]>
Commit | Line | Data |
---|---|---|
ea0cb7ae | 1 | /* Alias analysis for GNU C |
d4473c84 | 2 | Copyright (C) 1997, 1998, 1999, 2000, 2001, 2002, 2003, 2004, 2005, 2006, |
7cf0dbf3 | 3 | 2007, 2008, 2009, 2010 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 | |
8c4c00c1 | 10 | Software Foundation; either version 3, or (at your option) any later |
f12b58b3 | 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 | |
8c4c00c1 | 19 | along with GCC; see the file COPYING3. If not see |
20 | <http://www.gnu.org/licenses/>. */ | |
ea0cb7ae | 21 | |
22 | #include "config.h" | |
405711de | 23 | #include "system.h" |
805e22b2 | 24 | #include "coretypes.h" |
25 | #include "tm.h" | |
ea0cb7ae | 26 | #include "rtl.h" |
142131ed | 27 | #include "tree.h" |
7953c610 | 28 | #include "tm_p.h" |
0a893c29 | 29 | #include "function.h" |
4c74e6d9 | 30 | #include "alias.h" |
31 | #include "emit-rtl.h" | |
ea0cb7ae | 32 | #include "regs.h" |
33 | #include "hard-reg-set.h" | |
080f162d | 34 | #include "basic-block.h" |
ea0cb7ae | 35 | #include "flags.h" |
a068b2a9 | 36 | #include "output.h" |
0b205f4c | 37 | #include "diagnostic-core.h" |
1617c276 | 38 | #include "cselib.h" |
a5b1863e | 39 | #include "splay-tree.h" |
c2e94b25 | 40 | #include "ggc.h" |
b0278d39 | 41 | #include "langhooks.h" |
376c21d1 | 42 | #include "timevar.h" |
4f8e9d5c | 43 | #include "target.h" |
28992b23 | 44 | #include "cgraph.h" |
77fce4cd | 45 | #include "tree-pass.h" |
3072d30e | 46 | #include "df.h" |
3a443843 | 47 | #include "tree-ssa-alias.h" |
48 | #include "pointer-set.h" | |
49 | #include "tree-flow.h" | |
f7d118a9 | 50 | |
51 | /* The aliasing API provided here solves related but different problems: | |
52 | ||
a0c938f0 | 53 | Say there exists (in c) |
f7d118a9 | 54 | |
55 | struct X { | |
56 | struct Y y1; | |
57 | struct Z z2; | |
58 | } x1, *px1, *px2; | |
59 | ||
60 | struct Y y2, *py; | |
61 | struct Z z2, *pz; | |
62 | ||
63 | ||
64 | py = &px1.y1; | |
65 | px2 = &x1; | |
66 | ||
67 | Consider the four questions: | |
68 | ||
69 | Can a store to x1 interfere with px2->y1? | |
70 | Can a store to x1 interfere with px2->z2? | |
71 | (*px2).z2 | |
72 | Can a store to x1 change the value pointed to by with py? | |
73 | Can a store to x1 change the value pointed to by with pz? | |
74 | ||
75 | The answer to these questions can be yes, yes, yes, and maybe. | |
76 | ||
77 | The first two questions can be answered with a simple examination | |
78 | of the type system. If structure X contains a field of type Y then | |
79 | a store thru a pointer to an X can overwrite any field that is | |
80 | contained (recursively) in an X (unless we know that px1 != px2). | |
81 | ||
82 | The last two of the questions can be solved in the same way as the | |
83 | first two questions but this is too conservative. The observation | |
84 | is that in some cases analysis we can know if which (if any) fields | |
85 | are addressed and if those addresses are used in bad ways. This | |
86 | analysis may be language specific. In C, arbitrary operations may | |
87 | be applied to pointers. However, there is some indication that | |
88 | this may be too conservative for some C++ types. | |
89 | ||
90 | The pass ipa-type-escape does this analysis for the types whose | |
a0c938f0 | 91 | instances do not escape across the compilation boundary. |
f7d118a9 | 92 | |
93 | Historically in GCC, these two problems were combined and a single | |
94 | data structure was used to represent the solution to these | |
95 | problems. We now have two similar but different data structures, | |
96 | The data structure to solve the last two question is similar to the | |
97 | first, but does not contain have the fields in it whose address are | |
98 | never taken. For types that do escape the compilation unit, the | |
99 | data structures will have identical information. | |
100 | */ | |
a5b1863e | 101 | |
102 | /* The alias sets assigned to MEMs assist the back-end in determining | |
103 | which MEMs can alias which other MEMs. In general, two MEMs in | |
5badbbc0 | 104 | different alias sets cannot alias each other, with one important |
105 | exception. Consider something like: | |
a5b1863e | 106 | |
8e2ef041 | 107 | struct S { int i; double d; }; |
a5b1863e | 108 | |
109 | a store to an `S' can alias something of either type `int' or type | |
110 | `double'. (However, a store to an `int' cannot alias a `double' | |
111 | and vice versa.) We indicate this via a tree structure that looks | |
112 | like: | |
a0c938f0 | 113 | struct S |
114 | / \ | |
a5b1863e | 115 | / \ |
a0c938f0 | 116 | |/_ _\| |
117 | int double | |
a5b1863e | 118 | |
5badbbc0 | 119 | (The arrows are directed and point downwards.) |
120 | In this situation we say the alias set for `struct S' is the | |
121 | `superset' and that those for `int' and `double' are `subsets'. | |
122 | ||
f7c44134 | 123 | To see whether two alias sets can point to the same memory, we must |
124 | see if either alias set is a subset of the other. We need not trace | |
d716ce75 | 125 | past immediate descendants, however, since we propagate all |
f7c44134 | 126 | grandchildren up one level. |
a5b1863e | 127 | |
128 | Alias set zero is implicitly a superset of all other alias sets. | |
129 | However, this is no actual entry for alias set zero. It is an | |
130 | error to attempt to explicitly construct a subset of zero. */ | |
131 | ||
26dbec0a | 132 | struct GTY(()) alias_set_entry_d { |
a5b1863e | 133 | /* The alias set number, as stored in MEM_ALIAS_SET. */ |
32c2fdea | 134 | alias_set_type alias_set; |
a5b1863e | 135 | |
1e020227 | 136 | /* Nonzero if would have a child of zero: this effectively makes this |
137 | alias set the same as alias set zero. */ | |
138 | int has_zero_child; | |
139 | ||
a5b1863e | 140 | /* The children of the alias set. These are not just the immediate |
d716ce75 | 141 | children, but, in fact, all descendants. So, if we have: |
a5b1863e | 142 | |
9342ee68 | 143 | struct T { struct S s; float f; } |
a5b1863e | 144 | |
145 | continuing our example above, the children here will be all of | |
146 | `int', `double', `float', and `struct S'. */ | |
318dcdd8 | 147 | splay_tree GTY((param1_is (int), param2_is (int))) children; |
318dcdd8 | 148 | }; |
26dbec0a | 149 | typedef struct alias_set_entry_d *alias_set_entry; |
ea0cb7ae | 150 | |
7ecb5bb2 | 151 | static int rtx_equal_for_memref_p (const_rtx, const_rtx); |
aecda0d6 | 152 | static int memrefs_conflict_p (int, rtx, int, rtx, HOST_WIDE_INT); |
81a410b1 | 153 | static void record_set (rtx, const_rtx, void *); |
aecda0d6 | 154 | static int base_alias_check (rtx, rtx, enum machine_mode, |
155 | enum machine_mode); | |
156 | static rtx find_base_value (rtx); | |
52d07779 | 157 | static int mems_in_disjoint_alias_sets_p (const_rtx, const_rtx); |
aecda0d6 | 158 | static int insert_subset_children (splay_tree_node, void*); |
32c2fdea | 159 | static alias_set_entry get_alias_set_entry (alias_set_type); |
52d07779 | 160 | static int aliases_everything_p (const_rtx); |
161 | static bool nonoverlapping_component_refs_p (const_tree, const_tree); | |
aecda0d6 | 162 | static tree decl_for_component_ref (tree); |
52d07779 | 163 | static int write_dependence_p (const_rtx, const_rtx, int); |
aecda0d6 | 164 | |
a80f1c6c | 165 | static void memory_modified_1 (rtx, const_rtx, void *); |
ea0cb7ae | 166 | |
167 | /* Set up all info needed to perform alias analysis on memory references. */ | |
168 | ||
083a2b5e | 169 | /* Returns the size in bytes of the mode of X. */ |
ea0cb7ae | 170 | #define SIZE_FOR_MODE(X) (GET_MODE_SIZE (GET_MODE (X))) |
171 | ||
b5ba9f3a | 172 | /* Returns nonzero if MEM1 and MEM2 do not alias because they are in |
a068b2a9 | 173 | different alias sets. We ignore alias sets in functions making use |
174 | of variable arguments because the va_arg macros on some systems are | |
175 | not legal ANSI C. */ | |
176 | #define DIFFERENT_ALIAS_SETS_P(MEM1, MEM2) \ | |
a5b1863e | 177 | mems_in_disjoint_alias_sets_p (MEM1, MEM2) |
b5ba9f3a | 178 | |
37b864db | 179 | /* Cap the number of passes we make over the insns propagating alias |
5badbbc0 | 180 | information through set chains. 10 is a completely arbitrary choice. */ |
37b864db | 181 | #define MAX_ALIAS_LOOP_PASSES 10 |
9342ee68 | 182 | |
ea0cb7ae | 183 | /* reg_base_value[N] gives an address to which register N is related. |
184 | If all sets after the first add or subtract to the current value | |
185 | or otherwise modify it so it does not point to a different top level | |
186 | object, reg_base_value[N] is equal to the address part of the source | |
e7c893ba | 187 | of the first set. |
188 | ||
189 | A base address can be an ADDRESS, SYMBOL_REF, or LABEL_REF. ADDRESS | |
4e56ceb1 | 190 | expressions represent three types of base: |
3a935f06 | 191 | |
4e56ceb1 | 192 | 1. incoming arguments. There is just one ADDRESS to represent all |
193 | arguments, since we do not know at this level whether accesses | |
194 | based on different arguments can alias. The ADDRESS has id 0. | |
3a935f06 | 195 | |
4e56ceb1 | 196 | 2. stack_pointer_rtx, frame_pointer_rtx, hard_frame_pointer_rtx |
197 | (if distinct from frame_pointer_rtx) and arg_pointer_rtx. | |
198 | Each of these rtxes has a separate ADDRESS associated with it, | |
199 | each with a negative id. | |
200 | ||
201 | GCC is (and is required to be) precise in which register it | |
202 | chooses to access a particular region of stack. We can therefore | |
203 | assume that accesses based on one of these rtxes do not alias | |
204 | accesses based on another of these rtxes. | |
205 | ||
206 | 3. bases that are derived from malloc()ed memory (REG_NOALIAS). | |
207 | Each such piece of memory has a separate ADDRESS associated | |
208 | with it, each with an id greater than 0. | |
209 | ||
210 | Accesses based on one ADDRESS do not alias accesses based on other | |
211 | ADDRESSes. Accesses based on ADDRESSes in groups (2) and (3) do not | |
212 | alias globals either; the ADDRESSes have Pmode to indicate this. | |
213 | The ADDRESS in group (1) _may_ alias globals; it has VOIDmode to | |
214 | indicate this. */ | |
e7c893ba | 215 | |
980f3285 | 216 | static GTY(()) VEC(rtx,gc) *reg_base_value; |
c2e94b25 | 217 | static rtx *new_reg_base_value; |
e5b19a8c | 218 | |
4e56ceb1 | 219 | /* The single VOIDmode ADDRESS that represents all argument bases. |
220 | It has id 0. */ | |
221 | static GTY(()) rtx arg_base_value; | |
222 | ||
223 | /* Used to allocate unique ids to each REG_NOALIAS ADDRESS. */ | |
224 | static int unique_id; | |
225 | ||
e5b19a8c | 226 | /* We preserve the copy of old array around to avoid amount of garbage |
227 | produced. About 8% of garbage produced were attributed to this | |
228 | array. */ | |
980f3285 | 229 | static GTY((deletable)) VEC(rtx,gc) *old_reg_base_value; |
083a2b5e | 230 | |
eb2c25b5 | 231 | #define static_reg_base_value \ |
232 | (this_target_rtl->x_static_reg_base_value) | |
0c7f5242 | 233 | |
980f3285 | 234 | #define REG_BASE_VALUE(X) \ |
235 | (REGNO (X) < VEC_length (rtx, reg_base_value) \ | |
236 | ? VEC_index (rtx, reg_base_value, REGNO (X)) : 0) | |
ea0cb7ae | 237 | |
73f5c1e3 | 238 | /* Vector indexed by N giving the initial (unchanging) value known for |
12cb06ad | 239 | pseudo-register N. This array is initialized in init_alias_analysis, |
240 | and does not change until end_alias_analysis is called. */ | |
241 | static GTY((length("reg_known_value_size"))) rtx *reg_known_value; | |
ea0cb7ae | 242 | |
243 | /* Indicates number of valid entries in reg_known_value. */ | |
12cb06ad | 244 | static GTY(()) unsigned int reg_known_value_size; |
ea0cb7ae | 245 | |
246 | /* Vector recording for each reg_known_value whether it is due to a | |
247 | REG_EQUIV note. Future passes (viz., reload) may replace the | |
248 | pseudo with the equivalent expression and so we account for the | |
5badbbc0 | 249 | dependences that would be introduced if that happens. |
250 | ||
251 | The REG_EQUIV notes created in assign_parms may mention the arg | |
252 | pointer, and there are explicit insns in the RTL that modify the | |
253 | arg pointer. Thus we must ensure that such insns don't get | |
254 | scheduled across each other because that would invalidate the | |
255 | REG_EQUIV notes. One could argue that the REG_EQUIV notes are | |
256 | wrong, but solving the problem in the scheduler will likely give | |
257 | better code, so we do it here. */ | |
12cb06ad | 258 | static bool *reg_known_equiv_p; |
ea0cb7ae | 259 | |
e7c893ba | 260 | /* True when scanning insns from the start of the rtl to the |
261 | NOTE_INSN_FUNCTION_BEG note. */ | |
e5be0cc6 | 262 | static bool copying_arguments; |
ea0cb7ae | 263 | |
8b435339 | 264 | DEF_VEC_P(alias_set_entry); |
265 | DEF_VEC_ALLOC_P(alias_set_entry,gc); | |
266 | ||
a5b1863e | 267 | /* The splay-tree used to store the various alias set entries. */ |
8b435339 | 268 | static GTY (()) VEC(alias_set_entry,gc) *alias_sets; |
5badbbc0 | 269 | \f |
3a443843 | 270 | /* Build a decomposed reference object for querying the alias-oracle |
271 | from the MEM rtx and store it in *REF. | |
272 | Returns false if MEM is not suitable for the alias-oracle. */ | |
273 | ||
274 | static bool | |
275 | ao_ref_from_mem (ao_ref *ref, const_rtx mem) | |
276 | { | |
277 | tree expr = MEM_EXPR (mem); | |
278 | tree base; | |
279 | ||
280 | if (!expr) | |
281 | return false; | |
282 | ||
283 | ao_ref_init (ref, expr); | |
284 | ||
285 | /* Get the base of the reference and see if we have to reject or | |
286 | adjust it. */ | |
287 | base = ao_ref_base (ref); | |
288 | if (base == NULL_TREE) | |
289 | return false; | |
290 | ||
1f1e1860 | 291 | /* The tree oracle doesn't like to have these. */ |
292 | if (TREE_CODE (base) == FUNCTION_DECL | |
293 | || TREE_CODE (base) == LABEL_DECL) | |
294 | return false; | |
295 | ||
3a443843 | 296 | /* If this is a pointer dereference of a non-SSA_NAME punt. |
297 | ??? We could replace it with a pointer to anything. */ | |
182cf5a9 | 298 | if ((INDIRECT_REF_P (base) |
299 | || TREE_CODE (base) == MEM_REF) | |
3a443843 | 300 | && TREE_CODE (TREE_OPERAND (base, 0)) != SSA_NAME) |
301 | return false; | |
7e9c660e | 302 | if (TREE_CODE (base) == TARGET_MEM_REF |
303 | && TMR_BASE (base) | |
304 | && TREE_CODE (TMR_BASE (base)) != SSA_NAME) | |
305 | return false; | |
3a443843 | 306 | |
307 | /* If this is a reference based on a partitioned decl replace the | |
308 | base with an INDIRECT_REF of the pointer representative we | |
309 | created during stack slot partitioning. */ | |
310 | if (TREE_CODE (base) == VAR_DECL | |
311 | && ! TREE_STATIC (base) | |
312 | && cfun->gimple_df->decls_to_pointers != NULL) | |
313 | { | |
314 | void *namep; | |
315 | namep = pointer_map_contains (cfun->gimple_df->decls_to_pointers, base); | |
316 | if (namep) | |
182cf5a9 | 317 | ref->base = build_simple_mem_ref (*(tree *)namep); |
3a443843 | 318 | } |
7e9c660e | 319 | else if (TREE_CODE (base) == TARGET_MEM_REF |
28daba6f | 320 | && TREE_CODE (TMR_BASE (base)) == ADDR_EXPR |
321 | && TREE_CODE (TREE_OPERAND (TMR_BASE (base), 0)) == VAR_DECL | |
322 | && ! TREE_STATIC (TREE_OPERAND (TMR_BASE (base), 0)) | |
7e9c660e | 323 | && cfun->gimple_df->decls_to_pointers != NULL) |
324 | { | |
325 | void *namep; | |
326 | namep = pointer_map_contains (cfun->gimple_df->decls_to_pointers, | |
28daba6f | 327 | TREE_OPERAND (TMR_BASE (base), 0)); |
7e9c660e | 328 | if (namep) |
329 | ref->base = build_simple_mem_ref (*(tree *)namep); | |
330 | } | |
3a443843 | 331 | |
332 | ref->ref_alias_set = MEM_ALIAS_SET (mem); | |
333 | ||
5b2a69fa | 334 | /* If MEM_OFFSET or MEM_SIZE are unknown we have to punt. |
1f1e1860 | 335 | Keep points-to related information though. */ |
da443c27 | 336 | if (!MEM_OFFSET_KNOWN_P (mem) |
5b2a69fa | 337 | || !MEM_SIZE_KNOWN_P (mem)) |
1f1e1860 | 338 | { |
339 | ref->ref = NULL_TREE; | |
340 | ref->offset = 0; | |
341 | ref->size = -1; | |
342 | ref->max_size = -1; | |
343 | return true; | |
344 | } | |
345 | ||
3804203d | 346 | /* If the base decl is a parameter we can have negative MEM_OFFSET in |
347 | case of promoted subregs on bigendian targets. Trust the MEM_EXPR | |
348 | here. */ | |
da443c27 | 349 | if (MEM_OFFSET (mem) < 0 |
350 | && (MEM_SIZE (mem) + MEM_OFFSET (mem)) * BITS_PER_UNIT == ref->size) | |
3804203d | 351 | return true; |
352 | ||
da443c27 | 353 | ref->offset += MEM_OFFSET (mem) * BITS_PER_UNIT; |
5b2a69fa | 354 | ref->size = MEM_SIZE (mem) * BITS_PER_UNIT; |
3804203d | 355 | |
356 | /* The MEM may extend into adjacent fields, so adjust max_size if | |
357 | necessary. */ | |
358 | if (ref->max_size != -1 | |
359 | && ref->size > ref->max_size) | |
360 | ref->max_size = ref->size; | |
361 | ||
362 | /* If MEM_OFFSET and MEM_SIZE get us outside of the base object of | |
363 | the MEM_EXPR punt. This happens for STRICT_ALIGNMENT targets a lot. */ | |
364 | if (MEM_EXPR (mem) != get_spill_slot_decl (false) | |
365 | && (ref->offset < 0 | |
366 | || (DECL_P (ref->base) | |
367 | && (!host_integerp (DECL_SIZE (ref->base), 1) | |
368 | || (TREE_INT_CST_LOW (DECL_SIZE ((ref->base))) | |
369 | < (unsigned HOST_WIDE_INT)(ref->offset + ref->size)))))) | |
370 | return false; | |
3a443843 | 371 | |
372 | return true; | |
373 | } | |
374 | ||
375 | /* Query the alias-oracle on whether the two memory rtx X and MEM may | |
376 | alias. If TBAA_P is set also apply TBAA. Returns true if the | |
377 | two rtxen may alias, false otherwise. */ | |
378 | ||
379 | static bool | |
380 | rtx_refs_may_alias_p (const_rtx x, const_rtx mem, bool tbaa_p) | |
381 | { | |
382 | ao_ref ref1, ref2; | |
383 | ||
384 | if (!ao_ref_from_mem (&ref1, x) | |
385 | || !ao_ref_from_mem (&ref2, mem)) | |
386 | return true; | |
387 | ||
33dda198 | 388 | return refs_may_alias_p_1 (&ref1, &ref2, |
389 | tbaa_p | |
390 | && MEM_ALIAS_SET (x) != 0 | |
391 | && MEM_ALIAS_SET (mem) != 0); | |
3a443843 | 392 | } |
393 | ||
a5b1863e | 394 | /* Returns a pointer to the alias set entry for ALIAS_SET, if there is |
395 | such an entry, or NULL otherwise. */ | |
396 | ||
56bbdce4 | 397 | static inline alias_set_entry |
32c2fdea | 398 | get_alias_set_entry (alias_set_type alias_set) |
a5b1863e | 399 | { |
8b435339 | 400 | return VEC_index (alias_set_entry, alias_sets, alias_set); |
a5b1863e | 401 | } |
402 | ||
5badbbc0 | 403 | /* Returns nonzero if the alias sets for MEM1 and MEM2 are such that |
404 | the two MEMs cannot alias each other. */ | |
a5b1863e | 405 | |
56bbdce4 | 406 | static inline int |
52d07779 | 407 | mems_in_disjoint_alias_sets_p (const_rtx mem1, const_rtx mem2) |
a5b1863e | 408 | { |
a5b1863e | 409 | /* Perform a basic sanity check. Namely, that there are no alias sets |
410 | if we're not using strict aliasing. This helps to catch bugs | |
411 | whereby someone uses PUT_CODE, but doesn't clear MEM_ALIAS_SET, or | |
412 | where a MEM is allocated in some way other than by the use of | |
413 | gen_rtx_MEM, and the MEM_ALIAS_SET is not cleared. If we begin to | |
414 | use alias sets to indicate that spilled registers cannot alias each | |
415 | other, we might need to remove this check. */ | |
64db345d | 416 | gcc_assert (flag_strict_aliasing |
417 | || (!MEM_ALIAS_SET (mem1) && !MEM_ALIAS_SET (mem2))); | |
a5b1863e | 418 | |
387bc205 | 419 | return ! alias_sets_conflict_p (MEM_ALIAS_SET (mem1), MEM_ALIAS_SET (mem2)); |
420 | } | |
a5b1863e | 421 | |
387bc205 | 422 | /* Insert the NODE into the splay tree given by DATA. Used by |
423 | record_alias_subset via splay_tree_foreach. */ | |
424 | ||
425 | static int | |
aecda0d6 | 426 | insert_subset_children (splay_tree_node node, void *data) |
387bc205 | 427 | { |
428 | splay_tree_insert ((splay_tree) data, node->key, node->value); | |
429 | ||
430 | return 0; | |
431 | } | |
432 | ||
7d1f52b2 | 433 | /* Return true if the first alias set is a subset of the second. */ |
434 | ||
435 | bool | |
32c2fdea | 436 | alias_set_subset_of (alias_set_type set1, alias_set_type set2) |
7d1f52b2 | 437 | { |
438 | alias_set_entry ase; | |
439 | ||
440 | /* Everything is a subset of the "aliases everything" set. */ | |
441 | if (set2 == 0) | |
442 | return true; | |
443 | ||
444 | /* Otherwise, check if set1 is a subset of set2. */ | |
445 | ase = get_alias_set_entry (set2); | |
446 | if (ase != 0 | |
a5b373b8 | 447 | && (ase->has_zero_child |
6c711ccf | 448 | || splay_tree_lookup (ase->children, |
449 | (splay_tree_key) set1))) | |
7d1f52b2 | 450 | return true; |
451 | return false; | |
452 | } | |
453 | ||
387bc205 | 454 | /* Return 1 if the two specified alias sets may conflict. */ |
455 | ||
456 | int | |
32c2fdea | 457 | alias_sets_conflict_p (alias_set_type set1, alias_set_type set2) |
387bc205 | 458 | { |
459 | alias_set_entry ase; | |
460 | ||
879f881c | 461 | /* The easy case. */ |
462 | if (alias_sets_must_conflict_p (set1, set2)) | |
387bc205 | 463 | return 1; |
a5b1863e | 464 | |
f7c44134 | 465 | /* See if the first alias set is a subset of the second. */ |
387bc205 | 466 | ase = get_alias_set_entry (set1); |
851dfbff | 467 | if (ase != 0 |
468 | && (ase->has_zero_child | |
469 | || splay_tree_lookup (ase->children, | |
387bc205 | 470 | (splay_tree_key) set2))) |
471 | return 1; | |
a5b1863e | 472 | |
473 | /* Now do the same, but with the alias sets reversed. */ | |
387bc205 | 474 | ase = get_alias_set_entry (set2); |
851dfbff | 475 | if (ase != 0 |
476 | && (ase->has_zero_child | |
477 | || splay_tree_lookup (ase->children, | |
387bc205 | 478 | (splay_tree_key) set1))) |
479 | return 1; | |
a5b1863e | 480 | |
387bc205 | 481 | /* The two alias sets are distinct and neither one is the |
879f881c | 482 | child of the other. Therefore, they cannot conflict. */ |
387bc205 | 483 | return 0; |
a5b1863e | 484 | } |
c1628b55 | 485 | |
879f881c | 486 | /* Return 1 if the two specified alias sets will always conflict. */ |
c1628b55 | 487 | |
488 | int | |
32c2fdea | 489 | alias_sets_must_conflict_p (alias_set_type set1, alias_set_type set2) |
c1628b55 | 490 | { |
491 | if (set1 == 0 || set2 == 0 || set1 == set2) | |
492 | return 1; | |
493 | ||
494 | return 0; | |
495 | } | |
496 | ||
387bc205 | 497 | /* Return 1 if any MEM object of type T1 will always conflict (using the |
498 | dependency routines in this file) with any MEM object of type T2. | |
499 | This is used when allocating temporary storage. If T1 and/or T2 are | |
500 | NULL_TREE, it means we know nothing about the storage. */ | |
501 | ||
502 | int | |
aecda0d6 | 503 | objects_must_conflict_p (tree t1, tree t2) |
387bc205 | 504 | { |
32c2fdea | 505 | alias_set_type set1, set2; |
2bedf233 | 506 | |
218eb8d0 | 507 | /* If neither has a type specified, we don't know if they'll conflict |
508 | because we may be using them to store objects of various types, for | |
509 | example the argument and local variables areas of inlined functions. */ | |
8f2ba9c9 | 510 | if (t1 == 0 && t2 == 0) |
218eb8d0 | 511 | return 0; |
512 | ||
387bc205 | 513 | /* If they are the same type, they must conflict. */ |
514 | if (t1 == t2 | |
515 | /* Likewise if both are volatile. */ | |
516 | || (t1 != 0 && TYPE_VOLATILE (t1) && t2 != 0 && TYPE_VOLATILE (t2))) | |
517 | return 1; | |
518 | ||
2bedf233 | 519 | set1 = t1 ? get_alias_set (t1) : 0; |
520 | set2 = t2 ? get_alias_set (t2) : 0; | |
387bc205 | 521 | |
879f881c | 522 | /* We can't use alias_sets_conflict_p because we must make sure |
523 | that every subtype of t1 will conflict with every subtype of | |
2bedf233 | 524 | t2 for which a pair of subobjects of these respective subtypes |
525 | overlaps on the stack. */ | |
879f881c | 526 | return alias_sets_must_conflict_p (set1, set2); |
387bc205 | 527 | } |
528 | \f | |
2b02580f | 529 | /* Return true if all nested component references handled by |
530 | get_inner_reference in T are such that we should use the alias set | |
531 | provided by the object at the heart of T. | |
532 | ||
533 | This is true for non-addressable components (which don't have their | |
534 | own alias set), as well as components of objects in alias set zero. | |
535 | This later point is a special case wherein we wish to override the | |
536 | alias set used by the component, but we don't have per-FIELD_DECL | |
537 | assignable alias sets. */ | |
538 | ||
539 | bool | |
1f1872fd | 540 | component_uses_parent_alias_set (const_tree t) |
bd32979f | 541 | { |
1f9b622b | 542 | while (1) |
543 | { | |
2b02580f | 544 | /* If we're at the end, it vacuously uses its own alias set. */ |
1f9b622b | 545 | if (!handled_component_p (t)) |
2b02580f | 546 | return false; |
1f9b622b | 547 | |
548 | switch (TREE_CODE (t)) | |
549 | { | |
550 | case COMPONENT_REF: | |
551 | if (DECL_NONADDRESSABLE_P (TREE_OPERAND (t, 1))) | |
2b02580f | 552 | return true; |
1f9b622b | 553 | break; |
554 | ||
555 | case ARRAY_REF: | |
556 | case ARRAY_RANGE_REF: | |
557 | if (TYPE_NONALIASED_COMPONENT (TREE_TYPE (TREE_OPERAND (t, 0)))) | |
2b02580f | 558 | return true; |
1f9b622b | 559 | break; |
560 | ||
561 | case REALPART_EXPR: | |
562 | case IMAGPART_EXPR: | |
563 | break; | |
564 | ||
565 | default: | |
566 | /* Bitfields and casts are never addressable. */ | |
2b02580f | 567 | return true; |
1f9b622b | 568 | } |
569 | ||
570 | t = TREE_OPERAND (t, 0); | |
2b02580f | 571 | if (get_alias_set (TREE_TYPE (t)) == 0) |
572 | return true; | |
1f9b622b | 573 | } |
bd32979f | 574 | } |
575 | ||
dd277d48 | 576 | /* Return the alias set for the memory pointed to by T, which may be |
577 | either a type or an expression. Return -1 if there is nothing | |
578 | special about dereferencing T. */ | |
579 | ||
580 | static alias_set_type | |
581 | get_deref_alias_set_1 (tree t) | |
582 | { | |
583 | /* If we're not doing any alias analysis, just assume everything | |
584 | aliases everything else. */ | |
585 | if (!flag_strict_aliasing) | |
586 | return 0; | |
587 | ||
8115f0af | 588 | /* All we care about is the type. */ |
dd277d48 | 589 | if (! TYPE_P (t)) |
8115f0af | 590 | t = TREE_TYPE (t); |
dd277d48 | 591 | |
592 | /* If we have an INDIRECT_REF via a void pointer, we don't | |
593 | know anything about what that might alias. Likewise if the | |
594 | pointer is marked that way. */ | |
595 | if (TREE_CODE (TREE_TYPE (t)) == VOID_TYPE | |
596 | || TYPE_REF_CAN_ALIAS_ALL (t)) | |
597 | return 0; | |
598 | ||
599 | return -1; | |
600 | } | |
601 | ||
602 | /* Return the alias set for the memory pointed to by T, which may be | |
603 | either a type or an expression. */ | |
604 | ||
605 | alias_set_type | |
606 | get_deref_alias_set (tree t) | |
607 | { | |
608 | alias_set_type set = get_deref_alias_set_1 (t); | |
609 | ||
610 | /* Fall back to the alias-set of the pointed-to type. */ | |
611 | if (set == -1) | |
612 | { | |
613 | if (! TYPE_P (t)) | |
614 | t = TREE_TYPE (t); | |
615 | set = get_alias_set (TREE_TYPE (t)); | |
616 | } | |
617 | ||
618 | return set; | |
619 | } | |
620 | ||
f7c44134 | 621 | /* Return the alias set for T, which may be either a type or an |
622 | expression. Call language-specific routine for help, if needed. */ | |
623 | ||
32c2fdea | 624 | alias_set_type |
aecda0d6 | 625 | get_alias_set (tree t) |
f7c44134 | 626 | { |
32c2fdea | 627 | alias_set_type set; |
f7c44134 | 628 | |
629 | /* If we're not doing any alias analysis, just assume everything | |
630 | aliases everything else. Also return 0 if this or its type is | |
631 | an error. */ | |
632 | if (! flag_strict_aliasing || t == error_mark_node | |
633 | || (! TYPE_P (t) | |
634 | && (TREE_TYPE (t) == 0 || TREE_TYPE (t) == error_mark_node))) | |
635 | return 0; | |
636 | ||
637 | /* We can be passed either an expression or a type. This and the | |
c3a9c149 | 638 | language-specific routine may make mutually-recursive calls to each other |
639 | to figure out what to do. At each juncture, we see if this is a tree | |
640 | that the language may need to handle specially. First handle things that | |
96216d37 | 641 | aren't types. */ |
1607663f | 642 | if (! TYPE_P (t)) |
f7c44134 | 643 | { |
38e63987 | 644 | tree inner; |
96216d37 | 645 | |
182cf5a9 | 646 | /* Give the language a chance to do something with this tree |
647 | before we look at it. */ | |
2a631e19 | 648 | STRIP_NOPS (t); |
dc24ddbd | 649 | set = lang_hooks.get_alias_set (t); |
2a631e19 | 650 | if (set != -1) |
651 | return set; | |
652 | ||
182cf5a9 | 653 | /* Get the base object of the reference. */ |
38e63987 | 654 | inner = t; |
55f9d7dc | 655 | while (handled_component_p (inner)) |
96216d37 | 656 | { |
182cf5a9 | 657 | /* If there is a VIEW_CONVERT_EXPR in the chain we cannot use |
658 | the type of any component references that wrap it to | |
659 | determine the alias-set. */ | |
660 | if (TREE_CODE (inner) == VIEW_CONVERT_EXPR) | |
661 | t = TREE_OPERAND (inner, 0); | |
55f9d7dc | 662 | inner = TREE_OPERAND (inner, 0); |
96216d37 | 663 | } |
664 | ||
182cf5a9 | 665 | /* Handle pointer dereferences here, they can override the |
666 | alias-set. */ | |
2a448a75 | 667 | if (INDIRECT_REF_P (inner)) |
96216d37 | 668 | { |
dd277d48 | 669 | set = get_deref_alias_set_1 (TREE_OPERAND (inner, 0)); |
670 | if (set != -1) | |
671 | return set; | |
96216d37 | 672 | } |
9a14ba4f | 673 | else if (TREE_CODE (inner) == TARGET_MEM_REF) |
674 | return get_deref_alias_set (TMR_OFFSET (inner)); | |
182cf5a9 | 675 | else if (TREE_CODE (inner) == MEM_REF) |
676 | { | |
677 | set = get_deref_alias_set_1 (TREE_OPERAND (inner, 1)); | |
678 | if (set != -1) | |
679 | return set; | |
680 | } | |
681 | ||
682 | /* If the innermost reference is a MEM_REF that has a | |
683 | conversion embedded treat it like a VIEW_CONVERT_EXPR above, | |
684 | using the memory access type for determining the alias-set. */ | |
685 | if (TREE_CODE (inner) == MEM_REF | |
ec7b61e0 | 686 | && TYPE_MAIN_VARIANT (TREE_TYPE (inner)) |
687 | != TYPE_MAIN_VARIANT | |
688 | (TREE_TYPE (TREE_TYPE (TREE_OPERAND (inner, 1))))) | |
182cf5a9 | 689 | return get_deref_alias_set (TREE_OPERAND (inner, 1)); |
96216d37 | 690 | |
691 | /* Otherwise, pick up the outermost object that we could have a pointer | |
55f9d7dc | 692 | to, processing conversions as above. */ |
2b02580f | 693 | while (component_uses_parent_alias_set (t)) |
c3a9c149 | 694 | { |
55f9d7dc | 695 | t = TREE_OPERAND (t, 0); |
2a631e19 | 696 | STRIP_NOPS (t); |
697 | } | |
1607663f | 698 | |
96216d37 | 699 | /* If we've already determined the alias set for a decl, just return |
700 | it. This is necessary for C++ anonymous unions, whose component | |
701 | variables don't look like union members (boo!). */ | |
afce1f57 | 702 | if (TREE_CODE (t) == VAR_DECL |
e16ceb8e | 703 | && DECL_RTL_SET_P (t) && MEM_P (DECL_RTL (t))) |
afce1f57 | 704 | return MEM_ALIAS_SET (DECL_RTL (t)); |
705 | ||
1607663f | 706 | /* Now all we care about is the type. */ |
707 | t = TREE_TYPE (t); | |
f7c44134 | 708 | } |
709 | ||
f7c44134 | 710 | /* Variant qualifiers don't affect the alias set, so get the main |
dfe8e806 | 711 | variant. */ |
f7c44134 | 712 | t = TYPE_MAIN_VARIANT (t); |
dfe8e806 | 713 | |
714 | /* Always use the canonical type as well. If this is a type that | |
715 | requires structural comparisons to identify compatible types | |
716 | use alias set zero. */ | |
717 | if (TYPE_STRUCTURAL_EQUALITY_P (t)) | |
e58c17e7 | 718 | { |
719 | /* Allow the language to specify another alias set for this | |
720 | type. */ | |
721 | set = lang_hooks.get_alias_set (t); | |
722 | if (set != -1) | |
723 | return set; | |
724 | return 0; | |
725 | } | |
ec7b61e0 | 726 | |
dfe8e806 | 727 | t = TYPE_CANONICAL (t); |
ec7b61e0 | 728 | |
45014c84 | 729 | /* The canonical type should not require structural equality checks. */ |
730 | gcc_checking_assert (!TYPE_STRUCTURAL_EQUALITY_P (t)); | |
dfe8e806 | 731 | |
732 | /* If this is a type with a known alias set, return it. */ | |
96216d37 | 733 | if (TYPE_ALIAS_SET_KNOWN_P (t)) |
f7c44134 | 734 | return TYPE_ALIAS_SET (t); |
735 | ||
b35a8f48 | 736 | /* We don't want to set TYPE_ALIAS_SET for incomplete types. */ |
737 | if (!COMPLETE_TYPE_P (t)) | |
738 | { | |
739 | /* For arrays with unknown size the conservative answer is the | |
740 | alias set of the element type. */ | |
741 | if (TREE_CODE (t) == ARRAY_TYPE) | |
742 | return get_alias_set (TREE_TYPE (t)); | |
743 | ||
744 | /* But return zero as a conservative answer for incomplete types. */ | |
745 | return 0; | |
746 | } | |
747 | ||
f7c44134 | 748 | /* See if the language has special handling for this type. */ |
dc24ddbd | 749 | set = lang_hooks.get_alias_set (t); |
2a631e19 | 750 | if (set != -1) |
96216d37 | 751 | return set; |
851dfbff | 752 | |
f7c44134 | 753 | /* There are no objects of FUNCTION_TYPE, so there's no point in |
754 | using up an alias set for them. (There are, of course, pointers | |
755 | and references to functions, but that's different.) */ | |
ec7b61e0 | 756 | else if (TREE_CODE (t) == FUNCTION_TYPE || TREE_CODE (t) == METHOD_TYPE) |
f7c44134 | 757 | set = 0; |
9fcc3e54 | 758 | |
759 | /* Unless the language specifies otherwise, let vector types alias | |
760 | their components. This avoids some nasty type punning issues in | |
761 | normal usage. And indeed lets vectors be treated more like an | |
762 | array slice. */ | |
763 | else if (TREE_CODE (t) == VECTOR_TYPE) | |
764 | set = get_alias_set (TREE_TYPE (t)); | |
765 | ||
62d823d0 | 766 | /* Unless the language specifies otherwise, treat array types the |
767 | same as their components. This avoids the asymmetry we get | |
768 | through recording the components. Consider accessing a | |
769 | character(kind=1) through a reference to a character(kind=1)[1:1]. | |
770 | Or consider if we want to assign integer(kind=4)[0:D.1387] and | |
771 | integer(kind=4)[4] the same alias set or not. | |
772 | Just be pragmatic here and make sure the array and its element | |
773 | type get the same alias set assigned. */ | |
ec7b61e0 | 774 | else if (TREE_CODE (t) == ARRAY_TYPE && !TYPE_NONALIASED_COMPONENT (t)) |
62d823d0 | 775 | set = get_alias_set (TREE_TYPE (t)); |
776 | ||
2af087f2 | 777 | /* From the former common C and C++ langhook implementation: |
778 | ||
779 | Unfortunately, there is no canonical form of a pointer type. | |
780 | In particular, if we have `typedef int I', then `int *', and | |
781 | `I *' are different types. So, we have to pick a canonical | |
782 | representative. We do this below. | |
783 | ||
784 | Technically, this approach is actually more conservative that | |
785 | it needs to be. In particular, `const int *' and `int *' | |
786 | should be in different alias sets, according to the C and C++ | |
787 | standard, since their types are not the same, and so, | |
788 | technically, an `int **' and `const int **' cannot point at | |
789 | the same thing. | |
790 | ||
791 | But, the standard is wrong. In particular, this code is | |
792 | legal C++: | |
793 | ||
794 | int *ip; | |
795 | int **ipp = &ip; | |
796 | const int* const* cipp = ipp; | |
797 | And, it doesn't make sense for that to be legal unless you | |
798 | can dereference IPP and CIPP. So, we ignore cv-qualifiers on | |
799 | the pointed-to types. This issue has been reported to the | |
800 | C++ committee. | |
801 | ||
802 | In addition to the above canonicalization issue, with LTO | |
803 | we should also canonicalize `T (*)[]' to `T *' avoiding | |
804 | alias issues with pointer-to element types and pointer-to | |
805 | array types. | |
806 | ||
807 | Likewise we need to deal with the situation of incomplete | |
808 | pointed-to types and make `*(struct X **)&a' and | |
809 | `*(struct X {} **)&a' alias. Otherwise we will have to | |
810 | guarantee that all pointer-to incomplete type variants | |
811 | will be replaced by pointer-to complete type variants if | |
812 | they are available. | |
813 | ||
814 | With LTO the convenient situation of using `void *' to | |
815 | access and store any pointer type will also become | |
816 | more apparent (and `void *' is just another pointer-to | |
817 | incomplete type). Assigning alias-set zero to `void *' | |
818 | and all pointer-to incomplete types is a not appealing | |
819 | solution. Assigning an effective alias-set zero only | |
820 | affecting pointers might be - by recording proper subset | |
821 | relationships of all pointer alias-sets. | |
822 | ||
823 | Pointer-to function types are another grey area which | |
824 | needs caution. Globbing them all into one alias-set | |
825 | or the above effective zero set would work. | |
826 | ||
827 | For now just assign the same alias-set to all pointers. | |
828 | That's simple and avoids all the above problems. */ | |
829 | else if (POINTER_TYPE_P (t) | |
830 | && t != ptr_type_node) | |
45014c84 | 831 | set = get_alias_set (ptr_type_node); |
2af087f2 | 832 | |
ec7b61e0 | 833 | /* Otherwise make a new alias set for this type. */ |
f7c44134 | 834 | else |
45014c84 | 835 | { |
836 | /* Each canonical type gets its own alias set, so canonical types | |
837 | shouldn't form a tree. It doesn't really matter for types | |
838 | we handle specially above, so only check it where it possibly | |
839 | would result in a bogus alias set. */ | |
840 | gcc_checking_assert (TYPE_CANONICAL (t) == t); | |
841 | ||
842 | set = new_alias_set (); | |
843 | } | |
f7c44134 | 844 | |
845 | TYPE_ALIAS_SET (t) = set; | |
851dfbff | 846 | |
ec7b61e0 | 847 | /* If this is an aggregate type or a complex type, we must record any |
848 | component aliasing information. */ | |
680210f2 | 849 | if (AGGREGATE_TYPE_P (t) || TREE_CODE (t) == COMPLEX_TYPE) |
851dfbff | 850 | record_component_aliases (t); |
851 | ||
f7c44134 | 852 | return set; |
853 | } | |
854 | ||
855 | /* Return a brand-new alias set. */ | |
856 | ||
32c2fdea | 857 | alias_set_type |
aecda0d6 | 858 | new_alias_set (void) |
f7c44134 | 859 | { |
f7c44134 | 860 | if (flag_strict_aliasing) |
56bbdce4 | 861 | { |
8b435339 | 862 | if (alias_sets == 0) |
863 | VEC_safe_push (alias_set_entry, gc, alias_sets, 0); | |
864 | VEC_safe_push (alias_set_entry, gc, alias_sets, 0); | |
865 | return VEC_length (alias_set_entry, alias_sets) - 1; | |
56bbdce4 | 866 | } |
f7c44134 | 867 | else |
868 | return 0; | |
869 | } | |
a5b1863e | 870 | |
8e2ef041 | 871 | /* Indicate that things in SUBSET can alias things in SUPERSET, but that |
872 | not everything that aliases SUPERSET also aliases SUBSET. For example, | |
873 | in C, a store to an `int' can alias a load of a structure containing an | |
874 | `int', and vice versa. But it can't alias a load of a 'double' member | |
875 | of the same structure. Here, the structure would be the SUPERSET and | |
876 | `int' the SUBSET. This relationship is also described in the comment at | |
877 | the beginning of this file. | |
878 | ||
879 | This function should be called only once per SUPERSET/SUBSET pair. | |
a5b1863e | 880 | |
881 | It is illegal for SUPERSET to be zero; everything is implicitly a | |
882 | subset of alias set zero. */ | |
883 | ||
892fdb03 | 884 | void |
32c2fdea | 885 | record_alias_subset (alias_set_type superset, alias_set_type subset) |
a5b1863e | 886 | { |
887 | alias_set_entry superset_entry; | |
888 | alias_set_entry subset_entry; | |
889 | ||
c3a9c149 | 890 | /* It is possible in complex type situations for both sets to be the same, |
891 | in which case we can ignore this operation. */ | |
892 | if (superset == subset) | |
893 | return; | |
894 | ||
64db345d | 895 | gcc_assert (superset); |
a5b1863e | 896 | |
897 | superset_entry = get_alias_set_entry (superset); | |
9342ee68 | 898 | if (superset_entry == 0) |
a5b1863e | 899 | { |
900 | /* Create an entry for the SUPERSET, so that we have a place to | |
901 | attach the SUBSET. */ | |
ba72912a | 902 | superset_entry = ggc_alloc_cleared_alias_set_entry_d (); |
a5b1863e | 903 | superset_entry->alias_set = superset; |
9342ee68 | 904 | superset_entry->children |
ba72912a | 905 | = splay_tree_new_ggc (splay_tree_compare_ints, |
906 | ggc_alloc_splay_tree_scalar_scalar_splay_tree_s, | |
907 | ggc_alloc_splay_tree_scalar_scalar_splay_tree_node_s); | |
f2e5de80 | 908 | superset_entry->has_zero_child = 0; |
8b435339 | 909 | VEC_replace (alias_set_entry, alias_sets, superset, superset_entry); |
a5b1863e | 910 | } |
911 | ||
851dfbff | 912 | if (subset == 0) |
913 | superset_entry->has_zero_child = 1; | |
914 | else | |
915 | { | |
916 | subset_entry = get_alias_set_entry (subset); | |
917 | /* If there is an entry for the subset, enter all of its children | |
918 | (if they are not already present) as children of the SUPERSET. */ | |
9342ee68 | 919 | if (subset_entry) |
851dfbff | 920 | { |
921 | if (subset_entry->has_zero_child) | |
922 | superset_entry->has_zero_child = 1; | |
083a2b5e | 923 | |
851dfbff | 924 | splay_tree_foreach (subset_entry->children, insert_subset_children, |
925 | superset_entry->children); | |
926 | } | |
a5b1863e | 927 | |
851dfbff | 928 | /* Enter the SUBSET itself as a child of the SUPERSET. */ |
9342ee68 | 929 | splay_tree_insert (superset_entry->children, |
851dfbff | 930 | (splay_tree_key) subset, 0); |
931 | } | |
a5b1863e | 932 | } |
933 | ||
9501a338 | 934 | /* Record that component types of TYPE, if any, are part of that type for |
935 | aliasing purposes. For record types, we only record component types | |
c67bf066 | 936 | for fields that are not marked non-addressable. For array types, we |
937 | only record the component type if it is not marked non-aliased. */ | |
9501a338 | 938 | |
939 | void | |
aecda0d6 | 940 | record_component_aliases (tree type) |
9501a338 | 941 | { |
32c2fdea | 942 | alias_set_type superset = get_alias_set (type); |
9501a338 | 943 | tree field; |
944 | ||
945 | if (superset == 0) | |
946 | return; | |
947 | ||
948 | switch (TREE_CODE (type)) | |
949 | { | |
9501a338 | 950 | case RECORD_TYPE: |
951 | case UNION_TYPE: | |
952 | case QUAL_UNION_TYPE: | |
8b332087 | 953 | /* Recursively record aliases for the base classes, if there are any. */ |
f6cc6a08 | 954 | if (TYPE_BINFO (type)) |
9342ee68 | 955 | { |
956 | int i; | |
f6cc6a08 | 957 | tree binfo, base_binfo; |
a0c938f0 | 958 | |
f6cc6a08 | 959 | for (binfo = TYPE_BINFO (type), i = 0; |
960 | BINFO_BASE_ITERATE (binfo, i, base_binfo); i++) | |
961 | record_alias_subset (superset, | |
962 | get_alias_set (BINFO_TYPE (base_binfo))); | |
9342ee68 | 963 | } |
1767a056 | 964 | for (field = TYPE_FIELDS (type); field != 0; field = DECL_CHAIN (field)) |
c67bf066 | 965 | if (TREE_CODE (field) == FIELD_DECL && !DECL_NONADDRESSABLE_P (field)) |
851dfbff | 966 | record_alias_subset (superset, get_alias_set (TREE_TYPE (field))); |
9501a338 | 967 | break; |
968 | ||
680210f2 | 969 | case COMPLEX_TYPE: |
970 | record_alias_subset (superset, get_alias_set (TREE_TYPE (type))); | |
971 | break; | |
972 | ||
62d823d0 | 973 | /* VECTOR_TYPE and ARRAY_TYPE share the alias set with their |
974 | element type. */ | |
975 | ||
9501a338 | 976 | default: |
977 | break; | |
978 | } | |
979 | } | |
980 | ||
f7c44134 | 981 | /* Allocate an alias set for use in storing and reading from the varargs |
982 | spill area. */ | |
983 | ||
32c2fdea | 984 | static GTY(()) alias_set_type varargs_set = -1; |
4ec9406e | 985 | |
32c2fdea | 986 | alias_set_type |
aecda0d6 | 987 | get_varargs_alias_set (void) |
f7c44134 | 988 | { |
fcdd3ab3 | 989 | #if 1 |
990 | /* We now lower VA_ARG_EXPR, and there's currently no way to attach the | |
991 | varargs alias set to an INDIRECT_REF (FIXME!), so we can't | |
992 | consistently use the varargs alias set for loads from the varargs | |
993 | area. So don't use it anywhere. */ | |
994 | return 0; | |
995 | #else | |
4ec9406e | 996 | if (varargs_set == -1) |
997 | varargs_set = new_alias_set (); | |
f7c44134 | 998 | |
4ec9406e | 999 | return varargs_set; |
fcdd3ab3 | 1000 | #endif |
f7c44134 | 1001 | } |
1002 | ||
1003 | /* Likewise, but used for the fixed portions of the frame, e.g., register | |
1004 | save areas. */ | |
1005 | ||
32c2fdea | 1006 | static GTY(()) alias_set_type frame_set = -1; |
4ec9406e | 1007 | |
32c2fdea | 1008 | alias_set_type |
aecda0d6 | 1009 | get_frame_alias_set (void) |
f7c44134 | 1010 | { |
4ec9406e | 1011 | if (frame_set == -1) |
1012 | frame_set = new_alias_set (); | |
f7c44134 | 1013 | |
4ec9406e | 1014 | return frame_set; |
f7c44134 | 1015 | } |
1016 | ||
4e56ceb1 | 1017 | /* Create a new, unique base with id ID. */ |
1018 | ||
1019 | static rtx | |
1020 | unique_base_value (HOST_WIDE_INT id) | |
1021 | { | |
1022 | return gen_rtx_ADDRESS (Pmode, id); | |
1023 | } | |
1024 | ||
1025 | /* Return true if accesses based on any other base value cannot alias | |
1026 | those based on X. */ | |
1027 | ||
1028 | static bool | |
1029 | unique_base_value_p (rtx x) | |
1030 | { | |
1031 | return GET_CODE (x) == ADDRESS && GET_MODE (x) == Pmode; | |
1032 | } | |
1033 | ||
1034 | /* Return true if X is known to be a base value. */ | |
1035 | ||
1036 | static bool | |
1037 | known_base_value_p (rtx x) | |
1038 | { | |
1039 | switch (GET_CODE (x)) | |
1040 | { | |
1041 | case LABEL_REF: | |
1042 | case SYMBOL_REF: | |
1043 | return true; | |
1044 | ||
1045 | case ADDRESS: | |
1046 | /* Arguments may or may not be bases; we don't know for sure. */ | |
1047 | return GET_MODE (x) != VOIDmode; | |
1048 | ||
1049 | default: | |
1050 | return false; | |
1051 | } | |
1052 | } | |
1053 | ||
e7c893ba | 1054 | /* Inside SRC, the source of a SET, find a base address. */ |
1055 | ||
ea0cb7ae | 1056 | static rtx |
aecda0d6 | 1057 | find_base_value (rtx src) |
ea0cb7ae | 1058 | { |
970c0c06 | 1059 | unsigned int regno; |
d01410d3 | 1060 | |
2e984a7a | 1061 | #if defined (FIND_BASE_TERM) |
1062 | /* Try machine-dependent ways to find the base term. */ | |
1063 | src = FIND_BASE_TERM (src); | |
1064 | #endif | |
1065 | ||
ea0cb7ae | 1066 | switch (GET_CODE (src)) |
1067 | { | |
1068 | case SYMBOL_REF: | |
1069 | case LABEL_REF: | |
1070 | return src; | |
1071 | ||
1072 | case REG: | |
9db10305 | 1073 | regno = REGNO (src); |
083a2b5e | 1074 | /* At the start of a function, argument registers have known base |
e7c893ba | 1075 | values which may be lost later. Returning an ADDRESS |
1076 | expression here allows optimization based on argument values | |
1077 | even when the argument registers are used for other purposes. */ | |
970c0c06 | 1078 | if (regno < FIRST_PSEUDO_REGISTER && copying_arguments) |
1079 | return new_reg_base_value[regno]; | |
1a538f67 | 1080 | |
ebd13eec | 1081 | /* If a pseudo has a known base value, return it. Do not do this |
9464d436 | 1082 | for non-fixed hard regs since it can result in a circular |
1083 | dependency chain for registers which have values at function entry. | |
ebd13eec | 1084 | |
1085 | The test above is not sufficient because the scheduler may move | |
1086 | a copy out of an arg reg past the NOTE_INSN_FUNCTION_BEGIN. */ | |
9464d436 | 1087 | if ((regno >= FIRST_PSEUDO_REGISTER || fixed_regs[regno]) |
980f3285 | 1088 | && regno < VEC_length (rtx, reg_base_value)) |
e5be0cc6 | 1089 | { |
1090 | /* If we're inside init_alias_analysis, use new_reg_base_value | |
1091 | to reduce the number of relaxation iterations. */ | |
a303ed8c | 1092 | if (new_reg_base_value && new_reg_base_value[regno] |
3072d30e | 1093 | && DF_REG_DEF_COUNT (regno) == 1) |
e5be0cc6 | 1094 | return new_reg_base_value[regno]; |
1095 | ||
980f3285 | 1096 | if (VEC_index (rtx, reg_base_value, regno)) |
1097 | return VEC_index (rtx, reg_base_value, regno); | |
e5be0cc6 | 1098 | } |
1a538f67 | 1099 | |
1c5cb56c | 1100 | return 0; |
ea0cb7ae | 1101 | |
1102 | case MEM: | |
1103 | /* Check for an argument passed in memory. Only record in the | |
1104 | copying-arguments block; it is too hard to track changes | |
1105 | otherwise. */ | |
1106 | if (copying_arguments | |
1107 | && (XEXP (src, 0) == arg_pointer_rtx | |
1108 | || (GET_CODE (XEXP (src, 0)) == PLUS | |
1109 | && XEXP (XEXP (src, 0), 0) == arg_pointer_rtx))) | |
4e56ceb1 | 1110 | return arg_base_value; |
ea0cb7ae | 1111 | return 0; |
1112 | ||
1113 | case CONST: | |
1114 | src = XEXP (src, 0); | |
1115 | if (GET_CODE (src) != PLUS && GET_CODE (src) != MINUS) | |
1116 | break; | |
083a2b5e | 1117 | |
2c0e001b | 1118 | /* ... fall through ... */ |
e7c893ba | 1119 | |
ea0cb7ae | 1120 | case PLUS: |
1121 | case MINUS: | |
e7c893ba | 1122 | { |
06c5c535 | 1123 | rtx temp, src_0 = XEXP (src, 0), src_1 = XEXP (src, 1); |
1124 | ||
560de2f8 | 1125 | /* If either operand is a REG that is a known pointer, then it |
1126 | is the base. */ | |
1127 | if (REG_P (src_0) && REG_POINTER (src_0)) | |
1128 | return find_base_value (src_0); | |
1129 | if (REG_P (src_1) && REG_POINTER (src_1)) | |
1130 | return find_base_value (src_1); | |
1131 | ||
06c5c535 | 1132 | /* If either operand is a REG, then see if we already have |
1133 | a known value for it. */ | |
560de2f8 | 1134 | if (REG_P (src_0)) |
06c5c535 | 1135 | { |
1136 | temp = find_base_value (src_0); | |
083a2b5e | 1137 | if (temp != 0) |
06c5c535 | 1138 | src_0 = temp; |
1139 | } | |
1140 | ||
560de2f8 | 1141 | if (REG_P (src_1)) |
06c5c535 | 1142 | { |
1143 | temp = find_base_value (src_1); | |
083a2b5e | 1144 | if (temp!= 0) |
06c5c535 | 1145 | src_1 = temp; |
1146 | } | |
e7c893ba | 1147 | |
560de2f8 | 1148 | /* If either base is named object or a special address |
1149 | (like an argument or stack reference), then use it for the | |
1150 | base term. */ | |
4e56ceb1 | 1151 | if (src_0 != 0 && known_base_value_p (src_0)) |
560de2f8 | 1152 | return src_0; |
1153 | ||
4e56ceb1 | 1154 | if (src_1 != 0 && known_base_value_p (src_1)) |
560de2f8 | 1155 | return src_1; |
1156 | ||
083a2b5e | 1157 | /* Guess which operand is the base address: |
06c5c535 | 1158 | If either operand is a symbol, then it is the base. If |
1159 | either operand is a CONST_INT, then the other is the base. */ | |
971ba038 | 1160 | if (CONST_INT_P (src_1) || CONSTANT_P (src_0)) |
e7c893ba | 1161 | return find_base_value (src_0); |
971ba038 | 1162 | else if (CONST_INT_P (src_0) || CONSTANT_P (src_1)) |
06c5c535 | 1163 | return find_base_value (src_1); |
1164 | ||
ea0cb7ae | 1165 | return 0; |
e7c893ba | 1166 | } |
1167 | ||
1168 | case LO_SUM: | |
1169 | /* The standard form is (lo_sum reg sym) so look only at the | |
1170 | second operand. */ | |
1171 | return find_base_value (XEXP (src, 1)); | |
ea0cb7ae | 1172 | |
1173 | case AND: | |
1174 | /* If the second operand is constant set the base | |
2c0e001b | 1175 | address to the first operand. */ |
971ba038 | 1176 | if (CONST_INT_P (XEXP (src, 1)) && INTVAL (XEXP (src, 1)) != 0) |
e7c893ba | 1177 | return find_base_value (XEXP (src, 0)); |
ea0cb7ae | 1178 | return 0; |
1179 | ||
43d45282 | 1180 | case TRUNCATE: |
04ec15fa | 1181 | /* As we do not know which address space the pointer is referring to, we can |
98155838 | 1182 | handle this only if the target does not support different pointer or |
1183 | address modes depending on the address space. */ | |
1184 | if (!target_default_pointer_address_modes_p ()) | |
1185 | break; | |
43d45282 | 1186 | if (GET_MODE_SIZE (GET_MODE (src)) < GET_MODE_SIZE (Pmode)) |
1187 | break; | |
1188 | /* Fall through. */ | |
ea0cb7ae | 1189 | case HIGH: |
2be64ef0 | 1190 | case PRE_INC: |
1191 | case PRE_DEC: | |
1192 | case POST_INC: | |
1193 | case POST_DEC: | |
1194 | case PRE_MODIFY: | |
1195 | case POST_MODIFY: | |
e7c893ba | 1196 | return find_base_value (XEXP (src, 0)); |
99c14947 | 1197 | |
d01410d3 | 1198 | case ZERO_EXTEND: |
1199 | case SIGN_EXTEND: /* used for NT/Alpha pointers */ | |
04ec15fa | 1200 | /* As we do not know which address space the pointer is referring to, we can |
98155838 | 1201 | handle this only if the target does not support different pointer or |
1202 | address modes depending on the address space. */ | |
1203 | if (!target_default_pointer_address_modes_p ()) | |
1204 | break; | |
1205 | ||
d01410d3 | 1206 | { |
1207 | rtx temp = find_base_value (XEXP (src, 0)); | |
1208 | ||
85d654dd | 1209 | if (temp != 0 && CONSTANT_P (temp)) |
d01410d3 | 1210 | temp = convert_memory_address (Pmode, temp); |
d01410d3 | 1211 | |
1212 | return temp; | |
1213 | } | |
1214 | ||
99c14947 | 1215 | default: |
1216 | break; | |
ea0cb7ae | 1217 | } |
1218 | ||
1219 | return 0; | |
1220 | } | |
1221 | ||
4e56ceb1 | 1222 | /* Called from init_alias_analysis indirectly through note_stores, |
1223 | or directly if DEST is a register with a REG_NOALIAS note attached. | |
1224 | SET is null in the latter case. */ | |
ea0cb7ae | 1225 | |
083a2b5e | 1226 | /* While scanning insns to find base values, reg_seen[N] is nonzero if |
ea0cb7ae | 1227 | register N has been set in this function. */ |
1228 | static char *reg_seen; | |
1229 | ||
e7c893ba | 1230 | static void |
81a410b1 | 1231 | record_set (rtx dest, const_rtx set, void *data ATTRIBUTE_UNUSED) |
ea0cb7ae | 1232 | { |
19cb6b50 | 1233 | unsigned regno; |
ea0cb7ae | 1234 | rtx src; |
3721f260 | 1235 | int n; |
ea0cb7ae | 1236 | |
8ad4c111 | 1237 | if (!REG_P (dest)) |
ea0cb7ae | 1238 | return; |
1239 | ||
9db10305 | 1240 | regno = REGNO (dest); |
ea0cb7ae | 1241 | |
0ea2d350 | 1242 | gcc_checking_assert (regno < VEC_length (rtx, reg_base_value)); |
c2e94b25 | 1243 | |
3721f260 | 1244 | /* If this spans multiple hard registers, then we must indicate that every |
1245 | register has an unusable value. */ | |
1246 | if (regno < FIRST_PSEUDO_REGISTER) | |
67d6c12b | 1247 | n = hard_regno_nregs[regno][GET_MODE (dest)]; |
3721f260 | 1248 | else |
1249 | n = 1; | |
1250 | if (n != 1) | |
1251 | { | |
1252 | while (--n >= 0) | |
1253 | { | |
1254 | reg_seen[regno + n] = 1; | |
1255 | new_reg_base_value[regno + n] = 0; | |
1256 | } | |
1257 | return; | |
1258 | } | |
1259 | ||
ea0cb7ae | 1260 | if (set) |
1261 | { | |
1262 | /* A CLOBBER wipes out any old value but does not prevent a previously | |
1263 | unset register from acquiring a base address (i.e. reg_seen is not | |
1264 | set). */ | |
1265 | if (GET_CODE (set) == CLOBBER) | |
1266 | { | |
06c5c535 | 1267 | new_reg_base_value[regno] = 0; |
ea0cb7ae | 1268 | return; |
1269 | } | |
1270 | src = SET_SRC (set); | |
1271 | } | |
1272 | else | |
1273 | { | |
4e56ceb1 | 1274 | /* There's a REG_NOALIAS note against DEST. */ |
ea0cb7ae | 1275 | if (reg_seen[regno]) |
1276 | { | |
06c5c535 | 1277 | new_reg_base_value[regno] = 0; |
ea0cb7ae | 1278 | return; |
1279 | } | |
1280 | reg_seen[regno] = 1; | |
4e56ceb1 | 1281 | new_reg_base_value[regno] = unique_base_value (unique_id++); |
ea0cb7ae | 1282 | return; |
1283 | } | |
1284 | ||
ad6b1fe1 | 1285 | /* If this is not the first set of REGNO, see whether the new value |
1286 | is related to the old one. There are two cases of interest: | |
1287 | ||
1288 | (1) The register might be assigned an entirely new value | |
1289 | that has the same base term as the original set. | |
1290 | ||
1291 | (2) The set might be a simple self-modification that | |
1292 | cannot change REGNO's base value. | |
1293 | ||
1294 | If neither case holds, reject the original base value as invalid. | |
1295 | Note that the following situation is not detected: | |
1296 | ||
a0c938f0 | 1297 | extern int x, y; int *p = &x; p += (&y-&x); |
ad6b1fe1 | 1298 | |
ea0cb7ae | 1299 | ANSI C does not allow computing the difference of addresses |
1300 | of distinct top level objects. */ | |
ad6b1fe1 | 1301 | if (new_reg_base_value[regno] != 0 |
1302 | && find_base_value (src) != new_reg_base_value[regno]) | |
ea0cb7ae | 1303 | switch (GET_CODE (src)) |
1304 | { | |
e7c893ba | 1305 | case LO_SUM: |
ea0cb7ae | 1306 | case MINUS: |
1307 | if (XEXP (src, 0) != dest && XEXP (src, 1) != dest) | |
06c5c535 | 1308 | new_reg_base_value[regno] = 0; |
ea0cb7ae | 1309 | break; |
43d45282 | 1310 | case PLUS: |
1311 | /* If the value we add in the PLUS is also a valid base value, | |
1312 | this might be the actual base value, and the original value | |
1313 | an index. */ | |
1314 | { | |
1315 | rtx other = NULL_RTX; | |
1316 | ||
1317 | if (XEXP (src, 0) == dest) | |
1318 | other = XEXP (src, 1); | |
1319 | else if (XEXP (src, 1) == dest) | |
1320 | other = XEXP (src, 0); | |
1321 | ||
1322 | if (! other || find_base_value (other)) | |
1323 | new_reg_base_value[regno] = 0; | |
1324 | break; | |
1325 | } | |
ea0cb7ae | 1326 | case AND: |
971ba038 | 1327 | if (XEXP (src, 0) != dest || !CONST_INT_P (XEXP (src, 1))) |
06c5c535 | 1328 | new_reg_base_value[regno] = 0; |
ea0cb7ae | 1329 | break; |
ea0cb7ae | 1330 | default: |
06c5c535 | 1331 | new_reg_base_value[regno] = 0; |
ea0cb7ae | 1332 | break; |
1333 | } | |
1334 | /* If this is the first set of a register, record the value. */ | |
1335 | else if ((regno >= FIRST_PSEUDO_REGISTER || ! fixed_regs[regno]) | |
06c5c535 | 1336 | && ! reg_seen[regno] && new_reg_base_value[regno] == 0) |
1337 | new_reg_base_value[regno] = find_base_value (src); | |
ea0cb7ae | 1338 | |
1339 | reg_seen[regno] = 1; | |
1340 | } | |
1341 | ||
0cfef6c5 | 1342 | /* Return REG_BASE_VALUE for REGNO. Selective scheduler uses this to avoid |
1343 | using hard registers with non-null REG_BASE_VALUE for renaming. */ | |
1344 | rtx | |
1345 | get_reg_base_value (unsigned int regno) | |
1346 | { | |
1347 | return VEC_index (rtx, reg_base_value, regno); | |
1348 | } | |
1349 | ||
12cb06ad | 1350 | /* If a value is known for REGNO, return it. */ |
1351 | ||
a0c938f0 | 1352 | rtx |
12cb06ad | 1353 | get_reg_known_value (unsigned int regno) |
1354 | { | |
1355 | if (regno >= FIRST_PSEUDO_REGISTER) | |
1356 | { | |
1357 | regno -= FIRST_PSEUDO_REGISTER; | |
1358 | if (regno < reg_known_value_size) | |
1359 | return reg_known_value[regno]; | |
1360 | } | |
1361 | return NULL; | |
9f5e1a5a | 1362 | } |
1363 | ||
12cb06ad | 1364 | /* Set it. */ |
1365 | ||
1366 | static void | |
1367 | set_reg_known_value (unsigned int regno, rtx val) | |
1368 | { | |
1369 | if (regno >= FIRST_PSEUDO_REGISTER) | |
1370 | { | |
1371 | regno -= FIRST_PSEUDO_REGISTER; | |
1372 | if (regno < reg_known_value_size) | |
1373 | reg_known_value[regno] = val; | |
1374 | } | |
1375 | } | |
1376 | ||
1377 | /* Similarly for reg_known_equiv_p. */ | |
1378 | ||
1379 | bool | |
1380 | get_reg_known_equiv_p (unsigned int regno) | |
1381 | { | |
1382 | if (regno >= FIRST_PSEUDO_REGISTER) | |
1383 | { | |
1384 | regno -= FIRST_PSEUDO_REGISTER; | |
1385 | if (regno < reg_known_value_size) | |
1386 | return reg_known_equiv_p[regno]; | |
1387 | } | |
1388 | return false; | |
1389 | } | |
1390 | ||
1391 | static void | |
1392 | set_reg_known_equiv_p (unsigned int regno, bool val) | |
1393 | { | |
1394 | if (regno >= FIRST_PSEUDO_REGISTER) | |
1395 | { | |
1396 | regno -= FIRST_PSEUDO_REGISTER; | |
1397 | if (regno < reg_known_value_size) | |
1398 | reg_known_equiv_p[regno] = val; | |
1399 | } | |
1400 | } | |
1401 | ||
1402 | ||
7cfb9bcf | 1403 | /* Returns a canonical version of X, from the point of view alias |
1404 | analysis. (For example, if X is a MEM whose address is a register, | |
1405 | and the register has a known value (say a SYMBOL_REF), then a MEM | |
1406 | whose address is the SYMBOL_REF is returned.) */ | |
1407 | ||
1408 | rtx | |
aecda0d6 | 1409 | canon_rtx (rtx x) |
ea0cb7ae | 1410 | { |
1411 | /* Recursively look for equivalences. */ | |
8ad4c111 | 1412 | if (REG_P (x) && REGNO (x) >= FIRST_PSEUDO_REGISTER) |
12cb06ad | 1413 | { |
1414 | rtx t = get_reg_known_value (REGNO (x)); | |
1415 | if (t == x) | |
1416 | return x; | |
1417 | if (t) | |
1418 | return canon_rtx (t); | |
1419 | } | |
1420 | ||
1421 | if (GET_CODE (x) == PLUS) | |
ea0cb7ae | 1422 | { |
1423 | rtx x0 = canon_rtx (XEXP (x, 0)); | |
1424 | rtx x1 = canon_rtx (XEXP (x, 1)); | |
1425 | ||
1426 | if (x0 != XEXP (x, 0) || x1 != XEXP (x, 1)) | |
1427 | { | |
971ba038 | 1428 | if (CONST_INT_P (x0)) |
b244d4c7 | 1429 | return plus_constant (x1, INTVAL (x0)); |
971ba038 | 1430 | else if (CONST_INT_P (x1)) |
b244d4c7 | 1431 | return plus_constant (x0, INTVAL (x1)); |
941522d6 | 1432 | return gen_rtx_PLUS (GET_MODE (x), x0, x1); |
ea0cb7ae | 1433 | } |
1434 | } | |
083a2b5e | 1435 | |
ea0cb7ae | 1436 | /* This gives us much better alias analysis when called from |
1437 | the loop optimizer. Note we want to leave the original | |
1438 | MEM alone, but need to return the canonicalized MEM with | |
1439 | all the flags with their original values. */ | |
e16ceb8e | 1440 | else if (MEM_P (x)) |
e4e86ec5 | 1441 | x = replace_equiv_address_nv (x, canon_rtx (XEXP (x, 0))); |
083a2b5e | 1442 | |
ea0cb7ae | 1443 | return x; |
1444 | } | |
1445 | ||
1446 | /* Return 1 if X and Y are identical-looking rtx's. | |
6285298a | 1447 | Expect that X and Y has been already canonicalized. |
ea0cb7ae | 1448 | |
1449 | We use the data in reg_known_value above to see if two registers with | |
1450 | different numbers are, in fact, equivalent. */ | |
1451 | ||
1452 | static int | |
7ecb5bb2 | 1453 | rtx_equal_for_memref_p (const_rtx x, const_rtx y) |
ea0cb7ae | 1454 | { |
19cb6b50 | 1455 | int i; |
1456 | int j; | |
1457 | enum rtx_code code; | |
1458 | const char *fmt; | |
ea0cb7ae | 1459 | |
1460 | if (x == 0 && y == 0) | |
1461 | return 1; | |
1462 | if (x == 0 || y == 0) | |
1463 | return 0; | |
083a2b5e | 1464 | |
ea0cb7ae | 1465 | if (x == y) |
1466 | return 1; | |
1467 | ||
1468 | code = GET_CODE (x); | |
1469 | /* Rtx's of different codes cannot be equal. */ | |
1470 | if (code != GET_CODE (y)) | |
1471 | return 0; | |
1472 | ||
1473 | /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. | |
1474 | (REG:SI x) and (REG:HI x) are NOT equivalent. */ | |
1475 | ||
1476 | if (GET_MODE (x) != GET_MODE (y)) | |
1477 | return 0; | |
1478 | ||
7cfb9bcf | 1479 | /* Some RTL can be compared without a recursive examination. */ |
1480 | switch (code) | |
1481 | { | |
1482 | case REG: | |
1483 | return REGNO (x) == REGNO (y); | |
1484 | ||
1485 | case LABEL_REF: | |
1486 | return XEXP (x, 0) == XEXP (y, 0); | |
9342ee68 | 1487 | |
7cfb9bcf | 1488 | case SYMBOL_REF: |
1489 | return XSTR (x, 0) == XSTR (y, 0); | |
1490 | ||
a2e238f7 | 1491 | case VALUE: |
7cfb9bcf | 1492 | case CONST_INT: |
1493 | case CONST_DOUBLE: | |
e397ad8e | 1494 | case CONST_FIXED: |
7cfb9bcf | 1495 | /* There's no need to compare the contents of CONST_DOUBLEs or |
1496 | CONST_INTs because pointer equality is a good enough | |
1497 | comparison for these nodes. */ | |
1498 | return 0; | |
1499 | ||
7cfb9bcf | 1500 | default: |
1501 | break; | |
1502 | } | |
ea0cb7ae | 1503 | |
6285298a | 1504 | /* canon_rtx knows how to handle plus. No need to canonicalize. */ |
1505 | if (code == PLUS) | |
ea0cb7ae | 1506 | return ((rtx_equal_for_memref_p (XEXP (x, 0), XEXP (y, 0)) |
1507 | && rtx_equal_for_memref_p (XEXP (x, 1), XEXP (y, 1))) | |
1508 | || (rtx_equal_for_memref_p (XEXP (x, 0), XEXP (y, 1)) | |
1509 | && rtx_equal_for_memref_p (XEXP (x, 1), XEXP (y, 0)))); | |
6285298a | 1510 | /* For commutative operations, the RTX match if the operand match in any |
1511 | order. Also handle the simple binary and unary cases without a loop. */ | |
6720e96c | 1512 | if (COMMUTATIVE_P (x)) |
6285298a | 1513 | { |
1514 | rtx xop0 = canon_rtx (XEXP (x, 0)); | |
1515 | rtx yop0 = canon_rtx (XEXP (y, 0)); | |
1516 | rtx yop1 = canon_rtx (XEXP (y, 1)); | |
1517 | ||
1518 | return ((rtx_equal_for_memref_p (xop0, yop0) | |
1519 | && rtx_equal_for_memref_p (canon_rtx (XEXP (x, 1)), yop1)) | |
1520 | || (rtx_equal_for_memref_p (xop0, yop1) | |
1521 | && rtx_equal_for_memref_p (canon_rtx (XEXP (x, 1)), yop0))); | |
1522 | } | |
6720e96c | 1523 | else if (NON_COMMUTATIVE_P (x)) |
6285298a | 1524 | { |
1525 | return (rtx_equal_for_memref_p (canon_rtx (XEXP (x, 0)), | |
aecda0d6 | 1526 | canon_rtx (XEXP (y, 0))) |
6285298a | 1527 | && rtx_equal_for_memref_p (canon_rtx (XEXP (x, 1)), |
1528 | canon_rtx (XEXP (y, 1)))); | |
1529 | } | |
6720e96c | 1530 | else if (UNARY_P (x)) |
6285298a | 1531 | return rtx_equal_for_memref_p (canon_rtx (XEXP (x, 0)), |
aecda0d6 | 1532 | canon_rtx (XEXP (y, 0))); |
ea0cb7ae | 1533 | |
1534 | /* Compare the elements. If any pair of corresponding elements | |
1a6c3add | 1535 | fail to match, return 0 for the whole things. |
1536 | ||
1537 | Limit cases to types which actually appear in addresses. */ | |
ea0cb7ae | 1538 | |
1539 | fmt = GET_RTX_FORMAT (code); | |
1540 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
1541 | { | |
1542 | switch (fmt[i]) | |
1543 | { | |
ea0cb7ae | 1544 | case 'i': |
1545 | if (XINT (x, i) != XINT (y, i)) | |
1546 | return 0; | |
1547 | break; | |
1548 | ||
ea0cb7ae | 1549 | case 'E': |
1550 | /* Two vectors must have the same length. */ | |
1551 | if (XVECLEN (x, i) != XVECLEN (y, i)) | |
1552 | return 0; | |
1553 | ||
1554 | /* And the corresponding elements must match. */ | |
1555 | for (j = 0; j < XVECLEN (x, i); j++) | |
6285298a | 1556 | if (rtx_equal_for_memref_p (canon_rtx (XVECEXP (x, i, j)), |
1557 | canon_rtx (XVECEXP (y, i, j))) == 0) | |
ea0cb7ae | 1558 | return 0; |
1559 | break; | |
1560 | ||
1561 | case 'e': | |
6285298a | 1562 | if (rtx_equal_for_memref_p (canon_rtx (XEXP (x, i)), |
1563 | canon_rtx (XEXP (y, i))) == 0) | |
ea0cb7ae | 1564 | return 0; |
1565 | break; | |
1566 | ||
83ec925c | 1567 | /* This can happen for asm operands. */ |
1568 | case 's': | |
1569 | if (strcmp (XSTR (x, i), XSTR (y, i))) | |
1570 | return 0; | |
1571 | break; | |
1572 | ||
23c04f6b | 1573 | /* This can happen for an asm which clobbers memory. */ |
1574 | case '0': | |
1575 | break; | |
1576 | ||
ea0cb7ae | 1577 | /* It is believed that rtx's at this level will never |
1578 | contain anything but integers and other rtx's, | |
1579 | except for within LABEL_REFs and SYMBOL_REFs. */ | |
1580 | default: | |
64db345d | 1581 | gcc_unreachable (); |
ea0cb7ae | 1582 | } |
1583 | } | |
1584 | return 1; | |
1585 | } | |
1586 | ||
5a37abfd | 1587 | rtx |
aecda0d6 | 1588 | find_base_term (rtx x) |
ea0cb7ae | 1589 | { |
1617c276 | 1590 | cselib_val *val; |
8081d3a6 | 1591 | struct elt_loc_list *l, *f; |
1592 | rtx ret; | |
1617c276 | 1593 | |
15750b1e | 1594 | #if defined (FIND_BASE_TERM) |
1595 | /* Try machine-dependent ways to find the base term. */ | |
1596 | x = FIND_BASE_TERM (x); | |
1597 | #endif | |
1598 | ||
ea0cb7ae | 1599 | switch (GET_CODE (x)) |
1600 | { | |
1601 | case REG: | |
1602 | return REG_BASE_VALUE (x); | |
1603 | ||
2be64ef0 | 1604 | case TRUNCATE: |
04ec15fa | 1605 | /* As we do not know which address space the pointer is referring to, we can |
98155838 | 1606 | handle this only if the target does not support different pointer or |
1607 | address modes depending on the address space. */ | |
1608 | if (!target_default_pointer_address_modes_p ()) | |
1609 | return 0; | |
2be64ef0 | 1610 | if (GET_MODE_SIZE (GET_MODE (x)) < GET_MODE_SIZE (Pmode)) |
9342ee68 | 1611 | return 0; |
2be64ef0 | 1612 | /* Fall through. */ |
ea0cb7ae | 1613 | case HIGH: |
22dfec7d | 1614 | case PRE_INC: |
1615 | case PRE_DEC: | |
1616 | case POST_INC: | |
1617 | case POST_DEC: | |
2be64ef0 | 1618 | case PRE_MODIFY: |
1619 | case POST_MODIFY: | |
22dfec7d | 1620 | return find_base_term (XEXP (x, 0)); |
1621 | ||
5104ef69 | 1622 | case ZERO_EXTEND: |
1623 | case SIGN_EXTEND: /* Used for Alpha/NT pointers */ | |
04ec15fa | 1624 | /* As we do not know which address space the pointer is referring to, we can |
98155838 | 1625 | handle this only if the target does not support different pointer or |
1626 | address modes depending on the address space. */ | |
1627 | if (!target_default_pointer_address_modes_p ()) | |
1628 | return 0; | |
1629 | ||
5104ef69 | 1630 | { |
1631 | rtx temp = find_base_term (XEXP (x, 0)); | |
1632 | ||
85d654dd | 1633 | if (temp != 0 && CONSTANT_P (temp)) |
5104ef69 | 1634 | temp = convert_memory_address (Pmode, temp); |
5104ef69 | 1635 | |
1636 | return temp; | |
1637 | } | |
1638 | ||
1617c276 | 1639 | case VALUE: |
1640 | val = CSELIB_VAL_PTR (x); | |
8081d3a6 | 1641 | ret = NULL_RTX; |
1642 | ||
a2e238f7 | 1643 | if (!val) |
8081d3a6 | 1644 | return ret; |
1645 | ||
1646 | f = val->locs; | |
1647 | /* Temporarily reset val->locs to avoid infinite recursion. */ | |
1648 | val->locs = NULL; | |
1649 | ||
1650 | for (l = f; l; l = l->next) | |
1651 | if (GET_CODE (l->loc) == VALUE | |
1652 | && CSELIB_VAL_PTR (l->loc)->locs | |
1653 | && !CSELIB_VAL_PTR (l->loc)->locs->next | |
1654 | && CSELIB_VAL_PTR (l->loc)->locs->loc == x) | |
1655 | continue; | |
1656 | else if ((ret = find_base_term (l->loc)) != 0) | |
1657 | break; | |
1658 | ||
1659 | val->locs = f; | |
1660 | return ret; | |
1617c276 | 1661 | |
a7cf5bbf | 1662 | case LO_SUM: |
1663 | /* The standard form is (lo_sum reg sym) so look only at the | |
1664 | second operand. */ | |
1665 | return find_base_term (XEXP (x, 1)); | |
1666 | ||
ea0cb7ae | 1667 | case CONST: |
1668 | x = XEXP (x, 0); | |
1669 | if (GET_CODE (x) != PLUS && GET_CODE (x) != MINUS) | |
1670 | return 0; | |
b4b174c3 | 1671 | /* Fall through. */ |
ea0cb7ae | 1672 | case PLUS: |
1673 | case MINUS: | |
1674 | { | |
1b5bbc4c | 1675 | rtx tmp1 = XEXP (x, 0); |
1676 | rtx tmp2 = XEXP (x, 1); | |
1677 | ||
35a3065a | 1678 | /* This is a little bit tricky since we have to determine which of |
1b5bbc4c | 1679 | the two operands represents the real base address. Otherwise this |
1680 | routine may return the index register instead of the base register. | |
1681 | ||
1682 | That may cause us to believe no aliasing was possible, when in | |
1683 | fact aliasing is possible. | |
1684 | ||
1685 | We use a few simple tests to guess the base register. Additional | |
1686 | tests can certainly be added. For example, if one of the operands | |
1687 | is a shift or multiply, then it must be the index register and the | |
1688 | other operand is the base register. */ | |
9342ee68 | 1689 | |
15750b1e | 1690 | if (tmp1 == pic_offset_table_rtx && CONSTANT_P (tmp2)) |
1691 | return find_base_term (tmp2); | |
1692 | ||
1b5bbc4c | 1693 | /* If either operand is known to be a pointer, then use it |
1694 | to determine the base term. */ | |
e61a0a7f | 1695 | if (REG_P (tmp1) && REG_POINTER (tmp1)) |
02b0d167 | 1696 | { |
1697 | rtx base = find_base_term (tmp1); | |
1698 | if (base) | |
1699 | return base; | |
1700 | } | |
1b5bbc4c | 1701 | |
e61a0a7f | 1702 | if (REG_P (tmp2) && REG_POINTER (tmp2)) |
02b0d167 | 1703 | { |
1704 | rtx base = find_base_term (tmp2); | |
1705 | if (base) | |
1706 | return base; | |
1707 | } | |
1b5bbc4c | 1708 | |
1709 | /* Neither operand was known to be a pointer. Go ahead and find the | |
1710 | base term for both operands. */ | |
1711 | tmp1 = find_base_term (tmp1); | |
1712 | tmp2 = find_base_term (tmp2); | |
1713 | ||
1714 | /* If either base term is named object or a special address | |
1715 | (like an argument or stack reference), then use it for the | |
1716 | base term. */ | |
4e56ceb1 | 1717 | if (tmp1 != 0 && known_base_value_p (tmp1)) |
1b5bbc4c | 1718 | return tmp1; |
1719 | ||
4e56ceb1 | 1720 | if (tmp2 != 0 && known_base_value_p (tmp2)) |
1b5bbc4c | 1721 | return tmp2; |
1722 | ||
1723 | /* We could not determine which of the two operands was the | |
1724 | base register and which was the index. So we can determine | |
1725 | nothing from the base alias check. */ | |
1726 | return 0; | |
ea0cb7ae | 1727 | } |
1728 | ||
1729 | case AND: | |
971ba038 | 1730 | if (CONST_INT_P (XEXP (x, 1)) && INTVAL (XEXP (x, 1)) != 0) |
2be64ef0 | 1731 | return find_base_term (XEXP (x, 0)); |
ea0cb7ae | 1732 | return 0; |
1733 | ||
1734 | case SYMBOL_REF: | |
1735 | case LABEL_REF: | |
1736 | return x; | |
1737 | ||
1738 | default: | |
1739 | return 0; | |
1740 | } | |
1741 | } | |
1742 | ||
1743 | /* Return 0 if the addresses X and Y are known to point to different | |
1744 | objects, 1 if they might be pointers to the same object. */ | |
1745 | ||
1746 | static int | |
aecda0d6 | 1747 | base_alias_check (rtx x, rtx y, enum machine_mode x_mode, |
1748 | enum machine_mode y_mode) | |
ea0cb7ae | 1749 | { |
1750 | rtx x_base = find_base_term (x); | |
1751 | rtx y_base = find_base_term (y); | |
1752 | ||
3ad3a0ed | 1753 | /* If the address itself has no known base see if a known equivalent |
1754 | value has one. If either address still has no known base, nothing | |
1755 | is known about aliasing. */ | |
1756 | if (x_base == 0) | |
1757 | { | |
1758 | rtx x_c; | |
083a2b5e | 1759 | |
3ad3a0ed | 1760 | if (! flag_expensive_optimizations || (x_c = canon_rtx (x)) == x) |
1761 | return 1; | |
083a2b5e | 1762 | |
3ad3a0ed | 1763 | x_base = find_base_term (x_c); |
1764 | if (x_base == 0) | |
1765 | return 1; | |
1766 | } | |
ea0cb7ae | 1767 | |
3ad3a0ed | 1768 | if (y_base == 0) |
1769 | { | |
1770 | rtx y_c; | |
1771 | if (! flag_expensive_optimizations || (y_c = canon_rtx (y)) == y) | |
1772 | return 1; | |
083a2b5e | 1773 | |
3ad3a0ed | 1774 | y_base = find_base_term (y_c); |
1775 | if (y_base == 0) | |
1776 | return 1; | |
1777 | } | |
1778 | ||
1779 | /* If the base addresses are equal nothing is known about aliasing. */ | |
1780 | if (rtx_equal_p (x_base, y_base)) | |
ea0cb7ae | 1781 | return 1; |
1782 | ||
f92492f7 | 1783 | /* The base addresses are different expressions. If they are not accessed |
1784 | via AND, there is no conflict. We can bring knowledge of object | |
1785 | alignment into play here. For example, on alpha, "char a, b;" can | |
1786 | alias one another, though "char a; long b;" cannot. AND addesses may | |
1787 | implicitly alias surrounding objects; i.e. unaligned access in DImode | |
1788 | via AND address can alias all surrounding object types except those | |
1789 | with aligment 8 or higher. */ | |
1790 | if (GET_CODE (x) == AND && GET_CODE (y) == AND) | |
1791 | return 1; | |
1792 | if (GET_CODE (x) == AND | |
971ba038 | 1793 | && (!CONST_INT_P (XEXP (x, 1)) |
f92492f7 | 1794 | || (int) GET_MODE_UNIT_SIZE (y_mode) < -INTVAL (XEXP (x, 1)))) |
1795 | return 1; | |
1796 | if (GET_CODE (y) == AND | |
971ba038 | 1797 | && (!CONST_INT_P (XEXP (y, 1)) |
f92492f7 | 1798 | || (int) GET_MODE_UNIT_SIZE (x_mode) < -INTVAL (XEXP (y, 1)))) |
1799 | return 1; | |
1800 | ||
1801 | /* Differing symbols not accessed via AND never alias. */ | |
ea0cb7ae | 1802 | if (GET_CODE (x_base) != ADDRESS && GET_CODE (y_base) != ADDRESS) |
f92492f7 | 1803 | return 0; |
ea0cb7ae | 1804 | |
4e56ceb1 | 1805 | if (unique_base_value_p (x_base) || unique_base_value_p (y_base)) |
ea0cb7ae | 1806 | return 0; |
1807 | ||
e190b18a | 1808 | return 1; |
ea0cb7ae | 1809 | } |
1810 | ||
fa5e407e | 1811 | /* Callback for for_each_rtx, that returns 1 upon encountering a VALUE |
1812 | whose UID is greater than the int uid that D points to. */ | |
1813 | ||
1814 | static int | |
1815 | refs_newer_value_cb (rtx *x, void *d) | |
1816 | { | |
1817 | if (GET_CODE (*x) == VALUE && CSELIB_VAL_PTR (*x)->uid > *(int *)d) | |
1818 | return 1; | |
1819 | ||
1820 | return 0; | |
1821 | } | |
1822 | ||
1823 | /* Return TRUE if EXPR refers to a VALUE whose uid is greater than | |
1824 | that of V. */ | |
1825 | ||
1826 | static bool | |
1827 | refs_newer_value_p (rtx expr, rtx v) | |
1828 | { | |
1829 | int minuid = CSELIB_VAL_PTR (v)->uid; | |
1830 | ||
1831 | return for_each_rtx (&expr, refs_newer_value_cb, &minuid); | |
1832 | } | |
1833 | ||
1617c276 | 1834 | /* Convert the address X into something we can use. This is done by returning |
1835 | it unchanged unless it is a value; in the latter case we call cselib to get | |
1836 | a more useful rtx. */ | |
f7c44134 | 1837 | |
8e802be9 | 1838 | rtx |
aecda0d6 | 1839 | get_addr (rtx x) |
1617c276 | 1840 | { |
1841 | cselib_val *v; | |
1842 | struct elt_loc_list *l; | |
1843 | ||
1844 | if (GET_CODE (x) != VALUE) | |
1845 | return x; | |
1846 | v = CSELIB_VAL_PTR (x); | |
a2e238f7 | 1847 | if (v) |
1848 | { | |
85b6e75b | 1849 | bool have_equivs = cselib_have_permanent_equivalences (); |
1850 | if (have_equivs) | |
1851 | v = canonical_cselib_val (v); | |
a2e238f7 | 1852 | for (l = v->locs; l; l = l->next) |
1853 | if (CONSTANT_P (l->loc)) | |
1854 | return l->loc; | |
1855 | for (l = v->locs; l; l = l->next) | |
85b6e75b | 1856 | if (!REG_P (l->loc) && !MEM_P (l->loc) |
1857 | /* Avoid infinite recursion when potentially dealing with | |
1858 | var-tracking artificial equivalences, by skipping the | |
1859 | equivalences themselves, and not choosing expressions | |
1860 | that refer to newer VALUEs. */ | |
1861 | && (!have_equivs | |
1862 | || (GET_CODE (l->loc) != VALUE | |
1863 | && !refs_newer_value_p (l->loc, x)))) | |
fa5e407e | 1864 | return l->loc; |
85b6e75b | 1865 | if (have_equivs) |
1866 | { | |
1867 | for (l = v->locs; l; l = l->next) | |
1868 | if (REG_P (l->loc) | |
1869 | || (GET_CODE (l->loc) != VALUE | |
1870 | && !refs_newer_value_p (l->loc, x))) | |
1871 | return l->loc; | |
1872 | /* Return the canonical value. */ | |
1873 | return v->val_rtx; | |
1874 | } | |
1875 | if (v->locs) | |
1876 | return v->locs->loc; | |
a2e238f7 | 1877 | } |
1617c276 | 1878 | return x; |
1879 | } | |
1880 | ||
b9f92976 | 1881 | /* Return the address of the (N_REFS + 1)th memory reference to ADDR |
1882 | where SIZE is the size in bytes of the memory reference. If ADDR | |
1883 | is not modified by the memory reference then ADDR is returned. */ | |
1884 | ||
0aa9ffd4 | 1885 | static rtx |
aecda0d6 | 1886 | addr_side_effect_eval (rtx addr, int size, int n_refs) |
b9f92976 | 1887 | { |
1888 | int offset = 0; | |
9342ee68 | 1889 | |
b9f92976 | 1890 | switch (GET_CODE (addr)) |
1891 | { | |
1892 | case PRE_INC: | |
1893 | offset = (n_refs + 1) * size; | |
1894 | break; | |
1895 | case PRE_DEC: | |
1896 | offset = -(n_refs + 1) * size; | |
1897 | break; | |
1898 | case POST_INC: | |
1899 | offset = n_refs * size; | |
1900 | break; | |
1901 | case POST_DEC: | |
1902 | offset = -n_refs * size; | |
1903 | break; | |
1904 | ||
1905 | default: | |
1906 | return addr; | |
1907 | } | |
9342ee68 | 1908 | |
b9f92976 | 1909 | if (offset) |
6285298a | 1910 | addr = gen_rtx_PLUS (GET_MODE (addr), XEXP (addr, 0), |
a0c938f0 | 1911 | GEN_INT (offset)); |
b9f92976 | 1912 | else |
1913 | addr = XEXP (addr, 0); | |
6285298a | 1914 | addr = canon_rtx (addr); |
b9f92976 | 1915 | |
1916 | return addr; | |
1917 | } | |
1918 | ||
40f9e245 | 1919 | /* Return one if X and Y (memory addresses) reference the |
1920 | same location in memory or if the references overlap. | |
1921 | Return zero if they do not overlap, else return | |
1922 | minus one in which case they still might reference the same location. | |
1923 | ||
1924 | C is an offset accumulator. When | |
ea0cb7ae | 1925 | C is nonzero, we are testing aliases between X and Y + C. |
1926 | XSIZE is the size in bytes of the X reference, | |
1927 | similarly YSIZE is the size in bytes for Y. | |
6285298a | 1928 | Expect that canon_rtx has been already called for X and Y. |
ea0cb7ae | 1929 | |
1930 | If XSIZE or YSIZE is zero, we do not know the amount of memory being | |
1931 | referenced (the reference was BLKmode), so make the most pessimistic | |
1932 | assumptions. | |
1933 | ||
1e38bf49 | 1934 | If XSIZE or YSIZE is negative, we may access memory outside the object |
1935 | being referenced as a side effect. This can happen when using AND to | |
1936 | align memory references, as is done on the Alpha. | |
1937 | ||
ea0cb7ae | 1938 | Nice to notice that varying addresses cannot conflict with fp if no |
40f9e245 | 1939 | local variables had their addresses taken, but that's too hard now. |
1940 | ||
1941 | ??? Contrary to the tree alias oracle this does not return | |
1942 | one for X + non-constant and Y + non-constant when X and Y are equal. | |
1943 | If that is fixed the TBAA hack for union type-punning can be removed. */ | |
ea0cb7ae | 1944 | |
ea0cb7ae | 1945 | static int |
aecda0d6 | 1946 | memrefs_conflict_p (int xsize, rtx x, int ysize, rtx y, HOST_WIDE_INT c) |
ea0cb7ae | 1947 | { |
1617c276 | 1948 | if (GET_CODE (x) == VALUE) |
6f9688ce | 1949 | { |
1950 | if (REG_P (y)) | |
1951 | { | |
4e15f806 | 1952 | struct elt_loc_list *l = NULL; |
1953 | if (CSELIB_VAL_PTR (x)) | |
fa5e407e | 1954 | for (l = canonical_cselib_val (CSELIB_VAL_PTR (x))->locs; |
1955 | l; l = l->next) | |
4e15f806 | 1956 | if (REG_P (l->loc) && rtx_equal_for_memref_p (l->loc, y)) |
1957 | break; | |
6f9688ce | 1958 | if (l) |
1959 | x = y; | |
1960 | else | |
1961 | x = get_addr (x); | |
1962 | } | |
1963 | /* Don't call get_addr if y is the same VALUE. */ | |
1964 | else if (x != y) | |
1965 | x = get_addr (x); | |
1966 | } | |
1617c276 | 1967 | if (GET_CODE (y) == VALUE) |
6f9688ce | 1968 | { |
1969 | if (REG_P (x)) | |
1970 | { | |
4e15f806 | 1971 | struct elt_loc_list *l = NULL; |
1972 | if (CSELIB_VAL_PTR (y)) | |
fa5e407e | 1973 | for (l = canonical_cselib_val (CSELIB_VAL_PTR (y))->locs; |
1974 | l; l = l->next) | |
4e15f806 | 1975 | if (REG_P (l->loc) && rtx_equal_for_memref_p (l->loc, x)) |
1976 | break; | |
6f9688ce | 1977 | if (l) |
1978 | y = x; | |
1979 | else | |
1980 | y = get_addr (y); | |
1981 | } | |
1982 | /* Don't call get_addr if x is the same VALUE. */ | |
1983 | else if (y != x) | |
1984 | y = get_addr (y); | |
1985 | } | |
ea0cb7ae | 1986 | if (GET_CODE (x) == HIGH) |
1987 | x = XEXP (x, 0); | |
1988 | else if (GET_CODE (x) == LO_SUM) | |
1989 | x = XEXP (x, 1); | |
1990 | else | |
6285298a | 1991 | x = addr_side_effect_eval (x, xsize, 0); |
ea0cb7ae | 1992 | if (GET_CODE (y) == HIGH) |
1993 | y = XEXP (y, 0); | |
1994 | else if (GET_CODE (y) == LO_SUM) | |
1995 | y = XEXP (y, 1); | |
1996 | else | |
6285298a | 1997 | y = addr_side_effect_eval (y, ysize, 0); |
ea0cb7ae | 1998 | |
1999 | if (rtx_equal_for_memref_p (x, y)) | |
2000 | { | |
1e38bf49 | 2001 | if (xsize <= 0 || ysize <= 0) |
ea0cb7ae | 2002 | return 1; |
2003 | if (c >= 0 && xsize > c) | |
2004 | return 1; | |
2005 | if (c < 0 && ysize+c > 0) | |
2006 | return 1; | |
2007 | return 0; | |
2008 | } | |
2009 | ||
35d3eedd | 2010 | /* This code used to check for conflicts involving stack references and |
2011 | globals but the base address alias code now handles these cases. */ | |
ea0cb7ae | 2012 | |
2013 | if (GET_CODE (x) == PLUS) | |
2014 | { | |
2015 | /* The fact that X is canonicalized means that this | |
2016 | PLUS rtx is canonicalized. */ | |
2017 | rtx x0 = XEXP (x, 0); | |
2018 | rtx x1 = XEXP (x, 1); | |
2019 | ||
2020 | if (GET_CODE (y) == PLUS) | |
2021 | { | |
2022 | /* The fact that Y is canonicalized means that this | |
2023 | PLUS rtx is canonicalized. */ | |
2024 | rtx y0 = XEXP (y, 0); | |
2025 | rtx y1 = XEXP (y, 1); | |
2026 | ||
2027 | if (rtx_equal_for_memref_p (x1, y1)) | |
2028 | return memrefs_conflict_p (xsize, x0, ysize, y0, c); | |
2029 | if (rtx_equal_for_memref_p (x0, y0)) | |
2030 | return memrefs_conflict_p (xsize, x1, ysize, y1, c); | |
971ba038 | 2031 | if (CONST_INT_P (x1)) |
5f59c49b | 2032 | { |
971ba038 | 2033 | if (CONST_INT_P (y1)) |
5f59c49b | 2034 | return memrefs_conflict_p (xsize, x0, ysize, y0, |
2035 | c - INTVAL (x1) + INTVAL (y1)); | |
2036 | else | |
2037 | return memrefs_conflict_p (xsize, x0, ysize, y, | |
2038 | c - INTVAL (x1)); | |
2039 | } | |
971ba038 | 2040 | else if (CONST_INT_P (y1)) |
ea0cb7ae | 2041 | return memrefs_conflict_p (xsize, x, ysize, y0, c + INTVAL (y1)); |
2042 | ||
40f9e245 | 2043 | return -1; |
ea0cb7ae | 2044 | } |
971ba038 | 2045 | else if (CONST_INT_P (x1)) |
ea0cb7ae | 2046 | return memrefs_conflict_p (xsize, x0, ysize, y, c - INTVAL (x1)); |
2047 | } | |
2048 | else if (GET_CODE (y) == PLUS) | |
2049 | { | |
2050 | /* The fact that Y is canonicalized means that this | |
2051 | PLUS rtx is canonicalized. */ | |
2052 | rtx y0 = XEXP (y, 0); | |
2053 | rtx y1 = XEXP (y, 1); | |
2054 | ||
971ba038 | 2055 | if (CONST_INT_P (y1)) |
ea0cb7ae | 2056 | return memrefs_conflict_p (xsize, x, ysize, y0, c + INTVAL (y1)); |
2057 | else | |
40f9e245 | 2058 | return -1; |
ea0cb7ae | 2059 | } |
2060 | ||
2061 | if (GET_CODE (x) == GET_CODE (y)) | |
2062 | switch (GET_CODE (x)) | |
2063 | { | |
2064 | case MULT: | |
2065 | { | |
2066 | /* Handle cases where we expect the second operands to be the | |
2067 | same, and check only whether the first operand would conflict | |
2068 | or not. */ | |
2069 | rtx x0, y0; | |
2070 | rtx x1 = canon_rtx (XEXP (x, 1)); | |
2071 | rtx y1 = canon_rtx (XEXP (y, 1)); | |
2072 | if (! rtx_equal_for_memref_p (x1, y1)) | |
40f9e245 | 2073 | return -1; |
ea0cb7ae | 2074 | x0 = canon_rtx (XEXP (x, 0)); |
2075 | y0 = canon_rtx (XEXP (y, 0)); | |
2076 | if (rtx_equal_for_memref_p (x0, y0)) | |
2077 | return (xsize == 0 || ysize == 0 | |
2078 | || (c >= 0 && xsize > c) || (c < 0 && ysize+c > 0)); | |
2079 | ||
2080 | /* Can't properly adjust our sizes. */ | |
971ba038 | 2081 | if (!CONST_INT_P (x1)) |
40f9e245 | 2082 | return -1; |
ea0cb7ae | 2083 | xsize /= INTVAL (x1); |
2084 | ysize /= INTVAL (x1); | |
2085 | c /= INTVAL (x1); | |
2086 | return memrefs_conflict_p (xsize, x0, ysize, y0, c); | |
2087 | } | |
99c14947 | 2088 | |
2089 | default: | |
2090 | break; | |
ea0cb7ae | 2091 | } |
2092 | ||
2093 | /* Treat an access through an AND (e.g. a subword access on an Alpha) | |
9342ee68 | 2094 | as an access with indeterminate size. Assume that references |
9fb1c31b | 2095 | besides AND are aligned, so if the size of the other reference is |
2096 | at least as large as the alignment, assume no other overlap. */ | |
971ba038 | 2097 | if (GET_CODE (x) == AND && CONST_INT_P (XEXP (x, 1))) |
9fb1c31b | 2098 | { |
bcdb334e | 2099 | if (GET_CODE (y) == AND || ysize < -INTVAL (XEXP (x, 1))) |
9fb1c31b | 2100 | xsize = -1; |
6285298a | 2101 | return memrefs_conflict_p (xsize, canon_rtx (XEXP (x, 0)), ysize, y, c); |
9fb1c31b | 2102 | } |
971ba038 | 2103 | if (GET_CODE (y) == AND && CONST_INT_P (XEXP (y, 1))) |
1e38bf49 | 2104 | { |
9fb1c31b | 2105 | /* ??? If we are indexing far enough into the array/structure, we |
9342ee68 | 2106 | may yet be able to determine that we can not overlap. But we |
1e38bf49 | 2107 | also need to that we are far enough from the end not to overlap |
9fb1c31b | 2108 | a following reference, so we do nothing with that for now. */ |
bcdb334e | 2109 | if (GET_CODE (x) == AND || xsize < -INTVAL (XEXP (y, 1))) |
9fb1c31b | 2110 | ysize = -1; |
6285298a | 2111 | return memrefs_conflict_p (xsize, x, ysize, canon_rtx (XEXP (y, 0)), c); |
1e38bf49 | 2112 | } |
ea0cb7ae | 2113 | |
2114 | if (CONSTANT_P (x)) | |
2115 | { | |
971ba038 | 2116 | if (CONST_INT_P (x) && CONST_INT_P (y)) |
ea0cb7ae | 2117 | { |
2118 | c += (INTVAL (y) - INTVAL (x)); | |
1e38bf49 | 2119 | return (xsize <= 0 || ysize <= 0 |
ea0cb7ae | 2120 | || (c >= 0 && xsize > c) || (c < 0 && ysize+c > 0)); |
2121 | } | |
2122 | ||
2123 | if (GET_CODE (x) == CONST) | |
2124 | { | |
2125 | if (GET_CODE (y) == CONST) | |
2126 | return memrefs_conflict_p (xsize, canon_rtx (XEXP (x, 0)), | |
2127 | ysize, canon_rtx (XEXP (y, 0)), c); | |
2128 | else | |
2129 | return memrefs_conflict_p (xsize, canon_rtx (XEXP (x, 0)), | |
2130 | ysize, y, c); | |
2131 | } | |
2132 | if (GET_CODE (y) == CONST) | |
2133 | return memrefs_conflict_p (xsize, x, ysize, | |
2134 | canon_rtx (XEXP (y, 0)), c); | |
2135 | ||
2136 | if (CONSTANT_P (y)) | |
15750b1e | 2137 | return (xsize <= 0 || ysize <= 0 |
1e38bf49 | 2138 | || (rtx_equal_for_memref_p (x, y) |
15750b1e | 2139 | && ((c >= 0 && xsize > c) || (c < 0 && ysize+c > 0)))); |
ea0cb7ae | 2140 | |
40f9e245 | 2141 | return -1; |
ea0cb7ae | 2142 | } |
40f9e245 | 2143 | |
2144 | return -1; | |
ea0cb7ae | 2145 | } |
2146 | ||
2147 | /* Functions to compute memory dependencies. | |
2148 | ||
2149 | Since we process the insns in execution order, we can build tables | |
2150 | to keep track of what registers are fixed (and not aliased), what registers | |
2151 | are varying in known ways, and what registers are varying in unknown | |
2152 | ways. | |
2153 | ||
2154 | If both memory references are volatile, then there must always be a | |
2155 | dependence between the two references, since their order can not be | |
2156 | changed. A volatile and non-volatile reference can be interchanged | |
9342ee68 | 2157 | though. |
ea0cb7ae | 2158 | |
376a287d | 2159 | We also must allow AND addresses, because they may generate accesses |
2160 | outside the object being referenced. This is used to generate aligned | |
2161 | addresses from unaligned addresses, for instance, the alpha | |
71d9e995 | 2162 | storeqi_unaligned pattern. */ |
ea0cb7ae | 2163 | |
2164 | /* Read dependence: X is read after read in MEM takes place. There can | |
2165 | only be a dependence here if both reads are volatile. */ | |
2166 | ||
2167 | int | |
52d07779 | 2168 | read_dependence (const_rtx mem, const_rtx x) |
ea0cb7ae | 2169 | { |
2170 | return MEM_VOLATILE_P (x) && MEM_VOLATILE_P (mem); | |
2171 | } | |
2172 | ||
6a0934dd | 2173 | /* Returns nonzero if something about the mode or address format MEM1 |
2174 | indicates that it might well alias *anything*. */ | |
2175 | ||
fe69e4fe | 2176 | static int |
52d07779 | 2177 | aliases_everything_p (const_rtx mem) |
6a0934dd | 2178 | { |
6a0934dd | 2179 | if (GET_CODE (XEXP (mem, 0)) == AND) |
0975351b | 2180 | /* If the address is an AND, it's very hard to know at what it is |
6a0934dd | 2181 | actually pointing. */ |
2182 | return 1; | |
9342ee68 | 2183 | |
6a0934dd | 2184 | return 0; |
2185 | } | |
2186 | ||
b10dbbca | 2187 | /* Return true if we can determine that the fields referenced cannot |
2188 | overlap for any pair of objects. */ | |
2189 | ||
2190 | static bool | |
52d07779 | 2191 | nonoverlapping_component_refs_p (const_tree x, const_tree y) |
b10dbbca | 2192 | { |
52d07779 | 2193 | const_tree fieldx, fieldy, typex, typey, orig_y; |
b10dbbca | 2194 | |
b95fdaae | 2195 | if (!flag_strict_aliasing) |
2196 | return false; | |
2197 | ||
b10dbbca | 2198 | do |
2199 | { | |
2200 | /* The comparison has to be done at a common type, since we don't | |
cb0ccc1e | 2201 | know how the inheritance hierarchy works. */ |
b10dbbca | 2202 | orig_y = y; |
2203 | do | |
2204 | { | |
2205 | fieldx = TREE_OPERAND (x, 1); | |
a769a0a7 | 2206 | typex = TYPE_MAIN_VARIANT (DECL_FIELD_CONTEXT (fieldx)); |
b10dbbca | 2207 | |
2208 | y = orig_y; | |
2209 | do | |
2210 | { | |
2211 | fieldy = TREE_OPERAND (y, 1); | |
a769a0a7 | 2212 | typey = TYPE_MAIN_VARIANT (DECL_FIELD_CONTEXT (fieldy)); |
b10dbbca | 2213 | |
2214 | if (typex == typey) | |
2215 | goto found; | |
2216 | ||
2217 | y = TREE_OPERAND (y, 0); | |
2218 | } | |
2219 | while (y && TREE_CODE (y) == COMPONENT_REF); | |
2220 | ||
2221 | x = TREE_OPERAND (x, 0); | |
2222 | } | |
2223 | while (x && TREE_CODE (x) == COMPONENT_REF); | |
b10dbbca | 2224 | /* Never found a common type. */ |
a769a0a7 | 2225 | return false; |
b10dbbca | 2226 | |
2227 | found: | |
2228 | /* If we're left with accessing different fields of a structure, | |
2229 | then no overlap. */ | |
2230 | if (TREE_CODE (typex) == RECORD_TYPE | |
2231 | && fieldx != fieldy) | |
2232 | return true; | |
2233 | ||
2234 | /* The comparison on the current field failed. If we're accessing | |
2235 | a very nested structure, look at the next outer level. */ | |
2236 | x = TREE_OPERAND (x, 0); | |
2237 | y = TREE_OPERAND (y, 0); | |
2238 | } | |
2239 | while (x && y | |
2240 | && TREE_CODE (x) == COMPONENT_REF | |
2241 | && TREE_CODE (y) == COMPONENT_REF); | |
9342ee68 | 2242 | |
b10dbbca | 2243 | return false; |
2244 | } | |
2245 | ||
2246 | /* Look at the bottom of the COMPONENT_REF list for a DECL, and return it. */ | |
2247 | ||
2248 | static tree | |
aecda0d6 | 2249 | decl_for_component_ref (tree x) |
b10dbbca | 2250 | { |
2251 | do | |
2252 | { | |
2253 | x = TREE_OPERAND (x, 0); | |
2254 | } | |
2255 | while (x && TREE_CODE (x) == COMPONENT_REF); | |
2256 | ||
2257 | return x && DECL_P (x) ? x : NULL_TREE; | |
2258 | } | |
2259 | ||
da443c27 | 2260 | /* Walk up the COMPONENT_REF list in X and adjust *OFFSET to compensate |
2261 | for the offset of the field reference. *KNOWN_P says whether the | |
2262 | offset is known. */ | |
b10dbbca | 2263 | |
da443c27 | 2264 | static void |
2265 | adjust_offset_for_component_ref (tree x, bool *known_p, | |
2266 | HOST_WIDE_INT *offset) | |
b10dbbca | 2267 | { |
da443c27 | 2268 | if (!*known_p) |
2269 | return; | |
9342ee68 | 2270 | do |
b10dbbca | 2271 | { |
da443c27 | 2272 | tree xoffset = component_ref_field_offset (x); |
b10dbbca | 2273 | tree field = TREE_OPERAND (x, 1); |
2274 | ||
da443c27 | 2275 | if (! host_integerp (xoffset, 1)) |
2276 | { | |
2277 | *known_p = false; | |
2278 | return; | |
2279 | } | |
2280 | *offset += (tree_low_cst (xoffset, 1) | |
b10dbbca | 2281 | + (tree_low_cst (DECL_FIELD_BIT_OFFSET (field), 1) |
2282 | / BITS_PER_UNIT)); | |
2283 | ||
2284 | x = TREE_OPERAND (x, 0); | |
2285 | } | |
2286 | while (x && TREE_CODE (x) == COMPONENT_REF); | |
b10dbbca | 2287 | } |
2288 | ||
d716ce75 | 2289 | /* Return nonzero if we can determine the exprs corresponding to memrefs |
a84256aa | 2290 | X and Y and they do not overlap. |
2291 | If LOOP_VARIANT is set, skip offset-based disambiguation */ | |
ddb3d16f | 2292 | |
5fe61d21 | 2293 | int |
a84256aa | 2294 | nonoverlapping_memrefs_p (const_rtx x, const_rtx y, bool loop_invariant) |
ddb3d16f | 2295 | { |
b10dbbca | 2296 | tree exprx = MEM_EXPR (x), expry = MEM_EXPR (y); |
ddb3d16f | 2297 | rtx rtlx, rtly; |
2298 | rtx basex, basey; | |
da443c27 | 2299 | bool moffsetx_known_p, moffsety_known_p; |
2300 | HOST_WIDE_INT moffsetx = 0, moffsety = 0; | |
ddb3d16f | 2301 | HOST_WIDE_INT offsetx = 0, offsety = 0, sizex, sizey, tem; |
2302 | ||
b10dbbca | 2303 | /* Unless both have exprs, we can't tell anything. */ |
2304 | if (exprx == 0 || expry == 0) | |
2305 | return 0; | |
c6b80e5d | 2306 | |
2307 | /* For spill-slot accesses make sure we have valid offsets. */ | |
2308 | if ((exprx == get_spill_slot_decl (false) | |
da443c27 | 2309 | && ! MEM_OFFSET_KNOWN_P (x)) |
c6b80e5d | 2310 | || (expry == get_spill_slot_decl (false) |
da443c27 | 2311 | && ! MEM_OFFSET_KNOWN_P (y))) |
c6b80e5d | 2312 | return 0; |
a0c938f0 | 2313 | |
b10dbbca | 2314 | /* If both are field references, we may be able to determine something. */ |
2315 | if (TREE_CODE (exprx) == COMPONENT_REF | |
2316 | && TREE_CODE (expry) == COMPONENT_REF | |
2317 | && nonoverlapping_component_refs_p (exprx, expry)) | |
2318 | return 1; | |
2319 | ||
a0c938f0 | 2320 | |
b10dbbca | 2321 | /* If the field reference test failed, look at the DECLs involved. */ |
da443c27 | 2322 | moffsetx_known_p = MEM_OFFSET_KNOWN_P (x); |
2323 | if (moffsetx_known_p) | |
2324 | moffsetx = MEM_OFFSET (x); | |
b10dbbca | 2325 | if (TREE_CODE (exprx) == COMPONENT_REF) |
2326 | { | |
b50c6e49 | 2327 | tree t = decl_for_component_ref (exprx); |
2328 | if (! t) | |
2329 | return 0; | |
da443c27 | 2330 | adjust_offset_for_component_ref (exprx, &moffsetx_known_p, &moffsetx); |
b50c6e49 | 2331 | exprx = t; |
b10dbbca | 2332 | } |
2d8fe5d0 | 2333 | |
da443c27 | 2334 | moffsety_known_p = MEM_OFFSET_KNOWN_P (y); |
2335 | if (moffsety_known_p) | |
2336 | moffsety = MEM_OFFSET (y); | |
b10dbbca | 2337 | if (TREE_CODE (expry) == COMPONENT_REF) |
2338 | { | |
b50c6e49 | 2339 | tree t = decl_for_component_ref (expry); |
2340 | if (! t) | |
2341 | return 0; | |
da443c27 | 2342 | adjust_offset_for_component_ref (expry, &moffsety_known_p, &moffsety); |
b50c6e49 | 2343 | expry = t; |
b10dbbca | 2344 | } |
2345 | ||
2346 | if (! DECL_P (exprx) || ! DECL_P (expry)) | |
ddb3d16f | 2347 | return 0; |
2348 | ||
119147d1 | 2349 | /* With invalid code we can end up storing into the constant pool. |
2350 | Bail out to avoid ICEing when creating RTL for this. | |
2351 | See gfortran.dg/lto/20091028-2_0.f90. */ | |
2352 | if (TREE_CODE (exprx) == CONST_DECL | |
2353 | || TREE_CODE (expry) == CONST_DECL) | |
2354 | return 1; | |
2355 | ||
b10dbbca | 2356 | rtlx = DECL_RTL (exprx); |
2357 | rtly = DECL_RTL (expry); | |
ddb3d16f | 2358 | |
10e949c1 | 2359 | /* If either RTL is not a MEM, it must be a REG or CONCAT, meaning they |
2360 | can't overlap unless they are the same because we never reuse that part | |
2361 | of the stack frame used for locals for spilled pseudos. */ | |
e16ceb8e | 2362 | if ((!MEM_P (rtlx) || !MEM_P (rtly)) |
10e949c1 | 2363 | && ! rtx_equal_p (rtlx, rtly)) |
ddb3d16f | 2364 | return 1; |
2365 | ||
04ec15fa | 2366 | /* If we have MEMs referring to different address spaces (which can |
bd1a81f7 | 2367 | potentially overlap), we cannot easily tell from the addresses |
2368 | whether the references overlap. */ | |
2369 | if (MEM_P (rtlx) && MEM_P (rtly) | |
2370 | && MEM_ADDR_SPACE (rtlx) != MEM_ADDR_SPACE (rtly)) | |
2371 | return 0; | |
2372 | ||
ddb3d16f | 2373 | /* Get the base and offsets of both decls. If either is a register, we |
2374 | know both are and are the same, so use that as the base. The only | |
2375 | we can avoid overlap is if we can deduce that they are nonoverlapping | |
2376 | pieces of that decl, which is very rare. */ | |
e16ceb8e | 2377 | basex = MEM_P (rtlx) ? XEXP (rtlx, 0) : rtlx; |
971ba038 | 2378 | if (GET_CODE (basex) == PLUS && CONST_INT_P (XEXP (basex, 1))) |
ddb3d16f | 2379 | offsetx = INTVAL (XEXP (basex, 1)), basex = XEXP (basex, 0); |
2380 | ||
e16ceb8e | 2381 | basey = MEM_P (rtly) ? XEXP (rtly, 0) : rtly; |
971ba038 | 2382 | if (GET_CODE (basey) == PLUS && CONST_INT_P (XEXP (basey, 1))) |
ddb3d16f | 2383 | offsety = INTVAL (XEXP (basey, 1)), basey = XEXP (basey, 0); |
2384 | ||
c9c48480 | 2385 | /* If the bases are different, we know they do not overlap if both |
9342ee68 | 2386 | are constants or if one is a constant and the other a pointer into the |
c9c48480 | 2387 | stack frame. Otherwise a different base means we can't tell if they |
2388 | overlap or not. */ | |
2389 | if (! rtx_equal_p (basex, basey)) | |
9342ee68 | 2390 | return ((CONSTANT_P (basex) && CONSTANT_P (basey)) |
2391 | || (CONSTANT_P (basex) && REG_P (basey) | |
2392 | && REGNO_PTR_FRAME_P (REGNO (basey))) | |
2393 | || (CONSTANT_P (basey) && REG_P (basex) | |
2394 | && REGNO_PTR_FRAME_P (REGNO (basex)))); | |
ddb3d16f | 2395 | |
a84256aa | 2396 | /* Offset based disambiguation not appropriate for loop invariant */ |
2397 | if (loop_invariant) | |
2398 | return 0; | |
2399 | ||
e16ceb8e | 2400 | sizex = (!MEM_P (rtlx) ? (int) GET_MODE_SIZE (GET_MODE (rtlx)) |
5b2a69fa | 2401 | : MEM_SIZE_KNOWN_P (rtlx) ? MEM_SIZE (rtlx) |
ddb3d16f | 2402 | : -1); |
e16ceb8e | 2403 | sizey = (!MEM_P (rtly) ? (int) GET_MODE_SIZE (GET_MODE (rtly)) |
5b2a69fa | 2404 | : MEM_SIZE_KNOWN_P (rtly) ? MEM_SIZE (rtly) |
2405 | : -1); | |
ddb3d16f | 2406 | |
dc2dfdb6 | 2407 | /* If we have an offset for either memref, it can update the values computed |
2408 | above. */ | |
da443c27 | 2409 | if (moffsetx_known_p) |
2410 | offsetx += moffsetx, sizex -= moffsetx; | |
2411 | if (moffsety_known_p) | |
2412 | offsety += moffsety, sizey -= moffsety; | |
ddb3d16f | 2413 | |
dc2dfdb6 | 2414 | /* If a memref has both a size and an offset, we can use the smaller size. |
203c9fee | 2415 | We can't do this if the offset isn't known because we must view this |
dc2dfdb6 | 2416 | memref as being anywhere inside the DECL's MEM. */ |
da443c27 | 2417 | if (MEM_SIZE_KNOWN_P (x) && moffsetx_known_p) |
5b2a69fa | 2418 | sizex = MEM_SIZE (x); |
da443c27 | 2419 | if (MEM_SIZE_KNOWN_P (y) && moffsety_known_p) |
5b2a69fa | 2420 | sizey = MEM_SIZE (y); |
ddb3d16f | 2421 | |
2422 | /* Put the values of the memref with the lower offset in X's values. */ | |
2423 | if (offsetx > offsety) | |
2424 | { | |
2425 | tem = offsetx, offsetx = offsety, offsety = tem; | |
2426 | tem = sizex, sizex = sizey, sizey = tem; | |
2427 | } | |
2428 | ||
2429 | /* If we don't know the size of the lower-offset value, we can't tell | |
2430 | if they conflict. Otherwise, we do the test. */ | |
906923ba | 2431 | return sizex >= 0 && offsety >= offsetx + sizex; |
ddb3d16f | 2432 | } |
2433 | ||
02f77cc4 | 2434 | /* Helper for true_dependence and canon_true_dependence. |
2435 | Checks for true dependence: X is read after store in MEM takes place. | |
ea0cb7ae | 2436 | |
02f77cc4 | 2437 | If MEM_CANONICALIZED is FALSE, then X_ADDR and MEM_ADDR should be |
2438 | NULL_RTX, and the canonical addresses of MEM and X are both computed | |
2439 | here. If MEM_CANONICALIZED, then MEM must be already canonicalized. | |
2440 | ||
2441 | If X_ADDR is non-NULL, it is used in preference of XEXP (x, 0). | |
2442 | ||
2443 | Returns 1 if there is a true dependence, 0 otherwise. */ | |
2444 | ||
2445 | static int | |
2446 | true_dependence_1 (const_rtx mem, enum machine_mode mem_mode, rtx mem_addr, | |
376a287d | 2447 | const_rtx x, rtx x_addr, bool mem_canonicalized) |
ea0cb7ae | 2448 | { |
9ee7cac5 | 2449 | rtx base; |
40f9e245 | 2450 | int ret; |
ea0cb7ae | 2451 | |
02f77cc4 | 2452 | gcc_checking_assert (mem_canonicalized ? (mem_addr != NULL_RTX) |
2453 | : (mem_addr == NULL_RTX && x_addr == NULL_RTX)); | |
2454 | ||
ea0cb7ae | 2455 | if (MEM_VOLATILE_P (x) && MEM_VOLATILE_P (mem)) |
2456 | return 1; | |
2457 | ||
a57f4c4d | 2458 | /* (mem:BLK (scratch)) is a special mechanism to conflict with everything. |
67837170 | 2459 | This is used in epilogue deallocation functions, and in cselib. */ |
a57f4c4d | 2460 | if (GET_MODE (x) == BLKmode && GET_CODE (XEXP (x, 0)) == SCRATCH) |
2461 | return 1; | |
2462 | if (GET_MODE (mem) == BLKmode && GET_CODE (XEXP (mem, 0)) == SCRATCH) | |
2463 | return 1; | |
c94cfd1c | 2464 | if (MEM_ALIAS_SET (x) == ALIAS_SET_MEMORY_BARRIER |
2465 | || MEM_ALIAS_SET (mem) == ALIAS_SET_MEMORY_BARRIER) | |
2466 | return 1; | |
a57f4c4d | 2467 | |
b04fab2a | 2468 | /* Read-only memory is by definition never modified, and therefore can't |
2469 | conflict with anything. We don't expect to find read-only set on MEM, | |
a53ff4c1 | 2470 | but stupid user tricks can produce them, so don't die. */ |
b04fab2a | 2471 | if (MEM_READONLY_P (x)) |
ea0cb7ae | 2472 | return 0; |
2473 | ||
04ec15fa | 2474 | /* If we have MEMs referring to different address spaces (which can |
bd1a81f7 | 2475 | potentially overlap), we cannot easily tell from the addresses |
2476 | whether the references overlap. */ | |
2477 | if (MEM_ADDR_SPACE (mem) != MEM_ADDR_SPACE (x)) | |
2478 | return 1; | |
2479 | ||
02f77cc4 | 2480 | if (! mem_addr) |
2481 | { | |
2482 | mem_addr = XEXP (mem, 0); | |
2483 | if (mem_mode == VOIDmode) | |
2484 | mem_mode = GET_MODE (mem); | |
2485 | } | |
2486 | ||
2487 | if (! x_addr) | |
288debed | 2488 | { |
62696140 | 2489 | x_addr = XEXP (x, 0); |
2490 | if (!((GET_CODE (x_addr) == VALUE | |
2491 | && GET_CODE (mem_addr) != VALUE | |
2492 | && reg_mentioned_p (x_addr, mem_addr)) | |
2493 | || (GET_CODE (x_addr) != VALUE | |
2494 | && GET_CODE (mem_addr) == VALUE | |
2495 | && reg_mentioned_p (mem_addr, x_addr)))) | |
2496 | { | |
2497 | x_addr = get_addr (x_addr); | |
2498 | if (! mem_canonicalized) | |
2499 | mem_addr = get_addr (mem_addr); | |
2500 | } | |
288debed | 2501 | } |
1617c276 | 2502 | |
002fe3cb | 2503 | base = find_base_term (x_addr); |
2504 | if (base && (GET_CODE (base) == LABEL_REF | |
2505 | || (GET_CODE (base) == SYMBOL_REF | |
2506 | && CONSTANT_POOL_ADDRESS_P (base)))) | |
2507 | return 0; | |
2508 | ||
1617c276 | 2509 | if (! base_alias_check (x_addr, mem_addr, GET_MODE (x), mem_mode)) |
3ad3a0ed | 2510 | return 0; |
2511 | ||
1617c276 | 2512 | x_addr = canon_rtx (x_addr); |
02f77cc4 | 2513 | if (!mem_canonicalized) |
2514 | mem_addr = canon_rtx (mem_addr); | |
35d3eedd | 2515 | |
40f9e245 | 2516 | if ((ret = memrefs_conflict_p (GET_MODE_SIZE (mem_mode), mem_addr, |
2517 | SIZE_FOR_MODE (x), x_addr, 0)) != -1) | |
2518 | return ret; | |
2519 | ||
2520 | if (DIFFERENT_ALIAS_SETS_P (x, mem)) | |
2521 | return 0; | |
2522 | ||
a84256aa | 2523 | if (nonoverlapping_memrefs_p (mem, x, false)) |
4958fe6e | 2524 | return 0; |
2525 | ||
6a0934dd | 2526 | if (aliases_everything_p (x)) |
4958fe6e | 2527 | return 1; |
2528 | ||
35a3065a | 2529 | /* We cannot use aliases_everything_p to test MEM, since we must look |
02f77cc4 | 2530 | at MEM_ADDR, rather than XEXP (mem, 0). */ |
3bddda63 | 2531 | if (GET_CODE (mem_addr) == AND) |
8e802be9 | 2532 | return 1; |
2533 | ||
02f77cc4 | 2534 | /* ??? In true_dependence we also allow BLKmode to alias anything. Why |
8e802be9 | 2535 | don't we do this in anti_dependence and output_dependence? */ |
2536 | if (mem_mode == BLKmode || GET_MODE (x) == BLKmode) | |
2537 | return 1; | |
2538 | ||
3a443843 | 2539 | return rtx_refs_may_alias_p (x, mem, true); |
8e802be9 | 2540 | } |
2541 | ||
02f77cc4 | 2542 | /* True dependence: X is read after store in MEM takes place. */ |
2543 | ||
2544 | int | |
376a287d | 2545 | true_dependence (const_rtx mem, enum machine_mode mem_mode, const_rtx x) |
02f77cc4 | 2546 | { |
2547 | return true_dependence_1 (mem, mem_mode, NULL_RTX, | |
376a287d | 2548 | x, NULL_RTX, /*mem_canonicalized=*/false); |
02f77cc4 | 2549 | } |
2550 | ||
8e802be9 | 2551 | /* Canonical true dependence: X is read after store in MEM takes place. |
9342ee68 | 2552 | Variant of true_dependence which assumes MEM has already been |
2553 | canonicalized (hence we no longer do that here). | |
02f77cc4 | 2554 | The mem_addr argument has been added, since true_dependence_1 computed |
2555 | this value prior to canonicalizing. */ | |
8e802be9 | 2556 | |
2557 | int | |
52d07779 | 2558 | canon_true_dependence (const_rtx mem, enum machine_mode mem_mode, rtx mem_addr, |
376a287d | 2559 | const_rtx x, rtx x_addr) |
8e802be9 | 2560 | { |
02f77cc4 | 2561 | return true_dependence_1 (mem, mem_mode, mem_addr, |
376a287d | 2562 | x, x_addr, /*mem_canonicalized=*/true); |
ea0cb7ae | 2563 | } |
2564 | ||
d10cfa8d | 2565 | /* Returns nonzero if a write to X might alias a previous read from |
b04fab2a | 2566 | (or, if WRITEP is nonzero, a write to) MEM. */ |
ea0cb7ae | 2567 | |
fe69e4fe | 2568 | static int |
52d07779 | 2569 | write_dependence_p (const_rtx mem, const_rtx x, int writep) |
ea0cb7ae | 2570 | { |
35d3eedd | 2571 | rtx x_addr, mem_addr; |
9ee7cac5 | 2572 | rtx base; |
40f9e245 | 2573 | int ret; |
35d3eedd | 2574 | |
ea0cb7ae | 2575 | if (MEM_VOLATILE_P (x) && MEM_VOLATILE_P (mem)) |
2576 | return 1; | |
2577 | ||
a57f4c4d | 2578 | /* (mem:BLK (scratch)) is a special mechanism to conflict with everything. |
2579 | This is used in epilogue deallocation functions. */ | |
2580 | if (GET_MODE (x) == BLKmode && GET_CODE (XEXP (x, 0)) == SCRATCH) | |
2581 | return 1; | |
2582 | if (GET_MODE (mem) == BLKmode && GET_CODE (XEXP (mem, 0)) == SCRATCH) | |
2583 | return 1; | |
c94cfd1c | 2584 | if (MEM_ALIAS_SET (x) == ALIAS_SET_MEMORY_BARRIER |
2585 | || MEM_ALIAS_SET (mem) == ALIAS_SET_MEMORY_BARRIER) | |
2586 | return 1; | |
a57f4c4d | 2587 | |
b04fab2a | 2588 | /* A read from read-only memory can't conflict with read-write memory. */ |
2589 | if (!writep && MEM_READONLY_P (mem)) | |
2590 | return 0; | |
002fe3cb | 2591 | |
04ec15fa | 2592 | /* If we have MEMs referring to different address spaces (which can |
bd1a81f7 | 2593 | potentially overlap), we cannot easily tell from the addresses |
2594 | whether the references overlap. */ | |
2595 | if (MEM_ADDR_SPACE (mem) != MEM_ADDR_SPACE (x)) | |
2596 | return 1; | |
2597 | ||
288debed | 2598 | x_addr = XEXP (x, 0); |
2599 | mem_addr = XEXP (mem, 0); | |
2600 | if (!((GET_CODE (x_addr) == VALUE | |
2601 | && GET_CODE (mem_addr) != VALUE | |
2602 | && reg_mentioned_p (x_addr, mem_addr)) | |
2603 | || (GET_CODE (x_addr) != VALUE | |
2604 | && GET_CODE (mem_addr) == VALUE | |
2605 | && reg_mentioned_p (mem_addr, x_addr)))) | |
2606 | { | |
2607 | x_addr = get_addr (x_addr); | |
2608 | mem_addr = get_addr (mem_addr); | |
2609 | } | |
002fe3cb | 2610 | |
9ee7cac5 | 2611 | if (! writep) |
2612 | { | |
002fe3cb | 2613 | base = find_base_term (mem_addr); |
9ee7cac5 | 2614 | if (base && (GET_CODE (base) == LABEL_REF |
2615 | || (GET_CODE (base) == SYMBOL_REF | |
2616 | && CONSTANT_POOL_ADDRESS_P (base)))) | |
2617 | return 0; | |
2618 | } | |
2619 | ||
1617c276 | 2620 | if (! base_alias_check (x_addr, mem_addr, GET_MODE (x), |
2621 | GET_MODE (mem))) | |
b5ba9f3a | 2622 | return 0; |
2623 | ||
1617c276 | 2624 | x_addr = canon_rtx (x_addr); |
2625 | mem_addr = canon_rtx (mem_addr); | |
35d3eedd | 2626 | |
40f9e245 | 2627 | if ((ret = memrefs_conflict_p (SIZE_FOR_MODE (mem), mem_addr, |
2628 | SIZE_FOR_MODE (x), x_addr, 0)) != -1) | |
2629 | return ret; | |
2630 | ||
a84256aa | 2631 | if (nonoverlapping_memrefs_p (x, mem, false)) |
6a0934dd | 2632 | return 0; |
2633 | ||
3a443843 | 2634 | return rtx_refs_may_alias_p (x, mem, false); |
6a0934dd | 2635 | } |
2636 | ||
2637 | /* Anti dependence: X is written after read in MEM takes place. */ | |
2638 | ||
2639 | int | |
52d07779 | 2640 | anti_dependence (const_rtx mem, const_rtx x) |
6a0934dd | 2641 | { |
b04fab2a | 2642 | return write_dependence_p (mem, x, /*writep=*/0); |
ea0cb7ae | 2643 | } |
2644 | ||
2645 | /* Output dependence: X is written after store in MEM takes place. */ | |
2646 | ||
2647 | int | |
52d07779 | 2648 | output_dependence (const_rtx mem, const_rtx x) |
ea0cb7ae | 2649 | { |
b04fab2a | 2650 | return write_dependence_p (mem, x, /*writep=*/1); |
ea0cb7ae | 2651 | } |
c352cbc7 | 2652 | \f |
35d3eedd | 2653 | |
a84256aa | 2654 | |
2655 | /* Check whether X may be aliased with MEM. Don't do offset-based | |
2656 | memory disambiguation & TBAA. */ | |
2657 | int | |
2658 | may_alias_p (const_rtx mem, const_rtx x) | |
2659 | { | |
2660 | rtx x_addr, mem_addr; | |
a84256aa | 2661 | |
2662 | if (MEM_VOLATILE_P (x) && MEM_VOLATILE_P (mem)) | |
2663 | return 1; | |
2664 | ||
2665 | /* ??? In true_dependence we also allow BLKmode to alias anything. */ | |
2666 | if (GET_MODE (mem) == BLKmode || GET_MODE (x) == BLKmode) | |
2667 | return 1; | |
2668 | ||
2669 | if (MEM_ALIAS_SET (x) == ALIAS_SET_MEMORY_BARRIER | |
2670 | || MEM_ALIAS_SET (mem) == ALIAS_SET_MEMORY_BARRIER) | |
2671 | return 1; | |
2672 | ||
2673 | /* Read-only memory is by definition never modified, and therefore can't | |
2674 | conflict with anything. We don't expect to find read-only set on MEM, | |
2675 | but stupid user tricks can produce them, so don't die. */ | |
2676 | if (MEM_READONLY_P (x)) | |
2677 | return 0; | |
2678 | ||
04ec15fa | 2679 | /* If we have MEMs referring to different address spaces (which can |
a84256aa | 2680 | potentially overlap), we cannot easily tell from the addresses |
2681 | whether the references overlap. */ | |
2682 | if (MEM_ADDR_SPACE (mem) != MEM_ADDR_SPACE (x)) | |
2683 | return 1; | |
2684 | ||
2685 | x_addr = XEXP (x, 0); | |
2686 | mem_addr = XEXP (mem, 0); | |
2687 | if (!((GET_CODE (x_addr) == VALUE | |
2688 | && GET_CODE (mem_addr) != VALUE | |
2689 | && reg_mentioned_p (x_addr, mem_addr)) | |
2690 | || (GET_CODE (x_addr) != VALUE | |
2691 | && GET_CODE (mem_addr) == VALUE | |
2692 | && reg_mentioned_p (mem_addr, x_addr)))) | |
2693 | { | |
2694 | x_addr = get_addr (x_addr); | |
2695 | mem_addr = get_addr (mem_addr); | |
2696 | } | |
2697 | ||
2698 | if (! base_alias_check (x_addr, mem_addr, GET_MODE (x), GET_MODE (mem_addr))) | |
2699 | return 0; | |
2700 | ||
2701 | x_addr = canon_rtx (x_addr); | |
2702 | mem_addr = canon_rtx (mem_addr); | |
2703 | ||
2704 | if (nonoverlapping_memrefs_p (mem, x, true)) | |
2705 | return 0; | |
2706 | ||
2707 | if (aliases_everything_p (x)) | |
2708 | return 1; | |
2709 | ||
2710 | /* We cannot use aliases_everything_p to test MEM, since we must look | |
2711 | at MEM_ADDR, rather than XEXP (mem, 0). */ | |
3bddda63 | 2712 | if (GET_CODE (mem_addr) == AND) |
a84256aa | 2713 | return 1; |
2714 | ||
a84256aa | 2715 | /* TBAA not valid for loop_invarint */ |
2716 | return rtx_refs_may_alias_p (x, mem, false); | |
2717 | } | |
2718 | ||
35d3eedd | 2719 | void |
6d8b68a3 | 2720 | init_alias_target (void) |
35d3eedd | 2721 | { |
19cb6b50 | 2722 | int i; |
35d3eedd | 2723 | |
4e56ceb1 | 2724 | if (!arg_base_value) |
2725 | arg_base_value = gen_rtx_ADDRESS (VOIDmode, 0); | |
2726 | ||
6d8b68a3 | 2727 | memset (static_reg_base_value, 0, sizeof static_reg_base_value); |
2728 | ||
35d3eedd | 2729 | for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) |
2730 | /* Check whether this register can hold an incoming pointer | |
2731 | argument. FUNCTION_ARG_REGNO_P tests outgoing register | |
2c0e001b | 2732 | numbers, so translate if necessary due to register windows. */ |
35d3eedd | 2733 | if (FUNCTION_ARG_REGNO_P (OUTGOING_REGNO (i)) |
2734 | && HARD_REGNO_MODE_OK (i, Pmode)) | |
4e56ceb1 | 2735 | static_reg_base_value[i] = arg_base_value; |
2736 | ||
2737 | static_reg_base_value[STACK_POINTER_REGNUM] = unique_base_value (-1); | |
2738 | static_reg_base_value[ARG_POINTER_REGNUM] = unique_base_value (-2); | |
2739 | static_reg_base_value[FRAME_POINTER_REGNUM] = unique_base_value (-3); | |
5ae82d58 | 2740 | #if !HARD_FRAME_POINTER_IS_FRAME_POINTER |
4e56ceb1 | 2741 | static_reg_base_value[HARD_FRAME_POINTER_REGNUM] = unique_base_value (-4); |
0c7f5242 | 2742 | #endif |
2743 | } | |
2744 | ||
c7bf7428 | 2745 | /* Set MEMORY_MODIFIED when X modifies DATA (that is assumed |
2746 | to be memory reference. */ | |
2747 | static bool memory_modified; | |
2748 | static void | |
a80f1c6c | 2749 | memory_modified_1 (rtx x, const_rtx pat ATTRIBUTE_UNUSED, void *data) |
c7bf7428 | 2750 | { |
e16ceb8e | 2751 | if (MEM_P (x)) |
c7bf7428 | 2752 | { |
5493cb9a | 2753 | if (anti_dependence (x, (const_rtx)data) || output_dependence (x, (const_rtx)data)) |
c7bf7428 | 2754 | memory_modified = true; |
2755 | } | |
2756 | } | |
2757 | ||
2758 | ||
2759 | /* Return true when INSN possibly modify memory contents of MEM | |
0c6d8c36 | 2760 | (i.e. address can be modified). */ |
c7bf7428 | 2761 | bool |
5493cb9a | 2762 | memory_modified_in_insn_p (const_rtx mem, const_rtx insn) |
c7bf7428 | 2763 | { |
2764 | if (!INSN_P (insn)) | |
2765 | return false; | |
2766 | memory_modified = false; | |
a80f1c6c | 2767 | note_stores (PATTERN (insn), memory_modified_1, CONST_CAST_RTX(mem)); |
c7bf7428 | 2768 | return memory_modified; |
2769 | } | |
2770 | ||
73f5c1e3 | 2771 | /* Initialize the aliasing machinery. Initialize the REG_KNOWN_VALUE |
2772 | array. */ | |
2773 | ||
ea0cb7ae | 2774 | void |
aecda0d6 | 2775 | init_alias_analysis (void) |
ea0cb7ae | 2776 | { |
e5b19a8c | 2777 | unsigned int maxreg = max_reg_num (); |
37b864db | 2778 | int changed, pass; |
19cb6b50 | 2779 | int i; |
2780 | unsigned int ui; | |
2781 | rtx insn; | |
ea0cb7ae | 2782 | |
376c21d1 | 2783 | timevar_push (TV_ALIAS_ANALYSIS); |
2784 | ||
12cb06ad | 2785 | reg_known_value_size = maxreg - FIRST_PSEUDO_REGISTER; |
ba72912a | 2786 | reg_known_value = ggc_alloc_cleared_vec_rtx (reg_known_value_size); |
364c0c59 | 2787 | reg_known_equiv_p = XCNEWVEC (bool, reg_known_value_size); |
ea0cb7ae | 2788 | |
980f3285 | 2789 | /* If we have memory allocated from the previous run, use it. */ |
e5b19a8c | 2790 | if (old_reg_base_value) |
980f3285 | 2791 | reg_base_value = old_reg_base_value; |
2792 | ||
2793 | if (reg_base_value) | |
2794 | VEC_truncate (rtx, reg_base_value, 0); | |
2795 | ||
e85c2c2d | 2796 | VEC_safe_grow_cleared (rtx, gc, reg_base_value, maxreg); |
c2e94b25 | 2797 | |
4c36ffe6 | 2798 | new_reg_base_value = XNEWVEC (rtx, maxreg); |
2799 | reg_seen = XNEWVEC (char, maxreg); | |
06c5c535 | 2800 | |
2801 | /* The basic idea is that each pass through this loop will use the | |
2802 | "constant" information from the previous pass to propagate alias | |
2803 | information through another level of assignments. | |
2804 | ||
2805 | This could get expensive if the assignment chains are long. Maybe | |
2806 | we should throttle the number of iterations, possibly based on | |
35d3eedd | 2807 | the optimization level or flag_expensive_optimizations. |
06c5c535 | 2808 | |
2809 | We could propagate more information in the first pass by making use | |
3072d30e | 2810 | of DF_REG_DEF_COUNT to determine immediately that the alias information |
37b864db | 2811 | for a pseudo is "constant". |
2812 | ||
2813 | A program with an uninitialized variable can cause an infinite loop | |
2814 | here. Instead of doing a full dataflow analysis to detect such problems | |
2815 | we just cap the number of iterations for the loop. | |
2816 | ||
2817 | The state of the arrays for the set chain in question does not matter | |
2818 | since the program has undefined behavior. */ | |
35d3eedd | 2819 | |
37b864db | 2820 | pass = 0; |
35d3eedd | 2821 | do |
06c5c535 | 2822 | { |
2823 | /* Assume nothing will change this iteration of the loop. */ | |
2824 | changed = 0; | |
2825 | ||
06c5c535 | 2826 | /* We want to assign the same IDs each iteration of this loop, so |
4e56ceb1 | 2827 | start counting from one each iteration of the loop. */ |
2828 | unique_id = 1; | |
06c5c535 | 2829 | |
35a3065a | 2830 | /* We're at the start of the function each iteration through the |
06c5c535 | 2831 | loop, so we're copying arguments. */ |
e5be0cc6 | 2832 | copying_arguments = true; |
ea0cb7ae | 2833 | |
35d3eedd | 2834 | /* Wipe the potential alias information clean for this pass. */ |
e5b19a8c | 2835 | memset (new_reg_base_value, 0, maxreg * sizeof (rtx)); |
f0ad0992 | 2836 | |
35d3eedd | 2837 | /* Wipe the reg_seen array clean. */ |
e5b19a8c | 2838 | memset (reg_seen, 0, maxreg); |
ea0cb7ae | 2839 | |
35d3eedd | 2840 | /* Mark all hard registers which may contain an address. |
2841 | The stack, frame and argument pointers may contain an address. | |
2842 | An argument register which can hold a Pmode value may contain | |
2843 | an address even if it is not in BASE_REGS. | |
f0ad0992 | 2844 | |
35d3eedd | 2845 | The address expression is VOIDmode for an argument and |
2846 | Pmode for other registers. */ | |
2847 | ||
e26fd9bc | 2848 | memcpy (new_reg_base_value, static_reg_base_value, |
2849 | FIRST_PSEUDO_REGISTER * sizeof (rtx)); | |
35d3eedd | 2850 | |
06c5c535 | 2851 | /* Walk the insns adding values to the new_reg_base_value array. */ |
2852 | for (insn = get_insns (); insn; insn = NEXT_INSN (insn)) | |
ea0cb7ae | 2853 | { |
9204e736 | 2854 | if (INSN_P (insn)) |
06c5c535 | 2855 | { |
35d3eedd | 2856 | rtx note, set; |
d6cb6164 | 2857 | |
2858 | #if defined (HAVE_prologue) || defined (HAVE_epilogue) | |
35a3065a | 2859 | /* The prologue/epilogue insns are not threaded onto the |
fd037eda | 2860 | insn chain until after reload has completed. Thus, |
2861 | there is no sense wasting time checking if INSN is in | |
2862 | the prologue/epilogue until after reload has completed. */ | |
2863 | if (reload_completed | |
2864 | && prologue_epilogue_contains (insn)) | |
d6cb6164 | 2865 | continue; |
2866 | #endif | |
2867 | ||
06c5c535 | 2868 | /* If this insn has a noalias note, process it, Otherwise, |
a0c938f0 | 2869 | scan for sets. A simple set will have no side effects |
2870 | which could change the base value of any other register. */ | |
35d3eedd | 2871 | |
06c5c535 | 2872 | if (GET_CODE (PATTERN (insn)) == SET |
d6cb6164 | 2873 | && REG_NOTES (insn) != 0 |
2874 | && find_reg_note (insn, REG_NOALIAS, NULL_RTX)) | |
ec8895d7 | 2875 | record_set (SET_DEST (PATTERN (insn)), NULL_RTX, NULL); |
06c5c535 | 2876 | else |
ec8895d7 | 2877 | note_stores (PATTERN (insn), record_set, NULL); |
35d3eedd | 2878 | |
2879 | set = single_set (insn); | |
2880 | ||
2881 | if (set != 0 | |
8ad4c111 | 2882 | && REG_P (SET_DEST (set)) |
9db10305 | 2883 | && REGNO (SET_DEST (set)) >= FIRST_PSEUDO_REGISTER) |
35d3eedd | 2884 | { |
9db10305 | 2885 | unsigned int regno = REGNO (SET_DEST (set)); |
970c0c06 | 2886 | rtx src = SET_SRC (set); |
12cb06ad | 2887 | rtx t; |
970c0c06 | 2888 | |
750a330e | 2889 | note = find_reg_equal_equiv_note (insn); |
2890 | if (note && REG_NOTE_KIND (note) == REG_EQUAL | |
3072d30e | 2891 | && DF_REG_DEF_COUNT (regno) != 1) |
750a330e | 2892 | note = NULL_RTX; |
2893 | ||
2894 | if (note != NULL_RTX | |
970c0c06 | 2895 | && GET_CODE (XEXP (note, 0)) != EXPR_LIST |
97fb4325 | 2896 | && ! rtx_varies_p (XEXP (note, 0), 1) |
12cb06ad | 2897 | && ! reg_overlap_mentioned_p (SET_DEST (set), |
2898 | XEXP (note, 0))) | |
970c0c06 | 2899 | { |
12cb06ad | 2900 | set_reg_known_value (regno, XEXP (note, 0)); |
2901 | set_reg_known_equiv_p (regno, | |
2902 | REG_NOTE_KIND (note) == REG_EQUIV); | |
970c0c06 | 2903 | } |
3072d30e | 2904 | else if (DF_REG_DEF_COUNT (regno) == 1 |
970c0c06 | 2905 | && GET_CODE (src) == PLUS |
8ad4c111 | 2906 | && REG_P (XEXP (src, 0)) |
12cb06ad | 2907 | && (t = get_reg_known_value (REGNO (XEXP (src, 0)))) |
971ba038 | 2908 | && CONST_INT_P (XEXP (src, 1))) |
970c0c06 | 2909 | { |
12cb06ad | 2910 | t = plus_constant (t, INTVAL (XEXP (src, 1))); |
2911 | set_reg_known_value (regno, t); | |
2912 | set_reg_known_equiv_p (regno, 0); | |
970c0c06 | 2913 | } |
3072d30e | 2914 | else if (DF_REG_DEF_COUNT (regno) == 1 |
970c0c06 | 2915 | && ! rtx_varies_p (src, 1)) |
2916 | { | |
12cb06ad | 2917 | set_reg_known_value (regno, src); |
2918 | set_reg_known_equiv_p (regno, 0); | |
970c0c06 | 2919 | } |
35d3eedd | 2920 | } |
06c5c535 | 2921 | } |
6d7dc5b9 | 2922 | else if (NOTE_P (insn) |
ad4583d9 | 2923 | && NOTE_KIND (insn) == NOTE_INSN_FUNCTION_BEG) |
e5be0cc6 | 2924 | copying_arguments = false; |
35d3eedd | 2925 | } |
06c5c535 | 2926 | |
35d3eedd | 2927 | /* Now propagate values from new_reg_base_value to reg_base_value. */ |
d4473c84 | 2928 | gcc_assert (maxreg == (unsigned int) max_reg_num ()); |
a0c938f0 | 2929 | |
e5b19a8c | 2930 | for (ui = 0; ui < maxreg; ui++) |
35d3eedd | 2931 | { |
274c11d8 | 2932 | if (new_reg_base_value[ui] |
980f3285 | 2933 | && new_reg_base_value[ui] != VEC_index (rtx, reg_base_value, ui) |
e5b19a8c | 2934 | && ! rtx_equal_p (new_reg_base_value[ui], |
980f3285 | 2935 | VEC_index (rtx, reg_base_value, ui))) |
06c5c535 | 2936 | { |
980f3285 | 2937 | VEC_replace (rtx, reg_base_value, ui, new_reg_base_value[ui]); |
35d3eedd | 2938 | changed = 1; |
06c5c535 | 2939 | } |
ea0cb7ae | 2940 | } |
ea0cb7ae | 2941 | } |
35d3eedd | 2942 | while (changed && ++pass < MAX_ALIAS_LOOP_PASSES); |
ea0cb7ae | 2943 | |
2944 | /* Fill in the remaining entries. */ | |
12cb06ad | 2945 | for (i = 0; i < (int)reg_known_value_size; i++) |
ea0cb7ae | 2946 | if (reg_known_value[i] == 0) |
12cb06ad | 2947 | reg_known_value[i] = regno_reg_rtx[i + FIRST_PSEUDO_REGISTER]; |
ea0cb7ae | 2948 | |
7098ef4a | 2949 | /* Clean up. */ |
2950 | free (new_reg_base_value); | |
06c5c535 | 2951 | new_reg_base_value = 0; |
7098ef4a | 2952 | free (reg_seen); |
ea0cb7ae | 2953 | reg_seen = 0; |
376c21d1 | 2954 | timevar_pop (TV_ALIAS_ANALYSIS); |
ea0cb7ae | 2955 | } |
2956 | ||
1837e966 | 2957 | /* Equate REG_BASE_VALUE (reg1) to REG_BASE_VALUE (reg2). |
2958 | Special API for var-tracking pass purposes. */ | |
2959 | ||
2960 | void | |
2961 | vt_equate_reg_base_value (const_rtx reg1, const_rtx reg2) | |
2962 | { | |
2963 | VEC_replace (rtx, reg_base_value, REGNO (reg1), REG_BASE_VALUE (reg2)); | |
2964 | } | |
2965 | ||
ea0cb7ae | 2966 | void |
aecda0d6 | 2967 | end_alias_analysis (void) |
ea0cb7ae | 2968 | { |
e5b19a8c | 2969 | old_reg_base_value = reg_base_value; |
12cb06ad | 2970 | ggc_free (reg_known_value); |
ea0cb7ae | 2971 | reg_known_value = 0; |
c2e94b25 | 2972 | reg_known_value_size = 0; |
12cb06ad | 2973 | free (reg_known_equiv_p); |
7098ef4a | 2974 | reg_known_equiv_p = 0; |
ea0cb7ae | 2975 | } |
1f3233d1 | 2976 | |
2977 | #include "gt-alias.h" |