]>
Commit | Line | Data |
---|---|---|
b8698a0f | 1 | /* Vectorizer Specific Loop Manipulations |
5624e564 | 2 | Copyright (C) 2003-2015 Free Software Foundation, Inc. |
b8698a0f | 3 | Contributed by Dorit Naishlos <dorit@il.ibm.com> |
ebfd146a IR |
4 | and Ira Rosen <irar@il.ibm.com> |
5 | ||
6 | This file is part of GCC. | |
7 | ||
8 | GCC is free software; you can redistribute it and/or modify it under | |
9 | the terms of the GNU General Public License as published by the Free | |
10 | Software Foundation; either version 3, or (at your option) any later | |
11 | version. | |
12 | ||
13 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
14 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
15 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
16 | for more details. | |
17 | ||
18 | You should have received a copy of the GNU General Public License | |
19 | along with GCC; see the file COPYING3. If not see | |
20 | <http://www.gnu.org/licenses/>. */ | |
21 | ||
22 | #include "config.h" | |
23 | #include "system.h" | |
24 | #include "coretypes.h" | |
c7131fb2 | 25 | #include "backend.h" |
957060b5 | 26 | #include "hard-reg-set.h" |
40e23961 | 27 | #include "tree.h" |
c7131fb2 | 28 | #include "gimple.h" |
957060b5 AM |
29 | #include "cfghooks.h" |
30 | #include "tree-pass.h" | |
c7131fb2 | 31 | #include "ssa.h" |
957060b5 AM |
32 | #include "gimple-pretty-print.h" |
33 | #include "diagnostic-core.h" | |
34 | #include "dumpfile.h" | |
c7131fb2 AM |
35 | #include "alias.h" |
36 | #include "fold-const.h" | |
60393bbc | 37 | #include "cfganal.h" |
2fb9a547 | 38 | #include "internal-fn.h" |
45b0be94 | 39 | #include "gimplify.h" |
5be5c238 | 40 | #include "gimple-iterator.h" |
18f429e2 | 41 | #include "gimplify-me.h" |
442b4905 | 42 | #include "tree-cfg.h" |
e28030cf | 43 | #include "tree-ssa-loop-manip.h" |
442b4905 | 44 | #include "tree-into-ssa.h" |
7a300452 | 45 | #include "tree-ssa.h" |
ebfd146a | 46 | #include "cfgloop.h" |
ebfd146a IR |
47 | #include "tree-scalar-evolution.h" |
48 | #include "tree-vectorizer.h" | |
49 | #include "langhooks.h" | |
50 | ||
51 | /************************************************************************* | |
52 | Simple Loop Peeling Utilities | |
53 | ||
54 | Utilities to support loop peeling for vectorization purposes. | |
55 | *************************************************************************/ | |
56 | ||
57 | ||
58 | /* Renames the use *OP_P. */ | |
59 | ||
60 | static void | |
61 | rename_use_op (use_operand_p op_p) | |
62 | { | |
63 | tree new_name; | |
64 | ||
65 | if (TREE_CODE (USE_FROM_PTR (op_p)) != SSA_NAME) | |
66 | return; | |
67 | ||
68 | new_name = get_current_def (USE_FROM_PTR (op_p)); | |
69 | ||
70 | /* Something defined outside of the loop. */ | |
71 | if (!new_name) | |
72 | return; | |
73 | ||
74 | /* An ordinary ssa name defined in the loop. */ | |
75 | ||
76 | SET_USE (op_p, new_name); | |
77 | } | |
78 | ||
79 | ||
a6c51a12 YR |
80 | /* Renames the variables in basic block BB. Allow renaming of PHI argumnets |
81 | on edges incoming from outer-block header if RENAME_FROM_OUTER_LOOP is | |
82 | true. */ | |
ebfd146a | 83 | |
2cfc56b9 | 84 | static void |
a6c51a12 | 85 | rename_variables_in_bb (basic_block bb, bool rename_from_outer_loop) |
ebfd146a | 86 | { |
355fe088 | 87 | gimple *stmt; |
ebfd146a IR |
88 | use_operand_p use_p; |
89 | ssa_op_iter iter; | |
90 | edge e; | |
91 | edge_iterator ei; | |
92 | struct loop *loop = bb->loop_father; | |
a6c51a12 YR |
93 | struct loop *outer_loop = NULL; |
94 | ||
95 | if (rename_from_outer_loop) | |
96 | { | |
97 | gcc_assert (loop); | |
98 | outer_loop = loop_outer (loop); | |
99 | } | |
ebfd146a | 100 | |
538dd0b7 DM |
101 | for (gimple_stmt_iterator gsi = gsi_start_bb (bb); !gsi_end_p (gsi); |
102 | gsi_next (&gsi)) | |
ebfd146a IR |
103 | { |
104 | stmt = gsi_stmt (gsi); | |
105 | FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_ALL_USES) | |
106 | rename_use_op (use_p); | |
107 | } | |
108 | ||
2cfc56b9 | 109 | FOR_EACH_EDGE (e, ei, bb->preds) |
ebfd146a | 110 | { |
a6c51a12 YR |
111 | if (!flow_bb_inside_loop_p (loop, e->src) |
112 | && (!rename_from_outer_loop || e->src != outer_loop->header)) | |
ebfd146a | 113 | continue; |
538dd0b7 DM |
114 | for (gphi_iterator gsi = gsi_start_phis (bb); !gsi_end_p (gsi); |
115 | gsi_next (&gsi)) | |
116 | rename_use_op (PHI_ARG_DEF_PTR_FROM_EDGE (gsi.phi (), e)); | |
ebfd146a IR |
117 | } |
118 | } | |
119 | ||
120 | ||
a79683d5 | 121 | struct adjust_info |
684f25f4 AO |
122 | { |
123 | tree from, to; | |
124 | basic_block bb; | |
a79683d5 | 125 | }; |
684f25f4 | 126 | |
684f25f4 AO |
127 | /* A stack of values to be adjusted in debug stmts. We have to |
128 | process them LIFO, so that the closest substitution applies. If we | |
129 | processed them FIFO, without the stack, we might substitute uses | |
130 | with a PHI DEF that would soon become non-dominant, and when we got | |
131 | to the suitable one, it wouldn't have anything to substitute any | |
132 | more. */ | |
ff4c81cc | 133 | static vec<adjust_info, va_heap> adjust_vec; |
684f25f4 AO |
134 | |
135 | /* Adjust any debug stmts that referenced AI->from values to use the | |
136 | loop-closed AI->to, if the references are dominated by AI->bb and | |
137 | not by the definition of AI->from. */ | |
138 | ||
139 | static void | |
140 | adjust_debug_stmts_now (adjust_info *ai) | |
141 | { | |
142 | basic_block bbphi = ai->bb; | |
143 | tree orig_def = ai->from; | |
144 | tree new_def = ai->to; | |
145 | imm_use_iterator imm_iter; | |
355fe088 | 146 | gimple *stmt; |
684f25f4 AO |
147 | basic_block bbdef = gimple_bb (SSA_NAME_DEF_STMT (orig_def)); |
148 | ||
149 | gcc_assert (dom_info_available_p (CDI_DOMINATORS)); | |
150 | ||
151 | /* Adjust any debug stmts that held onto non-loop-closed | |
152 | references. */ | |
153 | FOR_EACH_IMM_USE_STMT (stmt, imm_iter, orig_def) | |
154 | { | |
155 | use_operand_p use_p; | |
156 | basic_block bbuse; | |
157 | ||
158 | if (!is_gimple_debug (stmt)) | |
159 | continue; | |
160 | ||
161 | gcc_assert (gimple_debug_bind_p (stmt)); | |
162 | ||
163 | bbuse = gimple_bb (stmt); | |
164 | ||
165 | if ((bbuse == bbphi | |
166 | || dominated_by_p (CDI_DOMINATORS, bbuse, bbphi)) | |
167 | && !(bbuse == bbdef | |
168 | || dominated_by_p (CDI_DOMINATORS, bbuse, bbdef))) | |
169 | { | |
170 | if (new_def) | |
171 | FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter) | |
172 | SET_USE (use_p, new_def); | |
173 | else | |
174 | { | |
175 | gimple_debug_bind_reset_value (stmt); | |
176 | update_stmt (stmt); | |
177 | } | |
178 | } | |
179 | } | |
180 | } | |
181 | ||
182 | /* Adjust debug stmts as scheduled before. */ | |
183 | ||
184 | static void | |
185 | adjust_vec_debug_stmts (void) | |
186 | { | |
187 | if (!MAY_HAVE_DEBUG_STMTS) | |
188 | return; | |
189 | ||
9771b263 | 190 | gcc_assert (adjust_vec.exists ()); |
684f25f4 | 191 | |
9771b263 | 192 | while (!adjust_vec.is_empty ()) |
684f25f4 | 193 | { |
9771b263 DN |
194 | adjust_debug_stmts_now (&adjust_vec.last ()); |
195 | adjust_vec.pop (); | |
684f25f4 AO |
196 | } |
197 | ||
9771b263 | 198 | adjust_vec.release (); |
684f25f4 AO |
199 | } |
200 | ||
201 | /* Adjust any debug stmts that referenced FROM values to use the | |
202 | loop-closed TO, if the references are dominated by BB and not by | |
203 | the definition of FROM. If adjust_vec is non-NULL, adjustments | |
204 | will be postponed until adjust_vec_debug_stmts is called. */ | |
205 | ||
206 | static void | |
207 | adjust_debug_stmts (tree from, tree to, basic_block bb) | |
208 | { | |
209 | adjust_info ai; | |
210 | ||
a471762f RG |
211 | if (MAY_HAVE_DEBUG_STMTS |
212 | && TREE_CODE (from) == SSA_NAME | |
a52ca739 | 213 | && ! SSA_NAME_IS_DEFAULT_DEF (from) |
a471762f | 214 | && ! virtual_operand_p (from)) |
684f25f4 AO |
215 | { |
216 | ai.from = from; | |
217 | ai.to = to; | |
218 | ai.bb = bb; | |
219 | ||
9771b263 DN |
220 | if (adjust_vec.exists ()) |
221 | adjust_vec.safe_push (ai); | |
684f25f4 AO |
222 | else |
223 | adjust_debug_stmts_now (&ai); | |
224 | } | |
225 | } | |
226 | ||
227 | /* Change E's phi arg in UPDATE_PHI to NEW_DEF, and record information | |
228 | to adjust any debug stmts that referenced the old phi arg, | |
229 | presumably non-loop-closed references left over from other | |
230 | transformations. */ | |
231 | ||
232 | static void | |
355fe088 | 233 | adjust_phi_and_debug_stmts (gimple *update_phi, edge e, tree new_def) |
684f25f4 AO |
234 | { |
235 | tree orig_def = PHI_ARG_DEF_FROM_EDGE (update_phi, e); | |
236 | ||
237 | SET_PHI_ARG_DEF (update_phi, e->dest_idx, new_def); | |
238 | ||
239 | if (MAY_HAVE_DEBUG_STMTS) | |
240 | adjust_debug_stmts (orig_def, PHI_RESULT (update_phi), | |
241 | gimple_bb (update_phi)); | |
242 | } | |
243 | ||
ebfd146a | 244 | |
ebfd146a IR |
245 | /* Update PHI nodes for a guard of the LOOP. |
246 | ||
247 | Input: | |
248 | - LOOP, GUARD_EDGE: LOOP is a loop for which we added guard code that | |
249 | controls whether LOOP is to be executed. GUARD_EDGE is the edge that | |
250 | originates from the guard-bb, skips LOOP and reaches the (unique) exit | |
251 | bb of LOOP. This loop-exit-bb is an empty bb with one successor. | |
252 | We denote this bb NEW_MERGE_BB because before the guard code was added | |
253 | it had a single predecessor (the LOOP header), and now it became a merge | |
254 | point of two paths - the path that ends with the LOOP exit-edge, and | |
255 | the path that ends with GUARD_EDGE. | |
256 | - NEW_EXIT_BB: New basic block that is added by this function between LOOP | |
257 | and NEW_MERGE_BB. It is used to place loop-closed-ssa-form exit-phis. | |
258 | ||
259 | ===> The CFG before the guard-code was added: | |
260 | LOOP_header_bb: | |
261 | loop_body | |
262 | if (exit_loop) goto update_bb | |
263 | else goto LOOP_header_bb | |
264 | update_bb: | |
265 | ||
266 | ==> The CFG after the guard-code was added: | |
267 | guard_bb: | |
268 | if (LOOP_guard_condition) goto new_merge_bb | |
269 | else goto LOOP_header_bb | |
270 | LOOP_header_bb: | |
271 | loop_body | |
272 | if (exit_loop_condition) goto new_merge_bb | |
273 | else goto LOOP_header_bb | |
274 | new_merge_bb: | |
275 | goto update_bb | |
276 | update_bb: | |
277 | ||
278 | ==> The CFG after this function: | |
279 | guard_bb: | |
280 | if (LOOP_guard_condition) goto new_merge_bb | |
281 | else goto LOOP_header_bb | |
282 | LOOP_header_bb: | |
283 | loop_body | |
284 | if (exit_loop_condition) goto new_exit_bb | |
285 | else goto LOOP_header_bb | |
286 | new_exit_bb: | |
287 | new_merge_bb: | |
288 | goto update_bb | |
289 | update_bb: | |
290 | ||
291 | This function: | |
292 | 1. creates and updates the relevant phi nodes to account for the new | |
293 | incoming edge (GUARD_EDGE) into NEW_MERGE_BB. This involves: | |
294 | 1.1. Create phi nodes at NEW_MERGE_BB. | |
295 | 1.2. Update the phi nodes at the successor of NEW_MERGE_BB (denoted | |
296 | UPDATE_BB). UPDATE_BB was the exit-bb of LOOP before NEW_MERGE_BB | |
297 | 2. preserves loop-closed-ssa-form by creating the required phi nodes | |
298 | at the exit of LOOP (i.e, in NEW_EXIT_BB). | |
299 | ||
300 | There are two flavors to this function: | |
301 | ||
302 | slpeel_update_phi_nodes_for_guard1: | |
303 | Here the guard controls whether we enter or skip LOOP, where LOOP is a | |
304 | prolog_loop (loop1 below), and the new phis created in NEW_MERGE_BB are | |
305 | for variables that have phis in the loop header. | |
306 | ||
307 | slpeel_update_phi_nodes_for_guard2: | |
308 | Here the guard controls whether we enter or skip LOOP, where LOOP is an | |
309 | epilog_loop (loop2 below), and the new phis created in NEW_MERGE_BB are | |
310 | for variables that have phis in the loop exit. | |
311 | ||
312 | I.E., the overall structure is: | |
313 | ||
314 | loop1_preheader_bb: | |
315 | guard1 (goto loop1/merge1_bb) | |
316 | loop1 | |
317 | loop1_exit_bb: | |
318 | guard2 (goto merge1_bb/merge2_bb) | |
319 | merge1_bb | |
320 | loop2 | |
321 | loop2_exit_bb | |
322 | merge2_bb | |
323 | next_bb | |
324 | ||
325 | slpeel_update_phi_nodes_for_guard1 takes care of creating phis in | |
326 | loop1_exit_bb and merge1_bb. These are entry phis (phis for the vars | |
327 | that have phis in loop1->header). | |
328 | ||
329 | slpeel_update_phi_nodes_for_guard2 takes care of creating phis in | |
330 | loop2_exit_bb and merge2_bb. These are exit phis (phis for the vars | |
331 | that have phis in next_bb). It also adds some of these phis to | |
332 | loop1_exit_bb. | |
333 | ||
334 | slpeel_update_phi_nodes_for_guard1 is always called before | |
335 | slpeel_update_phi_nodes_for_guard2. They are both needed in order | |
336 | to create correct data-flow and loop-closed-ssa-form. | |
337 | ||
338 | Generally slpeel_update_phi_nodes_for_guard1 creates phis for variables | |
339 | that change between iterations of a loop (and therefore have a phi-node | |
340 | at the loop entry), whereas slpeel_update_phi_nodes_for_guard2 creates | |
b8698a0f L |
341 | phis for variables that are used out of the loop (and therefore have |
342 | loop-closed exit phis). Some variables may be both updated between | |
ebfd146a IR |
343 | iterations and used after the loop. This is why in loop1_exit_bb we |
344 | may need both entry_phis (created by slpeel_update_phi_nodes_for_guard1) | |
345 | and exit phis (created by slpeel_update_phi_nodes_for_guard2). | |
346 | ||
347 | - IS_NEW_LOOP: if IS_NEW_LOOP is true, then LOOP is a newly created copy of | |
348 | an original loop. i.e., we have: | |
349 | ||
350 | orig_loop | |
351 | guard_bb (goto LOOP/new_merge) | |
352 | new_loop <-- LOOP | |
353 | new_exit | |
354 | new_merge | |
355 | next_bb | |
356 | ||
357 | If IS_NEW_LOOP is false, then LOOP is an original loop, in which case we | |
358 | have: | |
359 | ||
360 | new_loop | |
361 | guard_bb (goto LOOP/new_merge) | |
362 | orig_loop <-- LOOP | |
363 | new_exit | |
364 | new_merge | |
365 | next_bb | |
366 | ||
367 | The SSA names defined in the original loop have a current | |
026c3cfd AH |
368 | reaching definition that records the corresponding new ssa-name |
369 | used in the new duplicated loop copy. | |
ebfd146a IR |
370 | */ |
371 | ||
372 | /* Function slpeel_update_phi_nodes_for_guard1 | |
b8698a0f | 373 | |
ebfd146a IR |
374 | Input: |
375 | - GUARD_EDGE, LOOP, IS_NEW_LOOP, NEW_EXIT_BB - as explained above. | |
376 | - DEFS - a bitmap of ssa names to mark new names for which we recorded | |
b8698a0f L |
377 | information. |
378 | ||
ebfd146a IR |
379 | In the context of the overall structure, we have: |
380 | ||
b8698a0f | 381 | loop1_preheader_bb: |
ebfd146a IR |
382 | guard1 (goto loop1/merge1_bb) |
383 | LOOP-> loop1 | |
384 | loop1_exit_bb: | |
385 | guard2 (goto merge1_bb/merge2_bb) | |
386 | merge1_bb | |
387 | loop2 | |
388 | loop2_exit_bb | |
389 | merge2_bb | |
390 | next_bb | |
391 | ||
392 | For each name updated between loop iterations (i.e - for each name that has | |
393 | an entry (loop-header) phi in LOOP) we create a new phi in: | |
394 | 1. merge1_bb (to account for the edge from guard1) | |
395 | 2. loop1_exit_bb (an exit-phi to keep LOOP in loop-closed form) | |
396 | */ | |
397 | ||
398 | static void | |
399 | slpeel_update_phi_nodes_for_guard1 (edge guard_edge, struct loop *loop, | |
c334023f | 400 | bool is_new_loop, basic_block *new_exit_bb) |
ebfd146a | 401 | { |
538dd0b7 DM |
402 | gphi *orig_phi, *new_phi; |
403 | gphi *update_phi, *update_phi2; | |
ebfd146a IR |
404 | tree guard_arg, loop_arg; |
405 | basic_block new_merge_bb = guard_edge->dest; | |
406 | edge e = EDGE_SUCC (new_merge_bb, 0); | |
407 | basic_block update_bb = e->dest; | |
408 | basic_block orig_bb = loop->header; | |
409 | edge new_exit_e; | |
410 | tree current_new_name; | |
538dd0b7 | 411 | gphi_iterator gsi_orig, gsi_update; |
ebfd146a IR |
412 | |
413 | /* Create new bb between loop and new_merge_bb. */ | |
414 | *new_exit_bb = split_edge (single_exit (loop)); | |
415 | ||
416 | new_exit_e = EDGE_SUCC (*new_exit_bb, 0); | |
417 | ||
418 | for (gsi_orig = gsi_start_phis (orig_bb), | |
419 | gsi_update = gsi_start_phis (update_bb); | |
420 | !gsi_end_p (gsi_orig) && !gsi_end_p (gsi_update); | |
421 | gsi_next (&gsi_orig), gsi_next (&gsi_update)) | |
422 | { | |
e20f6b4b | 423 | source_location loop_locus, guard_locus; |
070ecdfd | 424 | tree new_res; |
538dd0b7 DM |
425 | orig_phi = gsi_orig.phi (); |
426 | update_phi = gsi_update.phi (); | |
ebfd146a | 427 | |
ebfd146a IR |
428 | /** 1. Handle new-merge-point phis **/ |
429 | ||
430 | /* 1.1. Generate new phi node in NEW_MERGE_BB: */ | |
b731b390 | 431 | new_res = copy_ssa_name (PHI_RESULT (orig_phi)); |
070ecdfd | 432 | new_phi = create_phi_node (new_res, new_merge_bb); |
ebfd146a IR |
433 | |
434 | /* 1.2. NEW_MERGE_BB has two incoming edges: GUARD_EDGE and the exit-edge | |
435 | of LOOP. Set the two phi args in NEW_PHI for these edges: */ | |
436 | loop_arg = PHI_ARG_DEF_FROM_EDGE (orig_phi, EDGE_SUCC (loop->latch, 0)); | |
b8698a0f L |
437 | loop_locus = gimple_phi_arg_location_from_edge (orig_phi, |
438 | EDGE_SUCC (loop->latch, | |
f5045c96 | 439 | 0)); |
ebfd146a | 440 | guard_arg = PHI_ARG_DEF_FROM_EDGE (orig_phi, loop_preheader_edge (loop)); |
b8698a0f L |
441 | guard_locus |
442 | = gimple_phi_arg_location_from_edge (orig_phi, | |
f5045c96 | 443 | loop_preheader_edge (loop)); |
ebfd146a | 444 | |
9e227d60 DC |
445 | add_phi_arg (new_phi, loop_arg, new_exit_e, loop_locus); |
446 | add_phi_arg (new_phi, guard_arg, guard_edge, guard_locus); | |
ebfd146a IR |
447 | |
448 | /* 1.3. Update phi in successor block. */ | |
449 | gcc_assert (PHI_ARG_DEF_FROM_EDGE (update_phi, e) == loop_arg | |
450 | || PHI_ARG_DEF_FROM_EDGE (update_phi, e) == guard_arg); | |
684f25f4 | 451 | adjust_phi_and_debug_stmts (update_phi, e, PHI_RESULT (new_phi)); |
ebfd146a IR |
452 | update_phi2 = new_phi; |
453 | ||
454 | ||
455 | /** 2. Handle loop-closed-ssa-form phis **/ | |
456 | ||
ea057359 | 457 | if (virtual_operand_p (PHI_RESULT (orig_phi))) |
ebfd146a IR |
458 | continue; |
459 | ||
460 | /* 2.1. Generate new phi node in NEW_EXIT_BB: */ | |
b731b390 | 461 | new_res = copy_ssa_name (PHI_RESULT (orig_phi)); |
070ecdfd | 462 | new_phi = create_phi_node (new_res, *new_exit_bb); |
ebfd146a IR |
463 | |
464 | /* 2.2. NEW_EXIT_BB has one incoming edge: the exit-edge of the loop. */ | |
9e227d60 | 465 | add_phi_arg (new_phi, loop_arg, single_exit (loop), loop_locus); |
ebfd146a IR |
466 | |
467 | /* 2.3. Update phi in successor of NEW_EXIT_BB: */ | |
468 | gcc_assert (PHI_ARG_DEF_FROM_EDGE (update_phi2, new_exit_e) == loop_arg); | |
684f25f4 AO |
469 | adjust_phi_and_debug_stmts (update_phi2, new_exit_e, |
470 | PHI_RESULT (new_phi)); | |
ebfd146a IR |
471 | |
472 | /* 2.4. Record the newly created name with set_current_def. | |
473 | We want to find a name such that | |
474 | name = get_current_def (orig_loop_name) | |
475 | and to set its current definition as follows: | |
476 | set_current_def (name, new_phi_name) | |
477 | ||
478 | If LOOP is a new loop then loop_arg is already the name we're | |
479 | looking for. If LOOP is the original loop, then loop_arg is | |
480 | the orig_loop_name and the relevant name is recorded in its | |
481 | current reaching definition. */ | |
482 | if (is_new_loop) | |
483 | current_new_name = loop_arg; | |
484 | else | |
485 | { | |
486 | current_new_name = get_current_def (loop_arg); | |
487 | /* current_def is not available only if the variable does not | |
488 | change inside the loop, in which case we also don't care | |
489 | about recording a current_def for it because we won't be | |
490 | trying to create loop-exit-phis for it. */ | |
491 | if (!current_new_name) | |
492 | continue; | |
493 | } | |
39719c84 JJ |
494 | tree new_name = get_current_def (current_new_name); |
495 | /* Because of peeled_chrec optimization it is possible that we have | |
496 | set this earlier. Verify the PHI has the same value. */ | |
497 | if (new_name) | |
498 | { | |
355fe088 | 499 | gimple *phi = SSA_NAME_DEF_STMT (new_name); |
39719c84 JJ |
500 | gcc_assert (gimple_code (phi) == GIMPLE_PHI |
501 | && gimple_bb (phi) == *new_exit_bb | |
502 | && (PHI_ARG_DEF_FROM_EDGE (phi, single_exit (loop)) | |
503 | == loop_arg)); | |
504 | continue; | |
505 | } | |
ebfd146a IR |
506 | |
507 | set_current_def (current_new_name, PHI_RESULT (new_phi)); | |
ebfd146a IR |
508 | } |
509 | } | |
510 | ||
511 | ||
512 | /* Function slpeel_update_phi_nodes_for_guard2 | |
513 | ||
514 | Input: | |
515 | - GUARD_EDGE, LOOP, IS_NEW_LOOP, NEW_EXIT_BB - as explained above. | |
516 | ||
517 | In the context of the overall structure, we have: | |
518 | ||
b8698a0f | 519 | loop1_preheader_bb: |
ebfd146a IR |
520 | guard1 (goto loop1/merge1_bb) |
521 | loop1 | |
b8698a0f | 522 | loop1_exit_bb: |
ebfd146a IR |
523 | guard2 (goto merge1_bb/merge2_bb) |
524 | merge1_bb | |
525 | LOOP-> loop2 | |
526 | loop2_exit_bb | |
527 | merge2_bb | |
528 | next_bb | |
529 | ||
530 | For each name used out side the loop (i.e - for each name that has an exit | |
531 | phi in next_bb) we create a new phi in: | |
b8698a0f | 532 | 1. merge2_bb (to account for the edge from guard_bb) |
ebfd146a IR |
533 | 2. loop2_exit_bb (an exit-phi to keep LOOP in loop-closed form) |
534 | 3. guard2 bb (an exit phi to keep the preceding loop in loop-closed form), | |
535 | if needed (if it wasn't handled by slpeel_update_phis_nodes_for_phi1). | |
536 | */ | |
537 | ||
538 | static void | |
539 | slpeel_update_phi_nodes_for_guard2 (edge guard_edge, struct loop *loop, | |
540 | bool is_new_loop, basic_block *new_exit_bb) | |
541 | { | |
538dd0b7 DM |
542 | gphi *orig_phi, *new_phi; |
543 | gphi *update_phi, *update_phi2; | |
ebfd146a IR |
544 | tree guard_arg, loop_arg; |
545 | basic_block new_merge_bb = guard_edge->dest; | |
546 | edge e = EDGE_SUCC (new_merge_bb, 0); | |
547 | basic_block update_bb = e->dest; | |
548 | edge new_exit_e; | |
549 | tree orig_def, orig_def_new_name; | |
550 | tree new_name, new_name2; | |
551 | tree arg; | |
538dd0b7 | 552 | gphi_iterator gsi; |
ebfd146a IR |
553 | |
554 | /* Create new bb between loop and new_merge_bb. */ | |
555 | *new_exit_bb = split_edge (single_exit (loop)); | |
556 | ||
557 | new_exit_e = EDGE_SUCC (*new_exit_bb, 0); | |
558 | ||
559 | for (gsi = gsi_start_phis (update_bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
560 | { | |
070ecdfd | 561 | tree new_res; |
538dd0b7 | 562 | update_phi = gsi.phi (); |
ebfd146a IR |
563 | orig_phi = update_phi; |
564 | orig_def = PHI_ARG_DEF_FROM_EDGE (orig_phi, e); | |
565 | /* This loop-closed-phi actually doesn't represent a use | |
b8698a0f | 566 | out of the loop - the phi arg is a constant. */ |
ebfd146a IR |
567 | if (TREE_CODE (orig_def) != SSA_NAME) |
568 | continue; | |
569 | orig_def_new_name = get_current_def (orig_def); | |
570 | arg = NULL_TREE; | |
571 | ||
572 | /** 1. Handle new-merge-point phis **/ | |
573 | ||
574 | /* 1.1. Generate new phi node in NEW_MERGE_BB: */ | |
b731b390 | 575 | new_res = copy_ssa_name (PHI_RESULT (orig_phi)); |
070ecdfd | 576 | new_phi = create_phi_node (new_res, new_merge_bb); |
ebfd146a IR |
577 | |
578 | /* 1.2. NEW_MERGE_BB has two incoming edges: GUARD_EDGE and the exit-edge | |
579 | of LOOP. Set the two PHI args in NEW_PHI for these edges: */ | |
580 | new_name = orig_def; | |
581 | new_name2 = NULL_TREE; | |
582 | if (orig_def_new_name) | |
583 | { | |
584 | new_name = orig_def_new_name; | |
585 | /* Some variables have both loop-entry-phis and loop-exit-phis. | |
586 | Such variables were given yet newer names by phis placed in | |
587 | guard_bb by slpeel_update_phi_nodes_for_guard1. I.e: | |
588 | new_name2 = get_current_def (get_current_def (orig_name)). */ | |
589 | new_name2 = get_current_def (new_name); | |
590 | } | |
b8698a0f | 591 | |
ebfd146a IR |
592 | if (is_new_loop) |
593 | { | |
594 | guard_arg = orig_def; | |
595 | loop_arg = new_name; | |
596 | } | |
597 | else | |
598 | { | |
599 | guard_arg = new_name; | |
600 | loop_arg = orig_def; | |
601 | } | |
602 | if (new_name2) | |
603 | guard_arg = new_name2; | |
b8698a0f | 604 | |
9e227d60 DC |
605 | add_phi_arg (new_phi, loop_arg, new_exit_e, UNKNOWN_LOCATION); |
606 | add_phi_arg (new_phi, guard_arg, guard_edge, UNKNOWN_LOCATION); | |
ebfd146a IR |
607 | |
608 | /* 1.3. Update phi in successor block. */ | |
609 | gcc_assert (PHI_ARG_DEF_FROM_EDGE (update_phi, e) == orig_def); | |
684f25f4 | 610 | adjust_phi_and_debug_stmts (update_phi, e, PHI_RESULT (new_phi)); |
ebfd146a IR |
611 | update_phi2 = new_phi; |
612 | ||
613 | ||
614 | /** 2. Handle loop-closed-ssa-form phis **/ | |
615 | ||
616 | /* 2.1. Generate new phi node in NEW_EXIT_BB: */ | |
b731b390 | 617 | new_res = copy_ssa_name (PHI_RESULT (orig_phi)); |
070ecdfd | 618 | new_phi = create_phi_node (new_res, *new_exit_bb); |
ebfd146a IR |
619 | |
620 | /* 2.2. NEW_EXIT_BB has one incoming edge: the exit-edge of the loop. */ | |
9e227d60 | 621 | add_phi_arg (new_phi, loop_arg, single_exit (loop), UNKNOWN_LOCATION); |
ebfd146a IR |
622 | |
623 | /* 2.3. Update phi in successor of NEW_EXIT_BB: */ | |
624 | gcc_assert (PHI_ARG_DEF_FROM_EDGE (update_phi2, new_exit_e) == loop_arg); | |
684f25f4 AO |
625 | adjust_phi_and_debug_stmts (update_phi2, new_exit_e, |
626 | PHI_RESULT (new_phi)); | |
ebfd146a IR |
627 | |
628 | ||
629 | /** 3. Handle loop-closed-ssa-form phis for first loop **/ | |
630 | ||
631 | /* 3.1. Find the relevant names that need an exit-phi in | |
632 | GUARD_BB, i.e. names for which | |
633 | slpeel_update_phi_nodes_for_guard1 had not already created a | |
634 | phi node. This is the case for names that are used outside | |
635 | the loop (and therefore need an exit phi) but are not updated | |
636 | across loop iterations (and therefore don't have a | |
637 | loop-header-phi). | |
638 | ||
639 | slpeel_update_phi_nodes_for_guard1 is responsible for | |
640 | creating loop-exit phis in GUARD_BB for names that have a | |
641 | loop-header-phi. When such a phi is created we also record | |
642 | the new name in its current definition. If this new name | |
643 | exists, then guard_arg was set to this new name (see 1.2 | |
644 | above). Therefore, if guard_arg is not this new name, this | |
645 | is an indication that an exit-phi in GUARD_BB was not yet | |
646 | created, so we take care of it here. */ | |
647 | if (guard_arg == new_name2) | |
648 | continue; | |
649 | arg = guard_arg; | |
650 | ||
651 | /* 3.2. Generate new phi node in GUARD_BB: */ | |
b731b390 | 652 | new_res = copy_ssa_name (PHI_RESULT (orig_phi)); |
070ecdfd | 653 | new_phi = create_phi_node (new_res, guard_edge->src); |
ebfd146a IR |
654 | |
655 | /* 3.3. GUARD_BB has one incoming edge: */ | |
656 | gcc_assert (EDGE_COUNT (guard_edge->src->preds) == 1); | |
b8698a0f | 657 | add_phi_arg (new_phi, arg, EDGE_PRED (guard_edge->src, 0), |
9e227d60 | 658 | UNKNOWN_LOCATION); |
ebfd146a IR |
659 | |
660 | /* 3.4. Update phi in successor of GUARD_BB: */ | |
661 | gcc_assert (PHI_ARG_DEF_FROM_EDGE (update_phi2, guard_edge) | |
662 | == guard_arg); | |
684f25f4 AO |
663 | adjust_phi_and_debug_stmts (update_phi2, guard_edge, |
664 | PHI_RESULT (new_phi)); | |
ebfd146a IR |
665 | } |
666 | } | |
667 | ||
668 | ||
669 | /* Make the LOOP iterate NITERS times. This is done by adding a new IV | |
670 | that starts at zero, increases by one and its limit is NITERS. | |
671 | ||
672 | Assumption: the exit-condition of LOOP is the last stmt in the loop. */ | |
673 | ||
674 | void | |
675 | slpeel_make_loop_iterate_ntimes (struct loop *loop, tree niters) | |
676 | { | |
677 | tree indx_before_incr, indx_after_incr; | |
538dd0b7 DM |
678 | gcond *cond_stmt; |
679 | gcond *orig_cond; | |
ebfd146a IR |
680 | edge exit_edge = single_exit (loop); |
681 | gimple_stmt_iterator loop_cond_gsi; | |
682 | gimple_stmt_iterator incr_gsi; | |
683 | bool insert_after; | |
684 | tree init = build_int_cst (TREE_TYPE (niters), 0); | |
685 | tree step = build_int_cst (TREE_TYPE (niters), 1); | |
b05e0233 | 686 | source_location loop_loc; |
ebfd146a IR |
687 | enum tree_code code; |
688 | ||
689 | orig_cond = get_loop_exit_condition (loop); | |
690 | gcc_assert (orig_cond); | |
691 | loop_cond_gsi = gsi_for_stmt (orig_cond); | |
692 | ||
693 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
694 | create_iv (init, step, NULL_TREE, loop, | |
695 | &incr_gsi, insert_after, &indx_before_incr, &indx_after_incr); | |
696 | ||
697 | indx_after_incr = force_gimple_operand_gsi (&loop_cond_gsi, indx_after_incr, | |
698 | true, NULL_TREE, true, | |
699 | GSI_SAME_STMT); | |
700 | niters = force_gimple_operand_gsi (&loop_cond_gsi, niters, true, NULL_TREE, | |
701 | true, GSI_SAME_STMT); | |
702 | ||
703 | code = (exit_edge->flags & EDGE_TRUE_VALUE) ? GE_EXPR : LT_EXPR; | |
704 | cond_stmt = gimple_build_cond (code, indx_after_incr, niters, NULL_TREE, | |
705 | NULL_TREE); | |
706 | ||
707 | gsi_insert_before (&loop_cond_gsi, cond_stmt, GSI_SAME_STMT); | |
708 | ||
709 | /* Remove old loop exit test: */ | |
710 | gsi_remove (&loop_cond_gsi, true); | |
6f723d33 | 711 | free_stmt_vec_info (orig_cond); |
ebfd146a IR |
712 | |
713 | loop_loc = find_loop_location (loop); | |
73fbfcad | 714 | if (dump_enabled_p ()) |
ebfd146a | 715 | { |
b05e0233 RB |
716 | if (LOCATION_LOCUS (loop_loc) != UNKNOWN_LOCATION) |
717 | dump_printf (MSG_NOTE, "\nloop at %s:%d: ", LOCATION_FILE (loop_loc), | |
718 | LOCATION_LINE (loop_loc)); | |
78c60e3d | 719 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, cond_stmt, 0); |
e645e942 | 720 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a | 721 | } |
ebfd146a IR |
722 | loop->nb_iterations = niters; |
723 | } | |
724 | ||
5ce9450f JJ |
725 | /* Helper routine of slpeel_tree_duplicate_loop_to_edge_cfg. |
726 | For all PHI arguments in FROM->dest and TO->dest from those | |
727 | edges ensure that TO->dest PHI arguments have current_def | |
728 | to that in from. */ | |
729 | ||
730 | static void | |
731 | slpeel_duplicate_current_defs_from_edges (edge from, edge to) | |
732 | { | |
733 | gimple_stmt_iterator gsi_from, gsi_to; | |
734 | ||
735 | for (gsi_from = gsi_start_phis (from->dest), | |
736 | gsi_to = gsi_start_phis (to->dest); | |
737 | !gsi_end_p (gsi_from) && !gsi_end_p (gsi_to); | |
738 | gsi_next (&gsi_from), gsi_next (&gsi_to)) | |
739 | { | |
355fe088 TS |
740 | gimple *from_phi = gsi_stmt (gsi_from); |
741 | gimple *to_phi = gsi_stmt (gsi_to); | |
5ce9450f JJ |
742 | tree from_arg = PHI_ARG_DEF_FROM_EDGE (from_phi, from); |
743 | tree to_arg = PHI_ARG_DEF_FROM_EDGE (to_phi, to); | |
744 | if (TREE_CODE (from_arg) == SSA_NAME | |
745 | && TREE_CODE (to_arg) == SSA_NAME | |
746 | && get_current_def (to_arg) == NULL_TREE) | |
747 | set_current_def (to_arg, get_current_def (from_arg)); | |
748 | } | |
749 | } | |
750 | ||
ebfd146a | 751 | |
b8698a0f | 752 | /* Given LOOP this function generates a new copy of it and puts it |
5ce9450f JJ |
753 | on E which is either the entry or exit of LOOP. If SCALAR_LOOP is |
754 | non-NULL, assume LOOP and SCALAR_LOOP are equivalent and copy the | |
755 | basic blocks from SCALAR_LOOP instead of LOOP, but to either the | |
756 | entry or exit of LOOP. */ | |
ebfd146a IR |
757 | |
758 | struct loop * | |
5ce9450f JJ |
759 | slpeel_tree_duplicate_loop_to_edge_cfg (struct loop *loop, |
760 | struct loop *scalar_loop, edge e) | |
ebfd146a IR |
761 | { |
762 | struct loop *new_loop; | |
763 | basic_block *new_bbs, *bbs; | |
764 | bool at_exit; | |
765 | bool was_imm_dom; | |
b8698a0f | 766 | basic_block exit_dest; |
ebfd146a | 767 | edge exit, new_exit; |
a6c51a12 | 768 | bool duplicate_outer_loop = false; |
ebfd146a | 769 | |
2cfc56b9 RB |
770 | exit = single_exit (loop); |
771 | at_exit = (e == exit); | |
ebfd146a IR |
772 | if (!at_exit && e != loop_preheader_edge (loop)) |
773 | return NULL; | |
774 | ||
5ce9450f JJ |
775 | if (scalar_loop == NULL) |
776 | scalar_loop = loop; | |
777 | ||
778 | bbs = XNEWVEC (basic_block, scalar_loop->num_nodes + 1); | |
779 | get_loop_body_with_size (scalar_loop, bbs, scalar_loop->num_nodes); | |
a6c51a12 YR |
780 | /* Allow duplication of outer loops. */ |
781 | if (scalar_loop->inner) | |
782 | duplicate_outer_loop = true; | |
ebfd146a | 783 | /* Check whether duplication is possible. */ |
5ce9450f | 784 | if (!can_copy_bbs_p (bbs, scalar_loop->num_nodes)) |
ebfd146a IR |
785 | { |
786 | free (bbs); | |
787 | return NULL; | |
788 | } | |
789 | ||
790 | /* Generate new loop structure. */ | |
5ce9450f JJ |
791 | new_loop = duplicate_loop (scalar_loop, loop_outer (scalar_loop)); |
792 | duplicate_subloops (scalar_loop, new_loop); | |
ebfd146a | 793 | |
2cfc56b9 | 794 | exit_dest = exit->dest; |
b8698a0f L |
795 | was_imm_dom = (get_immediate_dominator (CDI_DOMINATORS, |
796 | exit_dest) == loop->header ? | |
ebfd146a IR |
797 | true : false); |
798 | ||
2cfc56b9 RB |
799 | /* Also copy the pre-header, this avoids jumping through hoops to |
800 | duplicate the loop entry PHI arguments. Create an empty | |
801 | pre-header unconditionally for this. */ | |
5ce9450f | 802 | basic_block preheader = split_edge (loop_preheader_edge (scalar_loop)); |
2cfc56b9 | 803 | edge entry_e = single_pred_edge (preheader); |
5ce9450f JJ |
804 | bbs[scalar_loop->num_nodes] = preheader; |
805 | new_bbs = XNEWVEC (basic_block, scalar_loop->num_nodes + 1); | |
ebfd146a | 806 | |
5ce9450f JJ |
807 | exit = single_exit (scalar_loop); |
808 | copy_bbs (bbs, scalar_loop->num_nodes + 1, new_bbs, | |
ebfd146a | 809 | &exit, 1, &new_exit, NULL, |
f14540b6 | 810 | e->src, true); |
5ce9450f JJ |
811 | exit = single_exit (loop); |
812 | basic_block new_preheader = new_bbs[scalar_loop->num_nodes]; | |
ebfd146a | 813 | |
5ce9450f JJ |
814 | add_phi_args_after_copy (new_bbs, scalar_loop->num_nodes + 1, NULL); |
815 | ||
816 | if (scalar_loop != loop) | |
817 | { | |
818 | /* If we copied from SCALAR_LOOP rather than LOOP, SSA_NAMEs from | |
819 | SCALAR_LOOP will have current_def set to SSA_NAMEs in the new_loop, | |
820 | but LOOP will not. slpeel_update_phi_nodes_for_guard{1,2} expects | |
821 | the LOOP SSA_NAMEs (on the exit edge and edge from latch to | |
822 | header) to have current_def set, so copy them over. */ | |
823 | slpeel_duplicate_current_defs_from_edges (single_exit (scalar_loop), | |
824 | exit); | |
825 | slpeel_duplicate_current_defs_from_edges (EDGE_SUCC (scalar_loop->latch, | |
826 | 0), | |
827 | EDGE_SUCC (loop->latch, 0)); | |
828 | } | |
b8698a0f | 829 | |
ebfd146a IR |
830 | if (at_exit) /* Add the loop copy at exit. */ |
831 | { | |
5ce9450f JJ |
832 | if (scalar_loop != loop) |
833 | { | |
538dd0b7 | 834 | gphi_iterator gsi; |
5ce9450f JJ |
835 | new_exit = redirect_edge_and_branch (new_exit, exit_dest); |
836 | ||
837 | for (gsi = gsi_start_phis (exit_dest); !gsi_end_p (gsi); | |
838 | gsi_next (&gsi)) | |
839 | { | |
538dd0b7 | 840 | gphi *phi = gsi.phi (); |
5ce9450f JJ |
841 | tree orig_arg = PHI_ARG_DEF_FROM_EDGE (phi, e); |
842 | location_t orig_locus | |
843 | = gimple_phi_arg_location_from_edge (phi, e); | |
844 | ||
845 | add_phi_arg (phi, orig_arg, new_exit, orig_locus); | |
846 | } | |
847 | } | |
2cfc56b9 RB |
848 | redirect_edge_and_branch_force (e, new_preheader); |
849 | flush_pending_stmts (e); | |
850 | set_immediate_dominator (CDI_DOMINATORS, new_preheader, e->src); | |
a6c51a12 | 851 | if (was_imm_dom || duplicate_outer_loop) |
5ce9450f | 852 | set_immediate_dominator (CDI_DOMINATORS, exit_dest, new_exit->src); |
2cfc56b9 RB |
853 | |
854 | /* And remove the non-necessary forwarder again. Keep the other | |
855 | one so we have a proper pre-header for the loop at the exit edge. */ | |
5ce9450f JJ |
856 | redirect_edge_pred (single_succ_edge (preheader), |
857 | single_pred (preheader)); | |
2cfc56b9 | 858 | delete_basic_block (preheader); |
5ce9450f JJ |
859 | set_immediate_dominator (CDI_DOMINATORS, scalar_loop->header, |
860 | loop_preheader_edge (scalar_loop)->src); | |
ebfd146a IR |
861 | } |
862 | else /* Add the copy at entry. */ | |
863 | { | |
5ce9450f JJ |
864 | if (scalar_loop != loop) |
865 | { | |
866 | /* Remove the non-necessary forwarder of scalar_loop again. */ | |
867 | redirect_edge_pred (single_succ_edge (preheader), | |
868 | single_pred (preheader)); | |
869 | delete_basic_block (preheader); | |
870 | set_immediate_dominator (CDI_DOMINATORS, scalar_loop->header, | |
871 | loop_preheader_edge (scalar_loop)->src); | |
872 | preheader = split_edge (loop_preheader_edge (loop)); | |
873 | entry_e = single_pred_edge (preheader); | |
874 | } | |
875 | ||
2cfc56b9 RB |
876 | redirect_edge_and_branch_force (entry_e, new_preheader); |
877 | flush_pending_stmts (entry_e); | |
878 | set_immediate_dominator (CDI_DOMINATORS, new_preheader, entry_e->src); | |
879 | ||
880 | redirect_edge_and_branch_force (new_exit, preheader); | |
881 | flush_pending_stmts (new_exit); | |
882 | set_immediate_dominator (CDI_DOMINATORS, preheader, new_exit->src); | |
883 | ||
884 | /* And remove the non-necessary forwarder again. Keep the other | |
885 | one so we have a proper pre-header for the loop at the exit edge. */ | |
5ce9450f JJ |
886 | redirect_edge_pred (single_succ_edge (new_preheader), |
887 | single_pred (new_preheader)); | |
2cfc56b9 RB |
888 | delete_basic_block (new_preheader); |
889 | set_immediate_dominator (CDI_DOMINATORS, new_loop->header, | |
890 | loop_preheader_edge (new_loop)->src); | |
ebfd146a IR |
891 | } |
892 | ||
5ce9450f | 893 | for (unsigned i = 0; i < scalar_loop->num_nodes + 1; i++) |
a6c51a12 | 894 | rename_variables_in_bb (new_bbs[i], duplicate_outer_loop); |
2cfc56b9 | 895 | |
5ce9450f JJ |
896 | if (scalar_loop != loop) |
897 | { | |
898 | /* Update new_loop->header PHIs, so that on the preheader | |
899 | edge they are the ones from loop rather than scalar_loop. */ | |
538dd0b7 | 900 | gphi_iterator gsi_orig, gsi_new; |
5ce9450f JJ |
901 | edge orig_e = loop_preheader_edge (loop); |
902 | edge new_e = loop_preheader_edge (new_loop); | |
903 | ||
904 | for (gsi_orig = gsi_start_phis (loop->header), | |
905 | gsi_new = gsi_start_phis (new_loop->header); | |
906 | !gsi_end_p (gsi_orig) && !gsi_end_p (gsi_new); | |
907 | gsi_next (&gsi_orig), gsi_next (&gsi_new)) | |
908 | { | |
538dd0b7 DM |
909 | gphi *orig_phi = gsi_orig.phi (); |
910 | gphi *new_phi = gsi_new.phi (); | |
5ce9450f JJ |
911 | tree orig_arg = PHI_ARG_DEF_FROM_EDGE (orig_phi, orig_e); |
912 | location_t orig_locus | |
913 | = gimple_phi_arg_location_from_edge (orig_phi, orig_e); | |
914 | ||
915 | add_phi_arg (new_phi, orig_arg, new_e, orig_locus); | |
916 | } | |
917 | } | |
918 | ||
ebfd146a IR |
919 | free (new_bbs); |
920 | free (bbs); | |
921 | ||
b2b29377 | 922 | checking_verify_dominators (CDI_DOMINATORS); |
2cfc56b9 | 923 | |
ebfd146a IR |
924 | return new_loop; |
925 | } | |
926 | ||
927 | ||
928 | /* Given the condition statement COND, put it as the last statement | |
929 | of GUARD_BB; EXIT_BB is the basic block to skip the loop; | |
b8698a0f | 930 | Assumes that this is the single exit of the guarded loop. |
86290011 | 931 | Returns the skip edge, inserts new stmts on the COND_EXPR_STMT_LIST. */ |
ebfd146a IR |
932 | |
933 | static edge | |
86290011 RG |
934 | slpeel_add_loop_guard (basic_block guard_bb, tree cond, |
935 | gimple_seq cond_expr_stmt_list, | |
e78410bf JH |
936 | basic_block exit_bb, basic_block dom_bb, |
937 | int probability) | |
ebfd146a IR |
938 | { |
939 | gimple_stmt_iterator gsi; | |
940 | edge new_e, enter_e; | |
538dd0b7 | 941 | gcond *cond_stmt; |
ebfd146a IR |
942 | gimple_seq gimplify_stmt_list = NULL; |
943 | ||
944 | enter_e = EDGE_SUCC (guard_bb, 0); | |
945 | enter_e->flags &= ~EDGE_FALLTHRU; | |
946 | enter_e->flags |= EDGE_FALSE_VALUE; | |
947 | gsi = gsi_last_bb (guard_bb); | |
948 | ||
f7a06a98 RG |
949 | cond = force_gimple_operand_1 (cond, &gimplify_stmt_list, is_gimple_condexpr, |
950 | NULL_TREE); | |
86290011 RG |
951 | if (gimplify_stmt_list) |
952 | gimple_seq_add_seq (&cond_expr_stmt_list, gimplify_stmt_list); | |
f7a06a98 | 953 | cond_stmt = gimple_build_cond_from_tree (cond, NULL_TREE, NULL_TREE); |
86290011 RG |
954 | if (cond_expr_stmt_list) |
955 | gsi_insert_seq_after (&gsi, cond_expr_stmt_list, GSI_NEW_STMT); | |
ebfd146a IR |
956 | |
957 | gsi = gsi_last_bb (guard_bb); | |
958 | gsi_insert_after (&gsi, cond_stmt, GSI_NEW_STMT); | |
959 | ||
960 | /* Add new edge to connect guard block to the merge/loop-exit block. */ | |
961 | new_e = make_edge (guard_bb, exit_bb, EDGE_TRUE_VALUE); | |
e78410bf JH |
962 | |
963 | new_e->count = guard_bb->count; | |
964 | new_e->probability = probability; | |
965 | new_e->count = apply_probability (enter_e->count, probability); | |
966 | enter_e->count -= new_e->count; | |
967 | enter_e->probability = inverse_probability (probability); | |
ebfd146a IR |
968 | set_immediate_dominator (CDI_DOMINATORS, exit_bb, dom_bb); |
969 | return new_e; | |
970 | } | |
971 | ||
972 | ||
973 | /* This function verifies that the following restrictions apply to LOOP: | |
a6c51a12 YR |
974 | (1) it consists of exactly 2 basic blocks - header, and an empty latch |
975 | for innermost loop and 5 basic blocks for outer-loop. | |
976 | (2) it is single entry, single exit | |
977 | (3) its exit condition is the last stmt in the header | |
978 | (4) E is the entry/exit edge of LOOP. | |
ebfd146a IR |
979 | */ |
980 | ||
981 | bool | |
982 | slpeel_can_duplicate_loop_p (const struct loop *loop, const_edge e) | |
983 | { | |
984 | edge exit_e = single_exit (loop); | |
985 | edge entry_e = loop_preheader_edge (loop); | |
538dd0b7 | 986 | gcond *orig_cond = get_loop_exit_condition (loop); |
ebfd146a | 987 | gimple_stmt_iterator loop_exit_gsi = gsi_last_bb (exit_e->src); |
a6c51a12 | 988 | unsigned int num_bb = loop->inner? 5 : 2; |
ebfd146a | 989 | |
ebfd146a IR |
990 | /* All loops have an outer scope; the only case loop->outer is NULL is for |
991 | the function itself. */ | |
a6c51a12 YR |
992 | if (!loop_outer (loop) |
993 | || loop->num_nodes != num_bb | |
ebfd146a IR |
994 | || !empty_block_p (loop->latch) |
995 | || !single_exit (loop) | |
996 | /* Verify that new loop exit condition can be trivially modified. */ | |
997 | || (!orig_cond || orig_cond != gsi_stmt (loop_exit_gsi)) | |
998 | || (e != exit_e && e != entry_e)) | |
999 | return false; | |
1000 | ||
1001 | return true; | |
1002 | } | |
1003 | ||
ebfd146a | 1004 | static void |
b2b29377 MM |
1005 | slpeel_checking_verify_cfg_after_peeling (struct loop *first_loop, |
1006 | struct loop *second_loop) | |
ebfd146a | 1007 | { |
b2b29377 MM |
1008 | if (!flag_checking) |
1009 | return; | |
1010 | ||
ebfd146a IR |
1011 | basic_block loop1_exit_bb = single_exit (first_loop)->dest; |
1012 | basic_block loop2_entry_bb = loop_preheader_edge (second_loop)->src; | |
1013 | basic_block loop1_entry_bb = loop_preheader_edge (first_loop)->src; | |
1014 | ||
1015 | /* A guard that controls whether the second_loop is to be executed or skipped | |
1016 | is placed in first_loop->exit. first_loop->exit therefore has two | |
1017 | successors - one is the preheader of second_loop, and the other is a bb | |
1018 | after second_loop. | |
1019 | */ | |
1020 | gcc_assert (EDGE_COUNT (loop1_exit_bb->succs) == 2); | |
b8698a0f | 1021 | |
ebfd146a IR |
1022 | /* 1. Verify that one of the successors of first_loop->exit is the preheader |
1023 | of second_loop. */ | |
b8698a0f | 1024 | |
ebfd146a IR |
1025 | /* The preheader of new_loop is expected to have two predecessors: |
1026 | first_loop->exit and the block that precedes first_loop. */ | |
1027 | ||
b8698a0f | 1028 | gcc_assert (EDGE_COUNT (loop2_entry_bb->preds) == 2 |
ebfd146a IR |
1029 | && ((EDGE_PRED (loop2_entry_bb, 0)->src == loop1_exit_bb |
1030 | && EDGE_PRED (loop2_entry_bb, 1)->src == loop1_entry_bb) | |
1031 | || (EDGE_PRED (loop2_entry_bb, 1)->src == loop1_exit_bb | |
1032 | && EDGE_PRED (loop2_entry_bb, 0)->src == loop1_entry_bb))); | |
b8698a0f | 1033 | |
ebfd146a IR |
1034 | /* Verify that the other successor of first_loop->exit is after the |
1035 | second_loop. */ | |
1036 | /* TODO */ | |
1037 | } | |
ebfd146a IR |
1038 | |
1039 | /* If the run time cost model check determines that vectorization is | |
1040 | not profitable and hence scalar loop should be generated then set | |
1041 | FIRST_NITERS to prologue peeled iterations. This will allow all the | |
1042 | iterations to be executed in the prologue peeled scalar loop. */ | |
1043 | ||
1044 | static void | |
1045 | set_prologue_iterations (basic_block bb_before_first_loop, | |
5d2eb24b | 1046 | tree *first_niters, |
ebfd146a | 1047 | struct loop *loop, |
e78410bf JH |
1048 | unsigned int th, |
1049 | int probability) | |
ebfd146a IR |
1050 | { |
1051 | edge e; | |
1052 | basic_block cond_bb, then_bb; | |
1053 | tree var, prologue_after_cost_adjust_name; | |
1054 | gimple_stmt_iterator gsi; | |
538dd0b7 | 1055 | gphi *newphi; |
ebfd146a | 1056 | edge e_true, e_false, e_fallthru; |
538dd0b7 | 1057 | gcond *cond_stmt; |
f7a06a98 | 1058 | gimple_seq stmts = NULL; |
ebfd146a | 1059 | tree cost_pre_condition = NULL_TREE; |
b8698a0f | 1060 | tree scalar_loop_iters = |
ebfd146a IR |
1061 | unshare_expr (LOOP_VINFO_NITERS_UNCHANGED (loop_vec_info_for_loop (loop))); |
1062 | ||
1063 | e = single_pred_edge (bb_before_first_loop); | |
c3284718 | 1064 | cond_bb = split_edge (e); |
ebfd146a IR |
1065 | |
1066 | e = single_pred_edge (bb_before_first_loop); | |
c3284718 | 1067 | then_bb = split_edge (e); |
ebfd146a IR |
1068 | set_immediate_dominator (CDI_DOMINATORS, then_bb, cond_bb); |
1069 | ||
1070 | e_false = make_single_succ_edge (cond_bb, bb_before_first_loop, | |
1071 | EDGE_FALSE_VALUE); | |
1072 | set_immediate_dominator (CDI_DOMINATORS, bb_before_first_loop, cond_bb); | |
1073 | ||
1074 | e_true = EDGE_PRED (then_bb, 0); | |
1075 | e_true->flags &= ~EDGE_FALLTHRU; | |
1076 | e_true->flags |= EDGE_TRUE_VALUE; | |
1077 | ||
e78410bf JH |
1078 | e_true->probability = probability; |
1079 | e_false->probability = inverse_probability (probability); | |
1080 | e_true->count = apply_probability (cond_bb->count, probability); | |
1081 | e_false->count = cond_bb->count - e_true->count; | |
1082 | then_bb->frequency = EDGE_FREQUENCY (e_true); | |
1083 | then_bb->count = e_true->count; | |
1084 | ||
ebfd146a | 1085 | e_fallthru = EDGE_SUCC (then_bb, 0); |
e78410bf | 1086 | e_fallthru->count = then_bb->count; |
ebfd146a | 1087 | |
f7a06a98 | 1088 | gsi = gsi_last_bb (cond_bb); |
ebfd146a | 1089 | cost_pre_condition = |
b8698a0f | 1090 | fold_build2 (LE_EXPR, boolean_type_node, scalar_loop_iters, |
ebfd146a IR |
1091 | build_int_cst (TREE_TYPE (scalar_loop_iters), th)); |
1092 | cost_pre_condition = | |
f7a06a98 RG |
1093 | force_gimple_operand_gsi_1 (&gsi, cost_pre_condition, is_gimple_condexpr, |
1094 | NULL_TREE, false, GSI_CONTINUE_LINKING); | |
1095 | cond_stmt = gimple_build_cond_from_tree (cost_pre_condition, | |
1096 | NULL_TREE, NULL_TREE); | |
ebfd146a | 1097 | gsi_insert_after (&gsi, cond_stmt, GSI_NEW_STMT); |
b8698a0f | 1098 | |
ebfd146a IR |
1099 | var = create_tmp_var (TREE_TYPE (scalar_loop_iters), |
1100 | "prologue_after_cost_adjust"); | |
b8698a0f | 1101 | prologue_after_cost_adjust_name = |
ebfd146a IR |
1102 | force_gimple_operand (scalar_loop_iters, &stmts, false, var); |
1103 | ||
1104 | gsi = gsi_last_bb (then_bb); | |
1105 | if (stmts) | |
1106 | gsi_insert_seq_after (&gsi, stmts, GSI_NEW_STMT); | |
1107 | ||
1108 | newphi = create_phi_node (var, bb_before_first_loop); | |
b8698a0f | 1109 | add_phi_arg (newphi, prologue_after_cost_adjust_name, e_fallthru, |
9e227d60 DC |
1110 | UNKNOWN_LOCATION); |
1111 | add_phi_arg (newphi, *first_niters, e_false, UNKNOWN_LOCATION); | |
ebfd146a | 1112 | |
5d2eb24b | 1113 | *first_niters = PHI_RESULT (newphi); |
ebfd146a IR |
1114 | } |
1115 | ||
ebfd146a IR |
1116 | /* Function slpeel_tree_peel_loop_to_edge. |
1117 | ||
1118 | Peel the first (last) iterations of LOOP into a new prolog (epilog) loop | |
1119 | that is placed on the entry (exit) edge E of LOOP. After this transformation | |
1120 | we have two loops one after the other - first-loop iterates FIRST_NITERS | |
1121 | times, and second-loop iterates the remainder NITERS - FIRST_NITERS times. | |
b8698a0f | 1122 | If the cost model indicates that it is profitable to emit a scalar |
ebfd146a IR |
1123 | loop instead of the vector one, then the prolog (epilog) loop will iterate |
1124 | for the entire unchanged scalar iterations of the loop. | |
1125 | ||
1126 | Input: | |
1127 | - LOOP: the loop to be peeled. | |
5ce9450f JJ |
1128 | - SCALAR_LOOP: if non-NULL, the alternate loop from which basic blocks |
1129 | should be copied. | |
ebfd146a IR |
1130 | - E: the exit or entry edge of LOOP. |
1131 | If it is the entry edge, we peel the first iterations of LOOP. In this | |
1132 | case first-loop is LOOP, and second-loop is the newly created loop. | |
1133 | If it is the exit edge, we peel the last iterations of LOOP. In this | |
1134 | case, first-loop is the newly created loop, and second-loop is LOOP. | |
1135 | - NITERS: the number of iterations that LOOP iterates. | |
1136 | - FIRST_NITERS: the number of iterations that the first-loop should iterate. | |
1137 | - UPDATE_FIRST_LOOP_COUNT: specified whether this function is responsible | |
1138 | for updating the loop bound of the first-loop to FIRST_NITERS. If it | |
1139 | is false, the caller of this function may want to take care of this | |
1140 | (this can be useful if we don't want new stmts added to first-loop). | |
1141 | - TH: cost model profitability threshold of iterations for vectorization. | |
1142 | - CHECK_PROFITABILITY: specify whether cost model check has not occurred | |
1143 | during versioning and hence needs to occur during | |
b8698a0f | 1144 | prologue generation or whether cost model check |
ebfd146a IR |
1145 | has not occurred during prologue generation and hence |
1146 | needs to occur during epilogue generation. | |
e78410bf JH |
1147 | - BOUND1 is the upper bound on number of iterations of the first loop (if known) |
1148 | - BOUND2 is the upper bound on number of iterations of the second loop (if known) | |
b8698a0f | 1149 | |
ebfd146a IR |
1150 | |
1151 | Output: | |
1152 | The function returns a pointer to the new loop-copy, or NULL if it failed | |
1153 | to perform the transformation. | |
1154 | ||
1155 | The function generates two if-then-else guards: one before the first loop, | |
1156 | and the other before the second loop: | |
1157 | The first guard is: | |
1158 | if (FIRST_NITERS == 0) then skip the first loop, | |
1159 | and go directly to the second loop. | |
1160 | The second guard is: | |
1161 | if (FIRST_NITERS == NITERS) then skip the second loop. | |
1162 | ||
86290011 RG |
1163 | If the optional COND_EXPR and COND_EXPR_STMT_LIST arguments are given |
1164 | then the generated condition is combined with COND_EXPR and the | |
1165 | statements in COND_EXPR_STMT_LIST are emitted together with it. | |
1166 | ||
ebfd146a IR |
1167 | FORNOW only simple loops are supported (see slpeel_can_duplicate_loop_p). |
1168 | FORNOW the resulting code will not be in loop-closed-ssa form. | |
1169 | */ | |
1170 | ||
5ce9450f JJ |
1171 | static struct loop * |
1172 | slpeel_tree_peel_loop_to_edge (struct loop *loop, struct loop *scalar_loop, | |
5d2eb24b | 1173 | edge e, tree *first_niters, |
ebfd146a | 1174 | tree niters, bool update_first_loop_count, |
86290011 | 1175 | unsigned int th, bool check_profitability, |
e78410bf JH |
1176 | tree cond_expr, gimple_seq cond_expr_stmt_list, |
1177 | int bound1, int bound2) | |
ebfd146a IR |
1178 | { |
1179 | struct loop *new_loop = NULL, *first_loop, *second_loop; | |
1180 | edge skip_e; | |
1181 | tree pre_condition = NULL_TREE; | |
ebfd146a IR |
1182 | basic_block bb_before_second_loop, bb_after_second_loop; |
1183 | basic_block bb_before_first_loop; | |
1184 | basic_block bb_between_loops; | |
1185 | basic_block new_exit_bb; | |
538dd0b7 | 1186 | gphi_iterator gsi; |
ebfd146a | 1187 | edge exit_e = single_exit (loop); |
b05e0233 | 1188 | source_location loop_loc; |
e78410bf JH |
1189 | /* There are many aspects to how likely the first loop is going to be executed. |
1190 | Without histogram we can't really do good job. Simply set it to | |
1191 | 2/3, so the first loop is not reordered to the end of function and | |
1192 | the hot path through stays short. */ | |
1193 | int first_guard_probability = 2 * REG_BR_PROB_BASE / 3; | |
1194 | int second_guard_probability = 2 * REG_BR_PROB_BASE / 3; | |
1195 | int probability_of_second_loop; | |
b8698a0f | 1196 | |
ebfd146a IR |
1197 | if (!slpeel_can_duplicate_loop_p (loop, e)) |
1198 | return NULL; | |
b8698a0f | 1199 | |
141310ef RB |
1200 | /* We might have a queued need to update virtual SSA form. As we |
1201 | delete the update SSA machinery below after doing a regular | |
1202 | incremental SSA update during loop copying make sure we don't | |
1203 | lose that fact. | |
1204 | ??? Needing to update virtual SSA form by renaming is unfortunate | |
1205 | but not all of the vectorizer code inserting new loads / stores | |
1206 | properly assigns virtual operands to those statements. */ | |
1207 | update_ssa (TODO_update_ssa_only_virtuals); | |
1208 | ||
e20f6b4b JJ |
1209 | /* If the loop has a virtual PHI, but exit bb doesn't, create a virtual PHI |
1210 | in the exit bb and rename all the uses after the loop. This simplifies | |
1211 | the *guard[12] routines, which assume loop closed SSA form for all PHIs | |
1212 | (but normally loop closed SSA form doesn't require virtual PHIs to be | |
1213 | in the same form). Doing this early simplifies the checking what | |
1214 | uses should be renamed. */ | |
1215 | for (gsi = gsi_start_phis (loop->header); !gsi_end_p (gsi); gsi_next (&gsi)) | |
ea057359 | 1216 | if (virtual_operand_p (gimple_phi_result (gsi_stmt (gsi)))) |
e20f6b4b | 1217 | { |
538dd0b7 | 1218 | gphi *phi = gsi.phi (); |
e20f6b4b JJ |
1219 | for (gsi = gsi_start_phis (exit_e->dest); |
1220 | !gsi_end_p (gsi); gsi_next (&gsi)) | |
ea057359 | 1221 | if (virtual_operand_p (gimple_phi_result (gsi_stmt (gsi)))) |
e20f6b4b JJ |
1222 | break; |
1223 | if (gsi_end_p (gsi)) | |
1224 | { | |
b731b390 | 1225 | tree new_vop = copy_ssa_name (PHI_RESULT (phi)); |
538dd0b7 | 1226 | gphi *new_phi = create_phi_node (new_vop, exit_e->dest); |
e20f6b4b JJ |
1227 | tree vop = PHI_ARG_DEF_FROM_EDGE (phi, EDGE_SUCC (loop->latch, 0)); |
1228 | imm_use_iterator imm_iter; | |
355fe088 | 1229 | gimple *stmt; |
e20f6b4b JJ |
1230 | use_operand_p use_p; |
1231 | ||
9e227d60 | 1232 | add_phi_arg (new_phi, vop, exit_e, UNKNOWN_LOCATION); |
e20f6b4b JJ |
1233 | gimple_phi_set_result (new_phi, new_vop); |
1234 | FOR_EACH_IMM_USE_STMT (stmt, imm_iter, vop) | |
b5fd0440 YR |
1235 | if (stmt != new_phi |
1236 | && !flow_bb_inside_loop_p (loop, gimple_bb (stmt))) | |
e20f6b4b JJ |
1237 | FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter) |
1238 | SET_USE (use_p, new_vop); | |
1239 | } | |
1240 | break; | |
1241 | } | |
ebfd146a IR |
1242 | |
1243 | /* 1. Generate a copy of LOOP and put it on E (E is the entry/exit of LOOP). | |
1244 | Resulting CFG would be: | |
1245 | ||
1246 | first_loop: | |
1247 | do { | |
1248 | } while ... | |
1249 | ||
1250 | second_loop: | |
1251 | do { | |
1252 | } while ... | |
1253 | ||
1254 | orig_exit_bb: | |
1255 | */ | |
b8698a0f | 1256 | |
5ce9450f JJ |
1257 | if (!(new_loop = slpeel_tree_duplicate_loop_to_edge_cfg (loop, scalar_loop, |
1258 | e))) | |
ebfd146a IR |
1259 | { |
1260 | loop_loc = find_loop_location (loop); | |
78c60e3d SS |
1261 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, loop_loc, |
1262 | "tree_duplicate_loop_to_edge_cfg failed.\n"); | |
ebfd146a IR |
1263 | return NULL; |
1264 | } | |
b8698a0f | 1265 | |
684f25f4 AO |
1266 | if (MAY_HAVE_DEBUG_STMTS) |
1267 | { | |
9771b263 | 1268 | gcc_assert (!adjust_vec.exists ()); |
ff4c81cc | 1269 | adjust_vec.create (32); |
684f25f4 AO |
1270 | } |
1271 | ||
ebfd146a IR |
1272 | if (e == exit_e) |
1273 | { | |
1274 | /* NEW_LOOP was placed after LOOP. */ | |
1275 | first_loop = loop; | |
1276 | second_loop = new_loop; | |
1277 | } | |
1278 | else | |
1279 | { | |
1280 | /* NEW_LOOP was placed before LOOP. */ | |
1281 | first_loop = new_loop; | |
1282 | second_loop = loop; | |
1283 | } | |
1284 | ||
ebfd146a IR |
1285 | /* 2. Add the guard code in one of the following ways: |
1286 | ||
1287 | 2.a Add the guard that controls whether the first loop is executed. | |
1288 | This occurs when this function is invoked for prologue or epilogue | |
1289 | generation and when the cost model check can be done at compile time. | |
1290 | ||
1291 | Resulting CFG would be: | |
1292 | ||
1293 | bb_before_first_loop: | |
1294 | if (FIRST_NITERS == 0) GOTO bb_before_second_loop | |
1295 | GOTO first-loop | |
1296 | ||
1297 | first_loop: | |
1298 | do { | |
1299 | } while ... | |
1300 | ||
1301 | bb_before_second_loop: | |
1302 | ||
1303 | second_loop: | |
1304 | do { | |
1305 | } while ... | |
1306 | ||
1307 | orig_exit_bb: | |
1308 | ||
1309 | 2.b Add the cost model check that allows the prologue | |
1310 | to iterate for the entire unchanged scalar | |
1311 | iterations of the loop in the event that the cost | |
1312 | model indicates that the scalar loop is more | |
1313 | profitable than the vector one. This occurs when | |
1314 | this function is invoked for prologue generation | |
1315 | and the cost model check needs to be done at run | |
1316 | time. | |
1317 | ||
1318 | Resulting CFG after prologue peeling would be: | |
1319 | ||
1320 | if (scalar_loop_iterations <= th) | |
1321 | FIRST_NITERS = scalar_loop_iterations | |
1322 | ||
1323 | bb_before_first_loop: | |
1324 | if (FIRST_NITERS == 0) GOTO bb_before_second_loop | |
1325 | GOTO first-loop | |
1326 | ||
1327 | first_loop: | |
1328 | do { | |
1329 | } while ... | |
1330 | ||
1331 | bb_before_second_loop: | |
1332 | ||
1333 | second_loop: | |
1334 | do { | |
1335 | } while ... | |
1336 | ||
1337 | orig_exit_bb: | |
1338 | ||
1339 | 2.c Add the cost model check that allows the epilogue | |
1340 | to iterate for the entire unchanged scalar | |
1341 | iterations of the loop in the event that the cost | |
1342 | model indicates that the scalar loop is more | |
1343 | profitable than the vector one. This occurs when | |
1344 | this function is invoked for epilogue generation | |
1345 | and the cost model check needs to be done at run | |
86290011 RG |
1346 | time. This check is combined with any pre-existing |
1347 | check in COND_EXPR to avoid versioning. | |
ebfd146a IR |
1348 | |
1349 | Resulting CFG after prologue peeling would be: | |
1350 | ||
1351 | bb_before_first_loop: | |
1352 | if ((scalar_loop_iterations <= th) | |
1353 | || | |
1354 | FIRST_NITERS == 0) GOTO bb_before_second_loop | |
1355 | GOTO first-loop | |
1356 | ||
1357 | first_loop: | |
1358 | do { | |
1359 | } while ... | |
1360 | ||
1361 | bb_before_second_loop: | |
1362 | ||
1363 | second_loop: | |
1364 | do { | |
1365 | } while ... | |
1366 | ||
1367 | orig_exit_bb: | |
1368 | */ | |
1369 | ||
1370 | bb_before_first_loop = split_edge (loop_preheader_edge (first_loop)); | |
2cfc56b9 RB |
1371 | /* Loop copying insterted a forwarder block for us here. */ |
1372 | bb_before_second_loop = single_exit (first_loop)->dest; | |
ebfd146a | 1373 | |
e78410bf JH |
1374 | probability_of_second_loop = (inverse_probability (first_guard_probability) |
1375 | + combine_probabilities (second_guard_probability, | |
1376 | first_guard_probability)); | |
1377 | /* Theoretically preheader edge of first loop and exit edge should have | |
1378 | same frequencies. Loop exit probablities are however easy to get wrong. | |
1379 | It is safer to copy value from original loop entry. */ | |
1380 | bb_before_second_loop->frequency | |
8ddb5a29 TJ |
1381 | = combine_probabilities (bb_before_first_loop->frequency, |
1382 | probability_of_second_loop); | |
e78410bf JH |
1383 | bb_before_second_loop->count |
1384 | = apply_probability (bb_before_first_loop->count, | |
1385 | probability_of_second_loop); | |
1386 | single_succ_edge (bb_before_second_loop)->count | |
1387 | = bb_before_second_loop->count; | |
1388 | ||
ebfd146a IR |
1389 | /* Epilogue peeling. */ |
1390 | if (!update_first_loop_count) | |
1391 | { | |
95b3eff3 RB |
1392 | loop_vec_info loop_vinfo = loop_vec_info_for_loop (loop); |
1393 | tree scalar_loop_iters = LOOP_VINFO_NITERSM1 (loop_vinfo); | |
1394 | unsigned limit = LOOP_VINFO_VECT_FACTOR (loop_vinfo) - 1; | |
1395 | if (LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo)) | |
1396 | limit = limit + 1; | |
1397 | if (check_profitability | |
1398 | && th > limit) | |
1399 | limit = th; | |
ebfd146a | 1400 | pre_condition = |
95b3eff3 RB |
1401 | fold_build2 (LT_EXPR, boolean_type_node, scalar_loop_iters, |
1402 | build_int_cst (TREE_TYPE (scalar_loop_iters), limit)); | |
86290011 RG |
1403 | if (cond_expr) |
1404 | { | |
1405 | pre_condition = | |
1406 | fold_build2 (TRUTH_OR_EXPR, boolean_type_node, | |
1407 | pre_condition, | |
1408 | fold_build1 (TRUTH_NOT_EXPR, boolean_type_node, | |
1409 | cond_expr)); | |
1410 | } | |
ebfd146a IR |
1411 | } |
1412 | ||
b8698a0f | 1413 | /* Prologue peeling. */ |
ebfd146a IR |
1414 | else |
1415 | { | |
1416 | if (check_profitability) | |
1417 | set_prologue_iterations (bb_before_first_loop, first_niters, | |
e78410bf | 1418 | loop, th, first_guard_probability); |
ebfd146a IR |
1419 | |
1420 | pre_condition = | |
5d2eb24b IR |
1421 | fold_build2 (LE_EXPR, boolean_type_node, *first_niters, |
1422 | build_int_cst (TREE_TYPE (*first_niters), 0)); | |
ebfd146a IR |
1423 | } |
1424 | ||
1425 | skip_e = slpeel_add_loop_guard (bb_before_first_loop, pre_condition, | |
86290011 | 1426 | cond_expr_stmt_list, |
e78410bf JH |
1427 | bb_before_second_loop, bb_before_first_loop, |
1428 | inverse_probability (first_guard_probability)); | |
1429 | scale_loop_profile (first_loop, first_guard_probability, | |
1430 | check_profitability && (int)th > bound1 ? th : bound1); | |
ebfd146a IR |
1431 | slpeel_update_phi_nodes_for_guard1 (skip_e, first_loop, |
1432 | first_loop == new_loop, | |
c334023f | 1433 | &new_exit_bb); |
ebfd146a IR |
1434 | |
1435 | ||
1436 | /* 3. Add the guard that controls whether the second loop is executed. | |
1437 | Resulting CFG would be: | |
1438 | ||
1439 | bb_before_first_loop: | |
1440 | if (FIRST_NITERS == 0) GOTO bb_before_second_loop (skip first loop) | |
1441 | GOTO first-loop | |
1442 | ||
1443 | first_loop: | |
1444 | do { | |
1445 | } while ... | |
1446 | ||
1447 | bb_between_loops: | |
1448 | if (FIRST_NITERS == NITERS) GOTO bb_after_second_loop (skip second loop) | |
1449 | GOTO bb_before_second_loop | |
1450 | ||
1451 | bb_before_second_loop: | |
1452 | ||
1453 | second_loop: | |
1454 | do { | |
1455 | } while ... | |
1456 | ||
1457 | bb_after_second_loop: | |
1458 | ||
1459 | orig_exit_bb: | |
1460 | */ | |
1461 | ||
1462 | bb_between_loops = new_exit_bb; | |
1463 | bb_after_second_loop = split_edge (single_exit (second_loop)); | |
1464 | ||
b8698a0f | 1465 | pre_condition = |
5d2eb24b | 1466 | fold_build2 (EQ_EXPR, boolean_type_node, *first_niters, niters); |
86290011 | 1467 | skip_e = slpeel_add_loop_guard (bb_between_loops, pre_condition, NULL, |
e78410bf JH |
1468 | bb_after_second_loop, bb_before_first_loop, |
1469 | inverse_probability (second_guard_probability)); | |
1470 | scale_loop_profile (second_loop, probability_of_second_loop, bound2); | |
ebfd146a IR |
1471 | slpeel_update_phi_nodes_for_guard2 (skip_e, second_loop, |
1472 | second_loop == new_loop, &new_exit_bb); | |
1473 | ||
1474 | /* 4. Make first-loop iterate FIRST_NITERS times, if requested. | |
1475 | */ | |
1476 | if (update_first_loop_count) | |
5d2eb24b | 1477 | slpeel_make_loop_iterate_ntimes (first_loop, *first_niters); |
ebfd146a | 1478 | |
040d39ee RG |
1479 | delete_update_ssa (); |
1480 | ||
684f25f4 AO |
1481 | adjust_vec_debug_stmts (); |
1482 | ||
ebfd146a IR |
1483 | return new_loop; |
1484 | } | |
1485 | ||
1486 | /* Function vect_get_loop_location. | |
1487 | ||
1488 | Extract the location of the loop in the source code. | |
1489 | If the loop is not well formed for vectorization, an estimated | |
1490 | location is calculated. | |
1491 | Return the loop location if succeed and NULL if not. */ | |
1492 | ||
b05e0233 | 1493 | source_location |
ebfd146a IR |
1494 | find_loop_location (struct loop *loop) |
1495 | { | |
355fe088 | 1496 | gimple *stmt = NULL; |
ebfd146a IR |
1497 | basic_block bb; |
1498 | gimple_stmt_iterator si; | |
1499 | ||
1500 | if (!loop) | |
b05e0233 | 1501 | return UNKNOWN_LOCATION; |
ebfd146a IR |
1502 | |
1503 | stmt = get_loop_exit_condition (loop); | |
1504 | ||
502498d5 JJ |
1505 | if (stmt |
1506 | && LOCATION_LOCUS (gimple_location (stmt)) > BUILTINS_LOCATION) | |
ebfd146a IR |
1507 | return gimple_location (stmt); |
1508 | ||
1509 | /* If we got here the loop is probably not "well formed", | |
1510 | try to estimate the loop location */ | |
1511 | ||
1512 | if (!loop->header) | |
b05e0233 | 1513 | return UNKNOWN_LOCATION; |
ebfd146a IR |
1514 | |
1515 | bb = loop->header; | |
1516 | ||
1517 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
1518 | { | |
1519 | stmt = gsi_stmt (si); | |
502498d5 | 1520 | if (LOCATION_LOCUS (gimple_location (stmt)) > BUILTINS_LOCATION) |
ebfd146a IR |
1521 | return gimple_location (stmt); |
1522 | } | |
1523 | ||
b05e0233 | 1524 | return UNKNOWN_LOCATION; |
ebfd146a IR |
1525 | } |
1526 | ||
1527 | ||
ebfd146a IR |
1528 | /* Function vect_can_advance_ivs_p |
1529 | ||
b8698a0f L |
1530 | In case the number of iterations that LOOP iterates is unknown at compile |
1531 | time, an epilog loop will be generated, and the loop induction variables | |
1532 | (IVs) will be "advanced" to the value they are supposed to take just before | |
ebfd146a IR |
1533 | the epilog loop. Here we check that the access function of the loop IVs |
1534 | and the expression that represents the loop bound are simple enough. | |
1535 | These restrictions will be relaxed in the future. */ | |
1536 | ||
b8698a0f | 1537 | bool |
ebfd146a IR |
1538 | vect_can_advance_ivs_p (loop_vec_info loop_vinfo) |
1539 | { | |
1540 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
1541 | basic_block bb = loop->header; | |
355fe088 | 1542 | gimple *phi; |
538dd0b7 | 1543 | gphi_iterator gsi; |
ebfd146a IR |
1544 | |
1545 | /* Analyze phi functions of the loop header. */ | |
1546 | ||
73fbfcad | 1547 | if (dump_enabled_p ()) |
e645e942 | 1548 | dump_printf_loc (MSG_NOTE, vect_location, "vect_can_advance_ivs_p:\n"); |
ebfd146a IR |
1549 | for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
1550 | { | |
ebfd146a IR |
1551 | tree evolution_part; |
1552 | ||
538dd0b7 | 1553 | phi = gsi.phi (); |
73fbfcad | 1554 | if (dump_enabled_p ()) |
ebfd146a | 1555 | { |
78c60e3d SS |
1556 | dump_printf_loc (MSG_NOTE, vect_location, "Analyze phi: "); |
1557 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0); | |
e645e942 | 1558 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
1559 | } |
1560 | ||
1561 | /* Skip virtual phi's. The data dependences that are associated with | |
1562 | virtual defs/uses (i.e., memory accesses) are analyzed elsewhere. */ | |
1563 | ||
ea057359 | 1564 | if (virtual_operand_p (PHI_RESULT (phi))) |
ebfd146a | 1565 | { |
73fbfcad | 1566 | if (dump_enabled_p ()) |
78c60e3d | 1567 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1568 | "virtual phi. skip.\n"); |
ebfd146a IR |
1569 | continue; |
1570 | } | |
1571 | ||
1572 | /* Skip reduction phis. */ | |
1573 | ||
1574 | if (STMT_VINFO_DEF_TYPE (vinfo_for_stmt (phi)) == vect_reduction_def) | |
1575 | { | |
73fbfcad | 1576 | if (dump_enabled_p ()) |
78c60e3d | 1577 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1578 | "reduc phi. skip.\n"); |
ebfd146a IR |
1579 | continue; |
1580 | } | |
1581 | ||
1582 | /* Analyze the evolution function. */ | |
1583 | ||
afb119be RB |
1584 | evolution_part |
1585 | = STMT_VINFO_LOOP_PHI_EVOLUTION_PART (vinfo_for_stmt (phi)); | |
ebfd146a IR |
1586 | if (evolution_part == NULL_TREE) |
1587 | { | |
73fbfcad | 1588 | if (dump_enabled_p ()) |
afb119be | 1589 | dump_printf (MSG_MISSED_OPTIMIZATION, |
e645e942 | 1590 | "No access function or evolution.\n"); |
ebfd146a IR |
1591 | return false; |
1592 | } | |
b8698a0f L |
1593 | |
1594 | /* FORNOW: We do not transform initial conditions of IVs | |
ebfd146a IR |
1595 | which evolution functions are a polynomial of degree >= 2. */ |
1596 | ||
1597 | if (tree_is_chrec (evolution_part)) | |
b8698a0f | 1598 | return false; |
ebfd146a IR |
1599 | } |
1600 | ||
1601 | return true; | |
1602 | } | |
1603 | ||
1604 | ||
1605 | /* Function vect_update_ivs_after_vectorizer. | |
1606 | ||
1607 | "Advance" the induction variables of LOOP to the value they should take | |
1608 | after the execution of LOOP. This is currently necessary because the | |
1609 | vectorizer does not handle induction variables that are used after the | |
1610 | loop. Such a situation occurs when the last iterations of LOOP are | |
1611 | peeled, because: | |
1612 | 1. We introduced new uses after LOOP for IVs that were not originally used | |
1613 | after LOOP: the IVs of LOOP are now used by an epilog loop. | |
1614 | 2. LOOP is going to be vectorized; this means that it will iterate N/VF | |
1615 | times, whereas the loop IVs should be bumped N times. | |
1616 | ||
1617 | Input: | |
1618 | - LOOP - a loop that is going to be vectorized. The last few iterations | |
1619 | of LOOP were peeled. | |
1620 | - NITERS - the number of iterations that LOOP executes (before it is | |
1621 | vectorized). i.e, the number of times the ivs should be bumped. | |
1622 | - UPDATE_E - a successor edge of LOOP->exit that is on the (only) path | |
1623 | coming out from LOOP on which there are uses of the LOOP ivs | |
1624 | (this is the path from LOOP->exit to epilog_loop->preheader). | |
1625 | ||
1626 | The new definitions of the ivs are placed in LOOP->exit. | |
1627 | The phi args associated with the edge UPDATE_E in the bb | |
1628 | UPDATE_E->dest are updated accordingly. | |
1629 | ||
1630 | Assumption 1: Like the rest of the vectorizer, this function assumes | |
1631 | a single loop exit that has a single predecessor. | |
1632 | ||
1633 | Assumption 2: The phi nodes in the LOOP header and in update_bb are | |
1634 | organized in the same order. | |
1635 | ||
1636 | Assumption 3: The access function of the ivs is simple enough (see | |
1637 | vect_can_advance_ivs_p). This assumption will be relaxed in the future. | |
1638 | ||
1639 | Assumption 4: Exactly one of the successors of LOOP exit-bb is on a path | |
b8698a0f | 1640 | coming out of LOOP on which the ivs of LOOP are used (this is the path |
ebfd146a IR |
1641 | that leads to the epilog loop; other paths skip the epilog loop). This |
1642 | path starts with the edge UPDATE_E, and its destination (denoted update_bb) | |
1643 | needs to have its phis updated. | |
1644 | */ | |
1645 | ||
1646 | static void | |
b8698a0f | 1647 | vect_update_ivs_after_vectorizer (loop_vec_info loop_vinfo, tree niters, |
ebfd146a IR |
1648 | edge update_e) |
1649 | { | |
1650 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
1651 | basic_block exit_bb = single_exit (loop)->dest; | |
538dd0b7 DM |
1652 | gphi *phi, *phi1; |
1653 | gphi_iterator gsi, gsi1; | |
ebfd146a IR |
1654 | basic_block update_bb = update_e->dest; |
1655 | ||
47c32082 | 1656 | gcc_checking_assert (vect_can_advance_ivs_p (loop_vinfo)); |
ebfd146a IR |
1657 | |
1658 | /* Make sure there exists a single-predecessor exit bb: */ | |
1659 | gcc_assert (single_pred_p (exit_bb)); | |
1660 | ||
1661 | for (gsi = gsi_start_phis (loop->header), gsi1 = gsi_start_phis (update_bb); | |
1662 | !gsi_end_p (gsi) && !gsi_end_p (gsi1); | |
1663 | gsi_next (&gsi), gsi_next (&gsi1)) | |
1664 | { | |
ebfd146a | 1665 | tree init_expr; |
550918ca RG |
1666 | tree step_expr, off; |
1667 | tree type; | |
ebfd146a IR |
1668 | tree var, ni, ni_name; |
1669 | gimple_stmt_iterator last_gsi; | |
0ac168a1 | 1670 | stmt_vec_info stmt_info; |
ebfd146a | 1671 | |
538dd0b7 DM |
1672 | phi = gsi.phi (); |
1673 | phi1 = gsi1.phi (); | |
73fbfcad | 1674 | if (dump_enabled_p ()) |
ebfd146a | 1675 | { |
78c60e3d SS |
1676 | dump_printf_loc (MSG_NOTE, vect_location, |
1677 | "vect_update_ivs_after_vectorizer: phi: "); | |
1678 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0); | |
e645e942 | 1679 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
1680 | } |
1681 | ||
1682 | /* Skip virtual phi's. */ | |
ea057359 | 1683 | if (virtual_operand_p (PHI_RESULT (phi))) |
ebfd146a | 1684 | { |
73fbfcad | 1685 | if (dump_enabled_p ()) |
78c60e3d | 1686 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1687 | "virtual phi. skip.\n"); |
ebfd146a IR |
1688 | continue; |
1689 | } | |
1690 | ||
1691 | /* Skip reduction phis. */ | |
0ac168a1 RG |
1692 | stmt_info = vinfo_for_stmt (phi); |
1693 | if (STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def) | |
b8698a0f | 1694 | { |
73fbfcad | 1695 | if (dump_enabled_p ()) |
78c60e3d | 1696 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
e645e942 | 1697 | "reduc phi. skip.\n"); |
ebfd146a | 1698 | continue; |
b8698a0f | 1699 | } |
ebfd146a | 1700 | |
0ac168a1 RG |
1701 | type = TREE_TYPE (gimple_phi_result (phi)); |
1702 | step_expr = STMT_VINFO_LOOP_PHI_EVOLUTION_PART (stmt_info); | |
1703 | step_expr = unshare_expr (step_expr); | |
b8698a0f | 1704 | |
ebfd146a IR |
1705 | /* FORNOW: We do not support IVs whose evolution function is a polynomial |
1706 | of degree >= 2 or exponential. */ | |
0ac168a1 | 1707 | gcc_assert (!tree_is_chrec (step_expr)); |
ebfd146a | 1708 | |
0ac168a1 | 1709 | init_expr = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop)); |
ebfd146a | 1710 | |
550918ca RG |
1711 | off = fold_build2 (MULT_EXPR, TREE_TYPE (step_expr), |
1712 | fold_convert (TREE_TYPE (step_expr), niters), | |
1713 | step_expr); | |
0ac168a1 | 1714 | if (POINTER_TYPE_P (type)) |
5d49b6a7 | 1715 | ni = fold_build_pointer_plus (init_expr, off); |
ebfd146a | 1716 | else |
0ac168a1 RG |
1717 | ni = fold_build2 (PLUS_EXPR, type, |
1718 | init_expr, fold_convert (type, off)); | |
ebfd146a | 1719 | |
0ac168a1 | 1720 | var = create_tmp_var (type, "tmp"); |
ebfd146a IR |
1721 | |
1722 | last_gsi = gsi_last_bb (exit_bb); | |
1723 | ni_name = force_gimple_operand_gsi (&last_gsi, ni, false, var, | |
1724 | true, GSI_SAME_STMT); | |
b8698a0f | 1725 | |
ebfd146a | 1726 | /* Fix phi expressions in the successor bb. */ |
684f25f4 | 1727 | adjust_phi_and_debug_stmts (phi1, update_e, ni_name); |
ebfd146a IR |
1728 | } |
1729 | } | |
1730 | ||
ebfd146a IR |
1731 | /* Function vect_do_peeling_for_loop_bound |
1732 | ||
1733 | Peel the last iterations of the loop represented by LOOP_VINFO. | |
b8698a0f | 1734 | The peeled iterations form a new epilog loop. Given that the loop now |
ebfd146a IR |
1735 | iterates NITERS times, the new epilog loop iterates |
1736 | NITERS % VECTORIZATION_FACTOR times. | |
b8698a0f L |
1737 | |
1738 | The original loop will later be made to iterate | |
86290011 RG |
1739 | NITERS / VECTORIZATION_FACTOR times (this value is placed into RATIO). |
1740 | ||
1741 | COND_EXPR and COND_EXPR_STMT_LIST are combined with a new generated | |
1742 | test. */ | |
ebfd146a | 1743 | |
b8698a0f | 1744 | void |
f3c92486 RB |
1745 | vect_do_peeling_for_loop_bound (loop_vec_info loop_vinfo, |
1746 | tree ni_name, tree ratio_mult_vf_name, | |
368117e8 | 1747 | unsigned int th, bool check_profitability) |
ebfd146a | 1748 | { |
ebfd146a | 1749 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
5ce9450f | 1750 | struct loop *scalar_loop = LOOP_VINFO_SCALAR_LOOP (loop_vinfo); |
ebfd146a IR |
1751 | struct loop *new_loop; |
1752 | edge update_e; | |
1753 | basic_block preheader; | |
1754 | int loop_num; | |
d68d56b5 | 1755 | int max_iter; |
368117e8 RG |
1756 | tree cond_expr = NULL_TREE; |
1757 | gimple_seq cond_expr_stmt_list = NULL; | |
ebfd146a | 1758 | |
73fbfcad | 1759 | if (dump_enabled_p ()) |
ccb3ad87 | 1760 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1761 | "=== vect_do_peeling_for_loop_bound ===\n"); |
ebfd146a IR |
1762 | |
1763 | initialize_original_copy_tables (); | |
1764 | ||
b8698a0f | 1765 | loop_num = loop->num; |
ebfd146a | 1766 | |
5ce9450f JJ |
1767 | new_loop |
1768 | = slpeel_tree_peel_loop_to_edge (loop, scalar_loop, single_exit (loop), | |
1769 | &ratio_mult_vf_name, ni_name, false, | |
1770 | th, check_profitability, | |
1771 | cond_expr, cond_expr_stmt_list, | |
1772 | 0, LOOP_VINFO_VECT_FACTOR (loop_vinfo)); | |
ebfd146a IR |
1773 | gcc_assert (new_loop); |
1774 | gcc_assert (loop_num == loop->num); | |
b2b29377 | 1775 | slpeel_checking_verify_cfg_after_peeling (loop, new_loop); |
ebfd146a IR |
1776 | |
1777 | /* A guard that controls whether the new_loop is to be executed or skipped | |
1778 | is placed in LOOP->exit. LOOP->exit therefore has two successors - one | |
1779 | is the preheader of NEW_LOOP, where the IVs from LOOP are used. The other | |
1780 | is a bb after NEW_LOOP, where these IVs are not used. Find the edge that | |
1781 | is on the path where the LOOP IVs are used and need to be updated. */ | |
1782 | ||
1783 | preheader = loop_preheader_edge (new_loop)->src; | |
1784 | if (EDGE_PRED (preheader, 0)->src == single_exit (loop)->dest) | |
1785 | update_e = EDGE_PRED (preheader, 0); | |
1786 | else | |
1787 | update_e = EDGE_PRED (preheader, 1); | |
1788 | ||
b8698a0f | 1789 | /* Update IVs of original loop as if they were advanced |
ebfd146a | 1790 | by ratio_mult_vf_name steps. */ |
b8698a0f | 1791 | vect_update_ivs_after_vectorizer (loop_vinfo, ratio_mult_vf_name, update_e); |
ebfd146a | 1792 | |
22458c5a JH |
1793 | /* For vectorization factor N, we need to copy last N-1 values in epilogue |
1794 | and this means N-2 loopback edge executions. | |
1795 | ||
1796 | PEELING_FOR_GAPS works by subtracting last iteration and thus the epilogue | |
1797 | will execute at least LOOP_VINFO_VECT_FACTOR times. */ | |
1798 | max_iter = (LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) | |
1799 | ? LOOP_VINFO_VECT_FACTOR (loop_vinfo) * 2 | |
1800 | : LOOP_VINFO_VECT_FACTOR (loop_vinfo)) - 2; | |
368117e8 | 1801 | if (check_profitability) |
22458c5a | 1802 | max_iter = MAX (max_iter, (int) th - 1); |
807e902e | 1803 | record_niter_bound (new_loop, max_iter, false, true); |
ccb3ad87 | 1804 | dump_printf (MSG_NOTE, |
78c60e3d SS |
1805 | "Setting upper bound of nb iterations for epilogue " |
1806 | "loop to %d\n", max_iter); | |
7d5a99f4 | 1807 | |
ebfd146a IR |
1808 | /* After peeling we have to reset scalar evolution analyzer. */ |
1809 | scev_reset (); | |
1810 | ||
1811 | free_original_copy_tables (); | |
1812 | } | |
1813 | ||
1814 | ||
1815 | /* Function vect_gen_niters_for_prolog_loop | |
1816 | ||
1817 | Set the number of iterations for the loop represented by LOOP_VINFO | |
1818 | to the minimum between LOOP_NITERS (the original iteration count of the loop) | |
1819 | and the misalignment of DR - the data reference recorded in | |
b8698a0f | 1820 | LOOP_VINFO_UNALIGNED_DR (LOOP_VINFO). As a result, after the execution of |
ebfd146a IR |
1821 | this loop, the data reference DR will refer to an aligned location. |
1822 | ||
1823 | The following computation is generated: | |
1824 | ||
1825 | If the misalignment of DR is known at compile time: | |
1826 | addr_mis = int mis = DR_MISALIGNMENT (dr); | |
1827 | Else, compute address misalignment in bytes: | |
5aea1e76 | 1828 | addr_mis = addr & (vectype_align - 1) |
ebfd146a IR |
1829 | |
1830 | prolog_niters = min (LOOP_NITERS, ((VF - addr_mis/elem_size)&(VF-1))/step) | |
1831 | ||
1832 | (elem_size = element type size; an element is the scalar element whose type | |
1833 | is the inner type of the vectype) | |
1834 | ||
1835 | When the step of the data-ref in the loop is not 1 (as in interleaved data | |
1836 | and SLP), the number of iterations of the prolog must be divided by the step | |
1837 | (which is equal to the size of interleaved group). | |
1838 | ||
1839 | The above formulas assume that VF == number of elements in the vector. This | |
1840 | may not hold when there are multiple-types in the loop. | |
1841 | In this case, for some data-references in the loop the VF does not represent | |
1842 | the number of elements that fit in the vector. Therefore, instead of VF we | |
1843 | use TYPE_VECTOR_SUBPARTS. */ | |
1844 | ||
b8698a0f | 1845 | static tree |
e78410bf | 1846 | vect_gen_niters_for_prolog_loop (loop_vec_info loop_vinfo, tree loop_niters, int *bound) |
ebfd146a IR |
1847 | { |
1848 | struct data_reference *dr = LOOP_VINFO_UNALIGNED_DR (loop_vinfo); | |
1849 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
1850 | tree var; | |
1851 | gimple_seq stmts; | |
1852 | tree iters, iters_name; | |
1853 | edge pe; | |
1854 | basic_block new_bb; | |
355fe088 | 1855 | gimple *dr_stmt = DR_STMT (dr); |
ebfd146a IR |
1856 | stmt_vec_info stmt_info = vinfo_for_stmt (dr_stmt); |
1857 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
1858 | int vectype_align = TYPE_ALIGN (vectype) / BITS_PER_UNIT; | |
1859 | tree niters_type = TREE_TYPE (loop_niters); | |
ebfd146a IR |
1860 | int nelements = TYPE_VECTOR_SUBPARTS (vectype); |
1861 | ||
b8698a0f | 1862 | pe = loop_preheader_edge (loop); |
ebfd146a | 1863 | |
15e693cc | 1864 | if (LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo) > 0) |
ebfd146a | 1865 | { |
15e693cc | 1866 | int npeel = LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo); |
ebfd146a | 1867 | |
73fbfcad | 1868 | if (dump_enabled_p ()) |
ccb3ad87 | 1869 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1870 | "known peeling = %d.\n", npeel); |
ebfd146a | 1871 | |
720f5239 | 1872 | iters = build_int_cst (niters_type, npeel); |
15e693cc | 1873 | *bound = LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo); |
ebfd146a IR |
1874 | } |
1875 | else | |
1876 | { | |
1877 | gimple_seq new_stmts = NULL; | |
d8ba5b19 RG |
1878 | bool negative = tree_int_cst_compare (DR_STEP (dr), size_zero_node) < 0; |
1879 | tree offset = negative | |
5fa79de8 | 1880 | ? size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1) : size_zero_node; |
b8698a0f | 1881 | tree start_addr = vect_create_addr_base_for_vector_ref (dr_stmt, |
d8ba5b19 | 1882 | &new_stmts, offset, loop); |
96f9265a | 1883 | tree type = unsigned_type_for (TREE_TYPE (start_addr)); |
5aea1e76 UW |
1884 | tree vectype_align_minus_1 = build_int_cst (type, vectype_align - 1); |
1885 | HOST_WIDE_INT elem_size = | |
1886 | int_cst_value (TYPE_SIZE_UNIT (TREE_TYPE (vectype))); | |
1887 | tree elem_size_log = build_int_cst (type, exact_log2 (elem_size)); | |
ebfd146a IR |
1888 | tree nelements_minus_1 = build_int_cst (type, nelements - 1); |
1889 | tree nelements_tree = build_int_cst (type, nelements); | |
1890 | tree byte_misalign; | |
1891 | tree elem_misalign; | |
1892 | ||
1893 | new_bb = gsi_insert_seq_on_edge_immediate (pe, new_stmts); | |
1894 | gcc_assert (!new_bb); | |
b8698a0f | 1895 | |
5aea1e76 | 1896 | /* Create: byte_misalign = addr & (vectype_align - 1) */ |
b8698a0f | 1897 | byte_misalign = |
720f5239 | 1898 | fold_build2 (BIT_AND_EXPR, type, fold_convert (type, start_addr), |
5aea1e76 | 1899 | vectype_align_minus_1); |
b8698a0f | 1900 | |
ebfd146a IR |
1901 | /* Create: elem_misalign = byte_misalign / element_size */ |
1902 | elem_misalign = | |
1903 | fold_build2 (RSHIFT_EXPR, type, byte_misalign, elem_size_log); | |
1904 | ||
1905 | /* Create: (niters_type) (nelements - elem_misalign)&(nelements - 1) */ | |
d8ba5b19 RG |
1906 | if (negative) |
1907 | iters = fold_build2 (MINUS_EXPR, type, elem_misalign, nelements_tree); | |
1908 | else | |
1909 | iters = fold_build2 (MINUS_EXPR, type, nelements_tree, elem_misalign); | |
ebfd146a IR |
1910 | iters = fold_build2 (BIT_AND_EXPR, type, iters, nelements_minus_1); |
1911 | iters = fold_convert (niters_type, iters); | |
e78410bf | 1912 | *bound = nelements; |
ebfd146a IR |
1913 | } |
1914 | ||
1915 | /* Create: prolog_loop_niters = min (iters, loop_niters) */ | |
1916 | /* If the loop bound is known at compile time we already verified that it is | |
1917 | greater than vf; since the misalignment ('iters') is at most vf, there's | |
1918 | no need to generate the MIN_EXPR in this case. */ | |
1919 | if (TREE_CODE (loop_niters) != INTEGER_CST) | |
1920 | iters = fold_build2 (MIN_EXPR, niters_type, iters, loop_niters); | |
1921 | ||
73fbfcad | 1922 | if (dump_enabled_p ()) |
ebfd146a | 1923 | { |
ccb3ad87 | 1924 | dump_printf_loc (MSG_NOTE, vect_location, |
78c60e3d | 1925 | "niters for prolog loop: "); |
ccb3ad87 | 1926 | dump_generic_expr (MSG_NOTE, TDF_SLIM, iters); |
e645e942 | 1927 | dump_printf (MSG_NOTE, "\n"); |
ebfd146a IR |
1928 | } |
1929 | ||
1930 | var = create_tmp_var (niters_type, "prolog_loop_niters"); | |
ebfd146a IR |
1931 | stmts = NULL; |
1932 | iters_name = force_gimple_operand (iters, &stmts, false, var); | |
1933 | ||
1934 | /* Insert stmt on loop preheader edge. */ | |
1935 | if (stmts) | |
1936 | { | |
1937 | basic_block new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
1938 | gcc_assert (!new_bb); | |
1939 | } | |
1940 | ||
b8698a0f | 1941 | return iters_name; |
ebfd146a IR |
1942 | } |
1943 | ||
1944 | ||
1945 | /* Function vect_update_init_of_dr | |
1946 | ||
1947 | NITERS iterations were peeled from LOOP. DR represents a data reference | |
1948 | in LOOP. This function updates the information recorded in DR to | |
b8698a0f | 1949 | account for the fact that the first NITERS iterations had already been |
ebfd146a IR |
1950 | executed. Specifically, it updates the OFFSET field of DR. */ |
1951 | ||
1952 | static void | |
1953 | vect_update_init_of_dr (struct data_reference *dr, tree niters) | |
1954 | { | |
1955 | tree offset = DR_OFFSET (dr); | |
b8698a0f | 1956 | |
ebfd146a IR |
1957 | niters = fold_build2 (MULT_EXPR, sizetype, |
1958 | fold_convert (sizetype, niters), | |
1959 | fold_convert (sizetype, DR_STEP (dr))); | |
587aa063 RG |
1960 | offset = fold_build2 (PLUS_EXPR, sizetype, |
1961 | fold_convert (sizetype, offset), niters); | |
ebfd146a IR |
1962 | DR_OFFSET (dr) = offset; |
1963 | } | |
1964 | ||
1965 | ||
1966 | /* Function vect_update_inits_of_drs | |
1967 | ||
b8698a0f L |
1968 | NITERS iterations were peeled from the loop represented by LOOP_VINFO. |
1969 | This function updates the information recorded for the data references in | |
1970 | the loop to account for the fact that the first NITERS iterations had | |
ebfd146a IR |
1971 | already been executed. Specifically, it updates the initial_condition of |
1972 | the access_function of all the data_references in the loop. */ | |
1973 | ||
1974 | static void | |
1975 | vect_update_inits_of_drs (loop_vec_info loop_vinfo, tree niters) | |
1976 | { | |
1977 | unsigned int i; | |
9771b263 | 1978 | vec<data_reference_p> datarefs = LOOP_VINFO_DATAREFS (loop_vinfo); |
ebfd146a | 1979 | struct data_reference *dr; |
78c60e3d | 1980 | |
73fbfcad | 1981 | if (dump_enabled_p ()) |
ccb3ad87 | 1982 | dump_printf_loc (MSG_NOTE, vect_location, |
e645e942 | 1983 | "=== vect_update_inits_of_dr ===\n"); |
ebfd146a | 1984 | |
9771b263 | 1985 | FOR_EACH_VEC_ELT (datarefs, i, dr) |
ebfd146a IR |
1986 | vect_update_init_of_dr (dr, niters); |
1987 | } | |
1988 | ||
1989 | ||
1990 | /* Function vect_do_peeling_for_alignment | |
1991 | ||
1992 | Peel the first 'niters' iterations of the loop represented by LOOP_VINFO. | |
1993 | 'niters' is set to the misalignment of one of the data references in the | |
1994 | loop, thereby forcing it to refer to an aligned location at the beginning | |
1995 | of the execution of this loop. The data reference for which we are | |
1996 | peeling is recorded in LOOP_VINFO_UNALIGNED_DR. */ | |
1997 | ||
1998 | void | |
f3c92486 | 1999 | vect_do_peeling_for_alignment (loop_vec_info loop_vinfo, tree ni_name, |
368117e8 | 2000 | unsigned int th, bool check_profitability) |
ebfd146a IR |
2001 | { |
2002 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
5ce9450f | 2003 | struct loop *scalar_loop = LOOP_VINFO_SCALAR_LOOP (loop_vinfo); |
f3c92486 | 2004 | tree niters_of_prolog_loop; |
b61b1f17 | 2005 | tree wide_prolog_niters; |
ebfd146a | 2006 | struct loop *new_loop; |
03fd03d5 | 2007 | int max_iter; |
e78410bf | 2008 | int bound = 0; |
ebfd146a | 2009 | |
73fbfcad | 2010 | if (dump_enabled_p ()) |
9cc1fb4b XDL |
2011 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, vect_location, |
2012 | "loop peeled for vectorization to enhance" | |
2013 | " alignment\n"); | |
ebfd146a IR |
2014 | |
2015 | initialize_original_copy_tables (); | |
2016 | ||
f3c92486 RB |
2017 | gimple_seq stmts = NULL; |
2018 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
5d2eb24b | 2019 | niters_of_prolog_loop = vect_gen_niters_for_prolog_loop (loop_vinfo, |
e78410bf JH |
2020 | ni_name, |
2021 | &bound); | |
ebfd146a | 2022 | |
ebfd146a IR |
2023 | /* Peel the prolog loop and iterate it niters_of_prolog_loop. */ |
2024 | new_loop = | |
5ce9450f JJ |
2025 | slpeel_tree_peel_loop_to_edge (loop, scalar_loop, |
2026 | loop_preheader_edge (loop), | |
5d2eb24b | 2027 | &niters_of_prolog_loop, ni_name, true, |
e78410bf | 2028 | th, check_profitability, NULL_TREE, NULL, |
5ce9450f | 2029 | bound, 0); |
ebfd146a IR |
2030 | |
2031 | gcc_assert (new_loop); | |
b2b29377 | 2032 | slpeel_checking_verify_cfg_after_peeling (new_loop, loop); |
22458c5a JH |
2033 | /* For vectorization factor N, we need to copy at most N-1 values |
2034 | for alignment and this means N-2 loopback edge executions. */ | |
2035 | max_iter = LOOP_VINFO_VECT_FACTOR (loop_vinfo) - 2; | |
368117e8 | 2036 | if (check_profitability) |
22458c5a | 2037 | max_iter = MAX (max_iter, (int) th - 1); |
807e902e | 2038 | record_niter_bound (new_loop, max_iter, false, true); |
ccb3ad87 | 2039 | dump_printf (MSG_NOTE, |
78c60e3d SS |
2040 | "Setting upper bound of nb iterations for prologue " |
2041 | "loop to %d\n", max_iter); | |
ebfd146a IR |
2042 | |
2043 | /* Update number of times loop executes. */ | |
ebfd146a | 2044 | LOOP_VINFO_NITERS (loop_vinfo) = fold_build2 (MINUS_EXPR, |
15e693cc | 2045 | TREE_TYPE (ni_name), ni_name, niters_of_prolog_loop); |
95b3eff3 RB |
2046 | LOOP_VINFO_NITERSM1 (loop_vinfo) = fold_build2 (MINUS_EXPR, |
2047 | TREE_TYPE (ni_name), | |
2048 | LOOP_VINFO_NITERSM1 (loop_vinfo), niters_of_prolog_loop); | |
ebfd146a | 2049 | |
5d2eb24b IR |
2050 | if (types_compatible_p (sizetype, TREE_TYPE (niters_of_prolog_loop))) |
2051 | wide_prolog_niters = niters_of_prolog_loop; | |
2052 | else | |
2053 | { | |
2054 | gimple_seq seq = NULL; | |
2055 | edge pe = loop_preheader_edge (loop); | |
2056 | tree wide_iters = fold_convert (sizetype, niters_of_prolog_loop); | |
2057 | tree var = create_tmp_var (sizetype, "prolog_loop_adjusted_niters"); | |
5d2eb24b IR |
2058 | wide_prolog_niters = force_gimple_operand (wide_iters, &seq, false, |
2059 | var); | |
2060 | if (seq) | |
2061 | { | |
2062 | /* Insert stmt on loop preheader edge. */ | |
2063 | basic_block new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
2064 | gcc_assert (!new_bb); | |
2065 | } | |
2066 | } | |
2067 | ||
ebfd146a | 2068 | /* Update the init conditions of the access functions of all data refs. */ |
b61b1f17 | 2069 | vect_update_inits_of_drs (loop_vinfo, wide_prolog_niters); |
ebfd146a IR |
2070 | |
2071 | /* After peeling we have to reset scalar evolution analyzer. */ | |
2072 | scev_reset (); | |
2073 | ||
2074 | free_original_copy_tables (); | |
2075 | } | |
2076 | ||
2077 | ||
2078 | /* Function vect_create_cond_for_align_checks. | |
2079 | ||
2080 | Create a conditional expression that represents the alignment checks for | |
2081 | all of data references (array element references) whose alignment must be | |
2082 | checked at runtime. | |
2083 | ||
2084 | Input: | |
2085 | COND_EXPR - input conditional expression. New conditions will be chained | |
2086 | with logical AND operation. | |
2087 | LOOP_VINFO - two fields of the loop information are used. | |
2088 | LOOP_VINFO_PTR_MASK is the mask used to check the alignment. | |
2089 | LOOP_VINFO_MAY_MISALIGN_STMTS contains the refs to be checked. | |
2090 | ||
2091 | Output: | |
2092 | COND_EXPR_STMT_LIST - statements needed to construct the conditional | |
2093 | expression. | |
2094 | The returned value is the conditional expression to be used in the if | |
2095 | statement that controls which version of the loop gets executed at runtime. | |
2096 | ||
2097 | The algorithm makes two assumptions: | |
2098 | 1) The number of bytes "n" in a vector is a power of 2. | |
2099 | 2) An address "a" is aligned if a%n is zero and that this | |
2100 | test can be done as a&(n-1) == 0. For example, for 16 | |
2101 | byte vectors the test is a&0xf == 0. */ | |
2102 | ||
2103 | static void | |
2104 | vect_create_cond_for_align_checks (loop_vec_info loop_vinfo, | |
2105 | tree *cond_expr, | |
2106 | gimple_seq *cond_expr_stmt_list) | |
2107 | { | |
2108 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
355fe088 | 2109 | vec<gimple *> may_misalign_stmts |
ebfd146a | 2110 | = LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo); |
355fe088 | 2111 | gimple *ref_stmt; |
ebfd146a IR |
2112 | int mask = LOOP_VINFO_PTR_MASK (loop_vinfo); |
2113 | tree mask_cst; | |
2114 | unsigned int i; | |
ebfd146a IR |
2115 | tree int_ptrsize_type; |
2116 | char tmp_name[20]; | |
2117 | tree or_tmp_name = NULL_TREE; | |
83d5977e | 2118 | tree and_tmp_name; |
355fe088 | 2119 | gimple *and_stmt; |
ebfd146a IR |
2120 | tree ptrsize_zero; |
2121 | tree part_cond_expr; | |
2122 | ||
2123 | /* Check that mask is one less than a power of 2, i.e., mask is | |
2124 | all zeros followed by all ones. */ | |
2125 | gcc_assert ((mask != 0) && ((mask & (mask+1)) == 0)); | |
2126 | ||
96f9265a | 2127 | int_ptrsize_type = signed_type_for (ptr_type_node); |
ebfd146a IR |
2128 | |
2129 | /* Create expression (mask & (dr_1 || ... || dr_n)) where dr_i is the address | |
2130 | of the first vector of the i'th data reference. */ | |
2131 | ||
9771b263 | 2132 | FOR_EACH_VEC_ELT (may_misalign_stmts, i, ref_stmt) |
ebfd146a IR |
2133 | { |
2134 | gimple_seq new_stmt_list = NULL; | |
2135 | tree addr_base; | |
83d5977e RG |
2136 | tree addr_tmp_name; |
2137 | tree new_or_tmp_name; | |
355fe088 | 2138 | gimple *addr_stmt, *or_stmt; |
d8ba5b19 RG |
2139 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (ref_stmt); |
2140 | tree vectype = STMT_VINFO_VECTYPE (stmt_vinfo); | |
2141 | bool negative = tree_int_cst_compare | |
2142 | (DR_STEP (STMT_VINFO_DATA_REF (stmt_vinfo)), size_zero_node) < 0; | |
2143 | tree offset = negative | |
aad83b7c | 2144 | ? size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1) : size_zero_node; |
ebfd146a IR |
2145 | |
2146 | /* create: addr_tmp = (int)(address_of_first_vector) */ | |
2147 | addr_base = | |
2148 | vect_create_addr_base_for_vector_ref (ref_stmt, &new_stmt_list, | |
d8ba5b19 | 2149 | offset, loop); |
ebfd146a IR |
2150 | if (new_stmt_list != NULL) |
2151 | gimple_seq_add_seq (cond_expr_stmt_list, new_stmt_list); | |
2152 | ||
83d5977e RG |
2153 | sprintf (tmp_name, "addr2int%d", i); |
2154 | addr_tmp_name = make_temp_ssa_name (int_ptrsize_type, NULL, tmp_name); | |
0d0e4a03 | 2155 | addr_stmt = gimple_build_assign (addr_tmp_name, NOP_EXPR, addr_base); |
ebfd146a IR |
2156 | gimple_seq_add_stmt (cond_expr_stmt_list, addr_stmt); |
2157 | ||
2158 | /* The addresses are OR together. */ | |
2159 | ||
2160 | if (or_tmp_name != NULL_TREE) | |
2161 | { | |
2162 | /* create: or_tmp = or_tmp | addr_tmp */ | |
83d5977e RG |
2163 | sprintf (tmp_name, "orptrs%d", i); |
2164 | new_or_tmp_name = make_temp_ssa_name (int_ptrsize_type, NULL, tmp_name); | |
0d0e4a03 JJ |
2165 | or_stmt = gimple_build_assign (new_or_tmp_name, BIT_IOR_EXPR, |
2166 | or_tmp_name, addr_tmp_name); | |
ebfd146a IR |
2167 | gimple_seq_add_stmt (cond_expr_stmt_list, or_stmt); |
2168 | or_tmp_name = new_or_tmp_name; | |
2169 | } | |
2170 | else | |
2171 | or_tmp_name = addr_tmp_name; | |
2172 | ||
2173 | } /* end for i */ | |
2174 | ||
2175 | mask_cst = build_int_cst (int_ptrsize_type, mask); | |
2176 | ||
2177 | /* create: and_tmp = or_tmp & mask */ | |
83d5977e | 2178 | and_tmp_name = make_temp_ssa_name (int_ptrsize_type, NULL, "andmask"); |
ebfd146a | 2179 | |
0d0e4a03 JJ |
2180 | and_stmt = gimple_build_assign (and_tmp_name, BIT_AND_EXPR, |
2181 | or_tmp_name, mask_cst); | |
ebfd146a IR |
2182 | gimple_seq_add_stmt (cond_expr_stmt_list, and_stmt); |
2183 | ||
2184 | /* Make and_tmp the left operand of the conditional test against zero. | |
2185 | if and_tmp has a nonzero bit then some address is unaligned. */ | |
2186 | ptrsize_zero = build_int_cst (int_ptrsize_type, 0); | |
2187 | part_cond_expr = fold_build2 (EQ_EXPR, boolean_type_node, | |
2188 | and_tmp_name, ptrsize_zero); | |
2189 | if (*cond_expr) | |
2190 | *cond_expr = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, | |
2191 | *cond_expr, part_cond_expr); | |
2192 | else | |
2193 | *cond_expr = part_cond_expr; | |
2194 | } | |
2195 | ||
ebfd146a IR |
2196 | /* Function vect_create_cond_for_alias_checks. |
2197 | ||
2198 | Create a conditional expression that represents the run-time checks for | |
2199 | overlapping of address ranges represented by a list of data references | |
2200 | relations passed as input. | |
2201 | ||
2202 | Input: | |
2203 | COND_EXPR - input conditional expression. New conditions will be chained | |
a05a89fa CH |
2204 | with logical AND operation. If it is NULL, then the function |
2205 | is used to return the number of alias checks. | |
ebfd146a IR |
2206 | LOOP_VINFO - field LOOP_VINFO_MAY_ALIAS_STMTS contains the list of ddrs |
2207 | to be checked. | |
2208 | ||
2209 | Output: | |
2210 | COND_EXPR - conditional expression. | |
ebfd146a | 2211 | |
a05a89fa | 2212 | The returned COND_EXPR is the conditional expression to be used in the if |
ebfd146a IR |
2213 | statement that controls which version of the loop gets executed at runtime. |
2214 | */ | |
2215 | ||
a05a89fa | 2216 | void |
4bdd44c4 | 2217 | vect_create_cond_for_alias_checks (loop_vec_info loop_vinfo, tree * cond_expr) |
ebfd146a | 2218 | { |
93bdc3ed | 2219 | vec<dr_with_seg_len_pair_t> comp_alias_ddrs = |
a05a89fa CH |
2220 | LOOP_VINFO_COMP_ALIAS_DDRS (loop_vinfo); |
2221 | tree part_cond_expr; | |
ebfd146a IR |
2222 | |
2223 | /* Create expression | |
36fc3799 RS |
2224 | ((store_ptr_0 + store_segment_length_0) <= load_ptr_0) |
2225 | || (load_ptr_0 + load_segment_length_0) <= store_ptr_0)) | |
b8698a0f | 2226 | && |
ebfd146a IR |
2227 | ... |
2228 | && | |
36fc3799 RS |
2229 | ((store_ptr_n + store_segment_length_n) <= load_ptr_n) |
2230 | || (load_ptr_n + load_segment_length_n) <= store_ptr_n)) */ | |
ebfd146a | 2231 | |
a05a89fa | 2232 | if (comp_alias_ddrs.is_empty ()) |
ebfd146a IR |
2233 | return; |
2234 | ||
a05a89fa | 2235 | for (size_t i = 0, s = comp_alias_ddrs.length (); i < s; ++i) |
ebfd146a | 2236 | { |
93bdc3ed CH |
2237 | const dr_with_seg_len& dr_a = comp_alias_ddrs[i].first; |
2238 | const dr_with_seg_len& dr_b = comp_alias_ddrs[i].second; | |
a05a89fa CH |
2239 | tree segment_length_a = dr_a.seg_len; |
2240 | tree segment_length_b = dr_b.seg_len; | |
ebfd146a | 2241 | |
a05a89fa | 2242 | tree addr_base_a |
93bdc3ed | 2243 | = fold_build_pointer_plus (DR_BASE_ADDRESS (dr_a.dr), dr_a.offset); |
a05a89fa | 2244 | tree addr_base_b |
93bdc3ed | 2245 | = fold_build_pointer_plus (DR_BASE_ADDRESS (dr_b.dr), dr_b.offset); |
ebfd146a | 2246 | |
73fbfcad | 2247 | if (dump_enabled_p ()) |
ebfd146a | 2248 | { |
ccb3ad87 | 2249 | dump_printf_loc (MSG_NOTE, vect_location, |
a05a89fa CH |
2250 | "create runtime check for data references "); |
2251 | dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (dr_a.dr)); | |
ccb3ad87 | 2252 | dump_printf (MSG_NOTE, " and "); |
a05a89fa CH |
2253 | dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (dr_b.dr)); |
2254 | dump_printf (MSG_NOTE, "\n"); | |
ebfd146a IR |
2255 | } |
2256 | ||
a05a89fa CH |
2257 | tree seg_a_min = addr_base_a; |
2258 | tree seg_a_max = fold_build_pointer_plus (addr_base_a, segment_length_a); | |
82d89471 BM |
2259 | /* For negative step, we need to adjust address range by TYPE_SIZE_UNIT |
2260 | bytes, e.g., int a[3] -> a[1] range is [a+4, a+16) instead of | |
2261 | [a, a+12) */ | |
a05a89fa | 2262 | if (tree_int_cst_compare (DR_STEP (dr_a.dr), size_zero_node) < 0) |
82d89471 BM |
2263 | { |
2264 | tree unit_size = TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dr_a.dr))); | |
2265 | seg_a_min = fold_build_pointer_plus (seg_a_max, unit_size); | |
2266 | seg_a_max = fold_build_pointer_plus (addr_base_a, unit_size); | |
2267 | } | |
d8ba5b19 | 2268 | |
a05a89fa CH |
2269 | tree seg_b_min = addr_base_b; |
2270 | tree seg_b_max = fold_build_pointer_plus (addr_base_b, segment_length_b); | |
2271 | if (tree_int_cst_compare (DR_STEP (dr_b.dr), size_zero_node) < 0) | |
82d89471 BM |
2272 | { |
2273 | tree unit_size = TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dr_b.dr))); | |
2274 | seg_b_min = fold_build_pointer_plus (seg_b_max, unit_size); | |
2275 | seg_b_max = fold_build_pointer_plus (addr_base_b, unit_size); | |
2276 | } | |
ebfd146a | 2277 | |
b8698a0f | 2278 | part_cond_expr = |
ebfd146a | 2279 | fold_build2 (TRUTH_OR_EXPR, boolean_type_node, |
36fc3799 RS |
2280 | fold_build2 (LE_EXPR, boolean_type_node, seg_a_max, seg_b_min), |
2281 | fold_build2 (LE_EXPR, boolean_type_node, seg_b_max, seg_a_min)); | |
b8698a0f | 2282 | |
ebfd146a IR |
2283 | if (*cond_expr) |
2284 | *cond_expr = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, | |
2285 | *cond_expr, part_cond_expr); | |
2286 | else | |
2287 | *cond_expr = part_cond_expr; | |
2288 | } | |
ebfd146a | 2289 | |
73fbfcad | 2290 | if (dump_enabled_p ()) |
ccb3ad87 | 2291 | dump_printf_loc (MSG_NOTE, vect_location, |
78c60e3d | 2292 | "created %u versioning for alias checks.\n", |
a05a89fa CH |
2293 | comp_alias_ddrs.length ()); |
2294 | ||
2295 | comp_alias_ddrs.release (); | |
ebfd146a IR |
2296 | } |
2297 | ||
2298 | ||
2299 | /* Function vect_loop_versioning. | |
b8698a0f | 2300 | |
ebfd146a IR |
2301 | If the loop has data references that may or may not be aligned or/and |
2302 | has data reference relations whose independence was not proven then | |
2303 | two versions of the loop need to be generated, one which is vectorized | |
2304 | and one which isn't. A test is then generated to control which of the | |
2305 | loops is executed. The test checks for the alignment of all of the | |
2306 | data references that may or may not be aligned. An additional | |
2307 | sequence of runtime tests is generated for each pairs of DDRs whose | |
b8698a0f L |
2308 | independence was not proven. The vectorized version of loop is |
2309 | executed only if both alias and alignment tests are passed. | |
2310 | ||
ebfd146a | 2311 | The test generated to check which version of loop is executed |
b8698a0f | 2312 | is modified to also check for profitability as indicated by the |
86290011 RG |
2313 | cost model initially. |
2314 | ||
2315 | The versioning precondition(s) are placed in *COND_EXPR and | |
d68d56b5 | 2316 | *COND_EXPR_STMT_LIST. */ |
ebfd146a IR |
2317 | |
2318 | void | |
368117e8 RG |
2319 | vect_loop_versioning (loop_vec_info loop_vinfo, |
2320 | unsigned int th, bool check_profitability) | |
ebfd146a IR |
2321 | { |
2322 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
5ce9450f | 2323 | struct loop *scalar_loop = LOOP_VINFO_SCALAR_LOOP (loop_vinfo); |
ebfd146a | 2324 | basic_block condition_bb; |
538dd0b7 DM |
2325 | gphi_iterator gsi; |
2326 | gimple_stmt_iterator cond_exp_gsi; | |
ebfd146a IR |
2327 | basic_block merge_bb; |
2328 | basic_block new_exit_bb; | |
2329 | edge new_exit_e, e; | |
538dd0b7 | 2330 | gphi *orig_phi, *new_phi; |
368117e8 | 2331 | tree cond_expr = NULL_TREE; |
d68d56b5 | 2332 | gimple_seq cond_expr_stmt_list = NULL; |
ebfd146a IR |
2333 | tree arg; |
2334 | unsigned prob = 4 * REG_BR_PROB_BASE / 5; | |
2335 | gimple_seq gimplify_stmt_list = NULL; | |
2336 | tree scalar_loop_iters = LOOP_VINFO_NITERS (loop_vinfo); | |
9cc1fb4b XDL |
2337 | bool version_align = LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo); |
2338 | bool version_alias = LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo); | |
ebfd146a | 2339 | |
368117e8 RG |
2340 | if (check_profitability) |
2341 | { | |
2342 | cond_expr = fold_build2 (GT_EXPR, boolean_type_node, scalar_loop_iters, | |
2343 | build_int_cst (TREE_TYPE (scalar_loop_iters), th)); | |
2344 | cond_expr = force_gimple_operand_1 (cond_expr, &cond_expr_stmt_list, | |
2345 | is_gimple_condexpr, NULL_TREE); | |
2346 | } | |
ebfd146a | 2347 | |
9cc1fb4b | 2348 | if (version_align) |
d68d56b5 RG |
2349 | vect_create_cond_for_align_checks (loop_vinfo, &cond_expr, |
2350 | &cond_expr_stmt_list); | |
ebfd146a | 2351 | |
9cc1fb4b | 2352 | if (version_alias) |
4bdd44c4 | 2353 | vect_create_cond_for_alias_checks (loop_vinfo, &cond_expr); |
86290011 | 2354 | |
d68d56b5 RG |
2355 | cond_expr = force_gimple_operand_1 (cond_expr, &gimplify_stmt_list, |
2356 | is_gimple_condexpr, NULL_TREE); | |
2357 | gimple_seq_add_seq (&cond_expr_stmt_list, gimplify_stmt_list); | |
ebfd146a IR |
2358 | |
2359 | initialize_original_copy_tables (); | |
5ce9450f JJ |
2360 | if (scalar_loop) |
2361 | { | |
2362 | edge scalar_e; | |
2363 | basic_block preheader, scalar_preheader; | |
2364 | ||
2365 | /* We don't want to scale SCALAR_LOOP's frequencies, we need to | |
2366 | scale LOOP's frequencies instead. */ | |
2367 | loop_version (scalar_loop, cond_expr, &condition_bb, | |
2368 | prob, REG_BR_PROB_BASE, REG_BR_PROB_BASE - prob, true); | |
2369 | scale_loop_frequencies (loop, prob, REG_BR_PROB_BASE); | |
2370 | /* CONDITION_BB was created above SCALAR_LOOP's preheader, | |
2371 | while we need to move it above LOOP's preheader. */ | |
2372 | e = loop_preheader_edge (loop); | |
2373 | scalar_e = loop_preheader_edge (scalar_loop); | |
2374 | gcc_assert (empty_block_p (e->src) | |
2375 | && single_pred_p (e->src)); | |
2376 | gcc_assert (empty_block_p (scalar_e->src) | |
2377 | && single_pred_p (scalar_e->src)); | |
2378 | gcc_assert (single_pred_p (condition_bb)); | |
2379 | preheader = e->src; | |
2380 | scalar_preheader = scalar_e->src; | |
2381 | scalar_e = find_edge (condition_bb, scalar_preheader); | |
2382 | e = single_pred_edge (preheader); | |
2383 | redirect_edge_and_branch_force (single_pred_edge (condition_bb), | |
2384 | scalar_preheader); | |
2385 | redirect_edge_and_branch_force (scalar_e, preheader); | |
2386 | redirect_edge_and_branch_force (e, condition_bb); | |
2387 | set_immediate_dominator (CDI_DOMINATORS, condition_bb, | |
2388 | single_pred (condition_bb)); | |
2389 | set_immediate_dominator (CDI_DOMINATORS, scalar_preheader, | |
2390 | single_pred (scalar_preheader)); | |
2391 | set_immediate_dominator (CDI_DOMINATORS, preheader, | |
2392 | condition_bb); | |
2393 | } | |
2394 | else | |
2395 | loop_version (loop, cond_expr, &condition_bb, | |
2396 | prob, prob, REG_BR_PROB_BASE - prob, true); | |
9cc1fb4b | 2397 | |
b05e0233 | 2398 | if (LOCATION_LOCUS (vect_location) != UNKNOWN_LOCATION |
9cc1fb4b XDL |
2399 | && dump_enabled_p ()) |
2400 | { | |
2401 | if (version_alias) | |
2402 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, vect_location, | |
2403 | "loop versioned for vectorization because of " | |
2404 | "possible aliasing\n"); | |
2405 | if (version_align) | |
2406 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, vect_location, | |
2407 | "loop versioned for vectorization to enhance " | |
2408 | "alignment\n"); | |
2409 | ||
2410 | } | |
c3284718 | 2411 | free_original_copy_tables (); |
ebfd146a | 2412 | |
b8698a0f | 2413 | /* Loop versioning violates an assumption we try to maintain during |
ebfd146a IR |
2414 | vectorization - that the loop exit block has a single predecessor. |
2415 | After versioning, the exit block of both loop versions is the same | |
2416 | basic block (i.e. it has two predecessors). Just in order to simplify | |
2417 | following transformations in the vectorizer, we fix this situation | |
2418 | here by adding a new (empty) block on the exit-edge of the loop, | |
5ce9450f JJ |
2419 | with the proper loop-exit phis to maintain loop-closed-form. |
2420 | If loop versioning wasn't done from loop, but scalar_loop instead, | |
2421 | merge_bb will have already just a single successor. */ | |
b8698a0f | 2422 | |
ebfd146a | 2423 | merge_bb = single_exit (loop)->dest; |
5ce9450f | 2424 | if (scalar_loop == NULL || EDGE_COUNT (merge_bb->preds) >= 2) |
ebfd146a | 2425 | { |
5ce9450f JJ |
2426 | gcc_assert (EDGE_COUNT (merge_bb->preds) >= 2); |
2427 | new_exit_bb = split_edge (single_exit (loop)); | |
2428 | new_exit_e = single_exit (loop); | |
2429 | e = EDGE_SUCC (new_exit_bb, 0); | |
2430 | ||
2431 | for (gsi = gsi_start_phis (merge_bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
2432 | { | |
2433 | tree new_res; | |
538dd0b7 | 2434 | orig_phi = gsi.phi (); |
b731b390 | 2435 | new_res = copy_ssa_name (PHI_RESULT (orig_phi)); |
5ce9450f JJ |
2436 | new_phi = create_phi_node (new_res, new_exit_bb); |
2437 | arg = PHI_ARG_DEF_FROM_EDGE (orig_phi, e); | |
2438 | add_phi_arg (new_phi, arg, new_exit_e, | |
2439 | gimple_phi_arg_location_from_edge (orig_phi, e)); | |
2440 | adjust_phi_and_debug_stmts (orig_phi, e, PHI_RESULT (new_phi)); | |
2441 | } | |
b8698a0f | 2442 | } |
ebfd146a IR |
2443 | |
2444 | /* End loop-exit-fixes after versioning. */ | |
2445 | ||
d68d56b5 | 2446 | if (cond_expr_stmt_list) |
ebfd146a IR |
2447 | { |
2448 | cond_exp_gsi = gsi_last_bb (condition_bb); | |
d68d56b5 | 2449 | gsi_insert_seq_before (&cond_exp_gsi, cond_expr_stmt_list, |
86290011 | 2450 | GSI_SAME_STMT); |
ebfd146a | 2451 | } |
90eb75f2 | 2452 | update_ssa (TODO_update_ssa); |
ebfd146a | 2453 | } |