1 /* Alias analysis for GNU C
2 Copyright (C) 1997, 1998, 1999, 2000 Free Software Foundation, Inc.
3 Contributed by John Carr (jfc@mit.edu).
5 This file is part of GNU CC.
7 GNU CC is free software; you can redistribute it and/or modify
8 it under the terms of the GNU General Public License as published by
9 the Free Software Foundation; either version 2, or (at your option)
12 GNU CC is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License for more details.
17 You should have received a copy of the GNU General Public License
18 along with GNU CC; see the file COPYING. If not, write to
19 the Free Software Foundation, 59 Temple Place - Suite 330,
20 Boston, MA 02111-1307, USA. */
28 #include "insn-flags.h"
31 #include "hard-reg-set.h"
32 #include "basic-block.h"
37 #include "splay-tree.h"
40 /* The alias sets assigned to MEMs assist the back-end in determining
41 which MEMs can alias which other MEMs. In general, two MEMs in
42 different alias sets cannot alias each other, with one important
43 exception. Consider something like:
45 struct S {int i; double d; };
47 a store to an `S' can alias something of either type `int' or type
48 `double'. (However, a store to an `int' cannot alias a `double'
49 and vice versa.) We indicate this via a tree structure that looks
57 (The arrows are directed and point downwards.)
58 In this situation we say the alias set for `struct S' is the
59 `superset' and that those for `int' and `double' are `subsets'.
61 To see whether two alias sets can point to the same memory, we must
62 see if either alias set is a subset of the other. We need not trace
63 past immediate decendents, however, since we propagate all
64 grandchildren up one level.
66 Alias set zero is implicitly a superset of all other alias sets.
67 However, this is no actual entry for alias set zero. It is an
68 error to attempt to explicitly construct a subset of zero. */
70 typedef struct alias_set_entry
72 /* The alias set number, as stored in MEM_ALIAS_SET. */
73 HOST_WIDE_INT alias_set;
75 /* The children of the alias set. These are not just the immediate
76 children, but, in fact, all decendents. So, if we have:
78 struct T { struct S s; float f; }
80 continuing our example above, the children here will be all of
81 `int', `double', `float', and `struct S'. */
84 /* Nonzero if would have a child of zero: this effectively makes this
85 alias set the same as alias set zero. */
89 static int rtx_equal_for_memref_p PARAMS ((rtx, rtx));
90 static rtx find_symbolic_term PARAMS ((rtx));
91 static rtx get_addr PARAMS ((rtx));
92 static int memrefs_conflict_p PARAMS ((int, rtx, int, rtx,
94 static void record_set PARAMS ((rtx, rtx, void *));
95 static rtx find_base_term PARAMS ((rtx));
96 static int base_alias_check PARAMS ((rtx, rtx, enum machine_mode,
98 static rtx find_base_value PARAMS ((rtx));
99 static int mems_in_disjoint_alias_sets_p PARAMS ((rtx, rtx));
100 static int insert_subset_children PARAMS ((splay_tree_node, void*));
101 static tree find_base_decl PARAMS ((tree));
102 static alias_set_entry get_alias_set_entry PARAMS ((HOST_WIDE_INT));
103 static rtx fixed_scalar_and_varying_struct_p PARAMS ((rtx, rtx, rtx, rtx,
104 int (*) (rtx, int)));
105 static int aliases_everything_p PARAMS ((rtx));
106 static int write_dependence_p PARAMS ((rtx, rtx, int));
107 static int nonlocal_mentioned_p PARAMS ((rtx));
109 static int loop_p PARAMS ((void));
111 /* Set up all info needed to perform alias analysis on memory references. */
113 /* Returns the size in bytes of the mode of X. */
114 #define SIZE_FOR_MODE(X) (GET_MODE_SIZE (GET_MODE (X)))
116 /* Returns nonzero if MEM1 and MEM2 do not alias because they are in
117 different alias sets. We ignore alias sets in functions making use
118 of variable arguments because the va_arg macros on some systems are
120 #define DIFFERENT_ALIAS_SETS_P(MEM1, MEM2) \
121 mems_in_disjoint_alias_sets_p (MEM1, MEM2)
123 /* Cap the number of passes we make over the insns propagating alias
124 information through set chains. 10 is a completely arbitrary choice. */
125 #define MAX_ALIAS_LOOP_PASSES 10
127 /* reg_base_value[N] gives an address to which register N is related.
128 If all sets after the first add or subtract to the current value
129 or otherwise modify it so it does not point to a different top level
130 object, reg_base_value[N] is equal to the address part of the source
133 A base address can be an ADDRESS, SYMBOL_REF, or LABEL_REF. ADDRESS
134 expressions represent certain special values: function arguments and
135 the stack, frame, and argument pointers.
137 The contents of an ADDRESS is not normally used, the mode of the
138 ADDRESS determines whether the ADDRESS is a function argument or some
139 other special value. Pointer equality, not rtx_equal_p, determines whether
140 two ADDRESS expressions refer to the same base address.
142 The only use of the contents of an ADDRESS is for determining if the
143 current function performs nonlocal memory memory references for the
144 purposes of marking the function as a constant function. */
146 static rtx *reg_base_value;
147 static rtx *new_reg_base_value;
148 static unsigned int reg_base_value_size; /* size of reg_base_value array */
150 #define REG_BASE_VALUE(X) \
151 (REGNO (X) < reg_base_value_size \
152 ? reg_base_value[REGNO (X)] : 0)
154 /* Vector of known invariant relationships between registers. Set in
155 loop unrolling. Indexed by register number, if nonzero the value
156 is an expression describing this register in terms of another.
158 The length of this array is REG_BASE_VALUE_SIZE.
160 Because this array contains only pseudo registers it has no effect
162 static rtx *alias_invariant;
164 /* Vector indexed by N giving the initial (unchanging) value known for
165 pseudo-register N. This array is initialized in
166 init_alias_analysis, and does not change until end_alias_analysis
168 rtx *reg_known_value;
170 /* Indicates number of valid entries in reg_known_value. */
171 static unsigned int reg_known_value_size;
173 /* Vector recording for each reg_known_value whether it is due to a
174 REG_EQUIV note. Future passes (viz., reload) may replace the
175 pseudo with the equivalent expression and so we account for the
176 dependences that would be introduced if that happens.
178 The REG_EQUIV notes created in assign_parms may mention the arg
179 pointer, and there are explicit insns in the RTL that modify the
180 arg pointer. Thus we must ensure that such insns don't get
181 scheduled across each other because that would invalidate the
182 REG_EQUIV notes. One could argue that the REG_EQUIV notes are
183 wrong, but solving the problem in the scheduler will likely give
184 better code, so we do it here. */
185 char *reg_known_equiv_p;
187 /* True when scanning insns from the start of the rtl to the
188 NOTE_INSN_FUNCTION_BEG note. */
189 static int copying_arguments;
191 /* The splay-tree used to store the various alias set entries. */
192 static splay_tree alias_sets;
194 /* Returns a pointer to the alias set entry for ALIAS_SET, if there is
195 such an entry, or NULL otherwise. */
197 static alias_set_entry
198 get_alias_set_entry (alias_set)
199 HOST_WIDE_INT alias_set;
202 = splay_tree_lookup (alias_sets, (splay_tree_key) alias_set);
204 return sn != 0 ? ((alias_set_entry) sn->value) : 0;
207 /* Returns nonzero if the alias sets for MEM1 and MEM2 are such that
208 the two MEMs cannot alias each other. */
211 mems_in_disjoint_alias_sets_p (mem1, mem2)
215 #ifdef ENABLE_CHECKING
216 /* Perform a basic sanity check. Namely, that there are no alias sets
217 if we're not using strict aliasing. This helps to catch bugs
218 whereby someone uses PUT_CODE, but doesn't clear MEM_ALIAS_SET, or
219 where a MEM is allocated in some way other than by the use of
220 gen_rtx_MEM, and the MEM_ALIAS_SET is not cleared. If we begin to
221 use alias sets to indicate that spilled registers cannot alias each
222 other, we might need to remove this check. */
223 if (! flag_strict_aliasing
224 && (MEM_ALIAS_SET (mem1) != 0 || MEM_ALIAS_SET (mem2) != 0))
228 return ! alias_sets_conflict_p (MEM_ALIAS_SET (mem1), MEM_ALIAS_SET (mem2));
231 /* Insert the NODE into the splay tree given by DATA. Used by
232 record_alias_subset via splay_tree_foreach. */
235 insert_subset_children (node, data)
236 splay_tree_node node;
239 splay_tree_insert ((splay_tree) data, node->key, node->value);
244 /* Return 1 if the two specified alias sets may conflict. */
247 alias_sets_conflict_p (set1, set2)
248 HOST_WIDE_INT set1, set2;
252 /* If have no alias set information for one of the operands, we have
253 to assume it can alias anything. */
254 if (set1 == 0 || set2 == 0
255 /* If the two alias sets are the same, they may alias. */
259 /* See if the first alias set is a subset of the second. */
260 ase = get_alias_set_entry (set1);
262 && (ase->has_zero_child
263 || splay_tree_lookup (ase->children,
264 (splay_tree_key) set2)))
267 /* Now do the same, but with the alias sets reversed. */
268 ase = get_alias_set_entry (set2);
270 && (ase->has_zero_child
271 || splay_tree_lookup (ase->children,
272 (splay_tree_key) set1)))
275 /* The two alias sets are distinct and neither one is the
276 child of the other. Therefore, they cannot alias. */
280 /* Return 1 if TYPE is a RECORD_TYPE, UNION_TYPE, or QUAL_UNION_TYPE and has
281 has any readonly fields. If any of the fields have types that
282 contain readonly fields, return true as well. */
285 readonly_fields_p (type)
290 if (TREE_CODE (type) != RECORD_TYPE && TREE_CODE (type) != UNION_TYPE
291 && TREE_CODE (type) != QUAL_UNION_TYPE)
294 for (field = TYPE_FIELDS (type); field != 0; field = TREE_CHAIN (field))
295 if (TREE_CODE (field) == FIELD_DECL
296 && (TREE_READONLY (field)
297 || readonly_fields_p (TREE_TYPE (field))))
303 /* Return 1 if any MEM object of type T1 will always conflict (using the
304 dependency routines in this file) with any MEM object of type T2.
305 This is used when allocating temporary storage. If T1 and/or T2 are
306 NULL_TREE, it means we know nothing about the storage. */
309 objects_must_conflict_p (t1, t2)
312 /* If they are the same type, they must conflict. */
314 /* Likewise if both are volatile. */
315 || (t1 != 0 && TYPE_VOLATILE (t1) && t2 != 0 && TYPE_VOLATILE (t2)))
318 /* We now know they are different types. If one or both has readonly fields
319 or if one is readonly and the other not, they may not conflict.
320 Likewise if one is aggregate and the other is scalar. */
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)))
325 || ((t1 != 0 && AGGREGATE_TYPE_P (t1))
326 != (t2 != 0 && AGGREGATE_TYPE_P (t2))))
329 /* Otherwise they conflict only if the alias sets conflict. */
330 return alias_sets_conflict_p (t1 ? get_alias_set (t1) : 0,
331 t2 ? get_alias_set (t2) : 0);
334 /* T is an expression with pointer type. Find the DECL on which this
335 expression is based. (For example, in `a[i]' this would be `a'.)
336 If there is no such DECL, or a unique decl cannot be determined,
337 NULL_TREE is retured. */
345 if (t == 0 || t == error_mark_node || ! POINTER_TYPE_P (TREE_TYPE (t)))
348 /* If this is a declaration, return it. */
349 if (TREE_CODE_CLASS (TREE_CODE (t)) == 'd')
352 /* Handle general expressions. It would be nice to deal with
353 COMPONENT_REFs here. If we could tell that `a' and `b' were the
354 same, then `a->f' and `b->f' are also the same. */
355 switch (TREE_CODE_CLASS (TREE_CODE (t)))
358 return find_base_decl (TREE_OPERAND (t, 0));
361 /* Return 0 if found in neither or both are the same. */
362 d0 = find_base_decl (TREE_OPERAND (t, 0));
363 d1 = find_base_decl (TREE_OPERAND (t, 1));
374 d0 = find_base_decl (TREE_OPERAND (t, 0));
375 d1 = find_base_decl (TREE_OPERAND (t, 1));
376 d0 = find_base_decl (TREE_OPERAND (t, 0));
377 d2 = find_base_decl (TREE_OPERAND (t, 2));
379 /* Set any nonzero values from the last, then from the first. */
380 if (d1 == 0) d1 = d2;
381 if (d0 == 0) d0 = d1;
382 if (d1 == 0) d1 = d0;
383 if (d2 == 0) d2 = d1;
385 /* At this point all are nonzero or all are zero. If all three are the
386 same, return it. Otherwise, return zero. */
387 return (d0 == d1 && d1 == d2) ? d0 : 0;
394 /* Return the alias set for T, which may be either a type or an
395 expression. Call language-specific routine for help, if needed. */
404 /* If we're not doing any alias analysis, just assume everything
405 aliases everything else. Also return 0 if this or its type is
407 if (! flag_strict_aliasing || t == error_mark_node
409 && (TREE_TYPE (t) == 0 || TREE_TYPE (t) == error_mark_node)))
412 /* We can be passed either an expression or a type. This and the
413 language-specific routine may make mutually-recursive calls to
414 each other to figure out what to do. At each juncture, we see if
415 this is a tree that the language may need to handle specially.
416 First handle things that aren't types and start by removing nops
417 since we care only about the actual object. */
420 while (TREE_CODE (t) == NOP_EXPR || TREE_CODE (t) == CONVERT_EXPR
421 || TREE_CODE (t) == NON_LVALUE_EXPR)
422 t = TREE_OPERAND (t, 0);
424 /* Now give the language a chance to do something but record what we
425 gave it this time. */
427 if ((set = lang_get_alias_set (t)) != -1)
430 /* Now loop the same way as get_inner_reference and get the alias
431 set to use. Pick up the outermost object that we could have
435 /* Unnamed bitfields are not an addressable object. */
436 if (TREE_CODE (t) == BIT_FIELD_REF)
438 else if (TREE_CODE (t) == COMPONENT_REF)
440 if (! DECL_NONADDRESSABLE_P (TREE_OPERAND (t, 1)))
441 /* Stop at an adressable decl. */
444 else if (TREE_CODE (t) == ARRAY_REF)
446 if (! TYPE_NONALIASED_COMPONENT
447 (TREE_TYPE (TREE_OPERAND (t, 0))))
448 /* Stop at an addresssable array element. */
451 else if (TREE_CODE (t) != NON_LVALUE_EXPR
452 && ! ((TREE_CODE (t) == NOP_EXPR
453 || TREE_CODE (t) == CONVERT_EXPR)
454 && (TYPE_MODE (TREE_TYPE (t))
455 == TYPE_MODE (TREE_TYPE (TREE_OPERAND (t, 0))))))
456 /* Stop if not one of above and not mode-preserving conversion. */
459 t = TREE_OPERAND (t, 0);
462 if (TREE_CODE (t) == INDIRECT_REF)
464 /* Check for accesses through restrict-qualified pointers. */
465 tree decl = find_base_decl (TREE_OPERAND (t, 0));
467 if (decl && DECL_POINTER_ALIAS_SET_KNOWN_P (decl))
468 /* We use the alias set indicated in the declaration. */
469 return DECL_POINTER_ALIAS_SET (decl);
471 /* If we have an INDIRECT_REF via a void pointer, we don't
472 know anything about what that might alias. */
473 if (TREE_CODE (TREE_TYPE (t)) == VOID_TYPE)
477 /* Give the language another chance to do something special. */
479 && (set = lang_get_alias_set (t)) != -1)
482 /* Now all we care about is the type. */
486 /* Variant qualifiers don't affect the alias set, so get the main
487 variant. If this is a type with a known alias set, return it. */
488 t = TYPE_MAIN_VARIANT (t);
489 if (TYPE_P (t) && TYPE_ALIAS_SET_KNOWN_P (t))
490 return TYPE_ALIAS_SET (t);
492 /* See if the language has special handling for this type. */
493 if ((set = lang_get_alias_set (t)) != -1)
495 /* If the alias set is now known, we are done. */
496 if (TYPE_ALIAS_SET_KNOWN_P (t))
497 return TYPE_ALIAS_SET (t);
500 /* There are no objects of FUNCTION_TYPE, so there's no point in
501 using up an alias set for them. (There are, of course, pointers
502 and references to functions, but that's different.) */
503 else if (TREE_CODE (t) == FUNCTION_TYPE)
506 /* Otherwise make a new alias set for this type. */
507 set = new_alias_set ();
509 TYPE_ALIAS_SET (t) = set;
511 /* If this is an aggregate type, we must record any component aliasing
513 if (AGGREGATE_TYPE_P (t) || TREE_CODE (t) == COMPLEX_TYPE)
514 record_component_aliases (t);
519 /* Return a brand-new alias set. */
524 static HOST_WIDE_INT last_alias_set;
526 if (flag_strict_aliasing)
527 return ++last_alias_set;
532 /* Indicate that things in SUBSET can alias things in SUPERSET, but
533 not vice versa. For example, in C, a store to an `int' can alias a
534 structure containing an `int', but not vice versa. Here, the
535 structure would be the SUPERSET and `int' the SUBSET. This
536 function should be called only once per SUPERSET/SUBSET pair.
538 It is illegal for SUPERSET to be zero; everything is implicitly a
539 subset of alias set zero. */
542 record_alias_subset (superset, subset)
543 HOST_WIDE_INT superset;
544 HOST_WIDE_INT subset;
546 alias_set_entry superset_entry;
547 alias_set_entry subset_entry;
552 superset_entry = get_alias_set_entry (superset);
553 if (superset_entry == 0)
555 /* Create an entry for the SUPERSET, so that we have a place to
556 attach the SUBSET. */
558 = (alias_set_entry) xmalloc (sizeof (struct alias_set_entry));
559 superset_entry->alias_set = superset;
560 superset_entry->children
561 = splay_tree_new (splay_tree_compare_ints, 0, 0);
562 superset_entry->has_zero_child = 0;
563 splay_tree_insert (alias_sets, (splay_tree_key) superset,
564 (splay_tree_value) superset_entry);
568 superset_entry->has_zero_child = 1;
571 subset_entry = get_alias_set_entry (subset);
572 /* If there is an entry for the subset, enter all of its children
573 (if they are not already present) as children of the SUPERSET. */
576 if (subset_entry->has_zero_child)
577 superset_entry->has_zero_child = 1;
579 splay_tree_foreach (subset_entry->children, insert_subset_children,
580 superset_entry->children);
583 /* Enter the SUBSET itself as a child of the SUPERSET. */
584 splay_tree_insert (superset_entry->children,
585 (splay_tree_key) subset, 0);
589 /* Record that component types of TYPE, if any, are part of that type for
590 aliasing purposes. For record types, we only record component types
591 for fields that are marked addressable. For array types, we always
592 record the component types, so the front end should not call this
593 function if the individual component aren't addressable. */
596 record_component_aliases (type)
599 HOST_WIDE_INT superset = get_alias_set (type);
605 switch (TREE_CODE (type))
608 if (! TYPE_NONALIASED_COMPONENT (type))
609 record_alias_subset (superset, get_alias_set (TREE_TYPE (type)));
614 case QUAL_UNION_TYPE:
615 for (field = TYPE_FIELDS (type); field != 0; field = TREE_CHAIN (field))
616 if (TREE_CODE (field) == FIELD_DECL && ! DECL_NONADDRESSABLE_P (field))
617 record_alias_subset (superset, get_alias_set (TREE_TYPE (field)));
621 record_alias_subset (superset, get_alias_set (TREE_TYPE (type)));
629 /* Allocate an alias set for use in storing and reading from the varargs
633 get_varargs_alias_set ()
635 static HOST_WIDE_INT set = -1;
638 set = new_alias_set ();
643 /* Likewise, but used for the fixed portions of the frame, e.g., register
647 get_frame_alias_set ()
649 static HOST_WIDE_INT set = -1;
652 set = new_alias_set ();
657 /* Inside SRC, the source of a SET, find a base address. */
660 find_base_value (src)
664 switch (GET_CODE (src))
672 /* At the start of a function, argument registers have known base
673 values which may be lost later. Returning an ADDRESS
674 expression here allows optimization based on argument values
675 even when the argument registers are used for other purposes. */
676 if (regno < FIRST_PSEUDO_REGISTER && copying_arguments)
677 return new_reg_base_value[regno];
679 /* If a pseudo has a known base value, return it. Do not do this
680 for hard regs since it can result in a circular dependency
681 chain for registers which have values at function entry.
683 The test above is not sufficient because the scheduler may move
684 a copy out of an arg reg past the NOTE_INSN_FUNCTION_BEGIN. */
685 if (regno >= FIRST_PSEUDO_REGISTER
686 && regno < reg_base_value_size
687 && reg_base_value[regno])
688 return reg_base_value[regno];
693 /* Check for an argument passed in memory. Only record in the
694 copying-arguments block; it is too hard to track changes
696 if (copying_arguments
697 && (XEXP (src, 0) == arg_pointer_rtx
698 || (GET_CODE (XEXP (src, 0)) == PLUS
699 && XEXP (XEXP (src, 0), 0) == arg_pointer_rtx)))
700 return gen_rtx_ADDRESS (VOIDmode, src);
705 if (GET_CODE (src) != PLUS && GET_CODE (src) != MINUS)
708 /* ... fall through ... */
713 rtx temp, src_0 = XEXP (src, 0), src_1 = XEXP (src, 1);
715 /* If either operand is a REG, then see if we already have
716 a known value for it. */
717 if (GET_CODE (src_0) == REG)
719 temp = find_base_value (src_0);
724 if (GET_CODE (src_1) == REG)
726 temp = find_base_value (src_1);
731 /* Guess which operand is the base address:
732 If either operand is a symbol, then it is the base. If
733 either operand is a CONST_INT, then the other is the base. */
734 if (GET_CODE (src_1) == CONST_INT || CONSTANT_P (src_0))
735 return find_base_value (src_0);
736 else if (GET_CODE (src_0) == CONST_INT || CONSTANT_P (src_1))
737 return find_base_value (src_1);
739 /* This might not be necessary anymore:
740 If either operand is a REG that is a known pointer, then it
742 else if (GET_CODE (src_0) == REG && REG_POINTER (src_0))
743 return find_base_value (src_0);
744 else if (GET_CODE (src_1) == REG && REG_POINTER (src_1))
745 return find_base_value (src_1);
751 /* The standard form is (lo_sum reg sym) so look only at the
753 return find_base_value (XEXP (src, 1));
756 /* If the second operand is constant set the base
757 address to the first operand. */
758 if (GET_CODE (XEXP (src, 1)) == CONST_INT && INTVAL (XEXP (src, 1)) != 0)
759 return find_base_value (XEXP (src, 0));
763 case SIGN_EXTEND: /* used for NT/Alpha pointers */
765 return find_base_value (XEXP (src, 0));
774 /* Called from init_alias_analysis indirectly through note_stores. */
776 /* While scanning insns to find base values, reg_seen[N] is nonzero if
777 register N has been set in this function. */
778 static char *reg_seen;
780 /* Addresses which are known not to alias anything else are identified
781 by a unique integer. */
782 static int unique_id;
785 record_set (dest, set, data)
787 void *data ATTRIBUTE_UNUSED;
789 register unsigned regno;
792 if (GET_CODE (dest) != REG)
795 regno = REGNO (dest);
797 if (regno >= reg_base_value_size)
802 /* A CLOBBER wipes out any old value but does not prevent a previously
803 unset register from acquiring a base address (i.e. reg_seen is not
805 if (GET_CODE (set) == CLOBBER)
807 new_reg_base_value[regno] = 0;
816 new_reg_base_value[regno] = 0;
820 new_reg_base_value[regno] = gen_rtx_ADDRESS (Pmode,
821 GEN_INT (unique_id++));
825 /* This is not the first set. If the new value is not related to the
826 old value, forget the base value. Note that the following code is
828 extern int x, y; int *p = &x; p += (&y-&x);
829 ANSI C does not allow computing the difference of addresses
830 of distinct top level objects. */
831 if (new_reg_base_value[regno])
832 switch (GET_CODE (src))
837 if (XEXP (src, 0) != dest && XEXP (src, 1) != dest)
838 new_reg_base_value[regno] = 0;
841 if (XEXP (src, 0) != dest || GET_CODE (XEXP (src, 1)) != CONST_INT)
842 new_reg_base_value[regno] = 0;
845 new_reg_base_value[regno] = 0;
848 /* If this is the first set of a register, record the value. */
849 else if ((regno >= FIRST_PSEUDO_REGISTER || ! fixed_regs[regno])
850 && ! reg_seen[regno] && new_reg_base_value[regno] == 0)
851 new_reg_base_value[regno] = find_base_value (src);
856 /* Called from loop optimization when a new pseudo-register is
857 created. It indicates that REGNO is being set to VAL. f INVARIANT
858 is true then this value also describes an invariant relationship
859 which can be used to deduce that two registers with unknown values
863 record_base_value (regno, val, invariant)
868 if (regno >= reg_base_value_size)
871 if (invariant && alias_invariant)
872 alias_invariant[regno] = val;
874 if (GET_CODE (val) == REG)
876 if (REGNO (val) < reg_base_value_size)
877 reg_base_value[regno] = reg_base_value[REGNO (val)];
882 reg_base_value[regno] = find_base_value (val);
885 /* Returns a canonical version of X, from the point of view alias
886 analysis. (For example, if X is a MEM whose address is a register,
887 and the register has a known value (say a SYMBOL_REF), then a MEM
888 whose address is the SYMBOL_REF is returned.) */
894 /* Recursively look for equivalences. */
895 if (GET_CODE (x) == REG && REGNO (x) >= FIRST_PSEUDO_REGISTER
896 && REGNO (x) < reg_known_value_size)
897 return reg_known_value[REGNO (x)] == x
898 ? x : canon_rtx (reg_known_value[REGNO (x)]);
899 else if (GET_CODE (x) == PLUS)
901 rtx x0 = canon_rtx (XEXP (x, 0));
902 rtx x1 = canon_rtx (XEXP (x, 1));
904 if (x0 != XEXP (x, 0) || x1 != XEXP (x, 1))
906 /* We can tolerate LO_SUMs being offset here; these
907 rtl are used for nothing other than comparisons. */
908 if (GET_CODE (x0) == CONST_INT)
909 return plus_constant_for_output (x1, INTVAL (x0));
910 else if (GET_CODE (x1) == CONST_INT)
911 return plus_constant_for_output (x0, INTVAL (x1));
912 return gen_rtx_PLUS (GET_MODE (x), x0, x1);
916 /* This gives us much better alias analysis when called from
917 the loop optimizer. Note we want to leave the original
918 MEM alone, but need to return the canonicalized MEM with
919 all the flags with their original values. */
920 else if (GET_CODE (x) == MEM)
922 rtx addr = canon_rtx (XEXP (x, 0));
924 if (addr != XEXP (x, 0))
926 rtx new = gen_rtx_MEM (GET_MODE (x), addr);
928 MEM_COPY_ATTRIBUTES (new, x);
935 /* Return 1 if X and Y are identical-looking rtx's.
937 We use the data in reg_known_value above to see if two registers with
938 different numbers are, in fact, equivalent. */
941 rtx_equal_for_memref_p (x, y)
946 register enum rtx_code code;
947 register const char *fmt;
949 if (x == 0 && y == 0)
951 if (x == 0 || y == 0)
961 /* Rtx's of different codes cannot be equal. */
962 if (code != GET_CODE (y))
965 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent.
966 (REG:SI x) and (REG:HI x) are NOT equivalent. */
968 if (GET_MODE (x) != GET_MODE (y))
971 /* Some RTL can be compared without a recursive examination. */
975 return REGNO (x) == REGNO (y);
978 return XEXP (x, 0) == XEXP (y, 0);
981 return XSTR (x, 0) == XSTR (y, 0);
985 /* There's no need to compare the contents of CONST_DOUBLEs or
986 CONST_INTs because pointer equality is a good enough
987 comparison for these nodes. */
991 return (REGNO (XEXP (x, 0)) == REGNO (XEXP (y, 0))
992 && XINT (x, 1) == XINT (y, 1));
998 /* For commutative operations, the RTX match if the operand match in any
999 order. Also handle the simple binary and unary cases without a loop. */
1000 if (code == EQ || code == NE || GET_RTX_CLASS (code) == 'c')
1001 return ((rtx_equal_for_memref_p (XEXP (x, 0), XEXP (y, 0))
1002 && rtx_equal_for_memref_p (XEXP (x, 1), XEXP (y, 1)))
1003 || (rtx_equal_for_memref_p (XEXP (x, 0), XEXP (y, 1))
1004 && rtx_equal_for_memref_p (XEXP (x, 1), XEXP (y, 0))));
1005 else if (GET_RTX_CLASS (code) == '<' || GET_RTX_CLASS (code) == '2')
1006 return (rtx_equal_for_memref_p (XEXP (x, 0), XEXP (y, 0))
1007 && rtx_equal_for_memref_p (XEXP (x, 1), XEXP (y, 1)));
1008 else if (GET_RTX_CLASS (code) == '1')
1009 return rtx_equal_for_memref_p (XEXP (x, 0), XEXP (y, 0));
1011 /* Compare the elements. If any pair of corresponding elements
1012 fail to match, return 0 for the whole things.
1014 Limit cases to types which actually appear in addresses. */
1016 fmt = GET_RTX_FORMAT (code);
1017 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1022 if (XINT (x, i) != XINT (y, i))
1027 /* Two vectors must have the same length. */
1028 if (XVECLEN (x, i) != XVECLEN (y, i))
1031 /* And the corresponding elements must match. */
1032 for (j = 0; j < XVECLEN (x, i); j++)
1033 if (rtx_equal_for_memref_p (XVECEXP (x, i, j),
1034 XVECEXP (y, i, j)) == 0)
1039 if (rtx_equal_for_memref_p (XEXP (x, i), XEXP (y, i)) == 0)
1043 /* This can happen for an asm which clobbers memory. */
1047 /* It is believed that rtx's at this level will never
1048 contain anything but integers and other rtx's,
1049 except for within LABEL_REFs and SYMBOL_REFs. */
1057 /* Given an rtx X, find a SYMBOL_REF or LABEL_REF within
1058 X and return it, or return 0 if none found. */
1061 find_symbolic_term (x)
1065 register enum rtx_code code;
1066 register const char *fmt;
1068 code = GET_CODE (x);
1069 if (code == SYMBOL_REF || code == LABEL_REF)
1071 if (GET_RTX_CLASS (code) == 'o')
1074 fmt = GET_RTX_FORMAT (code);
1075 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1081 t = find_symbolic_term (XEXP (x, i));
1085 else if (fmt[i] == 'E')
1096 struct elt_loc_list *l;
1098 #if defined (FIND_BASE_TERM)
1099 /* Try machine-dependent ways to find the base term. */
1100 x = FIND_BASE_TERM (x);
1103 switch (GET_CODE (x))
1106 return REG_BASE_VALUE (x);
1109 case SIGN_EXTEND: /* Used for Alpha/NT pointers */
1115 return find_base_term (XEXP (x, 0));
1118 val = CSELIB_VAL_PTR (x);
1119 for (l = val->locs; l; l = l->next)
1120 if ((x = find_base_term (l->loc)) != 0)
1126 if (GET_CODE (x) != PLUS && GET_CODE (x) != MINUS)
1133 rtx tmp1 = XEXP (x, 0);
1134 rtx tmp2 = XEXP (x, 1);
1136 /* This is a litle bit tricky since we have to determine which of
1137 the two operands represents the real base address. Otherwise this
1138 routine may return the index register instead of the base register.
1140 That may cause us to believe no aliasing was possible, when in
1141 fact aliasing is possible.
1143 We use a few simple tests to guess the base register. Additional
1144 tests can certainly be added. For example, if one of the operands
1145 is a shift or multiply, then it must be the index register and the
1146 other operand is the base register. */
1148 if (tmp1 == pic_offset_table_rtx && CONSTANT_P (tmp2))
1149 return find_base_term (tmp2);
1151 /* If either operand is known to be a pointer, then use it
1152 to determine the base term. */
1153 if (REG_P (tmp1) && REG_POINTER (tmp1))
1154 return find_base_term (tmp1);
1156 if (REG_P (tmp2) && REG_POINTER (tmp2))
1157 return find_base_term (tmp2);
1159 /* Neither operand was known to be a pointer. Go ahead and find the
1160 base term for both operands. */
1161 tmp1 = find_base_term (tmp1);
1162 tmp2 = find_base_term (tmp2);
1164 /* If either base term is named object or a special address
1165 (like an argument or stack reference), then use it for the
1168 && (GET_CODE (tmp1) == SYMBOL_REF
1169 || GET_CODE (tmp1) == LABEL_REF
1170 || (GET_CODE (tmp1) == ADDRESS
1171 && GET_MODE (tmp1) != VOIDmode)))
1175 && (GET_CODE (tmp2) == SYMBOL_REF
1176 || GET_CODE (tmp2) == LABEL_REF
1177 || (GET_CODE (tmp2) == ADDRESS
1178 && GET_MODE (tmp2) != VOIDmode)))
1181 /* We could not determine which of the two operands was the
1182 base register and which was the index. So we can determine
1183 nothing from the base alias check. */
1188 if (GET_CODE (XEXP (x, 0)) == REG && GET_CODE (XEXP (x, 1)) == CONST_INT)
1189 return REG_BASE_VALUE (XEXP (x, 0));
1197 return REG_BASE_VALUE (frame_pointer_rtx);
1204 /* Return 0 if the addresses X and Y are known to point to different
1205 objects, 1 if they might be pointers to the same object. */
1208 base_alias_check (x, y, x_mode, y_mode)
1210 enum machine_mode x_mode, y_mode;
1212 rtx x_base = find_base_term (x);
1213 rtx y_base = find_base_term (y);
1215 /* If the address itself has no known base see if a known equivalent
1216 value has one. If either address still has no known base, nothing
1217 is known about aliasing. */
1222 if (! flag_expensive_optimizations || (x_c = canon_rtx (x)) == x)
1225 x_base = find_base_term (x_c);
1233 if (! flag_expensive_optimizations || (y_c = canon_rtx (y)) == y)
1236 y_base = find_base_term (y_c);
1241 /* If the base addresses are equal nothing is known about aliasing. */
1242 if (rtx_equal_p (x_base, y_base))
1245 /* The base addresses of the read and write are different expressions.
1246 If they are both symbols and they are not accessed via AND, there is
1247 no conflict. We can bring knowledge of object alignment into play
1248 here. For example, on alpha, "char a, b;" can alias one another,
1249 though "char a; long b;" cannot. */
1250 if (GET_CODE (x_base) != ADDRESS && GET_CODE (y_base) != ADDRESS)
1252 if (GET_CODE (x) == AND && GET_CODE (y) == AND)
1254 if (GET_CODE (x) == AND
1255 && (GET_CODE (XEXP (x, 1)) != CONST_INT
1256 || GET_MODE_UNIT_SIZE (y_mode) < -INTVAL (XEXP (x, 1))))
1258 if (GET_CODE (y) == AND
1259 && (GET_CODE (XEXP (y, 1)) != CONST_INT
1260 || GET_MODE_UNIT_SIZE (x_mode) < -INTVAL (XEXP (y, 1))))
1262 /* Differing symbols never alias. */
1266 /* If one address is a stack reference there can be no alias:
1267 stack references using different base registers do not alias,
1268 a stack reference can not alias a parameter, and a stack reference
1269 can not alias a global. */
1270 if ((GET_CODE (x_base) == ADDRESS && GET_MODE (x_base) == Pmode)
1271 || (GET_CODE (y_base) == ADDRESS && GET_MODE (y_base) == Pmode))
1274 if (! flag_argument_noalias)
1277 if (flag_argument_noalias > 1)
1280 /* Weak noalias assertion (arguments are distinct, but may match globals). */
1281 return ! (GET_MODE (x_base) == VOIDmode && GET_MODE (y_base) == VOIDmode);
1284 /* Convert the address X into something we can use. This is done by returning
1285 it unchanged unless it is a value; in the latter case we call cselib to get
1286 a more useful rtx. */
1293 struct elt_loc_list *l;
1295 if (GET_CODE (x) != VALUE)
1297 v = CSELIB_VAL_PTR (x);
1298 for (l = v->locs; l; l = l->next)
1299 if (CONSTANT_P (l->loc))
1301 for (l = v->locs; l; l = l->next)
1302 if (GET_CODE (l->loc) != REG && GET_CODE (l->loc) != MEM)
1305 return v->locs->loc;
1309 /* Return the address of the (N_REFS + 1)th memory reference to ADDR
1310 where SIZE is the size in bytes of the memory reference. If ADDR
1311 is not modified by the memory reference then ADDR is returned. */
1314 addr_side_effect_eval (addr, size, n_refs)
1321 switch (GET_CODE (addr))
1324 offset = (n_refs + 1) * size;
1327 offset = -(n_refs + 1) * size;
1330 offset = n_refs * size;
1333 offset = -n_refs * size;
1341 addr = gen_rtx_PLUS (GET_MODE (addr), XEXP (addr, 0), GEN_INT (offset));
1343 addr = XEXP (addr, 0);
1348 /* Return nonzero if X and Y (memory addresses) could reference the
1349 same location in memory. C is an offset accumulator. When
1350 C is nonzero, we are testing aliases between X and Y + C.
1351 XSIZE is the size in bytes of the X reference,
1352 similarly YSIZE is the size in bytes for Y.
1354 If XSIZE or YSIZE is zero, we do not know the amount of memory being
1355 referenced (the reference was BLKmode), so make the most pessimistic
1358 If XSIZE or YSIZE is negative, we may access memory outside the object
1359 being referenced as a side effect. This can happen when using AND to
1360 align memory references, as is done on the Alpha.
1362 Nice to notice that varying addresses cannot conflict with fp if no
1363 local variables had their addresses taken, but that's too hard now. */
1366 memrefs_conflict_p (xsize, x, ysize, y, c)
1371 if (GET_CODE (x) == VALUE)
1373 if (GET_CODE (y) == VALUE)
1375 if (GET_CODE (x) == HIGH)
1377 else if (GET_CODE (x) == LO_SUM)
1380 x = canon_rtx (addr_side_effect_eval (x, xsize, 0));
1381 if (GET_CODE (y) == HIGH)
1383 else if (GET_CODE (y) == LO_SUM)
1386 y = canon_rtx (addr_side_effect_eval (y, ysize, 0));
1388 if (rtx_equal_for_memref_p (x, y))
1390 if (xsize <= 0 || ysize <= 0)
1392 if (c >= 0 && xsize > c)
1394 if (c < 0 && ysize+c > 0)
1399 /* This code used to check for conflicts involving stack references and
1400 globals but the base address alias code now handles these cases. */
1402 if (GET_CODE (x) == PLUS)
1404 /* The fact that X is canonicalized means that this
1405 PLUS rtx is canonicalized. */
1406 rtx x0 = XEXP (x, 0);
1407 rtx x1 = XEXP (x, 1);
1409 if (GET_CODE (y) == PLUS)
1411 /* The fact that Y is canonicalized means that this
1412 PLUS rtx is canonicalized. */
1413 rtx y0 = XEXP (y, 0);
1414 rtx y1 = XEXP (y, 1);
1416 if (rtx_equal_for_memref_p (x1, y1))
1417 return memrefs_conflict_p (xsize, x0, ysize, y0, c);
1418 if (rtx_equal_for_memref_p (x0, y0))
1419 return memrefs_conflict_p (xsize, x1, ysize, y1, c);
1420 if (GET_CODE (x1) == CONST_INT)
1422 if (GET_CODE (y1) == CONST_INT)
1423 return memrefs_conflict_p (xsize, x0, ysize, y0,
1424 c - INTVAL (x1) + INTVAL (y1));
1426 return memrefs_conflict_p (xsize, x0, ysize, y,
1429 else if (GET_CODE (y1) == CONST_INT)
1430 return memrefs_conflict_p (xsize, x, ysize, y0, c + INTVAL (y1));
1434 else if (GET_CODE (x1) == CONST_INT)
1435 return memrefs_conflict_p (xsize, x0, ysize, y, c - INTVAL (x1));
1437 else if (GET_CODE (y) == PLUS)
1439 /* The fact that Y is canonicalized means that this
1440 PLUS rtx is canonicalized. */
1441 rtx y0 = XEXP (y, 0);
1442 rtx y1 = XEXP (y, 1);
1444 if (GET_CODE (y1) == CONST_INT)
1445 return memrefs_conflict_p (xsize, x, ysize, y0, c + INTVAL (y1));
1450 if (GET_CODE (x) == GET_CODE (y))
1451 switch (GET_CODE (x))
1455 /* Handle cases where we expect the second operands to be the
1456 same, and check only whether the first operand would conflict
1459 rtx x1 = canon_rtx (XEXP (x, 1));
1460 rtx y1 = canon_rtx (XEXP (y, 1));
1461 if (! rtx_equal_for_memref_p (x1, y1))
1463 x0 = canon_rtx (XEXP (x, 0));
1464 y0 = canon_rtx (XEXP (y, 0));
1465 if (rtx_equal_for_memref_p (x0, y0))
1466 return (xsize == 0 || ysize == 0
1467 || (c >= 0 && xsize > c) || (c < 0 && ysize+c > 0));
1469 /* Can't properly adjust our sizes. */
1470 if (GET_CODE (x1) != CONST_INT)
1472 xsize /= INTVAL (x1);
1473 ysize /= INTVAL (x1);
1475 return memrefs_conflict_p (xsize, x0, ysize, y0, c);
1479 /* Are these registers known not to be equal? */
1480 if (alias_invariant)
1482 unsigned int r_x = REGNO (x), r_y = REGNO (y);
1483 rtx i_x, i_y; /* invariant relationships of X and Y */
1485 i_x = r_x >= reg_base_value_size ? 0 : alias_invariant[r_x];
1486 i_y = r_y >= reg_base_value_size ? 0 : alias_invariant[r_y];
1488 if (i_x == 0 && i_y == 0)
1491 if (! memrefs_conflict_p (xsize, i_x ? i_x : x,
1492 ysize, i_y ? i_y : y, c))
1501 /* Treat an access through an AND (e.g. a subword access on an Alpha)
1502 as an access with indeterminate size. Assume that references
1503 besides AND are aligned, so if the size of the other reference is
1504 at least as large as the alignment, assume no other overlap. */
1505 if (GET_CODE (x) == AND && GET_CODE (XEXP (x, 1)) == CONST_INT)
1507 if (GET_CODE (y) == AND || ysize < -INTVAL (XEXP (x, 1)))
1509 return memrefs_conflict_p (xsize, XEXP (x, 0), ysize, y, c);
1511 if (GET_CODE (y) == AND && GET_CODE (XEXP (y, 1)) == CONST_INT)
1513 /* ??? If we are indexing far enough into the array/structure, we
1514 may yet be able to determine that we can not overlap. But we
1515 also need to that we are far enough from the end not to overlap
1516 a following reference, so we do nothing with that for now. */
1517 if (GET_CODE (x) == AND || xsize < -INTVAL (XEXP (y, 1)))
1519 return memrefs_conflict_p (xsize, x, ysize, XEXP (y, 0), c);
1522 if (GET_CODE (x) == ADDRESSOF)
1524 if (y == frame_pointer_rtx
1525 || GET_CODE (y) == ADDRESSOF)
1526 return xsize <= 0 || ysize <= 0;
1528 if (GET_CODE (y) == ADDRESSOF)
1530 if (x == frame_pointer_rtx)
1531 return xsize <= 0 || ysize <= 0;
1536 if (GET_CODE (x) == CONST_INT && GET_CODE (y) == CONST_INT)
1538 c += (INTVAL (y) - INTVAL (x));
1539 return (xsize <= 0 || ysize <= 0
1540 || (c >= 0 && xsize > c) || (c < 0 && ysize+c > 0));
1543 if (GET_CODE (x) == CONST)
1545 if (GET_CODE (y) == CONST)
1546 return memrefs_conflict_p (xsize, canon_rtx (XEXP (x, 0)),
1547 ysize, canon_rtx (XEXP (y, 0)), c);
1549 return memrefs_conflict_p (xsize, canon_rtx (XEXP (x, 0)),
1552 if (GET_CODE (y) == CONST)
1553 return memrefs_conflict_p (xsize, x, ysize,
1554 canon_rtx (XEXP (y, 0)), c);
1557 return (xsize <= 0 || ysize <= 0
1558 || (rtx_equal_for_memref_p (x, y)
1559 && ((c >= 0 && xsize > c) || (c < 0 && ysize+c > 0))));
1566 /* Functions to compute memory dependencies.
1568 Since we process the insns in execution order, we can build tables
1569 to keep track of what registers are fixed (and not aliased), what registers
1570 are varying in known ways, and what registers are varying in unknown
1573 If both memory references are volatile, then there must always be a
1574 dependence between the two references, since their order can not be
1575 changed. A volatile and non-volatile reference can be interchanged
1578 A MEM_IN_STRUCT reference at a non-AND varying address can never
1579 conflict with a non-MEM_IN_STRUCT reference at a fixed address. We
1580 also must allow AND addresses, because they may generate accesses
1581 outside the object being referenced. This is used to generate
1582 aligned addresses from unaligned addresses, for instance, the alpha
1583 storeqi_unaligned pattern. */
1585 /* Read dependence: X is read after read in MEM takes place. There can
1586 only be a dependence here if both reads are volatile. */
1589 read_dependence (mem, x)
1593 return MEM_VOLATILE_P (x) && MEM_VOLATILE_P (mem);
1596 /* Returns MEM1 if and only if MEM1 is a scalar at a fixed address and
1597 MEM2 is a reference to a structure at a varying address, or returns
1598 MEM2 if vice versa. Otherwise, returns NULL_RTX. If a non-NULL
1599 value is returned MEM1 and MEM2 can never alias. VARIES_P is used
1600 to decide whether or not an address may vary; it should return
1601 nonzero whenever variation is possible.
1602 MEM1_ADDR and MEM2_ADDR are the addresses of MEM1 and MEM2. */
1605 fixed_scalar_and_varying_struct_p (mem1, mem2, mem1_addr, mem2_addr, varies_p)
1607 rtx mem1_addr, mem2_addr;
1608 int (*varies_p) PARAMS ((rtx, int));
1610 if (! flag_strict_aliasing)
1613 if (MEM_SCALAR_P (mem1) && MEM_IN_STRUCT_P (mem2)
1614 && !varies_p (mem1_addr, 1) && varies_p (mem2_addr, 1))
1615 /* MEM1 is a scalar at a fixed address; MEM2 is a struct at a
1619 if (MEM_IN_STRUCT_P (mem1) && MEM_SCALAR_P (mem2)
1620 && varies_p (mem1_addr, 1) && !varies_p (mem2_addr, 1))
1621 /* MEM2 is a scalar at a fixed address; MEM1 is a struct at a
1628 /* Returns nonzero if something about the mode or address format MEM1
1629 indicates that it might well alias *anything*. */
1632 aliases_everything_p (mem)
1635 if (GET_CODE (XEXP (mem, 0)) == AND)
1636 /* If the address is an AND, its very hard to know at what it is
1637 actually pointing. */
1643 /* True dependence: X is read after store in MEM takes place. */
1646 true_dependence (mem, mem_mode, x, varies)
1648 enum machine_mode mem_mode;
1650 int (*varies) PARAMS ((rtx, int));
1652 register rtx x_addr, mem_addr;
1655 if (MEM_VOLATILE_P (x) && MEM_VOLATILE_P (mem))
1658 if (DIFFERENT_ALIAS_SETS_P (x, mem))
1661 /* Unchanging memory can't conflict with non-unchanging memory.
1662 A non-unchanging read can conflict with a non-unchanging write.
1663 An unchanging read can conflict with an unchanging write since
1664 there may be a single store to this address to initialize it.
1665 Note that an unchanging store can conflict with a non-unchanging read
1666 since we have to make conservative assumptions when we have a
1667 record with readonly fields and we are copying the whole thing.
1668 Just fall through to the code below to resolve potential conflicts.
1669 This won't handle all cases optimally, but the possible performance
1670 loss should be negligible. */
1671 if (RTX_UNCHANGING_P (x) && ! RTX_UNCHANGING_P (mem))
1674 if (mem_mode == VOIDmode)
1675 mem_mode = GET_MODE (mem);
1677 x_addr = get_addr (XEXP (x, 0));
1678 mem_addr = get_addr (XEXP (mem, 0));
1680 base = find_base_term (x_addr);
1681 if (base && (GET_CODE (base) == LABEL_REF
1682 || (GET_CODE (base) == SYMBOL_REF
1683 && CONSTANT_POOL_ADDRESS_P (base))))
1686 if (! base_alias_check (x_addr, mem_addr, GET_MODE (x), mem_mode))
1689 x_addr = canon_rtx (x_addr);
1690 mem_addr = canon_rtx (mem_addr);
1692 if (! memrefs_conflict_p (GET_MODE_SIZE (mem_mode), mem_addr,
1693 SIZE_FOR_MODE (x), x_addr, 0))
1696 if (aliases_everything_p (x))
1699 /* We cannot use aliases_everyting_p to test MEM, since we must look
1700 at MEM_MODE, rather than GET_MODE (MEM). */
1701 if (mem_mode == QImode || GET_CODE (mem_addr) == AND)
1704 /* In true_dependence we also allow BLKmode to alias anything. Why
1705 don't we do this in anti_dependence and output_dependence? */
1706 if (mem_mode == BLKmode || GET_MODE (x) == BLKmode)
1709 return ! fixed_scalar_and_varying_struct_p (mem, x, mem_addr, x_addr,
1713 /* Returns non-zero if a write to X might alias a previous read from
1714 (or, if WRITEP is non-zero, a write to) MEM. */
1717 write_dependence_p (mem, x, writep)
1722 rtx x_addr, mem_addr;
1726 if (MEM_VOLATILE_P (x) && MEM_VOLATILE_P (mem))
1729 if (DIFFERENT_ALIAS_SETS_P (x, mem))
1732 /* Unchanging memory can't conflict with non-unchanging memory. */
1733 if (RTX_UNCHANGING_P (x) != RTX_UNCHANGING_P (mem))
1736 /* If MEM is an unchanging read, then it can't possibly conflict with
1737 the store to X, because there is at most one store to MEM, and it must
1738 have occurred somewhere before MEM. */
1739 if (! writep && RTX_UNCHANGING_P (mem))
1742 x_addr = get_addr (XEXP (x, 0));
1743 mem_addr = get_addr (XEXP (mem, 0));
1747 base = find_base_term (mem_addr);
1748 if (base && (GET_CODE (base) == LABEL_REF
1749 || (GET_CODE (base) == SYMBOL_REF
1750 && CONSTANT_POOL_ADDRESS_P (base))))
1754 if (! base_alias_check (x_addr, mem_addr, GET_MODE (x),
1758 x_addr = canon_rtx (x_addr);
1759 mem_addr = canon_rtx (mem_addr);
1761 if (!memrefs_conflict_p (SIZE_FOR_MODE (mem), mem_addr,
1762 SIZE_FOR_MODE (x), x_addr, 0))
1766 = fixed_scalar_and_varying_struct_p (mem, x, mem_addr, x_addr,
1769 return (!(fixed_scalar == mem && !aliases_everything_p (x))
1770 && !(fixed_scalar == x && !aliases_everything_p (mem)));
1773 /* Anti dependence: X is written after read in MEM takes place. */
1776 anti_dependence (mem, x)
1780 return write_dependence_p (mem, x, /*writep=*/0);
1783 /* Output dependence: X is written after store in MEM takes place. */
1786 output_dependence (mem, x)
1790 return write_dependence_p (mem, x, /*writep=*/1);
1793 /* Returns non-zero if X mentions something which is not
1794 local to the function and is not constant. */
1797 nonlocal_mentioned_p (x)
1801 register RTX_CODE code;
1804 code = GET_CODE (x);
1806 if (GET_RTX_CLASS (code) == 'i')
1808 /* Constant functions can be constant if they don't use
1809 scratch memory used to mark function w/o side effects. */
1810 if (code == CALL_INSN && CONST_CALL_P (x))
1812 x = CALL_INSN_FUNCTION_USAGE (x);
1818 code = GET_CODE (x);
1824 if (GET_CODE (SUBREG_REG (x)) == REG)
1826 /* Global registers are not local. */
1827 if (REGNO (SUBREG_REG (x)) < FIRST_PSEUDO_REGISTER
1828 && global_regs[REGNO (SUBREG_REG (x)) + SUBREG_WORD (x)])
1836 /* Global registers are not local. */
1837 if (regno < FIRST_PSEUDO_REGISTER && global_regs[regno])
1851 /* Constants in the function's constants pool are constant. */
1852 if (CONSTANT_POOL_ADDRESS_P (x))
1857 /* Non-constant calls and recursion are not local. */
1861 /* Be overly conservative and consider any volatile memory
1862 reference as not local. */
1863 if (MEM_VOLATILE_P (x))
1865 base = find_base_term (XEXP (x, 0));
1868 /* A Pmode ADDRESS could be a reference via the structure value
1869 address or static chain. Such memory references are nonlocal.
1871 Thus, we have to examine the contents of the ADDRESS to find
1872 out if this is a local reference or not. */
1873 if (GET_CODE (base) == ADDRESS
1874 && GET_MODE (base) == Pmode
1875 && (XEXP (base, 0) == stack_pointer_rtx
1876 || XEXP (base, 0) == arg_pointer_rtx
1877 #if HARD_FRAME_POINTER_REGNUM != FRAME_POINTER_REGNUM
1878 || XEXP (base, 0) == hard_frame_pointer_rtx
1880 || XEXP (base, 0) == frame_pointer_rtx))
1882 /* Constants in the function's constant pool are constant. */
1883 if (GET_CODE (base) == SYMBOL_REF && CONSTANT_POOL_ADDRESS_P (base))
1888 case UNSPEC_VOLATILE:
1893 if (MEM_VOLATILE_P (x))
1902 /* Recursively scan the operands of this expression. */
1905 register const char *fmt = GET_RTX_FORMAT (code);
1908 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1910 if (fmt[i] == 'e' && XEXP (x, i))
1912 if (nonlocal_mentioned_p (XEXP (x, i)))
1915 else if (fmt[i] == 'E')
1918 for (j = 0; j < XVECLEN (x, i); j++)
1919 if (nonlocal_mentioned_p (XVECEXP (x, i, j)))
1928 /* Return non-zero if a loop (natural or otherwise) is present.
1929 Inspired by Depth_First_Search_PP described in:
1931 Advanced Compiler Design and Implementation
1933 Morgan Kaufmann, 1997
1935 and heavily borrowed from flow_depth_first_order_compute. */
1948 /* Allocate the preorder and postorder number arrays. */
1949 pre = (int *) xcalloc (n_basic_blocks, sizeof (int));
1950 post = (int *) xcalloc (n_basic_blocks, sizeof (int));
1952 /* Allocate stack for back-tracking up CFG. */
1953 stack = (edge *) xmalloc ((n_basic_blocks + 1) * sizeof (edge));
1956 /* Allocate bitmap to track nodes that have been visited. */
1957 visited = sbitmap_alloc (n_basic_blocks);
1959 /* None of the nodes in the CFG have been visited yet. */
1960 sbitmap_zero (visited);
1962 /* Push the first edge on to the stack. */
1963 stack[sp++] = ENTRY_BLOCK_PTR->succ;
1971 /* Look at the edge on the top of the stack. */
1976 /* Check if the edge destination has been visited yet. */
1977 if (dest != EXIT_BLOCK_PTR && ! TEST_BIT (visited, dest->index))
1979 /* Mark that we have visited the destination. */
1980 SET_BIT (visited, dest->index);
1982 pre[dest->index] = prenum++;
1986 /* Since the DEST node has been visited for the first
1987 time, check its successors. */
1988 stack[sp++] = dest->succ;
1991 post[dest->index] = postnum++;
1995 if (dest != EXIT_BLOCK_PTR
1996 && pre[src->index] >= pre[dest->index]
1997 && post[dest->index] == 0)
2000 if (! e->succ_next && src != ENTRY_BLOCK_PTR)
2001 post[src->index] = postnum++;
2004 stack[sp - 1] = e->succ_next;
2013 sbitmap_free (visited);
2018 /* Mark the function if it is constant. */
2021 mark_constant_function ()
2024 int nonlocal_mentioned;
2026 if (TREE_PUBLIC (current_function_decl)
2027 || TREE_READONLY (current_function_decl)
2028 || DECL_IS_PURE (current_function_decl)
2029 || TREE_THIS_VOLATILE (current_function_decl)
2030 || TYPE_MODE (TREE_TYPE (current_function_decl)) == VOIDmode)
2033 /* A loop might not return which counts as a side effect. */
2037 nonlocal_mentioned = 0;
2039 init_alias_analysis ();
2041 /* Determine if this is a constant function. */
2043 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
2044 if (INSN_P (insn) && nonlocal_mentioned_p (insn))
2046 nonlocal_mentioned = 1;
2050 end_alias_analysis ();
2052 /* Mark the function. */
2054 if (! nonlocal_mentioned)
2055 TREE_READONLY (current_function_decl) = 1;
2059 static HARD_REG_SET argument_registers;
2066 #ifndef OUTGOING_REGNO
2067 #define OUTGOING_REGNO(N) N
2069 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
2070 /* Check whether this register can hold an incoming pointer
2071 argument. FUNCTION_ARG_REGNO_P tests outgoing register
2072 numbers, so translate if necessary due to register windows. */
2073 if (FUNCTION_ARG_REGNO_P (OUTGOING_REGNO (i))
2074 && HARD_REGNO_MODE_OK (i, Pmode))
2075 SET_HARD_REG_BIT (argument_registers, i);
2077 alias_sets = splay_tree_new (splay_tree_compare_ints, 0, 0);
2080 /* Initialize the aliasing machinery. Initialize the REG_KNOWN_VALUE
2084 init_alias_analysis ()
2086 int maxreg = max_reg_num ();
2089 register unsigned int ui;
2092 reg_known_value_size = maxreg;
2095 = (rtx *) xcalloc ((maxreg - FIRST_PSEUDO_REGISTER), sizeof (rtx))
2096 - FIRST_PSEUDO_REGISTER;
2098 = (char*) xcalloc ((maxreg - FIRST_PSEUDO_REGISTER), sizeof (char))
2099 - FIRST_PSEUDO_REGISTER;
2101 /* Overallocate reg_base_value to allow some growth during loop
2102 optimization. Loop unrolling can create a large number of
2104 reg_base_value_size = maxreg * 2;
2105 reg_base_value = (rtx *) xcalloc (reg_base_value_size, sizeof (rtx));
2106 ggc_add_rtx_root (reg_base_value, reg_base_value_size);
2108 new_reg_base_value = (rtx *) xmalloc (reg_base_value_size * sizeof (rtx));
2109 reg_seen = (char *) xmalloc (reg_base_value_size);
2110 if (! reload_completed && flag_unroll_loops)
2112 /* ??? Why are we realloc'ing if we're just going to zero it? */
2113 alias_invariant = (rtx *)xrealloc (alias_invariant,
2114 reg_base_value_size * sizeof (rtx));
2115 memset ((char *)alias_invariant, 0, reg_base_value_size * sizeof (rtx));
2118 /* The basic idea is that each pass through this loop will use the
2119 "constant" information from the previous pass to propagate alias
2120 information through another level of assignments.
2122 This could get expensive if the assignment chains are long. Maybe
2123 we should throttle the number of iterations, possibly based on
2124 the optimization level or flag_expensive_optimizations.
2126 We could propagate more information in the first pass by making use
2127 of REG_N_SETS to determine immediately that the alias information
2128 for a pseudo is "constant".
2130 A program with an uninitialized variable can cause an infinite loop
2131 here. Instead of doing a full dataflow analysis to detect such problems
2132 we just cap the number of iterations for the loop.
2134 The state of the arrays for the set chain in question does not matter
2135 since the program has undefined behavior. */
2140 /* Assume nothing will change this iteration of the loop. */
2143 /* We want to assign the same IDs each iteration of this loop, so
2144 start counting from zero each iteration of the loop. */
2147 /* We're at the start of the funtion each iteration through the
2148 loop, so we're copying arguments. */
2149 copying_arguments = 1;
2151 /* Wipe the potential alias information clean for this pass. */
2152 memset ((char *) new_reg_base_value, 0, reg_base_value_size * sizeof (rtx));
2154 /* Wipe the reg_seen array clean. */
2155 memset ((char *) reg_seen, 0, reg_base_value_size);
2157 /* Mark all hard registers which may contain an address.
2158 The stack, frame and argument pointers may contain an address.
2159 An argument register which can hold a Pmode value may contain
2160 an address even if it is not in BASE_REGS.
2162 The address expression is VOIDmode for an argument and
2163 Pmode for other registers. */
2165 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
2166 if (TEST_HARD_REG_BIT (argument_registers, i))
2167 new_reg_base_value[i] = gen_rtx_ADDRESS (VOIDmode,
2168 gen_rtx_REG (Pmode, i));
2170 new_reg_base_value[STACK_POINTER_REGNUM]
2171 = gen_rtx_ADDRESS (Pmode, stack_pointer_rtx);
2172 new_reg_base_value[ARG_POINTER_REGNUM]
2173 = gen_rtx_ADDRESS (Pmode, arg_pointer_rtx);
2174 new_reg_base_value[FRAME_POINTER_REGNUM]
2175 = gen_rtx_ADDRESS (Pmode, frame_pointer_rtx);
2176 #if HARD_FRAME_POINTER_REGNUM != FRAME_POINTER_REGNUM
2177 new_reg_base_value[HARD_FRAME_POINTER_REGNUM]
2178 = gen_rtx_ADDRESS (Pmode, hard_frame_pointer_rtx);
2181 /* Walk the insns adding values to the new_reg_base_value array. */
2182 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
2188 #if defined (HAVE_prologue) || defined (HAVE_epilogue)
2189 /* The prologue/epilouge insns are not threaded onto the
2190 insn chain until after reload has completed. Thus,
2191 there is no sense wasting time checking if INSN is in
2192 the prologue/epilogue until after reload has completed. */
2193 if (reload_completed
2194 && prologue_epilogue_contains (insn))
2198 /* If this insn has a noalias note, process it, Otherwise,
2199 scan for sets. A simple set will have no side effects
2200 which could change the base value of any other register. */
2202 if (GET_CODE (PATTERN (insn)) == SET
2203 && REG_NOTES (insn) != 0
2204 && find_reg_note (insn, REG_NOALIAS, NULL_RTX))
2205 record_set (SET_DEST (PATTERN (insn)), NULL_RTX, NULL);
2207 note_stores (PATTERN (insn), record_set, NULL);
2209 set = single_set (insn);
2212 && GET_CODE (SET_DEST (set)) == REG
2213 && REGNO (SET_DEST (set)) >= FIRST_PSEUDO_REGISTER)
2215 unsigned int regno = REGNO (SET_DEST (set));
2216 rtx src = SET_SRC (set);
2218 if (REG_NOTES (insn) != 0
2219 && (((note = find_reg_note (insn, REG_EQUAL, 0)) != 0
2220 && REG_N_SETS (regno) == 1)
2221 || (note = find_reg_note (insn, REG_EQUIV, NULL_RTX)) != 0)
2222 && GET_CODE (XEXP (note, 0)) != EXPR_LIST
2223 && ! reg_overlap_mentioned_p (SET_DEST (set), XEXP (note, 0)))
2225 reg_known_value[regno] = XEXP (note, 0);
2226 reg_known_equiv_p[regno] = REG_NOTE_KIND (note) == REG_EQUIV;
2228 else if (REG_N_SETS (regno) == 1
2229 && GET_CODE (src) == PLUS
2230 && GET_CODE (XEXP (src, 0)) == REG
2231 && REGNO (XEXP (src, 0)) >= FIRST_PSEUDO_REGISTER
2232 && (reg_known_value[REGNO (XEXP (src, 0))])
2233 && GET_CODE (XEXP (src, 1)) == CONST_INT)
2235 rtx op0 = XEXP (src, 0);
2236 if (reg_known_value[REGNO (op0)])
2237 op0 = reg_known_value[REGNO (op0)];
2238 reg_known_value[regno]
2239 = plus_constant_for_output (op0,
2240 INTVAL (XEXP (src, 1)));
2241 reg_known_equiv_p[regno] = 0;
2243 else if (REG_N_SETS (regno) == 1
2244 && ! rtx_varies_p (src, 1))
2246 reg_known_value[regno] = src;
2247 reg_known_equiv_p[regno] = 0;
2251 else if (GET_CODE (insn) == NOTE
2252 && NOTE_LINE_NUMBER (insn) == NOTE_INSN_FUNCTION_BEG)
2253 copying_arguments = 0;
2256 /* Now propagate values from new_reg_base_value to reg_base_value. */
2257 for (ui = 0; ui < reg_base_value_size; ui++)
2259 if (new_reg_base_value[ui]
2260 && new_reg_base_value[ui] != reg_base_value[ui]
2261 && ! rtx_equal_p (new_reg_base_value[ui], reg_base_value[ui]))
2263 reg_base_value[ui] = new_reg_base_value[ui];
2268 while (changed && ++pass < MAX_ALIAS_LOOP_PASSES);
2270 /* Fill in the remaining entries. */
2271 for (i = FIRST_PSEUDO_REGISTER; i < maxreg; i++)
2272 if (reg_known_value[i] == 0)
2273 reg_known_value[i] = regno_reg_rtx[i];
2275 /* Simplify the reg_base_value array so that no register refers to
2276 another register, except to special registers indirectly through
2277 ADDRESS expressions.
2279 In theory this loop can take as long as O(registers^2), but unless
2280 there are very long dependency chains it will run in close to linear
2283 This loop may not be needed any longer now that the main loop does
2284 a better job at propagating alias information. */
2290 for (ui = 0; ui < reg_base_value_size; ui++)
2292 rtx base = reg_base_value[ui];
2293 if (base && GET_CODE (base) == REG)
2295 unsigned int base_regno = REGNO (base);
2296 if (base_regno == ui) /* register set from itself */
2297 reg_base_value[ui] = 0;
2299 reg_base_value[ui] = reg_base_value[base_regno];
2304 while (changed && pass < MAX_ALIAS_LOOP_PASSES);
2307 free (new_reg_base_value);
2308 new_reg_base_value = 0;
2314 end_alias_analysis ()
2316 free (reg_known_value + FIRST_PSEUDO_REGISTER);
2317 reg_known_value = 0;
2318 reg_known_value_size = 0;
2319 free (reg_known_equiv_p + FIRST_PSEUDO_REGISTER);
2320 reg_known_equiv_p = 0;
2323 ggc_del_root (reg_base_value);
2324 free (reg_base_value);
2327 reg_base_value_size = 0;
2328 if (alias_invariant)
2330 free (alias_invariant);
2331 alias_invariant = 0;