1 /* Alias analysis for GNU C
2 Copyright (C) 1997, 1998, 1999, 2000 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 rtx find_base_value
PARAMS ((rtx
));
99 static int mems_in_disjoint_alias_sets_p
PARAMS ((rtx
, rtx
));
100 static int insert_subset_children
PARAMS ((splay_tree_node
, void*));
101 static tree find_base_decl
PARAMS ((tree
));
102 static alias_set_entry get_alias_set_entry
PARAMS ((HOST_WIDE_INT
));
103 static rtx fixed_scalar_and_varying_struct_p
PARAMS ((rtx
, rtx
, rtx
, rtx
,
104 int (*) (rtx
, int)));
105 static int aliases_everything_p
PARAMS ((rtx
));
106 static int write_dependence_p
PARAMS ((rtx
, rtx
, int));
107 static int nonlocal_mentioned_p
PARAMS ((rtx
));
109 static int loop_p
PARAMS ((void));
111 /* Set up all info needed to perform alias analysis on memory references. */
113 /* Returns the size in bytes of the mode of X. */
114 #define SIZE_FOR_MODE(X) (GET_MODE_SIZE (GET_MODE (X)))
116 /* Returns nonzero if MEM1 and MEM2 do not alias because they are in
117 different alias sets. We ignore alias sets in functions making use
118 of variable arguments because the va_arg macros on some systems are
120 #define DIFFERENT_ALIAS_SETS_P(MEM1, MEM2) \
121 mems_in_disjoint_alias_sets_p (MEM1, MEM2)
123 /* Cap the number of passes we make over the insns propagating alias
124 information through set chains. 10 is a completely arbitrary choice. */
125 #define MAX_ALIAS_LOOP_PASSES 10
127 /* reg_base_value[N] gives an address to which register N is related.
128 If all sets after the first add or subtract to the current value
129 or otherwise modify it so it does not point to a different top level
130 object, reg_base_value[N] is equal to the address part of the source
133 A base address can be an ADDRESS, SYMBOL_REF, or LABEL_REF. ADDRESS
134 expressions represent certain special values: function arguments and
135 the stack, frame, and argument pointers.
137 The contents of an ADDRESS is not normally used, the mode of the
138 ADDRESS determines whether the ADDRESS is a function argument or some
139 other special value. Pointer equality, not rtx_equal_p, determines whether
140 two ADDRESS expressions refer to the same base address.
142 The only use of the contents of an ADDRESS is for determining if the
143 current function performs nonlocal memory memory references for the
144 purposes of marking the function as a constant function. */
146 static rtx
*reg_base_value
;
147 static rtx
*new_reg_base_value
;
148 static unsigned int reg_base_value_size
; /* size of reg_base_value array */
150 #define REG_BASE_VALUE(X) \
151 (REGNO (X) < reg_base_value_size \
152 ? reg_base_value[REGNO (X)] : 0)
154 /* Vector of known invariant relationships between registers. Set in
155 loop unrolling. Indexed by register number, if nonzero the value
156 is an expression describing this register in terms of another.
158 The length of this array is REG_BASE_VALUE_SIZE.
160 Because this array contains only pseudo registers it has no effect
162 static rtx
*alias_invariant
;
164 /* Vector indexed by N giving the initial (unchanging) value known for
165 pseudo-register N. This array is initialized in
166 init_alias_analysis, and does not change until end_alias_analysis
168 rtx
*reg_known_value
;
170 /* Indicates number of valid entries in reg_known_value. */
171 static unsigned int reg_known_value_size
;
173 /* Vector recording for each reg_known_value whether it is due to a
174 REG_EQUIV note. Future passes (viz., reload) may replace the
175 pseudo with the equivalent expression and so we account for the
176 dependences that would be introduced if that happens.
178 The REG_EQUIV notes created in assign_parms may mention the arg
179 pointer, and there are explicit insns in the RTL that modify the
180 arg pointer. Thus we must ensure that such insns don't get
181 scheduled across each other because that would invalidate the
182 REG_EQUIV notes. One could argue that the REG_EQUIV notes are
183 wrong, but solving the problem in the scheduler will likely give
184 better code, so we do it here. */
185 char *reg_known_equiv_p
;
187 /* True when scanning insns from the start of the rtl to the
188 NOTE_INSN_FUNCTION_BEG note. */
189 static int copying_arguments
;
191 /* The splay-tree used to store the various alias set entries. */
192 static splay_tree alias_sets
;
194 /* Returns a pointer to the alias set entry for ALIAS_SET, if there is
195 such an entry, or NULL otherwise. */
197 static alias_set_entry
198 get_alias_set_entry (alias_set
)
199 HOST_WIDE_INT alias_set
;
202 = splay_tree_lookup (alias_sets
, (splay_tree_key
) alias_set
);
204 return sn
!= 0 ? ((alias_set_entry
) sn
->value
) : 0;
207 /* Returns nonzero if the alias sets for MEM1 and MEM2 are such that
208 the two MEMs cannot alias each other. */
211 mems_in_disjoint_alias_sets_p (mem1
, mem2
)
215 #ifdef ENABLE_CHECKING
216 /* Perform a basic sanity check. Namely, that there are no alias sets
217 if we're not using strict aliasing. This helps to catch bugs
218 whereby someone uses PUT_CODE, but doesn't clear MEM_ALIAS_SET, or
219 where a MEM is allocated in some way other than by the use of
220 gen_rtx_MEM, and the MEM_ALIAS_SET is not cleared. If we begin to
221 use alias sets to indicate that spilled registers cannot alias each
222 other, we might need to remove this check. */
223 if (! flag_strict_aliasing
224 && (MEM_ALIAS_SET (mem1
) != 0 || MEM_ALIAS_SET (mem2
) != 0))
228 return ! alias_sets_conflict_p (MEM_ALIAS_SET (mem1
), MEM_ALIAS_SET (mem2
));
231 /* Insert the NODE into the splay tree given by DATA. Used by
232 record_alias_subset via splay_tree_foreach. */
235 insert_subset_children (node
, data
)
236 splay_tree_node node
;
239 splay_tree_insert ((splay_tree
) data
, node
->key
, node
->value
);
244 /* Return 1 if the two specified alias sets may conflict. */
247 alias_sets_conflict_p (set1
, set2
)
248 HOST_WIDE_INT set1
, set2
;
252 /* If have no alias set information for one of the operands, we have
253 to assume it can alias anything. */
254 if (set1
== 0 || set2
== 0
255 /* If the two alias sets are the same, they may alias. */
259 /* See if the first alias set is a subset of the second. */
260 ase
= get_alias_set_entry (set1
);
262 && (ase
->has_zero_child
263 || splay_tree_lookup (ase
->children
,
264 (splay_tree_key
) set2
)))
267 /* Now do the same, but with the alias sets reversed. */
268 ase
= get_alias_set_entry (set2
);
270 && (ase
->has_zero_child
271 || splay_tree_lookup (ase
->children
,
272 (splay_tree_key
) set1
)))
275 /* The two alias sets are distinct and neither one is the
276 child of the other. Therefore, they cannot alias. */
280 /* Return 1 if TYPE is a RECORD_TYPE, UNION_TYPE, or QUAL_UNION_TYPE and has
281 has any readonly fields. If any of the fields have types that
282 contain readonly fields, return true as well. */
285 readonly_fields_p (type
)
290 if (TREE_CODE (type
) != RECORD_TYPE
&& TREE_CODE (type
) != UNION_TYPE
291 && TREE_CODE (type
) != QUAL_UNION_TYPE
)
294 for (field
= TYPE_FIELDS (type
); field
!= 0; field
= TREE_CHAIN (field
))
295 if (TREE_CODE (field
) == FIELD_DECL
296 && (TREE_READONLY (field
)
297 || readonly_fields_p (TREE_TYPE (field
))))
303 /* Return 1 if any MEM object of type T1 will always conflict (using the
304 dependency routines in this file) with any MEM object of type T2.
305 This is used when allocating temporary storage. If T1 and/or T2 are
306 NULL_TREE, it means we know nothing about the storage. */
309 objects_must_conflict_p (t1
, t2
)
312 /* If one or the other has readonly fields or is readonly,
313 then they may not conflict. */
314 if ((t1
!= 0 && readonly_fields_p (t1
))
315 || (t2
!= 0 && readonly_fields_p (t2
))
316 || (t1
!= 0 && TYPE_READONLY (t1
))
317 || (t2
!= 0 && TYPE_READONLY (t2
)))
320 /* If they are the same type, they must conflict. */
322 /* Likewise if both are volatile. */
323 || (t1
!= 0 && TYPE_VOLATILE (t1
) && t2
!= 0 && TYPE_VOLATILE (t2
)))
326 /* If one is aggregate and the other is scalar then they may not
328 if ((t1
!= 0 && AGGREGATE_TYPE_P (t1
))
329 != (t2
!= 0 && AGGREGATE_TYPE_P (t2
)))
332 /* Otherwise they conflict only if the alias sets conflict. */
333 return alias_sets_conflict_p (t1
? get_alias_set (t1
) : 0,
334 t2
? get_alias_set (t2
) : 0);
337 /* T is an expression with pointer type. Find the DECL on which this
338 expression is based. (For example, in `a[i]' this would be `a'.)
339 If there is no such DECL, or a unique decl cannot be determined,
340 NULL_TREE is retured. */
348 if (t
== 0 || t
== error_mark_node
|| ! POINTER_TYPE_P (TREE_TYPE (t
)))
351 /* If this is a declaration, return it. */
352 if (TREE_CODE_CLASS (TREE_CODE (t
)) == 'd')
355 /* Handle general expressions. It would be nice to deal with
356 COMPONENT_REFs here. If we could tell that `a' and `b' were the
357 same, then `a->f' and `b->f' are also the same. */
358 switch (TREE_CODE_CLASS (TREE_CODE (t
)))
361 return find_base_decl (TREE_OPERAND (t
, 0));
364 /* Return 0 if found in neither or both are the same. */
365 d0
= find_base_decl (TREE_OPERAND (t
, 0));
366 d1
= find_base_decl (TREE_OPERAND (t
, 1));
377 d0
= find_base_decl (TREE_OPERAND (t
, 0));
378 d1
= find_base_decl (TREE_OPERAND (t
, 1));
379 d0
= find_base_decl (TREE_OPERAND (t
, 0));
380 d2
= find_base_decl (TREE_OPERAND (t
, 2));
382 /* Set any nonzero values from the last, then from the first. */
383 if (d1
== 0) d1
= d2
;
384 if (d0
== 0) d0
= d1
;
385 if (d1
== 0) d1
= d0
;
386 if (d2
== 0) d2
= d1
;
388 /* At this point all are nonzero or all are zero. If all three are the
389 same, return it. Otherwise, return zero. */
390 return (d0
== d1
&& d1
== d2
) ? d0
: 0;
397 /* Return the alias set for T, which may be either a type or an
398 expression. Call language-specific routine for help, if needed. */
407 /* If we're not doing any alias analysis, just assume everything
408 aliases everything else. Also return 0 if this or its type is
410 if (! flag_strict_aliasing
|| t
== error_mark_node
412 && (TREE_TYPE (t
) == 0 || TREE_TYPE (t
) == error_mark_node
)))
415 /* We can be passed either an expression or a type. This and the
416 language-specific routine may make mutually-recursive calls to
417 each other to figure out what to do. At each juncture, we see if
418 this is a tree that the language may need to handle specially.
419 First handle things that aren't types and start by removing nops
420 since we care only about the actual object. */
423 while (TREE_CODE (t
) == NOP_EXPR
|| TREE_CODE (t
) == CONVERT_EXPR
424 || TREE_CODE (t
) == NON_LVALUE_EXPR
)
425 t
= TREE_OPERAND (t
, 0);
427 /* Now give the language a chance to do something but record what we
428 gave it this time. */
430 if ((set
= lang_get_alias_set (t
)) != -1)
433 /* Now loop the same way as get_inner_reference and get the alias
434 set to use. Pick up the outermost object that we could have
438 /* Unnamed bitfields are not an addressable object. */
439 if (TREE_CODE (t
) == BIT_FIELD_REF
)
441 else if (TREE_CODE (t
) == COMPONENT_REF
)
443 if (! DECL_NONADDRESSABLE_P (TREE_OPERAND (t
, 1)))
444 /* Stop at an adressable decl. */
447 else if (TREE_CODE (t
) == ARRAY_REF
)
449 if (! TYPE_NONALIASED_COMPONENT
450 (TREE_TYPE (TREE_OPERAND (t
, 0))))
451 /* Stop at an addresssable array element. */
454 else if (TREE_CODE (t
) != NON_LVALUE_EXPR
455 && ! ((TREE_CODE (t
) == NOP_EXPR
456 || TREE_CODE (t
) == CONVERT_EXPR
)
457 && (TYPE_MODE (TREE_TYPE (t
))
458 == TYPE_MODE (TREE_TYPE (TREE_OPERAND (t
, 0))))))
459 /* Stop if not one of above and not mode-preserving conversion. */
462 t
= TREE_OPERAND (t
, 0);
465 if (TREE_CODE (t
) == INDIRECT_REF
)
467 /* Check for accesses through restrict-qualified pointers. */
468 tree decl
= find_base_decl (TREE_OPERAND (t
, 0));
470 if (decl
&& DECL_POINTER_ALIAS_SET_KNOWN_P (decl
))
471 /* We use the alias set indicated in the declaration. */
472 return DECL_POINTER_ALIAS_SET (decl
);
474 /* If we have an INDIRECT_REF via a void pointer, we don't
475 know anything about what that might alias. */
476 if (TREE_CODE (TREE_TYPE (t
)) == VOID_TYPE
)
480 /* Give the language another chance to do something special. */
482 && (set
= lang_get_alias_set (t
)) != -1)
485 /* Now all we care about is the type. */
489 /* Variant qualifiers don't affect the alias set, so get the main
490 variant. If this is a type with a known alias set, return it. */
491 t
= TYPE_MAIN_VARIANT (t
);
492 if (TYPE_P (t
) && TYPE_ALIAS_SET_KNOWN_P (t
))
493 return TYPE_ALIAS_SET (t
);
495 /* See if the language has special handling for this type. */
496 if ((set
= lang_get_alias_set (t
)) != -1)
498 /* If the alias set is now known, we are done. */
499 if (TYPE_ALIAS_SET_KNOWN_P (t
))
500 return TYPE_ALIAS_SET (t
);
503 /* There are no objects of FUNCTION_TYPE, so there's no point in
504 using up an alias set for them. (There are, of course, pointers
505 and references to functions, but that's different.) */
506 else if (TREE_CODE (t
) == FUNCTION_TYPE
)
509 /* Otherwise make a new alias set for this type. */
510 set
= new_alias_set ();
512 TYPE_ALIAS_SET (t
) = set
;
514 /* If this is an aggregate type, we must record any component aliasing
516 if (AGGREGATE_TYPE_P (t
) || TREE_CODE (t
) == COMPLEX_TYPE
)
517 record_component_aliases (t
);
522 /* Return a brand-new alias set. */
527 static HOST_WIDE_INT last_alias_set
;
529 if (flag_strict_aliasing
)
530 return ++last_alias_set
;
535 /* Indicate that things in SUBSET can alias things in SUPERSET, but
536 not vice versa. For example, in C, a store to an `int' can alias a
537 structure containing an `int', but not vice versa. Here, the
538 structure would be the SUPERSET and `int' the SUBSET. This
539 function should be called only once per SUPERSET/SUBSET pair.
541 It is illegal for SUPERSET to be zero; everything is implicitly a
542 subset of alias set zero. */
545 record_alias_subset (superset
, subset
)
546 HOST_WIDE_INT superset
;
547 HOST_WIDE_INT subset
;
549 alias_set_entry superset_entry
;
550 alias_set_entry subset_entry
;
555 superset_entry
= get_alias_set_entry (superset
);
556 if (superset_entry
== 0)
558 /* Create an entry for the SUPERSET, so that we have a place to
559 attach the SUBSET. */
561 = (alias_set_entry
) xmalloc (sizeof (struct alias_set_entry
));
562 superset_entry
->alias_set
= superset
;
563 superset_entry
->children
564 = splay_tree_new (splay_tree_compare_ints
, 0, 0);
565 superset_entry
->has_zero_child
= 0;
566 splay_tree_insert (alias_sets
, (splay_tree_key
) superset
,
567 (splay_tree_value
) superset_entry
);
571 superset_entry
->has_zero_child
= 1;
574 subset_entry
= get_alias_set_entry (subset
);
575 /* If there is an entry for the subset, enter all of its children
576 (if they are not already present) as children of the SUPERSET. */
579 if (subset_entry
->has_zero_child
)
580 superset_entry
->has_zero_child
= 1;
582 splay_tree_foreach (subset_entry
->children
, insert_subset_children
,
583 superset_entry
->children
);
586 /* Enter the SUBSET itself as a child of the SUPERSET. */
587 splay_tree_insert (superset_entry
->children
,
588 (splay_tree_key
) subset
, 0);
592 /* Record that component types of TYPE, if any, are part of that type for
593 aliasing purposes. For record types, we only record component types
594 for fields that are marked addressable. For array types, we always
595 record the component types, so the front end should not call this
596 function if the individual component aren't addressable. */
599 record_component_aliases (type
)
602 HOST_WIDE_INT superset
= get_alias_set (type
);
608 switch (TREE_CODE (type
))
611 if (! TYPE_NONALIASED_COMPONENT (type
))
612 record_alias_subset (superset
, get_alias_set (TREE_TYPE (type
)));
617 case QUAL_UNION_TYPE
:
618 for (field
= TYPE_FIELDS (type
); field
!= 0; field
= TREE_CHAIN (field
))
619 if (TREE_CODE (field
) == FIELD_DECL
&& ! DECL_NONADDRESSABLE_P (field
))
620 record_alias_subset (superset
, get_alias_set (TREE_TYPE (field
)));
624 record_alias_subset (superset
, get_alias_set (TREE_TYPE (type
)));
632 /* Allocate an alias set for use in storing and reading from the varargs
636 get_varargs_alias_set ()
638 static HOST_WIDE_INT set
= -1;
641 set
= new_alias_set ();
646 /* Likewise, but used for the fixed portions of the frame, e.g., register
650 get_frame_alias_set ()
652 static HOST_WIDE_INT set
= -1;
655 set
= new_alias_set ();
660 /* Inside SRC, the source of a SET, find a base address. */
663 find_base_value (src
)
667 switch (GET_CODE (src
))
675 /* At the start of a function, argument registers have known base
676 values which may be lost later. Returning an ADDRESS
677 expression here allows optimization based on argument values
678 even when the argument registers are used for other purposes. */
679 if (regno
< FIRST_PSEUDO_REGISTER
&& copying_arguments
)
680 return new_reg_base_value
[regno
];
682 /* If a pseudo has a known base value, return it. Do not do this
683 for hard regs since it can result in a circular dependency
684 chain for registers which have values at function entry.
686 The test above is not sufficient because the scheduler may move
687 a copy out of an arg reg past the NOTE_INSN_FUNCTION_BEGIN. */
688 if (regno
>= FIRST_PSEUDO_REGISTER
689 && regno
< reg_base_value_size
690 && reg_base_value
[regno
])
691 return reg_base_value
[regno
];
696 /* Check for an argument passed in memory. Only record in the
697 copying-arguments block; it is too hard to track changes
699 if (copying_arguments
700 && (XEXP (src
, 0) == arg_pointer_rtx
701 || (GET_CODE (XEXP (src
, 0)) == PLUS
702 && XEXP (XEXP (src
, 0), 0) == arg_pointer_rtx
)))
703 return gen_rtx_ADDRESS (VOIDmode
, src
);
708 if (GET_CODE (src
) != PLUS
&& GET_CODE (src
) != MINUS
)
711 /* ... fall through ... */
716 rtx temp
, src_0
= XEXP (src
, 0), src_1
= XEXP (src
, 1);
718 /* If either operand is a REG, then see if we already have
719 a known value for it. */
720 if (GET_CODE (src_0
) == REG
)
722 temp
= find_base_value (src_0
);
727 if (GET_CODE (src_1
) == REG
)
729 temp
= find_base_value (src_1
);
734 /* Guess which operand is the base address:
735 If either operand is a symbol, then it is the base. If
736 either operand is a CONST_INT, then the other is the base. */
737 if (GET_CODE (src_1
) == CONST_INT
|| CONSTANT_P (src_0
))
738 return find_base_value (src_0
);
739 else if (GET_CODE (src_0
) == CONST_INT
|| CONSTANT_P (src_1
))
740 return find_base_value (src_1
);
742 /* This might not be necessary anymore:
743 If either operand is a REG that is a known pointer, then it
745 else if (GET_CODE (src_0
) == REG
&& REG_POINTER (src_0
))
746 return find_base_value (src_0
);
747 else if (GET_CODE (src_1
) == REG
&& REG_POINTER (src_1
))
748 return find_base_value (src_1
);
754 /* The standard form is (lo_sum reg sym) so look only at the
756 return find_base_value (XEXP (src
, 1));
759 /* If the second operand is constant set the base
760 address to the first operand. */
761 if (GET_CODE (XEXP (src
, 1)) == CONST_INT
&& INTVAL (XEXP (src
, 1)) != 0)
762 return find_base_value (XEXP (src
, 0));
766 if (GET_MODE_SIZE (GET_MODE (src
)) < GET_MODE_SIZE (Pmode
))
770 case SIGN_EXTEND
: /* used for NT/Alpha pointers */
772 return find_base_value (XEXP (src
, 0));
781 /* Called from init_alias_analysis indirectly through note_stores. */
783 /* While scanning insns to find base values, reg_seen[N] is nonzero if
784 register N has been set in this function. */
785 static char *reg_seen
;
787 /* Addresses which are known not to alias anything else are identified
788 by a unique integer. */
789 static int unique_id
;
792 record_set (dest
, set
, data
)
794 void *data ATTRIBUTE_UNUSED
;
796 register unsigned regno
;
799 if (GET_CODE (dest
) != REG
)
802 regno
= REGNO (dest
);
804 if (regno
>= reg_base_value_size
)
809 /* A CLOBBER wipes out any old value but does not prevent a previously
810 unset register from acquiring a base address (i.e. reg_seen is not
812 if (GET_CODE (set
) == CLOBBER
)
814 new_reg_base_value
[regno
] = 0;
823 new_reg_base_value
[regno
] = 0;
827 new_reg_base_value
[regno
] = gen_rtx_ADDRESS (Pmode
,
828 GEN_INT (unique_id
++));
832 /* This is not the first set. If the new value is not related to the
833 old value, forget the base value. Note that the following code is
835 extern int x, y; int *p = &x; p += (&y-&x);
836 ANSI C does not allow computing the difference of addresses
837 of distinct top level objects. */
838 if (new_reg_base_value
[regno
])
839 switch (GET_CODE (src
))
843 if (XEXP (src
, 0) != dest
&& XEXP (src
, 1) != dest
)
844 new_reg_base_value
[regno
] = 0;
847 /* If the value we add in the PLUS is also a valid base value,
848 this might be the actual base value, and the original value
851 rtx other
= NULL_RTX
;
853 if (XEXP (src
, 0) == dest
)
854 other
= XEXP (src
, 1);
855 else if (XEXP (src
, 1) == dest
)
856 other
= XEXP (src
, 0);
858 if (! other
|| find_base_value (other
))
859 new_reg_base_value
[regno
] = 0;
863 if (XEXP (src
, 0) != dest
|| GET_CODE (XEXP (src
, 1)) != CONST_INT
)
864 new_reg_base_value
[regno
] = 0;
867 new_reg_base_value
[regno
] = 0;
870 /* If this is the first set of a register, record the value. */
871 else if ((regno
>= FIRST_PSEUDO_REGISTER
|| ! fixed_regs
[regno
])
872 && ! reg_seen
[regno
] && new_reg_base_value
[regno
] == 0)
873 new_reg_base_value
[regno
] = find_base_value (src
);
878 /* Called from loop optimization when a new pseudo-register is
879 created. It indicates that REGNO is being set to VAL. f INVARIANT
880 is true then this value also describes an invariant relationship
881 which can be used to deduce that two registers with unknown values
885 record_base_value (regno
, val
, invariant
)
890 if (regno
>= reg_base_value_size
)
893 if (invariant
&& alias_invariant
)
894 alias_invariant
[regno
] = val
;
896 if (GET_CODE (val
) == REG
)
898 if (REGNO (val
) < reg_base_value_size
)
899 reg_base_value
[regno
] = reg_base_value
[REGNO (val
)];
904 reg_base_value
[regno
] = find_base_value (val
);
907 /* Returns a canonical version of X, from the point of view alias
908 analysis. (For example, if X is a MEM whose address is a register,
909 and the register has a known value (say a SYMBOL_REF), then a MEM
910 whose address is the SYMBOL_REF is returned.) */
916 /* Recursively look for equivalences. */
917 if (GET_CODE (x
) == REG
&& REGNO (x
) >= FIRST_PSEUDO_REGISTER
918 && REGNO (x
) < reg_known_value_size
)
919 return reg_known_value
[REGNO (x
)] == x
920 ? x
: canon_rtx (reg_known_value
[REGNO (x
)]);
921 else if (GET_CODE (x
) == PLUS
)
923 rtx x0
= canon_rtx (XEXP (x
, 0));
924 rtx x1
= canon_rtx (XEXP (x
, 1));
926 if (x0
!= XEXP (x
, 0) || x1
!= XEXP (x
, 1))
928 /* We can tolerate LO_SUMs being offset here; these
929 rtl are used for nothing other than comparisons. */
930 if (GET_CODE (x0
) == CONST_INT
)
931 return plus_constant_for_output (x1
, INTVAL (x0
));
932 else if (GET_CODE (x1
) == CONST_INT
)
933 return plus_constant_for_output (x0
, INTVAL (x1
));
934 return gen_rtx_PLUS (GET_MODE (x
), x0
, x1
);
938 /* This gives us much better alias analysis when called from
939 the loop optimizer. Note we want to leave the original
940 MEM alone, but need to return the canonicalized MEM with
941 all the flags with their original values. */
942 else if (GET_CODE (x
) == MEM
)
944 rtx addr
= canon_rtx (XEXP (x
, 0));
946 if (addr
!= XEXP (x
, 0))
948 rtx
new = gen_rtx_MEM (GET_MODE (x
), addr
);
950 MEM_COPY_ATTRIBUTES (new, x
);
957 /* Return 1 if X and Y are identical-looking rtx's.
959 We use the data in reg_known_value above to see if two registers with
960 different numbers are, in fact, equivalent. */
963 rtx_equal_for_memref_p (x
, y
)
968 register enum rtx_code code
;
969 register const char *fmt
;
971 if (x
== 0 && y
== 0)
973 if (x
== 0 || y
== 0)
983 /* Rtx's of different codes cannot be equal. */
984 if (code
!= GET_CODE (y
))
987 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent.
988 (REG:SI x) and (REG:HI x) are NOT equivalent. */
990 if (GET_MODE (x
) != GET_MODE (y
))
993 /* Some RTL can be compared without a recursive examination. */
997 return REGNO (x
) == REGNO (y
);
1000 return XEXP (x
, 0) == XEXP (y
, 0);
1003 return XSTR (x
, 0) == XSTR (y
, 0);
1007 /* There's no need to compare the contents of CONST_DOUBLEs or
1008 CONST_INTs because pointer equality is a good enough
1009 comparison for these nodes. */
1013 return (XINT (x
, 1) == XINT (y
, 1)
1014 && rtx_equal_for_memref_p (XEXP (x
, 0), XEXP (y
, 0)));
1020 /* For commutative operations, the RTX match if the operand match in any
1021 order. Also handle the simple binary and unary cases without a loop. */
1022 if (code
== EQ
|| code
== NE
|| GET_RTX_CLASS (code
) == 'c')
1023 return ((rtx_equal_for_memref_p (XEXP (x
, 0), XEXP (y
, 0))
1024 && rtx_equal_for_memref_p (XEXP (x
, 1), XEXP (y
, 1)))
1025 || (rtx_equal_for_memref_p (XEXP (x
, 0), XEXP (y
, 1))
1026 && rtx_equal_for_memref_p (XEXP (x
, 1), XEXP (y
, 0))));
1027 else if (GET_RTX_CLASS (code
) == '<' || GET_RTX_CLASS (code
) == '2')
1028 return (rtx_equal_for_memref_p (XEXP (x
, 0), XEXP (y
, 0))
1029 && rtx_equal_for_memref_p (XEXP (x
, 1), XEXP (y
, 1)));
1030 else if (GET_RTX_CLASS (code
) == '1')
1031 return rtx_equal_for_memref_p (XEXP (x
, 0), XEXP (y
, 0));
1033 /* Compare the elements. If any pair of corresponding elements
1034 fail to match, return 0 for the whole things.
1036 Limit cases to types which actually appear in addresses. */
1038 fmt
= GET_RTX_FORMAT (code
);
1039 for (i
= GET_RTX_LENGTH (code
) - 1; i
>= 0; i
--)
1044 if (XINT (x
, i
) != XINT (y
, i
))
1049 /* Two vectors must have the same length. */
1050 if (XVECLEN (x
, i
) != XVECLEN (y
, i
))
1053 /* And the corresponding elements must match. */
1054 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
1055 if (rtx_equal_for_memref_p (XVECEXP (x
, i
, j
),
1056 XVECEXP (y
, i
, j
)) == 0)
1061 if (rtx_equal_for_memref_p (XEXP (x
, i
), XEXP (y
, i
)) == 0)
1065 /* This can happen for an asm which clobbers memory. */
1069 /* It is believed that rtx's at this level will never
1070 contain anything but integers and other rtx's,
1071 except for within LABEL_REFs and SYMBOL_REFs. */
1079 /* Given an rtx X, find a SYMBOL_REF or LABEL_REF within
1080 X and return it, or return 0 if none found. */
1083 find_symbolic_term (x
)
1087 register enum rtx_code code
;
1088 register const char *fmt
;
1090 code
= GET_CODE (x
);
1091 if (code
== SYMBOL_REF
|| code
== LABEL_REF
)
1093 if (GET_RTX_CLASS (code
) == 'o')
1096 fmt
= GET_RTX_FORMAT (code
);
1097 for (i
= GET_RTX_LENGTH (code
) - 1; i
>= 0; i
--)
1103 t
= find_symbolic_term (XEXP (x
, i
));
1107 else if (fmt
[i
] == 'E')
1118 struct elt_loc_list
*l
;
1120 #if defined (FIND_BASE_TERM)
1121 /* Try machine-dependent ways to find the base term. */
1122 x
= FIND_BASE_TERM (x
);
1125 switch (GET_CODE (x
))
1128 return REG_BASE_VALUE (x
);
1131 case SIGN_EXTEND
: /* Used for Alpha/NT pointers */
1137 return find_base_term (XEXP (x
, 0));
1140 val
= CSELIB_VAL_PTR (x
);
1141 for (l
= val
->locs
; l
; l
= l
->next
)
1142 if ((x
= find_base_term (l
->loc
)) != 0)
1148 if (GET_CODE (x
) != PLUS
&& GET_CODE (x
) != MINUS
)
1155 rtx tmp1
= XEXP (x
, 0);
1156 rtx tmp2
= XEXP (x
, 1);
1158 /* This is a litle bit tricky since we have to determine which of
1159 the two operands represents the real base address. Otherwise this
1160 routine may return the index register instead of the base register.
1162 That may cause us to believe no aliasing was possible, when in
1163 fact aliasing is possible.
1165 We use a few simple tests to guess the base register. Additional
1166 tests can certainly be added. For example, if one of the operands
1167 is a shift or multiply, then it must be the index register and the
1168 other operand is the base register. */
1170 if (tmp1
== pic_offset_table_rtx
&& CONSTANT_P (tmp2
))
1171 return find_base_term (tmp2
);
1173 /* If either operand is known to be a pointer, then use it
1174 to determine the base term. */
1175 if (REG_P (tmp1
) && REG_POINTER (tmp1
))
1176 return find_base_term (tmp1
);
1178 if (REG_P (tmp2
) && REG_POINTER (tmp2
))
1179 return find_base_term (tmp2
);
1181 /* Neither operand was known to be a pointer. Go ahead and find the
1182 base term for both operands. */
1183 tmp1
= find_base_term (tmp1
);
1184 tmp2
= find_base_term (tmp2
);
1186 /* If either base term is named object or a special address
1187 (like an argument or stack reference), then use it for the
1190 && (GET_CODE (tmp1
) == SYMBOL_REF
1191 || GET_CODE (tmp1
) == LABEL_REF
1192 || (GET_CODE (tmp1
) == ADDRESS
1193 && GET_MODE (tmp1
) != VOIDmode
)))
1197 && (GET_CODE (tmp2
) == SYMBOL_REF
1198 || GET_CODE (tmp2
) == LABEL_REF
1199 || (GET_CODE (tmp2
) == ADDRESS
1200 && GET_MODE (tmp2
) != VOIDmode
)))
1203 /* We could not determine which of the two operands was the
1204 base register and which was the index. So we can determine
1205 nothing from the base alias check. */
1210 if (GET_CODE (XEXP (x
, 0)) == REG
&& GET_CODE (XEXP (x
, 1)) == CONST_INT
)
1211 return REG_BASE_VALUE (XEXP (x
, 0));
1219 return REG_BASE_VALUE (frame_pointer_rtx
);
1226 /* Return 0 if the addresses X and Y are known to point to different
1227 objects, 1 if they might be pointers to the same object. */
1230 base_alias_check (x
, y
, x_mode
, y_mode
)
1232 enum machine_mode x_mode
, y_mode
;
1234 rtx x_base
= find_base_term (x
);
1235 rtx y_base
= find_base_term (y
);
1237 /* If the address itself has no known base see if a known equivalent
1238 value has one. If either address still has no known base, nothing
1239 is known about aliasing. */
1244 if (! flag_expensive_optimizations
|| (x_c
= canon_rtx (x
)) == x
)
1247 x_base
= find_base_term (x_c
);
1255 if (! flag_expensive_optimizations
|| (y_c
= canon_rtx (y
)) == y
)
1258 y_base
= find_base_term (y_c
);
1263 /* If the base addresses are equal nothing is known about aliasing. */
1264 if (rtx_equal_p (x_base
, y_base
))
1267 /* The base addresses of the read and write are different expressions.
1268 If they are both symbols and they are not accessed via AND, there is
1269 no conflict. We can bring knowledge of object alignment into play
1270 here. For example, on alpha, "char a, b;" can alias one another,
1271 though "char a; long b;" cannot. */
1272 if (GET_CODE (x_base
) != ADDRESS
&& GET_CODE (y_base
) != ADDRESS
)
1274 if (GET_CODE (x
) == AND
&& GET_CODE (y
) == AND
)
1276 if (GET_CODE (x
) == AND
1277 && (GET_CODE (XEXP (x
, 1)) != CONST_INT
1278 || GET_MODE_UNIT_SIZE (y_mode
) < -INTVAL (XEXP (x
, 1))))
1280 if (GET_CODE (y
) == AND
1281 && (GET_CODE (XEXP (y
, 1)) != CONST_INT
1282 || GET_MODE_UNIT_SIZE (x_mode
) < -INTVAL (XEXP (y
, 1))))
1284 /* Differing symbols never alias. */
1288 /* If one address is a stack reference there can be no alias:
1289 stack references using different base registers do not alias,
1290 a stack reference can not alias a parameter, and a stack reference
1291 can not alias a global. */
1292 if ((GET_CODE (x_base
) == ADDRESS
&& GET_MODE (x_base
) == Pmode
)
1293 || (GET_CODE (y_base
) == ADDRESS
&& GET_MODE (y_base
) == Pmode
))
1296 if (! flag_argument_noalias
)
1299 if (flag_argument_noalias
> 1)
1302 /* Weak noalias assertion (arguments are distinct, but may match globals). */
1303 return ! (GET_MODE (x_base
) == VOIDmode
&& GET_MODE (y_base
) == VOIDmode
);
1306 /* Convert the address X into something we can use. This is done by returning
1307 it unchanged unless it is a value; in the latter case we call cselib to get
1308 a more useful rtx. */
1315 struct elt_loc_list
*l
;
1317 if (GET_CODE (x
) != VALUE
)
1319 v
= CSELIB_VAL_PTR (x
);
1320 for (l
= v
->locs
; l
; l
= l
->next
)
1321 if (CONSTANT_P (l
->loc
))
1323 for (l
= v
->locs
; l
; l
= l
->next
)
1324 if (GET_CODE (l
->loc
) != REG
&& GET_CODE (l
->loc
) != MEM
)
1327 return v
->locs
->loc
;
1331 /* Return the address of the (N_REFS + 1)th memory reference to ADDR
1332 where SIZE is the size in bytes of the memory reference. If ADDR
1333 is not modified by the memory reference then ADDR is returned. */
1336 addr_side_effect_eval (addr
, size
, n_refs
)
1343 switch (GET_CODE (addr
))
1346 offset
= (n_refs
+ 1) * size
;
1349 offset
= -(n_refs
+ 1) * size
;
1352 offset
= n_refs
* size
;
1355 offset
= -n_refs
* size
;
1363 addr
= gen_rtx_PLUS (GET_MODE (addr
), XEXP (addr
, 0), GEN_INT (offset
));
1365 addr
= XEXP (addr
, 0);
1370 /* Return nonzero if X and Y (memory addresses) could reference the
1371 same location in memory. C is an offset accumulator. When
1372 C is nonzero, we are testing aliases between X and Y + C.
1373 XSIZE is the size in bytes of the X reference,
1374 similarly YSIZE is the size in bytes for Y.
1376 If XSIZE or YSIZE is zero, we do not know the amount of memory being
1377 referenced (the reference was BLKmode), so make the most pessimistic
1380 If XSIZE or YSIZE is negative, we may access memory outside the object
1381 being referenced as a side effect. This can happen when using AND to
1382 align memory references, as is done on the Alpha.
1384 Nice to notice that varying addresses cannot conflict with fp if no
1385 local variables had their addresses taken, but that's too hard now. */
1388 memrefs_conflict_p (xsize
, x
, ysize
, y
, c
)
1393 if (GET_CODE (x
) == VALUE
)
1395 if (GET_CODE (y
) == VALUE
)
1397 if (GET_CODE (x
) == HIGH
)
1399 else if (GET_CODE (x
) == LO_SUM
)
1402 x
= canon_rtx (addr_side_effect_eval (x
, xsize
, 0));
1403 if (GET_CODE (y
) == HIGH
)
1405 else if (GET_CODE (y
) == LO_SUM
)
1408 y
= canon_rtx (addr_side_effect_eval (y
, ysize
, 0));
1410 if (rtx_equal_for_memref_p (x
, y
))
1412 if (xsize
<= 0 || ysize
<= 0)
1414 if (c
>= 0 && xsize
> c
)
1416 if (c
< 0 && ysize
+c
> 0)
1421 /* This code used to check for conflicts involving stack references and
1422 globals but the base address alias code now handles these cases. */
1424 if (GET_CODE (x
) == PLUS
)
1426 /* The fact that X is canonicalized means that this
1427 PLUS rtx is canonicalized. */
1428 rtx x0
= XEXP (x
, 0);
1429 rtx x1
= XEXP (x
, 1);
1431 if (GET_CODE (y
) == PLUS
)
1433 /* The fact that Y is canonicalized means that this
1434 PLUS rtx is canonicalized. */
1435 rtx y0
= XEXP (y
, 0);
1436 rtx y1
= XEXP (y
, 1);
1438 if (rtx_equal_for_memref_p (x1
, y1
))
1439 return memrefs_conflict_p (xsize
, x0
, ysize
, y0
, c
);
1440 if (rtx_equal_for_memref_p (x0
, y0
))
1441 return memrefs_conflict_p (xsize
, x1
, ysize
, y1
, c
);
1442 if (GET_CODE (x1
) == CONST_INT
)
1444 if (GET_CODE (y1
) == CONST_INT
)
1445 return memrefs_conflict_p (xsize
, x0
, ysize
, y0
,
1446 c
- INTVAL (x1
) + INTVAL (y1
));
1448 return memrefs_conflict_p (xsize
, x0
, ysize
, y
,
1451 else if (GET_CODE (y1
) == CONST_INT
)
1452 return memrefs_conflict_p (xsize
, x
, ysize
, y0
, c
+ INTVAL (y1
));
1456 else if (GET_CODE (x1
) == CONST_INT
)
1457 return memrefs_conflict_p (xsize
, x0
, ysize
, y
, c
- INTVAL (x1
));
1459 else if (GET_CODE (y
) == PLUS
)
1461 /* The fact that Y is canonicalized means that this
1462 PLUS rtx is canonicalized. */
1463 rtx y0
= XEXP (y
, 0);
1464 rtx y1
= XEXP (y
, 1);
1466 if (GET_CODE (y1
) == CONST_INT
)
1467 return memrefs_conflict_p (xsize
, x
, ysize
, y0
, c
+ INTVAL (y1
));
1472 if (GET_CODE (x
) == GET_CODE (y
))
1473 switch (GET_CODE (x
))
1477 /* Handle cases where we expect the second operands to be the
1478 same, and check only whether the first operand would conflict
1481 rtx x1
= canon_rtx (XEXP (x
, 1));
1482 rtx y1
= canon_rtx (XEXP (y
, 1));
1483 if (! rtx_equal_for_memref_p (x1
, y1
))
1485 x0
= canon_rtx (XEXP (x
, 0));
1486 y0
= canon_rtx (XEXP (y
, 0));
1487 if (rtx_equal_for_memref_p (x0
, y0
))
1488 return (xsize
== 0 || ysize
== 0
1489 || (c
>= 0 && xsize
> c
) || (c
< 0 && ysize
+c
> 0));
1491 /* Can't properly adjust our sizes. */
1492 if (GET_CODE (x1
) != CONST_INT
)
1494 xsize
/= INTVAL (x1
);
1495 ysize
/= INTVAL (x1
);
1497 return memrefs_conflict_p (xsize
, x0
, ysize
, y0
, c
);
1501 /* Are these registers known not to be equal? */
1502 if (alias_invariant
)
1504 unsigned int r_x
= REGNO (x
), r_y
= REGNO (y
);
1505 rtx i_x
, i_y
; /* invariant relationships of X and Y */
1507 i_x
= r_x
>= reg_base_value_size
? 0 : alias_invariant
[r_x
];
1508 i_y
= r_y
>= reg_base_value_size
? 0 : alias_invariant
[r_y
];
1510 if (i_x
== 0 && i_y
== 0)
1513 if (! memrefs_conflict_p (xsize
, i_x
? i_x
: x
,
1514 ysize
, i_y
? i_y
: y
, c
))
1523 /* Treat an access through an AND (e.g. a subword access on an Alpha)
1524 as an access with indeterminate size. Assume that references
1525 besides AND are aligned, so if the size of the other reference is
1526 at least as large as the alignment, assume no other overlap. */
1527 if (GET_CODE (x
) == AND
&& GET_CODE (XEXP (x
, 1)) == CONST_INT
)
1529 if (GET_CODE (y
) == AND
|| ysize
< -INTVAL (XEXP (x
, 1)))
1531 return memrefs_conflict_p (xsize
, XEXP (x
, 0), ysize
, y
, c
);
1533 if (GET_CODE (y
) == AND
&& GET_CODE (XEXP (y
, 1)) == CONST_INT
)
1535 /* ??? If we are indexing far enough into the array/structure, we
1536 may yet be able to determine that we can not overlap. But we
1537 also need to that we are far enough from the end not to overlap
1538 a following reference, so we do nothing with that for now. */
1539 if (GET_CODE (x
) == AND
|| xsize
< -INTVAL (XEXP (y
, 1)))
1541 return memrefs_conflict_p (xsize
, x
, ysize
, XEXP (y
, 0), c
);
1544 if (GET_CODE (x
) == ADDRESSOF
)
1546 if (y
== frame_pointer_rtx
1547 || GET_CODE (y
) == ADDRESSOF
)
1548 return xsize
<= 0 || ysize
<= 0;
1550 if (GET_CODE (y
) == ADDRESSOF
)
1552 if (x
== frame_pointer_rtx
)
1553 return xsize
<= 0 || ysize
<= 0;
1558 if (GET_CODE (x
) == CONST_INT
&& GET_CODE (y
) == CONST_INT
)
1560 c
+= (INTVAL (y
) - INTVAL (x
));
1561 return (xsize
<= 0 || ysize
<= 0
1562 || (c
>= 0 && xsize
> c
) || (c
< 0 && ysize
+c
> 0));
1565 if (GET_CODE (x
) == CONST
)
1567 if (GET_CODE (y
) == CONST
)
1568 return memrefs_conflict_p (xsize
, canon_rtx (XEXP (x
, 0)),
1569 ysize
, canon_rtx (XEXP (y
, 0)), c
);
1571 return memrefs_conflict_p (xsize
, canon_rtx (XEXP (x
, 0)),
1574 if (GET_CODE (y
) == CONST
)
1575 return memrefs_conflict_p (xsize
, x
, ysize
,
1576 canon_rtx (XEXP (y
, 0)), c
);
1579 return (xsize
<= 0 || ysize
<= 0
1580 || (rtx_equal_for_memref_p (x
, y
)
1581 && ((c
>= 0 && xsize
> c
) || (c
< 0 && ysize
+c
> 0))));
1588 /* Functions to compute memory dependencies.
1590 Since we process the insns in execution order, we can build tables
1591 to keep track of what registers are fixed (and not aliased), what registers
1592 are varying in known ways, and what registers are varying in unknown
1595 If both memory references are volatile, then there must always be a
1596 dependence between the two references, since their order can not be
1597 changed. A volatile and non-volatile reference can be interchanged
1600 A MEM_IN_STRUCT reference at a non-AND varying address can never
1601 conflict with a non-MEM_IN_STRUCT reference at a fixed address. We
1602 also must allow AND addresses, because they may generate accesses
1603 outside the object being referenced. This is used to generate
1604 aligned addresses from unaligned addresses, for instance, the alpha
1605 storeqi_unaligned pattern. */
1607 /* Read dependence: X is read after read in MEM takes place. There can
1608 only be a dependence here if both reads are volatile. */
1611 read_dependence (mem
, x
)
1615 return MEM_VOLATILE_P (x
) && MEM_VOLATILE_P (mem
);
1618 /* Returns MEM1 if and only if MEM1 is a scalar at a fixed address and
1619 MEM2 is a reference to a structure at a varying address, or returns
1620 MEM2 if vice versa. Otherwise, returns NULL_RTX. If a non-NULL
1621 value is returned MEM1 and MEM2 can never alias. VARIES_P is used
1622 to decide whether or not an address may vary; it should return
1623 nonzero whenever variation is possible.
1624 MEM1_ADDR and MEM2_ADDR are the addresses of MEM1 and MEM2. */
1627 fixed_scalar_and_varying_struct_p (mem1
, mem2
, mem1_addr
, mem2_addr
, varies_p
)
1629 rtx mem1_addr
, mem2_addr
;
1630 int (*varies_p
) PARAMS ((rtx
, int));
1632 if (! flag_strict_aliasing
)
1635 if (MEM_SCALAR_P (mem1
) && MEM_IN_STRUCT_P (mem2
)
1636 && !varies_p (mem1_addr
, 1) && varies_p (mem2_addr
, 1))
1637 /* MEM1 is a scalar at a fixed address; MEM2 is a struct at a
1641 if (MEM_IN_STRUCT_P (mem1
) && MEM_SCALAR_P (mem2
)
1642 && varies_p (mem1_addr
, 1) && !varies_p (mem2_addr
, 1))
1643 /* MEM2 is a scalar at a fixed address; MEM1 is a struct at a
1650 /* Returns nonzero if something about the mode or address format MEM1
1651 indicates that it might well alias *anything*. */
1654 aliases_everything_p (mem
)
1657 if (GET_CODE (XEXP (mem
, 0)) == AND
)
1658 /* If the address is an AND, its very hard to know at what it is
1659 actually pointing. */
1665 /* True dependence: X is read after store in MEM takes place. */
1668 true_dependence (mem
, mem_mode
, x
, varies
)
1670 enum machine_mode mem_mode
;
1672 int (*varies
) PARAMS ((rtx
, int));
1674 register rtx x_addr
, mem_addr
;
1677 if (MEM_VOLATILE_P (x
) && MEM_VOLATILE_P (mem
))
1680 if (DIFFERENT_ALIAS_SETS_P (x
, mem
))
1683 /* Unchanging memory can't conflict with non-unchanging memory.
1684 A non-unchanging read can conflict with a non-unchanging write.
1685 An unchanging read can conflict with an unchanging write since
1686 there may be a single store to this address to initialize it.
1687 Note that an unchanging store can conflict with a non-unchanging read
1688 since we have to make conservative assumptions when we have a
1689 record with readonly fields and we are copying the whole thing.
1690 Just fall through to the code below to resolve potential conflicts.
1691 This won't handle all cases optimally, but the possible performance
1692 loss should be negligible. */
1693 if (RTX_UNCHANGING_P (x
) && ! RTX_UNCHANGING_P (mem
))
1696 if (mem_mode
== VOIDmode
)
1697 mem_mode
= GET_MODE (mem
);
1699 x_addr
= get_addr (XEXP (x
, 0));
1700 mem_addr
= get_addr (XEXP (mem
, 0));
1702 base
= find_base_term (x_addr
);
1703 if (base
&& (GET_CODE (base
) == LABEL_REF
1704 || (GET_CODE (base
) == SYMBOL_REF
1705 && CONSTANT_POOL_ADDRESS_P (base
))))
1708 if (! base_alias_check (x_addr
, mem_addr
, GET_MODE (x
), mem_mode
))
1711 x_addr
= canon_rtx (x_addr
);
1712 mem_addr
= canon_rtx (mem_addr
);
1714 if (! memrefs_conflict_p (GET_MODE_SIZE (mem_mode
), mem_addr
,
1715 SIZE_FOR_MODE (x
), x_addr
, 0))
1718 if (aliases_everything_p (x
))
1721 /* We cannot use aliases_everyting_p to test MEM, since we must look
1722 at MEM_MODE, rather than GET_MODE (MEM). */
1723 if (mem_mode
== QImode
|| GET_CODE (mem_addr
) == AND
)
1726 /* In true_dependence we also allow BLKmode to alias anything. Why
1727 don't we do this in anti_dependence and output_dependence? */
1728 if (mem_mode
== BLKmode
|| GET_MODE (x
) == BLKmode
)
1731 return ! fixed_scalar_and_varying_struct_p (mem
, x
, mem_addr
, x_addr
,
1735 /* Returns non-zero if a write to X might alias a previous read from
1736 (or, if WRITEP is non-zero, a write to) MEM. */
1739 write_dependence_p (mem
, x
, writep
)
1744 rtx x_addr
, mem_addr
;
1748 if (MEM_VOLATILE_P (x
) && MEM_VOLATILE_P (mem
))
1751 if (DIFFERENT_ALIAS_SETS_P (x
, mem
))
1754 /* Unchanging memory can't conflict with non-unchanging memory. */
1755 if (RTX_UNCHANGING_P (x
) != RTX_UNCHANGING_P (mem
))
1758 /* If MEM is an unchanging read, then it can't possibly conflict with
1759 the store to X, because there is at most one store to MEM, and it must
1760 have occurred somewhere before MEM. */
1761 if (! writep
&& RTX_UNCHANGING_P (mem
))
1764 x_addr
= get_addr (XEXP (x
, 0));
1765 mem_addr
= get_addr (XEXP (mem
, 0));
1769 base
= find_base_term (mem_addr
);
1770 if (base
&& (GET_CODE (base
) == LABEL_REF
1771 || (GET_CODE (base
) == SYMBOL_REF
1772 && CONSTANT_POOL_ADDRESS_P (base
))))
1776 if (! base_alias_check (x_addr
, mem_addr
, GET_MODE (x
),
1780 x_addr
= canon_rtx (x_addr
);
1781 mem_addr
= canon_rtx (mem_addr
);
1783 if (!memrefs_conflict_p (SIZE_FOR_MODE (mem
), mem_addr
,
1784 SIZE_FOR_MODE (x
), x_addr
, 0))
1788 = fixed_scalar_and_varying_struct_p (mem
, x
, mem_addr
, x_addr
,
1791 return (!(fixed_scalar
== mem
&& !aliases_everything_p (x
))
1792 && !(fixed_scalar
== x
&& !aliases_everything_p (mem
)));
1795 /* Anti dependence: X is written after read in MEM takes place. */
1798 anti_dependence (mem
, x
)
1802 return write_dependence_p (mem
, x
, /*writep=*/0);
1805 /* Output dependence: X is written after store in MEM takes place. */
1808 output_dependence (mem
, x
)
1812 return write_dependence_p (mem
, x
, /*writep=*/1);
1815 /* Returns non-zero if X mentions something which is not
1816 local to the function and is not constant. */
1819 nonlocal_mentioned_p (x
)
1823 register RTX_CODE code
;
1826 code
= GET_CODE (x
);
1828 if (GET_RTX_CLASS (code
) == 'i')
1830 /* Constant functions can be constant if they don't use
1831 scratch memory used to mark function w/o side effects. */
1832 if (code
== CALL_INSN
&& CONST_CALL_P (x
))
1834 x
= CALL_INSN_FUNCTION_USAGE (x
);
1840 code
= GET_CODE (x
);
1846 if (GET_CODE (SUBREG_REG (x
)) == REG
)
1848 /* Global registers are not local. */
1849 if (REGNO (SUBREG_REG (x
)) < FIRST_PSEUDO_REGISTER
1850 && global_regs
[REGNO (SUBREG_REG (x
)) + SUBREG_WORD (x
)])
1858 /* Global registers are not local. */
1859 if (regno
< FIRST_PSEUDO_REGISTER
&& global_regs
[regno
])
1873 /* Constants in the function's constants pool are constant. */
1874 if (CONSTANT_POOL_ADDRESS_P (x
))
1879 /* Non-constant calls and recursion are not local. */
1883 /* Be overly conservative and consider any volatile memory
1884 reference as not local. */
1885 if (MEM_VOLATILE_P (x
))
1887 base
= find_base_term (XEXP (x
, 0));
1890 /* A Pmode ADDRESS could be a reference via the structure value
1891 address or static chain. Such memory references are nonlocal.
1893 Thus, we have to examine the contents of the ADDRESS to find
1894 out if this is a local reference or not. */
1895 if (GET_CODE (base
) == ADDRESS
1896 && GET_MODE (base
) == Pmode
1897 && (XEXP (base
, 0) == stack_pointer_rtx
1898 || XEXP (base
, 0) == arg_pointer_rtx
1899 #if HARD_FRAME_POINTER_REGNUM != FRAME_POINTER_REGNUM
1900 || XEXP (base
, 0) == hard_frame_pointer_rtx
1902 || XEXP (base
, 0) == frame_pointer_rtx
))
1904 /* Constants in the function's constant pool are constant. */
1905 if (GET_CODE (base
) == SYMBOL_REF
&& CONSTANT_POOL_ADDRESS_P (base
))
1910 case UNSPEC_VOLATILE
:
1915 if (MEM_VOLATILE_P (x
))
1924 /* Recursively scan the operands of this expression. */
1927 register const char *fmt
= GET_RTX_FORMAT (code
);
1930 for (i
= GET_RTX_LENGTH (code
) - 1; i
>= 0; i
--)
1932 if (fmt
[i
] == 'e' && XEXP (x
, i
))
1934 if (nonlocal_mentioned_p (XEXP (x
, i
)))
1937 else if (fmt
[i
] == 'E')
1940 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
1941 if (nonlocal_mentioned_p (XVECEXP (x
, i
, j
)))
1950 /* Return non-zero if a loop (natural or otherwise) is present.
1951 Inspired by Depth_First_Search_PP described in:
1953 Advanced Compiler Design and Implementation
1955 Morgan Kaufmann, 1997
1957 and heavily borrowed from flow_depth_first_order_compute. */
1970 /* Allocate the preorder and postorder number arrays. */
1971 pre
= (int *) xcalloc (n_basic_blocks
, sizeof (int));
1972 post
= (int *) xcalloc (n_basic_blocks
, sizeof (int));
1974 /* Allocate stack for back-tracking up CFG. */
1975 stack
= (edge
*) xmalloc ((n_basic_blocks
+ 1) * sizeof (edge
));
1978 /* Allocate bitmap to track nodes that have been visited. */
1979 visited
= sbitmap_alloc (n_basic_blocks
);
1981 /* None of the nodes in the CFG have been visited yet. */
1982 sbitmap_zero (visited
);
1984 /* Push the first edge on to the stack. */
1985 stack
[sp
++] = ENTRY_BLOCK_PTR
->succ
;
1993 /* Look at the edge on the top of the stack. */
1998 /* Check if the edge destination has been visited yet. */
1999 if (dest
!= EXIT_BLOCK_PTR
&& ! TEST_BIT (visited
, dest
->index
))
2001 /* Mark that we have visited the destination. */
2002 SET_BIT (visited
, dest
->index
);
2004 pre
[dest
->index
] = prenum
++;
2008 /* Since the DEST node has been visited for the first
2009 time, check its successors. */
2010 stack
[sp
++] = dest
->succ
;
2013 post
[dest
->index
] = postnum
++;
2017 if (dest
!= EXIT_BLOCK_PTR
2018 && pre
[src
->index
] >= pre
[dest
->index
]
2019 && post
[dest
->index
] == 0)
2022 if (! e
->succ_next
&& src
!= ENTRY_BLOCK_PTR
)
2023 post
[src
->index
] = postnum
++;
2026 stack
[sp
- 1] = e
->succ_next
;
2035 sbitmap_free (visited
);
2040 /* Mark the function if it is constant. */
2043 mark_constant_function ()
2046 int nonlocal_mentioned
;
2048 if (TREE_PUBLIC (current_function_decl
)
2049 || TREE_READONLY (current_function_decl
)
2050 || DECL_IS_PURE (current_function_decl
)
2051 || TREE_THIS_VOLATILE (current_function_decl
)
2052 || TYPE_MODE (TREE_TYPE (current_function_decl
)) == VOIDmode
)
2055 /* A loop might not return which counts as a side effect. */
2059 nonlocal_mentioned
= 0;
2061 init_alias_analysis ();
2063 /* Determine if this is a constant function. */
2065 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
2066 if (INSN_P (insn
) && nonlocal_mentioned_p (insn
))
2068 nonlocal_mentioned
= 1;
2072 end_alias_analysis ();
2074 /* Mark the function. */
2076 if (! nonlocal_mentioned
)
2077 TREE_READONLY (current_function_decl
) = 1;
2081 static HARD_REG_SET argument_registers
;
2088 #ifndef OUTGOING_REGNO
2089 #define OUTGOING_REGNO(N) N
2091 for (i
= 0; i
< FIRST_PSEUDO_REGISTER
; i
++)
2092 /* Check whether this register can hold an incoming pointer
2093 argument. FUNCTION_ARG_REGNO_P tests outgoing register
2094 numbers, so translate if necessary due to register windows. */
2095 if (FUNCTION_ARG_REGNO_P (OUTGOING_REGNO (i
))
2096 && HARD_REGNO_MODE_OK (i
, Pmode
))
2097 SET_HARD_REG_BIT (argument_registers
, i
);
2099 alias_sets
= splay_tree_new (splay_tree_compare_ints
, 0, 0);
2102 /* Initialize the aliasing machinery. Initialize the REG_KNOWN_VALUE
2106 init_alias_analysis ()
2108 int maxreg
= max_reg_num ();
2111 register unsigned int ui
;
2114 reg_known_value_size
= maxreg
;
2117 = (rtx
*) xcalloc ((maxreg
- FIRST_PSEUDO_REGISTER
), sizeof (rtx
))
2118 - FIRST_PSEUDO_REGISTER
;
2120 = (char*) xcalloc ((maxreg
- FIRST_PSEUDO_REGISTER
), sizeof (char))
2121 - FIRST_PSEUDO_REGISTER
;
2123 /* Overallocate reg_base_value to allow some growth during loop
2124 optimization. Loop unrolling can create a large number of
2126 reg_base_value_size
= maxreg
* 2;
2127 reg_base_value
= (rtx
*) xcalloc (reg_base_value_size
, sizeof (rtx
));
2128 ggc_add_rtx_root (reg_base_value
, reg_base_value_size
);
2130 new_reg_base_value
= (rtx
*) xmalloc (reg_base_value_size
* sizeof (rtx
));
2131 reg_seen
= (char *) xmalloc (reg_base_value_size
);
2132 if (! reload_completed
&& flag_unroll_loops
)
2134 /* ??? Why are we realloc'ing if we're just going to zero it? */
2135 alias_invariant
= (rtx
*)xrealloc (alias_invariant
,
2136 reg_base_value_size
* sizeof (rtx
));
2137 memset ((char *)alias_invariant
, 0, reg_base_value_size
* sizeof (rtx
));
2140 /* The basic idea is that each pass through this loop will use the
2141 "constant" information from the previous pass to propagate alias
2142 information through another level of assignments.
2144 This could get expensive if the assignment chains are long. Maybe
2145 we should throttle the number of iterations, possibly based on
2146 the optimization level or flag_expensive_optimizations.
2148 We could propagate more information in the first pass by making use
2149 of REG_N_SETS to determine immediately that the alias information
2150 for a pseudo is "constant".
2152 A program with an uninitialized variable can cause an infinite loop
2153 here. Instead of doing a full dataflow analysis to detect such problems
2154 we just cap the number of iterations for the loop.
2156 The state of the arrays for the set chain in question does not matter
2157 since the program has undefined behavior. */
2162 /* Assume nothing will change this iteration of the loop. */
2165 /* We want to assign the same IDs each iteration of this loop, so
2166 start counting from zero each iteration of the loop. */
2169 /* We're at the start of the funtion each iteration through the
2170 loop, so we're copying arguments. */
2171 copying_arguments
= 1;
2173 /* Wipe the potential alias information clean for this pass. */
2174 memset ((char *) new_reg_base_value
, 0, reg_base_value_size
* sizeof (rtx
));
2176 /* Wipe the reg_seen array clean. */
2177 memset ((char *) reg_seen
, 0, reg_base_value_size
);
2179 /* Mark all hard registers which may contain an address.
2180 The stack, frame and argument pointers may contain an address.
2181 An argument register which can hold a Pmode value may contain
2182 an address even if it is not in BASE_REGS.
2184 The address expression is VOIDmode for an argument and
2185 Pmode for other registers. */
2187 for (i
= 0; i
< FIRST_PSEUDO_REGISTER
; i
++)
2188 if (TEST_HARD_REG_BIT (argument_registers
, i
))
2189 new_reg_base_value
[i
] = gen_rtx_ADDRESS (VOIDmode
,
2190 gen_rtx_REG (Pmode
, i
));
2192 new_reg_base_value
[STACK_POINTER_REGNUM
]
2193 = gen_rtx_ADDRESS (Pmode
, stack_pointer_rtx
);
2194 new_reg_base_value
[ARG_POINTER_REGNUM
]
2195 = gen_rtx_ADDRESS (Pmode
, arg_pointer_rtx
);
2196 new_reg_base_value
[FRAME_POINTER_REGNUM
]
2197 = gen_rtx_ADDRESS (Pmode
, frame_pointer_rtx
);
2198 #if HARD_FRAME_POINTER_REGNUM != FRAME_POINTER_REGNUM
2199 new_reg_base_value
[HARD_FRAME_POINTER_REGNUM
]
2200 = gen_rtx_ADDRESS (Pmode
, hard_frame_pointer_rtx
);
2203 /* Walk the insns adding values to the new_reg_base_value array. */
2204 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
2210 #if defined (HAVE_prologue) || defined (HAVE_epilogue)
2211 /* The prologue/epilouge insns are not threaded onto the
2212 insn chain until after reload has completed. Thus,
2213 there is no sense wasting time checking if INSN is in
2214 the prologue/epilogue until after reload has completed. */
2215 if (reload_completed
2216 && prologue_epilogue_contains (insn
))
2220 /* If this insn has a noalias note, process it, Otherwise,
2221 scan for sets. A simple set will have no side effects
2222 which could change the base value of any other register. */
2224 if (GET_CODE (PATTERN (insn
)) == SET
2225 && REG_NOTES (insn
) != 0
2226 && find_reg_note (insn
, REG_NOALIAS
, NULL_RTX
))
2227 record_set (SET_DEST (PATTERN (insn
)), NULL_RTX
, NULL
);
2229 note_stores (PATTERN (insn
), record_set
, NULL
);
2231 set
= single_set (insn
);
2234 && GET_CODE (SET_DEST (set
)) == REG
2235 && REGNO (SET_DEST (set
)) >= FIRST_PSEUDO_REGISTER
)
2237 unsigned int regno
= REGNO (SET_DEST (set
));
2238 rtx src
= SET_SRC (set
);
2240 if (REG_NOTES (insn
) != 0
2241 && (((note
= find_reg_note (insn
, REG_EQUAL
, 0)) != 0
2242 && REG_N_SETS (regno
) == 1)
2243 || (note
= find_reg_note (insn
, REG_EQUIV
, NULL_RTX
)) != 0)
2244 && GET_CODE (XEXP (note
, 0)) != EXPR_LIST
2245 && ! reg_overlap_mentioned_p (SET_DEST (set
), XEXP (note
, 0)))
2247 reg_known_value
[regno
] = XEXP (note
, 0);
2248 reg_known_equiv_p
[regno
] = REG_NOTE_KIND (note
) == REG_EQUIV
;
2250 else if (REG_N_SETS (regno
) == 1
2251 && GET_CODE (src
) == PLUS
2252 && GET_CODE (XEXP (src
, 0)) == REG
2253 && REGNO (XEXP (src
, 0)) >= FIRST_PSEUDO_REGISTER
2254 && (reg_known_value
[REGNO (XEXP (src
, 0))])
2255 && GET_CODE (XEXP (src
, 1)) == CONST_INT
)
2257 rtx op0
= XEXP (src
, 0);
2258 if (reg_known_value
[REGNO (op0
)])
2259 op0
= reg_known_value
[REGNO (op0
)];
2260 reg_known_value
[regno
]
2261 = plus_constant_for_output (op0
,
2262 INTVAL (XEXP (src
, 1)));
2263 reg_known_equiv_p
[regno
] = 0;
2265 else if (REG_N_SETS (regno
) == 1
2266 && ! rtx_varies_p (src
, 1))
2268 reg_known_value
[regno
] = src
;
2269 reg_known_equiv_p
[regno
] = 0;
2273 else if (GET_CODE (insn
) == NOTE
2274 && NOTE_LINE_NUMBER (insn
) == NOTE_INSN_FUNCTION_BEG
)
2275 copying_arguments
= 0;
2278 /* Now propagate values from new_reg_base_value to reg_base_value. */
2279 for (ui
= 0; ui
< reg_base_value_size
; ui
++)
2281 if (new_reg_base_value
[ui
]
2282 && new_reg_base_value
[ui
] != reg_base_value
[ui
]
2283 && ! rtx_equal_p (new_reg_base_value
[ui
], reg_base_value
[ui
]))
2285 reg_base_value
[ui
] = new_reg_base_value
[ui
];
2290 while (changed
&& ++pass
< MAX_ALIAS_LOOP_PASSES
);
2292 /* Fill in the remaining entries. */
2293 for (i
= FIRST_PSEUDO_REGISTER
; i
< maxreg
; i
++)
2294 if (reg_known_value
[i
] == 0)
2295 reg_known_value
[i
] = regno_reg_rtx
[i
];
2297 /* Simplify the reg_base_value array so that no register refers to
2298 another register, except to special registers indirectly through
2299 ADDRESS expressions.
2301 In theory this loop can take as long as O(registers^2), but unless
2302 there are very long dependency chains it will run in close to linear
2305 This loop may not be needed any longer now that the main loop does
2306 a better job at propagating alias information. */
2312 for (ui
= 0; ui
< reg_base_value_size
; ui
++)
2314 rtx base
= reg_base_value
[ui
];
2315 if (base
&& GET_CODE (base
) == REG
)
2317 unsigned int base_regno
= REGNO (base
);
2318 if (base_regno
== ui
) /* register set from itself */
2319 reg_base_value
[ui
] = 0;
2321 reg_base_value
[ui
] = reg_base_value
[base_regno
];
2326 while (changed
&& pass
< MAX_ALIAS_LOOP_PASSES
);
2329 free (new_reg_base_value
);
2330 new_reg_base_value
= 0;
2336 end_alias_analysis ()
2338 free (reg_known_value
+ FIRST_PSEUDO_REGISTER
);
2339 reg_known_value
= 0;
2340 reg_known_value_size
= 0;
2341 free (reg_known_equiv_p
+ FIRST_PSEUDO_REGISTER
);
2342 reg_known_equiv_p
= 0;
2345 ggc_del_root (reg_base_value
);
2346 free (reg_base_value
);
2349 reg_base_value_size
= 0;
2350 if (alias_invariant
)
2352 free (alias_invariant
);
2353 alias_invariant
= 0;