]>
Commit | Line | Data |
---|---|---|
ea0cb7ae | 1 | /* Alias analysis for GNU C |
3aea1f79 | 2 | Copyright (C) 1997-2014 Free Software Foundation, Inc. |
ea0cb7ae | 3 | Contributed by John Carr (jfc@mit.edu). |
4 | ||
f12b58b3 | 5 | This file is part of GCC. |
ea0cb7ae | 6 | |
f12b58b3 | 7 | GCC is free software; you can redistribute it and/or modify it under |
8 | the terms of the GNU General Public License as published by the Free | |
8c4c00c1 | 9 | Software Foundation; either version 3, or (at your option) any later |
f12b58b3 | 10 | version. |
ea0cb7ae | 11 | |
f12b58b3 | 12 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY |
13 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
14 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
15 | for more details. | |
ea0cb7ae | 16 | |
17 | You should have received a copy of the GNU General Public License | |
8c4c00c1 | 18 | along with GCC; see the file COPYING3. If not see |
19 | <http://www.gnu.org/licenses/>. */ | |
ea0cb7ae | 20 | |
21 | #include "config.h" | |
405711de | 22 | #include "system.h" |
805e22b2 | 23 | #include "coretypes.h" |
24 | #include "tm.h" | |
ea0cb7ae | 25 | #include "rtl.h" |
142131ed | 26 | #include "tree.h" |
9ed99284 | 27 | #include "varasm.h" |
28 | #include "expr.h" | |
7953c610 | 29 | #include "tm_p.h" |
0a893c29 | 30 | #include "function.h" |
4c74e6d9 | 31 | #include "alias.h" |
32 | #include "emit-rtl.h" | |
ea0cb7ae | 33 | #include "regs.h" |
34 | #include "hard-reg-set.h" | |
35 | #include "flags.h" | |
0b205f4c | 36 | #include "diagnostic-core.h" |
1617c276 | 37 | #include "cselib.h" |
a5b1863e | 38 | #include "splay-tree.h" |
b0278d39 | 39 | #include "langhooks.h" |
376c21d1 | 40 | #include "timevar.h" |
b9ed1410 | 41 | #include "dumpfile.h" |
4f8e9d5c | 42 | #include "target.h" |
3072d30e | 43 | #include "df.h" |
3a443843 | 44 | #include "tree-ssa-alias.h" |
45 | #include "pointer-set.h" | |
bc61cadb | 46 | #include "internal-fn.h" |
47 | #include "gimple-expr.h" | |
48 | #include "is-a.h" | |
073c1fd5 | 49 | #include "gimple.h" |
50 | #include "gimple-ssa.h" | |
f7d118a9 | 51 | |
52 | /* The aliasing API provided here solves related but different problems: | |
53 | ||
a0c938f0 | 54 | Say there exists (in c) |
f7d118a9 | 55 | |
56 | struct X { | |
57 | struct Y y1; | |
58 | struct Z z2; | |
59 | } x1, *px1, *px2; | |
60 | ||
61 | struct Y y2, *py; | |
62 | struct Z z2, *pz; | |
63 | ||
64 | ||
3c62ebbd | 65 | py = &x1.y1; |
f7d118a9 | 66 | px2 = &x1; |
67 | ||
68 | Consider the four questions: | |
69 | ||
70 | Can a store to x1 interfere with px2->y1? | |
71 | Can a store to x1 interfere with px2->z2? | |
f7d118a9 | 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 | |
9d75589a | 79 | a store through a pointer to an X can overwrite any field that is |
f7d118a9 | 80 | contained (recursively) in an X (unless we know that px1 != px2). |
81 | ||
3c62ebbd | 82 | The last two questions can be solved in the same way as the first |
83 | two questions but this is too conservative. The observation is | |
84 | that in some cases we can know which (if any) fields are addressed | |
85 | and if those addresses are used in bad ways. This analysis may be | |
86 | language specific. In C, arbitrary operations may be applied to | |
87 | pointers. However, there is some indication that this may be too | |
88 | conservative for some C++ types. | |
f7d118a9 | 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 | |
3c62ebbd | 94 | data structure that was used to represent the solution to these |
f7d118a9 | 95 | problems. We now have two similar but different data structures, |
3c62ebbd | 96 | The data structure to solve the last two questions is similar to |
97 | the first, but does not contain the fields whose address are never | |
98 | taken. For types that do escape the compilation unit, the data | |
99 | structures will have identical information. | |
f7d118a9 | 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 *); |
5abc99e4 | 154 | static int base_alias_check (rtx, rtx, rtx, rtx, enum machine_mode, |
aecda0d6 | 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); |
aecda0d6 | 160 | static tree decl_for_component_ref (tree); |
cb456db5 | 161 | static int write_dependence_p (const_rtx, |
162 | const_rtx, enum machine_mode, rtx, | |
163 | bool, bool, bool); | |
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 | ||
37b864db | 172 | /* Cap the number of passes we make over the insns propagating alias |
2808fda8 | 173 | information through set chains. |
174 | ??? 10 is a completely arbitrary choice. This should be based on the | |
175 | maximum loop depth in the CFG, but we do not have this information | |
176 | available (even if current_loops _is_ available). */ | |
37b864db | 177 | #define MAX_ALIAS_LOOP_PASSES 10 |
9342ee68 | 178 | |
ea0cb7ae | 179 | /* reg_base_value[N] gives an address to which register N is related. |
180 | If all sets after the first add or subtract to the current value | |
181 | or otherwise modify it so it does not point to a different top level | |
182 | object, reg_base_value[N] is equal to the address part of the source | |
e7c893ba | 183 | of the first set. |
184 | ||
185 | A base address can be an ADDRESS, SYMBOL_REF, or LABEL_REF. ADDRESS | |
4e56ceb1 | 186 | expressions represent three types of base: |
3a935f06 | 187 | |
4e56ceb1 | 188 | 1. incoming arguments. There is just one ADDRESS to represent all |
189 | arguments, since we do not know at this level whether accesses | |
190 | based on different arguments can alias. The ADDRESS has id 0. | |
3a935f06 | 191 | |
4e56ceb1 | 192 | 2. stack_pointer_rtx, frame_pointer_rtx, hard_frame_pointer_rtx |
193 | (if distinct from frame_pointer_rtx) and arg_pointer_rtx. | |
194 | Each of these rtxes has a separate ADDRESS associated with it, | |
195 | each with a negative id. | |
196 | ||
197 | GCC is (and is required to be) precise in which register it | |
198 | chooses to access a particular region of stack. We can therefore | |
199 | assume that accesses based on one of these rtxes do not alias | |
200 | accesses based on another of these rtxes. | |
201 | ||
202 | 3. bases that are derived from malloc()ed memory (REG_NOALIAS). | |
203 | Each such piece of memory has a separate ADDRESS associated | |
204 | with it, each with an id greater than 0. | |
205 | ||
206 | Accesses based on one ADDRESS do not alias accesses based on other | |
207 | ADDRESSes. Accesses based on ADDRESSes in groups (2) and (3) do not | |
208 | alias globals either; the ADDRESSes have Pmode to indicate this. | |
209 | The ADDRESS in group (1) _may_ alias globals; it has VOIDmode to | |
210 | indicate this. */ | |
e7c893ba | 211 | |
f1f41a6c | 212 | static GTY(()) vec<rtx, va_gc> *reg_base_value; |
c2e94b25 | 213 | static rtx *new_reg_base_value; |
e5b19a8c | 214 | |
4e56ceb1 | 215 | /* The single VOIDmode ADDRESS that represents all argument bases. |
216 | It has id 0. */ | |
217 | static GTY(()) rtx arg_base_value; | |
218 | ||
219 | /* Used to allocate unique ids to each REG_NOALIAS ADDRESS. */ | |
220 | static int unique_id; | |
221 | ||
e5b19a8c | 222 | /* We preserve the copy of old array around to avoid amount of garbage |
223 | produced. About 8% of garbage produced were attributed to this | |
224 | array. */ | |
f1f41a6c | 225 | static GTY((deletable)) vec<rtx, va_gc> *old_reg_base_value; |
083a2b5e | 226 | |
86e87ef6 | 227 | /* Values of XINT (address, 0) of Pmode ADDRESS rtxes for special |
228 | registers. */ | |
229 | #define UNIQUE_BASE_VALUE_SP -1 | |
230 | #define UNIQUE_BASE_VALUE_ARGP -2 | |
231 | #define UNIQUE_BASE_VALUE_FP -3 | |
232 | #define UNIQUE_BASE_VALUE_HFP -4 | |
233 | ||
eb2c25b5 | 234 | #define static_reg_base_value \ |
235 | (this_target_rtl->x_static_reg_base_value) | |
0c7f5242 | 236 | |
f1f41a6c | 237 | #define REG_BASE_VALUE(X) \ |
238 | (REGNO (X) < vec_safe_length (reg_base_value) \ | |
239 | ? (*reg_base_value)[REGNO (X)] : 0) | |
ea0cb7ae | 240 | |
73f5c1e3 | 241 | /* Vector indexed by N giving the initial (unchanging) value known for |
d6443ebe | 242 | pseudo-register N. This vector is initialized in init_alias_analysis, |
12cb06ad | 243 | and does not change until end_alias_analysis is called. */ |
f1f41a6c | 244 | static GTY(()) vec<rtx, va_gc> *reg_known_value; |
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. */ | |
d6443ebe | 258 | static sbitmap 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 | |
a5b1863e | 265 | /* The splay-tree used to store the various alias set entries. */ |
f1f41a6c | 266 | static GTY (()) vec<alias_set_entry, va_gc> *alias_sets; |
5badbbc0 | 267 | \f |
3a443843 | 268 | /* Build a decomposed reference object for querying the alias-oracle |
269 | from the MEM rtx and store it in *REF. | |
270 | Returns false if MEM is not suitable for the alias-oracle. */ | |
271 | ||
272 | static bool | |
273 | ao_ref_from_mem (ao_ref *ref, const_rtx mem) | |
274 | { | |
275 | tree expr = MEM_EXPR (mem); | |
276 | tree base; | |
277 | ||
278 | if (!expr) | |
279 | return false; | |
280 | ||
281 | ao_ref_init (ref, expr); | |
282 | ||
283 | /* Get the base of the reference and see if we have to reject or | |
284 | adjust it. */ | |
285 | base = ao_ref_base (ref); | |
286 | if (base == NULL_TREE) | |
287 | return false; | |
288 | ||
9bbb37b8 | 289 | /* The tree oracle doesn't like bases that are neither decls |
290 | nor indirect references of SSA names. */ | |
291 | if (!(DECL_P (base) | |
292 | || (TREE_CODE (base) == MEM_REF | |
293 | && TREE_CODE (TREE_OPERAND (base, 0)) == SSA_NAME) | |
294 | || (TREE_CODE (base) == TARGET_MEM_REF | |
295 | && TREE_CODE (TMR_BASE (base)) == SSA_NAME))) | |
7e9c660e | 296 | return false; |
3a443843 | 297 | |
298 | /* If this is a reference based on a partitioned decl replace the | |
9bbb37b8 | 299 | base with a MEM_REF of the pointer representative we |
3a443843 | 300 | created during stack slot partitioning. */ |
301 | if (TREE_CODE (base) == VAR_DECL | |
9bbb37b8 | 302 | && ! is_global_var (base) |
3a443843 | 303 | && cfun->gimple_df->decls_to_pointers != NULL) |
304 | { | |
305 | void *namep; | |
306 | namep = pointer_map_contains (cfun->gimple_df->decls_to_pointers, base); | |
307 | if (namep) | |
182cf5a9 | 308 | ref->base = build_simple_mem_ref (*(tree *)namep); |
7e9c660e | 309 | } |
3a443843 | 310 | |
311 | ref->ref_alias_set = MEM_ALIAS_SET (mem); | |
312 | ||
cbc80ce5 | 313 | /* If MEM_OFFSET or MEM_SIZE are unknown what we got from MEM_EXPR |
314 | is conservative, so trust it. */ | |
da443c27 | 315 | if (!MEM_OFFSET_KNOWN_P (mem) |
5b2a69fa | 316 | || !MEM_SIZE_KNOWN_P (mem)) |
cbc80ce5 | 317 | return true; |
1f1e1860 | 318 | |
3804203d | 319 | /* If the base decl is a parameter we can have negative MEM_OFFSET in |
320 | case of promoted subregs on bigendian targets. Trust the MEM_EXPR | |
321 | here. */ | |
da443c27 | 322 | if (MEM_OFFSET (mem) < 0 |
323 | && (MEM_SIZE (mem) + MEM_OFFSET (mem)) * BITS_PER_UNIT == ref->size) | |
3804203d | 324 | return true; |
325 | ||
cbc80ce5 | 326 | /* Otherwise continue and refine size and offset we got from analyzing |
327 | MEM_EXPR by using MEM_SIZE and MEM_OFFSET. */ | |
328 | ||
da443c27 | 329 | ref->offset += MEM_OFFSET (mem) * BITS_PER_UNIT; |
5b2a69fa | 330 | ref->size = MEM_SIZE (mem) * BITS_PER_UNIT; |
3804203d | 331 | |
332 | /* The MEM may extend into adjacent fields, so adjust max_size if | |
333 | necessary. */ | |
334 | if (ref->max_size != -1 | |
335 | && ref->size > ref->max_size) | |
336 | ref->max_size = ref->size; | |
337 | ||
338 | /* If MEM_OFFSET and MEM_SIZE get us outside of the base object of | |
339 | the MEM_EXPR punt. This happens for STRICT_ALIGNMENT targets a lot. */ | |
340 | if (MEM_EXPR (mem) != get_spill_slot_decl (false) | |
341 | && (ref->offset < 0 | |
342 | || (DECL_P (ref->base) | |
cd4547bf | 343 | && (!tree_fits_uhwi_p (DECL_SIZE (ref->base)) |
8c53c46c | 344 | || (tree_to_uhwi (DECL_SIZE (ref->base)) |
345 | < (unsigned HOST_WIDE_INT) (ref->offset + ref->size)))))) | |
3804203d | 346 | return false; |
3a443843 | 347 | |
348 | return true; | |
349 | } | |
350 | ||
351 | /* Query the alias-oracle on whether the two memory rtx X and MEM may | |
352 | alias. If TBAA_P is set also apply TBAA. Returns true if the | |
353 | two rtxen may alias, false otherwise. */ | |
354 | ||
355 | static bool | |
356 | rtx_refs_may_alias_p (const_rtx x, const_rtx mem, bool tbaa_p) | |
357 | { | |
358 | ao_ref ref1, ref2; | |
359 | ||
360 | if (!ao_ref_from_mem (&ref1, x) | |
361 | || !ao_ref_from_mem (&ref2, mem)) | |
362 | return true; | |
363 | ||
33dda198 | 364 | return refs_may_alias_p_1 (&ref1, &ref2, |
365 | tbaa_p | |
366 | && MEM_ALIAS_SET (x) != 0 | |
367 | && MEM_ALIAS_SET (mem) != 0); | |
3a443843 | 368 | } |
369 | ||
a5b1863e | 370 | /* Returns a pointer to the alias set entry for ALIAS_SET, if there is |
371 | such an entry, or NULL otherwise. */ | |
372 | ||
56bbdce4 | 373 | static inline alias_set_entry |
32c2fdea | 374 | get_alias_set_entry (alias_set_type alias_set) |
a5b1863e | 375 | { |
f1f41a6c | 376 | return (*alias_sets)[alias_set]; |
a5b1863e | 377 | } |
378 | ||
5badbbc0 | 379 | /* Returns nonzero if the alias sets for MEM1 and MEM2 are such that |
380 | the two MEMs cannot alias each other. */ | |
a5b1863e | 381 | |
56bbdce4 | 382 | static inline int |
52d07779 | 383 | mems_in_disjoint_alias_sets_p (const_rtx mem1, const_rtx mem2) |
a5b1863e | 384 | { |
a5b1863e | 385 | /* Perform a basic sanity check. Namely, that there are no alias sets |
386 | if we're not using strict aliasing. This helps to catch bugs | |
387 | whereby someone uses PUT_CODE, but doesn't clear MEM_ALIAS_SET, or | |
388 | where a MEM is allocated in some way other than by the use of | |
389 | gen_rtx_MEM, and the MEM_ALIAS_SET is not cleared. If we begin to | |
390 | use alias sets to indicate that spilled registers cannot alias each | |
391 | other, we might need to remove this check. */ | |
64db345d | 392 | gcc_assert (flag_strict_aliasing |
393 | || (!MEM_ALIAS_SET (mem1) && !MEM_ALIAS_SET (mem2))); | |
a5b1863e | 394 | |
387bc205 | 395 | return ! alias_sets_conflict_p (MEM_ALIAS_SET (mem1), MEM_ALIAS_SET (mem2)); |
396 | } | |
a5b1863e | 397 | |
387bc205 | 398 | /* Insert the NODE into the splay tree given by DATA. Used by |
399 | record_alias_subset via splay_tree_foreach. */ | |
400 | ||
401 | static int | |
aecda0d6 | 402 | insert_subset_children (splay_tree_node node, void *data) |
387bc205 | 403 | { |
404 | splay_tree_insert ((splay_tree) data, node->key, node->value); | |
405 | ||
406 | return 0; | |
407 | } | |
408 | ||
7d1f52b2 | 409 | /* Return true if the first alias set is a subset of the second. */ |
410 | ||
411 | bool | |
32c2fdea | 412 | alias_set_subset_of (alias_set_type set1, alias_set_type set2) |
7d1f52b2 | 413 | { |
414 | alias_set_entry ase; | |
415 | ||
416 | /* Everything is a subset of the "aliases everything" set. */ | |
417 | if (set2 == 0) | |
418 | return true; | |
419 | ||
420 | /* Otherwise, check if set1 is a subset of set2. */ | |
421 | ase = get_alias_set_entry (set2); | |
422 | if (ase != 0 | |
a5b373b8 | 423 | && (ase->has_zero_child |
6c711ccf | 424 | || splay_tree_lookup (ase->children, |
425 | (splay_tree_key) set1))) | |
7d1f52b2 | 426 | return true; |
427 | return false; | |
428 | } | |
429 | ||
387bc205 | 430 | /* Return 1 if the two specified alias sets may conflict. */ |
431 | ||
432 | int | |
32c2fdea | 433 | alias_sets_conflict_p (alias_set_type set1, alias_set_type set2) |
387bc205 | 434 | { |
435 | alias_set_entry ase; | |
436 | ||
879f881c | 437 | /* The easy case. */ |
438 | if (alias_sets_must_conflict_p (set1, set2)) | |
387bc205 | 439 | return 1; |
a5b1863e | 440 | |
f7c44134 | 441 | /* See if the first alias set is a subset of the second. */ |
387bc205 | 442 | ase = get_alias_set_entry (set1); |
851dfbff | 443 | if (ase != 0 |
444 | && (ase->has_zero_child | |
445 | || splay_tree_lookup (ase->children, | |
387bc205 | 446 | (splay_tree_key) set2))) |
447 | return 1; | |
a5b1863e | 448 | |
449 | /* Now do the same, but with the alias sets reversed. */ | |
387bc205 | 450 | ase = get_alias_set_entry (set2); |
851dfbff | 451 | if (ase != 0 |
452 | && (ase->has_zero_child | |
453 | || splay_tree_lookup (ase->children, | |
387bc205 | 454 | (splay_tree_key) set1))) |
455 | return 1; | |
a5b1863e | 456 | |
387bc205 | 457 | /* The two alias sets are distinct and neither one is the |
879f881c | 458 | child of the other. Therefore, they cannot conflict. */ |
387bc205 | 459 | return 0; |
a5b1863e | 460 | } |
c1628b55 | 461 | |
879f881c | 462 | /* Return 1 if the two specified alias sets will always conflict. */ |
c1628b55 | 463 | |
464 | int | |
32c2fdea | 465 | alias_sets_must_conflict_p (alias_set_type set1, alias_set_type set2) |
c1628b55 | 466 | { |
467 | if (set1 == 0 || set2 == 0 || set1 == set2) | |
468 | return 1; | |
469 | ||
470 | return 0; | |
471 | } | |
472 | ||
387bc205 | 473 | /* Return 1 if any MEM object of type T1 will always conflict (using the |
474 | dependency routines in this file) with any MEM object of type T2. | |
475 | This is used when allocating temporary storage. If T1 and/or T2 are | |
476 | NULL_TREE, it means we know nothing about the storage. */ | |
477 | ||
478 | int | |
aecda0d6 | 479 | objects_must_conflict_p (tree t1, tree t2) |
387bc205 | 480 | { |
32c2fdea | 481 | alias_set_type set1, set2; |
2bedf233 | 482 | |
218eb8d0 | 483 | /* If neither has a type specified, we don't know if they'll conflict |
484 | because we may be using them to store objects of various types, for | |
485 | example the argument and local variables areas of inlined functions. */ | |
8f2ba9c9 | 486 | if (t1 == 0 && t2 == 0) |
218eb8d0 | 487 | return 0; |
488 | ||
387bc205 | 489 | /* If they are the same type, they must conflict. */ |
490 | if (t1 == t2 | |
491 | /* Likewise if both are volatile. */ | |
492 | || (t1 != 0 && TYPE_VOLATILE (t1) && t2 != 0 && TYPE_VOLATILE (t2))) | |
493 | return 1; | |
494 | ||
2bedf233 | 495 | set1 = t1 ? get_alias_set (t1) : 0; |
496 | set2 = t2 ? get_alias_set (t2) : 0; | |
387bc205 | 497 | |
879f881c | 498 | /* We can't use alias_sets_conflict_p because we must make sure |
499 | that every subtype of t1 will conflict with every subtype of | |
2bedf233 | 500 | t2 for which a pair of subobjects of these respective subtypes |
501 | overlaps on the stack. */ | |
879f881c | 502 | return alias_sets_must_conflict_p (set1, set2); |
387bc205 | 503 | } |
504 | \f | |
d400f5e1 | 505 | /* Return the outermost parent of component present in the chain of |
506 | component references handled by get_inner_reference in T with the | |
507 | following property: | |
508 | - the component is non-addressable, or | |
509 | - the parent has alias set zero, | |
510 | or NULL_TREE if no such parent exists. In the former cases, the alias | |
511 | set of this parent is the alias set that must be used for T itself. */ | |
512 | ||
513 | tree | |
514 | component_uses_parent_alias_set_from (const_tree t) | |
bd32979f | 515 | { |
d400f5e1 | 516 | const_tree found = NULL_TREE; |
1f9b622b | 517 | |
d400f5e1 | 518 | while (handled_component_p (t)) |
519 | { | |
1f9b622b | 520 | switch (TREE_CODE (t)) |
521 | { | |
522 | case COMPONENT_REF: | |
523 | if (DECL_NONADDRESSABLE_P (TREE_OPERAND (t, 1))) | |
d400f5e1 | 524 | found = t; |
1f9b622b | 525 | break; |
526 | ||
527 | case ARRAY_REF: | |
528 | case ARRAY_RANGE_REF: | |
529 | if (TYPE_NONALIASED_COMPONENT (TREE_TYPE (TREE_OPERAND (t, 0)))) | |
d400f5e1 | 530 | found = t; |
1f9b622b | 531 | break; |
532 | ||
533 | case REALPART_EXPR: | |
534 | case IMAGPART_EXPR: | |
535 | break; | |
536 | ||
d400f5e1 | 537 | case BIT_FIELD_REF: |
538 | case VIEW_CONVERT_EXPR: | |
1f9b622b | 539 | /* Bitfields and casts are never addressable. */ |
d400f5e1 | 540 | found = t; |
541 | break; | |
542 | ||
543 | default: | |
544 | gcc_unreachable (); | |
1f9b622b | 545 | } |
546 | ||
d400f5e1 | 547 | if (get_alias_set (TREE_TYPE (TREE_OPERAND (t, 0))) == 0) |
548 | found = t; | |
549 | ||
1f9b622b | 550 | t = TREE_OPERAND (t, 0); |
551 | } | |
d400f5e1 | 552 | |
553 | if (found) | |
554 | return TREE_OPERAND (found, 0); | |
555 | ||
556 | return NULL_TREE; | |
bd32979f | 557 | } |
558 | ||
bd93e4f8 | 559 | |
560 | /* Return whether the pointer-type T effective for aliasing may | |
561 | access everything and thus the reference has to be assigned | |
562 | alias-set zero. */ | |
563 | ||
564 | static bool | |
565 | ref_all_alias_ptr_type_p (const_tree t) | |
566 | { | |
567 | return (TREE_CODE (TREE_TYPE (t)) == VOID_TYPE | |
568 | || TYPE_REF_CAN_ALIAS_ALL (t)); | |
569 | } | |
570 | ||
dd277d48 | 571 | /* Return the alias set for the memory pointed to by T, which may be |
572 | either a type or an expression. Return -1 if there is nothing | |
573 | special about dereferencing T. */ | |
574 | ||
575 | static alias_set_type | |
576 | get_deref_alias_set_1 (tree t) | |
577 | { | |
8115f0af | 578 | /* All we care about is the type. */ |
dd277d48 | 579 | if (! TYPE_P (t)) |
8115f0af | 580 | t = TREE_TYPE (t); |
dd277d48 | 581 | |
582 | /* If we have an INDIRECT_REF via a void pointer, we don't | |
583 | know anything about what that might alias. Likewise if the | |
584 | pointer is marked that way. */ | |
bd93e4f8 | 585 | if (ref_all_alias_ptr_type_p (t)) |
dd277d48 | 586 | return 0; |
587 | ||
588 | return -1; | |
589 | } | |
590 | ||
591 | /* Return the alias set for the memory pointed to by T, which may be | |
592 | either a type or an expression. */ | |
593 | ||
594 | alias_set_type | |
595 | get_deref_alias_set (tree t) | |
596 | { | |
bd93e4f8 | 597 | /* If we're not doing any alias analysis, just assume everything |
598 | aliases everything else. */ | |
599 | if (!flag_strict_aliasing) | |
600 | return 0; | |
601 | ||
dd277d48 | 602 | alias_set_type set = get_deref_alias_set_1 (t); |
603 | ||
604 | /* Fall back to the alias-set of the pointed-to type. */ | |
605 | if (set == -1) | |
606 | { | |
607 | if (! TYPE_P (t)) | |
608 | t = TREE_TYPE (t); | |
609 | set = get_alias_set (TREE_TYPE (t)); | |
610 | } | |
611 | ||
612 | return set; | |
613 | } | |
614 | ||
bd93e4f8 | 615 | /* Return the pointer-type relevant for TBAA purposes from the |
616 | memory reference tree *T or NULL_TREE in which case *T is | |
617 | adjusted to point to the outermost component reference that | |
618 | can be used for assigning an alias set. */ | |
619 | ||
620 | static tree | |
621 | reference_alias_ptr_type_1 (tree *t) | |
622 | { | |
623 | tree inner; | |
624 | ||
625 | /* Get the base object of the reference. */ | |
626 | inner = *t; | |
627 | while (handled_component_p (inner)) | |
628 | { | |
629 | /* If there is a VIEW_CONVERT_EXPR in the chain we cannot use | |
630 | the type of any component references that wrap it to | |
631 | determine the alias-set. */ | |
632 | if (TREE_CODE (inner) == VIEW_CONVERT_EXPR) | |
633 | *t = TREE_OPERAND (inner, 0); | |
634 | inner = TREE_OPERAND (inner, 0); | |
635 | } | |
636 | ||
637 | /* Handle pointer dereferences here, they can override the | |
638 | alias-set. */ | |
639 | if (INDIRECT_REF_P (inner) | |
640 | && ref_all_alias_ptr_type_p (TREE_TYPE (TREE_OPERAND (inner, 0)))) | |
641 | return TREE_TYPE (TREE_OPERAND (inner, 0)); | |
642 | else if (TREE_CODE (inner) == TARGET_MEM_REF) | |
643 | return TREE_TYPE (TMR_OFFSET (inner)); | |
644 | else if (TREE_CODE (inner) == MEM_REF | |
645 | && ref_all_alias_ptr_type_p (TREE_TYPE (TREE_OPERAND (inner, 1)))) | |
646 | return TREE_TYPE (TREE_OPERAND (inner, 1)); | |
647 | ||
648 | /* If the innermost reference is a MEM_REF that has a | |
649 | conversion embedded treat it like a VIEW_CONVERT_EXPR above, | |
650 | using the memory access type for determining the alias-set. */ | |
651 | if (TREE_CODE (inner) == MEM_REF | |
652 | && (TYPE_MAIN_VARIANT (TREE_TYPE (inner)) | |
653 | != TYPE_MAIN_VARIANT | |
654 | (TREE_TYPE (TREE_TYPE (TREE_OPERAND (inner, 1)))))) | |
655 | return TREE_TYPE (TREE_OPERAND (inner, 1)); | |
656 | ||
d400f5e1 | 657 | /* Otherwise, pick up the outermost object that we could have |
658 | a pointer to. */ | |
659 | tree tem = component_uses_parent_alias_set_from (*t); | |
660 | if (tem) | |
661 | *t = tem; | |
bd93e4f8 | 662 | |
663 | return NULL_TREE; | |
664 | } | |
665 | ||
666 | /* Return the pointer-type relevant for TBAA purposes from the | |
667 | gimple memory reference tree T. This is the type to be used for | |
668 | the offset operand of MEM_REF or TARGET_MEM_REF replacements of T | |
669 | and guarantees that get_alias_set will return the same alias | |
670 | set for T and the replacement. */ | |
671 | ||
672 | tree | |
673 | reference_alias_ptr_type (tree t) | |
674 | { | |
675 | tree ptype = reference_alias_ptr_type_1 (&t); | |
676 | /* If there is a given pointer type for aliasing purposes, return it. */ | |
677 | if (ptype != NULL_TREE) | |
678 | return ptype; | |
679 | ||
680 | /* Otherwise build one from the outermost component reference we | |
681 | may use. */ | |
682 | if (TREE_CODE (t) == MEM_REF | |
683 | || TREE_CODE (t) == TARGET_MEM_REF) | |
684 | return TREE_TYPE (TREE_OPERAND (t, 1)); | |
685 | else | |
686 | return build_pointer_type (TYPE_MAIN_VARIANT (TREE_TYPE (t))); | |
687 | } | |
688 | ||
689 | /* Return whether the pointer-types T1 and T2 used to determine | |
690 | two alias sets of two references will yield the same answer | |
691 | from get_deref_alias_set. */ | |
692 | ||
693 | bool | |
694 | alias_ptr_types_compatible_p (tree t1, tree t2) | |
695 | { | |
696 | if (TYPE_MAIN_VARIANT (t1) == TYPE_MAIN_VARIANT (t2)) | |
697 | return true; | |
698 | ||
699 | if (ref_all_alias_ptr_type_p (t1) | |
700 | || ref_all_alias_ptr_type_p (t2)) | |
701 | return false; | |
702 | ||
703 | return (TYPE_MAIN_VARIANT (TREE_TYPE (t1)) | |
704 | == TYPE_MAIN_VARIANT (TREE_TYPE (t2))); | |
705 | } | |
706 | ||
f7c44134 | 707 | /* Return the alias set for T, which may be either a type or an |
708 | expression. Call language-specific routine for help, if needed. */ | |
709 | ||
32c2fdea | 710 | alias_set_type |
aecda0d6 | 711 | get_alias_set (tree t) |
f7c44134 | 712 | { |
32c2fdea | 713 | alias_set_type set; |
f7c44134 | 714 | |
715 | /* If we're not doing any alias analysis, just assume everything | |
716 | aliases everything else. Also return 0 if this or its type is | |
717 | an error. */ | |
718 | if (! flag_strict_aliasing || t == error_mark_node | |
719 | || (! TYPE_P (t) | |
720 | && (TREE_TYPE (t) == 0 || TREE_TYPE (t) == error_mark_node))) | |
721 | return 0; | |
722 | ||
723 | /* We can be passed either an expression or a type. This and the | |
c3a9c149 | 724 | language-specific routine may make mutually-recursive calls to each other |
725 | to figure out what to do. At each juncture, we see if this is a tree | |
726 | that the language may need to handle specially. First handle things that | |
96216d37 | 727 | aren't types. */ |
1607663f | 728 | if (! TYPE_P (t)) |
f7c44134 | 729 | { |
182cf5a9 | 730 | /* Give the language a chance to do something with this tree |
731 | before we look at it. */ | |
2a631e19 | 732 | STRIP_NOPS (t); |
dc24ddbd | 733 | set = lang_hooks.get_alias_set (t); |
2a631e19 | 734 | if (set != -1) |
735 | return set; | |
736 | ||
bd93e4f8 | 737 | /* Get the alias pointer-type to use or the outermost object |
738 | that we could have a pointer to. */ | |
739 | tree ptype = reference_alias_ptr_type_1 (&t); | |
740 | if (ptype != NULL) | |
741 | return get_deref_alias_set (ptype); | |
1607663f | 742 | |
96216d37 | 743 | /* If we've already determined the alias set for a decl, just return |
744 | it. This is necessary for C++ anonymous unions, whose component | |
745 | variables don't look like union members (boo!). */ | |
afce1f57 | 746 | if (TREE_CODE (t) == VAR_DECL |
e16ceb8e | 747 | && DECL_RTL_SET_P (t) && MEM_P (DECL_RTL (t))) |
afce1f57 | 748 | return MEM_ALIAS_SET (DECL_RTL (t)); |
749 | ||
1607663f | 750 | /* Now all we care about is the type. */ |
751 | t = TREE_TYPE (t); | |
f7c44134 | 752 | } |
753 | ||
f7c44134 | 754 | /* Variant qualifiers don't affect the alias set, so get the main |
dfe8e806 | 755 | variant. */ |
f7c44134 | 756 | t = TYPE_MAIN_VARIANT (t); |
dfe8e806 | 757 | |
758 | /* Always use the canonical type as well. If this is a type that | |
759 | requires structural comparisons to identify compatible types | |
760 | use alias set zero. */ | |
761 | if (TYPE_STRUCTURAL_EQUALITY_P (t)) | |
e58c17e7 | 762 | { |
763 | /* Allow the language to specify another alias set for this | |
764 | type. */ | |
765 | set = lang_hooks.get_alias_set (t); | |
766 | if (set != -1) | |
767 | return set; | |
768 | return 0; | |
769 | } | |
ec7b61e0 | 770 | |
dfe8e806 | 771 | t = TYPE_CANONICAL (t); |
ec7b61e0 | 772 | |
45014c84 | 773 | /* The canonical type should not require structural equality checks. */ |
774 | gcc_checking_assert (!TYPE_STRUCTURAL_EQUALITY_P (t)); | |
dfe8e806 | 775 | |
776 | /* If this is a type with a known alias set, return it. */ | |
96216d37 | 777 | if (TYPE_ALIAS_SET_KNOWN_P (t)) |
f7c44134 | 778 | return TYPE_ALIAS_SET (t); |
779 | ||
b35a8f48 | 780 | /* We don't want to set TYPE_ALIAS_SET for incomplete types. */ |
781 | if (!COMPLETE_TYPE_P (t)) | |
782 | { | |
783 | /* For arrays with unknown size the conservative answer is the | |
784 | alias set of the element type. */ | |
785 | if (TREE_CODE (t) == ARRAY_TYPE) | |
786 | return get_alias_set (TREE_TYPE (t)); | |
787 | ||
788 | /* But return zero as a conservative answer for incomplete types. */ | |
789 | return 0; | |
790 | } | |
791 | ||
f7c44134 | 792 | /* See if the language has special handling for this type. */ |
dc24ddbd | 793 | set = lang_hooks.get_alias_set (t); |
2a631e19 | 794 | if (set != -1) |
96216d37 | 795 | return set; |
851dfbff | 796 | |
f7c44134 | 797 | /* There are no objects of FUNCTION_TYPE, so there's no point in |
798 | using up an alias set for them. (There are, of course, pointers | |
799 | and references to functions, but that's different.) */ | |
ec7b61e0 | 800 | else if (TREE_CODE (t) == FUNCTION_TYPE || TREE_CODE (t) == METHOD_TYPE) |
f7c44134 | 801 | set = 0; |
9fcc3e54 | 802 | |
803 | /* Unless the language specifies otherwise, let vector types alias | |
804 | their components. This avoids some nasty type punning issues in | |
805 | normal usage. And indeed lets vectors be treated more like an | |
806 | array slice. */ | |
807 | else if (TREE_CODE (t) == VECTOR_TYPE) | |
808 | set = get_alias_set (TREE_TYPE (t)); | |
809 | ||
62d823d0 | 810 | /* Unless the language specifies otherwise, treat array types the |
811 | same as their components. This avoids the asymmetry we get | |
812 | through recording the components. Consider accessing a | |
813 | character(kind=1) through a reference to a character(kind=1)[1:1]. | |
814 | Or consider if we want to assign integer(kind=4)[0:D.1387] and | |
815 | integer(kind=4)[4] the same alias set or not. | |
816 | Just be pragmatic here and make sure the array and its element | |
817 | type get the same alias set assigned. */ | |
ec7b61e0 | 818 | else if (TREE_CODE (t) == ARRAY_TYPE && !TYPE_NONALIASED_COMPONENT (t)) |
62d823d0 | 819 | set = get_alias_set (TREE_TYPE (t)); |
820 | ||
2af087f2 | 821 | /* From the former common C and C++ langhook implementation: |
822 | ||
823 | Unfortunately, there is no canonical form of a pointer type. | |
824 | In particular, if we have `typedef int I', then `int *', and | |
825 | `I *' are different types. So, we have to pick a canonical | |
826 | representative. We do this below. | |
827 | ||
828 | Technically, this approach is actually more conservative that | |
829 | it needs to be. In particular, `const int *' and `int *' | |
830 | should be in different alias sets, according to the C and C++ | |
831 | standard, since their types are not the same, and so, | |
832 | technically, an `int **' and `const int **' cannot point at | |
833 | the same thing. | |
834 | ||
835 | But, the standard is wrong. In particular, this code is | |
836 | legal C++: | |
837 | ||
838 | int *ip; | |
839 | int **ipp = &ip; | |
840 | const int* const* cipp = ipp; | |
841 | And, it doesn't make sense for that to be legal unless you | |
842 | can dereference IPP and CIPP. So, we ignore cv-qualifiers on | |
843 | the pointed-to types. This issue has been reported to the | |
844 | C++ committee. | |
845 | ||
846 | In addition to the above canonicalization issue, with LTO | |
847 | we should also canonicalize `T (*)[]' to `T *' avoiding | |
848 | alias issues with pointer-to element types and pointer-to | |
849 | array types. | |
850 | ||
851 | Likewise we need to deal with the situation of incomplete | |
852 | pointed-to types and make `*(struct X **)&a' and | |
853 | `*(struct X {} **)&a' alias. Otherwise we will have to | |
854 | guarantee that all pointer-to incomplete type variants | |
855 | will be replaced by pointer-to complete type variants if | |
856 | they are available. | |
857 | ||
858 | With LTO the convenient situation of using `void *' to | |
859 | access and store any pointer type will also become | |
860 | more apparent (and `void *' is just another pointer-to | |
861 | incomplete type). Assigning alias-set zero to `void *' | |
862 | and all pointer-to incomplete types is a not appealing | |
863 | solution. Assigning an effective alias-set zero only | |
864 | affecting pointers might be - by recording proper subset | |
865 | relationships of all pointer alias-sets. | |
866 | ||
867 | Pointer-to function types are another grey area which | |
868 | needs caution. Globbing them all into one alias-set | |
869 | or the above effective zero set would work. | |
870 | ||
871 | For now just assign the same alias-set to all pointers. | |
872 | That's simple and avoids all the above problems. */ | |
873 | else if (POINTER_TYPE_P (t) | |
874 | && t != ptr_type_node) | |
45014c84 | 875 | set = get_alias_set (ptr_type_node); |
2af087f2 | 876 | |
ec7b61e0 | 877 | /* Otherwise make a new alias set for this type. */ |
f7c44134 | 878 | else |
45014c84 | 879 | { |
880 | /* Each canonical type gets its own alias set, so canonical types | |
881 | shouldn't form a tree. It doesn't really matter for types | |
882 | we handle specially above, so only check it where it possibly | |
883 | would result in a bogus alias set. */ | |
884 | gcc_checking_assert (TYPE_CANONICAL (t) == t); | |
885 | ||
886 | set = new_alias_set (); | |
887 | } | |
f7c44134 | 888 | |
889 | TYPE_ALIAS_SET (t) = set; | |
851dfbff | 890 | |
ec7b61e0 | 891 | /* If this is an aggregate type or a complex type, we must record any |
892 | component aliasing information. */ | |
680210f2 | 893 | if (AGGREGATE_TYPE_P (t) || TREE_CODE (t) == COMPLEX_TYPE) |
851dfbff | 894 | record_component_aliases (t); |
895 | ||
f7c44134 | 896 | return set; |
897 | } | |
898 | ||
899 | /* Return a brand-new alias set. */ | |
900 | ||
32c2fdea | 901 | alias_set_type |
aecda0d6 | 902 | new_alias_set (void) |
f7c44134 | 903 | { |
f7c44134 | 904 | if (flag_strict_aliasing) |
56bbdce4 | 905 | { |
8b435339 | 906 | if (alias_sets == 0) |
f1f41a6c | 907 | vec_safe_push (alias_sets, (alias_set_entry) 0); |
908 | vec_safe_push (alias_sets, (alias_set_entry) 0); | |
909 | return alias_sets->length () - 1; | |
56bbdce4 | 910 | } |
f7c44134 | 911 | else |
912 | return 0; | |
913 | } | |
a5b1863e | 914 | |
8e2ef041 | 915 | /* Indicate that things in SUBSET can alias things in SUPERSET, but that |
916 | not everything that aliases SUPERSET also aliases SUBSET. For example, | |
917 | in C, a store to an `int' can alias a load of a structure containing an | |
918 | `int', and vice versa. But it can't alias a load of a 'double' member | |
919 | of the same structure. Here, the structure would be the SUPERSET and | |
920 | `int' the SUBSET. This relationship is also described in the comment at | |
921 | the beginning of this file. | |
922 | ||
923 | This function should be called only once per SUPERSET/SUBSET pair. | |
a5b1863e | 924 | |
925 | It is illegal for SUPERSET to be zero; everything is implicitly a | |
926 | subset of alias set zero. */ | |
927 | ||
892fdb03 | 928 | void |
32c2fdea | 929 | record_alias_subset (alias_set_type superset, alias_set_type subset) |
a5b1863e | 930 | { |
931 | alias_set_entry superset_entry; | |
932 | alias_set_entry subset_entry; | |
933 | ||
c3a9c149 | 934 | /* It is possible in complex type situations for both sets to be the same, |
935 | in which case we can ignore this operation. */ | |
936 | if (superset == subset) | |
937 | return; | |
938 | ||
64db345d | 939 | gcc_assert (superset); |
a5b1863e | 940 | |
941 | superset_entry = get_alias_set_entry (superset); | |
9342ee68 | 942 | if (superset_entry == 0) |
a5b1863e | 943 | { |
944 | /* Create an entry for the SUPERSET, so that we have a place to | |
945 | attach the SUBSET. */ | |
ba72912a | 946 | superset_entry = ggc_alloc_cleared_alias_set_entry_d (); |
a5b1863e | 947 | superset_entry->alias_set = superset; |
9342ee68 | 948 | superset_entry->children |
ba72912a | 949 | = splay_tree_new_ggc (splay_tree_compare_ints, |
950 | ggc_alloc_splay_tree_scalar_scalar_splay_tree_s, | |
951 | ggc_alloc_splay_tree_scalar_scalar_splay_tree_node_s); | |
f2e5de80 | 952 | superset_entry->has_zero_child = 0; |
f1f41a6c | 953 | (*alias_sets)[superset] = superset_entry; |
a5b1863e | 954 | } |
955 | ||
851dfbff | 956 | if (subset == 0) |
957 | superset_entry->has_zero_child = 1; | |
958 | else | |
959 | { | |
960 | subset_entry = get_alias_set_entry (subset); | |
961 | /* If there is an entry for the subset, enter all of its children | |
962 | (if they are not already present) as children of the SUPERSET. */ | |
9342ee68 | 963 | if (subset_entry) |
851dfbff | 964 | { |
965 | if (subset_entry->has_zero_child) | |
966 | superset_entry->has_zero_child = 1; | |
083a2b5e | 967 | |
851dfbff | 968 | splay_tree_foreach (subset_entry->children, insert_subset_children, |
969 | superset_entry->children); | |
970 | } | |
a5b1863e | 971 | |
851dfbff | 972 | /* Enter the SUBSET itself as a child of the SUPERSET. */ |
9342ee68 | 973 | splay_tree_insert (superset_entry->children, |
851dfbff | 974 | (splay_tree_key) subset, 0); |
975 | } | |
a5b1863e | 976 | } |
977 | ||
9501a338 | 978 | /* Record that component types of TYPE, if any, are part of that type for |
979 | aliasing purposes. For record types, we only record component types | |
c67bf066 | 980 | for fields that are not marked non-addressable. For array types, we |
981 | only record the component type if it is not marked non-aliased. */ | |
9501a338 | 982 | |
983 | void | |
aecda0d6 | 984 | record_component_aliases (tree type) |
9501a338 | 985 | { |
32c2fdea | 986 | alias_set_type superset = get_alias_set (type); |
9501a338 | 987 | tree field; |
988 | ||
989 | if (superset == 0) | |
990 | return; | |
991 | ||
992 | switch (TREE_CODE (type)) | |
993 | { | |
9501a338 | 994 | case RECORD_TYPE: |
995 | case UNION_TYPE: | |
996 | case QUAL_UNION_TYPE: | |
1767a056 | 997 | for (field = TYPE_FIELDS (type); field != 0; field = DECL_CHAIN (field)) |
c67bf066 | 998 | if (TREE_CODE (field) == FIELD_DECL && !DECL_NONADDRESSABLE_P (field)) |
851dfbff | 999 | record_alias_subset (superset, get_alias_set (TREE_TYPE (field))); |
9501a338 | 1000 | break; |
1001 | ||
680210f2 | 1002 | case COMPLEX_TYPE: |
1003 | record_alias_subset (superset, get_alias_set (TREE_TYPE (type))); | |
1004 | break; | |
1005 | ||
62d823d0 | 1006 | /* VECTOR_TYPE and ARRAY_TYPE share the alias set with their |
1007 | element type. */ | |
1008 | ||
9501a338 | 1009 | default: |
1010 | break; | |
1011 | } | |
1012 | } | |
1013 | ||
f7c44134 | 1014 | /* Allocate an alias set for use in storing and reading from the varargs |
1015 | spill area. */ | |
1016 | ||
32c2fdea | 1017 | static GTY(()) alias_set_type varargs_set = -1; |
4ec9406e | 1018 | |
32c2fdea | 1019 | alias_set_type |
aecda0d6 | 1020 | get_varargs_alias_set (void) |
f7c44134 | 1021 | { |
fcdd3ab3 | 1022 | #if 1 |
1023 | /* We now lower VA_ARG_EXPR, and there's currently no way to attach the | |
1024 | varargs alias set to an INDIRECT_REF (FIXME!), so we can't | |
1025 | consistently use the varargs alias set for loads from the varargs | |
1026 | area. So don't use it anywhere. */ | |
1027 | return 0; | |
1028 | #else | |
4ec9406e | 1029 | if (varargs_set == -1) |
1030 | varargs_set = new_alias_set (); | |
f7c44134 | 1031 | |
4ec9406e | 1032 | return varargs_set; |
fcdd3ab3 | 1033 | #endif |
f7c44134 | 1034 | } |
1035 | ||
1036 | /* Likewise, but used for the fixed portions of the frame, e.g., register | |
1037 | save areas. */ | |
1038 | ||
32c2fdea | 1039 | static GTY(()) alias_set_type frame_set = -1; |
4ec9406e | 1040 | |
32c2fdea | 1041 | alias_set_type |
aecda0d6 | 1042 | get_frame_alias_set (void) |
f7c44134 | 1043 | { |
4ec9406e | 1044 | if (frame_set == -1) |
1045 | frame_set = new_alias_set (); | |
f7c44134 | 1046 | |
4ec9406e | 1047 | return frame_set; |
f7c44134 | 1048 | } |
1049 | ||
4e56ceb1 | 1050 | /* Create a new, unique base with id ID. */ |
1051 | ||
1052 | static rtx | |
1053 | unique_base_value (HOST_WIDE_INT id) | |
1054 | { | |
1055 | return gen_rtx_ADDRESS (Pmode, id); | |
1056 | } | |
1057 | ||
1058 | /* Return true if accesses based on any other base value cannot alias | |
1059 | those based on X. */ | |
1060 | ||
1061 | static bool | |
1062 | unique_base_value_p (rtx x) | |
1063 | { | |
1064 | return GET_CODE (x) == ADDRESS && GET_MODE (x) == Pmode; | |
1065 | } | |
1066 | ||
1067 | /* Return true if X is known to be a base value. */ | |
1068 | ||
1069 | static bool | |
1070 | known_base_value_p (rtx x) | |
1071 | { | |
1072 | switch (GET_CODE (x)) | |
1073 | { | |
1074 | case LABEL_REF: | |
1075 | case SYMBOL_REF: | |
1076 | return true; | |
1077 | ||
1078 | case ADDRESS: | |
1079 | /* Arguments may or may not be bases; we don't know for sure. */ | |
1080 | return GET_MODE (x) != VOIDmode; | |
1081 | ||
1082 | default: | |
1083 | return false; | |
1084 | } | |
1085 | } | |
1086 | ||
e7c893ba | 1087 | /* Inside SRC, the source of a SET, find a base address. */ |
1088 | ||
ea0cb7ae | 1089 | static rtx |
aecda0d6 | 1090 | find_base_value (rtx src) |
ea0cb7ae | 1091 | { |
970c0c06 | 1092 | unsigned int regno; |
d01410d3 | 1093 | |
2e984a7a | 1094 | #if defined (FIND_BASE_TERM) |
1095 | /* Try machine-dependent ways to find the base term. */ | |
1096 | src = FIND_BASE_TERM (src); | |
1097 | #endif | |
1098 | ||
ea0cb7ae | 1099 | switch (GET_CODE (src)) |
1100 | { | |
1101 | case SYMBOL_REF: | |
1102 | case LABEL_REF: | |
1103 | return src; | |
1104 | ||
1105 | case REG: | |
9db10305 | 1106 | regno = REGNO (src); |
083a2b5e | 1107 | /* At the start of a function, argument registers have known base |
e7c893ba | 1108 | values which may be lost later. Returning an ADDRESS |
1109 | expression here allows optimization based on argument values | |
1110 | even when the argument registers are used for other purposes. */ | |
970c0c06 | 1111 | if (regno < FIRST_PSEUDO_REGISTER && copying_arguments) |
1112 | return new_reg_base_value[regno]; | |
1a538f67 | 1113 | |
ebd13eec | 1114 | /* If a pseudo has a known base value, return it. Do not do this |
9464d436 | 1115 | for non-fixed hard regs since it can result in a circular |
1116 | dependency chain for registers which have values at function entry. | |
ebd13eec | 1117 | |
1118 | The test above is not sufficient because the scheduler may move | |
1119 | a copy out of an arg reg past the NOTE_INSN_FUNCTION_BEGIN. */ | |
9464d436 | 1120 | if ((regno >= FIRST_PSEUDO_REGISTER || fixed_regs[regno]) |
f1f41a6c | 1121 | && regno < vec_safe_length (reg_base_value)) |
e5be0cc6 | 1122 | { |
1123 | /* If we're inside init_alias_analysis, use new_reg_base_value | |
1124 | to reduce the number of relaxation iterations. */ | |
a303ed8c | 1125 | if (new_reg_base_value && new_reg_base_value[regno] |
3072d30e | 1126 | && DF_REG_DEF_COUNT (regno) == 1) |
e5be0cc6 | 1127 | return new_reg_base_value[regno]; |
1128 | ||
f1f41a6c | 1129 | if ((*reg_base_value)[regno]) |
1130 | return (*reg_base_value)[regno]; | |
e5be0cc6 | 1131 | } |
1a538f67 | 1132 | |
1c5cb56c | 1133 | return 0; |
ea0cb7ae | 1134 | |
1135 | case MEM: | |
1136 | /* Check for an argument passed in memory. Only record in the | |
1137 | copying-arguments block; it is too hard to track changes | |
1138 | otherwise. */ | |
1139 | if (copying_arguments | |
1140 | && (XEXP (src, 0) == arg_pointer_rtx | |
1141 | || (GET_CODE (XEXP (src, 0)) == PLUS | |
1142 | && XEXP (XEXP (src, 0), 0) == arg_pointer_rtx))) | |
4e56ceb1 | 1143 | return arg_base_value; |
ea0cb7ae | 1144 | return 0; |
1145 | ||
1146 | case CONST: | |
1147 | src = XEXP (src, 0); | |
1148 | if (GET_CODE (src) != PLUS && GET_CODE (src) != MINUS) | |
1149 | break; | |
083a2b5e | 1150 | |
2c0e001b | 1151 | /* ... fall through ... */ |
e7c893ba | 1152 | |
ea0cb7ae | 1153 | case PLUS: |
1154 | case MINUS: | |
e7c893ba | 1155 | { |
06c5c535 | 1156 | rtx temp, src_0 = XEXP (src, 0), src_1 = XEXP (src, 1); |
1157 | ||
560de2f8 | 1158 | /* If either operand is a REG that is a known pointer, then it |
1159 | is the base. */ | |
1160 | if (REG_P (src_0) && REG_POINTER (src_0)) | |
1161 | return find_base_value (src_0); | |
1162 | if (REG_P (src_1) && REG_POINTER (src_1)) | |
1163 | return find_base_value (src_1); | |
1164 | ||
06c5c535 | 1165 | /* If either operand is a REG, then see if we already have |
1166 | a known value for it. */ | |
560de2f8 | 1167 | if (REG_P (src_0)) |
06c5c535 | 1168 | { |
1169 | temp = find_base_value (src_0); | |
083a2b5e | 1170 | if (temp != 0) |
06c5c535 | 1171 | src_0 = temp; |
1172 | } | |
1173 | ||
560de2f8 | 1174 | if (REG_P (src_1)) |
06c5c535 | 1175 | { |
1176 | temp = find_base_value (src_1); | |
083a2b5e | 1177 | if (temp!= 0) |
06c5c535 | 1178 | src_1 = temp; |
1179 | } | |
e7c893ba | 1180 | |
560de2f8 | 1181 | /* If either base is named object or a special address |
1182 | (like an argument or stack reference), then use it for the | |
1183 | base term. */ | |
4e56ceb1 | 1184 | if (src_0 != 0 && known_base_value_p (src_0)) |
560de2f8 | 1185 | return src_0; |
1186 | ||
4e56ceb1 | 1187 | if (src_1 != 0 && known_base_value_p (src_1)) |
560de2f8 | 1188 | return src_1; |
1189 | ||
083a2b5e | 1190 | /* Guess which operand is the base address: |
06c5c535 | 1191 | If either operand is a symbol, then it is the base. If |
1192 | either operand is a CONST_INT, then the other is the base. */ | |
971ba038 | 1193 | if (CONST_INT_P (src_1) || CONSTANT_P (src_0)) |
e7c893ba | 1194 | return find_base_value (src_0); |
971ba038 | 1195 | else if (CONST_INT_P (src_0) || CONSTANT_P (src_1)) |
06c5c535 | 1196 | return find_base_value (src_1); |
1197 | ||
ea0cb7ae | 1198 | return 0; |
e7c893ba | 1199 | } |
1200 | ||
1201 | case LO_SUM: | |
1202 | /* The standard form is (lo_sum reg sym) so look only at the | |
1203 | second operand. */ | |
1204 | return find_base_value (XEXP (src, 1)); | |
ea0cb7ae | 1205 | |
1206 | case AND: | |
1207 | /* If the second operand is constant set the base | |
2c0e001b | 1208 | address to the first operand. */ |
971ba038 | 1209 | if (CONST_INT_P (XEXP (src, 1)) && INTVAL (XEXP (src, 1)) != 0) |
e7c893ba | 1210 | return find_base_value (XEXP (src, 0)); |
ea0cb7ae | 1211 | return 0; |
1212 | ||
43d45282 | 1213 | case TRUNCATE: |
04ec15fa | 1214 | /* As we do not know which address space the pointer is referring to, we can |
98155838 | 1215 | handle this only if the target does not support different pointer or |
1216 | address modes depending on the address space. */ | |
1217 | if (!target_default_pointer_address_modes_p ()) | |
1218 | break; | |
43d45282 | 1219 | if (GET_MODE_SIZE (GET_MODE (src)) < GET_MODE_SIZE (Pmode)) |
1220 | break; | |
1221 | /* Fall through. */ | |
ea0cb7ae | 1222 | case HIGH: |
2be64ef0 | 1223 | case PRE_INC: |
1224 | case PRE_DEC: | |
1225 | case POST_INC: | |
1226 | case POST_DEC: | |
1227 | case PRE_MODIFY: | |
1228 | case POST_MODIFY: | |
e7c893ba | 1229 | return find_base_value (XEXP (src, 0)); |
99c14947 | 1230 | |
d01410d3 | 1231 | case ZERO_EXTEND: |
1232 | case SIGN_EXTEND: /* used for NT/Alpha pointers */ | |
04ec15fa | 1233 | /* As we do not know which address space the pointer is referring to, we can |
98155838 | 1234 | handle this only if the target does not support different pointer or |
1235 | address modes depending on the address space. */ | |
1236 | if (!target_default_pointer_address_modes_p ()) | |
1237 | break; | |
1238 | ||
d01410d3 | 1239 | { |
1240 | rtx temp = find_base_value (XEXP (src, 0)); | |
1241 | ||
85d654dd | 1242 | if (temp != 0 && CONSTANT_P (temp)) |
d01410d3 | 1243 | temp = convert_memory_address (Pmode, temp); |
d01410d3 | 1244 | |
1245 | return temp; | |
1246 | } | |
1247 | ||
99c14947 | 1248 | default: |
1249 | break; | |
ea0cb7ae | 1250 | } |
1251 | ||
1252 | return 0; | |
1253 | } | |
1254 | ||
4e56ceb1 | 1255 | /* Called from init_alias_analysis indirectly through note_stores, |
1256 | or directly if DEST is a register with a REG_NOALIAS note attached. | |
1257 | SET is null in the latter case. */ | |
ea0cb7ae | 1258 | |
083a2b5e | 1259 | /* While scanning insns to find base values, reg_seen[N] is nonzero if |
ea0cb7ae | 1260 | register N has been set in this function. */ |
7bf60644 | 1261 | static sbitmap reg_seen; |
ea0cb7ae | 1262 | |
e7c893ba | 1263 | static void |
81a410b1 | 1264 | record_set (rtx dest, const_rtx set, void *data ATTRIBUTE_UNUSED) |
ea0cb7ae | 1265 | { |
19cb6b50 | 1266 | unsigned regno; |
ea0cb7ae | 1267 | rtx src; |
3721f260 | 1268 | int n; |
ea0cb7ae | 1269 | |
8ad4c111 | 1270 | if (!REG_P (dest)) |
ea0cb7ae | 1271 | return; |
1272 | ||
9db10305 | 1273 | regno = REGNO (dest); |
ea0cb7ae | 1274 | |
f1f41a6c | 1275 | gcc_checking_assert (regno < reg_base_value->length ()); |
c2e94b25 | 1276 | |
3721f260 | 1277 | /* If this spans multiple hard registers, then we must indicate that every |
1278 | register has an unusable value. */ | |
1279 | if (regno < FIRST_PSEUDO_REGISTER) | |
67d6c12b | 1280 | n = hard_regno_nregs[regno][GET_MODE (dest)]; |
3721f260 | 1281 | else |
1282 | n = 1; | |
1283 | if (n != 1) | |
1284 | { | |
1285 | while (--n >= 0) | |
1286 | { | |
08b7917c | 1287 | bitmap_set_bit (reg_seen, regno + n); |
3721f260 | 1288 | new_reg_base_value[regno + n] = 0; |
1289 | } | |
1290 | return; | |
1291 | } | |
1292 | ||
ea0cb7ae | 1293 | if (set) |
1294 | { | |
1295 | /* A CLOBBER wipes out any old value but does not prevent a previously | |
1296 | unset register from acquiring a base address (i.e. reg_seen is not | |
1297 | set). */ | |
1298 | if (GET_CODE (set) == CLOBBER) | |
1299 | { | |
06c5c535 | 1300 | new_reg_base_value[regno] = 0; |
ea0cb7ae | 1301 | return; |
1302 | } | |
1303 | src = SET_SRC (set); | |
1304 | } | |
1305 | else | |
1306 | { | |
4e56ceb1 | 1307 | /* There's a REG_NOALIAS note against DEST. */ |
08b7917c | 1308 | if (bitmap_bit_p (reg_seen, regno)) |
ea0cb7ae | 1309 | { |
06c5c535 | 1310 | new_reg_base_value[regno] = 0; |
ea0cb7ae | 1311 | return; |
1312 | } | |
08b7917c | 1313 | bitmap_set_bit (reg_seen, regno); |
4e56ceb1 | 1314 | new_reg_base_value[regno] = unique_base_value (unique_id++); |
ea0cb7ae | 1315 | return; |
1316 | } | |
1317 | ||
ad6b1fe1 | 1318 | /* If this is not the first set of REGNO, see whether the new value |
1319 | is related to the old one. There are two cases of interest: | |
1320 | ||
1321 | (1) The register might be assigned an entirely new value | |
1322 | that has the same base term as the original set. | |
1323 | ||
1324 | (2) The set might be a simple self-modification that | |
1325 | cannot change REGNO's base value. | |
1326 | ||
1327 | If neither case holds, reject the original base value as invalid. | |
1328 | Note that the following situation is not detected: | |
1329 | ||
a0c938f0 | 1330 | extern int x, y; int *p = &x; p += (&y-&x); |
ad6b1fe1 | 1331 | |
ea0cb7ae | 1332 | ANSI C does not allow computing the difference of addresses |
1333 | of distinct top level objects. */ | |
ad6b1fe1 | 1334 | if (new_reg_base_value[regno] != 0 |
1335 | && find_base_value (src) != new_reg_base_value[regno]) | |
ea0cb7ae | 1336 | switch (GET_CODE (src)) |
1337 | { | |
e7c893ba | 1338 | case LO_SUM: |
ea0cb7ae | 1339 | case MINUS: |
1340 | if (XEXP (src, 0) != dest && XEXP (src, 1) != dest) | |
06c5c535 | 1341 | new_reg_base_value[regno] = 0; |
ea0cb7ae | 1342 | break; |
43d45282 | 1343 | case PLUS: |
1344 | /* If the value we add in the PLUS is also a valid base value, | |
1345 | this might be the actual base value, and the original value | |
1346 | an index. */ | |
1347 | { | |
1348 | rtx other = NULL_RTX; | |
1349 | ||
1350 | if (XEXP (src, 0) == dest) | |
1351 | other = XEXP (src, 1); | |
1352 | else if (XEXP (src, 1) == dest) | |
1353 | other = XEXP (src, 0); | |
1354 | ||
1355 | if (! other || find_base_value (other)) | |
1356 | new_reg_base_value[regno] = 0; | |
1357 | break; | |
1358 | } | |
ea0cb7ae | 1359 | case AND: |
971ba038 | 1360 | if (XEXP (src, 0) != dest || !CONST_INT_P (XEXP (src, 1))) |
06c5c535 | 1361 | new_reg_base_value[regno] = 0; |
ea0cb7ae | 1362 | break; |
ea0cb7ae | 1363 | default: |
06c5c535 | 1364 | new_reg_base_value[regno] = 0; |
ea0cb7ae | 1365 | break; |
1366 | } | |
1367 | /* If this is the first set of a register, record the value. */ | |
1368 | else if ((regno >= FIRST_PSEUDO_REGISTER || ! fixed_regs[regno]) | |
08b7917c | 1369 | && ! bitmap_bit_p (reg_seen, regno) && new_reg_base_value[regno] == 0) |
06c5c535 | 1370 | new_reg_base_value[regno] = find_base_value (src); |
ea0cb7ae | 1371 | |
08b7917c | 1372 | bitmap_set_bit (reg_seen, regno); |
ea0cb7ae | 1373 | } |
1374 | ||
0cfef6c5 | 1375 | /* Return REG_BASE_VALUE for REGNO. Selective scheduler uses this to avoid |
1376 | using hard registers with non-null REG_BASE_VALUE for renaming. */ | |
1377 | rtx | |
1378 | get_reg_base_value (unsigned int regno) | |
1379 | { | |
f1f41a6c | 1380 | return (*reg_base_value)[regno]; |
0cfef6c5 | 1381 | } |
1382 | ||
12cb06ad | 1383 | /* If a value is known for REGNO, return it. */ |
1384 | ||
a0c938f0 | 1385 | rtx |
12cb06ad | 1386 | get_reg_known_value (unsigned int regno) |
1387 | { | |
1388 | if (regno >= FIRST_PSEUDO_REGISTER) | |
1389 | { | |
1390 | regno -= FIRST_PSEUDO_REGISTER; | |
f1f41a6c | 1391 | if (regno < vec_safe_length (reg_known_value)) |
1392 | return (*reg_known_value)[regno]; | |
12cb06ad | 1393 | } |
1394 | return NULL; | |
9f5e1a5a | 1395 | } |
1396 | ||
12cb06ad | 1397 | /* Set it. */ |
1398 | ||
1399 | static void | |
1400 | set_reg_known_value (unsigned int regno, rtx val) | |
1401 | { | |
1402 | if (regno >= FIRST_PSEUDO_REGISTER) | |
1403 | { | |
1404 | regno -= FIRST_PSEUDO_REGISTER; | |
f1f41a6c | 1405 | if (regno < vec_safe_length (reg_known_value)) |
1406 | (*reg_known_value)[regno] = val; | |
12cb06ad | 1407 | } |
1408 | } | |
1409 | ||
1410 | /* Similarly for reg_known_equiv_p. */ | |
1411 | ||
1412 | bool | |
1413 | get_reg_known_equiv_p (unsigned int regno) | |
1414 | { | |
1415 | if (regno >= FIRST_PSEUDO_REGISTER) | |
1416 | { | |
1417 | regno -= FIRST_PSEUDO_REGISTER; | |
f1f41a6c | 1418 | if (regno < vec_safe_length (reg_known_value)) |
08b7917c | 1419 | return bitmap_bit_p (reg_known_equiv_p, regno); |
12cb06ad | 1420 | } |
1421 | return false; | |
1422 | } | |
1423 | ||
1424 | static void | |
1425 | set_reg_known_equiv_p (unsigned int regno, bool val) | |
1426 | { | |
1427 | if (regno >= FIRST_PSEUDO_REGISTER) | |
1428 | { | |
1429 | regno -= FIRST_PSEUDO_REGISTER; | |
f1f41a6c | 1430 | if (regno < vec_safe_length (reg_known_value)) |
d6443ebe | 1431 | { |
1432 | if (val) | |
08b7917c | 1433 | bitmap_set_bit (reg_known_equiv_p, regno); |
d6443ebe | 1434 | else |
08b7917c | 1435 | bitmap_clear_bit (reg_known_equiv_p, regno); |
d6443ebe | 1436 | } |
12cb06ad | 1437 | } |
1438 | } | |
1439 | ||
1440 | ||
7cfb9bcf | 1441 | /* Returns a canonical version of X, from the point of view alias |
1442 | analysis. (For example, if X is a MEM whose address is a register, | |
1443 | and the register has a known value (say a SYMBOL_REF), then a MEM | |
1444 | whose address is the SYMBOL_REF is returned.) */ | |
1445 | ||
1446 | rtx | |
aecda0d6 | 1447 | canon_rtx (rtx x) |
ea0cb7ae | 1448 | { |
1449 | /* Recursively look for equivalences. */ | |
8ad4c111 | 1450 | if (REG_P (x) && REGNO (x) >= FIRST_PSEUDO_REGISTER) |
12cb06ad | 1451 | { |
1452 | rtx t = get_reg_known_value (REGNO (x)); | |
1453 | if (t == x) | |
1454 | return x; | |
1455 | if (t) | |
1456 | return canon_rtx (t); | |
1457 | } | |
1458 | ||
1459 | if (GET_CODE (x) == PLUS) | |
ea0cb7ae | 1460 | { |
1461 | rtx x0 = canon_rtx (XEXP (x, 0)); | |
1462 | rtx x1 = canon_rtx (XEXP (x, 1)); | |
1463 | ||
1464 | if (x0 != XEXP (x, 0) || x1 != XEXP (x, 1)) | |
1465 | { | |
971ba038 | 1466 | if (CONST_INT_P (x0)) |
29c05e22 | 1467 | return plus_constant (GET_MODE (x), x1, INTVAL (x0)); |
971ba038 | 1468 | else if (CONST_INT_P (x1)) |
29c05e22 | 1469 | return plus_constant (GET_MODE (x), x0, INTVAL (x1)); |
941522d6 | 1470 | return gen_rtx_PLUS (GET_MODE (x), x0, x1); |
ea0cb7ae | 1471 | } |
1472 | } | |
083a2b5e | 1473 | |
ea0cb7ae | 1474 | /* This gives us much better alias analysis when called from |
1475 | the loop optimizer. Note we want to leave the original | |
1476 | MEM alone, but need to return the canonicalized MEM with | |
1477 | all the flags with their original values. */ | |
e16ceb8e | 1478 | else if (MEM_P (x)) |
e4e86ec5 | 1479 | x = replace_equiv_address_nv (x, canon_rtx (XEXP (x, 0))); |
083a2b5e | 1480 | |
ea0cb7ae | 1481 | return x; |
1482 | } | |
1483 | ||
1484 | /* Return 1 if X and Y are identical-looking rtx's. | |
6285298a | 1485 | Expect that X and Y has been already canonicalized. |
ea0cb7ae | 1486 | |
1487 | We use the data in reg_known_value above to see if two registers with | |
1488 | different numbers are, in fact, equivalent. */ | |
1489 | ||
1490 | static int | |
7ecb5bb2 | 1491 | rtx_equal_for_memref_p (const_rtx x, const_rtx y) |
ea0cb7ae | 1492 | { |
19cb6b50 | 1493 | int i; |
1494 | int j; | |
1495 | enum rtx_code code; | |
1496 | const char *fmt; | |
ea0cb7ae | 1497 | |
1498 | if (x == 0 && y == 0) | |
1499 | return 1; | |
1500 | if (x == 0 || y == 0) | |
1501 | return 0; | |
083a2b5e | 1502 | |
ea0cb7ae | 1503 | if (x == y) |
1504 | return 1; | |
1505 | ||
1506 | code = GET_CODE (x); | |
1507 | /* Rtx's of different codes cannot be equal. */ | |
1508 | if (code != GET_CODE (y)) | |
1509 | return 0; | |
1510 | ||
1511 | /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. | |
1512 | (REG:SI x) and (REG:HI x) are NOT equivalent. */ | |
1513 | ||
1514 | if (GET_MODE (x) != GET_MODE (y)) | |
1515 | return 0; | |
1516 | ||
7cfb9bcf | 1517 | /* Some RTL can be compared without a recursive examination. */ |
1518 | switch (code) | |
1519 | { | |
1520 | case REG: | |
1521 | return REGNO (x) == REGNO (y); | |
1522 | ||
1523 | case LABEL_REF: | |
1524 | return XEXP (x, 0) == XEXP (y, 0); | |
9342ee68 | 1525 | |
7cfb9bcf | 1526 | case SYMBOL_REF: |
1527 | return XSTR (x, 0) == XSTR (y, 0); | |
1528 | ||
861a6de4 | 1529 | case ENTRY_VALUE: |
1530 | /* This is magic, don't go through canonicalization et al. */ | |
1531 | return rtx_equal_p (ENTRY_VALUE_EXP (x), ENTRY_VALUE_EXP (y)); | |
1532 | ||
a2e238f7 | 1533 | case VALUE: |
0349edce | 1534 | CASE_CONST_UNIQUE: |
7cfb9bcf | 1535 | /* There's no need to compare the contents of CONST_DOUBLEs or |
1536 | CONST_INTs because pointer equality is a good enough | |
1537 | comparison for these nodes. */ | |
1538 | return 0; | |
1539 | ||
7cfb9bcf | 1540 | default: |
1541 | break; | |
1542 | } | |
ea0cb7ae | 1543 | |
6285298a | 1544 | /* canon_rtx knows how to handle plus. No need to canonicalize. */ |
1545 | if (code == PLUS) | |
ea0cb7ae | 1546 | return ((rtx_equal_for_memref_p (XEXP (x, 0), XEXP (y, 0)) |
1547 | && rtx_equal_for_memref_p (XEXP (x, 1), XEXP (y, 1))) | |
1548 | || (rtx_equal_for_memref_p (XEXP (x, 0), XEXP (y, 1)) | |
1549 | && rtx_equal_for_memref_p (XEXP (x, 1), XEXP (y, 0)))); | |
6285298a | 1550 | /* For commutative operations, the RTX match if the operand match in any |
1551 | order. Also handle the simple binary and unary cases without a loop. */ | |
6720e96c | 1552 | if (COMMUTATIVE_P (x)) |
6285298a | 1553 | { |
1554 | rtx xop0 = canon_rtx (XEXP (x, 0)); | |
1555 | rtx yop0 = canon_rtx (XEXP (y, 0)); | |
1556 | rtx yop1 = canon_rtx (XEXP (y, 1)); | |
1557 | ||
1558 | return ((rtx_equal_for_memref_p (xop0, yop0) | |
1559 | && rtx_equal_for_memref_p (canon_rtx (XEXP (x, 1)), yop1)) | |
1560 | || (rtx_equal_for_memref_p (xop0, yop1) | |
1561 | && rtx_equal_for_memref_p (canon_rtx (XEXP (x, 1)), yop0))); | |
1562 | } | |
6720e96c | 1563 | else if (NON_COMMUTATIVE_P (x)) |
6285298a | 1564 | { |
1565 | return (rtx_equal_for_memref_p (canon_rtx (XEXP (x, 0)), | |
aecda0d6 | 1566 | canon_rtx (XEXP (y, 0))) |
6285298a | 1567 | && rtx_equal_for_memref_p (canon_rtx (XEXP (x, 1)), |
1568 | canon_rtx (XEXP (y, 1)))); | |
1569 | } | |
6720e96c | 1570 | else if (UNARY_P (x)) |
6285298a | 1571 | return rtx_equal_for_memref_p (canon_rtx (XEXP (x, 0)), |
aecda0d6 | 1572 | canon_rtx (XEXP (y, 0))); |
ea0cb7ae | 1573 | |
1574 | /* Compare the elements. If any pair of corresponding elements | |
1a6c3add | 1575 | fail to match, return 0 for the whole things. |
1576 | ||
1577 | Limit cases to types which actually appear in addresses. */ | |
ea0cb7ae | 1578 | |
1579 | fmt = GET_RTX_FORMAT (code); | |
1580 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
1581 | { | |
1582 | switch (fmt[i]) | |
1583 | { | |
ea0cb7ae | 1584 | case 'i': |
1585 | if (XINT (x, i) != XINT (y, i)) | |
1586 | return 0; | |
1587 | break; | |
1588 | ||
ea0cb7ae | 1589 | case 'E': |
1590 | /* Two vectors must have the same length. */ | |
1591 | if (XVECLEN (x, i) != XVECLEN (y, i)) | |
1592 | return 0; | |
1593 | ||
1594 | /* And the corresponding elements must match. */ | |
1595 | for (j = 0; j < XVECLEN (x, i); j++) | |
6285298a | 1596 | if (rtx_equal_for_memref_p (canon_rtx (XVECEXP (x, i, j)), |
1597 | canon_rtx (XVECEXP (y, i, j))) == 0) | |
ea0cb7ae | 1598 | return 0; |
1599 | break; | |
1600 | ||
1601 | case 'e': | |
6285298a | 1602 | if (rtx_equal_for_memref_p (canon_rtx (XEXP (x, i)), |
1603 | canon_rtx (XEXP (y, i))) == 0) | |
ea0cb7ae | 1604 | return 0; |
1605 | break; | |
1606 | ||
83ec925c | 1607 | /* This can happen for asm operands. */ |
1608 | case 's': | |
1609 | if (strcmp (XSTR (x, i), XSTR (y, i))) | |
1610 | return 0; | |
1611 | break; | |
1612 | ||
23c04f6b | 1613 | /* This can happen for an asm which clobbers memory. */ |
1614 | case '0': | |
1615 | break; | |
1616 | ||
ea0cb7ae | 1617 | /* It is believed that rtx's at this level will never |
1618 | contain anything but integers and other rtx's, | |
1619 | except for within LABEL_REFs and SYMBOL_REFs. */ | |
1620 | default: | |
64db345d | 1621 | gcc_unreachable (); |
ea0cb7ae | 1622 | } |
1623 | } | |
1624 | return 1; | |
1625 | } | |
1626 | ||
86e87ef6 | 1627 | static rtx |
aecda0d6 | 1628 | find_base_term (rtx x) |
ea0cb7ae | 1629 | { |
1617c276 | 1630 | cselib_val *val; |
8081d3a6 | 1631 | struct elt_loc_list *l, *f; |
1632 | rtx ret; | |
1617c276 | 1633 | |
15750b1e | 1634 | #if defined (FIND_BASE_TERM) |
1635 | /* Try machine-dependent ways to find the base term. */ | |
1636 | x = FIND_BASE_TERM (x); | |
1637 | #endif | |
1638 | ||
ea0cb7ae | 1639 | switch (GET_CODE (x)) |
1640 | { | |
1641 | case REG: | |
1642 | return REG_BASE_VALUE (x); | |
1643 | ||
2be64ef0 | 1644 | case TRUNCATE: |
04ec15fa | 1645 | /* As we do not know which address space the pointer is referring to, we can |
98155838 | 1646 | handle this only if the target does not support different pointer or |
1647 | address modes depending on the address space. */ | |
1648 | if (!target_default_pointer_address_modes_p ()) | |
1649 | return 0; | |
2be64ef0 | 1650 | if (GET_MODE_SIZE (GET_MODE (x)) < GET_MODE_SIZE (Pmode)) |
9342ee68 | 1651 | return 0; |
2be64ef0 | 1652 | /* Fall through. */ |
ea0cb7ae | 1653 | case HIGH: |
22dfec7d | 1654 | case PRE_INC: |
1655 | case PRE_DEC: | |
1656 | case POST_INC: | |
1657 | case POST_DEC: | |
2be64ef0 | 1658 | case PRE_MODIFY: |
1659 | case POST_MODIFY: | |
22dfec7d | 1660 | return find_base_term (XEXP (x, 0)); |
1661 | ||
5104ef69 | 1662 | case ZERO_EXTEND: |
1663 | case SIGN_EXTEND: /* Used for Alpha/NT pointers */ | |
04ec15fa | 1664 | /* As we do not know which address space the pointer is referring to, we can |
98155838 | 1665 | handle this only if the target does not support different pointer or |
1666 | address modes depending on the address space. */ | |
1667 | if (!target_default_pointer_address_modes_p ()) | |
1668 | return 0; | |
1669 | ||
5104ef69 | 1670 | { |
1671 | rtx temp = find_base_term (XEXP (x, 0)); | |
1672 | ||
85d654dd | 1673 | if (temp != 0 && CONSTANT_P (temp)) |
5104ef69 | 1674 | temp = convert_memory_address (Pmode, temp); |
5104ef69 | 1675 | |
1676 | return temp; | |
1677 | } | |
1678 | ||
1617c276 | 1679 | case VALUE: |
1680 | val = CSELIB_VAL_PTR (x); | |
8081d3a6 | 1681 | ret = NULL_RTX; |
1682 | ||
a2e238f7 | 1683 | if (!val) |
8081d3a6 | 1684 | return ret; |
1685 | ||
bb75e83f | 1686 | if (cselib_sp_based_value_p (val)) |
1687 | return static_reg_base_value[STACK_POINTER_REGNUM]; | |
1688 | ||
8081d3a6 | 1689 | f = val->locs; |
1690 | /* Temporarily reset val->locs to avoid infinite recursion. */ | |
1691 | val->locs = NULL; | |
1692 | ||
1693 | for (l = f; l; l = l->next) | |
1694 | if (GET_CODE (l->loc) == VALUE | |
1695 | && CSELIB_VAL_PTR (l->loc)->locs | |
1696 | && !CSELIB_VAL_PTR (l->loc)->locs->next | |
1697 | && CSELIB_VAL_PTR (l->loc)->locs->loc == x) | |
1698 | continue; | |
1699 | else if ((ret = find_base_term (l->loc)) != 0) | |
1700 | break; | |
1701 | ||
1702 | val->locs = f; | |
1703 | return ret; | |
1617c276 | 1704 | |
a7cf5bbf | 1705 | case LO_SUM: |
1706 | /* The standard form is (lo_sum reg sym) so look only at the | |
1707 | second operand. */ | |
1708 | return find_base_term (XEXP (x, 1)); | |
1709 | ||
ea0cb7ae | 1710 | case CONST: |
1711 | x = XEXP (x, 0); | |
1712 | if (GET_CODE (x) != PLUS && GET_CODE (x) != MINUS) | |
1713 | return 0; | |
b4b174c3 | 1714 | /* Fall through. */ |
ea0cb7ae | 1715 | case PLUS: |
1716 | case MINUS: | |
1717 | { | |
1b5bbc4c | 1718 | rtx tmp1 = XEXP (x, 0); |
1719 | rtx tmp2 = XEXP (x, 1); | |
1720 | ||
35a3065a | 1721 | /* This is a little bit tricky since we have to determine which of |
1b5bbc4c | 1722 | the two operands represents the real base address. Otherwise this |
1723 | routine may return the index register instead of the base register. | |
1724 | ||
1725 | That may cause us to believe no aliasing was possible, when in | |
1726 | fact aliasing is possible. | |
1727 | ||
1728 | We use a few simple tests to guess the base register. Additional | |
1729 | tests can certainly be added. For example, if one of the operands | |
1730 | is a shift or multiply, then it must be the index register and the | |
1731 | other operand is the base register. */ | |
9342ee68 | 1732 | |
15750b1e | 1733 | if (tmp1 == pic_offset_table_rtx && CONSTANT_P (tmp2)) |
1734 | return find_base_term (tmp2); | |
1735 | ||
5abc99e4 | 1736 | /* If either operand is known to be a pointer, then prefer it |
1b5bbc4c | 1737 | to determine the base term. */ |
e61a0a7f | 1738 | if (REG_P (tmp1) && REG_POINTER (tmp1)) |
5abc99e4 | 1739 | ; |
1740 | else if (REG_P (tmp2) && REG_POINTER (tmp2)) | |
02b0d167 | 1741 | { |
5abc99e4 | 1742 | rtx tem = tmp1; |
1743 | tmp1 = tmp2; | |
1744 | tmp2 = tem; | |
02b0d167 | 1745 | } |
1b5bbc4c | 1746 | |
5abc99e4 | 1747 | /* Go ahead and find the base term for both operands. If either base |
1748 | term is from a pointer or is a named object or a special address | |
1b5bbc4c | 1749 | (like an argument or stack reference), then use it for the |
1750 | base term. */ | |
ce0deccc | 1751 | rtx base = find_base_term (tmp1); |
1752 | if (base != NULL_RTX | |
5abc99e4 | 1753 | && ((REG_P (tmp1) && REG_POINTER (tmp1)) |
ce0deccc | 1754 | || known_base_value_p (base))) |
1755 | return base; | |
1756 | base = find_base_term (tmp2); | |
1757 | if (base != NULL_RTX | |
5abc99e4 | 1758 | && ((REG_P (tmp2) && REG_POINTER (tmp2)) |
ce0deccc | 1759 | || known_base_value_p (base))) |
1760 | return base; | |
1b5bbc4c | 1761 | |
1762 | /* We could not determine which of the two operands was the | |
1763 | base register and which was the index. So we can determine | |
1764 | nothing from the base alias check. */ | |
1765 | return 0; | |
ea0cb7ae | 1766 | } |
1767 | ||
1768 | case AND: | |
971ba038 | 1769 | if (CONST_INT_P (XEXP (x, 1)) && INTVAL (XEXP (x, 1)) != 0) |
2be64ef0 | 1770 | return find_base_term (XEXP (x, 0)); |
ea0cb7ae | 1771 | return 0; |
1772 | ||
1773 | case SYMBOL_REF: | |
1774 | case LABEL_REF: | |
1775 | return x; | |
1776 | ||
1777 | default: | |
1778 | return 0; | |
1779 | } | |
1780 | } | |
1781 | ||
86e87ef6 | 1782 | /* Return true if accesses to address X may alias accesses based |
1783 | on the stack pointer. */ | |
1784 | ||
1785 | bool | |
1786 | may_be_sp_based_p (rtx x) | |
1787 | { | |
1788 | rtx base = find_base_term (x); | |
1789 | return !base || base == static_reg_base_value[STACK_POINTER_REGNUM]; | |
1790 | } | |
1791 | ||
ea0cb7ae | 1792 | /* Return 0 if the addresses X and Y are known to point to different |
1793 | objects, 1 if they might be pointers to the same object. */ | |
1794 | ||
1795 | static int | |
5abc99e4 | 1796 | base_alias_check (rtx x, rtx x_base, rtx y, rtx y_base, |
1797 | enum machine_mode x_mode, enum machine_mode y_mode) | |
ea0cb7ae | 1798 | { |
3ad3a0ed | 1799 | /* If the address itself has no known base see if a known equivalent |
1800 | value has one. If either address still has no known base, nothing | |
1801 | is known about aliasing. */ | |
1802 | if (x_base == 0) | |
1803 | { | |
1804 | rtx x_c; | |
083a2b5e | 1805 | |
3ad3a0ed | 1806 | if (! flag_expensive_optimizations || (x_c = canon_rtx (x)) == x) |
1807 | return 1; | |
083a2b5e | 1808 | |
3ad3a0ed | 1809 | x_base = find_base_term (x_c); |
1810 | if (x_base == 0) | |
1811 | return 1; | |
1812 | } | |
ea0cb7ae | 1813 | |
3ad3a0ed | 1814 | if (y_base == 0) |
1815 | { | |
1816 | rtx y_c; | |
1817 | if (! flag_expensive_optimizations || (y_c = canon_rtx (y)) == y) | |
1818 | return 1; | |
083a2b5e | 1819 | |
3ad3a0ed | 1820 | y_base = find_base_term (y_c); |
1821 | if (y_base == 0) | |
1822 | return 1; | |
1823 | } | |
1824 | ||
1825 | /* If the base addresses are equal nothing is known about aliasing. */ | |
1826 | if (rtx_equal_p (x_base, y_base)) | |
ea0cb7ae | 1827 | return 1; |
1828 | ||
f92492f7 | 1829 | /* The base addresses are different expressions. If they are not accessed |
1830 | via AND, there is no conflict. We can bring knowledge of object | |
1831 | alignment into play here. For example, on alpha, "char a, b;" can | |
1832 | alias one another, though "char a; long b;" cannot. AND addesses may | |
1833 | implicitly alias surrounding objects; i.e. unaligned access in DImode | |
1834 | via AND address can alias all surrounding object types except those | |
1835 | with aligment 8 or higher. */ | |
1836 | if (GET_CODE (x) == AND && GET_CODE (y) == AND) | |
1837 | return 1; | |
1838 | if (GET_CODE (x) == AND | |
971ba038 | 1839 | && (!CONST_INT_P (XEXP (x, 1)) |
f92492f7 | 1840 | || (int) GET_MODE_UNIT_SIZE (y_mode) < -INTVAL (XEXP (x, 1)))) |
1841 | return 1; | |
1842 | if (GET_CODE (y) == AND | |
971ba038 | 1843 | && (!CONST_INT_P (XEXP (y, 1)) |
f92492f7 | 1844 | || (int) GET_MODE_UNIT_SIZE (x_mode) < -INTVAL (XEXP (y, 1)))) |
1845 | return 1; | |
1846 | ||
1847 | /* Differing symbols not accessed via AND never alias. */ | |
ea0cb7ae | 1848 | if (GET_CODE (x_base) != ADDRESS && GET_CODE (y_base) != ADDRESS) |
f92492f7 | 1849 | return 0; |
ea0cb7ae | 1850 | |
4e56ceb1 | 1851 | if (unique_base_value_p (x_base) || unique_base_value_p (y_base)) |
ea0cb7ae | 1852 | return 0; |
1853 | ||
e190b18a | 1854 | return 1; |
ea0cb7ae | 1855 | } |
1856 | ||
fa5e407e | 1857 | /* Callback for for_each_rtx, that returns 1 upon encountering a VALUE |
1858 | whose UID is greater than the int uid that D points to. */ | |
1859 | ||
1860 | static int | |
1861 | refs_newer_value_cb (rtx *x, void *d) | |
1862 | { | |
1863 | if (GET_CODE (*x) == VALUE && CSELIB_VAL_PTR (*x)->uid > *(int *)d) | |
1864 | return 1; | |
1865 | ||
1866 | return 0; | |
1867 | } | |
1868 | ||
1869 | /* Return TRUE if EXPR refers to a VALUE whose uid is greater than | |
1870 | that of V. */ | |
1871 | ||
1872 | static bool | |
1873 | refs_newer_value_p (rtx expr, rtx v) | |
1874 | { | |
1875 | int minuid = CSELIB_VAL_PTR (v)->uid; | |
1876 | ||
1877 | return for_each_rtx (&expr, refs_newer_value_cb, &minuid); | |
1878 | } | |
1879 | ||
1617c276 | 1880 | /* Convert the address X into something we can use. This is done by returning |
1881 | it unchanged unless it is a value; in the latter case we call cselib to get | |
1882 | a more useful rtx. */ | |
f7c44134 | 1883 | |
8e802be9 | 1884 | rtx |
aecda0d6 | 1885 | get_addr (rtx x) |
1617c276 | 1886 | { |
1887 | cselib_val *v; | |
1888 | struct elt_loc_list *l; | |
1889 | ||
1890 | if (GET_CODE (x) != VALUE) | |
1891 | return x; | |
1892 | v = CSELIB_VAL_PTR (x); | |
a2e238f7 | 1893 | if (v) |
1894 | { | |
85b6e75b | 1895 | bool have_equivs = cselib_have_permanent_equivalences (); |
1896 | if (have_equivs) | |
1897 | v = canonical_cselib_val (v); | |
a2e238f7 | 1898 | for (l = v->locs; l; l = l->next) |
1899 | if (CONSTANT_P (l->loc)) | |
1900 | return l->loc; | |
1901 | for (l = v->locs; l; l = l->next) | |
85b6e75b | 1902 | if (!REG_P (l->loc) && !MEM_P (l->loc) |
1903 | /* Avoid infinite recursion when potentially dealing with | |
1904 | var-tracking artificial equivalences, by skipping the | |
1905 | equivalences themselves, and not choosing expressions | |
1906 | that refer to newer VALUEs. */ | |
1907 | && (!have_equivs | |
1908 | || (GET_CODE (l->loc) != VALUE | |
1909 | && !refs_newer_value_p (l->loc, x)))) | |
fa5e407e | 1910 | return l->loc; |
85b6e75b | 1911 | if (have_equivs) |
1912 | { | |
1913 | for (l = v->locs; l; l = l->next) | |
1914 | if (REG_P (l->loc) | |
1915 | || (GET_CODE (l->loc) != VALUE | |
1916 | && !refs_newer_value_p (l->loc, x))) | |
1917 | return l->loc; | |
1918 | /* Return the canonical value. */ | |
1919 | return v->val_rtx; | |
1920 | } | |
1921 | if (v->locs) | |
1922 | return v->locs->loc; | |
a2e238f7 | 1923 | } |
1617c276 | 1924 | return x; |
1925 | } | |
1926 | ||
b9f92976 | 1927 | /* Return the address of the (N_REFS + 1)th memory reference to ADDR |
1928 | where SIZE is the size in bytes of the memory reference. If ADDR | |
1929 | is not modified by the memory reference then ADDR is returned. */ | |
1930 | ||
0aa9ffd4 | 1931 | static rtx |
aecda0d6 | 1932 | addr_side_effect_eval (rtx addr, int size, int n_refs) |
b9f92976 | 1933 | { |
1934 | int offset = 0; | |
9342ee68 | 1935 | |
b9f92976 | 1936 | switch (GET_CODE (addr)) |
1937 | { | |
1938 | case PRE_INC: | |
1939 | offset = (n_refs + 1) * size; | |
1940 | break; | |
1941 | case PRE_DEC: | |
1942 | offset = -(n_refs + 1) * size; | |
1943 | break; | |
1944 | case POST_INC: | |
1945 | offset = n_refs * size; | |
1946 | break; | |
1947 | case POST_DEC: | |
1948 | offset = -n_refs * size; | |
1949 | break; | |
1950 | ||
1951 | default: | |
1952 | return addr; | |
1953 | } | |
9342ee68 | 1954 | |
b9f92976 | 1955 | if (offset) |
6285298a | 1956 | addr = gen_rtx_PLUS (GET_MODE (addr), XEXP (addr, 0), |
c338f2e3 | 1957 | gen_int_mode (offset, GET_MODE (addr))); |
b9f92976 | 1958 | else |
1959 | addr = XEXP (addr, 0); | |
6285298a | 1960 | addr = canon_rtx (addr); |
b9f92976 | 1961 | |
1962 | return addr; | |
1963 | } | |
1964 | ||
340ad617 | 1965 | /* Return TRUE if an object X sized at XSIZE bytes and another object |
1966 | Y sized at YSIZE bytes, starting C bytes after X, may overlap. If | |
1967 | any of the sizes is zero, assume an overlap, otherwise use the | |
1968 | absolute value of the sizes as the actual sizes. */ | |
1969 | ||
1970 | static inline bool | |
1971 | offset_overlap_p (HOST_WIDE_INT c, int xsize, int ysize) | |
1972 | { | |
1973 | return (xsize == 0 || ysize == 0 | |
1974 | || (c >= 0 | |
1975 | ? (abs (xsize) > c) | |
1976 | : (abs (ysize) > -c))); | |
1977 | } | |
1978 | ||
40f9e245 | 1979 | /* Return one if X and Y (memory addresses) reference the |
1980 | same location in memory or if the references overlap. | |
1981 | Return zero if they do not overlap, else return | |
1982 | minus one in which case they still might reference the same location. | |
1983 | ||
1984 | C is an offset accumulator. When | |
ea0cb7ae | 1985 | C is nonzero, we are testing aliases between X and Y + C. |
1986 | XSIZE is the size in bytes of the X reference, | |
1987 | similarly YSIZE is the size in bytes for Y. | |
6285298a | 1988 | Expect that canon_rtx has been already called for X and Y. |
ea0cb7ae | 1989 | |
1990 | If XSIZE or YSIZE is zero, we do not know the amount of memory being | |
1991 | referenced (the reference was BLKmode), so make the most pessimistic | |
1992 | assumptions. | |
1993 | ||
1e38bf49 | 1994 | If XSIZE or YSIZE is negative, we may access memory outside the object |
1995 | being referenced as a side effect. This can happen when using AND to | |
1996 | align memory references, as is done on the Alpha. | |
1997 | ||
ea0cb7ae | 1998 | Nice to notice that varying addresses cannot conflict with fp if no |
40f9e245 | 1999 | local variables had their addresses taken, but that's too hard now. |
2000 | ||
2001 | ??? Contrary to the tree alias oracle this does not return | |
2002 | one for X + non-constant and Y + non-constant when X and Y are equal. | |
2003 | If that is fixed the TBAA hack for union type-punning can be removed. */ | |
ea0cb7ae | 2004 | |
ea0cb7ae | 2005 | static int |
aecda0d6 | 2006 | memrefs_conflict_p (int xsize, rtx x, int ysize, rtx y, HOST_WIDE_INT c) |
ea0cb7ae | 2007 | { |
1617c276 | 2008 | if (GET_CODE (x) == VALUE) |
6f9688ce | 2009 | { |
2010 | if (REG_P (y)) | |
2011 | { | |
4e15f806 | 2012 | struct elt_loc_list *l = NULL; |
2013 | if (CSELIB_VAL_PTR (x)) | |
fa5e407e | 2014 | for (l = canonical_cselib_val (CSELIB_VAL_PTR (x))->locs; |
2015 | l; l = l->next) | |
4e15f806 | 2016 | if (REG_P (l->loc) && rtx_equal_for_memref_p (l->loc, y)) |
2017 | break; | |
6f9688ce | 2018 | if (l) |
2019 | x = y; | |
2020 | else | |
2021 | x = get_addr (x); | |
2022 | } | |
2023 | /* Don't call get_addr if y is the same VALUE. */ | |
2024 | else if (x != y) | |
2025 | x = get_addr (x); | |
2026 | } | |
1617c276 | 2027 | if (GET_CODE (y) == VALUE) |
6f9688ce | 2028 | { |
2029 | if (REG_P (x)) | |
2030 | { | |
4e15f806 | 2031 | struct elt_loc_list *l = NULL; |
2032 | if (CSELIB_VAL_PTR (y)) | |
fa5e407e | 2033 | for (l = canonical_cselib_val (CSELIB_VAL_PTR (y))->locs; |
2034 | l; l = l->next) | |
4e15f806 | 2035 | if (REG_P (l->loc) && rtx_equal_for_memref_p (l->loc, x)) |
2036 | break; | |
6f9688ce | 2037 | if (l) |
2038 | y = x; | |
2039 | else | |
2040 | y = get_addr (y); | |
2041 | } | |
2042 | /* Don't call get_addr if x is the same VALUE. */ | |
2043 | else if (y != x) | |
2044 | y = get_addr (y); | |
2045 | } | |
ea0cb7ae | 2046 | if (GET_CODE (x) == HIGH) |
2047 | x = XEXP (x, 0); | |
2048 | else if (GET_CODE (x) == LO_SUM) | |
2049 | x = XEXP (x, 1); | |
2050 | else | |
340ad617 | 2051 | x = addr_side_effect_eval (x, abs (xsize), 0); |
ea0cb7ae | 2052 | if (GET_CODE (y) == HIGH) |
2053 | y = XEXP (y, 0); | |
2054 | else if (GET_CODE (y) == LO_SUM) | |
2055 | y = XEXP (y, 1); | |
2056 | else | |
340ad617 | 2057 | y = addr_side_effect_eval (y, abs (ysize), 0); |
ea0cb7ae | 2058 | |
2059 | if (rtx_equal_for_memref_p (x, y)) | |
2060 | { | |
340ad617 | 2061 | return offset_overlap_p (c, xsize, ysize); |
ea0cb7ae | 2062 | } |
2063 | ||
35d3eedd | 2064 | /* This code used to check for conflicts involving stack references and |
2065 | globals but the base address alias code now handles these cases. */ | |
ea0cb7ae | 2066 | |
2067 | if (GET_CODE (x) == PLUS) | |
2068 | { | |
2069 | /* The fact that X is canonicalized means that this | |
2070 | PLUS rtx is canonicalized. */ | |
2071 | rtx x0 = XEXP (x, 0); | |
2072 | rtx x1 = XEXP (x, 1); | |
2073 | ||
2074 | if (GET_CODE (y) == PLUS) | |
2075 | { | |
2076 | /* The fact that Y is canonicalized means that this | |
2077 | PLUS rtx is canonicalized. */ | |
2078 | rtx y0 = XEXP (y, 0); | |
2079 | rtx y1 = XEXP (y, 1); | |
2080 | ||
2081 | if (rtx_equal_for_memref_p (x1, y1)) | |
2082 | return memrefs_conflict_p (xsize, x0, ysize, y0, c); | |
2083 | if (rtx_equal_for_memref_p (x0, y0)) | |
2084 | return memrefs_conflict_p (xsize, x1, ysize, y1, c); | |
971ba038 | 2085 | if (CONST_INT_P (x1)) |
5f59c49b | 2086 | { |
971ba038 | 2087 | if (CONST_INT_P (y1)) |
5f59c49b | 2088 | return memrefs_conflict_p (xsize, x0, ysize, y0, |
2089 | c - INTVAL (x1) + INTVAL (y1)); | |
2090 | else | |
2091 | return memrefs_conflict_p (xsize, x0, ysize, y, | |
2092 | c - INTVAL (x1)); | |
2093 | } | |
971ba038 | 2094 | else if (CONST_INT_P (y1)) |
ea0cb7ae | 2095 | return memrefs_conflict_p (xsize, x, ysize, y0, c + INTVAL (y1)); |
2096 | ||
40f9e245 | 2097 | return -1; |
ea0cb7ae | 2098 | } |
971ba038 | 2099 | else if (CONST_INT_P (x1)) |
ea0cb7ae | 2100 | return memrefs_conflict_p (xsize, x0, ysize, y, c - INTVAL (x1)); |
2101 | } | |
2102 | else if (GET_CODE (y) == PLUS) | |
2103 | { | |
2104 | /* The fact that Y is canonicalized means that this | |
2105 | PLUS rtx is canonicalized. */ | |
2106 | rtx y0 = XEXP (y, 0); | |
2107 | rtx y1 = XEXP (y, 1); | |
2108 | ||
971ba038 | 2109 | if (CONST_INT_P (y1)) |
ea0cb7ae | 2110 | return memrefs_conflict_p (xsize, x, ysize, y0, c + INTVAL (y1)); |
2111 | else | |
40f9e245 | 2112 | return -1; |
ea0cb7ae | 2113 | } |
2114 | ||
2115 | if (GET_CODE (x) == GET_CODE (y)) | |
2116 | switch (GET_CODE (x)) | |
2117 | { | |
2118 | case MULT: | |
2119 | { | |
2120 | /* Handle cases where we expect the second operands to be the | |
2121 | same, and check only whether the first operand would conflict | |
2122 | or not. */ | |
2123 | rtx x0, y0; | |
2124 | rtx x1 = canon_rtx (XEXP (x, 1)); | |
2125 | rtx y1 = canon_rtx (XEXP (y, 1)); | |
2126 | if (! rtx_equal_for_memref_p (x1, y1)) | |
40f9e245 | 2127 | return -1; |
ea0cb7ae | 2128 | x0 = canon_rtx (XEXP (x, 0)); |
2129 | y0 = canon_rtx (XEXP (y, 0)); | |
2130 | if (rtx_equal_for_memref_p (x0, y0)) | |
340ad617 | 2131 | return offset_overlap_p (c, xsize, ysize); |
ea0cb7ae | 2132 | |
2133 | /* Can't properly adjust our sizes. */ | |
971ba038 | 2134 | if (!CONST_INT_P (x1)) |
40f9e245 | 2135 | return -1; |
ea0cb7ae | 2136 | xsize /= INTVAL (x1); |
2137 | ysize /= INTVAL (x1); | |
2138 | c /= INTVAL (x1); | |
2139 | return memrefs_conflict_p (xsize, x0, ysize, y0, c); | |
2140 | } | |
99c14947 | 2141 | |
2142 | default: | |
2143 | break; | |
ea0cb7ae | 2144 | } |
2145 | ||
f78d5f53 | 2146 | /* Deal with alignment ANDs by adjusting offset and size so as to |
2147 | cover the maximum range, without taking any previously known | |
bb2bc4f3 | 2148 | alignment into account. Make a size negative after such an |
2149 | adjustments, so that, if we end up with e.g. two SYMBOL_REFs, we | |
2150 | assume a potential overlap, because they may end up in contiguous | |
2151 | memory locations and the stricter-alignment access may span over | |
2152 | part of both. */ | |
971ba038 | 2153 | if (GET_CODE (x) == AND && CONST_INT_P (XEXP (x, 1))) |
9fb1c31b | 2154 | { |
f78d5f53 | 2155 | HOST_WIDE_INT sc = INTVAL (XEXP (x, 1)); |
2156 | unsigned HOST_WIDE_INT uc = sc; | |
bb2bc4f3 | 2157 | if (sc < 0 && -uc == (uc & -uc)) |
f78d5f53 | 2158 | { |
bb2bc4f3 | 2159 | if (xsize > 0) |
2160 | xsize = -xsize; | |
340ad617 | 2161 | if (xsize) |
2162 | xsize += sc + 1; | |
686d491a | 2163 | c -= sc + 1; |
f78d5f53 | 2164 | return memrefs_conflict_p (xsize, canon_rtx (XEXP (x, 0)), |
2165 | ysize, y, c); | |
2166 | } | |
9fb1c31b | 2167 | } |
971ba038 | 2168 | if (GET_CODE (y) == AND && CONST_INT_P (XEXP (y, 1))) |
1e38bf49 | 2169 | { |
f78d5f53 | 2170 | HOST_WIDE_INT sc = INTVAL (XEXP (y, 1)); |
2171 | unsigned HOST_WIDE_INT uc = sc; | |
bb2bc4f3 | 2172 | if (sc < 0 && -uc == (uc & -uc)) |
f78d5f53 | 2173 | { |
bb2bc4f3 | 2174 | if (ysize > 0) |
2175 | ysize = -ysize; | |
340ad617 | 2176 | if (ysize) |
2177 | ysize += sc + 1; | |
686d491a | 2178 | c += sc + 1; |
f78d5f53 | 2179 | return memrefs_conflict_p (xsize, x, |
2180 | ysize, canon_rtx (XEXP (y, 0)), c); | |
2181 | } | |
1e38bf49 | 2182 | } |
ea0cb7ae | 2183 | |
2184 | if (CONSTANT_P (x)) | |
2185 | { | |
971ba038 | 2186 | if (CONST_INT_P (x) && CONST_INT_P (y)) |
ea0cb7ae | 2187 | { |
2188 | c += (INTVAL (y) - INTVAL (x)); | |
340ad617 | 2189 | return offset_overlap_p (c, xsize, ysize); |
ea0cb7ae | 2190 | } |
2191 | ||
2192 | if (GET_CODE (x) == CONST) | |
2193 | { | |
2194 | if (GET_CODE (y) == CONST) | |
2195 | return memrefs_conflict_p (xsize, canon_rtx (XEXP (x, 0)), | |
2196 | ysize, canon_rtx (XEXP (y, 0)), c); | |
2197 | else | |
2198 | return memrefs_conflict_p (xsize, canon_rtx (XEXP (x, 0)), | |
2199 | ysize, y, c); | |
2200 | } | |
2201 | if (GET_CODE (y) == CONST) | |
2202 | return memrefs_conflict_p (xsize, x, ysize, | |
2203 | canon_rtx (XEXP (y, 0)), c); | |
2204 | ||
340ad617 | 2205 | /* Assume a potential overlap for symbolic addresses that went |
2206 | through alignment adjustments (i.e., that have negative | |
2207 | sizes), because we can't know how far they are from each | |
2208 | other. */ | |
ea0cb7ae | 2209 | if (CONSTANT_P (y)) |
340ad617 | 2210 | return (xsize < 0 || ysize < 0 || offset_overlap_p (c, xsize, ysize)); |
ea0cb7ae | 2211 | |
40f9e245 | 2212 | return -1; |
ea0cb7ae | 2213 | } |
40f9e245 | 2214 | |
2215 | return -1; | |
ea0cb7ae | 2216 | } |
2217 | ||
2218 | /* Functions to compute memory dependencies. | |
2219 | ||
2220 | Since we process the insns in execution order, we can build tables | |
2221 | to keep track of what registers are fixed (and not aliased), what registers | |
2222 | are varying in known ways, and what registers are varying in unknown | |
2223 | ways. | |
2224 | ||
2225 | If both memory references are volatile, then there must always be a | |
2226 | dependence between the two references, since their order can not be | |
2227 | changed. A volatile and non-volatile reference can be interchanged | |
9342ee68 | 2228 | though. |
ea0cb7ae | 2229 | |
376a287d | 2230 | We also must allow AND addresses, because they may generate accesses |
2231 | outside the object being referenced. This is used to generate aligned | |
2232 | addresses from unaligned addresses, for instance, the alpha | |
71d9e995 | 2233 | storeqi_unaligned pattern. */ |
ea0cb7ae | 2234 | |
2235 | /* Read dependence: X is read after read in MEM takes place. There can | |
eae726e9 | 2236 | only be a dependence here if both reads are volatile, or if either is |
2237 | an explicit barrier. */ | |
ea0cb7ae | 2238 | |
2239 | int | |
52d07779 | 2240 | read_dependence (const_rtx mem, const_rtx x) |
ea0cb7ae | 2241 | { |
eae726e9 | 2242 | if (MEM_VOLATILE_P (x) && MEM_VOLATILE_P (mem)) |
2243 | return true; | |
2244 | if (MEM_ALIAS_SET (x) == ALIAS_SET_MEMORY_BARRIER | |
2245 | || MEM_ALIAS_SET (mem) == ALIAS_SET_MEMORY_BARRIER) | |
2246 | return true; | |
2247 | return false; | |
ea0cb7ae | 2248 | } |
2249 | ||
b10dbbca | 2250 | /* Look at the bottom of the COMPONENT_REF list for a DECL, and return it. */ |
2251 | ||
2252 | static tree | |
aecda0d6 | 2253 | decl_for_component_ref (tree x) |
b10dbbca | 2254 | { |
2255 | do | |
2256 | { | |
2257 | x = TREE_OPERAND (x, 0); | |
2258 | } | |
2259 | while (x && TREE_CODE (x) == COMPONENT_REF); | |
2260 | ||
2261 | return x && DECL_P (x) ? x : NULL_TREE; | |
2262 | } | |
2263 | ||
da443c27 | 2264 | /* Walk up the COMPONENT_REF list in X and adjust *OFFSET to compensate |
2265 | for the offset of the field reference. *KNOWN_P says whether the | |
2266 | offset is known. */ | |
b10dbbca | 2267 | |
da443c27 | 2268 | static void |
2269 | adjust_offset_for_component_ref (tree x, bool *known_p, | |
2270 | HOST_WIDE_INT *offset) | |
b10dbbca | 2271 | { |
da443c27 | 2272 | if (!*known_p) |
2273 | return; | |
9342ee68 | 2274 | do |
b10dbbca | 2275 | { |
da443c27 | 2276 | tree xoffset = component_ref_field_offset (x); |
b10dbbca | 2277 | tree field = TREE_OPERAND (x, 1); |
2278 | ||
cd4547bf | 2279 | if (! tree_fits_uhwi_p (xoffset)) |
da443c27 | 2280 | { |
2281 | *known_p = false; | |
2282 | return; | |
2283 | } | |
6a0712d4 | 2284 | *offset += (tree_to_uhwi (xoffset) |
2285 | + (tree_to_uhwi (DECL_FIELD_BIT_OFFSET (field)) | |
b10dbbca | 2286 | / BITS_PER_UNIT)); |
2287 | ||
2288 | x = TREE_OPERAND (x, 0); | |
2289 | } | |
2290 | while (x && TREE_CODE (x) == COMPONENT_REF); | |
b10dbbca | 2291 | } |
2292 | ||
d716ce75 | 2293 | /* Return nonzero if we can determine the exprs corresponding to memrefs |
a84256aa | 2294 | X and Y and they do not overlap. |
2295 | If LOOP_VARIANT is set, skip offset-based disambiguation */ | |
ddb3d16f | 2296 | |
5fe61d21 | 2297 | int |
a84256aa | 2298 | nonoverlapping_memrefs_p (const_rtx x, const_rtx y, bool loop_invariant) |
ddb3d16f | 2299 | { |
b10dbbca | 2300 | tree exprx = MEM_EXPR (x), expry = MEM_EXPR (y); |
ddb3d16f | 2301 | rtx rtlx, rtly; |
2302 | rtx basex, basey; | |
da443c27 | 2303 | bool moffsetx_known_p, moffsety_known_p; |
2304 | HOST_WIDE_INT moffsetx = 0, moffsety = 0; | |
ddb3d16f | 2305 | HOST_WIDE_INT offsetx = 0, offsety = 0, sizex, sizey, tem; |
2306 | ||
b10dbbca | 2307 | /* Unless both have exprs, we can't tell anything. */ |
2308 | if (exprx == 0 || expry == 0) | |
2309 | return 0; | |
c6b80e5d | 2310 | |
2311 | /* For spill-slot accesses make sure we have valid offsets. */ | |
2312 | if ((exprx == get_spill_slot_decl (false) | |
da443c27 | 2313 | && ! MEM_OFFSET_KNOWN_P (x)) |
c6b80e5d | 2314 | || (expry == get_spill_slot_decl (false) |
da443c27 | 2315 | && ! MEM_OFFSET_KNOWN_P (y))) |
c6b80e5d | 2316 | return 0; |
a0c938f0 | 2317 | |
b10dbbca | 2318 | /* If the field reference test failed, look at the DECLs involved. */ |
da443c27 | 2319 | moffsetx_known_p = MEM_OFFSET_KNOWN_P (x); |
2320 | if (moffsetx_known_p) | |
2321 | moffsetx = MEM_OFFSET (x); | |
b10dbbca | 2322 | if (TREE_CODE (exprx) == COMPONENT_REF) |
2323 | { | |
b50c6e49 | 2324 | tree t = decl_for_component_ref (exprx); |
2325 | if (! t) | |
2326 | return 0; | |
da443c27 | 2327 | adjust_offset_for_component_ref (exprx, &moffsetx_known_p, &moffsetx); |
b50c6e49 | 2328 | exprx = t; |
b10dbbca | 2329 | } |
2d8fe5d0 | 2330 | |
da443c27 | 2331 | moffsety_known_p = MEM_OFFSET_KNOWN_P (y); |
2332 | if (moffsety_known_p) | |
2333 | moffsety = MEM_OFFSET (y); | |
b10dbbca | 2334 | if (TREE_CODE (expry) == COMPONENT_REF) |
2335 | { | |
b50c6e49 | 2336 | tree t = decl_for_component_ref (expry); |
2337 | if (! t) | |
2338 | return 0; | |
da443c27 | 2339 | adjust_offset_for_component_ref (expry, &moffsety_known_p, &moffsety); |
b50c6e49 | 2340 | expry = t; |
b10dbbca | 2341 | } |
2342 | ||
2343 | if (! DECL_P (exprx) || ! DECL_P (expry)) | |
ddb3d16f | 2344 | return 0; |
2345 | ||
119147d1 | 2346 | /* With invalid code we can end up storing into the constant pool. |
2347 | Bail out to avoid ICEing when creating RTL for this. | |
2348 | See gfortran.dg/lto/20091028-2_0.f90. */ | |
2349 | if (TREE_CODE (exprx) == CONST_DECL | |
2350 | || TREE_CODE (expry) == CONST_DECL) | |
2351 | return 1; | |
2352 | ||
b10dbbca | 2353 | rtlx = DECL_RTL (exprx); |
2354 | rtly = DECL_RTL (expry); | |
ddb3d16f | 2355 | |
10e949c1 | 2356 | /* If either RTL is not a MEM, it must be a REG or CONCAT, meaning they |
2357 | can't overlap unless they are the same because we never reuse that part | |
2358 | of the stack frame used for locals for spilled pseudos. */ | |
e16ceb8e | 2359 | if ((!MEM_P (rtlx) || !MEM_P (rtly)) |
10e949c1 | 2360 | && ! rtx_equal_p (rtlx, rtly)) |
ddb3d16f | 2361 | return 1; |
2362 | ||
04ec15fa | 2363 | /* If we have MEMs referring to different address spaces (which can |
bd1a81f7 | 2364 | potentially overlap), we cannot easily tell from the addresses |
2365 | whether the references overlap. */ | |
2366 | if (MEM_P (rtlx) && MEM_P (rtly) | |
2367 | && MEM_ADDR_SPACE (rtlx) != MEM_ADDR_SPACE (rtly)) | |
2368 | return 0; | |
2369 | ||
ddb3d16f | 2370 | /* Get the base and offsets of both decls. If either is a register, we |
2371 | know both are and are the same, so use that as the base. The only | |
2372 | we can avoid overlap is if we can deduce that they are nonoverlapping | |
2373 | pieces of that decl, which is very rare. */ | |
e16ceb8e | 2374 | basex = MEM_P (rtlx) ? XEXP (rtlx, 0) : rtlx; |
971ba038 | 2375 | if (GET_CODE (basex) == PLUS && CONST_INT_P (XEXP (basex, 1))) |
ddb3d16f | 2376 | offsetx = INTVAL (XEXP (basex, 1)), basex = XEXP (basex, 0); |
2377 | ||
e16ceb8e | 2378 | basey = MEM_P (rtly) ? XEXP (rtly, 0) : rtly; |
971ba038 | 2379 | if (GET_CODE (basey) == PLUS && CONST_INT_P (XEXP (basey, 1))) |
ddb3d16f | 2380 | offsety = INTVAL (XEXP (basey, 1)), basey = XEXP (basey, 0); |
2381 | ||
c9c48480 | 2382 | /* If the bases are different, we know they do not overlap if both |
9342ee68 | 2383 | are constants or if one is a constant and the other a pointer into the |
c9c48480 | 2384 | stack frame. Otherwise a different base means we can't tell if they |
2385 | overlap or not. */ | |
2386 | if (! rtx_equal_p (basex, basey)) | |
9342ee68 | 2387 | return ((CONSTANT_P (basex) && CONSTANT_P (basey)) |
2388 | || (CONSTANT_P (basex) && REG_P (basey) | |
2389 | && REGNO_PTR_FRAME_P (REGNO (basey))) | |
2390 | || (CONSTANT_P (basey) && REG_P (basex) | |
2391 | && REGNO_PTR_FRAME_P (REGNO (basex)))); | |
ddb3d16f | 2392 | |
a84256aa | 2393 | /* Offset based disambiguation not appropriate for loop invariant */ |
2394 | if (loop_invariant) | |
2395 | return 0; | |
2396 | ||
e16ceb8e | 2397 | sizex = (!MEM_P (rtlx) ? (int) GET_MODE_SIZE (GET_MODE (rtlx)) |
5b2a69fa | 2398 | : MEM_SIZE_KNOWN_P (rtlx) ? MEM_SIZE (rtlx) |
ddb3d16f | 2399 | : -1); |
e16ceb8e | 2400 | sizey = (!MEM_P (rtly) ? (int) GET_MODE_SIZE (GET_MODE (rtly)) |
5b2a69fa | 2401 | : MEM_SIZE_KNOWN_P (rtly) ? MEM_SIZE (rtly) |
2402 | : -1); | |
ddb3d16f | 2403 | |
dc2dfdb6 | 2404 | /* If we have an offset for either memref, it can update the values computed |
2405 | above. */ | |
da443c27 | 2406 | if (moffsetx_known_p) |
2407 | offsetx += moffsetx, sizex -= moffsetx; | |
2408 | if (moffsety_known_p) | |
2409 | offsety += moffsety, sizey -= moffsety; | |
ddb3d16f | 2410 | |
dc2dfdb6 | 2411 | /* If a memref has both a size and an offset, we can use the smaller size. |
203c9fee | 2412 | We can't do this if the offset isn't known because we must view this |
dc2dfdb6 | 2413 | memref as being anywhere inside the DECL's MEM. */ |
da443c27 | 2414 | if (MEM_SIZE_KNOWN_P (x) && moffsetx_known_p) |
5b2a69fa | 2415 | sizex = MEM_SIZE (x); |
da443c27 | 2416 | if (MEM_SIZE_KNOWN_P (y) && moffsety_known_p) |
5b2a69fa | 2417 | sizey = MEM_SIZE (y); |
ddb3d16f | 2418 | |
2419 | /* Put the values of the memref with the lower offset in X's values. */ | |
2420 | if (offsetx > offsety) | |
2421 | { | |
2422 | tem = offsetx, offsetx = offsety, offsety = tem; | |
2423 | tem = sizex, sizex = sizey, sizey = tem; | |
2424 | } | |
2425 | ||
2426 | /* If we don't know the size of the lower-offset value, we can't tell | |
2427 | if they conflict. Otherwise, we do the test. */ | |
906923ba | 2428 | return sizex >= 0 && offsety >= offsetx + sizex; |
ddb3d16f | 2429 | } |
2430 | ||
02f77cc4 | 2431 | /* Helper for true_dependence and canon_true_dependence. |
2432 | Checks for true dependence: X is read after store in MEM takes place. | |
ea0cb7ae | 2433 | |
02f77cc4 | 2434 | If MEM_CANONICALIZED is FALSE, then X_ADDR and MEM_ADDR should be |
2435 | NULL_RTX, and the canonical addresses of MEM and X are both computed | |
2436 | here. If MEM_CANONICALIZED, then MEM must be already canonicalized. | |
2437 | ||
2438 | If X_ADDR is non-NULL, it is used in preference of XEXP (x, 0). | |
2439 | ||
2440 | Returns 1 if there is a true dependence, 0 otherwise. */ | |
2441 | ||
2442 | static int | |
2443 | true_dependence_1 (const_rtx mem, enum machine_mode mem_mode, rtx mem_addr, | |
376a287d | 2444 | const_rtx x, rtx x_addr, bool mem_canonicalized) |
ea0cb7ae | 2445 | { |
9ee7cac5 | 2446 | rtx base; |
40f9e245 | 2447 | int ret; |
ea0cb7ae | 2448 | |
02f77cc4 | 2449 | gcc_checking_assert (mem_canonicalized ? (mem_addr != NULL_RTX) |
2450 | : (mem_addr == NULL_RTX && x_addr == NULL_RTX)); | |
2451 | ||
ea0cb7ae | 2452 | if (MEM_VOLATILE_P (x) && MEM_VOLATILE_P (mem)) |
2453 | return 1; | |
2454 | ||
a57f4c4d | 2455 | /* (mem:BLK (scratch)) is a special mechanism to conflict with everything. |
67837170 | 2456 | This is used in epilogue deallocation functions, and in cselib. */ |
a57f4c4d | 2457 | if (GET_MODE (x) == BLKmode && GET_CODE (XEXP (x, 0)) == SCRATCH) |
2458 | return 1; | |
2459 | if (GET_MODE (mem) == BLKmode && GET_CODE (XEXP (mem, 0)) == SCRATCH) | |
2460 | return 1; | |
c94cfd1c | 2461 | if (MEM_ALIAS_SET (x) == ALIAS_SET_MEMORY_BARRIER |
2462 | || MEM_ALIAS_SET (mem) == ALIAS_SET_MEMORY_BARRIER) | |
2463 | return 1; | |
a57f4c4d | 2464 | |
b04fab2a | 2465 | /* Read-only memory is by definition never modified, and therefore can't |
2466 | conflict with anything. We don't expect to find read-only set on MEM, | |
a53ff4c1 | 2467 | but stupid user tricks can produce them, so don't die. */ |
b04fab2a | 2468 | if (MEM_READONLY_P (x)) |
ea0cb7ae | 2469 | return 0; |
2470 | ||
04ec15fa | 2471 | /* If we have MEMs referring to different address spaces (which can |
bd1a81f7 | 2472 | potentially overlap), we cannot easily tell from the addresses |
2473 | whether the references overlap. */ | |
2474 | if (MEM_ADDR_SPACE (mem) != MEM_ADDR_SPACE (x)) | |
2475 | return 1; | |
2476 | ||
02f77cc4 | 2477 | if (! mem_addr) |
2478 | { | |
2479 | mem_addr = XEXP (mem, 0); | |
2480 | if (mem_mode == VOIDmode) | |
2481 | mem_mode = GET_MODE (mem); | |
2482 | } | |
2483 | ||
2484 | if (! x_addr) | |
288debed | 2485 | { |
62696140 | 2486 | x_addr = XEXP (x, 0); |
2487 | if (!((GET_CODE (x_addr) == VALUE | |
2488 | && GET_CODE (mem_addr) != VALUE | |
2489 | && reg_mentioned_p (x_addr, mem_addr)) | |
2490 | || (GET_CODE (x_addr) != VALUE | |
2491 | && GET_CODE (mem_addr) == VALUE | |
2492 | && reg_mentioned_p (mem_addr, x_addr)))) | |
2493 | { | |
2494 | x_addr = get_addr (x_addr); | |
2495 | if (! mem_canonicalized) | |
2496 | mem_addr = get_addr (mem_addr); | |
2497 | } | |
288debed | 2498 | } |
1617c276 | 2499 | |
002fe3cb | 2500 | base = find_base_term (x_addr); |
2501 | if (base && (GET_CODE (base) == LABEL_REF | |
2502 | || (GET_CODE (base) == SYMBOL_REF | |
2503 | && CONSTANT_POOL_ADDRESS_P (base)))) | |
2504 | return 0; | |
2505 | ||
5abc99e4 | 2506 | rtx mem_base = find_base_term (mem_addr); |
2507 | if (! base_alias_check (x_addr, base, mem_addr, mem_base, | |
2508 | GET_MODE (x), mem_mode)) | |
3ad3a0ed | 2509 | return 0; |
2510 | ||
1617c276 | 2511 | x_addr = canon_rtx (x_addr); |
02f77cc4 | 2512 | if (!mem_canonicalized) |
2513 | mem_addr = canon_rtx (mem_addr); | |
35d3eedd | 2514 | |
40f9e245 | 2515 | if ((ret = memrefs_conflict_p (GET_MODE_SIZE (mem_mode), mem_addr, |
2516 | SIZE_FOR_MODE (x), x_addr, 0)) != -1) | |
2517 | return ret; | |
2518 | ||
557f718e | 2519 | if (mems_in_disjoint_alias_sets_p (x, mem)) |
40f9e245 | 2520 | return 0; |
2521 | ||
a84256aa | 2522 | if (nonoverlapping_memrefs_p (mem, x, false)) |
4958fe6e | 2523 | return 0; |
658f1736 | 2524 | |
3a443843 | 2525 | return rtx_refs_may_alias_p (x, mem, true); |
8e802be9 | 2526 | } |
2527 | ||
02f77cc4 | 2528 | /* True dependence: X is read after store in MEM takes place. */ |
2529 | ||
2530 | int | |
376a287d | 2531 | true_dependence (const_rtx mem, enum machine_mode mem_mode, const_rtx x) |
02f77cc4 | 2532 | { |
2533 | return true_dependence_1 (mem, mem_mode, NULL_RTX, | |
376a287d | 2534 | x, NULL_RTX, /*mem_canonicalized=*/false); |
02f77cc4 | 2535 | } |
2536 | ||
8e802be9 | 2537 | /* Canonical true dependence: X is read after store in MEM takes place. |
9342ee68 | 2538 | Variant of true_dependence which assumes MEM has already been |
2539 | canonicalized (hence we no longer do that here). | |
02f77cc4 | 2540 | The mem_addr argument has been added, since true_dependence_1 computed |
2541 | this value prior to canonicalizing. */ | |
8e802be9 | 2542 | |
2543 | int | |
52d07779 | 2544 | canon_true_dependence (const_rtx mem, enum machine_mode mem_mode, rtx mem_addr, |
376a287d | 2545 | const_rtx x, rtx x_addr) |
8e802be9 | 2546 | { |
02f77cc4 | 2547 | return true_dependence_1 (mem, mem_mode, mem_addr, |
376a287d | 2548 | x, x_addr, /*mem_canonicalized=*/true); |
ea0cb7ae | 2549 | } |
2550 | ||
d10cfa8d | 2551 | /* Returns nonzero if a write to X might alias a previous read from |
ddba76b8 | 2552 | (or, if WRITEP is true, a write to) MEM. |
cb456db5 | 2553 | If X_CANONCALIZED is true, then X_ADDR is the canonicalized address of X, |
2554 | and X_MODE the mode for that access. | |
2555 | If MEM_CANONICALIZED is true, MEM is canonicalized. */ | |
ea0cb7ae | 2556 | |
fe69e4fe | 2557 | static int |
cb456db5 | 2558 | write_dependence_p (const_rtx mem, |
2559 | const_rtx x, enum machine_mode x_mode, rtx x_addr, | |
2560 | bool mem_canonicalized, bool x_canonicalized, bool writep) | |
ea0cb7ae | 2561 | { |
cb456db5 | 2562 | rtx mem_addr; |
9ee7cac5 | 2563 | rtx base; |
40f9e245 | 2564 | int ret; |
35d3eedd | 2565 | |
cb456db5 | 2566 | gcc_checking_assert (x_canonicalized |
2567 | ? (x_addr != NULL_RTX && x_mode != VOIDmode) | |
2568 | : (x_addr == NULL_RTX && x_mode == VOIDmode)); | |
ddba76b8 | 2569 | |
ea0cb7ae | 2570 | if (MEM_VOLATILE_P (x) && MEM_VOLATILE_P (mem)) |
2571 | return 1; | |
2572 | ||
a57f4c4d | 2573 | /* (mem:BLK (scratch)) is a special mechanism to conflict with everything. |
2574 | This is used in epilogue deallocation functions. */ | |
2575 | if (GET_MODE (x) == BLKmode && GET_CODE (XEXP (x, 0)) == SCRATCH) | |
2576 | return 1; | |
2577 | if (GET_MODE (mem) == BLKmode && GET_CODE (XEXP (mem, 0)) == SCRATCH) | |
2578 | return 1; | |
c94cfd1c | 2579 | if (MEM_ALIAS_SET (x) == ALIAS_SET_MEMORY_BARRIER |
2580 | || MEM_ALIAS_SET (mem) == ALIAS_SET_MEMORY_BARRIER) | |
2581 | return 1; | |
a57f4c4d | 2582 | |
b04fab2a | 2583 | /* A read from read-only memory can't conflict with read-write memory. */ |
2584 | if (!writep && MEM_READONLY_P (mem)) | |
2585 | return 0; | |
002fe3cb | 2586 | |
04ec15fa | 2587 | /* If we have MEMs referring to different address spaces (which can |
bd1a81f7 | 2588 | potentially overlap), we cannot easily tell from the addresses |
2589 | whether the references overlap. */ | |
2590 | if (MEM_ADDR_SPACE (mem) != MEM_ADDR_SPACE (x)) | |
2591 | return 1; | |
2592 | ||
288debed | 2593 | mem_addr = XEXP (mem, 0); |
cb456db5 | 2594 | if (!x_addr) |
288debed | 2595 | { |
cb456db5 | 2596 | x_addr = XEXP (x, 0); |
2597 | if (!((GET_CODE (x_addr) == VALUE | |
2598 | && GET_CODE (mem_addr) != VALUE | |
2599 | && reg_mentioned_p (x_addr, mem_addr)) | |
2600 | || (GET_CODE (x_addr) != VALUE | |
2601 | && GET_CODE (mem_addr) == VALUE | |
2602 | && reg_mentioned_p (mem_addr, x_addr)))) | |
2603 | { | |
2604 | x_addr = get_addr (x_addr); | |
2605 | if (!mem_canonicalized) | |
2606 | mem_addr = get_addr (mem_addr); | |
2607 | } | |
288debed | 2608 | } |
002fe3cb | 2609 | |
5abc99e4 | 2610 | base = find_base_term (mem_addr); |
2611 | if (! writep | |
2612 | && base | |
2613 | && (GET_CODE (base) == LABEL_REF | |
2614 | || (GET_CODE (base) == SYMBOL_REF | |
2615 | && CONSTANT_POOL_ADDRESS_P (base)))) | |
2616 | return 0; | |
9ee7cac5 | 2617 | |
5abc99e4 | 2618 | rtx x_base = find_base_term (x_addr); |
2619 | if (! base_alias_check (x_addr, x_base, mem_addr, base, GET_MODE (x), | |
1617c276 | 2620 | GET_MODE (mem))) |
b5ba9f3a | 2621 | return 0; |
2622 | ||
cb456db5 | 2623 | if (!x_canonicalized) |
ddba76b8 | 2624 | { |
cb456db5 | 2625 | x_addr = canon_rtx (x_addr); |
2626 | x_mode = GET_MODE (x); | |
ddba76b8 | 2627 | } |
cb456db5 | 2628 | if (!mem_canonicalized) |
2629 | mem_addr = canon_rtx (mem_addr); | |
35d3eedd | 2630 | |
cb456db5 | 2631 | if ((ret = memrefs_conflict_p (SIZE_FOR_MODE (mem), mem_addr, |
2632 | GET_MODE_SIZE (x_mode), x_addr, 0)) != -1) | |
40f9e245 | 2633 | return ret; |
2634 | ||
a84256aa | 2635 | if (nonoverlapping_memrefs_p (x, mem, false)) |
6a0934dd | 2636 | return 0; |
2637 | ||
3a443843 | 2638 | return rtx_refs_may_alias_p (x, mem, false); |
6a0934dd | 2639 | } |
2640 | ||
2641 | /* Anti dependence: X is written after read in MEM takes place. */ | |
2642 | ||
2643 | int | |
52d07779 | 2644 | anti_dependence (const_rtx mem, const_rtx x) |
6a0934dd | 2645 | { |
cb456db5 | 2646 | return write_dependence_p (mem, x, VOIDmode, NULL_RTX, |
2647 | /*mem_canonicalized=*/false, | |
2648 | /*x_canonicalized*/false, /*writep=*/false); | |
ddba76b8 | 2649 | } |
2650 | ||
cb456db5 | 2651 | /* Likewise, but we already have a canonicalized MEM, and X_ADDR for X. |
2652 | Also, consider X in X_MODE (which might be from an enclosing | |
2653 | STRICT_LOW_PART / ZERO_EXTRACT). | |
2654 | If MEM_CANONICALIZED is true, MEM is canonicalized. */ | |
ddba76b8 | 2655 | |
2656 | int | |
cb456db5 | 2657 | canon_anti_dependence (const_rtx mem, bool mem_canonicalized, |
2658 | const_rtx x, enum machine_mode x_mode, rtx x_addr) | |
ddba76b8 | 2659 | { |
cb456db5 | 2660 | return write_dependence_p (mem, x, x_mode, x_addr, |
2661 | mem_canonicalized, /*x_canonicalized=*/true, | |
2662 | /*writep=*/false); | |
ea0cb7ae | 2663 | } |
2664 | ||
2665 | /* Output dependence: X is written after store in MEM takes place. */ | |
2666 | ||
2667 | int | |
52d07779 | 2668 | output_dependence (const_rtx mem, const_rtx x) |
ea0cb7ae | 2669 | { |
cb456db5 | 2670 | return write_dependence_p (mem, x, VOIDmode, NULL_RTX, |
2671 | /*mem_canonicalized=*/false, | |
2672 | /*x_canonicalized*/false, /*writep=*/true); | |
ea0cb7ae | 2673 | } |
c352cbc7 | 2674 | \f |
35d3eedd | 2675 | |
a84256aa | 2676 | |
2677 | /* Check whether X may be aliased with MEM. Don't do offset-based | |
2678 | memory disambiguation & TBAA. */ | |
2679 | int | |
2680 | may_alias_p (const_rtx mem, const_rtx x) | |
2681 | { | |
2682 | rtx x_addr, mem_addr; | |
a84256aa | 2683 | |
2684 | if (MEM_VOLATILE_P (x) && MEM_VOLATILE_P (mem)) | |
2685 | return 1; | |
2686 | ||
557f718e | 2687 | /* (mem:BLK (scratch)) is a special mechanism to conflict with everything. |
2688 | This is used in epilogue deallocation functions. */ | |
2689 | if (GET_MODE (x) == BLKmode && GET_CODE (XEXP (x, 0)) == SCRATCH) | |
2690 | return 1; | |
2691 | if (GET_MODE (mem) == BLKmode && GET_CODE (XEXP (mem, 0)) == SCRATCH) | |
a84256aa | 2692 | return 1; |
a84256aa | 2693 | if (MEM_ALIAS_SET (x) == ALIAS_SET_MEMORY_BARRIER |
2694 | || MEM_ALIAS_SET (mem) == ALIAS_SET_MEMORY_BARRIER) | |
2695 | return 1; | |
2696 | ||
2697 | /* Read-only memory is by definition never modified, and therefore can't | |
2698 | conflict with anything. We don't expect to find read-only set on MEM, | |
2699 | but stupid user tricks can produce them, so don't die. */ | |
2700 | if (MEM_READONLY_P (x)) | |
2701 | return 0; | |
2702 | ||
04ec15fa | 2703 | /* If we have MEMs referring to different address spaces (which can |
a84256aa | 2704 | potentially overlap), we cannot easily tell from the addresses |
2705 | whether the references overlap. */ | |
2706 | if (MEM_ADDR_SPACE (mem) != MEM_ADDR_SPACE (x)) | |
2707 | return 1; | |
2708 | ||
2709 | x_addr = XEXP (x, 0); | |
2710 | mem_addr = XEXP (mem, 0); | |
2711 | if (!((GET_CODE (x_addr) == VALUE | |
2712 | && GET_CODE (mem_addr) != VALUE | |
2713 | && reg_mentioned_p (x_addr, mem_addr)) | |
2714 | || (GET_CODE (x_addr) != VALUE | |
2715 | && GET_CODE (mem_addr) == VALUE | |
2716 | && reg_mentioned_p (mem_addr, x_addr)))) | |
2717 | { | |
2718 | x_addr = get_addr (x_addr); | |
2719 | mem_addr = get_addr (mem_addr); | |
2720 | } | |
2721 | ||
5abc99e4 | 2722 | rtx x_base = find_base_term (x_addr); |
2723 | rtx mem_base = find_base_term (mem_addr); | |
2724 | if (! base_alias_check (x_addr, x_base, mem_addr, mem_base, | |
2725 | GET_MODE (x), GET_MODE (mem_addr))) | |
a84256aa | 2726 | return 0; |
2727 | ||
2728 | x_addr = canon_rtx (x_addr); | |
2729 | mem_addr = canon_rtx (mem_addr); | |
2730 | ||
2731 | if (nonoverlapping_memrefs_p (mem, x, true)) | |
2732 | return 0; | |
2733 | ||
a84256aa | 2734 | /* TBAA not valid for loop_invarint */ |
2735 | return rtx_refs_may_alias_p (x, mem, false); | |
2736 | } | |
2737 | ||
35d3eedd | 2738 | void |
6d8b68a3 | 2739 | init_alias_target (void) |
35d3eedd | 2740 | { |
19cb6b50 | 2741 | int i; |
35d3eedd | 2742 | |
4e56ceb1 | 2743 | if (!arg_base_value) |
2744 | arg_base_value = gen_rtx_ADDRESS (VOIDmode, 0); | |
2745 | ||
6d8b68a3 | 2746 | memset (static_reg_base_value, 0, sizeof static_reg_base_value); |
2747 | ||
35d3eedd | 2748 | for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) |
2749 | /* Check whether this register can hold an incoming pointer | |
2750 | argument. FUNCTION_ARG_REGNO_P tests outgoing register | |
2c0e001b | 2751 | numbers, so translate if necessary due to register windows. */ |
35d3eedd | 2752 | if (FUNCTION_ARG_REGNO_P (OUTGOING_REGNO (i)) |
2753 | && HARD_REGNO_MODE_OK (i, Pmode)) | |
4e56ceb1 | 2754 | static_reg_base_value[i] = arg_base_value; |
2755 | ||
df3e5a67 | 2756 | static_reg_base_value[STACK_POINTER_REGNUM] |
2757 | = unique_base_value (UNIQUE_BASE_VALUE_SP); | |
2758 | static_reg_base_value[ARG_POINTER_REGNUM] | |
2759 | = unique_base_value (UNIQUE_BASE_VALUE_ARGP); | |
2760 | static_reg_base_value[FRAME_POINTER_REGNUM] | |
2761 | = unique_base_value (UNIQUE_BASE_VALUE_FP); | |
5ae82d58 | 2762 | #if !HARD_FRAME_POINTER_IS_FRAME_POINTER |
df3e5a67 | 2763 | static_reg_base_value[HARD_FRAME_POINTER_REGNUM] |
2764 | = unique_base_value (UNIQUE_BASE_VALUE_HFP); | |
0c7f5242 | 2765 | #endif |
2766 | } | |
2767 | ||
c7bf7428 | 2768 | /* Set MEMORY_MODIFIED when X modifies DATA (that is assumed |
2769 | to be memory reference. */ | |
2770 | static bool memory_modified; | |
2771 | static void | |
a80f1c6c | 2772 | memory_modified_1 (rtx x, const_rtx pat ATTRIBUTE_UNUSED, void *data) |
c7bf7428 | 2773 | { |
e16ceb8e | 2774 | if (MEM_P (x)) |
c7bf7428 | 2775 | { |
5493cb9a | 2776 | if (anti_dependence (x, (const_rtx)data) || output_dependence (x, (const_rtx)data)) |
c7bf7428 | 2777 | memory_modified = true; |
2778 | } | |
2779 | } | |
2780 | ||
2781 | ||
2782 | /* Return true when INSN possibly modify memory contents of MEM | |
0c6d8c36 | 2783 | (i.e. address can be modified). */ |
c7bf7428 | 2784 | bool |
5493cb9a | 2785 | memory_modified_in_insn_p (const_rtx mem, const_rtx insn) |
c7bf7428 | 2786 | { |
2787 | if (!INSN_P (insn)) | |
2788 | return false; | |
2789 | memory_modified = false; | |
a80f1c6c | 2790 | note_stores (PATTERN (insn), memory_modified_1, CONST_CAST_RTX(mem)); |
c7bf7428 | 2791 | return memory_modified; |
2792 | } | |
2793 | ||
43b6ba45 | 2794 | /* Return TRUE if the destination of a set is rtx identical to |
2795 | ITEM. */ | |
2796 | static inline bool | |
2797 | set_dest_equal_p (const_rtx set, const_rtx item) | |
2798 | { | |
2799 | rtx dest = SET_DEST (set); | |
2800 | return rtx_equal_p (dest, item); | |
2801 | } | |
2802 | ||
2803 | /* Like memory_modified_in_insn_p, but return TRUE if INSN will | |
2804 | *DEFINITELY* modify the memory contents of MEM. */ | |
2805 | bool | |
2806 | memory_must_be_modified_in_insn_p (const_rtx mem, const_rtx insn) | |
2807 | { | |
2808 | if (!INSN_P (insn)) | |
2809 | return false; | |
2810 | insn = PATTERN (insn); | |
2811 | if (GET_CODE (insn) == SET) | |
2812 | return set_dest_equal_p (insn, mem); | |
2813 | else if (GET_CODE (insn) == PARALLEL) | |
2814 | { | |
2815 | int i; | |
2816 | for (i = 0; i < XVECLEN (insn, 0); i++) | |
2817 | { | |
2818 | rtx sub = XVECEXP (insn, 0, i); | |
2819 | if (GET_CODE (sub) == SET | |
2820 | && set_dest_equal_p (sub, mem)) | |
2821 | return true; | |
2822 | } | |
2823 | } | |
2824 | return false; | |
2825 | } | |
2826 | ||
73f5c1e3 | 2827 | /* Initialize the aliasing machinery. Initialize the REG_KNOWN_VALUE |
2828 | array. */ | |
2829 | ||
ea0cb7ae | 2830 | void |
aecda0d6 | 2831 | init_alias_analysis (void) |
ea0cb7ae | 2832 | { |
e5b19a8c | 2833 | unsigned int maxreg = max_reg_num (); |
37b864db | 2834 | int changed, pass; |
19cb6b50 | 2835 | int i; |
2836 | unsigned int ui; | |
d6443ebe | 2837 | rtx insn, val; |
2808fda8 | 2838 | int rpo_cnt; |
2839 | int *rpo; | |
ea0cb7ae | 2840 | |
376c21d1 | 2841 | timevar_push (TV_ALIAS_ANALYSIS); |
2842 | ||
5d3b0682 | 2843 | vec_safe_grow_cleared (reg_known_value, maxreg - FIRST_PSEUDO_REGISTER); |
d6443ebe | 2844 | reg_known_equiv_p = sbitmap_alloc (maxreg - FIRST_PSEUDO_REGISTER); |
7a2f311c | 2845 | bitmap_clear (reg_known_equiv_p); |
ea0cb7ae | 2846 | |
980f3285 | 2847 | /* If we have memory allocated from the previous run, use it. */ |
e5b19a8c | 2848 | if (old_reg_base_value) |
980f3285 | 2849 | reg_base_value = old_reg_base_value; |
2850 | ||
2851 | if (reg_base_value) | |
f1f41a6c | 2852 | reg_base_value->truncate (0); |
980f3285 | 2853 | |
f1f41a6c | 2854 | vec_safe_grow_cleared (reg_base_value, maxreg); |
c2e94b25 | 2855 | |
4c36ffe6 | 2856 | new_reg_base_value = XNEWVEC (rtx, maxreg); |
7bf60644 | 2857 | reg_seen = sbitmap_alloc (maxreg); |
06c5c535 | 2858 | |
2859 | /* The basic idea is that each pass through this loop will use the | |
2860 | "constant" information from the previous pass to propagate alias | |
2861 | information through another level of assignments. | |
2862 | ||
2808fda8 | 2863 | The propagation is done on the CFG in reverse post-order, to propagate |
2864 | things forward as far as possible in each iteration. | |
2865 | ||
06c5c535 | 2866 | This could get expensive if the assignment chains are long. Maybe |
2867 | we should throttle the number of iterations, possibly based on | |
35d3eedd | 2868 | the optimization level or flag_expensive_optimizations. |
06c5c535 | 2869 | |
2870 | We could propagate more information in the first pass by making use | |
3072d30e | 2871 | of DF_REG_DEF_COUNT to determine immediately that the alias information |
37b864db | 2872 | for a pseudo is "constant". |
2873 | ||
2874 | A program with an uninitialized variable can cause an infinite loop | |
2875 | here. Instead of doing a full dataflow analysis to detect such problems | |
2876 | we just cap the number of iterations for the loop. | |
2877 | ||
2878 | The state of the arrays for the set chain in question does not matter | |
2879 | since the program has undefined behavior. */ | |
35d3eedd | 2880 | |
a28770e1 | 2881 | rpo = XNEWVEC (int, n_basic_blocks_for_fn (cfun)); |
2808fda8 | 2882 | rpo_cnt = pre_and_rev_post_order_compute (NULL, rpo, false); |
2883 | ||
37b864db | 2884 | pass = 0; |
35d3eedd | 2885 | do |
06c5c535 | 2886 | { |
2887 | /* Assume nothing will change this iteration of the loop. */ | |
2888 | changed = 0; | |
2889 | ||
06c5c535 | 2890 | /* We want to assign the same IDs each iteration of this loop, so |
4e56ceb1 | 2891 | start counting from one each iteration of the loop. */ |
2892 | unique_id = 1; | |
06c5c535 | 2893 | |
35a3065a | 2894 | /* We're at the start of the function each iteration through the |
06c5c535 | 2895 | loop, so we're copying arguments. */ |
e5be0cc6 | 2896 | copying_arguments = true; |
ea0cb7ae | 2897 | |
35d3eedd | 2898 | /* Wipe the potential alias information clean for this pass. */ |
e5b19a8c | 2899 | memset (new_reg_base_value, 0, maxreg * sizeof (rtx)); |
f0ad0992 | 2900 | |
35d3eedd | 2901 | /* Wipe the reg_seen array clean. */ |
53c5d9d4 | 2902 | bitmap_clear (reg_seen); |
ea0cb7ae | 2903 | |
a15c38cd | 2904 | /* Initialize the alias information for this pass. */ |
2905 | for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) | |
2906 | if (static_reg_base_value[i]) | |
2907 | { | |
2908 | new_reg_base_value[i] = static_reg_base_value[i]; | |
2909 | bitmap_set_bit (reg_seen, i); | |
2910 | } | |
35d3eedd | 2911 | |
06c5c535 | 2912 | /* Walk the insns adding values to the new_reg_base_value array. */ |
2808fda8 | 2913 | for (i = 0; i < rpo_cnt; i++) |
ea0cb7ae | 2914 | { |
f5a6b05f | 2915 | basic_block bb = BASIC_BLOCK_FOR_FN (cfun, rpo[i]); |
2808fda8 | 2916 | FOR_BB_INSNS (bb, insn) |
06c5c535 | 2917 | { |
2808fda8 | 2918 | if (NONDEBUG_INSN_P (insn)) |
2919 | { | |
2920 | rtx note, set; | |
d6cb6164 | 2921 | |
2922 | #if defined (HAVE_prologue) || defined (HAVE_epilogue) | |
2808fda8 | 2923 | /* The prologue/epilogue insns are not threaded onto the |
2924 | insn chain until after reload has completed. Thus, | |
2925 | there is no sense wasting time checking if INSN is in | |
2926 | the prologue/epilogue until after reload has completed. */ | |
2927 | if (reload_completed | |
2928 | && prologue_epilogue_contains (insn)) | |
2929 | continue; | |
d6cb6164 | 2930 | #endif |
2931 | ||
2808fda8 | 2932 | /* If this insn has a noalias note, process it, Otherwise, |
2933 | scan for sets. A simple set will have no side effects | |
2934 | which could change the base value of any other register. */ | |
35d3eedd | 2935 | |
2808fda8 | 2936 | if (GET_CODE (PATTERN (insn)) == SET |
2937 | && REG_NOTES (insn) != 0 | |
2938 | && find_reg_note (insn, REG_NOALIAS, NULL_RTX)) | |
2939 | record_set (SET_DEST (PATTERN (insn)), NULL_RTX, NULL); | |
2940 | else | |
2941 | note_stores (PATTERN (insn), record_set, NULL); | |
35d3eedd | 2942 | |
2808fda8 | 2943 | set = single_set (insn); |
35d3eedd | 2944 | |
2808fda8 | 2945 | if (set != 0 |
2946 | && REG_P (SET_DEST (set)) | |
2947 | && REGNO (SET_DEST (set)) >= FIRST_PSEUDO_REGISTER) | |
970c0c06 | 2948 | { |
2808fda8 | 2949 | unsigned int regno = REGNO (SET_DEST (set)); |
2950 | rtx src = SET_SRC (set); | |
2951 | rtx t; | |
2952 | ||
2953 | note = find_reg_equal_equiv_note (insn); | |
2954 | if (note && REG_NOTE_KIND (note) == REG_EQUAL | |
2955 | && DF_REG_DEF_COUNT (regno) != 1) | |
2956 | note = NULL_RTX; | |
2957 | ||
2958 | if (note != NULL_RTX | |
2959 | && GET_CODE (XEXP (note, 0)) != EXPR_LIST | |
2960 | && ! rtx_varies_p (XEXP (note, 0), 1) | |
2961 | && ! reg_overlap_mentioned_p (SET_DEST (set), | |
2962 | XEXP (note, 0))) | |
2963 | { | |
2964 | set_reg_known_value (regno, XEXP (note, 0)); | |
2965 | set_reg_known_equiv_p (regno, | |
2966 | REG_NOTE_KIND (note) == REG_EQUIV); | |
2967 | } | |
2968 | else if (DF_REG_DEF_COUNT (regno) == 1 | |
2969 | && GET_CODE (src) == PLUS | |
2970 | && REG_P (XEXP (src, 0)) | |
2971 | && (t = get_reg_known_value (REGNO (XEXP (src, 0)))) | |
2972 | && CONST_INT_P (XEXP (src, 1))) | |
2973 | { | |
2974 | t = plus_constant (GET_MODE (src), t, | |
2975 | INTVAL (XEXP (src, 1))); | |
2976 | set_reg_known_value (regno, t); | |
2977 | set_reg_known_equiv_p (regno, false); | |
2978 | } | |
2979 | else if (DF_REG_DEF_COUNT (regno) == 1 | |
2980 | && ! rtx_varies_p (src, 1)) | |
2981 | { | |
2982 | set_reg_known_value (regno, src); | |
2983 | set_reg_known_equiv_p (regno, false); | |
2984 | } | |
970c0c06 | 2985 | } |
35d3eedd | 2986 | } |
2808fda8 | 2987 | else if (NOTE_P (insn) |
2988 | && NOTE_KIND (insn) == NOTE_INSN_FUNCTION_BEG) | |
2989 | copying_arguments = false; | |
06c5c535 | 2990 | } |
35d3eedd | 2991 | } |
06c5c535 | 2992 | |
35d3eedd | 2993 | /* Now propagate values from new_reg_base_value to reg_base_value. */ |
d4473c84 | 2994 | gcc_assert (maxreg == (unsigned int) max_reg_num ()); |
a0c938f0 | 2995 | |
e5b19a8c | 2996 | for (ui = 0; ui < maxreg; ui++) |
35d3eedd | 2997 | { |
274c11d8 | 2998 | if (new_reg_base_value[ui] |
f1f41a6c | 2999 | && new_reg_base_value[ui] != (*reg_base_value)[ui] |
3000 | && ! rtx_equal_p (new_reg_base_value[ui], (*reg_base_value)[ui])) | |
06c5c535 | 3001 | { |
f1f41a6c | 3002 | (*reg_base_value)[ui] = new_reg_base_value[ui]; |
35d3eedd | 3003 | changed = 1; |
06c5c535 | 3004 | } |
ea0cb7ae | 3005 | } |
ea0cb7ae | 3006 | } |
35d3eedd | 3007 | while (changed && ++pass < MAX_ALIAS_LOOP_PASSES); |
2808fda8 | 3008 | XDELETEVEC (rpo); |
ea0cb7ae | 3009 | |
3010 | /* Fill in the remaining entries. */ | |
f1f41a6c | 3011 | FOR_EACH_VEC_ELT (*reg_known_value, i, val) |
d6443ebe | 3012 | { |
3013 | int regno = i + FIRST_PSEUDO_REGISTER; | |
3014 | if (! val) | |
3015 | set_reg_known_value (regno, regno_reg_rtx[regno]); | |
3016 | } | |
ea0cb7ae | 3017 | |
7098ef4a | 3018 | /* Clean up. */ |
3019 | free (new_reg_base_value); | |
06c5c535 | 3020 | new_reg_base_value = 0; |
7bf60644 | 3021 | sbitmap_free (reg_seen); |
ea0cb7ae | 3022 | reg_seen = 0; |
376c21d1 | 3023 | timevar_pop (TV_ALIAS_ANALYSIS); |
ea0cb7ae | 3024 | } |
3025 | ||
1837e966 | 3026 | /* Equate REG_BASE_VALUE (reg1) to REG_BASE_VALUE (reg2). |
3027 | Special API for var-tracking pass purposes. */ | |
3028 | ||
3029 | void | |
3030 | vt_equate_reg_base_value (const_rtx reg1, const_rtx reg2) | |
3031 | { | |
f1f41a6c | 3032 | (*reg_base_value)[REGNO (reg1)] = REG_BASE_VALUE (reg2); |
1837e966 | 3033 | } |
3034 | ||
ea0cb7ae | 3035 | void |
aecda0d6 | 3036 | end_alias_analysis (void) |
ea0cb7ae | 3037 | { |
e5b19a8c | 3038 | old_reg_base_value = reg_base_value; |
f1f41a6c | 3039 | vec_free (reg_known_value); |
d6443ebe | 3040 | sbitmap_free (reg_known_equiv_p); |
ea0cb7ae | 3041 | } |
1f3233d1 | 3042 | |
3043 | #include "gt-alias.h" |