]>
Commit | Line | Data |
---|---|---|
c9e93168 | 1 | /* Tail merging for gimple. |
5624e564 | 2 | Copyright (C) 2011-2015 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) | |
daf347dd | 95 | 6 [100.0%] (fallthru,exec) |
c9e93168 TV |
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), | |
daf347dd | 100 | .MEMD.3923_18(6)> |
c9e93168 TV |
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" | |
c7131fb2 | 191 | #include "backend.h" |
9fdcd34e | 192 | #include "cfghooks.h" |
c9e93168 | 193 | #include "tree.h" |
c7131fb2 AM |
194 | #include "gimple.h" |
195 | #include "hard-reg-set.h" | |
196 | #include "ssa.h" | |
197 | #include "alias.h" | |
40e23961 | 198 | #include "fold-const.h" |
d8a2d370 DN |
199 | #include "stor-layout.h" |
200 | #include "trans-mem.h" | |
c9e93168 | 201 | #include "tm_p.h" |
60393bbc AM |
202 | #include "cfganal.h" |
203 | #include "cfgcleanup.h" | |
60393bbc | 204 | #include "flags.h" |
2fb9a547 AM |
205 | #include "internal-fn.h" |
206 | #include "tree-eh.h" | |
5be5c238 | 207 | #include "gimple-iterator.h" |
442b4905 | 208 | #include "tree-cfg.h" |
442b4905 | 209 | #include "tree-into-ssa.h" |
c9e93168 | 210 | #include "params.h" |
c9e93168 TV |
211 | #include "gimple-pretty-print.h" |
212 | #include "tree-ssa-sccvn.h" | |
213 | #include "tree-dump.h" | |
a9e0d843 | 214 | #include "cfgloop.h" |
7ee2468b | 215 | #include "tree-pass.h" |
1fe37220 | 216 | #include "trans-mem.h" |
7ee2468b | 217 | |
c9e93168 TV |
218 | /* Describes a group of bbs with the same successors. The successor bbs are |
219 | cached in succs, and the successor edge flags are cached in succ_flags. | |
f0cbb0af | 220 | If a bb has the EDGE_TRUE/FALSE_VALUE flags swapped compared to succ_flags, |
c9e93168 TV |
221 | it's marked in inverse. |
222 | Additionally, the hash value for the struct is cached in hashval, and | |
223 | in_worklist indicates whether it's currently part of worklist. */ | |
224 | ||
7edd9b15 | 225 | struct same_succ_def : pointer_hash <same_succ_def> |
c9e93168 TV |
226 | { |
227 | /* The bbs that have the same successor bbs. */ | |
228 | bitmap bbs; | |
229 | /* The successor bbs. */ | |
230 | bitmap succs; | |
231 | /* Indicates whether the EDGE_TRUE/FALSE_VALUEs of succ_flags are swapped for | |
232 | bb. */ | |
233 | bitmap inverse; | |
234 | /* The edge flags for each of the successor bbs. */ | |
9771b263 | 235 | vec<int> succ_flags; |
c9e93168 TV |
236 | /* Indicates whether the struct is currently in the worklist. */ |
237 | bool in_worklist; | |
238 | /* The hash value of the struct. */ | |
239 | hashval_t hashval; | |
5deac340 RG |
240 | |
241 | /* hash_table support. */ | |
67f58944 TS |
242 | static inline hashval_t hash (const same_succ_def *); |
243 | static int equal (const same_succ_def *, const same_succ_def *); | |
244 | static void remove (same_succ_def *); | |
c9e93168 TV |
245 | }; |
246 | typedef struct same_succ_def *same_succ; | |
247 | typedef const struct same_succ_def *const_same_succ; | |
248 | ||
5deac340 RG |
249 | /* hash routine for hash_table support, returns hashval of E. */ |
250 | ||
251 | inline hashval_t | |
67f58944 | 252 | same_succ_def::hash (const same_succ_def *e) |
5deac340 RG |
253 | { |
254 | return e->hashval; | |
255 | } | |
256 | ||
c9e93168 TV |
257 | /* A group of bbs where 1 bb from bbs can replace the other bbs. */ |
258 | ||
259 | struct bb_cluster_def | |
260 | { | |
261 | /* The bbs in the cluster. */ | |
262 | bitmap bbs; | |
263 | /* The preds of the bbs in the cluster. */ | |
264 | bitmap preds; | |
265 | /* Index in all_clusters vector. */ | |
266 | int index; | |
267 | /* The bb to replace the cluster with. */ | |
268 | basic_block rep_bb; | |
269 | }; | |
270 | typedef struct bb_cluster_def *bb_cluster; | |
271 | typedef const struct bb_cluster_def *const_bb_cluster; | |
272 | ||
273 | /* Per bb-info. */ | |
274 | ||
275 | struct aux_bb_info | |
276 | { | |
277 | /* The number of non-debug statements in the bb. */ | |
278 | int size; | |
279 | /* The same_succ that this bb is a member of. */ | |
280 | same_succ bb_same_succ; | |
281 | /* The cluster that this bb is a member of. */ | |
282 | bb_cluster cluster; | |
283 | /* The vop state at the exit of a bb. This is shortlived data, used to | |
284 | communicate data between update_block_by and update_vuses. */ | |
285 | tree vop_at_exit; | |
286 | /* The bb that either contains or is dominated by the dependencies of the | |
287 | bb. */ | |
288 | basic_block dep_bb; | |
289 | }; | |
290 | ||
291 | /* Macros to access the fields of struct aux_bb_info. */ | |
292 | ||
293 | #define BB_SIZE(bb) (((struct aux_bb_info *)bb->aux)->size) | |
294 | #define BB_SAME_SUCC(bb) (((struct aux_bb_info *)bb->aux)->bb_same_succ) | |
295 | #define BB_CLUSTER(bb) (((struct aux_bb_info *)bb->aux)->cluster) | |
296 | #define BB_VOP_AT_EXIT(bb) (((struct aux_bb_info *)bb->aux)->vop_at_exit) | |
297 | #define BB_DEP_BB(bb) (((struct aux_bb_info *)bb->aux)->dep_bb) | |
298 | ||
b2b2f160 TV |
299 | /* Returns true if the only effect a statement STMT has, is to define locally |
300 | used SSA_NAMEs. */ | |
301 | ||
302 | static bool | |
303 | stmt_local_def (gimple stmt) | |
304 | { | |
305 | basic_block bb, def_bb; | |
306 | imm_use_iterator iter; | |
307 | use_operand_p use_p; | |
308 | tree val; | |
309 | def_operand_p def_p; | |
310 | ||
12d3031b RB |
311 | if (gimple_vdef (stmt) != NULL_TREE |
312 | || gimple_has_side_effects (stmt) | |
8024d4b6 TV |
313 | || gimple_could_trap_p_1 (stmt, false, false) |
314 | || gimple_vuse (stmt) != NULL_TREE) | |
b2b2f160 TV |
315 | return false; |
316 | ||
317 | def_p = SINGLE_SSA_DEF_OPERAND (stmt, SSA_OP_DEF); | |
318 | if (def_p == NULL) | |
319 | return false; | |
320 | ||
321 | val = DEF_FROM_PTR (def_p); | |
322 | if (val == NULL_TREE || TREE_CODE (val) != SSA_NAME) | |
323 | return false; | |
324 | ||
325 | def_bb = gimple_bb (stmt); | |
326 | ||
327 | FOR_EACH_IMM_USE_FAST (use_p, iter, val) | |
328 | { | |
329 | if (is_gimple_debug (USE_STMT (use_p))) | |
330 | continue; | |
331 | bb = gimple_bb (USE_STMT (use_p)); | |
332 | if (bb == def_bb) | |
333 | continue; | |
334 | ||
335 | if (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI | |
336 | && EDGE_PRED (bb, PHI_ARG_INDEX_FROM_USE (use_p))->src == def_bb) | |
337 | continue; | |
338 | ||
339 | return false; | |
340 | } | |
341 | ||
342 | return true; | |
343 | } | |
344 | ||
345 | /* Let GSI skip forwards over local defs. */ | |
346 | ||
347 | static void | |
348 | gsi_advance_fw_nondebug_nonlocal (gimple_stmt_iterator *gsi) | |
349 | { | |
350 | gimple stmt; | |
351 | ||
352 | while (true) | |
353 | { | |
354 | if (gsi_end_p (*gsi)) | |
355 | return; | |
356 | stmt = gsi_stmt (*gsi); | |
357 | if (!stmt_local_def (stmt)) | |
358 | return; | |
21c0a521 | 359 | gsi_next_nondebug (gsi); |
b2b2f160 TV |
360 | } |
361 | } | |
362 | ||
c9e93168 TV |
363 | /* VAL1 and VAL2 are either: |
364 | - uses in BB1 and BB2, or | |
365 | - phi alternatives for BB1 and BB2. | |
366 | Return true if the uses have the same gvn value. */ | |
367 | ||
368 | static bool | |
369 | gvn_uses_equal (tree val1, tree val2) | |
370 | { | |
371 | gcc_checking_assert (val1 != NULL_TREE && val2 != NULL_TREE); | |
372 | ||
373 | if (val1 == val2) | |
374 | return true; | |
375 | ||
376 | if (vn_valueize (val1) != vn_valueize (val2)) | |
377 | return false; | |
378 | ||
379 | return ((TREE_CODE (val1) == SSA_NAME || CONSTANT_CLASS_P (val1)) | |
380 | && (TREE_CODE (val2) == SSA_NAME || CONSTANT_CLASS_P (val2))); | |
381 | } | |
382 | ||
383 | /* Prints E to FILE. */ | |
384 | ||
385 | static void | |
386 | same_succ_print (FILE *file, const same_succ e) | |
387 | { | |
388 | unsigned int i; | |
389 | bitmap_print (file, e->bbs, "bbs:", "\n"); | |
390 | bitmap_print (file, e->succs, "succs:", "\n"); | |
391 | bitmap_print (file, e->inverse, "inverse:", "\n"); | |
392 | fprintf (file, "flags:"); | |
9771b263 DN |
393 | for (i = 0; i < e->succ_flags.length (); ++i) |
394 | fprintf (file, " %x", e->succ_flags[i]); | |
c9e93168 TV |
395 | fprintf (file, "\n"); |
396 | } | |
397 | ||
398 | /* Prints same_succ VE to VFILE. */ | |
399 | ||
0823efed DN |
400 | inline int |
401 | ssa_same_succ_print_traverse (same_succ *pe, FILE *file) | |
c9e93168 | 402 | { |
0823efed | 403 | const same_succ e = *pe; |
c9e93168 TV |
404 | same_succ_print (file, e); |
405 | return 1; | |
406 | } | |
407 | ||
408 | /* Update BB_DEP_BB (USE_BB), given a use of VAL in USE_BB. */ | |
409 | ||
410 | static void | |
411 | update_dep_bb (basic_block use_bb, tree val) | |
412 | { | |
413 | basic_block dep_bb; | |
414 | ||
415 | /* Not a dep. */ | |
416 | if (TREE_CODE (val) != SSA_NAME) | |
417 | return; | |
418 | ||
419 | /* Skip use of global def. */ | |
420 | if (SSA_NAME_IS_DEFAULT_DEF (val)) | |
421 | return; | |
422 | ||
423 | /* Skip use of local def. */ | |
424 | dep_bb = gimple_bb (SSA_NAME_DEF_STMT (val)); | |
425 | if (dep_bb == use_bb) | |
426 | return; | |
427 | ||
428 | if (BB_DEP_BB (use_bb) == NULL | |
429 | || dominated_by_p (CDI_DOMINATORS, dep_bb, BB_DEP_BB (use_bb))) | |
430 | BB_DEP_BB (use_bb) = dep_bb; | |
431 | } | |
432 | ||
433 | /* Update BB_DEP_BB, given the dependencies in STMT. */ | |
434 | ||
435 | static void | |
436 | stmt_update_dep_bb (gimple stmt) | |
437 | { | |
438 | ssa_op_iter iter; | |
439 | use_operand_p use; | |
440 | ||
441 | FOR_EACH_SSA_USE_OPERAND (use, stmt, iter, SSA_OP_USE) | |
442 | update_dep_bb (gimple_bb (stmt), USE_FROM_PTR (use)); | |
443 | } | |
444 | ||
c9e93168 TV |
445 | /* Calculates hash value for same_succ VE. */ |
446 | ||
5deac340 RG |
447 | static hashval_t |
448 | same_succ_hash (const_same_succ e) | |
c9e93168 | 449 | { |
bb7e83fc | 450 | inchash::hash hstate (bitmap_hash (e->succs)); |
c9e93168 TV |
451 | int flags; |
452 | unsigned int i; | |
453 | unsigned int first = bitmap_first_set_bit (e->bbs); | |
06e28de2 | 454 | basic_block bb = BASIC_BLOCK_FOR_FN (cfun, first); |
c9e93168 | 455 | int size = 0; |
c9e93168 TV |
456 | gimple stmt; |
457 | tree arg; | |
458 | unsigned int s; | |
459 | bitmap_iterator bs; | |
460 | ||
538dd0b7 | 461 | for (gimple_stmt_iterator gsi = gsi_start_nondebug_bb (bb); |
c9e93168 TV |
462 | !gsi_end_p (gsi); gsi_next_nondebug (&gsi)) |
463 | { | |
464 | stmt = gsi_stmt (gsi); | |
465 | stmt_update_dep_bb (stmt); | |
b2b2f160 | 466 | if (stmt_local_def (stmt)) |
c9e93168 TV |
467 | continue; |
468 | size++; | |
469 | ||
bb7e83fc | 470 | hstate.add_int (gimple_code (stmt)); |
c9e93168 | 471 | if (is_gimple_assign (stmt)) |
bb7e83fc | 472 | hstate.add_int (gimple_assign_rhs_code (stmt)); |
c9e93168 TV |
473 | if (!is_gimple_call (stmt)) |
474 | continue; | |
475 | if (gimple_call_internal_p (stmt)) | |
daf347dd | 476 | hstate.add_int (gimple_call_internal_fn (stmt)); |
c9e93168 | 477 | else |
51d2abeb | 478 | { |
bb7e83fc | 479 | inchash::add_expr (gimple_call_fn (stmt), hstate); |
51d2abeb | 480 | if (gimple_call_chain (stmt)) |
bb7e83fc | 481 | inchash::add_expr (gimple_call_chain (stmt), hstate); |
51d2abeb | 482 | } |
c9e93168 TV |
483 | for (i = 0; i < gimple_call_num_args (stmt); i++) |
484 | { | |
485 | arg = gimple_call_arg (stmt, i); | |
486 | arg = vn_valueize (arg); | |
bb7e83fc | 487 | inchash::add_expr (arg, hstate); |
c9e93168 TV |
488 | } |
489 | } | |
490 | ||
bb7e83fc | 491 | hstate.add_int (size); |
c9e93168 TV |
492 | BB_SIZE (bb) = size; |
493 | ||
9771b263 | 494 | for (i = 0; i < e->succ_flags.length (); ++i) |
c9e93168 | 495 | { |
9771b263 | 496 | flags = e->succ_flags[i]; |
c9e93168 | 497 | flags = flags & ~(EDGE_TRUE_VALUE | EDGE_FALSE_VALUE); |
bb7e83fc | 498 | hstate.add_int (flags); |
c9e93168 TV |
499 | } |
500 | ||
501 | EXECUTE_IF_SET_IN_BITMAP (e->succs, 0, s, bs) | |
502 | { | |
06e28de2 | 503 | int n = find_edge (bb, BASIC_BLOCK_FOR_FN (cfun, s))->dest_idx; |
538dd0b7 DM |
504 | for (gphi_iterator gsi = gsi_start_phis (BASIC_BLOCK_FOR_FN (cfun, s)); |
505 | !gsi_end_p (gsi); | |
c9e93168 TV |
506 | gsi_next (&gsi)) |
507 | { | |
538dd0b7 | 508 | gphi *phi = gsi.phi (); |
c9e93168 TV |
509 | tree lhs = gimple_phi_result (phi); |
510 | tree val = gimple_phi_arg_def (phi, n); | |
511 | ||
ea057359 | 512 | if (virtual_operand_p (lhs)) |
c9e93168 TV |
513 | continue; |
514 | update_dep_bb (bb, val); | |
515 | } | |
516 | } | |
517 | ||
bb7e83fc | 518 | return hstate.end (); |
c9e93168 TV |
519 | } |
520 | ||
521 | /* Returns true if E1 and E2 have 2 successors, and if the successor flags | |
522 | are inverse for the EDGE_TRUE_VALUE and EDGE_FALSE_VALUE flags, and equal for | |
523 | the other edge flags. */ | |
524 | ||
525 | static bool | |
526 | inverse_flags (const_same_succ e1, const_same_succ e2) | |
527 | { | |
528 | int f1a, f1b, f2a, f2b; | |
529 | int mask = ~(EDGE_TRUE_VALUE | EDGE_FALSE_VALUE); | |
530 | ||
9771b263 | 531 | if (e1->succ_flags.length () != 2) |
c9e93168 TV |
532 | return false; |
533 | ||
9771b263 DN |
534 | f1a = e1->succ_flags[0]; |
535 | f1b = e1->succ_flags[1]; | |
536 | f2a = e2->succ_flags[0]; | |
537 | f2b = e2->succ_flags[1]; | |
c9e93168 TV |
538 | |
539 | if (f1a == f2a && f1b == f2b) | |
540 | return false; | |
541 | ||
542 | return (f1a & mask) == (f2a & mask) && (f1b & mask) == (f2b & mask); | |
543 | } | |
544 | ||
5deac340 | 545 | /* Compares SAME_SUCCs E1 and E2. */ |
c9e93168 | 546 | |
0823efed | 547 | int |
67f58944 | 548 | same_succ_def::equal (const same_succ_def *e1, const same_succ_def *e2) |
c9e93168 | 549 | { |
c9e93168 TV |
550 | unsigned int i, first1, first2; |
551 | gimple_stmt_iterator gsi1, gsi2; | |
552 | gimple s1, s2; | |
553 | basic_block bb1, bb2; | |
554 | ||
555 | if (e1->hashval != e2->hashval) | |
556 | return 0; | |
557 | ||
9771b263 | 558 | if (e1->succ_flags.length () != e2->succ_flags.length ()) |
c9e93168 TV |
559 | return 0; |
560 | ||
561 | if (!bitmap_equal_p (e1->succs, e2->succs)) | |
562 | return 0; | |
563 | ||
564 | if (!inverse_flags (e1, e2)) | |
565 | { | |
9771b263 | 566 | for (i = 0; i < e1->succ_flags.length (); ++i) |
707550e4 | 567 | if (e1->succ_flags[i] != e2->succ_flags[i]) |
c9e93168 TV |
568 | return 0; |
569 | } | |
570 | ||
571 | first1 = bitmap_first_set_bit (e1->bbs); | |
572 | first2 = bitmap_first_set_bit (e2->bbs); | |
573 | ||
06e28de2 DM |
574 | bb1 = BASIC_BLOCK_FOR_FN (cfun, first1); |
575 | bb2 = BASIC_BLOCK_FOR_FN (cfun, first2); | |
c9e93168 TV |
576 | |
577 | if (BB_SIZE (bb1) != BB_SIZE (bb2)) | |
578 | return 0; | |
579 | ||
580 | gsi1 = gsi_start_nondebug_bb (bb1); | |
581 | gsi2 = gsi_start_nondebug_bb (bb2); | |
b2b2f160 TV |
582 | gsi_advance_fw_nondebug_nonlocal (&gsi1); |
583 | gsi_advance_fw_nondebug_nonlocal (&gsi2); | |
c9e93168 TV |
584 | while (!(gsi_end_p (gsi1) || gsi_end_p (gsi2))) |
585 | { | |
586 | s1 = gsi_stmt (gsi1); | |
587 | s2 = gsi_stmt (gsi2); | |
588 | if (gimple_code (s1) != gimple_code (s2)) | |
589 | return 0; | |
590 | if (is_gimple_call (s1) && !gimple_call_same_target_p (s1, s2)) | |
591 | return 0; | |
592 | gsi_next_nondebug (&gsi1); | |
593 | gsi_next_nondebug (&gsi2); | |
b2b2f160 TV |
594 | gsi_advance_fw_nondebug_nonlocal (&gsi1); |
595 | gsi_advance_fw_nondebug_nonlocal (&gsi2); | |
c9e93168 TV |
596 | } |
597 | ||
598 | return 1; | |
599 | } | |
600 | ||
601 | /* Alloc and init a new SAME_SUCC. */ | |
602 | ||
603 | static same_succ | |
604 | same_succ_alloc (void) | |
605 | { | |
606 | same_succ same = XNEW (struct same_succ_def); | |
607 | ||
608 | same->bbs = BITMAP_ALLOC (NULL); | |
609 | same->succs = BITMAP_ALLOC (NULL); | |
610 | same->inverse = BITMAP_ALLOC (NULL); | |
9771b263 | 611 | same->succ_flags.create (10); |
c9e93168 TV |
612 | same->in_worklist = false; |
613 | ||
614 | return same; | |
615 | } | |
616 | ||
5deac340 | 617 | /* Delete same_succ E. */ |
c9e93168 | 618 | |
5deac340 RG |
619 | void |
620 | same_succ_def::remove (same_succ e) | |
c9e93168 | 621 | { |
c9e93168 TV |
622 | BITMAP_FREE (e->bbs); |
623 | BITMAP_FREE (e->succs); | |
624 | BITMAP_FREE (e->inverse); | |
9771b263 | 625 | e->succ_flags.release (); |
c9e93168 | 626 | |
0823efed | 627 | XDELETE (e); |
c9e93168 TV |
628 | } |
629 | ||
630 | /* Reset same_succ SAME. */ | |
631 | ||
632 | static void | |
633 | same_succ_reset (same_succ same) | |
634 | { | |
635 | bitmap_clear (same->bbs); | |
636 | bitmap_clear (same->succs); | |
637 | bitmap_clear (same->inverse); | |
9771b263 | 638 | same->succ_flags.truncate (0); |
c9e93168 TV |
639 | } |
640 | ||
c203e8a7 | 641 | static hash_table<same_succ_def> *same_succ_htab; |
c9e93168 TV |
642 | |
643 | /* Array that is used to store the edge flags for a successor. */ | |
644 | ||
645 | static int *same_succ_edge_flags; | |
646 | ||
647 | /* Bitmap that is used to mark bbs that are recently deleted. */ | |
648 | ||
649 | static bitmap deleted_bbs; | |
650 | ||
651 | /* Bitmap that is used to mark predecessors of bbs that are | |
652 | deleted. */ | |
653 | ||
654 | static bitmap deleted_bb_preds; | |
655 | ||
656 | /* Prints same_succ_htab to stderr. */ | |
657 | ||
658 | extern void debug_same_succ (void); | |
659 | DEBUG_FUNCTION void | |
660 | debug_same_succ ( void) | |
661 | { | |
c203e8a7 | 662 | same_succ_htab->traverse <FILE *, ssa_same_succ_print_traverse> (stderr); |
c9e93168 TV |
663 | } |
664 | ||
c9e93168 TV |
665 | |
666 | /* Vector of bbs to process. */ | |
667 | ||
9771b263 | 668 | static vec<same_succ> worklist; |
c9e93168 TV |
669 | |
670 | /* Prints worklist to FILE. */ | |
671 | ||
672 | static void | |
673 | print_worklist (FILE *file) | |
674 | { | |
675 | unsigned int i; | |
9771b263 DN |
676 | for (i = 0; i < worklist.length (); ++i) |
677 | same_succ_print (file, worklist[i]); | |
c9e93168 TV |
678 | } |
679 | ||
680 | /* Adds SAME to worklist. */ | |
681 | ||
682 | static void | |
683 | add_to_worklist (same_succ same) | |
684 | { | |
685 | if (same->in_worklist) | |
686 | return; | |
687 | ||
688 | if (bitmap_count_bits (same->bbs) < 2) | |
689 | return; | |
690 | ||
691 | same->in_worklist = true; | |
9771b263 | 692 | worklist.safe_push (same); |
c9e93168 TV |
693 | } |
694 | ||
695 | /* Add BB to same_succ_htab. */ | |
696 | ||
697 | static void | |
698 | find_same_succ_bb (basic_block bb, same_succ *same_p) | |
699 | { | |
700 | unsigned int j; | |
701 | bitmap_iterator bj; | |
702 | same_succ same = *same_p; | |
703 | same_succ *slot; | |
704 | edge_iterator ei; | |
705 | edge e; | |
706 | ||
315bbd2e TV |
707 | if (bb == NULL |
708 | /* Be conservative with loop structure. It's not evident that this test | |
709 | is sufficient. Before tail-merge, we've just called | |
710 | loop_optimizer_finalize, and LOOPS_MAY_HAVE_MULTIPLE_LATCHES is now | |
711 | set, so there's no guarantee that the loop->latch value is still valid. | |
712 | But we assume that, since we've forced LOOPS_HAVE_SIMPLE_LATCHES at the | |
713 | start of pre, we've kept that property intact throughout pre, and are | |
714 | keeping it throughout tail-merge using this test. */ | |
715 | || bb->loop_father->latch == bb) | |
c9e93168 TV |
716 | return; |
717 | bitmap_set_bit (same->bbs, bb->index); | |
718 | FOR_EACH_EDGE (e, ei, bb->succs) | |
719 | { | |
720 | int index = e->dest->index; | |
721 | bitmap_set_bit (same->succs, index); | |
722 | same_succ_edge_flags[index] = e->flags; | |
723 | } | |
724 | EXECUTE_IF_SET_IN_BITMAP (same->succs, 0, j, bj) | |
9771b263 | 725 | same->succ_flags.safe_push (same_succ_edge_flags[j]); |
c9e93168 | 726 | |
5deac340 | 727 | same->hashval = same_succ_hash (same); |
c9e93168 | 728 | |
c203e8a7 | 729 | slot = same_succ_htab->find_slot_with_hash (same, same->hashval, INSERT); |
c9e93168 TV |
730 | if (*slot == NULL) |
731 | { | |
732 | *slot = same; | |
733 | BB_SAME_SUCC (bb) = same; | |
734 | add_to_worklist (same); | |
735 | *same_p = NULL; | |
736 | } | |
737 | else | |
738 | { | |
739 | bitmap_set_bit ((*slot)->bbs, bb->index); | |
740 | BB_SAME_SUCC (bb) = *slot; | |
741 | add_to_worklist (*slot); | |
742 | if (inverse_flags (same, *slot)) | |
743 | bitmap_set_bit ((*slot)->inverse, bb->index); | |
744 | same_succ_reset (same); | |
745 | } | |
746 | } | |
747 | ||
748 | /* Find bbs with same successors. */ | |
749 | ||
750 | static void | |
751 | find_same_succ (void) | |
752 | { | |
753 | same_succ same = same_succ_alloc (); | |
754 | basic_block bb; | |
755 | ||
11cd3bed | 756 | FOR_EACH_BB_FN (bb, cfun) |
c9e93168 TV |
757 | { |
758 | find_same_succ_bb (bb, &same); | |
759 | if (same == NULL) | |
760 | same = same_succ_alloc (); | |
761 | } | |
762 | ||
5deac340 | 763 | same_succ_def::remove (same); |
c9e93168 TV |
764 | } |
765 | ||
766 | /* Initializes worklist administration. */ | |
767 | ||
768 | static void | |
769 | init_worklist (void) | |
770 | { | |
771 | alloc_aux_for_blocks (sizeof (struct aux_bb_info)); | |
c203e8a7 | 772 | same_succ_htab = new hash_table<same_succ_def> (n_basic_blocks_for_fn (cfun)); |
8b1c6fd7 | 773 | same_succ_edge_flags = XCNEWVEC (int, last_basic_block_for_fn (cfun)); |
c9e93168 TV |
774 | deleted_bbs = BITMAP_ALLOC (NULL); |
775 | deleted_bb_preds = BITMAP_ALLOC (NULL); | |
0cae8d31 | 776 | worklist.create (n_basic_blocks_for_fn (cfun)); |
c9e93168 TV |
777 | find_same_succ (); |
778 | ||
779 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
780 | { | |
781 | fprintf (dump_file, "initial worklist:\n"); | |
782 | print_worklist (dump_file); | |
783 | } | |
784 | } | |
785 | ||
786 | /* Deletes worklist administration. */ | |
787 | ||
788 | static void | |
789 | delete_worklist (void) | |
790 | { | |
791 | free_aux_for_blocks (); | |
c203e8a7 TS |
792 | delete same_succ_htab; |
793 | same_succ_htab = NULL; | |
c9e93168 TV |
794 | XDELETEVEC (same_succ_edge_flags); |
795 | same_succ_edge_flags = NULL; | |
796 | BITMAP_FREE (deleted_bbs); | |
797 | BITMAP_FREE (deleted_bb_preds); | |
9771b263 | 798 | worklist.release (); |
c9e93168 TV |
799 | } |
800 | ||
801 | /* Mark BB as deleted, and mark its predecessors. */ | |
802 | ||
803 | static void | |
643400b8 | 804 | mark_basic_block_deleted (basic_block bb) |
c9e93168 TV |
805 | { |
806 | edge e; | |
807 | edge_iterator ei; | |
808 | ||
809 | bitmap_set_bit (deleted_bbs, bb->index); | |
810 | ||
811 | FOR_EACH_EDGE (e, ei, bb->preds) | |
812 | bitmap_set_bit (deleted_bb_preds, e->src->index); | |
813 | } | |
814 | ||
4cbdcd40 TV |
815 | /* Removes BB from its corresponding same_succ. */ |
816 | ||
817 | static void | |
818 | same_succ_flush_bb (basic_block bb) | |
819 | { | |
820 | same_succ same = BB_SAME_SUCC (bb); | |
821 | BB_SAME_SUCC (bb) = NULL; | |
822 | if (bitmap_single_bit_set_p (same->bbs)) | |
c203e8a7 | 823 | same_succ_htab->remove_elt_with_hash (same, same->hashval); |
4cbdcd40 TV |
824 | else |
825 | bitmap_clear_bit (same->bbs, bb->index); | |
826 | } | |
827 | ||
c9e93168 TV |
828 | /* Removes all bbs in BBS from their corresponding same_succ. */ |
829 | ||
830 | static void | |
831 | same_succ_flush_bbs (bitmap bbs) | |
832 | { | |
833 | unsigned int i; | |
834 | bitmap_iterator bi; | |
835 | ||
836 | EXECUTE_IF_SET_IN_BITMAP (bbs, 0, i, bi) | |
06e28de2 | 837 | same_succ_flush_bb (BASIC_BLOCK_FOR_FN (cfun, i)); |
c9e93168 TV |
838 | } |
839 | ||
fcddd80e RG |
840 | /* Release the last vdef in BB, either normal or phi result. */ |
841 | ||
842 | static void | |
843 | release_last_vdef (basic_block bb) | |
844 | { | |
538dd0b7 DM |
845 | for (gimple_stmt_iterator i = gsi_last_bb (bb); !gsi_end_p (i); |
846 | gsi_prev_nondebug (&i)) | |
fcddd80e RG |
847 | { |
848 | gimple stmt = gsi_stmt (i); | |
849 | if (gimple_vdef (stmt) == NULL_TREE) | |
850 | continue; | |
851 | ||
852 | mark_virtual_operand_for_renaming (gimple_vdef (stmt)); | |
853 | return; | |
854 | } | |
855 | ||
538dd0b7 DM |
856 | for (gphi_iterator i = gsi_start_phis (bb); !gsi_end_p (i); |
857 | gsi_next (&i)) | |
fcddd80e | 858 | { |
538dd0b7 | 859 | gphi *phi = i.phi (); |
fcddd80e RG |
860 | tree res = gimple_phi_result (phi); |
861 | ||
ea057359 | 862 | if (!virtual_operand_p (res)) |
fcddd80e RG |
863 | continue; |
864 | ||
865 | mark_virtual_phi_result_for_renaming (phi); | |
866 | return; | |
867 | } | |
fcddd80e RG |
868 | } |
869 | ||
c9e93168 TV |
870 | /* For deleted_bb_preds, find bbs with same successors. */ |
871 | ||
872 | static void | |
873 | update_worklist (void) | |
874 | { | |
875 | unsigned int i; | |
876 | bitmap_iterator bi; | |
877 | basic_block bb; | |
878 | same_succ same; | |
879 | ||
643400b8 TV |
880 | bitmap_and_compl_into (deleted_bb_preds, deleted_bbs); |
881 | bitmap_clear (deleted_bbs); | |
882 | ||
c9e93168 TV |
883 | bitmap_clear_bit (deleted_bb_preds, ENTRY_BLOCK); |
884 | same_succ_flush_bbs (deleted_bb_preds); | |
885 | ||
886 | same = same_succ_alloc (); | |
887 | EXECUTE_IF_SET_IN_BITMAP (deleted_bb_preds, 0, i, bi) | |
888 | { | |
06e28de2 | 889 | bb = BASIC_BLOCK_FOR_FN (cfun, i); |
c9e93168 TV |
890 | gcc_assert (bb != NULL); |
891 | find_same_succ_bb (bb, &same); | |
892 | if (same == NULL) | |
893 | same = same_succ_alloc (); | |
894 | } | |
5deac340 | 895 | same_succ_def::remove (same); |
c9e93168 TV |
896 | bitmap_clear (deleted_bb_preds); |
897 | } | |
898 | ||
899 | /* Prints cluster C to FILE. */ | |
900 | ||
901 | static void | |
902 | print_cluster (FILE *file, bb_cluster c) | |
903 | { | |
904 | if (c == NULL) | |
905 | return; | |
906 | bitmap_print (file, c->bbs, "bbs:", "\n"); | |
907 | bitmap_print (file, c->preds, "preds:", "\n"); | |
908 | } | |
909 | ||
910 | /* Prints cluster C to stderr. */ | |
911 | ||
912 | extern void debug_cluster (bb_cluster); | |
913 | DEBUG_FUNCTION void | |
914 | debug_cluster (bb_cluster c) | |
915 | { | |
916 | print_cluster (stderr, c); | |
917 | } | |
918 | ||
919 | /* Update C->rep_bb, given that BB is added to the cluster. */ | |
920 | ||
921 | static void | |
922 | update_rep_bb (bb_cluster c, basic_block bb) | |
923 | { | |
924 | /* Initial. */ | |
925 | if (c->rep_bb == NULL) | |
926 | { | |
927 | c->rep_bb = bb; | |
928 | return; | |
929 | } | |
930 | ||
931 | /* Current needs no deps, keep it. */ | |
932 | if (BB_DEP_BB (c->rep_bb) == NULL) | |
933 | return; | |
934 | ||
935 | /* Bb needs no deps, change rep_bb. */ | |
936 | if (BB_DEP_BB (bb) == NULL) | |
937 | { | |
938 | c->rep_bb = bb; | |
939 | return; | |
940 | } | |
941 | ||
942 | /* Bb needs last deps earlier than current, change rep_bb. A potential | |
943 | problem with this, is that the first deps might also be earlier, which | |
944 | would mean we prefer longer lifetimes for the deps. To be able to check | |
945 | for this, we would have to trace BB_FIRST_DEP_BB as well, besides | |
946 | BB_DEP_BB, which is really BB_LAST_DEP_BB. | |
947 | The benefit of choosing the bb with last deps earlier, is that it can | |
948 | potentially be used as replacement for more bbs. */ | |
949 | if (dominated_by_p (CDI_DOMINATORS, BB_DEP_BB (c->rep_bb), BB_DEP_BB (bb))) | |
950 | c->rep_bb = bb; | |
951 | } | |
952 | ||
953 | /* Add BB to cluster C. Sets BB in C->bbs, and preds of BB in C->preds. */ | |
954 | ||
955 | static void | |
956 | add_bb_to_cluster (bb_cluster c, basic_block bb) | |
957 | { | |
958 | edge e; | |
959 | edge_iterator ei; | |
960 | ||
961 | bitmap_set_bit (c->bbs, bb->index); | |
962 | ||
963 | FOR_EACH_EDGE (e, ei, bb->preds) | |
964 | bitmap_set_bit (c->preds, e->src->index); | |
965 | ||
966 | update_rep_bb (c, bb); | |
967 | } | |
968 | ||
969 | /* Allocate and init new cluster. */ | |
970 | ||
971 | static bb_cluster | |
972 | new_cluster (void) | |
973 | { | |
974 | bb_cluster c; | |
975 | c = XCNEW (struct bb_cluster_def); | |
976 | c->bbs = BITMAP_ALLOC (NULL); | |
977 | c->preds = BITMAP_ALLOC (NULL); | |
978 | c->rep_bb = NULL; | |
979 | return c; | |
980 | } | |
981 | ||
982 | /* Delete clusters. */ | |
983 | ||
984 | static void | |
985 | delete_cluster (bb_cluster c) | |
986 | { | |
987 | if (c == NULL) | |
988 | return; | |
989 | BITMAP_FREE (c->bbs); | |
990 | BITMAP_FREE (c->preds); | |
991 | XDELETE (c); | |
992 | } | |
993 | ||
c9e93168 TV |
994 | |
995 | /* Array that contains all clusters. */ | |
996 | ||
9771b263 | 997 | static vec<bb_cluster> all_clusters; |
c9e93168 TV |
998 | |
999 | /* Allocate all cluster vectors. */ | |
1000 | ||
1001 | static void | |
1002 | alloc_cluster_vectors (void) | |
1003 | { | |
0cae8d31 | 1004 | all_clusters.create (n_basic_blocks_for_fn (cfun)); |
c9e93168 TV |
1005 | } |
1006 | ||
1007 | /* Reset all cluster vectors. */ | |
1008 | ||
1009 | static void | |
1010 | reset_cluster_vectors (void) | |
1011 | { | |
1012 | unsigned int i; | |
1013 | basic_block bb; | |
9771b263 DN |
1014 | for (i = 0; i < all_clusters.length (); ++i) |
1015 | delete_cluster (all_clusters[i]); | |
1016 | all_clusters.truncate (0); | |
11cd3bed | 1017 | FOR_EACH_BB_FN (bb, cfun) |
c9e93168 TV |
1018 | BB_CLUSTER (bb) = NULL; |
1019 | } | |
1020 | ||
1021 | /* Delete all cluster vectors. */ | |
1022 | ||
1023 | static void | |
1024 | delete_cluster_vectors (void) | |
1025 | { | |
1026 | unsigned int i; | |
9771b263 DN |
1027 | for (i = 0; i < all_clusters.length (); ++i) |
1028 | delete_cluster (all_clusters[i]); | |
1029 | all_clusters.release (); | |
c9e93168 TV |
1030 | } |
1031 | ||
1032 | /* Merge cluster C2 into C1. */ | |
1033 | ||
1034 | static void | |
1035 | merge_clusters (bb_cluster c1, bb_cluster c2) | |
1036 | { | |
1037 | bitmap_ior_into (c1->bbs, c2->bbs); | |
1038 | bitmap_ior_into (c1->preds, c2->preds); | |
1039 | } | |
1040 | ||
1041 | /* Register equivalence of BB1 and BB2 (members of cluster C). Store c in | |
1042 | all_clusters, or merge c with existing cluster. */ | |
1043 | ||
1044 | static void | |
1045 | set_cluster (basic_block bb1, basic_block bb2) | |
1046 | { | |
1047 | basic_block merge_bb, other_bb; | |
1048 | bb_cluster merge, old, c; | |
1049 | ||
1050 | if (BB_CLUSTER (bb1) == NULL && BB_CLUSTER (bb2) == NULL) | |
1051 | { | |
1052 | c = new_cluster (); | |
1053 | add_bb_to_cluster (c, bb1); | |
1054 | add_bb_to_cluster (c, bb2); | |
1055 | BB_CLUSTER (bb1) = c; | |
1056 | BB_CLUSTER (bb2) = c; | |
9771b263 DN |
1057 | c->index = all_clusters.length (); |
1058 | all_clusters.safe_push (c); | |
c9e93168 TV |
1059 | } |
1060 | else if (BB_CLUSTER (bb1) == NULL || BB_CLUSTER (bb2) == NULL) | |
1061 | { | |
1062 | merge_bb = BB_CLUSTER (bb1) == NULL ? bb2 : bb1; | |
1063 | other_bb = BB_CLUSTER (bb1) == NULL ? bb1 : bb2; | |
1064 | merge = BB_CLUSTER (merge_bb); | |
1065 | add_bb_to_cluster (merge, other_bb); | |
1066 | BB_CLUSTER (other_bb) = merge; | |
1067 | } | |
1068 | else if (BB_CLUSTER (bb1) != BB_CLUSTER (bb2)) | |
1069 | { | |
1070 | unsigned int i; | |
1071 | bitmap_iterator bi; | |
1072 | ||
1073 | old = BB_CLUSTER (bb2); | |
1074 | merge = BB_CLUSTER (bb1); | |
1075 | merge_clusters (merge, old); | |
1076 | EXECUTE_IF_SET_IN_BITMAP (old->bbs, 0, i, bi) | |
06e28de2 | 1077 | BB_CLUSTER (BASIC_BLOCK_FOR_FN (cfun, i)) = merge; |
9771b263 | 1078 | all_clusters[old->index] = NULL; |
c9e93168 TV |
1079 | update_rep_bb (merge, old->rep_bb); |
1080 | delete_cluster (old); | |
1081 | } | |
1082 | else | |
1083 | gcc_unreachable (); | |
1084 | } | |
1085 | ||
b6366520 TV |
1086 | /* Return true if gimple operands T1 and T2 have the same value. */ |
1087 | ||
1088 | static bool | |
1089 | gimple_operand_equal_value_p (tree t1, tree t2) | |
1090 | { | |
1091 | if (t1 == t2) | |
1092 | return true; | |
1093 | ||
1094 | if (t1 == NULL_TREE | |
1095 | || t2 == NULL_TREE) | |
1096 | return false; | |
1097 | ||
1098 | if (operand_equal_p (t1, t2, 0)) | |
1099 | return true; | |
1100 | ||
1101 | return gvn_uses_equal (t1, t2); | |
1102 | } | |
1103 | ||
c9e93168 TV |
1104 | /* Return true if gimple statements S1 and S2 are equal. Gimple_bb (s1) and |
1105 | gimple_bb (s2) are members of SAME_SUCC. */ | |
1106 | ||
1107 | static bool | |
1108 | gimple_equal_p (same_succ same_succ, gimple s1, gimple s2) | |
1109 | { | |
1110 | unsigned int i; | |
1111 | tree lhs1, lhs2; | |
1112 | basic_block bb1 = gimple_bb (s1), bb2 = gimple_bb (s2); | |
1113 | tree t1, t2; | |
2f1aee04 | 1114 | bool inv_cond; |
c9e93168 TV |
1115 | enum tree_code code1, code2; |
1116 | ||
1117 | if (gimple_code (s1) != gimple_code (s2)) | |
1118 | return false; | |
1119 | ||
1120 | switch (gimple_code (s1)) | |
1121 | { | |
1122 | case GIMPLE_CALL: | |
c9e93168 | 1123 | if (!gimple_call_same_target_p (s1, s2)) |
daf347dd | 1124 | return false; |
c9e93168 | 1125 | |
51d2abeb EB |
1126 | t1 = gimple_call_chain (s1); |
1127 | t2 = gimple_call_chain (s2); | |
1128 | if (!gimple_operand_equal_value_p (t1, t2)) | |
1129 | return false; | |
1130 | ||
1131 | if (gimple_call_num_args (s1) != gimple_call_num_args (s2)) | |
1132 | return false; | |
1133 | ||
c9e93168 TV |
1134 | for (i = 0; i < gimple_call_num_args (s1); ++i) |
1135 | { | |
1136 | t1 = gimple_call_arg (s1, i); | |
1137 | t2 = gimple_call_arg (s2, i); | |
51d2abeb EB |
1138 | if (!gimple_operand_equal_value_p (t1, t2)) |
1139 | return false; | |
c9e93168 | 1140 | } |
c9e93168 TV |
1141 | |
1142 | lhs1 = gimple_get_lhs (s1); | |
1143 | lhs2 = gimple_get_lhs (s2); | |
e6fa9204 JJ |
1144 | if (lhs1 == NULL_TREE && lhs2 == NULL_TREE) |
1145 | return true; | |
1146 | if (lhs1 == NULL_TREE || lhs2 == NULL_TREE) | |
1147 | return false; | |
1148 | if (TREE_CODE (lhs1) == SSA_NAME && TREE_CODE (lhs2) == SSA_NAME) | |
1149 | return vn_valueize (lhs1) == vn_valueize (lhs2); | |
1150 | return operand_equal_p (lhs1, lhs2, 0); | |
c9e93168 TV |
1151 | |
1152 | case GIMPLE_ASSIGN: | |
1153 | lhs1 = gimple_get_lhs (s1); | |
1154 | lhs2 = gimple_get_lhs (s2); | |
fcfa87ac RB |
1155 | if (TREE_CODE (lhs1) != SSA_NAME |
1156 | && TREE_CODE (lhs2) != SSA_NAME) | |
030e321a RB |
1157 | return (operand_equal_p (lhs1, lhs2, 0) |
1158 | && gimple_operand_equal_value_p (gimple_assign_rhs1 (s1), | |
1159 | gimple_assign_rhs1 (s2))); | |
fcfa87ac RB |
1160 | else if (TREE_CODE (lhs1) == SSA_NAME |
1161 | && TREE_CODE (lhs2) == SSA_NAME) | |
8024d4b6 TV |
1162 | return operand_equal_p (gimple_assign_rhs1 (s1), |
1163 | gimple_assign_rhs1 (s2), 0); | |
fcfa87ac | 1164 | return false; |
c9e93168 TV |
1165 | |
1166 | case GIMPLE_COND: | |
1167 | t1 = gimple_cond_lhs (s1); | |
1168 | t2 = gimple_cond_lhs (s2); | |
b6366520 | 1169 | if (!gimple_operand_equal_value_p (t1, t2)) |
c9e93168 TV |
1170 | return false; |
1171 | ||
1172 | t1 = gimple_cond_rhs (s1); | |
1173 | t2 = gimple_cond_rhs (s2); | |
b6366520 | 1174 | if (!gimple_operand_equal_value_p (t1, t2)) |
c9e93168 TV |
1175 | return false; |
1176 | ||
1177 | code1 = gimple_expr_code (s1); | |
1178 | code2 = gimple_expr_code (s2); | |
1179 | inv_cond = (bitmap_bit_p (same_succ->inverse, bb1->index) | |
1180 | != bitmap_bit_p (same_succ->inverse, bb2->index)); | |
1181 | if (inv_cond) | |
1182 | { | |
1b457aa4 | 1183 | bool honor_nans = HONOR_NANS (t1); |
c9e93168 TV |
1184 | code2 = invert_tree_comparison (code2, honor_nans); |
1185 | } | |
1186 | return code1 == code2; | |
1187 | ||
1188 | default: | |
1189 | return false; | |
1190 | } | |
1191 | } | |
1192 | ||
46301137 TV |
1193 | /* Let GSI skip backwards over local defs. Return the earliest vuse in VUSE. |
1194 | Return true in VUSE_ESCAPED if the vuse influenced a SSA_OP_DEF of one of the | |
1195 | processed statements. */ | |
c9e93168 TV |
1196 | |
1197 | static void | |
46301137 TV |
1198 | gsi_advance_bw_nondebug_nonlocal (gimple_stmt_iterator *gsi, tree *vuse, |
1199 | bool *vuse_escaped) | |
c9e93168 TV |
1200 | { |
1201 | gimple stmt; | |
46301137 | 1202 | tree lvuse; |
c9e93168 TV |
1203 | |
1204 | while (true) | |
1205 | { | |
1206 | if (gsi_end_p (*gsi)) | |
1207 | return; | |
1208 | stmt = gsi_stmt (*gsi); | |
46301137 TV |
1209 | |
1210 | lvuse = gimple_vuse (stmt); | |
1211 | if (lvuse != NULL_TREE) | |
1212 | { | |
1213 | *vuse = lvuse; | |
1214 | if (!ZERO_SSA_OPERANDS (stmt, SSA_OP_DEF)) | |
1215 | *vuse_escaped = true; | |
1216 | } | |
1217 | ||
b2b2f160 | 1218 | if (!stmt_local_def (stmt)) |
c9e93168 TV |
1219 | return; |
1220 | gsi_prev_nondebug (gsi); | |
1221 | } | |
1222 | } | |
1223 | ||
1224 | /* Determines whether BB1 and BB2 (members of same_succ) are duplicates. If so, | |
1225 | clusters them. */ | |
1226 | ||
1227 | static void | |
1228 | find_duplicate (same_succ same_succ, basic_block bb1, basic_block bb2) | |
1229 | { | |
1230 | gimple_stmt_iterator gsi1 = gsi_last_nondebug_bb (bb1); | |
1231 | gimple_stmt_iterator gsi2 = gsi_last_nondebug_bb (bb2); | |
46301137 TV |
1232 | tree vuse1 = NULL_TREE, vuse2 = NULL_TREE; |
1233 | bool vuse_escaped = false; | |
c9e93168 | 1234 | |
46301137 TV |
1235 | gsi_advance_bw_nondebug_nonlocal (&gsi1, &vuse1, &vuse_escaped); |
1236 | gsi_advance_bw_nondebug_nonlocal (&gsi2, &vuse2, &vuse_escaped); | |
c9e93168 TV |
1237 | |
1238 | while (!gsi_end_p (gsi1) && !gsi_end_p (gsi2)) | |
1239 | { | |
7ec88701 RH |
1240 | gimple stmt1 = gsi_stmt (gsi1); |
1241 | gimple stmt2 = gsi_stmt (gsi2); | |
1242 | ||
026c3cfd | 1243 | /* What could be better than this here is to blacklist the bb |
d5ae1c25 TV |
1244 | containing the stmt, when encountering the stmt f.i. in |
1245 | same_succ_hash. */ | |
1246 | if (is_tm_ending (stmt1) | |
1247 | || is_tm_ending (stmt2)) | |
7ec88701 RH |
1248 | return; |
1249 | ||
d5ae1c25 | 1250 | if (!gimple_equal_p (same_succ, stmt1, stmt2)) |
c9e93168 TV |
1251 | return; |
1252 | ||
1253 | gsi_prev_nondebug (&gsi1); | |
1254 | gsi_prev_nondebug (&gsi2); | |
46301137 TV |
1255 | gsi_advance_bw_nondebug_nonlocal (&gsi1, &vuse1, &vuse_escaped); |
1256 | gsi_advance_bw_nondebug_nonlocal (&gsi2, &vuse2, &vuse_escaped); | |
c9e93168 TV |
1257 | } |
1258 | ||
1259 | if (!(gsi_end_p (gsi1) && gsi_end_p (gsi2))) | |
1260 | return; | |
1261 | ||
46301137 TV |
1262 | /* If the incoming vuses are not the same, and the vuse escaped into an |
1263 | SSA_OP_DEF, then merging the 2 blocks will change the value of the def, | |
1264 | which potentially means the semantics of one of the blocks will be changed. | |
1265 | TODO: make this check more precise. */ | |
1266 | if (vuse_escaped && vuse1 != vuse2) | |
1267 | return; | |
1268 | ||
c9e93168 TV |
1269 | if (dump_file) |
1270 | fprintf (dump_file, "find_duplicates: <bb %d> duplicate of <bb %d>\n", | |
1271 | bb1->index, bb2->index); | |
1272 | ||
1273 | set_cluster (bb1, bb2); | |
1274 | } | |
1275 | ||
1276 | /* Returns whether for all phis in DEST the phi alternatives for E1 and | |
1277 | E2 are equal. */ | |
1278 | ||
1279 | static bool | |
1280 | same_phi_alternatives_1 (basic_block dest, edge e1, edge e2) | |
1281 | { | |
1282 | int n1 = e1->dest_idx, n2 = e2->dest_idx; | |
538dd0b7 | 1283 | gphi_iterator gsi; |
c9e93168 TV |
1284 | |
1285 | for (gsi = gsi_start_phis (dest); !gsi_end_p (gsi); gsi_next (&gsi)) | |
1286 | { | |
538dd0b7 | 1287 | gphi *phi = gsi.phi (); |
c9e93168 TV |
1288 | tree lhs = gimple_phi_result (phi); |
1289 | tree val1 = gimple_phi_arg_def (phi, n1); | |
1290 | tree val2 = gimple_phi_arg_def (phi, n2); | |
1291 | ||
ea057359 | 1292 | if (virtual_operand_p (lhs)) |
c9e93168 TV |
1293 | continue; |
1294 | ||
1295 | if (operand_equal_for_phi_arg_p (val1, val2)) | |
daf347dd | 1296 | continue; |
c9e93168 TV |
1297 | if (gvn_uses_equal (val1, val2)) |
1298 | continue; | |
1299 | ||
1300 | return false; | |
1301 | } | |
1302 | ||
1303 | return true; | |
1304 | } | |
1305 | ||
1306 | /* Returns whether for all successors of BB1 and BB2 (members of SAME_SUCC), the | |
1307 | phi alternatives for BB1 and BB2 are equal. */ | |
1308 | ||
1309 | static bool | |
1310 | same_phi_alternatives (same_succ same_succ, basic_block bb1, basic_block bb2) | |
1311 | { | |
1312 | unsigned int s; | |
1313 | bitmap_iterator bs; | |
1314 | edge e1, e2; | |
1315 | basic_block succ; | |
1316 | ||
1317 | EXECUTE_IF_SET_IN_BITMAP (same_succ->succs, 0, s, bs) | |
1318 | { | |
06e28de2 | 1319 | succ = BASIC_BLOCK_FOR_FN (cfun, s); |
c9e93168 TV |
1320 | e1 = find_edge (bb1, succ); |
1321 | e2 = find_edge (bb2, succ); | |
1322 | if (e1->flags & EDGE_COMPLEX | |
1323 | || e2->flags & EDGE_COMPLEX) | |
1324 | return false; | |
1325 | ||
1326 | /* For all phis in bb, the phi alternatives for e1 and e2 need to have | |
1327 | the same value. */ | |
1328 | if (!same_phi_alternatives_1 (succ, e1, e2)) | |
1329 | return false; | |
1330 | } | |
1331 | ||
1332 | return true; | |
1333 | } | |
1334 | ||
1335 | /* Return true if BB has non-vop phis. */ | |
1336 | ||
1337 | static bool | |
1338 | bb_has_non_vop_phi (basic_block bb) | |
1339 | { | |
1340 | gimple_seq phis = phi_nodes (bb); | |
1341 | gimple phi; | |
1342 | ||
1343 | if (phis == NULL) | |
1344 | return false; | |
1345 | ||
1346 | if (!gimple_seq_singleton_p (phis)) | |
1347 | return true; | |
1348 | ||
1349 | phi = gimple_seq_first_stmt (phis); | |
ea057359 | 1350 | return !virtual_operand_p (gimple_phi_result (phi)); |
c9e93168 TV |
1351 | } |
1352 | ||
1353 | /* Returns true if redirecting the incoming edges of FROM to TO maintains the | |
1354 | invariant that uses in FROM are dominates by their defs. */ | |
1355 | ||
1356 | static bool | |
1357 | deps_ok_for_redirect_from_bb_to_bb (basic_block from, basic_block to) | |
1358 | { | |
1359 | basic_block cd, dep_bb = BB_DEP_BB (to); | |
1360 | edge_iterator ei; | |
1361 | edge e; | |
1362 | bitmap from_preds = BITMAP_ALLOC (NULL); | |
1363 | ||
1364 | if (dep_bb == NULL) | |
1365 | return true; | |
1366 | ||
1367 | FOR_EACH_EDGE (e, ei, from->preds) | |
1368 | bitmap_set_bit (from_preds, e->src->index); | |
1369 | cd = nearest_common_dominator_for_set (CDI_DOMINATORS, from_preds); | |
1370 | BITMAP_FREE (from_preds); | |
1371 | ||
1372 | return dominated_by_p (CDI_DOMINATORS, dep_bb, cd); | |
1373 | } | |
1374 | ||
1375 | /* Returns true if replacing BB1 (or its replacement bb) by BB2 (or its | |
1376 | replacement bb) and vice versa maintains the invariant that uses in the | |
1377 | replacement are dominates by their defs. */ | |
1378 | ||
1379 | static bool | |
1380 | deps_ok_for_redirect (basic_block bb1, basic_block bb2) | |
1381 | { | |
1382 | if (BB_CLUSTER (bb1) != NULL) | |
1383 | bb1 = BB_CLUSTER (bb1)->rep_bb; | |
1384 | ||
1385 | if (BB_CLUSTER (bb2) != NULL) | |
1386 | bb2 = BB_CLUSTER (bb2)->rep_bb; | |
1387 | ||
1388 | return (deps_ok_for_redirect_from_bb_to_bb (bb1, bb2) | |
1389 | && deps_ok_for_redirect_from_bb_to_bb (bb2, bb1)); | |
1390 | } | |
1391 | ||
1392 | /* Within SAME_SUCC->bbs, find clusters of bbs which can be merged. */ | |
1393 | ||
1394 | static void | |
1395 | find_clusters_1 (same_succ same_succ) | |
1396 | { | |
1397 | basic_block bb1, bb2; | |
1398 | unsigned int i, j; | |
1399 | bitmap_iterator bi, bj; | |
1400 | int nr_comparisons; | |
1401 | int max_comparisons = PARAM_VALUE (PARAM_MAX_TAIL_MERGE_COMPARISONS); | |
1402 | ||
1403 | EXECUTE_IF_SET_IN_BITMAP (same_succ->bbs, 0, i, bi) | |
1404 | { | |
06e28de2 | 1405 | bb1 = BASIC_BLOCK_FOR_FN (cfun, i); |
c9e93168 TV |
1406 | |
1407 | /* TODO: handle blocks with phi-nodes. We'll have to find corresponding | |
1408 | phi-nodes in bb1 and bb2, with the same alternatives for the same | |
1409 | preds. */ | |
1410 | if (bb_has_non_vop_phi (bb1)) | |
1411 | continue; | |
1412 | ||
1413 | nr_comparisons = 0; | |
1414 | EXECUTE_IF_SET_IN_BITMAP (same_succ->bbs, i + 1, j, bj) | |
1415 | { | |
06e28de2 | 1416 | bb2 = BASIC_BLOCK_FOR_FN (cfun, j); |
c9e93168 TV |
1417 | |
1418 | if (bb_has_non_vop_phi (bb2)) | |
1419 | continue; | |
1420 | ||
1421 | if (BB_CLUSTER (bb1) != NULL && BB_CLUSTER (bb1) == BB_CLUSTER (bb2)) | |
1422 | continue; | |
1423 | ||
1424 | /* Limit quadratic behaviour. */ | |
1425 | nr_comparisons++; | |
1426 | if (nr_comparisons > max_comparisons) | |
1427 | break; | |
1428 | ||
1429 | /* This is a conservative dependency check. We could test more | |
1430 | precise for allowed replacement direction. */ | |
1431 | if (!deps_ok_for_redirect (bb1, bb2)) | |
1432 | continue; | |
1433 | ||
1434 | if (!(same_phi_alternatives (same_succ, bb1, bb2))) | |
1435 | continue; | |
1436 | ||
1437 | find_duplicate (same_succ, bb1, bb2); | |
daf347dd | 1438 | } |
c9e93168 TV |
1439 | } |
1440 | } | |
1441 | ||
1442 | /* Find clusters of bbs which can be merged. */ | |
1443 | ||
1444 | static void | |
1445 | find_clusters (void) | |
1446 | { | |
1447 | same_succ same; | |
1448 | ||
9771b263 | 1449 | while (!worklist.is_empty ()) |
c9e93168 | 1450 | { |
9771b263 | 1451 | same = worklist.pop (); |
c9e93168 TV |
1452 | same->in_worklist = false; |
1453 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1454 | { | |
1455 | fprintf (dump_file, "processing worklist entry\n"); | |
1456 | same_succ_print (dump_file, same); | |
1457 | } | |
1458 | find_clusters_1 (same); | |
1459 | } | |
1460 | } | |
1461 | ||
c9e93168 TV |
1462 | /* Returns the vop phi of BB, if any. */ |
1463 | ||
538dd0b7 | 1464 | static gphi * |
c9e93168 TV |
1465 | vop_phi (basic_block bb) |
1466 | { | |
538dd0b7 DM |
1467 | gphi *stmt; |
1468 | gphi_iterator gsi; | |
c9e93168 TV |
1469 | for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
1470 | { | |
538dd0b7 | 1471 | stmt = gsi.phi (); |
ea057359 | 1472 | if (! virtual_operand_p (gimple_phi_result (stmt))) |
c9e93168 TV |
1473 | continue; |
1474 | return stmt; | |
1475 | } | |
1476 | return NULL; | |
1477 | } | |
1478 | ||
643400b8 | 1479 | /* Redirect all edges from BB1 to BB2, removes BB1 and marks it as removed. */ |
c9e93168 TV |
1480 | |
1481 | static void | |
643400b8 | 1482 | replace_block_by (basic_block bb1, basic_block bb2) |
c9e93168 TV |
1483 | { |
1484 | edge pred_edge; | |
b6c00353 | 1485 | edge e1, e2; |
adc7a812 | 1486 | edge_iterator ei; |
c9e93168 | 1487 | unsigned int i; |
538dd0b7 | 1488 | gphi *bb2_phi; |
40f73edd | 1489 | |
643400b8 | 1490 | bb2_phi = vop_phi (bb2); |
c9e93168 | 1491 | |
643400b8 TV |
1492 | /* Mark the basic block as deleted. */ |
1493 | mark_basic_block_deleted (bb1); | |
c9e93168 TV |
1494 | |
1495 | /* Redirect the incoming edges of bb1 to bb2. */ | |
1496 | for (i = EDGE_COUNT (bb1->preds); i > 0 ; --i) | |
1497 | { | |
1498 | pred_edge = EDGE_PRED (bb1, i - 1); | |
1499 | pred_edge = redirect_edge_and_branch (pred_edge, bb2); | |
1500 | gcc_assert (pred_edge != NULL); | |
643400b8 TV |
1501 | |
1502 | if (bb2_phi == NULL) | |
1503 | continue; | |
1504 | ||
1505 | /* The phi might have run out of capacity when the redirect added an | |
1506 | argument, which means it could have been replaced. Refresh it. */ | |
1507 | bb2_phi = vop_phi (bb2); | |
1508 | ||
1509 | add_phi_arg (bb2_phi, SSA_NAME_VAR (gimple_phi_result (bb2_phi)), | |
9e227d60 | 1510 | pred_edge, UNKNOWN_LOCATION); |
c9e93168 TV |
1511 | } |
1512 | ||
fa405d7b TV |
1513 | bb2->frequency += bb1->frequency; |
1514 | if (bb2->frequency > BB_FREQ_MAX) | |
1515 | bb2->frequency = BB_FREQ_MAX; | |
55fcb901 CB |
1516 | |
1517 | bb2->count += bb1->count; | |
fa405d7b | 1518 | |
adc7a812 | 1519 | /* Merge the outgoing edge counts from bb1 onto bb2. */ |
b6c00353 | 1520 | gcov_type out_sum = 0; |
adc7a812 TJ |
1521 | FOR_EACH_EDGE (e1, ei, bb1->succs) |
1522 | { | |
adc7a812 TJ |
1523 | e2 = find_edge (bb2, e1->dest); |
1524 | gcc_assert (e2); | |
1525 | e2->count += e1->count; | |
b6c00353 TJ |
1526 | out_sum += e2->count; |
1527 | } | |
1528 | /* Recompute the edge probabilities from the new merged edge count. | |
1529 | Use the sum of the new merged edge counts computed above instead | |
1530 | of bb2's merged count, in case there are profile count insanities | |
1531 | making the bb count inconsistent with the edge weights. */ | |
1532 | FOR_EACH_EDGE (e2, ei, bb2->succs) | |
1533 | { | |
1534 | e2->probability = GCOV_COMPUTE_SCALE (e2->count, out_sum); | |
adc7a812 TJ |
1535 | } |
1536 | ||
a31895d7 | 1537 | /* Do updates that use bb1, before deleting bb1. */ |
fcddd80e | 1538 | release_last_vdef (bb1); |
a31895d7 TV |
1539 | same_succ_flush_bb (bb1); |
1540 | ||
643400b8 | 1541 | delete_basic_block (bb1); |
c9e93168 TV |
1542 | } |
1543 | ||
1544 | /* Bbs for which update_debug_stmt need to be called. */ | |
1545 | ||
1546 | static bitmap update_bbs; | |
1547 | ||
1548 | /* For each cluster in all_clusters, merge all cluster->bbs. Returns | |
643400b8 | 1549 | number of bbs removed. */ |
c9e93168 TV |
1550 | |
1551 | static int | |
643400b8 | 1552 | apply_clusters (void) |
c9e93168 TV |
1553 | { |
1554 | basic_block bb1, bb2; | |
1555 | bb_cluster c; | |
1556 | unsigned int i, j; | |
1557 | bitmap_iterator bj; | |
1558 | int nr_bbs_removed = 0; | |
1559 | ||
9771b263 | 1560 | for (i = 0; i < all_clusters.length (); ++i) |
c9e93168 | 1561 | { |
9771b263 | 1562 | c = all_clusters[i]; |
c9e93168 TV |
1563 | if (c == NULL) |
1564 | continue; | |
1565 | ||
1566 | bb2 = c->rep_bb; | |
1567 | bitmap_set_bit (update_bbs, bb2->index); | |
1568 | ||
1569 | bitmap_clear_bit (c->bbs, bb2->index); | |
1570 | EXECUTE_IF_SET_IN_BITMAP (c->bbs, 0, j, bj) | |
1571 | { | |
06e28de2 | 1572 | bb1 = BASIC_BLOCK_FOR_FN (cfun, j); |
c9e93168 TV |
1573 | bitmap_clear_bit (update_bbs, bb1->index); |
1574 | ||
643400b8 | 1575 | replace_block_by (bb1, bb2); |
c9e93168 TV |
1576 | nr_bbs_removed++; |
1577 | } | |
1578 | } | |
1579 | ||
1580 | return nr_bbs_removed; | |
1581 | } | |
1582 | ||
1583 | /* Resets debug statement STMT if it has uses that are not dominated by their | |
1584 | defs. */ | |
1585 | ||
1586 | static void | |
1587 | update_debug_stmt (gimple stmt) | |
1588 | { | |
1589 | use_operand_p use_p; | |
1590 | ssa_op_iter oi; | |
a92bf1b1 | 1591 | basic_block bbuse; |
c9e93168 TV |
1592 | |
1593 | if (!gimple_debug_bind_p (stmt)) | |
1594 | return; | |
1595 | ||
1596 | bbuse = gimple_bb (stmt); | |
1597 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, oi, SSA_OP_USE) | |
1598 | { | |
a92bf1b1 RB |
1599 | tree name = USE_FROM_PTR (use_p); |
1600 | gimple def_stmt = SSA_NAME_DEF_STMT (name); | |
1601 | basic_block bbdef = gimple_bb (def_stmt); | |
c9e93168 TV |
1602 | if (bbdef == NULL || bbuse == bbdef |
1603 | || dominated_by_p (CDI_DOMINATORS, bbuse, bbdef)) | |
1604 | continue; | |
1605 | ||
1606 | gimple_debug_bind_reset_value (stmt); | |
1607 | update_stmt (stmt); | |
a92bf1b1 | 1608 | break; |
c9e93168 TV |
1609 | } |
1610 | } | |
1611 | ||
1612 | /* Resets all debug statements that have uses that are not | |
1613 | dominated by their defs. */ | |
1614 | ||
1615 | static void | |
1616 | update_debug_stmts (void) | |
1617 | { | |
1618 | basic_block bb; | |
1619 | bitmap_iterator bi; | |
1620 | unsigned int i; | |
1621 | ||
c9e93168 TV |
1622 | EXECUTE_IF_SET_IN_BITMAP (update_bbs, 0, i, bi) |
1623 | { | |
1624 | gimple stmt; | |
1625 | gimple_stmt_iterator gsi; | |
1626 | ||
06e28de2 | 1627 | bb = BASIC_BLOCK_FOR_FN (cfun, i); |
c9e93168 TV |
1628 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
1629 | { | |
1630 | stmt = gsi_stmt (gsi); | |
1631 | if (!is_gimple_debug (stmt)) | |
1632 | continue; | |
1633 | update_debug_stmt (stmt); | |
1634 | } | |
1635 | } | |
1636 | } | |
1637 | ||
1638 | /* Runs tail merge optimization. */ | |
1639 | ||
1640 | unsigned int | |
1641 | tail_merge_optimize (unsigned int todo) | |
1642 | { | |
1643 | int nr_bbs_removed_total = 0; | |
1644 | int nr_bbs_removed; | |
1645 | bool loop_entered = false; | |
1646 | int iteration_nr = 0; | |
c9e93168 TV |
1647 | int max_iterations = PARAM_VALUE (PARAM_MAX_TAIL_MERGE_ITERATIONS); |
1648 | ||
315bbd2e | 1649 | if (!flag_tree_tail_merge |
726338f4 | 1650 | || max_iterations == 0) |
c9e93168 TV |
1651 | return 0; |
1652 | ||
1653 | timevar_push (TV_TREE_TAIL_MERGE); | |
1654 | ||
9c370032 RB |
1655 | if (!dom_info_available_p (CDI_DOMINATORS)) |
1656 | { | |
1657 | /* PRE can leave us with unreachable blocks, remove them now. */ | |
1658 | delete_unreachable_blocks (); | |
1659 | calculate_dominance_info (CDI_DOMINATORS); | |
1660 | } | |
c9e93168 TV |
1661 | init_worklist (); |
1662 | ||
9771b263 | 1663 | while (!worklist.is_empty ()) |
c9e93168 TV |
1664 | { |
1665 | if (!loop_entered) | |
1666 | { | |
1667 | loop_entered = true; | |
1668 | alloc_cluster_vectors (); | |
1669 | update_bbs = BITMAP_ALLOC (NULL); | |
1670 | } | |
1671 | else | |
1672 | reset_cluster_vectors (); | |
1673 | ||
1674 | iteration_nr++; | |
1675 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1676 | fprintf (dump_file, "worklist iteration #%d\n", iteration_nr); | |
1677 | ||
1678 | find_clusters (); | |
9771b263 DN |
1679 | gcc_assert (worklist.is_empty ()); |
1680 | if (all_clusters.is_empty ()) | |
c9e93168 TV |
1681 | break; |
1682 | ||
643400b8 | 1683 | nr_bbs_removed = apply_clusters (); |
c9e93168 TV |
1684 | nr_bbs_removed_total += nr_bbs_removed; |
1685 | if (nr_bbs_removed == 0) | |
1686 | break; | |
1687 | ||
643400b8 | 1688 | free_dominance_info (CDI_DOMINATORS); |
c9e93168 TV |
1689 | |
1690 | if (iteration_nr == max_iterations) | |
1691 | break; | |
1692 | ||
643400b8 | 1693 | calculate_dominance_info (CDI_DOMINATORS); |
c9e93168 TV |
1694 | update_worklist (); |
1695 | } | |
1696 | ||
1697 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1698 | fprintf (dump_file, "htab collision / search: %f\n", | |
c203e8a7 | 1699 | same_succ_htab->collisions ()); |
c9e93168 TV |
1700 | |
1701 | if (nr_bbs_removed_total > 0) | |
1702 | { | |
643400b8 TV |
1703 | if (MAY_HAVE_DEBUG_STMTS) |
1704 | { | |
1705 | calculate_dominance_info (CDI_DOMINATORS); | |
1706 | update_debug_stmts (); | |
1707 | } | |
c9e93168 TV |
1708 | |
1709 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1710 | { | |
1711 | fprintf (dump_file, "Before TODOs.\n"); | |
1712 | dump_function_to_file (current_function_decl, dump_file, dump_flags); | |
1713 | } | |
1714 | ||
525174a2 | 1715 | mark_virtual_operands_for_renaming (cfun); |
c9e93168 TV |
1716 | } |
1717 | ||
1718 | delete_worklist (); | |
1719 | if (loop_entered) | |
1720 | { | |
1721 | delete_cluster_vectors (); | |
1722 | BITMAP_FREE (update_bbs); | |
1723 | } | |
1724 | ||
1725 | timevar_pop (TV_TREE_TAIL_MERGE); | |
1726 | ||
1727 | return todo; | |
1728 | } |