]>
Commit | Line | Data |
---|---|---|
c9e93168 | 1 | /* Tail merging for gimple. |
23a5b65a | 2 | Copyright (C) 2011-2014 Free Software Foundation, Inc. |
c9e93168 TV |
3 | Contributed by Tom de Vries (tom@codesourcery.com) |
4 | ||
5 | This file is part of GCC. | |
6 | ||
7 | GCC is free software; you can redistribute it and/or modify | |
8 | it under the terms of the GNU General Public License as published by | |
9 | the Free Software Foundation; either version 3, or (at your option) | |
10 | any later version. | |
11 | ||
12 | GCC is distributed in the hope that it will be useful, | |
13 | but WITHOUT ANY WARRANTY; without even the implied warranty of | |
14 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
15 | GNU General Public License for more details. | |
16 | ||
17 | You should have received a copy of the GNU General Public License | |
18 | along with GCC; see the file COPYING3. If not see | |
19 | <http://www.gnu.org/licenses/>. */ | |
20 | ||
21 | /* Pass overview. | |
22 | ||
23 | ||
24 | MOTIVATIONAL EXAMPLE | |
25 | ||
26 | gimple representation of gcc/testsuite/gcc.dg/pr43864.c at | |
27 | ||
28 | hprofStartupp (charD.1 * outputFileNameD.2600, charD.1 * ctxD.2601) | |
29 | { | |
30 | struct FILED.1638 * fpD.2605; | |
31 | charD.1 fileNameD.2604[1000]; | |
32 | intD.0 D.3915; | |
33 | const charD.1 * restrict outputFileName.0D.3914; | |
34 | ||
35 | # BLOCK 2 freq:10000 | |
36 | # PRED: ENTRY [100.0%] (fallthru,exec) | |
37 | # PT = nonlocal { D.3926 } (restr) | |
38 | outputFileName.0D.3914_3 | |
39 | = (const charD.1 * restrict) outputFileNameD.2600_2(D); | |
40 | # .MEMD.3923_13 = VDEF <.MEMD.3923_12(D)> | |
41 | # USE = nonlocal null { fileNameD.2604 D.3926 } (restr) | |
42 | # CLB = nonlocal null { fileNameD.2604 D.3926 } (restr) | |
43 | sprintfD.759 (&fileNameD.2604, outputFileName.0D.3914_3); | |
44 | # .MEMD.3923_14 = VDEF <.MEMD.3923_13> | |
45 | # USE = nonlocal null { fileNameD.2604 D.3926 } (restr) | |
46 | # CLB = nonlocal null { fileNameD.2604 D.3926 } (restr) | |
47 | D.3915_4 = accessD.2606 (&fileNameD.2604, 1); | |
48 | if (D.3915_4 == 0) | |
49 | goto <bb 3>; | |
50 | else | |
51 | goto <bb 4>; | |
52 | # SUCC: 3 [10.0%] (true,exec) 4 [90.0%] (false,exec) | |
53 | ||
54 | # BLOCK 3 freq:1000 | |
55 | # PRED: 2 [10.0%] (true,exec) | |
56 | # .MEMD.3923_15 = VDEF <.MEMD.3923_14> | |
57 | # USE = nonlocal null { fileNameD.2604 D.3926 } (restr) | |
58 | # CLB = nonlocal null { fileNameD.2604 D.3926 } (restr) | |
59 | freeD.898 (ctxD.2601_5(D)); | |
60 | goto <bb 7>; | |
61 | # SUCC: 7 [100.0%] (fallthru,exec) | |
62 | ||
63 | # BLOCK 4 freq:9000 | |
64 | # PRED: 2 [90.0%] (false,exec) | |
65 | # .MEMD.3923_16 = VDEF <.MEMD.3923_14> | |
66 | # PT = nonlocal escaped | |
67 | # USE = nonlocal null { fileNameD.2604 D.3926 } (restr) | |
68 | # CLB = nonlocal null { fileNameD.2604 D.3926 } (restr) | |
69 | fpD.2605_8 = fopenD.1805 (&fileNameD.2604[0], 0B); | |
70 | if (fpD.2605_8 == 0B) | |
71 | goto <bb 5>; | |
72 | else | |
73 | goto <bb 6>; | |
74 | # SUCC: 5 [1.9%] (true,exec) 6 [98.1%] (false,exec) | |
75 | ||
76 | # BLOCK 5 freq:173 | |
77 | # PRED: 4 [1.9%] (true,exec) | |
78 | # .MEMD.3923_17 = VDEF <.MEMD.3923_16> | |
79 | # USE = nonlocal null { fileNameD.2604 D.3926 } (restr) | |
80 | # CLB = nonlocal null { fileNameD.2604 D.3926 } (restr) | |
81 | freeD.898 (ctxD.2601_5(D)); | |
82 | goto <bb 7>; | |
83 | # SUCC: 7 [100.0%] (fallthru,exec) | |
84 | ||
85 | # BLOCK 6 freq:8827 | |
86 | # PRED: 4 [98.1%] (false,exec) | |
87 | # .MEMD.3923_18 = VDEF <.MEMD.3923_16> | |
88 | # USE = nonlocal null { fileNameD.2604 D.3926 } (restr) | |
89 | # CLB = nonlocal null { fileNameD.2604 D.3926 } (restr) | |
90 | fooD.2599 (outputFileNameD.2600_2(D), fpD.2605_8); | |
91 | # SUCC: 7 [100.0%] (fallthru,exec) | |
92 | ||
93 | # BLOCK 7 freq:10000 | |
94 | # PRED: 3 [100.0%] (fallthru,exec) 5 [100.0%] (fallthru,exec) | |
95 | 6 [100.0%] (fallthru,exec) | |
96 | # PT = nonlocal null | |
97 | ||
98 | # ctxD.2601_1 = PHI <0B(3), 0B(5), ctxD.2601_5(D)(6)> | |
99 | # .MEMD.3923_11 = PHI <.MEMD.3923_15(3), .MEMD.3923_17(5), | |
100 | .MEMD.3923_18(6)> | |
101 | # VUSE <.MEMD.3923_11> | |
102 | return ctxD.2601_1; | |
103 | # SUCC: EXIT [100.0%] | |
104 | } | |
105 | ||
106 | bb 3 and bb 5 can be merged. The blocks have different predecessors, but the | |
107 | same successors, and the same operations. | |
108 | ||
109 | ||
110 | CONTEXT | |
111 | ||
112 | A technique called tail merging (or cross jumping) can fix the example | |
113 | above. For a block, we look for common code at the end (the tail) of the | |
114 | predecessor blocks, and insert jumps from one block to the other. | |
115 | The example is a special case for tail merging, in that 2 whole blocks | |
116 | can be merged, rather than just the end parts of it. | |
117 | We currently only focus on whole block merging, so in that sense | |
118 | calling this pass tail merge is a bit of a misnomer. | |
119 | ||
120 | We distinguish 2 kinds of situations in which blocks can be merged: | |
121 | - same operations, same predecessors. The successor edges coming from one | |
122 | block are redirected to come from the other block. | |
123 | - same operations, same successors. The predecessor edges entering one block | |
124 | are redirected to enter the other block. Note that this operation might | |
125 | involve introducing phi operations. | |
126 | ||
127 | For efficient implementation, we would like to value numbers the blocks, and | |
128 | have a comparison operator that tells us whether the blocks are equal. | |
129 | Besides being runtime efficient, block value numbering should also abstract | |
130 | from irrelevant differences in order of operations, much like normal value | |
131 | numbering abstracts from irrelevant order of operations. | |
132 | ||
133 | For the first situation (same_operations, same predecessors), normal value | |
134 | numbering fits well. We can calculate a block value number based on the | |
135 | value numbers of the defs and vdefs. | |
136 | ||
137 | For the second situation (same operations, same successors), this approach | |
138 | doesn't work so well. We can illustrate this using the example. The calls | |
139 | to free use different vdefs: MEMD.3923_16 and MEMD.3923_14, and these will | |
140 | remain different in value numbering, since they represent different memory | |
141 | states. So the resulting vdefs of the frees will be different in value | |
142 | numbering, so the block value numbers will be different. | |
143 | ||
144 | The reason why we call the blocks equal is not because they define the same | |
145 | values, but because uses in the blocks use (possibly different) defs in the | |
146 | same way. To be able to detect this efficiently, we need to do some kind of | |
147 | reverse value numbering, meaning number the uses rather than the defs, and | |
148 | calculate a block value number based on the value number of the uses. | |
149 | Ideally, a block comparison operator will also indicate which phis are needed | |
150 | to merge the blocks. | |
151 | ||
152 | For the moment, we don't do block value numbering, but we do insn-by-insn | |
153 | matching, using scc value numbers to match operations with results, and | |
154 | structural comparison otherwise, while ignoring vop mismatches. | |
155 | ||
156 | ||
157 | IMPLEMENTATION | |
158 | ||
159 | 1. The pass first determines all groups of blocks with the same successor | |
160 | blocks. | |
161 | 2. Within each group, it tries to determine clusters of equal basic blocks. | |
162 | 3. The clusters are applied. | |
163 | 4. The same successor groups are updated. | |
164 | 5. This process is repeated from 2 onwards, until no more changes. | |
165 | ||
166 | ||
167 | LIMITATIONS/TODO | |
168 | ||
169 | - block only | |
170 | - handles only 'same operations, same successors'. | |
171 | It handles same predecessors as a special subcase though. | |
172 | - does not implement the reverse value numbering and block value numbering. | |
173 | - improve memory allocation: use garbage collected memory, obstacks, | |
174 | allocpools where appropriate. | |
175 | - no insertion of gimple_reg phis, We only introduce vop-phis. | |
176 | - handle blocks with gimple_reg phi_nodes. | |
177 | ||
178 | ||
8dee4479 TV |
179 | PASS PLACEMENT |
180 | This 'pass' is not a stand-alone gimple pass, but runs as part of | |
181 | pass_pre, in order to share the value numbering. | |
182 | ||
183 | ||
c9e93168 TV |
184 | SWITCHES |
185 | ||
186 | - ftree-tail-merge. On at -O2. We may have to enable it only at -Os. */ | |
187 | ||
188 | #include "config.h" | |
189 | #include "system.h" | |
190 | #include "coretypes.h" | |
191 | #include "tm.h" | |
192 | #include "tree.h" | |
d8a2d370 DN |
193 | #include "stor-layout.h" |
194 | #include "trans-mem.h" | |
6d8eb96b | 195 | #include "inchash.h" |
c9e93168 | 196 | #include "tm_p.h" |
60393bbc AM |
197 | #include "predict.h" |
198 | #include "vec.h" | |
83685514 AM |
199 | #include "hashtab.h" |
200 | #include "hash-set.h" | |
83685514 AM |
201 | #include "machmode.h" |
202 | #include "hard-reg-set.h" | |
203 | #include "input.h" | |
c9e93168 | 204 | #include "function.h" |
60393bbc AM |
205 | #include "dominance.h" |
206 | #include "cfg.h" | |
207 | #include "cfganal.h" | |
208 | #include "cfgcleanup.h" | |
209 | #include "basic-block.h" | |
210 | #include "flags.h" | |
2fb9a547 AM |
211 | #include "hash-table.h" |
212 | #include "tree-ssa-alias.h" | |
213 | #include "internal-fn.h" | |
214 | #include "tree-eh.h" | |
215 | #include "gimple-expr.h" | |
216 | #include "is-a.h" | |
442b4905 | 217 | #include "gimple.h" |
5be5c238 | 218 | #include "gimple-iterator.h" |
442b4905 AM |
219 | #include "gimple-ssa.h" |
220 | #include "tree-cfg.h" | |
221 | #include "tree-phinodes.h" | |
222 | #include "ssa-iterators.h" | |
223 | #include "tree-into-ssa.h" | |
c9e93168 | 224 | #include "params.h" |
c9e93168 TV |
225 | #include "gimple-pretty-print.h" |
226 | #include "tree-ssa-sccvn.h" | |
227 | #include "tree-dump.h" | |
a9e0d843 | 228 | #include "cfgloop.h" |
7ee2468b | 229 | #include "tree-pass.h" |
1fe37220 | 230 | #include "trans-mem.h" |
7ee2468b | 231 | |
c9e93168 TV |
232 | /* Describes a group of bbs with the same successors. The successor bbs are |
233 | cached in succs, and the successor edge flags are cached in succ_flags. | |
234 | If a bb has the EDGE_TRUE/VALSE_VALUE flags swapped compared to succ_flags, | |
235 | it's marked in inverse. | |
236 | Additionally, the hash value for the struct is cached in hashval, and | |
237 | in_worklist indicates whether it's currently part of worklist. */ | |
238 | ||
239 | struct same_succ_def | |
240 | { | |
241 | /* The bbs that have the same successor bbs. */ | |
242 | bitmap bbs; | |
243 | /* The successor bbs. */ | |
244 | bitmap succs; | |
245 | /* Indicates whether the EDGE_TRUE/FALSE_VALUEs of succ_flags are swapped for | |
246 | bb. */ | |
247 | bitmap inverse; | |
248 | /* The edge flags for each of the successor bbs. */ | |
9771b263 | 249 | vec<int> succ_flags; |
c9e93168 TV |
250 | /* Indicates whether the struct is currently in the worklist. */ |
251 | bool in_worklist; | |
252 | /* The hash value of the struct. */ | |
253 | hashval_t hashval; | |
5deac340 RG |
254 | |
255 | /* hash_table support. */ | |
5831a5f0 LC |
256 | typedef same_succ_def value_type; |
257 | typedef same_succ_def compare_type; | |
258 | static inline hashval_t hash (const value_type *); | |
259 | static int equal (const value_type *, const compare_type *); | |
260 | static void remove (value_type *); | |
c9e93168 TV |
261 | }; |
262 | typedef struct same_succ_def *same_succ; | |
263 | typedef const struct same_succ_def *const_same_succ; | |
264 | ||
5deac340 RG |
265 | /* hash routine for hash_table support, returns hashval of E. */ |
266 | ||
267 | inline hashval_t | |
5831a5f0 | 268 | same_succ_def::hash (const value_type *e) |
5deac340 RG |
269 | { |
270 | return e->hashval; | |
271 | } | |
272 | ||
c9e93168 TV |
273 | /* A group of bbs where 1 bb from bbs can replace the other bbs. */ |
274 | ||
275 | struct bb_cluster_def | |
276 | { | |
277 | /* The bbs in the cluster. */ | |
278 | bitmap bbs; | |
279 | /* The preds of the bbs in the cluster. */ | |
280 | bitmap preds; | |
281 | /* Index in all_clusters vector. */ | |
282 | int index; | |
283 | /* The bb to replace the cluster with. */ | |
284 | basic_block rep_bb; | |
285 | }; | |
286 | typedef struct bb_cluster_def *bb_cluster; | |
287 | typedef const struct bb_cluster_def *const_bb_cluster; | |
288 | ||
289 | /* Per bb-info. */ | |
290 | ||
291 | struct aux_bb_info | |
292 | { | |
293 | /* The number of non-debug statements in the bb. */ | |
294 | int size; | |
295 | /* The same_succ that this bb is a member of. */ | |
296 | same_succ bb_same_succ; | |
297 | /* The cluster that this bb is a member of. */ | |
298 | bb_cluster cluster; | |
299 | /* The vop state at the exit of a bb. This is shortlived data, used to | |
300 | communicate data between update_block_by and update_vuses. */ | |
301 | tree vop_at_exit; | |
302 | /* The bb that either contains or is dominated by the dependencies of the | |
303 | bb. */ | |
304 | basic_block dep_bb; | |
305 | }; | |
306 | ||
307 | /* Macros to access the fields of struct aux_bb_info. */ | |
308 | ||
309 | #define BB_SIZE(bb) (((struct aux_bb_info *)bb->aux)->size) | |
310 | #define BB_SAME_SUCC(bb) (((struct aux_bb_info *)bb->aux)->bb_same_succ) | |
311 | #define BB_CLUSTER(bb) (((struct aux_bb_info *)bb->aux)->cluster) | |
312 | #define BB_VOP_AT_EXIT(bb) (((struct aux_bb_info *)bb->aux)->vop_at_exit) | |
313 | #define BB_DEP_BB(bb) (((struct aux_bb_info *)bb->aux)->dep_bb) | |
314 | ||
b2b2f160 TV |
315 | /* Returns true if the only effect a statement STMT has, is to define locally |
316 | used SSA_NAMEs. */ | |
317 | ||
318 | static bool | |
319 | stmt_local_def (gimple stmt) | |
320 | { | |
321 | basic_block bb, def_bb; | |
322 | imm_use_iterator iter; | |
323 | use_operand_p use_p; | |
324 | tree val; | |
325 | def_operand_p def_p; | |
326 | ||
12d3031b RB |
327 | if (gimple_vdef (stmt) != NULL_TREE |
328 | || gimple_has_side_effects (stmt) | |
8024d4b6 TV |
329 | || gimple_could_trap_p_1 (stmt, false, false) |
330 | || gimple_vuse (stmt) != NULL_TREE) | |
b2b2f160 TV |
331 | return false; |
332 | ||
333 | def_p = SINGLE_SSA_DEF_OPERAND (stmt, SSA_OP_DEF); | |
334 | if (def_p == NULL) | |
335 | return false; | |
336 | ||
337 | val = DEF_FROM_PTR (def_p); | |
338 | if (val == NULL_TREE || TREE_CODE (val) != SSA_NAME) | |
339 | return false; | |
340 | ||
341 | def_bb = gimple_bb (stmt); | |
342 | ||
343 | FOR_EACH_IMM_USE_FAST (use_p, iter, val) | |
344 | { | |
345 | if (is_gimple_debug (USE_STMT (use_p))) | |
346 | continue; | |
347 | bb = gimple_bb (USE_STMT (use_p)); | |
348 | if (bb == def_bb) | |
349 | continue; | |
350 | ||
351 | if (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI | |
352 | && EDGE_PRED (bb, PHI_ARG_INDEX_FROM_USE (use_p))->src == def_bb) | |
353 | continue; | |
354 | ||
355 | return false; | |
356 | } | |
357 | ||
358 | return true; | |
359 | } | |
360 | ||
361 | /* Let GSI skip forwards over local defs. */ | |
362 | ||
363 | static void | |
364 | gsi_advance_fw_nondebug_nonlocal (gimple_stmt_iterator *gsi) | |
365 | { | |
366 | gimple stmt; | |
367 | ||
368 | while (true) | |
369 | { | |
370 | if (gsi_end_p (*gsi)) | |
371 | return; | |
372 | stmt = gsi_stmt (*gsi); | |
373 | if (!stmt_local_def (stmt)) | |
374 | return; | |
375 | gsi_next_nondebug (gsi); | |
376 | } | |
377 | } | |
378 | ||
c9e93168 TV |
379 | /* VAL1 and VAL2 are either: |
380 | - uses in BB1 and BB2, or | |
381 | - phi alternatives for BB1 and BB2. | |
382 | Return true if the uses have the same gvn value. */ | |
383 | ||
384 | static bool | |
385 | gvn_uses_equal (tree val1, tree val2) | |
386 | { | |
387 | gcc_checking_assert (val1 != NULL_TREE && val2 != NULL_TREE); | |
388 | ||
389 | if (val1 == val2) | |
390 | return true; | |
391 | ||
392 | if (vn_valueize (val1) != vn_valueize (val2)) | |
393 | return false; | |
394 | ||
395 | return ((TREE_CODE (val1) == SSA_NAME || CONSTANT_CLASS_P (val1)) | |
396 | && (TREE_CODE (val2) == SSA_NAME || CONSTANT_CLASS_P (val2))); | |
397 | } | |
398 | ||
399 | /* Prints E to FILE. */ | |
400 | ||
401 | static void | |
402 | same_succ_print (FILE *file, const same_succ e) | |
403 | { | |
404 | unsigned int i; | |
405 | bitmap_print (file, e->bbs, "bbs:", "\n"); | |
406 | bitmap_print (file, e->succs, "succs:", "\n"); | |
407 | bitmap_print (file, e->inverse, "inverse:", "\n"); | |
408 | fprintf (file, "flags:"); | |
9771b263 DN |
409 | for (i = 0; i < e->succ_flags.length (); ++i) |
410 | fprintf (file, " %x", e->succ_flags[i]); | |
c9e93168 TV |
411 | fprintf (file, "\n"); |
412 | } | |
413 | ||
414 | /* Prints same_succ VE to VFILE. */ | |
415 | ||
0823efed DN |
416 | inline int |
417 | ssa_same_succ_print_traverse (same_succ *pe, FILE *file) | |
c9e93168 | 418 | { |
0823efed | 419 | const same_succ e = *pe; |
c9e93168 TV |
420 | same_succ_print (file, e); |
421 | return 1; | |
422 | } | |
423 | ||
424 | /* Update BB_DEP_BB (USE_BB), given a use of VAL in USE_BB. */ | |
425 | ||
426 | static void | |
427 | update_dep_bb (basic_block use_bb, tree val) | |
428 | { | |
429 | basic_block dep_bb; | |
430 | ||
431 | /* Not a dep. */ | |
432 | if (TREE_CODE (val) != SSA_NAME) | |
433 | return; | |
434 | ||
435 | /* Skip use of global def. */ | |
436 | if (SSA_NAME_IS_DEFAULT_DEF (val)) | |
437 | return; | |
438 | ||
439 | /* Skip use of local def. */ | |
440 | dep_bb = gimple_bb (SSA_NAME_DEF_STMT (val)); | |
441 | if (dep_bb == use_bb) | |
442 | return; | |
443 | ||
444 | if (BB_DEP_BB (use_bb) == NULL | |
445 | || dominated_by_p (CDI_DOMINATORS, dep_bb, BB_DEP_BB (use_bb))) | |
446 | BB_DEP_BB (use_bb) = dep_bb; | |
447 | } | |
448 | ||
449 | /* Update BB_DEP_BB, given the dependencies in STMT. */ | |
450 | ||
451 | static void | |
452 | stmt_update_dep_bb (gimple stmt) | |
453 | { | |
454 | ssa_op_iter iter; | |
455 | use_operand_p use; | |
456 | ||
457 | FOR_EACH_SSA_USE_OPERAND (use, stmt, iter, SSA_OP_USE) | |
458 | update_dep_bb (gimple_bb (stmt), USE_FROM_PTR (use)); | |
459 | } | |
460 | ||
c9e93168 TV |
461 | /* Calculates hash value for same_succ VE. */ |
462 | ||
5deac340 RG |
463 | static hashval_t |
464 | same_succ_hash (const_same_succ e) | |
c9e93168 | 465 | { |
bb7e83fc | 466 | inchash::hash hstate (bitmap_hash (e->succs)); |
c9e93168 TV |
467 | int flags; |
468 | unsigned int i; | |
469 | unsigned int first = bitmap_first_set_bit (e->bbs); | |
06e28de2 | 470 | basic_block bb = BASIC_BLOCK_FOR_FN (cfun, first); |
c9e93168 | 471 | int size = 0; |
c9e93168 TV |
472 | gimple stmt; |
473 | tree arg; | |
474 | unsigned int s; | |
475 | bitmap_iterator bs; | |
476 | ||
538dd0b7 | 477 | for (gimple_stmt_iterator gsi = gsi_start_nondebug_bb (bb); |
c9e93168 TV |
478 | !gsi_end_p (gsi); gsi_next_nondebug (&gsi)) |
479 | { | |
480 | stmt = gsi_stmt (gsi); | |
481 | stmt_update_dep_bb (stmt); | |
b2b2f160 | 482 | if (stmt_local_def (stmt)) |
c9e93168 TV |
483 | continue; |
484 | size++; | |
485 | ||
bb7e83fc | 486 | hstate.add_int (gimple_code (stmt)); |
c9e93168 | 487 | if (is_gimple_assign (stmt)) |
bb7e83fc | 488 | hstate.add_int (gimple_assign_rhs_code (stmt)); |
c9e93168 TV |
489 | if (!is_gimple_call (stmt)) |
490 | continue; | |
491 | if (gimple_call_internal_p (stmt)) | |
bb7e83fc | 492 | hstate.add_int (gimple_call_internal_fn (stmt)); |
c9e93168 | 493 | else |
51d2abeb | 494 | { |
bb7e83fc | 495 | inchash::add_expr (gimple_call_fn (stmt), hstate); |
51d2abeb | 496 | if (gimple_call_chain (stmt)) |
bb7e83fc | 497 | inchash::add_expr (gimple_call_chain (stmt), hstate); |
51d2abeb | 498 | } |
c9e93168 TV |
499 | for (i = 0; i < gimple_call_num_args (stmt); i++) |
500 | { | |
501 | arg = gimple_call_arg (stmt, i); | |
502 | arg = vn_valueize (arg); | |
bb7e83fc | 503 | inchash::add_expr (arg, hstate); |
c9e93168 TV |
504 | } |
505 | } | |
506 | ||
bb7e83fc | 507 | hstate.add_int (size); |
c9e93168 TV |
508 | BB_SIZE (bb) = size; |
509 | ||
9771b263 | 510 | for (i = 0; i < e->succ_flags.length (); ++i) |
c9e93168 | 511 | { |
9771b263 | 512 | flags = e->succ_flags[i]; |
c9e93168 | 513 | flags = flags & ~(EDGE_TRUE_VALUE | EDGE_FALSE_VALUE); |
bb7e83fc | 514 | hstate.add_int (flags); |
c9e93168 TV |
515 | } |
516 | ||
517 | EXECUTE_IF_SET_IN_BITMAP (e->succs, 0, s, bs) | |
518 | { | |
06e28de2 | 519 | int n = find_edge (bb, BASIC_BLOCK_FOR_FN (cfun, s))->dest_idx; |
538dd0b7 DM |
520 | for (gphi_iterator gsi = gsi_start_phis (BASIC_BLOCK_FOR_FN (cfun, s)); |
521 | !gsi_end_p (gsi); | |
c9e93168 TV |
522 | gsi_next (&gsi)) |
523 | { | |
538dd0b7 | 524 | gphi *phi = gsi.phi (); |
c9e93168 TV |
525 | tree lhs = gimple_phi_result (phi); |
526 | tree val = gimple_phi_arg_def (phi, n); | |
527 | ||
ea057359 | 528 | if (virtual_operand_p (lhs)) |
c9e93168 TV |
529 | continue; |
530 | update_dep_bb (bb, val); | |
531 | } | |
532 | } | |
533 | ||
bb7e83fc | 534 | return hstate.end (); |
c9e93168 TV |
535 | } |
536 | ||
537 | /* Returns true if E1 and E2 have 2 successors, and if the successor flags | |
538 | are inverse for the EDGE_TRUE_VALUE and EDGE_FALSE_VALUE flags, and equal for | |
539 | the other edge flags. */ | |
540 | ||
541 | static bool | |
542 | inverse_flags (const_same_succ e1, const_same_succ e2) | |
543 | { | |
544 | int f1a, f1b, f2a, f2b; | |
545 | int mask = ~(EDGE_TRUE_VALUE | EDGE_FALSE_VALUE); | |
546 | ||
9771b263 | 547 | if (e1->succ_flags.length () != 2) |
c9e93168 TV |
548 | return false; |
549 | ||
9771b263 DN |
550 | f1a = e1->succ_flags[0]; |
551 | f1b = e1->succ_flags[1]; | |
552 | f2a = e2->succ_flags[0]; | |
553 | f2b = e2->succ_flags[1]; | |
c9e93168 TV |
554 | |
555 | if (f1a == f2a && f1b == f2b) | |
556 | return false; | |
557 | ||
558 | return (f1a & mask) == (f2a & mask) && (f1b & mask) == (f2b & mask); | |
559 | } | |
560 | ||
5deac340 | 561 | /* Compares SAME_SUCCs E1 and E2. */ |
c9e93168 | 562 | |
0823efed | 563 | int |
5831a5f0 | 564 | same_succ_def::equal (const value_type *e1, const compare_type *e2) |
c9e93168 | 565 | { |
c9e93168 TV |
566 | unsigned int i, first1, first2; |
567 | gimple_stmt_iterator gsi1, gsi2; | |
568 | gimple s1, s2; | |
569 | basic_block bb1, bb2; | |
570 | ||
571 | if (e1->hashval != e2->hashval) | |
572 | return 0; | |
573 | ||
9771b263 | 574 | if (e1->succ_flags.length () != e2->succ_flags.length ()) |
c9e93168 TV |
575 | return 0; |
576 | ||
577 | if (!bitmap_equal_p (e1->succs, e2->succs)) | |
578 | return 0; | |
579 | ||
580 | if (!inverse_flags (e1, e2)) | |
581 | { | |
9771b263 DN |
582 | for (i = 0; i < e1->succ_flags.length (); ++i) |
583 | if (e1->succ_flags[i] != e1->succ_flags[i]) | |
c9e93168 TV |
584 | return 0; |
585 | } | |
586 | ||
587 | first1 = bitmap_first_set_bit (e1->bbs); | |
588 | first2 = bitmap_first_set_bit (e2->bbs); | |
589 | ||
06e28de2 DM |
590 | bb1 = BASIC_BLOCK_FOR_FN (cfun, first1); |
591 | bb2 = BASIC_BLOCK_FOR_FN (cfun, first2); | |
c9e93168 TV |
592 | |
593 | if (BB_SIZE (bb1) != BB_SIZE (bb2)) | |
594 | return 0; | |
595 | ||
596 | gsi1 = gsi_start_nondebug_bb (bb1); | |
597 | gsi2 = gsi_start_nondebug_bb (bb2); | |
b2b2f160 TV |
598 | gsi_advance_fw_nondebug_nonlocal (&gsi1); |
599 | gsi_advance_fw_nondebug_nonlocal (&gsi2); | |
c9e93168 TV |
600 | while (!(gsi_end_p (gsi1) || gsi_end_p (gsi2))) |
601 | { | |
602 | s1 = gsi_stmt (gsi1); | |
603 | s2 = gsi_stmt (gsi2); | |
604 | if (gimple_code (s1) != gimple_code (s2)) | |
605 | return 0; | |
606 | if (is_gimple_call (s1) && !gimple_call_same_target_p (s1, s2)) | |
607 | return 0; | |
608 | gsi_next_nondebug (&gsi1); | |
609 | gsi_next_nondebug (&gsi2); | |
b2b2f160 TV |
610 | gsi_advance_fw_nondebug_nonlocal (&gsi1); |
611 | gsi_advance_fw_nondebug_nonlocal (&gsi2); | |
c9e93168 TV |
612 | } |
613 | ||
614 | return 1; | |
615 | } | |
616 | ||
617 | /* Alloc and init a new SAME_SUCC. */ | |
618 | ||
619 | static same_succ | |
620 | same_succ_alloc (void) | |
621 | { | |
622 | same_succ same = XNEW (struct same_succ_def); | |
623 | ||
624 | same->bbs = BITMAP_ALLOC (NULL); | |
625 | same->succs = BITMAP_ALLOC (NULL); | |
626 | same->inverse = BITMAP_ALLOC (NULL); | |
9771b263 | 627 | same->succ_flags.create (10); |
c9e93168 TV |
628 | same->in_worklist = false; |
629 | ||
630 | return same; | |
631 | } | |
632 | ||
5deac340 | 633 | /* Delete same_succ E. */ |
c9e93168 | 634 | |
5deac340 RG |
635 | void |
636 | same_succ_def::remove (same_succ e) | |
c9e93168 | 637 | { |
c9e93168 TV |
638 | BITMAP_FREE (e->bbs); |
639 | BITMAP_FREE (e->succs); | |
640 | BITMAP_FREE (e->inverse); | |
9771b263 | 641 | e->succ_flags.release (); |
c9e93168 | 642 | |
0823efed | 643 | XDELETE (e); |
c9e93168 TV |
644 | } |
645 | ||
646 | /* Reset same_succ SAME. */ | |
647 | ||
648 | static void | |
649 | same_succ_reset (same_succ same) | |
650 | { | |
651 | bitmap_clear (same->bbs); | |
652 | bitmap_clear (same->succs); | |
653 | bitmap_clear (same->inverse); | |
9771b263 | 654 | same->succ_flags.truncate (0); |
c9e93168 TV |
655 | } |
656 | ||
c203e8a7 | 657 | static hash_table<same_succ_def> *same_succ_htab; |
c9e93168 TV |
658 | |
659 | /* Array that is used to store the edge flags for a successor. */ | |
660 | ||
661 | static int *same_succ_edge_flags; | |
662 | ||
663 | /* Bitmap that is used to mark bbs that are recently deleted. */ | |
664 | ||
665 | static bitmap deleted_bbs; | |
666 | ||
667 | /* Bitmap that is used to mark predecessors of bbs that are | |
668 | deleted. */ | |
669 | ||
670 | static bitmap deleted_bb_preds; | |
671 | ||
672 | /* Prints same_succ_htab to stderr. */ | |
673 | ||
674 | extern void debug_same_succ (void); | |
675 | DEBUG_FUNCTION void | |
676 | debug_same_succ ( void) | |
677 | { | |
c203e8a7 | 678 | same_succ_htab->traverse <FILE *, ssa_same_succ_print_traverse> (stderr); |
c9e93168 TV |
679 | } |
680 | ||
c9e93168 TV |
681 | |
682 | /* Vector of bbs to process. */ | |
683 | ||
9771b263 | 684 | static vec<same_succ> worklist; |
c9e93168 TV |
685 | |
686 | /* Prints worklist to FILE. */ | |
687 | ||
688 | static void | |
689 | print_worklist (FILE *file) | |
690 | { | |
691 | unsigned int i; | |
9771b263 DN |
692 | for (i = 0; i < worklist.length (); ++i) |
693 | same_succ_print (file, worklist[i]); | |
c9e93168 TV |
694 | } |
695 | ||
696 | /* Adds SAME to worklist. */ | |
697 | ||
698 | static void | |
699 | add_to_worklist (same_succ same) | |
700 | { | |
701 | if (same->in_worklist) | |
702 | return; | |
703 | ||
704 | if (bitmap_count_bits (same->bbs) < 2) | |
705 | return; | |
706 | ||
707 | same->in_worklist = true; | |
9771b263 | 708 | worklist.safe_push (same); |
c9e93168 TV |
709 | } |
710 | ||
711 | /* Add BB to same_succ_htab. */ | |
712 | ||
713 | static void | |
714 | find_same_succ_bb (basic_block bb, same_succ *same_p) | |
715 | { | |
716 | unsigned int j; | |
717 | bitmap_iterator bj; | |
718 | same_succ same = *same_p; | |
719 | same_succ *slot; | |
720 | edge_iterator ei; | |
721 | edge e; | |
722 | ||
315bbd2e TV |
723 | if (bb == NULL |
724 | /* Be conservative with loop structure. It's not evident that this test | |
725 | is sufficient. Before tail-merge, we've just called | |
726 | loop_optimizer_finalize, and LOOPS_MAY_HAVE_MULTIPLE_LATCHES is now | |
727 | set, so there's no guarantee that the loop->latch value is still valid. | |
728 | But we assume that, since we've forced LOOPS_HAVE_SIMPLE_LATCHES at the | |
729 | start of pre, we've kept that property intact throughout pre, and are | |
730 | keeping it throughout tail-merge using this test. */ | |
731 | || bb->loop_father->latch == bb) | |
c9e93168 TV |
732 | return; |
733 | bitmap_set_bit (same->bbs, bb->index); | |
734 | FOR_EACH_EDGE (e, ei, bb->succs) | |
735 | { | |
736 | int index = e->dest->index; | |
737 | bitmap_set_bit (same->succs, index); | |
738 | same_succ_edge_flags[index] = e->flags; | |
739 | } | |
740 | EXECUTE_IF_SET_IN_BITMAP (same->succs, 0, j, bj) | |
9771b263 | 741 | same->succ_flags.safe_push (same_succ_edge_flags[j]); |
c9e93168 | 742 | |
5deac340 | 743 | same->hashval = same_succ_hash (same); |
c9e93168 | 744 | |
c203e8a7 | 745 | slot = same_succ_htab->find_slot_with_hash (same, same->hashval, INSERT); |
c9e93168 TV |
746 | if (*slot == NULL) |
747 | { | |
748 | *slot = same; | |
749 | BB_SAME_SUCC (bb) = same; | |
750 | add_to_worklist (same); | |
751 | *same_p = NULL; | |
752 | } | |
753 | else | |
754 | { | |
755 | bitmap_set_bit ((*slot)->bbs, bb->index); | |
756 | BB_SAME_SUCC (bb) = *slot; | |
757 | add_to_worklist (*slot); | |
758 | if (inverse_flags (same, *slot)) | |
759 | bitmap_set_bit ((*slot)->inverse, bb->index); | |
760 | same_succ_reset (same); | |
761 | } | |
762 | } | |
763 | ||
764 | /* Find bbs with same successors. */ | |
765 | ||
766 | static void | |
767 | find_same_succ (void) | |
768 | { | |
769 | same_succ same = same_succ_alloc (); | |
770 | basic_block bb; | |
771 | ||
11cd3bed | 772 | FOR_EACH_BB_FN (bb, cfun) |
c9e93168 TV |
773 | { |
774 | find_same_succ_bb (bb, &same); | |
775 | if (same == NULL) | |
776 | same = same_succ_alloc (); | |
777 | } | |
778 | ||
5deac340 | 779 | same_succ_def::remove (same); |
c9e93168 TV |
780 | } |
781 | ||
782 | /* Initializes worklist administration. */ | |
783 | ||
784 | static void | |
785 | init_worklist (void) | |
786 | { | |
787 | alloc_aux_for_blocks (sizeof (struct aux_bb_info)); | |
c203e8a7 | 788 | same_succ_htab = new hash_table<same_succ_def> (n_basic_blocks_for_fn (cfun)); |
8b1c6fd7 | 789 | same_succ_edge_flags = XCNEWVEC (int, last_basic_block_for_fn (cfun)); |
c9e93168 TV |
790 | deleted_bbs = BITMAP_ALLOC (NULL); |
791 | deleted_bb_preds = BITMAP_ALLOC (NULL); | |
0cae8d31 | 792 | worklist.create (n_basic_blocks_for_fn (cfun)); |
c9e93168 TV |
793 | find_same_succ (); |
794 | ||
795 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
796 | { | |
797 | fprintf (dump_file, "initial worklist:\n"); | |
798 | print_worklist (dump_file); | |
799 | } | |
800 | } | |
801 | ||
802 | /* Deletes worklist administration. */ | |
803 | ||
804 | static void | |
805 | delete_worklist (void) | |
806 | { | |
807 | free_aux_for_blocks (); | |
c203e8a7 TS |
808 | delete same_succ_htab; |
809 | same_succ_htab = NULL; | |
c9e93168 TV |
810 | XDELETEVEC (same_succ_edge_flags); |
811 | same_succ_edge_flags = NULL; | |
812 | BITMAP_FREE (deleted_bbs); | |
813 | BITMAP_FREE (deleted_bb_preds); | |
9771b263 | 814 | worklist.release (); |
c9e93168 TV |
815 | } |
816 | ||
817 | /* Mark BB as deleted, and mark its predecessors. */ | |
818 | ||
819 | static void | |
643400b8 | 820 | mark_basic_block_deleted (basic_block bb) |
c9e93168 TV |
821 | { |
822 | edge e; | |
823 | edge_iterator ei; | |
824 | ||
825 | bitmap_set_bit (deleted_bbs, bb->index); | |
826 | ||
827 | FOR_EACH_EDGE (e, ei, bb->preds) | |
828 | bitmap_set_bit (deleted_bb_preds, e->src->index); | |
829 | } | |
830 | ||
4cbdcd40 TV |
831 | /* Removes BB from its corresponding same_succ. */ |
832 | ||
833 | static void | |
834 | same_succ_flush_bb (basic_block bb) | |
835 | { | |
836 | same_succ same = BB_SAME_SUCC (bb); | |
837 | BB_SAME_SUCC (bb) = NULL; | |
838 | if (bitmap_single_bit_set_p (same->bbs)) | |
c203e8a7 | 839 | same_succ_htab->remove_elt_with_hash (same, same->hashval); |
4cbdcd40 TV |
840 | else |
841 | bitmap_clear_bit (same->bbs, bb->index); | |
842 | } | |
843 | ||
c9e93168 TV |
844 | /* Removes all bbs in BBS from their corresponding same_succ. */ |
845 | ||
846 | static void | |
847 | same_succ_flush_bbs (bitmap bbs) | |
848 | { | |
849 | unsigned int i; | |
850 | bitmap_iterator bi; | |
851 | ||
852 | EXECUTE_IF_SET_IN_BITMAP (bbs, 0, i, bi) | |
06e28de2 | 853 | same_succ_flush_bb (BASIC_BLOCK_FOR_FN (cfun, i)); |
c9e93168 TV |
854 | } |
855 | ||
fcddd80e RG |
856 | /* Release the last vdef in BB, either normal or phi result. */ |
857 | ||
858 | static void | |
859 | release_last_vdef (basic_block bb) | |
860 | { | |
538dd0b7 DM |
861 | for (gimple_stmt_iterator i = gsi_last_bb (bb); !gsi_end_p (i); |
862 | gsi_prev_nondebug (&i)) | |
fcddd80e RG |
863 | { |
864 | gimple stmt = gsi_stmt (i); | |
865 | if (gimple_vdef (stmt) == NULL_TREE) | |
866 | continue; | |
867 | ||
868 | mark_virtual_operand_for_renaming (gimple_vdef (stmt)); | |
869 | return; | |
870 | } | |
871 | ||
538dd0b7 DM |
872 | for (gphi_iterator i = gsi_start_phis (bb); !gsi_end_p (i); |
873 | gsi_next (&i)) | |
fcddd80e | 874 | { |
538dd0b7 | 875 | gphi *phi = i.phi (); |
fcddd80e RG |
876 | tree res = gimple_phi_result (phi); |
877 | ||
ea057359 | 878 | if (!virtual_operand_p (res)) |
fcddd80e RG |
879 | continue; |
880 | ||
881 | mark_virtual_phi_result_for_renaming (phi); | |
882 | return; | |
883 | } | |
884 | ||
885 | } | |
886 | ||
c9e93168 TV |
887 | /* For deleted_bb_preds, find bbs with same successors. */ |
888 | ||
889 | static void | |
890 | update_worklist (void) | |
891 | { | |
892 | unsigned int i; | |
893 | bitmap_iterator bi; | |
894 | basic_block bb; | |
895 | same_succ same; | |
896 | ||
643400b8 TV |
897 | bitmap_and_compl_into (deleted_bb_preds, deleted_bbs); |
898 | bitmap_clear (deleted_bbs); | |
899 | ||
c9e93168 TV |
900 | bitmap_clear_bit (deleted_bb_preds, ENTRY_BLOCK); |
901 | same_succ_flush_bbs (deleted_bb_preds); | |
902 | ||
903 | same = same_succ_alloc (); | |
904 | EXECUTE_IF_SET_IN_BITMAP (deleted_bb_preds, 0, i, bi) | |
905 | { | |
06e28de2 | 906 | bb = BASIC_BLOCK_FOR_FN (cfun, i); |
c9e93168 TV |
907 | gcc_assert (bb != NULL); |
908 | find_same_succ_bb (bb, &same); | |
909 | if (same == NULL) | |
910 | same = same_succ_alloc (); | |
911 | } | |
5deac340 | 912 | same_succ_def::remove (same); |
c9e93168 TV |
913 | bitmap_clear (deleted_bb_preds); |
914 | } | |
915 | ||
916 | /* Prints cluster C to FILE. */ | |
917 | ||
918 | static void | |
919 | print_cluster (FILE *file, bb_cluster c) | |
920 | { | |
921 | if (c == NULL) | |
922 | return; | |
923 | bitmap_print (file, c->bbs, "bbs:", "\n"); | |
924 | bitmap_print (file, c->preds, "preds:", "\n"); | |
925 | } | |
926 | ||
927 | /* Prints cluster C to stderr. */ | |
928 | ||
929 | extern void debug_cluster (bb_cluster); | |
930 | DEBUG_FUNCTION void | |
931 | debug_cluster (bb_cluster c) | |
932 | { | |
933 | print_cluster (stderr, c); | |
934 | } | |
935 | ||
936 | /* Update C->rep_bb, given that BB is added to the cluster. */ | |
937 | ||
938 | static void | |
939 | update_rep_bb (bb_cluster c, basic_block bb) | |
940 | { | |
941 | /* Initial. */ | |
942 | if (c->rep_bb == NULL) | |
943 | { | |
944 | c->rep_bb = bb; | |
945 | return; | |
946 | } | |
947 | ||
948 | /* Current needs no deps, keep it. */ | |
949 | if (BB_DEP_BB (c->rep_bb) == NULL) | |
950 | return; | |
951 | ||
952 | /* Bb needs no deps, change rep_bb. */ | |
953 | if (BB_DEP_BB (bb) == NULL) | |
954 | { | |
955 | c->rep_bb = bb; | |
956 | return; | |
957 | } | |
958 | ||
959 | /* Bb needs last deps earlier than current, change rep_bb. A potential | |
960 | problem with this, is that the first deps might also be earlier, which | |
961 | would mean we prefer longer lifetimes for the deps. To be able to check | |
962 | for this, we would have to trace BB_FIRST_DEP_BB as well, besides | |
963 | BB_DEP_BB, which is really BB_LAST_DEP_BB. | |
964 | The benefit of choosing the bb with last deps earlier, is that it can | |
965 | potentially be used as replacement for more bbs. */ | |
966 | if (dominated_by_p (CDI_DOMINATORS, BB_DEP_BB (c->rep_bb), BB_DEP_BB (bb))) | |
967 | c->rep_bb = bb; | |
968 | } | |
969 | ||
970 | /* Add BB to cluster C. Sets BB in C->bbs, and preds of BB in C->preds. */ | |
971 | ||
972 | static void | |
973 | add_bb_to_cluster (bb_cluster c, basic_block bb) | |
974 | { | |
975 | edge e; | |
976 | edge_iterator ei; | |
977 | ||
978 | bitmap_set_bit (c->bbs, bb->index); | |
979 | ||
980 | FOR_EACH_EDGE (e, ei, bb->preds) | |
981 | bitmap_set_bit (c->preds, e->src->index); | |
982 | ||
983 | update_rep_bb (c, bb); | |
984 | } | |
985 | ||
986 | /* Allocate and init new cluster. */ | |
987 | ||
988 | static bb_cluster | |
989 | new_cluster (void) | |
990 | { | |
991 | bb_cluster c; | |
992 | c = XCNEW (struct bb_cluster_def); | |
993 | c->bbs = BITMAP_ALLOC (NULL); | |
994 | c->preds = BITMAP_ALLOC (NULL); | |
995 | c->rep_bb = NULL; | |
996 | return c; | |
997 | } | |
998 | ||
999 | /* Delete clusters. */ | |
1000 | ||
1001 | static void | |
1002 | delete_cluster (bb_cluster c) | |
1003 | { | |
1004 | if (c == NULL) | |
1005 | return; | |
1006 | BITMAP_FREE (c->bbs); | |
1007 | BITMAP_FREE (c->preds); | |
1008 | XDELETE (c); | |
1009 | } | |
1010 | ||
c9e93168 TV |
1011 | |
1012 | /* Array that contains all clusters. */ | |
1013 | ||
9771b263 | 1014 | static vec<bb_cluster> all_clusters; |
c9e93168 TV |
1015 | |
1016 | /* Allocate all cluster vectors. */ | |
1017 | ||
1018 | static void | |
1019 | alloc_cluster_vectors (void) | |
1020 | { | |
0cae8d31 | 1021 | all_clusters.create (n_basic_blocks_for_fn (cfun)); |
c9e93168 TV |
1022 | } |
1023 | ||
1024 | /* Reset all cluster vectors. */ | |
1025 | ||
1026 | static void | |
1027 | reset_cluster_vectors (void) | |
1028 | { | |
1029 | unsigned int i; | |
1030 | basic_block bb; | |
9771b263 DN |
1031 | for (i = 0; i < all_clusters.length (); ++i) |
1032 | delete_cluster (all_clusters[i]); | |
1033 | all_clusters.truncate (0); | |
11cd3bed | 1034 | FOR_EACH_BB_FN (bb, cfun) |
c9e93168 TV |
1035 | BB_CLUSTER (bb) = NULL; |
1036 | } | |
1037 | ||
1038 | /* Delete all cluster vectors. */ | |
1039 | ||
1040 | static void | |
1041 | delete_cluster_vectors (void) | |
1042 | { | |
1043 | unsigned int i; | |
9771b263 DN |
1044 | for (i = 0; i < all_clusters.length (); ++i) |
1045 | delete_cluster (all_clusters[i]); | |
1046 | all_clusters.release (); | |
c9e93168 TV |
1047 | } |
1048 | ||
1049 | /* Merge cluster C2 into C1. */ | |
1050 | ||
1051 | static void | |
1052 | merge_clusters (bb_cluster c1, bb_cluster c2) | |
1053 | { | |
1054 | bitmap_ior_into (c1->bbs, c2->bbs); | |
1055 | bitmap_ior_into (c1->preds, c2->preds); | |
1056 | } | |
1057 | ||
1058 | /* Register equivalence of BB1 and BB2 (members of cluster C). Store c in | |
1059 | all_clusters, or merge c with existing cluster. */ | |
1060 | ||
1061 | static void | |
1062 | set_cluster (basic_block bb1, basic_block bb2) | |
1063 | { | |
1064 | basic_block merge_bb, other_bb; | |
1065 | bb_cluster merge, old, c; | |
1066 | ||
1067 | if (BB_CLUSTER (bb1) == NULL && BB_CLUSTER (bb2) == NULL) | |
1068 | { | |
1069 | c = new_cluster (); | |
1070 | add_bb_to_cluster (c, bb1); | |
1071 | add_bb_to_cluster (c, bb2); | |
1072 | BB_CLUSTER (bb1) = c; | |
1073 | BB_CLUSTER (bb2) = c; | |
9771b263 DN |
1074 | c->index = all_clusters.length (); |
1075 | all_clusters.safe_push (c); | |
c9e93168 TV |
1076 | } |
1077 | else if (BB_CLUSTER (bb1) == NULL || BB_CLUSTER (bb2) == NULL) | |
1078 | { | |
1079 | merge_bb = BB_CLUSTER (bb1) == NULL ? bb2 : bb1; | |
1080 | other_bb = BB_CLUSTER (bb1) == NULL ? bb1 : bb2; | |
1081 | merge = BB_CLUSTER (merge_bb); | |
1082 | add_bb_to_cluster (merge, other_bb); | |
1083 | BB_CLUSTER (other_bb) = merge; | |
1084 | } | |
1085 | else if (BB_CLUSTER (bb1) != BB_CLUSTER (bb2)) | |
1086 | { | |
1087 | unsigned int i; | |
1088 | bitmap_iterator bi; | |
1089 | ||
1090 | old = BB_CLUSTER (bb2); | |
1091 | merge = BB_CLUSTER (bb1); | |
1092 | merge_clusters (merge, old); | |
1093 | EXECUTE_IF_SET_IN_BITMAP (old->bbs, 0, i, bi) | |
06e28de2 | 1094 | BB_CLUSTER (BASIC_BLOCK_FOR_FN (cfun, i)) = merge; |
9771b263 | 1095 | all_clusters[old->index] = NULL; |
c9e93168 TV |
1096 | update_rep_bb (merge, old->rep_bb); |
1097 | delete_cluster (old); | |
1098 | } | |
1099 | else | |
1100 | gcc_unreachable (); | |
1101 | } | |
1102 | ||
b6366520 TV |
1103 | /* Return true if gimple operands T1 and T2 have the same value. */ |
1104 | ||
1105 | static bool | |
1106 | gimple_operand_equal_value_p (tree t1, tree t2) | |
1107 | { | |
1108 | if (t1 == t2) | |
1109 | return true; | |
1110 | ||
1111 | if (t1 == NULL_TREE | |
1112 | || t2 == NULL_TREE) | |
1113 | return false; | |
1114 | ||
1115 | if (operand_equal_p (t1, t2, 0)) | |
1116 | return true; | |
1117 | ||
1118 | return gvn_uses_equal (t1, t2); | |
1119 | } | |
1120 | ||
c9e93168 TV |
1121 | /* Return true if gimple statements S1 and S2 are equal. Gimple_bb (s1) and |
1122 | gimple_bb (s2) are members of SAME_SUCC. */ | |
1123 | ||
1124 | static bool | |
1125 | gimple_equal_p (same_succ same_succ, gimple s1, gimple s2) | |
1126 | { | |
1127 | unsigned int i; | |
1128 | tree lhs1, lhs2; | |
1129 | basic_block bb1 = gimple_bb (s1), bb2 = gimple_bb (s2); | |
1130 | tree t1, t2; | |
2f1aee04 | 1131 | bool inv_cond; |
c9e93168 TV |
1132 | enum tree_code code1, code2; |
1133 | ||
1134 | if (gimple_code (s1) != gimple_code (s2)) | |
1135 | return false; | |
1136 | ||
1137 | switch (gimple_code (s1)) | |
1138 | { | |
1139 | case GIMPLE_CALL: | |
c9e93168 TV |
1140 | if (!gimple_call_same_target_p (s1, s2)) |
1141 | return false; | |
1142 | ||
51d2abeb EB |
1143 | t1 = gimple_call_chain (s1); |
1144 | t2 = gimple_call_chain (s2); | |
1145 | if (!gimple_operand_equal_value_p (t1, t2)) | |
1146 | return false; | |
1147 | ||
1148 | if (gimple_call_num_args (s1) != gimple_call_num_args (s2)) | |
1149 | return false; | |
1150 | ||
c9e93168 TV |
1151 | for (i = 0; i < gimple_call_num_args (s1); ++i) |
1152 | { | |
1153 | t1 = gimple_call_arg (s1, i); | |
1154 | t2 = gimple_call_arg (s2, i); | |
51d2abeb EB |
1155 | if (!gimple_operand_equal_value_p (t1, t2)) |
1156 | return false; | |
c9e93168 | 1157 | } |
c9e93168 TV |
1158 | |
1159 | lhs1 = gimple_get_lhs (s1); | |
1160 | lhs2 = gimple_get_lhs (s2); | |
e6fa9204 JJ |
1161 | if (lhs1 == NULL_TREE && lhs2 == NULL_TREE) |
1162 | return true; | |
1163 | if (lhs1 == NULL_TREE || lhs2 == NULL_TREE) | |
1164 | return false; | |
1165 | if (TREE_CODE (lhs1) == SSA_NAME && TREE_CODE (lhs2) == SSA_NAME) | |
1166 | return vn_valueize (lhs1) == vn_valueize (lhs2); | |
1167 | return operand_equal_p (lhs1, lhs2, 0); | |
c9e93168 TV |
1168 | |
1169 | case GIMPLE_ASSIGN: | |
1170 | lhs1 = gimple_get_lhs (s1); | |
1171 | lhs2 = gimple_get_lhs (s2); | |
fcfa87ac RB |
1172 | if (TREE_CODE (lhs1) != SSA_NAME |
1173 | && TREE_CODE (lhs2) != SSA_NAME) | |
030e321a RB |
1174 | return (operand_equal_p (lhs1, lhs2, 0) |
1175 | && gimple_operand_equal_value_p (gimple_assign_rhs1 (s1), | |
1176 | gimple_assign_rhs1 (s2))); | |
fcfa87ac RB |
1177 | else if (TREE_CODE (lhs1) == SSA_NAME |
1178 | && TREE_CODE (lhs2) == SSA_NAME) | |
8024d4b6 TV |
1179 | return operand_equal_p (gimple_assign_rhs1 (s1), |
1180 | gimple_assign_rhs1 (s2), 0); | |
fcfa87ac | 1181 | return false; |
c9e93168 TV |
1182 | |
1183 | case GIMPLE_COND: | |
1184 | t1 = gimple_cond_lhs (s1); | |
1185 | t2 = gimple_cond_lhs (s2); | |
b6366520 | 1186 | if (!gimple_operand_equal_value_p (t1, t2)) |
c9e93168 TV |
1187 | return false; |
1188 | ||
1189 | t1 = gimple_cond_rhs (s1); | |
1190 | t2 = gimple_cond_rhs (s2); | |
b6366520 | 1191 | if (!gimple_operand_equal_value_p (t1, t2)) |
c9e93168 TV |
1192 | return false; |
1193 | ||
1194 | code1 = gimple_expr_code (s1); | |
1195 | code2 = gimple_expr_code (s2); | |
1196 | inv_cond = (bitmap_bit_p (same_succ->inverse, bb1->index) | |
1197 | != bitmap_bit_p (same_succ->inverse, bb2->index)); | |
1198 | if (inv_cond) | |
1199 | { | |
1200 | bool honor_nans | |
1201 | = HONOR_NANS (TYPE_MODE (TREE_TYPE (gimple_cond_lhs (s1)))); | |
1202 | code2 = invert_tree_comparison (code2, honor_nans); | |
1203 | } | |
1204 | return code1 == code2; | |
1205 | ||
1206 | default: | |
1207 | return false; | |
1208 | } | |
1209 | } | |
1210 | ||
46301137 TV |
1211 | /* Let GSI skip backwards over local defs. Return the earliest vuse in VUSE. |
1212 | Return true in VUSE_ESCAPED if the vuse influenced a SSA_OP_DEF of one of the | |
1213 | processed statements. */ | |
c9e93168 TV |
1214 | |
1215 | static void | |
46301137 TV |
1216 | gsi_advance_bw_nondebug_nonlocal (gimple_stmt_iterator *gsi, tree *vuse, |
1217 | bool *vuse_escaped) | |
c9e93168 TV |
1218 | { |
1219 | gimple stmt; | |
46301137 | 1220 | tree lvuse; |
c9e93168 TV |
1221 | |
1222 | while (true) | |
1223 | { | |
1224 | if (gsi_end_p (*gsi)) | |
1225 | return; | |
1226 | stmt = gsi_stmt (*gsi); | |
46301137 TV |
1227 | |
1228 | lvuse = gimple_vuse (stmt); | |
1229 | if (lvuse != NULL_TREE) | |
1230 | { | |
1231 | *vuse = lvuse; | |
1232 | if (!ZERO_SSA_OPERANDS (stmt, SSA_OP_DEF)) | |
1233 | *vuse_escaped = true; | |
1234 | } | |
1235 | ||
b2b2f160 | 1236 | if (!stmt_local_def (stmt)) |
c9e93168 TV |
1237 | return; |
1238 | gsi_prev_nondebug (gsi); | |
1239 | } | |
1240 | } | |
1241 | ||
1242 | /* Determines whether BB1 and BB2 (members of same_succ) are duplicates. If so, | |
1243 | clusters them. */ | |
1244 | ||
1245 | static void | |
1246 | find_duplicate (same_succ same_succ, basic_block bb1, basic_block bb2) | |
1247 | { | |
1248 | gimple_stmt_iterator gsi1 = gsi_last_nondebug_bb (bb1); | |
1249 | gimple_stmt_iterator gsi2 = gsi_last_nondebug_bb (bb2); | |
46301137 TV |
1250 | tree vuse1 = NULL_TREE, vuse2 = NULL_TREE; |
1251 | bool vuse_escaped = false; | |
c9e93168 | 1252 | |
46301137 TV |
1253 | gsi_advance_bw_nondebug_nonlocal (&gsi1, &vuse1, &vuse_escaped); |
1254 | gsi_advance_bw_nondebug_nonlocal (&gsi2, &vuse2, &vuse_escaped); | |
c9e93168 TV |
1255 | |
1256 | while (!gsi_end_p (gsi1) && !gsi_end_p (gsi2)) | |
1257 | { | |
7ec88701 RH |
1258 | gimple stmt1 = gsi_stmt (gsi1); |
1259 | gimple stmt2 = gsi_stmt (gsi2); | |
1260 | ||
d5ae1c25 TV |
1261 | /* What could be better than to this this here is to blacklist the bb |
1262 | containing the stmt, when encountering the stmt f.i. in | |
1263 | same_succ_hash. */ | |
1264 | if (is_tm_ending (stmt1) | |
1265 | || is_tm_ending (stmt2)) | |
7ec88701 RH |
1266 | return; |
1267 | ||
d5ae1c25 | 1268 | if (!gimple_equal_p (same_succ, stmt1, stmt2)) |
c9e93168 TV |
1269 | return; |
1270 | ||
1271 | gsi_prev_nondebug (&gsi1); | |
1272 | gsi_prev_nondebug (&gsi2); | |
46301137 TV |
1273 | gsi_advance_bw_nondebug_nonlocal (&gsi1, &vuse1, &vuse_escaped); |
1274 | gsi_advance_bw_nondebug_nonlocal (&gsi2, &vuse2, &vuse_escaped); | |
c9e93168 TV |
1275 | } |
1276 | ||
1277 | if (!(gsi_end_p (gsi1) && gsi_end_p (gsi2))) | |
1278 | return; | |
1279 | ||
46301137 TV |
1280 | /* If the incoming vuses are not the same, and the vuse escaped into an |
1281 | SSA_OP_DEF, then merging the 2 blocks will change the value of the def, | |
1282 | which potentially means the semantics of one of the blocks will be changed. | |
1283 | TODO: make this check more precise. */ | |
1284 | if (vuse_escaped && vuse1 != vuse2) | |
1285 | return; | |
1286 | ||
c9e93168 TV |
1287 | if (dump_file) |
1288 | fprintf (dump_file, "find_duplicates: <bb %d> duplicate of <bb %d>\n", | |
1289 | bb1->index, bb2->index); | |
1290 | ||
1291 | set_cluster (bb1, bb2); | |
1292 | } | |
1293 | ||
1294 | /* Returns whether for all phis in DEST the phi alternatives for E1 and | |
1295 | E2 are equal. */ | |
1296 | ||
1297 | static bool | |
1298 | same_phi_alternatives_1 (basic_block dest, edge e1, edge e2) | |
1299 | { | |
1300 | int n1 = e1->dest_idx, n2 = e2->dest_idx; | |
538dd0b7 | 1301 | gphi_iterator gsi; |
c9e93168 TV |
1302 | |
1303 | for (gsi = gsi_start_phis (dest); !gsi_end_p (gsi); gsi_next (&gsi)) | |
1304 | { | |
538dd0b7 | 1305 | gphi *phi = gsi.phi (); |
c9e93168 TV |
1306 | tree lhs = gimple_phi_result (phi); |
1307 | tree val1 = gimple_phi_arg_def (phi, n1); | |
1308 | tree val2 = gimple_phi_arg_def (phi, n2); | |
1309 | ||
ea057359 | 1310 | if (virtual_operand_p (lhs)) |
c9e93168 TV |
1311 | continue; |
1312 | ||
1313 | if (operand_equal_for_phi_arg_p (val1, val2)) | |
1314 | continue; | |
1315 | if (gvn_uses_equal (val1, val2)) | |
1316 | continue; | |
1317 | ||
1318 | return false; | |
1319 | } | |
1320 | ||
1321 | return true; | |
1322 | } | |
1323 | ||
1324 | /* Returns whether for all successors of BB1 and BB2 (members of SAME_SUCC), the | |
1325 | phi alternatives for BB1 and BB2 are equal. */ | |
1326 | ||
1327 | static bool | |
1328 | same_phi_alternatives (same_succ same_succ, basic_block bb1, basic_block bb2) | |
1329 | { | |
1330 | unsigned int s; | |
1331 | bitmap_iterator bs; | |
1332 | edge e1, e2; | |
1333 | basic_block succ; | |
1334 | ||
1335 | EXECUTE_IF_SET_IN_BITMAP (same_succ->succs, 0, s, bs) | |
1336 | { | |
06e28de2 | 1337 | succ = BASIC_BLOCK_FOR_FN (cfun, s); |
c9e93168 TV |
1338 | e1 = find_edge (bb1, succ); |
1339 | e2 = find_edge (bb2, succ); | |
1340 | if (e1->flags & EDGE_COMPLEX | |
1341 | || e2->flags & EDGE_COMPLEX) | |
1342 | return false; | |
1343 | ||
1344 | /* For all phis in bb, the phi alternatives for e1 and e2 need to have | |
1345 | the same value. */ | |
1346 | if (!same_phi_alternatives_1 (succ, e1, e2)) | |
1347 | return false; | |
1348 | } | |
1349 | ||
1350 | return true; | |
1351 | } | |
1352 | ||
1353 | /* Return true if BB has non-vop phis. */ | |
1354 | ||
1355 | static bool | |
1356 | bb_has_non_vop_phi (basic_block bb) | |
1357 | { | |
1358 | gimple_seq phis = phi_nodes (bb); | |
1359 | gimple phi; | |
1360 | ||
1361 | if (phis == NULL) | |
1362 | return false; | |
1363 | ||
1364 | if (!gimple_seq_singleton_p (phis)) | |
1365 | return true; | |
1366 | ||
1367 | phi = gimple_seq_first_stmt (phis); | |
ea057359 | 1368 | return !virtual_operand_p (gimple_phi_result (phi)); |
c9e93168 TV |
1369 | } |
1370 | ||
1371 | /* Returns true if redirecting the incoming edges of FROM to TO maintains the | |
1372 | invariant that uses in FROM are dominates by their defs. */ | |
1373 | ||
1374 | static bool | |
1375 | deps_ok_for_redirect_from_bb_to_bb (basic_block from, basic_block to) | |
1376 | { | |
1377 | basic_block cd, dep_bb = BB_DEP_BB (to); | |
1378 | edge_iterator ei; | |
1379 | edge e; | |
1380 | bitmap from_preds = BITMAP_ALLOC (NULL); | |
1381 | ||
1382 | if (dep_bb == NULL) | |
1383 | return true; | |
1384 | ||
1385 | FOR_EACH_EDGE (e, ei, from->preds) | |
1386 | bitmap_set_bit (from_preds, e->src->index); | |
1387 | cd = nearest_common_dominator_for_set (CDI_DOMINATORS, from_preds); | |
1388 | BITMAP_FREE (from_preds); | |
1389 | ||
1390 | return dominated_by_p (CDI_DOMINATORS, dep_bb, cd); | |
1391 | } | |
1392 | ||
1393 | /* Returns true if replacing BB1 (or its replacement bb) by BB2 (or its | |
1394 | replacement bb) and vice versa maintains the invariant that uses in the | |
1395 | replacement are dominates by their defs. */ | |
1396 | ||
1397 | static bool | |
1398 | deps_ok_for_redirect (basic_block bb1, basic_block bb2) | |
1399 | { | |
1400 | if (BB_CLUSTER (bb1) != NULL) | |
1401 | bb1 = BB_CLUSTER (bb1)->rep_bb; | |
1402 | ||
1403 | if (BB_CLUSTER (bb2) != NULL) | |
1404 | bb2 = BB_CLUSTER (bb2)->rep_bb; | |
1405 | ||
1406 | return (deps_ok_for_redirect_from_bb_to_bb (bb1, bb2) | |
1407 | && deps_ok_for_redirect_from_bb_to_bb (bb2, bb1)); | |
1408 | } | |
1409 | ||
1410 | /* Within SAME_SUCC->bbs, find clusters of bbs which can be merged. */ | |
1411 | ||
1412 | static void | |
1413 | find_clusters_1 (same_succ same_succ) | |
1414 | { | |
1415 | basic_block bb1, bb2; | |
1416 | unsigned int i, j; | |
1417 | bitmap_iterator bi, bj; | |
1418 | int nr_comparisons; | |
1419 | int max_comparisons = PARAM_VALUE (PARAM_MAX_TAIL_MERGE_COMPARISONS); | |
1420 | ||
1421 | EXECUTE_IF_SET_IN_BITMAP (same_succ->bbs, 0, i, bi) | |
1422 | { | |
06e28de2 | 1423 | bb1 = BASIC_BLOCK_FOR_FN (cfun, i); |
c9e93168 TV |
1424 | |
1425 | /* TODO: handle blocks with phi-nodes. We'll have to find corresponding | |
1426 | phi-nodes in bb1 and bb2, with the same alternatives for the same | |
1427 | preds. */ | |
1428 | if (bb_has_non_vop_phi (bb1)) | |
1429 | continue; | |
1430 | ||
1431 | nr_comparisons = 0; | |
1432 | EXECUTE_IF_SET_IN_BITMAP (same_succ->bbs, i + 1, j, bj) | |
1433 | { | |
06e28de2 | 1434 | bb2 = BASIC_BLOCK_FOR_FN (cfun, j); |
c9e93168 TV |
1435 | |
1436 | if (bb_has_non_vop_phi (bb2)) | |
1437 | continue; | |
1438 | ||
1439 | if (BB_CLUSTER (bb1) != NULL && BB_CLUSTER (bb1) == BB_CLUSTER (bb2)) | |
1440 | continue; | |
1441 | ||
1442 | /* Limit quadratic behaviour. */ | |
1443 | nr_comparisons++; | |
1444 | if (nr_comparisons > max_comparisons) | |
1445 | break; | |
1446 | ||
1447 | /* This is a conservative dependency check. We could test more | |
1448 | precise for allowed replacement direction. */ | |
1449 | if (!deps_ok_for_redirect (bb1, bb2)) | |
1450 | continue; | |
1451 | ||
1452 | if (!(same_phi_alternatives (same_succ, bb1, bb2))) | |
1453 | continue; | |
1454 | ||
1455 | find_duplicate (same_succ, bb1, bb2); | |
1456 | } | |
1457 | } | |
1458 | } | |
1459 | ||
1460 | /* Find clusters of bbs which can be merged. */ | |
1461 | ||
1462 | static void | |
1463 | find_clusters (void) | |
1464 | { | |
1465 | same_succ same; | |
1466 | ||
9771b263 | 1467 | while (!worklist.is_empty ()) |
c9e93168 | 1468 | { |
9771b263 | 1469 | same = worklist.pop (); |
c9e93168 TV |
1470 | same->in_worklist = false; |
1471 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1472 | { | |
1473 | fprintf (dump_file, "processing worklist entry\n"); | |
1474 | same_succ_print (dump_file, same); | |
1475 | } | |
1476 | find_clusters_1 (same); | |
1477 | } | |
1478 | } | |
1479 | ||
c9e93168 TV |
1480 | /* Returns the vop phi of BB, if any. */ |
1481 | ||
538dd0b7 | 1482 | static gphi * |
c9e93168 TV |
1483 | vop_phi (basic_block bb) |
1484 | { | |
538dd0b7 DM |
1485 | gphi *stmt; |
1486 | gphi_iterator gsi; | |
c9e93168 TV |
1487 | for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
1488 | { | |
538dd0b7 | 1489 | stmt = gsi.phi (); |
ea057359 | 1490 | if (! virtual_operand_p (gimple_phi_result (stmt))) |
c9e93168 TV |
1491 | continue; |
1492 | return stmt; | |
1493 | } | |
1494 | return NULL; | |
1495 | } | |
1496 | ||
643400b8 | 1497 | /* Redirect all edges from BB1 to BB2, removes BB1 and marks it as removed. */ |
c9e93168 TV |
1498 | |
1499 | static void | |
643400b8 | 1500 | replace_block_by (basic_block bb1, basic_block bb2) |
c9e93168 TV |
1501 | { |
1502 | edge pred_edge; | |
b6c00353 | 1503 | edge e1, e2; |
adc7a812 | 1504 | edge_iterator ei; |
c9e93168 | 1505 | unsigned int i; |
538dd0b7 | 1506 | gphi *bb2_phi; |
40f73edd | 1507 | |
643400b8 | 1508 | bb2_phi = vop_phi (bb2); |
c9e93168 | 1509 | |
643400b8 TV |
1510 | /* Mark the basic block as deleted. */ |
1511 | mark_basic_block_deleted (bb1); | |
c9e93168 TV |
1512 | |
1513 | /* Redirect the incoming edges of bb1 to bb2. */ | |
1514 | for (i = EDGE_COUNT (bb1->preds); i > 0 ; --i) | |
1515 | { | |
1516 | pred_edge = EDGE_PRED (bb1, i - 1); | |
1517 | pred_edge = redirect_edge_and_branch (pred_edge, bb2); | |
1518 | gcc_assert (pred_edge != NULL); | |
643400b8 TV |
1519 | |
1520 | if (bb2_phi == NULL) | |
1521 | continue; | |
1522 | ||
1523 | /* The phi might have run out of capacity when the redirect added an | |
1524 | argument, which means it could have been replaced. Refresh it. */ | |
1525 | bb2_phi = vop_phi (bb2); | |
1526 | ||
1527 | add_phi_arg (bb2_phi, SSA_NAME_VAR (gimple_phi_result (bb2_phi)), | |
9e227d60 | 1528 | pred_edge, UNKNOWN_LOCATION); |
c9e93168 TV |
1529 | } |
1530 | ||
fa405d7b TV |
1531 | bb2->frequency += bb1->frequency; |
1532 | if (bb2->frequency > BB_FREQ_MAX) | |
1533 | bb2->frequency = BB_FREQ_MAX; | |
55fcb901 CB |
1534 | |
1535 | bb2->count += bb1->count; | |
fa405d7b | 1536 | |
adc7a812 | 1537 | /* Merge the outgoing edge counts from bb1 onto bb2. */ |
b6c00353 | 1538 | gcov_type out_sum = 0; |
adc7a812 TJ |
1539 | FOR_EACH_EDGE (e1, ei, bb1->succs) |
1540 | { | |
adc7a812 TJ |
1541 | e2 = find_edge (bb2, e1->dest); |
1542 | gcc_assert (e2); | |
1543 | e2->count += e1->count; | |
b6c00353 TJ |
1544 | out_sum += e2->count; |
1545 | } | |
1546 | /* Recompute the edge probabilities from the new merged edge count. | |
1547 | Use the sum of the new merged edge counts computed above instead | |
1548 | of bb2's merged count, in case there are profile count insanities | |
1549 | making the bb count inconsistent with the edge weights. */ | |
1550 | FOR_EACH_EDGE (e2, ei, bb2->succs) | |
1551 | { | |
1552 | e2->probability = GCOV_COMPUTE_SCALE (e2->count, out_sum); | |
adc7a812 TJ |
1553 | } |
1554 | ||
a31895d7 | 1555 | /* Do updates that use bb1, before deleting bb1. */ |
fcddd80e | 1556 | release_last_vdef (bb1); |
a31895d7 TV |
1557 | same_succ_flush_bb (bb1); |
1558 | ||
643400b8 | 1559 | delete_basic_block (bb1); |
c9e93168 TV |
1560 | } |
1561 | ||
1562 | /* Bbs for which update_debug_stmt need to be called. */ | |
1563 | ||
1564 | static bitmap update_bbs; | |
1565 | ||
1566 | /* For each cluster in all_clusters, merge all cluster->bbs. Returns | |
643400b8 | 1567 | number of bbs removed. */ |
c9e93168 TV |
1568 | |
1569 | static int | |
643400b8 | 1570 | apply_clusters (void) |
c9e93168 TV |
1571 | { |
1572 | basic_block bb1, bb2; | |
1573 | bb_cluster c; | |
1574 | unsigned int i, j; | |
1575 | bitmap_iterator bj; | |
1576 | int nr_bbs_removed = 0; | |
1577 | ||
9771b263 | 1578 | for (i = 0; i < all_clusters.length (); ++i) |
c9e93168 | 1579 | { |
9771b263 | 1580 | c = all_clusters[i]; |
c9e93168 TV |
1581 | if (c == NULL) |
1582 | continue; | |
1583 | ||
1584 | bb2 = c->rep_bb; | |
1585 | bitmap_set_bit (update_bbs, bb2->index); | |
1586 | ||
1587 | bitmap_clear_bit (c->bbs, bb2->index); | |
1588 | EXECUTE_IF_SET_IN_BITMAP (c->bbs, 0, j, bj) | |
1589 | { | |
06e28de2 | 1590 | bb1 = BASIC_BLOCK_FOR_FN (cfun, j); |
c9e93168 TV |
1591 | bitmap_clear_bit (update_bbs, bb1->index); |
1592 | ||
643400b8 | 1593 | replace_block_by (bb1, bb2); |
c9e93168 TV |
1594 | nr_bbs_removed++; |
1595 | } | |
1596 | } | |
1597 | ||
1598 | return nr_bbs_removed; | |
1599 | } | |
1600 | ||
1601 | /* Resets debug statement STMT if it has uses that are not dominated by their | |
1602 | defs. */ | |
1603 | ||
1604 | static void | |
1605 | update_debug_stmt (gimple stmt) | |
1606 | { | |
1607 | use_operand_p use_p; | |
1608 | ssa_op_iter oi; | |
a92bf1b1 | 1609 | basic_block bbuse; |
c9e93168 TV |
1610 | |
1611 | if (!gimple_debug_bind_p (stmt)) | |
1612 | return; | |
1613 | ||
1614 | bbuse = gimple_bb (stmt); | |
1615 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, oi, SSA_OP_USE) | |
1616 | { | |
a92bf1b1 RB |
1617 | tree name = USE_FROM_PTR (use_p); |
1618 | gimple def_stmt = SSA_NAME_DEF_STMT (name); | |
1619 | basic_block bbdef = gimple_bb (def_stmt); | |
c9e93168 TV |
1620 | if (bbdef == NULL || bbuse == bbdef |
1621 | || dominated_by_p (CDI_DOMINATORS, bbuse, bbdef)) | |
1622 | continue; | |
1623 | ||
1624 | gimple_debug_bind_reset_value (stmt); | |
1625 | update_stmt (stmt); | |
a92bf1b1 | 1626 | break; |
c9e93168 TV |
1627 | } |
1628 | } | |
1629 | ||
1630 | /* Resets all debug statements that have uses that are not | |
1631 | dominated by their defs. */ | |
1632 | ||
1633 | static void | |
1634 | update_debug_stmts (void) | |
1635 | { | |
1636 | basic_block bb; | |
1637 | bitmap_iterator bi; | |
1638 | unsigned int i; | |
1639 | ||
c9e93168 TV |
1640 | EXECUTE_IF_SET_IN_BITMAP (update_bbs, 0, i, bi) |
1641 | { | |
1642 | gimple stmt; | |
1643 | gimple_stmt_iterator gsi; | |
1644 | ||
06e28de2 | 1645 | bb = BASIC_BLOCK_FOR_FN (cfun, i); |
c9e93168 TV |
1646 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
1647 | { | |
1648 | stmt = gsi_stmt (gsi); | |
1649 | if (!is_gimple_debug (stmt)) | |
1650 | continue; | |
1651 | update_debug_stmt (stmt); | |
1652 | } | |
1653 | } | |
1654 | } | |
1655 | ||
1656 | /* Runs tail merge optimization. */ | |
1657 | ||
1658 | unsigned int | |
1659 | tail_merge_optimize (unsigned int todo) | |
1660 | { | |
1661 | int nr_bbs_removed_total = 0; | |
1662 | int nr_bbs_removed; | |
1663 | bool loop_entered = false; | |
1664 | int iteration_nr = 0; | |
c9e93168 TV |
1665 | int max_iterations = PARAM_VALUE (PARAM_MAX_TAIL_MERGE_ITERATIONS); |
1666 | ||
315bbd2e | 1667 | if (!flag_tree_tail_merge |
726338f4 | 1668 | || max_iterations == 0) |
c9e93168 TV |
1669 | return 0; |
1670 | ||
1671 | timevar_push (TV_TREE_TAIL_MERGE); | |
1672 | ||
9c370032 RB |
1673 | if (!dom_info_available_p (CDI_DOMINATORS)) |
1674 | { | |
1675 | /* PRE can leave us with unreachable blocks, remove them now. */ | |
1676 | delete_unreachable_blocks (); | |
1677 | calculate_dominance_info (CDI_DOMINATORS); | |
1678 | } | |
c9e93168 TV |
1679 | init_worklist (); |
1680 | ||
9771b263 | 1681 | while (!worklist.is_empty ()) |
c9e93168 TV |
1682 | { |
1683 | if (!loop_entered) | |
1684 | { | |
1685 | loop_entered = true; | |
1686 | alloc_cluster_vectors (); | |
1687 | update_bbs = BITMAP_ALLOC (NULL); | |
1688 | } | |
1689 | else | |
1690 | reset_cluster_vectors (); | |
1691 | ||
1692 | iteration_nr++; | |
1693 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1694 | fprintf (dump_file, "worklist iteration #%d\n", iteration_nr); | |
1695 | ||
1696 | find_clusters (); | |
9771b263 DN |
1697 | gcc_assert (worklist.is_empty ()); |
1698 | if (all_clusters.is_empty ()) | |
c9e93168 TV |
1699 | break; |
1700 | ||
643400b8 | 1701 | nr_bbs_removed = apply_clusters (); |
c9e93168 TV |
1702 | nr_bbs_removed_total += nr_bbs_removed; |
1703 | if (nr_bbs_removed == 0) | |
1704 | break; | |
1705 | ||
643400b8 | 1706 | free_dominance_info (CDI_DOMINATORS); |
c9e93168 TV |
1707 | |
1708 | if (iteration_nr == max_iterations) | |
1709 | break; | |
1710 | ||
643400b8 | 1711 | calculate_dominance_info (CDI_DOMINATORS); |
c9e93168 TV |
1712 | update_worklist (); |
1713 | } | |
1714 | ||
1715 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1716 | fprintf (dump_file, "htab collision / search: %f\n", | |
c203e8a7 | 1717 | same_succ_htab->collisions ()); |
c9e93168 TV |
1718 | |
1719 | if (nr_bbs_removed_total > 0) | |
1720 | { | |
643400b8 TV |
1721 | if (MAY_HAVE_DEBUG_STMTS) |
1722 | { | |
1723 | calculate_dominance_info (CDI_DOMINATORS); | |
1724 | update_debug_stmts (); | |
1725 | } | |
c9e93168 TV |
1726 | |
1727 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1728 | { | |
1729 | fprintf (dump_file, "Before TODOs.\n"); | |
1730 | dump_function_to_file (current_function_decl, dump_file, dump_flags); | |
1731 | } | |
1732 | ||
525174a2 | 1733 | mark_virtual_operands_for_renaming (cfun); |
c9e93168 TV |
1734 | } |
1735 | ||
1736 | delete_worklist (); | |
1737 | if (loop_entered) | |
1738 | { | |
1739 | delete_cluster_vectors (); | |
1740 | BITMAP_FREE (update_bbs); | |
1741 | } | |
1742 | ||
1743 | timevar_pop (TV_TREE_TAIL_MERGE); | |
1744 | ||
1745 | return todo; | |
1746 | } |