gcc/ChangeLog:
[official-gcc.git] / gcc / lra-constraints.c
blob25869134bd0d1fe5d16078888b303f8928324759
1 /* Code for RTL transformations to satisfy insn constraints.
2 Copyright (C) 2010-2017 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 "params.h"
131 #include "lra.h"
132 #include "lra-int.h"
133 #include "print-rtl.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 lra_eliminate_reg_if_possible (m_base_loc);
363 if (m_ad->base_term2 != NULL)
364 *m_ad->base_term2 = *m_ad->base_term;
366 if (m_index_loc != NULL)
368 m_index_reg = *m_index_loc;
369 lra_eliminate_reg_if_possible (m_index_loc);
373 address_eliminator::~address_eliminator ()
375 if (m_base_loc && *m_base_loc != m_base_reg)
377 *m_base_loc = m_base_reg;
378 if (m_ad->base_term2 != NULL)
379 *m_ad->base_term2 = *m_ad->base_term;
381 if (m_index_loc && *m_index_loc != m_index_reg)
382 *m_index_loc = m_index_reg;
385 /* Return true if the eliminated form of AD is a legitimate target address. */
386 static bool
387 valid_address_p (struct address_info *ad)
389 address_eliminator eliminator (ad);
390 return valid_address_p (ad->mode, *ad->outer, ad->as);
393 /* Return true if the eliminated form of memory reference OP satisfies
394 extra (special) memory constraint CONSTRAINT. */
395 static bool
396 satisfies_memory_constraint_p (rtx op, enum constraint_num constraint)
398 struct address_info ad;
400 decompose_mem_address (&ad, op);
401 address_eliminator eliminator (&ad);
402 return constraint_satisfied_p (op, constraint);
405 /* Return true if the eliminated form of address AD satisfies extra
406 address constraint CONSTRAINT. */
407 static bool
408 satisfies_address_constraint_p (struct address_info *ad,
409 enum constraint_num constraint)
411 address_eliminator eliminator (ad);
412 return constraint_satisfied_p (*ad->outer, constraint);
415 /* Return true if the eliminated form of address OP satisfies extra
416 address constraint CONSTRAINT. */
417 static bool
418 satisfies_address_constraint_p (rtx op, enum constraint_num constraint)
420 struct address_info ad;
422 decompose_lea_address (&ad, &op);
423 return satisfies_address_constraint_p (&ad, constraint);
426 /* Initiate equivalences for LRA. As we keep original equivalences
427 before any elimination, we need to make copies otherwise any change
428 in insns might change the equivalences. */
429 void
430 lra_init_equiv (void)
432 ira_expand_reg_equiv ();
433 for (int i = FIRST_PSEUDO_REGISTER; i < max_reg_num (); i++)
435 rtx res;
437 if ((res = ira_reg_equiv[i].memory) != NULL_RTX)
438 ira_reg_equiv[i].memory = copy_rtx (res);
439 if ((res = ira_reg_equiv[i].invariant) != NULL_RTX)
440 ira_reg_equiv[i].invariant = copy_rtx (res);
444 static rtx loc_equivalence_callback (rtx, const_rtx, void *);
446 /* Update equivalence for REGNO. We need to this as the equivalence
447 might contain other pseudos which are changed by their
448 equivalences. */
449 static void
450 update_equiv (int regno)
452 rtx x;
454 if ((x = ira_reg_equiv[regno].memory) != NULL_RTX)
455 ira_reg_equiv[regno].memory
456 = simplify_replace_fn_rtx (x, NULL_RTX, loc_equivalence_callback,
457 NULL_RTX);
458 if ((x = ira_reg_equiv[regno].invariant) != NULL_RTX)
459 ira_reg_equiv[regno].invariant
460 = simplify_replace_fn_rtx (x, NULL_RTX, loc_equivalence_callback,
461 NULL_RTX);
464 /* If we have decided to substitute X with another value, return that
465 value, otherwise return X. */
466 static rtx
467 get_equiv (rtx x)
469 int regno;
470 rtx res;
472 if (! REG_P (x) || (regno = REGNO (x)) < FIRST_PSEUDO_REGISTER
473 || ! ira_reg_equiv[regno].defined_p
474 || ! ira_reg_equiv[regno].profitable_p
475 || lra_get_regno_hard_regno (regno) >= 0)
476 return x;
477 if ((res = ira_reg_equiv[regno].memory) != NULL_RTX)
479 if (targetm.cannot_substitute_mem_equiv_p (res))
480 return x;
481 return res;
483 if ((res = ira_reg_equiv[regno].constant) != NULL_RTX)
484 return res;
485 if ((res = ira_reg_equiv[regno].invariant) != NULL_RTX)
486 return res;
487 gcc_unreachable ();
490 /* If we have decided to substitute X with the equivalent value,
491 return that value after elimination for INSN, otherwise return
492 X. */
493 static rtx
494 get_equiv_with_elimination (rtx x, rtx_insn *insn)
496 rtx res = get_equiv (x);
498 if (x == res || CONSTANT_P (res))
499 return res;
500 return lra_eliminate_regs_1 (insn, res, GET_MODE (res),
501 false, false, 0, true);
504 /* Set up curr_operand_mode. */
505 static void
506 init_curr_operand_mode (void)
508 int nop = curr_static_id->n_operands;
509 for (int i = 0; i < nop; i++)
511 machine_mode mode = GET_MODE (*curr_id->operand_loc[i]);
512 if (mode == VOIDmode)
514 /* The .md mode for address operands is the mode of the
515 addressed value rather than the mode of the address itself. */
516 if (curr_id->icode >= 0 && curr_static_id->operand[i].is_address)
517 mode = Pmode;
518 else
519 mode = curr_static_id->operand[i].mode;
521 curr_operand_mode[i] = mode;
527 /* The page contains code to reuse input reloads. */
529 /* Structure describes input reload of the current insns. */
530 struct input_reload
532 /* True for input reload of matched operands. */
533 bool match_p;
534 /* Reloaded value. */
535 rtx input;
536 /* Reload pseudo used. */
537 rtx reg;
540 /* The number of elements in the following array. */
541 static int curr_insn_input_reloads_num;
542 /* Array containing info about input reloads. It is used to find the
543 same input reload and reuse the reload pseudo in this case. */
544 static struct input_reload curr_insn_input_reloads[LRA_MAX_INSN_RELOADS];
546 /* Initiate data concerning reuse of input reloads for the current
547 insn. */
548 static void
549 init_curr_insn_input_reloads (void)
551 curr_insn_input_reloads_num = 0;
554 /* Create a new pseudo using MODE, RCLASS, ORIGINAL or reuse already
555 created input reload pseudo (only if TYPE is not OP_OUT). Don't
556 reuse pseudo if IN_SUBREG_P is true and the reused pseudo should be
557 wrapped up in SUBREG. The result pseudo is returned through
558 RESULT_REG. Return TRUE if we created a new pseudo, FALSE if we
559 reused the already created input reload pseudo. Use TITLE to
560 describe new registers for debug purposes. */
561 static bool
562 get_reload_reg (enum op_type type, machine_mode mode, rtx original,
563 enum reg_class rclass, bool in_subreg_p,
564 const char *title, rtx *result_reg)
566 int i, regno;
567 enum reg_class new_class;
568 bool unique_p = false;
570 if (type == OP_OUT)
572 *result_reg
573 = lra_create_new_reg_with_unique_value (mode, original, rclass, title);
574 return true;
576 /* Prevent reuse value of expression with side effects,
577 e.g. volatile memory. */
578 if (! side_effects_p (original))
579 for (i = 0; i < curr_insn_input_reloads_num; i++)
581 if (! curr_insn_input_reloads[i].match_p
582 && rtx_equal_p (curr_insn_input_reloads[i].input, original)
583 && in_class_p (curr_insn_input_reloads[i].reg, rclass, &new_class))
585 rtx reg = curr_insn_input_reloads[i].reg;
586 regno = REGNO (reg);
587 /* If input is equal to original and both are VOIDmode,
588 GET_MODE (reg) might be still different from mode.
589 Ensure we don't return *result_reg with wrong mode. */
590 if (GET_MODE (reg) != mode)
592 if (in_subreg_p)
593 continue;
594 if (GET_MODE_SIZE (GET_MODE (reg)) < GET_MODE_SIZE (mode))
595 continue;
596 reg = lowpart_subreg (mode, reg, GET_MODE (reg));
597 if (reg == NULL_RTX || GET_CODE (reg) != SUBREG)
598 continue;
600 *result_reg = reg;
601 if (lra_dump_file != NULL)
603 fprintf (lra_dump_file, " Reuse r%d for reload ", regno);
604 dump_value_slim (lra_dump_file, original, 1);
606 if (new_class != lra_get_allocno_class (regno))
607 lra_change_class (regno, new_class, ", change to", false);
608 if (lra_dump_file != NULL)
609 fprintf (lra_dump_file, "\n");
610 return false;
612 /* If we have an input reload with a different mode, make sure it
613 will get a different hard reg. */
614 else if (REG_P (original)
615 && REG_P (curr_insn_input_reloads[i].input)
616 && REGNO (original) == REGNO (curr_insn_input_reloads[i].input)
617 && (GET_MODE (original)
618 != GET_MODE (curr_insn_input_reloads[i].input)))
619 unique_p = true;
621 *result_reg = (unique_p
622 ? lra_create_new_reg_with_unique_value
623 : lra_create_new_reg) (mode, original, rclass, title);
624 lra_assert (curr_insn_input_reloads_num < LRA_MAX_INSN_RELOADS);
625 curr_insn_input_reloads[curr_insn_input_reloads_num].input = original;
626 curr_insn_input_reloads[curr_insn_input_reloads_num].match_p = false;
627 curr_insn_input_reloads[curr_insn_input_reloads_num++].reg = *result_reg;
628 return true;
633 /* The page contains code to extract memory address parts. */
635 /* Wrapper around REGNO_OK_FOR_INDEX_P, to allow pseudos. */
636 static inline bool
637 ok_for_index_p_nonstrict (rtx reg)
639 unsigned regno = REGNO (reg);
641 return regno >= FIRST_PSEUDO_REGISTER || REGNO_OK_FOR_INDEX_P (regno);
644 /* A version of regno_ok_for_base_p for use here, when all pseudos
645 should count as OK. Arguments as for regno_ok_for_base_p. */
646 static inline bool
647 ok_for_base_p_nonstrict (rtx reg, machine_mode mode, addr_space_t as,
648 enum rtx_code outer_code, enum rtx_code index_code)
650 unsigned regno = REGNO (reg);
652 if (regno >= FIRST_PSEUDO_REGISTER)
653 return true;
654 return ok_for_base_p_1 (regno, mode, as, outer_code, index_code);
659 /* The page contains major code to choose the current insn alternative
660 and generate reloads for it. */
662 /* Return the offset from REGNO of the least significant register
663 in (reg:MODE REGNO).
665 This function is used to tell whether two registers satisfy
666 a matching constraint. (reg:MODE1 REGNO1) matches (reg:MODE2 REGNO2) if:
668 REGNO1 + lra_constraint_offset (REGNO1, MODE1)
669 == REGNO2 + lra_constraint_offset (REGNO2, MODE2) */
671 lra_constraint_offset (int regno, machine_mode mode)
673 lra_assert (regno < FIRST_PSEUDO_REGISTER);
674 if (WORDS_BIG_ENDIAN && GET_MODE_SIZE (mode) > UNITS_PER_WORD
675 && SCALAR_INT_MODE_P (mode))
676 return hard_regno_nregs[regno][mode] - 1;
677 return 0;
680 /* Like rtx_equal_p except that it allows a REG and a SUBREG to match
681 if they are the same hard reg, and has special hacks for
682 auto-increment and auto-decrement. This is specifically intended for
683 process_alt_operands to use in determining whether two operands
684 match. X is the operand whose number is the lower of the two.
686 It is supposed that X is the output operand and Y is the input
687 operand. Y_HARD_REGNO is the final hard regno of register Y or
688 register in subreg Y as we know it now. Otherwise, it is a
689 negative value. */
690 static bool
691 operands_match_p (rtx x, rtx y, int y_hard_regno)
693 int i;
694 RTX_CODE code = GET_CODE (x);
695 const char *fmt;
697 if (x == y)
698 return true;
699 if ((code == REG || (code == SUBREG && REG_P (SUBREG_REG (x))))
700 && (REG_P (y) || (GET_CODE (y) == SUBREG && REG_P (SUBREG_REG (y)))))
702 int j;
704 i = get_hard_regno (x, false);
705 if (i < 0)
706 goto slow;
708 if ((j = y_hard_regno) < 0)
709 goto slow;
711 i += lra_constraint_offset (i, GET_MODE (x));
712 j += lra_constraint_offset (j, GET_MODE (y));
714 return i == j;
717 /* If two operands must match, because they are really a single
718 operand of an assembler insn, then two post-increments are invalid
719 because the assembler insn would increment only once. On the
720 other hand, a post-increment matches ordinary indexing if the
721 post-increment is the output operand. */
722 if (code == POST_DEC || code == POST_INC || code == POST_MODIFY)
723 return operands_match_p (XEXP (x, 0), y, y_hard_regno);
725 /* Two pre-increments are invalid because the assembler insn would
726 increment only once. On the other hand, a pre-increment matches
727 ordinary indexing if the pre-increment is the input operand. */
728 if (GET_CODE (y) == PRE_DEC || GET_CODE (y) == PRE_INC
729 || GET_CODE (y) == PRE_MODIFY)
730 return operands_match_p (x, XEXP (y, 0), -1);
732 slow:
734 if (code == REG && REG_P (y))
735 return REGNO (x) == REGNO (y);
737 if (code == REG && GET_CODE (y) == SUBREG && REG_P (SUBREG_REG (y))
738 && x == SUBREG_REG (y))
739 return true;
740 if (GET_CODE (y) == REG && code == SUBREG && REG_P (SUBREG_REG (x))
741 && SUBREG_REG (x) == y)
742 return true;
744 /* Now we have disposed of all the cases in which different rtx
745 codes can match. */
746 if (code != GET_CODE (y))
747 return false;
749 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
750 if (GET_MODE (x) != GET_MODE (y))
751 return false;
753 switch (code)
755 CASE_CONST_UNIQUE:
756 return false;
758 case LABEL_REF:
759 return label_ref_label (x) == label_ref_label (y);
760 case SYMBOL_REF:
761 return XSTR (x, 0) == XSTR (y, 0);
763 default:
764 break;
767 /* Compare the elements. If any pair of corresponding elements fail
768 to match, return false for the whole things. */
770 fmt = GET_RTX_FORMAT (code);
771 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
773 int val, j;
774 switch (fmt[i])
776 case 'w':
777 if (XWINT (x, i) != XWINT (y, i))
778 return false;
779 break;
781 case 'i':
782 if (XINT (x, i) != XINT (y, i))
783 return false;
784 break;
786 case 'e':
787 val = operands_match_p (XEXP (x, i), XEXP (y, i), -1);
788 if (val == 0)
789 return false;
790 break;
792 case '0':
793 break;
795 case 'E':
796 if (XVECLEN (x, i) != XVECLEN (y, i))
797 return false;
798 for (j = XVECLEN (x, i) - 1; j >= 0; --j)
800 val = operands_match_p (XVECEXP (x, i, j), XVECEXP (y, i, j), -1);
801 if (val == 0)
802 return false;
804 break;
806 /* It is believed that rtx's at this level will never
807 contain anything but integers and other rtx's, except for
808 within LABEL_REFs and SYMBOL_REFs. */
809 default:
810 gcc_unreachable ();
813 return true;
816 /* True if X is a constant that can be forced into the constant pool.
817 MODE is the mode of the operand, or VOIDmode if not known. */
818 #define CONST_POOL_OK_P(MODE, X) \
819 ((MODE) != VOIDmode \
820 && CONSTANT_P (X) \
821 && GET_CODE (X) != HIGH \
822 && !targetm.cannot_force_const_mem (MODE, X))
824 /* True if C is a non-empty register class that has too few registers
825 to be safely used as a reload target class. */
826 #define SMALL_REGISTER_CLASS_P(C) \
827 (ira_class_hard_regs_num [(C)] == 1 \
828 || (ira_class_hard_regs_num [(C)] >= 1 \
829 && targetm.class_likely_spilled_p (C)))
831 /* If REG is a reload pseudo, try to make its class satisfying CL. */
832 static void
833 narrow_reload_pseudo_class (rtx reg, enum reg_class cl)
835 enum reg_class rclass;
837 /* Do not make more accurate class from reloads generated. They are
838 mostly moves with a lot of constraints. Making more accurate
839 class may results in very narrow class and impossibility of find
840 registers for several reloads of one insn. */
841 if (INSN_UID (curr_insn) >= new_insn_uid_start)
842 return;
843 if (GET_CODE (reg) == SUBREG)
844 reg = SUBREG_REG (reg);
845 if (! REG_P (reg) || (int) REGNO (reg) < new_regno_start)
846 return;
847 if (in_class_p (reg, cl, &rclass) && rclass != cl)
848 lra_change_class (REGNO (reg), rclass, " Change to", true);
851 /* Searches X for any reference to a reg with the same value as REGNO,
852 returning the rtx of the reference found if any. Otherwise,
853 returns NULL_RTX. */
854 static rtx
855 regno_val_use_in (unsigned int regno, rtx x)
857 const char *fmt;
858 int i, j;
859 rtx tem;
861 if (REG_P (x) && lra_reg_info[REGNO (x)].val == lra_reg_info[regno].val)
862 return x;
864 fmt = GET_RTX_FORMAT (GET_CODE (x));
865 for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
867 if (fmt[i] == 'e')
869 if ((tem = regno_val_use_in (regno, XEXP (x, i))))
870 return tem;
872 else if (fmt[i] == 'E')
873 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
874 if ((tem = regno_val_use_in (regno , XVECEXP (x, i, j))))
875 return tem;
878 return NULL_RTX;
881 /* Generate reloads for matching OUT and INS (array of input operand
882 numbers with end marker -1) with reg class GOAL_CLASS, considering
883 output operands OUTS (similar array to INS) needing to be in different
884 registers. Add input and output reloads correspondingly to the lists
885 *BEFORE and *AFTER. OUT might be negative. In this case we generate
886 input reloads for matched input operands INS. EARLY_CLOBBER_P is a flag
887 that the output operand is early clobbered for chosen alternative. */
888 static void
889 match_reload (signed char out, signed char *ins, signed char *outs,
890 enum reg_class goal_class, rtx_insn **before,
891 rtx_insn **after, bool early_clobber_p)
893 bool out_conflict;
894 int i, in;
895 rtx new_in_reg, new_out_reg, reg;
896 machine_mode inmode, outmode;
897 rtx in_rtx = *curr_id->operand_loc[ins[0]];
898 rtx out_rtx = out < 0 ? in_rtx : *curr_id->operand_loc[out];
900 inmode = curr_operand_mode[ins[0]];
901 outmode = out < 0 ? inmode : curr_operand_mode[out];
902 push_to_sequence (*before);
903 if (inmode != outmode)
905 if (GET_MODE_SIZE (inmode) > GET_MODE_SIZE (outmode))
907 reg = new_in_reg
908 = lra_create_new_reg_with_unique_value (inmode, in_rtx,
909 goal_class, "");
910 if (SCALAR_INT_MODE_P (inmode))
911 new_out_reg = gen_lowpart_SUBREG (outmode, reg);
912 else
913 new_out_reg = gen_rtx_SUBREG (outmode, reg, 0);
914 LRA_SUBREG_P (new_out_reg) = 1;
915 /* If the input reg is dying here, we can use the same hard
916 register for REG and IN_RTX. We do it only for original
917 pseudos as reload pseudos can die although original
918 pseudos still live where reload pseudos dies. */
919 if (REG_P (in_rtx) && (int) REGNO (in_rtx) < lra_new_regno_start
920 && find_regno_note (curr_insn, REG_DEAD, REGNO (in_rtx)))
921 lra_assign_reg_val (REGNO (in_rtx), REGNO (reg));
923 else
925 reg = new_out_reg
926 = lra_create_new_reg_with_unique_value (outmode, out_rtx,
927 goal_class, "");
928 if (SCALAR_INT_MODE_P (outmode))
929 new_in_reg = gen_lowpart_SUBREG (inmode, reg);
930 else
931 new_in_reg = gen_rtx_SUBREG (inmode, reg, 0);
932 /* NEW_IN_REG is non-paradoxical subreg. We don't want
933 NEW_OUT_REG living above. We add clobber clause for
934 this. This is just a temporary clobber. We can remove
935 it at the end of LRA work. */
936 rtx_insn *clobber = emit_clobber (new_out_reg);
937 LRA_TEMP_CLOBBER_P (PATTERN (clobber)) = 1;
938 LRA_SUBREG_P (new_in_reg) = 1;
939 if (GET_CODE (in_rtx) == SUBREG)
941 rtx subreg_reg = SUBREG_REG (in_rtx);
943 /* If SUBREG_REG is dying here and sub-registers IN_RTX
944 and NEW_IN_REG are similar, we can use the same hard
945 register for REG and SUBREG_REG. */
946 if (REG_P (subreg_reg)
947 && (int) REGNO (subreg_reg) < lra_new_regno_start
948 && GET_MODE (subreg_reg) == outmode
949 && SUBREG_BYTE (in_rtx) == SUBREG_BYTE (new_in_reg)
950 && find_regno_note (curr_insn, REG_DEAD, REGNO (subreg_reg)))
951 lra_assign_reg_val (REGNO (subreg_reg), REGNO (reg));
955 else
957 /* Pseudos have values -- see comments for lra_reg_info.
958 Different pseudos with the same value do not conflict even if
959 they live in the same place. When we create a pseudo we
960 assign value of original pseudo (if any) from which we
961 created the new pseudo. If we create the pseudo from the
962 input pseudo, the new pseudo will have no conflict with the
963 input pseudo which is wrong when the input pseudo lives after
964 the insn and as the new pseudo value is changed by the insn
965 output. Therefore we create the new pseudo from the output
966 except the case when we have single matched dying input
967 pseudo.
969 We cannot reuse the current output register because we might
970 have a situation like "a <- a op b", where the constraints
971 force the second input operand ("b") to match the output
972 operand ("a"). "b" must then be copied into a new register
973 so that it doesn't clobber the current value of "a".
975 We can not use the same value if the output pseudo is
976 early clobbered or the input pseudo is mentioned in the
977 output, e.g. as an address part in memory, because
978 output reload will actually extend the pseudo liveness.
979 We don't care about eliminable hard regs here as we are
980 interesting only in pseudos. */
982 /* Matching input's register value is the same as one of the other
983 output operand. Output operands in a parallel insn must be in
984 different registers. */
985 out_conflict = false;
986 if (REG_P (in_rtx))
988 for (i = 0; outs[i] >= 0; i++)
990 rtx other_out_rtx = *curr_id->operand_loc[outs[i]];
991 if (REG_P (other_out_rtx)
992 && (regno_val_use_in (REGNO (in_rtx), other_out_rtx)
993 != NULL_RTX))
995 out_conflict = true;
996 break;
1001 new_in_reg = new_out_reg
1002 = (! early_clobber_p && ins[1] < 0 && REG_P (in_rtx)
1003 && (int) REGNO (in_rtx) < lra_new_regno_start
1004 && find_regno_note (curr_insn, REG_DEAD, REGNO (in_rtx))
1005 && (out < 0
1006 || regno_val_use_in (REGNO (in_rtx), out_rtx) == NULL_RTX)
1007 && !out_conflict
1008 ? lra_create_new_reg (inmode, in_rtx, goal_class, "")
1009 : lra_create_new_reg_with_unique_value (outmode, out_rtx,
1010 goal_class, ""));
1012 /* In operand can be got from transformations before processing insn
1013 constraints. One example of such transformations is subreg
1014 reloading (see function simplify_operand_subreg). The new
1015 pseudos created by the transformations might have inaccurate
1016 class (ALL_REGS) and we should make their classes more
1017 accurate. */
1018 narrow_reload_pseudo_class (in_rtx, goal_class);
1019 lra_emit_move (copy_rtx (new_in_reg), in_rtx);
1020 *before = get_insns ();
1021 end_sequence ();
1022 /* Add the new pseudo to consider values of subsequent input reload
1023 pseudos. */
1024 lra_assert (curr_insn_input_reloads_num < LRA_MAX_INSN_RELOADS);
1025 curr_insn_input_reloads[curr_insn_input_reloads_num].input = in_rtx;
1026 curr_insn_input_reloads[curr_insn_input_reloads_num].match_p = true;
1027 curr_insn_input_reloads[curr_insn_input_reloads_num++].reg = new_in_reg;
1028 for (i = 0; (in = ins[i]) >= 0; i++)
1030 lra_assert
1031 (GET_MODE (*curr_id->operand_loc[in]) == VOIDmode
1032 || GET_MODE (new_in_reg) == GET_MODE (*curr_id->operand_loc[in]));
1033 *curr_id->operand_loc[in] = new_in_reg;
1035 lra_update_dups (curr_id, ins);
1036 if (out < 0)
1037 return;
1038 /* See a comment for the input operand above. */
1039 narrow_reload_pseudo_class (out_rtx, goal_class);
1040 if (find_reg_note (curr_insn, REG_UNUSED, out_rtx) == NULL_RTX)
1042 start_sequence ();
1043 lra_emit_move (out_rtx, copy_rtx (new_out_reg));
1044 emit_insn (*after);
1045 *after = get_insns ();
1046 end_sequence ();
1048 *curr_id->operand_loc[out] = new_out_reg;
1049 lra_update_dup (curr_id, out);
1052 /* Return register class which is union of all reg classes in insn
1053 constraint alternative string starting with P. */
1054 static enum reg_class
1055 reg_class_from_constraints (const char *p)
1057 int c, len;
1058 enum reg_class op_class = NO_REGS;
1061 switch ((c = *p, len = CONSTRAINT_LEN (c, p)), c)
1063 case '#':
1064 case ',':
1065 return op_class;
1067 case 'g':
1068 op_class = reg_class_subunion[op_class][GENERAL_REGS];
1069 break;
1071 default:
1072 enum constraint_num cn = lookup_constraint (p);
1073 enum reg_class cl = reg_class_for_constraint (cn);
1074 if (cl == NO_REGS)
1076 if (insn_extra_address_constraint (cn))
1077 op_class
1078 = (reg_class_subunion
1079 [op_class][base_reg_class (VOIDmode, ADDR_SPACE_GENERIC,
1080 ADDRESS, SCRATCH)]);
1081 break;
1084 op_class = reg_class_subunion[op_class][cl];
1085 break;
1087 while ((p += len), c);
1088 return op_class;
1091 /* If OP is a register, return the class of the register as per
1092 get_reg_class, otherwise return NO_REGS. */
1093 static inline enum reg_class
1094 get_op_class (rtx op)
1096 return REG_P (op) ? get_reg_class (REGNO (op)) : NO_REGS;
1099 /* Return generated insn mem_pseudo:=val if TO_P or val:=mem_pseudo
1100 otherwise. If modes of MEM_PSEUDO and VAL are different, use
1101 SUBREG for VAL to make them equal. */
1102 static rtx_insn *
1103 emit_spill_move (bool to_p, rtx mem_pseudo, rtx val)
1105 if (GET_MODE (mem_pseudo) != GET_MODE (val))
1107 /* Usually size of mem_pseudo is greater than val size but in
1108 rare cases it can be less as it can be defined by target
1109 dependent macro HARD_REGNO_CALLER_SAVE_MODE. */
1110 if (! MEM_P (val))
1112 val = gen_lowpart_SUBREG (GET_MODE (mem_pseudo),
1113 GET_CODE (val) == SUBREG
1114 ? SUBREG_REG (val) : val);
1115 LRA_SUBREG_P (val) = 1;
1117 else
1119 mem_pseudo = gen_lowpart_SUBREG (GET_MODE (val), mem_pseudo);
1120 LRA_SUBREG_P (mem_pseudo) = 1;
1123 return to_p ? gen_move_insn (mem_pseudo, val)
1124 : gen_move_insn (val, mem_pseudo);
1127 /* Process a special case insn (register move), return true if we
1128 don't need to process it anymore. INSN should be a single set
1129 insn. Set up that RTL was changed through CHANGE_P and macro
1130 SECONDARY_MEMORY_NEEDED says to use secondary memory through
1131 SEC_MEM_P. */
1132 static bool
1133 check_and_process_move (bool *change_p, bool *sec_mem_p ATTRIBUTE_UNUSED)
1135 int sregno, dregno;
1136 rtx dest, src, dreg, sreg, new_reg, scratch_reg;
1137 rtx_insn *before;
1138 enum reg_class dclass, sclass, secondary_class;
1139 secondary_reload_info sri;
1141 lra_assert (curr_insn_set != NULL_RTX);
1142 dreg = dest = SET_DEST (curr_insn_set);
1143 sreg = src = SET_SRC (curr_insn_set);
1144 if (GET_CODE (dest) == SUBREG)
1145 dreg = SUBREG_REG (dest);
1146 if (GET_CODE (src) == SUBREG)
1147 sreg = SUBREG_REG (src);
1148 if (! (REG_P (dreg) || MEM_P (dreg)) || ! (REG_P (sreg) || MEM_P (sreg)))
1149 return false;
1150 sclass = dclass = NO_REGS;
1151 if (REG_P (dreg))
1152 dclass = get_reg_class (REGNO (dreg));
1153 gcc_assert (dclass < LIM_REG_CLASSES);
1154 if (dclass == ALL_REGS)
1155 /* ALL_REGS is used for new pseudos created by transformations
1156 like reload of SUBREG_REG (see function
1157 simplify_operand_subreg). We don't know their class yet. We
1158 should figure out the class from processing the insn
1159 constraints not in this fast path function. Even if ALL_REGS
1160 were a right class for the pseudo, secondary_... hooks usually
1161 are not define for ALL_REGS. */
1162 return false;
1163 if (REG_P (sreg))
1164 sclass = get_reg_class (REGNO (sreg));
1165 gcc_assert (sclass < LIM_REG_CLASSES);
1166 if (sclass == ALL_REGS)
1167 /* See comments above. */
1168 return false;
1169 if (sclass == NO_REGS && dclass == NO_REGS)
1170 return false;
1171 #ifdef SECONDARY_MEMORY_NEEDED
1172 if (SECONDARY_MEMORY_NEEDED (sclass, dclass, GET_MODE (src))
1173 #ifdef SECONDARY_MEMORY_NEEDED_MODE
1174 && ((sclass != NO_REGS && dclass != NO_REGS)
1175 || GET_MODE (src) != SECONDARY_MEMORY_NEEDED_MODE (GET_MODE (src)))
1176 #endif
1179 *sec_mem_p = true;
1180 return false;
1182 #endif
1183 if (! REG_P (dreg) || ! REG_P (sreg))
1184 return false;
1185 sri.prev_sri = NULL;
1186 sri.icode = CODE_FOR_nothing;
1187 sri.extra_cost = 0;
1188 secondary_class = NO_REGS;
1189 /* Set up hard register for a reload pseudo for hook
1190 secondary_reload because some targets just ignore unassigned
1191 pseudos in the hook. */
1192 if (dclass != NO_REGS && lra_get_regno_hard_regno (REGNO (dreg)) < 0)
1194 dregno = REGNO (dreg);
1195 reg_renumber[dregno] = ira_class_hard_regs[dclass][0];
1197 else
1198 dregno = -1;
1199 if (sclass != NO_REGS && lra_get_regno_hard_regno (REGNO (sreg)) < 0)
1201 sregno = REGNO (sreg);
1202 reg_renumber[sregno] = ira_class_hard_regs[sclass][0];
1204 else
1205 sregno = -1;
1206 if (sclass != NO_REGS)
1207 secondary_class
1208 = (enum reg_class) targetm.secondary_reload (false, dest,
1209 (reg_class_t) sclass,
1210 GET_MODE (src), &sri);
1211 if (sclass == NO_REGS
1212 || ((secondary_class != NO_REGS || sri.icode != CODE_FOR_nothing)
1213 && dclass != NO_REGS))
1215 enum reg_class old_sclass = secondary_class;
1216 secondary_reload_info old_sri = sri;
1218 sri.prev_sri = NULL;
1219 sri.icode = CODE_FOR_nothing;
1220 sri.extra_cost = 0;
1221 secondary_class
1222 = (enum reg_class) targetm.secondary_reload (true, src,
1223 (reg_class_t) dclass,
1224 GET_MODE (src), &sri);
1225 /* Check the target hook consistency. */
1226 lra_assert
1227 ((secondary_class == NO_REGS && sri.icode == CODE_FOR_nothing)
1228 || (old_sclass == NO_REGS && old_sri.icode == CODE_FOR_nothing)
1229 || (secondary_class == old_sclass && sri.icode == old_sri.icode));
1231 if (sregno >= 0)
1232 reg_renumber [sregno] = -1;
1233 if (dregno >= 0)
1234 reg_renumber [dregno] = -1;
1235 if (secondary_class == NO_REGS && sri.icode == CODE_FOR_nothing)
1236 return false;
1237 *change_p = true;
1238 new_reg = NULL_RTX;
1239 if (secondary_class != NO_REGS)
1240 new_reg = lra_create_new_reg_with_unique_value (GET_MODE (src), NULL_RTX,
1241 secondary_class,
1242 "secondary");
1243 start_sequence ();
1244 if (sri.icode == CODE_FOR_nothing)
1245 lra_emit_move (new_reg, src);
1246 else
1248 enum reg_class scratch_class;
1250 scratch_class = (reg_class_from_constraints
1251 (insn_data[sri.icode].operand[2].constraint));
1252 scratch_reg = (lra_create_new_reg_with_unique_value
1253 (insn_data[sri.icode].operand[2].mode, NULL_RTX,
1254 scratch_class, "scratch"));
1255 emit_insn (GEN_FCN (sri.icode) (new_reg != NULL_RTX ? new_reg : dest,
1256 src, scratch_reg));
1258 before = get_insns ();
1259 end_sequence ();
1260 lra_process_new_insns (curr_insn, before, NULL, "Inserting the move");
1261 if (new_reg != NULL_RTX)
1262 SET_SRC (curr_insn_set) = new_reg;
1263 else
1265 if (lra_dump_file != NULL)
1267 fprintf (lra_dump_file, "Deleting move %u\n", INSN_UID (curr_insn));
1268 dump_insn_slim (lra_dump_file, curr_insn);
1270 lra_set_insn_deleted (curr_insn);
1271 return true;
1273 return false;
1276 /* The following data describe the result of process_alt_operands.
1277 The data are used in curr_insn_transform to generate reloads. */
1279 /* The chosen reg classes which should be used for the corresponding
1280 operands. */
1281 static enum reg_class goal_alt[MAX_RECOG_OPERANDS];
1282 /* True if the operand should be the same as another operand and that
1283 other operand does not need a reload. */
1284 static bool goal_alt_match_win[MAX_RECOG_OPERANDS];
1285 /* True if the operand does not need a reload. */
1286 static bool goal_alt_win[MAX_RECOG_OPERANDS];
1287 /* True if the operand can be offsetable memory. */
1288 static bool goal_alt_offmemok[MAX_RECOG_OPERANDS];
1289 /* The number of an operand to which given operand can be matched to. */
1290 static int goal_alt_matches[MAX_RECOG_OPERANDS];
1291 /* The number of elements in the following array. */
1292 static int goal_alt_dont_inherit_ops_num;
1293 /* Numbers of operands whose reload pseudos should not be inherited. */
1294 static int goal_alt_dont_inherit_ops[MAX_RECOG_OPERANDS];
1295 /* True if the insn commutative operands should be swapped. */
1296 static bool goal_alt_swapped;
1297 /* The chosen insn alternative. */
1298 static int goal_alt_number;
1300 /* True if the corresponding operand is the result of an equivalence
1301 substitution. */
1302 static bool equiv_substition_p[MAX_RECOG_OPERANDS];
1304 /* The following five variables are used to choose the best insn
1305 alternative. They reflect final characteristics of the best
1306 alternative. */
1308 /* Number of necessary reloads and overall cost reflecting the
1309 previous value and other unpleasantness of the best alternative. */
1310 static int best_losers, best_overall;
1311 /* Overall number hard registers used for reloads. For example, on
1312 some targets we need 2 general registers to reload DFmode and only
1313 one floating point register. */
1314 static int best_reload_nregs;
1315 /* Overall number reflecting distances of previous reloading the same
1316 value. The distances are counted from the current BB start. It is
1317 used to improve inheritance chances. */
1318 static int best_reload_sum;
1320 /* True if the current insn should have no correspondingly input or
1321 output reloads. */
1322 static bool no_input_reloads_p, no_output_reloads_p;
1324 /* True if we swapped the commutative operands in the current
1325 insn. */
1326 static int curr_swapped;
1328 /* if CHECK_ONLY_P is false, arrange for address element *LOC to be a
1329 register of class CL. Add any input reloads to list BEFORE. AFTER
1330 is nonnull if *LOC is an automodified value; handle that case by
1331 adding the required output reloads to list AFTER. Return true if
1332 the RTL was changed.
1334 if CHECK_ONLY_P is true, check that the *LOC is a correct address
1335 register. Return false if the address register is correct. */
1336 static bool
1337 process_addr_reg (rtx *loc, bool check_only_p, rtx_insn **before, rtx_insn **after,
1338 enum reg_class cl)
1340 int regno;
1341 enum reg_class rclass, new_class;
1342 rtx reg;
1343 rtx new_reg;
1344 machine_mode mode;
1345 bool subreg_p, before_p = false;
1347 subreg_p = GET_CODE (*loc) == SUBREG;
1348 if (subreg_p)
1350 reg = SUBREG_REG (*loc);
1351 mode = GET_MODE (reg);
1353 /* For mode with size bigger than ptr_mode, there unlikely to be "mov"
1354 between two registers with different classes, but there normally will
1355 be "mov" which transfers element of vector register into the general
1356 register, and this normally will be a subreg which should be reloaded
1357 as a whole. This is particularly likely to be triggered when
1358 -fno-split-wide-types specified. */
1359 if (!REG_P (reg)
1360 || in_class_p (reg, cl, &new_class)
1361 || GET_MODE_SIZE (mode) <= GET_MODE_SIZE (ptr_mode))
1362 loc = &SUBREG_REG (*loc);
1365 reg = *loc;
1366 mode = GET_MODE (reg);
1367 if (! REG_P (reg))
1369 if (check_only_p)
1370 return true;
1371 /* Always reload memory in an address even if the target supports
1372 such addresses. */
1373 new_reg = lra_create_new_reg_with_unique_value (mode, reg, cl, "address");
1374 before_p = true;
1376 else
1378 regno = REGNO (reg);
1379 rclass = get_reg_class (regno);
1380 if (! check_only_p
1381 && (*loc = get_equiv_with_elimination (reg, curr_insn)) != reg)
1383 if (lra_dump_file != NULL)
1385 fprintf (lra_dump_file,
1386 "Changing pseudo %d in address of insn %u on equiv ",
1387 REGNO (reg), INSN_UID (curr_insn));
1388 dump_value_slim (lra_dump_file, *loc, 1);
1389 fprintf (lra_dump_file, "\n");
1391 *loc = copy_rtx (*loc);
1393 if (*loc != reg || ! in_class_p (reg, cl, &new_class))
1395 if (check_only_p)
1396 return true;
1397 reg = *loc;
1398 if (get_reload_reg (after == NULL ? OP_IN : OP_INOUT,
1399 mode, reg, cl, subreg_p, "address", &new_reg))
1400 before_p = true;
1402 else if (new_class != NO_REGS && rclass != new_class)
1404 if (check_only_p)
1405 return true;
1406 lra_change_class (regno, new_class, " Change to", true);
1407 return false;
1409 else
1410 return false;
1412 if (before_p)
1414 push_to_sequence (*before);
1415 lra_emit_move (new_reg, reg);
1416 *before = get_insns ();
1417 end_sequence ();
1419 *loc = new_reg;
1420 if (after != NULL)
1422 start_sequence ();
1423 lra_emit_move (before_p ? copy_rtx (reg) : reg, new_reg);
1424 emit_insn (*after);
1425 *after = get_insns ();
1426 end_sequence ();
1428 return true;
1431 /* Insert move insn in simplify_operand_subreg. BEFORE returns
1432 the insn to be inserted before curr insn. AFTER returns the
1433 the insn to be inserted after curr insn. ORIGREG and NEWREG
1434 are the original reg and new reg for reload. */
1435 static void
1436 insert_move_for_subreg (rtx_insn **before, rtx_insn **after, rtx origreg,
1437 rtx newreg)
1439 if (before)
1441 push_to_sequence (*before);
1442 lra_emit_move (newreg, origreg);
1443 *before = get_insns ();
1444 end_sequence ();
1446 if (after)
1448 start_sequence ();
1449 lra_emit_move (origreg, newreg);
1450 emit_insn (*after);
1451 *after = get_insns ();
1452 end_sequence ();
1456 static int valid_address_p (machine_mode mode, rtx addr, addr_space_t as);
1457 static bool process_address (int, bool, rtx_insn **, rtx_insn **);
1459 /* Make reloads for subreg in operand NOP with internal subreg mode
1460 REG_MODE, add new reloads for further processing. Return true if
1461 any change was done. */
1462 static bool
1463 simplify_operand_subreg (int nop, machine_mode reg_mode)
1465 int hard_regno;
1466 rtx_insn *before, *after;
1467 machine_mode mode, innermode;
1468 rtx reg, new_reg;
1469 rtx operand = *curr_id->operand_loc[nop];
1470 enum reg_class regclass;
1471 enum op_type type;
1473 before = after = NULL;
1475 if (GET_CODE (operand) != SUBREG)
1476 return false;
1478 mode = GET_MODE (operand);
1479 reg = SUBREG_REG (operand);
1480 innermode = GET_MODE (reg);
1481 type = curr_static_id->operand[nop].type;
1482 if (MEM_P (reg))
1484 const bool addr_was_valid
1485 = valid_address_p (innermode, XEXP (reg, 0), MEM_ADDR_SPACE (reg));
1486 alter_subreg (curr_id->operand_loc[nop], false);
1487 rtx subst = *curr_id->operand_loc[nop];
1488 lra_assert (MEM_P (subst));
1490 if (!addr_was_valid
1491 || valid_address_p (GET_MODE (subst), XEXP (subst, 0),
1492 MEM_ADDR_SPACE (subst))
1493 || ((get_constraint_type (lookup_constraint
1494 (curr_static_id->operand[nop].constraint))
1495 != CT_SPECIAL_MEMORY)
1496 /* We still can reload address and if the address is
1497 valid, we can remove subreg without reloading its
1498 inner memory. */
1499 && valid_address_p (GET_MODE (subst),
1500 regno_reg_rtx
1501 [ira_class_hard_regs
1502 [base_reg_class (GET_MODE (subst),
1503 MEM_ADDR_SPACE (subst),
1504 ADDRESS, SCRATCH)][0]],
1505 MEM_ADDR_SPACE (subst))))
1507 /* If we change the address for a paradoxical subreg of memory, the
1508 new address might violate the necessary alignment or the access
1509 might be slow; take this into consideration. We need not worry
1510 about accesses beyond allocated memory for paradoxical memory
1511 subregs as we don't substitute such equiv memory (see processing
1512 equivalences in function lra_constraints) and because for spilled
1513 pseudos we allocate stack memory enough for the biggest
1514 corresponding paradoxical subreg. */
1515 if (!(MEM_ALIGN (subst) < GET_MODE_ALIGNMENT (mode)
1516 && SLOW_UNALIGNED_ACCESS (mode, MEM_ALIGN (subst)))
1517 || (MEM_ALIGN (reg) < GET_MODE_ALIGNMENT (innermode)
1518 && SLOW_UNALIGNED_ACCESS (innermode, MEM_ALIGN (reg))))
1519 return true;
1521 *curr_id->operand_loc[nop] = operand;
1523 /* But if the address was not valid, we cannot reload the MEM without
1524 reloading the address first. */
1525 if (!addr_was_valid)
1526 process_address (nop, false, &before, &after);
1528 /* INNERMODE is fast, MODE slow. Reload the mem in INNERMODE. */
1529 enum reg_class rclass
1530 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1531 if (get_reload_reg (curr_static_id->operand[nop].type, innermode,
1532 reg, rclass, TRUE, "slow mem", &new_reg))
1534 bool insert_before, insert_after;
1535 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1537 insert_before = (type != OP_OUT
1538 || GET_MODE_SIZE (innermode)
1539 > GET_MODE_SIZE (mode));
1540 insert_after = type != OP_IN;
1541 insert_move_for_subreg (insert_before ? &before : NULL,
1542 insert_after ? &after : NULL,
1543 reg, new_reg);
1545 SUBREG_REG (operand) = new_reg;
1547 /* Convert to MODE. */
1548 reg = operand;
1549 rclass
1550 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1551 if (get_reload_reg (curr_static_id->operand[nop].type, mode, reg,
1552 rclass, TRUE, "slow mem", &new_reg))
1554 bool insert_before, insert_after;
1555 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1557 insert_before = type != OP_OUT;
1558 insert_after = type != OP_IN;
1559 insert_move_for_subreg (insert_before ? &before : NULL,
1560 insert_after ? &after : NULL,
1561 reg, new_reg);
1563 *curr_id->operand_loc[nop] = new_reg;
1564 lra_process_new_insns (curr_insn, before, after,
1565 "Inserting slow mem reload");
1566 return true;
1569 /* If the address was valid and became invalid, prefer to reload
1570 the memory. Typical case is when the index scale should
1571 correspond the memory. */
1572 *curr_id->operand_loc[nop] = operand;
1573 return false;
1575 else if (REG_P (reg) && REGNO (reg) < FIRST_PSEUDO_REGISTER)
1577 alter_subreg (curr_id->operand_loc[nop], false);
1578 return true;
1580 else if (CONSTANT_P (reg))
1582 /* Try to simplify subreg of constant. It is usually result of
1583 equivalence substitution. */
1584 if (innermode == VOIDmode
1585 && (innermode = original_subreg_reg_mode[nop]) == VOIDmode)
1586 innermode = curr_static_id->operand[nop].mode;
1587 if ((new_reg = simplify_subreg (mode, reg, innermode,
1588 SUBREG_BYTE (operand))) != NULL_RTX)
1590 *curr_id->operand_loc[nop] = new_reg;
1591 return true;
1594 /* Put constant into memory when we have mixed modes. It generates
1595 a better code in most cases as it does not need a secondary
1596 reload memory. It also prevents LRA looping when LRA is using
1597 secondary reload memory again and again. */
1598 if (CONSTANT_P (reg) && CONST_POOL_OK_P (reg_mode, reg)
1599 && SCALAR_INT_MODE_P (reg_mode) != SCALAR_INT_MODE_P (mode))
1601 SUBREG_REG (operand) = force_const_mem (reg_mode, reg);
1602 alter_subreg (curr_id->operand_loc[nop], false);
1603 return true;
1605 /* Force a reload of the SUBREG_REG if this is a constant or PLUS or
1606 if there may be a problem accessing OPERAND in the outer
1607 mode. */
1608 if ((REG_P (reg)
1609 && REGNO (reg) >= FIRST_PSEUDO_REGISTER
1610 && (hard_regno = lra_get_regno_hard_regno (REGNO (reg))) >= 0
1611 /* Don't reload paradoxical subregs because we could be looping
1612 having repeatedly final regno out of hard regs range. */
1613 && (hard_regno_nregs[hard_regno][innermode]
1614 >= hard_regno_nregs[hard_regno][mode])
1615 && simplify_subreg_regno (hard_regno, innermode,
1616 SUBREG_BYTE (operand), mode) < 0
1617 /* Don't reload subreg for matching reload. It is actually
1618 valid subreg in LRA. */
1619 && ! LRA_SUBREG_P (operand))
1620 || CONSTANT_P (reg) || GET_CODE (reg) == PLUS || MEM_P (reg))
1622 enum reg_class rclass;
1624 if (REG_P (reg))
1625 /* There is a big probability that we will get the same class
1626 for the new pseudo and we will get the same insn which
1627 means infinite looping. So spill the new pseudo. */
1628 rclass = NO_REGS;
1629 else
1630 /* The class will be defined later in curr_insn_transform. */
1631 rclass
1632 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1634 if (get_reload_reg (curr_static_id->operand[nop].type, reg_mode, reg,
1635 rclass, TRUE, "subreg reg", &new_reg))
1637 bool insert_before, insert_after;
1638 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1640 insert_before = (type != OP_OUT
1641 || GET_MODE_SIZE (innermode) > GET_MODE_SIZE (mode));
1642 insert_after = (type != OP_IN);
1643 insert_move_for_subreg (insert_before ? &before : NULL,
1644 insert_after ? &after : NULL,
1645 reg, new_reg);
1647 SUBREG_REG (operand) = new_reg;
1648 lra_process_new_insns (curr_insn, before, after,
1649 "Inserting subreg reload");
1650 return true;
1652 /* Force a reload for a paradoxical subreg. For paradoxical subreg,
1653 IRA allocates hardreg to the inner pseudo reg according to its mode
1654 instead of the outermode, so the size of the hardreg may not be enough
1655 to contain the outermode operand, in that case we may need to insert
1656 reload for the reg. For the following two types of paradoxical subreg,
1657 we need to insert reload:
1658 1. If the op_type is OP_IN, and the hardreg could not be paired with
1659 other hardreg to contain the outermode operand
1660 (checked by in_hard_reg_set_p), we need to insert the reload.
1661 2. If the op_type is OP_OUT or OP_INOUT.
1663 Here is a paradoxical subreg example showing how the reload is generated:
1665 (insn 5 4 7 2 (set (reg:TI 106 [ __comp ])
1666 (subreg:TI (reg:DI 107 [ __comp ]) 0)) {*movti_internal_rex64}
1668 In IRA, reg107 is allocated to a DImode hardreg. We use x86-64 as example
1669 here, if reg107 is assigned to hardreg R15, because R15 is the last
1670 hardreg, compiler cannot find another hardreg to pair with R15 to
1671 contain TImode data. So we insert a TImode reload reg180 for it.
1672 After reload is inserted:
1674 (insn 283 0 0 (set (subreg:DI (reg:TI 180 [orig:107 __comp ] [107]) 0)
1675 (reg:DI 107 [ __comp ])) -1
1676 (insn 5 4 7 2 (set (reg:TI 106 [ __comp ])
1677 (subreg:TI (reg:TI 180 [orig:107 __comp ] [107]) 0)) {*movti_internal_rex64}
1679 Two reload hard registers will be allocated to reg180 to save TImode data
1680 in LRA_assign. */
1681 else if (REG_P (reg)
1682 && REGNO (reg) >= FIRST_PSEUDO_REGISTER
1683 && (hard_regno = lra_get_regno_hard_regno (REGNO (reg))) >= 0
1684 && (hard_regno_nregs[hard_regno][innermode]
1685 < hard_regno_nregs[hard_regno][mode])
1686 && (regclass = lra_get_allocno_class (REGNO (reg)))
1687 && (type != OP_IN
1688 || !in_hard_reg_set_p (reg_class_contents[regclass],
1689 mode, hard_regno)))
1691 /* The class will be defined later in curr_insn_transform. */
1692 enum reg_class rclass
1693 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1695 if (get_reload_reg (curr_static_id->operand[nop].type, mode, reg,
1696 rclass, TRUE, "paradoxical subreg", &new_reg))
1698 rtx subreg;
1699 bool insert_before, insert_after;
1701 PUT_MODE (new_reg, mode);
1702 subreg = gen_lowpart_SUBREG (innermode, new_reg);
1703 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1705 insert_before = (type != OP_OUT);
1706 insert_after = (type != OP_IN);
1707 insert_move_for_subreg (insert_before ? &before : NULL,
1708 insert_after ? &after : NULL,
1709 reg, subreg);
1711 SUBREG_REG (operand) = new_reg;
1712 lra_process_new_insns (curr_insn, before, after,
1713 "Inserting paradoxical subreg reload");
1714 return true;
1716 return false;
1719 /* Return TRUE if X refers for a hard register from SET. */
1720 static bool
1721 uses_hard_regs_p (rtx x, HARD_REG_SET set)
1723 int i, j, x_hard_regno;
1724 machine_mode mode;
1725 const char *fmt;
1726 enum rtx_code code;
1728 if (x == NULL_RTX)
1729 return false;
1730 code = GET_CODE (x);
1731 mode = GET_MODE (x);
1732 if (code == SUBREG)
1734 x = SUBREG_REG (x);
1735 code = GET_CODE (x);
1736 if (GET_MODE_SIZE (GET_MODE (x)) > GET_MODE_SIZE (mode))
1737 mode = GET_MODE (x);
1740 if (REG_P (x))
1742 x_hard_regno = get_hard_regno (x, true);
1743 return (x_hard_regno >= 0
1744 && overlaps_hard_reg_set_p (set, mode, x_hard_regno));
1746 if (MEM_P (x))
1748 struct address_info ad;
1750 decompose_mem_address (&ad, x);
1751 if (ad.base_term != NULL && uses_hard_regs_p (*ad.base_term, set))
1752 return true;
1753 if (ad.index_term != NULL && uses_hard_regs_p (*ad.index_term, set))
1754 return true;
1756 fmt = GET_RTX_FORMAT (code);
1757 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1759 if (fmt[i] == 'e')
1761 if (uses_hard_regs_p (XEXP (x, i), set))
1762 return true;
1764 else if (fmt[i] == 'E')
1766 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
1767 if (uses_hard_regs_p (XVECEXP (x, i, j), set))
1768 return true;
1771 return false;
1774 /* Return true if OP is a spilled pseudo. */
1775 static inline bool
1776 spilled_pseudo_p (rtx op)
1778 return (REG_P (op)
1779 && REGNO (op) >= FIRST_PSEUDO_REGISTER && in_mem_p (REGNO (op)));
1782 /* Return true if X is a general constant. */
1783 static inline bool
1784 general_constant_p (rtx x)
1786 return CONSTANT_P (x) && (! flag_pic || LEGITIMATE_PIC_OPERAND_P (x));
1789 static bool
1790 reg_in_class_p (rtx reg, enum reg_class cl)
1792 if (cl == NO_REGS)
1793 return get_reg_class (REGNO (reg)) == NO_REGS;
1794 return in_class_p (reg, cl, NULL);
1797 /* Return true if SET of RCLASS contains no hard regs which can be
1798 used in MODE. */
1799 static bool
1800 prohibited_class_reg_set_mode_p (enum reg_class rclass,
1801 HARD_REG_SET &set,
1802 enum machine_mode mode)
1804 HARD_REG_SET temp;
1806 lra_assert (hard_reg_set_subset_p (reg_class_contents[rclass], set));
1807 COPY_HARD_REG_SET (temp, set);
1808 AND_COMPL_HARD_REG_SET (temp, lra_no_alloc_regs);
1809 return (hard_reg_set_subset_p
1810 (temp, ira_prohibited_class_mode_regs[rclass][mode]));
1813 /* Major function to choose the current insn alternative and what
1814 operands should be reloaded and how. If ONLY_ALTERNATIVE is not
1815 negative we should consider only this alternative. Return false if
1816 we can not choose the alternative or find how to reload the
1817 operands. */
1818 static bool
1819 process_alt_operands (int only_alternative)
1821 bool ok_p = false;
1822 int nop, overall, nalt;
1823 int n_alternatives = curr_static_id->n_alternatives;
1824 int n_operands = curr_static_id->n_operands;
1825 /* LOSERS counts the operands that don't fit this alternative and
1826 would require loading. */
1827 int losers;
1828 /* REJECT is a count of how undesirable this alternative says it is
1829 if any reloading is required. If the alternative matches exactly
1830 then REJECT is ignored, but otherwise it gets this much counted
1831 against it in addition to the reloading needed. */
1832 int reject;
1833 int op_reject;
1834 /* The number of elements in the following array. */
1835 int early_clobbered_regs_num;
1836 /* Numbers of operands which are early clobber registers. */
1837 int early_clobbered_nops[MAX_RECOG_OPERANDS];
1838 enum reg_class curr_alt[MAX_RECOG_OPERANDS];
1839 HARD_REG_SET curr_alt_set[MAX_RECOG_OPERANDS];
1840 bool curr_alt_match_win[MAX_RECOG_OPERANDS];
1841 bool curr_alt_win[MAX_RECOG_OPERANDS];
1842 bool curr_alt_offmemok[MAX_RECOG_OPERANDS];
1843 int curr_alt_matches[MAX_RECOG_OPERANDS];
1844 /* The number of elements in the following array. */
1845 int curr_alt_dont_inherit_ops_num;
1846 /* Numbers of operands whose reload pseudos should not be inherited. */
1847 int curr_alt_dont_inherit_ops[MAX_RECOG_OPERANDS];
1848 rtx op;
1849 /* The register when the operand is a subreg of register, otherwise the
1850 operand itself. */
1851 rtx no_subreg_reg_operand[MAX_RECOG_OPERANDS];
1852 /* The register if the operand is a register or subreg of register,
1853 otherwise NULL. */
1854 rtx operand_reg[MAX_RECOG_OPERANDS];
1855 int hard_regno[MAX_RECOG_OPERANDS];
1856 machine_mode biggest_mode[MAX_RECOG_OPERANDS];
1857 int reload_nregs, reload_sum;
1858 bool costly_p;
1859 enum reg_class cl;
1861 /* Calculate some data common for all alternatives to speed up the
1862 function. */
1863 for (nop = 0; nop < n_operands; nop++)
1865 rtx reg;
1867 op = no_subreg_reg_operand[nop] = *curr_id->operand_loc[nop];
1868 /* The real hard regno of the operand after the allocation. */
1869 hard_regno[nop] = get_hard_regno (op, true);
1871 operand_reg[nop] = reg = op;
1872 biggest_mode[nop] = GET_MODE (op);
1873 if (GET_CODE (op) == SUBREG)
1875 operand_reg[nop] = reg = SUBREG_REG (op);
1876 if (GET_MODE_SIZE (biggest_mode[nop])
1877 < GET_MODE_SIZE (GET_MODE (reg)))
1878 biggest_mode[nop] = GET_MODE (reg);
1880 if (! REG_P (reg))
1881 operand_reg[nop] = NULL_RTX;
1882 else if (REGNO (reg) >= FIRST_PSEUDO_REGISTER
1883 || ((int) REGNO (reg)
1884 == lra_get_elimination_hard_regno (REGNO (reg))))
1885 no_subreg_reg_operand[nop] = reg;
1886 else
1887 operand_reg[nop] = no_subreg_reg_operand[nop]
1888 /* Just use natural mode for elimination result. It should
1889 be enough for extra constraints hooks. */
1890 = regno_reg_rtx[hard_regno[nop]];
1893 /* The constraints are made of several alternatives. Each operand's
1894 constraint looks like foo,bar,... with commas separating the
1895 alternatives. The first alternatives for all operands go
1896 together, the second alternatives go together, etc.
1898 First loop over alternatives. */
1899 alternative_mask preferred = curr_id->preferred_alternatives;
1900 if (only_alternative >= 0)
1901 preferred &= ALTERNATIVE_BIT (only_alternative);
1903 for (nalt = 0; nalt < n_alternatives; nalt++)
1905 /* Loop over operands for one constraint alternative. */
1906 if (!TEST_BIT (preferred, nalt))
1907 continue;
1909 overall = losers = reject = reload_nregs = reload_sum = 0;
1910 for (nop = 0; nop < n_operands; nop++)
1912 int inc = (curr_static_id
1913 ->operand_alternative[nalt * n_operands + nop].reject);
1914 if (lra_dump_file != NULL && inc != 0)
1915 fprintf (lra_dump_file,
1916 " Staticly defined alt reject+=%d\n", inc);
1917 reject += inc;
1919 early_clobbered_regs_num = 0;
1921 for (nop = 0; nop < n_operands; nop++)
1923 const char *p;
1924 char *end;
1925 int len, c, m, i, opalt_num, this_alternative_matches;
1926 bool win, did_match, offmemok, early_clobber_p;
1927 /* false => this operand can be reloaded somehow for this
1928 alternative. */
1929 bool badop;
1930 /* true => this operand can be reloaded if the alternative
1931 allows regs. */
1932 bool winreg;
1933 /* True if a constant forced into memory would be OK for
1934 this operand. */
1935 bool constmemok;
1936 enum reg_class this_alternative, this_costly_alternative;
1937 HARD_REG_SET this_alternative_set, this_costly_alternative_set;
1938 bool this_alternative_match_win, this_alternative_win;
1939 bool this_alternative_offmemok;
1940 bool scratch_p;
1941 machine_mode mode;
1942 enum constraint_num cn;
1944 opalt_num = nalt * n_operands + nop;
1945 if (curr_static_id->operand_alternative[opalt_num].anything_ok)
1947 /* Fast track for no constraints at all. */
1948 curr_alt[nop] = NO_REGS;
1949 CLEAR_HARD_REG_SET (curr_alt_set[nop]);
1950 curr_alt_win[nop] = true;
1951 curr_alt_match_win[nop] = false;
1952 curr_alt_offmemok[nop] = false;
1953 curr_alt_matches[nop] = -1;
1954 continue;
1957 op = no_subreg_reg_operand[nop];
1958 mode = curr_operand_mode[nop];
1960 win = did_match = winreg = offmemok = constmemok = false;
1961 badop = true;
1963 early_clobber_p = false;
1964 p = curr_static_id->operand_alternative[opalt_num].constraint;
1966 this_costly_alternative = this_alternative = NO_REGS;
1967 /* We update set of possible hard regs besides its class
1968 because reg class might be inaccurate. For example,
1969 union of LO_REGS (l), HI_REGS(h), and STACK_REG(k) in ARM
1970 is translated in HI_REGS because classes are merged by
1971 pairs and there is no accurate intermediate class. */
1972 CLEAR_HARD_REG_SET (this_alternative_set);
1973 CLEAR_HARD_REG_SET (this_costly_alternative_set);
1974 this_alternative_win = false;
1975 this_alternative_match_win = false;
1976 this_alternative_offmemok = false;
1977 this_alternative_matches = -1;
1979 /* An empty constraint should be excluded by the fast
1980 track. */
1981 lra_assert (*p != 0 && *p != ',');
1983 op_reject = 0;
1984 /* Scan this alternative's specs for this operand; set WIN
1985 if the operand fits any letter in this alternative.
1986 Otherwise, clear BADOP if this operand could fit some
1987 letter after reloads, or set WINREG if this operand could
1988 fit after reloads provided the constraint allows some
1989 registers. */
1990 costly_p = false;
1993 switch ((c = *p, len = CONSTRAINT_LEN (c, p)), c)
1995 case '\0':
1996 len = 0;
1997 break;
1998 case ',':
1999 c = '\0';
2000 break;
2002 case '&':
2003 early_clobber_p = true;
2004 break;
2006 case '$':
2007 op_reject += LRA_MAX_REJECT;
2008 break;
2009 case '^':
2010 op_reject += LRA_LOSER_COST_FACTOR;
2011 break;
2013 case '#':
2014 /* Ignore rest of this alternative. */
2015 c = '\0';
2016 break;
2018 case '0': case '1': case '2': case '3': case '4':
2019 case '5': case '6': case '7': case '8': case '9':
2021 int m_hregno;
2022 bool match_p;
2024 m = strtoul (p, &end, 10);
2025 p = end;
2026 len = 0;
2027 lra_assert (nop > m);
2029 this_alternative_matches = m;
2030 m_hregno = get_hard_regno (*curr_id->operand_loc[m], false);
2031 /* We are supposed to match a previous operand.
2032 If we do, we win if that one did. If we do
2033 not, count both of the operands as losers.
2034 (This is too conservative, since most of the
2035 time only a single reload insn will be needed
2036 to make the two operands win. As a result,
2037 this alternative may be rejected when it is
2038 actually desirable.) */
2039 match_p = false;
2040 if (operands_match_p (*curr_id->operand_loc[nop],
2041 *curr_id->operand_loc[m], m_hregno))
2043 /* We should reject matching of an early
2044 clobber operand if the matching operand is
2045 not dying in the insn. */
2046 if (! curr_static_id->operand[m].early_clobber
2047 || operand_reg[nop] == NULL_RTX
2048 || (find_regno_note (curr_insn, REG_DEAD,
2049 REGNO (op))
2050 || REGNO (op) == REGNO (operand_reg[m])))
2051 match_p = true;
2053 if (match_p)
2055 /* If we are matching a non-offsettable
2056 address where an offsettable address was
2057 expected, then we must reject this
2058 combination, because we can't reload
2059 it. */
2060 if (curr_alt_offmemok[m]
2061 && MEM_P (*curr_id->operand_loc[m])
2062 && curr_alt[m] == NO_REGS && ! curr_alt_win[m])
2063 continue;
2065 else
2067 /* Operands don't match. Both operands must
2068 allow a reload register, otherwise we
2069 cannot make them match. */
2070 if (curr_alt[m] == NO_REGS)
2071 break;
2072 /* Retroactively mark the operand we had to
2073 match as a loser, if it wasn't already and
2074 it wasn't matched to a register constraint
2075 (e.g it might be matched by memory). */
2076 if (curr_alt_win[m]
2077 && (operand_reg[m] == NULL_RTX
2078 || hard_regno[m] < 0))
2080 losers++;
2081 reload_nregs
2082 += (ira_reg_class_max_nregs[curr_alt[m]]
2083 [GET_MODE (*curr_id->operand_loc[m])]);
2086 /* Prefer matching earlyclobber alternative as
2087 it results in less hard regs required for
2088 the insn than a non-matching earlyclobber
2089 alternative. */
2090 if (curr_static_id->operand[m].early_clobber)
2092 if (lra_dump_file != NULL)
2093 fprintf
2094 (lra_dump_file,
2095 " %d Matching earlyclobber alt:"
2096 " reject--\n",
2097 nop);
2098 reject--;
2100 /* Otherwise we prefer no matching
2101 alternatives because it gives more freedom
2102 in RA. */
2103 else if (operand_reg[nop] == NULL_RTX
2104 || (find_regno_note (curr_insn, REG_DEAD,
2105 REGNO (operand_reg[nop]))
2106 == NULL_RTX))
2108 if (lra_dump_file != NULL)
2109 fprintf
2110 (lra_dump_file,
2111 " %d Matching alt: reject+=2\n",
2112 nop);
2113 reject += 2;
2116 /* If we have to reload this operand and some
2117 previous operand also had to match the same
2118 thing as this operand, we don't know how to do
2119 that. */
2120 if (!match_p || !curr_alt_win[m])
2122 for (i = 0; i < nop; i++)
2123 if (curr_alt_matches[i] == m)
2124 break;
2125 if (i < nop)
2126 break;
2128 else
2129 did_match = true;
2131 /* This can be fixed with reloads if the operand
2132 we are supposed to match can be fixed with
2133 reloads. */
2134 badop = false;
2135 this_alternative = curr_alt[m];
2136 COPY_HARD_REG_SET (this_alternative_set, curr_alt_set[m]);
2137 winreg = this_alternative != NO_REGS;
2138 break;
2141 case 'g':
2142 if (MEM_P (op)
2143 || general_constant_p (op)
2144 || spilled_pseudo_p (op))
2145 win = true;
2146 cl = GENERAL_REGS;
2147 goto reg;
2149 default:
2150 cn = lookup_constraint (p);
2151 switch (get_constraint_type (cn))
2153 case CT_REGISTER:
2154 cl = reg_class_for_constraint (cn);
2155 if (cl != NO_REGS)
2156 goto reg;
2157 break;
2159 case CT_CONST_INT:
2160 if (CONST_INT_P (op)
2161 && insn_const_int_ok_for_constraint (INTVAL (op), cn))
2162 win = true;
2163 break;
2165 case CT_MEMORY:
2166 if (MEM_P (op)
2167 && satisfies_memory_constraint_p (op, cn))
2168 win = true;
2169 else if (spilled_pseudo_p (op))
2170 win = true;
2172 /* If we didn't already win, we can reload constants
2173 via force_const_mem or put the pseudo value into
2174 memory, or make other memory by reloading the
2175 address like for 'o'. */
2176 if (CONST_POOL_OK_P (mode, op)
2177 || MEM_P (op) || REG_P (op)
2178 /* We can restore the equiv insn by a
2179 reload. */
2180 || equiv_substition_p[nop])
2181 badop = false;
2182 constmemok = true;
2183 offmemok = true;
2184 break;
2186 case CT_ADDRESS:
2187 /* If we didn't already win, we can reload the address
2188 into a base register. */
2189 if (satisfies_address_constraint_p (op, cn))
2190 win = true;
2191 cl = base_reg_class (VOIDmode, ADDR_SPACE_GENERIC,
2192 ADDRESS, SCRATCH);
2193 badop = false;
2194 goto reg;
2196 case CT_FIXED_FORM:
2197 if (constraint_satisfied_p (op, cn))
2198 win = true;
2199 break;
2201 case CT_SPECIAL_MEMORY:
2202 if (MEM_P (op)
2203 && satisfies_memory_constraint_p (op, cn))
2204 win = true;
2205 else if (spilled_pseudo_p (op))
2206 win = true;
2207 break;
2209 break;
2211 reg:
2212 this_alternative = reg_class_subunion[this_alternative][cl];
2213 IOR_HARD_REG_SET (this_alternative_set,
2214 reg_class_contents[cl]);
2215 if (costly_p)
2217 this_costly_alternative
2218 = reg_class_subunion[this_costly_alternative][cl];
2219 IOR_HARD_REG_SET (this_costly_alternative_set,
2220 reg_class_contents[cl]);
2222 if (mode == BLKmode)
2223 break;
2224 winreg = true;
2225 if (REG_P (op))
2227 if (hard_regno[nop] >= 0
2228 && in_hard_reg_set_p (this_alternative_set,
2229 mode, hard_regno[nop]))
2230 win = true;
2231 else if (hard_regno[nop] < 0
2232 && in_class_p (op, this_alternative, NULL))
2233 win = true;
2235 break;
2237 if (c != ' ' && c != '\t')
2238 costly_p = c == '*';
2240 while ((p += len), c);
2242 scratch_p = (operand_reg[nop] != NULL_RTX
2243 && lra_former_scratch_p (REGNO (operand_reg[nop])));
2244 /* Record which operands fit this alternative. */
2245 if (win)
2247 this_alternative_win = true;
2248 if (operand_reg[nop] != NULL_RTX)
2250 if (hard_regno[nop] >= 0)
2252 if (in_hard_reg_set_p (this_costly_alternative_set,
2253 mode, hard_regno[nop]))
2255 if (lra_dump_file != NULL)
2256 fprintf (lra_dump_file,
2257 " %d Costly set: reject++\n",
2258 nop);
2259 reject++;
2262 else
2264 /* Prefer won reg to spilled pseudo under other
2265 equal conditions for possibe inheritance. */
2266 if (! scratch_p)
2268 if (lra_dump_file != NULL)
2269 fprintf
2270 (lra_dump_file,
2271 " %d Non pseudo reload: reject++\n",
2272 nop);
2273 reject++;
2275 if (in_class_p (operand_reg[nop],
2276 this_costly_alternative, NULL))
2278 if (lra_dump_file != NULL)
2279 fprintf
2280 (lra_dump_file,
2281 " %d Non pseudo costly reload:"
2282 " reject++\n",
2283 nop);
2284 reject++;
2287 /* We simulate the behavior of old reload here.
2288 Although scratches need hard registers and it
2289 might result in spilling other pseudos, no reload
2290 insns are generated for the scratches. So it
2291 might cost something but probably less than old
2292 reload pass believes. */
2293 if (scratch_p)
2295 if (lra_dump_file != NULL)
2296 fprintf (lra_dump_file,
2297 " %d Scratch win: reject+=2\n",
2298 nop);
2299 reject += 2;
2303 else if (did_match)
2304 this_alternative_match_win = true;
2305 else
2307 int const_to_mem = 0;
2308 bool no_regs_p;
2310 reject += op_reject;
2311 /* Never do output reload of stack pointer. It makes
2312 impossible to do elimination when SP is changed in
2313 RTL. */
2314 if (op == stack_pointer_rtx && ! frame_pointer_needed
2315 && curr_static_id->operand[nop].type != OP_IN)
2316 goto fail;
2318 /* If this alternative asks for a specific reg class, see if there
2319 is at least one allocatable register in that class. */
2320 no_regs_p
2321 = (this_alternative == NO_REGS
2322 || (hard_reg_set_subset_p
2323 (reg_class_contents[this_alternative],
2324 lra_no_alloc_regs)));
2326 /* For asms, verify that the class for this alternative is possible
2327 for the mode that is specified. */
2328 if (!no_regs_p && INSN_CODE (curr_insn) < 0)
2330 int i;
2331 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
2332 if (HARD_REGNO_MODE_OK (i, mode)
2333 && in_hard_reg_set_p (reg_class_contents[this_alternative],
2334 mode, i))
2335 break;
2336 if (i == FIRST_PSEUDO_REGISTER)
2337 winreg = false;
2340 /* If this operand accepts a register, and if the
2341 register class has at least one allocatable register,
2342 then this operand can be reloaded. */
2343 if (winreg && !no_regs_p)
2344 badop = false;
2346 if (badop)
2348 if (lra_dump_file != NULL)
2349 fprintf (lra_dump_file,
2350 " alt=%d: Bad operand -- refuse\n",
2351 nalt);
2352 goto fail;
2355 if (this_alternative != NO_REGS)
2357 HARD_REG_SET available_regs;
2359 COPY_HARD_REG_SET (available_regs,
2360 reg_class_contents[this_alternative]);
2361 AND_COMPL_HARD_REG_SET
2362 (available_regs,
2363 ira_prohibited_class_mode_regs[this_alternative][mode]);
2364 AND_COMPL_HARD_REG_SET (available_regs, lra_no_alloc_regs);
2365 if (hard_reg_set_empty_p (available_regs))
2367 /* There are no hard regs holding a value of given
2368 mode. */
2369 if (offmemok)
2371 this_alternative = NO_REGS;
2372 if (lra_dump_file != NULL)
2373 fprintf (lra_dump_file,
2374 " %d Using memory because of"
2375 " a bad mode: reject+=2\n",
2376 nop);
2377 reject += 2;
2379 else
2381 if (lra_dump_file != NULL)
2382 fprintf (lra_dump_file,
2383 " alt=%d: Wrong mode -- refuse\n",
2384 nalt);
2385 goto fail;
2390 /* If not assigned pseudo has a class which a subset of
2391 required reg class, it is a less costly alternative
2392 as the pseudo still can get a hard reg of necessary
2393 class. */
2394 if (! no_regs_p && REG_P (op) && hard_regno[nop] < 0
2395 && (cl = get_reg_class (REGNO (op))) != NO_REGS
2396 && ira_class_subset_p[this_alternative][cl])
2398 if (lra_dump_file != NULL)
2399 fprintf
2400 (lra_dump_file,
2401 " %d Super set class reg: reject-=3\n", nop);
2402 reject -= 3;
2405 this_alternative_offmemok = offmemok;
2406 if (this_costly_alternative != NO_REGS)
2408 if (lra_dump_file != NULL)
2409 fprintf (lra_dump_file,
2410 " %d Costly loser: reject++\n", nop);
2411 reject++;
2413 /* If the operand is dying, has a matching constraint,
2414 and satisfies constraints of the matched operand
2415 which failed to satisfy the own constraints, most probably
2416 the reload for this operand will be gone. */
2417 if (this_alternative_matches >= 0
2418 && !curr_alt_win[this_alternative_matches]
2419 && REG_P (op)
2420 && find_regno_note (curr_insn, REG_DEAD, REGNO (op))
2421 && (hard_regno[nop] >= 0
2422 ? in_hard_reg_set_p (this_alternative_set,
2423 mode, hard_regno[nop])
2424 : in_class_p (op, this_alternative, NULL)))
2426 if (lra_dump_file != NULL)
2427 fprintf
2428 (lra_dump_file,
2429 " %d Dying matched operand reload: reject++\n",
2430 nop);
2431 reject++;
2433 else
2435 /* Strict_low_part requires to reload the register
2436 not the sub-register. In this case we should
2437 check that a final reload hard reg can hold the
2438 value mode. */
2439 if (curr_static_id->operand[nop].strict_low
2440 && REG_P (op)
2441 && hard_regno[nop] < 0
2442 && GET_CODE (*curr_id->operand_loc[nop]) == SUBREG
2443 && ira_class_hard_regs_num[this_alternative] > 0
2444 && ! HARD_REGNO_MODE_OK (ira_class_hard_regs
2445 [this_alternative][0],
2446 GET_MODE
2447 (*curr_id->operand_loc[nop])))
2449 if (lra_dump_file != NULL)
2450 fprintf
2451 (lra_dump_file,
2452 " alt=%d: Strict low subreg reload -- refuse\n",
2453 nalt);
2454 goto fail;
2456 losers++;
2458 if (operand_reg[nop] != NULL_RTX
2459 /* Output operands and matched input operands are
2460 not inherited. The following conditions do not
2461 exactly describe the previous statement but they
2462 are pretty close. */
2463 && curr_static_id->operand[nop].type != OP_OUT
2464 && (this_alternative_matches < 0
2465 || curr_static_id->operand[nop].type != OP_IN))
2467 int last_reload = (lra_reg_info[ORIGINAL_REGNO
2468 (operand_reg[nop])]
2469 .last_reload);
2471 /* The value of reload_sum has sense only if we
2472 process insns in their order. It happens only on
2473 the first constraints sub-pass when we do most of
2474 reload work. */
2475 if (lra_constraint_iter == 1 && last_reload > bb_reload_num)
2476 reload_sum += last_reload - bb_reload_num;
2478 /* If this is a constant that is reloaded into the
2479 desired class by copying it to memory first, count
2480 that as another reload. This is consistent with
2481 other code and is required to avoid choosing another
2482 alternative when the constant is moved into memory.
2483 Note that the test here is precisely the same as in
2484 the code below that calls force_const_mem. */
2485 if (CONST_POOL_OK_P (mode, op)
2486 && ((targetm.preferred_reload_class
2487 (op, this_alternative) == NO_REGS)
2488 || no_input_reloads_p))
2490 const_to_mem = 1;
2491 if (! no_regs_p)
2492 losers++;
2495 /* Alternative loses if it requires a type of reload not
2496 permitted for this insn. We can always reload
2497 objects with a REG_UNUSED note. */
2498 if ((curr_static_id->operand[nop].type != OP_IN
2499 && no_output_reloads_p
2500 && ! find_reg_note (curr_insn, REG_UNUSED, op))
2501 || (curr_static_id->operand[nop].type != OP_OUT
2502 && no_input_reloads_p && ! const_to_mem)
2503 || (this_alternative_matches >= 0
2504 && (no_input_reloads_p
2505 || (no_output_reloads_p
2506 && (curr_static_id->operand
2507 [this_alternative_matches].type != OP_IN)
2508 && ! find_reg_note (curr_insn, REG_UNUSED,
2509 no_subreg_reg_operand
2510 [this_alternative_matches])))))
2512 if (lra_dump_file != NULL)
2513 fprintf
2514 (lra_dump_file,
2515 " alt=%d: No input/otput reload -- refuse\n",
2516 nalt);
2517 goto fail;
2520 /* Alternative loses if it required class pseudo can not
2521 hold value of required mode. Such insns can be
2522 described by insn definitions with mode iterators. */
2523 if (GET_MODE (*curr_id->operand_loc[nop]) != VOIDmode
2524 && ! hard_reg_set_empty_p (this_alternative_set)
2525 /* It is common practice for constraints to use a
2526 class which does not have actually enough regs to
2527 hold the value (e.g. x86 AREG for mode requiring
2528 more one general reg). Therefore we have 2
2529 conditions to check that the reload pseudo can
2530 not hold the mode value. */
2531 && ! HARD_REGNO_MODE_OK (ira_class_hard_regs
2532 [this_alternative][0],
2533 GET_MODE (*curr_id->operand_loc[nop]))
2534 /* The above condition is not enough as the first
2535 reg in ira_class_hard_regs can be not aligned for
2536 multi-words mode values. */
2537 && (prohibited_class_reg_set_mode_p
2538 (this_alternative, this_alternative_set,
2539 GET_MODE (*curr_id->operand_loc[nop]))))
2541 if (lra_dump_file != NULL)
2542 fprintf (lra_dump_file,
2543 " alt=%d: reload pseudo for op %d "
2544 " can not hold the mode value -- refuse\n",
2545 nalt, nop);
2546 goto fail;
2549 /* Check strong discouragement of reload of non-constant
2550 into class THIS_ALTERNATIVE. */
2551 if (! CONSTANT_P (op) && ! no_regs_p
2552 && (targetm.preferred_reload_class
2553 (op, this_alternative) == NO_REGS
2554 || (curr_static_id->operand[nop].type == OP_OUT
2555 && (targetm.preferred_output_reload_class
2556 (op, this_alternative) == NO_REGS))))
2558 if (lra_dump_file != NULL)
2559 fprintf (lra_dump_file,
2560 " %d Non-prefered reload: reject+=%d\n",
2561 nop, LRA_MAX_REJECT);
2562 reject += LRA_MAX_REJECT;
2565 if (! (MEM_P (op) && offmemok)
2566 && ! (const_to_mem && constmemok))
2568 /* We prefer to reload pseudos over reloading other
2569 things, since such reloads may be able to be
2570 eliminated later. So bump REJECT in other cases.
2571 Don't do this in the case where we are forcing a
2572 constant into memory and it will then win since
2573 we don't want to have a different alternative
2574 match then. */
2575 if (! (REG_P (op) && REGNO (op) >= FIRST_PSEUDO_REGISTER))
2577 if (lra_dump_file != NULL)
2578 fprintf
2579 (lra_dump_file,
2580 " %d Non-pseudo reload: reject+=2\n",
2581 nop);
2582 reject += 2;
2585 if (! no_regs_p)
2586 reload_nregs
2587 += ira_reg_class_max_nregs[this_alternative][mode];
2589 if (SMALL_REGISTER_CLASS_P (this_alternative))
2591 if (lra_dump_file != NULL)
2592 fprintf
2593 (lra_dump_file,
2594 " %d Small class reload: reject+=%d\n",
2595 nop, LRA_LOSER_COST_FACTOR / 2);
2596 reject += LRA_LOSER_COST_FACTOR / 2;
2600 /* We are trying to spill pseudo into memory. It is
2601 usually more costly than moving to a hard register
2602 although it might takes the same number of
2603 reloads.
2605 Non-pseudo spill may happen also. Suppose a target allows both
2606 register and memory in the operand constraint alternatives,
2607 then it's typical that an eliminable register has a substition
2608 of "base + offset" which can either be reloaded by a simple
2609 "new_reg <= base + offset" which will match the register
2610 constraint, or a similar reg addition followed by further spill
2611 to and reload from memory which will match the memory
2612 constraint, but this memory spill will be much more costly
2613 usually.
2615 Code below increases the reject for both pseudo and non-pseudo
2616 spill. */
2617 if (no_regs_p
2618 && !(MEM_P (op) && offmemok)
2619 && !(REG_P (op) && hard_regno[nop] < 0))
2621 if (lra_dump_file != NULL)
2622 fprintf
2623 (lra_dump_file,
2624 " %d Spill %spseudo into memory: reject+=3\n",
2625 nop, REG_P (op) ? "" : "Non-");
2626 reject += 3;
2627 if (VECTOR_MODE_P (mode))
2629 /* Spilling vectors into memory is usually more
2630 costly as they contain big values. */
2631 if (lra_dump_file != NULL)
2632 fprintf
2633 (lra_dump_file,
2634 " %d Spill vector pseudo: reject+=2\n",
2635 nop);
2636 reject += 2;
2640 #ifdef SECONDARY_MEMORY_NEEDED
2641 /* If reload requires moving value through secondary
2642 memory, it will need one more insn at least. */
2643 if (this_alternative != NO_REGS
2644 && REG_P (op) && (cl = get_reg_class (REGNO (op))) != NO_REGS
2645 && ((curr_static_id->operand[nop].type != OP_OUT
2646 && SECONDARY_MEMORY_NEEDED (cl, this_alternative,
2647 GET_MODE (op)))
2648 || (curr_static_id->operand[nop].type != OP_IN
2649 && SECONDARY_MEMORY_NEEDED (this_alternative, cl,
2650 GET_MODE (op)))))
2651 losers++;
2652 #endif
2653 /* Input reloads can be inherited more often than output
2654 reloads can be removed, so penalize output
2655 reloads. */
2656 if (!REG_P (op) || curr_static_id->operand[nop].type != OP_IN)
2658 if (lra_dump_file != NULL)
2659 fprintf
2660 (lra_dump_file,
2661 " %d Non input pseudo reload: reject++\n",
2662 nop);
2663 reject++;
2667 if (early_clobber_p && ! scratch_p)
2669 if (lra_dump_file != NULL)
2670 fprintf (lra_dump_file,
2671 " %d Early clobber: reject++\n", nop);
2672 reject++;
2674 /* ??? We check early clobbers after processing all operands
2675 (see loop below) and there we update the costs more.
2676 Should we update the cost (may be approximately) here
2677 because of early clobber register reloads or it is a rare
2678 or non-important thing to be worth to do it. */
2679 overall = losers * LRA_LOSER_COST_FACTOR + reject;
2680 if ((best_losers == 0 || losers != 0) && best_overall < overall)
2682 if (lra_dump_file != NULL)
2683 fprintf (lra_dump_file,
2684 " alt=%d,overall=%d,losers=%d -- refuse\n",
2685 nalt, overall, losers);
2686 goto fail;
2689 curr_alt[nop] = this_alternative;
2690 COPY_HARD_REG_SET (curr_alt_set[nop], this_alternative_set);
2691 curr_alt_win[nop] = this_alternative_win;
2692 curr_alt_match_win[nop] = this_alternative_match_win;
2693 curr_alt_offmemok[nop] = this_alternative_offmemok;
2694 curr_alt_matches[nop] = this_alternative_matches;
2696 if (this_alternative_matches >= 0
2697 && !did_match && !this_alternative_win)
2698 curr_alt_win[this_alternative_matches] = false;
2700 if (early_clobber_p && operand_reg[nop] != NULL_RTX)
2701 early_clobbered_nops[early_clobbered_regs_num++] = nop;
2703 if (curr_insn_set != NULL_RTX && n_operands == 2
2704 /* Prevent processing non-move insns. */
2705 && (GET_CODE (SET_SRC (curr_insn_set)) == SUBREG
2706 || SET_SRC (curr_insn_set) == no_subreg_reg_operand[1])
2707 && ((! curr_alt_win[0] && ! curr_alt_win[1]
2708 && REG_P (no_subreg_reg_operand[0])
2709 && REG_P (no_subreg_reg_operand[1])
2710 && (reg_in_class_p (no_subreg_reg_operand[0], curr_alt[1])
2711 || reg_in_class_p (no_subreg_reg_operand[1], curr_alt[0])))
2712 || (! curr_alt_win[0] && curr_alt_win[1]
2713 && REG_P (no_subreg_reg_operand[1])
2714 && reg_in_class_p (no_subreg_reg_operand[1], curr_alt[0]))
2715 || (curr_alt_win[0] && ! curr_alt_win[1]
2716 && REG_P (no_subreg_reg_operand[0])
2717 && reg_in_class_p (no_subreg_reg_operand[0], curr_alt[1])
2718 && (! CONST_POOL_OK_P (curr_operand_mode[1],
2719 no_subreg_reg_operand[1])
2720 || (targetm.preferred_reload_class
2721 (no_subreg_reg_operand[1],
2722 (enum reg_class) curr_alt[1]) != NO_REGS))
2723 /* If it is a result of recent elimination in move
2724 insn we can transform it into an add still by
2725 using this alternative. */
2726 && GET_CODE (no_subreg_reg_operand[1]) != PLUS)))
2728 /* We have a move insn and a new reload insn will be similar
2729 to the current insn. We should avoid such situation as it
2730 results in LRA cycling. */
2731 overall += LRA_MAX_REJECT;
2733 ok_p = true;
2734 curr_alt_dont_inherit_ops_num = 0;
2735 for (nop = 0; nop < early_clobbered_regs_num; nop++)
2737 int i, j, clobbered_hard_regno, first_conflict_j, last_conflict_j;
2738 HARD_REG_SET temp_set;
2740 i = early_clobbered_nops[nop];
2741 if ((! curr_alt_win[i] && ! curr_alt_match_win[i])
2742 || hard_regno[i] < 0)
2743 continue;
2744 lra_assert (operand_reg[i] != NULL_RTX);
2745 clobbered_hard_regno = hard_regno[i];
2746 CLEAR_HARD_REG_SET (temp_set);
2747 add_to_hard_reg_set (&temp_set, biggest_mode[i], clobbered_hard_regno);
2748 first_conflict_j = last_conflict_j = -1;
2749 for (j = 0; j < n_operands; j++)
2750 if (j == i
2751 /* We don't want process insides of match_operator and
2752 match_parallel because otherwise we would process
2753 their operands once again generating a wrong
2754 code. */
2755 || curr_static_id->operand[j].is_operator)
2756 continue;
2757 else if ((curr_alt_matches[j] == i && curr_alt_match_win[j])
2758 || (curr_alt_matches[i] == j && curr_alt_match_win[i]))
2759 continue;
2760 /* If we don't reload j-th operand, check conflicts. */
2761 else if ((curr_alt_win[j] || curr_alt_match_win[j])
2762 && uses_hard_regs_p (*curr_id->operand_loc[j], temp_set))
2764 if (first_conflict_j < 0)
2765 first_conflict_j = j;
2766 last_conflict_j = j;
2768 if (last_conflict_j < 0)
2769 continue;
2770 /* If earlyclobber operand conflicts with another
2771 non-matching operand which is actually the same register
2772 as the earlyclobber operand, it is better to reload the
2773 another operand as an operand matching the earlyclobber
2774 operand can be also the same. */
2775 if (first_conflict_j == last_conflict_j
2776 && operand_reg[last_conflict_j] != NULL_RTX
2777 && ! curr_alt_match_win[last_conflict_j]
2778 && REGNO (operand_reg[i]) == REGNO (operand_reg[last_conflict_j]))
2780 curr_alt_win[last_conflict_j] = false;
2781 curr_alt_dont_inherit_ops[curr_alt_dont_inherit_ops_num++]
2782 = last_conflict_j;
2783 losers++;
2784 /* Early clobber was already reflected in REJECT. */
2785 lra_assert (reject > 0);
2786 if (lra_dump_file != NULL)
2787 fprintf
2788 (lra_dump_file,
2789 " %d Conflict early clobber reload: reject--\n",
2791 reject--;
2792 overall += LRA_LOSER_COST_FACTOR - 1;
2794 else
2796 /* We need to reload early clobbered register and the
2797 matched registers. */
2798 for (j = 0; j < n_operands; j++)
2799 if (curr_alt_matches[j] == i)
2801 curr_alt_match_win[j] = false;
2802 losers++;
2803 overall += LRA_LOSER_COST_FACTOR;
2805 if (! curr_alt_match_win[i])
2806 curr_alt_dont_inherit_ops[curr_alt_dont_inherit_ops_num++] = i;
2807 else
2809 /* Remember pseudos used for match reloads are never
2810 inherited. */
2811 lra_assert (curr_alt_matches[i] >= 0);
2812 curr_alt_win[curr_alt_matches[i]] = false;
2814 curr_alt_win[i] = curr_alt_match_win[i] = false;
2815 losers++;
2816 /* Early clobber was already reflected in REJECT. */
2817 lra_assert (reject > 0);
2818 if (lra_dump_file != NULL)
2819 fprintf
2820 (lra_dump_file,
2821 " %d Matched conflict early clobber reloads:"
2822 "reject--\n",
2824 reject--;
2825 overall += LRA_LOSER_COST_FACTOR - 1;
2828 if (lra_dump_file != NULL)
2829 fprintf (lra_dump_file, " alt=%d,overall=%d,losers=%d,rld_nregs=%d\n",
2830 nalt, overall, losers, reload_nregs);
2832 /* If this alternative can be made to work by reloading, and it
2833 needs less reloading than the others checked so far, record
2834 it as the chosen goal for reloading. */
2835 if ((best_losers != 0 && losers == 0)
2836 || (((best_losers == 0 && losers == 0)
2837 || (best_losers != 0 && losers != 0))
2838 && (best_overall > overall
2839 || (best_overall == overall
2840 /* If the cost of the reloads is the same,
2841 prefer alternative which requires minimal
2842 number of reload regs. */
2843 && (reload_nregs < best_reload_nregs
2844 || (reload_nregs == best_reload_nregs
2845 && (best_reload_sum < reload_sum
2846 || (best_reload_sum == reload_sum
2847 && nalt < goal_alt_number))))))))
2849 for (nop = 0; nop < n_operands; nop++)
2851 goal_alt_win[nop] = curr_alt_win[nop];
2852 goal_alt_match_win[nop] = curr_alt_match_win[nop];
2853 goal_alt_matches[nop] = curr_alt_matches[nop];
2854 goal_alt[nop] = curr_alt[nop];
2855 goal_alt_offmemok[nop] = curr_alt_offmemok[nop];
2857 goal_alt_dont_inherit_ops_num = curr_alt_dont_inherit_ops_num;
2858 for (nop = 0; nop < curr_alt_dont_inherit_ops_num; nop++)
2859 goal_alt_dont_inherit_ops[nop] = curr_alt_dont_inherit_ops[nop];
2860 goal_alt_swapped = curr_swapped;
2861 best_overall = overall;
2862 best_losers = losers;
2863 best_reload_nregs = reload_nregs;
2864 best_reload_sum = reload_sum;
2865 goal_alt_number = nalt;
2867 if (losers == 0)
2868 /* Everything is satisfied. Do not process alternatives
2869 anymore. */
2870 break;
2871 fail:
2874 return ok_p;
2877 /* Make reload base reg from address AD. */
2878 static rtx
2879 base_to_reg (struct address_info *ad)
2881 enum reg_class cl;
2882 int code = -1;
2883 rtx new_inner = NULL_RTX;
2884 rtx new_reg = NULL_RTX;
2885 rtx_insn *insn;
2886 rtx_insn *last_insn = get_last_insn();
2888 lra_assert (ad->base == ad->base_term && ad->disp == ad->disp_term);
2889 cl = base_reg_class (ad->mode, ad->as, ad->base_outer_code,
2890 get_index_code (ad));
2891 new_reg = lra_create_new_reg (GET_MODE (*ad->base_term), NULL_RTX,
2892 cl, "base");
2893 new_inner = simplify_gen_binary (PLUS, GET_MODE (new_reg), new_reg,
2894 ad->disp_term == NULL
2895 ? gen_int_mode (0, ad->mode)
2896 : *ad->disp_term);
2897 if (!valid_address_p (ad->mode, new_inner, ad->as))
2898 return NULL_RTX;
2899 insn = emit_insn (gen_rtx_SET (new_reg, *ad->base_term));
2900 code = recog_memoized (insn);
2901 if (code < 0)
2903 delete_insns_since (last_insn);
2904 return NULL_RTX;
2907 return new_inner;
2910 /* Make reload base reg + disp from address AD. Return the new pseudo. */
2911 static rtx
2912 base_plus_disp_to_reg (struct address_info *ad)
2914 enum reg_class cl;
2915 rtx new_reg;
2917 lra_assert (ad->base == ad->base_term && ad->disp == ad->disp_term);
2918 cl = base_reg_class (ad->mode, ad->as, ad->base_outer_code,
2919 get_index_code (ad));
2920 new_reg = lra_create_new_reg (GET_MODE (*ad->base_term), NULL_RTX,
2921 cl, "base + disp");
2922 lra_emit_add (new_reg, *ad->base_term, *ad->disp_term);
2923 return new_reg;
2926 /* Make reload of index part of address AD. Return the new
2927 pseudo. */
2928 static rtx
2929 index_part_to_reg (struct address_info *ad)
2931 rtx new_reg;
2933 new_reg = lra_create_new_reg (GET_MODE (*ad->index), NULL_RTX,
2934 INDEX_REG_CLASS, "index term");
2935 expand_mult (GET_MODE (*ad->index), *ad->index_term,
2936 GEN_INT (get_index_scale (ad)), new_reg, 1);
2937 return new_reg;
2940 /* Return true if we can add a displacement to address AD, even if that
2941 makes the address invalid. The fix-up code requires any new address
2942 to be the sum of the BASE_TERM, INDEX and DISP_TERM fields. */
2943 static bool
2944 can_add_disp_p (struct address_info *ad)
2946 return (!ad->autoinc_p
2947 && ad->segment == NULL
2948 && ad->base == ad->base_term
2949 && ad->disp == ad->disp_term);
2952 /* Make equiv substitution in address AD. Return true if a substitution
2953 was made. */
2954 static bool
2955 equiv_address_substitution (struct address_info *ad)
2957 rtx base_reg, new_base_reg, index_reg, new_index_reg, *base_term, *index_term;
2958 HOST_WIDE_INT disp, scale;
2959 bool change_p;
2961 base_term = strip_subreg (ad->base_term);
2962 if (base_term == NULL)
2963 base_reg = new_base_reg = NULL_RTX;
2964 else
2966 base_reg = *base_term;
2967 new_base_reg = get_equiv_with_elimination (base_reg, curr_insn);
2969 index_term = strip_subreg (ad->index_term);
2970 if (index_term == NULL)
2971 index_reg = new_index_reg = NULL_RTX;
2972 else
2974 index_reg = *index_term;
2975 new_index_reg = get_equiv_with_elimination (index_reg, curr_insn);
2977 if (base_reg == new_base_reg && index_reg == new_index_reg)
2978 return false;
2979 disp = 0;
2980 change_p = false;
2981 if (lra_dump_file != NULL)
2983 fprintf (lra_dump_file, "Changing address in insn %d ",
2984 INSN_UID (curr_insn));
2985 dump_value_slim (lra_dump_file, *ad->outer, 1);
2987 if (base_reg != new_base_reg)
2989 if (REG_P (new_base_reg))
2991 *base_term = new_base_reg;
2992 change_p = true;
2994 else if (GET_CODE (new_base_reg) == PLUS
2995 && REG_P (XEXP (new_base_reg, 0))
2996 && CONST_INT_P (XEXP (new_base_reg, 1))
2997 && can_add_disp_p (ad))
2999 disp += INTVAL (XEXP (new_base_reg, 1));
3000 *base_term = XEXP (new_base_reg, 0);
3001 change_p = true;
3003 if (ad->base_term2 != NULL)
3004 *ad->base_term2 = *ad->base_term;
3006 if (index_reg != new_index_reg)
3008 if (REG_P (new_index_reg))
3010 *index_term = new_index_reg;
3011 change_p = true;
3013 else if (GET_CODE (new_index_reg) == PLUS
3014 && REG_P (XEXP (new_index_reg, 0))
3015 && CONST_INT_P (XEXP (new_index_reg, 1))
3016 && can_add_disp_p (ad)
3017 && (scale = get_index_scale (ad)))
3019 disp += INTVAL (XEXP (new_index_reg, 1)) * scale;
3020 *index_term = XEXP (new_index_reg, 0);
3021 change_p = true;
3024 if (disp != 0)
3026 if (ad->disp != NULL)
3027 *ad->disp = plus_constant (GET_MODE (*ad->inner), *ad->disp, disp);
3028 else
3030 *ad->inner = plus_constant (GET_MODE (*ad->inner), *ad->inner, disp);
3031 update_address (ad);
3033 change_p = true;
3035 if (lra_dump_file != NULL)
3037 if (! change_p)
3038 fprintf (lra_dump_file, " -- no change\n");
3039 else
3041 fprintf (lra_dump_file, " on equiv ");
3042 dump_value_slim (lra_dump_file, *ad->outer, 1);
3043 fprintf (lra_dump_file, "\n");
3046 return change_p;
3049 /* Major function to make reloads for an address in operand NOP or
3050 check its correctness (If CHECK_ONLY_P is true). The supported
3051 cases are:
3053 1) an address that existed before LRA started, at which point it
3054 must have been valid. These addresses are subject to elimination
3055 and may have become invalid due to the elimination offset being out
3056 of range.
3058 2) an address created by forcing a constant to memory
3059 (force_const_to_mem). The initial form of these addresses might
3060 not be valid, and it is this function's job to make them valid.
3062 3) a frame address formed from a register and a (possibly zero)
3063 constant offset. As above, these addresses might not be valid and
3064 this function must make them so.
3066 Add reloads to the lists *BEFORE and *AFTER. We might need to add
3067 reloads to *AFTER because of inc/dec, {pre, post} modify in the
3068 address. Return true for any RTL change.
3070 The function is a helper function which does not produce all
3071 transformations (when CHECK_ONLY_P is false) which can be
3072 necessary. It does just basic steps. To do all necessary
3073 transformations use function process_address. */
3074 static bool
3075 process_address_1 (int nop, bool check_only_p,
3076 rtx_insn **before, rtx_insn **after)
3078 struct address_info ad;
3079 rtx new_reg;
3080 HOST_WIDE_INT scale;
3081 rtx op = *curr_id->operand_loc[nop];
3082 const char *constraint = curr_static_id->operand[nop].constraint;
3083 enum constraint_num cn = lookup_constraint (constraint);
3084 bool change_p = false;
3086 if (MEM_P (op)
3087 && GET_MODE (op) == BLKmode
3088 && GET_CODE (XEXP (op, 0)) == SCRATCH)
3089 return false;
3091 if (insn_extra_address_constraint (cn))
3092 decompose_lea_address (&ad, curr_id->operand_loc[nop]);
3093 /* Do not attempt to decompose arbitrary addresses generated by combine
3094 for asm operands with loose constraints, e.g 'X'. */
3095 else if (MEM_P (op)
3096 && !(get_constraint_type (cn) == CT_FIXED_FORM
3097 && constraint_satisfied_p (op, cn)))
3098 decompose_mem_address (&ad, op);
3099 else if (GET_CODE (op) == SUBREG
3100 && MEM_P (SUBREG_REG (op)))
3101 decompose_mem_address (&ad, SUBREG_REG (op));
3102 else
3103 return false;
3104 /* If INDEX_REG_CLASS is assigned to base_term already and isn't to
3105 index_term, swap them so to avoid assigning INDEX_REG_CLASS to both
3106 when INDEX_REG_CLASS is a single register class. */
3107 if (ad.base_term != NULL
3108 && ad.index_term != NULL
3109 && ira_class_hard_regs_num[INDEX_REG_CLASS] == 1
3110 && REG_P (*ad.base_term)
3111 && REG_P (*ad.index_term)
3112 && in_class_p (*ad.base_term, INDEX_REG_CLASS, NULL)
3113 && ! in_class_p (*ad.index_term, INDEX_REG_CLASS, NULL))
3115 std::swap (ad.base, ad.index);
3116 std::swap (ad.base_term, ad.index_term);
3118 if (! check_only_p)
3119 change_p = equiv_address_substitution (&ad);
3120 if (ad.base_term != NULL
3121 && (process_addr_reg
3122 (ad.base_term, check_only_p, before,
3123 (ad.autoinc_p
3124 && !(REG_P (*ad.base_term)
3125 && find_regno_note (curr_insn, REG_DEAD,
3126 REGNO (*ad.base_term)) != NULL_RTX)
3127 ? after : NULL),
3128 base_reg_class (ad.mode, ad.as, ad.base_outer_code,
3129 get_index_code (&ad)))))
3131 change_p = true;
3132 if (ad.base_term2 != NULL)
3133 *ad.base_term2 = *ad.base_term;
3135 if (ad.index_term != NULL
3136 && process_addr_reg (ad.index_term, check_only_p,
3137 before, NULL, INDEX_REG_CLASS))
3138 change_p = true;
3140 /* Target hooks sometimes don't treat extra-constraint addresses as
3141 legitimate address_operands, so handle them specially. */
3142 if (insn_extra_address_constraint (cn)
3143 && satisfies_address_constraint_p (&ad, cn))
3144 return change_p;
3146 if (check_only_p)
3147 return change_p;
3149 /* There are three cases where the shape of *AD.INNER may now be invalid:
3151 1) the original address was valid, but either elimination or
3152 equiv_address_substitution was applied and that made
3153 the address invalid.
3155 2) the address is an invalid symbolic address created by
3156 force_const_to_mem.
3158 3) the address is a frame address with an invalid offset.
3160 4) the address is a frame address with an invalid base.
3162 All these cases involve a non-autoinc address, so there is no
3163 point revalidating other types. */
3164 if (ad.autoinc_p || valid_address_p (&ad))
3165 return change_p;
3167 /* Any index existed before LRA started, so we can assume that the
3168 presence and shape of the index is valid. */
3169 push_to_sequence (*before);
3170 lra_assert (ad.disp == ad.disp_term);
3171 if (ad.base == NULL)
3173 if (ad.index == NULL)
3175 rtx_insn *insn;
3176 rtx_insn *last = get_last_insn ();
3177 int code = -1;
3178 enum reg_class cl = base_reg_class (ad.mode, ad.as,
3179 SCRATCH, SCRATCH);
3180 rtx addr = *ad.inner;
3182 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "addr");
3183 if (HAVE_lo_sum)
3185 /* addr => lo_sum (new_base, addr), case (2) above. */
3186 insn = emit_insn (gen_rtx_SET
3187 (new_reg,
3188 gen_rtx_HIGH (Pmode, copy_rtx (addr))));
3189 code = recog_memoized (insn);
3190 if (code >= 0)
3192 *ad.inner = gen_rtx_LO_SUM (Pmode, new_reg, addr);
3193 if (! valid_address_p (ad.mode, *ad.outer, ad.as))
3195 /* Try to put lo_sum into register. */
3196 insn = emit_insn (gen_rtx_SET
3197 (new_reg,
3198 gen_rtx_LO_SUM (Pmode, new_reg, addr)));
3199 code = recog_memoized (insn);
3200 if (code >= 0)
3202 *ad.inner = new_reg;
3203 if (! valid_address_p (ad.mode, *ad.outer, ad.as))
3205 *ad.inner = addr;
3206 code = -1;
3212 if (code < 0)
3213 delete_insns_since (last);
3216 if (code < 0)
3218 /* addr => new_base, case (2) above. */
3219 lra_emit_move (new_reg, addr);
3221 for (insn = last == NULL_RTX ? get_insns () : NEXT_INSN (last);
3222 insn != NULL_RTX;
3223 insn = NEXT_INSN (insn))
3224 if (recog_memoized (insn) < 0)
3225 break;
3226 if (insn != NULL_RTX)
3228 /* Do nothing if we cannot generate right insns.
3229 This is analogous to reload pass behavior. */
3230 delete_insns_since (last);
3231 end_sequence ();
3232 return false;
3234 *ad.inner = new_reg;
3237 else
3239 /* index * scale + disp => new base + index * scale,
3240 case (1) above. */
3241 enum reg_class cl = base_reg_class (ad.mode, ad.as, PLUS,
3242 GET_CODE (*ad.index));
3244 lra_assert (INDEX_REG_CLASS != NO_REGS);
3245 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "disp");
3246 lra_emit_move (new_reg, *ad.disp);
3247 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3248 new_reg, *ad.index);
3251 else if (ad.index == NULL)
3253 int regno;
3254 enum reg_class cl;
3255 rtx set;
3256 rtx_insn *insns, *last_insn;
3257 /* Try to reload base into register only if the base is invalid
3258 for the address but with valid offset, case (4) above. */
3259 start_sequence ();
3260 new_reg = base_to_reg (&ad);
3262 /* base + disp => new base, cases (1) and (3) above. */
3263 /* Another option would be to reload the displacement into an
3264 index register. However, postreload has code to optimize
3265 address reloads that have the same base and different
3266 displacements, so reloading into an index register would
3267 not necessarily be a win. */
3268 if (new_reg == NULL_RTX)
3269 new_reg = base_plus_disp_to_reg (&ad);
3270 insns = get_insns ();
3271 last_insn = get_last_insn ();
3272 /* If we generated at least two insns, try last insn source as
3273 an address. If we succeed, we generate one less insn. */
3274 if (last_insn != insns && (set = single_set (last_insn)) != NULL_RTX
3275 && GET_CODE (SET_SRC (set)) == PLUS
3276 && REG_P (XEXP (SET_SRC (set), 0))
3277 && CONSTANT_P (XEXP (SET_SRC (set), 1)))
3279 *ad.inner = SET_SRC (set);
3280 if (valid_address_p (ad.mode, *ad.outer, ad.as))
3282 *ad.base_term = XEXP (SET_SRC (set), 0);
3283 *ad.disp_term = XEXP (SET_SRC (set), 1);
3284 cl = base_reg_class (ad.mode, ad.as, ad.base_outer_code,
3285 get_index_code (&ad));
3286 regno = REGNO (*ad.base_term);
3287 if (regno >= FIRST_PSEUDO_REGISTER
3288 && cl != lra_get_allocno_class (regno))
3289 lra_change_class (regno, cl, " Change to", true);
3290 new_reg = SET_SRC (set);
3291 delete_insns_since (PREV_INSN (last_insn));
3294 /* Try if target can split displacement into legitimite new disp
3295 and offset. If it's the case, we replace the last insn with
3296 insns for base + offset => new_reg and set new_reg + new disp
3297 to *ad.inner. */
3298 last_insn = get_last_insn ();
3299 if ((set = single_set (last_insn)) != NULL_RTX
3300 && GET_CODE (SET_SRC (set)) == PLUS
3301 && REG_P (XEXP (SET_SRC (set), 0))
3302 && REGNO (XEXP (SET_SRC (set), 0)) < FIRST_PSEUDO_REGISTER
3303 && CONST_INT_P (XEXP (SET_SRC (set), 1)))
3305 rtx addend, disp = XEXP (SET_SRC (set), 1);
3306 if (targetm.legitimize_address_displacement (&disp, &addend,
3307 ad.mode))
3309 rtx_insn *new_insns;
3310 start_sequence ();
3311 lra_emit_add (new_reg, XEXP (SET_SRC (set), 0), addend);
3312 new_insns = get_insns ();
3313 end_sequence ();
3314 new_reg = gen_rtx_PLUS (Pmode, new_reg, disp);
3315 delete_insns_since (PREV_INSN (last_insn));
3316 add_insn (new_insns);
3317 insns = get_insns ();
3320 end_sequence ();
3321 emit_insn (insns);
3322 *ad.inner = new_reg;
3324 else if (ad.disp_term != NULL)
3326 /* base + scale * index + disp => new base + scale * index,
3327 case (1) above. */
3328 new_reg = base_plus_disp_to_reg (&ad);
3329 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3330 new_reg, *ad.index);
3332 else if ((scale = get_index_scale (&ad)) == 1)
3334 /* The last transformation to one reg will be made in
3335 curr_insn_transform function. */
3336 end_sequence ();
3337 return false;
3339 else if (scale != 0)
3341 /* base + scale * index => base + new_reg,
3342 case (1) above.
3343 Index part of address may become invalid. For example, we
3344 changed pseudo on the equivalent memory and a subreg of the
3345 pseudo onto the memory of different mode for which the scale is
3346 prohibitted. */
3347 new_reg = index_part_to_reg (&ad);
3348 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3349 *ad.base_term, new_reg);
3351 else
3353 enum reg_class cl = base_reg_class (ad.mode, ad.as,
3354 SCRATCH, SCRATCH);
3355 rtx addr = *ad.inner;
3357 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "addr");
3358 /* addr => new_base. */
3359 lra_emit_move (new_reg, addr);
3360 *ad.inner = new_reg;
3362 *before = get_insns ();
3363 end_sequence ();
3364 return true;
3367 /* If CHECK_ONLY_P is false, do address reloads until it is necessary.
3368 Use process_address_1 as a helper function. Return true for any
3369 RTL changes.
3371 If CHECK_ONLY_P is true, just check address correctness. Return
3372 false if the address correct. */
3373 static bool
3374 process_address (int nop, bool check_only_p,
3375 rtx_insn **before, rtx_insn **after)
3377 bool res = false;
3379 while (process_address_1 (nop, check_only_p, before, after))
3381 if (check_only_p)
3382 return true;
3383 res = true;
3385 return res;
3388 /* Emit insns to reload VALUE into a new register. VALUE is an
3389 auto-increment or auto-decrement RTX whose operand is a register or
3390 memory location; so reloading involves incrementing that location.
3391 IN is either identical to VALUE, or some cheaper place to reload
3392 value being incremented/decremented from.
3394 INC_AMOUNT is the number to increment or decrement by (always
3395 positive and ignored for POST_MODIFY/PRE_MODIFY).
3397 Return pseudo containing the result. */
3398 static rtx
3399 emit_inc (enum reg_class new_rclass, rtx in, rtx value, int inc_amount)
3401 /* REG or MEM to be copied and incremented. */
3402 rtx incloc = XEXP (value, 0);
3403 /* Nonzero if increment after copying. */
3404 int post = (GET_CODE (value) == POST_DEC || GET_CODE (value) == POST_INC
3405 || GET_CODE (value) == POST_MODIFY);
3406 rtx_insn *last;
3407 rtx inc;
3408 rtx_insn *add_insn;
3409 int code;
3410 rtx real_in = in == value ? incloc : in;
3411 rtx result;
3412 bool plus_p = true;
3414 if (GET_CODE (value) == PRE_MODIFY || GET_CODE (value) == POST_MODIFY)
3416 lra_assert (GET_CODE (XEXP (value, 1)) == PLUS
3417 || GET_CODE (XEXP (value, 1)) == MINUS);
3418 lra_assert (rtx_equal_p (XEXP (XEXP (value, 1), 0), XEXP (value, 0)));
3419 plus_p = GET_CODE (XEXP (value, 1)) == PLUS;
3420 inc = XEXP (XEXP (value, 1), 1);
3422 else
3424 if (GET_CODE (value) == PRE_DEC || GET_CODE (value) == POST_DEC)
3425 inc_amount = -inc_amount;
3427 inc = GEN_INT (inc_amount);
3430 if (! post && REG_P (incloc))
3431 result = incloc;
3432 else
3433 result = lra_create_new_reg (GET_MODE (value), value, new_rclass,
3434 "INC/DEC result");
3436 if (real_in != result)
3438 /* First copy the location to the result register. */
3439 lra_assert (REG_P (result));
3440 emit_insn (gen_move_insn (result, real_in));
3443 /* We suppose that there are insns to add/sub with the constant
3444 increment permitted in {PRE/POST)_{DEC/INC/MODIFY}. At least the
3445 old reload worked with this assumption. If the assumption
3446 becomes wrong, we should use approach in function
3447 base_plus_disp_to_reg. */
3448 if (in == value)
3450 /* See if we can directly increment INCLOC. */
3451 last = get_last_insn ();
3452 add_insn = emit_insn (plus_p
3453 ? gen_add2_insn (incloc, inc)
3454 : gen_sub2_insn (incloc, inc));
3456 code = recog_memoized (add_insn);
3457 if (code >= 0)
3459 if (! post && result != incloc)
3460 emit_insn (gen_move_insn (result, incloc));
3461 return result;
3463 delete_insns_since (last);
3466 /* If couldn't do the increment directly, must increment in RESULT.
3467 The way we do this depends on whether this is pre- or
3468 post-increment. For pre-increment, copy INCLOC to the reload
3469 register, increment it there, then save back. */
3470 if (! post)
3472 if (real_in != result)
3473 emit_insn (gen_move_insn (result, real_in));
3474 if (plus_p)
3475 emit_insn (gen_add2_insn (result, inc));
3476 else
3477 emit_insn (gen_sub2_insn (result, inc));
3478 if (result != incloc)
3479 emit_insn (gen_move_insn (incloc, result));
3481 else
3483 /* Post-increment.
3485 Because this might be a jump insn or a compare, and because
3486 RESULT may not be available after the insn in an input
3487 reload, we must do the incrementing before the insn being
3488 reloaded for.
3490 We have already copied IN to RESULT. Increment the copy in
3491 RESULT, save that back, then decrement RESULT so it has
3492 the original value. */
3493 if (plus_p)
3494 emit_insn (gen_add2_insn (result, inc));
3495 else
3496 emit_insn (gen_sub2_insn (result, inc));
3497 emit_insn (gen_move_insn (incloc, result));
3498 /* Restore non-modified value for the result. We prefer this
3499 way because it does not require an additional hard
3500 register. */
3501 if (plus_p)
3503 if (CONST_INT_P (inc))
3504 emit_insn (gen_add2_insn (result,
3505 gen_int_mode (-INTVAL (inc),
3506 GET_MODE (result))));
3507 else
3508 emit_insn (gen_sub2_insn (result, inc));
3510 else
3511 emit_insn (gen_add2_insn (result, inc));
3513 return result;
3516 /* Return true if the current move insn does not need processing as we
3517 already know that it satisfies its constraints. */
3518 static bool
3519 simple_move_p (void)
3521 rtx dest, src;
3522 enum reg_class dclass, sclass;
3524 lra_assert (curr_insn_set != NULL_RTX);
3525 dest = SET_DEST (curr_insn_set);
3526 src = SET_SRC (curr_insn_set);
3528 /* If the instruction has multiple sets we need to process it even if it
3529 is single_set. This can happen if one or more of the SETs are dead.
3530 See PR73650. */
3531 if (multiple_sets (curr_insn))
3532 return false;
3534 return ((dclass = get_op_class (dest)) != NO_REGS
3535 && (sclass = get_op_class (src)) != NO_REGS
3536 /* The backend guarantees that register moves of cost 2
3537 never need reloads. */
3538 && targetm.register_move_cost (GET_MODE (src), sclass, dclass) == 2);
3541 /* Swap operands NOP and NOP + 1. */
3542 static inline void
3543 swap_operands (int nop)
3545 std::swap (curr_operand_mode[nop], curr_operand_mode[nop + 1]);
3546 std::swap (original_subreg_reg_mode[nop], original_subreg_reg_mode[nop + 1]);
3547 std::swap (*curr_id->operand_loc[nop], *curr_id->operand_loc[nop + 1]);
3548 std::swap (equiv_substition_p[nop], equiv_substition_p[nop + 1]);
3549 /* Swap the duplicates too. */
3550 lra_update_dup (curr_id, nop);
3551 lra_update_dup (curr_id, nop + 1);
3554 /* Main entry point of the constraint code: search the body of the
3555 current insn to choose the best alternative. It is mimicking insn
3556 alternative cost calculation model of former reload pass. That is
3557 because machine descriptions were written to use this model. This
3558 model can be changed in future. Make commutative operand exchange
3559 if it is chosen.
3561 if CHECK_ONLY_P is false, do RTL changes to satisfy the
3562 constraints. Return true if any change happened during function
3563 call.
3565 If CHECK_ONLY_P is true then don't do any transformation. Just
3566 check that the insn satisfies all constraints. If the insn does
3567 not satisfy any constraint, return true. */
3568 static bool
3569 curr_insn_transform (bool check_only_p)
3571 int i, j, k;
3572 int n_operands;
3573 int n_alternatives;
3574 int n_outputs;
3575 int commutative;
3576 signed char goal_alt_matched[MAX_RECOG_OPERANDS][MAX_RECOG_OPERANDS];
3577 signed char match_inputs[MAX_RECOG_OPERANDS + 1];
3578 signed char outputs[MAX_RECOG_OPERANDS + 1];
3579 rtx_insn *before, *after;
3580 bool alt_p = false;
3581 /* Flag that the insn has been changed through a transformation. */
3582 bool change_p;
3583 bool sec_mem_p;
3584 #ifdef SECONDARY_MEMORY_NEEDED
3585 bool use_sec_mem_p;
3586 #endif
3587 int max_regno_before;
3588 int reused_alternative_num;
3590 curr_insn_set = single_set (curr_insn);
3591 if (curr_insn_set != NULL_RTX && simple_move_p ())
3592 return false;
3594 no_input_reloads_p = no_output_reloads_p = false;
3595 goal_alt_number = -1;
3596 change_p = sec_mem_p = false;
3597 /* JUMP_INSNs and CALL_INSNs are not allowed to have any output
3598 reloads; neither are insns that SET cc0. Insns that use CC0 are
3599 not allowed to have any input reloads. */
3600 if (JUMP_P (curr_insn) || CALL_P (curr_insn))
3601 no_output_reloads_p = true;
3603 if (HAVE_cc0 && reg_referenced_p (cc0_rtx, PATTERN (curr_insn)))
3604 no_input_reloads_p = true;
3605 if (HAVE_cc0 && reg_set_p (cc0_rtx, PATTERN (curr_insn)))
3606 no_output_reloads_p = true;
3608 n_operands = curr_static_id->n_operands;
3609 n_alternatives = curr_static_id->n_alternatives;
3611 /* Just return "no reloads" if insn has no operands with
3612 constraints. */
3613 if (n_operands == 0 || n_alternatives == 0)
3614 return false;
3616 max_regno_before = max_reg_num ();
3618 for (i = 0; i < n_operands; i++)
3620 goal_alt_matched[i][0] = -1;
3621 goal_alt_matches[i] = -1;
3624 commutative = curr_static_id->commutative;
3626 /* Now see what we need for pseudos that didn't get hard regs or got
3627 the wrong kind of hard reg. For this, we must consider all the
3628 operands together against the register constraints. */
3630 best_losers = best_overall = INT_MAX;
3631 best_reload_sum = 0;
3633 curr_swapped = false;
3634 goal_alt_swapped = false;
3636 if (! check_only_p)
3637 /* Make equivalence substitution and memory subreg elimination
3638 before address processing because an address legitimacy can
3639 depend on memory mode. */
3640 for (i = 0; i < n_operands; i++)
3642 rtx op, subst, old;
3643 bool op_change_p = false;
3645 if (curr_static_id->operand[i].is_operator)
3646 continue;
3648 old = op = *curr_id->operand_loc[i];
3649 if (GET_CODE (old) == SUBREG)
3650 old = SUBREG_REG (old);
3651 subst = get_equiv_with_elimination (old, curr_insn);
3652 original_subreg_reg_mode[i] = VOIDmode;
3653 equiv_substition_p[i] = false;
3654 if (subst != old)
3656 equiv_substition_p[i] = true;
3657 subst = copy_rtx (subst);
3658 lra_assert (REG_P (old));
3659 if (GET_CODE (op) != SUBREG)
3660 *curr_id->operand_loc[i] = subst;
3661 else
3663 SUBREG_REG (op) = subst;
3664 if (GET_MODE (subst) == VOIDmode)
3665 original_subreg_reg_mode[i] = GET_MODE (old);
3667 if (lra_dump_file != NULL)
3669 fprintf (lra_dump_file,
3670 "Changing pseudo %d in operand %i of insn %u on equiv ",
3671 REGNO (old), i, INSN_UID (curr_insn));
3672 dump_value_slim (lra_dump_file, subst, 1);
3673 fprintf (lra_dump_file, "\n");
3675 op_change_p = change_p = true;
3677 if (simplify_operand_subreg (i, GET_MODE (old)) || op_change_p)
3679 change_p = true;
3680 lra_update_dup (curr_id, i);
3684 /* Reload address registers and displacements. We do it before
3685 finding an alternative because of memory constraints. */
3686 before = after = NULL;
3687 for (i = 0; i < n_operands; i++)
3688 if (! curr_static_id->operand[i].is_operator
3689 && process_address (i, check_only_p, &before, &after))
3691 if (check_only_p)
3692 return true;
3693 change_p = true;
3694 lra_update_dup (curr_id, i);
3697 if (change_p)
3698 /* If we've changed the instruction then any alternative that
3699 we chose previously may no longer be valid. */
3700 lra_set_used_insn_alternative (curr_insn, -1);
3702 if (! check_only_p && curr_insn_set != NULL_RTX
3703 && check_and_process_move (&change_p, &sec_mem_p))
3704 return change_p;
3706 try_swapped:
3708 reused_alternative_num = check_only_p ? -1 : curr_id->used_insn_alternative;
3709 if (lra_dump_file != NULL && reused_alternative_num >= 0)
3710 fprintf (lra_dump_file, "Reusing alternative %d for insn #%u\n",
3711 reused_alternative_num, INSN_UID (curr_insn));
3713 if (process_alt_operands (reused_alternative_num))
3714 alt_p = true;
3716 if (check_only_p)
3717 return ! alt_p || best_losers != 0;
3719 /* If insn is commutative (it's safe to exchange a certain pair of
3720 operands) then we need to try each alternative twice, the second
3721 time matching those two operands as if we had exchanged them. To
3722 do this, really exchange them in operands.
3724 If we have just tried the alternatives the second time, return
3725 operands to normal and drop through. */
3727 if (reused_alternative_num < 0 && commutative >= 0)
3729 curr_swapped = !curr_swapped;
3730 if (curr_swapped)
3732 swap_operands (commutative);
3733 goto try_swapped;
3735 else
3736 swap_operands (commutative);
3739 if (! alt_p && ! sec_mem_p)
3741 /* No alternative works with reloads?? */
3742 if (INSN_CODE (curr_insn) >= 0)
3743 fatal_insn ("unable to generate reloads for:", curr_insn);
3744 error_for_asm (curr_insn,
3745 "inconsistent operand constraints in an %<asm%>");
3746 /* Avoid further trouble with this insn. */
3747 PATTERN (curr_insn) = gen_rtx_USE (VOIDmode, const0_rtx);
3748 lra_invalidate_insn_data (curr_insn);
3749 return true;
3752 /* If the best alternative is with operands 1 and 2 swapped, swap
3753 them. Update the operand numbers of any reloads already
3754 pushed. */
3756 if (goal_alt_swapped)
3758 if (lra_dump_file != NULL)
3759 fprintf (lra_dump_file, " Commutative operand exchange in insn %u\n",
3760 INSN_UID (curr_insn));
3762 /* Swap the duplicates too. */
3763 swap_operands (commutative);
3764 change_p = true;
3767 #ifdef SECONDARY_MEMORY_NEEDED
3768 /* Some target macros SECONDARY_MEMORY_NEEDED (e.g. x86) are defined
3769 too conservatively. So we use the secondary memory only if there
3770 is no any alternative without reloads. */
3771 use_sec_mem_p = false;
3772 if (! alt_p)
3773 use_sec_mem_p = true;
3774 else if (sec_mem_p)
3776 for (i = 0; i < n_operands; i++)
3777 if (! goal_alt_win[i] && ! goal_alt_match_win[i])
3778 break;
3779 use_sec_mem_p = i < n_operands;
3782 if (use_sec_mem_p)
3784 int in = -1, out = -1;
3785 rtx new_reg, src, dest, rld;
3786 machine_mode sec_mode, rld_mode;
3788 lra_assert (curr_insn_set != NULL_RTX && sec_mem_p);
3789 dest = SET_DEST (curr_insn_set);
3790 src = SET_SRC (curr_insn_set);
3791 for (i = 0; i < n_operands; i++)
3792 if (*curr_id->operand_loc[i] == dest)
3793 out = i;
3794 else if (*curr_id->operand_loc[i] == src)
3795 in = i;
3796 for (i = 0; i < curr_static_id->n_dups; i++)
3797 if (out < 0 && *curr_id->dup_loc[i] == dest)
3798 out = curr_static_id->dup_num[i];
3799 else if (in < 0 && *curr_id->dup_loc[i] == src)
3800 in = curr_static_id->dup_num[i];
3801 lra_assert (out >= 0 && in >= 0
3802 && curr_static_id->operand[out].type == OP_OUT
3803 && curr_static_id->operand[in].type == OP_IN);
3804 rld = (GET_MODE_SIZE (GET_MODE (dest)) <= GET_MODE_SIZE (GET_MODE (src))
3805 ? dest : src);
3806 rld_mode = GET_MODE (rld);
3807 #ifdef SECONDARY_MEMORY_NEEDED_MODE
3808 sec_mode = SECONDARY_MEMORY_NEEDED_MODE (rld_mode);
3809 #else
3810 sec_mode = rld_mode;
3811 #endif
3812 new_reg = lra_create_new_reg (sec_mode, NULL_RTX,
3813 NO_REGS, "secondary");
3814 /* If the mode is changed, it should be wider. */
3815 lra_assert (GET_MODE_SIZE (sec_mode) >= GET_MODE_SIZE (rld_mode));
3816 if (sec_mode != rld_mode)
3818 /* If the target says specifically to use another mode for
3819 secondary memory moves we can not reuse the original
3820 insn. */
3821 after = emit_spill_move (false, new_reg, dest);
3822 lra_process_new_insns (curr_insn, NULL, after,
3823 "Inserting the sec. move");
3824 /* We may have non null BEFORE here (e.g. after address
3825 processing. */
3826 push_to_sequence (before);
3827 before = emit_spill_move (true, new_reg, src);
3828 emit_insn (before);
3829 before = get_insns ();
3830 end_sequence ();
3831 lra_process_new_insns (curr_insn, before, NULL, "Changing on");
3832 lra_set_insn_deleted (curr_insn);
3834 else if (dest == rld)
3836 *curr_id->operand_loc[out] = new_reg;
3837 lra_update_dup (curr_id, out);
3838 after = emit_spill_move (false, new_reg, dest);
3839 lra_process_new_insns (curr_insn, NULL, after,
3840 "Inserting the sec. move");
3842 else
3844 *curr_id->operand_loc[in] = new_reg;
3845 lra_update_dup (curr_id, in);
3846 /* See comments above. */
3847 push_to_sequence (before);
3848 before = emit_spill_move (true, new_reg, src);
3849 emit_insn (before);
3850 before = get_insns ();
3851 end_sequence ();
3852 lra_process_new_insns (curr_insn, before, NULL,
3853 "Inserting the sec. move");
3855 lra_update_insn_regno_info (curr_insn);
3856 return true;
3858 #endif
3860 lra_assert (goal_alt_number >= 0);
3861 lra_set_used_insn_alternative (curr_insn, goal_alt_number);
3863 if (lra_dump_file != NULL)
3865 const char *p;
3867 fprintf (lra_dump_file, " Choosing alt %d in insn %u:",
3868 goal_alt_number, INSN_UID (curr_insn));
3869 for (i = 0; i < n_operands; i++)
3871 p = (curr_static_id->operand_alternative
3872 [goal_alt_number * n_operands + i].constraint);
3873 if (*p == '\0')
3874 continue;
3875 fprintf (lra_dump_file, " (%d) ", i);
3876 for (; *p != '\0' && *p != ',' && *p != '#'; p++)
3877 fputc (*p, lra_dump_file);
3879 if (INSN_CODE (curr_insn) >= 0
3880 && (p = get_insn_name (INSN_CODE (curr_insn))) != NULL)
3881 fprintf (lra_dump_file, " {%s}", p);
3882 if (curr_id->sp_offset != 0)
3883 fprintf (lra_dump_file, " (sp_off=%" HOST_WIDE_INT_PRINT "d)",
3884 curr_id->sp_offset);
3885 fprintf (lra_dump_file, "\n");
3888 /* Right now, for any pair of operands I and J that are required to
3889 match, with J < I, goal_alt_matches[I] is J. Add I to
3890 goal_alt_matched[J]. */
3892 for (i = 0; i < n_operands; i++)
3893 if ((j = goal_alt_matches[i]) >= 0)
3895 for (k = 0; goal_alt_matched[j][k] >= 0; k++)
3897 /* We allow matching one output operand and several input
3898 operands. */
3899 lra_assert (k == 0
3900 || (curr_static_id->operand[j].type == OP_OUT
3901 && curr_static_id->operand[i].type == OP_IN
3902 && (curr_static_id->operand
3903 [goal_alt_matched[j][0]].type == OP_IN)));
3904 goal_alt_matched[j][k] = i;
3905 goal_alt_matched[j][k + 1] = -1;
3908 for (i = 0; i < n_operands; i++)
3909 goal_alt_win[i] |= goal_alt_match_win[i];
3911 /* Any constants that aren't allowed and can't be reloaded into
3912 registers are here changed into memory references. */
3913 for (i = 0; i < n_operands; i++)
3914 if (goal_alt_win[i])
3916 int regno;
3917 enum reg_class new_class;
3918 rtx reg = *curr_id->operand_loc[i];
3920 if (GET_CODE (reg) == SUBREG)
3921 reg = SUBREG_REG (reg);
3923 if (REG_P (reg) && (regno = REGNO (reg)) >= FIRST_PSEUDO_REGISTER)
3925 bool ok_p = in_class_p (reg, goal_alt[i], &new_class);
3927 if (new_class != NO_REGS && get_reg_class (regno) != new_class)
3929 lra_assert (ok_p);
3930 lra_change_class (regno, new_class, " Change to", true);
3934 else
3936 const char *constraint;
3937 char c;
3938 rtx op = *curr_id->operand_loc[i];
3939 rtx subreg = NULL_RTX;
3940 machine_mode mode = curr_operand_mode[i];
3942 if (GET_CODE (op) == SUBREG)
3944 subreg = op;
3945 op = SUBREG_REG (op);
3946 mode = GET_MODE (op);
3949 if (CONST_POOL_OK_P (mode, op)
3950 && ((targetm.preferred_reload_class
3951 (op, (enum reg_class) goal_alt[i]) == NO_REGS)
3952 || no_input_reloads_p))
3954 rtx tem = force_const_mem (mode, op);
3956 change_p = true;
3957 if (subreg != NULL_RTX)
3958 tem = gen_rtx_SUBREG (mode, tem, SUBREG_BYTE (subreg));
3960 *curr_id->operand_loc[i] = tem;
3961 lra_update_dup (curr_id, i);
3962 process_address (i, false, &before, &after);
3964 /* If the alternative accepts constant pool refs directly
3965 there will be no reload needed at all. */
3966 if (subreg != NULL_RTX)
3967 continue;
3968 /* Skip alternatives before the one requested. */
3969 constraint = (curr_static_id->operand_alternative
3970 [goal_alt_number * n_operands + i].constraint);
3971 for (;
3972 (c = *constraint) && c != ',' && c != '#';
3973 constraint += CONSTRAINT_LEN (c, constraint))
3975 enum constraint_num cn = lookup_constraint (constraint);
3976 if ((insn_extra_memory_constraint (cn)
3977 || insn_extra_special_memory_constraint (cn))
3978 && satisfies_memory_constraint_p (tem, cn))
3979 break;
3981 if (c == '\0' || c == ',' || c == '#')
3982 continue;
3984 goal_alt_win[i] = true;
3988 n_outputs = 0;
3989 outputs[0] = -1;
3990 for (i = 0; i < n_operands; i++)
3992 int regno;
3993 bool optional_p = false;
3994 rtx old, new_reg;
3995 rtx op = *curr_id->operand_loc[i];
3997 if (goal_alt_win[i])
3999 if (goal_alt[i] == NO_REGS
4000 && REG_P (op)
4001 /* When we assign NO_REGS it means that we will not
4002 assign a hard register to the scratch pseudo by
4003 assigment pass and the scratch pseudo will be
4004 spilled. Spilled scratch pseudos are transformed
4005 back to scratches at the LRA end. */
4006 && lra_former_scratch_operand_p (curr_insn, i)
4007 && lra_former_scratch_p (REGNO (op)))
4009 int regno = REGNO (op);
4010 lra_change_class (regno, NO_REGS, " Change to", true);
4011 if (lra_get_regno_hard_regno (regno) >= 0)
4012 /* We don't have to mark all insn affected by the
4013 spilled pseudo as there is only one such insn, the
4014 current one. */
4015 reg_renumber[regno] = -1;
4016 lra_assert (bitmap_single_bit_set_p
4017 (&lra_reg_info[REGNO (op)].insn_bitmap));
4019 /* We can do an optional reload. If the pseudo got a hard
4020 reg, we might improve the code through inheritance. If
4021 it does not get a hard register we coalesce memory/memory
4022 moves later. Ignore move insns to avoid cycling. */
4023 if (! lra_simple_p
4024 && lra_undo_inheritance_iter < LRA_MAX_INHERITANCE_PASSES
4025 && goal_alt[i] != NO_REGS && REG_P (op)
4026 && (regno = REGNO (op)) >= FIRST_PSEUDO_REGISTER
4027 && regno < new_regno_start
4028 && ! lra_former_scratch_p (regno)
4029 && reg_renumber[regno] < 0
4030 /* Check that the optional reload pseudo will be able to
4031 hold given mode value. */
4032 && ! (prohibited_class_reg_set_mode_p
4033 (goal_alt[i], reg_class_contents[goal_alt[i]],
4034 PSEUDO_REGNO_MODE (regno)))
4035 && (curr_insn_set == NULL_RTX
4036 || !((REG_P (SET_SRC (curr_insn_set))
4037 || MEM_P (SET_SRC (curr_insn_set))
4038 || GET_CODE (SET_SRC (curr_insn_set)) == SUBREG)
4039 && (REG_P (SET_DEST (curr_insn_set))
4040 || MEM_P (SET_DEST (curr_insn_set))
4041 || GET_CODE (SET_DEST (curr_insn_set)) == SUBREG))))
4042 optional_p = true;
4043 else
4044 continue;
4047 /* Operands that match previous ones have already been handled. */
4048 if (goal_alt_matches[i] >= 0)
4049 continue;
4051 /* We should not have an operand with a non-offsettable address
4052 appearing where an offsettable address will do. It also may
4053 be a case when the address should be special in other words
4054 not a general one (e.g. it needs no index reg). */
4055 if (goal_alt_matched[i][0] == -1 && goal_alt_offmemok[i] && MEM_P (op))
4057 enum reg_class rclass;
4058 rtx *loc = &XEXP (op, 0);
4059 enum rtx_code code = GET_CODE (*loc);
4061 push_to_sequence (before);
4062 rclass = base_reg_class (GET_MODE (op), MEM_ADDR_SPACE (op),
4063 MEM, SCRATCH);
4064 if (GET_RTX_CLASS (code) == RTX_AUTOINC)
4065 new_reg = emit_inc (rclass, *loc, *loc,
4066 /* This value does not matter for MODIFY. */
4067 GET_MODE_SIZE (GET_MODE (op)));
4068 else if (get_reload_reg (OP_IN, Pmode, *loc, rclass, FALSE,
4069 "offsetable address", &new_reg))
4070 lra_emit_move (new_reg, *loc);
4071 before = get_insns ();
4072 end_sequence ();
4073 *loc = new_reg;
4074 lra_update_dup (curr_id, i);
4076 else if (goal_alt_matched[i][0] == -1)
4078 machine_mode mode;
4079 rtx reg, *loc;
4080 int hard_regno, byte;
4081 enum op_type type = curr_static_id->operand[i].type;
4083 loc = curr_id->operand_loc[i];
4084 mode = curr_operand_mode[i];
4085 if (GET_CODE (*loc) == SUBREG)
4087 reg = SUBREG_REG (*loc);
4088 byte = SUBREG_BYTE (*loc);
4089 if (REG_P (reg)
4090 /* Strict_low_part requires reload the register not
4091 the sub-register. */
4092 && (curr_static_id->operand[i].strict_low
4093 || (GET_MODE_SIZE (mode)
4094 <= GET_MODE_SIZE (GET_MODE (reg))
4095 && (hard_regno
4096 = get_try_hard_regno (REGNO (reg))) >= 0
4097 && (simplify_subreg_regno
4098 (hard_regno,
4099 GET_MODE (reg), byte, mode) < 0)
4100 && (goal_alt[i] == NO_REGS
4101 || (simplify_subreg_regno
4102 (ira_class_hard_regs[goal_alt[i]][0],
4103 GET_MODE (reg), byte, mode) >= 0)))))
4105 if (type == OP_OUT)
4106 type = OP_INOUT;
4107 loc = &SUBREG_REG (*loc);
4108 mode = GET_MODE (*loc);
4111 old = *loc;
4112 if (get_reload_reg (type, mode, old, goal_alt[i],
4113 loc != curr_id->operand_loc[i], "", &new_reg)
4114 && type != OP_OUT)
4116 push_to_sequence (before);
4117 lra_emit_move (new_reg, old);
4118 before = get_insns ();
4119 end_sequence ();
4121 *loc = new_reg;
4122 if (type != OP_IN
4123 && find_reg_note (curr_insn, REG_UNUSED, old) == NULL_RTX)
4125 start_sequence ();
4126 lra_emit_move (type == OP_INOUT ? copy_rtx (old) : old, new_reg);
4127 emit_insn (after);
4128 after = get_insns ();
4129 end_sequence ();
4130 *loc = new_reg;
4132 for (j = 0; j < goal_alt_dont_inherit_ops_num; j++)
4133 if (goal_alt_dont_inherit_ops[j] == i)
4135 lra_set_regno_unique_value (REGNO (new_reg));
4136 break;
4138 lra_update_dup (curr_id, i);
4140 else if (curr_static_id->operand[i].type == OP_IN
4141 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4142 == OP_OUT))
4144 /* generate reloads for input and matched outputs. */
4145 match_inputs[0] = i;
4146 match_inputs[1] = -1;
4147 match_reload (goal_alt_matched[i][0], match_inputs, outputs,
4148 goal_alt[i], &before, &after,
4149 curr_static_id->operand_alternative
4150 [goal_alt_number * n_operands + goal_alt_matched[i][0]]
4151 .earlyclobber);
4153 else if (curr_static_id->operand[i].type == OP_OUT
4154 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4155 == OP_IN))
4156 /* Generate reloads for output and matched inputs. */
4157 match_reload (i, goal_alt_matched[i], outputs, goal_alt[i], &before,
4158 &after, curr_static_id->operand_alternative
4159 [goal_alt_number * n_operands + i].earlyclobber);
4160 else if (curr_static_id->operand[i].type == OP_IN
4161 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4162 == OP_IN))
4164 /* Generate reloads for matched inputs. */
4165 match_inputs[0] = i;
4166 for (j = 0; (k = goal_alt_matched[i][j]) >= 0; j++)
4167 match_inputs[j + 1] = k;
4168 match_inputs[j + 1] = -1;
4169 match_reload (-1, match_inputs, outputs, goal_alt[i], &before,
4170 &after, false);
4172 else
4173 /* We must generate code in any case when function
4174 process_alt_operands decides that it is possible. */
4175 gcc_unreachable ();
4177 /* Memorise processed outputs so that output remaining to be processed
4178 can avoid using the same register value (see match_reload). */
4179 if (curr_static_id->operand[i].type == OP_OUT)
4181 outputs[n_outputs++] = i;
4182 outputs[n_outputs] = -1;
4185 if (optional_p)
4187 rtx reg = op;
4189 lra_assert (REG_P (reg));
4190 regno = REGNO (reg);
4191 op = *curr_id->operand_loc[i]; /* Substitution. */
4192 if (GET_CODE (op) == SUBREG)
4193 op = SUBREG_REG (op);
4194 gcc_assert (REG_P (op) && (int) REGNO (op) >= new_regno_start);
4195 bitmap_set_bit (&lra_optional_reload_pseudos, REGNO (op));
4196 lra_reg_info[REGNO (op)].restore_rtx = reg;
4197 if (lra_dump_file != NULL)
4198 fprintf (lra_dump_file,
4199 " Making reload reg %d for reg %d optional\n",
4200 REGNO (op), regno);
4203 if (before != NULL_RTX || after != NULL_RTX
4204 || max_regno_before != max_reg_num ())
4205 change_p = true;
4206 if (change_p)
4208 lra_update_operator_dups (curr_id);
4209 /* Something changes -- process the insn. */
4210 lra_update_insn_regno_info (curr_insn);
4212 lra_process_new_insns (curr_insn, before, after, "Inserting insn reload");
4213 return change_p;
4216 /* Return true if INSN satisfies all constraints. In other words, no
4217 reload insns are needed. */
4218 bool
4219 lra_constrain_insn (rtx_insn *insn)
4221 int saved_new_regno_start = new_regno_start;
4222 int saved_new_insn_uid_start = new_insn_uid_start;
4223 bool change_p;
4225 curr_insn = insn;
4226 curr_id = lra_get_insn_recog_data (curr_insn);
4227 curr_static_id = curr_id->insn_static_data;
4228 new_insn_uid_start = get_max_uid ();
4229 new_regno_start = max_reg_num ();
4230 change_p = curr_insn_transform (true);
4231 new_regno_start = saved_new_regno_start;
4232 new_insn_uid_start = saved_new_insn_uid_start;
4233 return ! change_p;
4236 /* Return true if X is in LIST. */
4237 static bool
4238 in_list_p (rtx x, rtx list)
4240 for (; list != NULL_RTX; list = XEXP (list, 1))
4241 if (XEXP (list, 0) == x)
4242 return true;
4243 return false;
4246 /* Return true if X contains an allocatable hard register (if
4247 HARD_REG_P) or a (spilled if SPILLED_P) pseudo. */
4248 static bool
4249 contains_reg_p (rtx x, bool hard_reg_p, bool spilled_p)
4251 int i, j;
4252 const char *fmt;
4253 enum rtx_code code;
4255 code = GET_CODE (x);
4256 if (REG_P (x))
4258 int regno = REGNO (x);
4259 HARD_REG_SET alloc_regs;
4261 if (hard_reg_p)
4263 if (regno >= FIRST_PSEUDO_REGISTER)
4264 regno = lra_get_regno_hard_regno (regno);
4265 if (regno < 0)
4266 return false;
4267 COMPL_HARD_REG_SET (alloc_regs, lra_no_alloc_regs);
4268 return overlaps_hard_reg_set_p (alloc_regs, GET_MODE (x), regno);
4270 else
4272 if (regno < FIRST_PSEUDO_REGISTER)
4273 return false;
4274 if (! spilled_p)
4275 return true;
4276 return lra_get_regno_hard_regno (regno) < 0;
4279 fmt = GET_RTX_FORMAT (code);
4280 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4282 if (fmt[i] == 'e')
4284 if (contains_reg_p (XEXP (x, i), hard_reg_p, spilled_p))
4285 return true;
4287 else if (fmt[i] == 'E')
4289 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4290 if (contains_reg_p (XVECEXP (x, i, j), hard_reg_p, spilled_p))
4291 return true;
4294 return false;
4297 /* Process all regs in location *LOC and change them on equivalent
4298 substitution. Return true if any change was done. */
4299 static bool
4300 loc_equivalence_change_p (rtx *loc)
4302 rtx subst, reg, x = *loc;
4303 bool result = false;
4304 enum rtx_code code = GET_CODE (x);
4305 const char *fmt;
4306 int i, j;
4308 if (code == SUBREG)
4310 reg = SUBREG_REG (x);
4311 if ((subst = get_equiv_with_elimination (reg, curr_insn)) != reg
4312 && GET_MODE (subst) == VOIDmode)
4314 /* We cannot reload debug location. Simplify subreg here
4315 while we know the inner mode. */
4316 *loc = simplify_gen_subreg (GET_MODE (x), subst,
4317 GET_MODE (reg), SUBREG_BYTE (x));
4318 return true;
4321 if (code == REG && (subst = get_equiv_with_elimination (x, curr_insn)) != x)
4323 *loc = subst;
4324 return true;
4327 /* Scan all the operand sub-expressions. */
4328 fmt = GET_RTX_FORMAT (code);
4329 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4331 if (fmt[i] == 'e')
4332 result = loc_equivalence_change_p (&XEXP (x, i)) || result;
4333 else if (fmt[i] == 'E')
4334 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4335 result
4336 = loc_equivalence_change_p (&XVECEXP (x, i, j)) || result;
4338 return result;
4341 /* Similar to loc_equivalence_change_p, but for use as
4342 simplify_replace_fn_rtx callback. DATA is insn for which the
4343 elimination is done. If it null we don't do the elimination. */
4344 static rtx
4345 loc_equivalence_callback (rtx loc, const_rtx, void *data)
4347 if (!REG_P (loc))
4348 return NULL_RTX;
4350 rtx subst = (data == NULL
4351 ? get_equiv (loc) : get_equiv_with_elimination (loc, (rtx_insn *) data));
4352 if (subst != loc)
4353 return subst;
4355 return NULL_RTX;
4358 /* Maximum number of generated reload insns per an insn. It is for
4359 preventing this pass cycling in a bug case. */
4360 #define MAX_RELOAD_INSNS_NUMBER LRA_MAX_INSN_RELOADS
4362 /* The current iteration number of this LRA pass. */
4363 int lra_constraint_iter;
4365 /* True if we substituted equiv which needs checking register
4366 allocation correctness because the equivalent value contains
4367 allocatable hard registers or when we restore multi-register
4368 pseudo. */
4369 bool lra_risky_transformations_p;
4371 /* Return true if REGNO is referenced in more than one block. */
4372 static bool
4373 multi_block_pseudo_p (int regno)
4375 basic_block bb = NULL;
4376 unsigned int uid;
4377 bitmap_iterator bi;
4379 if (regno < FIRST_PSEUDO_REGISTER)
4380 return false;
4382 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi)
4383 if (bb == NULL)
4384 bb = BLOCK_FOR_INSN (lra_insn_recog_data[uid]->insn);
4385 else if (BLOCK_FOR_INSN (lra_insn_recog_data[uid]->insn) != bb)
4386 return true;
4387 return false;
4390 /* Return true if LIST contains a deleted insn. */
4391 static bool
4392 contains_deleted_insn_p (rtx_insn_list *list)
4394 for (; list != NULL_RTX; list = list->next ())
4395 if (NOTE_P (list->insn ())
4396 && NOTE_KIND (list->insn ()) == NOTE_INSN_DELETED)
4397 return true;
4398 return false;
4401 /* Return true if X contains a pseudo dying in INSN. */
4402 static bool
4403 dead_pseudo_p (rtx x, rtx_insn *insn)
4405 int i, j;
4406 const char *fmt;
4407 enum rtx_code code;
4409 if (REG_P (x))
4410 return (insn != NULL_RTX
4411 && find_regno_note (insn, REG_DEAD, REGNO (x)) != NULL_RTX);
4412 code = GET_CODE (x);
4413 fmt = GET_RTX_FORMAT (code);
4414 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4416 if (fmt[i] == 'e')
4418 if (dead_pseudo_p (XEXP (x, i), insn))
4419 return true;
4421 else if (fmt[i] == 'E')
4423 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4424 if (dead_pseudo_p (XVECEXP (x, i, j), insn))
4425 return true;
4428 return false;
4431 /* Return true if INSN contains a dying pseudo in INSN right hand
4432 side. */
4433 static bool
4434 insn_rhs_dead_pseudo_p (rtx_insn *insn)
4436 rtx set = single_set (insn);
4438 gcc_assert (set != NULL);
4439 return dead_pseudo_p (SET_SRC (set), insn);
4442 /* Return true if any init insn of REGNO contains a dying pseudo in
4443 insn right hand side. */
4444 static bool
4445 init_insn_rhs_dead_pseudo_p (int regno)
4447 rtx_insn_list *insns = ira_reg_equiv[regno].init_insns;
4449 if (insns == NULL)
4450 return false;
4451 for (; insns != NULL_RTX; insns = insns->next ())
4452 if (insn_rhs_dead_pseudo_p (insns->insn ()))
4453 return true;
4454 return false;
4457 /* Return TRUE if REGNO has a reverse equivalence. The equivalence is
4458 reverse only if we have one init insn with given REGNO as a
4459 source. */
4460 static bool
4461 reverse_equiv_p (int regno)
4463 rtx_insn_list *insns = ira_reg_equiv[regno].init_insns;
4464 rtx set;
4466 if (insns == NULL)
4467 return false;
4468 if (! INSN_P (insns->insn ())
4469 || insns->next () != NULL)
4470 return false;
4471 if ((set = single_set (insns->insn ())) == NULL_RTX)
4472 return false;
4473 return REG_P (SET_SRC (set)) && (int) REGNO (SET_SRC (set)) == regno;
4476 /* Return TRUE if REGNO was reloaded in an equivalence init insn. We
4477 call this function only for non-reverse equivalence. */
4478 static bool
4479 contains_reloaded_insn_p (int regno)
4481 rtx set;
4482 rtx_insn_list *list = ira_reg_equiv[regno].init_insns;
4484 for (; list != NULL; list = list->next ())
4485 if ((set = single_set (list->insn ())) == NULL_RTX
4486 || ! REG_P (SET_DEST (set))
4487 || (int) REGNO (SET_DEST (set)) != regno)
4488 return true;
4489 return false;
4492 /* Entry function of LRA constraint pass. Return true if the
4493 constraint pass did change the code. */
4494 bool
4495 lra_constraints (bool first_p)
4497 bool changed_p;
4498 int i, hard_regno, new_insns_num;
4499 unsigned int min_len, new_min_len, uid;
4500 rtx set, x, reg, dest_reg;
4501 basic_block last_bb;
4502 bitmap_head equiv_insn_bitmap;
4503 bitmap_iterator bi;
4505 lra_constraint_iter++;
4506 if (lra_dump_file != NULL)
4507 fprintf (lra_dump_file, "\n********** Local #%d: **********\n\n",
4508 lra_constraint_iter);
4509 changed_p = false;
4510 if (pic_offset_table_rtx
4511 && REGNO (pic_offset_table_rtx) >= FIRST_PSEUDO_REGISTER)
4512 lra_risky_transformations_p = true;
4513 else
4514 /* On the first iteration we should check IRA assignment
4515 correctness. In rare cases, the assignments can be wrong as
4516 early clobbers operands are ignored in IRA. */
4517 lra_risky_transformations_p = first_p;
4518 new_insn_uid_start = get_max_uid ();
4519 new_regno_start = first_p ? lra_constraint_new_regno_start : max_reg_num ();
4520 /* Mark used hard regs for target stack size calulations. */
4521 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4522 if (lra_reg_info[i].nrefs != 0
4523 && (hard_regno = lra_get_regno_hard_regno (i)) >= 0)
4525 int j, nregs;
4527 nregs = hard_regno_nregs[hard_regno][lra_reg_info[i].biggest_mode];
4528 for (j = 0; j < nregs; j++)
4529 df_set_regs_ever_live (hard_regno + j, true);
4531 /* Do elimination before the equivalence processing as we can spill
4532 some pseudos during elimination. */
4533 lra_eliminate (false, first_p);
4534 bitmap_initialize (&equiv_insn_bitmap, &reg_obstack);
4535 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4536 if (lra_reg_info[i].nrefs != 0)
4538 ira_reg_equiv[i].profitable_p = true;
4539 reg = regno_reg_rtx[i];
4540 if (lra_get_regno_hard_regno (i) < 0 && (x = get_equiv (reg)) != reg)
4542 bool pseudo_p = contains_reg_p (x, false, false);
4544 /* After RTL transformation, we can not guarantee that
4545 pseudo in the substitution was not reloaded which might
4546 make equivalence invalid. For example, in reverse
4547 equiv of p0
4549 p0 <- ...
4551 equiv_mem <- p0
4553 the memory address register was reloaded before the 2nd
4554 insn. */
4555 if ((! first_p && pseudo_p)
4556 /* We don't use DF for compilation speed sake. So it
4557 is problematic to update live info when we use an
4558 equivalence containing pseudos in more than one
4559 BB. */
4560 || (pseudo_p && multi_block_pseudo_p (i))
4561 /* If an init insn was deleted for some reason, cancel
4562 the equiv. We could update the equiv insns after
4563 transformations including an equiv insn deletion
4564 but it is not worthy as such cases are extremely
4565 rare. */
4566 || contains_deleted_insn_p (ira_reg_equiv[i].init_insns)
4567 /* If it is not a reverse equivalence, we check that a
4568 pseudo in rhs of the init insn is not dying in the
4569 insn. Otherwise, the live info at the beginning of
4570 the corresponding BB might be wrong after we
4571 removed the insn. When the equiv can be a
4572 constant, the right hand side of the init insn can
4573 be a pseudo. */
4574 || (! reverse_equiv_p (i)
4575 && (init_insn_rhs_dead_pseudo_p (i)
4576 /* If we reloaded the pseudo in an equivalence
4577 init insn, we can not remove the equiv init
4578 insns and the init insns might write into
4579 const memory in this case. */
4580 || contains_reloaded_insn_p (i)))
4581 /* Prevent access beyond equivalent memory for
4582 paradoxical subregs. */
4583 || (MEM_P (x)
4584 && (GET_MODE_SIZE (lra_reg_info[i].biggest_mode)
4585 > GET_MODE_SIZE (GET_MODE (x))))
4586 || (pic_offset_table_rtx
4587 && ((CONST_POOL_OK_P (PSEUDO_REGNO_MODE (i), x)
4588 && (targetm.preferred_reload_class
4589 (x, lra_get_allocno_class (i)) == NO_REGS))
4590 || contains_symbol_ref_p (x))))
4591 ira_reg_equiv[i].defined_p = false;
4592 if (contains_reg_p (x, false, true))
4593 ira_reg_equiv[i].profitable_p = false;
4594 if (get_equiv (reg) != reg)
4595 bitmap_ior_into (&equiv_insn_bitmap, &lra_reg_info[i].insn_bitmap);
4598 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4599 update_equiv (i);
4600 /* We should add all insns containing pseudos which should be
4601 substituted by their equivalences. */
4602 EXECUTE_IF_SET_IN_BITMAP (&equiv_insn_bitmap, 0, uid, bi)
4603 lra_push_insn_by_uid (uid);
4604 min_len = lra_insn_stack_length ();
4605 new_insns_num = 0;
4606 last_bb = NULL;
4607 changed_p = false;
4608 while ((new_min_len = lra_insn_stack_length ()) != 0)
4610 curr_insn = lra_pop_insn ();
4611 --new_min_len;
4612 curr_bb = BLOCK_FOR_INSN (curr_insn);
4613 if (curr_bb != last_bb)
4615 last_bb = curr_bb;
4616 bb_reload_num = lra_curr_reload_num;
4618 if (min_len > new_min_len)
4620 min_len = new_min_len;
4621 new_insns_num = 0;
4623 if (new_insns_num > MAX_RELOAD_INSNS_NUMBER)
4624 internal_error
4625 ("Max. number of generated reload insns per insn is achieved (%d)\n",
4626 MAX_RELOAD_INSNS_NUMBER);
4627 new_insns_num++;
4628 if (DEBUG_INSN_P (curr_insn))
4630 /* We need to check equivalence in debug insn and change
4631 pseudo to the equivalent value if necessary. */
4632 curr_id = lra_get_insn_recog_data (curr_insn);
4633 if (bitmap_bit_p (&equiv_insn_bitmap, INSN_UID (curr_insn)))
4635 rtx old = *curr_id->operand_loc[0];
4636 *curr_id->operand_loc[0]
4637 = simplify_replace_fn_rtx (old, NULL_RTX,
4638 loc_equivalence_callback, curr_insn);
4639 if (old != *curr_id->operand_loc[0])
4641 lra_update_insn_regno_info (curr_insn);
4642 changed_p = true;
4646 else if (INSN_P (curr_insn))
4648 if ((set = single_set (curr_insn)) != NULL_RTX)
4650 dest_reg = SET_DEST (set);
4651 /* The equivalence pseudo could be set up as SUBREG in a
4652 case when it is a call restore insn in a mode
4653 different from the pseudo mode. */
4654 if (GET_CODE (dest_reg) == SUBREG)
4655 dest_reg = SUBREG_REG (dest_reg);
4656 if ((REG_P (dest_reg)
4657 && (x = get_equiv (dest_reg)) != dest_reg
4658 /* Remove insns which set up a pseudo whose value
4659 can not be changed. Such insns might be not in
4660 init_insns because we don't update equiv data
4661 during insn transformations.
4663 As an example, let suppose that a pseudo got
4664 hard register and on the 1st pass was not
4665 changed to equivalent constant. We generate an
4666 additional insn setting up the pseudo because of
4667 secondary memory movement. Then the pseudo is
4668 spilled and we use the equiv constant. In this
4669 case we should remove the additional insn and
4670 this insn is not init_insns list. */
4671 && (! MEM_P (x) || MEM_READONLY_P (x)
4672 /* Check that this is actually an insn setting
4673 up the equivalence. */
4674 || in_list_p (curr_insn,
4675 ira_reg_equiv
4676 [REGNO (dest_reg)].init_insns)))
4677 || (((x = get_equiv (SET_SRC (set))) != SET_SRC (set))
4678 && in_list_p (curr_insn,
4679 ira_reg_equiv
4680 [REGNO (SET_SRC (set))].init_insns)))
4682 /* This is equiv init insn of pseudo which did not get a
4683 hard register -- remove the insn. */
4684 if (lra_dump_file != NULL)
4686 fprintf (lra_dump_file,
4687 " Removing equiv init insn %i (freq=%d)\n",
4688 INSN_UID (curr_insn),
4689 REG_FREQ_FROM_BB (BLOCK_FOR_INSN (curr_insn)));
4690 dump_insn_slim (lra_dump_file, curr_insn);
4692 if (contains_reg_p (x, true, false))
4693 lra_risky_transformations_p = true;
4694 lra_set_insn_deleted (curr_insn);
4695 continue;
4698 curr_id = lra_get_insn_recog_data (curr_insn);
4699 curr_static_id = curr_id->insn_static_data;
4700 init_curr_insn_input_reloads ();
4701 init_curr_operand_mode ();
4702 if (curr_insn_transform (false))
4703 changed_p = true;
4704 /* Check non-transformed insns too for equiv change as USE
4705 or CLOBBER don't need reloads but can contain pseudos
4706 being changed on their equivalences. */
4707 else if (bitmap_bit_p (&equiv_insn_bitmap, INSN_UID (curr_insn))
4708 && loc_equivalence_change_p (&PATTERN (curr_insn)))
4710 lra_update_insn_regno_info (curr_insn);
4711 changed_p = true;
4715 bitmap_clear (&equiv_insn_bitmap);
4716 /* If we used a new hard regno, changed_p should be true because the
4717 hard reg is assigned to a new pseudo. */
4718 if (flag_checking && !changed_p)
4720 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4721 if (lra_reg_info[i].nrefs != 0
4722 && (hard_regno = lra_get_regno_hard_regno (i)) >= 0)
4724 int j, nregs = hard_regno_nregs[hard_regno][PSEUDO_REGNO_MODE (i)];
4726 for (j = 0; j < nregs; j++)
4727 lra_assert (df_regs_ever_live_p (hard_regno + j));
4730 return changed_p;
4733 static void initiate_invariants (void);
4734 static void finish_invariants (void);
4736 /* Initiate the LRA constraint pass. It is done once per
4737 function. */
4738 void
4739 lra_constraints_init (void)
4741 initiate_invariants ();
4744 /* Finalize the LRA constraint pass. It is done once per
4745 function. */
4746 void
4747 lra_constraints_finish (void)
4749 finish_invariants ();
4754 /* Structure describes invariants for ineheritance. */
4755 struct lra_invariant
4757 /* The order number of the invariant. */
4758 int num;
4759 /* The invariant RTX. */
4760 rtx invariant_rtx;
4761 /* The origin insn of the invariant. */
4762 rtx_insn *insn;
4765 typedef lra_invariant invariant_t;
4766 typedef invariant_t *invariant_ptr_t;
4767 typedef const invariant_t *const_invariant_ptr_t;
4769 /* Pointer to the inheritance invariants. */
4770 static vec<invariant_ptr_t> invariants;
4772 /* Allocation pool for the invariants. */
4773 static object_allocator<lra_invariant> *invariants_pool;
4775 /* Hash table for the invariants. */
4776 static htab_t invariant_table;
4778 /* Hash function for INVARIANT. */
4779 static hashval_t
4780 invariant_hash (const void *invariant)
4782 rtx inv = ((const_invariant_ptr_t) invariant)->invariant_rtx;
4783 return lra_rtx_hash (inv);
4786 /* Equal function for invariants INVARIANT1 and INVARIANT2. */
4787 static int
4788 invariant_eq_p (const void *invariant1, const void *invariant2)
4790 rtx inv1 = ((const_invariant_ptr_t) invariant1)->invariant_rtx;
4791 rtx inv2 = ((const_invariant_ptr_t) invariant2)->invariant_rtx;
4793 return rtx_equal_p (inv1, inv2);
4796 /* Insert INVARIANT_RTX into the table if it is not there yet. Return
4797 invariant which is in the table. */
4798 static invariant_ptr_t
4799 insert_invariant (rtx invariant_rtx)
4801 void **entry_ptr;
4802 invariant_t invariant;
4803 invariant_ptr_t invariant_ptr;
4805 invariant.invariant_rtx = invariant_rtx;
4806 entry_ptr = htab_find_slot (invariant_table, &invariant, INSERT);
4807 if (*entry_ptr == NULL)
4809 invariant_ptr = invariants_pool->allocate ();
4810 invariant_ptr->invariant_rtx = invariant_rtx;
4811 invariant_ptr->insn = NULL;
4812 invariants.safe_push (invariant_ptr);
4813 *entry_ptr = (void *) invariant_ptr;
4815 return (invariant_ptr_t) *entry_ptr;
4818 /* Initiate the invariant table. */
4819 static void
4820 initiate_invariants (void)
4822 invariants.create (100);
4823 invariants_pool
4824 = new object_allocator<lra_invariant> ("Inheritance invariants");
4825 invariant_table = htab_create (100, invariant_hash, invariant_eq_p, NULL);
4828 /* Finish the invariant table. */
4829 static void
4830 finish_invariants (void)
4832 htab_delete (invariant_table);
4833 delete invariants_pool;
4834 invariants.release ();
4837 /* Make the invariant table empty. */
4838 static void
4839 clear_invariants (void)
4841 htab_empty (invariant_table);
4842 invariants_pool->release ();
4843 invariants.truncate (0);
4848 /* This page contains code to do inheritance/split
4849 transformations. */
4851 /* Number of reloads passed so far in current EBB. */
4852 static int reloads_num;
4854 /* Number of calls passed so far in current EBB. */
4855 static int calls_num;
4857 /* Current reload pseudo check for validity of elements in
4858 USAGE_INSNS. */
4859 static int curr_usage_insns_check;
4861 /* Info about last usage of registers in EBB to do inheritance/split
4862 transformation. Inheritance transformation is done from a spilled
4863 pseudo and split transformations from a hard register or a pseudo
4864 assigned to a hard register. */
4865 struct usage_insns
4867 /* If the value is equal to CURR_USAGE_INSNS_CHECK, then the member
4868 value INSNS is valid. The insns is chain of optional debug insns
4869 and a finishing non-debug insn using the corresponding reg. The
4870 value is also used to mark the registers which are set up in the
4871 current insn. The negated insn uid is used for this. */
4872 int check;
4873 /* Value of global reloads_num at the last insn in INSNS. */
4874 int reloads_num;
4875 /* Value of global reloads_nums at the last insn in INSNS. */
4876 int calls_num;
4877 /* It can be true only for splitting. And it means that the restore
4878 insn should be put after insn given by the following member. */
4879 bool after_p;
4880 /* Next insns in the current EBB which use the original reg and the
4881 original reg value is not changed between the current insn and
4882 the next insns. In order words, e.g. for inheritance, if we need
4883 to use the original reg value again in the next insns we can try
4884 to use the value in a hard register from a reload insn of the
4885 current insn. */
4886 rtx insns;
4889 /* Map: regno -> corresponding pseudo usage insns. */
4890 static struct usage_insns *usage_insns;
4892 static void
4893 setup_next_usage_insn (int regno, rtx insn, int reloads_num, bool after_p)
4895 usage_insns[regno].check = curr_usage_insns_check;
4896 usage_insns[regno].insns = insn;
4897 usage_insns[regno].reloads_num = reloads_num;
4898 usage_insns[regno].calls_num = calls_num;
4899 usage_insns[regno].after_p = after_p;
4902 /* The function is used to form list REGNO usages which consists of
4903 optional debug insns finished by a non-debug insn using REGNO.
4904 RELOADS_NUM is current number of reload insns processed so far. */
4905 static void
4906 add_next_usage_insn (int regno, rtx_insn *insn, int reloads_num)
4908 rtx next_usage_insns;
4910 if (usage_insns[regno].check == curr_usage_insns_check
4911 && (next_usage_insns = usage_insns[regno].insns) != NULL_RTX
4912 && DEBUG_INSN_P (insn))
4914 /* Check that we did not add the debug insn yet. */
4915 if (next_usage_insns != insn
4916 && (GET_CODE (next_usage_insns) != INSN_LIST
4917 || XEXP (next_usage_insns, 0) != insn))
4918 usage_insns[regno].insns = gen_rtx_INSN_LIST (VOIDmode, insn,
4919 next_usage_insns);
4921 else if (NONDEBUG_INSN_P (insn))
4922 setup_next_usage_insn (regno, insn, reloads_num, false);
4923 else
4924 usage_insns[regno].check = 0;
4927 /* Return first non-debug insn in list USAGE_INSNS. */
4928 static rtx_insn *
4929 skip_usage_debug_insns (rtx usage_insns)
4931 rtx insn;
4933 /* Skip debug insns. */
4934 for (insn = usage_insns;
4935 insn != NULL_RTX && GET_CODE (insn) == INSN_LIST;
4936 insn = XEXP (insn, 1))
4938 return safe_as_a <rtx_insn *> (insn);
4941 /* Return true if we need secondary memory moves for insn in
4942 USAGE_INSNS after inserting inherited pseudo of class INHER_CL
4943 into the insn. */
4944 static bool
4945 check_secondary_memory_needed_p (enum reg_class inher_cl ATTRIBUTE_UNUSED,
4946 rtx usage_insns ATTRIBUTE_UNUSED)
4948 #ifndef SECONDARY_MEMORY_NEEDED
4949 return false;
4950 #else
4951 rtx_insn *insn;
4952 rtx set, dest;
4953 enum reg_class cl;
4955 if (inher_cl == ALL_REGS
4956 || (insn = skip_usage_debug_insns (usage_insns)) == NULL_RTX)
4957 return false;
4958 lra_assert (INSN_P (insn));
4959 if ((set = single_set (insn)) == NULL_RTX || ! REG_P (SET_DEST (set)))
4960 return false;
4961 dest = SET_DEST (set);
4962 if (! REG_P (dest))
4963 return false;
4964 lra_assert (inher_cl != NO_REGS);
4965 cl = get_reg_class (REGNO (dest));
4966 return (cl != NO_REGS && cl != ALL_REGS
4967 && SECONDARY_MEMORY_NEEDED (inher_cl, cl, GET_MODE (dest)));
4968 #endif
4971 /* Registers involved in inheritance/split in the current EBB
4972 (inheritance/split pseudos and original registers). */
4973 static bitmap_head check_only_regs;
4975 /* Reload pseudos can not be involded in invariant inheritance in the
4976 current EBB. */
4977 static bitmap_head invalid_invariant_regs;
4979 /* Do inheritance transformations for insn INSN, which defines (if
4980 DEF_P) or uses ORIGINAL_REGNO. NEXT_USAGE_INSNS specifies which
4981 instruction in the EBB next uses ORIGINAL_REGNO; it has the same
4982 form as the "insns" field of usage_insns. Return true if we
4983 succeed in such transformation.
4985 The transformations look like:
4987 p <- ... i <- ...
4988 ... p <- i (new insn)
4989 ... =>
4990 <- ... p ... <- ... i ...
4992 ... i <- p (new insn)
4993 <- ... p ... <- ... i ...
4994 ... =>
4995 <- ... p ... <- ... i ...
4996 where p is a spilled original pseudo and i is a new inheritance pseudo.
4999 The inheritance pseudo has the smallest class of two classes CL and
5000 class of ORIGINAL REGNO. */
5001 static bool
5002 inherit_reload_reg (bool def_p, int original_regno,
5003 enum reg_class cl, rtx_insn *insn, rtx next_usage_insns)
5005 if (optimize_function_for_size_p (cfun))
5006 return false;
5008 enum reg_class rclass = lra_get_allocno_class (original_regno);
5009 rtx original_reg = regno_reg_rtx[original_regno];
5010 rtx new_reg, usage_insn;
5011 rtx_insn *new_insns;
5013 lra_assert (! usage_insns[original_regno].after_p);
5014 if (lra_dump_file != NULL)
5015 fprintf (lra_dump_file,
5016 " <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<\n");
5017 if (! ira_reg_classes_intersect_p[cl][rclass])
5019 if (lra_dump_file != NULL)
5021 fprintf (lra_dump_file,
5022 " Rejecting inheritance for %d "
5023 "because of disjoint classes %s and %s\n",
5024 original_regno, reg_class_names[cl],
5025 reg_class_names[rclass]);
5026 fprintf (lra_dump_file,
5027 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5029 return false;
5031 if ((ira_class_subset_p[cl][rclass] && cl != rclass)
5032 /* We don't use a subset of two classes because it can be
5033 NO_REGS. This transformation is still profitable in most
5034 cases even if the classes are not intersected as register
5035 move is probably cheaper than a memory load. */
5036 || ira_class_hard_regs_num[cl] < ira_class_hard_regs_num[rclass])
5038 if (lra_dump_file != NULL)
5039 fprintf (lra_dump_file, " Use smallest class of %s and %s\n",
5040 reg_class_names[cl], reg_class_names[rclass]);
5042 rclass = cl;
5044 if (check_secondary_memory_needed_p (rclass, next_usage_insns))
5046 /* Reject inheritance resulting in secondary memory moves.
5047 Otherwise, there is a danger in LRA cycling. Also such
5048 transformation will be unprofitable. */
5049 if (lra_dump_file != NULL)
5051 rtx_insn *insn = skip_usage_debug_insns (next_usage_insns);
5052 rtx set = single_set (insn);
5054 lra_assert (set != NULL_RTX);
5056 rtx dest = SET_DEST (set);
5058 lra_assert (REG_P (dest));
5059 fprintf (lra_dump_file,
5060 " Rejecting inheritance for insn %d(%s)<-%d(%s) "
5061 "as secondary mem is needed\n",
5062 REGNO (dest), reg_class_names[get_reg_class (REGNO (dest))],
5063 original_regno, reg_class_names[rclass]);
5064 fprintf (lra_dump_file,
5065 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5067 return false;
5069 new_reg = lra_create_new_reg (GET_MODE (original_reg), original_reg,
5070 rclass, "inheritance");
5071 start_sequence ();
5072 if (def_p)
5073 lra_emit_move (original_reg, new_reg);
5074 else
5075 lra_emit_move (new_reg, original_reg);
5076 new_insns = get_insns ();
5077 end_sequence ();
5078 if (NEXT_INSN (new_insns) != NULL_RTX)
5080 if (lra_dump_file != NULL)
5082 fprintf (lra_dump_file,
5083 " Rejecting inheritance %d->%d "
5084 "as it results in 2 or more insns:\n",
5085 original_regno, REGNO (new_reg));
5086 dump_rtl_slim (lra_dump_file, new_insns, NULL, -1, 0);
5087 fprintf (lra_dump_file,
5088 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5090 return false;
5092 lra_substitute_pseudo_within_insn (insn, original_regno, new_reg, false);
5093 lra_update_insn_regno_info (insn);
5094 if (! def_p)
5095 /* We now have a new usage insn for original regno. */
5096 setup_next_usage_insn (original_regno, new_insns, reloads_num, false);
5097 if (lra_dump_file != NULL)
5098 fprintf (lra_dump_file, " Original reg change %d->%d (bb%d):\n",
5099 original_regno, REGNO (new_reg), BLOCK_FOR_INSN (insn)->index);
5100 lra_reg_info[REGNO (new_reg)].restore_rtx = regno_reg_rtx[original_regno];
5101 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
5102 bitmap_set_bit (&check_only_regs, original_regno);
5103 bitmap_set_bit (&lra_inheritance_pseudos, REGNO (new_reg));
5104 if (def_p)
5105 lra_process_new_insns (insn, NULL, new_insns,
5106 "Add original<-inheritance");
5107 else
5108 lra_process_new_insns (insn, new_insns, NULL,
5109 "Add inheritance<-original");
5110 while (next_usage_insns != NULL_RTX)
5112 if (GET_CODE (next_usage_insns) != INSN_LIST)
5114 usage_insn = next_usage_insns;
5115 lra_assert (NONDEBUG_INSN_P (usage_insn));
5116 next_usage_insns = NULL;
5118 else
5120 usage_insn = XEXP (next_usage_insns, 0);
5121 lra_assert (DEBUG_INSN_P (usage_insn));
5122 next_usage_insns = XEXP (next_usage_insns, 1);
5124 lra_substitute_pseudo (&usage_insn, original_regno, new_reg, false);
5125 lra_update_insn_regno_info (as_a <rtx_insn *> (usage_insn));
5126 if (lra_dump_file != NULL)
5128 fprintf (lra_dump_file,
5129 " Inheritance reuse change %d->%d (bb%d):\n",
5130 original_regno, REGNO (new_reg),
5131 BLOCK_FOR_INSN (usage_insn)->index);
5132 dump_insn_slim (lra_dump_file, as_a <rtx_insn *> (usage_insn));
5135 if (lra_dump_file != NULL)
5136 fprintf (lra_dump_file,
5137 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5138 return true;
5141 /* Return true if we need a caller save/restore for pseudo REGNO which
5142 was assigned to a hard register. */
5143 static inline bool
5144 need_for_call_save_p (int regno)
5146 lra_assert (regno >= FIRST_PSEUDO_REGISTER && reg_renumber[regno] >= 0);
5147 return (usage_insns[regno].calls_num < calls_num
5148 && (overlaps_hard_reg_set_p
5149 ((flag_ipa_ra &&
5150 ! hard_reg_set_empty_p (lra_reg_info[regno].actual_call_used_reg_set))
5151 ? lra_reg_info[regno].actual_call_used_reg_set
5152 : call_used_reg_set,
5153 PSEUDO_REGNO_MODE (regno), reg_renumber[regno])
5154 || HARD_REGNO_CALL_PART_CLOBBERED (reg_renumber[regno],
5155 PSEUDO_REGNO_MODE (regno))));
5158 /* Global registers occurring in the current EBB. */
5159 static bitmap_head ebb_global_regs;
5161 /* Return true if we need a split for hard register REGNO or pseudo
5162 REGNO which was assigned to a hard register.
5163 POTENTIAL_RELOAD_HARD_REGS contains hard registers which might be
5164 used for reloads since the EBB end. It is an approximation of the
5165 used hard registers in the split range. The exact value would
5166 require expensive calculations. If we were aggressive with
5167 splitting because of the approximation, the split pseudo will save
5168 the same hard register assignment and will be removed in the undo
5169 pass. We still need the approximation because too aggressive
5170 splitting would result in too inaccurate cost calculation in the
5171 assignment pass because of too many generated moves which will be
5172 probably removed in the undo pass. */
5173 static inline bool
5174 need_for_split_p (HARD_REG_SET potential_reload_hard_regs, int regno)
5176 int hard_regno = regno < FIRST_PSEUDO_REGISTER ? regno : reg_renumber[regno];
5178 lra_assert (hard_regno >= 0);
5179 return ((TEST_HARD_REG_BIT (potential_reload_hard_regs, hard_regno)
5180 /* Don't split eliminable hard registers, otherwise we can
5181 split hard registers like hard frame pointer, which
5182 lives on BB start/end according to DF-infrastructure,
5183 when there is a pseudo assigned to the register and
5184 living in the same BB. */
5185 && (regno >= FIRST_PSEUDO_REGISTER
5186 || ! TEST_HARD_REG_BIT (eliminable_regset, hard_regno))
5187 && ! TEST_HARD_REG_BIT (lra_no_alloc_regs, hard_regno)
5188 /* Don't split call clobbered hard regs living through
5189 calls, otherwise we might have a check problem in the
5190 assign sub-pass as in the most cases (exception is a
5191 situation when lra_risky_transformations_p value is
5192 true) the assign pass assumes that all pseudos living
5193 through calls are assigned to call saved hard regs. */
5194 && (regno >= FIRST_PSEUDO_REGISTER
5195 || ! TEST_HARD_REG_BIT (call_used_reg_set, regno)
5196 || usage_insns[regno].calls_num == calls_num)
5197 /* We need at least 2 reloads to make pseudo splitting
5198 profitable. We should provide hard regno splitting in
5199 any case to solve 1st insn scheduling problem when
5200 moving hard register definition up might result in
5201 impossibility to find hard register for reload pseudo of
5202 small register class. */
5203 && (usage_insns[regno].reloads_num
5204 + (regno < FIRST_PSEUDO_REGISTER ? 0 : 3) < reloads_num)
5205 && (regno < FIRST_PSEUDO_REGISTER
5206 /* For short living pseudos, spilling + inheritance can
5207 be considered a substitution for splitting.
5208 Therefore we do not splitting for local pseudos. It
5209 decreases also aggressiveness of splitting. The
5210 minimal number of references is chosen taking into
5211 account that for 2 references splitting has no sense
5212 as we can just spill the pseudo. */
5213 || (regno >= FIRST_PSEUDO_REGISTER
5214 && lra_reg_info[regno].nrefs > 3
5215 && bitmap_bit_p (&ebb_global_regs, regno))))
5216 || (regno >= FIRST_PSEUDO_REGISTER && need_for_call_save_p (regno)));
5219 /* Return class for the split pseudo created from original pseudo with
5220 ALLOCNO_CLASS and MODE which got a hard register HARD_REGNO. We
5221 choose subclass of ALLOCNO_CLASS which contains HARD_REGNO and
5222 results in no secondary memory movements. */
5223 static enum reg_class
5224 choose_split_class (enum reg_class allocno_class,
5225 int hard_regno ATTRIBUTE_UNUSED,
5226 machine_mode mode ATTRIBUTE_UNUSED)
5228 #ifndef SECONDARY_MEMORY_NEEDED
5229 return allocno_class;
5230 #else
5231 int i;
5232 enum reg_class cl, best_cl = NO_REGS;
5233 enum reg_class hard_reg_class ATTRIBUTE_UNUSED
5234 = REGNO_REG_CLASS (hard_regno);
5236 if (! SECONDARY_MEMORY_NEEDED (allocno_class, allocno_class, mode)
5237 && TEST_HARD_REG_BIT (reg_class_contents[allocno_class], hard_regno))
5238 return allocno_class;
5239 for (i = 0;
5240 (cl = reg_class_subclasses[allocno_class][i]) != LIM_REG_CLASSES;
5241 i++)
5242 if (! SECONDARY_MEMORY_NEEDED (cl, hard_reg_class, mode)
5243 && ! SECONDARY_MEMORY_NEEDED (hard_reg_class, cl, mode)
5244 && TEST_HARD_REG_BIT (reg_class_contents[cl], hard_regno)
5245 && (best_cl == NO_REGS
5246 || ira_class_hard_regs_num[best_cl] < ira_class_hard_regs_num[cl]))
5247 best_cl = cl;
5248 return best_cl;
5249 #endif
5252 /* Do split transformations for insn INSN, which defines or uses
5253 ORIGINAL_REGNO. NEXT_USAGE_INSNS specifies which instruction in
5254 the EBB next uses ORIGINAL_REGNO; it has the same form as the
5255 "insns" field of usage_insns.
5257 The transformations look like:
5259 p <- ... p <- ...
5260 ... s <- p (new insn -- save)
5261 ... =>
5262 ... p <- s (new insn -- restore)
5263 <- ... p ... <- ... p ...
5265 <- ... p ... <- ... p ...
5266 ... s <- p (new insn -- save)
5267 ... =>
5268 ... p <- s (new insn -- restore)
5269 <- ... p ... <- ... p ...
5271 where p is an original pseudo got a hard register or a hard
5272 register and s is a new split pseudo. The save is put before INSN
5273 if BEFORE_P is true. Return true if we succeed in such
5274 transformation. */
5275 static bool
5276 split_reg (bool before_p, int original_regno, rtx_insn *insn,
5277 rtx next_usage_insns)
5279 enum reg_class rclass;
5280 rtx original_reg;
5281 int hard_regno, nregs;
5282 rtx new_reg, usage_insn;
5283 rtx_insn *restore, *save;
5284 bool after_p;
5285 bool call_save_p;
5286 machine_mode mode;
5288 if (original_regno < FIRST_PSEUDO_REGISTER)
5290 rclass = ira_allocno_class_translate[REGNO_REG_CLASS (original_regno)];
5291 hard_regno = original_regno;
5292 call_save_p = false;
5293 nregs = 1;
5294 mode = lra_reg_info[hard_regno].biggest_mode;
5295 machine_mode reg_rtx_mode = GET_MODE (regno_reg_rtx[hard_regno]);
5296 /* A reg can have a biggest_mode of VOIDmode if it was only ever seen
5297 as part of a multi-word register. In that case, or if the biggest
5298 mode was larger than a register, just use the reg_rtx. Otherwise,
5299 limit the size to that of the biggest access in the function. */
5300 if (mode == VOIDmode
5301 || GET_MODE_SIZE (mode) > GET_MODE_SIZE (reg_rtx_mode))
5303 original_reg = regno_reg_rtx[hard_regno];
5304 mode = reg_rtx_mode;
5306 else
5307 original_reg = gen_rtx_REG (mode, hard_regno);
5309 else
5311 mode = PSEUDO_REGNO_MODE (original_regno);
5312 hard_regno = reg_renumber[original_regno];
5313 nregs = hard_regno_nregs[hard_regno][mode];
5314 rclass = lra_get_allocno_class (original_regno);
5315 original_reg = regno_reg_rtx[original_regno];
5316 call_save_p = need_for_call_save_p (original_regno);
5318 lra_assert (hard_regno >= 0);
5319 if (lra_dump_file != NULL)
5320 fprintf (lra_dump_file,
5321 " ((((((((((((((((((((((((((((((((((((((((((((((((\n");
5323 if (call_save_p)
5325 mode = HARD_REGNO_CALLER_SAVE_MODE (hard_regno,
5326 hard_regno_nregs[hard_regno][mode],
5327 mode);
5328 new_reg = lra_create_new_reg (mode, NULL_RTX, NO_REGS, "save");
5330 else
5332 rclass = choose_split_class (rclass, hard_regno, mode);
5333 if (rclass == NO_REGS)
5335 if (lra_dump_file != NULL)
5337 fprintf (lra_dump_file,
5338 " Rejecting split of %d(%s): "
5339 "no good reg class for %d(%s)\n",
5340 original_regno,
5341 reg_class_names[lra_get_allocno_class (original_regno)],
5342 hard_regno,
5343 reg_class_names[REGNO_REG_CLASS (hard_regno)]);
5344 fprintf
5345 (lra_dump_file,
5346 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5348 return false;
5350 new_reg = lra_create_new_reg (mode, original_reg, rclass, "split");
5351 reg_renumber[REGNO (new_reg)] = hard_regno;
5353 save = emit_spill_move (true, new_reg, original_reg);
5354 if (NEXT_INSN (save) != NULL_RTX && !call_save_p)
5356 if (lra_dump_file != NULL)
5358 fprintf
5359 (lra_dump_file,
5360 " Rejecting split %d->%d resulting in > 2 save insns:\n",
5361 original_regno, REGNO (new_reg));
5362 dump_rtl_slim (lra_dump_file, save, NULL, -1, 0);
5363 fprintf (lra_dump_file,
5364 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5366 return false;
5368 restore = emit_spill_move (false, new_reg, original_reg);
5369 if (NEXT_INSN (restore) != NULL_RTX && !call_save_p)
5371 if (lra_dump_file != NULL)
5373 fprintf (lra_dump_file,
5374 " Rejecting split %d->%d "
5375 "resulting in > 2 restore insns:\n",
5376 original_regno, REGNO (new_reg));
5377 dump_rtl_slim (lra_dump_file, restore, NULL, -1, 0);
5378 fprintf (lra_dump_file,
5379 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5381 return false;
5383 after_p = usage_insns[original_regno].after_p;
5384 lra_reg_info[REGNO (new_reg)].restore_rtx = regno_reg_rtx[original_regno];
5385 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
5386 bitmap_set_bit (&check_only_regs, original_regno);
5387 bitmap_set_bit (&lra_split_regs, REGNO (new_reg));
5388 for (;;)
5390 if (GET_CODE (next_usage_insns) != INSN_LIST)
5392 usage_insn = next_usage_insns;
5393 break;
5395 usage_insn = XEXP (next_usage_insns, 0);
5396 lra_assert (DEBUG_INSN_P (usage_insn));
5397 next_usage_insns = XEXP (next_usage_insns, 1);
5398 lra_substitute_pseudo (&usage_insn, original_regno, new_reg, false);
5399 lra_update_insn_regno_info (as_a <rtx_insn *> (usage_insn));
5400 if (lra_dump_file != NULL)
5402 fprintf (lra_dump_file, " Split reuse change %d->%d:\n",
5403 original_regno, REGNO (new_reg));
5404 dump_insn_slim (lra_dump_file, as_a <rtx_insn *> (usage_insn));
5407 lra_assert (NOTE_P (usage_insn) || NONDEBUG_INSN_P (usage_insn));
5408 lra_assert (usage_insn != insn || (after_p && before_p));
5409 lra_process_new_insns (as_a <rtx_insn *> (usage_insn),
5410 after_p ? NULL : restore,
5411 after_p ? restore : NULL,
5412 call_save_p
5413 ? "Add reg<-save" : "Add reg<-split");
5414 lra_process_new_insns (insn, before_p ? save : NULL,
5415 before_p ? NULL : save,
5416 call_save_p
5417 ? "Add save<-reg" : "Add split<-reg");
5418 if (nregs > 1)
5419 /* If we are trying to split multi-register. We should check
5420 conflicts on the next assignment sub-pass. IRA can allocate on
5421 sub-register levels, LRA do this on pseudos level right now and
5422 this discrepancy may create allocation conflicts after
5423 splitting. */
5424 lra_risky_transformations_p = true;
5425 if (lra_dump_file != NULL)
5426 fprintf (lra_dump_file,
5427 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5428 return true;
5431 /* Recognize that we need a split transformation for insn INSN, which
5432 defines or uses REGNO in its insn biggest MODE (we use it only if
5433 REGNO is a hard register). POTENTIAL_RELOAD_HARD_REGS contains
5434 hard registers which might be used for reloads since the EBB end.
5435 Put the save before INSN if BEFORE_P is true. MAX_UID is maximla
5436 uid before starting INSN processing. Return true if we succeed in
5437 such transformation. */
5438 static bool
5439 split_if_necessary (int regno, machine_mode mode,
5440 HARD_REG_SET potential_reload_hard_regs,
5441 bool before_p, rtx_insn *insn, int max_uid)
5443 bool res = false;
5444 int i, nregs = 1;
5445 rtx next_usage_insns;
5447 if (regno < FIRST_PSEUDO_REGISTER)
5448 nregs = hard_regno_nregs[regno][mode];
5449 for (i = 0; i < nregs; i++)
5450 if (usage_insns[regno + i].check == curr_usage_insns_check
5451 && (next_usage_insns = usage_insns[regno + i].insns) != NULL_RTX
5452 /* To avoid processing the register twice or more. */
5453 && ((GET_CODE (next_usage_insns) != INSN_LIST
5454 && INSN_UID (next_usage_insns) < max_uid)
5455 || (GET_CODE (next_usage_insns) == INSN_LIST
5456 && (INSN_UID (XEXP (next_usage_insns, 0)) < max_uid)))
5457 && need_for_split_p (potential_reload_hard_regs, regno + i)
5458 && split_reg (before_p, regno + i, insn, next_usage_insns))
5459 res = true;
5460 return res;
5463 /* Return TRUE if rtx X is considered as an invariant for
5464 inheritance. */
5465 static bool
5466 invariant_p (const_rtx x)
5468 machine_mode mode;
5469 const char *fmt;
5470 enum rtx_code code;
5471 int i, j;
5473 code = GET_CODE (x);
5474 mode = GET_MODE (x);
5475 if (code == SUBREG)
5477 x = SUBREG_REG (x);
5478 code = GET_CODE (x);
5479 if (GET_MODE_SIZE (GET_MODE (x)) > GET_MODE_SIZE (mode))
5480 mode = GET_MODE (x);
5483 if (MEM_P (x))
5484 return false;
5486 if (REG_P (x))
5488 int i, nregs, regno = REGNO (x);
5490 if (regno >= FIRST_PSEUDO_REGISTER || regno == STACK_POINTER_REGNUM
5491 || TEST_HARD_REG_BIT (eliminable_regset, regno)
5492 || GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
5493 return false;
5494 nregs = hard_regno_nregs[regno][mode];
5495 for (i = 0; i < nregs; i++)
5496 if (! fixed_regs[regno + i]
5497 /* A hard register may be clobbered in the current insn
5498 but we can ignore this case because if the hard
5499 register is used it should be set somewhere after the
5500 clobber. */
5501 || bitmap_bit_p (&invalid_invariant_regs, regno + i))
5502 return false;
5504 fmt = GET_RTX_FORMAT (code);
5505 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
5507 if (fmt[i] == 'e')
5509 if (! invariant_p (XEXP (x, i)))
5510 return false;
5512 else if (fmt[i] == 'E')
5514 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
5515 if (! invariant_p (XVECEXP (x, i, j)))
5516 return false;
5519 return true;
5522 /* We have 'dest_reg <- invariant'. Let us try to make an invariant
5523 inheritance transformation (using dest_reg instead invariant in a
5524 subsequent insn). */
5525 static bool
5526 process_invariant_for_inheritance (rtx dst_reg, rtx invariant_rtx)
5528 invariant_ptr_t invariant_ptr;
5529 rtx_insn *insn, *new_insns;
5530 rtx insn_set, insn_reg, new_reg;
5531 int insn_regno;
5532 bool succ_p = false;
5533 int dst_regno = REGNO (dst_reg);
5534 enum machine_mode dst_mode = GET_MODE (dst_reg);
5535 enum reg_class cl = lra_get_allocno_class (dst_regno), insn_reg_cl;
5537 invariant_ptr = insert_invariant (invariant_rtx);
5538 if ((insn = invariant_ptr->insn) != NULL_RTX)
5540 /* We have a subsequent insn using the invariant. */
5541 insn_set = single_set (insn);
5542 lra_assert (insn_set != NULL);
5543 insn_reg = SET_DEST (insn_set);
5544 lra_assert (REG_P (insn_reg));
5545 insn_regno = REGNO (insn_reg);
5546 insn_reg_cl = lra_get_allocno_class (insn_regno);
5548 if (dst_mode == GET_MODE (insn_reg)
5549 /* We should consider only result move reg insns which are
5550 cheap. */
5551 && targetm.register_move_cost (dst_mode, cl, insn_reg_cl) == 2
5552 && targetm.register_move_cost (dst_mode, cl, cl) == 2)
5554 if (lra_dump_file != NULL)
5555 fprintf (lra_dump_file,
5556 " [[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[\n");
5557 new_reg = lra_create_new_reg (dst_mode, dst_reg,
5558 cl, "invariant inheritance");
5559 bitmap_set_bit (&lra_inheritance_pseudos, REGNO (new_reg));
5560 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
5561 lra_reg_info[REGNO (new_reg)].restore_rtx = PATTERN (insn);
5562 start_sequence ();
5563 lra_emit_move (new_reg, dst_reg);
5564 new_insns = get_insns ();
5565 end_sequence ();
5566 lra_process_new_insns (curr_insn, NULL, new_insns,
5567 "Add invariant inheritance<-original");
5568 start_sequence ();
5569 lra_emit_move (SET_DEST (insn_set), new_reg);
5570 new_insns = get_insns ();
5571 end_sequence ();
5572 lra_process_new_insns (insn, NULL, new_insns,
5573 "Changing reload<-inheritance");
5574 lra_set_insn_deleted (insn);
5575 succ_p = true;
5576 if (lra_dump_file != NULL)
5578 fprintf (lra_dump_file,
5579 " Invariant inheritance reuse change %d (bb%d):\n",
5580 REGNO (new_reg), BLOCK_FOR_INSN (insn)->index);
5581 dump_insn_slim (lra_dump_file, insn);
5582 fprintf (lra_dump_file,
5583 " ]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]\n");
5587 invariant_ptr->insn = curr_insn;
5588 return succ_p;
5591 /* Check only registers living at the current program point in the
5592 current EBB. */
5593 static bitmap_head live_regs;
5595 /* Update live info in EBB given by its HEAD and TAIL insns after
5596 inheritance/split transformation. The function removes dead moves
5597 too. */
5598 static void
5599 update_ebb_live_info (rtx_insn *head, rtx_insn *tail)
5601 unsigned int j;
5602 int i, regno;
5603 bool live_p;
5604 rtx_insn *prev_insn;
5605 rtx set;
5606 bool remove_p;
5607 basic_block last_bb, prev_bb, curr_bb;
5608 bitmap_iterator bi;
5609 struct lra_insn_reg *reg;
5610 edge e;
5611 edge_iterator ei;
5613 last_bb = BLOCK_FOR_INSN (tail);
5614 prev_bb = NULL;
5615 for (curr_insn = tail;
5616 curr_insn != PREV_INSN (head);
5617 curr_insn = prev_insn)
5619 prev_insn = PREV_INSN (curr_insn);
5620 /* We need to process empty blocks too. They contain
5621 NOTE_INSN_BASIC_BLOCK referring for the basic block. */
5622 if (NOTE_P (curr_insn) && NOTE_KIND (curr_insn) != NOTE_INSN_BASIC_BLOCK)
5623 continue;
5624 curr_bb = BLOCK_FOR_INSN (curr_insn);
5625 if (curr_bb != prev_bb)
5627 if (prev_bb != NULL)
5629 /* Update df_get_live_in (prev_bb): */
5630 EXECUTE_IF_SET_IN_BITMAP (&check_only_regs, 0, j, bi)
5631 if (bitmap_bit_p (&live_regs, j))
5632 bitmap_set_bit (df_get_live_in (prev_bb), j);
5633 else
5634 bitmap_clear_bit (df_get_live_in (prev_bb), j);
5636 if (curr_bb != last_bb)
5638 /* Update df_get_live_out (curr_bb): */
5639 EXECUTE_IF_SET_IN_BITMAP (&check_only_regs, 0, j, bi)
5641 live_p = bitmap_bit_p (&live_regs, j);
5642 if (! live_p)
5643 FOR_EACH_EDGE (e, ei, curr_bb->succs)
5644 if (bitmap_bit_p (df_get_live_in (e->dest), j))
5646 live_p = true;
5647 break;
5649 if (live_p)
5650 bitmap_set_bit (df_get_live_out (curr_bb), j);
5651 else
5652 bitmap_clear_bit (df_get_live_out (curr_bb), j);
5655 prev_bb = curr_bb;
5656 bitmap_and (&live_regs, &check_only_regs, df_get_live_out (curr_bb));
5658 if (! NONDEBUG_INSN_P (curr_insn))
5659 continue;
5660 curr_id = lra_get_insn_recog_data (curr_insn);
5661 curr_static_id = curr_id->insn_static_data;
5662 remove_p = false;
5663 if ((set = single_set (curr_insn)) != NULL_RTX
5664 && REG_P (SET_DEST (set))
5665 && (regno = REGNO (SET_DEST (set))) >= FIRST_PSEUDO_REGISTER
5666 && SET_DEST (set) != pic_offset_table_rtx
5667 && bitmap_bit_p (&check_only_regs, regno)
5668 && ! bitmap_bit_p (&live_regs, regno))
5669 remove_p = true;
5670 /* See which defined values die here. */
5671 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5672 if (reg->type == OP_OUT && ! reg->subreg_p)
5673 bitmap_clear_bit (&live_regs, reg->regno);
5674 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
5675 if (reg->type == OP_OUT && ! reg->subreg_p)
5676 bitmap_clear_bit (&live_regs, reg->regno);
5677 if (curr_id->arg_hard_regs != NULL)
5678 /* Make clobbered argument hard registers die. */
5679 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
5680 if (regno >= FIRST_PSEUDO_REGISTER)
5681 bitmap_clear_bit (&live_regs, regno - FIRST_PSEUDO_REGISTER);
5682 /* Mark each used value as live. */
5683 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5684 if (reg->type != OP_OUT
5685 && bitmap_bit_p (&check_only_regs, reg->regno))
5686 bitmap_set_bit (&live_regs, reg->regno);
5687 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
5688 if (reg->type != OP_OUT
5689 && bitmap_bit_p (&check_only_regs, reg->regno))
5690 bitmap_set_bit (&live_regs, reg->regno);
5691 if (curr_id->arg_hard_regs != NULL)
5692 /* Make used argument hard registers live. */
5693 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
5694 if (regno < FIRST_PSEUDO_REGISTER
5695 && bitmap_bit_p (&check_only_regs, regno))
5696 bitmap_set_bit (&live_regs, regno);
5697 /* It is quite important to remove dead move insns because it
5698 means removing dead store. We don't need to process them for
5699 constraints. */
5700 if (remove_p)
5702 if (lra_dump_file != NULL)
5704 fprintf (lra_dump_file, " Removing dead insn:\n ");
5705 dump_insn_slim (lra_dump_file, curr_insn);
5707 lra_set_insn_deleted (curr_insn);
5712 /* The structure describes info to do an inheritance for the current
5713 insn. We need to collect such info first before doing the
5714 transformations because the transformations change the insn
5715 internal representation. */
5716 struct to_inherit
5718 /* Original regno. */
5719 int regno;
5720 /* Subsequent insns which can inherit original reg value. */
5721 rtx insns;
5724 /* Array containing all info for doing inheritance from the current
5725 insn. */
5726 static struct to_inherit to_inherit[LRA_MAX_INSN_RELOADS];
5728 /* Number elements in the previous array. */
5729 static int to_inherit_num;
5731 /* Add inheritance info REGNO and INSNS. Their meaning is described in
5732 structure to_inherit. */
5733 static void
5734 add_to_inherit (int regno, rtx insns)
5736 int i;
5738 for (i = 0; i < to_inherit_num; i++)
5739 if (to_inherit[i].regno == regno)
5740 return;
5741 lra_assert (to_inherit_num < LRA_MAX_INSN_RELOADS);
5742 to_inherit[to_inherit_num].regno = regno;
5743 to_inherit[to_inherit_num++].insns = insns;
5746 /* Return the last non-debug insn in basic block BB, or the block begin
5747 note if none. */
5748 static rtx_insn *
5749 get_last_insertion_point (basic_block bb)
5751 rtx_insn *insn;
5753 FOR_BB_INSNS_REVERSE (bb, insn)
5754 if (NONDEBUG_INSN_P (insn) || NOTE_INSN_BASIC_BLOCK_P (insn))
5755 return insn;
5756 gcc_unreachable ();
5759 /* Set up RES by registers living on edges FROM except the edge (FROM,
5760 TO) or by registers set up in a jump insn in BB FROM. */
5761 static void
5762 get_live_on_other_edges (basic_block from, basic_block to, bitmap res)
5764 rtx_insn *last;
5765 struct lra_insn_reg *reg;
5766 edge e;
5767 edge_iterator ei;
5769 lra_assert (to != NULL);
5770 bitmap_clear (res);
5771 FOR_EACH_EDGE (e, ei, from->succs)
5772 if (e->dest != to)
5773 bitmap_ior_into (res, df_get_live_in (e->dest));
5774 last = get_last_insertion_point (from);
5775 if (! JUMP_P (last))
5776 return;
5777 curr_id = lra_get_insn_recog_data (last);
5778 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5779 if (reg->type != OP_IN)
5780 bitmap_set_bit (res, reg->regno);
5783 /* Used as a temporary results of some bitmap calculations. */
5784 static bitmap_head temp_bitmap;
5786 /* We split for reloads of small class of hard regs. The following
5787 defines how many hard regs the class should have to be qualified as
5788 small. The code is mostly oriented to x86/x86-64 architecture
5789 where some insns need to use only specific register or pair of
5790 registers and these register can live in RTL explicitly, e.g. for
5791 parameter passing. */
5792 static const int max_small_class_regs_num = 2;
5794 /* Do inheritance/split transformations in EBB starting with HEAD and
5795 finishing on TAIL. We process EBB insns in the reverse order.
5796 Return true if we did any inheritance/split transformation in the
5797 EBB.
5799 We should avoid excessive splitting which results in worse code
5800 because of inaccurate cost calculations for spilling new split
5801 pseudos in such case. To achieve this we do splitting only if
5802 register pressure is high in given basic block and there are reload
5803 pseudos requiring hard registers. We could do more register
5804 pressure calculations at any given program point to avoid necessary
5805 splitting even more but it is to expensive and the current approach
5806 works well enough. */
5807 static bool
5808 inherit_in_ebb (rtx_insn *head, rtx_insn *tail)
5810 int i, src_regno, dst_regno, nregs;
5811 bool change_p, succ_p, update_reloads_num_p;
5812 rtx_insn *prev_insn, *last_insn;
5813 rtx next_usage_insns, curr_set;
5814 enum reg_class cl;
5815 struct lra_insn_reg *reg;
5816 basic_block last_processed_bb, curr_bb = NULL;
5817 HARD_REG_SET potential_reload_hard_regs, live_hard_regs;
5818 bitmap to_process;
5819 unsigned int j;
5820 bitmap_iterator bi;
5821 bool head_p, after_p;
5823 change_p = false;
5824 curr_usage_insns_check++;
5825 clear_invariants ();
5826 reloads_num = calls_num = 0;
5827 bitmap_clear (&check_only_regs);
5828 bitmap_clear (&invalid_invariant_regs);
5829 last_processed_bb = NULL;
5830 CLEAR_HARD_REG_SET (potential_reload_hard_regs);
5831 COPY_HARD_REG_SET (live_hard_regs, eliminable_regset);
5832 IOR_HARD_REG_SET (live_hard_regs, lra_no_alloc_regs);
5833 /* We don't process new insns generated in the loop. */
5834 for (curr_insn = tail; curr_insn != PREV_INSN (head); curr_insn = prev_insn)
5836 prev_insn = PREV_INSN (curr_insn);
5837 if (BLOCK_FOR_INSN (curr_insn) != NULL)
5838 curr_bb = BLOCK_FOR_INSN (curr_insn);
5839 if (last_processed_bb != curr_bb)
5841 /* We are at the end of BB. Add qualified living
5842 pseudos for potential splitting. */
5843 to_process = df_get_live_out (curr_bb);
5844 if (last_processed_bb != NULL)
5846 /* We are somewhere in the middle of EBB. */
5847 get_live_on_other_edges (curr_bb, last_processed_bb,
5848 &temp_bitmap);
5849 to_process = &temp_bitmap;
5851 last_processed_bb = curr_bb;
5852 last_insn = get_last_insertion_point (curr_bb);
5853 after_p = (! JUMP_P (last_insn)
5854 && (! CALL_P (last_insn)
5855 || (find_reg_note (last_insn,
5856 REG_NORETURN, NULL_RTX) == NULL_RTX
5857 && ! SIBLING_CALL_P (last_insn))));
5858 CLEAR_HARD_REG_SET (potential_reload_hard_regs);
5859 EXECUTE_IF_SET_IN_BITMAP (to_process, 0, j, bi)
5861 if ((int) j >= lra_constraint_new_regno_start)
5862 break;
5863 if (j < FIRST_PSEUDO_REGISTER || reg_renumber[j] >= 0)
5865 if (j < FIRST_PSEUDO_REGISTER)
5866 SET_HARD_REG_BIT (live_hard_regs, j);
5867 else
5868 add_to_hard_reg_set (&live_hard_regs,
5869 PSEUDO_REGNO_MODE (j),
5870 reg_renumber[j]);
5871 setup_next_usage_insn (j, last_insn, reloads_num, after_p);
5875 src_regno = dst_regno = -1;
5876 curr_set = single_set (curr_insn);
5877 if (curr_set != NULL_RTX && REG_P (SET_DEST (curr_set)))
5878 dst_regno = REGNO (SET_DEST (curr_set));
5879 if (curr_set != NULL_RTX && REG_P (SET_SRC (curr_set)))
5880 src_regno = REGNO (SET_SRC (curr_set));
5881 update_reloads_num_p = true;
5882 if (src_regno < lra_constraint_new_regno_start
5883 && src_regno >= FIRST_PSEUDO_REGISTER
5884 && reg_renumber[src_regno] < 0
5885 && dst_regno >= lra_constraint_new_regno_start
5886 && (cl = lra_get_allocno_class (dst_regno)) != NO_REGS)
5888 /* 'reload_pseudo <- original_pseudo'. */
5889 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
5890 reloads_num++;
5891 update_reloads_num_p = false;
5892 succ_p = false;
5893 if (usage_insns[src_regno].check == curr_usage_insns_check
5894 && (next_usage_insns = usage_insns[src_regno].insns) != NULL_RTX)
5895 succ_p = inherit_reload_reg (false, src_regno, cl,
5896 curr_insn, next_usage_insns);
5897 if (succ_p)
5898 change_p = true;
5899 else
5900 setup_next_usage_insn (src_regno, curr_insn, reloads_num, false);
5901 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
5902 IOR_HARD_REG_SET (potential_reload_hard_regs,
5903 reg_class_contents[cl]);
5905 else if (src_regno < 0
5906 && dst_regno >= lra_constraint_new_regno_start
5907 && invariant_p (SET_SRC (curr_set))
5908 && (cl = lra_get_allocno_class (dst_regno)) != NO_REGS
5909 && ! bitmap_bit_p (&invalid_invariant_regs, dst_regno)
5910 && ! bitmap_bit_p (&invalid_invariant_regs,
5911 ORIGINAL_REGNO(regno_reg_rtx[dst_regno])))
5913 /* 'reload_pseudo <- invariant'. */
5914 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
5915 reloads_num++;
5916 update_reloads_num_p = false;
5917 if (process_invariant_for_inheritance (SET_DEST (curr_set), SET_SRC (curr_set)))
5918 change_p = true;
5919 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
5920 IOR_HARD_REG_SET (potential_reload_hard_regs,
5921 reg_class_contents[cl]);
5923 else if (src_regno >= lra_constraint_new_regno_start
5924 && dst_regno < lra_constraint_new_regno_start
5925 && dst_regno >= FIRST_PSEUDO_REGISTER
5926 && reg_renumber[dst_regno] < 0
5927 && (cl = lra_get_allocno_class (src_regno)) != NO_REGS
5928 && usage_insns[dst_regno].check == curr_usage_insns_check
5929 && (next_usage_insns
5930 = usage_insns[dst_regno].insns) != NULL_RTX)
5932 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
5933 reloads_num++;
5934 update_reloads_num_p = false;
5935 /* 'original_pseudo <- reload_pseudo'. */
5936 if (! JUMP_P (curr_insn)
5937 && inherit_reload_reg (true, dst_regno, cl,
5938 curr_insn, next_usage_insns))
5939 change_p = true;
5940 /* Invalidate. */
5941 usage_insns[dst_regno].check = 0;
5942 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
5943 IOR_HARD_REG_SET (potential_reload_hard_regs,
5944 reg_class_contents[cl]);
5946 else if (INSN_P (curr_insn))
5948 int iter;
5949 int max_uid = get_max_uid ();
5951 curr_id = lra_get_insn_recog_data (curr_insn);
5952 curr_static_id = curr_id->insn_static_data;
5953 to_inherit_num = 0;
5954 /* Process insn definitions. */
5955 for (iter = 0; iter < 2; iter++)
5956 for (reg = iter == 0 ? curr_id->regs : curr_static_id->hard_regs;
5957 reg != NULL;
5958 reg = reg->next)
5959 if (reg->type != OP_IN
5960 && (dst_regno = reg->regno) < lra_constraint_new_regno_start)
5962 if (dst_regno >= FIRST_PSEUDO_REGISTER && reg->type == OP_OUT
5963 && reg_renumber[dst_regno] < 0 && ! reg->subreg_p
5964 && usage_insns[dst_regno].check == curr_usage_insns_check
5965 && (next_usage_insns
5966 = usage_insns[dst_regno].insns) != NULL_RTX)
5968 struct lra_insn_reg *r;
5970 for (r = curr_id->regs; r != NULL; r = r->next)
5971 if (r->type != OP_OUT && r->regno == dst_regno)
5972 break;
5973 /* Don't do inheritance if the pseudo is also
5974 used in the insn. */
5975 if (r == NULL)
5976 /* We can not do inheritance right now
5977 because the current insn reg info (chain
5978 regs) can change after that. */
5979 add_to_inherit (dst_regno, next_usage_insns);
5981 /* We can not process one reg twice here because of
5982 usage_insns invalidation. */
5983 if ((dst_regno < FIRST_PSEUDO_REGISTER
5984 || reg_renumber[dst_regno] >= 0)
5985 && ! reg->subreg_p && reg->type != OP_IN)
5987 HARD_REG_SET s;
5989 if (split_if_necessary (dst_regno, reg->biggest_mode,
5990 potential_reload_hard_regs,
5991 false, curr_insn, max_uid))
5992 change_p = true;
5993 CLEAR_HARD_REG_SET (s);
5994 if (dst_regno < FIRST_PSEUDO_REGISTER)
5995 add_to_hard_reg_set (&s, reg->biggest_mode, dst_regno);
5996 else
5997 add_to_hard_reg_set (&s, PSEUDO_REGNO_MODE (dst_regno),
5998 reg_renumber[dst_regno]);
5999 AND_COMPL_HARD_REG_SET (live_hard_regs, s);
6001 /* We should invalidate potential inheritance or
6002 splitting for the current insn usages to the next
6003 usage insns (see code below) as the output pseudo
6004 prevents this. */
6005 if ((dst_regno >= FIRST_PSEUDO_REGISTER
6006 && reg_renumber[dst_regno] < 0)
6007 || (reg->type == OP_OUT && ! reg->subreg_p
6008 && (dst_regno < FIRST_PSEUDO_REGISTER
6009 || reg_renumber[dst_regno] >= 0)))
6011 /* Invalidate and mark definitions. */
6012 if (dst_regno >= FIRST_PSEUDO_REGISTER)
6013 usage_insns[dst_regno].check = -(int) INSN_UID (curr_insn);
6014 else
6016 nregs = hard_regno_nregs[dst_regno][reg->biggest_mode];
6017 for (i = 0; i < nregs; i++)
6018 usage_insns[dst_regno + i].check
6019 = -(int) INSN_UID (curr_insn);
6023 /* Process clobbered call regs. */
6024 if (curr_id->arg_hard_regs != NULL)
6025 for (i = 0; (dst_regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6026 if (dst_regno >= FIRST_PSEUDO_REGISTER)
6027 usage_insns[dst_regno - FIRST_PSEUDO_REGISTER].check
6028 = -(int) INSN_UID (curr_insn);
6029 if (! JUMP_P (curr_insn))
6030 for (i = 0; i < to_inherit_num; i++)
6031 if (inherit_reload_reg (true, to_inherit[i].regno,
6032 ALL_REGS, curr_insn,
6033 to_inherit[i].insns))
6034 change_p = true;
6035 if (CALL_P (curr_insn))
6037 rtx cheap, pat, dest;
6038 rtx_insn *restore;
6039 int regno, hard_regno;
6041 calls_num++;
6042 if ((cheap = find_reg_note (curr_insn,
6043 REG_RETURNED, NULL_RTX)) != NULL_RTX
6044 && ((cheap = XEXP (cheap, 0)), true)
6045 && (regno = REGNO (cheap)) >= FIRST_PSEUDO_REGISTER
6046 && (hard_regno = reg_renumber[regno]) >= 0
6047 /* If there are pending saves/restores, the
6048 optimization is not worth. */
6049 && usage_insns[regno].calls_num == calls_num - 1
6050 && TEST_HARD_REG_BIT (call_used_reg_set, hard_regno))
6052 /* Restore the pseudo from the call result as
6053 REG_RETURNED note says that the pseudo value is
6054 in the call result and the pseudo is an argument
6055 of the call. */
6056 pat = PATTERN (curr_insn);
6057 if (GET_CODE (pat) == PARALLEL)
6058 pat = XVECEXP (pat, 0, 0);
6059 dest = SET_DEST (pat);
6060 /* For multiple return values dest is PARALLEL.
6061 Currently we handle only single return value case. */
6062 if (REG_P (dest))
6064 start_sequence ();
6065 emit_move_insn (cheap, copy_rtx (dest));
6066 restore = get_insns ();
6067 end_sequence ();
6068 lra_process_new_insns (curr_insn, NULL, restore,
6069 "Inserting call parameter restore");
6070 /* We don't need to save/restore of the pseudo from
6071 this call. */
6072 usage_insns[regno].calls_num = calls_num;
6073 bitmap_set_bit (&check_only_regs, regno);
6077 to_inherit_num = 0;
6078 /* Process insn usages. */
6079 for (iter = 0; iter < 2; iter++)
6080 for (reg = iter == 0 ? curr_id->regs : curr_static_id->hard_regs;
6081 reg != NULL;
6082 reg = reg->next)
6083 if ((reg->type != OP_OUT
6084 || (reg->type == OP_OUT && reg->subreg_p))
6085 && (src_regno = reg->regno) < lra_constraint_new_regno_start)
6087 if (src_regno >= FIRST_PSEUDO_REGISTER
6088 && reg_renumber[src_regno] < 0 && reg->type == OP_IN)
6090 if (usage_insns[src_regno].check == curr_usage_insns_check
6091 && (next_usage_insns
6092 = usage_insns[src_regno].insns) != NULL_RTX
6093 && NONDEBUG_INSN_P (curr_insn))
6094 add_to_inherit (src_regno, next_usage_insns);
6095 else if (usage_insns[src_regno].check
6096 != -(int) INSN_UID (curr_insn))
6097 /* Add usages but only if the reg is not set up
6098 in the same insn. */
6099 add_next_usage_insn (src_regno, curr_insn, reloads_num);
6101 else if (src_regno < FIRST_PSEUDO_REGISTER
6102 || reg_renumber[src_regno] >= 0)
6104 bool before_p;
6105 rtx_insn *use_insn = curr_insn;
6107 before_p = (JUMP_P (curr_insn)
6108 || (CALL_P (curr_insn) && reg->type == OP_IN));
6109 if (NONDEBUG_INSN_P (curr_insn)
6110 && (! JUMP_P (curr_insn) || reg->type == OP_IN)
6111 && split_if_necessary (src_regno, reg->biggest_mode,
6112 potential_reload_hard_regs,
6113 before_p, curr_insn, max_uid))
6115 if (reg->subreg_p)
6116 lra_risky_transformations_p = true;
6117 change_p = true;
6118 /* Invalidate. */
6119 usage_insns[src_regno].check = 0;
6120 if (before_p)
6121 use_insn = PREV_INSN (curr_insn);
6123 if (NONDEBUG_INSN_P (curr_insn))
6125 if (src_regno < FIRST_PSEUDO_REGISTER)
6126 add_to_hard_reg_set (&live_hard_regs,
6127 reg->biggest_mode, src_regno);
6128 else
6129 add_to_hard_reg_set (&live_hard_regs,
6130 PSEUDO_REGNO_MODE (src_regno),
6131 reg_renumber[src_regno]);
6133 add_next_usage_insn (src_regno, use_insn, reloads_num);
6136 /* Process used call regs. */
6137 if (curr_id->arg_hard_regs != NULL)
6138 for (i = 0; (src_regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6139 if (src_regno < FIRST_PSEUDO_REGISTER)
6141 SET_HARD_REG_BIT (live_hard_regs, src_regno);
6142 add_next_usage_insn (src_regno, curr_insn, reloads_num);
6144 for (i = 0; i < to_inherit_num; i++)
6146 src_regno = to_inherit[i].regno;
6147 if (inherit_reload_reg (false, src_regno, ALL_REGS,
6148 curr_insn, to_inherit[i].insns))
6149 change_p = true;
6150 else
6151 setup_next_usage_insn (src_regno, curr_insn, reloads_num, false);
6154 if (update_reloads_num_p
6155 && NONDEBUG_INSN_P (curr_insn) && curr_set != NULL_RTX)
6157 int regno = -1;
6158 if ((REG_P (SET_DEST (curr_set))
6159 && (regno = REGNO (SET_DEST (curr_set))) >= lra_constraint_new_regno_start
6160 && reg_renumber[regno] < 0
6161 && (cl = lra_get_allocno_class (regno)) != NO_REGS)
6162 || (REG_P (SET_SRC (curr_set))
6163 && (regno = REGNO (SET_SRC (curr_set))) >= lra_constraint_new_regno_start
6164 && reg_renumber[regno] < 0
6165 && (cl = lra_get_allocno_class (regno)) != NO_REGS))
6167 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6168 reloads_num++;
6169 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6170 IOR_HARD_REG_SET (potential_reload_hard_regs,
6171 reg_class_contents[cl]);
6174 if (NONDEBUG_INSN_P (curr_insn))
6176 int regno;
6178 /* Invalidate invariants with changed regs. */
6179 curr_id = lra_get_insn_recog_data (curr_insn);
6180 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
6181 if (reg->type != OP_IN)
6183 bitmap_set_bit (&invalid_invariant_regs, reg->regno);
6184 bitmap_set_bit (&invalid_invariant_regs,
6185 ORIGINAL_REGNO (regno_reg_rtx[reg->regno]));
6187 curr_static_id = curr_id->insn_static_data;
6188 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
6189 if (reg->type != OP_IN)
6190 bitmap_set_bit (&invalid_invariant_regs, reg->regno);
6191 if (curr_id->arg_hard_regs != NULL)
6192 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6193 if (regno >= FIRST_PSEUDO_REGISTER)
6194 bitmap_set_bit (&invalid_invariant_regs,
6195 regno - FIRST_PSEUDO_REGISTER);
6197 /* We reached the start of the current basic block. */
6198 if (prev_insn == NULL_RTX || prev_insn == PREV_INSN (head)
6199 || BLOCK_FOR_INSN (prev_insn) != curr_bb)
6201 /* We reached the beginning of the current block -- do
6202 rest of spliting in the current BB. */
6203 to_process = df_get_live_in (curr_bb);
6204 if (BLOCK_FOR_INSN (head) != curr_bb)
6206 /* We are somewhere in the middle of EBB. */
6207 get_live_on_other_edges (EDGE_PRED (curr_bb, 0)->src,
6208 curr_bb, &temp_bitmap);
6209 to_process = &temp_bitmap;
6211 head_p = true;
6212 EXECUTE_IF_SET_IN_BITMAP (to_process, 0, j, bi)
6214 if ((int) j >= lra_constraint_new_regno_start)
6215 break;
6216 if (((int) j < FIRST_PSEUDO_REGISTER || reg_renumber[j] >= 0)
6217 && usage_insns[j].check == curr_usage_insns_check
6218 && (next_usage_insns = usage_insns[j].insns) != NULL_RTX)
6220 if (need_for_split_p (potential_reload_hard_regs, j))
6222 if (lra_dump_file != NULL && head_p)
6224 fprintf (lra_dump_file,
6225 " ----------------------------------\n");
6226 head_p = false;
6228 if (split_reg (false, j, bb_note (curr_bb),
6229 next_usage_insns))
6230 change_p = true;
6232 usage_insns[j].check = 0;
6237 return change_p;
6240 /* This value affects EBB forming. If probability of edge from EBB to
6241 a BB is not greater than the following value, we don't add the BB
6242 to EBB. */
6243 #define EBB_PROBABILITY_CUTOFF \
6244 ((REG_BR_PROB_BASE * LRA_INHERITANCE_EBB_PROBABILITY_CUTOFF) / 100)
6246 /* Current number of inheritance/split iteration. */
6247 int lra_inheritance_iter;
6249 /* Entry function for inheritance/split pass. */
6250 void
6251 lra_inheritance (void)
6253 int i;
6254 basic_block bb, start_bb;
6255 edge e;
6257 lra_inheritance_iter++;
6258 if (lra_inheritance_iter > LRA_MAX_INHERITANCE_PASSES)
6259 return;
6260 timevar_push (TV_LRA_INHERITANCE);
6261 if (lra_dump_file != NULL)
6262 fprintf (lra_dump_file, "\n********** Inheritance #%d: **********\n\n",
6263 lra_inheritance_iter);
6264 curr_usage_insns_check = 0;
6265 usage_insns = XNEWVEC (struct usage_insns, lra_constraint_new_regno_start);
6266 for (i = 0; i < lra_constraint_new_regno_start; i++)
6267 usage_insns[i].check = 0;
6268 bitmap_initialize (&check_only_regs, &reg_obstack);
6269 bitmap_initialize (&invalid_invariant_regs, &reg_obstack);
6270 bitmap_initialize (&live_regs, &reg_obstack);
6271 bitmap_initialize (&temp_bitmap, &reg_obstack);
6272 bitmap_initialize (&ebb_global_regs, &reg_obstack);
6273 FOR_EACH_BB_FN (bb, cfun)
6275 start_bb = bb;
6276 if (lra_dump_file != NULL)
6277 fprintf (lra_dump_file, "EBB");
6278 /* Form a EBB starting with BB. */
6279 bitmap_clear (&ebb_global_regs);
6280 bitmap_ior_into (&ebb_global_regs, df_get_live_in (bb));
6281 for (;;)
6283 if (lra_dump_file != NULL)
6284 fprintf (lra_dump_file, " %d", bb->index);
6285 if (bb->next_bb == EXIT_BLOCK_PTR_FOR_FN (cfun)
6286 || LABEL_P (BB_HEAD (bb->next_bb)))
6287 break;
6288 e = find_fallthru_edge (bb->succs);
6289 if (! e)
6290 break;
6291 if (e->probability < EBB_PROBABILITY_CUTOFF)
6292 break;
6293 bb = bb->next_bb;
6295 bitmap_ior_into (&ebb_global_regs, df_get_live_out (bb));
6296 if (lra_dump_file != NULL)
6297 fprintf (lra_dump_file, "\n");
6298 if (inherit_in_ebb (BB_HEAD (start_bb), BB_END (bb)))
6299 /* Remember that the EBB head and tail can change in
6300 inherit_in_ebb. */
6301 update_ebb_live_info (BB_HEAD (start_bb), BB_END (bb));
6303 bitmap_clear (&ebb_global_regs);
6304 bitmap_clear (&temp_bitmap);
6305 bitmap_clear (&live_regs);
6306 bitmap_clear (&invalid_invariant_regs);
6307 bitmap_clear (&check_only_regs);
6308 free (usage_insns);
6310 timevar_pop (TV_LRA_INHERITANCE);
6315 /* This page contains code to undo failed inheritance/split
6316 transformations. */
6318 /* Current number of iteration undoing inheritance/split. */
6319 int lra_undo_inheritance_iter;
6321 /* Fix BB live info LIVE after removing pseudos created on pass doing
6322 inheritance/split which are REMOVED_PSEUDOS. */
6323 static void
6324 fix_bb_live_info (bitmap live, bitmap removed_pseudos)
6326 unsigned int regno;
6327 bitmap_iterator bi;
6329 EXECUTE_IF_SET_IN_BITMAP (removed_pseudos, 0, regno, bi)
6330 if (bitmap_clear_bit (live, regno)
6331 && REG_P (lra_reg_info[regno].restore_rtx))
6332 bitmap_set_bit (live, REGNO (lra_reg_info[regno].restore_rtx));
6335 /* Return regno of the (subreg of) REG. Otherwise, return a negative
6336 number. */
6337 static int
6338 get_regno (rtx reg)
6340 if (GET_CODE (reg) == SUBREG)
6341 reg = SUBREG_REG (reg);
6342 if (REG_P (reg))
6343 return REGNO (reg);
6344 return -1;
6347 /* Delete a move INSN with destination reg DREGNO and a previous
6348 clobber insn with the same regno. The inheritance/split code can
6349 generate moves with preceding clobber and when we delete such moves
6350 we should delete the clobber insn too to keep the correct life
6351 info. */
6352 static void
6353 delete_move_and_clobber (rtx_insn *insn, int dregno)
6355 rtx_insn *prev_insn = PREV_INSN (insn);
6357 lra_set_insn_deleted (insn);
6358 lra_assert (dregno >= 0);
6359 if (prev_insn != NULL && NONDEBUG_INSN_P (prev_insn)
6360 && GET_CODE (PATTERN (prev_insn)) == CLOBBER
6361 && dregno == get_regno (XEXP (PATTERN (prev_insn), 0)))
6362 lra_set_insn_deleted (prev_insn);
6365 /* Remove inheritance/split pseudos which are in REMOVE_PSEUDOS and
6366 return true if we did any change. The undo transformations for
6367 inheritance looks like
6368 i <- i2
6369 p <- i => p <- i2
6370 or removing
6371 p <- i, i <- p, and i <- i3
6372 where p is original pseudo from which inheritance pseudo i was
6373 created, i and i3 are removed inheritance pseudos, i2 is another
6374 not removed inheritance pseudo. All split pseudos or other
6375 occurrences of removed inheritance pseudos are changed on the
6376 corresponding original pseudos.
6378 The function also schedules insns changed and created during
6379 inheritance/split pass for processing by the subsequent constraint
6380 pass. */
6381 static bool
6382 remove_inheritance_pseudos (bitmap remove_pseudos)
6384 basic_block bb;
6385 int regno, sregno, prev_sregno, dregno;
6386 rtx restore_rtx;
6387 rtx set, prev_set;
6388 rtx_insn *prev_insn;
6389 bool change_p, done_p;
6391 change_p = ! bitmap_empty_p (remove_pseudos);
6392 /* We can not finish the function right away if CHANGE_P is true
6393 because we need to marks insns affected by previous
6394 inheritance/split pass for processing by the subsequent
6395 constraint pass. */
6396 FOR_EACH_BB_FN (bb, cfun)
6398 fix_bb_live_info (df_get_live_in (bb), remove_pseudos);
6399 fix_bb_live_info (df_get_live_out (bb), remove_pseudos);
6400 FOR_BB_INSNS_REVERSE (bb, curr_insn)
6402 if (! INSN_P (curr_insn))
6403 continue;
6404 done_p = false;
6405 sregno = dregno = -1;
6406 if (change_p && NONDEBUG_INSN_P (curr_insn)
6407 && (set = single_set (curr_insn)) != NULL_RTX)
6409 dregno = get_regno (SET_DEST (set));
6410 sregno = get_regno (SET_SRC (set));
6413 if (sregno >= 0 && dregno >= 0)
6415 if (bitmap_bit_p (remove_pseudos, dregno)
6416 && ! REG_P (lra_reg_info[dregno].restore_rtx))
6418 /* invariant inheritance pseudo <- original pseudo */
6419 if (lra_dump_file != NULL)
6421 fprintf (lra_dump_file, " Removing invariant inheritance:\n");
6422 dump_insn_slim (lra_dump_file, curr_insn);
6423 fprintf (lra_dump_file, "\n");
6425 delete_move_and_clobber (curr_insn, dregno);
6426 done_p = true;
6428 else if (bitmap_bit_p (remove_pseudos, sregno)
6429 && ! REG_P (lra_reg_info[sregno].restore_rtx))
6431 /* reload pseudo <- invariant inheritance pseudo */
6432 start_sequence ();
6433 /* We can not just change the source. It might be
6434 an insn different from the move. */
6435 emit_insn (lra_reg_info[sregno].restore_rtx);
6436 rtx_insn *new_insns = get_insns ();
6437 end_sequence ();
6438 lra_assert (single_set (new_insns) != NULL
6439 && SET_DEST (set) == SET_DEST (single_set (new_insns)));
6440 lra_process_new_insns (curr_insn, NULL, new_insns,
6441 "Changing reload<-invariant inheritance");
6442 delete_move_and_clobber (curr_insn, dregno);
6443 done_p = true;
6445 else if ((bitmap_bit_p (remove_pseudos, sregno)
6446 && (get_regno (lra_reg_info[sregno].restore_rtx) == dregno
6447 || (bitmap_bit_p (remove_pseudos, dregno)
6448 && get_regno (lra_reg_info[sregno].restore_rtx) >= 0
6449 && (get_regno (lra_reg_info[sregno].restore_rtx)
6450 == get_regno (lra_reg_info[dregno].restore_rtx)))))
6451 || (bitmap_bit_p (remove_pseudos, dregno)
6452 && get_regno (lra_reg_info[dregno].restore_rtx) == sregno))
6453 /* One of the following cases:
6454 original <- removed inheritance pseudo
6455 removed inherit pseudo <- another removed inherit pseudo
6456 removed inherit pseudo <- original pseudo
6458 removed_split_pseudo <- original_reg
6459 original_reg <- removed_split_pseudo */
6461 if (lra_dump_file != NULL)
6463 fprintf (lra_dump_file, " Removing %s:\n",
6464 bitmap_bit_p (&lra_split_regs, sregno)
6465 || bitmap_bit_p (&lra_split_regs, dregno)
6466 ? "split" : "inheritance");
6467 dump_insn_slim (lra_dump_file, curr_insn);
6469 delete_move_and_clobber (curr_insn, dregno);
6470 done_p = true;
6472 else if (bitmap_bit_p (remove_pseudos, sregno)
6473 && bitmap_bit_p (&lra_inheritance_pseudos, sregno))
6475 /* Search the following pattern:
6476 inherit_or_split_pseudo1 <- inherit_or_split_pseudo2
6477 original_pseudo <- inherit_or_split_pseudo1
6478 where the 2nd insn is the current insn and
6479 inherit_or_split_pseudo2 is not removed. If it is found,
6480 change the current insn onto:
6481 original_pseudo <- inherit_or_split_pseudo2. */
6482 for (prev_insn = PREV_INSN (curr_insn);
6483 prev_insn != NULL_RTX && ! NONDEBUG_INSN_P (prev_insn);
6484 prev_insn = PREV_INSN (prev_insn))
6486 if (prev_insn != NULL_RTX && BLOCK_FOR_INSN (prev_insn) == bb
6487 && (prev_set = single_set (prev_insn)) != NULL_RTX
6488 /* There should be no subregs in insn we are
6489 searching because only the original reg might
6490 be in subreg when we changed the mode of
6491 load/store for splitting. */
6492 && REG_P (SET_DEST (prev_set))
6493 && REG_P (SET_SRC (prev_set))
6494 && (int) REGNO (SET_DEST (prev_set)) == sregno
6495 && ((prev_sregno = REGNO (SET_SRC (prev_set)))
6496 >= FIRST_PSEUDO_REGISTER)
6497 && (lra_reg_info[prev_sregno].restore_rtx == NULL_RTX
6499 /* As we consider chain of inheritance or
6500 splitting described in above comment we should
6501 check that sregno and prev_sregno were
6502 inheritance/split pseudos created from the
6503 same original regno. */
6504 (get_regno (lra_reg_info[sregno].restore_rtx) >= 0
6505 && (get_regno (lra_reg_info[sregno].restore_rtx)
6506 == get_regno (lra_reg_info[prev_sregno].restore_rtx))))
6507 && ! bitmap_bit_p (remove_pseudos, prev_sregno))
6509 lra_assert (GET_MODE (SET_SRC (prev_set))
6510 == GET_MODE (regno_reg_rtx[sregno]));
6511 if (GET_CODE (SET_SRC (set)) == SUBREG)
6512 SUBREG_REG (SET_SRC (set)) = SET_SRC (prev_set);
6513 else
6514 SET_SRC (set) = SET_SRC (prev_set);
6515 /* As we are finishing with processing the insn
6516 here, check the destination too as it might
6517 inheritance pseudo for another pseudo. */
6518 if (bitmap_bit_p (remove_pseudos, dregno)
6519 && bitmap_bit_p (&lra_inheritance_pseudos, dregno)
6520 && (restore_rtx
6521 = lra_reg_info[dregno].restore_rtx) != NULL_RTX)
6523 if (GET_CODE (SET_DEST (set)) == SUBREG)
6524 SUBREG_REG (SET_DEST (set)) = restore_rtx;
6525 else
6526 SET_DEST (set) = restore_rtx;
6528 lra_push_insn_and_update_insn_regno_info (curr_insn);
6529 lra_set_used_insn_alternative_by_uid
6530 (INSN_UID (curr_insn), -1);
6531 done_p = true;
6532 if (lra_dump_file != NULL)
6534 fprintf (lra_dump_file, " Change reload insn:\n");
6535 dump_insn_slim (lra_dump_file, curr_insn);
6540 if (! done_p)
6542 struct lra_insn_reg *reg;
6543 bool restored_regs_p = false;
6544 bool kept_regs_p = false;
6546 curr_id = lra_get_insn_recog_data (curr_insn);
6547 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
6549 regno = reg->regno;
6550 restore_rtx = lra_reg_info[regno].restore_rtx;
6551 if (restore_rtx != NULL_RTX)
6553 if (change_p && bitmap_bit_p (remove_pseudos, regno))
6555 lra_substitute_pseudo_within_insn
6556 (curr_insn, regno, restore_rtx, false);
6557 restored_regs_p = true;
6559 else
6560 kept_regs_p = true;
6563 if (NONDEBUG_INSN_P (curr_insn) && kept_regs_p)
6565 /* The instruction has changed since the previous
6566 constraints pass. */
6567 lra_push_insn_and_update_insn_regno_info (curr_insn);
6568 lra_set_used_insn_alternative_by_uid
6569 (INSN_UID (curr_insn), -1);
6571 else if (restored_regs_p)
6572 /* The instruction has been restored to the form that
6573 it had during the previous constraints pass. */
6574 lra_update_insn_regno_info (curr_insn);
6575 if (restored_regs_p && lra_dump_file != NULL)
6577 fprintf (lra_dump_file, " Insn after restoring regs:\n");
6578 dump_insn_slim (lra_dump_file, curr_insn);
6583 return change_p;
6586 /* If optional reload pseudos failed to get a hard register or was not
6587 inherited, it is better to remove optional reloads. We do this
6588 transformation after undoing inheritance to figure out necessity to
6589 remove optional reloads easier. Return true if we do any
6590 change. */
6591 static bool
6592 undo_optional_reloads (void)
6594 bool change_p, keep_p;
6595 unsigned int regno, uid;
6596 bitmap_iterator bi, bi2;
6597 rtx_insn *insn;
6598 rtx set, src, dest;
6599 bitmap_head removed_optional_reload_pseudos, insn_bitmap;
6601 bitmap_initialize (&removed_optional_reload_pseudos, &reg_obstack);
6602 bitmap_copy (&removed_optional_reload_pseudos, &lra_optional_reload_pseudos);
6603 EXECUTE_IF_SET_IN_BITMAP (&lra_optional_reload_pseudos, 0, regno, bi)
6605 keep_p = false;
6606 /* Keep optional reloads from previous subpasses. */
6607 if (lra_reg_info[regno].restore_rtx == NULL_RTX
6608 /* If the original pseudo changed its allocation, just
6609 removing the optional pseudo is dangerous as the original
6610 pseudo will have longer live range. */
6611 || reg_renumber[REGNO (lra_reg_info[regno].restore_rtx)] >= 0)
6612 keep_p = true;
6613 else if (reg_renumber[regno] >= 0)
6614 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi2)
6616 insn = lra_insn_recog_data[uid]->insn;
6617 if ((set = single_set (insn)) == NULL_RTX)
6618 continue;
6619 src = SET_SRC (set);
6620 dest = SET_DEST (set);
6621 if (! REG_P (src) || ! REG_P (dest))
6622 continue;
6623 if (REGNO (dest) == regno
6624 /* Ignore insn for optional reloads itself. */
6625 && REGNO (lra_reg_info[regno].restore_rtx) != REGNO (src)
6626 /* Check only inheritance on last inheritance pass. */
6627 && (int) REGNO (src) >= new_regno_start
6628 /* Check that the optional reload was inherited. */
6629 && bitmap_bit_p (&lra_inheritance_pseudos, REGNO (src)))
6631 keep_p = true;
6632 break;
6635 if (keep_p)
6637 bitmap_clear_bit (&removed_optional_reload_pseudos, regno);
6638 if (lra_dump_file != NULL)
6639 fprintf (lra_dump_file, "Keep optional reload reg %d\n", regno);
6642 change_p = ! bitmap_empty_p (&removed_optional_reload_pseudos);
6643 bitmap_initialize (&insn_bitmap, &reg_obstack);
6644 EXECUTE_IF_SET_IN_BITMAP (&removed_optional_reload_pseudos, 0, regno, bi)
6646 if (lra_dump_file != NULL)
6647 fprintf (lra_dump_file, "Remove optional reload reg %d\n", regno);
6648 bitmap_copy (&insn_bitmap, &lra_reg_info[regno].insn_bitmap);
6649 EXECUTE_IF_SET_IN_BITMAP (&insn_bitmap, 0, uid, bi2)
6651 insn = lra_insn_recog_data[uid]->insn;
6652 if ((set = single_set (insn)) != NULL_RTX)
6654 src = SET_SRC (set);
6655 dest = SET_DEST (set);
6656 if (REG_P (src) && REG_P (dest)
6657 && ((REGNO (src) == regno
6658 && (REGNO (lra_reg_info[regno].restore_rtx)
6659 == REGNO (dest)))
6660 || (REGNO (dest) == regno
6661 && (REGNO (lra_reg_info[regno].restore_rtx)
6662 == REGNO (src)))))
6664 if (lra_dump_file != NULL)
6666 fprintf (lra_dump_file, " Deleting move %u\n",
6667 INSN_UID (insn));
6668 dump_insn_slim (lra_dump_file, insn);
6670 delete_move_and_clobber (insn, REGNO (dest));
6671 continue;
6673 /* We should not worry about generation memory-memory
6674 moves here as if the corresponding inheritance did
6675 not work (inheritance pseudo did not get a hard reg),
6676 we remove the inheritance pseudo and the optional
6677 reload. */
6679 lra_substitute_pseudo_within_insn
6680 (insn, regno, lra_reg_info[regno].restore_rtx, false);
6681 lra_update_insn_regno_info (insn);
6682 if (lra_dump_file != NULL)
6684 fprintf (lra_dump_file,
6685 " Restoring original insn:\n");
6686 dump_insn_slim (lra_dump_file, insn);
6690 /* Clear restore_regnos. */
6691 EXECUTE_IF_SET_IN_BITMAP (&lra_optional_reload_pseudos, 0, regno, bi)
6692 lra_reg_info[regno].restore_rtx = NULL_RTX;
6693 bitmap_clear (&insn_bitmap);
6694 bitmap_clear (&removed_optional_reload_pseudos);
6695 return change_p;
6698 /* Entry function for undoing inheritance/split transformation. Return true
6699 if we did any RTL change in this pass. */
6700 bool
6701 lra_undo_inheritance (void)
6703 unsigned int regno;
6704 int hard_regno;
6705 int n_all_inherit, n_inherit, n_all_split, n_split;
6706 rtx restore_rtx;
6707 bitmap_head remove_pseudos;
6708 bitmap_iterator bi;
6709 bool change_p;
6711 lra_undo_inheritance_iter++;
6712 if (lra_undo_inheritance_iter > LRA_MAX_INHERITANCE_PASSES)
6713 return false;
6714 if (lra_dump_file != NULL)
6715 fprintf (lra_dump_file,
6716 "\n********** Undoing inheritance #%d: **********\n\n",
6717 lra_undo_inheritance_iter);
6718 bitmap_initialize (&remove_pseudos, &reg_obstack);
6719 n_inherit = n_all_inherit = 0;
6720 EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, regno, bi)
6721 if (lra_reg_info[regno].restore_rtx != NULL_RTX)
6723 n_all_inherit++;
6724 if (reg_renumber[regno] < 0
6725 /* If the original pseudo changed its allocation, just
6726 removing inheritance is dangerous as for changing
6727 allocation we used shorter live-ranges. */
6728 && (! REG_P (lra_reg_info[regno].restore_rtx)
6729 || reg_renumber[REGNO (lra_reg_info[regno].restore_rtx)] < 0))
6730 bitmap_set_bit (&remove_pseudos, regno);
6731 else
6732 n_inherit++;
6734 if (lra_dump_file != NULL && n_all_inherit != 0)
6735 fprintf (lra_dump_file, "Inherit %d out of %d (%.2f%%)\n",
6736 n_inherit, n_all_inherit,
6737 (double) n_inherit / n_all_inherit * 100);
6738 n_split = n_all_split = 0;
6739 EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, regno, bi)
6740 if ((restore_rtx = lra_reg_info[regno].restore_rtx) != NULL_RTX)
6742 int restore_regno = REGNO (restore_rtx);
6744 n_all_split++;
6745 hard_regno = (restore_regno >= FIRST_PSEUDO_REGISTER
6746 ? reg_renumber[restore_regno] : restore_regno);
6747 if (hard_regno < 0 || reg_renumber[regno] == hard_regno)
6748 bitmap_set_bit (&remove_pseudos, regno);
6749 else
6751 n_split++;
6752 if (lra_dump_file != NULL)
6753 fprintf (lra_dump_file, " Keep split r%d (orig=r%d)\n",
6754 regno, restore_regno);
6757 if (lra_dump_file != NULL && n_all_split != 0)
6758 fprintf (lra_dump_file, "Split %d out of %d (%.2f%%)\n",
6759 n_split, n_all_split,
6760 (double) n_split / n_all_split * 100);
6761 change_p = remove_inheritance_pseudos (&remove_pseudos);
6762 bitmap_clear (&remove_pseudos);
6763 /* Clear restore_regnos. */
6764 EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, regno, bi)
6765 lra_reg_info[regno].restore_rtx = NULL_RTX;
6766 EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, regno, bi)
6767 lra_reg_info[regno].restore_rtx = NULL_RTX;
6768 change_p = undo_optional_reloads () || change_p;
6769 return change_p;