]>
Commit | Line | Data |
---|---|---|
8f8cadbc | 1 | /* Perform simple optimizations to clean up the result of reload. |
fbd26352 | 2 | Copyright (C) 1987-2019 Free Software Foundation, Inc. |
8f8cadbc | 3 | |
4 | This file is part of GCC. | |
5 | ||
6 | GCC is free software; you can redistribute it and/or modify it under | |
7 | the terms of the GNU General Public License as published by the Free | |
8c4c00c1 | 8 | Software Foundation; either version 3, or (at your option) any later |
8f8cadbc | 9 | version. |
10 | ||
11 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
12 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
13 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
14 | for more details. | |
15 | ||
16 | You should have received a copy of the GNU General Public License | |
8c4c00c1 | 17 | along with GCC; see the file COPYING3. If not see |
18 | <http://www.gnu.org/licenses/>. */ | |
8f8cadbc | 19 | |
20 | #include "config.h" | |
21 | #include "system.h" | |
22 | #include "coretypes.h" | |
9ef16211 | 23 | #include "backend.h" |
7c29e30e | 24 | #include "target.h" |
8f8cadbc | 25 | #include "rtl.h" |
7c29e30e | 26 | #include "tree.h" |
27 | #include "predict.h" | |
9ef16211 | 28 | #include "df.h" |
ad7b10a2 | 29 | #include "memmodel.h" |
8f8cadbc | 30 | #include "tm_p.h" |
7c29e30e | 31 | #include "optabs.h" |
32 | #include "regs.h" | |
33 | #include "emit-rtl.h" | |
34 | #include "recog.h" | |
7c29e30e | 35 | |
94ea8568 | 36 | #include "cfgrtl.h" |
37 | #include "cfgbuild.h" | |
38 | #include "cfgcleanup.h" | |
8f8cadbc | 39 | #include "reload.h" |
8f8cadbc | 40 | #include "cselib.h" |
77fce4cd | 41 | #include "tree-pass.h" |
3072d30e | 42 | #include "dbgcnt.h" |
8f8cadbc | 43 | |
3ad4992f | 44 | static int reload_cse_noop_set_p (rtx); |
3aeaa53f | 45 | static bool reload_cse_simplify (rtx_insn *, rtx); |
26709122 | 46 | static void reload_cse_regs_1 (void); |
3aeaa53f | 47 | static int reload_cse_simplify_set (rtx, rtx_insn *); |
48 | static int reload_cse_simplify_operands (rtx_insn *, rtx); | |
8f8cadbc | 49 | |
3ad4992f | 50 | static void reload_combine (void); |
3aeaa53f | 51 | static void reload_combine_note_use (rtx *, rtx_insn *, int, rtx); |
81a410b1 | 52 | static void reload_combine_note_store (rtx, const_rtx, void *); |
8f8cadbc | 53 | |
3aeaa53f | 54 | static bool reload_cse_move2add (rtx_insn *); |
81a410b1 | 55 | static void move2add_note_store (rtx, const_rtx, void *); |
8f8cadbc | 56 | |
57 | /* Call cse / combine like post-reload optimization phases. | |
58 | FIRST is the first instruction. */ | |
98799adc | 59 | |
60 | static void | |
3aeaa53f | 61 | reload_cse_regs (rtx_insn *first ATTRIBUTE_UNUSED) |
8f8cadbc | 62 | { |
d83ccc81 | 63 | bool moves_converted; |
26709122 | 64 | reload_cse_regs_1 (); |
8f8cadbc | 65 | reload_combine (); |
d83ccc81 | 66 | moves_converted = reload_cse_move2add (first); |
8f8cadbc | 67 | if (flag_expensive_optimizations) |
d83ccc81 | 68 | { |
69 | if (moves_converted) | |
70 | reload_combine (); | |
26709122 | 71 | reload_cse_regs_1 (); |
d83ccc81 | 72 | } |
8f8cadbc | 73 | } |
74 | ||
75 | /* See whether a single set SET is a noop. */ | |
76 | static int | |
3ad4992f | 77 | reload_cse_noop_set_p (rtx set) |
8f8cadbc | 78 | { |
79 | if (cselib_reg_set_mode (SET_DEST (set)) != GET_MODE (SET_DEST (set))) | |
80 | return 0; | |
81 | ||
82 | return rtx_equal_for_cselib_p (SET_DEST (set), SET_SRC (set)); | |
83 | } | |
84 | ||
26709122 | 85 | /* Try to simplify INSN. Return true if the CFG may have changed. */ |
86 | static bool | |
3aeaa53f | 87 | reload_cse_simplify (rtx_insn *insn, rtx testreg) |
8f8cadbc | 88 | { |
89 | rtx body = PATTERN (insn); | |
26709122 | 90 | basic_block insn_bb = BLOCK_FOR_INSN (insn); |
91 | unsigned insn_bb_succs = EDGE_COUNT (insn_bb->succs); | |
8f8cadbc | 92 | |
94f4da1b | 93 | /* If NO_FUNCTION_CSE has been set by the target, then we should not try |
94 | to cse function calls. */ | |
95 | if (NO_FUNCTION_CSE && CALL_P (insn)) | |
96 | return false; | |
97 | ||
8f8cadbc | 98 | if (GET_CODE (body) == SET) |
99 | { | |
100 | int count = 0; | |
101 | ||
102 | /* Simplify even if we may think it is a no-op. | |
103 | We may think a memory load of a value smaller than WORD_SIZE | |
104 | is redundant because we haven't taken into account possible | |
105 | implicit extension. reload_cse_simplify_set() will bring | |
106 | this out, so it's safer to simplify before we delete. */ | |
107 | count += reload_cse_simplify_set (body, insn); | |
108 | ||
109 | if (!count && reload_cse_noop_set_p (body)) | |
110 | { | |
5a9ecd4a | 111 | if (check_for_inc_dec (insn)) |
112 | delete_insn_and_edges (insn); | |
26709122 | 113 | /* We're done with this insn. */ |
114 | goto done; | |
8f8cadbc | 115 | } |
116 | ||
117 | if (count > 0) | |
118 | apply_change_group (); | |
119 | else | |
120 | reload_cse_simplify_operands (insn, testreg); | |
121 | } | |
122 | else if (GET_CODE (body) == PARALLEL) | |
123 | { | |
124 | int i; | |
125 | int count = 0; | |
126 | rtx value = NULL_RTX; | |
127 | ||
17883489 | 128 | /* Registers mentioned in the clobber list for an asm cannot be reused |
129 | within the body of the asm. Invalidate those registers now so that | |
130 | we don't try to substitute values for them. */ | |
131 | if (asm_noperands (body) >= 0) | |
132 | { | |
133 | for (i = XVECLEN (body, 0) - 1; i >= 0; --i) | |
134 | { | |
135 | rtx part = XVECEXP (body, 0, i); | |
70bdfe23 | 136 | /* asms can only have full clobbers, not clobber_highs. */ |
137 | gcc_assert (GET_CODE (part) != CLOBBER_HIGH); | |
17883489 | 138 | if (GET_CODE (part) == CLOBBER && REG_P (XEXP (part, 0))) |
139 | cselib_invalidate_rtx (XEXP (part, 0)); | |
140 | } | |
141 | } | |
142 | ||
8f8cadbc | 143 | /* If every action in a PARALLEL is a noop, we can delete |
144 | the entire PARALLEL. */ | |
145 | for (i = XVECLEN (body, 0) - 1; i >= 0; --i) | |
146 | { | |
147 | rtx part = XVECEXP (body, 0, i); | |
148 | if (GET_CODE (part) == SET) | |
149 | { | |
150 | if (! reload_cse_noop_set_p (part)) | |
151 | break; | |
152 | if (REG_P (SET_DEST (part)) | |
153 | && REG_FUNCTION_VALUE_P (SET_DEST (part))) | |
154 | { | |
155 | if (value) | |
156 | break; | |
157 | value = SET_DEST (part); | |
158 | } | |
159 | } | |
430d0b16 | 160 | else if (GET_CODE (part) != CLOBBER |
70bdfe23 | 161 | && GET_CODE (part) != CLOBBER_HIGH |
430d0b16 | 162 | && GET_CODE (part) != USE) |
8f8cadbc | 163 | break; |
164 | } | |
165 | ||
166 | if (i < 0) | |
167 | { | |
5a9ecd4a | 168 | if (check_for_inc_dec (insn)) |
169 | delete_insn_and_edges (insn); | |
8f8cadbc | 170 | /* We're done with this insn. */ |
26709122 | 171 | goto done; |
8f8cadbc | 172 | } |
173 | ||
174 | /* It's not a no-op, but we can try to simplify it. */ | |
175 | for (i = XVECLEN (body, 0) - 1; i >= 0; --i) | |
176 | if (GET_CODE (XVECEXP (body, 0, i)) == SET) | |
177 | count += reload_cse_simplify_set (XVECEXP (body, 0, i), insn); | |
178 | ||
179 | if (count > 0) | |
180 | apply_change_group (); | |
181 | else | |
182 | reload_cse_simplify_operands (insn, testreg); | |
183 | } | |
26709122 | 184 | |
185 | done: | |
186 | return (EDGE_COUNT (insn_bb->succs) != insn_bb_succs); | |
8f8cadbc | 187 | } |
188 | ||
189 | /* Do a very simple CSE pass over the hard registers. | |
190 | ||
191 | This function detects no-op moves where we happened to assign two | |
192 | different pseudo-registers to the same hard register, and then | |
193 | copied one to the other. Reload will generate a useless | |
194 | instruction copying a register to itself. | |
195 | ||
196 | This function also detects cases where we load a value from memory | |
197 | into two different registers, and (if memory is more expensive than | |
198 | registers) changes it to simply copy the first register into the | |
199 | second register. | |
200 | ||
201 | Another optimization is performed that scans the operands of each | |
202 | instruction to see whether the value is already available in a | |
203 | hard register. It then replaces the operand with the hard register | |
204 | if possible, much like an optional reload would. */ | |
205 | ||
206 | static void | |
26709122 | 207 | reload_cse_regs_1 (void) |
8f8cadbc | 208 | { |
26709122 | 209 | bool cfg_changed = false; |
210 | basic_block bb; | |
3aeaa53f | 211 | rtx_insn *insn; |
dcd6d0f4 | 212 | rtx testreg = gen_rtx_REG (word_mode, LAST_VIRTUAL_REGISTER + 1); |
8f8cadbc | 213 | |
35af0188 | 214 | cselib_init (CSELIB_RECORD_MEMORY); |
8f8cadbc | 215 | init_alias_analysis (); |
216 | ||
fc00614f | 217 | FOR_EACH_BB_FN (bb, cfun) |
26709122 | 218 | FOR_BB_INSNS (bb, insn) |
219 | { | |
220 | if (INSN_P (insn)) | |
221 | cfg_changed |= reload_cse_simplify (insn, testreg); | |
8f8cadbc | 222 | |
26709122 | 223 | cselib_process_insn (insn); |
224 | } | |
8f8cadbc | 225 | |
226 | /* Clean up. */ | |
227 | end_alias_analysis (); | |
228 | cselib_finish (); | |
26709122 | 229 | if (cfg_changed) |
230 | cleanup_cfg (0); | |
8f8cadbc | 231 | } |
232 | ||
233 | /* Try to simplify a single SET instruction. SET is the set pattern. | |
234 | INSN is the instruction it came from. | |
235 | This function only handles one case: if we set a register to a value | |
236 | which is not a register, we try to find that value in some other register | |
237 | and change the set into a register copy. */ | |
238 | ||
239 | static int | |
3aeaa53f | 240 | reload_cse_simplify_set (rtx set, rtx_insn *insn) |
8f8cadbc | 241 | { |
242 | int did_change = 0; | |
243 | int dreg; | |
244 | rtx src; | |
ade444a4 | 245 | reg_class_t dclass; |
8f8cadbc | 246 | int old_cost; |
247 | cselib_val *val; | |
248 | struct elt_loc_list *l; | |
21f1e711 | 249 | enum rtx_code extend_op = UNKNOWN; |
f529eb25 | 250 | bool speed = optimize_bb_for_speed_p (BLOCK_FOR_INSN (insn)); |
8f8cadbc | 251 | |
252 | dreg = true_regnum (SET_DEST (set)); | |
253 | if (dreg < 0) | |
254 | return 0; | |
255 | ||
256 | src = SET_SRC (set); | |
257 | if (side_effects_p (src) || true_regnum (src) >= 0) | |
258 | return 0; | |
259 | ||
260 | dclass = REGNO_REG_CLASS (dreg); | |
261 | ||
8f8cadbc | 262 | /* When replacing a memory with a register, we need to honor assumptions |
263 | that combine made wrt the contents of sign bits. We'll do this by | |
264 | generating an extend instruction instead of a reg->reg copy. Thus | |
265 | the destination must be a register that we can widen. */ | |
e16ceb8e | 266 | if (MEM_P (src) |
e73fe78f | 267 | && (extend_op = load_extend_op (GET_MODE (src))) != UNKNOWN |
8ad4c111 | 268 | && !REG_P (SET_DEST (set))) |
8f8cadbc | 269 | return 0; |
8f8cadbc | 270 | |
1f864115 | 271 | val = cselib_lookup (src, GET_MODE (SET_DEST (set)), 0, VOIDmode); |
3be01943 | 272 | if (! val) |
273 | return 0; | |
274 | ||
8f8cadbc | 275 | /* If memory loads are cheaper than register copies, don't change them. */ |
e16ceb8e | 276 | if (MEM_P (src)) |
251a613e | 277 | old_cost = memory_move_cost (GET_MODE (src), dclass, true); |
8ad4c111 | 278 | else if (REG_P (src)) |
e6078fbb | 279 | old_cost = register_move_cost (GET_MODE (src), |
8f8cadbc | 280 | REGNO_REG_CLASS (REGNO (src)), dclass); |
281 | else | |
5ae4887d | 282 | old_cost = set_src_cost (src, GET_MODE (SET_DEST (set)), speed); |
8f8cadbc | 283 | |
8f8cadbc | 284 | for (l = val->locs; l; l = l->next) |
285 | { | |
286 | rtx this_rtx = l->loc; | |
287 | int this_cost; | |
288 | ||
289 | if (CONSTANT_P (this_rtx) && ! references_value_p (this_rtx, 0)) | |
290 | { | |
21f1e711 | 291 | if (extend_op != UNKNOWN) |
8f8cadbc | 292 | { |
e913b5cd | 293 | wide_int result; |
8f8cadbc | 294 | |
e913b5cd | 295 | if (!CONST_SCALAR_INT_P (this_rtx)) |
8f8cadbc | 296 | continue; |
297 | ||
8f8cadbc | 298 | switch (extend_op) |
299 | { | |
300 | case ZERO_EXTEND: | |
c67875ad | 301 | result = wide_int::from (rtx_mode_t (this_rtx, |
302 | GET_MODE (src)), | |
ecc41f48 | 303 | BITS_PER_WORD, UNSIGNED); |
8f8cadbc | 304 | break; |
305 | case SIGN_EXTEND: | |
c67875ad | 306 | result = wide_int::from (rtx_mode_t (this_rtx, |
307 | GET_MODE (src)), | |
ecc41f48 | 308 | BITS_PER_WORD, SIGNED); |
e913b5cd | 309 | break; |
8f8cadbc | 310 | default: |
876760f6 | 311 | gcc_unreachable (); |
8f8cadbc | 312 | } |
ecc41f48 | 313 | this_rtx = immed_wide_int_const (result, word_mode); |
8f8cadbc | 314 | } |
5fe18e78 | 315 | |
5ae4887d | 316 | this_cost = set_src_cost (this_rtx, GET_MODE (SET_DEST (set)), speed); |
8f8cadbc | 317 | } |
8ad4c111 | 318 | else if (REG_P (this_rtx)) |
8f8cadbc | 319 | { |
21f1e711 | 320 | if (extend_op != UNKNOWN) |
8f8cadbc | 321 | { |
322 | this_rtx = gen_rtx_fmt_e (extend_op, word_mode, this_rtx); | |
5ae4887d | 323 | this_cost = set_src_cost (this_rtx, word_mode, speed); |
8f8cadbc | 324 | } |
325 | else | |
e6078fbb | 326 | this_cost = register_move_cost (GET_MODE (this_rtx), |
8f8cadbc | 327 | REGNO_REG_CLASS (REGNO (this_rtx)), |
328 | dclass); | |
329 | } | |
330 | else | |
331 | continue; | |
332 | ||
333 | /* If equal costs, prefer registers over anything else. That | |
334 | tends to lead to smaller instructions on some machines. */ | |
335 | if (this_cost < old_cost | |
336 | || (this_cost == old_cost | |
8ad4c111 | 337 | && REG_P (this_rtx) |
338 | && !REG_P (SET_SRC (set)))) | |
8f8cadbc | 339 | { |
e73fe78f | 340 | if (extend_op != UNKNOWN |
b56a9dbc | 341 | && REG_CAN_CHANGE_MODE_P (REGNO (SET_DEST (set)), |
342 | GET_MODE (SET_DEST (set)), word_mode)) | |
8f8cadbc | 343 | { |
344 | rtx wide_dest = gen_rtx_REG (word_mode, REGNO (SET_DEST (set))); | |
345 | ORIGINAL_REGNO (wide_dest) = ORIGINAL_REGNO (SET_DEST (set)); | |
346 | validate_change (insn, &SET_DEST (set), wide_dest, 1); | |
347 | } | |
8f8cadbc | 348 | |
11d686e2 | 349 | validate_unshare_change (insn, &SET_SRC (set), this_rtx, 1); |
8f8cadbc | 350 | old_cost = this_cost, did_change = 1; |
351 | } | |
352 | } | |
353 | ||
354 | return did_change; | |
355 | } | |
356 | ||
357 | /* Try to replace operands in INSN with equivalent values that are already | |
358 | in registers. This can be viewed as optional reloading. | |
359 | ||
360 | For each non-register operand in the insn, see if any hard regs are | |
361 | known to be equivalent to that operand. Record the alternatives which | |
362 | can accept these hard registers. Among all alternatives, select the | |
363 | ones which are better or equal to the one currently matching, where | |
364 | "better" is in terms of '?' and '!' constraints. Among the remaining | |
365 | alternatives, select the one which replaces most operands with | |
366 | hard registers. */ | |
367 | ||
368 | static int | |
3aeaa53f | 369 | reload_cse_simplify_operands (rtx_insn *insn, rtx testreg) |
8f8cadbc | 370 | { |
371 | int i, j; | |
372 | ||
373 | /* For each operand, all registers that are equivalent to it. */ | |
374 | HARD_REG_SET equiv_regs[MAX_RECOG_OPERANDS]; | |
375 | ||
376 | const char *constraints[MAX_RECOG_OPERANDS]; | |
377 | ||
378 | /* Vector recording how bad an alternative is. */ | |
379 | int *alternative_reject; | |
380 | /* Vector recording how many registers can be introduced by choosing | |
381 | this alternative. */ | |
382 | int *alternative_nregs; | |
383 | /* Array of vectors recording, for each operand and each alternative, | |
384 | which hard register to substitute, or -1 if the operand should be | |
385 | left as it is. */ | |
386 | int *op_alt_regno[MAX_RECOG_OPERANDS]; | |
387 | /* Array of alternatives, sorted in order of decreasing desirability. */ | |
388 | int *alternative_order; | |
389 | ||
835b8178 | 390 | extract_constrain_insn (insn); |
8f8cadbc | 391 | |
392 | if (recog_data.n_alternatives == 0 || recog_data.n_operands == 0) | |
393 | return 0; | |
394 | ||
4077bf7a | 395 | alternative_reject = XALLOCAVEC (int, recog_data.n_alternatives); |
396 | alternative_nregs = XALLOCAVEC (int, recog_data.n_alternatives); | |
397 | alternative_order = XALLOCAVEC (int, recog_data.n_alternatives); | |
f0af5a88 | 398 | memset (alternative_reject, 0, recog_data.n_alternatives * sizeof (int)); |
399 | memset (alternative_nregs, 0, recog_data.n_alternatives * sizeof (int)); | |
8f8cadbc | 400 | |
401 | /* For each operand, find out which regs are equivalent. */ | |
402 | for (i = 0; i < recog_data.n_operands; i++) | |
403 | { | |
404 | cselib_val *v; | |
405 | struct elt_loc_list *l; | |
9d9e3c81 | 406 | rtx op; |
8f8cadbc | 407 | |
408 | CLEAR_HARD_REG_SET (equiv_regs[i]); | |
409 | ||
410 | /* cselib blows up on CODE_LABELs. Trying to fix that doesn't seem | |
dca0c3a8 | 411 | right, so avoid the problem here. Similarly NOTE_INSN_DELETED_LABEL. |
412 | Likewise if we have a constant and the insn pattern doesn't tell us | |
413 | the mode we need. */ | |
6d7dc5b9 | 414 | if (LABEL_P (recog_data.operand[i]) |
dca0c3a8 | 415 | || (NOTE_P (recog_data.operand[i]) |
416 | && NOTE_KIND (recog_data.operand[i]) == NOTE_INSN_DELETED_LABEL) | |
8f8cadbc | 417 | || (CONSTANT_P (recog_data.operand[i]) |
418 | && recog_data.operand_mode[i] == VOIDmode)) | |
419 | continue; | |
420 | ||
9d9e3c81 | 421 | op = recog_data.operand[i]; |
e73fe78f | 422 | if (MEM_P (op) && load_extend_op (GET_MODE (op)) != UNKNOWN) |
9d9e3c81 | 423 | { |
424 | rtx set = single_set (insn); | |
425 | ||
4885b286 | 426 | /* We might have multiple sets, some of which do implicit |
9d9e3c81 | 427 | extension. Punt on this for now. */ |
428 | if (! set) | |
429 | continue; | |
86481e89 | 430 | /* If the destination is also a MEM or a STRICT_LOW_PART, no |
9d9e3c81 | 431 | extension applies. |
432 | Also, if there is an explicit extension, we don't have to | |
433 | worry about an implicit one. */ | |
e16ceb8e | 434 | else if (MEM_P (SET_DEST (set)) |
9d9e3c81 | 435 | || GET_CODE (SET_DEST (set)) == STRICT_LOW_PART |
436 | || GET_CODE (SET_SRC (set)) == ZERO_EXTEND | |
437 | || GET_CODE (SET_SRC (set)) == SIGN_EXTEND) | |
438 | ; /* Continue ordinary processing. */ | |
a091e4f5 | 439 | /* If the register cannot change mode to word_mode, it follows that |
440 | it cannot have been used in word_mode. */ | |
8ad4c111 | 441 | else if (REG_P (SET_DEST (set)) |
b56a9dbc | 442 | && !REG_CAN_CHANGE_MODE_P (REGNO (SET_DEST (set)), |
443 | GET_MODE (SET_DEST (set)), | |
444 | word_mode)) | |
a091e4f5 | 445 | ; /* Continue ordinary processing. */ |
9d9e3c81 | 446 | /* If this is a straight load, make the extension explicit. */ |
8ad4c111 | 447 | else if (REG_P (SET_DEST (set)) |
9d9e3c81 | 448 | && recog_data.n_operands == 2 |
449 | && SET_SRC (set) == op | |
450 | && SET_DEST (set) == recog_data.operand[1-i]) | |
451 | { | |
452 | validate_change (insn, recog_data.operand_loc[i], | |
e73fe78f | 453 | gen_rtx_fmt_e (load_extend_op (GET_MODE (op)), |
9d9e3c81 | 454 | word_mode, op), |
455 | 1); | |
456 | validate_change (insn, recog_data.operand_loc[1-i], | |
457 | gen_rtx_REG (word_mode, REGNO (SET_DEST (set))), | |
458 | 1); | |
459 | if (! apply_change_group ()) | |
460 | return 0; | |
461 | return reload_cse_simplify_operands (insn, testreg); | |
462 | } | |
463 | else | |
464 | /* ??? There might be arithmetic operations with memory that are | |
465 | safe to optimize, but is it worth the trouble? */ | |
466 | continue; | |
467 | } | |
5fe18e78 | 468 | |
017b7047 | 469 | if (side_effects_p (op)) |
470 | continue; | |
1f864115 | 471 | v = cselib_lookup (op, recog_data.operand_mode[i], 0, VOIDmode); |
8f8cadbc | 472 | if (! v) |
473 | continue; | |
474 | ||
475 | for (l = v->locs; l; l = l->next) | |
8ad4c111 | 476 | if (REG_P (l->loc)) |
8f8cadbc | 477 | SET_HARD_REG_BIT (equiv_regs[i], REGNO (l->loc)); |
478 | } | |
479 | ||
e1a797ad | 480 | alternative_mask preferred = get_preferred_alternatives (insn); |
8f8cadbc | 481 | for (i = 0; i < recog_data.n_operands; i++) |
482 | { | |
3754d046 | 483 | machine_mode mode; |
8f8cadbc | 484 | int regno; |
485 | const char *p; | |
486 | ||
4077bf7a | 487 | op_alt_regno[i] = XALLOCAVEC (int, recog_data.n_alternatives); |
8f8cadbc | 488 | for (j = 0; j < recog_data.n_alternatives; j++) |
489 | op_alt_regno[i][j] = -1; | |
490 | ||
491 | p = constraints[i] = recog_data.constraints[i]; | |
492 | mode = recog_data.operand_mode[i]; | |
493 | ||
494 | /* Add the reject values for each alternative given by the constraints | |
495 | for this operand. */ | |
496 | j = 0; | |
497 | while (*p != '\0') | |
498 | { | |
499 | char c = *p++; | |
500 | if (c == ',') | |
501 | j++; | |
502 | else if (c == '?') | |
503 | alternative_reject[j] += 3; | |
504 | else if (c == '!') | |
505 | alternative_reject[j] += 300; | |
506 | } | |
507 | ||
508 | /* We won't change operands which are already registers. We | |
509 | also don't want to modify output operands. */ | |
510 | regno = true_regnum (recog_data.operand[i]); | |
511 | if (regno >= 0 | |
512 | || constraints[i][0] == '=' | |
513 | || constraints[i][0] == '+') | |
514 | continue; | |
515 | ||
516 | for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++) | |
517 | { | |
b9c74b4d | 518 | enum reg_class rclass = NO_REGS; |
8f8cadbc | 519 | |
520 | if (! TEST_HARD_REG_BIT (equiv_regs[i], regno)) | |
521 | continue; | |
522 | ||
937ca48e | 523 | set_mode_and_regno (testreg, mode, regno); |
8f8cadbc | 524 | |
525 | /* We found a register equal to this operand. Now look for all | |
526 | alternatives that can accept this register and have not been | |
527 | assigned a register they can use yet. */ | |
528 | j = 0; | |
529 | p = constraints[i]; | |
530 | for (;;) | |
531 | { | |
532 | char c = *p; | |
533 | ||
534 | switch (c) | |
535 | { | |
69449463 | 536 | case 'g': |
537 | rclass = reg_class_subunion[rclass][GENERAL_REGS]; | |
8f8cadbc | 538 | break; |
539 | ||
540 | default: | |
6659485c | 541 | rclass |
8f8cadbc | 542 | = (reg_class_subunion |
79bc09fb | 543 | [rclass] |
544 | [reg_class_for_constraint (lookup_constraint (p))]); | |
8f8cadbc | 545 | break; |
546 | ||
547 | case ',': case '\0': | |
548 | /* See if REGNO fits this alternative, and set it up as the | |
549 | replacement register if we don't have one for this | |
550 | alternative yet and the operand being replaced is not | |
551 | a cheap CONST_INT. */ | |
552 | if (op_alt_regno[i][j] == -1 | |
e1a797ad | 553 | && TEST_BIT (preferred, j) |
6659485c | 554 | && reg_fits_class_p (testreg, rclass, 0, mode) |
971ba038 | 555 | && (!CONST_INT_P (recog_data.operand[i]) |
5ae4887d | 556 | || (set_src_cost (recog_data.operand[i], mode, |
7013e87c | 557 | optimize_bb_for_speed_p |
558 | (BLOCK_FOR_INSN (insn))) | |
5ae4887d | 559 | > set_src_cost (testreg, mode, |
7013e87c | 560 | optimize_bb_for_speed_p |
561 | (BLOCK_FOR_INSN (insn)))))) | |
8f8cadbc | 562 | { |
563 | alternative_nregs[j]++; | |
564 | op_alt_regno[i][j] = regno; | |
565 | } | |
566 | j++; | |
b9c74b4d | 567 | rclass = NO_REGS; |
8f8cadbc | 568 | break; |
569 | } | |
570 | p += CONSTRAINT_LEN (c, p); | |
571 | ||
572 | if (c == '\0') | |
573 | break; | |
574 | } | |
575 | } | |
576 | } | |
577 | ||
578 | /* Record all alternatives which are better or equal to the currently | |
579 | matching one in the alternative_order array. */ | |
580 | for (i = j = 0; i < recog_data.n_alternatives; i++) | |
581 | if (alternative_reject[i] <= alternative_reject[which_alternative]) | |
582 | alternative_order[j++] = i; | |
583 | recog_data.n_alternatives = j; | |
584 | ||
585 | /* Sort it. Given a small number of alternatives, a dumb algorithm | |
586 | won't hurt too much. */ | |
587 | for (i = 0; i < recog_data.n_alternatives - 1; i++) | |
588 | { | |
589 | int best = i; | |
590 | int best_reject = alternative_reject[alternative_order[i]]; | |
591 | int best_nregs = alternative_nregs[alternative_order[i]]; | |
8f8cadbc | 592 | |
593 | for (j = i + 1; j < recog_data.n_alternatives; j++) | |
594 | { | |
595 | int this_reject = alternative_reject[alternative_order[j]]; | |
596 | int this_nregs = alternative_nregs[alternative_order[j]]; | |
597 | ||
598 | if (this_reject < best_reject | |
c2d0cf41 | 599 | || (this_reject == best_reject && this_nregs > best_nregs)) |
8f8cadbc | 600 | { |
601 | best = j; | |
602 | best_reject = this_reject; | |
603 | best_nregs = this_nregs; | |
604 | } | |
605 | } | |
606 | ||
dfcf26a5 | 607 | std::swap (alternative_order[best], alternative_order[i]); |
8f8cadbc | 608 | } |
609 | ||
610 | /* Substitute the operands as determined by op_alt_regno for the best | |
611 | alternative. */ | |
612 | j = alternative_order[0]; | |
613 | ||
614 | for (i = 0; i < recog_data.n_operands; i++) | |
615 | { | |
3754d046 | 616 | machine_mode mode = recog_data.operand_mode[i]; |
8f8cadbc | 617 | if (op_alt_regno[i][j] == -1) |
618 | continue; | |
619 | ||
620 | validate_change (insn, recog_data.operand_loc[i], | |
621 | gen_rtx_REG (mode, op_alt_regno[i][j]), 1); | |
622 | } | |
623 | ||
624 | for (i = recog_data.n_dups - 1; i >= 0; i--) | |
625 | { | |
626 | int op = recog_data.dup_num[i]; | |
3754d046 | 627 | machine_mode mode = recog_data.operand_mode[op]; |
8f8cadbc | 628 | |
629 | if (op_alt_regno[op][j] == -1) | |
630 | continue; | |
631 | ||
632 | validate_change (insn, recog_data.dup_loc[i], | |
633 | gen_rtx_REG (mode, op_alt_regno[op][j]), 1); | |
634 | } | |
635 | ||
636 | return apply_change_group (); | |
637 | } | |
638 | \f | |
639 | /* If reload couldn't use reg+reg+offset addressing, try to use reg+reg | |
640 | addressing now. | |
641 | This code might also be useful when reload gave up on reg+reg addressing | |
642 | because of clashes between the return register and INDEX_REG_CLASS. */ | |
643 | ||
644 | /* The maximum number of uses of a register we can keep track of to | |
645 | replace them with reg+reg addressing. */ | |
d83ccc81 | 646 | #define RELOAD_COMBINE_MAX_USES 16 |
8f8cadbc | 647 | |
d83ccc81 | 648 | /* Describes a recorded use of a register. */ |
649 | struct reg_use | |
650 | { | |
651 | /* The insn where a register has been used. */ | |
3aeaa53f | 652 | rtx_insn *insn; |
d83ccc81 | 653 | /* Points to the memory reference enclosing the use, if any, NULL_RTX |
654 | otherwise. */ | |
655 | rtx containing_mem; | |
9d75589a | 656 | /* Location of the register within INSN. */ |
d83ccc81 | 657 | rtx *usep; |
658 | /* The reverse uid of the insn. */ | |
659 | int ruid; | |
660 | }; | |
8f8cadbc | 661 | |
662 | /* If the register is used in some unknown fashion, USE_INDEX is negative. | |
663 | If it is dead, USE_INDEX is RELOAD_COMBINE_MAX_USES, and STORE_RUID | |
d83ccc81 | 664 | indicates where it is first set or clobbered. |
8f8cadbc | 665 | Otherwise, USE_INDEX is the index of the last encountered use of the |
d83ccc81 | 666 | register (which is first among these we have seen since we scan backwards). |
667 | USE_RUID indicates the first encountered, i.e. last, of these uses. | |
668 | If ALL_OFFSETS_MATCH is true, all encountered uses were inside a PLUS | |
669 | with a constant offset; OFFSET contains this constant in that case. | |
8f8cadbc | 670 | STORE_RUID is always meaningful if we only want to use a value in a |
671 | register in a different place: it denotes the next insn in the insn | |
d83ccc81 | 672 | stream (i.e. the last encountered) that sets or clobbers the register. |
70bdfe23 | 673 | REAL_STORE_RUID is similar, but clobbers are ignored when updating it. |
674 | EXPR is the expression used when storing the register. */ | |
8f8cadbc | 675 | static struct |
676 | { | |
677 | struct reg_use reg_use[RELOAD_COMBINE_MAX_USES]; | |
8f8cadbc | 678 | rtx offset; |
d83ccc81 | 679 | int use_index; |
8f8cadbc | 680 | int store_ruid; |
d83ccc81 | 681 | int real_store_ruid; |
8f8cadbc | 682 | int use_ruid; |
d83ccc81 | 683 | bool all_offsets_match; |
70bdfe23 | 684 | rtx expr; |
8f8cadbc | 685 | } reg_state[FIRST_PSEUDO_REGISTER]; |
686 | ||
687 | /* Reverse linear uid. This is increased in reload_combine while scanning | |
688 | the instructions from last to first. It is used to set last_label_ruid | |
689 | and the store_ruid / use_ruid fields in reg_state. */ | |
690 | static int reload_combine_ruid; | |
691 | ||
fb79f695 | 692 | /* The RUID of the last label we encountered in reload_combine. */ |
693 | static int last_label_ruid; | |
694 | ||
d83ccc81 | 695 | /* The RUID of the last jump we encountered in reload_combine. */ |
696 | static int last_jump_ruid; | |
697 | ||
fb79f695 | 698 | /* The register numbers of the first and last index register. A value of |
699 | -1 in LAST_INDEX_REG indicates that we've previously computed these | |
700 | values and found no suitable index registers. */ | |
701 | static int first_index_reg = -1; | |
702 | static int last_index_reg; | |
703 | ||
8f8cadbc | 704 | #define LABEL_LIVE(LABEL) \ |
705 | (label_live[CODE_LABEL_NUMBER (LABEL) - min_labelno]) | |
706 | ||
d83ccc81 | 707 | /* Subroutine of reload_combine_split_ruids, called to fix up a single |
708 | ruid pointed to by *PRUID if it is higher than SPLIT_RUID. */ | |
709 | ||
710 | static inline void | |
711 | reload_combine_split_one_ruid (int *pruid, int split_ruid) | |
712 | { | |
713 | if (*pruid > split_ruid) | |
714 | (*pruid)++; | |
715 | } | |
716 | ||
717 | /* Called when we insert a new insn in a position we've already passed in | |
718 | the scan. Examine all our state, increasing all ruids that are higher | |
719 | than SPLIT_RUID by one in order to make room for a new insn. */ | |
720 | ||
721 | static void | |
722 | reload_combine_split_ruids (int split_ruid) | |
723 | { | |
724 | unsigned i; | |
725 | ||
726 | reload_combine_split_one_ruid (&reload_combine_ruid, split_ruid); | |
727 | reload_combine_split_one_ruid (&last_label_ruid, split_ruid); | |
728 | reload_combine_split_one_ruid (&last_jump_ruid, split_ruid); | |
729 | ||
730 | for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) | |
731 | { | |
732 | int j, idx = reg_state[i].use_index; | |
733 | reload_combine_split_one_ruid (®_state[i].use_ruid, split_ruid); | |
734 | reload_combine_split_one_ruid (®_state[i].store_ruid, split_ruid); | |
735 | reload_combine_split_one_ruid (®_state[i].real_store_ruid, | |
736 | split_ruid); | |
737 | if (idx < 0) | |
738 | continue; | |
739 | for (j = idx; j < RELOAD_COMBINE_MAX_USES; j++) | |
740 | { | |
741 | reload_combine_split_one_ruid (®_state[i].reg_use[j].ruid, | |
742 | split_ruid); | |
743 | } | |
744 | } | |
745 | } | |
746 | ||
747 | /* Called when we are about to rescan a previously encountered insn with | |
748 | reload_combine_note_use after modifying some part of it. This clears all | |
749 | information about uses in that particular insn. */ | |
750 | ||
751 | static void | |
3aeaa53f | 752 | reload_combine_purge_insn_uses (rtx_insn *insn) |
d83ccc81 | 753 | { |
754 | unsigned i; | |
755 | ||
756 | for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) | |
757 | { | |
758 | int j, k, idx = reg_state[i].use_index; | |
759 | if (idx < 0) | |
760 | continue; | |
761 | j = k = RELOAD_COMBINE_MAX_USES; | |
762 | while (j-- > idx) | |
763 | { | |
764 | if (reg_state[i].reg_use[j].insn != insn) | |
765 | { | |
766 | k--; | |
767 | if (k != j) | |
768 | reg_state[i].reg_use[k] = reg_state[i].reg_use[j]; | |
769 | } | |
770 | } | |
771 | reg_state[i].use_index = k; | |
772 | } | |
773 | } | |
774 | ||
775 | /* Called when we need to forget about all uses of REGNO after an insn | |
776 | which is identified by RUID. */ | |
777 | ||
778 | static void | |
779 | reload_combine_purge_reg_uses_after_ruid (unsigned regno, int ruid) | |
780 | { | |
781 | int j, k, idx = reg_state[regno].use_index; | |
782 | if (idx < 0) | |
783 | return; | |
784 | j = k = RELOAD_COMBINE_MAX_USES; | |
785 | while (j-- > idx) | |
786 | { | |
787 | if (reg_state[regno].reg_use[j].ruid >= ruid) | |
788 | { | |
789 | k--; | |
790 | if (k != j) | |
791 | reg_state[regno].reg_use[k] = reg_state[regno].reg_use[j]; | |
792 | } | |
793 | } | |
794 | reg_state[regno].use_index = k; | |
795 | } | |
796 | ||
797 | /* Find the use of REGNO with the ruid that is highest among those | |
798 | lower than RUID_LIMIT, and return it if it is the only use of this | |
727047d0 | 799 | reg in the insn. Return NULL otherwise. */ |
d83ccc81 | 800 | |
801 | static struct reg_use * | |
802 | reload_combine_closest_single_use (unsigned regno, int ruid_limit) | |
803 | { | |
804 | int i, best_ruid = 0; | |
805 | int use_idx = reg_state[regno].use_index; | |
806 | struct reg_use *retval; | |
807 | ||
808 | if (use_idx < 0) | |
809 | return NULL; | |
810 | retval = NULL; | |
811 | for (i = use_idx; i < RELOAD_COMBINE_MAX_USES; i++) | |
812 | { | |
0ead6a7d | 813 | struct reg_use *use = reg_state[regno].reg_use + i; |
814 | int this_ruid = use->ruid; | |
d83ccc81 | 815 | if (this_ruid >= ruid_limit) |
816 | continue; | |
817 | if (this_ruid > best_ruid) | |
818 | { | |
819 | best_ruid = this_ruid; | |
727047d0 | 820 | retval = use; |
d83ccc81 | 821 | } |
727047d0 | 822 | else if (this_ruid == best_ruid) |
d83ccc81 | 823 | retval = NULL; |
824 | } | |
825 | if (last_label_ruid >= best_ruid) | |
826 | return NULL; | |
827 | return retval; | |
828 | } | |
829 | ||
65069495 | 830 | /* After we've moved an add insn, fix up any debug insns that occur |
831 | between the old location of the add and the new location. REG is | |
832 | the destination register of the add insn; REPLACEMENT is the | |
833 | SET_SRC of the add. FROM and TO specify the range in which we | |
834 | should make this change on debug insns. */ | |
0ead6a7d | 835 | |
836 | static void | |
3aeaa53f | 837 | fixup_debug_insns (rtx reg, rtx replacement, rtx_insn *from, rtx_insn *to) |
0ead6a7d | 838 | { |
3aeaa53f | 839 | rtx_insn *insn; |
65069495 | 840 | for (insn = from; insn != to; insn = NEXT_INSN (insn)) |
0ead6a7d | 841 | { |
842 | rtx t; | |
65069495 | 843 | |
bce107d7 | 844 | if (!DEBUG_BIND_INSN_P (insn)) |
0ead6a7d | 845 | continue; |
65069495 | 846 | |
847 | t = INSN_VAR_LOCATION_LOC (insn); | |
727047d0 | 848 | t = simplify_replace_rtx (t, reg, replacement); |
65069495 | 849 | validate_change (insn, &INSN_VAR_LOCATION_LOC (insn), t, 0); |
0ead6a7d | 850 | } |
851 | } | |
852 | ||
692ec7c8 | 853 | /* Subroutine of reload_combine_recognize_const_pattern. Try to replace REG |
854 | with SRC in the insn described by USE, taking costs into account. Return | |
855 | true if we made the replacement. */ | |
856 | ||
857 | static bool | |
858 | try_replace_in_use (struct reg_use *use, rtx reg, rtx src) | |
859 | { | |
3aeaa53f | 860 | rtx_insn *use_insn = use->insn; |
692ec7c8 | 861 | rtx mem = use->containing_mem; |
862 | bool speed = optimize_bb_for_speed_p (BLOCK_FOR_INSN (use_insn)); | |
863 | ||
864 | if (mem != NULL_RTX) | |
865 | { | |
866 | addr_space_t as = MEM_ADDR_SPACE (mem); | |
867 | rtx oldaddr = XEXP (mem, 0); | |
868 | rtx newaddr = NULL_RTX; | |
869 | int old_cost = address_cost (oldaddr, GET_MODE (mem), as, speed); | |
870 | int new_cost; | |
871 | ||
872 | newaddr = simplify_replace_rtx (oldaddr, reg, src); | |
873 | if (memory_address_addr_space_p (GET_MODE (mem), newaddr, as)) | |
874 | { | |
875 | XEXP (mem, 0) = newaddr; | |
876 | new_cost = address_cost (newaddr, GET_MODE (mem), as, speed); | |
877 | XEXP (mem, 0) = oldaddr; | |
878 | if (new_cost <= old_cost | |
879 | && validate_change (use_insn, | |
880 | &XEXP (mem, 0), newaddr, 0)) | |
881 | return true; | |
882 | } | |
883 | } | |
884 | else | |
885 | { | |
886 | rtx new_set = single_set (use_insn); | |
887 | if (new_set | |
888 | && REG_P (SET_DEST (new_set)) | |
889 | && GET_CODE (SET_SRC (new_set)) == PLUS | |
890 | && REG_P (XEXP (SET_SRC (new_set), 0)) | |
891 | && CONSTANT_P (XEXP (SET_SRC (new_set), 1))) | |
892 | { | |
893 | rtx new_src; | |
5ae4887d | 894 | machine_mode mode = GET_MODE (SET_DEST (new_set)); |
895 | int old_cost = set_src_cost (SET_SRC (new_set), mode, speed); | |
692ec7c8 | 896 | |
897 | gcc_assert (rtx_equal_p (XEXP (SET_SRC (new_set), 0), reg)); | |
898 | new_src = simplify_replace_rtx (SET_SRC (new_set), reg, src); | |
899 | ||
5ae4887d | 900 | if (set_src_cost (new_src, mode, speed) <= old_cost |
692ec7c8 | 901 | && validate_change (use_insn, &SET_SRC (new_set), |
902 | new_src, 0)) | |
903 | return true; | |
904 | } | |
905 | } | |
906 | return false; | |
907 | } | |
908 | ||
d83ccc81 | 909 | /* Called by reload_combine when scanning INSN. This function tries to detect |
910 | patterns where a constant is added to a register, and the result is used | |
911 | in an address. | |
912 | Return true if no further processing is needed on INSN; false if it wasn't | |
913 | recognized and should be handled normally. */ | |
914 | ||
915 | static bool | |
3aeaa53f | 916 | reload_combine_recognize_const_pattern (rtx_insn *insn) |
d83ccc81 | 917 | { |
918 | int from_ruid = reload_combine_ruid; | |
919 | rtx set, pat, reg, src, addreg; | |
920 | unsigned int regno; | |
921 | struct reg_use *use; | |
922 | bool must_move_add; | |
3aeaa53f | 923 | rtx_insn *add_moved_after_insn = NULL; |
d83ccc81 | 924 | int add_moved_after_ruid = 0; |
925 | int clobbered_regno = -1; | |
926 | ||
927 | set = single_set (insn); | |
928 | if (set == NULL_RTX) | |
929 | return false; | |
930 | ||
931 | reg = SET_DEST (set); | |
932 | src = SET_SRC (set); | |
933 | if (!REG_P (reg) | |
0933f1d9 | 934 | || REG_NREGS (reg) != 1 |
d83ccc81 | 935 | || GET_MODE (reg) != Pmode |
936 | || reg == stack_pointer_rtx) | |
937 | return false; | |
938 | ||
939 | regno = REGNO (reg); | |
940 | ||
941 | /* We look for a REG1 = REG2 + CONSTANT insn, followed by either | |
942 | uses of REG1 inside an address, or inside another add insn. If | |
943 | possible and profitable, merge the addition into subsequent | |
944 | uses. */ | |
945 | if (GET_CODE (src) != PLUS | |
946 | || !REG_P (XEXP (src, 0)) | |
947 | || !CONSTANT_P (XEXP (src, 1))) | |
948 | return false; | |
949 | ||
950 | addreg = XEXP (src, 0); | |
951 | must_move_add = rtx_equal_p (reg, addreg); | |
952 | ||
953 | pat = PATTERN (insn); | |
954 | if (must_move_add && set != pat) | |
955 | { | |
956 | /* We have to be careful when moving the add; apart from the | |
957 | single_set there may also be clobbers. Recognize one special | |
958 | case, that of one clobber alongside the set (likely a clobber | |
959 | of the CC register). */ | |
960 | gcc_assert (GET_CODE (PATTERN (insn)) == PARALLEL); | |
961 | if (XVECLEN (pat, 0) != 2 || XVECEXP (pat, 0, 0) != set | |
962 | || GET_CODE (XVECEXP (pat, 0, 1)) != CLOBBER | |
963 | || !REG_P (XEXP (XVECEXP (pat, 0, 1), 0))) | |
964 | return false; | |
965 | clobbered_regno = REGNO (XEXP (XVECEXP (pat, 0, 1), 0)); | |
966 | } | |
967 | ||
968 | do | |
969 | { | |
970 | use = reload_combine_closest_single_use (regno, from_ruid); | |
971 | ||
972 | if (use) | |
973 | /* Start the search for the next use from here. */ | |
974 | from_ruid = use->ruid; | |
975 | ||
976 | if (use && GET_MODE (*use->usep) == Pmode) | |
977 | { | |
692ec7c8 | 978 | bool delete_add = false; |
3aeaa53f | 979 | rtx_insn *use_insn = use->insn; |
d83ccc81 | 980 | int use_ruid = use->ruid; |
d83ccc81 | 981 | |
982 | /* Avoid moving the add insn past a jump. */ | |
0ead6a7d | 983 | if (must_move_add && use_ruid <= last_jump_ruid) |
d83ccc81 | 984 | break; |
985 | ||
986 | /* If the add clobbers another hard reg in parallel, don't move | |
987 | it past a real set of this hard reg. */ | |
988 | if (must_move_add && clobbered_regno >= 0 | |
989 | && reg_state[clobbered_regno].real_store_ruid >= use_ruid) | |
990 | break; | |
991 | ||
33b7314b | 992 | /* Do not separate cc0 setter and cc0 user on HAVE_cc0 targets. */ |
ff900b8e | 993 | if (HAVE_cc0 && must_move_add && sets_cc0_p (PATTERN (use_insn))) |
33b7314b | 994 | break; |
33b7314b | 995 | |
6aba0ea1 | 996 | gcc_assert (reg_state[regno].store_ruid <= use_ruid); |
997 | /* Avoid moving a use of ADDREG past a point where it is stored. */ | |
692ec7c8 | 998 | if (reg_state[REGNO (addreg)].store_ruid > use_ruid) |
d83ccc81 | 999 | break; |
1000 | ||
692ec7c8 | 1001 | /* We also must not move the addition past an insn that sets |
1002 | the same register, unless we can combine two add insns. */ | |
1003 | if (must_move_add && reg_state[regno].store_ruid == use_ruid) | |
d83ccc81 | 1004 | { |
692ec7c8 | 1005 | if (use->containing_mem == NULL_RTX) |
1006 | delete_add = true; | |
1007 | else | |
1008 | break; | |
d83ccc81 | 1009 | } |
d83ccc81 | 1010 | |
692ec7c8 | 1011 | if (try_replace_in_use (use, reg, src)) |
1012 | { | |
1013 | reload_combine_purge_insn_uses (use_insn); | |
1014 | reload_combine_note_use (&PATTERN (use_insn), use_insn, | |
1015 | use_ruid, NULL_RTX); | |
d83ccc81 | 1016 | |
692ec7c8 | 1017 | if (delete_add) |
1018 | { | |
1019 | fixup_debug_insns (reg, src, insn, use_insn); | |
1020 | delete_insn (insn); | |
1021 | return true; | |
1022 | } | |
1023 | if (must_move_add) | |
1024 | { | |
1025 | add_moved_after_insn = use_insn; | |
1026 | add_moved_after_ruid = use_ruid; | |
d83ccc81 | 1027 | } |
692ec7c8 | 1028 | continue; |
d83ccc81 | 1029 | } |
d83ccc81 | 1030 | } |
6aba0ea1 | 1031 | /* If we get here, we couldn't handle this use. */ |
1032 | if (must_move_add) | |
1033 | break; | |
d83ccc81 | 1034 | } |
1035 | while (use); | |
1036 | ||
1037 | if (!must_move_add || add_moved_after_insn == NULL_RTX) | |
1038 | /* Process the add normally. */ | |
1039 | return false; | |
1040 | ||
65069495 | 1041 | fixup_debug_insns (reg, src, insn, add_moved_after_insn); |
1042 | ||
d83ccc81 | 1043 | reorder_insns (insn, insn, add_moved_after_insn); |
1044 | reload_combine_purge_reg_uses_after_ruid (regno, add_moved_after_ruid); | |
1045 | reload_combine_split_ruids (add_moved_after_ruid - 1); | |
1046 | reload_combine_note_use (&PATTERN (insn), insn, | |
1047 | add_moved_after_ruid, NULL_RTX); | |
1048 | reg_state[regno].store_ruid = add_moved_after_ruid; | |
1049 | ||
1050 | return true; | |
1051 | } | |
1052 | ||
fb79f695 | 1053 | /* Called by reload_combine when scanning INSN. Try to detect a pattern we |
1054 | can handle and improve. Return true if no further processing is needed on | |
1055 | INSN; false if it wasn't recognized and should be handled normally. */ | |
1056 | ||
1057 | static bool | |
3aeaa53f | 1058 | reload_combine_recognize_pattern (rtx_insn *insn) |
fb79f695 | 1059 | { |
1060 | rtx set, reg, src; | |
fb79f695 | 1061 | |
d83ccc81 | 1062 | set = single_set (insn); |
1063 | if (set == NULL_RTX) | |
1064 | return false; | |
1065 | ||
1066 | reg = SET_DEST (set); | |
1067 | src = SET_SRC (set); | |
0933f1d9 | 1068 | if (!REG_P (reg) || REG_NREGS (reg) != 1) |
d83ccc81 | 1069 | return false; |
1070 | ||
d8ec06ae | 1071 | unsigned int regno = REGNO (reg); |
1072 | machine_mode mode = GET_MODE (reg); | |
1073 | ||
1074 | if (reg_state[regno].use_index < 0 | |
1075 | || reg_state[regno].use_index >= RELOAD_COMBINE_MAX_USES) | |
1076 | return false; | |
1077 | ||
1078 | for (int i = reg_state[regno].use_index; | |
1079 | i < RELOAD_COMBINE_MAX_USES; i++) | |
1080 | { | |
1081 | struct reg_use *use = reg_state[regno].reg_use + i; | |
1082 | if (GET_MODE (*use->usep) != mode) | |
1083 | return false; | |
1084 | } | |
d83ccc81 | 1085 | |
fb79f695 | 1086 | /* Look for (set (REGX) (CONST_INT)) |
1087 | (set (REGX) (PLUS (REGX) (REGY))) | |
1088 | ... | |
1089 | ... (MEM (REGX)) ... | |
1090 | and convert it to | |
1091 | (set (REGZ) (CONST_INT)) | |
1092 | ... | |
1093 | ... (MEM (PLUS (REGZ) (REGY)))... . | |
1094 | ||
1095 | First, check that we have (set (REGX) (PLUS (REGX) (REGY))) | |
1096 | and that we know all uses of REGX before it dies. | |
1097 | Also, explicitly check that REGX != REGY; our life information | |
1098 | does not yet show whether REGY changes in this insn. */ | |
fb79f695 | 1099 | |
1100 | if (GET_CODE (src) == PLUS | |
d83ccc81 | 1101 | && reg_state[regno].all_offsets_match |
1102 | && last_index_reg != -1 | |
fb79f695 | 1103 | && REG_P (XEXP (src, 1)) |
1104 | && rtx_equal_p (XEXP (src, 0), reg) | |
1105 | && !rtx_equal_p (XEXP (src, 1), reg) | |
1106 | && last_label_ruid < reg_state[regno].use_ruid) | |
1107 | { | |
1108 | rtx base = XEXP (src, 1); | |
3aeaa53f | 1109 | rtx_insn *prev = prev_nonnote_nondebug_insn (insn); |
fb79f695 | 1110 | rtx prev_set = prev ? single_set (prev) : NULL_RTX; |
1111 | rtx index_reg = NULL_RTX; | |
1112 | rtx reg_sum = NULL_RTX; | |
1113 | int i; | |
1114 | ||
1115 | /* Now we need to set INDEX_REG to an index register (denoted as | |
1116 | REGZ in the illustration above) and REG_SUM to the expression | |
1117 | register+register that we want to use to substitute uses of REG | |
1118 | (typically in MEMs) with. First check REG and BASE for being | |
1119 | index registers; we can use them even if they are not dead. */ | |
1120 | if (TEST_HARD_REG_BIT (reg_class_contents[INDEX_REG_CLASS], regno) | |
1121 | || TEST_HARD_REG_BIT (reg_class_contents[INDEX_REG_CLASS], | |
1122 | REGNO (base))) | |
1123 | { | |
1124 | index_reg = reg; | |
1125 | reg_sum = src; | |
1126 | } | |
1127 | else | |
1128 | { | |
1129 | /* Otherwise, look for a free index register. Since we have | |
1130 | checked above that neither REG nor BASE are index registers, | |
1131 | if we find anything at all, it will be different from these | |
1132 | two registers. */ | |
1133 | for (i = first_index_reg; i <= last_index_reg; i++) | |
1134 | { | |
1135 | if (TEST_HARD_REG_BIT (reg_class_contents[INDEX_REG_CLASS], i) | |
1136 | && reg_state[i].use_index == RELOAD_COMBINE_MAX_USES | |
1137 | && reg_state[i].store_ruid <= reg_state[regno].use_ruid | |
727047d0 | 1138 | && (call_used_regs[i] || df_regs_ever_live_p (i)) |
1139 | && (!frame_pointer_needed || i != HARD_FRAME_POINTER_REGNUM) | |
1140 | && !fixed_regs[i] && !global_regs[i] | |
92d2aec3 | 1141 | && hard_regno_nregs (i, GET_MODE (reg)) == 1 |
727047d0 | 1142 | && targetm.hard_regno_scratch_ok (i)) |
fb79f695 | 1143 | { |
1144 | index_reg = gen_rtx_REG (GET_MODE (reg), i); | |
1145 | reg_sum = gen_rtx_PLUS (GET_MODE (reg), index_reg, base); | |
1146 | break; | |
1147 | } | |
1148 | } | |
1149 | } | |
1150 | ||
1151 | /* Check that PREV_SET is indeed (set (REGX) (CONST_INT)) and that | |
1152 | (REGY), i.e. BASE, is not clobbered before the last use we'll | |
1153 | create. */ | |
1154 | if (reg_sum | |
1155 | && prev_set | |
1156 | && CONST_INT_P (SET_SRC (prev_set)) | |
1157 | && rtx_equal_p (SET_DEST (prev_set), reg) | |
fb79f695 | 1158 | && (reg_state[REGNO (base)].store_ruid |
1159 | <= reg_state[regno].use_ruid)) | |
1160 | { | |
1161 | /* Change destination register and, if necessary, the constant | |
1162 | value in PREV, the constant loading instruction. */ | |
1163 | validate_change (prev, &SET_DEST (prev_set), index_reg, 1); | |
1164 | if (reg_state[regno].offset != const0_rtx) | |
31c1512f | 1165 | { |
1166 | HOST_WIDE_INT c | |
1167 | = trunc_int_for_mode (UINTVAL (SET_SRC (prev_set)) | |
1168 | + UINTVAL (reg_state[regno].offset), | |
1169 | GET_MODE (index_reg)); | |
1170 | validate_change (prev, &SET_SRC (prev_set), GEN_INT (c), 1); | |
1171 | } | |
fb79f695 | 1172 | |
1173 | /* Now for every use of REG that we have recorded, replace REG | |
1174 | with REG_SUM. */ | |
1175 | for (i = reg_state[regno].use_index; | |
1176 | i < RELOAD_COMBINE_MAX_USES; i++) | |
1177 | validate_unshare_change (reg_state[regno].reg_use[i].insn, | |
1178 | reg_state[regno].reg_use[i].usep, | |
1179 | /* Each change must have its own | |
1180 | replacement. */ | |
1181 | reg_sum, 1); | |
1182 | ||
1183 | if (apply_change_group ()) | |
1184 | { | |
65069495 | 1185 | struct reg_use *lowest_ruid = NULL; |
1186 | ||
fb79f695 | 1187 | /* For every new use of REG_SUM, we have to record the use |
1188 | of BASE therein, i.e. operand 1. */ | |
1189 | for (i = reg_state[regno].use_index; | |
1190 | i < RELOAD_COMBINE_MAX_USES; i++) | |
65069495 | 1191 | { |
1192 | struct reg_use *use = reg_state[regno].reg_use + i; | |
1193 | reload_combine_note_use (&XEXP (*use->usep, 1), use->insn, | |
1194 | use->ruid, use->containing_mem); | |
1195 | if (lowest_ruid == NULL || use->ruid < lowest_ruid->ruid) | |
1196 | lowest_ruid = use; | |
1197 | } | |
1198 | ||
1199 | fixup_debug_insns (reg, reg_sum, insn, lowest_ruid->insn); | |
fb79f695 | 1200 | |
fb79f695 | 1201 | /* Delete the reg-reg addition. */ |
1202 | delete_insn (insn); | |
1203 | ||
d20ae451 | 1204 | if (reg_state[regno].offset != const0_rtx |
1205 | /* Previous REG_EQUIV / REG_EQUAL notes for PREV | |
1206 | are now invalid. */ | |
1207 | && remove_reg_equal_equiv_notes (prev)) | |
1208 | df_notes_rescan (prev); | |
fb79f695 | 1209 | |
1210 | reg_state[regno].use_index = RELOAD_COMBINE_MAX_USES; | |
fb79f695 | 1211 | return true; |
1212 | } | |
1213 | } | |
1214 | } | |
1215 | return false; | |
1216 | } | |
1217 | ||
8f8cadbc | 1218 | static void |
3ad4992f | 1219 | reload_combine (void) |
8f8cadbc | 1220 | { |
3aeaa53f | 1221 | rtx_insn *insn, *prev; |
8f8cadbc | 1222 | basic_block bb; |
1223 | unsigned int r; | |
8f8cadbc | 1224 | int min_labelno, n_labels; |
1225 | HARD_REG_SET ever_live_at_start, *label_live; | |
1226 | ||
8f8cadbc | 1227 | /* To avoid wasting too much time later searching for an index register, |
1228 | determine the minimum and maximum index register numbers. */ | |
fb79f695 | 1229 | if (INDEX_REG_CLASS == NO_REGS) |
1230 | last_index_reg = -1; | |
1231 | else if (first_index_reg == -1 && last_index_reg == 0) | |
1232 | { | |
1233 | for (r = 0; r < FIRST_PSEUDO_REGISTER; r++) | |
1234 | if (TEST_HARD_REG_BIT (reg_class_contents[INDEX_REG_CLASS], r)) | |
1235 | { | |
1236 | if (first_index_reg == -1) | |
1237 | first_index_reg = r; | |
1238 | ||
1239 | last_index_reg = r; | |
1240 | } | |
1241 | ||
1242 | /* If no index register is available, we can quit now. Set LAST_INDEX_REG | |
1243 | to -1 so we'll know to quit early the next time we get here. */ | |
1244 | if (first_index_reg == -1) | |
1245 | { | |
1246 | last_index_reg = -1; | |
1247 | return; | |
1248 | } | |
1249 | } | |
8f8cadbc | 1250 | |
8f8cadbc | 1251 | /* Set up LABEL_LIVE and EVER_LIVE_AT_START. The register lifetime |
1252 | information is a bit fuzzy immediately after reload, but it's | |
1253 | still good enough to determine which registers are live at a jump | |
1254 | destination. */ | |
1255 | min_labelno = get_first_label_num (); | |
1256 | n_labels = max_label_num () - min_labelno; | |
4c36ffe6 | 1257 | label_live = XNEWVEC (HARD_REG_SET, n_labels); |
8f8cadbc | 1258 | CLEAR_HARD_REG_SET (ever_live_at_start); |
1259 | ||
7a46197b | 1260 | FOR_EACH_BB_REVERSE_FN (bb, cfun) |
8f8cadbc | 1261 | { |
5496dbfc | 1262 | insn = BB_HEAD (bb); |
6d7dc5b9 | 1263 | if (LABEL_P (insn)) |
8f8cadbc | 1264 | { |
1265 | HARD_REG_SET live; | |
deb2741b | 1266 | bitmap live_in = df_get_live_in (bb); |
8f8cadbc | 1267 | |
deb2741b | 1268 | REG_SET_TO_HARD_REG_SET (live, live_in); |
1269 | compute_use_by_pseudos (&live, live_in); | |
8f8cadbc | 1270 | COPY_HARD_REG_SET (LABEL_LIVE (insn), live); |
1271 | IOR_HARD_REG_SET (ever_live_at_start, live); | |
1272 | } | |
1273 | } | |
1274 | ||
1275 | /* Initialize last_label_ruid, reload_combine_ruid and reg_state. */ | |
d83ccc81 | 1276 | last_label_ruid = last_jump_ruid = reload_combine_ruid = 0; |
8f8cadbc | 1277 | for (r = 0; r < FIRST_PSEUDO_REGISTER; r++) |
1278 | { | |
d83ccc81 | 1279 | reg_state[r].store_ruid = 0; |
1280 | reg_state[r].real_store_ruid = 0; | |
8f8cadbc | 1281 | if (fixed_regs[r]) |
1282 | reg_state[r].use_index = -1; | |
1283 | else | |
1284 | reg_state[r].use_index = RELOAD_COMBINE_MAX_USES; | |
1285 | } | |
1286 | ||
d83ccc81 | 1287 | for (insn = get_last_insn (); insn; insn = prev) |
8f8cadbc | 1288 | { |
8b52f64e | 1289 | bool control_flow_insn; |
8f8cadbc | 1290 | rtx note; |
1291 | ||
d83ccc81 | 1292 | prev = PREV_INSN (insn); |
1293 | ||
8f8cadbc | 1294 | /* We cannot do our optimization across labels. Invalidating all the use |
1295 | information we have would be costly, so we just note where the label | |
1296 | is and then later disable any optimization that would cross it. */ | |
6d7dc5b9 | 1297 | if (LABEL_P (insn)) |
8f8cadbc | 1298 | last_label_ruid = reload_combine_ruid; |
19f69355 | 1299 | else if (BARRIER_P (insn)) |
1300 | { | |
1301 | /* Crossing a barrier resets all the use information. */ | |
1302 | for (r = 0; r < FIRST_PSEUDO_REGISTER; r++) | |
1303 | if (! fixed_regs[r]) | |
8f8cadbc | 1304 | reg_state[r].use_index = RELOAD_COMBINE_MAX_USES; |
19f69355 | 1305 | } |
1306 | else if (INSN_P (insn) && volatile_insn_p (PATTERN (insn))) | |
1307 | /* Optimizations across insns being marked as volatile must be | |
1308 | prevented. All the usage information is invalidated | |
1309 | here. */ | |
1310 | for (r = 0; r < FIRST_PSEUDO_REGISTER; r++) | |
1311 | if (! fixed_regs[r] | |
1312 | && reg_state[r].use_index != RELOAD_COMBINE_MAX_USES) | |
1313 | reg_state[r].use_index = -1; | |
8f8cadbc | 1314 | |
65069495 | 1315 | if (! NONDEBUG_INSN_P (insn)) |
8f8cadbc | 1316 | continue; |
1317 | ||
1318 | reload_combine_ruid++; | |
1319 | ||
8b52f64e | 1320 | control_flow_insn = control_flow_insn_p (insn); |
1321 | if (control_flow_insn) | |
d83ccc81 | 1322 | last_jump_ruid = reload_combine_ruid; |
1323 | ||
1324 | if (reload_combine_recognize_const_pattern (insn) | |
1325 | || reload_combine_recognize_pattern (insn)) | |
fb79f695 | 1326 | continue; |
8f8cadbc | 1327 | |
1328 | note_stores (PATTERN (insn), reload_combine_note_store, NULL); | |
1329 | ||
6d7dc5b9 | 1330 | if (CALL_P (insn)) |
8f8cadbc | 1331 | { |
1332 | rtx link; | |
30326fda | 1333 | HARD_REG_SET used_regs; |
1334 | ||
1335 | get_call_reg_set_usage (insn, &used_regs, call_used_reg_set); | |
8f8cadbc | 1336 | |
1337 | for (r = 0; r < FIRST_PSEUDO_REGISTER; r++) | |
30326fda | 1338 | if (TEST_HARD_REG_BIT (used_regs, r)) |
8f8cadbc | 1339 | { |
1340 | reg_state[r].use_index = RELOAD_COMBINE_MAX_USES; | |
1341 | reg_state[r].store_ruid = reload_combine_ruid; | |
1342 | } | |
1343 | ||
1344 | for (link = CALL_INSN_FUNCTION_USAGE (insn); link; | |
1345 | link = XEXP (link, 1)) | |
1346 | { | |
c8010b80 | 1347 | rtx setuse = XEXP (link, 0); |
1348 | rtx usage_rtx = XEXP (setuse, 0); | |
70bdfe23 | 1349 | /* We could support CLOBBER_HIGH and treat it in the same way as |
1350 | HARD_REGNO_CALL_PART_CLOBBERED, but no port needs that yet. */ | |
1351 | gcc_assert (GET_CODE (setuse) != CLOBBER_HIGH); | |
1352 | ||
c8010b80 | 1353 | if ((GET_CODE (setuse) == USE || GET_CODE (setuse) == CLOBBER) |
1354 | && REG_P (usage_rtx)) | |
8f8cadbc | 1355 | { |
6a298741 | 1356 | unsigned int end_regno = END_REGNO (usage_rtx); |
1357 | for (unsigned int i = REGNO (usage_rtx); i < end_regno; ++i) | |
8f8cadbc | 1358 | if (GET_CODE (XEXP (link, 0)) == CLOBBER) |
1359 | { | |
1360 | reg_state[i].use_index = RELOAD_COMBINE_MAX_USES; | |
1361 | reg_state[i].store_ruid = reload_combine_ruid; | |
1362 | } | |
1363 | else | |
1364 | reg_state[i].use_index = -1; | |
1365 | } | |
1366 | } | |
8f8cadbc | 1367 | } |
f4979459 | 1368 | |
7777a939 | 1369 | if (control_flow_insn && !ANY_RETURN_P (PATTERN (insn))) |
8f8cadbc | 1370 | { |
1371 | /* Non-spill registers might be used at the call destination in | |
1372 | some unknown fashion, so we have to mark the unknown use. */ | |
1373 | HARD_REG_SET *live; | |
1374 | ||
1375 | if ((condjump_p (insn) || condjump_in_parallel_p (insn)) | |
1376 | && JUMP_LABEL (insn)) | |
7777a939 | 1377 | { |
1378 | if (ANY_RETURN_P (JUMP_LABEL (insn))) | |
1379 | live = NULL; | |
1380 | else | |
1381 | live = &LABEL_LIVE (JUMP_LABEL (insn)); | |
1382 | } | |
8f8cadbc | 1383 | else |
1384 | live = &ever_live_at_start; | |
1385 | ||
7777a939 | 1386 | if (live) |
1387 | for (r = 0; r < FIRST_PSEUDO_REGISTER; r++) | |
1388 | if (TEST_HARD_REG_BIT (*live, r)) | |
1389 | reg_state[r].use_index = -1; | |
8f8cadbc | 1390 | } |
1391 | ||
8b52f64e | 1392 | reload_combine_note_use (&PATTERN (insn), insn, reload_combine_ruid, |
1393 | NULL_RTX); | |
1394 | ||
8f8cadbc | 1395 | for (note = REG_NOTES (insn); note; note = XEXP (note, 1)) |
1396 | { | |
8b52f64e | 1397 | if (REG_NOTE_KIND (note) == REG_INC && REG_P (XEXP (note, 0))) |
8f8cadbc | 1398 | { |
1399 | int regno = REGNO (XEXP (note, 0)); | |
8f8cadbc | 1400 | reg_state[regno].store_ruid = reload_combine_ruid; |
d83ccc81 | 1401 | reg_state[regno].real_store_ruid = reload_combine_ruid; |
8f8cadbc | 1402 | reg_state[regno].use_index = -1; |
1403 | } | |
1404 | } | |
1405 | } | |
1406 | ||
1407 | free (label_live); | |
1408 | } | |
1409 | ||
1410 | /* Check if DST is a register or a subreg of a register; if it is, | |
d83ccc81 | 1411 | update store_ruid, real_store_ruid and use_index in the reg_state |
1412 | structure accordingly. Called via note_stores from reload_combine. */ | |
8f8cadbc | 1413 | |
1414 | static void | |
81a410b1 | 1415 | reload_combine_note_store (rtx dst, const_rtx set, void *data ATTRIBUTE_UNUSED) |
8f8cadbc | 1416 | { |
1417 | int regno = 0; | |
1418 | int i; | |
3754d046 | 1419 | machine_mode mode = GET_MODE (dst); |
8f8cadbc | 1420 | |
1421 | if (GET_CODE (dst) == SUBREG) | |
1422 | { | |
1423 | regno = subreg_regno_offset (REGNO (SUBREG_REG (dst)), | |
1424 | GET_MODE (SUBREG_REG (dst)), | |
1425 | SUBREG_BYTE (dst), | |
1426 | GET_MODE (dst)); | |
1427 | dst = SUBREG_REG (dst); | |
1428 | } | |
fe6524b0 | 1429 | |
1430 | /* Some targets do argument pushes without adding REG_INC notes. */ | |
1431 | ||
1432 | if (MEM_P (dst)) | |
1433 | { | |
1434 | dst = XEXP (dst, 0); | |
1435 | if (GET_CODE (dst) == PRE_INC || GET_CODE (dst) == POST_INC | |
a5dda0b9 | 1436 | || GET_CODE (dst) == PRE_DEC || GET_CODE (dst) == POST_DEC |
1437 | || GET_CODE (dst) == PRE_MODIFY || GET_CODE (dst) == POST_MODIFY) | |
fe6524b0 | 1438 | { |
6a298741 | 1439 | unsigned int end_regno = END_REGNO (XEXP (dst, 0)); |
1440 | for (unsigned int i = REGNO (XEXP (dst, 0)); i < end_regno; ++i) | |
fe6524b0 | 1441 | { |
1442 | /* We could probably do better, but for now mark the register | |
1443 | as used in an unknown fashion and set/clobbered at this | |
1444 | insn. */ | |
1445 | reg_state[i].use_index = -1; | |
1446 | reg_state[i].store_ruid = reload_combine_ruid; | |
1447 | reg_state[i].real_store_ruid = reload_combine_ruid; | |
1448 | } | |
1449 | } | |
1450 | else | |
1451 | return; | |
1452 | } | |
1453 | ||
8ad4c111 | 1454 | if (!REG_P (dst)) |
8f8cadbc | 1455 | return; |
1456 | regno += REGNO (dst); | |
1457 | ||
1458 | /* note_stores might have stripped a STRICT_LOW_PART, so we have to be | |
1459 | careful with registers / register parts that are not full words. | |
476d094d | 1460 | Similarly for ZERO_EXTRACT. */ |
d83ccc81 | 1461 | if (GET_CODE (SET_DEST (set)) == ZERO_EXTRACT |
8f8cadbc | 1462 | || GET_CODE (SET_DEST (set)) == STRICT_LOW_PART) |
1463 | { | |
16b9e38b | 1464 | for (i = end_hard_regno (mode, regno) - 1; i >= regno; i--) |
8f8cadbc | 1465 | { |
1466 | reg_state[i].use_index = -1; | |
1467 | reg_state[i].store_ruid = reload_combine_ruid; | |
d83ccc81 | 1468 | reg_state[i].real_store_ruid = reload_combine_ruid; |
8f8cadbc | 1469 | } |
1470 | } | |
1471 | else | |
1472 | { | |
16b9e38b | 1473 | for (i = end_hard_regno (mode, regno) - 1; i >= regno; i--) |
8f8cadbc | 1474 | { |
1475 | reg_state[i].store_ruid = reload_combine_ruid; | |
d83ccc81 | 1476 | if (GET_CODE (set) == SET) |
1477 | reg_state[i].real_store_ruid = reload_combine_ruid; | |
8f8cadbc | 1478 | reg_state[i].use_index = RELOAD_COMBINE_MAX_USES; |
1479 | } | |
1480 | } | |
1481 | } | |
1482 | ||
1483 | /* XP points to a piece of rtl that has to be checked for any uses of | |
1484 | registers. | |
1485 | *XP is the pattern of INSN, or a part of it. | |
1486 | Called from reload_combine, and recursively by itself. */ | |
1487 | static void | |
3aeaa53f | 1488 | reload_combine_note_use (rtx *xp, rtx_insn *insn, int ruid, rtx containing_mem) |
8f8cadbc | 1489 | { |
1490 | rtx x = *xp; | |
1491 | enum rtx_code code = x->code; | |
1492 | const char *fmt; | |
1493 | int i, j; | |
1494 | rtx offset = const0_rtx; /* For the REG case below. */ | |
1495 | ||
1496 | switch (code) | |
1497 | { | |
1498 | case SET: | |
8ad4c111 | 1499 | if (REG_P (SET_DEST (x))) |
8f8cadbc | 1500 | { |
d83ccc81 | 1501 | reload_combine_note_use (&SET_SRC (x), insn, ruid, NULL_RTX); |
8f8cadbc | 1502 | return; |
1503 | } | |
1504 | break; | |
1505 | ||
1506 | case USE: | |
1507 | /* If this is the USE of a return value, we can't change it. */ | |
8ad4c111 | 1508 | if (REG_P (XEXP (x, 0)) && REG_FUNCTION_VALUE_P (XEXP (x, 0))) |
8f8cadbc | 1509 | { |
6a298741 | 1510 | /* Mark the return register as used in an unknown fashion. */ |
8f8cadbc | 1511 | rtx reg = XEXP (x, 0); |
6a298741 | 1512 | unsigned int end_regno = END_REGNO (reg); |
1513 | for (unsigned int regno = REGNO (reg); regno < end_regno; ++regno) | |
1514 | reg_state[regno].use_index = -1; | |
8f8cadbc | 1515 | return; |
1516 | } | |
1517 | break; | |
1518 | ||
1519 | case CLOBBER: | |
8ad4c111 | 1520 | if (REG_P (SET_DEST (x))) |
8f8cadbc | 1521 | { |
1522 | /* No spurious CLOBBERs of pseudo registers may remain. */ | |
876760f6 | 1523 | gcc_assert (REGNO (SET_DEST (x)) < FIRST_PSEUDO_REGISTER); |
8f8cadbc | 1524 | return; |
1525 | } | |
1526 | break; | |
1527 | ||
70bdfe23 | 1528 | case CLOBBER_HIGH: |
1529 | gcc_assert (REG_P (SET_DEST (x))); | |
1530 | return; | |
1531 | ||
8f8cadbc | 1532 | case PLUS: |
1533 | /* We are interested in (plus (reg) (const_int)) . */ | |
8ad4c111 | 1534 | if (!REG_P (XEXP (x, 0)) |
971ba038 | 1535 | || !CONST_INT_P (XEXP (x, 1))) |
8f8cadbc | 1536 | break; |
1537 | offset = XEXP (x, 1); | |
1538 | x = XEXP (x, 0); | |
1539 | /* Fall through. */ | |
1540 | case REG: | |
1541 | { | |
1542 | int regno = REGNO (x); | |
1543 | int use_index; | |
1544 | int nregs; | |
1545 | ||
1546 | /* No spurious USEs of pseudo registers may remain. */ | |
876760f6 | 1547 | gcc_assert (regno < FIRST_PSEUDO_REGISTER); |
8f8cadbc | 1548 | |
0933f1d9 | 1549 | nregs = REG_NREGS (x); |
8f8cadbc | 1550 | |
1551 | /* We can't substitute into multi-hard-reg uses. */ | |
1552 | if (nregs > 1) | |
1553 | { | |
1554 | while (--nregs >= 0) | |
1555 | reg_state[regno + nregs].use_index = -1; | |
1556 | return; | |
1557 | } | |
1558 | ||
727047d0 | 1559 | /* We may be called to update uses in previously seen insns. |
1560 | Don't add uses beyond the last store we saw. */ | |
1561 | if (ruid < reg_state[regno].store_ruid) | |
1562 | return; | |
1563 | ||
8f8cadbc | 1564 | /* If this register is already used in some unknown fashion, we |
1565 | can't do anything. | |
1566 | If we decrement the index from zero to -1, we can't store more | |
1567 | uses, so this register becomes used in an unknown fashion. */ | |
1568 | use_index = --reg_state[regno].use_index; | |
1569 | if (use_index < 0) | |
1570 | return; | |
1571 | ||
d83ccc81 | 1572 | if (use_index == RELOAD_COMBINE_MAX_USES - 1) |
8f8cadbc | 1573 | { |
1574 | /* This is the first use of this register we have seen since we | |
1575 | marked it as dead. */ | |
1576 | reg_state[regno].offset = offset; | |
d83ccc81 | 1577 | reg_state[regno].all_offsets_match = true; |
1578 | reg_state[regno].use_ruid = ruid; | |
8f8cadbc | 1579 | } |
0ead6a7d | 1580 | else |
1581 | { | |
1582 | if (reg_state[regno].use_ruid > ruid) | |
1583 | reg_state[regno].use_ruid = ruid; | |
1584 | ||
1585 | if (! rtx_equal_p (offset, reg_state[regno].offset)) | |
1586 | reg_state[regno].all_offsets_match = false; | |
1587 | } | |
d83ccc81 | 1588 | |
8f8cadbc | 1589 | reg_state[regno].reg_use[use_index].insn = insn; |
d83ccc81 | 1590 | reg_state[regno].reg_use[use_index].ruid = ruid; |
1591 | reg_state[regno].reg_use[use_index].containing_mem = containing_mem; | |
8f8cadbc | 1592 | reg_state[regno].reg_use[use_index].usep = xp; |
1593 | return; | |
1594 | } | |
1595 | ||
d83ccc81 | 1596 | case MEM: |
1597 | containing_mem = x; | |
1598 | break; | |
1599 | ||
8f8cadbc | 1600 | default: |
1601 | break; | |
1602 | } | |
1603 | ||
1604 | /* Recursively process the components of X. */ | |
1605 | fmt = GET_RTX_FORMAT (code); | |
1606 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
1607 | { | |
1608 | if (fmt[i] == 'e') | |
d83ccc81 | 1609 | reload_combine_note_use (&XEXP (x, i), insn, ruid, containing_mem); |
8f8cadbc | 1610 | else if (fmt[i] == 'E') |
1611 | { | |
1612 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
d83ccc81 | 1613 | reload_combine_note_use (&XVECEXP (x, i, j), insn, ruid, |
1614 | containing_mem); | |
8f8cadbc | 1615 | } |
1616 | } | |
1617 | } | |
1618 | \f | |
1619 | /* See if we can reduce the cost of a constant by replacing a move | |
1620 | with an add. We track situations in which a register is set to a | |
1621 | constant or to a register plus a constant. */ | |
1622 | /* We cannot do our optimization across labels. Invalidating all the | |
1623 | information about register contents we have would be costly, so we | |
1624 | use move2add_last_label_luid to note where the label is and then | |
1625 | later disable any optimization that would cross it. | |
6132c0d0 | 1626 | reg_offset[n] / reg_base_reg[n] / reg_symbol_ref[n] / reg_mode[n] |
1627 | are only valid if reg_set_luid[n] is greater than | |
b6b86e87 | 1628 | move2add_last_label_luid. |
1629 | For a set that established a new (potential) base register with | |
1630 | non-constant value, we use move2add_luid from the place where the | |
1631 | setting insn is encountered; registers based off that base then | |
1632 | get the same reg_set_luid. Constants all get | |
1633 | move2add_last_label_luid + 1 as their reg_set_luid. */ | |
8f8cadbc | 1634 | static int reg_set_luid[FIRST_PSEUDO_REGISTER]; |
1635 | ||
1636 | /* If reg_base_reg[n] is negative, register n has been set to | |
6132c0d0 | 1637 | reg_offset[n] or reg_symbol_ref[n] + reg_offset[n] in mode reg_mode[n]. |
8f8cadbc | 1638 | If reg_base_reg[n] is non-negative, register n has been set to the |
1639 | sum of reg_offset[n] and the value of register reg_base_reg[n] | |
b6b86e87 | 1640 | before reg_set_luid[n], calculated in mode reg_mode[n] . |
1641 | For multi-hard-register registers, all but the first one are | |
1642 | recorded as BLKmode in reg_mode. Setting reg_mode to VOIDmode | |
1643 | marks it as invalid. */ | |
8f8cadbc | 1644 | static HOST_WIDE_INT reg_offset[FIRST_PSEUDO_REGISTER]; |
1645 | static int reg_base_reg[FIRST_PSEUDO_REGISTER]; | |
6132c0d0 | 1646 | static rtx reg_symbol_ref[FIRST_PSEUDO_REGISTER]; |
3754d046 | 1647 | static machine_mode reg_mode[FIRST_PSEUDO_REGISTER]; |
8f8cadbc | 1648 | |
1649 | /* move2add_luid is linearly increased while scanning the instructions | |
1650 | from first to last. It is used to set reg_set_luid in | |
1651 | reload_cse_move2add and move2add_note_store. */ | |
1652 | static int move2add_luid; | |
1653 | ||
1654 | /* move2add_last_label_luid is set whenever a label is found. Labels | |
1655 | invalidate all previously collected reg_offset data. */ | |
1656 | static int move2add_last_label_luid; | |
1657 | ||
1658 | /* ??? We don't know how zero / sign extension is handled, hence we | |
1659 | can't go from a narrower to a wider mode. */ | |
1660 | #define MODES_OK_FOR_MOVE2ADD(OUTMODE, INMODE) \ | |
1661 | (GET_MODE_SIZE (OUTMODE) == GET_MODE_SIZE (INMODE) \ | |
1662 | || (GET_MODE_SIZE (OUTMODE) <= GET_MODE_SIZE (INMODE) \ | |
396f2130 | 1663 | && TRULY_NOOP_TRUNCATION_MODES_P (OUTMODE, INMODE))) |
8f8cadbc | 1664 | |
b6b86e87 | 1665 | /* Record that REG is being set to a value with the mode of REG. */ |
1666 | ||
1667 | static void | |
1668 | move2add_record_mode (rtx reg) | |
1669 | { | |
1670 | int regno, nregs; | |
3754d046 | 1671 | machine_mode mode = GET_MODE (reg); |
b6b86e87 | 1672 | |
1673 | if (GET_CODE (reg) == SUBREG) | |
1674 | { | |
1675 | regno = subreg_regno (reg); | |
1676 | nregs = subreg_nregs (reg); | |
1677 | } | |
1678 | else if (REG_P (reg)) | |
1679 | { | |
1680 | regno = REGNO (reg); | |
0933f1d9 | 1681 | nregs = REG_NREGS (reg); |
b6b86e87 | 1682 | } |
1683 | else | |
1684 | gcc_unreachable (); | |
1685 | for (int i = nregs - 1; i > 0; i--) | |
1686 | reg_mode[regno + i] = BLKmode; | |
1687 | reg_mode[regno] = mode; | |
1688 | } | |
1689 | ||
1690 | /* Record that REG is being set to the sum of SYM and OFF. */ | |
1691 | ||
1692 | static void | |
1693 | move2add_record_sym_value (rtx reg, rtx sym, rtx off) | |
1694 | { | |
1695 | int regno = REGNO (reg); | |
1696 | ||
1697 | move2add_record_mode (reg); | |
1698 | reg_set_luid[regno] = move2add_luid; | |
1699 | reg_base_reg[regno] = -1; | |
1700 | reg_symbol_ref[regno] = sym; | |
1701 | reg_offset[regno] = INTVAL (off); | |
1702 | } | |
1703 | ||
1704 | /* Check if REGNO contains a valid value in MODE. */ | |
1705 | ||
1706 | static bool | |
18426c5b | 1707 | move2add_valid_value_p (int regno, scalar_int_mode mode) |
b6b86e87 | 1708 | { |
5bea3269 | 1709 | if (reg_set_luid[regno] <= move2add_last_label_luid) |
b6b86e87 | 1710 | return false; |
1711 | ||
5bea3269 | 1712 | if (mode != reg_mode[regno]) |
1713 | { | |
4c53345c | 1714 | scalar_int_mode old_mode; |
1715 | if (!is_a <scalar_int_mode> (reg_mode[regno], &old_mode) | |
1716 | || !MODES_OK_FOR_MOVE2ADD (mode, old_mode)) | |
5bea3269 | 1717 | return false; |
1718 | /* The value loaded into regno in reg_mode[regno] is also valid in | |
1719 | mode after truncation only if (REG:mode regno) is the lowpart of | |
1720 | (REG:reg_mode[regno] regno). Now, for big endian, the starting | |
1721 | regno of the lowpart might be different. */ | |
9edf7ea8 | 1722 | poly_int64 s_off = subreg_lowpart_offset (mode, old_mode); |
4c53345c | 1723 | s_off = subreg_regno_offset (regno, old_mode, s_off, mode); |
9edf7ea8 | 1724 | if (maybe_ne (s_off, 0)) |
5bea3269 | 1725 | /* We could in principle adjust regno, check reg_mode[regno] to be |
1726 | BLKmode, and return s_off to the caller (vs. -1 for failure), | |
1727 | but we currently have no callers that could make use of this | |
1728 | information. */ | |
1729 | return false; | |
1730 | } | |
1731 | ||
16b9e38b | 1732 | for (int i = end_hard_regno (mode, regno) - 1; i > regno; i--) |
1733 | if (reg_mode[i] != BLKmode) | |
b6b86e87 | 1734 | return false; |
1735 | return true; | |
1736 | } | |
1737 | ||
18426c5b | 1738 | /* This function is called with INSN that sets REG (of mode MODE) |
1739 | to (SYM + OFF), while REG is known to already have value (SYM + offset). | |
6132c0d0 | 1740 | This function tries to change INSN into an add instruction |
1741 | (set (REG) (plus (REG) (OFF - offset))) using the known value. | |
d83ccc81 | 1742 | It also updates the information about REG's known value. |
1743 | Return true if we made a change. */ | |
6132c0d0 | 1744 | |
d83ccc81 | 1745 | static bool |
18426c5b | 1746 | move2add_use_add2_insn (scalar_int_mode mode, rtx reg, rtx sym, rtx off, |
1747 | rtx_insn *insn) | |
6132c0d0 | 1748 | { |
1749 | rtx pat = PATTERN (insn); | |
1750 | rtx src = SET_SRC (pat); | |
1751 | int regno = REGNO (reg); | |
18426c5b | 1752 | rtx new_src = gen_int_mode (UINTVAL (off) - reg_offset[regno], mode); |
6132c0d0 | 1753 | bool speed = optimize_bb_for_speed_p (BLOCK_FOR_INSN (insn)); |
d83ccc81 | 1754 | bool changed = false; |
6132c0d0 | 1755 | |
1756 | /* (set (reg) (plus (reg) (const_int 0))) is not canonical; | |
1757 | use (set (reg) (reg)) instead. | |
1758 | We don't delete this insn, nor do we convert it into a | |
1759 | note, to avoid losing register notes or the return | |
1760 | value flag. jump2 already knows how to get rid of | |
1761 | no-op moves. */ | |
1762 | if (new_src == const0_rtx) | |
1763 | { | |
1764 | /* If the constants are different, this is a | |
1765 | truncation, that, if turned into (set (reg) | |
1766 | (reg)), would be discarded. Maybe we should | |
1767 | try a truncMN pattern? */ | |
1768 | if (INTVAL (off) == reg_offset [regno]) | |
d83ccc81 | 1769 | changed = validate_change (insn, &SET_SRC (pat), reg, 0); |
6132c0d0 | 1770 | } |
c9a03487 | 1771 | else |
6132c0d0 | 1772 | { |
c9a03487 | 1773 | struct full_rtx_costs oldcst, newcst; |
18426c5b | 1774 | rtx tem = gen_rtx_PLUS (mode, reg, new_src); |
c9a03487 | 1775 | |
b72d459f | 1776 | get_full_set_rtx_cost (pat, &oldcst); |
c9a03487 | 1777 | SET_SRC (pat) = tem; |
b72d459f | 1778 | get_full_set_rtx_cost (pat, &newcst); |
c9a03487 | 1779 | SET_SRC (pat) = src; |
1780 | ||
1781 | if (costs_lt_p (&newcst, &oldcst, speed) | |
1782 | && have_add2_insn (reg, new_src)) | |
1783 | changed = validate_change (insn, &SET_SRC (pat), tem, 0); | |
18426c5b | 1784 | else if (sym == NULL_RTX && mode != BImode) |
6132c0d0 | 1785 | { |
18426c5b | 1786 | scalar_int_mode narrow_mode; |
1787 | FOR_EACH_MODE_UNTIL (narrow_mode, mode) | |
6132c0d0 | 1788 | { |
c9a03487 | 1789 | if (have_insn_for (STRICT_LOW_PART, narrow_mode) |
1790 | && ((reg_offset[regno] & ~GET_MODE_MASK (narrow_mode)) | |
1791 | == (INTVAL (off) & ~GET_MODE_MASK (narrow_mode)))) | |
1792 | { | |
17ce39e3 | 1793 | rtx narrow_reg = gen_lowpart_common (narrow_mode, reg); |
c9a03487 | 1794 | rtx narrow_src = gen_int_mode (INTVAL (off), |
1795 | narrow_mode); | |
1796 | rtx new_set | |
d1f9b275 | 1797 | = gen_rtx_SET (gen_rtx_STRICT_LOW_PART (VOIDmode, |
c9a03487 | 1798 | narrow_reg), |
1799 | narrow_src); | |
d4177981 | 1800 | get_full_set_rtx_cost (new_set, &newcst); |
1801 | if (costs_lt_p (&newcst, &oldcst, speed)) | |
1802 | { | |
1803 | changed = validate_change (insn, &PATTERN (insn), | |
1804 | new_set, 0); | |
1805 | if (changed) | |
1806 | break; | |
1807 | } | |
c9a03487 | 1808 | } |
6132c0d0 | 1809 | } |
1810 | } | |
1811 | } | |
b6b86e87 | 1812 | move2add_record_sym_value (reg, sym, off); |
d83ccc81 | 1813 | return changed; |
6132c0d0 | 1814 | } |
1815 | ||
1816 | ||
18426c5b | 1817 | /* This function is called with INSN that sets REG (of mode MODE) to |
1818 | (SYM + OFF), but REG doesn't have known value (SYM + offset). This | |
1819 | function tries to find another register which is known to already have | |
6132c0d0 | 1820 | value (SYM + offset) and change INSN into an add instruction |
1821 | (set (REG) (plus (the found register) (OFF - offset))) if such | |
1822 | a register is found. It also updates the information about | |
d83ccc81 | 1823 | REG's known value. |
1824 | Return true iff we made a change. */ | |
6132c0d0 | 1825 | |
d83ccc81 | 1826 | static bool |
18426c5b | 1827 | move2add_use_add3_insn (scalar_int_mode mode, rtx reg, rtx sym, rtx off, |
1828 | rtx_insn *insn) | |
6132c0d0 | 1829 | { |
1830 | rtx pat = PATTERN (insn); | |
1831 | rtx src = SET_SRC (pat); | |
1832 | int regno = REGNO (reg); | |
c2130a4b | 1833 | int min_regno = 0; |
6132c0d0 | 1834 | bool speed = optimize_bb_for_speed_p (BLOCK_FOR_INSN (insn)); |
1835 | int i; | |
d83ccc81 | 1836 | bool changed = false; |
c9a03487 | 1837 | struct full_rtx_costs oldcst, newcst, mincst; |
1838 | rtx plus_expr; | |
1839 | ||
1840 | init_costs_to_max (&mincst); | |
b72d459f | 1841 | get_full_set_rtx_cost (pat, &oldcst); |
c9a03487 | 1842 | |
1843 | plus_expr = gen_rtx_PLUS (GET_MODE (reg), reg, const0_rtx); | |
1844 | SET_SRC (pat) = plus_expr; | |
6132c0d0 | 1845 | |
1846 | for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) | |
18426c5b | 1847 | if (move2add_valid_value_p (i, mode) |
6132c0d0 | 1848 | && reg_base_reg[i] < 0 |
1849 | && reg_symbol_ref[i] != NULL_RTX | |
1850 | && rtx_equal_p (sym, reg_symbol_ref[i])) | |
1851 | { | |
60141df0 | 1852 | rtx new_src = gen_int_mode (UINTVAL (off) - reg_offset[i], |
6132c0d0 | 1853 | GET_MODE (reg)); |
1854 | /* (set (reg) (plus (reg) (const_int 0))) is not canonical; | |
1855 | use (set (reg) (reg)) instead. | |
1856 | We don't delete this insn, nor do we convert it into a | |
1857 | note, to avoid losing register notes or the return | |
1858 | value flag. jump2 already knows how to get rid of | |
1859 | no-op moves. */ | |
1860 | if (new_src == const0_rtx) | |
1861 | { | |
c9a03487 | 1862 | init_costs_to_zero (&mincst); |
6132c0d0 | 1863 | min_regno = i; |
1864 | break; | |
1865 | } | |
1866 | else | |
1867 | { | |
c9a03487 | 1868 | XEXP (plus_expr, 1) = new_src; |
b72d459f | 1869 | get_full_set_rtx_cost (pat, &newcst); |
c9a03487 | 1870 | |
1871 | if (costs_lt_p (&newcst, &mincst, speed)) | |
6132c0d0 | 1872 | { |
c9a03487 | 1873 | mincst = newcst; |
6132c0d0 | 1874 | min_regno = i; |
1875 | } | |
1876 | } | |
1877 | } | |
c9a03487 | 1878 | SET_SRC (pat) = src; |
6132c0d0 | 1879 | |
c9a03487 | 1880 | if (costs_lt_p (&mincst, &oldcst, speed)) |
6132c0d0 | 1881 | { |
1882 | rtx tem; | |
1883 | ||
1884 | tem = gen_rtx_REG (GET_MODE (reg), min_regno); | |
1885 | if (i != min_regno) | |
1886 | { | |
60141df0 | 1887 | rtx new_src = gen_int_mode (UINTVAL (off) - reg_offset[min_regno], |
6132c0d0 | 1888 | GET_MODE (reg)); |
1889 | tem = gen_rtx_PLUS (GET_MODE (reg), tem, new_src); | |
1890 | } | |
d83ccc81 | 1891 | if (validate_change (insn, &SET_SRC (pat), tem, 0)) |
1892 | changed = true; | |
6132c0d0 | 1893 | } |
1894 | reg_set_luid[regno] = move2add_luid; | |
b6b86e87 | 1895 | move2add_record_sym_value (reg, sym, off); |
d83ccc81 | 1896 | return changed; |
6132c0d0 | 1897 | } |
1898 | ||
d83ccc81 | 1899 | /* Convert move insns with constant inputs to additions if they are cheaper. |
1900 | Return true if any changes were made. */ | |
1901 | static bool | |
3aeaa53f | 1902 | reload_cse_move2add (rtx_insn *first) |
8f8cadbc | 1903 | { |
1904 | int i; | |
3aeaa53f | 1905 | rtx_insn *insn; |
d83ccc81 | 1906 | bool changed = false; |
8f8cadbc | 1907 | |
1908 | for (i = FIRST_PSEUDO_REGISTER - 1; i >= 0; i--) | |
6132c0d0 | 1909 | { |
1910 | reg_set_luid[i] = 0; | |
1911 | reg_offset[i] = 0; | |
1912 | reg_base_reg[i] = 0; | |
1913 | reg_symbol_ref[i] = NULL_RTX; | |
1914 | reg_mode[i] = VOIDmode; | |
1915 | } | |
8f8cadbc | 1916 | |
1917 | move2add_last_label_luid = 0; | |
1918 | move2add_luid = 2; | |
1919 | for (insn = first; insn; insn = NEXT_INSN (insn), move2add_luid++) | |
1920 | { | |
1921 | rtx pat, note; | |
1922 | ||
6d7dc5b9 | 1923 | if (LABEL_P (insn)) |
8f8cadbc | 1924 | { |
1925 | move2add_last_label_luid = move2add_luid; | |
1926 | /* We're going to increment move2add_luid twice after a | |
1927 | label, so that we can use move2add_last_label_luid + 1 as | |
1928 | the luid for constants. */ | |
1929 | move2add_luid++; | |
1930 | continue; | |
1931 | } | |
1932 | if (! INSN_P (insn)) | |
1933 | continue; | |
1934 | pat = PATTERN (insn); | |
1935 | /* For simplicity, we only perform this optimization on | |
1936 | straightforward SETs. */ | |
18426c5b | 1937 | scalar_int_mode mode; |
8f8cadbc | 1938 | if (GET_CODE (pat) == SET |
18426c5b | 1939 | && REG_P (SET_DEST (pat)) |
1940 | && is_a <scalar_int_mode> (GET_MODE (SET_DEST (pat)), &mode)) | |
8f8cadbc | 1941 | { |
1942 | rtx reg = SET_DEST (pat); | |
1943 | int regno = REGNO (reg); | |
1944 | rtx src = SET_SRC (pat); | |
1945 | ||
1946 | /* Check if we have valid information on the contents of this | |
1947 | register in the mode of REG. */ | |
18426c5b | 1948 | if (move2add_valid_value_p (regno, mode) |
3072d30e | 1949 | && dbg_cnt (cse2_move2add)) |
8f8cadbc | 1950 | { |
1951 | /* Try to transform (set (REGX) (CONST_INT A)) | |
1952 | ... | |
1953 | (set (REGX) (CONST_INT B)) | |
1954 | to | |
1955 | (set (REGX) (CONST_INT A)) | |
1956 | ... | |
1957 | (set (REGX) (plus (REGX) (CONST_INT B-A))) | |
1958 | or | |
1959 | (set (REGX) (CONST_INT A)) | |
1960 | ... | |
1961 | (set (STRICT_LOW_PART (REGX)) (CONST_INT B)) | |
1962 | */ | |
1963 | ||
6132c0d0 | 1964 | if (CONST_INT_P (src) |
1965 | && reg_base_reg[regno] < 0 | |
1966 | && reg_symbol_ref[regno] == NULL_RTX) | |
8f8cadbc | 1967 | { |
18426c5b | 1968 | changed |= move2add_use_add2_insn (mode, reg, NULL_RTX, |
1969 | src, insn); | |
8f8cadbc | 1970 | continue; |
1971 | } | |
1972 | ||
1973 | /* Try to transform (set (REGX) (REGY)) | |
1974 | (set (REGX) (PLUS (REGX) (CONST_INT A))) | |
1975 | ... | |
1976 | (set (REGX) (REGY)) | |
1977 | (set (REGX) (PLUS (REGX) (CONST_INT B))) | |
1978 | to | |
1979 | (set (REGX) (REGY)) | |
1980 | (set (REGX) (PLUS (REGX) (CONST_INT A))) | |
1981 | ... | |
1982 | (set (REGX) (plus (REGX) (CONST_INT B-A))) */ | |
8ad4c111 | 1983 | else if (REG_P (src) |
8f8cadbc | 1984 | && reg_set_luid[regno] == reg_set_luid[REGNO (src)] |
1985 | && reg_base_reg[regno] == reg_base_reg[REGNO (src)] | |
18426c5b | 1986 | && move2add_valid_value_p (REGNO (src), mode)) |
8f8cadbc | 1987 | { |
3aeaa53f | 1988 | rtx_insn *next = next_nonnote_nondebug_insn (insn); |
8f8cadbc | 1989 | rtx set = NULL_RTX; |
1990 | if (next) | |
1991 | set = single_set (next); | |
1992 | if (set | |
1993 | && SET_DEST (set) == reg | |
1994 | && GET_CODE (SET_SRC (set)) == PLUS | |
1995 | && XEXP (SET_SRC (set), 0) == reg | |
971ba038 | 1996 | && CONST_INT_P (XEXP (SET_SRC (set), 1))) |
8f8cadbc | 1997 | { |
1998 | rtx src3 = XEXP (SET_SRC (set), 1); | |
60141df0 | 1999 | unsigned HOST_WIDE_INT added_offset = UINTVAL (src3); |
8f8cadbc | 2000 | HOST_WIDE_INT base_offset = reg_offset[REGNO (src)]; |
2001 | HOST_WIDE_INT regno_offset = reg_offset[regno]; | |
2002 | rtx new_src = | |
69e41517 | 2003 | gen_int_mode (added_offset |
2004 | + base_offset | |
2005 | - regno_offset, | |
18426c5b | 2006 | mode); |
f529eb25 | 2007 | bool success = false; |
2008 | bool speed = optimize_bb_for_speed_p (BLOCK_FOR_INSN (insn)); | |
8f8cadbc | 2009 | |
2010 | if (new_src == const0_rtx) | |
2011 | /* See above why we create (set (reg) (reg)) here. */ | |
2012 | success | |
2013 | = validate_change (next, &SET_SRC (set), reg, 0); | |
c9a03487 | 2014 | else |
8f8cadbc | 2015 | { |
c9a03487 | 2016 | rtx old_src = SET_SRC (set); |
2017 | struct full_rtx_costs oldcst, newcst; | |
18426c5b | 2018 | rtx tem = gen_rtx_PLUS (mode, reg, new_src); |
c9a03487 | 2019 | |
b72d459f | 2020 | get_full_set_rtx_cost (set, &oldcst); |
c9a03487 | 2021 | SET_SRC (set) = tem; |
18426c5b | 2022 | get_full_set_src_cost (tem, mode, &newcst); |
c9a03487 | 2023 | SET_SRC (set) = old_src; |
2024 | costs_add_n_insns (&oldcst, 1); | |
2025 | ||
2026 | if (costs_lt_p (&newcst, &oldcst, speed) | |
2027 | && have_add2_insn (reg, new_src)) | |
2028 | { | |
d1f9b275 | 2029 | rtx newpat = gen_rtx_SET (reg, tem); |
c9a03487 | 2030 | success |
2031 | = validate_change (next, &PATTERN (next), | |
2032 | newpat, 0); | |
2033 | } | |
8f8cadbc | 2034 | } |
2035 | if (success) | |
2036 | delete_insn (insn); | |
d83ccc81 | 2037 | changed |= success; |
8f8cadbc | 2038 | insn = next; |
b6b86e87 | 2039 | move2add_record_mode (reg); |
2040 | reg_offset[regno] | |
2041 | = trunc_int_for_mode (added_offset + base_offset, | |
18426c5b | 2042 | mode); |
8f8cadbc | 2043 | continue; |
2044 | } | |
2045 | } | |
2046 | } | |
6132c0d0 | 2047 | |
2048 | /* Try to transform | |
2049 | (set (REGX) (CONST (PLUS (SYMBOL_REF) (CONST_INT A)))) | |
2050 | ... | |
2051 | (set (REGY) (CONST (PLUS (SYMBOL_REF) (CONST_INT B)))) | |
2052 | to | |
2053 | (set (REGX) (CONST (PLUS (SYMBOL_REF) (CONST_INT A)))) | |
2054 | ... | |
2055 | (set (REGY) (CONST (PLUS (REGX) (CONST_INT B-A)))) */ | |
2056 | if ((GET_CODE (src) == SYMBOL_REF | |
2057 | || (GET_CODE (src) == CONST | |
2058 | && GET_CODE (XEXP (src, 0)) == PLUS | |
2059 | && GET_CODE (XEXP (XEXP (src, 0), 0)) == SYMBOL_REF | |
2060 | && CONST_INT_P (XEXP (XEXP (src, 0), 1)))) | |
2061 | && dbg_cnt (cse2_move2add)) | |
2062 | { | |
2063 | rtx sym, off; | |
2064 | ||
2065 | if (GET_CODE (src) == SYMBOL_REF) | |
2066 | { | |
2067 | sym = src; | |
2068 | off = const0_rtx; | |
2069 | } | |
2070 | else | |
2071 | { | |
2072 | sym = XEXP (XEXP (src, 0), 0); | |
2073 | off = XEXP (XEXP (src, 0), 1); | |
2074 | } | |
2075 | ||
2076 | /* If the reg already contains the value which is sum of | |
2077 | sym and some constant value, we can use an add2 insn. */ | |
18426c5b | 2078 | if (move2add_valid_value_p (regno, mode) |
6132c0d0 | 2079 | && reg_base_reg[regno] < 0 |
2080 | && reg_symbol_ref[regno] != NULL_RTX | |
2081 | && rtx_equal_p (sym, reg_symbol_ref[regno])) | |
18426c5b | 2082 | changed |= move2add_use_add2_insn (mode, reg, sym, off, insn); |
6132c0d0 | 2083 | |
2084 | /* Otherwise, we have to find a register whose value is sum | |
2085 | of sym and some constant value. */ | |
2086 | else | |
18426c5b | 2087 | changed |= move2add_use_add3_insn (mode, reg, sym, off, insn); |
6132c0d0 | 2088 | |
2089 | continue; | |
2090 | } | |
8f8cadbc | 2091 | } |
2092 | ||
2093 | for (note = REG_NOTES (insn); note; note = XEXP (note, 1)) | |
2094 | { | |
2095 | if (REG_NOTE_KIND (note) == REG_INC | |
8ad4c111 | 2096 | && REG_P (XEXP (note, 0))) |
8f8cadbc | 2097 | { |
2098 | /* Reset the information about this register. */ | |
2099 | int regno = REGNO (XEXP (note, 0)); | |
2100 | if (regno < FIRST_PSEUDO_REGISTER) | |
b6b86e87 | 2101 | { |
2102 | move2add_record_mode (XEXP (note, 0)); | |
2103 | reg_mode[regno] = VOIDmode; | |
2104 | } | |
8f8cadbc | 2105 | } |
2106 | } | |
6132c0d0 | 2107 | note_stores (PATTERN (insn), move2add_note_store, insn); |
8f8cadbc | 2108 | |
2109 | /* If INSN is a conditional branch, we try to extract an | |
2110 | implicit set out of it. */ | |
f222bc3b | 2111 | if (any_condjump_p (insn)) |
8f8cadbc | 2112 | { |
2113 | rtx cnd = fis_get_condition (insn); | |
2114 | ||
2115 | if (cnd != NULL_RTX | |
2116 | && GET_CODE (cnd) == NE | |
8ad4c111 | 2117 | && REG_P (XEXP (cnd, 0)) |
f222bc3b | 2118 | && !reg_set_p (XEXP (cnd, 0), insn) |
8f8cadbc | 2119 | /* The following two checks, which are also in |
2120 | move2add_note_store, are intended to reduce the | |
2121 | number of calls to gen_rtx_SET to avoid memory | |
2122 | allocation if possible. */ | |
2123 | && SCALAR_INT_MODE_P (GET_MODE (XEXP (cnd, 0))) | |
0933f1d9 | 2124 | && REG_NREGS (XEXP (cnd, 0)) == 1 |
971ba038 | 2125 | && CONST_INT_P (XEXP (cnd, 1))) |
8f8cadbc | 2126 | { |
2127 | rtx implicit_set = | |
d1f9b275 | 2128 | gen_rtx_SET (XEXP (cnd, 0), XEXP (cnd, 1)); |
6132c0d0 | 2129 | move2add_note_store (SET_DEST (implicit_set), implicit_set, insn); |
8f8cadbc | 2130 | } |
2131 | } | |
2132 | ||
2133 | /* If this is a CALL_INSN, all call used registers are stored with | |
2134 | unknown values. */ | |
6d7dc5b9 | 2135 | if (CALL_P (insn)) |
8f8cadbc | 2136 | { |
39bde736 | 2137 | rtx link; |
2138 | ||
8f8cadbc | 2139 | for (i = FIRST_PSEUDO_REGISTER - 1; i >= 0; i--) |
2140 | { | |
2141 | if (call_used_regs[i]) | |
2142 | /* Reset the information about this register. */ | |
b6b86e87 | 2143 | reg_mode[i] = VOIDmode; |
8f8cadbc | 2144 | } |
39bde736 | 2145 | |
2146 | for (link = CALL_INSN_FUNCTION_USAGE (insn); link; | |
2147 | link = XEXP (link, 1)) | |
2148 | { | |
2149 | rtx setuse = XEXP (link, 0); | |
2150 | rtx usage_rtx = XEXP (setuse, 0); | |
70bdfe23 | 2151 | /* CALL_INSN_FUNCTION_USAGEs can only have full clobbers, not |
2152 | clobber_highs. */ | |
2153 | gcc_assert (GET_CODE (setuse) != CLOBBER_HIGH); | |
39bde736 | 2154 | if (GET_CODE (setuse) == CLOBBER |
2155 | && REG_P (usage_rtx)) | |
2156 | { | |
2157 | unsigned int end_regno = END_REGNO (usage_rtx); | |
2158 | for (unsigned int r = REGNO (usage_rtx); r < end_regno; ++r) | |
2159 | /* Reset the information about this register. */ | |
2160 | reg_mode[r] = VOIDmode; | |
2161 | } | |
2162 | } | |
8f8cadbc | 2163 | } |
2164 | } | |
d83ccc81 | 2165 | return changed; |
8f8cadbc | 2166 | } |
2167 | ||
6132c0d0 | 2168 | /* SET is a SET or CLOBBER that sets DST. DATA is the insn which |
2169 | contains SET. | |
8f8cadbc | 2170 | Update reg_set_luid, reg_offset and reg_base_reg accordingly. |
2171 | Called from reload_cse_move2add via note_stores. */ | |
2172 | ||
2173 | static void | |
6132c0d0 | 2174 | move2add_note_store (rtx dst, const_rtx set, void *data) |
8f8cadbc | 2175 | { |
3aeaa53f | 2176 | rtx_insn *insn = (rtx_insn *) data; |
8f8cadbc | 2177 | unsigned int regno = 0; |
8974b7a3 | 2178 | scalar_int_mode mode; |
8f8cadbc | 2179 | |
8f8cadbc | 2180 | /* Some targets do argument pushes without adding REG_INC notes. */ |
2181 | ||
e16ceb8e | 2182 | if (MEM_P (dst)) |
8f8cadbc | 2183 | { |
2184 | dst = XEXP (dst, 0); | |
2185 | if (GET_CODE (dst) == PRE_INC || GET_CODE (dst) == POST_INC | |
2186 | || GET_CODE (dst) == PRE_DEC || GET_CODE (dst) == POST_DEC) | |
b6b86e87 | 2187 | reg_mode[REGNO (XEXP (dst, 0))] = VOIDmode; |
8f8cadbc | 2188 | return; |
2189 | } | |
8f8cadbc | 2190 | |
b6b86e87 | 2191 | if (GET_CODE (dst) == SUBREG) |
2192 | regno = subreg_regno (dst); | |
2193 | else if (REG_P (dst)) | |
2194 | regno = REGNO (dst); | |
2195 | else | |
2196 | return; | |
8f8cadbc | 2197 | |
8974b7a3 | 2198 | if (!is_a <scalar_int_mode> (GET_MODE (dst), &mode)) |
2199 | goto invalidate; | |
2200 | ||
2201 | if (GET_CODE (set) == SET) | |
6132c0d0 | 2202 | { |
2203 | rtx note, sym = NULL_RTX; | |
b6b86e87 | 2204 | rtx off; |
6132c0d0 | 2205 | |
2206 | note = find_reg_equal_equiv_note (insn); | |
2207 | if (note && GET_CODE (XEXP (note, 0)) == SYMBOL_REF) | |
2208 | { | |
2209 | sym = XEXP (note, 0); | |
b6b86e87 | 2210 | off = const0_rtx; |
6132c0d0 | 2211 | } |
2212 | else if (note && GET_CODE (XEXP (note, 0)) == CONST | |
2213 | && GET_CODE (XEXP (XEXP (note, 0), 0)) == PLUS | |
2214 | && GET_CODE (XEXP (XEXP (XEXP (note, 0), 0), 0)) == SYMBOL_REF | |
2215 | && CONST_INT_P (XEXP (XEXP (XEXP (note, 0), 0), 1))) | |
2216 | { | |
2217 | sym = XEXP (XEXP (XEXP (note, 0), 0), 0); | |
b6b86e87 | 2218 | off = XEXP (XEXP (XEXP (note, 0), 0), 1); |
6132c0d0 | 2219 | } |
2220 | ||
2221 | if (sym != NULL_RTX) | |
2222 | { | |
b6b86e87 | 2223 | move2add_record_sym_value (dst, sym, off); |
6132c0d0 | 2224 | return; |
2225 | } | |
2226 | } | |
2227 | ||
8974b7a3 | 2228 | if (GET_CODE (set) == SET |
8f8cadbc | 2229 | && GET_CODE (SET_DEST (set)) != ZERO_EXTRACT |
8f8cadbc | 2230 | && GET_CODE (SET_DEST (set)) != STRICT_LOW_PART) |
2231 | { | |
2232 | rtx src = SET_SRC (set); | |
2233 | rtx base_reg; | |
60141df0 | 2234 | unsigned HOST_WIDE_INT offset; |
8f8cadbc | 2235 | int base_regno; |
8f8cadbc | 2236 | |
2237 | switch (GET_CODE (src)) | |
2238 | { | |
2239 | case PLUS: | |
8ad4c111 | 2240 | if (REG_P (XEXP (src, 0))) |
8f8cadbc | 2241 | { |
2242 | base_reg = XEXP (src, 0); | |
2243 | ||
971ba038 | 2244 | if (CONST_INT_P (XEXP (src, 1))) |
60141df0 | 2245 | offset = UINTVAL (XEXP (src, 1)); |
8ad4c111 | 2246 | else if (REG_P (XEXP (src, 1)) |
b6b86e87 | 2247 | && move2add_valid_value_p (REGNO (XEXP (src, 1)), mode)) |
8f8cadbc | 2248 | { |
c389f975 | 2249 | if (reg_base_reg[REGNO (XEXP (src, 1))] < 0 |
2250 | && reg_symbol_ref[REGNO (XEXP (src, 1))] == NULL_RTX) | |
8f8cadbc | 2251 | offset = reg_offset[REGNO (XEXP (src, 1))]; |
2252 | /* Maybe the first register is known to be a | |
2253 | constant. */ | |
b6b86e87 | 2254 | else if (move2add_valid_value_p (REGNO (base_reg), mode) |
c389f975 | 2255 | && reg_base_reg[REGNO (base_reg)] < 0 |
2256 | && reg_symbol_ref[REGNO (base_reg)] == NULL_RTX) | |
8f8cadbc | 2257 | { |
2258 | offset = reg_offset[REGNO (base_reg)]; | |
2259 | base_reg = XEXP (src, 1); | |
2260 | } | |
2261 | else | |
2262 | goto invalidate; | |
2263 | } | |
2264 | else | |
2265 | goto invalidate; | |
2266 | ||
2267 | break; | |
2268 | } | |
2269 | ||
2270 | goto invalidate; | |
2271 | ||
2272 | case REG: | |
2273 | base_reg = src; | |
2274 | offset = 0; | |
2275 | break; | |
2276 | ||
2277 | case CONST_INT: | |
2278 | /* Start tracking the register as a constant. */ | |
2279 | reg_base_reg[regno] = -1; | |
6132c0d0 | 2280 | reg_symbol_ref[regno] = NULL_RTX; |
8f8cadbc | 2281 | reg_offset[regno] = INTVAL (SET_SRC (set)); |
2282 | /* We assign the same luid to all registers set to constants. */ | |
2283 | reg_set_luid[regno] = move2add_last_label_luid + 1; | |
b6b86e87 | 2284 | move2add_record_mode (dst); |
8f8cadbc | 2285 | return; |
2286 | ||
2287 | default: | |
b6b86e87 | 2288 | goto invalidate; |
8f8cadbc | 2289 | } |
2290 | ||
2291 | base_regno = REGNO (base_reg); | |
2292 | /* If information about the base register is not valid, set it | |
2293 | up as a new base register, pretending its value is known | |
2294 | starting from the current insn. */ | |
b6b86e87 | 2295 | if (!move2add_valid_value_p (base_regno, mode)) |
8f8cadbc | 2296 | { |
2297 | reg_base_reg[base_regno] = base_regno; | |
6132c0d0 | 2298 | reg_symbol_ref[base_regno] = NULL_RTX; |
8f8cadbc | 2299 | reg_offset[base_regno] = 0; |
2300 | reg_set_luid[base_regno] = move2add_luid; | |
b6b86e87 | 2301 | gcc_assert (GET_MODE (base_reg) == mode); |
2302 | move2add_record_mode (base_reg); | |
8f8cadbc | 2303 | } |
8f8cadbc | 2304 | |
2305 | /* Copy base information from our base register. */ | |
2306 | reg_set_luid[regno] = reg_set_luid[base_regno]; | |
2307 | reg_base_reg[regno] = reg_base_reg[base_regno]; | |
6132c0d0 | 2308 | reg_symbol_ref[regno] = reg_symbol_ref[base_regno]; |
8f8cadbc | 2309 | |
2310 | /* Compute the sum of the offsets or constants. */ | |
b6b86e87 | 2311 | reg_offset[regno] |
2312 | = trunc_int_for_mode (offset + reg_offset[base_regno], mode); | |
2313 | ||
2314 | move2add_record_mode (dst); | |
8f8cadbc | 2315 | } |
70bdfe23 | 2316 | else if (GET_CODE (set) == CLOBBER_HIGH) |
2317 | { | |
2318 | /* Only invalidate if actually clobbered. */ | |
2319 | if (reg_mode[regno] == BLKmode | |
2320 | || reg_is_clobbered_by_clobber_high (regno, reg_mode[regno], dst)) | |
2321 | goto invalidate; | |
2322 | } | |
8f8cadbc | 2323 | else |
2324 | { | |
b6b86e87 | 2325 | invalidate: |
2326 | /* Invalidate the contents of the register. */ | |
2327 | move2add_record_mode (dst); | |
2328 | reg_mode[regno] = VOIDmode; | |
8f8cadbc | 2329 | } |
2330 | } | |
77fce4cd | 2331 | \f |
cbe8bda8 | 2332 | namespace { |
2333 | ||
2334 | const pass_data pass_data_postreload_cse = | |
77fce4cd | 2335 | { |
cbe8bda8 | 2336 | RTL_PASS, /* type */ |
2337 | "postreload", /* name */ | |
2338 | OPTGROUP_NONE, /* optinfo_flags */ | |
cbe8bda8 | 2339 | TV_RELOAD_CSE_REGS, /* tv_id */ |
2340 | 0, /* properties_required */ | |
2341 | 0, /* properties_provided */ | |
2342 | 0, /* properties_destroyed */ | |
2343 | 0, /* todo_flags_start */ | |
8b88439e | 2344 | TODO_df_finish, /* todo_flags_finish */ |
77fce4cd | 2345 | }; |
cbe8bda8 | 2346 | |
2347 | class pass_postreload_cse : public rtl_opt_pass | |
2348 | { | |
2349 | public: | |
9af5ce0c | 2350 | pass_postreload_cse (gcc::context *ctxt) |
2351 | : rtl_opt_pass (pass_data_postreload_cse, ctxt) | |
cbe8bda8 | 2352 | {} |
2353 | ||
2354 | /* opt_pass methods: */ | |
31315c24 | 2355 | virtual bool gate (function *) { return (optimize > 0 && reload_completed); } |
2356 | ||
65b0537f | 2357 | virtual unsigned int execute (function *); |
cbe8bda8 | 2358 | |
2359 | }; // class pass_postreload_cse | |
2360 | ||
65b0537f | 2361 | unsigned int |
2362 | pass_postreload_cse::execute (function *fun) | |
2363 | { | |
2364 | if (!dbg_cnt (postreload_cse)) | |
2365 | return 0; | |
2366 | ||
2367 | /* Do a very simple CSE pass over just the hard registers. */ | |
2368 | reload_cse_regs (get_insns ()); | |
2369 | /* Reload_cse_regs can eliminate potentially-trapping MEMs. | |
2370 | Remove any EH edges associated with them. */ | |
2371 | if (fun->can_throw_non_call_exceptions | |
2372 | && purge_all_dead_edges ()) | |
2373 | cleanup_cfg (0); | |
2374 | ||
2375 | return 0; | |
2376 | } | |
2377 | ||
cbe8bda8 | 2378 | } // anon namespace |
2379 | ||
2380 | rtl_opt_pass * | |
2381 | make_pass_postreload_cse (gcc::context *ctxt) | |
2382 | { | |
2383 | return new pass_postreload_cse (ctxt); | |
2384 | } |