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)
976 rtx addr = canon_rtx (XEXP (x, 0));
978 if (addr != XEXP (x, 0))
980 rtx new = gen_rtx_MEM (GET_MODE (x), addr);
982 MEM_COPY_ATTRIBUTES (new, x);
989 /* Return 1 if X and Y are identical-looking rtx's.
991 We use the data in reg_known_value above to see if two registers with
992 different numbers are, in fact, equivalent. */
995 rtx_equal_for_memref_p (x, y)
1000 register enum rtx_code code;
1001 register const char *fmt;
1003 if (x == 0 && y == 0)
1005 if (x == 0 || y == 0)
1014 code = GET_CODE (x);
1015 /* Rtx's of different codes cannot be equal. */
1016 if (code != GET_CODE (y))
1019 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent.
1020 (REG:SI x) and (REG:HI x) are NOT equivalent. */
1022 if (GET_MODE (x) != GET_MODE (y))
1025 /* Some RTL can be compared without a recursive examination. */
1029 return REGNO (x) == REGNO (y);
1032 return XEXP (x, 0) == XEXP (y, 0);
1035 return XSTR (x, 0) == XSTR (y, 0);
1039 /* There's no need to compare the contents of CONST_DOUBLEs or
1040 CONST_INTs because pointer equality is a good enough
1041 comparison for these nodes. */
1045 return (XINT (x, 1) == XINT (y, 1)
1046 && rtx_equal_for_memref_p (XEXP (x, 0), XEXP (y, 0)));
1052 /* For commutative operations, the RTX match if the operand match in any
1053 order. Also handle the simple binary and unary cases without a loop. */
1054 if (code == EQ || code == NE || GET_RTX_CLASS (code) == 'c')
1055 return ((rtx_equal_for_memref_p (XEXP (x, 0), XEXP (y, 0))
1056 && rtx_equal_for_memref_p (XEXP (x, 1), XEXP (y, 1)))
1057 || (rtx_equal_for_memref_p (XEXP (x, 0), XEXP (y, 1))
1058 && rtx_equal_for_memref_p (XEXP (x, 1), XEXP (y, 0))));
1059 else if (GET_RTX_CLASS (code) == '<' || GET_RTX_CLASS (code) == '2')
1060 return (rtx_equal_for_memref_p (XEXP (x, 0), XEXP (y, 0))
1061 && rtx_equal_for_memref_p (XEXP (x, 1), XEXP (y, 1)));
1062 else if (GET_RTX_CLASS (code) == '1')
1063 return rtx_equal_for_memref_p (XEXP (x, 0), XEXP (y, 0));
1065 /* Compare the elements. If any pair of corresponding elements
1066 fail to match, return 0 for the whole things.
1068 Limit cases to types which actually appear in addresses. */
1070 fmt = GET_RTX_FORMAT (code);
1071 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1076 if (XINT (x, i) != XINT (y, i))
1081 /* Two vectors must have the same length. */
1082 if (XVECLEN (x, i) != XVECLEN (y, i))
1085 /* And the corresponding elements must match. */
1086 for (j = 0; j < XVECLEN (x, i); j++)
1087 if (rtx_equal_for_memref_p (XVECEXP (x, i, j),
1088 XVECEXP (y, i, j)) == 0)
1093 if (rtx_equal_for_memref_p (XEXP (x, i), XEXP (y, i)) == 0)
1097 /* This can happen for an asm which clobbers memory. */
1101 /* It is believed that rtx's at this level will never
1102 contain anything but integers and other rtx's,
1103 except for within LABEL_REFs and SYMBOL_REFs. */
1111 /* Given an rtx X, find a SYMBOL_REF or LABEL_REF within
1112 X and return it, or return 0 if none found. */
1115 find_symbolic_term (x)
1119 register enum rtx_code code;
1120 register const char *fmt;
1122 code = GET_CODE (x);
1123 if (code == SYMBOL_REF || code == LABEL_REF)
1125 if (GET_RTX_CLASS (code) == 'o')
1128 fmt = GET_RTX_FORMAT (code);
1129 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1135 t = find_symbolic_term (XEXP (x, i));
1139 else if (fmt[i] == 'E')
1150 struct elt_loc_list *l;
1152 #if defined (FIND_BASE_TERM)
1153 /* Try machine-dependent ways to find the base term. */
1154 x = FIND_BASE_TERM (x);
1157 switch (GET_CODE (x))
1160 return REG_BASE_VALUE (x);
1163 case SIGN_EXTEND: /* Used for Alpha/NT pointers */
1169 return find_base_term (XEXP (x, 0));
1172 val = CSELIB_VAL_PTR (x);
1173 for (l = val->locs; l; l = l->next)
1174 if ((x = find_base_term (l->loc)) != 0)
1180 if (GET_CODE (x) != PLUS && GET_CODE (x) != MINUS)
1187 rtx tmp1 = XEXP (x, 0);
1188 rtx tmp2 = XEXP (x, 1);
1190 /* This is a litle bit tricky since we have to determine which of
1191 the two operands represents the real base address. Otherwise this
1192 routine may return the index register instead of the base register.
1194 That may cause us to believe no aliasing was possible, when in
1195 fact aliasing is possible.
1197 We use a few simple tests to guess the base register. Additional
1198 tests can certainly be added. For example, if one of the operands
1199 is a shift or multiply, then it must be the index register and the
1200 other operand is the base register. */
1202 if (tmp1 == pic_offset_table_rtx && CONSTANT_P (tmp2))
1203 return find_base_term (tmp2);
1205 /* If either operand is known to be a pointer, then use it
1206 to determine the base term. */
1207 if (REG_P (tmp1) && REG_POINTER (tmp1))
1208 return find_base_term (tmp1);
1210 if (REG_P (tmp2) && REG_POINTER (tmp2))
1211 return find_base_term (tmp2);
1213 /* Neither operand was known to be a pointer. Go ahead and find the
1214 base term for both operands. */
1215 tmp1 = find_base_term (tmp1);
1216 tmp2 = find_base_term (tmp2);
1218 /* If either base term is named object or a special address
1219 (like an argument or stack reference), then use it for the
1222 && (GET_CODE (tmp1) == SYMBOL_REF
1223 || GET_CODE (tmp1) == LABEL_REF
1224 || (GET_CODE (tmp1) == ADDRESS
1225 && GET_MODE (tmp1) != VOIDmode)))
1229 && (GET_CODE (tmp2) == SYMBOL_REF
1230 || GET_CODE (tmp2) == LABEL_REF
1231 || (GET_CODE (tmp2) == ADDRESS
1232 && GET_MODE (tmp2) != VOIDmode)))
1235 /* We could not determine which of the two operands was the
1236 base register and which was the index. So we can determine
1237 nothing from the base alias check. */
1242 if (GET_CODE (XEXP (x, 0)) == REG && GET_CODE (XEXP (x, 1)) == CONST_INT)
1243 return REG_BASE_VALUE (XEXP (x, 0));
1251 return REG_BASE_VALUE (frame_pointer_rtx);
1258 /* Return 0 if the addresses X and Y are known to point to different
1259 objects, 1 if they might be pointers to the same object. */
1262 base_alias_check (x, y, x_mode, y_mode)
1264 enum machine_mode x_mode, y_mode;
1266 rtx x_base = find_base_term (x);
1267 rtx y_base = find_base_term (y);
1269 /* If the address itself has no known base see if a known equivalent
1270 value has one. If either address still has no known base, nothing
1271 is known about aliasing. */
1276 if (! flag_expensive_optimizations || (x_c = canon_rtx (x)) == x)
1279 x_base = find_base_term (x_c);
1287 if (! flag_expensive_optimizations || (y_c = canon_rtx (y)) == y)
1290 y_base = find_base_term (y_c);
1295 /* If the base addresses are equal nothing is known about aliasing. */
1296 if (rtx_equal_p (x_base, y_base))
1299 /* The base addresses of the read and write are different expressions.
1300 If they are both symbols and they are not accessed via AND, there is
1301 no conflict. We can bring knowledge of object alignment into play
1302 here. For example, on alpha, "char a, b;" can alias one another,
1303 though "char a; long b;" cannot. */
1304 if (GET_CODE (x_base) != ADDRESS && GET_CODE (y_base) != ADDRESS)
1306 if (GET_CODE (x) == AND && GET_CODE (y) == AND)
1308 if (GET_CODE (x) == AND
1309 && (GET_CODE (XEXP (x, 1)) != CONST_INT
1310 || GET_MODE_UNIT_SIZE (y_mode) < -INTVAL (XEXP (x, 1))))
1312 if (GET_CODE (y) == AND
1313 && (GET_CODE (XEXP (y, 1)) != CONST_INT
1314 || GET_MODE_UNIT_SIZE (x_mode) < -INTVAL (XEXP (y, 1))))
1316 /* Differing symbols never alias. */
1320 /* If one address is a stack reference there can be no alias:
1321 stack references using different base registers do not alias,
1322 a stack reference can not alias a parameter, and a stack reference
1323 can not alias a global. */
1324 if ((GET_CODE (x_base) == ADDRESS && GET_MODE (x_base) == Pmode)
1325 || (GET_CODE (y_base) == ADDRESS && GET_MODE (y_base) == Pmode))
1328 if (! flag_argument_noalias)
1331 if (flag_argument_noalias > 1)
1334 /* Weak noalias assertion (arguments are distinct, but may match globals). */
1335 return ! (GET_MODE (x_base) == VOIDmode && GET_MODE (y_base) == VOIDmode);
1338 /* Convert the address X into something we can use. This is done by returning
1339 it unchanged unless it is a value; in the latter case we call cselib to get
1340 a more useful rtx. */
1347 struct elt_loc_list *l;
1349 if (GET_CODE (x) != VALUE)
1351 v = CSELIB_VAL_PTR (x);
1352 for (l = v->locs; l; l = l->next)
1353 if (CONSTANT_P (l->loc))
1355 for (l = v->locs; l; l = l->next)
1356 if (GET_CODE (l->loc) != REG && GET_CODE (l->loc) != MEM)
1359 return v->locs->loc;
1363 /* Return the address of the (N_REFS + 1)th memory reference to ADDR
1364 where SIZE is the size in bytes of the memory reference. If ADDR
1365 is not modified by the memory reference then ADDR is returned. */
1368 addr_side_effect_eval (addr, size, n_refs)
1375 switch (GET_CODE (addr))
1378 offset = (n_refs + 1) * size;
1381 offset = -(n_refs + 1) * size;
1384 offset = n_refs * size;
1387 offset = -n_refs * size;
1395 addr = gen_rtx_PLUS (GET_MODE (addr), XEXP (addr, 0), GEN_INT (offset));
1397 addr = XEXP (addr, 0);
1402 /* Return nonzero if X and Y (memory addresses) could reference the
1403 same location in memory. C is an offset accumulator. When
1404 C is nonzero, we are testing aliases between X and Y + C.
1405 XSIZE is the size in bytes of the X reference,
1406 similarly YSIZE is the size in bytes for Y.
1408 If XSIZE or YSIZE is zero, we do not know the amount of memory being
1409 referenced (the reference was BLKmode), so make the most pessimistic
1412 If XSIZE or YSIZE is negative, we may access memory outside the object
1413 being referenced as a side effect. This can happen when using AND to
1414 align memory references, as is done on the Alpha.
1416 Nice to notice that varying addresses cannot conflict with fp if no
1417 local variables had their addresses taken, but that's too hard now. */
1420 memrefs_conflict_p (xsize, x, ysize, y, c)
1425 if (GET_CODE (x) == VALUE)
1427 if (GET_CODE (y) == VALUE)
1429 if (GET_CODE (x) == HIGH)
1431 else if (GET_CODE (x) == LO_SUM)
1434 x = canon_rtx (addr_side_effect_eval (x, xsize, 0));
1435 if (GET_CODE (y) == HIGH)
1437 else if (GET_CODE (y) == LO_SUM)
1440 y = canon_rtx (addr_side_effect_eval (y, ysize, 0));
1442 if (rtx_equal_for_memref_p (x, y))
1444 if (xsize <= 0 || ysize <= 0)
1446 if (c >= 0 && xsize > c)
1448 if (c < 0 && ysize+c > 0)
1453 /* This code used to check for conflicts involving stack references and
1454 globals but the base address alias code now handles these cases. */
1456 if (GET_CODE (x) == PLUS)
1458 /* The fact that X is canonicalized means that this
1459 PLUS rtx is canonicalized. */
1460 rtx x0 = XEXP (x, 0);
1461 rtx x1 = XEXP (x, 1);
1463 if (GET_CODE (y) == PLUS)
1465 /* The fact that Y is canonicalized means that this
1466 PLUS rtx is canonicalized. */
1467 rtx y0 = XEXP (y, 0);
1468 rtx y1 = XEXP (y, 1);
1470 if (rtx_equal_for_memref_p (x1, y1))
1471 return memrefs_conflict_p (xsize, x0, ysize, y0, c);
1472 if (rtx_equal_for_memref_p (x0, y0))
1473 return memrefs_conflict_p (xsize, x1, ysize, y1, c);
1474 if (GET_CODE (x1) == CONST_INT)
1476 if (GET_CODE (y1) == CONST_INT)
1477 return memrefs_conflict_p (xsize, x0, ysize, y0,
1478 c - INTVAL (x1) + INTVAL (y1));
1480 return memrefs_conflict_p (xsize, x0, ysize, y,
1483 else if (GET_CODE (y1) == CONST_INT)
1484 return memrefs_conflict_p (xsize, x, ysize, y0, c + INTVAL (y1));
1488 else if (GET_CODE (x1) == CONST_INT)
1489 return memrefs_conflict_p (xsize, x0, ysize, y, c - INTVAL (x1));
1491 else if (GET_CODE (y) == PLUS)
1493 /* The fact that Y is canonicalized means that this
1494 PLUS rtx is canonicalized. */
1495 rtx y0 = XEXP (y, 0);
1496 rtx y1 = XEXP (y, 1);
1498 if (GET_CODE (y1) == CONST_INT)
1499 return memrefs_conflict_p (xsize, x, ysize, y0, c + INTVAL (y1));
1504 if (GET_CODE (x) == GET_CODE (y))
1505 switch (GET_CODE (x))
1509 /* Handle cases where we expect the second operands to be the
1510 same, and check only whether the first operand would conflict
1513 rtx x1 = canon_rtx (XEXP (x, 1));
1514 rtx y1 = canon_rtx (XEXP (y, 1));
1515 if (! rtx_equal_for_memref_p (x1, y1))
1517 x0 = canon_rtx (XEXP (x, 0));
1518 y0 = canon_rtx (XEXP (y, 0));
1519 if (rtx_equal_for_memref_p (x0, y0))
1520 return (xsize == 0 || ysize == 0
1521 || (c >= 0 && xsize > c) || (c < 0 && ysize+c > 0));
1523 /* Can't properly adjust our sizes. */
1524 if (GET_CODE (x1) != CONST_INT)
1526 xsize /= INTVAL (x1);
1527 ysize /= INTVAL (x1);
1529 return memrefs_conflict_p (xsize, x0, ysize, y0, c);
1533 /* Are these registers known not to be equal? */
1534 if (alias_invariant)
1536 unsigned int r_x = REGNO (x), r_y = REGNO (y);
1537 rtx i_x, i_y; /* invariant relationships of X and Y */
1539 i_x = r_x >= reg_base_value_size ? 0 : alias_invariant[r_x];
1540 i_y = r_y >= reg_base_value_size ? 0 : alias_invariant[r_y];
1542 if (i_x == 0 && i_y == 0)
1545 if (! memrefs_conflict_p (xsize, i_x ? i_x : x,
1546 ysize, i_y ? i_y : y, c))
1555 /* Treat an access through an AND (e.g. a subword access on an Alpha)
1556 as an access with indeterminate size. Assume that references
1557 besides AND are aligned, so if the size of the other reference is
1558 at least as large as the alignment, assume no other overlap. */
1559 if (GET_CODE (x) == AND && GET_CODE (XEXP (x, 1)) == CONST_INT)
1561 if (GET_CODE (y) == AND || ysize < -INTVAL (XEXP (x, 1)))
1563 return memrefs_conflict_p (xsize, XEXP (x, 0), ysize, y, c);
1565 if (GET_CODE (y) == AND && GET_CODE (XEXP (y, 1)) == CONST_INT)
1567 /* ??? If we are indexing far enough into the array/structure, we
1568 may yet be able to determine that we can not overlap. But we
1569 also need to that we are far enough from the end not to overlap
1570 a following reference, so we do nothing with that for now. */
1571 if (GET_CODE (x) == AND || xsize < -INTVAL (XEXP (y, 1)))
1573 return memrefs_conflict_p (xsize, x, ysize, XEXP (y, 0), c);
1576 if (GET_CODE (x) == ADDRESSOF)
1578 if (y == frame_pointer_rtx
1579 || GET_CODE (y) == ADDRESSOF)
1580 return xsize <= 0 || ysize <= 0;
1582 if (GET_CODE (y) == ADDRESSOF)
1584 if (x == frame_pointer_rtx)
1585 return xsize <= 0 || ysize <= 0;
1590 if (GET_CODE (x) == CONST_INT && GET_CODE (y) == CONST_INT)
1592 c += (INTVAL (y) - INTVAL (x));
1593 return (xsize <= 0 || ysize <= 0
1594 || (c >= 0 && xsize > c) || (c < 0 && ysize+c > 0));
1597 if (GET_CODE (x) == CONST)
1599 if (GET_CODE (y) == CONST)
1600 return memrefs_conflict_p (xsize, canon_rtx (XEXP (x, 0)),
1601 ysize, canon_rtx (XEXP (y, 0)), c);
1603 return memrefs_conflict_p (xsize, canon_rtx (XEXP (x, 0)),
1606 if (GET_CODE (y) == CONST)
1607 return memrefs_conflict_p (xsize, x, ysize,
1608 canon_rtx (XEXP (y, 0)), c);
1611 return (xsize <= 0 || ysize <= 0
1612 || (rtx_equal_for_memref_p (x, y)
1613 && ((c >= 0 && xsize > c) || (c < 0 && ysize+c > 0))));
1620 /* Functions to compute memory dependencies.
1622 Since we process the insns in execution order, we can build tables
1623 to keep track of what registers are fixed (and not aliased), what registers
1624 are varying in known ways, and what registers are varying in unknown
1627 If both memory references are volatile, then there must always be a
1628 dependence between the two references, since their order can not be
1629 changed. A volatile and non-volatile reference can be interchanged
1632 A MEM_IN_STRUCT reference at a non-AND varying address can never
1633 conflict with a non-MEM_IN_STRUCT reference at a fixed address. We
1634 also must allow AND addresses, because they may generate accesses
1635 outside the object being referenced. This is used to generate
1636 aligned addresses from unaligned addresses, for instance, the alpha
1637 storeqi_unaligned pattern. */
1639 /* Read dependence: X is read after read in MEM takes place. There can
1640 only be a dependence here if both reads are volatile. */
1643 read_dependence (mem, x)
1647 return MEM_VOLATILE_P (x) && MEM_VOLATILE_P (mem);
1650 /* Returns MEM1 if and only if MEM1 is a scalar at a fixed address and
1651 MEM2 is a reference to a structure at a varying address, or returns
1652 MEM2 if vice versa. Otherwise, returns NULL_RTX. If a non-NULL
1653 value is returned MEM1 and MEM2 can never alias. VARIES_P is used
1654 to decide whether or not an address may vary; it should return
1655 nonzero whenever variation is possible.
1656 MEM1_ADDR and MEM2_ADDR are the addresses of MEM1 and MEM2. */
1659 fixed_scalar_and_varying_struct_p (mem1, mem2, mem1_addr, mem2_addr, varies_p)
1661 rtx mem1_addr, mem2_addr;
1662 int (*varies_p) PARAMS ((rtx, int));
1664 if (! flag_strict_aliasing)
1667 if (MEM_SCALAR_P (mem1) && MEM_IN_STRUCT_P (mem2)
1668 && !varies_p (mem1_addr, 1) && varies_p (mem2_addr, 1))
1669 /* MEM1 is a scalar at a fixed address; MEM2 is a struct at a
1673 if (MEM_IN_STRUCT_P (mem1) && MEM_SCALAR_P (mem2)
1674 && varies_p (mem1_addr, 1) && !varies_p (mem2_addr, 1))
1675 /* MEM2 is a scalar at a fixed address; MEM1 is a struct at a
1682 /* Returns nonzero if something about the mode or address format MEM1
1683 indicates that it might well alias *anything*. */
1686 aliases_everything_p (mem)
1689 if (GET_CODE (XEXP (mem, 0)) == AND)
1690 /* If the address is an AND, its very hard to know at what it is
1691 actually pointing. */
1697 /* True dependence: X is read after store in MEM takes place. */
1700 true_dependence (mem, mem_mode, x, varies)
1702 enum machine_mode mem_mode;
1704 int (*varies) PARAMS ((rtx, int));
1706 register rtx x_addr, mem_addr;
1709 if (MEM_VOLATILE_P (x) && MEM_VOLATILE_P (mem))
1712 if (DIFFERENT_ALIAS_SETS_P (x, mem))
1715 /* Unchanging memory can't conflict with non-unchanging memory.
1716 A non-unchanging read can conflict with a non-unchanging write.
1717 An unchanging read can conflict with an unchanging write since
1718 there may be a single store to this address to initialize it.
1719 Note that an unchanging store can conflict with a non-unchanging read
1720 since we have to make conservative assumptions when we have a
1721 record with readonly fields and we are copying the whole thing.
1722 Just fall through to the code below to resolve potential conflicts.
1723 This won't handle all cases optimally, but the possible performance
1724 loss should be negligible. */
1725 if (RTX_UNCHANGING_P (x) && ! RTX_UNCHANGING_P (mem))
1728 if (mem_mode == VOIDmode)
1729 mem_mode = GET_MODE (mem);
1731 x_addr = get_addr (XEXP (x, 0));
1732 mem_addr = get_addr (XEXP (mem, 0));
1734 base = find_base_term (x_addr);
1735 if (base && (GET_CODE (base) == LABEL_REF
1736 || (GET_CODE (base) == SYMBOL_REF
1737 && CONSTANT_POOL_ADDRESS_P (base))))
1740 if (! base_alias_check (x_addr, mem_addr, GET_MODE (x), mem_mode))
1743 x_addr = canon_rtx (x_addr);
1744 mem_addr = canon_rtx (mem_addr);
1746 if (! memrefs_conflict_p (GET_MODE_SIZE (mem_mode), mem_addr,
1747 SIZE_FOR_MODE (x), x_addr, 0))
1750 if (aliases_everything_p (x))
1753 /* We cannot use aliases_everyting_p to test MEM, since we must look
1754 at MEM_MODE, rather than GET_MODE (MEM). */
1755 if (mem_mode == QImode || GET_CODE (mem_addr) == AND)
1758 /* In true_dependence we also allow BLKmode to alias anything. Why
1759 don't we do this in anti_dependence and output_dependence? */
1760 if (mem_mode == BLKmode || GET_MODE (x) == BLKmode)
1763 return ! fixed_scalar_and_varying_struct_p (mem, x, mem_addr, x_addr,
1767 /* Canonical true dependence: X is read after store in MEM takes place.
1768 Variant of true_dependece which assumes MEM has already been
1769 canonicalized (hence we no longer do that here).
1770 The mem_addr argument has been added, since true_dependence computed
1771 this value prior to canonicalizing. */
1774 canon_true_dependence (mem, mem_mode, mem_addr, x, varies)
1775 rtx mem, mem_addr, x;
1776 enum machine_mode mem_mode;
1777 int (*varies) PARAMS ((rtx, int));
1779 register rtx x_addr;
1781 if (MEM_VOLATILE_P (x) && MEM_VOLATILE_P (mem))
1784 if (DIFFERENT_ALIAS_SETS_P (x, mem))
1787 /* If X is an unchanging read, then it can't possibly conflict with any
1788 non-unchanging store. It may conflict with an unchanging write though,
1789 because there may be a single store to this address to initialize it.
1790 Just fall through to the code below to resolve the case where we have
1791 both an unchanging read and an unchanging write. This won't handle all
1792 cases optimally, but the possible performance loss should be
1794 if (RTX_UNCHANGING_P (x) && ! RTX_UNCHANGING_P (mem))
1797 x_addr = get_addr (XEXP (x, 0));
1799 if (! base_alias_check (x_addr, mem_addr, GET_MODE (x), mem_mode))
1802 x_addr = canon_rtx (x_addr);
1803 if (! memrefs_conflict_p (GET_MODE_SIZE (mem_mode), mem_addr,
1804 SIZE_FOR_MODE (x), x_addr, 0))
1807 if (aliases_everything_p (x))
1810 /* We cannot use aliases_everyting_p to test MEM, since we must look
1811 at MEM_MODE, rather than GET_MODE (MEM). */
1812 if (mem_mode == QImode || GET_CODE (mem_addr) == AND)
1815 /* In true_dependence we also allow BLKmode to alias anything. Why
1816 don't we do this in anti_dependence and output_dependence? */
1817 if (mem_mode == BLKmode || GET_MODE (x) == BLKmode)
1820 return ! fixed_scalar_and_varying_struct_p (mem, x, mem_addr, x_addr,
1824 /* Returns non-zero if a write to X might alias a previous read from
1825 (or, if WRITEP is non-zero, a write to) MEM. */
1828 write_dependence_p (mem, x, writep)
1833 rtx x_addr, mem_addr;
1837 if (MEM_VOLATILE_P (x) && MEM_VOLATILE_P (mem))
1840 if (DIFFERENT_ALIAS_SETS_P (x, mem))
1843 /* Unchanging memory can't conflict with non-unchanging memory. */
1844 if (RTX_UNCHANGING_P (x) != RTX_UNCHANGING_P (mem))
1847 /* If MEM is an unchanging read, then it can't possibly conflict with
1848 the store to X, because there is at most one store to MEM, and it must
1849 have occurred somewhere before MEM. */
1850 if (! writep && RTX_UNCHANGING_P (mem))
1853 x_addr = get_addr (XEXP (x, 0));
1854 mem_addr = get_addr (XEXP (mem, 0));
1858 base = find_base_term (mem_addr);
1859 if (base && (GET_CODE (base) == LABEL_REF
1860 || (GET_CODE (base) == SYMBOL_REF
1861 && CONSTANT_POOL_ADDRESS_P (base))))
1865 if (! base_alias_check (x_addr, mem_addr, GET_MODE (x),
1869 x_addr = canon_rtx (x_addr);
1870 mem_addr = canon_rtx (mem_addr);
1872 if (!memrefs_conflict_p (SIZE_FOR_MODE (mem), mem_addr,
1873 SIZE_FOR_MODE (x), x_addr, 0))
1877 = fixed_scalar_and_varying_struct_p (mem, x, mem_addr, x_addr,
1880 return (!(fixed_scalar == mem && !aliases_everything_p (x))
1881 && !(fixed_scalar == x && !aliases_everything_p (mem)));
1884 /* Anti dependence: X is written after read in MEM takes place. */
1887 anti_dependence (mem, x)
1891 return write_dependence_p (mem, x, /*writep=*/0);
1894 /* Output dependence: X is written after store in MEM takes place. */
1897 output_dependence (mem, x)
1901 return write_dependence_p (mem, x, /*writep=*/1);
1904 /* Returns non-zero if X mentions something which is not
1905 local to the function and is not constant. */
1908 nonlocal_mentioned_p (x)
1912 register RTX_CODE code;
1915 code = GET_CODE (x);
1917 if (GET_RTX_CLASS (code) == 'i')
1919 /* Constant functions can be constant if they don't use
1920 scratch memory used to mark function w/o side effects. */
1921 if (code == CALL_INSN && CONST_CALL_P (x))
1923 x = CALL_INSN_FUNCTION_USAGE (x);
1929 code = GET_CODE (x);
1935 if (GET_CODE (SUBREG_REG (x)) == REG)
1937 /* Global registers are not local. */
1938 if (REGNO (SUBREG_REG (x)) < FIRST_PSEUDO_REGISTER
1939 && global_regs[subreg_regno (x)])
1947 /* Global registers are not local. */
1948 if (regno < FIRST_PSEUDO_REGISTER && global_regs[regno])
1962 /* Constants in the function's constants pool are constant. */
1963 if (CONSTANT_POOL_ADDRESS_P (x))
1968 /* Non-constant calls and recursion are not local. */
1972 /* Be overly conservative and consider any volatile memory
1973 reference as not local. */
1974 if (MEM_VOLATILE_P (x))
1976 base = find_base_term (XEXP (x, 0));
1979 /* A Pmode ADDRESS could be a reference via the structure value
1980 address or static chain. Such memory references are nonlocal.
1982 Thus, we have to examine the contents of the ADDRESS to find
1983 out if this is a local reference or not. */
1984 if (GET_CODE (base) == ADDRESS
1985 && GET_MODE (base) == Pmode
1986 && (XEXP (base, 0) == stack_pointer_rtx
1987 || XEXP (base, 0) == arg_pointer_rtx
1988 #if HARD_FRAME_POINTER_REGNUM != FRAME_POINTER_REGNUM
1989 || XEXP (base, 0) == hard_frame_pointer_rtx
1991 || XEXP (base, 0) == frame_pointer_rtx))
1993 /* Constants in the function's constant pool are constant. */
1994 if (GET_CODE (base) == SYMBOL_REF && CONSTANT_POOL_ADDRESS_P (base))
1999 case UNSPEC_VOLATILE:
2004 if (MEM_VOLATILE_P (x))
2013 /* Recursively scan the operands of this expression. */
2016 register const char *fmt = GET_RTX_FORMAT (code);
2019 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2021 if (fmt[i] == 'e' && XEXP (x, i))
2023 if (nonlocal_mentioned_p (XEXP (x, i)))
2026 else if (fmt[i] == 'E')
2029 for (j = 0; j < XVECLEN (x, i); j++)
2030 if (nonlocal_mentioned_p (XVECEXP (x, i, j)))
2039 /* Return non-zero if a loop (natural or otherwise) is present.
2040 Inspired by Depth_First_Search_PP described in:
2042 Advanced Compiler Design and Implementation
2044 Morgan Kaufmann, 1997
2046 and heavily borrowed from flow_depth_first_order_compute. */
2059 /* Allocate the preorder and postorder number arrays. */
2060 pre = (int *) xcalloc (n_basic_blocks, sizeof (int));
2061 post = (int *) xcalloc (n_basic_blocks, sizeof (int));
2063 /* Allocate stack for back-tracking up CFG. */
2064 stack = (edge *) xmalloc ((n_basic_blocks + 1) * sizeof (edge));
2067 /* Allocate bitmap to track nodes that have been visited. */
2068 visited = sbitmap_alloc (n_basic_blocks);
2070 /* None of the nodes in the CFG have been visited yet. */
2071 sbitmap_zero (visited);
2073 /* Push the first edge on to the stack. */
2074 stack[sp++] = ENTRY_BLOCK_PTR->succ;
2082 /* Look at the edge on the top of the stack. */
2087 /* Check if the edge destination has been visited yet. */
2088 if (dest != EXIT_BLOCK_PTR && ! TEST_BIT (visited, dest->index))
2090 /* Mark that we have visited the destination. */
2091 SET_BIT (visited, dest->index);
2093 pre[dest->index] = prenum++;
2097 /* Since the DEST node has been visited for the first
2098 time, check its successors. */
2099 stack[sp++] = dest->succ;
2102 post[dest->index] = postnum++;
2106 if (dest != EXIT_BLOCK_PTR
2107 && pre[src->index] >= pre[dest->index]
2108 && post[dest->index] == 0)
2111 if (! e->succ_next && src != ENTRY_BLOCK_PTR)
2112 post[src->index] = postnum++;
2115 stack[sp - 1] = e->succ_next;
2124 sbitmap_free (visited);
2129 /* Mark the function if it is constant. */
2132 mark_constant_function ()
2135 int nonlocal_mentioned;
2137 if (TREE_PUBLIC (current_function_decl)
2138 || TREE_READONLY (current_function_decl)
2139 || DECL_IS_PURE (current_function_decl)
2140 || TREE_THIS_VOLATILE (current_function_decl)
2141 || TYPE_MODE (TREE_TYPE (current_function_decl)) == VOIDmode)
2144 /* A loop might not return which counts as a side effect. */
2148 nonlocal_mentioned = 0;
2150 init_alias_analysis ();
2152 /* Determine if this is a constant function. */
2154 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
2155 if (INSN_P (insn) && nonlocal_mentioned_p (insn))
2157 nonlocal_mentioned = 1;
2161 end_alias_analysis ();
2163 /* Mark the function. */
2165 if (! nonlocal_mentioned)
2166 TREE_READONLY (current_function_decl) = 1;
2170 static HARD_REG_SET argument_registers;
2177 #ifndef OUTGOING_REGNO
2178 #define OUTGOING_REGNO(N) N
2180 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
2181 /* Check whether this register can hold an incoming pointer
2182 argument. FUNCTION_ARG_REGNO_P tests outgoing register
2183 numbers, so translate if necessary due to register windows. */
2184 if (FUNCTION_ARG_REGNO_P (OUTGOING_REGNO (i))
2185 && HARD_REGNO_MODE_OK (i, Pmode))
2186 SET_HARD_REG_BIT (argument_registers, i);
2188 alias_sets = splay_tree_new (splay_tree_compare_ints, 0, 0);
2191 /* Initialize the aliasing machinery. Initialize the REG_KNOWN_VALUE
2195 init_alias_analysis ()
2197 int maxreg = max_reg_num ();
2200 register unsigned int ui;
2203 reg_known_value_size = maxreg;
2206 = (rtx *) xcalloc ((maxreg - FIRST_PSEUDO_REGISTER), sizeof (rtx))
2207 - FIRST_PSEUDO_REGISTER;
2209 = (char*) xcalloc ((maxreg - FIRST_PSEUDO_REGISTER), sizeof (char))
2210 - FIRST_PSEUDO_REGISTER;
2212 /* Overallocate reg_base_value to allow some growth during loop
2213 optimization. Loop unrolling can create a large number of
2215 reg_base_value_size = maxreg * 2;
2216 reg_base_value = (rtx *) xcalloc (reg_base_value_size, sizeof (rtx));
2217 ggc_add_rtx_root (reg_base_value, reg_base_value_size);
2219 new_reg_base_value = (rtx *) xmalloc (reg_base_value_size * sizeof (rtx));
2220 reg_seen = (char *) xmalloc (reg_base_value_size);
2221 if (! reload_completed && flag_unroll_loops)
2223 /* ??? Why are we realloc'ing if we're just going to zero it? */
2224 alias_invariant = (rtx *)xrealloc (alias_invariant,
2225 reg_base_value_size * sizeof (rtx));
2226 memset ((char *)alias_invariant, 0, reg_base_value_size * sizeof (rtx));
2229 /* The basic idea is that each pass through this loop will use the
2230 "constant" information from the previous pass to propagate alias
2231 information through another level of assignments.
2233 This could get expensive if the assignment chains are long. Maybe
2234 we should throttle the number of iterations, possibly based on
2235 the optimization level or flag_expensive_optimizations.
2237 We could propagate more information in the first pass by making use
2238 of REG_N_SETS to determine immediately that the alias information
2239 for a pseudo is "constant".
2241 A program with an uninitialized variable can cause an infinite loop
2242 here. Instead of doing a full dataflow analysis to detect such problems
2243 we just cap the number of iterations for the loop.
2245 The state of the arrays for the set chain in question does not matter
2246 since the program has undefined behavior. */
2251 /* Assume nothing will change this iteration of the loop. */
2254 /* We want to assign the same IDs each iteration of this loop, so
2255 start counting from zero each iteration of the loop. */
2258 /* We're at the start of the funtion each iteration through the
2259 loop, so we're copying arguments. */
2260 copying_arguments = 1;
2262 /* Wipe the potential alias information clean for this pass. */
2263 memset ((char *) new_reg_base_value, 0, reg_base_value_size * sizeof (rtx));
2265 /* Wipe the reg_seen array clean. */
2266 memset ((char *) reg_seen, 0, reg_base_value_size);
2268 /* Mark all hard registers which may contain an address.
2269 The stack, frame and argument pointers may contain an address.
2270 An argument register which can hold a Pmode value may contain
2271 an address even if it is not in BASE_REGS.
2273 The address expression is VOIDmode for an argument and
2274 Pmode for other registers. */
2276 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
2277 if (TEST_HARD_REG_BIT (argument_registers, i))
2278 new_reg_base_value[i] = gen_rtx_ADDRESS (VOIDmode,
2279 gen_rtx_REG (Pmode, i));
2281 new_reg_base_value[STACK_POINTER_REGNUM]
2282 = gen_rtx_ADDRESS (Pmode, stack_pointer_rtx);
2283 new_reg_base_value[ARG_POINTER_REGNUM]
2284 = gen_rtx_ADDRESS (Pmode, arg_pointer_rtx);
2285 new_reg_base_value[FRAME_POINTER_REGNUM]
2286 = gen_rtx_ADDRESS (Pmode, frame_pointer_rtx);
2287 #if HARD_FRAME_POINTER_REGNUM != FRAME_POINTER_REGNUM
2288 new_reg_base_value[HARD_FRAME_POINTER_REGNUM]
2289 = gen_rtx_ADDRESS (Pmode, hard_frame_pointer_rtx);
2292 /* Walk the insns adding values to the new_reg_base_value array. */
2293 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
2299 #if defined (HAVE_prologue) || defined (HAVE_epilogue)
2300 /* The prologue/epilouge insns are not threaded onto the
2301 insn chain until after reload has completed. Thus,
2302 there is no sense wasting time checking if INSN is in
2303 the prologue/epilogue until after reload has completed. */
2304 if (reload_completed
2305 && prologue_epilogue_contains (insn))
2309 /* If this insn has a noalias note, process it, Otherwise,
2310 scan for sets. A simple set will have no side effects
2311 which could change the base value of any other register. */
2313 if (GET_CODE (PATTERN (insn)) == SET
2314 && REG_NOTES (insn) != 0
2315 && find_reg_note (insn, REG_NOALIAS, NULL_RTX))
2316 record_set (SET_DEST (PATTERN (insn)), NULL_RTX, NULL);
2318 note_stores (PATTERN (insn), record_set, NULL);
2320 set = single_set (insn);
2323 && GET_CODE (SET_DEST (set)) == REG
2324 && REGNO (SET_DEST (set)) >= FIRST_PSEUDO_REGISTER)
2326 unsigned int regno = REGNO (SET_DEST (set));
2327 rtx src = SET_SRC (set);
2329 if (REG_NOTES (insn) != 0
2330 && (((note = find_reg_note (insn, REG_EQUAL, 0)) != 0
2331 && REG_N_SETS (regno) == 1)
2332 || (note = find_reg_note (insn, REG_EQUIV, NULL_RTX)) != 0)
2333 && GET_CODE (XEXP (note, 0)) != EXPR_LIST
2334 && ! rtx_varies_p (XEXP (note, 0), 1)
2335 && ! reg_overlap_mentioned_p (SET_DEST (set), XEXP (note, 0)))
2337 reg_known_value[regno] = XEXP (note, 0);
2338 reg_known_equiv_p[regno] = REG_NOTE_KIND (note) == REG_EQUIV;
2340 else if (REG_N_SETS (regno) == 1
2341 && GET_CODE (src) == PLUS
2342 && GET_CODE (XEXP (src, 0)) == REG
2343 && REGNO (XEXP (src, 0)) >= FIRST_PSEUDO_REGISTER
2344 && (reg_known_value[REGNO (XEXP (src, 0))])
2345 && GET_CODE (XEXP (src, 1)) == CONST_INT)
2347 rtx op0 = XEXP (src, 0);
2348 op0 = reg_known_value[REGNO (op0)];
2349 reg_known_value[regno]
2350 = plus_constant (op0, INTVAL (XEXP (src, 1)));
2351 reg_known_equiv_p[regno] = 0;
2353 else if (REG_N_SETS (regno) == 1
2354 && ! rtx_varies_p (src, 1))
2356 reg_known_value[regno] = src;
2357 reg_known_equiv_p[regno] = 0;
2361 else if (GET_CODE (insn) == NOTE
2362 && NOTE_LINE_NUMBER (insn) == NOTE_INSN_FUNCTION_BEG)
2363 copying_arguments = 0;
2366 /* Now propagate values from new_reg_base_value to reg_base_value. */
2367 for (ui = 0; ui < reg_base_value_size; ui++)
2369 if (new_reg_base_value[ui]
2370 && new_reg_base_value[ui] != reg_base_value[ui]
2371 && ! rtx_equal_p (new_reg_base_value[ui], reg_base_value[ui]))
2373 reg_base_value[ui] = new_reg_base_value[ui];
2378 while (changed && ++pass < MAX_ALIAS_LOOP_PASSES);
2380 /* Fill in the remaining entries. */
2381 for (i = FIRST_PSEUDO_REGISTER; i < maxreg; i++)
2382 if (reg_known_value[i] == 0)
2383 reg_known_value[i] = regno_reg_rtx[i];
2385 /* Simplify the reg_base_value array so that no register refers to
2386 another register, except to special registers indirectly through
2387 ADDRESS expressions.
2389 In theory this loop can take as long as O(registers^2), but unless
2390 there are very long dependency chains it will run in close to linear
2393 This loop may not be needed any longer now that the main loop does
2394 a better job at propagating alias information. */
2400 for (ui = 0; ui < reg_base_value_size; ui++)
2402 rtx base = reg_base_value[ui];
2403 if (base && GET_CODE (base) == REG)
2405 unsigned int base_regno = REGNO (base);
2406 if (base_regno == ui) /* register set from itself */
2407 reg_base_value[ui] = 0;
2409 reg_base_value[ui] = reg_base_value[base_regno];
2414 while (changed && pass < MAX_ALIAS_LOOP_PASSES);
2417 free (new_reg_base_value);
2418 new_reg_base_value = 0;
2424 end_alias_analysis ()
2426 free (reg_known_value + FIRST_PSEUDO_REGISTER);
2427 reg_known_value = 0;
2428 reg_known_value_size = 0;
2429 free (reg_known_equiv_p + FIRST_PSEUDO_REGISTER);
2430 reg_known_equiv_p = 0;
2433 ggc_del_root (reg_base_value);
2434 free (reg_base_value);
2437 reg_base_value_size = 0;
2438 if (alias_invariant)
2440 free (alias_invariant);
2441 alias_invariant = 0;