]>
Commit | Line | Data |
---|---|---|
ea0cb7ae | 1 | /* Alias analysis for GNU C |
d353bf18 | 2 | Copyright (C) 1997-2015 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" |
b20a8bb4 | 26 | #include "hash-set.h" |
b20a8bb4 | 27 | #include "vec.h" |
b20a8bb4 | 28 | #include "input.h" |
29 | #include "alias.h" | |
30 | #include "symtab.h" | |
b20a8bb4 | 31 | #include "inchash.h" |
142131ed | 32 | #include "tree.h" |
b20a8bb4 | 33 | #include "fold-const.h" |
9ed99284 | 34 | #include "varasm.h" |
a3020f2f | 35 | #include "hashtab.h" |
a3020f2f | 36 | #include "hard-reg-set.h" |
0a893c29 | 37 | #include "function.h" |
d53441c8 | 38 | #include "flags.h" |
39 | #include "statistics.h" | |
d53441c8 | 40 | #include "insn-config.h" |
41 | #include "expmed.h" | |
42 | #include "dojump.h" | |
43 | #include "explow.h" | |
44 | #include "calls.h" | |
4c74e6d9 | 45 | #include "emit-rtl.h" |
d53441c8 | 46 | #include "stmt.h" |
47 | #include "expr.h" | |
48 | #include "tm_p.h" | |
ea0cb7ae | 49 | #include "regs.h" |
0b205f4c | 50 | #include "diagnostic-core.h" |
f7d27fdc | 51 | #include "alloc-pool.h" |
1617c276 | 52 | #include "cselib.h" |
d4786b13 | 53 | #include "hash-map.h" |
b0278d39 | 54 | #include "langhooks.h" |
376c21d1 | 55 | #include "timevar.h" |
b9ed1410 | 56 | #include "dumpfile.h" |
4f8e9d5c | 57 | #include "target.h" |
94ea8568 | 58 | #include "dominance.h" |
59 | #include "cfg.h" | |
60 | #include "cfganal.h" | |
61 | #include "predict.h" | |
62 | #include "basic-block.h" | |
3072d30e | 63 | #include "df.h" |
3a443843 | 64 | #include "tree-ssa-alias.h" |
bc61cadb | 65 | #include "internal-fn.h" |
66 | #include "gimple-expr.h" | |
67 | #include "is-a.h" | |
073c1fd5 | 68 | #include "gimple.h" |
69 | #include "gimple-ssa.h" | |
794f1ac3 | 70 | #include "rtl-iter.h" |
f7d118a9 | 71 | |
72 | /* The aliasing API provided here solves related but different problems: | |
73 | ||
a0c938f0 | 74 | Say there exists (in c) |
f7d118a9 | 75 | |
76 | struct X { | |
77 | struct Y y1; | |
78 | struct Z z2; | |
79 | } x1, *px1, *px2; | |
80 | ||
81 | struct Y y2, *py; | |
82 | struct Z z2, *pz; | |
83 | ||
84 | ||
3c62ebbd | 85 | py = &x1.y1; |
f7d118a9 | 86 | px2 = &x1; |
87 | ||
88 | Consider the four questions: | |
89 | ||
90 | Can a store to x1 interfere with px2->y1? | |
91 | Can a store to x1 interfere with px2->z2? | |
f7d118a9 | 92 | Can a store to x1 change the value pointed to by with py? |
93 | Can a store to x1 change the value pointed to by with pz? | |
94 | ||
95 | The answer to these questions can be yes, yes, yes, and maybe. | |
96 | ||
97 | The first two questions can be answered with a simple examination | |
98 | of the type system. If structure X contains a field of type Y then | |
9d75589a | 99 | a store through a pointer to an X can overwrite any field that is |
f7d118a9 | 100 | contained (recursively) in an X (unless we know that px1 != px2). |
101 | ||
3c62ebbd | 102 | The last two questions can be solved in the same way as the first |
103 | two questions but this is too conservative. The observation is | |
104 | that in some cases we can know which (if any) fields are addressed | |
105 | and if those addresses are used in bad ways. This analysis may be | |
106 | language specific. In C, arbitrary operations may be applied to | |
107 | pointers. However, there is some indication that this may be too | |
108 | conservative for some C++ types. | |
f7d118a9 | 109 | |
110 | The pass ipa-type-escape does this analysis for the types whose | |
a0c938f0 | 111 | instances do not escape across the compilation boundary. |
f7d118a9 | 112 | |
113 | Historically in GCC, these two problems were combined and a single | |
3c62ebbd | 114 | data structure that was used to represent the solution to these |
f7d118a9 | 115 | problems. We now have two similar but different data structures, |
3c62ebbd | 116 | The data structure to solve the last two questions is similar to |
117 | the first, but does not contain the fields whose address are never | |
118 | taken. For types that do escape the compilation unit, the data | |
119 | structures will have identical information. | |
f7d118a9 | 120 | */ |
a5b1863e | 121 | |
122 | /* The alias sets assigned to MEMs assist the back-end in determining | |
123 | which MEMs can alias which other MEMs. In general, two MEMs in | |
5badbbc0 | 124 | different alias sets cannot alias each other, with one important |
125 | exception. Consider something like: | |
a5b1863e | 126 | |
8e2ef041 | 127 | struct S { int i; double d; }; |
a5b1863e | 128 | |
129 | a store to an `S' can alias something of either type `int' or type | |
130 | `double'. (However, a store to an `int' cannot alias a `double' | |
131 | and vice versa.) We indicate this via a tree structure that looks | |
132 | like: | |
a0c938f0 | 133 | struct S |
134 | / \ | |
a5b1863e | 135 | / \ |
a0c938f0 | 136 | |/_ _\| |
137 | int double | |
a5b1863e | 138 | |
5badbbc0 | 139 | (The arrows are directed and point downwards.) |
140 | In this situation we say the alias set for `struct S' is the | |
141 | `superset' and that those for `int' and `double' are `subsets'. | |
142 | ||
f7c44134 | 143 | To see whether two alias sets can point to the same memory, we must |
144 | see if either alias set is a subset of the other. We need not trace | |
d716ce75 | 145 | past immediate descendants, however, since we propagate all |
f7c44134 | 146 | grandchildren up one level. |
a5b1863e | 147 | |
148 | Alias set zero is implicitly a superset of all other alias sets. | |
149 | However, this is no actual entry for alias set zero. It is an | |
150 | error to attempt to explicitly construct a subset of zero. */ | |
151 | ||
d4786b13 | 152 | struct alias_set_traits : default_hashmap_traits |
153 | { | |
154 | template<typename T> | |
155 | static bool | |
156 | is_empty (T &e) | |
157 | { | |
158 | return e.m_key == INT_MIN; | |
159 | } | |
160 | ||
161 | template<typename T> | |
162 | static bool | |
163 | is_deleted (T &e) | |
164 | { | |
165 | return e.m_key == (INT_MIN + 1); | |
166 | } | |
167 | ||
168 | template<typename T> static void mark_empty (T &e) { e.m_key = INT_MIN; } | |
169 | ||
170 | template<typename T> | |
171 | static void | |
172 | mark_deleted (T &e) | |
173 | { | |
174 | e.m_key = INT_MIN + 1; | |
175 | } | |
176 | }; | |
177 | ||
26dbec0a | 178 | struct GTY(()) alias_set_entry_d { |
a5b1863e | 179 | /* The alias set number, as stored in MEM_ALIAS_SET. */ |
32c2fdea | 180 | alias_set_type alias_set; |
a5b1863e | 181 | |
182 | /* The children of the alias set. These are not just the immediate | |
d716ce75 | 183 | children, but, in fact, all descendants. So, if we have: |
a5b1863e | 184 | |
9342ee68 | 185 | struct T { struct S s; float f; } |
a5b1863e | 186 | |
187 | continuing our example above, the children here will be all of | |
188 | `int', `double', `float', and `struct S'. */ | |
d4786b13 | 189 | hash_map<int, int, alias_set_traits> *children; |
fdc195e1 | 190 | |
191 | /* Nonzero if would have a child of zero: this effectively makes this | |
192 | alias set the same as alias set zero. */ | |
193 | bool has_zero_child; | |
194 | /* Nonzero if alias set corresponds to pointer type itself (i.e. not to | |
195 | aggregate contaiing pointer. | |
196 | This is used for a special case where we need an universal pointer type | |
197 | compatible with all other pointer types. */ | |
198 | bool is_pointer; | |
199 | /* Nonzero if is_pointer or if one of childs have has_pointer set. */ | |
200 | bool has_pointer; | |
318dcdd8 | 201 | }; |
26dbec0a | 202 | typedef struct alias_set_entry_d *alias_set_entry; |
ea0cb7ae | 203 | |
7ecb5bb2 | 204 | static int rtx_equal_for_memref_p (const_rtx, const_rtx); |
aecda0d6 | 205 | static int memrefs_conflict_p (int, rtx, int, rtx, HOST_WIDE_INT); |
81a410b1 | 206 | static void record_set (rtx, const_rtx, void *); |
3754d046 | 207 | static int base_alias_check (rtx, rtx, rtx, rtx, machine_mode, |
208 | machine_mode); | |
aecda0d6 | 209 | static rtx find_base_value (rtx); |
52d07779 | 210 | static int mems_in_disjoint_alias_sets_p (const_rtx, const_rtx); |
32c2fdea | 211 | static alias_set_entry get_alias_set_entry (alias_set_type); |
aecda0d6 | 212 | static tree decl_for_component_ref (tree); |
cb456db5 | 213 | static int write_dependence_p (const_rtx, |
3754d046 | 214 | const_rtx, machine_mode, rtx, |
cb456db5 | 215 | bool, bool, bool); |
aecda0d6 | 216 | |
a80f1c6c | 217 | static void memory_modified_1 (rtx, const_rtx, void *); |
ea0cb7ae | 218 | |
5b6bf098 | 219 | /* Query statistics for the different low-level disambiguators. |
220 | A high-level query may trigger multiple of them. */ | |
221 | ||
222 | static struct { | |
223 | unsigned long long num_alias_zero; | |
224 | unsigned long long num_same_alias_set; | |
225 | unsigned long long num_same_objects; | |
226 | unsigned long long num_volatile; | |
227 | unsigned long long num_dag; | |
fdc195e1 | 228 | unsigned long long num_universal; |
5b6bf098 | 229 | unsigned long long num_disambiguated; |
230 | } alias_stats; | |
231 | ||
232 | ||
ea0cb7ae | 233 | /* Set up all info needed to perform alias analysis on memory references. */ |
234 | ||
083a2b5e | 235 | /* Returns the size in bytes of the mode of X. */ |
ea0cb7ae | 236 | #define SIZE_FOR_MODE(X) (GET_MODE_SIZE (GET_MODE (X))) |
237 | ||
37b864db | 238 | /* Cap the number of passes we make over the insns propagating alias |
2808fda8 | 239 | information through set chains. |
240 | ??? 10 is a completely arbitrary choice. This should be based on the | |
241 | maximum loop depth in the CFG, but we do not have this information | |
242 | available (even if current_loops _is_ available). */ | |
37b864db | 243 | #define MAX_ALIAS_LOOP_PASSES 10 |
9342ee68 | 244 | |
ea0cb7ae | 245 | /* reg_base_value[N] gives an address to which register N is related. |
246 | If all sets after the first add or subtract to the current value | |
247 | or otherwise modify it so it does not point to a different top level | |
248 | object, reg_base_value[N] is equal to the address part of the source | |
e7c893ba | 249 | of the first set. |
250 | ||
251 | A base address can be an ADDRESS, SYMBOL_REF, or LABEL_REF. ADDRESS | |
4e56ceb1 | 252 | expressions represent three types of base: |
3a935f06 | 253 | |
4e56ceb1 | 254 | 1. incoming arguments. There is just one ADDRESS to represent all |
255 | arguments, since we do not know at this level whether accesses | |
256 | based on different arguments can alias. The ADDRESS has id 0. | |
3a935f06 | 257 | |
4e56ceb1 | 258 | 2. stack_pointer_rtx, frame_pointer_rtx, hard_frame_pointer_rtx |
259 | (if distinct from frame_pointer_rtx) and arg_pointer_rtx. | |
260 | Each of these rtxes has a separate ADDRESS associated with it, | |
261 | each with a negative id. | |
262 | ||
263 | GCC is (and is required to be) precise in which register it | |
264 | chooses to access a particular region of stack. We can therefore | |
265 | assume that accesses based on one of these rtxes do not alias | |
266 | accesses based on another of these rtxes. | |
267 | ||
268 | 3. bases that are derived from malloc()ed memory (REG_NOALIAS). | |
269 | Each such piece of memory has a separate ADDRESS associated | |
270 | with it, each with an id greater than 0. | |
271 | ||
272 | Accesses based on one ADDRESS do not alias accesses based on other | |
273 | ADDRESSes. Accesses based on ADDRESSes in groups (2) and (3) do not | |
274 | alias globals either; the ADDRESSes have Pmode to indicate this. | |
275 | The ADDRESS in group (1) _may_ alias globals; it has VOIDmode to | |
276 | indicate this. */ | |
e7c893ba | 277 | |
f1f41a6c | 278 | static GTY(()) vec<rtx, va_gc> *reg_base_value; |
c2e94b25 | 279 | static rtx *new_reg_base_value; |
e5b19a8c | 280 | |
4e56ceb1 | 281 | /* The single VOIDmode ADDRESS that represents all argument bases. |
282 | It has id 0. */ | |
283 | static GTY(()) rtx arg_base_value; | |
284 | ||
285 | /* Used to allocate unique ids to each REG_NOALIAS ADDRESS. */ | |
286 | static int unique_id; | |
287 | ||
e5b19a8c | 288 | /* We preserve the copy of old array around to avoid amount of garbage |
289 | produced. About 8% of garbage produced were attributed to this | |
290 | array. */ | |
f1f41a6c | 291 | static GTY((deletable)) vec<rtx, va_gc> *old_reg_base_value; |
083a2b5e | 292 | |
86e87ef6 | 293 | /* Values of XINT (address, 0) of Pmode ADDRESS rtxes for special |
294 | registers. */ | |
295 | #define UNIQUE_BASE_VALUE_SP -1 | |
296 | #define UNIQUE_BASE_VALUE_ARGP -2 | |
297 | #define UNIQUE_BASE_VALUE_FP -3 | |
298 | #define UNIQUE_BASE_VALUE_HFP -4 | |
299 | ||
eb2c25b5 | 300 | #define static_reg_base_value \ |
301 | (this_target_rtl->x_static_reg_base_value) | |
0c7f5242 | 302 | |
f1f41a6c | 303 | #define REG_BASE_VALUE(X) \ |
304 | (REGNO (X) < vec_safe_length (reg_base_value) \ | |
305 | ? (*reg_base_value)[REGNO (X)] : 0) | |
ea0cb7ae | 306 | |
73f5c1e3 | 307 | /* Vector indexed by N giving the initial (unchanging) value known for |
d6443ebe | 308 | pseudo-register N. This vector is initialized in init_alias_analysis, |
12cb06ad | 309 | and does not change until end_alias_analysis is called. */ |
f1f41a6c | 310 | static GTY(()) vec<rtx, va_gc> *reg_known_value; |
ea0cb7ae | 311 | |
312 | /* Vector recording for each reg_known_value whether it is due to a | |
313 | REG_EQUIV note. Future passes (viz., reload) may replace the | |
314 | pseudo with the equivalent expression and so we account for the | |
5badbbc0 | 315 | dependences that would be introduced if that happens. |
316 | ||
317 | The REG_EQUIV notes created in assign_parms may mention the arg | |
318 | pointer, and there are explicit insns in the RTL that modify the | |
319 | arg pointer. Thus we must ensure that such insns don't get | |
320 | scheduled across each other because that would invalidate the | |
321 | REG_EQUIV notes. One could argue that the REG_EQUIV notes are | |
322 | wrong, but solving the problem in the scheduler will likely give | |
323 | better code, so we do it here. */ | |
d6443ebe | 324 | static sbitmap reg_known_equiv_p; |
ea0cb7ae | 325 | |
e7c893ba | 326 | /* True when scanning insns from the start of the rtl to the |
327 | NOTE_INSN_FUNCTION_BEG note. */ | |
e5be0cc6 | 328 | static bool copying_arguments; |
ea0cb7ae | 329 | |
8b435339 | 330 | |
a5b1863e | 331 | /* The splay-tree used to store the various alias set entries. */ |
f1f41a6c | 332 | static GTY (()) vec<alias_set_entry, va_gc> *alias_sets; |
5badbbc0 | 333 | \f |
3a443843 | 334 | /* Build a decomposed reference object for querying the alias-oracle |
335 | from the MEM rtx and store it in *REF. | |
336 | Returns false if MEM is not suitable for the alias-oracle. */ | |
337 | ||
338 | static bool | |
339 | ao_ref_from_mem (ao_ref *ref, const_rtx mem) | |
340 | { | |
341 | tree expr = MEM_EXPR (mem); | |
342 | tree base; | |
343 | ||
344 | if (!expr) | |
345 | return false; | |
346 | ||
347 | ao_ref_init (ref, expr); | |
348 | ||
349 | /* Get the base of the reference and see if we have to reject or | |
350 | adjust it. */ | |
351 | base = ao_ref_base (ref); | |
352 | if (base == NULL_TREE) | |
353 | return false; | |
354 | ||
9bbb37b8 | 355 | /* The tree oracle doesn't like bases that are neither decls |
356 | nor indirect references of SSA names. */ | |
357 | if (!(DECL_P (base) | |
358 | || (TREE_CODE (base) == MEM_REF | |
359 | && TREE_CODE (TREE_OPERAND (base, 0)) == SSA_NAME) | |
360 | || (TREE_CODE (base) == TARGET_MEM_REF | |
361 | && TREE_CODE (TMR_BASE (base)) == SSA_NAME))) | |
7e9c660e | 362 | return false; |
3a443843 | 363 | |
364 | /* If this is a reference based on a partitioned decl replace the | |
9bbb37b8 | 365 | base with a MEM_REF of the pointer representative we |
3a443843 | 366 | created during stack slot partitioning. */ |
367 | if (TREE_CODE (base) == VAR_DECL | |
9bbb37b8 | 368 | && ! is_global_var (base) |
3a443843 | 369 | && cfun->gimple_df->decls_to_pointers != NULL) |
370 | { | |
5f8841a5 | 371 | tree *namep = cfun->gimple_df->decls_to_pointers->get (base); |
3a443843 | 372 | if (namep) |
5f8841a5 | 373 | ref->base = build_simple_mem_ref (*namep); |
7e9c660e | 374 | } |
3a443843 | 375 | |
376 | ref->ref_alias_set = MEM_ALIAS_SET (mem); | |
377 | ||
cbc80ce5 | 378 | /* If MEM_OFFSET or MEM_SIZE are unknown what we got from MEM_EXPR |
379 | is conservative, so trust it. */ | |
da443c27 | 380 | if (!MEM_OFFSET_KNOWN_P (mem) |
5b2a69fa | 381 | || !MEM_SIZE_KNOWN_P (mem)) |
cbc80ce5 | 382 | return true; |
1f1e1860 | 383 | |
3804203d | 384 | /* If the base decl is a parameter we can have negative MEM_OFFSET in |
385 | case of promoted subregs on bigendian targets. Trust the MEM_EXPR | |
386 | here. */ | |
da443c27 | 387 | if (MEM_OFFSET (mem) < 0 |
388 | && (MEM_SIZE (mem) + MEM_OFFSET (mem)) * BITS_PER_UNIT == ref->size) | |
3804203d | 389 | return true; |
390 | ||
cbc80ce5 | 391 | /* Otherwise continue and refine size and offset we got from analyzing |
392 | MEM_EXPR by using MEM_SIZE and MEM_OFFSET. */ | |
393 | ||
da443c27 | 394 | ref->offset += MEM_OFFSET (mem) * BITS_PER_UNIT; |
5b2a69fa | 395 | ref->size = MEM_SIZE (mem) * BITS_PER_UNIT; |
3804203d | 396 | |
397 | /* The MEM may extend into adjacent fields, so adjust max_size if | |
398 | necessary. */ | |
399 | if (ref->max_size != -1 | |
400 | && ref->size > ref->max_size) | |
401 | ref->max_size = ref->size; | |
402 | ||
403 | /* If MEM_OFFSET and MEM_SIZE get us outside of the base object of | |
404 | the MEM_EXPR punt. This happens for STRICT_ALIGNMENT targets a lot. */ | |
405 | if (MEM_EXPR (mem) != get_spill_slot_decl (false) | |
406 | && (ref->offset < 0 | |
407 | || (DECL_P (ref->base) | |
e913b5cd | 408 | && (DECL_SIZE (ref->base) == NULL_TREE |
409 | || TREE_CODE (DECL_SIZE (ref->base)) != INTEGER_CST | |
f9e2a585 | 410 | || wi::ltu_p (wi::to_offset (DECL_SIZE (ref->base)), |
796b6678 | 411 | ref->offset + ref->size))))) |
3804203d | 412 | return false; |
3a443843 | 413 | |
414 | return true; | |
415 | } | |
416 | ||
417 | /* Query the alias-oracle on whether the two memory rtx X and MEM may | |
418 | alias. If TBAA_P is set also apply TBAA. Returns true if the | |
419 | two rtxen may alias, false otherwise. */ | |
420 | ||
421 | static bool | |
422 | rtx_refs_may_alias_p (const_rtx x, const_rtx mem, bool tbaa_p) | |
423 | { | |
424 | ao_ref ref1, ref2; | |
425 | ||
426 | if (!ao_ref_from_mem (&ref1, x) | |
427 | || !ao_ref_from_mem (&ref2, mem)) | |
428 | return true; | |
429 | ||
33dda198 | 430 | return refs_may_alias_p_1 (&ref1, &ref2, |
431 | tbaa_p | |
432 | && MEM_ALIAS_SET (x) != 0 | |
433 | && MEM_ALIAS_SET (mem) != 0); | |
3a443843 | 434 | } |
435 | ||
a5b1863e | 436 | /* Returns a pointer to the alias set entry for ALIAS_SET, if there is |
437 | such an entry, or NULL otherwise. */ | |
438 | ||
56bbdce4 | 439 | static inline alias_set_entry |
32c2fdea | 440 | get_alias_set_entry (alias_set_type alias_set) |
a5b1863e | 441 | { |
f1f41a6c | 442 | return (*alias_sets)[alias_set]; |
a5b1863e | 443 | } |
444 | ||
5badbbc0 | 445 | /* Returns nonzero if the alias sets for MEM1 and MEM2 are such that |
446 | the two MEMs cannot alias each other. */ | |
a5b1863e | 447 | |
56bbdce4 | 448 | static inline int |
52d07779 | 449 | mems_in_disjoint_alias_sets_p (const_rtx mem1, const_rtx mem2) |
a5b1863e | 450 | { |
832f8005 | 451 | return (flag_strict_aliasing |
452 | && ! alias_sets_conflict_p (MEM_ALIAS_SET (mem1), | |
453 | MEM_ALIAS_SET (mem2))); | |
387bc205 | 454 | } |
a5b1863e | 455 | |
7d1f52b2 | 456 | /* Return true if the first alias set is a subset of the second. */ |
457 | ||
458 | bool | |
32c2fdea | 459 | alias_set_subset_of (alias_set_type set1, alias_set_type set2) |
7d1f52b2 | 460 | { |
fdc195e1 | 461 | alias_set_entry ase2; |
7d1f52b2 | 462 | |
463 | /* Everything is a subset of the "aliases everything" set. */ | |
464 | if (set2 == 0) | |
465 | return true; | |
466 | ||
fdc195e1 | 467 | /* Check if set1 is a subset of set2. */ |
468 | ase2 = get_alias_set_entry (set2); | |
469 | if (ase2 != 0 | |
470 | && (ase2->has_zero_child | |
471 | || (ase2->children && ase2->children->get (set1)))) | |
7d1f52b2 | 472 | return true; |
fdc195e1 | 473 | |
474 | /* As a special case we consider alias set of "void *" to be both subset | |
475 | and superset of every alias set of a pointer. This extra symmetry does | |
476 | not matter for alias_sets_conflict_p but it makes aliasing_component_refs_p | |
477 | to return true on the following testcase: | |
478 | ||
479 | void *ptr; | |
480 | char **ptr2=(char **)&ptr; | |
481 | *ptr2 = ... | |
482 | ||
483 | Additionally if a set contains universal pointer, we consider every pointer | |
484 | to be a subset of it, but we do not represent this explicitely - doing so | |
485 | would require us to update transitive closure each time we introduce new | |
486 | pointer type. This makes aliasing_component_refs_p to return true | |
487 | on the following testcase: | |
488 | ||
489 | struct a {void *ptr;} | |
490 | char **ptr = (char **)&a.ptr; | |
491 | ptr = ... | |
492 | ||
493 | This makes void * truly universal pointer type. See pointer handling in | |
494 | get_alias_set for more details. */ | |
495 | if (ase2 && ase2->has_pointer) | |
496 | { | |
497 | alias_set_entry ase1 = get_alias_set_entry (set1); | |
498 | ||
499 | if (ase1 && ase1->is_pointer) | |
500 | { | |
501 | alias_set_type voidptr_set = TYPE_ALIAS_SET (ptr_type_node); | |
502 | /* If one is ptr_type_node and other is pointer, then we consider | |
503 | them subset of each other. */ | |
504 | if (set1 == voidptr_set || set2 == voidptr_set) | |
505 | return true; | |
506 | /* If SET2 contains universal pointer's alias set, then we consdier | |
507 | every (non-universal) pointer. */ | |
508 | if (ase2->children && set1 != voidptr_set | |
509 | && ase2->children->get (voidptr_set)) | |
510 | return true; | |
511 | } | |
512 | } | |
7d1f52b2 | 513 | return false; |
514 | } | |
515 | ||
387bc205 | 516 | /* Return 1 if the two specified alias sets may conflict. */ |
517 | ||
518 | int | |
32c2fdea | 519 | alias_sets_conflict_p (alias_set_type set1, alias_set_type set2) |
387bc205 | 520 | { |
fdc195e1 | 521 | alias_set_entry ase1; |
522 | alias_set_entry ase2; | |
387bc205 | 523 | |
879f881c | 524 | /* The easy case. */ |
525 | if (alias_sets_must_conflict_p (set1, set2)) | |
387bc205 | 526 | return 1; |
a5b1863e | 527 | |
f7c44134 | 528 | /* See if the first alias set is a subset of the second. */ |
fdc195e1 | 529 | ase1 = get_alias_set_entry (set1); |
530 | if (ase1 != 0 | |
531 | && ase1->children && ase1->children->get (set2)) | |
5b6bf098 | 532 | { |
533 | ++alias_stats.num_dag; | |
534 | return 1; | |
535 | } | |
a5b1863e | 536 | |
537 | /* Now do the same, but with the alias sets reversed. */ | |
fdc195e1 | 538 | ase2 = get_alias_set_entry (set2); |
539 | if (ase2 != 0 | |
540 | && ase2->children && ase2->children->get (set1)) | |
5b6bf098 | 541 | { |
542 | ++alias_stats.num_dag; | |
543 | return 1; | |
544 | } | |
fdc195e1 | 545 | |
546 | /* We want void * to be compatible with any other pointer without | |
547 | really dropping it to alias set 0. Doing so would make it | |
548 | compatible with all non-pointer types too. | |
549 | ||
550 | This is not strictly necessary by the C/C++ language | |
551 | standards, but avoids common type punning mistakes. In | |
552 | addition to that, we need the existence of such universal | |
553 | pointer to implement Fortran's C_PTR type (which is defined as | |
554 | type compatible with all C pointers). */ | |
555 | if (ase1 && ase2 && ase1->has_pointer && ase2->has_pointer) | |
556 | { | |
557 | alias_set_type voidptr_set = TYPE_ALIAS_SET (ptr_type_node); | |
558 | ||
559 | /* If one of the sets corresponds to universal pointer, | |
560 | we consider it to conflict with anything that is | |
561 | or contains pointer. */ | |
562 | if (set1 == voidptr_set || set2 == voidptr_set) | |
563 | { | |
564 | ++alias_stats.num_universal; | |
565 | return true; | |
566 | } | |
567 | /* If one of sets is (non-universal) pointer and the other | |
568 | contains universal pointer, we also get conflict. */ | |
569 | if (ase1->is_pointer && set2 != voidptr_set | |
570 | && ase2->children && ase2->children->get (voidptr_set)) | |
571 | { | |
572 | ++alias_stats.num_universal; | |
573 | return true; | |
574 | } | |
575 | if (ase2->is_pointer && set1 != voidptr_set | |
576 | && ase1->children && ase1->children->get (voidptr_set)) | |
577 | { | |
578 | ++alias_stats.num_universal; | |
579 | return true; | |
580 | } | |
581 | } | |
582 | ||
5b6bf098 | 583 | ++alias_stats.num_disambiguated; |
a5b1863e | 584 | |
387bc205 | 585 | /* The two alias sets are distinct and neither one is the |
879f881c | 586 | child of the other. Therefore, they cannot conflict. */ |
387bc205 | 587 | return 0; |
a5b1863e | 588 | } |
c1628b55 | 589 | |
879f881c | 590 | /* Return 1 if the two specified alias sets will always conflict. */ |
c1628b55 | 591 | |
592 | int | |
32c2fdea | 593 | alias_sets_must_conflict_p (alias_set_type set1, alias_set_type set2) |
c1628b55 | 594 | { |
5b6bf098 | 595 | if (set1 == 0 || set2 == 0) |
596 | { | |
597 | ++alias_stats.num_alias_zero; | |
598 | return 1; | |
599 | } | |
600 | if (set1 == set2) | |
601 | { | |
602 | ++alias_stats.num_same_alias_set; | |
603 | return 1; | |
604 | } | |
c1628b55 | 605 | |
606 | return 0; | |
607 | } | |
608 | ||
387bc205 | 609 | /* Return 1 if any MEM object of type T1 will always conflict (using the |
610 | dependency routines in this file) with any MEM object of type T2. | |
611 | This is used when allocating temporary storage. If T1 and/or T2 are | |
612 | NULL_TREE, it means we know nothing about the storage. */ | |
613 | ||
614 | int | |
aecda0d6 | 615 | objects_must_conflict_p (tree t1, tree t2) |
387bc205 | 616 | { |
32c2fdea | 617 | alias_set_type set1, set2; |
2bedf233 | 618 | |
218eb8d0 | 619 | /* If neither has a type specified, we don't know if they'll conflict |
620 | because we may be using them to store objects of various types, for | |
621 | example the argument and local variables areas of inlined functions. */ | |
8f2ba9c9 | 622 | if (t1 == 0 && t2 == 0) |
218eb8d0 | 623 | return 0; |
624 | ||
387bc205 | 625 | /* If they are the same type, they must conflict. */ |
5b6bf098 | 626 | if (t1 == t2) |
627 | { | |
628 | ++alias_stats.num_same_objects; | |
629 | return 1; | |
630 | } | |
631 | /* Likewise if both are volatile. */ | |
632 | if (t1 != 0 && TYPE_VOLATILE (t1) && t2 != 0 && TYPE_VOLATILE (t2)) | |
633 | { | |
634 | ++alias_stats.num_volatile; | |
635 | return 1; | |
636 | } | |
387bc205 | 637 | |
2bedf233 | 638 | set1 = t1 ? get_alias_set (t1) : 0; |
639 | set2 = t2 ? get_alias_set (t2) : 0; | |
387bc205 | 640 | |
879f881c | 641 | /* We can't use alias_sets_conflict_p because we must make sure |
642 | that every subtype of t1 will conflict with every subtype of | |
2bedf233 | 643 | t2 for which a pair of subobjects of these respective subtypes |
644 | overlaps on the stack. */ | |
879f881c | 645 | return alias_sets_must_conflict_p (set1, set2); |
387bc205 | 646 | } |
647 | \f | |
d400f5e1 | 648 | /* Return the outermost parent of component present in the chain of |
649 | component references handled by get_inner_reference in T with the | |
650 | following property: | |
651 | - the component is non-addressable, or | |
652 | - the parent has alias set zero, | |
653 | or NULL_TREE if no such parent exists. In the former cases, the alias | |
654 | set of this parent is the alias set that must be used for T itself. */ | |
655 | ||
656 | tree | |
657 | component_uses_parent_alias_set_from (const_tree t) | |
bd32979f | 658 | { |
d400f5e1 | 659 | const_tree found = NULL_TREE; |
1f9b622b | 660 | |
d400f5e1 | 661 | while (handled_component_p (t)) |
662 | { | |
1f9b622b | 663 | switch (TREE_CODE (t)) |
664 | { | |
665 | case COMPONENT_REF: | |
666 | if (DECL_NONADDRESSABLE_P (TREE_OPERAND (t, 1))) | |
d400f5e1 | 667 | found = t; |
1f9b622b | 668 | break; |
669 | ||
670 | case ARRAY_REF: | |
671 | case ARRAY_RANGE_REF: | |
672 | if (TYPE_NONALIASED_COMPONENT (TREE_TYPE (TREE_OPERAND (t, 0)))) | |
d400f5e1 | 673 | found = t; |
1f9b622b | 674 | break; |
675 | ||
676 | case REALPART_EXPR: | |
677 | case IMAGPART_EXPR: | |
678 | break; | |
679 | ||
d400f5e1 | 680 | case BIT_FIELD_REF: |
681 | case VIEW_CONVERT_EXPR: | |
1f9b622b | 682 | /* Bitfields and casts are never addressable. */ |
d400f5e1 | 683 | found = t; |
684 | break; | |
685 | ||
686 | default: | |
687 | gcc_unreachable (); | |
1f9b622b | 688 | } |
689 | ||
d400f5e1 | 690 | if (get_alias_set (TREE_TYPE (TREE_OPERAND (t, 0))) == 0) |
691 | found = t; | |
692 | ||
1f9b622b | 693 | t = TREE_OPERAND (t, 0); |
694 | } | |
d400f5e1 | 695 | |
696 | if (found) | |
697 | return TREE_OPERAND (found, 0); | |
698 | ||
699 | return NULL_TREE; | |
bd32979f | 700 | } |
701 | ||
bd93e4f8 | 702 | |
703 | /* Return whether the pointer-type T effective for aliasing may | |
704 | access everything and thus the reference has to be assigned | |
705 | alias-set zero. */ | |
706 | ||
707 | static bool | |
708 | ref_all_alias_ptr_type_p (const_tree t) | |
709 | { | |
710 | return (TREE_CODE (TREE_TYPE (t)) == VOID_TYPE | |
711 | || TYPE_REF_CAN_ALIAS_ALL (t)); | |
bd32979f | 712 | } |
713 | ||
dd277d48 | 714 | /* Return the alias set for the memory pointed to by T, which may be |
715 | either a type or an expression. Return -1 if there is nothing | |
716 | special about dereferencing T. */ | |
717 | ||
718 | static alias_set_type | |
719 | get_deref_alias_set_1 (tree t) | |
720 | { | |
8115f0af | 721 | /* All we care about is the type. */ |
dd277d48 | 722 | if (! TYPE_P (t)) |
8115f0af | 723 | t = TREE_TYPE (t); |
dd277d48 | 724 | |
725 | /* If we have an INDIRECT_REF via a void pointer, we don't | |
726 | know anything about what that might alias. Likewise if the | |
727 | pointer is marked that way. */ | |
bd93e4f8 | 728 | if (ref_all_alias_ptr_type_p (t)) |
dd277d48 | 729 | return 0; |
730 | ||
731 | return -1; | |
732 | } | |
733 | ||
734 | /* Return the alias set for the memory pointed to by T, which may be | |
735 | either a type or an expression. */ | |
736 | ||
737 | alias_set_type | |
738 | get_deref_alias_set (tree t) | |
739 | { | |
bd93e4f8 | 740 | /* If we're not doing any alias analysis, just assume everything |
741 | aliases everything else. */ | |
742 | if (!flag_strict_aliasing) | |
743 | return 0; | |
744 | ||
dd277d48 | 745 | alias_set_type set = get_deref_alias_set_1 (t); |
746 | ||
747 | /* Fall back to the alias-set of the pointed-to type. */ | |
748 | if (set == -1) | |
749 | { | |
750 | if (! TYPE_P (t)) | |
751 | t = TREE_TYPE (t); | |
752 | set = get_alias_set (TREE_TYPE (t)); | |
753 | } | |
754 | ||
755 | return set; | |
756 | } | |
757 | ||
bd93e4f8 | 758 | /* Return the pointer-type relevant for TBAA purposes from the |
759 | memory reference tree *T or NULL_TREE in which case *T is | |
760 | adjusted to point to the outermost component reference that | |
761 | can be used for assigning an alias set. */ | |
762 | ||
763 | static tree | |
764 | reference_alias_ptr_type_1 (tree *t) | |
765 | { | |
766 | tree inner; | |
767 | ||
768 | /* Get the base object of the reference. */ | |
769 | inner = *t; | |
770 | while (handled_component_p (inner)) | |
771 | { | |
772 | /* If there is a VIEW_CONVERT_EXPR in the chain we cannot use | |
773 | the type of any component references that wrap it to | |
774 | determine the alias-set. */ | |
775 | if (TREE_CODE (inner) == VIEW_CONVERT_EXPR) | |
776 | *t = TREE_OPERAND (inner, 0); | |
777 | inner = TREE_OPERAND (inner, 0); | |
778 | } | |
779 | ||
780 | /* Handle pointer dereferences here, they can override the | |
781 | alias-set. */ | |
782 | if (INDIRECT_REF_P (inner) | |
783 | && ref_all_alias_ptr_type_p (TREE_TYPE (TREE_OPERAND (inner, 0)))) | |
784 | return TREE_TYPE (TREE_OPERAND (inner, 0)); | |
785 | else if (TREE_CODE (inner) == TARGET_MEM_REF) | |
786 | return TREE_TYPE (TMR_OFFSET (inner)); | |
787 | else if (TREE_CODE (inner) == MEM_REF | |
788 | && ref_all_alias_ptr_type_p (TREE_TYPE (TREE_OPERAND (inner, 1)))) | |
789 | return TREE_TYPE (TREE_OPERAND (inner, 1)); | |
790 | ||
791 | /* If the innermost reference is a MEM_REF that has a | |
792 | conversion embedded treat it like a VIEW_CONVERT_EXPR above, | |
793 | using the memory access type for determining the alias-set. */ | |
794 | if (TREE_CODE (inner) == MEM_REF | |
795 | && (TYPE_MAIN_VARIANT (TREE_TYPE (inner)) | |
796 | != TYPE_MAIN_VARIANT | |
797 | (TREE_TYPE (TREE_TYPE (TREE_OPERAND (inner, 1)))))) | |
798 | return TREE_TYPE (TREE_OPERAND (inner, 1)); | |
799 | ||
d400f5e1 | 800 | /* Otherwise, pick up the outermost object that we could have |
801 | a pointer to. */ | |
802 | tree tem = component_uses_parent_alias_set_from (*t); | |
803 | if (tem) | |
804 | *t = tem; | |
bd93e4f8 | 805 | |
806 | return NULL_TREE; | |
807 | } | |
808 | ||
809 | /* Return the pointer-type relevant for TBAA purposes from the | |
810 | gimple memory reference tree T. This is the type to be used for | |
811 | the offset operand of MEM_REF or TARGET_MEM_REF replacements of T | |
812 | and guarantees that get_alias_set will return the same alias | |
813 | set for T and the replacement. */ | |
814 | ||
815 | tree | |
816 | reference_alias_ptr_type (tree t) | |
817 | { | |
818 | tree ptype = reference_alias_ptr_type_1 (&t); | |
819 | /* If there is a given pointer type for aliasing purposes, return it. */ | |
820 | if (ptype != NULL_TREE) | |
821 | return ptype; | |
822 | ||
823 | /* Otherwise build one from the outermost component reference we | |
824 | may use. */ | |
825 | if (TREE_CODE (t) == MEM_REF | |
826 | || TREE_CODE (t) == TARGET_MEM_REF) | |
827 | return TREE_TYPE (TREE_OPERAND (t, 1)); | |
828 | else | |
829 | return build_pointer_type (TYPE_MAIN_VARIANT (TREE_TYPE (t))); | |
830 | } | |
831 | ||
832 | /* Return whether the pointer-types T1 and T2 used to determine | |
833 | two alias sets of two references will yield the same answer | |
834 | from get_deref_alias_set. */ | |
835 | ||
836 | bool | |
837 | alias_ptr_types_compatible_p (tree t1, tree t2) | |
838 | { | |
839 | if (TYPE_MAIN_VARIANT (t1) == TYPE_MAIN_VARIANT (t2)) | |
840 | return true; | |
841 | ||
842 | if (ref_all_alias_ptr_type_p (t1) | |
843 | || ref_all_alias_ptr_type_p (t2)) | |
844 | return false; | |
845 | ||
846 | return (TYPE_MAIN_VARIANT (TREE_TYPE (t1)) | |
847 | == TYPE_MAIN_VARIANT (TREE_TYPE (t2))); | |
848 | } | |
849 | ||
fdc195e1 | 850 | /* Create emptry alias set entry. */ |
851 | ||
852 | alias_set_entry | |
853 | init_alias_set_entry (alias_set_type set) | |
854 | { | |
855 | alias_set_entry ase = ggc_alloc<alias_set_entry_d> (); | |
856 | ase->alias_set = set; | |
857 | ase->children = NULL; | |
858 | ase->has_zero_child = false; | |
859 | ase->is_pointer = false; | |
860 | ase->has_pointer = false; | |
861 | gcc_checking_assert (!get_alias_set_entry (set)); | |
862 | (*alias_sets)[set] = ase; | |
863 | return ase; | |
864 | } | |
865 | ||
f7c44134 | 866 | /* Return the alias set for T, which may be either a type or an |
867 | expression. Call language-specific routine for help, if needed. */ | |
868 | ||
32c2fdea | 869 | alias_set_type |
aecda0d6 | 870 | get_alias_set (tree t) |
f7c44134 | 871 | { |
32c2fdea | 872 | alias_set_type set; |
f7c44134 | 873 | |
874 | /* If we're not doing any alias analysis, just assume everything | |
875 | aliases everything else. Also return 0 if this or its type is | |
876 | an error. */ | |
877 | if (! flag_strict_aliasing || t == error_mark_node | |
878 | || (! TYPE_P (t) | |
879 | && (TREE_TYPE (t) == 0 || TREE_TYPE (t) == error_mark_node))) | |
880 | return 0; | |
881 | ||
882 | /* We can be passed either an expression or a type. This and the | |
c3a9c149 | 883 | language-specific routine may make mutually-recursive calls to each other |
884 | to figure out what to do. At each juncture, we see if this is a tree | |
885 | that the language may need to handle specially. First handle things that | |
96216d37 | 886 | aren't types. */ |
1607663f | 887 | if (! TYPE_P (t)) |
f7c44134 | 888 | { |
182cf5a9 | 889 | /* Give the language a chance to do something with this tree |
890 | before we look at it. */ | |
2a631e19 | 891 | STRIP_NOPS (t); |
dc24ddbd | 892 | set = lang_hooks.get_alias_set (t); |
2a631e19 | 893 | if (set != -1) |
894 | return set; | |
895 | ||
bd93e4f8 | 896 | /* Get the alias pointer-type to use or the outermost object |
897 | that we could have a pointer to. */ | |
898 | tree ptype = reference_alias_ptr_type_1 (&t); | |
899 | if (ptype != NULL) | |
900 | return get_deref_alias_set (ptype); | |
1607663f | 901 | |
96216d37 | 902 | /* If we've already determined the alias set for a decl, just return |
903 | it. This is necessary for C++ anonymous unions, whose component | |
904 | variables don't look like union members (boo!). */ | |
afce1f57 | 905 | if (TREE_CODE (t) == VAR_DECL |
e16ceb8e | 906 | && DECL_RTL_SET_P (t) && MEM_P (DECL_RTL (t))) |
afce1f57 | 907 | return MEM_ALIAS_SET (DECL_RTL (t)); |
908 | ||
1607663f | 909 | /* Now all we care about is the type. */ |
910 | t = TREE_TYPE (t); | |
f7c44134 | 911 | } |
912 | ||
f7c44134 | 913 | /* Variant qualifiers don't affect the alias set, so get the main |
dfe8e806 | 914 | variant. */ |
f7c44134 | 915 | t = TYPE_MAIN_VARIANT (t); |
dfe8e806 | 916 | |
917 | /* Always use the canonical type as well. If this is a type that | |
918 | requires structural comparisons to identify compatible types | |
919 | use alias set zero. */ | |
920 | if (TYPE_STRUCTURAL_EQUALITY_P (t)) | |
e58c17e7 | 921 | { |
922 | /* Allow the language to specify another alias set for this | |
923 | type. */ | |
924 | set = lang_hooks.get_alias_set (t); | |
925 | if (set != -1) | |
926 | return set; | |
927 | return 0; | |
928 | } | |
ec7b61e0 | 929 | |
dfe8e806 | 930 | t = TYPE_CANONICAL (t); |
ec7b61e0 | 931 | |
45014c84 | 932 | /* The canonical type should not require structural equality checks. */ |
933 | gcc_checking_assert (!TYPE_STRUCTURAL_EQUALITY_P (t)); | |
dfe8e806 | 934 | |
935 | /* If this is a type with a known alias set, return it. */ | |
96216d37 | 936 | if (TYPE_ALIAS_SET_KNOWN_P (t)) |
f7c44134 | 937 | return TYPE_ALIAS_SET (t); |
938 | ||
b35a8f48 | 939 | /* We don't want to set TYPE_ALIAS_SET for incomplete types. */ |
940 | if (!COMPLETE_TYPE_P (t)) | |
941 | { | |
942 | /* For arrays with unknown size the conservative answer is the | |
943 | alias set of the element type. */ | |
944 | if (TREE_CODE (t) == ARRAY_TYPE) | |
945 | return get_alias_set (TREE_TYPE (t)); | |
946 | ||
947 | /* But return zero as a conservative answer for incomplete types. */ | |
948 | return 0; | |
949 | } | |
950 | ||
f7c44134 | 951 | /* See if the language has special handling for this type. */ |
dc24ddbd | 952 | set = lang_hooks.get_alias_set (t); |
2a631e19 | 953 | if (set != -1) |
96216d37 | 954 | return set; |
851dfbff | 955 | |
f7c44134 | 956 | /* There are no objects of FUNCTION_TYPE, so there's no point in |
957 | using up an alias set for them. (There are, of course, pointers | |
958 | and references to functions, but that's different.) */ | |
ec7b61e0 | 959 | else if (TREE_CODE (t) == FUNCTION_TYPE || TREE_CODE (t) == METHOD_TYPE) |
f7c44134 | 960 | set = 0; |
9fcc3e54 | 961 | |
962 | /* Unless the language specifies otherwise, let vector types alias | |
963 | their components. This avoids some nasty type punning issues in | |
964 | normal usage. And indeed lets vectors be treated more like an | |
965 | array slice. */ | |
966 | else if (TREE_CODE (t) == VECTOR_TYPE) | |
967 | set = get_alias_set (TREE_TYPE (t)); | |
968 | ||
62d823d0 | 969 | /* Unless the language specifies otherwise, treat array types the |
970 | same as their components. This avoids the asymmetry we get | |
971 | through recording the components. Consider accessing a | |
972 | character(kind=1) through a reference to a character(kind=1)[1:1]. | |
973 | Or consider if we want to assign integer(kind=4)[0:D.1387] and | |
974 | integer(kind=4)[4] the same alias set or not. | |
975 | Just be pragmatic here and make sure the array and its element | |
976 | type get the same alias set assigned. */ | |
ec7b61e0 | 977 | else if (TREE_CODE (t) == ARRAY_TYPE && !TYPE_NONALIASED_COMPONENT (t)) |
62d823d0 | 978 | set = get_alias_set (TREE_TYPE (t)); |
979 | ||
2af087f2 | 980 | /* From the former common C and C++ langhook implementation: |
981 | ||
982 | Unfortunately, there is no canonical form of a pointer type. | |
983 | In particular, if we have `typedef int I', then `int *', and | |
984 | `I *' are different types. So, we have to pick a canonical | |
985 | representative. We do this below. | |
986 | ||
987 | Technically, this approach is actually more conservative that | |
988 | it needs to be. In particular, `const int *' and `int *' | |
989 | should be in different alias sets, according to the C and C++ | |
990 | standard, since their types are not the same, and so, | |
991 | technically, an `int **' and `const int **' cannot point at | |
992 | the same thing. | |
993 | ||
994 | But, the standard is wrong. In particular, this code is | |
995 | legal C++: | |
996 | ||
997 | int *ip; | |
998 | int **ipp = &ip; | |
999 | const int* const* cipp = ipp; | |
1000 | And, it doesn't make sense for that to be legal unless you | |
1001 | can dereference IPP and CIPP. So, we ignore cv-qualifiers on | |
1002 | the pointed-to types. This issue has been reported to the | |
1003 | C++ committee. | |
1004 | ||
fdc195e1 | 1005 | For this reason go to canonical type of the unqalified pointer type. |
1006 | Until GCC 6 this code set all pointers sets to have alias set of | |
1007 | ptr_type_node but that is a bad idea, because it prevents disabiguations | |
1008 | in between pointers. For Firefox this accounts about 20% of all | |
1009 | disambiguations in the program. */ | |
1010 | else if (POINTER_TYPE_P (t) && t != ptr_type_node && !in_lto_p) | |
1011 | { | |
1012 | tree p; | |
1013 | auto_vec <bool, 8> reference; | |
1014 | ||
1015 | /* Unnest all pointers and references. | |
1016 | We also want to make pointer to array equivalent to pointer to its | |
1017 | element. So skip all array types, too. */ | |
1018 | for (p = t; POINTER_TYPE_P (p) | |
1019 | || (TREE_CODE (p) == ARRAY_TYPE && !TYPE_NONALIASED_COMPONENT (p)); | |
1020 | p = TREE_TYPE (p)) | |
1021 | { | |
1022 | if (TREE_CODE (p) == REFERENCE_TYPE) | |
1023 | reference.safe_push (true); | |
1024 | if (TREE_CODE (p) == POINTER_TYPE) | |
1025 | reference.safe_push (false); | |
1026 | } | |
1027 | p = TYPE_MAIN_VARIANT (p); | |
1028 | ||
1029 | /* Make void * compatible with char * and also void **. | |
1030 | Programs are commonly violating TBAA by this. | |
1031 | ||
1032 | We also make void * to conflict with every pointer | |
1033 | (see record_component_aliases) and thus it is safe it to use it for | |
1034 | pointers to types with TYPE_STRUCTURAL_EQUALITY_P. */ | |
1035 | if (TREE_CODE (p) == VOID_TYPE || TYPE_STRUCTURAL_EQUALITY_P (p)) | |
1036 | set = get_alias_set (ptr_type_node); | |
1037 | else | |
1038 | { | |
1039 | /* Rebuild pointer type from starting from canonical types using | |
1040 | unqualified pointers and references only. This way all such | |
1041 | pointers will have the same alias set and will conflict with | |
1042 | each other. | |
1043 | ||
1044 | Most of time we already have pointers or references of a given type. | |
1045 | If not we build new one just to be sure that if someone later | |
1046 | (probably only middle-end can, as we should assign all alias | |
1047 | classes only after finishing translation unit) builds the pointer | |
1048 | type, the canonical type will match. */ | |
1049 | p = TYPE_CANONICAL (p); | |
1050 | while (!reference.is_empty ()) | |
1051 | { | |
1052 | if (reference.pop ()) | |
1053 | p = build_reference_type (p); | |
1054 | else | |
1055 | p = build_pointer_type (p); | |
1056 | p = TYPE_CANONICAL (TYPE_MAIN_VARIANT (p)); | |
1057 | } | |
1058 | gcc_checking_assert (TYPE_CANONICAL (p) == p); | |
1059 | ||
1060 | /* Assign the alias set to both p and t. | |
1061 | We can not call get_alias_set (p) here as that would trigger | |
1062 | infinite recursion when p == t. In other cases it would just | |
1063 | trigger unnecesary legwork of rebuilding the pointer again. */ | |
1064 | if (TYPE_ALIAS_SET_KNOWN_P (p)) | |
1065 | set = TYPE_ALIAS_SET (p); | |
1066 | else | |
1067 | { | |
1068 | set = new_alias_set (); | |
1069 | TYPE_ALIAS_SET (p) = set; | |
1070 | } | |
1071 | } | |
1072 | } | |
1073 | /* In LTO the rules above needs to be part of canonical type machinery. | |
1074 | For now just punt. */ | |
1075 | else if (POINTER_TYPE_P (t) && t != ptr_type_node && in_lto_p) | |
45014c84 | 1076 | set = get_alias_set (ptr_type_node); |
2af087f2 | 1077 | |
ec7b61e0 | 1078 | /* Otherwise make a new alias set for this type. */ |
f7c44134 | 1079 | else |
45014c84 | 1080 | { |
1081 | /* Each canonical type gets its own alias set, so canonical types | |
1082 | shouldn't form a tree. It doesn't really matter for types | |
1083 | we handle specially above, so only check it where it possibly | |
1084 | would result in a bogus alias set. */ | |
1085 | gcc_checking_assert (TYPE_CANONICAL (t) == t); | |
1086 | ||
1087 | set = new_alias_set (); | |
1088 | } | |
f7c44134 | 1089 | |
1090 | TYPE_ALIAS_SET (t) = set; | |
851dfbff | 1091 | |
ec7b61e0 | 1092 | /* If this is an aggregate type or a complex type, we must record any |
1093 | component aliasing information. */ | |
680210f2 | 1094 | if (AGGREGATE_TYPE_P (t) || TREE_CODE (t) == COMPLEX_TYPE) |
851dfbff | 1095 | record_component_aliases (t); |
1096 | ||
fdc195e1 | 1097 | /* We treat pointer types specially in alias_set_subset_of. */ |
1098 | if (POINTER_TYPE_P (t) && set) | |
1099 | { | |
1100 | alias_set_entry ase = get_alias_set_entry (set); | |
1101 | if (!ase) | |
1102 | ase = init_alias_set_entry (set); | |
1103 | ase->is_pointer = true; | |
1104 | ase->has_pointer = true; | |
1105 | } | |
1106 | ||
f7c44134 | 1107 | return set; |
1108 | } | |
1109 | ||
1110 | /* Return a brand-new alias set. */ | |
1111 | ||
32c2fdea | 1112 | alias_set_type |
aecda0d6 | 1113 | new_alias_set (void) |
f7c44134 | 1114 | { |
f7c44134 | 1115 | if (flag_strict_aliasing) |
56bbdce4 | 1116 | { |
8b435339 | 1117 | if (alias_sets == 0) |
f1f41a6c | 1118 | vec_safe_push (alias_sets, (alias_set_entry) 0); |
1119 | vec_safe_push (alias_sets, (alias_set_entry) 0); | |
1120 | return alias_sets->length () - 1; | |
56bbdce4 | 1121 | } |
f7c44134 | 1122 | else |
1123 | return 0; | |
1124 | } | |
a5b1863e | 1125 | |
8e2ef041 | 1126 | /* Indicate that things in SUBSET can alias things in SUPERSET, but that |
1127 | not everything that aliases SUPERSET also aliases SUBSET. For example, | |
1128 | in C, a store to an `int' can alias a load of a structure containing an | |
1129 | `int', and vice versa. But it can't alias a load of a 'double' member | |
1130 | of the same structure. Here, the structure would be the SUPERSET and | |
1131 | `int' the SUBSET. This relationship is also described in the comment at | |
1132 | the beginning of this file. | |
1133 | ||
1134 | This function should be called only once per SUPERSET/SUBSET pair. | |
a5b1863e | 1135 | |
1136 | It is illegal for SUPERSET to be zero; everything is implicitly a | |
1137 | subset of alias set zero. */ | |
1138 | ||
892fdb03 | 1139 | void |
32c2fdea | 1140 | record_alias_subset (alias_set_type superset, alias_set_type subset) |
a5b1863e | 1141 | { |
1142 | alias_set_entry superset_entry; | |
1143 | alias_set_entry subset_entry; | |
1144 | ||
c3a9c149 | 1145 | /* It is possible in complex type situations for both sets to be the same, |
1146 | in which case we can ignore this operation. */ | |
1147 | if (superset == subset) | |
1148 | return; | |
1149 | ||
64db345d | 1150 | gcc_assert (superset); |
a5b1863e | 1151 | |
1152 | superset_entry = get_alias_set_entry (superset); | |
9342ee68 | 1153 | if (superset_entry == 0) |
a5b1863e | 1154 | { |
1155 | /* Create an entry for the SUPERSET, so that we have a place to | |
1156 | attach the SUBSET. */ | |
fdc195e1 | 1157 | superset_entry = init_alias_set_entry (superset); |
a5b1863e | 1158 | } |
1159 | ||
851dfbff | 1160 | if (subset == 0) |
1161 | superset_entry->has_zero_child = 1; | |
1162 | else | |
1163 | { | |
1164 | subset_entry = get_alias_set_entry (subset); | |
fdc195e1 | 1165 | if (!superset_entry->children) |
1166 | superset_entry->children | |
1167 | = hash_map<int, int, alias_set_traits>::create_ggc (64); | |
851dfbff | 1168 | /* If there is an entry for the subset, enter all of its children |
1169 | (if they are not already present) as children of the SUPERSET. */ | |
9342ee68 | 1170 | if (subset_entry) |
851dfbff | 1171 | { |
1172 | if (subset_entry->has_zero_child) | |
fdc195e1 | 1173 | superset_entry->has_zero_child = true; |
1174 | if (subset_entry->has_pointer) | |
1175 | superset_entry->has_pointer = true; | |
083a2b5e | 1176 | |
fdc195e1 | 1177 | if (subset_entry->children) |
1178 | { | |
1179 | hash_map<int, int, alias_set_traits>::iterator iter | |
1180 | = subset_entry->children->begin (); | |
1181 | for (; iter != subset_entry->children->end (); ++iter) | |
1182 | superset_entry->children->put ((*iter).first, (*iter).second); | |
1183 | } | |
851dfbff | 1184 | } |
a5b1863e | 1185 | |
851dfbff | 1186 | /* Enter the SUBSET itself as a child of the SUPERSET. */ |
d4786b13 | 1187 | superset_entry->children->put (subset, 0); |
851dfbff | 1188 | } |
a5b1863e | 1189 | } |
1190 | ||
9501a338 | 1191 | /* Record that component types of TYPE, if any, are part of that type for |
1192 | aliasing purposes. For record types, we only record component types | |
c67bf066 | 1193 | for fields that are not marked non-addressable. For array types, we |
1194 | only record the component type if it is not marked non-aliased. */ | |
9501a338 | 1195 | |
1196 | void | |
aecda0d6 | 1197 | record_component_aliases (tree type) |
9501a338 | 1198 | { |
32c2fdea | 1199 | alias_set_type superset = get_alias_set (type); |
9501a338 | 1200 | tree field; |
1201 | ||
1202 | if (superset == 0) | |
1203 | return; | |
1204 | ||
1205 | switch (TREE_CODE (type)) | |
1206 | { | |
9501a338 | 1207 | case RECORD_TYPE: |
1208 | case UNION_TYPE: | |
1209 | case QUAL_UNION_TYPE: | |
1767a056 | 1210 | for (field = TYPE_FIELDS (type); field != 0; field = DECL_CHAIN (field)) |
c67bf066 | 1211 | if (TREE_CODE (field) == FIELD_DECL && !DECL_NONADDRESSABLE_P (field)) |
851dfbff | 1212 | record_alias_subset (superset, get_alias_set (TREE_TYPE (field))); |
9501a338 | 1213 | break; |
1214 | ||
680210f2 | 1215 | case COMPLEX_TYPE: |
1216 | record_alias_subset (superset, get_alias_set (TREE_TYPE (type))); | |
1217 | break; | |
1218 | ||
62d823d0 | 1219 | /* VECTOR_TYPE and ARRAY_TYPE share the alias set with their |
1220 | element type. */ | |
1221 | ||
9501a338 | 1222 | default: |
1223 | break; | |
1224 | } | |
1225 | } | |
1226 | ||
f7c44134 | 1227 | /* Allocate an alias set for use in storing and reading from the varargs |
1228 | spill area. */ | |
1229 | ||
32c2fdea | 1230 | static GTY(()) alias_set_type varargs_set = -1; |
4ec9406e | 1231 | |
32c2fdea | 1232 | alias_set_type |
aecda0d6 | 1233 | get_varargs_alias_set (void) |
f7c44134 | 1234 | { |
fcdd3ab3 | 1235 | #if 1 |
1236 | /* We now lower VA_ARG_EXPR, and there's currently no way to attach the | |
1237 | varargs alias set to an INDIRECT_REF (FIXME!), so we can't | |
1238 | consistently use the varargs alias set for loads from the varargs | |
1239 | area. So don't use it anywhere. */ | |
1240 | return 0; | |
1241 | #else | |
4ec9406e | 1242 | if (varargs_set == -1) |
1243 | varargs_set = new_alias_set (); | |
f7c44134 | 1244 | |
4ec9406e | 1245 | return varargs_set; |
fcdd3ab3 | 1246 | #endif |
f7c44134 | 1247 | } |
1248 | ||
1249 | /* Likewise, but used for the fixed portions of the frame, e.g., register | |
1250 | save areas. */ | |
1251 | ||
32c2fdea | 1252 | static GTY(()) alias_set_type frame_set = -1; |
4ec9406e | 1253 | |
32c2fdea | 1254 | alias_set_type |
aecda0d6 | 1255 | get_frame_alias_set (void) |
f7c44134 | 1256 | { |
4ec9406e | 1257 | if (frame_set == -1) |
1258 | frame_set = new_alias_set (); | |
f7c44134 | 1259 | |
4ec9406e | 1260 | return frame_set; |
f7c44134 | 1261 | } |
1262 | ||
4e56ceb1 | 1263 | /* Create a new, unique base with id ID. */ |
1264 | ||
1265 | static rtx | |
1266 | unique_base_value (HOST_WIDE_INT id) | |
1267 | { | |
1268 | return gen_rtx_ADDRESS (Pmode, id); | |
1269 | } | |
1270 | ||
1271 | /* Return true if accesses based on any other base value cannot alias | |
1272 | those based on X. */ | |
1273 | ||
1274 | static bool | |
1275 | unique_base_value_p (rtx x) | |
1276 | { | |
1277 | return GET_CODE (x) == ADDRESS && GET_MODE (x) == Pmode; | |
1278 | } | |
1279 | ||
1280 | /* Return true if X is known to be a base value. */ | |
1281 | ||
1282 | static bool | |
1283 | known_base_value_p (rtx x) | |
1284 | { | |
1285 | switch (GET_CODE (x)) | |
1286 | { | |
1287 | case LABEL_REF: | |
1288 | case SYMBOL_REF: | |
1289 | return true; | |
1290 | ||
1291 | case ADDRESS: | |
1292 | /* Arguments may or may not be bases; we don't know for sure. */ | |
1293 | return GET_MODE (x) != VOIDmode; | |
1294 | ||
1295 | default: | |
1296 | return false; | |
1297 | } | |
1298 | } | |
1299 | ||
e7c893ba | 1300 | /* Inside SRC, the source of a SET, find a base address. */ |
1301 | ||
ea0cb7ae | 1302 | static rtx |
aecda0d6 | 1303 | find_base_value (rtx src) |
ea0cb7ae | 1304 | { |
970c0c06 | 1305 | unsigned int regno; |
d01410d3 | 1306 | |
2e984a7a | 1307 | #if defined (FIND_BASE_TERM) |
1308 | /* Try machine-dependent ways to find the base term. */ | |
1309 | src = FIND_BASE_TERM (src); | |
1310 | #endif | |
1311 | ||
ea0cb7ae | 1312 | switch (GET_CODE (src)) |
1313 | { | |
1314 | case SYMBOL_REF: | |
1315 | case LABEL_REF: | |
1316 | return src; | |
1317 | ||
1318 | case REG: | |
9db10305 | 1319 | regno = REGNO (src); |
083a2b5e | 1320 | /* At the start of a function, argument registers have known base |
e7c893ba | 1321 | values which may be lost later. Returning an ADDRESS |
1322 | expression here allows optimization based on argument values | |
1323 | even when the argument registers are used for other purposes. */ | |
970c0c06 | 1324 | if (regno < FIRST_PSEUDO_REGISTER && copying_arguments) |
1325 | return new_reg_base_value[regno]; | |
1a538f67 | 1326 | |
ebd13eec | 1327 | /* If a pseudo has a known base value, return it. Do not do this |
9464d436 | 1328 | for non-fixed hard regs since it can result in a circular |
1329 | dependency chain for registers which have values at function entry. | |
ebd13eec | 1330 | |
1331 | The test above is not sufficient because the scheduler may move | |
1332 | a copy out of an arg reg past the NOTE_INSN_FUNCTION_BEGIN. */ | |
9464d436 | 1333 | if ((regno >= FIRST_PSEUDO_REGISTER || fixed_regs[regno]) |
f1f41a6c | 1334 | && regno < vec_safe_length (reg_base_value)) |
e5be0cc6 | 1335 | { |
1336 | /* If we're inside init_alias_analysis, use new_reg_base_value | |
1337 | to reduce the number of relaxation iterations. */ | |
a303ed8c | 1338 | if (new_reg_base_value && new_reg_base_value[regno] |
3072d30e | 1339 | && DF_REG_DEF_COUNT (regno) == 1) |
e5be0cc6 | 1340 | return new_reg_base_value[regno]; |
1341 | ||
f1f41a6c | 1342 | if ((*reg_base_value)[regno]) |
1343 | return (*reg_base_value)[regno]; | |
e5be0cc6 | 1344 | } |
1a538f67 | 1345 | |
1c5cb56c | 1346 | return 0; |
ea0cb7ae | 1347 | |
1348 | case MEM: | |
1349 | /* Check for an argument passed in memory. Only record in the | |
1350 | copying-arguments block; it is too hard to track changes | |
1351 | otherwise. */ | |
1352 | if (copying_arguments | |
1353 | && (XEXP (src, 0) == arg_pointer_rtx | |
1354 | || (GET_CODE (XEXP (src, 0)) == PLUS | |
1355 | && XEXP (XEXP (src, 0), 0) == arg_pointer_rtx))) | |
4e56ceb1 | 1356 | return arg_base_value; |
ea0cb7ae | 1357 | return 0; |
1358 | ||
1359 | case CONST: | |
1360 | src = XEXP (src, 0); | |
1361 | if (GET_CODE (src) != PLUS && GET_CODE (src) != MINUS) | |
1362 | break; | |
083a2b5e | 1363 | |
2c0e001b | 1364 | /* ... fall through ... */ |
e7c893ba | 1365 | |
ea0cb7ae | 1366 | case PLUS: |
1367 | case MINUS: | |
e7c893ba | 1368 | { |
06c5c535 | 1369 | rtx temp, src_0 = XEXP (src, 0), src_1 = XEXP (src, 1); |
1370 | ||
560de2f8 | 1371 | /* If either operand is a REG that is a known pointer, then it |
1372 | is the base. */ | |
1373 | if (REG_P (src_0) && REG_POINTER (src_0)) | |
1374 | return find_base_value (src_0); | |
1375 | if (REG_P (src_1) && REG_POINTER (src_1)) | |
1376 | return find_base_value (src_1); | |
1377 | ||
06c5c535 | 1378 | /* If either operand is a REG, then see if we already have |
1379 | a known value for it. */ | |
560de2f8 | 1380 | if (REG_P (src_0)) |
06c5c535 | 1381 | { |
1382 | temp = find_base_value (src_0); | |
083a2b5e | 1383 | if (temp != 0) |
06c5c535 | 1384 | src_0 = temp; |
1385 | } | |
1386 | ||
560de2f8 | 1387 | if (REG_P (src_1)) |
06c5c535 | 1388 | { |
1389 | temp = find_base_value (src_1); | |
083a2b5e | 1390 | if (temp!= 0) |
06c5c535 | 1391 | src_1 = temp; |
1392 | } | |
e7c893ba | 1393 | |
560de2f8 | 1394 | /* If either base is named object or a special address |
1395 | (like an argument or stack reference), then use it for the | |
1396 | base term. */ | |
4e56ceb1 | 1397 | if (src_0 != 0 && known_base_value_p (src_0)) |
560de2f8 | 1398 | return src_0; |
1399 | ||
4e56ceb1 | 1400 | if (src_1 != 0 && known_base_value_p (src_1)) |
560de2f8 | 1401 | return src_1; |
1402 | ||
083a2b5e | 1403 | /* Guess which operand is the base address: |
06c5c535 | 1404 | If either operand is a symbol, then it is the base. If |
1405 | either operand is a CONST_INT, then the other is the base. */ | |
971ba038 | 1406 | if (CONST_INT_P (src_1) || CONSTANT_P (src_0)) |
e7c893ba | 1407 | return find_base_value (src_0); |
971ba038 | 1408 | else if (CONST_INT_P (src_0) || CONSTANT_P (src_1)) |
06c5c535 | 1409 | return find_base_value (src_1); |
1410 | ||
ea0cb7ae | 1411 | return 0; |
e7c893ba | 1412 | } |
1413 | ||
1414 | case LO_SUM: | |
1415 | /* The standard form is (lo_sum reg sym) so look only at the | |
1416 | second operand. */ | |
1417 | return find_base_value (XEXP (src, 1)); | |
ea0cb7ae | 1418 | |
1419 | case AND: | |
1420 | /* If the second operand is constant set the base | |
2c0e001b | 1421 | address to the first operand. */ |
971ba038 | 1422 | if (CONST_INT_P (XEXP (src, 1)) && INTVAL (XEXP (src, 1)) != 0) |
e7c893ba | 1423 | return find_base_value (XEXP (src, 0)); |
ea0cb7ae | 1424 | return 0; |
1425 | ||
43d45282 | 1426 | case TRUNCATE: |
04ec15fa | 1427 | /* As we do not know which address space the pointer is referring to, we can |
98155838 | 1428 | handle this only if the target does not support different pointer or |
1429 | address modes depending on the address space. */ | |
1430 | if (!target_default_pointer_address_modes_p ()) | |
1431 | break; | |
43d45282 | 1432 | if (GET_MODE_SIZE (GET_MODE (src)) < GET_MODE_SIZE (Pmode)) |
1433 | break; | |
1434 | /* Fall through. */ | |
ea0cb7ae | 1435 | case HIGH: |
2be64ef0 | 1436 | case PRE_INC: |
1437 | case PRE_DEC: | |
1438 | case POST_INC: | |
1439 | case POST_DEC: | |
1440 | case PRE_MODIFY: | |
1441 | case POST_MODIFY: | |
e7c893ba | 1442 | return find_base_value (XEXP (src, 0)); |
99c14947 | 1443 | |
d01410d3 | 1444 | case ZERO_EXTEND: |
1445 | case SIGN_EXTEND: /* used for NT/Alpha pointers */ | |
04ec15fa | 1446 | /* As we do not know which address space the pointer is referring to, we can |
98155838 | 1447 | handle this only if the target does not support different pointer or |
1448 | address modes depending on the address space. */ | |
1449 | if (!target_default_pointer_address_modes_p ()) | |
1450 | break; | |
1451 | ||
d01410d3 | 1452 | { |
1453 | rtx temp = find_base_value (XEXP (src, 0)); | |
1454 | ||
85d654dd | 1455 | if (temp != 0 && CONSTANT_P (temp)) |
d01410d3 | 1456 | temp = convert_memory_address (Pmode, temp); |
d01410d3 | 1457 | |
1458 | return temp; | |
1459 | } | |
1460 | ||
99c14947 | 1461 | default: |
1462 | break; | |
ea0cb7ae | 1463 | } |
1464 | ||
1465 | return 0; | |
1466 | } | |
1467 | ||
4e56ceb1 | 1468 | /* Called from init_alias_analysis indirectly through note_stores, |
1469 | or directly if DEST is a register with a REG_NOALIAS note attached. | |
1470 | SET is null in the latter case. */ | |
ea0cb7ae | 1471 | |
083a2b5e | 1472 | /* While scanning insns to find base values, reg_seen[N] is nonzero if |
ea0cb7ae | 1473 | register N has been set in this function. */ |
7bf60644 | 1474 | static sbitmap reg_seen; |
ea0cb7ae | 1475 | |
e7c893ba | 1476 | static void |
81a410b1 | 1477 | record_set (rtx dest, const_rtx set, void *data ATTRIBUTE_UNUSED) |
ea0cb7ae | 1478 | { |
19cb6b50 | 1479 | unsigned regno; |
ea0cb7ae | 1480 | rtx src; |
3721f260 | 1481 | int n; |
ea0cb7ae | 1482 | |
8ad4c111 | 1483 | if (!REG_P (dest)) |
ea0cb7ae | 1484 | return; |
1485 | ||
9db10305 | 1486 | regno = REGNO (dest); |
ea0cb7ae | 1487 | |
f1f41a6c | 1488 | gcc_checking_assert (regno < reg_base_value->length ()); |
c2e94b25 | 1489 | |
0933f1d9 | 1490 | n = REG_NREGS (dest); |
3721f260 | 1491 | if (n != 1) |
1492 | { | |
1493 | while (--n >= 0) | |
1494 | { | |
08b7917c | 1495 | bitmap_set_bit (reg_seen, regno + n); |
3721f260 | 1496 | new_reg_base_value[regno + n] = 0; |
1497 | } | |
1498 | return; | |
1499 | } | |
1500 | ||
ea0cb7ae | 1501 | if (set) |
1502 | { | |
1503 | /* A CLOBBER wipes out any old value but does not prevent a previously | |
1504 | unset register from acquiring a base address (i.e. reg_seen is not | |
1505 | set). */ | |
1506 | if (GET_CODE (set) == CLOBBER) | |
1507 | { | |
06c5c535 | 1508 | new_reg_base_value[regno] = 0; |
ea0cb7ae | 1509 | return; |
1510 | } | |
1511 | src = SET_SRC (set); | |
1512 | } | |
1513 | else | |
1514 | { | |
4e56ceb1 | 1515 | /* There's a REG_NOALIAS note against DEST. */ |
08b7917c | 1516 | if (bitmap_bit_p (reg_seen, regno)) |
ea0cb7ae | 1517 | { |
06c5c535 | 1518 | new_reg_base_value[regno] = 0; |
ea0cb7ae | 1519 | return; |
1520 | } | |
08b7917c | 1521 | bitmap_set_bit (reg_seen, regno); |
4e56ceb1 | 1522 | new_reg_base_value[regno] = unique_base_value (unique_id++); |
ea0cb7ae | 1523 | return; |
1524 | } | |
1525 | ||
ad6b1fe1 | 1526 | /* If this is not the first set of REGNO, see whether the new value |
1527 | is related to the old one. There are two cases of interest: | |
1528 | ||
1529 | (1) The register might be assigned an entirely new value | |
1530 | that has the same base term as the original set. | |
1531 | ||
1532 | (2) The set might be a simple self-modification that | |
1533 | cannot change REGNO's base value. | |
1534 | ||
1535 | If neither case holds, reject the original base value as invalid. | |
1536 | Note that the following situation is not detected: | |
1537 | ||
a0c938f0 | 1538 | extern int x, y; int *p = &x; p += (&y-&x); |
ad6b1fe1 | 1539 | |
ea0cb7ae | 1540 | ANSI C does not allow computing the difference of addresses |
1541 | of distinct top level objects. */ | |
ad6b1fe1 | 1542 | if (new_reg_base_value[regno] != 0 |
1543 | && find_base_value (src) != new_reg_base_value[regno]) | |
ea0cb7ae | 1544 | switch (GET_CODE (src)) |
1545 | { | |
e7c893ba | 1546 | case LO_SUM: |
ea0cb7ae | 1547 | case MINUS: |
1548 | if (XEXP (src, 0) != dest && XEXP (src, 1) != dest) | |
06c5c535 | 1549 | new_reg_base_value[regno] = 0; |
ea0cb7ae | 1550 | break; |
43d45282 | 1551 | case PLUS: |
1552 | /* If the value we add in the PLUS is also a valid base value, | |
1553 | this might be the actual base value, and the original value | |
1554 | an index. */ | |
1555 | { | |
1556 | rtx other = NULL_RTX; | |
1557 | ||
1558 | if (XEXP (src, 0) == dest) | |
1559 | other = XEXP (src, 1); | |
1560 | else if (XEXP (src, 1) == dest) | |
1561 | other = XEXP (src, 0); | |
1562 | ||
1563 | if (! other || find_base_value (other)) | |
1564 | new_reg_base_value[regno] = 0; | |
1565 | break; | |
1566 | } | |
ea0cb7ae | 1567 | case AND: |
971ba038 | 1568 | if (XEXP (src, 0) != dest || !CONST_INT_P (XEXP (src, 1))) |
06c5c535 | 1569 | new_reg_base_value[regno] = 0; |
ea0cb7ae | 1570 | break; |
ea0cb7ae | 1571 | default: |
06c5c535 | 1572 | new_reg_base_value[regno] = 0; |
ea0cb7ae | 1573 | break; |
1574 | } | |
1575 | /* If this is the first set of a register, record the value. */ | |
1576 | else if ((regno >= FIRST_PSEUDO_REGISTER || ! fixed_regs[regno]) | |
08b7917c | 1577 | && ! bitmap_bit_p (reg_seen, regno) && new_reg_base_value[regno] == 0) |
06c5c535 | 1578 | new_reg_base_value[regno] = find_base_value (src); |
ea0cb7ae | 1579 | |
08b7917c | 1580 | bitmap_set_bit (reg_seen, regno); |
ea0cb7ae | 1581 | } |
1582 | ||
0cfef6c5 | 1583 | /* Return REG_BASE_VALUE for REGNO. Selective scheduler uses this to avoid |
1584 | using hard registers with non-null REG_BASE_VALUE for renaming. */ | |
1585 | rtx | |
1586 | get_reg_base_value (unsigned int regno) | |
1587 | { | |
f1f41a6c | 1588 | return (*reg_base_value)[regno]; |
0cfef6c5 | 1589 | } |
1590 | ||
12cb06ad | 1591 | /* If a value is known for REGNO, return it. */ |
1592 | ||
a0c938f0 | 1593 | rtx |
12cb06ad | 1594 | get_reg_known_value (unsigned int regno) |
1595 | { | |
1596 | if (regno >= FIRST_PSEUDO_REGISTER) | |
1597 | { | |
1598 | regno -= FIRST_PSEUDO_REGISTER; | |
f1f41a6c | 1599 | if (regno < vec_safe_length (reg_known_value)) |
1600 | return (*reg_known_value)[regno]; | |
12cb06ad | 1601 | } |
1602 | return NULL; | |
9f5e1a5a | 1603 | } |
1604 | ||
12cb06ad | 1605 | /* Set it. */ |
1606 | ||
1607 | static void | |
1608 | set_reg_known_value (unsigned int regno, rtx val) | |
1609 | { | |
1610 | if (regno >= FIRST_PSEUDO_REGISTER) | |
1611 | { | |
1612 | regno -= FIRST_PSEUDO_REGISTER; | |
f1f41a6c | 1613 | if (regno < vec_safe_length (reg_known_value)) |
1614 | (*reg_known_value)[regno] = val; | |
12cb06ad | 1615 | } |
1616 | } | |
1617 | ||
1618 | /* Similarly for reg_known_equiv_p. */ | |
1619 | ||
1620 | bool | |
1621 | get_reg_known_equiv_p (unsigned int regno) | |
1622 | { | |
1623 | if (regno >= FIRST_PSEUDO_REGISTER) | |
1624 | { | |
1625 | regno -= FIRST_PSEUDO_REGISTER; | |
f1f41a6c | 1626 | if (regno < vec_safe_length (reg_known_value)) |
08b7917c | 1627 | return bitmap_bit_p (reg_known_equiv_p, regno); |
12cb06ad | 1628 | } |
1629 | return false; | |
1630 | } | |
1631 | ||
1632 | static void | |
1633 | set_reg_known_equiv_p (unsigned int regno, bool val) | |
1634 | { | |
1635 | if (regno >= FIRST_PSEUDO_REGISTER) | |
1636 | { | |
1637 | regno -= FIRST_PSEUDO_REGISTER; | |
f1f41a6c | 1638 | if (regno < vec_safe_length (reg_known_value)) |
d6443ebe | 1639 | { |
1640 | if (val) | |
08b7917c | 1641 | bitmap_set_bit (reg_known_equiv_p, regno); |
d6443ebe | 1642 | else |
08b7917c | 1643 | bitmap_clear_bit (reg_known_equiv_p, regno); |
d6443ebe | 1644 | } |
12cb06ad | 1645 | } |
1646 | } | |
1647 | ||
1648 | ||
7cfb9bcf | 1649 | /* Returns a canonical version of X, from the point of view alias |
1650 | analysis. (For example, if X is a MEM whose address is a register, | |
1651 | and the register has a known value (say a SYMBOL_REF), then a MEM | |
1652 | whose address is the SYMBOL_REF is returned.) */ | |
1653 | ||
1654 | rtx | |
aecda0d6 | 1655 | canon_rtx (rtx x) |
ea0cb7ae | 1656 | { |
1657 | /* Recursively look for equivalences. */ | |
8ad4c111 | 1658 | if (REG_P (x) && REGNO (x) >= FIRST_PSEUDO_REGISTER) |
12cb06ad | 1659 | { |
1660 | rtx t = get_reg_known_value (REGNO (x)); | |
1661 | if (t == x) | |
1662 | return x; | |
1663 | if (t) | |
1664 | return canon_rtx (t); | |
1665 | } | |
1666 | ||
1667 | if (GET_CODE (x) == PLUS) | |
ea0cb7ae | 1668 | { |
1669 | rtx x0 = canon_rtx (XEXP (x, 0)); | |
1670 | rtx x1 = canon_rtx (XEXP (x, 1)); | |
1671 | ||
1672 | if (x0 != XEXP (x, 0) || x1 != XEXP (x, 1)) | |
1673 | { | |
971ba038 | 1674 | if (CONST_INT_P (x0)) |
29c05e22 | 1675 | return plus_constant (GET_MODE (x), x1, INTVAL (x0)); |
971ba038 | 1676 | else if (CONST_INT_P (x1)) |
29c05e22 | 1677 | return plus_constant (GET_MODE (x), x0, INTVAL (x1)); |
941522d6 | 1678 | return gen_rtx_PLUS (GET_MODE (x), x0, x1); |
ea0cb7ae | 1679 | } |
1680 | } | |
083a2b5e | 1681 | |
ea0cb7ae | 1682 | /* This gives us much better alias analysis when called from |
1683 | the loop optimizer. Note we want to leave the original | |
1684 | MEM alone, but need to return the canonicalized MEM with | |
1685 | all the flags with their original values. */ | |
e16ceb8e | 1686 | else if (MEM_P (x)) |
e4e86ec5 | 1687 | x = replace_equiv_address_nv (x, canon_rtx (XEXP (x, 0))); |
083a2b5e | 1688 | |
ea0cb7ae | 1689 | return x; |
1690 | } | |
1691 | ||
1692 | /* Return 1 if X and Y are identical-looking rtx's. | |
6285298a | 1693 | Expect that X and Y has been already canonicalized. |
ea0cb7ae | 1694 | |
1695 | We use the data in reg_known_value above to see if two registers with | |
1696 | different numbers are, in fact, equivalent. */ | |
1697 | ||
1698 | static int | |
7ecb5bb2 | 1699 | rtx_equal_for_memref_p (const_rtx x, const_rtx y) |
ea0cb7ae | 1700 | { |
19cb6b50 | 1701 | int i; |
1702 | int j; | |
1703 | enum rtx_code code; | |
1704 | const char *fmt; | |
ea0cb7ae | 1705 | |
1706 | if (x == 0 && y == 0) | |
1707 | return 1; | |
1708 | if (x == 0 || y == 0) | |
1709 | return 0; | |
083a2b5e | 1710 | |
ea0cb7ae | 1711 | if (x == y) |
1712 | return 1; | |
1713 | ||
1714 | code = GET_CODE (x); | |
1715 | /* Rtx's of different codes cannot be equal. */ | |
1716 | if (code != GET_CODE (y)) | |
1717 | return 0; | |
1718 | ||
1719 | /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. | |
1720 | (REG:SI x) and (REG:HI x) are NOT equivalent. */ | |
1721 | ||
1722 | if (GET_MODE (x) != GET_MODE (y)) | |
1723 | return 0; | |
1724 | ||
7cfb9bcf | 1725 | /* Some RTL can be compared without a recursive examination. */ |
1726 | switch (code) | |
1727 | { | |
1728 | case REG: | |
1729 | return REGNO (x) == REGNO (y); | |
1730 | ||
1731 | case LABEL_REF: | |
b49f2e4b | 1732 | return LABEL_REF_LABEL (x) == LABEL_REF_LABEL (y); |
9342ee68 | 1733 | |
7cfb9bcf | 1734 | case SYMBOL_REF: |
1735 | return XSTR (x, 0) == XSTR (y, 0); | |
1736 | ||
861a6de4 | 1737 | case ENTRY_VALUE: |
1738 | /* This is magic, don't go through canonicalization et al. */ | |
1739 | return rtx_equal_p (ENTRY_VALUE_EXP (x), ENTRY_VALUE_EXP (y)); | |
1740 | ||
a2e238f7 | 1741 | case VALUE: |
0349edce | 1742 | CASE_CONST_UNIQUE: |
e913b5cd | 1743 | /* Pointer equality guarantees equality for these nodes. */ |
7cfb9bcf | 1744 | return 0; |
1745 | ||
7cfb9bcf | 1746 | default: |
1747 | break; | |
1748 | } | |
ea0cb7ae | 1749 | |
6285298a | 1750 | /* canon_rtx knows how to handle plus. No need to canonicalize. */ |
1751 | if (code == PLUS) | |
ea0cb7ae | 1752 | return ((rtx_equal_for_memref_p (XEXP (x, 0), XEXP (y, 0)) |
1753 | && rtx_equal_for_memref_p (XEXP (x, 1), XEXP (y, 1))) | |
1754 | || (rtx_equal_for_memref_p (XEXP (x, 0), XEXP (y, 1)) | |
1755 | && rtx_equal_for_memref_p (XEXP (x, 1), XEXP (y, 0)))); | |
6285298a | 1756 | /* For commutative operations, the RTX match if the operand match in any |
1757 | order. Also handle the simple binary and unary cases without a loop. */ | |
6720e96c | 1758 | if (COMMUTATIVE_P (x)) |
6285298a | 1759 | { |
1760 | rtx xop0 = canon_rtx (XEXP (x, 0)); | |
1761 | rtx yop0 = canon_rtx (XEXP (y, 0)); | |
1762 | rtx yop1 = canon_rtx (XEXP (y, 1)); | |
1763 | ||
1764 | return ((rtx_equal_for_memref_p (xop0, yop0) | |
1765 | && rtx_equal_for_memref_p (canon_rtx (XEXP (x, 1)), yop1)) | |
1766 | || (rtx_equal_for_memref_p (xop0, yop1) | |
1767 | && rtx_equal_for_memref_p (canon_rtx (XEXP (x, 1)), yop0))); | |
1768 | } | |
6720e96c | 1769 | else if (NON_COMMUTATIVE_P (x)) |
6285298a | 1770 | { |
1771 | return (rtx_equal_for_memref_p (canon_rtx (XEXP (x, 0)), | |
aecda0d6 | 1772 | canon_rtx (XEXP (y, 0))) |
6285298a | 1773 | && rtx_equal_for_memref_p (canon_rtx (XEXP (x, 1)), |
1774 | canon_rtx (XEXP (y, 1)))); | |
1775 | } | |
6720e96c | 1776 | else if (UNARY_P (x)) |
6285298a | 1777 | return rtx_equal_for_memref_p (canon_rtx (XEXP (x, 0)), |
aecda0d6 | 1778 | canon_rtx (XEXP (y, 0))); |
ea0cb7ae | 1779 | |
1780 | /* Compare the elements. If any pair of corresponding elements | |
1a6c3add | 1781 | fail to match, return 0 for the whole things. |
1782 | ||
1783 | Limit cases to types which actually appear in addresses. */ | |
ea0cb7ae | 1784 | |
1785 | fmt = GET_RTX_FORMAT (code); | |
1786 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
1787 | { | |
1788 | switch (fmt[i]) | |
1789 | { | |
ea0cb7ae | 1790 | case 'i': |
1791 | if (XINT (x, i) != XINT (y, i)) | |
1792 | return 0; | |
1793 | break; | |
1794 | ||
ea0cb7ae | 1795 | case 'E': |
1796 | /* Two vectors must have the same length. */ | |
1797 | if (XVECLEN (x, i) != XVECLEN (y, i)) | |
1798 | return 0; | |
1799 | ||
1800 | /* And the corresponding elements must match. */ | |
1801 | for (j = 0; j < XVECLEN (x, i); j++) | |
6285298a | 1802 | if (rtx_equal_for_memref_p (canon_rtx (XVECEXP (x, i, j)), |
1803 | canon_rtx (XVECEXP (y, i, j))) == 0) | |
ea0cb7ae | 1804 | return 0; |
1805 | break; | |
1806 | ||
1807 | case 'e': | |
6285298a | 1808 | if (rtx_equal_for_memref_p (canon_rtx (XEXP (x, i)), |
1809 | canon_rtx (XEXP (y, i))) == 0) | |
ea0cb7ae | 1810 | return 0; |
1811 | break; | |
1812 | ||
83ec925c | 1813 | /* This can happen for asm operands. */ |
1814 | case 's': | |
1815 | if (strcmp (XSTR (x, i), XSTR (y, i))) | |
1816 | return 0; | |
1817 | break; | |
1818 | ||
23c04f6b | 1819 | /* This can happen for an asm which clobbers memory. */ |
1820 | case '0': | |
1821 | break; | |
1822 | ||
ea0cb7ae | 1823 | /* It is believed that rtx's at this level will never |
1824 | contain anything but integers and other rtx's, | |
1825 | except for within LABEL_REFs and SYMBOL_REFs. */ | |
1826 | default: | |
64db345d | 1827 | gcc_unreachable (); |
ea0cb7ae | 1828 | } |
1829 | } | |
1830 | return 1; | |
1831 | } | |
1832 | ||
86e87ef6 | 1833 | static rtx |
aecda0d6 | 1834 | find_base_term (rtx x) |
ea0cb7ae | 1835 | { |
1617c276 | 1836 | cselib_val *val; |
8081d3a6 | 1837 | struct elt_loc_list *l, *f; |
1838 | rtx ret; | |
1617c276 | 1839 | |
15750b1e | 1840 | #if defined (FIND_BASE_TERM) |
1841 | /* Try machine-dependent ways to find the base term. */ | |
1842 | x = FIND_BASE_TERM (x); | |
1843 | #endif | |
1844 | ||
ea0cb7ae | 1845 | switch (GET_CODE (x)) |
1846 | { | |
1847 | case REG: | |
1848 | return REG_BASE_VALUE (x); | |
1849 | ||
2be64ef0 | 1850 | case TRUNCATE: |
04ec15fa | 1851 | /* As we do not know which address space the pointer is referring to, we can |
98155838 | 1852 | handle this only if the target does not support different pointer or |
1853 | address modes depending on the address space. */ | |
1854 | if (!target_default_pointer_address_modes_p ()) | |
1855 | return 0; | |
2be64ef0 | 1856 | if (GET_MODE_SIZE (GET_MODE (x)) < GET_MODE_SIZE (Pmode)) |
9342ee68 | 1857 | return 0; |
2be64ef0 | 1858 | /* Fall through. */ |
ea0cb7ae | 1859 | case HIGH: |
22dfec7d | 1860 | case PRE_INC: |
1861 | case PRE_DEC: | |
1862 | case POST_INC: | |
1863 | case POST_DEC: | |
2be64ef0 | 1864 | case PRE_MODIFY: |
1865 | case POST_MODIFY: | |
22dfec7d | 1866 | return find_base_term (XEXP (x, 0)); |
1867 | ||
5104ef69 | 1868 | case ZERO_EXTEND: |
1869 | case SIGN_EXTEND: /* Used for Alpha/NT pointers */ | |
04ec15fa | 1870 | /* As we do not know which address space the pointer is referring to, we can |
98155838 | 1871 | handle this only if the target does not support different pointer or |
1872 | address modes depending on the address space. */ | |
1873 | if (!target_default_pointer_address_modes_p ()) | |
1874 | return 0; | |
1875 | ||
5104ef69 | 1876 | { |
1877 | rtx temp = find_base_term (XEXP (x, 0)); | |
1878 | ||
85d654dd | 1879 | if (temp != 0 && CONSTANT_P (temp)) |
5104ef69 | 1880 | temp = convert_memory_address (Pmode, temp); |
5104ef69 | 1881 | |
1882 | return temp; | |
1883 | } | |
1884 | ||
1617c276 | 1885 | case VALUE: |
1886 | val = CSELIB_VAL_PTR (x); | |
8081d3a6 | 1887 | ret = NULL_RTX; |
1888 | ||
a2e238f7 | 1889 | if (!val) |
8081d3a6 | 1890 | return ret; |
1891 | ||
bb75e83f | 1892 | if (cselib_sp_based_value_p (val)) |
1893 | return static_reg_base_value[STACK_POINTER_REGNUM]; | |
1894 | ||
8081d3a6 | 1895 | f = val->locs; |
1896 | /* Temporarily reset val->locs to avoid infinite recursion. */ | |
1897 | val->locs = NULL; | |
1898 | ||
1899 | for (l = f; l; l = l->next) | |
1900 | if (GET_CODE (l->loc) == VALUE | |
1901 | && CSELIB_VAL_PTR (l->loc)->locs | |
1902 | && !CSELIB_VAL_PTR (l->loc)->locs->next | |
1903 | && CSELIB_VAL_PTR (l->loc)->locs->loc == x) | |
1904 | continue; | |
1905 | else if ((ret = find_base_term (l->loc)) != 0) | |
1906 | break; | |
1907 | ||
1908 | val->locs = f; | |
1909 | return ret; | |
1617c276 | 1910 | |
a7cf5bbf | 1911 | case LO_SUM: |
1912 | /* The standard form is (lo_sum reg sym) so look only at the | |
1913 | second operand. */ | |
1914 | return find_base_term (XEXP (x, 1)); | |
1915 | ||
ea0cb7ae | 1916 | case CONST: |
1917 | x = XEXP (x, 0); | |
1918 | if (GET_CODE (x) != PLUS && GET_CODE (x) != MINUS) | |
1919 | return 0; | |
b4b174c3 | 1920 | /* Fall through. */ |
ea0cb7ae | 1921 | case PLUS: |
1922 | case MINUS: | |
1923 | { | |
1b5bbc4c | 1924 | rtx tmp1 = XEXP (x, 0); |
1925 | rtx tmp2 = XEXP (x, 1); | |
1926 | ||
35a3065a | 1927 | /* This is a little bit tricky since we have to determine which of |
1b5bbc4c | 1928 | the two operands represents the real base address. Otherwise this |
1929 | routine may return the index register instead of the base register. | |
1930 | ||
1931 | That may cause us to believe no aliasing was possible, when in | |
1932 | fact aliasing is possible. | |
1933 | ||
1934 | We use a few simple tests to guess the base register. Additional | |
1935 | tests can certainly be added. For example, if one of the operands | |
1936 | is a shift or multiply, then it must be the index register and the | |
1937 | other operand is the base register. */ | |
9342ee68 | 1938 | |
15750b1e | 1939 | if (tmp1 == pic_offset_table_rtx && CONSTANT_P (tmp2)) |
1940 | return find_base_term (tmp2); | |
1941 | ||
5abc99e4 | 1942 | /* If either operand is known to be a pointer, then prefer it |
1b5bbc4c | 1943 | to determine the base term. */ |
e61a0a7f | 1944 | if (REG_P (tmp1) && REG_POINTER (tmp1)) |
5abc99e4 | 1945 | ; |
1946 | else if (REG_P (tmp2) && REG_POINTER (tmp2)) | |
da584ec3 | 1947 | std::swap (tmp1, tmp2); |
1948 | /* If second argument is constant which has base term, prefer it | |
1949 | over variable tmp1. See PR64025. */ | |
1950 | else if (CONSTANT_P (tmp2) && !CONST_INT_P (tmp2)) | |
1951 | std::swap (tmp1, tmp2); | |
1b5bbc4c | 1952 | |
5abc99e4 | 1953 | /* Go ahead and find the base term for both operands. If either base |
1954 | term is from a pointer or is a named object or a special address | |
1b5bbc4c | 1955 | (like an argument or stack reference), then use it for the |
1956 | base term. */ | |
ce0deccc | 1957 | rtx base = find_base_term (tmp1); |
1958 | if (base != NULL_RTX | |
5abc99e4 | 1959 | && ((REG_P (tmp1) && REG_POINTER (tmp1)) |
ce0deccc | 1960 | || known_base_value_p (base))) |
1961 | return base; | |
1962 | base = find_base_term (tmp2); | |
1963 | if (base != NULL_RTX | |
5abc99e4 | 1964 | && ((REG_P (tmp2) && REG_POINTER (tmp2)) |
ce0deccc | 1965 | || known_base_value_p (base))) |
1966 | return base; | |
1b5bbc4c | 1967 | |
1968 | /* We could not determine which of the two operands was the | |
1969 | base register and which was the index. So we can determine | |
1970 | nothing from the base alias check. */ | |
1971 | return 0; | |
ea0cb7ae | 1972 | } |
1973 | ||
1974 | case AND: | |
971ba038 | 1975 | if (CONST_INT_P (XEXP (x, 1)) && INTVAL (XEXP (x, 1)) != 0) |
2be64ef0 | 1976 | return find_base_term (XEXP (x, 0)); |
ea0cb7ae | 1977 | return 0; |
1978 | ||
1979 | case SYMBOL_REF: | |
1980 | case LABEL_REF: | |
1981 | return x; | |
1982 | ||
1983 | default: | |
1984 | return 0; | |
1985 | } | |
1986 | } | |
1987 | ||
86e87ef6 | 1988 | /* Return true if accesses to address X may alias accesses based |
1989 | on the stack pointer. */ | |
1990 | ||
1991 | bool | |
1992 | may_be_sp_based_p (rtx x) | |
1993 | { | |
1994 | rtx base = find_base_term (x); | |
1995 | return !base || base == static_reg_base_value[STACK_POINTER_REGNUM]; | |
1996 | } | |
1997 | ||
ea0cb7ae | 1998 | /* Return 0 if the addresses X and Y are known to point to different |
1999 | objects, 1 if they might be pointers to the same object. */ | |
2000 | ||
2001 | static int | |
5abc99e4 | 2002 | base_alias_check (rtx x, rtx x_base, rtx y, rtx y_base, |
3754d046 | 2003 | machine_mode x_mode, machine_mode y_mode) |
ea0cb7ae | 2004 | { |
3ad3a0ed | 2005 | /* If the address itself has no known base see if a known equivalent |
2006 | value has one. If either address still has no known base, nothing | |
2007 | is known about aliasing. */ | |
2008 | if (x_base == 0) | |
2009 | { | |
2010 | rtx x_c; | |
083a2b5e | 2011 | |
3ad3a0ed | 2012 | if (! flag_expensive_optimizations || (x_c = canon_rtx (x)) == x) |
2013 | return 1; | |
083a2b5e | 2014 | |
3ad3a0ed | 2015 | x_base = find_base_term (x_c); |
2016 | if (x_base == 0) | |
2017 | return 1; | |
2018 | } | |
ea0cb7ae | 2019 | |
3ad3a0ed | 2020 | if (y_base == 0) |
2021 | { | |
2022 | rtx y_c; | |
2023 | if (! flag_expensive_optimizations || (y_c = canon_rtx (y)) == y) | |
2024 | return 1; | |
083a2b5e | 2025 | |
3ad3a0ed | 2026 | y_base = find_base_term (y_c); |
2027 | if (y_base == 0) | |
2028 | return 1; | |
2029 | } | |
2030 | ||
2031 | /* If the base addresses are equal nothing is known about aliasing. */ | |
2032 | if (rtx_equal_p (x_base, y_base)) | |
ea0cb7ae | 2033 | return 1; |
2034 | ||
f92492f7 | 2035 | /* The base addresses are different expressions. If they are not accessed |
2036 | via AND, there is no conflict. We can bring knowledge of object | |
2037 | alignment into play here. For example, on alpha, "char a, b;" can | |
2038 | alias one another, though "char a; long b;" cannot. AND addesses may | |
2039 | implicitly alias surrounding objects; i.e. unaligned access in DImode | |
2040 | via AND address can alias all surrounding object types except those | |
2041 | with aligment 8 or higher. */ | |
2042 | if (GET_CODE (x) == AND && GET_CODE (y) == AND) | |
2043 | return 1; | |
2044 | if (GET_CODE (x) == AND | |
971ba038 | 2045 | && (!CONST_INT_P (XEXP (x, 1)) |
f92492f7 | 2046 | || (int) GET_MODE_UNIT_SIZE (y_mode) < -INTVAL (XEXP (x, 1)))) |
2047 | return 1; | |
2048 | if (GET_CODE (y) == AND | |
971ba038 | 2049 | && (!CONST_INT_P (XEXP (y, 1)) |
f92492f7 | 2050 | || (int) GET_MODE_UNIT_SIZE (x_mode) < -INTVAL (XEXP (y, 1)))) |
2051 | return 1; | |
2052 | ||
2053 | /* Differing symbols not accessed via AND never alias. */ | |
ea0cb7ae | 2054 | if (GET_CODE (x_base) != ADDRESS && GET_CODE (y_base) != ADDRESS) |
f92492f7 | 2055 | return 0; |
ea0cb7ae | 2056 | |
4e56ceb1 | 2057 | if (unique_base_value_p (x_base) || unique_base_value_p (y_base)) |
ea0cb7ae | 2058 | return 0; |
2059 | ||
e190b18a | 2060 | return 1; |
ea0cb7ae | 2061 | } |
2062 | ||
fa5e407e | 2063 | /* Return TRUE if EXPR refers to a VALUE whose uid is greater than |
2064 | that of V. */ | |
2065 | ||
2066 | static bool | |
794f1ac3 | 2067 | refs_newer_value_p (const_rtx expr, rtx v) |
fa5e407e | 2068 | { |
2069 | int minuid = CSELIB_VAL_PTR (v)->uid; | |
794f1ac3 | 2070 | subrtx_iterator::array_type array; |
2071 | FOR_EACH_SUBRTX (iter, array, expr, NONCONST) | |
2072 | if (GET_CODE (*iter) == VALUE && CSELIB_VAL_PTR (*iter)->uid > minuid) | |
2073 | return true; | |
2074 | return false; | |
fa5e407e | 2075 | } |
2076 | ||
1617c276 | 2077 | /* Convert the address X into something we can use. This is done by returning |
2078 | it unchanged unless it is a value; in the latter case we call cselib to get | |
2079 | a more useful rtx. */ | |
f7c44134 | 2080 | |
8e802be9 | 2081 | rtx |
aecda0d6 | 2082 | get_addr (rtx x) |
1617c276 | 2083 | { |
2084 | cselib_val *v; | |
2085 | struct elt_loc_list *l; | |
2086 | ||
2087 | if (GET_CODE (x) != VALUE) | |
2088 | return x; | |
2089 | v = CSELIB_VAL_PTR (x); | |
a2e238f7 | 2090 | if (v) |
2091 | { | |
85b6e75b | 2092 | bool have_equivs = cselib_have_permanent_equivalences (); |
2093 | if (have_equivs) | |
2094 | v = canonical_cselib_val (v); | |
a2e238f7 | 2095 | for (l = v->locs; l; l = l->next) |
2096 | if (CONSTANT_P (l->loc)) | |
2097 | return l->loc; | |
2098 | for (l = v->locs; l; l = l->next) | |
85b6e75b | 2099 | if (!REG_P (l->loc) && !MEM_P (l->loc) |
2100 | /* Avoid infinite recursion when potentially dealing with | |
2101 | var-tracking artificial equivalences, by skipping the | |
2102 | equivalences themselves, and not choosing expressions | |
2103 | that refer to newer VALUEs. */ | |
2104 | && (!have_equivs | |
2105 | || (GET_CODE (l->loc) != VALUE | |
2106 | && !refs_newer_value_p (l->loc, x)))) | |
fa5e407e | 2107 | return l->loc; |
85b6e75b | 2108 | if (have_equivs) |
2109 | { | |
2110 | for (l = v->locs; l; l = l->next) | |
2111 | if (REG_P (l->loc) | |
2112 | || (GET_CODE (l->loc) != VALUE | |
2113 | && !refs_newer_value_p (l->loc, x))) | |
2114 | return l->loc; | |
2115 | /* Return the canonical value. */ | |
2116 | return v->val_rtx; | |
2117 | } | |
2118 | if (v->locs) | |
2119 | return v->locs->loc; | |
a2e238f7 | 2120 | } |
1617c276 | 2121 | return x; |
2122 | } | |
2123 | ||
b9f92976 | 2124 | /* Return the address of the (N_REFS + 1)th memory reference to ADDR |
2125 | where SIZE is the size in bytes of the memory reference. If ADDR | |
2126 | is not modified by the memory reference then ADDR is returned. */ | |
2127 | ||
0aa9ffd4 | 2128 | static rtx |
aecda0d6 | 2129 | addr_side_effect_eval (rtx addr, int size, int n_refs) |
b9f92976 | 2130 | { |
2131 | int offset = 0; | |
9342ee68 | 2132 | |
b9f92976 | 2133 | switch (GET_CODE (addr)) |
2134 | { | |
2135 | case PRE_INC: | |
2136 | offset = (n_refs + 1) * size; | |
2137 | break; | |
2138 | case PRE_DEC: | |
2139 | offset = -(n_refs + 1) * size; | |
2140 | break; | |
2141 | case POST_INC: | |
2142 | offset = n_refs * size; | |
2143 | break; | |
2144 | case POST_DEC: | |
2145 | offset = -n_refs * size; | |
2146 | break; | |
2147 | ||
2148 | default: | |
2149 | return addr; | |
2150 | } | |
9342ee68 | 2151 | |
b9f92976 | 2152 | if (offset) |
6285298a | 2153 | addr = gen_rtx_PLUS (GET_MODE (addr), XEXP (addr, 0), |
c338f2e3 | 2154 | gen_int_mode (offset, GET_MODE (addr))); |
b9f92976 | 2155 | else |
2156 | addr = XEXP (addr, 0); | |
6285298a | 2157 | addr = canon_rtx (addr); |
b9f92976 | 2158 | |
2159 | return addr; | |
2160 | } | |
2161 | ||
340ad617 | 2162 | /* Return TRUE if an object X sized at XSIZE bytes and another object |
2163 | Y sized at YSIZE bytes, starting C bytes after X, may overlap. If | |
2164 | any of the sizes is zero, assume an overlap, otherwise use the | |
2165 | absolute value of the sizes as the actual sizes. */ | |
2166 | ||
2167 | static inline bool | |
2168 | offset_overlap_p (HOST_WIDE_INT c, int xsize, int ysize) | |
2169 | { | |
2170 | return (xsize == 0 || ysize == 0 | |
2171 | || (c >= 0 | |
2172 | ? (abs (xsize) > c) | |
2173 | : (abs (ysize) > -c))); | |
2174 | } | |
2175 | ||
40f9e245 | 2176 | /* Return one if X and Y (memory addresses) reference the |
2177 | same location in memory or if the references overlap. | |
2178 | Return zero if they do not overlap, else return | |
2179 | minus one in which case they still might reference the same location. | |
2180 | ||
2181 | C is an offset accumulator. When | |
ea0cb7ae | 2182 | C is nonzero, we are testing aliases between X and Y + C. |
2183 | XSIZE is the size in bytes of the X reference, | |
2184 | similarly YSIZE is the size in bytes for Y. | |
6285298a | 2185 | Expect that canon_rtx has been already called for X and Y. |
ea0cb7ae | 2186 | |
2187 | If XSIZE or YSIZE is zero, we do not know the amount of memory being | |
2188 | referenced (the reference was BLKmode), so make the most pessimistic | |
2189 | assumptions. | |
2190 | ||
1e38bf49 | 2191 | If XSIZE or YSIZE is negative, we may access memory outside the object |
2192 | being referenced as a side effect. This can happen when using AND to | |
2193 | align memory references, as is done on the Alpha. | |
2194 | ||
ea0cb7ae | 2195 | Nice to notice that varying addresses cannot conflict with fp if no |
40f9e245 | 2196 | local variables had their addresses taken, but that's too hard now. |
2197 | ||
2198 | ??? Contrary to the tree alias oracle this does not return | |
2199 | one for X + non-constant and Y + non-constant when X and Y are equal. | |
2200 | If that is fixed the TBAA hack for union type-punning can be removed. */ | |
ea0cb7ae | 2201 | |
ea0cb7ae | 2202 | static int |
aecda0d6 | 2203 | memrefs_conflict_p (int xsize, rtx x, int ysize, rtx y, HOST_WIDE_INT c) |
ea0cb7ae | 2204 | { |
1617c276 | 2205 | if (GET_CODE (x) == VALUE) |
6f9688ce | 2206 | { |
2207 | if (REG_P (y)) | |
2208 | { | |
4e15f806 | 2209 | struct elt_loc_list *l = NULL; |
2210 | if (CSELIB_VAL_PTR (x)) | |
fa5e407e | 2211 | for (l = canonical_cselib_val (CSELIB_VAL_PTR (x))->locs; |
2212 | l; l = l->next) | |
4e15f806 | 2213 | if (REG_P (l->loc) && rtx_equal_for_memref_p (l->loc, y)) |
2214 | break; | |
6f9688ce | 2215 | if (l) |
2216 | x = y; | |
2217 | else | |
2218 | x = get_addr (x); | |
2219 | } | |
2220 | /* Don't call get_addr if y is the same VALUE. */ | |
2221 | else if (x != y) | |
2222 | x = get_addr (x); | |
2223 | } | |
1617c276 | 2224 | if (GET_CODE (y) == VALUE) |
6f9688ce | 2225 | { |
2226 | if (REG_P (x)) | |
2227 | { | |
4e15f806 | 2228 | struct elt_loc_list *l = NULL; |
2229 | if (CSELIB_VAL_PTR (y)) | |
fa5e407e | 2230 | for (l = canonical_cselib_val (CSELIB_VAL_PTR (y))->locs; |
2231 | l; l = l->next) | |
4e15f806 | 2232 | if (REG_P (l->loc) && rtx_equal_for_memref_p (l->loc, x)) |
2233 | break; | |
6f9688ce | 2234 | if (l) |
2235 | y = x; | |
2236 | else | |
2237 | y = get_addr (y); | |
2238 | } | |
2239 | /* Don't call get_addr if x is the same VALUE. */ | |
2240 | else if (y != x) | |
2241 | y = get_addr (y); | |
2242 | } | |
ea0cb7ae | 2243 | if (GET_CODE (x) == HIGH) |
2244 | x = XEXP (x, 0); | |
2245 | else if (GET_CODE (x) == LO_SUM) | |
2246 | x = XEXP (x, 1); | |
2247 | else | |
340ad617 | 2248 | x = addr_side_effect_eval (x, abs (xsize), 0); |
ea0cb7ae | 2249 | if (GET_CODE (y) == HIGH) |
2250 | y = XEXP (y, 0); | |
2251 | else if (GET_CODE (y) == LO_SUM) | |
2252 | y = XEXP (y, 1); | |
2253 | else | |
340ad617 | 2254 | y = addr_side_effect_eval (y, abs (ysize), 0); |
ea0cb7ae | 2255 | |
2256 | if (rtx_equal_for_memref_p (x, y)) | |
2257 | { | |
340ad617 | 2258 | return offset_overlap_p (c, xsize, ysize); |
ea0cb7ae | 2259 | } |
2260 | ||
35d3eedd | 2261 | /* This code used to check for conflicts involving stack references and |
2262 | globals but the base address alias code now handles these cases. */ | |
ea0cb7ae | 2263 | |
2264 | if (GET_CODE (x) == PLUS) | |
2265 | { | |
2266 | /* The fact that X is canonicalized means that this | |
2267 | PLUS rtx is canonicalized. */ | |
2268 | rtx x0 = XEXP (x, 0); | |
2269 | rtx x1 = XEXP (x, 1); | |
2270 | ||
2271 | if (GET_CODE (y) == PLUS) | |
2272 | { | |
2273 | /* The fact that Y is canonicalized means that this | |
2274 | PLUS rtx is canonicalized. */ | |
2275 | rtx y0 = XEXP (y, 0); | |
2276 | rtx y1 = XEXP (y, 1); | |
2277 | ||
2278 | if (rtx_equal_for_memref_p (x1, y1)) | |
2279 | return memrefs_conflict_p (xsize, x0, ysize, y0, c); | |
2280 | if (rtx_equal_for_memref_p (x0, y0)) | |
2281 | return memrefs_conflict_p (xsize, x1, ysize, y1, c); | |
971ba038 | 2282 | if (CONST_INT_P (x1)) |
5f59c49b | 2283 | { |
971ba038 | 2284 | if (CONST_INT_P (y1)) |
5f59c49b | 2285 | return memrefs_conflict_p (xsize, x0, ysize, y0, |
2286 | c - INTVAL (x1) + INTVAL (y1)); | |
2287 | else | |
2288 | return memrefs_conflict_p (xsize, x0, ysize, y, | |
2289 | c - INTVAL (x1)); | |
2290 | } | |
971ba038 | 2291 | else if (CONST_INT_P (y1)) |
ea0cb7ae | 2292 | return memrefs_conflict_p (xsize, x, ysize, y0, c + INTVAL (y1)); |
2293 | ||
40f9e245 | 2294 | return -1; |
ea0cb7ae | 2295 | } |
971ba038 | 2296 | else if (CONST_INT_P (x1)) |
ea0cb7ae | 2297 | return memrefs_conflict_p (xsize, x0, ysize, y, c - INTVAL (x1)); |
2298 | } | |
2299 | else if (GET_CODE (y) == PLUS) | |
2300 | { | |
2301 | /* The fact that Y is canonicalized means that this | |
2302 | PLUS rtx is canonicalized. */ | |
2303 | rtx y0 = XEXP (y, 0); | |
2304 | rtx y1 = XEXP (y, 1); | |
2305 | ||
971ba038 | 2306 | if (CONST_INT_P (y1)) |
ea0cb7ae | 2307 | return memrefs_conflict_p (xsize, x, ysize, y0, c + INTVAL (y1)); |
2308 | else | |
40f9e245 | 2309 | return -1; |
ea0cb7ae | 2310 | } |
2311 | ||
2312 | if (GET_CODE (x) == GET_CODE (y)) | |
2313 | switch (GET_CODE (x)) | |
2314 | { | |
2315 | case MULT: | |
2316 | { | |
2317 | /* Handle cases where we expect the second operands to be the | |
2318 | same, and check only whether the first operand would conflict | |
2319 | or not. */ | |
2320 | rtx x0, y0; | |
2321 | rtx x1 = canon_rtx (XEXP (x, 1)); | |
2322 | rtx y1 = canon_rtx (XEXP (y, 1)); | |
2323 | if (! rtx_equal_for_memref_p (x1, y1)) | |
40f9e245 | 2324 | return -1; |
ea0cb7ae | 2325 | x0 = canon_rtx (XEXP (x, 0)); |
2326 | y0 = canon_rtx (XEXP (y, 0)); | |
2327 | if (rtx_equal_for_memref_p (x0, y0)) | |
340ad617 | 2328 | return offset_overlap_p (c, xsize, ysize); |
ea0cb7ae | 2329 | |
2330 | /* Can't properly adjust our sizes. */ | |
971ba038 | 2331 | if (!CONST_INT_P (x1)) |
40f9e245 | 2332 | return -1; |
ea0cb7ae | 2333 | xsize /= INTVAL (x1); |
2334 | ysize /= INTVAL (x1); | |
2335 | c /= INTVAL (x1); | |
2336 | return memrefs_conflict_p (xsize, x0, ysize, y0, c); | |
2337 | } | |
99c14947 | 2338 | |
2339 | default: | |
2340 | break; | |
ea0cb7ae | 2341 | } |
2342 | ||
f78d5f53 | 2343 | /* Deal with alignment ANDs by adjusting offset and size so as to |
2344 | cover the maximum range, without taking any previously known | |
bb2bc4f3 | 2345 | alignment into account. Make a size negative after such an |
2346 | adjustments, so that, if we end up with e.g. two SYMBOL_REFs, we | |
2347 | assume a potential overlap, because they may end up in contiguous | |
2348 | memory locations and the stricter-alignment access may span over | |
2349 | part of both. */ | |
971ba038 | 2350 | if (GET_CODE (x) == AND && CONST_INT_P (XEXP (x, 1))) |
9fb1c31b | 2351 | { |
f78d5f53 | 2352 | HOST_WIDE_INT sc = INTVAL (XEXP (x, 1)); |
2353 | unsigned HOST_WIDE_INT uc = sc; | |
bb2bc4f3 | 2354 | if (sc < 0 && -uc == (uc & -uc)) |
f78d5f53 | 2355 | { |
bb2bc4f3 | 2356 | if (xsize > 0) |
2357 | xsize = -xsize; | |
340ad617 | 2358 | if (xsize) |
2359 | xsize += sc + 1; | |
686d491a | 2360 | c -= sc + 1; |
f78d5f53 | 2361 | return memrefs_conflict_p (xsize, canon_rtx (XEXP (x, 0)), |
2362 | ysize, y, c); | |
2363 | } | |
9fb1c31b | 2364 | } |
971ba038 | 2365 | if (GET_CODE (y) == AND && CONST_INT_P (XEXP (y, 1))) |
1e38bf49 | 2366 | { |
f78d5f53 | 2367 | HOST_WIDE_INT sc = INTVAL (XEXP (y, 1)); |
2368 | unsigned HOST_WIDE_INT uc = sc; | |
bb2bc4f3 | 2369 | if (sc < 0 && -uc == (uc & -uc)) |
f78d5f53 | 2370 | { |
bb2bc4f3 | 2371 | if (ysize > 0) |
2372 | ysize = -ysize; | |
340ad617 | 2373 | if (ysize) |
2374 | ysize += sc + 1; | |
686d491a | 2375 | c += sc + 1; |
f78d5f53 | 2376 | return memrefs_conflict_p (xsize, x, |
2377 | ysize, canon_rtx (XEXP (y, 0)), c); | |
2378 | } | |
1e38bf49 | 2379 | } |
ea0cb7ae | 2380 | |
2381 | if (CONSTANT_P (x)) | |
2382 | { | |
971ba038 | 2383 | if (CONST_INT_P (x) && CONST_INT_P (y)) |
ea0cb7ae | 2384 | { |
2385 | c += (INTVAL (y) - INTVAL (x)); | |
340ad617 | 2386 | return offset_overlap_p (c, xsize, ysize); |
ea0cb7ae | 2387 | } |
2388 | ||
2389 | if (GET_CODE (x) == CONST) | |
2390 | { | |
2391 | if (GET_CODE (y) == CONST) | |
2392 | return memrefs_conflict_p (xsize, canon_rtx (XEXP (x, 0)), | |
2393 | ysize, canon_rtx (XEXP (y, 0)), c); | |
2394 | else | |
2395 | return memrefs_conflict_p (xsize, canon_rtx (XEXP (x, 0)), | |
2396 | ysize, y, c); | |
2397 | } | |
2398 | if (GET_CODE (y) == CONST) | |
2399 | return memrefs_conflict_p (xsize, x, ysize, | |
2400 | canon_rtx (XEXP (y, 0)), c); | |
2401 | ||
340ad617 | 2402 | /* Assume a potential overlap for symbolic addresses that went |
2403 | through alignment adjustments (i.e., that have negative | |
2404 | sizes), because we can't know how far they are from each | |
2405 | other. */ | |
ea0cb7ae | 2406 | if (CONSTANT_P (y)) |
340ad617 | 2407 | return (xsize < 0 || ysize < 0 || offset_overlap_p (c, xsize, ysize)); |
ea0cb7ae | 2408 | |
40f9e245 | 2409 | return -1; |
ea0cb7ae | 2410 | } |
40f9e245 | 2411 | |
2412 | return -1; | |
ea0cb7ae | 2413 | } |
2414 | ||
2415 | /* Functions to compute memory dependencies. | |
2416 | ||
2417 | Since we process the insns in execution order, we can build tables | |
2418 | to keep track of what registers are fixed (and not aliased), what registers | |
2419 | are varying in known ways, and what registers are varying in unknown | |
2420 | ways. | |
2421 | ||
2422 | If both memory references are volatile, then there must always be a | |
2423 | dependence between the two references, since their order can not be | |
2424 | changed. A volatile and non-volatile reference can be interchanged | |
9342ee68 | 2425 | though. |
ea0cb7ae | 2426 | |
376a287d | 2427 | We also must allow AND addresses, because they may generate accesses |
2428 | outside the object being referenced. This is used to generate aligned | |
2429 | addresses from unaligned addresses, for instance, the alpha | |
71d9e995 | 2430 | storeqi_unaligned pattern. */ |
ea0cb7ae | 2431 | |
2432 | /* Read dependence: X is read after read in MEM takes place. There can | |
eae726e9 | 2433 | only be a dependence here if both reads are volatile, or if either is |
2434 | an explicit barrier. */ | |
ea0cb7ae | 2435 | |
2436 | int | |
52d07779 | 2437 | read_dependence (const_rtx mem, const_rtx x) |
ea0cb7ae | 2438 | { |
eae726e9 | 2439 | if (MEM_VOLATILE_P (x) && MEM_VOLATILE_P (mem)) |
2440 | return true; | |
2441 | if (MEM_ALIAS_SET (x) == ALIAS_SET_MEMORY_BARRIER | |
2442 | || MEM_ALIAS_SET (mem) == ALIAS_SET_MEMORY_BARRIER) | |
2443 | return true; | |
2444 | return false; | |
ea0cb7ae | 2445 | } |
2446 | ||
b10dbbca | 2447 | /* Look at the bottom of the COMPONENT_REF list for a DECL, and return it. */ |
2448 | ||
2449 | static tree | |
aecda0d6 | 2450 | decl_for_component_ref (tree x) |
b10dbbca | 2451 | { |
2452 | do | |
2453 | { | |
2454 | x = TREE_OPERAND (x, 0); | |
2455 | } | |
2456 | while (x && TREE_CODE (x) == COMPONENT_REF); | |
2457 | ||
2458 | return x && DECL_P (x) ? x : NULL_TREE; | |
2459 | } | |
2460 | ||
da443c27 | 2461 | /* Walk up the COMPONENT_REF list in X and adjust *OFFSET to compensate |
2462 | for the offset of the field reference. *KNOWN_P says whether the | |
2463 | offset is known. */ | |
b10dbbca | 2464 | |
da443c27 | 2465 | static void |
2466 | adjust_offset_for_component_ref (tree x, bool *known_p, | |
2467 | HOST_WIDE_INT *offset) | |
b10dbbca | 2468 | { |
da443c27 | 2469 | if (!*known_p) |
2470 | return; | |
9342ee68 | 2471 | do |
b10dbbca | 2472 | { |
da443c27 | 2473 | tree xoffset = component_ref_field_offset (x); |
b10dbbca | 2474 | tree field = TREE_OPERAND (x, 1); |
e913b5cd | 2475 | if (TREE_CODE (xoffset) != INTEGER_CST) |
2476 | { | |
2477 | *known_p = false; | |
2478 | return; | |
2479 | } | |
2480 | ||
ab2c1de8 | 2481 | offset_int woffset |
2482 | = (wi::to_offset (xoffset) | |
2315e038 | 2483 | + wi::lrshift (wi::to_offset (DECL_FIELD_BIT_OFFSET (field)), |
2484 | LOG2_BITS_PER_UNIT)); | |
796b6678 | 2485 | if (!wi::fits_uhwi_p (woffset)) |
da443c27 | 2486 | { |
2487 | *known_p = false; | |
2488 | return; | |
2489 | } | |
e913b5cd | 2490 | *offset += woffset.to_uhwi (); |
b10dbbca | 2491 | |
2492 | x = TREE_OPERAND (x, 0); | |
2493 | } | |
2494 | while (x && TREE_CODE (x) == COMPONENT_REF); | |
b10dbbca | 2495 | } |
2496 | ||
d716ce75 | 2497 | /* Return nonzero if we can determine the exprs corresponding to memrefs |
a84256aa | 2498 | X and Y and they do not overlap. |
2499 | If LOOP_VARIANT is set, skip offset-based disambiguation */ | |
ddb3d16f | 2500 | |
5fe61d21 | 2501 | int |
a84256aa | 2502 | nonoverlapping_memrefs_p (const_rtx x, const_rtx y, bool loop_invariant) |
ddb3d16f | 2503 | { |
b10dbbca | 2504 | tree exprx = MEM_EXPR (x), expry = MEM_EXPR (y); |
ddb3d16f | 2505 | rtx rtlx, rtly; |
2506 | rtx basex, basey; | |
da443c27 | 2507 | bool moffsetx_known_p, moffsety_known_p; |
2508 | HOST_WIDE_INT moffsetx = 0, moffsety = 0; | |
ddb3d16f | 2509 | HOST_WIDE_INT offsetx = 0, offsety = 0, sizex, sizey, tem; |
2510 | ||
b10dbbca | 2511 | /* Unless both have exprs, we can't tell anything. */ |
2512 | if (exprx == 0 || expry == 0) | |
2513 | return 0; | |
c6b80e5d | 2514 | |
2515 | /* For spill-slot accesses make sure we have valid offsets. */ | |
2516 | if ((exprx == get_spill_slot_decl (false) | |
da443c27 | 2517 | && ! MEM_OFFSET_KNOWN_P (x)) |
c6b80e5d | 2518 | || (expry == get_spill_slot_decl (false) |
da443c27 | 2519 | && ! MEM_OFFSET_KNOWN_P (y))) |
c6b80e5d | 2520 | return 0; |
a0c938f0 | 2521 | |
b10dbbca | 2522 | /* If the field reference test failed, look at the DECLs involved. */ |
da443c27 | 2523 | moffsetx_known_p = MEM_OFFSET_KNOWN_P (x); |
2524 | if (moffsetx_known_p) | |
2525 | moffsetx = MEM_OFFSET (x); | |
b10dbbca | 2526 | if (TREE_CODE (exprx) == COMPONENT_REF) |
2527 | { | |
b50c6e49 | 2528 | tree t = decl_for_component_ref (exprx); |
2529 | if (! t) | |
2530 | return 0; | |
da443c27 | 2531 | adjust_offset_for_component_ref (exprx, &moffsetx_known_p, &moffsetx); |
b50c6e49 | 2532 | exprx = t; |
b10dbbca | 2533 | } |
2d8fe5d0 | 2534 | |
da443c27 | 2535 | moffsety_known_p = MEM_OFFSET_KNOWN_P (y); |
2536 | if (moffsety_known_p) | |
2537 | moffsety = MEM_OFFSET (y); | |
b10dbbca | 2538 | if (TREE_CODE (expry) == COMPONENT_REF) |
2539 | { | |
b50c6e49 | 2540 | tree t = decl_for_component_ref (expry); |
2541 | if (! t) | |
2542 | return 0; | |
da443c27 | 2543 | adjust_offset_for_component_ref (expry, &moffsety_known_p, &moffsety); |
b50c6e49 | 2544 | expry = t; |
b10dbbca | 2545 | } |
2546 | ||
2547 | if (! DECL_P (exprx) || ! DECL_P (expry)) | |
ddb3d16f | 2548 | return 0; |
2549 | ||
119147d1 | 2550 | /* With invalid code we can end up storing into the constant pool. |
2551 | Bail out to avoid ICEing when creating RTL for this. | |
2552 | See gfortran.dg/lto/20091028-2_0.f90. */ | |
2553 | if (TREE_CODE (exprx) == CONST_DECL | |
2554 | || TREE_CODE (expry) == CONST_DECL) | |
2555 | return 1; | |
2556 | ||
b10dbbca | 2557 | rtlx = DECL_RTL (exprx); |
2558 | rtly = DECL_RTL (expry); | |
ddb3d16f | 2559 | |
10e949c1 | 2560 | /* If either RTL is not a MEM, it must be a REG or CONCAT, meaning they |
2561 | can't overlap unless they are the same because we never reuse that part | |
2562 | of the stack frame used for locals for spilled pseudos. */ | |
e16ceb8e | 2563 | if ((!MEM_P (rtlx) || !MEM_P (rtly)) |
10e949c1 | 2564 | && ! rtx_equal_p (rtlx, rtly)) |
ddb3d16f | 2565 | return 1; |
2566 | ||
04ec15fa | 2567 | /* If we have MEMs referring to different address spaces (which can |
bd1a81f7 | 2568 | potentially overlap), we cannot easily tell from the addresses |
2569 | whether the references overlap. */ | |
2570 | if (MEM_P (rtlx) && MEM_P (rtly) | |
2571 | && MEM_ADDR_SPACE (rtlx) != MEM_ADDR_SPACE (rtly)) | |
2572 | return 0; | |
2573 | ||
ddb3d16f | 2574 | /* Get the base and offsets of both decls. If either is a register, we |
2575 | know both are and are the same, so use that as the base. The only | |
2576 | we can avoid overlap is if we can deduce that they are nonoverlapping | |
2577 | pieces of that decl, which is very rare. */ | |
e16ceb8e | 2578 | basex = MEM_P (rtlx) ? XEXP (rtlx, 0) : rtlx; |
971ba038 | 2579 | if (GET_CODE (basex) == PLUS && CONST_INT_P (XEXP (basex, 1))) |
ddb3d16f | 2580 | offsetx = INTVAL (XEXP (basex, 1)), basex = XEXP (basex, 0); |
2581 | ||
e16ceb8e | 2582 | basey = MEM_P (rtly) ? XEXP (rtly, 0) : rtly; |
971ba038 | 2583 | if (GET_CODE (basey) == PLUS && CONST_INT_P (XEXP (basey, 1))) |
ddb3d16f | 2584 | offsety = INTVAL (XEXP (basey, 1)), basey = XEXP (basey, 0); |
2585 | ||
c9c48480 | 2586 | /* If the bases are different, we know they do not overlap if both |
9342ee68 | 2587 | are constants or if one is a constant and the other a pointer into the |
c9c48480 | 2588 | stack frame. Otherwise a different base means we can't tell if they |
2589 | overlap or not. */ | |
2590 | if (! rtx_equal_p (basex, basey)) | |
9342ee68 | 2591 | return ((CONSTANT_P (basex) && CONSTANT_P (basey)) |
2592 | || (CONSTANT_P (basex) && REG_P (basey) | |
2593 | && REGNO_PTR_FRAME_P (REGNO (basey))) | |
2594 | || (CONSTANT_P (basey) && REG_P (basex) | |
2595 | && REGNO_PTR_FRAME_P (REGNO (basex)))); | |
ddb3d16f | 2596 | |
a84256aa | 2597 | /* Offset based disambiguation not appropriate for loop invariant */ |
2598 | if (loop_invariant) | |
2599 | return 0; | |
2600 | ||
e16ceb8e | 2601 | sizex = (!MEM_P (rtlx) ? (int) GET_MODE_SIZE (GET_MODE (rtlx)) |
5b2a69fa | 2602 | : MEM_SIZE_KNOWN_P (rtlx) ? MEM_SIZE (rtlx) |
ddb3d16f | 2603 | : -1); |
e16ceb8e | 2604 | sizey = (!MEM_P (rtly) ? (int) GET_MODE_SIZE (GET_MODE (rtly)) |
5b2a69fa | 2605 | : MEM_SIZE_KNOWN_P (rtly) ? MEM_SIZE (rtly) |
2606 | : -1); | |
ddb3d16f | 2607 | |
dc2dfdb6 | 2608 | /* If we have an offset for either memref, it can update the values computed |
2609 | above. */ | |
da443c27 | 2610 | if (moffsetx_known_p) |
2611 | offsetx += moffsetx, sizex -= moffsetx; | |
2612 | if (moffsety_known_p) | |
2613 | offsety += moffsety, sizey -= moffsety; | |
ddb3d16f | 2614 | |
dc2dfdb6 | 2615 | /* If a memref has both a size and an offset, we can use the smaller size. |
203c9fee | 2616 | We can't do this if the offset isn't known because we must view this |
dc2dfdb6 | 2617 | memref as being anywhere inside the DECL's MEM. */ |
da443c27 | 2618 | if (MEM_SIZE_KNOWN_P (x) && moffsetx_known_p) |
5b2a69fa | 2619 | sizex = MEM_SIZE (x); |
da443c27 | 2620 | if (MEM_SIZE_KNOWN_P (y) && moffsety_known_p) |
5b2a69fa | 2621 | sizey = MEM_SIZE (y); |
ddb3d16f | 2622 | |
2623 | /* Put the values of the memref with the lower offset in X's values. */ | |
2624 | if (offsetx > offsety) | |
2625 | { | |
2626 | tem = offsetx, offsetx = offsety, offsety = tem; | |
2627 | tem = sizex, sizex = sizey, sizey = tem; | |
2628 | } | |
2629 | ||
2630 | /* If we don't know the size of the lower-offset value, we can't tell | |
2631 | if they conflict. Otherwise, we do the test. */ | |
906923ba | 2632 | return sizex >= 0 && offsety >= offsetx + sizex; |
ddb3d16f | 2633 | } |
2634 | ||
02f77cc4 | 2635 | /* Helper for true_dependence and canon_true_dependence. |
2636 | Checks for true dependence: X is read after store in MEM takes place. | |
ea0cb7ae | 2637 | |
02f77cc4 | 2638 | If MEM_CANONICALIZED is FALSE, then X_ADDR and MEM_ADDR should be |
2639 | NULL_RTX, and the canonical addresses of MEM and X are both computed | |
2640 | here. If MEM_CANONICALIZED, then MEM must be already canonicalized. | |
2641 | ||
2642 | If X_ADDR is non-NULL, it is used in preference of XEXP (x, 0). | |
2643 | ||
2644 | Returns 1 if there is a true dependence, 0 otherwise. */ | |
2645 | ||
2646 | static int | |
3754d046 | 2647 | true_dependence_1 (const_rtx mem, machine_mode mem_mode, rtx mem_addr, |
376a287d | 2648 | const_rtx x, rtx x_addr, bool mem_canonicalized) |
ea0cb7ae | 2649 | { |
0d46a25a | 2650 | rtx true_mem_addr; |
9ee7cac5 | 2651 | rtx base; |
40f9e245 | 2652 | int ret; |
ea0cb7ae | 2653 | |
02f77cc4 | 2654 | gcc_checking_assert (mem_canonicalized ? (mem_addr != NULL_RTX) |
2655 | : (mem_addr == NULL_RTX && x_addr == NULL_RTX)); | |
2656 | ||
ea0cb7ae | 2657 | if (MEM_VOLATILE_P (x) && MEM_VOLATILE_P (mem)) |
2658 | return 1; | |
2659 | ||
a57f4c4d | 2660 | /* (mem:BLK (scratch)) is a special mechanism to conflict with everything. |
67837170 | 2661 | This is used in epilogue deallocation functions, and in cselib. */ |
a57f4c4d | 2662 | if (GET_MODE (x) == BLKmode && GET_CODE (XEXP (x, 0)) == SCRATCH) |
2663 | return 1; | |
2664 | if (GET_MODE (mem) == BLKmode && GET_CODE (XEXP (mem, 0)) == SCRATCH) | |
2665 | return 1; | |
c94cfd1c | 2666 | if (MEM_ALIAS_SET (x) == ALIAS_SET_MEMORY_BARRIER |
2667 | || MEM_ALIAS_SET (mem) == ALIAS_SET_MEMORY_BARRIER) | |
2668 | return 1; | |
a57f4c4d | 2669 | |
0d46a25a | 2670 | if (! x_addr) |
2671 | x_addr = XEXP (x, 0); | |
2672 | x_addr = get_addr (x_addr); | |
2673 | ||
02f77cc4 | 2674 | if (! mem_addr) |
2675 | { | |
2676 | mem_addr = XEXP (mem, 0); | |
2677 | if (mem_mode == VOIDmode) | |
2678 | mem_mode = GET_MODE (mem); | |
2679 | } | |
0d46a25a | 2680 | true_mem_addr = get_addr (mem_addr); |
1617c276 | 2681 | |
41e22474 | 2682 | /* Read-only memory is by definition never modified, and therefore can't |
2683 | conflict with anything. However, don't assume anything when AND | |
2684 | addresses are involved and leave to the code below to determine | |
2685 | dependence. We don't expect to find read-only set on MEM, but | |
2686 | stupid user tricks can produce them, so don't die. */ | |
2687 | if (MEM_READONLY_P (x) | |
2688 | && GET_CODE (x_addr) != AND | |
0d46a25a | 2689 | && GET_CODE (true_mem_addr) != AND) |
41e22474 | 2690 | return 0; |
2691 | ||
2692 | /* If we have MEMs referring to different address spaces (which can | |
2693 | potentially overlap), we cannot easily tell from the addresses | |
2694 | whether the references overlap. */ | |
2695 | if (MEM_ADDR_SPACE (mem) != MEM_ADDR_SPACE (x)) | |
2696 | return 1; | |
2697 | ||
002fe3cb | 2698 | base = find_base_term (x_addr); |
2699 | if (base && (GET_CODE (base) == LABEL_REF | |
2700 | || (GET_CODE (base) == SYMBOL_REF | |
2701 | && CONSTANT_POOL_ADDRESS_P (base)))) | |
2702 | return 0; | |
2703 | ||
0d46a25a | 2704 | rtx mem_base = find_base_term (true_mem_addr); |
2705 | if (! base_alias_check (x_addr, base, true_mem_addr, mem_base, | |
5abc99e4 | 2706 | GET_MODE (x), mem_mode)) |
3ad3a0ed | 2707 | return 0; |
2708 | ||
1617c276 | 2709 | x_addr = canon_rtx (x_addr); |
02f77cc4 | 2710 | if (!mem_canonicalized) |
0d46a25a | 2711 | mem_addr = canon_rtx (true_mem_addr); |
35d3eedd | 2712 | |
40f9e245 | 2713 | if ((ret = memrefs_conflict_p (GET_MODE_SIZE (mem_mode), mem_addr, |
2714 | SIZE_FOR_MODE (x), x_addr, 0)) != -1) | |
2715 | return ret; | |
2716 | ||
557f718e | 2717 | if (mems_in_disjoint_alias_sets_p (x, mem)) |
40f9e245 | 2718 | return 0; |
2719 | ||
a84256aa | 2720 | if (nonoverlapping_memrefs_p (mem, x, false)) |
4958fe6e | 2721 | return 0; |
658f1736 | 2722 | |
3a443843 | 2723 | return rtx_refs_may_alias_p (x, mem, true); |
8e802be9 | 2724 | } |
2725 | ||
02f77cc4 | 2726 | /* True dependence: X is read after store in MEM takes place. */ |
2727 | ||
2728 | int | |
3754d046 | 2729 | true_dependence (const_rtx mem, machine_mode mem_mode, const_rtx x) |
02f77cc4 | 2730 | { |
2731 | return true_dependence_1 (mem, mem_mode, NULL_RTX, | |
376a287d | 2732 | x, NULL_RTX, /*mem_canonicalized=*/false); |
02f77cc4 | 2733 | } |
2734 | ||
8e802be9 | 2735 | /* Canonical true dependence: X is read after store in MEM takes place. |
9342ee68 | 2736 | Variant of true_dependence which assumes MEM has already been |
2737 | canonicalized (hence we no longer do that here). | |
02f77cc4 | 2738 | The mem_addr argument has been added, since true_dependence_1 computed |
2739 | this value prior to canonicalizing. */ | |
8e802be9 | 2740 | |
2741 | int | |
3754d046 | 2742 | canon_true_dependence (const_rtx mem, machine_mode mem_mode, rtx mem_addr, |
376a287d | 2743 | const_rtx x, rtx x_addr) |
8e802be9 | 2744 | { |
02f77cc4 | 2745 | return true_dependence_1 (mem, mem_mode, mem_addr, |
376a287d | 2746 | x, x_addr, /*mem_canonicalized=*/true); |
ea0cb7ae | 2747 | } |
2748 | ||
d10cfa8d | 2749 | /* Returns nonzero if a write to X might alias a previous read from |
ddba76b8 | 2750 | (or, if WRITEP is true, a write to) MEM. |
cb456db5 | 2751 | If X_CANONCALIZED is true, then X_ADDR is the canonicalized address of X, |
2752 | and X_MODE the mode for that access. | |
2753 | If MEM_CANONICALIZED is true, MEM is canonicalized. */ | |
ea0cb7ae | 2754 | |
fe69e4fe | 2755 | static int |
cb456db5 | 2756 | write_dependence_p (const_rtx mem, |
3754d046 | 2757 | const_rtx x, machine_mode x_mode, rtx x_addr, |
cb456db5 | 2758 | bool mem_canonicalized, bool x_canonicalized, bool writep) |
ea0cb7ae | 2759 | { |
cb456db5 | 2760 | rtx mem_addr; |
0d46a25a | 2761 | rtx true_mem_addr, true_x_addr; |
9ee7cac5 | 2762 | rtx base; |
40f9e245 | 2763 | int ret; |
35d3eedd | 2764 | |
cb456db5 | 2765 | gcc_checking_assert (x_canonicalized |
2766 | ? (x_addr != NULL_RTX && x_mode != VOIDmode) | |
2767 | : (x_addr == NULL_RTX && x_mode == VOIDmode)); | |
ddba76b8 | 2768 | |
ea0cb7ae | 2769 | if (MEM_VOLATILE_P (x) && MEM_VOLATILE_P (mem)) |
2770 | return 1; | |
2771 | ||
a57f4c4d | 2772 | /* (mem:BLK (scratch)) is a special mechanism to conflict with everything. |
2773 | This is used in epilogue deallocation functions. */ | |
2774 | if (GET_MODE (x) == BLKmode && GET_CODE (XEXP (x, 0)) == SCRATCH) | |
2775 | return 1; | |
2776 | if (GET_MODE (mem) == BLKmode && GET_CODE (XEXP (mem, 0)) == SCRATCH) | |
2777 | return 1; | |
c94cfd1c | 2778 | if (MEM_ALIAS_SET (x) == ALIAS_SET_MEMORY_BARRIER |
2779 | || MEM_ALIAS_SET (mem) == ALIAS_SET_MEMORY_BARRIER) | |
2780 | return 1; | |
a57f4c4d | 2781 | |
cb456db5 | 2782 | if (!x_addr) |
0d46a25a | 2783 | x_addr = XEXP (x, 0); |
2784 | true_x_addr = get_addr (x_addr); | |
2785 | ||
2786 | mem_addr = XEXP (mem, 0); | |
2787 | true_mem_addr = get_addr (mem_addr); | |
002fe3cb | 2788 | |
41e22474 | 2789 | /* A read from read-only memory can't conflict with read-write memory. |
2790 | Don't assume anything when AND addresses are involved and leave to | |
2791 | the code below to determine dependence. */ | |
2792 | if (!writep | |
2793 | && MEM_READONLY_P (mem) | |
0d46a25a | 2794 | && GET_CODE (true_x_addr) != AND |
2795 | && GET_CODE (true_mem_addr) != AND) | |
41e22474 | 2796 | return 0; |
2797 | ||
2798 | /* If we have MEMs referring to different address spaces (which can | |
2799 | potentially overlap), we cannot easily tell from the addresses | |
2800 | whether the references overlap. */ | |
2801 | if (MEM_ADDR_SPACE (mem) != MEM_ADDR_SPACE (x)) | |
2802 | return 1; | |
2803 | ||
0d46a25a | 2804 | base = find_base_term (true_mem_addr); |
5abc99e4 | 2805 | if (! writep |
2806 | && base | |
2807 | && (GET_CODE (base) == LABEL_REF | |
2808 | || (GET_CODE (base) == SYMBOL_REF | |
2809 | && CONSTANT_POOL_ADDRESS_P (base)))) | |
2810 | return 0; | |
9ee7cac5 | 2811 | |
0d46a25a | 2812 | rtx x_base = find_base_term (true_x_addr); |
2813 | if (! base_alias_check (true_x_addr, x_base, true_mem_addr, base, | |
2814 | GET_MODE (x), GET_MODE (mem))) | |
b5ba9f3a | 2815 | return 0; |
2816 | ||
cb456db5 | 2817 | if (!x_canonicalized) |
ddba76b8 | 2818 | { |
0d46a25a | 2819 | x_addr = canon_rtx (true_x_addr); |
cb456db5 | 2820 | x_mode = GET_MODE (x); |
ddba76b8 | 2821 | } |
cb456db5 | 2822 | if (!mem_canonicalized) |
0d46a25a | 2823 | mem_addr = canon_rtx (true_mem_addr); |
35d3eedd | 2824 | |
cb456db5 | 2825 | if ((ret = memrefs_conflict_p (SIZE_FOR_MODE (mem), mem_addr, |
2826 | GET_MODE_SIZE (x_mode), x_addr, 0)) != -1) | |
40f9e245 | 2827 | return ret; |
2828 | ||
a84256aa | 2829 | if (nonoverlapping_memrefs_p (x, mem, false)) |
6a0934dd | 2830 | return 0; |
2831 | ||
3a443843 | 2832 | return rtx_refs_may_alias_p (x, mem, false); |
6a0934dd | 2833 | } |
2834 | ||
2835 | /* Anti dependence: X is written after read in MEM takes place. */ | |
2836 | ||
2837 | int | |
52d07779 | 2838 | anti_dependence (const_rtx mem, const_rtx x) |
6a0934dd | 2839 | { |
cb456db5 | 2840 | return write_dependence_p (mem, x, VOIDmode, NULL_RTX, |
2841 | /*mem_canonicalized=*/false, | |
2842 | /*x_canonicalized*/false, /*writep=*/false); | |
ddba76b8 | 2843 | } |
2844 | ||
cb456db5 | 2845 | /* Likewise, but we already have a canonicalized MEM, and X_ADDR for X. |
2846 | Also, consider X in X_MODE (which might be from an enclosing | |
2847 | STRICT_LOW_PART / ZERO_EXTRACT). | |
2848 | If MEM_CANONICALIZED is true, MEM is canonicalized. */ | |
ddba76b8 | 2849 | |
2850 | int | |
cb456db5 | 2851 | canon_anti_dependence (const_rtx mem, bool mem_canonicalized, |
3754d046 | 2852 | const_rtx x, machine_mode x_mode, rtx x_addr) |
ddba76b8 | 2853 | { |
cb456db5 | 2854 | return write_dependence_p (mem, x, x_mode, x_addr, |
2855 | mem_canonicalized, /*x_canonicalized=*/true, | |
2856 | /*writep=*/false); | |
ea0cb7ae | 2857 | } |
2858 | ||
2859 | /* Output dependence: X is written after store in MEM takes place. */ | |
2860 | ||
2861 | int | |
52d07779 | 2862 | output_dependence (const_rtx mem, const_rtx x) |
ea0cb7ae | 2863 | { |
cb456db5 | 2864 | return write_dependence_p (mem, x, VOIDmode, NULL_RTX, |
2865 | /*mem_canonicalized=*/false, | |
2866 | /*x_canonicalized*/false, /*writep=*/true); | |
ea0cb7ae | 2867 | } |
c352cbc7 | 2868 | \f |
35d3eedd | 2869 | |
a84256aa | 2870 | |
2871 | /* Check whether X may be aliased with MEM. Don't do offset-based | |
2872 | memory disambiguation & TBAA. */ | |
2873 | int | |
2874 | may_alias_p (const_rtx mem, const_rtx x) | |
2875 | { | |
2876 | rtx x_addr, mem_addr; | |
a84256aa | 2877 | |
2878 | if (MEM_VOLATILE_P (x) && MEM_VOLATILE_P (mem)) | |
2879 | return 1; | |
2880 | ||
557f718e | 2881 | /* (mem:BLK (scratch)) is a special mechanism to conflict with everything. |
2882 | This is used in epilogue deallocation functions. */ | |
2883 | if (GET_MODE (x) == BLKmode && GET_CODE (XEXP (x, 0)) == SCRATCH) | |
2884 | return 1; | |
2885 | if (GET_MODE (mem) == BLKmode && GET_CODE (XEXP (mem, 0)) == SCRATCH) | |
a84256aa | 2886 | return 1; |
a84256aa | 2887 | if (MEM_ALIAS_SET (x) == ALIAS_SET_MEMORY_BARRIER |
2888 | || MEM_ALIAS_SET (mem) == ALIAS_SET_MEMORY_BARRIER) | |
2889 | return 1; | |
2890 | ||
a84256aa | 2891 | x_addr = XEXP (x, 0); |
0d46a25a | 2892 | x_addr = get_addr (x_addr); |
2893 | ||
a84256aa | 2894 | mem_addr = XEXP (mem, 0); |
0d46a25a | 2895 | mem_addr = get_addr (mem_addr); |
a84256aa | 2896 | |
41e22474 | 2897 | /* Read-only memory is by definition never modified, and therefore can't |
2898 | conflict with anything. However, don't assume anything when AND | |
2899 | addresses are involved and leave to the code below to determine | |
2900 | dependence. We don't expect to find read-only set on MEM, but | |
2901 | stupid user tricks can produce them, so don't die. */ | |
2902 | if (MEM_READONLY_P (x) | |
2903 | && GET_CODE (x_addr) != AND | |
2904 | && GET_CODE (mem_addr) != AND) | |
2905 | return 0; | |
2906 | ||
2907 | /* If we have MEMs referring to different address spaces (which can | |
2908 | potentially overlap), we cannot easily tell from the addresses | |
2909 | whether the references overlap. */ | |
2910 | if (MEM_ADDR_SPACE (mem) != MEM_ADDR_SPACE (x)) | |
2911 | return 1; | |
2912 | ||
5abc99e4 | 2913 | rtx x_base = find_base_term (x_addr); |
2914 | rtx mem_base = find_base_term (mem_addr); | |
2915 | if (! base_alias_check (x_addr, x_base, mem_addr, mem_base, | |
2916 | GET_MODE (x), GET_MODE (mem_addr))) | |
a84256aa | 2917 | return 0; |
2918 | ||
a84256aa | 2919 | if (nonoverlapping_memrefs_p (mem, x, true)) |
2920 | return 0; | |
2921 | ||
a84256aa | 2922 | /* TBAA not valid for loop_invarint */ |
2923 | return rtx_refs_may_alias_p (x, mem, false); | |
2924 | } | |
2925 | ||
35d3eedd | 2926 | void |
6d8b68a3 | 2927 | init_alias_target (void) |
35d3eedd | 2928 | { |
19cb6b50 | 2929 | int i; |
35d3eedd | 2930 | |
4e56ceb1 | 2931 | if (!arg_base_value) |
2932 | arg_base_value = gen_rtx_ADDRESS (VOIDmode, 0); | |
2933 | ||
6d8b68a3 | 2934 | memset (static_reg_base_value, 0, sizeof static_reg_base_value); |
2935 | ||
35d3eedd | 2936 | for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) |
2937 | /* Check whether this register can hold an incoming pointer | |
2938 | argument. FUNCTION_ARG_REGNO_P tests outgoing register | |
2c0e001b | 2939 | numbers, so translate if necessary due to register windows. */ |
35d3eedd | 2940 | if (FUNCTION_ARG_REGNO_P (OUTGOING_REGNO (i)) |
2941 | && HARD_REGNO_MODE_OK (i, Pmode)) | |
4e56ceb1 | 2942 | static_reg_base_value[i] = arg_base_value; |
2943 | ||
df3e5a67 | 2944 | static_reg_base_value[STACK_POINTER_REGNUM] |
2945 | = unique_base_value (UNIQUE_BASE_VALUE_SP); | |
2946 | static_reg_base_value[ARG_POINTER_REGNUM] | |
2947 | = unique_base_value (UNIQUE_BASE_VALUE_ARGP); | |
2948 | static_reg_base_value[FRAME_POINTER_REGNUM] | |
2949 | = unique_base_value (UNIQUE_BASE_VALUE_FP); | |
f703b3d6 | 2950 | if (!HARD_FRAME_POINTER_IS_FRAME_POINTER) |
2951 | static_reg_base_value[HARD_FRAME_POINTER_REGNUM] | |
2952 | = unique_base_value (UNIQUE_BASE_VALUE_HFP); | |
0c7f5242 | 2953 | } |
2954 | ||
c7bf7428 | 2955 | /* Set MEMORY_MODIFIED when X modifies DATA (that is assumed |
2956 | to be memory reference. */ | |
2957 | static bool memory_modified; | |
2958 | static void | |
a80f1c6c | 2959 | memory_modified_1 (rtx x, const_rtx pat ATTRIBUTE_UNUSED, void *data) |
c7bf7428 | 2960 | { |
e16ceb8e | 2961 | if (MEM_P (x)) |
c7bf7428 | 2962 | { |
5493cb9a | 2963 | if (anti_dependence (x, (const_rtx)data) || output_dependence (x, (const_rtx)data)) |
c7bf7428 | 2964 | memory_modified = true; |
2965 | } | |
2966 | } | |
2967 | ||
2968 | ||
2969 | /* Return true when INSN possibly modify memory contents of MEM | |
0c6d8c36 | 2970 | (i.e. address can be modified). */ |
c7bf7428 | 2971 | bool |
5493cb9a | 2972 | memory_modified_in_insn_p (const_rtx mem, const_rtx insn) |
c7bf7428 | 2973 | { |
2974 | if (!INSN_P (insn)) | |
2975 | return false; | |
2976 | memory_modified = false; | |
a80f1c6c | 2977 | note_stores (PATTERN (insn), memory_modified_1, CONST_CAST_RTX(mem)); |
c7bf7428 | 2978 | return memory_modified; |
2979 | } | |
2980 | ||
43b6ba45 | 2981 | /* Return TRUE if the destination of a set is rtx identical to |
2982 | ITEM. */ | |
2983 | static inline bool | |
2984 | set_dest_equal_p (const_rtx set, const_rtx item) | |
2985 | { | |
2986 | rtx dest = SET_DEST (set); | |
2987 | return rtx_equal_p (dest, item); | |
2988 | } | |
2989 | ||
2990 | /* Like memory_modified_in_insn_p, but return TRUE if INSN will | |
2991 | *DEFINITELY* modify the memory contents of MEM. */ | |
2992 | bool | |
2993 | memory_must_be_modified_in_insn_p (const_rtx mem, const_rtx insn) | |
2994 | { | |
2995 | if (!INSN_P (insn)) | |
2996 | return false; | |
2997 | insn = PATTERN (insn); | |
2998 | if (GET_CODE (insn) == SET) | |
2999 | return set_dest_equal_p (insn, mem); | |
3000 | else if (GET_CODE (insn) == PARALLEL) | |
3001 | { | |
3002 | int i; | |
3003 | for (i = 0; i < XVECLEN (insn, 0); i++) | |
3004 | { | |
3005 | rtx sub = XVECEXP (insn, 0, i); | |
3006 | if (GET_CODE (sub) == SET | |
3007 | && set_dest_equal_p (sub, mem)) | |
3008 | return true; | |
3009 | } | |
3010 | } | |
3011 | return false; | |
3012 | } | |
3013 | ||
73f5c1e3 | 3014 | /* Initialize the aliasing machinery. Initialize the REG_KNOWN_VALUE |
3015 | array. */ | |
3016 | ||
ea0cb7ae | 3017 | void |
aecda0d6 | 3018 | init_alias_analysis (void) |
ea0cb7ae | 3019 | { |
e5b19a8c | 3020 | unsigned int maxreg = max_reg_num (); |
37b864db | 3021 | int changed, pass; |
19cb6b50 | 3022 | int i; |
3023 | unsigned int ui; | |
a7e77bb7 | 3024 | rtx_insn *insn; |
3025 | rtx val; | |
2808fda8 | 3026 | int rpo_cnt; |
3027 | int *rpo; | |
ea0cb7ae | 3028 | |
376c21d1 | 3029 | timevar_push (TV_ALIAS_ANALYSIS); |
3030 | ||
5d3b0682 | 3031 | vec_safe_grow_cleared (reg_known_value, maxreg - FIRST_PSEUDO_REGISTER); |
d6443ebe | 3032 | reg_known_equiv_p = sbitmap_alloc (maxreg - FIRST_PSEUDO_REGISTER); |
7a2f311c | 3033 | bitmap_clear (reg_known_equiv_p); |
ea0cb7ae | 3034 | |
980f3285 | 3035 | /* If we have memory allocated from the previous run, use it. */ |
e5b19a8c | 3036 | if (old_reg_base_value) |
980f3285 | 3037 | reg_base_value = old_reg_base_value; |
3038 | ||
3039 | if (reg_base_value) | |
f1f41a6c | 3040 | reg_base_value->truncate (0); |
980f3285 | 3041 | |
f1f41a6c | 3042 | vec_safe_grow_cleared (reg_base_value, maxreg); |
c2e94b25 | 3043 | |
4c36ffe6 | 3044 | new_reg_base_value = XNEWVEC (rtx, maxreg); |
7bf60644 | 3045 | reg_seen = sbitmap_alloc (maxreg); |
06c5c535 | 3046 | |
3047 | /* The basic idea is that each pass through this loop will use the | |
3048 | "constant" information from the previous pass to propagate alias | |
3049 | information through another level of assignments. | |
3050 | ||
2808fda8 | 3051 | The propagation is done on the CFG in reverse post-order, to propagate |
3052 | things forward as far as possible in each iteration. | |
3053 | ||
06c5c535 | 3054 | This could get expensive if the assignment chains are long. Maybe |
3055 | we should throttle the number of iterations, possibly based on | |
35d3eedd | 3056 | the optimization level or flag_expensive_optimizations. |
06c5c535 | 3057 | |
3058 | We could propagate more information in the first pass by making use | |
3072d30e | 3059 | of DF_REG_DEF_COUNT to determine immediately that the alias information |
37b864db | 3060 | for a pseudo is "constant". |
3061 | ||
3062 | A program with an uninitialized variable can cause an infinite loop | |
3063 | here. Instead of doing a full dataflow analysis to detect such problems | |
3064 | we just cap the number of iterations for the loop. | |
3065 | ||
3066 | The state of the arrays for the set chain in question does not matter | |
3067 | since the program has undefined behavior. */ | |
35d3eedd | 3068 | |
a28770e1 | 3069 | rpo = XNEWVEC (int, n_basic_blocks_for_fn (cfun)); |
2808fda8 | 3070 | rpo_cnt = pre_and_rev_post_order_compute (NULL, rpo, false); |
3071 | ||
37b864db | 3072 | pass = 0; |
35d3eedd | 3073 | do |
06c5c535 | 3074 | { |
3075 | /* Assume nothing will change this iteration of the loop. */ | |
3076 | changed = 0; | |
3077 | ||
06c5c535 | 3078 | /* We want to assign the same IDs each iteration of this loop, so |
4e56ceb1 | 3079 | start counting from one each iteration of the loop. */ |
3080 | unique_id = 1; | |
06c5c535 | 3081 | |
35a3065a | 3082 | /* We're at the start of the function each iteration through the |
06c5c535 | 3083 | loop, so we're copying arguments. */ |
e5be0cc6 | 3084 | copying_arguments = true; |
ea0cb7ae | 3085 | |
35d3eedd | 3086 | /* Wipe the potential alias information clean for this pass. */ |
e5b19a8c | 3087 | memset (new_reg_base_value, 0, maxreg * sizeof (rtx)); |
f0ad0992 | 3088 | |
35d3eedd | 3089 | /* Wipe the reg_seen array clean. */ |
53c5d9d4 | 3090 | bitmap_clear (reg_seen); |
ea0cb7ae | 3091 | |
a15c38cd | 3092 | /* Initialize the alias information for this pass. */ |
3093 | for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) | |
3094 | if (static_reg_base_value[i]) | |
3095 | { | |
3096 | new_reg_base_value[i] = static_reg_base_value[i]; | |
3097 | bitmap_set_bit (reg_seen, i); | |
3098 | } | |
35d3eedd | 3099 | |
06c5c535 | 3100 | /* Walk the insns adding values to the new_reg_base_value array. */ |
2808fda8 | 3101 | for (i = 0; i < rpo_cnt; i++) |
ea0cb7ae | 3102 | { |
f5a6b05f | 3103 | basic_block bb = BASIC_BLOCK_FOR_FN (cfun, rpo[i]); |
2808fda8 | 3104 | FOR_BB_INSNS (bb, insn) |
06c5c535 | 3105 | { |
2808fda8 | 3106 | if (NONDEBUG_INSN_P (insn)) |
3107 | { | |
3108 | rtx note, set; | |
d6cb6164 | 3109 | |
317443b3 | 3110 | #if defined (HAVE_prologue) |
3111 | static const bool prologue = true; | |
3112 | #else | |
3113 | static const bool prologue = false; | |
3114 | #endif | |
3115 | ||
2808fda8 | 3116 | /* The prologue/epilogue insns are not threaded onto the |
3117 | insn chain until after reload has completed. Thus, | |
3118 | there is no sense wasting time checking if INSN is in | |
3119 | the prologue/epilogue until after reload has completed. */ | |
317443b3 | 3120 | if ((prologue || HAVE_epilogue) && reload_completed |
2808fda8 | 3121 | && prologue_epilogue_contains (insn)) |
3122 | continue; | |
d6cb6164 | 3123 | |
2808fda8 | 3124 | /* If this insn has a noalias note, process it, Otherwise, |
3125 | scan for sets. A simple set will have no side effects | |
3126 | which could change the base value of any other register. */ | |
35d3eedd | 3127 | |
2808fda8 | 3128 | if (GET_CODE (PATTERN (insn)) == SET |
3129 | && REG_NOTES (insn) != 0 | |
3130 | && find_reg_note (insn, REG_NOALIAS, NULL_RTX)) | |
3131 | record_set (SET_DEST (PATTERN (insn)), NULL_RTX, NULL); | |
3132 | else | |
3133 | note_stores (PATTERN (insn), record_set, NULL); | |
35d3eedd | 3134 | |
2808fda8 | 3135 | set = single_set (insn); |
35d3eedd | 3136 | |
2808fda8 | 3137 | if (set != 0 |
3138 | && REG_P (SET_DEST (set)) | |
3139 | && REGNO (SET_DEST (set)) >= FIRST_PSEUDO_REGISTER) | |
970c0c06 | 3140 | { |
2808fda8 | 3141 | unsigned int regno = REGNO (SET_DEST (set)); |
3142 | rtx src = SET_SRC (set); | |
3143 | rtx t; | |
3144 | ||
3145 | note = find_reg_equal_equiv_note (insn); | |
3146 | if (note && REG_NOTE_KIND (note) == REG_EQUAL | |
3147 | && DF_REG_DEF_COUNT (regno) != 1) | |
3148 | note = NULL_RTX; | |
3149 | ||
3150 | if (note != NULL_RTX | |
3151 | && GET_CODE (XEXP (note, 0)) != EXPR_LIST | |
3152 | && ! rtx_varies_p (XEXP (note, 0), 1) | |
3153 | && ! reg_overlap_mentioned_p (SET_DEST (set), | |
3154 | XEXP (note, 0))) | |
3155 | { | |
3156 | set_reg_known_value (regno, XEXP (note, 0)); | |
3157 | set_reg_known_equiv_p (regno, | |
3158 | REG_NOTE_KIND (note) == REG_EQUIV); | |
3159 | } | |
3160 | else if (DF_REG_DEF_COUNT (regno) == 1 | |
3161 | && GET_CODE (src) == PLUS | |
3162 | && REG_P (XEXP (src, 0)) | |
3163 | && (t = get_reg_known_value (REGNO (XEXP (src, 0)))) | |
3164 | && CONST_INT_P (XEXP (src, 1))) | |
3165 | { | |
3166 | t = plus_constant (GET_MODE (src), t, | |
3167 | INTVAL (XEXP (src, 1))); | |
3168 | set_reg_known_value (regno, t); | |
3169 | set_reg_known_equiv_p (regno, false); | |
3170 | } | |
3171 | else if (DF_REG_DEF_COUNT (regno) == 1 | |
3172 | && ! rtx_varies_p (src, 1)) | |
3173 | { | |
3174 | set_reg_known_value (regno, src); | |
3175 | set_reg_known_equiv_p (regno, false); | |
3176 | } | |
970c0c06 | 3177 | } |
35d3eedd | 3178 | } |
2808fda8 | 3179 | else if (NOTE_P (insn) |
3180 | && NOTE_KIND (insn) == NOTE_INSN_FUNCTION_BEG) | |
3181 | copying_arguments = false; | |
06c5c535 | 3182 | } |
35d3eedd | 3183 | } |
06c5c535 | 3184 | |
35d3eedd | 3185 | /* Now propagate values from new_reg_base_value to reg_base_value. */ |
d4473c84 | 3186 | gcc_assert (maxreg == (unsigned int) max_reg_num ()); |
a0c938f0 | 3187 | |
e5b19a8c | 3188 | for (ui = 0; ui < maxreg; ui++) |
35d3eedd | 3189 | { |
274c11d8 | 3190 | if (new_reg_base_value[ui] |
f1f41a6c | 3191 | && new_reg_base_value[ui] != (*reg_base_value)[ui] |
3192 | && ! rtx_equal_p (new_reg_base_value[ui], (*reg_base_value)[ui])) | |
06c5c535 | 3193 | { |
f1f41a6c | 3194 | (*reg_base_value)[ui] = new_reg_base_value[ui]; |
35d3eedd | 3195 | changed = 1; |
06c5c535 | 3196 | } |
ea0cb7ae | 3197 | } |
ea0cb7ae | 3198 | } |
35d3eedd | 3199 | while (changed && ++pass < MAX_ALIAS_LOOP_PASSES); |
2808fda8 | 3200 | XDELETEVEC (rpo); |
ea0cb7ae | 3201 | |
3202 | /* Fill in the remaining entries. */ | |
f1f41a6c | 3203 | FOR_EACH_VEC_ELT (*reg_known_value, i, val) |
d6443ebe | 3204 | { |
3205 | int regno = i + FIRST_PSEUDO_REGISTER; | |
3206 | if (! val) | |
3207 | set_reg_known_value (regno, regno_reg_rtx[regno]); | |
3208 | } | |
ea0cb7ae | 3209 | |
7098ef4a | 3210 | /* Clean up. */ |
3211 | free (new_reg_base_value); | |
06c5c535 | 3212 | new_reg_base_value = 0; |
7bf60644 | 3213 | sbitmap_free (reg_seen); |
ea0cb7ae | 3214 | reg_seen = 0; |
376c21d1 | 3215 | timevar_pop (TV_ALIAS_ANALYSIS); |
ea0cb7ae | 3216 | } |
3217 | ||
1837e966 | 3218 | /* Equate REG_BASE_VALUE (reg1) to REG_BASE_VALUE (reg2). |
3219 | Special API for var-tracking pass purposes. */ | |
3220 | ||
3221 | void | |
3222 | vt_equate_reg_base_value (const_rtx reg1, const_rtx reg2) | |
3223 | { | |
f1f41a6c | 3224 | (*reg_base_value)[REGNO (reg1)] = REG_BASE_VALUE (reg2); |
1837e966 | 3225 | } |
3226 | ||
ea0cb7ae | 3227 | void |
aecda0d6 | 3228 | end_alias_analysis (void) |
ea0cb7ae | 3229 | { |
e5b19a8c | 3230 | old_reg_base_value = reg_base_value; |
f1f41a6c | 3231 | vec_free (reg_known_value); |
d6443ebe | 3232 | sbitmap_free (reg_known_equiv_p); |
ea0cb7ae | 3233 | } |
1f3233d1 | 3234 | |
5b6bf098 | 3235 | void |
3236 | dump_alias_stats_in_alias_c (FILE *s) | |
3237 | { | |
3238 | fprintf (s, " TBAA oracle: %llu disambiguations %llu queries\n" | |
3239 | " %llu are in alias set 0\n" | |
3240 | " %llu queries asked about the same object\n" | |
3241 | " %llu queries asked about the same alias set\n" | |
3242 | " %llu access volatile\n" | |
fdc195e1 | 3243 | " %llu are dependent in the DAG\n" |
3244 | " %llu are aritificially in conflict with void *\n", | |
5b6bf098 | 3245 | alias_stats.num_disambiguated, |
3246 | alias_stats.num_alias_zero + alias_stats.num_same_alias_set | |
3247 | + alias_stats.num_same_objects + alias_stats.num_volatile | |
fdc195e1 | 3248 | + alias_stats.num_dag + alias_stats.num_disambiguated |
3249 | + alias_stats.num_universal, | |
5b6bf098 | 3250 | alias_stats.num_alias_zero, alias_stats.num_same_alias_set, |
fdc195e1 | 3251 | alias_stats.num_same_objects, alias_stats.num_volatile, |
3252 | alias_stats.num_dag, alias_stats.num_universal); | |
5b6bf098 | 3253 | } |
1f3233d1 | 3254 | #include "gt-alias.h" |