]>
Commit | Line | Data |
---|---|---|
b8698a0f | 1 | /* Vectorizer Specific Loop Manipulations |
5d2eb24b | 2 | Copyright (C) 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010, 2012 |
7c028163 | 3 | Free Software Foundation, Inc. |
b8698a0f | 4 | Contributed by Dorit Naishlos <dorit@il.ibm.com> |
ebfd146a IR |
5 | and Ira Rosen <irar@il.ibm.com> |
6 | ||
7 | This file is part of GCC. | |
8 | ||
9 | GCC is free software; you can redistribute it and/or modify it under | |
10 | the terms of the GNU General Public License as published by the Free | |
11 | Software Foundation; either version 3, or (at your option) any later | |
12 | version. | |
13 | ||
14 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
15 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
16 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
17 | for more details. | |
18 | ||
19 | You should have received a copy of the GNU General Public License | |
20 | along with GCC; see the file COPYING3. If not see | |
21 | <http://www.gnu.org/licenses/>. */ | |
22 | ||
23 | #include "config.h" | |
24 | #include "system.h" | |
25 | #include "coretypes.h" | |
78c60e3d | 26 | #include "dumpfile.h" |
ebfd146a IR |
27 | #include "tm.h" |
28 | #include "ggc.h" | |
29 | #include "tree.h" | |
30 | #include "basic-block.h" | |
cf835838 | 31 | #include "gimple-pretty-print.h" |
ebfd146a | 32 | #include "tree-flow.h" |
7ee2468b | 33 | #include "tree-pass.h" |
ebfd146a | 34 | #include "cfgloop.h" |
718f9c0f | 35 | #include "diagnostic-core.h" |
ebfd146a IR |
36 | #include "tree-scalar-evolution.h" |
37 | #include "tree-vectorizer.h" | |
38 | #include "langhooks.h" | |
39 | ||
40 | /************************************************************************* | |
41 | Simple Loop Peeling Utilities | |
42 | ||
43 | Utilities to support loop peeling for vectorization purposes. | |
44 | *************************************************************************/ | |
45 | ||
46 | ||
47 | /* Renames the use *OP_P. */ | |
48 | ||
49 | static void | |
50 | rename_use_op (use_operand_p op_p) | |
51 | { | |
52 | tree new_name; | |
53 | ||
54 | if (TREE_CODE (USE_FROM_PTR (op_p)) != SSA_NAME) | |
55 | return; | |
56 | ||
57 | new_name = get_current_def (USE_FROM_PTR (op_p)); | |
58 | ||
59 | /* Something defined outside of the loop. */ | |
60 | if (!new_name) | |
61 | return; | |
62 | ||
63 | /* An ordinary ssa name defined in the loop. */ | |
64 | ||
65 | SET_USE (op_p, new_name); | |
66 | } | |
67 | ||
68 | ||
69 | /* Renames the variables in basic block BB. */ | |
70 | ||
71 | void | |
72 | rename_variables_in_bb (basic_block bb) | |
73 | { | |
74 | gimple_stmt_iterator gsi; | |
75 | gimple stmt; | |
76 | use_operand_p use_p; | |
77 | ssa_op_iter iter; | |
78 | edge e; | |
79 | edge_iterator ei; | |
80 | struct loop *loop = bb->loop_father; | |
81 | ||
82 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
83 | { | |
84 | stmt = gsi_stmt (gsi); | |
85 | FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_ALL_USES) | |
86 | rename_use_op (use_p); | |
87 | } | |
88 | ||
89 | FOR_EACH_EDGE (e, ei, bb->succs) | |
90 | { | |
91 | if (!flow_bb_inside_loop_p (loop, e->dest)) | |
92 | continue; | |
93 | for (gsi = gsi_start_phis (e->dest); !gsi_end_p (gsi); gsi_next (&gsi)) | |
94 | rename_use_op (PHI_ARG_DEF_PTR_FROM_EDGE (gsi_stmt (gsi), e)); | |
95 | } | |
96 | } | |
97 | ||
98 | ||
99 | /* Renames variables in new generated LOOP. */ | |
100 | ||
101 | void | |
102 | rename_variables_in_loop (struct loop *loop) | |
103 | { | |
104 | unsigned i; | |
105 | basic_block *bbs; | |
106 | ||
107 | bbs = get_loop_body (loop); | |
108 | ||
109 | for (i = 0; i < loop->num_nodes; i++) | |
110 | rename_variables_in_bb (bbs[i]); | |
111 | ||
112 | free (bbs); | |
113 | } | |
114 | ||
684f25f4 AO |
115 | typedef struct |
116 | { | |
117 | tree from, to; | |
118 | basic_block bb; | |
119 | } adjust_info; | |
120 | ||
121 | DEF_VEC_O(adjust_info); | |
122 | DEF_VEC_ALLOC_O_STACK(adjust_info); | |
123 | #define VEC_adjust_info_stack_alloc(alloc) VEC_stack_alloc (adjust_info, alloc) | |
124 | ||
125 | /* A stack of values to be adjusted in debug stmts. We have to | |
126 | process them LIFO, so that the closest substitution applies. If we | |
127 | processed them FIFO, without the stack, we might substitute uses | |
128 | with a PHI DEF that would soon become non-dominant, and when we got | |
129 | to the suitable one, it wouldn't have anything to substitute any | |
130 | more. */ | |
131 | static VEC(adjust_info, stack) *adjust_vec; | |
132 | ||
133 | /* Adjust any debug stmts that referenced AI->from values to use the | |
134 | loop-closed AI->to, if the references are dominated by AI->bb and | |
135 | not by the definition of AI->from. */ | |
136 | ||
137 | static void | |
138 | adjust_debug_stmts_now (adjust_info *ai) | |
139 | { | |
140 | basic_block bbphi = ai->bb; | |
141 | tree orig_def = ai->from; | |
142 | tree new_def = ai->to; | |
143 | imm_use_iterator imm_iter; | |
144 | gimple stmt; | |
145 | basic_block bbdef = gimple_bb (SSA_NAME_DEF_STMT (orig_def)); | |
146 | ||
147 | gcc_assert (dom_info_available_p (CDI_DOMINATORS)); | |
148 | ||
149 | /* Adjust any debug stmts that held onto non-loop-closed | |
150 | references. */ | |
151 | FOR_EACH_IMM_USE_STMT (stmt, imm_iter, orig_def) | |
152 | { | |
153 | use_operand_p use_p; | |
154 | basic_block bbuse; | |
155 | ||
156 | if (!is_gimple_debug (stmt)) | |
157 | continue; | |
158 | ||
159 | gcc_assert (gimple_debug_bind_p (stmt)); | |
160 | ||
161 | bbuse = gimple_bb (stmt); | |
162 | ||
163 | if ((bbuse == bbphi | |
164 | || dominated_by_p (CDI_DOMINATORS, bbuse, bbphi)) | |
165 | && !(bbuse == bbdef | |
166 | || dominated_by_p (CDI_DOMINATORS, bbuse, bbdef))) | |
167 | { | |
168 | if (new_def) | |
169 | FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter) | |
170 | SET_USE (use_p, new_def); | |
171 | else | |
172 | { | |
173 | gimple_debug_bind_reset_value (stmt); | |
174 | update_stmt (stmt); | |
175 | } | |
176 | } | |
177 | } | |
178 | } | |
179 | ||
180 | /* Adjust debug stmts as scheduled before. */ | |
181 | ||
182 | static void | |
183 | adjust_vec_debug_stmts (void) | |
184 | { | |
185 | if (!MAY_HAVE_DEBUG_STMTS) | |
186 | return; | |
187 | ||
188 | gcc_assert (adjust_vec); | |
189 | ||
190 | while (!VEC_empty (adjust_info, adjust_vec)) | |
191 | { | |
0823efed | 192 | adjust_debug_stmts_now (&VEC_last (adjust_info, adjust_vec)); |
684f25f4 AO |
193 | VEC_pop (adjust_info, adjust_vec); |
194 | } | |
195 | ||
196 | VEC_free (adjust_info, stack, adjust_vec); | |
197 | } | |
198 | ||
199 | /* Adjust any debug stmts that referenced FROM values to use the | |
200 | loop-closed TO, if the references are dominated by BB and not by | |
201 | the definition of FROM. If adjust_vec is non-NULL, adjustments | |
202 | will be postponed until adjust_vec_debug_stmts is called. */ | |
203 | ||
204 | static void | |
205 | adjust_debug_stmts (tree from, tree to, basic_block bb) | |
206 | { | |
207 | adjust_info ai; | |
208 | ||
a471762f RG |
209 | if (MAY_HAVE_DEBUG_STMTS |
210 | && TREE_CODE (from) == SSA_NAME | |
211 | && ! virtual_operand_p (from)) | |
684f25f4 AO |
212 | { |
213 | ai.from = from; | |
214 | ai.to = to; | |
215 | ai.bb = bb; | |
216 | ||
217 | if (adjust_vec) | |
f32682ca | 218 | VEC_safe_push (adjust_info, stack, adjust_vec, ai); |
684f25f4 AO |
219 | else |
220 | adjust_debug_stmts_now (&ai); | |
221 | } | |
222 | } | |
223 | ||
224 | /* Change E's phi arg in UPDATE_PHI to NEW_DEF, and record information | |
225 | to adjust any debug stmts that referenced the old phi arg, | |
226 | presumably non-loop-closed references left over from other | |
227 | transformations. */ | |
228 | ||
229 | static void | |
230 | adjust_phi_and_debug_stmts (gimple update_phi, edge e, tree new_def) | |
231 | { | |
232 | tree orig_def = PHI_ARG_DEF_FROM_EDGE (update_phi, e); | |
233 | ||
234 | SET_PHI_ARG_DEF (update_phi, e->dest_idx, new_def); | |
235 | ||
236 | if (MAY_HAVE_DEBUG_STMTS) | |
237 | adjust_debug_stmts (orig_def, PHI_RESULT (update_phi), | |
238 | gimple_bb (update_phi)); | |
239 | } | |
240 | ||
ebfd146a IR |
241 | |
242 | /* Update the PHI nodes of NEW_LOOP. | |
243 | ||
244 | NEW_LOOP is a duplicate of ORIG_LOOP. | |
245 | AFTER indicates whether NEW_LOOP executes before or after ORIG_LOOP: | |
246 | AFTER is true if NEW_LOOP executes after ORIG_LOOP, and false if it | |
247 | executes before it. */ | |
248 | ||
249 | static void | |
250 | slpeel_update_phis_for_duplicate_loop (struct loop *orig_loop, | |
251 | struct loop *new_loop, bool after) | |
252 | { | |
253 | tree new_ssa_name; | |
254 | gimple phi_new, phi_orig; | |
255 | tree def; | |
256 | edge orig_loop_latch = loop_latch_edge (orig_loop); | |
257 | edge orig_entry_e = loop_preheader_edge (orig_loop); | |
258 | edge new_loop_exit_e = single_exit (new_loop); | |
259 | edge new_loop_entry_e = loop_preheader_edge (new_loop); | |
260 | edge entry_arg_e = (after ? orig_loop_latch : orig_entry_e); | |
261 | gimple_stmt_iterator gsi_new, gsi_orig; | |
262 | ||
263 | /* | |
264 | step 1. For each loop-header-phi: | |
265 | Add the first phi argument for the phi in NEW_LOOP | |
266 | (the one associated with the entry of NEW_LOOP) | |
267 | ||
268 | step 2. For each loop-header-phi: | |
269 | Add the second phi argument for the phi in NEW_LOOP | |
270 | (the one associated with the latch of NEW_LOOP) | |
271 | ||
272 | step 3. Update the phis in the successor block of NEW_LOOP. | |
273 | ||
274 | case 1: NEW_LOOP was placed before ORIG_LOOP: | |
275 | The successor block of NEW_LOOP is the header of ORIG_LOOP. | |
276 | Updating the phis in the successor block can therefore be done | |
277 | along with the scanning of the loop header phis, because the | |
278 | header blocks of ORIG_LOOP and NEW_LOOP have exactly the same | |
279 | phi nodes, organized in the same order. | |
280 | ||
281 | case 2: NEW_LOOP was placed after ORIG_LOOP: | |
b8698a0f | 282 | The successor block of NEW_LOOP is the original exit block of |
ebfd146a IR |
283 | ORIG_LOOP - the phis to be updated are the loop-closed-ssa phis. |
284 | We postpone updating these phis to a later stage (when | |
285 | loop guards are added). | |
286 | */ | |
287 | ||
288 | ||
289 | /* Scan the phis in the headers of the old and new loops | |
290 | (they are organized in exactly the same order). */ | |
291 | ||
292 | for (gsi_new = gsi_start_phis (new_loop->header), | |
293 | gsi_orig = gsi_start_phis (orig_loop->header); | |
294 | !gsi_end_p (gsi_new) && !gsi_end_p (gsi_orig); | |
295 | gsi_next (&gsi_new), gsi_next (&gsi_orig)) | |
296 | { | |
f5045c96 | 297 | source_location locus; |
ebfd146a IR |
298 | phi_new = gsi_stmt (gsi_new); |
299 | phi_orig = gsi_stmt (gsi_orig); | |
300 | ||
301 | /* step 1. */ | |
302 | def = PHI_ARG_DEF_FROM_EDGE (phi_orig, entry_arg_e); | |
f5045c96 | 303 | locus = gimple_phi_arg_location_from_edge (phi_orig, entry_arg_e); |
9e227d60 | 304 | add_phi_arg (phi_new, def, new_loop_entry_e, locus); |
ebfd146a IR |
305 | |
306 | /* step 2. */ | |
307 | def = PHI_ARG_DEF_FROM_EDGE (phi_orig, orig_loop_latch); | |
f5045c96 | 308 | locus = gimple_phi_arg_location_from_edge (phi_orig, orig_loop_latch); |
ebfd146a IR |
309 | if (TREE_CODE (def) != SSA_NAME) |
310 | continue; | |
311 | ||
312 | new_ssa_name = get_current_def (def); | |
313 | if (!new_ssa_name) | |
314 | { | |
315 | /* This only happens if there are no definitions | |
316 | inside the loop. use the phi_result in this case. */ | |
317 | new_ssa_name = PHI_RESULT (phi_new); | |
318 | } | |
319 | ||
320 | /* An ordinary ssa name defined in the loop. */ | |
9e227d60 | 321 | add_phi_arg (phi_new, new_ssa_name, loop_latch_edge (new_loop), locus); |
ebfd146a | 322 | |
684f25f4 AO |
323 | /* Drop any debug references outside the loop, if they would |
324 | become ill-formed SSA. */ | |
325 | adjust_debug_stmts (def, NULL, single_exit (orig_loop)->dest); | |
326 | ||
ebfd146a IR |
327 | /* step 3 (case 1). */ |
328 | if (!after) | |
329 | { | |
330 | gcc_assert (new_loop_exit_e == orig_entry_e); | |
684f25f4 | 331 | adjust_phi_and_debug_stmts (phi_orig, new_loop_exit_e, new_ssa_name); |
ebfd146a IR |
332 | } |
333 | } | |
334 | } | |
335 | ||
336 | ||
337 | /* Update PHI nodes for a guard of the LOOP. | |
338 | ||
339 | Input: | |
340 | - LOOP, GUARD_EDGE: LOOP is a loop for which we added guard code that | |
341 | controls whether LOOP is to be executed. GUARD_EDGE is the edge that | |
342 | originates from the guard-bb, skips LOOP and reaches the (unique) exit | |
343 | bb of LOOP. This loop-exit-bb is an empty bb with one successor. | |
344 | We denote this bb NEW_MERGE_BB because before the guard code was added | |
345 | it had a single predecessor (the LOOP header), and now it became a merge | |
346 | point of two paths - the path that ends with the LOOP exit-edge, and | |
347 | the path that ends with GUARD_EDGE. | |
348 | - NEW_EXIT_BB: New basic block that is added by this function between LOOP | |
349 | and NEW_MERGE_BB. It is used to place loop-closed-ssa-form exit-phis. | |
350 | ||
351 | ===> The CFG before the guard-code was added: | |
352 | LOOP_header_bb: | |
353 | loop_body | |
354 | if (exit_loop) goto update_bb | |
355 | else goto LOOP_header_bb | |
356 | update_bb: | |
357 | ||
358 | ==> The CFG after the guard-code was added: | |
359 | guard_bb: | |
360 | if (LOOP_guard_condition) goto new_merge_bb | |
361 | else goto LOOP_header_bb | |
362 | LOOP_header_bb: | |
363 | loop_body | |
364 | if (exit_loop_condition) goto new_merge_bb | |
365 | else goto LOOP_header_bb | |
366 | new_merge_bb: | |
367 | goto update_bb | |
368 | update_bb: | |
369 | ||
370 | ==> The CFG after this function: | |
371 | guard_bb: | |
372 | if (LOOP_guard_condition) goto new_merge_bb | |
373 | else goto LOOP_header_bb | |
374 | LOOP_header_bb: | |
375 | loop_body | |
376 | if (exit_loop_condition) goto new_exit_bb | |
377 | else goto LOOP_header_bb | |
378 | new_exit_bb: | |
379 | new_merge_bb: | |
380 | goto update_bb | |
381 | update_bb: | |
382 | ||
383 | This function: | |
384 | 1. creates and updates the relevant phi nodes to account for the new | |
385 | incoming edge (GUARD_EDGE) into NEW_MERGE_BB. This involves: | |
386 | 1.1. Create phi nodes at NEW_MERGE_BB. | |
387 | 1.2. Update the phi nodes at the successor of NEW_MERGE_BB (denoted | |
388 | UPDATE_BB). UPDATE_BB was the exit-bb of LOOP before NEW_MERGE_BB | |
389 | 2. preserves loop-closed-ssa-form by creating the required phi nodes | |
390 | at the exit of LOOP (i.e, in NEW_EXIT_BB). | |
391 | ||
392 | There are two flavors to this function: | |
393 | ||
394 | slpeel_update_phi_nodes_for_guard1: | |
395 | Here the guard controls whether we enter or skip LOOP, where LOOP is a | |
396 | prolog_loop (loop1 below), and the new phis created in NEW_MERGE_BB are | |
397 | for variables that have phis in the loop header. | |
398 | ||
399 | slpeel_update_phi_nodes_for_guard2: | |
400 | Here the guard controls whether we enter or skip LOOP, where LOOP is an | |
401 | epilog_loop (loop2 below), and the new phis created in NEW_MERGE_BB are | |
402 | for variables that have phis in the loop exit. | |
403 | ||
404 | I.E., the overall structure is: | |
405 | ||
406 | loop1_preheader_bb: | |
407 | guard1 (goto loop1/merge1_bb) | |
408 | loop1 | |
409 | loop1_exit_bb: | |
410 | guard2 (goto merge1_bb/merge2_bb) | |
411 | merge1_bb | |
412 | loop2 | |
413 | loop2_exit_bb | |
414 | merge2_bb | |
415 | next_bb | |
416 | ||
417 | slpeel_update_phi_nodes_for_guard1 takes care of creating phis in | |
418 | loop1_exit_bb and merge1_bb. These are entry phis (phis for the vars | |
419 | that have phis in loop1->header). | |
420 | ||
421 | slpeel_update_phi_nodes_for_guard2 takes care of creating phis in | |
422 | loop2_exit_bb and merge2_bb. These are exit phis (phis for the vars | |
423 | that have phis in next_bb). It also adds some of these phis to | |
424 | loop1_exit_bb. | |
425 | ||
426 | slpeel_update_phi_nodes_for_guard1 is always called before | |
427 | slpeel_update_phi_nodes_for_guard2. They are both needed in order | |
428 | to create correct data-flow and loop-closed-ssa-form. | |
429 | ||
430 | Generally slpeel_update_phi_nodes_for_guard1 creates phis for variables | |
431 | that change between iterations of a loop (and therefore have a phi-node | |
432 | at the loop entry), whereas slpeel_update_phi_nodes_for_guard2 creates | |
b8698a0f L |
433 | phis for variables that are used out of the loop (and therefore have |
434 | loop-closed exit phis). Some variables may be both updated between | |
ebfd146a IR |
435 | iterations and used after the loop. This is why in loop1_exit_bb we |
436 | may need both entry_phis (created by slpeel_update_phi_nodes_for_guard1) | |
437 | and exit phis (created by slpeel_update_phi_nodes_for_guard2). | |
438 | ||
439 | - IS_NEW_LOOP: if IS_NEW_LOOP is true, then LOOP is a newly created copy of | |
440 | an original loop. i.e., we have: | |
441 | ||
442 | orig_loop | |
443 | guard_bb (goto LOOP/new_merge) | |
444 | new_loop <-- LOOP | |
445 | new_exit | |
446 | new_merge | |
447 | next_bb | |
448 | ||
449 | If IS_NEW_LOOP is false, then LOOP is an original loop, in which case we | |
450 | have: | |
451 | ||
452 | new_loop | |
453 | guard_bb (goto LOOP/new_merge) | |
454 | orig_loop <-- LOOP | |
455 | new_exit | |
456 | new_merge | |
457 | next_bb | |
458 | ||
459 | The SSA names defined in the original loop have a current | |
460 | reaching definition that that records the corresponding new | |
461 | ssa-name used in the new duplicated loop copy. | |
462 | */ | |
463 | ||
464 | /* Function slpeel_update_phi_nodes_for_guard1 | |
b8698a0f | 465 | |
ebfd146a IR |
466 | Input: |
467 | - GUARD_EDGE, LOOP, IS_NEW_LOOP, NEW_EXIT_BB - as explained above. | |
468 | - DEFS - a bitmap of ssa names to mark new names for which we recorded | |
b8698a0f L |
469 | information. |
470 | ||
ebfd146a IR |
471 | In the context of the overall structure, we have: |
472 | ||
b8698a0f | 473 | loop1_preheader_bb: |
ebfd146a IR |
474 | guard1 (goto loop1/merge1_bb) |
475 | LOOP-> loop1 | |
476 | loop1_exit_bb: | |
477 | guard2 (goto merge1_bb/merge2_bb) | |
478 | merge1_bb | |
479 | loop2 | |
480 | loop2_exit_bb | |
481 | merge2_bb | |
482 | next_bb | |
483 | ||
484 | For each name updated between loop iterations (i.e - for each name that has | |
485 | an entry (loop-header) phi in LOOP) we create a new phi in: | |
486 | 1. merge1_bb (to account for the edge from guard1) | |
487 | 2. loop1_exit_bb (an exit-phi to keep LOOP in loop-closed form) | |
488 | */ | |
489 | ||
490 | static void | |
491 | slpeel_update_phi_nodes_for_guard1 (edge guard_edge, struct loop *loop, | |
c334023f | 492 | bool is_new_loop, basic_block *new_exit_bb) |
ebfd146a IR |
493 | { |
494 | gimple orig_phi, new_phi; | |
495 | gimple update_phi, update_phi2; | |
496 | tree guard_arg, loop_arg; | |
497 | basic_block new_merge_bb = guard_edge->dest; | |
498 | edge e = EDGE_SUCC (new_merge_bb, 0); | |
499 | basic_block update_bb = e->dest; | |
500 | basic_block orig_bb = loop->header; | |
501 | edge new_exit_e; | |
502 | tree current_new_name; | |
ebfd146a IR |
503 | gimple_stmt_iterator gsi_orig, gsi_update; |
504 | ||
505 | /* Create new bb between loop and new_merge_bb. */ | |
506 | *new_exit_bb = split_edge (single_exit (loop)); | |
507 | ||
508 | new_exit_e = EDGE_SUCC (*new_exit_bb, 0); | |
509 | ||
510 | for (gsi_orig = gsi_start_phis (orig_bb), | |
511 | gsi_update = gsi_start_phis (update_bb); | |
512 | !gsi_end_p (gsi_orig) && !gsi_end_p (gsi_update); | |
513 | gsi_next (&gsi_orig), gsi_next (&gsi_update)) | |
514 | { | |
e20f6b4b | 515 | source_location loop_locus, guard_locus; |
070ecdfd | 516 | tree new_res; |
ebfd146a IR |
517 | orig_phi = gsi_stmt (gsi_orig); |
518 | update_phi = gsi_stmt (gsi_update); | |
519 | ||
ebfd146a IR |
520 | /** 1. Handle new-merge-point phis **/ |
521 | ||
522 | /* 1.1. Generate new phi node in NEW_MERGE_BB: */ | |
070ecdfd RG |
523 | new_res = copy_ssa_name (PHI_RESULT (orig_phi), NULL); |
524 | new_phi = create_phi_node (new_res, new_merge_bb); | |
ebfd146a IR |
525 | |
526 | /* 1.2. NEW_MERGE_BB has two incoming edges: GUARD_EDGE and the exit-edge | |
527 | of LOOP. Set the two phi args in NEW_PHI for these edges: */ | |
528 | loop_arg = PHI_ARG_DEF_FROM_EDGE (orig_phi, EDGE_SUCC (loop->latch, 0)); | |
b8698a0f L |
529 | loop_locus = gimple_phi_arg_location_from_edge (orig_phi, |
530 | EDGE_SUCC (loop->latch, | |
f5045c96 | 531 | 0)); |
ebfd146a | 532 | guard_arg = PHI_ARG_DEF_FROM_EDGE (orig_phi, loop_preheader_edge (loop)); |
b8698a0f L |
533 | guard_locus |
534 | = gimple_phi_arg_location_from_edge (orig_phi, | |
f5045c96 | 535 | loop_preheader_edge (loop)); |
ebfd146a | 536 | |
9e227d60 DC |
537 | add_phi_arg (new_phi, loop_arg, new_exit_e, loop_locus); |
538 | add_phi_arg (new_phi, guard_arg, guard_edge, guard_locus); | |
ebfd146a IR |
539 | |
540 | /* 1.3. Update phi in successor block. */ | |
541 | gcc_assert (PHI_ARG_DEF_FROM_EDGE (update_phi, e) == loop_arg | |
542 | || PHI_ARG_DEF_FROM_EDGE (update_phi, e) == guard_arg); | |
684f25f4 | 543 | adjust_phi_and_debug_stmts (update_phi, e, PHI_RESULT (new_phi)); |
ebfd146a IR |
544 | update_phi2 = new_phi; |
545 | ||
546 | ||
547 | /** 2. Handle loop-closed-ssa-form phis **/ | |
548 | ||
ea057359 | 549 | if (virtual_operand_p (PHI_RESULT (orig_phi))) |
ebfd146a IR |
550 | continue; |
551 | ||
552 | /* 2.1. Generate new phi node in NEW_EXIT_BB: */ | |
070ecdfd RG |
553 | new_res = copy_ssa_name (PHI_RESULT (orig_phi), NULL); |
554 | new_phi = create_phi_node (new_res, *new_exit_bb); | |
ebfd146a IR |
555 | |
556 | /* 2.2. NEW_EXIT_BB has one incoming edge: the exit-edge of the loop. */ | |
9e227d60 | 557 | add_phi_arg (new_phi, loop_arg, single_exit (loop), loop_locus); |
ebfd146a IR |
558 | |
559 | /* 2.3. Update phi in successor of NEW_EXIT_BB: */ | |
560 | gcc_assert (PHI_ARG_DEF_FROM_EDGE (update_phi2, new_exit_e) == loop_arg); | |
684f25f4 AO |
561 | adjust_phi_and_debug_stmts (update_phi2, new_exit_e, |
562 | PHI_RESULT (new_phi)); | |
ebfd146a IR |
563 | |
564 | /* 2.4. Record the newly created name with set_current_def. | |
565 | We want to find a name such that | |
566 | name = get_current_def (orig_loop_name) | |
567 | and to set its current definition as follows: | |
568 | set_current_def (name, new_phi_name) | |
569 | ||
570 | If LOOP is a new loop then loop_arg is already the name we're | |
571 | looking for. If LOOP is the original loop, then loop_arg is | |
572 | the orig_loop_name and the relevant name is recorded in its | |
573 | current reaching definition. */ | |
574 | if (is_new_loop) | |
575 | current_new_name = loop_arg; | |
576 | else | |
577 | { | |
578 | current_new_name = get_current_def (loop_arg); | |
579 | /* current_def is not available only if the variable does not | |
580 | change inside the loop, in which case we also don't care | |
581 | about recording a current_def for it because we won't be | |
582 | trying to create loop-exit-phis for it. */ | |
583 | if (!current_new_name) | |
584 | continue; | |
585 | } | |
586 | gcc_assert (get_current_def (current_new_name) == NULL_TREE); | |
587 | ||
588 | set_current_def (current_new_name, PHI_RESULT (new_phi)); | |
ebfd146a IR |
589 | } |
590 | } | |
591 | ||
592 | ||
593 | /* Function slpeel_update_phi_nodes_for_guard2 | |
594 | ||
595 | Input: | |
596 | - GUARD_EDGE, LOOP, IS_NEW_LOOP, NEW_EXIT_BB - as explained above. | |
597 | ||
598 | In the context of the overall structure, we have: | |
599 | ||
b8698a0f | 600 | loop1_preheader_bb: |
ebfd146a IR |
601 | guard1 (goto loop1/merge1_bb) |
602 | loop1 | |
b8698a0f | 603 | loop1_exit_bb: |
ebfd146a IR |
604 | guard2 (goto merge1_bb/merge2_bb) |
605 | merge1_bb | |
606 | LOOP-> loop2 | |
607 | loop2_exit_bb | |
608 | merge2_bb | |
609 | next_bb | |
610 | ||
611 | For each name used out side the loop (i.e - for each name that has an exit | |
612 | phi in next_bb) we create a new phi in: | |
b8698a0f | 613 | 1. merge2_bb (to account for the edge from guard_bb) |
ebfd146a IR |
614 | 2. loop2_exit_bb (an exit-phi to keep LOOP in loop-closed form) |
615 | 3. guard2 bb (an exit phi to keep the preceding loop in loop-closed form), | |
616 | if needed (if it wasn't handled by slpeel_update_phis_nodes_for_phi1). | |
617 | */ | |
618 | ||
619 | static void | |
620 | slpeel_update_phi_nodes_for_guard2 (edge guard_edge, struct loop *loop, | |
621 | bool is_new_loop, basic_block *new_exit_bb) | |
622 | { | |
623 | gimple orig_phi, new_phi; | |
624 | gimple update_phi, update_phi2; | |
625 | tree guard_arg, loop_arg; | |
626 | basic_block new_merge_bb = guard_edge->dest; | |
627 | edge e = EDGE_SUCC (new_merge_bb, 0); | |
628 | basic_block update_bb = e->dest; | |
629 | edge new_exit_e; | |
630 | tree orig_def, orig_def_new_name; | |
631 | tree new_name, new_name2; | |
632 | tree arg; | |
633 | gimple_stmt_iterator gsi; | |
634 | ||
635 | /* Create new bb between loop and new_merge_bb. */ | |
636 | *new_exit_bb = split_edge (single_exit (loop)); | |
637 | ||
638 | new_exit_e = EDGE_SUCC (*new_exit_bb, 0); | |
639 | ||
640 | for (gsi = gsi_start_phis (update_bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
641 | { | |
070ecdfd | 642 | tree new_res; |
ebfd146a IR |
643 | update_phi = gsi_stmt (gsi); |
644 | orig_phi = update_phi; | |
645 | orig_def = PHI_ARG_DEF_FROM_EDGE (orig_phi, e); | |
646 | /* This loop-closed-phi actually doesn't represent a use | |
b8698a0f | 647 | out of the loop - the phi arg is a constant. */ |
ebfd146a IR |
648 | if (TREE_CODE (orig_def) != SSA_NAME) |
649 | continue; | |
650 | orig_def_new_name = get_current_def (orig_def); | |
651 | arg = NULL_TREE; | |
652 | ||
653 | /** 1. Handle new-merge-point phis **/ | |
654 | ||
655 | /* 1.1. Generate new phi node in NEW_MERGE_BB: */ | |
070ecdfd RG |
656 | new_res = copy_ssa_name (PHI_RESULT (orig_phi), NULL); |
657 | new_phi = create_phi_node (new_res, new_merge_bb); | |
ebfd146a IR |
658 | |
659 | /* 1.2. NEW_MERGE_BB has two incoming edges: GUARD_EDGE and the exit-edge | |
660 | of LOOP. Set the two PHI args in NEW_PHI for these edges: */ | |
661 | new_name = orig_def; | |
662 | new_name2 = NULL_TREE; | |
663 | if (orig_def_new_name) | |
664 | { | |
665 | new_name = orig_def_new_name; | |
666 | /* Some variables have both loop-entry-phis and loop-exit-phis. | |
667 | Such variables were given yet newer names by phis placed in | |
668 | guard_bb by slpeel_update_phi_nodes_for_guard1. I.e: | |
669 | new_name2 = get_current_def (get_current_def (orig_name)). */ | |
670 | new_name2 = get_current_def (new_name); | |
671 | } | |
b8698a0f | 672 | |
ebfd146a IR |
673 | if (is_new_loop) |
674 | { | |
675 | guard_arg = orig_def; | |
676 | loop_arg = new_name; | |
677 | } | |
678 | else | |
679 | { | |
680 | guard_arg = new_name; | |
681 | loop_arg = orig_def; | |
682 | } | |
683 | if (new_name2) | |
684 | guard_arg = new_name2; | |
b8698a0f | 685 | |
9e227d60 DC |
686 | add_phi_arg (new_phi, loop_arg, new_exit_e, UNKNOWN_LOCATION); |
687 | add_phi_arg (new_phi, guard_arg, guard_edge, UNKNOWN_LOCATION); | |
ebfd146a IR |
688 | |
689 | /* 1.3. Update phi in successor block. */ | |
690 | gcc_assert (PHI_ARG_DEF_FROM_EDGE (update_phi, e) == orig_def); | |
684f25f4 | 691 | adjust_phi_and_debug_stmts (update_phi, e, PHI_RESULT (new_phi)); |
ebfd146a IR |
692 | update_phi2 = new_phi; |
693 | ||
694 | ||
695 | /** 2. Handle loop-closed-ssa-form phis **/ | |
696 | ||
697 | /* 2.1. Generate new phi node in NEW_EXIT_BB: */ | |
070ecdfd RG |
698 | new_res = copy_ssa_name (PHI_RESULT (orig_phi), NULL); |
699 | new_phi = create_phi_node (new_res, *new_exit_bb); | |
ebfd146a IR |
700 | |
701 | /* 2.2. NEW_EXIT_BB has one incoming edge: the exit-edge of the loop. */ | |
9e227d60 | 702 | add_phi_arg (new_phi, loop_arg, single_exit (loop), UNKNOWN_LOCATION); |
ebfd146a IR |
703 | |
704 | /* 2.3. Update phi in successor of NEW_EXIT_BB: */ | |
705 | gcc_assert (PHI_ARG_DEF_FROM_EDGE (update_phi2, new_exit_e) == loop_arg); | |
684f25f4 AO |
706 | adjust_phi_and_debug_stmts (update_phi2, new_exit_e, |
707 | PHI_RESULT (new_phi)); | |
ebfd146a IR |
708 | |
709 | ||
710 | /** 3. Handle loop-closed-ssa-form phis for first loop **/ | |
711 | ||
712 | /* 3.1. Find the relevant names that need an exit-phi in | |
713 | GUARD_BB, i.e. names for which | |
714 | slpeel_update_phi_nodes_for_guard1 had not already created a | |
715 | phi node. This is the case for names that are used outside | |
716 | the loop (and therefore need an exit phi) but are not updated | |
717 | across loop iterations (and therefore don't have a | |
718 | loop-header-phi). | |
719 | ||
720 | slpeel_update_phi_nodes_for_guard1 is responsible for | |
721 | creating loop-exit phis in GUARD_BB for names that have a | |
722 | loop-header-phi. When such a phi is created we also record | |
723 | the new name in its current definition. If this new name | |
724 | exists, then guard_arg was set to this new name (see 1.2 | |
725 | above). Therefore, if guard_arg is not this new name, this | |
726 | is an indication that an exit-phi in GUARD_BB was not yet | |
727 | created, so we take care of it here. */ | |
728 | if (guard_arg == new_name2) | |
729 | continue; | |
730 | arg = guard_arg; | |
731 | ||
732 | /* 3.2. Generate new phi node in GUARD_BB: */ | |
070ecdfd RG |
733 | new_res = copy_ssa_name (PHI_RESULT (orig_phi), NULL); |
734 | new_phi = create_phi_node (new_res, guard_edge->src); | |
ebfd146a IR |
735 | |
736 | /* 3.3. GUARD_BB has one incoming edge: */ | |
737 | gcc_assert (EDGE_COUNT (guard_edge->src->preds) == 1); | |
b8698a0f | 738 | add_phi_arg (new_phi, arg, EDGE_PRED (guard_edge->src, 0), |
9e227d60 | 739 | UNKNOWN_LOCATION); |
ebfd146a IR |
740 | |
741 | /* 3.4. Update phi in successor of GUARD_BB: */ | |
742 | gcc_assert (PHI_ARG_DEF_FROM_EDGE (update_phi2, guard_edge) | |
743 | == guard_arg); | |
684f25f4 AO |
744 | adjust_phi_and_debug_stmts (update_phi2, guard_edge, |
745 | PHI_RESULT (new_phi)); | |
ebfd146a IR |
746 | } |
747 | } | |
748 | ||
749 | ||
750 | /* Make the LOOP iterate NITERS times. This is done by adding a new IV | |
751 | that starts at zero, increases by one and its limit is NITERS. | |
752 | ||
753 | Assumption: the exit-condition of LOOP is the last stmt in the loop. */ | |
754 | ||
755 | void | |
756 | slpeel_make_loop_iterate_ntimes (struct loop *loop, tree niters) | |
757 | { | |
758 | tree indx_before_incr, indx_after_incr; | |
759 | gimple cond_stmt; | |
760 | gimple orig_cond; | |
761 | edge exit_edge = single_exit (loop); | |
762 | gimple_stmt_iterator loop_cond_gsi; | |
763 | gimple_stmt_iterator incr_gsi; | |
764 | bool insert_after; | |
765 | tree init = build_int_cst (TREE_TYPE (niters), 0); | |
766 | tree step = build_int_cst (TREE_TYPE (niters), 1); | |
767 | LOC loop_loc; | |
768 | enum tree_code code; | |
769 | ||
770 | orig_cond = get_loop_exit_condition (loop); | |
771 | gcc_assert (orig_cond); | |
772 | loop_cond_gsi = gsi_for_stmt (orig_cond); | |
773 | ||
774 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
775 | create_iv (init, step, NULL_TREE, loop, | |
776 | &incr_gsi, insert_after, &indx_before_incr, &indx_after_incr); | |
777 | ||
778 | indx_after_incr = force_gimple_operand_gsi (&loop_cond_gsi, indx_after_incr, | |
779 | true, NULL_TREE, true, | |
780 | GSI_SAME_STMT); | |
781 | niters = force_gimple_operand_gsi (&loop_cond_gsi, niters, true, NULL_TREE, | |
782 | true, GSI_SAME_STMT); | |
783 | ||
784 | code = (exit_edge->flags & EDGE_TRUE_VALUE) ? GE_EXPR : LT_EXPR; | |
785 | cond_stmt = gimple_build_cond (code, indx_after_incr, niters, NULL_TREE, | |
786 | NULL_TREE); | |
787 | ||
788 | gsi_insert_before (&loop_cond_gsi, cond_stmt, GSI_SAME_STMT); | |
789 | ||
790 | /* Remove old loop exit test: */ | |
791 | gsi_remove (&loop_cond_gsi, true); | |
6f723d33 | 792 | free_stmt_vec_info (orig_cond); |
ebfd146a IR |
793 | |
794 | loop_loc = find_loop_location (loop); | |
78c60e3d | 795 | if (dump_kind_p (MSG_NOTE)) |
ebfd146a | 796 | { |
84df911b DC |
797 | if (LOCATION_LOCUS (loop_loc) != UNKNOWN_LOC) |
798 | dump_printf (MSG_NOTE, "\nloop at %s:%d: ", LOC_FILE (loop_loc), | |
799 | LOC_LINE (loop_loc)); | |
78c60e3d | 800 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, cond_stmt, 0); |
ebfd146a | 801 | } |
ebfd146a IR |
802 | loop->nb_iterations = niters; |
803 | } | |
804 | ||
805 | ||
b8698a0f | 806 | /* Given LOOP this function generates a new copy of it and puts it |
ebfd146a IR |
807 | on E which is either the entry or exit of LOOP. */ |
808 | ||
809 | struct loop * | |
810 | slpeel_tree_duplicate_loop_to_edge_cfg (struct loop *loop, edge e) | |
811 | { | |
812 | struct loop *new_loop; | |
813 | basic_block *new_bbs, *bbs; | |
814 | bool at_exit; | |
815 | bool was_imm_dom; | |
b8698a0f | 816 | basic_block exit_dest; |
ebfd146a IR |
817 | gimple phi; |
818 | tree phi_arg; | |
819 | edge exit, new_exit; | |
820 | gimple_stmt_iterator gsi; | |
821 | ||
b8698a0f | 822 | at_exit = (e == single_exit (loop)); |
ebfd146a IR |
823 | if (!at_exit && e != loop_preheader_edge (loop)) |
824 | return NULL; | |
825 | ||
826 | bbs = get_loop_body (loop); | |
827 | ||
828 | /* Check whether duplication is possible. */ | |
829 | if (!can_copy_bbs_p (bbs, loop->num_nodes)) | |
830 | { | |
831 | free (bbs); | |
832 | return NULL; | |
833 | } | |
834 | ||
835 | /* Generate new loop structure. */ | |
836 | new_loop = duplicate_loop (loop, loop_outer (loop)); | |
837 | if (!new_loop) | |
838 | { | |
839 | free (bbs); | |
840 | return NULL; | |
841 | } | |
842 | ||
843 | exit_dest = single_exit (loop)->dest; | |
b8698a0f L |
844 | was_imm_dom = (get_immediate_dominator (CDI_DOMINATORS, |
845 | exit_dest) == loop->header ? | |
ebfd146a IR |
846 | true : false); |
847 | ||
848 | new_bbs = XNEWVEC (basic_block, loop->num_nodes); | |
849 | ||
850 | exit = single_exit (loop); | |
851 | copy_bbs (bbs, loop->num_nodes, new_bbs, | |
852 | &exit, 1, &new_exit, NULL, | |
853 | e->src); | |
854 | ||
b8698a0f | 855 | /* Duplicating phi args at exit bbs as coming |
ebfd146a IR |
856 | also from exit of duplicated loop. */ |
857 | for (gsi = gsi_start_phis (exit_dest); !gsi_end_p (gsi); gsi_next (&gsi)) | |
858 | { | |
859 | phi = gsi_stmt (gsi); | |
860 | phi_arg = PHI_ARG_DEF_FROM_EDGE (phi, single_exit (loop)); | |
861 | if (phi_arg) | |
862 | { | |
863 | edge new_loop_exit_edge; | |
f5045c96 | 864 | source_location locus; |
ebfd146a | 865 | |
f5045c96 | 866 | locus = gimple_phi_arg_location_from_edge (phi, single_exit (loop)); |
ebfd146a IR |
867 | if (EDGE_SUCC (new_loop->header, 0)->dest == new_loop->latch) |
868 | new_loop_exit_edge = EDGE_SUCC (new_loop->header, 1); | |
869 | else | |
870 | new_loop_exit_edge = EDGE_SUCC (new_loop->header, 0); | |
b8698a0f | 871 | |
9e227d60 | 872 | add_phi_arg (phi, phi_arg, new_loop_exit_edge, locus); |
ebfd146a | 873 | } |
b8698a0f L |
874 | } |
875 | ||
ebfd146a IR |
876 | if (at_exit) /* Add the loop copy at exit. */ |
877 | { | |
878 | redirect_edge_and_branch_force (e, new_loop->header); | |
879 | PENDING_STMT (e) = NULL; | |
880 | set_immediate_dominator (CDI_DOMINATORS, new_loop->header, e->src); | |
881 | if (was_imm_dom) | |
882 | set_immediate_dominator (CDI_DOMINATORS, exit_dest, new_loop->header); | |
883 | } | |
884 | else /* Add the copy at entry. */ | |
885 | { | |
886 | edge new_exit_e; | |
887 | edge entry_e = loop_preheader_edge (loop); | |
888 | basic_block preheader = entry_e->src; | |
b8698a0f L |
889 | |
890 | if (!flow_bb_inside_loop_p (new_loop, | |
ebfd146a IR |
891 | EDGE_SUCC (new_loop->header, 0)->dest)) |
892 | new_exit_e = EDGE_SUCC (new_loop->header, 0); | |
893 | else | |
b8698a0f | 894 | new_exit_e = EDGE_SUCC (new_loop->header, 1); |
ebfd146a IR |
895 | |
896 | redirect_edge_and_branch_force (new_exit_e, loop->header); | |
897 | PENDING_STMT (new_exit_e) = NULL; | |
898 | set_immediate_dominator (CDI_DOMINATORS, loop->header, | |
899 | new_exit_e->src); | |
900 | ||
b8698a0f | 901 | /* We have to add phi args to the loop->header here as coming |
ebfd146a IR |
902 | from new_exit_e edge. */ |
903 | for (gsi = gsi_start_phis (loop->header); | |
904 | !gsi_end_p (gsi); | |
905 | gsi_next (&gsi)) | |
906 | { | |
907 | phi = gsi_stmt (gsi); | |
908 | phi_arg = PHI_ARG_DEF_FROM_EDGE (phi, entry_e); | |
909 | if (phi_arg) | |
f5045c96 | 910 | add_phi_arg (phi, phi_arg, new_exit_e, |
9e227d60 | 911 | gimple_phi_arg_location_from_edge (phi, entry_e)); |
b8698a0f | 912 | } |
ebfd146a IR |
913 | |
914 | redirect_edge_and_branch_force (entry_e, new_loop->header); | |
915 | PENDING_STMT (entry_e) = NULL; | |
916 | set_immediate_dominator (CDI_DOMINATORS, new_loop->header, preheader); | |
917 | } | |
918 | ||
919 | free (new_bbs); | |
920 | free (bbs); | |
921 | ||
922 | return new_loop; | |
923 | } | |
924 | ||
925 | ||
926 | /* Given the condition statement COND, put it as the last statement | |
927 | of GUARD_BB; EXIT_BB is the basic block to skip the loop; | |
b8698a0f | 928 | Assumes that this is the single exit of the guarded loop. |
86290011 | 929 | Returns the skip edge, inserts new stmts on the COND_EXPR_STMT_LIST. */ |
ebfd146a IR |
930 | |
931 | static edge | |
86290011 RG |
932 | slpeel_add_loop_guard (basic_block guard_bb, tree cond, |
933 | gimple_seq cond_expr_stmt_list, | |
e78410bf JH |
934 | basic_block exit_bb, basic_block dom_bb, |
935 | int probability) | |
ebfd146a IR |
936 | { |
937 | gimple_stmt_iterator gsi; | |
938 | edge new_e, enter_e; | |
939 | gimple cond_stmt; | |
940 | gimple_seq gimplify_stmt_list = NULL; | |
941 | ||
942 | enter_e = EDGE_SUCC (guard_bb, 0); | |
943 | enter_e->flags &= ~EDGE_FALLTHRU; | |
944 | enter_e->flags |= EDGE_FALSE_VALUE; | |
945 | gsi = gsi_last_bb (guard_bb); | |
946 | ||
f7a06a98 RG |
947 | cond = force_gimple_operand_1 (cond, &gimplify_stmt_list, is_gimple_condexpr, |
948 | NULL_TREE); | |
86290011 RG |
949 | if (gimplify_stmt_list) |
950 | gimple_seq_add_seq (&cond_expr_stmt_list, gimplify_stmt_list); | |
f7a06a98 | 951 | cond_stmt = gimple_build_cond_from_tree (cond, NULL_TREE, NULL_TREE); |
86290011 RG |
952 | if (cond_expr_stmt_list) |
953 | gsi_insert_seq_after (&gsi, cond_expr_stmt_list, GSI_NEW_STMT); | |
ebfd146a IR |
954 | |
955 | gsi = gsi_last_bb (guard_bb); | |
956 | gsi_insert_after (&gsi, cond_stmt, GSI_NEW_STMT); | |
957 | ||
958 | /* Add new edge to connect guard block to the merge/loop-exit block. */ | |
959 | new_e = make_edge (guard_bb, exit_bb, EDGE_TRUE_VALUE); | |
e78410bf JH |
960 | |
961 | new_e->count = guard_bb->count; | |
962 | new_e->probability = probability; | |
963 | new_e->count = apply_probability (enter_e->count, probability); | |
964 | enter_e->count -= new_e->count; | |
965 | enter_e->probability = inverse_probability (probability); | |
ebfd146a IR |
966 | set_immediate_dominator (CDI_DOMINATORS, exit_bb, dom_bb); |
967 | return new_e; | |
968 | } | |
969 | ||
970 | ||
971 | /* This function verifies that the following restrictions apply to LOOP: | |
972 | (1) it is innermost | |
973 | (2) it consists of exactly 2 basic blocks - header, and an empty latch. | |
974 | (3) it is single entry, single exit | |
975 | (4) its exit condition is the last stmt in the header | |
976 | (5) E is the entry/exit edge of LOOP. | |
977 | */ | |
978 | ||
979 | bool | |
980 | slpeel_can_duplicate_loop_p (const struct loop *loop, const_edge e) | |
981 | { | |
982 | edge exit_e = single_exit (loop); | |
983 | edge entry_e = loop_preheader_edge (loop); | |
984 | gimple orig_cond = get_loop_exit_condition (loop); | |
985 | gimple_stmt_iterator loop_exit_gsi = gsi_last_bb (exit_e->src); | |
986 | ||
5006671f | 987 | if (need_ssa_update_p (cfun)) |
ebfd146a IR |
988 | return false; |
989 | ||
990 | if (loop->inner | |
991 | /* All loops have an outer scope; the only case loop->outer is NULL is for | |
992 | the function itself. */ | |
993 | || !loop_outer (loop) | |
994 | || loop->num_nodes != 2 | |
995 | || !empty_block_p (loop->latch) | |
996 | || !single_exit (loop) | |
997 | /* Verify that new loop exit condition can be trivially modified. */ | |
998 | || (!orig_cond || orig_cond != gsi_stmt (loop_exit_gsi)) | |
999 | || (e != exit_e && e != entry_e)) | |
1000 | return false; | |
1001 | ||
1002 | return true; | |
1003 | } | |
1004 | ||
1005 | #ifdef ENABLE_CHECKING | |
1006 | static void | |
1007 | slpeel_verify_cfg_after_peeling (struct loop *first_loop, | |
1008 | struct loop *second_loop) | |
1009 | { | |
1010 | basic_block loop1_exit_bb = single_exit (first_loop)->dest; | |
1011 | basic_block loop2_entry_bb = loop_preheader_edge (second_loop)->src; | |
1012 | basic_block loop1_entry_bb = loop_preheader_edge (first_loop)->src; | |
1013 | ||
1014 | /* A guard that controls whether the second_loop is to be executed or skipped | |
1015 | is placed in first_loop->exit. first_loop->exit therefore has two | |
1016 | successors - one is the preheader of second_loop, and the other is a bb | |
1017 | after second_loop. | |
1018 | */ | |
1019 | gcc_assert (EDGE_COUNT (loop1_exit_bb->succs) == 2); | |
b8698a0f | 1020 | |
ebfd146a IR |
1021 | /* 1. Verify that one of the successors of first_loop->exit is the preheader |
1022 | of second_loop. */ | |
b8698a0f | 1023 | |
ebfd146a IR |
1024 | /* The preheader of new_loop is expected to have two predecessors: |
1025 | first_loop->exit and the block that precedes first_loop. */ | |
1026 | ||
b8698a0f | 1027 | gcc_assert (EDGE_COUNT (loop2_entry_bb->preds) == 2 |
ebfd146a IR |
1028 | && ((EDGE_PRED (loop2_entry_bb, 0)->src == loop1_exit_bb |
1029 | && EDGE_PRED (loop2_entry_bb, 1)->src == loop1_entry_bb) | |
1030 | || (EDGE_PRED (loop2_entry_bb, 1)->src == loop1_exit_bb | |
1031 | && EDGE_PRED (loop2_entry_bb, 0)->src == loop1_entry_bb))); | |
b8698a0f | 1032 | |
ebfd146a IR |
1033 | /* Verify that the other successor of first_loop->exit is after the |
1034 | second_loop. */ | |
1035 | /* TODO */ | |
1036 | } | |
1037 | #endif | |
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; | |
1055 | gimple newphi; | |
1056 | edge e_true, e_false, e_fallthru; | |
1057 | gimple 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); | |
1064 | cond_bb = split_edge(e); | |
1065 | ||
1066 | e = single_pred_edge (bb_before_first_loop); | |
1067 | then_bb = split_edge(e); | |
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. | |
1128 | - E: the exit or entry edge of LOOP. | |
1129 | If it is the entry edge, we peel the first iterations of LOOP. In this | |
1130 | case first-loop is LOOP, and second-loop is the newly created loop. | |
1131 | If it is the exit edge, we peel the last iterations of LOOP. In this | |
1132 | case, first-loop is the newly created loop, and second-loop is LOOP. | |
1133 | - NITERS: the number of iterations that LOOP iterates. | |
1134 | - FIRST_NITERS: the number of iterations that the first-loop should iterate. | |
1135 | - UPDATE_FIRST_LOOP_COUNT: specified whether this function is responsible | |
1136 | for updating the loop bound of the first-loop to FIRST_NITERS. If it | |
1137 | is false, the caller of this function may want to take care of this | |
1138 | (this can be useful if we don't want new stmts added to first-loop). | |
1139 | - TH: cost model profitability threshold of iterations for vectorization. | |
1140 | - CHECK_PROFITABILITY: specify whether cost model check has not occurred | |
1141 | during versioning and hence needs to occur during | |
b8698a0f | 1142 | prologue generation or whether cost model check |
ebfd146a IR |
1143 | has not occurred during prologue generation and hence |
1144 | needs to occur during epilogue generation. | |
e78410bf JH |
1145 | - BOUND1 is the upper bound on number of iterations of the first loop (if known) |
1146 | - BOUND2 is the upper bound on number of iterations of the second loop (if known) | |
b8698a0f | 1147 | |
ebfd146a IR |
1148 | |
1149 | Output: | |
1150 | The function returns a pointer to the new loop-copy, or NULL if it failed | |
1151 | to perform the transformation. | |
1152 | ||
1153 | The function generates two if-then-else guards: one before the first loop, | |
1154 | and the other before the second loop: | |
1155 | The first guard is: | |
1156 | if (FIRST_NITERS == 0) then skip the first loop, | |
1157 | and go directly to the second loop. | |
1158 | The second guard is: | |
1159 | if (FIRST_NITERS == NITERS) then skip the second loop. | |
1160 | ||
86290011 RG |
1161 | If the optional COND_EXPR and COND_EXPR_STMT_LIST arguments are given |
1162 | then the generated condition is combined with COND_EXPR and the | |
1163 | statements in COND_EXPR_STMT_LIST are emitted together with it. | |
1164 | ||
ebfd146a IR |
1165 | FORNOW only simple loops are supported (see slpeel_can_duplicate_loop_p). |
1166 | FORNOW the resulting code will not be in loop-closed-ssa form. | |
1167 | */ | |
1168 | ||
1169 | static struct loop* | |
b8698a0f | 1170 | slpeel_tree_peel_loop_to_edge (struct loop *loop, |
5d2eb24b | 1171 | edge e, tree *first_niters, |
ebfd146a | 1172 | tree niters, bool update_first_loop_count, |
86290011 | 1173 | unsigned int th, bool check_profitability, |
e78410bf JH |
1174 | tree cond_expr, gimple_seq cond_expr_stmt_list, |
1175 | int bound1, int bound2) | |
ebfd146a IR |
1176 | { |
1177 | struct loop *new_loop = NULL, *first_loop, *second_loop; | |
1178 | edge skip_e; | |
1179 | tree pre_condition = NULL_TREE; | |
ebfd146a IR |
1180 | basic_block bb_before_second_loop, bb_after_second_loop; |
1181 | basic_block bb_before_first_loop; | |
1182 | basic_block bb_between_loops; | |
1183 | basic_block new_exit_bb; | |
e20f6b4b | 1184 | gimple_stmt_iterator gsi; |
ebfd146a IR |
1185 | edge exit_e = single_exit (loop); |
1186 | LOC loop_loc; | |
1187 | tree cost_pre_condition = NULL_TREE; | |
e78410bf JH |
1188 | /* There are many aspects to how likely the first loop is going to be executed. |
1189 | Without histogram we can't really do good job. Simply set it to | |
1190 | 2/3, so the first loop is not reordered to the end of function and | |
1191 | the hot path through stays short. */ | |
1192 | int first_guard_probability = 2 * REG_BR_PROB_BASE / 3; | |
1193 | int second_guard_probability = 2 * REG_BR_PROB_BASE / 3; | |
1194 | int probability_of_second_loop; | |
b8698a0f | 1195 | |
ebfd146a IR |
1196 | if (!slpeel_can_duplicate_loop_p (loop, e)) |
1197 | return NULL; | |
b8698a0f | 1198 | |
e20f6b4b JJ |
1199 | /* If the loop has a virtual PHI, but exit bb doesn't, create a virtual PHI |
1200 | in the exit bb and rename all the uses after the loop. This simplifies | |
1201 | the *guard[12] routines, which assume loop closed SSA form for all PHIs | |
1202 | (but normally loop closed SSA form doesn't require virtual PHIs to be | |
1203 | in the same form). Doing this early simplifies the checking what | |
1204 | uses should be renamed. */ | |
1205 | for (gsi = gsi_start_phis (loop->header); !gsi_end_p (gsi); gsi_next (&gsi)) | |
ea057359 | 1206 | if (virtual_operand_p (gimple_phi_result (gsi_stmt (gsi)))) |
e20f6b4b JJ |
1207 | { |
1208 | gimple phi = gsi_stmt (gsi); | |
1209 | for (gsi = gsi_start_phis (exit_e->dest); | |
1210 | !gsi_end_p (gsi); gsi_next (&gsi)) | |
ea057359 | 1211 | if (virtual_operand_p (gimple_phi_result (gsi_stmt (gsi)))) |
e20f6b4b JJ |
1212 | break; |
1213 | if (gsi_end_p (gsi)) | |
1214 | { | |
070ecdfd RG |
1215 | tree new_vop = copy_ssa_name (PHI_RESULT (phi), NULL); |
1216 | gimple new_phi = create_phi_node (new_vop, exit_e->dest); | |
e20f6b4b JJ |
1217 | tree vop = PHI_ARG_DEF_FROM_EDGE (phi, EDGE_SUCC (loop->latch, 0)); |
1218 | imm_use_iterator imm_iter; | |
1219 | gimple stmt; | |
e20f6b4b JJ |
1220 | use_operand_p use_p; |
1221 | ||
9e227d60 | 1222 | add_phi_arg (new_phi, vop, exit_e, UNKNOWN_LOCATION); |
e20f6b4b JJ |
1223 | gimple_phi_set_result (new_phi, new_vop); |
1224 | FOR_EACH_IMM_USE_STMT (stmt, imm_iter, vop) | |
1225 | if (stmt != new_phi && gimple_bb (stmt) != loop->header) | |
1226 | FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter) | |
1227 | SET_USE (use_p, new_vop); | |
1228 | } | |
1229 | break; | |
1230 | } | |
ebfd146a IR |
1231 | |
1232 | /* 1. Generate a copy of LOOP and put it on E (E is the entry/exit of LOOP). | |
1233 | Resulting CFG would be: | |
1234 | ||
1235 | first_loop: | |
1236 | do { | |
1237 | } while ... | |
1238 | ||
1239 | second_loop: | |
1240 | do { | |
1241 | } while ... | |
1242 | ||
1243 | orig_exit_bb: | |
1244 | */ | |
b8698a0f | 1245 | |
ebfd146a IR |
1246 | if (!(new_loop = slpeel_tree_duplicate_loop_to_edge_cfg (loop, e))) |
1247 | { | |
1248 | loop_loc = find_loop_location (loop); | |
78c60e3d SS |
1249 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, loop_loc, |
1250 | "tree_duplicate_loop_to_edge_cfg failed.\n"); | |
ebfd146a IR |
1251 | return NULL; |
1252 | } | |
b8698a0f | 1253 | |
684f25f4 AO |
1254 | if (MAY_HAVE_DEBUG_STMTS) |
1255 | { | |
1256 | gcc_assert (!adjust_vec); | |
1257 | adjust_vec = VEC_alloc (adjust_info, stack, 32); | |
1258 | } | |
1259 | ||
ebfd146a IR |
1260 | if (e == exit_e) |
1261 | { | |
1262 | /* NEW_LOOP was placed after LOOP. */ | |
1263 | first_loop = loop; | |
1264 | second_loop = new_loop; | |
1265 | } | |
1266 | else | |
1267 | { | |
1268 | /* NEW_LOOP was placed before LOOP. */ | |
1269 | first_loop = new_loop; | |
1270 | second_loop = loop; | |
1271 | } | |
1272 | ||
ebfd146a IR |
1273 | slpeel_update_phis_for_duplicate_loop (loop, new_loop, e == exit_e); |
1274 | rename_variables_in_loop (new_loop); | |
1275 | ||
1276 | ||
1277 | /* 2. Add the guard code in one of the following ways: | |
1278 | ||
1279 | 2.a Add the guard that controls whether the first loop is executed. | |
1280 | This occurs when this function is invoked for prologue or epilogue | |
1281 | generation and when the cost model check can be done at compile time. | |
1282 | ||
1283 | Resulting CFG would be: | |
1284 | ||
1285 | bb_before_first_loop: | |
1286 | if (FIRST_NITERS == 0) GOTO bb_before_second_loop | |
1287 | GOTO first-loop | |
1288 | ||
1289 | first_loop: | |
1290 | do { | |
1291 | } while ... | |
1292 | ||
1293 | bb_before_second_loop: | |
1294 | ||
1295 | second_loop: | |
1296 | do { | |
1297 | } while ... | |
1298 | ||
1299 | orig_exit_bb: | |
1300 | ||
1301 | 2.b Add the cost model check that allows the prologue | |
1302 | to iterate for the entire unchanged scalar | |
1303 | iterations of the loop in the event that the cost | |
1304 | model indicates that the scalar loop is more | |
1305 | profitable than the vector one. This occurs when | |
1306 | this function is invoked for prologue generation | |
1307 | and the cost model check needs to be done at run | |
1308 | time. | |
1309 | ||
1310 | Resulting CFG after prologue peeling would be: | |
1311 | ||
1312 | if (scalar_loop_iterations <= th) | |
1313 | FIRST_NITERS = scalar_loop_iterations | |
1314 | ||
1315 | bb_before_first_loop: | |
1316 | if (FIRST_NITERS == 0) GOTO bb_before_second_loop | |
1317 | GOTO first-loop | |
1318 | ||
1319 | first_loop: | |
1320 | do { | |
1321 | } while ... | |
1322 | ||
1323 | bb_before_second_loop: | |
1324 | ||
1325 | second_loop: | |
1326 | do { | |
1327 | } while ... | |
1328 | ||
1329 | orig_exit_bb: | |
1330 | ||
1331 | 2.c Add the cost model check that allows the epilogue | |
1332 | to iterate for the entire unchanged scalar | |
1333 | iterations of the loop in the event that the cost | |
1334 | model indicates that the scalar loop is more | |
1335 | profitable than the vector one. This occurs when | |
1336 | this function is invoked for epilogue generation | |
1337 | and the cost model check needs to be done at run | |
86290011 RG |
1338 | time. This check is combined with any pre-existing |
1339 | check in COND_EXPR to avoid versioning. | |
ebfd146a IR |
1340 | |
1341 | Resulting CFG after prologue peeling would be: | |
1342 | ||
1343 | bb_before_first_loop: | |
1344 | if ((scalar_loop_iterations <= th) | |
1345 | || | |
1346 | FIRST_NITERS == 0) GOTO bb_before_second_loop | |
1347 | GOTO first-loop | |
1348 | ||
1349 | first_loop: | |
1350 | do { | |
1351 | } while ... | |
1352 | ||
1353 | bb_before_second_loop: | |
1354 | ||
1355 | second_loop: | |
1356 | do { | |
1357 | } while ... | |
1358 | ||
1359 | orig_exit_bb: | |
1360 | */ | |
1361 | ||
1362 | bb_before_first_loop = split_edge (loop_preheader_edge (first_loop)); | |
1363 | bb_before_second_loop = split_edge (single_exit (first_loop)); | |
1364 | ||
e78410bf JH |
1365 | probability_of_second_loop = (inverse_probability (first_guard_probability) |
1366 | + combine_probabilities (second_guard_probability, | |
1367 | first_guard_probability)); | |
1368 | /* Theoretically preheader edge of first loop and exit edge should have | |
1369 | same frequencies. Loop exit probablities are however easy to get wrong. | |
1370 | It is safer to copy value from original loop entry. */ | |
1371 | bb_before_second_loop->frequency | |
1372 | = apply_probability (bb_before_first_loop->frequency, | |
1373 | probability_of_second_loop); | |
1374 | bb_before_second_loop->count | |
1375 | = apply_probability (bb_before_first_loop->count, | |
1376 | probability_of_second_loop); | |
1377 | single_succ_edge (bb_before_second_loop)->count | |
1378 | = bb_before_second_loop->count; | |
1379 | ||
ebfd146a IR |
1380 | /* Epilogue peeling. */ |
1381 | if (!update_first_loop_count) | |
1382 | { | |
1383 | pre_condition = | |
5d2eb24b IR |
1384 | fold_build2 (LE_EXPR, boolean_type_node, *first_niters, |
1385 | build_int_cst (TREE_TYPE (*first_niters), 0)); | |
ebfd146a IR |
1386 | if (check_profitability) |
1387 | { | |
1388 | tree scalar_loop_iters | |
1389 | = unshare_expr (LOOP_VINFO_NITERS_UNCHANGED | |
1390 | (loop_vec_info_for_loop (loop))); | |
b8698a0f L |
1391 | cost_pre_condition = |
1392 | fold_build2 (LE_EXPR, boolean_type_node, scalar_loop_iters, | |
ebfd146a IR |
1393 | build_int_cst (TREE_TYPE (scalar_loop_iters), th)); |
1394 | ||
1395 | pre_condition = fold_build2 (TRUTH_OR_EXPR, boolean_type_node, | |
1396 | cost_pre_condition, pre_condition); | |
1397 | } | |
86290011 RG |
1398 | if (cond_expr) |
1399 | { | |
1400 | pre_condition = | |
1401 | fold_build2 (TRUTH_OR_EXPR, boolean_type_node, | |
1402 | pre_condition, | |
1403 | fold_build1 (TRUTH_NOT_EXPR, boolean_type_node, | |
1404 | cond_expr)); | |
1405 | } | |
ebfd146a IR |
1406 | } |
1407 | ||
b8698a0f | 1408 | /* Prologue peeling. */ |
ebfd146a IR |
1409 | else |
1410 | { | |
1411 | if (check_profitability) | |
1412 | set_prologue_iterations (bb_before_first_loop, first_niters, | |
e78410bf | 1413 | loop, th, first_guard_probability); |
ebfd146a IR |
1414 | |
1415 | pre_condition = | |
5d2eb24b IR |
1416 | fold_build2 (LE_EXPR, boolean_type_node, *first_niters, |
1417 | build_int_cst (TREE_TYPE (*first_niters), 0)); | |
ebfd146a IR |
1418 | } |
1419 | ||
1420 | skip_e = slpeel_add_loop_guard (bb_before_first_loop, pre_condition, | |
86290011 | 1421 | cond_expr_stmt_list, |
e78410bf JH |
1422 | bb_before_second_loop, bb_before_first_loop, |
1423 | inverse_probability (first_guard_probability)); | |
1424 | scale_loop_profile (first_loop, first_guard_probability, | |
1425 | check_profitability && (int)th > bound1 ? th : bound1); | |
ebfd146a IR |
1426 | slpeel_update_phi_nodes_for_guard1 (skip_e, first_loop, |
1427 | first_loop == new_loop, | |
c334023f | 1428 | &new_exit_bb); |
ebfd146a IR |
1429 | |
1430 | ||
1431 | /* 3. Add the guard that controls whether the second loop is executed. | |
1432 | Resulting CFG would be: | |
1433 | ||
1434 | bb_before_first_loop: | |
1435 | if (FIRST_NITERS == 0) GOTO bb_before_second_loop (skip first loop) | |
1436 | GOTO first-loop | |
1437 | ||
1438 | first_loop: | |
1439 | do { | |
1440 | } while ... | |
1441 | ||
1442 | bb_between_loops: | |
1443 | if (FIRST_NITERS == NITERS) GOTO bb_after_second_loop (skip second loop) | |
1444 | GOTO bb_before_second_loop | |
1445 | ||
1446 | bb_before_second_loop: | |
1447 | ||
1448 | second_loop: | |
1449 | do { | |
1450 | } while ... | |
1451 | ||
1452 | bb_after_second_loop: | |
1453 | ||
1454 | orig_exit_bb: | |
1455 | */ | |
1456 | ||
1457 | bb_between_loops = new_exit_bb; | |
1458 | bb_after_second_loop = split_edge (single_exit (second_loop)); | |
1459 | ||
b8698a0f | 1460 | pre_condition = |
5d2eb24b | 1461 | fold_build2 (EQ_EXPR, boolean_type_node, *first_niters, niters); |
86290011 | 1462 | skip_e = slpeel_add_loop_guard (bb_between_loops, pre_condition, NULL, |
e78410bf JH |
1463 | bb_after_second_loop, bb_before_first_loop, |
1464 | inverse_probability (second_guard_probability)); | |
1465 | scale_loop_profile (second_loop, probability_of_second_loop, bound2); | |
ebfd146a IR |
1466 | slpeel_update_phi_nodes_for_guard2 (skip_e, second_loop, |
1467 | second_loop == new_loop, &new_exit_bb); | |
1468 | ||
1469 | /* 4. Make first-loop iterate FIRST_NITERS times, if requested. | |
1470 | */ | |
1471 | if (update_first_loop_count) | |
5d2eb24b | 1472 | slpeel_make_loop_iterate_ntimes (first_loop, *first_niters); |
ebfd146a | 1473 | |
040d39ee RG |
1474 | delete_update_ssa (); |
1475 | ||
684f25f4 AO |
1476 | adjust_vec_debug_stmts (); |
1477 | ||
ebfd146a IR |
1478 | return new_loop; |
1479 | } | |
1480 | ||
1481 | /* Function vect_get_loop_location. | |
1482 | ||
1483 | Extract the location of the loop in the source code. | |
1484 | If the loop is not well formed for vectorization, an estimated | |
1485 | location is calculated. | |
1486 | Return the loop location if succeed and NULL if not. */ | |
1487 | ||
1488 | LOC | |
1489 | find_loop_location (struct loop *loop) | |
1490 | { | |
1491 | gimple stmt = NULL; | |
1492 | basic_block bb; | |
1493 | gimple_stmt_iterator si; | |
1494 | ||
1495 | if (!loop) | |
1496 | return UNKNOWN_LOC; | |
1497 | ||
1498 | stmt = get_loop_exit_condition (loop); | |
1499 | ||
1500 | if (stmt && gimple_location (stmt) != UNKNOWN_LOC) | |
1501 | return gimple_location (stmt); | |
1502 | ||
1503 | /* If we got here the loop is probably not "well formed", | |
1504 | try to estimate the loop location */ | |
1505 | ||
1506 | if (!loop->header) | |
1507 | return UNKNOWN_LOC; | |
1508 | ||
1509 | bb = loop->header; | |
1510 | ||
1511 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
1512 | { | |
1513 | stmt = gsi_stmt (si); | |
1514 | if (gimple_location (stmt) != UNKNOWN_LOC) | |
1515 | return gimple_location (stmt); | |
1516 | } | |
1517 | ||
1518 | return UNKNOWN_LOC; | |
1519 | } | |
1520 | ||
1521 | ||
1522 | /* This function builds ni_name = number of iterations loop executes | |
86290011 RG |
1523 | on the loop preheader. If SEQ is given the stmt is instead emitted |
1524 | there. */ | |
ebfd146a IR |
1525 | |
1526 | static tree | |
86290011 | 1527 | vect_build_loop_niters (loop_vec_info loop_vinfo, gimple_seq seq) |
ebfd146a IR |
1528 | { |
1529 | tree ni_name, var; | |
1530 | gimple_seq stmts = NULL; | |
1531 | edge pe; | |
1532 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
1533 | tree ni = unshare_expr (LOOP_VINFO_NITERS (loop_vinfo)); | |
1534 | ||
1535 | var = create_tmp_var (TREE_TYPE (ni), "niters"); | |
ebfd146a IR |
1536 | ni_name = force_gimple_operand (ni, &stmts, false, var); |
1537 | ||
1538 | pe = loop_preheader_edge (loop); | |
1539 | if (stmts) | |
1540 | { | |
86290011 RG |
1541 | if (seq) |
1542 | gimple_seq_add_seq (&seq, stmts); | |
1543 | else | |
1544 | { | |
1545 | basic_block new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
1546 | gcc_assert (!new_bb); | |
1547 | } | |
ebfd146a IR |
1548 | } |
1549 | ||
1550 | return ni_name; | |
1551 | } | |
1552 | ||
1553 | ||
1554 | /* This function generates the following statements: | |
1555 | ||
1556 | ni_name = number of iterations loop executes | |
1557 | ratio = ni_name / vf | |
1558 | ratio_mult_vf_name = ratio * vf | |
1559 | ||
86290011 RG |
1560 | and places them at the loop preheader edge or in COND_EXPR_STMT_LIST |
1561 | if that is non-NULL. */ | |
ebfd146a | 1562 | |
b8698a0f L |
1563 | static void |
1564 | vect_generate_tmps_on_preheader (loop_vec_info loop_vinfo, | |
ebfd146a | 1565 | tree *ni_name_ptr, |
b8698a0f | 1566 | tree *ratio_mult_vf_name_ptr, |
86290011 RG |
1567 | tree *ratio_name_ptr, |
1568 | gimple_seq cond_expr_stmt_list) | |
ebfd146a IR |
1569 | { |
1570 | ||
1571 | edge pe; | |
1572 | basic_block new_bb; | |
1573 | gimple_seq stmts; | |
48df3fa6 | 1574 | tree ni_name, ni_minus_gap_name; |
ebfd146a IR |
1575 | tree var; |
1576 | tree ratio_name; | |
1577 | tree ratio_mult_vf_name; | |
1578 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
1579 | tree ni = LOOP_VINFO_NITERS (loop_vinfo); | |
1580 | int vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
1581 | tree log_vf; | |
1582 | ||
1583 | pe = loop_preheader_edge (loop); | |
1584 | ||
b8698a0f | 1585 | /* Generate temporary variable that contains |
ebfd146a IR |
1586 | number of iterations loop executes. */ |
1587 | ||
86290011 | 1588 | ni_name = vect_build_loop_niters (loop_vinfo, cond_expr_stmt_list); |
ebfd146a IR |
1589 | log_vf = build_int_cst (TREE_TYPE (ni), exact_log2 (vf)); |
1590 | ||
48df3fa6 IR |
1591 | /* If epilogue loop is required because of data accesses with gaps, we |
1592 | subtract one iteration from the total number of iterations here for | |
1593 | correct calculation of RATIO. */ | |
1594 | if (LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo)) | |
1595 | { | |
1596 | ni_minus_gap_name = fold_build2 (MINUS_EXPR, TREE_TYPE (ni_name), | |
1597 | ni_name, | |
1598 | build_one_cst (TREE_TYPE (ni_name))); | |
1599 | if (!is_gimple_val (ni_minus_gap_name)) | |
1600 | { | |
1601 | var = create_tmp_var (TREE_TYPE (ni), "ni_gap"); | |
48df3fa6 IR |
1602 | |
1603 | stmts = NULL; | |
1604 | ni_minus_gap_name = force_gimple_operand (ni_minus_gap_name, &stmts, | |
1605 | true, var); | |
1606 | if (cond_expr_stmt_list) | |
1607 | gimple_seq_add_seq (&cond_expr_stmt_list, stmts); | |
1608 | else | |
1609 | { | |
1610 | pe = loop_preheader_edge (loop); | |
1611 | new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
1612 | gcc_assert (!new_bb); | |
1613 | } | |
1614 | } | |
1615 | } | |
1616 | else | |
1617 | ni_minus_gap_name = ni_name; | |
1618 | ||
ebfd146a IR |
1619 | /* Create: ratio = ni >> log2(vf) */ |
1620 | ||
48df3fa6 IR |
1621 | ratio_name = fold_build2 (RSHIFT_EXPR, TREE_TYPE (ni_minus_gap_name), |
1622 | ni_minus_gap_name, log_vf); | |
ebfd146a IR |
1623 | if (!is_gimple_val (ratio_name)) |
1624 | { | |
1625 | var = create_tmp_var (TREE_TYPE (ni), "bnd"); | |
ebfd146a IR |
1626 | |
1627 | stmts = NULL; | |
1628 | ratio_name = force_gimple_operand (ratio_name, &stmts, true, var); | |
86290011 RG |
1629 | if (cond_expr_stmt_list) |
1630 | gimple_seq_add_seq (&cond_expr_stmt_list, stmts); | |
1631 | else | |
1632 | { | |
1633 | pe = loop_preheader_edge (loop); | |
1634 | new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
1635 | gcc_assert (!new_bb); | |
1636 | } | |
ebfd146a | 1637 | } |
b8698a0f | 1638 | |
ebfd146a IR |
1639 | /* Create: ratio_mult_vf = ratio << log2 (vf). */ |
1640 | ||
1641 | ratio_mult_vf_name = fold_build2 (LSHIFT_EXPR, TREE_TYPE (ratio_name), | |
1642 | ratio_name, log_vf); | |
1643 | if (!is_gimple_val (ratio_mult_vf_name)) | |
1644 | { | |
1645 | var = create_tmp_var (TREE_TYPE (ni), "ratio_mult_vf"); | |
ebfd146a IR |
1646 | |
1647 | stmts = NULL; | |
1648 | ratio_mult_vf_name = force_gimple_operand (ratio_mult_vf_name, &stmts, | |
1649 | true, var); | |
86290011 RG |
1650 | if (cond_expr_stmt_list) |
1651 | gimple_seq_add_seq (&cond_expr_stmt_list, stmts); | |
1652 | else | |
1653 | { | |
1654 | pe = loop_preheader_edge (loop); | |
1655 | new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
1656 | gcc_assert (!new_bb); | |
1657 | } | |
ebfd146a IR |
1658 | } |
1659 | ||
1660 | *ni_name_ptr = ni_name; | |
1661 | *ratio_mult_vf_name_ptr = ratio_mult_vf_name; | |
1662 | *ratio_name_ptr = ratio_name; | |
b8698a0f L |
1663 | |
1664 | return; | |
ebfd146a IR |
1665 | } |
1666 | ||
1667 | /* Function vect_can_advance_ivs_p | |
1668 | ||
b8698a0f L |
1669 | In case the number of iterations that LOOP iterates is unknown at compile |
1670 | time, an epilog loop will be generated, and the loop induction variables | |
1671 | (IVs) will be "advanced" to the value they are supposed to take just before | |
ebfd146a IR |
1672 | the epilog loop. Here we check that the access function of the loop IVs |
1673 | and the expression that represents the loop bound are simple enough. | |
1674 | These restrictions will be relaxed in the future. */ | |
1675 | ||
b8698a0f | 1676 | bool |
ebfd146a IR |
1677 | vect_can_advance_ivs_p (loop_vec_info loop_vinfo) |
1678 | { | |
1679 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
1680 | basic_block bb = loop->header; | |
1681 | gimple phi; | |
1682 | gimple_stmt_iterator gsi; | |
1683 | ||
1684 | /* Analyze phi functions of the loop header. */ | |
1685 | ||
78c60e3d SS |
1686 | if (dump_kind_p (MSG_NOTE)) |
1687 | dump_printf_loc (MSG_NOTE, vect_location, "vect_can_advance_ivs_p:"); | |
ebfd146a IR |
1688 | for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
1689 | { | |
1690 | tree access_fn = NULL; | |
1691 | tree evolution_part; | |
1692 | ||
1693 | phi = gsi_stmt (gsi); | |
78c60e3d | 1694 | if (dump_kind_p (MSG_NOTE)) |
ebfd146a | 1695 | { |
78c60e3d SS |
1696 | dump_printf_loc (MSG_NOTE, vect_location, "Analyze phi: "); |
1697 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0); | |
ebfd146a IR |
1698 | } |
1699 | ||
1700 | /* Skip virtual phi's. The data dependences that are associated with | |
1701 | virtual defs/uses (i.e., memory accesses) are analyzed elsewhere. */ | |
1702 | ||
ea057359 | 1703 | if (virtual_operand_p (PHI_RESULT (phi))) |
ebfd146a | 1704 | { |
78c60e3d SS |
1705 | if (dump_kind_p (MSG_MISSED_OPTIMIZATION)) |
1706 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1707 | "virtual phi. skip."); | |
ebfd146a IR |
1708 | continue; |
1709 | } | |
1710 | ||
1711 | /* Skip reduction phis. */ | |
1712 | ||
1713 | if (STMT_VINFO_DEF_TYPE (vinfo_for_stmt (phi)) == vect_reduction_def) | |
1714 | { | |
78c60e3d SS |
1715 | if (dump_kind_p (MSG_MISSED_OPTIMIZATION)) |
1716 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1717 | "reduc phi. skip."); | |
ebfd146a IR |
1718 | continue; |
1719 | } | |
1720 | ||
1721 | /* Analyze the evolution function. */ | |
1722 | ||
1723 | access_fn = instantiate_parameters | |
1724 | (loop, analyze_scalar_evolution (loop, PHI_RESULT (phi))); | |
1725 | ||
1726 | if (!access_fn) | |
1727 | { | |
78c60e3d SS |
1728 | if (dump_kind_p (MSG_MISSED_OPTIMIZATION)) |
1729 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1730 | "No Access function."); | |
ebfd146a IR |
1731 | return false; |
1732 | } | |
1733 | ||
78c60e3d | 1734 | if (dump_kind_p (MSG_NOTE)) |
ebfd146a | 1735 | { |
78c60e3d SS |
1736 | dump_printf_loc (MSG_NOTE, vect_location, |
1737 | "Access function of PHI: "); | |
1738 | dump_generic_expr (MSG_NOTE, TDF_SLIM, access_fn); | |
ebfd146a IR |
1739 | } |
1740 | ||
1741 | evolution_part = evolution_part_in_loop_num (access_fn, loop->num); | |
b8698a0f | 1742 | |
ebfd146a IR |
1743 | if (evolution_part == NULL_TREE) |
1744 | { | |
78c60e3d SS |
1745 | if (dump_kind_p (MSG_MISSED_OPTIMIZATION)) |
1746 | dump_printf (MSG_MISSED_OPTIMIZATION, "No evolution."); | |
ebfd146a IR |
1747 | return false; |
1748 | } | |
b8698a0f L |
1749 | |
1750 | /* FORNOW: We do not transform initial conditions of IVs | |
ebfd146a IR |
1751 | which evolution functions are a polynomial of degree >= 2. */ |
1752 | ||
1753 | if (tree_is_chrec (evolution_part)) | |
b8698a0f | 1754 | return false; |
ebfd146a IR |
1755 | } |
1756 | ||
1757 | return true; | |
1758 | } | |
1759 | ||
1760 | ||
1761 | /* Function vect_update_ivs_after_vectorizer. | |
1762 | ||
1763 | "Advance" the induction variables of LOOP to the value they should take | |
1764 | after the execution of LOOP. This is currently necessary because the | |
1765 | vectorizer does not handle induction variables that are used after the | |
1766 | loop. Such a situation occurs when the last iterations of LOOP are | |
1767 | peeled, because: | |
1768 | 1. We introduced new uses after LOOP for IVs that were not originally used | |
1769 | after LOOP: the IVs of LOOP are now used by an epilog loop. | |
1770 | 2. LOOP is going to be vectorized; this means that it will iterate N/VF | |
1771 | times, whereas the loop IVs should be bumped N times. | |
1772 | ||
1773 | Input: | |
1774 | - LOOP - a loop that is going to be vectorized. The last few iterations | |
1775 | of LOOP were peeled. | |
1776 | - NITERS - the number of iterations that LOOP executes (before it is | |
1777 | vectorized). i.e, the number of times the ivs should be bumped. | |
1778 | - UPDATE_E - a successor edge of LOOP->exit that is on the (only) path | |
1779 | coming out from LOOP on which there are uses of the LOOP ivs | |
1780 | (this is the path from LOOP->exit to epilog_loop->preheader). | |
1781 | ||
1782 | The new definitions of the ivs are placed in LOOP->exit. | |
1783 | The phi args associated with the edge UPDATE_E in the bb | |
1784 | UPDATE_E->dest are updated accordingly. | |
1785 | ||
1786 | Assumption 1: Like the rest of the vectorizer, this function assumes | |
1787 | a single loop exit that has a single predecessor. | |
1788 | ||
1789 | Assumption 2: The phi nodes in the LOOP header and in update_bb are | |
1790 | organized in the same order. | |
1791 | ||
1792 | Assumption 3: The access function of the ivs is simple enough (see | |
1793 | vect_can_advance_ivs_p). This assumption will be relaxed in the future. | |
1794 | ||
1795 | Assumption 4: Exactly one of the successors of LOOP exit-bb is on a path | |
b8698a0f | 1796 | coming out of LOOP on which the ivs of LOOP are used (this is the path |
ebfd146a IR |
1797 | that leads to the epilog loop; other paths skip the epilog loop). This |
1798 | path starts with the edge UPDATE_E, and its destination (denoted update_bb) | |
1799 | needs to have its phis updated. | |
1800 | */ | |
1801 | ||
1802 | static void | |
b8698a0f | 1803 | vect_update_ivs_after_vectorizer (loop_vec_info loop_vinfo, tree niters, |
ebfd146a IR |
1804 | edge update_e) |
1805 | { | |
1806 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
1807 | basic_block exit_bb = single_exit (loop)->dest; | |
1808 | gimple phi, phi1; | |
1809 | gimple_stmt_iterator gsi, gsi1; | |
1810 | basic_block update_bb = update_e->dest; | |
1811 | ||
1812 | /* gcc_assert (vect_can_advance_ivs_p (loop_vinfo)); */ | |
1813 | ||
1814 | /* Make sure there exists a single-predecessor exit bb: */ | |
1815 | gcc_assert (single_pred_p (exit_bb)); | |
1816 | ||
1817 | for (gsi = gsi_start_phis (loop->header), gsi1 = gsi_start_phis (update_bb); | |
1818 | !gsi_end_p (gsi) && !gsi_end_p (gsi1); | |
1819 | gsi_next (&gsi), gsi_next (&gsi1)) | |
1820 | { | |
ebfd146a | 1821 | tree init_expr; |
550918ca RG |
1822 | tree step_expr, off; |
1823 | tree type; | |
ebfd146a IR |
1824 | tree var, ni, ni_name; |
1825 | gimple_stmt_iterator last_gsi; | |
0ac168a1 | 1826 | stmt_vec_info stmt_info; |
ebfd146a IR |
1827 | |
1828 | phi = gsi_stmt (gsi); | |
1829 | phi1 = gsi_stmt (gsi1); | |
78c60e3d | 1830 | if (dump_kind_p (MSG_NOTE)) |
ebfd146a | 1831 | { |
78c60e3d SS |
1832 | dump_printf_loc (MSG_NOTE, vect_location, |
1833 | "vect_update_ivs_after_vectorizer: phi: "); | |
1834 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0); | |
ebfd146a IR |
1835 | } |
1836 | ||
1837 | /* Skip virtual phi's. */ | |
ea057359 | 1838 | if (virtual_operand_p (PHI_RESULT (phi))) |
ebfd146a | 1839 | { |
78c60e3d SS |
1840 | if (dump_kind_p (MSG_MISSED_OPTIMIZATION)) |
1841 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1842 | "virtual phi. skip."); | |
ebfd146a IR |
1843 | continue; |
1844 | } | |
1845 | ||
1846 | /* Skip reduction phis. */ | |
0ac168a1 RG |
1847 | stmt_info = vinfo_for_stmt (phi); |
1848 | if (STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def) | |
b8698a0f | 1849 | { |
78c60e3d SS |
1850 | if (dump_kind_p (MSG_MISSED_OPTIMIZATION)) |
1851 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1852 | "reduc phi. skip."); | |
ebfd146a | 1853 | continue; |
b8698a0f | 1854 | } |
ebfd146a | 1855 | |
0ac168a1 RG |
1856 | type = TREE_TYPE (gimple_phi_result (phi)); |
1857 | step_expr = STMT_VINFO_LOOP_PHI_EVOLUTION_PART (stmt_info); | |
1858 | step_expr = unshare_expr (step_expr); | |
b8698a0f | 1859 | |
ebfd146a IR |
1860 | /* FORNOW: We do not support IVs whose evolution function is a polynomial |
1861 | of degree >= 2 or exponential. */ | |
0ac168a1 | 1862 | gcc_assert (!tree_is_chrec (step_expr)); |
ebfd146a | 1863 | |
0ac168a1 | 1864 | init_expr = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop)); |
ebfd146a | 1865 | |
550918ca RG |
1866 | off = fold_build2 (MULT_EXPR, TREE_TYPE (step_expr), |
1867 | fold_convert (TREE_TYPE (step_expr), niters), | |
1868 | step_expr); | |
0ac168a1 | 1869 | if (POINTER_TYPE_P (type)) |
5d49b6a7 | 1870 | ni = fold_build_pointer_plus (init_expr, off); |
ebfd146a | 1871 | else |
0ac168a1 RG |
1872 | ni = fold_build2 (PLUS_EXPR, type, |
1873 | init_expr, fold_convert (type, off)); | |
ebfd146a | 1874 | |
0ac168a1 | 1875 | var = create_tmp_var (type, "tmp"); |
ebfd146a IR |
1876 | |
1877 | last_gsi = gsi_last_bb (exit_bb); | |
1878 | ni_name = force_gimple_operand_gsi (&last_gsi, ni, false, var, | |
1879 | true, GSI_SAME_STMT); | |
b8698a0f | 1880 | |
ebfd146a | 1881 | /* Fix phi expressions in the successor bb. */ |
684f25f4 | 1882 | adjust_phi_and_debug_stmts (phi1, update_e, ni_name); |
ebfd146a IR |
1883 | } |
1884 | } | |
1885 | ||
ebfd146a IR |
1886 | /* Function vect_do_peeling_for_loop_bound |
1887 | ||
1888 | Peel the last iterations of the loop represented by LOOP_VINFO. | |
b8698a0f | 1889 | The peeled iterations form a new epilog loop. Given that the loop now |
ebfd146a IR |
1890 | iterates NITERS times, the new epilog loop iterates |
1891 | NITERS % VECTORIZATION_FACTOR times. | |
b8698a0f L |
1892 | |
1893 | The original loop will later be made to iterate | |
86290011 RG |
1894 | NITERS / VECTORIZATION_FACTOR times (this value is placed into RATIO). |
1895 | ||
1896 | COND_EXPR and COND_EXPR_STMT_LIST are combined with a new generated | |
1897 | test. */ | |
ebfd146a | 1898 | |
b8698a0f | 1899 | void |
86290011 | 1900 | vect_do_peeling_for_loop_bound (loop_vec_info loop_vinfo, tree *ratio, |
368117e8 | 1901 | unsigned int th, bool check_profitability) |
ebfd146a IR |
1902 | { |
1903 | tree ni_name, ratio_mult_vf_name; | |
1904 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
1905 | struct loop *new_loop; | |
1906 | edge update_e; | |
1907 | basic_block preheader; | |
1908 | int loop_num; | |
d68d56b5 | 1909 | int max_iter; |
368117e8 RG |
1910 | tree cond_expr = NULL_TREE; |
1911 | gimple_seq cond_expr_stmt_list = NULL; | |
ebfd146a | 1912 | |
78c60e3d SS |
1913 | if (dump_kind_p (MSG_OPTIMIZED_LOCATIONS)) |
1914 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, vect_location, | |
1915 | "=== vect_do_peeling_for_loop_bound ==="); | |
ebfd146a IR |
1916 | |
1917 | initialize_original_copy_tables (); | |
1918 | ||
1919 | /* Generate the following variables on the preheader of original loop: | |
b8698a0f | 1920 | |
ebfd146a IR |
1921 | ni_name = number of iteration the original loop executes |
1922 | ratio = ni_name / vf | |
1923 | ratio_mult_vf_name = ratio * vf */ | |
1924 | vect_generate_tmps_on_preheader (loop_vinfo, &ni_name, | |
86290011 RG |
1925 | &ratio_mult_vf_name, ratio, |
1926 | cond_expr_stmt_list); | |
ebfd146a | 1927 | |
b8698a0f | 1928 | loop_num = loop->num; |
ebfd146a | 1929 | |
ebfd146a | 1930 | new_loop = slpeel_tree_peel_loop_to_edge (loop, single_exit (loop), |
5d2eb24b | 1931 | &ratio_mult_vf_name, ni_name, false, |
86290011 | 1932 | th, check_profitability, |
e78410bf JH |
1933 | cond_expr, cond_expr_stmt_list, |
1934 | 0, LOOP_VINFO_VECT_FACTOR (loop_vinfo)); | |
ebfd146a IR |
1935 | gcc_assert (new_loop); |
1936 | gcc_assert (loop_num == loop->num); | |
1937 | #ifdef ENABLE_CHECKING | |
1938 | slpeel_verify_cfg_after_peeling (loop, new_loop); | |
1939 | #endif | |
1940 | ||
1941 | /* A guard that controls whether the new_loop is to be executed or skipped | |
1942 | is placed in LOOP->exit. LOOP->exit therefore has two successors - one | |
1943 | is the preheader of NEW_LOOP, where the IVs from LOOP are used. The other | |
1944 | is a bb after NEW_LOOP, where these IVs are not used. Find the edge that | |
1945 | is on the path where the LOOP IVs are used and need to be updated. */ | |
1946 | ||
1947 | preheader = loop_preheader_edge (new_loop)->src; | |
1948 | if (EDGE_PRED (preheader, 0)->src == single_exit (loop)->dest) | |
1949 | update_e = EDGE_PRED (preheader, 0); | |
1950 | else | |
1951 | update_e = EDGE_PRED (preheader, 1); | |
1952 | ||
b8698a0f | 1953 | /* Update IVs of original loop as if they were advanced |
ebfd146a | 1954 | by ratio_mult_vf_name steps. */ |
b8698a0f | 1955 | vect_update_ivs_after_vectorizer (loop_vinfo, ratio_mult_vf_name, update_e); |
ebfd146a | 1956 | |
368117e8 RG |
1957 | max_iter = LOOP_VINFO_VECT_FACTOR (loop_vinfo) - 1; |
1958 | if (check_profitability) | |
1959 | max_iter = MAX (max_iter, (int) th); | |
27bcd47c | 1960 | record_niter_bound (new_loop, double_int::from_shwi (max_iter), false, true); |
78c60e3d SS |
1961 | dump_printf (MSG_OPTIMIZED_LOCATIONS, |
1962 | "Setting upper bound of nb iterations for epilogue " | |
1963 | "loop to %d\n", max_iter); | |
7d5a99f4 | 1964 | |
ebfd146a IR |
1965 | /* After peeling we have to reset scalar evolution analyzer. */ |
1966 | scev_reset (); | |
1967 | ||
1968 | free_original_copy_tables (); | |
1969 | } | |
1970 | ||
1971 | ||
1972 | /* Function vect_gen_niters_for_prolog_loop | |
1973 | ||
1974 | Set the number of iterations for the loop represented by LOOP_VINFO | |
1975 | to the minimum between LOOP_NITERS (the original iteration count of the loop) | |
1976 | and the misalignment of DR - the data reference recorded in | |
b8698a0f | 1977 | LOOP_VINFO_UNALIGNED_DR (LOOP_VINFO). As a result, after the execution of |
ebfd146a IR |
1978 | this loop, the data reference DR will refer to an aligned location. |
1979 | ||
1980 | The following computation is generated: | |
1981 | ||
1982 | If the misalignment of DR is known at compile time: | |
1983 | addr_mis = int mis = DR_MISALIGNMENT (dr); | |
1984 | Else, compute address misalignment in bytes: | |
5aea1e76 | 1985 | addr_mis = addr & (vectype_align - 1) |
ebfd146a IR |
1986 | |
1987 | prolog_niters = min (LOOP_NITERS, ((VF - addr_mis/elem_size)&(VF-1))/step) | |
1988 | ||
1989 | (elem_size = element type size; an element is the scalar element whose type | |
1990 | is the inner type of the vectype) | |
1991 | ||
1992 | When the step of the data-ref in the loop is not 1 (as in interleaved data | |
1993 | and SLP), the number of iterations of the prolog must be divided by the step | |
1994 | (which is equal to the size of interleaved group). | |
1995 | ||
1996 | The above formulas assume that VF == number of elements in the vector. This | |
1997 | may not hold when there are multiple-types in the loop. | |
1998 | In this case, for some data-references in the loop the VF does not represent | |
1999 | the number of elements that fit in the vector. Therefore, instead of VF we | |
2000 | use TYPE_VECTOR_SUBPARTS. */ | |
2001 | ||
b8698a0f | 2002 | static tree |
e78410bf | 2003 | vect_gen_niters_for_prolog_loop (loop_vec_info loop_vinfo, tree loop_niters, int *bound) |
ebfd146a IR |
2004 | { |
2005 | struct data_reference *dr = LOOP_VINFO_UNALIGNED_DR (loop_vinfo); | |
2006 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
2007 | tree var; | |
2008 | gimple_seq stmts; | |
2009 | tree iters, iters_name; | |
2010 | edge pe; | |
2011 | basic_block new_bb; | |
2012 | gimple dr_stmt = DR_STMT (dr); | |
2013 | stmt_vec_info stmt_info = vinfo_for_stmt (dr_stmt); | |
2014 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2015 | int vectype_align = TYPE_ALIGN (vectype) / BITS_PER_UNIT; | |
2016 | tree niters_type = TREE_TYPE (loop_niters); | |
ebfd146a IR |
2017 | int nelements = TYPE_VECTOR_SUBPARTS (vectype); |
2018 | ||
b8698a0f | 2019 | pe = loop_preheader_edge (loop); |
ebfd146a IR |
2020 | |
2021 | if (LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo) > 0) | |
2022 | { | |
720f5239 | 2023 | int npeel = LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo); |
ebfd146a | 2024 | |
78c60e3d SS |
2025 | if (dump_kind_p (MSG_OPTIMIZED_LOCATIONS)) |
2026 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, vect_location, | |
2027 | "known peeling = %d.", npeel); | |
ebfd146a | 2028 | |
720f5239 | 2029 | iters = build_int_cst (niters_type, npeel); |
e78410bf | 2030 | *bound = LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo); |
ebfd146a IR |
2031 | } |
2032 | else | |
2033 | { | |
2034 | gimple_seq new_stmts = NULL; | |
d8ba5b19 RG |
2035 | bool negative = tree_int_cst_compare (DR_STEP (dr), size_zero_node) < 0; |
2036 | tree offset = negative | |
2037 | ? size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1) : NULL_TREE; | |
b8698a0f | 2038 | tree start_addr = vect_create_addr_base_for_vector_ref (dr_stmt, |
d8ba5b19 | 2039 | &new_stmts, offset, loop); |
96f9265a | 2040 | tree type = unsigned_type_for (TREE_TYPE (start_addr)); |
5aea1e76 UW |
2041 | tree vectype_align_minus_1 = build_int_cst (type, vectype_align - 1); |
2042 | HOST_WIDE_INT elem_size = | |
2043 | int_cst_value (TYPE_SIZE_UNIT (TREE_TYPE (vectype))); | |
2044 | tree elem_size_log = build_int_cst (type, exact_log2 (elem_size)); | |
ebfd146a IR |
2045 | tree nelements_minus_1 = build_int_cst (type, nelements - 1); |
2046 | tree nelements_tree = build_int_cst (type, nelements); | |
2047 | tree byte_misalign; | |
2048 | tree elem_misalign; | |
2049 | ||
2050 | new_bb = gsi_insert_seq_on_edge_immediate (pe, new_stmts); | |
2051 | gcc_assert (!new_bb); | |
b8698a0f | 2052 | |
5aea1e76 | 2053 | /* Create: byte_misalign = addr & (vectype_align - 1) */ |
b8698a0f | 2054 | byte_misalign = |
720f5239 | 2055 | fold_build2 (BIT_AND_EXPR, type, fold_convert (type, start_addr), |
5aea1e76 | 2056 | vectype_align_minus_1); |
b8698a0f | 2057 | |
ebfd146a IR |
2058 | /* Create: elem_misalign = byte_misalign / element_size */ |
2059 | elem_misalign = | |
2060 | fold_build2 (RSHIFT_EXPR, type, byte_misalign, elem_size_log); | |
2061 | ||
2062 | /* Create: (niters_type) (nelements - elem_misalign)&(nelements - 1) */ | |
d8ba5b19 RG |
2063 | if (negative) |
2064 | iters = fold_build2 (MINUS_EXPR, type, elem_misalign, nelements_tree); | |
2065 | else | |
2066 | iters = fold_build2 (MINUS_EXPR, type, nelements_tree, elem_misalign); | |
ebfd146a IR |
2067 | iters = fold_build2 (BIT_AND_EXPR, type, iters, nelements_minus_1); |
2068 | iters = fold_convert (niters_type, iters); | |
e78410bf | 2069 | *bound = nelements; |
ebfd146a IR |
2070 | } |
2071 | ||
2072 | /* Create: prolog_loop_niters = min (iters, loop_niters) */ | |
2073 | /* If the loop bound is known at compile time we already verified that it is | |
2074 | greater than vf; since the misalignment ('iters') is at most vf, there's | |
2075 | no need to generate the MIN_EXPR in this case. */ | |
2076 | if (TREE_CODE (loop_niters) != INTEGER_CST) | |
2077 | iters = fold_build2 (MIN_EXPR, niters_type, iters, loop_niters); | |
2078 | ||
78c60e3d | 2079 | if (dump_kind_p (MSG_OPTIMIZED_LOCATIONS)) |
ebfd146a | 2080 | { |
78c60e3d SS |
2081 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, vect_location, |
2082 | "niters for prolog loop: "); | |
2083 | dump_generic_expr (MSG_OPTIMIZED_LOCATIONS, TDF_SLIM, iters); | |
ebfd146a IR |
2084 | } |
2085 | ||
2086 | var = create_tmp_var (niters_type, "prolog_loop_niters"); | |
ebfd146a IR |
2087 | stmts = NULL; |
2088 | iters_name = force_gimple_operand (iters, &stmts, false, var); | |
2089 | ||
2090 | /* Insert stmt on loop preheader edge. */ | |
2091 | if (stmts) | |
2092 | { | |
2093 | basic_block new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
2094 | gcc_assert (!new_bb); | |
2095 | } | |
2096 | ||
b8698a0f | 2097 | return iters_name; |
ebfd146a IR |
2098 | } |
2099 | ||
2100 | ||
2101 | /* Function vect_update_init_of_dr | |
2102 | ||
2103 | NITERS iterations were peeled from LOOP. DR represents a data reference | |
2104 | in LOOP. This function updates the information recorded in DR to | |
b8698a0f | 2105 | account for the fact that the first NITERS iterations had already been |
ebfd146a IR |
2106 | executed. Specifically, it updates the OFFSET field of DR. */ |
2107 | ||
2108 | static void | |
2109 | vect_update_init_of_dr (struct data_reference *dr, tree niters) | |
2110 | { | |
2111 | tree offset = DR_OFFSET (dr); | |
b8698a0f | 2112 | |
ebfd146a IR |
2113 | niters = fold_build2 (MULT_EXPR, sizetype, |
2114 | fold_convert (sizetype, niters), | |
2115 | fold_convert (sizetype, DR_STEP (dr))); | |
587aa063 RG |
2116 | offset = fold_build2 (PLUS_EXPR, sizetype, |
2117 | fold_convert (sizetype, offset), niters); | |
ebfd146a IR |
2118 | DR_OFFSET (dr) = offset; |
2119 | } | |
2120 | ||
2121 | ||
2122 | /* Function vect_update_inits_of_drs | |
2123 | ||
b8698a0f L |
2124 | NITERS iterations were peeled from the loop represented by LOOP_VINFO. |
2125 | This function updates the information recorded for the data references in | |
2126 | the loop to account for the fact that the first NITERS iterations had | |
ebfd146a IR |
2127 | already been executed. Specifically, it updates the initial_condition of |
2128 | the access_function of all the data_references in the loop. */ | |
2129 | ||
2130 | static void | |
2131 | vect_update_inits_of_drs (loop_vec_info loop_vinfo, tree niters) | |
2132 | { | |
2133 | unsigned int i; | |
2134 | VEC (data_reference_p, heap) *datarefs = LOOP_VINFO_DATAREFS (loop_vinfo); | |
2135 | struct data_reference *dr; | |
78c60e3d SS |
2136 | |
2137 | if (dump_kind_p (MSG_OPTIMIZED_LOCATIONS)) | |
2138 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, vect_location, | |
2139 | "=== vect_update_inits_of_dr ==="); | |
ebfd146a | 2140 | |
ac47786e | 2141 | FOR_EACH_VEC_ELT (data_reference_p, datarefs, i, dr) |
ebfd146a IR |
2142 | vect_update_init_of_dr (dr, niters); |
2143 | } | |
2144 | ||
2145 | ||
2146 | /* Function vect_do_peeling_for_alignment | |
2147 | ||
2148 | Peel the first 'niters' iterations of the loop represented by LOOP_VINFO. | |
2149 | 'niters' is set to the misalignment of one of the data references in the | |
2150 | loop, thereby forcing it to refer to an aligned location at the beginning | |
2151 | of the execution of this loop. The data reference for which we are | |
2152 | peeling is recorded in LOOP_VINFO_UNALIGNED_DR. */ | |
2153 | ||
2154 | void | |
368117e8 RG |
2155 | vect_do_peeling_for_alignment (loop_vec_info loop_vinfo, |
2156 | unsigned int th, bool check_profitability) | |
ebfd146a IR |
2157 | { |
2158 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
2159 | tree niters_of_prolog_loop, ni_name; | |
2160 | tree n_iters; | |
b61b1f17 | 2161 | tree wide_prolog_niters; |
ebfd146a | 2162 | struct loop *new_loop; |
03fd03d5 | 2163 | int max_iter; |
e78410bf | 2164 | int bound = 0; |
ebfd146a | 2165 | |
78c60e3d SS |
2166 | if (dump_kind_p (MSG_OPTIMIZED_LOCATIONS)) |
2167 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, vect_location, | |
2168 | "=== vect_do_peeling_for_alignment ==="); | |
ebfd146a IR |
2169 | |
2170 | initialize_original_copy_tables (); | |
2171 | ||
86290011 | 2172 | ni_name = vect_build_loop_niters (loop_vinfo, NULL); |
5d2eb24b | 2173 | niters_of_prolog_loop = vect_gen_niters_for_prolog_loop (loop_vinfo, |
e78410bf JH |
2174 | ni_name, |
2175 | &bound); | |
ebfd146a | 2176 | |
ebfd146a IR |
2177 | /* Peel the prolog loop and iterate it niters_of_prolog_loop. */ |
2178 | new_loop = | |
2179 | slpeel_tree_peel_loop_to_edge (loop, loop_preheader_edge (loop), | |
5d2eb24b | 2180 | &niters_of_prolog_loop, ni_name, true, |
e78410bf JH |
2181 | th, check_profitability, NULL_TREE, NULL, |
2182 | bound, | |
2183 | 0); | |
ebfd146a IR |
2184 | |
2185 | gcc_assert (new_loop); | |
2186 | #ifdef ENABLE_CHECKING | |
2187 | slpeel_verify_cfg_after_peeling (new_loop, loop); | |
2188 | #endif | |
368117e8 RG |
2189 | max_iter = LOOP_VINFO_VECT_FACTOR (loop_vinfo) - 1; |
2190 | if (check_profitability) | |
2191 | max_iter = MAX (max_iter, (int) th); | |
27bcd47c | 2192 | record_niter_bound (new_loop, double_int::from_shwi (max_iter), false, true); |
78c60e3d SS |
2193 | dump_printf (MSG_OPTIMIZED_LOCATIONS, |
2194 | "Setting upper bound of nb iterations for prologue " | |
2195 | "loop to %d\n", max_iter); | |
ebfd146a IR |
2196 | |
2197 | /* Update number of times loop executes. */ | |
2198 | n_iters = LOOP_VINFO_NITERS (loop_vinfo); | |
2199 | LOOP_VINFO_NITERS (loop_vinfo) = fold_build2 (MINUS_EXPR, | |
2200 | TREE_TYPE (n_iters), n_iters, niters_of_prolog_loop); | |
2201 | ||
5d2eb24b IR |
2202 | if (types_compatible_p (sizetype, TREE_TYPE (niters_of_prolog_loop))) |
2203 | wide_prolog_niters = niters_of_prolog_loop; | |
2204 | else | |
2205 | { | |
2206 | gimple_seq seq = NULL; | |
2207 | edge pe = loop_preheader_edge (loop); | |
2208 | tree wide_iters = fold_convert (sizetype, niters_of_prolog_loop); | |
2209 | tree var = create_tmp_var (sizetype, "prolog_loop_adjusted_niters"); | |
5d2eb24b IR |
2210 | wide_prolog_niters = force_gimple_operand (wide_iters, &seq, false, |
2211 | var); | |
2212 | if (seq) | |
2213 | { | |
2214 | /* Insert stmt on loop preheader edge. */ | |
2215 | basic_block new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
2216 | gcc_assert (!new_bb); | |
2217 | } | |
2218 | } | |
2219 | ||
ebfd146a | 2220 | /* Update the init conditions of the access functions of all data refs. */ |
b61b1f17 | 2221 | vect_update_inits_of_drs (loop_vinfo, wide_prolog_niters); |
ebfd146a IR |
2222 | |
2223 | /* After peeling we have to reset scalar evolution analyzer. */ | |
2224 | scev_reset (); | |
2225 | ||
2226 | free_original_copy_tables (); | |
2227 | } | |
2228 | ||
2229 | ||
2230 | /* Function vect_create_cond_for_align_checks. | |
2231 | ||
2232 | Create a conditional expression that represents the alignment checks for | |
2233 | all of data references (array element references) whose alignment must be | |
2234 | checked at runtime. | |
2235 | ||
2236 | Input: | |
2237 | COND_EXPR - input conditional expression. New conditions will be chained | |
2238 | with logical AND operation. | |
2239 | LOOP_VINFO - two fields of the loop information are used. | |
2240 | LOOP_VINFO_PTR_MASK is the mask used to check the alignment. | |
2241 | LOOP_VINFO_MAY_MISALIGN_STMTS contains the refs to be checked. | |
2242 | ||
2243 | Output: | |
2244 | COND_EXPR_STMT_LIST - statements needed to construct the conditional | |
2245 | expression. | |
2246 | The returned value is the conditional expression to be used in the if | |
2247 | statement that controls which version of the loop gets executed at runtime. | |
2248 | ||
2249 | The algorithm makes two assumptions: | |
2250 | 1) The number of bytes "n" in a vector is a power of 2. | |
2251 | 2) An address "a" is aligned if a%n is zero and that this | |
2252 | test can be done as a&(n-1) == 0. For example, for 16 | |
2253 | byte vectors the test is a&0xf == 0. */ | |
2254 | ||
2255 | static void | |
2256 | vect_create_cond_for_align_checks (loop_vec_info loop_vinfo, | |
2257 | tree *cond_expr, | |
2258 | gimple_seq *cond_expr_stmt_list) | |
2259 | { | |
2260 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
2261 | VEC(gimple,heap) *may_misalign_stmts | |
2262 | = LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo); | |
2263 | gimple ref_stmt; | |
2264 | int mask = LOOP_VINFO_PTR_MASK (loop_vinfo); | |
2265 | tree mask_cst; | |
2266 | unsigned int i; | |
ebfd146a IR |
2267 | tree int_ptrsize_type; |
2268 | char tmp_name[20]; | |
2269 | tree or_tmp_name = NULL_TREE; | |
83d5977e | 2270 | tree and_tmp_name; |
ebfd146a IR |
2271 | gimple and_stmt; |
2272 | tree ptrsize_zero; | |
2273 | tree part_cond_expr; | |
2274 | ||
2275 | /* Check that mask is one less than a power of 2, i.e., mask is | |
2276 | all zeros followed by all ones. */ | |
2277 | gcc_assert ((mask != 0) && ((mask & (mask+1)) == 0)); | |
2278 | ||
96f9265a | 2279 | int_ptrsize_type = signed_type_for (ptr_type_node); |
ebfd146a IR |
2280 | |
2281 | /* Create expression (mask & (dr_1 || ... || dr_n)) where dr_i is the address | |
2282 | of the first vector of the i'th data reference. */ | |
2283 | ||
ac47786e | 2284 | FOR_EACH_VEC_ELT (gimple, may_misalign_stmts, i, ref_stmt) |
ebfd146a IR |
2285 | { |
2286 | gimple_seq new_stmt_list = NULL; | |
2287 | tree addr_base; | |
83d5977e RG |
2288 | tree addr_tmp_name; |
2289 | tree new_or_tmp_name; | |
ebfd146a | 2290 | gimple addr_stmt, or_stmt; |
d8ba5b19 RG |
2291 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (ref_stmt); |
2292 | tree vectype = STMT_VINFO_VECTYPE (stmt_vinfo); | |
2293 | bool negative = tree_int_cst_compare | |
2294 | (DR_STEP (STMT_VINFO_DATA_REF (stmt_vinfo)), size_zero_node) < 0; | |
2295 | tree offset = negative | |
2296 | ? size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1) : NULL_TREE; | |
ebfd146a IR |
2297 | |
2298 | /* create: addr_tmp = (int)(address_of_first_vector) */ | |
2299 | addr_base = | |
2300 | vect_create_addr_base_for_vector_ref (ref_stmt, &new_stmt_list, | |
d8ba5b19 | 2301 | offset, loop); |
ebfd146a IR |
2302 | if (new_stmt_list != NULL) |
2303 | gimple_seq_add_seq (cond_expr_stmt_list, new_stmt_list); | |
2304 | ||
83d5977e RG |
2305 | sprintf (tmp_name, "addr2int%d", i); |
2306 | addr_tmp_name = make_temp_ssa_name (int_ptrsize_type, NULL, tmp_name); | |
ebfd146a IR |
2307 | addr_stmt = gimple_build_assign_with_ops (NOP_EXPR, addr_tmp_name, |
2308 | addr_base, NULL_TREE); | |
ebfd146a IR |
2309 | gimple_seq_add_stmt (cond_expr_stmt_list, addr_stmt); |
2310 | ||
2311 | /* The addresses are OR together. */ | |
2312 | ||
2313 | if (or_tmp_name != NULL_TREE) | |
2314 | { | |
2315 | /* create: or_tmp = or_tmp | addr_tmp */ | |
83d5977e RG |
2316 | sprintf (tmp_name, "orptrs%d", i); |
2317 | new_or_tmp_name = make_temp_ssa_name (int_ptrsize_type, NULL, tmp_name); | |
ebfd146a IR |
2318 | or_stmt = gimple_build_assign_with_ops (BIT_IOR_EXPR, |
2319 | new_or_tmp_name, | |
2320 | or_tmp_name, addr_tmp_name); | |
ebfd146a IR |
2321 | gimple_seq_add_stmt (cond_expr_stmt_list, or_stmt); |
2322 | or_tmp_name = new_or_tmp_name; | |
2323 | } | |
2324 | else | |
2325 | or_tmp_name = addr_tmp_name; | |
2326 | ||
2327 | } /* end for i */ | |
2328 | ||
2329 | mask_cst = build_int_cst (int_ptrsize_type, mask); | |
2330 | ||
2331 | /* create: and_tmp = or_tmp & mask */ | |
83d5977e | 2332 | and_tmp_name = make_temp_ssa_name (int_ptrsize_type, NULL, "andmask"); |
ebfd146a IR |
2333 | |
2334 | and_stmt = gimple_build_assign_with_ops (BIT_AND_EXPR, and_tmp_name, | |
2335 | or_tmp_name, mask_cst); | |
ebfd146a IR |
2336 | gimple_seq_add_stmt (cond_expr_stmt_list, and_stmt); |
2337 | ||
2338 | /* Make and_tmp the left operand of the conditional test against zero. | |
2339 | if and_tmp has a nonzero bit then some address is unaligned. */ | |
2340 | ptrsize_zero = build_int_cst (int_ptrsize_type, 0); | |
2341 | part_cond_expr = fold_build2 (EQ_EXPR, boolean_type_node, | |
2342 | and_tmp_name, ptrsize_zero); | |
2343 | if (*cond_expr) | |
2344 | *cond_expr = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, | |
2345 | *cond_expr, part_cond_expr); | |
2346 | else | |
2347 | *cond_expr = part_cond_expr; | |
2348 | } | |
2349 | ||
2350 | ||
2351 | /* Function vect_vfa_segment_size. | |
2352 | ||
2353 | Create an expression that computes the size of segment | |
2354 | that will be accessed for a data reference. The functions takes into | |
2355 | account that realignment loads may access one more vector. | |
2356 | ||
2357 | Input: | |
2358 | DR: The data reference. | |
208cb8cb | 2359 | LENGTH_FACTOR: segment length to consider. |
ebfd146a IR |
2360 | |
2361 | Return an expression whose value is the size of segment which will be | |
2362 | accessed by DR. */ | |
2363 | ||
2364 | static tree | |
208cb8cb | 2365 | vect_vfa_segment_size (struct data_reference *dr, tree length_factor) |
ebfd146a | 2366 | { |
e2a3a5f1 | 2367 | tree segment_length; |
338f655d | 2368 | |
319e6439 | 2369 | if (integer_zerop (DR_STEP (dr))) |
338f655d IR |
2370 | segment_length = TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dr))); |
2371 | else | |
2372 | segment_length = size_binop (MULT_EXPR, | |
2373 | fold_convert (sizetype, DR_STEP (dr)), | |
2374 | fold_convert (sizetype, length_factor)); | |
2375 | ||
720f5239 IR |
2376 | if (vect_supportable_dr_alignment (dr, false) |
2377 | == dr_explicit_realign_optimized) | |
ebfd146a IR |
2378 | { |
2379 | tree vector_size = TYPE_SIZE_UNIT | |
2380 | (STMT_VINFO_VECTYPE (vinfo_for_stmt (DR_STMT (dr)))); | |
2381 | ||
e2a3a5f1 | 2382 | segment_length = size_binop (PLUS_EXPR, segment_length, vector_size); |
ebfd146a | 2383 | } |
e2a3a5f1 | 2384 | return segment_length; |
ebfd146a IR |
2385 | } |
2386 | ||
2387 | ||
2388 | /* Function vect_create_cond_for_alias_checks. | |
2389 | ||
2390 | Create a conditional expression that represents the run-time checks for | |
2391 | overlapping of address ranges represented by a list of data references | |
2392 | relations passed as input. | |
2393 | ||
2394 | Input: | |
2395 | COND_EXPR - input conditional expression. New conditions will be chained | |
2396 | with logical AND operation. | |
2397 | LOOP_VINFO - field LOOP_VINFO_MAY_ALIAS_STMTS contains the list of ddrs | |
2398 | to be checked. | |
2399 | ||
2400 | Output: | |
2401 | COND_EXPR - conditional expression. | |
2402 | COND_EXPR_STMT_LIST - statements needed to construct the conditional | |
2403 | expression. | |
2404 | ||
2405 | ||
2406 | The returned value is the conditional expression to be used in the if | |
2407 | statement that controls which version of the loop gets executed at runtime. | |
2408 | */ | |
2409 | ||
2410 | static void | |
2411 | vect_create_cond_for_alias_checks (loop_vec_info loop_vinfo, | |
2412 | tree * cond_expr, | |
2413 | gimple_seq * cond_expr_stmt_list) | |
2414 | { | |
2415 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
2416 | VEC (ddr_p, heap) * may_alias_ddrs = | |
2417 | LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo); | |
e2a3a5f1 RG |
2418 | int vect_factor = LOOP_VINFO_VECT_FACTOR (loop_vinfo); |
2419 | tree scalar_loop_iters = LOOP_VINFO_NITERS (loop_vinfo); | |
ebfd146a IR |
2420 | |
2421 | ddr_p ddr; | |
2422 | unsigned int i; | |
208cb8cb | 2423 | tree part_cond_expr, length_factor; |
ebfd146a IR |
2424 | |
2425 | /* Create expression | |
36fc3799 RS |
2426 | ((store_ptr_0 + store_segment_length_0) <= load_ptr_0) |
2427 | || (load_ptr_0 + load_segment_length_0) <= store_ptr_0)) | |
b8698a0f | 2428 | && |
ebfd146a IR |
2429 | ... |
2430 | && | |
36fc3799 RS |
2431 | ((store_ptr_n + store_segment_length_n) <= load_ptr_n) |
2432 | || (load_ptr_n + load_segment_length_n) <= store_ptr_n)) */ | |
ebfd146a IR |
2433 | |
2434 | if (VEC_empty (ddr_p, may_alias_ddrs)) | |
2435 | return; | |
2436 | ||
ac47786e | 2437 | FOR_EACH_VEC_ELT (ddr_p, may_alias_ddrs, i, ddr) |
ebfd146a IR |
2438 | { |
2439 | struct data_reference *dr_a, *dr_b; | |
2440 | gimple dr_group_first_a, dr_group_first_b; | |
2441 | tree addr_base_a, addr_base_b; | |
2442 | tree segment_length_a, segment_length_b; | |
2443 | gimple stmt_a, stmt_b; | |
d8ba5b19 | 2444 | tree seg_a_min, seg_a_max, seg_b_min, seg_b_max; |
ebfd146a IR |
2445 | |
2446 | dr_a = DDR_A (ddr); | |
2447 | stmt_a = DR_STMT (DDR_A (ddr)); | |
e14c1050 | 2448 | dr_group_first_a = GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt_a)); |
ebfd146a IR |
2449 | if (dr_group_first_a) |
2450 | { | |
2451 | stmt_a = dr_group_first_a; | |
2452 | dr_a = STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt_a)); | |
2453 | } | |
2454 | ||
2455 | dr_b = DDR_B (ddr); | |
2456 | stmt_b = DR_STMT (DDR_B (ddr)); | |
e14c1050 | 2457 | dr_group_first_b = GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt_b)); |
ebfd146a IR |
2458 | if (dr_group_first_b) |
2459 | { | |
2460 | stmt_b = dr_group_first_b; | |
2461 | dr_b = STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt_b)); | |
2462 | } | |
2463 | ||
2464 | addr_base_a = | |
2465 | vect_create_addr_base_for_vector_ref (stmt_a, cond_expr_stmt_list, | |
2466 | NULL_TREE, loop); | |
2467 | addr_base_b = | |
2468 | vect_create_addr_base_for_vector_ref (stmt_b, cond_expr_stmt_list, | |
2469 | NULL_TREE, loop); | |
2470 | ||
208cb8cb RG |
2471 | if (!operand_equal_p (DR_STEP (dr_a), DR_STEP (dr_b), 0)) |
2472 | length_factor = scalar_loop_iters; | |
2473 | else | |
2474 | length_factor = size_int (vect_factor); | |
2475 | segment_length_a = vect_vfa_segment_size (dr_a, length_factor); | |
2476 | segment_length_b = vect_vfa_segment_size (dr_b, length_factor); | |
ebfd146a | 2477 | |
78c60e3d | 2478 | if (dump_kind_p (MSG_OPTIMIZED_LOCATIONS)) |
ebfd146a | 2479 | { |
78c60e3d SS |
2480 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, vect_location, |
2481 | "create runtime check for data references "); | |
2482 | dump_generic_expr (MSG_OPTIMIZED_LOCATIONS, TDF_SLIM, DR_REF (dr_a)); | |
2483 | dump_printf (MSG_OPTIMIZED_LOCATIONS, " and "); | |
2484 | dump_generic_expr (MSG_OPTIMIZED_LOCATIONS, TDF_SLIM, DR_REF (dr_b)); | |
ebfd146a IR |
2485 | } |
2486 | ||
d8ba5b19 | 2487 | seg_a_min = addr_base_a; |
5d49b6a7 | 2488 | seg_a_max = fold_build_pointer_plus (addr_base_a, segment_length_a); |
d8ba5b19 RG |
2489 | if (tree_int_cst_compare (DR_STEP (dr_a), size_zero_node) < 0) |
2490 | seg_a_min = seg_a_max, seg_a_max = addr_base_a; | |
2491 | ||
2492 | seg_b_min = addr_base_b; | |
5d49b6a7 | 2493 | seg_b_max = fold_build_pointer_plus (addr_base_b, segment_length_b); |
d8ba5b19 RG |
2494 | if (tree_int_cst_compare (DR_STEP (dr_b), size_zero_node) < 0) |
2495 | seg_b_min = seg_b_max, seg_b_max = addr_base_b; | |
ebfd146a | 2496 | |
b8698a0f | 2497 | part_cond_expr = |
ebfd146a | 2498 | fold_build2 (TRUTH_OR_EXPR, boolean_type_node, |
36fc3799 RS |
2499 | fold_build2 (LE_EXPR, boolean_type_node, seg_a_max, seg_b_min), |
2500 | fold_build2 (LE_EXPR, boolean_type_node, seg_b_max, seg_a_min)); | |
b8698a0f | 2501 | |
ebfd146a IR |
2502 | if (*cond_expr) |
2503 | *cond_expr = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, | |
2504 | *cond_expr, part_cond_expr); | |
2505 | else | |
2506 | *cond_expr = part_cond_expr; | |
2507 | } | |
ebfd146a | 2508 | |
78c60e3d SS |
2509 | if (dump_kind_p (MSG_OPTIMIZED_LOCATIONS)) |
2510 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, vect_location, | |
2511 | "created %u versioning for alias checks.\n", | |
2512 | VEC_length (ddr_p, may_alias_ddrs)); | |
ebfd146a IR |
2513 | } |
2514 | ||
2515 | ||
2516 | /* Function vect_loop_versioning. | |
b8698a0f | 2517 | |
ebfd146a IR |
2518 | If the loop has data references that may or may not be aligned or/and |
2519 | has data reference relations whose independence was not proven then | |
2520 | two versions of the loop need to be generated, one which is vectorized | |
2521 | and one which isn't. A test is then generated to control which of the | |
2522 | loops is executed. The test checks for the alignment of all of the | |
2523 | data references that may or may not be aligned. An additional | |
2524 | sequence of runtime tests is generated for each pairs of DDRs whose | |
b8698a0f L |
2525 | independence was not proven. The vectorized version of loop is |
2526 | executed only if both alias and alignment tests are passed. | |
2527 | ||
ebfd146a | 2528 | The test generated to check which version of loop is executed |
b8698a0f | 2529 | is modified to also check for profitability as indicated by the |
86290011 RG |
2530 | cost model initially. |
2531 | ||
2532 | The versioning precondition(s) are placed in *COND_EXPR and | |
d68d56b5 | 2533 | *COND_EXPR_STMT_LIST. */ |
ebfd146a IR |
2534 | |
2535 | void | |
368117e8 RG |
2536 | vect_loop_versioning (loop_vec_info loop_vinfo, |
2537 | unsigned int th, bool check_profitability) | |
ebfd146a IR |
2538 | { |
2539 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a IR |
2540 | basic_block condition_bb; |
2541 | gimple_stmt_iterator gsi, cond_exp_gsi; | |
2542 | basic_block merge_bb; | |
2543 | basic_block new_exit_bb; | |
2544 | edge new_exit_e, e; | |
2545 | gimple orig_phi, new_phi; | |
368117e8 | 2546 | tree cond_expr = NULL_TREE; |
d68d56b5 | 2547 | gimple_seq cond_expr_stmt_list = NULL; |
ebfd146a IR |
2548 | tree arg; |
2549 | unsigned prob = 4 * REG_BR_PROB_BASE / 5; | |
2550 | gimple_seq gimplify_stmt_list = NULL; | |
2551 | tree scalar_loop_iters = LOOP_VINFO_NITERS (loop_vinfo); | |
ebfd146a | 2552 | |
368117e8 RG |
2553 | if (check_profitability) |
2554 | { | |
2555 | cond_expr = fold_build2 (GT_EXPR, boolean_type_node, scalar_loop_iters, | |
2556 | build_int_cst (TREE_TYPE (scalar_loop_iters), th)); | |
2557 | cond_expr = force_gimple_operand_1 (cond_expr, &cond_expr_stmt_list, | |
2558 | is_gimple_condexpr, NULL_TREE); | |
2559 | } | |
ebfd146a | 2560 | |
e9dbe7bb | 2561 | if (LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo)) |
d68d56b5 RG |
2562 | vect_create_cond_for_align_checks (loop_vinfo, &cond_expr, |
2563 | &cond_expr_stmt_list); | |
ebfd146a | 2564 | |
e9dbe7bb | 2565 | if (LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo)) |
d68d56b5 RG |
2566 | vect_create_cond_for_alias_checks (loop_vinfo, &cond_expr, |
2567 | &cond_expr_stmt_list); | |
86290011 | 2568 | |
d68d56b5 RG |
2569 | cond_expr = force_gimple_operand_1 (cond_expr, &gimplify_stmt_list, |
2570 | is_gimple_condexpr, NULL_TREE); | |
2571 | gimple_seq_add_seq (&cond_expr_stmt_list, gimplify_stmt_list); | |
ebfd146a IR |
2572 | |
2573 | initialize_original_copy_tables (); | |
d68d56b5 | 2574 | loop_version (loop, cond_expr, &condition_bb, |
0f900dfa | 2575 | prob, prob, REG_BR_PROB_BASE - prob, true); |
ebfd146a IR |
2576 | free_original_copy_tables(); |
2577 | ||
b8698a0f | 2578 | /* Loop versioning violates an assumption we try to maintain during |
ebfd146a IR |
2579 | vectorization - that the loop exit block has a single predecessor. |
2580 | After versioning, the exit block of both loop versions is the same | |
2581 | basic block (i.e. it has two predecessors). Just in order to simplify | |
2582 | following transformations in the vectorizer, we fix this situation | |
2583 | here by adding a new (empty) block on the exit-edge of the loop, | |
2584 | with the proper loop-exit phis to maintain loop-closed-form. */ | |
b8698a0f | 2585 | |
ebfd146a IR |
2586 | merge_bb = single_exit (loop)->dest; |
2587 | gcc_assert (EDGE_COUNT (merge_bb->preds) == 2); | |
2588 | new_exit_bb = split_edge (single_exit (loop)); | |
2589 | new_exit_e = single_exit (loop); | |
2590 | e = EDGE_SUCC (new_exit_bb, 0); | |
2591 | ||
2592 | for (gsi = gsi_start_phis (merge_bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
2593 | { | |
070ecdfd | 2594 | tree new_res; |
ebfd146a | 2595 | orig_phi = gsi_stmt (gsi); |
070ecdfd RG |
2596 | new_res = copy_ssa_name (PHI_RESULT (orig_phi), NULL); |
2597 | new_phi = create_phi_node (new_res, new_exit_bb); | |
ebfd146a | 2598 | arg = PHI_ARG_DEF_FROM_EDGE (orig_phi, e); |
b8698a0f | 2599 | add_phi_arg (new_phi, arg, new_exit_e, |
9e227d60 | 2600 | gimple_phi_arg_location_from_edge (orig_phi, e)); |
684f25f4 | 2601 | adjust_phi_and_debug_stmts (orig_phi, e, PHI_RESULT (new_phi)); |
b8698a0f | 2602 | } |
ebfd146a IR |
2603 | |
2604 | /* End loop-exit-fixes after versioning. */ | |
2605 | ||
2606 | update_ssa (TODO_update_ssa); | |
d68d56b5 | 2607 | if (cond_expr_stmt_list) |
ebfd146a IR |
2608 | { |
2609 | cond_exp_gsi = gsi_last_bb (condition_bb); | |
d68d56b5 | 2610 | gsi_insert_seq_before (&cond_exp_gsi, cond_expr_stmt_list, |
86290011 | 2611 | GSI_SAME_STMT); |
ebfd146a IR |
2612 | } |
2613 | } |