1 /* Dead store elimination
2 Copyright (C) 2004, 2005, 2006, 2007 Free Software Foundation, Inc.
4 This file is part of GCC.
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 3, or (at your option)
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.
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
22 #include "coretypes.h"
28 #include "basic-block.h"
30 #include "diagnostic.h"
31 #include "tree-flow.h"
32 #include "tree-pass.h"
33 #include "tree-dump.h"
37 /* This file implements dead store elimination.
39 A dead store is a store into a memory location which will later be
40 overwritten by another store without any intervening loads. In this
41 case the earlier store can be deleted.
43 In our SSA + virtual operand world we use immediate uses of virtual
44 operands to detect dead stores. If a store's virtual definition
45 is used precisely once by a later store to the same location which
46 post dominates the first store, then the first store is dead.
48 The single use of the store's virtual definition ensures that
49 there are no intervening aliased loads and the requirement that
50 the second load post dominate the first ensures that if the earlier
51 store executes, then the later stores will execute before the function
54 It may help to think of this as first moving the earlier store to
55 the point immediately before the later store. Again, the single
56 use of the virtual definition and the post-dominance relationship
57 ensure that such movement would be safe. Clearly if there are
58 back to back stores, then the second is redundant.
60 Reviewing section 10.7.2 in Morgan's "Building an Optimizing Compiler"
61 may also help in understanding this code since it discusses the
62 relationship between dead store and redundant load elimination. In
63 fact, they are the same transformation applied to different views of
67 struct dse_global_data
69 /* This is the global bitmap for store statements.
71 Each statement has a unique ID. When we encounter a store statement
72 that we want to record, set the bit corresponding to the statement's
73 unique ID in this bitmap. */
77 /* We allocate a bitmap-per-block for stores which are encountered
78 during the scan of that block. This allows us to restore the
79 global bitmap of stores when we finish processing a block. */
80 struct dse_block_local_data
85 /* Basic blocks of the potentially dead store and the following
86 store, for memory_address_same. */
87 struct address_walk_data
89 basic_block store1_bb, store2_bb;
92 static bool gate_dse (void);
93 static unsigned int tree_ssa_dse (void);
94 static void dse_initialize_block_local_data (struct dom_walk_data *,
97 static void dse_optimize_stmt (struct dom_walk_data *,
100 static void dse_record_phis (struct dom_walk_data *, basic_block);
101 static void dse_finalize_block (struct dom_walk_data *, basic_block);
102 static void record_voperand_set (bitmap, bitmap *, unsigned int);
104 static unsigned max_stmt_uid; /* Maximal uid of a statement. Uids to phi
105 nodes are assigned using the versions of
106 ssa names they define. */
108 /* Returns uid of statement STMT. */
111 get_stmt_uid (tree stmt)
113 if (TREE_CODE (stmt) == PHI_NODE)
114 return SSA_NAME_VERSION (PHI_RESULT (stmt)) + max_stmt_uid;
116 return stmt_ann (stmt)->uid;
119 /* Set bit UID in bitmaps GLOBAL and *LOCAL, creating *LOCAL as needed. */
122 record_voperand_set (bitmap global, bitmap *local, unsigned int uid)
124 /* Lazily allocate the bitmap. Note that we do not get a notification
125 when the block local data structures die, so we allocate the local
126 bitmap backed by the GC system. */
128 *local = BITMAP_GGC_ALLOC ();
130 /* Set the bit in the local and global bitmaps. */
131 bitmap_set_bit (*local, uid);
132 bitmap_set_bit (global, uid);
135 /* Initialize block local data structures. */
138 dse_initialize_block_local_data (struct dom_walk_data *walk_data,
139 basic_block bb ATTRIBUTE_UNUSED,
142 struct dse_block_local_data *bd
143 = (struct dse_block_local_data *)
144 VEC_last (void_p, walk_data->block_data_stack);
146 /* If we are given a recycled block local data structure, ensure any
147 bitmap associated with the block is cleared. */
151 bitmap_clear (bd->stores);
155 /* Helper function for memory_address_same via walk_tree. Returns
156 non-NULL if it finds an SSA_NAME which is part of the address,
157 such that the definition of the SSA_NAME post-dominates the store
158 we want to delete but not the store that we believe makes it
159 redundant. This indicates that the address may change between
163 memory_ssa_name_same (tree *expr_p, int *walk_subtrees ATTRIBUTE_UNUSED,
166 struct address_walk_data *walk_data = (struct address_walk_data *) data;
171 if (TREE_CODE (expr) != SSA_NAME)
174 /* If we've found a default definition, then there's no problem. Both
175 stores will post-dominate it. And def_bb will be NULL. */
176 if (SSA_NAME_IS_DEFAULT_DEF (expr))
179 def_stmt = SSA_NAME_DEF_STMT (expr);
180 def_bb = bb_for_stmt (def_stmt);
182 /* DEF_STMT must dominate both stores. So if it is in the same
183 basic block as one, it does not post-dominate that store. */
184 if (walk_data->store1_bb != def_bb
185 && dominated_by_p (CDI_POST_DOMINATORS, walk_data->store1_bb, def_bb))
187 if (walk_data->store2_bb == def_bb
188 || !dominated_by_p (CDI_POST_DOMINATORS, walk_data->store2_bb,
190 /* Return non-NULL to stop the walk. */
197 /* Return TRUE if the destination memory address in STORE1 and STORE2
198 might be modified after STORE1, before control reaches STORE2. */
201 memory_address_same (tree store1, tree store2)
203 struct address_walk_data walk_data;
205 walk_data.store1_bb = bb_for_stmt (store1);
206 walk_data.store2_bb = bb_for_stmt (store2);
208 return (walk_tree (&GIMPLE_STMT_OPERAND (store1, 0), memory_ssa_name_same,
213 /* Return the use stmt for the lhs of STMT following the virtual
214 def-use chains. Returns the MODIFY_EXPR stmt which lhs is equal to
215 the lhs of STMT or NULL_TREE if no such stmt can be found. */
217 get_use_of_stmt_lhs (tree stmt,
218 use_operand_p * first_use_p,
219 use_operand_p * use_p, tree * use_stmt)
224 if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
227 lhs = GIMPLE_STMT_OPERAND (stmt, 0);
229 /* The stmt must have a single VDEF. */
230 def_p = SINGLE_SSA_DEF_OPERAND (stmt, SSA_OP_VDEF);
231 if (def_p == NULL_DEF_OPERAND_P)
234 if (!has_single_use (DEF_FROM_PTR (def_p)))
236 /* Get the immediate use of the def. */
237 single_imm_use (DEF_FROM_PTR (def_p), use_p, use_stmt);
238 gcc_assert (*use_p != NULL_USE_OPERAND_P);
241 /* If the use is not simple, give up. */
242 if (TREE_CODE (*use_stmt) != GIMPLE_MODIFY_STMT
243 || get_call_expr_in (*use_stmt))
248 /* Look at the use stmt and see if it's LHS matches
249 stmt's lhs SSA_NAME. */
250 def_p = SINGLE_SSA_DEF_OPERAND (*use_stmt, SSA_OP_VDEF);
251 if (def_p == NULL_DEF_OPERAND_P)
254 usevar = GIMPLE_STMT_OPERAND (*use_stmt, 0);
255 if (operand_equal_p (usevar, lhs, 0))
258 if (!has_single_use (DEF_FROM_PTR (def_p)))
260 single_imm_use (DEF_FROM_PTR (def_p), use_p, use_stmt);
261 gcc_assert (*use_p != NULL_USE_OPERAND_P);
262 if (TREE_CODE (*use_stmt) != GIMPLE_MODIFY_STMT
263 || get_call_expr_in (*use_stmt))
271 /* A helper of dse_optimize_stmt.
272 Given a GIMPLE_MODIFY_STMT in STMT, check that each VDEF has one
273 use, and that one use is another VDEF clobbering the first one.
275 Return TRUE if the above conditions are met, otherwise FALSE. */
278 dse_possible_dead_store_p (tree stmt,
279 use_operand_p *first_use_p,
280 use_operand_p *use_p,
282 struct dse_global_data *dse_gd,
283 struct dse_block_local_data *bd)
289 tree defvar = NULL_TREE, temp;
290 tree prev_defvar = NULL_TREE;
291 stmt_ann_t ann = stmt_ann (stmt);
293 /* We want to verify that each virtual definition in STMT has
294 precisely one use and that all the virtual definitions are
295 used by the same single statement. When complete, we
296 want USE_STMT to refer to the one statement which uses
297 all of the virtual definitions from STMT. */
299 FOR_EACH_SSA_VDEF_OPERAND (var1, vv, stmt, op_iter)
301 defvar = DEF_FROM_PTR (var1);
303 /* If this virtual def does not have precisely one use, then
304 we will not be able to eliminate STMT. */
305 if (!has_single_use (defvar))
311 /* Get the one and only immediate use of DEFVAR. */
312 single_imm_use (defvar, use_p, &temp);
313 gcc_assert (*use_p != NULL_USE_OPERAND_P);
314 *first_use_p = *use_p;
316 /* In the case of memory partitions, we may get:
318 # MPT.764_162 = VDEF <MPT.764_161(D)>
320 # MPT.764_167 = VDEF <MPT.764_162>
323 So we must make sure we're talking about the same LHS.
325 if (TREE_CODE (temp) == GIMPLE_MODIFY_STMT)
327 tree base1 = get_base_address (GIMPLE_STMT_OPERAND (stmt, 0));
328 tree base2 = get_base_address (GIMPLE_STMT_OPERAND (temp, 0));
330 while (base1 && INDIRECT_REF_P (base1))
331 base1 = TREE_OPERAND (base1, 0);
332 while (base2 && INDIRECT_REF_P (base2))
333 base2 = TREE_OPERAND (base2, 0);
342 /* If the immediate use of DEF_VAR is not the same as the
343 previously find immediate uses, then we will not be able
344 to eliminate STMT. */
345 if (*use_stmt == NULL)
348 prev_defvar = defvar;
350 else if (temp != *use_stmt)
359 record_voperand_set (dse_gd->stores, &bd->stores, ann->uid);
363 /* Skip through any PHI nodes we have already seen if the PHI
364 represents the only use of this store.
366 Note this does not handle the case where the store has
367 multiple VDEFs which all reach a set of PHI nodes in the same block. */
368 while (*use_p != NULL_USE_OPERAND_P
369 && TREE_CODE (*use_stmt) == PHI_NODE
370 && bitmap_bit_p (dse_gd->stores, get_stmt_uid (*use_stmt)))
372 /* A PHI node can both define and use the same SSA_NAME if
373 the PHI is at the top of a loop and the PHI_RESULT is
374 a loop invariant and copies have not been fully propagated.
376 The safe thing to do is exit assuming no optimization is
378 if (SSA_NAME_DEF_STMT (PHI_RESULT (*use_stmt)) == *use_stmt)
381 /* Skip past this PHI and loop again in case we had a PHI
383 single_imm_use (PHI_RESULT (*use_stmt), use_p, use_stmt);
390 /* Attempt to eliminate dead stores in the statement referenced by BSI.
392 A dead store is a store into a memory location which will later be
393 overwritten by another store without any intervening loads. In this
394 case the earlier store can be deleted.
396 In our SSA + virtual operand world we use immediate uses of virtual
397 operands to detect dead stores. If a store's virtual definition
398 is used precisely once by a later store to the same location which
399 post dominates the first store, then the first store is dead. */
402 dse_optimize_stmt (struct dom_walk_data *walk_data,
403 basic_block bb ATTRIBUTE_UNUSED,
404 block_stmt_iterator bsi)
406 struct dse_block_local_data *bd
407 = (struct dse_block_local_data *)
408 VEC_last (void_p, walk_data->block_data_stack);
409 struct dse_global_data *dse_gd
410 = (struct dse_global_data *) walk_data->global_data;
411 tree stmt = bsi_stmt (bsi);
412 stmt_ann_t ann = stmt_ann (stmt);
414 /* If this statement has no virtual defs, then there is nothing
416 if (ZERO_SSA_OPERANDS (stmt, SSA_OP_VDEF))
419 /* We know we have virtual definitions. If this is a GIMPLE_MODIFY_STMT
420 that's not also a function call, then record it into our table. */
421 if (get_call_expr_in (stmt))
424 if (ann->has_volatile_ops)
427 if (TREE_CODE (stmt) == GIMPLE_MODIFY_STMT)
429 use_operand_p first_use_p = NULL_USE_OPERAND_P;
430 use_operand_p use_p = NULL;
433 if (!dse_possible_dead_store_p (stmt, &first_use_p, &use_p, &use_stmt,
437 /* If we have precisely one immediate use at this point, then we may
438 have found redundant store. Make sure that the stores are to
439 the same memory location. This includes checking that any
440 SSA-form variables in the address will have the same values. */
441 if (use_p != NULL_USE_OPERAND_P
442 && bitmap_bit_p (dse_gd->stores, get_stmt_uid (use_stmt))
443 && !operand_equal_p (GIMPLE_STMT_OPERAND (stmt, 0),
444 GIMPLE_STMT_OPERAND (use_stmt, 0), 0)
445 && memory_address_same (stmt, use_stmt))
447 /* If we have precisely one immediate use at this point, but
448 the stores are not to the same memory location then walk the
449 virtual def-use chain to get the stmt which stores to that same
451 if (get_use_of_stmt_lhs (stmt, &first_use_p, &use_p, &use_stmt) ==
454 record_voperand_set (dse_gd->stores, &bd->stores, ann->uid);
459 /* If we have precisely one immediate use at this point and the
460 stores are to the same memory location or there is a chain of
461 virtual uses from stmt and the stmt which stores to that same
462 memory location, then we may have found redundant store. */
463 if (use_p != NULL_USE_OPERAND_P
464 && bitmap_bit_p (dse_gd->stores, get_stmt_uid (use_stmt))
465 && operand_equal_p (GIMPLE_STMT_OPERAND (stmt, 0),
466 GIMPLE_STMT_OPERAND (use_stmt, 0), 0)
467 && memory_address_same (stmt, use_stmt))
474 if (dump_file && (dump_flags & TDF_DETAILS))
476 fprintf (dump_file, " Deleted dead store '");
477 print_generic_expr (dump_file, bsi_stmt (bsi), dump_flags);
478 fprintf (dump_file, "'\n");
481 /* Then we need to fix the operand of the consuming stmt. */
482 stmt_lhs = USE_FROM_PTR (first_use_p);
483 FOR_EACH_SSA_VDEF_OPERAND (var1, vv, stmt, op_iter)
487 single_imm_use (DEF_FROM_PTR (var1), &use_p, &temp);
488 gcc_assert (VUSE_VECT_NUM_ELEM (*vv) == 1);
489 usevar = VUSE_ELEMENT_VAR (*vv, 0);
490 SET_USE (use_p, usevar);
492 /* Make sure we propagate the ABNORMAL bit setting. */
493 if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (stmt_lhs))
494 SSA_NAME_OCCURS_IN_ABNORMAL_PHI (usevar) = 1;
497 /* Remove the dead store. */
498 bsi_remove (&bsi, true);
500 /* And release any SSA_NAMEs set in this statement back to the
505 record_voperand_set (dse_gd->stores, &bd->stores, ann->uid);
509 /* Record that we have seen the PHIs at the start of BB which correspond
510 to virtual operands. */
512 dse_record_phis (struct dom_walk_data *walk_data, basic_block bb)
514 struct dse_block_local_data *bd
515 = (struct dse_block_local_data *)
516 VEC_last (void_p, walk_data->block_data_stack);
517 struct dse_global_data *dse_gd
518 = (struct dse_global_data *) walk_data->global_data;
521 for (phi = phi_nodes (bb); phi; phi = PHI_CHAIN (phi))
522 if (!is_gimple_reg (PHI_RESULT (phi)))
523 record_voperand_set (dse_gd->stores,
529 dse_finalize_block (struct dom_walk_data *walk_data,
530 basic_block bb ATTRIBUTE_UNUSED)
532 struct dse_block_local_data *bd
533 = (struct dse_block_local_data *)
534 VEC_last (void_p, walk_data->block_data_stack);
535 struct dse_global_data *dse_gd
536 = (struct dse_global_data *) walk_data->global_data;
537 bitmap stores = dse_gd->stores;
541 /* Unwind the stores noted in this basic block. */
543 EXECUTE_IF_SET_IN_BITMAP (bd->stores, 0, i, bi)
545 bitmap_clear_bit (stores, i);
549 /* Main entry point. */
554 struct dom_walk_data walk_data;
555 struct dse_global_data dse_gd;
558 /* Create a UID for each statement in the function. Ordering of the
559 UIDs is not important for this pass. */
563 block_stmt_iterator bsi;
565 for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
566 stmt_ann (bsi_stmt (bsi))->uid = max_stmt_uid++;
569 /* We might consider making this a property of each pass so that it
570 can be [re]computed on an as-needed basis. Particularly since
571 this pass could be seen as an extension of DCE which needs post
573 calculate_dominance_info (CDI_POST_DOMINATORS);
575 /* Dead store elimination is fundamentally a walk of the post-dominator
576 tree and a backwards walk of statements within each block. */
577 walk_data.walk_stmts_backward = true;
578 walk_data.dom_direction = CDI_POST_DOMINATORS;
579 walk_data.initialize_block_local_data = dse_initialize_block_local_data;
580 walk_data.before_dom_children_before_stmts = NULL;
581 walk_data.before_dom_children_walk_stmts = dse_optimize_stmt;
582 walk_data.before_dom_children_after_stmts = dse_record_phis;
583 walk_data.after_dom_children_before_stmts = NULL;
584 walk_data.after_dom_children_walk_stmts = NULL;
585 walk_data.after_dom_children_after_stmts = dse_finalize_block;
586 walk_data.interesting_blocks = NULL;
588 walk_data.block_local_data_size = sizeof (struct dse_block_local_data);
590 /* This is the main hash table for the dead store elimination pass. */
591 dse_gd.stores = BITMAP_ALLOC (NULL);
592 walk_data.global_data = &dse_gd;
594 /* Initialize the dominator walker. */
595 init_walk_dominator_tree (&walk_data);
597 /* Recursively walk the dominator tree. */
598 walk_dominator_tree (&walk_data, EXIT_BLOCK_PTR);
600 /* Finalize the dominator walker. */
601 fini_walk_dominator_tree (&walk_data);
603 /* Release the main bitmap. */
604 BITMAP_FREE (dse_gd.stores);
606 /* For now, just wipe the post-dominator information. */
607 free_dominance_info (CDI_POST_DOMINATORS);
614 return flag_tree_dse != 0;
617 struct tree_opt_pass pass_dse = {
620 tree_ssa_dse, /* execute */
623 0, /* static_pass_number */
624 TV_TREE_DSE, /* tv_id */
627 | PROP_alias, /* properties_required */
628 0, /* properties_provided */
629 0, /* properties_destroyed */
630 0, /* todo_flags_start */
633 | TODO_verify_ssa, /* todo_flags_finish */
637 /* A very simple dead store pass eliminating write only local variables.
638 The pass does not require alias information and thus can be run before
639 inlining to quickly eliminate artifacts of some common C++ constructs. */
642 execute_simple_dse (void)
644 block_stmt_iterator bsi;
646 bitmap variables_loaded = BITMAP_ALLOC (NULL);
647 unsigned int todo = 0;
649 /* Collect into VARIABLES LOADED all variables that are read in function
652 for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
653 if (LOADED_SYMS (bsi_stmt (bsi)))
654 bitmap_ior_into (variables_loaded,
655 LOADED_SYMS (bsi_stmt (bsi)));
657 /* Look for statements writting into the write only variables.
658 And try to remove them. */
661 for (bsi = bsi_start (bb); !bsi_end_p (bsi);)
663 tree stmt = bsi_stmt (bsi), op;
664 bool removed = false;
667 if (STORED_SYMS (stmt) && TREE_CODE (stmt) == GIMPLE_MODIFY_STMT
668 && TREE_CODE (stmt) != RETURN_EXPR
669 && !bitmap_intersect_p (STORED_SYMS (stmt), variables_loaded))
677 /* See if STMT only stores to write-only variables and
678 verify that there are no volatile operands. tree-ssa-operands
679 sets has_volatile_ops flag for all statements involving
680 reads and writes when aliases are not built to prevent passes
681 from removing them as dead. The flag thus has no use for us
682 and we need to look into all operands. */
684 EXECUTE_IF_SET_IN_BITMAP (STORED_SYMS (stmt), 0, i, bi)
686 tree var = referenced_var_lookup (i);
687 if (TREE_ADDRESSABLE (var)
688 || is_global_var (var)
689 || TREE_THIS_VOLATILE (var))
693 if (dead && LOADED_SYMS (stmt))
694 EXECUTE_IF_SET_IN_BITMAP (LOADED_SYMS (stmt), 0, i, bi)
695 if (TREE_THIS_VOLATILE (referenced_var_lookup (i)))
699 FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_ALL_OPERANDS)
700 if (TREE_THIS_VOLATILE (op))
703 /* Look for possible occurence var = indirect_ref (...) where
704 indirect_ref itself is volatile. */
706 if (dead && TREE_THIS_VOLATILE (GIMPLE_STMT_OPERAND (stmt, 1)))
711 tree call = get_call_expr_in (stmt);
713 /* When LHS of var = call (); is dead, simplify it into
714 call (); saving one operand. */
715 if (TREE_CODE (stmt) == GIMPLE_MODIFY_STMT
717 && TREE_SIDE_EFFECTS (call))
719 if (dump_file && (dump_flags & TDF_DETAILS))
721 fprintf (dump_file, "Deleted LHS of call: ");
722 print_generic_stmt (dump_file, stmt, TDF_SLIM);
723 fprintf (dump_file, "\n");
725 push_stmt_changes (bsi_stmt_ptr (bsi));
726 TREE_BLOCK (call) = TREE_BLOCK (stmt);
727 bsi_replace (&bsi, call, false);
728 maybe_clean_or_replace_eh_stmt (stmt, call);
729 mark_symbols_for_renaming (call);
730 pop_stmt_changes (bsi_stmt_ptr (bsi));
734 if (dump_file && (dump_flags & TDF_DETAILS))
736 fprintf (dump_file, " Deleted dead store '");
737 print_generic_expr (dump_file, stmt, dump_flags);
738 fprintf (dump_file, "'\n");
741 bsi_remove (&bsi, true);
742 todo |= TODO_cleanup_cfg;
744 todo |= TODO_remove_unused_locals | TODO_ggc_collect;
750 BITMAP_FREE (variables_loaded);
754 struct tree_opt_pass pass_simple_dse =
758 execute_simple_dse, /* execute */
761 0, /* static_pass_number */
763 PROP_ssa, /* properties_required */
764 0, /* properties_provided */
765 0, /* properties_destroyed */
766 0, /* todo_flags_start */
767 TODO_dump_func, /* todo_flags_finish */