]>
Commit | Line | Data |
---|---|---|
9ae8ffe7 | 1 | /* Alias analysis for GNU C |
8d9254fc | 2 | Copyright (C) 1997-2020 Free Software Foundation, Inc. |
9ae8ffe7 JL |
3 | Contributed by John Carr (jfc@mit.edu). |
4 | ||
1322177d | 5 | This file is part of GCC. |
9ae8ffe7 | 6 | |
1322177d LB |
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 | |
9dcd6f09 | 9 | Software Foundation; either version 3, or (at your option) any later |
1322177d | 10 | version. |
9ae8ffe7 | 11 | |
1322177d LB |
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. | |
9ae8ffe7 JL |
16 | |
17 | You should have received a copy of the GNU General Public License | |
9dcd6f09 NC |
18 | along with GCC; see the file COPYING3. If not see |
19 | <http://www.gnu.org/licenses/>. */ | |
9ae8ffe7 JL |
20 | |
21 | #include "config.h" | |
670ee920 | 22 | #include "system.h" |
4977bab6 | 23 | #include "coretypes.h" |
c7131fb2 | 24 | #include "backend.h" |
957060b5 AM |
25 | #include "target.h" |
26 | #include "rtl.h" | |
c7131fb2 AM |
27 | #include "tree.h" |
28 | #include "gimple.h" | |
c7131fb2 | 29 | #include "df.h" |
4d0cdd0c | 30 | #include "memmodel.h" |
957060b5 AM |
31 | #include "tm_p.h" |
32 | #include "gimple-ssa.h" | |
957060b5 | 33 | #include "emit-rtl.h" |
40e23961 | 34 | #include "alias.h" |
40e23961 | 35 | #include "fold-const.h" |
d8a2d370 | 36 | #include "varasm.h" |
eab5c70a | 37 | #include "cselib.h" |
d23c55c2 | 38 | #include "langhooks.h" |
60393bbc | 39 | #include "cfganal.h" |
403837b4 | 40 | #include "rtl-iter.h" |
54363f8a | 41 | #include "cgraph.h" |
e4b44fd7 | 42 | #include "ipa-utils.h" |
ea900239 DB |
43 | |
44 | /* The aliasing API provided here solves related but different problems: | |
45 | ||
c22cacf3 | 46 | Say there exists (in c) |
ea900239 DB |
47 | |
48 | struct X { | |
49 | struct Y y1; | |
50 | struct Z z2; | |
51 | } x1, *px1, *px2; | |
52 | ||
53 | struct Y y2, *py; | |
54 | struct Z z2, *pz; | |
55 | ||
56 | ||
308a3fe2 | 57 | py = &x1.y1; |
ea900239 DB |
58 | px2 = &x1; |
59 | ||
60 | Consider the four questions: | |
61 | ||
62 | Can a store to x1 interfere with px2->y1? | |
63 | Can a store to x1 interfere with px2->z2? | |
ea900239 DB |
64 | Can a store to x1 change the value pointed to by with py? |
65 | Can a store to x1 change the value pointed to by with pz? | |
66 | ||
67 | The answer to these questions can be yes, yes, yes, and maybe. | |
68 | ||
69 | The first two questions can be answered with a simple examination | |
70 | of the type system. If structure X contains a field of type Y then | |
073a8998 | 71 | a store through a pointer to an X can overwrite any field that is |
ea900239 DB |
72 | contained (recursively) in an X (unless we know that px1 != px2). |
73 | ||
308a3fe2 DS |
74 | The last two questions can be solved in the same way as the first |
75 | two questions but this is too conservative. The observation is | |
76 | that in some cases we can know which (if any) fields are addressed | |
77 | and if those addresses are used in bad ways. This analysis may be | |
78 | language specific. In C, arbitrary operations may be applied to | |
79 | pointers. However, there is some indication that this may be too | |
80 | conservative for some C++ types. | |
ea900239 DB |
81 | |
82 | The pass ipa-type-escape does this analysis for the types whose | |
c22cacf3 | 83 | instances do not escape across the compilation boundary. |
ea900239 DB |
84 | |
85 | Historically in GCC, these two problems were combined and a single | |
308a3fe2 | 86 | data structure that was used to represent the solution to these |
ea900239 | 87 | problems. We now have two similar but different data structures, |
308a3fe2 DS |
88 | The data structure to solve the last two questions is similar to |
89 | the first, but does not contain the fields whose address are never | |
90 | taken. For types that do escape the compilation unit, the data | |
91 | structures will have identical information. | |
ea900239 | 92 | */ |
3932261a MM |
93 | |
94 | /* The alias sets assigned to MEMs assist the back-end in determining | |
95 | which MEMs can alias which other MEMs. In general, two MEMs in | |
ac3d9668 RK |
96 | different alias sets cannot alias each other, with one important |
97 | exception. Consider something like: | |
3932261a | 98 | |
01d28c3f | 99 | struct S { int i; double d; }; |
3932261a MM |
100 | |
101 | a store to an `S' can alias something of either type `int' or type | |
102 | `double'. (However, a store to an `int' cannot alias a `double' | |
103 | and vice versa.) We indicate this via a tree structure that looks | |
104 | like: | |
c22cacf3 MS |
105 | struct S |
106 | / \ | |
3932261a | 107 | / \ |
c22cacf3 MS |
108 | |/_ _\| |
109 | int double | |
3932261a | 110 | |
ac3d9668 RK |
111 | (The arrows are directed and point downwards.) |
112 | In this situation we say the alias set for `struct S' is the | |
113 | `superset' and that those for `int' and `double' are `subsets'. | |
114 | ||
3bdf5ad1 RK |
115 | To see whether two alias sets can point to the same memory, we must |
116 | see if either alias set is a subset of the other. We need not trace | |
95bd1dd7 | 117 | past immediate descendants, however, since we propagate all |
3bdf5ad1 | 118 | grandchildren up one level. |
3932261a MM |
119 | |
120 | Alias set zero is implicitly a superset of all other alias sets. | |
121 | However, this is no actual entry for alias set zero. It is an | |
122 | error to attempt to explicitly construct a subset of zero. */ | |
123 | ||
e0702244 | 124 | struct alias_set_hash : int_hash <int, INT_MIN, INT_MIN + 1> {}; |
de144fb2 | 125 | |
02ced957 | 126 | struct GTY(()) alias_set_entry { |
3932261a | 127 | /* The alias set number, as stored in MEM_ALIAS_SET. */ |
4862826d | 128 | alias_set_type alias_set; |
3932261a | 129 | |
6e042ef4 JH |
130 | /* Nonzero if would have a child of zero: this effectively makes this |
131 | alias set the same as alias set zero. */ | |
132 | bool has_zero_child; | |
133 | /* Nonzero if alias set corresponds to pointer type itself (i.e. not to | |
134 | aggregate contaiing pointer. | |
135 | This is used for a special case where we need an universal pointer type | |
136 | compatible with all other pointer types. */ | |
137 | bool is_pointer; | |
138 | /* Nonzero if is_pointer or if one of childs have has_pointer set. */ | |
139 | bool has_pointer; | |
34e82342 RB |
140 | |
141 | /* The children of the alias set. These are not just the immediate | |
142 | children, but, in fact, all descendants. So, if we have: | |
143 | ||
144 | struct T { struct S s; float f; } | |
145 | ||
146 | continuing our example above, the children here will be all of | |
147 | `int', `double', `float', and `struct S'. */ | |
148 | hash_map<alias_set_hash, int> *children; | |
b604074c | 149 | }; |
9ae8ffe7 | 150 | |
ed7a4b4b | 151 | static int rtx_equal_for_memref_p (const_rtx, const_rtx); |
7bc980e1 | 152 | static void record_set (rtx, const_rtx, void *); |
ef4bddc2 RS |
153 | static int base_alias_check (rtx, rtx, rtx, rtx, machine_mode, |
154 | machine_mode); | |
4682ae04 | 155 | static rtx find_base_value (rtx); |
4f588890 | 156 | static int mems_in_disjoint_alias_sets_p (const_rtx, const_rtx); |
02ced957 | 157 | static alias_set_entry *get_alias_set_entry (alias_set_type); |
4682ae04 | 158 | static tree decl_for_component_ref (tree); |
bd280792 | 159 | static int write_dependence_p (const_rtx, |
ef4bddc2 | 160 | const_rtx, machine_mode, rtx, |
bd280792 | 161 | bool, bool, bool); |
73e48cb3 | 162 | static int compare_base_symbol_refs (const_rtx, const_rtx); |
4682ae04 | 163 | |
aa317c97 | 164 | static void memory_modified_1 (rtx, const_rtx, void *); |
9ae8ffe7 | 165 | |
3ecf9d13 JH |
166 | /* Query statistics for the different low-level disambiguators. |
167 | A high-level query may trigger multiple of them. */ | |
168 | ||
169 | static struct { | |
170 | unsigned long long num_alias_zero; | |
171 | unsigned long long num_same_alias_set; | |
172 | unsigned long long num_same_objects; | |
173 | unsigned long long num_volatile; | |
174 | unsigned long long num_dag; | |
6e042ef4 | 175 | unsigned long long num_universal; |
3ecf9d13 JH |
176 | unsigned long long num_disambiguated; |
177 | } alias_stats; | |
178 | ||
179 | ||
9ae8ffe7 JL |
180 | /* Set up all info needed to perform alias analysis on memory references. */ |
181 | ||
d4b60170 | 182 | /* Returns the size in bytes of the mode of X. */ |
9ae8ffe7 JL |
183 | #define SIZE_FOR_MODE(X) (GET_MODE_SIZE (GET_MODE (X))) |
184 | ||
ea64ef27 | 185 | /* Cap the number of passes we make over the insns propagating alias |
131db6b8 SB |
186 | information through set chains. |
187 | ??? 10 is a completely arbitrary choice. This should be based on the | |
188 | maximum loop depth in the CFG, but we do not have this information | |
189 | available (even if current_loops _is_ available). */ | |
ea64ef27 | 190 | #define MAX_ALIAS_LOOP_PASSES 10 |
ca7fd9cd | 191 | |
9ae8ffe7 JL |
192 | /* reg_base_value[N] gives an address to which register N is related. |
193 | If all sets after the first add or subtract to the current value | |
194 | or otherwise modify it so it does not point to a different top level | |
195 | object, reg_base_value[N] is equal to the address part of the source | |
2a2c8203 JC |
196 | of the first set. |
197 | ||
198 | A base address can be an ADDRESS, SYMBOL_REF, or LABEL_REF. ADDRESS | |
9fc37b2b | 199 | expressions represent three types of base: |
b3b5ad95 | 200 | |
9fc37b2b RS |
201 | 1. incoming arguments. There is just one ADDRESS to represent all |
202 | arguments, since we do not know at this level whether accesses | |
203 | based on different arguments can alias. The ADDRESS has id 0. | |
b3b5ad95 | 204 | |
9fc37b2b RS |
205 | 2. stack_pointer_rtx, frame_pointer_rtx, hard_frame_pointer_rtx |
206 | (if distinct from frame_pointer_rtx) and arg_pointer_rtx. | |
207 | Each of these rtxes has a separate ADDRESS associated with it, | |
208 | each with a negative id. | |
209 | ||
210 | GCC is (and is required to be) precise in which register it | |
211 | chooses to access a particular region of stack. We can therefore | |
212 | assume that accesses based on one of these rtxes do not alias | |
213 | accesses based on another of these rtxes. | |
214 | ||
215 | 3. bases that are derived from malloc()ed memory (REG_NOALIAS). | |
216 | Each such piece of memory has a separate ADDRESS associated | |
217 | with it, each with an id greater than 0. | |
218 | ||
219 | Accesses based on one ADDRESS do not alias accesses based on other | |
220 | ADDRESSes. Accesses based on ADDRESSes in groups (2) and (3) do not | |
221 | alias globals either; the ADDRESSes have Pmode to indicate this. | |
222 | The ADDRESS in group (1) _may_ alias globals; it has VOIDmode to | |
223 | indicate this. */ | |
2a2c8203 | 224 | |
9771b263 | 225 | static GTY(()) vec<rtx, va_gc> *reg_base_value; |
ac606739 | 226 | static rtx *new_reg_base_value; |
c582d54a | 227 | |
9fc37b2b RS |
228 | /* The single VOIDmode ADDRESS that represents all argument bases. |
229 | It has id 0. */ | |
230 | static GTY(()) rtx arg_base_value; | |
231 | ||
232 | /* Used to allocate unique ids to each REG_NOALIAS ADDRESS. */ | |
233 | static int unique_id; | |
234 | ||
c582d54a JH |
235 | /* We preserve the copy of old array around to avoid amount of garbage |
236 | produced. About 8% of garbage produced were attributed to this | |
237 | array. */ | |
9771b263 | 238 | static GTY((deletable)) vec<rtx, va_gc> *old_reg_base_value; |
d4b60170 | 239 | |
9e412ca3 RS |
240 | /* Values of XINT (address, 0) of Pmode ADDRESS rtxes for special |
241 | registers. */ | |
242 | #define UNIQUE_BASE_VALUE_SP -1 | |
243 | #define UNIQUE_BASE_VALUE_ARGP -2 | |
244 | #define UNIQUE_BASE_VALUE_FP -3 | |
245 | #define UNIQUE_BASE_VALUE_HFP -4 | |
246 | ||
7bf84454 RS |
247 | #define static_reg_base_value \ |
248 | (this_target_rtl->x_static_reg_base_value) | |
bf1660a6 | 249 | |
9771b263 DN |
250 | #define REG_BASE_VALUE(X) \ |
251 | (REGNO (X) < vec_safe_length (reg_base_value) \ | |
252 | ? (*reg_base_value)[REGNO (X)] : 0) | |
9ae8ffe7 | 253 | |
c13e8210 | 254 | /* Vector indexed by N giving the initial (unchanging) value known for |
9ff3c7ca | 255 | pseudo-register N. This vector is initialized in init_alias_analysis, |
bb1acb3e | 256 | and does not change until end_alias_analysis is called. */ |
9771b263 | 257 | static GTY(()) vec<rtx, va_gc> *reg_known_value; |
9ae8ffe7 JL |
258 | |
259 | /* Vector recording for each reg_known_value whether it is due to a | |
260 | REG_EQUIV note. Future passes (viz., reload) may replace the | |
261 | pseudo with the equivalent expression and so we account for the | |
ac3d9668 RK |
262 | dependences that would be introduced if that happens. |
263 | ||
264 | The REG_EQUIV notes created in assign_parms may mention the arg | |
265 | pointer, and there are explicit insns in the RTL that modify the | |
266 | arg pointer. Thus we must ensure that such insns don't get | |
267 | scheduled across each other because that would invalidate the | |
268 | REG_EQUIV notes. One could argue that the REG_EQUIV notes are | |
269 | wrong, but solving the problem in the scheduler will likely give | |
270 | better code, so we do it here. */ | |
9ff3c7ca | 271 | static sbitmap reg_known_equiv_p; |
9ae8ffe7 | 272 | |
2a2c8203 JC |
273 | /* True when scanning insns from the start of the rtl to the |
274 | NOTE_INSN_FUNCTION_BEG note. */ | |
83bbd9b6 | 275 | static bool copying_arguments; |
9ae8ffe7 | 276 | |
1a5640b4 | 277 | |
3932261a | 278 | /* The splay-tree used to store the various alias set entries. */ |
02ced957 | 279 | static GTY (()) vec<alias_set_entry *, va_gc> *alias_sets; |
ac3d9668 | 280 | \f |
55b34b5f RG |
281 | /* Build a decomposed reference object for querying the alias-oracle |
282 | from the MEM rtx and store it in *REF. | |
283 | Returns false if MEM is not suitable for the alias-oracle. */ | |
284 | ||
285 | static bool | |
286 | ao_ref_from_mem (ao_ref *ref, const_rtx mem) | |
287 | { | |
288 | tree expr = MEM_EXPR (mem); | |
289 | tree base; | |
290 | ||
291 | if (!expr) | |
292 | return false; | |
293 | ||
294 | ao_ref_init (ref, expr); | |
295 | ||
296 | /* Get the base of the reference and see if we have to reject or | |
297 | adjust it. */ | |
298 | base = ao_ref_base (ref); | |
299 | if (base == NULL_TREE) | |
300 | return false; | |
301 | ||
ef7a9fb8 RB |
302 | /* The tree oracle doesn't like bases that are neither decls |
303 | nor indirect references of SSA names. */ | |
304 | if (!(DECL_P (base) | |
305 | || (TREE_CODE (base) == MEM_REF | |
306 | && TREE_CODE (TREE_OPERAND (base, 0)) == SSA_NAME) | |
307 | || (TREE_CODE (base) == TARGET_MEM_REF | |
308 | && TREE_CODE (TMR_BASE (base)) == SSA_NAME))) | |
d15adbeb | 309 | return false; |
55b34b5f | 310 | |
55b34b5f RG |
311 | ref->ref_alias_set = MEM_ALIAS_SET (mem); |
312 | ||
f68396a1 RG |
313 | /* If MEM_OFFSET or MEM_SIZE are unknown what we got from MEM_EXPR |
314 | is conservative, so trust it. */ | |
527210c4 | 315 | if (!MEM_OFFSET_KNOWN_P (mem) |
f5541398 | 316 | || !MEM_SIZE_KNOWN_P (mem)) |
f68396a1 | 317 | return true; |
366f945f | 318 | |
e8024441 RB |
319 | /* If MEM_OFFSET/MEM_SIZE get us outside of ref->offset/ref->max_size |
320 | drop ref->ref. */ | |
d05d7551 | 321 | if (maybe_lt (MEM_OFFSET (mem), 0) |
b9c25734 RS |
322 | || (ref->max_size_known_p () |
323 | && maybe_gt ((MEM_OFFSET (mem) + MEM_SIZE (mem)) * BITS_PER_UNIT, | |
324 | ref->max_size))) | |
e8024441 | 325 | ref->ref = NULL_TREE; |
b0e96404 | 326 | |
e8024441 RB |
327 | /* Refine size and offset we got from analyzing MEM_EXPR by using |
328 | MEM_SIZE and MEM_OFFSET. */ | |
f68396a1 | 329 | |
527210c4 | 330 | ref->offset += MEM_OFFSET (mem) * BITS_PER_UNIT; |
f5541398 | 331 | ref->size = MEM_SIZE (mem) * BITS_PER_UNIT; |
b0e96404 RG |
332 | |
333 | /* The MEM may extend into adjacent fields, so adjust max_size if | |
334 | necessary. */ | |
b9c25734 RS |
335 | if (ref->max_size_known_p ()) |
336 | ref->max_size = upper_bound (ref->max_size, ref->size); | |
b0e96404 | 337 | |
b9c25734 | 338 | /* If MEM_OFFSET and MEM_SIZE might get us outside of the base object of |
b0e96404 RG |
339 | the MEM_EXPR punt. This happens for STRICT_ALIGNMENT targets a lot. */ |
340 | if (MEM_EXPR (mem) != get_spill_slot_decl (false) | |
b9c25734 | 341 | && (maybe_lt (ref->offset, 0) |
b0e96404 | 342 | || (DECL_P (ref->base) |
807e902e | 343 | && (DECL_SIZE (ref->base) == NULL_TREE |
b9c25734 RS |
344 | || !poly_int_tree_p (DECL_SIZE (ref->base)) |
345 | || maybe_lt (wi::to_poly_offset (DECL_SIZE (ref->base)), | |
346 | ref->offset + ref->size))))) | |
b0e96404 | 347 | return false; |
55b34b5f RG |
348 | |
349 | return true; | |
350 | } | |
351 | ||
352 | /* Query the alias-oracle on whether the two memory rtx X and MEM may | |
353 | alias. If TBAA_P is set also apply TBAA. Returns true if the | |
354 | two rtxen may alias, false otherwise. */ | |
355 | ||
356 | static bool | |
357 | rtx_refs_may_alias_p (const_rtx x, const_rtx mem, bool tbaa_p) | |
358 | { | |
359 | ao_ref ref1, ref2; | |
360 | ||
361 | if (!ao_ref_from_mem (&ref1, x) | |
362 | || !ao_ref_from_mem (&ref2, mem)) | |
363 | return true; | |
364 | ||
55e3bc4c RG |
365 | return refs_may_alias_p_1 (&ref1, &ref2, |
366 | tbaa_p | |
367 | && MEM_ALIAS_SET (x) != 0 | |
368 | && MEM_ALIAS_SET (mem) != 0); | |
55b34b5f RG |
369 | } |
370 | ||
3d6fd7ce RB |
371 | /* Return true if the ref EARLIER behaves the same as LATER with respect |
372 | to TBAA for every memory reference that might follow LATER. */ | |
373 | ||
374 | bool | |
375 | refs_same_for_tbaa_p (tree earlier, tree later) | |
376 | { | |
377 | ao_ref earlier_ref, later_ref; | |
378 | ao_ref_init (&earlier_ref, earlier); | |
379 | ao_ref_init (&later_ref, later); | |
380 | alias_set_type earlier_set = ao_ref_alias_set (&earlier_ref); | |
381 | alias_set_type later_set = ao_ref_alias_set (&later_ref); | |
382 | if (!(earlier_set == later_set | |
383 | || alias_set_subset_of (later_set, earlier_set))) | |
384 | return false; | |
385 | alias_set_type later_base_set = ao_ref_base_alias_set (&later_ref); | |
386 | alias_set_type earlier_base_set = ao_ref_base_alias_set (&earlier_ref); | |
387 | return (earlier_base_set == later_base_set | |
388 | || alias_set_subset_of (later_base_set, earlier_base_set)); | |
389 | } | |
390 | ||
3932261a MM |
391 | /* Returns a pointer to the alias set entry for ALIAS_SET, if there is |
392 | such an entry, or NULL otherwise. */ | |
393 | ||
02ced957 | 394 | static inline alias_set_entry * |
4862826d | 395 | get_alias_set_entry (alias_set_type alias_set) |
3932261a | 396 | { |
9771b263 | 397 | return (*alias_sets)[alias_set]; |
3932261a MM |
398 | } |
399 | ||
ac3d9668 RK |
400 | /* Returns nonzero if the alias sets for MEM1 and MEM2 are such that |
401 | the two MEMs cannot alias each other. */ | |
3932261a | 402 | |
9ddb66ca | 403 | static inline int |
4f588890 | 404 | mems_in_disjoint_alias_sets_p (const_rtx mem1, const_rtx mem2) |
3932261a | 405 | { |
598f8eca RB |
406 | return (flag_strict_aliasing |
407 | && ! alias_sets_conflict_p (MEM_ALIAS_SET (mem1), | |
408 | MEM_ALIAS_SET (mem2))); | |
1da68f56 | 409 | } |
3932261a | 410 | |
c58936b6 DB |
411 | /* Return true if the first alias set is a subset of the second. */ |
412 | ||
413 | bool | |
4862826d | 414 | alias_set_subset_of (alias_set_type set1, alias_set_type set2) |
c58936b6 | 415 | { |
02ced957 | 416 | alias_set_entry *ase2; |
c58936b6 | 417 | |
bd04cddf JH |
418 | /* Disable TBAA oracle with !flag_strict_aliasing. */ |
419 | if (!flag_strict_aliasing) | |
420 | return true; | |
421 | ||
c58936b6 DB |
422 | /* Everything is a subset of the "aliases everything" set. */ |
423 | if (set2 == 0) | |
424 | return true; | |
425 | ||
6e042ef4 JH |
426 | /* Check if set1 is a subset of set2. */ |
427 | ase2 = get_alias_set_entry (set2); | |
428 | if (ase2 != 0 | |
429 | && (ase2->has_zero_child | |
430 | || (ase2->children && ase2->children->get (set1)))) | |
c58936b6 | 431 | return true; |
6e042ef4 JH |
432 | |
433 | /* As a special case we consider alias set of "void *" to be both subset | |
434 | and superset of every alias set of a pointer. This extra symmetry does | |
435 | not matter for alias_sets_conflict_p but it makes aliasing_component_refs_p | |
436 | to return true on the following testcase: | |
437 | ||
438 | void *ptr; | |
439 | char **ptr2=(char **)&ptr; | |
440 | *ptr2 = ... | |
441 | ||
442 | Additionally if a set contains universal pointer, we consider every pointer | |
443 | to be a subset of it, but we do not represent this explicitely - doing so | |
444 | would require us to update transitive closure each time we introduce new | |
445 | pointer type. This makes aliasing_component_refs_p to return true | |
446 | on the following testcase: | |
447 | ||
448 | struct a {void *ptr;} | |
449 | char **ptr = (char **)&a.ptr; | |
450 | ptr = ... | |
451 | ||
452 | This makes void * truly universal pointer type. See pointer handling in | |
453 | get_alias_set for more details. */ | |
454 | if (ase2 && ase2->has_pointer) | |
455 | { | |
02ced957 | 456 | alias_set_entry *ase1 = get_alias_set_entry (set1); |
6e042ef4 JH |
457 | |
458 | if (ase1 && ase1->is_pointer) | |
459 | { | |
460 | alias_set_type voidptr_set = TYPE_ALIAS_SET (ptr_type_node); | |
461 | /* If one is ptr_type_node and other is pointer, then we consider | |
462 | them subset of each other. */ | |
463 | if (set1 == voidptr_set || set2 == voidptr_set) | |
464 | return true; | |
465 | /* If SET2 contains universal pointer's alias set, then we consdier | |
466 | every (non-universal) pointer. */ | |
467 | if (ase2->children && set1 != voidptr_set | |
468 | && ase2->children->get (voidptr_set)) | |
469 | return true; | |
470 | } | |
471 | } | |
c58936b6 DB |
472 | return false; |
473 | } | |
474 | ||
1da68f56 RK |
475 | /* Return 1 if the two specified alias sets may conflict. */ |
476 | ||
477 | int | |
4862826d | 478 | alias_sets_conflict_p (alias_set_type set1, alias_set_type set2) |
1da68f56 | 479 | { |
02ced957 TS |
480 | alias_set_entry *ase1; |
481 | alias_set_entry *ase2; | |
1da68f56 | 482 | |
836f7794 EB |
483 | /* The easy case. */ |
484 | if (alias_sets_must_conflict_p (set1, set2)) | |
1da68f56 | 485 | return 1; |
3932261a | 486 | |
3bdf5ad1 | 487 | /* See if the first alias set is a subset of the second. */ |
6e042ef4 JH |
488 | ase1 = get_alias_set_entry (set1); |
489 | if (ase1 != 0 | |
490 | && ase1->children && ase1->children->get (set2)) | |
3ecf9d13 JH |
491 | { |
492 | ++alias_stats.num_dag; | |
493 | return 1; | |
494 | } | |
3932261a MM |
495 | |
496 | /* Now do the same, but with the alias sets reversed. */ | |
6e042ef4 JH |
497 | ase2 = get_alias_set_entry (set2); |
498 | if (ase2 != 0 | |
499 | && ase2->children && ase2->children->get (set1)) | |
3ecf9d13 JH |
500 | { |
501 | ++alias_stats.num_dag; | |
502 | return 1; | |
503 | } | |
6e042ef4 JH |
504 | |
505 | /* We want void * to be compatible with any other pointer without | |
506 | really dropping it to alias set 0. Doing so would make it | |
507 | compatible with all non-pointer types too. | |
508 | ||
509 | This is not strictly necessary by the C/C++ language | |
510 | standards, but avoids common type punning mistakes. In | |
511 | addition to that, we need the existence of such universal | |
512 | pointer to implement Fortran's C_PTR type (which is defined as | |
513 | type compatible with all C pointers). */ | |
514 | if (ase1 && ase2 && ase1->has_pointer && ase2->has_pointer) | |
515 | { | |
516 | alias_set_type voidptr_set = TYPE_ALIAS_SET (ptr_type_node); | |
517 | ||
518 | /* If one of the sets corresponds to universal pointer, | |
519 | we consider it to conflict with anything that is | |
520 | or contains pointer. */ | |
521 | if (set1 == voidptr_set || set2 == voidptr_set) | |
522 | { | |
523 | ++alias_stats.num_universal; | |
524 | return true; | |
525 | } | |
526 | /* If one of sets is (non-universal) pointer and the other | |
527 | contains universal pointer, we also get conflict. */ | |
528 | if (ase1->is_pointer && set2 != voidptr_set | |
529 | && ase2->children && ase2->children->get (voidptr_set)) | |
530 | { | |
531 | ++alias_stats.num_universal; | |
532 | return true; | |
533 | } | |
534 | if (ase2->is_pointer && set1 != voidptr_set | |
535 | && ase1->children && ase1->children->get (voidptr_set)) | |
536 | { | |
537 | ++alias_stats.num_universal; | |
538 | return true; | |
539 | } | |
540 | } | |
541 | ||
3ecf9d13 | 542 | ++alias_stats.num_disambiguated; |
3932261a | 543 | |
1da68f56 | 544 | /* The two alias sets are distinct and neither one is the |
836f7794 | 545 | child of the other. Therefore, they cannot conflict. */ |
1da68f56 | 546 | return 0; |
3932261a | 547 | } |
5399d643 | 548 | |
836f7794 | 549 | /* Return 1 if the two specified alias sets will always conflict. */ |
5399d643 JW |
550 | |
551 | int | |
4862826d | 552 | alias_sets_must_conflict_p (alias_set_type set1, alias_set_type set2) |
5399d643 | 553 | { |
bd04cddf JH |
554 | /* Disable TBAA oracle with !flag_strict_aliasing. */ |
555 | if (!flag_strict_aliasing) | |
556 | return 1; | |
3ecf9d13 JH |
557 | if (set1 == 0 || set2 == 0) |
558 | { | |
559 | ++alias_stats.num_alias_zero; | |
560 | return 1; | |
561 | } | |
562 | if (set1 == set2) | |
563 | { | |
564 | ++alias_stats.num_same_alias_set; | |
565 | return 1; | |
566 | } | |
5399d643 JW |
567 | |
568 | return 0; | |
569 | } | |
570 | ||
1da68f56 RK |
571 | /* Return 1 if any MEM object of type T1 will always conflict (using the |
572 | dependency routines in this file) with any MEM object of type T2. | |
573 | This is used when allocating temporary storage. If T1 and/or T2 are | |
574 | NULL_TREE, it means we know nothing about the storage. */ | |
575 | ||
576 | int | |
4682ae04 | 577 | objects_must_conflict_p (tree t1, tree t2) |
1da68f56 | 578 | { |
4862826d | 579 | alias_set_type set1, set2; |
82d610ec | 580 | |
e8ea2809 RK |
581 | /* If neither has a type specified, we don't know if they'll conflict |
582 | because we may be using them to store objects of various types, for | |
583 | example the argument and local variables areas of inlined functions. */ | |
981a4c34 | 584 | if (t1 == 0 && t2 == 0) |
e8ea2809 RK |
585 | return 0; |
586 | ||
1da68f56 | 587 | /* If they are the same type, they must conflict. */ |
3ecf9d13 JH |
588 | if (t1 == t2) |
589 | { | |
590 | ++alias_stats.num_same_objects; | |
591 | return 1; | |
592 | } | |
593 | /* Likewise if both are volatile. */ | |
594 | if (t1 != 0 && TYPE_VOLATILE (t1) && t2 != 0 && TYPE_VOLATILE (t2)) | |
595 | { | |
596 | ++alias_stats.num_volatile; | |
597 | return 1; | |
598 | } | |
1da68f56 | 599 | |
82d610ec RK |
600 | set1 = t1 ? get_alias_set (t1) : 0; |
601 | set2 = t2 ? get_alias_set (t2) : 0; | |
1da68f56 | 602 | |
836f7794 EB |
603 | /* We can't use alias_sets_conflict_p because we must make sure |
604 | that every subtype of t1 will conflict with every subtype of | |
82d610ec RK |
605 | t2 for which a pair of subobjects of these respective subtypes |
606 | overlaps on the stack. */ | |
836f7794 | 607 | return alias_sets_must_conflict_p (set1, set2); |
1da68f56 RK |
608 | } |
609 | \f | |
9640ff5a JH |
610 | /* Return true if T is an end of the access path which can be used |
611 | by type based alias oracle. */ | |
612 | ||
613 | bool | |
614 | ends_tbaa_access_path_p (const_tree t) | |
615 | { | |
616 | switch (TREE_CODE (t)) | |
617 | { | |
618 | case COMPONENT_REF: | |
619 | if (DECL_NONADDRESSABLE_P (TREE_OPERAND (t, 1))) | |
620 | return true; | |
621 | /* Permit type-punning when accessing a union, provided the access | |
622 | is directly through the union. For example, this code does not | |
623 | permit taking the address of a union member and then storing | |
624 | through it. Even the type-punning allowed here is a GCC | |
625 | extension, albeit a common and useful one; the C standard says | |
626 | that such accesses have implementation-defined behavior. */ | |
627 | else if (TREE_CODE (TREE_TYPE (TREE_OPERAND (t, 0))) == UNION_TYPE) | |
628 | return true; | |
629 | break; | |
630 | ||
631 | case ARRAY_REF: | |
632 | case ARRAY_RANGE_REF: | |
633 | if (TYPE_NONALIASED_COMPONENT (TREE_TYPE (TREE_OPERAND (t, 0)))) | |
634 | return true; | |
635 | break; | |
636 | ||
637 | case REALPART_EXPR: | |
638 | case IMAGPART_EXPR: | |
639 | break; | |
640 | ||
641 | case BIT_FIELD_REF: | |
642 | case VIEW_CONVERT_EXPR: | |
643 | /* Bitfields and casts are never addressable. */ | |
644 | return true; | |
645 | break; | |
646 | ||
647 | default: | |
648 | gcc_unreachable (); | |
649 | } | |
650 | return false; | |
651 | } | |
652 | ||
b4ada065 RB |
653 | /* Return the outermost parent of component present in the chain of |
654 | component references handled by get_inner_reference in T with the | |
655 | following property: | |
dc3221e1 | 656 | - the component is non-addressable |
b4ada065 RB |
657 | or NULL_TREE if no such parent exists. In the former cases, the alias |
658 | set of this parent is the alias set that must be used for T itself. */ | |
659 | ||
660 | tree | |
661 | component_uses_parent_alias_set_from (const_tree t) | |
6e24b709 | 662 | { |
b4ada065 | 663 | const_tree found = NULL_TREE; |
afe84921 | 664 | |
b4ada065 RB |
665 | while (handled_component_p (t)) |
666 | { | |
9640ff5a JH |
667 | if (ends_tbaa_access_path_p (t)) |
668 | found = t; | |
afe84921 RH |
669 | |
670 | t = TREE_OPERAND (t, 0); | |
671 | } | |
b4ada065 RB |
672 | |
673 | if (found) | |
674 | return TREE_OPERAND (found, 0); | |
675 | ||
676 | return NULL_TREE; | |
6e24b709 RK |
677 | } |
678 | ||
f40333af RB |
679 | |
680 | /* Return whether the pointer-type T effective for aliasing may | |
681 | access everything and thus the reference has to be assigned | |
682 | alias-set zero. */ | |
683 | ||
684 | static bool | |
685 | ref_all_alias_ptr_type_p (const_tree t) | |
686 | { | |
687 | return (TREE_CODE (TREE_TYPE (t)) == VOID_TYPE | |
688 | || TYPE_REF_CAN_ALIAS_ALL (t)); | |
689 | } | |
690 | ||
5006671f RG |
691 | /* Return the alias set for the memory pointed to by T, which may be |
692 | either a type or an expression. Return -1 if there is nothing | |
693 | special about dereferencing T. */ | |
694 | ||
695 | static alias_set_type | |
696 | get_deref_alias_set_1 (tree t) | |
697 | { | |
5b21f0f3 | 698 | /* All we care about is the type. */ |
5006671f | 699 | if (! TYPE_P (t)) |
5b21f0f3 | 700 | t = TREE_TYPE (t); |
5006671f RG |
701 | |
702 | /* If we have an INDIRECT_REF via a void pointer, we don't | |
703 | know anything about what that might alias. Likewise if the | |
704 | pointer is marked that way. */ | |
f40333af | 705 | if (ref_all_alias_ptr_type_p (t)) |
5006671f RG |
706 | return 0; |
707 | ||
708 | return -1; | |
709 | } | |
710 | ||
711 | /* Return the alias set for the memory pointed to by T, which may be | |
712 | either a type or an expression. */ | |
713 | ||
714 | alias_set_type | |
715 | get_deref_alias_set (tree t) | |
716 | { | |
f40333af RB |
717 | /* If we're not doing any alias analysis, just assume everything |
718 | aliases everything else. */ | |
719 | if (!flag_strict_aliasing) | |
720 | return 0; | |
721 | ||
5006671f RG |
722 | alias_set_type set = get_deref_alias_set_1 (t); |
723 | ||
724 | /* Fall back to the alias-set of the pointed-to type. */ | |
725 | if (set == -1) | |
726 | { | |
727 | if (! TYPE_P (t)) | |
728 | t = TREE_TYPE (t); | |
729 | set = get_alias_set (TREE_TYPE (t)); | |
730 | } | |
731 | ||
732 | return set; | |
733 | } | |
734 | ||
f40333af RB |
735 | /* Return the pointer-type relevant for TBAA purposes from the |
736 | memory reference tree *T or NULL_TREE in which case *T is | |
737 | adjusted to point to the outermost component reference that | |
738 | can be used for assigning an alias set. */ | |
739 | ||
740 | static tree | |
741 | reference_alias_ptr_type_1 (tree *t) | |
742 | { | |
743 | tree inner; | |
744 | ||
745 | /* Get the base object of the reference. */ | |
746 | inner = *t; | |
747 | while (handled_component_p (inner)) | |
748 | { | |
749 | /* If there is a VIEW_CONVERT_EXPR in the chain we cannot use | |
750 | the type of any component references that wrap it to | |
751 | determine the alias-set. */ | |
752 | if (TREE_CODE (inner) == VIEW_CONVERT_EXPR) | |
753 | *t = TREE_OPERAND (inner, 0); | |
754 | inner = TREE_OPERAND (inner, 0); | |
755 | } | |
756 | ||
757 | /* Handle pointer dereferences here, they can override the | |
758 | alias-set. */ | |
759 | if (INDIRECT_REF_P (inner) | |
760 | && ref_all_alias_ptr_type_p (TREE_TYPE (TREE_OPERAND (inner, 0)))) | |
761 | return TREE_TYPE (TREE_OPERAND (inner, 0)); | |
762 | else if (TREE_CODE (inner) == TARGET_MEM_REF) | |
763 | return TREE_TYPE (TMR_OFFSET (inner)); | |
764 | else if (TREE_CODE (inner) == MEM_REF | |
765 | && ref_all_alias_ptr_type_p (TREE_TYPE (TREE_OPERAND (inner, 1)))) | |
766 | return TREE_TYPE (TREE_OPERAND (inner, 1)); | |
767 | ||
768 | /* If the innermost reference is a MEM_REF that has a | |
769 | conversion embedded treat it like a VIEW_CONVERT_EXPR above, | |
770 | using the memory access type for determining the alias-set. */ | |
771 | if (TREE_CODE (inner) == MEM_REF | |
772 | && (TYPE_MAIN_VARIANT (TREE_TYPE (inner)) | |
773 | != TYPE_MAIN_VARIANT | |
774 | (TREE_TYPE (TREE_TYPE (TREE_OPERAND (inner, 1)))))) | |
775 | return TREE_TYPE (TREE_OPERAND (inner, 1)); | |
776 | ||
b4ada065 RB |
777 | /* Otherwise, pick up the outermost object that we could have |
778 | a pointer to. */ | |
779 | tree tem = component_uses_parent_alias_set_from (*t); | |
780 | if (tem) | |
781 | *t = tem; | |
f40333af RB |
782 | |
783 | return NULL_TREE; | |
784 | } | |
785 | ||
786 | /* Return the pointer-type relevant for TBAA purposes from the | |
787 | gimple memory reference tree T. This is the type to be used for | |
788 | the offset operand of MEM_REF or TARGET_MEM_REF replacements of T | |
789 | and guarantees that get_alias_set will return the same alias | |
790 | set for T and the replacement. */ | |
791 | ||
792 | tree | |
793 | reference_alias_ptr_type (tree t) | |
794 | { | |
ebc1b29e RB |
795 | /* If the frontend assigns this alias-set zero, preserve that. */ |
796 | if (lang_hooks.get_alias_set (t) == 0) | |
797 | return ptr_type_node; | |
798 | ||
f40333af RB |
799 | tree ptype = reference_alias_ptr_type_1 (&t); |
800 | /* If there is a given pointer type for aliasing purposes, return it. */ | |
801 | if (ptype != NULL_TREE) | |
802 | return ptype; | |
803 | ||
804 | /* Otherwise build one from the outermost component reference we | |
805 | may use. */ | |
806 | if (TREE_CODE (t) == MEM_REF | |
807 | || TREE_CODE (t) == TARGET_MEM_REF) | |
808 | return TREE_TYPE (TREE_OPERAND (t, 1)); | |
809 | else | |
810 | return build_pointer_type (TYPE_MAIN_VARIANT (TREE_TYPE (t))); | |
811 | } | |
812 | ||
813 | /* Return whether the pointer-types T1 and T2 used to determine | |
814 | two alias sets of two references will yield the same answer | |
815 | from get_deref_alias_set. */ | |
816 | ||
817 | bool | |
818 | alias_ptr_types_compatible_p (tree t1, tree t2) | |
819 | { | |
820 | if (TYPE_MAIN_VARIANT (t1) == TYPE_MAIN_VARIANT (t2)) | |
821 | return true; | |
822 | ||
823 | if (ref_all_alias_ptr_type_p (t1) | |
824 | || ref_all_alias_ptr_type_p (t2)) | |
825 | return false; | |
826 | ||
33e8e0ee ML |
827 | /* This function originally abstracts from simply comparing |
828 | get_deref_alias_set so that we are sure this still computes | |
829 | the same result after LTO type merging is applied. | |
830 | When in LTO type merging is done we can actually do this compare. | |
831 | */ | |
832 | if (in_lto_p) | |
833 | return get_deref_alias_set (t1) == get_deref_alias_set (t2); | |
834 | else | |
835 | return (TYPE_MAIN_VARIANT (TREE_TYPE (t1)) | |
836 | == TYPE_MAIN_VARIANT (TREE_TYPE (t2))); | |
f40333af RB |
837 | } |
838 | ||
6e042ef4 JH |
839 | /* Create emptry alias set entry. */ |
840 | ||
02ced957 | 841 | alias_set_entry * |
6e042ef4 JH |
842 | init_alias_set_entry (alias_set_type set) |
843 | { | |
02ced957 | 844 | alias_set_entry *ase = ggc_alloc<alias_set_entry> (); |
6e042ef4 JH |
845 | ase->alias_set = set; |
846 | ase->children = NULL; | |
847 | ase->has_zero_child = false; | |
848 | ase->is_pointer = false; | |
849 | ase->has_pointer = false; | |
850 | gcc_checking_assert (!get_alias_set_entry (set)); | |
851 | (*alias_sets)[set] = ase; | |
852 | return ase; | |
853 | } | |
854 | ||
3bdf5ad1 RK |
855 | /* Return the alias set for T, which may be either a type or an |
856 | expression. Call language-specific routine for help, if needed. */ | |
857 | ||
4862826d | 858 | alias_set_type |
4682ae04 | 859 | get_alias_set (tree t) |
3bdf5ad1 | 860 | { |
4862826d | 861 | alias_set_type set; |
3bdf5ad1 | 862 | |
67914693 | 863 | /* We cannot give up with -fno-strict-aliasing because we need to build |
63f56527 | 864 | proper type representations for possible functions which are built with |
e8444ca6 | 865 | -fstrict-aliasing. */ |
bd04cddf JH |
866 | |
867 | /* return 0 if this or its type is an error. */ | |
868 | if (t == error_mark_node | |
3bdf5ad1 RK |
869 | || (! TYPE_P (t) |
870 | && (TREE_TYPE (t) == 0 || TREE_TYPE (t) == error_mark_node))) | |
871 | return 0; | |
872 | ||
873 | /* We can be passed either an expression or a type. This and the | |
f47e9b4e RK |
874 | language-specific routine may make mutually-recursive calls to each other |
875 | to figure out what to do. At each juncture, we see if this is a tree | |
876 | that the language may need to handle specially. First handle things that | |
738cc472 | 877 | aren't types. */ |
f824e5c3 | 878 | if (! TYPE_P (t)) |
3bdf5ad1 | 879 | { |
70f34814 RG |
880 | /* Give the language a chance to do something with this tree |
881 | before we look at it. */ | |
8ac61af7 | 882 | STRIP_NOPS (t); |
ae2bcd98 | 883 | set = lang_hooks.get_alias_set (t); |
8ac61af7 RK |
884 | if (set != -1) |
885 | return set; | |
886 | ||
f40333af RB |
887 | /* Get the alias pointer-type to use or the outermost object |
888 | that we could have a pointer to. */ | |
889 | tree ptype = reference_alias_ptr_type_1 (&t); | |
890 | if (ptype != NULL) | |
891 | return get_deref_alias_set (ptype); | |
f824e5c3 | 892 | |
738cc472 RK |
893 | /* If we've already determined the alias set for a decl, just return |
894 | it. This is necessary for C++ anonymous unions, whose component | |
895 | variables don't look like union members (boo!). */ | |
8813a647 | 896 | if (VAR_P (t) |
3c0cb5de | 897 | && DECL_RTL_SET_P (t) && MEM_P (DECL_RTL (t))) |
5755cd38 JM |
898 | return MEM_ALIAS_SET (DECL_RTL (t)); |
899 | ||
f824e5c3 RK |
900 | /* Now all we care about is the type. */ |
901 | t = TREE_TYPE (t); | |
3bdf5ad1 RK |
902 | } |
903 | ||
3bdf5ad1 | 904 | /* Variant qualifiers don't affect the alias set, so get the main |
daad0278 | 905 | variant. */ |
3bdf5ad1 | 906 | t = TYPE_MAIN_VARIANT (t); |
daad0278 | 907 | |
350792ff RB |
908 | if (AGGREGATE_TYPE_P (t) |
909 | && TYPE_TYPELESS_STORAGE (t)) | |
910 | return 0; | |
911 | ||
daad0278 RG |
912 | /* Always use the canonical type as well. If this is a type that |
913 | requires structural comparisons to identify compatible types | |
914 | use alias set zero. */ | |
915 | if (TYPE_STRUCTURAL_EQUALITY_P (t)) | |
cb9c2485 JM |
916 | { |
917 | /* Allow the language to specify another alias set for this | |
918 | type. */ | |
919 | set = lang_hooks.get_alias_set (t); | |
920 | if (set != -1) | |
921 | return set; | |
aea50b45 | 922 | /* Handle structure type equality for pointer types, arrays and vectors. |
63f56527 | 923 | This is easy to do, because the code below ignores canonical types on |
aea50b45 | 924 | these anyway. This is important for LTO, where TYPE_CANONICAL for |
63f56527 | 925 | pointers cannot be meaningfully computed by the frontend. */ |
aea50b45 | 926 | if (canonical_type_used_p (t)) |
f85d2487 JH |
927 | { |
928 | /* In LTO we set canonical types for all types where it makes | |
929 | sense to do so. Double check we did not miss some type. */ | |
930 | gcc_checking_assert (!in_lto_p || !type_with_alias_set_p (t)); | |
931 | return 0; | |
932 | } | |
933 | } | |
934 | else | |
935 | { | |
936 | t = TYPE_CANONICAL (t); | |
937 | gcc_checking_assert (!TYPE_STRUCTURAL_EQUALITY_P (t)); | |
cb9c2485 | 938 | } |
daad0278 RG |
939 | |
940 | /* If this is a type with a known alias set, return it. */ | |
ba6a6a1d | 941 | gcc_checking_assert (t == TYPE_MAIN_VARIANT (t)); |
738cc472 | 942 | if (TYPE_ALIAS_SET_KNOWN_P (t)) |
3bdf5ad1 RK |
943 | return TYPE_ALIAS_SET (t); |
944 | ||
36784d0e RG |
945 | /* We don't want to set TYPE_ALIAS_SET for incomplete types. */ |
946 | if (!COMPLETE_TYPE_P (t)) | |
947 | { | |
948 | /* For arrays with unknown size the conservative answer is the | |
949 | alias set of the element type. */ | |
950 | if (TREE_CODE (t) == ARRAY_TYPE) | |
951 | return get_alias_set (TREE_TYPE (t)); | |
952 | ||
953 | /* But return zero as a conservative answer for incomplete types. */ | |
954 | return 0; | |
955 | } | |
956 | ||
3bdf5ad1 | 957 | /* See if the language has special handling for this type. */ |
ae2bcd98 | 958 | set = lang_hooks.get_alias_set (t); |
8ac61af7 | 959 | if (set != -1) |
738cc472 | 960 | return set; |
2bf105ab | 961 | |
3bdf5ad1 RK |
962 | /* There are no objects of FUNCTION_TYPE, so there's no point in |
963 | using up an alias set for them. (There are, of course, pointers | |
964 | and references to functions, but that's different.) */ | |
7be7d292 | 965 | else if (TREE_CODE (t) == FUNCTION_TYPE || TREE_CODE (t) == METHOD_TYPE) |
3bdf5ad1 | 966 | set = 0; |
74d86f4f RH |
967 | |
968 | /* Unless the language specifies otherwise, let vector types alias | |
969 | their components. This avoids some nasty type punning issues in | |
970 | normal usage. And indeed lets vectors be treated more like an | |
971 | array slice. */ | |
972 | else if (TREE_CODE (t) == VECTOR_TYPE) | |
973 | set = get_alias_set (TREE_TYPE (t)); | |
974 | ||
4653cae5 RG |
975 | /* Unless the language specifies otherwise, treat array types the |
976 | same as their components. This avoids the asymmetry we get | |
977 | through recording the components. Consider accessing a | |
978 | character(kind=1) through a reference to a character(kind=1)[1:1]. | |
979 | Or consider if we want to assign integer(kind=4)[0:D.1387] and | |
980 | integer(kind=4)[4] the same alias set or not. | |
981 | Just be pragmatic here and make sure the array and its element | |
982 | type get the same alias set assigned. */ | |
aea50b45 JH |
983 | else if (TREE_CODE (t) == ARRAY_TYPE |
984 | && (!TYPE_NONALIASED_COMPONENT (t) | |
985 | || TYPE_STRUCTURAL_EQUALITY_P (t))) | |
4653cae5 RG |
986 | set = get_alias_set (TREE_TYPE (t)); |
987 | ||
0ceb0201 RG |
988 | /* From the former common C and C++ langhook implementation: |
989 | ||
990 | Unfortunately, there is no canonical form of a pointer type. | |
991 | In particular, if we have `typedef int I', then `int *', and | |
992 | `I *' are different types. So, we have to pick a canonical | |
993 | representative. We do this below. | |
994 | ||
995 | Technically, this approach is actually more conservative that | |
996 | it needs to be. In particular, `const int *' and `int *' | |
997 | should be in different alias sets, according to the C and C++ | |
998 | standard, since their types are not the same, and so, | |
999 | technically, an `int **' and `const int **' cannot point at | |
1000 | the same thing. | |
1001 | ||
1002 | But, the standard is wrong. In particular, this code is | |
1003 | legal C++: | |
1004 | ||
1005 | int *ip; | |
1006 | int **ipp = &ip; | |
1007 | const int* const* cipp = ipp; | |
1008 | And, it doesn't make sense for that to be legal unless you | |
1009 | can dereference IPP and CIPP. So, we ignore cv-qualifiers on | |
1010 | the pointed-to types. This issue has been reported to the | |
1011 | C++ committee. | |
1012 | ||
6e042ef4 JH |
1013 | For this reason go to canonical type of the unqalified pointer type. |
1014 | Until GCC 6 this code set all pointers sets to have alias set of | |
1015 | ptr_type_node but that is a bad idea, because it prevents disabiguations | |
1016 | in between pointers. For Firefox this accounts about 20% of all | |
1017 | disambiguations in the program. */ | |
f85d2487 | 1018 | else if (POINTER_TYPE_P (t) && t != ptr_type_node) |
6e042ef4 JH |
1019 | { |
1020 | tree p; | |
1021 | auto_vec <bool, 8> reference; | |
1022 | ||
1023 | /* Unnest all pointers and references. | |
f85d2487 JH |
1024 | We also want to make pointer to array/vector equivalent to pointer to |
1025 | its element (see the reasoning above). Skip all those types, too. */ | |
6e042ef4 | 1026 | for (p = t; POINTER_TYPE_P (p) |
aea50b45 JH |
1027 | || (TREE_CODE (p) == ARRAY_TYPE |
1028 | && (!TYPE_NONALIASED_COMPONENT (p) | |
1029 | || !COMPLETE_TYPE_P (p) | |
1030 | || TYPE_STRUCTURAL_EQUALITY_P (p))) | |
f85d2487 | 1031 | || TREE_CODE (p) == VECTOR_TYPE; |
6e042ef4 JH |
1032 | p = TREE_TYPE (p)) |
1033 | { | |
63f56527 SL |
1034 | /* Ada supports recursive pointers. Instead of doing recursion |
1035 | check, just give up once the preallocated space of 8 elements | |
1036 | is up. In this case just punt to void * alias set. */ | |
54363f8a JH |
1037 | if (reference.length () == 8) |
1038 | { | |
1039 | p = ptr_type_node; | |
1040 | break; | |
1041 | } | |
6e042ef4 | 1042 | if (TREE_CODE (p) == REFERENCE_TYPE) |
f85d2487 JH |
1043 | /* In LTO we want languages that use references to be compatible |
1044 | with languages that use pointers. */ | |
1045 | reference.safe_push (true && !in_lto_p); | |
6e042ef4 JH |
1046 | if (TREE_CODE (p) == POINTER_TYPE) |
1047 | reference.safe_push (false); | |
1048 | } | |
1049 | p = TYPE_MAIN_VARIANT (p); | |
1050 | ||
63f56527 | 1051 | /* In LTO for C++ programs we can turn incomplete types to complete |
e4b44fd7 JH |
1052 | using ODR name lookup. */ |
1053 | if (in_lto_p && TYPE_STRUCTURAL_EQUALITY_P (p) && odr_type_p (p)) | |
1054 | { | |
1055 | p = prevailing_odr_type (p); | |
1056 | gcc_checking_assert (TYPE_MAIN_VARIANT (p) == p); | |
1057 | } | |
1058 | ||
6e042ef4 JH |
1059 | /* Make void * compatible with char * and also void **. |
1060 | Programs are commonly violating TBAA by this. | |
1061 | ||
1062 | We also make void * to conflict with every pointer | |
1063 | (see record_component_aliases) and thus it is safe it to use it for | |
1064 | pointers to types with TYPE_STRUCTURAL_EQUALITY_P. */ | |
1065 | if (TREE_CODE (p) == VOID_TYPE || TYPE_STRUCTURAL_EQUALITY_P (p)) | |
1066 | set = get_alias_set (ptr_type_node); | |
1067 | else | |
1068 | { | |
f85d2487 | 1069 | /* Rebuild pointer type starting from canonical types using |
6e042ef4 JH |
1070 | unqualified pointers and references only. This way all such |
1071 | pointers will have the same alias set and will conflict with | |
1072 | each other. | |
1073 | ||
1074 | Most of time we already have pointers or references of a given type. | |
1075 | If not we build new one just to be sure that if someone later | |
1076 | (probably only middle-end can, as we should assign all alias | |
1077 | classes only after finishing translation unit) builds the pointer | |
1078 | type, the canonical type will match. */ | |
1079 | p = TYPE_CANONICAL (p); | |
1080 | while (!reference.is_empty ()) | |
1081 | { | |
1082 | if (reference.pop ()) | |
1083 | p = build_reference_type (p); | |
1084 | else | |
1085 | p = build_pointer_type (p); | |
f85d2487 JH |
1086 | gcc_checking_assert (p == TYPE_MAIN_VARIANT (p)); |
1087 | /* build_pointer_type should always return the canonical type. | |
1088 | For LTO TYPE_CANOINCAL may be NULL, because we do not compute | |
1089 | them. Be sure that frontends do not glob canonical types of | |
1090 | pointers in unexpected way and that p == TYPE_CANONICAL (p) | |
1091 | in all other cases. */ | |
1092 | gcc_checking_assert (!TYPE_CANONICAL (p) | |
1093 | || p == TYPE_CANONICAL (p)); | |
6e042ef4 | 1094 | } |
6e042ef4 JH |
1095 | |
1096 | /* Assign the alias set to both p and t. | |
67914693 | 1097 | We cannot call get_alias_set (p) here as that would trigger |
6e042ef4 JH |
1098 | infinite recursion when p == t. In other cases it would just |
1099 | trigger unnecesary legwork of rebuilding the pointer again. */ | |
ba6a6a1d | 1100 | gcc_checking_assert (p == TYPE_MAIN_VARIANT (p)); |
6e042ef4 JH |
1101 | if (TYPE_ALIAS_SET_KNOWN_P (p)) |
1102 | set = TYPE_ALIAS_SET (p); | |
1103 | else | |
1104 | { | |
1105 | set = new_alias_set (); | |
1106 | TYPE_ALIAS_SET (p) = set; | |
1107 | } | |
1108 | } | |
1109 | } | |
f85d2487 JH |
1110 | /* Alias set of ptr_type_node is special and serve as universal pointer which |
1111 | is TBAA compatible with every other pointer type. Be sure we have the | |
1112 | alias set built even for LTO which otherwise keeps all TYPE_CANONICAL | |
1113 | of pointer types NULL. */ | |
1114 | else if (t == ptr_type_node) | |
1115 | set = new_alias_set (); | |
0ceb0201 | 1116 | |
7be7d292 | 1117 | /* Otherwise make a new alias set for this type. */ |
3bdf5ad1 | 1118 | else |
96d91dcf RG |
1119 | { |
1120 | /* Each canonical type gets its own alias set, so canonical types | |
1121 | shouldn't form a tree. It doesn't really matter for types | |
1122 | we handle specially above, so only check it where it possibly | |
1123 | would result in a bogus alias set. */ | |
1124 | gcc_checking_assert (TYPE_CANONICAL (t) == t); | |
1125 | ||
1126 | set = new_alias_set (); | |
1127 | } | |
3bdf5ad1 RK |
1128 | |
1129 | TYPE_ALIAS_SET (t) = set; | |
2bf105ab | 1130 | |
7be7d292 EB |
1131 | /* If this is an aggregate type or a complex type, we must record any |
1132 | component aliasing information. */ | |
1d79fd2c | 1133 | if (AGGREGATE_TYPE_P (t) || TREE_CODE (t) == COMPLEX_TYPE) |
2bf105ab RK |
1134 | record_component_aliases (t); |
1135 | ||
6e042ef4 JH |
1136 | /* We treat pointer types specially in alias_set_subset_of. */ |
1137 | if (POINTER_TYPE_P (t) && set) | |
1138 | { | |
02ced957 | 1139 | alias_set_entry *ase = get_alias_set_entry (set); |
6e042ef4 JH |
1140 | if (!ase) |
1141 | ase = init_alias_set_entry (set); | |
1142 | ase->is_pointer = true; | |
1143 | ase->has_pointer = true; | |
1144 | } | |
1145 | ||
3bdf5ad1 RK |
1146 | return set; |
1147 | } | |
1148 | ||
1149 | /* Return a brand-new alias set. */ | |
1150 | ||
4862826d | 1151 | alias_set_type |
4682ae04 | 1152 | new_alias_set (void) |
3bdf5ad1 | 1153 | { |
bd04cddf JH |
1154 | if (alias_sets == 0) |
1155 | vec_safe_push (alias_sets, (alias_set_entry *) NULL); | |
1156 | vec_safe_push (alias_sets, (alias_set_entry *) NULL); | |
1157 | return alias_sets->length () - 1; | |
3bdf5ad1 | 1158 | } |
3932261a | 1159 | |
01d28c3f JM |
1160 | /* Indicate that things in SUBSET can alias things in SUPERSET, but that |
1161 | not everything that aliases SUPERSET also aliases SUBSET. For example, | |
1162 | in C, a store to an `int' can alias a load of a structure containing an | |
1163 | `int', and vice versa. But it can't alias a load of a 'double' member | |
1164 | of the same structure. Here, the structure would be the SUPERSET and | |
1165 | `int' the SUBSET. This relationship is also described in the comment at | |
1166 | the beginning of this file. | |
1167 | ||
1168 | This function should be called only once per SUPERSET/SUBSET pair. | |
3932261a MM |
1169 | |
1170 | It is illegal for SUPERSET to be zero; everything is implicitly a | |
1171 | subset of alias set zero. */ | |
1172 | ||
794511d2 | 1173 | void |
4862826d | 1174 | record_alias_subset (alias_set_type superset, alias_set_type subset) |
3932261a | 1175 | { |
02ced957 TS |
1176 | alias_set_entry *superset_entry; |
1177 | alias_set_entry *subset_entry; | |
3932261a | 1178 | |
f47e9b4e RK |
1179 | /* It is possible in complex type situations for both sets to be the same, |
1180 | in which case we can ignore this operation. */ | |
1181 | if (superset == subset) | |
1182 | return; | |
1183 | ||
298e6adc | 1184 | gcc_assert (superset); |
3932261a MM |
1185 | |
1186 | superset_entry = get_alias_set_entry (superset); | |
ca7fd9cd | 1187 | if (superset_entry == 0) |
3932261a MM |
1188 | { |
1189 | /* Create an entry for the SUPERSET, so that we have a place to | |
1190 | attach the SUBSET. */ | |
6e042ef4 | 1191 | superset_entry = init_alias_set_entry (superset); |
3932261a MM |
1192 | } |
1193 | ||
2bf105ab RK |
1194 | if (subset == 0) |
1195 | superset_entry->has_zero_child = 1; | |
1196 | else | |
1197 | { | |
6e042ef4 JH |
1198 | if (!superset_entry->children) |
1199 | superset_entry->children | |
fb5c464a | 1200 | = hash_map<alias_set_hash, int>::create_ggc (64); |
6c577667 RB |
1201 | |
1202 | /* Enter the SUBSET itself as a child of the SUPERSET. If it was | |
1203 | already there we're done. */ | |
1204 | if (superset_entry->children->put (subset, 0)) | |
1205 | return; | |
1206 | ||
1207 | subset_entry = get_alias_set_entry (subset); | |
2bf105ab RK |
1208 | /* If there is an entry for the subset, enter all of its children |
1209 | (if they are not already present) as children of the SUPERSET. */ | |
ca7fd9cd | 1210 | if (subset_entry) |
2bf105ab RK |
1211 | { |
1212 | if (subset_entry->has_zero_child) | |
6e042ef4 JH |
1213 | superset_entry->has_zero_child = true; |
1214 | if (subset_entry->has_pointer) | |
1215 | superset_entry->has_pointer = true; | |
d4b60170 | 1216 | |
6e042ef4 JH |
1217 | if (subset_entry->children) |
1218 | { | |
fb5c464a | 1219 | hash_map<alias_set_hash, int>::iterator iter |
6e042ef4 JH |
1220 | = subset_entry->children->begin (); |
1221 | for (; iter != subset_entry->children->end (); ++iter) | |
1222 | superset_entry->children->put ((*iter).first, (*iter).second); | |
1223 | } | |
2bf105ab | 1224 | } |
2bf105ab | 1225 | } |
3932261a MM |
1226 | } |
1227 | ||
6b8df3e4 | 1228 | /* Record that component types of TYPE, if any, are part of SUPERSET for |
a0c33338 | 1229 | aliasing purposes. For record types, we only record component types |
b5487346 EB |
1230 | for fields that are not marked non-addressable. For array types, we |
1231 | only record the component type if it is not marked non-aliased. */ | |
a0c33338 RK |
1232 | |
1233 | void | |
6b8df3e4 | 1234 | record_component_aliases (tree type, alias_set_type superset) |
a0c33338 | 1235 | { |
a0c33338 RK |
1236 | tree field; |
1237 | ||
1238 | if (superset == 0) | |
1239 | return; | |
1240 | ||
1241 | switch (TREE_CODE (type)) | |
1242 | { | |
a0c33338 RK |
1243 | case RECORD_TYPE: |
1244 | case UNION_TYPE: | |
1245 | case QUAL_UNION_TYPE: | |
17b99c98 JH |
1246 | { |
1247 | /* LTO non-ODR type merging does not make any difference between | |
1248 | component pointer types. We may have | |
1249 | ||
1250 | struct foo {int *a;}; | |
1251 | ||
1252 | as TYPE_CANONICAL of | |
1253 | ||
1254 | struct bar {float *a;}; | |
1255 | ||
1256 | Because accesses to int * and float * do not alias, we would get | |
1257 | false negative when accessing the same memory location by | |
1258 | float ** and bar *. We thus record the canonical type as: | |
1259 | ||
1260 | struct {void *a;}; | |
1261 | ||
1262 | void * is special cased and works as a universal pointer type. | |
1263 | Accesses to it conflicts with accesses to any other pointer | |
1264 | type. */ | |
1265 | bool void_pointers = in_lto_p | |
1266 | && (!odr_type_p (type) | |
1267 | || !odr_based_tbaa_p (type)); | |
1268 | for (field = TYPE_FIELDS (type); field != 0; field = DECL_CHAIN (field)) | |
1269 | if (TREE_CODE (field) == FIELD_DECL && !DECL_NONADDRESSABLE_P (field)) | |
1270 | { | |
1271 | tree t = TREE_TYPE (field); | |
1272 | if (void_pointers) | |
1273 | { | |
1274 | /* VECTOR_TYPE and ARRAY_TYPE share the alias set with their | |
1275 | element type and that type has to be normalized to void *, | |
1276 | too, in the case it is a pointer. */ | |
1277 | while (!canonical_type_used_p (t) && !POINTER_TYPE_P (t)) | |
1278 | { | |
1279 | gcc_checking_assert (TYPE_STRUCTURAL_EQUALITY_P (t)); | |
1280 | t = TREE_TYPE (t); | |
1281 | } | |
1282 | if (POINTER_TYPE_P (t)) | |
1283 | t = ptr_type_node; | |
1284 | else if (flag_checking) | |
1285 | gcc_checking_assert (get_alias_set (t) | |
1286 | == get_alias_set (TREE_TYPE (field))); | |
1287 | } | |
1288 | ||
6b8df3e4 RB |
1289 | alias_set_type set = get_alias_set (t); |
1290 | record_alias_subset (superset, set); | |
1291 | /* If the field has alias-set zero make sure to still record | |
1292 | any componets of it. This makes sure that for | |
1293 | struct A { | |
1294 | struct B { | |
1295 | int i; | |
1296 | char c[4]; | |
1297 | } b; | |
1298 | }; | |
1299 | in C++ even though 'B' has alias-set zero because | |
1300 | TYPE_TYPELESS_STORAGE is set, 'A' has the alias-set of | |
1301 | 'int' as subset. */ | |
1302 | if (set == 0) | |
1303 | record_component_aliases (t, superset); | |
17b99c98 JH |
1304 | } |
1305 | } | |
a0c33338 RK |
1306 | break; |
1307 | ||
1d79fd2c JW |
1308 | case COMPLEX_TYPE: |
1309 | record_alias_subset (superset, get_alias_set (TREE_TYPE (type))); | |
1310 | break; | |
1311 | ||
4653cae5 RG |
1312 | /* VECTOR_TYPE and ARRAY_TYPE share the alias set with their |
1313 | element type. */ | |
1314 | ||
a0c33338 RK |
1315 | default: |
1316 | break; | |
1317 | } | |
1318 | } | |
1319 | ||
6b8df3e4 RB |
1320 | /* Record that component types of TYPE, if any, are part of that type for |
1321 | aliasing purposes. For record types, we only record component types | |
1322 | for fields that are not marked non-addressable. For array types, we | |
1323 | only record the component type if it is not marked non-aliased. */ | |
1324 | ||
1325 | void | |
1326 | record_component_aliases (tree type) | |
1327 | { | |
1328 | alias_set_type superset = get_alias_set (type); | |
1329 | record_component_aliases (type, superset); | |
1330 | } | |
1331 | ||
1332 | ||
3bdf5ad1 RK |
1333 | /* Allocate an alias set for use in storing and reading from the varargs |
1334 | spill area. */ | |
1335 | ||
4862826d | 1336 | static GTY(()) alias_set_type varargs_set = -1; |
f103e34d | 1337 | |
4862826d | 1338 | alias_set_type |
4682ae04 | 1339 | get_varargs_alias_set (void) |
3bdf5ad1 | 1340 | { |
cd3ce9b4 JM |
1341 | #if 1 |
1342 | /* We now lower VA_ARG_EXPR, and there's currently no way to attach the | |
1343 | varargs alias set to an INDIRECT_REF (FIXME!), so we can't | |
1344 | consistently use the varargs alias set for loads from the varargs | |
1345 | area. So don't use it anywhere. */ | |
1346 | return 0; | |
1347 | #else | |
f103e34d GK |
1348 | if (varargs_set == -1) |
1349 | varargs_set = new_alias_set (); | |
3bdf5ad1 | 1350 | |
f103e34d | 1351 | return varargs_set; |
cd3ce9b4 | 1352 | #endif |
3bdf5ad1 RK |
1353 | } |
1354 | ||
1355 | /* Likewise, but used for the fixed portions of the frame, e.g., register | |
1356 | save areas. */ | |
1357 | ||
4862826d | 1358 | static GTY(()) alias_set_type frame_set = -1; |
f103e34d | 1359 | |
4862826d | 1360 | alias_set_type |
4682ae04 | 1361 | get_frame_alias_set (void) |
3bdf5ad1 | 1362 | { |
f103e34d GK |
1363 | if (frame_set == -1) |
1364 | frame_set = new_alias_set (); | |
3bdf5ad1 | 1365 | |
f103e34d | 1366 | return frame_set; |
3bdf5ad1 RK |
1367 | } |
1368 | ||
9fc37b2b RS |
1369 | /* Create a new, unique base with id ID. */ |
1370 | ||
1371 | static rtx | |
1372 | unique_base_value (HOST_WIDE_INT id) | |
1373 | { | |
1374 | return gen_rtx_ADDRESS (Pmode, id); | |
1375 | } | |
1376 | ||
1377 | /* Return true if accesses based on any other base value cannot alias | |
1378 | those based on X. */ | |
1379 | ||
1380 | static bool | |
1381 | unique_base_value_p (rtx x) | |
1382 | { | |
1383 | return GET_CODE (x) == ADDRESS && GET_MODE (x) == Pmode; | |
1384 | } | |
1385 | ||
1386 | /* Return true if X is known to be a base value. */ | |
1387 | ||
1388 | static bool | |
1389 | known_base_value_p (rtx x) | |
1390 | { | |
1391 | switch (GET_CODE (x)) | |
1392 | { | |
1393 | case LABEL_REF: | |
1394 | case SYMBOL_REF: | |
1395 | return true; | |
1396 | ||
1397 | case ADDRESS: | |
1398 | /* Arguments may or may not be bases; we don't know for sure. */ | |
1399 | return GET_MODE (x) != VOIDmode; | |
1400 | ||
1401 | default: | |
1402 | return false; | |
1403 | } | |
1404 | } | |
1405 | ||
2a2c8203 JC |
1406 | /* Inside SRC, the source of a SET, find a base address. */ |
1407 | ||
9ae8ffe7 | 1408 | static rtx |
4682ae04 | 1409 | find_base_value (rtx src) |
9ae8ffe7 | 1410 | { |
713f41f9 | 1411 | unsigned int regno; |
6645d841 | 1412 | scalar_int_mode int_mode; |
0aacc8ed | 1413 | |
53451050 RS |
1414 | #if defined (FIND_BASE_TERM) |
1415 | /* Try machine-dependent ways to find the base term. */ | |
1416 | src = FIND_BASE_TERM (src); | |
1417 | #endif | |
1418 | ||
9ae8ffe7 JL |
1419 | switch (GET_CODE (src)) |
1420 | { | |
1421 | case SYMBOL_REF: | |
1422 | case LABEL_REF: | |
1423 | return src; | |
1424 | ||
1425 | case REG: | |
fb6754f0 | 1426 | regno = REGNO (src); |
d4b60170 | 1427 | /* At the start of a function, argument registers have known base |
2a2c8203 JC |
1428 | values which may be lost later. Returning an ADDRESS |
1429 | expression here allows optimization based on argument values | |
1430 | even when the argument registers are used for other purposes. */ | |
713f41f9 BS |
1431 | if (regno < FIRST_PSEUDO_REGISTER && copying_arguments) |
1432 | return new_reg_base_value[regno]; | |
73774bc7 | 1433 | |
eaf407a5 | 1434 | /* If a pseudo has a known base value, return it. Do not do this |
9b462c42 RH |
1435 | for non-fixed hard regs since it can result in a circular |
1436 | dependency chain for registers which have values at function entry. | |
eaf407a5 JL |
1437 | |
1438 | The test above is not sufficient because the scheduler may move | |
1439 | a copy out of an arg reg past the NOTE_INSN_FUNCTION_BEGIN. */ | |
9b462c42 | 1440 | if ((regno >= FIRST_PSEUDO_REGISTER || fixed_regs[regno]) |
9771b263 | 1441 | && regno < vec_safe_length (reg_base_value)) |
83bbd9b6 RH |
1442 | { |
1443 | /* If we're inside init_alias_analysis, use new_reg_base_value | |
1444 | to reduce the number of relaxation iterations. */ | |
1afdf91c | 1445 | if (new_reg_base_value && new_reg_base_value[regno] |
6fb5fa3c | 1446 | && DF_REG_DEF_COUNT (regno) == 1) |
83bbd9b6 RH |
1447 | return new_reg_base_value[regno]; |
1448 | ||
9771b263 DN |
1449 | if ((*reg_base_value)[regno]) |
1450 | return (*reg_base_value)[regno]; | |
83bbd9b6 | 1451 | } |
73774bc7 | 1452 | |
e3f049a8 | 1453 | return 0; |
9ae8ffe7 JL |
1454 | |
1455 | case MEM: | |
1456 | /* Check for an argument passed in memory. Only record in the | |
1457 | copying-arguments block; it is too hard to track changes | |
1458 | otherwise. */ | |
1459 | if (copying_arguments | |
1460 | && (XEXP (src, 0) == arg_pointer_rtx | |
1461 | || (GET_CODE (XEXP (src, 0)) == PLUS | |
1462 | && XEXP (XEXP (src, 0), 0) == arg_pointer_rtx))) | |
9fc37b2b | 1463 | return arg_base_value; |
9ae8ffe7 JL |
1464 | return 0; |
1465 | ||
1466 | case CONST: | |
1467 | src = XEXP (src, 0); | |
1468 | if (GET_CODE (src) != PLUS && GET_CODE (src) != MINUS) | |
1469 | break; | |
d4b60170 | 1470 | |
191816a3 | 1471 | /* fall through */ |
2a2c8203 | 1472 | |
9ae8ffe7 JL |
1473 | case PLUS: |
1474 | case MINUS: | |
2a2c8203 | 1475 | { |
ec907dd8 JL |
1476 | rtx temp, src_0 = XEXP (src, 0), src_1 = XEXP (src, 1); |
1477 | ||
0134bf2d DE |
1478 | /* If either operand is a REG that is a known pointer, then it |
1479 | is the base. */ | |
1480 | if (REG_P (src_0) && REG_POINTER (src_0)) | |
1481 | return find_base_value (src_0); | |
1482 | if (REG_P (src_1) && REG_POINTER (src_1)) | |
1483 | return find_base_value (src_1); | |
1484 | ||
ec907dd8 JL |
1485 | /* If either operand is a REG, then see if we already have |
1486 | a known value for it. */ | |
0134bf2d | 1487 | if (REG_P (src_0)) |
ec907dd8 JL |
1488 | { |
1489 | temp = find_base_value (src_0); | |
d4b60170 | 1490 | if (temp != 0) |
ec907dd8 JL |
1491 | src_0 = temp; |
1492 | } | |
1493 | ||
0134bf2d | 1494 | if (REG_P (src_1)) |
ec907dd8 JL |
1495 | { |
1496 | temp = find_base_value (src_1); | |
d4b60170 | 1497 | if (temp!= 0) |
ec907dd8 JL |
1498 | src_1 = temp; |
1499 | } | |
2a2c8203 | 1500 | |
0134bf2d DE |
1501 | /* If either base is named object or a special address |
1502 | (like an argument or stack reference), then use it for the | |
1503 | base term. */ | |
9fc37b2b | 1504 | if (src_0 != 0 && known_base_value_p (src_0)) |
0134bf2d DE |
1505 | return src_0; |
1506 | ||
9fc37b2b | 1507 | if (src_1 != 0 && known_base_value_p (src_1)) |
0134bf2d DE |
1508 | return src_1; |
1509 | ||
d4b60170 | 1510 | /* Guess which operand is the base address: |
ec907dd8 JL |
1511 | If either operand is a symbol, then it is the base. If |
1512 | either operand is a CONST_INT, then the other is the base. */ | |
481683e1 | 1513 | if (CONST_INT_P (src_1) || CONSTANT_P (src_0)) |
2a2c8203 | 1514 | return find_base_value (src_0); |
481683e1 | 1515 | else if (CONST_INT_P (src_0) || CONSTANT_P (src_1)) |
ec907dd8 JL |
1516 | return find_base_value (src_1); |
1517 | ||
9ae8ffe7 | 1518 | return 0; |
2a2c8203 JC |
1519 | } |
1520 | ||
1521 | case LO_SUM: | |
1522 | /* The standard form is (lo_sum reg sym) so look only at the | |
1523 | second operand. */ | |
1524 | return find_base_value (XEXP (src, 1)); | |
9ae8ffe7 JL |
1525 | |
1526 | case AND: | |
ef50b972 RB |
1527 | /* Look through aligning ANDs. And AND with zero or one with |
1528 | the LSB set isn't one (see for example PR92462). */ | |
1529 | if (CONST_INT_P (XEXP (src, 1)) | |
1530 | && INTVAL (XEXP (src, 1)) != 0 | |
1531 | && (INTVAL (XEXP (src, 1)) & 1) == 0) | |
2a2c8203 | 1532 | return find_base_value (XEXP (src, 0)); |
9ae8ffe7 JL |
1533 | return 0; |
1534 | ||
61f0131c | 1535 | case TRUNCATE: |
5932a4d4 | 1536 | /* As we do not know which address space the pointer is referring to, we can |
d4ebfa65 BE |
1537 | handle this only if the target does not support different pointer or |
1538 | address modes depending on the address space. */ | |
1539 | if (!target_default_pointer_address_modes_p ()) | |
1540 | break; | |
6645d841 RS |
1541 | if (!is_a <scalar_int_mode> (GET_MODE (src), &int_mode) |
1542 | || GET_MODE_PRECISION (int_mode) < GET_MODE_PRECISION (Pmode)) | |
61f0131c R |
1543 | break; |
1544 | /* Fall through. */ | |
9ae8ffe7 | 1545 | case HIGH: |
d288e53d DE |
1546 | case PRE_INC: |
1547 | case PRE_DEC: | |
1548 | case POST_INC: | |
1549 | case POST_DEC: | |
1550 | case PRE_MODIFY: | |
1551 | case POST_MODIFY: | |
2a2c8203 | 1552 | return find_base_value (XEXP (src, 0)); |
1d300e19 | 1553 | |
0aacc8ed RK |
1554 | case ZERO_EXTEND: |
1555 | case SIGN_EXTEND: /* used for NT/Alpha pointers */ | |
5932a4d4 | 1556 | /* As we do not know which address space the pointer is referring to, we can |
d4ebfa65 BE |
1557 | handle this only if the target does not support different pointer or |
1558 | address modes depending on the address space. */ | |
1559 | if (!target_default_pointer_address_modes_p ()) | |
1560 | break; | |
1561 | ||
0aacc8ed RK |
1562 | { |
1563 | rtx temp = find_base_value (XEXP (src, 0)); | |
1564 | ||
5ae6cd0d | 1565 | if (temp != 0 && CONSTANT_P (temp)) |
0aacc8ed | 1566 | temp = convert_memory_address (Pmode, temp); |
0aacc8ed RK |
1567 | |
1568 | return temp; | |
1569 | } | |
1570 | ||
1d300e19 KG |
1571 | default: |
1572 | break; | |
9ae8ffe7 JL |
1573 | } |
1574 | ||
1575 | return 0; | |
1576 | } | |
1577 | ||
9fc37b2b RS |
1578 | /* Called from init_alias_analysis indirectly through note_stores, |
1579 | or directly if DEST is a register with a REG_NOALIAS note attached. | |
1580 | SET is null in the latter case. */ | |
9ae8ffe7 | 1581 | |
d4b60170 | 1582 | /* While scanning insns to find base values, reg_seen[N] is nonzero if |
9ae8ffe7 | 1583 | register N has been set in this function. */ |
d630245f | 1584 | static sbitmap reg_seen; |
9ae8ffe7 | 1585 | |
2a2c8203 | 1586 | static void |
7bc980e1 | 1587 | record_set (rtx dest, const_rtx set, void *data ATTRIBUTE_UNUSED) |
9ae8ffe7 | 1588 | { |
b3694847 | 1589 | unsigned regno; |
9ae8ffe7 | 1590 | rtx src; |
c28b4e40 | 1591 | int n; |
9ae8ffe7 | 1592 | |
f8cfc6aa | 1593 | if (!REG_P (dest)) |
9ae8ffe7 JL |
1594 | return; |
1595 | ||
fb6754f0 | 1596 | regno = REGNO (dest); |
9ae8ffe7 | 1597 | |
9771b263 | 1598 | gcc_checking_assert (regno < reg_base_value->length ()); |
ac606739 | 1599 | |
dc8afb70 | 1600 | n = REG_NREGS (dest); |
c28b4e40 JW |
1601 | if (n != 1) |
1602 | { | |
1603 | while (--n >= 0) | |
1604 | { | |
d7c028c0 | 1605 | bitmap_set_bit (reg_seen, regno + n); |
c28b4e40 JW |
1606 | new_reg_base_value[regno + n] = 0; |
1607 | } | |
1608 | return; | |
1609 | } | |
1610 | ||
9ae8ffe7 JL |
1611 | if (set) |
1612 | { | |
1613 | /* A CLOBBER wipes out any old value but does not prevent a previously | |
1614 | unset register from acquiring a base address (i.e. reg_seen is not | |
1615 | set). */ | |
1616 | if (GET_CODE (set) == CLOBBER) | |
1617 | { | |
ec907dd8 | 1618 | new_reg_base_value[regno] = 0; |
9ae8ffe7 JL |
1619 | return; |
1620 | } | |
8df47bdf | 1621 | |
9ae8ffe7 JL |
1622 | src = SET_SRC (set); |
1623 | } | |
1624 | else | |
1625 | { | |
9fc37b2b | 1626 | /* There's a REG_NOALIAS note against DEST. */ |
d7c028c0 | 1627 | if (bitmap_bit_p (reg_seen, regno)) |
9ae8ffe7 | 1628 | { |
ec907dd8 | 1629 | new_reg_base_value[regno] = 0; |
9ae8ffe7 JL |
1630 | return; |
1631 | } | |
d7c028c0 | 1632 | bitmap_set_bit (reg_seen, regno); |
9fc37b2b | 1633 | new_reg_base_value[regno] = unique_base_value (unique_id++); |
9ae8ffe7 JL |
1634 | return; |
1635 | } | |
1636 | ||
5da6f168 RS |
1637 | /* If this is not the first set of REGNO, see whether the new value |
1638 | is related to the old one. There are two cases of interest: | |
1639 | ||
1640 | (1) The register might be assigned an entirely new value | |
1641 | that has the same base term as the original set. | |
1642 | ||
1643 | (2) The set might be a simple self-modification that | |
1644 | cannot change REGNO's base value. | |
1645 | ||
1646 | If neither case holds, reject the original base value as invalid. | |
1647 | Note that the following situation is not detected: | |
1648 | ||
c22cacf3 | 1649 | extern int x, y; int *p = &x; p += (&y-&x); |
5da6f168 | 1650 | |
9ae8ffe7 JL |
1651 | ANSI C does not allow computing the difference of addresses |
1652 | of distinct top level objects. */ | |
5da6f168 RS |
1653 | if (new_reg_base_value[regno] != 0 |
1654 | && find_base_value (src) != new_reg_base_value[regno]) | |
9ae8ffe7 JL |
1655 | switch (GET_CODE (src)) |
1656 | { | |
2a2c8203 | 1657 | case LO_SUM: |
9ae8ffe7 JL |
1658 | case MINUS: |
1659 | if (XEXP (src, 0) != dest && XEXP (src, 1) != dest) | |
ec907dd8 | 1660 | new_reg_base_value[regno] = 0; |
9ae8ffe7 | 1661 | break; |
61f0131c R |
1662 | case PLUS: |
1663 | /* If the value we add in the PLUS is also a valid base value, | |
1664 | this might be the actual base value, and the original value | |
1665 | an index. */ | |
1666 | { | |
1667 | rtx other = NULL_RTX; | |
1668 | ||
1669 | if (XEXP (src, 0) == dest) | |
1670 | other = XEXP (src, 1); | |
1671 | else if (XEXP (src, 1) == dest) | |
1672 | other = XEXP (src, 0); | |
1673 | ||
1674 | if (! other || find_base_value (other)) | |
1675 | new_reg_base_value[regno] = 0; | |
1676 | break; | |
1677 | } | |
9ae8ffe7 | 1678 | case AND: |
481683e1 | 1679 | if (XEXP (src, 0) != dest || !CONST_INT_P (XEXP (src, 1))) |
ec907dd8 | 1680 | new_reg_base_value[regno] = 0; |
9ae8ffe7 | 1681 | break; |
9ae8ffe7 | 1682 | default: |
ec907dd8 | 1683 | new_reg_base_value[regno] = 0; |
9ae8ffe7 JL |
1684 | break; |
1685 | } | |
1686 | /* If this is the first set of a register, record the value. */ | |
1687 | else if ((regno >= FIRST_PSEUDO_REGISTER || ! fixed_regs[regno]) | |
d7c028c0 | 1688 | && ! bitmap_bit_p (reg_seen, regno) && new_reg_base_value[regno] == 0) |
ec907dd8 | 1689 | new_reg_base_value[regno] = find_base_value (src); |
9ae8ffe7 | 1690 | |
d7c028c0 | 1691 | bitmap_set_bit (reg_seen, regno); |
9ae8ffe7 JL |
1692 | } |
1693 | ||
8fd0a474 AM |
1694 | /* Return REG_BASE_VALUE for REGNO. Selective scheduler uses this to avoid |
1695 | using hard registers with non-null REG_BASE_VALUE for renaming. */ | |
1696 | rtx | |
1697 | get_reg_base_value (unsigned int regno) | |
1698 | { | |
9771b263 | 1699 | return (*reg_base_value)[regno]; |
8fd0a474 AM |
1700 | } |
1701 | ||
bb1acb3e RH |
1702 | /* If a value is known for REGNO, return it. */ |
1703 | ||
c22cacf3 | 1704 | rtx |
bb1acb3e RH |
1705 | get_reg_known_value (unsigned int regno) |
1706 | { | |
1707 | if (regno >= FIRST_PSEUDO_REGISTER) | |
1708 | { | |
1709 | regno -= FIRST_PSEUDO_REGISTER; | |
9771b263 DN |
1710 | if (regno < vec_safe_length (reg_known_value)) |
1711 | return (*reg_known_value)[regno]; | |
bb1acb3e RH |
1712 | } |
1713 | return NULL; | |
43fe47ca JW |
1714 | } |
1715 | ||
bb1acb3e RH |
1716 | /* Set it. */ |
1717 | ||
1718 | static void | |
1719 | set_reg_known_value (unsigned int regno, rtx val) | |
1720 | { | |
1721 | if (regno >= FIRST_PSEUDO_REGISTER) | |
1722 | { | |
1723 | regno -= FIRST_PSEUDO_REGISTER; | |
9771b263 DN |
1724 | if (regno < vec_safe_length (reg_known_value)) |
1725 | (*reg_known_value)[regno] = val; | |
bb1acb3e RH |
1726 | } |
1727 | } | |
1728 | ||
1729 | /* Similarly for reg_known_equiv_p. */ | |
1730 | ||
1731 | bool | |
1732 | get_reg_known_equiv_p (unsigned int regno) | |
1733 | { | |
1734 | if (regno >= FIRST_PSEUDO_REGISTER) | |
1735 | { | |
1736 | regno -= FIRST_PSEUDO_REGISTER; | |
9771b263 | 1737 | if (regno < vec_safe_length (reg_known_value)) |
d7c028c0 | 1738 | return bitmap_bit_p (reg_known_equiv_p, regno); |
bb1acb3e RH |
1739 | } |
1740 | return false; | |
1741 | } | |
1742 | ||
1743 | static void | |
1744 | set_reg_known_equiv_p (unsigned int regno, bool val) | |
1745 | { | |
1746 | if (regno >= FIRST_PSEUDO_REGISTER) | |
1747 | { | |
1748 | regno -= FIRST_PSEUDO_REGISTER; | |
9771b263 | 1749 | if (regno < vec_safe_length (reg_known_value)) |
9ff3c7ca SB |
1750 | { |
1751 | if (val) | |
d7c028c0 | 1752 | bitmap_set_bit (reg_known_equiv_p, regno); |
9ff3c7ca | 1753 | else |
d7c028c0 | 1754 | bitmap_clear_bit (reg_known_equiv_p, regno); |
9ff3c7ca | 1755 | } |
bb1acb3e RH |
1756 | } |
1757 | } | |
1758 | ||
1759 | ||
db048faf MM |
1760 | /* Returns a canonical version of X, from the point of view alias |
1761 | analysis. (For example, if X is a MEM whose address is a register, | |
1762 | and the register has a known value (say a SYMBOL_REF), then a MEM | |
1763 | whose address is the SYMBOL_REF is returned.) */ | |
1764 | ||
1765 | rtx | |
4682ae04 | 1766 | canon_rtx (rtx x) |
9ae8ffe7 JL |
1767 | { |
1768 | /* Recursively look for equivalences. */ | |
f8cfc6aa | 1769 | if (REG_P (x) && REGNO (x) >= FIRST_PSEUDO_REGISTER) |
bb1acb3e RH |
1770 | { |
1771 | rtx t = get_reg_known_value (REGNO (x)); | |
1772 | if (t == x) | |
1773 | return x; | |
1774 | if (t) | |
1775 | return canon_rtx (t); | |
1776 | } | |
1777 | ||
1778 | if (GET_CODE (x) == PLUS) | |
9ae8ffe7 JL |
1779 | { |
1780 | rtx x0 = canon_rtx (XEXP (x, 0)); | |
1781 | rtx x1 = canon_rtx (XEXP (x, 1)); | |
1782 | ||
1783 | if (x0 != XEXP (x, 0) || x1 != XEXP (x, 1)) | |
231314e3 | 1784 | return simplify_gen_binary (PLUS, GET_MODE (x), x0, x1); |
9ae8ffe7 | 1785 | } |
d4b60170 | 1786 | |
9ae8ffe7 JL |
1787 | /* This gives us much better alias analysis when called from |
1788 | the loop optimizer. Note we want to leave the original | |
1789 | MEM alone, but need to return the canonicalized MEM with | |
1790 | all the flags with their original values. */ | |
3c0cb5de | 1791 | else if (MEM_P (x)) |
f1ec5147 | 1792 | x = replace_equiv_address_nv (x, canon_rtx (XEXP (x, 0))); |
d4b60170 | 1793 | |
9ae8ffe7 JL |
1794 | return x; |
1795 | } | |
1796 | ||
1797 | /* Return 1 if X and Y are identical-looking rtx's. | |
45183e03 | 1798 | Expect that X and Y has been already canonicalized. |
9ae8ffe7 JL |
1799 | |
1800 | We use the data in reg_known_value above to see if two registers with | |
1801 | different numbers are, in fact, equivalent. */ | |
1802 | ||
1803 | static int | |
ed7a4b4b | 1804 | rtx_equal_for_memref_p (const_rtx x, const_rtx y) |
9ae8ffe7 | 1805 | { |
b3694847 SS |
1806 | int i; |
1807 | int j; | |
1808 | enum rtx_code code; | |
1809 | const char *fmt; | |
9ae8ffe7 JL |
1810 | |
1811 | if (x == 0 && y == 0) | |
1812 | return 1; | |
1813 | if (x == 0 || y == 0) | |
1814 | return 0; | |
d4b60170 | 1815 | |
9ae8ffe7 JL |
1816 | if (x == y) |
1817 | return 1; | |
1818 | ||
1819 | code = GET_CODE (x); | |
1820 | /* Rtx's of different codes cannot be equal. */ | |
1821 | if (code != GET_CODE (y)) | |
1822 | return 0; | |
1823 | ||
1824 | /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. | |
1825 | (REG:SI x) and (REG:HI x) are NOT equivalent. */ | |
1826 | ||
1827 | if (GET_MODE (x) != GET_MODE (y)) | |
1828 | return 0; | |
1829 | ||
db048faf MM |
1830 | /* Some RTL can be compared without a recursive examination. */ |
1831 | switch (code) | |
1832 | { | |
1833 | case REG: | |
1834 | return REGNO (x) == REGNO (y); | |
1835 | ||
1836 | case LABEL_REF: | |
04a121a7 | 1837 | return label_ref_label (x) == label_ref_label (y); |
ca7fd9cd | 1838 | |
db048faf | 1839 | case SYMBOL_REF: |
73e48cb3 | 1840 | return compare_base_symbol_refs (x, y) == 1; |
db048faf | 1841 | |
af6236c1 AO |
1842 | case ENTRY_VALUE: |
1843 | /* This is magic, don't go through canonicalization et al. */ | |
1844 | return rtx_equal_p (ENTRY_VALUE_EXP (x), ENTRY_VALUE_EXP (y)); | |
1845 | ||
40e02b4a | 1846 | case VALUE: |
d8116890 | 1847 | CASE_CONST_UNIQUE: |
807e902e | 1848 | /* Pointer equality guarantees equality for these nodes. */ |
db048faf MM |
1849 | return 0; |
1850 | ||
db048faf MM |
1851 | default: |
1852 | break; | |
1853 | } | |
9ae8ffe7 | 1854 | |
45183e03 JH |
1855 | /* canon_rtx knows how to handle plus. No need to canonicalize. */ |
1856 | if (code == PLUS) | |
9ae8ffe7 JL |
1857 | return ((rtx_equal_for_memref_p (XEXP (x, 0), XEXP (y, 0)) |
1858 | && rtx_equal_for_memref_p (XEXP (x, 1), XEXP (y, 1))) | |
1859 | || (rtx_equal_for_memref_p (XEXP (x, 0), XEXP (y, 1)) | |
1860 | && rtx_equal_for_memref_p (XEXP (x, 1), XEXP (y, 0)))); | |
45183e03 JH |
1861 | /* For commutative operations, the RTX match if the operand match in any |
1862 | order. Also handle the simple binary and unary cases without a loop. */ | |
ec8e098d | 1863 | if (COMMUTATIVE_P (x)) |
45183e03 JH |
1864 | { |
1865 | rtx xop0 = canon_rtx (XEXP (x, 0)); | |
1866 | rtx yop0 = canon_rtx (XEXP (y, 0)); | |
1867 | rtx yop1 = canon_rtx (XEXP (y, 1)); | |
1868 | ||
1869 | return ((rtx_equal_for_memref_p (xop0, yop0) | |
1870 | && rtx_equal_for_memref_p (canon_rtx (XEXP (x, 1)), yop1)) | |
1871 | || (rtx_equal_for_memref_p (xop0, yop1) | |
1872 | && rtx_equal_for_memref_p (canon_rtx (XEXP (x, 1)), yop0))); | |
1873 | } | |
ec8e098d | 1874 | else if (NON_COMMUTATIVE_P (x)) |
45183e03 JH |
1875 | { |
1876 | return (rtx_equal_for_memref_p (canon_rtx (XEXP (x, 0)), | |
4682ae04 | 1877 | canon_rtx (XEXP (y, 0))) |
45183e03 JH |
1878 | && rtx_equal_for_memref_p (canon_rtx (XEXP (x, 1)), |
1879 | canon_rtx (XEXP (y, 1)))); | |
1880 | } | |
ec8e098d | 1881 | else if (UNARY_P (x)) |
45183e03 | 1882 | return rtx_equal_for_memref_p (canon_rtx (XEXP (x, 0)), |
4682ae04 | 1883 | canon_rtx (XEXP (y, 0))); |
9ae8ffe7 JL |
1884 | |
1885 | /* Compare the elements. If any pair of corresponding elements | |
de12be17 JC |
1886 | fail to match, return 0 for the whole things. |
1887 | ||
1888 | Limit cases to types which actually appear in addresses. */ | |
9ae8ffe7 JL |
1889 | |
1890 | fmt = GET_RTX_FORMAT (code); | |
1891 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
1892 | { | |
1893 | switch (fmt[i]) | |
1894 | { | |
9ae8ffe7 JL |
1895 | case 'i': |
1896 | if (XINT (x, i) != XINT (y, i)) | |
1897 | return 0; | |
1898 | break; | |
1899 | ||
91914e56 RS |
1900 | case 'p': |
1901 | if (maybe_ne (SUBREG_BYTE (x), SUBREG_BYTE (y))) | |
1902 | return 0; | |
1903 | break; | |
1904 | ||
9ae8ffe7 JL |
1905 | case 'E': |
1906 | /* Two vectors must have the same length. */ | |
1907 | if (XVECLEN (x, i) != XVECLEN (y, i)) | |
1908 | return 0; | |
1909 | ||
1910 | /* And the corresponding elements must match. */ | |
1911 | for (j = 0; j < XVECLEN (x, i); j++) | |
45183e03 JH |
1912 | if (rtx_equal_for_memref_p (canon_rtx (XVECEXP (x, i, j)), |
1913 | canon_rtx (XVECEXP (y, i, j))) == 0) | |
9ae8ffe7 JL |
1914 | return 0; |
1915 | break; | |
1916 | ||
1917 | case 'e': | |
45183e03 JH |
1918 | if (rtx_equal_for_memref_p (canon_rtx (XEXP (x, i)), |
1919 | canon_rtx (XEXP (y, i))) == 0) | |
9ae8ffe7 JL |
1920 | return 0; |
1921 | break; | |
1922 | ||
3237ac18 AH |
1923 | /* This can happen for asm operands. */ |
1924 | case 's': | |
1925 | if (strcmp (XSTR (x, i), XSTR (y, i))) | |
1926 | return 0; | |
1927 | break; | |
1928 | ||
aee21ba9 JL |
1929 | /* This can happen for an asm which clobbers memory. */ |
1930 | case '0': | |
1931 | break; | |
1932 | ||
9ae8ffe7 JL |
1933 | /* It is believed that rtx's at this level will never |
1934 | contain anything but integers and other rtx's, | |
1935 | except for within LABEL_REFs and SYMBOL_REFs. */ | |
1936 | default: | |
298e6adc | 1937 | gcc_unreachable (); |
9ae8ffe7 JL |
1938 | } |
1939 | } | |
1940 | return 1; | |
1941 | } | |
1942 | ||
9e412ca3 | 1943 | static rtx |
6ca83833 RB |
1944 | find_base_term (rtx x, vec<std::pair<cselib_val *, |
1945 | struct elt_loc_list *> > &visited_vals) | |
9ae8ffe7 | 1946 | { |
eab5c70a | 1947 | cselib_val *val; |
6f2ffb4b AO |
1948 | struct elt_loc_list *l, *f; |
1949 | rtx ret; | |
6645d841 | 1950 | scalar_int_mode int_mode; |
eab5c70a | 1951 | |
b949ea8b JW |
1952 | #if defined (FIND_BASE_TERM) |
1953 | /* Try machine-dependent ways to find the base term. */ | |
1954 | x = FIND_BASE_TERM (x); | |
1955 | #endif | |
1956 | ||
9ae8ffe7 JL |
1957 | switch (GET_CODE (x)) |
1958 | { | |
1959 | case REG: | |
1960 | return REG_BASE_VALUE (x); | |
1961 | ||
d288e53d | 1962 | case TRUNCATE: |
5932a4d4 | 1963 | /* As we do not know which address space the pointer is referring to, we can |
d4ebfa65 BE |
1964 | handle this only if the target does not support different pointer or |
1965 | address modes depending on the address space. */ | |
1966 | if (!target_default_pointer_address_modes_p ()) | |
1967 | return 0; | |
6645d841 RS |
1968 | if (!is_a <scalar_int_mode> (GET_MODE (x), &int_mode) |
1969 | || GET_MODE_PRECISION (int_mode) < GET_MODE_PRECISION (Pmode)) | |
ca7fd9cd | 1970 | return 0; |
d288e53d | 1971 | /* Fall through. */ |
9ae8ffe7 | 1972 | case HIGH: |
6d849a2a JL |
1973 | case PRE_INC: |
1974 | case PRE_DEC: | |
1975 | case POST_INC: | |
1976 | case POST_DEC: | |
d288e53d DE |
1977 | case PRE_MODIFY: |
1978 | case POST_MODIFY: | |
6ca83833 | 1979 | return find_base_term (XEXP (x, 0), visited_vals); |
6d849a2a | 1980 | |
1abade85 RK |
1981 | case ZERO_EXTEND: |
1982 | case SIGN_EXTEND: /* Used for Alpha/NT pointers */ | |
5932a4d4 | 1983 | /* As we do not know which address space the pointer is referring to, we can |
d4ebfa65 BE |
1984 | handle this only if the target does not support different pointer or |
1985 | address modes depending on the address space. */ | |
1986 | if (!target_default_pointer_address_modes_p ()) | |
1987 | return 0; | |
1988 | ||
1abade85 | 1989 | { |
6ca83833 | 1990 | rtx temp = find_base_term (XEXP (x, 0), visited_vals); |
1abade85 | 1991 | |
5ae6cd0d | 1992 | if (temp != 0 && CONSTANT_P (temp)) |
1abade85 | 1993 | temp = convert_memory_address (Pmode, temp); |
1abade85 RK |
1994 | |
1995 | return temp; | |
1996 | } | |
1997 | ||
eab5c70a BS |
1998 | case VALUE: |
1999 | val = CSELIB_VAL_PTR (x); | |
6f2ffb4b AO |
2000 | ret = NULL_RTX; |
2001 | ||
40e02b4a | 2002 | if (!val) |
6f2ffb4b AO |
2003 | return ret; |
2004 | ||
0fe03ac3 JJ |
2005 | if (cselib_sp_based_value_p (val)) |
2006 | return static_reg_base_value[STACK_POINTER_REGNUM]; | |
2007 | ||
2e94d3ee JJ |
2008 | if (visited_vals.length () > (unsigned) param_max_find_base_term_values) |
2009 | return ret; | |
2010 | ||
6f2ffb4b | 2011 | f = val->locs; |
6ca83833 RB |
2012 | /* Reset val->locs to avoid infinite recursion. */ |
2013 | if (f) | |
2014 | visited_vals.safe_push (std::make_pair (val, f)); | |
6f2ffb4b AO |
2015 | val->locs = NULL; |
2016 | ||
2017 | for (l = f; l; l = l->next) | |
2018 | if (GET_CODE (l->loc) == VALUE | |
2019 | && CSELIB_VAL_PTR (l->loc)->locs | |
2020 | && !CSELIB_VAL_PTR (l->loc)->locs->next | |
2021 | && CSELIB_VAL_PTR (l->loc)->locs->loc == x) | |
2022 | continue; | |
6ca83833 | 2023 | else if ((ret = find_base_term (l->loc, visited_vals)) != 0) |
6f2ffb4b AO |
2024 | break; |
2025 | ||
6f2ffb4b | 2026 | return ret; |
eab5c70a | 2027 | |
023f059b JJ |
2028 | case LO_SUM: |
2029 | /* The standard form is (lo_sum reg sym) so look only at the | |
2030 | second operand. */ | |
6ca83833 | 2031 | return find_base_term (XEXP (x, 1), visited_vals); |
023f059b | 2032 | |
9ae8ffe7 JL |
2033 | case CONST: |
2034 | x = XEXP (x, 0); | |
2035 | if (GET_CODE (x) != PLUS && GET_CODE (x) != MINUS) | |
2036 | return 0; | |
938d968e | 2037 | /* Fall through. */ |
9ae8ffe7 JL |
2038 | case PLUS: |
2039 | case MINUS: | |
2040 | { | |
3c567fae JL |
2041 | rtx tmp1 = XEXP (x, 0); |
2042 | rtx tmp2 = XEXP (x, 1); | |
2043 | ||
f5143c46 | 2044 | /* This is a little bit tricky since we have to determine which of |
3c567fae JL |
2045 | the two operands represents the real base address. Otherwise this |
2046 | routine may return the index register instead of the base register. | |
2047 | ||
2048 | That may cause us to believe no aliasing was possible, when in | |
2049 | fact aliasing is possible. | |
2050 | ||
2051 | We use a few simple tests to guess the base register. Additional | |
2052 | tests can certainly be added. For example, if one of the operands | |
2053 | is a shift or multiply, then it must be the index register and the | |
2054 | other operand is the base register. */ | |
ca7fd9cd | 2055 | |
b949ea8b | 2056 | if (tmp1 == pic_offset_table_rtx && CONSTANT_P (tmp2)) |
6ca83833 | 2057 | return find_base_term (tmp2, visited_vals); |
b949ea8b | 2058 | |
31b0a960 | 2059 | /* If either operand is known to be a pointer, then prefer it |
3c567fae | 2060 | to determine the base term. */ |
3502dc9c | 2061 | if (REG_P (tmp1) && REG_POINTER (tmp1)) |
31b0a960 RB |
2062 | ; |
2063 | else if (REG_P (tmp2) && REG_POINTER (tmp2)) | |
a7c75343 JJ |
2064 | std::swap (tmp1, tmp2); |
2065 | /* If second argument is constant which has base term, prefer it | |
2066 | over variable tmp1. See PR64025. */ | |
2067 | else if (CONSTANT_P (tmp2) && !CONST_INT_P (tmp2)) | |
2068 | std::swap (tmp1, tmp2); | |
3c567fae | 2069 | |
31b0a960 RB |
2070 | /* Go ahead and find the base term for both operands. If either base |
2071 | term is from a pointer or is a named object or a special address | |
3c567fae JL |
2072 | (like an argument or stack reference), then use it for the |
2073 | base term. */ | |
6ca83833 | 2074 | rtx base = find_base_term (tmp1, visited_vals); |
481be1c4 | 2075 | if (base != NULL_RTX |
31b0a960 | 2076 | && ((REG_P (tmp1) && REG_POINTER (tmp1)) |
481be1c4 RB |
2077 | || known_base_value_p (base))) |
2078 | return base; | |
6ca83833 | 2079 | base = find_base_term (tmp2, visited_vals); |
481be1c4 | 2080 | if (base != NULL_RTX |
31b0a960 | 2081 | && ((REG_P (tmp2) && REG_POINTER (tmp2)) |
481be1c4 RB |
2082 | || known_base_value_p (base))) |
2083 | return base; | |
3c567fae JL |
2084 | |
2085 | /* We could not determine which of the two operands was the | |
2086 | base register and which was the index. So we can determine | |
2087 | nothing from the base alias check. */ | |
2088 | return 0; | |
9ae8ffe7 JL |
2089 | } |
2090 | ||
2091 | case AND: | |
ef50b972 RB |
2092 | /* Look through aligning ANDs. And AND with zero or one with |
2093 | the LSB set isn't one (see for example PR92462). */ | |
2094 | if (CONST_INT_P (XEXP (x, 1)) | |
2095 | && INTVAL (XEXP (x, 1)) != 0 | |
2096 | && (INTVAL (XEXP (x, 1)) & 1) == 0) | |
6ca83833 | 2097 | return find_base_term (XEXP (x, 0), visited_vals); |
9ae8ffe7 JL |
2098 | return 0; |
2099 | ||
2100 | case SYMBOL_REF: | |
2101 | case LABEL_REF: | |
2102 | return x; | |
2103 | ||
2104 | default: | |
2105 | return 0; | |
2106 | } | |
2107 | } | |
2108 | ||
6ca83833 RB |
2109 | /* Wrapper around the worker above which removes locs from visited VALUEs |
2110 | to avoid visiting them multiple times. We unwind that changes here. */ | |
2111 | ||
2112 | static rtx | |
2113 | find_base_term (rtx x) | |
2114 | { | |
2115 | auto_vec<std::pair<cselib_val *, struct elt_loc_list *>, 32> visited_vals; | |
2116 | rtx res = find_base_term (x, visited_vals); | |
2117 | for (unsigned i = 0; i < visited_vals.length (); ++i) | |
2118 | visited_vals[i].first->locs = visited_vals[i].second; | |
2119 | return res; | |
2120 | } | |
2121 | ||
9e412ca3 RS |
2122 | /* Return true if accesses to address X may alias accesses based |
2123 | on the stack pointer. */ | |
2124 | ||
2125 | bool | |
2126 | may_be_sp_based_p (rtx x) | |
2127 | { | |
2128 | rtx base = find_base_term (x); | |
2129 | return !base || base == static_reg_base_value[STACK_POINTER_REGNUM]; | |
2130 | } | |
2131 | ||
54363f8a | 2132 | /* BASE1 and BASE2 are decls. Return 1 if they refer to same object, 0 |
67914693 | 2133 | if they refer to different objects and -1 if we cannot decide. */ |
54363f8a JH |
2134 | |
2135 | int | |
2136 | compare_base_decls (tree base1, tree base2) | |
2137 | { | |
2138 | int ret; | |
2139 | gcc_checking_assert (DECL_P (base1) && DECL_P (base2)); | |
2140 | if (base1 == base2) | |
2141 | return 1; | |
2142 | ||
bed3fd46 | 2143 | /* If we have two register decls with register specification we |
816c4ba2 | 2144 | cannot decide unless their assembler names are the same. */ |
bed3fd46 RB |
2145 | if (DECL_REGISTER (base1) |
2146 | && DECL_REGISTER (base2) | |
816c4ba2 NS |
2147 | && HAS_DECL_ASSEMBLER_NAME_P (base1) |
2148 | && HAS_DECL_ASSEMBLER_NAME_P (base2) | |
bed3fd46 RB |
2149 | && DECL_ASSEMBLER_NAME_SET_P (base1) |
2150 | && DECL_ASSEMBLER_NAME_SET_P (base2)) | |
2151 | { | |
816c4ba2 | 2152 | if (DECL_ASSEMBLER_NAME_RAW (base1) == DECL_ASSEMBLER_NAME_RAW (base2)) |
bed3fd46 RB |
2153 | return 1; |
2154 | return -1; | |
2155 | } | |
2156 | ||
54363f8a JH |
2157 | /* Declarations of non-automatic variables may have aliases. All other |
2158 | decls are unique. */ | |
7ec4f343 NS |
2159 | if (!decl_in_symtab_p (base1) |
2160 | || !decl_in_symtab_p (base2)) | |
54363f8a | 2161 | return 0; |
7ec4f343 | 2162 | |
929710d9 NS |
2163 | /* Don't cause symbols to be inserted by the act of checking. */ |
2164 | symtab_node *node1 = symtab_node::get (base1); | |
2165 | if (!node1) | |
2166 | return 0; | |
2167 | symtab_node *node2 = symtab_node::get (base2); | |
2168 | if (!node2) | |
2169 | return 0; | |
2170 | ||
2171 | ret = node1->equal_address_to (node2, true); | |
54363f8a JH |
2172 | return ret; |
2173 | } | |
2174 | ||
73e48cb3 JH |
2175 | /* Same as compare_base_decls but for SYMBOL_REF. */ |
2176 | ||
2177 | static int | |
2178 | compare_base_symbol_refs (const_rtx x_base, const_rtx y_base) | |
2179 | { | |
2180 | tree x_decl = SYMBOL_REF_DECL (x_base); | |
2181 | tree y_decl = SYMBOL_REF_DECL (y_base); | |
2182 | bool binds_def = true; | |
2183 | ||
2184 | if (XSTR (x_base, 0) == XSTR (y_base, 0)) | |
2185 | return 1; | |
2186 | if (x_decl && y_decl) | |
2187 | return compare_base_decls (x_decl, y_decl); | |
2188 | if (x_decl || y_decl) | |
2189 | { | |
2190 | if (!x_decl) | |
2191 | { | |
2192 | std::swap (x_decl, y_decl); | |
2193 | std::swap (x_base, y_base); | |
2194 | } | |
2195 | /* We handle specially only section anchors and assume that other | |
2196 | labels may overlap with user variables in an arbitrary way. */ | |
2197 | if (!SYMBOL_REF_HAS_BLOCK_INFO_P (y_base)) | |
2198 | return -1; | |
2199 | /* Anchors contains static VAR_DECLs and CONST_DECLs. We are safe | |
2200 | to ignore CONST_DECLs because they are readonly. */ | |
8813a647 | 2201 | if (!VAR_P (x_decl) |
73e48cb3 JH |
2202 | || (!TREE_STATIC (x_decl) && !TREE_PUBLIC (x_decl))) |
2203 | return 0; | |
2204 | ||
2205 | symtab_node *x_node = symtab_node::get_create (x_decl) | |
2206 | ->ultimate_alias_target (); | |
67914693 | 2207 | /* External variable cannot be in section anchor. */ |
73e48cb3 JH |
2208 | if (!x_node->definition) |
2209 | return 0; | |
2210 | x_base = XEXP (DECL_RTL (x_node->decl), 0); | |
2211 | /* If not in anchor, we can disambiguate. */ | |
2212 | if (!SYMBOL_REF_HAS_BLOCK_INFO_P (x_base)) | |
2213 | return 0; | |
2214 | ||
2215 | /* We have an alias of anchored variable. If it can be interposed; | |
2216 | we must assume it may or may not alias its anchor. */ | |
2217 | binds_def = decl_binds_to_current_def_p (x_decl); | |
2218 | } | |
2219 | /* If we have variable in section anchor, we can compare by offset. */ | |
2220 | if (SYMBOL_REF_HAS_BLOCK_INFO_P (x_base) | |
2221 | && SYMBOL_REF_HAS_BLOCK_INFO_P (y_base)) | |
2222 | { | |
2223 | if (SYMBOL_REF_BLOCK (x_base) != SYMBOL_REF_BLOCK (y_base)) | |
2224 | return 0; | |
2225 | if (SYMBOL_REF_BLOCK_OFFSET (x_base) == SYMBOL_REF_BLOCK_OFFSET (y_base)) | |
2226 | return binds_def ? 1 : -1; | |
2227 | if (SYMBOL_REF_ANCHOR_P (x_base) != SYMBOL_REF_ANCHOR_P (y_base)) | |
2228 | return -1; | |
2229 | return 0; | |
2230 | } | |
2231 | /* In general we assume that memory locations pointed to by different labels | |
2232 | may overlap in undefined ways. */ | |
2233 | return -1; | |
2234 | } | |
2235 | ||
9ae8ffe7 JL |
2236 | /* Return 0 if the addresses X and Y are known to point to different |
2237 | objects, 1 if they might be pointers to the same object. */ | |
2238 | ||
2239 | static int | |
31b0a960 | 2240 | base_alias_check (rtx x, rtx x_base, rtx y, rtx y_base, |
ef4bddc2 | 2241 | machine_mode x_mode, machine_mode y_mode) |
9ae8ffe7 | 2242 | { |
1c72c7f6 JC |
2243 | /* If the address itself has no known base see if a known equivalent |
2244 | value has one. If either address still has no known base, nothing | |
2245 | is known about aliasing. */ | |
2246 | if (x_base == 0) | |
2247 | { | |
2248 | rtx x_c; | |
d4b60170 | 2249 | |
1c72c7f6 JC |
2250 | if (! flag_expensive_optimizations || (x_c = canon_rtx (x)) == x) |
2251 | return 1; | |
d4b60170 | 2252 | |
1c72c7f6 JC |
2253 | x_base = find_base_term (x_c); |
2254 | if (x_base == 0) | |
2255 | return 1; | |
2256 | } | |
9ae8ffe7 | 2257 | |
1c72c7f6 JC |
2258 | if (y_base == 0) |
2259 | { | |
2260 | rtx y_c; | |
2261 | if (! flag_expensive_optimizations || (y_c = canon_rtx (y)) == y) | |
2262 | return 1; | |
d4b60170 | 2263 | |
1c72c7f6 JC |
2264 | y_base = find_base_term (y_c); |
2265 | if (y_base == 0) | |
2266 | return 1; | |
2267 | } | |
2268 | ||
2269 | /* If the base addresses are equal nothing is known about aliasing. */ | |
2270 | if (rtx_equal_p (x_base, y_base)) | |
9ae8ffe7 JL |
2271 | return 1; |
2272 | ||
435da628 UB |
2273 | /* The base addresses are different expressions. If they are not accessed |
2274 | via AND, there is no conflict. We can bring knowledge of object | |
2275 | alignment into play here. For example, on alpha, "char a, b;" can | |
5764ee3c | 2276 | alias one another, though "char a; long b;" cannot. AND addresses may |
435da628 UB |
2277 | implicitly alias surrounding objects; i.e. unaligned access in DImode |
2278 | via AND address can alias all surrounding object types except those | |
2279 | with aligment 8 or higher. */ | |
2280 | if (GET_CODE (x) == AND && GET_CODE (y) == AND) | |
2281 | return 1; | |
2282 | if (GET_CODE (x) == AND | |
481683e1 | 2283 | && (!CONST_INT_P (XEXP (x, 1)) |
435da628 UB |
2284 | || (int) GET_MODE_UNIT_SIZE (y_mode) < -INTVAL (XEXP (x, 1)))) |
2285 | return 1; | |
2286 | if (GET_CODE (y) == AND | |
481683e1 | 2287 | && (!CONST_INT_P (XEXP (y, 1)) |
435da628 UB |
2288 | || (int) GET_MODE_UNIT_SIZE (x_mode) < -INTVAL (XEXP (y, 1)))) |
2289 | return 1; | |
2290 | ||
73e48cb3 | 2291 | /* Differing symbols not accessed via AND never alias. */ |
3a28db46 | 2292 | if (GET_CODE (x_base) == SYMBOL_REF && GET_CODE (y_base) == SYMBOL_REF) |
73e48cb3 | 2293 | return compare_base_symbol_refs (x_base, y_base) != 0; |
3a28db46 | 2294 | |
9ae8ffe7 | 2295 | if (GET_CODE (x_base) != ADDRESS && GET_CODE (y_base) != ADDRESS) |
435da628 | 2296 | return 0; |
9ae8ffe7 | 2297 | |
9fc37b2b | 2298 | if (unique_base_value_p (x_base) || unique_base_value_p (y_base)) |
9ae8ffe7 JL |
2299 | return 0; |
2300 | ||
0d3c82d6 | 2301 | return 1; |
9ae8ffe7 JL |
2302 | } |
2303 | ||
a5628378 | 2304 | /* Return TRUE if EXPR refers to a VALUE whose uid is greater than |
c779924e | 2305 | (or equal to) that of V. */ |
a5628378 AO |
2306 | |
2307 | static bool | |
403837b4 | 2308 | refs_newer_value_p (const_rtx expr, rtx v) |
a5628378 AO |
2309 | { |
2310 | int minuid = CSELIB_VAL_PTR (v)->uid; | |
403837b4 RS |
2311 | subrtx_iterator::array_type array; |
2312 | FOR_EACH_SUBRTX (iter, array, expr, NONCONST) | |
c779924e | 2313 | if (GET_CODE (*iter) == VALUE && CSELIB_VAL_PTR (*iter)->uid >= minuid) |
403837b4 RS |
2314 | return true; |
2315 | return false; | |
a5628378 AO |
2316 | } |
2317 | ||
eab5c70a | 2318 | /* Convert the address X into something we can use. This is done by returning |
569efc34 JJ |
2319 | it unchanged unless it is a VALUE or VALUE +/- constant; for VALUE |
2320 | we call cselib to get a more useful rtx. */ | |
3bdf5ad1 | 2321 | |
a13d4ebf | 2322 | rtx |
4682ae04 | 2323 | get_addr (rtx x) |
eab5c70a BS |
2324 | { |
2325 | cselib_val *v; | |
2326 | struct elt_loc_list *l; | |
2327 | ||
2328 | if (GET_CODE (x) != VALUE) | |
569efc34 JJ |
2329 | { |
2330 | if ((GET_CODE (x) == PLUS || GET_CODE (x) == MINUS) | |
2331 | && GET_CODE (XEXP (x, 0)) == VALUE | |
2332 | && CONST_SCALAR_INT_P (XEXP (x, 1))) | |
2333 | { | |
2334 | rtx op0 = get_addr (XEXP (x, 0)); | |
2335 | if (op0 != XEXP (x, 0)) | |
2336 | { | |
5284e559 | 2337 | poly_int64 c; |
569efc34 | 2338 | if (GET_CODE (x) == PLUS |
5284e559 RS |
2339 | && poly_int_rtx_p (XEXP (x, 1), &c)) |
2340 | return plus_constant (GET_MODE (x), op0, c); | |
569efc34 JJ |
2341 | return simplify_gen_binary (GET_CODE (x), GET_MODE (x), |
2342 | op0, XEXP (x, 1)); | |
2343 | } | |
2344 | } | |
2345 | return x; | |
2346 | } | |
eab5c70a | 2347 | v = CSELIB_VAL_PTR (x); |
40e02b4a JH |
2348 | if (v) |
2349 | { | |
0f68ba3e AO |
2350 | bool have_equivs = cselib_have_permanent_equivalences (); |
2351 | if (have_equivs) | |
2352 | v = canonical_cselib_val (v); | |
40e02b4a JH |
2353 | for (l = v->locs; l; l = l->next) |
2354 | if (CONSTANT_P (l->loc)) | |
2355 | return l->loc; | |
2356 | for (l = v->locs; l; l = l->next) | |
0f68ba3e AO |
2357 | if (!REG_P (l->loc) && !MEM_P (l->loc) |
2358 | /* Avoid infinite recursion when potentially dealing with | |
2359 | var-tracking artificial equivalences, by skipping the | |
2360 | equivalences themselves, and not choosing expressions | |
2361 | that refer to newer VALUEs. */ | |
2362 | && (!have_equivs | |
2363 | || (GET_CODE (l->loc) != VALUE | |
2364 | && !refs_newer_value_p (l->loc, x)))) | |
a5628378 | 2365 | return l->loc; |
0f68ba3e AO |
2366 | if (have_equivs) |
2367 | { | |
2368 | for (l = v->locs; l; l = l->next) | |
2369 | if (REG_P (l->loc) | |
2370 | || (GET_CODE (l->loc) != VALUE | |
2371 | && !refs_newer_value_p (l->loc, x))) | |
2372 | return l->loc; | |
2373 | /* Return the canonical value. */ | |
2374 | return v->val_rtx; | |
2375 | } | |
2376 | if (v->locs) | |
2377 | return v->locs->loc; | |
40e02b4a | 2378 | } |
eab5c70a BS |
2379 | return x; |
2380 | } | |
2381 | ||
39cec1ac MH |
2382 | /* Return the address of the (N_REFS + 1)th memory reference to ADDR |
2383 | where SIZE is the size in bytes of the memory reference. If ADDR | |
2384 | is not modified by the memory reference then ADDR is returned. */ | |
2385 | ||
04e2b4d3 | 2386 | static rtx |
9f61be58 | 2387 | addr_side_effect_eval (rtx addr, poly_int64 size, int n_refs) |
39cec1ac | 2388 | { |
9f61be58 | 2389 | poly_int64 offset = 0; |
ca7fd9cd | 2390 | |
39cec1ac MH |
2391 | switch (GET_CODE (addr)) |
2392 | { | |
2393 | case PRE_INC: | |
2394 | offset = (n_refs + 1) * size; | |
2395 | break; | |
2396 | case PRE_DEC: | |
2397 | offset = -(n_refs + 1) * size; | |
2398 | break; | |
2399 | case POST_INC: | |
2400 | offset = n_refs * size; | |
2401 | break; | |
2402 | case POST_DEC: | |
2403 | offset = -n_refs * size; | |
2404 | break; | |
2405 | ||
2406 | default: | |
2407 | return addr; | |
2408 | } | |
ca7fd9cd | 2409 | |
9f61be58 | 2410 | addr = plus_constant (GET_MODE (addr), XEXP (addr, 0), offset); |
45183e03 | 2411 | addr = canon_rtx (addr); |
39cec1ac MH |
2412 | |
2413 | return addr; | |
2414 | } | |
2415 | ||
3aa03517 AO |
2416 | /* Return TRUE if an object X sized at XSIZE bytes and another object |
2417 | Y sized at YSIZE bytes, starting C bytes after X, may overlap. If | |
2418 | any of the sizes is zero, assume an overlap, otherwise use the | |
2419 | absolute value of the sizes as the actual sizes. */ | |
2420 | ||
2421 | static inline bool | |
d05d7551 | 2422 | offset_overlap_p (poly_int64 c, poly_int64 xsize, poly_int64 ysize) |
3aa03517 | 2423 | { |
d05d7551 RS |
2424 | if (known_eq (xsize, 0) || known_eq (ysize, 0)) |
2425 | return true; | |
2426 | ||
2427 | if (maybe_ge (c, 0)) | |
2428 | return maybe_gt (maybe_lt (xsize, 0) ? -xsize : xsize, c); | |
2429 | else | |
2430 | return maybe_gt (maybe_lt (ysize, 0) ? -ysize : ysize, -c); | |
3aa03517 AO |
2431 | } |
2432 | ||
f47e08d9 RG |
2433 | /* Return one if X and Y (memory addresses) reference the |
2434 | same location in memory or if the references overlap. | |
2435 | Return zero if they do not overlap, else return | |
2436 | minus one in which case they still might reference the same location. | |
2437 | ||
2438 | C is an offset accumulator. When | |
9ae8ffe7 JL |
2439 | C is nonzero, we are testing aliases between X and Y + C. |
2440 | XSIZE is the size in bytes of the X reference, | |
2441 | similarly YSIZE is the size in bytes for Y. | |
45183e03 | 2442 | Expect that canon_rtx has been already called for X and Y. |
9ae8ffe7 JL |
2443 | |
2444 | If XSIZE or YSIZE is zero, we do not know the amount of memory being | |
2445 | referenced (the reference was BLKmode), so make the most pessimistic | |
2446 | assumptions. | |
2447 | ||
c02f035f RH |
2448 | If XSIZE or YSIZE is negative, we may access memory outside the object |
2449 | being referenced as a side effect. This can happen when using AND to | |
2450 | align memory references, as is done on the Alpha. | |
2451 | ||
9ae8ffe7 | 2452 | Nice to notice that varying addresses cannot conflict with fp if no |
f47e08d9 RG |
2453 | local variables had their addresses taken, but that's too hard now. |
2454 | ||
2455 | ??? Contrary to the tree alias oracle this does not return | |
2456 | one for X + non-constant and Y + non-constant when X and Y are equal. | |
2457 | If that is fixed the TBAA hack for union type-punning can be removed. */ | |
9ae8ffe7 | 2458 | |
9ae8ffe7 | 2459 | static int |
9f61be58 RS |
2460 | memrefs_conflict_p (poly_int64 xsize, rtx x, poly_int64 ysize, rtx y, |
2461 | poly_int64 c) | |
9ae8ffe7 | 2462 | { |
eab5c70a | 2463 | if (GET_CODE (x) == VALUE) |
5312b066 JJ |
2464 | { |
2465 | if (REG_P (y)) | |
2466 | { | |
24f8d71e JJ |
2467 | struct elt_loc_list *l = NULL; |
2468 | if (CSELIB_VAL_PTR (x)) | |
a5628378 AO |
2469 | for (l = canonical_cselib_val (CSELIB_VAL_PTR (x))->locs; |
2470 | l; l = l->next) | |
24f8d71e JJ |
2471 | if (REG_P (l->loc) && rtx_equal_for_memref_p (l->loc, y)) |
2472 | break; | |
5312b066 JJ |
2473 | if (l) |
2474 | x = y; | |
2475 | else | |
2476 | x = get_addr (x); | |
2477 | } | |
2478 | /* Don't call get_addr if y is the same VALUE. */ | |
2479 | else if (x != y) | |
2480 | x = get_addr (x); | |
2481 | } | |
eab5c70a | 2482 | if (GET_CODE (y) == VALUE) |
5312b066 JJ |
2483 | { |
2484 | if (REG_P (x)) | |
2485 | { | |
24f8d71e JJ |
2486 | struct elt_loc_list *l = NULL; |
2487 | if (CSELIB_VAL_PTR (y)) | |
a5628378 AO |
2488 | for (l = canonical_cselib_val (CSELIB_VAL_PTR (y))->locs; |
2489 | l; l = l->next) | |
24f8d71e JJ |
2490 | if (REG_P (l->loc) && rtx_equal_for_memref_p (l->loc, x)) |
2491 | break; | |
5312b066 JJ |
2492 | if (l) |
2493 | y = x; | |
2494 | else | |
2495 | y = get_addr (y); | |
2496 | } | |
2497 | /* Don't call get_addr if x is the same VALUE. */ | |
2498 | else if (y != x) | |
2499 | y = get_addr (y); | |
2500 | } | |
9ae8ffe7 JL |
2501 | if (GET_CODE (x) == HIGH) |
2502 | x = XEXP (x, 0); | |
2503 | else if (GET_CODE (x) == LO_SUM) | |
2504 | x = XEXP (x, 1); | |
2505 | else | |
9f61be58 | 2506 | x = addr_side_effect_eval (x, maybe_lt (xsize, 0) ? -xsize : xsize, 0); |
9ae8ffe7 JL |
2507 | if (GET_CODE (y) == HIGH) |
2508 | y = XEXP (y, 0); | |
2509 | else if (GET_CODE (y) == LO_SUM) | |
2510 | y = XEXP (y, 1); | |
2511 | else | |
9f61be58 | 2512 | y = addr_side_effect_eval (y, maybe_lt (ysize, 0) ? -ysize : ysize, 0); |
9ae8ffe7 | 2513 | |
54363f8a JH |
2514 | if (GET_CODE (x) == SYMBOL_REF && GET_CODE (y) == SYMBOL_REF) |
2515 | { | |
73e48cb3 | 2516 | int cmp = compare_base_symbol_refs (x,y); |
54363f8a JH |
2517 | |
2518 | /* If both decls are the same, decide by offsets. */ | |
2519 | if (cmp == 1) | |
2520 | return offset_overlap_p (c, xsize, ysize); | |
3a28db46 UB |
2521 | /* Assume a potential overlap for symbolic addresses that went |
2522 | through alignment adjustments (i.e., that have negative | |
2523 | sizes), because we can't know how far they are from each | |
2524 | other. */ | |
9f61be58 | 2525 | if (maybe_lt (xsize, 0) || maybe_lt (ysize, 0)) |
3a28db46 | 2526 | return -1; |
54363f8a JH |
2527 | /* If decls are different or we know by offsets that there is no overlap, |
2528 | we win. */ | |
2529 | if (!cmp || !offset_overlap_p (c, xsize, ysize)) | |
2530 | return 0; | |
2531 | /* Decls may or may not be different and offsets overlap....*/ | |
2532 | return -1; | |
2533 | } | |
2534 | else if (rtx_equal_for_memref_p (x, y)) | |
9ae8ffe7 | 2535 | { |
3aa03517 | 2536 | return offset_overlap_p (c, xsize, ysize); |
9ae8ffe7 JL |
2537 | } |
2538 | ||
6e73e666 JC |
2539 | /* This code used to check for conflicts involving stack references and |
2540 | globals but the base address alias code now handles these cases. */ | |
9ae8ffe7 JL |
2541 | |
2542 | if (GET_CODE (x) == PLUS) | |
2543 | { | |
2544 | /* The fact that X is canonicalized means that this | |
2545 | PLUS rtx is canonicalized. */ | |
2546 | rtx x0 = XEXP (x, 0); | |
2547 | rtx x1 = XEXP (x, 1); | |
2548 | ||
2d88904a AO |
2549 | /* However, VALUEs might end up in different positions even in |
2550 | canonical PLUSes. Comparing their addresses is enough. */ | |
2551 | if (x0 == y) | |
2552 | return memrefs_conflict_p (xsize, x1, ysize, const0_rtx, c); | |
2553 | else if (x1 == y) | |
2554 | return memrefs_conflict_p (xsize, x0, ysize, const0_rtx, c); | |
2555 | ||
9f61be58 | 2556 | poly_int64 cx1, cy1; |
9ae8ffe7 JL |
2557 | if (GET_CODE (y) == PLUS) |
2558 | { | |
2559 | /* The fact that Y is canonicalized means that this | |
2560 | PLUS rtx is canonicalized. */ | |
2561 | rtx y0 = XEXP (y, 0); | |
2562 | rtx y1 = XEXP (y, 1); | |
2563 | ||
2d88904a AO |
2564 | if (x0 == y1) |
2565 | return memrefs_conflict_p (xsize, x1, ysize, y0, c); | |
2566 | if (x1 == y0) | |
2567 | return memrefs_conflict_p (xsize, x0, ysize, y1, c); | |
2568 | ||
9ae8ffe7 JL |
2569 | if (rtx_equal_for_memref_p (x1, y1)) |
2570 | return memrefs_conflict_p (xsize, x0, ysize, y0, c); | |
2571 | if (rtx_equal_for_memref_p (x0, y0)) | |
2572 | return memrefs_conflict_p (xsize, x1, ysize, y1, c); | |
9f61be58 | 2573 | if (poly_int_rtx_p (x1, &cx1)) |
63be02db | 2574 | { |
9f61be58 | 2575 | if (poly_int_rtx_p (y1, &cy1)) |
63be02db | 2576 | return memrefs_conflict_p (xsize, x0, ysize, y0, |
9f61be58 | 2577 | c - cx1 + cy1); |
63be02db | 2578 | else |
9f61be58 | 2579 | return memrefs_conflict_p (xsize, x0, ysize, y, c - cx1); |
63be02db | 2580 | } |
9f61be58 RS |
2581 | else if (poly_int_rtx_p (y1, &cy1)) |
2582 | return memrefs_conflict_p (xsize, x, ysize, y0, c + cy1); | |
9ae8ffe7 | 2583 | |
f47e08d9 | 2584 | return -1; |
9ae8ffe7 | 2585 | } |
9f61be58 RS |
2586 | else if (poly_int_rtx_p (x1, &cx1)) |
2587 | return memrefs_conflict_p (xsize, x0, ysize, y, c - cx1); | |
9ae8ffe7 JL |
2588 | } |
2589 | else if (GET_CODE (y) == PLUS) | |
2590 | { | |
2591 | /* The fact that Y is canonicalized means that this | |
2592 | PLUS rtx is canonicalized. */ | |
2593 | rtx y0 = XEXP (y, 0); | |
2594 | rtx y1 = XEXP (y, 1); | |
2595 | ||
2d88904a AO |
2596 | if (x == y0) |
2597 | return memrefs_conflict_p (xsize, const0_rtx, ysize, y1, c); | |
2598 | if (x == y1) | |
2599 | return memrefs_conflict_p (xsize, const0_rtx, ysize, y0, c); | |
2600 | ||
9f61be58 RS |
2601 | poly_int64 cy1; |
2602 | if (poly_int_rtx_p (y1, &cy1)) | |
2603 | return memrefs_conflict_p (xsize, x, ysize, y0, c + cy1); | |
9ae8ffe7 | 2604 | else |
f47e08d9 | 2605 | return -1; |
9ae8ffe7 JL |
2606 | } |
2607 | ||
2608 | if (GET_CODE (x) == GET_CODE (y)) | |
2609 | switch (GET_CODE (x)) | |
2610 | { | |
2611 | case MULT: | |
2612 | { | |
2613 | /* Handle cases where we expect the second operands to be the | |
2614 | same, and check only whether the first operand would conflict | |
2615 | or not. */ | |
2616 | rtx x0, y0; | |
2617 | rtx x1 = canon_rtx (XEXP (x, 1)); | |
2618 | rtx y1 = canon_rtx (XEXP (y, 1)); | |
2619 | if (! rtx_equal_for_memref_p (x1, y1)) | |
f47e08d9 | 2620 | return -1; |
9ae8ffe7 JL |
2621 | x0 = canon_rtx (XEXP (x, 0)); |
2622 | y0 = canon_rtx (XEXP (y, 0)); | |
2623 | if (rtx_equal_for_memref_p (x0, y0)) | |
3aa03517 | 2624 | return offset_overlap_p (c, xsize, ysize); |
9ae8ffe7 JL |
2625 | |
2626 | /* Can't properly adjust our sizes. */ | |
5284e559 RS |
2627 | poly_int64 c1; |
2628 | if (!poly_int_rtx_p (x1, &c1) | |
2629 | || !can_div_trunc_p (xsize, c1, &xsize) | |
2630 | || !can_div_trunc_p (ysize, c1, &ysize) | |
2631 | || !can_div_trunc_p (c, c1, &c)) | |
f47e08d9 | 2632 | return -1; |
9ae8ffe7 JL |
2633 | return memrefs_conflict_p (xsize, x0, ysize, y0, c); |
2634 | } | |
1d300e19 KG |
2635 | |
2636 | default: | |
2637 | break; | |
9ae8ffe7 JL |
2638 | } |
2639 | ||
a9bf4fe2 AO |
2640 | /* Deal with alignment ANDs by adjusting offset and size so as to |
2641 | cover the maximum range, without taking any previously known | |
5147bf6a AO |
2642 | alignment into account. Make a size negative after such an |
2643 | adjustments, so that, if we end up with e.g. two SYMBOL_REFs, we | |
2644 | assume a potential overlap, because they may end up in contiguous | |
2645 | memory locations and the stricter-alignment access may span over | |
2646 | part of both. */ | |
481683e1 | 2647 | if (GET_CODE (x) == AND && CONST_INT_P (XEXP (x, 1))) |
56ee9281 | 2648 | { |
a9bf4fe2 AO |
2649 | HOST_WIDE_INT sc = INTVAL (XEXP (x, 1)); |
2650 | unsigned HOST_WIDE_INT uc = sc; | |
146ec50f | 2651 | if (sc < 0 && pow2_or_zerop (-uc)) |
a9bf4fe2 | 2652 | { |
9f61be58 | 2653 | if (maybe_gt (xsize, 0)) |
5147bf6a | 2654 | xsize = -xsize; |
9f61be58 | 2655 | if (maybe_ne (xsize, 0)) |
3aa03517 | 2656 | xsize += sc + 1; |
fe8fb1c4 | 2657 | c -= sc + 1; |
a9bf4fe2 AO |
2658 | return memrefs_conflict_p (xsize, canon_rtx (XEXP (x, 0)), |
2659 | ysize, y, c); | |
2660 | } | |
56ee9281 | 2661 | } |
481683e1 | 2662 | if (GET_CODE (y) == AND && CONST_INT_P (XEXP (y, 1))) |
c02f035f | 2663 | { |
a9bf4fe2 AO |
2664 | HOST_WIDE_INT sc = INTVAL (XEXP (y, 1)); |
2665 | unsigned HOST_WIDE_INT uc = sc; | |
146ec50f | 2666 | if (sc < 0 && pow2_or_zerop (-uc)) |
a9bf4fe2 | 2667 | { |
9f61be58 | 2668 | if (maybe_gt (ysize, 0)) |
5147bf6a | 2669 | ysize = -ysize; |
9f61be58 | 2670 | if (maybe_ne (ysize, 0)) |
3aa03517 | 2671 | ysize += sc + 1; |
fe8fb1c4 | 2672 | c += sc + 1; |
a9bf4fe2 AO |
2673 | return memrefs_conflict_p (xsize, x, |
2674 | ysize, canon_rtx (XEXP (y, 0)), c); | |
2675 | } | |
c02f035f | 2676 | } |
9ae8ffe7 JL |
2677 | |
2678 | if (CONSTANT_P (x)) | |
2679 | { | |
9f61be58 RS |
2680 | poly_int64 cx, cy; |
2681 | if (poly_int_rtx_p (x, &cx) && poly_int_rtx_p (y, &cy)) | |
9ae8ffe7 | 2682 | { |
9f61be58 | 2683 | c += cy - cx; |
3aa03517 | 2684 | return offset_overlap_p (c, xsize, ysize); |
9ae8ffe7 JL |
2685 | } |
2686 | ||
2687 | if (GET_CODE (x) == CONST) | |
2688 | { | |
2689 | if (GET_CODE (y) == CONST) | |
2690 | return memrefs_conflict_p (xsize, canon_rtx (XEXP (x, 0)), | |
2691 | ysize, canon_rtx (XEXP (y, 0)), c); | |
2692 | else | |
2693 | return memrefs_conflict_p (xsize, canon_rtx (XEXP (x, 0)), | |
2694 | ysize, y, c); | |
2695 | } | |
2696 | if (GET_CODE (y) == CONST) | |
2697 | return memrefs_conflict_p (xsize, x, ysize, | |
2698 | canon_rtx (XEXP (y, 0)), c); | |
2699 | ||
3aa03517 AO |
2700 | /* Assume a potential overlap for symbolic addresses that went |
2701 | through alignment adjustments (i.e., that have negative | |
2702 | sizes), because we can't know how far they are from each | |
2703 | other. */ | |
9ae8ffe7 | 2704 | if (CONSTANT_P (y)) |
9f61be58 RS |
2705 | return (maybe_lt (xsize, 0) |
2706 | || maybe_lt (ysize, 0) | |
2707 | || offset_overlap_p (c, xsize, ysize)); | |
9ae8ffe7 | 2708 | |
f47e08d9 | 2709 | return -1; |
9ae8ffe7 | 2710 | } |
f47e08d9 RG |
2711 | |
2712 | return -1; | |
9ae8ffe7 JL |
2713 | } |
2714 | ||
2715 | /* Functions to compute memory dependencies. | |
2716 | ||
2717 | Since we process the insns in execution order, we can build tables | |
2718 | to keep track of what registers are fixed (and not aliased), what registers | |
2719 | are varying in known ways, and what registers are varying in unknown | |
2720 | ways. | |
2721 | ||
2722 | If both memory references are volatile, then there must always be a | |
67914693 | 2723 | dependence between the two references, since their order cannot be |
9ae8ffe7 | 2724 | changed. A volatile and non-volatile reference can be interchanged |
ca7fd9cd | 2725 | though. |
9ae8ffe7 | 2726 | |
53d9622b RS |
2727 | We also must allow AND addresses, because they may generate accesses |
2728 | outside the object being referenced. This is used to generate aligned | |
2729 | addresses from unaligned addresses, for instance, the alpha | |
dc1618bc | 2730 | storeqi_unaligned pattern. */ |
9ae8ffe7 JL |
2731 | |
2732 | /* Read dependence: X is read after read in MEM takes place. There can | |
96672a3e RH |
2733 | only be a dependence here if both reads are volatile, or if either is |
2734 | an explicit barrier. */ | |
9ae8ffe7 JL |
2735 | |
2736 | int | |
4f588890 | 2737 | read_dependence (const_rtx mem, const_rtx x) |
9ae8ffe7 | 2738 | { |
96672a3e RH |
2739 | if (MEM_VOLATILE_P (x) && MEM_VOLATILE_P (mem)) |
2740 | return true; | |
2741 | if (MEM_ALIAS_SET (x) == ALIAS_SET_MEMORY_BARRIER | |
2742 | || MEM_ALIAS_SET (mem) == ALIAS_SET_MEMORY_BARRIER) | |
2743 | return true; | |
2744 | return false; | |
9ae8ffe7 JL |
2745 | } |
2746 | ||
998d7deb RH |
2747 | /* Look at the bottom of the COMPONENT_REF list for a DECL, and return it. */ |
2748 | ||
2749 | static tree | |
4682ae04 | 2750 | decl_for_component_ref (tree x) |
998d7deb RH |
2751 | { |
2752 | do | |
2753 | { | |
2754 | x = TREE_OPERAND (x, 0); | |
2755 | } | |
2756 | while (x && TREE_CODE (x) == COMPONENT_REF); | |
2757 | ||
2758 | return x && DECL_P (x) ? x : NULL_TREE; | |
2759 | } | |
2760 | ||
527210c4 RS |
2761 | /* Walk up the COMPONENT_REF list in X and adjust *OFFSET to compensate |
2762 | for the offset of the field reference. *KNOWN_P says whether the | |
2763 | offset is known. */ | |
998d7deb | 2764 | |
527210c4 RS |
2765 | static void |
2766 | adjust_offset_for_component_ref (tree x, bool *known_p, | |
d05d7551 | 2767 | poly_int64 *offset) |
998d7deb | 2768 | { |
527210c4 RS |
2769 | if (!*known_p) |
2770 | return; | |
ca7fd9cd | 2771 | do |
998d7deb | 2772 | { |
527210c4 | 2773 | tree xoffset = component_ref_field_offset (x); |
998d7deb | 2774 | tree field = TREE_OPERAND (x, 1); |
6e246559 | 2775 | if (!poly_int_tree_p (xoffset)) |
807e902e KZ |
2776 | { |
2777 | *known_p = false; | |
2778 | return; | |
2779 | } | |
998d7deb | 2780 | |
6e246559 RS |
2781 | poly_offset_int woffset |
2782 | = (wi::to_poly_offset (xoffset) | |
8de73453 | 2783 | + (wi::to_offset (DECL_FIELD_BIT_OFFSET (field)) |
6e246559 RS |
2784 | >> LOG2_BITS_PER_UNIT) |
2785 | + *offset); | |
2786 | if (!woffset.to_shwi (offset)) | |
527210c4 RS |
2787 | { |
2788 | *known_p = false; | |
2789 | return; | |
2790 | } | |
998d7deb RH |
2791 | |
2792 | x = TREE_OPERAND (x, 0); | |
2793 | } | |
2794 | while (x && TREE_CODE (x) == COMPONENT_REF); | |
998d7deb RH |
2795 | } |
2796 | ||
95bd1dd7 | 2797 | /* Return nonzero if we can determine the exprs corresponding to memrefs |
c6ea834c BM |
2798 | X and Y and they do not overlap. |
2799 | If LOOP_VARIANT is set, skip offset-based disambiguation */ | |
a4311dfe | 2800 | |
2e4e39f6 | 2801 | int |
c6ea834c | 2802 | nonoverlapping_memrefs_p (const_rtx x, const_rtx y, bool loop_invariant) |
a4311dfe | 2803 | { |
998d7deb | 2804 | tree exprx = MEM_EXPR (x), expry = MEM_EXPR (y); |
a4311dfe RK |
2805 | rtx rtlx, rtly; |
2806 | rtx basex, basey; | |
527210c4 | 2807 | bool moffsetx_known_p, moffsety_known_p; |
d05d7551 RS |
2808 | poly_int64 moffsetx = 0, moffsety = 0; |
2809 | poly_int64 offsetx = 0, offsety = 0, sizex, sizey; | |
a4311dfe | 2810 | |
998d7deb RH |
2811 | /* Unless both have exprs, we can't tell anything. */ |
2812 | if (exprx == 0 || expry == 0) | |
2813 | return 0; | |
2b22e382 RG |
2814 | |
2815 | /* For spill-slot accesses make sure we have valid offsets. */ | |
2816 | if ((exprx == get_spill_slot_decl (false) | |
527210c4 | 2817 | && ! MEM_OFFSET_KNOWN_P (x)) |
2b22e382 | 2818 | || (expry == get_spill_slot_decl (false) |
527210c4 | 2819 | && ! MEM_OFFSET_KNOWN_P (y))) |
2b22e382 | 2820 | return 0; |
c22cacf3 | 2821 | |
998d7deb | 2822 | /* If the field reference test failed, look at the DECLs involved. */ |
527210c4 RS |
2823 | moffsetx_known_p = MEM_OFFSET_KNOWN_P (x); |
2824 | if (moffsetx_known_p) | |
2825 | moffsetx = MEM_OFFSET (x); | |
998d7deb RH |
2826 | if (TREE_CODE (exprx) == COMPONENT_REF) |
2827 | { | |
2e0c984c RG |
2828 | tree t = decl_for_component_ref (exprx); |
2829 | if (! t) | |
2830 | return 0; | |
527210c4 | 2831 | adjust_offset_for_component_ref (exprx, &moffsetx_known_p, &moffsetx); |
2e0c984c | 2832 | exprx = t; |
998d7deb | 2833 | } |
c67a1cf6 | 2834 | |
527210c4 RS |
2835 | moffsety_known_p = MEM_OFFSET_KNOWN_P (y); |
2836 | if (moffsety_known_p) | |
2837 | moffsety = MEM_OFFSET (y); | |
998d7deb RH |
2838 | if (TREE_CODE (expry) == COMPONENT_REF) |
2839 | { | |
2e0c984c RG |
2840 | tree t = decl_for_component_ref (expry); |
2841 | if (! t) | |
2842 | return 0; | |
527210c4 | 2843 | adjust_offset_for_component_ref (expry, &moffsety_known_p, &moffsety); |
2e0c984c | 2844 | expry = t; |
998d7deb RH |
2845 | } |
2846 | ||
2847 | if (! DECL_P (exprx) || ! DECL_P (expry)) | |
a4311dfe RK |
2848 | return 0; |
2849 | ||
1f9ceff1 AO |
2850 | /* If we refer to different gimple registers, or one gimple register |
2851 | and one non-gimple-register, we know they can't overlap. First, | |
2852 | gimple registers don't have their addresses taken. Now, there | |
2853 | could be more than one stack slot for (different versions of) the | |
2854 | same gimple register, but we can presumably tell they don't | |
2855 | overlap based on offsets from stack base addresses elsewhere. | |
2856 | It's important that we don't proceed to DECL_RTL, because gimple | |
2857 | registers may not pass DECL_RTL_SET_P, and make_decl_rtl won't be | |
2858 | able to do anything about them since no SSA information will have | |
2859 | remained to guide it. */ | |
2860 | if (is_gimple_reg (exprx) || is_gimple_reg (expry)) | |
2d88904a AO |
2861 | return exprx != expry |
2862 | || (moffsetx_known_p && moffsety_known_p | |
2863 | && MEM_SIZE_KNOWN_P (x) && MEM_SIZE_KNOWN_P (y) | |
2864 | && !offset_overlap_p (moffsety - moffsetx, | |
2865 | MEM_SIZE (x), MEM_SIZE (y))); | |
1f9ceff1 | 2866 | |
1307c758 RG |
2867 | /* With invalid code we can end up storing into the constant pool. |
2868 | Bail out to avoid ICEing when creating RTL for this. | |
2869 | See gfortran.dg/lto/20091028-2_0.f90. */ | |
2870 | if (TREE_CODE (exprx) == CONST_DECL | |
2871 | || TREE_CODE (expry) == CONST_DECL) | |
2872 | return 1; | |
2873 | ||
dca16798 JJ |
2874 | /* If one decl is known to be a function or label in a function and |
2875 | the other is some kind of data, they can't overlap. */ | |
2876 | if ((TREE_CODE (exprx) == FUNCTION_DECL | |
2877 | || TREE_CODE (exprx) == LABEL_DECL) | |
2878 | != (TREE_CODE (expry) == FUNCTION_DECL | |
2879 | || TREE_CODE (expry) == LABEL_DECL)) | |
2880 | return 1; | |
2881 | ||
5f4cebba JJ |
2882 | /* If either of the decls doesn't have DECL_RTL set (e.g. marked as |
2883 | living in multiple places), we can't tell anything. Exception | |
2884 | are FUNCTION_DECLs for which we can create DECL_RTL on demand. */ | |
2885 | if ((!DECL_RTL_SET_P (exprx) && TREE_CODE (exprx) != FUNCTION_DECL) | |
2886 | || (!DECL_RTL_SET_P (expry) && TREE_CODE (expry) != FUNCTION_DECL)) | |
2887 | return 0; | |
2888 | ||
998d7deb RH |
2889 | rtlx = DECL_RTL (exprx); |
2890 | rtly = DECL_RTL (expry); | |
a4311dfe | 2891 | |
1edcd60b RK |
2892 | /* If either RTL is not a MEM, it must be a REG or CONCAT, meaning they |
2893 | can't overlap unless they are the same because we never reuse that part | |
2894 | of the stack frame used for locals for spilled pseudos. */ | |
3c0cb5de | 2895 | if ((!MEM_P (rtlx) || !MEM_P (rtly)) |
1edcd60b | 2896 | && ! rtx_equal_p (rtlx, rtly)) |
a4311dfe RK |
2897 | return 1; |
2898 | ||
5932a4d4 | 2899 | /* If we have MEMs referring to different address spaces (which can |
09e881c9 BE |
2900 | potentially overlap), we cannot easily tell from the addresses |
2901 | whether the references overlap. */ | |
2902 | if (MEM_P (rtlx) && MEM_P (rtly) | |
2903 | && MEM_ADDR_SPACE (rtlx) != MEM_ADDR_SPACE (rtly)) | |
2904 | return 0; | |
2905 | ||
a4311dfe RK |
2906 | /* Get the base and offsets of both decls. If either is a register, we |
2907 | know both are and are the same, so use that as the base. The only | |
2908 | we can avoid overlap is if we can deduce that they are nonoverlapping | |
2909 | pieces of that decl, which is very rare. */ | |
3c0cb5de | 2910 | basex = MEM_P (rtlx) ? XEXP (rtlx, 0) : rtlx; |
d05d7551 | 2911 | basex = strip_offset_and_add (basex, &offsetx); |
a4311dfe | 2912 | |
3c0cb5de | 2913 | basey = MEM_P (rtly) ? XEXP (rtly, 0) : rtly; |
d05d7551 | 2914 | basey = strip_offset_and_add (basey, &offsety); |
a4311dfe | 2915 | |
d746694a | 2916 | /* If the bases are different, we know they do not overlap if both |
ca7fd9cd | 2917 | are constants or if one is a constant and the other a pointer into the |
d746694a RK |
2918 | stack frame. Otherwise a different base means we can't tell if they |
2919 | overlap or not. */ | |
54363f8a | 2920 | if (compare_base_decls (exprx, expry) == 0) |
ca7fd9cd KH |
2921 | return ((CONSTANT_P (basex) && CONSTANT_P (basey)) |
2922 | || (CONSTANT_P (basex) && REG_P (basey) | |
2923 | && REGNO_PTR_FRAME_P (REGNO (basey))) | |
2924 | || (CONSTANT_P (basey) && REG_P (basex) | |
2925 | && REGNO_PTR_FRAME_P (REGNO (basex)))); | |
a4311dfe | 2926 | |
c6ea834c BM |
2927 | /* Offset based disambiguation not appropriate for loop invariant */ |
2928 | if (loop_invariant) | |
dca16798 | 2929 | return 0; |
c6ea834c | 2930 | |
54363f8a JH |
2931 | /* Offset based disambiguation is OK even if we do not know that the |
2932 | declarations are necessarily different | |
2933 | (i.e. compare_base_decls (exprx, expry) == -1) */ | |
2934 | ||
d05d7551 | 2935 | sizex = (!MEM_P (rtlx) ? poly_int64 (GET_MODE_SIZE (GET_MODE (rtlx))) |
f5541398 | 2936 | : MEM_SIZE_KNOWN_P (rtlx) ? MEM_SIZE (rtlx) |
a4311dfe | 2937 | : -1); |
d05d7551 | 2938 | sizey = (!MEM_P (rtly) ? poly_int64 (GET_MODE_SIZE (GET_MODE (rtly))) |
f5541398 RS |
2939 | : MEM_SIZE_KNOWN_P (rtly) ? MEM_SIZE (rtly) |
2940 | : -1); | |
a4311dfe | 2941 | |
0af5bc3e RK |
2942 | /* If we have an offset for either memref, it can update the values computed |
2943 | above. */ | |
527210c4 RS |
2944 | if (moffsetx_known_p) |
2945 | offsetx += moffsetx, sizex -= moffsetx; | |
2946 | if (moffsety_known_p) | |
2947 | offsety += moffsety, sizey -= moffsety; | |
a4311dfe | 2948 | |
0af5bc3e | 2949 | /* If a memref has both a size and an offset, we can use the smaller size. |
efc981bb | 2950 | We can't do this if the offset isn't known because we must view this |
0af5bc3e | 2951 | memref as being anywhere inside the DECL's MEM. */ |
527210c4 | 2952 | if (MEM_SIZE_KNOWN_P (x) && moffsetx_known_p) |
f5541398 | 2953 | sizex = MEM_SIZE (x); |
527210c4 | 2954 | if (MEM_SIZE_KNOWN_P (y) && moffsety_known_p) |
f5541398 | 2955 | sizey = MEM_SIZE (y); |
a4311dfe | 2956 | |
d05d7551 | 2957 | return !ranges_maybe_overlap_p (offsetx, sizex, offsety, sizey); |
a4311dfe RK |
2958 | } |
2959 | ||
9362286d SB |
2960 | /* Helper for true_dependence and canon_true_dependence. |
2961 | Checks for true dependence: X is read after store in MEM takes place. | |
9ae8ffe7 | 2962 | |
9362286d SB |
2963 | If MEM_CANONICALIZED is FALSE, then X_ADDR and MEM_ADDR should be |
2964 | NULL_RTX, and the canonical addresses of MEM and X are both computed | |
2965 | here. If MEM_CANONICALIZED, then MEM must be already canonicalized. | |
2966 | ||
2967 | If X_ADDR is non-NULL, it is used in preference of XEXP (x, 0). | |
2968 | ||
2969 | Returns 1 if there is a true dependence, 0 otherwise. */ | |
2970 | ||
2971 | static int | |
ef4bddc2 | 2972 | true_dependence_1 (const_rtx mem, machine_mode mem_mode, rtx mem_addr, |
53d9622b | 2973 | const_rtx x, rtx x_addr, bool mem_canonicalized) |
9ae8ffe7 | 2974 | { |
0777fc02 | 2975 | rtx true_mem_addr; |
49982682 | 2976 | rtx base; |
f47e08d9 | 2977 | int ret; |
9ae8ffe7 | 2978 | |
9362286d SB |
2979 | gcc_checking_assert (mem_canonicalized ? (mem_addr != NULL_RTX) |
2980 | : (mem_addr == NULL_RTX && x_addr == NULL_RTX)); | |
2981 | ||
9ae8ffe7 JL |
2982 | if (MEM_VOLATILE_P (x) && MEM_VOLATILE_P (mem)) |
2983 | return 1; | |
2984 | ||
c4484b8f | 2985 | /* (mem:BLK (scratch)) is a special mechanism to conflict with everything. |
ac3768f6 | 2986 | This is used in epilogue deallocation functions, and in cselib. */ |
c4484b8f RH |
2987 | if (GET_MODE (x) == BLKmode && GET_CODE (XEXP (x, 0)) == SCRATCH) |
2988 | return 1; | |
2989 | if (GET_MODE (mem) == BLKmode && GET_CODE (XEXP (mem, 0)) == SCRATCH) | |
2990 | return 1; | |
9cd9e512 RH |
2991 | if (MEM_ALIAS_SET (x) == ALIAS_SET_MEMORY_BARRIER |
2992 | || MEM_ALIAS_SET (mem) == ALIAS_SET_MEMORY_BARRIER) | |
2993 | return 1; | |
c4484b8f | 2994 | |
0777fc02 UB |
2995 | if (! x_addr) |
2996 | x_addr = XEXP (x, 0); | |
2997 | x_addr = get_addr (x_addr); | |
2998 | ||
9362286d SB |
2999 | if (! mem_addr) |
3000 | { | |
3001 | mem_addr = XEXP (mem, 0); | |
3002 | if (mem_mode == VOIDmode) | |
3003 | mem_mode = GET_MODE (mem); | |
3004 | } | |
0777fc02 | 3005 | true_mem_addr = get_addr (mem_addr); |
eab5c70a | 3006 | |
878f5596 UB |
3007 | /* Read-only memory is by definition never modified, and therefore can't |
3008 | conflict with anything. However, don't assume anything when AND | |
3009 | addresses are involved and leave to the code below to determine | |
3010 | dependence. We don't expect to find read-only set on MEM, but | |
3011 | stupid user tricks can produce them, so don't die. */ | |
3012 | if (MEM_READONLY_P (x) | |
3013 | && GET_CODE (x_addr) != AND | |
0777fc02 | 3014 | && GET_CODE (true_mem_addr) != AND) |
878f5596 UB |
3015 | return 0; |
3016 | ||
3017 | /* If we have MEMs referring to different address spaces (which can | |
3018 | potentially overlap), we cannot easily tell from the addresses | |
3019 | whether the references overlap. */ | |
3020 | if (MEM_ADDR_SPACE (mem) != MEM_ADDR_SPACE (x)) | |
3021 | return 1; | |
3022 | ||
55efb413 JW |
3023 | base = find_base_term (x_addr); |
3024 | if (base && (GET_CODE (base) == LABEL_REF | |
3025 | || (GET_CODE (base) == SYMBOL_REF | |
3026 | && CONSTANT_POOL_ADDRESS_P (base)))) | |
3027 | return 0; | |
3028 | ||
0777fc02 UB |
3029 | rtx mem_base = find_base_term (true_mem_addr); |
3030 | if (! base_alias_check (x_addr, base, true_mem_addr, mem_base, | |
31b0a960 | 3031 | GET_MODE (x), mem_mode)) |
1c72c7f6 JC |
3032 | return 0; |
3033 | ||
eab5c70a | 3034 | x_addr = canon_rtx (x_addr); |
9362286d | 3035 | if (!mem_canonicalized) |
0777fc02 | 3036 | mem_addr = canon_rtx (true_mem_addr); |
6e73e666 | 3037 | |
f47e08d9 RG |
3038 | if ((ret = memrefs_conflict_p (GET_MODE_SIZE (mem_mode), mem_addr, |
3039 | SIZE_FOR_MODE (x), x_addr, 0)) != -1) | |
3040 | return ret; | |
3041 | ||
a95b3cc7 | 3042 | if (mems_in_disjoint_alias_sets_p (x, mem)) |
f47e08d9 RG |
3043 | return 0; |
3044 | ||
c6ea834c | 3045 | if (nonoverlapping_memrefs_p (mem, x, false)) |
0211b6ab | 3046 | return 0; |
175a7536 | 3047 | |
55b34b5f | 3048 | return rtx_refs_may_alias_p (x, mem, true); |
a13d4ebf AM |
3049 | } |
3050 | ||
9362286d SB |
3051 | /* True dependence: X is read after store in MEM takes place. */ |
3052 | ||
3053 | int | |
ef4bddc2 | 3054 | true_dependence (const_rtx mem, machine_mode mem_mode, const_rtx x) |
9362286d SB |
3055 | { |
3056 | return true_dependence_1 (mem, mem_mode, NULL_RTX, | |
53d9622b | 3057 | x, NULL_RTX, /*mem_canonicalized=*/false); |
9362286d SB |
3058 | } |
3059 | ||
a13d4ebf | 3060 | /* Canonical true dependence: X is read after store in MEM takes place. |
ca7fd9cd KH |
3061 | Variant of true_dependence which assumes MEM has already been |
3062 | canonicalized (hence we no longer do that here). | |
9362286d SB |
3063 | The mem_addr argument has been added, since true_dependence_1 computed |
3064 | this value prior to canonicalizing. */ | |
a13d4ebf AM |
3065 | |
3066 | int | |
ef4bddc2 | 3067 | canon_true_dependence (const_rtx mem, machine_mode mem_mode, rtx mem_addr, |
53d9622b | 3068 | const_rtx x, rtx x_addr) |
a13d4ebf | 3069 | { |
9362286d | 3070 | return true_dependence_1 (mem, mem_mode, mem_addr, |
53d9622b | 3071 | x, x_addr, /*mem_canonicalized=*/true); |
9ae8ffe7 JL |
3072 | } |
3073 | ||
da7d8304 | 3074 | /* Returns nonzero if a write to X might alias a previous read from |
393f9fed | 3075 | (or, if WRITEP is true, a write to) MEM. |
bd280792 JR |
3076 | If X_CANONCALIZED is true, then X_ADDR is the canonicalized address of X, |
3077 | and X_MODE the mode for that access. | |
3078 | If MEM_CANONICALIZED is true, MEM is canonicalized. */ | |
9ae8ffe7 | 3079 | |
2c72b78f | 3080 | static int |
bd280792 | 3081 | write_dependence_p (const_rtx mem, |
ef4bddc2 | 3082 | const_rtx x, machine_mode x_mode, rtx x_addr, |
bd280792 | 3083 | bool mem_canonicalized, bool x_canonicalized, bool writep) |
9ae8ffe7 | 3084 | { |
bd280792 | 3085 | rtx mem_addr; |
0777fc02 | 3086 | rtx true_mem_addr, true_x_addr; |
49982682 | 3087 | rtx base; |
f47e08d9 | 3088 | int ret; |
6e73e666 | 3089 | |
bd280792 | 3090 | gcc_checking_assert (x_canonicalized |
6f5799be JJ |
3091 | ? (x_addr != NULL_RTX |
3092 | && (x_mode != VOIDmode || GET_MODE (x) == VOIDmode)) | |
bd280792 | 3093 | : (x_addr == NULL_RTX && x_mode == VOIDmode)); |
393f9fed | 3094 | |
9ae8ffe7 JL |
3095 | if (MEM_VOLATILE_P (x) && MEM_VOLATILE_P (mem)) |
3096 | return 1; | |
3097 | ||
c4484b8f RH |
3098 | /* (mem:BLK (scratch)) is a special mechanism to conflict with everything. |
3099 | This is used in epilogue deallocation functions. */ | |
3100 | if (GET_MODE (x) == BLKmode && GET_CODE (XEXP (x, 0)) == SCRATCH) | |
3101 | return 1; | |
3102 | if (GET_MODE (mem) == BLKmode && GET_CODE (XEXP (mem, 0)) == SCRATCH) | |
3103 | return 1; | |
9cd9e512 RH |
3104 | if (MEM_ALIAS_SET (x) == ALIAS_SET_MEMORY_BARRIER |
3105 | || MEM_ALIAS_SET (mem) == ALIAS_SET_MEMORY_BARRIER) | |
3106 | return 1; | |
c4484b8f | 3107 | |
bd280792 | 3108 | if (!x_addr) |
0777fc02 UB |
3109 | x_addr = XEXP (x, 0); |
3110 | true_x_addr = get_addr (x_addr); | |
3111 | ||
3112 | mem_addr = XEXP (mem, 0); | |
3113 | true_mem_addr = get_addr (mem_addr); | |
55efb413 | 3114 | |
878f5596 UB |
3115 | /* A read from read-only memory can't conflict with read-write memory. |
3116 | Don't assume anything when AND addresses are involved and leave to | |
3117 | the code below to determine dependence. */ | |
3118 | if (!writep | |
3119 | && MEM_READONLY_P (mem) | |
0777fc02 UB |
3120 | && GET_CODE (true_x_addr) != AND |
3121 | && GET_CODE (true_mem_addr) != AND) | |
878f5596 UB |
3122 | return 0; |
3123 | ||
3124 | /* If we have MEMs referring to different address spaces (which can | |
3125 | potentially overlap), we cannot easily tell from the addresses | |
3126 | whether the references overlap. */ | |
3127 | if (MEM_ADDR_SPACE (mem) != MEM_ADDR_SPACE (x)) | |
3128 | return 1; | |
3129 | ||
0777fc02 | 3130 | base = find_base_term (true_mem_addr); |
31b0a960 RB |
3131 | if (! writep |
3132 | && base | |
3133 | && (GET_CODE (base) == LABEL_REF | |
3134 | || (GET_CODE (base) == SYMBOL_REF | |
3135 | && CONSTANT_POOL_ADDRESS_P (base)))) | |
3136 | return 0; | |
49982682 | 3137 | |
0777fc02 UB |
3138 | rtx x_base = find_base_term (true_x_addr); |
3139 | if (! base_alias_check (true_x_addr, x_base, true_mem_addr, base, | |
3140 | GET_MODE (x), GET_MODE (mem))) | |
41472af8 MM |
3141 | return 0; |
3142 | ||
bd280792 | 3143 | if (!x_canonicalized) |
393f9fed | 3144 | { |
0777fc02 | 3145 | x_addr = canon_rtx (true_x_addr); |
bd280792 | 3146 | x_mode = GET_MODE (x); |
393f9fed | 3147 | } |
bd280792 | 3148 | if (!mem_canonicalized) |
0777fc02 | 3149 | mem_addr = canon_rtx (true_mem_addr); |
6e73e666 | 3150 | |
bd280792 JR |
3151 | if ((ret = memrefs_conflict_p (SIZE_FOR_MODE (mem), mem_addr, |
3152 | GET_MODE_SIZE (x_mode), x_addr, 0)) != -1) | |
f47e08d9 RG |
3153 | return ret; |
3154 | ||
c6ea834c | 3155 | if (nonoverlapping_memrefs_p (x, mem, false)) |
c6df88cb MM |
3156 | return 0; |
3157 | ||
55b34b5f | 3158 | return rtx_refs_may_alias_p (x, mem, false); |
c6df88cb MM |
3159 | } |
3160 | ||
3161 | /* Anti dependence: X is written after read in MEM takes place. */ | |
3162 | ||
3163 | int | |
4f588890 | 3164 | anti_dependence (const_rtx mem, const_rtx x) |
c6df88cb | 3165 | { |
bd280792 JR |
3166 | return write_dependence_p (mem, x, VOIDmode, NULL_RTX, |
3167 | /*mem_canonicalized=*/false, | |
3168 | /*x_canonicalized*/false, /*writep=*/false); | |
393f9fed JR |
3169 | } |
3170 | ||
bd280792 JR |
3171 | /* Likewise, but we already have a canonicalized MEM, and X_ADDR for X. |
3172 | Also, consider X in X_MODE (which might be from an enclosing | |
3173 | STRICT_LOW_PART / ZERO_EXTRACT). | |
3174 | If MEM_CANONICALIZED is true, MEM is canonicalized. */ | |
393f9fed JR |
3175 | |
3176 | int | |
bd280792 | 3177 | canon_anti_dependence (const_rtx mem, bool mem_canonicalized, |
ef4bddc2 | 3178 | const_rtx x, machine_mode x_mode, rtx x_addr) |
393f9fed | 3179 | { |
bd280792 JR |
3180 | return write_dependence_p (mem, x, x_mode, x_addr, |
3181 | mem_canonicalized, /*x_canonicalized=*/true, | |
3182 | /*writep=*/false); | |
9ae8ffe7 JL |
3183 | } |
3184 | ||
3185 | /* Output dependence: X is written after store in MEM takes place. */ | |
3186 | ||
3187 | int | |
4f588890 | 3188 | output_dependence (const_rtx mem, const_rtx x) |
9ae8ffe7 | 3189 | { |
bd280792 JR |
3190 | return write_dependence_p (mem, x, VOIDmode, NULL_RTX, |
3191 | /*mem_canonicalized=*/false, | |
3192 | /*x_canonicalized*/false, /*writep=*/true); | |
9ae8ffe7 | 3193 | } |
43b9f499 RB |
3194 | |
3195 | /* Likewise, but we already have a canonicalized MEM, and X_ADDR for X. | |
3196 | Also, consider X in X_MODE (which might be from an enclosing | |
3197 | STRICT_LOW_PART / ZERO_EXTRACT). | |
3198 | If MEM_CANONICALIZED is true, MEM is canonicalized. */ | |
3199 | ||
3200 | int | |
3201 | canon_output_dependence (const_rtx mem, bool mem_canonicalized, | |
3202 | const_rtx x, machine_mode x_mode, rtx x_addr) | |
3203 | { | |
3204 | return write_dependence_p (mem, x, x_mode, x_addr, | |
3205 | mem_canonicalized, /*x_canonicalized=*/true, | |
3206 | /*writep=*/true); | |
3207 | } | |
c14b9960 | 3208 | \f |
6e73e666 | 3209 | |
c6ea834c BM |
3210 | |
3211 | /* Check whether X may be aliased with MEM. Don't do offset-based | |
3212 | memory disambiguation & TBAA. */ | |
3213 | int | |
3214 | may_alias_p (const_rtx mem, const_rtx x) | |
3215 | { | |
3216 | rtx x_addr, mem_addr; | |
c6ea834c BM |
3217 | |
3218 | if (MEM_VOLATILE_P (x) && MEM_VOLATILE_P (mem)) | |
3219 | return 1; | |
3220 | ||
a95b3cc7 RG |
3221 | /* (mem:BLK (scratch)) is a special mechanism to conflict with everything. |
3222 | This is used in epilogue deallocation functions. */ | |
3223 | if (GET_MODE (x) == BLKmode && GET_CODE (XEXP (x, 0)) == SCRATCH) | |
3224 | return 1; | |
3225 | if (GET_MODE (mem) == BLKmode && GET_CODE (XEXP (mem, 0)) == SCRATCH) | |
c6ea834c | 3226 | return 1; |
c6ea834c BM |
3227 | if (MEM_ALIAS_SET (x) == ALIAS_SET_MEMORY_BARRIER |
3228 | || MEM_ALIAS_SET (mem) == ALIAS_SET_MEMORY_BARRIER) | |
3229 | return 1; | |
3230 | ||
c6ea834c | 3231 | x_addr = XEXP (x, 0); |
0777fc02 UB |
3232 | x_addr = get_addr (x_addr); |
3233 | ||
c6ea834c | 3234 | mem_addr = XEXP (mem, 0); |
0777fc02 | 3235 | mem_addr = get_addr (mem_addr); |
c6ea834c | 3236 | |
878f5596 UB |
3237 | /* Read-only memory is by definition never modified, and therefore can't |
3238 | conflict with anything. However, don't assume anything when AND | |
3239 | addresses are involved and leave to the code below to determine | |
3240 | dependence. We don't expect to find read-only set on MEM, but | |
3241 | stupid user tricks can produce them, so don't die. */ | |
3242 | if (MEM_READONLY_P (x) | |
3243 | && GET_CODE (x_addr) != AND | |
3244 | && GET_CODE (mem_addr) != AND) | |
3245 | return 0; | |
3246 | ||
3247 | /* If we have MEMs referring to different address spaces (which can | |
3248 | potentially overlap), we cannot easily tell from the addresses | |
3249 | whether the references overlap. */ | |
3250 | if (MEM_ADDR_SPACE (mem) != MEM_ADDR_SPACE (x)) | |
3251 | return 1; | |
3252 | ||
31b0a960 RB |
3253 | rtx x_base = find_base_term (x_addr); |
3254 | rtx mem_base = find_base_term (mem_addr); | |
3255 | if (! base_alias_check (x_addr, x_base, mem_addr, mem_base, | |
3256 | GET_MODE (x), GET_MODE (mem_addr))) | |
c6ea834c BM |
3257 | return 0; |
3258 | ||
c6ea834c BM |
3259 | if (nonoverlapping_memrefs_p (mem, x, true)) |
3260 | return 0; | |
3261 | ||
c6ea834c BM |
3262 | /* TBAA not valid for loop_invarint */ |
3263 | return rtx_refs_may_alias_p (x, mem, false); | |
3264 | } | |
3265 | ||
6e73e666 | 3266 | void |
b5deb7b6 | 3267 | init_alias_target (void) |
6e73e666 | 3268 | { |
b3694847 | 3269 | int i; |
6e73e666 | 3270 | |
9fc37b2b RS |
3271 | if (!arg_base_value) |
3272 | arg_base_value = gen_rtx_ADDRESS (VOIDmode, 0); | |
3273 | ||
b5deb7b6 SL |
3274 | memset (static_reg_base_value, 0, sizeof static_reg_base_value); |
3275 | ||
6e73e666 JC |
3276 | for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) |
3277 | /* Check whether this register can hold an incoming pointer | |
3278 | argument. FUNCTION_ARG_REGNO_P tests outgoing register | |
ec5c56db | 3279 | numbers, so translate if necessary due to register windows. */ |
6e73e666 | 3280 | if (FUNCTION_ARG_REGNO_P (OUTGOING_REGNO (i)) |
f939c3e6 | 3281 | && targetm.hard_regno_mode_ok (i, Pmode)) |
9fc37b2b RS |
3282 | static_reg_base_value[i] = arg_base_value; |
3283 | ||
e6eacdc9 RS |
3284 | /* RTL code is required to be consistent about whether it uses the |
3285 | stack pointer, the frame pointer or the argument pointer to | |
3286 | access a given area of the frame. We can therefore use the | |
3287 | base address to distinguish between the different areas. */ | |
757e8ba2 JJ |
3288 | static_reg_base_value[STACK_POINTER_REGNUM] |
3289 | = unique_base_value (UNIQUE_BASE_VALUE_SP); | |
3290 | static_reg_base_value[ARG_POINTER_REGNUM] | |
3291 | = unique_base_value (UNIQUE_BASE_VALUE_ARGP); | |
3292 | static_reg_base_value[FRAME_POINTER_REGNUM] | |
3293 | = unique_base_value (UNIQUE_BASE_VALUE_FP); | |
e6eacdc9 RS |
3294 | |
3295 | /* The above rules extend post-reload, with eliminations applying | |
3296 | consistently to each of the three pointers. Cope with cases in | |
3297 | which the frame pointer is eliminated to the hard frame pointer | |
3298 | rather than the stack pointer. */ | |
c3e08036 TS |
3299 | if (!HARD_FRAME_POINTER_IS_FRAME_POINTER) |
3300 | static_reg_base_value[HARD_FRAME_POINTER_REGNUM] | |
3301 | = unique_base_value (UNIQUE_BASE_VALUE_HFP); | |
bf1660a6 JL |
3302 | } |
3303 | ||
7b52eede JH |
3304 | /* Set MEMORY_MODIFIED when X modifies DATA (that is assumed |
3305 | to be memory reference. */ | |
3306 | static bool memory_modified; | |
3307 | static void | |
aa317c97 | 3308 | memory_modified_1 (rtx x, const_rtx pat ATTRIBUTE_UNUSED, void *data) |
7b52eede | 3309 | { |
3c0cb5de | 3310 | if (MEM_P (x)) |
7b52eede | 3311 | { |
9678086d | 3312 | if (anti_dependence (x, (const_rtx)data) || output_dependence (x, (const_rtx)data)) |
7b52eede JH |
3313 | memory_modified = true; |
3314 | } | |
3315 | } | |
3316 | ||
3317 | ||
3318 | /* Return true when INSN possibly modify memory contents of MEM | |
454ff5cb | 3319 | (i.e. address can be modified). */ |
7b52eede | 3320 | bool |
9678086d | 3321 | memory_modified_in_insn_p (const_rtx mem, const_rtx insn) |
7b52eede JH |
3322 | { |
3323 | if (!INSN_P (insn)) | |
3324 | return false; | |
bc36c711 JJ |
3325 | /* Conservatively assume all non-readonly MEMs might be modified in |
3326 | calls. */ | |
3327 | if (CALL_P (insn)) | |
3328 | return true; | |
7b52eede | 3329 | memory_modified = false; |
e8448ba5 RS |
3330 | note_stores (as_a<const rtx_insn *> (insn), memory_modified_1, |
3331 | CONST_CAST_RTX(mem)); | |
7b52eede JH |
3332 | return memory_modified; |
3333 | } | |
3334 | ||
c13e8210 MM |
3335 | /* Initialize the aliasing machinery. Initialize the REG_KNOWN_VALUE |
3336 | array. */ | |
3337 | ||
9ae8ffe7 | 3338 | void |
4682ae04 | 3339 | init_alias_analysis (void) |
9ae8ffe7 | 3340 | { |
c582d54a | 3341 | unsigned int maxreg = max_reg_num (); |
ea64ef27 | 3342 | int changed, pass; |
b3694847 SS |
3343 | int i; |
3344 | unsigned int ui; | |
d36a28b8 DM |
3345 | rtx_insn *insn; |
3346 | rtx val; | |
131db6b8 SB |
3347 | int rpo_cnt; |
3348 | int *rpo; | |
9ae8ffe7 | 3349 | |
0d446150 JH |
3350 | timevar_push (TV_ALIAS_ANALYSIS); |
3351 | ||
92390dd1 | 3352 | vec_safe_grow_cleared (reg_known_value, maxreg - FIRST_PSEUDO_REGISTER); |
9ff3c7ca | 3353 | reg_known_equiv_p = sbitmap_alloc (maxreg - FIRST_PSEUDO_REGISTER); |
dd3d1ec0 | 3354 | bitmap_clear (reg_known_equiv_p); |
9ae8ffe7 | 3355 | |
08c79682 | 3356 | /* If we have memory allocated from the previous run, use it. */ |
c582d54a | 3357 | if (old_reg_base_value) |
08c79682 KH |
3358 | reg_base_value = old_reg_base_value; |
3359 | ||
3360 | if (reg_base_value) | |
9771b263 | 3361 | reg_base_value->truncate (0); |
08c79682 | 3362 | |
9771b263 | 3363 | vec_safe_grow_cleared (reg_base_value, maxreg); |
ac606739 | 3364 | |
5ed6ace5 | 3365 | new_reg_base_value = XNEWVEC (rtx, maxreg); |
d630245f | 3366 | reg_seen = sbitmap_alloc (maxreg); |
ec907dd8 JL |
3367 | |
3368 | /* The basic idea is that each pass through this loop will use the | |
3369 | "constant" information from the previous pass to propagate alias | |
3370 | information through another level of assignments. | |
3371 | ||
131db6b8 SB |
3372 | The propagation is done on the CFG in reverse post-order, to propagate |
3373 | things forward as far as possible in each iteration. | |
3374 | ||
ec907dd8 JL |
3375 | This could get expensive if the assignment chains are long. Maybe |
3376 | we should throttle the number of iterations, possibly based on | |
6e73e666 | 3377 | the optimization level or flag_expensive_optimizations. |
ec907dd8 JL |
3378 | |
3379 | We could propagate more information in the first pass by making use | |
6fb5fa3c | 3380 | of DF_REG_DEF_COUNT to determine immediately that the alias information |
ea64ef27 JL |
3381 | for a pseudo is "constant". |
3382 | ||
3383 | A program with an uninitialized variable can cause an infinite loop | |
3384 | here. Instead of doing a full dataflow analysis to detect such problems | |
3385 | we just cap the number of iterations for the loop. | |
3386 | ||
3387 | The state of the arrays for the set chain in question does not matter | |
3388 | since the program has undefined behavior. */ | |
6e73e666 | 3389 | |
0cae8d31 | 3390 | rpo = XNEWVEC (int, n_basic_blocks_for_fn (cfun)); |
131db6b8 SB |
3391 | rpo_cnt = pre_and_rev_post_order_compute (NULL, rpo, false); |
3392 | ||
e86a9946 RS |
3393 | /* The prologue/epilogue insns are not threaded onto the |
3394 | insn chain until after reload has completed. Thus, | |
3395 | there is no sense wasting time checking if INSN is in | |
3396 | the prologue/epilogue until after reload has completed. */ | |
3397 | bool could_be_prologue_epilogue = ((targetm.have_prologue () | |
3398 | || targetm.have_epilogue ()) | |
3399 | && reload_completed); | |
3400 | ||
ea64ef27 | 3401 | pass = 0; |
6e73e666 | 3402 | do |
ec907dd8 JL |
3403 | { |
3404 | /* Assume nothing will change this iteration of the loop. */ | |
3405 | changed = 0; | |
3406 | ||
ec907dd8 | 3407 | /* We want to assign the same IDs each iteration of this loop, so |
9fc37b2b RS |
3408 | start counting from one each iteration of the loop. */ |
3409 | unique_id = 1; | |
ec907dd8 | 3410 | |
f5143c46 | 3411 | /* We're at the start of the function each iteration through the |
ec907dd8 | 3412 | loop, so we're copying arguments. */ |
83bbd9b6 | 3413 | copying_arguments = true; |
9ae8ffe7 | 3414 | |
6e73e666 | 3415 | /* Wipe the potential alias information clean for this pass. */ |
c582d54a | 3416 | memset (new_reg_base_value, 0, maxreg * sizeof (rtx)); |
8072f69c | 3417 | |
6e73e666 | 3418 | /* Wipe the reg_seen array clean. */ |
f61e445a | 3419 | bitmap_clear (reg_seen); |
9ae8ffe7 | 3420 | |
356610cb EB |
3421 | /* Initialize the alias information for this pass. */ |
3422 | for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) | |
e6eacdc9 RS |
3423 | if (static_reg_base_value[i] |
3424 | /* Don't treat the hard frame pointer as special if we | |
3425 | eliminated the frame pointer to the stack pointer instead. */ | |
3426 | && !(i == HARD_FRAME_POINTER_REGNUM | |
3427 | && reload_completed | |
3428 | && !frame_pointer_needed | |
3429 | && targetm.can_eliminate (FRAME_POINTER_REGNUM, | |
3430 | STACK_POINTER_REGNUM))) | |
356610cb EB |
3431 | { |
3432 | new_reg_base_value[i] = static_reg_base_value[i]; | |
3433 | bitmap_set_bit (reg_seen, i); | |
3434 | } | |
6e73e666 | 3435 | |
ec907dd8 | 3436 | /* Walk the insns adding values to the new_reg_base_value array. */ |
131db6b8 | 3437 | for (i = 0; i < rpo_cnt; i++) |
9ae8ffe7 | 3438 | { |
06e28de2 | 3439 | basic_block bb = BASIC_BLOCK_FOR_FN (cfun, rpo[i]); |
131db6b8 | 3440 | FOR_BB_INSNS (bb, insn) |
ec907dd8 | 3441 | { |
131db6b8 SB |
3442 | if (NONDEBUG_INSN_P (insn)) |
3443 | { | |
3444 | rtx note, set; | |
efc9bd41 | 3445 | |
e86a9946 | 3446 | if (could_be_prologue_epilogue |
131db6b8 SB |
3447 | && prologue_epilogue_contains (insn)) |
3448 | continue; | |
efc9bd41 | 3449 | |
131db6b8 SB |
3450 | /* If this insn has a noalias note, process it, Otherwise, |
3451 | scan for sets. A simple set will have no side effects | |
3452 | which could change the base value of any other register. */ | |
6e73e666 | 3453 | |
131db6b8 SB |
3454 | if (GET_CODE (PATTERN (insn)) == SET |
3455 | && REG_NOTES (insn) != 0 | |
3456 | && find_reg_note (insn, REG_NOALIAS, NULL_RTX)) | |
3457 | record_set (SET_DEST (PATTERN (insn)), NULL_RTX, NULL); | |
3458 | else | |
e8448ba5 | 3459 | note_stores (insn, record_set, NULL); |
6e73e666 | 3460 | |
131db6b8 | 3461 | set = single_set (insn); |
6e73e666 | 3462 | |
131db6b8 SB |
3463 | if (set != 0 |
3464 | && REG_P (SET_DEST (set)) | |
3465 | && REGNO (SET_DEST (set)) >= FIRST_PSEUDO_REGISTER) | |
713f41f9 | 3466 | { |
131db6b8 SB |
3467 | unsigned int regno = REGNO (SET_DEST (set)); |
3468 | rtx src = SET_SRC (set); | |
3469 | rtx t; | |
3470 | ||
3471 | note = find_reg_equal_equiv_note (insn); | |
3472 | if (note && REG_NOTE_KIND (note) == REG_EQUAL | |
3473 | && DF_REG_DEF_COUNT (regno) != 1) | |
3474 | note = NULL_RTX; | |
3475 | ||
5284e559 | 3476 | poly_int64 offset; |
131db6b8 SB |
3477 | if (note != NULL_RTX |
3478 | && GET_CODE (XEXP (note, 0)) != EXPR_LIST | |
3479 | && ! rtx_varies_p (XEXP (note, 0), 1) | |
3480 | && ! reg_overlap_mentioned_p (SET_DEST (set), | |
3481 | XEXP (note, 0))) | |
3482 | { | |
3483 | set_reg_known_value (regno, XEXP (note, 0)); | |
3484 | set_reg_known_equiv_p (regno, | |
3485 | REG_NOTE_KIND (note) == REG_EQUIV); | |
3486 | } | |
3487 | else if (DF_REG_DEF_COUNT (regno) == 1 | |
3488 | && GET_CODE (src) == PLUS | |
3489 | && REG_P (XEXP (src, 0)) | |
3490 | && (t = get_reg_known_value (REGNO (XEXP (src, 0)))) | |
5284e559 | 3491 | && poly_int_rtx_p (XEXP (src, 1), &offset)) |
131db6b8 | 3492 | { |
5284e559 | 3493 | t = plus_constant (GET_MODE (src), t, offset); |
131db6b8 SB |
3494 | set_reg_known_value (regno, t); |
3495 | set_reg_known_equiv_p (regno, false); | |
3496 | } | |
3497 | else if (DF_REG_DEF_COUNT (regno) == 1 | |
3498 | && ! rtx_varies_p (src, 1)) | |
3499 | { | |
3500 | set_reg_known_value (regno, src); | |
3501 | set_reg_known_equiv_p (regno, false); | |
3502 | } | |
713f41f9 | 3503 | } |
6e73e666 | 3504 | } |
131db6b8 SB |
3505 | else if (NOTE_P (insn) |
3506 | && NOTE_KIND (insn) == NOTE_INSN_FUNCTION_BEG) | |
3507 | copying_arguments = false; | |
ec907dd8 | 3508 | } |
6e73e666 | 3509 | } |
ec907dd8 | 3510 | |
6e73e666 | 3511 | /* Now propagate values from new_reg_base_value to reg_base_value. */ |
62e5bf5d | 3512 | gcc_assert (maxreg == (unsigned int) max_reg_num ()); |
c22cacf3 | 3513 | |
c582d54a | 3514 | for (ui = 0; ui < maxreg; ui++) |
6e73e666 | 3515 | { |
e51712db | 3516 | if (new_reg_base_value[ui] |
9771b263 DN |
3517 | && new_reg_base_value[ui] != (*reg_base_value)[ui] |
3518 | && ! rtx_equal_p (new_reg_base_value[ui], (*reg_base_value)[ui])) | |
ec907dd8 | 3519 | { |
9771b263 | 3520 | (*reg_base_value)[ui] = new_reg_base_value[ui]; |
6e73e666 | 3521 | changed = 1; |
ec907dd8 | 3522 | } |
9ae8ffe7 | 3523 | } |
9ae8ffe7 | 3524 | } |
6e73e666 | 3525 | while (changed && ++pass < MAX_ALIAS_LOOP_PASSES); |
131db6b8 | 3526 | XDELETEVEC (rpo); |
9ae8ffe7 JL |
3527 | |
3528 | /* Fill in the remaining entries. */ | |
9771b263 | 3529 | FOR_EACH_VEC_ELT (*reg_known_value, i, val) |
9ff3c7ca SB |
3530 | { |
3531 | int regno = i + FIRST_PSEUDO_REGISTER; | |
3532 | if (! val) | |
3533 | set_reg_known_value (regno, regno_reg_rtx[regno]); | |
3534 | } | |
9ae8ffe7 | 3535 | |
e05e2395 MM |
3536 | /* Clean up. */ |
3537 | free (new_reg_base_value); | |
ec907dd8 | 3538 | new_reg_base_value = 0; |
d630245f | 3539 | sbitmap_free (reg_seen); |
9ae8ffe7 | 3540 | reg_seen = 0; |
0d446150 | 3541 | timevar_pop (TV_ALIAS_ANALYSIS); |
9ae8ffe7 JL |
3542 | } |
3543 | ||
61630b27 JJ |
3544 | /* Equate REG_BASE_VALUE (reg1) to REG_BASE_VALUE (reg2). |
3545 | Special API for var-tracking pass purposes. */ | |
3546 | ||
3547 | void | |
3548 | vt_equate_reg_base_value (const_rtx reg1, const_rtx reg2) | |
3549 | { | |
9771b263 | 3550 | (*reg_base_value)[REGNO (reg1)] = REG_BASE_VALUE (reg2); |
61630b27 JJ |
3551 | } |
3552 | ||
9ae8ffe7 | 3553 | void |
4682ae04 | 3554 | end_alias_analysis (void) |
9ae8ffe7 | 3555 | { |
c582d54a | 3556 | old_reg_base_value = reg_base_value; |
9771b263 | 3557 | vec_free (reg_known_value); |
9ff3c7ca | 3558 | sbitmap_free (reg_known_equiv_p); |
9ae8ffe7 | 3559 | } |
e2500fed | 3560 | |
3ecf9d13 JH |
3561 | void |
3562 | dump_alias_stats_in_alias_c (FILE *s) | |
3563 | { | |
3564 | fprintf (s, " TBAA oracle: %llu disambiguations %llu queries\n" | |
3565 | " %llu are in alias set 0\n" | |
3566 | " %llu queries asked about the same object\n" | |
3567 | " %llu queries asked about the same alias set\n" | |
3568 | " %llu access volatile\n" | |
6e042ef4 JH |
3569 | " %llu are dependent in the DAG\n" |
3570 | " %llu are aritificially in conflict with void *\n", | |
3ecf9d13 JH |
3571 | alias_stats.num_disambiguated, |
3572 | alias_stats.num_alias_zero + alias_stats.num_same_alias_set | |
3573 | + alias_stats.num_same_objects + alias_stats.num_volatile | |
6e042ef4 JH |
3574 | + alias_stats.num_dag + alias_stats.num_disambiguated |
3575 | + alias_stats.num_universal, | |
3ecf9d13 | 3576 | alias_stats.num_alias_zero, alias_stats.num_same_alias_set, |
6e042ef4 JH |
3577 | alias_stats.num_same_objects, alias_stats.num_volatile, |
3578 | alias_stats.num_dag, alias_stats.num_universal); | |
3ecf9d13 | 3579 | } |
e2500fed | 3580 | #include "gt-alias.h" |