typeck.c (cp_truthvalue_conversion): Add tsubst_flags_t parameter and use it in calls...
[official-gcc.git] / gcc / lra-constraints.c
blob54b5ae5cffed9bbf0a58f0a46f2d8e1c6b0092b4
1 /* Code for RTL transformations to satisfy insn constraints.
2 Copyright (C) 2010-2019 Free Software Foundation, Inc.
3 Contributed by Vladimir Makarov <vmakarov@redhat.com>.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
10 version.
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
22 /* This file contains code for 3 passes: constraint pass,
23 inheritance/split pass, and pass for undoing failed inheritance and
24 split.
26 The major goal of constraint pass is to transform RTL to satisfy
27 insn and address constraints by:
28 o choosing insn alternatives;
29 o generating *reload insns* (or reloads in brief) and *reload
30 pseudos* which will get necessary hard registers later;
31 o substituting pseudos with equivalent values and removing the
32 instructions that initialized those pseudos.
34 The constraint pass has biggest and most complicated code in LRA.
35 There are a lot of important details like:
36 o reuse of input reload pseudos to simplify reload pseudo
37 allocations;
38 o some heuristics to choose insn alternative to improve the
39 inheritance;
40 o early clobbers etc.
42 The pass is mimicking former reload pass in alternative choosing
43 because the reload pass is oriented to current machine description
44 model. It might be changed if the machine description model is
45 changed.
47 There is special code for preventing all LRA and this pass cycling
48 in case of bugs.
50 On the first iteration of the pass we process every instruction and
51 choose an alternative for each one. On subsequent iterations we try
52 to avoid reprocessing instructions if we can be sure that the old
53 choice is still valid.
55 The inheritance/spilt pass is to transform code to achieve
56 ineheritance and live range splitting. It is done on backward
57 traversal of EBBs.
59 The inheritance optimization goal is to reuse values in hard
60 registers. There is analogous optimization in old reload pass. The
61 inheritance is achieved by following transformation:
63 reload_p1 <- p reload_p1 <- p
64 ... new_p <- reload_p1
65 ... => ...
66 reload_p2 <- p reload_p2 <- new_p
68 where p is spilled and not changed between the insns. Reload_p1 is
69 also called *original pseudo* and new_p is called *inheritance
70 pseudo*.
72 The subsequent assignment pass will try to assign the same (or
73 another if it is not possible) hard register to new_p as to
74 reload_p1 or reload_p2.
76 If the assignment pass fails to assign a hard register to new_p,
77 this file will undo the inheritance and restore the original code.
78 This is because implementing the above sequence with a spilled
79 new_p would make the code much worse. The inheritance is done in
80 EBB scope. The above is just a simplified example to get an idea
81 of the inheritance as the inheritance is also done for non-reload
82 insns.
84 Splitting (transformation) is also done in EBB scope on the same
85 pass as the inheritance:
87 r <- ... or ... <- r r <- ... or ... <- r
88 ... s <- r (new insn -- save)
89 ... =>
90 ... r <- s (new insn -- restore)
91 ... <- r ... <- r
93 The *split pseudo* s is assigned to the hard register of the
94 original pseudo or hard register r.
96 Splitting is done:
97 o In EBBs with high register pressure for global pseudos (living
98 in at least 2 BBs) and assigned to hard registers when there
99 are more one reloads needing the hard registers;
100 o for pseudos needing save/restore code around calls.
102 If the split pseudo still has the same hard register as the
103 original pseudo after the subsequent assignment pass or the
104 original pseudo was split, the opposite transformation is done on
105 the same pass for undoing inheritance. */
107 #undef REG_OK_STRICT
109 #include "config.h"
110 #include "system.h"
111 #include "coretypes.h"
112 #include "backend.h"
113 #include "target.h"
114 #include "rtl.h"
115 #include "tree.h"
116 #include "predict.h"
117 #include "df.h"
118 #include "memmodel.h"
119 #include "tm_p.h"
120 #include "expmed.h"
121 #include "optabs.h"
122 #include "regs.h"
123 #include "ira.h"
124 #include "recog.h"
125 #include "output.h"
126 #include "addresses.h"
127 #include "expr.h"
128 #include "cfgrtl.h"
129 #include "rtl-error.h"
130 #include "lra.h"
131 #include "lra-int.h"
132 #include "print-rtl.h"
133 #include "function-abi.h"
135 /* Value of LRA_CURR_RELOAD_NUM at the beginning of BB of the current
136 insn. Remember that LRA_CURR_RELOAD_NUM is the number of emitted
137 reload insns. */
138 static int bb_reload_num;
140 /* The current insn being processed and corresponding its single set
141 (NULL otherwise), its data (basic block, the insn data, the insn
142 static data, and the mode of each operand). */
143 static rtx_insn *curr_insn;
144 static rtx curr_insn_set;
145 static basic_block curr_bb;
146 static lra_insn_recog_data_t curr_id;
147 static struct lra_static_insn_data *curr_static_id;
148 static machine_mode curr_operand_mode[MAX_RECOG_OPERANDS];
149 /* Mode of the register substituted by its equivalence with VOIDmode
150 (e.g. constant) and whose subreg is given operand of the current
151 insn. VOIDmode in all other cases. */
152 static machine_mode original_subreg_reg_mode[MAX_RECOG_OPERANDS];
156 /* Start numbers for new registers and insns at the current constraints
157 pass start. */
158 static int new_regno_start;
159 static int new_insn_uid_start;
161 /* If LOC is nonnull, strip any outer subreg from it. */
162 static inline rtx *
163 strip_subreg (rtx *loc)
165 return loc && GET_CODE (*loc) == SUBREG ? &SUBREG_REG (*loc) : loc;
168 /* Return hard regno of REGNO or if it is was not assigned to a hard
169 register, use a hard register from its allocno class. */
170 static int
171 get_try_hard_regno (int regno)
173 int hard_regno;
174 enum reg_class rclass;
176 if ((hard_regno = regno) >= FIRST_PSEUDO_REGISTER)
177 hard_regno = lra_get_regno_hard_regno (regno);
178 if (hard_regno >= 0)
179 return hard_regno;
180 rclass = lra_get_allocno_class (regno);
181 if (rclass == NO_REGS)
182 return -1;
183 return ira_class_hard_regs[rclass][0];
186 /* Return the hard regno of X after removing its subreg. If X is not
187 a register or a subreg of a register, return -1. If X is a pseudo,
188 use its assignment. If FINAL_P return the final hard regno which will
189 be after elimination. */
190 static int
191 get_hard_regno (rtx x, bool final_p)
193 rtx reg;
194 int hard_regno;
196 reg = x;
197 if (SUBREG_P (x))
198 reg = SUBREG_REG (x);
199 if (! REG_P (reg))
200 return -1;
201 if (! HARD_REGISTER_NUM_P (hard_regno = REGNO (reg)))
202 hard_regno = lra_get_regno_hard_regno (hard_regno);
203 if (hard_regno < 0)
204 return -1;
205 if (final_p)
206 hard_regno = lra_get_elimination_hard_regno (hard_regno);
207 if (SUBREG_P (x))
208 hard_regno += subreg_regno_offset (hard_regno, GET_MODE (reg),
209 SUBREG_BYTE (x), GET_MODE (x));
210 return hard_regno;
213 /* If REGNO is a hard register or has been allocated a hard register,
214 return the class of that register. If REGNO is a reload pseudo
215 created by the current constraints pass, return its allocno class.
216 Return NO_REGS otherwise. */
217 static enum reg_class
218 get_reg_class (int regno)
220 int hard_regno;
222 if (! HARD_REGISTER_NUM_P (hard_regno = regno))
223 hard_regno = lra_get_regno_hard_regno (regno);
224 if (hard_regno >= 0)
226 hard_regno = lra_get_elimination_hard_regno (hard_regno);
227 return REGNO_REG_CLASS (hard_regno);
229 if (regno >= new_regno_start)
230 return lra_get_allocno_class (regno);
231 return NO_REGS;
234 /* Return true if REG satisfies (or will satisfy) reg class constraint
235 CL. Use elimination first if REG is a hard register. If REG is a
236 reload pseudo created by this constraints pass, assume that it will
237 be allocated a hard register from its allocno class, but allow that
238 class to be narrowed to CL if it is currently a superset of CL.
240 If NEW_CLASS is nonnull, set *NEW_CLASS to the new allocno class of
241 REGNO (reg), or NO_REGS if no change in its class was needed. */
242 static bool
243 in_class_p (rtx reg, enum reg_class cl, enum reg_class *new_class)
245 enum reg_class rclass, common_class;
246 machine_mode reg_mode;
247 int class_size, hard_regno, nregs, i, j;
248 int regno = REGNO (reg);
250 if (new_class != NULL)
251 *new_class = NO_REGS;
252 if (regno < FIRST_PSEUDO_REGISTER)
254 rtx final_reg = reg;
255 rtx *final_loc = &final_reg;
257 lra_eliminate_reg_if_possible (final_loc);
258 return TEST_HARD_REG_BIT (reg_class_contents[cl], REGNO (*final_loc));
260 reg_mode = GET_MODE (reg);
261 rclass = get_reg_class (regno);
262 if (regno < new_regno_start
263 /* Do not allow the constraints for reload instructions to
264 influence the classes of new pseudos. These reloads are
265 typically moves that have many alternatives, and restricting
266 reload pseudos for one alternative may lead to situations
267 where other reload pseudos are no longer allocatable. */
268 || (INSN_UID (curr_insn) >= new_insn_uid_start
269 && curr_insn_set != NULL
270 && ((OBJECT_P (SET_SRC (curr_insn_set))
271 && ! CONSTANT_P (SET_SRC (curr_insn_set)))
272 || (GET_CODE (SET_SRC (curr_insn_set)) == SUBREG
273 && OBJECT_P (SUBREG_REG (SET_SRC (curr_insn_set)))
274 && ! CONSTANT_P (SUBREG_REG (SET_SRC (curr_insn_set)))))))
275 /* When we don't know what class will be used finally for reload
276 pseudos, we use ALL_REGS. */
277 return ((regno >= new_regno_start && rclass == ALL_REGS)
278 || (rclass != NO_REGS && ira_class_subset_p[rclass][cl]
279 && ! hard_reg_set_subset_p (reg_class_contents[cl],
280 lra_no_alloc_regs)));
281 else
283 common_class = ira_reg_class_subset[rclass][cl];
284 if (new_class != NULL)
285 *new_class = common_class;
286 if (hard_reg_set_subset_p (reg_class_contents[common_class],
287 lra_no_alloc_regs))
288 return false;
289 /* Check that there are enough allocatable regs. */
290 class_size = ira_class_hard_regs_num[common_class];
291 for (i = 0; i < class_size; i++)
293 hard_regno = ira_class_hard_regs[common_class][i];
294 nregs = hard_regno_nregs (hard_regno, reg_mode);
295 if (nregs == 1)
296 return true;
297 for (j = 0; j < nregs; j++)
298 if (TEST_HARD_REG_BIT (lra_no_alloc_regs, hard_regno + j)
299 || ! TEST_HARD_REG_BIT (reg_class_contents[common_class],
300 hard_regno + j))
301 break;
302 if (j >= nregs)
303 return true;
305 return false;
309 /* Return true if REGNO satisfies a memory constraint. */
310 static bool
311 in_mem_p (int regno)
313 return get_reg_class (regno) == NO_REGS;
316 /* Return 1 if ADDR is a valid memory address for mode MODE in address
317 space AS, and check that each pseudo has the proper kind of hard
318 reg. */
319 static int
320 valid_address_p (machine_mode mode ATTRIBUTE_UNUSED,
321 rtx addr, addr_space_t as)
323 #ifdef GO_IF_LEGITIMATE_ADDRESS
324 lra_assert (ADDR_SPACE_GENERIC_P (as));
325 GO_IF_LEGITIMATE_ADDRESS (mode, addr, win);
326 return 0;
328 win:
329 return 1;
330 #else
331 return targetm.addr_space.legitimate_address_p (mode, addr, 0, as);
332 #endif
335 namespace {
336 /* Temporarily eliminates registers in an address (for the lifetime of
337 the object). */
338 class address_eliminator {
339 public:
340 address_eliminator (struct address_info *ad);
341 ~address_eliminator ();
343 private:
344 struct address_info *m_ad;
345 rtx *m_base_loc;
346 rtx m_base_reg;
347 rtx *m_index_loc;
348 rtx m_index_reg;
352 address_eliminator::address_eliminator (struct address_info *ad)
353 : m_ad (ad),
354 m_base_loc (strip_subreg (ad->base_term)),
355 m_base_reg (NULL_RTX),
356 m_index_loc (strip_subreg (ad->index_term)),
357 m_index_reg (NULL_RTX)
359 if (m_base_loc != NULL)
361 m_base_reg = *m_base_loc;
362 /* If we have non-legitimate address which is decomposed not in
363 the way we expected, don't do elimination here. In such case
364 the address will be reloaded and elimination will be done in
365 reload insn finally. */
366 if (REG_P (m_base_reg))
367 lra_eliminate_reg_if_possible (m_base_loc);
368 if (m_ad->base_term2 != NULL)
369 *m_ad->base_term2 = *m_ad->base_term;
371 if (m_index_loc != NULL)
373 m_index_reg = *m_index_loc;
374 if (REG_P (m_index_reg))
375 lra_eliminate_reg_if_possible (m_index_loc);
379 address_eliminator::~address_eliminator ()
381 if (m_base_loc && *m_base_loc != m_base_reg)
383 *m_base_loc = m_base_reg;
384 if (m_ad->base_term2 != NULL)
385 *m_ad->base_term2 = *m_ad->base_term;
387 if (m_index_loc && *m_index_loc != m_index_reg)
388 *m_index_loc = m_index_reg;
391 /* Return true if the eliminated form of AD is a legitimate target address. */
392 static bool
393 valid_address_p (struct address_info *ad)
395 address_eliminator eliminator (ad);
396 return valid_address_p (ad->mode, *ad->outer, ad->as);
399 /* Return true if the eliminated form of memory reference OP satisfies
400 extra (special) memory constraint CONSTRAINT. */
401 static bool
402 satisfies_memory_constraint_p (rtx op, enum constraint_num constraint)
404 struct address_info ad;
406 decompose_mem_address (&ad, op);
407 address_eliminator eliminator (&ad);
408 return constraint_satisfied_p (op, constraint);
411 /* Return true if the eliminated form of address AD satisfies extra
412 address constraint CONSTRAINT. */
413 static bool
414 satisfies_address_constraint_p (struct address_info *ad,
415 enum constraint_num constraint)
417 address_eliminator eliminator (ad);
418 return constraint_satisfied_p (*ad->outer, constraint);
421 /* Return true if the eliminated form of address OP satisfies extra
422 address constraint CONSTRAINT. */
423 static bool
424 satisfies_address_constraint_p (rtx op, enum constraint_num constraint)
426 struct address_info ad;
428 decompose_lea_address (&ad, &op);
429 return satisfies_address_constraint_p (&ad, constraint);
432 /* Initiate equivalences for LRA. As we keep original equivalences
433 before any elimination, we need to make copies otherwise any change
434 in insns might change the equivalences. */
435 void
436 lra_init_equiv (void)
438 ira_expand_reg_equiv ();
439 for (int i = FIRST_PSEUDO_REGISTER; i < max_reg_num (); i++)
441 rtx res;
443 if ((res = ira_reg_equiv[i].memory) != NULL_RTX)
444 ira_reg_equiv[i].memory = copy_rtx (res);
445 if ((res = ira_reg_equiv[i].invariant) != NULL_RTX)
446 ira_reg_equiv[i].invariant = copy_rtx (res);
450 static rtx loc_equivalence_callback (rtx, const_rtx, void *);
452 /* Update equivalence for REGNO. We need to this as the equivalence
453 might contain other pseudos which are changed by their
454 equivalences. */
455 static void
456 update_equiv (int regno)
458 rtx x;
460 if ((x = ira_reg_equiv[regno].memory) != NULL_RTX)
461 ira_reg_equiv[regno].memory
462 = simplify_replace_fn_rtx (x, NULL_RTX, loc_equivalence_callback,
463 NULL_RTX);
464 if ((x = ira_reg_equiv[regno].invariant) != NULL_RTX)
465 ira_reg_equiv[regno].invariant
466 = simplify_replace_fn_rtx (x, NULL_RTX, loc_equivalence_callback,
467 NULL_RTX);
470 /* If we have decided to substitute X with another value, return that
471 value, otherwise return X. */
472 static rtx
473 get_equiv (rtx x)
475 int regno;
476 rtx res;
478 if (! REG_P (x) || (regno = REGNO (x)) < FIRST_PSEUDO_REGISTER
479 || ! ira_reg_equiv[regno].defined_p
480 || ! ira_reg_equiv[regno].profitable_p
481 || lra_get_regno_hard_regno (regno) >= 0)
482 return x;
483 if ((res = ira_reg_equiv[regno].memory) != NULL_RTX)
485 if (targetm.cannot_substitute_mem_equiv_p (res))
486 return x;
487 return res;
489 if ((res = ira_reg_equiv[regno].constant) != NULL_RTX)
490 return res;
491 if ((res = ira_reg_equiv[regno].invariant) != NULL_RTX)
492 return res;
493 gcc_unreachable ();
496 /* If we have decided to substitute X with the equivalent value,
497 return that value after elimination for INSN, otherwise return
498 X. */
499 static rtx
500 get_equiv_with_elimination (rtx x, rtx_insn *insn)
502 rtx res = get_equiv (x);
504 if (x == res || CONSTANT_P (res))
505 return res;
506 return lra_eliminate_regs_1 (insn, res, GET_MODE (res),
507 false, false, 0, true);
510 /* Set up curr_operand_mode. */
511 static void
512 init_curr_operand_mode (void)
514 int nop = curr_static_id->n_operands;
515 for (int i = 0; i < nop; i++)
517 machine_mode mode = GET_MODE (*curr_id->operand_loc[i]);
518 if (mode == VOIDmode)
520 /* The .md mode for address operands is the mode of the
521 addressed value rather than the mode of the address itself. */
522 if (curr_id->icode >= 0 && curr_static_id->operand[i].is_address)
523 mode = Pmode;
524 else
525 mode = curr_static_id->operand[i].mode;
527 curr_operand_mode[i] = mode;
533 /* The page contains code to reuse input reloads. */
535 /* Structure describes input reload of the current insns. */
536 struct input_reload
538 /* True for input reload of matched operands. */
539 bool match_p;
540 /* Reloaded value. */
541 rtx input;
542 /* Reload pseudo used. */
543 rtx reg;
546 /* The number of elements in the following array. */
547 static int curr_insn_input_reloads_num;
548 /* Array containing info about input reloads. It is used to find the
549 same input reload and reuse the reload pseudo in this case. */
550 static struct input_reload curr_insn_input_reloads[LRA_MAX_INSN_RELOADS];
552 /* Initiate data concerning reuse of input reloads for the current
553 insn. */
554 static void
555 init_curr_insn_input_reloads (void)
557 curr_insn_input_reloads_num = 0;
560 /* Create a new pseudo using MODE, RCLASS, ORIGINAL or reuse already
561 created input reload pseudo (only if TYPE is not OP_OUT). Don't
562 reuse pseudo if IN_SUBREG_P is true and the reused pseudo should be
563 wrapped up in SUBREG. The result pseudo is returned through
564 RESULT_REG. Return TRUE if we created a new pseudo, FALSE if we
565 reused the already created input reload pseudo. Use TITLE to
566 describe new registers for debug purposes. */
567 static bool
568 get_reload_reg (enum op_type type, machine_mode mode, rtx original,
569 enum reg_class rclass, bool in_subreg_p,
570 const char *title, rtx *result_reg)
572 int i, regno;
573 enum reg_class new_class;
574 bool unique_p = false;
576 if (type == OP_OUT)
578 *result_reg
579 = lra_create_new_reg_with_unique_value (mode, original, rclass, title);
580 return true;
582 /* Prevent reuse value of expression with side effects,
583 e.g. volatile memory. */
584 if (! side_effects_p (original))
585 for (i = 0; i < curr_insn_input_reloads_num; i++)
587 if (! curr_insn_input_reloads[i].match_p
588 && rtx_equal_p (curr_insn_input_reloads[i].input, original)
589 && in_class_p (curr_insn_input_reloads[i].reg, rclass, &new_class))
591 rtx reg = curr_insn_input_reloads[i].reg;
592 regno = REGNO (reg);
593 /* If input is equal to original and both are VOIDmode,
594 GET_MODE (reg) might be still different from mode.
595 Ensure we don't return *result_reg with wrong mode. */
596 if (GET_MODE (reg) != mode)
598 if (in_subreg_p)
599 continue;
600 if (maybe_lt (GET_MODE_SIZE (GET_MODE (reg)),
601 GET_MODE_SIZE (mode)))
602 continue;
603 reg = lowpart_subreg (mode, reg, GET_MODE (reg));
604 if (reg == NULL_RTX || GET_CODE (reg) != SUBREG)
605 continue;
607 *result_reg = reg;
608 if (lra_dump_file != NULL)
610 fprintf (lra_dump_file, " Reuse r%d for reload ", regno);
611 dump_value_slim (lra_dump_file, original, 1);
613 if (new_class != lra_get_allocno_class (regno))
614 lra_change_class (regno, new_class, ", change to", false);
615 if (lra_dump_file != NULL)
616 fprintf (lra_dump_file, "\n");
617 return false;
619 /* If we have an input reload with a different mode, make sure it
620 will get a different hard reg. */
621 else if (REG_P (original)
622 && REG_P (curr_insn_input_reloads[i].input)
623 && REGNO (original) == REGNO (curr_insn_input_reloads[i].input)
624 && (GET_MODE (original)
625 != GET_MODE (curr_insn_input_reloads[i].input)))
626 unique_p = true;
628 *result_reg = (unique_p
629 ? lra_create_new_reg_with_unique_value
630 : lra_create_new_reg) (mode, original, rclass, title);
631 lra_assert (curr_insn_input_reloads_num < LRA_MAX_INSN_RELOADS);
632 curr_insn_input_reloads[curr_insn_input_reloads_num].input = original;
633 curr_insn_input_reloads[curr_insn_input_reloads_num].match_p = false;
634 curr_insn_input_reloads[curr_insn_input_reloads_num++].reg = *result_reg;
635 return true;
639 /* The page contains major code to choose the current insn alternative
640 and generate reloads for it. */
642 /* Return the offset from REGNO of the least significant register
643 in (reg:MODE REGNO).
645 This function is used to tell whether two registers satisfy
646 a matching constraint. (reg:MODE1 REGNO1) matches (reg:MODE2 REGNO2) if:
648 REGNO1 + lra_constraint_offset (REGNO1, MODE1)
649 == REGNO2 + lra_constraint_offset (REGNO2, MODE2) */
651 lra_constraint_offset (int regno, machine_mode mode)
653 lra_assert (regno < FIRST_PSEUDO_REGISTER);
655 scalar_int_mode int_mode;
656 if (WORDS_BIG_ENDIAN
657 && is_a <scalar_int_mode> (mode, &int_mode)
658 && GET_MODE_SIZE (int_mode) > UNITS_PER_WORD)
659 return hard_regno_nregs (regno, mode) - 1;
660 return 0;
663 /* Like rtx_equal_p except that it allows a REG and a SUBREG to match
664 if they are the same hard reg, and has special hacks for
665 auto-increment and auto-decrement. This is specifically intended for
666 process_alt_operands to use in determining whether two operands
667 match. X is the operand whose number is the lower of the two.
669 It is supposed that X is the output operand and Y is the input
670 operand. Y_HARD_REGNO is the final hard regno of register Y or
671 register in subreg Y as we know it now. Otherwise, it is a
672 negative value. */
673 static bool
674 operands_match_p (rtx x, rtx y, int y_hard_regno)
676 int i;
677 RTX_CODE code = GET_CODE (x);
678 const char *fmt;
680 if (x == y)
681 return true;
682 if ((code == REG || (code == SUBREG && REG_P (SUBREG_REG (x))))
683 && (REG_P (y) || (GET_CODE (y) == SUBREG && REG_P (SUBREG_REG (y)))))
685 int j;
687 i = get_hard_regno (x, false);
688 if (i < 0)
689 goto slow;
691 if ((j = y_hard_regno) < 0)
692 goto slow;
694 i += lra_constraint_offset (i, GET_MODE (x));
695 j += lra_constraint_offset (j, GET_MODE (y));
697 return i == j;
700 /* If two operands must match, because they are really a single
701 operand of an assembler insn, then two post-increments are invalid
702 because the assembler insn would increment only once. On the
703 other hand, a post-increment matches ordinary indexing if the
704 post-increment is the output operand. */
705 if (code == POST_DEC || code == POST_INC || code == POST_MODIFY)
706 return operands_match_p (XEXP (x, 0), y, y_hard_regno);
708 /* Two pre-increments are invalid because the assembler insn would
709 increment only once. On the other hand, a pre-increment matches
710 ordinary indexing if the pre-increment is the input operand. */
711 if (GET_CODE (y) == PRE_DEC || GET_CODE (y) == PRE_INC
712 || GET_CODE (y) == PRE_MODIFY)
713 return operands_match_p (x, XEXP (y, 0), -1);
715 slow:
717 if (code == REG && REG_P (y))
718 return REGNO (x) == REGNO (y);
720 if (code == REG && GET_CODE (y) == SUBREG && REG_P (SUBREG_REG (y))
721 && x == SUBREG_REG (y))
722 return true;
723 if (GET_CODE (y) == REG && code == SUBREG && REG_P (SUBREG_REG (x))
724 && SUBREG_REG (x) == y)
725 return true;
727 /* Now we have disposed of all the cases in which different rtx
728 codes can match. */
729 if (code != GET_CODE (y))
730 return false;
732 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
733 if (GET_MODE (x) != GET_MODE (y))
734 return false;
736 switch (code)
738 CASE_CONST_UNIQUE:
739 return false;
741 case LABEL_REF:
742 return label_ref_label (x) == label_ref_label (y);
743 case SYMBOL_REF:
744 return XSTR (x, 0) == XSTR (y, 0);
746 default:
747 break;
750 /* Compare the elements. If any pair of corresponding elements fail
751 to match, return false for the whole things. */
753 fmt = GET_RTX_FORMAT (code);
754 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
756 int val, j;
757 switch (fmt[i])
759 case 'w':
760 if (XWINT (x, i) != XWINT (y, i))
761 return false;
762 break;
764 case 'i':
765 if (XINT (x, i) != XINT (y, i))
766 return false;
767 break;
769 case 'p':
770 if (maybe_ne (SUBREG_BYTE (x), SUBREG_BYTE (y)))
771 return false;
772 break;
774 case 'e':
775 val = operands_match_p (XEXP (x, i), XEXP (y, i), -1);
776 if (val == 0)
777 return false;
778 break;
780 case '0':
781 break;
783 case 'E':
784 if (XVECLEN (x, i) != XVECLEN (y, i))
785 return false;
786 for (j = XVECLEN (x, i) - 1; j >= 0; --j)
788 val = operands_match_p (XVECEXP (x, i, j), XVECEXP (y, i, j), -1);
789 if (val == 0)
790 return false;
792 break;
794 /* It is believed that rtx's at this level will never
795 contain anything but integers and other rtx's, except for
796 within LABEL_REFs and SYMBOL_REFs. */
797 default:
798 gcc_unreachable ();
801 return true;
804 /* True if X is a constant that can be forced into the constant pool.
805 MODE is the mode of the operand, or VOIDmode if not known. */
806 #define CONST_POOL_OK_P(MODE, X) \
807 ((MODE) != VOIDmode \
808 && CONSTANT_P (X) \
809 && GET_CODE (X) != HIGH \
810 && GET_MODE_SIZE (MODE).is_constant () \
811 && !targetm.cannot_force_const_mem (MODE, X))
813 /* True if C is a non-empty register class that has too few registers
814 to be safely used as a reload target class. */
815 #define SMALL_REGISTER_CLASS_P(C) \
816 (ira_class_hard_regs_num [(C)] == 1 \
817 || (ira_class_hard_regs_num [(C)] >= 1 \
818 && targetm.class_likely_spilled_p (C)))
820 /* If REG is a reload pseudo, try to make its class satisfying CL. */
821 static void
822 narrow_reload_pseudo_class (rtx reg, enum reg_class cl)
824 enum reg_class rclass;
826 /* Do not make more accurate class from reloads generated. They are
827 mostly moves with a lot of constraints. Making more accurate
828 class may results in very narrow class and impossibility of find
829 registers for several reloads of one insn. */
830 if (INSN_UID (curr_insn) >= new_insn_uid_start)
831 return;
832 if (GET_CODE (reg) == SUBREG)
833 reg = SUBREG_REG (reg);
834 if (! REG_P (reg) || (int) REGNO (reg) < new_regno_start)
835 return;
836 if (in_class_p (reg, cl, &rclass) && rclass != cl)
837 lra_change_class (REGNO (reg), rclass, " Change to", true);
840 /* Searches X for any reference to a reg with the same value as REGNO,
841 returning the rtx of the reference found if any. Otherwise,
842 returns NULL_RTX. */
843 static rtx
844 regno_val_use_in (unsigned int regno, rtx x)
846 const char *fmt;
847 int i, j;
848 rtx tem;
850 if (REG_P (x) && lra_reg_info[REGNO (x)].val == lra_reg_info[regno].val)
851 return x;
853 fmt = GET_RTX_FORMAT (GET_CODE (x));
854 for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
856 if (fmt[i] == 'e')
858 if ((tem = regno_val_use_in (regno, XEXP (x, i))))
859 return tem;
861 else if (fmt[i] == 'E')
862 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
863 if ((tem = regno_val_use_in (regno , XVECEXP (x, i, j))))
864 return tem;
867 return NULL_RTX;
870 /* Return true if all current insn non-output operands except INS (it
871 has a negaitve end marker) do not use pseudos with the same value
872 as REGNO. */
873 static bool
874 check_conflict_input_operands (int regno, signed char *ins)
876 int in;
877 int n_operands = curr_static_id->n_operands;
879 for (int nop = 0; nop < n_operands; nop++)
880 if (! curr_static_id->operand[nop].is_operator
881 && curr_static_id->operand[nop].type != OP_OUT)
883 for (int i = 0; (in = ins[i]) >= 0; i++)
884 if (in == nop)
885 break;
886 if (in < 0
887 && regno_val_use_in (regno, *curr_id->operand_loc[nop]) != NULL_RTX)
888 return false;
890 return true;
893 /* Generate reloads for matching OUT and INS (array of input operand
894 numbers with end marker -1) with reg class GOAL_CLASS, considering
895 output operands OUTS (similar array to INS) needing to be in different
896 registers. Add input and output reloads correspondingly to the lists
897 *BEFORE and *AFTER. OUT might be negative. In this case we generate
898 input reloads for matched input operands INS. EARLY_CLOBBER_P is a flag
899 that the output operand is early clobbered for chosen alternative. */
900 static void
901 match_reload (signed char out, signed char *ins, signed char *outs,
902 enum reg_class goal_class, rtx_insn **before,
903 rtx_insn **after, bool early_clobber_p)
905 bool out_conflict;
906 int i, in;
907 rtx new_in_reg, new_out_reg, reg;
908 machine_mode inmode, outmode;
909 rtx in_rtx = *curr_id->operand_loc[ins[0]];
910 rtx out_rtx = out < 0 ? in_rtx : *curr_id->operand_loc[out];
912 inmode = curr_operand_mode[ins[0]];
913 outmode = out < 0 ? inmode : curr_operand_mode[out];
914 push_to_sequence (*before);
915 if (inmode != outmode)
917 /* process_alt_operands has already checked that the mode sizes
918 are ordered. */
919 if (partial_subreg_p (outmode, inmode))
921 reg = new_in_reg
922 = lra_create_new_reg_with_unique_value (inmode, in_rtx,
923 goal_class, "");
924 new_out_reg = gen_lowpart_SUBREG (outmode, reg);
925 LRA_SUBREG_P (new_out_reg) = 1;
926 /* If the input reg is dying here, we can use the same hard
927 register for REG and IN_RTX. We do it only for original
928 pseudos as reload pseudos can die although original
929 pseudos still live where reload pseudos dies. */
930 if (REG_P (in_rtx) && (int) REGNO (in_rtx) < lra_new_regno_start
931 && find_regno_note (curr_insn, REG_DEAD, REGNO (in_rtx))
932 && (!early_clobber_p
933 || check_conflict_input_operands(REGNO (in_rtx), ins)))
934 lra_assign_reg_val (REGNO (in_rtx), REGNO (reg));
936 else
938 reg = new_out_reg
939 = lra_create_new_reg_with_unique_value (outmode, out_rtx,
940 goal_class, "");
941 new_in_reg = gen_lowpart_SUBREG (inmode, reg);
942 /* NEW_IN_REG is non-paradoxical subreg. We don't want
943 NEW_OUT_REG living above. We add clobber clause for
944 this. This is just a temporary clobber. We can remove
945 it at the end of LRA work. */
946 rtx_insn *clobber = emit_clobber (new_out_reg);
947 LRA_TEMP_CLOBBER_P (PATTERN (clobber)) = 1;
948 LRA_SUBREG_P (new_in_reg) = 1;
949 if (GET_CODE (in_rtx) == SUBREG)
951 rtx subreg_reg = SUBREG_REG (in_rtx);
953 /* If SUBREG_REG is dying here and sub-registers IN_RTX
954 and NEW_IN_REG are similar, we can use the same hard
955 register for REG and SUBREG_REG. */
956 if (REG_P (subreg_reg)
957 && (int) REGNO (subreg_reg) < lra_new_regno_start
958 && GET_MODE (subreg_reg) == outmode
959 && known_eq (SUBREG_BYTE (in_rtx), SUBREG_BYTE (new_in_reg))
960 && find_regno_note (curr_insn, REG_DEAD, REGNO (subreg_reg))
961 && (! early_clobber_p
962 || check_conflict_input_operands (REGNO (subreg_reg),
963 ins)))
964 lra_assign_reg_val (REGNO (subreg_reg), REGNO (reg));
968 else
970 /* Pseudos have values -- see comments for lra_reg_info.
971 Different pseudos with the same value do not conflict even if
972 they live in the same place. When we create a pseudo we
973 assign value of original pseudo (if any) from which we
974 created the new pseudo. If we create the pseudo from the
975 input pseudo, the new pseudo will have no conflict with the
976 input pseudo which is wrong when the input pseudo lives after
977 the insn and as the new pseudo value is changed by the insn
978 output. Therefore we create the new pseudo from the output
979 except the case when we have single matched dying input
980 pseudo.
982 We cannot reuse the current output register because we might
983 have a situation like "a <- a op b", where the constraints
984 force the second input operand ("b") to match the output
985 operand ("a"). "b" must then be copied into a new register
986 so that it doesn't clobber the current value of "a".
988 We cannot use the same value if the output pseudo is
989 early clobbered or the input pseudo is mentioned in the
990 output, e.g. as an address part in memory, because
991 output reload will actually extend the pseudo liveness.
992 We don't care about eliminable hard regs here as we are
993 interesting only in pseudos. */
995 /* Matching input's register value is the same as one of the other
996 output operand. Output operands in a parallel insn must be in
997 different registers. */
998 out_conflict = false;
999 if (REG_P (in_rtx))
1001 for (i = 0; outs[i] >= 0; i++)
1003 rtx other_out_rtx = *curr_id->operand_loc[outs[i]];
1004 if (REG_P (other_out_rtx)
1005 && (regno_val_use_in (REGNO (in_rtx), other_out_rtx)
1006 != NULL_RTX))
1008 out_conflict = true;
1009 break;
1014 new_in_reg = new_out_reg
1015 = (! early_clobber_p && ins[1] < 0 && REG_P (in_rtx)
1016 && (int) REGNO (in_rtx) < lra_new_regno_start
1017 && find_regno_note (curr_insn, REG_DEAD, REGNO (in_rtx))
1018 && (! early_clobber_p
1019 || check_conflict_input_operands (REGNO (in_rtx), ins))
1020 && (out < 0
1021 || regno_val_use_in (REGNO (in_rtx), out_rtx) == NULL_RTX)
1022 && !out_conflict
1023 ? lra_create_new_reg (inmode, in_rtx, goal_class, "")
1024 : lra_create_new_reg_with_unique_value (outmode, out_rtx,
1025 goal_class, ""));
1027 /* In operand can be got from transformations before processing insn
1028 constraints. One example of such transformations is subreg
1029 reloading (see function simplify_operand_subreg). The new
1030 pseudos created by the transformations might have inaccurate
1031 class (ALL_REGS) and we should make their classes more
1032 accurate. */
1033 narrow_reload_pseudo_class (in_rtx, goal_class);
1034 lra_emit_move (copy_rtx (new_in_reg), in_rtx);
1035 *before = get_insns ();
1036 end_sequence ();
1037 /* Add the new pseudo to consider values of subsequent input reload
1038 pseudos. */
1039 lra_assert (curr_insn_input_reloads_num < LRA_MAX_INSN_RELOADS);
1040 curr_insn_input_reloads[curr_insn_input_reloads_num].input = in_rtx;
1041 curr_insn_input_reloads[curr_insn_input_reloads_num].match_p = true;
1042 curr_insn_input_reloads[curr_insn_input_reloads_num++].reg = new_in_reg;
1043 for (i = 0; (in = ins[i]) >= 0; i++)
1045 lra_assert
1046 (GET_MODE (*curr_id->operand_loc[in]) == VOIDmode
1047 || GET_MODE (new_in_reg) == GET_MODE (*curr_id->operand_loc[in]));
1048 *curr_id->operand_loc[in] = new_in_reg;
1050 lra_update_dups (curr_id, ins);
1051 if (out < 0)
1052 return;
1053 /* See a comment for the input operand above. */
1054 narrow_reload_pseudo_class (out_rtx, goal_class);
1055 if (find_reg_note (curr_insn, REG_UNUSED, out_rtx) == NULL_RTX)
1057 start_sequence ();
1058 lra_emit_move (out_rtx, copy_rtx (new_out_reg));
1059 emit_insn (*after);
1060 *after = get_insns ();
1061 end_sequence ();
1063 *curr_id->operand_loc[out] = new_out_reg;
1064 lra_update_dup (curr_id, out);
1067 /* Return register class which is union of all reg classes in insn
1068 constraint alternative string starting with P. */
1069 static enum reg_class
1070 reg_class_from_constraints (const char *p)
1072 int c, len;
1073 enum reg_class op_class = NO_REGS;
1076 switch ((c = *p, len = CONSTRAINT_LEN (c, p)), c)
1078 case '#':
1079 case ',':
1080 return op_class;
1082 case 'g':
1083 op_class = reg_class_subunion[op_class][GENERAL_REGS];
1084 break;
1086 default:
1087 enum constraint_num cn = lookup_constraint (p);
1088 enum reg_class cl = reg_class_for_constraint (cn);
1089 if (cl == NO_REGS)
1091 if (insn_extra_address_constraint (cn))
1092 op_class
1093 = (reg_class_subunion
1094 [op_class][base_reg_class (VOIDmode, ADDR_SPACE_GENERIC,
1095 ADDRESS, SCRATCH)]);
1096 break;
1099 op_class = reg_class_subunion[op_class][cl];
1100 break;
1102 while ((p += len), c);
1103 return op_class;
1106 /* If OP is a register, return the class of the register as per
1107 get_reg_class, otherwise return NO_REGS. */
1108 static inline enum reg_class
1109 get_op_class (rtx op)
1111 return REG_P (op) ? get_reg_class (REGNO (op)) : NO_REGS;
1114 /* Return generated insn mem_pseudo:=val if TO_P or val:=mem_pseudo
1115 otherwise. If modes of MEM_PSEUDO and VAL are different, use
1116 SUBREG for VAL to make them equal. */
1117 static rtx_insn *
1118 emit_spill_move (bool to_p, rtx mem_pseudo, rtx val)
1120 if (GET_MODE (mem_pseudo) != GET_MODE (val))
1122 /* Usually size of mem_pseudo is greater than val size but in
1123 rare cases it can be less as it can be defined by target
1124 dependent macro HARD_REGNO_CALLER_SAVE_MODE. */
1125 if (! MEM_P (val))
1127 val = gen_lowpart_SUBREG (GET_MODE (mem_pseudo),
1128 GET_CODE (val) == SUBREG
1129 ? SUBREG_REG (val) : val);
1130 LRA_SUBREG_P (val) = 1;
1132 else
1134 mem_pseudo = gen_lowpart_SUBREG (GET_MODE (val), mem_pseudo);
1135 LRA_SUBREG_P (mem_pseudo) = 1;
1138 return to_p ? gen_move_insn (mem_pseudo, val)
1139 : gen_move_insn (val, mem_pseudo);
1142 /* Process a special case insn (register move), return true if we
1143 don't need to process it anymore. INSN should be a single set
1144 insn. Set up that RTL was changed through CHANGE_P and that hook
1145 TARGET_SECONDARY_MEMORY_NEEDED says to use secondary memory through
1146 SEC_MEM_P. */
1147 static bool
1148 check_and_process_move (bool *change_p, bool *sec_mem_p ATTRIBUTE_UNUSED)
1150 int sregno, dregno;
1151 rtx dest, src, dreg, sreg, new_reg, scratch_reg;
1152 rtx_insn *before;
1153 enum reg_class dclass, sclass, secondary_class;
1154 secondary_reload_info sri;
1156 lra_assert (curr_insn_set != NULL_RTX);
1157 dreg = dest = SET_DEST (curr_insn_set);
1158 sreg = src = SET_SRC (curr_insn_set);
1159 if (GET_CODE (dest) == SUBREG)
1160 dreg = SUBREG_REG (dest);
1161 if (GET_CODE (src) == SUBREG)
1162 sreg = SUBREG_REG (src);
1163 if (! (REG_P (dreg) || MEM_P (dreg)) || ! (REG_P (sreg) || MEM_P (sreg)))
1164 return false;
1165 sclass = dclass = NO_REGS;
1166 if (REG_P (dreg))
1167 dclass = get_reg_class (REGNO (dreg));
1168 gcc_assert (dclass < LIM_REG_CLASSES);
1169 if (dclass == ALL_REGS)
1170 /* ALL_REGS is used for new pseudos created by transformations
1171 like reload of SUBREG_REG (see function
1172 simplify_operand_subreg). We don't know their class yet. We
1173 should figure out the class from processing the insn
1174 constraints not in this fast path function. Even if ALL_REGS
1175 were a right class for the pseudo, secondary_... hooks usually
1176 are not define for ALL_REGS. */
1177 return false;
1178 if (REG_P (sreg))
1179 sclass = get_reg_class (REGNO (sreg));
1180 gcc_assert (sclass < LIM_REG_CLASSES);
1181 if (sclass == ALL_REGS)
1182 /* See comments above. */
1183 return false;
1184 if (sclass == NO_REGS && dclass == NO_REGS)
1185 return false;
1186 if (targetm.secondary_memory_needed (GET_MODE (src), sclass, dclass)
1187 && ((sclass != NO_REGS && dclass != NO_REGS)
1188 || (GET_MODE (src)
1189 != targetm.secondary_memory_needed_mode (GET_MODE (src)))))
1191 *sec_mem_p = true;
1192 return false;
1194 if (! REG_P (dreg) || ! REG_P (sreg))
1195 return false;
1196 sri.prev_sri = NULL;
1197 sri.icode = CODE_FOR_nothing;
1198 sri.extra_cost = 0;
1199 secondary_class = NO_REGS;
1200 /* Set up hard register for a reload pseudo for hook
1201 secondary_reload because some targets just ignore unassigned
1202 pseudos in the hook. */
1203 if (dclass != NO_REGS && lra_get_regno_hard_regno (REGNO (dreg)) < 0)
1205 dregno = REGNO (dreg);
1206 reg_renumber[dregno] = ira_class_hard_regs[dclass][0];
1208 else
1209 dregno = -1;
1210 if (sclass != NO_REGS && lra_get_regno_hard_regno (REGNO (sreg)) < 0)
1212 sregno = REGNO (sreg);
1213 reg_renumber[sregno] = ira_class_hard_regs[sclass][0];
1215 else
1216 sregno = -1;
1217 if (sclass != NO_REGS)
1218 secondary_class
1219 = (enum reg_class) targetm.secondary_reload (false, dest,
1220 (reg_class_t) sclass,
1221 GET_MODE (src), &sri);
1222 if (sclass == NO_REGS
1223 || ((secondary_class != NO_REGS || sri.icode != CODE_FOR_nothing)
1224 && dclass != NO_REGS))
1226 enum reg_class old_sclass = secondary_class;
1227 secondary_reload_info old_sri = sri;
1229 sri.prev_sri = NULL;
1230 sri.icode = CODE_FOR_nothing;
1231 sri.extra_cost = 0;
1232 secondary_class
1233 = (enum reg_class) targetm.secondary_reload (true, src,
1234 (reg_class_t) dclass,
1235 GET_MODE (src), &sri);
1236 /* Check the target hook consistency. */
1237 lra_assert
1238 ((secondary_class == NO_REGS && sri.icode == CODE_FOR_nothing)
1239 || (old_sclass == NO_REGS && old_sri.icode == CODE_FOR_nothing)
1240 || (secondary_class == old_sclass && sri.icode == old_sri.icode));
1242 if (sregno >= 0)
1243 reg_renumber [sregno] = -1;
1244 if (dregno >= 0)
1245 reg_renumber [dregno] = -1;
1246 if (secondary_class == NO_REGS && sri.icode == CODE_FOR_nothing)
1247 return false;
1248 *change_p = true;
1249 new_reg = NULL_RTX;
1250 if (secondary_class != NO_REGS)
1251 new_reg = lra_create_new_reg_with_unique_value (GET_MODE (src), NULL_RTX,
1252 secondary_class,
1253 "secondary");
1254 start_sequence ();
1255 if (sri.icode == CODE_FOR_nothing)
1256 lra_emit_move (new_reg, src);
1257 else
1259 enum reg_class scratch_class;
1261 scratch_class = (reg_class_from_constraints
1262 (insn_data[sri.icode].operand[2].constraint));
1263 scratch_reg = (lra_create_new_reg_with_unique_value
1264 (insn_data[sri.icode].operand[2].mode, NULL_RTX,
1265 scratch_class, "scratch"));
1266 emit_insn (GEN_FCN (sri.icode) (new_reg != NULL_RTX ? new_reg : dest,
1267 src, scratch_reg));
1269 before = get_insns ();
1270 end_sequence ();
1271 lra_process_new_insns (curr_insn, before, NULL, "Inserting the move");
1272 if (new_reg != NULL_RTX)
1273 SET_SRC (curr_insn_set) = new_reg;
1274 else
1276 if (lra_dump_file != NULL)
1278 fprintf (lra_dump_file, "Deleting move %u\n", INSN_UID (curr_insn));
1279 dump_insn_slim (lra_dump_file, curr_insn);
1281 lra_set_insn_deleted (curr_insn);
1282 return true;
1284 return false;
1287 /* The following data describe the result of process_alt_operands.
1288 The data are used in curr_insn_transform to generate reloads. */
1290 /* The chosen reg classes which should be used for the corresponding
1291 operands. */
1292 static enum reg_class goal_alt[MAX_RECOG_OPERANDS];
1293 /* True if the operand should be the same as another operand and that
1294 other operand does not need a reload. */
1295 static bool goal_alt_match_win[MAX_RECOG_OPERANDS];
1296 /* True if the operand does not need a reload. */
1297 static bool goal_alt_win[MAX_RECOG_OPERANDS];
1298 /* True if the operand can be offsetable memory. */
1299 static bool goal_alt_offmemok[MAX_RECOG_OPERANDS];
1300 /* The number of an operand to which given operand can be matched to. */
1301 static int goal_alt_matches[MAX_RECOG_OPERANDS];
1302 /* The number of elements in the following array. */
1303 static int goal_alt_dont_inherit_ops_num;
1304 /* Numbers of operands whose reload pseudos should not be inherited. */
1305 static int goal_alt_dont_inherit_ops[MAX_RECOG_OPERANDS];
1306 /* True if the insn commutative operands should be swapped. */
1307 static bool goal_alt_swapped;
1308 /* The chosen insn alternative. */
1309 static int goal_alt_number;
1311 /* True if the corresponding operand is the result of an equivalence
1312 substitution. */
1313 static bool equiv_substition_p[MAX_RECOG_OPERANDS];
1315 /* The following five variables are used to choose the best insn
1316 alternative. They reflect final characteristics of the best
1317 alternative. */
1319 /* Number of necessary reloads and overall cost reflecting the
1320 previous value and other unpleasantness of the best alternative. */
1321 static int best_losers, best_overall;
1322 /* Overall number hard registers used for reloads. For example, on
1323 some targets we need 2 general registers to reload DFmode and only
1324 one floating point register. */
1325 static int best_reload_nregs;
1326 /* Overall number reflecting distances of previous reloading the same
1327 value. The distances are counted from the current BB start. It is
1328 used to improve inheritance chances. */
1329 static int best_reload_sum;
1331 /* True if the current insn should have no correspondingly input or
1332 output reloads. */
1333 static bool no_input_reloads_p, no_output_reloads_p;
1335 /* True if we swapped the commutative operands in the current
1336 insn. */
1337 static int curr_swapped;
1339 /* if CHECK_ONLY_P is false, arrange for address element *LOC to be a
1340 register of class CL. Add any input reloads to list BEFORE. AFTER
1341 is nonnull if *LOC is an automodified value; handle that case by
1342 adding the required output reloads to list AFTER. Return true if
1343 the RTL was changed.
1345 if CHECK_ONLY_P is true, check that the *LOC is a correct address
1346 register. Return false if the address register is correct. */
1347 static bool
1348 process_addr_reg (rtx *loc, bool check_only_p, rtx_insn **before, rtx_insn **after,
1349 enum reg_class cl)
1351 int regno;
1352 enum reg_class rclass, new_class;
1353 rtx reg;
1354 rtx new_reg;
1355 machine_mode mode;
1356 bool subreg_p, before_p = false;
1358 subreg_p = GET_CODE (*loc) == SUBREG;
1359 if (subreg_p)
1361 reg = SUBREG_REG (*loc);
1362 mode = GET_MODE (reg);
1364 /* For mode with size bigger than ptr_mode, there unlikely to be "mov"
1365 between two registers with different classes, but there normally will
1366 be "mov" which transfers element of vector register into the general
1367 register, and this normally will be a subreg which should be reloaded
1368 as a whole. This is particularly likely to be triggered when
1369 -fno-split-wide-types specified. */
1370 if (!REG_P (reg)
1371 || in_class_p (reg, cl, &new_class)
1372 || known_le (GET_MODE_SIZE (mode), GET_MODE_SIZE (ptr_mode)))
1373 loc = &SUBREG_REG (*loc);
1376 reg = *loc;
1377 mode = GET_MODE (reg);
1378 if (! REG_P (reg))
1380 if (check_only_p)
1381 return true;
1382 /* Always reload memory in an address even if the target supports
1383 such addresses. */
1384 new_reg = lra_create_new_reg_with_unique_value (mode, reg, cl, "address");
1385 before_p = true;
1387 else
1389 regno = REGNO (reg);
1390 rclass = get_reg_class (regno);
1391 if (! check_only_p
1392 && (*loc = get_equiv_with_elimination (reg, curr_insn)) != reg)
1394 if (lra_dump_file != NULL)
1396 fprintf (lra_dump_file,
1397 "Changing pseudo %d in address of insn %u on equiv ",
1398 REGNO (reg), INSN_UID (curr_insn));
1399 dump_value_slim (lra_dump_file, *loc, 1);
1400 fprintf (lra_dump_file, "\n");
1402 *loc = copy_rtx (*loc);
1404 if (*loc != reg || ! in_class_p (reg, cl, &new_class))
1406 if (check_only_p)
1407 return true;
1408 reg = *loc;
1409 if (get_reload_reg (after == NULL ? OP_IN : OP_INOUT,
1410 mode, reg, cl, subreg_p, "address", &new_reg))
1411 before_p = true;
1413 else if (new_class != NO_REGS && rclass != new_class)
1415 if (check_only_p)
1416 return true;
1417 lra_change_class (regno, new_class, " Change to", true);
1418 return false;
1420 else
1421 return false;
1423 if (before_p)
1425 push_to_sequence (*before);
1426 lra_emit_move (new_reg, reg);
1427 *before = get_insns ();
1428 end_sequence ();
1430 *loc = new_reg;
1431 if (after != NULL)
1433 start_sequence ();
1434 lra_emit_move (before_p ? copy_rtx (reg) : reg, new_reg);
1435 emit_insn (*after);
1436 *after = get_insns ();
1437 end_sequence ();
1439 return true;
1442 /* Insert move insn in simplify_operand_subreg. BEFORE returns
1443 the insn to be inserted before curr insn. AFTER returns the
1444 the insn to be inserted after curr insn. ORIGREG and NEWREG
1445 are the original reg and new reg for reload. */
1446 static void
1447 insert_move_for_subreg (rtx_insn **before, rtx_insn **after, rtx origreg,
1448 rtx newreg)
1450 if (before)
1452 push_to_sequence (*before);
1453 lra_emit_move (newreg, origreg);
1454 *before = get_insns ();
1455 end_sequence ();
1457 if (after)
1459 start_sequence ();
1460 lra_emit_move (origreg, newreg);
1461 emit_insn (*after);
1462 *after = get_insns ();
1463 end_sequence ();
1467 static int valid_address_p (machine_mode mode, rtx addr, addr_space_t as);
1468 static bool process_address (int, bool, rtx_insn **, rtx_insn **);
1470 /* Make reloads for subreg in operand NOP with internal subreg mode
1471 REG_MODE, add new reloads for further processing. Return true if
1472 any change was done. */
1473 static bool
1474 simplify_operand_subreg (int nop, machine_mode reg_mode)
1476 int hard_regno;
1477 rtx_insn *before, *after;
1478 machine_mode mode, innermode;
1479 rtx reg, new_reg;
1480 rtx operand = *curr_id->operand_loc[nop];
1481 enum reg_class regclass;
1482 enum op_type type;
1484 before = after = NULL;
1486 if (GET_CODE (operand) != SUBREG)
1487 return false;
1489 mode = GET_MODE (operand);
1490 reg = SUBREG_REG (operand);
1491 innermode = GET_MODE (reg);
1492 type = curr_static_id->operand[nop].type;
1493 if (MEM_P (reg))
1495 const bool addr_was_valid
1496 = valid_address_p (innermode, XEXP (reg, 0), MEM_ADDR_SPACE (reg));
1497 alter_subreg (curr_id->operand_loc[nop], false);
1498 rtx subst = *curr_id->operand_loc[nop];
1499 lra_assert (MEM_P (subst));
1500 const bool addr_is_valid = valid_address_p (GET_MODE (subst),
1501 XEXP (subst, 0),
1502 MEM_ADDR_SPACE (subst));
1503 if (!addr_was_valid
1504 || addr_is_valid
1505 || ((get_constraint_type (lookup_constraint
1506 (curr_static_id->operand[nop].constraint))
1507 != CT_SPECIAL_MEMORY)
1508 /* We still can reload address and if the address is
1509 valid, we can remove subreg without reloading its
1510 inner memory. */
1511 && valid_address_p (GET_MODE (subst),
1512 regno_reg_rtx
1513 [ira_class_hard_regs
1514 [base_reg_class (GET_MODE (subst),
1515 MEM_ADDR_SPACE (subst),
1516 ADDRESS, SCRATCH)][0]],
1517 MEM_ADDR_SPACE (subst))))
1519 /* If we change the address for a paradoxical subreg of memory, the
1520 new address might violate the necessary alignment or the access
1521 might be slow; take this into consideration. We need not worry
1522 about accesses beyond allocated memory for paradoxical memory
1523 subregs as we don't substitute such equiv memory (see processing
1524 equivalences in function lra_constraints) and because for spilled
1525 pseudos we allocate stack memory enough for the biggest
1526 corresponding paradoxical subreg.
1528 However, do not blindly simplify a (subreg (mem ...)) for
1529 WORD_REGISTER_OPERATIONS targets as this may lead to loading junk
1530 data into a register when the inner is narrower than outer or
1531 missing important data from memory when the inner is wider than
1532 outer. This rule only applies to modes that are no wider than
1533 a word.
1535 If valid memory becomes invalid after subreg elimination
1536 and address might be different we still have to reload
1537 memory.
1539 if ((! addr_was_valid
1540 || addr_is_valid
1541 || known_eq (GET_MODE_SIZE (mode), GET_MODE_SIZE (innermode)))
1542 && !(maybe_ne (GET_MODE_PRECISION (mode),
1543 GET_MODE_PRECISION (innermode))
1544 && known_le (GET_MODE_SIZE (mode), UNITS_PER_WORD)
1545 && known_le (GET_MODE_SIZE (innermode), UNITS_PER_WORD)
1546 && WORD_REGISTER_OPERATIONS)
1547 && (!(MEM_ALIGN (subst) < GET_MODE_ALIGNMENT (mode)
1548 && targetm.slow_unaligned_access (mode, MEM_ALIGN (subst)))
1549 || (MEM_ALIGN (reg) < GET_MODE_ALIGNMENT (innermode)
1550 && targetm.slow_unaligned_access (innermode,
1551 MEM_ALIGN (reg)))))
1552 return true;
1554 *curr_id->operand_loc[nop] = operand;
1556 /* But if the address was not valid, we cannot reload the MEM without
1557 reloading the address first. */
1558 if (!addr_was_valid)
1559 process_address (nop, false, &before, &after);
1561 /* INNERMODE is fast, MODE slow. Reload the mem in INNERMODE. */
1562 enum reg_class rclass
1563 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1564 if (get_reload_reg (curr_static_id->operand[nop].type, innermode,
1565 reg, rclass, TRUE, "slow/invalid mem", &new_reg))
1567 bool insert_before, insert_after;
1568 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1570 insert_before = (type != OP_OUT
1571 || partial_subreg_p (mode, innermode));
1572 insert_after = type != OP_IN;
1573 insert_move_for_subreg (insert_before ? &before : NULL,
1574 insert_after ? &after : NULL,
1575 reg, new_reg);
1577 SUBREG_REG (operand) = new_reg;
1579 /* Convert to MODE. */
1580 reg = operand;
1581 rclass
1582 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1583 if (get_reload_reg (curr_static_id->operand[nop].type, mode, reg,
1584 rclass, TRUE, "slow/invalid mem", &new_reg))
1586 bool insert_before, insert_after;
1587 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1589 insert_before = type != OP_OUT;
1590 insert_after = type != OP_IN;
1591 insert_move_for_subreg (insert_before ? &before : NULL,
1592 insert_after ? &after : NULL,
1593 reg, new_reg);
1595 *curr_id->operand_loc[nop] = new_reg;
1596 lra_process_new_insns (curr_insn, before, after,
1597 "Inserting slow/invalid mem reload");
1598 return true;
1601 /* If the address was valid and became invalid, prefer to reload
1602 the memory. Typical case is when the index scale should
1603 correspond the memory. */
1604 *curr_id->operand_loc[nop] = operand;
1605 /* Do not return false here as the MEM_P (reg) will be processed
1606 later in this function. */
1608 else if (REG_P (reg) && REGNO (reg) < FIRST_PSEUDO_REGISTER)
1610 alter_subreg (curr_id->operand_loc[nop], false);
1611 return true;
1613 else if (CONSTANT_P (reg))
1615 /* Try to simplify subreg of constant. It is usually result of
1616 equivalence substitution. */
1617 if (innermode == VOIDmode
1618 && (innermode = original_subreg_reg_mode[nop]) == VOIDmode)
1619 innermode = curr_static_id->operand[nop].mode;
1620 if ((new_reg = simplify_subreg (mode, reg, innermode,
1621 SUBREG_BYTE (operand))) != NULL_RTX)
1623 *curr_id->operand_loc[nop] = new_reg;
1624 return true;
1627 /* Put constant into memory when we have mixed modes. It generates
1628 a better code in most cases as it does not need a secondary
1629 reload memory. It also prevents LRA looping when LRA is using
1630 secondary reload memory again and again. */
1631 if (CONSTANT_P (reg) && CONST_POOL_OK_P (reg_mode, reg)
1632 && SCALAR_INT_MODE_P (reg_mode) != SCALAR_INT_MODE_P (mode))
1634 SUBREG_REG (operand) = force_const_mem (reg_mode, reg);
1635 alter_subreg (curr_id->operand_loc[nop], false);
1636 return true;
1638 /* Force a reload of the SUBREG_REG if this is a constant or PLUS or
1639 if there may be a problem accessing OPERAND in the outer
1640 mode. */
1641 if ((REG_P (reg)
1642 && REGNO (reg) >= FIRST_PSEUDO_REGISTER
1643 && (hard_regno = lra_get_regno_hard_regno (REGNO (reg))) >= 0
1644 /* Don't reload paradoxical subregs because we could be looping
1645 having repeatedly final regno out of hard regs range. */
1646 && (hard_regno_nregs (hard_regno, innermode)
1647 >= hard_regno_nregs (hard_regno, mode))
1648 && simplify_subreg_regno (hard_regno, innermode,
1649 SUBREG_BYTE (operand), mode) < 0
1650 /* Don't reload subreg for matching reload. It is actually
1651 valid subreg in LRA. */
1652 && ! LRA_SUBREG_P (operand))
1653 || CONSTANT_P (reg) || GET_CODE (reg) == PLUS || MEM_P (reg))
1655 enum reg_class rclass;
1657 if (REG_P (reg))
1658 /* There is a big probability that we will get the same class
1659 for the new pseudo and we will get the same insn which
1660 means infinite looping. So spill the new pseudo. */
1661 rclass = NO_REGS;
1662 else
1663 /* The class will be defined later in curr_insn_transform. */
1664 rclass
1665 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1667 if (get_reload_reg (curr_static_id->operand[nop].type, reg_mode, reg,
1668 rclass, TRUE, "subreg reg", &new_reg))
1670 bool insert_before, insert_after;
1671 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1673 insert_before = (type != OP_OUT
1674 || read_modify_subreg_p (operand));
1675 insert_after = (type != OP_IN);
1676 insert_move_for_subreg (insert_before ? &before : NULL,
1677 insert_after ? &after : NULL,
1678 reg, new_reg);
1680 SUBREG_REG (operand) = new_reg;
1681 lra_process_new_insns (curr_insn, before, after,
1682 "Inserting subreg reload");
1683 return true;
1685 /* Force a reload for a paradoxical subreg. For paradoxical subreg,
1686 IRA allocates hardreg to the inner pseudo reg according to its mode
1687 instead of the outermode, so the size of the hardreg may not be enough
1688 to contain the outermode operand, in that case we may need to insert
1689 reload for the reg. For the following two types of paradoxical subreg,
1690 we need to insert reload:
1691 1. If the op_type is OP_IN, and the hardreg could not be paired with
1692 other hardreg to contain the outermode operand
1693 (checked by in_hard_reg_set_p), we need to insert the reload.
1694 2. If the op_type is OP_OUT or OP_INOUT.
1696 Here is a paradoxical subreg example showing how the reload is generated:
1698 (insn 5 4 7 2 (set (reg:TI 106 [ __comp ])
1699 (subreg:TI (reg:DI 107 [ __comp ]) 0)) {*movti_internal_rex64}
1701 In IRA, reg107 is allocated to a DImode hardreg. We use x86-64 as example
1702 here, if reg107 is assigned to hardreg R15, because R15 is the last
1703 hardreg, compiler cannot find another hardreg to pair with R15 to
1704 contain TImode data. So we insert a TImode reload reg180 for it.
1705 After reload is inserted:
1707 (insn 283 0 0 (set (subreg:DI (reg:TI 180 [orig:107 __comp ] [107]) 0)
1708 (reg:DI 107 [ __comp ])) -1
1709 (insn 5 4 7 2 (set (reg:TI 106 [ __comp ])
1710 (subreg:TI (reg:TI 180 [orig:107 __comp ] [107]) 0)) {*movti_internal_rex64}
1712 Two reload hard registers will be allocated to reg180 to save TImode data
1713 in LRA_assign.
1715 For LRA pseudos this should normally be handled by the biggest_mode
1716 mechanism. However, it's possible for new uses of an LRA pseudo
1717 to be introduced after we've allocated it, such as when undoing
1718 inheritance, and the allocated register might not then be appropriate
1719 for the new uses. */
1720 else if (REG_P (reg)
1721 && REGNO (reg) >= FIRST_PSEUDO_REGISTER
1722 && (hard_regno = lra_get_regno_hard_regno (REGNO (reg))) >= 0
1723 && (hard_regno_nregs (hard_regno, innermode)
1724 < hard_regno_nregs (hard_regno, mode))
1725 && (regclass = lra_get_allocno_class (REGNO (reg)))
1726 && (type != OP_IN
1727 || !in_hard_reg_set_p (reg_class_contents[regclass],
1728 mode, hard_regno)
1729 || overlaps_hard_reg_set_p (lra_no_alloc_regs,
1730 mode, hard_regno)))
1732 /* The class will be defined later in curr_insn_transform. */
1733 enum reg_class rclass
1734 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1736 if (get_reload_reg (curr_static_id->operand[nop].type, mode, reg,
1737 rclass, TRUE, "paradoxical subreg", &new_reg))
1739 rtx subreg;
1740 bool insert_before, insert_after;
1742 PUT_MODE (new_reg, mode);
1743 subreg = gen_lowpart_SUBREG (innermode, new_reg);
1744 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1746 insert_before = (type != OP_OUT);
1747 insert_after = (type != OP_IN);
1748 insert_move_for_subreg (insert_before ? &before : NULL,
1749 insert_after ? &after : NULL,
1750 reg, subreg);
1752 SUBREG_REG (operand) = new_reg;
1753 lra_process_new_insns (curr_insn, before, after,
1754 "Inserting paradoxical subreg reload");
1755 return true;
1757 return false;
1760 /* Return TRUE if X refers for a hard register from SET. */
1761 static bool
1762 uses_hard_regs_p (rtx x, HARD_REG_SET set)
1764 int i, j, x_hard_regno;
1765 machine_mode mode;
1766 const char *fmt;
1767 enum rtx_code code;
1769 if (x == NULL_RTX)
1770 return false;
1771 code = GET_CODE (x);
1772 mode = GET_MODE (x);
1774 if (code == SUBREG)
1776 /* For all SUBREGs we want to check whether the full multi-register
1777 overlaps the set. For normal SUBREGs this means 'get_hard_regno' of
1778 the inner register, for paradoxical SUBREGs this means the
1779 'get_hard_regno' of the full SUBREG and for complete SUBREGs either is
1780 fine. Use the wider mode for all cases. */
1781 rtx subreg = SUBREG_REG (x);
1782 mode = wider_subreg_mode (x);
1783 if (mode == GET_MODE (subreg))
1785 x = subreg;
1786 code = GET_CODE (x);
1790 if (REG_P (x) || SUBREG_P (x))
1792 x_hard_regno = get_hard_regno (x, true);
1793 return (x_hard_regno >= 0
1794 && overlaps_hard_reg_set_p (set, mode, x_hard_regno));
1796 if (MEM_P (x))
1798 struct address_info ad;
1800 decompose_mem_address (&ad, x);
1801 if (ad.base_term != NULL && uses_hard_regs_p (*ad.base_term, set))
1802 return true;
1803 if (ad.index_term != NULL && uses_hard_regs_p (*ad.index_term, set))
1804 return true;
1806 fmt = GET_RTX_FORMAT (code);
1807 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1809 if (fmt[i] == 'e')
1811 if (uses_hard_regs_p (XEXP (x, i), set))
1812 return true;
1814 else if (fmt[i] == 'E')
1816 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
1817 if (uses_hard_regs_p (XVECEXP (x, i, j), set))
1818 return true;
1821 return false;
1824 /* Return true if OP is a spilled pseudo. */
1825 static inline bool
1826 spilled_pseudo_p (rtx op)
1828 return (REG_P (op)
1829 && REGNO (op) >= FIRST_PSEUDO_REGISTER && in_mem_p (REGNO (op)));
1832 /* Return true if X is a general constant. */
1833 static inline bool
1834 general_constant_p (rtx x)
1836 return CONSTANT_P (x) && (! flag_pic || LEGITIMATE_PIC_OPERAND_P (x));
1839 static bool
1840 reg_in_class_p (rtx reg, enum reg_class cl)
1842 if (cl == NO_REGS)
1843 return get_reg_class (REGNO (reg)) == NO_REGS;
1844 return in_class_p (reg, cl, NULL);
1847 /* Return true if SET of RCLASS contains no hard regs which can be
1848 used in MODE. */
1849 static bool
1850 prohibited_class_reg_set_mode_p (enum reg_class rclass,
1851 HARD_REG_SET &set,
1852 machine_mode mode)
1854 HARD_REG_SET temp;
1856 lra_assert (hard_reg_set_subset_p (reg_class_contents[rclass], set));
1857 temp = set & ~lra_no_alloc_regs;
1858 return (hard_reg_set_subset_p
1859 (temp, ira_prohibited_class_mode_regs[rclass][mode]));
1863 /* Used to check validity info about small class input operands. It
1864 should be incremented at start of processing an insn
1865 alternative. */
1866 static unsigned int curr_small_class_check = 0;
1868 /* Update number of used inputs of class OP_CLASS for operand NOP
1869 of alternative NALT. Return true if we have more such class operands
1870 than the number of available regs. */
1871 static bool
1872 update_and_check_small_class_inputs (int nop, int nalt,
1873 enum reg_class op_class)
1875 static unsigned int small_class_check[LIM_REG_CLASSES];
1876 static int small_class_input_nums[LIM_REG_CLASSES];
1878 if (SMALL_REGISTER_CLASS_P (op_class)
1879 /* We are interesting in classes became small because of fixing
1880 some hard regs, e.g. by an user through GCC options. */
1881 && hard_reg_set_intersect_p (reg_class_contents[op_class],
1882 ira_no_alloc_regs)
1883 && (curr_static_id->operand[nop].type != OP_OUT
1884 || TEST_BIT (curr_static_id->operand[nop].early_clobber_alts, nalt)))
1886 if (small_class_check[op_class] == curr_small_class_check)
1887 small_class_input_nums[op_class]++;
1888 else
1890 small_class_check[op_class] = curr_small_class_check;
1891 small_class_input_nums[op_class] = 1;
1893 if (small_class_input_nums[op_class] > ira_class_hard_regs_num[op_class])
1894 return true;
1896 return false;
1899 /* Major function to choose the current insn alternative and what
1900 operands should be reloaded and how. If ONLY_ALTERNATIVE is not
1901 negative we should consider only this alternative. Return false if
1902 we cannot choose the alternative or find how to reload the
1903 operands. */
1904 static bool
1905 process_alt_operands (int only_alternative)
1907 bool ok_p = false;
1908 int nop, overall, nalt;
1909 int n_alternatives = curr_static_id->n_alternatives;
1910 int n_operands = curr_static_id->n_operands;
1911 /* LOSERS counts the operands that don't fit this alternative and
1912 would require loading. */
1913 int losers;
1914 int addr_losers;
1915 /* REJECT is a count of how undesirable this alternative says it is
1916 if any reloading is required. If the alternative matches exactly
1917 then REJECT is ignored, but otherwise it gets this much counted
1918 against it in addition to the reloading needed. */
1919 int reject;
1920 /* This is defined by '!' or '?' alternative constraint and added to
1921 reject. But in some cases it can be ignored. */
1922 int static_reject;
1923 int op_reject;
1924 /* The number of elements in the following array. */
1925 int early_clobbered_regs_num;
1926 /* Numbers of operands which are early clobber registers. */
1927 int early_clobbered_nops[MAX_RECOG_OPERANDS];
1928 enum reg_class curr_alt[MAX_RECOG_OPERANDS];
1929 HARD_REG_SET curr_alt_set[MAX_RECOG_OPERANDS];
1930 bool curr_alt_match_win[MAX_RECOG_OPERANDS];
1931 bool curr_alt_win[MAX_RECOG_OPERANDS];
1932 bool curr_alt_offmemok[MAX_RECOG_OPERANDS];
1933 int curr_alt_matches[MAX_RECOG_OPERANDS];
1934 /* The number of elements in the following array. */
1935 int curr_alt_dont_inherit_ops_num;
1936 /* Numbers of operands whose reload pseudos should not be inherited. */
1937 int curr_alt_dont_inherit_ops[MAX_RECOG_OPERANDS];
1938 rtx op;
1939 /* The register when the operand is a subreg of register, otherwise the
1940 operand itself. */
1941 rtx no_subreg_reg_operand[MAX_RECOG_OPERANDS];
1942 /* The register if the operand is a register or subreg of register,
1943 otherwise NULL. */
1944 rtx operand_reg[MAX_RECOG_OPERANDS];
1945 int hard_regno[MAX_RECOG_OPERANDS];
1946 machine_mode biggest_mode[MAX_RECOG_OPERANDS];
1947 int reload_nregs, reload_sum;
1948 bool costly_p;
1949 enum reg_class cl;
1951 /* Calculate some data common for all alternatives to speed up the
1952 function. */
1953 for (nop = 0; nop < n_operands; nop++)
1955 rtx reg;
1957 op = no_subreg_reg_operand[nop] = *curr_id->operand_loc[nop];
1958 /* The real hard regno of the operand after the allocation. */
1959 hard_regno[nop] = get_hard_regno (op, true);
1961 operand_reg[nop] = reg = op;
1962 biggest_mode[nop] = GET_MODE (op);
1963 if (GET_CODE (op) == SUBREG)
1965 biggest_mode[nop] = wider_subreg_mode (op);
1966 operand_reg[nop] = reg = SUBREG_REG (op);
1968 if (! REG_P (reg))
1969 operand_reg[nop] = NULL_RTX;
1970 else if (REGNO (reg) >= FIRST_PSEUDO_REGISTER
1971 || ((int) REGNO (reg)
1972 == lra_get_elimination_hard_regno (REGNO (reg))))
1973 no_subreg_reg_operand[nop] = reg;
1974 else
1975 operand_reg[nop] = no_subreg_reg_operand[nop]
1976 /* Just use natural mode for elimination result. It should
1977 be enough for extra constraints hooks. */
1978 = regno_reg_rtx[hard_regno[nop]];
1981 /* The constraints are made of several alternatives. Each operand's
1982 constraint looks like foo,bar,... with commas separating the
1983 alternatives. The first alternatives for all operands go
1984 together, the second alternatives go together, etc.
1986 First loop over alternatives. */
1987 alternative_mask preferred = curr_id->preferred_alternatives;
1988 if (only_alternative >= 0)
1989 preferred &= ALTERNATIVE_BIT (only_alternative);
1991 for (nalt = 0; nalt < n_alternatives; nalt++)
1993 /* Loop over operands for one constraint alternative. */
1994 if (!TEST_BIT (preferred, nalt))
1995 continue;
1997 bool matching_early_clobber[MAX_RECOG_OPERANDS];
1998 curr_small_class_check++;
1999 overall = losers = addr_losers = 0;
2000 static_reject = reject = reload_nregs = reload_sum = 0;
2001 for (nop = 0; nop < n_operands; nop++)
2003 int inc = (curr_static_id
2004 ->operand_alternative[nalt * n_operands + nop].reject);
2005 if (lra_dump_file != NULL && inc != 0)
2006 fprintf (lra_dump_file,
2007 " Staticly defined alt reject+=%d\n", inc);
2008 static_reject += inc;
2009 matching_early_clobber[nop] = 0;
2011 reject += static_reject;
2012 early_clobbered_regs_num = 0;
2014 for (nop = 0; nop < n_operands; nop++)
2016 const char *p;
2017 char *end;
2018 int len, c, m, i, opalt_num, this_alternative_matches;
2019 bool win, did_match, offmemok, early_clobber_p;
2020 /* false => this operand can be reloaded somehow for this
2021 alternative. */
2022 bool badop;
2023 /* true => this operand can be reloaded if the alternative
2024 allows regs. */
2025 bool winreg;
2026 /* True if a constant forced into memory would be OK for
2027 this operand. */
2028 bool constmemok;
2029 enum reg_class this_alternative, this_costly_alternative;
2030 HARD_REG_SET this_alternative_set, this_costly_alternative_set;
2031 bool this_alternative_match_win, this_alternative_win;
2032 bool this_alternative_offmemok;
2033 bool scratch_p;
2034 machine_mode mode;
2035 enum constraint_num cn;
2037 opalt_num = nalt * n_operands + nop;
2038 if (curr_static_id->operand_alternative[opalt_num].anything_ok)
2040 /* Fast track for no constraints at all. */
2041 curr_alt[nop] = NO_REGS;
2042 CLEAR_HARD_REG_SET (curr_alt_set[nop]);
2043 curr_alt_win[nop] = true;
2044 curr_alt_match_win[nop] = false;
2045 curr_alt_offmemok[nop] = false;
2046 curr_alt_matches[nop] = -1;
2047 continue;
2050 op = no_subreg_reg_operand[nop];
2051 mode = curr_operand_mode[nop];
2053 win = did_match = winreg = offmemok = constmemok = false;
2054 badop = true;
2056 early_clobber_p = false;
2057 p = curr_static_id->operand_alternative[opalt_num].constraint;
2059 this_costly_alternative = this_alternative = NO_REGS;
2060 /* We update set of possible hard regs besides its class
2061 because reg class might be inaccurate. For example,
2062 union of LO_REGS (l), HI_REGS(h), and STACK_REG(k) in ARM
2063 is translated in HI_REGS because classes are merged by
2064 pairs and there is no accurate intermediate class. */
2065 CLEAR_HARD_REG_SET (this_alternative_set);
2066 CLEAR_HARD_REG_SET (this_costly_alternative_set);
2067 this_alternative_win = false;
2068 this_alternative_match_win = false;
2069 this_alternative_offmemok = false;
2070 this_alternative_matches = -1;
2072 /* An empty constraint should be excluded by the fast
2073 track. */
2074 lra_assert (*p != 0 && *p != ',');
2076 op_reject = 0;
2077 /* Scan this alternative's specs for this operand; set WIN
2078 if the operand fits any letter in this alternative.
2079 Otherwise, clear BADOP if this operand could fit some
2080 letter after reloads, or set WINREG if this operand could
2081 fit after reloads provided the constraint allows some
2082 registers. */
2083 costly_p = false;
2086 switch ((c = *p, len = CONSTRAINT_LEN (c, p)), c)
2088 case '\0':
2089 len = 0;
2090 break;
2091 case ',':
2092 c = '\0';
2093 break;
2095 case '&':
2096 early_clobber_p = true;
2097 break;
2099 case '$':
2100 op_reject += LRA_MAX_REJECT;
2101 break;
2102 case '^':
2103 op_reject += LRA_LOSER_COST_FACTOR;
2104 break;
2106 case '#':
2107 /* Ignore rest of this alternative. */
2108 c = '\0';
2109 break;
2111 case '0': case '1': case '2': case '3': case '4':
2112 case '5': case '6': case '7': case '8': case '9':
2114 int m_hregno;
2115 bool match_p;
2117 m = strtoul (p, &end, 10);
2118 p = end;
2119 len = 0;
2120 lra_assert (nop > m);
2122 /* Reject matches if we don't know which operand is
2123 bigger. This situation would arguably be a bug in
2124 an .md pattern, but could also occur in a user asm. */
2125 if (!ordered_p (GET_MODE_SIZE (biggest_mode[m]),
2126 GET_MODE_SIZE (biggest_mode[nop])))
2127 break;
2129 /* Don't match wrong asm insn operands for proper
2130 diagnostic later. */
2131 if (INSN_CODE (curr_insn) < 0
2132 && (curr_operand_mode[m] == BLKmode
2133 || curr_operand_mode[nop] == BLKmode)
2134 && curr_operand_mode[m] != curr_operand_mode[nop])
2135 break;
2137 m_hregno = get_hard_regno (*curr_id->operand_loc[m], false);
2138 /* We are supposed to match a previous operand.
2139 If we do, we win if that one did. If we do
2140 not, count both of the operands as losers.
2141 (This is too conservative, since most of the
2142 time only a single reload insn will be needed
2143 to make the two operands win. As a result,
2144 this alternative may be rejected when it is
2145 actually desirable.) */
2146 match_p = false;
2147 if (operands_match_p (*curr_id->operand_loc[nop],
2148 *curr_id->operand_loc[m], m_hregno))
2150 /* We should reject matching of an early
2151 clobber operand if the matching operand is
2152 not dying in the insn. */
2153 if (!TEST_BIT (curr_static_id->operand[m]
2154 .early_clobber_alts, nalt)
2155 || operand_reg[nop] == NULL_RTX
2156 || (find_regno_note (curr_insn, REG_DEAD,
2157 REGNO (op))
2158 || REGNO (op) == REGNO (operand_reg[m])))
2159 match_p = true;
2161 if (match_p)
2163 /* If we are matching a non-offsettable
2164 address where an offsettable address was
2165 expected, then we must reject this
2166 combination, because we can't reload
2167 it. */
2168 if (curr_alt_offmemok[m]
2169 && MEM_P (*curr_id->operand_loc[m])
2170 && curr_alt[m] == NO_REGS && ! curr_alt_win[m])
2171 continue;
2173 else
2175 /* If the operands do not match and one
2176 operand is INOUT, we can not match them.
2177 Try other possibilities, e.g. other
2178 alternatives or commutative operand
2179 exchange. */
2180 if (curr_static_id->operand[nop].type == OP_INOUT
2181 || curr_static_id->operand[m].type == OP_INOUT)
2182 break;
2183 /* Operands don't match. If the operands are
2184 different user defined explicit hard
2185 registers, then we cannot make them match
2186 when one is early clobber operand. */
2187 if ((REG_P (*curr_id->operand_loc[nop])
2188 || SUBREG_P (*curr_id->operand_loc[nop]))
2189 && (REG_P (*curr_id->operand_loc[m])
2190 || SUBREG_P (*curr_id->operand_loc[m])))
2192 rtx nop_reg = *curr_id->operand_loc[nop];
2193 if (SUBREG_P (nop_reg))
2194 nop_reg = SUBREG_REG (nop_reg);
2195 rtx m_reg = *curr_id->operand_loc[m];
2196 if (SUBREG_P (m_reg))
2197 m_reg = SUBREG_REG (m_reg);
2199 if (REG_P (nop_reg)
2200 && HARD_REGISTER_P (nop_reg)
2201 && REG_USERVAR_P (nop_reg)
2202 && REG_P (m_reg)
2203 && HARD_REGISTER_P (m_reg)
2204 && REG_USERVAR_P (m_reg))
2206 int i;
2208 for (i = 0; i < early_clobbered_regs_num; i++)
2209 if (m == early_clobbered_nops[i])
2210 break;
2211 if (i < early_clobbered_regs_num
2212 || early_clobber_p)
2213 break;
2216 /* Both operands must allow a reload register,
2217 otherwise we cannot make them match. */
2218 if (curr_alt[m] == NO_REGS)
2219 break;
2220 /* Retroactively mark the operand we had to
2221 match as a loser, if it wasn't already and
2222 it wasn't matched to a register constraint
2223 (e.g it might be matched by memory). */
2224 if (curr_alt_win[m]
2225 && (operand_reg[m] == NULL_RTX
2226 || hard_regno[m] < 0))
2228 losers++;
2229 reload_nregs
2230 += (ira_reg_class_max_nregs[curr_alt[m]]
2231 [GET_MODE (*curr_id->operand_loc[m])]);
2234 /* Prefer matching earlyclobber alternative as
2235 it results in less hard regs required for
2236 the insn than a non-matching earlyclobber
2237 alternative. */
2238 if (TEST_BIT (curr_static_id->operand[m]
2239 .early_clobber_alts, nalt))
2241 if (lra_dump_file != NULL)
2242 fprintf
2243 (lra_dump_file,
2244 " %d Matching earlyclobber alt:"
2245 " reject--\n",
2246 nop);
2247 if (!matching_early_clobber[m])
2249 reject--;
2250 matching_early_clobber[m] = 1;
2253 /* Otherwise we prefer no matching
2254 alternatives because it gives more freedom
2255 in RA. */
2256 else if (operand_reg[nop] == NULL_RTX
2257 || (find_regno_note (curr_insn, REG_DEAD,
2258 REGNO (operand_reg[nop]))
2259 == NULL_RTX))
2261 if (lra_dump_file != NULL)
2262 fprintf
2263 (lra_dump_file,
2264 " %d Matching alt: reject+=2\n",
2265 nop);
2266 reject += 2;
2269 /* If we have to reload this operand and some
2270 previous operand also had to match the same
2271 thing as this operand, we don't know how to do
2272 that. */
2273 if (!match_p || !curr_alt_win[m])
2275 for (i = 0; i < nop; i++)
2276 if (curr_alt_matches[i] == m)
2277 break;
2278 if (i < nop)
2279 break;
2281 else
2282 did_match = true;
2284 this_alternative_matches = m;
2285 /* This can be fixed with reloads if the operand
2286 we are supposed to match can be fixed with
2287 reloads. */
2288 badop = false;
2289 this_alternative = curr_alt[m];
2290 this_alternative_set = curr_alt_set[m];
2291 winreg = this_alternative != NO_REGS;
2292 break;
2295 case 'g':
2296 if (MEM_P (op)
2297 || general_constant_p (op)
2298 || spilled_pseudo_p (op))
2299 win = true;
2300 cl = GENERAL_REGS;
2301 goto reg;
2303 default:
2304 cn = lookup_constraint (p);
2305 switch (get_constraint_type (cn))
2307 case CT_REGISTER:
2308 cl = reg_class_for_constraint (cn);
2309 if (cl != NO_REGS)
2310 goto reg;
2311 break;
2313 case CT_CONST_INT:
2314 if (CONST_INT_P (op)
2315 && insn_const_int_ok_for_constraint (INTVAL (op), cn))
2316 win = true;
2317 break;
2319 case CT_MEMORY:
2320 if (MEM_P (op)
2321 && satisfies_memory_constraint_p (op, cn))
2322 win = true;
2323 else if (spilled_pseudo_p (op))
2324 win = true;
2326 /* If we didn't already win, we can reload constants
2327 via force_const_mem or put the pseudo value into
2328 memory, or make other memory by reloading the
2329 address like for 'o'. */
2330 if (CONST_POOL_OK_P (mode, op)
2331 || MEM_P (op) || REG_P (op)
2332 /* We can restore the equiv insn by a
2333 reload. */
2334 || equiv_substition_p[nop])
2335 badop = false;
2336 constmemok = true;
2337 offmemok = true;
2338 break;
2340 case CT_ADDRESS:
2341 /* An asm operand with an address constraint
2342 that doesn't satisfy address_operand has
2343 is_address cleared, so that we don't try to
2344 make a non-address fit. */
2345 if (!curr_static_id->operand[nop].is_address)
2346 break;
2347 /* If we didn't already win, we can reload the address
2348 into a base register. */
2349 if (satisfies_address_constraint_p (op, cn))
2350 win = true;
2351 cl = base_reg_class (VOIDmode, ADDR_SPACE_GENERIC,
2352 ADDRESS, SCRATCH);
2353 badop = false;
2354 goto reg;
2356 case CT_FIXED_FORM:
2357 if (constraint_satisfied_p (op, cn))
2358 win = true;
2359 break;
2361 case CT_SPECIAL_MEMORY:
2362 if (MEM_P (op)
2363 && satisfies_memory_constraint_p (op, cn))
2364 win = true;
2365 else if (spilled_pseudo_p (op))
2366 win = true;
2367 break;
2369 break;
2371 reg:
2372 if (mode == BLKmode)
2373 break;
2374 this_alternative = reg_class_subunion[this_alternative][cl];
2375 this_alternative_set |= reg_class_contents[cl];
2376 if (costly_p)
2378 this_costly_alternative
2379 = reg_class_subunion[this_costly_alternative][cl];
2380 this_costly_alternative_set |= reg_class_contents[cl];
2382 winreg = true;
2383 if (REG_P (op))
2385 if (hard_regno[nop] >= 0
2386 && in_hard_reg_set_p (this_alternative_set,
2387 mode, hard_regno[nop]))
2388 win = true;
2389 else if (hard_regno[nop] < 0
2390 && in_class_p (op, this_alternative, NULL))
2391 win = true;
2393 break;
2395 if (c != ' ' && c != '\t')
2396 costly_p = c == '*';
2398 while ((p += len), c);
2400 scratch_p = (operand_reg[nop] != NULL_RTX
2401 && lra_former_scratch_p (REGNO (operand_reg[nop])));
2402 /* Record which operands fit this alternative. */
2403 if (win)
2405 this_alternative_win = true;
2406 if (operand_reg[nop] != NULL_RTX)
2408 if (hard_regno[nop] >= 0)
2410 if (in_hard_reg_set_p (this_costly_alternative_set,
2411 mode, hard_regno[nop]))
2413 if (lra_dump_file != NULL)
2414 fprintf (lra_dump_file,
2415 " %d Costly set: reject++\n",
2416 nop);
2417 reject++;
2420 else
2422 /* Prefer won reg to spilled pseudo under other
2423 equal conditions for possibe inheritance. */
2424 if (! scratch_p)
2426 if (lra_dump_file != NULL)
2427 fprintf
2428 (lra_dump_file,
2429 " %d Non pseudo reload: reject++\n",
2430 nop);
2431 reject++;
2433 if (in_class_p (operand_reg[nop],
2434 this_costly_alternative, NULL))
2436 if (lra_dump_file != NULL)
2437 fprintf
2438 (lra_dump_file,
2439 " %d Non pseudo costly reload:"
2440 " reject++\n",
2441 nop);
2442 reject++;
2445 /* We simulate the behavior of old reload here.
2446 Although scratches need hard registers and it
2447 might result in spilling other pseudos, no reload
2448 insns are generated for the scratches. So it
2449 might cost something but probably less than old
2450 reload pass believes. */
2451 if (scratch_p)
2453 if (lra_dump_file != NULL)
2454 fprintf (lra_dump_file,
2455 " %d Scratch win: reject+=2\n",
2456 nop);
2457 reject += 2;
2461 else if (did_match)
2462 this_alternative_match_win = true;
2463 else
2465 int const_to_mem = 0;
2466 bool no_regs_p;
2468 reject += op_reject;
2469 /* Never do output reload of stack pointer. It makes
2470 impossible to do elimination when SP is changed in
2471 RTL. */
2472 if (op == stack_pointer_rtx && ! frame_pointer_needed
2473 && curr_static_id->operand[nop].type != OP_IN)
2474 goto fail;
2476 /* If this alternative asks for a specific reg class, see if there
2477 is at least one allocatable register in that class. */
2478 no_regs_p
2479 = (this_alternative == NO_REGS
2480 || (hard_reg_set_subset_p
2481 (reg_class_contents[this_alternative],
2482 lra_no_alloc_regs)));
2484 /* For asms, verify that the class for this alternative is possible
2485 for the mode that is specified. */
2486 if (!no_regs_p && INSN_CODE (curr_insn) < 0)
2488 int i;
2489 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
2490 if (targetm.hard_regno_mode_ok (i, mode)
2491 && in_hard_reg_set_p (reg_class_contents[this_alternative],
2492 mode, i))
2493 break;
2494 if (i == FIRST_PSEUDO_REGISTER)
2495 winreg = false;
2498 /* If this operand accepts a register, and if the
2499 register class has at least one allocatable register,
2500 then this operand can be reloaded. */
2501 if (winreg && !no_regs_p)
2502 badop = false;
2504 if (badop)
2506 if (lra_dump_file != NULL)
2507 fprintf (lra_dump_file,
2508 " alt=%d: Bad operand -- refuse\n",
2509 nalt);
2510 goto fail;
2513 if (this_alternative != NO_REGS)
2515 HARD_REG_SET available_regs
2516 = (reg_class_contents[this_alternative]
2517 & ~((ira_prohibited_class_mode_regs
2518 [this_alternative][mode])
2519 | lra_no_alloc_regs));
2520 if (hard_reg_set_empty_p (available_regs))
2522 /* There are no hard regs holding a value of given
2523 mode. */
2524 if (offmemok)
2526 this_alternative = NO_REGS;
2527 if (lra_dump_file != NULL)
2528 fprintf (lra_dump_file,
2529 " %d Using memory because of"
2530 " a bad mode: reject+=2\n",
2531 nop);
2532 reject += 2;
2534 else
2536 if (lra_dump_file != NULL)
2537 fprintf (lra_dump_file,
2538 " alt=%d: Wrong mode -- refuse\n",
2539 nalt);
2540 goto fail;
2545 /* If not assigned pseudo has a class which a subset of
2546 required reg class, it is a less costly alternative
2547 as the pseudo still can get a hard reg of necessary
2548 class. */
2549 if (! no_regs_p && REG_P (op) && hard_regno[nop] < 0
2550 && (cl = get_reg_class (REGNO (op))) != NO_REGS
2551 && ira_class_subset_p[this_alternative][cl])
2553 if (lra_dump_file != NULL)
2554 fprintf
2555 (lra_dump_file,
2556 " %d Super set class reg: reject-=3\n", nop);
2557 reject -= 3;
2560 this_alternative_offmemok = offmemok;
2561 if (this_costly_alternative != NO_REGS)
2563 if (lra_dump_file != NULL)
2564 fprintf (lra_dump_file,
2565 " %d Costly loser: reject++\n", nop);
2566 reject++;
2568 /* If the operand is dying, has a matching constraint,
2569 and satisfies constraints of the matched operand
2570 which failed to satisfy the own constraints, most probably
2571 the reload for this operand will be gone. */
2572 if (this_alternative_matches >= 0
2573 && !curr_alt_win[this_alternative_matches]
2574 && REG_P (op)
2575 && find_regno_note (curr_insn, REG_DEAD, REGNO (op))
2576 && (hard_regno[nop] >= 0
2577 ? in_hard_reg_set_p (this_alternative_set,
2578 mode, hard_regno[nop])
2579 : in_class_p (op, this_alternative, NULL)))
2581 if (lra_dump_file != NULL)
2582 fprintf
2583 (lra_dump_file,
2584 " %d Dying matched operand reload: reject++\n",
2585 nop);
2586 reject++;
2588 else
2590 /* Strict_low_part requires to reload the register
2591 not the sub-register. In this case we should
2592 check that a final reload hard reg can hold the
2593 value mode. */
2594 if (curr_static_id->operand[nop].strict_low
2595 && REG_P (op)
2596 && hard_regno[nop] < 0
2597 && GET_CODE (*curr_id->operand_loc[nop]) == SUBREG
2598 && ira_class_hard_regs_num[this_alternative] > 0
2599 && (!targetm.hard_regno_mode_ok
2600 (ira_class_hard_regs[this_alternative][0],
2601 GET_MODE (*curr_id->operand_loc[nop]))))
2603 if (lra_dump_file != NULL)
2604 fprintf
2605 (lra_dump_file,
2606 " alt=%d: Strict low subreg reload -- refuse\n",
2607 nalt);
2608 goto fail;
2610 losers++;
2612 if (operand_reg[nop] != NULL_RTX
2613 /* Output operands and matched input operands are
2614 not inherited. The following conditions do not
2615 exactly describe the previous statement but they
2616 are pretty close. */
2617 && curr_static_id->operand[nop].type != OP_OUT
2618 && (this_alternative_matches < 0
2619 || curr_static_id->operand[nop].type != OP_IN))
2621 int last_reload = (lra_reg_info[ORIGINAL_REGNO
2622 (operand_reg[nop])]
2623 .last_reload);
2625 /* The value of reload_sum has sense only if we
2626 process insns in their order. It happens only on
2627 the first constraints sub-pass when we do most of
2628 reload work. */
2629 if (lra_constraint_iter == 1 && last_reload > bb_reload_num)
2630 reload_sum += last_reload - bb_reload_num;
2632 /* If this is a constant that is reloaded into the
2633 desired class by copying it to memory first, count
2634 that as another reload. This is consistent with
2635 other code and is required to avoid choosing another
2636 alternative when the constant is moved into memory.
2637 Note that the test here is precisely the same as in
2638 the code below that calls force_const_mem. */
2639 if (CONST_POOL_OK_P (mode, op)
2640 && ((targetm.preferred_reload_class
2641 (op, this_alternative) == NO_REGS)
2642 || no_input_reloads_p))
2644 const_to_mem = 1;
2645 if (! no_regs_p)
2646 losers++;
2649 /* Alternative loses if it requires a type of reload not
2650 permitted for this insn. We can always reload
2651 objects with a REG_UNUSED note. */
2652 if ((curr_static_id->operand[nop].type != OP_IN
2653 && no_output_reloads_p
2654 && ! find_reg_note (curr_insn, REG_UNUSED, op))
2655 || (curr_static_id->operand[nop].type != OP_OUT
2656 && no_input_reloads_p && ! const_to_mem)
2657 || (this_alternative_matches >= 0
2658 && (no_input_reloads_p
2659 || (no_output_reloads_p
2660 && (curr_static_id->operand
2661 [this_alternative_matches].type != OP_IN)
2662 && ! find_reg_note (curr_insn, REG_UNUSED,
2663 no_subreg_reg_operand
2664 [this_alternative_matches])))))
2666 if (lra_dump_file != NULL)
2667 fprintf
2668 (lra_dump_file,
2669 " alt=%d: No input/otput reload -- refuse\n",
2670 nalt);
2671 goto fail;
2674 /* Alternative loses if it required class pseudo cannot
2675 hold value of required mode. Such insns can be
2676 described by insn definitions with mode iterators. */
2677 if (GET_MODE (*curr_id->operand_loc[nop]) != VOIDmode
2678 && ! hard_reg_set_empty_p (this_alternative_set)
2679 /* It is common practice for constraints to use a
2680 class which does not have actually enough regs to
2681 hold the value (e.g. x86 AREG for mode requiring
2682 more one general reg). Therefore we have 2
2683 conditions to check that the reload pseudo cannot
2684 hold the mode value. */
2685 && (!targetm.hard_regno_mode_ok
2686 (ira_class_hard_regs[this_alternative][0],
2687 GET_MODE (*curr_id->operand_loc[nop])))
2688 /* The above condition is not enough as the first
2689 reg in ira_class_hard_regs can be not aligned for
2690 multi-words mode values. */
2691 && (prohibited_class_reg_set_mode_p
2692 (this_alternative, this_alternative_set,
2693 GET_MODE (*curr_id->operand_loc[nop]))))
2695 if (lra_dump_file != NULL)
2696 fprintf (lra_dump_file,
2697 " alt=%d: reload pseudo for op %d "
2698 "cannot hold the mode value -- refuse\n",
2699 nalt, nop);
2700 goto fail;
2703 /* Check strong discouragement of reload of non-constant
2704 into class THIS_ALTERNATIVE. */
2705 if (! CONSTANT_P (op) && ! no_regs_p
2706 && (targetm.preferred_reload_class
2707 (op, this_alternative) == NO_REGS
2708 || (curr_static_id->operand[nop].type == OP_OUT
2709 && (targetm.preferred_output_reload_class
2710 (op, this_alternative) == NO_REGS))))
2712 if (lra_dump_file != NULL)
2713 fprintf (lra_dump_file,
2714 " %d Non-prefered reload: reject+=%d\n",
2715 nop, LRA_MAX_REJECT);
2716 reject += LRA_MAX_REJECT;
2719 if (! (MEM_P (op) && offmemok)
2720 && ! (const_to_mem && constmemok))
2722 /* We prefer to reload pseudos over reloading other
2723 things, since such reloads may be able to be
2724 eliminated later. So bump REJECT in other cases.
2725 Don't do this in the case where we are forcing a
2726 constant into memory and it will then win since
2727 we don't want to have a different alternative
2728 match then. */
2729 if (! (REG_P (op) && REGNO (op) >= FIRST_PSEUDO_REGISTER))
2731 if (lra_dump_file != NULL)
2732 fprintf
2733 (lra_dump_file,
2734 " %d Non-pseudo reload: reject+=2\n",
2735 nop);
2736 reject += 2;
2739 if (! no_regs_p)
2740 reload_nregs
2741 += ira_reg_class_max_nregs[this_alternative][mode];
2743 if (SMALL_REGISTER_CLASS_P (this_alternative))
2745 if (lra_dump_file != NULL)
2746 fprintf
2747 (lra_dump_file,
2748 " %d Small class reload: reject+=%d\n",
2749 nop, LRA_LOSER_COST_FACTOR / 2);
2750 reject += LRA_LOSER_COST_FACTOR / 2;
2754 /* We are trying to spill pseudo into memory. It is
2755 usually more costly than moving to a hard register
2756 although it might takes the same number of
2757 reloads.
2759 Non-pseudo spill may happen also. Suppose a target allows both
2760 register and memory in the operand constraint alternatives,
2761 then it's typical that an eliminable register has a substition
2762 of "base + offset" which can either be reloaded by a simple
2763 "new_reg <= base + offset" which will match the register
2764 constraint, or a similar reg addition followed by further spill
2765 to and reload from memory which will match the memory
2766 constraint, but this memory spill will be much more costly
2767 usually.
2769 Code below increases the reject for both pseudo and non-pseudo
2770 spill. */
2771 if (no_regs_p
2772 && !(MEM_P (op) && offmemok)
2773 && !(REG_P (op) && hard_regno[nop] < 0))
2775 if (lra_dump_file != NULL)
2776 fprintf
2777 (lra_dump_file,
2778 " %d Spill %spseudo into memory: reject+=3\n",
2779 nop, REG_P (op) ? "" : "Non-");
2780 reject += 3;
2781 if (VECTOR_MODE_P (mode))
2783 /* Spilling vectors into memory is usually more
2784 costly as they contain big values. */
2785 if (lra_dump_file != NULL)
2786 fprintf
2787 (lra_dump_file,
2788 " %d Spill vector pseudo: reject+=2\n",
2789 nop);
2790 reject += 2;
2794 /* When we use an operand requiring memory in given
2795 alternative, the insn should write *and* read the
2796 value to/from memory it is costly in comparison with
2797 an insn alternative which does not use memory
2798 (e.g. register or immediate operand). We exclude
2799 memory operand for such case as we can satisfy the
2800 memory constraints by reloading address. */
2801 if (no_regs_p && offmemok && !MEM_P (op))
2803 if (lra_dump_file != NULL)
2804 fprintf
2805 (lra_dump_file,
2806 " Using memory insn operand %d: reject+=3\n",
2807 nop);
2808 reject += 3;
2811 /* If reload requires moving value through secondary
2812 memory, it will need one more insn at least. */
2813 if (this_alternative != NO_REGS
2814 && REG_P (op) && (cl = get_reg_class (REGNO (op))) != NO_REGS
2815 && ((curr_static_id->operand[nop].type != OP_OUT
2816 && targetm.secondary_memory_needed (GET_MODE (op), cl,
2817 this_alternative))
2818 || (curr_static_id->operand[nop].type != OP_IN
2819 && (targetm.secondary_memory_needed
2820 (GET_MODE (op), this_alternative, cl)))))
2821 losers++;
2823 if (MEM_P (op) && offmemok)
2824 addr_losers++;
2825 else
2827 /* Input reloads can be inherited more often than
2828 output reloads can be removed, so penalize output
2829 reloads. */
2830 if (!REG_P (op) || curr_static_id->operand[nop].type != OP_IN)
2832 if (lra_dump_file != NULL)
2833 fprintf
2834 (lra_dump_file,
2835 " %d Non input pseudo reload: reject++\n",
2836 nop);
2837 reject++;
2840 if (curr_static_id->operand[nop].type == OP_INOUT)
2842 if (lra_dump_file != NULL)
2843 fprintf
2844 (lra_dump_file,
2845 " %d Input/Output reload: reject+=%d\n",
2846 nop, LRA_LOSER_COST_FACTOR);
2847 reject += LRA_LOSER_COST_FACTOR;
2852 if (early_clobber_p && ! scratch_p)
2854 if (lra_dump_file != NULL)
2855 fprintf (lra_dump_file,
2856 " %d Early clobber: reject++\n", nop);
2857 reject++;
2859 /* ??? We check early clobbers after processing all operands
2860 (see loop below) and there we update the costs more.
2861 Should we update the cost (may be approximately) here
2862 because of early clobber register reloads or it is a rare
2863 or non-important thing to be worth to do it. */
2864 overall = (losers * LRA_LOSER_COST_FACTOR + reject
2865 - (addr_losers == losers ? static_reject : 0));
2866 if ((best_losers == 0 || losers != 0) && best_overall < overall)
2868 if (lra_dump_file != NULL)
2869 fprintf (lra_dump_file,
2870 " alt=%d,overall=%d,losers=%d -- refuse\n",
2871 nalt, overall, losers);
2872 goto fail;
2875 if (update_and_check_small_class_inputs (nop, nalt,
2876 this_alternative))
2878 if (lra_dump_file != NULL)
2879 fprintf (lra_dump_file,
2880 " alt=%d, not enough small class regs -- refuse\n",
2881 nalt);
2882 goto fail;
2884 curr_alt[nop] = this_alternative;
2885 curr_alt_set[nop] = this_alternative_set;
2886 curr_alt_win[nop] = this_alternative_win;
2887 curr_alt_match_win[nop] = this_alternative_match_win;
2888 curr_alt_offmemok[nop] = this_alternative_offmemok;
2889 curr_alt_matches[nop] = this_alternative_matches;
2891 if (this_alternative_matches >= 0
2892 && !did_match && !this_alternative_win)
2893 curr_alt_win[this_alternative_matches] = false;
2895 if (early_clobber_p && operand_reg[nop] != NULL_RTX)
2896 early_clobbered_nops[early_clobbered_regs_num++] = nop;
2899 if (curr_insn_set != NULL_RTX && n_operands == 2
2900 /* Prevent processing non-move insns. */
2901 && (GET_CODE (SET_SRC (curr_insn_set)) == SUBREG
2902 || SET_SRC (curr_insn_set) == no_subreg_reg_operand[1])
2903 && ((! curr_alt_win[0] && ! curr_alt_win[1]
2904 && REG_P (no_subreg_reg_operand[0])
2905 && REG_P (no_subreg_reg_operand[1])
2906 && (reg_in_class_p (no_subreg_reg_operand[0], curr_alt[1])
2907 || reg_in_class_p (no_subreg_reg_operand[1], curr_alt[0])))
2908 || (! curr_alt_win[0] && curr_alt_win[1]
2909 && REG_P (no_subreg_reg_operand[1])
2910 /* Check that we reload memory not the memory
2911 address. */
2912 && ! (curr_alt_offmemok[0]
2913 && MEM_P (no_subreg_reg_operand[0]))
2914 && reg_in_class_p (no_subreg_reg_operand[1], curr_alt[0]))
2915 || (curr_alt_win[0] && ! curr_alt_win[1]
2916 && REG_P (no_subreg_reg_operand[0])
2917 /* Check that we reload memory not the memory
2918 address. */
2919 && ! (curr_alt_offmemok[1]
2920 && MEM_P (no_subreg_reg_operand[1]))
2921 && reg_in_class_p (no_subreg_reg_operand[0], curr_alt[1])
2922 && (! CONST_POOL_OK_P (curr_operand_mode[1],
2923 no_subreg_reg_operand[1])
2924 || (targetm.preferred_reload_class
2925 (no_subreg_reg_operand[1],
2926 (enum reg_class) curr_alt[1]) != NO_REGS))
2927 /* If it is a result of recent elimination in move
2928 insn we can transform it into an add still by
2929 using this alternative. */
2930 && GET_CODE (no_subreg_reg_operand[1]) != PLUS
2931 /* Likewise if the source has been replaced with an
2932 equivalent value. This only happens once -- the reload
2933 will use the equivalent value instead of the register it
2934 replaces -- so there should be no danger of cycling. */
2935 && !equiv_substition_p[1])))
2937 /* We have a move insn and a new reload insn will be similar
2938 to the current insn. We should avoid such situation as
2939 it results in LRA cycling. */
2940 if (lra_dump_file != NULL)
2941 fprintf (lra_dump_file,
2942 " Cycle danger: overall += LRA_MAX_REJECT\n");
2943 overall += LRA_MAX_REJECT;
2945 ok_p = true;
2946 curr_alt_dont_inherit_ops_num = 0;
2947 for (nop = 0; nop < early_clobbered_regs_num; nop++)
2949 int i, j, clobbered_hard_regno, first_conflict_j, last_conflict_j;
2950 HARD_REG_SET temp_set;
2952 i = early_clobbered_nops[nop];
2953 if ((! curr_alt_win[i] && ! curr_alt_match_win[i])
2954 || hard_regno[i] < 0)
2955 continue;
2956 lra_assert (operand_reg[i] != NULL_RTX);
2957 clobbered_hard_regno = hard_regno[i];
2958 CLEAR_HARD_REG_SET (temp_set);
2959 add_to_hard_reg_set (&temp_set, biggest_mode[i], clobbered_hard_regno);
2960 first_conflict_j = last_conflict_j = -1;
2961 for (j = 0; j < n_operands; j++)
2962 if (j == i
2963 /* We don't want process insides of match_operator and
2964 match_parallel because otherwise we would process
2965 their operands once again generating a wrong
2966 code. */
2967 || curr_static_id->operand[j].is_operator)
2968 continue;
2969 else if ((curr_alt_matches[j] == i && curr_alt_match_win[j])
2970 || (curr_alt_matches[i] == j && curr_alt_match_win[i]))
2971 continue;
2972 /* If we don't reload j-th operand, check conflicts. */
2973 else if ((curr_alt_win[j] || curr_alt_match_win[j])
2974 && uses_hard_regs_p (*curr_id->operand_loc[j], temp_set))
2976 if (first_conflict_j < 0)
2977 first_conflict_j = j;
2978 last_conflict_j = j;
2979 /* Both the earlyclobber operand and conflicting operand
2980 cannot both be user defined hard registers. */
2981 if (HARD_REGISTER_P (operand_reg[i])
2982 && REG_USERVAR_P (operand_reg[i])
2983 && operand_reg[j] != NULL_RTX
2984 && HARD_REGISTER_P (operand_reg[j])
2985 && REG_USERVAR_P (operand_reg[j]))
2986 fatal_insn ("unable to generate reloads for "
2987 "impossible constraints:", curr_insn);
2989 if (last_conflict_j < 0)
2990 continue;
2992 /* If an earlyclobber operand conflicts with another non-matching
2993 operand (ie, they have been assigned the same hard register),
2994 then it is better to reload the other operand, as there may
2995 exist yet another operand with a matching constraint associated
2996 with the earlyclobber operand. However, if one of the operands
2997 is an explicit use of a hard register, then we must reload the
2998 other non-hard register operand. */
2999 if (HARD_REGISTER_P (operand_reg[i])
3000 || (first_conflict_j == last_conflict_j
3001 && operand_reg[last_conflict_j] != NULL_RTX
3002 && !curr_alt_match_win[last_conflict_j]
3003 && !HARD_REGISTER_P (operand_reg[last_conflict_j])))
3005 curr_alt_win[last_conflict_j] = false;
3006 curr_alt_dont_inherit_ops[curr_alt_dont_inherit_ops_num++]
3007 = last_conflict_j;
3008 losers++;
3009 if (lra_dump_file != NULL)
3010 fprintf
3011 (lra_dump_file,
3012 " %d Conflict early clobber reload: reject--\n",
3015 else
3017 /* We need to reload early clobbered register and the
3018 matched registers. */
3019 for (j = 0; j < n_operands; j++)
3020 if (curr_alt_matches[j] == i)
3022 curr_alt_match_win[j] = false;
3023 losers++;
3024 overall += LRA_LOSER_COST_FACTOR;
3026 if (! curr_alt_match_win[i])
3027 curr_alt_dont_inherit_ops[curr_alt_dont_inherit_ops_num++] = i;
3028 else
3030 /* Remember pseudos used for match reloads are never
3031 inherited. */
3032 lra_assert (curr_alt_matches[i] >= 0);
3033 curr_alt_win[curr_alt_matches[i]] = false;
3035 curr_alt_win[i] = curr_alt_match_win[i] = false;
3036 losers++;
3037 if (lra_dump_file != NULL)
3038 fprintf
3039 (lra_dump_file,
3040 " %d Matched conflict early clobber reloads: "
3041 "reject--\n",
3044 /* Early clobber was already reflected in REJECT. */
3045 if (!matching_early_clobber[i])
3047 lra_assert (reject > 0);
3048 reject--;
3049 matching_early_clobber[i] = 1;
3051 overall += LRA_LOSER_COST_FACTOR - 1;
3053 if (lra_dump_file != NULL)
3054 fprintf (lra_dump_file, " alt=%d,overall=%d,losers=%d,rld_nregs=%d\n",
3055 nalt, overall, losers, reload_nregs);
3057 /* If this alternative can be made to work by reloading, and it
3058 needs less reloading than the others checked so far, record
3059 it as the chosen goal for reloading. */
3060 if ((best_losers != 0 && losers == 0)
3061 || (((best_losers == 0 && losers == 0)
3062 || (best_losers != 0 && losers != 0))
3063 && (best_overall > overall
3064 || (best_overall == overall
3065 /* If the cost of the reloads is the same,
3066 prefer alternative which requires minimal
3067 number of reload regs. */
3068 && (reload_nregs < best_reload_nregs
3069 || (reload_nregs == best_reload_nregs
3070 && (best_reload_sum < reload_sum
3071 || (best_reload_sum == reload_sum
3072 && nalt < goal_alt_number))))))))
3074 for (nop = 0; nop < n_operands; nop++)
3076 goal_alt_win[nop] = curr_alt_win[nop];
3077 goal_alt_match_win[nop] = curr_alt_match_win[nop];
3078 goal_alt_matches[nop] = curr_alt_matches[nop];
3079 goal_alt[nop] = curr_alt[nop];
3080 goal_alt_offmemok[nop] = curr_alt_offmemok[nop];
3082 goal_alt_dont_inherit_ops_num = curr_alt_dont_inherit_ops_num;
3083 for (nop = 0; nop < curr_alt_dont_inherit_ops_num; nop++)
3084 goal_alt_dont_inherit_ops[nop] = curr_alt_dont_inherit_ops[nop];
3085 goal_alt_swapped = curr_swapped;
3086 best_overall = overall;
3087 best_losers = losers;
3088 best_reload_nregs = reload_nregs;
3089 best_reload_sum = reload_sum;
3090 goal_alt_number = nalt;
3092 if (losers == 0)
3093 /* Everything is satisfied. Do not process alternatives
3094 anymore. */
3095 break;
3096 fail:
3099 return ok_p;
3102 /* Make reload base reg from address AD. */
3103 static rtx
3104 base_to_reg (struct address_info *ad)
3106 enum reg_class cl;
3107 int code = -1;
3108 rtx new_inner = NULL_RTX;
3109 rtx new_reg = NULL_RTX;
3110 rtx_insn *insn;
3111 rtx_insn *last_insn = get_last_insn();
3113 lra_assert (ad->disp == ad->disp_term);
3114 cl = base_reg_class (ad->mode, ad->as, ad->base_outer_code,
3115 get_index_code (ad));
3116 new_reg = lra_create_new_reg (GET_MODE (*ad->base), NULL_RTX,
3117 cl, "base");
3118 new_inner = simplify_gen_binary (PLUS, GET_MODE (new_reg), new_reg,
3119 ad->disp_term == NULL
3120 ? const0_rtx
3121 : *ad->disp_term);
3122 if (!valid_address_p (ad->mode, new_inner, ad->as))
3123 return NULL_RTX;
3124 insn = emit_insn (gen_rtx_SET (new_reg, *ad->base));
3125 code = recog_memoized (insn);
3126 if (code < 0)
3128 delete_insns_since (last_insn);
3129 return NULL_RTX;
3132 return new_inner;
3135 /* Make reload base reg + DISP from address AD. Return the new pseudo. */
3136 static rtx
3137 base_plus_disp_to_reg (struct address_info *ad, rtx disp)
3139 enum reg_class cl;
3140 rtx new_reg;
3142 lra_assert (ad->base == ad->base_term);
3143 cl = base_reg_class (ad->mode, ad->as, ad->base_outer_code,
3144 get_index_code (ad));
3145 new_reg = lra_create_new_reg (GET_MODE (*ad->base_term), NULL_RTX,
3146 cl, "base + disp");
3147 lra_emit_add (new_reg, *ad->base_term, disp);
3148 return new_reg;
3151 /* Make reload of index part of address AD. Return the new
3152 pseudo. */
3153 static rtx
3154 index_part_to_reg (struct address_info *ad)
3156 rtx new_reg;
3158 new_reg = lra_create_new_reg (GET_MODE (*ad->index), NULL_RTX,
3159 INDEX_REG_CLASS, "index term");
3160 expand_mult (GET_MODE (*ad->index), *ad->index_term,
3161 GEN_INT (get_index_scale (ad)), new_reg, 1);
3162 return new_reg;
3165 /* Return true if we can add a displacement to address AD, even if that
3166 makes the address invalid. The fix-up code requires any new address
3167 to be the sum of the BASE_TERM, INDEX and DISP_TERM fields. */
3168 static bool
3169 can_add_disp_p (struct address_info *ad)
3171 return (!ad->autoinc_p
3172 && ad->segment == NULL
3173 && ad->base == ad->base_term
3174 && ad->disp == ad->disp_term);
3177 /* Make equiv substitution in address AD. Return true if a substitution
3178 was made. */
3179 static bool
3180 equiv_address_substitution (struct address_info *ad)
3182 rtx base_reg, new_base_reg, index_reg, new_index_reg, *base_term, *index_term;
3183 poly_int64 disp;
3184 HOST_WIDE_INT scale;
3185 bool change_p;
3187 base_term = strip_subreg (ad->base_term);
3188 if (base_term == NULL)
3189 base_reg = new_base_reg = NULL_RTX;
3190 else
3192 base_reg = *base_term;
3193 new_base_reg = get_equiv_with_elimination (base_reg, curr_insn);
3195 index_term = strip_subreg (ad->index_term);
3196 if (index_term == NULL)
3197 index_reg = new_index_reg = NULL_RTX;
3198 else
3200 index_reg = *index_term;
3201 new_index_reg = get_equiv_with_elimination (index_reg, curr_insn);
3203 if (base_reg == new_base_reg && index_reg == new_index_reg)
3204 return false;
3205 disp = 0;
3206 change_p = false;
3207 if (lra_dump_file != NULL)
3209 fprintf (lra_dump_file, "Changing address in insn %d ",
3210 INSN_UID (curr_insn));
3211 dump_value_slim (lra_dump_file, *ad->outer, 1);
3213 if (base_reg != new_base_reg)
3215 poly_int64 offset;
3216 if (REG_P (new_base_reg))
3218 *base_term = new_base_reg;
3219 change_p = true;
3221 else if (GET_CODE (new_base_reg) == PLUS
3222 && REG_P (XEXP (new_base_reg, 0))
3223 && poly_int_rtx_p (XEXP (new_base_reg, 1), &offset)
3224 && can_add_disp_p (ad))
3226 disp += offset;
3227 *base_term = XEXP (new_base_reg, 0);
3228 change_p = true;
3230 if (ad->base_term2 != NULL)
3231 *ad->base_term2 = *ad->base_term;
3233 if (index_reg != new_index_reg)
3235 poly_int64 offset;
3236 if (REG_P (new_index_reg))
3238 *index_term = new_index_reg;
3239 change_p = true;
3241 else if (GET_CODE (new_index_reg) == PLUS
3242 && REG_P (XEXP (new_index_reg, 0))
3243 && poly_int_rtx_p (XEXP (new_index_reg, 1), &offset)
3244 && can_add_disp_p (ad)
3245 && (scale = get_index_scale (ad)))
3247 disp += offset * scale;
3248 *index_term = XEXP (new_index_reg, 0);
3249 change_p = true;
3252 if (maybe_ne (disp, 0))
3254 if (ad->disp != NULL)
3255 *ad->disp = plus_constant (GET_MODE (*ad->inner), *ad->disp, disp);
3256 else
3258 *ad->inner = plus_constant (GET_MODE (*ad->inner), *ad->inner, disp);
3259 update_address (ad);
3261 change_p = true;
3263 if (lra_dump_file != NULL)
3265 if (! change_p)
3266 fprintf (lra_dump_file, " -- no change\n");
3267 else
3269 fprintf (lra_dump_file, " on equiv ");
3270 dump_value_slim (lra_dump_file, *ad->outer, 1);
3271 fprintf (lra_dump_file, "\n");
3274 return change_p;
3277 /* Major function to make reloads for an address in operand NOP or
3278 check its correctness (If CHECK_ONLY_P is true). The supported
3279 cases are:
3281 1) an address that existed before LRA started, at which point it
3282 must have been valid. These addresses are subject to elimination
3283 and may have become invalid due to the elimination offset being out
3284 of range.
3286 2) an address created by forcing a constant to memory
3287 (force_const_to_mem). The initial form of these addresses might
3288 not be valid, and it is this function's job to make them valid.
3290 3) a frame address formed from a register and a (possibly zero)
3291 constant offset. As above, these addresses might not be valid and
3292 this function must make them so.
3294 Add reloads to the lists *BEFORE and *AFTER. We might need to add
3295 reloads to *AFTER because of inc/dec, {pre, post} modify in the
3296 address. Return true for any RTL change.
3298 The function is a helper function which does not produce all
3299 transformations (when CHECK_ONLY_P is false) which can be
3300 necessary. It does just basic steps. To do all necessary
3301 transformations use function process_address. */
3302 static bool
3303 process_address_1 (int nop, bool check_only_p,
3304 rtx_insn **before, rtx_insn **after)
3306 struct address_info ad;
3307 rtx new_reg;
3308 HOST_WIDE_INT scale;
3309 rtx op = *curr_id->operand_loc[nop];
3310 const char *constraint = curr_static_id->operand[nop].constraint;
3311 enum constraint_num cn = lookup_constraint (constraint);
3312 bool change_p = false;
3314 if (MEM_P (op)
3315 && GET_MODE (op) == BLKmode
3316 && GET_CODE (XEXP (op, 0)) == SCRATCH)
3317 return false;
3319 if (insn_extra_address_constraint (cn)
3320 /* When we find an asm operand with an address constraint that
3321 doesn't satisfy address_operand to begin with, we clear
3322 is_address, so that we don't try to make a non-address fit.
3323 If the asm statement got this far, it's because other
3324 constraints are available, and we'll use them, disregarding
3325 the unsatisfiable address ones. */
3326 && curr_static_id->operand[nop].is_address)
3327 decompose_lea_address (&ad, curr_id->operand_loc[nop]);
3328 /* Do not attempt to decompose arbitrary addresses generated by combine
3329 for asm operands with loose constraints, e.g 'X'. */
3330 else if (MEM_P (op)
3331 && !(INSN_CODE (curr_insn) < 0
3332 && get_constraint_type (cn) == CT_FIXED_FORM
3333 && constraint_satisfied_p (op, cn)))
3334 decompose_mem_address (&ad, op);
3335 else if (GET_CODE (op) == SUBREG
3336 && MEM_P (SUBREG_REG (op)))
3337 decompose_mem_address (&ad, SUBREG_REG (op));
3338 else
3339 return false;
3340 /* If INDEX_REG_CLASS is assigned to base_term already and isn't to
3341 index_term, swap them so to avoid assigning INDEX_REG_CLASS to both
3342 when INDEX_REG_CLASS is a single register class. */
3343 if (ad.base_term != NULL
3344 && ad.index_term != NULL
3345 && ira_class_hard_regs_num[INDEX_REG_CLASS] == 1
3346 && REG_P (*ad.base_term)
3347 && REG_P (*ad.index_term)
3348 && in_class_p (*ad.base_term, INDEX_REG_CLASS, NULL)
3349 && ! in_class_p (*ad.index_term, INDEX_REG_CLASS, NULL))
3351 std::swap (ad.base, ad.index);
3352 std::swap (ad.base_term, ad.index_term);
3354 if (! check_only_p)
3355 change_p = equiv_address_substitution (&ad);
3356 if (ad.base_term != NULL
3357 && (process_addr_reg
3358 (ad.base_term, check_only_p, before,
3359 (ad.autoinc_p
3360 && !(REG_P (*ad.base_term)
3361 && find_regno_note (curr_insn, REG_DEAD,
3362 REGNO (*ad.base_term)) != NULL_RTX)
3363 ? after : NULL),
3364 base_reg_class (ad.mode, ad.as, ad.base_outer_code,
3365 get_index_code (&ad)))))
3367 change_p = true;
3368 if (ad.base_term2 != NULL)
3369 *ad.base_term2 = *ad.base_term;
3371 if (ad.index_term != NULL
3372 && process_addr_reg (ad.index_term, check_only_p,
3373 before, NULL, INDEX_REG_CLASS))
3374 change_p = true;
3376 /* Target hooks sometimes don't treat extra-constraint addresses as
3377 legitimate address_operands, so handle them specially. */
3378 if (insn_extra_address_constraint (cn)
3379 && satisfies_address_constraint_p (&ad, cn))
3380 return change_p;
3382 if (check_only_p)
3383 return change_p;
3385 /* There are three cases where the shape of *AD.INNER may now be invalid:
3387 1) the original address was valid, but either elimination or
3388 equiv_address_substitution was applied and that made
3389 the address invalid.
3391 2) the address is an invalid symbolic address created by
3392 force_const_to_mem.
3394 3) the address is a frame address with an invalid offset.
3396 4) the address is a frame address with an invalid base.
3398 All these cases involve a non-autoinc address, so there is no
3399 point revalidating other types. */
3400 if (ad.autoinc_p || valid_address_p (&ad))
3401 return change_p;
3403 /* Any index existed before LRA started, so we can assume that the
3404 presence and shape of the index is valid. */
3405 push_to_sequence (*before);
3406 lra_assert (ad.disp == ad.disp_term);
3407 if (ad.base == NULL)
3409 if (ad.index == NULL)
3411 rtx_insn *insn;
3412 rtx_insn *last = get_last_insn ();
3413 int code = -1;
3414 enum reg_class cl = base_reg_class (ad.mode, ad.as,
3415 SCRATCH, SCRATCH);
3416 rtx addr = *ad.inner;
3418 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "addr");
3419 if (HAVE_lo_sum)
3421 /* addr => lo_sum (new_base, addr), case (2) above. */
3422 insn = emit_insn (gen_rtx_SET
3423 (new_reg,
3424 gen_rtx_HIGH (Pmode, copy_rtx (addr))));
3425 code = recog_memoized (insn);
3426 if (code >= 0)
3428 *ad.inner = gen_rtx_LO_SUM (Pmode, new_reg, addr);
3429 if (! valid_address_p (ad.mode, *ad.outer, ad.as))
3431 /* Try to put lo_sum into register. */
3432 insn = emit_insn (gen_rtx_SET
3433 (new_reg,
3434 gen_rtx_LO_SUM (Pmode, new_reg, addr)));
3435 code = recog_memoized (insn);
3436 if (code >= 0)
3438 *ad.inner = new_reg;
3439 if (! valid_address_p (ad.mode, *ad.outer, ad.as))
3441 *ad.inner = addr;
3442 code = -1;
3448 if (code < 0)
3449 delete_insns_since (last);
3452 if (code < 0)
3454 /* addr => new_base, case (2) above. */
3455 lra_emit_move (new_reg, addr);
3457 for (insn = last == NULL_RTX ? get_insns () : NEXT_INSN (last);
3458 insn != NULL_RTX;
3459 insn = NEXT_INSN (insn))
3460 if (recog_memoized (insn) < 0)
3461 break;
3462 if (insn != NULL_RTX)
3464 /* Do nothing if we cannot generate right insns.
3465 This is analogous to reload pass behavior. */
3466 delete_insns_since (last);
3467 end_sequence ();
3468 return false;
3470 *ad.inner = new_reg;
3473 else
3475 /* index * scale + disp => new base + index * scale,
3476 case (1) above. */
3477 enum reg_class cl = base_reg_class (ad.mode, ad.as, PLUS,
3478 GET_CODE (*ad.index));
3480 lra_assert (INDEX_REG_CLASS != NO_REGS);
3481 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "disp");
3482 lra_emit_move (new_reg, *ad.disp);
3483 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3484 new_reg, *ad.index);
3487 else if (ad.index == NULL)
3489 int regno;
3490 enum reg_class cl;
3491 rtx set;
3492 rtx_insn *insns, *last_insn;
3493 /* Try to reload base into register only if the base is invalid
3494 for the address but with valid offset, case (4) above. */
3495 start_sequence ();
3496 new_reg = base_to_reg (&ad);
3498 /* base + disp => new base, cases (1) and (3) above. */
3499 /* Another option would be to reload the displacement into an
3500 index register. However, postreload has code to optimize
3501 address reloads that have the same base and different
3502 displacements, so reloading into an index register would
3503 not necessarily be a win. */
3504 if (new_reg == NULL_RTX)
3506 /* See if the target can split the displacement into a
3507 legitimate new displacement from a local anchor. */
3508 gcc_assert (ad.disp == ad.disp_term);
3509 poly_int64 orig_offset;
3510 rtx offset1, offset2;
3511 if (poly_int_rtx_p (*ad.disp, &orig_offset)
3512 && targetm.legitimize_address_displacement (&offset1, &offset2,
3513 orig_offset,
3514 ad.mode))
3516 new_reg = base_plus_disp_to_reg (&ad, offset1);
3517 new_reg = gen_rtx_PLUS (GET_MODE (new_reg), new_reg, offset2);
3519 else
3520 new_reg = base_plus_disp_to_reg (&ad, *ad.disp);
3522 insns = get_insns ();
3523 last_insn = get_last_insn ();
3524 /* If we generated at least two insns, try last insn source as
3525 an address. If we succeed, we generate one less insn. */
3526 if (REG_P (new_reg)
3527 && last_insn != insns
3528 && (set = single_set (last_insn)) != NULL_RTX
3529 && GET_CODE (SET_SRC (set)) == PLUS
3530 && REG_P (XEXP (SET_SRC (set), 0))
3531 && CONSTANT_P (XEXP (SET_SRC (set), 1)))
3533 *ad.inner = SET_SRC (set);
3534 if (valid_address_p (ad.mode, *ad.outer, ad.as))
3536 *ad.base_term = XEXP (SET_SRC (set), 0);
3537 *ad.disp_term = XEXP (SET_SRC (set), 1);
3538 cl = base_reg_class (ad.mode, ad.as, ad.base_outer_code,
3539 get_index_code (&ad));
3540 regno = REGNO (*ad.base_term);
3541 if (regno >= FIRST_PSEUDO_REGISTER
3542 && cl != lra_get_allocno_class (regno))
3543 lra_change_class (regno, cl, " Change to", true);
3544 new_reg = SET_SRC (set);
3545 delete_insns_since (PREV_INSN (last_insn));
3548 end_sequence ();
3549 emit_insn (insns);
3550 *ad.inner = new_reg;
3552 else if (ad.disp_term != NULL)
3554 /* base + scale * index + disp => new base + scale * index,
3555 case (1) above. */
3556 gcc_assert (ad.disp == ad.disp_term);
3557 new_reg = base_plus_disp_to_reg (&ad, *ad.disp);
3558 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3559 new_reg, *ad.index);
3561 else if ((scale = get_index_scale (&ad)) == 1)
3563 /* The last transformation to one reg will be made in
3564 curr_insn_transform function. */
3565 end_sequence ();
3566 return false;
3568 else if (scale != 0)
3570 /* base + scale * index => base + new_reg,
3571 case (1) above.
3572 Index part of address may become invalid. For example, we
3573 changed pseudo on the equivalent memory and a subreg of the
3574 pseudo onto the memory of different mode for which the scale is
3575 prohibitted. */
3576 new_reg = index_part_to_reg (&ad);
3577 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3578 *ad.base_term, new_reg);
3580 else
3582 enum reg_class cl = base_reg_class (ad.mode, ad.as,
3583 SCRATCH, SCRATCH);
3584 rtx addr = *ad.inner;
3586 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "addr");
3587 /* addr => new_base. */
3588 lra_emit_move (new_reg, addr);
3589 *ad.inner = new_reg;
3591 *before = get_insns ();
3592 end_sequence ();
3593 return true;
3596 /* If CHECK_ONLY_P is false, do address reloads until it is necessary.
3597 Use process_address_1 as a helper function. Return true for any
3598 RTL changes.
3600 If CHECK_ONLY_P is true, just check address correctness. Return
3601 false if the address correct. */
3602 static bool
3603 process_address (int nop, bool check_only_p,
3604 rtx_insn **before, rtx_insn **after)
3606 bool res = false;
3608 while (process_address_1 (nop, check_only_p, before, after))
3610 if (check_only_p)
3611 return true;
3612 res = true;
3614 return res;
3617 /* Emit insns to reload VALUE into a new register. VALUE is an
3618 auto-increment or auto-decrement RTX whose operand is a register or
3619 memory location; so reloading involves incrementing that location.
3620 IN is either identical to VALUE, or some cheaper place to reload
3621 value being incremented/decremented from.
3623 INC_AMOUNT is the number to increment or decrement by (always
3624 positive and ignored for POST_MODIFY/PRE_MODIFY).
3626 Return pseudo containing the result. */
3627 static rtx
3628 emit_inc (enum reg_class new_rclass, rtx in, rtx value, poly_int64 inc_amount)
3630 /* REG or MEM to be copied and incremented. */
3631 rtx incloc = XEXP (value, 0);
3632 /* Nonzero if increment after copying. */
3633 int post = (GET_CODE (value) == POST_DEC || GET_CODE (value) == POST_INC
3634 || GET_CODE (value) == POST_MODIFY);
3635 rtx_insn *last;
3636 rtx inc;
3637 rtx_insn *add_insn;
3638 int code;
3639 rtx real_in = in == value ? incloc : in;
3640 rtx result;
3641 bool plus_p = true;
3643 if (GET_CODE (value) == PRE_MODIFY || GET_CODE (value) == POST_MODIFY)
3645 lra_assert (GET_CODE (XEXP (value, 1)) == PLUS
3646 || GET_CODE (XEXP (value, 1)) == MINUS);
3647 lra_assert (rtx_equal_p (XEXP (XEXP (value, 1), 0), XEXP (value, 0)));
3648 plus_p = GET_CODE (XEXP (value, 1)) == PLUS;
3649 inc = XEXP (XEXP (value, 1), 1);
3651 else
3653 if (GET_CODE (value) == PRE_DEC || GET_CODE (value) == POST_DEC)
3654 inc_amount = -inc_amount;
3656 inc = gen_int_mode (inc_amount, GET_MODE (value));
3659 if (! post && REG_P (incloc))
3660 result = incloc;
3661 else
3662 result = lra_create_new_reg (GET_MODE (value), value, new_rclass,
3663 "INC/DEC result");
3665 if (real_in != result)
3667 /* First copy the location to the result register. */
3668 lra_assert (REG_P (result));
3669 emit_insn (gen_move_insn (result, real_in));
3672 /* We suppose that there are insns to add/sub with the constant
3673 increment permitted in {PRE/POST)_{DEC/INC/MODIFY}. At least the
3674 old reload worked with this assumption. If the assumption
3675 becomes wrong, we should use approach in function
3676 base_plus_disp_to_reg. */
3677 if (in == value)
3679 /* See if we can directly increment INCLOC. */
3680 last = get_last_insn ();
3681 add_insn = emit_insn (plus_p
3682 ? gen_add2_insn (incloc, inc)
3683 : gen_sub2_insn (incloc, inc));
3685 code = recog_memoized (add_insn);
3686 if (code >= 0)
3688 if (! post && result != incloc)
3689 emit_insn (gen_move_insn (result, incloc));
3690 return result;
3692 delete_insns_since (last);
3695 /* If couldn't do the increment directly, must increment in RESULT.
3696 The way we do this depends on whether this is pre- or
3697 post-increment. For pre-increment, copy INCLOC to the reload
3698 register, increment it there, then save back. */
3699 if (! post)
3701 if (real_in != result)
3702 emit_insn (gen_move_insn (result, real_in));
3703 if (plus_p)
3704 emit_insn (gen_add2_insn (result, inc));
3705 else
3706 emit_insn (gen_sub2_insn (result, inc));
3707 if (result != incloc)
3708 emit_insn (gen_move_insn (incloc, result));
3710 else
3712 /* Post-increment.
3714 Because this might be a jump insn or a compare, and because
3715 RESULT may not be available after the insn in an input
3716 reload, we must do the incrementing before the insn being
3717 reloaded for.
3719 We have already copied IN to RESULT. Increment the copy in
3720 RESULT, save that back, then decrement RESULT so it has
3721 the original value. */
3722 if (plus_p)
3723 emit_insn (gen_add2_insn (result, inc));
3724 else
3725 emit_insn (gen_sub2_insn (result, inc));
3726 emit_insn (gen_move_insn (incloc, result));
3727 /* Restore non-modified value for the result. We prefer this
3728 way because it does not require an additional hard
3729 register. */
3730 if (plus_p)
3732 poly_int64 offset;
3733 if (poly_int_rtx_p (inc, &offset))
3734 emit_insn (gen_add2_insn (result,
3735 gen_int_mode (-offset,
3736 GET_MODE (result))));
3737 else
3738 emit_insn (gen_sub2_insn (result, inc));
3740 else
3741 emit_insn (gen_add2_insn (result, inc));
3743 return result;
3746 /* Return true if the current move insn does not need processing as we
3747 already know that it satisfies its constraints. */
3748 static bool
3749 simple_move_p (void)
3751 rtx dest, src;
3752 enum reg_class dclass, sclass;
3754 lra_assert (curr_insn_set != NULL_RTX);
3755 dest = SET_DEST (curr_insn_set);
3756 src = SET_SRC (curr_insn_set);
3758 /* If the instruction has multiple sets we need to process it even if it
3759 is single_set. This can happen if one or more of the SETs are dead.
3760 See PR73650. */
3761 if (multiple_sets (curr_insn))
3762 return false;
3764 return ((dclass = get_op_class (dest)) != NO_REGS
3765 && (sclass = get_op_class (src)) != NO_REGS
3766 /* The backend guarantees that register moves of cost 2
3767 never need reloads. */
3768 && targetm.register_move_cost (GET_MODE (src), sclass, dclass) == 2);
3771 /* Swap operands NOP and NOP + 1. */
3772 static inline void
3773 swap_operands (int nop)
3775 std::swap (curr_operand_mode[nop], curr_operand_mode[nop + 1]);
3776 std::swap (original_subreg_reg_mode[nop], original_subreg_reg_mode[nop + 1]);
3777 std::swap (*curr_id->operand_loc[nop], *curr_id->operand_loc[nop + 1]);
3778 std::swap (equiv_substition_p[nop], equiv_substition_p[nop + 1]);
3779 /* Swap the duplicates too. */
3780 lra_update_dup (curr_id, nop);
3781 lra_update_dup (curr_id, nop + 1);
3784 /* Main entry point of the constraint code: search the body of the
3785 current insn to choose the best alternative. It is mimicking insn
3786 alternative cost calculation model of former reload pass. That is
3787 because machine descriptions were written to use this model. This
3788 model can be changed in future. Make commutative operand exchange
3789 if it is chosen.
3791 if CHECK_ONLY_P is false, do RTL changes to satisfy the
3792 constraints. Return true if any change happened during function
3793 call.
3795 If CHECK_ONLY_P is true then don't do any transformation. Just
3796 check that the insn satisfies all constraints. If the insn does
3797 not satisfy any constraint, return true. */
3798 static bool
3799 curr_insn_transform (bool check_only_p)
3801 int i, j, k;
3802 int n_operands;
3803 int n_alternatives;
3804 int n_outputs;
3805 int commutative;
3806 signed char goal_alt_matched[MAX_RECOG_OPERANDS][MAX_RECOG_OPERANDS];
3807 signed char match_inputs[MAX_RECOG_OPERANDS + 1];
3808 signed char outputs[MAX_RECOG_OPERANDS + 1];
3809 rtx_insn *before, *after;
3810 bool alt_p = false;
3811 /* Flag that the insn has been changed through a transformation. */
3812 bool change_p;
3813 bool sec_mem_p;
3814 bool use_sec_mem_p;
3815 int max_regno_before;
3816 int reused_alternative_num;
3818 curr_insn_set = single_set (curr_insn);
3819 if (curr_insn_set != NULL_RTX && simple_move_p ())
3821 /* We assume that the corresponding insn alternative has no
3822 earlier clobbers. If it is not the case, don't define move
3823 cost equal to 2 for the corresponding register classes. */
3824 lra_set_used_insn_alternative (curr_insn, LRA_NON_CLOBBERED_ALT);
3825 return false;
3828 no_input_reloads_p = no_output_reloads_p = false;
3829 goal_alt_number = -1;
3830 change_p = sec_mem_p = false;
3831 /* JUMP_INSNs and CALL_INSNs are not allowed to have any output
3832 reloads; neither are insns that SET cc0. Insns that use CC0 are
3833 not allowed to have any input reloads. */
3834 if (JUMP_P (curr_insn) || CALL_P (curr_insn))
3835 no_output_reloads_p = true;
3837 if (HAVE_cc0 && reg_referenced_p (cc0_rtx, PATTERN (curr_insn)))
3838 no_input_reloads_p = true;
3839 if (HAVE_cc0 && reg_set_p (cc0_rtx, PATTERN (curr_insn)))
3840 no_output_reloads_p = true;
3842 n_operands = curr_static_id->n_operands;
3843 n_alternatives = curr_static_id->n_alternatives;
3845 /* Just return "no reloads" if insn has no operands with
3846 constraints. */
3847 if (n_operands == 0 || n_alternatives == 0)
3848 return false;
3850 max_regno_before = max_reg_num ();
3852 for (i = 0; i < n_operands; i++)
3854 goal_alt_matched[i][0] = -1;
3855 goal_alt_matches[i] = -1;
3858 commutative = curr_static_id->commutative;
3860 /* Now see what we need for pseudos that didn't get hard regs or got
3861 the wrong kind of hard reg. For this, we must consider all the
3862 operands together against the register constraints. */
3864 best_losers = best_overall = INT_MAX;
3865 best_reload_sum = 0;
3867 curr_swapped = false;
3868 goal_alt_swapped = false;
3870 if (! check_only_p)
3871 /* Make equivalence substitution and memory subreg elimination
3872 before address processing because an address legitimacy can
3873 depend on memory mode. */
3874 for (i = 0; i < n_operands; i++)
3876 rtx op, subst, old;
3877 bool op_change_p = false;
3879 if (curr_static_id->operand[i].is_operator)
3880 continue;
3882 old = op = *curr_id->operand_loc[i];
3883 if (GET_CODE (old) == SUBREG)
3884 old = SUBREG_REG (old);
3885 subst = get_equiv_with_elimination (old, curr_insn);
3886 original_subreg_reg_mode[i] = VOIDmode;
3887 equiv_substition_p[i] = false;
3888 if (subst != old)
3890 equiv_substition_p[i] = true;
3891 subst = copy_rtx (subst);
3892 lra_assert (REG_P (old));
3893 if (GET_CODE (op) != SUBREG)
3894 *curr_id->operand_loc[i] = subst;
3895 else
3897 SUBREG_REG (op) = subst;
3898 if (GET_MODE (subst) == VOIDmode)
3899 original_subreg_reg_mode[i] = GET_MODE (old);
3901 if (lra_dump_file != NULL)
3903 fprintf (lra_dump_file,
3904 "Changing pseudo %d in operand %i of insn %u on equiv ",
3905 REGNO (old), i, INSN_UID (curr_insn));
3906 dump_value_slim (lra_dump_file, subst, 1);
3907 fprintf (lra_dump_file, "\n");
3909 op_change_p = change_p = true;
3911 if (simplify_operand_subreg (i, GET_MODE (old)) || op_change_p)
3913 change_p = true;
3914 lra_update_dup (curr_id, i);
3918 /* Reload address registers and displacements. We do it before
3919 finding an alternative because of memory constraints. */
3920 before = after = NULL;
3921 for (i = 0; i < n_operands; i++)
3922 if (! curr_static_id->operand[i].is_operator
3923 && process_address (i, check_only_p, &before, &after))
3925 if (check_only_p)
3926 return true;
3927 change_p = true;
3928 lra_update_dup (curr_id, i);
3931 if (change_p)
3932 /* If we've changed the instruction then any alternative that
3933 we chose previously may no longer be valid. */
3934 lra_set_used_insn_alternative (curr_insn, LRA_UNKNOWN_ALT);
3936 if (! check_only_p && curr_insn_set != NULL_RTX
3937 && check_and_process_move (&change_p, &sec_mem_p))
3938 return change_p;
3940 try_swapped:
3942 reused_alternative_num = check_only_p ? LRA_UNKNOWN_ALT : curr_id->used_insn_alternative;
3943 if (lra_dump_file != NULL && reused_alternative_num >= 0)
3944 fprintf (lra_dump_file, "Reusing alternative %d for insn #%u\n",
3945 reused_alternative_num, INSN_UID (curr_insn));
3947 if (process_alt_operands (reused_alternative_num))
3948 alt_p = true;
3950 if (check_only_p)
3951 return ! alt_p || best_losers != 0;
3953 /* If insn is commutative (it's safe to exchange a certain pair of
3954 operands) then we need to try each alternative twice, the second
3955 time matching those two operands as if we had exchanged them. To
3956 do this, really exchange them in operands.
3958 If we have just tried the alternatives the second time, return
3959 operands to normal and drop through. */
3961 if (reused_alternative_num < 0 && commutative >= 0)
3963 curr_swapped = !curr_swapped;
3964 if (curr_swapped)
3966 swap_operands (commutative);
3967 goto try_swapped;
3969 else
3970 swap_operands (commutative);
3973 if (! alt_p && ! sec_mem_p)
3975 /* No alternative works with reloads?? */
3976 if (INSN_CODE (curr_insn) >= 0)
3977 fatal_insn ("unable to generate reloads for:", curr_insn);
3978 error_for_asm (curr_insn,
3979 "inconsistent operand constraints in an %<asm%>");
3980 lra_asm_error_p = true;
3981 /* Avoid further trouble with this insn. Don't generate use
3982 pattern here as we could use the insn SP offset. */
3983 lra_set_insn_deleted (curr_insn);
3984 return true;
3987 /* If the best alternative is with operands 1 and 2 swapped, swap
3988 them. Update the operand numbers of any reloads already
3989 pushed. */
3991 if (goal_alt_swapped)
3993 if (lra_dump_file != NULL)
3994 fprintf (lra_dump_file, " Commutative operand exchange in insn %u\n",
3995 INSN_UID (curr_insn));
3997 /* Swap the duplicates too. */
3998 swap_operands (commutative);
3999 change_p = true;
4002 /* Some targets' TARGET_SECONDARY_MEMORY_NEEDED (e.g. x86) are defined
4003 too conservatively. So we use the secondary memory only if there
4004 is no any alternative without reloads. */
4005 use_sec_mem_p = false;
4006 if (! alt_p)
4007 use_sec_mem_p = true;
4008 else if (sec_mem_p)
4010 for (i = 0; i < n_operands; i++)
4011 if (! goal_alt_win[i] && ! goal_alt_match_win[i])
4012 break;
4013 use_sec_mem_p = i < n_operands;
4016 if (use_sec_mem_p)
4018 int in = -1, out = -1;
4019 rtx new_reg, src, dest, rld;
4020 machine_mode sec_mode, rld_mode;
4022 lra_assert (curr_insn_set != NULL_RTX && sec_mem_p);
4023 dest = SET_DEST (curr_insn_set);
4024 src = SET_SRC (curr_insn_set);
4025 for (i = 0; i < n_operands; i++)
4026 if (*curr_id->operand_loc[i] == dest)
4027 out = i;
4028 else if (*curr_id->operand_loc[i] == src)
4029 in = i;
4030 for (i = 0; i < curr_static_id->n_dups; i++)
4031 if (out < 0 && *curr_id->dup_loc[i] == dest)
4032 out = curr_static_id->dup_num[i];
4033 else if (in < 0 && *curr_id->dup_loc[i] == src)
4034 in = curr_static_id->dup_num[i];
4035 lra_assert (out >= 0 && in >= 0
4036 && curr_static_id->operand[out].type == OP_OUT
4037 && curr_static_id->operand[in].type == OP_IN);
4038 rld = partial_subreg_p (GET_MODE (src), GET_MODE (dest)) ? src : dest;
4039 rld_mode = GET_MODE (rld);
4040 sec_mode = targetm.secondary_memory_needed_mode (rld_mode);
4041 new_reg = lra_create_new_reg (sec_mode, NULL_RTX,
4042 NO_REGS, "secondary");
4043 /* If the mode is changed, it should be wider. */
4044 lra_assert (!partial_subreg_p (sec_mode, rld_mode));
4045 if (sec_mode != rld_mode)
4047 /* If the target says specifically to use another mode for
4048 secondary memory moves we cannot reuse the original
4049 insn. */
4050 after = emit_spill_move (false, new_reg, dest);
4051 lra_process_new_insns (curr_insn, NULL, after,
4052 "Inserting the sec. move");
4053 /* We may have non null BEFORE here (e.g. after address
4054 processing. */
4055 push_to_sequence (before);
4056 before = emit_spill_move (true, new_reg, src);
4057 emit_insn (before);
4058 before = get_insns ();
4059 end_sequence ();
4060 lra_process_new_insns (curr_insn, before, NULL, "Changing on");
4061 lra_set_insn_deleted (curr_insn);
4063 else if (dest == rld)
4065 *curr_id->operand_loc[out] = new_reg;
4066 lra_update_dup (curr_id, out);
4067 after = emit_spill_move (false, new_reg, dest);
4068 lra_process_new_insns (curr_insn, NULL, after,
4069 "Inserting the sec. move");
4071 else
4073 *curr_id->operand_loc[in] = new_reg;
4074 lra_update_dup (curr_id, in);
4075 /* See comments above. */
4076 push_to_sequence (before);
4077 before = emit_spill_move (true, new_reg, src);
4078 emit_insn (before);
4079 before = get_insns ();
4080 end_sequence ();
4081 lra_process_new_insns (curr_insn, before, NULL,
4082 "Inserting the sec. move");
4084 lra_update_insn_regno_info (curr_insn);
4085 return true;
4088 lra_assert (goal_alt_number >= 0);
4089 lra_set_used_insn_alternative (curr_insn, goal_alt_number);
4091 if (lra_dump_file != NULL)
4093 const char *p;
4095 fprintf (lra_dump_file, " Choosing alt %d in insn %u:",
4096 goal_alt_number, INSN_UID (curr_insn));
4097 for (i = 0; i < n_operands; i++)
4099 p = (curr_static_id->operand_alternative
4100 [goal_alt_number * n_operands + i].constraint);
4101 if (*p == '\0')
4102 continue;
4103 fprintf (lra_dump_file, " (%d) ", i);
4104 for (; *p != '\0' && *p != ',' && *p != '#'; p++)
4105 fputc (*p, lra_dump_file);
4107 if (INSN_CODE (curr_insn) >= 0
4108 && (p = get_insn_name (INSN_CODE (curr_insn))) != NULL)
4109 fprintf (lra_dump_file, " {%s}", p);
4110 if (maybe_ne (curr_id->sp_offset, 0))
4112 fprintf (lra_dump_file, " (sp_off=");
4113 print_dec (curr_id->sp_offset, lra_dump_file);
4114 fprintf (lra_dump_file, ")");
4116 fprintf (lra_dump_file, "\n");
4119 /* Right now, for any pair of operands I and J that are required to
4120 match, with J < I, goal_alt_matches[I] is J. Add I to
4121 goal_alt_matched[J]. */
4123 for (i = 0; i < n_operands; i++)
4124 if ((j = goal_alt_matches[i]) >= 0)
4126 for (k = 0; goal_alt_matched[j][k] >= 0; k++)
4128 /* We allow matching one output operand and several input
4129 operands. */
4130 lra_assert (k == 0
4131 || (curr_static_id->operand[j].type == OP_OUT
4132 && curr_static_id->operand[i].type == OP_IN
4133 && (curr_static_id->operand
4134 [goal_alt_matched[j][0]].type == OP_IN)));
4135 goal_alt_matched[j][k] = i;
4136 goal_alt_matched[j][k + 1] = -1;
4139 for (i = 0; i < n_operands; i++)
4140 goal_alt_win[i] |= goal_alt_match_win[i];
4142 /* Any constants that aren't allowed and can't be reloaded into
4143 registers are here changed into memory references. */
4144 for (i = 0; i < n_operands; i++)
4145 if (goal_alt_win[i])
4147 int regno;
4148 enum reg_class new_class;
4149 rtx reg = *curr_id->operand_loc[i];
4151 if (GET_CODE (reg) == SUBREG)
4152 reg = SUBREG_REG (reg);
4154 if (REG_P (reg) && (regno = REGNO (reg)) >= FIRST_PSEUDO_REGISTER)
4156 bool ok_p = in_class_p (reg, goal_alt[i], &new_class);
4158 if (new_class != NO_REGS && get_reg_class (regno) != new_class)
4160 lra_assert (ok_p);
4161 lra_change_class (regno, new_class, " Change to", true);
4165 else
4167 const char *constraint;
4168 char c;
4169 rtx op = *curr_id->operand_loc[i];
4170 rtx subreg = NULL_RTX;
4171 machine_mode mode = curr_operand_mode[i];
4173 if (GET_CODE (op) == SUBREG)
4175 subreg = op;
4176 op = SUBREG_REG (op);
4177 mode = GET_MODE (op);
4180 if (CONST_POOL_OK_P (mode, op)
4181 && ((targetm.preferred_reload_class
4182 (op, (enum reg_class) goal_alt[i]) == NO_REGS)
4183 || no_input_reloads_p))
4185 rtx tem = force_const_mem (mode, op);
4187 change_p = true;
4188 if (subreg != NULL_RTX)
4189 tem = gen_rtx_SUBREG (mode, tem, SUBREG_BYTE (subreg));
4191 *curr_id->operand_loc[i] = tem;
4192 lra_update_dup (curr_id, i);
4193 process_address (i, false, &before, &after);
4195 /* If the alternative accepts constant pool refs directly
4196 there will be no reload needed at all. */
4197 if (subreg != NULL_RTX)
4198 continue;
4199 /* Skip alternatives before the one requested. */
4200 constraint = (curr_static_id->operand_alternative
4201 [goal_alt_number * n_operands + i].constraint);
4202 for (;
4203 (c = *constraint) && c != ',' && c != '#';
4204 constraint += CONSTRAINT_LEN (c, constraint))
4206 enum constraint_num cn = lookup_constraint (constraint);
4207 if ((insn_extra_memory_constraint (cn)
4208 || insn_extra_special_memory_constraint (cn))
4209 && satisfies_memory_constraint_p (tem, cn))
4210 break;
4212 if (c == '\0' || c == ',' || c == '#')
4213 continue;
4215 goal_alt_win[i] = true;
4219 n_outputs = 0;
4220 outputs[0] = -1;
4221 for (i = 0; i < n_operands; i++)
4223 int regno;
4224 bool optional_p = false;
4225 rtx old, new_reg;
4226 rtx op = *curr_id->operand_loc[i];
4228 if (goal_alt_win[i])
4230 if (goal_alt[i] == NO_REGS
4231 && REG_P (op)
4232 /* When we assign NO_REGS it means that we will not
4233 assign a hard register to the scratch pseudo by
4234 assigment pass and the scratch pseudo will be
4235 spilled. Spilled scratch pseudos are transformed
4236 back to scratches at the LRA end. */
4237 && lra_former_scratch_operand_p (curr_insn, i)
4238 && lra_former_scratch_p (REGNO (op)))
4240 int regno = REGNO (op);
4241 lra_change_class (regno, NO_REGS, " Change to", true);
4242 if (lra_get_regno_hard_regno (regno) >= 0)
4243 /* We don't have to mark all insn affected by the
4244 spilled pseudo as there is only one such insn, the
4245 current one. */
4246 reg_renumber[regno] = -1;
4247 lra_assert (bitmap_single_bit_set_p
4248 (&lra_reg_info[REGNO (op)].insn_bitmap));
4250 /* We can do an optional reload. If the pseudo got a hard
4251 reg, we might improve the code through inheritance. If
4252 it does not get a hard register we coalesce memory/memory
4253 moves later. Ignore move insns to avoid cycling. */
4254 if (! lra_simple_p
4255 && lra_undo_inheritance_iter < LRA_MAX_INHERITANCE_PASSES
4256 && goal_alt[i] != NO_REGS && REG_P (op)
4257 && (regno = REGNO (op)) >= FIRST_PSEUDO_REGISTER
4258 && regno < new_regno_start
4259 && ! lra_former_scratch_p (regno)
4260 && reg_renumber[regno] < 0
4261 /* Check that the optional reload pseudo will be able to
4262 hold given mode value. */
4263 && ! (prohibited_class_reg_set_mode_p
4264 (goal_alt[i], reg_class_contents[goal_alt[i]],
4265 PSEUDO_REGNO_MODE (regno)))
4266 && (curr_insn_set == NULL_RTX
4267 || !((REG_P (SET_SRC (curr_insn_set))
4268 || MEM_P (SET_SRC (curr_insn_set))
4269 || GET_CODE (SET_SRC (curr_insn_set)) == SUBREG)
4270 && (REG_P (SET_DEST (curr_insn_set))
4271 || MEM_P (SET_DEST (curr_insn_set))
4272 || GET_CODE (SET_DEST (curr_insn_set)) == SUBREG))))
4273 optional_p = true;
4274 else if (goal_alt_matched[i][0] != -1
4275 && curr_static_id->operand[i].type == OP_OUT
4276 && (curr_static_id->operand_alternative
4277 [goal_alt_number * n_operands + i].earlyclobber)
4278 && REG_P (op))
4280 for (j = 0; goal_alt_matched[i][j] != -1; j++)
4282 rtx op2 = *curr_id->operand_loc[goal_alt_matched[i][j]];
4284 if (REG_P (op2) && REGNO (op) != REGNO (op2))
4285 break;
4287 if (goal_alt_matched[i][j] != -1)
4289 /* Generate reloads for different output and matched
4290 input registers. This is the easiest way to avoid
4291 creation of non-existing register conflicts in
4292 lra-lives.c. */
4293 match_reload (i, goal_alt_matched[i], outputs, goal_alt[i], &before,
4294 &after, TRUE);
4295 outputs[n_outputs++] = i;
4296 outputs[n_outputs] = -1;
4298 continue;
4300 else
4301 continue;
4304 /* Operands that match previous ones have already been handled. */
4305 if (goal_alt_matches[i] >= 0)
4306 continue;
4308 /* We should not have an operand with a non-offsettable address
4309 appearing where an offsettable address will do. It also may
4310 be a case when the address should be special in other words
4311 not a general one (e.g. it needs no index reg). */
4312 if (goal_alt_matched[i][0] == -1 && goal_alt_offmemok[i] && MEM_P (op))
4314 enum reg_class rclass;
4315 rtx *loc = &XEXP (op, 0);
4316 enum rtx_code code = GET_CODE (*loc);
4318 push_to_sequence (before);
4319 rclass = base_reg_class (GET_MODE (op), MEM_ADDR_SPACE (op),
4320 MEM, SCRATCH);
4321 if (GET_RTX_CLASS (code) == RTX_AUTOINC)
4322 new_reg = emit_inc (rclass, *loc, *loc,
4323 /* This value does not matter for MODIFY. */
4324 GET_MODE_SIZE (GET_MODE (op)));
4325 else if (get_reload_reg (OP_IN, Pmode, *loc, rclass, FALSE,
4326 "offsetable address", &new_reg))
4328 rtx addr = *loc;
4329 enum rtx_code code = GET_CODE (addr);
4331 if (code == AND && CONST_INT_P (XEXP (addr, 1)))
4332 /* (and ... (const_int -X)) is used to align to X bytes. */
4333 addr = XEXP (*loc, 0);
4334 lra_emit_move (new_reg, addr);
4335 if (addr != *loc)
4336 emit_move_insn (new_reg, gen_rtx_AND (GET_MODE (new_reg), new_reg, XEXP (*loc, 1)));
4338 before = get_insns ();
4339 end_sequence ();
4340 *loc = new_reg;
4341 lra_update_dup (curr_id, i);
4343 else if (goal_alt_matched[i][0] == -1)
4345 machine_mode mode;
4346 rtx reg, *loc;
4347 int hard_regno;
4348 enum op_type type = curr_static_id->operand[i].type;
4350 loc = curr_id->operand_loc[i];
4351 mode = curr_operand_mode[i];
4352 if (GET_CODE (*loc) == SUBREG)
4354 reg = SUBREG_REG (*loc);
4355 poly_int64 byte = SUBREG_BYTE (*loc);
4356 if (REG_P (reg)
4357 /* Strict_low_part requires reloading the register and not
4358 just the subreg. Likewise for a strict subreg no wider
4359 than a word for WORD_REGISTER_OPERATIONS targets. */
4360 && (curr_static_id->operand[i].strict_low
4361 || (!paradoxical_subreg_p (mode, GET_MODE (reg))
4362 && (hard_regno
4363 = get_try_hard_regno (REGNO (reg))) >= 0
4364 && (simplify_subreg_regno
4365 (hard_regno,
4366 GET_MODE (reg), byte, mode) < 0)
4367 && (goal_alt[i] == NO_REGS
4368 || (simplify_subreg_regno
4369 (ira_class_hard_regs[goal_alt[i]][0],
4370 GET_MODE (reg), byte, mode) >= 0)))
4371 || (partial_subreg_p (mode, GET_MODE (reg))
4372 && known_le (GET_MODE_SIZE (GET_MODE (reg)),
4373 UNITS_PER_WORD)
4374 && WORD_REGISTER_OPERATIONS)))
4376 /* An OP_INOUT is required when reloading a subreg of a
4377 mode wider than a word to ensure that data beyond the
4378 word being reloaded is preserved. Also automatically
4379 ensure that strict_low_part reloads are made into
4380 OP_INOUT which should already be true from the backend
4381 constraints. */
4382 if (type == OP_OUT
4383 && (curr_static_id->operand[i].strict_low
4384 || read_modify_subreg_p (*loc)))
4385 type = OP_INOUT;
4386 loc = &SUBREG_REG (*loc);
4387 mode = GET_MODE (*loc);
4390 old = *loc;
4391 if (get_reload_reg (type, mode, old, goal_alt[i],
4392 loc != curr_id->operand_loc[i], "", &new_reg)
4393 && type != OP_OUT)
4395 push_to_sequence (before);
4396 lra_emit_move (new_reg, old);
4397 before = get_insns ();
4398 end_sequence ();
4400 *loc = new_reg;
4401 if (type != OP_IN
4402 && find_reg_note (curr_insn, REG_UNUSED, old) == NULL_RTX)
4404 start_sequence ();
4405 lra_emit_move (type == OP_INOUT ? copy_rtx (old) : old, new_reg);
4406 emit_insn (after);
4407 after = get_insns ();
4408 end_sequence ();
4409 *loc = new_reg;
4411 for (j = 0; j < goal_alt_dont_inherit_ops_num; j++)
4412 if (goal_alt_dont_inherit_ops[j] == i)
4414 lra_set_regno_unique_value (REGNO (new_reg));
4415 break;
4417 lra_update_dup (curr_id, i);
4419 else if (curr_static_id->operand[i].type == OP_IN
4420 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4421 == OP_OUT
4422 || (curr_static_id->operand[goal_alt_matched[i][0]].type
4423 == OP_INOUT
4424 && (operands_match_p
4425 (*curr_id->operand_loc[i],
4426 *curr_id->operand_loc[goal_alt_matched[i][0]],
4427 -1)))))
4429 /* generate reloads for input and matched outputs. */
4430 match_inputs[0] = i;
4431 match_inputs[1] = -1;
4432 match_reload (goal_alt_matched[i][0], match_inputs, outputs,
4433 goal_alt[i], &before, &after,
4434 curr_static_id->operand_alternative
4435 [goal_alt_number * n_operands + goal_alt_matched[i][0]]
4436 .earlyclobber);
4438 else if ((curr_static_id->operand[i].type == OP_OUT
4439 || (curr_static_id->operand[i].type == OP_INOUT
4440 && (operands_match_p
4441 (*curr_id->operand_loc[i],
4442 *curr_id->operand_loc[goal_alt_matched[i][0]],
4443 -1))))
4444 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4445 == OP_IN))
4446 /* Generate reloads for output and matched inputs. */
4447 match_reload (i, goal_alt_matched[i], outputs, goal_alt[i], &before,
4448 &after, curr_static_id->operand_alternative
4449 [goal_alt_number * n_operands + i].earlyclobber);
4450 else if (curr_static_id->operand[i].type == OP_IN
4451 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4452 == OP_IN))
4454 /* Generate reloads for matched inputs. */
4455 match_inputs[0] = i;
4456 for (j = 0; (k = goal_alt_matched[i][j]) >= 0; j++)
4457 match_inputs[j + 1] = k;
4458 match_inputs[j + 1] = -1;
4459 match_reload (-1, match_inputs, outputs, goal_alt[i], &before,
4460 &after, false);
4462 else
4463 /* We must generate code in any case when function
4464 process_alt_operands decides that it is possible. */
4465 gcc_unreachable ();
4467 /* Memorise processed outputs so that output remaining to be processed
4468 can avoid using the same register value (see match_reload). */
4469 if (curr_static_id->operand[i].type == OP_OUT)
4471 outputs[n_outputs++] = i;
4472 outputs[n_outputs] = -1;
4475 if (optional_p)
4477 rtx reg = op;
4479 lra_assert (REG_P (reg));
4480 regno = REGNO (reg);
4481 op = *curr_id->operand_loc[i]; /* Substitution. */
4482 if (GET_CODE (op) == SUBREG)
4483 op = SUBREG_REG (op);
4484 gcc_assert (REG_P (op) && (int) REGNO (op) >= new_regno_start);
4485 bitmap_set_bit (&lra_optional_reload_pseudos, REGNO (op));
4486 lra_reg_info[REGNO (op)].restore_rtx = reg;
4487 if (lra_dump_file != NULL)
4488 fprintf (lra_dump_file,
4489 " Making reload reg %d for reg %d optional\n",
4490 REGNO (op), regno);
4493 if (before != NULL_RTX || after != NULL_RTX
4494 || max_regno_before != max_reg_num ())
4495 change_p = true;
4496 if (change_p)
4498 lra_update_operator_dups (curr_id);
4499 /* Something changes -- process the insn. */
4500 lra_update_insn_regno_info (curr_insn);
4502 lra_process_new_insns (curr_insn, before, after, "Inserting insn reload");
4503 return change_p;
4506 /* Return true if INSN satisfies all constraints. In other words, no
4507 reload insns are needed. */
4508 bool
4509 lra_constrain_insn (rtx_insn *insn)
4511 int saved_new_regno_start = new_regno_start;
4512 int saved_new_insn_uid_start = new_insn_uid_start;
4513 bool change_p;
4515 curr_insn = insn;
4516 curr_id = lra_get_insn_recog_data (curr_insn);
4517 curr_static_id = curr_id->insn_static_data;
4518 new_insn_uid_start = get_max_uid ();
4519 new_regno_start = max_reg_num ();
4520 change_p = curr_insn_transform (true);
4521 new_regno_start = saved_new_regno_start;
4522 new_insn_uid_start = saved_new_insn_uid_start;
4523 return ! change_p;
4526 /* Return true if X is in LIST. */
4527 static bool
4528 in_list_p (rtx x, rtx list)
4530 for (; list != NULL_RTX; list = XEXP (list, 1))
4531 if (XEXP (list, 0) == x)
4532 return true;
4533 return false;
4536 /* Return true if X contains an allocatable hard register (if
4537 HARD_REG_P) or a (spilled if SPILLED_P) pseudo. */
4538 static bool
4539 contains_reg_p (rtx x, bool hard_reg_p, bool spilled_p)
4541 int i, j;
4542 const char *fmt;
4543 enum rtx_code code;
4545 code = GET_CODE (x);
4546 if (REG_P (x))
4548 int regno = REGNO (x);
4549 HARD_REG_SET alloc_regs;
4551 if (hard_reg_p)
4553 if (regno >= FIRST_PSEUDO_REGISTER)
4554 regno = lra_get_regno_hard_regno (regno);
4555 if (regno < 0)
4556 return false;
4557 alloc_regs = ~lra_no_alloc_regs;
4558 return overlaps_hard_reg_set_p (alloc_regs, GET_MODE (x), regno);
4560 else
4562 if (regno < FIRST_PSEUDO_REGISTER)
4563 return false;
4564 if (! spilled_p)
4565 return true;
4566 return lra_get_regno_hard_regno (regno) < 0;
4569 fmt = GET_RTX_FORMAT (code);
4570 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4572 if (fmt[i] == 'e')
4574 if (contains_reg_p (XEXP (x, i), hard_reg_p, spilled_p))
4575 return true;
4577 else if (fmt[i] == 'E')
4579 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4580 if (contains_reg_p (XVECEXP (x, i, j), hard_reg_p, spilled_p))
4581 return true;
4584 return false;
4587 /* Process all regs in location *LOC and change them on equivalent
4588 substitution. Return true if any change was done. */
4589 static bool
4590 loc_equivalence_change_p (rtx *loc)
4592 rtx subst, reg, x = *loc;
4593 bool result = false;
4594 enum rtx_code code = GET_CODE (x);
4595 const char *fmt;
4596 int i, j;
4598 if (code == SUBREG)
4600 reg = SUBREG_REG (x);
4601 if ((subst = get_equiv_with_elimination (reg, curr_insn)) != reg
4602 && GET_MODE (subst) == VOIDmode)
4604 /* We cannot reload debug location. Simplify subreg here
4605 while we know the inner mode. */
4606 *loc = simplify_gen_subreg (GET_MODE (x), subst,
4607 GET_MODE (reg), SUBREG_BYTE (x));
4608 return true;
4611 if (code == REG && (subst = get_equiv_with_elimination (x, curr_insn)) != x)
4613 *loc = subst;
4614 return true;
4617 /* Scan all the operand sub-expressions. */
4618 fmt = GET_RTX_FORMAT (code);
4619 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4621 if (fmt[i] == 'e')
4622 result = loc_equivalence_change_p (&XEXP (x, i)) || result;
4623 else if (fmt[i] == 'E')
4624 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4625 result
4626 = loc_equivalence_change_p (&XVECEXP (x, i, j)) || result;
4628 return result;
4631 /* Similar to loc_equivalence_change_p, but for use as
4632 simplify_replace_fn_rtx callback. DATA is insn for which the
4633 elimination is done. If it null we don't do the elimination. */
4634 static rtx
4635 loc_equivalence_callback (rtx loc, const_rtx, void *data)
4637 if (!REG_P (loc))
4638 return NULL_RTX;
4640 rtx subst = (data == NULL
4641 ? get_equiv (loc) : get_equiv_with_elimination (loc, (rtx_insn *) data));
4642 if (subst != loc)
4643 return subst;
4645 return NULL_RTX;
4648 /* Maximum number of generated reload insns per an insn. It is for
4649 preventing this pass cycling in a bug case. */
4650 #define MAX_RELOAD_INSNS_NUMBER LRA_MAX_INSN_RELOADS
4652 /* The current iteration number of this LRA pass. */
4653 int lra_constraint_iter;
4655 /* True if we substituted equiv which needs checking register
4656 allocation correctness because the equivalent value contains
4657 allocatable hard registers or when we restore multi-register
4658 pseudo. */
4659 bool lra_risky_transformations_p;
4661 /* Return true if REGNO is referenced in more than one block. */
4662 static bool
4663 multi_block_pseudo_p (int regno)
4665 basic_block bb = NULL;
4666 unsigned int uid;
4667 bitmap_iterator bi;
4669 if (regno < FIRST_PSEUDO_REGISTER)
4670 return false;
4672 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi)
4673 if (bb == NULL)
4674 bb = BLOCK_FOR_INSN (lra_insn_recog_data[uid]->insn);
4675 else if (BLOCK_FOR_INSN (lra_insn_recog_data[uid]->insn) != bb)
4676 return true;
4677 return false;
4680 /* Return true if LIST contains a deleted insn. */
4681 static bool
4682 contains_deleted_insn_p (rtx_insn_list *list)
4684 for (; list != NULL_RTX; list = list->next ())
4685 if (NOTE_P (list->insn ())
4686 && NOTE_KIND (list->insn ()) == NOTE_INSN_DELETED)
4687 return true;
4688 return false;
4691 /* Return true if X contains a pseudo dying in INSN. */
4692 static bool
4693 dead_pseudo_p (rtx x, rtx_insn *insn)
4695 int i, j;
4696 const char *fmt;
4697 enum rtx_code code;
4699 if (REG_P (x))
4700 return (insn != NULL_RTX
4701 && find_regno_note (insn, REG_DEAD, REGNO (x)) != NULL_RTX);
4702 code = GET_CODE (x);
4703 fmt = GET_RTX_FORMAT (code);
4704 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4706 if (fmt[i] == 'e')
4708 if (dead_pseudo_p (XEXP (x, i), insn))
4709 return true;
4711 else if (fmt[i] == 'E')
4713 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4714 if (dead_pseudo_p (XVECEXP (x, i, j), insn))
4715 return true;
4718 return false;
4721 /* Return true if INSN contains a dying pseudo in INSN right hand
4722 side. */
4723 static bool
4724 insn_rhs_dead_pseudo_p (rtx_insn *insn)
4726 rtx set = single_set (insn);
4728 gcc_assert (set != NULL);
4729 return dead_pseudo_p (SET_SRC (set), insn);
4732 /* Return true if any init insn of REGNO contains a dying pseudo in
4733 insn right hand side. */
4734 static bool
4735 init_insn_rhs_dead_pseudo_p (int regno)
4737 rtx_insn_list *insns = ira_reg_equiv[regno].init_insns;
4739 if (insns == NULL)
4740 return false;
4741 for (; insns != NULL_RTX; insns = insns->next ())
4742 if (insn_rhs_dead_pseudo_p (insns->insn ()))
4743 return true;
4744 return false;
4747 /* Return TRUE if REGNO has a reverse equivalence. The equivalence is
4748 reverse only if we have one init insn with given REGNO as a
4749 source. */
4750 static bool
4751 reverse_equiv_p (int regno)
4753 rtx_insn_list *insns = ira_reg_equiv[regno].init_insns;
4754 rtx set;
4756 if (insns == NULL)
4757 return false;
4758 if (! INSN_P (insns->insn ())
4759 || insns->next () != NULL)
4760 return false;
4761 if ((set = single_set (insns->insn ())) == NULL_RTX)
4762 return false;
4763 return REG_P (SET_SRC (set)) && (int) REGNO (SET_SRC (set)) == regno;
4766 /* Return TRUE if REGNO was reloaded in an equivalence init insn. We
4767 call this function only for non-reverse equivalence. */
4768 static bool
4769 contains_reloaded_insn_p (int regno)
4771 rtx set;
4772 rtx_insn_list *list = ira_reg_equiv[regno].init_insns;
4774 for (; list != NULL; list = list->next ())
4775 if ((set = single_set (list->insn ())) == NULL_RTX
4776 || ! REG_P (SET_DEST (set))
4777 || (int) REGNO (SET_DEST (set)) != regno)
4778 return true;
4779 return false;
4782 /* Entry function of LRA constraint pass. Return true if the
4783 constraint pass did change the code. */
4784 bool
4785 lra_constraints (bool first_p)
4787 bool changed_p;
4788 int i, hard_regno, new_insns_num;
4789 unsigned int min_len, new_min_len, uid;
4790 rtx set, x, reg, dest_reg;
4791 basic_block last_bb;
4792 bitmap_iterator bi;
4794 lra_constraint_iter++;
4795 if (lra_dump_file != NULL)
4796 fprintf (lra_dump_file, "\n********** Local #%d: **********\n\n",
4797 lra_constraint_iter);
4798 changed_p = false;
4799 if (pic_offset_table_rtx
4800 && REGNO (pic_offset_table_rtx) >= FIRST_PSEUDO_REGISTER)
4801 lra_risky_transformations_p = true;
4802 else
4803 /* On the first iteration we should check IRA assignment
4804 correctness. In rare cases, the assignments can be wrong as
4805 early clobbers operands are ignored in IRA or usages of
4806 paradoxical sub-registers are not taken into account by
4807 IRA. */
4808 lra_risky_transformations_p = first_p;
4809 new_insn_uid_start = get_max_uid ();
4810 new_regno_start = first_p ? lra_constraint_new_regno_start : max_reg_num ();
4811 /* Mark used hard regs for target stack size calulations. */
4812 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4813 if (lra_reg_info[i].nrefs != 0
4814 && (hard_regno = lra_get_regno_hard_regno (i)) >= 0)
4816 int j, nregs;
4818 nregs = hard_regno_nregs (hard_regno, lra_reg_info[i].biggest_mode);
4819 for (j = 0; j < nregs; j++)
4820 df_set_regs_ever_live (hard_regno + j, true);
4822 /* Do elimination before the equivalence processing as we can spill
4823 some pseudos during elimination. */
4824 lra_eliminate (false, first_p);
4825 auto_bitmap equiv_insn_bitmap (&reg_obstack);
4826 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4827 if (lra_reg_info[i].nrefs != 0)
4829 ira_reg_equiv[i].profitable_p = true;
4830 reg = regno_reg_rtx[i];
4831 if (lra_get_regno_hard_regno (i) < 0 && (x = get_equiv (reg)) != reg)
4833 bool pseudo_p = contains_reg_p (x, false, false);
4835 /* After RTL transformation, we cannot guarantee that
4836 pseudo in the substitution was not reloaded which might
4837 make equivalence invalid. For example, in reverse
4838 equiv of p0
4840 p0 <- ...
4842 equiv_mem <- p0
4844 the memory address register was reloaded before the 2nd
4845 insn. */
4846 if ((! first_p && pseudo_p)
4847 /* We don't use DF for compilation speed sake. So it
4848 is problematic to update live info when we use an
4849 equivalence containing pseudos in more than one
4850 BB. */
4851 || (pseudo_p && multi_block_pseudo_p (i))
4852 /* If an init insn was deleted for some reason, cancel
4853 the equiv. We could update the equiv insns after
4854 transformations including an equiv insn deletion
4855 but it is not worthy as such cases are extremely
4856 rare. */
4857 || contains_deleted_insn_p (ira_reg_equiv[i].init_insns)
4858 /* If it is not a reverse equivalence, we check that a
4859 pseudo in rhs of the init insn is not dying in the
4860 insn. Otherwise, the live info at the beginning of
4861 the corresponding BB might be wrong after we
4862 removed the insn. When the equiv can be a
4863 constant, the right hand side of the init insn can
4864 be a pseudo. */
4865 || (! reverse_equiv_p (i)
4866 && (init_insn_rhs_dead_pseudo_p (i)
4867 /* If we reloaded the pseudo in an equivalence
4868 init insn, we cannot remove the equiv init
4869 insns and the init insns might write into
4870 const memory in this case. */
4871 || contains_reloaded_insn_p (i)))
4872 /* Prevent access beyond equivalent memory for
4873 paradoxical subregs. */
4874 || (MEM_P (x)
4875 && maybe_gt (GET_MODE_SIZE (lra_reg_info[i].biggest_mode),
4876 GET_MODE_SIZE (GET_MODE (x))))
4877 || (pic_offset_table_rtx
4878 && ((CONST_POOL_OK_P (PSEUDO_REGNO_MODE (i), x)
4879 && (targetm.preferred_reload_class
4880 (x, lra_get_allocno_class (i)) == NO_REGS))
4881 || contains_symbol_ref_p (x))))
4882 ira_reg_equiv[i].defined_p = false;
4883 if (contains_reg_p (x, false, true))
4884 ira_reg_equiv[i].profitable_p = false;
4885 if (get_equiv (reg) != reg)
4886 bitmap_ior_into (equiv_insn_bitmap, &lra_reg_info[i].insn_bitmap);
4889 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4890 update_equiv (i);
4891 /* We should add all insns containing pseudos which should be
4892 substituted by their equivalences. */
4893 EXECUTE_IF_SET_IN_BITMAP (equiv_insn_bitmap, 0, uid, bi)
4894 lra_push_insn_by_uid (uid);
4895 min_len = lra_insn_stack_length ();
4896 new_insns_num = 0;
4897 last_bb = NULL;
4898 changed_p = false;
4899 while ((new_min_len = lra_insn_stack_length ()) != 0)
4901 curr_insn = lra_pop_insn ();
4902 --new_min_len;
4903 curr_bb = BLOCK_FOR_INSN (curr_insn);
4904 if (curr_bb != last_bb)
4906 last_bb = curr_bb;
4907 bb_reload_num = lra_curr_reload_num;
4909 if (min_len > new_min_len)
4911 min_len = new_min_len;
4912 new_insns_num = 0;
4914 if (new_insns_num > MAX_RELOAD_INSNS_NUMBER)
4915 internal_error
4916 ("maximum number of generated reload insns per insn achieved (%d)",
4917 MAX_RELOAD_INSNS_NUMBER);
4918 new_insns_num++;
4919 if (DEBUG_INSN_P (curr_insn))
4921 /* We need to check equivalence in debug insn and change
4922 pseudo to the equivalent value if necessary. */
4923 curr_id = lra_get_insn_recog_data (curr_insn);
4924 if (bitmap_bit_p (equiv_insn_bitmap, INSN_UID (curr_insn)))
4926 rtx old = *curr_id->operand_loc[0];
4927 *curr_id->operand_loc[0]
4928 = simplify_replace_fn_rtx (old, NULL_RTX,
4929 loc_equivalence_callback, curr_insn);
4930 if (old != *curr_id->operand_loc[0])
4932 lra_update_insn_regno_info (curr_insn);
4933 changed_p = true;
4937 else if (INSN_P (curr_insn))
4939 if ((set = single_set (curr_insn)) != NULL_RTX)
4941 dest_reg = SET_DEST (set);
4942 /* The equivalence pseudo could be set up as SUBREG in a
4943 case when it is a call restore insn in a mode
4944 different from the pseudo mode. */
4945 if (GET_CODE (dest_reg) == SUBREG)
4946 dest_reg = SUBREG_REG (dest_reg);
4947 if ((REG_P (dest_reg)
4948 && (x = get_equiv (dest_reg)) != dest_reg
4949 /* Remove insns which set up a pseudo whose value
4950 cannot be changed. Such insns might be not in
4951 init_insns because we don't update equiv data
4952 during insn transformations.
4954 As an example, let suppose that a pseudo got
4955 hard register and on the 1st pass was not
4956 changed to equivalent constant. We generate an
4957 additional insn setting up the pseudo because of
4958 secondary memory movement. Then the pseudo is
4959 spilled and we use the equiv constant. In this
4960 case we should remove the additional insn and
4961 this insn is not init_insns list. */
4962 && (! MEM_P (x) || MEM_READONLY_P (x)
4963 /* Check that this is actually an insn setting
4964 up the equivalence. */
4965 || in_list_p (curr_insn,
4966 ira_reg_equiv
4967 [REGNO (dest_reg)].init_insns)))
4968 || (((x = get_equiv (SET_SRC (set))) != SET_SRC (set))
4969 && in_list_p (curr_insn,
4970 ira_reg_equiv
4971 [REGNO (SET_SRC (set))].init_insns)))
4973 /* This is equiv init insn of pseudo which did not get a
4974 hard register -- remove the insn. */
4975 if (lra_dump_file != NULL)
4977 fprintf (lra_dump_file,
4978 " Removing equiv init insn %i (freq=%d)\n",
4979 INSN_UID (curr_insn),
4980 REG_FREQ_FROM_BB (BLOCK_FOR_INSN (curr_insn)));
4981 dump_insn_slim (lra_dump_file, curr_insn);
4983 if (contains_reg_p (x, true, false))
4984 lra_risky_transformations_p = true;
4985 lra_set_insn_deleted (curr_insn);
4986 continue;
4989 curr_id = lra_get_insn_recog_data (curr_insn);
4990 curr_static_id = curr_id->insn_static_data;
4991 init_curr_insn_input_reloads ();
4992 init_curr_operand_mode ();
4993 if (curr_insn_transform (false))
4994 changed_p = true;
4995 /* Check non-transformed insns too for equiv change as USE
4996 or CLOBBER don't need reloads but can contain pseudos
4997 being changed on their equivalences. */
4998 else if (bitmap_bit_p (equiv_insn_bitmap, INSN_UID (curr_insn))
4999 && loc_equivalence_change_p (&PATTERN (curr_insn)))
5001 lra_update_insn_regno_info (curr_insn);
5002 changed_p = true;
5007 /* If we used a new hard regno, changed_p should be true because the
5008 hard reg is assigned to a new pseudo. */
5009 if (flag_checking && !changed_p)
5011 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
5012 if (lra_reg_info[i].nrefs != 0
5013 && (hard_regno = lra_get_regno_hard_regno (i)) >= 0)
5015 int j, nregs = hard_regno_nregs (hard_regno,
5016 PSEUDO_REGNO_MODE (i));
5018 for (j = 0; j < nregs; j++)
5019 lra_assert (df_regs_ever_live_p (hard_regno + j));
5022 return changed_p;
5025 static void initiate_invariants (void);
5026 static void finish_invariants (void);
5028 /* Initiate the LRA constraint pass. It is done once per
5029 function. */
5030 void
5031 lra_constraints_init (void)
5033 initiate_invariants ();
5036 /* Finalize the LRA constraint pass. It is done once per
5037 function. */
5038 void
5039 lra_constraints_finish (void)
5041 finish_invariants ();
5046 /* Structure describes invariants for ineheritance. */
5047 struct lra_invariant
5049 /* The order number of the invariant. */
5050 int num;
5051 /* The invariant RTX. */
5052 rtx invariant_rtx;
5053 /* The origin insn of the invariant. */
5054 rtx_insn *insn;
5057 typedef lra_invariant invariant_t;
5058 typedef invariant_t *invariant_ptr_t;
5059 typedef const invariant_t *const_invariant_ptr_t;
5061 /* Pointer to the inheritance invariants. */
5062 static vec<invariant_ptr_t> invariants;
5064 /* Allocation pool for the invariants. */
5065 static object_allocator<lra_invariant> *invariants_pool;
5067 /* Hash table for the invariants. */
5068 static htab_t invariant_table;
5070 /* Hash function for INVARIANT. */
5071 static hashval_t
5072 invariant_hash (const void *invariant)
5074 rtx inv = ((const_invariant_ptr_t) invariant)->invariant_rtx;
5075 return lra_rtx_hash (inv);
5078 /* Equal function for invariants INVARIANT1 and INVARIANT2. */
5079 static int
5080 invariant_eq_p (const void *invariant1, const void *invariant2)
5082 rtx inv1 = ((const_invariant_ptr_t) invariant1)->invariant_rtx;
5083 rtx inv2 = ((const_invariant_ptr_t) invariant2)->invariant_rtx;
5085 return rtx_equal_p (inv1, inv2);
5088 /* Insert INVARIANT_RTX into the table if it is not there yet. Return
5089 invariant which is in the table. */
5090 static invariant_ptr_t
5091 insert_invariant (rtx invariant_rtx)
5093 void **entry_ptr;
5094 invariant_t invariant;
5095 invariant_ptr_t invariant_ptr;
5097 invariant.invariant_rtx = invariant_rtx;
5098 entry_ptr = htab_find_slot (invariant_table, &invariant, INSERT);
5099 if (*entry_ptr == NULL)
5101 invariant_ptr = invariants_pool->allocate ();
5102 invariant_ptr->invariant_rtx = invariant_rtx;
5103 invariant_ptr->insn = NULL;
5104 invariants.safe_push (invariant_ptr);
5105 *entry_ptr = (void *) invariant_ptr;
5107 return (invariant_ptr_t) *entry_ptr;
5110 /* Initiate the invariant table. */
5111 static void
5112 initiate_invariants (void)
5114 invariants.create (100);
5115 invariants_pool
5116 = new object_allocator<lra_invariant> ("Inheritance invariants");
5117 invariant_table = htab_create (100, invariant_hash, invariant_eq_p, NULL);
5120 /* Finish the invariant table. */
5121 static void
5122 finish_invariants (void)
5124 htab_delete (invariant_table);
5125 delete invariants_pool;
5126 invariants.release ();
5129 /* Make the invariant table empty. */
5130 static void
5131 clear_invariants (void)
5133 htab_empty (invariant_table);
5134 invariants_pool->release ();
5135 invariants.truncate (0);
5140 /* This page contains code to do inheritance/split
5141 transformations. */
5143 /* Number of reloads passed so far in current EBB. */
5144 static int reloads_num;
5146 /* Number of calls passed so far in current EBB. */
5147 static int calls_num;
5149 /* Index ID is the CALLS_NUM associated the last call we saw with
5150 ABI identifier ID. */
5151 static int last_call_for_abi[NUM_ABI_IDS];
5153 /* Which registers have been fully or partially clobbered by a call
5154 since they were last used. */
5155 static HARD_REG_SET full_and_partial_call_clobbers;
5157 /* Current reload pseudo check for validity of elements in
5158 USAGE_INSNS. */
5159 static int curr_usage_insns_check;
5161 /* Info about last usage of registers in EBB to do inheritance/split
5162 transformation. Inheritance transformation is done from a spilled
5163 pseudo and split transformations from a hard register or a pseudo
5164 assigned to a hard register. */
5165 struct usage_insns
5167 /* If the value is equal to CURR_USAGE_INSNS_CHECK, then the member
5168 value INSNS is valid. The insns is chain of optional debug insns
5169 and a finishing non-debug insn using the corresponding reg. The
5170 value is also used to mark the registers which are set up in the
5171 current insn. The negated insn uid is used for this. */
5172 int check;
5173 /* Value of global reloads_num at the last insn in INSNS. */
5174 int reloads_num;
5175 /* Value of global reloads_nums at the last insn in INSNS. */
5176 int calls_num;
5177 /* It can be true only for splitting. And it means that the restore
5178 insn should be put after insn given by the following member. */
5179 bool after_p;
5180 /* Next insns in the current EBB which use the original reg and the
5181 original reg value is not changed between the current insn and
5182 the next insns. In order words, e.g. for inheritance, if we need
5183 to use the original reg value again in the next insns we can try
5184 to use the value in a hard register from a reload insn of the
5185 current insn. */
5186 rtx insns;
5189 /* Map: regno -> corresponding pseudo usage insns. */
5190 static struct usage_insns *usage_insns;
5192 static void
5193 setup_next_usage_insn (int regno, rtx insn, int reloads_num, bool after_p)
5195 usage_insns[regno].check = curr_usage_insns_check;
5196 usage_insns[regno].insns = insn;
5197 usage_insns[regno].reloads_num = reloads_num;
5198 usage_insns[regno].calls_num = calls_num;
5199 usage_insns[regno].after_p = after_p;
5200 if (regno >= FIRST_PSEUDO_REGISTER && reg_renumber[regno] >= 0)
5201 remove_from_hard_reg_set (&full_and_partial_call_clobbers,
5202 PSEUDO_REGNO_MODE (regno),
5203 reg_renumber[regno]);
5206 /* The function is used to form list REGNO usages which consists of
5207 optional debug insns finished by a non-debug insn using REGNO.
5208 RELOADS_NUM is current number of reload insns processed so far. */
5209 static void
5210 add_next_usage_insn (int regno, rtx_insn *insn, int reloads_num)
5212 rtx next_usage_insns;
5214 if (usage_insns[regno].check == curr_usage_insns_check
5215 && (next_usage_insns = usage_insns[regno].insns) != NULL_RTX
5216 && DEBUG_INSN_P (insn))
5218 /* Check that we did not add the debug insn yet. */
5219 if (next_usage_insns != insn
5220 && (GET_CODE (next_usage_insns) != INSN_LIST
5221 || XEXP (next_usage_insns, 0) != insn))
5222 usage_insns[regno].insns = gen_rtx_INSN_LIST (VOIDmode, insn,
5223 next_usage_insns);
5225 else if (NONDEBUG_INSN_P (insn))
5226 setup_next_usage_insn (regno, insn, reloads_num, false);
5227 else
5228 usage_insns[regno].check = 0;
5231 /* Return first non-debug insn in list USAGE_INSNS. */
5232 static rtx_insn *
5233 skip_usage_debug_insns (rtx usage_insns)
5235 rtx insn;
5237 /* Skip debug insns. */
5238 for (insn = usage_insns;
5239 insn != NULL_RTX && GET_CODE (insn) == INSN_LIST;
5240 insn = XEXP (insn, 1))
5242 return safe_as_a <rtx_insn *> (insn);
5245 /* Return true if we need secondary memory moves for insn in
5246 USAGE_INSNS after inserting inherited pseudo of class INHER_CL
5247 into the insn. */
5248 static bool
5249 check_secondary_memory_needed_p (enum reg_class inher_cl ATTRIBUTE_UNUSED,
5250 rtx usage_insns ATTRIBUTE_UNUSED)
5252 rtx_insn *insn;
5253 rtx set, dest;
5254 enum reg_class cl;
5256 if (inher_cl == ALL_REGS
5257 || (insn = skip_usage_debug_insns (usage_insns)) == NULL_RTX)
5258 return false;
5259 lra_assert (INSN_P (insn));
5260 if ((set = single_set (insn)) == NULL_RTX || ! REG_P (SET_DEST (set)))
5261 return false;
5262 dest = SET_DEST (set);
5263 if (! REG_P (dest))
5264 return false;
5265 lra_assert (inher_cl != NO_REGS);
5266 cl = get_reg_class (REGNO (dest));
5267 return (cl != NO_REGS && cl != ALL_REGS
5268 && targetm.secondary_memory_needed (GET_MODE (dest), inher_cl, cl));
5271 /* Registers involved in inheritance/split in the current EBB
5272 (inheritance/split pseudos and original registers). */
5273 static bitmap_head check_only_regs;
5275 /* Reload pseudos cannot be involded in invariant inheritance in the
5276 current EBB. */
5277 static bitmap_head invalid_invariant_regs;
5279 /* Do inheritance transformations for insn INSN, which defines (if
5280 DEF_P) or uses ORIGINAL_REGNO. NEXT_USAGE_INSNS specifies which
5281 instruction in the EBB next uses ORIGINAL_REGNO; it has the same
5282 form as the "insns" field of usage_insns. Return true if we
5283 succeed in such transformation.
5285 The transformations look like:
5287 p <- ... i <- ...
5288 ... p <- i (new insn)
5289 ... =>
5290 <- ... p ... <- ... i ...
5292 ... i <- p (new insn)
5293 <- ... p ... <- ... i ...
5294 ... =>
5295 <- ... p ... <- ... i ...
5296 where p is a spilled original pseudo and i is a new inheritance pseudo.
5299 The inheritance pseudo has the smallest class of two classes CL and
5300 class of ORIGINAL REGNO. */
5301 static bool
5302 inherit_reload_reg (bool def_p, int original_regno,
5303 enum reg_class cl, rtx_insn *insn, rtx next_usage_insns)
5305 if (optimize_function_for_size_p (cfun))
5306 return false;
5308 enum reg_class rclass = lra_get_allocno_class (original_regno);
5309 rtx original_reg = regno_reg_rtx[original_regno];
5310 rtx new_reg, usage_insn;
5311 rtx_insn *new_insns;
5313 lra_assert (! usage_insns[original_regno].after_p);
5314 if (lra_dump_file != NULL)
5315 fprintf (lra_dump_file,
5316 " <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<\n");
5317 if (! ira_reg_classes_intersect_p[cl][rclass])
5319 if (lra_dump_file != NULL)
5321 fprintf (lra_dump_file,
5322 " Rejecting inheritance for %d "
5323 "because of disjoint classes %s and %s\n",
5324 original_regno, reg_class_names[cl],
5325 reg_class_names[rclass]);
5326 fprintf (lra_dump_file,
5327 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5329 return false;
5331 if ((ira_class_subset_p[cl][rclass] && cl != rclass)
5332 /* We don't use a subset of two classes because it can be
5333 NO_REGS. This transformation is still profitable in most
5334 cases even if the classes are not intersected as register
5335 move is probably cheaper than a memory load. */
5336 || ira_class_hard_regs_num[cl] < ira_class_hard_regs_num[rclass])
5338 if (lra_dump_file != NULL)
5339 fprintf (lra_dump_file, " Use smallest class of %s and %s\n",
5340 reg_class_names[cl], reg_class_names[rclass]);
5342 rclass = cl;
5344 if (check_secondary_memory_needed_p (rclass, next_usage_insns))
5346 /* Reject inheritance resulting in secondary memory moves.
5347 Otherwise, there is a danger in LRA cycling. Also such
5348 transformation will be unprofitable. */
5349 if (lra_dump_file != NULL)
5351 rtx_insn *insn = skip_usage_debug_insns (next_usage_insns);
5352 rtx set = single_set (insn);
5354 lra_assert (set != NULL_RTX);
5356 rtx dest = SET_DEST (set);
5358 lra_assert (REG_P (dest));
5359 fprintf (lra_dump_file,
5360 " Rejecting inheritance for insn %d(%s)<-%d(%s) "
5361 "as secondary mem is needed\n",
5362 REGNO (dest), reg_class_names[get_reg_class (REGNO (dest))],
5363 original_regno, reg_class_names[rclass]);
5364 fprintf (lra_dump_file,
5365 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5367 return false;
5369 new_reg = lra_create_new_reg (GET_MODE (original_reg), original_reg,
5370 rclass, "inheritance");
5371 start_sequence ();
5372 if (def_p)
5373 lra_emit_move (original_reg, new_reg);
5374 else
5375 lra_emit_move (new_reg, original_reg);
5376 new_insns = get_insns ();
5377 end_sequence ();
5378 if (NEXT_INSN (new_insns) != NULL_RTX)
5380 if (lra_dump_file != NULL)
5382 fprintf (lra_dump_file,
5383 " Rejecting inheritance %d->%d "
5384 "as it results in 2 or more insns:\n",
5385 original_regno, REGNO (new_reg));
5386 dump_rtl_slim (lra_dump_file, new_insns, NULL, -1, 0);
5387 fprintf (lra_dump_file,
5388 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5390 return false;
5392 lra_substitute_pseudo_within_insn (insn, original_regno, new_reg, false);
5393 lra_update_insn_regno_info (insn);
5394 if (! def_p)
5395 /* We now have a new usage insn for original regno. */
5396 setup_next_usage_insn (original_regno, new_insns, reloads_num, false);
5397 if (lra_dump_file != NULL)
5398 fprintf (lra_dump_file, " Original reg change %d->%d (bb%d):\n",
5399 original_regno, REGNO (new_reg), BLOCK_FOR_INSN (insn)->index);
5400 lra_reg_info[REGNO (new_reg)].restore_rtx = regno_reg_rtx[original_regno];
5401 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
5402 bitmap_set_bit (&check_only_regs, original_regno);
5403 bitmap_set_bit (&lra_inheritance_pseudos, REGNO (new_reg));
5404 if (def_p)
5405 lra_process_new_insns (insn, NULL, new_insns,
5406 "Add original<-inheritance");
5407 else
5408 lra_process_new_insns (insn, new_insns, NULL,
5409 "Add inheritance<-original");
5410 while (next_usage_insns != NULL_RTX)
5412 if (GET_CODE (next_usage_insns) != INSN_LIST)
5414 usage_insn = next_usage_insns;
5415 lra_assert (NONDEBUG_INSN_P (usage_insn));
5416 next_usage_insns = NULL;
5418 else
5420 usage_insn = XEXP (next_usage_insns, 0);
5421 lra_assert (DEBUG_INSN_P (usage_insn));
5422 next_usage_insns = XEXP (next_usage_insns, 1);
5424 lra_substitute_pseudo (&usage_insn, original_regno, new_reg, false,
5425 DEBUG_INSN_P (usage_insn));
5426 lra_update_insn_regno_info (as_a <rtx_insn *> (usage_insn));
5427 if (lra_dump_file != NULL)
5429 basic_block bb = BLOCK_FOR_INSN (usage_insn);
5430 fprintf (lra_dump_file,
5431 " Inheritance reuse change %d->%d (bb%d):\n",
5432 original_regno, REGNO (new_reg),
5433 bb ? bb->index : -1);
5434 dump_insn_slim (lra_dump_file, as_a <rtx_insn *> (usage_insn));
5437 if (lra_dump_file != NULL)
5438 fprintf (lra_dump_file,
5439 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5440 return true;
5443 /* Return true if we need a caller save/restore for pseudo REGNO which
5444 was assigned to a hard register. */
5445 static inline bool
5446 need_for_call_save_p (int regno)
5448 lra_assert (regno >= FIRST_PSEUDO_REGISTER && reg_renumber[regno] >= 0);
5449 if (usage_insns[regno].calls_num < calls_num)
5451 unsigned int abis = 0;
5452 for (unsigned int i = 0; i < NUM_ABI_IDS; ++i)
5453 if (last_call_for_abi[i] > usage_insns[regno].calls_num)
5454 abis |= 1 << i;
5455 gcc_assert (abis);
5456 if (call_clobbered_in_region_p (abis, full_and_partial_call_clobbers,
5457 PSEUDO_REGNO_MODE (regno),
5458 reg_renumber[regno]))
5459 return true;
5461 return false;
5464 /* Global registers occurring in the current EBB. */
5465 static bitmap_head ebb_global_regs;
5467 /* Return true if we need a split for hard register REGNO or pseudo
5468 REGNO which was assigned to a hard register.
5469 POTENTIAL_RELOAD_HARD_REGS contains hard registers which might be
5470 used for reloads since the EBB end. It is an approximation of the
5471 used hard registers in the split range. The exact value would
5472 require expensive calculations. If we were aggressive with
5473 splitting because of the approximation, the split pseudo will save
5474 the same hard register assignment and will be removed in the undo
5475 pass. We still need the approximation because too aggressive
5476 splitting would result in too inaccurate cost calculation in the
5477 assignment pass because of too many generated moves which will be
5478 probably removed in the undo pass. */
5479 static inline bool
5480 need_for_split_p (HARD_REG_SET potential_reload_hard_regs, int regno)
5482 int hard_regno = regno < FIRST_PSEUDO_REGISTER ? regno : reg_renumber[regno];
5484 lra_assert (hard_regno >= 0);
5485 return ((TEST_HARD_REG_BIT (potential_reload_hard_regs, hard_regno)
5486 /* Don't split eliminable hard registers, otherwise we can
5487 split hard registers like hard frame pointer, which
5488 lives on BB start/end according to DF-infrastructure,
5489 when there is a pseudo assigned to the register and
5490 living in the same BB. */
5491 && (regno >= FIRST_PSEUDO_REGISTER
5492 || ! TEST_HARD_REG_BIT (eliminable_regset, hard_regno))
5493 && ! TEST_HARD_REG_BIT (lra_no_alloc_regs, hard_regno)
5494 /* Don't split call clobbered hard regs living through
5495 calls, otherwise we might have a check problem in the
5496 assign sub-pass as in the most cases (exception is a
5497 situation when lra_risky_transformations_p value is
5498 true) the assign pass assumes that all pseudos living
5499 through calls are assigned to call saved hard regs. */
5500 && (regno >= FIRST_PSEUDO_REGISTER
5501 || !TEST_HARD_REG_BIT (full_and_partial_call_clobbers, regno))
5502 /* We need at least 2 reloads to make pseudo splitting
5503 profitable. We should provide hard regno splitting in
5504 any case to solve 1st insn scheduling problem when
5505 moving hard register definition up might result in
5506 impossibility to find hard register for reload pseudo of
5507 small register class. */
5508 && (usage_insns[regno].reloads_num
5509 + (regno < FIRST_PSEUDO_REGISTER ? 0 : 3) < reloads_num)
5510 && (regno < FIRST_PSEUDO_REGISTER
5511 /* For short living pseudos, spilling + inheritance can
5512 be considered a substitution for splitting.
5513 Therefore we do not splitting for local pseudos. It
5514 decreases also aggressiveness of splitting. The
5515 minimal number of references is chosen taking into
5516 account that for 2 references splitting has no sense
5517 as we can just spill the pseudo. */
5518 || (regno >= FIRST_PSEUDO_REGISTER
5519 && lra_reg_info[regno].nrefs > 3
5520 && bitmap_bit_p (&ebb_global_regs, regno))))
5521 || (regno >= FIRST_PSEUDO_REGISTER && need_for_call_save_p (regno)));
5524 /* Return class for the split pseudo created from original pseudo with
5525 ALLOCNO_CLASS and MODE which got a hard register HARD_REGNO. We
5526 choose subclass of ALLOCNO_CLASS which contains HARD_REGNO and
5527 results in no secondary memory movements. */
5528 static enum reg_class
5529 choose_split_class (enum reg_class allocno_class,
5530 int hard_regno ATTRIBUTE_UNUSED,
5531 machine_mode mode ATTRIBUTE_UNUSED)
5533 int i;
5534 enum reg_class cl, best_cl = NO_REGS;
5535 enum reg_class hard_reg_class ATTRIBUTE_UNUSED
5536 = REGNO_REG_CLASS (hard_regno);
5538 if (! targetm.secondary_memory_needed (mode, allocno_class, allocno_class)
5539 && TEST_HARD_REG_BIT (reg_class_contents[allocno_class], hard_regno))
5540 return allocno_class;
5541 for (i = 0;
5542 (cl = reg_class_subclasses[allocno_class][i]) != LIM_REG_CLASSES;
5543 i++)
5544 if (! targetm.secondary_memory_needed (mode, cl, hard_reg_class)
5545 && ! targetm.secondary_memory_needed (mode, hard_reg_class, cl)
5546 && TEST_HARD_REG_BIT (reg_class_contents[cl], hard_regno)
5547 && (best_cl == NO_REGS
5548 || ira_class_hard_regs_num[best_cl] < ira_class_hard_regs_num[cl]))
5549 best_cl = cl;
5550 return best_cl;
5553 /* Copy any equivalence information from ORIGINAL_REGNO to NEW_REGNO.
5554 It only makes sense to call this function if NEW_REGNO is always
5555 equal to ORIGINAL_REGNO. */
5557 static void
5558 lra_copy_reg_equiv (unsigned int new_regno, unsigned int original_regno)
5560 if (!ira_reg_equiv[original_regno].defined_p)
5561 return;
5563 ira_expand_reg_equiv ();
5564 ira_reg_equiv[new_regno].defined_p = true;
5565 if (ira_reg_equiv[original_regno].memory)
5566 ira_reg_equiv[new_regno].memory
5567 = copy_rtx (ira_reg_equiv[original_regno].memory);
5568 if (ira_reg_equiv[original_regno].constant)
5569 ira_reg_equiv[new_regno].constant
5570 = copy_rtx (ira_reg_equiv[original_regno].constant);
5571 if (ira_reg_equiv[original_regno].invariant)
5572 ira_reg_equiv[new_regno].invariant
5573 = copy_rtx (ira_reg_equiv[original_regno].invariant);
5576 /* Do split transformations for insn INSN, which defines or uses
5577 ORIGINAL_REGNO. NEXT_USAGE_INSNS specifies which instruction in
5578 the EBB next uses ORIGINAL_REGNO; it has the same form as the
5579 "insns" field of usage_insns. If TO is not NULL, we don't use
5580 usage_insns, we put restore insns after TO insn. It is a case when
5581 we call it from lra_split_hard_reg_for, outside the inheritance
5582 pass.
5584 The transformations look like:
5586 p <- ... p <- ...
5587 ... s <- p (new insn -- save)
5588 ... =>
5589 ... p <- s (new insn -- restore)
5590 <- ... p ... <- ... p ...
5592 <- ... p ... <- ... p ...
5593 ... s <- p (new insn -- save)
5594 ... =>
5595 ... p <- s (new insn -- restore)
5596 <- ... p ... <- ... p ...
5598 where p is an original pseudo got a hard register or a hard
5599 register and s is a new split pseudo. The save is put before INSN
5600 if BEFORE_P is true. Return true if we succeed in such
5601 transformation. */
5602 static bool
5603 split_reg (bool before_p, int original_regno, rtx_insn *insn,
5604 rtx next_usage_insns, rtx_insn *to)
5606 enum reg_class rclass;
5607 rtx original_reg;
5608 int hard_regno, nregs;
5609 rtx new_reg, usage_insn;
5610 rtx_insn *restore, *save;
5611 bool after_p;
5612 bool call_save_p;
5613 machine_mode mode;
5615 if (original_regno < FIRST_PSEUDO_REGISTER)
5617 rclass = ira_allocno_class_translate[REGNO_REG_CLASS (original_regno)];
5618 hard_regno = original_regno;
5619 call_save_p = false;
5620 nregs = 1;
5621 mode = lra_reg_info[hard_regno].biggest_mode;
5622 machine_mode reg_rtx_mode = GET_MODE (regno_reg_rtx[hard_regno]);
5623 /* A reg can have a biggest_mode of VOIDmode if it was only ever seen
5624 as part of a multi-word register. In that case, or if the biggest
5625 mode was larger than a register, just use the reg_rtx. Otherwise,
5626 limit the size to that of the biggest access in the function. */
5627 if (mode == VOIDmode
5628 || paradoxical_subreg_p (mode, reg_rtx_mode))
5630 original_reg = regno_reg_rtx[hard_regno];
5631 mode = reg_rtx_mode;
5633 else
5634 original_reg = gen_rtx_REG (mode, hard_regno);
5636 else
5638 mode = PSEUDO_REGNO_MODE (original_regno);
5639 hard_regno = reg_renumber[original_regno];
5640 nregs = hard_regno_nregs (hard_regno, mode);
5641 rclass = lra_get_allocno_class (original_regno);
5642 original_reg = regno_reg_rtx[original_regno];
5643 call_save_p = need_for_call_save_p (original_regno);
5645 lra_assert (hard_regno >= 0);
5646 if (lra_dump_file != NULL)
5647 fprintf (lra_dump_file,
5648 " ((((((((((((((((((((((((((((((((((((((((((((((((\n");
5650 if (call_save_p)
5652 mode = HARD_REGNO_CALLER_SAVE_MODE (hard_regno,
5653 hard_regno_nregs (hard_regno, mode),
5654 mode);
5655 new_reg = lra_create_new_reg (mode, NULL_RTX, NO_REGS, "save");
5657 else
5659 rclass = choose_split_class (rclass, hard_regno, mode);
5660 if (rclass == NO_REGS)
5662 if (lra_dump_file != NULL)
5664 fprintf (lra_dump_file,
5665 " Rejecting split of %d(%s): "
5666 "no good reg class for %d(%s)\n",
5667 original_regno,
5668 reg_class_names[lra_get_allocno_class (original_regno)],
5669 hard_regno,
5670 reg_class_names[REGNO_REG_CLASS (hard_regno)]);
5671 fprintf
5672 (lra_dump_file,
5673 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5675 return false;
5677 /* Split_if_necessary can split hard registers used as part of a
5678 multi-register mode but splits each register individually. The
5679 mode used for each independent register may not be supported
5680 so reject the split. Splitting the wider mode should theoretically
5681 be possible but is not implemented. */
5682 if (!targetm.hard_regno_mode_ok (hard_regno, mode))
5684 if (lra_dump_file != NULL)
5686 fprintf (lra_dump_file,
5687 " Rejecting split of %d(%s): unsuitable mode %s\n",
5688 original_regno,
5689 reg_class_names[lra_get_allocno_class (original_regno)],
5690 GET_MODE_NAME (mode));
5691 fprintf
5692 (lra_dump_file,
5693 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5695 return false;
5697 new_reg = lra_create_new_reg (mode, original_reg, rclass, "split");
5698 reg_renumber[REGNO (new_reg)] = hard_regno;
5700 int new_regno = REGNO (new_reg);
5701 save = emit_spill_move (true, new_reg, original_reg);
5702 if (NEXT_INSN (save) != NULL_RTX && !call_save_p)
5704 if (lra_dump_file != NULL)
5706 fprintf
5707 (lra_dump_file,
5708 " Rejecting split %d->%d resulting in > 2 save insns:\n",
5709 original_regno, new_regno);
5710 dump_rtl_slim (lra_dump_file, save, NULL, -1, 0);
5711 fprintf (lra_dump_file,
5712 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5714 return false;
5716 restore = emit_spill_move (false, new_reg, original_reg);
5717 if (NEXT_INSN (restore) != NULL_RTX && !call_save_p)
5719 if (lra_dump_file != NULL)
5721 fprintf (lra_dump_file,
5722 " Rejecting split %d->%d "
5723 "resulting in > 2 restore insns:\n",
5724 original_regno, new_regno);
5725 dump_rtl_slim (lra_dump_file, restore, NULL, -1, 0);
5726 fprintf (lra_dump_file,
5727 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5729 return false;
5731 /* Transfer equivalence information to the spill register, so that
5732 if we fail to allocate the spill register, we have the option of
5733 rematerializing the original value instead of spilling to the stack. */
5734 if (!HARD_REGISTER_NUM_P (original_regno)
5735 && mode == PSEUDO_REGNO_MODE (original_regno))
5736 lra_copy_reg_equiv (new_regno, original_regno);
5737 lra_reg_info[new_regno].restore_rtx = regno_reg_rtx[original_regno];
5738 bitmap_set_bit (&lra_split_regs, new_regno);
5739 if (to != NULL)
5741 lra_assert (next_usage_insns == NULL);
5742 usage_insn = to;
5743 after_p = TRUE;
5745 else
5747 /* We need check_only_regs only inside the inheritance pass. */
5748 bitmap_set_bit (&check_only_regs, new_regno);
5749 bitmap_set_bit (&check_only_regs, original_regno);
5750 after_p = usage_insns[original_regno].after_p;
5751 for (;;)
5753 if (GET_CODE (next_usage_insns) != INSN_LIST)
5755 usage_insn = next_usage_insns;
5756 break;
5758 usage_insn = XEXP (next_usage_insns, 0);
5759 lra_assert (DEBUG_INSN_P (usage_insn));
5760 next_usage_insns = XEXP (next_usage_insns, 1);
5761 lra_substitute_pseudo (&usage_insn, original_regno, new_reg, false,
5762 true);
5763 lra_update_insn_regno_info (as_a <rtx_insn *> (usage_insn));
5764 if (lra_dump_file != NULL)
5766 fprintf (lra_dump_file, " Split reuse change %d->%d:\n",
5767 original_regno, new_regno);
5768 dump_insn_slim (lra_dump_file, as_a <rtx_insn *> (usage_insn));
5772 lra_assert (NOTE_P (usage_insn) || NONDEBUG_INSN_P (usage_insn));
5773 lra_assert (usage_insn != insn || (after_p && before_p));
5774 lra_process_new_insns (as_a <rtx_insn *> (usage_insn),
5775 after_p ? NULL : restore,
5776 after_p ? restore : NULL,
5777 call_save_p
5778 ? "Add reg<-save" : "Add reg<-split");
5779 lra_process_new_insns (insn, before_p ? save : NULL,
5780 before_p ? NULL : save,
5781 call_save_p
5782 ? "Add save<-reg" : "Add split<-reg");
5783 if (nregs > 1)
5784 /* If we are trying to split multi-register. We should check
5785 conflicts on the next assignment sub-pass. IRA can allocate on
5786 sub-register levels, LRA do this on pseudos level right now and
5787 this discrepancy may create allocation conflicts after
5788 splitting. */
5789 lra_risky_transformations_p = true;
5790 if (lra_dump_file != NULL)
5791 fprintf (lra_dump_file,
5792 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5793 return true;
5796 /* Split a hard reg for reload pseudo REGNO having RCLASS and living
5797 in the range [FROM, TO]. Return true if did a split. Otherwise,
5798 return false. */
5799 bool
5800 spill_hard_reg_in_range (int regno, enum reg_class rclass, rtx_insn *from, rtx_insn *to)
5802 int i, hard_regno;
5803 int rclass_size;
5804 rtx_insn *insn;
5805 unsigned int uid;
5806 bitmap_iterator bi;
5807 HARD_REG_SET ignore;
5809 lra_assert (from != NULL && to != NULL);
5810 CLEAR_HARD_REG_SET (ignore);
5811 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi)
5813 lra_insn_recog_data_t id = lra_insn_recog_data[uid];
5814 struct lra_static_insn_data *static_id = id->insn_static_data;
5815 struct lra_insn_reg *reg;
5817 for (reg = id->regs; reg != NULL; reg = reg->next)
5818 if (reg->regno < FIRST_PSEUDO_REGISTER)
5819 SET_HARD_REG_BIT (ignore, reg->regno);
5820 for (reg = static_id->hard_regs; reg != NULL; reg = reg->next)
5821 SET_HARD_REG_BIT (ignore, reg->regno);
5823 rclass_size = ira_class_hard_regs_num[rclass];
5824 for (i = 0; i < rclass_size; i++)
5826 hard_regno = ira_class_hard_regs[rclass][i];
5827 if (! TEST_HARD_REG_BIT (lra_reg_info[regno].conflict_hard_regs, hard_regno)
5828 || TEST_HARD_REG_BIT (ignore, hard_regno))
5829 continue;
5830 for (insn = from; insn != NEXT_INSN (to); insn = NEXT_INSN (insn))
5832 struct lra_static_insn_data *static_id;
5833 struct lra_insn_reg *reg;
5835 if (!INSN_P (insn))
5836 continue;
5837 if (bitmap_bit_p (&lra_reg_info[hard_regno].insn_bitmap,
5838 INSN_UID (insn)))
5839 break;
5840 static_id = lra_get_insn_recog_data (insn)->insn_static_data;
5841 for (reg = static_id->hard_regs; reg != NULL; reg = reg->next)
5842 if (reg->regno == hard_regno)
5843 break;
5844 if (reg != NULL)
5845 break;
5847 if (insn != NEXT_INSN (to))
5848 continue;
5849 if (split_reg (TRUE, hard_regno, from, NULL, to))
5850 return true;
5852 return false;
5855 /* Recognize that we need a split transformation for insn INSN, which
5856 defines or uses REGNO in its insn biggest MODE (we use it only if
5857 REGNO is a hard register). POTENTIAL_RELOAD_HARD_REGS contains
5858 hard registers which might be used for reloads since the EBB end.
5859 Put the save before INSN if BEFORE_P is true. MAX_UID is maximla
5860 uid before starting INSN processing. Return true if we succeed in
5861 such transformation. */
5862 static bool
5863 split_if_necessary (int regno, machine_mode mode,
5864 HARD_REG_SET potential_reload_hard_regs,
5865 bool before_p, rtx_insn *insn, int max_uid)
5867 bool res = false;
5868 int i, nregs = 1;
5869 rtx next_usage_insns;
5871 if (regno < FIRST_PSEUDO_REGISTER)
5872 nregs = hard_regno_nregs (regno, mode);
5873 for (i = 0; i < nregs; i++)
5874 if (usage_insns[regno + i].check == curr_usage_insns_check
5875 && (next_usage_insns = usage_insns[regno + i].insns) != NULL_RTX
5876 /* To avoid processing the register twice or more. */
5877 && ((GET_CODE (next_usage_insns) != INSN_LIST
5878 && INSN_UID (next_usage_insns) < max_uid)
5879 || (GET_CODE (next_usage_insns) == INSN_LIST
5880 && (INSN_UID (XEXP (next_usage_insns, 0)) < max_uid)))
5881 && need_for_split_p (potential_reload_hard_regs, regno + i)
5882 && split_reg (before_p, regno + i, insn, next_usage_insns, NULL))
5883 res = true;
5884 return res;
5887 /* Return TRUE if rtx X is considered as an invariant for
5888 inheritance. */
5889 static bool
5890 invariant_p (const_rtx x)
5892 machine_mode mode;
5893 const char *fmt;
5894 enum rtx_code code;
5895 int i, j;
5897 if (side_effects_p (x))
5898 return false;
5900 code = GET_CODE (x);
5901 mode = GET_MODE (x);
5902 if (code == SUBREG)
5904 x = SUBREG_REG (x);
5905 code = GET_CODE (x);
5906 mode = wider_subreg_mode (mode, GET_MODE (x));
5909 if (MEM_P (x))
5910 return false;
5912 if (REG_P (x))
5914 int i, nregs, regno = REGNO (x);
5916 if (regno >= FIRST_PSEUDO_REGISTER || regno == STACK_POINTER_REGNUM
5917 || TEST_HARD_REG_BIT (eliminable_regset, regno)
5918 || GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
5919 return false;
5920 nregs = hard_regno_nregs (regno, mode);
5921 for (i = 0; i < nregs; i++)
5922 if (! fixed_regs[regno + i]
5923 /* A hard register may be clobbered in the current insn
5924 but we can ignore this case because if the hard
5925 register is used it should be set somewhere after the
5926 clobber. */
5927 || bitmap_bit_p (&invalid_invariant_regs, regno + i))
5928 return false;
5930 fmt = GET_RTX_FORMAT (code);
5931 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
5933 if (fmt[i] == 'e')
5935 if (! invariant_p (XEXP (x, i)))
5936 return false;
5938 else if (fmt[i] == 'E')
5940 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
5941 if (! invariant_p (XVECEXP (x, i, j)))
5942 return false;
5945 return true;
5948 /* We have 'dest_reg <- invariant'. Let us try to make an invariant
5949 inheritance transformation (using dest_reg instead invariant in a
5950 subsequent insn). */
5951 static bool
5952 process_invariant_for_inheritance (rtx dst_reg, rtx invariant_rtx)
5954 invariant_ptr_t invariant_ptr;
5955 rtx_insn *insn, *new_insns;
5956 rtx insn_set, insn_reg, new_reg;
5957 int insn_regno;
5958 bool succ_p = false;
5959 int dst_regno = REGNO (dst_reg);
5960 machine_mode dst_mode = GET_MODE (dst_reg);
5961 enum reg_class cl = lra_get_allocno_class (dst_regno), insn_reg_cl;
5963 invariant_ptr = insert_invariant (invariant_rtx);
5964 if ((insn = invariant_ptr->insn) != NULL_RTX)
5966 /* We have a subsequent insn using the invariant. */
5967 insn_set = single_set (insn);
5968 lra_assert (insn_set != NULL);
5969 insn_reg = SET_DEST (insn_set);
5970 lra_assert (REG_P (insn_reg));
5971 insn_regno = REGNO (insn_reg);
5972 insn_reg_cl = lra_get_allocno_class (insn_regno);
5974 if (dst_mode == GET_MODE (insn_reg)
5975 /* We should consider only result move reg insns which are
5976 cheap. */
5977 && targetm.register_move_cost (dst_mode, cl, insn_reg_cl) == 2
5978 && targetm.register_move_cost (dst_mode, cl, cl) == 2)
5980 if (lra_dump_file != NULL)
5981 fprintf (lra_dump_file,
5982 " [[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[\n");
5983 new_reg = lra_create_new_reg (dst_mode, dst_reg,
5984 cl, "invariant inheritance");
5985 bitmap_set_bit (&lra_inheritance_pseudos, REGNO (new_reg));
5986 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
5987 lra_reg_info[REGNO (new_reg)].restore_rtx = PATTERN (insn);
5988 start_sequence ();
5989 lra_emit_move (new_reg, dst_reg);
5990 new_insns = get_insns ();
5991 end_sequence ();
5992 lra_process_new_insns (curr_insn, NULL, new_insns,
5993 "Add invariant inheritance<-original");
5994 start_sequence ();
5995 lra_emit_move (SET_DEST (insn_set), new_reg);
5996 new_insns = get_insns ();
5997 end_sequence ();
5998 lra_process_new_insns (insn, NULL, new_insns,
5999 "Changing reload<-inheritance");
6000 lra_set_insn_deleted (insn);
6001 succ_p = true;
6002 if (lra_dump_file != NULL)
6004 fprintf (lra_dump_file,
6005 " Invariant inheritance reuse change %d (bb%d):\n",
6006 REGNO (new_reg), BLOCK_FOR_INSN (insn)->index);
6007 dump_insn_slim (lra_dump_file, insn);
6008 fprintf (lra_dump_file,
6009 " ]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]\n");
6013 invariant_ptr->insn = curr_insn;
6014 return succ_p;
6017 /* Check only registers living at the current program point in the
6018 current EBB. */
6019 static bitmap_head live_regs;
6021 /* Update live info in EBB given by its HEAD and TAIL insns after
6022 inheritance/split transformation. The function removes dead moves
6023 too. */
6024 static void
6025 update_ebb_live_info (rtx_insn *head, rtx_insn *tail)
6027 unsigned int j;
6028 int i, regno;
6029 bool live_p;
6030 rtx_insn *prev_insn;
6031 rtx set;
6032 bool remove_p;
6033 basic_block last_bb, prev_bb, curr_bb;
6034 bitmap_iterator bi;
6035 struct lra_insn_reg *reg;
6036 edge e;
6037 edge_iterator ei;
6039 last_bb = BLOCK_FOR_INSN (tail);
6040 prev_bb = NULL;
6041 for (curr_insn = tail;
6042 curr_insn != PREV_INSN (head);
6043 curr_insn = prev_insn)
6045 prev_insn = PREV_INSN (curr_insn);
6046 /* We need to process empty blocks too. They contain
6047 NOTE_INSN_BASIC_BLOCK referring for the basic block. */
6048 if (NOTE_P (curr_insn) && NOTE_KIND (curr_insn) != NOTE_INSN_BASIC_BLOCK)
6049 continue;
6050 curr_bb = BLOCK_FOR_INSN (curr_insn);
6051 if (curr_bb != prev_bb)
6053 if (prev_bb != NULL)
6055 /* Update df_get_live_in (prev_bb): */
6056 EXECUTE_IF_SET_IN_BITMAP (&check_only_regs, 0, j, bi)
6057 if (bitmap_bit_p (&live_regs, j))
6058 bitmap_set_bit (df_get_live_in (prev_bb), j);
6059 else
6060 bitmap_clear_bit (df_get_live_in (prev_bb), j);
6062 if (curr_bb != last_bb)
6064 /* Update df_get_live_out (curr_bb): */
6065 EXECUTE_IF_SET_IN_BITMAP (&check_only_regs, 0, j, bi)
6067 live_p = bitmap_bit_p (&live_regs, j);
6068 if (! live_p)
6069 FOR_EACH_EDGE (e, ei, curr_bb->succs)
6070 if (bitmap_bit_p (df_get_live_in (e->dest), j))
6072 live_p = true;
6073 break;
6075 if (live_p)
6076 bitmap_set_bit (df_get_live_out (curr_bb), j);
6077 else
6078 bitmap_clear_bit (df_get_live_out (curr_bb), j);
6081 prev_bb = curr_bb;
6082 bitmap_and (&live_regs, &check_only_regs, df_get_live_out (curr_bb));
6084 if (! NONDEBUG_INSN_P (curr_insn))
6085 continue;
6086 curr_id = lra_get_insn_recog_data (curr_insn);
6087 curr_static_id = curr_id->insn_static_data;
6088 remove_p = false;
6089 if ((set = single_set (curr_insn)) != NULL_RTX
6090 && REG_P (SET_DEST (set))
6091 && (regno = REGNO (SET_DEST (set))) >= FIRST_PSEUDO_REGISTER
6092 && SET_DEST (set) != pic_offset_table_rtx
6093 && bitmap_bit_p (&check_only_regs, regno)
6094 && ! bitmap_bit_p (&live_regs, regno))
6095 remove_p = true;
6096 /* See which defined values die here. */
6097 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
6098 if (reg->type == OP_OUT && ! reg->subreg_p)
6099 bitmap_clear_bit (&live_regs, reg->regno);
6100 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
6101 if (reg->type == OP_OUT && ! reg->subreg_p)
6102 bitmap_clear_bit (&live_regs, reg->regno);
6103 if (curr_id->arg_hard_regs != NULL)
6104 /* Make clobbered argument hard registers die. */
6105 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6106 if (regno >= FIRST_PSEUDO_REGISTER)
6107 bitmap_clear_bit (&live_regs, regno - FIRST_PSEUDO_REGISTER);
6108 /* Mark each used value as live. */
6109 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
6110 if (reg->type != OP_OUT
6111 && bitmap_bit_p (&check_only_regs, reg->regno))
6112 bitmap_set_bit (&live_regs, reg->regno);
6113 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
6114 if (reg->type != OP_OUT
6115 && bitmap_bit_p (&check_only_regs, reg->regno))
6116 bitmap_set_bit (&live_regs, reg->regno);
6117 if (curr_id->arg_hard_regs != NULL)
6118 /* Make used argument hard registers live. */
6119 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6120 if (regno < FIRST_PSEUDO_REGISTER
6121 && bitmap_bit_p (&check_only_regs, regno))
6122 bitmap_set_bit (&live_regs, regno);
6123 /* It is quite important to remove dead move insns because it
6124 means removing dead store. We don't need to process them for
6125 constraints. */
6126 if (remove_p)
6128 if (lra_dump_file != NULL)
6130 fprintf (lra_dump_file, " Removing dead insn:\n ");
6131 dump_insn_slim (lra_dump_file, curr_insn);
6133 lra_set_insn_deleted (curr_insn);
6138 /* The structure describes info to do an inheritance for the current
6139 insn. We need to collect such info first before doing the
6140 transformations because the transformations change the insn
6141 internal representation. */
6142 struct to_inherit
6144 /* Original regno. */
6145 int regno;
6146 /* Subsequent insns which can inherit original reg value. */
6147 rtx insns;
6150 /* Array containing all info for doing inheritance from the current
6151 insn. */
6152 static struct to_inherit to_inherit[LRA_MAX_INSN_RELOADS];
6154 /* Number elements in the previous array. */
6155 static int to_inherit_num;
6157 /* Add inheritance info REGNO and INSNS. Their meaning is described in
6158 structure to_inherit. */
6159 static void
6160 add_to_inherit (int regno, rtx insns)
6162 int i;
6164 for (i = 0; i < to_inherit_num; i++)
6165 if (to_inherit[i].regno == regno)
6166 return;
6167 lra_assert (to_inherit_num < LRA_MAX_INSN_RELOADS);
6168 to_inherit[to_inherit_num].regno = regno;
6169 to_inherit[to_inherit_num++].insns = insns;
6172 /* Return the last non-debug insn in basic block BB, or the block begin
6173 note if none. */
6174 static rtx_insn *
6175 get_last_insertion_point (basic_block bb)
6177 rtx_insn *insn;
6179 FOR_BB_INSNS_REVERSE (bb, insn)
6180 if (NONDEBUG_INSN_P (insn) || NOTE_INSN_BASIC_BLOCK_P (insn))
6181 return insn;
6182 gcc_unreachable ();
6185 /* Set up RES by registers living on edges FROM except the edge (FROM,
6186 TO) or by registers set up in a jump insn in BB FROM. */
6187 static void
6188 get_live_on_other_edges (basic_block from, basic_block to, bitmap res)
6190 rtx_insn *last;
6191 struct lra_insn_reg *reg;
6192 edge e;
6193 edge_iterator ei;
6195 lra_assert (to != NULL);
6196 bitmap_clear (res);
6197 FOR_EACH_EDGE (e, ei, from->succs)
6198 if (e->dest != to)
6199 bitmap_ior_into (res, df_get_live_in (e->dest));
6200 last = get_last_insertion_point (from);
6201 if (! JUMP_P (last))
6202 return;
6203 curr_id = lra_get_insn_recog_data (last);
6204 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
6205 if (reg->type != OP_IN)
6206 bitmap_set_bit (res, reg->regno);
6209 /* Used as a temporary results of some bitmap calculations. */
6210 static bitmap_head temp_bitmap;
6212 /* We split for reloads of small class of hard regs. The following
6213 defines how many hard regs the class should have to be qualified as
6214 small. The code is mostly oriented to x86/x86-64 architecture
6215 where some insns need to use only specific register or pair of
6216 registers and these register can live in RTL explicitly, e.g. for
6217 parameter passing. */
6218 static const int max_small_class_regs_num = 2;
6220 /* Do inheritance/split transformations in EBB starting with HEAD and
6221 finishing on TAIL. We process EBB insns in the reverse order.
6222 Return true if we did any inheritance/split transformation in the
6223 EBB.
6225 We should avoid excessive splitting which results in worse code
6226 because of inaccurate cost calculations for spilling new split
6227 pseudos in such case. To achieve this we do splitting only if
6228 register pressure is high in given basic block and there are reload
6229 pseudos requiring hard registers. We could do more register
6230 pressure calculations at any given program point to avoid necessary
6231 splitting even more but it is to expensive and the current approach
6232 works well enough. */
6233 static bool
6234 inherit_in_ebb (rtx_insn *head, rtx_insn *tail)
6236 int i, src_regno, dst_regno, nregs;
6237 bool change_p, succ_p, update_reloads_num_p;
6238 rtx_insn *prev_insn, *last_insn;
6239 rtx next_usage_insns, curr_set;
6240 enum reg_class cl;
6241 struct lra_insn_reg *reg;
6242 basic_block last_processed_bb, curr_bb = NULL;
6243 HARD_REG_SET potential_reload_hard_regs, live_hard_regs;
6244 bitmap to_process;
6245 unsigned int j;
6246 bitmap_iterator bi;
6247 bool head_p, after_p;
6249 change_p = false;
6250 curr_usage_insns_check++;
6251 clear_invariants ();
6252 reloads_num = calls_num = 0;
6253 for (unsigned int i = 0; i < NUM_ABI_IDS; ++i)
6254 last_call_for_abi[i] = 0;
6255 CLEAR_HARD_REG_SET (full_and_partial_call_clobbers);
6256 bitmap_clear (&check_only_regs);
6257 bitmap_clear (&invalid_invariant_regs);
6258 last_processed_bb = NULL;
6259 CLEAR_HARD_REG_SET (potential_reload_hard_regs);
6260 live_hard_regs = eliminable_regset | lra_no_alloc_regs;
6261 /* We don't process new insns generated in the loop. */
6262 for (curr_insn = tail; curr_insn != PREV_INSN (head); curr_insn = prev_insn)
6264 prev_insn = PREV_INSN (curr_insn);
6265 if (BLOCK_FOR_INSN (curr_insn) != NULL)
6266 curr_bb = BLOCK_FOR_INSN (curr_insn);
6267 if (last_processed_bb != curr_bb)
6269 /* We are at the end of BB. Add qualified living
6270 pseudos for potential splitting. */
6271 to_process = df_get_live_out (curr_bb);
6272 if (last_processed_bb != NULL)
6274 /* We are somewhere in the middle of EBB. */
6275 get_live_on_other_edges (curr_bb, last_processed_bb,
6276 &temp_bitmap);
6277 to_process = &temp_bitmap;
6279 last_processed_bb = curr_bb;
6280 last_insn = get_last_insertion_point (curr_bb);
6281 after_p = (! JUMP_P (last_insn)
6282 && (! CALL_P (last_insn)
6283 || (find_reg_note (last_insn,
6284 REG_NORETURN, NULL_RTX) == NULL_RTX
6285 && ! SIBLING_CALL_P (last_insn))));
6286 CLEAR_HARD_REG_SET (potential_reload_hard_regs);
6287 EXECUTE_IF_SET_IN_BITMAP (to_process, 0, j, bi)
6289 if ((int) j >= lra_constraint_new_regno_start)
6290 break;
6291 if (j < FIRST_PSEUDO_REGISTER || reg_renumber[j] >= 0)
6293 if (j < FIRST_PSEUDO_REGISTER)
6294 SET_HARD_REG_BIT (live_hard_regs, j);
6295 else
6296 add_to_hard_reg_set (&live_hard_regs,
6297 PSEUDO_REGNO_MODE (j),
6298 reg_renumber[j]);
6299 setup_next_usage_insn (j, last_insn, reloads_num, after_p);
6303 src_regno = dst_regno = -1;
6304 curr_set = single_set (curr_insn);
6305 if (curr_set != NULL_RTX && REG_P (SET_DEST (curr_set)))
6306 dst_regno = REGNO (SET_DEST (curr_set));
6307 if (curr_set != NULL_RTX && REG_P (SET_SRC (curr_set)))
6308 src_regno = REGNO (SET_SRC (curr_set));
6309 update_reloads_num_p = true;
6310 if (src_regno < lra_constraint_new_regno_start
6311 && src_regno >= FIRST_PSEUDO_REGISTER
6312 && reg_renumber[src_regno] < 0
6313 && dst_regno >= lra_constraint_new_regno_start
6314 && (cl = lra_get_allocno_class (dst_regno)) != NO_REGS)
6316 /* 'reload_pseudo <- original_pseudo'. */
6317 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6318 reloads_num++;
6319 update_reloads_num_p = false;
6320 succ_p = false;
6321 if (usage_insns[src_regno].check == curr_usage_insns_check
6322 && (next_usage_insns = usage_insns[src_regno].insns) != NULL_RTX)
6323 succ_p = inherit_reload_reg (false, src_regno, cl,
6324 curr_insn, next_usage_insns);
6325 if (succ_p)
6326 change_p = true;
6327 else
6328 setup_next_usage_insn (src_regno, curr_insn, reloads_num, false);
6329 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6330 potential_reload_hard_regs |= reg_class_contents[cl];
6332 else if (src_regno < 0
6333 && dst_regno >= lra_constraint_new_regno_start
6334 && invariant_p (SET_SRC (curr_set))
6335 && (cl = lra_get_allocno_class (dst_regno)) != NO_REGS
6336 && ! bitmap_bit_p (&invalid_invariant_regs, dst_regno)
6337 && ! bitmap_bit_p (&invalid_invariant_regs,
6338 ORIGINAL_REGNO(regno_reg_rtx[dst_regno])))
6340 /* 'reload_pseudo <- invariant'. */
6341 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6342 reloads_num++;
6343 update_reloads_num_p = false;
6344 if (process_invariant_for_inheritance (SET_DEST (curr_set), SET_SRC (curr_set)))
6345 change_p = true;
6346 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6347 potential_reload_hard_regs |= reg_class_contents[cl];
6349 else if (src_regno >= lra_constraint_new_regno_start
6350 && dst_regno < lra_constraint_new_regno_start
6351 && dst_regno >= FIRST_PSEUDO_REGISTER
6352 && reg_renumber[dst_regno] < 0
6353 && (cl = lra_get_allocno_class (src_regno)) != NO_REGS
6354 && usage_insns[dst_regno].check == curr_usage_insns_check
6355 && (next_usage_insns
6356 = usage_insns[dst_regno].insns) != NULL_RTX)
6358 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6359 reloads_num++;
6360 update_reloads_num_p = false;
6361 /* 'original_pseudo <- reload_pseudo'. */
6362 if (! JUMP_P (curr_insn)
6363 && inherit_reload_reg (true, dst_regno, cl,
6364 curr_insn, next_usage_insns))
6365 change_p = true;
6366 /* Invalidate. */
6367 usage_insns[dst_regno].check = 0;
6368 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6369 potential_reload_hard_regs |= reg_class_contents[cl];
6371 else if (INSN_P (curr_insn))
6373 int iter;
6374 int max_uid = get_max_uid ();
6376 curr_id = lra_get_insn_recog_data (curr_insn);
6377 curr_static_id = curr_id->insn_static_data;
6378 to_inherit_num = 0;
6379 /* Process insn definitions. */
6380 for (iter = 0; iter < 2; iter++)
6381 for (reg = iter == 0 ? curr_id->regs : curr_static_id->hard_regs;
6382 reg != NULL;
6383 reg = reg->next)
6384 if (reg->type != OP_IN
6385 && (dst_regno = reg->regno) < lra_constraint_new_regno_start)
6387 if (dst_regno >= FIRST_PSEUDO_REGISTER && reg->type == OP_OUT
6388 && reg_renumber[dst_regno] < 0 && ! reg->subreg_p
6389 && usage_insns[dst_regno].check == curr_usage_insns_check
6390 && (next_usage_insns
6391 = usage_insns[dst_regno].insns) != NULL_RTX)
6393 struct lra_insn_reg *r;
6395 for (r = curr_id->regs; r != NULL; r = r->next)
6396 if (r->type != OP_OUT && r->regno == dst_regno)
6397 break;
6398 /* Don't do inheritance if the pseudo is also
6399 used in the insn. */
6400 if (r == NULL)
6401 /* We cannot do inheritance right now
6402 because the current insn reg info (chain
6403 regs) can change after that. */
6404 add_to_inherit (dst_regno, next_usage_insns);
6406 /* We cannot process one reg twice here because of
6407 usage_insns invalidation. */
6408 if ((dst_regno < FIRST_PSEUDO_REGISTER
6409 || reg_renumber[dst_regno] >= 0)
6410 && ! reg->subreg_p && reg->type != OP_IN)
6412 HARD_REG_SET s;
6414 if (split_if_necessary (dst_regno, reg->biggest_mode,
6415 potential_reload_hard_regs,
6416 false, curr_insn, max_uid))
6417 change_p = true;
6418 CLEAR_HARD_REG_SET (s);
6419 if (dst_regno < FIRST_PSEUDO_REGISTER)
6420 add_to_hard_reg_set (&s, reg->biggest_mode, dst_regno);
6421 else
6422 add_to_hard_reg_set (&s, PSEUDO_REGNO_MODE (dst_regno),
6423 reg_renumber[dst_regno]);
6424 live_hard_regs &= ~s;
6425 potential_reload_hard_regs &= ~s;
6427 /* We should invalidate potential inheritance or
6428 splitting for the current insn usages to the next
6429 usage insns (see code below) as the output pseudo
6430 prevents this. */
6431 if ((dst_regno >= FIRST_PSEUDO_REGISTER
6432 && reg_renumber[dst_regno] < 0)
6433 || (reg->type == OP_OUT && ! reg->subreg_p
6434 && (dst_regno < FIRST_PSEUDO_REGISTER
6435 || reg_renumber[dst_regno] >= 0)))
6437 /* Invalidate and mark definitions. */
6438 if (dst_regno >= FIRST_PSEUDO_REGISTER)
6439 usage_insns[dst_regno].check = -(int) INSN_UID (curr_insn);
6440 else
6442 nregs = hard_regno_nregs (dst_regno,
6443 reg->biggest_mode);
6444 for (i = 0; i < nregs; i++)
6445 usage_insns[dst_regno + i].check
6446 = -(int) INSN_UID (curr_insn);
6450 /* Process clobbered call regs. */
6451 if (curr_id->arg_hard_regs != NULL)
6452 for (i = 0; (dst_regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6453 if (dst_regno >= FIRST_PSEUDO_REGISTER)
6454 usage_insns[dst_regno - FIRST_PSEUDO_REGISTER].check
6455 = -(int) INSN_UID (curr_insn);
6456 if (! JUMP_P (curr_insn))
6457 for (i = 0; i < to_inherit_num; i++)
6458 if (inherit_reload_reg (true, to_inherit[i].regno,
6459 ALL_REGS, curr_insn,
6460 to_inherit[i].insns))
6461 change_p = true;
6462 if (CALL_P (curr_insn))
6464 rtx cheap, pat, dest;
6465 rtx_insn *restore;
6466 int regno, hard_regno;
6468 calls_num++;
6469 function_abi callee_abi = insn_callee_abi (curr_insn);
6470 last_call_for_abi[callee_abi.id ()] = calls_num;
6471 full_and_partial_call_clobbers
6472 |= callee_abi.full_and_partial_reg_clobbers ();
6473 if ((cheap = find_reg_note (curr_insn,
6474 REG_RETURNED, NULL_RTX)) != NULL_RTX
6475 && ((cheap = XEXP (cheap, 0)), true)
6476 && (regno = REGNO (cheap)) >= FIRST_PSEUDO_REGISTER
6477 && (hard_regno = reg_renumber[regno]) >= 0
6478 && usage_insns[regno].check == curr_usage_insns_check
6479 /* If there are pending saves/restores, the
6480 optimization is not worth. */
6481 && usage_insns[regno].calls_num == calls_num - 1
6482 && callee_abi.clobbers_reg_p (GET_MODE (cheap), hard_regno))
6484 /* Restore the pseudo from the call result as
6485 REG_RETURNED note says that the pseudo value is
6486 in the call result and the pseudo is an argument
6487 of the call. */
6488 pat = PATTERN (curr_insn);
6489 if (GET_CODE (pat) == PARALLEL)
6490 pat = XVECEXP (pat, 0, 0);
6491 dest = SET_DEST (pat);
6492 /* For multiple return values dest is PARALLEL.
6493 Currently we handle only single return value case. */
6494 if (REG_P (dest))
6496 start_sequence ();
6497 emit_move_insn (cheap, copy_rtx (dest));
6498 restore = get_insns ();
6499 end_sequence ();
6500 lra_process_new_insns (curr_insn, NULL, restore,
6501 "Inserting call parameter restore");
6502 /* We don't need to save/restore of the pseudo from
6503 this call. */
6504 usage_insns[regno].calls_num = calls_num;
6505 remove_from_hard_reg_set
6506 (&full_and_partial_call_clobbers,
6507 GET_MODE (cheap), hard_regno);
6508 bitmap_set_bit (&check_only_regs, regno);
6512 to_inherit_num = 0;
6513 /* Process insn usages. */
6514 for (iter = 0; iter < 2; iter++)
6515 for (reg = iter == 0 ? curr_id->regs : curr_static_id->hard_regs;
6516 reg != NULL;
6517 reg = reg->next)
6518 if ((reg->type != OP_OUT
6519 || (reg->type == OP_OUT && reg->subreg_p))
6520 && (src_regno = reg->regno) < lra_constraint_new_regno_start)
6522 if (src_regno >= FIRST_PSEUDO_REGISTER
6523 && reg_renumber[src_regno] < 0 && reg->type == OP_IN)
6525 if (usage_insns[src_regno].check == curr_usage_insns_check
6526 && (next_usage_insns
6527 = usage_insns[src_regno].insns) != NULL_RTX
6528 && NONDEBUG_INSN_P (curr_insn))
6529 add_to_inherit (src_regno, next_usage_insns);
6530 else if (usage_insns[src_regno].check
6531 != -(int) INSN_UID (curr_insn))
6532 /* Add usages but only if the reg is not set up
6533 in the same insn. */
6534 add_next_usage_insn (src_regno, curr_insn, reloads_num);
6536 else if (src_regno < FIRST_PSEUDO_REGISTER
6537 || reg_renumber[src_regno] >= 0)
6539 bool before_p;
6540 rtx_insn *use_insn = curr_insn;
6542 before_p = (JUMP_P (curr_insn)
6543 || (CALL_P (curr_insn) && reg->type == OP_IN));
6544 if (NONDEBUG_INSN_P (curr_insn)
6545 && (! JUMP_P (curr_insn) || reg->type == OP_IN)
6546 && split_if_necessary (src_regno, reg->biggest_mode,
6547 potential_reload_hard_regs,
6548 before_p, curr_insn, max_uid))
6550 if (reg->subreg_p)
6551 lra_risky_transformations_p = true;
6552 change_p = true;
6553 /* Invalidate. */
6554 usage_insns[src_regno].check = 0;
6555 if (before_p)
6556 use_insn = PREV_INSN (curr_insn);
6558 if (NONDEBUG_INSN_P (curr_insn))
6560 if (src_regno < FIRST_PSEUDO_REGISTER)
6561 add_to_hard_reg_set (&live_hard_regs,
6562 reg->biggest_mode, src_regno);
6563 else
6564 add_to_hard_reg_set (&live_hard_regs,
6565 PSEUDO_REGNO_MODE (src_regno),
6566 reg_renumber[src_regno]);
6568 if (src_regno >= FIRST_PSEUDO_REGISTER)
6569 add_next_usage_insn (src_regno, use_insn, reloads_num);
6570 else
6572 for (i = 0; i < hard_regno_nregs (src_regno, reg->biggest_mode); i++)
6573 add_next_usage_insn (src_regno + i, use_insn, reloads_num);
6577 /* Process used call regs. */
6578 if (curr_id->arg_hard_regs != NULL)
6579 for (i = 0; (src_regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6580 if (src_regno < FIRST_PSEUDO_REGISTER)
6582 SET_HARD_REG_BIT (live_hard_regs, src_regno);
6583 add_next_usage_insn (src_regno, curr_insn, reloads_num);
6585 for (i = 0; i < to_inherit_num; i++)
6587 src_regno = to_inherit[i].regno;
6588 if (inherit_reload_reg (false, src_regno, ALL_REGS,
6589 curr_insn, to_inherit[i].insns))
6590 change_p = true;
6591 else
6592 setup_next_usage_insn (src_regno, curr_insn, reloads_num, false);
6595 if (update_reloads_num_p
6596 && NONDEBUG_INSN_P (curr_insn) && curr_set != NULL_RTX)
6598 int regno = -1;
6599 if ((REG_P (SET_DEST (curr_set))
6600 && (regno = REGNO (SET_DEST (curr_set))) >= lra_constraint_new_regno_start
6601 && reg_renumber[regno] < 0
6602 && (cl = lra_get_allocno_class (regno)) != NO_REGS)
6603 || (REG_P (SET_SRC (curr_set))
6604 && (regno = REGNO (SET_SRC (curr_set))) >= lra_constraint_new_regno_start
6605 && reg_renumber[regno] < 0
6606 && (cl = lra_get_allocno_class (regno)) != NO_REGS))
6608 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6609 reloads_num++;
6610 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6611 potential_reload_hard_regs |= reg_class_contents[cl];
6614 if (NONDEBUG_INSN_P (curr_insn))
6616 int regno;
6618 /* Invalidate invariants with changed regs. */
6619 curr_id = lra_get_insn_recog_data (curr_insn);
6620 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
6621 if (reg->type != OP_IN)
6623 bitmap_set_bit (&invalid_invariant_regs, reg->regno);
6624 bitmap_set_bit (&invalid_invariant_regs,
6625 ORIGINAL_REGNO (regno_reg_rtx[reg->regno]));
6627 curr_static_id = curr_id->insn_static_data;
6628 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
6629 if (reg->type != OP_IN)
6630 bitmap_set_bit (&invalid_invariant_regs, reg->regno);
6631 if (curr_id->arg_hard_regs != NULL)
6632 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6633 if (regno >= FIRST_PSEUDO_REGISTER)
6634 bitmap_set_bit (&invalid_invariant_regs,
6635 regno - FIRST_PSEUDO_REGISTER);
6637 /* We reached the start of the current basic block. */
6638 if (prev_insn == NULL_RTX || prev_insn == PREV_INSN (head)
6639 || BLOCK_FOR_INSN (prev_insn) != curr_bb)
6641 /* We reached the beginning of the current block -- do
6642 rest of spliting in the current BB. */
6643 to_process = df_get_live_in (curr_bb);
6644 if (BLOCK_FOR_INSN (head) != curr_bb)
6646 /* We are somewhere in the middle of EBB. */
6647 get_live_on_other_edges (EDGE_PRED (curr_bb, 0)->src,
6648 curr_bb, &temp_bitmap);
6649 to_process = &temp_bitmap;
6651 head_p = true;
6652 EXECUTE_IF_SET_IN_BITMAP (to_process, 0, j, bi)
6654 if ((int) j >= lra_constraint_new_regno_start)
6655 break;
6656 if (((int) j < FIRST_PSEUDO_REGISTER || reg_renumber[j] >= 0)
6657 && usage_insns[j].check == curr_usage_insns_check
6658 && (next_usage_insns = usage_insns[j].insns) != NULL_RTX)
6660 if (need_for_split_p (potential_reload_hard_regs, j))
6662 if (lra_dump_file != NULL && head_p)
6664 fprintf (lra_dump_file,
6665 " ----------------------------------\n");
6666 head_p = false;
6668 if (split_reg (false, j, bb_note (curr_bb),
6669 next_usage_insns, NULL))
6670 change_p = true;
6672 usage_insns[j].check = 0;
6677 return change_p;
6680 /* This value affects EBB forming. If probability of edge from EBB to
6681 a BB is not greater than the following value, we don't add the BB
6682 to EBB. */
6683 #define EBB_PROBABILITY_CUTOFF \
6684 ((REG_BR_PROB_BASE * param_lra_inheritance_ebb_probability_cutoff) / 100)
6686 /* Current number of inheritance/split iteration. */
6687 int lra_inheritance_iter;
6689 /* Entry function for inheritance/split pass. */
6690 void
6691 lra_inheritance (void)
6693 int i;
6694 basic_block bb, start_bb;
6695 edge e;
6697 lra_inheritance_iter++;
6698 if (lra_inheritance_iter > LRA_MAX_INHERITANCE_PASSES)
6699 return;
6700 timevar_push (TV_LRA_INHERITANCE);
6701 if (lra_dump_file != NULL)
6702 fprintf (lra_dump_file, "\n********** Inheritance #%d: **********\n\n",
6703 lra_inheritance_iter);
6704 curr_usage_insns_check = 0;
6705 usage_insns = XNEWVEC (struct usage_insns, lra_constraint_new_regno_start);
6706 for (i = 0; i < lra_constraint_new_regno_start; i++)
6707 usage_insns[i].check = 0;
6708 bitmap_initialize (&check_only_regs, &reg_obstack);
6709 bitmap_initialize (&invalid_invariant_regs, &reg_obstack);
6710 bitmap_initialize (&live_regs, &reg_obstack);
6711 bitmap_initialize (&temp_bitmap, &reg_obstack);
6712 bitmap_initialize (&ebb_global_regs, &reg_obstack);
6713 FOR_EACH_BB_FN (bb, cfun)
6715 start_bb = bb;
6716 if (lra_dump_file != NULL)
6717 fprintf (lra_dump_file, "EBB");
6718 /* Form a EBB starting with BB. */
6719 bitmap_clear (&ebb_global_regs);
6720 bitmap_ior_into (&ebb_global_regs, df_get_live_in (bb));
6721 for (;;)
6723 if (lra_dump_file != NULL)
6724 fprintf (lra_dump_file, " %d", bb->index);
6725 if (bb->next_bb == EXIT_BLOCK_PTR_FOR_FN (cfun)
6726 || LABEL_P (BB_HEAD (bb->next_bb)))
6727 break;
6728 e = find_fallthru_edge (bb->succs);
6729 if (! e)
6730 break;
6731 if (e->probability.initialized_p ()
6732 && e->probability.to_reg_br_prob_base () < EBB_PROBABILITY_CUTOFF)
6733 break;
6734 bb = bb->next_bb;
6736 bitmap_ior_into (&ebb_global_regs, df_get_live_out (bb));
6737 if (lra_dump_file != NULL)
6738 fprintf (lra_dump_file, "\n");
6739 if (inherit_in_ebb (BB_HEAD (start_bb), BB_END (bb)))
6740 /* Remember that the EBB head and tail can change in
6741 inherit_in_ebb. */
6742 update_ebb_live_info (BB_HEAD (start_bb), BB_END (bb));
6744 bitmap_release (&ebb_global_regs);
6745 bitmap_release (&temp_bitmap);
6746 bitmap_release (&live_regs);
6747 bitmap_release (&invalid_invariant_regs);
6748 bitmap_release (&check_only_regs);
6749 free (usage_insns);
6751 timevar_pop (TV_LRA_INHERITANCE);
6756 /* This page contains code to undo failed inheritance/split
6757 transformations. */
6759 /* Current number of iteration undoing inheritance/split. */
6760 int lra_undo_inheritance_iter;
6762 /* Fix BB live info LIVE after removing pseudos created on pass doing
6763 inheritance/split which are REMOVED_PSEUDOS. */
6764 static void
6765 fix_bb_live_info (bitmap live, bitmap removed_pseudos)
6767 unsigned int regno;
6768 bitmap_iterator bi;
6770 EXECUTE_IF_SET_IN_BITMAP (removed_pseudos, 0, regno, bi)
6771 if (bitmap_clear_bit (live, regno)
6772 && REG_P (lra_reg_info[regno].restore_rtx))
6773 bitmap_set_bit (live, REGNO (lra_reg_info[regno].restore_rtx));
6776 /* Return regno of the (subreg of) REG. Otherwise, return a negative
6777 number. */
6778 static int
6779 get_regno (rtx reg)
6781 if (GET_CODE (reg) == SUBREG)
6782 reg = SUBREG_REG (reg);
6783 if (REG_P (reg))
6784 return REGNO (reg);
6785 return -1;
6788 /* Delete a move INSN with destination reg DREGNO and a previous
6789 clobber insn with the same regno. The inheritance/split code can
6790 generate moves with preceding clobber and when we delete such moves
6791 we should delete the clobber insn too to keep the correct life
6792 info. */
6793 static void
6794 delete_move_and_clobber (rtx_insn *insn, int dregno)
6796 rtx_insn *prev_insn = PREV_INSN (insn);
6798 lra_set_insn_deleted (insn);
6799 lra_assert (dregno >= 0);
6800 if (prev_insn != NULL && NONDEBUG_INSN_P (prev_insn)
6801 && GET_CODE (PATTERN (prev_insn)) == CLOBBER
6802 && dregno == get_regno (XEXP (PATTERN (prev_insn), 0)))
6803 lra_set_insn_deleted (prev_insn);
6806 /* Remove inheritance/split pseudos which are in REMOVE_PSEUDOS and
6807 return true if we did any change. The undo transformations for
6808 inheritance looks like
6809 i <- i2
6810 p <- i => p <- i2
6811 or removing
6812 p <- i, i <- p, and i <- i3
6813 where p is original pseudo from which inheritance pseudo i was
6814 created, i and i3 are removed inheritance pseudos, i2 is another
6815 not removed inheritance pseudo. All split pseudos or other
6816 occurrences of removed inheritance pseudos are changed on the
6817 corresponding original pseudos.
6819 The function also schedules insns changed and created during
6820 inheritance/split pass for processing by the subsequent constraint
6821 pass. */
6822 static bool
6823 remove_inheritance_pseudos (bitmap remove_pseudos)
6825 basic_block bb;
6826 int regno, sregno, prev_sregno, dregno;
6827 rtx restore_rtx;
6828 rtx set, prev_set;
6829 rtx_insn *prev_insn;
6830 bool change_p, done_p;
6832 change_p = ! bitmap_empty_p (remove_pseudos);
6833 /* We cannot finish the function right away if CHANGE_P is true
6834 because we need to marks insns affected by previous
6835 inheritance/split pass for processing by the subsequent
6836 constraint pass. */
6837 FOR_EACH_BB_FN (bb, cfun)
6839 fix_bb_live_info (df_get_live_in (bb), remove_pseudos);
6840 fix_bb_live_info (df_get_live_out (bb), remove_pseudos);
6841 FOR_BB_INSNS_REVERSE (bb, curr_insn)
6843 if (! INSN_P (curr_insn))
6844 continue;
6845 done_p = false;
6846 sregno = dregno = -1;
6847 if (change_p && NONDEBUG_INSN_P (curr_insn)
6848 && (set = single_set (curr_insn)) != NULL_RTX)
6850 dregno = get_regno (SET_DEST (set));
6851 sregno = get_regno (SET_SRC (set));
6854 if (sregno >= 0 && dregno >= 0)
6856 if (bitmap_bit_p (remove_pseudos, dregno)
6857 && ! REG_P (lra_reg_info[dregno].restore_rtx))
6859 /* invariant inheritance pseudo <- original pseudo */
6860 if (lra_dump_file != NULL)
6862 fprintf (lra_dump_file, " Removing invariant inheritance:\n");
6863 dump_insn_slim (lra_dump_file, curr_insn);
6864 fprintf (lra_dump_file, "\n");
6866 delete_move_and_clobber (curr_insn, dregno);
6867 done_p = true;
6869 else if (bitmap_bit_p (remove_pseudos, sregno)
6870 && ! REG_P (lra_reg_info[sregno].restore_rtx))
6872 /* reload pseudo <- invariant inheritance pseudo */
6873 start_sequence ();
6874 /* We cannot just change the source. It might be
6875 an insn different from the move. */
6876 emit_insn (lra_reg_info[sregno].restore_rtx);
6877 rtx_insn *new_insns = get_insns ();
6878 end_sequence ();
6879 lra_assert (single_set (new_insns) != NULL
6880 && SET_DEST (set) == SET_DEST (single_set (new_insns)));
6881 lra_process_new_insns (curr_insn, NULL, new_insns,
6882 "Changing reload<-invariant inheritance");
6883 delete_move_and_clobber (curr_insn, dregno);
6884 done_p = true;
6886 else if ((bitmap_bit_p (remove_pseudos, sregno)
6887 && (get_regno (lra_reg_info[sregno].restore_rtx) == dregno
6888 || (bitmap_bit_p (remove_pseudos, dregno)
6889 && get_regno (lra_reg_info[sregno].restore_rtx) >= 0
6890 && (get_regno (lra_reg_info[sregno].restore_rtx)
6891 == get_regno (lra_reg_info[dregno].restore_rtx)))))
6892 || (bitmap_bit_p (remove_pseudos, dregno)
6893 && get_regno (lra_reg_info[dregno].restore_rtx) == sregno))
6894 /* One of the following cases:
6895 original <- removed inheritance pseudo
6896 removed inherit pseudo <- another removed inherit pseudo
6897 removed inherit pseudo <- original pseudo
6899 removed_split_pseudo <- original_reg
6900 original_reg <- removed_split_pseudo */
6902 if (lra_dump_file != NULL)
6904 fprintf (lra_dump_file, " Removing %s:\n",
6905 bitmap_bit_p (&lra_split_regs, sregno)
6906 || bitmap_bit_p (&lra_split_regs, dregno)
6907 ? "split" : "inheritance");
6908 dump_insn_slim (lra_dump_file, curr_insn);
6910 delete_move_and_clobber (curr_insn, dregno);
6911 done_p = true;
6913 else if (bitmap_bit_p (remove_pseudos, sregno)
6914 && bitmap_bit_p (&lra_inheritance_pseudos, sregno))
6916 /* Search the following pattern:
6917 inherit_or_split_pseudo1 <- inherit_or_split_pseudo2
6918 original_pseudo <- inherit_or_split_pseudo1
6919 where the 2nd insn is the current insn and
6920 inherit_or_split_pseudo2 is not removed. If it is found,
6921 change the current insn onto:
6922 original_pseudo <- inherit_or_split_pseudo2. */
6923 for (prev_insn = PREV_INSN (curr_insn);
6924 prev_insn != NULL_RTX && ! NONDEBUG_INSN_P (prev_insn);
6925 prev_insn = PREV_INSN (prev_insn))
6927 if (prev_insn != NULL_RTX && BLOCK_FOR_INSN (prev_insn) == bb
6928 && (prev_set = single_set (prev_insn)) != NULL_RTX
6929 /* There should be no subregs in insn we are
6930 searching because only the original reg might
6931 be in subreg when we changed the mode of
6932 load/store for splitting. */
6933 && REG_P (SET_DEST (prev_set))
6934 && REG_P (SET_SRC (prev_set))
6935 && (int) REGNO (SET_DEST (prev_set)) == sregno
6936 && ((prev_sregno = REGNO (SET_SRC (prev_set)))
6937 >= FIRST_PSEUDO_REGISTER)
6938 && (lra_reg_info[prev_sregno].restore_rtx == NULL_RTX
6940 /* As we consider chain of inheritance or
6941 splitting described in above comment we should
6942 check that sregno and prev_sregno were
6943 inheritance/split pseudos created from the
6944 same original regno. */
6945 (get_regno (lra_reg_info[sregno].restore_rtx) >= 0
6946 && (get_regno (lra_reg_info[sregno].restore_rtx)
6947 == get_regno (lra_reg_info[prev_sregno].restore_rtx))))
6948 && ! bitmap_bit_p (remove_pseudos, prev_sregno))
6950 lra_assert (GET_MODE (SET_SRC (prev_set))
6951 == GET_MODE (regno_reg_rtx[sregno]));
6952 /* Although we have a single set, the insn can
6953 contain more one sregno register occurrence
6954 as a source. Change all occurrences. */
6955 lra_substitute_pseudo_within_insn (curr_insn, sregno,
6956 SET_SRC (prev_set),
6957 false);
6958 /* As we are finishing with processing the insn
6959 here, check the destination too as it might
6960 inheritance pseudo for another pseudo. */
6961 if (bitmap_bit_p (remove_pseudos, dregno)
6962 && bitmap_bit_p (&lra_inheritance_pseudos, dregno)
6963 && (restore_rtx
6964 = lra_reg_info[dregno].restore_rtx) != NULL_RTX)
6966 if (GET_CODE (SET_DEST (set)) == SUBREG)
6967 SUBREG_REG (SET_DEST (set)) = restore_rtx;
6968 else
6969 SET_DEST (set) = restore_rtx;
6971 lra_push_insn_and_update_insn_regno_info (curr_insn);
6972 lra_set_used_insn_alternative_by_uid
6973 (INSN_UID (curr_insn), LRA_UNKNOWN_ALT);
6974 done_p = true;
6975 if (lra_dump_file != NULL)
6977 fprintf (lra_dump_file, " Change reload insn:\n");
6978 dump_insn_slim (lra_dump_file, curr_insn);
6983 if (! done_p)
6985 struct lra_insn_reg *reg;
6986 bool restored_regs_p = false;
6987 bool kept_regs_p = false;
6989 curr_id = lra_get_insn_recog_data (curr_insn);
6990 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
6992 regno = reg->regno;
6993 restore_rtx = lra_reg_info[regno].restore_rtx;
6994 if (restore_rtx != NULL_RTX)
6996 if (change_p && bitmap_bit_p (remove_pseudos, regno))
6998 lra_substitute_pseudo_within_insn
6999 (curr_insn, regno, restore_rtx, false);
7000 restored_regs_p = true;
7002 else
7003 kept_regs_p = true;
7006 if (NONDEBUG_INSN_P (curr_insn) && kept_regs_p)
7008 /* The instruction has changed since the previous
7009 constraints pass. */
7010 lra_push_insn_and_update_insn_regno_info (curr_insn);
7011 lra_set_used_insn_alternative_by_uid
7012 (INSN_UID (curr_insn), LRA_UNKNOWN_ALT);
7014 else if (restored_regs_p)
7015 /* The instruction has been restored to the form that
7016 it had during the previous constraints pass. */
7017 lra_update_insn_regno_info (curr_insn);
7018 if (restored_regs_p && lra_dump_file != NULL)
7020 fprintf (lra_dump_file, " Insn after restoring regs:\n");
7021 dump_insn_slim (lra_dump_file, curr_insn);
7026 return change_p;
7029 /* If optional reload pseudos failed to get a hard register or was not
7030 inherited, it is better to remove optional reloads. We do this
7031 transformation after undoing inheritance to figure out necessity to
7032 remove optional reloads easier. Return true if we do any
7033 change. */
7034 static bool
7035 undo_optional_reloads (void)
7037 bool change_p, keep_p;
7038 unsigned int regno, uid;
7039 bitmap_iterator bi, bi2;
7040 rtx_insn *insn;
7041 rtx set, src, dest;
7042 auto_bitmap removed_optional_reload_pseudos (&reg_obstack);
7044 bitmap_copy (removed_optional_reload_pseudos, &lra_optional_reload_pseudos);
7045 EXECUTE_IF_SET_IN_BITMAP (&lra_optional_reload_pseudos, 0, regno, bi)
7047 keep_p = false;
7048 /* Keep optional reloads from previous subpasses. */
7049 if (lra_reg_info[regno].restore_rtx == NULL_RTX
7050 /* If the original pseudo changed its allocation, just
7051 removing the optional pseudo is dangerous as the original
7052 pseudo will have longer live range. */
7053 || reg_renumber[REGNO (lra_reg_info[regno].restore_rtx)] >= 0)
7054 keep_p = true;
7055 else if (reg_renumber[regno] >= 0)
7056 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi2)
7058 insn = lra_insn_recog_data[uid]->insn;
7059 if ((set = single_set (insn)) == NULL_RTX)
7060 continue;
7061 src = SET_SRC (set);
7062 dest = SET_DEST (set);
7063 if (! REG_P (src) || ! REG_P (dest))
7064 continue;
7065 if (REGNO (dest) == regno
7066 /* Ignore insn for optional reloads itself. */
7067 && REGNO (lra_reg_info[regno].restore_rtx) != REGNO (src)
7068 /* Check only inheritance on last inheritance pass. */
7069 && (int) REGNO (src) >= new_regno_start
7070 /* Check that the optional reload was inherited. */
7071 && bitmap_bit_p (&lra_inheritance_pseudos, REGNO (src)))
7073 keep_p = true;
7074 break;
7077 if (keep_p)
7079 bitmap_clear_bit (removed_optional_reload_pseudos, regno);
7080 if (lra_dump_file != NULL)
7081 fprintf (lra_dump_file, "Keep optional reload reg %d\n", regno);
7084 change_p = ! bitmap_empty_p (removed_optional_reload_pseudos);
7085 auto_bitmap insn_bitmap (&reg_obstack);
7086 EXECUTE_IF_SET_IN_BITMAP (removed_optional_reload_pseudos, 0, regno, bi)
7088 if (lra_dump_file != NULL)
7089 fprintf (lra_dump_file, "Remove optional reload reg %d\n", regno);
7090 bitmap_copy (insn_bitmap, &lra_reg_info[regno].insn_bitmap);
7091 EXECUTE_IF_SET_IN_BITMAP (insn_bitmap, 0, uid, bi2)
7093 insn = lra_insn_recog_data[uid]->insn;
7094 if ((set = single_set (insn)) != NULL_RTX)
7096 src = SET_SRC (set);
7097 dest = SET_DEST (set);
7098 if (REG_P (src) && REG_P (dest)
7099 && ((REGNO (src) == regno
7100 && (REGNO (lra_reg_info[regno].restore_rtx)
7101 == REGNO (dest)))
7102 || (REGNO (dest) == regno
7103 && (REGNO (lra_reg_info[regno].restore_rtx)
7104 == REGNO (src)))))
7106 if (lra_dump_file != NULL)
7108 fprintf (lra_dump_file, " Deleting move %u\n",
7109 INSN_UID (insn));
7110 dump_insn_slim (lra_dump_file, insn);
7112 delete_move_and_clobber (insn, REGNO (dest));
7113 continue;
7115 /* We should not worry about generation memory-memory
7116 moves here as if the corresponding inheritance did
7117 not work (inheritance pseudo did not get a hard reg),
7118 we remove the inheritance pseudo and the optional
7119 reload. */
7121 lra_substitute_pseudo_within_insn
7122 (insn, regno, lra_reg_info[regno].restore_rtx, false);
7123 lra_update_insn_regno_info (insn);
7124 if (lra_dump_file != NULL)
7126 fprintf (lra_dump_file,
7127 " Restoring original insn:\n");
7128 dump_insn_slim (lra_dump_file, insn);
7132 /* Clear restore_regnos. */
7133 EXECUTE_IF_SET_IN_BITMAP (&lra_optional_reload_pseudos, 0, regno, bi)
7134 lra_reg_info[regno].restore_rtx = NULL_RTX;
7135 return change_p;
7138 /* Entry function for undoing inheritance/split transformation. Return true
7139 if we did any RTL change in this pass. */
7140 bool
7141 lra_undo_inheritance (void)
7143 unsigned int regno;
7144 int hard_regno;
7145 int n_all_inherit, n_inherit, n_all_split, n_split;
7146 rtx restore_rtx;
7147 bitmap_iterator bi;
7148 bool change_p;
7150 lra_undo_inheritance_iter++;
7151 if (lra_undo_inheritance_iter > LRA_MAX_INHERITANCE_PASSES)
7152 return false;
7153 if (lra_dump_file != NULL)
7154 fprintf (lra_dump_file,
7155 "\n********** Undoing inheritance #%d: **********\n\n",
7156 lra_undo_inheritance_iter);
7157 auto_bitmap remove_pseudos (&reg_obstack);
7158 n_inherit = n_all_inherit = 0;
7159 EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, regno, bi)
7160 if (lra_reg_info[regno].restore_rtx != NULL_RTX)
7162 n_all_inherit++;
7163 if (reg_renumber[regno] < 0
7164 /* If the original pseudo changed its allocation, just
7165 removing inheritance is dangerous as for changing
7166 allocation we used shorter live-ranges. */
7167 && (! REG_P (lra_reg_info[regno].restore_rtx)
7168 || reg_renumber[REGNO (lra_reg_info[regno].restore_rtx)] < 0))
7169 bitmap_set_bit (remove_pseudos, regno);
7170 else
7171 n_inherit++;
7173 if (lra_dump_file != NULL && n_all_inherit != 0)
7174 fprintf (lra_dump_file, "Inherit %d out of %d (%.2f%%)\n",
7175 n_inherit, n_all_inherit,
7176 (double) n_inherit / n_all_inherit * 100);
7177 n_split = n_all_split = 0;
7178 EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, regno, bi)
7179 if ((restore_rtx = lra_reg_info[regno].restore_rtx) != NULL_RTX)
7181 int restore_regno = REGNO (restore_rtx);
7183 n_all_split++;
7184 hard_regno = (restore_regno >= FIRST_PSEUDO_REGISTER
7185 ? reg_renumber[restore_regno] : restore_regno);
7186 if (hard_regno < 0 || reg_renumber[regno] == hard_regno)
7187 bitmap_set_bit (remove_pseudos, regno);
7188 else
7190 n_split++;
7191 if (lra_dump_file != NULL)
7192 fprintf (lra_dump_file, " Keep split r%d (orig=r%d)\n",
7193 regno, restore_regno);
7196 if (lra_dump_file != NULL && n_all_split != 0)
7197 fprintf (lra_dump_file, "Split %d out of %d (%.2f%%)\n",
7198 n_split, n_all_split,
7199 (double) n_split / n_all_split * 100);
7200 change_p = remove_inheritance_pseudos (remove_pseudos);
7201 /* Clear restore_regnos. */
7202 EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, regno, bi)
7203 lra_reg_info[regno].restore_rtx = NULL_RTX;
7204 EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, regno, bi)
7205 lra_reg_info[regno].restore_rtx = NULL_RTX;
7206 change_p = undo_optional_reloads () || change_p;
7207 return change_p;