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. */
30 #include "hard-reg-set.h"
31 #include "basic-block.h"
36 #include "splay-tree.h"
39 /* The alias sets assigned to MEMs assist the back-end in determining
40 which MEMs can alias which other MEMs. In general, two MEMs in
41 different alias sets cannot alias each other, with one important
42 exception. Consider something like:
44 struct S {int i; double d; };
46 a store to an `S' can alias something of either type `int' or type
47 `double'. (However, a store to an `int' cannot alias a `double'
48 and vice versa.) We indicate this via a tree structure that looks
56 (The arrows are directed and point downwards.)
57 In this situation we say the alias set for `struct S' is the
58 `superset' and that those for `int' and `double' are `subsets'.
60 To see whether two alias sets can point to the same memory, we must
61 see if either alias set is a subset of the other. We need not trace
62 past immediate decendents, however, since we propagate all
63 grandchildren up one level.
65 Alias set zero is implicitly a superset of all other alias sets.
66 However, this is no actual entry for alias set zero. It is an
67 error to attempt to explicitly construct a subset of zero. */
69 typedef struct alias_set_entry
71 /* The alias set number, as stored in MEM_ALIAS_SET. */
72 HOST_WIDE_INT alias_set;
74 /* The children of the alias set. These are not just the immediate
75 children, but, in fact, all decendents. So, if we have:
77 struct T { struct S s; float f; }
79 continuing our example above, the children here will be all of
80 `int', `double', `float', and `struct S'. */
83 /* Nonzero if would have a child of zero: this effectively makes this
84 alias set the same as alias set zero. */
88 static int rtx_equal_for_memref_p PARAMS ((rtx, rtx));
89 static rtx find_symbolic_term PARAMS ((rtx));
90 rtx get_addr PARAMS ((rtx));
91 static int memrefs_conflict_p PARAMS ((int, rtx, int, rtx,
93 static void record_set PARAMS ((rtx, rtx, void *));
94 static rtx find_base_term PARAMS ((rtx));
95 static int base_alias_check PARAMS ((rtx, rtx, enum machine_mode,
97 static int handled_component_p PARAMS ((tree));
98 static int can_address_p PARAMS ((tree));
99 static rtx find_base_value PARAMS ((rtx));
100 static int mems_in_disjoint_alias_sets_p PARAMS ((rtx, rtx));
101 static int insert_subset_children PARAMS ((splay_tree_node, void*));
102 static tree find_base_decl PARAMS ((tree));
103 static alias_set_entry get_alias_set_entry PARAMS ((HOST_WIDE_INT));
104 static rtx fixed_scalar_and_varying_struct_p PARAMS ((rtx, rtx, rtx, rtx,
105 int (*) (rtx, int)));
106 static int aliases_everything_p PARAMS ((rtx));
107 static int write_dependence_p PARAMS ((rtx, rtx, int));
108 static int nonlocal_mentioned_p PARAMS ((rtx));
110 static int loop_p PARAMS ((void));
112 /* Set up all info needed to perform alias analysis on memory references. */
114 /* Returns the size in bytes of the mode of X. */
115 #define SIZE_FOR_MODE(X) (GET_MODE_SIZE (GET_MODE (X)))
117 /* Returns nonzero if MEM1 and MEM2 do not alias because they are in
118 different alias sets. We ignore alias sets in functions making use
119 of variable arguments because the va_arg macros on some systems are
121 #define DIFFERENT_ALIAS_SETS_P(MEM1, MEM2) \
122 mems_in_disjoint_alias_sets_p (MEM1, MEM2)
124 /* Cap the number of passes we make over the insns propagating alias
125 information through set chains. 10 is a completely arbitrary choice. */
126 #define MAX_ALIAS_LOOP_PASSES 10
128 /* reg_base_value[N] gives an address to which register N is related.
129 If all sets after the first add or subtract to the current value
130 or otherwise modify it so it does not point to a different top level
131 object, reg_base_value[N] is equal to the address part of the source
134 A base address can be an ADDRESS, SYMBOL_REF, or LABEL_REF. ADDRESS
135 expressions represent certain special values: function arguments and
136 the stack, frame, and argument pointers.
138 The contents of an ADDRESS is not normally used, the mode of the
139 ADDRESS determines whether the ADDRESS is a function argument or some
140 other special value. Pointer equality, not rtx_equal_p, determines whether
141 two ADDRESS expressions refer to the same base address.
143 The only use of the contents of an ADDRESS is for determining if the
144 current function performs nonlocal memory memory references for the
145 purposes of marking the function as a constant function. */
147 static rtx *reg_base_value;
148 static rtx *new_reg_base_value;
149 static unsigned int reg_base_value_size; /* size of reg_base_value array */
151 #define REG_BASE_VALUE(X) \
152 (REGNO (X) < reg_base_value_size \
153 ? reg_base_value[REGNO (X)] : 0)
155 /* Vector of known invariant relationships between registers. Set in
156 loop unrolling. Indexed by register number, if nonzero the value
157 is an expression describing this register in terms of another.
159 The length of this array is REG_BASE_VALUE_SIZE.
161 Because this array contains only pseudo registers it has no effect
163 static rtx *alias_invariant;
165 /* Vector indexed by N giving the initial (unchanging) value known for
166 pseudo-register N. This array is initialized in
167 init_alias_analysis, and does not change until end_alias_analysis
169 rtx *reg_known_value;
171 /* Indicates number of valid entries in reg_known_value. */
172 static unsigned int reg_known_value_size;
174 /* Vector recording for each reg_known_value whether it is due to a
175 REG_EQUIV note. Future passes (viz., reload) may replace the
176 pseudo with the equivalent expression and so we account for the
177 dependences that would be introduced if that happens.
179 The REG_EQUIV notes created in assign_parms may mention the arg
180 pointer, and there are explicit insns in the RTL that modify the
181 arg pointer. Thus we must ensure that such insns don't get
182 scheduled across each other because that would invalidate the
183 REG_EQUIV notes. One could argue that the REG_EQUIV notes are
184 wrong, but solving the problem in the scheduler will likely give
185 better code, so we do it here. */
186 char *reg_known_equiv_p;
188 /* True when scanning insns from the start of the rtl to the
189 NOTE_INSN_FUNCTION_BEG note. */
190 static int copying_arguments;
192 /* The splay-tree used to store the various alias set entries. */
193 static splay_tree alias_sets;
195 /* Returns a pointer to the alias set entry for ALIAS_SET, if there is
196 such an entry, or NULL otherwise. */
198 static alias_set_entry
199 get_alias_set_entry (alias_set)
200 HOST_WIDE_INT alias_set;
203 = splay_tree_lookup (alias_sets, (splay_tree_key) alias_set);
205 return sn != 0 ? ((alias_set_entry) sn->value) : 0;
208 /* Returns nonzero if the alias sets for MEM1 and MEM2 are such that
209 the two MEMs cannot alias each other. */
212 mems_in_disjoint_alias_sets_p (mem1, mem2)
216 #ifdef ENABLE_CHECKING
217 /* Perform a basic sanity check. Namely, that there are no alias sets
218 if we're not using strict aliasing. This helps to catch bugs
219 whereby someone uses PUT_CODE, but doesn't clear MEM_ALIAS_SET, or
220 where a MEM is allocated in some way other than by the use of
221 gen_rtx_MEM, and the MEM_ALIAS_SET is not cleared. If we begin to
222 use alias sets to indicate that spilled registers cannot alias each
223 other, we might need to remove this check. */
224 if (! flag_strict_aliasing
225 && (MEM_ALIAS_SET (mem1) != 0 || MEM_ALIAS_SET (mem2) != 0))
229 return ! alias_sets_conflict_p (MEM_ALIAS_SET (mem1), MEM_ALIAS_SET (mem2));
232 /* Insert the NODE into the splay tree given by DATA. Used by
233 record_alias_subset via splay_tree_foreach. */
236 insert_subset_children (node, data)
237 splay_tree_node node;
240 splay_tree_insert ((splay_tree) data, node->key, node->value);
245 /* Return 1 if the two specified alias sets may conflict. */
248 alias_sets_conflict_p (set1, set2)
249 HOST_WIDE_INT set1, set2;
253 /* If have no alias set information for one of the operands, we have
254 to assume it can alias anything. */
255 if (set1 == 0 || set2 == 0
256 /* If the two alias sets are the same, they may alias. */
260 /* See if the first alias set is a subset of the second. */
261 ase = get_alias_set_entry (set1);
263 && (ase->has_zero_child
264 || splay_tree_lookup (ase->children,
265 (splay_tree_key) set2)))
268 /* Now do the same, but with the alias sets reversed. */
269 ase = get_alias_set_entry (set2);
271 && (ase->has_zero_child
272 || splay_tree_lookup (ase->children,
273 (splay_tree_key) set1)))
276 /* The two alias sets are distinct and neither one is the
277 child of the other. Therefore, they cannot alias. */
281 /* Return 1 if TYPE is a RECORD_TYPE, UNION_TYPE, or QUAL_UNION_TYPE and has
282 has any readonly fields. If any of the fields have types that
283 contain readonly fields, return true as well. */
286 readonly_fields_p (type)
291 if (TREE_CODE (type) != RECORD_TYPE && TREE_CODE (type) != UNION_TYPE
292 && TREE_CODE (type) != QUAL_UNION_TYPE)
295 for (field = TYPE_FIELDS (type); field != 0; field = TREE_CHAIN (field))
296 if (TREE_CODE (field) == FIELD_DECL
297 && (TREE_READONLY (field)
298 || readonly_fields_p (TREE_TYPE (field))))
304 /* Return 1 if any MEM object of type T1 will always conflict (using the
305 dependency routines in this file) with any MEM object of type T2.
306 This is used when allocating temporary storage. If T1 and/or T2 are
307 NULL_TREE, it means we know nothing about the storage. */
310 objects_must_conflict_p (t1, t2)
313 /* If neither has a type specified, we don't know if they'll conflict
314 because we may be using them to store objects of various types, for
315 example the argument and local variables areas of inlined functions. */
316 if (t1 == 0 && t2 == 0)
319 /* If one or the other has readonly fields or is readonly,
320 then they may not conflict. */
321 if ((t1 != 0 && readonly_fields_p (t1))
322 || (t2 != 0 && readonly_fields_p (t2))
323 || (t1 != 0 && TYPE_READONLY (t1))
324 || (t2 != 0 && TYPE_READONLY (t2)))
327 /* If they are the same type, they must conflict. */
329 /* Likewise if both are volatile. */
330 || (t1 != 0 && TYPE_VOLATILE (t1) && t2 != 0 && TYPE_VOLATILE (t2)))
333 /* If one is aggregate and the other is scalar then they may not
335 if ((t1 != 0 && AGGREGATE_TYPE_P (t1))
336 != (t2 != 0 && AGGREGATE_TYPE_P (t2)))
339 /* Otherwise they conflict only if the alias sets conflict. */
340 return alias_sets_conflict_p (t1 ? get_alias_set (t1) : 0,
341 t2 ? get_alias_set (t2) : 0);
344 /* T is an expression with pointer type. Find the DECL on which this
345 expression is based. (For example, in `a[i]' this would be `a'.)
346 If there is no such DECL, or a unique decl cannot be determined,
347 NULL_TREE is retured. */
355 if (t == 0 || t == error_mark_node || ! POINTER_TYPE_P (TREE_TYPE (t)))
358 /* If this is a declaration, return it. */
359 if (TREE_CODE_CLASS (TREE_CODE (t)) == 'd')
362 /* Handle general expressions. It would be nice to deal with
363 COMPONENT_REFs here. If we could tell that `a' and `b' were the
364 same, then `a->f' and `b->f' are also the same. */
365 switch (TREE_CODE_CLASS (TREE_CODE (t)))
368 return find_base_decl (TREE_OPERAND (t, 0));
371 /* Return 0 if found in neither or both are the same. */
372 d0 = find_base_decl (TREE_OPERAND (t, 0));
373 d1 = find_base_decl (TREE_OPERAND (t, 1));
384 d0 = find_base_decl (TREE_OPERAND (t, 0));
385 d1 = find_base_decl (TREE_OPERAND (t, 1));
386 d0 = find_base_decl (TREE_OPERAND (t, 0));
387 d2 = find_base_decl (TREE_OPERAND (t, 2));
389 /* Set any nonzero values from the last, then from the first. */
390 if (d1 == 0) d1 = d2;
391 if (d0 == 0) d0 = d1;
392 if (d1 == 0) d1 = d0;
393 if (d2 == 0) d2 = d1;
395 /* At this point all are nonzero or all are zero. If all three are the
396 same, return it. Otherwise, return zero. */
397 return (d0 == d1 && d1 == d2) ? d0 : 0;
404 /* Return 1 if T is an expression that get_inner_reference handles. */
407 handled_component_p (t)
410 switch (TREE_CODE (t))
415 case ARRAY_RANGE_REF:
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 || TREE_CODE (t) == ARRAY_RANGE_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 if (GET_CODE (x0) == CONST_INT)
963 return plus_constant (x1, INTVAL (x0));
964 else if (GET_CODE (x1) == CONST_INT)
965 return plus_constant (x0, INTVAL (x1));
966 return gen_rtx_PLUS (GET_MODE (x), x0, x1);
970 /* This gives us much better alias analysis when called from
971 the loop optimizer. Note we want to leave the original
972 MEM alone, but need to return the canonicalized MEM with
973 all the flags with their original values. */
974 else if (GET_CODE (x) == MEM)
975 x = replace_equiv_address_nv (x, canon_rtx (XEXP (x, 0)));
980 /* Return 1 if X and Y are identical-looking rtx's.
982 We use the data in reg_known_value above to see if two registers with
983 different numbers are, in fact, equivalent. */
986 rtx_equal_for_memref_p (x, y)
991 register enum rtx_code code;
992 register const char *fmt;
994 if (x == 0 && y == 0)
996 if (x == 0 || y == 0)
1005 code = GET_CODE (x);
1006 /* Rtx's of different codes cannot be equal. */
1007 if (code != GET_CODE (y))
1010 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent.
1011 (REG:SI x) and (REG:HI x) are NOT equivalent. */
1013 if (GET_MODE (x) != GET_MODE (y))
1016 /* Some RTL can be compared without a recursive examination. */
1020 return REGNO (x) == REGNO (y);
1023 return XEXP (x, 0) == XEXP (y, 0);
1026 return XSTR (x, 0) == XSTR (y, 0);
1030 /* There's no need to compare the contents of CONST_DOUBLEs or
1031 CONST_INTs because pointer equality is a good enough
1032 comparison for these nodes. */
1036 return (XINT (x, 1) == XINT (y, 1)
1037 && rtx_equal_for_memref_p (XEXP (x, 0), XEXP (y, 0)));
1043 /* For commutative operations, the RTX match if the operand match in any
1044 order. Also handle the simple binary and unary cases without a loop. */
1045 if (code == EQ || code == NE || GET_RTX_CLASS (code) == 'c')
1046 return ((rtx_equal_for_memref_p (XEXP (x, 0), XEXP (y, 0))
1047 && rtx_equal_for_memref_p (XEXP (x, 1), XEXP (y, 1)))
1048 || (rtx_equal_for_memref_p (XEXP (x, 0), XEXP (y, 1))
1049 && rtx_equal_for_memref_p (XEXP (x, 1), XEXP (y, 0))));
1050 else if (GET_RTX_CLASS (code) == '<' || GET_RTX_CLASS (code) == '2')
1051 return (rtx_equal_for_memref_p (XEXP (x, 0), XEXP (y, 0))
1052 && rtx_equal_for_memref_p (XEXP (x, 1), XEXP (y, 1)));
1053 else if (GET_RTX_CLASS (code) == '1')
1054 return rtx_equal_for_memref_p (XEXP (x, 0), XEXP (y, 0));
1056 /* Compare the elements. If any pair of corresponding elements
1057 fail to match, return 0 for the whole things.
1059 Limit cases to types which actually appear in addresses. */
1061 fmt = GET_RTX_FORMAT (code);
1062 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1067 if (XINT (x, i) != XINT (y, i))
1072 /* Two vectors must have the same length. */
1073 if (XVECLEN (x, i) != XVECLEN (y, i))
1076 /* And the corresponding elements must match. */
1077 for (j = 0; j < XVECLEN (x, i); j++)
1078 if (rtx_equal_for_memref_p (XVECEXP (x, i, j),
1079 XVECEXP (y, i, j)) == 0)
1084 if (rtx_equal_for_memref_p (XEXP (x, i), XEXP (y, i)) == 0)
1088 /* This can happen for an asm which clobbers memory. */
1092 /* It is believed that rtx's at this level will never
1093 contain anything but integers and other rtx's,
1094 except for within LABEL_REFs and SYMBOL_REFs. */
1102 /* Given an rtx X, find a SYMBOL_REF or LABEL_REF within
1103 X and return it, or return 0 if none found. */
1106 find_symbolic_term (x)
1110 register enum rtx_code code;
1111 register const char *fmt;
1113 code = GET_CODE (x);
1114 if (code == SYMBOL_REF || code == LABEL_REF)
1116 if (GET_RTX_CLASS (code) == 'o')
1119 fmt = GET_RTX_FORMAT (code);
1120 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1126 t = find_symbolic_term (XEXP (x, i));
1130 else if (fmt[i] == 'E')
1141 struct elt_loc_list *l;
1143 #if defined (FIND_BASE_TERM)
1144 /* Try machine-dependent ways to find the base term. */
1145 x = FIND_BASE_TERM (x);
1148 switch (GET_CODE (x))
1151 return REG_BASE_VALUE (x);
1154 case SIGN_EXTEND: /* Used for Alpha/NT pointers */
1160 return find_base_term (XEXP (x, 0));
1163 val = CSELIB_VAL_PTR (x);
1164 for (l = val->locs; l; l = l->next)
1165 if ((x = find_base_term (l->loc)) != 0)
1171 if (GET_CODE (x) != PLUS && GET_CODE (x) != MINUS)
1178 rtx tmp1 = XEXP (x, 0);
1179 rtx tmp2 = XEXP (x, 1);
1181 /* This is a litle bit tricky since we have to determine which of
1182 the two operands represents the real base address. Otherwise this
1183 routine may return the index register instead of the base register.
1185 That may cause us to believe no aliasing was possible, when in
1186 fact aliasing is possible.
1188 We use a few simple tests to guess the base register. Additional
1189 tests can certainly be added. For example, if one of the operands
1190 is a shift or multiply, then it must be the index register and the
1191 other operand is the base register. */
1193 if (tmp1 == pic_offset_table_rtx && CONSTANT_P (tmp2))
1194 return find_base_term (tmp2);
1196 /* If either operand is known to be a pointer, then use it
1197 to determine the base term. */
1198 if (REG_P (tmp1) && REG_POINTER (tmp1))
1199 return find_base_term (tmp1);
1201 if (REG_P (tmp2) && REG_POINTER (tmp2))
1202 return find_base_term (tmp2);
1204 /* Neither operand was known to be a pointer. Go ahead and find the
1205 base term for both operands. */
1206 tmp1 = find_base_term (tmp1);
1207 tmp2 = find_base_term (tmp2);
1209 /* If either base term is named object or a special address
1210 (like an argument or stack reference), then use it for the
1213 && (GET_CODE (tmp1) == SYMBOL_REF
1214 || GET_CODE (tmp1) == LABEL_REF
1215 || (GET_CODE (tmp1) == ADDRESS
1216 && GET_MODE (tmp1) != VOIDmode)))
1220 && (GET_CODE (tmp2) == SYMBOL_REF
1221 || GET_CODE (tmp2) == LABEL_REF
1222 || (GET_CODE (tmp2) == ADDRESS
1223 && GET_MODE (tmp2) != VOIDmode)))
1226 /* We could not determine which of the two operands was the
1227 base register and which was the index. So we can determine
1228 nothing from the base alias check. */
1233 if (GET_CODE (XEXP (x, 0)) == REG && GET_CODE (XEXP (x, 1)) == CONST_INT)
1234 return REG_BASE_VALUE (XEXP (x, 0));
1242 return REG_BASE_VALUE (frame_pointer_rtx);
1249 /* Return 0 if the addresses X and Y are known to point to different
1250 objects, 1 if they might be pointers to the same object. */
1253 base_alias_check (x, y, x_mode, y_mode)
1255 enum machine_mode x_mode, y_mode;
1257 rtx x_base = find_base_term (x);
1258 rtx y_base = find_base_term (y);
1260 /* If the address itself has no known base see if a known equivalent
1261 value has one. If either address still has no known base, nothing
1262 is known about aliasing. */
1267 if (! flag_expensive_optimizations || (x_c = canon_rtx (x)) == x)
1270 x_base = find_base_term (x_c);
1278 if (! flag_expensive_optimizations || (y_c = canon_rtx (y)) == y)
1281 y_base = find_base_term (y_c);
1286 /* If the base addresses are equal nothing is known about aliasing. */
1287 if (rtx_equal_p (x_base, y_base))
1290 /* The base addresses of the read and write are different expressions.
1291 If they are both symbols and they are not accessed via AND, there is
1292 no conflict. We can bring knowledge of object alignment into play
1293 here. For example, on alpha, "char a, b;" can alias one another,
1294 though "char a; long b;" cannot. */
1295 if (GET_CODE (x_base) != ADDRESS && GET_CODE (y_base) != ADDRESS)
1297 if (GET_CODE (x) == AND && GET_CODE (y) == AND)
1299 if (GET_CODE (x) == AND
1300 && (GET_CODE (XEXP (x, 1)) != CONST_INT
1301 || (int) GET_MODE_UNIT_SIZE (y_mode) < -INTVAL (XEXP (x, 1))))
1303 if (GET_CODE (y) == AND
1304 && (GET_CODE (XEXP (y, 1)) != CONST_INT
1305 || (int) GET_MODE_UNIT_SIZE (x_mode) < -INTVAL (XEXP (y, 1))))
1307 /* Differing symbols never alias. */
1311 /* If one address is a stack reference there can be no alias:
1312 stack references using different base registers do not alias,
1313 a stack reference can not alias a parameter, and a stack reference
1314 can not alias a global. */
1315 if ((GET_CODE (x_base) == ADDRESS && GET_MODE (x_base) == Pmode)
1316 || (GET_CODE (y_base) == ADDRESS && GET_MODE (y_base) == Pmode))
1319 if (! flag_argument_noalias)
1322 if (flag_argument_noalias > 1)
1325 /* Weak noalias assertion (arguments are distinct, but may match globals). */
1326 return ! (GET_MODE (x_base) == VOIDmode && GET_MODE (y_base) == VOIDmode);
1329 /* Convert the address X into something we can use. This is done by returning
1330 it unchanged unless it is a value; in the latter case we call cselib to get
1331 a more useful rtx. */
1338 struct elt_loc_list *l;
1340 if (GET_CODE (x) != VALUE)
1342 v = CSELIB_VAL_PTR (x);
1343 for (l = v->locs; l; l = l->next)
1344 if (CONSTANT_P (l->loc))
1346 for (l = v->locs; l; l = l->next)
1347 if (GET_CODE (l->loc) != REG && GET_CODE (l->loc) != MEM)
1350 return v->locs->loc;
1354 /* Return the address of the (N_REFS + 1)th memory reference to ADDR
1355 where SIZE is the size in bytes of the memory reference. If ADDR
1356 is not modified by the memory reference then ADDR is returned. */
1359 addr_side_effect_eval (addr, size, n_refs)
1366 switch (GET_CODE (addr))
1369 offset = (n_refs + 1) * size;
1372 offset = -(n_refs + 1) * size;
1375 offset = n_refs * size;
1378 offset = -n_refs * size;
1386 addr = gen_rtx_PLUS (GET_MODE (addr), XEXP (addr, 0), GEN_INT (offset));
1388 addr = XEXP (addr, 0);
1393 /* Return nonzero if X and Y (memory addresses) could reference the
1394 same location in memory. C is an offset accumulator. When
1395 C is nonzero, we are testing aliases between X and Y + C.
1396 XSIZE is the size in bytes of the X reference,
1397 similarly YSIZE is the size in bytes for Y.
1399 If XSIZE or YSIZE is zero, we do not know the amount of memory being
1400 referenced (the reference was BLKmode), so make the most pessimistic
1403 If XSIZE or YSIZE is negative, we may access memory outside the object
1404 being referenced as a side effect. This can happen when using AND to
1405 align memory references, as is done on the Alpha.
1407 Nice to notice that varying addresses cannot conflict with fp if no
1408 local variables had their addresses taken, but that's too hard now. */
1411 memrefs_conflict_p (xsize, x, ysize, y, c)
1416 if (GET_CODE (x) == VALUE)
1418 if (GET_CODE (y) == VALUE)
1420 if (GET_CODE (x) == HIGH)
1422 else if (GET_CODE (x) == LO_SUM)
1425 x = canon_rtx (addr_side_effect_eval (x, xsize, 0));
1426 if (GET_CODE (y) == HIGH)
1428 else if (GET_CODE (y) == LO_SUM)
1431 y = canon_rtx (addr_side_effect_eval (y, ysize, 0));
1433 if (rtx_equal_for_memref_p (x, y))
1435 if (xsize <= 0 || ysize <= 0)
1437 if (c >= 0 && xsize > c)
1439 if (c < 0 && ysize+c > 0)
1444 /* This code used to check for conflicts involving stack references and
1445 globals but the base address alias code now handles these cases. */
1447 if (GET_CODE (x) == PLUS)
1449 /* The fact that X is canonicalized means that this
1450 PLUS rtx is canonicalized. */
1451 rtx x0 = XEXP (x, 0);
1452 rtx x1 = XEXP (x, 1);
1454 if (GET_CODE (y) == PLUS)
1456 /* The fact that Y is canonicalized means that this
1457 PLUS rtx is canonicalized. */
1458 rtx y0 = XEXP (y, 0);
1459 rtx y1 = XEXP (y, 1);
1461 if (rtx_equal_for_memref_p (x1, y1))
1462 return memrefs_conflict_p (xsize, x0, ysize, y0, c);
1463 if (rtx_equal_for_memref_p (x0, y0))
1464 return memrefs_conflict_p (xsize, x1, ysize, y1, c);
1465 if (GET_CODE (x1) == CONST_INT)
1467 if (GET_CODE (y1) == CONST_INT)
1468 return memrefs_conflict_p (xsize, x0, ysize, y0,
1469 c - INTVAL (x1) + INTVAL (y1));
1471 return memrefs_conflict_p (xsize, x0, ysize, y,
1474 else if (GET_CODE (y1) == CONST_INT)
1475 return memrefs_conflict_p (xsize, x, ysize, y0, c + INTVAL (y1));
1479 else if (GET_CODE (x1) == CONST_INT)
1480 return memrefs_conflict_p (xsize, x0, ysize, y, c - INTVAL (x1));
1482 else if (GET_CODE (y) == PLUS)
1484 /* The fact that Y is canonicalized means that this
1485 PLUS rtx is canonicalized. */
1486 rtx y0 = XEXP (y, 0);
1487 rtx y1 = XEXP (y, 1);
1489 if (GET_CODE (y1) == CONST_INT)
1490 return memrefs_conflict_p (xsize, x, ysize, y0, c + INTVAL (y1));
1495 if (GET_CODE (x) == GET_CODE (y))
1496 switch (GET_CODE (x))
1500 /* Handle cases where we expect the second operands to be the
1501 same, and check only whether the first operand would conflict
1504 rtx x1 = canon_rtx (XEXP (x, 1));
1505 rtx y1 = canon_rtx (XEXP (y, 1));
1506 if (! rtx_equal_for_memref_p (x1, y1))
1508 x0 = canon_rtx (XEXP (x, 0));
1509 y0 = canon_rtx (XEXP (y, 0));
1510 if (rtx_equal_for_memref_p (x0, y0))
1511 return (xsize == 0 || ysize == 0
1512 || (c >= 0 && xsize > c) || (c < 0 && ysize+c > 0));
1514 /* Can't properly adjust our sizes. */
1515 if (GET_CODE (x1) != CONST_INT)
1517 xsize /= INTVAL (x1);
1518 ysize /= INTVAL (x1);
1520 return memrefs_conflict_p (xsize, x0, ysize, y0, c);
1524 /* Are these registers known not to be equal? */
1525 if (alias_invariant)
1527 unsigned int r_x = REGNO (x), r_y = REGNO (y);
1528 rtx i_x, i_y; /* invariant relationships of X and Y */
1530 i_x = r_x >= reg_base_value_size ? 0 : alias_invariant[r_x];
1531 i_y = r_y >= reg_base_value_size ? 0 : alias_invariant[r_y];
1533 if (i_x == 0 && i_y == 0)
1536 if (! memrefs_conflict_p (xsize, i_x ? i_x : x,
1537 ysize, i_y ? i_y : y, c))
1546 /* Treat an access through an AND (e.g. a subword access on an Alpha)
1547 as an access with indeterminate size. Assume that references
1548 besides AND are aligned, so if the size of the other reference is
1549 at least as large as the alignment, assume no other overlap. */
1550 if (GET_CODE (x) == AND && GET_CODE (XEXP (x, 1)) == CONST_INT)
1552 if (GET_CODE (y) == AND || ysize < -INTVAL (XEXP (x, 1)))
1554 return memrefs_conflict_p (xsize, XEXP (x, 0), ysize, y, c);
1556 if (GET_CODE (y) == AND && GET_CODE (XEXP (y, 1)) == CONST_INT)
1558 /* ??? If we are indexing far enough into the array/structure, we
1559 may yet be able to determine that we can not overlap. But we
1560 also need to that we are far enough from the end not to overlap
1561 a following reference, so we do nothing with that for now. */
1562 if (GET_CODE (x) == AND || xsize < -INTVAL (XEXP (y, 1)))
1564 return memrefs_conflict_p (xsize, x, ysize, XEXP (y, 0), c);
1567 if (GET_CODE (x) == ADDRESSOF)
1569 if (y == frame_pointer_rtx
1570 || GET_CODE (y) == ADDRESSOF)
1571 return xsize <= 0 || ysize <= 0;
1573 if (GET_CODE (y) == ADDRESSOF)
1575 if (x == frame_pointer_rtx)
1576 return xsize <= 0 || ysize <= 0;
1581 if (GET_CODE (x) == CONST_INT && GET_CODE (y) == CONST_INT)
1583 c += (INTVAL (y) - INTVAL (x));
1584 return (xsize <= 0 || ysize <= 0
1585 || (c >= 0 && xsize > c) || (c < 0 && ysize+c > 0));
1588 if (GET_CODE (x) == CONST)
1590 if (GET_CODE (y) == CONST)
1591 return memrefs_conflict_p (xsize, canon_rtx (XEXP (x, 0)),
1592 ysize, canon_rtx (XEXP (y, 0)), c);
1594 return memrefs_conflict_p (xsize, canon_rtx (XEXP (x, 0)),
1597 if (GET_CODE (y) == CONST)
1598 return memrefs_conflict_p (xsize, x, ysize,
1599 canon_rtx (XEXP (y, 0)), c);
1602 return (xsize <= 0 || ysize <= 0
1603 || (rtx_equal_for_memref_p (x, y)
1604 && ((c >= 0 && xsize > c) || (c < 0 && ysize+c > 0))));
1611 /* Functions to compute memory dependencies.
1613 Since we process the insns in execution order, we can build tables
1614 to keep track of what registers are fixed (and not aliased), what registers
1615 are varying in known ways, and what registers are varying in unknown
1618 If both memory references are volatile, then there must always be a
1619 dependence between the two references, since their order can not be
1620 changed. A volatile and non-volatile reference can be interchanged
1623 A MEM_IN_STRUCT reference at a non-AND varying address can never
1624 conflict with a non-MEM_IN_STRUCT reference at a fixed address. We
1625 also must allow AND addresses, because they may generate accesses
1626 outside the object being referenced. This is used to generate
1627 aligned addresses from unaligned addresses, for instance, the alpha
1628 storeqi_unaligned pattern. */
1630 /* Read dependence: X is read after read in MEM takes place. There can
1631 only be a dependence here if both reads are volatile. */
1634 read_dependence (mem, x)
1638 return MEM_VOLATILE_P (x) && MEM_VOLATILE_P (mem);
1641 /* Returns MEM1 if and only if MEM1 is a scalar at a fixed address and
1642 MEM2 is a reference to a structure at a varying address, or returns
1643 MEM2 if vice versa. Otherwise, returns NULL_RTX. If a non-NULL
1644 value is returned MEM1 and MEM2 can never alias. VARIES_P is used
1645 to decide whether or not an address may vary; it should return
1646 nonzero whenever variation is possible.
1647 MEM1_ADDR and MEM2_ADDR are the addresses of MEM1 and MEM2. */
1650 fixed_scalar_and_varying_struct_p (mem1, mem2, mem1_addr, mem2_addr, varies_p)
1652 rtx mem1_addr, mem2_addr;
1653 int (*varies_p) PARAMS ((rtx, int));
1655 if (! flag_strict_aliasing)
1658 if (MEM_SCALAR_P (mem1) && MEM_IN_STRUCT_P (mem2)
1659 && !varies_p (mem1_addr, 1) && varies_p (mem2_addr, 1))
1660 /* MEM1 is a scalar at a fixed address; MEM2 is a struct at a
1664 if (MEM_IN_STRUCT_P (mem1) && MEM_SCALAR_P (mem2)
1665 && varies_p (mem1_addr, 1) && !varies_p (mem2_addr, 1))
1666 /* MEM2 is a scalar at a fixed address; MEM1 is a struct at a
1673 /* Returns nonzero if something about the mode or address format MEM1
1674 indicates that it might well alias *anything*. */
1677 aliases_everything_p (mem)
1680 if (GET_CODE (XEXP (mem, 0)) == AND)
1681 /* If the address is an AND, its very hard to know at what it is
1682 actually pointing. */
1688 /* True dependence: X is read after store in MEM takes place. */
1691 true_dependence (mem, mem_mode, x, varies)
1693 enum machine_mode mem_mode;
1695 int (*varies) PARAMS ((rtx, int));
1697 register rtx x_addr, mem_addr;
1700 if (MEM_VOLATILE_P (x) && MEM_VOLATILE_P (mem))
1703 if (DIFFERENT_ALIAS_SETS_P (x, mem))
1706 /* Unchanging memory can't conflict with non-unchanging memory.
1707 A non-unchanging read can conflict with a non-unchanging write.
1708 An unchanging read can conflict with an unchanging write since
1709 there may be a single store to this address to initialize it.
1710 Note that an unchanging store can conflict with a non-unchanging read
1711 since we have to make conservative assumptions when we have a
1712 record with readonly fields and we are copying the whole thing.
1713 Just fall through to the code below to resolve potential conflicts.
1714 This won't handle all cases optimally, but the possible performance
1715 loss should be negligible. */
1716 if (RTX_UNCHANGING_P (x) && ! RTX_UNCHANGING_P (mem))
1719 if (mem_mode == VOIDmode)
1720 mem_mode = GET_MODE (mem);
1722 x_addr = get_addr (XEXP (x, 0));
1723 mem_addr = get_addr (XEXP (mem, 0));
1725 base = find_base_term (x_addr);
1726 if (base && (GET_CODE (base) == LABEL_REF
1727 || (GET_CODE (base) == SYMBOL_REF
1728 && CONSTANT_POOL_ADDRESS_P (base))))
1731 if (! base_alias_check (x_addr, mem_addr, GET_MODE (x), mem_mode))
1734 x_addr = canon_rtx (x_addr);
1735 mem_addr = canon_rtx (mem_addr);
1737 if (! memrefs_conflict_p (GET_MODE_SIZE (mem_mode), mem_addr,
1738 SIZE_FOR_MODE (x), x_addr, 0))
1741 if (aliases_everything_p (x))
1744 /* We cannot use aliases_everyting_p to test MEM, since we must look
1745 at MEM_MODE, rather than GET_MODE (MEM). */
1746 if (mem_mode == QImode || GET_CODE (mem_addr) == AND)
1749 /* In true_dependence we also allow BLKmode to alias anything. Why
1750 don't we do this in anti_dependence and output_dependence? */
1751 if (mem_mode == BLKmode || GET_MODE (x) == BLKmode)
1754 return ! fixed_scalar_and_varying_struct_p (mem, x, mem_addr, x_addr,
1758 /* Canonical true dependence: X is read after store in MEM takes place.
1759 Variant of true_dependece which assumes MEM has already been
1760 canonicalized (hence we no longer do that here).
1761 The mem_addr argument has been added, since true_dependence computed
1762 this value prior to canonicalizing. */
1765 canon_true_dependence (mem, mem_mode, mem_addr, x, varies)
1766 rtx mem, mem_addr, x;
1767 enum machine_mode mem_mode;
1768 int (*varies) PARAMS ((rtx, int));
1770 register rtx x_addr;
1772 if (MEM_VOLATILE_P (x) && MEM_VOLATILE_P (mem))
1775 if (DIFFERENT_ALIAS_SETS_P (x, mem))
1778 /* If X is an unchanging read, then it can't possibly conflict with any
1779 non-unchanging store. It may conflict with an unchanging write though,
1780 because there may be a single store to this address to initialize it.
1781 Just fall through to the code below to resolve the case where we have
1782 both an unchanging read and an unchanging write. This won't handle all
1783 cases optimally, but the possible performance loss should be
1785 if (RTX_UNCHANGING_P (x) && ! RTX_UNCHANGING_P (mem))
1788 x_addr = get_addr (XEXP (x, 0));
1790 if (! base_alias_check (x_addr, mem_addr, GET_MODE (x), mem_mode))
1793 x_addr = canon_rtx (x_addr);
1794 if (! memrefs_conflict_p (GET_MODE_SIZE (mem_mode), mem_addr,
1795 SIZE_FOR_MODE (x), x_addr, 0))
1798 if (aliases_everything_p (x))
1801 /* We cannot use aliases_everyting_p to test MEM, since we must look
1802 at MEM_MODE, rather than GET_MODE (MEM). */
1803 if (mem_mode == QImode || GET_CODE (mem_addr) == AND)
1806 /* In true_dependence we also allow BLKmode to alias anything. Why
1807 don't we do this in anti_dependence and output_dependence? */
1808 if (mem_mode == BLKmode || GET_MODE (x) == BLKmode)
1811 return ! fixed_scalar_and_varying_struct_p (mem, x, mem_addr, x_addr,
1815 /* Returns non-zero if a write to X might alias a previous read from
1816 (or, if WRITEP is non-zero, a write to) MEM. */
1819 write_dependence_p (mem, x, writep)
1824 rtx x_addr, mem_addr;
1828 if (MEM_VOLATILE_P (x) && MEM_VOLATILE_P (mem))
1831 if (DIFFERENT_ALIAS_SETS_P (x, mem))
1834 /* Unchanging memory can't conflict with non-unchanging memory. */
1835 if (RTX_UNCHANGING_P (x) != RTX_UNCHANGING_P (mem))
1838 /* If MEM is an unchanging read, then it can't possibly conflict with
1839 the store to X, because there is at most one store to MEM, and it must
1840 have occurred somewhere before MEM. */
1841 if (! writep && RTX_UNCHANGING_P (mem))
1844 x_addr = get_addr (XEXP (x, 0));
1845 mem_addr = get_addr (XEXP (mem, 0));
1849 base = find_base_term (mem_addr);
1850 if (base && (GET_CODE (base) == LABEL_REF
1851 || (GET_CODE (base) == SYMBOL_REF
1852 && CONSTANT_POOL_ADDRESS_P (base))))
1856 if (! base_alias_check (x_addr, mem_addr, GET_MODE (x),
1860 x_addr = canon_rtx (x_addr);
1861 mem_addr = canon_rtx (mem_addr);
1863 if (!memrefs_conflict_p (SIZE_FOR_MODE (mem), mem_addr,
1864 SIZE_FOR_MODE (x), x_addr, 0))
1868 = fixed_scalar_and_varying_struct_p (mem, x, mem_addr, x_addr,
1871 return (!(fixed_scalar == mem && !aliases_everything_p (x))
1872 && !(fixed_scalar == x && !aliases_everything_p (mem)));
1875 /* Anti dependence: X is written after read in MEM takes place. */
1878 anti_dependence (mem, x)
1882 return write_dependence_p (mem, x, /*writep=*/0);
1885 /* Output dependence: X is written after store in MEM takes place. */
1888 output_dependence (mem, x)
1892 return write_dependence_p (mem, x, /*writep=*/1);
1895 /* Returns non-zero if X mentions something which is not
1896 local to the function and is not constant. */
1899 nonlocal_mentioned_p (x)
1903 register RTX_CODE code;
1906 code = GET_CODE (x);
1908 if (GET_RTX_CLASS (code) == 'i')
1910 /* Constant functions can be constant if they don't use
1911 scratch memory used to mark function w/o side effects. */
1912 if (code == CALL_INSN && CONST_CALL_P (x))
1914 x = CALL_INSN_FUNCTION_USAGE (x);
1920 code = GET_CODE (x);
1926 if (GET_CODE (SUBREG_REG (x)) == REG)
1928 /* Global registers are not local. */
1929 if (REGNO (SUBREG_REG (x)) < FIRST_PSEUDO_REGISTER
1930 && global_regs[subreg_regno (x)])
1938 /* Global registers are not local. */
1939 if (regno < FIRST_PSEUDO_REGISTER && global_regs[regno])
1953 /* Constants in the function's constants pool are constant. */
1954 if (CONSTANT_POOL_ADDRESS_P (x))
1959 /* Non-constant calls and recursion are not local. */
1963 /* Be overly conservative and consider any volatile memory
1964 reference as not local. */
1965 if (MEM_VOLATILE_P (x))
1967 base = find_base_term (XEXP (x, 0));
1970 /* A Pmode ADDRESS could be a reference via the structure value
1971 address or static chain. Such memory references are nonlocal.
1973 Thus, we have to examine the contents of the ADDRESS to find
1974 out if this is a local reference or not. */
1975 if (GET_CODE (base) == ADDRESS
1976 && GET_MODE (base) == Pmode
1977 && (XEXP (base, 0) == stack_pointer_rtx
1978 || XEXP (base, 0) == arg_pointer_rtx
1979 #if HARD_FRAME_POINTER_REGNUM != FRAME_POINTER_REGNUM
1980 || XEXP (base, 0) == hard_frame_pointer_rtx
1982 || XEXP (base, 0) == frame_pointer_rtx))
1984 /* Constants in the function's constant pool are constant. */
1985 if (GET_CODE (base) == SYMBOL_REF && CONSTANT_POOL_ADDRESS_P (base))
1990 case UNSPEC_VOLATILE:
1995 if (MEM_VOLATILE_P (x))
2004 /* Recursively scan the operands of this expression. */
2007 register const char *fmt = GET_RTX_FORMAT (code);
2010 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2012 if (fmt[i] == 'e' && XEXP (x, i))
2014 if (nonlocal_mentioned_p (XEXP (x, i)))
2017 else if (fmt[i] == 'E')
2020 for (j = 0; j < XVECLEN (x, i); j++)
2021 if (nonlocal_mentioned_p (XVECEXP (x, i, j)))
2030 /* Return non-zero if a loop (natural or otherwise) is present.
2031 Inspired by Depth_First_Search_PP described in:
2033 Advanced Compiler Design and Implementation
2035 Morgan Kaufmann, 1997
2037 and heavily borrowed from flow_depth_first_order_compute. */
2050 /* Allocate the preorder and postorder number arrays. */
2051 pre = (int *) xcalloc (n_basic_blocks, sizeof (int));
2052 post = (int *) xcalloc (n_basic_blocks, sizeof (int));
2054 /* Allocate stack for back-tracking up CFG. */
2055 stack = (edge *) xmalloc ((n_basic_blocks + 1) * sizeof (edge));
2058 /* Allocate bitmap to track nodes that have been visited. */
2059 visited = sbitmap_alloc (n_basic_blocks);
2061 /* None of the nodes in the CFG have been visited yet. */
2062 sbitmap_zero (visited);
2064 /* Push the first edge on to the stack. */
2065 stack[sp++] = ENTRY_BLOCK_PTR->succ;
2073 /* Look at the edge on the top of the stack. */
2078 /* Check if the edge destination has been visited yet. */
2079 if (dest != EXIT_BLOCK_PTR && ! TEST_BIT (visited, dest->index))
2081 /* Mark that we have visited the destination. */
2082 SET_BIT (visited, dest->index);
2084 pre[dest->index] = prenum++;
2088 /* Since the DEST node has been visited for the first
2089 time, check its successors. */
2090 stack[sp++] = dest->succ;
2093 post[dest->index] = postnum++;
2097 if (dest != EXIT_BLOCK_PTR
2098 && pre[src->index] >= pre[dest->index]
2099 && post[dest->index] == 0)
2102 if (! e->succ_next && src != ENTRY_BLOCK_PTR)
2103 post[src->index] = postnum++;
2106 stack[sp - 1] = e->succ_next;
2115 sbitmap_free (visited);
2120 /* Mark the function if it is constant. */
2123 mark_constant_function ()
2126 int nonlocal_mentioned;
2128 if (TREE_PUBLIC (current_function_decl)
2129 || TREE_READONLY (current_function_decl)
2130 || DECL_IS_PURE (current_function_decl)
2131 || TREE_THIS_VOLATILE (current_function_decl)
2132 || TYPE_MODE (TREE_TYPE (current_function_decl)) == VOIDmode)
2135 /* A loop might not return which counts as a side effect. */
2139 nonlocal_mentioned = 0;
2141 init_alias_analysis ();
2143 /* Determine if this is a constant function. */
2145 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
2146 if (INSN_P (insn) && nonlocal_mentioned_p (insn))
2148 nonlocal_mentioned = 1;
2152 end_alias_analysis ();
2154 /* Mark the function. */
2156 if (! nonlocal_mentioned)
2157 TREE_READONLY (current_function_decl) = 1;
2161 static HARD_REG_SET argument_registers;
2168 #ifndef OUTGOING_REGNO
2169 #define OUTGOING_REGNO(N) N
2171 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
2172 /* Check whether this register can hold an incoming pointer
2173 argument. FUNCTION_ARG_REGNO_P tests outgoing register
2174 numbers, so translate if necessary due to register windows. */
2175 if (FUNCTION_ARG_REGNO_P (OUTGOING_REGNO (i))
2176 && HARD_REGNO_MODE_OK (i, Pmode))
2177 SET_HARD_REG_BIT (argument_registers, i);
2179 alias_sets = splay_tree_new (splay_tree_compare_ints, 0, 0);
2182 /* Initialize the aliasing machinery. Initialize the REG_KNOWN_VALUE
2186 init_alias_analysis ()
2188 int maxreg = max_reg_num ();
2191 register unsigned int ui;
2194 reg_known_value_size = maxreg;
2197 = (rtx *) xcalloc ((maxreg - FIRST_PSEUDO_REGISTER), sizeof (rtx))
2198 - FIRST_PSEUDO_REGISTER;
2200 = (char*) xcalloc ((maxreg - FIRST_PSEUDO_REGISTER), sizeof (char))
2201 - FIRST_PSEUDO_REGISTER;
2203 /* Overallocate reg_base_value to allow some growth during loop
2204 optimization. Loop unrolling can create a large number of
2206 reg_base_value_size = maxreg * 2;
2207 reg_base_value = (rtx *) xcalloc (reg_base_value_size, sizeof (rtx));
2208 ggc_add_rtx_root (reg_base_value, reg_base_value_size);
2210 new_reg_base_value = (rtx *) xmalloc (reg_base_value_size * sizeof (rtx));
2211 reg_seen = (char *) xmalloc (reg_base_value_size);
2212 if (! reload_completed && flag_unroll_loops)
2214 /* ??? Why are we realloc'ing if we're just going to zero it? */
2215 alias_invariant = (rtx *)xrealloc (alias_invariant,
2216 reg_base_value_size * sizeof (rtx));
2217 memset ((char *)alias_invariant, 0, reg_base_value_size * sizeof (rtx));
2220 /* The basic idea is that each pass through this loop will use the
2221 "constant" information from the previous pass to propagate alias
2222 information through another level of assignments.
2224 This could get expensive if the assignment chains are long. Maybe
2225 we should throttle the number of iterations, possibly based on
2226 the optimization level or flag_expensive_optimizations.
2228 We could propagate more information in the first pass by making use
2229 of REG_N_SETS to determine immediately that the alias information
2230 for a pseudo is "constant".
2232 A program with an uninitialized variable can cause an infinite loop
2233 here. Instead of doing a full dataflow analysis to detect such problems
2234 we just cap the number of iterations for the loop.
2236 The state of the arrays for the set chain in question does not matter
2237 since the program has undefined behavior. */
2242 /* Assume nothing will change this iteration of the loop. */
2245 /* We want to assign the same IDs each iteration of this loop, so
2246 start counting from zero each iteration of the loop. */
2249 /* We're at the start of the funtion each iteration through the
2250 loop, so we're copying arguments. */
2251 copying_arguments = 1;
2253 /* Wipe the potential alias information clean for this pass. */
2254 memset ((char *) new_reg_base_value, 0, reg_base_value_size * sizeof (rtx));
2256 /* Wipe the reg_seen array clean. */
2257 memset ((char *) reg_seen, 0, reg_base_value_size);
2259 /* Mark all hard registers which may contain an address.
2260 The stack, frame and argument pointers may contain an address.
2261 An argument register which can hold a Pmode value may contain
2262 an address even if it is not in BASE_REGS.
2264 The address expression is VOIDmode for an argument and
2265 Pmode for other registers. */
2267 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
2268 if (TEST_HARD_REG_BIT (argument_registers, i))
2269 new_reg_base_value[i] = gen_rtx_ADDRESS (VOIDmode,
2270 gen_rtx_REG (Pmode, i));
2272 new_reg_base_value[STACK_POINTER_REGNUM]
2273 = gen_rtx_ADDRESS (Pmode, stack_pointer_rtx);
2274 new_reg_base_value[ARG_POINTER_REGNUM]
2275 = gen_rtx_ADDRESS (Pmode, arg_pointer_rtx);
2276 new_reg_base_value[FRAME_POINTER_REGNUM]
2277 = gen_rtx_ADDRESS (Pmode, frame_pointer_rtx);
2278 #if HARD_FRAME_POINTER_REGNUM != FRAME_POINTER_REGNUM
2279 new_reg_base_value[HARD_FRAME_POINTER_REGNUM]
2280 = gen_rtx_ADDRESS (Pmode, hard_frame_pointer_rtx);
2283 /* Walk the insns adding values to the new_reg_base_value array. */
2284 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
2290 #if defined (HAVE_prologue) || defined (HAVE_epilogue)
2291 /* The prologue/epilouge insns are not threaded onto the
2292 insn chain until after reload has completed. Thus,
2293 there is no sense wasting time checking if INSN is in
2294 the prologue/epilogue until after reload has completed. */
2295 if (reload_completed
2296 && prologue_epilogue_contains (insn))
2300 /* If this insn has a noalias note, process it, Otherwise,
2301 scan for sets. A simple set will have no side effects
2302 which could change the base value of any other register. */
2304 if (GET_CODE (PATTERN (insn)) == SET
2305 && REG_NOTES (insn) != 0
2306 && find_reg_note (insn, REG_NOALIAS, NULL_RTX))
2307 record_set (SET_DEST (PATTERN (insn)), NULL_RTX, NULL);
2309 note_stores (PATTERN (insn), record_set, NULL);
2311 set = single_set (insn);
2314 && GET_CODE (SET_DEST (set)) == REG
2315 && REGNO (SET_DEST (set)) >= FIRST_PSEUDO_REGISTER)
2317 unsigned int regno = REGNO (SET_DEST (set));
2318 rtx src = SET_SRC (set);
2320 if (REG_NOTES (insn) != 0
2321 && (((note = find_reg_note (insn, REG_EQUAL, 0)) != 0
2322 && REG_N_SETS (regno) == 1)
2323 || (note = find_reg_note (insn, REG_EQUIV, NULL_RTX)) != 0)
2324 && GET_CODE (XEXP (note, 0)) != EXPR_LIST
2325 && ! rtx_varies_p (XEXP (note, 0), 1)
2326 && ! reg_overlap_mentioned_p (SET_DEST (set), XEXP (note, 0)))
2328 reg_known_value[regno] = XEXP (note, 0);
2329 reg_known_equiv_p[regno] = REG_NOTE_KIND (note) == REG_EQUIV;
2331 else if (REG_N_SETS (regno) == 1
2332 && GET_CODE (src) == PLUS
2333 && GET_CODE (XEXP (src, 0)) == REG
2334 && REGNO (XEXP (src, 0)) >= FIRST_PSEUDO_REGISTER
2335 && (reg_known_value[REGNO (XEXP (src, 0))])
2336 && GET_CODE (XEXP (src, 1)) == CONST_INT)
2338 rtx op0 = XEXP (src, 0);
2339 op0 = reg_known_value[REGNO (op0)];
2340 reg_known_value[regno]
2341 = plus_constant (op0, INTVAL (XEXP (src, 1)));
2342 reg_known_equiv_p[regno] = 0;
2344 else if (REG_N_SETS (regno) == 1
2345 && ! rtx_varies_p (src, 1))
2347 reg_known_value[regno] = src;
2348 reg_known_equiv_p[regno] = 0;
2352 else if (GET_CODE (insn) == NOTE
2353 && NOTE_LINE_NUMBER (insn) == NOTE_INSN_FUNCTION_BEG)
2354 copying_arguments = 0;
2357 /* Now propagate values from new_reg_base_value to reg_base_value. */
2358 for (ui = 0; ui < reg_base_value_size; ui++)
2360 if (new_reg_base_value[ui]
2361 && new_reg_base_value[ui] != reg_base_value[ui]
2362 && ! rtx_equal_p (new_reg_base_value[ui], reg_base_value[ui]))
2364 reg_base_value[ui] = new_reg_base_value[ui];
2369 while (changed && ++pass < MAX_ALIAS_LOOP_PASSES);
2371 /* Fill in the remaining entries. */
2372 for (i = FIRST_PSEUDO_REGISTER; i < maxreg; i++)
2373 if (reg_known_value[i] == 0)
2374 reg_known_value[i] = regno_reg_rtx[i];
2376 /* Simplify the reg_base_value array so that no register refers to
2377 another register, except to special registers indirectly through
2378 ADDRESS expressions.
2380 In theory this loop can take as long as O(registers^2), but unless
2381 there are very long dependency chains it will run in close to linear
2384 This loop may not be needed any longer now that the main loop does
2385 a better job at propagating alias information. */
2391 for (ui = 0; ui < reg_base_value_size; ui++)
2393 rtx base = reg_base_value[ui];
2394 if (base && GET_CODE (base) == REG)
2396 unsigned int base_regno = REGNO (base);
2397 if (base_regno == ui) /* register set from itself */
2398 reg_base_value[ui] = 0;
2400 reg_base_value[ui] = reg_base_value[base_regno];
2405 while (changed && pass < MAX_ALIAS_LOOP_PASSES);
2408 free (new_reg_base_value);
2409 new_reg_base_value = 0;
2415 end_alias_analysis ()
2417 free (reg_known_value + FIRST_PSEUDO_REGISTER);
2418 reg_known_value = 0;
2419 reg_known_value_size = 0;
2420 free (reg_known_equiv_p + FIRST_PSEUDO_REGISTER);
2421 reg_known_equiv_p = 0;
2424 ggc_del_root (reg_base_value);
2425 free (reg_base_value);
2428 reg_base_value_size = 0;
2429 if (alias_invariant)
2431 free (alias_invariant);
2432 alias_invariant = 0;