]>
Commit | Line | Data |
---|---|---|
b8698a0f | 1 | /* Vectorizer Specific Loop Manipulations |
d1e082c2 | 2 | Copyright (C) 2003-2013 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" | |
78c60e3d | 25 | #include "dumpfile.h" |
ebfd146a IR |
26 | #include "tm.h" |
27 | #include "ggc.h" | |
28 | #include "tree.h" | |
29 | #include "basic-block.h" | |
cf835838 | 30 | #include "gimple-pretty-print.h" |
45b0be94 | 31 | #include "gimplify.h" |
442b4905 AM |
32 | #include "gimple-ssa.h" |
33 | #include "tree-cfg.h" | |
34 | #include "tree-phinodes.h" | |
35 | #include "ssa-iterators.h" | |
36 | #include "tree-ssanames.h" | |
e28030cf | 37 | #include "tree-ssa-loop-manip.h" |
442b4905 | 38 | #include "tree-into-ssa.h" |
7a300452 | 39 | #include "tree-ssa.h" |
7ee2468b | 40 | #include "tree-pass.h" |
ebfd146a | 41 | #include "cfgloop.h" |
718f9c0f | 42 | #include "diagnostic-core.h" |
ebfd146a IR |
43 | #include "tree-scalar-evolution.h" |
44 | #include "tree-vectorizer.h" | |
45 | #include "langhooks.h" | |
46 | ||
47 | /************************************************************************* | |
48 | Simple Loop Peeling Utilities | |
49 | ||
50 | Utilities to support loop peeling for vectorization purposes. | |
51 | *************************************************************************/ | |
52 | ||
53 | ||
54 | /* Renames the use *OP_P. */ | |
55 | ||
56 | static void | |
57 | rename_use_op (use_operand_p op_p) | |
58 | { | |
59 | tree new_name; | |
60 | ||
61 | if (TREE_CODE (USE_FROM_PTR (op_p)) != SSA_NAME) | |
62 | return; | |
63 | ||
64 | new_name = get_current_def (USE_FROM_PTR (op_p)); | |
65 | ||
66 | /* Something defined outside of the loop. */ | |
67 | if (!new_name) | |
68 | return; | |
69 | ||
70 | /* An ordinary ssa name defined in the loop. */ | |
71 | ||
72 | SET_USE (op_p, new_name); | |
73 | } | |
74 | ||
75 | ||
76 | /* Renames the variables in basic block BB. */ | |
77 | ||
2cfc56b9 | 78 | static void |
ebfd146a IR |
79 | rename_variables_in_bb (basic_block bb) |
80 | { | |
81 | gimple_stmt_iterator gsi; | |
82 | gimple stmt; | |
83 | use_operand_p use_p; | |
84 | ssa_op_iter iter; | |
85 | edge e; | |
86 | edge_iterator ei; | |
87 | struct loop *loop = bb->loop_father; | |
88 | ||
89 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
90 | { | |
91 | stmt = gsi_stmt (gsi); | |
92 | FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_ALL_USES) | |
93 | rename_use_op (use_p); | |
94 | } | |
95 | ||
2cfc56b9 | 96 | FOR_EACH_EDGE (e, ei, bb->preds) |
ebfd146a | 97 | { |
2cfc56b9 | 98 | if (!flow_bb_inside_loop_p (loop, e->src)) |
ebfd146a | 99 | continue; |
2cfc56b9 | 100 | for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
ebfd146a IR |
101 | rename_use_op (PHI_ARG_DEF_PTR_FROM_EDGE (gsi_stmt (gsi), e)); |
102 | } | |
103 | } | |
104 | ||
105 | ||
684f25f4 AO |
106 | typedef struct |
107 | { | |
108 | tree from, to; | |
109 | basic_block bb; | |
110 | } adjust_info; | |
111 | ||
684f25f4 AO |
112 | /* A stack of values to be adjusted in debug stmts. We have to |
113 | process them LIFO, so that the closest substitution applies. If we | |
114 | processed them FIFO, without the stack, we might substitute uses | |
115 | with a PHI DEF that would soon become non-dominant, and when we got | |
116 | to the suitable one, it wouldn't have anything to substitute any | |
117 | more. */ | |
ff4c81cc | 118 | static vec<adjust_info, va_heap> adjust_vec; |
684f25f4 AO |
119 | |
120 | /* Adjust any debug stmts that referenced AI->from values to use the | |
121 | loop-closed AI->to, if the references are dominated by AI->bb and | |
122 | not by the definition of AI->from. */ | |
123 | ||
124 | static void | |
125 | adjust_debug_stmts_now (adjust_info *ai) | |
126 | { | |
127 | basic_block bbphi = ai->bb; | |
128 | tree orig_def = ai->from; | |
129 | tree new_def = ai->to; | |
130 | imm_use_iterator imm_iter; | |
131 | gimple stmt; | |
132 | basic_block bbdef = gimple_bb (SSA_NAME_DEF_STMT (orig_def)); | |
133 | ||
134 | gcc_assert (dom_info_available_p (CDI_DOMINATORS)); | |
135 | ||
136 | /* Adjust any debug stmts that held onto non-loop-closed | |
137 | references. */ | |
138 | FOR_EACH_IMM_USE_STMT (stmt, imm_iter, orig_def) | |
139 | { | |
140 | use_operand_p use_p; | |
141 | basic_block bbuse; | |
142 | ||
143 | if (!is_gimple_debug (stmt)) | |
144 | continue; | |
145 | ||
146 | gcc_assert (gimple_debug_bind_p (stmt)); | |
147 | ||
148 | bbuse = gimple_bb (stmt); | |
149 | ||
150 | if ((bbuse == bbphi | |
151 | || dominated_by_p (CDI_DOMINATORS, bbuse, bbphi)) | |
152 | && !(bbuse == bbdef | |
153 | || dominated_by_p (CDI_DOMINATORS, bbuse, bbdef))) | |
154 | { | |
155 | if (new_def) | |
156 | FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter) | |
157 | SET_USE (use_p, new_def); | |
158 | else | |
159 | { | |
160 | gimple_debug_bind_reset_value (stmt); | |
161 | update_stmt (stmt); | |
162 | } | |
163 | } | |
164 | } | |
165 | } | |
166 | ||
167 | /* Adjust debug stmts as scheduled before. */ | |
168 | ||
169 | static void | |
170 | adjust_vec_debug_stmts (void) | |
171 | { | |
172 | if (!MAY_HAVE_DEBUG_STMTS) | |
173 | return; | |
174 | ||
9771b263 | 175 | gcc_assert (adjust_vec.exists ()); |
684f25f4 | 176 | |
9771b263 | 177 | while (!adjust_vec.is_empty ()) |
684f25f4 | 178 | { |
9771b263 DN |
179 | adjust_debug_stmts_now (&adjust_vec.last ()); |
180 | adjust_vec.pop (); | |
684f25f4 AO |
181 | } |
182 | ||
9771b263 | 183 | adjust_vec.release (); |
684f25f4 AO |
184 | } |
185 | ||
186 | /* Adjust any debug stmts that referenced FROM values to use the | |
187 | loop-closed TO, if the references are dominated by BB and not by | |
188 | the definition of FROM. If adjust_vec is non-NULL, adjustments | |
189 | will be postponed until adjust_vec_debug_stmts is called. */ | |
190 | ||
191 | static void | |
192 | adjust_debug_stmts (tree from, tree to, basic_block bb) | |
193 | { | |
194 | adjust_info ai; | |
195 | ||
a471762f RG |
196 | if (MAY_HAVE_DEBUG_STMTS |
197 | && TREE_CODE (from) == SSA_NAME | |
a52ca739 | 198 | && ! SSA_NAME_IS_DEFAULT_DEF (from) |
a471762f | 199 | && ! virtual_operand_p (from)) |
684f25f4 AO |
200 | { |
201 | ai.from = from; | |
202 | ai.to = to; | |
203 | ai.bb = bb; | |
204 | ||
9771b263 DN |
205 | if (adjust_vec.exists ()) |
206 | adjust_vec.safe_push (ai); | |
684f25f4 AO |
207 | else |
208 | adjust_debug_stmts_now (&ai); | |
209 | } | |
210 | } | |
211 | ||
212 | /* Change E's phi arg in UPDATE_PHI to NEW_DEF, and record information | |
213 | to adjust any debug stmts that referenced the old phi arg, | |
214 | presumably non-loop-closed references left over from other | |
215 | transformations. */ | |
216 | ||
217 | static void | |
218 | adjust_phi_and_debug_stmts (gimple update_phi, edge e, tree new_def) | |
219 | { | |
220 | tree orig_def = PHI_ARG_DEF_FROM_EDGE (update_phi, e); | |
221 | ||
222 | SET_PHI_ARG_DEF (update_phi, e->dest_idx, new_def); | |
223 | ||
224 | if (MAY_HAVE_DEBUG_STMTS) | |
225 | adjust_debug_stmts (orig_def, PHI_RESULT (update_phi), | |
226 | gimple_bb (update_phi)); | |
227 | } | |
228 | ||
ebfd146a | 229 | |
ebfd146a IR |
230 | /* Update PHI nodes for a guard of the LOOP. |
231 | ||
232 | Input: | |
233 | - LOOP, GUARD_EDGE: LOOP is a loop for which we added guard code that | |
234 | controls whether LOOP is to be executed. GUARD_EDGE is the edge that | |
235 | originates from the guard-bb, skips LOOP and reaches the (unique) exit | |
236 | bb of LOOP. This loop-exit-bb is an empty bb with one successor. | |
237 | We denote this bb NEW_MERGE_BB because before the guard code was added | |
238 | it had a single predecessor (the LOOP header), and now it became a merge | |
239 | point of two paths - the path that ends with the LOOP exit-edge, and | |
240 | the path that ends with GUARD_EDGE. | |
241 | - NEW_EXIT_BB: New basic block that is added by this function between LOOP | |
242 | and NEW_MERGE_BB. It is used to place loop-closed-ssa-form exit-phis. | |
243 | ||
244 | ===> The CFG before the guard-code was added: | |
245 | LOOP_header_bb: | |
246 | loop_body | |
247 | if (exit_loop) goto update_bb | |
248 | else goto LOOP_header_bb | |
249 | update_bb: | |
250 | ||
251 | ==> The CFG after the guard-code was added: | |
252 | guard_bb: | |
253 | if (LOOP_guard_condition) goto new_merge_bb | |
254 | else goto LOOP_header_bb | |
255 | LOOP_header_bb: | |
256 | loop_body | |
257 | if (exit_loop_condition) goto new_merge_bb | |
258 | else goto LOOP_header_bb | |
259 | new_merge_bb: | |
260 | goto update_bb | |
261 | update_bb: | |
262 | ||
263 | ==> The CFG after this function: | |
264 | guard_bb: | |
265 | if (LOOP_guard_condition) goto new_merge_bb | |
266 | else goto LOOP_header_bb | |
267 | LOOP_header_bb: | |
268 | loop_body | |
269 | if (exit_loop_condition) goto new_exit_bb | |
270 | else goto LOOP_header_bb | |
271 | new_exit_bb: | |
272 | new_merge_bb: | |
273 | goto update_bb | |
274 | update_bb: | |
275 | ||
276 | This function: | |
277 | 1. creates and updates the relevant phi nodes to account for the new | |
278 | incoming edge (GUARD_EDGE) into NEW_MERGE_BB. This involves: | |
279 | 1.1. Create phi nodes at NEW_MERGE_BB. | |
280 | 1.2. Update the phi nodes at the successor of NEW_MERGE_BB (denoted | |
281 | UPDATE_BB). UPDATE_BB was the exit-bb of LOOP before NEW_MERGE_BB | |
282 | 2. preserves loop-closed-ssa-form by creating the required phi nodes | |
283 | at the exit of LOOP (i.e, in NEW_EXIT_BB). | |
284 | ||
285 | There are two flavors to this function: | |
286 | ||
287 | slpeel_update_phi_nodes_for_guard1: | |
288 | Here the guard controls whether we enter or skip LOOP, where LOOP is a | |
289 | prolog_loop (loop1 below), and the new phis created in NEW_MERGE_BB are | |
290 | for variables that have phis in the loop header. | |
291 | ||
292 | slpeel_update_phi_nodes_for_guard2: | |
293 | Here the guard controls whether we enter or skip LOOP, where LOOP is an | |
294 | epilog_loop (loop2 below), and the new phis created in NEW_MERGE_BB are | |
295 | for variables that have phis in the loop exit. | |
296 | ||
297 | I.E., the overall structure is: | |
298 | ||
299 | loop1_preheader_bb: | |
300 | guard1 (goto loop1/merge1_bb) | |
301 | loop1 | |
302 | loop1_exit_bb: | |
303 | guard2 (goto merge1_bb/merge2_bb) | |
304 | merge1_bb | |
305 | loop2 | |
306 | loop2_exit_bb | |
307 | merge2_bb | |
308 | next_bb | |
309 | ||
310 | slpeel_update_phi_nodes_for_guard1 takes care of creating phis in | |
311 | loop1_exit_bb and merge1_bb. These are entry phis (phis for the vars | |
312 | that have phis in loop1->header). | |
313 | ||
314 | slpeel_update_phi_nodes_for_guard2 takes care of creating phis in | |
315 | loop2_exit_bb and merge2_bb. These are exit phis (phis for the vars | |
316 | that have phis in next_bb). It also adds some of these phis to | |
317 | loop1_exit_bb. | |
318 | ||
319 | slpeel_update_phi_nodes_for_guard1 is always called before | |
320 | slpeel_update_phi_nodes_for_guard2. They are both needed in order | |
321 | to create correct data-flow and loop-closed-ssa-form. | |
322 | ||
323 | Generally slpeel_update_phi_nodes_for_guard1 creates phis for variables | |
324 | that change between iterations of a loop (and therefore have a phi-node | |
325 | at the loop entry), whereas slpeel_update_phi_nodes_for_guard2 creates | |
b8698a0f L |
326 | phis for variables that are used out of the loop (and therefore have |
327 | loop-closed exit phis). Some variables may be both updated between | |
ebfd146a IR |
328 | iterations and used after the loop. This is why in loop1_exit_bb we |
329 | may need both entry_phis (created by slpeel_update_phi_nodes_for_guard1) | |
330 | and exit phis (created by slpeel_update_phi_nodes_for_guard2). | |
331 | ||
332 | - IS_NEW_LOOP: if IS_NEW_LOOP is true, then LOOP is a newly created copy of | |
333 | an original loop. i.e., we have: | |
334 | ||
335 | orig_loop | |
336 | guard_bb (goto LOOP/new_merge) | |
337 | new_loop <-- LOOP | |
338 | new_exit | |
339 | new_merge | |
340 | next_bb | |
341 | ||
342 | If IS_NEW_LOOP is false, then LOOP is an original loop, in which case we | |
343 | have: | |
344 | ||
345 | new_loop | |
346 | guard_bb (goto LOOP/new_merge) | |
347 | orig_loop <-- LOOP | |
348 | new_exit | |
349 | new_merge | |
350 | next_bb | |
351 | ||
352 | The SSA names defined in the original loop have a current | |
353 | reaching definition that that records the corresponding new | |
354 | ssa-name used in the new duplicated loop copy. | |
355 | */ | |
356 | ||
357 | /* Function slpeel_update_phi_nodes_for_guard1 | |
b8698a0f | 358 | |
ebfd146a IR |
359 | Input: |
360 | - GUARD_EDGE, LOOP, IS_NEW_LOOP, NEW_EXIT_BB - as explained above. | |
361 | - DEFS - a bitmap of ssa names to mark new names for which we recorded | |
b8698a0f L |
362 | information. |
363 | ||
ebfd146a IR |
364 | In the context of the overall structure, we have: |
365 | ||
b8698a0f | 366 | loop1_preheader_bb: |
ebfd146a IR |
367 | guard1 (goto loop1/merge1_bb) |
368 | LOOP-> loop1 | |
369 | loop1_exit_bb: | |
370 | guard2 (goto merge1_bb/merge2_bb) | |
371 | merge1_bb | |
372 | loop2 | |
373 | loop2_exit_bb | |
374 | merge2_bb | |
375 | next_bb | |
376 | ||
377 | For each name updated between loop iterations (i.e - for each name that has | |
378 | an entry (loop-header) phi in LOOP) we create a new phi in: | |
379 | 1. merge1_bb (to account for the edge from guard1) | |
380 | 2. loop1_exit_bb (an exit-phi to keep LOOP in loop-closed form) | |
381 | */ | |
382 | ||
383 | static void | |
384 | slpeel_update_phi_nodes_for_guard1 (edge guard_edge, struct loop *loop, | |
c334023f | 385 | bool is_new_loop, basic_block *new_exit_bb) |
ebfd146a IR |
386 | { |
387 | gimple orig_phi, new_phi; | |
388 | gimple update_phi, update_phi2; | |
389 | tree guard_arg, loop_arg; | |
390 | basic_block new_merge_bb = guard_edge->dest; | |
391 | edge e = EDGE_SUCC (new_merge_bb, 0); | |
392 | basic_block update_bb = e->dest; | |
393 | basic_block orig_bb = loop->header; | |
394 | edge new_exit_e; | |
395 | tree current_new_name; | |
ebfd146a IR |
396 | gimple_stmt_iterator gsi_orig, gsi_update; |
397 | ||
398 | /* Create new bb between loop and new_merge_bb. */ | |
399 | *new_exit_bb = split_edge (single_exit (loop)); | |
400 | ||
401 | new_exit_e = EDGE_SUCC (*new_exit_bb, 0); | |
402 | ||
403 | for (gsi_orig = gsi_start_phis (orig_bb), | |
404 | gsi_update = gsi_start_phis (update_bb); | |
405 | !gsi_end_p (gsi_orig) && !gsi_end_p (gsi_update); | |
406 | gsi_next (&gsi_orig), gsi_next (&gsi_update)) | |
407 | { | |
e20f6b4b | 408 | source_location loop_locus, guard_locus; |
070ecdfd | 409 | tree new_res; |
ebfd146a IR |
410 | orig_phi = gsi_stmt (gsi_orig); |
411 | update_phi = gsi_stmt (gsi_update); | |
412 | ||
ebfd146a IR |
413 | /** 1. Handle new-merge-point phis **/ |
414 | ||
415 | /* 1.1. Generate new phi node in NEW_MERGE_BB: */ | |
070ecdfd RG |
416 | new_res = copy_ssa_name (PHI_RESULT (orig_phi), NULL); |
417 | new_phi = create_phi_node (new_res, new_merge_bb); | |
ebfd146a IR |
418 | |
419 | /* 1.2. NEW_MERGE_BB has two incoming edges: GUARD_EDGE and the exit-edge | |
420 | of LOOP. Set the two phi args in NEW_PHI for these edges: */ | |
421 | loop_arg = PHI_ARG_DEF_FROM_EDGE (orig_phi, EDGE_SUCC (loop->latch, 0)); | |
b8698a0f L |
422 | loop_locus = gimple_phi_arg_location_from_edge (orig_phi, |
423 | EDGE_SUCC (loop->latch, | |
f5045c96 | 424 | 0)); |
ebfd146a | 425 | guard_arg = PHI_ARG_DEF_FROM_EDGE (orig_phi, loop_preheader_edge (loop)); |
b8698a0f L |
426 | guard_locus |
427 | = gimple_phi_arg_location_from_edge (orig_phi, | |
f5045c96 | 428 | loop_preheader_edge (loop)); |
ebfd146a | 429 | |
9e227d60 DC |
430 | add_phi_arg (new_phi, loop_arg, new_exit_e, loop_locus); |
431 | add_phi_arg (new_phi, guard_arg, guard_edge, guard_locus); | |
ebfd146a IR |
432 | |
433 | /* 1.3. Update phi in successor block. */ | |
434 | gcc_assert (PHI_ARG_DEF_FROM_EDGE (update_phi, e) == loop_arg | |
435 | || PHI_ARG_DEF_FROM_EDGE (update_phi, e) == guard_arg); | |
684f25f4 | 436 | adjust_phi_and_debug_stmts (update_phi, e, PHI_RESULT (new_phi)); |
ebfd146a IR |
437 | update_phi2 = new_phi; |
438 | ||
439 | ||
440 | /** 2. Handle loop-closed-ssa-form phis **/ | |
441 | ||
ea057359 | 442 | if (virtual_operand_p (PHI_RESULT (orig_phi))) |
ebfd146a IR |
443 | continue; |
444 | ||
445 | /* 2.1. Generate new phi node in NEW_EXIT_BB: */ | |
070ecdfd RG |
446 | new_res = copy_ssa_name (PHI_RESULT (orig_phi), NULL); |
447 | new_phi = create_phi_node (new_res, *new_exit_bb); | |
ebfd146a IR |
448 | |
449 | /* 2.2. NEW_EXIT_BB has one incoming edge: the exit-edge of the loop. */ | |
9e227d60 | 450 | add_phi_arg (new_phi, loop_arg, single_exit (loop), loop_locus); |
ebfd146a IR |
451 | |
452 | /* 2.3. Update phi in successor of NEW_EXIT_BB: */ | |
453 | gcc_assert (PHI_ARG_DEF_FROM_EDGE (update_phi2, new_exit_e) == loop_arg); | |
684f25f4 AO |
454 | adjust_phi_and_debug_stmts (update_phi2, new_exit_e, |
455 | PHI_RESULT (new_phi)); | |
ebfd146a IR |
456 | |
457 | /* 2.4. Record the newly created name with set_current_def. | |
458 | We want to find a name such that | |
459 | name = get_current_def (orig_loop_name) | |
460 | and to set its current definition as follows: | |
461 | set_current_def (name, new_phi_name) | |
462 | ||
463 | If LOOP is a new loop then loop_arg is already the name we're | |
464 | looking for. If LOOP is the original loop, then loop_arg is | |
465 | the orig_loop_name and the relevant name is recorded in its | |
466 | current reaching definition. */ | |
467 | if (is_new_loop) | |
468 | current_new_name = loop_arg; | |
469 | else | |
470 | { | |
471 | current_new_name = get_current_def (loop_arg); | |
472 | /* current_def is not available only if the variable does not | |
473 | change inside the loop, in which case we also don't care | |
474 | about recording a current_def for it because we won't be | |
475 | trying to create loop-exit-phis for it. */ | |
476 | if (!current_new_name) | |
477 | continue; | |
478 | } | |
479 | gcc_assert (get_current_def (current_new_name) == NULL_TREE); | |
480 | ||
481 | set_current_def (current_new_name, PHI_RESULT (new_phi)); | |
ebfd146a IR |
482 | } |
483 | } | |
484 | ||
485 | ||
486 | /* Function slpeel_update_phi_nodes_for_guard2 | |
487 | ||
488 | Input: | |
489 | - GUARD_EDGE, LOOP, IS_NEW_LOOP, NEW_EXIT_BB - as explained above. | |
490 | ||
491 | In the context of the overall structure, we have: | |
492 | ||
b8698a0f | 493 | loop1_preheader_bb: |
ebfd146a IR |
494 | guard1 (goto loop1/merge1_bb) |
495 | loop1 | |
b8698a0f | 496 | loop1_exit_bb: |
ebfd146a IR |
497 | guard2 (goto merge1_bb/merge2_bb) |
498 | merge1_bb | |
499 | LOOP-> loop2 | |
500 | loop2_exit_bb | |
501 | merge2_bb | |
502 | next_bb | |
503 | ||
504 | For each name used out side the loop (i.e - for each name that has an exit | |
505 | phi in next_bb) we create a new phi in: | |
b8698a0f | 506 | 1. merge2_bb (to account for the edge from guard_bb) |
ebfd146a IR |
507 | 2. loop2_exit_bb (an exit-phi to keep LOOP in loop-closed form) |
508 | 3. guard2 bb (an exit phi to keep the preceding loop in loop-closed form), | |
509 | if needed (if it wasn't handled by slpeel_update_phis_nodes_for_phi1). | |
510 | */ | |
511 | ||
512 | static void | |
513 | slpeel_update_phi_nodes_for_guard2 (edge guard_edge, struct loop *loop, | |
514 | bool is_new_loop, basic_block *new_exit_bb) | |
515 | { | |
516 | gimple orig_phi, new_phi; | |
517 | gimple update_phi, update_phi2; | |
518 | tree guard_arg, loop_arg; | |
519 | basic_block new_merge_bb = guard_edge->dest; | |
520 | edge e = EDGE_SUCC (new_merge_bb, 0); | |
521 | basic_block update_bb = e->dest; | |
522 | edge new_exit_e; | |
523 | tree orig_def, orig_def_new_name; | |
524 | tree new_name, new_name2; | |
525 | tree arg; | |
526 | gimple_stmt_iterator gsi; | |
527 | ||
528 | /* Create new bb between loop and new_merge_bb. */ | |
529 | *new_exit_bb = split_edge (single_exit (loop)); | |
530 | ||
531 | new_exit_e = EDGE_SUCC (*new_exit_bb, 0); | |
532 | ||
533 | for (gsi = gsi_start_phis (update_bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
534 | { | |
070ecdfd | 535 | tree new_res; |
ebfd146a IR |
536 | update_phi = gsi_stmt (gsi); |
537 | orig_phi = update_phi; | |
538 | orig_def = PHI_ARG_DEF_FROM_EDGE (orig_phi, e); | |
539 | /* This loop-closed-phi actually doesn't represent a use | |
b8698a0f | 540 | out of the loop - the phi arg is a constant. */ |
ebfd146a IR |
541 | if (TREE_CODE (orig_def) != SSA_NAME) |
542 | continue; | |
543 | orig_def_new_name = get_current_def (orig_def); | |
544 | arg = NULL_TREE; | |
545 | ||
546 | /** 1. Handle new-merge-point phis **/ | |
547 | ||
548 | /* 1.1. Generate new phi node in NEW_MERGE_BB: */ | |
070ecdfd RG |
549 | new_res = copy_ssa_name (PHI_RESULT (orig_phi), NULL); |
550 | new_phi = create_phi_node (new_res, new_merge_bb); | |
ebfd146a IR |
551 | |
552 | /* 1.2. NEW_MERGE_BB has two incoming edges: GUARD_EDGE and the exit-edge | |
553 | of LOOP. Set the two PHI args in NEW_PHI for these edges: */ | |
554 | new_name = orig_def; | |
555 | new_name2 = NULL_TREE; | |
556 | if (orig_def_new_name) | |
557 | { | |
558 | new_name = orig_def_new_name; | |
559 | /* Some variables have both loop-entry-phis and loop-exit-phis. | |
560 | Such variables were given yet newer names by phis placed in | |
561 | guard_bb by slpeel_update_phi_nodes_for_guard1. I.e: | |
562 | new_name2 = get_current_def (get_current_def (orig_name)). */ | |
563 | new_name2 = get_current_def (new_name); | |
564 | } | |
b8698a0f | 565 | |
ebfd146a IR |
566 | if (is_new_loop) |
567 | { | |
568 | guard_arg = orig_def; | |
569 | loop_arg = new_name; | |
570 | } | |
571 | else | |
572 | { | |
573 | guard_arg = new_name; | |
574 | loop_arg = orig_def; | |
575 | } | |
576 | if (new_name2) | |
577 | guard_arg = new_name2; | |
b8698a0f | 578 | |
9e227d60 DC |
579 | add_phi_arg (new_phi, loop_arg, new_exit_e, UNKNOWN_LOCATION); |
580 | add_phi_arg (new_phi, guard_arg, guard_edge, UNKNOWN_LOCATION); | |
ebfd146a IR |
581 | |
582 | /* 1.3. Update phi in successor block. */ | |
583 | gcc_assert (PHI_ARG_DEF_FROM_EDGE (update_phi, e) == orig_def); | |
684f25f4 | 584 | adjust_phi_and_debug_stmts (update_phi, e, PHI_RESULT (new_phi)); |
ebfd146a IR |
585 | update_phi2 = new_phi; |
586 | ||
587 | ||
588 | /** 2. Handle loop-closed-ssa-form phis **/ | |
589 | ||
590 | /* 2.1. Generate new phi node in NEW_EXIT_BB: */ | |
070ecdfd RG |
591 | new_res = copy_ssa_name (PHI_RESULT (orig_phi), NULL); |
592 | new_phi = create_phi_node (new_res, *new_exit_bb); | |
ebfd146a IR |
593 | |
594 | /* 2.2. NEW_EXIT_BB has one incoming edge: the exit-edge of the loop. */ | |
9e227d60 | 595 | add_phi_arg (new_phi, loop_arg, single_exit (loop), UNKNOWN_LOCATION); |
ebfd146a IR |
596 | |
597 | /* 2.3. Update phi in successor of NEW_EXIT_BB: */ | |
598 | gcc_assert (PHI_ARG_DEF_FROM_EDGE (update_phi2, new_exit_e) == loop_arg); | |
684f25f4 AO |
599 | adjust_phi_and_debug_stmts (update_phi2, new_exit_e, |
600 | PHI_RESULT (new_phi)); | |
ebfd146a IR |
601 | |
602 | ||
603 | /** 3. Handle loop-closed-ssa-form phis for first loop **/ | |
604 | ||
605 | /* 3.1. Find the relevant names that need an exit-phi in | |
606 | GUARD_BB, i.e. names for which | |
607 | slpeel_update_phi_nodes_for_guard1 had not already created a | |
608 | phi node. This is the case for names that are used outside | |
609 | the loop (and therefore need an exit phi) but are not updated | |
610 | across loop iterations (and therefore don't have a | |
611 | loop-header-phi). | |
612 | ||
613 | slpeel_update_phi_nodes_for_guard1 is responsible for | |
614 | creating loop-exit phis in GUARD_BB for names that have a | |
615 | loop-header-phi. When such a phi is created we also record | |
616 | the new name in its current definition. If this new name | |
617 | exists, then guard_arg was set to this new name (see 1.2 | |
618 | above). Therefore, if guard_arg is not this new name, this | |
619 | is an indication that an exit-phi in GUARD_BB was not yet | |
620 | created, so we take care of it here. */ | |
621 | if (guard_arg == new_name2) | |
622 | continue; | |
623 | arg = guard_arg; | |
624 | ||
625 | /* 3.2. Generate new phi node in GUARD_BB: */ | |
070ecdfd RG |
626 | new_res = copy_ssa_name (PHI_RESULT (orig_phi), NULL); |
627 | new_phi = create_phi_node (new_res, guard_edge->src); | |
ebfd146a IR |
628 | |
629 | /* 3.3. GUARD_BB has one incoming edge: */ | |
630 | gcc_assert (EDGE_COUNT (guard_edge->src->preds) == 1); | |
b8698a0f | 631 | add_phi_arg (new_phi, arg, EDGE_PRED (guard_edge->src, 0), |
9e227d60 | 632 | UNKNOWN_LOCATION); |
ebfd146a IR |
633 | |
634 | /* 3.4. Update phi in successor of GUARD_BB: */ | |
635 | gcc_assert (PHI_ARG_DEF_FROM_EDGE (update_phi2, guard_edge) | |
636 | == guard_arg); | |
684f25f4 AO |
637 | adjust_phi_and_debug_stmts (update_phi2, guard_edge, |
638 | PHI_RESULT (new_phi)); | |
ebfd146a IR |
639 | } |
640 | } | |
641 | ||
642 | ||
643 | /* Make the LOOP iterate NITERS times. This is done by adding a new IV | |
644 | that starts at zero, increases by one and its limit is NITERS. | |
645 | ||
646 | Assumption: the exit-condition of LOOP is the last stmt in the loop. */ | |
647 | ||
648 | void | |
649 | slpeel_make_loop_iterate_ntimes (struct loop *loop, tree niters) | |
650 | { | |
651 | tree indx_before_incr, indx_after_incr; | |
652 | gimple cond_stmt; | |
653 | gimple orig_cond; | |
654 | edge exit_edge = single_exit (loop); | |
655 | gimple_stmt_iterator loop_cond_gsi; | |
656 | gimple_stmt_iterator incr_gsi; | |
657 | bool insert_after; | |
658 | tree init = build_int_cst (TREE_TYPE (niters), 0); | |
659 | tree step = build_int_cst (TREE_TYPE (niters), 1); | |
660 | LOC loop_loc; | |
661 | enum tree_code code; | |
662 | ||
663 | orig_cond = get_loop_exit_condition (loop); | |
664 | gcc_assert (orig_cond); | |
665 | loop_cond_gsi = gsi_for_stmt (orig_cond); | |
666 | ||
667 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
668 | create_iv (init, step, NULL_TREE, loop, | |
669 | &incr_gsi, insert_after, &indx_before_incr, &indx_after_incr); | |
670 | ||
671 | indx_after_incr = force_gimple_operand_gsi (&loop_cond_gsi, indx_after_incr, | |
672 | true, NULL_TREE, true, | |
673 | GSI_SAME_STMT); | |
674 | niters = force_gimple_operand_gsi (&loop_cond_gsi, niters, true, NULL_TREE, | |
675 | true, GSI_SAME_STMT); | |
676 | ||
677 | code = (exit_edge->flags & EDGE_TRUE_VALUE) ? GE_EXPR : LT_EXPR; | |
678 | cond_stmt = gimple_build_cond (code, indx_after_incr, niters, NULL_TREE, | |
679 | NULL_TREE); | |
680 | ||
681 | gsi_insert_before (&loop_cond_gsi, cond_stmt, GSI_SAME_STMT); | |
682 | ||
683 | /* Remove old loop exit test: */ | |
684 | gsi_remove (&loop_cond_gsi, true); | |
6f723d33 | 685 | free_stmt_vec_info (orig_cond); |
ebfd146a IR |
686 | |
687 | loop_loc = find_loop_location (loop); | |
73fbfcad | 688 | if (dump_enabled_p ()) |
ebfd146a | 689 | { |
84df911b DC |
690 | if (LOCATION_LOCUS (loop_loc) != UNKNOWN_LOC) |
691 | dump_printf (MSG_NOTE, "\nloop at %s:%d: ", LOC_FILE (loop_loc), | |
692 | LOC_LINE (loop_loc)); | |
78c60e3d | 693 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, cond_stmt, 0); |
e645e942 | 694 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a | 695 | } |
ebfd146a IR |
696 | loop->nb_iterations = niters; |
697 | } | |
698 | ||
699 | ||
b8698a0f | 700 | /* Given LOOP this function generates a new copy of it and puts it |
ebfd146a IR |
701 | on E which is either the entry or exit of LOOP. */ |
702 | ||
703 | struct loop * | |
704 | slpeel_tree_duplicate_loop_to_edge_cfg (struct loop *loop, edge e) | |
705 | { | |
706 | struct loop *new_loop; | |
707 | basic_block *new_bbs, *bbs; | |
708 | bool at_exit; | |
709 | bool was_imm_dom; | |
b8698a0f | 710 | basic_block exit_dest; |
ebfd146a | 711 | edge exit, new_exit; |
ebfd146a | 712 | |
2cfc56b9 RB |
713 | exit = single_exit (loop); |
714 | at_exit = (e == exit); | |
ebfd146a IR |
715 | if (!at_exit && e != loop_preheader_edge (loop)) |
716 | return NULL; | |
717 | ||
2cfc56b9 RB |
718 | bbs = XNEWVEC (basic_block, loop->num_nodes + 1); |
719 | get_loop_body_with_size (loop, bbs, loop->num_nodes); | |
ebfd146a IR |
720 | |
721 | /* Check whether duplication is possible. */ | |
722 | if (!can_copy_bbs_p (bbs, loop->num_nodes)) | |
723 | { | |
724 | free (bbs); | |
725 | return NULL; | |
726 | } | |
727 | ||
728 | /* Generate new loop structure. */ | |
729 | new_loop = duplicate_loop (loop, loop_outer (loop)); | |
2cfc56b9 | 730 | duplicate_subloops (loop, new_loop); |
ebfd146a | 731 | |
2cfc56b9 | 732 | exit_dest = exit->dest; |
b8698a0f L |
733 | was_imm_dom = (get_immediate_dominator (CDI_DOMINATORS, |
734 | exit_dest) == loop->header ? | |
ebfd146a IR |
735 | true : false); |
736 | ||
2cfc56b9 RB |
737 | /* Also copy the pre-header, this avoids jumping through hoops to |
738 | duplicate the loop entry PHI arguments. Create an empty | |
739 | pre-header unconditionally for this. */ | |
740 | basic_block preheader = split_edge (loop_preheader_edge (loop)); | |
741 | edge entry_e = single_pred_edge (preheader); | |
742 | bbs[loop->num_nodes] = preheader; | |
743 | new_bbs = XNEWVEC (basic_block, loop->num_nodes + 1); | |
ebfd146a | 744 | |
2cfc56b9 | 745 | copy_bbs (bbs, loop->num_nodes + 1, new_bbs, |
ebfd146a | 746 | &exit, 1, &new_exit, NULL, |
f14540b6 | 747 | e->src, true); |
2cfc56b9 | 748 | basic_block new_preheader = new_bbs[loop->num_nodes]; |
ebfd146a | 749 | |
2cfc56b9 | 750 | add_phi_args_after_copy (new_bbs, loop->num_nodes + 1, NULL); |
b8698a0f | 751 | |
ebfd146a IR |
752 | if (at_exit) /* Add the loop copy at exit. */ |
753 | { | |
2cfc56b9 RB |
754 | redirect_edge_and_branch_force (e, new_preheader); |
755 | flush_pending_stmts (e); | |
756 | set_immediate_dominator (CDI_DOMINATORS, new_preheader, e->src); | |
ebfd146a IR |
757 | if (was_imm_dom) |
758 | set_immediate_dominator (CDI_DOMINATORS, exit_dest, new_loop->header); | |
2cfc56b9 RB |
759 | |
760 | /* And remove the non-necessary forwarder again. Keep the other | |
761 | one so we have a proper pre-header for the loop at the exit edge. */ | |
762 | redirect_edge_pred (single_succ_edge (preheader), single_pred (preheader)); | |
763 | delete_basic_block (preheader); | |
764 | set_immediate_dominator (CDI_DOMINATORS, loop->header, | |
765 | loop_preheader_edge (loop)->src); | |
ebfd146a IR |
766 | } |
767 | else /* Add the copy at entry. */ | |
768 | { | |
2cfc56b9 RB |
769 | redirect_edge_and_branch_force (entry_e, new_preheader); |
770 | flush_pending_stmts (entry_e); | |
771 | set_immediate_dominator (CDI_DOMINATORS, new_preheader, entry_e->src); | |
772 | ||
773 | redirect_edge_and_branch_force (new_exit, preheader); | |
774 | flush_pending_stmts (new_exit); | |
775 | set_immediate_dominator (CDI_DOMINATORS, preheader, new_exit->src); | |
776 | ||
777 | /* And remove the non-necessary forwarder again. Keep the other | |
778 | one so we have a proper pre-header for the loop at the exit edge. */ | |
779 | redirect_edge_pred (single_succ_edge (new_preheader), single_pred (new_preheader)); | |
780 | delete_basic_block (new_preheader); | |
781 | set_immediate_dominator (CDI_DOMINATORS, new_loop->header, | |
782 | loop_preheader_edge (new_loop)->src); | |
ebfd146a IR |
783 | } |
784 | ||
2cfc56b9 RB |
785 | for (unsigned i = 0; i < loop->num_nodes+1; i++) |
786 | rename_variables_in_bb (new_bbs[i]); | |
787 | ||
ebfd146a IR |
788 | free (new_bbs); |
789 | free (bbs); | |
790 | ||
2cfc56b9 RB |
791 | #ifdef ENABLE_CHECKING |
792 | verify_dominators (CDI_DOMINATORS); | |
793 | #endif | |
794 | ||
ebfd146a IR |
795 | return new_loop; |
796 | } | |
797 | ||
798 | ||
799 | /* Given the condition statement COND, put it as the last statement | |
800 | of GUARD_BB; EXIT_BB is the basic block to skip the loop; | |
b8698a0f | 801 | Assumes that this is the single exit of the guarded loop. |
86290011 | 802 | Returns the skip edge, inserts new stmts on the COND_EXPR_STMT_LIST. */ |
ebfd146a IR |
803 | |
804 | static edge | |
86290011 RG |
805 | slpeel_add_loop_guard (basic_block guard_bb, tree cond, |
806 | gimple_seq cond_expr_stmt_list, | |
e78410bf JH |
807 | basic_block exit_bb, basic_block dom_bb, |
808 | int probability) | |
ebfd146a IR |
809 | { |
810 | gimple_stmt_iterator gsi; | |
811 | edge new_e, enter_e; | |
812 | gimple cond_stmt; | |
813 | gimple_seq gimplify_stmt_list = NULL; | |
814 | ||
815 | enter_e = EDGE_SUCC (guard_bb, 0); | |
816 | enter_e->flags &= ~EDGE_FALLTHRU; | |
817 | enter_e->flags |= EDGE_FALSE_VALUE; | |
818 | gsi = gsi_last_bb (guard_bb); | |
819 | ||
f7a06a98 RG |
820 | cond = force_gimple_operand_1 (cond, &gimplify_stmt_list, is_gimple_condexpr, |
821 | NULL_TREE); | |
86290011 RG |
822 | if (gimplify_stmt_list) |
823 | gimple_seq_add_seq (&cond_expr_stmt_list, gimplify_stmt_list); | |
f7a06a98 | 824 | cond_stmt = gimple_build_cond_from_tree (cond, NULL_TREE, NULL_TREE); |
86290011 RG |
825 | if (cond_expr_stmt_list) |
826 | gsi_insert_seq_after (&gsi, cond_expr_stmt_list, GSI_NEW_STMT); | |
ebfd146a IR |
827 | |
828 | gsi = gsi_last_bb (guard_bb); | |
829 | gsi_insert_after (&gsi, cond_stmt, GSI_NEW_STMT); | |
830 | ||
831 | /* Add new edge to connect guard block to the merge/loop-exit block. */ | |
832 | new_e = make_edge (guard_bb, exit_bb, EDGE_TRUE_VALUE); | |
e78410bf JH |
833 | |
834 | new_e->count = guard_bb->count; | |
835 | new_e->probability = probability; | |
836 | new_e->count = apply_probability (enter_e->count, probability); | |
837 | enter_e->count -= new_e->count; | |
838 | enter_e->probability = inverse_probability (probability); | |
ebfd146a IR |
839 | set_immediate_dominator (CDI_DOMINATORS, exit_bb, dom_bb); |
840 | return new_e; | |
841 | } | |
842 | ||
843 | ||
844 | /* This function verifies that the following restrictions apply to LOOP: | |
845 | (1) it is innermost | |
846 | (2) it consists of exactly 2 basic blocks - header, and an empty latch. | |
847 | (3) it is single entry, single exit | |
848 | (4) its exit condition is the last stmt in the header | |
849 | (5) E is the entry/exit edge of LOOP. | |
850 | */ | |
851 | ||
852 | bool | |
853 | slpeel_can_duplicate_loop_p (const struct loop *loop, const_edge e) | |
854 | { | |
855 | edge exit_e = single_exit (loop); | |
856 | edge entry_e = loop_preheader_edge (loop); | |
857 | gimple orig_cond = get_loop_exit_condition (loop); | |
858 | gimple_stmt_iterator loop_exit_gsi = gsi_last_bb (exit_e->src); | |
859 | ||
ebfd146a IR |
860 | if (loop->inner |
861 | /* All loops have an outer scope; the only case loop->outer is NULL is for | |
862 | the function itself. */ | |
863 | || !loop_outer (loop) | |
864 | || loop->num_nodes != 2 | |
865 | || !empty_block_p (loop->latch) | |
866 | || !single_exit (loop) | |
867 | /* Verify that new loop exit condition can be trivially modified. */ | |
868 | || (!orig_cond || orig_cond != gsi_stmt (loop_exit_gsi)) | |
869 | || (e != exit_e && e != entry_e)) | |
870 | return false; | |
871 | ||
872 | return true; | |
873 | } | |
874 | ||
875 | #ifdef ENABLE_CHECKING | |
876 | static void | |
877 | slpeel_verify_cfg_after_peeling (struct loop *first_loop, | |
878 | struct loop *second_loop) | |
879 | { | |
880 | basic_block loop1_exit_bb = single_exit (first_loop)->dest; | |
881 | basic_block loop2_entry_bb = loop_preheader_edge (second_loop)->src; | |
882 | basic_block loop1_entry_bb = loop_preheader_edge (first_loop)->src; | |
883 | ||
884 | /* A guard that controls whether the second_loop is to be executed or skipped | |
885 | is placed in first_loop->exit. first_loop->exit therefore has two | |
886 | successors - one is the preheader of second_loop, and the other is a bb | |
887 | after second_loop. | |
888 | */ | |
889 | gcc_assert (EDGE_COUNT (loop1_exit_bb->succs) == 2); | |
b8698a0f | 890 | |
ebfd146a IR |
891 | /* 1. Verify that one of the successors of first_loop->exit is the preheader |
892 | of second_loop. */ | |
b8698a0f | 893 | |
ebfd146a IR |
894 | /* The preheader of new_loop is expected to have two predecessors: |
895 | first_loop->exit and the block that precedes first_loop. */ | |
896 | ||
b8698a0f | 897 | gcc_assert (EDGE_COUNT (loop2_entry_bb->preds) == 2 |
ebfd146a IR |
898 | && ((EDGE_PRED (loop2_entry_bb, 0)->src == loop1_exit_bb |
899 | && EDGE_PRED (loop2_entry_bb, 1)->src == loop1_entry_bb) | |
900 | || (EDGE_PRED (loop2_entry_bb, 1)->src == loop1_exit_bb | |
901 | && EDGE_PRED (loop2_entry_bb, 0)->src == loop1_entry_bb))); | |
b8698a0f | 902 | |
ebfd146a IR |
903 | /* Verify that the other successor of first_loop->exit is after the |
904 | second_loop. */ | |
905 | /* TODO */ | |
906 | } | |
907 | #endif | |
908 | ||
909 | /* If the run time cost model check determines that vectorization is | |
910 | not profitable and hence scalar loop should be generated then set | |
911 | FIRST_NITERS to prologue peeled iterations. This will allow all the | |
912 | iterations to be executed in the prologue peeled scalar loop. */ | |
913 | ||
914 | static void | |
915 | set_prologue_iterations (basic_block bb_before_first_loop, | |
5d2eb24b | 916 | tree *first_niters, |
ebfd146a | 917 | struct loop *loop, |
e78410bf JH |
918 | unsigned int th, |
919 | int probability) | |
ebfd146a IR |
920 | { |
921 | edge e; | |
922 | basic_block cond_bb, then_bb; | |
923 | tree var, prologue_after_cost_adjust_name; | |
924 | gimple_stmt_iterator gsi; | |
925 | gimple newphi; | |
926 | edge e_true, e_false, e_fallthru; | |
927 | gimple cond_stmt; | |
f7a06a98 | 928 | gimple_seq stmts = NULL; |
ebfd146a | 929 | tree cost_pre_condition = NULL_TREE; |
b8698a0f | 930 | tree scalar_loop_iters = |
ebfd146a IR |
931 | unshare_expr (LOOP_VINFO_NITERS_UNCHANGED (loop_vec_info_for_loop (loop))); |
932 | ||
933 | e = single_pred_edge (bb_before_first_loop); | |
c3284718 | 934 | cond_bb = split_edge (e); |
ebfd146a IR |
935 | |
936 | e = single_pred_edge (bb_before_first_loop); | |
c3284718 | 937 | then_bb = split_edge (e); |
ebfd146a IR |
938 | set_immediate_dominator (CDI_DOMINATORS, then_bb, cond_bb); |
939 | ||
940 | e_false = make_single_succ_edge (cond_bb, bb_before_first_loop, | |
941 | EDGE_FALSE_VALUE); | |
942 | set_immediate_dominator (CDI_DOMINATORS, bb_before_first_loop, cond_bb); | |
943 | ||
944 | e_true = EDGE_PRED (then_bb, 0); | |
945 | e_true->flags &= ~EDGE_FALLTHRU; | |
946 | e_true->flags |= EDGE_TRUE_VALUE; | |
947 | ||
e78410bf JH |
948 | e_true->probability = probability; |
949 | e_false->probability = inverse_probability (probability); | |
950 | e_true->count = apply_probability (cond_bb->count, probability); | |
951 | e_false->count = cond_bb->count - e_true->count; | |
952 | then_bb->frequency = EDGE_FREQUENCY (e_true); | |
953 | then_bb->count = e_true->count; | |
954 | ||
ebfd146a | 955 | e_fallthru = EDGE_SUCC (then_bb, 0); |
e78410bf | 956 | e_fallthru->count = then_bb->count; |
ebfd146a | 957 | |
f7a06a98 | 958 | gsi = gsi_last_bb (cond_bb); |
ebfd146a | 959 | cost_pre_condition = |
b8698a0f | 960 | fold_build2 (LE_EXPR, boolean_type_node, scalar_loop_iters, |
ebfd146a IR |
961 | build_int_cst (TREE_TYPE (scalar_loop_iters), th)); |
962 | cost_pre_condition = | |
f7a06a98 RG |
963 | force_gimple_operand_gsi_1 (&gsi, cost_pre_condition, is_gimple_condexpr, |
964 | NULL_TREE, false, GSI_CONTINUE_LINKING); | |
965 | cond_stmt = gimple_build_cond_from_tree (cost_pre_condition, | |
966 | NULL_TREE, NULL_TREE); | |
ebfd146a | 967 | gsi_insert_after (&gsi, cond_stmt, GSI_NEW_STMT); |
b8698a0f | 968 | |
ebfd146a IR |
969 | var = create_tmp_var (TREE_TYPE (scalar_loop_iters), |
970 | "prologue_after_cost_adjust"); | |
b8698a0f | 971 | prologue_after_cost_adjust_name = |
ebfd146a IR |
972 | force_gimple_operand (scalar_loop_iters, &stmts, false, var); |
973 | ||
974 | gsi = gsi_last_bb (then_bb); | |
975 | if (stmts) | |
976 | gsi_insert_seq_after (&gsi, stmts, GSI_NEW_STMT); | |
977 | ||
978 | newphi = create_phi_node (var, bb_before_first_loop); | |
b8698a0f | 979 | add_phi_arg (newphi, prologue_after_cost_adjust_name, e_fallthru, |
9e227d60 DC |
980 | UNKNOWN_LOCATION); |
981 | add_phi_arg (newphi, *first_niters, e_false, UNKNOWN_LOCATION); | |
ebfd146a | 982 | |
5d2eb24b | 983 | *first_niters = PHI_RESULT (newphi); |
ebfd146a IR |
984 | } |
985 | ||
ebfd146a IR |
986 | /* Function slpeel_tree_peel_loop_to_edge. |
987 | ||
988 | Peel the first (last) iterations of LOOP into a new prolog (epilog) loop | |
989 | that is placed on the entry (exit) edge E of LOOP. After this transformation | |
990 | we have two loops one after the other - first-loop iterates FIRST_NITERS | |
991 | times, and second-loop iterates the remainder NITERS - FIRST_NITERS times. | |
b8698a0f | 992 | If the cost model indicates that it is profitable to emit a scalar |
ebfd146a IR |
993 | loop instead of the vector one, then the prolog (epilog) loop will iterate |
994 | for the entire unchanged scalar iterations of the loop. | |
995 | ||
996 | Input: | |
997 | - LOOP: the loop to be peeled. | |
998 | - E: the exit or entry edge of LOOP. | |
999 | If it is the entry edge, we peel the first iterations of LOOP. In this | |
1000 | case first-loop is LOOP, and second-loop is the newly created loop. | |
1001 | If it is the exit edge, we peel the last iterations of LOOP. In this | |
1002 | case, first-loop is the newly created loop, and second-loop is LOOP. | |
1003 | - NITERS: the number of iterations that LOOP iterates. | |
1004 | - FIRST_NITERS: the number of iterations that the first-loop should iterate. | |
1005 | - UPDATE_FIRST_LOOP_COUNT: specified whether this function is responsible | |
1006 | for updating the loop bound of the first-loop to FIRST_NITERS. If it | |
1007 | is false, the caller of this function may want to take care of this | |
1008 | (this can be useful if we don't want new stmts added to first-loop). | |
1009 | - TH: cost model profitability threshold of iterations for vectorization. | |
1010 | - CHECK_PROFITABILITY: specify whether cost model check has not occurred | |
1011 | during versioning and hence needs to occur during | |
b8698a0f | 1012 | prologue generation or whether cost model check |
ebfd146a IR |
1013 | has not occurred during prologue generation and hence |
1014 | needs to occur during epilogue generation. | |
e78410bf JH |
1015 | - BOUND1 is the upper bound on number of iterations of the first loop (if known) |
1016 | - BOUND2 is the upper bound on number of iterations of the second loop (if known) | |
b8698a0f | 1017 | |
ebfd146a IR |
1018 | |
1019 | Output: | |
1020 | The function returns a pointer to the new loop-copy, or NULL if it failed | |
1021 | to perform the transformation. | |
1022 | ||
1023 | The function generates two if-then-else guards: one before the first loop, | |
1024 | and the other before the second loop: | |
1025 | The first guard is: | |
1026 | if (FIRST_NITERS == 0) then skip the first loop, | |
1027 | and go directly to the second loop. | |
1028 | The second guard is: | |
1029 | if (FIRST_NITERS == NITERS) then skip the second loop. | |
1030 | ||
86290011 RG |
1031 | If the optional COND_EXPR and COND_EXPR_STMT_LIST arguments are given |
1032 | then the generated condition is combined with COND_EXPR and the | |
1033 | statements in COND_EXPR_STMT_LIST are emitted together with it. | |
1034 | ||
ebfd146a IR |
1035 | FORNOW only simple loops are supported (see slpeel_can_duplicate_loop_p). |
1036 | FORNOW the resulting code will not be in loop-closed-ssa form. | |
1037 | */ | |
1038 | ||
1039 | static struct loop* | |
b8698a0f | 1040 | slpeel_tree_peel_loop_to_edge (struct loop *loop, |
5d2eb24b | 1041 | edge e, tree *first_niters, |
ebfd146a | 1042 | tree niters, bool update_first_loop_count, |
86290011 | 1043 | unsigned int th, bool check_profitability, |
e78410bf JH |
1044 | tree cond_expr, gimple_seq cond_expr_stmt_list, |
1045 | int bound1, int bound2) | |
ebfd146a IR |
1046 | { |
1047 | struct loop *new_loop = NULL, *first_loop, *second_loop; | |
1048 | edge skip_e; | |
1049 | tree pre_condition = NULL_TREE; | |
ebfd146a IR |
1050 | basic_block bb_before_second_loop, bb_after_second_loop; |
1051 | basic_block bb_before_first_loop; | |
1052 | basic_block bb_between_loops; | |
1053 | basic_block new_exit_bb; | |
e20f6b4b | 1054 | gimple_stmt_iterator gsi; |
ebfd146a IR |
1055 | edge exit_e = single_exit (loop); |
1056 | LOC loop_loc; | |
1057 | tree cost_pre_condition = NULL_TREE; | |
e78410bf JH |
1058 | /* There are many aspects to how likely the first loop is going to be executed. |
1059 | Without histogram we can't really do good job. Simply set it to | |
1060 | 2/3, so the first loop is not reordered to the end of function and | |
1061 | the hot path through stays short. */ | |
1062 | int first_guard_probability = 2 * REG_BR_PROB_BASE / 3; | |
1063 | int second_guard_probability = 2 * REG_BR_PROB_BASE / 3; | |
1064 | int probability_of_second_loop; | |
b8698a0f | 1065 | |
ebfd146a IR |
1066 | if (!slpeel_can_duplicate_loop_p (loop, e)) |
1067 | return NULL; | |
b8698a0f | 1068 | |
141310ef RB |
1069 | /* We might have a queued need to update virtual SSA form. As we |
1070 | delete the update SSA machinery below after doing a regular | |
1071 | incremental SSA update during loop copying make sure we don't | |
1072 | lose that fact. | |
1073 | ??? Needing to update virtual SSA form by renaming is unfortunate | |
1074 | but not all of the vectorizer code inserting new loads / stores | |
1075 | properly assigns virtual operands to those statements. */ | |
1076 | update_ssa (TODO_update_ssa_only_virtuals); | |
1077 | ||
e20f6b4b JJ |
1078 | /* If the loop has a virtual PHI, but exit bb doesn't, create a virtual PHI |
1079 | in the exit bb and rename all the uses after the loop. This simplifies | |
1080 | the *guard[12] routines, which assume loop closed SSA form for all PHIs | |
1081 | (but normally loop closed SSA form doesn't require virtual PHIs to be | |
1082 | in the same form). Doing this early simplifies the checking what | |
1083 | uses should be renamed. */ | |
1084 | for (gsi = gsi_start_phis (loop->header); !gsi_end_p (gsi); gsi_next (&gsi)) | |
ea057359 | 1085 | if (virtual_operand_p (gimple_phi_result (gsi_stmt (gsi)))) |
e20f6b4b JJ |
1086 | { |
1087 | gimple phi = gsi_stmt (gsi); | |
1088 | for (gsi = gsi_start_phis (exit_e->dest); | |
1089 | !gsi_end_p (gsi); gsi_next (&gsi)) | |
ea057359 | 1090 | if (virtual_operand_p (gimple_phi_result (gsi_stmt (gsi)))) |
e20f6b4b JJ |
1091 | break; |
1092 | if (gsi_end_p (gsi)) | |
1093 | { | |
070ecdfd RG |
1094 | tree new_vop = copy_ssa_name (PHI_RESULT (phi), NULL); |
1095 | gimple new_phi = create_phi_node (new_vop, exit_e->dest); | |
e20f6b4b JJ |
1096 | tree vop = PHI_ARG_DEF_FROM_EDGE (phi, EDGE_SUCC (loop->latch, 0)); |
1097 | imm_use_iterator imm_iter; | |
1098 | gimple stmt; | |
e20f6b4b JJ |
1099 | use_operand_p use_p; |
1100 | ||
9e227d60 | 1101 | add_phi_arg (new_phi, vop, exit_e, UNKNOWN_LOCATION); |
e20f6b4b JJ |
1102 | gimple_phi_set_result (new_phi, new_vop); |
1103 | FOR_EACH_IMM_USE_STMT (stmt, imm_iter, vop) | |
1104 | if (stmt != new_phi && gimple_bb (stmt) != loop->header) | |
1105 | FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter) | |
1106 | SET_USE (use_p, new_vop); | |
1107 | } | |
1108 | break; | |
1109 | } | |
ebfd146a IR |
1110 | |
1111 | /* 1. Generate a copy of LOOP and put it on E (E is the entry/exit of LOOP). | |
1112 | Resulting CFG would be: | |
1113 | ||
1114 | first_loop: | |
1115 | do { | |
1116 | } while ... | |
1117 | ||
1118 | second_loop: | |
1119 | do { | |
1120 | } while ... | |
1121 | ||
1122 | orig_exit_bb: | |
1123 | */ | |
b8698a0f | 1124 | |
ebfd146a IR |
1125 | if (!(new_loop = slpeel_tree_duplicate_loop_to_edge_cfg (loop, e))) |
1126 | { | |
1127 | loop_loc = find_loop_location (loop); | |
78c60e3d SS |
1128 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, loop_loc, |
1129 | "tree_duplicate_loop_to_edge_cfg failed.\n"); | |
ebfd146a IR |
1130 | return NULL; |
1131 | } | |
b8698a0f | 1132 | |
684f25f4 AO |
1133 | if (MAY_HAVE_DEBUG_STMTS) |
1134 | { | |
9771b263 | 1135 | gcc_assert (!adjust_vec.exists ()); |
ff4c81cc | 1136 | adjust_vec.create (32); |
684f25f4 AO |
1137 | } |
1138 | ||
ebfd146a IR |
1139 | if (e == exit_e) |
1140 | { | |
1141 | /* NEW_LOOP was placed after LOOP. */ | |
1142 | first_loop = loop; | |
1143 | second_loop = new_loop; | |
1144 | } | |
1145 | else | |
1146 | { | |
1147 | /* NEW_LOOP was placed before LOOP. */ | |
1148 | first_loop = new_loop; | |
1149 | second_loop = loop; | |
1150 | } | |
1151 | ||
ebfd146a IR |
1152 | /* 2. Add the guard code in one of the following ways: |
1153 | ||
1154 | 2.a Add the guard that controls whether the first loop is executed. | |
1155 | This occurs when this function is invoked for prologue or epilogue | |
1156 | generation and when the cost model check can be done at compile time. | |
1157 | ||
1158 | Resulting CFG would be: | |
1159 | ||
1160 | bb_before_first_loop: | |
1161 | if (FIRST_NITERS == 0) GOTO bb_before_second_loop | |
1162 | GOTO first-loop | |
1163 | ||
1164 | first_loop: | |
1165 | do { | |
1166 | } while ... | |
1167 | ||
1168 | bb_before_second_loop: | |
1169 | ||
1170 | second_loop: | |
1171 | do { | |
1172 | } while ... | |
1173 | ||
1174 | orig_exit_bb: | |
1175 | ||
1176 | 2.b Add the cost model check that allows the prologue | |
1177 | to iterate for the entire unchanged scalar | |
1178 | iterations of the loop in the event that the cost | |
1179 | model indicates that the scalar loop is more | |
1180 | profitable than the vector one. This occurs when | |
1181 | this function is invoked for prologue generation | |
1182 | and the cost model check needs to be done at run | |
1183 | time. | |
1184 | ||
1185 | Resulting CFG after prologue peeling would be: | |
1186 | ||
1187 | if (scalar_loop_iterations <= th) | |
1188 | FIRST_NITERS = scalar_loop_iterations | |
1189 | ||
1190 | bb_before_first_loop: | |
1191 | if (FIRST_NITERS == 0) GOTO bb_before_second_loop | |
1192 | GOTO first-loop | |
1193 | ||
1194 | first_loop: | |
1195 | do { | |
1196 | } while ... | |
1197 | ||
1198 | bb_before_second_loop: | |
1199 | ||
1200 | second_loop: | |
1201 | do { | |
1202 | } while ... | |
1203 | ||
1204 | orig_exit_bb: | |
1205 | ||
1206 | 2.c Add the cost model check that allows the epilogue | |
1207 | to iterate for the entire unchanged scalar | |
1208 | iterations of the loop in the event that the cost | |
1209 | model indicates that the scalar loop is more | |
1210 | profitable than the vector one. This occurs when | |
1211 | this function is invoked for epilogue generation | |
1212 | and the cost model check needs to be done at run | |
86290011 RG |
1213 | time. This check is combined with any pre-existing |
1214 | check in COND_EXPR to avoid versioning. | |
ebfd146a IR |
1215 | |
1216 | Resulting CFG after prologue peeling would be: | |
1217 | ||
1218 | bb_before_first_loop: | |
1219 | if ((scalar_loop_iterations <= th) | |
1220 | || | |
1221 | FIRST_NITERS == 0) GOTO bb_before_second_loop | |
1222 | GOTO first-loop | |
1223 | ||
1224 | first_loop: | |
1225 | do { | |
1226 | } while ... | |
1227 | ||
1228 | bb_before_second_loop: | |
1229 | ||
1230 | second_loop: | |
1231 | do { | |
1232 | } while ... | |
1233 | ||
1234 | orig_exit_bb: | |
1235 | */ | |
1236 | ||
1237 | bb_before_first_loop = split_edge (loop_preheader_edge (first_loop)); | |
2cfc56b9 RB |
1238 | /* Loop copying insterted a forwarder block for us here. */ |
1239 | bb_before_second_loop = single_exit (first_loop)->dest; | |
ebfd146a | 1240 | |
e78410bf JH |
1241 | probability_of_second_loop = (inverse_probability (first_guard_probability) |
1242 | + combine_probabilities (second_guard_probability, | |
1243 | first_guard_probability)); | |
1244 | /* Theoretically preheader edge of first loop and exit edge should have | |
1245 | same frequencies. Loop exit probablities are however easy to get wrong. | |
1246 | It is safer to copy value from original loop entry. */ | |
1247 | bb_before_second_loop->frequency | |
8ddb5a29 TJ |
1248 | = combine_probabilities (bb_before_first_loop->frequency, |
1249 | probability_of_second_loop); | |
e78410bf JH |
1250 | bb_before_second_loop->count |
1251 | = apply_probability (bb_before_first_loop->count, | |
1252 | probability_of_second_loop); | |
1253 | single_succ_edge (bb_before_second_loop)->count | |
1254 | = bb_before_second_loop->count; | |
1255 | ||
ebfd146a IR |
1256 | /* Epilogue peeling. */ |
1257 | if (!update_first_loop_count) | |
1258 | { | |
1259 | pre_condition = | |
5d2eb24b IR |
1260 | fold_build2 (LE_EXPR, boolean_type_node, *first_niters, |
1261 | build_int_cst (TREE_TYPE (*first_niters), 0)); | |
ebfd146a IR |
1262 | if (check_profitability) |
1263 | { | |
1264 | tree scalar_loop_iters | |
1265 | = unshare_expr (LOOP_VINFO_NITERS_UNCHANGED | |
1266 | (loop_vec_info_for_loop (loop))); | |
b8698a0f L |
1267 | cost_pre_condition = |
1268 | fold_build2 (LE_EXPR, boolean_type_node, scalar_loop_iters, | |
ebfd146a IR |
1269 | build_int_cst (TREE_TYPE (scalar_loop_iters), th)); |
1270 | ||
1271 | pre_condition = fold_build2 (TRUTH_OR_EXPR, boolean_type_node, | |
1272 | cost_pre_condition, pre_condition); | |
1273 | } | |
86290011 RG |
1274 | if (cond_expr) |
1275 | { | |
1276 | pre_condition = | |
1277 | fold_build2 (TRUTH_OR_EXPR, boolean_type_node, | |
1278 | pre_condition, | |
1279 | fold_build1 (TRUTH_NOT_EXPR, boolean_type_node, | |
1280 | cond_expr)); | |
1281 | } | |
ebfd146a IR |
1282 | } |
1283 | ||
b8698a0f | 1284 | /* Prologue peeling. */ |
ebfd146a IR |
1285 | else |
1286 | { | |
1287 | if (check_profitability) | |
1288 | set_prologue_iterations (bb_before_first_loop, first_niters, | |
e78410bf | 1289 | loop, th, first_guard_probability); |
ebfd146a IR |
1290 | |
1291 | pre_condition = | |
5d2eb24b IR |
1292 | fold_build2 (LE_EXPR, boolean_type_node, *first_niters, |
1293 | build_int_cst (TREE_TYPE (*first_niters), 0)); | |
ebfd146a IR |
1294 | } |
1295 | ||
1296 | skip_e = slpeel_add_loop_guard (bb_before_first_loop, pre_condition, | |
86290011 | 1297 | cond_expr_stmt_list, |
e78410bf JH |
1298 | bb_before_second_loop, bb_before_first_loop, |
1299 | inverse_probability (first_guard_probability)); | |
1300 | scale_loop_profile (first_loop, first_guard_probability, | |
1301 | check_profitability && (int)th > bound1 ? th : bound1); | |
ebfd146a IR |
1302 | slpeel_update_phi_nodes_for_guard1 (skip_e, first_loop, |
1303 | first_loop == new_loop, | |
c334023f | 1304 | &new_exit_bb); |
ebfd146a IR |
1305 | |
1306 | ||
1307 | /* 3. Add the guard that controls whether the second loop is executed. | |
1308 | Resulting CFG would be: | |
1309 | ||
1310 | bb_before_first_loop: | |
1311 | if (FIRST_NITERS == 0) GOTO bb_before_second_loop (skip first loop) | |
1312 | GOTO first-loop | |
1313 | ||
1314 | first_loop: | |
1315 | do { | |
1316 | } while ... | |
1317 | ||
1318 | bb_between_loops: | |
1319 | if (FIRST_NITERS == NITERS) GOTO bb_after_second_loop (skip second loop) | |
1320 | GOTO bb_before_second_loop | |
1321 | ||
1322 | bb_before_second_loop: | |
1323 | ||
1324 | second_loop: | |
1325 | do { | |
1326 | } while ... | |
1327 | ||
1328 | bb_after_second_loop: | |
1329 | ||
1330 | orig_exit_bb: | |
1331 | */ | |
1332 | ||
1333 | bb_between_loops = new_exit_bb; | |
1334 | bb_after_second_loop = split_edge (single_exit (second_loop)); | |
1335 | ||
b8698a0f | 1336 | pre_condition = |
5d2eb24b | 1337 | fold_build2 (EQ_EXPR, boolean_type_node, *first_niters, niters); |
86290011 | 1338 | skip_e = slpeel_add_loop_guard (bb_between_loops, pre_condition, NULL, |
e78410bf JH |
1339 | bb_after_second_loop, bb_before_first_loop, |
1340 | inverse_probability (second_guard_probability)); | |
1341 | scale_loop_profile (second_loop, probability_of_second_loop, bound2); | |
ebfd146a IR |
1342 | slpeel_update_phi_nodes_for_guard2 (skip_e, second_loop, |
1343 | second_loop == new_loop, &new_exit_bb); | |
1344 | ||
1345 | /* 4. Make first-loop iterate FIRST_NITERS times, if requested. | |
1346 | */ | |
1347 | if (update_first_loop_count) | |
5d2eb24b | 1348 | slpeel_make_loop_iterate_ntimes (first_loop, *first_niters); |
ebfd146a | 1349 | |
040d39ee RG |
1350 | delete_update_ssa (); |
1351 | ||
684f25f4 AO |
1352 | adjust_vec_debug_stmts (); |
1353 | ||
ebfd146a IR |
1354 | return new_loop; |
1355 | } | |
1356 | ||
1357 | /* Function vect_get_loop_location. | |
1358 | ||
1359 | Extract the location of the loop in the source code. | |
1360 | If the loop is not well formed for vectorization, an estimated | |
1361 | location is calculated. | |
1362 | Return the loop location if succeed and NULL if not. */ | |
1363 | ||
1364 | LOC | |
1365 | find_loop_location (struct loop *loop) | |
1366 | { | |
1367 | gimple stmt = NULL; | |
1368 | basic_block bb; | |
1369 | gimple_stmt_iterator si; | |
1370 | ||
1371 | if (!loop) | |
1372 | return UNKNOWN_LOC; | |
1373 | ||
1374 | stmt = get_loop_exit_condition (loop); | |
1375 | ||
502498d5 JJ |
1376 | if (stmt |
1377 | && LOCATION_LOCUS (gimple_location (stmt)) > BUILTINS_LOCATION) | |
ebfd146a IR |
1378 | return gimple_location (stmt); |
1379 | ||
1380 | /* If we got here the loop is probably not "well formed", | |
1381 | try to estimate the loop location */ | |
1382 | ||
1383 | if (!loop->header) | |
1384 | return UNKNOWN_LOC; | |
1385 | ||
1386 | bb = loop->header; | |
1387 | ||
1388 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
1389 | { | |
1390 | stmt = gsi_stmt (si); | |
502498d5 | 1391 | if (LOCATION_LOCUS (gimple_location (stmt)) > BUILTINS_LOCATION) |
ebfd146a IR |
1392 | return gimple_location (stmt); |
1393 | } | |
1394 | ||
1395 | return UNKNOWN_LOC; | |
1396 | } | |
1397 | ||
1398 | ||
1399 | /* This function builds ni_name = number of iterations loop executes | |
86290011 RG |
1400 | on the loop preheader. If SEQ is given the stmt is instead emitted |
1401 | there. */ | |
ebfd146a IR |
1402 | |
1403 | static tree | |
86290011 | 1404 | vect_build_loop_niters (loop_vec_info loop_vinfo, gimple_seq seq) |
ebfd146a IR |
1405 | { |
1406 | tree ni_name, var; | |
1407 | gimple_seq stmts = NULL; | |
1408 | edge pe; | |
1409 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
1410 | tree ni = unshare_expr (LOOP_VINFO_NITERS (loop_vinfo)); | |
1411 | ||
1412 | var = create_tmp_var (TREE_TYPE (ni), "niters"); | |
ebfd146a IR |
1413 | ni_name = force_gimple_operand (ni, &stmts, false, var); |
1414 | ||
1415 | pe = loop_preheader_edge (loop); | |
1416 | if (stmts) | |
1417 | { | |
86290011 RG |
1418 | if (seq) |
1419 | gimple_seq_add_seq (&seq, stmts); | |
1420 | else | |
1421 | { | |
1422 | basic_block new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
1423 | gcc_assert (!new_bb); | |
1424 | } | |
ebfd146a IR |
1425 | } |
1426 | ||
1427 | return ni_name; | |
1428 | } | |
1429 | ||
1430 | ||
1431 | /* This function generates the following statements: | |
1432 | ||
1433 | ni_name = number of iterations loop executes | |
1434 | ratio = ni_name / vf | |
1435 | ratio_mult_vf_name = ratio * vf | |
1436 | ||
86290011 RG |
1437 | and places them at the loop preheader edge or in COND_EXPR_STMT_LIST |
1438 | if that is non-NULL. */ | |
ebfd146a | 1439 | |
eae76e53 | 1440 | void |
b8698a0f | 1441 | vect_generate_tmps_on_preheader (loop_vec_info loop_vinfo, |
ebfd146a | 1442 | tree *ni_name_ptr, |
b8698a0f | 1443 | tree *ratio_mult_vf_name_ptr, |
86290011 RG |
1444 | tree *ratio_name_ptr, |
1445 | gimple_seq cond_expr_stmt_list) | |
ebfd146a IR |
1446 | { |
1447 | ||
1448 | edge pe; | |
1449 | basic_block new_bb; | |
1450 | gimple_seq stmts; | |
48df3fa6 | 1451 | tree ni_name, ni_minus_gap_name; |
ebfd146a IR |
1452 | tree var; |
1453 | tree ratio_name; | |
1454 | tree ratio_mult_vf_name; | |
1455 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
1456 | tree ni = LOOP_VINFO_NITERS (loop_vinfo); | |
1457 | int vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
1458 | tree log_vf; | |
1459 | ||
1460 | pe = loop_preheader_edge (loop); | |
1461 | ||
b8698a0f | 1462 | /* Generate temporary variable that contains |
ebfd146a IR |
1463 | number of iterations loop executes. */ |
1464 | ||
86290011 | 1465 | ni_name = vect_build_loop_niters (loop_vinfo, cond_expr_stmt_list); |
ebfd146a IR |
1466 | log_vf = build_int_cst (TREE_TYPE (ni), exact_log2 (vf)); |
1467 | ||
48df3fa6 IR |
1468 | /* If epilogue loop is required because of data accesses with gaps, we |
1469 | subtract one iteration from the total number of iterations here for | |
1470 | correct calculation of RATIO. */ | |
1471 | if (LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo)) | |
1472 | { | |
1473 | ni_minus_gap_name = fold_build2 (MINUS_EXPR, TREE_TYPE (ni_name), | |
1474 | ni_name, | |
1475 | build_one_cst (TREE_TYPE (ni_name))); | |
1476 | if (!is_gimple_val (ni_minus_gap_name)) | |
1477 | { | |
1478 | var = create_tmp_var (TREE_TYPE (ni), "ni_gap"); | |
48df3fa6 IR |
1479 | |
1480 | stmts = NULL; | |
1481 | ni_minus_gap_name = force_gimple_operand (ni_minus_gap_name, &stmts, | |
1482 | true, var); | |
1483 | if (cond_expr_stmt_list) | |
1484 | gimple_seq_add_seq (&cond_expr_stmt_list, stmts); | |
1485 | else | |
1486 | { | |
1487 | pe = loop_preheader_edge (loop); | |
1488 | new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
1489 | gcc_assert (!new_bb); | |
1490 | } | |
1491 | } | |
1492 | } | |
1493 | else | |
1494 | ni_minus_gap_name = ni_name; | |
1495 | ||
ebfd146a IR |
1496 | /* Create: ratio = ni >> log2(vf) */ |
1497 | ||
48df3fa6 IR |
1498 | ratio_name = fold_build2 (RSHIFT_EXPR, TREE_TYPE (ni_minus_gap_name), |
1499 | ni_minus_gap_name, log_vf); | |
ebfd146a IR |
1500 | if (!is_gimple_val (ratio_name)) |
1501 | { | |
1502 | var = create_tmp_var (TREE_TYPE (ni), "bnd"); | |
ebfd146a IR |
1503 | |
1504 | stmts = NULL; | |
1505 | ratio_name = force_gimple_operand (ratio_name, &stmts, true, var); | |
86290011 RG |
1506 | if (cond_expr_stmt_list) |
1507 | gimple_seq_add_seq (&cond_expr_stmt_list, stmts); | |
1508 | else | |
1509 | { | |
1510 | pe = loop_preheader_edge (loop); | |
1511 | new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
1512 | gcc_assert (!new_bb); | |
1513 | } | |
ebfd146a | 1514 | } |
b8698a0f | 1515 | |
ebfd146a IR |
1516 | /* Create: ratio_mult_vf = ratio << log2 (vf). */ |
1517 | ||
1518 | ratio_mult_vf_name = fold_build2 (LSHIFT_EXPR, TREE_TYPE (ratio_name), | |
1519 | ratio_name, log_vf); | |
1520 | if (!is_gimple_val (ratio_mult_vf_name)) | |
1521 | { | |
1522 | var = create_tmp_var (TREE_TYPE (ni), "ratio_mult_vf"); | |
ebfd146a IR |
1523 | |
1524 | stmts = NULL; | |
1525 | ratio_mult_vf_name = force_gimple_operand (ratio_mult_vf_name, &stmts, | |
1526 | true, var); | |
86290011 RG |
1527 | if (cond_expr_stmt_list) |
1528 | gimple_seq_add_seq (&cond_expr_stmt_list, stmts); | |
1529 | else | |
1530 | { | |
1531 | pe = loop_preheader_edge (loop); | |
1532 | new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
1533 | gcc_assert (!new_bb); | |
1534 | } | |
ebfd146a IR |
1535 | } |
1536 | ||
1537 | *ni_name_ptr = ni_name; | |
1538 | *ratio_mult_vf_name_ptr = ratio_mult_vf_name; | |
1539 | *ratio_name_ptr = ratio_name; | |
b8698a0f L |
1540 | |
1541 | return; | |
ebfd146a IR |
1542 | } |
1543 | ||
1544 | /* Function vect_can_advance_ivs_p | |
1545 | ||
b8698a0f L |
1546 | In case the number of iterations that LOOP iterates is unknown at compile |
1547 | time, an epilog loop will be generated, and the loop induction variables | |
1548 | (IVs) will be "advanced" to the value they are supposed to take just before | |
ebfd146a IR |
1549 | the epilog loop. Here we check that the access function of the loop IVs |
1550 | and the expression that represents the loop bound are simple enough. | |
1551 | These restrictions will be relaxed in the future. */ | |
1552 | ||
b8698a0f | 1553 | bool |
ebfd146a IR |
1554 | vect_can_advance_ivs_p (loop_vec_info loop_vinfo) |
1555 | { | |
1556 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
1557 | basic_block bb = loop->header; | |
1558 | gimple phi; | |
1559 | gimple_stmt_iterator gsi; | |
1560 | ||
1561 | /* Analyze phi functions of the loop header. */ | |
1562 | ||
73fbfcad | 1563 | if (dump_enabled_p ()) |
e645e942 | 1564 | dump_printf_loc (MSG_NOTE, vect_location, "vect_can_advance_ivs_p:\n"); |
ebfd146a IR |
1565 | for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
1566 | { | |
ebfd146a IR |
1567 | tree evolution_part; |
1568 | ||
1569 | phi = gsi_stmt (gsi); | |
73fbfcad | 1570 | if (dump_enabled_p ()) |
ebfd146a | 1571 | { |
78c60e3d SS |
1572 | dump_printf_loc (MSG_NOTE, vect_location, "Analyze phi: "); |
1573 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0); | |
e645e942 | 1574 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
1575 | } |
1576 | ||
1577 | /* Skip virtual phi's. The data dependences that are associated with | |
1578 | virtual defs/uses (i.e., memory accesses) are analyzed elsewhere. */ | |
1579 | ||
ea057359 | 1580 | if (virtual_operand_p (PHI_RESULT (phi))) |
ebfd146a | 1581 | { |
73fbfcad | 1582 | if (dump_enabled_p ()) |
78c60e3d | 1583 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1584 | "virtual phi. skip.\n"); |
ebfd146a IR |
1585 | continue; |
1586 | } | |
1587 | ||
1588 | /* Skip reduction phis. */ | |
1589 | ||
1590 | if (STMT_VINFO_DEF_TYPE (vinfo_for_stmt (phi)) == vect_reduction_def) | |
1591 | { | |
73fbfcad | 1592 | if (dump_enabled_p ()) |
78c60e3d | 1593 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1594 | "reduc phi. skip.\n"); |
ebfd146a IR |
1595 | continue; |
1596 | } | |
1597 | ||
1598 | /* Analyze the evolution function. */ | |
1599 | ||
afb119be RB |
1600 | evolution_part |
1601 | = STMT_VINFO_LOOP_PHI_EVOLUTION_PART (vinfo_for_stmt (phi)); | |
ebfd146a IR |
1602 | if (evolution_part == NULL_TREE) |
1603 | { | |
73fbfcad | 1604 | if (dump_enabled_p ()) |
afb119be | 1605 | dump_printf (MSG_MISSED_OPTIMIZATION, |
e645e942 | 1606 | "No access function or evolution.\n"); |
ebfd146a IR |
1607 | return false; |
1608 | } | |
b8698a0f L |
1609 | |
1610 | /* FORNOW: We do not transform initial conditions of IVs | |
ebfd146a IR |
1611 | which evolution functions are a polynomial of degree >= 2. */ |
1612 | ||
1613 | if (tree_is_chrec (evolution_part)) | |
b8698a0f | 1614 | return false; |
ebfd146a IR |
1615 | } |
1616 | ||
1617 | return true; | |
1618 | } | |
1619 | ||
1620 | ||
1621 | /* Function vect_update_ivs_after_vectorizer. | |
1622 | ||
1623 | "Advance" the induction variables of LOOP to the value they should take | |
1624 | after the execution of LOOP. This is currently necessary because the | |
1625 | vectorizer does not handle induction variables that are used after the | |
1626 | loop. Such a situation occurs when the last iterations of LOOP are | |
1627 | peeled, because: | |
1628 | 1. We introduced new uses after LOOP for IVs that were not originally used | |
1629 | after LOOP: the IVs of LOOP are now used by an epilog loop. | |
1630 | 2. LOOP is going to be vectorized; this means that it will iterate N/VF | |
1631 | times, whereas the loop IVs should be bumped N times. | |
1632 | ||
1633 | Input: | |
1634 | - LOOP - a loop that is going to be vectorized. The last few iterations | |
1635 | of LOOP were peeled. | |
1636 | - NITERS - the number of iterations that LOOP executes (before it is | |
1637 | vectorized). i.e, the number of times the ivs should be bumped. | |
1638 | - UPDATE_E - a successor edge of LOOP->exit that is on the (only) path | |
1639 | coming out from LOOP on which there are uses of the LOOP ivs | |
1640 | (this is the path from LOOP->exit to epilog_loop->preheader). | |
1641 | ||
1642 | The new definitions of the ivs are placed in LOOP->exit. | |
1643 | The phi args associated with the edge UPDATE_E in the bb | |
1644 | UPDATE_E->dest are updated accordingly. | |
1645 | ||
1646 | Assumption 1: Like the rest of the vectorizer, this function assumes | |
1647 | a single loop exit that has a single predecessor. | |
1648 | ||
1649 | Assumption 2: The phi nodes in the LOOP header and in update_bb are | |
1650 | organized in the same order. | |
1651 | ||
1652 | Assumption 3: The access function of the ivs is simple enough (see | |
1653 | vect_can_advance_ivs_p). This assumption will be relaxed in the future. | |
1654 | ||
1655 | Assumption 4: Exactly one of the successors of LOOP exit-bb is on a path | |
b8698a0f | 1656 | coming out of LOOP on which the ivs of LOOP are used (this is the path |
ebfd146a IR |
1657 | that leads to the epilog loop; other paths skip the epilog loop). This |
1658 | path starts with the edge UPDATE_E, and its destination (denoted update_bb) | |
1659 | needs to have its phis updated. | |
1660 | */ | |
1661 | ||
1662 | static void | |
b8698a0f | 1663 | vect_update_ivs_after_vectorizer (loop_vec_info loop_vinfo, tree niters, |
ebfd146a IR |
1664 | edge update_e) |
1665 | { | |
1666 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
1667 | basic_block exit_bb = single_exit (loop)->dest; | |
1668 | gimple phi, phi1; | |
1669 | gimple_stmt_iterator gsi, gsi1; | |
1670 | basic_block update_bb = update_e->dest; | |
1671 | ||
1672 | /* gcc_assert (vect_can_advance_ivs_p (loop_vinfo)); */ | |
1673 | ||
1674 | /* Make sure there exists a single-predecessor exit bb: */ | |
1675 | gcc_assert (single_pred_p (exit_bb)); | |
1676 | ||
1677 | for (gsi = gsi_start_phis (loop->header), gsi1 = gsi_start_phis (update_bb); | |
1678 | !gsi_end_p (gsi) && !gsi_end_p (gsi1); | |
1679 | gsi_next (&gsi), gsi_next (&gsi1)) | |
1680 | { | |
ebfd146a | 1681 | tree init_expr; |
550918ca RG |
1682 | tree step_expr, off; |
1683 | tree type; | |
ebfd146a IR |
1684 | tree var, ni, ni_name; |
1685 | gimple_stmt_iterator last_gsi; | |
0ac168a1 | 1686 | stmt_vec_info stmt_info; |
ebfd146a IR |
1687 | |
1688 | phi = gsi_stmt (gsi); | |
1689 | phi1 = gsi_stmt (gsi1); | |
73fbfcad | 1690 | if (dump_enabled_p ()) |
ebfd146a | 1691 | { |
78c60e3d SS |
1692 | dump_printf_loc (MSG_NOTE, vect_location, |
1693 | "vect_update_ivs_after_vectorizer: phi: "); | |
1694 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0); | |
e645e942 | 1695 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
1696 | } |
1697 | ||
1698 | /* Skip virtual phi's. */ | |
ea057359 | 1699 | if (virtual_operand_p (PHI_RESULT (phi))) |
ebfd146a | 1700 | { |
73fbfcad | 1701 | if (dump_enabled_p ()) |
78c60e3d | 1702 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1703 | "virtual phi. skip.\n"); |
ebfd146a IR |
1704 | continue; |
1705 | } | |
1706 | ||
1707 | /* Skip reduction phis. */ | |
0ac168a1 RG |
1708 | stmt_info = vinfo_for_stmt (phi); |
1709 | if (STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def) | |
b8698a0f | 1710 | { |
73fbfcad | 1711 | if (dump_enabled_p ()) |
78c60e3d | 1712 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1713 | "reduc phi. skip.\n"); |
ebfd146a | 1714 | continue; |
b8698a0f | 1715 | } |
ebfd146a | 1716 | |
0ac168a1 RG |
1717 | type = TREE_TYPE (gimple_phi_result (phi)); |
1718 | step_expr = STMT_VINFO_LOOP_PHI_EVOLUTION_PART (stmt_info); | |
1719 | step_expr = unshare_expr (step_expr); | |
b8698a0f | 1720 | |
ebfd146a IR |
1721 | /* FORNOW: We do not support IVs whose evolution function is a polynomial |
1722 | of degree >= 2 or exponential. */ | |
0ac168a1 | 1723 | gcc_assert (!tree_is_chrec (step_expr)); |
ebfd146a | 1724 | |
0ac168a1 | 1725 | init_expr = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop)); |
ebfd146a | 1726 | |
550918ca RG |
1727 | off = fold_build2 (MULT_EXPR, TREE_TYPE (step_expr), |
1728 | fold_convert (TREE_TYPE (step_expr), niters), | |
1729 | step_expr); | |
0ac168a1 | 1730 | if (POINTER_TYPE_P (type)) |
5d49b6a7 | 1731 | ni = fold_build_pointer_plus (init_expr, off); |
ebfd146a | 1732 | else |
0ac168a1 RG |
1733 | ni = fold_build2 (PLUS_EXPR, type, |
1734 | init_expr, fold_convert (type, off)); | |
ebfd146a | 1735 | |
0ac168a1 | 1736 | var = create_tmp_var (type, "tmp"); |
ebfd146a IR |
1737 | |
1738 | last_gsi = gsi_last_bb (exit_bb); | |
1739 | ni_name = force_gimple_operand_gsi (&last_gsi, ni, false, var, | |
1740 | true, GSI_SAME_STMT); | |
b8698a0f | 1741 | |
ebfd146a | 1742 | /* Fix phi expressions in the successor bb. */ |
684f25f4 | 1743 | adjust_phi_and_debug_stmts (phi1, update_e, ni_name); |
ebfd146a IR |
1744 | } |
1745 | } | |
1746 | ||
ebfd146a IR |
1747 | /* Function vect_do_peeling_for_loop_bound |
1748 | ||
1749 | Peel the last iterations of the loop represented by LOOP_VINFO. | |
b8698a0f | 1750 | The peeled iterations form a new epilog loop. Given that the loop now |
ebfd146a IR |
1751 | iterates NITERS times, the new epilog loop iterates |
1752 | NITERS % VECTORIZATION_FACTOR times. | |
b8698a0f L |
1753 | |
1754 | The original loop will later be made to iterate | |
86290011 RG |
1755 | NITERS / VECTORIZATION_FACTOR times (this value is placed into RATIO). |
1756 | ||
1757 | COND_EXPR and COND_EXPR_STMT_LIST are combined with a new generated | |
1758 | test. */ | |
ebfd146a | 1759 | |
b8698a0f | 1760 | void |
86290011 | 1761 | vect_do_peeling_for_loop_bound (loop_vec_info loop_vinfo, tree *ratio, |
368117e8 | 1762 | unsigned int th, bool check_profitability) |
ebfd146a IR |
1763 | { |
1764 | tree ni_name, ratio_mult_vf_name; | |
1765 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
1766 | struct loop *new_loop; | |
1767 | edge update_e; | |
1768 | basic_block preheader; | |
1769 | int loop_num; | |
d68d56b5 | 1770 | int max_iter; |
368117e8 RG |
1771 | tree cond_expr = NULL_TREE; |
1772 | gimple_seq cond_expr_stmt_list = NULL; | |
ebfd146a | 1773 | |
73fbfcad | 1774 | if (dump_enabled_p ()) |
ccb3ad87 | 1775 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1776 | "=== vect_do_peeling_for_loop_bound ===\n"); |
ebfd146a IR |
1777 | |
1778 | initialize_original_copy_tables (); | |
1779 | ||
1780 | /* Generate the following variables on the preheader of original loop: | |
b8698a0f | 1781 | |
ebfd146a IR |
1782 | ni_name = number of iteration the original loop executes |
1783 | ratio = ni_name / vf | |
1784 | ratio_mult_vf_name = ratio * vf */ | |
1785 | vect_generate_tmps_on_preheader (loop_vinfo, &ni_name, | |
86290011 RG |
1786 | &ratio_mult_vf_name, ratio, |
1787 | cond_expr_stmt_list); | |
ebfd146a | 1788 | |
b8698a0f | 1789 | loop_num = loop->num; |
ebfd146a | 1790 | |
ebfd146a | 1791 | new_loop = slpeel_tree_peel_loop_to_edge (loop, single_exit (loop), |
5d2eb24b | 1792 | &ratio_mult_vf_name, ni_name, false, |
86290011 | 1793 | th, check_profitability, |
e78410bf JH |
1794 | cond_expr, cond_expr_stmt_list, |
1795 | 0, LOOP_VINFO_VECT_FACTOR (loop_vinfo)); | |
ebfd146a IR |
1796 | gcc_assert (new_loop); |
1797 | gcc_assert (loop_num == loop->num); | |
1798 | #ifdef ENABLE_CHECKING | |
1799 | slpeel_verify_cfg_after_peeling (loop, new_loop); | |
1800 | #endif | |
1801 | ||
1802 | /* A guard that controls whether the new_loop is to be executed or skipped | |
1803 | is placed in LOOP->exit. LOOP->exit therefore has two successors - one | |
1804 | is the preheader of NEW_LOOP, where the IVs from LOOP are used. The other | |
1805 | is a bb after NEW_LOOP, where these IVs are not used. Find the edge that | |
1806 | is on the path where the LOOP IVs are used and need to be updated. */ | |
1807 | ||
1808 | preheader = loop_preheader_edge (new_loop)->src; | |
1809 | if (EDGE_PRED (preheader, 0)->src == single_exit (loop)->dest) | |
1810 | update_e = EDGE_PRED (preheader, 0); | |
1811 | else | |
1812 | update_e = EDGE_PRED (preheader, 1); | |
1813 | ||
b8698a0f | 1814 | /* Update IVs of original loop as if they were advanced |
ebfd146a | 1815 | by ratio_mult_vf_name steps. */ |
b8698a0f | 1816 | vect_update_ivs_after_vectorizer (loop_vinfo, ratio_mult_vf_name, update_e); |
ebfd146a | 1817 | |
22458c5a JH |
1818 | /* For vectorization factor N, we need to copy last N-1 values in epilogue |
1819 | and this means N-2 loopback edge executions. | |
1820 | ||
1821 | PEELING_FOR_GAPS works by subtracting last iteration and thus the epilogue | |
1822 | will execute at least LOOP_VINFO_VECT_FACTOR times. */ | |
1823 | max_iter = (LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) | |
1824 | ? LOOP_VINFO_VECT_FACTOR (loop_vinfo) * 2 | |
1825 | : LOOP_VINFO_VECT_FACTOR (loop_vinfo)) - 2; | |
368117e8 | 1826 | if (check_profitability) |
22458c5a | 1827 | max_iter = MAX (max_iter, (int) th - 1); |
27bcd47c | 1828 | record_niter_bound (new_loop, double_int::from_shwi (max_iter), false, true); |
ccb3ad87 | 1829 | dump_printf (MSG_NOTE, |
78c60e3d SS |
1830 | "Setting upper bound of nb iterations for epilogue " |
1831 | "loop to %d\n", max_iter); | |
7d5a99f4 | 1832 | |
ebfd146a IR |
1833 | /* After peeling we have to reset scalar evolution analyzer. */ |
1834 | scev_reset (); | |
1835 | ||
1836 | free_original_copy_tables (); | |
1837 | } | |
1838 | ||
1839 | ||
1840 | /* Function vect_gen_niters_for_prolog_loop | |
1841 | ||
1842 | Set the number of iterations for the loop represented by LOOP_VINFO | |
1843 | to the minimum between LOOP_NITERS (the original iteration count of the loop) | |
1844 | and the misalignment of DR - the data reference recorded in | |
b8698a0f | 1845 | LOOP_VINFO_UNALIGNED_DR (LOOP_VINFO). As a result, after the execution of |
ebfd146a IR |
1846 | this loop, the data reference DR will refer to an aligned location. |
1847 | ||
1848 | The following computation is generated: | |
1849 | ||
1850 | If the misalignment of DR is known at compile time: | |
1851 | addr_mis = int mis = DR_MISALIGNMENT (dr); | |
1852 | Else, compute address misalignment in bytes: | |
5aea1e76 | 1853 | addr_mis = addr & (vectype_align - 1) |
ebfd146a IR |
1854 | |
1855 | prolog_niters = min (LOOP_NITERS, ((VF - addr_mis/elem_size)&(VF-1))/step) | |
1856 | ||
1857 | (elem_size = element type size; an element is the scalar element whose type | |
1858 | is the inner type of the vectype) | |
1859 | ||
1860 | When the step of the data-ref in the loop is not 1 (as in interleaved data | |
1861 | and SLP), the number of iterations of the prolog must be divided by the step | |
1862 | (which is equal to the size of interleaved group). | |
1863 | ||
1864 | The above formulas assume that VF == number of elements in the vector. This | |
1865 | may not hold when there are multiple-types in the loop. | |
1866 | In this case, for some data-references in the loop the VF does not represent | |
1867 | the number of elements that fit in the vector. Therefore, instead of VF we | |
1868 | use TYPE_VECTOR_SUBPARTS. */ | |
1869 | ||
b8698a0f | 1870 | static tree |
e78410bf | 1871 | vect_gen_niters_for_prolog_loop (loop_vec_info loop_vinfo, tree loop_niters, int *bound) |
ebfd146a IR |
1872 | { |
1873 | struct data_reference *dr = LOOP_VINFO_UNALIGNED_DR (loop_vinfo); | |
1874 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
1875 | tree var; | |
1876 | gimple_seq stmts; | |
1877 | tree iters, iters_name; | |
1878 | edge pe; | |
1879 | basic_block new_bb; | |
1880 | gimple dr_stmt = DR_STMT (dr); | |
1881 | stmt_vec_info stmt_info = vinfo_for_stmt (dr_stmt); | |
1882 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
1883 | int vectype_align = TYPE_ALIGN (vectype) / BITS_PER_UNIT; | |
1884 | tree niters_type = TREE_TYPE (loop_niters); | |
ebfd146a IR |
1885 | int nelements = TYPE_VECTOR_SUBPARTS (vectype); |
1886 | ||
b8698a0f | 1887 | pe = loop_preheader_edge (loop); |
ebfd146a IR |
1888 | |
1889 | if (LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo) > 0) | |
1890 | { | |
720f5239 | 1891 | int npeel = LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo); |
ebfd146a | 1892 | |
73fbfcad | 1893 | if (dump_enabled_p ()) |
ccb3ad87 | 1894 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1895 | "known peeling = %d.\n", npeel); |
ebfd146a | 1896 | |
720f5239 | 1897 | iters = build_int_cst (niters_type, npeel); |
e78410bf | 1898 | *bound = LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo); |
ebfd146a IR |
1899 | } |
1900 | else | |
1901 | { | |
1902 | gimple_seq new_stmts = NULL; | |
d8ba5b19 RG |
1903 | bool negative = tree_int_cst_compare (DR_STEP (dr), size_zero_node) < 0; |
1904 | tree offset = negative | |
1905 | ? size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1) : NULL_TREE; | |
b8698a0f | 1906 | tree start_addr = vect_create_addr_base_for_vector_ref (dr_stmt, |
d8ba5b19 | 1907 | &new_stmts, offset, loop); |
96f9265a | 1908 | tree type = unsigned_type_for (TREE_TYPE (start_addr)); |
5aea1e76 UW |
1909 | tree vectype_align_minus_1 = build_int_cst (type, vectype_align - 1); |
1910 | HOST_WIDE_INT elem_size = | |
1911 | int_cst_value (TYPE_SIZE_UNIT (TREE_TYPE (vectype))); | |
1912 | tree elem_size_log = build_int_cst (type, exact_log2 (elem_size)); | |
ebfd146a IR |
1913 | tree nelements_minus_1 = build_int_cst (type, nelements - 1); |
1914 | tree nelements_tree = build_int_cst (type, nelements); | |
1915 | tree byte_misalign; | |
1916 | tree elem_misalign; | |
1917 | ||
1918 | new_bb = gsi_insert_seq_on_edge_immediate (pe, new_stmts); | |
1919 | gcc_assert (!new_bb); | |
b8698a0f | 1920 | |
5aea1e76 | 1921 | /* Create: byte_misalign = addr & (vectype_align - 1) */ |
b8698a0f | 1922 | byte_misalign = |
720f5239 | 1923 | fold_build2 (BIT_AND_EXPR, type, fold_convert (type, start_addr), |
5aea1e76 | 1924 | vectype_align_minus_1); |
b8698a0f | 1925 | |
ebfd146a IR |
1926 | /* Create: elem_misalign = byte_misalign / element_size */ |
1927 | elem_misalign = | |
1928 | fold_build2 (RSHIFT_EXPR, type, byte_misalign, elem_size_log); | |
1929 | ||
1930 | /* Create: (niters_type) (nelements - elem_misalign)&(nelements - 1) */ | |
d8ba5b19 RG |
1931 | if (negative) |
1932 | iters = fold_build2 (MINUS_EXPR, type, elem_misalign, nelements_tree); | |
1933 | else | |
1934 | iters = fold_build2 (MINUS_EXPR, type, nelements_tree, elem_misalign); | |
ebfd146a IR |
1935 | iters = fold_build2 (BIT_AND_EXPR, type, iters, nelements_minus_1); |
1936 | iters = fold_convert (niters_type, iters); | |
e78410bf | 1937 | *bound = nelements; |
ebfd146a IR |
1938 | } |
1939 | ||
1940 | /* Create: prolog_loop_niters = min (iters, loop_niters) */ | |
1941 | /* If the loop bound is known at compile time we already verified that it is | |
1942 | greater than vf; since the misalignment ('iters') is at most vf, there's | |
1943 | no need to generate the MIN_EXPR in this case. */ | |
1944 | if (TREE_CODE (loop_niters) != INTEGER_CST) | |
1945 | iters = fold_build2 (MIN_EXPR, niters_type, iters, loop_niters); | |
1946 | ||
73fbfcad | 1947 | if (dump_enabled_p ()) |
ebfd146a | 1948 | { |
ccb3ad87 | 1949 | dump_printf_loc (MSG_NOTE, vect_location, |
78c60e3d | 1950 | "niters for prolog loop: "); |
ccb3ad87 | 1951 | dump_generic_expr (MSG_NOTE, TDF_SLIM, iters); |
e645e942 | 1952 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
1953 | } |
1954 | ||
1955 | var = create_tmp_var (niters_type, "prolog_loop_niters"); | |
ebfd146a IR |
1956 | stmts = NULL; |
1957 | iters_name = force_gimple_operand (iters, &stmts, false, var); | |
1958 | ||
1959 | /* Insert stmt on loop preheader edge. */ | |
1960 | if (stmts) | |
1961 | { | |
1962 | basic_block new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
1963 | gcc_assert (!new_bb); | |
1964 | } | |
1965 | ||
b8698a0f | 1966 | return iters_name; |
ebfd146a IR |
1967 | } |
1968 | ||
1969 | ||
1970 | /* Function vect_update_init_of_dr | |
1971 | ||
1972 | NITERS iterations were peeled from LOOP. DR represents a data reference | |
1973 | in LOOP. This function updates the information recorded in DR to | |
b8698a0f | 1974 | account for the fact that the first NITERS iterations had already been |
ebfd146a IR |
1975 | executed. Specifically, it updates the OFFSET field of DR. */ |
1976 | ||
1977 | static void | |
1978 | vect_update_init_of_dr (struct data_reference *dr, tree niters) | |
1979 | { | |
1980 | tree offset = DR_OFFSET (dr); | |
b8698a0f | 1981 | |
ebfd146a IR |
1982 | niters = fold_build2 (MULT_EXPR, sizetype, |
1983 | fold_convert (sizetype, niters), | |
1984 | fold_convert (sizetype, DR_STEP (dr))); | |
587aa063 RG |
1985 | offset = fold_build2 (PLUS_EXPR, sizetype, |
1986 | fold_convert (sizetype, offset), niters); | |
ebfd146a IR |
1987 | DR_OFFSET (dr) = offset; |
1988 | } | |
1989 | ||
1990 | ||
1991 | /* Function vect_update_inits_of_drs | |
1992 | ||
b8698a0f L |
1993 | NITERS iterations were peeled from the loop represented by LOOP_VINFO. |
1994 | This function updates the information recorded for the data references in | |
1995 | the loop to account for the fact that the first NITERS iterations had | |
ebfd146a IR |
1996 | already been executed. Specifically, it updates the initial_condition of |
1997 | the access_function of all the data_references in the loop. */ | |
1998 | ||
1999 | static void | |
2000 | vect_update_inits_of_drs (loop_vec_info loop_vinfo, tree niters) | |
2001 | { | |
2002 | unsigned int i; | |
9771b263 | 2003 | vec<data_reference_p> datarefs = LOOP_VINFO_DATAREFS (loop_vinfo); |
ebfd146a | 2004 | struct data_reference *dr; |
78c60e3d | 2005 | |
73fbfcad | 2006 | if (dump_enabled_p ()) |
ccb3ad87 | 2007 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 2008 | "=== vect_update_inits_of_dr ===\n"); |
ebfd146a | 2009 | |
9771b263 | 2010 | FOR_EACH_VEC_ELT (datarefs, i, dr) |
ebfd146a IR |
2011 | vect_update_init_of_dr (dr, niters); |
2012 | } | |
2013 | ||
2014 | ||
2015 | /* Function vect_do_peeling_for_alignment | |
2016 | ||
2017 | Peel the first 'niters' iterations of the loop represented by LOOP_VINFO. | |
2018 | 'niters' is set to the misalignment of one of the data references in the | |
2019 | loop, thereby forcing it to refer to an aligned location at the beginning | |
2020 | of the execution of this loop. The data reference for which we are | |
2021 | peeling is recorded in LOOP_VINFO_UNALIGNED_DR. */ | |
2022 | ||
2023 | void | |
368117e8 RG |
2024 | vect_do_peeling_for_alignment (loop_vec_info loop_vinfo, |
2025 | unsigned int th, bool check_profitability) | |
ebfd146a IR |
2026 | { |
2027 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
2028 | tree niters_of_prolog_loop, ni_name; | |
2029 | tree n_iters; | |
b61b1f17 | 2030 | tree wide_prolog_niters; |
ebfd146a | 2031 | struct loop *new_loop; |
03fd03d5 | 2032 | int max_iter; |
e78410bf | 2033 | int bound = 0; |
ebfd146a | 2034 | |
73fbfcad | 2035 | if (dump_enabled_p ()) |
9cc1fb4b XDL |
2036 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, vect_location, |
2037 | "loop peeled for vectorization to enhance" | |
2038 | " alignment\n"); | |
ebfd146a IR |
2039 | |
2040 | initialize_original_copy_tables (); | |
2041 | ||
86290011 | 2042 | ni_name = vect_build_loop_niters (loop_vinfo, NULL); |
5d2eb24b | 2043 | niters_of_prolog_loop = vect_gen_niters_for_prolog_loop (loop_vinfo, |
e78410bf JH |
2044 | ni_name, |
2045 | &bound); | |
ebfd146a | 2046 | |
ebfd146a IR |
2047 | /* Peel the prolog loop and iterate it niters_of_prolog_loop. */ |
2048 | new_loop = | |
2049 | slpeel_tree_peel_loop_to_edge (loop, loop_preheader_edge (loop), | |
5d2eb24b | 2050 | &niters_of_prolog_loop, ni_name, true, |
e78410bf JH |
2051 | th, check_profitability, NULL_TREE, NULL, |
2052 | bound, | |
2053 | 0); | |
ebfd146a IR |
2054 | |
2055 | gcc_assert (new_loop); | |
2056 | #ifdef ENABLE_CHECKING | |
2057 | slpeel_verify_cfg_after_peeling (new_loop, loop); | |
2058 | #endif | |
22458c5a JH |
2059 | /* For vectorization factor N, we need to copy at most N-1 values |
2060 | for alignment and this means N-2 loopback edge executions. */ | |
2061 | max_iter = LOOP_VINFO_VECT_FACTOR (loop_vinfo) - 2; | |
368117e8 | 2062 | if (check_profitability) |
22458c5a | 2063 | max_iter = MAX (max_iter, (int) th - 1); |
27bcd47c | 2064 | record_niter_bound (new_loop, double_int::from_shwi (max_iter), false, true); |
ccb3ad87 | 2065 | dump_printf (MSG_NOTE, |
78c60e3d SS |
2066 | "Setting upper bound of nb iterations for prologue " |
2067 | "loop to %d\n", max_iter); | |
ebfd146a IR |
2068 | |
2069 | /* Update number of times loop executes. */ | |
2070 | n_iters = LOOP_VINFO_NITERS (loop_vinfo); | |
2071 | LOOP_VINFO_NITERS (loop_vinfo) = fold_build2 (MINUS_EXPR, | |
2072 | TREE_TYPE (n_iters), n_iters, niters_of_prolog_loop); | |
2073 | ||
5d2eb24b IR |
2074 | if (types_compatible_p (sizetype, TREE_TYPE (niters_of_prolog_loop))) |
2075 | wide_prolog_niters = niters_of_prolog_loop; | |
2076 | else | |
2077 | { | |
2078 | gimple_seq seq = NULL; | |
2079 | edge pe = loop_preheader_edge (loop); | |
2080 | tree wide_iters = fold_convert (sizetype, niters_of_prolog_loop); | |
2081 | tree var = create_tmp_var (sizetype, "prolog_loop_adjusted_niters"); | |
5d2eb24b IR |
2082 | wide_prolog_niters = force_gimple_operand (wide_iters, &seq, false, |
2083 | var); | |
2084 | if (seq) | |
2085 | { | |
2086 | /* Insert stmt on loop preheader edge. */ | |
2087 | basic_block new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
2088 | gcc_assert (!new_bb); | |
2089 | } | |
2090 | } | |
2091 | ||
ebfd146a | 2092 | /* Update the init conditions of the access functions of all data refs. */ |
b61b1f17 | 2093 | vect_update_inits_of_drs (loop_vinfo, wide_prolog_niters); |
ebfd146a IR |
2094 | |
2095 | /* After peeling we have to reset scalar evolution analyzer. */ | |
2096 | scev_reset (); | |
2097 | ||
2098 | free_original_copy_tables (); | |
2099 | } | |
2100 | ||
2101 | ||
2102 | /* Function vect_create_cond_for_align_checks. | |
2103 | ||
2104 | Create a conditional expression that represents the alignment checks for | |
2105 | all of data references (array element references) whose alignment must be | |
2106 | checked at runtime. | |
2107 | ||
2108 | Input: | |
2109 | COND_EXPR - input conditional expression. New conditions will be chained | |
2110 | with logical AND operation. | |
2111 | LOOP_VINFO - two fields of the loop information are used. | |
2112 | LOOP_VINFO_PTR_MASK is the mask used to check the alignment. | |
2113 | LOOP_VINFO_MAY_MISALIGN_STMTS contains the refs to be checked. | |
2114 | ||
2115 | Output: | |
2116 | COND_EXPR_STMT_LIST - statements needed to construct the conditional | |
2117 | expression. | |
2118 | The returned value is the conditional expression to be used in the if | |
2119 | statement that controls which version of the loop gets executed at runtime. | |
2120 | ||
2121 | The algorithm makes two assumptions: | |
2122 | 1) The number of bytes "n" in a vector is a power of 2. | |
2123 | 2) An address "a" is aligned if a%n is zero and that this | |
2124 | test can be done as a&(n-1) == 0. For example, for 16 | |
2125 | byte vectors the test is a&0xf == 0. */ | |
2126 | ||
2127 | static void | |
2128 | vect_create_cond_for_align_checks (loop_vec_info loop_vinfo, | |
2129 | tree *cond_expr, | |
2130 | gimple_seq *cond_expr_stmt_list) | |
2131 | { | |
2132 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
9771b263 | 2133 | vec<gimple> may_misalign_stmts |
ebfd146a IR |
2134 | = LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo); |
2135 | gimple ref_stmt; | |
2136 | int mask = LOOP_VINFO_PTR_MASK (loop_vinfo); | |
2137 | tree mask_cst; | |
2138 | unsigned int i; | |
ebfd146a IR |
2139 | tree int_ptrsize_type; |
2140 | char tmp_name[20]; | |
2141 | tree or_tmp_name = NULL_TREE; | |
83d5977e | 2142 | tree and_tmp_name; |
ebfd146a IR |
2143 | gimple and_stmt; |
2144 | tree ptrsize_zero; | |
2145 | tree part_cond_expr; | |
2146 | ||
2147 | /* Check that mask is one less than a power of 2, i.e., mask is | |
2148 | all zeros followed by all ones. */ | |
2149 | gcc_assert ((mask != 0) && ((mask & (mask+1)) == 0)); | |
2150 | ||
96f9265a | 2151 | int_ptrsize_type = signed_type_for (ptr_type_node); |
ebfd146a IR |
2152 | |
2153 | /* Create expression (mask & (dr_1 || ... || dr_n)) where dr_i is the address | |
2154 | of the first vector of the i'th data reference. */ | |
2155 | ||
9771b263 | 2156 | FOR_EACH_VEC_ELT (may_misalign_stmts, i, ref_stmt) |
ebfd146a IR |
2157 | { |
2158 | gimple_seq new_stmt_list = NULL; | |
2159 | tree addr_base; | |
83d5977e RG |
2160 | tree addr_tmp_name; |
2161 | tree new_or_tmp_name; | |
ebfd146a | 2162 | gimple addr_stmt, or_stmt; |
d8ba5b19 RG |
2163 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (ref_stmt); |
2164 | tree vectype = STMT_VINFO_VECTYPE (stmt_vinfo); | |
2165 | bool negative = tree_int_cst_compare | |
2166 | (DR_STEP (STMT_VINFO_DATA_REF (stmt_vinfo)), size_zero_node) < 0; | |
2167 | tree offset = negative | |
2168 | ? size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1) : NULL_TREE; | |
ebfd146a IR |
2169 | |
2170 | /* create: addr_tmp = (int)(address_of_first_vector) */ | |
2171 | addr_base = | |
2172 | vect_create_addr_base_for_vector_ref (ref_stmt, &new_stmt_list, | |
d8ba5b19 | 2173 | offset, loop); |
ebfd146a IR |
2174 | if (new_stmt_list != NULL) |
2175 | gimple_seq_add_seq (cond_expr_stmt_list, new_stmt_list); | |
2176 | ||
83d5977e RG |
2177 | sprintf (tmp_name, "addr2int%d", i); |
2178 | addr_tmp_name = make_temp_ssa_name (int_ptrsize_type, NULL, tmp_name); | |
ebfd146a IR |
2179 | addr_stmt = gimple_build_assign_with_ops (NOP_EXPR, addr_tmp_name, |
2180 | addr_base, NULL_TREE); | |
ebfd146a IR |
2181 | gimple_seq_add_stmt (cond_expr_stmt_list, addr_stmt); |
2182 | ||
2183 | /* The addresses are OR together. */ | |
2184 | ||
2185 | if (or_tmp_name != NULL_TREE) | |
2186 | { | |
2187 | /* create: or_tmp = or_tmp | addr_tmp */ | |
83d5977e RG |
2188 | sprintf (tmp_name, "orptrs%d", i); |
2189 | new_or_tmp_name = make_temp_ssa_name (int_ptrsize_type, NULL, tmp_name); | |
ebfd146a IR |
2190 | or_stmt = gimple_build_assign_with_ops (BIT_IOR_EXPR, |
2191 | new_or_tmp_name, | |
2192 | or_tmp_name, addr_tmp_name); | |
ebfd146a IR |
2193 | gimple_seq_add_stmt (cond_expr_stmt_list, or_stmt); |
2194 | or_tmp_name = new_or_tmp_name; | |
2195 | } | |
2196 | else | |
2197 | or_tmp_name = addr_tmp_name; | |
2198 | ||
2199 | } /* end for i */ | |
2200 | ||
2201 | mask_cst = build_int_cst (int_ptrsize_type, mask); | |
2202 | ||
2203 | /* create: and_tmp = or_tmp & mask */ | |
83d5977e | 2204 | and_tmp_name = make_temp_ssa_name (int_ptrsize_type, NULL, "andmask"); |
ebfd146a IR |
2205 | |
2206 | and_stmt = gimple_build_assign_with_ops (BIT_AND_EXPR, and_tmp_name, | |
2207 | or_tmp_name, mask_cst); | |
ebfd146a IR |
2208 | gimple_seq_add_stmt (cond_expr_stmt_list, and_stmt); |
2209 | ||
2210 | /* Make and_tmp the left operand of the conditional test against zero. | |
2211 | if and_tmp has a nonzero bit then some address is unaligned. */ | |
2212 | ptrsize_zero = build_int_cst (int_ptrsize_type, 0); | |
2213 | part_cond_expr = fold_build2 (EQ_EXPR, boolean_type_node, | |
2214 | and_tmp_name, ptrsize_zero); | |
2215 | if (*cond_expr) | |
2216 | *cond_expr = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, | |
2217 | *cond_expr, part_cond_expr); | |
2218 | else | |
2219 | *cond_expr = part_cond_expr; | |
2220 | } | |
2221 | ||
ebfd146a IR |
2222 | /* Function vect_create_cond_for_alias_checks. |
2223 | ||
2224 | Create a conditional expression that represents the run-time checks for | |
2225 | overlapping of address ranges represented by a list of data references | |
2226 | relations passed as input. | |
2227 | ||
2228 | Input: | |
2229 | COND_EXPR - input conditional expression. New conditions will be chained | |
a05a89fa CH |
2230 | with logical AND operation. If it is NULL, then the function |
2231 | is used to return the number of alias checks. | |
ebfd146a IR |
2232 | LOOP_VINFO - field LOOP_VINFO_MAY_ALIAS_STMTS contains the list of ddrs |
2233 | to be checked. | |
2234 | ||
2235 | Output: | |
2236 | COND_EXPR - conditional expression. | |
ebfd146a | 2237 | |
a05a89fa | 2238 | The returned COND_EXPR is the conditional expression to be used in the if |
ebfd146a IR |
2239 | statement that controls which version of the loop gets executed at runtime. |
2240 | */ | |
2241 | ||
a05a89fa | 2242 | void |
4bdd44c4 | 2243 | vect_create_cond_for_alias_checks (loop_vec_info loop_vinfo, tree * cond_expr) |
ebfd146a | 2244 | { |
a05a89fa CH |
2245 | vec<dr_addr_with_seg_len_pair_t> comp_alias_ddrs = |
2246 | LOOP_VINFO_COMP_ALIAS_DDRS (loop_vinfo); | |
2247 | tree part_cond_expr; | |
ebfd146a IR |
2248 | |
2249 | /* Create expression | |
36fc3799 RS |
2250 | ((store_ptr_0 + store_segment_length_0) <= load_ptr_0) |
2251 | || (load_ptr_0 + load_segment_length_0) <= store_ptr_0)) | |
b8698a0f | 2252 | && |
ebfd146a IR |
2253 | ... |
2254 | && | |
36fc3799 RS |
2255 | ((store_ptr_n + store_segment_length_n) <= load_ptr_n) |
2256 | || (load_ptr_n + load_segment_length_n) <= store_ptr_n)) */ | |
ebfd146a | 2257 | |
a05a89fa | 2258 | if (comp_alias_ddrs.is_empty ()) |
ebfd146a IR |
2259 | return; |
2260 | ||
a05a89fa | 2261 | for (size_t i = 0, s = comp_alias_ddrs.length (); i < s; ++i) |
ebfd146a | 2262 | { |
a05a89fa CH |
2263 | const dr_addr_with_seg_len& dr_a = comp_alias_ddrs[i].first; |
2264 | const dr_addr_with_seg_len& dr_b = comp_alias_ddrs[i].second; | |
2265 | tree segment_length_a = dr_a.seg_len; | |
2266 | tree segment_length_b = dr_b.seg_len; | |
ebfd146a | 2267 | |
a05a89fa CH |
2268 | tree addr_base_a |
2269 | = fold_build_pointer_plus (dr_a.basic_addr, dr_a.offset); | |
2270 | tree addr_base_b | |
2271 | = fold_build_pointer_plus (dr_b.basic_addr, dr_b.offset); | |
ebfd146a | 2272 | |
73fbfcad | 2273 | if (dump_enabled_p ()) |
ebfd146a | 2274 | { |
ccb3ad87 | 2275 | dump_printf_loc (MSG_NOTE, vect_location, |
a05a89fa CH |
2276 | "create runtime check for data references "); |
2277 | dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (dr_a.dr)); | |
ccb3ad87 | 2278 | dump_printf (MSG_NOTE, " and "); |
a05a89fa CH |
2279 | dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (dr_b.dr)); |
2280 | dump_printf (MSG_NOTE, "\n"); | |
ebfd146a IR |
2281 | } |
2282 | ||
a05a89fa CH |
2283 | tree seg_a_min = addr_base_a; |
2284 | tree seg_a_max = fold_build_pointer_plus (addr_base_a, segment_length_a); | |
2285 | if (tree_int_cst_compare (DR_STEP (dr_a.dr), size_zero_node) < 0) | |
d8ba5b19 RG |
2286 | seg_a_min = seg_a_max, seg_a_max = addr_base_a; |
2287 | ||
a05a89fa CH |
2288 | tree seg_b_min = addr_base_b; |
2289 | tree seg_b_max = fold_build_pointer_plus (addr_base_b, segment_length_b); | |
2290 | if (tree_int_cst_compare (DR_STEP (dr_b.dr), size_zero_node) < 0) | |
d8ba5b19 | 2291 | seg_b_min = seg_b_max, seg_b_max = addr_base_b; |
ebfd146a | 2292 | |
b8698a0f | 2293 | part_cond_expr = |
ebfd146a | 2294 | fold_build2 (TRUTH_OR_EXPR, boolean_type_node, |
36fc3799 RS |
2295 | fold_build2 (LE_EXPR, boolean_type_node, seg_a_max, seg_b_min), |
2296 | fold_build2 (LE_EXPR, boolean_type_node, seg_b_max, seg_a_min)); | |
b8698a0f | 2297 | |
ebfd146a IR |
2298 | if (*cond_expr) |
2299 | *cond_expr = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, | |
2300 | *cond_expr, part_cond_expr); | |
2301 | else | |
2302 | *cond_expr = part_cond_expr; | |
2303 | } | |
ebfd146a | 2304 | |
73fbfcad | 2305 | if (dump_enabled_p ()) |
ccb3ad87 | 2306 | dump_printf_loc (MSG_NOTE, vect_location, |
78c60e3d | 2307 | "created %u versioning for alias checks.\n", |
a05a89fa CH |
2308 | comp_alias_ddrs.length ()); |
2309 | ||
2310 | comp_alias_ddrs.release (); | |
ebfd146a IR |
2311 | } |
2312 | ||
2313 | ||
2314 | /* Function vect_loop_versioning. | |
b8698a0f | 2315 | |
ebfd146a IR |
2316 | If the loop has data references that may or may not be aligned or/and |
2317 | has data reference relations whose independence was not proven then | |
2318 | two versions of the loop need to be generated, one which is vectorized | |
2319 | and one which isn't. A test is then generated to control which of the | |
2320 | loops is executed. The test checks for the alignment of all of the | |
2321 | data references that may or may not be aligned. An additional | |
2322 | sequence of runtime tests is generated for each pairs of DDRs whose | |
b8698a0f L |
2323 | independence was not proven. The vectorized version of loop is |
2324 | executed only if both alias and alignment tests are passed. | |
2325 | ||
ebfd146a | 2326 | The test generated to check which version of loop is executed |
b8698a0f | 2327 | is modified to also check for profitability as indicated by the |
86290011 RG |
2328 | cost model initially. |
2329 | ||
2330 | The versioning precondition(s) are placed in *COND_EXPR and | |
d68d56b5 | 2331 | *COND_EXPR_STMT_LIST. */ |
ebfd146a IR |
2332 | |
2333 | void | |
368117e8 RG |
2334 | vect_loop_versioning (loop_vec_info loop_vinfo, |
2335 | unsigned int th, bool check_profitability) | |
ebfd146a IR |
2336 | { |
2337 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a IR |
2338 | basic_block condition_bb; |
2339 | gimple_stmt_iterator gsi, cond_exp_gsi; | |
2340 | basic_block merge_bb; | |
2341 | basic_block new_exit_bb; | |
2342 | edge new_exit_e, e; | |
2343 | gimple orig_phi, new_phi; | |
368117e8 | 2344 | tree cond_expr = NULL_TREE; |
d68d56b5 | 2345 | gimple_seq cond_expr_stmt_list = NULL; |
ebfd146a IR |
2346 | tree arg; |
2347 | unsigned prob = 4 * REG_BR_PROB_BASE / 5; | |
2348 | gimple_seq gimplify_stmt_list = NULL; | |
2349 | tree scalar_loop_iters = LOOP_VINFO_NITERS (loop_vinfo); | |
9cc1fb4b XDL |
2350 | bool version_align = LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo); |
2351 | bool version_alias = LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo); | |
ebfd146a | 2352 | |
368117e8 RG |
2353 | if (check_profitability) |
2354 | { | |
2355 | cond_expr = fold_build2 (GT_EXPR, boolean_type_node, scalar_loop_iters, | |
2356 | build_int_cst (TREE_TYPE (scalar_loop_iters), th)); | |
2357 | cond_expr = force_gimple_operand_1 (cond_expr, &cond_expr_stmt_list, | |
2358 | is_gimple_condexpr, NULL_TREE); | |
2359 | } | |
ebfd146a | 2360 | |
9cc1fb4b | 2361 | if (version_align) |
d68d56b5 RG |
2362 | vect_create_cond_for_align_checks (loop_vinfo, &cond_expr, |
2363 | &cond_expr_stmt_list); | |
ebfd146a | 2364 | |
9cc1fb4b | 2365 | if (version_alias) |
4bdd44c4 | 2366 | vect_create_cond_for_alias_checks (loop_vinfo, &cond_expr); |
86290011 | 2367 | |
d68d56b5 RG |
2368 | cond_expr = force_gimple_operand_1 (cond_expr, &gimplify_stmt_list, |
2369 | is_gimple_condexpr, NULL_TREE); | |
2370 | gimple_seq_add_seq (&cond_expr_stmt_list, gimplify_stmt_list); | |
ebfd146a IR |
2371 | |
2372 | initialize_original_copy_tables (); | |
d68d56b5 | 2373 | loop_version (loop, cond_expr, &condition_bb, |
0f900dfa | 2374 | prob, prob, REG_BR_PROB_BASE - prob, true); |
9cc1fb4b XDL |
2375 | |
2376 | if (LOCATION_LOCUS (vect_location) != UNKNOWN_LOC | |
2377 | && dump_enabled_p ()) | |
2378 | { | |
2379 | if (version_alias) | |
2380 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, vect_location, | |
2381 | "loop versioned for vectorization because of " | |
2382 | "possible aliasing\n"); | |
2383 | if (version_align) | |
2384 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, vect_location, | |
2385 | "loop versioned for vectorization to enhance " | |
2386 | "alignment\n"); | |
2387 | ||
2388 | } | |
c3284718 | 2389 | free_original_copy_tables (); |
ebfd146a | 2390 | |
b8698a0f | 2391 | /* Loop versioning violates an assumption we try to maintain during |
ebfd146a IR |
2392 | vectorization - that the loop exit block has a single predecessor. |
2393 | After versioning, the exit block of both loop versions is the same | |
2394 | basic block (i.e. it has two predecessors). Just in order to simplify | |
2395 | following transformations in the vectorizer, we fix this situation | |
2396 | here by adding a new (empty) block on the exit-edge of the loop, | |
2397 | with the proper loop-exit phis to maintain loop-closed-form. */ | |
b8698a0f | 2398 | |
ebfd146a IR |
2399 | merge_bb = single_exit (loop)->dest; |
2400 | gcc_assert (EDGE_COUNT (merge_bb->preds) == 2); | |
2401 | new_exit_bb = split_edge (single_exit (loop)); | |
2402 | new_exit_e = single_exit (loop); | |
2403 | e = EDGE_SUCC (new_exit_bb, 0); | |
2404 | ||
2405 | for (gsi = gsi_start_phis (merge_bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
2406 | { | |
070ecdfd | 2407 | tree new_res; |
ebfd146a | 2408 | orig_phi = gsi_stmt (gsi); |
070ecdfd RG |
2409 | new_res = copy_ssa_name (PHI_RESULT (orig_phi), NULL); |
2410 | new_phi = create_phi_node (new_res, new_exit_bb); | |
ebfd146a | 2411 | arg = PHI_ARG_DEF_FROM_EDGE (orig_phi, e); |
b8698a0f | 2412 | add_phi_arg (new_phi, arg, new_exit_e, |
9e227d60 | 2413 | gimple_phi_arg_location_from_edge (orig_phi, e)); |
684f25f4 | 2414 | adjust_phi_and_debug_stmts (orig_phi, e, PHI_RESULT (new_phi)); |
b8698a0f | 2415 | } |
ebfd146a | 2416 | |
6f978a2a CH |
2417 | |
2418 | /* Extract load statements on memrefs with zero-stride accesses. */ | |
2419 | ||
2420 | if (LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo)) | |
2421 | { | |
2422 | /* In the loop body, we iterate each statement to check if it is a load. | |
2423 | Then we check the DR_STEP of the data reference. If DR_STEP is zero, | |
2424 | then we will hoist the load statement to the loop preheader. */ | |
2425 | ||
2426 | basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); | |
2427 | int nbbs = loop->num_nodes; | |
2428 | ||
2429 | for (int i = 0; i < nbbs; ++i) | |
2430 | { | |
2431 | for (gimple_stmt_iterator si = gsi_start_bb (bbs[i]); | |
2432 | !gsi_end_p (si);) | |
2433 | { | |
2434 | gimple stmt = gsi_stmt (si); | |
2435 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2436 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
2437 | ||
2438 | if (is_gimple_assign (stmt) | |
2439 | && (!dr | |
2440 | || (DR_IS_READ (dr) && integer_zerop (DR_STEP (dr))))) | |
2441 | { | |
2442 | bool hoist = true; | |
2443 | ssa_op_iter iter; | |
2444 | tree var; | |
2445 | ||
2446 | /* We hoist a statement if all SSA uses in it are defined | |
2447 | outside of the loop. */ | |
2448 | FOR_EACH_SSA_TREE_OPERAND (var, stmt, iter, SSA_OP_USE) | |
2449 | { | |
2450 | gimple def = SSA_NAME_DEF_STMT (var); | |
2451 | if (!gimple_nop_p (def) | |
2452 | && flow_bb_inside_loop_p (loop, gimple_bb (def))) | |
2453 | { | |
2454 | hoist = false; | |
2455 | break; | |
2456 | } | |
2457 | } | |
2458 | ||
2459 | if (hoist) | |
2460 | { | |
2461 | if (dr) | |
2462 | gimple_set_vuse (stmt, NULL); | |
2463 | ||
2464 | gsi_remove (&si, false); | |
2465 | gsi_insert_on_edge_immediate (loop_preheader_edge (loop), | |
2466 | stmt); | |
2467 | ||
2468 | if (dump_enabled_p ()) | |
2469 | { | |
2470 | dump_printf_loc | |
2471 | (MSG_NOTE, vect_location, | |
2472 | "hoisting out of the vectorized loop: "); | |
2473 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); | |
2474 | dump_printf (MSG_NOTE, "\n"); | |
2475 | } | |
2476 | continue; | |
2477 | } | |
2478 | } | |
2479 | gsi_next (&si); | |
2480 | } | |
2481 | } | |
2482 | } | |
2483 | ||
ebfd146a IR |
2484 | /* End loop-exit-fixes after versioning. */ |
2485 | ||
d68d56b5 | 2486 | if (cond_expr_stmt_list) |
ebfd146a IR |
2487 | { |
2488 | cond_exp_gsi = gsi_last_bb (condition_bb); | |
d68d56b5 | 2489 | gsi_insert_seq_before (&cond_exp_gsi, cond_expr_stmt_list, |
86290011 | 2490 | GSI_SAME_STMT); |
ebfd146a | 2491 | } |
90eb75f2 | 2492 | update_ssa (TODO_update_ssa); |
ebfd146a | 2493 | } |