]>
Commit | Line | Data |
---|---|---|
b8698a0f | 1 | /* Vectorizer Specific Loop Manipulations |
d1e082c2 | 2 | Copyright (C) 2003-2013 Free Software Foundation, Inc. |
b8698a0f | 3 | Contributed by Dorit Naishlos <dorit@il.ibm.com> |
ebfd146a IR |
4 | and Ira Rosen <irar@il.ibm.com> |
5 | ||
6 | This file is part of GCC. | |
7 | ||
8 | GCC is free software; you can redistribute it and/or modify it under | |
9 | the terms of the GNU General Public License as published by the Free | |
10 | Software Foundation; either version 3, or (at your option) any later | |
11 | version. | |
12 | ||
13 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
14 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
15 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
16 | for more details. | |
17 | ||
18 | You should have received a copy of the GNU General Public License | |
19 | along with GCC; see the file COPYING3. If not see | |
20 | <http://www.gnu.org/licenses/>. */ | |
21 | ||
22 | #include "config.h" | |
23 | #include "system.h" | |
24 | #include "coretypes.h" | |
78c60e3d | 25 | #include "dumpfile.h" |
ebfd146a IR |
26 | #include "tm.h" |
27 | #include "ggc.h" | |
28 | #include "tree.h" | |
29 | #include "basic-block.h" | |
cf835838 | 30 | #include "gimple-pretty-print.h" |
7a300452 | 31 | #include "tree-ssa.h" |
7ee2468b | 32 | #include "tree-pass.h" |
ebfd146a | 33 | #include "cfgloop.h" |
718f9c0f | 34 | #include "diagnostic-core.h" |
ebfd146a IR |
35 | #include "tree-scalar-evolution.h" |
36 | #include "tree-vectorizer.h" | |
37 | #include "langhooks.h" | |
38 | ||
39 | /************************************************************************* | |
40 | Simple Loop Peeling Utilities | |
41 | ||
42 | Utilities to support loop peeling for vectorization purposes. | |
43 | *************************************************************************/ | |
44 | ||
45 | ||
46 | /* Renames the use *OP_P. */ | |
47 | ||
48 | static void | |
49 | rename_use_op (use_operand_p op_p) | |
50 | { | |
51 | tree new_name; | |
52 | ||
53 | if (TREE_CODE (USE_FROM_PTR (op_p)) != SSA_NAME) | |
54 | return; | |
55 | ||
56 | new_name = get_current_def (USE_FROM_PTR (op_p)); | |
57 | ||
58 | /* Something defined outside of the loop. */ | |
59 | if (!new_name) | |
60 | return; | |
61 | ||
62 | /* An ordinary ssa name defined in the loop. */ | |
63 | ||
64 | SET_USE (op_p, new_name); | |
65 | } | |
66 | ||
67 | ||
68 | /* Renames the variables in basic block BB. */ | |
69 | ||
2cfc56b9 | 70 | static void |
ebfd146a IR |
71 | rename_variables_in_bb (basic_block bb) |
72 | { | |
73 | gimple_stmt_iterator gsi; | |
74 | gimple stmt; | |
75 | use_operand_p use_p; | |
76 | ssa_op_iter iter; | |
77 | edge e; | |
78 | edge_iterator ei; | |
79 | struct loop *loop = bb->loop_father; | |
80 | ||
81 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
82 | { | |
83 | stmt = gsi_stmt (gsi); | |
84 | FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_ALL_USES) | |
85 | rename_use_op (use_p); | |
86 | } | |
87 | ||
2cfc56b9 | 88 | FOR_EACH_EDGE (e, ei, bb->preds) |
ebfd146a | 89 | { |
2cfc56b9 | 90 | if (!flow_bb_inside_loop_p (loop, e->src)) |
ebfd146a | 91 | continue; |
2cfc56b9 | 92 | for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
ebfd146a IR |
93 | rename_use_op (PHI_ARG_DEF_PTR_FROM_EDGE (gsi_stmt (gsi), e)); |
94 | } | |
95 | } | |
96 | ||
97 | ||
684f25f4 AO |
98 | typedef struct |
99 | { | |
100 | tree from, to; | |
101 | basic_block bb; | |
102 | } adjust_info; | |
103 | ||
684f25f4 AO |
104 | /* A stack of values to be adjusted in debug stmts. We have to |
105 | process them LIFO, so that the closest substitution applies. If we | |
106 | processed them FIFO, without the stack, we might substitute uses | |
107 | with a PHI DEF that would soon become non-dominant, and when we got | |
108 | to the suitable one, it wouldn't have anything to substitute any | |
109 | more. */ | |
9771b263 | 110 | static vec<adjust_info, va_stack> adjust_vec; |
684f25f4 AO |
111 | |
112 | /* Adjust any debug stmts that referenced AI->from values to use the | |
113 | loop-closed AI->to, if the references are dominated by AI->bb and | |
114 | not by the definition of AI->from. */ | |
115 | ||
116 | static void | |
117 | adjust_debug_stmts_now (adjust_info *ai) | |
118 | { | |
119 | basic_block bbphi = ai->bb; | |
120 | tree orig_def = ai->from; | |
121 | tree new_def = ai->to; | |
122 | imm_use_iterator imm_iter; | |
123 | gimple stmt; | |
124 | basic_block bbdef = gimple_bb (SSA_NAME_DEF_STMT (orig_def)); | |
125 | ||
126 | gcc_assert (dom_info_available_p (CDI_DOMINATORS)); | |
127 | ||
128 | /* Adjust any debug stmts that held onto non-loop-closed | |
129 | references. */ | |
130 | FOR_EACH_IMM_USE_STMT (stmt, imm_iter, orig_def) | |
131 | { | |
132 | use_operand_p use_p; | |
133 | basic_block bbuse; | |
134 | ||
135 | if (!is_gimple_debug (stmt)) | |
136 | continue; | |
137 | ||
138 | gcc_assert (gimple_debug_bind_p (stmt)); | |
139 | ||
140 | bbuse = gimple_bb (stmt); | |
141 | ||
142 | if ((bbuse == bbphi | |
143 | || dominated_by_p (CDI_DOMINATORS, bbuse, bbphi)) | |
144 | && !(bbuse == bbdef | |
145 | || dominated_by_p (CDI_DOMINATORS, bbuse, bbdef))) | |
146 | { | |
147 | if (new_def) | |
148 | FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter) | |
149 | SET_USE (use_p, new_def); | |
150 | else | |
151 | { | |
152 | gimple_debug_bind_reset_value (stmt); | |
153 | update_stmt (stmt); | |
154 | } | |
155 | } | |
156 | } | |
157 | } | |
158 | ||
159 | /* Adjust debug stmts as scheduled before. */ | |
160 | ||
161 | static void | |
162 | adjust_vec_debug_stmts (void) | |
163 | { | |
164 | if (!MAY_HAVE_DEBUG_STMTS) | |
165 | return; | |
166 | ||
9771b263 | 167 | gcc_assert (adjust_vec.exists ()); |
684f25f4 | 168 | |
9771b263 | 169 | while (!adjust_vec.is_empty ()) |
684f25f4 | 170 | { |
9771b263 DN |
171 | adjust_debug_stmts_now (&adjust_vec.last ()); |
172 | adjust_vec.pop (); | |
684f25f4 AO |
173 | } |
174 | ||
9771b263 | 175 | adjust_vec.release (); |
684f25f4 AO |
176 | } |
177 | ||
178 | /* Adjust any debug stmts that referenced FROM values to use the | |
179 | loop-closed TO, if the references are dominated by BB and not by | |
180 | the definition of FROM. If adjust_vec is non-NULL, adjustments | |
181 | will be postponed until adjust_vec_debug_stmts is called. */ | |
182 | ||
183 | static void | |
184 | adjust_debug_stmts (tree from, tree to, basic_block bb) | |
185 | { | |
186 | adjust_info ai; | |
187 | ||
a471762f RG |
188 | if (MAY_HAVE_DEBUG_STMTS |
189 | && TREE_CODE (from) == SSA_NAME | |
a52ca739 | 190 | && ! SSA_NAME_IS_DEFAULT_DEF (from) |
a471762f | 191 | && ! virtual_operand_p (from)) |
684f25f4 AO |
192 | { |
193 | ai.from = from; | |
194 | ai.to = to; | |
195 | ai.bb = bb; | |
196 | ||
9771b263 DN |
197 | if (adjust_vec.exists ()) |
198 | adjust_vec.safe_push (ai); | |
684f25f4 AO |
199 | else |
200 | adjust_debug_stmts_now (&ai); | |
201 | } | |
202 | } | |
203 | ||
204 | /* Change E's phi arg in UPDATE_PHI to NEW_DEF, and record information | |
205 | to adjust any debug stmts that referenced the old phi arg, | |
206 | presumably non-loop-closed references left over from other | |
207 | transformations. */ | |
208 | ||
209 | static void | |
210 | adjust_phi_and_debug_stmts (gimple update_phi, edge e, tree new_def) | |
211 | { | |
212 | tree orig_def = PHI_ARG_DEF_FROM_EDGE (update_phi, e); | |
213 | ||
214 | SET_PHI_ARG_DEF (update_phi, e->dest_idx, new_def); | |
215 | ||
216 | if (MAY_HAVE_DEBUG_STMTS) | |
217 | adjust_debug_stmts (orig_def, PHI_RESULT (update_phi), | |
218 | gimple_bb (update_phi)); | |
219 | } | |
220 | ||
ebfd146a | 221 | |
ebfd146a IR |
222 | /* Update PHI nodes for a guard of the LOOP. |
223 | ||
224 | Input: | |
225 | - LOOP, GUARD_EDGE: LOOP is a loop for which we added guard code that | |
226 | controls whether LOOP is to be executed. GUARD_EDGE is the edge that | |
227 | originates from the guard-bb, skips LOOP and reaches the (unique) exit | |
228 | bb of LOOP. This loop-exit-bb is an empty bb with one successor. | |
229 | We denote this bb NEW_MERGE_BB because before the guard code was added | |
230 | it had a single predecessor (the LOOP header), and now it became a merge | |
231 | point of two paths - the path that ends with the LOOP exit-edge, and | |
232 | the path that ends with GUARD_EDGE. | |
233 | - NEW_EXIT_BB: New basic block that is added by this function between LOOP | |
234 | and NEW_MERGE_BB. It is used to place loop-closed-ssa-form exit-phis. | |
235 | ||
236 | ===> The CFG before the guard-code was added: | |
237 | LOOP_header_bb: | |
238 | loop_body | |
239 | if (exit_loop) goto update_bb | |
240 | else goto LOOP_header_bb | |
241 | update_bb: | |
242 | ||
243 | ==> The CFG after the guard-code was added: | |
244 | guard_bb: | |
245 | if (LOOP_guard_condition) goto new_merge_bb | |
246 | else goto LOOP_header_bb | |
247 | LOOP_header_bb: | |
248 | loop_body | |
249 | if (exit_loop_condition) goto new_merge_bb | |
250 | else goto LOOP_header_bb | |
251 | new_merge_bb: | |
252 | goto update_bb | |
253 | update_bb: | |
254 | ||
255 | ==> The CFG after this function: | |
256 | guard_bb: | |
257 | if (LOOP_guard_condition) goto new_merge_bb | |
258 | else goto LOOP_header_bb | |
259 | LOOP_header_bb: | |
260 | loop_body | |
261 | if (exit_loop_condition) goto new_exit_bb | |
262 | else goto LOOP_header_bb | |
263 | new_exit_bb: | |
264 | new_merge_bb: | |
265 | goto update_bb | |
266 | update_bb: | |
267 | ||
268 | This function: | |
269 | 1. creates and updates the relevant phi nodes to account for the new | |
270 | incoming edge (GUARD_EDGE) into NEW_MERGE_BB. This involves: | |
271 | 1.1. Create phi nodes at NEW_MERGE_BB. | |
272 | 1.2. Update the phi nodes at the successor of NEW_MERGE_BB (denoted | |
273 | UPDATE_BB). UPDATE_BB was the exit-bb of LOOP before NEW_MERGE_BB | |
274 | 2. preserves loop-closed-ssa-form by creating the required phi nodes | |
275 | at the exit of LOOP (i.e, in NEW_EXIT_BB). | |
276 | ||
277 | There are two flavors to this function: | |
278 | ||
279 | slpeel_update_phi_nodes_for_guard1: | |
280 | Here the guard controls whether we enter or skip LOOP, where LOOP is a | |
281 | prolog_loop (loop1 below), and the new phis created in NEW_MERGE_BB are | |
282 | for variables that have phis in the loop header. | |
283 | ||
284 | slpeel_update_phi_nodes_for_guard2: | |
285 | Here the guard controls whether we enter or skip LOOP, where LOOP is an | |
286 | epilog_loop (loop2 below), and the new phis created in NEW_MERGE_BB are | |
287 | for variables that have phis in the loop exit. | |
288 | ||
289 | I.E., the overall structure is: | |
290 | ||
291 | loop1_preheader_bb: | |
292 | guard1 (goto loop1/merge1_bb) | |
293 | loop1 | |
294 | loop1_exit_bb: | |
295 | guard2 (goto merge1_bb/merge2_bb) | |
296 | merge1_bb | |
297 | loop2 | |
298 | loop2_exit_bb | |
299 | merge2_bb | |
300 | next_bb | |
301 | ||
302 | slpeel_update_phi_nodes_for_guard1 takes care of creating phis in | |
303 | loop1_exit_bb and merge1_bb. These are entry phis (phis for the vars | |
304 | that have phis in loop1->header). | |
305 | ||
306 | slpeel_update_phi_nodes_for_guard2 takes care of creating phis in | |
307 | loop2_exit_bb and merge2_bb. These are exit phis (phis for the vars | |
308 | that have phis in next_bb). It also adds some of these phis to | |
309 | loop1_exit_bb. | |
310 | ||
311 | slpeel_update_phi_nodes_for_guard1 is always called before | |
312 | slpeel_update_phi_nodes_for_guard2. They are both needed in order | |
313 | to create correct data-flow and loop-closed-ssa-form. | |
314 | ||
315 | Generally slpeel_update_phi_nodes_for_guard1 creates phis for variables | |
316 | that change between iterations of a loop (and therefore have a phi-node | |
317 | at the loop entry), whereas slpeel_update_phi_nodes_for_guard2 creates | |
b8698a0f L |
318 | phis for variables that are used out of the loop (and therefore have |
319 | loop-closed exit phis). Some variables may be both updated between | |
ebfd146a IR |
320 | iterations and used after the loop. This is why in loop1_exit_bb we |
321 | may need both entry_phis (created by slpeel_update_phi_nodes_for_guard1) | |
322 | and exit phis (created by slpeel_update_phi_nodes_for_guard2). | |
323 | ||
324 | - IS_NEW_LOOP: if IS_NEW_LOOP is true, then LOOP is a newly created copy of | |
325 | an original loop. i.e., we have: | |
326 | ||
327 | orig_loop | |
328 | guard_bb (goto LOOP/new_merge) | |
329 | new_loop <-- LOOP | |
330 | new_exit | |
331 | new_merge | |
332 | next_bb | |
333 | ||
334 | If IS_NEW_LOOP is false, then LOOP is an original loop, in which case we | |
335 | have: | |
336 | ||
337 | new_loop | |
338 | guard_bb (goto LOOP/new_merge) | |
339 | orig_loop <-- LOOP | |
340 | new_exit | |
341 | new_merge | |
342 | next_bb | |
343 | ||
344 | The SSA names defined in the original loop have a current | |
345 | reaching definition that that records the corresponding new | |
346 | ssa-name used in the new duplicated loop copy. | |
347 | */ | |
348 | ||
349 | /* Function slpeel_update_phi_nodes_for_guard1 | |
b8698a0f | 350 | |
ebfd146a IR |
351 | Input: |
352 | - GUARD_EDGE, LOOP, IS_NEW_LOOP, NEW_EXIT_BB - as explained above. | |
353 | - DEFS - a bitmap of ssa names to mark new names for which we recorded | |
b8698a0f L |
354 | information. |
355 | ||
ebfd146a IR |
356 | In the context of the overall structure, we have: |
357 | ||
b8698a0f | 358 | loop1_preheader_bb: |
ebfd146a IR |
359 | guard1 (goto loop1/merge1_bb) |
360 | LOOP-> loop1 | |
361 | loop1_exit_bb: | |
362 | guard2 (goto merge1_bb/merge2_bb) | |
363 | merge1_bb | |
364 | loop2 | |
365 | loop2_exit_bb | |
366 | merge2_bb | |
367 | next_bb | |
368 | ||
369 | For each name updated between loop iterations (i.e - for each name that has | |
370 | an entry (loop-header) phi in LOOP) we create a new phi in: | |
371 | 1. merge1_bb (to account for the edge from guard1) | |
372 | 2. loop1_exit_bb (an exit-phi to keep LOOP in loop-closed form) | |
373 | */ | |
374 | ||
375 | static void | |
376 | slpeel_update_phi_nodes_for_guard1 (edge guard_edge, struct loop *loop, | |
c334023f | 377 | bool is_new_loop, basic_block *new_exit_bb) |
ebfd146a IR |
378 | { |
379 | gimple orig_phi, new_phi; | |
380 | gimple update_phi, update_phi2; | |
381 | tree guard_arg, loop_arg; | |
382 | basic_block new_merge_bb = guard_edge->dest; | |
383 | edge e = EDGE_SUCC (new_merge_bb, 0); | |
384 | basic_block update_bb = e->dest; | |
385 | basic_block orig_bb = loop->header; | |
386 | edge new_exit_e; | |
387 | tree current_new_name; | |
ebfd146a IR |
388 | gimple_stmt_iterator gsi_orig, gsi_update; |
389 | ||
390 | /* Create new bb between loop and new_merge_bb. */ | |
391 | *new_exit_bb = split_edge (single_exit (loop)); | |
392 | ||
393 | new_exit_e = EDGE_SUCC (*new_exit_bb, 0); | |
394 | ||
395 | for (gsi_orig = gsi_start_phis (orig_bb), | |
396 | gsi_update = gsi_start_phis (update_bb); | |
397 | !gsi_end_p (gsi_orig) && !gsi_end_p (gsi_update); | |
398 | gsi_next (&gsi_orig), gsi_next (&gsi_update)) | |
399 | { | |
e20f6b4b | 400 | source_location loop_locus, guard_locus; |
070ecdfd | 401 | tree new_res; |
ebfd146a IR |
402 | orig_phi = gsi_stmt (gsi_orig); |
403 | update_phi = gsi_stmt (gsi_update); | |
404 | ||
ebfd146a IR |
405 | /** 1. Handle new-merge-point phis **/ |
406 | ||
407 | /* 1.1. Generate new phi node in NEW_MERGE_BB: */ | |
070ecdfd RG |
408 | new_res = copy_ssa_name (PHI_RESULT (orig_phi), NULL); |
409 | new_phi = create_phi_node (new_res, new_merge_bb); | |
ebfd146a IR |
410 | |
411 | /* 1.2. NEW_MERGE_BB has two incoming edges: GUARD_EDGE and the exit-edge | |
412 | of LOOP. Set the two phi args in NEW_PHI for these edges: */ | |
413 | loop_arg = PHI_ARG_DEF_FROM_EDGE (orig_phi, EDGE_SUCC (loop->latch, 0)); | |
b8698a0f L |
414 | loop_locus = gimple_phi_arg_location_from_edge (orig_phi, |
415 | EDGE_SUCC (loop->latch, | |
f5045c96 | 416 | 0)); |
ebfd146a | 417 | guard_arg = PHI_ARG_DEF_FROM_EDGE (orig_phi, loop_preheader_edge (loop)); |
b8698a0f L |
418 | guard_locus |
419 | = gimple_phi_arg_location_from_edge (orig_phi, | |
f5045c96 | 420 | loop_preheader_edge (loop)); |
ebfd146a | 421 | |
9e227d60 DC |
422 | add_phi_arg (new_phi, loop_arg, new_exit_e, loop_locus); |
423 | add_phi_arg (new_phi, guard_arg, guard_edge, guard_locus); | |
ebfd146a IR |
424 | |
425 | /* 1.3. Update phi in successor block. */ | |
426 | gcc_assert (PHI_ARG_DEF_FROM_EDGE (update_phi, e) == loop_arg | |
427 | || PHI_ARG_DEF_FROM_EDGE (update_phi, e) == guard_arg); | |
684f25f4 | 428 | adjust_phi_and_debug_stmts (update_phi, e, PHI_RESULT (new_phi)); |
ebfd146a IR |
429 | update_phi2 = new_phi; |
430 | ||
431 | ||
432 | /** 2. Handle loop-closed-ssa-form phis **/ | |
433 | ||
ea057359 | 434 | if (virtual_operand_p (PHI_RESULT (orig_phi))) |
ebfd146a IR |
435 | continue; |
436 | ||
437 | /* 2.1. Generate new phi node in NEW_EXIT_BB: */ | |
070ecdfd RG |
438 | new_res = copy_ssa_name (PHI_RESULT (orig_phi), NULL); |
439 | new_phi = create_phi_node (new_res, *new_exit_bb); | |
ebfd146a IR |
440 | |
441 | /* 2.2. NEW_EXIT_BB has one incoming edge: the exit-edge of the loop. */ | |
9e227d60 | 442 | add_phi_arg (new_phi, loop_arg, single_exit (loop), loop_locus); |
ebfd146a IR |
443 | |
444 | /* 2.3. Update phi in successor of NEW_EXIT_BB: */ | |
445 | gcc_assert (PHI_ARG_DEF_FROM_EDGE (update_phi2, new_exit_e) == loop_arg); | |
684f25f4 AO |
446 | adjust_phi_and_debug_stmts (update_phi2, new_exit_e, |
447 | PHI_RESULT (new_phi)); | |
ebfd146a IR |
448 | |
449 | /* 2.4. Record the newly created name with set_current_def. | |
450 | We want to find a name such that | |
451 | name = get_current_def (orig_loop_name) | |
452 | and to set its current definition as follows: | |
453 | set_current_def (name, new_phi_name) | |
454 | ||
455 | If LOOP is a new loop then loop_arg is already the name we're | |
456 | looking for. If LOOP is the original loop, then loop_arg is | |
457 | the orig_loop_name and the relevant name is recorded in its | |
458 | current reaching definition. */ | |
459 | if (is_new_loop) | |
460 | current_new_name = loop_arg; | |
461 | else | |
462 | { | |
463 | current_new_name = get_current_def (loop_arg); | |
464 | /* current_def is not available only if the variable does not | |
465 | change inside the loop, in which case we also don't care | |
466 | about recording a current_def for it because we won't be | |
467 | trying to create loop-exit-phis for it. */ | |
468 | if (!current_new_name) | |
469 | continue; | |
470 | } | |
471 | gcc_assert (get_current_def (current_new_name) == NULL_TREE); | |
472 | ||
473 | set_current_def (current_new_name, PHI_RESULT (new_phi)); | |
ebfd146a IR |
474 | } |
475 | } | |
476 | ||
477 | ||
478 | /* Function slpeel_update_phi_nodes_for_guard2 | |
479 | ||
480 | Input: | |
481 | - GUARD_EDGE, LOOP, IS_NEW_LOOP, NEW_EXIT_BB - as explained above. | |
482 | ||
483 | In the context of the overall structure, we have: | |
484 | ||
b8698a0f | 485 | loop1_preheader_bb: |
ebfd146a IR |
486 | guard1 (goto loop1/merge1_bb) |
487 | loop1 | |
b8698a0f | 488 | loop1_exit_bb: |
ebfd146a IR |
489 | guard2 (goto merge1_bb/merge2_bb) |
490 | merge1_bb | |
491 | LOOP-> loop2 | |
492 | loop2_exit_bb | |
493 | merge2_bb | |
494 | next_bb | |
495 | ||
496 | For each name used out side the loop (i.e - for each name that has an exit | |
497 | phi in next_bb) we create a new phi in: | |
b8698a0f | 498 | 1. merge2_bb (to account for the edge from guard_bb) |
ebfd146a IR |
499 | 2. loop2_exit_bb (an exit-phi to keep LOOP in loop-closed form) |
500 | 3. guard2 bb (an exit phi to keep the preceding loop in loop-closed form), | |
501 | if needed (if it wasn't handled by slpeel_update_phis_nodes_for_phi1). | |
502 | */ | |
503 | ||
504 | static void | |
505 | slpeel_update_phi_nodes_for_guard2 (edge guard_edge, struct loop *loop, | |
506 | bool is_new_loop, basic_block *new_exit_bb) | |
507 | { | |
508 | gimple orig_phi, new_phi; | |
509 | gimple update_phi, update_phi2; | |
510 | tree guard_arg, loop_arg; | |
511 | basic_block new_merge_bb = guard_edge->dest; | |
512 | edge e = EDGE_SUCC (new_merge_bb, 0); | |
513 | basic_block update_bb = e->dest; | |
514 | edge new_exit_e; | |
515 | tree orig_def, orig_def_new_name; | |
516 | tree new_name, new_name2; | |
517 | tree arg; | |
518 | gimple_stmt_iterator gsi; | |
519 | ||
520 | /* Create new bb between loop and new_merge_bb. */ | |
521 | *new_exit_bb = split_edge (single_exit (loop)); | |
522 | ||
523 | new_exit_e = EDGE_SUCC (*new_exit_bb, 0); | |
524 | ||
525 | for (gsi = gsi_start_phis (update_bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
526 | { | |
070ecdfd | 527 | tree new_res; |
ebfd146a IR |
528 | update_phi = gsi_stmt (gsi); |
529 | orig_phi = update_phi; | |
530 | orig_def = PHI_ARG_DEF_FROM_EDGE (orig_phi, e); | |
531 | /* This loop-closed-phi actually doesn't represent a use | |
b8698a0f | 532 | out of the loop - the phi arg is a constant. */ |
ebfd146a IR |
533 | if (TREE_CODE (orig_def) != SSA_NAME) |
534 | continue; | |
535 | orig_def_new_name = get_current_def (orig_def); | |
536 | arg = NULL_TREE; | |
537 | ||
538 | /** 1. Handle new-merge-point phis **/ | |
539 | ||
540 | /* 1.1. Generate new phi node in NEW_MERGE_BB: */ | |
070ecdfd RG |
541 | new_res = copy_ssa_name (PHI_RESULT (orig_phi), NULL); |
542 | new_phi = create_phi_node (new_res, new_merge_bb); | |
ebfd146a IR |
543 | |
544 | /* 1.2. NEW_MERGE_BB has two incoming edges: GUARD_EDGE and the exit-edge | |
545 | of LOOP. Set the two PHI args in NEW_PHI for these edges: */ | |
546 | new_name = orig_def; | |
547 | new_name2 = NULL_TREE; | |
548 | if (orig_def_new_name) | |
549 | { | |
550 | new_name = orig_def_new_name; | |
551 | /* Some variables have both loop-entry-phis and loop-exit-phis. | |
552 | Such variables were given yet newer names by phis placed in | |
553 | guard_bb by slpeel_update_phi_nodes_for_guard1. I.e: | |
554 | new_name2 = get_current_def (get_current_def (orig_name)). */ | |
555 | new_name2 = get_current_def (new_name); | |
556 | } | |
b8698a0f | 557 | |
ebfd146a IR |
558 | if (is_new_loop) |
559 | { | |
560 | guard_arg = orig_def; | |
561 | loop_arg = new_name; | |
562 | } | |
563 | else | |
564 | { | |
565 | guard_arg = new_name; | |
566 | loop_arg = orig_def; | |
567 | } | |
568 | if (new_name2) | |
569 | guard_arg = new_name2; | |
b8698a0f | 570 | |
9e227d60 DC |
571 | add_phi_arg (new_phi, loop_arg, new_exit_e, UNKNOWN_LOCATION); |
572 | add_phi_arg (new_phi, guard_arg, guard_edge, UNKNOWN_LOCATION); | |
ebfd146a IR |
573 | |
574 | /* 1.3. Update phi in successor block. */ | |
575 | gcc_assert (PHI_ARG_DEF_FROM_EDGE (update_phi, e) == orig_def); | |
684f25f4 | 576 | adjust_phi_and_debug_stmts (update_phi, e, PHI_RESULT (new_phi)); |
ebfd146a IR |
577 | update_phi2 = new_phi; |
578 | ||
579 | ||
580 | /** 2. Handle loop-closed-ssa-form phis **/ | |
581 | ||
582 | /* 2.1. Generate new phi node in NEW_EXIT_BB: */ | |
070ecdfd RG |
583 | new_res = copy_ssa_name (PHI_RESULT (orig_phi), NULL); |
584 | new_phi = create_phi_node (new_res, *new_exit_bb); | |
ebfd146a IR |
585 | |
586 | /* 2.2. NEW_EXIT_BB has one incoming edge: the exit-edge of the loop. */ | |
9e227d60 | 587 | add_phi_arg (new_phi, loop_arg, single_exit (loop), UNKNOWN_LOCATION); |
ebfd146a IR |
588 | |
589 | /* 2.3. Update phi in successor of NEW_EXIT_BB: */ | |
590 | gcc_assert (PHI_ARG_DEF_FROM_EDGE (update_phi2, new_exit_e) == loop_arg); | |
684f25f4 AO |
591 | adjust_phi_and_debug_stmts (update_phi2, new_exit_e, |
592 | PHI_RESULT (new_phi)); | |
ebfd146a IR |
593 | |
594 | ||
595 | /** 3. Handle loop-closed-ssa-form phis for first loop **/ | |
596 | ||
597 | /* 3.1. Find the relevant names that need an exit-phi in | |
598 | GUARD_BB, i.e. names for which | |
599 | slpeel_update_phi_nodes_for_guard1 had not already created a | |
600 | phi node. This is the case for names that are used outside | |
601 | the loop (and therefore need an exit phi) but are not updated | |
602 | across loop iterations (and therefore don't have a | |
603 | loop-header-phi). | |
604 | ||
605 | slpeel_update_phi_nodes_for_guard1 is responsible for | |
606 | creating loop-exit phis in GUARD_BB for names that have a | |
607 | loop-header-phi. When such a phi is created we also record | |
608 | the new name in its current definition. If this new name | |
609 | exists, then guard_arg was set to this new name (see 1.2 | |
610 | above). Therefore, if guard_arg is not this new name, this | |
611 | is an indication that an exit-phi in GUARD_BB was not yet | |
612 | created, so we take care of it here. */ | |
613 | if (guard_arg == new_name2) | |
614 | continue; | |
615 | arg = guard_arg; | |
616 | ||
617 | /* 3.2. Generate new phi node in GUARD_BB: */ | |
070ecdfd RG |
618 | new_res = copy_ssa_name (PHI_RESULT (orig_phi), NULL); |
619 | new_phi = create_phi_node (new_res, guard_edge->src); | |
ebfd146a IR |
620 | |
621 | /* 3.3. GUARD_BB has one incoming edge: */ | |
622 | gcc_assert (EDGE_COUNT (guard_edge->src->preds) == 1); | |
b8698a0f | 623 | add_phi_arg (new_phi, arg, EDGE_PRED (guard_edge->src, 0), |
9e227d60 | 624 | UNKNOWN_LOCATION); |
ebfd146a IR |
625 | |
626 | /* 3.4. Update phi in successor of GUARD_BB: */ | |
627 | gcc_assert (PHI_ARG_DEF_FROM_EDGE (update_phi2, guard_edge) | |
628 | == guard_arg); | |
684f25f4 AO |
629 | adjust_phi_and_debug_stmts (update_phi2, guard_edge, |
630 | PHI_RESULT (new_phi)); | |
ebfd146a IR |
631 | } |
632 | } | |
633 | ||
634 | ||
635 | /* Make the LOOP iterate NITERS times. This is done by adding a new IV | |
636 | that starts at zero, increases by one and its limit is NITERS. | |
637 | ||
638 | Assumption: the exit-condition of LOOP is the last stmt in the loop. */ | |
639 | ||
640 | void | |
641 | slpeel_make_loop_iterate_ntimes (struct loop *loop, tree niters) | |
642 | { | |
643 | tree indx_before_incr, indx_after_incr; | |
644 | gimple cond_stmt; | |
645 | gimple orig_cond; | |
646 | edge exit_edge = single_exit (loop); | |
647 | gimple_stmt_iterator loop_cond_gsi; | |
648 | gimple_stmt_iterator incr_gsi; | |
649 | bool insert_after; | |
650 | tree init = build_int_cst (TREE_TYPE (niters), 0); | |
651 | tree step = build_int_cst (TREE_TYPE (niters), 1); | |
652 | LOC loop_loc; | |
653 | enum tree_code code; | |
654 | ||
655 | orig_cond = get_loop_exit_condition (loop); | |
656 | gcc_assert (orig_cond); | |
657 | loop_cond_gsi = gsi_for_stmt (orig_cond); | |
658 | ||
659 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
660 | create_iv (init, step, NULL_TREE, loop, | |
661 | &incr_gsi, insert_after, &indx_before_incr, &indx_after_incr); | |
662 | ||
663 | indx_after_incr = force_gimple_operand_gsi (&loop_cond_gsi, indx_after_incr, | |
664 | true, NULL_TREE, true, | |
665 | GSI_SAME_STMT); | |
666 | niters = force_gimple_operand_gsi (&loop_cond_gsi, niters, true, NULL_TREE, | |
667 | true, GSI_SAME_STMT); | |
668 | ||
669 | code = (exit_edge->flags & EDGE_TRUE_VALUE) ? GE_EXPR : LT_EXPR; | |
670 | cond_stmt = gimple_build_cond (code, indx_after_incr, niters, NULL_TREE, | |
671 | NULL_TREE); | |
672 | ||
673 | gsi_insert_before (&loop_cond_gsi, cond_stmt, GSI_SAME_STMT); | |
674 | ||
675 | /* Remove old loop exit test: */ | |
676 | gsi_remove (&loop_cond_gsi, true); | |
6f723d33 | 677 | free_stmt_vec_info (orig_cond); |
ebfd146a IR |
678 | |
679 | loop_loc = find_loop_location (loop); | |
73fbfcad | 680 | if (dump_enabled_p ()) |
ebfd146a | 681 | { |
84df911b DC |
682 | if (LOCATION_LOCUS (loop_loc) != UNKNOWN_LOC) |
683 | dump_printf (MSG_NOTE, "\nloop at %s:%d: ", LOC_FILE (loop_loc), | |
684 | LOC_LINE (loop_loc)); | |
78c60e3d | 685 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, cond_stmt, 0); |
ebfd146a | 686 | } |
ebfd146a IR |
687 | loop->nb_iterations = niters; |
688 | } | |
689 | ||
690 | ||
b8698a0f | 691 | /* Given LOOP this function generates a new copy of it and puts it |
ebfd146a IR |
692 | on E which is either the entry or exit of LOOP. */ |
693 | ||
694 | struct loop * | |
695 | slpeel_tree_duplicate_loop_to_edge_cfg (struct loop *loop, edge e) | |
696 | { | |
697 | struct loop *new_loop; | |
698 | basic_block *new_bbs, *bbs; | |
699 | bool at_exit; | |
700 | bool was_imm_dom; | |
b8698a0f | 701 | basic_block exit_dest; |
ebfd146a | 702 | edge exit, new_exit; |
ebfd146a | 703 | |
2cfc56b9 RB |
704 | exit = single_exit (loop); |
705 | at_exit = (e == exit); | |
ebfd146a IR |
706 | if (!at_exit && e != loop_preheader_edge (loop)) |
707 | return NULL; | |
708 | ||
2cfc56b9 RB |
709 | bbs = XNEWVEC (basic_block, loop->num_nodes + 1); |
710 | get_loop_body_with_size (loop, bbs, loop->num_nodes); | |
ebfd146a IR |
711 | |
712 | /* Check whether duplication is possible. */ | |
713 | if (!can_copy_bbs_p (bbs, loop->num_nodes)) | |
714 | { | |
715 | free (bbs); | |
716 | return NULL; | |
717 | } | |
718 | ||
719 | /* Generate new loop structure. */ | |
720 | new_loop = duplicate_loop (loop, loop_outer (loop)); | |
2cfc56b9 | 721 | duplicate_subloops (loop, new_loop); |
ebfd146a | 722 | |
2cfc56b9 | 723 | exit_dest = exit->dest; |
b8698a0f L |
724 | was_imm_dom = (get_immediate_dominator (CDI_DOMINATORS, |
725 | exit_dest) == loop->header ? | |
ebfd146a IR |
726 | true : false); |
727 | ||
2cfc56b9 RB |
728 | /* Also copy the pre-header, this avoids jumping through hoops to |
729 | duplicate the loop entry PHI arguments. Create an empty | |
730 | pre-header unconditionally for this. */ | |
731 | basic_block preheader = split_edge (loop_preheader_edge (loop)); | |
732 | edge entry_e = single_pred_edge (preheader); | |
733 | bbs[loop->num_nodes] = preheader; | |
734 | new_bbs = XNEWVEC (basic_block, loop->num_nodes + 1); | |
ebfd146a | 735 | |
2cfc56b9 | 736 | copy_bbs (bbs, loop->num_nodes + 1, new_bbs, |
ebfd146a | 737 | &exit, 1, &new_exit, NULL, |
f14540b6 | 738 | e->src, true); |
2cfc56b9 | 739 | basic_block new_preheader = new_bbs[loop->num_nodes]; |
ebfd146a | 740 | |
2cfc56b9 | 741 | add_phi_args_after_copy (new_bbs, loop->num_nodes + 1, NULL); |
b8698a0f | 742 | |
ebfd146a IR |
743 | if (at_exit) /* Add the loop copy at exit. */ |
744 | { | |
2cfc56b9 RB |
745 | redirect_edge_and_branch_force (e, new_preheader); |
746 | flush_pending_stmts (e); | |
747 | set_immediate_dominator (CDI_DOMINATORS, new_preheader, e->src); | |
ebfd146a IR |
748 | if (was_imm_dom) |
749 | set_immediate_dominator (CDI_DOMINATORS, exit_dest, new_loop->header); | |
2cfc56b9 RB |
750 | |
751 | /* And remove the non-necessary forwarder again. Keep the other | |
752 | one so we have a proper pre-header for the loop at the exit edge. */ | |
753 | redirect_edge_pred (single_succ_edge (preheader), single_pred (preheader)); | |
754 | delete_basic_block (preheader); | |
755 | set_immediate_dominator (CDI_DOMINATORS, loop->header, | |
756 | loop_preheader_edge (loop)->src); | |
ebfd146a IR |
757 | } |
758 | else /* Add the copy at entry. */ | |
759 | { | |
2cfc56b9 RB |
760 | redirect_edge_and_branch_force (entry_e, new_preheader); |
761 | flush_pending_stmts (entry_e); | |
762 | set_immediate_dominator (CDI_DOMINATORS, new_preheader, entry_e->src); | |
763 | ||
764 | redirect_edge_and_branch_force (new_exit, preheader); | |
765 | flush_pending_stmts (new_exit); | |
766 | set_immediate_dominator (CDI_DOMINATORS, preheader, new_exit->src); | |
767 | ||
768 | /* And remove the non-necessary forwarder again. Keep the other | |
769 | one so we have a proper pre-header for the loop at the exit edge. */ | |
770 | redirect_edge_pred (single_succ_edge (new_preheader), single_pred (new_preheader)); | |
771 | delete_basic_block (new_preheader); | |
772 | set_immediate_dominator (CDI_DOMINATORS, new_loop->header, | |
773 | loop_preheader_edge (new_loop)->src); | |
ebfd146a IR |
774 | } |
775 | ||
2cfc56b9 RB |
776 | for (unsigned i = 0; i < loop->num_nodes+1; i++) |
777 | rename_variables_in_bb (new_bbs[i]); | |
778 | ||
ebfd146a IR |
779 | free (new_bbs); |
780 | free (bbs); | |
781 | ||
2cfc56b9 RB |
782 | #ifdef ENABLE_CHECKING |
783 | verify_dominators (CDI_DOMINATORS); | |
784 | #endif | |
785 | ||
ebfd146a IR |
786 | return new_loop; |
787 | } | |
788 | ||
789 | ||
790 | /* Given the condition statement COND, put it as the last statement | |
791 | of GUARD_BB; EXIT_BB is the basic block to skip the loop; | |
b8698a0f | 792 | Assumes that this is the single exit of the guarded loop. |
86290011 | 793 | Returns the skip edge, inserts new stmts on the COND_EXPR_STMT_LIST. */ |
ebfd146a IR |
794 | |
795 | static edge | |
86290011 RG |
796 | slpeel_add_loop_guard (basic_block guard_bb, tree cond, |
797 | gimple_seq cond_expr_stmt_list, | |
e78410bf JH |
798 | basic_block exit_bb, basic_block dom_bb, |
799 | int probability) | |
ebfd146a IR |
800 | { |
801 | gimple_stmt_iterator gsi; | |
802 | edge new_e, enter_e; | |
803 | gimple cond_stmt; | |
804 | gimple_seq gimplify_stmt_list = NULL; | |
805 | ||
806 | enter_e = EDGE_SUCC (guard_bb, 0); | |
807 | enter_e->flags &= ~EDGE_FALLTHRU; | |
808 | enter_e->flags |= EDGE_FALSE_VALUE; | |
809 | gsi = gsi_last_bb (guard_bb); | |
810 | ||
f7a06a98 RG |
811 | cond = force_gimple_operand_1 (cond, &gimplify_stmt_list, is_gimple_condexpr, |
812 | NULL_TREE); | |
86290011 RG |
813 | if (gimplify_stmt_list) |
814 | gimple_seq_add_seq (&cond_expr_stmt_list, gimplify_stmt_list); | |
f7a06a98 | 815 | cond_stmt = gimple_build_cond_from_tree (cond, NULL_TREE, NULL_TREE); |
86290011 RG |
816 | if (cond_expr_stmt_list) |
817 | gsi_insert_seq_after (&gsi, cond_expr_stmt_list, GSI_NEW_STMT); | |
ebfd146a IR |
818 | |
819 | gsi = gsi_last_bb (guard_bb); | |
820 | gsi_insert_after (&gsi, cond_stmt, GSI_NEW_STMT); | |
821 | ||
822 | /* Add new edge to connect guard block to the merge/loop-exit block. */ | |
823 | new_e = make_edge (guard_bb, exit_bb, EDGE_TRUE_VALUE); | |
e78410bf JH |
824 | |
825 | new_e->count = guard_bb->count; | |
826 | new_e->probability = probability; | |
827 | new_e->count = apply_probability (enter_e->count, probability); | |
828 | enter_e->count -= new_e->count; | |
829 | enter_e->probability = inverse_probability (probability); | |
ebfd146a IR |
830 | set_immediate_dominator (CDI_DOMINATORS, exit_bb, dom_bb); |
831 | return new_e; | |
832 | } | |
833 | ||
834 | ||
835 | /* This function verifies that the following restrictions apply to LOOP: | |
836 | (1) it is innermost | |
837 | (2) it consists of exactly 2 basic blocks - header, and an empty latch. | |
838 | (3) it is single entry, single exit | |
839 | (4) its exit condition is the last stmt in the header | |
840 | (5) E is the entry/exit edge of LOOP. | |
841 | */ | |
842 | ||
843 | bool | |
844 | slpeel_can_duplicate_loop_p (const struct loop *loop, const_edge e) | |
845 | { | |
846 | edge exit_e = single_exit (loop); | |
847 | edge entry_e = loop_preheader_edge (loop); | |
848 | gimple orig_cond = get_loop_exit_condition (loop); | |
849 | gimple_stmt_iterator loop_exit_gsi = gsi_last_bb (exit_e->src); | |
850 | ||
ebfd146a IR |
851 | if (loop->inner |
852 | /* All loops have an outer scope; the only case loop->outer is NULL is for | |
853 | the function itself. */ | |
854 | || !loop_outer (loop) | |
855 | || loop->num_nodes != 2 | |
856 | || !empty_block_p (loop->latch) | |
857 | || !single_exit (loop) | |
858 | /* Verify that new loop exit condition can be trivially modified. */ | |
859 | || (!orig_cond || orig_cond != gsi_stmt (loop_exit_gsi)) | |
860 | || (e != exit_e && e != entry_e)) | |
861 | return false; | |
862 | ||
863 | return true; | |
864 | } | |
865 | ||
866 | #ifdef ENABLE_CHECKING | |
867 | static void | |
868 | slpeel_verify_cfg_after_peeling (struct loop *first_loop, | |
869 | struct loop *second_loop) | |
870 | { | |
871 | basic_block loop1_exit_bb = single_exit (first_loop)->dest; | |
872 | basic_block loop2_entry_bb = loop_preheader_edge (second_loop)->src; | |
873 | basic_block loop1_entry_bb = loop_preheader_edge (first_loop)->src; | |
874 | ||
875 | /* A guard that controls whether the second_loop is to be executed or skipped | |
876 | is placed in first_loop->exit. first_loop->exit therefore has two | |
877 | successors - one is the preheader of second_loop, and the other is a bb | |
878 | after second_loop. | |
879 | */ | |
880 | gcc_assert (EDGE_COUNT (loop1_exit_bb->succs) == 2); | |
b8698a0f | 881 | |
ebfd146a IR |
882 | /* 1. Verify that one of the successors of first_loop->exit is the preheader |
883 | of second_loop. */ | |
b8698a0f | 884 | |
ebfd146a IR |
885 | /* The preheader of new_loop is expected to have two predecessors: |
886 | first_loop->exit and the block that precedes first_loop. */ | |
887 | ||
b8698a0f | 888 | gcc_assert (EDGE_COUNT (loop2_entry_bb->preds) == 2 |
ebfd146a IR |
889 | && ((EDGE_PRED (loop2_entry_bb, 0)->src == loop1_exit_bb |
890 | && EDGE_PRED (loop2_entry_bb, 1)->src == loop1_entry_bb) | |
891 | || (EDGE_PRED (loop2_entry_bb, 1)->src == loop1_exit_bb | |
892 | && EDGE_PRED (loop2_entry_bb, 0)->src == loop1_entry_bb))); | |
b8698a0f | 893 | |
ebfd146a IR |
894 | /* Verify that the other successor of first_loop->exit is after the |
895 | second_loop. */ | |
896 | /* TODO */ | |
897 | } | |
898 | #endif | |
899 | ||
900 | /* If the run time cost model check determines that vectorization is | |
901 | not profitable and hence scalar loop should be generated then set | |
902 | FIRST_NITERS to prologue peeled iterations. This will allow all the | |
903 | iterations to be executed in the prologue peeled scalar loop. */ | |
904 | ||
905 | static void | |
906 | set_prologue_iterations (basic_block bb_before_first_loop, | |
5d2eb24b | 907 | tree *first_niters, |
ebfd146a | 908 | struct loop *loop, |
e78410bf JH |
909 | unsigned int th, |
910 | int probability) | |
ebfd146a IR |
911 | { |
912 | edge e; | |
913 | basic_block cond_bb, then_bb; | |
914 | tree var, prologue_after_cost_adjust_name; | |
915 | gimple_stmt_iterator gsi; | |
916 | gimple newphi; | |
917 | edge e_true, e_false, e_fallthru; | |
918 | gimple cond_stmt; | |
f7a06a98 | 919 | gimple_seq stmts = NULL; |
ebfd146a | 920 | tree cost_pre_condition = NULL_TREE; |
b8698a0f | 921 | tree scalar_loop_iters = |
ebfd146a IR |
922 | unshare_expr (LOOP_VINFO_NITERS_UNCHANGED (loop_vec_info_for_loop (loop))); |
923 | ||
924 | e = single_pred_edge (bb_before_first_loop); | |
925 | cond_bb = split_edge(e); | |
926 | ||
927 | e = single_pred_edge (bb_before_first_loop); | |
928 | then_bb = split_edge(e); | |
929 | set_immediate_dominator (CDI_DOMINATORS, then_bb, cond_bb); | |
930 | ||
931 | e_false = make_single_succ_edge (cond_bb, bb_before_first_loop, | |
932 | EDGE_FALSE_VALUE); | |
933 | set_immediate_dominator (CDI_DOMINATORS, bb_before_first_loop, cond_bb); | |
934 | ||
935 | e_true = EDGE_PRED (then_bb, 0); | |
936 | e_true->flags &= ~EDGE_FALLTHRU; | |
937 | e_true->flags |= EDGE_TRUE_VALUE; | |
938 | ||
e78410bf JH |
939 | e_true->probability = probability; |
940 | e_false->probability = inverse_probability (probability); | |
941 | e_true->count = apply_probability (cond_bb->count, probability); | |
942 | e_false->count = cond_bb->count - e_true->count; | |
943 | then_bb->frequency = EDGE_FREQUENCY (e_true); | |
944 | then_bb->count = e_true->count; | |
945 | ||
ebfd146a | 946 | e_fallthru = EDGE_SUCC (then_bb, 0); |
e78410bf | 947 | e_fallthru->count = then_bb->count; |
ebfd146a | 948 | |
f7a06a98 | 949 | gsi = gsi_last_bb (cond_bb); |
ebfd146a | 950 | cost_pre_condition = |
b8698a0f | 951 | fold_build2 (LE_EXPR, boolean_type_node, scalar_loop_iters, |
ebfd146a IR |
952 | build_int_cst (TREE_TYPE (scalar_loop_iters), th)); |
953 | cost_pre_condition = | |
f7a06a98 RG |
954 | force_gimple_operand_gsi_1 (&gsi, cost_pre_condition, is_gimple_condexpr, |
955 | NULL_TREE, false, GSI_CONTINUE_LINKING); | |
956 | cond_stmt = gimple_build_cond_from_tree (cost_pre_condition, | |
957 | NULL_TREE, NULL_TREE); | |
ebfd146a | 958 | gsi_insert_after (&gsi, cond_stmt, GSI_NEW_STMT); |
b8698a0f | 959 | |
ebfd146a IR |
960 | var = create_tmp_var (TREE_TYPE (scalar_loop_iters), |
961 | "prologue_after_cost_adjust"); | |
b8698a0f | 962 | prologue_after_cost_adjust_name = |
ebfd146a IR |
963 | force_gimple_operand (scalar_loop_iters, &stmts, false, var); |
964 | ||
965 | gsi = gsi_last_bb (then_bb); | |
966 | if (stmts) | |
967 | gsi_insert_seq_after (&gsi, stmts, GSI_NEW_STMT); | |
968 | ||
969 | newphi = create_phi_node (var, bb_before_first_loop); | |
b8698a0f | 970 | add_phi_arg (newphi, prologue_after_cost_adjust_name, e_fallthru, |
9e227d60 DC |
971 | UNKNOWN_LOCATION); |
972 | add_phi_arg (newphi, *first_niters, e_false, UNKNOWN_LOCATION); | |
ebfd146a | 973 | |
5d2eb24b | 974 | *first_niters = PHI_RESULT (newphi); |
ebfd146a IR |
975 | } |
976 | ||
ebfd146a IR |
977 | /* Function slpeel_tree_peel_loop_to_edge. |
978 | ||
979 | Peel the first (last) iterations of LOOP into a new prolog (epilog) loop | |
980 | that is placed on the entry (exit) edge E of LOOP. After this transformation | |
981 | we have two loops one after the other - first-loop iterates FIRST_NITERS | |
982 | times, and second-loop iterates the remainder NITERS - FIRST_NITERS times. | |
b8698a0f | 983 | If the cost model indicates that it is profitable to emit a scalar |
ebfd146a IR |
984 | loop instead of the vector one, then the prolog (epilog) loop will iterate |
985 | for the entire unchanged scalar iterations of the loop. | |
986 | ||
987 | Input: | |
988 | - LOOP: the loop to be peeled. | |
989 | - E: the exit or entry edge of LOOP. | |
990 | If it is the entry edge, we peel the first iterations of LOOP. In this | |
991 | case first-loop is LOOP, and second-loop is the newly created loop. | |
992 | If it is the exit edge, we peel the last iterations of LOOP. In this | |
993 | case, first-loop is the newly created loop, and second-loop is LOOP. | |
994 | - NITERS: the number of iterations that LOOP iterates. | |
995 | - FIRST_NITERS: the number of iterations that the first-loop should iterate. | |
996 | - UPDATE_FIRST_LOOP_COUNT: specified whether this function is responsible | |
997 | for updating the loop bound of the first-loop to FIRST_NITERS. If it | |
998 | is false, the caller of this function may want to take care of this | |
999 | (this can be useful if we don't want new stmts added to first-loop). | |
1000 | - TH: cost model profitability threshold of iterations for vectorization. | |
1001 | - CHECK_PROFITABILITY: specify whether cost model check has not occurred | |
1002 | during versioning and hence needs to occur during | |
b8698a0f | 1003 | prologue generation or whether cost model check |
ebfd146a IR |
1004 | has not occurred during prologue generation and hence |
1005 | needs to occur during epilogue generation. | |
e78410bf JH |
1006 | - BOUND1 is the upper bound on number of iterations of the first loop (if known) |
1007 | - BOUND2 is the upper bound on number of iterations of the second loop (if known) | |
b8698a0f | 1008 | |
ebfd146a IR |
1009 | |
1010 | Output: | |
1011 | The function returns a pointer to the new loop-copy, or NULL if it failed | |
1012 | to perform the transformation. | |
1013 | ||
1014 | The function generates two if-then-else guards: one before the first loop, | |
1015 | and the other before the second loop: | |
1016 | The first guard is: | |
1017 | if (FIRST_NITERS == 0) then skip the first loop, | |
1018 | and go directly to the second loop. | |
1019 | The second guard is: | |
1020 | if (FIRST_NITERS == NITERS) then skip the second loop. | |
1021 | ||
86290011 RG |
1022 | If the optional COND_EXPR and COND_EXPR_STMT_LIST arguments are given |
1023 | then the generated condition is combined with COND_EXPR and the | |
1024 | statements in COND_EXPR_STMT_LIST are emitted together with it. | |
1025 | ||
ebfd146a IR |
1026 | FORNOW only simple loops are supported (see slpeel_can_duplicate_loop_p). |
1027 | FORNOW the resulting code will not be in loop-closed-ssa form. | |
1028 | */ | |
1029 | ||
1030 | static struct loop* | |
b8698a0f | 1031 | slpeel_tree_peel_loop_to_edge (struct loop *loop, |
5d2eb24b | 1032 | edge e, tree *first_niters, |
ebfd146a | 1033 | tree niters, bool update_first_loop_count, |
86290011 | 1034 | unsigned int th, bool check_profitability, |
e78410bf JH |
1035 | tree cond_expr, gimple_seq cond_expr_stmt_list, |
1036 | int bound1, int bound2) | |
ebfd146a IR |
1037 | { |
1038 | struct loop *new_loop = NULL, *first_loop, *second_loop; | |
1039 | edge skip_e; | |
1040 | tree pre_condition = NULL_TREE; | |
ebfd146a IR |
1041 | basic_block bb_before_second_loop, bb_after_second_loop; |
1042 | basic_block bb_before_first_loop; | |
1043 | basic_block bb_between_loops; | |
1044 | basic_block new_exit_bb; | |
e20f6b4b | 1045 | gimple_stmt_iterator gsi; |
ebfd146a IR |
1046 | edge exit_e = single_exit (loop); |
1047 | LOC loop_loc; | |
1048 | tree cost_pre_condition = NULL_TREE; | |
e78410bf JH |
1049 | /* There are many aspects to how likely the first loop is going to be executed. |
1050 | Without histogram we can't really do good job. Simply set it to | |
1051 | 2/3, so the first loop is not reordered to the end of function and | |
1052 | the hot path through stays short. */ | |
1053 | int first_guard_probability = 2 * REG_BR_PROB_BASE / 3; | |
1054 | int second_guard_probability = 2 * REG_BR_PROB_BASE / 3; | |
1055 | int probability_of_second_loop; | |
b8698a0f | 1056 | |
ebfd146a IR |
1057 | if (!slpeel_can_duplicate_loop_p (loop, e)) |
1058 | return NULL; | |
b8698a0f | 1059 | |
141310ef RB |
1060 | /* We might have a queued need to update virtual SSA form. As we |
1061 | delete the update SSA machinery below after doing a regular | |
1062 | incremental SSA update during loop copying make sure we don't | |
1063 | lose that fact. | |
1064 | ??? Needing to update virtual SSA form by renaming is unfortunate | |
1065 | but not all of the vectorizer code inserting new loads / stores | |
1066 | properly assigns virtual operands to those statements. */ | |
1067 | update_ssa (TODO_update_ssa_only_virtuals); | |
1068 | ||
e20f6b4b JJ |
1069 | /* If the loop has a virtual PHI, but exit bb doesn't, create a virtual PHI |
1070 | in the exit bb and rename all the uses after the loop. This simplifies | |
1071 | the *guard[12] routines, which assume loop closed SSA form for all PHIs | |
1072 | (but normally loop closed SSA form doesn't require virtual PHIs to be | |
1073 | in the same form). Doing this early simplifies the checking what | |
1074 | uses should be renamed. */ | |
1075 | for (gsi = gsi_start_phis (loop->header); !gsi_end_p (gsi); gsi_next (&gsi)) | |
ea057359 | 1076 | if (virtual_operand_p (gimple_phi_result (gsi_stmt (gsi)))) |
e20f6b4b JJ |
1077 | { |
1078 | gimple phi = gsi_stmt (gsi); | |
1079 | for (gsi = gsi_start_phis (exit_e->dest); | |
1080 | !gsi_end_p (gsi); gsi_next (&gsi)) | |
ea057359 | 1081 | if (virtual_operand_p (gimple_phi_result (gsi_stmt (gsi)))) |
e20f6b4b JJ |
1082 | break; |
1083 | if (gsi_end_p (gsi)) | |
1084 | { | |
070ecdfd RG |
1085 | tree new_vop = copy_ssa_name (PHI_RESULT (phi), NULL); |
1086 | gimple new_phi = create_phi_node (new_vop, exit_e->dest); | |
e20f6b4b JJ |
1087 | tree vop = PHI_ARG_DEF_FROM_EDGE (phi, EDGE_SUCC (loop->latch, 0)); |
1088 | imm_use_iterator imm_iter; | |
1089 | gimple stmt; | |
e20f6b4b JJ |
1090 | use_operand_p use_p; |
1091 | ||
9e227d60 | 1092 | add_phi_arg (new_phi, vop, exit_e, UNKNOWN_LOCATION); |
e20f6b4b JJ |
1093 | gimple_phi_set_result (new_phi, new_vop); |
1094 | FOR_EACH_IMM_USE_STMT (stmt, imm_iter, vop) | |
1095 | if (stmt != new_phi && gimple_bb (stmt) != loop->header) | |
1096 | FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter) | |
1097 | SET_USE (use_p, new_vop); | |
1098 | } | |
1099 | break; | |
1100 | } | |
ebfd146a IR |
1101 | |
1102 | /* 1. Generate a copy of LOOP and put it on E (E is the entry/exit of LOOP). | |
1103 | Resulting CFG would be: | |
1104 | ||
1105 | first_loop: | |
1106 | do { | |
1107 | } while ... | |
1108 | ||
1109 | second_loop: | |
1110 | do { | |
1111 | } while ... | |
1112 | ||
1113 | orig_exit_bb: | |
1114 | */ | |
b8698a0f | 1115 | |
ebfd146a IR |
1116 | if (!(new_loop = slpeel_tree_duplicate_loop_to_edge_cfg (loop, e))) |
1117 | { | |
1118 | loop_loc = find_loop_location (loop); | |
78c60e3d SS |
1119 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, loop_loc, |
1120 | "tree_duplicate_loop_to_edge_cfg failed.\n"); | |
ebfd146a IR |
1121 | return NULL; |
1122 | } | |
b8698a0f | 1123 | |
684f25f4 AO |
1124 | if (MAY_HAVE_DEBUG_STMTS) |
1125 | { | |
9771b263 DN |
1126 | gcc_assert (!adjust_vec.exists ()); |
1127 | vec_stack_alloc (adjust_info, adjust_vec, 32); | |
684f25f4 AO |
1128 | } |
1129 | ||
ebfd146a IR |
1130 | if (e == exit_e) |
1131 | { | |
1132 | /* NEW_LOOP was placed after LOOP. */ | |
1133 | first_loop = loop; | |
1134 | second_loop = new_loop; | |
1135 | } | |
1136 | else | |
1137 | { | |
1138 | /* NEW_LOOP was placed before LOOP. */ | |
1139 | first_loop = new_loop; | |
1140 | second_loop = loop; | |
1141 | } | |
1142 | ||
ebfd146a IR |
1143 | /* 2. Add the guard code in one of the following ways: |
1144 | ||
1145 | 2.a Add the guard that controls whether the first loop is executed. | |
1146 | This occurs when this function is invoked for prologue or epilogue | |
1147 | generation and when the cost model check can be done at compile time. | |
1148 | ||
1149 | Resulting CFG would be: | |
1150 | ||
1151 | bb_before_first_loop: | |
1152 | if (FIRST_NITERS == 0) GOTO bb_before_second_loop | |
1153 | GOTO first-loop | |
1154 | ||
1155 | first_loop: | |
1156 | do { | |
1157 | } while ... | |
1158 | ||
1159 | bb_before_second_loop: | |
1160 | ||
1161 | second_loop: | |
1162 | do { | |
1163 | } while ... | |
1164 | ||
1165 | orig_exit_bb: | |
1166 | ||
1167 | 2.b Add the cost model check that allows the prologue | |
1168 | to iterate for the entire unchanged scalar | |
1169 | iterations of the loop in the event that the cost | |
1170 | model indicates that the scalar loop is more | |
1171 | profitable than the vector one. This occurs when | |
1172 | this function is invoked for prologue generation | |
1173 | and the cost model check needs to be done at run | |
1174 | time. | |
1175 | ||
1176 | Resulting CFG after prologue peeling would be: | |
1177 | ||
1178 | if (scalar_loop_iterations <= th) | |
1179 | FIRST_NITERS = scalar_loop_iterations | |
1180 | ||
1181 | bb_before_first_loop: | |
1182 | if (FIRST_NITERS == 0) GOTO bb_before_second_loop | |
1183 | GOTO first-loop | |
1184 | ||
1185 | first_loop: | |
1186 | do { | |
1187 | } while ... | |
1188 | ||
1189 | bb_before_second_loop: | |
1190 | ||
1191 | second_loop: | |
1192 | do { | |
1193 | } while ... | |
1194 | ||
1195 | orig_exit_bb: | |
1196 | ||
1197 | 2.c Add the cost model check that allows the epilogue | |
1198 | to iterate for the entire unchanged scalar | |
1199 | iterations of the loop in the event that the cost | |
1200 | model indicates that the scalar loop is more | |
1201 | profitable than the vector one. This occurs when | |
1202 | this function is invoked for epilogue generation | |
1203 | and the cost model check needs to be done at run | |
86290011 RG |
1204 | time. This check is combined with any pre-existing |
1205 | check in COND_EXPR to avoid versioning. | |
ebfd146a IR |
1206 | |
1207 | Resulting CFG after prologue peeling would be: | |
1208 | ||
1209 | bb_before_first_loop: | |
1210 | if ((scalar_loop_iterations <= th) | |
1211 | || | |
1212 | FIRST_NITERS == 0) GOTO bb_before_second_loop | |
1213 | GOTO first-loop | |
1214 | ||
1215 | first_loop: | |
1216 | do { | |
1217 | } while ... | |
1218 | ||
1219 | bb_before_second_loop: | |
1220 | ||
1221 | second_loop: | |
1222 | do { | |
1223 | } while ... | |
1224 | ||
1225 | orig_exit_bb: | |
1226 | */ | |
1227 | ||
1228 | bb_before_first_loop = split_edge (loop_preheader_edge (first_loop)); | |
2cfc56b9 RB |
1229 | /* Loop copying insterted a forwarder block for us here. */ |
1230 | bb_before_second_loop = single_exit (first_loop)->dest; | |
ebfd146a | 1231 | |
e78410bf JH |
1232 | probability_of_second_loop = (inverse_probability (first_guard_probability) |
1233 | + combine_probabilities (second_guard_probability, | |
1234 | first_guard_probability)); | |
1235 | /* Theoretically preheader edge of first loop and exit edge should have | |
1236 | same frequencies. Loop exit probablities are however easy to get wrong. | |
1237 | It is safer to copy value from original loop entry. */ | |
1238 | bb_before_second_loop->frequency | |
8ddb5a29 TJ |
1239 | = combine_probabilities (bb_before_first_loop->frequency, |
1240 | probability_of_second_loop); | |
e78410bf JH |
1241 | bb_before_second_loop->count |
1242 | = apply_probability (bb_before_first_loop->count, | |
1243 | probability_of_second_loop); | |
1244 | single_succ_edge (bb_before_second_loop)->count | |
1245 | = bb_before_second_loop->count; | |
1246 | ||
ebfd146a IR |
1247 | /* Epilogue peeling. */ |
1248 | if (!update_first_loop_count) | |
1249 | { | |
1250 | pre_condition = | |
5d2eb24b IR |
1251 | fold_build2 (LE_EXPR, boolean_type_node, *first_niters, |
1252 | build_int_cst (TREE_TYPE (*first_niters), 0)); | |
ebfd146a IR |
1253 | if (check_profitability) |
1254 | { | |
1255 | tree scalar_loop_iters | |
1256 | = unshare_expr (LOOP_VINFO_NITERS_UNCHANGED | |
1257 | (loop_vec_info_for_loop (loop))); | |
b8698a0f L |
1258 | cost_pre_condition = |
1259 | fold_build2 (LE_EXPR, boolean_type_node, scalar_loop_iters, | |
ebfd146a IR |
1260 | build_int_cst (TREE_TYPE (scalar_loop_iters), th)); |
1261 | ||
1262 | pre_condition = fold_build2 (TRUTH_OR_EXPR, boolean_type_node, | |
1263 | cost_pre_condition, pre_condition); | |
1264 | } | |
86290011 RG |
1265 | if (cond_expr) |
1266 | { | |
1267 | pre_condition = | |
1268 | fold_build2 (TRUTH_OR_EXPR, boolean_type_node, | |
1269 | pre_condition, | |
1270 | fold_build1 (TRUTH_NOT_EXPR, boolean_type_node, | |
1271 | cond_expr)); | |
1272 | } | |
ebfd146a IR |
1273 | } |
1274 | ||
b8698a0f | 1275 | /* Prologue peeling. */ |
ebfd146a IR |
1276 | else |
1277 | { | |
1278 | if (check_profitability) | |
1279 | set_prologue_iterations (bb_before_first_loop, first_niters, | |
e78410bf | 1280 | loop, th, first_guard_probability); |
ebfd146a IR |
1281 | |
1282 | pre_condition = | |
5d2eb24b IR |
1283 | fold_build2 (LE_EXPR, boolean_type_node, *first_niters, |
1284 | build_int_cst (TREE_TYPE (*first_niters), 0)); | |
ebfd146a IR |
1285 | } |
1286 | ||
1287 | skip_e = slpeel_add_loop_guard (bb_before_first_loop, pre_condition, | |
86290011 | 1288 | cond_expr_stmt_list, |
e78410bf JH |
1289 | bb_before_second_loop, bb_before_first_loop, |
1290 | inverse_probability (first_guard_probability)); | |
1291 | scale_loop_profile (first_loop, first_guard_probability, | |
1292 | check_profitability && (int)th > bound1 ? th : bound1); | |
ebfd146a IR |
1293 | slpeel_update_phi_nodes_for_guard1 (skip_e, first_loop, |
1294 | first_loop == new_loop, | |
c334023f | 1295 | &new_exit_bb); |
ebfd146a IR |
1296 | |
1297 | ||
1298 | /* 3. Add the guard that controls whether the second loop is executed. | |
1299 | Resulting CFG would be: | |
1300 | ||
1301 | bb_before_first_loop: | |
1302 | if (FIRST_NITERS == 0) GOTO bb_before_second_loop (skip first loop) | |
1303 | GOTO first-loop | |
1304 | ||
1305 | first_loop: | |
1306 | do { | |
1307 | } while ... | |
1308 | ||
1309 | bb_between_loops: | |
1310 | if (FIRST_NITERS == NITERS) GOTO bb_after_second_loop (skip second loop) | |
1311 | GOTO bb_before_second_loop | |
1312 | ||
1313 | bb_before_second_loop: | |
1314 | ||
1315 | second_loop: | |
1316 | do { | |
1317 | } while ... | |
1318 | ||
1319 | bb_after_second_loop: | |
1320 | ||
1321 | orig_exit_bb: | |
1322 | */ | |
1323 | ||
1324 | bb_between_loops = new_exit_bb; | |
1325 | bb_after_second_loop = split_edge (single_exit (second_loop)); | |
1326 | ||
b8698a0f | 1327 | pre_condition = |
5d2eb24b | 1328 | fold_build2 (EQ_EXPR, boolean_type_node, *first_niters, niters); |
86290011 | 1329 | skip_e = slpeel_add_loop_guard (bb_between_loops, pre_condition, NULL, |
e78410bf JH |
1330 | bb_after_second_loop, bb_before_first_loop, |
1331 | inverse_probability (second_guard_probability)); | |
1332 | scale_loop_profile (second_loop, probability_of_second_loop, bound2); | |
ebfd146a IR |
1333 | slpeel_update_phi_nodes_for_guard2 (skip_e, second_loop, |
1334 | second_loop == new_loop, &new_exit_bb); | |
1335 | ||
1336 | /* 4. Make first-loop iterate FIRST_NITERS times, if requested. | |
1337 | */ | |
1338 | if (update_first_loop_count) | |
5d2eb24b | 1339 | slpeel_make_loop_iterate_ntimes (first_loop, *first_niters); |
ebfd146a | 1340 | |
040d39ee RG |
1341 | delete_update_ssa (); |
1342 | ||
684f25f4 AO |
1343 | adjust_vec_debug_stmts (); |
1344 | ||
ebfd146a IR |
1345 | return new_loop; |
1346 | } | |
1347 | ||
1348 | /* Function vect_get_loop_location. | |
1349 | ||
1350 | Extract the location of the loop in the source code. | |
1351 | If the loop is not well formed for vectorization, an estimated | |
1352 | location is calculated. | |
1353 | Return the loop location if succeed and NULL if not. */ | |
1354 | ||
1355 | LOC | |
1356 | find_loop_location (struct loop *loop) | |
1357 | { | |
1358 | gimple stmt = NULL; | |
1359 | basic_block bb; | |
1360 | gimple_stmt_iterator si; | |
1361 | ||
1362 | if (!loop) | |
1363 | return UNKNOWN_LOC; | |
1364 | ||
1365 | stmt = get_loop_exit_condition (loop); | |
1366 | ||
502498d5 JJ |
1367 | if (stmt |
1368 | && LOCATION_LOCUS (gimple_location (stmt)) > BUILTINS_LOCATION) | |
ebfd146a IR |
1369 | return gimple_location (stmt); |
1370 | ||
1371 | /* If we got here the loop is probably not "well formed", | |
1372 | try to estimate the loop location */ | |
1373 | ||
1374 | if (!loop->header) | |
1375 | return UNKNOWN_LOC; | |
1376 | ||
1377 | bb = loop->header; | |
1378 | ||
1379 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
1380 | { | |
1381 | stmt = gsi_stmt (si); | |
502498d5 | 1382 | if (LOCATION_LOCUS (gimple_location (stmt)) > BUILTINS_LOCATION) |
ebfd146a IR |
1383 | return gimple_location (stmt); |
1384 | } | |
1385 | ||
1386 | return UNKNOWN_LOC; | |
1387 | } | |
1388 | ||
1389 | ||
1390 | /* This function builds ni_name = number of iterations loop executes | |
86290011 RG |
1391 | on the loop preheader. If SEQ is given the stmt is instead emitted |
1392 | there. */ | |
ebfd146a IR |
1393 | |
1394 | static tree | |
86290011 | 1395 | vect_build_loop_niters (loop_vec_info loop_vinfo, gimple_seq seq) |
ebfd146a IR |
1396 | { |
1397 | tree ni_name, var; | |
1398 | gimple_seq stmts = NULL; | |
1399 | edge pe; | |
1400 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
1401 | tree ni = unshare_expr (LOOP_VINFO_NITERS (loop_vinfo)); | |
1402 | ||
1403 | var = create_tmp_var (TREE_TYPE (ni), "niters"); | |
ebfd146a IR |
1404 | ni_name = force_gimple_operand (ni, &stmts, false, var); |
1405 | ||
1406 | pe = loop_preheader_edge (loop); | |
1407 | if (stmts) | |
1408 | { | |
86290011 RG |
1409 | if (seq) |
1410 | gimple_seq_add_seq (&seq, stmts); | |
1411 | else | |
1412 | { | |
1413 | basic_block new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
1414 | gcc_assert (!new_bb); | |
1415 | } | |
ebfd146a IR |
1416 | } |
1417 | ||
1418 | return ni_name; | |
1419 | } | |
1420 | ||
1421 | ||
1422 | /* This function generates the following statements: | |
1423 | ||
1424 | ni_name = number of iterations loop executes | |
1425 | ratio = ni_name / vf | |
1426 | ratio_mult_vf_name = ratio * vf | |
1427 | ||
86290011 RG |
1428 | and places them at the loop preheader edge or in COND_EXPR_STMT_LIST |
1429 | if that is non-NULL. */ | |
ebfd146a | 1430 | |
b8698a0f L |
1431 | static void |
1432 | vect_generate_tmps_on_preheader (loop_vec_info loop_vinfo, | |
ebfd146a | 1433 | tree *ni_name_ptr, |
b8698a0f | 1434 | tree *ratio_mult_vf_name_ptr, |
86290011 RG |
1435 | tree *ratio_name_ptr, |
1436 | gimple_seq cond_expr_stmt_list) | |
ebfd146a IR |
1437 | { |
1438 | ||
1439 | edge pe; | |
1440 | basic_block new_bb; | |
1441 | gimple_seq stmts; | |
48df3fa6 | 1442 | tree ni_name, ni_minus_gap_name; |
ebfd146a IR |
1443 | tree var; |
1444 | tree ratio_name; | |
1445 | tree ratio_mult_vf_name; | |
1446 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
1447 | tree ni = LOOP_VINFO_NITERS (loop_vinfo); | |
1448 | int vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
1449 | tree log_vf; | |
1450 | ||
1451 | pe = loop_preheader_edge (loop); | |
1452 | ||
b8698a0f | 1453 | /* Generate temporary variable that contains |
ebfd146a IR |
1454 | number of iterations loop executes. */ |
1455 | ||
86290011 | 1456 | ni_name = vect_build_loop_niters (loop_vinfo, cond_expr_stmt_list); |
ebfd146a IR |
1457 | log_vf = build_int_cst (TREE_TYPE (ni), exact_log2 (vf)); |
1458 | ||
48df3fa6 IR |
1459 | /* If epilogue loop is required because of data accesses with gaps, we |
1460 | subtract one iteration from the total number of iterations here for | |
1461 | correct calculation of RATIO. */ | |
1462 | if (LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo)) | |
1463 | { | |
1464 | ni_minus_gap_name = fold_build2 (MINUS_EXPR, TREE_TYPE (ni_name), | |
1465 | ni_name, | |
1466 | build_one_cst (TREE_TYPE (ni_name))); | |
1467 | if (!is_gimple_val (ni_minus_gap_name)) | |
1468 | { | |
1469 | var = create_tmp_var (TREE_TYPE (ni), "ni_gap"); | |
48df3fa6 IR |
1470 | |
1471 | stmts = NULL; | |
1472 | ni_minus_gap_name = force_gimple_operand (ni_minus_gap_name, &stmts, | |
1473 | true, var); | |
1474 | if (cond_expr_stmt_list) | |
1475 | gimple_seq_add_seq (&cond_expr_stmt_list, stmts); | |
1476 | else | |
1477 | { | |
1478 | pe = loop_preheader_edge (loop); | |
1479 | new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
1480 | gcc_assert (!new_bb); | |
1481 | } | |
1482 | } | |
1483 | } | |
1484 | else | |
1485 | ni_minus_gap_name = ni_name; | |
1486 | ||
ebfd146a IR |
1487 | /* Create: ratio = ni >> log2(vf) */ |
1488 | ||
48df3fa6 IR |
1489 | ratio_name = fold_build2 (RSHIFT_EXPR, TREE_TYPE (ni_minus_gap_name), |
1490 | ni_minus_gap_name, log_vf); | |
ebfd146a IR |
1491 | if (!is_gimple_val (ratio_name)) |
1492 | { | |
1493 | var = create_tmp_var (TREE_TYPE (ni), "bnd"); | |
ebfd146a IR |
1494 | |
1495 | stmts = NULL; | |
1496 | ratio_name = force_gimple_operand (ratio_name, &stmts, true, var); | |
86290011 RG |
1497 | if (cond_expr_stmt_list) |
1498 | gimple_seq_add_seq (&cond_expr_stmt_list, stmts); | |
1499 | else | |
1500 | { | |
1501 | pe = loop_preheader_edge (loop); | |
1502 | new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
1503 | gcc_assert (!new_bb); | |
1504 | } | |
ebfd146a | 1505 | } |
b8698a0f | 1506 | |
ebfd146a IR |
1507 | /* Create: ratio_mult_vf = ratio << log2 (vf). */ |
1508 | ||
1509 | ratio_mult_vf_name = fold_build2 (LSHIFT_EXPR, TREE_TYPE (ratio_name), | |
1510 | ratio_name, log_vf); | |
1511 | if (!is_gimple_val (ratio_mult_vf_name)) | |
1512 | { | |
1513 | var = create_tmp_var (TREE_TYPE (ni), "ratio_mult_vf"); | |
ebfd146a IR |
1514 | |
1515 | stmts = NULL; | |
1516 | ratio_mult_vf_name = force_gimple_operand (ratio_mult_vf_name, &stmts, | |
1517 | true, var); | |
86290011 RG |
1518 | if (cond_expr_stmt_list) |
1519 | gimple_seq_add_seq (&cond_expr_stmt_list, stmts); | |
1520 | else | |
1521 | { | |
1522 | pe = loop_preheader_edge (loop); | |
1523 | new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
1524 | gcc_assert (!new_bb); | |
1525 | } | |
ebfd146a IR |
1526 | } |
1527 | ||
1528 | *ni_name_ptr = ni_name; | |
1529 | *ratio_mult_vf_name_ptr = ratio_mult_vf_name; | |
1530 | *ratio_name_ptr = ratio_name; | |
b8698a0f L |
1531 | |
1532 | return; | |
ebfd146a IR |
1533 | } |
1534 | ||
1535 | /* Function vect_can_advance_ivs_p | |
1536 | ||
b8698a0f L |
1537 | In case the number of iterations that LOOP iterates is unknown at compile |
1538 | time, an epilog loop will be generated, and the loop induction variables | |
1539 | (IVs) will be "advanced" to the value they are supposed to take just before | |
ebfd146a IR |
1540 | the epilog loop. Here we check that the access function of the loop IVs |
1541 | and the expression that represents the loop bound are simple enough. | |
1542 | These restrictions will be relaxed in the future. */ | |
1543 | ||
b8698a0f | 1544 | bool |
ebfd146a IR |
1545 | vect_can_advance_ivs_p (loop_vec_info loop_vinfo) |
1546 | { | |
1547 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
1548 | basic_block bb = loop->header; | |
1549 | gimple phi; | |
1550 | gimple_stmt_iterator gsi; | |
1551 | ||
1552 | /* Analyze phi functions of the loop header. */ | |
1553 | ||
73fbfcad | 1554 | if (dump_enabled_p ()) |
78c60e3d | 1555 | dump_printf_loc (MSG_NOTE, vect_location, "vect_can_advance_ivs_p:"); |
ebfd146a IR |
1556 | for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
1557 | { | |
ebfd146a IR |
1558 | tree evolution_part; |
1559 | ||
1560 | phi = gsi_stmt (gsi); | |
73fbfcad | 1561 | if (dump_enabled_p ()) |
ebfd146a | 1562 | { |
78c60e3d SS |
1563 | dump_printf_loc (MSG_NOTE, vect_location, "Analyze phi: "); |
1564 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0); | |
ebfd146a IR |
1565 | } |
1566 | ||
1567 | /* Skip virtual phi's. The data dependences that are associated with | |
1568 | virtual defs/uses (i.e., memory accesses) are analyzed elsewhere. */ | |
1569 | ||
ea057359 | 1570 | if (virtual_operand_p (PHI_RESULT (phi))) |
ebfd146a | 1571 | { |
73fbfcad | 1572 | if (dump_enabled_p ()) |
78c60e3d SS |
1573 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1574 | "virtual phi. skip."); | |
ebfd146a IR |
1575 | continue; |
1576 | } | |
1577 | ||
1578 | /* Skip reduction phis. */ | |
1579 | ||
1580 | if (STMT_VINFO_DEF_TYPE (vinfo_for_stmt (phi)) == vect_reduction_def) | |
1581 | { | |
73fbfcad | 1582 | if (dump_enabled_p ()) |
78c60e3d SS |
1583 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1584 | "reduc phi. skip."); | |
ebfd146a IR |
1585 | continue; |
1586 | } | |
1587 | ||
1588 | /* Analyze the evolution function. */ | |
1589 | ||
afb119be RB |
1590 | evolution_part |
1591 | = STMT_VINFO_LOOP_PHI_EVOLUTION_PART (vinfo_for_stmt (phi)); | |
ebfd146a IR |
1592 | if (evolution_part == NULL_TREE) |
1593 | { | |
73fbfcad | 1594 | if (dump_enabled_p ()) |
afb119be RB |
1595 | dump_printf (MSG_MISSED_OPTIMIZATION, |
1596 | "No access function or evolution."); | |
ebfd146a IR |
1597 | return false; |
1598 | } | |
b8698a0f L |
1599 | |
1600 | /* FORNOW: We do not transform initial conditions of IVs | |
ebfd146a IR |
1601 | which evolution functions are a polynomial of degree >= 2. */ |
1602 | ||
1603 | if (tree_is_chrec (evolution_part)) | |
b8698a0f | 1604 | return false; |
ebfd146a IR |
1605 | } |
1606 | ||
1607 | return true; | |
1608 | } | |
1609 | ||
1610 | ||
1611 | /* Function vect_update_ivs_after_vectorizer. | |
1612 | ||
1613 | "Advance" the induction variables of LOOP to the value they should take | |
1614 | after the execution of LOOP. This is currently necessary because the | |
1615 | vectorizer does not handle induction variables that are used after the | |
1616 | loop. Such a situation occurs when the last iterations of LOOP are | |
1617 | peeled, because: | |
1618 | 1. We introduced new uses after LOOP for IVs that were not originally used | |
1619 | after LOOP: the IVs of LOOP are now used by an epilog loop. | |
1620 | 2. LOOP is going to be vectorized; this means that it will iterate N/VF | |
1621 | times, whereas the loop IVs should be bumped N times. | |
1622 | ||
1623 | Input: | |
1624 | - LOOP - a loop that is going to be vectorized. The last few iterations | |
1625 | of LOOP were peeled. | |
1626 | - NITERS - the number of iterations that LOOP executes (before it is | |
1627 | vectorized). i.e, the number of times the ivs should be bumped. | |
1628 | - UPDATE_E - a successor edge of LOOP->exit that is on the (only) path | |
1629 | coming out from LOOP on which there are uses of the LOOP ivs | |
1630 | (this is the path from LOOP->exit to epilog_loop->preheader). | |
1631 | ||
1632 | The new definitions of the ivs are placed in LOOP->exit. | |
1633 | The phi args associated with the edge UPDATE_E in the bb | |
1634 | UPDATE_E->dest are updated accordingly. | |
1635 | ||
1636 | Assumption 1: Like the rest of the vectorizer, this function assumes | |
1637 | a single loop exit that has a single predecessor. | |
1638 | ||
1639 | Assumption 2: The phi nodes in the LOOP header and in update_bb are | |
1640 | organized in the same order. | |
1641 | ||
1642 | Assumption 3: The access function of the ivs is simple enough (see | |
1643 | vect_can_advance_ivs_p). This assumption will be relaxed in the future. | |
1644 | ||
1645 | Assumption 4: Exactly one of the successors of LOOP exit-bb is on a path | |
b8698a0f | 1646 | coming out of LOOP on which the ivs of LOOP are used (this is the path |
ebfd146a IR |
1647 | that leads to the epilog loop; other paths skip the epilog loop). This |
1648 | path starts with the edge UPDATE_E, and its destination (denoted update_bb) | |
1649 | needs to have its phis updated. | |
1650 | */ | |
1651 | ||
1652 | static void | |
b8698a0f | 1653 | vect_update_ivs_after_vectorizer (loop_vec_info loop_vinfo, tree niters, |
ebfd146a IR |
1654 | edge update_e) |
1655 | { | |
1656 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
1657 | basic_block exit_bb = single_exit (loop)->dest; | |
1658 | gimple phi, phi1; | |
1659 | gimple_stmt_iterator gsi, gsi1; | |
1660 | basic_block update_bb = update_e->dest; | |
1661 | ||
1662 | /* gcc_assert (vect_can_advance_ivs_p (loop_vinfo)); */ | |
1663 | ||
1664 | /* Make sure there exists a single-predecessor exit bb: */ | |
1665 | gcc_assert (single_pred_p (exit_bb)); | |
1666 | ||
1667 | for (gsi = gsi_start_phis (loop->header), gsi1 = gsi_start_phis (update_bb); | |
1668 | !gsi_end_p (gsi) && !gsi_end_p (gsi1); | |
1669 | gsi_next (&gsi), gsi_next (&gsi1)) | |
1670 | { | |
ebfd146a | 1671 | tree init_expr; |
550918ca RG |
1672 | tree step_expr, off; |
1673 | tree type; | |
ebfd146a IR |
1674 | tree var, ni, ni_name; |
1675 | gimple_stmt_iterator last_gsi; | |
0ac168a1 | 1676 | stmt_vec_info stmt_info; |
ebfd146a IR |
1677 | |
1678 | phi = gsi_stmt (gsi); | |
1679 | phi1 = gsi_stmt (gsi1); | |
73fbfcad | 1680 | if (dump_enabled_p ()) |
ebfd146a | 1681 | { |
78c60e3d SS |
1682 | dump_printf_loc (MSG_NOTE, vect_location, |
1683 | "vect_update_ivs_after_vectorizer: phi: "); | |
1684 | dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0); | |
ebfd146a IR |
1685 | } |
1686 | ||
1687 | /* Skip virtual phi's. */ | |
ea057359 | 1688 | if (virtual_operand_p (PHI_RESULT (phi))) |
ebfd146a | 1689 | { |
73fbfcad | 1690 | if (dump_enabled_p ()) |
78c60e3d SS |
1691 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1692 | "virtual phi. skip."); | |
ebfd146a IR |
1693 | continue; |
1694 | } | |
1695 | ||
1696 | /* Skip reduction phis. */ | |
0ac168a1 RG |
1697 | stmt_info = vinfo_for_stmt (phi); |
1698 | if (STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def) | |
b8698a0f | 1699 | { |
73fbfcad | 1700 | if (dump_enabled_p ()) |
78c60e3d SS |
1701 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, |
1702 | "reduc phi. skip."); | |
ebfd146a | 1703 | continue; |
b8698a0f | 1704 | } |
ebfd146a | 1705 | |
0ac168a1 RG |
1706 | type = TREE_TYPE (gimple_phi_result (phi)); |
1707 | step_expr = STMT_VINFO_LOOP_PHI_EVOLUTION_PART (stmt_info); | |
1708 | step_expr = unshare_expr (step_expr); | |
b8698a0f | 1709 | |
ebfd146a IR |
1710 | /* FORNOW: We do not support IVs whose evolution function is a polynomial |
1711 | of degree >= 2 or exponential. */ | |
0ac168a1 | 1712 | gcc_assert (!tree_is_chrec (step_expr)); |
ebfd146a | 1713 | |
0ac168a1 | 1714 | init_expr = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop)); |
ebfd146a | 1715 | |
550918ca RG |
1716 | off = fold_build2 (MULT_EXPR, TREE_TYPE (step_expr), |
1717 | fold_convert (TREE_TYPE (step_expr), niters), | |
1718 | step_expr); | |
0ac168a1 | 1719 | if (POINTER_TYPE_P (type)) |
5d49b6a7 | 1720 | ni = fold_build_pointer_plus (init_expr, off); |
ebfd146a | 1721 | else |
0ac168a1 RG |
1722 | ni = fold_build2 (PLUS_EXPR, type, |
1723 | init_expr, fold_convert (type, off)); | |
ebfd146a | 1724 | |
0ac168a1 | 1725 | var = create_tmp_var (type, "tmp"); |
ebfd146a IR |
1726 | |
1727 | last_gsi = gsi_last_bb (exit_bb); | |
1728 | ni_name = force_gimple_operand_gsi (&last_gsi, ni, false, var, | |
1729 | true, GSI_SAME_STMT); | |
b8698a0f | 1730 | |
ebfd146a | 1731 | /* Fix phi expressions in the successor bb. */ |
684f25f4 | 1732 | adjust_phi_and_debug_stmts (phi1, update_e, ni_name); |
ebfd146a IR |
1733 | } |
1734 | } | |
1735 | ||
ebfd146a IR |
1736 | /* Function vect_do_peeling_for_loop_bound |
1737 | ||
1738 | Peel the last iterations of the loop represented by LOOP_VINFO. | |
b8698a0f | 1739 | The peeled iterations form a new epilog loop. Given that the loop now |
ebfd146a IR |
1740 | iterates NITERS times, the new epilog loop iterates |
1741 | NITERS % VECTORIZATION_FACTOR times. | |
b8698a0f L |
1742 | |
1743 | The original loop will later be made to iterate | |
86290011 RG |
1744 | NITERS / VECTORIZATION_FACTOR times (this value is placed into RATIO). |
1745 | ||
1746 | COND_EXPR and COND_EXPR_STMT_LIST are combined with a new generated | |
1747 | test. */ | |
ebfd146a | 1748 | |
b8698a0f | 1749 | void |
86290011 | 1750 | vect_do_peeling_for_loop_bound (loop_vec_info loop_vinfo, tree *ratio, |
368117e8 | 1751 | unsigned int th, bool check_profitability) |
ebfd146a IR |
1752 | { |
1753 | tree ni_name, ratio_mult_vf_name; | |
1754 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
1755 | struct loop *new_loop; | |
1756 | edge update_e; | |
1757 | basic_block preheader; | |
1758 | int loop_num; | |
d68d56b5 | 1759 | int max_iter; |
368117e8 RG |
1760 | tree cond_expr = NULL_TREE; |
1761 | gimple_seq cond_expr_stmt_list = NULL; | |
ebfd146a | 1762 | |
73fbfcad | 1763 | if (dump_enabled_p ()) |
ccb3ad87 | 1764 | dump_printf_loc (MSG_NOTE, vect_location, |
78c60e3d | 1765 | "=== vect_do_peeling_for_loop_bound ==="); |
ebfd146a IR |
1766 | |
1767 | initialize_original_copy_tables (); | |
1768 | ||
1769 | /* Generate the following variables on the preheader of original loop: | |
b8698a0f | 1770 | |
ebfd146a IR |
1771 | ni_name = number of iteration the original loop executes |
1772 | ratio = ni_name / vf | |
1773 | ratio_mult_vf_name = ratio * vf */ | |
1774 | vect_generate_tmps_on_preheader (loop_vinfo, &ni_name, | |
86290011 RG |
1775 | &ratio_mult_vf_name, ratio, |
1776 | cond_expr_stmt_list); | |
ebfd146a | 1777 | |
b8698a0f | 1778 | loop_num = loop->num; |
ebfd146a | 1779 | |
ebfd146a | 1780 | new_loop = slpeel_tree_peel_loop_to_edge (loop, single_exit (loop), |
5d2eb24b | 1781 | &ratio_mult_vf_name, ni_name, false, |
86290011 | 1782 | th, check_profitability, |
e78410bf JH |
1783 | cond_expr, cond_expr_stmt_list, |
1784 | 0, LOOP_VINFO_VECT_FACTOR (loop_vinfo)); | |
ebfd146a IR |
1785 | gcc_assert (new_loop); |
1786 | gcc_assert (loop_num == loop->num); | |
1787 | #ifdef ENABLE_CHECKING | |
1788 | slpeel_verify_cfg_after_peeling (loop, new_loop); | |
1789 | #endif | |
1790 | ||
1791 | /* A guard that controls whether the new_loop is to be executed or skipped | |
1792 | is placed in LOOP->exit. LOOP->exit therefore has two successors - one | |
1793 | is the preheader of NEW_LOOP, where the IVs from LOOP are used. The other | |
1794 | is a bb after NEW_LOOP, where these IVs are not used. Find the edge that | |
1795 | is on the path where the LOOP IVs are used and need to be updated. */ | |
1796 | ||
1797 | preheader = loop_preheader_edge (new_loop)->src; | |
1798 | if (EDGE_PRED (preheader, 0)->src == single_exit (loop)->dest) | |
1799 | update_e = EDGE_PRED (preheader, 0); | |
1800 | else | |
1801 | update_e = EDGE_PRED (preheader, 1); | |
1802 | ||
b8698a0f | 1803 | /* Update IVs of original loop as if they were advanced |
ebfd146a | 1804 | by ratio_mult_vf_name steps. */ |
b8698a0f | 1805 | vect_update_ivs_after_vectorizer (loop_vinfo, ratio_mult_vf_name, update_e); |
ebfd146a | 1806 | |
22458c5a JH |
1807 | /* For vectorization factor N, we need to copy last N-1 values in epilogue |
1808 | and this means N-2 loopback edge executions. | |
1809 | ||
1810 | PEELING_FOR_GAPS works by subtracting last iteration and thus the epilogue | |
1811 | will execute at least LOOP_VINFO_VECT_FACTOR times. */ | |
1812 | max_iter = (LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) | |
1813 | ? LOOP_VINFO_VECT_FACTOR (loop_vinfo) * 2 | |
1814 | : LOOP_VINFO_VECT_FACTOR (loop_vinfo)) - 2; | |
368117e8 | 1815 | if (check_profitability) |
22458c5a | 1816 | max_iter = MAX (max_iter, (int) th - 1); |
27bcd47c | 1817 | record_niter_bound (new_loop, double_int::from_shwi (max_iter), false, true); |
ccb3ad87 | 1818 | dump_printf (MSG_NOTE, |
78c60e3d SS |
1819 | "Setting upper bound of nb iterations for epilogue " |
1820 | "loop to %d\n", max_iter); | |
7d5a99f4 | 1821 | |
ebfd146a IR |
1822 | /* After peeling we have to reset scalar evolution analyzer. */ |
1823 | scev_reset (); | |
1824 | ||
1825 | free_original_copy_tables (); | |
1826 | } | |
1827 | ||
1828 | ||
1829 | /* Function vect_gen_niters_for_prolog_loop | |
1830 | ||
1831 | Set the number of iterations for the loop represented by LOOP_VINFO | |
1832 | to the minimum between LOOP_NITERS (the original iteration count of the loop) | |
1833 | and the misalignment of DR - the data reference recorded in | |
b8698a0f | 1834 | LOOP_VINFO_UNALIGNED_DR (LOOP_VINFO). As a result, after the execution of |
ebfd146a IR |
1835 | this loop, the data reference DR will refer to an aligned location. |
1836 | ||
1837 | The following computation is generated: | |
1838 | ||
1839 | If the misalignment of DR is known at compile time: | |
1840 | addr_mis = int mis = DR_MISALIGNMENT (dr); | |
1841 | Else, compute address misalignment in bytes: | |
5aea1e76 | 1842 | addr_mis = addr & (vectype_align - 1) |
ebfd146a IR |
1843 | |
1844 | prolog_niters = min (LOOP_NITERS, ((VF - addr_mis/elem_size)&(VF-1))/step) | |
1845 | ||
1846 | (elem_size = element type size; an element is the scalar element whose type | |
1847 | is the inner type of the vectype) | |
1848 | ||
1849 | When the step of the data-ref in the loop is not 1 (as in interleaved data | |
1850 | and SLP), the number of iterations of the prolog must be divided by the step | |
1851 | (which is equal to the size of interleaved group). | |
1852 | ||
1853 | The above formulas assume that VF == number of elements in the vector. This | |
1854 | may not hold when there are multiple-types in the loop. | |
1855 | In this case, for some data-references in the loop the VF does not represent | |
1856 | the number of elements that fit in the vector. Therefore, instead of VF we | |
1857 | use TYPE_VECTOR_SUBPARTS. */ | |
1858 | ||
b8698a0f | 1859 | static tree |
e78410bf | 1860 | vect_gen_niters_for_prolog_loop (loop_vec_info loop_vinfo, tree loop_niters, int *bound) |
ebfd146a IR |
1861 | { |
1862 | struct data_reference *dr = LOOP_VINFO_UNALIGNED_DR (loop_vinfo); | |
1863 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
1864 | tree var; | |
1865 | gimple_seq stmts; | |
1866 | tree iters, iters_name; | |
1867 | edge pe; | |
1868 | basic_block new_bb; | |
1869 | gimple dr_stmt = DR_STMT (dr); | |
1870 | stmt_vec_info stmt_info = vinfo_for_stmt (dr_stmt); | |
1871 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
1872 | int vectype_align = TYPE_ALIGN (vectype) / BITS_PER_UNIT; | |
1873 | tree niters_type = TREE_TYPE (loop_niters); | |
ebfd146a IR |
1874 | int nelements = TYPE_VECTOR_SUBPARTS (vectype); |
1875 | ||
b8698a0f | 1876 | pe = loop_preheader_edge (loop); |
ebfd146a IR |
1877 | |
1878 | if (LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo) > 0) | |
1879 | { | |
720f5239 | 1880 | int npeel = LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo); |
ebfd146a | 1881 | |
73fbfcad | 1882 | if (dump_enabled_p ()) |
ccb3ad87 | 1883 | dump_printf_loc (MSG_NOTE, vect_location, |
78c60e3d | 1884 | "known peeling = %d.", npeel); |
ebfd146a | 1885 | |
720f5239 | 1886 | iters = build_int_cst (niters_type, npeel); |
e78410bf | 1887 | *bound = LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo); |
ebfd146a IR |
1888 | } |
1889 | else | |
1890 | { | |
1891 | gimple_seq new_stmts = NULL; | |
d8ba5b19 RG |
1892 | bool negative = tree_int_cst_compare (DR_STEP (dr), size_zero_node) < 0; |
1893 | tree offset = negative | |
1894 | ? size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1) : NULL_TREE; | |
b8698a0f | 1895 | tree start_addr = vect_create_addr_base_for_vector_ref (dr_stmt, |
d8ba5b19 | 1896 | &new_stmts, offset, loop); |
96f9265a | 1897 | tree type = unsigned_type_for (TREE_TYPE (start_addr)); |
5aea1e76 UW |
1898 | tree vectype_align_minus_1 = build_int_cst (type, vectype_align - 1); |
1899 | HOST_WIDE_INT elem_size = | |
1900 | int_cst_value (TYPE_SIZE_UNIT (TREE_TYPE (vectype))); | |
1901 | tree elem_size_log = build_int_cst (type, exact_log2 (elem_size)); | |
ebfd146a IR |
1902 | tree nelements_minus_1 = build_int_cst (type, nelements - 1); |
1903 | tree nelements_tree = build_int_cst (type, nelements); | |
1904 | tree byte_misalign; | |
1905 | tree elem_misalign; | |
1906 | ||
1907 | new_bb = gsi_insert_seq_on_edge_immediate (pe, new_stmts); | |
1908 | gcc_assert (!new_bb); | |
b8698a0f | 1909 | |
5aea1e76 | 1910 | /* Create: byte_misalign = addr & (vectype_align - 1) */ |
b8698a0f | 1911 | byte_misalign = |
720f5239 | 1912 | fold_build2 (BIT_AND_EXPR, type, fold_convert (type, start_addr), |
5aea1e76 | 1913 | vectype_align_minus_1); |
b8698a0f | 1914 | |
ebfd146a IR |
1915 | /* Create: elem_misalign = byte_misalign / element_size */ |
1916 | elem_misalign = | |
1917 | fold_build2 (RSHIFT_EXPR, type, byte_misalign, elem_size_log); | |
1918 | ||
1919 | /* Create: (niters_type) (nelements - elem_misalign)&(nelements - 1) */ | |
d8ba5b19 RG |
1920 | if (negative) |
1921 | iters = fold_build2 (MINUS_EXPR, type, elem_misalign, nelements_tree); | |
1922 | else | |
1923 | iters = fold_build2 (MINUS_EXPR, type, nelements_tree, elem_misalign); | |
ebfd146a IR |
1924 | iters = fold_build2 (BIT_AND_EXPR, type, iters, nelements_minus_1); |
1925 | iters = fold_convert (niters_type, iters); | |
e78410bf | 1926 | *bound = nelements; |
ebfd146a IR |
1927 | } |
1928 | ||
1929 | /* Create: prolog_loop_niters = min (iters, loop_niters) */ | |
1930 | /* If the loop bound is known at compile time we already verified that it is | |
1931 | greater than vf; since the misalignment ('iters') is at most vf, there's | |
1932 | no need to generate the MIN_EXPR in this case. */ | |
1933 | if (TREE_CODE (loop_niters) != INTEGER_CST) | |
1934 | iters = fold_build2 (MIN_EXPR, niters_type, iters, loop_niters); | |
1935 | ||
73fbfcad | 1936 | if (dump_enabled_p ()) |
ebfd146a | 1937 | { |
ccb3ad87 | 1938 | dump_printf_loc (MSG_NOTE, vect_location, |
78c60e3d | 1939 | "niters for prolog loop: "); |
ccb3ad87 | 1940 | dump_generic_expr (MSG_NOTE, TDF_SLIM, iters); |
ebfd146a IR |
1941 | } |
1942 | ||
1943 | var = create_tmp_var (niters_type, "prolog_loop_niters"); | |
ebfd146a IR |
1944 | stmts = NULL; |
1945 | iters_name = force_gimple_operand (iters, &stmts, false, var); | |
1946 | ||
1947 | /* Insert stmt on loop preheader edge. */ | |
1948 | if (stmts) | |
1949 | { | |
1950 | basic_block new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
1951 | gcc_assert (!new_bb); | |
1952 | } | |
1953 | ||
b8698a0f | 1954 | return iters_name; |
ebfd146a IR |
1955 | } |
1956 | ||
1957 | ||
1958 | /* Function vect_update_init_of_dr | |
1959 | ||
1960 | NITERS iterations were peeled from LOOP. DR represents a data reference | |
1961 | in LOOP. This function updates the information recorded in DR to | |
b8698a0f | 1962 | account for the fact that the first NITERS iterations had already been |
ebfd146a IR |
1963 | executed. Specifically, it updates the OFFSET field of DR. */ |
1964 | ||
1965 | static void | |
1966 | vect_update_init_of_dr (struct data_reference *dr, tree niters) | |
1967 | { | |
1968 | tree offset = DR_OFFSET (dr); | |
b8698a0f | 1969 | |
ebfd146a IR |
1970 | niters = fold_build2 (MULT_EXPR, sizetype, |
1971 | fold_convert (sizetype, niters), | |
1972 | fold_convert (sizetype, DR_STEP (dr))); | |
587aa063 RG |
1973 | offset = fold_build2 (PLUS_EXPR, sizetype, |
1974 | fold_convert (sizetype, offset), niters); | |
ebfd146a IR |
1975 | DR_OFFSET (dr) = offset; |
1976 | } | |
1977 | ||
1978 | ||
1979 | /* Function vect_update_inits_of_drs | |
1980 | ||
b8698a0f L |
1981 | NITERS iterations were peeled from the loop represented by LOOP_VINFO. |
1982 | This function updates the information recorded for the data references in | |
1983 | the loop to account for the fact that the first NITERS iterations had | |
ebfd146a IR |
1984 | already been executed. Specifically, it updates the initial_condition of |
1985 | the access_function of all the data_references in the loop. */ | |
1986 | ||
1987 | static void | |
1988 | vect_update_inits_of_drs (loop_vec_info loop_vinfo, tree niters) | |
1989 | { | |
1990 | unsigned int i; | |
9771b263 | 1991 | vec<data_reference_p> datarefs = LOOP_VINFO_DATAREFS (loop_vinfo); |
ebfd146a | 1992 | struct data_reference *dr; |
78c60e3d | 1993 | |
73fbfcad | 1994 | if (dump_enabled_p ()) |
ccb3ad87 | 1995 | dump_printf_loc (MSG_NOTE, vect_location, |
78c60e3d | 1996 | "=== vect_update_inits_of_dr ==="); |
ebfd146a | 1997 | |
9771b263 | 1998 | FOR_EACH_VEC_ELT (datarefs, i, dr) |
ebfd146a IR |
1999 | vect_update_init_of_dr (dr, niters); |
2000 | } | |
2001 | ||
2002 | ||
2003 | /* Function vect_do_peeling_for_alignment | |
2004 | ||
2005 | Peel the first 'niters' iterations of the loop represented by LOOP_VINFO. | |
2006 | 'niters' is set to the misalignment of one of the data references in the | |
2007 | loop, thereby forcing it to refer to an aligned location at the beginning | |
2008 | of the execution of this loop. The data reference for which we are | |
2009 | peeling is recorded in LOOP_VINFO_UNALIGNED_DR. */ | |
2010 | ||
2011 | void | |
368117e8 RG |
2012 | vect_do_peeling_for_alignment (loop_vec_info loop_vinfo, |
2013 | unsigned int th, bool check_profitability) | |
ebfd146a IR |
2014 | { |
2015 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
2016 | tree niters_of_prolog_loop, ni_name; | |
2017 | tree n_iters; | |
b61b1f17 | 2018 | tree wide_prolog_niters; |
ebfd146a | 2019 | struct loop *new_loop; |
03fd03d5 | 2020 | int max_iter; |
e78410bf | 2021 | int bound = 0; |
ebfd146a | 2022 | |
73fbfcad | 2023 | if (dump_enabled_p ()) |
9cc1fb4b XDL |
2024 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, vect_location, |
2025 | "loop peeled for vectorization to enhance" | |
2026 | " alignment\n"); | |
ebfd146a IR |
2027 | |
2028 | initialize_original_copy_tables (); | |
2029 | ||
86290011 | 2030 | ni_name = vect_build_loop_niters (loop_vinfo, NULL); |
5d2eb24b | 2031 | niters_of_prolog_loop = vect_gen_niters_for_prolog_loop (loop_vinfo, |
e78410bf JH |
2032 | ni_name, |
2033 | &bound); | |
ebfd146a | 2034 | |
ebfd146a IR |
2035 | /* Peel the prolog loop and iterate it niters_of_prolog_loop. */ |
2036 | new_loop = | |
2037 | slpeel_tree_peel_loop_to_edge (loop, loop_preheader_edge (loop), | |
5d2eb24b | 2038 | &niters_of_prolog_loop, ni_name, true, |
e78410bf JH |
2039 | th, check_profitability, NULL_TREE, NULL, |
2040 | bound, | |
2041 | 0); | |
ebfd146a IR |
2042 | |
2043 | gcc_assert (new_loop); | |
2044 | #ifdef ENABLE_CHECKING | |
2045 | slpeel_verify_cfg_after_peeling (new_loop, loop); | |
2046 | #endif | |
22458c5a JH |
2047 | /* For vectorization factor N, we need to copy at most N-1 values |
2048 | for alignment and this means N-2 loopback edge executions. */ | |
2049 | max_iter = LOOP_VINFO_VECT_FACTOR (loop_vinfo) - 2; | |
368117e8 | 2050 | if (check_profitability) |
22458c5a | 2051 | max_iter = MAX (max_iter, (int) th - 1); |
27bcd47c | 2052 | record_niter_bound (new_loop, double_int::from_shwi (max_iter), false, true); |
ccb3ad87 | 2053 | dump_printf (MSG_NOTE, |
78c60e3d SS |
2054 | "Setting upper bound of nb iterations for prologue " |
2055 | "loop to %d\n", max_iter); | |
ebfd146a IR |
2056 | |
2057 | /* Update number of times loop executes. */ | |
2058 | n_iters = LOOP_VINFO_NITERS (loop_vinfo); | |
2059 | LOOP_VINFO_NITERS (loop_vinfo) = fold_build2 (MINUS_EXPR, | |
2060 | TREE_TYPE (n_iters), n_iters, niters_of_prolog_loop); | |
2061 | ||
5d2eb24b IR |
2062 | if (types_compatible_p (sizetype, TREE_TYPE (niters_of_prolog_loop))) |
2063 | wide_prolog_niters = niters_of_prolog_loop; | |
2064 | else | |
2065 | { | |
2066 | gimple_seq seq = NULL; | |
2067 | edge pe = loop_preheader_edge (loop); | |
2068 | tree wide_iters = fold_convert (sizetype, niters_of_prolog_loop); | |
2069 | tree var = create_tmp_var (sizetype, "prolog_loop_adjusted_niters"); | |
5d2eb24b IR |
2070 | wide_prolog_niters = force_gimple_operand (wide_iters, &seq, false, |
2071 | var); | |
2072 | if (seq) | |
2073 | { | |
2074 | /* Insert stmt on loop preheader edge. */ | |
2075 | basic_block new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
2076 | gcc_assert (!new_bb); | |
2077 | } | |
2078 | } | |
2079 | ||
ebfd146a | 2080 | /* Update the init conditions of the access functions of all data refs. */ |
b61b1f17 | 2081 | vect_update_inits_of_drs (loop_vinfo, wide_prolog_niters); |
ebfd146a IR |
2082 | |
2083 | /* After peeling we have to reset scalar evolution analyzer. */ | |
2084 | scev_reset (); | |
2085 | ||
2086 | free_original_copy_tables (); | |
2087 | } | |
2088 | ||
2089 | ||
2090 | /* Function vect_create_cond_for_align_checks. | |
2091 | ||
2092 | Create a conditional expression that represents the alignment checks for | |
2093 | all of data references (array element references) whose alignment must be | |
2094 | checked at runtime. | |
2095 | ||
2096 | Input: | |
2097 | COND_EXPR - input conditional expression. New conditions will be chained | |
2098 | with logical AND operation. | |
2099 | LOOP_VINFO - two fields of the loop information are used. | |
2100 | LOOP_VINFO_PTR_MASK is the mask used to check the alignment. | |
2101 | LOOP_VINFO_MAY_MISALIGN_STMTS contains the refs to be checked. | |
2102 | ||
2103 | Output: | |
2104 | COND_EXPR_STMT_LIST - statements needed to construct the conditional | |
2105 | expression. | |
2106 | The returned value is the conditional expression to be used in the if | |
2107 | statement that controls which version of the loop gets executed at runtime. | |
2108 | ||
2109 | The algorithm makes two assumptions: | |
2110 | 1) The number of bytes "n" in a vector is a power of 2. | |
2111 | 2) An address "a" is aligned if a%n is zero and that this | |
2112 | test can be done as a&(n-1) == 0. For example, for 16 | |
2113 | byte vectors the test is a&0xf == 0. */ | |
2114 | ||
2115 | static void | |
2116 | vect_create_cond_for_align_checks (loop_vec_info loop_vinfo, | |
2117 | tree *cond_expr, | |
2118 | gimple_seq *cond_expr_stmt_list) | |
2119 | { | |
2120 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
9771b263 | 2121 | vec<gimple> may_misalign_stmts |
ebfd146a IR |
2122 | = LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo); |
2123 | gimple ref_stmt; | |
2124 | int mask = LOOP_VINFO_PTR_MASK (loop_vinfo); | |
2125 | tree mask_cst; | |
2126 | unsigned int i; | |
ebfd146a IR |
2127 | tree int_ptrsize_type; |
2128 | char tmp_name[20]; | |
2129 | tree or_tmp_name = NULL_TREE; | |
83d5977e | 2130 | tree and_tmp_name; |
ebfd146a IR |
2131 | gimple and_stmt; |
2132 | tree ptrsize_zero; | |
2133 | tree part_cond_expr; | |
2134 | ||
2135 | /* Check that mask is one less than a power of 2, i.e., mask is | |
2136 | all zeros followed by all ones. */ | |
2137 | gcc_assert ((mask != 0) && ((mask & (mask+1)) == 0)); | |
2138 | ||
96f9265a | 2139 | int_ptrsize_type = signed_type_for (ptr_type_node); |
ebfd146a IR |
2140 | |
2141 | /* Create expression (mask & (dr_1 || ... || dr_n)) where dr_i is the address | |
2142 | of the first vector of the i'th data reference. */ | |
2143 | ||
9771b263 | 2144 | FOR_EACH_VEC_ELT (may_misalign_stmts, i, ref_stmt) |
ebfd146a IR |
2145 | { |
2146 | gimple_seq new_stmt_list = NULL; | |
2147 | tree addr_base; | |
83d5977e RG |
2148 | tree addr_tmp_name; |
2149 | tree new_or_tmp_name; | |
ebfd146a | 2150 | gimple addr_stmt, or_stmt; |
d8ba5b19 RG |
2151 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (ref_stmt); |
2152 | tree vectype = STMT_VINFO_VECTYPE (stmt_vinfo); | |
2153 | bool negative = tree_int_cst_compare | |
2154 | (DR_STEP (STMT_VINFO_DATA_REF (stmt_vinfo)), size_zero_node) < 0; | |
2155 | tree offset = negative | |
2156 | ? size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1) : NULL_TREE; | |
ebfd146a IR |
2157 | |
2158 | /* create: addr_tmp = (int)(address_of_first_vector) */ | |
2159 | addr_base = | |
2160 | vect_create_addr_base_for_vector_ref (ref_stmt, &new_stmt_list, | |
d8ba5b19 | 2161 | offset, loop); |
ebfd146a IR |
2162 | if (new_stmt_list != NULL) |
2163 | gimple_seq_add_seq (cond_expr_stmt_list, new_stmt_list); | |
2164 | ||
83d5977e RG |
2165 | sprintf (tmp_name, "addr2int%d", i); |
2166 | addr_tmp_name = make_temp_ssa_name (int_ptrsize_type, NULL, tmp_name); | |
ebfd146a IR |
2167 | addr_stmt = gimple_build_assign_with_ops (NOP_EXPR, addr_tmp_name, |
2168 | addr_base, NULL_TREE); | |
ebfd146a IR |
2169 | gimple_seq_add_stmt (cond_expr_stmt_list, addr_stmt); |
2170 | ||
2171 | /* The addresses are OR together. */ | |
2172 | ||
2173 | if (or_tmp_name != NULL_TREE) | |
2174 | { | |
2175 | /* create: or_tmp = or_tmp | addr_tmp */ | |
83d5977e RG |
2176 | sprintf (tmp_name, "orptrs%d", i); |
2177 | new_or_tmp_name = make_temp_ssa_name (int_ptrsize_type, NULL, tmp_name); | |
ebfd146a IR |
2178 | or_stmt = gimple_build_assign_with_ops (BIT_IOR_EXPR, |
2179 | new_or_tmp_name, | |
2180 | or_tmp_name, addr_tmp_name); | |
ebfd146a IR |
2181 | gimple_seq_add_stmt (cond_expr_stmt_list, or_stmt); |
2182 | or_tmp_name = new_or_tmp_name; | |
2183 | } | |
2184 | else | |
2185 | or_tmp_name = addr_tmp_name; | |
2186 | ||
2187 | } /* end for i */ | |
2188 | ||
2189 | mask_cst = build_int_cst (int_ptrsize_type, mask); | |
2190 | ||
2191 | /* create: and_tmp = or_tmp & mask */ | |
83d5977e | 2192 | and_tmp_name = make_temp_ssa_name (int_ptrsize_type, NULL, "andmask"); |
ebfd146a IR |
2193 | |
2194 | and_stmt = gimple_build_assign_with_ops (BIT_AND_EXPR, and_tmp_name, | |
2195 | or_tmp_name, mask_cst); | |
ebfd146a IR |
2196 | gimple_seq_add_stmt (cond_expr_stmt_list, and_stmt); |
2197 | ||
2198 | /* Make and_tmp the left operand of the conditional test against zero. | |
2199 | if and_tmp has a nonzero bit then some address is unaligned. */ | |
2200 | ptrsize_zero = build_int_cst (int_ptrsize_type, 0); | |
2201 | part_cond_expr = fold_build2 (EQ_EXPR, boolean_type_node, | |
2202 | and_tmp_name, ptrsize_zero); | |
2203 | if (*cond_expr) | |
2204 | *cond_expr = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, | |
2205 | *cond_expr, part_cond_expr); | |
2206 | else | |
2207 | *cond_expr = part_cond_expr; | |
2208 | } | |
2209 | ||
2210 | ||
2211 | /* Function vect_vfa_segment_size. | |
2212 | ||
2213 | Create an expression that computes the size of segment | |
2214 | that will be accessed for a data reference. The functions takes into | |
2215 | account that realignment loads may access one more vector. | |
2216 | ||
2217 | Input: | |
2218 | DR: The data reference. | |
208cb8cb | 2219 | LENGTH_FACTOR: segment length to consider. |
ebfd146a IR |
2220 | |
2221 | Return an expression whose value is the size of segment which will be | |
2222 | accessed by DR. */ | |
2223 | ||
2224 | static tree | |
208cb8cb | 2225 | vect_vfa_segment_size (struct data_reference *dr, tree length_factor) |
ebfd146a | 2226 | { |
e2a3a5f1 | 2227 | tree segment_length; |
338f655d | 2228 | |
319e6439 | 2229 | if (integer_zerop (DR_STEP (dr))) |
338f655d IR |
2230 | segment_length = TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dr))); |
2231 | else | |
2232 | segment_length = size_binop (MULT_EXPR, | |
2233 | fold_convert (sizetype, DR_STEP (dr)), | |
2234 | fold_convert (sizetype, length_factor)); | |
2235 | ||
720f5239 IR |
2236 | if (vect_supportable_dr_alignment (dr, false) |
2237 | == dr_explicit_realign_optimized) | |
ebfd146a IR |
2238 | { |
2239 | tree vector_size = TYPE_SIZE_UNIT | |
2240 | (STMT_VINFO_VECTYPE (vinfo_for_stmt (DR_STMT (dr)))); | |
2241 | ||
e2a3a5f1 | 2242 | segment_length = size_binop (PLUS_EXPR, segment_length, vector_size); |
ebfd146a | 2243 | } |
e2a3a5f1 | 2244 | return segment_length; |
ebfd146a IR |
2245 | } |
2246 | ||
2247 | ||
2248 | /* Function vect_create_cond_for_alias_checks. | |
2249 | ||
2250 | Create a conditional expression that represents the run-time checks for | |
2251 | overlapping of address ranges represented by a list of data references | |
2252 | relations passed as input. | |
2253 | ||
2254 | Input: | |
2255 | COND_EXPR - input conditional expression. New conditions will be chained | |
2256 | with logical AND operation. | |
2257 | LOOP_VINFO - field LOOP_VINFO_MAY_ALIAS_STMTS contains the list of ddrs | |
2258 | to be checked. | |
2259 | ||
2260 | Output: | |
2261 | COND_EXPR - conditional expression. | |
ebfd146a IR |
2262 | |
2263 | The returned value is the conditional expression to be used in the if | |
2264 | statement that controls which version of the loop gets executed at runtime. | |
2265 | */ | |
2266 | ||
2267 | static void | |
4bdd44c4 | 2268 | vect_create_cond_for_alias_checks (loop_vec_info loop_vinfo, tree * cond_expr) |
ebfd146a | 2269 | { |
9771b263 | 2270 | vec<ddr_p> may_alias_ddrs = |
ebfd146a | 2271 | LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo); |
e2a3a5f1 RG |
2272 | int vect_factor = LOOP_VINFO_VECT_FACTOR (loop_vinfo); |
2273 | tree scalar_loop_iters = LOOP_VINFO_NITERS (loop_vinfo); | |
ebfd146a IR |
2274 | |
2275 | ddr_p ddr; | |
2276 | unsigned int i; | |
208cb8cb | 2277 | tree part_cond_expr, length_factor; |
ebfd146a IR |
2278 | |
2279 | /* Create expression | |
36fc3799 RS |
2280 | ((store_ptr_0 + store_segment_length_0) <= load_ptr_0) |
2281 | || (load_ptr_0 + load_segment_length_0) <= store_ptr_0)) | |
b8698a0f | 2282 | && |
ebfd146a IR |
2283 | ... |
2284 | && | |
36fc3799 RS |
2285 | ((store_ptr_n + store_segment_length_n) <= load_ptr_n) |
2286 | || (load_ptr_n + load_segment_length_n) <= store_ptr_n)) */ | |
ebfd146a | 2287 | |
9771b263 | 2288 | if (may_alias_ddrs.is_empty ()) |
ebfd146a IR |
2289 | return; |
2290 | ||
9771b263 | 2291 | FOR_EACH_VEC_ELT (may_alias_ddrs, i, ddr) |
ebfd146a IR |
2292 | { |
2293 | struct data_reference *dr_a, *dr_b; | |
2294 | gimple dr_group_first_a, dr_group_first_b; | |
2295 | tree addr_base_a, addr_base_b; | |
2296 | tree segment_length_a, segment_length_b; | |
2297 | gimple stmt_a, stmt_b; | |
d8ba5b19 | 2298 | tree seg_a_min, seg_a_max, seg_b_min, seg_b_max; |
ebfd146a IR |
2299 | |
2300 | dr_a = DDR_A (ddr); | |
2301 | stmt_a = DR_STMT (DDR_A (ddr)); | |
e14c1050 | 2302 | dr_group_first_a = GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt_a)); |
ebfd146a IR |
2303 | if (dr_group_first_a) |
2304 | { | |
2305 | stmt_a = dr_group_first_a; | |
2306 | dr_a = STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt_a)); | |
2307 | } | |
2308 | ||
2309 | dr_b = DDR_B (ddr); | |
2310 | stmt_b = DR_STMT (DDR_B (ddr)); | |
e14c1050 | 2311 | dr_group_first_b = GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt_b)); |
ebfd146a IR |
2312 | if (dr_group_first_b) |
2313 | { | |
2314 | stmt_b = dr_group_first_b; | |
2315 | dr_b = STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt_b)); | |
2316 | } | |
2317 | ||
4bdd44c4 RB |
2318 | addr_base_a |
2319 | = fold_build_pointer_plus (DR_BASE_ADDRESS (dr_a), | |
2320 | size_binop (PLUS_EXPR, DR_OFFSET (dr_a), | |
2321 | DR_INIT (dr_a))); | |
2322 | addr_base_b | |
2323 | = fold_build_pointer_plus (DR_BASE_ADDRESS (dr_b), | |
2324 | size_binop (PLUS_EXPR, DR_OFFSET (dr_b), | |
2325 | DR_INIT (dr_b))); | |
ebfd146a | 2326 | |
208cb8cb RG |
2327 | if (!operand_equal_p (DR_STEP (dr_a), DR_STEP (dr_b), 0)) |
2328 | length_factor = scalar_loop_iters; | |
2329 | else | |
2330 | length_factor = size_int (vect_factor); | |
2331 | segment_length_a = vect_vfa_segment_size (dr_a, length_factor); | |
2332 | segment_length_b = vect_vfa_segment_size (dr_b, length_factor); | |
ebfd146a | 2333 | |
73fbfcad | 2334 | if (dump_enabled_p ()) |
ebfd146a | 2335 | { |
ccb3ad87 | 2336 | dump_printf_loc (MSG_NOTE, vect_location, |
78c60e3d | 2337 | "create runtime check for data references "); |
ccb3ad87 RB |
2338 | dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (dr_a)); |
2339 | dump_printf (MSG_NOTE, " and "); | |
2340 | dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (dr_b)); | |
ebfd146a IR |
2341 | } |
2342 | ||
d8ba5b19 | 2343 | seg_a_min = addr_base_a; |
5d49b6a7 | 2344 | seg_a_max = fold_build_pointer_plus (addr_base_a, segment_length_a); |
d8ba5b19 RG |
2345 | if (tree_int_cst_compare (DR_STEP (dr_a), size_zero_node) < 0) |
2346 | seg_a_min = seg_a_max, seg_a_max = addr_base_a; | |
2347 | ||
2348 | seg_b_min = addr_base_b; | |
5d49b6a7 | 2349 | seg_b_max = fold_build_pointer_plus (addr_base_b, segment_length_b); |
d8ba5b19 RG |
2350 | if (tree_int_cst_compare (DR_STEP (dr_b), size_zero_node) < 0) |
2351 | seg_b_min = seg_b_max, seg_b_max = addr_base_b; | |
ebfd146a | 2352 | |
b8698a0f | 2353 | part_cond_expr = |
ebfd146a | 2354 | fold_build2 (TRUTH_OR_EXPR, boolean_type_node, |
36fc3799 RS |
2355 | fold_build2 (LE_EXPR, boolean_type_node, seg_a_max, seg_b_min), |
2356 | fold_build2 (LE_EXPR, boolean_type_node, seg_b_max, seg_a_min)); | |
b8698a0f | 2357 | |
ebfd146a IR |
2358 | if (*cond_expr) |
2359 | *cond_expr = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, | |
2360 | *cond_expr, part_cond_expr); | |
2361 | else | |
2362 | *cond_expr = part_cond_expr; | |
2363 | } | |
ebfd146a | 2364 | |
73fbfcad | 2365 | if (dump_enabled_p ()) |
ccb3ad87 | 2366 | dump_printf_loc (MSG_NOTE, vect_location, |
78c60e3d | 2367 | "created %u versioning for alias checks.\n", |
9771b263 | 2368 | may_alias_ddrs.length ()); |
ebfd146a IR |
2369 | } |
2370 | ||
2371 | ||
2372 | /* Function vect_loop_versioning. | |
b8698a0f | 2373 | |
ebfd146a IR |
2374 | If the loop has data references that may or may not be aligned or/and |
2375 | has data reference relations whose independence was not proven then | |
2376 | two versions of the loop need to be generated, one which is vectorized | |
2377 | and one which isn't. A test is then generated to control which of the | |
2378 | loops is executed. The test checks for the alignment of all of the | |
2379 | data references that may or may not be aligned. An additional | |
2380 | sequence of runtime tests is generated for each pairs of DDRs whose | |
b8698a0f L |
2381 | independence was not proven. The vectorized version of loop is |
2382 | executed only if both alias and alignment tests are passed. | |
2383 | ||
ebfd146a | 2384 | The test generated to check which version of loop is executed |
b8698a0f | 2385 | is modified to also check for profitability as indicated by the |
86290011 RG |
2386 | cost model initially. |
2387 | ||
2388 | The versioning precondition(s) are placed in *COND_EXPR and | |
d68d56b5 | 2389 | *COND_EXPR_STMT_LIST. */ |
ebfd146a IR |
2390 | |
2391 | void | |
368117e8 RG |
2392 | vect_loop_versioning (loop_vec_info loop_vinfo, |
2393 | unsigned int th, bool check_profitability) | |
ebfd146a IR |
2394 | { |
2395 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
ebfd146a IR |
2396 | basic_block condition_bb; |
2397 | gimple_stmt_iterator gsi, cond_exp_gsi; | |
2398 | basic_block merge_bb; | |
2399 | basic_block new_exit_bb; | |
2400 | edge new_exit_e, e; | |
2401 | gimple orig_phi, new_phi; | |
368117e8 | 2402 | tree cond_expr = NULL_TREE; |
d68d56b5 | 2403 | gimple_seq cond_expr_stmt_list = NULL; |
ebfd146a IR |
2404 | tree arg; |
2405 | unsigned prob = 4 * REG_BR_PROB_BASE / 5; | |
2406 | gimple_seq gimplify_stmt_list = NULL; | |
2407 | tree scalar_loop_iters = LOOP_VINFO_NITERS (loop_vinfo); | |
9cc1fb4b XDL |
2408 | bool version_align = LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo); |
2409 | bool version_alias = LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo); | |
ebfd146a | 2410 | |
368117e8 RG |
2411 | if (check_profitability) |
2412 | { | |
2413 | cond_expr = fold_build2 (GT_EXPR, boolean_type_node, scalar_loop_iters, | |
2414 | build_int_cst (TREE_TYPE (scalar_loop_iters), th)); | |
2415 | cond_expr = force_gimple_operand_1 (cond_expr, &cond_expr_stmt_list, | |
2416 | is_gimple_condexpr, NULL_TREE); | |
2417 | } | |
ebfd146a | 2418 | |
9cc1fb4b | 2419 | if (version_align) |
d68d56b5 RG |
2420 | vect_create_cond_for_align_checks (loop_vinfo, &cond_expr, |
2421 | &cond_expr_stmt_list); | |
ebfd146a | 2422 | |
9cc1fb4b | 2423 | if (version_alias) |
4bdd44c4 | 2424 | vect_create_cond_for_alias_checks (loop_vinfo, &cond_expr); |
86290011 | 2425 | |
d68d56b5 RG |
2426 | cond_expr = force_gimple_operand_1 (cond_expr, &gimplify_stmt_list, |
2427 | is_gimple_condexpr, NULL_TREE); | |
2428 | gimple_seq_add_seq (&cond_expr_stmt_list, gimplify_stmt_list); | |
ebfd146a IR |
2429 | |
2430 | initialize_original_copy_tables (); | |
d68d56b5 | 2431 | loop_version (loop, cond_expr, &condition_bb, |
0f900dfa | 2432 | prob, prob, REG_BR_PROB_BASE - prob, true); |
9cc1fb4b XDL |
2433 | |
2434 | if (LOCATION_LOCUS (vect_location) != UNKNOWN_LOC | |
2435 | && dump_enabled_p ()) | |
2436 | { | |
2437 | if (version_alias) | |
2438 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, vect_location, | |
2439 | "loop versioned for vectorization because of " | |
2440 | "possible aliasing\n"); | |
2441 | if (version_align) | |
2442 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, vect_location, | |
2443 | "loop versioned for vectorization to enhance " | |
2444 | "alignment\n"); | |
2445 | ||
2446 | } | |
ebfd146a IR |
2447 | free_original_copy_tables(); |
2448 | ||
b8698a0f | 2449 | /* Loop versioning violates an assumption we try to maintain during |
ebfd146a IR |
2450 | vectorization - that the loop exit block has a single predecessor. |
2451 | After versioning, the exit block of both loop versions is the same | |
2452 | basic block (i.e. it has two predecessors). Just in order to simplify | |
2453 | following transformations in the vectorizer, we fix this situation | |
2454 | here by adding a new (empty) block on the exit-edge of the loop, | |
2455 | with the proper loop-exit phis to maintain loop-closed-form. */ | |
b8698a0f | 2456 | |
ebfd146a IR |
2457 | merge_bb = single_exit (loop)->dest; |
2458 | gcc_assert (EDGE_COUNT (merge_bb->preds) == 2); | |
2459 | new_exit_bb = split_edge (single_exit (loop)); | |
2460 | new_exit_e = single_exit (loop); | |
2461 | e = EDGE_SUCC (new_exit_bb, 0); | |
2462 | ||
2463 | for (gsi = gsi_start_phis (merge_bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
2464 | { | |
070ecdfd | 2465 | tree new_res; |
ebfd146a | 2466 | orig_phi = gsi_stmt (gsi); |
070ecdfd RG |
2467 | new_res = copy_ssa_name (PHI_RESULT (orig_phi), NULL); |
2468 | new_phi = create_phi_node (new_res, new_exit_bb); | |
ebfd146a | 2469 | arg = PHI_ARG_DEF_FROM_EDGE (orig_phi, e); |
b8698a0f | 2470 | add_phi_arg (new_phi, arg, new_exit_e, |
9e227d60 | 2471 | gimple_phi_arg_location_from_edge (orig_phi, e)); |
684f25f4 | 2472 | adjust_phi_and_debug_stmts (orig_phi, e, PHI_RESULT (new_phi)); |
b8698a0f | 2473 | } |
ebfd146a IR |
2474 | |
2475 | /* End loop-exit-fixes after versioning. */ | |
2476 | ||
d68d56b5 | 2477 | if (cond_expr_stmt_list) |
ebfd146a IR |
2478 | { |
2479 | cond_exp_gsi = gsi_last_bb (condition_bb); | |
d68d56b5 | 2480 | gsi_insert_seq_before (&cond_exp_gsi, cond_expr_stmt_list, |
86290011 | 2481 | GSI_SAME_STMT); |
ebfd146a | 2482 | } |
90eb75f2 | 2483 | update_ssa (TODO_update_ssa); |
ebfd146a | 2484 | } |