]>
Commit | Line | Data |
---|---|---|
9ae8ffe7 | 1 | /* Alias analysis for GNU C |
5624e564 | 2 | Copyright (C) 1997-2015 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 AM |
24 | #include "backend.h" |
25 | #include "tree.h" | |
26 | #include "gimple.h" | |
9ae8ffe7 | 27 | #include "rtl.h" |
c7131fb2 | 28 | #include "df.h" |
40e23961 | 29 | #include "alias.h" |
40e23961 | 30 | #include "fold-const.h" |
d8a2d370 | 31 | #include "varasm.h" |
36566b39 | 32 | #include "flags.h" |
36566b39 PK |
33 | #include "insn-config.h" |
34 | #include "expmed.h" | |
35 | #include "dojump.h" | |
36 | #include "explow.h" | |
37 | #include "calls.h" | |
78528714 | 38 | #include "emit-rtl.h" |
36566b39 PK |
39 | #include "stmt.h" |
40 | #include "expr.h" | |
41 | #include "tm_p.h" | |
9ae8ffe7 | 42 | #include "regs.h" |
718f9c0f | 43 | #include "diagnostic-core.h" |
a78a26f1 | 44 | #include "alloc-pool.h" |
eab5c70a | 45 | #include "cselib.h" |
d23c55c2 | 46 | #include "langhooks.h" |
0d446150 | 47 | #include "timevar.h" |
7ee2468b | 48 | #include "dumpfile.h" |
ab780373 | 49 | #include "target.h" |
60393bbc | 50 | #include "cfganal.h" |
2fb9a547 | 51 | #include "internal-fn.h" |
442b4905 | 52 | #include "gimple-ssa.h" |
403837b4 | 53 | #include "rtl-iter.h" |
ea900239 DB |
54 | |
55 | /* The aliasing API provided here solves related but different problems: | |
56 | ||
c22cacf3 | 57 | Say there exists (in c) |
ea900239 DB |
58 | |
59 | struct X { | |
60 | struct Y y1; | |
61 | struct Z z2; | |
62 | } x1, *px1, *px2; | |
63 | ||
64 | struct Y y2, *py; | |
65 | struct Z z2, *pz; | |
66 | ||
67 | ||
308a3fe2 | 68 | py = &x1.y1; |
ea900239 DB |
69 | px2 = &x1; |
70 | ||
71 | Consider the four questions: | |
72 | ||
73 | Can a store to x1 interfere with px2->y1? | |
74 | Can a store to x1 interfere with px2->z2? | |
ea900239 DB |
75 | Can a store to x1 change the value pointed to by with py? |
76 | Can a store to x1 change the value pointed to by with pz? | |
77 | ||
78 | The answer to these questions can be yes, yes, yes, and maybe. | |
79 | ||
80 | The first two questions can be answered with a simple examination | |
81 | of the type system. If structure X contains a field of type Y then | |
073a8998 | 82 | a store through a pointer to an X can overwrite any field that is |
ea900239 DB |
83 | contained (recursively) in an X (unless we know that px1 != px2). |
84 | ||
308a3fe2 DS |
85 | The last two questions can be solved in the same way as the first |
86 | two questions but this is too conservative. The observation is | |
87 | that in some cases we can know which (if any) fields are addressed | |
88 | and if those addresses are used in bad ways. This analysis may be | |
89 | language specific. In C, arbitrary operations may be applied to | |
90 | pointers. However, there is some indication that this may be too | |
91 | conservative for some C++ types. | |
ea900239 DB |
92 | |
93 | The pass ipa-type-escape does this analysis for the types whose | |
c22cacf3 | 94 | instances do not escape across the compilation boundary. |
ea900239 DB |
95 | |
96 | Historically in GCC, these two problems were combined and a single | |
308a3fe2 | 97 | data structure that was used to represent the solution to these |
ea900239 | 98 | problems. We now have two similar but different data structures, |
308a3fe2 DS |
99 | The data structure to solve the last two questions is similar to |
100 | the first, but does not contain the fields whose address are never | |
101 | taken. For types that do escape the compilation unit, the data | |
102 | structures will have identical information. | |
ea900239 | 103 | */ |
3932261a MM |
104 | |
105 | /* The alias sets assigned to MEMs assist the back-end in determining | |
106 | which MEMs can alias which other MEMs. In general, two MEMs in | |
ac3d9668 RK |
107 | different alias sets cannot alias each other, with one important |
108 | exception. Consider something like: | |
3932261a | 109 | |
01d28c3f | 110 | struct S { int i; double d; }; |
3932261a MM |
111 | |
112 | a store to an `S' can alias something of either type `int' or type | |
113 | `double'. (However, a store to an `int' cannot alias a `double' | |
114 | and vice versa.) We indicate this via a tree structure that looks | |
115 | like: | |
c22cacf3 MS |
116 | struct S |
117 | / \ | |
3932261a | 118 | / \ |
c22cacf3 MS |
119 | |/_ _\| |
120 | int double | |
3932261a | 121 | |
ac3d9668 RK |
122 | (The arrows are directed and point downwards.) |
123 | In this situation we say the alias set for `struct S' is the | |
124 | `superset' and that those for `int' and `double' are `subsets'. | |
125 | ||
3bdf5ad1 RK |
126 | To see whether two alias sets can point to the same memory, we must |
127 | see if either alias set is a subset of the other. We need not trace | |
95bd1dd7 | 128 | past immediate descendants, however, since we propagate all |
3bdf5ad1 | 129 | grandchildren up one level. |
3932261a MM |
130 | |
131 | Alias set zero is implicitly a superset of all other alias sets. | |
132 | However, this is no actual entry for alias set zero. It is an | |
133 | error to attempt to explicitly construct a subset of zero. */ | |
134 | ||
e0702244 | 135 | struct alias_set_hash : int_hash <int, INT_MIN, INT_MIN + 1> {}; |
de144fb2 | 136 | |
7e5487a2 | 137 | struct GTY(()) alias_set_entry_d { |
3932261a | 138 | /* The alias set number, as stored in MEM_ALIAS_SET. */ |
4862826d | 139 | alias_set_type alias_set; |
3932261a MM |
140 | |
141 | /* The children of the alias set. These are not just the immediate | |
95bd1dd7 | 142 | children, but, in fact, all descendants. So, if we have: |
3932261a | 143 | |
ca7fd9cd | 144 | struct T { struct S s; float f; } |
3932261a MM |
145 | |
146 | continuing our example above, the children here will be all of | |
147 | `int', `double', `float', and `struct S'. */ | |
fb5c464a | 148 | hash_map<alias_set_hash, int> *children; |
6e042ef4 JH |
149 | |
150 | /* Nonzero if would have a child of zero: this effectively makes this | |
151 | alias set the same as alias set zero. */ | |
152 | bool has_zero_child; | |
153 | /* Nonzero if alias set corresponds to pointer type itself (i.e. not to | |
154 | aggregate contaiing pointer. | |
155 | This is used for a special case where we need an universal pointer type | |
156 | compatible with all other pointer types. */ | |
157 | bool is_pointer; | |
158 | /* Nonzero if is_pointer or if one of childs have has_pointer set. */ | |
159 | bool has_pointer; | |
b604074c | 160 | }; |
7e5487a2 | 161 | typedef struct alias_set_entry_d *alias_set_entry; |
9ae8ffe7 | 162 | |
ed7a4b4b | 163 | static int rtx_equal_for_memref_p (const_rtx, const_rtx); |
4682ae04 | 164 | static int memrefs_conflict_p (int, rtx, int, rtx, HOST_WIDE_INT); |
7bc980e1 | 165 | static void record_set (rtx, const_rtx, void *); |
ef4bddc2 RS |
166 | static int base_alias_check (rtx, rtx, rtx, rtx, machine_mode, |
167 | machine_mode); | |
4682ae04 | 168 | static rtx find_base_value (rtx); |
4f588890 | 169 | static int mems_in_disjoint_alias_sets_p (const_rtx, const_rtx); |
4862826d | 170 | static alias_set_entry get_alias_set_entry (alias_set_type); |
4682ae04 | 171 | static tree decl_for_component_ref (tree); |
bd280792 | 172 | static int write_dependence_p (const_rtx, |
ef4bddc2 | 173 | const_rtx, machine_mode, rtx, |
bd280792 | 174 | bool, bool, bool); |
4682ae04 | 175 | |
aa317c97 | 176 | static void memory_modified_1 (rtx, const_rtx, void *); |
9ae8ffe7 | 177 | |
3ecf9d13 JH |
178 | /* Query statistics for the different low-level disambiguators. |
179 | A high-level query may trigger multiple of them. */ | |
180 | ||
181 | static struct { | |
182 | unsigned long long num_alias_zero; | |
183 | unsigned long long num_same_alias_set; | |
184 | unsigned long long num_same_objects; | |
185 | unsigned long long num_volatile; | |
186 | unsigned long long num_dag; | |
6e042ef4 | 187 | unsigned long long num_universal; |
3ecf9d13 JH |
188 | unsigned long long num_disambiguated; |
189 | } alias_stats; | |
190 | ||
191 | ||
9ae8ffe7 JL |
192 | /* Set up all info needed to perform alias analysis on memory references. */ |
193 | ||
d4b60170 | 194 | /* Returns the size in bytes of the mode of X. */ |
9ae8ffe7 JL |
195 | #define SIZE_FOR_MODE(X) (GET_MODE_SIZE (GET_MODE (X))) |
196 | ||
ea64ef27 | 197 | /* Cap the number of passes we make over the insns propagating alias |
131db6b8 SB |
198 | information through set chains. |
199 | ??? 10 is a completely arbitrary choice. This should be based on the | |
200 | maximum loop depth in the CFG, but we do not have this information | |
201 | available (even if current_loops _is_ available). */ | |
ea64ef27 | 202 | #define MAX_ALIAS_LOOP_PASSES 10 |
ca7fd9cd | 203 | |
9ae8ffe7 JL |
204 | /* reg_base_value[N] gives an address to which register N is related. |
205 | If all sets after the first add or subtract to the current value | |
206 | or otherwise modify it so it does not point to a different top level | |
207 | object, reg_base_value[N] is equal to the address part of the source | |
2a2c8203 JC |
208 | of the first set. |
209 | ||
210 | A base address can be an ADDRESS, SYMBOL_REF, or LABEL_REF. ADDRESS | |
9fc37b2b | 211 | expressions represent three types of base: |
b3b5ad95 | 212 | |
9fc37b2b RS |
213 | 1. incoming arguments. There is just one ADDRESS to represent all |
214 | arguments, since we do not know at this level whether accesses | |
215 | based on different arguments can alias. The ADDRESS has id 0. | |
b3b5ad95 | 216 | |
9fc37b2b RS |
217 | 2. stack_pointer_rtx, frame_pointer_rtx, hard_frame_pointer_rtx |
218 | (if distinct from frame_pointer_rtx) and arg_pointer_rtx. | |
219 | Each of these rtxes has a separate ADDRESS associated with it, | |
220 | each with a negative id. | |
221 | ||
222 | GCC is (and is required to be) precise in which register it | |
223 | chooses to access a particular region of stack. We can therefore | |
224 | assume that accesses based on one of these rtxes do not alias | |
225 | accesses based on another of these rtxes. | |
226 | ||
227 | 3. bases that are derived from malloc()ed memory (REG_NOALIAS). | |
228 | Each such piece of memory has a separate ADDRESS associated | |
229 | with it, each with an id greater than 0. | |
230 | ||
231 | Accesses based on one ADDRESS do not alias accesses based on other | |
232 | ADDRESSes. Accesses based on ADDRESSes in groups (2) and (3) do not | |
233 | alias globals either; the ADDRESSes have Pmode to indicate this. | |
234 | The ADDRESS in group (1) _may_ alias globals; it has VOIDmode to | |
235 | indicate this. */ | |
2a2c8203 | 236 | |
9771b263 | 237 | static GTY(()) vec<rtx, va_gc> *reg_base_value; |
ac606739 | 238 | static rtx *new_reg_base_value; |
c582d54a | 239 | |
9fc37b2b RS |
240 | /* The single VOIDmode ADDRESS that represents all argument bases. |
241 | It has id 0. */ | |
242 | static GTY(()) rtx arg_base_value; | |
243 | ||
244 | /* Used to allocate unique ids to each REG_NOALIAS ADDRESS. */ | |
245 | static int unique_id; | |
246 | ||
c582d54a JH |
247 | /* We preserve the copy of old array around to avoid amount of garbage |
248 | produced. About 8% of garbage produced were attributed to this | |
249 | array. */ | |
9771b263 | 250 | static GTY((deletable)) vec<rtx, va_gc> *old_reg_base_value; |
d4b60170 | 251 | |
9e412ca3 RS |
252 | /* Values of XINT (address, 0) of Pmode ADDRESS rtxes for special |
253 | registers. */ | |
254 | #define UNIQUE_BASE_VALUE_SP -1 | |
255 | #define UNIQUE_BASE_VALUE_ARGP -2 | |
256 | #define UNIQUE_BASE_VALUE_FP -3 | |
257 | #define UNIQUE_BASE_VALUE_HFP -4 | |
258 | ||
7bf84454 RS |
259 | #define static_reg_base_value \ |
260 | (this_target_rtl->x_static_reg_base_value) | |
bf1660a6 | 261 | |
9771b263 DN |
262 | #define REG_BASE_VALUE(X) \ |
263 | (REGNO (X) < vec_safe_length (reg_base_value) \ | |
264 | ? (*reg_base_value)[REGNO (X)] : 0) | |
9ae8ffe7 | 265 | |
c13e8210 | 266 | /* Vector indexed by N giving the initial (unchanging) value known for |
9ff3c7ca | 267 | pseudo-register N. This vector is initialized in init_alias_analysis, |
bb1acb3e | 268 | and does not change until end_alias_analysis is called. */ |
9771b263 | 269 | static GTY(()) vec<rtx, va_gc> *reg_known_value; |
9ae8ffe7 JL |
270 | |
271 | /* Vector recording for each reg_known_value whether it is due to a | |
272 | REG_EQUIV note. Future passes (viz., reload) may replace the | |
273 | pseudo with the equivalent expression and so we account for the | |
ac3d9668 RK |
274 | dependences that would be introduced if that happens. |
275 | ||
276 | The REG_EQUIV notes created in assign_parms may mention the arg | |
277 | pointer, and there are explicit insns in the RTL that modify the | |
278 | arg pointer. Thus we must ensure that such insns don't get | |
279 | scheduled across each other because that would invalidate the | |
280 | REG_EQUIV notes. One could argue that the REG_EQUIV notes are | |
281 | wrong, but solving the problem in the scheduler will likely give | |
282 | better code, so we do it here. */ | |
9ff3c7ca | 283 | static sbitmap reg_known_equiv_p; |
9ae8ffe7 | 284 | |
2a2c8203 JC |
285 | /* True when scanning insns from the start of the rtl to the |
286 | NOTE_INSN_FUNCTION_BEG note. */ | |
83bbd9b6 | 287 | static bool copying_arguments; |
9ae8ffe7 | 288 | |
1a5640b4 | 289 | |
3932261a | 290 | /* The splay-tree used to store the various alias set entries. */ |
9771b263 | 291 | static GTY (()) vec<alias_set_entry, va_gc> *alias_sets; |
ac3d9668 | 292 | \f |
55b34b5f RG |
293 | /* Build a decomposed reference object for querying the alias-oracle |
294 | from the MEM rtx and store it in *REF. | |
295 | Returns false if MEM is not suitable for the alias-oracle. */ | |
296 | ||
297 | static bool | |
298 | ao_ref_from_mem (ao_ref *ref, const_rtx mem) | |
299 | { | |
300 | tree expr = MEM_EXPR (mem); | |
301 | tree base; | |
302 | ||
303 | if (!expr) | |
304 | return false; | |
305 | ||
306 | ao_ref_init (ref, expr); | |
307 | ||
308 | /* Get the base of the reference and see if we have to reject or | |
309 | adjust it. */ | |
310 | base = ao_ref_base (ref); | |
311 | if (base == NULL_TREE) | |
312 | return false; | |
313 | ||
ef7a9fb8 RB |
314 | /* The tree oracle doesn't like bases that are neither decls |
315 | nor indirect references of SSA names. */ | |
316 | if (!(DECL_P (base) | |
317 | || (TREE_CODE (base) == MEM_REF | |
318 | && TREE_CODE (TREE_OPERAND (base, 0)) == SSA_NAME) | |
319 | || (TREE_CODE (base) == TARGET_MEM_REF | |
320 | && TREE_CODE (TMR_BASE (base)) == SSA_NAME))) | |
d15adbeb | 321 | return false; |
55b34b5f RG |
322 | |
323 | /* If this is a reference based on a partitioned decl replace the | |
ef7a9fb8 | 324 | base with a MEM_REF of the pointer representative we |
55b34b5f RG |
325 | created during stack slot partitioning. */ |
326 | if (TREE_CODE (base) == VAR_DECL | |
ef7a9fb8 | 327 | && ! is_global_var (base) |
55b34b5f RG |
328 | && cfun->gimple_df->decls_to_pointers != NULL) |
329 | { | |
39c8aaa4 | 330 | tree *namep = cfun->gimple_df->decls_to_pointers->get (base); |
55b34b5f | 331 | if (namep) |
39c8aaa4 | 332 | ref->base = build_simple_mem_ref (*namep); |
d15adbeb | 333 | } |
55b34b5f RG |
334 | |
335 | ref->ref_alias_set = MEM_ALIAS_SET (mem); | |
336 | ||
f68396a1 RG |
337 | /* If MEM_OFFSET or MEM_SIZE are unknown what we got from MEM_EXPR |
338 | is conservative, so trust it. */ | |
527210c4 | 339 | if (!MEM_OFFSET_KNOWN_P (mem) |
f5541398 | 340 | || !MEM_SIZE_KNOWN_P (mem)) |
f68396a1 | 341 | return true; |
366f945f | 342 | |
b0e96404 RG |
343 | /* If the base decl is a parameter we can have negative MEM_OFFSET in |
344 | case of promoted subregs on bigendian targets. Trust the MEM_EXPR | |
345 | here. */ | |
527210c4 RS |
346 | if (MEM_OFFSET (mem) < 0 |
347 | && (MEM_SIZE (mem) + MEM_OFFSET (mem)) * BITS_PER_UNIT == ref->size) | |
b0e96404 RG |
348 | return true; |
349 | ||
f68396a1 RG |
350 | /* Otherwise continue and refine size and offset we got from analyzing |
351 | MEM_EXPR by using MEM_SIZE and MEM_OFFSET. */ | |
352 | ||
527210c4 | 353 | ref->offset += MEM_OFFSET (mem) * BITS_PER_UNIT; |
f5541398 | 354 | ref->size = MEM_SIZE (mem) * BITS_PER_UNIT; |
b0e96404 RG |
355 | |
356 | /* The MEM may extend into adjacent fields, so adjust max_size if | |
357 | necessary. */ | |
358 | if (ref->max_size != -1 | |
359 | && ref->size > ref->max_size) | |
360 | ref->max_size = ref->size; | |
361 | ||
362 | /* If MEM_OFFSET and MEM_SIZE get us outside of the base object of | |
363 | the MEM_EXPR punt. This happens for STRICT_ALIGNMENT targets a lot. */ | |
364 | if (MEM_EXPR (mem) != get_spill_slot_decl (false) | |
365 | && (ref->offset < 0 | |
366 | || (DECL_P (ref->base) | |
807e902e KZ |
367 | && (DECL_SIZE (ref->base) == NULL_TREE |
368 | || TREE_CODE (DECL_SIZE (ref->base)) != INTEGER_CST | |
369 | || wi::ltu_p (wi::to_offset (DECL_SIZE (ref->base)), | |
370 | ref->offset + ref->size))))) | |
b0e96404 | 371 | return false; |
55b34b5f RG |
372 | |
373 | return true; | |
374 | } | |
375 | ||
376 | /* Query the alias-oracle on whether the two memory rtx X and MEM may | |
377 | alias. If TBAA_P is set also apply TBAA. Returns true if the | |
378 | two rtxen may alias, false otherwise. */ | |
379 | ||
380 | static bool | |
381 | rtx_refs_may_alias_p (const_rtx x, const_rtx mem, bool tbaa_p) | |
382 | { | |
383 | ao_ref ref1, ref2; | |
384 | ||
385 | if (!ao_ref_from_mem (&ref1, x) | |
386 | || !ao_ref_from_mem (&ref2, mem)) | |
387 | return true; | |
388 | ||
55e3bc4c RG |
389 | return refs_may_alias_p_1 (&ref1, &ref2, |
390 | tbaa_p | |
391 | && MEM_ALIAS_SET (x) != 0 | |
392 | && MEM_ALIAS_SET (mem) != 0); | |
55b34b5f RG |
393 | } |
394 | ||
3932261a MM |
395 | /* Returns a pointer to the alias set entry for ALIAS_SET, if there is |
396 | such an entry, or NULL otherwise. */ | |
397 | ||
9ddb66ca | 398 | static inline alias_set_entry |
4862826d | 399 | get_alias_set_entry (alias_set_type alias_set) |
3932261a | 400 | { |
9771b263 | 401 | return (*alias_sets)[alias_set]; |
3932261a MM |
402 | } |
403 | ||
ac3d9668 RK |
404 | /* Returns nonzero if the alias sets for MEM1 and MEM2 are such that |
405 | the two MEMs cannot alias each other. */ | |
3932261a | 406 | |
9ddb66ca | 407 | static inline int |
4f588890 | 408 | mems_in_disjoint_alias_sets_p (const_rtx mem1, const_rtx mem2) |
3932261a | 409 | { |
598f8eca RB |
410 | return (flag_strict_aliasing |
411 | && ! alias_sets_conflict_p (MEM_ALIAS_SET (mem1), | |
412 | MEM_ALIAS_SET (mem2))); | |
1da68f56 | 413 | } |
3932261a | 414 | |
c58936b6 DB |
415 | /* Return true if the first alias set is a subset of the second. */ |
416 | ||
417 | bool | |
4862826d | 418 | alias_set_subset_of (alias_set_type set1, alias_set_type set2) |
c58936b6 | 419 | { |
6e042ef4 | 420 | alias_set_entry ase2; |
c58936b6 DB |
421 | |
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 | { | |
456 | alias_set_entry ase1 = get_alias_set_entry (set1); | |
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 | { |
6e042ef4 JH |
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 | { |
3ecf9d13 JH |
554 | if (set1 == 0 || set2 == 0) |
555 | { | |
556 | ++alias_stats.num_alias_zero; | |
557 | return 1; | |
558 | } | |
559 | if (set1 == set2) | |
560 | { | |
561 | ++alias_stats.num_same_alias_set; | |
562 | return 1; | |
563 | } | |
5399d643 JW |
564 | |
565 | return 0; | |
566 | } | |
567 | ||
1da68f56 RK |
568 | /* Return 1 if any MEM object of type T1 will always conflict (using the |
569 | dependency routines in this file) with any MEM object of type T2. | |
570 | This is used when allocating temporary storage. If T1 and/or T2 are | |
571 | NULL_TREE, it means we know nothing about the storage. */ | |
572 | ||
573 | int | |
4682ae04 | 574 | objects_must_conflict_p (tree t1, tree t2) |
1da68f56 | 575 | { |
4862826d | 576 | alias_set_type set1, set2; |
82d610ec | 577 | |
e8ea2809 RK |
578 | /* If neither has a type specified, we don't know if they'll conflict |
579 | because we may be using them to store objects of various types, for | |
580 | example the argument and local variables areas of inlined functions. */ | |
981a4c34 | 581 | if (t1 == 0 && t2 == 0) |
e8ea2809 RK |
582 | return 0; |
583 | ||
1da68f56 | 584 | /* If they are the same type, they must conflict. */ |
3ecf9d13 JH |
585 | if (t1 == t2) |
586 | { | |
587 | ++alias_stats.num_same_objects; | |
588 | return 1; | |
589 | } | |
590 | /* Likewise if both are volatile. */ | |
591 | if (t1 != 0 && TYPE_VOLATILE (t1) && t2 != 0 && TYPE_VOLATILE (t2)) | |
592 | { | |
593 | ++alias_stats.num_volatile; | |
594 | return 1; | |
595 | } | |
1da68f56 | 596 | |
82d610ec RK |
597 | set1 = t1 ? get_alias_set (t1) : 0; |
598 | set2 = t2 ? get_alias_set (t2) : 0; | |
1da68f56 | 599 | |
836f7794 EB |
600 | /* We can't use alias_sets_conflict_p because we must make sure |
601 | that every subtype of t1 will conflict with every subtype of | |
82d610ec RK |
602 | t2 for which a pair of subobjects of these respective subtypes |
603 | overlaps on the stack. */ | |
836f7794 | 604 | return alias_sets_must_conflict_p (set1, set2); |
1da68f56 RK |
605 | } |
606 | \f | |
b4ada065 RB |
607 | /* Return the outermost parent of component present in the chain of |
608 | component references handled by get_inner_reference in T with the | |
609 | following property: | |
610 | - the component is non-addressable, or | |
611 | - the parent has alias set zero, | |
612 | or NULL_TREE if no such parent exists. In the former cases, the alias | |
613 | set of this parent is the alias set that must be used for T itself. */ | |
614 | ||
615 | tree | |
616 | component_uses_parent_alias_set_from (const_tree t) | |
6e24b709 | 617 | { |
b4ada065 | 618 | const_tree found = NULL_TREE; |
afe84921 | 619 | |
b4ada065 RB |
620 | while (handled_component_p (t)) |
621 | { | |
afe84921 RH |
622 | switch (TREE_CODE (t)) |
623 | { | |
624 | case COMPONENT_REF: | |
625 | if (DECL_NONADDRESSABLE_P (TREE_OPERAND (t, 1))) | |
b4ada065 | 626 | found = t; |
afe84921 RH |
627 | break; |
628 | ||
629 | case ARRAY_REF: | |
630 | case ARRAY_RANGE_REF: | |
631 | if (TYPE_NONALIASED_COMPONENT (TREE_TYPE (TREE_OPERAND (t, 0)))) | |
b4ada065 | 632 | found = t; |
afe84921 RH |
633 | break; |
634 | ||
635 | case REALPART_EXPR: | |
636 | case IMAGPART_EXPR: | |
637 | break; | |
638 | ||
b4ada065 RB |
639 | case BIT_FIELD_REF: |
640 | case VIEW_CONVERT_EXPR: | |
afe84921 | 641 | /* Bitfields and casts are never addressable. */ |
b4ada065 RB |
642 | found = t; |
643 | break; | |
644 | ||
645 | default: | |
646 | gcc_unreachable (); | |
afe84921 RH |
647 | } |
648 | ||
b4ada065 RB |
649 | if (get_alias_set (TREE_TYPE (TREE_OPERAND (t, 0))) == 0) |
650 | found = t; | |
651 | ||
afe84921 RH |
652 | t = TREE_OPERAND (t, 0); |
653 | } | |
b4ada065 RB |
654 | |
655 | if (found) | |
656 | return TREE_OPERAND (found, 0); | |
657 | ||
658 | return NULL_TREE; | |
6e24b709 RK |
659 | } |
660 | ||
f40333af RB |
661 | |
662 | /* Return whether the pointer-type T effective for aliasing may | |
663 | access everything and thus the reference has to be assigned | |
664 | alias-set zero. */ | |
665 | ||
666 | static bool | |
667 | ref_all_alias_ptr_type_p (const_tree t) | |
668 | { | |
669 | return (TREE_CODE (TREE_TYPE (t)) == VOID_TYPE | |
670 | || TYPE_REF_CAN_ALIAS_ALL (t)); | |
671 | } | |
672 | ||
5006671f RG |
673 | /* Return the alias set for the memory pointed to by T, which may be |
674 | either a type or an expression. Return -1 if there is nothing | |
675 | special about dereferencing T. */ | |
676 | ||
677 | static alias_set_type | |
678 | get_deref_alias_set_1 (tree t) | |
679 | { | |
5b21f0f3 | 680 | /* All we care about is the type. */ |
5006671f | 681 | if (! TYPE_P (t)) |
5b21f0f3 | 682 | t = TREE_TYPE (t); |
5006671f RG |
683 | |
684 | /* If we have an INDIRECT_REF via a void pointer, we don't | |
685 | know anything about what that might alias. Likewise if the | |
686 | pointer is marked that way. */ | |
f40333af | 687 | if (ref_all_alias_ptr_type_p (t)) |
5006671f RG |
688 | return 0; |
689 | ||
690 | return -1; | |
691 | } | |
692 | ||
693 | /* Return the alias set for the memory pointed to by T, which may be | |
694 | either a type or an expression. */ | |
695 | ||
696 | alias_set_type | |
697 | get_deref_alias_set (tree t) | |
698 | { | |
f40333af RB |
699 | /* If we're not doing any alias analysis, just assume everything |
700 | aliases everything else. */ | |
701 | if (!flag_strict_aliasing) | |
702 | return 0; | |
703 | ||
5006671f RG |
704 | alias_set_type set = get_deref_alias_set_1 (t); |
705 | ||
706 | /* Fall back to the alias-set of the pointed-to type. */ | |
707 | if (set == -1) | |
708 | { | |
709 | if (! TYPE_P (t)) | |
710 | t = TREE_TYPE (t); | |
711 | set = get_alias_set (TREE_TYPE (t)); | |
712 | } | |
713 | ||
714 | return set; | |
715 | } | |
716 | ||
f40333af RB |
717 | /* Return the pointer-type relevant for TBAA purposes from the |
718 | memory reference tree *T or NULL_TREE in which case *T is | |
719 | adjusted to point to the outermost component reference that | |
720 | can be used for assigning an alias set. */ | |
721 | ||
722 | static tree | |
723 | reference_alias_ptr_type_1 (tree *t) | |
724 | { | |
725 | tree inner; | |
726 | ||
727 | /* Get the base object of the reference. */ | |
728 | inner = *t; | |
729 | while (handled_component_p (inner)) | |
730 | { | |
731 | /* If there is a VIEW_CONVERT_EXPR in the chain we cannot use | |
732 | the type of any component references that wrap it to | |
733 | determine the alias-set. */ | |
734 | if (TREE_CODE (inner) == VIEW_CONVERT_EXPR) | |
735 | *t = TREE_OPERAND (inner, 0); | |
736 | inner = TREE_OPERAND (inner, 0); | |
737 | } | |
738 | ||
739 | /* Handle pointer dereferences here, they can override the | |
740 | alias-set. */ | |
741 | if (INDIRECT_REF_P (inner) | |
742 | && ref_all_alias_ptr_type_p (TREE_TYPE (TREE_OPERAND (inner, 0)))) | |
743 | return TREE_TYPE (TREE_OPERAND (inner, 0)); | |
744 | else if (TREE_CODE (inner) == TARGET_MEM_REF) | |
745 | return TREE_TYPE (TMR_OFFSET (inner)); | |
746 | else if (TREE_CODE (inner) == MEM_REF | |
747 | && ref_all_alias_ptr_type_p (TREE_TYPE (TREE_OPERAND (inner, 1)))) | |
748 | return TREE_TYPE (TREE_OPERAND (inner, 1)); | |
749 | ||
750 | /* If the innermost reference is a MEM_REF that has a | |
751 | conversion embedded treat it like a VIEW_CONVERT_EXPR above, | |
752 | using the memory access type for determining the alias-set. */ | |
753 | if (TREE_CODE (inner) == MEM_REF | |
754 | && (TYPE_MAIN_VARIANT (TREE_TYPE (inner)) | |
755 | != TYPE_MAIN_VARIANT | |
756 | (TREE_TYPE (TREE_TYPE (TREE_OPERAND (inner, 1)))))) | |
757 | return TREE_TYPE (TREE_OPERAND (inner, 1)); | |
758 | ||
b4ada065 RB |
759 | /* Otherwise, pick up the outermost object that we could have |
760 | a pointer to. */ | |
761 | tree tem = component_uses_parent_alias_set_from (*t); | |
762 | if (tem) | |
763 | *t = tem; | |
f40333af RB |
764 | |
765 | return NULL_TREE; | |
766 | } | |
767 | ||
768 | /* Return the pointer-type relevant for TBAA purposes from the | |
769 | gimple memory reference tree T. This is the type to be used for | |
770 | the offset operand of MEM_REF or TARGET_MEM_REF replacements of T | |
771 | and guarantees that get_alias_set will return the same alias | |
772 | set for T and the replacement. */ | |
773 | ||
774 | tree | |
775 | reference_alias_ptr_type (tree t) | |
776 | { | |
777 | tree ptype = reference_alias_ptr_type_1 (&t); | |
778 | /* If there is a given pointer type for aliasing purposes, return it. */ | |
779 | if (ptype != NULL_TREE) | |
780 | return ptype; | |
781 | ||
782 | /* Otherwise build one from the outermost component reference we | |
783 | may use. */ | |
784 | if (TREE_CODE (t) == MEM_REF | |
785 | || TREE_CODE (t) == TARGET_MEM_REF) | |
786 | return TREE_TYPE (TREE_OPERAND (t, 1)); | |
787 | else | |
788 | return build_pointer_type (TYPE_MAIN_VARIANT (TREE_TYPE (t))); | |
789 | } | |
790 | ||
791 | /* Return whether the pointer-types T1 and T2 used to determine | |
792 | two alias sets of two references will yield the same answer | |
793 | from get_deref_alias_set. */ | |
794 | ||
795 | bool | |
796 | alias_ptr_types_compatible_p (tree t1, tree t2) | |
797 | { | |
798 | if (TYPE_MAIN_VARIANT (t1) == TYPE_MAIN_VARIANT (t2)) | |
799 | return true; | |
800 | ||
801 | if (ref_all_alias_ptr_type_p (t1) | |
802 | || ref_all_alias_ptr_type_p (t2)) | |
803 | return false; | |
804 | ||
805 | return (TYPE_MAIN_VARIANT (TREE_TYPE (t1)) | |
806 | == TYPE_MAIN_VARIANT (TREE_TYPE (t2))); | |
807 | } | |
808 | ||
6e042ef4 JH |
809 | /* Create emptry alias set entry. */ |
810 | ||
811 | alias_set_entry | |
812 | init_alias_set_entry (alias_set_type set) | |
813 | { | |
814 | alias_set_entry ase = ggc_alloc<alias_set_entry_d> (); | |
815 | ase->alias_set = set; | |
816 | ase->children = NULL; | |
817 | ase->has_zero_child = false; | |
818 | ase->is_pointer = false; | |
819 | ase->has_pointer = false; | |
820 | gcc_checking_assert (!get_alias_set_entry (set)); | |
821 | (*alias_sets)[set] = ase; | |
822 | return ase; | |
823 | } | |
824 | ||
3bdf5ad1 RK |
825 | /* Return the alias set for T, which may be either a type or an |
826 | expression. Call language-specific routine for help, if needed. */ | |
827 | ||
4862826d | 828 | alias_set_type |
4682ae04 | 829 | get_alias_set (tree t) |
3bdf5ad1 | 830 | { |
4862826d | 831 | alias_set_type set; |
3bdf5ad1 RK |
832 | |
833 | /* If we're not doing any alias analysis, just assume everything | |
834 | aliases everything else. Also return 0 if this or its type is | |
835 | an error. */ | |
836 | if (! flag_strict_aliasing || t == error_mark_node | |
837 | || (! TYPE_P (t) | |
838 | && (TREE_TYPE (t) == 0 || TREE_TYPE (t) == error_mark_node))) | |
839 | return 0; | |
840 | ||
841 | /* We can be passed either an expression or a type. This and the | |
f47e9b4e RK |
842 | language-specific routine may make mutually-recursive calls to each other |
843 | to figure out what to do. At each juncture, we see if this is a tree | |
844 | that the language may need to handle specially. First handle things that | |
738cc472 | 845 | aren't types. */ |
f824e5c3 | 846 | if (! TYPE_P (t)) |
3bdf5ad1 | 847 | { |
70f34814 RG |
848 | /* Give the language a chance to do something with this tree |
849 | before we look at it. */ | |
8ac61af7 | 850 | STRIP_NOPS (t); |
ae2bcd98 | 851 | set = lang_hooks.get_alias_set (t); |
8ac61af7 RK |
852 | if (set != -1) |
853 | return set; | |
854 | ||
f40333af RB |
855 | /* Get the alias pointer-type to use or the outermost object |
856 | that we could have a pointer to. */ | |
857 | tree ptype = reference_alias_ptr_type_1 (&t); | |
858 | if (ptype != NULL) | |
859 | return get_deref_alias_set (ptype); | |
f824e5c3 | 860 | |
738cc472 RK |
861 | /* If we've already determined the alias set for a decl, just return |
862 | it. This is necessary for C++ anonymous unions, whose component | |
863 | variables don't look like union members (boo!). */ | |
5755cd38 | 864 | if (TREE_CODE (t) == VAR_DECL |
3c0cb5de | 865 | && DECL_RTL_SET_P (t) && MEM_P (DECL_RTL (t))) |
5755cd38 JM |
866 | return MEM_ALIAS_SET (DECL_RTL (t)); |
867 | ||
f824e5c3 RK |
868 | /* Now all we care about is the type. */ |
869 | t = TREE_TYPE (t); | |
3bdf5ad1 RK |
870 | } |
871 | ||
3bdf5ad1 | 872 | /* Variant qualifiers don't affect the alias set, so get the main |
daad0278 | 873 | variant. */ |
3bdf5ad1 | 874 | t = TYPE_MAIN_VARIANT (t); |
daad0278 RG |
875 | |
876 | /* Always use the canonical type as well. If this is a type that | |
877 | requires structural comparisons to identify compatible types | |
878 | use alias set zero. */ | |
879 | if (TYPE_STRUCTURAL_EQUALITY_P (t)) | |
cb9c2485 JM |
880 | { |
881 | /* Allow the language to specify another alias set for this | |
882 | type. */ | |
883 | set = lang_hooks.get_alias_set (t); | |
884 | if (set != -1) | |
885 | return set; | |
886 | return 0; | |
887 | } | |
7be7d292 | 888 | |
daad0278 | 889 | t = TYPE_CANONICAL (t); |
7be7d292 | 890 | |
96d91dcf RG |
891 | /* The canonical type should not require structural equality checks. */ |
892 | gcc_checking_assert (!TYPE_STRUCTURAL_EQUALITY_P (t)); | |
daad0278 RG |
893 | |
894 | /* If this is a type with a known alias set, return it. */ | |
738cc472 | 895 | if (TYPE_ALIAS_SET_KNOWN_P (t)) |
3bdf5ad1 RK |
896 | return TYPE_ALIAS_SET (t); |
897 | ||
36784d0e RG |
898 | /* We don't want to set TYPE_ALIAS_SET for incomplete types. */ |
899 | if (!COMPLETE_TYPE_P (t)) | |
900 | { | |
901 | /* For arrays with unknown size the conservative answer is the | |
902 | alias set of the element type. */ | |
903 | if (TREE_CODE (t) == ARRAY_TYPE) | |
904 | return get_alias_set (TREE_TYPE (t)); | |
905 | ||
906 | /* But return zero as a conservative answer for incomplete types. */ | |
907 | return 0; | |
908 | } | |
909 | ||
3bdf5ad1 | 910 | /* See if the language has special handling for this type. */ |
ae2bcd98 | 911 | set = lang_hooks.get_alias_set (t); |
8ac61af7 | 912 | if (set != -1) |
738cc472 | 913 | return set; |
2bf105ab | 914 | |
3bdf5ad1 RK |
915 | /* There are no objects of FUNCTION_TYPE, so there's no point in |
916 | using up an alias set for them. (There are, of course, pointers | |
917 | and references to functions, but that's different.) */ | |
7be7d292 | 918 | else if (TREE_CODE (t) == FUNCTION_TYPE || TREE_CODE (t) == METHOD_TYPE) |
3bdf5ad1 | 919 | set = 0; |
74d86f4f RH |
920 | |
921 | /* Unless the language specifies otherwise, let vector types alias | |
922 | their components. This avoids some nasty type punning issues in | |
923 | normal usage. And indeed lets vectors be treated more like an | |
924 | array slice. */ | |
925 | else if (TREE_CODE (t) == VECTOR_TYPE) | |
926 | set = get_alias_set (TREE_TYPE (t)); | |
927 | ||
4653cae5 RG |
928 | /* Unless the language specifies otherwise, treat array types the |
929 | same as their components. This avoids the asymmetry we get | |
930 | through recording the components. Consider accessing a | |
931 | character(kind=1) through a reference to a character(kind=1)[1:1]. | |
932 | Or consider if we want to assign integer(kind=4)[0:D.1387] and | |
933 | integer(kind=4)[4] the same alias set or not. | |
934 | Just be pragmatic here and make sure the array and its element | |
935 | type get the same alias set assigned. */ | |
7be7d292 | 936 | else if (TREE_CODE (t) == ARRAY_TYPE && !TYPE_NONALIASED_COMPONENT (t)) |
4653cae5 RG |
937 | set = get_alias_set (TREE_TYPE (t)); |
938 | ||
0ceb0201 RG |
939 | /* From the former common C and C++ langhook implementation: |
940 | ||
941 | Unfortunately, there is no canonical form of a pointer type. | |
942 | In particular, if we have `typedef int I', then `int *', and | |
943 | `I *' are different types. So, we have to pick a canonical | |
944 | representative. We do this below. | |
945 | ||
946 | Technically, this approach is actually more conservative that | |
947 | it needs to be. In particular, `const int *' and `int *' | |
948 | should be in different alias sets, according to the C and C++ | |
949 | standard, since their types are not the same, and so, | |
950 | technically, an `int **' and `const int **' cannot point at | |
951 | the same thing. | |
952 | ||
953 | But, the standard is wrong. In particular, this code is | |
954 | legal C++: | |
955 | ||
956 | int *ip; | |
957 | int **ipp = &ip; | |
958 | const int* const* cipp = ipp; | |
959 | And, it doesn't make sense for that to be legal unless you | |
960 | can dereference IPP and CIPP. So, we ignore cv-qualifiers on | |
961 | the pointed-to types. This issue has been reported to the | |
962 | C++ committee. | |
963 | ||
6e042ef4 JH |
964 | For this reason go to canonical type of the unqalified pointer type. |
965 | Until GCC 6 this code set all pointers sets to have alias set of | |
966 | ptr_type_node but that is a bad idea, because it prevents disabiguations | |
967 | in between pointers. For Firefox this accounts about 20% of all | |
968 | disambiguations in the program. */ | |
969 | else if (POINTER_TYPE_P (t) && t != ptr_type_node && !in_lto_p) | |
970 | { | |
971 | tree p; | |
972 | auto_vec <bool, 8> reference; | |
973 | ||
974 | /* Unnest all pointers and references. | |
975 | We also want to make pointer to array equivalent to pointer to its | |
976 | element. So skip all array types, too. */ | |
977 | for (p = t; POINTER_TYPE_P (p) | |
978 | || (TREE_CODE (p) == ARRAY_TYPE && !TYPE_NONALIASED_COMPONENT (p)); | |
979 | p = TREE_TYPE (p)) | |
980 | { | |
981 | if (TREE_CODE (p) == REFERENCE_TYPE) | |
982 | reference.safe_push (true); | |
983 | if (TREE_CODE (p) == POINTER_TYPE) | |
984 | reference.safe_push (false); | |
985 | } | |
986 | p = TYPE_MAIN_VARIANT (p); | |
987 | ||
988 | /* Make void * compatible with char * and also void **. | |
989 | Programs are commonly violating TBAA by this. | |
990 | ||
991 | We also make void * to conflict with every pointer | |
992 | (see record_component_aliases) and thus it is safe it to use it for | |
993 | pointers to types with TYPE_STRUCTURAL_EQUALITY_P. */ | |
994 | if (TREE_CODE (p) == VOID_TYPE || TYPE_STRUCTURAL_EQUALITY_P (p)) | |
995 | set = get_alias_set (ptr_type_node); | |
996 | else | |
997 | { | |
998 | /* Rebuild pointer type from starting from canonical types using | |
999 | unqualified pointers and references only. This way all such | |
1000 | pointers will have the same alias set and will conflict with | |
1001 | each other. | |
1002 | ||
1003 | Most of time we already have pointers or references of a given type. | |
1004 | If not we build new one just to be sure that if someone later | |
1005 | (probably only middle-end can, as we should assign all alias | |
1006 | classes only after finishing translation unit) builds the pointer | |
1007 | type, the canonical type will match. */ | |
1008 | p = TYPE_CANONICAL (p); | |
1009 | while (!reference.is_empty ()) | |
1010 | { | |
1011 | if (reference.pop ()) | |
1012 | p = build_reference_type (p); | |
1013 | else | |
1014 | p = build_pointer_type (p); | |
1015 | p = TYPE_CANONICAL (TYPE_MAIN_VARIANT (p)); | |
1016 | } | |
1017 | gcc_checking_assert (TYPE_CANONICAL (p) == p); | |
1018 | ||
1019 | /* Assign the alias set to both p and t. | |
1020 | We can not call get_alias_set (p) here as that would trigger | |
1021 | infinite recursion when p == t. In other cases it would just | |
1022 | trigger unnecesary legwork of rebuilding the pointer again. */ | |
1023 | if (TYPE_ALIAS_SET_KNOWN_P (p)) | |
1024 | set = TYPE_ALIAS_SET (p); | |
1025 | else | |
1026 | { | |
1027 | set = new_alias_set (); | |
1028 | TYPE_ALIAS_SET (p) = set; | |
1029 | } | |
1030 | } | |
1031 | } | |
1032 | /* In LTO the rules above needs to be part of canonical type machinery. | |
1033 | For now just punt. */ | |
ccb5ad37 JH |
1034 | else if (POINTER_TYPE_P (t) |
1035 | && t != TYPE_CANONICAL (ptr_type_node) && in_lto_p) | |
1036 | set = get_alias_set (TYPE_CANONICAL (ptr_type_node)); | |
0ceb0201 | 1037 | |
7be7d292 | 1038 | /* Otherwise make a new alias set for this type. */ |
3bdf5ad1 | 1039 | else |
96d91dcf RG |
1040 | { |
1041 | /* Each canonical type gets its own alias set, so canonical types | |
1042 | shouldn't form a tree. It doesn't really matter for types | |
1043 | we handle specially above, so only check it where it possibly | |
1044 | would result in a bogus alias set. */ | |
1045 | gcc_checking_assert (TYPE_CANONICAL (t) == t); | |
1046 | ||
1047 | set = new_alias_set (); | |
1048 | } | |
3bdf5ad1 RK |
1049 | |
1050 | TYPE_ALIAS_SET (t) = set; | |
2bf105ab | 1051 | |
7be7d292 EB |
1052 | /* If this is an aggregate type or a complex type, we must record any |
1053 | component aliasing information. */ | |
1d79fd2c | 1054 | if (AGGREGATE_TYPE_P (t) || TREE_CODE (t) == COMPLEX_TYPE) |
2bf105ab RK |
1055 | record_component_aliases (t); |
1056 | ||
6e042ef4 JH |
1057 | /* We treat pointer types specially in alias_set_subset_of. */ |
1058 | if (POINTER_TYPE_P (t) && set) | |
1059 | { | |
1060 | alias_set_entry ase = get_alias_set_entry (set); | |
1061 | if (!ase) | |
1062 | ase = init_alias_set_entry (set); | |
1063 | ase->is_pointer = true; | |
1064 | ase->has_pointer = true; | |
1065 | } | |
1066 | ||
3bdf5ad1 RK |
1067 | return set; |
1068 | } | |
1069 | ||
1070 | /* Return a brand-new alias set. */ | |
1071 | ||
4862826d | 1072 | alias_set_type |
4682ae04 | 1073 | new_alias_set (void) |
3bdf5ad1 | 1074 | { |
3bdf5ad1 | 1075 | if (flag_strict_aliasing) |
9ddb66ca | 1076 | { |
1a5640b4 | 1077 | if (alias_sets == 0) |
9771b263 DN |
1078 | vec_safe_push (alias_sets, (alias_set_entry) 0); |
1079 | vec_safe_push (alias_sets, (alias_set_entry) 0); | |
1080 | return alias_sets->length () - 1; | |
9ddb66ca | 1081 | } |
3bdf5ad1 RK |
1082 | else |
1083 | return 0; | |
1084 | } | |
3932261a | 1085 | |
01d28c3f JM |
1086 | /* Indicate that things in SUBSET can alias things in SUPERSET, but that |
1087 | not everything that aliases SUPERSET also aliases SUBSET. For example, | |
1088 | in C, a store to an `int' can alias a load of a structure containing an | |
1089 | `int', and vice versa. But it can't alias a load of a 'double' member | |
1090 | of the same structure. Here, the structure would be the SUPERSET and | |
1091 | `int' the SUBSET. This relationship is also described in the comment at | |
1092 | the beginning of this file. | |
1093 | ||
1094 | This function should be called only once per SUPERSET/SUBSET pair. | |
3932261a MM |
1095 | |
1096 | It is illegal for SUPERSET to be zero; everything is implicitly a | |
1097 | subset of alias set zero. */ | |
1098 | ||
794511d2 | 1099 | void |
4862826d | 1100 | record_alias_subset (alias_set_type superset, alias_set_type subset) |
3932261a MM |
1101 | { |
1102 | alias_set_entry superset_entry; | |
1103 | alias_set_entry subset_entry; | |
1104 | ||
f47e9b4e RK |
1105 | /* It is possible in complex type situations for both sets to be the same, |
1106 | in which case we can ignore this operation. */ | |
1107 | if (superset == subset) | |
1108 | return; | |
1109 | ||
298e6adc | 1110 | gcc_assert (superset); |
3932261a MM |
1111 | |
1112 | superset_entry = get_alias_set_entry (superset); | |
ca7fd9cd | 1113 | if (superset_entry == 0) |
3932261a MM |
1114 | { |
1115 | /* Create an entry for the SUPERSET, so that we have a place to | |
1116 | attach the SUBSET. */ | |
6e042ef4 | 1117 | superset_entry = init_alias_set_entry (superset); |
3932261a MM |
1118 | } |
1119 | ||
2bf105ab RK |
1120 | if (subset == 0) |
1121 | superset_entry->has_zero_child = 1; | |
1122 | else | |
1123 | { | |
1124 | subset_entry = get_alias_set_entry (subset); | |
6e042ef4 JH |
1125 | if (!superset_entry->children) |
1126 | superset_entry->children | |
fb5c464a | 1127 | = hash_map<alias_set_hash, int>::create_ggc (64); |
2bf105ab RK |
1128 | /* If there is an entry for the subset, enter all of its children |
1129 | (if they are not already present) as children of the SUPERSET. */ | |
ca7fd9cd | 1130 | if (subset_entry) |
2bf105ab RK |
1131 | { |
1132 | if (subset_entry->has_zero_child) | |
6e042ef4 JH |
1133 | superset_entry->has_zero_child = true; |
1134 | if (subset_entry->has_pointer) | |
1135 | superset_entry->has_pointer = true; | |
d4b60170 | 1136 | |
6e042ef4 JH |
1137 | if (subset_entry->children) |
1138 | { | |
fb5c464a | 1139 | hash_map<alias_set_hash, int>::iterator iter |
6e042ef4 JH |
1140 | = subset_entry->children->begin (); |
1141 | for (; iter != subset_entry->children->end (); ++iter) | |
1142 | superset_entry->children->put ((*iter).first, (*iter).second); | |
1143 | } | |
2bf105ab | 1144 | } |
3932261a | 1145 | |
2bf105ab | 1146 | /* Enter the SUBSET itself as a child of the SUPERSET. */ |
de144fb2 | 1147 | superset_entry->children->put (subset, 0); |
2bf105ab | 1148 | } |
3932261a MM |
1149 | } |
1150 | ||
a0c33338 RK |
1151 | /* Record that component types of TYPE, if any, are part of that type for |
1152 | aliasing purposes. For record types, we only record component types | |
b5487346 EB |
1153 | for fields that are not marked non-addressable. For array types, we |
1154 | only record the component type if it is not marked non-aliased. */ | |
a0c33338 RK |
1155 | |
1156 | void | |
4682ae04 | 1157 | record_component_aliases (tree type) |
a0c33338 | 1158 | { |
4862826d | 1159 | alias_set_type superset = get_alias_set (type); |
a0c33338 RK |
1160 | tree field; |
1161 | ||
1162 | if (superset == 0) | |
1163 | return; | |
1164 | ||
1165 | switch (TREE_CODE (type)) | |
1166 | { | |
a0c33338 RK |
1167 | case RECORD_TYPE: |
1168 | case UNION_TYPE: | |
1169 | case QUAL_UNION_TYPE: | |
910ad8de | 1170 | for (field = TYPE_FIELDS (type); field != 0; field = DECL_CHAIN (field)) |
b5487346 | 1171 | if (TREE_CODE (field) == FIELD_DECL && !DECL_NONADDRESSABLE_P (field)) |
2bf105ab | 1172 | record_alias_subset (superset, get_alias_set (TREE_TYPE (field))); |
a0c33338 RK |
1173 | break; |
1174 | ||
1d79fd2c JW |
1175 | case COMPLEX_TYPE: |
1176 | record_alias_subset (superset, get_alias_set (TREE_TYPE (type))); | |
1177 | break; | |
1178 | ||
4653cae5 RG |
1179 | /* VECTOR_TYPE and ARRAY_TYPE share the alias set with their |
1180 | element type. */ | |
1181 | ||
a0c33338 RK |
1182 | default: |
1183 | break; | |
1184 | } | |
1185 | } | |
1186 | ||
3bdf5ad1 RK |
1187 | /* Allocate an alias set for use in storing and reading from the varargs |
1188 | spill area. */ | |
1189 | ||
4862826d | 1190 | static GTY(()) alias_set_type varargs_set = -1; |
f103e34d | 1191 | |
4862826d | 1192 | alias_set_type |
4682ae04 | 1193 | get_varargs_alias_set (void) |
3bdf5ad1 | 1194 | { |
cd3ce9b4 JM |
1195 | #if 1 |
1196 | /* We now lower VA_ARG_EXPR, and there's currently no way to attach the | |
1197 | varargs alias set to an INDIRECT_REF (FIXME!), so we can't | |
1198 | consistently use the varargs alias set for loads from the varargs | |
1199 | area. So don't use it anywhere. */ | |
1200 | return 0; | |
1201 | #else | |
f103e34d GK |
1202 | if (varargs_set == -1) |
1203 | varargs_set = new_alias_set (); | |
3bdf5ad1 | 1204 | |
f103e34d | 1205 | return varargs_set; |
cd3ce9b4 | 1206 | #endif |
3bdf5ad1 RK |
1207 | } |
1208 | ||
1209 | /* Likewise, but used for the fixed portions of the frame, e.g., register | |
1210 | save areas. */ | |
1211 | ||
4862826d | 1212 | static GTY(()) alias_set_type frame_set = -1; |
f103e34d | 1213 | |
4862826d | 1214 | alias_set_type |
4682ae04 | 1215 | get_frame_alias_set (void) |
3bdf5ad1 | 1216 | { |
f103e34d GK |
1217 | if (frame_set == -1) |
1218 | frame_set = new_alias_set (); | |
3bdf5ad1 | 1219 | |
f103e34d | 1220 | return frame_set; |
3bdf5ad1 RK |
1221 | } |
1222 | ||
9fc37b2b RS |
1223 | /* Create a new, unique base with id ID. */ |
1224 | ||
1225 | static rtx | |
1226 | unique_base_value (HOST_WIDE_INT id) | |
1227 | { | |
1228 | return gen_rtx_ADDRESS (Pmode, id); | |
1229 | } | |
1230 | ||
1231 | /* Return true if accesses based on any other base value cannot alias | |
1232 | those based on X. */ | |
1233 | ||
1234 | static bool | |
1235 | unique_base_value_p (rtx x) | |
1236 | { | |
1237 | return GET_CODE (x) == ADDRESS && GET_MODE (x) == Pmode; | |
1238 | } | |
1239 | ||
1240 | /* Return true if X is known to be a base value. */ | |
1241 | ||
1242 | static bool | |
1243 | known_base_value_p (rtx x) | |
1244 | { | |
1245 | switch (GET_CODE (x)) | |
1246 | { | |
1247 | case LABEL_REF: | |
1248 | case SYMBOL_REF: | |
1249 | return true; | |
1250 | ||
1251 | case ADDRESS: | |
1252 | /* Arguments may or may not be bases; we don't know for sure. */ | |
1253 | return GET_MODE (x) != VOIDmode; | |
1254 | ||
1255 | default: | |
1256 | return false; | |
1257 | } | |
1258 | } | |
1259 | ||
2a2c8203 JC |
1260 | /* Inside SRC, the source of a SET, find a base address. */ |
1261 | ||
9ae8ffe7 | 1262 | static rtx |
4682ae04 | 1263 | find_base_value (rtx src) |
9ae8ffe7 | 1264 | { |
713f41f9 | 1265 | unsigned int regno; |
0aacc8ed | 1266 | |
53451050 RS |
1267 | #if defined (FIND_BASE_TERM) |
1268 | /* Try machine-dependent ways to find the base term. */ | |
1269 | src = FIND_BASE_TERM (src); | |
1270 | #endif | |
1271 | ||
9ae8ffe7 JL |
1272 | switch (GET_CODE (src)) |
1273 | { | |
1274 | case SYMBOL_REF: | |
1275 | case LABEL_REF: | |
1276 | return src; | |
1277 | ||
1278 | case REG: | |
fb6754f0 | 1279 | regno = REGNO (src); |
d4b60170 | 1280 | /* At the start of a function, argument registers have known base |
2a2c8203 JC |
1281 | values which may be lost later. Returning an ADDRESS |
1282 | expression here allows optimization based on argument values | |
1283 | even when the argument registers are used for other purposes. */ | |
713f41f9 BS |
1284 | if (regno < FIRST_PSEUDO_REGISTER && copying_arguments) |
1285 | return new_reg_base_value[regno]; | |
73774bc7 | 1286 | |
eaf407a5 | 1287 | /* If a pseudo has a known base value, return it. Do not do this |
9b462c42 RH |
1288 | for non-fixed hard regs since it can result in a circular |
1289 | dependency chain for registers which have values at function entry. | |
eaf407a5 JL |
1290 | |
1291 | The test above is not sufficient because the scheduler may move | |
1292 | a copy out of an arg reg past the NOTE_INSN_FUNCTION_BEGIN. */ | |
9b462c42 | 1293 | if ((regno >= FIRST_PSEUDO_REGISTER || fixed_regs[regno]) |
9771b263 | 1294 | && regno < vec_safe_length (reg_base_value)) |
83bbd9b6 RH |
1295 | { |
1296 | /* If we're inside init_alias_analysis, use new_reg_base_value | |
1297 | to reduce the number of relaxation iterations. */ | |
1afdf91c | 1298 | if (new_reg_base_value && new_reg_base_value[regno] |
6fb5fa3c | 1299 | && DF_REG_DEF_COUNT (regno) == 1) |
83bbd9b6 RH |
1300 | return new_reg_base_value[regno]; |
1301 | ||
9771b263 DN |
1302 | if ((*reg_base_value)[regno]) |
1303 | return (*reg_base_value)[regno]; | |
83bbd9b6 | 1304 | } |
73774bc7 | 1305 | |
e3f049a8 | 1306 | return 0; |
9ae8ffe7 JL |
1307 | |
1308 | case MEM: | |
1309 | /* Check for an argument passed in memory. Only record in the | |
1310 | copying-arguments block; it is too hard to track changes | |
1311 | otherwise. */ | |
1312 | if (copying_arguments | |
1313 | && (XEXP (src, 0) == arg_pointer_rtx | |
1314 | || (GET_CODE (XEXP (src, 0)) == PLUS | |
1315 | && XEXP (XEXP (src, 0), 0) == arg_pointer_rtx))) | |
9fc37b2b | 1316 | return arg_base_value; |
9ae8ffe7 JL |
1317 | return 0; |
1318 | ||
1319 | case CONST: | |
1320 | src = XEXP (src, 0); | |
1321 | if (GET_CODE (src) != PLUS && GET_CODE (src) != MINUS) | |
1322 | break; | |
d4b60170 | 1323 | |
ec5c56db | 1324 | /* ... fall through ... */ |
2a2c8203 | 1325 | |
9ae8ffe7 JL |
1326 | case PLUS: |
1327 | case MINUS: | |
2a2c8203 | 1328 | { |
ec907dd8 JL |
1329 | rtx temp, src_0 = XEXP (src, 0), src_1 = XEXP (src, 1); |
1330 | ||
0134bf2d DE |
1331 | /* If either operand is a REG that is a known pointer, then it |
1332 | is the base. */ | |
1333 | if (REG_P (src_0) && REG_POINTER (src_0)) | |
1334 | return find_base_value (src_0); | |
1335 | if (REG_P (src_1) && REG_POINTER (src_1)) | |
1336 | return find_base_value (src_1); | |
1337 | ||
ec907dd8 JL |
1338 | /* If either operand is a REG, then see if we already have |
1339 | a known value for it. */ | |
0134bf2d | 1340 | if (REG_P (src_0)) |
ec907dd8 JL |
1341 | { |
1342 | temp = find_base_value (src_0); | |
d4b60170 | 1343 | if (temp != 0) |
ec907dd8 JL |
1344 | src_0 = temp; |
1345 | } | |
1346 | ||
0134bf2d | 1347 | if (REG_P (src_1)) |
ec907dd8 JL |
1348 | { |
1349 | temp = find_base_value (src_1); | |
d4b60170 | 1350 | if (temp!= 0) |
ec907dd8 JL |
1351 | src_1 = temp; |
1352 | } | |
2a2c8203 | 1353 | |
0134bf2d DE |
1354 | /* If either base is named object or a special address |
1355 | (like an argument or stack reference), then use it for the | |
1356 | base term. */ | |
9fc37b2b | 1357 | if (src_0 != 0 && known_base_value_p (src_0)) |
0134bf2d DE |
1358 | return src_0; |
1359 | ||
9fc37b2b | 1360 | if (src_1 != 0 && known_base_value_p (src_1)) |
0134bf2d DE |
1361 | return src_1; |
1362 | ||
d4b60170 | 1363 | /* Guess which operand is the base address: |
ec907dd8 JL |
1364 | If either operand is a symbol, then it is the base. If |
1365 | either operand is a CONST_INT, then the other is the base. */ | |
481683e1 | 1366 | if (CONST_INT_P (src_1) || CONSTANT_P (src_0)) |
2a2c8203 | 1367 | return find_base_value (src_0); |
481683e1 | 1368 | else if (CONST_INT_P (src_0) || CONSTANT_P (src_1)) |
ec907dd8 JL |
1369 | return find_base_value (src_1); |
1370 | ||
9ae8ffe7 | 1371 | return 0; |
2a2c8203 JC |
1372 | } |
1373 | ||
1374 | case LO_SUM: | |
1375 | /* The standard form is (lo_sum reg sym) so look only at the | |
1376 | second operand. */ | |
1377 | return find_base_value (XEXP (src, 1)); | |
9ae8ffe7 JL |
1378 | |
1379 | case AND: | |
1380 | /* If the second operand is constant set the base | |
ec5c56db | 1381 | address to the first operand. */ |
481683e1 | 1382 | if (CONST_INT_P (XEXP (src, 1)) && INTVAL (XEXP (src, 1)) != 0) |
2a2c8203 | 1383 | return find_base_value (XEXP (src, 0)); |
9ae8ffe7 JL |
1384 | return 0; |
1385 | ||
61f0131c | 1386 | case TRUNCATE: |
5932a4d4 | 1387 | /* As we do not know which address space the pointer is referring to, we can |
d4ebfa65 BE |
1388 | handle this only if the target does not support different pointer or |
1389 | address modes depending on the address space. */ | |
1390 | if (!target_default_pointer_address_modes_p ()) | |
1391 | break; | |
61f0131c R |
1392 | if (GET_MODE_SIZE (GET_MODE (src)) < GET_MODE_SIZE (Pmode)) |
1393 | break; | |
1394 | /* Fall through. */ | |
9ae8ffe7 | 1395 | case HIGH: |
d288e53d DE |
1396 | case PRE_INC: |
1397 | case PRE_DEC: | |
1398 | case POST_INC: | |
1399 | case POST_DEC: | |
1400 | case PRE_MODIFY: | |
1401 | case POST_MODIFY: | |
2a2c8203 | 1402 | return find_base_value (XEXP (src, 0)); |
1d300e19 | 1403 | |
0aacc8ed RK |
1404 | case ZERO_EXTEND: |
1405 | case SIGN_EXTEND: /* used for NT/Alpha pointers */ | |
5932a4d4 | 1406 | /* As we do not know which address space the pointer is referring to, we can |
d4ebfa65 BE |
1407 | handle this only if the target does not support different pointer or |
1408 | address modes depending on the address space. */ | |
1409 | if (!target_default_pointer_address_modes_p ()) | |
1410 | break; | |
1411 | ||
0aacc8ed RK |
1412 | { |
1413 | rtx temp = find_base_value (XEXP (src, 0)); | |
1414 | ||
5ae6cd0d | 1415 | if (temp != 0 && CONSTANT_P (temp)) |
0aacc8ed | 1416 | temp = convert_memory_address (Pmode, temp); |
0aacc8ed RK |
1417 | |
1418 | return temp; | |
1419 | } | |
1420 | ||
1d300e19 KG |
1421 | default: |
1422 | break; | |
9ae8ffe7 JL |
1423 | } |
1424 | ||
1425 | return 0; | |
1426 | } | |
1427 | ||
9fc37b2b RS |
1428 | /* Called from init_alias_analysis indirectly through note_stores, |
1429 | or directly if DEST is a register with a REG_NOALIAS note attached. | |
1430 | SET is null in the latter case. */ | |
9ae8ffe7 | 1431 | |
d4b60170 | 1432 | /* While scanning insns to find base values, reg_seen[N] is nonzero if |
9ae8ffe7 | 1433 | register N has been set in this function. */ |
d630245f | 1434 | static sbitmap reg_seen; |
9ae8ffe7 | 1435 | |
2a2c8203 | 1436 | static void |
7bc980e1 | 1437 | record_set (rtx dest, const_rtx set, void *data ATTRIBUTE_UNUSED) |
9ae8ffe7 | 1438 | { |
b3694847 | 1439 | unsigned regno; |
9ae8ffe7 | 1440 | rtx src; |
c28b4e40 | 1441 | int n; |
9ae8ffe7 | 1442 | |
f8cfc6aa | 1443 | if (!REG_P (dest)) |
9ae8ffe7 JL |
1444 | return; |
1445 | ||
fb6754f0 | 1446 | regno = REGNO (dest); |
9ae8ffe7 | 1447 | |
9771b263 | 1448 | gcc_checking_assert (regno < reg_base_value->length ()); |
ac606739 | 1449 | |
dc8afb70 | 1450 | n = REG_NREGS (dest); |
c28b4e40 JW |
1451 | if (n != 1) |
1452 | { | |
1453 | while (--n >= 0) | |
1454 | { | |
d7c028c0 | 1455 | bitmap_set_bit (reg_seen, regno + n); |
c28b4e40 JW |
1456 | new_reg_base_value[regno + n] = 0; |
1457 | } | |
1458 | return; | |
1459 | } | |
1460 | ||
9ae8ffe7 JL |
1461 | if (set) |
1462 | { | |
1463 | /* A CLOBBER wipes out any old value but does not prevent a previously | |
1464 | unset register from acquiring a base address (i.e. reg_seen is not | |
1465 | set). */ | |
1466 | if (GET_CODE (set) == CLOBBER) | |
1467 | { | |
ec907dd8 | 1468 | new_reg_base_value[regno] = 0; |
9ae8ffe7 JL |
1469 | return; |
1470 | } | |
1471 | src = SET_SRC (set); | |
1472 | } | |
1473 | else | |
1474 | { | |
9fc37b2b | 1475 | /* There's a REG_NOALIAS note against DEST. */ |
d7c028c0 | 1476 | if (bitmap_bit_p (reg_seen, regno)) |
9ae8ffe7 | 1477 | { |
ec907dd8 | 1478 | new_reg_base_value[regno] = 0; |
9ae8ffe7 JL |
1479 | return; |
1480 | } | |
d7c028c0 | 1481 | bitmap_set_bit (reg_seen, regno); |
9fc37b2b | 1482 | new_reg_base_value[regno] = unique_base_value (unique_id++); |
9ae8ffe7 JL |
1483 | return; |
1484 | } | |
1485 | ||
5da6f168 RS |
1486 | /* If this is not the first set of REGNO, see whether the new value |
1487 | is related to the old one. There are two cases of interest: | |
1488 | ||
1489 | (1) The register might be assigned an entirely new value | |
1490 | that has the same base term as the original set. | |
1491 | ||
1492 | (2) The set might be a simple self-modification that | |
1493 | cannot change REGNO's base value. | |
1494 | ||
1495 | If neither case holds, reject the original base value as invalid. | |
1496 | Note that the following situation is not detected: | |
1497 | ||
c22cacf3 | 1498 | extern int x, y; int *p = &x; p += (&y-&x); |
5da6f168 | 1499 | |
9ae8ffe7 JL |
1500 | ANSI C does not allow computing the difference of addresses |
1501 | of distinct top level objects. */ | |
5da6f168 RS |
1502 | if (new_reg_base_value[regno] != 0 |
1503 | && find_base_value (src) != new_reg_base_value[regno]) | |
9ae8ffe7 JL |
1504 | switch (GET_CODE (src)) |
1505 | { | |
2a2c8203 | 1506 | case LO_SUM: |
9ae8ffe7 JL |
1507 | case MINUS: |
1508 | if (XEXP (src, 0) != dest && XEXP (src, 1) != dest) | |
ec907dd8 | 1509 | new_reg_base_value[regno] = 0; |
9ae8ffe7 | 1510 | break; |
61f0131c R |
1511 | case PLUS: |
1512 | /* If the value we add in the PLUS is also a valid base value, | |
1513 | this might be the actual base value, and the original value | |
1514 | an index. */ | |
1515 | { | |
1516 | rtx other = NULL_RTX; | |
1517 | ||
1518 | if (XEXP (src, 0) == dest) | |
1519 | other = XEXP (src, 1); | |
1520 | else if (XEXP (src, 1) == dest) | |
1521 | other = XEXP (src, 0); | |
1522 | ||
1523 | if (! other || find_base_value (other)) | |
1524 | new_reg_base_value[regno] = 0; | |
1525 | break; | |
1526 | } | |
9ae8ffe7 | 1527 | case AND: |
481683e1 | 1528 | if (XEXP (src, 0) != dest || !CONST_INT_P (XEXP (src, 1))) |
ec907dd8 | 1529 | new_reg_base_value[regno] = 0; |
9ae8ffe7 | 1530 | break; |
9ae8ffe7 | 1531 | default: |
ec907dd8 | 1532 | new_reg_base_value[regno] = 0; |
9ae8ffe7 JL |
1533 | break; |
1534 | } | |
1535 | /* If this is the first set of a register, record the value. */ | |
1536 | else if ((regno >= FIRST_PSEUDO_REGISTER || ! fixed_regs[regno]) | |
d7c028c0 | 1537 | && ! bitmap_bit_p (reg_seen, regno) && new_reg_base_value[regno] == 0) |
ec907dd8 | 1538 | new_reg_base_value[regno] = find_base_value (src); |
9ae8ffe7 | 1539 | |
d7c028c0 | 1540 | bitmap_set_bit (reg_seen, regno); |
9ae8ffe7 JL |
1541 | } |
1542 | ||
8fd0a474 AM |
1543 | /* Return REG_BASE_VALUE for REGNO. Selective scheduler uses this to avoid |
1544 | using hard registers with non-null REG_BASE_VALUE for renaming. */ | |
1545 | rtx | |
1546 | get_reg_base_value (unsigned int regno) | |
1547 | { | |
9771b263 | 1548 | return (*reg_base_value)[regno]; |
8fd0a474 AM |
1549 | } |
1550 | ||
bb1acb3e RH |
1551 | /* If a value is known for REGNO, return it. */ |
1552 | ||
c22cacf3 | 1553 | rtx |
bb1acb3e RH |
1554 | get_reg_known_value (unsigned int regno) |
1555 | { | |
1556 | if (regno >= FIRST_PSEUDO_REGISTER) | |
1557 | { | |
1558 | regno -= FIRST_PSEUDO_REGISTER; | |
9771b263 DN |
1559 | if (regno < vec_safe_length (reg_known_value)) |
1560 | return (*reg_known_value)[regno]; | |
bb1acb3e RH |
1561 | } |
1562 | return NULL; | |
43fe47ca JW |
1563 | } |
1564 | ||
bb1acb3e RH |
1565 | /* Set it. */ |
1566 | ||
1567 | static void | |
1568 | set_reg_known_value (unsigned int regno, rtx val) | |
1569 | { | |
1570 | if (regno >= FIRST_PSEUDO_REGISTER) | |
1571 | { | |
1572 | regno -= FIRST_PSEUDO_REGISTER; | |
9771b263 DN |
1573 | if (regno < vec_safe_length (reg_known_value)) |
1574 | (*reg_known_value)[regno] = val; | |
bb1acb3e RH |
1575 | } |
1576 | } | |
1577 | ||
1578 | /* Similarly for reg_known_equiv_p. */ | |
1579 | ||
1580 | bool | |
1581 | get_reg_known_equiv_p (unsigned int regno) | |
1582 | { | |
1583 | if (regno >= FIRST_PSEUDO_REGISTER) | |
1584 | { | |
1585 | regno -= FIRST_PSEUDO_REGISTER; | |
9771b263 | 1586 | if (regno < vec_safe_length (reg_known_value)) |
d7c028c0 | 1587 | return bitmap_bit_p (reg_known_equiv_p, regno); |
bb1acb3e RH |
1588 | } |
1589 | return false; | |
1590 | } | |
1591 | ||
1592 | static void | |
1593 | set_reg_known_equiv_p (unsigned int regno, bool val) | |
1594 | { | |
1595 | if (regno >= FIRST_PSEUDO_REGISTER) | |
1596 | { | |
1597 | regno -= FIRST_PSEUDO_REGISTER; | |
9771b263 | 1598 | if (regno < vec_safe_length (reg_known_value)) |
9ff3c7ca SB |
1599 | { |
1600 | if (val) | |
d7c028c0 | 1601 | bitmap_set_bit (reg_known_equiv_p, regno); |
9ff3c7ca | 1602 | else |
d7c028c0 | 1603 | bitmap_clear_bit (reg_known_equiv_p, regno); |
9ff3c7ca | 1604 | } |
bb1acb3e RH |
1605 | } |
1606 | } | |
1607 | ||
1608 | ||
db048faf MM |
1609 | /* Returns a canonical version of X, from the point of view alias |
1610 | analysis. (For example, if X is a MEM whose address is a register, | |
1611 | and the register has a known value (say a SYMBOL_REF), then a MEM | |
1612 | whose address is the SYMBOL_REF is returned.) */ | |
1613 | ||
1614 | rtx | |
4682ae04 | 1615 | canon_rtx (rtx x) |
9ae8ffe7 JL |
1616 | { |
1617 | /* Recursively look for equivalences. */ | |
f8cfc6aa | 1618 | if (REG_P (x) && REGNO (x) >= FIRST_PSEUDO_REGISTER) |
bb1acb3e RH |
1619 | { |
1620 | rtx t = get_reg_known_value (REGNO (x)); | |
1621 | if (t == x) | |
1622 | return x; | |
1623 | if (t) | |
1624 | return canon_rtx (t); | |
1625 | } | |
1626 | ||
1627 | if (GET_CODE (x) == PLUS) | |
9ae8ffe7 JL |
1628 | { |
1629 | rtx x0 = canon_rtx (XEXP (x, 0)); | |
1630 | rtx x1 = canon_rtx (XEXP (x, 1)); | |
1631 | ||
1632 | if (x0 != XEXP (x, 0) || x1 != XEXP (x, 1)) | |
1633 | { | |
481683e1 | 1634 | if (CONST_INT_P (x0)) |
0a81f074 | 1635 | return plus_constant (GET_MODE (x), x1, INTVAL (x0)); |
481683e1 | 1636 | else if (CONST_INT_P (x1)) |
0a81f074 | 1637 | return plus_constant (GET_MODE (x), x0, INTVAL (x1)); |
38a448ca | 1638 | return gen_rtx_PLUS (GET_MODE (x), x0, x1); |
9ae8ffe7 JL |
1639 | } |
1640 | } | |
d4b60170 | 1641 | |
9ae8ffe7 JL |
1642 | /* This gives us much better alias analysis when called from |
1643 | the loop optimizer. Note we want to leave the original | |
1644 | MEM alone, but need to return the canonicalized MEM with | |
1645 | all the flags with their original values. */ | |
3c0cb5de | 1646 | else if (MEM_P (x)) |
f1ec5147 | 1647 | x = replace_equiv_address_nv (x, canon_rtx (XEXP (x, 0))); |
d4b60170 | 1648 | |
9ae8ffe7 JL |
1649 | return x; |
1650 | } | |
1651 | ||
1652 | /* Return 1 if X and Y are identical-looking rtx's. | |
45183e03 | 1653 | Expect that X and Y has been already canonicalized. |
9ae8ffe7 JL |
1654 | |
1655 | We use the data in reg_known_value above to see if two registers with | |
1656 | different numbers are, in fact, equivalent. */ | |
1657 | ||
1658 | static int | |
ed7a4b4b | 1659 | rtx_equal_for_memref_p (const_rtx x, const_rtx y) |
9ae8ffe7 | 1660 | { |
b3694847 SS |
1661 | int i; |
1662 | int j; | |
1663 | enum rtx_code code; | |
1664 | const char *fmt; | |
9ae8ffe7 JL |
1665 | |
1666 | if (x == 0 && y == 0) | |
1667 | return 1; | |
1668 | if (x == 0 || y == 0) | |
1669 | return 0; | |
d4b60170 | 1670 | |
9ae8ffe7 JL |
1671 | if (x == y) |
1672 | return 1; | |
1673 | ||
1674 | code = GET_CODE (x); | |
1675 | /* Rtx's of different codes cannot be equal. */ | |
1676 | if (code != GET_CODE (y)) | |
1677 | return 0; | |
1678 | ||
1679 | /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. | |
1680 | (REG:SI x) and (REG:HI x) are NOT equivalent. */ | |
1681 | ||
1682 | if (GET_MODE (x) != GET_MODE (y)) | |
1683 | return 0; | |
1684 | ||
db048faf MM |
1685 | /* Some RTL can be compared without a recursive examination. */ |
1686 | switch (code) | |
1687 | { | |
1688 | case REG: | |
1689 | return REGNO (x) == REGNO (y); | |
1690 | ||
1691 | case LABEL_REF: | |
a827d9b1 | 1692 | return LABEL_REF_LABEL (x) == LABEL_REF_LABEL (y); |
ca7fd9cd | 1693 | |
db048faf MM |
1694 | case SYMBOL_REF: |
1695 | return XSTR (x, 0) == XSTR (y, 0); | |
1696 | ||
af6236c1 AO |
1697 | case ENTRY_VALUE: |
1698 | /* This is magic, don't go through canonicalization et al. */ | |
1699 | return rtx_equal_p (ENTRY_VALUE_EXP (x), ENTRY_VALUE_EXP (y)); | |
1700 | ||
40e02b4a | 1701 | case VALUE: |
d8116890 | 1702 | CASE_CONST_UNIQUE: |
807e902e | 1703 | /* Pointer equality guarantees equality for these nodes. */ |
db048faf MM |
1704 | return 0; |
1705 | ||
db048faf MM |
1706 | default: |
1707 | break; | |
1708 | } | |
9ae8ffe7 | 1709 | |
45183e03 JH |
1710 | /* canon_rtx knows how to handle plus. No need to canonicalize. */ |
1711 | if (code == PLUS) | |
9ae8ffe7 JL |
1712 | return ((rtx_equal_for_memref_p (XEXP (x, 0), XEXP (y, 0)) |
1713 | && rtx_equal_for_memref_p (XEXP (x, 1), XEXP (y, 1))) | |
1714 | || (rtx_equal_for_memref_p (XEXP (x, 0), XEXP (y, 1)) | |
1715 | && rtx_equal_for_memref_p (XEXP (x, 1), XEXP (y, 0)))); | |
45183e03 JH |
1716 | /* For commutative operations, the RTX match if the operand match in any |
1717 | order. Also handle the simple binary and unary cases without a loop. */ | |
ec8e098d | 1718 | if (COMMUTATIVE_P (x)) |
45183e03 JH |
1719 | { |
1720 | rtx xop0 = canon_rtx (XEXP (x, 0)); | |
1721 | rtx yop0 = canon_rtx (XEXP (y, 0)); | |
1722 | rtx yop1 = canon_rtx (XEXP (y, 1)); | |
1723 | ||
1724 | return ((rtx_equal_for_memref_p (xop0, yop0) | |
1725 | && rtx_equal_for_memref_p (canon_rtx (XEXP (x, 1)), yop1)) | |
1726 | || (rtx_equal_for_memref_p (xop0, yop1) | |
1727 | && rtx_equal_for_memref_p (canon_rtx (XEXP (x, 1)), yop0))); | |
1728 | } | |
ec8e098d | 1729 | else if (NON_COMMUTATIVE_P (x)) |
45183e03 JH |
1730 | { |
1731 | return (rtx_equal_for_memref_p (canon_rtx (XEXP (x, 0)), | |
4682ae04 | 1732 | canon_rtx (XEXP (y, 0))) |
45183e03 JH |
1733 | && rtx_equal_for_memref_p (canon_rtx (XEXP (x, 1)), |
1734 | canon_rtx (XEXP (y, 1)))); | |
1735 | } | |
ec8e098d | 1736 | else if (UNARY_P (x)) |
45183e03 | 1737 | return rtx_equal_for_memref_p (canon_rtx (XEXP (x, 0)), |
4682ae04 | 1738 | canon_rtx (XEXP (y, 0))); |
9ae8ffe7 JL |
1739 | |
1740 | /* Compare the elements. If any pair of corresponding elements | |
de12be17 JC |
1741 | fail to match, return 0 for the whole things. |
1742 | ||
1743 | Limit cases to types which actually appear in addresses. */ | |
9ae8ffe7 JL |
1744 | |
1745 | fmt = GET_RTX_FORMAT (code); | |
1746 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
1747 | { | |
1748 | switch (fmt[i]) | |
1749 | { | |
9ae8ffe7 JL |
1750 | case 'i': |
1751 | if (XINT (x, i) != XINT (y, i)) | |
1752 | return 0; | |
1753 | break; | |
1754 | ||
9ae8ffe7 JL |
1755 | case 'E': |
1756 | /* Two vectors must have the same length. */ | |
1757 | if (XVECLEN (x, i) != XVECLEN (y, i)) | |
1758 | return 0; | |
1759 | ||
1760 | /* And the corresponding elements must match. */ | |
1761 | for (j = 0; j < XVECLEN (x, i); j++) | |
45183e03 JH |
1762 | if (rtx_equal_for_memref_p (canon_rtx (XVECEXP (x, i, j)), |
1763 | canon_rtx (XVECEXP (y, i, j))) == 0) | |
9ae8ffe7 JL |
1764 | return 0; |
1765 | break; | |
1766 | ||
1767 | case 'e': | |
45183e03 JH |
1768 | if (rtx_equal_for_memref_p (canon_rtx (XEXP (x, i)), |
1769 | canon_rtx (XEXP (y, i))) == 0) | |
9ae8ffe7 JL |
1770 | return 0; |
1771 | break; | |
1772 | ||
3237ac18 AH |
1773 | /* This can happen for asm operands. */ |
1774 | case 's': | |
1775 | if (strcmp (XSTR (x, i), XSTR (y, i))) | |
1776 | return 0; | |
1777 | break; | |
1778 | ||
aee21ba9 JL |
1779 | /* This can happen for an asm which clobbers memory. */ |
1780 | case '0': | |
1781 | break; | |
1782 | ||
9ae8ffe7 JL |
1783 | /* It is believed that rtx's at this level will never |
1784 | contain anything but integers and other rtx's, | |
1785 | except for within LABEL_REFs and SYMBOL_REFs. */ | |
1786 | default: | |
298e6adc | 1787 | gcc_unreachable (); |
9ae8ffe7 JL |
1788 | } |
1789 | } | |
1790 | return 1; | |
1791 | } | |
1792 | ||
9e412ca3 | 1793 | static rtx |
4682ae04 | 1794 | find_base_term (rtx x) |
9ae8ffe7 | 1795 | { |
eab5c70a | 1796 | cselib_val *val; |
6f2ffb4b AO |
1797 | struct elt_loc_list *l, *f; |
1798 | rtx ret; | |
eab5c70a | 1799 | |
b949ea8b JW |
1800 | #if defined (FIND_BASE_TERM) |
1801 | /* Try machine-dependent ways to find the base term. */ | |
1802 | x = FIND_BASE_TERM (x); | |
1803 | #endif | |
1804 | ||
9ae8ffe7 JL |
1805 | switch (GET_CODE (x)) |
1806 | { | |
1807 | case REG: | |
1808 | return REG_BASE_VALUE (x); | |
1809 | ||
d288e53d | 1810 | case TRUNCATE: |
5932a4d4 | 1811 | /* As we do not know which address space the pointer is referring to, we can |
d4ebfa65 BE |
1812 | handle this only if the target does not support different pointer or |
1813 | address modes depending on the address space. */ | |
1814 | if (!target_default_pointer_address_modes_p ()) | |
1815 | return 0; | |
d288e53d | 1816 | if (GET_MODE_SIZE (GET_MODE (x)) < GET_MODE_SIZE (Pmode)) |
ca7fd9cd | 1817 | return 0; |
d288e53d | 1818 | /* Fall through. */ |
9ae8ffe7 | 1819 | case HIGH: |
6d849a2a JL |
1820 | case PRE_INC: |
1821 | case PRE_DEC: | |
1822 | case POST_INC: | |
1823 | case POST_DEC: | |
d288e53d DE |
1824 | case PRE_MODIFY: |
1825 | case POST_MODIFY: | |
6d849a2a JL |
1826 | return find_base_term (XEXP (x, 0)); |
1827 | ||
1abade85 RK |
1828 | case ZERO_EXTEND: |
1829 | case SIGN_EXTEND: /* Used for Alpha/NT pointers */ | |
5932a4d4 | 1830 | /* As we do not know which address space the pointer is referring to, we can |
d4ebfa65 BE |
1831 | handle this only if the target does not support different pointer or |
1832 | address modes depending on the address space. */ | |
1833 | if (!target_default_pointer_address_modes_p ()) | |
1834 | return 0; | |
1835 | ||
1abade85 RK |
1836 | { |
1837 | rtx temp = find_base_term (XEXP (x, 0)); | |
1838 | ||
5ae6cd0d | 1839 | if (temp != 0 && CONSTANT_P (temp)) |
1abade85 | 1840 | temp = convert_memory_address (Pmode, temp); |
1abade85 RK |
1841 | |
1842 | return temp; | |
1843 | } | |
1844 | ||
eab5c70a BS |
1845 | case VALUE: |
1846 | val = CSELIB_VAL_PTR (x); | |
6f2ffb4b AO |
1847 | ret = NULL_RTX; |
1848 | ||
40e02b4a | 1849 | if (!val) |
6f2ffb4b AO |
1850 | return ret; |
1851 | ||
0fe03ac3 JJ |
1852 | if (cselib_sp_based_value_p (val)) |
1853 | return static_reg_base_value[STACK_POINTER_REGNUM]; | |
1854 | ||
6f2ffb4b AO |
1855 | f = val->locs; |
1856 | /* Temporarily reset val->locs to avoid infinite recursion. */ | |
1857 | val->locs = NULL; | |
1858 | ||
1859 | for (l = f; l; l = l->next) | |
1860 | if (GET_CODE (l->loc) == VALUE | |
1861 | && CSELIB_VAL_PTR (l->loc)->locs | |
1862 | && !CSELIB_VAL_PTR (l->loc)->locs->next | |
1863 | && CSELIB_VAL_PTR (l->loc)->locs->loc == x) | |
1864 | continue; | |
1865 | else if ((ret = find_base_term (l->loc)) != 0) | |
1866 | break; | |
1867 | ||
1868 | val->locs = f; | |
1869 | return ret; | |
eab5c70a | 1870 | |
023f059b JJ |
1871 | case LO_SUM: |
1872 | /* The standard form is (lo_sum reg sym) so look only at the | |
1873 | second operand. */ | |
1874 | return find_base_term (XEXP (x, 1)); | |
1875 | ||
9ae8ffe7 JL |
1876 | case CONST: |
1877 | x = XEXP (x, 0); | |
1878 | if (GET_CODE (x) != PLUS && GET_CODE (x) != MINUS) | |
1879 | return 0; | |
938d968e | 1880 | /* Fall through. */ |
9ae8ffe7 JL |
1881 | case PLUS: |
1882 | case MINUS: | |
1883 | { | |
3c567fae JL |
1884 | rtx tmp1 = XEXP (x, 0); |
1885 | rtx tmp2 = XEXP (x, 1); | |
1886 | ||
f5143c46 | 1887 | /* This is a little bit tricky since we have to determine which of |
3c567fae JL |
1888 | the two operands represents the real base address. Otherwise this |
1889 | routine may return the index register instead of the base register. | |
1890 | ||
1891 | That may cause us to believe no aliasing was possible, when in | |
1892 | fact aliasing is possible. | |
1893 | ||
1894 | We use a few simple tests to guess the base register. Additional | |
1895 | tests can certainly be added. For example, if one of the operands | |
1896 | is a shift or multiply, then it must be the index register and the | |
1897 | other operand is the base register. */ | |
ca7fd9cd | 1898 | |
b949ea8b JW |
1899 | if (tmp1 == pic_offset_table_rtx && CONSTANT_P (tmp2)) |
1900 | return find_base_term (tmp2); | |
1901 | ||
31b0a960 | 1902 | /* If either operand is known to be a pointer, then prefer it |
3c567fae | 1903 | to determine the base term. */ |
3502dc9c | 1904 | if (REG_P (tmp1) && REG_POINTER (tmp1)) |
31b0a960 RB |
1905 | ; |
1906 | else if (REG_P (tmp2) && REG_POINTER (tmp2)) | |
a7c75343 JJ |
1907 | std::swap (tmp1, tmp2); |
1908 | /* If second argument is constant which has base term, prefer it | |
1909 | over variable tmp1. See PR64025. */ | |
1910 | else if (CONSTANT_P (tmp2) && !CONST_INT_P (tmp2)) | |
1911 | std::swap (tmp1, tmp2); | |
3c567fae | 1912 | |
31b0a960 RB |
1913 | /* Go ahead and find the base term for both operands. If either base |
1914 | term is from a pointer or is a named object or a special address | |
3c567fae JL |
1915 | (like an argument or stack reference), then use it for the |
1916 | base term. */ | |
481be1c4 RB |
1917 | rtx base = find_base_term (tmp1); |
1918 | if (base != NULL_RTX | |
31b0a960 | 1919 | && ((REG_P (tmp1) && REG_POINTER (tmp1)) |
481be1c4 RB |
1920 | || known_base_value_p (base))) |
1921 | return base; | |
1922 | base = find_base_term (tmp2); | |
1923 | if (base != NULL_RTX | |
31b0a960 | 1924 | && ((REG_P (tmp2) && REG_POINTER (tmp2)) |
481be1c4 RB |
1925 | || known_base_value_p (base))) |
1926 | return base; | |
3c567fae JL |
1927 | |
1928 | /* We could not determine which of the two operands was the | |
1929 | base register and which was the index. So we can determine | |
1930 | nothing from the base alias check. */ | |
1931 | return 0; | |
9ae8ffe7 JL |
1932 | } |
1933 | ||
1934 | case AND: | |
481683e1 | 1935 | if (CONST_INT_P (XEXP (x, 1)) && INTVAL (XEXP (x, 1)) != 0) |
d288e53d | 1936 | return find_base_term (XEXP (x, 0)); |
9ae8ffe7 JL |
1937 | return 0; |
1938 | ||
1939 | case SYMBOL_REF: | |
1940 | case LABEL_REF: | |
1941 | return x; | |
1942 | ||
1943 | default: | |
1944 | return 0; | |
1945 | } | |
1946 | } | |
1947 | ||
9e412ca3 RS |
1948 | /* Return true if accesses to address X may alias accesses based |
1949 | on the stack pointer. */ | |
1950 | ||
1951 | bool | |
1952 | may_be_sp_based_p (rtx x) | |
1953 | { | |
1954 | rtx base = find_base_term (x); | |
1955 | return !base || base == static_reg_base_value[STACK_POINTER_REGNUM]; | |
1956 | } | |
1957 | ||
9ae8ffe7 JL |
1958 | /* Return 0 if the addresses X and Y are known to point to different |
1959 | objects, 1 if they might be pointers to the same object. */ | |
1960 | ||
1961 | static int | |
31b0a960 | 1962 | base_alias_check (rtx x, rtx x_base, rtx y, rtx y_base, |
ef4bddc2 | 1963 | machine_mode x_mode, machine_mode y_mode) |
9ae8ffe7 | 1964 | { |
1c72c7f6 JC |
1965 | /* If the address itself has no known base see if a known equivalent |
1966 | value has one. If either address still has no known base, nothing | |
1967 | is known about aliasing. */ | |
1968 | if (x_base == 0) | |
1969 | { | |
1970 | rtx x_c; | |
d4b60170 | 1971 | |
1c72c7f6 JC |
1972 | if (! flag_expensive_optimizations || (x_c = canon_rtx (x)) == x) |
1973 | return 1; | |
d4b60170 | 1974 | |
1c72c7f6 JC |
1975 | x_base = find_base_term (x_c); |
1976 | if (x_base == 0) | |
1977 | return 1; | |
1978 | } | |
9ae8ffe7 | 1979 | |
1c72c7f6 JC |
1980 | if (y_base == 0) |
1981 | { | |
1982 | rtx y_c; | |
1983 | if (! flag_expensive_optimizations || (y_c = canon_rtx (y)) == y) | |
1984 | return 1; | |
d4b60170 | 1985 | |
1c72c7f6 JC |
1986 | y_base = find_base_term (y_c); |
1987 | if (y_base == 0) | |
1988 | return 1; | |
1989 | } | |
1990 | ||
1991 | /* If the base addresses are equal nothing is known about aliasing. */ | |
1992 | if (rtx_equal_p (x_base, y_base)) | |
9ae8ffe7 JL |
1993 | return 1; |
1994 | ||
435da628 UB |
1995 | /* The base addresses are different expressions. If they are not accessed |
1996 | via AND, there is no conflict. We can bring knowledge of object | |
1997 | alignment into play here. For example, on alpha, "char a, b;" can | |
1998 | alias one another, though "char a; long b;" cannot. AND addesses may | |
1999 | implicitly alias surrounding objects; i.e. unaligned access in DImode | |
2000 | via AND address can alias all surrounding object types except those | |
2001 | with aligment 8 or higher. */ | |
2002 | if (GET_CODE (x) == AND && GET_CODE (y) == AND) | |
2003 | return 1; | |
2004 | if (GET_CODE (x) == AND | |
481683e1 | 2005 | && (!CONST_INT_P (XEXP (x, 1)) |
435da628 UB |
2006 | || (int) GET_MODE_UNIT_SIZE (y_mode) < -INTVAL (XEXP (x, 1)))) |
2007 | return 1; | |
2008 | if (GET_CODE (y) == AND | |
481683e1 | 2009 | && (!CONST_INT_P (XEXP (y, 1)) |
435da628 UB |
2010 | || (int) GET_MODE_UNIT_SIZE (x_mode) < -INTVAL (XEXP (y, 1)))) |
2011 | return 1; | |
2012 | ||
2013 | /* Differing symbols not accessed via AND never alias. */ | |
9ae8ffe7 | 2014 | if (GET_CODE (x_base) != ADDRESS && GET_CODE (y_base) != ADDRESS) |
435da628 | 2015 | return 0; |
9ae8ffe7 | 2016 | |
9fc37b2b | 2017 | if (unique_base_value_p (x_base) || unique_base_value_p (y_base)) |
9ae8ffe7 JL |
2018 | return 0; |
2019 | ||
0d3c82d6 | 2020 | return 1; |
9ae8ffe7 JL |
2021 | } |
2022 | ||
a5628378 AO |
2023 | /* Return TRUE if EXPR refers to a VALUE whose uid is greater than |
2024 | that of V. */ | |
2025 | ||
2026 | static bool | |
403837b4 | 2027 | refs_newer_value_p (const_rtx expr, rtx v) |
a5628378 AO |
2028 | { |
2029 | int minuid = CSELIB_VAL_PTR (v)->uid; | |
403837b4 RS |
2030 | subrtx_iterator::array_type array; |
2031 | FOR_EACH_SUBRTX (iter, array, expr, NONCONST) | |
2032 | if (GET_CODE (*iter) == VALUE && CSELIB_VAL_PTR (*iter)->uid > minuid) | |
2033 | return true; | |
2034 | return false; | |
a5628378 AO |
2035 | } |
2036 | ||
eab5c70a BS |
2037 | /* Convert the address X into something we can use. This is done by returning |
2038 | it unchanged unless it is a value; in the latter case we call cselib to get | |
2039 | a more useful rtx. */ | |
3bdf5ad1 | 2040 | |
a13d4ebf | 2041 | rtx |
4682ae04 | 2042 | get_addr (rtx x) |
eab5c70a BS |
2043 | { |
2044 | cselib_val *v; | |
2045 | struct elt_loc_list *l; | |
2046 | ||
2047 | if (GET_CODE (x) != VALUE) | |
2048 | return x; | |
2049 | v = CSELIB_VAL_PTR (x); | |
40e02b4a JH |
2050 | if (v) |
2051 | { | |
0f68ba3e AO |
2052 | bool have_equivs = cselib_have_permanent_equivalences (); |
2053 | if (have_equivs) | |
2054 | v = canonical_cselib_val (v); | |
40e02b4a JH |
2055 | for (l = v->locs; l; l = l->next) |
2056 | if (CONSTANT_P (l->loc)) | |
2057 | return l->loc; | |
2058 | for (l = v->locs; l; l = l->next) | |
0f68ba3e AO |
2059 | if (!REG_P (l->loc) && !MEM_P (l->loc) |
2060 | /* Avoid infinite recursion when potentially dealing with | |
2061 | var-tracking artificial equivalences, by skipping the | |
2062 | equivalences themselves, and not choosing expressions | |
2063 | that refer to newer VALUEs. */ | |
2064 | && (!have_equivs | |
2065 | || (GET_CODE (l->loc) != VALUE | |
2066 | && !refs_newer_value_p (l->loc, x)))) | |
a5628378 | 2067 | return l->loc; |
0f68ba3e AO |
2068 | if (have_equivs) |
2069 | { | |
2070 | for (l = v->locs; l; l = l->next) | |
2071 | if (REG_P (l->loc) | |
2072 | || (GET_CODE (l->loc) != VALUE | |
2073 | && !refs_newer_value_p (l->loc, x))) | |
2074 | return l->loc; | |
2075 | /* Return the canonical value. */ | |
2076 | return v->val_rtx; | |
2077 | } | |
2078 | if (v->locs) | |
2079 | return v->locs->loc; | |
40e02b4a | 2080 | } |
eab5c70a BS |
2081 | return x; |
2082 | } | |
2083 | ||
39cec1ac MH |
2084 | /* Return the address of the (N_REFS + 1)th memory reference to ADDR |
2085 | where SIZE is the size in bytes of the memory reference. If ADDR | |
2086 | is not modified by the memory reference then ADDR is returned. */ | |
2087 | ||
04e2b4d3 | 2088 | static rtx |
4682ae04 | 2089 | addr_side_effect_eval (rtx addr, int size, int n_refs) |
39cec1ac MH |
2090 | { |
2091 | int offset = 0; | |
ca7fd9cd | 2092 | |
39cec1ac MH |
2093 | switch (GET_CODE (addr)) |
2094 | { | |
2095 | case PRE_INC: | |
2096 | offset = (n_refs + 1) * size; | |
2097 | break; | |
2098 | case PRE_DEC: | |
2099 | offset = -(n_refs + 1) * size; | |
2100 | break; | |
2101 | case POST_INC: | |
2102 | offset = n_refs * size; | |
2103 | break; | |
2104 | case POST_DEC: | |
2105 | offset = -n_refs * size; | |
2106 | break; | |
2107 | ||
2108 | default: | |
2109 | return addr; | |
2110 | } | |
ca7fd9cd | 2111 | |
39cec1ac | 2112 | if (offset) |
45183e03 | 2113 | addr = gen_rtx_PLUS (GET_MODE (addr), XEXP (addr, 0), |
4789c0ce | 2114 | gen_int_mode (offset, GET_MODE (addr))); |
39cec1ac MH |
2115 | else |
2116 | addr = XEXP (addr, 0); | |
45183e03 | 2117 | addr = canon_rtx (addr); |
39cec1ac MH |
2118 | |
2119 | return addr; | |
2120 | } | |
2121 | ||
3aa03517 AO |
2122 | /* Return TRUE if an object X sized at XSIZE bytes and another object |
2123 | Y sized at YSIZE bytes, starting C bytes after X, may overlap. If | |
2124 | any of the sizes is zero, assume an overlap, otherwise use the | |
2125 | absolute value of the sizes as the actual sizes. */ | |
2126 | ||
2127 | static inline bool | |
2128 | offset_overlap_p (HOST_WIDE_INT c, int xsize, int ysize) | |
2129 | { | |
2130 | return (xsize == 0 || ysize == 0 | |
2131 | || (c >= 0 | |
2132 | ? (abs (xsize) > c) | |
2133 | : (abs (ysize) > -c))); | |
2134 | } | |
2135 | ||
f47e08d9 RG |
2136 | /* Return one if X and Y (memory addresses) reference the |
2137 | same location in memory or if the references overlap. | |
2138 | Return zero if they do not overlap, else return | |
2139 | minus one in which case they still might reference the same location. | |
2140 | ||
2141 | C is an offset accumulator. When | |
9ae8ffe7 JL |
2142 | C is nonzero, we are testing aliases between X and Y + C. |
2143 | XSIZE is the size in bytes of the X reference, | |
2144 | similarly YSIZE is the size in bytes for Y. | |
45183e03 | 2145 | Expect that canon_rtx has been already called for X and Y. |
9ae8ffe7 JL |
2146 | |
2147 | If XSIZE or YSIZE is zero, we do not know the amount of memory being | |
2148 | referenced (the reference was BLKmode), so make the most pessimistic | |
2149 | assumptions. | |
2150 | ||
c02f035f RH |
2151 | If XSIZE or YSIZE is negative, we may access memory outside the object |
2152 | being referenced as a side effect. This can happen when using AND to | |
2153 | align memory references, as is done on the Alpha. | |
2154 | ||
9ae8ffe7 | 2155 | Nice to notice that varying addresses cannot conflict with fp if no |
f47e08d9 RG |
2156 | local variables had their addresses taken, but that's too hard now. |
2157 | ||
2158 | ??? Contrary to the tree alias oracle this does not return | |
2159 | one for X + non-constant and Y + non-constant when X and Y are equal. | |
2160 | If that is fixed the TBAA hack for union type-punning can be removed. */ | |
9ae8ffe7 | 2161 | |
9ae8ffe7 | 2162 | static int |
4682ae04 | 2163 | memrefs_conflict_p (int xsize, rtx x, int ysize, rtx y, HOST_WIDE_INT c) |
9ae8ffe7 | 2164 | { |
eab5c70a | 2165 | if (GET_CODE (x) == VALUE) |
5312b066 JJ |
2166 | { |
2167 | if (REG_P (y)) | |
2168 | { | |
24f8d71e JJ |
2169 | struct elt_loc_list *l = NULL; |
2170 | if (CSELIB_VAL_PTR (x)) | |
a5628378 AO |
2171 | for (l = canonical_cselib_val (CSELIB_VAL_PTR (x))->locs; |
2172 | l; l = l->next) | |
24f8d71e JJ |
2173 | if (REG_P (l->loc) && rtx_equal_for_memref_p (l->loc, y)) |
2174 | break; | |
5312b066 JJ |
2175 | if (l) |
2176 | x = y; | |
2177 | else | |
2178 | x = get_addr (x); | |
2179 | } | |
2180 | /* Don't call get_addr if y is the same VALUE. */ | |
2181 | else if (x != y) | |
2182 | x = get_addr (x); | |
2183 | } | |
eab5c70a | 2184 | if (GET_CODE (y) == VALUE) |
5312b066 JJ |
2185 | { |
2186 | if (REG_P (x)) | |
2187 | { | |
24f8d71e JJ |
2188 | struct elt_loc_list *l = NULL; |
2189 | if (CSELIB_VAL_PTR (y)) | |
a5628378 AO |
2190 | for (l = canonical_cselib_val (CSELIB_VAL_PTR (y))->locs; |
2191 | l; l = l->next) | |
24f8d71e JJ |
2192 | if (REG_P (l->loc) && rtx_equal_for_memref_p (l->loc, x)) |
2193 | break; | |
5312b066 JJ |
2194 | if (l) |
2195 | y = x; | |
2196 | else | |
2197 | y = get_addr (y); | |
2198 | } | |
2199 | /* Don't call get_addr if x is the same VALUE. */ | |
2200 | else if (y != x) | |
2201 | y = get_addr (y); | |
2202 | } | |
9ae8ffe7 JL |
2203 | if (GET_CODE (x) == HIGH) |
2204 | x = XEXP (x, 0); | |
2205 | else if (GET_CODE (x) == LO_SUM) | |
2206 | x = XEXP (x, 1); | |
2207 | else | |
3aa03517 | 2208 | x = addr_side_effect_eval (x, abs (xsize), 0); |
9ae8ffe7 JL |
2209 | if (GET_CODE (y) == HIGH) |
2210 | y = XEXP (y, 0); | |
2211 | else if (GET_CODE (y) == LO_SUM) | |
2212 | y = XEXP (y, 1); | |
2213 | else | |
3aa03517 | 2214 | y = addr_side_effect_eval (y, abs (ysize), 0); |
9ae8ffe7 JL |
2215 | |
2216 | if (rtx_equal_for_memref_p (x, y)) | |
2217 | { | |
3aa03517 | 2218 | return offset_overlap_p (c, xsize, ysize); |
9ae8ffe7 JL |
2219 | } |
2220 | ||
6e73e666 JC |
2221 | /* This code used to check for conflicts involving stack references and |
2222 | globals but the base address alias code now handles these cases. */ | |
9ae8ffe7 JL |
2223 | |
2224 | if (GET_CODE (x) == PLUS) | |
2225 | { | |
2226 | /* The fact that X is canonicalized means that this | |
2227 | PLUS rtx is canonicalized. */ | |
2228 | rtx x0 = XEXP (x, 0); | |
2229 | rtx x1 = XEXP (x, 1); | |
2230 | ||
2231 | if (GET_CODE (y) == PLUS) | |
2232 | { | |
2233 | /* The fact that Y is canonicalized means that this | |
2234 | PLUS rtx is canonicalized. */ | |
2235 | rtx y0 = XEXP (y, 0); | |
2236 | rtx y1 = XEXP (y, 1); | |
2237 | ||
2238 | if (rtx_equal_for_memref_p (x1, y1)) | |
2239 | return memrefs_conflict_p (xsize, x0, ysize, y0, c); | |
2240 | if (rtx_equal_for_memref_p (x0, y0)) | |
2241 | return memrefs_conflict_p (xsize, x1, ysize, y1, c); | |
481683e1 | 2242 | if (CONST_INT_P (x1)) |
63be02db | 2243 | { |
481683e1 | 2244 | if (CONST_INT_P (y1)) |
63be02db JM |
2245 | return memrefs_conflict_p (xsize, x0, ysize, y0, |
2246 | c - INTVAL (x1) + INTVAL (y1)); | |
2247 | else | |
2248 | return memrefs_conflict_p (xsize, x0, ysize, y, | |
2249 | c - INTVAL (x1)); | |
2250 | } | |
481683e1 | 2251 | else if (CONST_INT_P (y1)) |
9ae8ffe7 JL |
2252 | return memrefs_conflict_p (xsize, x, ysize, y0, c + INTVAL (y1)); |
2253 | ||
f47e08d9 | 2254 | return -1; |
9ae8ffe7 | 2255 | } |
481683e1 | 2256 | else if (CONST_INT_P (x1)) |
9ae8ffe7 JL |
2257 | return memrefs_conflict_p (xsize, x0, ysize, y, c - INTVAL (x1)); |
2258 | } | |
2259 | else if (GET_CODE (y) == PLUS) | |
2260 | { | |
2261 | /* The fact that Y is canonicalized means that this | |
2262 | PLUS rtx is canonicalized. */ | |
2263 | rtx y0 = XEXP (y, 0); | |
2264 | rtx y1 = XEXP (y, 1); | |
2265 | ||
481683e1 | 2266 | if (CONST_INT_P (y1)) |
9ae8ffe7 JL |
2267 | return memrefs_conflict_p (xsize, x, ysize, y0, c + INTVAL (y1)); |
2268 | else | |
f47e08d9 | 2269 | return -1; |
9ae8ffe7 JL |
2270 | } |
2271 | ||
2272 | if (GET_CODE (x) == GET_CODE (y)) | |
2273 | switch (GET_CODE (x)) | |
2274 | { | |
2275 | case MULT: | |
2276 | { | |
2277 | /* Handle cases where we expect the second operands to be the | |
2278 | same, and check only whether the first operand would conflict | |
2279 | or not. */ | |
2280 | rtx x0, y0; | |
2281 | rtx x1 = canon_rtx (XEXP (x, 1)); | |
2282 | rtx y1 = canon_rtx (XEXP (y, 1)); | |
2283 | if (! rtx_equal_for_memref_p (x1, y1)) | |
f47e08d9 | 2284 | return -1; |
9ae8ffe7 JL |
2285 | x0 = canon_rtx (XEXP (x, 0)); |
2286 | y0 = canon_rtx (XEXP (y, 0)); | |
2287 | if (rtx_equal_for_memref_p (x0, y0)) | |
3aa03517 | 2288 | return offset_overlap_p (c, xsize, ysize); |
9ae8ffe7 JL |
2289 | |
2290 | /* Can't properly adjust our sizes. */ | |
481683e1 | 2291 | if (!CONST_INT_P (x1)) |
f47e08d9 | 2292 | return -1; |
9ae8ffe7 JL |
2293 | xsize /= INTVAL (x1); |
2294 | ysize /= INTVAL (x1); | |
2295 | c /= INTVAL (x1); | |
2296 | return memrefs_conflict_p (xsize, x0, ysize, y0, c); | |
2297 | } | |
1d300e19 KG |
2298 | |
2299 | default: | |
2300 | break; | |
9ae8ffe7 JL |
2301 | } |
2302 | ||
a9bf4fe2 AO |
2303 | /* Deal with alignment ANDs by adjusting offset and size so as to |
2304 | cover the maximum range, without taking any previously known | |
5147bf6a AO |
2305 | alignment into account. Make a size negative after such an |
2306 | adjustments, so that, if we end up with e.g. two SYMBOL_REFs, we | |
2307 | assume a potential overlap, because they may end up in contiguous | |
2308 | memory locations and the stricter-alignment access may span over | |
2309 | part of both. */ | |
481683e1 | 2310 | if (GET_CODE (x) == AND && CONST_INT_P (XEXP (x, 1))) |
56ee9281 | 2311 | { |
a9bf4fe2 AO |
2312 | HOST_WIDE_INT sc = INTVAL (XEXP (x, 1)); |
2313 | unsigned HOST_WIDE_INT uc = sc; | |
5147bf6a | 2314 | if (sc < 0 && -uc == (uc & -uc)) |
a9bf4fe2 | 2315 | { |
5147bf6a AO |
2316 | if (xsize > 0) |
2317 | xsize = -xsize; | |
3aa03517 AO |
2318 | if (xsize) |
2319 | xsize += sc + 1; | |
fe8fb1c4 | 2320 | c -= sc + 1; |
a9bf4fe2 AO |
2321 | return memrefs_conflict_p (xsize, canon_rtx (XEXP (x, 0)), |
2322 | ysize, y, c); | |
2323 | } | |
56ee9281 | 2324 | } |
481683e1 | 2325 | if (GET_CODE (y) == AND && CONST_INT_P (XEXP (y, 1))) |
c02f035f | 2326 | { |
a9bf4fe2 AO |
2327 | HOST_WIDE_INT sc = INTVAL (XEXP (y, 1)); |
2328 | unsigned HOST_WIDE_INT uc = sc; | |
5147bf6a | 2329 | if (sc < 0 && -uc == (uc & -uc)) |
a9bf4fe2 | 2330 | { |
5147bf6a AO |
2331 | if (ysize > 0) |
2332 | ysize = -ysize; | |
3aa03517 AO |
2333 | if (ysize) |
2334 | ysize += sc + 1; | |
fe8fb1c4 | 2335 | c += sc + 1; |
a9bf4fe2 AO |
2336 | return memrefs_conflict_p (xsize, x, |
2337 | ysize, canon_rtx (XEXP (y, 0)), c); | |
2338 | } | |
c02f035f | 2339 | } |
9ae8ffe7 JL |
2340 | |
2341 | if (CONSTANT_P (x)) | |
2342 | { | |
481683e1 | 2343 | if (CONST_INT_P (x) && CONST_INT_P (y)) |
9ae8ffe7 JL |
2344 | { |
2345 | c += (INTVAL (y) - INTVAL (x)); | |
3aa03517 | 2346 | return offset_overlap_p (c, xsize, ysize); |
9ae8ffe7 JL |
2347 | } |
2348 | ||
2349 | if (GET_CODE (x) == CONST) | |
2350 | { | |
2351 | if (GET_CODE (y) == CONST) | |
2352 | return memrefs_conflict_p (xsize, canon_rtx (XEXP (x, 0)), | |
2353 | ysize, canon_rtx (XEXP (y, 0)), c); | |
2354 | else | |
2355 | return memrefs_conflict_p (xsize, canon_rtx (XEXP (x, 0)), | |
2356 | ysize, y, c); | |
2357 | } | |
2358 | if (GET_CODE (y) == CONST) | |
2359 | return memrefs_conflict_p (xsize, x, ysize, | |
2360 | canon_rtx (XEXP (y, 0)), c); | |
2361 | ||
3aa03517 AO |
2362 | /* Assume a potential overlap for symbolic addresses that went |
2363 | through alignment adjustments (i.e., that have negative | |
2364 | sizes), because we can't know how far they are from each | |
2365 | other. */ | |
9ae8ffe7 | 2366 | if (CONSTANT_P (y)) |
3aa03517 | 2367 | return (xsize < 0 || ysize < 0 || offset_overlap_p (c, xsize, ysize)); |
9ae8ffe7 | 2368 | |
f47e08d9 | 2369 | return -1; |
9ae8ffe7 | 2370 | } |
f47e08d9 RG |
2371 | |
2372 | return -1; | |
9ae8ffe7 JL |
2373 | } |
2374 | ||
2375 | /* Functions to compute memory dependencies. | |
2376 | ||
2377 | Since we process the insns in execution order, we can build tables | |
2378 | to keep track of what registers are fixed (and not aliased), what registers | |
2379 | are varying in known ways, and what registers are varying in unknown | |
2380 | ways. | |
2381 | ||
2382 | If both memory references are volatile, then there must always be a | |
2383 | dependence between the two references, since their order can not be | |
2384 | changed. A volatile and non-volatile reference can be interchanged | |
ca7fd9cd | 2385 | though. |
9ae8ffe7 | 2386 | |
53d9622b RS |
2387 | We also must allow AND addresses, because they may generate accesses |
2388 | outside the object being referenced. This is used to generate aligned | |
2389 | addresses from unaligned addresses, for instance, the alpha | |
dc1618bc | 2390 | storeqi_unaligned pattern. */ |
9ae8ffe7 JL |
2391 | |
2392 | /* Read dependence: X is read after read in MEM takes place. There can | |
96672a3e RH |
2393 | only be a dependence here if both reads are volatile, or if either is |
2394 | an explicit barrier. */ | |
9ae8ffe7 JL |
2395 | |
2396 | int | |
4f588890 | 2397 | read_dependence (const_rtx mem, const_rtx x) |
9ae8ffe7 | 2398 | { |
96672a3e RH |
2399 | if (MEM_VOLATILE_P (x) && MEM_VOLATILE_P (mem)) |
2400 | return true; | |
2401 | if (MEM_ALIAS_SET (x) == ALIAS_SET_MEMORY_BARRIER | |
2402 | || MEM_ALIAS_SET (mem) == ALIAS_SET_MEMORY_BARRIER) | |
2403 | return true; | |
2404 | return false; | |
9ae8ffe7 JL |
2405 | } |
2406 | ||
998d7deb RH |
2407 | /* Look at the bottom of the COMPONENT_REF list for a DECL, and return it. */ |
2408 | ||
2409 | static tree | |
4682ae04 | 2410 | decl_for_component_ref (tree x) |
998d7deb RH |
2411 | { |
2412 | do | |
2413 | { | |
2414 | x = TREE_OPERAND (x, 0); | |
2415 | } | |
2416 | while (x && TREE_CODE (x) == COMPONENT_REF); | |
2417 | ||
2418 | return x && DECL_P (x) ? x : NULL_TREE; | |
2419 | } | |
2420 | ||
527210c4 RS |
2421 | /* Walk up the COMPONENT_REF list in X and adjust *OFFSET to compensate |
2422 | for the offset of the field reference. *KNOWN_P says whether the | |
2423 | offset is known. */ | |
998d7deb | 2424 | |
527210c4 RS |
2425 | static void |
2426 | adjust_offset_for_component_ref (tree x, bool *known_p, | |
2427 | HOST_WIDE_INT *offset) | |
998d7deb | 2428 | { |
527210c4 RS |
2429 | if (!*known_p) |
2430 | return; | |
ca7fd9cd | 2431 | do |
998d7deb | 2432 | { |
527210c4 | 2433 | tree xoffset = component_ref_field_offset (x); |
998d7deb | 2434 | tree field = TREE_OPERAND (x, 1); |
807e902e KZ |
2435 | if (TREE_CODE (xoffset) != INTEGER_CST) |
2436 | { | |
2437 | *known_p = false; | |
2438 | return; | |
2439 | } | |
998d7deb | 2440 | |
807e902e KZ |
2441 | offset_int woffset |
2442 | = (wi::to_offset (xoffset) | |
2443 | + wi::lrshift (wi::to_offset (DECL_FIELD_BIT_OFFSET (field)), | |
2444 | LOG2_BITS_PER_UNIT)); | |
2445 | if (!wi::fits_uhwi_p (woffset)) | |
527210c4 RS |
2446 | { |
2447 | *known_p = false; | |
2448 | return; | |
2449 | } | |
807e902e | 2450 | *offset += woffset.to_uhwi (); |
998d7deb RH |
2451 | |
2452 | x = TREE_OPERAND (x, 0); | |
2453 | } | |
2454 | while (x && TREE_CODE (x) == COMPONENT_REF); | |
998d7deb RH |
2455 | } |
2456 | ||
95bd1dd7 | 2457 | /* Return nonzero if we can determine the exprs corresponding to memrefs |
c6ea834c BM |
2458 | X and Y and they do not overlap. |
2459 | If LOOP_VARIANT is set, skip offset-based disambiguation */ | |
a4311dfe | 2460 | |
2e4e39f6 | 2461 | int |
c6ea834c | 2462 | nonoverlapping_memrefs_p (const_rtx x, const_rtx y, bool loop_invariant) |
a4311dfe | 2463 | { |
998d7deb | 2464 | tree exprx = MEM_EXPR (x), expry = MEM_EXPR (y); |
a4311dfe RK |
2465 | rtx rtlx, rtly; |
2466 | rtx basex, basey; | |
527210c4 RS |
2467 | bool moffsetx_known_p, moffsety_known_p; |
2468 | HOST_WIDE_INT moffsetx = 0, moffsety = 0; | |
a4311dfe RK |
2469 | HOST_WIDE_INT offsetx = 0, offsety = 0, sizex, sizey, tem; |
2470 | ||
998d7deb RH |
2471 | /* Unless both have exprs, we can't tell anything. */ |
2472 | if (exprx == 0 || expry == 0) | |
2473 | return 0; | |
2b22e382 RG |
2474 | |
2475 | /* For spill-slot accesses make sure we have valid offsets. */ | |
2476 | if ((exprx == get_spill_slot_decl (false) | |
527210c4 | 2477 | && ! MEM_OFFSET_KNOWN_P (x)) |
2b22e382 | 2478 | || (expry == get_spill_slot_decl (false) |
527210c4 | 2479 | && ! MEM_OFFSET_KNOWN_P (y))) |
2b22e382 | 2480 | return 0; |
c22cacf3 | 2481 | |
998d7deb | 2482 | /* If the field reference test failed, look at the DECLs involved. */ |
527210c4 RS |
2483 | moffsetx_known_p = MEM_OFFSET_KNOWN_P (x); |
2484 | if (moffsetx_known_p) | |
2485 | moffsetx = MEM_OFFSET (x); | |
998d7deb RH |
2486 | if (TREE_CODE (exprx) == COMPONENT_REF) |
2487 | { | |
2e0c984c RG |
2488 | tree t = decl_for_component_ref (exprx); |
2489 | if (! t) | |
2490 | return 0; | |
527210c4 | 2491 | adjust_offset_for_component_ref (exprx, &moffsetx_known_p, &moffsetx); |
2e0c984c | 2492 | exprx = t; |
998d7deb | 2493 | } |
c67a1cf6 | 2494 | |
527210c4 RS |
2495 | moffsety_known_p = MEM_OFFSET_KNOWN_P (y); |
2496 | if (moffsety_known_p) | |
2497 | moffsety = MEM_OFFSET (y); | |
998d7deb RH |
2498 | if (TREE_CODE (expry) == COMPONENT_REF) |
2499 | { | |
2e0c984c RG |
2500 | tree t = decl_for_component_ref (expry); |
2501 | if (! t) | |
2502 | return 0; | |
527210c4 | 2503 | adjust_offset_for_component_ref (expry, &moffsety_known_p, &moffsety); |
2e0c984c | 2504 | expry = t; |
998d7deb RH |
2505 | } |
2506 | ||
2507 | if (! DECL_P (exprx) || ! DECL_P (expry)) | |
a4311dfe RK |
2508 | return 0; |
2509 | ||
1307c758 RG |
2510 | /* With invalid code we can end up storing into the constant pool. |
2511 | Bail out to avoid ICEing when creating RTL for this. | |
2512 | See gfortran.dg/lto/20091028-2_0.f90. */ | |
2513 | if (TREE_CODE (exprx) == CONST_DECL | |
2514 | || TREE_CODE (expry) == CONST_DECL) | |
2515 | return 1; | |
2516 | ||
998d7deb RH |
2517 | rtlx = DECL_RTL (exprx); |
2518 | rtly = DECL_RTL (expry); | |
a4311dfe | 2519 | |
1edcd60b RK |
2520 | /* If either RTL is not a MEM, it must be a REG or CONCAT, meaning they |
2521 | can't overlap unless they are the same because we never reuse that part | |
2522 | of the stack frame used for locals for spilled pseudos. */ | |
3c0cb5de | 2523 | if ((!MEM_P (rtlx) || !MEM_P (rtly)) |
1edcd60b | 2524 | && ! rtx_equal_p (rtlx, rtly)) |
a4311dfe RK |
2525 | return 1; |
2526 | ||
5932a4d4 | 2527 | /* If we have MEMs referring to different address spaces (which can |
09e881c9 BE |
2528 | potentially overlap), we cannot easily tell from the addresses |
2529 | whether the references overlap. */ | |
2530 | if (MEM_P (rtlx) && MEM_P (rtly) | |
2531 | && MEM_ADDR_SPACE (rtlx) != MEM_ADDR_SPACE (rtly)) | |
2532 | return 0; | |
2533 | ||
a4311dfe RK |
2534 | /* Get the base and offsets of both decls. If either is a register, we |
2535 | know both are and are the same, so use that as the base. The only | |
2536 | we can avoid overlap is if we can deduce that they are nonoverlapping | |
2537 | pieces of that decl, which is very rare. */ | |
3c0cb5de | 2538 | basex = MEM_P (rtlx) ? XEXP (rtlx, 0) : rtlx; |
481683e1 | 2539 | if (GET_CODE (basex) == PLUS && CONST_INT_P (XEXP (basex, 1))) |
a4311dfe RK |
2540 | offsetx = INTVAL (XEXP (basex, 1)), basex = XEXP (basex, 0); |
2541 | ||
3c0cb5de | 2542 | basey = MEM_P (rtly) ? XEXP (rtly, 0) : rtly; |
481683e1 | 2543 | if (GET_CODE (basey) == PLUS && CONST_INT_P (XEXP (basey, 1))) |
a4311dfe RK |
2544 | offsety = INTVAL (XEXP (basey, 1)), basey = XEXP (basey, 0); |
2545 | ||
d746694a | 2546 | /* If the bases are different, we know they do not overlap if both |
ca7fd9cd | 2547 | are constants or if one is a constant and the other a pointer into the |
d746694a RK |
2548 | stack frame. Otherwise a different base means we can't tell if they |
2549 | overlap or not. */ | |
2550 | if (! rtx_equal_p (basex, basey)) | |
ca7fd9cd KH |
2551 | return ((CONSTANT_P (basex) && CONSTANT_P (basey)) |
2552 | || (CONSTANT_P (basex) && REG_P (basey) | |
2553 | && REGNO_PTR_FRAME_P (REGNO (basey))) | |
2554 | || (CONSTANT_P (basey) && REG_P (basex) | |
2555 | && REGNO_PTR_FRAME_P (REGNO (basex)))); | |
a4311dfe | 2556 | |
c6ea834c BM |
2557 | /* Offset based disambiguation not appropriate for loop invariant */ |
2558 | if (loop_invariant) | |
2559 | return 0; | |
2560 | ||
3c0cb5de | 2561 | sizex = (!MEM_P (rtlx) ? (int) GET_MODE_SIZE (GET_MODE (rtlx)) |
f5541398 | 2562 | : MEM_SIZE_KNOWN_P (rtlx) ? MEM_SIZE (rtlx) |
a4311dfe | 2563 | : -1); |
3c0cb5de | 2564 | sizey = (!MEM_P (rtly) ? (int) GET_MODE_SIZE (GET_MODE (rtly)) |
f5541398 RS |
2565 | : MEM_SIZE_KNOWN_P (rtly) ? MEM_SIZE (rtly) |
2566 | : -1); | |
a4311dfe | 2567 | |
0af5bc3e RK |
2568 | /* If we have an offset for either memref, it can update the values computed |
2569 | above. */ | |
527210c4 RS |
2570 | if (moffsetx_known_p) |
2571 | offsetx += moffsetx, sizex -= moffsetx; | |
2572 | if (moffsety_known_p) | |
2573 | offsety += moffsety, sizey -= moffsety; | |
a4311dfe | 2574 | |
0af5bc3e | 2575 | /* If a memref has both a size and an offset, we can use the smaller size. |
efc981bb | 2576 | We can't do this if the offset isn't known because we must view this |
0af5bc3e | 2577 | memref as being anywhere inside the DECL's MEM. */ |
527210c4 | 2578 | if (MEM_SIZE_KNOWN_P (x) && moffsetx_known_p) |
f5541398 | 2579 | sizex = MEM_SIZE (x); |
527210c4 | 2580 | if (MEM_SIZE_KNOWN_P (y) && moffsety_known_p) |
f5541398 | 2581 | sizey = MEM_SIZE (y); |
a4311dfe RK |
2582 | |
2583 | /* Put the values of the memref with the lower offset in X's values. */ | |
2584 | if (offsetx > offsety) | |
2585 | { | |
2586 | tem = offsetx, offsetx = offsety, offsety = tem; | |
2587 | tem = sizex, sizex = sizey, sizey = tem; | |
2588 | } | |
2589 | ||
2590 | /* If we don't know the size of the lower-offset value, we can't tell | |
2591 | if they conflict. Otherwise, we do the test. */ | |
a6f7c915 | 2592 | return sizex >= 0 && offsety >= offsetx + sizex; |
a4311dfe RK |
2593 | } |
2594 | ||
9362286d SB |
2595 | /* Helper for true_dependence and canon_true_dependence. |
2596 | Checks for true dependence: X is read after store in MEM takes place. | |
9ae8ffe7 | 2597 | |
9362286d SB |
2598 | If MEM_CANONICALIZED is FALSE, then X_ADDR and MEM_ADDR should be |
2599 | NULL_RTX, and the canonical addresses of MEM and X are both computed | |
2600 | here. If MEM_CANONICALIZED, then MEM must be already canonicalized. | |
2601 | ||
2602 | If X_ADDR is non-NULL, it is used in preference of XEXP (x, 0). | |
2603 | ||
2604 | Returns 1 if there is a true dependence, 0 otherwise. */ | |
2605 | ||
2606 | static int | |
ef4bddc2 | 2607 | true_dependence_1 (const_rtx mem, machine_mode mem_mode, rtx mem_addr, |
53d9622b | 2608 | const_rtx x, rtx x_addr, bool mem_canonicalized) |
9ae8ffe7 | 2609 | { |
0777fc02 | 2610 | rtx true_mem_addr; |
49982682 | 2611 | rtx base; |
f47e08d9 | 2612 | int ret; |
9ae8ffe7 | 2613 | |
9362286d SB |
2614 | gcc_checking_assert (mem_canonicalized ? (mem_addr != NULL_RTX) |
2615 | : (mem_addr == NULL_RTX && x_addr == NULL_RTX)); | |
2616 | ||
9ae8ffe7 JL |
2617 | if (MEM_VOLATILE_P (x) && MEM_VOLATILE_P (mem)) |
2618 | return 1; | |
2619 | ||
c4484b8f | 2620 | /* (mem:BLK (scratch)) is a special mechanism to conflict with everything. |
ac3768f6 | 2621 | This is used in epilogue deallocation functions, and in cselib. */ |
c4484b8f RH |
2622 | if (GET_MODE (x) == BLKmode && GET_CODE (XEXP (x, 0)) == SCRATCH) |
2623 | return 1; | |
2624 | if (GET_MODE (mem) == BLKmode && GET_CODE (XEXP (mem, 0)) == SCRATCH) | |
2625 | return 1; | |
9cd9e512 RH |
2626 | if (MEM_ALIAS_SET (x) == ALIAS_SET_MEMORY_BARRIER |
2627 | || MEM_ALIAS_SET (mem) == ALIAS_SET_MEMORY_BARRIER) | |
2628 | return 1; | |
c4484b8f | 2629 | |
0777fc02 UB |
2630 | if (! x_addr) |
2631 | x_addr = XEXP (x, 0); | |
2632 | x_addr = get_addr (x_addr); | |
2633 | ||
9362286d SB |
2634 | if (! mem_addr) |
2635 | { | |
2636 | mem_addr = XEXP (mem, 0); | |
2637 | if (mem_mode == VOIDmode) | |
2638 | mem_mode = GET_MODE (mem); | |
2639 | } | |
0777fc02 | 2640 | true_mem_addr = get_addr (mem_addr); |
eab5c70a | 2641 | |
878f5596 UB |
2642 | /* Read-only memory is by definition never modified, and therefore can't |
2643 | conflict with anything. However, don't assume anything when AND | |
2644 | addresses are involved and leave to the code below to determine | |
2645 | dependence. We don't expect to find read-only set on MEM, but | |
2646 | stupid user tricks can produce them, so don't die. */ | |
2647 | if (MEM_READONLY_P (x) | |
2648 | && GET_CODE (x_addr) != AND | |
0777fc02 | 2649 | && GET_CODE (true_mem_addr) != AND) |
878f5596 UB |
2650 | return 0; |
2651 | ||
2652 | /* If we have MEMs referring to different address spaces (which can | |
2653 | potentially overlap), we cannot easily tell from the addresses | |
2654 | whether the references overlap. */ | |
2655 | if (MEM_ADDR_SPACE (mem) != MEM_ADDR_SPACE (x)) | |
2656 | return 1; | |
2657 | ||
55efb413 JW |
2658 | base = find_base_term (x_addr); |
2659 | if (base && (GET_CODE (base) == LABEL_REF | |
2660 | || (GET_CODE (base) == SYMBOL_REF | |
2661 | && CONSTANT_POOL_ADDRESS_P (base)))) | |
2662 | return 0; | |
2663 | ||
0777fc02 UB |
2664 | rtx mem_base = find_base_term (true_mem_addr); |
2665 | if (! base_alias_check (x_addr, base, true_mem_addr, mem_base, | |
31b0a960 | 2666 | GET_MODE (x), mem_mode)) |
1c72c7f6 JC |
2667 | return 0; |
2668 | ||
eab5c70a | 2669 | x_addr = canon_rtx (x_addr); |
9362286d | 2670 | if (!mem_canonicalized) |
0777fc02 | 2671 | mem_addr = canon_rtx (true_mem_addr); |
6e73e666 | 2672 | |
f47e08d9 RG |
2673 | if ((ret = memrefs_conflict_p (GET_MODE_SIZE (mem_mode), mem_addr, |
2674 | SIZE_FOR_MODE (x), x_addr, 0)) != -1) | |
2675 | return ret; | |
2676 | ||
a95b3cc7 | 2677 | if (mems_in_disjoint_alias_sets_p (x, mem)) |
f47e08d9 RG |
2678 | return 0; |
2679 | ||
c6ea834c | 2680 | if (nonoverlapping_memrefs_p (mem, x, false)) |
0211b6ab | 2681 | return 0; |
175a7536 | 2682 | |
55b34b5f | 2683 | return rtx_refs_may_alias_p (x, mem, true); |
a13d4ebf AM |
2684 | } |
2685 | ||
9362286d SB |
2686 | /* True dependence: X is read after store in MEM takes place. */ |
2687 | ||
2688 | int | |
ef4bddc2 | 2689 | true_dependence (const_rtx mem, machine_mode mem_mode, const_rtx x) |
9362286d SB |
2690 | { |
2691 | return true_dependence_1 (mem, mem_mode, NULL_RTX, | |
53d9622b | 2692 | x, NULL_RTX, /*mem_canonicalized=*/false); |
9362286d SB |
2693 | } |
2694 | ||
a13d4ebf | 2695 | /* Canonical true dependence: X is read after store in MEM takes place. |
ca7fd9cd KH |
2696 | Variant of true_dependence which assumes MEM has already been |
2697 | canonicalized (hence we no longer do that here). | |
9362286d SB |
2698 | The mem_addr argument has been added, since true_dependence_1 computed |
2699 | this value prior to canonicalizing. */ | |
a13d4ebf AM |
2700 | |
2701 | int | |
ef4bddc2 | 2702 | canon_true_dependence (const_rtx mem, machine_mode mem_mode, rtx mem_addr, |
53d9622b | 2703 | const_rtx x, rtx x_addr) |
a13d4ebf | 2704 | { |
9362286d | 2705 | return true_dependence_1 (mem, mem_mode, mem_addr, |
53d9622b | 2706 | x, x_addr, /*mem_canonicalized=*/true); |
9ae8ffe7 JL |
2707 | } |
2708 | ||
da7d8304 | 2709 | /* Returns nonzero if a write to X might alias a previous read from |
393f9fed | 2710 | (or, if WRITEP is true, a write to) MEM. |
bd280792 JR |
2711 | If X_CANONCALIZED is true, then X_ADDR is the canonicalized address of X, |
2712 | and X_MODE the mode for that access. | |
2713 | If MEM_CANONICALIZED is true, MEM is canonicalized. */ | |
9ae8ffe7 | 2714 | |
2c72b78f | 2715 | static int |
bd280792 | 2716 | write_dependence_p (const_rtx mem, |
ef4bddc2 | 2717 | const_rtx x, machine_mode x_mode, rtx x_addr, |
bd280792 | 2718 | bool mem_canonicalized, bool x_canonicalized, bool writep) |
9ae8ffe7 | 2719 | { |
bd280792 | 2720 | rtx mem_addr; |
0777fc02 | 2721 | rtx true_mem_addr, true_x_addr; |
49982682 | 2722 | rtx base; |
f47e08d9 | 2723 | int ret; |
6e73e666 | 2724 | |
bd280792 JR |
2725 | gcc_checking_assert (x_canonicalized |
2726 | ? (x_addr != NULL_RTX && x_mode != VOIDmode) | |
2727 | : (x_addr == NULL_RTX && x_mode == VOIDmode)); | |
393f9fed | 2728 | |
9ae8ffe7 JL |
2729 | if (MEM_VOLATILE_P (x) && MEM_VOLATILE_P (mem)) |
2730 | return 1; | |
2731 | ||
c4484b8f RH |
2732 | /* (mem:BLK (scratch)) is a special mechanism to conflict with everything. |
2733 | This is used in epilogue deallocation functions. */ | |
2734 | if (GET_MODE (x) == BLKmode && GET_CODE (XEXP (x, 0)) == SCRATCH) | |
2735 | return 1; | |
2736 | if (GET_MODE (mem) == BLKmode && GET_CODE (XEXP (mem, 0)) == SCRATCH) | |
2737 | return 1; | |
9cd9e512 RH |
2738 | if (MEM_ALIAS_SET (x) == ALIAS_SET_MEMORY_BARRIER |
2739 | || MEM_ALIAS_SET (mem) == ALIAS_SET_MEMORY_BARRIER) | |
2740 | return 1; | |
c4484b8f | 2741 | |
bd280792 | 2742 | if (!x_addr) |
0777fc02 UB |
2743 | x_addr = XEXP (x, 0); |
2744 | true_x_addr = get_addr (x_addr); | |
2745 | ||
2746 | mem_addr = XEXP (mem, 0); | |
2747 | true_mem_addr = get_addr (mem_addr); | |
55efb413 | 2748 | |
878f5596 UB |
2749 | /* A read from read-only memory can't conflict with read-write memory. |
2750 | Don't assume anything when AND addresses are involved and leave to | |
2751 | the code below to determine dependence. */ | |
2752 | if (!writep | |
2753 | && MEM_READONLY_P (mem) | |
0777fc02 UB |
2754 | && GET_CODE (true_x_addr) != AND |
2755 | && GET_CODE (true_mem_addr) != AND) | |
878f5596 UB |
2756 | return 0; |
2757 | ||
2758 | /* If we have MEMs referring to different address spaces (which can | |
2759 | potentially overlap), we cannot easily tell from the addresses | |
2760 | whether the references overlap. */ | |
2761 | if (MEM_ADDR_SPACE (mem) != MEM_ADDR_SPACE (x)) | |
2762 | return 1; | |
2763 | ||
0777fc02 | 2764 | base = find_base_term (true_mem_addr); |
31b0a960 RB |
2765 | if (! writep |
2766 | && base | |
2767 | && (GET_CODE (base) == LABEL_REF | |
2768 | || (GET_CODE (base) == SYMBOL_REF | |
2769 | && CONSTANT_POOL_ADDRESS_P (base)))) | |
2770 | return 0; | |
49982682 | 2771 | |
0777fc02 UB |
2772 | rtx x_base = find_base_term (true_x_addr); |
2773 | if (! base_alias_check (true_x_addr, x_base, true_mem_addr, base, | |
2774 | GET_MODE (x), GET_MODE (mem))) | |
41472af8 MM |
2775 | return 0; |
2776 | ||
bd280792 | 2777 | if (!x_canonicalized) |
393f9fed | 2778 | { |
0777fc02 | 2779 | x_addr = canon_rtx (true_x_addr); |
bd280792 | 2780 | x_mode = GET_MODE (x); |
393f9fed | 2781 | } |
bd280792 | 2782 | if (!mem_canonicalized) |
0777fc02 | 2783 | mem_addr = canon_rtx (true_mem_addr); |
6e73e666 | 2784 | |
bd280792 JR |
2785 | if ((ret = memrefs_conflict_p (SIZE_FOR_MODE (mem), mem_addr, |
2786 | GET_MODE_SIZE (x_mode), x_addr, 0)) != -1) | |
f47e08d9 RG |
2787 | return ret; |
2788 | ||
c6ea834c | 2789 | if (nonoverlapping_memrefs_p (x, mem, false)) |
c6df88cb MM |
2790 | return 0; |
2791 | ||
55b34b5f | 2792 | return rtx_refs_may_alias_p (x, mem, false); |
c6df88cb MM |
2793 | } |
2794 | ||
2795 | /* Anti dependence: X is written after read in MEM takes place. */ | |
2796 | ||
2797 | int | |
4f588890 | 2798 | anti_dependence (const_rtx mem, const_rtx x) |
c6df88cb | 2799 | { |
bd280792 JR |
2800 | return write_dependence_p (mem, x, VOIDmode, NULL_RTX, |
2801 | /*mem_canonicalized=*/false, | |
2802 | /*x_canonicalized*/false, /*writep=*/false); | |
393f9fed JR |
2803 | } |
2804 | ||
bd280792 JR |
2805 | /* Likewise, but we already have a canonicalized MEM, and X_ADDR for X. |
2806 | Also, consider X in X_MODE (which might be from an enclosing | |
2807 | STRICT_LOW_PART / ZERO_EXTRACT). | |
2808 | If MEM_CANONICALIZED is true, MEM is canonicalized. */ | |
393f9fed JR |
2809 | |
2810 | int | |
bd280792 | 2811 | canon_anti_dependence (const_rtx mem, bool mem_canonicalized, |
ef4bddc2 | 2812 | const_rtx x, machine_mode x_mode, rtx x_addr) |
393f9fed | 2813 | { |
bd280792 JR |
2814 | return write_dependence_p (mem, x, x_mode, x_addr, |
2815 | mem_canonicalized, /*x_canonicalized=*/true, | |
2816 | /*writep=*/false); | |
9ae8ffe7 JL |
2817 | } |
2818 | ||
2819 | /* Output dependence: X is written after store in MEM takes place. */ | |
2820 | ||
2821 | int | |
4f588890 | 2822 | output_dependence (const_rtx mem, const_rtx x) |
9ae8ffe7 | 2823 | { |
bd280792 JR |
2824 | return write_dependence_p (mem, x, VOIDmode, NULL_RTX, |
2825 | /*mem_canonicalized=*/false, | |
2826 | /*x_canonicalized*/false, /*writep=*/true); | |
9ae8ffe7 | 2827 | } |
c14b9960 | 2828 | \f |
6e73e666 | 2829 | |
c6ea834c BM |
2830 | |
2831 | /* Check whether X may be aliased with MEM. Don't do offset-based | |
2832 | memory disambiguation & TBAA. */ | |
2833 | int | |
2834 | may_alias_p (const_rtx mem, const_rtx x) | |
2835 | { | |
2836 | rtx x_addr, mem_addr; | |
c6ea834c BM |
2837 | |
2838 | if (MEM_VOLATILE_P (x) && MEM_VOLATILE_P (mem)) | |
2839 | return 1; | |
2840 | ||
a95b3cc7 RG |
2841 | /* (mem:BLK (scratch)) is a special mechanism to conflict with everything. |
2842 | This is used in epilogue deallocation functions. */ | |
2843 | if (GET_MODE (x) == BLKmode && GET_CODE (XEXP (x, 0)) == SCRATCH) | |
2844 | return 1; | |
2845 | if (GET_MODE (mem) == BLKmode && GET_CODE (XEXP (mem, 0)) == SCRATCH) | |
c6ea834c | 2846 | return 1; |
c6ea834c BM |
2847 | if (MEM_ALIAS_SET (x) == ALIAS_SET_MEMORY_BARRIER |
2848 | || MEM_ALIAS_SET (mem) == ALIAS_SET_MEMORY_BARRIER) | |
2849 | return 1; | |
2850 | ||
c6ea834c | 2851 | x_addr = XEXP (x, 0); |
0777fc02 UB |
2852 | x_addr = get_addr (x_addr); |
2853 | ||
c6ea834c | 2854 | mem_addr = XEXP (mem, 0); |
0777fc02 | 2855 | mem_addr = get_addr (mem_addr); |
c6ea834c | 2856 | |
878f5596 UB |
2857 | /* Read-only memory is by definition never modified, and therefore can't |
2858 | conflict with anything. However, don't assume anything when AND | |
2859 | addresses are involved and leave to the code below to determine | |
2860 | dependence. We don't expect to find read-only set on MEM, but | |
2861 | stupid user tricks can produce them, so don't die. */ | |
2862 | if (MEM_READONLY_P (x) | |
2863 | && GET_CODE (x_addr) != AND | |
2864 | && GET_CODE (mem_addr) != AND) | |
2865 | return 0; | |
2866 | ||
2867 | /* If we have MEMs referring to different address spaces (which can | |
2868 | potentially overlap), we cannot easily tell from the addresses | |
2869 | whether the references overlap. */ | |
2870 | if (MEM_ADDR_SPACE (mem) != MEM_ADDR_SPACE (x)) | |
2871 | return 1; | |
2872 | ||
31b0a960 RB |
2873 | rtx x_base = find_base_term (x_addr); |
2874 | rtx mem_base = find_base_term (mem_addr); | |
2875 | if (! base_alias_check (x_addr, x_base, mem_addr, mem_base, | |
2876 | GET_MODE (x), GET_MODE (mem_addr))) | |
c6ea834c BM |
2877 | return 0; |
2878 | ||
c6ea834c BM |
2879 | if (nonoverlapping_memrefs_p (mem, x, true)) |
2880 | return 0; | |
2881 | ||
c6ea834c BM |
2882 | /* TBAA not valid for loop_invarint */ |
2883 | return rtx_refs_may_alias_p (x, mem, false); | |
2884 | } | |
2885 | ||
6e73e666 | 2886 | void |
b5deb7b6 | 2887 | init_alias_target (void) |
6e73e666 | 2888 | { |
b3694847 | 2889 | int i; |
6e73e666 | 2890 | |
9fc37b2b RS |
2891 | if (!arg_base_value) |
2892 | arg_base_value = gen_rtx_ADDRESS (VOIDmode, 0); | |
2893 | ||
b5deb7b6 SL |
2894 | memset (static_reg_base_value, 0, sizeof static_reg_base_value); |
2895 | ||
6e73e666 JC |
2896 | for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) |
2897 | /* Check whether this register can hold an incoming pointer | |
2898 | argument. FUNCTION_ARG_REGNO_P tests outgoing register | |
ec5c56db | 2899 | numbers, so translate if necessary due to register windows. */ |
6e73e666 JC |
2900 | if (FUNCTION_ARG_REGNO_P (OUTGOING_REGNO (i)) |
2901 | && HARD_REGNO_MODE_OK (i, Pmode)) | |
9fc37b2b RS |
2902 | static_reg_base_value[i] = arg_base_value; |
2903 | ||
757e8ba2 JJ |
2904 | static_reg_base_value[STACK_POINTER_REGNUM] |
2905 | = unique_base_value (UNIQUE_BASE_VALUE_SP); | |
2906 | static_reg_base_value[ARG_POINTER_REGNUM] | |
2907 | = unique_base_value (UNIQUE_BASE_VALUE_ARGP); | |
2908 | static_reg_base_value[FRAME_POINTER_REGNUM] | |
2909 | = unique_base_value (UNIQUE_BASE_VALUE_FP); | |
c3e08036 TS |
2910 | if (!HARD_FRAME_POINTER_IS_FRAME_POINTER) |
2911 | static_reg_base_value[HARD_FRAME_POINTER_REGNUM] | |
2912 | = unique_base_value (UNIQUE_BASE_VALUE_HFP); | |
bf1660a6 JL |
2913 | } |
2914 | ||
7b52eede JH |
2915 | /* Set MEMORY_MODIFIED when X modifies DATA (that is assumed |
2916 | to be memory reference. */ | |
2917 | static bool memory_modified; | |
2918 | static void | |
aa317c97 | 2919 | memory_modified_1 (rtx x, const_rtx pat ATTRIBUTE_UNUSED, void *data) |
7b52eede | 2920 | { |
3c0cb5de | 2921 | if (MEM_P (x)) |
7b52eede | 2922 | { |
9678086d | 2923 | if (anti_dependence (x, (const_rtx)data) || output_dependence (x, (const_rtx)data)) |
7b52eede JH |
2924 | memory_modified = true; |
2925 | } | |
2926 | } | |
2927 | ||
2928 | ||
2929 | /* Return true when INSN possibly modify memory contents of MEM | |
454ff5cb | 2930 | (i.e. address can be modified). */ |
7b52eede | 2931 | bool |
9678086d | 2932 | memory_modified_in_insn_p (const_rtx mem, const_rtx insn) |
7b52eede JH |
2933 | { |
2934 | if (!INSN_P (insn)) | |
2935 | return false; | |
2936 | memory_modified = false; | |
aa317c97 | 2937 | note_stores (PATTERN (insn), memory_modified_1, CONST_CAST_RTX(mem)); |
7b52eede JH |
2938 | return memory_modified; |
2939 | } | |
2940 | ||
a7b159a4 AH |
2941 | /* Return TRUE if the destination of a set is rtx identical to |
2942 | ITEM. */ | |
2943 | static inline bool | |
2944 | set_dest_equal_p (const_rtx set, const_rtx item) | |
2945 | { | |
2946 | rtx dest = SET_DEST (set); | |
2947 | return rtx_equal_p (dest, item); | |
2948 | } | |
2949 | ||
2950 | /* Like memory_modified_in_insn_p, but return TRUE if INSN will | |
2951 | *DEFINITELY* modify the memory contents of MEM. */ | |
2952 | bool | |
2953 | memory_must_be_modified_in_insn_p (const_rtx mem, const_rtx insn) | |
2954 | { | |
2955 | if (!INSN_P (insn)) | |
2956 | return false; | |
2957 | insn = PATTERN (insn); | |
2958 | if (GET_CODE (insn) == SET) | |
2959 | return set_dest_equal_p (insn, mem); | |
2960 | else if (GET_CODE (insn) == PARALLEL) | |
2961 | { | |
2962 | int i; | |
2963 | for (i = 0; i < XVECLEN (insn, 0); i++) | |
2964 | { | |
2965 | rtx sub = XVECEXP (insn, 0, i); | |
2966 | if (GET_CODE (sub) == SET | |
2967 | && set_dest_equal_p (sub, mem)) | |
2968 | return true; | |
2969 | } | |
2970 | } | |
2971 | return false; | |
2972 | } | |
2973 | ||
c13e8210 MM |
2974 | /* Initialize the aliasing machinery. Initialize the REG_KNOWN_VALUE |
2975 | array. */ | |
2976 | ||
9ae8ffe7 | 2977 | void |
4682ae04 | 2978 | init_alias_analysis (void) |
9ae8ffe7 | 2979 | { |
c582d54a | 2980 | unsigned int maxreg = max_reg_num (); |
ea64ef27 | 2981 | int changed, pass; |
b3694847 SS |
2982 | int i; |
2983 | unsigned int ui; | |
d36a28b8 DM |
2984 | rtx_insn *insn; |
2985 | rtx val; | |
131db6b8 SB |
2986 | int rpo_cnt; |
2987 | int *rpo; | |
9ae8ffe7 | 2988 | |
0d446150 JH |
2989 | timevar_push (TV_ALIAS_ANALYSIS); |
2990 | ||
92390dd1 | 2991 | vec_safe_grow_cleared (reg_known_value, maxreg - FIRST_PSEUDO_REGISTER); |
9ff3c7ca | 2992 | reg_known_equiv_p = sbitmap_alloc (maxreg - FIRST_PSEUDO_REGISTER); |
dd3d1ec0 | 2993 | bitmap_clear (reg_known_equiv_p); |
9ae8ffe7 | 2994 | |
08c79682 | 2995 | /* If we have memory allocated from the previous run, use it. */ |
c582d54a | 2996 | if (old_reg_base_value) |
08c79682 KH |
2997 | reg_base_value = old_reg_base_value; |
2998 | ||
2999 | if (reg_base_value) | |
9771b263 | 3000 | reg_base_value->truncate (0); |
08c79682 | 3001 | |
9771b263 | 3002 | vec_safe_grow_cleared (reg_base_value, maxreg); |
ac606739 | 3003 | |
5ed6ace5 | 3004 | new_reg_base_value = XNEWVEC (rtx, maxreg); |
d630245f | 3005 | reg_seen = sbitmap_alloc (maxreg); |
ec907dd8 JL |
3006 | |
3007 | /* The basic idea is that each pass through this loop will use the | |
3008 | "constant" information from the previous pass to propagate alias | |
3009 | information through another level of assignments. | |
3010 | ||
131db6b8 SB |
3011 | The propagation is done on the CFG in reverse post-order, to propagate |
3012 | things forward as far as possible in each iteration. | |
3013 | ||
ec907dd8 JL |
3014 | This could get expensive if the assignment chains are long. Maybe |
3015 | we should throttle the number of iterations, possibly based on | |
6e73e666 | 3016 | the optimization level or flag_expensive_optimizations. |
ec907dd8 JL |
3017 | |
3018 | We could propagate more information in the first pass by making use | |
6fb5fa3c | 3019 | of DF_REG_DEF_COUNT to determine immediately that the alias information |
ea64ef27 JL |
3020 | for a pseudo is "constant". |
3021 | ||
3022 | A program with an uninitialized variable can cause an infinite loop | |
3023 | here. Instead of doing a full dataflow analysis to detect such problems | |
3024 | we just cap the number of iterations for the loop. | |
3025 | ||
3026 | The state of the arrays for the set chain in question does not matter | |
3027 | since the program has undefined behavior. */ | |
6e73e666 | 3028 | |
0cae8d31 | 3029 | rpo = XNEWVEC (int, n_basic_blocks_for_fn (cfun)); |
131db6b8 SB |
3030 | rpo_cnt = pre_and_rev_post_order_compute (NULL, rpo, false); |
3031 | ||
e86a9946 RS |
3032 | /* The prologue/epilogue insns are not threaded onto the |
3033 | insn chain until after reload has completed. Thus, | |
3034 | there is no sense wasting time checking if INSN is in | |
3035 | the prologue/epilogue until after reload has completed. */ | |
3036 | bool could_be_prologue_epilogue = ((targetm.have_prologue () | |
3037 | || targetm.have_epilogue ()) | |
3038 | && reload_completed); | |
3039 | ||
ea64ef27 | 3040 | pass = 0; |
6e73e666 | 3041 | do |
ec907dd8 JL |
3042 | { |
3043 | /* Assume nothing will change this iteration of the loop. */ | |
3044 | changed = 0; | |
3045 | ||
ec907dd8 | 3046 | /* We want to assign the same IDs each iteration of this loop, so |
9fc37b2b RS |
3047 | start counting from one each iteration of the loop. */ |
3048 | unique_id = 1; | |
ec907dd8 | 3049 | |
f5143c46 | 3050 | /* We're at the start of the function each iteration through the |
ec907dd8 | 3051 | loop, so we're copying arguments. */ |
83bbd9b6 | 3052 | copying_arguments = true; |
9ae8ffe7 | 3053 | |
6e73e666 | 3054 | /* Wipe the potential alias information clean for this pass. */ |
c582d54a | 3055 | memset (new_reg_base_value, 0, maxreg * sizeof (rtx)); |
8072f69c | 3056 | |
6e73e666 | 3057 | /* Wipe the reg_seen array clean. */ |
f61e445a | 3058 | bitmap_clear (reg_seen); |
9ae8ffe7 | 3059 | |
356610cb EB |
3060 | /* Initialize the alias information for this pass. */ |
3061 | for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) | |
3062 | if (static_reg_base_value[i]) | |
3063 | { | |
3064 | new_reg_base_value[i] = static_reg_base_value[i]; | |
3065 | bitmap_set_bit (reg_seen, i); | |
3066 | } | |
6e73e666 | 3067 | |
ec907dd8 | 3068 | /* Walk the insns adding values to the new_reg_base_value array. */ |
131db6b8 | 3069 | for (i = 0; i < rpo_cnt; i++) |
9ae8ffe7 | 3070 | { |
06e28de2 | 3071 | basic_block bb = BASIC_BLOCK_FOR_FN (cfun, rpo[i]); |
131db6b8 | 3072 | FOR_BB_INSNS (bb, insn) |
ec907dd8 | 3073 | { |
131db6b8 SB |
3074 | if (NONDEBUG_INSN_P (insn)) |
3075 | { | |
3076 | rtx note, set; | |
efc9bd41 | 3077 | |
e86a9946 | 3078 | if (could_be_prologue_epilogue |
131db6b8 SB |
3079 | && prologue_epilogue_contains (insn)) |
3080 | continue; | |
efc9bd41 | 3081 | |
131db6b8 SB |
3082 | /* If this insn has a noalias note, process it, Otherwise, |
3083 | scan for sets. A simple set will have no side effects | |
3084 | which could change the base value of any other register. */ | |
6e73e666 | 3085 | |
131db6b8 SB |
3086 | if (GET_CODE (PATTERN (insn)) == SET |
3087 | && REG_NOTES (insn) != 0 | |
3088 | && find_reg_note (insn, REG_NOALIAS, NULL_RTX)) | |
3089 | record_set (SET_DEST (PATTERN (insn)), NULL_RTX, NULL); | |
3090 | else | |
3091 | note_stores (PATTERN (insn), record_set, NULL); | |
6e73e666 | 3092 | |
131db6b8 | 3093 | set = single_set (insn); |
6e73e666 | 3094 | |
131db6b8 SB |
3095 | if (set != 0 |
3096 | && REG_P (SET_DEST (set)) | |
3097 | && REGNO (SET_DEST (set)) >= FIRST_PSEUDO_REGISTER) | |
713f41f9 | 3098 | { |
131db6b8 SB |
3099 | unsigned int regno = REGNO (SET_DEST (set)); |
3100 | rtx src = SET_SRC (set); | |
3101 | rtx t; | |
3102 | ||
3103 | note = find_reg_equal_equiv_note (insn); | |
3104 | if (note && REG_NOTE_KIND (note) == REG_EQUAL | |
3105 | && DF_REG_DEF_COUNT (regno) != 1) | |
3106 | note = NULL_RTX; | |
3107 | ||
3108 | if (note != NULL_RTX | |
3109 | && GET_CODE (XEXP (note, 0)) != EXPR_LIST | |
3110 | && ! rtx_varies_p (XEXP (note, 0), 1) | |
3111 | && ! reg_overlap_mentioned_p (SET_DEST (set), | |
3112 | XEXP (note, 0))) | |
3113 | { | |
3114 | set_reg_known_value (regno, XEXP (note, 0)); | |
3115 | set_reg_known_equiv_p (regno, | |
3116 | REG_NOTE_KIND (note) == REG_EQUIV); | |
3117 | } | |
3118 | else if (DF_REG_DEF_COUNT (regno) == 1 | |
3119 | && GET_CODE (src) == PLUS | |
3120 | && REG_P (XEXP (src, 0)) | |
3121 | && (t = get_reg_known_value (REGNO (XEXP (src, 0)))) | |
3122 | && CONST_INT_P (XEXP (src, 1))) | |
3123 | { | |
3124 | t = plus_constant (GET_MODE (src), t, | |
3125 | INTVAL (XEXP (src, 1))); | |
3126 | set_reg_known_value (regno, t); | |
3127 | set_reg_known_equiv_p (regno, false); | |
3128 | } | |
3129 | else if (DF_REG_DEF_COUNT (regno) == 1 | |
3130 | && ! rtx_varies_p (src, 1)) | |
3131 | { | |
3132 | set_reg_known_value (regno, src); | |
3133 | set_reg_known_equiv_p (regno, false); | |
3134 | } | |
713f41f9 | 3135 | } |
6e73e666 | 3136 | } |
131db6b8 SB |
3137 | else if (NOTE_P (insn) |
3138 | && NOTE_KIND (insn) == NOTE_INSN_FUNCTION_BEG) | |
3139 | copying_arguments = false; | |
ec907dd8 | 3140 | } |
6e73e666 | 3141 | } |
ec907dd8 | 3142 | |
6e73e666 | 3143 | /* Now propagate values from new_reg_base_value to reg_base_value. */ |
62e5bf5d | 3144 | gcc_assert (maxreg == (unsigned int) max_reg_num ()); |
c22cacf3 | 3145 | |
c582d54a | 3146 | for (ui = 0; ui < maxreg; ui++) |
6e73e666 | 3147 | { |
e51712db | 3148 | if (new_reg_base_value[ui] |
9771b263 DN |
3149 | && new_reg_base_value[ui] != (*reg_base_value)[ui] |
3150 | && ! rtx_equal_p (new_reg_base_value[ui], (*reg_base_value)[ui])) | |
ec907dd8 | 3151 | { |
9771b263 | 3152 | (*reg_base_value)[ui] = new_reg_base_value[ui]; |
6e73e666 | 3153 | changed = 1; |
ec907dd8 | 3154 | } |
9ae8ffe7 | 3155 | } |
9ae8ffe7 | 3156 | } |
6e73e666 | 3157 | while (changed && ++pass < MAX_ALIAS_LOOP_PASSES); |
131db6b8 | 3158 | XDELETEVEC (rpo); |
9ae8ffe7 JL |
3159 | |
3160 | /* Fill in the remaining entries. */ | |
9771b263 | 3161 | FOR_EACH_VEC_ELT (*reg_known_value, i, val) |
9ff3c7ca SB |
3162 | { |
3163 | int regno = i + FIRST_PSEUDO_REGISTER; | |
3164 | if (! val) | |
3165 | set_reg_known_value (regno, regno_reg_rtx[regno]); | |
3166 | } | |
9ae8ffe7 | 3167 | |
e05e2395 MM |
3168 | /* Clean up. */ |
3169 | free (new_reg_base_value); | |
ec907dd8 | 3170 | new_reg_base_value = 0; |
d630245f | 3171 | sbitmap_free (reg_seen); |
9ae8ffe7 | 3172 | reg_seen = 0; |
0d446150 | 3173 | timevar_pop (TV_ALIAS_ANALYSIS); |
9ae8ffe7 JL |
3174 | } |
3175 | ||
61630b27 JJ |
3176 | /* Equate REG_BASE_VALUE (reg1) to REG_BASE_VALUE (reg2). |
3177 | Special API for var-tracking pass purposes. */ | |
3178 | ||
3179 | void | |
3180 | vt_equate_reg_base_value (const_rtx reg1, const_rtx reg2) | |
3181 | { | |
9771b263 | 3182 | (*reg_base_value)[REGNO (reg1)] = REG_BASE_VALUE (reg2); |
61630b27 JJ |
3183 | } |
3184 | ||
9ae8ffe7 | 3185 | void |
4682ae04 | 3186 | end_alias_analysis (void) |
9ae8ffe7 | 3187 | { |
c582d54a | 3188 | old_reg_base_value = reg_base_value; |
9771b263 | 3189 | vec_free (reg_known_value); |
9ff3c7ca | 3190 | sbitmap_free (reg_known_equiv_p); |
9ae8ffe7 | 3191 | } |
e2500fed | 3192 | |
3ecf9d13 JH |
3193 | void |
3194 | dump_alias_stats_in_alias_c (FILE *s) | |
3195 | { | |
3196 | fprintf (s, " TBAA oracle: %llu disambiguations %llu queries\n" | |
3197 | " %llu are in alias set 0\n" | |
3198 | " %llu queries asked about the same object\n" | |
3199 | " %llu queries asked about the same alias set\n" | |
3200 | " %llu access volatile\n" | |
6e042ef4 JH |
3201 | " %llu are dependent in the DAG\n" |
3202 | " %llu are aritificially in conflict with void *\n", | |
3ecf9d13 JH |
3203 | alias_stats.num_disambiguated, |
3204 | alias_stats.num_alias_zero + alias_stats.num_same_alias_set | |
3205 | + alias_stats.num_same_objects + alias_stats.num_volatile | |
6e042ef4 JH |
3206 | + alias_stats.num_dag + alias_stats.num_disambiguated |
3207 | + alias_stats.num_universal, | |
3ecf9d13 | 3208 | alias_stats.num_alias_zero, alias_stats.num_same_alias_set, |
6e042ef4 JH |
3209 | alias_stats.num_same_objects, alias_stats.num_volatile, |
3210 | alias_stats.num_dag, alias_stats.num_universal); | |
3ecf9d13 | 3211 | } |
e2500fed | 3212 | #include "gt-alias.h" |