]>
Commit | Line | Data |
---|---|---|
6de9cd9a DN |
1 | /* Dead store elimination |
2 | Copyright (C) 2004 Free Software Foundation, Inc. | |
3 | ||
4 | This file is part of GCC. | |
5 | ||
6 | GCC is free software; you can redistribute it and/or modify | |
7 | it under the terms of the GNU General Public License as published by | |
8 | the Free Software Foundation; either version 2, or (at your option) | |
9 | any later version. | |
10 | ||
11 | GCC is distributed in the hope that it will be useful, | |
12 | but WITHOUT ANY WARRANTY; without even the implied warranty of | |
13 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
14 | GNU General Public License for more details. | |
15 | ||
16 | You should have received a copy of the GNU General Public License | |
17 | along with GCC; see the file COPYING. If not, write to | |
18 | the Free Software Foundation, 59 Temple Place - Suite 330, | |
19 | Boston, MA 02111-1307, USA. */ | |
20 | ||
21 | #include "config.h" | |
22 | #include "system.h" | |
23 | #include "coretypes.h" | |
24 | #include "tm.h" | |
25 | #include "errors.h" | |
26 | #include "ggc.h" | |
27 | #include "tree.h" | |
28 | #include "rtl.h" | |
29 | #include "tm_p.h" | |
30 | #include "basic-block.h" | |
31 | #include "timevar.h" | |
32 | #include "diagnostic.h" | |
33 | #include "tree-flow.h" | |
34 | #include "tree-pass.h" | |
35 | #include "tree-dump.h" | |
36 | #include "domwalk.h" | |
37 | #include "flags.h" | |
38 | ||
39 | /* This file implements dead store elimination. | |
40 | ||
41 | A dead store is a store into a memory location which will later be | |
42 | overwritten by another store without any intervening loads. In this | |
43 | case the earlier store can be deleted. | |
44 | ||
45 | In our SSA + virtual operand world we use immediate uses of virtual | |
46 | operands to detect dead stores. If a store's virtual definition | |
47 | is used precisely once by a later store to the same location which | |
48 | post dominates the first store, then the first store is dead. | |
49 | ||
50 | The single use of the store's virtual definition ensures that | |
51 | there are no intervening aliased loads and the requirement that | |
52 | the second load post dominate the first ensures that if the earlier | |
53 | store executes, then the later stores will execute before the function | |
54 | exits. | |
55 | ||
56 | It may help to think of this as first moving the earlier store to | |
57 | the point immediately before the later store. Again, the single | |
61ada8ae | 58 | use of the virtual definition and the post-dominance relationship |
6de9cd9a DN |
59 | ensure that such movement would be safe. Clearly if there are |
60 | back to back stores, then the second is redundant. | |
61 | ||
62 | Reviewing section 10.7.2 in Morgan's "Building an Optimizing Compiler" | |
63 | may also help in understanding this code since it discusses the | |
64 | relationship between dead store and redundant load elimination. In | |
65 | fact, they are the same transformation applied to different views of | |
66 | the CFG. */ | |
67 | ||
68 | ||
69 | struct dse_global_data | |
70 | { | |
71 | /* This is the global bitmap for store statements. | |
72 | ||
73 | Each statement has a unique ID. When we encounter a store statement | |
74 | that we want to record, set the bit corresponding to the statement's | |
75 | unique ID in this bitmap. */ | |
76 | bitmap stores; | |
77 | }; | |
78 | ||
79 | /* We allocate a bitmap-per-block for stores which are encountered | |
80 | during the scan of that block. This allows us to restore the | |
81 | global bitmap of stores when we finish processing a block. */ | |
82 | struct dse_block_local_data | |
83 | { | |
84 | bitmap stores; | |
85 | }; | |
86 | ||
87 | static bool gate_dse (void); | |
88 | static void tree_ssa_dse (void); | |
89 | static void dse_initialize_block_local_data (struct dom_walk_data *, | |
90 | basic_block, | |
91 | bool); | |
92 | static void dse_optimize_stmt (struct dom_walk_data *, | |
93 | basic_block, | |
94 | block_stmt_iterator); | |
95 | static void dse_record_phis (struct dom_walk_data *, basic_block); | |
96 | static void dse_finalize_block (struct dom_walk_data *, basic_block); | |
97 | static void fix_phi_uses (tree, tree); | |
a32b97a2 | 98 | static void fix_stmt_v_may_defs (tree, tree); |
6de9cd9a DN |
99 | static void record_voperand_set (bitmap, bitmap *, unsigned int); |
100 | ||
101 | /* Function indicating whether we ought to include information for 'var' | |
102 | when calculating immediate uses. For this pass we only want use | |
103 | information for virtual variables. */ | |
104 | ||
105 | static bool | |
106 | need_imm_uses_for (tree var) | |
107 | { | |
108 | return !is_gimple_reg (var); | |
109 | } | |
110 | ||
111 | ||
a32b97a2 BB |
112 | /* Replace uses in PHI which match V_MAY_DEF_RESULTs in STMT with the |
113 | corresponding V_MAY_DEF_OP in STMT. */ | |
6de9cd9a DN |
114 | |
115 | static void | |
116 | fix_phi_uses (tree phi, tree stmt) | |
117 | { | |
118 | stmt_ann_t ann = stmt_ann (stmt); | |
a32b97a2 | 119 | v_may_def_optype v_may_defs; |
6de9cd9a DN |
120 | unsigned int i; |
121 | int j; | |
122 | ||
123 | get_stmt_operands (stmt); | |
a32b97a2 | 124 | v_may_defs = V_MAY_DEF_OPS (ann); |
6de9cd9a | 125 | |
a32b97a2 BB |
126 | /* Walk each V_MAY_DEF in STMT. */ |
127 | for (i = 0; i < NUM_V_MAY_DEFS (v_may_defs); i++) | |
6de9cd9a | 128 | { |
a32b97a2 | 129 | tree v_may_def = V_MAY_DEF_RESULT (v_may_defs, i); |
6de9cd9a | 130 | |
a32b97a2 BB |
131 | /* Find any uses in the PHI which match V_MAY_DEF and replace |
132 | them with the appropriate V_MAY_DEF_OP. */ | |
6de9cd9a | 133 | for (j = 0; j < PHI_NUM_ARGS (phi); j++) |
a32b97a2 | 134 | if (v_may_def == PHI_ARG_DEF (phi, j)) |
d00ad49b | 135 | SET_PHI_ARG_DEF (phi, j, V_MAY_DEF_OP (v_may_defs, i)); |
6de9cd9a DN |
136 | } |
137 | } | |
138 | ||
a32b97a2 BB |
139 | /* Replace the V_MAY_DEF_OPs in STMT1 which match V_MAY_DEF_RESULTs |
140 | in STMT2 with the appropriate V_MAY_DEF_OPs from STMT2. */ | |
6de9cd9a DN |
141 | |
142 | static void | |
a32b97a2 | 143 | fix_stmt_v_may_defs (tree stmt1, tree stmt2) |
6de9cd9a DN |
144 | { |
145 | stmt_ann_t ann1 = stmt_ann (stmt1); | |
146 | stmt_ann_t ann2 = stmt_ann (stmt2); | |
a32b97a2 BB |
147 | v_may_def_optype v_may_defs1; |
148 | v_may_def_optype v_may_defs2; | |
6de9cd9a DN |
149 | unsigned int i, j; |
150 | ||
151 | get_stmt_operands (stmt1); | |
152 | get_stmt_operands (stmt2); | |
a32b97a2 BB |
153 | v_may_defs1 = V_MAY_DEF_OPS (ann1); |
154 | v_may_defs2 = V_MAY_DEF_OPS (ann2); | |
6de9cd9a | 155 | |
a32b97a2 BB |
156 | /* Walk each V_MAY_DEF_OP in stmt1. */ |
157 | for (i = 0; i < NUM_V_MAY_DEFS (v_may_defs1); i++) | |
6de9cd9a | 158 | { |
a32b97a2 | 159 | tree v_may_def1 = V_MAY_DEF_OP (v_may_defs1, i); |
6de9cd9a | 160 | |
a32b97a2 BB |
161 | /* Find the appropriate V_MAY_DEF_RESULT in STMT2. */ |
162 | for (j = 0; j < NUM_V_MAY_DEFS (v_may_defs2); j++) | |
6de9cd9a | 163 | { |
a32b97a2 | 164 | if (v_may_def1 == V_MAY_DEF_RESULT (v_may_defs2, j)) |
6de9cd9a DN |
165 | { |
166 | /* Update. */ | |
d00ad49b | 167 | SET_V_MAY_DEF_OP (v_may_defs1, i, V_MAY_DEF_OP (v_may_defs2, j)); |
6de9cd9a DN |
168 | break; |
169 | } | |
170 | } | |
171 | ||
172 | #ifdef ENABLE_CHECKING | |
a32b97a2 | 173 | /* If we did not find a corresponding V_MAY_DEF_RESULT, then something |
6de9cd9a | 174 | has gone terribly wrong. */ |
a32b97a2 | 175 | if (j == NUM_V_MAY_DEFS (v_may_defs2)) |
6de9cd9a DN |
176 | abort (); |
177 | #endif | |
178 | ||
179 | } | |
180 | } | |
181 | ||
182 | ||
183 | /* Set bit UID in bitmaps GLOBAL and *LOCAL, creating *LOCAL as needed. */ | |
184 | static void | |
185 | record_voperand_set (bitmap global, bitmap *local, unsigned int uid) | |
186 | { | |
187 | /* Lazily allocate the bitmap. Note that we do not get a notification | |
188 | when the block local data structures die, so we allocate the local | |
189 | bitmap backed by the GC system. */ | |
190 | if (*local == NULL) | |
191 | *local = BITMAP_GGC_ALLOC (); | |
192 | ||
193 | /* Set the bit in the local and global bitmaps. */ | |
194 | bitmap_set_bit (*local, uid); | |
195 | bitmap_set_bit (global, uid); | |
196 | } | |
197 | /* Initialize block local data structures. */ | |
198 | ||
199 | static void | |
200 | dse_initialize_block_local_data (struct dom_walk_data *walk_data, | |
201 | basic_block bb ATTRIBUTE_UNUSED, | |
202 | bool recycled) | |
203 | { | |
204 | struct dse_block_local_data *bd | |
205 | = VARRAY_TOP_GENERIC_PTR (walk_data->block_data_stack); | |
206 | ||
207 | /* If we are given a recycled block local data structure, ensure any | |
208 | bitmap associated with the block is cleared. */ | |
209 | if (recycled) | |
210 | { | |
211 | if (bd->stores) | |
212 | bitmap_clear (bd->stores); | |
213 | } | |
214 | } | |
215 | ||
216 | /* Attempt to eliminate dead stores in the statement referenced by BSI. | |
217 | ||
218 | A dead store is a store into a memory location which will later be | |
219 | overwritten by another store without any intervening loads. In this | |
220 | case the earlier store can be deleted. | |
221 | ||
222 | In our SSA + virtual operand world we use immediate uses of virtual | |
223 | operands to detect dead stores. If a store's virtual definition | |
224 | is used precisely once by a later store to the same location which | |
225 | post dominates the first store, then the first store is dead. */ | |
226 | ||
227 | static void | |
228 | dse_optimize_stmt (struct dom_walk_data *walk_data, | |
229 | basic_block bb ATTRIBUTE_UNUSED, | |
230 | block_stmt_iterator bsi) | |
231 | { | |
232 | struct dse_block_local_data *bd | |
233 | = VARRAY_TOP_GENERIC_PTR (walk_data->block_data_stack); | |
234 | struct dse_global_data *dse_gd = walk_data->global_data; | |
235 | tree stmt = bsi_stmt (bsi); | |
236 | stmt_ann_t ann = stmt_ann (stmt); | |
a32b97a2 | 237 | v_may_def_optype v_may_defs; |
6de9cd9a DN |
238 | |
239 | get_stmt_operands (stmt); | |
a32b97a2 | 240 | v_may_defs = V_MAY_DEF_OPS (ann); |
6de9cd9a DN |
241 | |
242 | /* If this statement has no virtual uses, then there is nothing | |
243 | to do. */ | |
a32b97a2 | 244 | if (NUM_V_MAY_DEFS (v_may_defs) == 0) |
6de9cd9a DN |
245 | return; |
246 | ||
cd709752 RH |
247 | /* We know we have virtual definitions. If this is a MODIFY_EXPR that's |
248 | not also a function call, then record it into our table. */ | |
249 | if (get_call_expr_in (stmt)) | |
250 | return; | |
251 | if (TREE_CODE (stmt) == MODIFY_EXPR) | |
6de9cd9a DN |
252 | { |
253 | dataflow_t df = get_immediate_uses (stmt); | |
254 | unsigned int num_uses = num_immediate_uses (df); | |
255 | tree use; | |
256 | tree skipped_phi; | |
257 | ||
258 | ||
259 | /* If there are no uses then there is nothing left to do. */ | |
260 | if (num_uses == 0) | |
261 | { | |
262 | record_voperand_set (dse_gd->stores, &bd->stores, ann->uid); | |
263 | return; | |
264 | } | |
265 | ||
266 | use = immediate_use (df, 0); | |
267 | skipped_phi = NULL; | |
268 | ||
269 | /* Skip through any PHI nodes we have already seen if the PHI | |
270 | represents the only use of this store. | |
271 | ||
272 | Note this does not handle the case where the store has | |
a32b97a2 | 273 | multiple V_MAY_DEFs which all reach a set of PHI nodes in the |
6de9cd9a DN |
274 | same block. */ |
275 | while (num_uses == 1 | |
276 | && TREE_CODE (use) == PHI_NODE | |
277 | && bitmap_bit_p (dse_gd->stores, stmt_ann (use)->uid)) | |
278 | { | |
279 | /* Record the first PHI we skip so that we can fix its | |
280 | uses if we find that STMT is a dead store. */ | |
281 | if (!skipped_phi) | |
282 | skipped_phi = use; | |
283 | ||
284 | /* Skip past this PHI and loop again in case we had a PHI | |
285 | chain. */ | |
286 | df = get_immediate_uses (use); | |
287 | num_uses = num_immediate_uses (df); | |
288 | use = immediate_use (df, 0); | |
289 | } | |
290 | ||
291 | /* If we have precisely one immediate use at this point, then we may | |
292 | have found redundant store. */ | |
293 | if (num_uses == 1 | |
294 | && bitmap_bit_p (dse_gd->stores, stmt_ann (use)->uid) | |
295 | && operand_equal_p (TREE_OPERAND (stmt, 0), | |
296 | TREE_OPERAND (use, 0), 0)) | |
297 | { | |
298 | /* We need to fix the operands if either the first PHI we | |
299 | skipped, or the store which we are not deleting if we did | |
300 | not skip any PHIs. */ | |
301 | if (skipped_phi) | |
302 | fix_phi_uses (skipped_phi, stmt); | |
303 | else | |
a32b97a2 | 304 | fix_stmt_v_may_defs (use, stmt); |
6de9cd9a DN |
305 | |
306 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
307 | { | |
308 | fprintf (dump_file, " Deleted dead store '"); | |
309 | print_generic_expr (dump_file, bsi_stmt (bsi), dump_flags); | |
310 | fprintf (dump_file, "'\n"); | |
311 | } | |
312 | ||
313 | /* Any immediate uses which reference STMT need to instead | |
314 | reference the new consumer, either SKIPPED_PHI or USE. | |
315 | This allows us to cascade dead stores. */ | |
316 | redirect_immediate_uses (stmt, skipped_phi ? skipped_phi : use); | |
317 | ||
318 | /* Finally remove the dead store. */ | |
319 | bsi_remove (&bsi); | |
320 | } | |
321 | ||
322 | record_voperand_set (dse_gd->stores, &bd->stores, ann->uid); | |
323 | } | |
324 | } | |
325 | ||
326 | /* Record that we have seen the PHIs at the start of BB which correspond | |
327 | to virtual operands. */ | |
328 | static void | |
329 | dse_record_phis (struct dom_walk_data *walk_data, basic_block bb) | |
330 | { | |
331 | struct dse_block_local_data *bd | |
332 | = VARRAY_TOP_GENERIC_PTR (walk_data->block_data_stack); | |
333 | struct dse_global_data *dse_gd = walk_data->global_data; | |
334 | tree phi; | |
335 | ||
17192884 | 336 | for (phi = phi_nodes (bb); phi; phi = PHI_CHAIN (phi)) |
6de9cd9a DN |
337 | if (need_imm_uses_for (PHI_RESULT (phi))) |
338 | record_voperand_set (dse_gd->stores, | |
339 | &bd->stores, | |
340 | get_stmt_ann (phi)->uid); | |
341 | } | |
342 | ||
343 | static void | |
344 | dse_finalize_block (struct dom_walk_data *walk_data, | |
345 | basic_block bb ATTRIBUTE_UNUSED) | |
346 | { | |
347 | struct dse_block_local_data *bd | |
348 | = VARRAY_TOP_GENERIC_PTR (walk_data->block_data_stack); | |
349 | struct dse_global_data *dse_gd = walk_data->global_data; | |
350 | bitmap stores = dse_gd->stores; | |
351 | unsigned int i; | |
352 | ||
353 | /* Unwind the stores noted in this basic block. */ | |
354 | if (bd->stores) | |
355 | EXECUTE_IF_SET_IN_BITMAP (bd->stores, 0, i, bitmap_clear_bit (stores, i);); | |
356 | } | |
357 | ||
358 | static void | |
359 | tree_ssa_dse (void) | |
360 | { | |
361 | struct dom_walk_data walk_data; | |
362 | struct dse_global_data dse_gd; | |
363 | unsigned int uid = 0; | |
364 | basic_block bb; | |
365 | ||
366 | /* Create a UID for each statement in the function. Ordering of the | |
367 | UIDs is not important for this pass. */ | |
368 | FOR_EACH_BB (bb) | |
369 | { | |
370 | block_stmt_iterator bsi; | |
371 | tree phi; | |
372 | ||
373 | for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi)) | |
374 | stmt_ann (bsi_stmt (bsi))->uid = uid++; | |
375 | ||
17192884 | 376 | for (phi = phi_nodes (bb); phi; phi = PHI_CHAIN (phi)) |
6de9cd9a DN |
377 | stmt_ann (phi)->uid = uid++; |
378 | } | |
379 | ||
380 | /* We might consider making this a property of each pass so that it | |
381 | can be [re]computed on an as-needed basis. Particularly since | |
382 | this pass could be seen as an extension of DCE which needs post | |
383 | dominators. */ | |
384 | calculate_dominance_info (CDI_POST_DOMINATORS); | |
385 | ||
386 | /* We also need immediate use information for virtual operands. */ | |
387 | compute_immediate_uses (TDFA_USE_VOPS, need_imm_uses_for); | |
388 | ||
389 | /* Dead store elimination is fundamentally a walk of the post-dominator | |
390 | tree and a backwards walk of statements within each block. */ | |
391 | walk_data.walk_stmts_backward = true; | |
392 | walk_data.dom_direction = CDI_POST_DOMINATORS; | |
393 | walk_data.initialize_block_local_data = dse_initialize_block_local_data; | |
394 | walk_data.before_dom_children_before_stmts = NULL; | |
395 | walk_data.before_dom_children_walk_stmts = dse_optimize_stmt; | |
396 | walk_data.before_dom_children_after_stmts = dse_record_phis; | |
397 | walk_data.after_dom_children_before_stmts = NULL; | |
398 | walk_data.after_dom_children_walk_stmts = NULL; | |
399 | walk_data.after_dom_children_after_stmts = dse_finalize_block; | |
400 | ||
401 | walk_data.block_local_data_size = sizeof (struct dse_block_local_data); | |
402 | ||
403 | /* This is the main hash table for the dead store elimination pass. */ | |
404 | dse_gd.stores = BITMAP_XMALLOC (); | |
405 | walk_data.global_data = &dse_gd; | |
406 | ||
407 | /* Initialize the dominator walker. */ | |
408 | init_walk_dominator_tree (&walk_data); | |
409 | ||
410 | /* Recursively walk the dominator tree. */ | |
411 | walk_dominator_tree (&walk_data, EXIT_BLOCK_PTR); | |
412 | ||
413 | /* Finalize the dominator walker. */ | |
414 | fini_walk_dominator_tree (&walk_data); | |
415 | ||
416 | /* Release the main bitmap. */ | |
417 | BITMAP_XFREE (dse_gd.stores); | |
418 | ||
419 | /* Free dataflow information. It's probably out of date now anyway. */ | |
420 | free_df (); | |
421 | ||
422 | /* For now, just wipe the post-dominator information. */ | |
423 | free_dominance_info (CDI_POST_DOMINATORS); | |
424 | } | |
425 | ||
426 | static bool | |
427 | gate_dse (void) | |
428 | { | |
429 | return flag_tree_dse != 0; | |
430 | } | |
431 | ||
432 | struct tree_opt_pass pass_dse = { | |
433 | "dse", /* name */ | |
434 | gate_dse, /* gate */ | |
435 | tree_ssa_dse, /* execute */ | |
436 | NULL, /* sub */ | |
437 | NULL, /* next */ | |
438 | 0, /* static_pass_number */ | |
439 | TV_TREE_DSE, /* tv_id */ | |
c1b763fa DN |
440 | PROP_cfg | PROP_ssa |
441 | | PROP_alias, /* properties_required */ | |
6de9cd9a DN |
442 | 0, /* properties_provided */ |
443 | 0, /* properties_destroyed */ | |
444 | 0, /* todo_flags_start */ | |
445 | TODO_dump_func | TODO_ggc_collect /* todo_flags_finish */ | |
9f8628ba PB |
446 | | TODO_verify_ssa, |
447 | 0 /* letter */ | |
6de9cd9a | 448 | }; |