]>
Commit | Line | Data |
---|---|---|
ea0cb7ae | 1 | /* Alias analysis for GNU C |
0975351b | 2 | Copyright (C) 1997, 1998, 1999, 2000, 2001, 2002, 2003, 2004, 2005 |
a303ed8c | 3 | Free Software Foundation, Inc. |
ea0cb7ae | 4 | Contributed by John Carr (jfc@mit.edu). |
5 | ||
f12b58b3 | 6 | This file is part of GCC. |
ea0cb7ae | 7 | |
f12b58b3 | 8 | GCC is free software; you can redistribute it and/or modify it under |
9 | the terms of the GNU General Public License as published by the Free | |
10 | Software Foundation; either version 2, or (at your option) any later | |
11 | version. | |
ea0cb7ae | 12 | |
f12b58b3 | 13 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY |
14 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
15 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
16 | for more details. | |
ea0cb7ae | 17 | |
18 | You should have received a copy of the GNU General Public License | |
f12b58b3 | 19 | along with GCC; see the file COPYING. If not, write to the Free |
67ce556b | 20 | Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA |
21 | 02110-1301, USA. */ | |
ea0cb7ae | 22 | |
23 | #include "config.h" | |
405711de | 24 | #include "system.h" |
805e22b2 | 25 | #include "coretypes.h" |
26 | #include "tm.h" | |
ea0cb7ae | 27 | #include "rtl.h" |
142131ed | 28 | #include "tree.h" |
7953c610 | 29 | #include "tm_p.h" |
0a893c29 | 30 | #include "function.h" |
4c74e6d9 | 31 | #include "alias.h" |
32 | #include "emit-rtl.h" | |
ea0cb7ae | 33 | #include "regs.h" |
34 | #include "hard-reg-set.h" | |
080f162d | 35 | #include "basic-block.h" |
ea0cb7ae | 36 | #include "flags.h" |
a068b2a9 | 37 | #include "output.h" |
ce1fd7fc | 38 | #include "toplev.h" |
1617c276 | 39 | #include "cselib.h" |
a5b1863e | 40 | #include "splay-tree.h" |
c2e94b25 | 41 | #include "ggc.h" |
b0278d39 | 42 | #include "langhooks.h" |
376c21d1 | 43 | #include "timevar.h" |
4f8e9d5c | 44 | #include "target.h" |
28992b23 | 45 | #include "cgraph.h" |
56bbdce4 | 46 | #include "varray.h" |
77fce4cd | 47 | #include "tree-pass.h" |
a5b1863e | 48 | |
49 | /* The alias sets assigned to MEMs assist the back-end in determining | |
50 | which MEMs can alias which other MEMs. In general, two MEMs in | |
5badbbc0 | 51 | different alias sets cannot alias each other, with one important |
52 | exception. Consider something like: | |
a5b1863e | 53 | |
8e2ef041 | 54 | struct S { int i; double d; }; |
a5b1863e | 55 | |
56 | a store to an `S' can alias something of either type `int' or type | |
57 | `double'. (However, a store to an `int' cannot alias a `double' | |
58 | and vice versa.) We indicate this via a tree structure that looks | |
59 | like: | |
60 | struct S | |
61 | / \ | |
62 | / \ | |
63 | |/_ _\| | |
64 | int double | |
65 | ||
5badbbc0 | 66 | (The arrows are directed and point downwards.) |
67 | In this situation we say the alias set for `struct S' is the | |
68 | `superset' and that those for `int' and `double' are `subsets'. | |
69 | ||
f7c44134 | 70 | To see whether two alias sets can point to the same memory, we must |
71 | see if either alias set is a subset of the other. We need not trace | |
d716ce75 | 72 | past immediate descendants, however, since we propagate all |
f7c44134 | 73 | grandchildren up one level. |
a5b1863e | 74 | |
75 | Alias set zero is implicitly a superset of all other alias sets. | |
76 | However, this is no actual entry for alias set zero. It is an | |
77 | error to attempt to explicitly construct a subset of zero. */ | |
78 | ||
318dcdd8 | 79 | struct alias_set_entry GTY(()) |
083a2b5e | 80 | { |
a5b1863e | 81 | /* The alias set number, as stored in MEM_ALIAS_SET. */ |
f7c44134 | 82 | HOST_WIDE_INT alias_set; |
a5b1863e | 83 | |
84 | /* The children of the alias set. These are not just the immediate | |
d716ce75 | 85 | children, but, in fact, all descendants. So, if we have: |
a5b1863e | 86 | |
9342ee68 | 87 | struct T { struct S s; float f; } |
a5b1863e | 88 | |
89 | continuing our example above, the children here will be all of | |
90 | `int', `double', `float', and `struct S'. */ | |
318dcdd8 | 91 | splay_tree GTY((param1_is (int), param2_is (int))) children; |
851dfbff | 92 | |
93 | /* Nonzero if would have a child of zero: this effectively makes this | |
94 | alias set the same as alias set zero. */ | |
95 | int has_zero_child; | |
318dcdd8 | 96 | }; |
97 | typedef struct alias_set_entry *alias_set_entry; | |
ea0cb7ae | 98 | |
aecda0d6 | 99 | static int rtx_equal_for_memref_p (rtx, rtx); |
100 | static rtx find_symbolic_term (rtx); | |
aecda0d6 | 101 | static int memrefs_conflict_p (int, rtx, int, rtx, HOST_WIDE_INT); |
102 | static void record_set (rtx, rtx, void *); | |
103 | static int base_alias_check (rtx, rtx, enum machine_mode, | |
104 | enum machine_mode); | |
105 | static rtx find_base_value (rtx); | |
106 | static int mems_in_disjoint_alias_sets_p (rtx, rtx); | |
107 | static int insert_subset_children (splay_tree_node, void*); | |
108 | static tree find_base_decl (tree); | |
109 | static alias_set_entry get_alias_set_entry (HOST_WIDE_INT); | |
110 | static rtx fixed_scalar_and_varying_struct_p (rtx, rtx, rtx, rtx, | |
111 | int (*) (rtx, int)); | |
112 | static int aliases_everything_p (rtx); | |
113 | static bool nonoverlapping_component_refs_p (tree, tree); | |
114 | static tree decl_for_component_ref (tree); | |
115 | static rtx adjust_offset_for_component_ref (tree, rtx); | |
116 | static int nonoverlapping_memrefs_p (rtx, rtx); | |
b04fab2a | 117 | static int write_dependence_p (rtx, rtx, int); |
aecda0d6 | 118 | |
119 | static int nonlocal_mentioned_p_1 (rtx *, void *); | |
120 | static int nonlocal_mentioned_p (rtx); | |
121 | static int nonlocal_referenced_p_1 (rtx *, void *); | |
122 | static int nonlocal_referenced_p (rtx); | |
123 | static int nonlocal_set_p_1 (rtx *, void *); | |
124 | static int nonlocal_set_p (rtx); | |
125 | static void memory_modified_1 (rtx, rtx, void *); | |
0aa9ffd4 | 126 | static void record_alias_subset (HOST_WIDE_INT, HOST_WIDE_INT); |
ea0cb7ae | 127 | |
128 | /* Set up all info needed to perform alias analysis on memory references. */ | |
129 | ||
083a2b5e | 130 | /* Returns the size in bytes of the mode of X. */ |
ea0cb7ae | 131 | #define SIZE_FOR_MODE(X) (GET_MODE_SIZE (GET_MODE (X))) |
132 | ||
b5ba9f3a | 133 | /* Returns nonzero if MEM1 and MEM2 do not alias because they are in |
a068b2a9 | 134 | different alias sets. We ignore alias sets in functions making use |
135 | of variable arguments because the va_arg macros on some systems are | |
136 | not legal ANSI C. */ | |
137 | #define DIFFERENT_ALIAS_SETS_P(MEM1, MEM2) \ | |
a5b1863e | 138 | mems_in_disjoint_alias_sets_p (MEM1, MEM2) |
b5ba9f3a | 139 | |
37b864db | 140 | /* Cap the number of passes we make over the insns propagating alias |
5badbbc0 | 141 | information through set chains. 10 is a completely arbitrary choice. */ |
37b864db | 142 | #define MAX_ALIAS_LOOP_PASSES 10 |
9342ee68 | 143 | |
ea0cb7ae | 144 | /* reg_base_value[N] gives an address to which register N is related. |
145 | If all sets after the first add or subtract to the current value | |
146 | or otherwise modify it so it does not point to a different top level | |
147 | object, reg_base_value[N] is equal to the address part of the source | |
e7c893ba | 148 | of the first set. |
149 | ||
150 | A base address can be an ADDRESS, SYMBOL_REF, or LABEL_REF. ADDRESS | |
151 | expressions represent certain special values: function arguments and | |
9342ee68 | 152 | the stack, frame, and argument pointers. |
3a935f06 | 153 | |
154 | The contents of an ADDRESS is not normally used, the mode of the | |
155 | ADDRESS determines whether the ADDRESS is a function argument or some | |
156 | other special value. Pointer equality, not rtx_equal_p, determines whether | |
157 | two ADDRESS expressions refer to the same base address. | |
158 | ||
159 | The only use of the contents of an ADDRESS is for determining if the | |
160 | current function performs nonlocal memory memory references for the | |
161 | purposes of marking the function as a constant function. */ | |
e7c893ba | 162 | |
e5b19a8c | 163 | static GTY(()) varray_type reg_base_value; |
c2e94b25 | 164 | static rtx *new_reg_base_value; |
e5b19a8c | 165 | |
166 | /* We preserve the copy of old array around to avoid amount of garbage | |
167 | produced. About 8% of garbage produced were attributed to this | |
168 | array. */ | |
7035b2ab | 169 | static GTY((deletable)) varray_type old_reg_base_value; |
083a2b5e | 170 | |
0c7f5242 | 171 | /* Static hunks of RTL used by the aliasing code; these are initialized |
172 | once per function to avoid unnecessary RTL allocations. */ | |
173 | static GTY (()) rtx static_reg_base_value[FIRST_PSEUDO_REGISTER]; | |
174 | ||
ea0cb7ae | 175 | #define REG_BASE_VALUE(X) \ |
e5b19a8c | 176 | (reg_base_value && REGNO (X) < VARRAY_SIZE (reg_base_value) \ |
177 | ? VARRAY_RTX (reg_base_value, REGNO (X)) : 0) | |
ea0cb7ae | 178 | |
1a6c3add | 179 | /* Vector of known invariant relationships between registers. Set in |
180 | loop unrolling. Indexed by register number, if nonzero the value | |
181 | is an expression describing this register in terms of another. | |
182 | ||
183 | The length of this array is REG_BASE_VALUE_SIZE. | |
184 | ||
185 | Because this array contains only pseudo registers it has no effect | |
186 | after reload. */ | |
12cb06ad | 187 | static GTY((length("alias_invariant_size"))) rtx *alias_invariant; |
8ca71ef2 | 188 | static GTY(()) unsigned int alias_invariant_size; |
1a6c3add | 189 | |
73f5c1e3 | 190 | /* Vector indexed by N giving the initial (unchanging) value known for |
12cb06ad | 191 | pseudo-register N. This array is initialized in init_alias_analysis, |
192 | and does not change until end_alias_analysis is called. */ | |
193 | static GTY((length("reg_known_value_size"))) rtx *reg_known_value; | |
ea0cb7ae | 194 | |
195 | /* Indicates number of valid entries in reg_known_value. */ | |
12cb06ad | 196 | static GTY(()) unsigned int reg_known_value_size; |
ea0cb7ae | 197 | |
198 | /* Vector recording for each reg_known_value whether it is due to a | |
199 | REG_EQUIV note. Future passes (viz., reload) may replace the | |
200 | pseudo with the equivalent expression and so we account for the | |
5badbbc0 | 201 | dependences that would be introduced if that happens. |
202 | ||
203 | The REG_EQUIV notes created in assign_parms may mention the arg | |
204 | pointer, and there are explicit insns in the RTL that modify the | |
205 | arg pointer. Thus we must ensure that such insns don't get | |
206 | scheduled across each other because that would invalidate the | |
207 | REG_EQUIV notes. One could argue that the REG_EQUIV notes are | |
208 | wrong, but solving the problem in the scheduler will likely give | |
209 | better code, so we do it here. */ | |
12cb06ad | 210 | static bool *reg_known_equiv_p; |
ea0cb7ae | 211 | |
e7c893ba | 212 | /* True when scanning insns from the start of the rtl to the |
213 | NOTE_INSN_FUNCTION_BEG note. */ | |
e5be0cc6 | 214 | static bool copying_arguments; |
ea0cb7ae | 215 | |
a5b1863e | 216 | /* The splay-tree used to store the various alias set entries. */ |
318dcdd8 | 217 | static GTY ((param_is (struct alias_set_entry))) varray_type alias_sets; |
5badbbc0 | 218 | \f |
a5b1863e | 219 | /* Returns a pointer to the alias set entry for ALIAS_SET, if there is |
220 | such an entry, or NULL otherwise. */ | |
221 | ||
56bbdce4 | 222 | static inline alias_set_entry |
aecda0d6 | 223 | get_alias_set_entry (HOST_WIDE_INT alias_set) |
a5b1863e | 224 | { |
56bbdce4 | 225 | return (alias_set_entry)VARRAY_GENERIC_PTR (alias_sets, alias_set); |
a5b1863e | 226 | } |
227 | ||
5badbbc0 | 228 | /* Returns nonzero if the alias sets for MEM1 and MEM2 are such that |
229 | the two MEMs cannot alias each other. */ | |
a5b1863e | 230 | |
56bbdce4 | 231 | static inline int |
aecda0d6 | 232 | mems_in_disjoint_alias_sets_p (rtx mem1, rtx mem2) |
a5b1863e | 233 | { |
a5b1863e | 234 | /* Perform a basic sanity check. Namely, that there are no alias sets |
235 | if we're not using strict aliasing. This helps to catch bugs | |
236 | whereby someone uses PUT_CODE, but doesn't clear MEM_ALIAS_SET, or | |
237 | where a MEM is allocated in some way other than by the use of | |
238 | gen_rtx_MEM, and the MEM_ALIAS_SET is not cleared. If we begin to | |
239 | use alias sets to indicate that spilled registers cannot alias each | |
240 | other, we might need to remove this check. */ | |
64db345d | 241 | gcc_assert (flag_strict_aliasing |
242 | || (!MEM_ALIAS_SET (mem1) && !MEM_ALIAS_SET (mem2))); | |
a5b1863e | 243 | |
387bc205 | 244 | return ! alias_sets_conflict_p (MEM_ALIAS_SET (mem1), MEM_ALIAS_SET (mem2)); |
245 | } | |
a5b1863e | 246 | |
387bc205 | 247 | /* Insert the NODE into the splay tree given by DATA. Used by |
248 | record_alias_subset via splay_tree_foreach. */ | |
249 | ||
250 | static int | |
aecda0d6 | 251 | insert_subset_children (splay_tree_node node, void *data) |
387bc205 | 252 | { |
253 | splay_tree_insert ((splay_tree) data, node->key, node->value); | |
254 | ||
255 | return 0; | |
256 | } | |
257 | ||
258 | /* Return 1 if the two specified alias sets may conflict. */ | |
259 | ||
260 | int | |
aecda0d6 | 261 | alias_sets_conflict_p (HOST_WIDE_INT set1, HOST_WIDE_INT set2) |
387bc205 | 262 | { |
263 | alias_set_entry ase; | |
264 | ||
265 | /* If have no alias set information for one of the operands, we have | |
266 | to assume it can alias anything. */ | |
267 | if (set1 == 0 || set2 == 0 | |
268 | /* If the two alias sets are the same, they may alias. */ | |
269 | || set1 == set2) | |
270 | return 1; | |
a5b1863e | 271 | |
f7c44134 | 272 | /* See if the first alias set is a subset of the second. */ |
387bc205 | 273 | ase = get_alias_set_entry (set1); |
851dfbff | 274 | if (ase != 0 |
275 | && (ase->has_zero_child | |
276 | || splay_tree_lookup (ase->children, | |
387bc205 | 277 | (splay_tree_key) set2))) |
278 | return 1; | |
a5b1863e | 279 | |
280 | /* Now do the same, but with the alias sets reversed. */ | |
387bc205 | 281 | ase = get_alias_set_entry (set2); |
851dfbff | 282 | if (ase != 0 |
283 | && (ase->has_zero_child | |
284 | || splay_tree_lookup (ase->children, | |
387bc205 | 285 | (splay_tree_key) set1))) |
286 | return 1; | |
a5b1863e | 287 | |
387bc205 | 288 | /* The two alias sets are distinct and neither one is the |
a5b1863e | 289 | child of the other. Therefore, they cannot alias. */ |
387bc205 | 290 | return 0; |
a5b1863e | 291 | } |
c1628b55 | 292 | |
293 | /* Return 1 if the two specified alias sets might conflict, or if any subtype | |
294 | of these alias sets might conflict. */ | |
295 | ||
296 | int | |
297 | alias_sets_might_conflict_p (HOST_WIDE_INT set1, HOST_WIDE_INT set2) | |
298 | { | |
299 | if (set1 == 0 || set2 == 0 || set1 == set2) | |
300 | return 1; | |
301 | ||
302 | return 0; | |
303 | } | |
304 | ||
f7c44134 | 305 | \f |
387bc205 | 306 | /* Return 1 if any MEM object of type T1 will always conflict (using the |
307 | dependency routines in this file) with any MEM object of type T2. | |
308 | This is used when allocating temporary storage. If T1 and/or T2 are | |
309 | NULL_TREE, it means we know nothing about the storage. */ | |
310 | ||
311 | int | |
aecda0d6 | 312 | objects_must_conflict_p (tree t1, tree t2) |
387bc205 | 313 | { |
2bedf233 | 314 | HOST_WIDE_INT set1, set2; |
315 | ||
218eb8d0 | 316 | /* If neither has a type specified, we don't know if they'll conflict |
317 | because we may be using them to store objects of various types, for | |
318 | example the argument and local variables areas of inlined functions. */ | |
8f2ba9c9 | 319 | if (t1 == 0 && t2 == 0) |
218eb8d0 | 320 | return 0; |
321 | ||
387bc205 | 322 | /* If they are the same type, they must conflict. */ |
323 | if (t1 == t2 | |
324 | /* Likewise if both are volatile. */ | |
325 | || (t1 != 0 && TYPE_VOLATILE (t1) && t2 != 0 && TYPE_VOLATILE (t2))) | |
326 | return 1; | |
327 | ||
2bedf233 | 328 | set1 = t1 ? get_alias_set (t1) : 0; |
329 | set2 = t2 ? get_alias_set (t2) : 0; | |
387bc205 | 330 | |
2bedf233 | 331 | /* Otherwise they conflict if they have no alias set or the same. We |
332 | can't simply use alias_sets_conflict_p here, because we must make | |
333 | sure that every subtype of t1 will conflict with every subtype of | |
334 | t2 for which a pair of subobjects of these respective subtypes | |
335 | overlaps on the stack. */ | |
336 | return set1 == 0 || set2 == 0 || set1 == set2; | |
387bc205 | 337 | } |
338 | \f | |
f7c44134 | 339 | /* T is an expression with pointer type. Find the DECL on which this |
340 | expression is based. (For example, in `a[i]' this would be `a'.) | |
341 | If there is no such DECL, or a unique decl cannot be determined, | |
35a3065a | 342 | NULL_TREE is returned. */ |
f7c44134 | 343 | |
344 | static tree | |
aecda0d6 | 345 | find_base_decl (tree t) |
f7c44134 | 346 | { |
ce45a448 | 347 | tree d0, d1; |
f7c44134 | 348 | |
349 | if (t == 0 || t == error_mark_node || ! POINTER_TYPE_P (TREE_TYPE (t))) | |
350 | return 0; | |
351 | ||
352 | /* If this is a declaration, return it. */ | |
ce45a448 | 353 | if (DECL_P (t)) |
f7c44134 | 354 | return t; |
355 | ||
356 | /* Handle general expressions. It would be nice to deal with | |
357 | COMPONENT_REFs here. If we could tell that `a' and `b' were the | |
358 | same, then `a->f' and `b->f' are also the same. */ | |
359 | switch (TREE_CODE_CLASS (TREE_CODE (t))) | |
360 | { | |
ce45a448 | 361 | case tcc_unary: |
f7c44134 | 362 | return find_base_decl (TREE_OPERAND (t, 0)); |
363 | ||
ce45a448 | 364 | case tcc_binary: |
f7c44134 | 365 | /* Return 0 if found in neither or both are the same. */ |
366 | d0 = find_base_decl (TREE_OPERAND (t, 0)); | |
367 | d1 = find_base_decl (TREE_OPERAND (t, 1)); | |
368 | if (d0 == d1) | |
369 | return d0; | |
370 | else if (d0 == 0) | |
371 | return d1; | |
372 | else if (d1 == 0) | |
373 | return d0; | |
374 | else | |
375 | return 0; | |
376 | ||
f7c44134 | 377 | default: |
378 | return 0; | |
379 | } | |
380 | } | |
381 | ||
2b02580f | 382 | /* Return true if all nested component references handled by |
383 | get_inner_reference in T are such that we should use the alias set | |
384 | provided by the object at the heart of T. | |
385 | ||
386 | This is true for non-addressable components (which don't have their | |
387 | own alias set), as well as components of objects in alias set zero. | |
388 | This later point is a special case wherein we wish to override the | |
389 | alias set used by the component, but we don't have per-FIELD_DECL | |
390 | assignable alias sets. */ | |
391 | ||
392 | bool | |
393 | component_uses_parent_alias_set (tree t) | |
bd32979f | 394 | { |
1f9b622b | 395 | while (1) |
396 | { | |
2b02580f | 397 | /* If we're at the end, it vacuously uses its own alias set. */ |
1f9b622b | 398 | if (!handled_component_p (t)) |
2b02580f | 399 | return false; |
1f9b622b | 400 | |
401 | switch (TREE_CODE (t)) | |
402 | { | |
403 | case COMPONENT_REF: | |
404 | if (DECL_NONADDRESSABLE_P (TREE_OPERAND (t, 1))) | |
2b02580f | 405 | return true; |
1f9b622b | 406 | break; |
407 | ||
408 | case ARRAY_REF: | |
409 | case ARRAY_RANGE_REF: | |
410 | if (TYPE_NONALIASED_COMPONENT (TREE_TYPE (TREE_OPERAND (t, 0)))) | |
2b02580f | 411 | return true; |
1f9b622b | 412 | break; |
413 | ||
414 | case REALPART_EXPR: | |
415 | case IMAGPART_EXPR: | |
416 | break; | |
417 | ||
418 | default: | |
419 | /* Bitfields and casts are never addressable. */ | |
2b02580f | 420 | return true; |
1f9b622b | 421 | } |
422 | ||
423 | t = TREE_OPERAND (t, 0); | |
2b02580f | 424 | if (get_alias_set (TREE_TYPE (t)) == 0) |
425 | return true; | |
1f9b622b | 426 | } |
bd32979f | 427 | } |
428 | ||
f7c44134 | 429 | /* Return the alias set for T, which may be either a type or an |
430 | expression. Call language-specific routine for help, if needed. */ | |
431 | ||
432 | HOST_WIDE_INT | |
aecda0d6 | 433 | get_alias_set (tree t) |
f7c44134 | 434 | { |
435 | HOST_WIDE_INT set; | |
f7c44134 | 436 | |
437 | /* If we're not doing any alias analysis, just assume everything | |
438 | aliases everything else. Also return 0 if this or its type is | |
439 | an error. */ | |
440 | if (! flag_strict_aliasing || t == error_mark_node | |
441 | || (! TYPE_P (t) | |
442 | && (TREE_TYPE (t) == 0 || TREE_TYPE (t) == error_mark_node))) | |
443 | return 0; | |
444 | ||
445 | /* We can be passed either an expression or a type. This and the | |
c3a9c149 | 446 | language-specific routine may make mutually-recursive calls to each other |
447 | to figure out what to do. At each juncture, we see if this is a tree | |
448 | that the language may need to handle specially. First handle things that | |
96216d37 | 449 | aren't types. */ |
1607663f | 450 | if (! TYPE_P (t)) |
f7c44134 | 451 | { |
96216d37 | 452 | tree inner = t; |
96216d37 | 453 | |
2a631e19 | 454 | /* Remove any nops, then give the language a chance to do |
455 | something with this tree before we look at it. */ | |
456 | STRIP_NOPS (t); | |
dc24ddbd | 457 | set = lang_hooks.get_alias_set (t); |
2a631e19 | 458 | if (set != -1) |
459 | return set; | |
460 | ||
96216d37 | 461 | /* First see if the actual object referenced is an INDIRECT_REF from a |
55f9d7dc | 462 | restrict-qualified pointer or a "void *". */ |
463 | while (handled_component_p (inner)) | |
96216d37 | 464 | { |
55f9d7dc | 465 | inner = TREE_OPERAND (inner, 0); |
2a631e19 | 466 | STRIP_NOPS (inner); |
96216d37 | 467 | } |
468 | ||
469 | /* Check for accesses through restrict-qualified pointers. */ | |
2a448a75 | 470 | if (INDIRECT_REF_P (inner)) |
96216d37 | 471 | { |
472 | tree decl = find_base_decl (TREE_OPERAND (inner, 0)); | |
473 | ||
474 | if (decl && DECL_POINTER_ALIAS_SET_KNOWN_P (decl)) | |
475 | { | |
4e69477b | 476 | /* If we haven't computed the actual alias set, do it now. */ |
96216d37 | 477 | if (DECL_POINTER_ALIAS_SET (decl) == -2) |
478 | { | |
a2ff5234 | 479 | tree pointed_to_type = TREE_TYPE (TREE_TYPE (decl)); |
480 | ||
96216d37 | 481 | /* No two restricted pointers can point at the same thing. |
482 | However, a restricted pointer can point at the same thing | |
483 | as an unrestricted pointer, if that unrestricted pointer | |
484 | is based on the restricted pointer. So, we make the | |
485 | alias set for the restricted pointer a subset of the | |
486 | alias set for the type pointed to by the type of the | |
487 | decl. */ | |
488 | HOST_WIDE_INT pointed_to_alias_set | |
a2ff5234 | 489 | = get_alias_set (pointed_to_type); |
96216d37 | 490 | |
491 | if (pointed_to_alias_set == 0) | |
492 | /* It's not legal to make a subset of alias set zero. */ | |
bcbf7cf4 | 493 | DECL_POINTER_ALIAS_SET (decl) = 0; |
a2ff5234 | 494 | else if (AGGREGATE_TYPE_P (pointed_to_type)) |
495 | /* For an aggregate, we must treat the restricted | |
496 | pointer the same as an ordinary pointer. If we | |
497 | were to make the type pointed to by the | |
498 | restricted pointer a subset of the pointed-to | |
499 | type, then we would believe that other subsets | |
500 | of the pointed-to type (such as fields of that | |
501 | type) do not conflict with the type pointed to | |
dac49aa5 | 502 | by the restricted pointer. */ |
a2ff5234 | 503 | DECL_POINTER_ALIAS_SET (decl) |
504 | = pointed_to_alias_set; | |
96216d37 | 505 | else |
506 | { | |
507 | DECL_POINTER_ALIAS_SET (decl) = new_alias_set (); | |
9342ee68 | 508 | record_alias_subset (pointed_to_alias_set, |
509 | DECL_POINTER_ALIAS_SET (decl)); | |
96216d37 | 510 | } |
511 | } | |
512 | ||
513 | /* We use the alias set indicated in the declaration. */ | |
514 | return DECL_POINTER_ALIAS_SET (decl); | |
515 | } | |
516 | ||
517 | /* If we have an INDIRECT_REF via a void pointer, we don't | |
b24423db | 518 | know anything about what that might alias. Likewise if the |
519 | pointer is marked that way. */ | |
520 | else if (TREE_CODE (TREE_TYPE (inner)) == VOID_TYPE | |
521 | || (TYPE_REF_CAN_ALIAS_ALL | |
522 | (TREE_TYPE (TREE_OPERAND (inner, 0))))) | |
96216d37 | 523 | return 0; |
524 | } | |
525 | ||
526 | /* Otherwise, pick up the outermost object that we could have a pointer | |
55f9d7dc | 527 | to, processing conversions as above. */ |
2b02580f | 528 | while (component_uses_parent_alias_set (t)) |
c3a9c149 | 529 | { |
55f9d7dc | 530 | t = TREE_OPERAND (t, 0); |
2a631e19 | 531 | STRIP_NOPS (t); |
532 | } | |
1607663f | 533 | |
96216d37 | 534 | /* If we've already determined the alias set for a decl, just return |
535 | it. This is necessary for C++ anonymous unions, whose component | |
536 | variables don't look like union members (boo!). */ | |
afce1f57 | 537 | if (TREE_CODE (t) == VAR_DECL |
e16ceb8e | 538 | && DECL_RTL_SET_P (t) && MEM_P (DECL_RTL (t))) |
afce1f57 | 539 | return MEM_ALIAS_SET (DECL_RTL (t)); |
540 | ||
1607663f | 541 | /* Now all we care about is the type. */ |
542 | t = TREE_TYPE (t); | |
f7c44134 | 543 | } |
544 | ||
f7c44134 | 545 | /* Variant qualifiers don't affect the alias set, so get the main |
546 | variant. If this is a type with a known alias set, return it. */ | |
547 | t = TYPE_MAIN_VARIANT (t); | |
96216d37 | 548 | if (TYPE_ALIAS_SET_KNOWN_P (t)) |
f7c44134 | 549 | return TYPE_ALIAS_SET (t); |
550 | ||
551 | /* See if the language has special handling for this type. */ | |
dc24ddbd | 552 | set = lang_hooks.get_alias_set (t); |
2a631e19 | 553 | if (set != -1) |
96216d37 | 554 | return set; |
851dfbff | 555 | |
f7c44134 | 556 | /* There are no objects of FUNCTION_TYPE, so there's no point in |
557 | using up an alias set for them. (There are, of course, pointers | |
558 | and references to functions, but that's different.) */ | |
559 | else if (TREE_CODE (t) == FUNCTION_TYPE) | |
560 | set = 0; | |
9fcc3e54 | 561 | |
562 | /* Unless the language specifies otherwise, let vector types alias | |
563 | their components. This avoids some nasty type punning issues in | |
564 | normal usage. And indeed lets vectors be treated more like an | |
565 | array slice. */ | |
566 | else if (TREE_CODE (t) == VECTOR_TYPE) | |
567 | set = get_alias_set (TREE_TYPE (t)); | |
568 | ||
f7c44134 | 569 | else |
570 | /* Otherwise make a new alias set for this type. */ | |
571 | set = new_alias_set (); | |
572 | ||
573 | TYPE_ALIAS_SET (t) = set; | |
851dfbff | 574 | |
575 | /* If this is an aggregate type, we must record any component aliasing | |
576 | information. */ | |
680210f2 | 577 | if (AGGREGATE_TYPE_P (t) || TREE_CODE (t) == COMPLEX_TYPE) |
851dfbff | 578 | record_component_aliases (t); |
579 | ||
f7c44134 | 580 | return set; |
581 | } | |
582 | ||
583 | /* Return a brand-new alias set. */ | |
584 | ||
4ec9406e | 585 | static GTY(()) HOST_WIDE_INT last_alias_set; |
586 | ||
f7c44134 | 587 | HOST_WIDE_INT |
aecda0d6 | 588 | new_alias_set (void) |
f7c44134 | 589 | { |
f7c44134 | 590 | if (flag_strict_aliasing) |
56bbdce4 | 591 | { |
243ec9ec | 592 | if (!alias_sets) |
593 | VARRAY_GENERIC_PTR_INIT (alias_sets, 10, "alias sets"); | |
594 | else | |
595 | VARRAY_GROW (alias_sets, last_alias_set + 2); | |
56bbdce4 | 596 | return ++last_alias_set; |
597 | } | |
f7c44134 | 598 | else |
599 | return 0; | |
600 | } | |
a5b1863e | 601 | |
8e2ef041 | 602 | /* Indicate that things in SUBSET can alias things in SUPERSET, but that |
603 | not everything that aliases SUPERSET also aliases SUBSET. For example, | |
604 | in C, a store to an `int' can alias a load of a structure containing an | |
605 | `int', and vice versa. But it can't alias a load of a 'double' member | |
606 | of the same structure. Here, the structure would be the SUPERSET and | |
607 | `int' the SUBSET. This relationship is also described in the comment at | |
608 | the beginning of this file. | |
609 | ||
610 | This function should be called only once per SUPERSET/SUBSET pair. | |
a5b1863e | 611 | |
612 | It is illegal for SUPERSET to be zero; everything is implicitly a | |
613 | subset of alias set zero. */ | |
614 | ||
0aa9ffd4 | 615 | static void |
aecda0d6 | 616 | record_alias_subset (HOST_WIDE_INT superset, HOST_WIDE_INT subset) |
a5b1863e | 617 | { |
618 | alias_set_entry superset_entry; | |
619 | alias_set_entry subset_entry; | |
620 | ||
c3a9c149 | 621 | /* It is possible in complex type situations for both sets to be the same, |
622 | in which case we can ignore this operation. */ | |
623 | if (superset == subset) | |
624 | return; | |
625 | ||
64db345d | 626 | gcc_assert (superset); |
a5b1863e | 627 | |
628 | superset_entry = get_alias_set_entry (superset); | |
9342ee68 | 629 | if (superset_entry == 0) |
a5b1863e | 630 | { |
631 | /* Create an entry for the SUPERSET, so that we have a place to | |
632 | attach the SUBSET. */ | |
318dcdd8 | 633 | superset_entry = ggc_alloc (sizeof (struct alias_set_entry)); |
a5b1863e | 634 | superset_entry->alias_set = superset; |
9342ee68 | 635 | superset_entry->children |
318dcdd8 | 636 | = splay_tree_new_ggc (splay_tree_compare_ints); |
f2e5de80 | 637 | superset_entry->has_zero_child = 0; |
56bbdce4 | 638 | VARRAY_GENERIC_PTR (alias_sets, superset) = superset_entry; |
a5b1863e | 639 | } |
640 | ||
851dfbff | 641 | if (subset == 0) |
642 | superset_entry->has_zero_child = 1; | |
643 | else | |
644 | { | |
645 | subset_entry = get_alias_set_entry (subset); | |
646 | /* If there is an entry for the subset, enter all of its children | |
647 | (if they are not already present) as children of the SUPERSET. */ | |
9342ee68 | 648 | if (subset_entry) |
851dfbff | 649 | { |
650 | if (subset_entry->has_zero_child) | |
651 | superset_entry->has_zero_child = 1; | |
083a2b5e | 652 | |
851dfbff | 653 | splay_tree_foreach (subset_entry->children, insert_subset_children, |
654 | superset_entry->children); | |
655 | } | |
a5b1863e | 656 | |
851dfbff | 657 | /* Enter the SUBSET itself as a child of the SUPERSET. */ |
9342ee68 | 658 | splay_tree_insert (superset_entry->children, |
851dfbff | 659 | (splay_tree_key) subset, 0); |
660 | } | |
a5b1863e | 661 | } |
662 | ||
9501a338 | 663 | /* Record that component types of TYPE, if any, are part of that type for |
664 | aliasing purposes. For record types, we only record component types | |
665 | for fields that are marked addressable. For array types, we always | |
666 | record the component types, so the front end should not call this | |
667 | function if the individual component aren't addressable. */ | |
668 | ||
669 | void | |
aecda0d6 | 670 | record_component_aliases (tree type) |
9501a338 | 671 | { |
f7c44134 | 672 | HOST_WIDE_INT superset = get_alias_set (type); |
9501a338 | 673 | tree field; |
674 | ||
675 | if (superset == 0) | |
676 | return; | |
677 | ||
678 | switch (TREE_CODE (type)) | |
679 | { | |
680 | case ARRAY_TYPE: | |
851dfbff | 681 | if (! TYPE_NONALIASED_COMPONENT (type)) |
682 | record_alias_subset (superset, get_alias_set (TREE_TYPE (type))); | |
9501a338 | 683 | break; |
684 | ||
685 | case RECORD_TYPE: | |
686 | case UNION_TYPE: | |
687 | case QUAL_UNION_TYPE: | |
8b332087 | 688 | /* Recursively record aliases for the base classes, if there are any. */ |
f6cc6a08 | 689 | if (TYPE_BINFO (type)) |
9342ee68 | 690 | { |
691 | int i; | |
f6cc6a08 | 692 | tree binfo, base_binfo; |
693 | ||
694 | for (binfo = TYPE_BINFO (type), i = 0; | |
695 | BINFO_BASE_ITERATE (binfo, i, base_binfo); i++) | |
696 | record_alias_subset (superset, | |
697 | get_alias_set (BINFO_TYPE (base_binfo))); | |
9342ee68 | 698 | } |
9501a338 | 699 | for (field = TYPE_FIELDS (type); field != 0; field = TREE_CHAIN (field)) |
e7bfabe2 | 700 | if (TREE_CODE (field) == FIELD_DECL && ! DECL_NONADDRESSABLE_P (field)) |
851dfbff | 701 | record_alias_subset (superset, get_alias_set (TREE_TYPE (field))); |
9501a338 | 702 | break; |
703 | ||
680210f2 | 704 | case COMPLEX_TYPE: |
705 | record_alias_subset (superset, get_alias_set (TREE_TYPE (type))); | |
706 | break; | |
707 | ||
9501a338 | 708 | default: |
709 | break; | |
710 | } | |
711 | } | |
712 | ||
f7c44134 | 713 | /* Allocate an alias set for use in storing and reading from the varargs |
714 | spill area. */ | |
715 | ||
4ec9406e | 716 | static GTY(()) HOST_WIDE_INT varargs_set = -1; |
717 | ||
f7c44134 | 718 | HOST_WIDE_INT |
aecda0d6 | 719 | get_varargs_alias_set (void) |
f7c44134 | 720 | { |
fcdd3ab3 | 721 | #if 1 |
722 | /* We now lower VA_ARG_EXPR, and there's currently no way to attach the | |
723 | varargs alias set to an INDIRECT_REF (FIXME!), so we can't | |
724 | consistently use the varargs alias set for loads from the varargs | |
725 | area. So don't use it anywhere. */ | |
726 | return 0; | |
727 | #else | |
4ec9406e | 728 | if (varargs_set == -1) |
729 | varargs_set = new_alias_set (); | |
f7c44134 | 730 | |
4ec9406e | 731 | return varargs_set; |
fcdd3ab3 | 732 | #endif |
f7c44134 | 733 | } |
734 | ||
735 | /* Likewise, but used for the fixed portions of the frame, e.g., register | |
736 | save areas. */ | |
737 | ||
4ec9406e | 738 | static GTY(()) HOST_WIDE_INT frame_set = -1; |
739 | ||
f7c44134 | 740 | HOST_WIDE_INT |
aecda0d6 | 741 | get_frame_alias_set (void) |
f7c44134 | 742 | { |
4ec9406e | 743 | if (frame_set == -1) |
744 | frame_set = new_alias_set (); | |
f7c44134 | 745 | |
4ec9406e | 746 | return frame_set; |
f7c44134 | 747 | } |
748 | ||
e7c893ba | 749 | /* Inside SRC, the source of a SET, find a base address. */ |
750 | ||
ea0cb7ae | 751 | static rtx |
aecda0d6 | 752 | find_base_value (rtx src) |
ea0cb7ae | 753 | { |
970c0c06 | 754 | unsigned int regno; |
d01410d3 | 755 | |
ea0cb7ae | 756 | switch (GET_CODE (src)) |
757 | { | |
758 | case SYMBOL_REF: | |
759 | case LABEL_REF: | |
760 | return src; | |
761 | ||
762 | case REG: | |
9db10305 | 763 | regno = REGNO (src); |
083a2b5e | 764 | /* At the start of a function, argument registers have known base |
e7c893ba | 765 | values which may be lost later. Returning an ADDRESS |
766 | expression here allows optimization based on argument values | |
767 | even when the argument registers are used for other purposes. */ | |
970c0c06 | 768 | if (regno < FIRST_PSEUDO_REGISTER && copying_arguments) |
769 | return new_reg_base_value[regno]; | |
1a538f67 | 770 | |
ebd13eec | 771 | /* If a pseudo has a known base value, return it. Do not do this |
9464d436 | 772 | for non-fixed hard regs since it can result in a circular |
773 | dependency chain for registers which have values at function entry. | |
ebd13eec | 774 | |
775 | The test above is not sufficient because the scheduler may move | |
776 | a copy out of an arg reg past the NOTE_INSN_FUNCTION_BEGIN. */ | |
9464d436 | 777 | if ((regno >= FIRST_PSEUDO_REGISTER || fixed_regs[regno]) |
e5b19a8c | 778 | && regno < VARRAY_SIZE (reg_base_value)) |
e5be0cc6 | 779 | { |
780 | /* If we're inside init_alias_analysis, use new_reg_base_value | |
781 | to reduce the number of relaxation iterations. */ | |
a303ed8c | 782 | if (new_reg_base_value && new_reg_base_value[regno] |
783 | && REG_N_SETS (regno) == 1) | |
e5be0cc6 | 784 | return new_reg_base_value[regno]; |
785 | ||
e5b19a8c | 786 | if (VARRAY_RTX (reg_base_value, regno)) |
787 | return VARRAY_RTX (reg_base_value, regno); | |
e5be0cc6 | 788 | } |
1a538f67 | 789 | |
1c5cb56c | 790 | return 0; |
ea0cb7ae | 791 | |
792 | case MEM: | |
793 | /* Check for an argument passed in memory. Only record in the | |
794 | copying-arguments block; it is too hard to track changes | |
795 | otherwise. */ | |
796 | if (copying_arguments | |
797 | && (XEXP (src, 0) == arg_pointer_rtx | |
798 | || (GET_CODE (XEXP (src, 0)) == PLUS | |
799 | && XEXP (XEXP (src, 0), 0) == arg_pointer_rtx))) | |
941522d6 | 800 | return gen_rtx_ADDRESS (VOIDmode, src); |
ea0cb7ae | 801 | return 0; |
802 | ||
803 | case CONST: | |
804 | src = XEXP (src, 0); | |
805 | if (GET_CODE (src) != PLUS && GET_CODE (src) != MINUS) | |
806 | break; | |
083a2b5e | 807 | |
2c0e001b | 808 | /* ... fall through ... */ |
e7c893ba | 809 | |
ea0cb7ae | 810 | case PLUS: |
811 | case MINUS: | |
e7c893ba | 812 | { |
06c5c535 | 813 | rtx temp, src_0 = XEXP (src, 0), src_1 = XEXP (src, 1); |
814 | ||
560de2f8 | 815 | /* If either operand is a REG that is a known pointer, then it |
816 | is the base. */ | |
817 | if (REG_P (src_0) && REG_POINTER (src_0)) | |
818 | return find_base_value (src_0); | |
819 | if (REG_P (src_1) && REG_POINTER (src_1)) | |
820 | return find_base_value (src_1); | |
821 | ||
06c5c535 | 822 | /* If either operand is a REG, then see if we already have |
823 | a known value for it. */ | |
560de2f8 | 824 | if (REG_P (src_0)) |
06c5c535 | 825 | { |
826 | temp = find_base_value (src_0); | |
083a2b5e | 827 | if (temp != 0) |
06c5c535 | 828 | src_0 = temp; |
829 | } | |
830 | ||
560de2f8 | 831 | if (REG_P (src_1)) |
06c5c535 | 832 | { |
833 | temp = find_base_value (src_1); | |
083a2b5e | 834 | if (temp!= 0) |
06c5c535 | 835 | src_1 = temp; |
836 | } | |
e7c893ba | 837 | |
560de2f8 | 838 | /* If either base is named object or a special address |
839 | (like an argument or stack reference), then use it for the | |
840 | base term. */ | |
841 | if (src_0 != 0 | |
842 | && (GET_CODE (src_0) == SYMBOL_REF | |
843 | || GET_CODE (src_0) == LABEL_REF | |
844 | || (GET_CODE (src_0) == ADDRESS | |
845 | && GET_MODE (src_0) != VOIDmode))) | |
846 | return src_0; | |
847 | ||
848 | if (src_1 != 0 | |
849 | && (GET_CODE (src_1) == SYMBOL_REF | |
850 | || GET_CODE (src_1) == LABEL_REF | |
851 | || (GET_CODE (src_1) == ADDRESS | |
852 | && GET_MODE (src_1) != VOIDmode))) | |
853 | return src_1; | |
854 | ||
083a2b5e | 855 | /* Guess which operand is the base address: |
06c5c535 | 856 | If either operand is a symbol, then it is the base. If |
857 | either operand is a CONST_INT, then the other is the base. */ | |
083a2b5e | 858 | if (GET_CODE (src_1) == CONST_INT || CONSTANT_P (src_0)) |
e7c893ba | 859 | return find_base_value (src_0); |
083a2b5e | 860 | else if (GET_CODE (src_0) == CONST_INT || CONSTANT_P (src_1)) |
06c5c535 | 861 | return find_base_value (src_1); |
862 | ||
ea0cb7ae | 863 | return 0; |
e7c893ba | 864 | } |
865 | ||
866 | case LO_SUM: | |
867 | /* The standard form is (lo_sum reg sym) so look only at the | |
868 | second operand. */ | |
869 | return find_base_value (XEXP (src, 1)); | |
ea0cb7ae | 870 | |
871 | case AND: | |
872 | /* If the second operand is constant set the base | |
2c0e001b | 873 | address to the first operand. */ |
e7c893ba | 874 | if (GET_CODE (XEXP (src, 1)) == CONST_INT && INTVAL (XEXP (src, 1)) != 0) |
875 | return find_base_value (XEXP (src, 0)); | |
ea0cb7ae | 876 | return 0; |
877 | ||
43d45282 | 878 | case TRUNCATE: |
879 | if (GET_MODE_SIZE (GET_MODE (src)) < GET_MODE_SIZE (Pmode)) | |
880 | break; | |
881 | /* Fall through. */ | |
ea0cb7ae | 882 | case HIGH: |
2be64ef0 | 883 | case PRE_INC: |
884 | case PRE_DEC: | |
885 | case POST_INC: | |
886 | case POST_DEC: | |
887 | case PRE_MODIFY: | |
888 | case POST_MODIFY: | |
e7c893ba | 889 | return find_base_value (XEXP (src, 0)); |
99c14947 | 890 | |
d01410d3 | 891 | case ZERO_EXTEND: |
892 | case SIGN_EXTEND: /* used for NT/Alpha pointers */ | |
893 | { | |
894 | rtx temp = find_base_value (XEXP (src, 0)); | |
895 | ||
85d654dd | 896 | if (temp != 0 && CONSTANT_P (temp)) |
d01410d3 | 897 | temp = convert_memory_address (Pmode, temp); |
d01410d3 | 898 | |
899 | return temp; | |
900 | } | |
901 | ||
99c14947 | 902 | default: |
903 | break; | |
ea0cb7ae | 904 | } |
905 | ||
906 | return 0; | |
907 | } | |
908 | ||
909 | /* Called from init_alias_analysis indirectly through note_stores. */ | |
910 | ||
083a2b5e | 911 | /* While scanning insns to find base values, reg_seen[N] is nonzero if |
ea0cb7ae | 912 | register N has been set in this function. */ |
913 | static char *reg_seen; | |
914 | ||
10882800 | 915 | /* Addresses which are known not to alias anything else are identified |
916 | by a unique integer. */ | |
06c5c535 | 917 | static int unique_id; |
918 | ||
e7c893ba | 919 | static void |
aecda0d6 | 920 | record_set (rtx dest, rtx set, void *data ATTRIBUTE_UNUSED) |
ea0cb7ae | 921 | { |
19cb6b50 | 922 | unsigned regno; |
ea0cb7ae | 923 | rtx src; |
3721f260 | 924 | int n; |
ea0cb7ae | 925 | |
8ad4c111 | 926 | if (!REG_P (dest)) |
ea0cb7ae | 927 | return; |
928 | ||
9db10305 | 929 | regno = REGNO (dest); |
ea0cb7ae | 930 | |
64db345d | 931 | gcc_assert (regno < VARRAY_SIZE (reg_base_value)); |
c2e94b25 | 932 | |
3721f260 | 933 | /* If this spans multiple hard registers, then we must indicate that every |
934 | register has an unusable value. */ | |
935 | if (regno < FIRST_PSEUDO_REGISTER) | |
67d6c12b | 936 | n = hard_regno_nregs[regno][GET_MODE (dest)]; |
3721f260 | 937 | else |
938 | n = 1; | |
939 | if (n != 1) | |
940 | { | |
941 | while (--n >= 0) | |
942 | { | |
943 | reg_seen[regno + n] = 1; | |
944 | new_reg_base_value[regno + n] = 0; | |
945 | } | |
946 | return; | |
947 | } | |
948 | ||
ea0cb7ae | 949 | if (set) |
950 | { | |
951 | /* A CLOBBER wipes out any old value but does not prevent a previously | |
952 | unset register from acquiring a base address (i.e. reg_seen is not | |
953 | set). */ | |
954 | if (GET_CODE (set) == CLOBBER) | |
955 | { | |
06c5c535 | 956 | new_reg_base_value[regno] = 0; |
ea0cb7ae | 957 | return; |
958 | } | |
959 | src = SET_SRC (set); | |
960 | } | |
961 | else | |
962 | { | |
ea0cb7ae | 963 | if (reg_seen[regno]) |
964 | { | |
06c5c535 | 965 | new_reg_base_value[regno] = 0; |
ea0cb7ae | 966 | return; |
967 | } | |
968 | reg_seen[regno] = 1; | |
941522d6 | 969 | new_reg_base_value[regno] = gen_rtx_ADDRESS (Pmode, |
970 | GEN_INT (unique_id++)); | |
ea0cb7ae | 971 | return; |
972 | } | |
973 | ||
ad6b1fe1 | 974 | /* If this is not the first set of REGNO, see whether the new value |
975 | is related to the old one. There are two cases of interest: | |
976 | ||
977 | (1) The register might be assigned an entirely new value | |
978 | that has the same base term as the original set. | |
979 | ||
980 | (2) The set might be a simple self-modification that | |
981 | cannot change REGNO's base value. | |
982 | ||
983 | If neither case holds, reject the original base value as invalid. | |
984 | Note that the following situation is not detected: | |
985 | ||
986 | extern int x, y; int *p = &x; p += (&y-&x); | |
987 | ||
ea0cb7ae | 988 | ANSI C does not allow computing the difference of addresses |
989 | of distinct top level objects. */ | |
ad6b1fe1 | 990 | if (new_reg_base_value[regno] != 0 |
991 | && find_base_value (src) != new_reg_base_value[regno]) | |
ea0cb7ae | 992 | switch (GET_CODE (src)) |
993 | { | |
e7c893ba | 994 | case LO_SUM: |
ea0cb7ae | 995 | case MINUS: |
996 | if (XEXP (src, 0) != dest && XEXP (src, 1) != dest) | |
06c5c535 | 997 | new_reg_base_value[regno] = 0; |
ea0cb7ae | 998 | break; |
43d45282 | 999 | case PLUS: |
1000 | /* If the value we add in the PLUS is also a valid base value, | |
1001 | this might be the actual base value, and the original value | |
1002 | an index. */ | |
1003 | { | |
1004 | rtx other = NULL_RTX; | |
1005 | ||
1006 | if (XEXP (src, 0) == dest) | |
1007 | other = XEXP (src, 1); | |
1008 | else if (XEXP (src, 1) == dest) | |
1009 | other = XEXP (src, 0); | |
1010 | ||
1011 | if (! other || find_base_value (other)) | |
1012 | new_reg_base_value[regno] = 0; | |
1013 | break; | |
1014 | } | |
ea0cb7ae | 1015 | case AND: |
1016 | if (XEXP (src, 0) != dest || GET_CODE (XEXP (src, 1)) != CONST_INT) | |
06c5c535 | 1017 | new_reg_base_value[regno] = 0; |
ea0cb7ae | 1018 | break; |
ea0cb7ae | 1019 | default: |
06c5c535 | 1020 | new_reg_base_value[regno] = 0; |
ea0cb7ae | 1021 | break; |
1022 | } | |
1023 | /* If this is the first set of a register, record the value. */ | |
1024 | else if ((regno >= FIRST_PSEUDO_REGISTER || ! fixed_regs[regno]) | |
06c5c535 | 1025 | && ! reg_seen[regno] && new_reg_base_value[regno] == 0) |
1026 | new_reg_base_value[regno] = find_base_value (src); | |
ea0cb7ae | 1027 | |
1028 | reg_seen[regno] = 1; | |
1029 | } | |
1030 | ||
5badbbc0 | 1031 | /* Called from loop optimization when a new pseudo-register is |
1032 | created. It indicates that REGNO is being set to VAL. f INVARIANT | |
1033 | is true then this value also describes an invariant relationship | |
1034 | which can be used to deduce that two registers with unknown values | |
1035 | are different. */ | |
083a2b5e | 1036 | |
ea0cb7ae | 1037 | void |
aecda0d6 | 1038 | record_base_value (unsigned int regno, rtx val, int invariant) |
ea0cb7ae | 1039 | { |
e5b19a8c | 1040 | if (invariant && alias_invariant && regno < alias_invariant_size) |
1a6c3add | 1041 | alias_invariant[regno] = val; |
1042 | ||
e5b19a8c | 1043 | if (regno >= VARRAY_SIZE (reg_base_value)) |
1044 | VARRAY_GROW (reg_base_value, max_reg_num ()); | |
1045 | ||
8ad4c111 | 1046 | if (REG_P (val)) |
ea0cb7ae | 1047 | { |
e5b19a8c | 1048 | VARRAY_RTX (reg_base_value, regno) |
1049 | = REG_BASE_VALUE (val); | |
ea0cb7ae | 1050 | return; |
1051 | } | |
e5b19a8c | 1052 | VARRAY_RTX (reg_base_value, regno) |
1053 | = find_base_value (val); | |
ea0cb7ae | 1054 | } |
1055 | ||
9f5e1a5a | 1056 | /* Clear alias info for a register. This is used if an RTL transformation |
1057 | changes the value of a register. This is used in flow by AUTO_INC_DEC | |
1058 | optimizations. We don't need to clear reg_base_value, since flow only | |
1059 | changes the offset. */ | |
1060 | ||
1061 | void | |
aecda0d6 | 1062 | clear_reg_alias_info (rtx reg) |
9f5e1a5a | 1063 | { |
26c8b134 | 1064 | unsigned int regno = REGNO (reg); |
1065 | ||
12cb06ad | 1066 | if (regno >= FIRST_PSEUDO_REGISTER) |
1067 | { | |
1068 | regno -= FIRST_PSEUDO_REGISTER; | |
1069 | if (regno < reg_known_value_size) | |
1070 | { | |
1071 | reg_known_value[regno] = reg; | |
1072 | reg_known_equiv_p[regno] = false; | |
1073 | } | |
1074 | } | |
1075 | } | |
1076 | ||
1077 | /* If a value is known for REGNO, return it. */ | |
1078 | ||
1079 | rtx | |
1080 | get_reg_known_value (unsigned int regno) | |
1081 | { | |
1082 | if (regno >= FIRST_PSEUDO_REGISTER) | |
1083 | { | |
1084 | regno -= FIRST_PSEUDO_REGISTER; | |
1085 | if (regno < reg_known_value_size) | |
1086 | return reg_known_value[regno]; | |
1087 | } | |
1088 | return NULL; | |
9f5e1a5a | 1089 | } |
1090 | ||
12cb06ad | 1091 | /* Set it. */ |
1092 | ||
1093 | static void | |
1094 | set_reg_known_value (unsigned int regno, rtx val) | |
1095 | { | |
1096 | if (regno >= FIRST_PSEUDO_REGISTER) | |
1097 | { | |
1098 | regno -= FIRST_PSEUDO_REGISTER; | |
1099 | if (regno < reg_known_value_size) | |
1100 | reg_known_value[regno] = val; | |
1101 | } | |
1102 | } | |
1103 | ||
1104 | /* Similarly for reg_known_equiv_p. */ | |
1105 | ||
1106 | bool | |
1107 | get_reg_known_equiv_p (unsigned int regno) | |
1108 | { | |
1109 | if (regno >= FIRST_PSEUDO_REGISTER) | |
1110 | { | |
1111 | regno -= FIRST_PSEUDO_REGISTER; | |
1112 | if (regno < reg_known_value_size) | |
1113 | return reg_known_equiv_p[regno]; | |
1114 | } | |
1115 | return false; | |
1116 | } | |
1117 | ||
1118 | static void | |
1119 | set_reg_known_equiv_p (unsigned int regno, bool val) | |
1120 | { | |
1121 | if (regno >= FIRST_PSEUDO_REGISTER) | |
1122 | { | |
1123 | regno -= FIRST_PSEUDO_REGISTER; | |
1124 | if (regno < reg_known_value_size) | |
1125 | reg_known_equiv_p[regno] = val; | |
1126 | } | |
1127 | } | |
1128 | ||
1129 | ||
7cfb9bcf | 1130 | /* Returns a canonical version of X, from the point of view alias |
1131 | analysis. (For example, if X is a MEM whose address is a register, | |
1132 | and the register has a known value (say a SYMBOL_REF), then a MEM | |
1133 | whose address is the SYMBOL_REF is returned.) */ | |
1134 | ||
1135 | rtx | |
aecda0d6 | 1136 | canon_rtx (rtx x) |
ea0cb7ae | 1137 | { |
1138 | /* Recursively look for equivalences. */ | |
8ad4c111 | 1139 | if (REG_P (x) && REGNO (x) >= FIRST_PSEUDO_REGISTER) |
12cb06ad | 1140 | { |
1141 | rtx t = get_reg_known_value (REGNO (x)); | |
1142 | if (t == x) | |
1143 | return x; | |
1144 | if (t) | |
1145 | return canon_rtx (t); | |
1146 | } | |
1147 | ||
1148 | if (GET_CODE (x) == PLUS) | |
ea0cb7ae | 1149 | { |
1150 | rtx x0 = canon_rtx (XEXP (x, 0)); | |
1151 | rtx x1 = canon_rtx (XEXP (x, 1)); | |
1152 | ||
1153 | if (x0 != XEXP (x, 0) || x1 != XEXP (x, 1)) | |
1154 | { | |
ea0cb7ae | 1155 | if (GET_CODE (x0) == CONST_INT) |
b244d4c7 | 1156 | return plus_constant (x1, INTVAL (x0)); |
ea0cb7ae | 1157 | else if (GET_CODE (x1) == CONST_INT) |
b244d4c7 | 1158 | return plus_constant (x0, INTVAL (x1)); |
941522d6 | 1159 | return gen_rtx_PLUS (GET_MODE (x), x0, x1); |
ea0cb7ae | 1160 | } |
1161 | } | |
083a2b5e | 1162 | |
ea0cb7ae | 1163 | /* This gives us much better alias analysis when called from |
1164 | the loop optimizer. Note we want to leave the original | |
1165 | MEM alone, but need to return the canonicalized MEM with | |
1166 | all the flags with their original values. */ | |
e16ceb8e | 1167 | else if (MEM_P (x)) |
e4e86ec5 | 1168 | x = replace_equiv_address_nv (x, canon_rtx (XEXP (x, 0))); |
083a2b5e | 1169 | |
ea0cb7ae | 1170 | return x; |
1171 | } | |
1172 | ||
1173 | /* Return 1 if X and Y are identical-looking rtx's. | |
6285298a | 1174 | Expect that X and Y has been already canonicalized. |
ea0cb7ae | 1175 | |
1176 | We use the data in reg_known_value above to see if two registers with | |
1177 | different numbers are, in fact, equivalent. */ | |
1178 | ||
1179 | static int | |
aecda0d6 | 1180 | rtx_equal_for_memref_p (rtx x, rtx y) |
ea0cb7ae | 1181 | { |
19cb6b50 | 1182 | int i; |
1183 | int j; | |
1184 | enum rtx_code code; | |
1185 | const char *fmt; | |
ea0cb7ae | 1186 | |
1187 | if (x == 0 && y == 0) | |
1188 | return 1; | |
1189 | if (x == 0 || y == 0) | |
1190 | return 0; | |
083a2b5e | 1191 | |
ea0cb7ae | 1192 | if (x == y) |
1193 | return 1; | |
1194 | ||
1195 | code = GET_CODE (x); | |
1196 | /* Rtx's of different codes cannot be equal. */ | |
1197 | if (code != GET_CODE (y)) | |
1198 | return 0; | |
1199 | ||
1200 | /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. | |
1201 | (REG:SI x) and (REG:HI x) are NOT equivalent. */ | |
1202 | ||
1203 | if (GET_MODE (x) != GET_MODE (y)) | |
1204 | return 0; | |
1205 | ||
7cfb9bcf | 1206 | /* Some RTL can be compared without a recursive examination. */ |
1207 | switch (code) | |
1208 | { | |
1209 | case REG: | |
1210 | return REGNO (x) == REGNO (y); | |
1211 | ||
1212 | case LABEL_REF: | |
1213 | return XEXP (x, 0) == XEXP (y, 0); | |
9342ee68 | 1214 | |
7cfb9bcf | 1215 | case SYMBOL_REF: |
1216 | return XSTR (x, 0) == XSTR (y, 0); | |
1217 | ||
a2e238f7 | 1218 | case VALUE: |
7cfb9bcf | 1219 | case CONST_INT: |
1220 | case CONST_DOUBLE: | |
1221 | /* There's no need to compare the contents of CONST_DOUBLEs or | |
1222 | CONST_INTs because pointer equality is a good enough | |
1223 | comparison for these nodes. */ | |
1224 | return 0; | |
1225 | ||
7cfb9bcf | 1226 | default: |
1227 | break; | |
1228 | } | |
ea0cb7ae | 1229 | |
6285298a | 1230 | /* canon_rtx knows how to handle plus. No need to canonicalize. */ |
1231 | if (code == PLUS) | |
ea0cb7ae | 1232 | return ((rtx_equal_for_memref_p (XEXP (x, 0), XEXP (y, 0)) |
1233 | && rtx_equal_for_memref_p (XEXP (x, 1), XEXP (y, 1))) | |
1234 | || (rtx_equal_for_memref_p (XEXP (x, 0), XEXP (y, 1)) | |
1235 | && rtx_equal_for_memref_p (XEXP (x, 1), XEXP (y, 0)))); | |
6285298a | 1236 | /* For commutative operations, the RTX match if the operand match in any |
1237 | order. Also handle the simple binary and unary cases without a loop. */ | |
6720e96c | 1238 | if (COMMUTATIVE_P (x)) |
6285298a | 1239 | { |
1240 | rtx xop0 = canon_rtx (XEXP (x, 0)); | |
1241 | rtx yop0 = canon_rtx (XEXP (y, 0)); | |
1242 | rtx yop1 = canon_rtx (XEXP (y, 1)); | |
1243 | ||
1244 | return ((rtx_equal_for_memref_p (xop0, yop0) | |
1245 | && rtx_equal_for_memref_p (canon_rtx (XEXP (x, 1)), yop1)) | |
1246 | || (rtx_equal_for_memref_p (xop0, yop1) | |
1247 | && rtx_equal_for_memref_p (canon_rtx (XEXP (x, 1)), yop0))); | |
1248 | } | |
6720e96c | 1249 | else if (NON_COMMUTATIVE_P (x)) |
6285298a | 1250 | { |
1251 | return (rtx_equal_for_memref_p (canon_rtx (XEXP (x, 0)), | |
aecda0d6 | 1252 | canon_rtx (XEXP (y, 0))) |
6285298a | 1253 | && rtx_equal_for_memref_p (canon_rtx (XEXP (x, 1)), |
1254 | canon_rtx (XEXP (y, 1)))); | |
1255 | } | |
6720e96c | 1256 | else if (UNARY_P (x)) |
6285298a | 1257 | return rtx_equal_for_memref_p (canon_rtx (XEXP (x, 0)), |
aecda0d6 | 1258 | canon_rtx (XEXP (y, 0))); |
ea0cb7ae | 1259 | |
1260 | /* Compare the elements. If any pair of corresponding elements | |
1a6c3add | 1261 | fail to match, return 0 for the whole things. |
1262 | ||
1263 | Limit cases to types which actually appear in addresses. */ | |
ea0cb7ae | 1264 | |
1265 | fmt = GET_RTX_FORMAT (code); | |
1266 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
1267 | { | |
1268 | switch (fmt[i]) | |
1269 | { | |
ea0cb7ae | 1270 | case 'i': |
1271 | if (XINT (x, i) != XINT (y, i)) | |
1272 | return 0; | |
1273 | break; | |
1274 | ||
ea0cb7ae | 1275 | case 'E': |
1276 | /* Two vectors must have the same length. */ | |
1277 | if (XVECLEN (x, i) != XVECLEN (y, i)) | |
1278 | return 0; | |
1279 | ||
1280 | /* And the corresponding elements must match. */ | |
1281 | for (j = 0; j < XVECLEN (x, i); j++) | |
6285298a | 1282 | if (rtx_equal_for_memref_p (canon_rtx (XVECEXP (x, i, j)), |
1283 | canon_rtx (XVECEXP (y, i, j))) == 0) | |
ea0cb7ae | 1284 | return 0; |
1285 | break; | |
1286 | ||
1287 | case 'e': | |
6285298a | 1288 | if (rtx_equal_for_memref_p (canon_rtx (XEXP (x, i)), |
1289 | canon_rtx (XEXP (y, i))) == 0) | |
ea0cb7ae | 1290 | return 0; |
1291 | break; | |
1292 | ||
83ec925c | 1293 | /* This can happen for asm operands. */ |
1294 | case 's': | |
1295 | if (strcmp (XSTR (x, i), XSTR (y, i))) | |
1296 | return 0; | |
1297 | break; | |
1298 | ||
23c04f6b | 1299 | /* This can happen for an asm which clobbers memory. */ |
1300 | case '0': | |
1301 | break; | |
1302 | ||
ea0cb7ae | 1303 | /* It is believed that rtx's at this level will never |
1304 | contain anything but integers and other rtx's, | |
1305 | except for within LABEL_REFs and SYMBOL_REFs. */ | |
1306 | default: | |
64db345d | 1307 | gcc_unreachable (); |
ea0cb7ae | 1308 | } |
1309 | } | |
1310 | return 1; | |
1311 | } | |
1312 | ||
1313 | /* Given an rtx X, find a SYMBOL_REF or LABEL_REF within | |
1314 | X and return it, or return 0 if none found. */ | |
1315 | ||
1316 | static rtx | |
aecda0d6 | 1317 | find_symbolic_term (rtx x) |
ea0cb7ae | 1318 | { |
19cb6b50 | 1319 | int i; |
1320 | enum rtx_code code; | |
1321 | const char *fmt; | |
ea0cb7ae | 1322 | |
1323 | code = GET_CODE (x); | |
1324 | if (code == SYMBOL_REF || code == LABEL_REF) | |
1325 | return x; | |
6720e96c | 1326 | if (OBJECT_P (x)) |
ea0cb7ae | 1327 | return 0; |
1328 | ||
1329 | fmt = GET_RTX_FORMAT (code); | |
1330 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
1331 | { | |
1332 | rtx t; | |
1333 | ||
1334 | if (fmt[i] == 'e') | |
1335 | { | |
1336 | t = find_symbolic_term (XEXP (x, i)); | |
1337 | if (t != 0) | |
1338 | return t; | |
1339 | } | |
1340 | else if (fmt[i] == 'E') | |
1341 | break; | |
1342 | } | |
1343 | return 0; | |
1344 | } | |
1345 | ||
5a37abfd | 1346 | rtx |
aecda0d6 | 1347 | find_base_term (rtx x) |
ea0cb7ae | 1348 | { |
1617c276 | 1349 | cselib_val *val; |
1350 | struct elt_loc_list *l; | |
1351 | ||
15750b1e | 1352 | #if defined (FIND_BASE_TERM) |
1353 | /* Try machine-dependent ways to find the base term. */ | |
1354 | x = FIND_BASE_TERM (x); | |
1355 | #endif | |
1356 | ||
ea0cb7ae | 1357 | switch (GET_CODE (x)) |
1358 | { | |
1359 | case REG: | |
1360 | return REG_BASE_VALUE (x); | |
1361 | ||
2be64ef0 | 1362 | case TRUNCATE: |
1363 | if (GET_MODE_SIZE (GET_MODE (x)) < GET_MODE_SIZE (Pmode)) | |
9342ee68 | 1364 | return 0; |
2be64ef0 | 1365 | /* Fall through. */ |
ea0cb7ae | 1366 | case HIGH: |
22dfec7d | 1367 | case PRE_INC: |
1368 | case PRE_DEC: | |
1369 | case POST_INC: | |
1370 | case POST_DEC: | |
2be64ef0 | 1371 | case PRE_MODIFY: |
1372 | case POST_MODIFY: | |
22dfec7d | 1373 | return find_base_term (XEXP (x, 0)); |
1374 | ||
5104ef69 | 1375 | case ZERO_EXTEND: |
1376 | case SIGN_EXTEND: /* Used for Alpha/NT pointers */ | |
1377 | { | |
1378 | rtx temp = find_base_term (XEXP (x, 0)); | |
1379 | ||
85d654dd | 1380 | if (temp != 0 && CONSTANT_P (temp)) |
5104ef69 | 1381 | temp = convert_memory_address (Pmode, temp); |
5104ef69 | 1382 | |
1383 | return temp; | |
1384 | } | |
1385 | ||
1617c276 | 1386 | case VALUE: |
1387 | val = CSELIB_VAL_PTR (x); | |
a2e238f7 | 1388 | if (!val) |
1389 | return 0; | |
1617c276 | 1390 | for (l = val->locs; l; l = l->next) |
1391 | if ((x = find_base_term (l->loc)) != 0) | |
1392 | return x; | |
1393 | return 0; | |
1394 | ||
ea0cb7ae | 1395 | case CONST: |
1396 | x = XEXP (x, 0); | |
1397 | if (GET_CODE (x) != PLUS && GET_CODE (x) != MINUS) | |
1398 | return 0; | |
b4b174c3 | 1399 | /* Fall through. */ |
ea0cb7ae | 1400 | case LO_SUM: |
1401 | case PLUS: | |
1402 | case MINUS: | |
1403 | { | |
1b5bbc4c | 1404 | rtx tmp1 = XEXP (x, 0); |
1405 | rtx tmp2 = XEXP (x, 1); | |
1406 | ||
35a3065a | 1407 | /* This is a little bit tricky since we have to determine which of |
1b5bbc4c | 1408 | the two operands represents the real base address. Otherwise this |
1409 | routine may return the index register instead of the base register. | |
1410 | ||
1411 | That may cause us to believe no aliasing was possible, when in | |
1412 | fact aliasing is possible. | |
1413 | ||
1414 | We use a few simple tests to guess the base register. Additional | |
1415 | tests can certainly be added. For example, if one of the operands | |
1416 | is a shift or multiply, then it must be the index register and the | |
1417 | other operand is the base register. */ | |
9342ee68 | 1418 | |
15750b1e | 1419 | if (tmp1 == pic_offset_table_rtx && CONSTANT_P (tmp2)) |
1420 | return find_base_term (tmp2); | |
1421 | ||
1b5bbc4c | 1422 | /* If either operand is known to be a pointer, then use it |
1423 | to determine the base term. */ | |
e61a0a7f | 1424 | if (REG_P (tmp1) && REG_POINTER (tmp1)) |
1b5bbc4c | 1425 | return find_base_term (tmp1); |
1426 | ||
e61a0a7f | 1427 | if (REG_P (tmp2) && REG_POINTER (tmp2)) |
1b5bbc4c | 1428 | return find_base_term (tmp2); |
1429 | ||
1430 | /* Neither operand was known to be a pointer. Go ahead and find the | |
1431 | base term for both operands. */ | |
1432 | tmp1 = find_base_term (tmp1); | |
1433 | tmp2 = find_base_term (tmp2); | |
1434 | ||
1435 | /* If either base term is named object or a special address | |
1436 | (like an argument or stack reference), then use it for the | |
1437 | base term. */ | |
083a2b5e | 1438 | if (tmp1 != 0 |
1b5bbc4c | 1439 | && (GET_CODE (tmp1) == SYMBOL_REF |
1440 | || GET_CODE (tmp1) == LABEL_REF | |
1441 | || (GET_CODE (tmp1) == ADDRESS | |
1442 | && GET_MODE (tmp1) != VOIDmode))) | |
1443 | return tmp1; | |
1444 | ||
083a2b5e | 1445 | if (tmp2 != 0 |
1b5bbc4c | 1446 | && (GET_CODE (tmp2) == SYMBOL_REF |
1447 | || GET_CODE (tmp2) == LABEL_REF | |
1448 | || (GET_CODE (tmp2) == ADDRESS | |
1449 | && GET_MODE (tmp2) != VOIDmode))) | |
1450 | return tmp2; | |
1451 | ||
1452 | /* We could not determine which of the two operands was the | |
1453 | base register and which was the index. So we can determine | |
1454 | nothing from the base alias check. */ | |
1455 | return 0; | |
ea0cb7ae | 1456 | } |
1457 | ||
1458 | case AND: | |
2be64ef0 | 1459 | if (GET_CODE (XEXP (x, 1)) == CONST_INT && INTVAL (XEXP (x, 1)) != 0) |
1460 | return find_base_term (XEXP (x, 0)); | |
ea0cb7ae | 1461 | return 0; |
1462 | ||
1463 | case SYMBOL_REF: | |
1464 | case LABEL_REF: | |
1465 | return x; | |
1466 | ||
1467 | default: | |
1468 | return 0; | |
1469 | } | |
1470 | } | |
1471 | ||
1472 | /* Return 0 if the addresses X and Y are known to point to different | |
1473 | objects, 1 if they might be pointers to the same object. */ | |
1474 | ||
1475 | static int | |
aecda0d6 | 1476 | base_alias_check (rtx x, rtx y, enum machine_mode x_mode, |
1477 | enum machine_mode y_mode) | |
ea0cb7ae | 1478 | { |
1479 | rtx x_base = find_base_term (x); | |
1480 | rtx y_base = find_base_term (y); | |
1481 | ||
3ad3a0ed | 1482 | /* If the address itself has no known base see if a known equivalent |
1483 | value has one. If either address still has no known base, nothing | |
1484 | is known about aliasing. */ | |
1485 | if (x_base == 0) | |
1486 | { | |
1487 | rtx x_c; | |
083a2b5e | 1488 | |
3ad3a0ed | 1489 | if (! flag_expensive_optimizations || (x_c = canon_rtx (x)) == x) |
1490 | return 1; | |
083a2b5e | 1491 | |
3ad3a0ed | 1492 | x_base = find_base_term (x_c); |
1493 | if (x_base == 0) | |
1494 | return 1; | |
1495 | } | |
ea0cb7ae | 1496 | |
3ad3a0ed | 1497 | if (y_base == 0) |
1498 | { | |
1499 | rtx y_c; | |
1500 | if (! flag_expensive_optimizations || (y_c = canon_rtx (y)) == y) | |
1501 | return 1; | |
083a2b5e | 1502 | |
3ad3a0ed | 1503 | y_base = find_base_term (y_c); |
1504 | if (y_base == 0) | |
1505 | return 1; | |
1506 | } | |
1507 | ||
1508 | /* If the base addresses are equal nothing is known about aliasing. */ | |
1509 | if (rtx_equal_p (x_base, y_base)) | |
ea0cb7ae | 1510 | return 1; |
1511 | ||
9342ee68 | 1512 | /* The base addresses of the read and write are different expressions. |
9fb1c31b | 1513 | If they are both symbols and they are not accessed via AND, there is |
1514 | no conflict. We can bring knowledge of object alignment into play | |
1515 | here. For example, on alpha, "char a, b;" can alias one another, | |
1516 | though "char a; long b;" cannot. */ | |
ea0cb7ae | 1517 | if (GET_CODE (x_base) != ADDRESS && GET_CODE (y_base) != ADDRESS) |
1e38bf49 | 1518 | { |
9fb1c31b | 1519 | if (GET_CODE (x) == AND && GET_CODE (y) == AND) |
1520 | return 1; | |
1521 | if (GET_CODE (x) == AND | |
1522 | && (GET_CODE (XEXP (x, 1)) != CONST_INT | |
8e6c5eac | 1523 | || (int) GET_MODE_UNIT_SIZE (y_mode) < -INTVAL (XEXP (x, 1)))) |
9fb1c31b | 1524 | return 1; |
1525 | if (GET_CODE (y) == AND | |
1526 | && (GET_CODE (XEXP (y, 1)) != CONST_INT | |
8e6c5eac | 1527 | || (int) GET_MODE_UNIT_SIZE (x_mode) < -INTVAL (XEXP (y, 1)))) |
9fb1c31b | 1528 | return 1; |
2b4e2cb9 | 1529 | /* Differing symbols never alias. */ |
1530 | return 0; | |
1e38bf49 | 1531 | } |
ea0cb7ae | 1532 | |
1533 | /* If one address is a stack reference there can be no alias: | |
1534 | stack references using different base registers do not alias, | |
1535 | a stack reference can not alias a parameter, and a stack reference | |
1536 | can not alias a global. */ | |
1537 | if ((GET_CODE (x_base) == ADDRESS && GET_MODE (x_base) == Pmode) | |
1538 | || (GET_CODE (y_base) == ADDRESS && GET_MODE (y_base) == Pmode)) | |
1539 | return 0; | |
1540 | ||
1541 | if (! flag_argument_noalias) | |
1542 | return 1; | |
1543 | ||
1544 | if (flag_argument_noalias > 1) | |
1545 | return 0; | |
1546 | ||
2c0e001b | 1547 | /* Weak noalias assertion (arguments are distinct, but may match globals). */ |
ea0cb7ae | 1548 | return ! (GET_MODE (x_base) == VOIDmode && GET_MODE (y_base) == VOIDmode); |
1549 | } | |
1550 | ||
1617c276 | 1551 | /* Convert the address X into something we can use. This is done by returning |
1552 | it unchanged unless it is a value; in the latter case we call cselib to get | |
1553 | a more useful rtx. */ | |
f7c44134 | 1554 | |
8e802be9 | 1555 | rtx |
aecda0d6 | 1556 | get_addr (rtx x) |
1617c276 | 1557 | { |
1558 | cselib_val *v; | |
1559 | struct elt_loc_list *l; | |
1560 | ||
1561 | if (GET_CODE (x) != VALUE) | |
1562 | return x; | |
1563 | v = CSELIB_VAL_PTR (x); | |
a2e238f7 | 1564 | if (v) |
1565 | { | |
1566 | for (l = v->locs; l; l = l->next) | |
1567 | if (CONSTANT_P (l->loc)) | |
1568 | return l->loc; | |
1569 | for (l = v->locs; l; l = l->next) | |
e16ceb8e | 1570 | if (!REG_P (l->loc) && !MEM_P (l->loc)) |
a2e238f7 | 1571 | return l->loc; |
1572 | if (v->locs) | |
1573 | return v->locs->loc; | |
1574 | } | |
1617c276 | 1575 | return x; |
1576 | } | |
1577 | ||
b9f92976 | 1578 | /* Return the address of the (N_REFS + 1)th memory reference to ADDR |
1579 | where SIZE is the size in bytes of the memory reference. If ADDR | |
1580 | is not modified by the memory reference then ADDR is returned. */ | |
1581 | ||
0aa9ffd4 | 1582 | static rtx |
aecda0d6 | 1583 | addr_side_effect_eval (rtx addr, int size, int n_refs) |
b9f92976 | 1584 | { |
1585 | int offset = 0; | |
9342ee68 | 1586 | |
b9f92976 | 1587 | switch (GET_CODE (addr)) |
1588 | { | |
1589 | case PRE_INC: | |
1590 | offset = (n_refs + 1) * size; | |
1591 | break; | |
1592 | case PRE_DEC: | |
1593 | offset = -(n_refs + 1) * size; | |
1594 | break; | |
1595 | case POST_INC: | |
1596 | offset = n_refs * size; | |
1597 | break; | |
1598 | case POST_DEC: | |
1599 | offset = -n_refs * size; | |
1600 | break; | |
1601 | ||
1602 | default: | |
1603 | return addr; | |
1604 | } | |
9342ee68 | 1605 | |
b9f92976 | 1606 | if (offset) |
6285298a | 1607 | addr = gen_rtx_PLUS (GET_MODE (addr), XEXP (addr, 0), |
1608 | GEN_INT (offset)); | |
b9f92976 | 1609 | else |
1610 | addr = XEXP (addr, 0); | |
6285298a | 1611 | addr = canon_rtx (addr); |
b9f92976 | 1612 | |
1613 | return addr; | |
1614 | } | |
1615 | ||
ea0cb7ae | 1616 | /* Return nonzero if X and Y (memory addresses) could reference the |
1617 | same location in memory. C is an offset accumulator. When | |
1618 | C is nonzero, we are testing aliases between X and Y + C. | |
1619 | XSIZE is the size in bytes of the X reference, | |
1620 | similarly YSIZE is the size in bytes for Y. | |
6285298a | 1621 | Expect that canon_rtx has been already called for X and Y. |
ea0cb7ae | 1622 | |
1623 | If XSIZE or YSIZE is zero, we do not know the amount of memory being | |
1624 | referenced (the reference was BLKmode), so make the most pessimistic | |
1625 | assumptions. | |
1626 | ||
1e38bf49 | 1627 | If XSIZE or YSIZE is negative, we may access memory outside the object |
1628 | being referenced as a side effect. This can happen when using AND to | |
1629 | align memory references, as is done on the Alpha. | |
1630 | ||
ea0cb7ae | 1631 | Nice to notice that varying addresses cannot conflict with fp if no |
4958fe6e | 1632 | local variables had their addresses taken, but that's too hard now. */ |
ea0cb7ae | 1633 | |
ea0cb7ae | 1634 | static int |
aecda0d6 | 1635 | memrefs_conflict_p (int xsize, rtx x, int ysize, rtx y, HOST_WIDE_INT c) |
ea0cb7ae | 1636 | { |
1617c276 | 1637 | if (GET_CODE (x) == VALUE) |
1638 | x = get_addr (x); | |
1639 | if (GET_CODE (y) == VALUE) | |
1640 | y = get_addr (y); | |
ea0cb7ae | 1641 | if (GET_CODE (x) == HIGH) |
1642 | x = XEXP (x, 0); | |
1643 | else if (GET_CODE (x) == LO_SUM) | |
1644 | x = XEXP (x, 1); | |
1645 | else | |
6285298a | 1646 | x = addr_side_effect_eval (x, xsize, 0); |
ea0cb7ae | 1647 | if (GET_CODE (y) == HIGH) |
1648 | y = XEXP (y, 0); | |
1649 | else if (GET_CODE (y) == LO_SUM) | |
1650 | y = XEXP (y, 1); | |
1651 | else | |
6285298a | 1652 | y = addr_side_effect_eval (y, ysize, 0); |
ea0cb7ae | 1653 | |
1654 | if (rtx_equal_for_memref_p (x, y)) | |
1655 | { | |
1e38bf49 | 1656 | if (xsize <= 0 || ysize <= 0) |
ea0cb7ae | 1657 | return 1; |
1658 | if (c >= 0 && xsize > c) | |
1659 | return 1; | |
1660 | if (c < 0 && ysize+c > 0) | |
1661 | return 1; | |
1662 | return 0; | |
1663 | } | |
1664 | ||
35d3eedd | 1665 | /* This code used to check for conflicts involving stack references and |
1666 | globals but the base address alias code now handles these cases. */ | |
ea0cb7ae | 1667 | |
1668 | if (GET_CODE (x) == PLUS) | |
1669 | { | |
1670 | /* The fact that X is canonicalized means that this | |
1671 | PLUS rtx is canonicalized. */ | |
1672 | rtx x0 = XEXP (x, 0); | |
1673 | rtx x1 = XEXP (x, 1); | |
1674 | ||
1675 | if (GET_CODE (y) == PLUS) | |
1676 | { | |
1677 | /* The fact that Y is canonicalized means that this | |
1678 | PLUS rtx is canonicalized. */ | |
1679 | rtx y0 = XEXP (y, 0); | |
1680 | rtx y1 = XEXP (y, 1); | |
1681 | ||
1682 | if (rtx_equal_for_memref_p (x1, y1)) | |
1683 | return memrefs_conflict_p (xsize, x0, ysize, y0, c); | |
1684 | if (rtx_equal_for_memref_p (x0, y0)) | |
1685 | return memrefs_conflict_p (xsize, x1, ysize, y1, c); | |
1686 | if (GET_CODE (x1) == CONST_INT) | |
5f59c49b | 1687 | { |
1688 | if (GET_CODE (y1) == CONST_INT) | |
1689 | return memrefs_conflict_p (xsize, x0, ysize, y0, | |
1690 | c - INTVAL (x1) + INTVAL (y1)); | |
1691 | else | |
1692 | return memrefs_conflict_p (xsize, x0, ysize, y, | |
1693 | c - INTVAL (x1)); | |
1694 | } | |
ea0cb7ae | 1695 | else if (GET_CODE (y1) == CONST_INT) |
1696 | return memrefs_conflict_p (xsize, x, ysize, y0, c + INTVAL (y1)); | |
1697 | ||
35d3eedd | 1698 | return 1; |
ea0cb7ae | 1699 | } |
1700 | else if (GET_CODE (x1) == CONST_INT) | |
1701 | return memrefs_conflict_p (xsize, x0, ysize, y, c - INTVAL (x1)); | |
1702 | } | |
1703 | else if (GET_CODE (y) == PLUS) | |
1704 | { | |
1705 | /* The fact that Y is canonicalized means that this | |
1706 | PLUS rtx is canonicalized. */ | |
1707 | rtx y0 = XEXP (y, 0); | |
1708 | rtx y1 = XEXP (y, 1); | |
1709 | ||
1710 | if (GET_CODE (y1) == CONST_INT) | |
1711 | return memrefs_conflict_p (xsize, x, ysize, y0, c + INTVAL (y1)); | |
1712 | else | |
1713 | return 1; | |
1714 | } | |
1715 | ||
1716 | if (GET_CODE (x) == GET_CODE (y)) | |
1717 | switch (GET_CODE (x)) | |
1718 | { | |
1719 | case MULT: | |
1720 | { | |
1721 | /* Handle cases where we expect the second operands to be the | |
1722 | same, and check only whether the first operand would conflict | |
1723 | or not. */ | |
1724 | rtx x0, y0; | |
1725 | rtx x1 = canon_rtx (XEXP (x, 1)); | |
1726 | rtx y1 = canon_rtx (XEXP (y, 1)); | |
1727 | if (! rtx_equal_for_memref_p (x1, y1)) | |
1728 | return 1; | |
1729 | x0 = canon_rtx (XEXP (x, 0)); | |
1730 | y0 = canon_rtx (XEXP (y, 0)); | |
1731 | if (rtx_equal_for_memref_p (x0, y0)) | |
1732 | return (xsize == 0 || ysize == 0 | |
1733 | || (c >= 0 && xsize > c) || (c < 0 && ysize+c > 0)); | |
1734 | ||
1735 | /* Can't properly adjust our sizes. */ | |
1736 | if (GET_CODE (x1) != CONST_INT) | |
1737 | return 1; | |
1738 | xsize /= INTVAL (x1); | |
1739 | ysize /= INTVAL (x1); | |
1740 | c /= INTVAL (x1); | |
1741 | return memrefs_conflict_p (xsize, x0, ysize, y0, c); | |
1742 | } | |
99c14947 | 1743 | |
1a6c3add | 1744 | case REG: |
1745 | /* Are these registers known not to be equal? */ | |
1746 | if (alias_invariant) | |
1747 | { | |
274c11d8 | 1748 | unsigned int r_x = REGNO (x), r_y = REGNO (y); |
1a6c3add | 1749 | rtx i_x, i_y; /* invariant relationships of X and Y */ |
1750 | ||
e5b19a8c | 1751 | i_x = r_x >= alias_invariant_size ? 0 : alias_invariant[r_x]; |
1752 | i_y = r_y >= alias_invariant_size ? 0 : alias_invariant[r_y]; | |
1a6c3add | 1753 | |
1754 | if (i_x == 0 && i_y == 0) | |
1755 | break; | |
1756 | ||
1757 | if (! memrefs_conflict_p (xsize, i_x ? i_x : x, | |
1758 | ysize, i_y ? i_y : y, c)) | |
1759 | return 0; | |
1760 | } | |
1761 | break; | |
1762 | ||
99c14947 | 1763 | default: |
1764 | break; | |
ea0cb7ae | 1765 | } |
1766 | ||
1767 | /* Treat an access through an AND (e.g. a subword access on an Alpha) | |
9342ee68 | 1768 | as an access with indeterminate size. Assume that references |
9fb1c31b | 1769 | besides AND are aligned, so if the size of the other reference is |
1770 | at least as large as the alignment, assume no other overlap. */ | |
ea0cb7ae | 1771 | if (GET_CODE (x) == AND && GET_CODE (XEXP (x, 1)) == CONST_INT) |
9fb1c31b | 1772 | { |
bcdb334e | 1773 | if (GET_CODE (y) == AND || ysize < -INTVAL (XEXP (x, 1))) |
9fb1c31b | 1774 | xsize = -1; |
6285298a | 1775 | return memrefs_conflict_p (xsize, canon_rtx (XEXP (x, 0)), ysize, y, c); |
9fb1c31b | 1776 | } |
ea0cb7ae | 1777 | if (GET_CODE (y) == AND && GET_CODE (XEXP (y, 1)) == CONST_INT) |
1e38bf49 | 1778 | { |
9fb1c31b | 1779 | /* ??? If we are indexing far enough into the array/structure, we |
9342ee68 | 1780 | may yet be able to determine that we can not overlap. But we |
1e38bf49 | 1781 | also need to that we are far enough from the end not to overlap |
9fb1c31b | 1782 | a following reference, so we do nothing with that for now. */ |
bcdb334e | 1783 | if (GET_CODE (x) == AND || xsize < -INTVAL (XEXP (y, 1))) |
9fb1c31b | 1784 | ysize = -1; |
6285298a | 1785 | return memrefs_conflict_p (xsize, x, ysize, canon_rtx (XEXP (y, 0)), c); |
1e38bf49 | 1786 | } |
ea0cb7ae | 1787 | |
1788 | if (CONSTANT_P (x)) | |
1789 | { | |
1790 | if (GET_CODE (x) == CONST_INT && GET_CODE (y) == CONST_INT) | |
1791 | { | |
1792 | c += (INTVAL (y) - INTVAL (x)); | |
1e38bf49 | 1793 | return (xsize <= 0 || ysize <= 0 |
ea0cb7ae | 1794 | || (c >= 0 && xsize > c) || (c < 0 && ysize+c > 0)); |
1795 | } | |
1796 | ||
1797 | if (GET_CODE (x) == CONST) | |
1798 | { | |
1799 | if (GET_CODE (y) == CONST) | |
1800 | return memrefs_conflict_p (xsize, canon_rtx (XEXP (x, 0)), | |
1801 | ysize, canon_rtx (XEXP (y, 0)), c); | |
1802 | else | |
1803 | return memrefs_conflict_p (xsize, canon_rtx (XEXP (x, 0)), | |
1804 | ysize, y, c); | |
1805 | } | |
1806 | if (GET_CODE (y) == CONST) | |
1807 | return memrefs_conflict_p (xsize, x, ysize, | |
1808 | canon_rtx (XEXP (y, 0)), c); | |
1809 | ||
1810 | if (CONSTANT_P (y)) | |
15750b1e | 1811 | return (xsize <= 0 || ysize <= 0 |
1e38bf49 | 1812 | || (rtx_equal_for_memref_p (x, y) |
15750b1e | 1813 | && ((c >= 0 && xsize > c) || (c < 0 && ysize+c > 0)))); |
ea0cb7ae | 1814 | |
1815 | return 1; | |
1816 | } | |
1817 | return 1; | |
1818 | } | |
1819 | ||
1820 | /* Functions to compute memory dependencies. | |
1821 | ||
1822 | Since we process the insns in execution order, we can build tables | |
1823 | to keep track of what registers are fixed (and not aliased), what registers | |
1824 | are varying in known ways, and what registers are varying in unknown | |
1825 | ways. | |
1826 | ||
1827 | If both memory references are volatile, then there must always be a | |
1828 | dependence between the two references, since their order can not be | |
1829 | changed. A volatile and non-volatile reference can be interchanged | |
9342ee68 | 1830 | though. |
ea0cb7ae | 1831 | |
71d9e995 | 1832 | A MEM_IN_STRUCT reference at a non-AND varying address can never |
1833 | conflict with a non-MEM_IN_STRUCT reference at a fixed address. We | |
1834 | also must allow AND addresses, because they may generate accesses | |
1835 | outside the object being referenced. This is used to generate | |
1836 | aligned addresses from unaligned addresses, for instance, the alpha | |
1837 | storeqi_unaligned pattern. */ | |
ea0cb7ae | 1838 | |
1839 | /* Read dependence: X is read after read in MEM takes place. There can | |
1840 | only be a dependence here if both reads are volatile. */ | |
1841 | ||
1842 | int | |
aecda0d6 | 1843 | read_dependence (rtx mem, rtx x) |
ea0cb7ae | 1844 | { |
1845 | return MEM_VOLATILE_P (x) && MEM_VOLATILE_P (mem); | |
1846 | } | |
1847 | ||
6a0934dd | 1848 | /* Returns MEM1 if and only if MEM1 is a scalar at a fixed address and |
1849 | MEM2 is a reference to a structure at a varying address, or returns | |
1850 | MEM2 if vice versa. Otherwise, returns NULL_RTX. If a non-NULL | |
1851 | value is returned MEM1 and MEM2 can never alias. VARIES_P is used | |
1852 | to decide whether or not an address may vary; it should return | |
1617c276 | 1853 | nonzero whenever variation is possible. |
1854 | MEM1_ADDR and MEM2_ADDR are the addresses of MEM1 and MEM2. */ | |
9342ee68 | 1855 | |
fe69e4fe | 1856 | static rtx |
aecda0d6 | 1857 | fixed_scalar_and_varying_struct_p (rtx mem1, rtx mem2, rtx mem1_addr, |
1858 | rtx mem2_addr, | |
1859 | int (*varies_p) (rtx, int)) | |
9342ee68 | 1860 | { |
c7f6877a | 1861 | if (! flag_strict_aliasing) |
1862 | return NULL_RTX; | |
1863 | ||
9342ee68 | 1864 | if (MEM_SCALAR_P (mem1) && MEM_IN_STRUCT_P (mem2) |
ea087693 | 1865 | && !varies_p (mem1_addr, 1) && varies_p (mem2_addr, 1)) |
6a0934dd | 1866 | /* MEM1 is a scalar at a fixed address; MEM2 is a struct at a |
1867 | varying address. */ | |
1868 | return mem1; | |
1869 | ||
9342ee68 | 1870 | if (MEM_IN_STRUCT_P (mem1) && MEM_SCALAR_P (mem2) |
ea087693 | 1871 | && varies_p (mem1_addr, 1) && !varies_p (mem2_addr, 1)) |
6a0934dd | 1872 | /* MEM2 is a scalar at a fixed address; MEM1 is a struct at a |
1873 | varying address. */ | |
1874 | return mem2; | |
1875 | ||
1876 | return NULL_RTX; | |
1877 | } | |
1878 | ||
1879 | /* Returns nonzero if something about the mode or address format MEM1 | |
1880 | indicates that it might well alias *anything*. */ | |
1881 | ||
fe69e4fe | 1882 | static int |
aecda0d6 | 1883 | aliases_everything_p (rtx mem) |
6a0934dd | 1884 | { |
6a0934dd | 1885 | if (GET_CODE (XEXP (mem, 0)) == AND) |
0975351b | 1886 | /* If the address is an AND, it's very hard to know at what it is |
6a0934dd | 1887 | actually pointing. */ |
1888 | return 1; | |
9342ee68 | 1889 | |
6a0934dd | 1890 | return 0; |
1891 | } | |
1892 | ||
b10dbbca | 1893 | /* Return true if we can determine that the fields referenced cannot |
1894 | overlap for any pair of objects. */ | |
1895 | ||
1896 | static bool | |
aecda0d6 | 1897 | nonoverlapping_component_refs_p (tree x, tree y) |
b10dbbca | 1898 | { |
1899 | tree fieldx, fieldy, typex, typey, orig_y; | |
1900 | ||
1901 | do | |
1902 | { | |
1903 | /* The comparison has to be done at a common type, since we don't | |
cb0ccc1e | 1904 | know how the inheritance hierarchy works. */ |
b10dbbca | 1905 | orig_y = y; |
1906 | do | |
1907 | { | |
1908 | fieldx = TREE_OPERAND (x, 1); | |
1909 | typex = DECL_FIELD_CONTEXT (fieldx); | |
1910 | ||
1911 | y = orig_y; | |
1912 | do | |
1913 | { | |
1914 | fieldy = TREE_OPERAND (y, 1); | |
1915 | typey = DECL_FIELD_CONTEXT (fieldy); | |
1916 | ||
1917 | if (typex == typey) | |
1918 | goto found; | |
1919 | ||
1920 | y = TREE_OPERAND (y, 0); | |
1921 | } | |
1922 | while (y && TREE_CODE (y) == COMPONENT_REF); | |
1923 | ||
1924 | x = TREE_OPERAND (x, 0); | |
1925 | } | |
1926 | while (x && TREE_CODE (x) == COMPONENT_REF); | |
1927 | ||
1928 | /* Never found a common type. */ | |
1929 | return false; | |
1930 | ||
1931 | found: | |
1932 | /* If we're left with accessing different fields of a structure, | |
1933 | then no overlap. */ | |
1934 | if (TREE_CODE (typex) == RECORD_TYPE | |
1935 | && fieldx != fieldy) | |
1936 | return true; | |
1937 | ||
1938 | /* The comparison on the current field failed. If we're accessing | |
1939 | a very nested structure, look at the next outer level. */ | |
1940 | x = TREE_OPERAND (x, 0); | |
1941 | y = TREE_OPERAND (y, 0); | |
1942 | } | |
1943 | while (x && y | |
1944 | && TREE_CODE (x) == COMPONENT_REF | |
1945 | && TREE_CODE (y) == COMPONENT_REF); | |
9342ee68 | 1946 | |
b10dbbca | 1947 | return false; |
1948 | } | |
1949 | ||
1950 | /* Look at the bottom of the COMPONENT_REF list for a DECL, and return it. */ | |
1951 | ||
1952 | static tree | |
aecda0d6 | 1953 | decl_for_component_ref (tree x) |
b10dbbca | 1954 | { |
1955 | do | |
1956 | { | |
1957 | x = TREE_OPERAND (x, 0); | |
1958 | } | |
1959 | while (x && TREE_CODE (x) == COMPONENT_REF); | |
1960 | ||
1961 | return x && DECL_P (x) ? x : NULL_TREE; | |
1962 | } | |
1963 | ||
1964 | /* Walk up the COMPONENT_REF list and adjust OFFSET to compensate for the | |
1965 | offset of the field reference. */ | |
1966 | ||
1967 | static rtx | |
aecda0d6 | 1968 | adjust_offset_for_component_ref (tree x, rtx offset) |
b10dbbca | 1969 | { |
1970 | HOST_WIDE_INT ioffset; | |
1971 | ||
1972 | if (! offset) | |
1973 | return NULL_RTX; | |
1974 | ||
1975 | ioffset = INTVAL (offset); | |
9342ee68 | 1976 | do |
b10dbbca | 1977 | { |
6374121b | 1978 | tree offset = component_ref_field_offset (x); |
b10dbbca | 1979 | tree field = TREE_OPERAND (x, 1); |
1980 | ||
6374121b | 1981 | if (! host_integerp (offset, 1)) |
b10dbbca | 1982 | return NULL_RTX; |
6374121b | 1983 | ioffset += (tree_low_cst (offset, 1) |
b10dbbca | 1984 | + (tree_low_cst (DECL_FIELD_BIT_OFFSET (field), 1) |
1985 | / BITS_PER_UNIT)); | |
1986 | ||
1987 | x = TREE_OPERAND (x, 0); | |
1988 | } | |
1989 | while (x && TREE_CODE (x) == COMPONENT_REF); | |
1990 | ||
1991 | return GEN_INT (ioffset); | |
1992 | } | |
1993 | ||
d716ce75 | 1994 | /* Return nonzero if we can determine the exprs corresponding to memrefs |
ddb3d16f | 1995 | X and Y and they do not overlap. */ |
1996 | ||
1997 | static int | |
aecda0d6 | 1998 | nonoverlapping_memrefs_p (rtx x, rtx y) |
ddb3d16f | 1999 | { |
b10dbbca | 2000 | tree exprx = MEM_EXPR (x), expry = MEM_EXPR (y); |
ddb3d16f | 2001 | rtx rtlx, rtly; |
2002 | rtx basex, basey; | |
b10dbbca | 2003 | rtx moffsetx, moffsety; |
ddb3d16f | 2004 | HOST_WIDE_INT offsetx = 0, offsety = 0, sizex, sizey, tem; |
2005 | ||
b10dbbca | 2006 | /* Unless both have exprs, we can't tell anything. */ |
2007 | if (exprx == 0 || expry == 0) | |
2008 | return 0; | |
2009 | ||
2010 | /* If both are field references, we may be able to determine something. */ | |
2011 | if (TREE_CODE (exprx) == COMPONENT_REF | |
2012 | && TREE_CODE (expry) == COMPONENT_REF | |
2013 | && nonoverlapping_component_refs_p (exprx, expry)) | |
2014 | return 1; | |
2015 | ||
2016 | /* If the field reference test failed, look at the DECLs involved. */ | |
2017 | moffsetx = MEM_OFFSET (x); | |
2018 | if (TREE_CODE (exprx) == COMPONENT_REF) | |
2019 | { | |
2020 | tree t = decl_for_component_ref (exprx); | |
2021 | if (! t) | |
2022 | return 0; | |
2023 | moffsetx = adjust_offset_for_component_ref (exprx, moffsetx); | |
2024 | exprx = t; | |
2025 | } | |
2a448a75 | 2026 | else if (INDIRECT_REF_P (exprx)) |
2d8fe5d0 | 2027 | { |
2028 | exprx = TREE_OPERAND (exprx, 0); | |
2029 | if (flag_argument_noalias < 2 | |
2030 | || TREE_CODE (exprx) != PARM_DECL) | |
2031 | return 0; | |
2032 | } | |
2033 | ||
b10dbbca | 2034 | moffsety = MEM_OFFSET (y); |
2035 | if (TREE_CODE (expry) == COMPONENT_REF) | |
2036 | { | |
2037 | tree t = decl_for_component_ref (expry); | |
2038 | if (! t) | |
2039 | return 0; | |
2040 | moffsety = adjust_offset_for_component_ref (expry, moffsety); | |
2041 | expry = t; | |
2042 | } | |
2a448a75 | 2043 | else if (INDIRECT_REF_P (expry)) |
2d8fe5d0 | 2044 | { |
2045 | expry = TREE_OPERAND (expry, 0); | |
2046 | if (flag_argument_noalias < 2 | |
2047 | || TREE_CODE (expry) != PARM_DECL) | |
2048 | return 0; | |
2049 | } | |
b10dbbca | 2050 | |
2051 | if (! DECL_P (exprx) || ! DECL_P (expry)) | |
ddb3d16f | 2052 | return 0; |
2053 | ||
b10dbbca | 2054 | rtlx = DECL_RTL (exprx); |
2055 | rtly = DECL_RTL (expry); | |
ddb3d16f | 2056 | |
10e949c1 | 2057 | /* If either RTL is not a MEM, it must be a REG or CONCAT, meaning they |
2058 | can't overlap unless they are the same because we never reuse that part | |
2059 | of the stack frame used for locals for spilled pseudos. */ | |
e16ceb8e | 2060 | if ((!MEM_P (rtlx) || !MEM_P (rtly)) |
10e949c1 | 2061 | && ! rtx_equal_p (rtlx, rtly)) |
ddb3d16f | 2062 | return 1; |
2063 | ||
2064 | /* Get the base and offsets of both decls. If either is a register, we | |
2065 | know both are and are the same, so use that as the base. The only | |
2066 | we can avoid overlap is if we can deduce that they are nonoverlapping | |
2067 | pieces of that decl, which is very rare. */ | |
e16ceb8e | 2068 | basex = MEM_P (rtlx) ? XEXP (rtlx, 0) : rtlx; |
ddb3d16f | 2069 | if (GET_CODE (basex) == PLUS && GET_CODE (XEXP (basex, 1)) == CONST_INT) |
2070 | offsetx = INTVAL (XEXP (basex, 1)), basex = XEXP (basex, 0); | |
2071 | ||
e16ceb8e | 2072 | basey = MEM_P (rtly) ? XEXP (rtly, 0) : rtly; |
ddb3d16f | 2073 | if (GET_CODE (basey) == PLUS && GET_CODE (XEXP (basey, 1)) == CONST_INT) |
2074 | offsety = INTVAL (XEXP (basey, 1)), basey = XEXP (basey, 0); | |
2075 | ||
c9c48480 | 2076 | /* If the bases are different, we know they do not overlap if both |
9342ee68 | 2077 | are constants or if one is a constant and the other a pointer into the |
c9c48480 | 2078 | stack frame. Otherwise a different base means we can't tell if they |
2079 | overlap or not. */ | |
2080 | if (! rtx_equal_p (basex, basey)) | |
9342ee68 | 2081 | return ((CONSTANT_P (basex) && CONSTANT_P (basey)) |
2082 | || (CONSTANT_P (basex) && REG_P (basey) | |
2083 | && REGNO_PTR_FRAME_P (REGNO (basey))) | |
2084 | || (CONSTANT_P (basey) && REG_P (basex) | |
2085 | && REGNO_PTR_FRAME_P (REGNO (basex)))); | |
ddb3d16f | 2086 | |
e16ceb8e | 2087 | sizex = (!MEM_P (rtlx) ? (int) GET_MODE_SIZE (GET_MODE (rtlx)) |
ddb3d16f | 2088 | : MEM_SIZE (rtlx) ? INTVAL (MEM_SIZE (rtlx)) |
2089 | : -1); | |
e16ceb8e | 2090 | sizey = (!MEM_P (rtly) ? (int) GET_MODE_SIZE (GET_MODE (rtly)) |
ddb3d16f | 2091 | : MEM_SIZE (rtly) ? INTVAL (MEM_SIZE (rtly)) : |
2092 | -1); | |
2093 | ||
dc2dfdb6 | 2094 | /* If we have an offset for either memref, it can update the values computed |
2095 | above. */ | |
b10dbbca | 2096 | if (moffsetx) |
2097 | offsetx += INTVAL (moffsetx), sizex -= INTVAL (moffsetx); | |
2098 | if (moffsety) | |
2099 | offsety += INTVAL (moffsety), sizey -= INTVAL (moffsety); | |
ddb3d16f | 2100 | |
dc2dfdb6 | 2101 | /* If a memref has both a size and an offset, we can use the smaller size. |
203c9fee | 2102 | We can't do this if the offset isn't known because we must view this |
dc2dfdb6 | 2103 | memref as being anywhere inside the DECL's MEM. */ |
b10dbbca | 2104 | if (MEM_SIZE (x) && moffsetx) |
ddb3d16f | 2105 | sizex = INTVAL (MEM_SIZE (x)); |
b10dbbca | 2106 | if (MEM_SIZE (y) && moffsety) |
ddb3d16f | 2107 | sizey = INTVAL (MEM_SIZE (y)); |
2108 | ||
2109 | /* Put the values of the memref with the lower offset in X's values. */ | |
2110 | if (offsetx > offsety) | |
2111 | { | |
2112 | tem = offsetx, offsetx = offsety, offsety = tem; | |
2113 | tem = sizex, sizex = sizey, sizey = tem; | |
2114 | } | |
2115 | ||
2116 | /* If we don't know the size of the lower-offset value, we can't tell | |
2117 | if they conflict. Otherwise, we do the test. */ | |
906923ba | 2118 | return sizex >= 0 && offsety >= offsetx + sizex; |
ddb3d16f | 2119 | } |
2120 | ||
ea0cb7ae | 2121 | /* True dependence: X is read after store in MEM takes place. */ |
2122 | ||
2123 | int | |
aecda0d6 | 2124 | true_dependence (rtx mem, enum machine_mode mem_mode, rtx x, |
2125 | int (*varies) (rtx, int)) | |
ea0cb7ae | 2126 | { |
19cb6b50 | 2127 | rtx x_addr, mem_addr; |
9ee7cac5 | 2128 | rtx base; |
ea0cb7ae | 2129 | |
2130 | if (MEM_VOLATILE_P (x) && MEM_VOLATILE_P (mem)) | |
2131 | return 1; | |
2132 | ||
a57f4c4d | 2133 | /* (mem:BLK (scratch)) is a special mechanism to conflict with everything. |
2134 | This is used in epilogue deallocation functions. */ | |
2135 | if (GET_MODE (x) == BLKmode && GET_CODE (XEXP (x, 0)) == SCRATCH) | |
2136 | return 1; | |
2137 | if (GET_MODE (mem) == BLKmode && GET_CODE (XEXP (mem, 0)) == SCRATCH) | |
2138 | return 1; | |
2139 | ||
b5ba9f3a | 2140 | if (DIFFERENT_ALIAS_SETS_P (x, mem)) |
2141 | return 0; | |
2142 | ||
b04fab2a | 2143 | /* Read-only memory is by definition never modified, and therefore can't |
2144 | conflict with anything. We don't expect to find read-only set on MEM, | |
a53ff4c1 | 2145 | but stupid user tricks can produce them, so don't die. */ |
b04fab2a | 2146 | if (MEM_READONLY_P (x)) |
ea0cb7ae | 2147 | return 0; |
2148 | ||
ddb3d16f | 2149 | if (nonoverlapping_memrefs_p (mem, x)) |
2150 | return 0; | |
2151 | ||
9fb1c31b | 2152 | if (mem_mode == VOIDmode) |
2153 | mem_mode = GET_MODE (mem); | |
2154 | ||
1617c276 | 2155 | x_addr = get_addr (XEXP (x, 0)); |
2156 | mem_addr = get_addr (XEXP (mem, 0)); | |
2157 | ||
002fe3cb | 2158 | base = find_base_term (x_addr); |
2159 | if (base && (GET_CODE (base) == LABEL_REF | |
2160 | || (GET_CODE (base) == SYMBOL_REF | |
2161 | && CONSTANT_POOL_ADDRESS_P (base)))) | |
2162 | return 0; | |
2163 | ||
1617c276 | 2164 | if (! base_alias_check (x_addr, mem_addr, GET_MODE (x), mem_mode)) |
3ad3a0ed | 2165 | return 0; |
2166 | ||
1617c276 | 2167 | x_addr = canon_rtx (x_addr); |
2168 | mem_addr = canon_rtx (mem_addr); | |
35d3eedd | 2169 | |
4958fe6e | 2170 | if (! memrefs_conflict_p (GET_MODE_SIZE (mem_mode), mem_addr, |
2171 | SIZE_FOR_MODE (x), x_addr, 0)) | |
2172 | return 0; | |
2173 | ||
6a0934dd | 2174 | if (aliases_everything_p (x)) |
4958fe6e | 2175 | return 1; |
2176 | ||
35a3065a | 2177 | /* We cannot use aliases_everything_p to test MEM, since we must look |
6a0934dd | 2178 | at MEM_MODE, rather than GET_MODE (MEM). */ |
2179 | if (mem_mode == QImode || GET_CODE (mem_addr) == AND) | |
8e802be9 | 2180 | return 1; |
2181 | ||
2182 | /* In true_dependence we also allow BLKmode to alias anything. Why | |
2183 | don't we do this in anti_dependence and output_dependence? */ | |
2184 | if (mem_mode == BLKmode || GET_MODE (x) == BLKmode) | |
2185 | return 1; | |
2186 | ||
2187 | return ! fixed_scalar_and_varying_struct_p (mem, x, mem_addr, x_addr, | |
2188 | varies); | |
2189 | } | |
2190 | ||
2191 | /* Canonical true dependence: X is read after store in MEM takes place. | |
9342ee68 | 2192 | Variant of true_dependence which assumes MEM has already been |
2193 | canonicalized (hence we no longer do that here). | |
2194 | The mem_addr argument has been added, since true_dependence computed | |
8e802be9 | 2195 | this value prior to canonicalizing. */ |
2196 | ||
2197 | int | |
aecda0d6 | 2198 | canon_true_dependence (rtx mem, enum machine_mode mem_mode, rtx mem_addr, |
2199 | rtx x, int (*varies) (rtx, int)) | |
8e802be9 | 2200 | { |
19cb6b50 | 2201 | rtx x_addr; |
8e802be9 | 2202 | |
2203 | if (MEM_VOLATILE_P (x) && MEM_VOLATILE_P (mem)) | |
2204 | return 1; | |
2205 | ||
d7b592b0 | 2206 | /* (mem:BLK (scratch)) is a special mechanism to conflict with everything. |
2207 | This is used in epilogue deallocation functions. */ | |
2208 | if (GET_MODE (x) == BLKmode && GET_CODE (XEXP (x, 0)) == SCRATCH) | |
2209 | return 1; | |
2210 | if (GET_MODE (mem) == BLKmode && GET_CODE (XEXP (mem, 0)) == SCRATCH) | |
2211 | return 1; | |
2212 | ||
8e802be9 | 2213 | if (DIFFERENT_ALIAS_SETS_P (x, mem)) |
2214 | return 0; | |
2215 | ||
b04fab2a | 2216 | /* Read-only memory is by definition never modified, and therefore can't |
2217 | conflict with anything. We don't expect to find read-only set on MEM, | |
a53ff4c1 | 2218 | but stupid user tricks can produce them, so don't die. */ |
b04fab2a | 2219 | if (MEM_READONLY_P (x)) |
8e802be9 | 2220 | return 0; |
2221 | ||
ddb3d16f | 2222 | if (nonoverlapping_memrefs_p (x, mem)) |
2223 | return 0; | |
2224 | ||
8e802be9 | 2225 | x_addr = get_addr (XEXP (x, 0)); |
2226 | ||
2227 | if (! base_alias_check (x_addr, mem_addr, GET_MODE (x), mem_mode)) | |
2228 | return 0; | |
2229 | ||
2230 | x_addr = canon_rtx (x_addr); | |
2231 | if (! memrefs_conflict_p (GET_MODE_SIZE (mem_mode), mem_addr, | |
2232 | SIZE_FOR_MODE (x), x_addr, 0)) | |
2233 | return 0; | |
2234 | ||
2235 | if (aliases_everything_p (x)) | |
2236 | return 1; | |
2237 | ||
35a3065a | 2238 | /* We cannot use aliases_everything_p to test MEM, since we must look |
8e802be9 | 2239 | at MEM_MODE, rather than GET_MODE (MEM). */ |
2240 | if (mem_mode == QImode || GET_CODE (mem_addr) == AND) | |
6a0934dd | 2241 | return 1; |
4958fe6e | 2242 | |
6a0934dd | 2243 | /* In true_dependence we also allow BLKmode to alias anything. Why |
2244 | don't we do this in anti_dependence and output_dependence? */ | |
2245 | if (mem_mode == BLKmode || GET_MODE (x) == BLKmode) | |
2246 | return 1; | |
4958fe6e | 2247 | |
1617c276 | 2248 | return ! fixed_scalar_and_varying_struct_p (mem, x, mem_addr, x_addr, |
2249 | varies); | |
ea0cb7ae | 2250 | } |
2251 | ||
d10cfa8d | 2252 | /* Returns nonzero if a write to X might alias a previous read from |
b04fab2a | 2253 | (or, if WRITEP is nonzero, a write to) MEM. */ |
ea0cb7ae | 2254 | |
fe69e4fe | 2255 | static int |
b04fab2a | 2256 | write_dependence_p (rtx mem, rtx x, int writep) |
ea0cb7ae | 2257 | { |
35d3eedd | 2258 | rtx x_addr, mem_addr; |
6a0934dd | 2259 | rtx fixed_scalar; |
9ee7cac5 | 2260 | rtx base; |
35d3eedd | 2261 | |
ea0cb7ae | 2262 | if (MEM_VOLATILE_P (x) && MEM_VOLATILE_P (mem)) |
2263 | return 1; | |
2264 | ||
a57f4c4d | 2265 | /* (mem:BLK (scratch)) is a special mechanism to conflict with everything. |
2266 | This is used in epilogue deallocation functions. */ | |
2267 | if (GET_MODE (x) == BLKmode && GET_CODE (XEXP (x, 0)) == SCRATCH) | |
2268 | return 1; | |
2269 | if (GET_MODE (mem) == BLKmode && GET_CODE (XEXP (mem, 0)) == SCRATCH) | |
2270 | return 1; | |
2271 | ||
1617c276 | 2272 | if (DIFFERENT_ALIAS_SETS_P (x, mem)) |
2273 | return 0; | |
2274 | ||
b04fab2a | 2275 | /* A read from read-only memory can't conflict with read-write memory. */ |
2276 | if (!writep && MEM_READONLY_P (mem)) | |
2277 | return 0; | |
002fe3cb | 2278 | |
ddb3d16f | 2279 | if (nonoverlapping_memrefs_p (x, mem)) |
2280 | return 0; | |
2281 | ||
002fe3cb | 2282 | x_addr = get_addr (XEXP (x, 0)); |
2283 | mem_addr = get_addr (XEXP (mem, 0)); | |
2284 | ||
9ee7cac5 | 2285 | if (! writep) |
2286 | { | |
002fe3cb | 2287 | base = find_base_term (mem_addr); |
9ee7cac5 | 2288 | if (base && (GET_CODE (base) == LABEL_REF |
2289 | || (GET_CODE (base) == SYMBOL_REF | |
2290 | && CONSTANT_POOL_ADDRESS_P (base)))) | |
2291 | return 0; | |
2292 | } | |
2293 | ||
1617c276 | 2294 | if (! base_alias_check (x_addr, mem_addr, GET_MODE (x), |
2295 | GET_MODE (mem))) | |
b5ba9f3a | 2296 | return 0; |
2297 | ||
1617c276 | 2298 | x_addr = canon_rtx (x_addr); |
2299 | mem_addr = canon_rtx (mem_addr); | |
35d3eedd | 2300 | |
6a0934dd | 2301 | if (!memrefs_conflict_p (SIZE_FOR_MODE (mem), mem_addr, |
2302 | SIZE_FOR_MODE (x), x_addr, 0)) | |
2303 | return 0; | |
2304 | ||
9342ee68 | 2305 | fixed_scalar |
1617c276 | 2306 | = fixed_scalar_and_varying_struct_p (mem, x, mem_addr, x_addr, |
2307 | rtx_addr_varies_p); | |
2308 | ||
6a0934dd | 2309 | return (!(fixed_scalar == mem && !aliases_everything_p (x)) |
2310 | && !(fixed_scalar == x && !aliases_everything_p (mem))); | |
2311 | } | |
2312 | ||
2313 | /* Anti dependence: X is written after read in MEM takes place. */ | |
2314 | ||
2315 | int | |
aecda0d6 | 2316 | anti_dependence (rtx mem, rtx x) |
6a0934dd | 2317 | { |
b04fab2a | 2318 | return write_dependence_p (mem, x, /*writep=*/0); |
ea0cb7ae | 2319 | } |
2320 | ||
2321 | /* Output dependence: X is written after store in MEM takes place. */ | |
2322 | ||
2323 | int | |
aecda0d6 | 2324 | output_dependence (rtx mem, rtx x) |
ea0cb7ae | 2325 | { |
b04fab2a | 2326 | return write_dependence_p (mem, x, /*writep=*/1); |
ea0cb7ae | 2327 | } |
c352cbc7 | 2328 | \f |
2329 | /* A subroutine of nonlocal_mentioned_p, returns 1 if *LOC mentions | |
2330 | something which is not local to the function and is not constant. */ | |
142131ed | 2331 | |
2332 | static int | |
aecda0d6 | 2333 | nonlocal_mentioned_p_1 (rtx *loc, void *data ATTRIBUTE_UNUSED) |
142131ed | 2334 | { |
c352cbc7 | 2335 | rtx x = *loc; |
142131ed | 2336 | rtx base; |
142131ed | 2337 | int regno; |
2338 | ||
c352cbc7 | 2339 | if (! x) |
2340 | return 0; | |
142131ed | 2341 | |
c352cbc7 | 2342 | switch (GET_CODE (x)) |
142131ed | 2343 | { |
2344 | case SUBREG: | |
8ad4c111 | 2345 | if (REG_P (SUBREG_REG (x))) |
142131ed | 2346 | { |
2347 | /* Global registers are not local. */ | |
2348 | if (REGNO (SUBREG_REG (x)) < FIRST_PSEUDO_REGISTER | |
701e46d0 | 2349 | && global_regs[subreg_regno (x)]) |
142131ed | 2350 | return 1; |
2351 | return 0; | |
2352 | } | |
2353 | break; | |
2354 | ||
2355 | case REG: | |
2356 | regno = REGNO (x); | |
2357 | /* Global registers are not local. */ | |
2358 | if (regno < FIRST_PSEUDO_REGISTER && global_regs[regno]) | |
2359 | return 1; | |
2360 | return 0; | |
2361 | ||
2362 | case SCRATCH: | |
2363 | case PC: | |
2364 | case CC0: | |
2365 | case CONST_INT: | |
2366 | case CONST_DOUBLE: | |
886cfd4f | 2367 | case CONST_VECTOR: |
142131ed | 2368 | case CONST: |
2369 | case LABEL_REF: | |
2370 | return 0; | |
2371 | ||
2372 | case SYMBOL_REF: | |
2373 | /* Constants in the function's constants pool are constant. */ | |
2374 | if (CONSTANT_POOL_ADDRESS_P (x)) | |
2375 | return 0; | |
2376 | return 1; | |
2377 | ||
2378 | case CALL: | |
1addbe73 | 2379 | /* Non-constant calls and recursion are not local. */ |
142131ed | 2380 | return 1; |
2381 | ||
2382 | case MEM: | |
2383 | /* Be overly conservative and consider any volatile memory | |
2384 | reference as not local. */ | |
2385 | if (MEM_VOLATILE_P (x)) | |
2386 | return 1; | |
2387 | base = find_base_term (XEXP (x, 0)); | |
2388 | if (base) | |
2389 | { | |
3a935f06 | 2390 | /* A Pmode ADDRESS could be a reference via the structure value |
2391 | address or static chain. Such memory references are nonlocal. | |
2392 | ||
2393 | Thus, we have to examine the contents of the ADDRESS to find | |
2394 | out if this is a local reference or not. */ | |
2395 | if (GET_CODE (base) == ADDRESS | |
2396 | && GET_MODE (base) == Pmode | |
2397 | && (XEXP (base, 0) == stack_pointer_rtx | |
2398 | || XEXP (base, 0) == arg_pointer_rtx | |
2399 | #if HARD_FRAME_POINTER_REGNUM != FRAME_POINTER_REGNUM | |
2400 | || XEXP (base, 0) == hard_frame_pointer_rtx | |
2401 | #endif | |
2402 | || XEXP (base, 0) == frame_pointer_rtx)) | |
142131ed | 2403 | return 0; |
2404 | /* Constants in the function's constant pool are constant. */ | |
2405 | if (GET_CODE (base) == SYMBOL_REF && CONSTANT_POOL_ADDRESS_P (base)) | |
2406 | return 0; | |
2407 | } | |
2408 | return 1; | |
2409 | ||
1addbe73 | 2410 | case UNSPEC_VOLATILE: |
142131ed | 2411 | case ASM_INPUT: |
142131ed | 2412 | return 1; |
2413 | ||
1addbe73 | 2414 | case ASM_OPERANDS: |
2415 | if (MEM_VOLATILE_P (x)) | |
2416 | return 1; | |
2417 | ||
d632b59a | 2418 | /* Fall through. */ |
1addbe73 | 2419 | |
142131ed | 2420 | default: |
2421 | break; | |
2422 | } | |
2423 | ||
c352cbc7 | 2424 | return 0; |
2425 | } | |
2426 | ||
d10cfa8d | 2427 | /* Returns nonzero if X might mention something which is not |
c352cbc7 | 2428 | local to the function and is not constant. */ |
142131ed | 2429 | |
c352cbc7 | 2430 | static int |
aecda0d6 | 2431 | nonlocal_mentioned_p (rtx x) |
c352cbc7 | 2432 | { |
c352cbc7 | 2433 | if (INSN_P (x)) |
2434 | { | |
6d7dc5b9 | 2435 | if (CALL_P (x)) |
c352cbc7 | 2436 | { |
2437 | if (! CONST_OR_PURE_CALL_P (x)) | |
2438 | return 1; | |
2439 | x = CALL_INSN_FUNCTION_USAGE (x); | |
2440 | if (x == 0) | |
2441 | return 0; | |
9342ee68 | 2442 | } |
c352cbc7 | 2443 | else |
9342ee68 | 2444 | x = PATTERN (x); |
c352cbc7 | 2445 | } |
2446 | ||
2447 | return for_each_rtx (&x, nonlocal_mentioned_p_1, NULL); | |
2448 | } | |
2449 | ||
2450 | /* A subroutine of nonlocal_referenced_p, returns 1 if *LOC references | |
2451 | something which is not local to the function and is not constant. */ | |
2452 | ||
2453 | static int | |
aecda0d6 | 2454 | nonlocal_referenced_p_1 (rtx *loc, void *data ATTRIBUTE_UNUSED) |
c352cbc7 | 2455 | { |
2456 | rtx x = *loc; | |
2457 | ||
2458 | if (! x) | |
2459 | return 0; | |
2460 | ||
2461 | switch (GET_CODE (x)) | |
2462 | { | |
2463 | case MEM: | |
2464 | case REG: | |
2465 | case SYMBOL_REF: | |
2466 | case SUBREG: | |
2467 | return nonlocal_mentioned_p (x); | |
2468 | ||
2469 | case CALL: | |
2470 | /* Non-constant calls and recursion are not local. */ | |
2471 | return 1; | |
2472 | ||
2473 | case SET: | |
2474 | if (nonlocal_mentioned_p (SET_SRC (x))) | |
2475 | return 1; | |
2476 | ||
e16ceb8e | 2477 | if (MEM_P (SET_DEST (x))) |
c352cbc7 | 2478 | return nonlocal_mentioned_p (XEXP (SET_DEST (x), 0)); |
2479 | ||
2480 | /* If the destination is anything other than a CC0, PC, | |
2481 | MEM, REG, or a SUBREG of a REG that occupies all of | |
2482 | the REG, then X references nonlocal memory if it is | |
2483 | mentioned in the destination. */ | |
2484 | if (GET_CODE (SET_DEST (x)) != CC0 | |
2485 | && GET_CODE (SET_DEST (x)) != PC | |
8ad4c111 | 2486 | && !REG_P (SET_DEST (x)) |
c352cbc7 | 2487 | && ! (GET_CODE (SET_DEST (x)) == SUBREG |
8ad4c111 | 2488 | && REG_P (SUBREG_REG (SET_DEST (x))) |
c352cbc7 | 2489 | && (((GET_MODE_SIZE (GET_MODE (SUBREG_REG (SET_DEST (x)))) |
2490 | + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD) | |
2491 | == ((GET_MODE_SIZE (GET_MODE (SET_DEST (x))) | |
2492 | + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD)))) | |
2493 | return nonlocal_mentioned_p (SET_DEST (x)); | |
2494 | return 0; | |
2495 | ||
2496 | case CLOBBER: | |
e16ceb8e | 2497 | if (MEM_P (XEXP (x, 0))) |
c352cbc7 | 2498 | return nonlocal_mentioned_p (XEXP (XEXP (x, 0), 0)); |
2499 | return 0; | |
2500 | ||
2501 | case USE: | |
2502 | return nonlocal_mentioned_p (XEXP (x, 0)); | |
2503 | ||
2504 | case ASM_INPUT: | |
2505 | case UNSPEC_VOLATILE: | |
2506 | return 1; | |
2507 | ||
2508 | case ASM_OPERANDS: | |
2509 | if (MEM_VOLATILE_P (x)) | |
2510 | return 1; | |
2511 | ||
d632b59a | 2512 | /* Fall through. */ |
c352cbc7 | 2513 | |
2514 | default: | |
2515 | break; | |
2516 | } | |
2517 | ||
2518 | return 0; | |
2519 | } | |
2520 | ||
d10cfa8d | 2521 | /* Returns nonzero if X might reference something which is not |
c352cbc7 | 2522 | local to the function and is not constant. */ |
2523 | ||
2524 | static int | |
aecda0d6 | 2525 | nonlocal_referenced_p (rtx x) |
c352cbc7 | 2526 | { |
c352cbc7 | 2527 | if (INSN_P (x)) |
2528 | { | |
6d7dc5b9 | 2529 | if (CALL_P (x)) |
c352cbc7 | 2530 | { |
2531 | if (! CONST_OR_PURE_CALL_P (x)) | |
2532 | return 1; | |
2533 | x = CALL_INSN_FUNCTION_USAGE (x); | |
2534 | if (x == 0) | |
2535 | return 0; | |
9342ee68 | 2536 | } |
c352cbc7 | 2537 | else |
9342ee68 | 2538 | x = PATTERN (x); |
c352cbc7 | 2539 | } |
2540 | ||
2541 | return for_each_rtx (&x, nonlocal_referenced_p_1, NULL); | |
2542 | } | |
2543 | ||
2544 | /* A subroutine of nonlocal_set_p, returns 1 if *LOC sets | |
2545 | something which is not local to the function and is not constant. */ | |
2546 | ||
2547 | static int | |
aecda0d6 | 2548 | nonlocal_set_p_1 (rtx *loc, void *data ATTRIBUTE_UNUSED) |
c352cbc7 | 2549 | { |
2550 | rtx x = *loc; | |
2551 | ||
2552 | if (! x) | |
2553 | return 0; | |
2554 | ||
2555 | switch (GET_CODE (x)) | |
2556 | { | |
2557 | case CALL: | |
2558 | /* Non-constant calls and recursion are not local. */ | |
2559 | return 1; | |
2560 | ||
2561 | case PRE_INC: | |
2562 | case PRE_DEC: | |
2563 | case POST_INC: | |
2564 | case POST_DEC: | |
2565 | case PRE_MODIFY: | |
2566 | case POST_MODIFY: | |
2567 | return nonlocal_mentioned_p (XEXP (x, 0)); | |
2568 | ||
2569 | case SET: | |
2570 | if (nonlocal_mentioned_p (SET_DEST (x))) | |
2571 | return 1; | |
2572 | return nonlocal_set_p (SET_SRC (x)); | |
2573 | ||
2574 | case CLOBBER: | |
2575 | return nonlocal_mentioned_p (XEXP (x, 0)); | |
2576 | ||
2577 | case USE: | |
2578 | return 0; | |
2579 | ||
2580 | case ASM_INPUT: | |
2581 | case UNSPEC_VOLATILE: | |
2582 | return 1; | |
2583 | ||
2584 | case ASM_OPERANDS: | |
2585 | if (MEM_VOLATILE_P (x)) | |
2586 | return 1; | |
2587 | ||
d632b59a | 2588 | /* Fall through. */ |
c352cbc7 | 2589 | |
2590 | default: | |
2591 | break; | |
2592 | } | |
142131ed | 2593 | |
2594 | return 0; | |
2595 | } | |
2596 | ||
d10cfa8d | 2597 | /* Returns nonzero if X might set something which is not |
c352cbc7 | 2598 | local to the function and is not constant. */ |
2599 | ||
2600 | static int | |
aecda0d6 | 2601 | nonlocal_set_p (rtx x) |
c352cbc7 | 2602 | { |
c352cbc7 | 2603 | if (INSN_P (x)) |
2604 | { | |
6d7dc5b9 | 2605 | if (CALL_P (x)) |
c352cbc7 | 2606 | { |
2607 | if (! CONST_OR_PURE_CALL_P (x)) | |
2608 | return 1; | |
2609 | x = CALL_INSN_FUNCTION_USAGE (x); | |
2610 | if (x == 0) | |
2611 | return 0; | |
9342ee68 | 2612 | } |
c352cbc7 | 2613 | else |
9342ee68 | 2614 | x = PATTERN (x); |
c352cbc7 | 2615 | } |
2616 | ||
2617 | return for_each_rtx (&x, nonlocal_set_p_1, NULL); | |
2618 | } | |
2619 | ||
01960f67 | 2620 | /* Mark the function if it is pure or constant. */ |
142131ed | 2621 | |
2622 | void | |
aecda0d6 | 2623 | mark_constant_function (void) |
142131ed | 2624 | { |
2625 | rtx insn; | |
c352cbc7 | 2626 | int nonlocal_memory_referenced; |
142131ed | 2627 | |
4f8e9d5c | 2628 | if (TREE_READONLY (current_function_decl) |
1addbe73 | 2629 | || DECL_IS_PURE (current_function_decl) |
142131ed | 2630 | || TREE_THIS_VOLATILE (current_function_decl) |
4f8e9d5c | 2631 | || current_function_has_nonlocal_goto |
883b2e73 | 2632 | || !targetm.binds_local_p (current_function_decl)) |
142131ed | 2633 | return; |
2634 | ||
080f162d | 2635 | /* A loop might not return which counts as a side effect. */ |
ecb7b891 | 2636 | if (mark_dfs_back_edges ()) |
080f162d | 2637 | return; |
2638 | ||
c352cbc7 | 2639 | nonlocal_memory_referenced = 0; |
1addbe73 | 2640 | |
2641 | init_alias_analysis (); | |
2642 | ||
c352cbc7 | 2643 | /* Determine if this is a constant or pure function. */ |
142131ed | 2644 | |
2645 | for (insn = get_insns (); insn; insn = NEXT_INSN (insn)) | |
c352cbc7 | 2646 | { |
2647 | if (! INSN_P (insn)) | |
2648 | continue; | |
1addbe73 | 2649 | |
c352cbc7 | 2650 | if (nonlocal_set_p (insn) || global_reg_mentioned_p (insn) |
2651 | || volatile_refs_p (PATTERN (insn))) | |
9342ee68 | 2652 | break; |
142131ed | 2653 | |
c352cbc7 | 2654 | if (! nonlocal_memory_referenced) |
2655 | nonlocal_memory_referenced = nonlocal_referenced_p (insn); | |
2656 | } | |
9342ee68 | 2657 | |
c352cbc7 | 2658 | end_alias_analysis (); |
9342ee68 | 2659 | |
142131ed | 2660 | /* Mark the function. */ |
9342ee68 | 2661 | |
c352cbc7 | 2662 | if (insn) |
2663 | ; | |
2664 | else if (nonlocal_memory_referenced) | |
01960f67 | 2665 | { |
2666 | cgraph_rtl_info (current_function_decl)->pure_function = 1; | |
2667 | DECL_IS_PURE (current_function_decl) = 1; | |
2668 | } | |
c352cbc7 | 2669 | else |
01960f67 | 2670 | { |
2671 | cgraph_rtl_info (current_function_decl)->const_function = 1; | |
2672 | TREE_READONLY (current_function_decl) = 1; | |
2673 | } | |
142131ed | 2674 | } |
c352cbc7 | 2675 | \f |
35d3eedd | 2676 | |
35d3eedd | 2677 | void |
aecda0d6 | 2678 | init_alias_once (void) |
35d3eedd | 2679 | { |
19cb6b50 | 2680 | int i; |
35d3eedd | 2681 | |
35d3eedd | 2682 | for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) |
2683 | /* Check whether this register can hold an incoming pointer | |
2684 | argument. FUNCTION_ARG_REGNO_P tests outgoing register | |
2c0e001b | 2685 | numbers, so translate if necessary due to register windows. */ |
35d3eedd | 2686 | if (FUNCTION_ARG_REGNO_P (OUTGOING_REGNO (i)) |
2687 | && HARD_REGNO_MODE_OK (i, Pmode)) | |
0c7f5242 | 2688 | static_reg_base_value[i] |
2689 | = gen_rtx_ADDRESS (VOIDmode, gen_rtx_REG (Pmode, i)); | |
2690 | ||
0c7f5242 | 2691 | static_reg_base_value[STACK_POINTER_REGNUM] |
2692 | = gen_rtx_ADDRESS (Pmode, stack_pointer_rtx); | |
2693 | static_reg_base_value[ARG_POINTER_REGNUM] | |
2694 | = gen_rtx_ADDRESS (Pmode, arg_pointer_rtx); | |
2695 | static_reg_base_value[FRAME_POINTER_REGNUM] | |
2696 | = gen_rtx_ADDRESS (Pmode, frame_pointer_rtx); | |
2697 | #if HARD_FRAME_POINTER_REGNUM != FRAME_POINTER_REGNUM | |
2698 | static_reg_base_value[HARD_FRAME_POINTER_REGNUM] | |
2699 | = gen_rtx_ADDRESS (Pmode, hard_frame_pointer_rtx); | |
2700 | #endif | |
2701 | } | |
2702 | ||
c7bf7428 | 2703 | /* Set MEMORY_MODIFIED when X modifies DATA (that is assumed |
2704 | to be memory reference. */ | |
2705 | static bool memory_modified; | |
2706 | static void | |
aecda0d6 | 2707 | memory_modified_1 (rtx x, rtx pat ATTRIBUTE_UNUSED, void *data) |
c7bf7428 | 2708 | { |
e16ceb8e | 2709 | if (MEM_P (x)) |
c7bf7428 | 2710 | { |
2711 | if (anti_dependence (x, (rtx)data) || output_dependence (x, (rtx)data)) | |
2712 | memory_modified = true; | |
2713 | } | |
2714 | } | |
2715 | ||
2716 | ||
2717 | /* Return true when INSN possibly modify memory contents of MEM | |
0c6d8c36 | 2718 | (i.e. address can be modified). */ |
c7bf7428 | 2719 | bool |
aecda0d6 | 2720 | memory_modified_in_insn_p (rtx mem, rtx insn) |
c7bf7428 | 2721 | { |
2722 | if (!INSN_P (insn)) | |
2723 | return false; | |
2724 | memory_modified = false; | |
2725 | note_stores (PATTERN (insn), memory_modified_1, mem); | |
2726 | return memory_modified; | |
2727 | } | |
2728 | ||
73f5c1e3 | 2729 | /* Initialize the aliasing machinery. Initialize the REG_KNOWN_VALUE |
2730 | array. */ | |
2731 | ||
ea0cb7ae | 2732 | void |
aecda0d6 | 2733 | init_alias_analysis (void) |
ea0cb7ae | 2734 | { |
e5b19a8c | 2735 | unsigned int maxreg = max_reg_num (); |
37b864db | 2736 | int changed, pass; |
19cb6b50 | 2737 | int i; |
2738 | unsigned int ui; | |
2739 | rtx insn; | |
ea0cb7ae | 2740 | |
376c21d1 | 2741 | timevar_push (TV_ALIAS_ANALYSIS); |
2742 | ||
12cb06ad | 2743 | reg_known_value_size = maxreg - FIRST_PSEUDO_REGISTER; |
2744 | reg_known_value = ggc_calloc (reg_known_value_size, sizeof (rtx)); | |
2745 | reg_known_equiv_p = xcalloc (reg_known_value_size, sizeof (bool)); | |
ea0cb7ae | 2746 | |
35d3eedd | 2747 | /* Overallocate reg_base_value to allow some growth during loop |
2748 | optimization. Loop unrolling can create a large number of | |
2749 | registers. */ | |
e5b19a8c | 2750 | if (old_reg_base_value) |
2751 | { | |
2752 | reg_base_value = old_reg_base_value; | |
2753 | /* If varray gets large zeroing cost may get important. */ | |
2754 | if (VARRAY_SIZE (reg_base_value) > 256 | |
2755 | && VARRAY_SIZE (reg_base_value) > 4 * maxreg) | |
2756 | VARRAY_GROW (reg_base_value, maxreg); | |
2757 | VARRAY_CLEAR (reg_base_value); | |
2758 | if (VARRAY_SIZE (reg_base_value) < maxreg) | |
2759 | VARRAY_GROW (reg_base_value, maxreg); | |
2760 | } | |
2761 | else | |
2762 | { | |
2763 | VARRAY_RTX_INIT (reg_base_value, maxreg, "reg_base_value"); | |
2764 | } | |
c2e94b25 | 2765 | |
e5b19a8c | 2766 | new_reg_base_value = xmalloc (maxreg * sizeof (rtx)); |
2767 | reg_seen = xmalloc (maxreg); | |
06c5c535 | 2768 | |
2769 | /* The basic idea is that each pass through this loop will use the | |
2770 | "constant" information from the previous pass to propagate alias | |
2771 | information through another level of assignments. | |
2772 | ||
2773 | This could get expensive if the assignment chains are long. Maybe | |
2774 | we should throttle the number of iterations, possibly based on | |
35d3eedd | 2775 | the optimization level or flag_expensive_optimizations. |
06c5c535 | 2776 | |
2777 | We could propagate more information in the first pass by making use | |
2778 | of REG_N_SETS to determine immediately that the alias information | |
37b864db | 2779 | for a pseudo is "constant". |
2780 | ||
2781 | A program with an uninitialized variable can cause an infinite loop | |
2782 | here. Instead of doing a full dataflow analysis to detect such problems | |
2783 | we just cap the number of iterations for the loop. | |
2784 | ||
2785 | The state of the arrays for the set chain in question does not matter | |
2786 | since the program has undefined behavior. */ | |
35d3eedd | 2787 | |
37b864db | 2788 | pass = 0; |
35d3eedd | 2789 | do |
06c5c535 | 2790 | { |
2791 | /* Assume nothing will change this iteration of the loop. */ | |
2792 | changed = 0; | |
2793 | ||
06c5c535 | 2794 | /* We want to assign the same IDs each iteration of this loop, so |
2795 | start counting from zero each iteration of the loop. */ | |
2796 | unique_id = 0; | |
2797 | ||
35a3065a | 2798 | /* We're at the start of the function each iteration through the |
06c5c535 | 2799 | loop, so we're copying arguments. */ |
e5be0cc6 | 2800 | copying_arguments = true; |
ea0cb7ae | 2801 | |
35d3eedd | 2802 | /* Wipe the potential alias information clean for this pass. */ |
e5b19a8c | 2803 | memset (new_reg_base_value, 0, maxreg * sizeof (rtx)); |
f0ad0992 | 2804 | |
35d3eedd | 2805 | /* Wipe the reg_seen array clean. */ |
e5b19a8c | 2806 | memset (reg_seen, 0, maxreg); |
ea0cb7ae | 2807 | |
35d3eedd | 2808 | /* Mark all hard registers which may contain an address. |
2809 | The stack, frame and argument pointers may contain an address. | |
2810 | An argument register which can hold a Pmode value may contain | |
2811 | an address even if it is not in BASE_REGS. | |
f0ad0992 | 2812 | |
35d3eedd | 2813 | The address expression is VOIDmode for an argument and |
2814 | Pmode for other registers. */ | |
2815 | ||
e26fd9bc | 2816 | memcpy (new_reg_base_value, static_reg_base_value, |
2817 | FIRST_PSEUDO_REGISTER * sizeof (rtx)); | |
35d3eedd | 2818 | |
06c5c535 | 2819 | /* Walk the insns adding values to the new_reg_base_value array. */ |
2820 | for (insn = get_insns (); insn; insn = NEXT_INSN (insn)) | |
ea0cb7ae | 2821 | { |
9204e736 | 2822 | if (INSN_P (insn)) |
06c5c535 | 2823 | { |
35d3eedd | 2824 | rtx note, set; |
d6cb6164 | 2825 | |
2826 | #if defined (HAVE_prologue) || defined (HAVE_epilogue) | |
35a3065a | 2827 | /* The prologue/epilogue insns are not threaded onto the |
fd037eda | 2828 | insn chain until after reload has completed. Thus, |
2829 | there is no sense wasting time checking if INSN is in | |
2830 | the prologue/epilogue until after reload has completed. */ | |
2831 | if (reload_completed | |
2832 | && prologue_epilogue_contains (insn)) | |
d6cb6164 | 2833 | continue; |
2834 | #endif | |
2835 | ||
06c5c535 | 2836 | /* If this insn has a noalias note, process it, Otherwise, |
2837 | scan for sets. A simple set will have no side effects | |
2c0e001b | 2838 | which could change the base value of any other register. */ |
35d3eedd | 2839 | |
06c5c535 | 2840 | if (GET_CODE (PATTERN (insn)) == SET |
d6cb6164 | 2841 | && REG_NOTES (insn) != 0 |
2842 | && find_reg_note (insn, REG_NOALIAS, NULL_RTX)) | |
ec8895d7 | 2843 | record_set (SET_DEST (PATTERN (insn)), NULL_RTX, NULL); |
06c5c535 | 2844 | else |
ec8895d7 | 2845 | note_stores (PATTERN (insn), record_set, NULL); |
35d3eedd | 2846 | |
2847 | set = single_set (insn); | |
2848 | ||
2849 | if (set != 0 | |
8ad4c111 | 2850 | && REG_P (SET_DEST (set)) |
9db10305 | 2851 | && REGNO (SET_DEST (set)) >= FIRST_PSEUDO_REGISTER) |
35d3eedd | 2852 | { |
9db10305 | 2853 | unsigned int regno = REGNO (SET_DEST (set)); |
970c0c06 | 2854 | rtx src = SET_SRC (set); |
12cb06ad | 2855 | rtx t; |
970c0c06 | 2856 | |
2857 | if (REG_NOTES (insn) != 0 | |
2858 | && (((note = find_reg_note (insn, REG_EQUAL, 0)) != 0 | |
2859 | && REG_N_SETS (regno) == 1) | |
2860 | || (note = find_reg_note (insn, REG_EQUIV, NULL_RTX)) != 0) | |
2861 | && GET_CODE (XEXP (note, 0)) != EXPR_LIST | |
97fb4325 | 2862 | && ! rtx_varies_p (XEXP (note, 0), 1) |
12cb06ad | 2863 | && ! reg_overlap_mentioned_p (SET_DEST (set), |
2864 | XEXP (note, 0))) | |
970c0c06 | 2865 | { |
12cb06ad | 2866 | set_reg_known_value (regno, XEXP (note, 0)); |
2867 | set_reg_known_equiv_p (regno, | |
2868 | REG_NOTE_KIND (note) == REG_EQUIV); | |
970c0c06 | 2869 | } |
2870 | else if (REG_N_SETS (regno) == 1 | |
2871 | && GET_CODE (src) == PLUS | |
8ad4c111 | 2872 | && REG_P (XEXP (src, 0)) |
12cb06ad | 2873 | && (t = get_reg_known_value (REGNO (XEXP (src, 0)))) |
970c0c06 | 2874 | && GET_CODE (XEXP (src, 1)) == CONST_INT) |
2875 | { | |
12cb06ad | 2876 | t = plus_constant (t, INTVAL (XEXP (src, 1))); |
2877 | set_reg_known_value (regno, t); | |
2878 | set_reg_known_equiv_p (regno, 0); | |
970c0c06 | 2879 | } |
2880 | else if (REG_N_SETS (regno) == 1 | |
2881 | && ! rtx_varies_p (src, 1)) | |
2882 | { | |
12cb06ad | 2883 | set_reg_known_value (regno, src); |
2884 | set_reg_known_equiv_p (regno, 0); | |
970c0c06 | 2885 | } |
35d3eedd | 2886 | } |
06c5c535 | 2887 | } |
6d7dc5b9 | 2888 | else if (NOTE_P (insn) |
06c5c535 | 2889 | && NOTE_LINE_NUMBER (insn) == NOTE_INSN_FUNCTION_BEG) |
e5be0cc6 | 2890 | copying_arguments = false; |
35d3eedd | 2891 | } |
06c5c535 | 2892 | |
35d3eedd | 2893 | /* Now propagate values from new_reg_base_value to reg_base_value. */ |
64db345d | 2894 | gcc_assert (maxreg == (unsigned int) max_reg_num()); |
2895 | ||
e5b19a8c | 2896 | for (ui = 0; ui < maxreg; ui++) |
35d3eedd | 2897 | { |
274c11d8 | 2898 | if (new_reg_base_value[ui] |
e5b19a8c | 2899 | && new_reg_base_value[ui] != VARRAY_RTX (reg_base_value, ui) |
2900 | && ! rtx_equal_p (new_reg_base_value[ui], | |
2901 | VARRAY_RTX (reg_base_value, ui))) | |
06c5c535 | 2902 | { |
e5b19a8c | 2903 | VARRAY_RTX (reg_base_value, ui) = new_reg_base_value[ui]; |
35d3eedd | 2904 | changed = 1; |
06c5c535 | 2905 | } |
ea0cb7ae | 2906 | } |
ea0cb7ae | 2907 | } |
35d3eedd | 2908 | while (changed && ++pass < MAX_ALIAS_LOOP_PASSES); |
ea0cb7ae | 2909 | |
2910 | /* Fill in the remaining entries. */ | |
12cb06ad | 2911 | for (i = 0; i < (int)reg_known_value_size; i++) |
ea0cb7ae | 2912 | if (reg_known_value[i] == 0) |
12cb06ad | 2913 | reg_known_value[i] = regno_reg_rtx[i + FIRST_PSEUDO_REGISTER]; |
ea0cb7ae | 2914 | |
ea0cb7ae | 2915 | /* Simplify the reg_base_value array so that no register refers to |
2916 | another register, except to special registers indirectly through | |
2917 | ADDRESS expressions. | |
2918 | ||
2919 | In theory this loop can take as long as O(registers^2), but unless | |
2920 | there are very long dependency chains it will run in close to linear | |
37b864db | 2921 | time. |
2922 | ||
2923 | This loop may not be needed any longer now that the main loop does | |
2924 | a better job at propagating alias information. */ | |
2925 | pass = 0; | |
ea0cb7ae | 2926 | do |
2927 | { | |
2928 | changed = 0; | |
37b864db | 2929 | pass++; |
e5b19a8c | 2930 | for (ui = 0; ui < maxreg; ui++) |
ea0cb7ae | 2931 | { |
e5b19a8c | 2932 | rtx base = VARRAY_RTX (reg_base_value, ui); |
8ad4c111 | 2933 | if (base && REG_P (base)) |
ea0cb7ae | 2934 | { |
9db10305 | 2935 | unsigned int base_regno = REGNO (base); |
274c11d8 | 2936 | if (base_regno == ui) /* register set from itself */ |
e5b19a8c | 2937 | VARRAY_RTX (reg_base_value, ui) = 0; |
ea0cb7ae | 2938 | else |
e5b19a8c | 2939 | VARRAY_RTX (reg_base_value, ui) |
2940 | = VARRAY_RTX (reg_base_value, base_regno); | |
ea0cb7ae | 2941 | changed = 1; |
2942 | } | |
2943 | } | |
2944 | } | |
37b864db | 2945 | while (changed && pass < MAX_ALIAS_LOOP_PASSES); |
ea0cb7ae | 2946 | |
7098ef4a | 2947 | /* Clean up. */ |
2948 | free (new_reg_base_value); | |
06c5c535 | 2949 | new_reg_base_value = 0; |
7098ef4a | 2950 | free (reg_seen); |
ea0cb7ae | 2951 | reg_seen = 0; |
376c21d1 | 2952 | timevar_pop (TV_ALIAS_ANALYSIS); |
ea0cb7ae | 2953 | } |
2954 | ||
2955 | void | |
aecda0d6 | 2956 | end_alias_analysis (void) |
ea0cb7ae | 2957 | { |
e5b19a8c | 2958 | old_reg_base_value = reg_base_value; |
12cb06ad | 2959 | ggc_free (reg_known_value); |
ea0cb7ae | 2960 | reg_known_value = 0; |
c2e94b25 | 2961 | reg_known_value_size = 0; |
12cb06ad | 2962 | free (reg_known_equiv_p); |
7098ef4a | 2963 | reg_known_equiv_p = 0; |
1a6c3add | 2964 | if (alias_invariant) |
2965 | { | |
0fde16b4 | 2966 | ggc_free (alias_invariant); |
1a6c3add | 2967 | alias_invariant = 0; |
e5b19a8c | 2968 | alias_invariant_size = 0; |
1a6c3add | 2969 | } |
ea0cb7ae | 2970 | } |
77fce4cd | 2971 | \f |
2972 | /* Do control and data flow analysis; write some of the results to the | |
2973 | dump file. */ | |
2974 | static void | |
2975 | rest_of_handle_cfg (void) | |
2976 | { | |
2977 | if (dump_file) | |
2978 | dump_flow_info (dump_file); | |
2979 | if (optimize) | |
2980 | cleanup_cfg (CLEANUP_EXPENSIVE | |
2981 | | (flag_thread_jumps ? CLEANUP_THREADING : 0)); | |
2982 | ||
2983 | /* It may make more sense to mark constant functions after dead code is | |
2984 | eliminated by life_analysis, but we need to do it early, as -fprofile-arcs | |
2985 | may insert code making function non-constant, but we still must consider | |
2986 | it as constant, otherwise -fbranch-probabilities will not read data back. | |
2987 | ||
2988 | life_analysis rarely eliminates modification of external memory. | |
2989 | ||
2990 | FIXME: now with tree based profiling we are in the trap described above | |
2991 | again. It seems to be easiest to disable the optimization for time | |
2992 | being before the problem is either solved by moving the transformation | |
2993 | to the IPA level (we need the CFG for this) or the very early optimization | |
2994 | passes are made to ignore the const/pure flags so code does not change. */ | |
2995 | if (optimize | |
2996 | && (!flag_tree_based_profiling | |
2997 | || (!profile_arc_flag && !flag_branch_probabilities))) | |
2998 | { | |
2999 | /* Alias analysis depends on this information and mark_constant_function | |
3000 | depends on alias analysis. */ | |
3001 | reg_scan (get_insns (), max_reg_num ()); | |
3002 | mark_constant_function (); | |
3003 | } | |
3004 | } | |
3005 | ||
3006 | struct tree_opt_pass pass_cfg = | |
3007 | { | |
3008 | "cfg", /* name */ | |
3009 | NULL, /* gate */ | |
3010 | rest_of_handle_cfg, /* execute */ | |
3011 | NULL, /* sub */ | |
3012 | NULL, /* next */ | |
3013 | 0, /* static_pass_number */ | |
3014 | TV_FLOW, /* tv_id */ | |
3015 | 0, /* properties_required */ | |
3016 | 0, /* properties_provided */ | |
3017 | 0, /* properties_destroyed */ | |
3018 | 0, /* todo_flags_start */ | |
3019 | TODO_dump_func, /* todo_flags_finish */ | |
3020 | 'f' /* letter */ | |
3021 | }; | |
3022 | ||
1f3233d1 | 3023 | |
3024 | #include "gt-alias.h" |