1 /* Dead store elimination
2 Copyright (C) 2004, 2005 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 2, 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 COPYING. If not, write to
18 the Free Software Foundation, 59 Temple Place - Suite 330,
19 Boston, MA 02111-1307, USA. */
23 #include "coretypes.h"
30 #include "basic-block.h"
32 #include "diagnostic.h"
33 #include "tree-flow.h"
34 #include "tree-pass.h"
35 #include "tree-dump.h"
39 /* This file implements dead store elimination.
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.
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.
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
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
58 use of the virtual definition and the post-dominance relationship
59 ensure that such movement would be safe. Clearly if there are
60 back to back stores, then the second is redundant.
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
69 struct dse_global_data
71 /* This is the global bitmap for store statements.
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. */
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
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 *,
92 static void dse_optimize_stmt (struct dom_walk_data *,
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 record_voperand_set (bitmap, bitmap *, unsigned int);
99 static unsigned max_stmt_uid; /* Maximal uid of a statement. Uids to phi
100 nodes are assigned using the versions of
101 ssa names they define. */
103 /* Returns uid of statement STMT. */
106 get_stmt_uid (tree stmt)
108 if (TREE_CODE (stmt) == PHI_NODE)
109 return SSA_NAME_VERSION (PHI_RESULT (stmt)) + max_stmt_uid;
111 return stmt_ann (stmt)->uid;
114 /* Function indicating whether we ought to include information for 'var'
115 when calculating immediate uses. For this pass we only want use
116 information for virtual variables. */
119 need_imm_uses_for (tree var)
121 return !is_gimple_reg (var);
125 /* Set bit UID in bitmaps GLOBAL and *LOCAL, creating *LOCAL as needed. */
127 record_voperand_set (bitmap global, bitmap *local, unsigned int uid)
129 /* Lazily allocate the bitmap. Note that we do not get a notification
130 when the block local data structures die, so we allocate the local
131 bitmap backed by the GC system. */
133 *local = BITMAP_GGC_ALLOC ();
135 /* Set the bit in the local and global bitmaps. */
136 bitmap_set_bit (*local, uid);
137 bitmap_set_bit (global, uid);
139 /* Initialize block local data structures. */
142 dse_initialize_block_local_data (struct dom_walk_data *walk_data,
143 basic_block bb ATTRIBUTE_UNUSED,
146 struct dse_block_local_data *bd
147 = VARRAY_TOP_GENERIC_PTR (walk_data->block_data_stack);
149 /* If we are given a recycled block local data structure, ensure any
150 bitmap associated with the block is cleared. */
154 bitmap_clear (bd->stores);
158 /* Attempt to eliminate dead stores in the statement referenced by BSI.
160 A dead store is a store into a memory location which will later be
161 overwritten by another store without any intervening loads. In this
162 case the earlier store can be deleted.
164 In our SSA + virtual operand world we use immediate uses of virtual
165 operands to detect dead stores. If a store's virtual definition
166 is used precisely once by a later store to the same location which
167 post dominates the first store, then the first store is dead. */
170 dse_optimize_stmt (struct dom_walk_data *walk_data,
171 basic_block bb ATTRIBUTE_UNUSED,
172 block_stmt_iterator bsi)
174 struct dse_block_local_data *bd
175 = VARRAY_TOP_GENERIC_PTR (walk_data->block_data_stack);
176 struct dse_global_data *dse_gd = walk_data->global_data;
177 tree stmt = bsi_stmt (bsi);
178 stmt_ann_t ann = stmt_ann (stmt);
179 v_may_def_optype v_may_defs;
181 get_stmt_operands (stmt);
182 v_may_defs = V_MAY_DEF_OPS (ann);
184 /* If this statement has no virtual uses, then there is nothing
186 if (NUM_V_MAY_DEFS (v_may_defs) == 0)
189 /* We know we have virtual definitions. If this is a MODIFY_EXPR that's
190 not also a function call, then record it into our table. */
191 if (get_call_expr_in (stmt))
194 if (ann->has_volatile_ops)
197 if (TREE_CODE (stmt) == MODIFY_EXPR)
199 unsigned int num_uses = 0, count = 0;
200 use_operand_p first_use_p = NULL_USE_OPERAND_P;
203 tree defvar = NULL_TREE, usevar = NULL_TREE;
208 FOR_EACH_SSA_MAYDEF_OPERAND (var1, var2, stmt, op_iter)
210 defvar = DEF_FROM_PTR (var1);
211 usevar = USE_FROM_PTR (var2);
212 num_uses += num_imm_uses (defvar);
214 if (num_uses > 1 || count > 1)
218 if (count == 1 && num_uses == 1)
220 single_imm_use (defvar, &use_p, &use_stmt);
221 gcc_assert (use_p != NULL_USE_OPERAND_P);
223 use = USE_FROM_PTR (use_p);
227 record_voperand_set (dse_gd->stores, &bd->stores, ann->uid);
231 /* Skip through any PHI nodes we have already seen if the PHI
232 represents the only use of this store.
234 Note this does not handle the case where the store has
235 multiple V_MAY_DEFs which all reach a set of PHI nodes in the
237 while (use_p != NULL_USE_OPERAND_P
238 && TREE_CODE (use_stmt) == PHI_NODE
239 && bitmap_bit_p (dse_gd->stores, get_stmt_uid (use_stmt)))
241 /* Skip past this PHI and loop again in case we had a PHI
243 if (single_imm_use (PHI_RESULT (use_stmt), &use_p, &use_stmt))
244 use = USE_FROM_PTR (use_p);
247 /* If we have precisely one immediate use at this point, then we may
248 have found redundant store. */
249 if (use_p != NULL_USE_OPERAND_P
250 && bitmap_bit_p (dse_gd->stores, get_stmt_uid (use_stmt))
251 && operand_equal_p (TREE_OPERAND (stmt, 0),
252 TREE_OPERAND (use_stmt, 0), 0))
254 /* Make sure we propagate the ABNORMAL bit setting. */
255 if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (USE_FROM_PTR (first_use_p)))
256 SSA_NAME_OCCURS_IN_ABNORMAL_PHI (usevar) = 1;
257 /* Then we need to fix the operand of the consuming stmt. */
258 SET_USE (first_use_p, usevar);
260 if (dump_file && (dump_flags & TDF_DETAILS))
262 fprintf (dump_file, " Deleted dead store '");
263 print_generic_expr (dump_file, bsi_stmt (bsi), dump_flags);
264 fprintf (dump_file, "'\n");
267 /* Remove the dead store. */
270 /* And release any SSA_NAMEs set in this statement back to the
275 record_voperand_set (dse_gd->stores, &bd->stores, ann->uid);
279 /* Record that we have seen the PHIs at the start of BB which correspond
280 to virtual operands. */
282 dse_record_phis (struct dom_walk_data *walk_data, basic_block bb)
284 struct dse_block_local_data *bd
285 = VARRAY_TOP_GENERIC_PTR (walk_data->block_data_stack);
286 struct dse_global_data *dse_gd = walk_data->global_data;
289 for (phi = phi_nodes (bb); phi; phi = PHI_CHAIN (phi))
290 if (need_imm_uses_for (PHI_RESULT (phi)))
291 record_voperand_set (dse_gd->stores,
297 dse_finalize_block (struct dom_walk_data *walk_data,
298 basic_block bb ATTRIBUTE_UNUSED)
300 struct dse_block_local_data *bd
301 = VARRAY_TOP_GENERIC_PTR (walk_data->block_data_stack);
302 struct dse_global_data *dse_gd = walk_data->global_data;
303 bitmap stores = dse_gd->stores;
307 /* Unwind the stores noted in this basic block. */
309 EXECUTE_IF_SET_IN_BITMAP (bd->stores, 0, i, bi)
311 bitmap_clear_bit (stores, i);
318 struct dom_walk_data walk_data;
319 struct dse_global_data dse_gd;
322 /* Create a UID for each statement in the function. Ordering of the
323 UIDs is not important for this pass. */
327 block_stmt_iterator bsi;
329 for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
330 stmt_ann (bsi_stmt (bsi))->uid = max_stmt_uid++;
333 /* We might consider making this a property of each pass so that it
334 can be [re]computed on an as-needed basis. Particularly since
335 this pass could be seen as an extension of DCE which needs post
337 calculate_dominance_info (CDI_POST_DOMINATORS);
339 /* Dead store elimination is fundamentally a walk of the post-dominator
340 tree and a backwards walk of statements within each block. */
341 walk_data.walk_stmts_backward = true;
342 walk_data.dom_direction = CDI_POST_DOMINATORS;
343 walk_data.initialize_block_local_data = dse_initialize_block_local_data;
344 walk_data.before_dom_children_before_stmts = NULL;
345 walk_data.before_dom_children_walk_stmts = dse_optimize_stmt;
346 walk_data.before_dom_children_after_stmts = dse_record_phis;
347 walk_data.after_dom_children_before_stmts = NULL;
348 walk_data.after_dom_children_walk_stmts = NULL;
349 walk_data.after_dom_children_after_stmts = dse_finalize_block;
351 walk_data.block_local_data_size = sizeof (struct dse_block_local_data);
353 /* This is the main hash table for the dead store elimination pass. */
354 dse_gd.stores = BITMAP_ALLOC (NULL);
355 walk_data.global_data = &dse_gd;
357 /* Initialize the dominator walker. */
358 init_walk_dominator_tree (&walk_data);
360 /* Recursively walk the dominator tree. */
361 walk_dominator_tree (&walk_data, EXIT_BLOCK_PTR);
363 /* Finalize the dominator walker. */
364 fini_walk_dominator_tree (&walk_data);
366 /* Release the main bitmap. */
367 BITMAP_FREE (dse_gd.stores);
369 /* For now, just wipe the post-dominator information. */
370 free_dominance_info (CDI_POST_DOMINATORS);
376 return flag_tree_dse != 0;
379 struct tree_opt_pass pass_dse = {
382 tree_ssa_dse, /* execute */
385 0, /* static_pass_number */
386 TV_TREE_DSE, /* tv_id */
388 | PROP_alias, /* properties_required */
389 0, /* properties_provided */
390 0, /* properties_destroyed */
391 0, /* todo_flags_start */
392 TODO_dump_func | TODO_ggc_collect /* todo_flags_finish */