]>
Commit | Line | Data |
---|---|---|
fa49fd0f RK |
1 | /* Common subexpression elimination library for GNU compiler. |
2 | Copyright (C) 1987, 1988, 1989, 1992, 1993, 1994, 1995, 1996, 1997, 1998, | |
0de3e43f | 3 | 1999, 2000, 2001, 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010 |
8dd5516b | 4 | Free Software Foundation, Inc. |
fa49fd0f | 5 | |
1322177d | 6 | This file is part of GCC. |
fa49fd0f | 7 | |
1322177d LB |
8 | GCC is free software; you can redistribute it and/or modify it under |
9 | the terms of the GNU General Public License as published by the Free | |
9dcd6f09 | 10 | Software Foundation; either version 3, or (at your option) any later |
1322177d | 11 | version. |
fa49fd0f | 12 | |
1322177d LB |
13 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY |
14 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
15 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
16 | for more details. | |
fa49fd0f RK |
17 | |
18 | You should have received a copy of the GNU General Public License | |
9dcd6f09 NC |
19 | along with GCC; see the file COPYING3. If not see |
20 | <http://www.gnu.org/licenses/>. */ | |
fa49fd0f RK |
21 | |
22 | #include "config.h" | |
23 | #include "system.h" | |
4977bab6 ZW |
24 | #include "coretypes.h" |
25 | #include "tm.h" | |
fa49fd0f RK |
26 | |
27 | #include "rtl.h" | |
28 | #include "tm_p.h" | |
29 | #include "regs.h" | |
30 | #include "hard-reg-set.h" | |
31 | #include "flags.h" | |
fa49fd0f RK |
32 | #include "insn-config.h" |
33 | #include "recog.h" | |
34 | #include "function.h" | |
78528714 | 35 | #include "emit-rtl.h" |
fa49fd0f RK |
36 | #include "toplev.h" |
37 | #include "output.h" | |
38 | #include "ggc.h" | |
fa49fd0f | 39 | #include "hashtab.h" |
b5b8b0ac | 40 | #include "tree-pass.h" |
fa49fd0f | 41 | #include "cselib.h" |
c65ecebc | 42 | #include "params.h" |
6a59927d | 43 | #include "alloc-pool.h" |
29c1846b | 44 | #include "target.h" |
7a8cba34 | 45 | #include "bitmap.h" |
fa49fd0f | 46 | |
463301c3 | 47 | static bool cselib_record_memory; |
457eeaae | 48 | static bool cselib_preserve_constants; |
7080f735 AJ |
49 | static int entry_and_rtx_equal_p (const void *, const void *); |
50 | static hashval_t get_value_hash (const void *); | |
51 | static struct elt_list *new_elt_list (struct elt_list *, cselib_val *); | |
52 | static struct elt_loc_list *new_elt_loc_list (struct elt_loc_list *, rtx); | |
53 | static void unchain_one_value (cselib_val *); | |
54 | static void unchain_one_elt_list (struct elt_list **); | |
55 | static void unchain_one_elt_loc_list (struct elt_loc_list **); | |
7080f735 AJ |
56 | static int discard_useless_locs (void **, void *); |
57 | static int discard_useless_values (void **, void *); | |
58 | static void remove_useless_values (void); | |
29c1846b | 59 | static unsigned int cselib_hash_rtx (rtx, int); |
b5b8b0ac | 60 | static cselib_val *new_cselib_val (unsigned int, enum machine_mode, rtx); |
7080f735 AJ |
61 | static void add_mem_for_addr (cselib_val *, cselib_val *, rtx); |
62 | static cselib_val *cselib_lookup_mem (rtx, int); | |
63 | static void cselib_invalidate_regno (unsigned int, enum machine_mode); | |
7080f735 | 64 | static void cselib_invalidate_mem (rtx); |
7080f735 AJ |
65 | static void cselib_record_set (rtx, cselib_val *, cselib_val *); |
66 | static void cselib_record_sets (rtx); | |
fa49fd0f | 67 | |
b5b8b0ac AO |
68 | struct expand_value_data |
69 | { | |
70 | bitmap regs_active; | |
71 | cselib_expand_callback callback; | |
72 | void *callback_arg; | |
864ddef7 | 73 | bool dummy; |
b5b8b0ac AO |
74 | }; |
75 | ||
76 | static rtx cselib_expand_value_rtx_1 (rtx, struct expand_value_data *, int); | |
77 | ||
fa49fd0f RK |
78 | /* There are three ways in which cselib can look up an rtx: |
79 | - for a REG, the reg_values table (which is indexed by regno) is used | |
80 | - for a MEM, we recursively look up its address and then follow the | |
81 | addr_list of that value | |
82 | - for everything else, we compute a hash value and go through the hash | |
83 | table. Since different rtx's can still have the same hash value, | |
84 | this involves walking the table entries for a given value and comparing | |
85 | the locations of the entries with the rtx we are looking up. */ | |
86 | ||
87 | /* A table that enables us to look up elts by their value. */ | |
7c514720 | 88 | static htab_t cselib_hash_table; |
fa49fd0f RK |
89 | |
90 | /* This is a global so we don't have to pass this through every function. | |
91 | It is used in new_elt_loc_list to set SETTING_INSN. */ | |
92 | static rtx cselib_current_insn; | |
93 | ||
5440c0e7 AO |
94 | /* The unique id that the next create value will take. */ |
95 | static unsigned int next_uid; | |
fa49fd0f RK |
96 | |
97 | /* The number of registers we had when the varrays were last resized. */ | |
98 | static unsigned int cselib_nregs; | |
99 | ||
5847e8da AO |
100 | /* Count values without known locations, or with only locations that |
101 | wouldn't have been known except for debug insns. Whenever this | |
102 | grows too big, we remove these useless values from the table. | |
103 | ||
104 | Counting values with only debug values is a bit tricky. We don't | |
105 | want to increment n_useless_values when we create a value for a | |
106 | debug insn, for this would get n_useless_values out of sync, but we | |
107 | want increment it if all locs in the list that were ever referenced | |
108 | in nondebug insns are removed from the list. | |
109 | ||
110 | In the general case, once we do that, we'd have to stop accepting | |
111 | nondebug expressions in the loc list, to avoid having two values | |
112 | equivalent that, without debug insns, would have been made into | |
113 | separate values. However, because debug insns never introduce | |
114 | equivalences themselves (no assignments), the only means for | |
115 | growing loc lists is through nondebug assignments. If the locs | |
116 | also happen to be referenced in debug insns, it will work just fine. | |
117 | ||
118 | A consequence of this is that there's at most one debug-only loc in | |
119 | each loc list. If we keep it in the first entry, testing whether | |
120 | we have a debug-only loc list takes O(1). | |
121 | ||
122 | Furthermore, since any additional entry in a loc list containing a | |
123 | debug loc would have to come from an assignment (nondebug) that | |
124 | references both the initial debug loc and the newly-equivalent loc, | |
125 | the initial debug loc would be promoted to a nondebug loc, and the | |
126 | loc list would not contain debug locs any more. | |
127 | ||
128 | So the only case we have to be careful with in order to keep | |
129 | n_useless_values in sync between debug and nondebug compilations is | |
130 | to avoid incrementing n_useless_values when removing the single loc | |
131 | from a value that turns out to not appear outside debug values. We | |
132 | increment n_useless_debug_values instead, and leave such values | |
133 | alone until, for other reasons, we garbage-collect useless | |
134 | values. */ | |
fa49fd0f | 135 | static int n_useless_values; |
5847e8da AO |
136 | static int n_useless_debug_values; |
137 | ||
138 | /* Count values whose locs have been taken exclusively from debug | |
139 | insns for the entire life of the value. */ | |
140 | static int n_debug_values; | |
fa49fd0f RK |
141 | |
142 | /* Number of useless values before we remove them from the hash table. */ | |
143 | #define MAX_USELESS_VALUES 32 | |
144 | ||
60fa6660 AO |
145 | /* This table maps from register number to values. It does not |
146 | contain pointers to cselib_val structures, but rather elt_lists. | |
147 | The purpose is to be able to refer to the same register in | |
148 | different modes. The first element of the list defines the mode in | |
149 | which the register was set; if the mode is unknown or the value is | |
150 | no longer valid in that mode, ELT will be NULL for the first | |
151 | element. */ | |
5211d65a KH |
152 | static struct elt_list **reg_values; |
153 | static unsigned int reg_values_size; | |
6790d1ab | 154 | #define REG_VALUES(i) reg_values[i] |
fa49fd0f | 155 | |
31825e57 | 156 | /* The largest number of hard regs used by any entry added to the |
eb232f4e | 157 | REG_VALUES table. Cleared on each cselib_clear_table() invocation. */ |
31825e57 DM |
158 | static unsigned int max_value_regs; |
159 | ||
fa49fd0f | 160 | /* Here the set of indices I with REG_VALUES(I) != 0 is saved. This is used |
eb232f4e | 161 | in cselib_clear_table() for fast emptying. */ |
6790d1ab JH |
162 | static unsigned int *used_regs; |
163 | static unsigned int n_used_regs; | |
fa49fd0f RK |
164 | |
165 | /* We pass this to cselib_invalidate_mem to invalidate all of | |
166 | memory for a non-const call instruction. */ | |
e2500fed | 167 | static GTY(()) rtx callmem; |
fa49fd0f | 168 | |
fa49fd0f RK |
169 | /* Set by discard_useless_locs if it deleted the last location of any |
170 | value. */ | |
171 | static int values_became_useless; | |
7101fb18 JH |
172 | |
173 | /* Used as stop element of the containing_mem list so we can check | |
174 | presence in the list by checking the next pointer. */ | |
175 | static cselib_val dummy_val; | |
176 | ||
457eeaae JJ |
177 | /* If non-NULL, value of the eliminated arg_pointer_rtx or frame_pointer_rtx |
178 | that is constant through the whole function and should never be | |
179 | eliminated. */ | |
180 | static cselib_val *cfa_base_preserved_val; | |
181 | ||
7080f735 | 182 | /* Used to list all values that contain memory reference. |
7101fb18 JH |
183 | May or may not contain the useless values - the list is compacted |
184 | each time memory is invalidated. */ | |
185 | static cselib_val *first_containing_mem = &dummy_val; | |
23bd7a93 | 186 | static alloc_pool elt_loc_list_pool, elt_list_pool, cselib_val_pool, value_pool; |
6fb5fa3c DB |
187 | |
188 | /* If nonnull, cselib will call this function before freeing useless | |
189 | VALUEs. A VALUE is deemed useless if its "locs" field is null. */ | |
190 | void (*cselib_discard_hook) (cselib_val *); | |
b5b8b0ac AO |
191 | |
192 | /* If nonnull, cselib will call this function before recording sets or | |
193 | even clobbering outputs of INSN. All the recorded sets will be | |
194 | represented in the array sets[n_sets]. new_val_min can be used to | |
195 | tell whether values present in sets are introduced by this | |
196 | instruction. */ | |
197 | void (*cselib_record_sets_hook) (rtx insn, struct cselib_set *sets, | |
198 | int n_sets); | |
199 | ||
200 | #define PRESERVED_VALUE_P(RTX) \ | |
201 | (RTL_FLAG_CHECK1("PRESERVED_VALUE_P", (RTX), VALUE)->unchanging) | |
b5b8b0ac | 202 | |
fa49fd0f RK |
203 | \f |
204 | ||
205 | /* Allocate a struct elt_list and fill in its two elements with the | |
206 | arguments. */ | |
207 | ||
6a59927d | 208 | static inline struct elt_list * |
7080f735 | 209 | new_elt_list (struct elt_list *next, cselib_val *elt) |
fa49fd0f | 210 | { |
6a59927d | 211 | struct elt_list *el; |
f883e0a7 | 212 | el = (struct elt_list *) pool_alloc (elt_list_pool); |
fa49fd0f RK |
213 | el->next = next; |
214 | el->elt = elt; | |
215 | return el; | |
216 | } | |
217 | ||
218 | /* Allocate a struct elt_loc_list and fill in its two elements with the | |
219 | arguments. */ | |
220 | ||
6a59927d | 221 | static inline struct elt_loc_list * |
7080f735 | 222 | new_elt_loc_list (struct elt_loc_list *next, rtx loc) |
fa49fd0f | 223 | { |
6a59927d | 224 | struct elt_loc_list *el; |
f883e0a7 | 225 | el = (struct elt_loc_list *) pool_alloc (elt_loc_list_pool); |
fa49fd0f RK |
226 | el->next = next; |
227 | el->loc = loc; | |
228 | el->setting_insn = cselib_current_insn; | |
5847e8da AO |
229 | gcc_assert (!next || !next->setting_insn |
230 | || !DEBUG_INSN_P (next->setting_insn)); | |
231 | ||
232 | /* If we're creating the first loc in a debug insn context, we've | |
233 | just created a debug value. Count it. */ | |
234 | if (!next && cselib_current_insn && DEBUG_INSN_P (cselib_current_insn)) | |
235 | n_debug_values++; | |
236 | ||
fa49fd0f RK |
237 | return el; |
238 | } | |
239 | ||
5847e8da AO |
240 | /* Promote loc L to a nondebug cselib_current_insn if L is marked as |
241 | originating from a debug insn, maintaining the debug values | |
242 | count. */ | |
243 | ||
244 | static inline void | |
245 | promote_debug_loc (struct elt_loc_list *l) | |
246 | { | |
247 | if (l->setting_insn && DEBUG_INSN_P (l->setting_insn) | |
248 | && (!cselib_current_insn || !DEBUG_INSN_P (cselib_current_insn))) | |
249 | { | |
250 | n_debug_values--; | |
251 | l->setting_insn = cselib_current_insn; | |
252 | gcc_assert (!l->next); | |
253 | } | |
254 | } | |
255 | ||
fa49fd0f RK |
256 | /* The elt_list at *PL is no longer needed. Unchain it and free its |
257 | storage. */ | |
258 | ||
6a59927d | 259 | static inline void |
7080f735 | 260 | unchain_one_elt_list (struct elt_list **pl) |
fa49fd0f RK |
261 | { |
262 | struct elt_list *l = *pl; | |
263 | ||
264 | *pl = l->next; | |
6a59927d | 265 | pool_free (elt_list_pool, l); |
fa49fd0f RK |
266 | } |
267 | ||
268 | /* Likewise for elt_loc_lists. */ | |
269 | ||
270 | static void | |
7080f735 | 271 | unchain_one_elt_loc_list (struct elt_loc_list **pl) |
fa49fd0f RK |
272 | { |
273 | struct elt_loc_list *l = *pl; | |
274 | ||
275 | *pl = l->next; | |
6a59927d | 276 | pool_free (elt_loc_list_pool, l); |
fa49fd0f RK |
277 | } |
278 | ||
279 | /* Likewise for cselib_vals. This also frees the addr_list associated with | |
280 | V. */ | |
281 | ||
282 | static void | |
7080f735 | 283 | unchain_one_value (cselib_val *v) |
fa49fd0f RK |
284 | { |
285 | while (v->addr_list) | |
286 | unchain_one_elt_list (&v->addr_list); | |
287 | ||
6a59927d | 288 | pool_free (cselib_val_pool, v); |
fa49fd0f RK |
289 | } |
290 | ||
291 | /* Remove all entries from the hash table. Also used during | |
b5b8b0ac | 292 | initialization. */ |
fa49fd0f | 293 | |
eb232f4e SB |
294 | void |
295 | cselib_clear_table (void) | |
b5b8b0ac | 296 | { |
5440c0e7 | 297 | cselib_reset_table (1); |
b5b8b0ac AO |
298 | } |
299 | ||
457eeaae JJ |
300 | /* Remove from hash table all VALUEs except constants. */ |
301 | ||
302 | static int | |
303 | preserve_only_constants (void **x, void *info ATTRIBUTE_UNUSED) | |
304 | { | |
305 | cselib_val *v = (cselib_val *)*x; | |
306 | ||
307 | if (v->locs != NULL | |
308 | && v->locs->next == NULL) | |
309 | { | |
310 | if (CONSTANT_P (v->locs->loc) | |
311 | && (GET_CODE (v->locs->loc) != CONST | |
312 | || !references_value_p (v->locs->loc, 0))) | |
313 | return 1; | |
314 | if (cfa_base_preserved_val) | |
315 | { | |
316 | if (v == cfa_base_preserved_val) | |
317 | return 1; | |
318 | if (GET_CODE (v->locs->loc) == PLUS | |
319 | && CONST_INT_P (XEXP (v->locs->loc, 1)) | |
320 | && XEXP (v->locs->loc, 0) == cfa_base_preserved_val->val_rtx) | |
321 | return 1; | |
322 | } | |
323 | } | |
324 | ||
325 | htab_clear_slot (cselib_hash_table, x); | |
326 | return 1; | |
327 | } | |
328 | ||
b5b8b0ac AO |
329 | /* Remove all entries from the hash table, arranging for the next |
330 | value to be numbered NUM. */ | |
331 | ||
332 | void | |
5440c0e7 | 333 | cselib_reset_table (unsigned int num) |
fa49fd0f RK |
334 | { |
335 | unsigned int i; | |
336 | ||
31825e57 DM |
337 | max_value_regs = 0; |
338 | ||
457eeaae JJ |
339 | if (cfa_base_preserved_val) |
340 | { | |
341 | unsigned int regno = REGNO (cfa_base_preserved_val->locs->loc); | |
342 | unsigned int new_used_regs = 0; | |
343 | for (i = 0; i < n_used_regs; i++) | |
344 | if (used_regs[i] == regno) | |
345 | { | |
346 | new_used_regs = 1; | |
347 | continue; | |
348 | } | |
349 | else | |
350 | REG_VALUES (used_regs[i]) = 0; | |
351 | gcc_assert (new_used_regs == 1); | |
352 | n_used_regs = new_used_regs; | |
353 | used_regs[0] = regno; | |
354 | max_value_regs | |
355 | = hard_regno_nregs[regno][GET_MODE (cfa_base_preserved_val->locs->loc)]; | |
356 | } | |
357 | else | |
358 | { | |
359 | for (i = 0; i < n_used_regs; i++) | |
360 | REG_VALUES (used_regs[i]) = 0; | |
361 | n_used_regs = 0; | |
362 | } | |
fa49fd0f | 363 | |
457eeaae JJ |
364 | if (cselib_preserve_constants) |
365 | htab_traverse (cselib_hash_table, preserve_only_constants, NULL); | |
366 | else | |
367 | htab_empty (cselib_hash_table); | |
fa49fd0f | 368 | |
fa49fd0f | 369 | n_useless_values = 0; |
5847e8da AO |
370 | n_useless_debug_values = 0; |
371 | n_debug_values = 0; | |
fa49fd0f | 372 | |
5440c0e7 | 373 | next_uid = num; |
7101fb18 JH |
374 | |
375 | first_containing_mem = &dummy_val; | |
fa49fd0f RK |
376 | } |
377 | ||
b5b8b0ac AO |
378 | /* Return the number of the next value that will be generated. */ |
379 | ||
380 | unsigned int | |
5440c0e7 | 381 | cselib_get_next_uid (void) |
b5b8b0ac | 382 | { |
5440c0e7 | 383 | return next_uid; |
b5b8b0ac AO |
384 | } |
385 | ||
fa49fd0f RK |
386 | /* The equality test for our hash table. The first argument ENTRY is a table |
387 | element (i.e. a cselib_val), while the second arg X is an rtx. We know | |
388 | that all callers of htab_find_slot_with_hash will wrap CONST_INTs into a | |
389 | CONST of an appropriate mode. */ | |
390 | ||
391 | static int | |
7080f735 | 392 | entry_and_rtx_equal_p (const void *entry, const void *x_arg) |
fa49fd0f RK |
393 | { |
394 | struct elt_loc_list *l; | |
e5cfc29f | 395 | const cselib_val *const v = (const cselib_val *) entry; |
f883e0a7 | 396 | rtx x = CONST_CAST_RTX ((const_rtx)x_arg); |
fa49fd0f RK |
397 | enum machine_mode mode = GET_MODE (x); |
398 | ||
481683e1 | 399 | gcc_assert (!CONST_INT_P (x) && GET_CODE (x) != CONST_FIXED |
341c100f | 400 | && (mode != VOIDmode || GET_CODE (x) != CONST_DOUBLE)); |
b8698a0f | 401 | |
757bbef8 | 402 | if (mode != GET_MODE (v->val_rtx)) |
fa49fd0f RK |
403 | return 0; |
404 | ||
405 | /* Unwrap X if necessary. */ | |
406 | if (GET_CODE (x) == CONST | |
481683e1 | 407 | && (CONST_INT_P (XEXP (x, 0)) |
091a3ac7 | 408 | || GET_CODE (XEXP (x, 0)) == CONST_FIXED |
fa49fd0f RK |
409 | || GET_CODE (XEXP (x, 0)) == CONST_DOUBLE)) |
410 | x = XEXP (x, 0); | |
7080f735 | 411 | |
fa49fd0f RK |
412 | /* We don't guarantee that distinct rtx's have different hash values, |
413 | so we need to do a comparison. */ | |
414 | for (l = v->locs; l; l = l->next) | |
415 | if (rtx_equal_for_cselib_p (l->loc, x)) | |
5847e8da AO |
416 | { |
417 | promote_debug_loc (l); | |
418 | return 1; | |
419 | } | |
fa49fd0f RK |
420 | |
421 | return 0; | |
422 | } | |
423 | ||
424 | /* The hash function for our hash table. The value is always computed with | |
0516f6fe SB |
425 | cselib_hash_rtx when adding an element; this function just extracts the |
426 | hash value from a cselib_val structure. */ | |
fa49fd0f | 427 | |
fb7e6024 | 428 | static hashval_t |
7080f735 | 429 | get_value_hash (const void *entry) |
fa49fd0f | 430 | { |
4f588890 | 431 | const cselib_val *const v = (const cselib_val *) entry; |
5440c0e7 | 432 | return v->hash; |
fa49fd0f RK |
433 | } |
434 | ||
435 | /* Return true if X contains a VALUE rtx. If ONLY_USELESS is set, we | |
436 | only return true for values which point to a cselib_val whose value | |
437 | element has been set to zero, which implies the cselib_val will be | |
438 | removed. */ | |
439 | ||
440 | int | |
4f588890 | 441 | references_value_p (const_rtx x, int only_useless) |
fa49fd0f | 442 | { |
4f588890 | 443 | const enum rtx_code code = GET_CODE (x); |
fa49fd0f RK |
444 | const char *fmt = GET_RTX_FORMAT (code); |
445 | int i, j; | |
446 | ||
447 | if (GET_CODE (x) == VALUE | |
448 | && (! only_useless || CSELIB_VAL_PTR (x)->locs == 0)) | |
449 | return 1; | |
450 | ||
451 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
452 | { | |
453 | if (fmt[i] == 'e' && references_value_p (XEXP (x, i), only_useless)) | |
454 | return 1; | |
455 | else if (fmt[i] == 'E') | |
456 | for (j = 0; j < XVECLEN (x, i); j++) | |
457 | if (references_value_p (XVECEXP (x, i, j), only_useless)) | |
458 | return 1; | |
459 | } | |
460 | ||
461 | return 0; | |
462 | } | |
463 | ||
464 | /* For all locations found in X, delete locations that reference useless | |
465 | values (i.e. values without any location). Called through | |
466 | htab_traverse. */ | |
467 | ||
468 | static int | |
7080f735 | 469 | discard_useless_locs (void **x, void *info ATTRIBUTE_UNUSED) |
fa49fd0f RK |
470 | { |
471 | cselib_val *v = (cselib_val *)*x; | |
472 | struct elt_loc_list **p = &v->locs; | |
5847e8da AO |
473 | bool had_locs = v->locs != NULL; |
474 | rtx setting_insn = v->locs ? v->locs->setting_insn : NULL; | |
fa49fd0f RK |
475 | |
476 | while (*p) | |
477 | { | |
478 | if (references_value_p ((*p)->loc, 1)) | |
479 | unchain_one_elt_loc_list (p); | |
480 | else | |
481 | p = &(*p)->next; | |
482 | } | |
483 | ||
b5b8b0ac | 484 | if (had_locs && v->locs == 0 && !PRESERVED_VALUE_P (v->val_rtx)) |
fa49fd0f | 485 | { |
5847e8da AO |
486 | if (setting_insn && DEBUG_INSN_P (setting_insn)) |
487 | n_useless_debug_values++; | |
488 | else | |
489 | n_useless_values++; | |
fa49fd0f RK |
490 | values_became_useless = 1; |
491 | } | |
492 | return 1; | |
493 | } | |
494 | ||
495 | /* If X is a value with no locations, remove it from the hashtable. */ | |
496 | ||
497 | static int | |
7080f735 | 498 | discard_useless_values (void **x, void *info ATTRIBUTE_UNUSED) |
fa49fd0f RK |
499 | { |
500 | cselib_val *v = (cselib_val *)*x; | |
501 | ||
b5b8b0ac | 502 | if (v->locs == 0 && !PRESERVED_VALUE_P (v->val_rtx)) |
fa49fd0f | 503 | { |
6fb5fa3c DB |
504 | if (cselib_discard_hook) |
505 | cselib_discard_hook (v); | |
506 | ||
757bbef8 | 507 | CSELIB_VAL_PTR (v->val_rtx) = NULL; |
7c514720 | 508 | htab_clear_slot (cselib_hash_table, x); |
fa49fd0f RK |
509 | unchain_one_value (v); |
510 | n_useless_values--; | |
511 | } | |
512 | ||
513 | return 1; | |
514 | } | |
515 | ||
516 | /* Clean out useless values (i.e. those which no longer have locations | |
517 | associated with them) from the hash table. */ | |
518 | ||
519 | static void | |
7080f735 | 520 | remove_useless_values (void) |
fa49fd0f | 521 | { |
7101fb18 | 522 | cselib_val **p, *v; |
5847e8da | 523 | |
fa49fd0f RK |
524 | /* First pass: eliminate locations that reference the value. That in |
525 | turn can make more values useless. */ | |
526 | do | |
527 | { | |
528 | values_became_useless = 0; | |
7c514720 | 529 | htab_traverse (cselib_hash_table, discard_useless_locs, 0); |
fa49fd0f RK |
530 | } |
531 | while (values_became_useless); | |
532 | ||
533 | /* Second pass: actually remove the values. */ | |
fa49fd0f | 534 | |
7101fb18 JH |
535 | p = &first_containing_mem; |
536 | for (v = *p; v != &dummy_val; v = v->next_containing_mem) | |
537 | if (v->locs) | |
538 | { | |
539 | *p = v; | |
540 | p = &(*p)->next_containing_mem; | |
541 | } | |
542 | *p = &dummy_val; | |
543 | ||
5847e8da AO |
544 | n_useless_values += n_useless_debug_values; |
545 | n_debug_values -= n_useless_debug_values; | |
546 | n_useless_debug_values = 0; | |
547 | ||
7c514720 | 548 | htab_traverse (cselib_hash_table, discard_useless_values, 0); |
3e2a0bd2 | 549 | |
341c100f | 550 | gcc_assert (!n_useless_values); |
fa49fd0f RK |
551 | } |
552 | ||
b5b8b0ac AO |
553 | /* Arrange for a value to not be removed from the hash table even if |
554 | it becomes useless. */ | |
555 | ||
556 | void | |
557 | cselib_preserve_value (cselib_val *v) | |
558 | { | |
559 | PRESERVED_VALUE_P (v->val_rtx) = 1; | |
560 | } | |
561 | ||
562 | /* Test whether a value is preserved. */ | |
563 | ||
564 | bool | |
565 | cselib_preserved_value_p (cselib_val *v) | |
566 | { | |
567 | return PRESERVED_VALUE_P (v->val_rtx); | |
568 | } | |
569 | ||
457eeaae JJ |
570 | /* Arrange for a REG value to be assumed constant through the whole function, |
571 | never invalidated and preserved across cselib_reset_table calls. */ | |
572 | ||
573 | void | |
574 | cselib_preserve_cfa_base_value (cselib_val *v) | |
575 | { | |
576 | if (cselib_preserve_constants | |
577 | && v->locs | |
578 | && REG_P (v->locs->loc)) | |
579 | cfa_base_preserved_val = v; | |
580 | } | |
581 | ||
b5b8b0ac AO |
582 | /* Clean all non-constant expressions in the hash table, but retain |
583 | their values. */ | |
584 | ||
585 | void | |
0de3e43f | 586 | cselib_preserve_only_values (void) |
b5b8b0ac AO |
587 | { |
588 | int i; | |
589 | ||
b5b8b0ac AO |
590 | for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) |
591 | cselib_invalidate_regno (i, reg_raw_mode[i]); | |
592 | ||
593 | cselib_invalidate_mem (callmem); | |
594 | ||
595 | remove_useless_values (); | |
596 | ||
597 | gcc_assert (first_containing_mem == &dummy_val); | |
598 | } | |
599 | ||
60fa6660 AO |
600 | /* Return the mode in which a register was last set. If X is not a |
601 | register, return its mode. If the mode in which the register was | |
602 | set is not known, or the value was already clobbered, return | |
603 | VOIDmode. */ | |
604 | ||
605 | enum machine_mode | |
4f588890 | 606 | cselib_reg_set_mode (const_rtx x) |
60fa6660 | 607 | { |
f8cfc6aa | 608 | if (!REG_P (x)) |
60fa6660 AO |
609 | return GET_MODE (x); |
610 | ||
611 | if (REG_VALUES (REGNO (x)) == NULL | |
612 | || REG_VALUES (REGNO (x))->elt == NULL) | |
613 | return VOIDmode; | |
614 | ||
757bbef8 | 615 | return GET_MODE (REG_VALUES (REGNO (x))->elt->val_rtx); |
60fa6660 AO |
616 | } |
617 | ||
fa49fd0f RK |
618 | /* Return nonzero if we can prove that X and Y contain the same value, taking |
619 | our gathered information into account. */ | |
620 | ||
621 | int | |
7080f735 | 622 | rtx_equal_for_cselib_p (rtx x, rtx y) |
fa49fd0f RK |
623 | { |
624 | enum rtx_code code; | |
625 | const char *fmt; | |
626 | int i; | |
7080f735 | 627 | |
f8cfc6aa | 628 | if (REG_P (x) || MEM_P (x)) |
fa49fd0f RK |
629 | { |
630 | cselib_val *e = cselib_lookup (x, GET_MODE (x), 0); | |
631 | ||
632 | if (e) | |
757bbef8 | 633 | x = e->val_rtx; |
fa49fd0f RK |
634 | } |
635 | ||
f8cfc6aa | 636 | if (REG_P (y) || MEM_P (y)) |
fa49fd0f RK |
637 | { |
638 | cselib_val *e = cselib_lookup (y, GET_MODE (y), 0); | |
639 | ||
640 | if (e) | |
757bbef8 | 641 | y = e->val_rtx; |
fa49fd0f RK |
642 | } |
643 | ||
644 | if (x == y) | |
645 | return 1; | |
646 | ||
647 | if (GET_CODE (x) == VALUE && GET_CODE (y) == VALUE) | |
648 | return CSELIB_VAL_PTR (x) == CSELIB_VAL_PTR (y); | |
649 | ||
650 | if (GET_CODE (x) == VALUE) | |
651 | { | |
652 | cselib_val *e = CSELIB_VAL_PTR (x); | |
653 | struct elt_loc_list *l; | |
654 | ||
655 | for (l = e->locs; l; l = l->next) | |
656 | { | |
657 | rtx t = l->loc; | |
658 | ||
659 | /* Avoid infinite recursion. */ | |
3c0cb5de | 660 | if (REG_P (t) || MEM_P (t)) |
fa49fd0f RK |
661 | continue; |
662 | else if (rtx_equal_for_cselib_p (t, y)) | |
663 | return 1; | |
664 | } | |
7080f735 | 665 | |
fa49fd0f RK |
666 | return 0; |
667 | } | |
668 | ||
669 | if (GET_CODE (y) == VALUE) | |
670 | { | |
671 | cselib_val *e = CSELIB_VAL_PTR (y); | |
672 | struct elt_loc_list *l; | |
673 | ||
674 | for (l = e->locs; l; l = l->next) | |
675 | { | |
676 | rtx t = l->loc; | |
677 | ||
3c0cb5de | 678 | if (REG_P (t) || MEM_P (t)) |
fa49fd0f RK |
679 | continue; |
680 | else if (rtx_equal_for_cselib_p (x, t)) | |
681 | return 1; | |
682 | } | |
7080f735 | 683 | |
fa49fd0f RK |
684 | return 0; |
685 | } | |
686 | ||
687 | if (GET_CODE (x) != GET_CODE (y) || GET_MODE (x) != GET_MODE (y)) | |
688 | return 0; | |
689 | ||
37cf6116 RH |
690 | /* These won't be handled correctly by the code below. */ |
691 | switch (GET_CODE (x)) | |
692 | { | |
693 | case CONST_DOUBLE: | |
091a3ac7 | 694 | case CONST_FIXED: |
0ca5af51 | 695 | case DEBUG_EXPR: |
37cf6116 RH |
696 | return 0; |
697 | ||
698 | case LABEL_REF: | |
699 | return XEXP (x, 0) == XEXP (y, 0); | |
700 | ||
701 | default: | |
702 | break; | |
703 | } | |
7080f735 | 704 | |
fa49fd0f RK |
705 | code = GET_CODE (x); |
706 | fmt = GET_RTX_FORMAT (code); | |
707 | ||
708 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
709 | { | |
710 | int j; | |
711 | ||
712 | switch (fmt[i]) | |
713 | { | |
714 | case 'w': | |
715 | if (XWINT (x, i) != XWINT (y, i)) | |
716 | return 0; | |
717 | break; | |
718 | ||
719 | case 'n': | |
720 | case 'i': | |
721 | if (XINT (x, i) != XINT (y, i)) | |
722 | return 0; | |
723 | break; | |
724 | ||
725 | case 'V': | |
726 | case 'E': | |
727 | /* Two vectors must have the same length. */ | |
728 | if (XVECLEN (x, i) != XVECLEN (y, i)) | |
729 | return 0; | |
730 | ||
731 | /* And the corresponding elements must match. */ | |
732 | for (j = 0; j < XVECLEN (x, i); j++) | |
733 | if (! rtx_equal_for_cselib_p (XVECEXP (x, i, j), | |
734 | XVECEXP (y, i, j))) | |
735 | return 0; | |
736 | break; | |
737 | ||
738 | case 'e': | |
29c1846b R |
739 | if (i == 1 |
740 | && targetm.commutative_p (x, UNKNOWN) | |
741 | && rtx_equal_for_cselib_p (XEXP (x, 1), XEXP (y, 0)) | |
742 | && rtx_equal_for_cselib_p (XEXP (x, 0), XEXP (y, 1))) | |
743 | return 1; | |
fa49fd0f RK |
744 | if (! rtx_equal_for_cselib_p (XEXP (x, i), XEXP (y, i))) |
745 | return 0; | |
746 | break; | |
747 | ||
748 | case 'S': | |
749 | case 's': | |
750 | if (strcmp (XSTR (x, i), XSTR (y, i))) | |
751 | return 0; | |
752 | break; | |
753 | ||
754 | case 'u': | |
755 | /* These are just backpointers, so they don't matter. */ | |
756 | break; | |
757 | ||
758 | case '0': | |
759 | case 't': | |
760 | break; | |
761 | ||
762 | /* It is believed that rtx's at this level will never | |
763 | contain anything but integers and other rtx's, | |
764 | except for within LABEL_REFs and SYMBOL_REFs. */ | |
765 | default: | |
341c100f | 766 | gcc_unreachable (); |
fa49fd0f RK |
767 | } |
768 | } | |
769 | return 1; | |
770 | } | |
771 | ||
3af4ba41 RS |
772 | /* We need to pass down the mode of constants through the hash table |
773 | functions. For that purpose, wrap them in a CONST of the appropriate | |
774 | mode. */ | |
775 | static rtx | |
776 | wrap_constant (enum machine_mode mode, rtx x) | |
777 | { | |
778 | if (!CONST_INT_P (x) && GET_CODE (x) != CONST_FIXED | |
779 | && (GET_CODE (x) != CONST_DOUBLE || GET_MODE (x) != VOIDmode)) | |
780 | return x; | |
781 | gcc_assert (mode != VOIDmode); | |
782 | return gen_rtx_CONST (mode, x); | |
783 | } | |
784 | ||
fa49fd0f RK |
785 | /* Hash an rtx. Return 0 if we couldn't hash the rtx. |
786 | For registers and memory locations, we look up their cselib_val structure | |
787 | and return its VALUE element. | |
788 | Possible reasons for return 0 are: the object is volatile, or we couldn't | |
789 | find a register or memory location in the table and CREATE is zero. If | |
790 | CREATE is nonzero, table elts are created for regs and mem. | |
29c1846b R |
791 | N.B. this hash function returns the same hash value for RTXes that |
792 | differ only in the order of operands, thus it is suitable for comparisons | |
793 | that take commutativity into account. | |
794 | If we wanted to also support associative rules, we'd have to use a different | |
795 | strategy to avoid returning spurious 0, e.g. return ~(~0U >> 1) . | |
796 | We used to have a MODE argument for hashing for CONST_INTs, but that | |
797 | didn't make sense, since it caused spurious hash differences between | |
798 | (set (reg:SI 1) (const_int)) | |
799 | (plus:SI (reg:SI 2) (reg:SI 1)) | |
800 | and | |
801 | (plus:SI (reg:SI 2) (const_int)) | |
802 | If the mode is important in any context, it must be checked specifically | |
803 | in a comparison anyway, since relying on hash differences is unsafe. */ | |
fa49fd0f RK |
804 | |
805 | static unsigned int | |
29c1846b | 806 | cselib_hash_rtx (rtx x, int create) |
fa49fd0f RK |
807 | { |
808 | cselib_val *e; | |
809 | int i, j; | |
810 | enum rtx_code code; | |
811 | const char *fmt; | |
812 | unsigned int hash = 0; | |
813 | ||
fa49fd0f RK |
814 | code = GET_CODE (x); |
815 | hash += (unsigned) code + (unsigned) GET_MODE (x); | |
816 | ||
817 | switch (code) | |
818 | { | |
819 | case MEM: | |
820 | case REG: | |
821 | e = cselib_lookup (x, GET_MODE (x), create); | |
822 | if (! e) | |
823 | return 0; | |
824 | ||
5440c0e7 | 825 | return e->hash; |
fa49fd0f | 826 | |
0ca5af51 | 827 | case DEBUG_EXPR: |
e4fb38bd JJ |
828 | hash += ((unsigned) DEBUG_EXPR << 7) |
829 | + DEBUG_TEMP_UID (DEBUG_EXPR_TREE_DECL (x)); | |
0ca5af51 AO |
830 | return hash ? hash : (unsigned int) DEBUG_EXPR; |
831 | ||
fa49fd0f | 832 | case CONST_INT: |
29c1846b | 833 | hash += ((unsigned) CONST_INT << 7) + INTVAL (x); |
dc76f41c | 834 | return hash ? hash : (unsigned int) CONST_INT; |
fa49fd0f RK |
835 | |
836 | case CONST_DOUBLE: | |
837 | /* This is like the general case, except that it only counts | |
838 | the integers representing the constant. */ | |
839 | hash += (unsigned) code + (unsigned) GET_MODE (x); | |
840 | if (GET_MODE (x) != VOIDmode) | |
46b33600 | 841 | hash += real_hash (CONST_DOUBLE_REAL_VALUE (x)); |
fa49fd0f RK |
842 | else |
843 | hash += ((unsigned) CONST_DOUBLE_LOW (x) | |
844 | + (unsigned) CONST_DOUBLE_HIGH (x)); | |
dc76f41c | 845 | return hash ? hash : (unsigned int) CONST_DOUBLE; |
fa49fd0f | 846 | |
091a3ac7 CF |
847 | case CONST_FIXED: |
848 | hash += (unsigned int) code + (unsigned int) GET_MODE (x); | |
849 | hash += fixed_hash (CONST_FIXED_VALUE (x)); | |
850 | return hash ? hash : (unsigned int) CONST_FIXED; | |
851 | ||
69ef87e2 AH |
852 | case CONST_VECTOR: |
853 | { | |
854 | int units; | |
855 | rtx elt; | |
856 | ||
857 | units = CONST_VECTOR_NUNITS (x); | |
858 | ||
859 | for (i = 0; i < units; ++i) | |
860 | { | |
861 | elt = CONST_VECTOR_ELT (x, i); | |
29c1846b | 862 | hash += cselib_hash_rtx (elt, 0); |
69ef87e2 AH |
863 | } |
864 | ||
865 | return hash; | |
866 | } | |
867 | ||
fa49fd0f RK |
868 | /* Assume there is only one rtx object for any given label. */ |
869 | case LABEL_REF: | |
4c6669c2 RS |
870 | /* We don't hash on the address of the CODE_LABEL to avoid bootstrap |
871 | differences and differences between each stage's debugging dumps. */ | |
872 | hash += (((unsigned int) LABEL_REF << 7) | |
873 | + CODE_LABEL_NUMBER (XEXP (x, 0))); | |
dc76f41c | 874 | return hash ? hash : (unsigned int) LABEL_REF; |
fa49fd0f RK |
875 | |
876 | case SYMBOL_REF: | |
4c6669c2 RS |
877 | { |
878 | /* Don't hash on the symbol's address to avoid bootstrap differences. | |
879 | Different hash values may cause expressions to be recorded in | |
880 | different orders and thus different registers to be used in the | |
881 | final assembler. This also avoids differences in the dump files | |
882 | between various stages. */ | |
883 | unsigned int h = 0; | |
884 | const unsigned char *p = (const unsigned char *) XSTR (x, 0); | |
885 | ||
886 | while (*p) | |
887 | h += (h << 7) + *p++; /* ??? revisit */ | |
888 | ||
889 | hash += ((unsigned int) SYMBOL_REF << 7) + h; | |
890 | return hash ? hash : (unsigned int) SYMBOL_REF; | |
891 | } | |
fa49fd0f RK |
892 | |
893 | case PRE_DEC: | |
894 | case PRE_INC: | |
895 | case POST_DEC: | |
896 | case POST_INC: | |
897 | case POST_MODIFY: | |
898 | case PRE_MODIFY: | |
899 | case PC: | |
900 | case CC0: | |
901 | case CALL: | |
902 | case UNSPEC_VOLATILE: | |
903 | return 0; | |
904 | ||
905 | case ASM_OPERANDS: | |
906 | if (MEM_VOLATILE_P (x)) | |
907 | return 0; | |
908 | ||
909 | break; | |
7080f735 | 910 | |
fa49fd0f RK |
911 | default: |
912 | break; | |
913 | } | |
914 | ||
915 | i = GET_RTX_LENGTH (code) - 1; | |
916 | fmt = GET_RTX_FORMAT (code); | |
917 | for (; i >= 0; i--) | |
918 | { | |
341c100f | 919 | switch (fmt[i]) |
fa49fd0f | 920 | { |
341c100f | 921 | case 'e': |
fa49fd0f | 922 | { |
341c100f | 923 | rtx tem = XEXP (x, i); |
29c1846b | 924 | unsigned int tem_hash = cselib_hash_rtx (tem, create); |
b8698a0f | 925 | |
fa49fd0f RK |
926 | if (tem_hash == 0) |
927 | return 0; | |
b8698a0f | 928 | |
fa49fd0f RK |
929 | hash += tem_hash; |
930 | } | |
341c100f NS |
931 | break; |
932 | case 'E': | |
933 | for (j = 0; j < XVECLEN (x, i); j++) | |
934 | { | |
935 | unsigned int tem_hash | |
29c1846b | 936 | = cselib_hash_rtx (XVECEXP (x, i, j), create); |
b8698a0f | 937 | |
341c100f NS |
938 | if (tem_hash == 0) |
939 | return 0; | |
b8698a0f | 940 | |
341c100f NS |
941 | hash += tem_hash; |
942 | } | |
943 | break; | |
fa49fd0f | 944 | |
341c100f NS |
945 | case 's': |
946 | { | |
947 | const unsigned char *p = (const unsigned char *) XSTR (x, i); | |
b8698a0f | 948 | |
341c100f NS |
949 | if (p) |
950 | while (*p) | |
951 | hash += *p++; | |
952 | break; | |
953 | } | |
b8698a0f | 954 | |
341c100f NS |
955 | case 'i': |
956 | hash += XINT (x, i); | |
957 | break; | |
958 | ||
959 | case '0': | |
960 | case 't': | |
961 | /* unused */ | |
962 | break; | |
b8698a0f | 963 | |
341c100f NS |
964 | default: |
965 | gcc_unreachable (); | |
fa49fd0f | 966 | } |
fa49fd0f RK |
967 | } |
968 | ||
dc76f41c | 969 | return hash ? hash : 1 + (unsigned int) GET_CODE (x); |
fa49fd0f RK |
970 | } |
971 | ||
972 | /* Create a new value structure for VALUE and initialize it. The mode of the | |
973 | value is MODE. */ | |
974 | ||
6a59927d | 975 | static inline cselib_val * |
5440c0e7 | 976 | new_cselib_val (unsigned int hash, enum machine_mode mode, rtx x) |
fa49fd0f | 977 | { |
f883e0a7 | 978 | cselib_val *e = (cselib_val *) pool_alloc (cselib_val_pool); |
fa49fd0f | 979 | |
5440c0e7 AO |
980 | gcc_assert (hash); |
981 | gcc_assert (next_uid); | |
fa49fd0f | 982 | |
5440c0e7 AO |
983 | e->hash = hash; |
984 | e->uid = next_uid++; | |
d67fb775 SB |
985 | /* We use an alloc pool to allocate this RTL construct because it |
986 | accounts for about 8% of the overall memory usage. We know | |
987 | precisely when we can have VALUE RTXen (when cselib is active) | |
daa956d0 | 988 | so we don't need to put them in garbage collected memory. |
d67fb775 | 989 | ??? Why should a VALUE be an RTX in the first place? */ |
f883e0a7 | 990 | e->val_rtx = (rtx) pool_alloc (value_pool); |
757bbef8 SB |
991 | memset (e->val_rtx, 0, RTX_HDR_SIZE); |
992 | PUT_CODE (e->val_rtx, VALUE); | |
993 | PUT_MODE (e->val_rtx, mode); | |
994 | CSELIB_VAL_PTR (e->val_rtx) = e; | |
fa49fd0f RK |
995 | e->addr_list = 0; |
996 | e->locs = 0; | |
7101fb18 | 997 | e->next_containing_mem = 0; |
b5b8b0ac AO |
998 | |
999 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1000 | { | |
5440c0e7 | 1001 | fprintf (dump_file, "cselib value %u:%u ", e->uid, hash); |
b5b8b0ac AO |
1002 | if (flag_dump_noaddr || flag_dump_unnumbered) |
1003 | fputs ("# ", dump_file); | |
1004 | else | |
1005 | fprintf (dump_file, "%p ", (void*)e); | |
1006 | print_rtl_single (dump_file, x); | |
1007 | fputc ('\n', dump_file); | |
1008 | } | |
1009 | ||
fa49fd0f RK |
1010 | return e; |
1011 | } | |
1012 | ||
1013 | /* ADDR_ELT is a value that is used as address. MEM_ELT is the value that | |
1014 | contains the data at this address. X is a MEM that represents the | |
1015 | value. Update the two value structures to represent this situation. */ | |
1016 | ||
1017 | static void | |
7080f735 | 1018 | add_mem_for_addr (cselib_val *addr_elt, cselib_val *mem_elt, rtx x) |
fa49fd0f | 1019 | { |
fa49fd0f RK |
1020 | struct elt_loc_list *l; |
1021 | ||
1022 | /* Avoid duplicates. */ | |
1023 | for (l = mem_elt->locs; l; l = l->next) | |
3c0cb5de | 1024 | if (MEM_P (l->loc) |
fa49fd0f | 1025 | && CSELIB_VAL_PTR (XEXP (l->loc, 0)) == addr_elt) |
5847e8da AO |
1026 | { |
1027 | promote_debug_loc (l); | |
1028 | return; | |
1029 | } | |
fa49fd0f | 1030 | |
fa49fd0f | 1031 | addr_elt->addr_list = new_elt_list (addr_elt->addr_list, mem_elt); |
f1ec5147 RK |
1032 | mem_elt->locs |
1033 | = new_elt_loc_list (mem_elt->locs, | |
757bbef8 | 1034 | replace_equiv_address_nv (x, addr_elt->val_rtx)); |
7101fb18 JH |
1035 | if (mem_elt->next_containing_mem == NULL) |
1036 | { | |
1037 | mem_elt->next_containing_mem = first_containing_mem; | |
1038 | first_containing_mem = mem_elt; | |
1039 | } | |
fa49fd0f RK |
1040 | } |
1041 | ||
1042 | /* Subroutine of cselib_lookup. Return a value for X, which is a MEM rtx. | |
1043 | If CREATE, make a new one if we haven't seen it before. */ | |
1044 | ||
1045 | static cselib_val * | |
7080f735 | 1046 | cselib_lookup_mem (rtx x, int create) |
fa49fd0f RK |
1047 | { |
1048 | enum machine_mode mode = GET_MODE (x); | |
1049 | void **slot; | |
1050 | cselib_val *addr; | |
1051 | cselib_val *mem_elt; | |
1052 | struct elt_list *l; | |
1053 | ||
1054 | if (MEM_VOLATILE_P (x) || mode == BLKmode | |
463301c3 | 1055 | || !cselib_record_memory |
fa49fd0f RK |
1056 | || (FLOAT_MODE_P (mode) && flag_float_store)) |
1057 | return 0; | |
1058 | ||
1059 | /* Look up the value for the address. */ | |
1060 | addr = cselib_lookup (XEXP (x, 0), mode, create); | |
1061 | if (! addr) | |
1062 | return 0; | |
1063 | ||
1064 | /* Find a value that describes a value of our mode at that address. */ | |
1065 | for (l = addr->addr_list; l; l = l->next) | |
757bbef8 | 1066 | if (GET_MODE (l->elt->val_rtx) == mode) |
5847e8da AO |
1067 | { |
1068 | promote_debug_loc (l->elt->locs); | |
1069 | return l->elt; | |
1070 | } | |
fa49fd0f RK |
1071 | |
1072 | if (! create) | |
1073 | return 0; | |
1074 | ||
5440c0e7 | 1075 | mem_elt = new_cselib_val (next_uid, mode, x); |
fa49fd0f | 1076 | add_mem_for_addr (addr, mem_elt, x); |
7c514720 | 1077 | slot = htab_find_slot_with_hash (cselib_hash_table, wrap_constant (mode, x), |
5440c0e7 | 1078 | mem_elt->hash, INSERT); |
fa49fd0f RK |
1079 | *slot = mem_elt; |
1080 | return mem_elt; | |
1081 | } | |
1082 | ||
6fb5fa3c DB |
1083 | /* Search thru the possible substitutions in P. We prefer a non reg |
1084 | substitution because this allows us to expand the tree further. If | |
1085 | we find, just a reg, take the lowest regno. There may be several | |
1086 | non-reg results, we just take the first one because they will all | |
1087 | expand to the same place. */ | |
1088 | ||
b8698a0f | 1089 | static rtx |
b5b8b0ac AO |
1090 | expand_loc (struct elt_loc_list *p, struct expand_value_data *evd, |
1091 | int max_depth) | |
6fb5fa3c DB |
1092 | { |
1093 | rtx reg_result = NULL; | |
1094 | unsigned int regno = UINT_MAX; | |
1095 | struct elt_loc_list *p_in = p; | |
1096 | ||
1097 | for (; p; p = p -> next) | |
1098 | { | |
1099 | /* Avoid infinite recursion trying to expand a reg into a | |
1100 | the same reg. */ | |
b8698a0f L |
1101 | if ((REG_P (p->loc)) |
1102 | && (REGNO (p->loc) < regno) | |
b5b8b0ac | 1103 | && !bitmap_bit_p (evd->regs_active, REGNO (p->loc))) |
6fb5fa3c DB |
1104 | { |
1105 | reg_result = p->loc; | |
1106 | regno = REGNO (p->loc); | |
1107 | } | |
1108 | /* Avoid infinite recursion and do not try to expand the | |
1109 | value. */ | |
b8698a0f | 1110 | else if (GET_CODE (p->loc) == VALUE |
6fb5fa3c DB |
1111 | && CSELIB_VAL_PTR (p->loc)->locs == p_in) |
1112 | continue; | |
1113 | else if (!REG_P (p->loc)) | |
1114 | { | |
8dd5516b | 1115 | rtx result, note; |
b5b8b0ac | 1116 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
1117 | { |
1118 | print_inline_rtx (dump_file, p->loc, 0); | |
1119 | fprintf (dump_file, "\n"); | |
1120 | } | |
8dd5516b JJ |
1121 | if (GET_CODE (p->loc) == LO_SUM |
1122 | && GET_CODE (XEXP (p->loc, 1)) == SYMBOL_REF | |
1123 | && p->setting_insn | |
1124 | && (note = find_reg_note (p->setting_insn, REG_EQUAL, NULL_RTX)) | |
1125 | && XEXP (note, 0) == XEXP (p->loc, 1)) | |
1126 | return XEXP (p->loc, 1); | |
b5b8b0ac | 1127 | result = cselib_expand_value_rtx_1 (p->loc, evd, max_depth - 1); |
6fb5fa3c DB |
1128 | if (result) |
1129 | return result; | |
1130 | } | |
b8698a0f | 1131 | |
6fb5fa3c | 1132 | } |
b8698a0f | 1133 | |
6fb5fa3c DB |
1134 | if (regno != UINT_MAX) |
1135 | { | |
1136 | rtx result; | |
b5b8b0ac | 1137 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
1138 | fprintf (dump_file, "r%d\n", regno); |
1139 | ||
b5b8b0ac | 1140 | result = cselib_expand_value_rtx_1 (reg_result, evd, max_depth - 1); |
6fb5fa3c DB |
1141 | if (result) |
1142 | return result; | |
1143 | } | |
1144 | ||
b5b8b0ac | 1145 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
1146 | { |
1147 | if (reg_result) | |
1148 | { | |
1149 | print_inline_rtx (dump_file, reg_result, 0); | |
1150 | fprintf (dump_file, "\n"); | |
1151 | } | |
b8698a0f | 1152 | else |
6fb5fa3c DB |
1153 | fprintf (dump_file, "NULL\n"); |
1154 | } | |
1155 | return reg_result; | |
1156 | } | |
1157 | ||
1158 | ||
1159 | /* Forward substitute and expand an expression out to its roots. | |
1160 | This is the opposite of common subexpression. Because local value | |
1161 | numbering is such a weak optimization, the expanded expression is | |
1162 | pretty much unique (not from a pointer equals point of view but | |
b8698a0f | 1163 | from a tree shape point of view. |
6fb5fa3c DB |
1164 | |
1165 | This function returns NULL if the expansion fails. The expansion | |
1166 | will fail if there is no value number for one of the operands or if | |
1167 | one of the operands has been overwritten between the current insn | |
1168 | and the beginning of the basic block. For instance x has no | |
1169 | expansion in: | |
1170 | ||
1171 | r1 <- r1 + 3 | |
1172 | x <- r1 + 8 | |
1173 | ||
1174 | REGS_ACTIVE is a scratch bitmap that should be clear when passing in. | |
1175 | It is clear on return. */ | |
1176 | ||
1177 | rtx | |
1178 | cselib_expand_value_rtx (rtx orig, bitmap regs_active, int max_depth) | |
b5b8b0ac AO |
1179 | { |
1180 | struct expand_value_data evd; | |
1181 | ||
1182 | evd.regs_active = regs_active; | |
1183 | evd.callback = NULL; | |
1184 | evd.callback_arg = NULL; | |
864ddef7 | 1185 | evd.dummy = false; |
b5b8b0ac AO |
1186 | |
1187 | return cselib_expand_value_rtx_1 (orig, &evd, max_depth); | |
1188 | } | |
1189 | ||
1190 | /* Same as cselib_expand_value_rtx, but using a callback to try to | |
0b7e34d7 AO |
1191 | resolve some expressions. The CB function should return ORIG if it |
1192 | can't or does not want to deal with a certain RTX. Any other | |
1193 | return value, including NULL, will be used as the expansion for | |
1194 | VALUE, without any further changes. */ | |
b5b8b0ac AO |
1195 | |
1196 | rtx | |
1197 | cselib_expand_value_rtx_cb (rtx orig, bitmap regs_active, int max_depth, | |
1198 | cselib_expand_callback cb, void *data) | |
1199 | { | |
1200 | struct expand_value_data evd; | |
1201 | ||
1202 | evd.regs_active = regs_active; | |
1203 | evd.callback = cb; | |
1204 | evd.callback_arg = data; | |
864ddef7 | 1205 | evd.dummy = false; |
b5b8b0ac AO |
1206 | |
1207 | return cselib_expand_value_rtx_1 (orig, &evd, max_depth); | |
1208 | } | |
1209 | ||
864ddef7 JJ |
1210 | /* Similar to cselib_expand_value_rtx_cb, but no rtxs are actually copied |
1211 | or simplified. Useful to find out whether cselib_expand_value_rtx_cb | |
1212 | would return NULL or non-NULL, without allocating new rtx. */ | |
1213 | ||
1214 | bool | |
1215 | cselib_dummy_expand_value_rtx_cb (rtx orig, bitmap regs_active, int max_depth, | |
1216 | cselib_expand_callback cb, void *data) | |
1217 | { | |
1218 | struct expand_value_data evd; | |
1219 | ||
1220 | evd.regs_active = regs_active; | |
1221 | evd.callback = cb; | |
1222 | evd.callback_arg = data; | |
1223 | evd.dummy = true; | |
1224 | ||
1225 | return cselib_expand_value_rtx_1 (orig, &evd, max_depth) != NULL; | |
1226 | } | |
1227 | ||
0b7e34d7 AO |
1228 | /* Internal implementation of cselib_expand_value_rtx and |
1229 | cselib_expand_value_rtx_cb. */ | |
1230 | ||
b5b8b0ac AO |
1231 | static rtx |
1232 | cselib_expand_value_rtx_1 (rtx orig, struct expand_value_data *evd, | |
1233 | int max_depth) | |
6fb5fa3c DB |
1234 | { |
1235 | rtx copy, scopy; | |
1236 | int i, j; | |
1237 | RTX_CODE code; | |
1238 | const char *format_ptr; | |
8dd5516b | 1239 | enum machine_mode mode; |
6fb5fa3c DB |
1240 | |
1241 | code = GET_CODE (orig); | |
1242 | ||
1243 | /* For the context of dse, if we end up expand into a huge tree, we | |
1244 | will not have a useful address, so we might as well just give up | |
1245 | quickly. */ | |
1246 | if (max_depth <= 0) | |
1247 | return NULL; | |
1248 | ||
1249 | switch (code) | |
1250 | { | |
1251 | case REG: | |
1252 | { | |
1253 | struct elt_list *l = REG_VALUES (REGNO (orig)); | |
1254 | ||
1255 | if (l && l->elt == NULL) | |
1256 | l = l->next; | |
1257 | for (; l; l = l->next) | |
1258 | if (GET_MODE (l->elt->val_rtx) == GET_MODE (orig)) | |
1259 | { | |
1260 | rtx result; | |
1261 | int regno = REGNO (orig); | |
b8698a0f | 1262 | |
6fb5fa3c | 1263 | /* The only thing that we are not willing to do (this |
6ed3da00 | 1264 | is requirement of dse and if others potential uses |
6fb5fa3c DB |
1265 | need this function we should add a parm to control |
1266 | it) is that we will not substitute the | |
1267 | STACK_POINTER_REGNUM, FRAME_POINTER or the | |
1268 | HARD_FRAME_POINTER. | |
1269 | ||
cea618ac | 1270 | These expansions confuses the code that notices that |
6fb5fa3c DB |
1271 | stores into the frame go dead at the end of the |
1272 | function and that the frame is not effected by calls | |
1273 | to subroutines. If you allow the | |
1274 | STACK_POINTER_REGNUM substitution, then dse will | |
1275 | think that parameter pushing also goes dead which is | |
1276 | wrong. If you allow the FRAME_POINTER or the | |
1277 | HARD_FRAME_POINTER then you lose the opportunity to | |
1278 | make the frame assumptions. */ | |
1279 | if (regno == STACK_POINTER_REGNUM | |
1280 | || regno == FRAME_POINTER_REGNUM | |
1281 | || regno == HARD_FRAME_POINTER_REGNUM) | |
1282 | return orig; | |
1283 | ||
b5b8b0ac | 1284 | bitmap_set_bit (evd->regs_active, regno); |
6fb5fa3c | 1285 | |
b5b8b0ac | 1286 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
1287 | fprintf (dump_file, "expanding: r%d into: ", regno); |
1288 | ||
b5b8b0ac AO |
1289 | result = expand_loc (l->elt->locs, evd, max_depth); |
1290 | bitmap_clear_bit (evd->regs_active, regno); | |
6fb5fa3c DB |
1291 | |
1292 | if (result) | |
1293 | return result; | |
b8698a0f | 1294 | else |
6fb5fa3c DB |
1295 | return orig; |
1296 | } | |
1297 | } | |
b8698a0f | 1298 | |
6fb5fa3c DB |
1299 | case CONST_INT: |
1300 | case CONST_DOUBLE: | |
1301 | case CONST_VECTOR: | |
1302 | case SYMBOL_REF: | |
1303 | case CODE_LABEL: | |
1304 | case PC: | |
1305 | case CC0: | |
1306 | case SCRATCH: | |
1307 | /* SCRATCH must be shared because they represent distinct values. */ | |
1308 | return orig; | |
1309 | case CLOBBER: | |
1310 | if (REG_P (XEXP (orig, 0)) && HARD_REGISTER_NUM_P (REGNO (XEXP (orig, 0)))) | |
1311 | return orig; | |
1312 | break; | |
1313 | ||
1314 | case CONST: | |
1315 | if (shared_const_p (orig)) | |
1316 | return orig; | |
1317 | break; | |
1318 | ||
8dd5516b | 1319 | case SUBREG: |
6fb5fa3c | 1320 | { |
0b7e34d7 AO |
1321 | rtx subreg; |
1322 | ||
1323 | if (evd->callback) | |
1324 | { | |
1325 | subreg = evd->callback (orig, evd->regs_active, max_depth, | |
1326 | evd->callback_arg); | |
1327 | if (subreg != orig) | |
1328 | return subreg; | |
1329 | } | |
1330 | ||
1331 | subreg = cselib_expand_value_rtx_1 (SUBREG_REG (orig), evd, | |
1332 | max_depth - 1); | |
8dd5516b JJ |
1333 | if (!subreg) |
1334 | return NULL; | |
1335 | scopy = simplify_gen_subreg (GET_MODE (orig), subreg, | |
1336 | GET_MODE (SUBREG_REG (orig)), | |
1337 | SUBREG_BYTE (orig)); | |
0b7e34d7 AO |
1338 | if (scopy == NULL |
1339 | || (GET_CODE (scopy) == SUBREG | |
1340 | && !REG_P (SUBREG_REG (scopy)) | |
1341 | && !MEM_P (SUBREG_REG (scopy)))) | |
1342 | return NULL; | |
1343 | ||
8dd5516b | 1344 | return scopy; |
6fb5fa3c | 1345 | } |
8dd5516b JJ |
1346 | |
1347 | case VALUE: | |
b5b8b0ac AO |
1348 | { |
1349 | rtx result; | |
0b7e34d7 | 1350 | |
b5b8b0ac AO |
1351 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1352 | { | |
1353 | fputs ("\nexpanding ", dump_file); | |
1354 | print_rtl_single (dump_file, orig); | |
1355 | fputs (" into...", dump_file); | |
1356 | } | |
8dd5516b | 1357 | |
0b7e34d7 | 1358 | if (evd->callback) |
b5b8b0ac AO |
1359 | { |
1360 | result = evd->callback (orig, evd->regs_active, max_depth, | |
1361 | evd->callback_arg); | |
0b7e34d7 AO |
1362 | |
1363 | if (result != orig) | |
1364 | return result; | |
b5b8b0ac | 1365 | } |
8dd5516b | 1366 | |
0b7e34d7 | 1367 | result = expand_loc (CSELIB_VAL_PTR (orig)->locs, evd, max_depth); |
b5b8b0ac AO |
1368 | return result; |
1369 | } | |
0ca5af51 AO |
1370 | |
1371 | case DEBUG_EXPR: | |
1372 | if (evd->callback) | |
1373 | return evd->callback (orig, evd->regs_active, max_depth, | |
1374 | evd->callback_arg); | |
1375 | return orig; | |
1376 | ||
6fb5fa3c DB |
1377 | default: |
1378 | break; | |
1379 | } | |
1380 | ||
1381 | /* Copy the various flags, fields, and other information. We assume | |
1382 | that all fields need copying, and then clear the fields that should | |
1383 | not be copied. That is the sensible default behavior, and forces | |
1384 | us to explicitly document why we are *not* copying a flag. */ | |
864ddef7 JJ |
1385 | if (evd->dummy) |
1386 | copy = NULL; | |
1387 | else | |
1388 | copy = shallow_copy_rtx (orig); | |
6fb5fa3c | 1389 | |
8dd5516b | 1390 | format_ptr = GET_RTX_FORMAT (code); |
6fb5fa3c | 1391 | |
8dd5516b | 1392 | for (i = 0; i < GET_RTX_LENGTH (code); i++) |
6fb5fa3c DB |
1393 | switch (*format_ptr++) |
1394 | { | |
1395 | case 'e': | |
1396 | if (XEXP (orig, i) != NULL) | |
1397 | { | |
b5b8b0ac AO |
1398 | rtx result = cselib_expand_value_rtx_1 (XEXP (orig, i), evd, |
1399 | max_depth - 1); | |
6fb5fa3c DB |
1400 | if (!result) |
1401 | return NULL; | |
864ddef7 JJ |
1402 | if (copy) |
1403 | XEXP (copy, i) = result; | |
6fb5fa3c DB |
1404 | } |
1405 | break; | |
1406 | ||
1407 | case 'E': | |
1408 | case 'V': | |
1409 | if (XVEC (orig, i) != NULL) | |
1410 | { | |
864ddef7 JJ |
1411 | if (copy) |
1412 | XVEC (copy, i) = rtvec_alloc (XVECLEN (orig, i)); | |
1413 | for (j = 0; j < XVECLEN (orig, i); j++) | |
6fb5fa3c | 1414 | { |
b5b8b0ac AO |
1415 | rtx result = cselib_expand_value_rtx_1 (XVECEXP (orig, i, j), |
1416 | evd, max_depth - 1); | |
6fb5fa3c DB |
1417 | if (!result) |
1418 | return NULL; | |
864ddef7 JJ |
1419 | if (copy) |
1420 | XVECEXP (copy, i, j) = result; | |
6fb5fa3c DB |
1421 | } |
1422 | } | |
1423 | break; | |
1424 | ||
1425 | case 't': | |
1426 | case 'w': | |
1427 | case 'i': | |
1428 | case 's': | |
1429 | case 'S': | |
1430 | case 'T': | |
1431 | case 'u': | |
1432 | case 'B': | |
1433 | case '0': | |
1434 | /* These are left unchanged. */ | |
1435 | break; | |
1436 | ||
1437 | default: | |
1438 | gcc_unreachable (); | |
1439 | } | |
1440 | ||
864ddef7 JJ |
1441 | if (evd->dummy) |
1442 | return orig; | |
1443 | ||
8dd5516b JJ |
1444 | mode = GET_MODE (copy); |
1445 | /* If an operand has been simplified into CONST_INT, which doesn't | |
1446 | have a mode and the mode isn't derivable from whole rtx's mode, | |
1447 | try simplify_*_operation first with mode from original's operand | |
1448 | and as a fallback wrap CONST_INT into gen_rtx_CONST. */ | |
1449 | scopy = copy; | |
1450 | switch (GET_RTX_CLASS (code)) | |
1451 | { | |
1452 | case RTX_UNARY: | |
1453 | if (CONST_INT_P (XEXP (copy, 0)) | |
1454 | && GET_MODE (XEXP (orig, 0)) != VOIDmode) | |
1455 | { | |
1456 | scopy = simplify_unary_operation (code, mode, XEXP (copy, 0), | |
1457 | GET_MODE (XEXP (orig, 0))); | |
1458 | if (scopy) | |
1459 | return scopy; | |
1460 | } | |
1461 | break; | |
1462 | case RTX_COMM_ARITH: | |
1463 | case RTX_BIN_ARITH: | |
1464 | /* These expressions can derive operand modes from the whole rtx's mode. */ | |
1465 | break; | |
1466 | case RTX_TERNARY: | |
1467 | case RTX_BITFIELD_OPS: | |
1468 | if (CONST_INT_P (XEXP (copy, 0)) | |
1469 | && GET_MODE (XEXP (orig, 0)) != VOIDmode) | |
1470 | { | |
1471 | scopy = simplify_ternary_operation (code, mode, | |
1472 | GET_MODE (XEXP (orig, 0)), | |
1473 | XEXP (copy, 0), XEXP (copy, 1), | |
1474 | XEXP (copy, 2)); | |
1475 | if (scopy) | |
1476 | return scopy; | |
1477 | } | |
1478 | break; | |
1479 | case RTX_COMPARE: | |
1480 | case RTX_COMM_COMPARE: | |
1481 | if (CONST_INT_P (XEXP (copy, 0)) | |
1482 | && GET_MODE (XEXP (copy, 1)) == VOIDmode | |
1483 | && (GET_MODE (XEXP (orig, 0)) != VOIDmode | |
1484 | || GET_MODE (XEXP (orig, 1)) != VOIDmode)) | |
1485 | { | |
1486 | scopy = simplify_relational_operation (code, mode, | |
1487 | (GET_MODE (XEXP (orig, 0)) | |
1488 | != VOIDmode) | |
1489 | ? GET_MODE (XEXP (orig, 0)) | |
1490 | : GET_MODE (XEXP (orig, 1)), | |
1491 | XEXP (copy, 0), | |
1492 | XEXP (copy, 1)); | |
1493 | if (scopy) | |
1494 | return scopy; | |
1495 | } | |
1496 | break; | |
1497 | default: | |
1498 | break; | |
1499 | } | |
6fb5fa3c DB |
1500 | scopy = simplify_rtx (copy); |
1501 | if (scopy) | |
3af4ba41 | 1502 | return scopy; |
6fb5fa3c DB |
1503 | return copy; |
1504 | } | |
1505 | ||
fa49fd0f RK |
1506 | /* Walk rtx X and replace all occurrences of REG and MEM subexpressions |
1507 | with VALUE expressions. This way, it becomes independent of changes | |
1508 | to registers and memory. | |
1509 | X isn't actually modified; if modifications are needed, new rtl is | |
1510 | allocated. However, the return value can share rtl with X. */ | |
1511 | ||
91700444 | 1512 | rtx |
7080f735 | 1513 | cselib_subst_to_values (rtx x) |
fa49fd0f RK |
1514 | { |
1515 | enum rtx_code code = GET_CODE (x); | |
1516 | const char *fmt = GET_RTX_FORMAT (code); | |
1517 | cselib_val *e; | |
1518 | struct elt_list *l; | |
1519 | rtx copy = x; | |
1520 | int i; | |
1521 | ||
1522 | switch (code) | |
1523 | { | |
1524 | case REG: | |
60fa6660 AO |
1525 | l = REG_VALUES (REGNO (x)); |
1526 | if (l && l->elt == NULL) | |
1527 | l = l->next; | |
1528 | for (; l; l = l->next) | |
757bbef8 SB |
1529 | if (GET_MODE (l->elt->val_rtx) == GET_MODE (x)) |
1530 | return l->elt->val_rtx; | |
fa49fd0f | 1531 | |
341c100f | 1532 | gcc_unreachable (); |
fa49fd0f RK |
1533 | |
1534 | case MEM: | |
1535 | e = cselib_lookup_mem (x, 0); | |
1536 | if (! e) | |
91700444 BS |
1537 | { |
1538 | /* This happens for autoincrements. Assign a value that doesn't | |
1539 | match any other. */ | |
5440c0e7 | 1540 | e = new_cselib_val (next_uid, GET_MODE (x), x); |
91700444 | 1541 | } |
757bbef8 | 1542 | return e->val_rtx; |
fa49fd0f | 1543 | |
fa49fd0f | 1544 | case CONST_DOUBLE: |
69ef87e2 | 1545 | case CONST_VECTOR: |
fa49fd0f | 1546 | case CONST_INT: |
091a3ac7 | 1547 | case CONST_FIXED: |
fa49fd0f RK |
1548 | return x; |
1549 | ||
91700444 BS |
1550 | case POST_INC: |
1551 | case PRE_INC: | |
1552 | case POST_DEC: | |
1553 | case PRE_DEC: | |
1554 | case POST_MODIFY: | |
1555 | case PRE_MODIFY: | |
5440c0e7 | 1556 | e = new_cselib_val (next_uid, GET_MODE (x), x); |
757bbef8 | 1557 | return e->val_rtx; |
7080f735 | 1558 | |
fa49fd0f RK |
1559 | default: |
1560 | break; | |
1561 | } | |
1562 | ||
1563 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
1564 | { | |
1565 | if (fmt[i] == 'e') | |
1566 | { | |
1567 | rtx t = cselib_subst_to_values (XEXP (x, i)); | |
1568 | ||
bd7960b1 RS |
1569 | if (t != XEXP (x, i)) |
1570 | { | |
1571 | if (x == copy) | |
1572 | copy = shallow_copy_rtx (x); | |
1573 | XEXP (copy, i) = t; | |
1574 | } | |
fa49fd0f RK |
1575 | } |
1576 | else if (fmt[i] == 'E') | |
1577 | { | |
bd7960b1 | 1578 | int j; |
fa49fd0f RK |
1579 | |
1580 | for (j = 0; j < XVECLEN (x, i); j++) | |
1581 | { | |
1582 | rtx t = cselib_subst_to_values (XVECEXP (x, i, j)); | |
1583 | ||
bd7960b1 | 1584 | if (t != XVECEXP (x, i, j)) |
fa49fd0f | 1585 | { |
bd7960b1 RS |
1586 | if (XVEC (x, i) == XVEC (copy, i)) |
1587 | { | |
1588 | if (x == copy) | |
1589 | copy = shallow_copy_rtx (x); | |
1590 | XVEC (copy, i) = shallow_copy_rtvec (XVEC (x, i)); | |
1591 | } | |
1592 | XVECEXP (copy, i, j) = t; | |
fa49fd0f | 1593 | } |
fa49fd0f RK |
1594 | } |
1595 | } | |
1596 | } | |
1597 | ||
1598 | return copy; | |
1599 | } | |
1600 | ||
1601 | /* Look up the rtl expression X in our tables and return the value it has. | |
1602 | If CREATE is zero, we return NULL if we don't know the value. Otherwise, | |
1603 | we create a new one if possible, using mode MODE if X doesn't have a mode | |
1604 | (i.e. because it's a constant). */ | |
1605 | ||
5847e8da AO |
1606 | static cselib_val * |
1607 | cselib_lookup_1 (rtx x, enum machine_mode mode, int create) | |
fa49fd0f RK |
1608 | { |
1609 | void **slot; | |
1610 | cselib_val *e; | |
1611 | unsigned int hashval; | |
1612 | ||
1613 | if (GET_MODE (x) != VOIDmode) | |
1614 | mode = GET_MODE (x); | |
1615 | ||
1616 | if (GET_CODE (x) == VALUE) | |
1617 | return CSELIB_VAL_PTR (x); | |
1618 | ||
f8cfc6aa | 1619 | if (REG_P (x)) |
fa49fd0f RK |
1620 | { |
1621 | struct elt_list *l; | |
1622 | unsigned int i = REGNO (x); | |
1623 | ||
60fa6660 AO |
1624 | l = REG_VALUES (i); |
1625 | if (l && l->elt == NULL) | |
1626 | l = l->next; | |
1627 | for (; l; l = l->next) | |
757bbef8 | 1628 | if (mode == GET_MODE (l->elt->val_rtx)) |
5847e8da AO |
1629 | { |
1630 | promote_debug_loc (l->elt->locs); | |
1631 | return l->elt; | |
1632 | } | |
fa49fd0f RK |
1633 | |
1634 | if (! create) | |
5847e8da | 1635 | return 0; |
fa49fd0f | 1636 | |
31825e57 DM |
1637 | if (i < FIRST_PSEUDO_REGISTER) |
1638 | { | |
66fd46b6 | 1639 | unsigned int n = hard_regno_nregs[i][mode]; |
31825e57 DM |
1640 | |
1641 | if (n > max_value_regs) | |
1642 | max_value_regs = n; | |
1643 | } | |
1644 | ||
5440c0e7 | 1645 | e = new_cselib_val (next_uid, GET_MODE (x), x); |
fa49fd0f RK |
1646 | e->locs = new_elt_loc_list (e->locs, x); |
1647 | if (REG_VALUES (i) == 0) | |
60fa6660 AO |
1648 | { |
1649 | /* Maintain the invariant that the first entry of | |
1650 | REG_VALUES, if present, must be the value used to set the | |
1651 | register, or NULL. */ | |
6790d1ab | 1652 | used_regs[n_used_regs++] = i; |
60fa6660 AO |
1653 | REG_VALUES (i) = new_elt_list (REG_VALUES (i), NULL); |
1654 | } | |
1655 | REG_VALUES (i)->next = new_elt_list (REG_VALUES (i)->next, e); | |
5440c0e7 | 1656 | slot = htab_find_slot_with_hash (cselib_hash_table, x, e->hash, INSERT); |
fa49fd0f | 1657 | *slot = e; |
5847e8da | 1658 | return e; |
fa49fd0f RK |
1659 | } |
1660 | ||
3c0cb5de | 1661 | if (MEM_P (x)) |
5847e8da | 1662 | return cselib_lookup_mem (x, create); |
fa49fd0f | 1663 | |
29c1846b | 1664 | hashval = cselib_hash_rtx (x, create); |
fa49fd0f RK |
1665 | /* Can't even create if hashing is not possible. */ |
1666 | if (! hashval) | |
5847e8da | 1667 | return 0; |
fa49fd0f | 1668 | |
7c514720 | 1669 | slot = htab_find_slot_with_hash (cselib_hash_table, wrap_constant (mode, x), |
fa49fd0f RK |
1670 | hashval, create ? INSERT : NO_INSERT); |
1671 | if (slot == 0) | |
5847e8da | 1672 | return 0; |
fa49fd0f RK |
1673 | |
1674 | e = (cselib_val *) *slot; | |
1675 | if (e) | |
5847e8da | 1676 | return e; |
fa49fd0f | 1677 | |
b5b8b0ac | 1678 | e = new_cselib_val (hashval, mode, x); |
fa49fd0f RK |
1679 | |
1680 | /* We have to fill the slot before calling cselib_subst_to_values: | |
1681 | the hash table is inconsistent until we do so, and | |
1682 | cselib_subst_to_values will need to do lookups. */ | |
1683 | *slot = (void *) e; | |
1684 | e->locs = new_elt_loc_list (e->locs, cselib_subst_to_values (x)); | |
5847e8da AO |
1685 | return e; |
1686 | } | |
1687 | ||
1688 | /* Wrapper for cselib_lookup, that indicates X is in INSN. */ | |
1689 | ||
1690 | cselib_val * | |
1691 | cselib_lookup_from_insn (rtx x, enum machine_mode mode, | |
1692 | int create, rtx insn) | |
1693 | { | |
1694 | cselib_val *ret; | |
1695 | ||
1696 | gcc_assert (!cselib_current_insn); | |
1697 | cselib_current_insn = insn; | |
1698 | ||
1699 | ret = cselib_lookup (x, mode, create); | |
1700 | ||
1701 | cselib_current_insn = NULL; | |
1702 | ||
1703 | return ret; | |
1704 | } | |
1705 | ||
1706 | /* Wrapper for cselib_lookup_1, that logs the lookup result and | |
1707 | maintains invariants related with debug insns. */ | |
1708 | ||
1709 | cselib_val * | |
1710 | cselib_lookup (rtx x, enum machine_mode mode, int create) | |
1711 | { | |
1712 | cselib_val *ret = cselib_lookup_1 (x, mode, create); | |
1713 | ||
1714 | /* ??? Should we return NULL if we're not to create an entry, the | |
1715 | found loc is a debug loc and cselib_current_insn is not DEBUG? | |
1716 | If so, we should also avoid converting val to non-DEBUG; probably | |
1717 | easiest setting cselib_current_insn to NULL before the call | |
1718 | above. */ | |
1719 | ||
1720 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1721 | { | |
1722 | fputs ("cselib lookup ", dump_file); | |
1723 | print_inline_rtx (dump_file, x, 2); | |
1724 | fprintf (dump_file, " => %u:%u\n", | |
1725 | ret ? ret->uid : 0, | |
1726 | ret ? ret->hash : 0); | |
1727 | } | |
1728 | ||
1729 | return ret; | |
fa49fd0f RK |
1730 | } |
1731 | ||
1732 | /* Invalidate any entries in reg_values that overlap REGNO. This is called | |
1733 | if REGNO is changing. MODE is the mode of the assignment to REGNO, which | |
1734 | is used to determine how many hard registers are being changed. If MODE | |
1735 | is VOIDmode, then only REGNO is being changed; this is used when | |
1736 | invalidating call clobbered registers across a call. */ | |
1737 | ||
1738 | static void | |
7080f735 | 1739 | cselib_invalidate_regno (unsigned int regno, enum machine_mode mode) |
fa49fd0f RK |
1740 | { |
1741 | unsigned int endregno; | |
1742 | unsigned int i; | |
1743 | ||
1744 | /* If we see pseudos after reload, something is _wrong_. */ | |
341c100f NS |
1745 | gcc_assert (!reload_completed || regno < FIRST_PSEUDO_REGISTER |
1746 | || reg_renumber[regno] < 0); | |
fa49fd0f RK |
1747 | |
1748 | /* Determine the range of registers that must be invalidated. For | |
1749 | pseudos, only REGNO is affected. For hard regs, we must take MODE | |
1750 | into account, and we must also invalidate lower register numbers | |
1751 | if they contain values that overlap REGNO. */ | |
291aac59 | 1752 | if (regno < FIRST_PSEUDO_REGISTER) |
31825e57 | 1753 | { |
341c100f | 1754 | gcc_assert (mode != VOIDmode); |
7080f735 | 1755 | |
31825e57 DM |
1756 | if (regno < max_value_regs) |
1757 | i = 0; | |
1758 | else | |
1759 | i = regno - max_value_regs; | |
fa49fd0f | 1760 | |
09e18274 | 1761 | endregno = end_hard_regno (mode, regno); |
31825e57 DM |
1762 | } |
1763 | else | |
1764 | { | |
1765 | i = regno; | |
1766 | endregno = regno + 1; | |
1767 | } | |
1768 | ||
1769 | for (; i < endregno; i++) | |
fa49fd0f RK |
1770 | { |
1771 | struct elt_list **l = ®_VALUES (i); | |
1772 | ||
1773 | /* Go through all known values for this reg; if it overlaps the range | |
1774 | we're invalidating, remove the value. */ | |
1775 | while (*l) | |
1776 | { | |
1777 | cselib_val *v = (*l)->elt; | |
5847e8da AO |
1778 | bool had_locs; |
1779 | rtx setting_insn; | |
fa49fd0f RK |
1780 | struct elt_loc_list **p; |
1781 | unsigned int this_last = i; | |
1782 | ||
60fa6660 | 1783 | if (i < FIRST_PSEUDO_REGISTER && v != NULL) |
09e18274 | 1784 | this_last = end_hard_regno (GET_MODE (v->val_rtx), i) - 1; |
fa49fd0f | 1785 | |
457eeaae | 1786 | if (this_last < regno || v == NULL || v == cfa_base_preserved_val) |
fa49fd0f RK |
1787 | { |
1788 | l = &(*l)->next; | |
1789 | continue; | |
1790 | } | |
1791 | ||
1792 | /* We have an overlap. */ | |
60fa6660 AO |
1793 | if (*l == REG_VALUES (i)) |
1794 | { | |
1795 | /* Maintain the invariant that the first entry of | |
1796 | REG_VALUES, if present, must be the value used to set | |
1797 | the register, or NULL. This is also nice because | |
1798 | then we won't push the same regno onto user_regs | |
1799 | multiple times. */ | |
1800 | (*l)->elt = NULL; | |
1801 | l = &(*l)->next; | |
1802 | } | |
1803 | else | |
1804 | unchain_one_elt_list (l); | |
fa49fd0f | 1805 | |
5847e8da AO |
1806 | had_locs = v->locs != NULL; |
1807 | setting_insn = v->locs ? v->locs->setting_insn : NULL; | |
1808 | ||
fa49fd0f RK |
1809 | /* Now, we clear the mapping from value to reg. It must exist, so |
1810 | this code will crash intentionally if it doesn't. */ | |
1811 | for (p = &v->locs; ; p = &(*p)->next) | |
1812 | { | |
1813 | rtx x = (*p)->loc; | |
1814 | ||
f8cfc6aa | 1815 | if (REG_P (x) && REGNO (x) == i) |
fa49fd0f RK |
1816 | { |
1817 | unchain_one_elt_loc_list (p); | |
1818 | break; | |
1819 | } | |
1820 | } | |
5847e8da AO |
1821 | |
1822 | if (had_locs && v->locs == 0 && !PRESERVED_VALUE_P (v->val_rtx)) | |
1823 | { | |
1824 | if (setting_insn && DEBUG_INSN_P (setting_insn)) | |
1825 | n_useless_debug_values++; | |
1826 | else | |
1827 | n_useless_values++; | |
1828 | } | |
fa49fd0f RK |
1829 | } |
1830 | } | |
1831 | } | |
9ddb66ca JH |
1832 | \f |
1833 | /* Return 1 if X has a value that can vary even between two | |
1834 | executions of the program. 0 means X can be compared reliably | |
1835 | against certain constants or near-constants. */ | |
fa49fd0f | 1836 | |
4f588890 KG |
1837 | static bool |
1838 | cselib_rtx_varies_p (const_rtx x ATTRIBUTE_UNUSED, bool from_alias ATTRIBUTE_UNUSED) | |
fa49fd0f | 1839 | { |
9ddb66ca JH |
1840 | /* We actually don't need to verify very hard. This is because |
1841 | if X has actually changed, we invalidate the memory anyway, | |
1842 | so assume that all common memory addresses are | |
1843 | invariant. */ | |
fa49fd0f RK |
1844 | return 0; |
1845 | } | |
1846 | ||
7101fb18 JH |
1847 | /* Invalidate any locations in the table which are changed because of a |
1848 | store to MEM_RTX. If this is called because of a non-const call | |
1849 | instruction, MEM_RTX is (mem:BLK const0_rtx). */ | |
fa49fd0f | 1850 | |
7101fb18 | 1851 | static void |
7080f735 | 1852 | cselib_invalidate_mem (rtx mem_rtx) |
fa49fd0f | 1853 | { |
7101fb18 | 1854 | cselib_val **vp, *v, *next; |
c65ecebc | 1855 | int num_mems = 0; |
9ddb66ca JH |
1856 | rtx mem_addr; |
1857 | ||
1858 | mem_addr = canon_rtx (get_addr (XEXP (mem_rtx, 0))); | |
1859 | mem_rtx = canon_rtx (mem_rtx); | |
fa49fd0f | 1860 | |
7101fb18 JH |
1861 | vp = &first_containing_mem; |
1862 | for (v = *vp; v != &dummy_val; v = next) | |
fa49fd0f | 1863 | { |
7101fb18 JH |
1864 | bool has_mem = false; |
1865 | struct elt_loc_list **p = &v->locs; | |
5847e8da AO |
1866 | bool had_locs = v->locs != NULL; |
1867 | rtx setting_insn = v->locs ? v->locs->setting_insn : NULL; | |
fa49fd0f | 1868 | |
7101fb18 | 1869 | while (*p) |
fa49fd0f | 1870 | { |
7101fb18 JH |
1871 | rtx x = (*p)->loc; |
1872 | cselib_val *addr; | |
1873 | struct elt_list **mem_chain; | |
1874 | ||
1875 | /* MEMs may occur in locations only at the top level; below | |
1876 | that every MEM or REG is substituted by its VALUE. */ | |
3c0cb5de | 1877 | if (!MEM_P (x)) |
fa49fd0f | 1878 | { |
7101fb18 JH |
1879 | p = &(*p)->next; |
1880 | continue; | |
1881 | } | |
c65ecebc | 1882 | if (num_mems < PARAM_VALUE (PARAM_MAX_CSELIB_MEMORY_LOCATIONS) |
9ddb66ca | 1883 | && ! canon_true_dependence (mem_rtx, GET_MODE (mem_rtx), mem_addr, |
6216f94e | 1884 | x, NULL_RTX, cselib_rtx_varies_p)) |
7101fb18 JH |
1885 | { |
1886 | has_mem = true; | |
c65ecebc | 1887 | num_mems++; |
7101fb18 JH |
1888 | p = &(*p)->next; |
1889 | continue; | |
fa49fd0f RK |
1890 | } |
1891 | ||
7101fb18 JH |
1892 | /* This one overlaps. */ |
1893 | /* We must have a mapping from this MEM's address to the | |
1894 | value (E). Remove that, too. */ | |
1895 | addr = cselib_lookup (XEXP (x, 0), VOIDmode, 0); | |
1896 | mem_chain = &addr->addr_list; | |
1897 | for (;;) | |
1898 | { | |
1899 | if ((*mem_chain)->elt == v) | |
1900 | { | |
1901 | unchain_one_elt_list (mem_chain); | |
1902 | break; | |
1903 | } | |
fa49fd0f | 1904 | |
7101fb18 JH |
1905 | mem_chain = &(*mem_chain)->next; |
1906 | } | |
fa49fd0f | 1907 | |
7101fb18 JH |
1908 | unchain_one_elt_loc_list (p); |
1909 | } | |
fa49fd0f | 1910 | |
b5b8b0ac | 1911 | if (had_locs && v->locs == 0 && !PRESERVED_VALUE_P (v->val_rtx)) |
5847e8da AO |
1912 | { |
1913 | if (setting_insn && DEBUG_INSN_P (setting_insn)) | |
1914 | n_useless_debug_values++; | |
1915 | else | |
1916 | n_useless_values++; | |
1917 | } | |
fa49fd0f | 1918 | |
7101fb18 JH |
1919 | next = v->next_containing_mem; |
1920 | if (has_mem) | |
1921 | { | |
1922 | *vp = v; | |
1923 | vp = &(*vp)->next_containing_mem; | |
1924 | } | |
1925 | else | |
1926 | v->next_containing_mem = NULL; | |
1927 | } | |
1928 | *vp = &dummy_val; | |
fa49fd0f RK |
1929 | } |
1930 | ||
0d87c765 | 1931 | /* Invalidate DEST, which is being assigned to or clobbered. */ |
fa49fd0f | 1932 | |
0d87c765 RH |
1933 | void |
1934 | cselib_invalidate_rtx (rtx dest) | |
fa49fd0f | 1935 | { |
46d096a3 SB |
1936 | while (GET_CODE (dest) == SUBREG |
1937 | || GET_CODE (dest) == ZERO_EXTRACT | |
1938 | || GET_CODE (dest) == STRICT_LOW_PART) | |
fa49fd0f RK |
1939 | dest = XEXP (dest, 0); |
1940 | ||
f8cfc6aa | 1941 | if (REG_P (dest)) |
fa49fd0f | 1942 | cselib_invalidate_regno (REGNO (dest), GET_MODE (dest)); |
3c0cb5de | 1943 | else if (MEM_P (dest)) |
fa49fd0f RK |
1944 | cselib_invalidate_mem (dest); |
1945 | ||
1946 | /* Some machines don't define AUTO_INC_DEC, but they still use push | |
1947 | instructions. We need to catch that case here in order to | |
1948 | invalidate the stack pointer correctly. Note that invalidating | |
1949 | the stack pointer is different from invalidating DEST. */ | |
1950 | if (push_operand (dest, GET_MODE (dest))) | |
0d87c765 RH |
1951 | cselib_invalidate_rtx (stack_pointer_rtx); |
1952 | } | |
1953 | ||
1954 | /* A wrapper for cselib_invalidate_rtx to be called via note_stores. */ | |
1955 | ||
1956 | static void | |
7bc980e1 | 1957 | cselib_invalidate_rtx_note_stores (rtx dest, const_rtx ignore ATTRIBUTE_UNUSED, |
0d87c765 RH |
1958 | void *data ATTRIBUTE_UNUSED) |
1959 | { | |
1960 | cselib_invalidate_rtx (dest); | |
fa49fd0f RK |
1961 | } |
1962 | ||
1963 | /* Record the result of a SET instruction. DEST is being set; the source | |
1964 | contains the value described by SRC_ELT. If DEST is a MEM, DEST_ADDR_ELT | |
1965 | describes its address. */ | |
1966 | ||
1967 | static void | |
7080f735 | 1968 | cselib_record_set (rtx dest, cselib_val *src_elt, cselib_val *dest_addr_elt) |
fa49fd0f | 1969 | { |
f8cfc6aa | 1970 | int dreg = REG_P (dest) ? (int) REGNO (dest) : -1; |
fa49fd0f RK |
1971 | |
1972 | if (src_elt == 0 || side_effects_p (dest)) | |
1973 | return; | |
1974 | ||
1975 | if (dreg >= 0) | |
1976 | { | |
31825e57 DM |
1977 | if (dreg < FIRST_PSEUDO_REGISTER) |
1978 | { | |
66fd46b6 | 1979 | unsigned int n = hard_regno_nregs[dreg][GET_MODE (dest)]; |
31825e57 DM |
1980 | |
1981 | if (n > max_value_regs) | |
1982 | max_value_regs = n; | |
1983 | } | |
1984 | ||
60fa6660 AO |
1985 | if (REG_VALUES (dreg) == 0) |
1986 | { | |
6790d1ab | 1987 | used_regs[n_used_regs++] = dreg; |
60fa6660 AO |
1988 | REG_VALUES (dreg) = new_elt_list (REG_VALUES (dreg), src_elt); |
1989 | } | |
1990 | else | |
1991 | { | |
341c100f NS |
1992 | /* The register should have been invalidated. */ |
1993 | gcc_assert (REG_VALUES (dreg)->elt == 0); | |
1994 | REG_VALUES (dreg)->elt = src_elt; | |
60fa6660 AO |
1995 | } |
1996 | ||
b5b8b0ac | 1997 | if (src_elt->locs == 0 && !PRESERVED_VALUE_P (src_elt->val_rtx)) |
fa49fd0f RK |
1998 | n_useless_values--; |
1999 | src_elt->locs = new_elt_loc_list (src_elt->locs, dest); | |
2000 | } | |
3c0cb5de | 2001 | else if (MEM_P (dest) && dest_addr_elt != 0 |
463301c3 | 2002 | && cselib_record_memory) |
fa49fd0f | 2003 | { |
b5b8b0ac | 2004 | if (src_elt->locs == 0 && !PRESERVED_VALUE_P (src_elt->val_rtx)) |
fa49fd0f RK |
2005 | n_useless_values--; |
2006 | add_mem_for_addr (dest_addr_elt, src_elt, dest); | |
2007 | } | |
2008 | } | |
2009 | ||
fa49fd0f RK |
2010 | /* There is no good way to determine how many elements there can be |
2011 | in a PARALLEL. Since it's fairly cheap, use a really large number. */ | |
2012 | #define MAX_SETS (FIRST_PSEUDO_REGISTER * 2) | |
2013 | ||
2014 | /* Record the effects of any sets in INSN. */ | |
2015 | static void | |
7080f735 | 2016 | cselib_record_sets (rtx insn) |
fa49fd0f RK |
2017 | { |
2018 | int n_sets = 0; | |
2019 | int i; | |
b5b8b0ac | 2020 | struct cselib_set sets[MAX_SETS]; |
fa49fd0f | 2021 | rtx body = PATTERN (insn); |
b7933c21 | 2022 | rtx cond = 0; |
fa49fd0f RK |
2023 | |
2024 | body = PATTERN (insn); | |
b7933c21 BS |
2025 | if (GET_CODE (body) == COND_EXEC) |
2026 | { | |
2027 | cond = COND_EXEC_TEST (body); | |
2028 | body = COND_EXEC_CODE (body); | |
2029 | } | |
2030 | ||
fa49fd0f RK |
2031 | /* Find all sets. */ |
2032 | if (GET_CODE (body) == SET) | |
2033 | { | |
2034 | sets[0].src = SET_SRC (body); | |
2035 | sets[0].dest = SET_DEST (body); | |
2036 | n_sets = 1; | |
2037 | } | |
2038 | else if (GET_CODE (body) == PARALLEL) | |
2039 | { | |
2040 | /* Look through the PARALLEL and record the values being | |
2041 | set, if possible. Also handle any CLOBBERs. */ | |
2042 | for (i = XVECLEN (body, 0) - 1; i >= 0; --i) | |
2043 | { | |
2044 | rtx x = XVECEXP (body, 0, i); | |
2045 | ||
2046 | if (GET_CODE (x) == SET) | |
2047 | { | |
2048 | sets[n_sets].src = SET_SRC (x); | |
2049 | sets[n_sets].dest = SET_DEST (x); | |
2050 | n_sets++; | |
2051 | } | |
2052 | } | |
2053 | } | |
2054 | ||
8dd5516b JJ |
2055 | if (n_sets == 1 |
2056 | && MEM_P (sets[0].src) | |
2057 | && !cselib_record_memory | |
2058 | && MEM_READONLY_P (sets[0].src)) | |
2059 | { | |
2060 | rtx note = find_reg_equal_equiv_note (insn); | |
2061 | ||
2062 | if (note && CONSTANT_P (XEXP (note, 0))) | |
2063 | sets[0].src = XEXP (note, 0); | |
2064 | } | |
2065 | ||
fa49fd0f RK |
2066 | /* Look up the values that are read. Do this before invalidating the |
2067 | locations that are written. */ | |
2068 | for (i = 0; i < n_sets; i++) | |
2069 | { | |
2070 | rtx dest = sets[i].dest; | |
2071 | ||
2072 | /* A STRICT_LOW_PART can be ignored; we'll record the equivalence for | |
2073 | the low part after invalidating any knowledge about larger modes. */ | |
2074 | if (GET_CODE (sets[i].dest) == STRICT_LOW_PART) | |
2075 | sets[i].dest = dest = XEXP (dest, 0); | |
2076 | ||
2077 | /* We don't know how to record anything but REG or MEM. */ | |
f8cfc6aa | 2078 | if (REG_P (dest) |
3c0cb5de | 2079 | || (MEM_P (dest) && cselib_record_memory)) |
fa49fd0f | 2080 | { |
b7933c21 BS |
2081 | rtx src = sets[i].src; |
2082 | if (cond) | |
be9ed5d5 | 2083 | src = gen_rtx_IF_THEN_ELSE (GET_MODE (dest), cond, src, dest); |
37060e78 | 2084 | sets[i].src_elt = cselib_lookup (src, GET_MODE (dest), 1); |
3c0cb5de | 2085 | if (MEM_P (dest)) |
d4ebfa65 BE |
2086 | { |
2087 | enum machine_mode address_mode | |
2088 | = targetm.addr_space.address_mode (MEM_ADDR_SPACE (dest)); | |
2089 | ||
2090 | sets[i].dest_addr_elt = cselib_lookup (XEXP (dest, 0), | |
2091 | address_mode, 1); | |
2092 | } | |
fa49fd0f RK |
2093 | else |
2094 | sets[i].dest_addr_elt = 0; | |
2095 | } | |
2096 | } | |
2097 | ||
b5b8b0ac AO |
2098 | if (cselib_record_sets_hook) |
2099 | cselib_record_sets_hook (insn, sets, n_sets); | |
2100 | ||
fa49fd0f RK |
2101 | /* Invalidate all locations written by this insn. Note that the elts we |
2102 | looked up in the previous loop aren't affected, just some of their | |
2103 | locations may go away. */ | |
0d87c765 | 2104 | note_stores (body, cselib_invalidate_rtx_note_stores, NULL); |
fa49fd0f | 2105 | |
b7048ab7 RH |
2106 | /* If this is an asm, look for duplicate sets. This can happen when the |
2107 | user uses the same value as an output multiple times. This is valid | |
2108 | if the outputs are not actually used thereafter. Treat this case as | |
2109 | if the value isn't actually set. We do this by smashing the destination | |
2110 | to pc_rtx, so that we won't record the value later. */ | |
2111 | if (n_sets >= 2 && asm_noperands (body) >= 0) | |
2112 | { | |
2113 | for (i = 0; i < n_sets; i++) | |
2114 | { | |
2115 | rtx dest = sets[i].dest; | |
3c0cb5de | 2116 | if (REG_P (dest) || MEM_P (dest)) |
b7048ab7 RH |
2117 | { |
2118 | int j; | |
2119 | for (j = i + 1; j < n_sets; j++) | |
2120 | if (rtx_equal_p (dest, sets[j].dest)) | |
2121 | { | |
2122 | sets[i].dest = pc_rtx; | |
2123 | sets[j].dest = pc_rtx; | |
2124 | } | |
2125 | } | |
2126 | } | |
2127 | } | |
2128 | ||
fa49fd0f RK |
2129 | /* Now enter the equivalences in our tables. */ |
2130 | for (i = 0; i < n_sets; i++) | |
2131 | { | |
2132 | rtx dest = sets[i].dest; | |
f8cfc6aa | 2133 | if (REG_P (dest) |
3c0cb5de | 2134 | || (MEM_P (dest) && cselib_record_memory)) |
fa49fd0f RK |
2135 | cselib_record_set (dest, sets[i].src_elt, sets[i].dest_addr_elt); |
2136 | } | |
2137 | } | |
2138 | ||
2139 | /* Record the effects of INSN. */ | |
2140 | ||
2141 | void | |
7080f735 | 2142 | cselib_process_insn (rtx insn) |
fa49fd0f RK |
2143 | { |
2144 | int i; | |
2145 | rtx x; | |
2146 | ||
2147 | cselib_current_insn = insn; | |
2148 | ||
2149 | /* Forget everything at a CODE_LABEL, a volatile asm, or a setjmp. */ | |
4b4bf941 JQ |
2150 | if (LABEL_P (insn) |
2151 | || (CALL_P (insn) | |
570a98eb | 2152 | && find_reg_note (insn, REG_SETJMP, NULL)) |
4b4bf941 | 2153 | || (NONJUMP_INSN_P (insn) |
fa49fd0f RK |
2154 | && GET_CODE (PATTERN (insn)) == ASM_OPERANDS |
2155 | && MEM_VOLATILE_P (PATTERN (insn)))) | |
2156 | { | |
5440c0e7 | 2157 | cselib_reset_table (next_uid); |
2080bd29 | 2158 | cselib_current_insn = NULL_RTX; |
fa49fd0f RK |
2159 | return; |
2160 | } | |
2161 | ||
2162 | if (! INSN_P (insn)) | |
2163 | { | |
2080bd29 | 2164 | cselib_current_insn = NULL_RTX; |
fa49fd0f RK |
2165 | return; |
2166 | } | |
2167 | ||
2168 | /* If this is a call instruction, forget anything stored in a | |
2169 | call clobbered register, or, if this is not a const call, in | |
2170 | memory. */ | |
4b4bf941 | 2171 | if (CALL_P (insn)) |
fa49fd0f RK |
2172 | { |
2173 | for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) | |
7e42db17 DJ |
2174 | if (call_used_regs[i] |
2175 | || (REG_VALUES (i) && REG_VALUES (i)->elt | |
b8698a0f | 2176 | && HARD_REGNO_CALL_PART_CLOBBERED (i, |
757bbef8 | 2177 | GET_MODE (REG_VALUES (i)->elt->val_rtx)))) |
291aac59 | 2178 | cselib_invalidate_regno (i, reg_raw_mode[i]); |
fa49fd0f | 2179 | |
becfd6e5 KZ |
2180 | /* Since it is not clear how cselib is going to be used, be |
2181 | conservative here and treat looping pure or const functions | |
2182 | as if they were regular functions. */ | |
2183 | if (RTL_LOOPING_CONST_OR_PURE_CALL_P (insn) | |
2184 | || !(RTL_CONST_OR_PURE_CALL_P (insn))) | |
fa49fd0f RK |
2185 | cselib_invalidate_mem (callmem); |
2186 | } | |
2187 | ||
2188 | cselib_record_sets (insn); | |
2189 | ||
2190 | #ifdef AUTO_INC_DEC | |
2191 | /* Clobber any registers which appear in REG_INC notes. We | |
2192 | could keep track of the changes to their values, but it is | |
2193 | unlikely to help. */ | |
2194 | for (x = REG_NOTES (insn); x; x = XEXP (x, 1)) | |
2195 | if (REG_NOTE_KIND (x) == REG_INC) | |
0d87c765 | 2196 | cselib_invalidate_rtx (XEXP (x, 0)); |
fa49fd0f RK |
2197 | #endif |
2198 | ||
2199 | /* Look for any CLOBBERs in CALL_INSN_FUNCTION_USAGE, but only | |
2200 | after we have processed the insn. */ | |
4b4bf941 | 2201 | if (CALL_P (insn)) |
fa49fd0f RK |
2202 | for (x = CALL_INSN_FUNCTION_USAGE (insn); x; x = XEXP (x, 1)) |
2203 | if (GET_CODE (XEXP (x, 0)) == CLOBBER) | |
0d87c765 | 2204 | cselib_invalidate_rtx (XEXP (XEXP (x, 0), 0)); |
fa49fd0f | 2205 | |
2080bd29 | 2206 | cselib_current_insn = NULL_RTX; |
fa49fd0f | 2207 | |
96d0cc81 JH |
2208 | if (n_useless_values > MAX_USELESS_VALUES |
2209 | /* remove_useless_values is linear in the hash table size. Avoid | |
9f5ed61a | 2210 | quadratic behavior for very large hashtables with very few |
96d0cc81 | 2211 | useless elements. */ |
5847e8da AO |
2212 | && ((unsigned int)n_useless_values |
2213 | > (cselib_hash_table->n_elements | |
2214 | - cselib_hash_table->n_deleted | |
2215 | - n_debug_values) / 4)) | |
fa49fd0f RK |
2216 | remove_useless_values (); |
2217 | } | |
2218 | ||
fa49fd0f RK |
2219 | /* Initialize cselib for one pass. The caller must also call |
2220 | init_alias_analysis. */ | |
2221 | ||
2222 | void | |
457eeaae | 2223 | cselib_init (int record_what) |
fa49fd0f | 2224 | { |
b8698a0f | 2225 | elt_list_pool = create_alloc_pool ("elt_list", |
6a59927d | 2226 | sizeof (struct elt_list), 10); |
b8698a0f | 2227 | elt_loc_list_pool = create_alloc_pool ("elt_loc_list", |
6a59927d | 2228 | sizeof (struct elt_loc_list), 10); |
b8698a0f | 2229 | cselib_val_pool = create_alloc_pool ("cselib_val_list", |
6a59927d | 2230 | sizeof (cselib_val), 10); |
aacd3885 | 2231 | value_pool = create_alloc_pool ("value", RTX_CODE_SIZE (VALUE), 100); |
457eeaae JJ |
2232 | cselib_record_memory = record_what & CSELIB_RECORD_MEMORY; |
2233 | cselib_preserve_constants = record_what & CSELIB_PRESERVE_CONSTANTS; | |
ac3768f6 SB |
2234 | |
2235 | /* (mem:BLK (scratch)) is a special mechanism to conflict with everything, | |
2236 | see canon_true_dependence. This is only created once. */ | |
fa49fd0f | 2237 | if (! callmem) |
ac3768f6 | 2238 | callmem = gen_rtx_MEM (BLKmode, gen_rtx_SCRATCH (VOIDmode)); |
fa49fd0f RK |
2239 | |
2240 | cselib_nregs = max_reg_num (); | |
6790d1ab JH |
2241 | |
2242 | /* We preserve reg_values to allow expensive clearing of the whole thing. | |
2243 | Reallocate it however if it happens to be too large. */ | |
2244 | if (!reg_values || reg_values_size < cselib_nregs | |
2245 | || (reg_values_size > 10 && reg_values_size > cselib_nregs * 4)) | |
e2500fed | 2246 | { |
6790d1ab JH |
2247 | if (reg_values) |
2248 | free (reg_values); | |
2249 | /* Some space for newly emit instructions so we don't end up | |
2250 | reallocating in between passes. */ | |
2251 | reg_values_size = cselib_nregs + (63 + cselib_nregs) / 16; | |
5ed6ace5 | 2252 | reg_values = XCNEWVEC (struct elt_list *, reg_values_size); |
e2500fed | 2253 | } |
5ed6ace5 | 2254 | used_regs = XNEWVEC (unsigned int, cselib_nregs); |
6790d1ab | 2255 | n_used_regs = 0; |
7c514720 KH |
2256 | cselib_hash_table = htab_create (31, get_value_hash, |
2257 | entry_and_rtx_equal_p, NULL); | |
5440c0e7 | 2258 | next_uid = 1; |
fa49fd0f RK |
2259 | } |
2260 | ||
2261 | /* Called when the current user is done with cselib. */ | |
2262 | ||
2263 | void | |
7080f735 | 2264 | cselib_finish (void) |
fa49fd0f | 2265 | { |
6fb5fa3c | 2266 | cselib_discard_hook = NULL; |
457eeaae JJ |
2267 | cselib_preserve_constants = false; |
2268 | cfa_base_preserved_val = NULL; | |
6a59927d JH |
2269 | free_alloc_pool (elt_list_pool); |
2270 | free_alloc_pool (elt_loc_list_pool); | |
2271 | free_alloc_pool (cselib_val_pool); | |
23bd7a93 | 2272 | free_alloc_pool (value_pool); |
eb232f4e | 2273 | cselib_clear_table (); |
7c514720 | 2274 | htab_delete (cselib_hash_table); |
0fc0c4c9 | 2275 | free (used_regs); |
e2500fed | 2276 | used_regs = 0; |
7c514720 | 2277 | cselib_hash_table = 0; |
e2500fed | 2278 | n_useless_values = 0; |
5847e8da AO |
2279 | n_useless_debug_values = 0; |
2280 | n_debug_values = 0; | |
5440c0e7 | 2281 | next_uid = 0; |
fa49fd0f | 2282 | } |
e2500fed | 2283 | |
b5b8b0ac AO |
2284 | /* Dump the cselib_val *X to FILE *info. */ |
2285 | ||
2286 | static int | |
2287 | dump_cselib_val (void **x, void *info) | |
2288 | { | |
2289 | cselib_val *v = (cselib_val *)*x; | |
2290 | FILE *out = (FILE *)info; | |
2291 | bool need_lf = true; | |
2292 | ||
2293 | print_inline_rtx (out, v->val_rtx, 0); | |
2294 | ||
2295 | if (v->locs) | |
2296 | { | |
2297 | struct elt_loc_list *l = v->locs; | |
2298 | if (need_lf) | |
2299 | { | |
2300 | fputc ('\n', out); | |
2301 | need_lf = false; | |
2302 | } | |
2303 | fputs (" locs:", out); | |
2304 | do | |
2305 | { | |
2306 | fprintf (out, "\n from insn %i ", | |
2307 | INSN_UID (l->setting_insn)); | |
2308 | print_inline_rtx (out, l->loc, 4); | |
2309 | } | |
2310 | while ((l = l->next)); | |
2311 | fputc ('\n', out); | |
2312 | } | |
2313 | else | |
2314 | { | |
2315 | fputs (" no locs", out); | |
2316 | need_lf = true; | |
2317 | } | |
2318 | ||
2319 | if (v->addr_list) | |
2320 | { | |
2321 | struct elt_list *e = v->addr_list; | |
2322 | if (need_lf) | |
2323 | { | |
2324 | fputc ('\n', out); | |
2325 | need_lf = false; | |
2326 | } | |
2327 | fputs (" addr list:", out); | |
2328 | do | |
2329 | { | |
2330 | fputs ("\n ", out); | |
2331 | print_inline_rtx (out, e->elt->val_rtx, 2); | |
2332 | } | |
2333 | while ((e = e->next)); | |
2334 | fputc ('\n', out); | |
2335 | } | |
2336 | else | |
2337 | { | |
2338 | fputs (" no addrs", out); | |
2339 | need_lf = true; | |
2340 | } | |
2341 | ||
2342 | if (v->next_containing_mem == &dummy_val) | |
2343 | fputs (" last mem\n", out); | |
2344 | else if (v->next_containing_mem) | |
2345 | { | |
2346 | fputs (" next mem ", out); | |
2347 | print_inline_rtx (out, v->next_containing_mem->val_rtx, 2); | |
2348 | fputc ('\n', out); | |
2349 | } | |
2350 | else if (need_lf) | |
2351 | fputc ('\n', out); | |
2352 | ||
2353 | return 1; | |
2354 | } | |
2355 | ||
2356 | /* Dump to OUT everything in the CSELIB table. */ | |
2357 | ||
2358 | void | |
2359 | dump_cselib_table (FILE *out) | |
2360 | { | |
2361 | fprintf (out, "cselib hash table:\n"); | |
2362 | htab_traverse (cselib_hash_table, dump_cselib_val, out); | |
2363 | if (first_containing_mem != &dummy_val) | |
2364 | { | |
2365 | fputs ("first mem ", out); | |
2366 | print_inline_rtx (out, first_containing_mem->val_rtx, 2); | |
2367 | fputc ('\n', out); | |
2368 | } | |
5440c0e7 | 2369 | fprintf (out, "next uid %i\n", next_uid); |
b5b8b0ac AO |
2370 | } |
2371 | ||
e2500fed | 2372 | #include "gt-cselib.h" |