1 /* Alias analysis for GNU C
2 Copyright (C) 1997, 1998, 1999, 2000, 2001 Free Software Foundation, Inc.
3 Contributed by John Carr (jfc@mit.edu).
5 This file is part of GNU CC.
7 GNU CC 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 2, or (at your option)
12 GNU CC 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.
17 You should have received a copy of the GNU General Public License
18 along with GNU CC; see the file COPYING. If not, write to
19 the Free Software Foundation, 59 Temple Place - Suite 330,
20 Boston, MA 02111-1307, USA. */
28 #include "insn-flags.h"
31 #include "hard-reg-set.h"
32 #include "basic-block.h"
37 #include "splay-tree.h"
40 /* The alias sets assigned to MEMs assist the back-end in determining
41 which MEMs can alias which other MEMs. In general, two MEMs in
42 different alias sets cannot alias each other, with one important
43 exception. Consider something like:
45 struct S {int i; double d; };
47 a store to an `S' can alias something of either type `int' or type
48 `double'. (However, a store to an `int' cannot alias a `double'
49 and vice versa.) We indicate this via a tree structure that looks
57 (The arrows are directed and point downwards.)
58 In this situation we say the alias set for `struct S' is the
59 `superset' and that those for `int' and `double' are `subsets'.
61 To see whether two alias sets can point to the same memory, we must
62 see if either alias set is a subset of the other. We need not trace
63 past immediate decendents, however, since we propagate all
64 grandchildren up one level.
66 Alias set zero is implicitly a superset of all other alias sets.
67 However, this is no actual entry for alias set zero. It is an
68 error to attempt to explicitly construct a subset of zero. */
70 typedef struct alias_set_entry
72 /* The alias set number, as stored in MEM_ALIAS_SET. */
73 HOST_WIDE_INT alias_set
;
75 /* The children of the alias set. These are not just the immediate
76 children, but, in fact, all decendents. So, if we have:
78 struct T { struct S s; float f; }
80 continuing our example above, the children here will be all of
81 `int', `double', `float', and `struct S'. */
84 /* Nonzero if would have a child of zero: this effectively makes this
85 alias set the same as alias set zero. */
89 static int rtx_equal_for_memref_p
PARAMS ((rtx
, rtx
));
90 static rtx find_symbolic_term
PARAMS ((rtx
));
91 static rtx get_addr
PARAMS ((rtx
));
92 static int memrefs_conflict_p
PARAMS ((int, rtx
, int, rtx
,
94 static void record_set
PARAMS ((rtx
, rtx
, void *));
95 static rtx find_base_term
PARAMS ((rtx
));
96 static int base_alias_check
PARAMS ((rtx
, rtx
, enum machine_mode
,
98 static int handled_component_p
PARAMS ((tree
));
99 static int can_address_p
PARAMS ((tree
));
100 static rtx find_base_value
PARAMS ((rtx
));
101 static int mems_in_disjoint_alias_sets_p
PARAMS ((rtx
, rtx
));
102 static int insert_subset_children
PARAMS ((splay_tree_node
, void*));
103 static tree find_base_decl
PARAMS ((tree
));
104 static alias_set_entry get_alias_set_entry
PARAMS ((HOST_WIDE_INT
));
105 static rtx fixed_scalar_and_varying_struct_p
PARAMS ((rtx
, rtx
, rtx
, rtx
,
106 int (*) (rtx
, int)));
107 static int aliases_everything_p
PARAMS ((rtx
));
108 static int write_dependence_p
PARAMS ((rtx
, rtx
, int));
109 static int nonlocal_mentioned_p
PARAMS ((rtx
));
111 static int loop_p
PARAMS ((void));
113 /* Set up all info needed to perform alias analysis on memory references. */
115 /* Returns the size in bytes of the mode of X. */
116 #define SIZE_FOR_MODE(X) (GET_MODE_SIZE (GET_MODE (X)))
118 /* Returns nonzero if MEM1 and MEM2 do not alias because they are in
119 different alias sets. We ignore alias sets in functions making use
120 of variable arguments because the va_arg macros on some systems are
122 #define DIFFERENT_ALIAS_SETS_P(MEM1, MEM2) \
123 mems_in_disjoint_alias_sets_p (MEM1, MEM2)
125 /* Cap the number of passes we make over the insns propagating alias
126 information through set chains. 10 is a completely arbitrary choice. */
127 #define MAX_ALIAS_LOOP_PASSES 10
129 /* reg_base_value[N] gives an address to which register N is related.
130 If all sets after the first add or subtract to the current value
131 or otherwise modify it so it does not point to a different top level
132 object, reg_base_value[N] is equal to the address part of the source
135 A base address can be an ADDRESS, SYMBOL_REF, or LABEL_REF. ADDRESS
136 expressions represent certain special values: function arguments and
137 the stack, frame, and argument pointers.
139 The contents of an ADDRESS is not normally used, the mode of the
140 ADDRESS determines whether the ADDRESS is a function argument or some
141 other special value. Pointer equality, not rtx_equal_p, determines whether
142 two ADDRESS expressions refer to the same base address.
144 The only use of the contents of an ADDRESS is for determining if the
145 current function performs nonlocal memory memory references for the
146 purposes of marking the function as a constant function. */
148 static rtx
*reg_base_value
;
149 static rtx
*new_reg_base_value
;
150 static unsigned int reg_base_value_size
; /* size of reg_base_value array */
152 #define REG_BASE_VALUE(X) \
153 (REGNO (X) < reg_base_value_size \
154 ? reg_base_value[REGNO (X)] : 0)
156 /* Vector of known invariant relationships between registers. Set in
157 loop unrolling. Indexed by register number, if nonzero the value
158 is an expression describing this register in terms of another.
160 The length of this array is REG_BASE_VALUE_SIZE.
162 Because this array contains only pseudo registers it has no effect
164 static rtx
*alias_invariant
;
166 /* Vector indexed by N giving the initial (unchanging) value known for
167 pseudo-register N. This array is initialized in
168 init_alias_analysis, and does not change until end_alias_analysis
170 rtx
*reg_known_value
;
172 /* Indicates number of valid entries in reg_known_value. */
173 static unsigned int reg_known_value_size
;
175 /* Vector recording for each reg_known_value whether it is due to a
176 REG_EQUIV note. Future passes (viz., reload) may replace the
177 pseudo with the equivalent expression and so we account for the
178 dependences that would be introduced if that happens.
180 The REG_EQUIV notes created in assign_parms may mention the arg
181 pointer, and there are explicit insns in the RTL that modify the
182 arg pointer. Thus we must ensure that such insns don't get
183 scheduled across each other because that would invalidate the
184 REG_EQUIV notes. One could argue that the REG_EQUIV notes are
185 wrong, but solving the problem in the scheduler will likely give
186 better code, so we do it here. */
187 char *reg_known_equiv_p
;
189 /* True when scanning insns from the start of the rtl to the
190 NOTE_INSN_FUNCTION_BEG note. */
191 static int copying_arguments
;
193 /* The splay-tree used to store the various alias set entries. */
194 static splay_tree alias_sets
;
196 /* Returns a pointer to the alias set entry for ALIAS_SET, if there is
197 such an entry, or NULL otherwise. */
199 static alias_set_entry
200 get_alias_set_entry (alias_set
)
201 HOST_WIDE_INT alias_set
;
204 = splay_tree_lookup (alias_sets
, (splay_tree_key
) alias_set
);
206 return sn
!= 0 ? ((alias_set_entry
) sn
->value
) : 0;
209 /* Returns nonzero if the alias sets for MEM1 and MEM2 are such that
210 the two MEMs cannot alias each other. */
213 mems_in_disjoint_alias_sets_p (mem1
, mem2
)
217 #ifdef ENABLE_CHECKING
218 /* Perform a basic sanity check. Namely, that there are no alias sets
219 if we're not using strict aliasing. This helps to catch bugs
220 whereby someone uses PUT_CODE, but doesn't clear MEM_ALIAS_SET, or
221 where a MEM is allocated in some way other than by the use of
222 gen_rtx_MEM, and the MEM_ALIAS_SET is not cleared. If we begin to
223 use alias sets to indicate that spilled registers cannot alias each
224 other, we might need to remove this check. */
225 if (! flag_strict_aliasing
226 && (MEM_ALIAS_SET (mem1
) != 0 || MEM_ALIAS_SET (mem2
) != 0))
230 return ! alias_sets_conflict_p (MEM_ALIAS_SET (mem1
), MEM_ALIAS_SET (mem2
));
233 /* Insert the NODE into the splay tree given by DATA. Used by
234 record_alias_subset via splay_tree_foreach. */
237 insert_subset_children (node
, data
)
238 splay_tree_node node
;
241 splay_tree_insert ((splay_tree
) data
, node
->key
, node
->value
);
246 /* Return 1 if the two specified alias sets may conflict. */
249 alias_sets_conflict_p (set1
, set2
)
250 HOST_WIDE_INT set1
, set2
;
254 /* If have no alias set information for one of the operands, we have
255 to assume it can alias anything. */
256 if (set1
== 0 || set2
== 0
257 /* If the two alias sets are the same, they may alias. */
261 /* See if the first alias set is a subset of the second. */
262 ase
= get_alias_set_entry (set1
);
264 && (ase
->has_zero_child
265 || splay_tree_lookup (ase
->children
,
266 (splay_tree_key
) set2
)))
269 /* Now do the same, but with the alias sets reversed. */
270 ase
= get_alias_set_entry (set2
);
272 && (ase
->has_zero_child
273 || splay_tree_lookup (ase
->children
,
274 (splay_tree_key
) set1
)))
277 /* The two alias sets are distinct and neither one is the
278 child of the other. Therefore, they cannot alias. */
282 /* Return 1 if TYPE is a RECORD_TYPE, UNION_TYPE, or QUAL_UNION_TYPE and has
283 has any readonly fields. If any of the fields have types that
284 contain readonly fields, return true as well. */
287 readonly_fields_p (type
)
292 if (TREE_CODE (type
) != RECORD_TYPE
&& TREE_CODE (type
) != UNION_TYPE
293 && TREE_CODE (type
) != QUAL_UNION_TYPE
)
296 for (field
= TYPE_FIELDS (type
); field
!= 0; field
= TREE_CHAIN (field
))
297 if (TREE_CODE (field
) == FIELD_DECL
298 && (TREE_READONLY (field
)
299 || readonly_fields_p (TREE_TYPE (field
))))
305 /* Return 1 if any MEM object of type T1 will always conflict (using the
306 dependency routines in this file) with any MEM object of type T2.
307 This is used when allocating temporary storage. If T1 and/or T2 are
308 NULL_TREE, it means we know nothing about the storage. */
311 objects_must_conflict_p (t1
, t2
)
314 /* If neither has a type specified, we don't know if they'll conflict
315 because we may be using them to store objects of various types, for
316 example the argument and local variables areas of inlined functions. */
317 if (t1
== 0 && t2
== 0)
320 /* If one or the other has readonly fields or is readonly,
321 then they may not conflict. */
322 if ((t1
!= 0 && readonly_fields_p (t1
))
323 || (t2
!= 0 && readonly_fields_p (t2
))
324 || (t1
!= 0 && TYPE_READONLY (t1
))
325 || (t2
!= 0 && TYPE_READONLY (t2
)))
328 /* If they are the same type, they must conflict. */
330 /* Likewise if both are volatile. */
331 || (t1
!= 0 && TYPE_VOLATILE (t1
) && t2
!= 0 && TYPE_VOLATILE (t2
)))
334 /* If one is aggregate and the other is scalar then they may not
336 if ((t1
!= 0 && AGGREGATE_TYPE_P (t1
))
337 != (t2
!= 0 && AGGREGATE_TYPE_P (t2
)))
340 /* Otherwise they conflict only if the alias sets conflict. */
341 return alias_sets_conflict_p (t1
? get_alias_set (t1
) : 0,
342 t2
? get_alias_set (t2
) : 0);
345 /* T is an expression with pointer type. Find the DECL on which this
346 expression is based. (For example, in `a[i]' this would be `a'.)
347 If there is no such DECL, or a unique decl cannot be determined,
348 NULL_TREE is retured. */
356 if (t
== 0 || t
== error_mark_node
|| ! POINTER_TYPE_P (TREE_TYPE (t
)))
359 /* If this is a declaration, return it. */
360 if (TREE_CODE_CLASS (TREE_CODE (t
)) == 'd')
363 /* Handle general expressions. It would be nice to deal with
364 COMPONENT_REFs here. If we could tell that `a' and `b' were the
365 same, then `a->f' and `b->f' are also the same. */
366 switch (TREE_CODE_CLASS (TREE_CODE (t
)))
369 return find_base_decl (TREE_OPERAND (t
, 0));
372 /* Return 0 if found in neither or both are the same. */
373 d0
= find_base_decl (TREE_OPERAND (t
, 0));
374 d1
= find_base_decl (TREE_OPERAND (t
, 1));
385 d0
= find_base_decl (TREE_OPERAND (t
, 0));
386 d1
= find_base_decl (TREE_OPERAND (t
, 1));
387 d0
= find_base_decl (TREE_OPERAND (t
, 0));
388 d2
= find_base_decl (TREE_OPERAND (t
, 2));
390 /* Set any nonzero values from the last, then from the first. */
391 if (d1
== 0) d1
= d2
;
392 if (d0
== 0) d0
= d1
;
393 if (d1
== 0) d1
= d0
;
394 if (d2
== 0) d2
= d1
;
396 /* At this point all are nonzero or all are zero. If all three are the
397 same, return it. Otherwise, return zero. */
398 return (d0
== d1
&& d1
== d2
) ? d0
: 0;
405 /* Return 1 if T is an expression that get_inner_reference handles. */
408 handled_component_p (t
)
411 switch (TREE_CODE (t
))
416 case NON_LVALUE_EXPR
:
421 return (TYPE_MODE (TREE_TYPE (t
))
422 == TYPE_MODE (TREE_TYPE (TREE_OPERAND (t
, 0))));
429 /* Return 1 if all the nested component references handled by
430 get_inner_reference in T are such that we can address the object in T. */
436 /* If we're at the end, it is vacuously addressable. */
437 if (! handled_component_p (t
))
440 /* Bitfields are never addressable. */
441 else if (TREE_CODE (t
) == BIT_FIELD_REF
)
444 else if (TREE_CODE (t
) == COMPONENT_REF
445 && ! DECL_NONADDRESSABLE_P (TREE_OPERAND (t
, 1))
446 && can_address_p (TREE_OPERAND (t
, 0)))
449 else if (TREE_CODE (t
) == ARRAY_REF
450 && ! TYPE_NONALIASED_COMPONENT (TREE_TYPE (TREE_OPERAND (t
, 0)))
451 && can_address_p (TREE_OPERAND (t
, 0)))
457 /* Return the alias set for T, which may be either a type or an
458 expression. Call language-specific routine for help, if needed. */
467 /* If we're not doing any alias analysis, just assume everything
468 aliases everything else. Also return 0 if this or its type is
470 if (! flag_strict_aliasing
|| t
== error_mark_node
472 && (TREE_TYPE (t
) == 0 || TREE_TYPE (t
) == error_mark_node
)))
475 /* We can be passed either an expression or a type. This and the
476 language-specific routine may make mutually-recursive calls to
477 each other to figure out what to do. At each juncture, we see if
478 this is a tree that the language may need to handle specially.
479 First handle things that aren't types and start by removing nops
480 since we care only about the actual object. */
483 while (TREE_CODE (t
) == NOP_EXPR
|| TREE_CODE (t
) == CONVERT_EXPR
484 || TREE_CODE (t
) == NON_LVALUE_EXPR
)
485 t
= TREE_OPERAND (t
, 0);
487 /* Now give the language a chance to do something but record what we
488 gave it this time. */
490 if ((set
= lang_get_alias_set (t
)) != -1)
493 /* Now loop the same way as get_inner_reference and get the alias
494 set to use. Pick up the outermost object that we could have
496 while (handled_component_p (t
) && ! can_address_p (t
))
497 t
= TREE_OPERAND (t
, 0);
499 if (TREE_CODE (t
) == INDIRECT_REF
)
501 /* Check for accesses through restrict-qualified pointers. */
502 tree decl
= find_base_decl (TREE_OPERAND (t
, 0));
504 if (decl
&& DECL_POINTER_ALIAS_SET_KNOWN_P (decl
))
505 /* We use the alias set indicated in the declaration. */
506 return DECL_POINTER_ALIAS_SET (decl
);
508 /* If we have an INDIRECT_REF via a void pointer, we don't
509 know anything about what that might alias. */
510 if (TREE_CODE (TREE_TYPE (t
)) == VOID_TYPE
)
514 /* Give the language another chance to do something special. */
516 && (set
= lang_get_alias_set (t
)) != -1)
519 /* Now all we care about is the type. */
523 /* Variant qualifiers don't affect the alias set, so get the main
524 variant. If this is a type with a known alias set, return it. */
525 t
= TYPE_MAIN_VARIANT (t
);
526 if (TYPE_P (t
) && TYPE_ALIAS_SET_KNOWN_P (t
))
527 return TYPE_ALIAS_SET (t
);
529 /* See if the language has special handling for this type. */
530 if ((set
= lang_get_alias_set (t
)) != -1)
532 /* If the alias set is now known, we are done. */
533 if (TYPE_ALIAS_SET_KNOWN_P (t
))
534 return TYPE_ALIAS_SET (t
);
537 /* There are no objects of FUNCTION_TYPE, so there's no point in
538 using up an alias set for them. (There are, of course, pointers
539 and references to functions, but that's different.) */
540 else if (TREE_CODE (t
) == FUNCTION_TYPE
)
543 /* Otherwise make a new alias set for this type. */
544 set
= new_alias_set ();
546 TYPE_ALIAS_SET (t
) = set
;
548 /* If this is an aggregate type, we must record any component aliasing
550 if (AGGREGATE_TYPE_P (t
) || TREE_CODE (t
) == COMPLEX_TYPE
)
551 record_component_aliases (t
);
556 /* Return a brand-new alias set. */
561 static HOST_WIDE_INT last_alias_set
;
563 if (flag_strict_aliasing
)
564 return ++last_alias_set
;
569 /* Indicate that things in SUBSET can alias things in SUPERSET, but
570 not vice versa. For example, in C, a store to an `int' can alias a
571 structure containing an `int', but not vice versa. Here, the
572 structure would be the SUPERSET and `int' the SUBSET. This
573 function should be called only once per SUPERSET/SUBSET pair.
575 It is illegal for SUPERSET to be zero; everything is implicitly a
576 subset of alias set zero. */
579 record_alias_subset (superset
, subset
)
580 HOST_WIDE_INT superset
;
581 HOST_WIDE_INT subset
;
583 alias_set_entry superset_entry
;
584 alias_set_entry subset_entry
;
589 superset_entry
= get_alias_set_entry (superset
);
590 if (superset_entry
== 0)
592 /* Create an entry for the SUPERSET, so that we have a place to
593 attach the SUBSET. */
595 = (alias_set_entry
) xmalloc (sizeof (struct alias_set_entry
));
596 superset_entry
->alias_set
= superset
;
597 superset_entry
->children
598 = splay_tree_new (splay_tree_compare_ints
, 0, 0);
599 superset_entry
->has_zero_child
= 0;
600 splay_tree_insert (alias_sets
, (splay_tree_key
) superset
,
601 (splay_tree_value
) superset_entry
);
605 superset_entry
->has_zero_child
= 1;
608 subset_entry
= get_alias_set_entry (subset
);
609 /* If there is an entry for the subset, enter all of its children
610 (if they are not already present) as children of the SUPERSET. */
613 if (subset_entry
->has_zero_child
)
614 superset_entry
->has_zero_child
= 1;
616 splay_tree_foreach (subset_entry
->children
, insert_subset_children
,
617 superset_entry
->children
);
620 /* Enter the SUBSET itself as a child of the SUPERSET. */
621 splay_tree_insert (superset_entry
->children
,
622 (splay_tree_key
) subset
, 0);
626 /* Record that component types of TYPE, if any, are part of that type for
627 aliasing purposes. For record types, we only record component types
628 for fields that are marked addressable. For array types, we always
629 record the component types, so the front end should not call this
630 function if the individual component aren't addressable. */
633 record_component_aliases (type
)
636 HOST_WIDE_INT superset
= get_alias_set (type
);
642 switch (TREE_CODE (type
))
645 if (! TYPE_NONALIASED_COMPONENT (type
))
646 record_alias_subset (superset
, get_alias_set (TREE_TYPE (type
)));
651 case QUAL_UNION_TYPE
:
652 for (field
= TYPE_FIELDS (type
); field
!= 0; field
= TREE_CHAIN (field
))
653 if (TREE_CODE (field
) == FIELD_DECL
&& ! DECL_NONADDRESSABLE_P (field
))
654 record_alias_subset (superset
, get_alias_set (TREE_TYPE (field
)));
658 record_alias_subset (superset
, get_alias_set (TREE_TYPE (type
)));
666 /* Allocate an alias set for use in storing and reading from the varargs
670 get_varargs_alias_set ()
672 static HOST_WIDE_INT set
= -1;
675 set
= new_alias_set ();
680 /* Likewise, but used for the fixed portions of the frame, e.g., register
684 get_frame_alias_set ()
686 static HOST_WIDE_INT set
= -1;
689 set
= new_alias_set ();
694 /* Inside SRC, the source of a SET, find a base address. */
697 find_base_value (src
)
701 switch (GET_CODE (src
))
709 /* At the start of a function, argument registers have known base
710 values which may be lost later. Returning an ADDRESS
711 expression here allows optimization based on argument values
712 even when the argument registers are used for other purposes. */
713 if (regno
< FIRST_PSEUDO_REGISTER
&& copying_arguments
)
714 return new_reg_base_value
[regno
];
716 /* If a pseudo has a known base value, return it. Do not do this
717 for hard regs since it can result in a circular dependency
718 chain for registers which have values at function entry.
720 The test above is not sufficient because the scheduler may move
721 a copy out of an arg reg past the NOTE_INSN_FUNCTION_BEGIN. */
722 if (regno
>= FIRST_PSEUDO_REGISTER
723 && regno
< reg_base_value_size
724 && reg_base_value
[regno
])
725 return reg_base_value
[regno
];
730 /* Check for an argument passed in memory. Only record in the
731 copying-arguments block; it is too hard to track changes
733 if (copying_arguments
734 && (XEXP (src
, 0) == arg_pointer_rtx
735 || (GET_CODE (XEXP (src
, 0)) == PLUS
736 && XEXP (XEXP (src
, 0), 0) == arg_pointer_rtx
)))
737 return gen_rtx_ADDRESS (VOIDmode
, src
);
742 if (GET_CODE (src
) != PLUS
&& GET_CODE (src
) != MINUS
)
745 /* ... fall through ... */
750 rtx temp
, src_0
= XEXP (src
, 0), src_1
= XEXP (src
, 1);
752 /* If either operand is a REG, then see if we already have
753 a known value for it. */
754 if (GET_CODE (src_0
) == REG
)
756 temp
= find_base_value (src_0
);
761 if (GET_CODE (src_1
) == REG
)
763 temp
= find_base_value (src_1
);
768 /* Guess which operand is the base address:
769 If either operand is a symbol, then it is the base. If
770 either operand is a CONST_INT, then the other is the base. */
771 if (GET_CODE (src_1
) == CONST_INT
|| CONSTANT_P (src_0
))
772 return find_base_value (src_0
);
773 else if (GET_CODE (src_0
) == CONST_INT
|| CONSTANT_P (src_1
))
774 return find_base_value (src_1
);
776 /* This might not be necessary anymore:
777 If either operand is a REG that is a known pointer, then it
779 else if (GET_CODE (src_0
) == REG
&& REG_POINTER (src_0
))
780 return find_base_value (src_0
);
781 else if (GET_CODE (src_1
) == REG
&& REG_POINTER (src_1
))
782 return find_base_value (src_1
);
788 /* The standard form is (lo_sum reg sym) so look only at the
790 return find_base_value (XEXP (src
, 1));
793 /* If the second operand is constant set the base
794 address to the first operand. */
795 if (GET_CODE (XEXP (src
, 1)) == CONST_INT
&& INTVAL (XEXP (src
, 1)) != 0)
796 return find_base_value (XEXP (src
, 0));
800 if (GET_MODE_SIZE (GET_MODE (src
)) < GET_MODE_SIZE (Pmode
))
804 case SIGN_EXTEND
: /* used for NT/Alpha pointers */
806 return find_base_value (XEXP (src
, 0));
815 /* Called from init_alias_analysis indirectly through note_stores. */
817 /* While scanning insns to find base values, reg_seen[N] is nonzero if
818 register N has been set in this function. */
819 static char *reg_seen
;
821 /* Addresses which are known not to alias anything else are identified
822 by a unique integer. */
823 static int unique_id
;
826 record_set (dest
, set
, data
)
828 void *data ATTRIBUTE_UNUSED
;
830 register unsigned regno
;
833 if (GET_CODE (dest
) != REG
)
836 regno
= REGNO (dest
);
838 if (regno
>= reg_base_value_size
)
843 /* A CLOBBER wipes out any old value but does not prevent a previously
844 unset register from acquiring a base address (i.e. reg_seen is not
846 if (GET_CODE (set
) == CLOBBER
)
848 new_reg_base_value
[regno
] = 0;
857 new_reg_base_value
[regno
] = 0;
861 new_reg_base_value
[regno
] = gen_rtx_ADDRESS (Pmode
,
862 GEN_INT (unique_id
++));
866 /* This is not the first set. If the new value is not related to the
867 old value, forget the base value. Note that the following code is
869 extern int x, y; int *p = &x; p += (&y-&x);
870 ANSI C does not allow computing the difference of addresses
871 of distinct top level objects. */
872 if (new_reg_base_value
[regno
])
873 switch (GET_CODE (src
))
877 if (XEXP (src
, 0) != dest
&& XEXP (src
, 1) != dest
)
878 new_reg_base_value
[regno
] = 0;
881 /* If the value we add in the PLUS is also a valid base value,
882 this might be the actual base value, and the original value
885 rtx other
= NULL_RTX
;
887 if (XEXP (src
, 0) == dest
)
888 other
= XEXP (src
, 1);
889 else if (XEXP (src
, 1) == dest
)
890 other
= XEXP (src
, 0);
892 if (! other
|| find_base_value (other
))
893 new_reg_base_value
[regno
] = 0;
897 if (XEXP (src
, 0) != dest
|| GET_CODE (XEXP (src
, 1)) != CONST_INT
)
898 new_reg_base_value
[regno
] = 0;
901 new_reg_base_value
[regno
] = 0;
904 /* If this is the first set of a register, record the value. */
905 else if ((regno
>= FIRST_PSEUDO_REGISTER
|| ! fixed_regs
[regno
])
906 && ! reg_seen
[regno
] && new_reg_base_value
[regno
] == 0)
907 new_reg_base_value
[regno
] = find_base_value (src
);
912 /* Called from loop optimization when a new pseudo-register is
913 created. It indicates that REGNO is being set to VAL. f INVARIANT
914 is true then this value also describes an invariant relationship
915 which can be used to deduce that two registers with unknown values
919 record_base_value (regno
, val
, invariant
)
924 if (regno
>= reg_base_value_size
)
927 if (invariant
&& alias_invariant
)
928 alias_invariant
[regno
] = val
;
930 if (GET_CODE (val
) == REG
)
932 if (REGNO (val
) < reg_base_value_size
)
933 reg_base_value
[regno
] = reg_base_value
[REGNO (val
)];
938 reg_base_value
[regno
] = find_base_value (val
);
941 /* Returns a canonical version of X, from the point of view alias
942 analysis. (For example, if X is a MEM whose address is a register,
943 and the register has a known value (say a SYMBOL_REF), then a MEM
944 whose address is the SYMBOL_REF is returned.) */
950 /* Recursively look for equivalences. */
951 if (GET_CODE (x
) == REG
&& REGNO (x
) >= FIRST_PSEUDO_REGISTER
952 && REGNO (x
) < reg_known_value_size
)
953 return reg_known_value
[REGNO (x
)] == x
954 ? x
: canon_rtx (reg_known_value
[REGNO (x
)]);
955 else if (GET_CODE (x
) == PLUS
)
957 rtx x0
= canon_rtx (XEXP (x
, 0));
958 rtx x1
= canon_rtx (XEXP (x
, 1));
960 if (x0
!= XEXP (x
, 0) || x1
!= XEXP (x
, 1))
962 /* We can tolerate LO_SUMs being offset here; these
963 rtl are used for nothing other than comparisons. */
964 if (GET_CODE (x0
) == CONST_INT
)
965 return plus_constant_for_output (x1
, INTVAL (x0
));
966 else if (GET_CODE (x1
) == CONST_INT
)
967 return plus_constant_for_output (x0
, INTVAL (x1
));
968 return gen_rtx_PLUS (GET_MODE (x
), x0
, x1
);
972 /* This gives us much better alias analysis when called from
973 the loop optimizer. Note we want to leave the original
974 MEM alone, but need to return the canonicalized MEM with
975 all the flags with their original values. */
976 else if (GET_CODE (x
) == MEM
)
978 rtx addr
= canon_rtx (XEXP (x
, 0));
980 if (addr
!= XEXP (x
, 0))
982 rtx
new = gen_rtx_MEM (GET_MODE (x
), addr
);
984 MEM_COPY_ATTRIBUTES (new, x
);
991 /* Return 1 if X and Y are identical-looking rtx's.
993 We use the data in reg_known_value above to see if two registers with
994 different numbers are, in fact, equivalent. */
997 rtx_equal_for_memref_p (x
, y
)
1002 register enum rtx_code code
;
1003 register const char *fmt
;
1005 if (x
== 0 && y
== 0)
1007 if (x
== 0 || y
== 0)
1016 code
= GET_CODE (x
);
1017 /* Rtx's of different codes cannot be equal. */
1018 if (code
!= GET_CODE (y
))
1021 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent.
1022 (REG:SI x) and (REG:HI x) are NOT equivalent. */
1024 if (GET_MODE (x
) != GET_MODE (y
))
1027 /* Some RTL can be compared without a recursive examination. */
1031 return REGNO (x
) == REGNO (y
);
1034 return XEXP (x
, 0) == XEXP (y
, 0);
1037 return XSTR (x
, 0) == XSTR (y
, 0);
1041 /* There's no need to compare the contents of CONST_DOUBLEs or
1042 CONST_INTs because pointer equality is a good enough
1043 comparison for these nodes. */
1047 return (XINT (x
, 1) == XINT (y
, 1)
1048 && rtx_equal_for_memref_p (XEXP (x
, 0), XEXP (y
, 0)));
1054 /* For commutative operations, the RTX match if the operand match in any
1055 order. Also handle the simple binary and unary cases without a loop. */
1056 if (code
== EQ
|| code
== NE
|| GET_RTX_CLASS (code
) == 'c')
1057 return ((rtx_equal_for_memref_p (XEXP (x
, 0), XEXP (y
, 0))
1058 && rtx_equal_for_memref_p (XEXP (x
, 1), XEXP (y
, 1)))
1059 || (rtx_equal_for_memref_p (XEXP (x
, 0), XEXP (y
, 1))
1060 && rtx_equal_for_memref_p (XEXP (x
, 1), XEXP (y
, 0))));
1061 else if (GET_RTX_CLASS (code
) == '<' || GET_RTX_CLASS (code
) == '2')
1062 return (rtx_equal_for_memref_p (XEXP (x
, 0), XEXP (y
, 0))
1063 && rtx_equal_for_memref_p (XEXP (x
, 1), XEXP (y
, 1)));
1064 else if (GET_RTX_CLASS (code
) == '1')
1065 return rtx_equal_for_memref_p (XEXP (x
, 0), XEXP (y
, 0));
1067 /* Compare the elements. If any pair of corresponding elements
1068 fail to match, return 0 for the whole things.
1070 Limit cases to types which actually appear in addresses. */
1072 fmt
= GET_RTX_FORMAT (code
);
1073 for (i
= GET_RTX_LENGTH (code
) - 1; i
>= 0; i
--)
1078 if (XINT (x
, i
) != XINT (y
, i
))
1083 /* Two vectors must have the same length. */
1084 if (XVECLEN (x
, i
) != XVECLEN (y
, i
))
1087 /* And the corresponding elements must match. */
1088 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
1089 if (rtx_equal_for_memref_p (XVECEXP (x
, i
, j
),
1090 XVECEXP (y
, i
, j
)) == 0)
1095 if (rtx_equal_for_memref_p (XEXP (x
, i
), XEXP (y
, i
)) == 0)
1099 /* This can happen for an asm which clobbers memory. */
1103 /* It is believed that rtx's at this level will never
1104 contain anything but integers and other rtx's,
1105 except for within LABEL_REFs and SYMBOL_REFs. */
1113 /* Given an rtx X, find a SYMBOL_REF or LABEL_REF within
1114 X and return it, or return 0 if none found. */
1117 find_symbolic_term (x
)
1121 register enum rtx_code code
;
1122 register const char *fmt
;
1124 code
= GET_CODE (x
);
1125 if (code
== SYMBOL_REF
|| code
== LABEL_REF
)
1127 if (GET_RTX_CLASS (code
) == 'o')
1130 fmt
= GET_RTX_FORMAT (code
);
1131 for (i
= GET_RTX_LENGTH (code
) - 1; i
>= 0; i
--)
1137 t
= find_symbolic_term (XEXP (x
, i
));
1141 else if (fmt
[i
] == 'E')
1152 struct elt_loc_list
*l
;
1154 #if defined (FIND_BASE_TERM)
1155 /* Try machine-dependent ways to find the base term. */
1156 x
= FIND_BASE_TERM (x
);
1159 switch (GET_CODE (x
))
1162 return REG_BASE_VALUE (x
);
1165 case SIGN_EXTEND
: /* Used for Alpha/NT pointers */
1171 return find_base_term (XEXP (x
, 0));
1174 val
= CSELIB_VAL_PTR (x
);
1175 for (l
= val
->locs
; l
; l
= l
->next
)
1176 if ((x
= find_base_term (l
->loc
)) != 0)
1182 if (GET_CODE (x
) != PLUS
&& GET_CODE (x
) != MINUS
)
1189 rtx tmp1
= XEXP (x
, 0);
1190 rtx tmp2
= XEXP (x
, 1);
1192 /* This is a litle bit tricky since we have to determine which of
1193 the two operands represents the real base address. Otherwise this
1194 routine may return the index register instead of the base register.
1196 That may cause us to believe no aliasing was possible, when in
1197 fact aliasing is possible.
1199 We use a few simple tests to guess the base register. Additional
1200 tests can certainly be added. For example, if one of the operands
1201 is a shift or multiply, then it must be the index register and the
1202 other operand is the base register. */
1204 if (tmp1
== pic_offset_table_rtx
&& CONSTANT_P (tmp2
))
1205 return find_base_term (tmp2
);
1207 /* If either operand is known to be a pointer, then use it
1208 to determine the base term. */
1209 if (REG_P (tmp1
) && REG_POINTER (tmp1
))
1210 return find_base_term (tmp1
);
1212 if (REG_P (tmp2
) && REG_POINTER (tmp2
))
1213 return find_base_term (tmp2
);
1215 /* Neither operand was known to be a pointer. Go ahead and find the
1216 base term for both operands. */
1217 tmp1
= find_base_term (tmp1
);
1218 tmp2
= find_base_term (tmp2
);
1220 /* If either base term is named object or a special address
1221 (like an argument or stack reference), then use it for the
1224 && (GET_CODE (tmp1
) == SYMBOL_REF
1225 || GET_CODE (tmp1
) == LABEL_REF
1226 || (GET_CODE (tmp1
) == ADDRESS
1227 && GET_MODE (tmp1
) != VOIDmode
)))
1231 && (GET_CODE (tmp2
) == SYMBOL_REF
1232 || GET_CODE (tmp2
) == LABEL_REF
1233 || (GET_CODE (tmp2
) == ADDRESS
1234 && GET_MODE (tmp2
) != VOIDmode
)))
1237 /* We could not determine which of the two operands was the
1238 base register and which was the index. So we can determine
1239 nothing from the base alias check. */
1244 if (GET_CODE (XEXP (x
, 0)) == REG
&& GET_CODE (XEXP (x
, 1)) == CONST_INT
)
1245 return REG_BASE_VALUE (XEXP (x
, 0));
1253 return REG_BASE_VALUE (frame_pointer_rtx
);
1260 /* Return 0 if the addresses X and Y are known to point to different
1261 objects, 1 if they might be pointers to the same object. */
1264 base_alias_check (x
, y
, x_mode
, y_mode
)
1266 enum machine_mode x_mode
, y_mode
;
1268 rtx x_base
= find_base_term (x
);
1269 rtx y_base
= find_base_term (y
);
1271 /* If the address itself has no known base see if a known equivalent
1272 value has one. If either address still has no known base, nothing
1273 is known about aliasing. */
1278 if (! flag_expensive_optimizations
|| (x_c
= canon_rtx (x
)) == x
)
1281 x_base
= find_base_term (x_c
);
1289 if (! flag_expensive_optimizations
|| (y_c
= canon_rtx (y
)) == y
)
1292 y_base
= find_base_term (y_c
);
1297 /* If the base addresses are equal nothing is known about aliasing. */
1298 if (rtx_equal_p (x_base
, y_base
))
1301 /* The base addresses of the read and write are different expressions.
1302 If they are both symbols and they are not accessed via AND, there is
1303 no conflict. We can bring knowledge of object alignment into play
1304 here. For example, on alpha, "char a, b;" can alias one another,
1305 though "char a; long b;" cannot. */
1306 if (GET_CODE (x_base
) != ADDRESS
&& GET_CODE (y_base
) != ADDRESS
)
1308 if (GET_CODE (x
) == AND
&& GET_CODE (y
) == AND
)
1310 if (GET_CODE (x
) == AND
1311 && (GET_CODE (XEXP (x
, 1)) != CONST_INT
1312 || GET_MODE_UNIT_SIZE (y_mode
) < -INTVAL (XEXP (x
, 1))))
1314 if (GET_CODE (y
) == AND
1315 && (GET_CODE (XEXP (y
, 1)) != CONST_INT
1316 || GET_MODE_UNIT_SIZE (x_mode
) < -INTVAL (XEXP (y
, 1))))
1318 /* Differing symbols never alias. */
1322 /* If one address is a stack reference there can be no alias:
1323 stack references using different base registers do not alias,
1324 a stack reference can not alias a parameter, and a stack reference
1325 can not alias a global. */
1326 if ((GET_CODE (x_base
) == ADDRESS
&& GET_MODE (x_base
) == Pmode
)
1327 || (GET_CODE (y_base
) == ADDRESS
&& GET_MODE (y_base
) == Pmode
))
1330 if (! flag_argument_noalias
)
1333 if (flag_argument_noalias
> 1)
1336 /* Weak noalias assertion (arguments are distinct, but may match globals). */
1337 return ! (GET_MODE (x_base
) == VOIDmode
&& GET_MODE (y_base
) == VOIDmode
);
1340 /* Convert the address X into something we can use. This is done by returning
1341 it unchanged unless it is a value; in the latter case we call cselib to get
1342 a more useful rtx. */
1349 struct elt_loc_list
*l
;
1351 if (GET_CODE (x
) != VALUE
)
1353 v
= CSELIB_VAL_PTR (x
);
1354 for (l
= v
->locs
; l
; l
= l
->next
)
1355 if (CONSTANT_P (l
->loc
))
1357 for (l
= v
->locs
; l
; l
= l
->next
)
1358 if (GET_CODE (l
->loc
) != REG
&& GET_CODE (l
->loc
) != MEM
)
1361 return v
->locs
->loc
;
1365 /* Return the address of the (N_REFS + 1)th memory reference to ADDR
1366 where SIZE is the size in bytes of the memory reference. If ADDR
1367 is not modified by the memory reference then ADDR is returned. */
1370 addr_side_effect_eval (addr
, size
, n_refs
)
1377 switch (GET_CODE (addr
))
1380 offset
= (n_refs
+ 1) * size
;
1383 offset
= -(n_refs
+ 1) * size
;
1386 offset
= n_refs
* size
;
1389 offset
= -n_refs
* size
;
1397 addr
= gen_rtx_PLUS (GET_MODE (addr
), XEXP (addr
, 0), GEN_INT (offset
));
1399 addr
= XEXP (addr
, 0);
1404 /* Return nonzero if X and Y (memory addresses) could reference the
1405 same location in memory. C is an offset accumulator. When
1406 C is nonzero, we are testing aliases between X and Y + C.
1407 XSIZE is the size in bytes of the X reference,
1408 similarly YSIZE is the size in bytes for Y.
1410 If XSIZE or YSIZE is zero, we do not know the amount of memory being
1411 referenced (the reference was BLKmode), so make the most pessimistic
1414 If XSIZE or YSIZE is negative, we may access memory outside the object
1415 being referenced as a side effect. This can happen when using AND to
1416 align memory references, as is done on the Alpha.
1418 Nice to notice that varying addresses cannot conflict with fp if no
1419 local variables had their addresses taken, but that's too hard now. */
1422 memrefs_conflict_p (xsize
, x
, ysize
, y
, c
)
1427 if (GET_CODE (x
) == VALUE
)
1429 if (GET_CODE (y
) == VALUE
)
1431 if (GET_CODE (x
) == HIGH
)
1433 else if (GET_CODE (x
) == LO_SUM
)
1436 x
= canon_rtx (addr_side_effect_eval (x
, xsize
, 0));
1437 if (GET_CODE (y
) == HIGH
)
1439 else if (GET_CODE (y
) == LO_SUM
)
1442 y
= canon_rtx (addr_side_effect_eval (y
, ysize
, 0));
1444 if (rtx_equal_for_memref_p (x
, y
))
1446 if (xsize
<= 0 || ysize
<= 0)
1448 if (c
>= 0 && xsize
> c
)
1450 if (c
< 0 && ysize
+c
> 0)
1455 /* This code used to check for conflicts involving stack references and
1456 globals but the base address alias code now handles these cases. */
1458 if (GET_CODE (x
) == PLUS
)
1460 /* The fact that X is canonicalized means that this
1461 PLUS rtx is canonicalized. */
1462 rtx x0
= XEXP (x
, 0);
1463 rtx x1
= XEXP (x
, 1);
1465 if (GET_CODE (y
) == PLUS
)
1467 /* The fact that Y is canonicalized means that this
1468 PLUS rtx is canonicalized. */
1469 rtx y0
= XEXP (y
, 0);
1470 rtx y1
= XEXP (y
, 1);
1472 if (rtx_equal_for_memref_p (x1
, y1
))
1473 return memrefs_conflict_p (xsize
, x0
, ysize
, y0
, c
);
1474 if (rtx_equal_for_memref_p (x0
, y0
))
1475 return memrefs_conflict_p (xsize
, x1
, ysize
, y1
, c
);
1476 if (GET_CODE (x1
) == CONST_INT
)
1478 if (GET_CODE (y1
) == CONST_INT
)
1479 return memrefs_conflict_p (xsize
, x0
, ysize
, y0
,
1480 c
- INTVAL (x1
) + INTVAL (y1
));
1482 return memrefs_conflict_p (xsize
, x0
, ysize
, y
,
1485 else if (GET_CODE (y1
) == CONST_INT
)
1486 return memrefs_conflict_p (xsize
, x
, ysize
, y0
, c
+ INTVAL (y1
));
1490 else if (GET_CODE (x1
) == CONST_INT
)
1491 return memrefs_conflict_p (xsize
, x0
, ysize
, y
, c
- INTVAL (x1
));
1493 else if (GET_CODE (y
) == PLUS
)
1495 /* The fact that Y is canonicalized means that this
1496 PLUS rtx is canonicalized. */
1497 rtx y0
= XEXP (y
, 0);
1498 rtx y1
= XEXP (y
, 1);
1500 if (GET_CODE (y1
) == CONST_INT
)
1501 return memrefs_conflict_p (xsize
, x
, ysize
, y0
, c
+ INTVAL (y1
));
1506 if (GET_CODE (x
) == GET_CODE (y
))
1507 switch (GET_CODE (x
))
1511 /* Handle cases where we expect the second operands to be the
1512 same, and check only whether the first operand would conflict
1515 rtx x1
= canon_rtx (XEXP (x
, 1));
1516 rtx y1
= canon_rtx (XEXP (y
, 1));
1517 if (! rtx_equal_for_memref_p (x1
, y1
))
1519 x0
= canon_rtx (XEXP (x
, 0));
1520 y0
= canon_rtx (XEXP (y
, 0));
1521 if (rtx_equal_for_memref_p (x0
, y0
))
1522 return (xsize
== 0 || ysize
== 0
1523 || (c
>= 0 && xsize
> c
) || (c
< 0 && ysize
+c
> 0));
1525 /* Can't properly adjust our sizes. */
1526 if (GET_CODE (x1
) != CONST_INT
)
1528 xsize
/= INTVAL (x1
);
1529 ysize
/= INTVAL (x1
);
1531 return memrefs_conflict_p (xsize
, x0
, ysize
, y0
, c
);
1535 /* Are these registers known not to be equal? */
1536 if (alias_invariant
)
1538 unsigned int r_x
= REGNO (x
), r_y
= REGNO (y
);
1539 rtx i_x
, i_y
; /* invariant relationships of X and Y */
1541 i_x
= r_x
>= reg_base_value_size
? 0 : alias_invariant
[r_x
];
1542 i_y
= r_y
>= reg_base_value_size
? 0 : alias_invariant
[r_y
];
1544 if (i_x
== 0 && i_y
== 0)
1547 if (! memrefs_conflict_p (xsize
, i_x
? i_x
: x
,
1548 ysize
, i_y
? i_y
: y
, c
))
1557 /* Treat an access through an AND (e.g. a subword access on an Alpha)
1558 as an access with indeterminate size. Assume that references
1559 besides AND are aligned, so if the size of the other reference is
1560 at least as large as the alignment, assume no other overlap. */
1561 if (GET_CODE (x
) == AND
&& GET_CODE (XEXP (x
, 1)) == CONST_INT
)
1563 if (GET_CODE (y
) == AND
|| ysize
< -INTVAL (XEXP (x
, 1)))
1565 return memrefs_conflict_p (xsize
, XEXP (x
, 0), ysize
, y
, c
);
1567 if (GET_CODE (y
) == AND
&& GET_CODE (XEXP (y
, 1)) == CONST_INT
)
1569 /* ??? If we are indexing far enough into the array/structure, we
1570 may yet be able to determine that we can not overlap. But we
1571 also need to that we are far enough from the end not to overlap
1572 a following reference, so we do nothing with that for now. */
1573 if (GET_CODE (x
) == AND
|| xsize
< -INTVAL (XEXP (y
, 1)))
1575 return memrefs_conflict_p (xsize
, x
, ysize
, XEXP (y
, 0), c
);
1578 if (GET_CODE (x
) == ADDRESSOF
)
1580 if (y
== frame_pointer_rtx
1581 || GET_CODE (y
) == ADDRESSOF
)
1582 return xsize
<= 0 || ysize
<= 0;
1584 if (GET_CODE (y
) == ADDRESSOF
)
1586 if (x
== frame_pointer_rtx
)
1587 return xsize
<= 0 || ysize
<= 0;
1592 if (GET_CODE (x
) == CONST_INT
&& GET_CODE (y
) == CONST_INT
)
1594 c
+= (INTVAL (y
) - INTVAL (x
));
1595 return (xsize
<= 0 || ysize
<= 0
1596 || (c
>= 0 && xsize
> c
) || (c
< 0 && ysize
+c
> 0));
1599 if (GET_CODE (x
) == CONST
)
1601 if (GET_CODE (y
) == CONST
)
1602 return memrefs_conflict_p (xsize
, canon_rtx (XEXP (x
, 0)),
1603 ysize
, canon_rtx (XEXP (y
, 0)), c
);
1605 return memrefs_conflict_p (xsize
, canon_rtx (XEXP (x
, 0)),
1608 if (GET_CODE (y
) == CONST
)
1609 return memrefs_conflict_p (xsize
, x
, ysize
,
1610 canon_rtx (XEXP (y
, 0)), c
);
1613 return (xsize
<= 0 || ysize
<= 0
1614 || (rtx_equal_for_memref_p (x
, y
)
1615 && ((c
>= 0 && xsize
> c
) || (c
< 0 && ysize
+c
> 0))));
1622 /* Functions to compute memory dependencies.
1624 Since we process the insns in execution order, we can build tables
1625 to keep track of what registers are fixed (and not aliased), what registers
1626 are varying in known ways, and what registers are varying in unknown
1629 If both memory references are volatile, then there must always be a
1630 dependence between the two references, since their order can not be
1631 changed. A volatile and non-volatile reference can be interchanged
1634 A MEM_IN_STRUCT reference at a non-AND varying address can never
1635 conflict with a non-MEM_IN_STRUCT reference at a fixed address. We
1636 also must allow AND addresses, because they may generate accesses
1637 outside the object being referenced. This is used to generate
1638 aligned addresses from unaligned addresses, for instance, the alpha
1639 storeqi_unaligned pattern. */
1641 /* Read dependence: X is read after read in MEM takes place. There can
1642 only be a dependence here if both reads are volatile. */
1645 read_dependence (mem
, x
)
1649 return MEM_VOLATILE_P (x
) && MEM_VOLATILE_P (mem
);
1652 /* Returns MEM1 if and only if MEM1 is a scalar at a fixed address and
1653 MEM2 is a reference to a structure at a varying address, or returns
1654 MEM2 if vice versa. Otherwise, returns NULL_RTX. If a non-NULL
1655 value is returned MEM1 and MEM2 can never alias. VARIES_P is used
1656 to decide whether or not an address may vary; it should return
1657 nonzero whenever variation is possible.
1658 MEM1_ADDR and MEM2_ADDR are the addresses of MEM1 and MEM2. */
1661 fixed_scalar_and_varying_struct_p (mem1
, mem2
, mem1_addr
, mem2_addr
, varies_p
)
1663 rtx mem1_addr
, mem2_addr
;
1664 int (*varies_p
) PARAMS ((rtx
, int));
1666 if (! flag_strict_aliasing
)
1669 if (MEM_SCALAR_P (mem1
) && MEM_IN_STRUCT_P (mem2
)
1670 && !varies_p (mem1_addr
, 1) && varies_p (mem2_addr
, 1))
1671 /* MEM1 is a scalar at a fixed address; MEM2 is a struct at a
1675 if (MEM_IN_STRUCT_P (mem1
) && MEM_SCALAR_P (mem2
)
1676 && varies_p (mem1_addr
, 1) && !varies_p (mem2_addr
, 1))
1677 /* MEM2 is a scalar at a fixed address; MEM1 is a struct at a
1684 /* Returns nonzero if something about the mode or address format MEM1
1685 indicates that it might well alias *anything*. */
1688 aliases_everything_p (mem
)
1691 if (GET_CODE (XEXP (mem
, 0)) == AND
)
1692 /* If the address is an AND, its very hard to know at what it is
1693 actually pointing. */
1699 /* True dependence: X is read after store in MEM takes place. */
1702 true_dependence (mem
, mem_mode
, x
, varies
)
1704 enum machine_mode mem_mode
;
1706 int (*varies
) PARAMS ((rtx
, int));
1708 register rtx x_addr
, mem_addr
;
1711 if (MEM_VOLATILE_P (x
) && MEM_VOLATILE_P (mem
))
1714 if (DIFFERENT_ALIAS_SETS_P (x
, mem
))
1717 /* Unchanging memory can't conflict with non-unchanging memory.
1718 A non-unchanging read can conflict with a non-unchanging write.
1719 An unchanging read can conflict with an unchanging write since
1720 there may be a single store to this address to initialize it.
1721 Note that an unchanging store can conflict with a non-unchanging read
1722 since we have to make conservative assumptions when we have a
1723 record with readonly fields and we are copying the whole thing.
1724 Just fall through to the code below to resolve potential conflicts.
1725 This won't handle all cases optimally, but the possible performance
1726 loss should be negligible. */
1727 if (RTX_UNCHANGING_P (x
) && ! RTX_UNCHANGING_P (mem
))
1730 if (mem_mode
== VOIDmode
)
1731 mem_mode
= GET_MODE (mem
);
1733 x_addr
= get_addr (XEXP (x
, 0));
1734 mem_addr
= get_addr (XEXP (mem
, 0));
1736 base
= find_base_term (x_addr
);
1737 if (base
&& (GET_CODE (base
) == LABEL_REF
1738 || (GET_CODE (base
) == SYMBOL_REF
1739 && CONSTANT_POOL_ADDRESS_P (base
))))
1742 if (! base_alias_check (x_addr
, mem_addr
, GET_MODE (x
), mem_mode
))
1745 x_addr
= canon_rtx (x_addr
);
1746 mem_addr
= canon_rtx (mem_addr
);
1748 if (! memrefs_conflict_p (GET_MODE_SIZE (mem_mode
), mem_addr
,
1749 SIZE_FOR_MODE (x
), x_addr
, 0))
1752 if (aliases_everything_p (x
))
1755 /* We cannot use aliases_everyting_p to test MEM, since we must look
1756 at MEM_MODE, rather than GET_MODE (MEM). */
1757 if (mem_mode
== QImode
|| GET_CODE (mem_addr
) == AND
)
1760 /* In true_dependence we also allow BLKmode to alias anything. Why
1761 don't we do this in anti_dependence and output_dependence? */
1762 if (mem_mode
== BLKmode
|| GET_MODE (x
) == BLKmode
)
1765 return ! fixed_scalar_and_varying_struct_p (mem
, x
, mem_addr
, x_addr
,
1769 /* Returns non-zero if a write to X might alias a previous read from
1770 (or, if WRITEP is non-zero, a write to) MEM. */
1773 write_dependence_p (mem
, x
, writep
)
1778 rtx x_addr
, mem_addr
;
1782 if (MEM_VOLATILE_P (x
) && MEM_VOLATILE_P (mem
))
1785 if (DIFFERENT_ALIAS_SETS_P (x
, mem
))
1788 /* Unchanging memory can't conflict with non-unchanging memory. */
1789 if (RTX_UNCHANGING_P (x
) != RTX_UNCHANGING_P (mem
))
1792 /* If MEM is an unchanging read, then it can't possibly conflict with
1793 the store to X, because there is at most one store to MEM, and it must
1794 have occurred somewhere before MEM. */
1795 if (! writep
&& RTX_UNCHANGING_P (mem
))
1798 x_addr
= get_addr (XEXP (x
, 0));
1799 mem_addr
= get_addr (XEXP (mem
, 0));
1803 base
= find_base_term (mem_addr
);
1804 if (base
&& (GET_CODE (base
) == LABEL_REF
1805 || (GET_CODE (base
) == SYMBOL_REF
1806 && CONSTANT_POOL_ADDRESS_P (base
))))
1810 if (! base_alias_check (x_addr
, mem_addr
, GET_MODE (x
),
1814 x_addr
= canon_rtx (x_addr
);
1815 mem_addr
= canon_rtx (mem_addr
);
1817 if (!memrefs_conflict_p (SIZE_FOR_MODE (mem
), mem_addr
,
1818 SIZE_FOR_MODE (x
), x_addr
, 0))
1822 = fixed_scalar_and_varying_struct_p (mem
, x
, mem_addr
, x_addr
,
1825 return (!(fixed_scalar
== mem
&& !aliases_everything_p (x
))
1826 && !(fixed_scalar
== x
&& !aliases_everything_p (mem
)));
1829 /* Anti dependence: X is written after read in MEM takes place. */
1832 anti_dependence (mem
, x
)
1836 return write_dependence_p (mem
, x
, /*writep=*/0);
1839 /* Output dependence: X is written after store in MEM takes place. */
1842 output_dependence (mem
, x
)
1846 return write_dependence_p (mem
, x
, /*writep=*/1);
1849 /* Returns non-zero if X mentions something which is not
1850 local to the function and is not constant. */
1853 nonlocal_mentioned_p (x
)
1857 register RTX_CODE code
;
1860 code
= GET_CODE (x
);
1862 if (GET_RTX_CLASS (code
) == 'i')
1864 /* Constant functions can be constant if they don't use
1865 scratch memory used to mark function w/o side effects. */
1866 if (code
== CALL_INSN
&& CONST_CALL_P (x
))
1868 x
= CALL_INSN_FUNCTION_USAGE (x
);
1874 code
= GET_CODE (x
);
1880 if (GET_CODE (SUBREG_REG (x
)) == REG
)
1882 /* Global registers are not local. */
1883 if (REGNO (SUBREG_REG (x
)) < FIRST_PSEUDO_REGISTER
1884 && global_regs
[REGNO (SUBREG_REG (x
)) + SUBREG_WORD (x
)])
1892 /* Global registers are not local. */
1893 if (regno
< FIRST_PSEUDO_REGISTER
&& global_regs
[regno
])
1907 /* Constants in the function's constants pool are constant. */
1908 if (CONSTANT_POOL_ADDRESS_P (x
))
1913 /* Non-constant calls and recursion are not local. */
1917 /* Be overly conservative and consider any volatile memory
1918 reference as not local. */
1919 if (MEM_VOLATILE_P (x
))
1921 base
= find_base_term (XEXP (x
, 0));
1924 /* A Pmode ADDRESS could be a reference via the structure value
1925 address or static chain. Such memory references are nonlocal.
1927 Thus, we have to examine the contents of the ADDRESS to find
1928 out if this is a local reference or not. */
1929 if (GET_CODE (base
) == ADDRESS
1930 && GET_MODE (base
) == Pmode
1931 && (XEXP (base
, 0) == stack_pointer_rtx
1932 || XEXP (base
, 0) == arg_pointer_rtx
1933 #if HARD_FRAME_POINTER_REGNUM != FRAME_POINTER_REGNUM
1934 || XEXP (base
, 0) == hard_frame_pointer_rtx
1936 || XEXP (base
, 0) == frame_pointer_rtx
))
1938 /* Constants in the function's constant pool are constant. */
1939 if (GET_CODE (base
) == SYMBOL_REF
&& CONSTANT_POOL_ADDRESS_P (base
))
1944 case UNSPEC_VOLATILE
:
1949 if (MEM_VOLATILE_P (x
))
1958 /* Recursively scan the operands of this expression. */
1961 register const char *fmt
= GET_RTX_FORMAT (code
);
1964 for (i
= GET_RTX_LENGTH (code
) - 1; i
>= 0; i
--)
1966 if (fmt
[i
] == 'e' && XEXP (x
, i
))
1968 if (nonlocal_mentioned_p (XEXP (x
, i
)))
1971 else if (fmt
[i
] == 'E')
1974 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
1975 if (nonlocal_mentioned_p (XVECEXP (x
, i
, j
)))
1984 /* Return non-zero if a loop (natural or otherwise) is present.
1985 Inspired by Depth_First_Search_PP described in:
1987 Advanced Compiler Design and Implementation
1989 Morgan Kaufmann, 1997
1991 and heavily borrowed from flow_depth_first_order_compute. */
2004 /* Allocate the preorder and postorder number arrays. */
2005 pre
= (int *) xcalloc (n_basic_blocks
, sizeof (int));
2006 post
= (int *) xcalloc (n_basic_blocks
, sizeof (int));
2008 /* Allocate stack for back-tracking up CFG. */
2009 stack
= (edge
*) xmalloc ((n_basic_blocks
+ 1) * sizeof (edge
));
2012 /* Allocate bitmap to track nodes that have been visited. */
2013 visited
= sbitmap_alloc (n_basic_blocks
);
2015 /* None of the nodes in the CFG have been visited yet. */
2016 sbitmap_zero (visited
);
2018 /* Push the first edge on to the stack. */
2019 stack
[sp
++] = ENTRY_BLOCK_PTR
->succ
;
2027 /* Look at the edge on the top of the stack. */
2032 /* Check if the edge destination has been visited yet. */
2033 if (dest
!= EXIT_BLOCK_PTR
&& ! TEST_BIT (visited
, dest
->index
))
2035 /* Mark that we have visited the destination. */
2036 SET_BIT (visited
, dest
->index
);
2038 pre
[dest
->index
] = prenum
++;
2042 /* Since the DEST node has been visited for the first
2043 time, check its successors. */
2044 stack
[sp
++] = dest
->succ
;
2047 post
[dest
->index
] = postnum
++;
2051 if (dest
!= EXIT_BLOCK_PTR
2052 && pre
[src
->index
] >= pre
[dest
->index
]
2053 && post
[dest
->index
] == 0)
2056 if (! e
->succ_next
&& src
!= ENTRY_BLOCK_PTR
)
2057 post
[src
->index
] = postnum
++;
2060 stack
[sp
- 1] = e
->succ_next
;
2069 sbitmap_free (visited
);
2074 /* Mark the function if it is constant. */
2077 mark_constant_function ()
2080 int nonlocal_mentioned
;
2082 if (TREE_PUBLIC (current_function_decl
)
2083 || TREE_READONLY (current_function_decl
)
2084 || DECL_IS_PURE (current_function_decl
)
2085 || TREE_THIS_VOLATILE (current_function_decl
)
2086 || TYPE_MODE (TREE_TYPE (current_function_decl
)) == VOIDmode
)
2089 /* A loop might not return which counts as a side effect. */
2093 nonlocal_mentioned
= 0;
2095 init_alias_analysis ();
2097 /* Determine if this is a constant function. */
2099 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
2100 if (INSN_P (insn
) && nonlocal_mentioned_p (insn
))
2102 nonlocal_mentioned
= 1;
2106 end_alias_analysis ();
2108 /* Mark the function. */
2110 if (! nonlocal_mentioned
)
2111 TREE_READONLY (current_function_decl
) = 1;
2115 static HARD_REG_SET argument_registers
;
2122 #ifndef OUTGOING_REGNO
2123 #define OUTGOING_REGNO(N) N
2125 for (i
= 0; i
< FIRST_PSEUDO_REGISTER
; i
++)
2126 /* Check whether this register can hold an incoming pointer
2127 argument. FUNCTION_ARG_REGNO_P tests outgoing register
2128 numbers, so translate if necessary due to register windows. */
2129 if (FUNCTION_ARG_REGNO_P (OUTGOING_REGNO (i
))
2130 && HARD_REGNO_MODE_OK (i
, Pmode
))
2131 SET_HARD_REG_BIT (argument_registers
, i
);
2133 alias_sets
= splay_tree_new (splay_tree_compare_ints
, 0, 0);
2136 /* Initialize the aliasing machinery. Initialize the REG_KNOWN_VALUE
2140 init_alias_analysis ()
2142 int maxreg
= max_reg_num ();
2145 register unsigned int ui
;
2148 reg_known_value_size
= maxreg
;
2151 = (rtx
*) xcalloc ((maxreg
- FIRST_PSEUDO_REGISTER
), sizeof (rtx
))
2152 - FIRST_PSEUDO_REGISTER
;
2154 = (char*) xcalloc ((maxreg
- FIRST_PSEUDO_REGISTER
), sizeof (char))
2155 - FIRST_PSEUDO_REGISTER
;
2157 /* Overallocate reg_base_value to allow some growth during loop
2158 optimization. Loop unrolling can create a large number of
2160 reg_base_value_size
= maxreg
* 2;
2161 reg_base_value
= (rtx
*) xcalloc (reg_base_value_size
, sizeof (rtx
));
2162 ggc_add_rtx_root (reg_base_value
, reg_base_value_size
);
2164 new_reg_base_value
= (rtx
*) xmalloc (reg_base_value_size
* sizeof (rtx
));
2165 reg_seen
= (char *) xmalloc (reg_base_value_size
);
2166 if (! reload_completed
&& flag_unroll_loops
)
2168 /* ??? Why are we realloc'ing if we're just going to zero it? */
2169 alias_invariant
= (rtx
*)xrealloc (alias_invariant
,
2170 reg_base_value_size
* sizeof (rtx
));
2171 memset ((char *)alias_invariant
, 0, reg_base_value_size
* sizeof (rtx
));
2174 /* The basic idea is that each pass through this loop will use the
2175 "constant" information from the previous pass to propagate alias
2176 information through another level of assignments.
2178 This could get expensive if the assignment chains are long. Maybe
2179 we should throttle the number of iterations, possibly based on
2180 the optimization level or flag_expensive_optimizations.
2182 We could propagate more information in the first pass by making use
2183 of REG_N_SETS to determine immediately that the alias information
2184 for a pseudo is "constant".
2186 A program with an uninitialized variable can cause an infinite loop
2187 here. Instead of doing a full dataflow analysis to detect such problems
2188 we just cap the number of iterations for the loop.
2190 The state of the arrays for the set chain in question does not matter
2191 since the program has undefined behavior. */
2196 /* Assume nothing will change this iteration of the loop. */
2199 /* We want to assign the same IDs each iteration of this loop, so
2200 start counting from zero each iteration of the loop. */
2203 /* We're at the start of the funtion each iteration through the
2204 loop, so we're copying arguments. */
2205 copying_arguments
= 1;
2207 /* Wipe the potential alias information clean for this pass. */
2208 memset ((char *) new_reg_base_value
, 0, reg_base_value_size
* sizeof (rtx
));
2210 /* Wipe the reg_seen array clean. */
2211 memset ((char *) reg_seen
, 0, reg_base_value_size
);
2213 /* Mark all hard registers which may contain an address.
2214 The stack, frame and argument pointers may contain an address.
2215 An argument register which can hold a Pmode value may contain
2216 an address even if it is not in BASE_REGS.
2218 The address expression is VOIDmode for an argument and
2219 Pmode for other registers. */
2221 for (i
= 0; i
< FIRST_PSEUDO_REGISTER
; i
++)
2222 if (TEST_HARD_REG_BIT (argument_registers
, i
))
2223 new_reg_base_value
[i
] = gen_rtx_ADDRESS (VOIDmode
,
2224 gen_rtx_REG (Pmode
, i
));
2226 new_reg_base_value
[STACK_POINTER_REGNUM
]
2227 = gen_rtx_ADDRESS (Pmode
, stack_pointer_rtx
);
2228 new_reg_base_value
[ARG_POINTER_REGNUM
]
2229 = gen_rtx_ADDRESS (Pmode
, arg_pointer_rtx
);
2230 new_reg_base_value
[FRAME_POINTER_REGNUM
]
2231 = gen_rtx_ADDRESS (Pmode
, frame_pointer_rtx
);
2232 #if HARD_FRAME_POINTER_REGNUM != FRAME_POINTER_REGNUM
2233 new_reg_base_value
[HARD_FRAME_POINTER_REGNUM
]
2234 = gen_rtx_ADDRESS (Pmode
, hard_frame_pointer_rtx
);
2237 /* Walk the insns adding values to the new_reg_base_value array. */
2238 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
2244 #if defined (HAVE_prologue) || defined (HAVE_epilogue)
2245 /* The prologue/epilouge insns are not threaded onto the
2246 insn chain until after reload has completed. Thus,
2247 there is no sense wasting time checking if INSN is in
2248 the prologue/epilogue until after reload has completed. */
2249 if (reload_completed
2250 && prologue_epilogue_contains (insn
))
2254 /* If this insn has a noalias note, process it, Otherwise,
2255 scan for sets. A simple set will have no side effects
2256 which could change the base value of any other register. */
2258 if (GET_CODE (PATTERN (insn
)) == SET
2259 && REG_NOTES (insn
) != 0
2260 && find_reg_note (insn
, REG_NOALIAS
, NULL_RTX
))
2261 record_set (SET_DEST (PATTERN (insn
)), NULL_RTX
, NULL
);
2263 note_stores (PATTERN (insn
), record_set
, NULL
);
2265 set
= single_set (insn
);
2268 && GET_CODE (SET_DEST (set
)) == REG
2269 && REGNO (SET_DEST (set
)) >= FIRST_PSEUDO_REGISTER
)
2271 unsigned int regno
= REGNO (SET_DEST (set
));
2272 rtx src
= SET_SRC (set
);
2274 if (REG_NOTES (insn
) != 0
2275 && (((note
= find_reg_note (insn
, REG_EQUAL
, 0)) != 0
2276 && REG_N_SETS (regno
) == 1)
2277 || (note
= find_reg_note (insn
, REG_EQUIV
, NULL_RTX
)) != 0)
2278 && GET_CODE (XEXP (note
, 0)) != EXPR_LIST
2279 && ! reg_overlap_mentioned_p (SET_DEST (set
), XEXP (note
, 0)))
2281 reg_known_value
[regno
] = XEXP (note
, 0);
2282 reg_known_equiv_p
[regno
] = REG_NOTE_KIND (note
) == REG_EQUIV
;
2284 else if (REG_N_SETS (regno
) == 1
2285 && GET_CODE (src
) == PLUS
2286 && GET_CODE (XEXP (src
, 0)) == REG
2287 && REGNO (XEXP (src
, 0)) >= FIRST_PSEUDO_REGISTER
2288 && (reg_known_value
[REGNO (XEXP (src
, 0))])
2289 && GET_CODE (XEXP (src
, 1)) == CONST_INT
)
2291 rtx op0
= XEXP (src
, 0);
2292 if (reg_known_value
[REGNO (op0
)])
2293 op0
= reg_known_value
[REGNO (op0
)];
2294 reg_known_value
[regno
]
2295 = plus_constant_for_output (op0
,
2296 INTVAL (XEXP (src
, 1)));
2297 reg_known_equiv_p
[regno
] = 0;
2299 else if (REG_N_SETS (regno
) == 1
2300 && ! rtx_varies_p (src
, 1))
2302 reg_known_value
[regno
] = src
;
2303 reg_known_equiv_p
[regno
] = 0;
2307 else if (GET_CODE (insn
) == NOTE
2308 && NOTE_LINE_NUMBER (insn
) == NOTE_INSN_FUNCTION_BEG
)
2309 copying_arguments
= 0;
2312 /* Now propagate values from new_reg_base_value to reg_base_value. */
2313 for (ui
= 0; ui
< reg_base_value_size
; ui
++)
2315 if (new_reg_base_value
[ui
]
2316 && new_reg_base_value
[ui
] != reg_base_value
[ui
]
2317 && ! rtx_equal_p (new_reg_base_value
[ui
], reg_base_value
[ui
]))
2319 reg_base_value
[ui
] = new_reg_base_value
[ui
];
2324 while (changed
&& ++pass
< MAX_ALIAS_LOOP_PASSES
);
2326 /* Fill in the remaining entries. */
2327 for (i
= FIRST_PSEUDO_REGISTER
; i
< maxreg
; i
++)
2328 if (reg_known_value
[i
] == 0)
2329 reg_known_value
[i
] = regno_reg_rtx
[i
];
2331 /* Simplify the reg_base_value array so that no register refers to
2332 another register, except to special registers indirectly through
2333 ADDRESS expressions.
2335 In theory this loop can take as long as O(registers^2), but unless
2336 there are very long dependency chains it will run in close to linear
2339 This loop may not be needed any longer now that the main loop does
2340 a better job at propagating alias information. */
2346 for (ui
= 0; ui
< reg_base_value_size
; ui
++)
2348 rtx base
= reg_base_value
[ui
];
2349 if (base
&& GET_CODE (base
) == REG
)
2351 unsigned int base_regno
= REGNO (base
);
2352 if (base_regno
== ui
) /* register set from itself */
2353 reg_base_value
[ui
] = 0;
2355 reg_base_value
[ui
] = reg_base_value
[base_regno
];
2360 while (changed
&& pass
< MAX_ALIAS_LOOP_PASSES
);
2363 free (new_reg_base_value
);
2364 new_reg_base_value
= 0;
2370 end_alias_analysis ()
2372 free (reg_known_value
+ FIRST_PSEUDO_REGISTER
);
2373 reg_known_value
= 0;
2374 reg_known_value_size
= 0;
2375 free (reg_known_equiv_p
+ FIRST_PSEUDO_REGISTER
);
2376 reg_known_equiv_p
= 0;
2379 ggc_del_root (reg_base_value
);
2380 free (reg_base_value
);
2383 reg_base_value_size
= 0;
2384 if (alias_invariant
)
2386 free (alias_invariant
);
2387 alias_invariant
= 0;