1 /* Dead and redundant store elimination
2 Copyright (C) 2004-2019 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"
27 #include "tree-pass.h"
29 #include "gimple-pretty-print.h"
30 #include "fold-const.h"
31 #include "gimple-iterator.h"
35 #include "tree-cfgcleanup.h"
38 #include "tree-ssa-loop.h"
40 /* This file implements dead store elimination.
42 A dead store is a store into a memory location which will later be
43 overwritten by another store without any intervening loads. In this
44 case the earlier store can be deleted or trimmed if the store
47 A redundant store is a store into a memory location which stores
48 the exact same value as a prior store to the same memory location.
49 While this can often be handled by dead store elimination, removing
50 the redundant store is often better than removing or trimming the
53 In our SSA + virtual operand world we use immediate uses of virtual
54 operands to detect these cases. If a store's virtual definition
55 is used precisely once by a later store to the same location which
56 post dominates the first store, then the first store is dead. If
57 the data stored is the same, then the second store is redundant.
59 The single use of the store's virtual definition ensures that
60 there are no intervening aliased loads and the requirement that
61 the second load post dominate the first ensures that if the earlier
62 store executes, then the later stores will execute before the function
65 It may help to think of this as first moving the earlier store to
66 the point immediately before the later store. Again, the single
67 use of the virtual definition and the post-dominance relationship
68 ensure that such movement would be safe. Clearly if there are
69 back to back stores, then the second is makes the first dead. If
70 the second store stores the same value, then the second store is
73 Reviewing section 10.7.2 in Morgan's "Building an Optimizing Compiler"
74 may also help in understanding this code since it discusses the
75 relationship between dead store and redundant load elimination. In
76 fact, they are the same transformation applied to different views of
79 static void delete_dead_or_redundant_assignment (gimple_stmt_iterator
*, const char *);
80 static void delete_dead_or_redundant_call (gimple_stmt_iterator
*, const char *);
82 /* Bitmap of blocks that have had EH statements cleaned. We should
83 remove their dead edges eventually. */
84 static bitmap need_eh_cleanup
;
86 /* Return value from dse_classify_store */
90 DSE_STORE_MAYBE_PARTIAL_DEAD
,
94 /* STMT is a statement that may write into memory. Analyze it and
95 initialize WRITE to describe how STMT affects memory.
97 Return TRUE if the the statement was analyzed, FALSE otherwise.
99 It is always safe to return FALSE. But typically better optimziation
100 can be achieved by analyzing more statements. */
103 initialize_ao_ref_for_dse (gimple
*stmt
, ao_ref
*write
)
105 /* It's advantageous to handle certain mem* functions. */
106 if (gimple_call_builtin_p (stmt
, BUILT_IN_NORMAL
))
108 switch (DECL_FUNCTION_CODE (gimple_call_fndecl (stmt
)))
110 case BUILT_IN_MEMCPY
:
111 case BUILT_IN_MEMMOVE
:
112 case BUILT_IN_MEMSET
:
113 case BUILT_IN_MEMCPY_CHK
:
114 case BUILT_IN_MEMMOVE_CHK
:
115 case BUILT_IN_MEMSET_CHK
:
116 case BUILT_IN_STRNCPY
:
117 case BUILT_IN_STRNCPY_CHK
:
119 tree size
= gimple_call_arg (stmt
, 2);
120 tree ptr
= gimple_call_arg (stmt
, 0);
121 ao_ref_init_from_ptr_and_size (write
, ptr
, size
);
125 /* A calloc call can never be dead, but it can make
126 subsequent stores redundant if they store 0 into
127 the same memory locations. */
128 case BUILT_IN_CALLOC
:
130 tree nelem
= gimple_call_arg (stmt
, 0);
131 tree selem
= gimple_call_arg (stmt
, 1);
133 if (TREE_CODE (nelem
) == INTEGER_CST
134 && TREE_CODE (selem
) == INTEGER_CST
135 && (lhs
= gimple_call_lhs (stmt
)) != NULL_TREE
)
137 tree size
= fold_build2 (MULT_EXPR
, TREE_TYPE (nelem
),
139 ao_ref_init_from_ptr_and_size (write
, lhs
, size
);
148 else if (is_gimple_assign (stmt
))
150 ao_ref_init (write
, gimple_assign_lhs (stmt
));
156 /* Given REF from the the alias oracle, return TRUE if it is a valid
157 memory reference for dead store elimination, false otherwise.
159 In particular, the reference must have a known base, known maximum
160 size, start at a byte offset and have a size that is one or more
164 valid_ao_ref_for_dse (ao_ref
*ref
)
166 return (ao_ref_base (ref
)
167 && known_size_p (ref
->max_size
)
168 && maybe_ne (ref
->size
, 0)
169 && known_eq (ref
->max_size
, ref
->size
)
170 && known_ge (ref
->offset
, 0)
171 && multiple_p (ref
->offset
, BITS_PER_UNIT
)
172 && multiple_p (ref
->size
, BITS_PER_UNIT
));
175 /* Try to normalize COPY (an ao_ref) relative to REF. Essentially when we are
176 done COPY will only refer bytes found within REF. Return true if COPY
177 is known to intersect at least one byte of REF. */
180 normalize_ref (ao_ref
*copy
, ao_ref
*ref
)
182 if (!ordered_p (copy
->offset
, ref
->offset
))
185 /* If COPY starts before REF, then reset the beginning of
186 COPY to match REF and decrease the size of COPY by the
187 number of bytes removed from COPY. */
188 if (maybe_lt (copy
->offset
, ref
->offset
))
190 poly_int64 diff
= ref
->offset
- copy
->offset
;
191 if (maybe_le (copy
->size
, diff
))
194 copy
->offset
= ref
->offset
;
197 poly_int64 diff
= copy
->offset
- ref
->offset
;
198 if (maybe_le (ref
->size
, diff
))
201 /* If COPY extends beyond REF, chop off its size appropriately. */
202 poly_int64 limit
= ref
->size
- diff
;
203 if (!ordered_p (limit
, copy
->size
))
206 if (maybe_gt (copy
->size
, limit
))
211 /* Clear any bytes written by STMT from the bitmap LIVE_BYTES. The base
212 address written by STMT must match the one found in REF, which must
213 have its base address previously initialized.
215 This routine must be conservative. If we don't know the offset or
216 actual size written, assume nothing was written. */
219 clear_bytes_written_by (sbitmap live_bytes
, gimple
*stmt
, ao_ref
*ref
)
222 if (!initialize_ao_ref_for_dse (stmt
, &write
))
225 /* Verify we have the same base memory address, the write
226 has a known size and overlaps with REF. */
227 HOST_WIDE_INT start
, size
;
228 if (valid_ao_ref_for_dse (&write
)
229 && operand_equal_p (write
.base
, ref
->base
, OEP_ADDRESS_OF
)
230 && known_eq (write
.size
, write
.max_size
)
231 && normalize_ref (&write
, ref
)
232 && (write
.offset
- ref
->offset
).is_constant (&start
)
233 && write
.size
.is_constant (&size
))
234 bitmap_clear_range (live_bytes
, start
/ BITS_PER_UNIT
,
235 size
/ BITS_PER_UNIT
);
238 /* REF is a memory write. Extract relevant information from it and
239 initialize the LIVE_BYTES bitmap. If successful, return TRUE.
240 Otherwise return FALSE. */
243 setup_live_bytes_from_ref (ao_ref
*ref
, sbitmap live_bytes
)
245 HOST_WIDE_INT const_size
;
246 if (valid_ao_ref_for_dse (ref
)
247 && ref
->size
.is_constant (&const_size
)
248 && (const_size
/ BITS_PER_UNIT
249 <= PARAM_VALUE (PARAM_DSE_MAX_OBJECT_SIZE
)))
251 bitmap_clear (live_bytes
);
252 bitmap_set_range (live_bytes
, 0, const_size
/ BITS_PER_UNIT
);
258 /* Compute the number of elements that we can trim from the head and
259 tail of ORIG resulting in a bitmap that is a superset of LIVE.
261 Store the number of elements trimmed from the head and tail in
262 TRIM_HEAD and TRIM_TAIL.
264 STMT is the statement being trimmed and is used for debugging dump
268 compute_trims (ao_ref
*ref
, sbitmap live
, int *trim_head
, int *trim_tail
,
271 /* We use sbitmaps biased such that ref->offset is bit zero and the bitmap
272 extends through ref->size. So we know that in the original bitmap
273 bits 0..ref->size were true. We don't actually need the bitmap, just
274 the REF to compute the trims. */
276 /* Now identify how much, if any of the tail we can chop off. */
277 HOST_WIDE_INT const_size
;
278 int last_live
= bitmap_last_set_bit (live
);
279 if (ref
->size
.is_constant (&const_size
))
281 int last_orig
= (const_size
/ BITS_PER_UNIT
) - 1;
282 /* We can leave inconvenient amounts on the tail as
283 residual handling in mem* and str* functions is usually
284 reasonably efficient. */
285 *trim_tail
= last_orig
- last_live
;
287 /* But don't trim away out of bounds accesses, as this defeats
290 We could have a type with no TYPE_SIZE_UNIT or we could have a VLA
291 where TYPE_SIZE_UNIT is not a constant. */
293 && TYPE_SIZE_UNIT (TREE_TYPE (ref
->base
))
294 && TREE_CODE (TYPE_SIZE_UNIT (TREE_TYPE (ref
->base
))) == INTEGER_CST
295 && compare_tree_int (TYPE_SIZE_UNIT (TREE_TYPE (ref
->base
)),
302 /* Identify how much, if any of the head we can chop off. */
304 int first_live
= bitmap_first_set_bit (live
);
305 *trim_head
= first_live
- first_orig
;
307 /* If more than a word remains, then make sure to keep the
308 starting point at least word aligned. */
309 if (last_live
- first_live
> UNITS_PER_WORD
)
310 *trim_head
&= ~(UNITS_PER_WORD
- 1);
312 if ((*trim_head
|| *trim_tail
)
313 && dump_file
&& (dump_flags
& TDF_DETAILS
))
315 fprintf (dump_file
, " Trimming statement (head = %d, tail = %d): ",
316 *trim_head
, *trim_tail
);
317 print_gimple_stmt (dump_file
, stmt
, 0, dump_flags
);
318 fprintf (dump_file
, "\n");
322 /* STMT initializes an object from COMPLEX_CST where one or more of the
323 bytes written may be dead stores. REF is a representation of the
324 memory written. LIVE is the bitmap of stores that are actually live.
326 Attempt to rewrite STMT so that only the real or imaginary part of
327 the object is actually stored. */
330 maybe_trim_complex_store (ao_ref
*ref
, sbitmap live
, gimple
*stmt
)
332 int trim_head
, trim_tail
;
333 compute_trims (ref
, live
, &trim_head
, &trim_tail
, stmt
);
335 /* The amount of data trimmed from the head or tail must be at
336 least half the size of the object to ensure we're trimming
337 the entire real or imaginary half. By writing things this
338 way we avoid more O(n) bitmap operations. */
339 if (known_ge (trim_tail
* 2 * BITS_PER_UNIT
, ref
->size
))
341 /* TREE_REALPART is live */
342 tree x
= TREE_REALPART (gimple_assign_rhs1 (stmt
));
343 tree y
= gimple_assign_lhs (stmt
);
344 y
= build1 (REALPART_EXPR
, TREE_TYPE (x
), y
);
345 gimple_assign_set_lhs (stmt
, y
);
346 gimple_assign_set_rhs1 (stmt
, x
);
348 else if (known_ge (trim_head
* 2 * BITS_PER_UNIT
, ref
->size
))
350 /* TREE_IMAGPART is live */
351 tree x
= TREE_IMAGPART (gimple_assign_rhs1 (stmt
));
352 tree y
= gimple_assign_lhs (stmt
);
353 y
= build1 (IMAGPART_EXPR
, TREE_TYPE (x
), y
);
354 gimple_assign_set_lhs (stmt
, y
);
355 gimple_assign_set_rhs1 (stmt
, x
);
358 /* Other cases indicate parts of both the real and imag subobjects
359 are live. We do not try to optimize those cases. */
362 /* STMT initializes an object using a CONSTRUCTOR where one or more of the
363 bytes written are dead stores. ORIG is the bitmap of bytes stored by
364 STMT. LIVE is the bitmap of stores that are actually live.
366 Attempt to rewrite STMT so that only the real or imaginary part of
367 the object is actually stored.
369 The most common case for getting here is a CONSTRUCTOR with no elements
370 being used to zero initialize an object. We do not try to handle other
371 cases as those would force us to fully cover the object with the
372 CONSTRUCTOR node except for the components that are dead. */
375 maybe_trim_constructor_store (ao_ref
*ref
, sbitmap live
, gimple
*stmt
)
377 tree ctor
= gimple_assign_rhs1 (stmt
);
379 /* This is the only case we currently handle. It actually seems to
380 catch most cases of actual interest. */
381 gcc_assert (CONSTRUCTOR_NELTS (ctor
) == 0);
385 compute_trims (ref
, live
, &head_trim
, &tail_trim
, stmt
);
387 /* Now we want to replace the constructor initializer
388 with memset (object + head_trim, 0, size - head_trim - tail_trim). */
389 if (head_trim
|| tail_trim
)
391 /* We want &lhs for the MEM_REF expression. */
392 tree lhs_addr
= build_fold_addr_expr (gimple_assign_lhs (stmt
));
394 if (! is_gimple_min_invariant (lhs_addr
))
397 /* The number of bytes for the new constructor. */
398 poly_int64 ref_bytes
= exact_div (ref
->size
, BITS_PER_UNIT
);
399 poly_int64 count
= ref_bytes
- head_trim
- tail_trim
;
401 /* And the new type for the CONSTRUCTOR. Essentially it's just
402 a char array large enough to cover the non-trimmed parts of
403 the original CONSTRUCTOR. Note we want explicit bounds here
404 so that we know how many bytes to clear when expanding the
406 tree type
= build_array_type_nelts (char_type_node
, count
);
408 /* Build a suitable alias type rather than using alias set zero
409 to avoid pessimizing. */
410 tree alias_type
= reference_alias_ptr_type (gimple_assign_lhs (stmt
));
412 /* Build a MEM_REF representing the whole accessed area, starting
413 at the first byte not trimmed. */
414 tree exp
= fold_build2 (MEM_REF
, type
, lhs_addr
,
415 build_int_cst (alias_type
, head_trim
));
417 /* Now update STMT with a new RHS and LHS. */
418 gimple_assign_set_lhs (stmt
, exp
);
419 gimple_assign_set_rhs1 (stmt
, build_constructor (type
, NULL
));
423 /* STMT is a memcpy, memmove or memset. Decrement the number of bytes
424 copied/set by DECREMENT. */
426 decrement_count (gimple
*stmt
, int decrement
)
428 tree
*countp
= gimple_call_arg_ptr (stmt
, 2);
429 gcc_assert (TREE_CODE (*countp
) == INTEGER_CST
);
430 *countp
= wide_int_to_tree (TREE_TYPE (*countp
), (TREE_INT_CST_LOW (*countp
)
436 increment_start_addr (gimple
*stmt
, tree
*where
, int increment
)
438 if (TREE_CODE (*where
) == SSA_NAME
)
440 tree tem
= make_ssa_name (TREE_TYPE (*where
));
442 = gimple_build_assign (tem
, POINTER_PLUS_EXPR
, *where
,
443 build_int_cst (sizetype
, increment
));
444 gimple_stmt_iterator gsi
= gsi_for_stmt (stmt
);
445 gsi_insert_before (&gsi
, newop
, GSI_SAME_STMT
);
447 update_stmt (gsi_stmt (gsi
));
451 *where
= build_fold_addr_expr (fold_build2 (MEM_REF
, char_type_node
,
453 build_int_cst (ptr_type_node
,
457 /* STMT is builtin call that writes bytes in bitmap ORIG, some bytes are dead
458 (ORIG & ~NEW) and need not be stored. Try to rewrite STMT to reduce
459 the amount of data it actually writes.
461 Right now we only support trimming from the head or the tail of the
462 memory region. In theory we could split the mem* call, but it's
463 likely of marginal value. */
466 maybe_trim_memstar_call (ao_ref
*ref
, sbitmap live
, gimple
*stmt
)
468 switch (DECL_FUNCTION_CODE (gimple_call_fndecl (stmt
)))
470 case BUILT_IN_MEMCPY
:
471 case BUILT_IN_MEMMOVE
:
472 case BUILT_IN_STRNCPY
:
473 case BUILT_IN_MEMCPY_CHK
:
474 case BUILT_IN_MEMMOVE_CHK
:
475 case BUILT_IN_STRNCPY_CHK
:
477 int head_trim
, tail_trim
;
478 compute_trims (ref
, live
, &head_trim
, &tail_trim
, stmt
);
480 /* Tail trimming is easy, we can just reduce the count. */
482 decrement_count (stmt
, tail_trim
);
484 /* Head trimming requires adjusting all the arguments. */
487 tree
*dst
= gimple_call_arg_ptr (stmt
, 0);
488 increment_start_addr (stmt
, dst
, head_trim
);
489 tree
*src
= gimple_call_arg_ptr (stmt
, 1);
490 increment_start_addr (stmt
, src
, head_trim
);
491 decrement_count (stmt
, head_trim
);
496 case BUILT_IN_MEMSET
:
497 case BUILT_IN_MEMSET_CHK
:
499 int head_trim
, tail_trim
;
500 compute_trims (ref
, live
, &head_trim
, &tail_trim
, stmt
);
502 /* Tail trimming is easy, we can just reduce the count. */
504 decrement_count (stmt
, tail_trim
);
506 /* Head trimming requires adjusting all the arguments. */
509 tree
*dst
= gimple_call_arg_ptr (stmt
, 0);
510 increment_start_addr (stmt
, dst
, head_trim
);
511 decrement_count (stmt
, head_trim
);
521 /* STMT is a memory write where one or more bytes written are dead
522 stores. ORIG is the bitmap of bytes stored by STMT. LIVE is the
523 bitmap of stores that are actually live.
525 Attempt to rewrite STMT so that it writes fewer memory locations. Right
526 now we only support trimming at the start or end of the memory region.
527 It's not clear how much there is to be gained by trimming from the middle
531 maybe_trim_partially_dead_store (ao_ref
*ref
, sbitmap live
, gimple
*stmt
)
533 if (is_gimple_assign (stmt
)
534 && TREE_CODE (gimple_assign_lhs (stmt
)) != TARGET_MEM_REF
)
536 switch (gimple_assign_rhs_code (stmt
))
539 maybe_trim_constructor_store (ref
, live
, stmt
);
542 maybe_trim_complex_store (ref
, live
, stmt
);
550 /* Return TRUE if USE_REF reads bytes from LIVE where live is
551 derived from REF, a write reference.
553 While this routine may modify USE_REF, it's passed by value, not
554 location. So callers do not see those modifications. */
557 live_bytes_read (ao_ref use_ref
, ao_ref
*ref
, sbitmap live
)
559 /* We have already verified that USE_REF and REF hit the same object.
560 Now verify that there's actually an overlap between USE_REF and REF. */
561 HOST_WIDE_INT start
, size
;
562 if (normalize_ref (&use_ref
, ref
)
563 && (use_ref
.offset
- ref
->offset
).is_constant (&start
)
564 && use_ref
.size
.is_constant (&size
))
566 /* If USE_REF covers all of REF, then it will hit one or more
567 live bytes. This avoids useless iteration over the bitmap
569 if (start
== 0 && known_eq (size
, ref
->size
))
572 /* Now check if any of the remaining bits in use_ref are set in LIVE. */
573 return bitmap_bit_in_range_p (live
, start
/ BITS_PER_UNIT
,
574 (start
+ size
- 1) / BITS_PER_UNIT
);
579 /* Callback for dse_classify_store calling for_each_index. Verify that
580 indices are invariant in the loop with backedge PHI in basic-block DATA. */
583 check_name (tree
, tree
*idx
, void *data
)
585 basic_block phi_bb
= (basic_block
) data
;
586 if (TREE_CODE (*idx
) == SSA_NAME
587 && !SSA_NAME_IS_DEFAULT_DEF (*idx
)
588 && dominated_by_p (CDI_DOMINATORS
, gimple_bb (SSA_NAME_DEF_STMT (*idx
)),
594 /* STMT stores the value 0 into one or more memory locations
595 (via memset, empty constructor, calloc call, etc).
597 See if there is a subsequent store of the value 0 to one
598 or more of the same memory location(s). If so, the subsequent
599 store is redundant and can be removed.
601 The subsequent stores could be via memset, empty constructors,
602 simple MEM stores, etc. */
605 dse_optimize_redundant_stores (gimple
*stmt
)
609 /* We could do something fairly complex and look through PHIs
610 like DSE_CLASSIFY_STORE, but it doesn't seem to be worth
613 Look at all the immediate uses of the VDEF (which are obviously
614 dominated by STMT). See if one or more stores 0 into the same
615 memory locations a STMT, if so remove the immediate use statements. */
616 tree defvar
= gimple_vdef (stmt
);
619 FOR_EACH_IMM_USE_STMT (use_stmt
, ui
, defvar
)
621 /* Limit stmt walking. */
622 if (++cnt
> PARAM_VALUE (PARAM_DSE_MAX_ALIAS_QUERIES_PER_STORE
))
623 BREAK_FROM_IMM_USE_STMT (ui
);
625 /* If USE_STMT stores 0 into one or more of the same locations
626 as STMT and STMT would kill USE_STMT, then we can just remove
629 if ((is_gimple_assign (use_stmt
)
630 && gimple_vdef (use_stmt
)
631 && ((gimple_assign_rhs_code (use_stmt
) == CONSTRUCTOR
632 && CONSTRUCTOR_NELTS (gimple_assign_rhs1 (use_stmt
)) == 0
633 && !gimple_clobber_p (stmt
))
634 || (gimple_assign_rhs_code (use_stmt
) == INTEGER_CST
635 && integer_zerop (gimple_assign_rhs1 (use_stmt
)))))
636 || (gimple_call_builtin_p (use_stmt
, BUILT_IN_NORMAL
)
637 && (fndecl
= gimple_call_fndecl (use_stmt
)) != NULL
638 && (DECL_FUNCTION_CODE (fndecl
) == BUILT_IN_MEMSET
639 || DECL_FUNCTION_CODE (fndecl
) == BUILT_IN_MEMSET_CHK
)
640 && integer_zerop (gimple_call_arg (use_stmt
, 1))))
644 if (!initialize_ao_ref_for_dse (use_stmt
, &write
))
645 BREAK_FROM_IMM_USE_STMT (ui
)
647 if (valid_ao_ref_for_dse (&write
)
648 && stmt_kills_ref_p (stmt
, &write
))
650 gimple_stmt_iterator gsi
= gsi_for_stmt (use_stmt
);
651 if (is_gimple_assign (use_stmt
))
652 delete_dead_or_redundant_assignment (&gsi
, "redundant");
653 else if (is_gimple_call (use_stmt
))
654 delete_dead_or_redundant_call (&gsi
, "redundant");
662 /* A helper of dse_optimize_stmt.
663 Given a GIMPLE_ASSIGN in STMT that writes to REF, classify it
664 according to downstream uses and defs. Sets *BY_CLOBBER_P to true
665 if only clobber statements influenced the classification result.
666 Returns the classification. */
668 static dse_store_status
669 dse_classify_store (ao_ref
*ref
, gimple
*stmt
,
670 bool byte_tracking_enabled
, sbitmap live_bytes
,
671 bool *by_clobber_p
= NULL
)
678 *by_clobber_p
= true;
680 /* Find the first dominated statement that clobbers (part of) the
681 memory stmt stores to with no intermediate statement that may use
682 part of the memory stmt stores. That is, find a store that may
683 prove stmt to be a dead store. */
692 if (gimple_code (temp
) == GIMPLE_PHI
)
694 /* If we visit this PHI by following a backedge then we have to
695 make sure ref->ref only refers to SSA names that are invariant
696 with respect to the loop represented by this PHI node. */
697 if (dominated_by_p (CDI_DOMINATORS
, gimple_bb (stmt
),
699 && !for_each_index (ref
->ref
? &ref
->ref
: &ref
->base
,
700 check_name
, gimple_bb (temp
)))
701 return DSE_STORE_LIVE
;
702 defvar
= PHI_RESULT (temp
);
703 bitmap_set_bit (visited
, SSA_NAME_VERSION (defvar
));
706 defvar
= gimple_vdef (temp
);
707 auto_vec
<gimple
*, 10> defs
;
708 gimple
*phi_def
= NULL
;
709 FOR_EACH_IMM_USE_STMT (use_stmt
, ui
, defvar
)
711 /* Limit stmt walking. */
712 if (++cnt
> PARAM_VALUE (PARAM_DSE_MAX_ALIAS_QUERIES_PER_STORE
))
715 BREAK_FROM_IMM_USE_STMT (ui
);
718 /* We have visited ourselves already so ignore STMT for the
719 purpose of chaining. */
720 if (use_stmt
== stmt
)
722 /* In simple cases we can look through PHI nodes, but we
723 have to be careful with loops and with memory references
724 containing operands that are also operands of PHI nodes.
725 See gcc.c-torture/execute/20051110-*.c. */
726 else if (gimple_code (use_stmt
) == GIMPLE_PHI
)
728 /* If we already visited this PHI ignore it for further
730 if (!bitmap_bit_p (visited
,
731 SSA_NAME_VERSION (PHI_RESULT (use_stmt
))))
733 defs
.safe_push (use_stmt
);
737 /* If the statement is a use the store is not dead. */
738 else if (ref_maybe_used_by_stmt_p (use_stmt
, ref
))
740 /* Handle common cases where we can easily build an ao_ref
741 structure for USE_STMT and in doing so we find that the
742 references hit non-live bytes and thus can be ignored. */
743 if (byte_tracking_enabled
744 && is_gimple_assign (use_stmt
))
747 ao_ref_init (&use_ref
, gimple_assign_rhs1 (use_stmt
));
748 if (valid_ao_ref_for_dse (&use_ref
)
749 && use_ref
.base
== ref
->base
750 && known_eq (use_ref
.size
, use_ref
.max_size
)
751 && !live_bytes_read (use_ref
, ref
, live_bytes
))
753 /* If this is a store, remember it as we possibly
754 need to walk the defs uses. */
755 if (gimple_vdef (use_stmt
))
756 defs
.safe_push (use_stmt
);
762 BREAK_FROM_IMM_USE_STMT (ui
);
764 /* If this is a store, remember it as we possibly need to walk the
766 else if (gimple_vdef (use_stmt
))
767 defs
.safe_push (use_stmt
);
772 /* STMT might be partially dead and we may be able to reduce
773 how many memory locations it stores into. */
774 if (byte_tracking_enabled
&& !gimple_clobber_p (stmt
))
775 return DSE_STORE_MAYBE_PARTIAL_DEAD
;
776 return DSE_STORE_LIVE
;
779 /* If we didn't find any definition this means the store is dead
780 if it isn't a store to global reachable memory. In this case
781 just pretend the stmt makes itself dead. Otherwise fail. */
782 if (defs
.is_empty ())
784 if (ref_may_alias_global_p (ref
))
785 return DSE_STORE_LIVE
;
788 *by_clobber_p
= false;
789 return DSE_STORE_DEAD
;
792 /* Process defs and remove those we need not process further. */
793 for (unsigned i
= 0; i
< defs
.length ();)
795 gimple
*def
= defs
[i
];
798 /* If the path to check starts with a kill we do not need to
800 ??? With byte tracking we need only kill the bytes currently
802 if (stmt_kills_ref_p (def
, ref
))
804 if (by_clobber_p
&& !gimple_clobber_p (def
))
805 *by_clobber_p
= false;
806 defs
.unordered_remove (i
);
808 /* In addition to kills we can remove defs whose only use
809 is another def in defs. That can only ever be PHIs of which
810 we track a single for simplicity reasons (we fail for multiple
811 PHIs anyways). We can also ignore defs that feed only into
812 already visited PHIs. */
813 else if (gimple_code (def
) != GIMPLE_PHI
814 && single_imm_use (gimple_vdef (def
), &use_p
, &use_stmt
)
815 && (use_stmt
== phi_def
816 || (gimple_code (use_stmt
) == GIMPLE_PHI
817 && bitmap_bit_p (visited
,
819 (PHI_RESULT (use_stmt
))))))
820 defs
.unordered_remove (i
);
825 /* If all defs kill the ref we are done. */
826 if (defs
.is_empty ())
827 return DSE_STORE_DEAD
;
828 /* If more than one def survives fail. */
829 if (defs
.length () > 1)
831 /* STMT might be partially dead and we may be able to reduce
832 how many memory locations it stores into. */
833 if (byte_tracking_enabled
&& !gimple_clobber_p (stmt
))
834 return DSE_STORE_MAYBE_PARTIAL_DEAD
;
835 return DSE_STORE_LIVE
;
839 /* Track partial kills. */
840 if (byte_tracking_enabled
)
842 clear_bytes_written_by (live_bytes
, temp
, ref
);
843 if (bitmap_empty_p (live_bytes
))
845 if (by_clobber_p
&& !gimple_clobber_p (temp
))
846 *by_clobber_p
= false;
847 return DSE_STORE_DEAD
;
851 /* Continue walking until there are no more live bytes. */
856 class dse_dom_walker
: public dom_walker
859 dse_dom_walker (cdi_direction direction
)
860 : dom_walker (direction
),
861 m_live_bytes (PARAM_VALUE (PARAM_DSE_MAX_OBJECT_SIZE
)),
862 m_byte_tracking_enabled (false) {}
864 virtual edge
before_dom_children (basic_block
);
867 auto_sbitmap m_live_bytes
;
868 bool m_byte_tracking_enabled
;
869 void dse_optimize_stmt (gimple_stmt_iterator
*);
872 /* Delete a dead call at GSI, which is mem* call of some kind. */
874 delete_dead_or_redundant_call (gimple_stmt_iterator
*gsi
, const char *type
)
876 gimple
*stmt
= gsi_stmt (*gsi
);
877 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
879 fprintf (dump_file
, " Deleted %s call: ", type
);
880 print_gimple_stmt (dump_file
, stmt
, 0, dump_flags
);
881 fprintf (dump_file
, "\n");
884 tree lhs
= gimple_call_lhs (stmt
);
887 tree ptr
= gimple_call_arg (stmt
, 0);
888 gimple
*new_stmt
= gimple_build_assign (lhs
, ptr
);
889 unlink_stmt_vdef (stmt
);
890 if (gsi_replace (gsi
, new_stmt
, true))
891 bitmap_set_bit (need_eh_cleanup
, gimple_bb (stmt
)->index
);
895 /* Then we need to fix the operand of the consuming stmt. */
896 unlink_stmt_vdef (stmt
);
898 /* Remove the dead store. */
899 if (gsi_remove (gsi
, true))
900 bitmap_set_bit (need_eh_cleanup
, gimple_bb (stmt
)->index
);
905 /* Delete a dead store at GSI, which is a gimple assignment. */
908 delete_dead_or_redundant_assignment (gimple_stmt_iterator
*gsi
, const char *type
)
910 gimple
*stmt
= gsi_stmt (*gsi
);
911 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
913 fprintf (dump_file
, " Deleted %s store: ", type
);
914 print_gimple_stmt (dump_file
, stmt
, 0, dump_flags
);
915 fprintf (dump_file
, "\n");
918 /* Then we need to fix the operand of the consuming stmt. */
919 unlink_stmt_vdef (stmt
);
921 /* Remove the dead store. */
922 basic_block bb
= gimple_bb (stmt
);
923 if (gsi_remove (gsi
, true))
924 bitmap_set_bit (need_eh_cleanup
, bb
->index
);
926 /* And release any SSA_NAMEs set in this statement back to the
931 /* Attempt to eliminate dead stores in the statement referenced by BSI.
933 A dead store is a store into a memory location which will later be
934 overwritten by another store without any intervening loads. In this
935 case the earlier store can be deleted.
937 In our SSA + virtual operand world we use immediate uses of virtual
938 operands to detect dead stores. If a store's virtual definition
939 is used precisely once by a later store to the same location which
940 post dominates the first store, then the first store is dead. */
943 dse_dom_walker::dse_optimize_stmt (gimple_stmt_iterator
*gsi
)
945 gimple
*stmt
= gsi_stmt (*gsi
);
947 /* If this statement has no virtual defs, then there is nothing
949 if (!gimple_vdef (stmt
))
952 /* Don't return early on *this_2(D) ={v} {CLOBBER}. */
953 if (gimple_has_volatile_ops (stmt
)
954 && (!gimple_clobber_p (stmt
)
955 || TREE_CODE (gimple_assign_lhs (stmt
)) != MEM_REF
))
959 if (!initialize_ao_ref_for_dse (stmt
, &ref
))
962 /* We know we have virtual definitions. We can handle assignments and
963 some builtin calls. */
964 if (gimple_call_builtin_p (stmt
, BUILT_IN_NORMAL
))
966 tree fndecl
= gimple_call_fndecl (stmt
);
967 switch (DECL_FUNCTION_CODE (fndecl
))
969 case BUILT_IN_MEMCPY
:
970 case BUILT_IN_MEMMOVE
:
971 case BUILT_IN_STRNCPY
:
972 case BUILT_IN_MEMSET
:
973 case BUILT_IN_MEMCPY_CHK
:
974 case BUILT_IN_MEMMOVE_CHK
:
975 case BUILT_IN_STRNCPY_CHK
:
976 case BUILT_IN_MEMSET_CHK
:
978 /* Occasionally calls with an explicit length of zero
979 show up in the IL. It's pointless to do analysis
980 on them, they're trivially dead. */
981 tree size
= gimple_call_arg (stmt
, 2);
982 if (integer_zerop (size
))
984 delete_dead_or_redundant_call (gsi
, "dead");
988 /* If this is a memset call that initializes an object
989 to zero, it may be redundant with an earlier memset
990 or empty CONSTRUCTOR of a larger object. */
991 if ((DECL_FUNCTION_CODE (fndecl
) == BUILT_IN_MEMSET
992 || DECL_FUNCTION_CODE (fndecl
) == BUILT_IN_MEMSET_CHK
)
993 && integer_zerop (gimple_call_arg (stmt
, 1)))
994 dse_optimize_redundant_stores (stmt
);
996 enum dse_store_status store_status
;
997 m_byte_tracking_enabled
998 = setup_live_bytes_from_ref (&ref
, m_live_bytes
);
999 store_status
= dse_classify_store (&ref
, stmt
,
1000 m_byte_tracking_enabled
,
1002 if (store_status
== DSE_STORE_LIVE
)
1005 if (store_status
== DSE_STORE_MAYBE_PARTIAL_DEAD
)
1007 maybe_trim_memstar_call (&ref
, m_live_bytes
, stmt
);
1011 if (store_status
== DSE_STORE_DEAD
)
1012 delete_dead_or_redundant_call (gsi
, "dead");
1016 case BUILT_IN_CALLOC
:
1017 /* We already know the arguments are integer constants. */
1018 dse_optimize_redundant_stores (stmt
);
1026 if (is_gimple_assign (stmt
))
1028 bool by_clobber_p
= false;
1030 /* First see if this store is a CONSTRUCTOR and if there
1031 are subsequent CONSTRUCTOR stores which are totally
1032 subsumed by this statement. If so remove the subsequent
1035 This will tend to make fewer calls into memset with longer
1037 if (gimple_assign_rhs_code (stmt
) == CONSTRUCTOR
1038 && CONSTRUCTOR_NELTS (gimple_assign_rhs1 (stmt
)) == 0
1039 && !gimple_clobber_p (stmt
))
1040 dse_optimize_redundant_stores (stmt
);
1042 /* Self-assignments are zombies. */
1043 if (operand_equal_p (gimple_assign_rhs1 (stmt
),
1044 gimple_assign_lhs (stmt
), 0))
1048 m_byte_tracking_enabled
1049 = setup_live_bytes_from_ref (&ref
, m_live_bytes
);
1050 enum dse_store_status store_status
;
1051 store_status
= dse_classify_store (&ref
, stmt
,
1052 m_byte_tracking_enabled
,
1053 m_live_bytes
, &by_clobber_p
);
1054 if (store_status
== DSE_STORE_LIVE
)
1057 if (store_status
== DSE_STORE_MAYBE_PARTIAL_DEAD
)
1059 maybe_trim_partially_dead_store (&ref
, m_live_bytes
, stmt
);
1064 /* Now we know that use_stmt kills the LHS of stmt. */
1066 /* But only remove *this_2(D) ={v} {CLOBBER} if killed by
1067 another clobber stmt. */
1068 if (gimple_clobber_p (stmt
)
1072 delete_dead_or_redundant_assignment (gsi
, "dead");
1077 dse_dom_walker::before_dom_children (basic_block bb
)
1079 gimple_stmt_iterator gsi
;
1081 for (gsi
= gsi_last_bb (bb
); !gsi_end_p (gsi
);)
1083 dse_optimize_stmt (&gsi
);
1084 if (gsi_end_p (gsi
))
1085 gsi
= gsi_last_bb (bb
);
1094 const pass_data pass_data_dse
=
1096 GIMPLE_PASS
, /* type */
1098 OPTGROUP_NONE
, /* optinfo_flags */
1099 TV_TREE_DSE
, /* tv_id */
1100 ( PROP_cfg
| PROP_ssa
), /* properties_required */
1101 0, /* properties_provided */
1102 0, /* properties_destroyed */
1103 0, /* todo_flags_start */
1104 0, /* todo_flags_finish */
1107 class pass_dse
: public gimple_opt_pass
1110 pass_dse (gcc::context
*ctxt
)
1111 : gimple_opt_pass (pass_data_dse
, ctxt
)
1114 /* opt_pass methods: */
1115 opt_pass
* clone () { return new pass_dse (m_ctxt
); }
1116 virtual bool gate (function
*) { return flag_tree_dse
!= 0; }
1117 virtual unsigned int execute (function
*);
1119 }; // class pass_dse
1122 pass_dse::execute (function
*fun
)
1124 need_eh_cleanup
= BITMAP_ALLOC (NULL
);
1126 renumber_gimple_stmt_uids ();
1128 /* We might consider making this a property of each pass so that it
1129 can be [re]computed on an as-needed basis. Particularly since
1130 this pass could be seen as an extension of DCE which needs post
1132 calculate_dominance_info (CDI_POST_DOMINATORS
);
1133 calculate_dominance_info (CDI_DOMINATORS
);
1135 /* Dead store elimination is fundamentally a walk of the post-dominator
1136 tree and a backwards walk of statements within each block. */
1137 dse_dom_walker (CDI_POST_DOMINATORS
).walk (fun
->cfg
->x_exit_block_ptr
);
1139 /* Removal of stores may make some EH edges dead. Purge such edges from
1140 the CFG as needed. */
1141 if (!bitmap_empty_p (need_eh_cleanup
))
1143 gimple_purge_all_dead_eh_edges (need_eh_cleanup
);
1144 cleanup_tree_cfg ();
1147 BITMAP_FREE (need_eh_cleanup
);
1149 /* For now, just wipe the post-dominator information. */
1150 free_dominance_info (CDI_POST_DOMINATORS
);
1157 make_pass_dse (gcc::context
*ctxt
)
1159 return new pass_dse (ctxt
);