2016-11-30 Vladimir Makarov <vmakarov@redhat.com>
[official-gcc.git] / gcc / lra-constraints.c
blob260591acb06bbc4d567d59d02743bc90daf13645
1 /* Code for RTL transformations to satisfy insn constraints.
2 Copyright (C) 2010-2016 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_rtx_SUBREG (GET_MODE (mem_pseudo),
1113 GET_CODE (val) == SUBREG ? 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);
1458 /* Make reloads for subreg in operand NOP with internal subreg mode
1459 REG_MODE, add new reloads for further processing. Return true if
1460 any change was done. */
1461 static bool
1462 simplify_operand_subreg (int nop, machine_mode reg_mode)
1464 int hard_regno;
1465 rtx_insn *before, *after;
1466 machine_mode mode, innermode;
1467 rtx reg, new_reg;
1468 rtx operand = *curr_id->operand_loc[nop];
1469 enum reg_class regclass;
1470 enum op_type type;
1472 before = after = NULL;
1474 if (GET_CODE (operand) != SUBREG)
1475 return false;
1477 mode = GET_MODE (operand);
1478 reg = SUBREG_REG (operand);
1479 innermode = GET_MODE (reg);
1480 type = curr_static_id->operand[nop].type;
1481 if (MEM_P (reg))
1483 rtx subst;
1485 alter_subreg (curr_id->operand_loc[nop], false);
1486 subst = *curr_id->operand_loc[nop];
1487 lra_assert (MEM_P (subst));
1488 if (! valid_address_p (innermode, XEXP (reg, 0),
1489 MEM_ADDR_SPACE (reg))
1490 || valid_address_p (GET_MODE (subst), XEXP (subst, 0),
1491 MEM_ADDR_SPACE (subst))
1492 || ((get_constraint_type (lookup_constraint
1493 (curr_static_id->operand[nop].constraint))
1494 != CT_SPECIAL_MEMORY)
1495 /* We still can reload address and if the address is
1496 valid, we can remove subreg without reloading its
1497 inner memory. */
1498 && valid_address_p (GET_MODE (subst),
1499 regno_reg_rtx
1500 [ira_class_hard_regs
1501 [base_reg_class (GET_MODE (subst),
1502 MEM_ADDR_SPACE (subst),
1503 ADDRESS, SCRATCH)][0]],
1504 MEM_ADDR_SPACE (subst))))
1506 /* If we change address for paradoxical subreg of memory, the
1507 address might violate the necessary alignment or the access might
1508 be slow. So take this into consideration. We should not worry
1509 about access beyond allocated memory for paradoxical memory
1510 subregs as we don't substitute such equiv memory (see processing
1511 equivalences in function lra_constraints) and because for spilled
1512 pseudos we allocate stack memory enough for the biggest
1513 corresponding paradoxical subreg. */
1514 if (!(MEM_ALIGN (reg) < GET_MODE_ALIGNMENT (mode)
1515 && SLOW_UNALIGNED_ACCESS (mode, MEM_ALIGN (reg)))
1516 || (MEM_ALIGN (reg) < GET_MODE_ALIGNMENT (innermode)
1517 && SLOW_UNALIGNED_ACCESS (innermode, MEM_ALIGN (reg))))
1518 return true;
1520 /* INNERMODE is fast, MODE slow. Reload the mem in INNERMODE. */
1521 enum reg_class rclass
1522 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1523 if (get_reload_reg (curr_static_id->operand[nop].type, innermode, reg,
1524 rclass, TRUE, "slow mem", &new_reg))
1526 bool insert_before, insert_after;
1527 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1529 insert_before = (type != OP_OUT
1530 || GET_MODE_SIZE (innermode) > GET_MODE_SIZE (mode));
1531 insert_after = type != OP_IN;
1532 insert_move_for_subreg (insert_before ? &before : NULL,
1533 insert_after ? &after : NULL,
1534 reg, new_reg);
1536 *curr_id->operand_loc[nop] = operand;
1537 SUBREG_REG (operand) = new_reg;
1539 /* Convert to MODE. */
1540 reg = operand;
1541 rclass = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1542 if (get_reload_reg (curr_static_id->operand[nop].type, mode, reg,
1543 rclass, TRUE, "slow mem", &new_reg))
1545 bool insert_before, insert_after;
1546 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1548 insert_before = type != OP_OUT;
1549 insert_after = type != OP_IN;
1550 insert_move_for_subreg (insert_before ? &before : NULL,
1551 insert_after ? &after : NULL,
1552 reg, new_reg);
1554 *curr_id->operand_loc[nop] = new_reg;
1555 lra_process_new_insns (curr_insn, before, after,
1556 "Inserting slow mem reload");
1557 return true;
1560 /* If the address was valid and became invalid, prefer to reload
1561 the memory. Typical case is when the index scale should
1562 correspond the memory. */
1563 *curr_id->operand_loc[nop] = operand;
1565 else if (REG_P (reg) && REGNO (reg) < FIRST_PSEUDO_REGISTER)
1567 alter_subreg (curr_id->operand_loc[nop], false);
1568 return true;
1570 else if (CONSTANT_P (reg))
1572 /* Try to simplify subreg of constant. It is usually result of
1573 equivalence substitution. */
1574 if (innermode == VOIDmode
1575 && (innermode = original_subreg_reg_mode[nop]) == VOIDmode)
1576 innermode = curr_static_id->operand[nop].mode;
1577 if ((new_reg = simplify_subreg (mode, reg, innermode,
1578 SUBREG_BYTE (operand))) != NULL_RTX)
1580 *curr_id->operand_loc[nop] = new_reg;
1581 return true;
1584 /* Put constant into memory when we have mixed modes. It generates
1585 a better code in most cases as it does not need a secondary
1586 reload memory. It also prevents LRA looping when LRA is using
1587 secondary reload memory again and again. */
1588 if (CONSTANT_P (reg) && CONST_POOL_OK_P (reg_mode, reg)
1589 && SCALAR_INT_MODE_P (reg_mode) != SCALAR_INT_MODE_P (mode))
1591 SUBREG_REG (operand) = force_const_mem (reg_mode, reg);
1592 alter_subreg (curr_id->operand_loc[nop], false);
1593 return true;
1595 /* Force a reload of the SUBREG_REG if this is a constant or PLUS or
1596 if there may be a problem accessing OPERAND in the outer
1597 mode. */
1598 if ((REG_P (reg)
1599 && REGNO (reg) >= FIRST_PSEUDO_REGISTER
1600 && (hard_regno = lra_get_regno_hard_regno (REGNO (reg))) >= 0
1601 /* Don't reload paradoxical subregs because we could be looping
1602 having repeatedly final regno out of hard regs range. */
1603 && (hard_regno_nregs[hard_regno][innermode]
1604 >= hard_regno_nregs[hard_regno][mode])
1605 && simplify_subreg_regno (hard_regno, innermode,
1606 SUBREG_BYTE (operand), mode) < 0
1607 /* Don't reload subreg for matching reload. It is actually
1608 valid subreg in LRA. */
1609 && ! LRA_SUBREG_P (operand))
1610 || CONSTANT_P (reg) || GET_CODE (reg) == PLUS || MEM_P (reg))
1612 enum reg_class rclass;
1614 if (REG_P (reg))
1615 /* There is a big probability that we will get the same class
1616 for the new pseudo and we will get the same insn which
1617 means infinite looping. So spill the new pseudo. */
1618 rclass = NO_REGS;
1619 else
1620 /* The class will be defined later in curr_insn_transform. */
1621 rclass
1622 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1624 if (get_reload_reg (curr_static_id->operand[nop].type, reg_mode, reg,
1625 rclass, TRUE, "subreg reg", &new_reg))
1627 bool insert_before, insert_after;
1628 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1630 insert_before = (type != OP_OUT
1631 || GET_MODE_SIZE (innermode) > GET_MODE_SIZE (mode));
1632 insert_after = (type != OP_IN);
1633 insert_move_for_subreg (insert_before ? &before : NULL,
1634 insert_after ? &after : NULL,
1635 reg, new_reg);
1637 SUBREG_REG (operand) = new_reg;
1638 lra_process_new_insns (curr_insn, before, after,
1639 "Inserting subreg reload");
1640 return true;
1642 /* Force a reload for a paradoxical subreg. For paradoxical subreg,
1643 IRA allocates hardreg to the inner pseudo reg according to its mode
1644 instead of the outermode, so the size of the hardreg may not be enough
1645 to contain the outermode operand, in that case we may need to insert
1646 reload for the reg. For the following two types of paradoxical subreg,
1647 we need to insert reload:
1648 1. If the op_type is OP_IN, and the hardreg could not be paired with
1649 other hardreg to contain the outermode operand
1650 (checked by in_hard_reg_set_p), we need to insert the reload.
1651 2. If the op_type is OP_OUT or OP_INOUT.
1653 Here is a paradoxical subreg example showing how the reload is generated:
1655 (insn 5 4 7 2 (set (reg:TI 106 [ __comp ])
1656 (subreg:TI (reg:DI 107 [ __comp ]) 0)) {*movti_internal_rex64}
1658 In IRA, reg107 is allocated to a DImode hardreg. We use x86-64 as example
1659 here, if reg107 is assigned to hardreg R15, because R15 is the last
1660 hardreg, compiler cannot find another hardreg to pair with R15 to
1661 contain TImode data. So we insert a TImode reload reg180 for it.
1662 After reload is inserted:
1664 (insn 283 0 0 (set (subreg:DI (reg:TI 180 [orig:107 __comp ] [107]) 0)
1665 (reg:DI 107 [ __comp ])) -1
1666 (insn 5 4 7 2 (set (reg:TI 106 [ __comp ])
1667 (subreg:TI (reg:TI 180 [orig:107 __comp ] [107]) 0)) {*movti_internal_rex64}
1669 Two reload hard registers will be allocated to reg180 to save TImode data
1670 in LRA_assign. */
1671 else if (REG_P (reg)
1672 && REGNO (reg) >= FIRST_PSEUDO_REGISTER
1673 && (hard_regno = lra_get_regno_hard_regno (REGNO (reg))) >= 0
1674 && (hard_regno_nregs[hard_regno][innermode]
1675 < hard_regno_nregs[hard_regno][mode])
1676 && (regclass = lra_get_allocno_class (REGNO (reg)))
1677 && (type != OP_IN
1678 || !in_hard_reg_set_p (reg_class_contents[regclass],
1679 mode, hard_regno)))
1681 /* The class will be defined later in curr_insn_transform. */
1682 enum reg_class rclass
1683 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1685 if (get_reload_reg (curr_static_id->operand[nop].type, mode, reg,
1686 rclass, TRUE, "paradoxical subreg", &new_reg))
1688 rtx subreg;
1689 bool insert_before, insert_after;
1691 PUT_MODE (new_reg, mode);
1692 subreg = gen_lowpart_SUBREG (innermode, new_reg);
1693 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1695 insert_before = (type != OP_OUT);
1696 insert_after = (type != OP_IN);
1697 insert_move_for_subreg (insert_before ? &before : NULL,
1698 insert_after ? &after : NULL,
1699 reg, subreg);
1701 SUBREG_REG (operand) = new_reg;
1702 lra_process_new_insns (curr_insn, before, after,
1703 "Inserting paradoxical subreg reload");
1704 return true;
1706 return false;
1709 /* Return TRUE if X refers for a hard register from SET. */
1710 static bool
1711 uses_hard_regs_p (rtx x, HARD_REG_SET set)
1713 int i, j, x_hard_regno;
1714 machine_mode mode;
1715 const char *fmt;
1716 enum rtx_code code;
1718 if (x == NULL_RTX)
1719 return false;
1720 code = GET_CODE (x);
1721 mode = GET_MODE (x);
1722 if (code == SUBREG)
1724 x = SUBREG_REG (x);
1725 code = GET_CODE (x);
1726 if (GET_MODE_SIZE (GET_MODE (x)) > GET_MODE_SIZE (mode))
1727 mode = GET_MODE (x);
1730 if (REG_P (x))
1732 x_hard_regno = get_hard_regno (x, true);
1733 return (x_hard_regno >= 0
1734 && overlaps_hard_reg_set_p (set, mode, x_hard_regno));
1736 if (MEM_P (x))
1738 struct address_info ad;
1740 decompose_mem_address (&ad, x);
1741 if (ad.base_term != NULL && uses_hard_regs_p (*ad.base_term, set))
1742 return true;
1743 if (ad.index_term != NULL && uses_hard_regs_p (*ad.index_term, set))
1744 return true;
1746 fmt = GET_RTX_FORMAT (code);
1747 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1749 if (fmt[i] == 'e')
1751 if (uses_hard_regs_p (XEXP (x, i), set))
1752 return true;
1754 else if (fmt[i] == 'E')
1756 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
1757 if (uses_hard_regs_p (XVECEXP (x, i, j), set))
1758 return true;
1761 return false;
1764 /* Return true if OP is a spilled pseudo. */
1765 static inline bool
1766 spilled_pseudo_p (rtx op)
1768 return (REG_P (op)
1769 && REGNO (op) >= FIRST_PSEUDO_REGISTER && in_mem_p (REGNO (op)));
1772 /* Return true if X is a general constant. */
1773 static inline bool
1774 general_constant_p (rtx x)
1776 return CONSTANT_P (x) && (! flag_pic || LEGITIMATE_PIC_OPERAND_P (x));
1779 static bool
1780 reg_in_class_p (rtx reg, enum reg_class cl)
1782 if (cl == NO_REGS)
1783 return get_reg_class (REGNO (reg)) == NO_REGS;
1784 return in_class_p (reg, cl, NULL);
1787 /* Return true if SET of RCLASS contains no hard regs which can be
1788 used in MODE. */
1789 static bool
1790 prohibited_class_reg_set_mode_p (enum reg_class rclass,
1791 HARD_REG_SET &set,
1792 enum machine_mode mode)
1794 HARD_REG_SET temp;
1796 lra_assert (hard_reg_set_subset_p (reg_class_contents[rclass], set));
1797 COPY_HARD_REG_SET (temp, set);
1798 AND_COMPL_HARD_REG_SET (temp, lra_no_alloc_regs);
1799 return (hard_reg_set_subset_p
1800 (temp, ira_prohibited_class_mode_regs[rclass][mode]));
1803 /* Major function to choose the current insn alternative and what
1804 operands should be reloaded and how. If ONLY_ALTERNATIVE is not
1805 negative we should consider only this alternative. Return false if
1806 we can not choose the alternative or find how to reload the
1807 operands. */
1808 static bool
1809 process_alt_operands (int only_alternative)
1811 bool ok_p = false;
1812 int nop, overall, nalt;
1813 int n_alternatives = curr_static_id->n_alternatives;
1814 int n_operands = curr_static_id->n_operands;
1815 /* LOSERS counts the operands that don't fit this alternative and
1816 would require loading. */
1817 int losers;
1818 /* REJECT is a count of how undesirable this alternative says it is
1819 if any reloading is required. If the alternative matches exactly
1820 then REJECT is ignored, but otherwise it gets this much counted
1821 against it in addition to the reloading needed. */
1822 int reject;
1823 int op_reject;
1824 /* The number of elements in the following array. */
1825 int early_clobbered_regs_num;
1826 /* Numbers of operands which are early clobber registers. */
1827 int early_clobbered_nops[MAX_RECOG_OPERANDS];
1828 enum reg_class curr_alt[MAX_RECOG_OPERANDS];
1829 HARD_REG_SET curr_alt_set[MAX_RECOG_OPERANDS];
1830 bool curr_alt_match_win[MAX_RECOG_OPERANDS];
1831 bool curr_alt_win[MAX_RECOG_OPERANDS];
1832 bool curr_alt_offmemok[MAX_RECOG_OPERANDS];
1833 int curr_alt_matches[MAX_RECOG_OPERANDS];
1834 /* The number of elements in the following array. */
1835 int curr_alt_dont_inherit_ops_num;
1836 /* Numbers of operands whose reload pseudos should not be inherited. */
1837 int curr_alt_dont_inherit_ops[MAX_RECOG_OPERANDS];
1838 rtx op;
1839 /* The register when the operand is a subreg of register, otherwise the
1840 operand itself. */
1841 rtx no_subreg_reg_operand[MAX_RECOG_OPERANDS];
1842 /* The register if the operand is a register or subreg of register,
1843 otherwise NULL. */
1844 rtx operand_reg[MAX_RECOG_OPERANDS];
1845 int hard_regno[MAX_RECOG_OPERANDS];
1846 machine_mode biggest_mode[MAX_RECOG_OPERANDS];
1847 int reload_nregs, reload_sum;
1848 bool costly_p;
1849 enum reg_class cl;
1851 /* Calculate some data common for all alternatives to speed up the
1852 function. */
1853 for (nop = 0; nop < n_operands; nop++)
1855 rtx reg;
1857 op = no_subreg_reg_operand[nop] = *curr_id->operand_loc[nop];
1858 /* The real hard regno of the operand after the allocation. */
1859 hard_regno[nop] = get_hard_regno (op, true);
1861 operand_reg[nop] = reg = op;
1862 biggest_mode[nop] = GET_MODE (op);
1863 if (GET_CODE (op) == SUBREG)
1865 operand_reg[nop] = reg = SUBREG_REG (op);
1866 if (GET_MODE_SIZE (biggest_mode[nop])
1867 < GET_MODE_SIZE (GET_MODE (reg)))
1868 biggest_mode[nop] = GET_MODE (reg);
1870 if (! REG_P (reg))
1871 operand_reg[nop] = NULL_RTX;
1872 else if (REGNO (reg) >= FIRST_PSEUDO_REGISTER
1873 || ((int) REGNO (reg)
1874 == lra_get_elimination_hard_regno (REGNO (reg))))
1875 no_subreg_reg_operand[nop] = reg;
1876 else
1877 operand_reg[nop] = no_subreg_reg_operand[nop]
1878 /* Just use natural mode for elimination result. It should
1879 be enough for extra constraints hooks. */
1880 = regno_reg_rtx[hard_regno[nop]];
1883 /* The constraints are made of several alternatives. Each operand's
1884 constraint looks like foo,bar,... with commas separating the
1885 alternatives. The first alternatives for all operands go
1886 together, the second alternatives go together, etc.
1888 First loop over alternatives. */
1889 alternative_mask preferred = curr_id->preferred_alternatives;
1890 if (only_alternative >= 0)
1891 preferred &= ALTERNATIVE_BIT (only_alternative);
1893 for (nalt = 0; nalt < n_alternatives; nalt++)
1895 /* Loop over operands for one constraint alternative. */
1896 if (!TEST_BIT (preferred, nalt))
1897 continue;
1899 overall = losers = reject = reload_nregs = reload_sum = 0;
1900 for (nop = 0; nop < n_operands; nop++)
1902 int inc = (curr_static_id
1903 ->operand_alternative[nalt * n_operands + nop].reject);
1904 if (lra_dump_file != NULL && inc != 0)
1905 fprintf (lra_dump_file,
1906 " Staticly defined alt reject+=%d\n", inc);
1907 reject += inc;
1909 early_clobbered_regs_num = 0;
1911 for (nop = 0; nop < n_operands; nop++)
1913 const char *p;
1914 char *end;
1915 int len, c, m, i, opalt_num, this_alternative_matches;
1916 bool win, did_match, offmemok, early_clobber_p;
1917 /* false => this operand can be reloaded somehow for this
1918 alternative. */
1919 bool badop;
1920 /* true => this operand can be reloaded if the alternative
1921 allows regs. */
1922 bool winreg;
1923 /* True if a constant forced into memory would be OK for
1924 this operand. */
1925 bool constmemok;
1926 enum reg_class this_alternative, this_costly_alternative;
1927 HARD_REG_SET this_alternative_set, this_costly_alternative_set;
1928 bool this_alternative_match_win, this_alternative_win;
1929 bool this_alternative_offmemok;
1930 bool scratch_p;
1931 machine_mode mode;
1932 enum constraint_num cn;
1934 opalt_num = nalt * n_operands + nop;
1935 if (curr_static_id->operand_alternative[opalt_num].anything_ok)
1937 /* Fast track for no constraints at all. */
1938 curr_alt[nop] = NO_REGS;
1939 CLEAR_HARD_REG_SET (curr_alt_set[nop]);
1940 curr_alt_win[nop] = true;
1941 curr_alt_match_win[nop] = false;
1942 curr_alt_offmemok[nop] = false;
1943 curr_alt_matches[nop] = -1;
1944 continue;
1947 op = no_subreg_reg_operand[nop];
1948 mode = curr_operand_mode[nop];
1950 win = did_match = winreg = offmemok = constmemok = false;
1951 badop = true;
1953 early_clobber_p = false;
1954 p = curr_static_id->operand_alternative[opalt_num].constraint;
1956 this_costly_alternative = this_alternative = NO_REGS;
1957 /* We update set of possible hard regs besides its class
1958 because reg class might be inaccurate. For example,
1959 union of LO_REGS (l), HI_REGS(h), and STACK_REG(k) in ARM
1960 is translated in HI_REGS because classes are merged by
1961 pairs and there is no accurate intermediate class. */
1962 CLEAR_HARD_REG_SET (this_alternative_set);
1963 CLEAR_HARD_REG_SET (this_costly_alternative_set);
1964 this_alternative_win = false;
1965 this_alternative_match_win = false;
1966 this_alternative_offmemok = false;
1967 this_alternative_matches = -1;
1969 /* An empty constraint should be excluded by the fast
1970 track. */
1971 lra_assert (*p != 0 && *p != ',');
1973 op_reject = 0;
1974 /* Scan this alternative's specs for this operand; set WIN
1975 if the operand fits any letter in this alternative.
1976 Otherwise, clear BADOP if this operand could fit some
1977 letter after reloads, or set WINREG if this operand could
1978 fit after reloads provided the constraint allows some
1979 registers. */
1980 costly_p = false;
1983 switch ((c = *p, len = CONSTRAINT_LEN (c, p)), c)
1985 case '\0':
1986 len = 0;
1987 break;
1988 case ',':
1989 c = '\0';
1990 break;
1992 case '&':
1993 early_clobber_p = true;
1994 break;
1996 case '$':
1997 op_reject += LRA_MAX_REJECT;
1998 break;
1999 case '^':
2000 op_reject += LRA_LOSER_COST_FACTOR;
2001 break;
2003 case '#':
2004 /* Ignore rest of this alternative. */
2005 c = '\0';
2006 break;
2008 case '0': case '1': case '2': case '3': case '4':
2009 case '5': case '6': case '7': case '8': case '9':
2011 int m_hregno;
2012 bool match_p;
2014 m = strtoul (p, &end, 10);
2015 p = end;
2016 len = 0;
2017 lra_assert (nop > m);
2019 this_alternative_matches = m;
2020 m_hregno = get_hard_regno (*curr_id->operand_loc[m], false);
2021 /* We are supposed to match a previous operand.
2022 If we do, we win if that one did. If we do
2023 not, count both of the operands as losers.
2024 (This is too conservative, since most of the
2025 time only a single reload insn will be needed
2026 to make the two operands win. As a result,
2027 this alternative may be rejected when it is
2028 actually desirable.) */
2029 match_p = false;
2030 if (operands_match_p (*curr_id->operand_loc[nop],
2031 *curr_id->operand_loc[m], m_hregno))
2033 /* We should reject matching of an early
2034 clobber operand if the matching operand is
2035 not dying in the insn. */
2036 if (! curr_static_id->operand[m].early_clobber
2037 || operand_reg[nop] == NULL_RTX
2038 || (find_regno_note (curr_insn, REG_DEAD,
2039 REGNO (op))
2040 || REGNO (op) == REGNO (operand_reg[m])))
2041 match_p = true;
2043 if (match_p)
2045 /* If we are matching a non-offsettable
2046 address where an offsettable address was
2047 expected, then we must reject this
2048 combination, because we can't reload
2049 it. */
2050 if (curr_alt_offmemok[m]
2051 && MEM_P (*curr_id->operand_loc[m])
2052 && curr_alt[m] == NO_REGS && ! curr_alt_win[m])
2053 continue;
2055 else
2057 /* Operands don't match. Both operands must
2058 allow a reload register, otherwise we
2059 cannot make them match. */
2060 if (curr_alt[m] == NO_REGS)
2061 break;
2062 /* Retroactively mark the operand we had to
2063 match as a loser, if it wasn't already and
2064 it wasn't matched to a register constraint
2065 (e.g it might be matched by memory). */
2066 if (curr_alt_win[m]
2067 && (operand_reg[m] == NULL_RTX
2068 || hard_regno[m] < 0))
2070 losers++;
2071 reload_nregs
2072 += (ira_reg_class_max_nregs[curr_alt[m]]
2073 [GET_MODE (*curr_id->operand_loc[m])]);
2076 /* Prefer matching earlyclobber alternative as
2077 it results in less hard regs required for
2078 the insn than a non-matching earlyclobber
2079 alternative. */
2080 if (curr_static_id->operand[m].early_clobber)
2082 if (lra_dump_file != NULL)
2083 fprintf
2084 (lra_dump_file,
2085 " %d Matching earlyclobber alt:"
2086 " reject--\n",
2087 nop);
2088 reject--;
2090 /* Otherwise we prefer no matching
2091 alternatives because it gives more freedom
2092 in RA. */
2093 else if (operand_reg[nop] == NULL_RTX
2094 || (find_regno_note (curr_insn, REG_DEAD,
2095 REGNO (operand_reg[nop]))
2096 == NULL_RTX))
2098 if (lra_dump_file != NULL)
2099 fprintf
2100 (lra_dump_file,
2101 " %d Matching alt: reject+=2\n",
2102 nop);
2103 reject += 2;
2106 /* If we have to reload this operand and some
2107 previous operand also had to match the same
2108 thing as this operand, we don't know how to do
2109 that. */
2110 if (!match_p || !curr_alt_win[m])
2112 for (i = 0; i < nop; i++)
2113 if (curr_alt_matches[i] == m)
2114 break;
2115 if (i < nop)
2116 break;
2118 else
2119 did_match = true;
2121 /* This can be fixed with reloads if the operand
2122 we are supposed to match can be fixed with
2123 reloads. */
2124 badop = false;
2125 this_alternative = curr_alt[m];
2126 COPY_HARD_REG_SET (this_alternative_set, curr_alt_set[m]);
2127 winreg = this_alternative != NO_REGS;
2128 break;
2131 case 'g':
2132 if (MEM_P (op)
2133 || general_constant_p (op)
2134 || spilled_pseudo_p (op))
2135 win = true;
2136 cl = GENERAL_REGS;
2137 goto reg;
2139 default:
2140 cn = lookup_constraint (p);
2141 switch (get_constraint_type (cn))
2143 case CT_REGISTER:
2144 cl = reg_class_for_constraint (cn);
2145 if (cl != NO_REGS)
2146 goto reg;
2147 break;
2149 case CT_CONST_INT:
2150 if (CONST_INT_P (op)
2151 && insn_const_int_ok_for_constraint (INTVAL (op), cn))
2152 win = true;
2153 break;
2155 case CT_MEMORY:
2156 if (MEM_P (op)
2157 && satisfies_memory_constraint_p (op, cn))
2158 win = true;
2159 else if (spilled_pseudo_p (op))
2160 win = true;
2162 /* If we didn't already win, we can reload constants
2163 via force_const_mem or put the pseudo value into
2164 memory, or make other memory by reloading the
2165 address like for 'o'. */
2166 if (CONST_POOL_OK_P (mode, op)
2167 || MEM_P (op) || REG_P (op)
2168 /* We can restore the equiv insn by a
2169 reload. */
2170 || equiv_substition_p[nop])
2171 badop = false;
2172 constmemok = true;
2173 offmemok = true;
2174 break;
2176 case CT_ADDRESS:
2177 /* If we didn't already win, we can reload the address
2178 into a base register. */
2179 if (satisfies_address_constraint_p (op, cn))
2180 win = true;
2181 cl = base_reg_class (VOIDmode, ADDR_SPACE_GENERIC,
2182 ADDRESS, SCRATCH);
2183 badop = false;
2184 goto reg;
2186 case CT_FIXED_FORM:
2187 if (constraint_satisfied_p (op, cn))
2188 win = true;
2189 break;
2191 case CT_SPECIAL_MEMORY:
2192 if (MEM_P (op)
2193 && satisfies_memory_constraint_p (op, cn))
2194 win = true;
2195 else if (spilled_pseudo_p (op))
2196 win = true;
2197 break;
2199 break;
2201 reg:
2202 this_alternative = reg_class_subunion[this_alternative][cl];
2203 IOR_HARD_REG_SET (this_alternative_set,
2204 reg_class_contents[cl]);
2205 if (costly_p)
2207 this_costly_alternative
2208 = reg_class_subunion[this_costly_alternative][cl];
2209 IOR_HARD_REG_SET (this_costly_alternative_set,
2210 reg_class_contents[cl]);
2212 if (mode == BLKmode)
2213 break;
2214 winreg = true;
2215 if (REG_P (op))
2217 if (hard_regno[nop] >= 0
2218 && in_hard_reg_set_p (this_alternative_set,
2219 mode, hard_regno[nop]))
2220 win = true;
2221 else if (hard_regno[nop] < 0
2222 && in_class_p (op, this_alternative, NULL))
2223 win = true;
2225 break;
2227 if (c != ' ' && c != '\t')
2228 costly_p = c == '*';
2230 while ((p += len), c);
2232 scratch_p = (operand_reg[nop] != NULL_RTX
2233 && lra_former_scratch_p (REGNO (operand_reg[nop])));
2234 /* Record which operands fit this alternative. */
2235 if (win)
2237 this_alternative_win = true;
2238 if (operand_reg[nop] != NULL_RTX)
2240 if (hard_regno[nop] >= 0)
2242 if (in_hard_reg_set_p (this_costly_alternative_set,
2243 mode, hard_regno[nop]))
2245 if (lra_dump_file != NULL)
2246 fprintf (lra_dump_file,
2247 " %d Costly set: reject++\n",
2248 nop);
2249 reject++;
2252 else
2254 /* Prefer won reg to spilled pseudo under other
2255 equal conditions for possibe inheritance. */
2256 if (! scratch_p)
2258 if (lra_dump_file != NULL)
2259 fprintf
2260 (lra_dump_file,
2261 " %d Non pseudo reload: reject++\n",
2262 nop);
2263 reject++;
2265 if (in_class_p (operand_reg[nop],
2266 this_costly_alternative, NULL))
2268 if (lra_dump_file != NULL)
2269 fprintf
2270 (lra_dump_file,
2271 " %d Non pseudo costly reload:"
2272 " reject++\n",
2273 nop);
2274 reject++;
2277 /* We simulate the behavior of old reload here.
2278 Although scratches need hard registers and it
2279 might result in spilling other pseudos, no reload
2280 insns are generated for the scratches. So it
2281 might cost something but probably less than old
2282 reload pass believes. */
2283 if (scratch_p)
2285 if (lra_dump_file != NULL)
2286 fprintf (lra_dump_file,
2287 " %d Scratch win: reject+=2\n",
2288 nop);
2289 reject += 2;
2293 else if (did_match)
2294 this_alternative_match_win = true;
2295 else
2297 int const_to_mem = 0;
2298 bool no_regs_p;
2300 reject += op_reject;
2301 /* Never do output reload of stack pointer. It makes
2302 impossible to do elimination when SP is changed in
2303 RTL. */
2304 if (op == stack_pointer_rtx && ! frame_pointer_needed
2305 && curr_static_id->operand[nop].type != OP_IN)
2306 goto fail;
2308 /* If this alternative asks for a specific reg class, see if there
2309 is at least one allocatable register in that class. */
2310 no_regs_p
2311 = (this_alternative == NO_REGS
2312 || (hard_reg_set_subset_p
2313 (reg_class_contents[this_alternative],
2314 lra_no_alloc_regs)));
2316 /* For asms, verify that the class for this alternative is possible
2317 for the mode that is specified. */
2318 if (!no_regs_p && INSN_CODE (curr_insn) < 0)
2320 int i;
2321 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
2322 if (HARD_REGNO_MODE_OK (i, mode)
2323 && in_hard_reg_set_p (reg_class_contents[this_alternative],
2324 mode, i))
2325 break;
2326 if (i == FIRST_PSEUDO_REGISTER)
2327 winreg = false;
2330 /* If this operand accepts a register, and if the
2331 register class has at least one allocatable register,
2332 then this operand can be reloaded. */
2333 if (winreg && !no_regs_p)
2334 badop = false;
2336 if (badop)
2338 if (lra_dump_file != NULL)
2339 fprintf (lra_dump_file,
2340 " alt=%d: Bad operand -- refuse\n",
2341 nalt);
2342 goto fail;
2345 if (this_alternative != NO_REGS)
2347 HARD_REG_SET available_regs;
2349 COPY_HARD_REG_SET (available_regs,
2350 reg_class_contents[this_alternative]);
2351 AND_COMPL_HARD_REG_SET
2352 (available_regs,
2353 ira_prohibited_class_mode_regs[this_alternative][mode]);
2354 AND_COMPL_HARD_REG_SET (available_regs, lra_no_alloc_regs);
2355 if (hard_reg_set_empty_p (available_regs))
2357 /* There are no hard regs holding a value of given
2358 mode. */
2359 if (offmemok)
2361 this_alternative = NO_REGS;
2362 if (lra_dump_file != NULL)
2363 fprintf (lra_dump_file,
2364 " %d Using memory because of"
2365 " a bad mode: reject+=2\n",
2366 nop);
2367 reject += 2;
2369 else
2371 if (lra_dump_file != NULL)
2372 fprintf (lra_dump_file,
2373 " alt=%d: Wrong mode -- refuse\n",
2374 nalt);
2375 goto fail;
2380 /* If not assigned pseudo has a class which a subset of
2381 required reg class, it is a less costly alternative
2382 as the pseudo still can get a hard reg of necessary
2383 class. */
2384 if (! no_regs_p && REG_P (op) && hard_regno[nop] < 0
2385 && (cl = get_reg_class (REGNO (op))) != NO_REGS
2386 && ira_class_subset_p[this_alternative][cl])
2388 if (lra_dump_file != NULL)
2389 fprintf
2390 (lra_dump_file,
2391 " %d Super set class reg: reject-=3\n", nop);
2392 reject -= 3;
2395 this_alternative_offmemok = offmemok;
2396 if (this_costly_alternative != NO_REGS)
2398 if (lra_dump_file != NULL)
2399 fprintf (lra_dump_file,
2400 " %d Costly loser: reject++\n", nop);
2401 reject++;
2403 /* If the operand is dying, has a matching constraint,
2404 and satisfies constraints of the matched operand
2405 which failed to satisfy the own constraints, most probably
2406 the reload for this operand will be gone. */
2407 if (this_alternative_matches >= 0
2408 && !curr_alt_win[this_alternative_matches]
2409 && REG_P (op)
2410 && find_regno_note (curr_insn, REG_DEAD, REGNO (op))
2411 && (hard_regno[nop] >= 0
2412 ? in_hard_reg_set_p (this_alternative_set,
2413 mode, hard_regno[nop])
2414 : in_class_p (op, this_alternative, NULL)))
2416 if (lra_dump_file != NULL)
2417 fprintf
2418 (lra_dump_file,
2419 " %d Dying matched operand reload: reject++\n",
2420 nop);
2421 reject++;
2423 else
2425 /* Strict_low_part requires to reload the register
2426 not the sub-register. In this case we should
2427 check that a final reload hard reg can hold the
2428 value mode. */
2429 if (curr_static_id->operand[nop].strict_low
2430 && REG_P (op)
2431 && hard_regno[nop] < 0
2432 && GET_CODE (*curr_id->operand_loc[nop]) == SUBREG
2433 && ira_class_hard_regs_num[this_alternative] > 0
2434 && ! HARD_REGNO_MODE_OK (ira_class_hard_regs
2435 [this_alternative][0],
2436 GET_MODE
2437 (*curr_id->operand_loc[nop])))
2439 if (lra_dump_file != NULL)
2440 fprintf
2441 (lra_dump_file,
2442 " alt=%d: Strict low subreg reload -- refuse\n",
2443 nalt);
2444 goto fail;
2446 losers++;
2448 if (operand_reg[nop] != NULL_RTX
2449 /* Output operands and matched input operands are
2450 not inherited. The following conditions do not
2451 exactly describe the previous statement but they
2452 are pretty close. */
2453 && curr_static_id->operand[nop].type != OP_OUT
2454 && (this_alternative_matches < 0
2455 || curr_static_id->operand[nop].type != OP_IN))
2457 int last_reload = (lra_reg_info[ORIGINAL_REGNO
2458 (operand_reg[nop])]
2459 .last_reload);
2461 /* The value of reload_sum has sense only if we
2462 process insns in their order. It happens only on
2463 the first constraints sub-pass when we do most of
2464 reload work. */
2465 if (lra_constraint_iter == 1 && last_reload > bb_reload_num)
2466 reload_sum += last_reload - bb_reload_num;
2468 /* If this is a constant that is reloaded into the
2469 desired class by copying it to memory first, count
2470 that as another reload. This is consistent with
2471 other code and is required to avoid choosing another
2472 alternative when the constant is moved into memory.
2473 Note that the test here is precisely the same as in
2474 the code below that calls force_const_mem. */
2475 if (CONST_POOL_OK_P (mode, op)
2476 && ((targetm.preferred_reload_class
2477 (op, this_alternative) == NO_REGS)
2478 || no_input_reloads_p))
2480 const_to_mem = 1;
2481 if (! no_regs_p)
2482 losers++;
2485 /* Alternative loses if it requires a type of reload not
2486 permitted for this insn. We can always reload
2487 objects with a REG_UNUSED note. */
2488 if ((curr_static_id->operand[nop].type != OP_IN
2489 && no_output_reloads_p
2490 && ! find_reg_note (curr_insn, REG_UNUSED, op))
2491 || (curr_static_id->operand[nop].type != OP_OUT
2492 && no_input_reloads_p && ! const_to_mem)
2493 || (this_alternative_matches >= 0
2494 && (no_input_reloads_p
2495 || (no_output_reloads_p
2496 && (curr_static_id->operand
2497 [this_alternative_matches].type != OP_IN)
2498 && ! find_reg_note (curr_insn, REG_UNUSED,
2499 no_subreg_reg_operand
2500 [this_alternative_matches])))))
2502 if (lra_dump_file != NULL)
2503 fprintf
2504 (lra_dump_file,
2505 " alt=%d: No input/otput reload -- refuse\n",
2506 nalt);
2507 goto fail;
2510 /* Alternative loses if it required class pseudo can not
2511 hold value of required mode. Such insns can be
2512 described by insn definitions with mode iterators. */
2513 if (GET_MODE (*curr_id->operand_loc[nop]) != VOIDmode
2514 && ! hard_reg_set_empty_p (this_alternative_set)
2515 /* It is common practice for constraints to use a
2516 class which does not have actually enough regs to
2517 hold the value (e.g. x86 AREG for mode requiring
2518 more one general reg). Therefore we have 2
2519 conditions to check that the reload pseudo can
2520 not hold the mode value. */
2521 && ! HARD_REGNO_MODE_OK (ira_class_hard_regs
2522 [this_alternative][0],
2523 GET_MODE (*curr_id->operand_loc[nop]))
2524 /* The above condition is not enough as the first
2525 reg in ira_class_hard_regs can be not aligned for
2526 multi-words mode values. */
2527 && (prohibited_class_reg_set_mode_p
2528 (this_alternative, this_alternative_set,
2529 GET_MODE (*curr_id->operand_loc[nop]))))
2531 if (lra_dump_file != NULL)
2532 fprintf (lra_dump_file,
2533 " alt=%d: reload pseudo for op %d "
2534 " can not hold the mode value -- refuse\n",
2535 nalt, nop);
2536 goto fail;
2539 /* Check strong discouragement of reload of non-constant
2540 into class THIS_ALTERNATIVE. */
2541 if (! CONSTANT_P (op) && ! no_regs_p
2542 && (targetm.preferred_reload_class
2543 (op, this_alternative) == NO_REGS
2544 || (curr_static_id->operand[nop].type == OP_OUT
2545 && (targetm.preferred_output_reload_class
2546 (op, this_alternative) == NO_REGS))))
2548 if (lra_dump_file != NULL)
2549 fprintf (lra_dump_file,
2550 " %d Non-prefered reload: reject+=%d\n",
2551 nop, LRA_MAX_REJECT);
2552 reject += LRA_MAX_REJECT;
2555 if (! (MEM_P (op) && offmemok)
2556 && ! (const_to_mem && constmemok))
2558 /* We prefer to reload pseudos over reloading other
2559 things, since such reloads may be able to be
2560 eliminated later. So bump REJECT in other cases.
2561 Don't do this in the case where we are forcing a
2562 constant into memory and it will then win since
2563 we don't want to have a different alternative
2564 match then. */
2565 if (! (REG_P (op) && REGNO (op) >= FIRST_PSEUDO_REGISTER))
2567 if (lra_dump_file != NULL)
2568 fprintf
2569 (lra_dump_file,
2570 " %d Non-pseudo reload: reject+=2\n",
2571 nop);
2572 reject += 2;
2575 if (! no_regs_p)
2576 reload_nregs
2577 += ira_reg_class_max_nregs[this_alternative][mode];
2579 if (SMALL_REGISTER_CLASS_P (this_alternative))
2581 if (lra_dump_file != NULL)
2582 fprintf
2583 (lra_dump_file,
2584 " %d Small class reload: reject+=%d\n",
2585 nop, LRA_LOSER_COST_FACTOR / 2);
2586 reject += LRA_LOSER_COST_FACTOR / 2;
2590 /* We are trying to spill pseudo into memory. It is
2591 usually more costly than moving to a hard register
2592 although it might takes the same number of
2593 reloads.
2595 Non-pseudo spill may happen also. Suppose a target allows both
2596 register and memory in the operand constraint alternatives,
2597 then it's typical that an eliminable register has a substition
2598 of "base + offset" which can either be reloaded by a simple
2599 "new_reg <= base + offset" which will match the register
2600 constraint, or a similar reg addition followed by further spill
2601 to and reload from memory which will match the memory
2602 constraint, but this memory spill will be much more costly
2603 usually.
2605 Code below increases the reject for both pseudo and non-pseudo
2606 spill. */
2607 if (no_regs_p
2608 && !(MEM_P (op) && offmemok)
2609 && !(REG_P (op) && hard_regno[nop] < 0))
2611 if (lra_dump_file != NULL)
2612 fprintf
2613 (lra_dump_file,
2614 " %d Spill %spseudo into memory: reject+=3\n",
2615 nop, REG_P (op) ? "" : "Non-");
2616 reject += 3;
2617 if (VECTOR_MODE_P (mode))
2619 /* Spilling vectors into memory is usually more
2620 costly as they contain big values. */
2621 if (lra_dump_file != NULL)
2622 fprintf
2623 (lra_dump_file,
2624 " %d Spill vector pseudo: reject+=2\n",
2625 nop);
2626 reject += 2;
2630 #ifdef SECONDARY_MEMORY_NEEDED
2631 /* If reload requires moving value through secondary
2632 memory, it will need one more insn at least. */
2633 if (this_alternative != NO_REGS
2634 && REG_P (op) && (cl = get_reg_class (REGNO (op))) != NO_REGS
2635 && ((curr_static_id->operand[nop].type != OP_OUT
2636 && SECONDARY_MEMORY_NEEDED (cl, this_alternative,
2637 GET_MODE (op)))
2638 || (curr_static_id->operand[nop].type != OP_IN
2639 && SECONDARY_MEMORY_NEEDED (this_alternative, cl,
2640 GET_MODE (op)))))
2641 losers++;
2642 #endif
2643 /* Input reloads can be inherited more often than output
2644 reloads can be removed, so penalize output
2645 reloads. */
2646 if (!REG_P (op) || curr_static_id->operand[nop].type != OP_IN)
2648 if (lra_dump_file != NULL)
2649 fprintf
2650 (lra_dump_file,
2651 " %d Non input pseudo reload: reject++\n",
2652 nop);
2653 reject++;
2657 if (early_clobber_p && ! scratch_p)
2659 if (lra_dump_file != NULL)
2660 fprintf (lra_dump_file,
2661 " %d Early clobber: reject++\n", nop);
2662 reject++;
2664 /* ??? We check early clobbers after processing all operands
2665 (see loop below) and there we update the costs more.
2666 Should we update the cost (may be approximately) here
2667 because of early clobber register reloads or it is a rare
2668 or non-important thing to be worth to do it. */
2669 overall = losers * LRA_LOSER_COST_FACTOR + reject;
2670 if ((best_losers == 0 || losers != 0) && best_overall < overall)
2672 if (lra_dump_file != NULL)
2673 fprintf (lra_dump_file,
2674 " alt=%d,overall=%d,losers=%d -- refuse\n",
2675 nalt, overall, losers);
2676 goto fail;
2679 curr_alt[nop] = this_alternative;
2680 COPY_HARD_REG_SET (curr_alt_set[nop], this_alternative_set);
2681 curr_alt_win[nop] = this_alternative_win;
2682 curr_alt_match_win[nop] = this_alternative_match_win;
2683 curr_alt_offmemok[nop] = this_alternative_offmemok;
2684 curr_alt_matches[nop] = this_alternative_matches;
2686 if (this_alternative_matches >= 0
2687 && !did_match && !this_alternative_win)
2688 curr_alt_win[this_alternative_matches] = false;
2690 if (early_clobber_p && operand_reg[nop] != NULL_RTX)
2691 early_clobbered_nops[early_clobbered_regs_num++] = nop;
2693 if (curr_insn_set != NULL_RTX && n_operands == 2
2694 /* Prevent processing non-move insns. */
2695 && (GET_CODE (SET_SRC (curr_insn_set)) == SUBREG
2696 || SET_SRC (curr_insn_set) == no_subreg_reg_operand[1])
2697 && ((! curr_alt_win[0] && ! curr_alt_win[1]
2698 && REG_P (no_subreg_reg_operand[0])
2699 && REG_P (no_subreg_reg_operand[1])
2700 && (reg_in_class_p (no_subreg_reg_operand[0], curr_alt[1])
2701 || reg_in_class_p (no_subreg_reg_operand[1], curr_alt[0])))
2702 || (! curr_alt_win[0] && curr_alt_win[1]
2703 && REG_P (no_subreg_reg_operand[1])
2704 && reg_in_class_p (no_subreg_reg_operand[1], curr_alt[0]))
2705 || (curr_alt_win[0] && ! curr_alt_win[1]
2706 && REG_P (no_subreg_reg_operand[0])
2707 && reg_in_class_p (no_subreg_reg_operand[0], curr_alt[1])
2708 && (! CONST_POOL_OK_P (curr_operand_mode[1],
2709 no_subreg_reg_operand[1])
2710 || (targetm.preferred_reload_class
2711 (no_subreg_reg_operand[1],
2712 (enum reg_class) curr_alt[1]) != NO_REGS))
2713 /* If it is a result of recent elimination in move
2714 insn we can transform it into an add still by
2715 using this alternative. */
2716 && GET_CODE (no_subreg_reg_operand[1]) != PLUS)))
2718 /* We have a move insn and a new reload insn will be similar
2719 to the current insn. We should avoid such situation as it
2720 results in LRA cycling. */
2721 overall += LRA_MAX_REJECT;
2723 ok_p = true;
2724 curr_alt_dont_inherit_ops_num = 0;
2725 for (nop = 0; nop < early_clobbered_regs_num; nop++)
2727 int i, j, clobbered_hard_regno, first_conflict_j, last_conflict_j;
2728 HARD_REG_SET temp_set;
2730 i = early_clobbered_nops[nop];
2731 if ((! curr_alt_win[i] && ! curr_alt_match_win[i])
2732 || hard_regno[i] < 0)
2733 continue;
2734 lra_assert (operand_reg[i] != NULL_RTX);
2735 clobbered_hard_regno = hard_regno[i];
2736 CLEAR_HARD_REG_SET (temp_set);
2737 add_to_hard_reg_set (&temp_set, biggest_mode[i], clobbered_hard_regno);
2738 first_conflict_j = last_conflict_j = -1;
2739 for (j = 0; j < n_operands; j++)
2740 if (j == i
2741 /* We don't want process insides of match_operator and
2742 match_parallel because otherwise we would process
2743 their operands once again generating a wrong
2744 code. */
2745 || curr_static_id->operand[j].is_operator)
2746 continue;
2747 else if ((curr_alt_matches[j] == i && curr_alt_match_win[j])
2748 || (curr_alt_matches[i] == j && curr_alt_match_win[i]))
2749 continue;
2750 /* If we don't reload j-th operand, check conflicts. */
2751 else if ((curr_alt_win[j] || curr_alt_match_win[j])
2752 && uses_hard_regs_p (*curr_id->operand_loc[j], temp_set))
2754 if (first_conflict_j < 0)
2755 first_conflict_j = j;
2756 last_conflict_j = j;
2758 if (last_conflict_j < 0)
2759 continue;
2760 /* If earlyclobber operand conflicts with another
2761 non-matching operand which is actually the same register
2762 as the earlyclobber operand, it is better to reload the
2763 another operand as an operand matching the earlyclobber
2764 operand can be also the same. */
2765 if (first_conflict_j == last_conflict_j
2766 && operand_reg[last_conflict_j] != NULL_RTX
2767 && ! curr_alt_match_win[last_conflict_j]
2768 && REGNO (operand_reg[i]) == REGNO (operand_reg[last_conflict_j]))
2770 curr_alt_win[last_conflict_j] = false;
2771 curr_alt_dont_inherit_ops[curr_alt_dont_inherit_ops_num++]
2772 = last_conflict_j;
2773 losers++;
2774 /* Early clobber was already reflected in REJECT. */
2775 lra_assert (reject > 0);
2776 if (lra_dump_file != NULL)
2777 fprintf
2778 (lra_dump_file,
2779 " %d Conflict early clobber reload: reject--\n",
2781 reject--;
2782 overall += LRA_LOSER_COST_FACTOR - 1;
2784 else
2786 /* We need to reload early clobbered register and the
2787 matched registers. */
2788 for (j = 0; j < n_operands; j++)
2789 if (curr_alt_matches[j] == i)
2791 curr_alt_match_win[j] = false;
2792 losers++;
2793 overall += LRA_LOSER_COST_FACTOR;
2795 if (! curr_alt_match_win[i])
2796 curr_alt_dont_inherit_ops[curr_alt_dont_inherit_ops_num++] = i;
2797 else
2799 /* Remember pseudos used for match reloads are never
2800 inherited. */
2801 lra_assert (curr_alt_matches[i] >= 0);
2802 curr_alt_win[curr_alt_matches[i]] = false;
2804 curr_alt_win[i] = curr_alt_match_win[i] = false;
2805 losers++;
2806 /* Early clobber was already reflected in REJECT. */
2807 lra_assert (reject > 0);
2808 if (lra_dump_file != NULL)
2809 fprintf
2810 (lra_dump_file,
2811 " %d Matched conflict early clobber reloads:"
2812 "reject--\n",
2814 reject--;
2815 overall += LRA_LOSER_COST_FACTOR - 1;
2818 if (lra_dump_file != NULL)
2819 fprintf (lra_dump_file, " alt=%d,overall=%d,losers=%d,rld_nregs=%d\n",
2820 nalt, overall, losers, reload_nregs);
2822 /* If this alternative can be made to work by reloading, and it
2823 needs less reloading than the others checked so far, record
2824 it as the chosen goal for reloading. */
2825 if ((best_losers != 0 && losers == 0)
2826 || (((best_losers == 0 && losers == 0)
2827 || (best_losers != 0 && losers != 0))
2828 && (best_overall > overall
2829 || (best_overall == overall
2830 /* If the cost of the reloads is the same,
2831 prefer alternative which requires minimal
2832 number of reload regs. */
2833 && (reload_nregs < best_reload_nregs
2834 || (reload_nregs == best_reload_nregs
2835 && (best_reload_sum < reload_sum
2836 || (best_reload_sum == reload_sum
2837 && nalt < goal_alt_number))))))))
2839 for (nop = 0; nop < n_operands; nop++)
2841 goal_alt_win[nop] = curr_alt_win[nop];
2842 goal_alt_match_win[nop] = curr_alt_match_win[nop];
2843 goal_alt_matches[nop] = curr_alt_matches[nop];
2844 goal_alt[nop] = curr_alt[nop];
2845 goal_alt_offmemok[nop] = curr_alt_offmemok[nop];
2847 goal_alt_dont_inherit_ops_num = curr_alt_dont_inherit_ops_num;
2848 for (nop = 0; nop < curr_alt_dont_inherit_ops_num; nop++)
2849 goal_alt_dont_inherit_ops[nop] = curr_alt_dont_inherit_ops[nop];
2850 goal_alt_swapped = curr_swapped;
2851 best_overall = overall;
2852 best_losers = losers;
2853 best_reload_nregs = reload_nregs;
2854 best_reload_sum = reload_sum;
2855 goal_alt_number = nalt;
2857 if (losers == 0)
2858 /* Everything is satisfied. Do not process alternatives
2859 anymore. */
2860 break;
2861 fail:
2864 return ok_p;
2867 /* Make reload base reg from address AD. */
2868 static rtx
2869 base_to_reg (struct address_info *ad)
2871 enum reg_class cl;
2872 int code = -1;
2873 rtx new_inner = NULL_RTX;
2874 rtx new_reg = NULL_RTX;
2875 rtx_insn *insn;
2876 rtx_insn *last_insn = get_last_insn();
2878 lra_assert (ad->base == ad->base_term && ad->disp == ad->disp_term);
2879 cl = base_reg_class (ad->mode, ad->as, ad->base_outer_code,
2880 get_index_code (ad));
2881 new_reg = lra_create_new_reg (GET_MODE (*ad->base_term), NULL_RTX,
2882 cl, "base");
2883 new_inner = simplify_gen_binary (PLUS, GET_MODE (new_reg), new_reg,
2884 ad->disp_term == NULL
2885 ? gen_int_mode (0, ad->mode)
2886 : *ad->disp_term);
2887 if (!valid_address_p (ad->mode, new_inner, ad->as))
2888 return NULL_RTX;
2889 insn = emit_insn (gen_rtx_SET (new_reg, *ad->base_term));
2890 code = recog_memoized (insn);
2891 if (code < 0)
2893 delete_insns_since (last_insn);
2894 return NULL_RTX;
2897 return new_inner;
2900 /* Make reload base reg + disp from address AD. Return the new pseudo. */
2901 static rtx
2902 base_plus_disp_to_reg (struct address_info *ad)
2904 enum reg_class cl;
2905 rtx new_reg;
2907 lra_assert (ad->base == ad->base_term && ad->disp == ad->disp_term);
2908 cl = base_reg_class (ad->mode, ad->as, ad->base_outer_code,
2909 get_index_code (ad));
2910 new_reg = lra_create_new_reg (GET_MODE (*ad->base_term), NULL_RTX,
2911 cl, "base + disp");
2912 lra_emit_add (new_reg, *ad->base_term, *ad->disp_term);
2913 return new_reg;
2916 /* Make reload of index part of address AD. Return the new
2917 pseudo. */
2918 static rtx
2919 index_part_to_reg (struct address_info *ad)
2921 rtx new_reg;
2923 new_reg = lra_create_new_reg (GET_MODE (*ad->index), NULL_RTX,
2924 INDEX_REG_CLASS, "index term");
2925 expand_mult (GET_MODE (*ad->index), *ad->index_term,
2926 GEN_INT (get_index_scale (ad)), new_reg, 1);
2927 return new_reg;
2930 /* Return true if we can add a displacement to address AD, even if that
2931 makes the address invalid. The fix-up code requires any new address
2932 to be the sum of the BASE_TERM, INDEX and DISP_TERM fields. */
2933 static bool
2934 can_add_disp_p (struct address_info *ad)
2936 return (!ad->autoinc_p
2937 && ad->segment == NULL
2938 && ad->base == ad->base_term
2939 && ad->disp == ad->disp_term);
2942 /* Make equiv substitution in address AD. Return true if a substitution
2943 was made. */
2944 static bool
2945 equiv_address_substitution (struct address_info *ad)
2947 rtx base_reg, new_base_reg, index_reg, new_index_reg, *base_term, *index_term;
2948 HOST_WIDE_INT disp, scale;
2949 bool change_p;
2951 base_term = strip_subreg (ad->base_term);
2952 if (base_term == NULL)
2953 base_reg = new_base_reg = NULL_RTX;
2954 else
2956 base_reg = *base_term;
2957 new_base_reg = get_equiv_with_elimination (base_reg, curr_insn);
2959 index_term = strip_subreg (ad->index_term);
2960 if (index_term == NULL)
2961 index_reg = new_index_reg = NULL_RTX;
2962 else
2964 index_reg = *index_term;
2965 new_index_reg = get_equiv_with_elimination (index_reg, curr_insn);
2967 if (base_reg == new_base_reg && index_reg == new_index_reg)
2968 return false;
2969 disp = 0;
2970 change_p = false;
2971 if (lra_dump_file != NULL)
2973 fprintf (lra_dump_file, "Changing address in insn %d ",
2974 INSN_UID (curr_insn));
2975 dump_value_slim (lra_dump_file, *ad->outer, 1);
2977 if (base_reg != new_base_reg)
2979 if (REG_P (new_base_reg))
2981 *base_term = new_base_reg;
2982 change_p = true;
2984 else if (GET_CODE (new_base_reg) == PLUS
2985 && REG_P (XEXP (new_base_reg, 0))
2986 && CONST_INT_P (XEXP (new_base_reg, 1))
2987 && can_add_disp_p (ad))
2989 disp += INTVAL (XEXP (new_base_reg, 1));
2990 *base_term = XEXP (new_base_reg, 0);
2991 change_p = true;
2993 if (ad->base_term2 != NULL)
2994 *ad->base_term2 = *ad->base_term;
2996 if (index_reg != new_index_reg)
2998 if (REG_P (new_index_reg))
3000 *index_term = new_index_reg;
3001 change_p = true;
3003 else if (GET_CODE (new_index_reg) == PLUS
3004 && REG_P (XEXP (new_index_reg, 0))
3005 && CONST_INT_P (XEXP (new_index_reg, 1))
3006 && can_add_disp_p (ad)
3007 && (scale = get_index_scale (ad)))
3009 disp += INTVAL (XEXP (new_index_reg, 1)) * scale;
3010 *index_term = XEXP (new_index_reg, 0);
3011 change_p = true;
3014 if (disp != 0)
3016 if (ad->disp != NULL)
3017 *ad->disp = plus_constant (GET_MODE (*ad->inner), *ad->disp, disp);
3018 else
3020 *ad->inner = plus_constant (GET_MODE (*ad->inner), *ad->inner, disp);
3021 update_address (ad);
3023 change_p = true;
3025 if (lra_dump_file != NULL)
3027 if (! change_p)
3028 fprintf (lra_dump_file, " -- no change\n");
3029 else
3031 fprintf (lra_dump_file, " on equiv ");
3032 dump_value_slim (lra_dump_file, *ad->outer, 1);
3033 fprintf (lra_dump_file, "\n");
3036 return change_p;
3039 /* Major function to make reloads for an address in operand NOP or
3040 check its correctness (If CHECK_ONLY_P is true). The supported
3041 cases are:
3043 1) an address that existed before LRA started, at which point it
3044 must have been valid. These addresses are subject to elimination
3045 and may have become invalid due to the elimination offset being out
3046 of range.
3048 2) an address created by forcing a constant to memory
3049 (force_const_to_mem). The initial form of these addresses might
3050 not be valid, and it is this function's job to make them valid.
3052 3) a frame address formed from a register and a (possibly zero)
3053 constant offset. As above, these addresses might not be valid and
3054 this function must make them so.
3056 Add reloads to the lists *BEFORE and *AFTER. We might need to add
3057 reloads to *AFTER because of inc/dec, {pre, post} modify in the
3058 address. Return true for any RTL change.
3060 The function is a helper function which does not produce all
3061 transformations (when CHECK_ONLY_P is false) which can be
3062 necessary. It does just basic steps. To do all necessary
3063 transformations use function process_address. */
3064 static bool
3065 process_address_1 (int nop, bool check_only_p,
3066 rtx_insn **before, rtx_insn **after)
3068 struct address_info ad;
3069 rtx new_reg;
3070 HOST_WIDE_INT scale;
3071 rtx op = *curr_id->operand_loc[nop];
3072 const char *constraint = curr_static_id->operand[nop].constraint;
3073 enum constraint_num cn = lookup_constraint (constraint);
3074 bool change_p = false;
3076 if (MEM_P (op)
3077 && GET_MODE (op) == BLKmode
3078 && GET_CODE (XEXP (op, 0)) == SCRATCH)
3079 return false;
3081 if (insn_extra_address_constraint (cn))
3082 decompose_lea_address (&ad, curr_id->operand_loc[nop]);
3083 else if (MEM_P (op))
3084 decompose_mem_address (&ad, op);
3085 else if (GET_CODE (op) == SUBREG
3086 && MEM_P (SUBREG_REG (op)))
3087 decompose_mem_address (&ad, SUBREG_REG (op));
3088 else
3089 return false;
3090 /* If INDEX_REG_CLASS is assigned to base_term already and isn't to
3091 index_term, swap them so to avoid assigning INDEX_REG_CLASS to both
3092 when INDEX_REG_CLASS is a single register class. */
3093 if (ad.base_term != NULL
3094 && ad.index_term != NULL
3095 && ira_class_hard_regs_num[INDEX_REG_CLASS] == 1
3096 && REG_P (*ad.base_term)
3097 && REG_P (*ad.index_term)
3098 && in_class_p (*ad.base_term, INDEX_REG_CLASS, NULL)
3099 && ! in_class_p (*ad.index_term, INDEX_REG_CLASS, NULL))
3101 std::swap (ad.base, ad.index);
3102 std::swap (ad.base_term, ad.index_term);
3104 if (! check_only_p)
3105 change_p = equiv_address_substitution (&ad);
3106 if (ad.base_term != NULL
3107 && (process_addr_reg
3108 (ad.base_term, check_only_p, before,
3109 (ad.autoinc_p
3110 && !(REG_P (*ad.base_term)
3111 && find_regno_note (curr_insn, REG_DEAD,
3112 REGNO (*ad.base_term)) != NULL_RTX)
3113 ? after : NULL),
3114 base_reg_class (ad.mode, ad.as, ad.base_outer_code,
3115 get_index_code (&ad)))))
3117 change_p = true;
3118 if (ad.base_term2 != NULL)
3119 *ad.base_term2 = *ad.base_term;
3121 if (ad.index_term != NULL
3122 && process_addr_reg (ad.index_term, check_only_p,
3123 before, NULL, INDEX_REG_CLASS))
3124 change_p = true;
3126 /* Target hooks sometimes don't treat extra-constraint addresses as
3127 legitimate address_operands, so handle them specially. */
3128 if (insn_extra_address_constraint (cn)
3129 && satisfies_address_constraint_p (&ad, cn))
3130 return change_p;
3132 if (check_only_p)
3133 return change_p;
3135 /* There are three cases where the shape of *AD.INNER may now be invalid:
3137 1) the original address was valid, but either elimination or
3138 equiv_address_substitution was applied and that made
3139 the address invalid.
3141 2) the address is an invalid symbolic address created by
3142 force_const_to_mem.
3144 3) the address is a frame address with an invalid offset.
3146 4) the address is a frame address with an invalid base.
3148 All these cases involve a non-autoinc address, so there is no
3149 point revalidating other types. */
3150 if (ad.autoinc_p || valid_address_p (&ad))
3151 return change_p;
3153 /* Any index existed before LRA started, so we can assume that the
3154 presence and shape of the index is valid. */
3155 push_to_sequence (*before);
3156 lra_assert (ad.disp == ad.disp_term);
3157 if (ad.base == NULL)
3159 if (ad.index == NULL)
3161 rtx_insn *insn;
3162 rtx_insn *last = get_last_insn ();
3163 int code = -1;
3164 enum reg_class cl = base_reg_class (ad.mode, ad.as,
3165 SCRATCH, SCRATCH);
3166 rtx addr = *ad.inner;
3168 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "addr");
3169 if (HAVE_lo_sum)
3171 /* addr => lo_sum (new_base, addr), case (2) above. */
3172 insn = emit_insn (gen_rtx_SET
3173 (new_reg,
3174 gen_rtx_HIGH (Pmode, copy_rtx (addr))));
3175 code = recog_memoized (insn);
3176 if (code >= 0)
3178 *ad.inner = gen_rtx_LO_SUM (Pmode, new_reg, addr);
3179 if (! valid_address_p (ad.mode, *ad.outer, ad.as))
3181 /* Try to put lo_sum into register. */
3182 insn = emit_insn (gen_rtx_SET
3183 (new_reg,
3184 gen_rtx_LO_SUM (Pmode, new_reg, addr)));
3185 code = recog_memoized (insn);
3186 if (code >= 0)
3188 *ad.inner = new_reg;
3189 if (! valid_address_p (ad.mode, *ad.outer, ad.as))
3191 *ad.inner = addr;
3192 code = -1;
3198 if (code < 0)
3199 delete_insns_since (last);
3202 if (code < 0)
3204 /* addr => new_base, case (2) above. */
3205 lra_emit_move (new_reg, addr);
3207 for (insn = last == NULL_RTX ? get_insns () : NEXT_INSN (last);
3208 insn != NULL_RTX;
3209 insn = NEXT_INSN (insn))
3210 if (recog_memoized (insn) < 0)
3211 break;
3212 if (insn != NULL_RTX)
3214 /* Do nothing if we cannot generate right insns.
3215 This is analogous to reload pass behavior. */
3216 delete_insns_since (last);
3217 end_sequence ();
3218 return false;
3220 *ad.inner = new_reg;
3223 else
3225 /* index * scale + disp => new base + index * scale,
3226 case (1) above. */
3227 enum reg_class cl = base_reg_class (ad.mode, ad.as, PLUS,
3228 GET_CODE (*ad.index));
3230 lra_assert (INDEX_REG_CLASS != NO_REGS);
3231 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "disp");
3232 lra_emit_move (new_reg, *ad.disp);
3233 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3234 new_reg, *ad.index);
3237 else if (ad.index == NULL)
3239 int regno;
3240 enum reg_class cl;
3241 rtx set;
3242 rtx_insn *insns, *last_insn;
3243 /* Try to reload base into register only if the base is invalid
3244 for the address but with valid offset, case (4) above. */
3245 start_sequence ();
3246 new_reg = base_to_reg (&ad);
3248 /* base + disp => new base, cases (1) and (3) above. */
3249 /* Another option would be to reload the displacement into an
3250 index register. However, postreload has code to optimize
3251 address reloads that have the same base and different
3252 displacements, so reloading into an index register would
3253 not necessarily be a win. */
3254 if (new_reg == NULL_RTX)
3255 new_reg = base_plus_disp_to_reg (&ad);
3256 insns = get_insns ();
3257 last_insn = get_last_insn ();
3258 /* If we generated at least two insns, try last insn source as
3259 an address. If we succeed, we generate one less insn. */
3260 if (last_insn != insns && (set = single_set (last_insn)) != NULL_RTX
3261 && GET_CODE (SET_SRC (set)) == PLUS
3262 && REG_P (XEXP (SET_SRC (set), 0))
3263 && CONSTANT_P (XEXP (SET_SRC (set), 1)))
3265 *ad.inner = SET_SRC (set);
3266 if (valid_address_p (ad.mode, *ad.outer, ad.as))
3268 *ad.base_term = XEXP (SET_SRC (set), 0);
3269 *ad.disp_term = XEXP (SET_SRC (set), 1);
3270 cl = base_reg_class (ad.mode, ad.as, ad.base_outer_code,
3271 get_index_code (&ad));
3272 regno = REGNO (*ad.base_term);
3273 if (regno >= FIRST_PSEUDO_REGISTER
3274 && cl != lra_get_allocno_class (regno))
3275 lra_change_class (regno, cl, " Change to", true);
3276 new_reg = SET_SRC (set);
3277 delete_insns_since (PREV_INSN (last_insn));
3280 /* Try if target can split displacement into legitimite new disp
3281 and offset. If it's the case, we replace the last insn with
3282 insns for base + offset => new_reg and set new_reg + new disp
3283 to *ad.inner. */
3284 last_insn = get_last_insn ();
3285 if ((set = single_set (last_insn)) != NULL_RTX
3286 && GET_CODE (SET_SRC (set)) == PLUS
3287 && REG_P (XEXP (SET_SRC (set), 0))
3288 && REGNO (XEXP (SET_SRC (set), 0)) < FIRST_PSEUDO_REGISTER
3289 && CONST_INT_P (XEXP (SET_SRC (set), 1)))
3291 rtx addend, disp = XEXP (SET_SRC (set), 1);
3292 if (targetm.legitimize_address_displacement (&disp, &addend,
3293 ad.mode))
3295 rtx_insn *new_insns;
3296 start_sequence ();
3297 lra_emit_add (new_reg, XEXP (SET_SRC (set), 0), addend);
3298 new_insns = get_insns ();
3299 end_sequence ();
3300 new_reg = gen_rtx_PLUS (Pmode, new_reg, disp);
3301 delete_insns_since (PREV_INSN (last_insn));
3302 add_insn (new_insns);
3303 insns = get_insns ();
3306 end_sequence ();
3307 emit_insn (insns);
3308 *ad.inner = new_reg;
3310 else if (ad.disp_term != NULL)
3312 /* base + scale * index + disp => new base + scale * index,
3313 case (1) above. */
3314 new_reg = base_plus_disp_to_reg (&ad);
3315 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3316 new_reg, *ad.index);
3318 else if ((scale = get_index_scale (&ad)) == 1)
3320 /* The last transformation to one reg will be made in
3321 curr_insn_transform function. */
3322 end_sequence ();
3323 return false;
3325 else if (scale != 0)
3327 /* base + scale * index => base + new_reg,
3328 case (1) above.
3329 Index part of address may become invalid. For example, we
3330 changed pseudo on the equivalent memory and a subreg of the
3331 pseudo onto the memory of different mode for which the scale is
3332 prohibitted. */
3333 new_reg = index_part_to_reg (&ad);
3334 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3335 *ad.base_term, new_reg);
3337 else
3339 enum reg_class cl = base_reg_class (ad.mode, ad.as,
3340 SCRATCH, SCRATCH);
3341 rtx addr = *ad.inner;
3343 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "addr");
3344 /* addr => new_base. */
3345 lra_emit_move (new_reg, addr);
3346 *ad.inner = new_reg;
3348 *before = get_insns ();
3349 end_sequence ();
3350 return true;
3353 /* If CHECK_ONLY_P is false, do address reloads until it is necessary.
3354 Use process_address_1 as a helper function. Return true for any
3355 RTL changes.
3357 If CHECK_ONLY_P is true, just check address correctness. Return
3358 false if the address correct. */
3359 static bool
3360 process_address (int nop, bool check_only_p,
3361 rtx_insn **before, rtx_insn **after)
3363 bool res = false;
3365 while (process_address_1 (nop, check_only_p, before, after))
3367 if (check_only_p)
3368 return true;
3369 res = true;
3371 return res;
3374 /* Emit insns to reload VALUE into a new register. VALUE is an
3375 auto-increment or auto-decrement RTX whose operand is a register or
3376 memory location; so reloading involves incrementing that location.
3377 IN is either identical to VALUE, or some cheaper place to reload
3378 value being incremented/decremented from.
3380 INC_AMOUNT is the number to increment or decrement by (always
3381 positive and ignored for POST_MODIFY/PRE_MODIFY).
3383 Return pseudo containing the result. */
3384 static rtx
3385 emit_inc (enum reg_class new_rclass, rtx in, rtx value, int inc_amount)
3387 /* REG or MEM to be copied and incremented. */
3388 rtx incloc = XEXP (value, 0);
3389 /* Nonzero if increment after copying. */
3390 int post = (GET_CODE (value) == POST_DEC || GET_CODE (value) == POST_INC
3391 || GET_CODE (value) == POST_MODIFY);
3392 rtx_insn *last;
3393 rtx inc;
3394 rtx_insn *add_insn;
3395 int code;
3396 rtx real_in = in == value ? incloc : in;
3397 rtx result;
3398 bool plus_p = true;
3400 if (GET_CODE (value) == PRE_MODIFY || GET_CODE (value) == POST_MODIFY)
3402 lra_assert (GET_CODE (XEXP (value, 1)) == PLUS
3403 || GET_CODE (XEXP (value, 1)) == MINUS);
3404 lra_assert (rtx_equal_p (XEXP (XEXP (value, 1), 0), XEXP (value, 0)));
3405 plus_p = GET_CODE (XEXP (value, 1)) == PLUS;
3406 inc = XEXP (XEXP (value, 1), 1);
3408 else
3410 if (GET_CODE (value) == PRE_DEC || GET_CODE (value) == POST_DEC)
3411 inc_amount = -inc_amount;
3413 inc = GEN_INT (inc_amount);
3416 if (! post && REG_P (incloc))
3417 result = incloc;
3418 else
3419 result = lra_create_new_reg (GET_MODE (value), value, new_rclass,
3420 "INC/DEC result");
3422 if (real_in != result)
3424 /* First copy the location to the result register. */
3425 lra_assert (REG_P (result));
3426 emit_insn (gen_move_insn (result, real_in));
3429 /* We suppose that there are insns to add/sub with the constant
3430 increment permitted in {PRE/POST)_{DEC/INC/MODIFY}. At least the
3431 old reload worked with this assumption. If the assumption
3432 becomes wrong, we should use approach in function
3433 base_plus_disp_to_reg. */
3434 if (in == value)
3436 /* See if we can directly increment INCLOC. */
3437 last = get_last_insn ();
3438 add_insn = emit_insn (plus_p
3439 ? gen_add2_insn (incloc, inc)
3440 : gen_sub2_insn (incloc, inc));
3442 code = recog_memoized (add_insn);
3443 if (code >= 0)
3445 if (! post && result != incloc)
3446 emit_insn (gen_move_insn (result, incloc));
3447 return result;
3449 delete_insns_since (last);
3452 /* If couldn't do the increment directly, must increment in RESULT.
3453 The way we do this depends on whether this is pre- or
3454 post-increment. For pre-increment, copy INCLOC to the reload
3455 register, increment it there, then save back. */
3456 if (! post)
3458 if (real_in != result)
3459 emit_insn (gen_move_insn (result, real_in));
3460 if (plus_p)
3461 emit_insn (gen_add2_insn (result, inc));
3462 else
3463 emit_insn (gen_sub2_insn (result, inc));
3464 if (result != incloc)
3465 emit_insn (gen_move_insn (incloc, result));
3467 else
3469 /* Post-increment.
3471 Because this might be a jump insn or a compare, and because
3472 RESULT may not be available after the insn in an input
3473 reload, we must do the incrementing before the insn being
3474 reloaded for.
3476 We have already copied IN to RESULT. Increment the copy in
3477 RESULT, save that back, then decrement RESULT so it has
3478 the original value. */
3479 if (plus_p)
3480 emit_insn (gen_add2_insn (result, inc));
3481 else
3482 emit_insn (gen_sub2_insn (result, inc));
3483 emit_insn (gen_move_insn (incloc, result));
3484 /* Restore non-modified value for the result. We prefer this
3485 way because it does not require an additional hard
3486 register. */
3487 if (plus_p)
3489 if (CONST_INT_P (inc))
3490 emit_insn (gen_add2_insn (result,
3491 gen_int_mode (-INTVAL (inc),
3492 GET_MODE (result))));
3493 else
3494 emit_insn (gen_sub2_insn (result, inc));
3496 else
3497 emit_insn (gen_add2_insn (result, inc));
3499 return result;
3502 /* Return true if the current move insn does not need processing as we
3503 already know that it satisfies its constraints. */
3504 static bool
3505 simple_move_p (void)
3507 rtx dest, src;
3508 enum reg_class dclass, sclass;
3510 lra_assert (curr_insn_set != NULL_RTX);
3511 dest = SET_DEST (curr_insn_set);
3512 src = SET_SRC (curr_insn_set);
3514 /* If the instruction has multiple sets we need to process it even if it
3515 is single_set. This can happen if one or more of the SETs are dead.
3516 See PR73650. */
3517 if (multiple_sets (curr_insn))
3518 return false;
3520 return ((dclass = get_op_class (dest)) != NO_REGS
3521 && (sclass = get_op_class (src)) != NO_REGS
3522 /* The backend guarantees that register moves of cost 2
3523 never need reloads. */
3524 && targetm.register_move_cost (GET_MODE (src), sclass, dclass) == 2);
3527 /* Swap operands NOP and NOP + 1. */
3528 static inline void
3529 swap_operands (int nop)
3531 std::swap (curr_operand_mode[nop], curr_operand_mode[nop + 1]);
3532 std::swap (original_subreg_reg_mode[nop], original_subreg_reg_mode[nop + 1]);
3533 std::swap (*curr_id->operand_loc[nop], *curr_id->operand_loc[nop + 1]);
3534 std::swap (equiv_substition_p[nop], equiv_substition_p[nop + 1]);
3535 /* Swap the duplicates too. */
3536 lra_update_dup (curr_id, nop);
3537 lra_update_dup (curr_id, nop + 1);
3540 /* Main entry point of the constraint code: search the body of the
3541 current insn to choose the best alternative. It is mimicking insn
3542 alternative cost calculation model of former reload pass. That is
3543 because machine descriptions were written to use this model. This
3544 model can be changed in future. Make commutative operand exchange
3545 if it is chosen.
3547 if CHECK_ONLY_P is false, do RTL changes to satisfy the
3548 constraints. Return true if any change happened during function
3549 call.
3551 If CHECK_ONLY_P is true then don't do any transformation. Just
3552 check that the insn satisfies all constraints. If the insn does
3553 not satisfy any constraint, return true. */
3554 static bool
3555 curr_insn_transform (bool check_only_p)
3557 int i, j, k;
3558 int n_operands;
3559 int n_alternatives;
3560 int n_outputs;
3561 int commutative;
3562 signed char goal_alt_matched[MAX_RECOG_OPERANDS][MAX_RECOG_OPERANDS];
3563 signed char match_inputs[MAX_RECOG_OPERANDS + 1];
3564 signed char outputs[MAX_RECOG_OPERANDS + 1];
3565 rtx_insn *before, *after;
3566 bool alt_p = false;
3567 /* Flag that the insn has been changed through a transformation. */
3568 bool change_p;
3569 bool sec_mem_p;
3570 #ifdef SECONDARY_MEMORY_NEEDED
3571 bool use_sec_mem_p;
3572 #endif
3573 int max_regno_before;
3574 int reused_alternative_num;
3576 curr_insn_set = single_set (curr_insn);
3577 if (curr_insn_set != NULL_RTX && simple_move_p ())
3578 return false;
3580 no_input_reloads_p = no_output_reloads_p = false;
3581 goal_alt_number = -1;
3582 change_p = sec_mem_p = false;
3583 /* JUMP_INSNs and CALL_INSNs are not allowed to have any output
3584 reloads; neither are insns that SET cc0. Insns that use CC0 are
3585 not allowed to have any input reloads. */
3586 if (JUMP_P (curr_insn) || CALL_P (curr_insn))
3587 no_output_reloads_p = true;
3589 if (HAVE_cc0 && reg_referenced_p (cc0_rtx, PATTERN (curr_insn)))
3590 no_input_reloads_p = true;
3591 if (HAVE_cc0 && reg_set_p (cc0_rtx, PATTERN (curr_insn)))
3592 no_output_reloads_p = true;
3594 n_operands = curr_static_id->n_operands;
3595 n_alternatives = curr_static_id->n_alternatives;
3597 /* Just return "no reloads" if insn has no operands with
3598 constraints. */
3599 if (n_operands == 0 || n_alternatives == 0)
3600 return false;
3602 max_regno_before = max_reg_num ();
3604 for (i = 0; i < n_operands; i++)
3606 goal_alt_matched[i][0] = -1;
3607 goal_alt_matches[i] = -1;
3610 commutative = curr_static_id->commutative;
3612 /* Now see what we need for pseudos that didn't get hard regs or got
3613 the wrong kind of hard reg. For this, we must consider all the
3614 operands together against the register constraints. */
3616 best_losers = best_overall = INT_MAX;
3617 best_reload_sum = 0;
3619 curr_swapped = false;
3620 goal_alt_swapped = false;
3622 if (! check_only_p)
3623 /* Make equivalence substitution and memory subreg elimination
3624 before address processing because an address legitimacy can
3625 depend on memory mode. */
3626 for (i = 0; i < n_operands; i++)
3628 rtx op, subst, old;
3629 bool op_change_p = false;
3631 if (curr_static_id->operand[i].is_operator)
3632 continue;
3634 old = op = *curr_id->operand_loc[i];
3635 if (GET_CODE (old) == SUBREG)
3636 old = SUBREG_REG (old);
3637 subst = get_equiv_with_elimination (old, curr_insn);
3638 original_subreg_reg_mode[i] = VOIDmode;
3639 equiv_substition_p[i] = false;
3640 if (subst != old)
3642 equiv_substition_p[i] = true;
3643 subst = copy_rtx (subst);
3644 lra_assert (REG_P (old));
3645 if (GET_CODE (op) != SUBREG)
3646 *curr_id->operand_loc[i] = subst;
3647 else
3649 SUBREG_REG (op) = subst;
3650 if (GET_MODE (subst) == VOIDmode)
3651 original_subreg_reg_mode[i] = GET_MODE (old);
3653 if (lra_dump_file != NULL)
3655 fprintf (lra_dump_file,
3656 "Changing pseudo %d in operand %i of insn %u on equiv ",
3657 REGNO (old), i, INSN_UID (curr_insn));
3658 dump_value_slim (lra_dump_file, subst, 1);
3659 fprintf (lra_dump_file, "\n");
3661 op_change_p = change_p = true;
3663 if (simplify_operand_subreg (i, GET_MODE (old)) || op_change_p)
3665 change_p = true;
3666 lra_update_dup (curr_id, i);
3670 /* Reload address registers and displacements. We do it before
3671 finding an alternative because of memory constraints. */
3672 before = after = NULL;
3673 for (i = 0; i < n_operands; i++)
3674 if (! curr_static_id->operand[i].is_operator
3675 && process_address (i, check_only_p, &before, &after))
3677 if (check_only_p)
3678 return true;
3679 change_p = true;
3680 lra_update_dup (curr_id, i);
3683 if (change_p)
3684 /* If we've changed the instruction then any alternative that
3685 we chose previously may no longer be valid. */
3686 lra_set_used_insn_alternative (curr_insn, -1);
3688 if (! check_only_p && curr_insn_set != NULL_RTX
3689 && check_and_process_move (&change_p, &sec_mem_p))
3690 return change_p;
3692 try_swapped:
3694 reused_alternative_num = check_only_p ? -1 : curr_id->used_insn_alternative;
3695 if (lra_dump_file != NULL && reused_alternative_num >= 0)
3696 fprintf (lra_dump_file, "Reusing alternative %d for insn #%u\n",
3697 reused_alternative_num, INSN_UID (curr_insn));
3699 if (process_alt_operands (reused_alternative_num))
3700 alt_p = true;
3702 if (check_only_p)
3703 return ! alt_p || best_losers != 0;
3705 /* If insn is commutative (it's safe to exchange a certain pair of
3706 operands) then we need to try each alternative twice, the second
3707 time matching those two operands as if we had exchanged them. To
3708 do this, really exchange them in operands.
3710 If we have just tried the alternatives the second time, return
3711 operands to normal and drop through. */
3713 if (reused_alternative_num < 0 && commutative >= 0)
3715 curr_swapped = !curr_swapped;
3716 if (curr_swapped)
3718 swap_operands (commutative);
3719 goto try_swapped;
3721 else
3722 swap_operands (commutative);
3725 if (! alt_p && ! sec_mem_p)
3727 /* No alternative works with reloads?? */
3728 if (INSN_CODE (curr_insn) >= 0)
3729 fatal_insn ("unable to generate reloads for:", curr_insn);
3730 error_for_asm (curr_insn,
3731 "inconsistent operand constraints in an %<asm%>");
3732 /* Avoid further trouble with this insn. */
3733 PATTERN (curr_insn) = gen_rtx_USE (VOIDmode, const0_rtx);
3734 lra_invalidate_insn_data (curr_insn);
3735 return true;
3738 /* If the best alternative is with operands 1 and 2 swapped, swap
3739 them. Update the operand numbers of any reloads already
3740 pushed. */
3742 if (goal_alt_swapped)
3744 if (lra_dump_file != NULL)
3745 fprintf (lra_dump_file, " Commutative operand exchange in insn %u\n",
3746 INSN_UID (curr_insn));
3748 /* Swap the duplicates too. */
3749 swap_operands (commutative);
3750 change_p = true;
3753 #ifdef SECONDARY_MEMORY_NEEDED
3754 /* Some target macros SECONDARY_MEMORY_NEEDED (e.g. x86) are defined
3755 too conservatively. So we use the secondary memory only if there
3756 is no any alternative without reloads. */
3757 use_sec_mem_p = false;
3758 if (! alt_p)
3759 use_sec_mem_p = true;
3760 else if (sec_mem_p)
3762 for (i = 0; i < n_operands; i++)
3763 if (! goal_alt_win[i] && ! goal_alt_match_win[i])
3764 break;
3765 use_sec_mem_p = i < n_operands;
3768 if (use_sec_mem_p)
3770 int in = -1, out = -1;
3771 rtx new_reg, src, dest, rld;
3772 machine_mode sec_mode, rld_mode;
3774 lra_assert (curr_insn_set != NULL_RTX && sec_mem_p);
3775 dest = SET_DEST (curr_insn_set);
3776 src = SET_SRC (curr_insn_set);
3777 for (i = 0; i < n_operands; i++)
3778 if (*curr_id->operand_loc[i] == dest)
3779 out = i;
3780 else if (*curr_id->operand_loc[i] == src)
3781 in = i;
3782 for (i = 0; i < curr_static_id->n_dups; i++)
3783 if (out < 0 && *curr_id->dup_loc[i] == dest)
3784 out = curr_static_id->dup_num[i];
3785 else if (in < 0 && *curr_id->dup_loc[i] == src)
3786 in = curr_static_id->dup_num[i];
3787 lra_assert (out >= 0 && in >= 0
3788 && curr_static_id->operand[out].type == OP_OUT
3789 && curr_static_id->operand[in].type == OP_IN);
3790 rld = (GET_MODE_SIZE (GET_MODE (dest)) <= GET_MODE_SIZE (GET_MODE (src))
3791 ? dest : src);
3792 rld_mode = GET_MODE (rld);
3793 #ifdef SECONDARY_MEMORY_NEEDED_MODE
3794 sec_mode = SECONDARY_MEMORY_NEEDED_MODE (rld_mode);
3795 #else
3796 sec_mode = rld_mode;
3797 #endif
3798 new_reg = lra_create_new_reg (sec_mode, NULL_RTX,
3799 NO_REGS, "secondary");
3800 /* If the mode is changed, it should be wider. */
3801 lra_assert (GET_MODE_SIZE (sec_mode) >= GET_MODE_SIZE (rld_mode));
3802 if (sec_mode != rld_mode)
3804 /* If the target says specifically to use another mode for
3805 secondary memory moves we can not reuse the original
3806 insn. */
3807 after = emit_spill_move (false, new_reg, dest);
3808 lra_process_new_insns (curr_insn, NULL, after,
3809 "Inserting the sec. move");
3810 /* We may have non null BEFORE here (e.g. after address
3811 processing. */
3812 push_to_sequence (before);
3813 before = emit_spill_move (true, new_reg, src);
3814 emit_insn (before);
3815 before = get_insns ();
3816 end_sequence ();
3817 lra_process_new_insns (curr_insn, before, NULL, "Changing on");
3818 lra_set_insn_deleted (curr_insn);
3820 else if (dest == rld)
3822 *curr_id->operand_loc[out] = new_reg;
3823 lra_update_dup (curr_id, out);
3824 after = emit_spill_move (false, new_reg, dest);
3825 lra_process_new_insns (curr_insn, NULL, after,
3826 "Inserting the sec. move");
3828 else
3830 *curr_id->operand_loc[in] = new_reg;
3831 lra_update_dup (curr_id, in);
3832 /* See comments above. */
3833 push_to_sequence (before);
3834 before = emit_spill_move (true, new_reg, src);
3835 emit_insn (before);
3836 before = get_insns ();
3837 end_sequence ();
3838 lra_process_new_insns (curr_insn, before, NULL,
3839 "Inserting the sec. move");
3841 lra_update_insn_regno_info (curr_insn);
3842 return true;
3844 #endif
3846 lra_assert (goal_alt_number >= 0);
3847 lra_set_used_insn_alternative (curr_insn, goal_alt_number);
3849 if (lra_dump_file != NULL)
3851 const char *p;
3853 fprintf (lra_dump_file, " Choosing alt %d in insn %u:",
3854 goal_alt_number, INSN_UID (curr_insn));
3855 for (i = 0; i < n_operands; i++)
3857 p = (curr_static_id->operand_alternative
3858 [goal_alt_number * n_operands + i].constraint);
3859 if (*p == '\0')
3860 continue;
3861 fprintf (lra_dump_file, " (%d) ", i);
3862 for (; *p != '\0' && *p != ',' && *p != '#'; p++)
3863 fputc (*p, lra_dump_file);
3865 if (INSN_CODE (curr_insn) >= 0
3866 && (p = get_insn_name (INSN_CODE (curr_insn))) != NULL)
3867 fprintf (lra_dump_file, " {%s}", p);
3868 if (curr_id->sp_offset != 0)
3869 fprintf (lra_dump_file, " (sp_off=%" HOST_WIDE_INT_PRINT "d)",
3870 curr_id->sp_offset);
3871 fprintf (lra_dump_file, "\n");
3874 /* Right now, for any pair of operands I and J that are required to
3875 match, with J < I, goal_alt_matches[I] is J. Add I to
3876 goal_alt_matched[J]. */
3878 for (i = 0; i < n_operands; i++)
3879 if ((j = goal_alt_matches[i]) >= 0)
3881 for (k = 0; goal_alt_matched[j][k] >= 0; k++)
3883 /* We allow matching one output operand and several input
3884 operands. */
3885 lra_assert (k == 0
3886 || (curr_static_id->operand[j].type == OP_OUT
3887 && curr_static_id->operand[i].type == OP_IN
3888 && (curr_static_id->operand
3889 [goal_alt_matched[j][0]].type == OP_IN)));
3890 goal_alt_matched[j][k] = i;
3891 goal_alt_matched[j][k + 1] = -1;
3894 for (i = 0; i < n_operands; i++)
3895 goal_alt_win[i] |= goal_alt_match_win[i];
3897 /* Any constants that aren't allowed and can't be reloaded into
3898 registers are here changed into memory references. */
3899 for (i = 0; i < n_operands; i++)
3900 if (goal_alt_win[i])
3902 int regno;
3903 enum reg_class new_class;
3904 rtx reg = *curr_id->operand_loc[i];
3906 if (GET_CODE (reg) == SUBREG)
3907 reg = SUBREG_REG (reg);
3909 if (REG_P (reg) && (regno = REGNO (reg)) >= FIRST_PSEUDO_REGISTER)
3911 bool ok_p = in_class_p (reg, goal_alt[i], &new_class);
3913 if (new_class != NO_REGS && get_reg_class (regno) != new_class)
3915 lra_assert (ok_p);
3916 lra_change_class (regno, new_class, " Change to", true);
3920 else
3922 const char *constraint;
3923 char c;
3924 rtx op = *curr_id->operand_loc[i];
3925 rtx subreg = NULL_RTX;
3926 machine_mode mode = curr_operand_mode[i];
3928 if (GET_CODE (op) == SUBREG)
3930 subreg = op;
3931 op = SUBREG_REG (op);
3932 mode = GET_MODE (op);
3935 if (CONST_POOL_OK_P (mode, op)
3936 && ((targetm.preferred_reload_class
3937 (op, (enum reg_class) goal_alt[i]) == NO_REGS)
3938 || no_input_reloads_p))
3940 rtx tem = force_const_mem (mode, op);
3942 change_p = true;
3943 if (subreg != NULL_RTX)
3944 tem = gen_rtx_SUBREG (mode, tem, SUBREG_BYTE (subreg));
3946 *curr_id->operand_loc[i] = tem;
3947 lra_update_dup (curr_id, i);
3948 process_address (i, false, &before, &after);
3950 /* If the alternative accepts constant pool refs directly
3951 there will be no reload needed at all. */
3952 if (subreg != NULL_RTX)
3953 continue;
3954 /* Skip alternatives before the one requested. */
3955 constraint = (curr_static_id->operand_alternative
3956 [goal_alt_number * n_operands + i].constraint);
3957 for (;
3958 (c = *constraint) && c != ',' && c != '#';
3959 constraint += CONSTRAINT_LEN (c, constraint))
3961 enum constraint_num cn = lookup_constraint (constraint);
3962 if ((insn_extra_memory_constraint (cn)
3963 || insn_extra_special_memory_constraint (cn))
3964 && satisfies_memory_constraint_p (tem, cn))
3965 break;
3967 if (c == '\0' || c == ',' || c == '#')
3968 continue;
3970 goal_alt_win[i] = true;
3974 n_outputs = 0;
3975 outputs[0] = -1;
3976 for (i = 0; i < n_operands; i++)
3978 int regno;
3979 bool optional_p = false;
3980 rtx old, new_reg;
3981 rtx op = *curr_id->operand_loc[i];
3983 if (goal_alt_win[i])
3985 if (goal_alt[i] == NO_REGS
3986 && REG_P (op)
3987 /* When we assign NO_REGS it means that we will not
3988 assign a hard register to the scratch pseudo by
3989 assigment pass and the scratch pseudo will be
3990 spilled. Spilled scratch pseudos are transformed
3991 back to scratches at the LRA end. */
3992 && lra_former_scratch_operand_p (curr_insn, i)
3993 && lra_former_scratch_p (REGNO (op)))
3995 int regno = REGNO (op);
3996 lra_change_class (regno, NO_REGS, " Change to", true);
3997 if (lra_get_regno_hard_regno (regno) >= 0)
3998 /* We don't have to mark all insn affected by the
3999 spilled pseudo as there is only one such insn, the
4000 current one. */
4001 reg_renumber[regno] = -1;
4002 lra_assert (bitmap_single_bit_set_p
4003 (&lra_reg_info[REGNO (op)].insn_bitmap));
4005 /* We can do an optional reload. If the pseudo got a hard
4006 reg, we might improve the code through inheritance. If
4007 it does not get a hard register we coalesce memory/memory
4008 moves later. Ignore move insns to avoid cycling. */
4009 if (! lra_simple_p
4010 && lra_undo_inheritance_iter < LRA_MAX_INHERITANCE_PASSES
4011 && goal_alt[i] != NO_REGS && REG_P (op)
4012 && (regno = REGNO (op)) >= FIRST_PSEUDO_REGISTER
4013 && regno < new_regno_start
4014 && ! lra_former_scratch_p (regno)
4015 && reg_renumber[regno] < 0
4016 /* Check that the optional reload pseudo will be able to
4017 hold given mode value. */
4018 && ! (prohibited_class_reg_set_mode_p
4019 (goal_alt[i], reg_class_contents[goal_alt[i]],
4020 PSEUDO_REGNO_MODE (regno)))
4021 && (curr_insn_set == NULL_RTX
4022 || !((REG_P (SET_SRC (curr_insn_set))
4023 || MEM_P (SET_SRC (curr_insn_set))
4024 || GET_CODE (SET_SRC (curr_insn_set)) == SUBREG)
4025 && (REG_P (SET_DEST (curr_insn_set))
4026 || MEM_P (SET_DEST (curr_insn_set))
4027 || GET_CODE (SET_DEST (curr_insn_set)) == SUBREG))))
4028 optional_p = true;
4029 else
4030 continue;
4033 /* Operands that match previous ones have already been handled. */
4034 if (goal_alt_matches[i] >= 0)
4035 continue;
4037 /* We should not have an operand with a non-offsettable address
4038 appearing where an offsettable address will do. It also may
4039 be a case when the address should be special in other words
4040 not a general one (e.g. it needs no index reg). */
4041 if (goal_alt_matched[i][0] == -1 && goal_alt_offmemok[i] && MEM_P (op))
4043 enum reg_class rclass;
4044 rtx *loc = &XEXP (op, 0);
4045 enum rtx_code code = GET_CODE (*loc);
4047 push_to_sequence (before);
4048 rclass = base_reg_class (GET_MODE (op), MEM_ADDR_SPACE (op),
4049 MEM, SCRATCH);
4050 if (GET_RTX_CLASS (code) == RTX_AUTOINC)
4051 new_reg = emit_inc (rclass, *loc, *loc,
4052 /* This value does not matter for MODIFY. */
4053 GET_MODE_SIZE (GET_MODE (op)));
4054 else if (get_reload_reg (OP_IN, Pmode, *loc, rclass, FALSE,
4055 "offsetable address", &new_reg))
4056 lra_emit_move (new_reg, *loc);
4057 before = get_insns ();
4058 end_sequence ();
4059 *loc = new_reg;
4060 lra_update_dup (curr_id, i);
4062 else if (goal_alt_matched[i][0] == -1)
4064 machine_mode mode;
4065 rtx reg, *loc;
4066 int hard_regno, byte;
4067 enum op_type type = curr_static_id->operand[i].type;
4069 loc = curr_id->operand_loc[i];
4070 mode = curr_operand_mode[i];
4071 if (GET_CODE (*loc) == SUBREG)
4073 reg = SUBREG_REG (*loc);
4074 byte = SUBREG_BYTE (*loc);
4075 if (REG_P (reg)
4076 /* Strict_low_part requires reload the register not
4077 the sub-register. */
4078 && (curr_static_id->operand[i].strict_low
4079 || (GET_MODE_SIZE (mode)
4080 <= GET_MODE_SIZE (GET_MODE (reg))
4081 && (hard_regno
4082 = get_try_hard_regno (REGNO (reg))) >= 0
4083 && (simplify_subreg_regno
4084 (hard_regno,
4085 GET_MODE (reg), byte, mode) < 0)
4086 && (goal_alt[i] == NO_REGS
4087 || (simplify_subreg_regno
4088 (ira_class_hard_regs[goal_alt[i]][0],
4089 GET_MODE (reg), byte, mode) >= 0)))))
4091 if (type == OP_OUT)
4092 type = OP_INOUT;
4093 loc = &SUBREG_REG (*loc);
4094 mode = GET_MODE (*loc);
4097 old = *loc;
4098 if (get_reload_reg (type, mode, old, goal_alt[i],
4099 loc != curr_id->operand_loc[i], "", &new_reg)
4100 && type != OP_OUT)
4102 push_to_sequence (before);
4103 lra_emit_move (new_reg, old);
4104 before = get_insns ();
4105 end_sequence ();
4107 *loc = new_reg;
4108 if (type != OP_IN
4109 && find_reg_note (curr_insn, REG_UNUSED, old) == NULL_RTX)
4111 start_sequence ();
4112 lra_emit_move (type == OP_INOUT ? copy_rtx (old) : old, new_reg);
4113 emit_insn (after);
4114 after = get_insns ();
4115 end_sequence ();
4116 *loc = new_reg;
4118 for (j = 0; j < goal_alt_dont_inherit_ops_num; j++)
4119 if (goal_alt_dont_inherit_ops[j] == i)
4121 lra_set_regno_unique_value (REGNO (new_reg));
4122 break;
4124 lra_update_dup (curr_id, i);
4126 else if (curr_static_id->operand[i].type == OP_IN
4127 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4128 == OP_OUT))
4130 /* generate reloads for input and matched outputs. */
4131 match_inputs[0] = i;
4132 match_inputs[1] = -1;
4133 match_reload (goal_alt_matched[i][0], match_inputs, outputs,
4134 goal_alt[i], &before, &after,
4135 curr_static_id->operand_alternative
4136 [goal_alt_number * n_operands + goal_alt_matched[i][0]]
4137 .earlyclobber);
4139 else if (curr_static_id->operand[i].type == OP_OUT
4140 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4141 == OP_IN))
4142 /* Generate reloads for output and matched inputs. */
4143 match_reload (i, goal_alt_matched[i], outputs, goal_alt[i], &before,
4144 &after, curr_static_id->operand_alternative
4145 [goal_alt_number * n_operands + i].earlyclobber);
4146 else if (curr_static_id->operand[i].type == OP_IN
4147 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4148 == OP_IN))
4150 /* Generate reloads for matched inputs. */
4151 match_inputs[0] = i;
4152 for (j = 0; (k = goal_alt_matched[i][j]) >= 0; j++)
4153 match_inputs[j + 1] = k;
4154 match_inputs[j + 1] = -1;
4155 match_reload (-1, match_inputs, outputs, goal_alt[i], &before,
4156 &after, false);
4158 else
4159 /* We must generate code in any case when function
4160 process_alt_operands decides that it is possible. */
4161 gcc_unreachable ();
4163 /* Memorise processed outputs so that output remaining to be processed
4164 can avoid using the same register value (see match_reload). */
4165 if (curr_static_id->operand[i].type == OP_OUT)
4167 outputs[n_outputs++] = i;
4168 outputs[n_outputs] = -1;
4171 if (optional_p)
4173 rtx reg = op;
4175 lra_assert (REG_P (reg));
4176 regno = REGNO (reg);
4177 op = *curr_id->operand_loc[i]; /* Substitution. */
4178 if (GET_CODE (op) == SUBREG)
4179 op = SUBREG_REG (op);
4180 gcc_assert (REG_P (op) && (int) REGNO (op) >= new_regno_start);
4181 bitmap_set_bit (&lra_optional_reload_pseudos, REGNO (op));
4182 lra_reg_info[REGNO (op)].restore_rtx = reg;
4183 if (lra_dump_file != NULL)
4184 fprintf (lra_dump_file,
4185 " Making reload reg %d for reg %d optional\n",
4186 REGNO (op), regno);
4189 if (before != NULL_RTX || after != NULL_RTX
4190 || max_regno_before != max_reg_num ())
4191 change_p = true;
4192 if (change_p)
4194 lra_update_operator_dups (curr_id);
4195 /* Something changes -- process the insn. */
4196 lra_update_insn_regno_info (curr_insn);
4198 lra_process_new_insns (curr_insn, before, after, "Inserting insn reload");
4199 return change_p;
4202 /* Return true if INSN satisfies all constraints. In other words, no
4203 reload insns are needed. */
4204 bool
4205 lra_constrain_insn (rtx_insn *insn)
4207 int saved_new_regno_start = new_regno_start;
4208 int saved_new_insn_uid_start = new_insn_uid_start;
4209 bool change_p;
4211 curr_insn = insn;
4212 curr_id = lra_get_insn_recog_data (curr_insn);
4213 curr_static_id = curr_id->insn_static_data;
4214 new_insn_uid_start = get_max_uid ();
4215 new_regno_start = max_reg_num ();
4216 change_p = curr_insn_transform (true);
4217 new_regno_start = saved_new_regno_start;
4218 new_insn_uid_start = saved_new_insn_uid_start;
4219 return ! change_p;
4222 /* Return true if X is in LIST. */
4223 static bool
4224 in_list_p (rtx x, rtx list)
4226 for (; list != NULL_RTX; list = XEXP (list, 1))
4227 if (XEXP (list, 0) == x)
4228 return true;
4229 return false;
4232 /* Return true if X contains an allocatable hard register (if
4233 HARD_REG_P) or a (spilled if SPILLED_P) pseudo. */
4234 static bool
4235 contains_reg_p (rtx x, bool hard_reg_p, bool spilled_p)
4237 int i, j;
4238 const char *fmt;
4239 enum rtx_code code;
4241 code = GET_CODE (x);
4242 if (REG_P (x))
4244 int regno = REGNO (x);
4245 HARD_REG_SET alloc_regs;
4247 if (hard_reg_p)
4249 if (regno >= FIRST_PSEUDO_REGISTER)
4250 regno = lra_get_regno_hard_regno (regno);
4251 if (regno < 0)
4252 return false;
4253 COMPL_HARD_REG_SET (alloc_regs, lra_no_alloc_regs);
4254 return overlaps_hard_reg_set_p (alloc_regs, GET_MODE (x), regno);
4256 else
4258 if (regno < FIRST_PSEUDO_REGISTER)
4259 return false;
4260 if (! spilled_p)
4261 return true;
4262 return lra_get_regno_hard_regno (regno) < 0;
4265 fmt = GET_RTX_FORMAT (code);
4266 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4268 if (fmt[i] == 'e')
4270 if (contains_reg_p (XEXP (x, i), hard_reg_p, spilled_p))
4271 return true;
4273 else if (fmt[i] == 'E')
4275 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4276 if (contains_reg_p (XVECEXP (x, i, j), hard_reg_p, spilled_p))
4277 return true;
4280 return false;
4283 /* Process all regs in location *LOC and change them on equivalent
4284 substitution. Return true if any change was done. */
4285 static bool
4286 loc_equivalence_change_p (rtx *loc)
4288 rtx subst, reg, x = *loc;
4289 bool result = false;
4290 enum rtx_code code = GET_CODE (x);
4291 const char *fmt;
4292 int i, j;
4294 if (code == SUBREG)
4296 reg = SUBREG_REG (x);
4297 if ((subst = get_equiv_with_elimination (reg, curr_insn)) != reg
4298 && GET_MODE (subst) == VOIDmode)
4300 /* We cannot reload debug location. Simplify subreg here
4301 while we know the inner mode. */
4302 *loc = simplify_gen_subreg (GET_MODE (x), subst,
4303 GET_MODE (reg), SUBREG_BYTE (x));
4304 return true;
4307 if (code == REG && (subst = get_equiv_with_elimination (x, curr_insn)) != x)
4309 *loc = subst;
4310 return true;
4313 /* Scan all the operand sub-expressions. */
4314 fmt = GET_RTX_FORMAT (code);
4315 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4317 if (fmt[i] == 'e')
4318 result = loc_equivalence_change_p (&XEXP (x, i)) || result;
4319 else if (fmt[i] == 'E')
4320 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4321 result
4322 = loc_equivalence_change_p (&XVECEXP (x, i, j)) || result;
4324 return result;
4327 /* Similar to loc_equivalence_change_p, but for use as
4328 simplify_replace_fn_rtx callback. DATA is insn for which the
4329 elimination is done. If it null we don't do the elimination. */
4330 static rtx
4331 loc_equivalence_callback (rtx loc, const_rtx, void *data)
4333 if (!REG_P (loc))
4334 return NULL_RTX;
4336 rtx subst = (data == NULL
4337 ? get_equiv (loc) : get_equiv_with_elimination (loc, (rtx_insn *) data));
4338 if (subst != loc)
4339 return subst;
4341 return NULL_RTX;
4344 /* Maximum number of generated reload insns per an insn. It is for
4345 preventing this pass cycling in a bug case. */
4346 #define MAX_RELOAD_INSNS_NUMBER LRA_MAX_INSN_RELOADS
4348 /* The current iteration number of this LRA pass. */
4349 int lra_constraint_iter;
4351 /* True if we substituted equiv which needs checking register
4352 allocation correctness because the equivalent value contains
4353 allocatable hard registers or when we restore multi-register
4354 pseudo. */
4355 bool lra_risky_transformations_p;
4357 /* Return true if REGNO is referenced in more than one block. */
4358 static bool
4359 multi_block_pseudo_p (int regno)
4361 basic_block bb = NULL;
4362 unsigned int uid;
4363 bitmap_iterator bi;
4365 if (regno < FIRST_PSEUDO_REGISTER)
4366 return false;
4368 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi)
4369 if (bb == NULL)
4370 bb = BLOCK_FOR_INSN (lra_insn_recog_data[uid]->insn);
4371 else if (BLOCK_FOR_INSN (lra_insn_recog_data[uid]->insn) != bb)
4372 return true;
4373 return false;
4376 /* Return true if LIST contains a deleted insn. */
4377 static bool
4378 contains_deleted_insn_p (rtx_insn_list *list)
4380 for (; list != NULL_RTX; list = list->next ())
4381 if (NOTE_P (list->insn ())
4382 && NOTE_KIND (list->insn ()) == NOTE_INSN_DELETED)
4383 return true;
4384 return false;
4387 /* Return true if X contains a pseudo dying in INSN. */
4388 static bool
4389 dead_pseudo_p (rtx x, rtx_insn *insn)
4391 int i, j;
4392 const char *fmt;
4393 enum rtx_code code;
4395 if (REG_P (x))
4396 return (insn != NULL_RTX
4397 && find_regno_note (insn, REG_DEAD, REGNO (x)) != NULL_RTX);
4398 code = GET_CODE (x);
4399 fmt = GET_RTX_FORMAT (code);
4400 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4402 if (fmt[i] == 'e')
4404 if (dead_pseudo_p (XEXP (x, i), insn))
4405 return true;
4407 else if (fmt[i] == 'E')
4409 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4410 if (dead_pseudo_p (XVECEXP (x, i, j), insn))
4411 return true;
4414 return false;
4417 /* Return true if INSN contains a dying pseudo in INSN right hand
4418 side. */
4419 static bool
4420 insn_rhs_dead_pseudo_p (rtx_insn *insn)
4422 rtx set = single_set (insn);
4424 gcc_assert (set != NULL);
4425 return dead_pseudo_p (SET_SRC (set), insn);
4428 /* Return true if any init insn of REGNO contains a dying pseudo in
4429 insn right hand side. */
4430 static bool
4431 init_insn_rhs_dead_pseudo_p (int regno)
4433 rtx_insn_list *insns = ira_reg_equiv[regno].init_insns;
4435 if (insns == NULL)
4436 return false;
4437 for (; insns != NULL_RTX; insns = insns->next ())
4438 if (insn_rhs_dead_pseudo_p (insns->insn ()))
4439 return true;
4440 return false;
4443 /* Return TRUE if REGNO has a reverse equivalence. The equivalence is
4444 reverse only if we have one init insn with given REGNO as a
4445 source. */
4446 static bool
4447 reverse_equiv_p (int regno)
4449 rtx_insn_list *insns = ira_reg_equiv[regno].init_insns;
4450 rtx set;
4452 if (insns == NULL)
4453 return false;
4454 if (! INSN_P (insns->insn ())
4455 || insns->next () != NULL)
4456 return false;
4457 if ((set = single_set (insns->insn ())) == NULL_RTX)
4458 return false;
4459 return REG_P (SET_SRC (set)) && (int) REGNO (SET_SRC (set)) == regno;
4462 /* Return TRUE if REGNO was reloaded in an equivalence init insn. We
4463 call this function only for non-reverse equivalence. */
4464 static bool
4465 contains_reloaded_insn_p (int regno)
4467 rtx set;
4468 rtx_insn_list *list = ira_reg_equiv[regno].init_insns;
4470 for (; list != NULL; list = list->next ())
4471 if ((set = single_set (list->insn ())) == NULL_RTX
4472 || ! REG_P (SET_DEST (set))
4473 || (int) REGNO (SET_DEST (set)) != regno)
4474 return true;
4475 return false;
4478 /* Entry function of LRA constraint pass. Return true if the
4479 constraint pass did change the code. */
4480 bool
4481 lra_constraints (bool first_p)
4483 bool changed_p;
4484 int i, hard_regno, new_insns_num;
4485 unsigned int min_len, new_min_len, uid;
4486 rtx set, x, reg, dest_reg;
4487 basic_block last_bb;
4488 bitmap_head equiv_insn_bitmap;
4489 bitmap_iterator bi;
4491 lra_constraint_iter++;
4492 if (lra_dump_file != NULL)
4493 fprintf (lra_dump_file, "\n********** Local #%d: **********\n\n",
4494 lra_constraint_iter);
4495 changed_p = false;
4496 if (pic_offset_table_rtx
4497 && REGNO (pic_offset_table_rtx) >= FIRST_PSEUDO_REGISTER)
4498 lra_risky_transformations_p = true;
4499 else
4500 lra_risky_transformations_p = false;
4501 new_insn_uid_start = get_max_uid ();
4502 new_regno_start = first_p ? lra_constraint_new_regno_start : max_reg_num ();
4503 /* Mark used hard regs for target stack size calulations. */
4504 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4505 if (lra_reg_info[i].nrefs != 0
4506 && (hard_regno = lra_get_regno_hard_regno (i)) >= 0)
4508 int j, nregs;
4510 nregs = hard_regno_nregs[hard_regno][lra_reg_info[i].biggest_mode];
4511 for (j = 0; j < nregs; j++)
4512 df_set_regs_ever_live (hard_regno + j, true);
4514 /* Do elimination before the equivalence processing as we can spill
4515 some pseudos during elimination. */
4516 lra_eliminate (false, first_p);
4517 bitmap_initialize (&equiv_insn_bitmap, &reg_obstack);
4518 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4519 if (lra_reg_info[i].nrefs != 0)
4521 ira_reg_equiv[i].profitable_p = true;
4522 reg = regno_reg_rtx[i];
4523 if (lra_get_regno_hard_regno (i) < 0 && (x = get_equiv (reg)) != reg)
4525 bool pseudo_p = contains_reg_p (x, false, false);
4527 /* After RTL transformation, we can not guarantee that
4528 pseudo in the substitution was not reloaded which might
4529 make equivalence invalid. For example, in reverse
4530 equiv of p0
4532 p0 <- ...
4534 equiv_mem <- p0
4536 the memory address register was reloaded before the 2nd
4537 insn. */
4538 if ((! first_p && pseudo_p)
4539 /* We don't use DF for compilation speed sake. So it
4540 is problematic to update live info when we use an
4541 equivalence containing pseudos in more than one
4542 BB. */
4543 || (pseudo_p && multi_block_pseudo_p (i))
4544 /* If an init insn was deleted for some reason, cancel
4545 the equiv. We could update the equiv insns after
4546 transformations including an equiv insn deletion
4547 but it is not worthy as such cases are extremely
4548 rare. */
4549 || contains_deleted_insn_p (ira_reg_equiv[i].init_insns)
4550 /* If it is not a reverse equivalence, we check that a
4551 pseudo in rhs of the init insn is not dying in the
4552 insn. Otherwise, the live info at the beginning of
4553 the corresponding BB might be wrong after we
4554 removed the insn. When the equiv can be a
4555 constant, the right hand side of the init insn can
4556 be a pseudo. */
4557 || (! reverse_equiv_p (i)
4558 && (init_insn_rhs_dead_pseudo_p (i)
4559 /* If we reloaded the pseudo in an equivalence
4560 init insn, we can not remove the equiv init
4561 insns and the init insns might write into
4562 const memory in this case. */
4563 || contains_reloaded_insn_p (i)))
4564 /* Prevent access beyond equivalent memory for
4565 paradoxical subregs. */
4566 || (MEM_P (x)
4567 && (GET_MODE_SIZE (lra_reg_info[i].biggest_mode)
4568 > GET_MODE_SIZE (GET_MODE (x))))
4569 || (pic_offset_table_rtx
4570 && ((CONST_POOL_OK_P (PSEUDO_REGNO_MODE (i), x)
4571 && (targetm.preferred_reload_class
4572 (x, lra_get_allocno_class (i)) == NO_REGS))
4573 || contains_symbol_ref_p (x))))
4574 ira_reg_equiv[i].defined_p = false;
4575 if (contains_reg_p (x, false, true))
4576 ira_reg_equiv[i].profitable_p = false;
4577 if (get_equiv (reg) != reg)
4578 bitmap_ior_into (&equiv_insn_bitmap, &lra_reg_info[i].insn_bitmap);
4581 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4582 update_equiv (i);
4583 /* We should add all insns containing pseudos which should be
4584 substituted by their equivalences. */
4585 EXECUTE_IF_SET_IN_BITMAP (&equiv_insn_bitmap, 0, uid, bi)
4586 lra_push_insn_by_uid (uid);
4587 min_len = lra_insn_stack_length ();
4588 new_insns_num = 0;
4589 last_bb = NULL;
4590 changed_p = false;
4591 while ((new_min_len = lra_insn_stack_length ()) != 0)
4593 curr_insn = lra_pop_insn ();
4594 --new_min_len;
4595 curr_bb = BLOCK_FOR_INSN (curr_insn);
4596 if (curr_bb != last_bb)
4598 last_bb = curr_bb;
4599 bb_reload_num = lra_curr_reload_num;
4601 if (min_len > new_min_len)
4603 min_len = new_min_len;
4604 new_insns_num = 0;
4606 if (new_insns_num > MAX_RELOAD_INSNS_NUMBER)
4607 internal_error
4608 ("Max. number of generated reload insns per insn is achieved (%d)\n",
4609 MAX_RELOAD_INSNS_NUMBER);
4610 new_insns_num++;
4611 if (DEBUG_INSN_P (curr_insn))
4613 /* We need to check equivalence in debug insn and change
4614 pseudo to the equivalent value if necessary. */
4615 curr_id = lra_get_insn_recog_data (curr_insn);
4616 if (bitmap_bit_p (&equiv_insn_bitmap, INSN_UID (curr_insn)))
4618 rtx old = *curr_id->operand_loc[0];
4619 *curr_id->operand_loc[0]
4620 = simplify_replace_fn_rtx (old, NULL_RTX,
4621 loc_equivalence_callback, curr_insn);
4622 if (old != *curr_id->operand_loc[0])
4624 lra_update_insn_regno_info (curr_insn);
4625 changed_p = true;
4629 else if (INSN_P (curr_insn))
4631 if ((set = single_set (curr_insn)) != NULL_RTX)
4633 dest_reg = SET_DEST (set);
4634 /* The equivalence pseudo could be set up as SUBREG in a
4635 case when it is a call restore insn in a mode
4636 different from the pseudo mode. */
4637 if (GET_CODE (dest_reg) == SUBREG)
4638 dest_reg = SUBREG_REG (dest_reg);
4639 if ((REG_P (dest_reg)
4640 && (x = get_equiv (dest_reg)) != dest_reg
4641 /* Remove insns which set up a pseudo whose value
4642 can not be changed. Such insns might be not in
4643 init_insns because we don't update equiv data
4644 during insn transformations.
4646 As an example, let suppose that a pseudo got
4647 hard register and on the 1st pass was not
4648 changed to equivalent constant. We generate an
4649 additional insn setting up the pseudo because of
4650 secondary memory movement. Then the pseudo is
4651 spilled and we use the equiv constant. In this
4652 case we should remove the additional insn and
4653 this insn is not init_insns list. */
4654 && (! MEM_P (x) || MEM_READONLY_P (x)
4655 /* Check that this is actually an insn setting
4656 up the equivalence. */
4657 || in_list_p (curr_insn,
4658 ira_reg_equiv
4659 [REGNO (dest_reg)].init_insns)))
4660 || (((x = get_equiv (SET_SRC (set))) != SET_SRC (set))
4661 && in_list_p (curr_insn,
4662 ira_reg_equiv
4663 [REGNO (SET_SRC (set))].init_insns)))
4665 /* This is equiv init insn of pseudo which did not get a
4666 hard register -- remove the insn. */
4667 if (lra_dump_file != NULL)
4669 fprintf (lra_dump_file,
4670 " Removing equiv init insn %i (freq=%d)\n",
4671 INSN_UID (curr_insn),
4672 REG_FREQ_FROM_BB (BLOCK_FOR_INSN (curr_insn)));
4673 dump_insn_slim (lra_dump_file, curr_insn);
4675 if (contains_reg_p (x, true, false))
4676 lra_risky_transformations_p = true;
4677 lra_set_insn_deleted (curr_insn);
4678 continue;
4681 curr_id = lra_get_insn_recog_data (curr_insn);
4682 curr_static_id = curr_id->insn_static_data;
4683 init_curr_insn_input_reloads ();
4684 init_curr_operand_mode ();
4685 if (curr_insn_transform (false))
4686 changed_p = true;
4687 /* Check non-transformed insns too for equiv change as USE
4688 or CLOBBER don't need reloads but can contain pseudos
4689 being changed on their equivalences. */
4690 else if (bitmap_bit_p (&equiv_insn_bitmap, INSN_UID (curr_insn))
4691 && loc_equivalence_change_p (&PATTERN (curr_insn)))
4693 lra_update_insn_regno_info (curr_insn);
4694 changed_p = true;
4698 bitmap_clear (&equiv_insn_bitmap);
4699 /* If we used a new hard regno, changed_p should be true because the
4700 hard reg is assigned to a new pseudo. */
4701 if (flag_checking && !changed_p)
4703 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4704 if (lra_reg_info[i].nrefs != 0
4705 && (hard_regno = lra_get_regno_hard_regno (i)) >= 0)
4707 int j, nregs = hard_regno_nregs[hard_regno][PSEUDO_REGNO_MODE (i)];
4709 for (j = 0; j < nregs; j++)
4710 lra_assert (df_regs_ever_live_p (hard_regno + j));
4713 return changed_p;
4716 static void initiate_invariants (void);
4717 static void finish_invariants (void);
4719 /* Initiate the LRA constraint pass. It is done once per
4720 function. */
4721 void
4722 lra_constraints_init (void)
4724 initiate_invariants ();
4727 /* Finalize the LRA constraint pass. It is done once per
4728 function. */
4729 void
4730 lra_constraints_finish (void)
4732 finish_invariants ();
4737 /* Structure describes invariants for ineheritance. */
4738 struct invariant
4740 /* The order number of the invariant. */
4741 int num;
4742 /* The invariant RTX. */
4743 rtx invariant_rtx;
4744 /* The origin insn of the invariant. */
4745 rtx_insn *insn;
4748 typedef struct invariant invariant_t;
4749 typedef invariant_t *invariant_ptr_t;
4750 typedef const invariant_t *const_invariant_ptr_t;
4752 /* Pointer to the inheritance invariants. */
4753 static vec<invariant_ptr_t> invariants;
4755 /* Allocation pool for the invariants. */
4756 static object_allocator<struct invariant> *invariants_pool;
4758 /* Hash table for the invariants. */
4759 static htab_t invariant_table;
4761 /* Hash function for INVARIANT. */
4762 static hashval_t
4763 invariant_hash (const void *invariant)
4765 rtx inv = ((const_invariant_ptr_t) invariant)->invariant_rtx;
4766 return lra_rtx_hash (inv);
4769 /* Equal function for invariants INVARIANT1 and INVARIANT2. */
4770 static int
4771 invariant_eq_p (const void *invariant1, const void *invariant2)
4773 rtx inv1 = ((const_invariant_ptr_t) invariant1)->invariant_rtx;
4774 rtx inv2 = ((const_invariant_ptr_t) invariant2)->invariant_rtx;
4776 return rtx_equal_p (inv1, inv2);
4779 /* Insert INVARIANT_RTX into the table if it is not there yet. Return
4780 invariant which is in the table. */
4781 static invariant_ptr_t
4782 insert_invariant (rtx invariant_rtx)
4784 void **entry_ptr;
4785 invariant_t invariant;
4786 invariant_ptr_t invariant_ptr;
4788 invariant.invariant_rtx = invariant_rtx;
4789 entry_ptr = htab_find_slot (invariant_table, &invariant, INSERT);
4790 if (*entry_ptr == NULL)
4792 invariant_ptr = invariants_pool->allocate ();
4793 invariant_ptr->invariant_rtx = invariant_rtx;
4794 invariant_ptr->insn = NULL;
4795 invariants.safe_push (invariant_ptr);
4796 *entry_ptr = (void *) invariant_ptr;
4798 return (invariant_ptr_t) *entry_ptr;
4801 /* Initiate the invariant table. */
4802 static void
4803 initiate_invariants (void)
4805 invariants.create (100);
4806 invariants_pool = new object_allocator<struct invariant> ("Inheritance invariants");
4807 invariant_table = htab_create (100, invariant_hash, invariant_eq_p, NULL);
4810 /* Finish the invariant table. */
4811 static void
4812 finish_invariants (void)
4814 htab_delete (invariant_table);
4815 delete invariants_pool;
4816 invariants.release ();
4819 /* Make the invariant table empty. */
4820 static void
4821 clear_invariants (void)
4823 htab_empty (invariant_table);
4824 invariants_pool->release ();
4825 invariants.truncate (0);
4830 /* This page contains code to do inheritance/split
4831 transformations. */
4833 /* Number of reloads passed so far in current EBB. */
4834 static int reloads_num;
4836 /* Number of calls passed so far in current EBB. */
4837 static int calls_num;
4839 /* Current reload pseudo check for validity of elements in
4840 USAGE_INSNS. */
4841 static int curr_usage_insns_check;
4843 /* Info about last usage of registers in EBB to do inheritance/split
4844 transformation. Inheritance transformation is done from a spilled
4845 pseudo and split transformations from a hard register or a pseudo
4846 assigned to a hard register. */
4847 struct usage_insns
4849 /* If the value is equal to CURR_USAGE_INSNS_CHECK, then the member
4850 value INSNS is valid. The insns is chain of optional debug insns
4851 and a finishing non-debug insn using the corresponding reg. The
4852 value is also used to mark the registers which are set up in the
4853 current insn. The negated insn uid is used for this. */
4854 int check;
4855 /* Value of global reloads_num at the last insn in INSNS. */
4856 int reloads_num;
4857 /* Value of global reloads_nums at the last insn in INSNS. */
4858 int calls_num;
4859 /* It can be true only for splitting. And it means that the restore
4860 insn should be put after insn given by the following member. */
4861 bool after_p;
4862 /* Next insns in the current EBB which use the original reg and the
4863 original reg value is not changed between the current insn and
4864 the next insns. In order words, e.g. for inheritance, if we need
4865 to use the original reg value again in the next insns we can try
4866 to use the value in a hard register from a reload insn of the
4867 current insn. */
4868 rtx insns;
4871 /* Map: regno -> corresponding pseudo usage insns. */
4872 static struct usage_insns *usage_insns;
4874 static void
4875 setup_next_usage_insn (int regno, rtx insn, int reloads_num, bool after_p)
4877 usage_insns[regno].check = curr_usage_insns_check;
4878 usage_insns[regno].insns = insn;
4879 usage_insns[regno].reloads_num = reloads_num;
4880 usage_insns[regno].calls_num = calls_num;
4881 usage_insns[regno].after_p = after_p;
4884 /* The function is used to form list REGNO usages which consists of
4885 optional debug insns finished by a non-debug insn using REGNO.
4886 RELOADS_NUM is current number of reload insns processed so far. */
4887 static void
4888 add_next_usage_insn (int regno, rtx_insn *insn, int reloads_num)
4890 rtx next_usage_insns;
4892 if (usage_insns[regno].check == curr_usage_insns_check
4893 && (next_usage_insns = usage_insns[regno].insns) != NULL_RTX
4894 && DEBUG_INSN_P (insn))
4896 /* Check that we did not add the debug insn yet. */
4897 if (next_usage_insns != insn
4898 && (GET_CODE (next_usage_insns) != INSN_LIST
4899 || XEXP (next_usage_insns, 0) != insn))
4900 usage_insns[regno].insns = gen_rtx_INSN_LIST (VOIDmode, insn,
4901 next_usage_insns);
4903 else if (NONDEBUG_INSN_P (insn))
4904 setup_next_usage_insn (regno, insn, reloads_num, false);
4905 else
4906 usage_insns[regno].check = 0;
4909 /* Return first non-debug insn in list USAGE_INSNS. */
4910 static rtx_insn *
4911 skip_usage_debug_insns (rtx usage_insns)
4913 rtx insn;
4915 /* Skip debug insns. */
4916 for (insn = usage_insns;
4917 insn != NULL_RTX && GET_CODE (insn) == INSN_LIST;
4918 insn = XEXP (insn, 1))
4920 return safe_as_a <rtx_insn *> (insn);
4923 /* Return true if we need secondary memory moves for insn in
4924 USAGE_INSNS after inserting inherited pseudo of class INHER_CL
4925 into the insn. */
4926 static bool
4927 check_secondary_memory_needed_p (enum reg_class inher_cl ATTRIBUTE_UNUSED,
4928 rtx usage_insns ATTRIBUTE_UNUSED)
4930 #ifndef SECONDARY_MEMORY_NEEDED
4931 return false;
4932 #else
4933 rtx_insn *insn;
4934 rtx set, dest;
4935 enum reg_class cl;
4937 if (inher_cl == ALL_REGS
4938 || (insn = skip_usage_debug_insns (usage_insns)) == NULL_RTX)
4939 return false;
4940 lra_assert (INSN_P (insn));
4941 if ((set = single_set (insn)) == NULL_RTX || ! REG_P (SET_DEST (set)))
4942 return false;
4943 dest = SET_DEST (set);
4944 if (! REG_P (dest))
4945 return false;
4946 lra_assert (inher_cl != NO_REGS);
4947 cl = get_reg_class (REGNO (dest));
4948 return (cl != NO_REGS && cl != ALL_REGS
4949 && SECONDARY_MEMORY_NEEDED (inher_cl, cl, GET_MODE (dest)));
4950 #endif
4953 /* Registers involved in inheritance/split in the current EBB
4954 (inheritance/split pseudos and original registers). */
4955 static bitmap_head check_only_regs;
4957 /* Reload pseudos can not be involded in invariant inheritance in the
4958 current EBB. */
4959 static bitmap_head invalid_invariant_regs;
4961 /* Do inheritance transformations for insn INSN, which defines (if
4962 DEF_P) or uses ORIGINAL_REGNO. NEXT_USAGE_INSNS specifies which
4963 instruction in the EBB next uses ORIGINAL_REGNO; it has the same
4964 form as the "insns" field of usage_insns. Return true if we
4965 succeed in such transformation.
4967 The transformations look like:
4969 p <- ... i <- ...
4970 ... p <- i (new insn)
4971 ... =>
4972 <- ... p ... <- ... i ...
4974 ... i <- p (new insn)
4975 <- ... p ... <- ... i ...
4976 ... =>
4977 <- ... p ... <- ... i ...
4978 where p is a spilled original pseudo and i is a new inheritance pseudo.
4981 The inheritance pseudo has the smallest class of two classes CL and
4982 class of ORIGINAL REGNO. */
4983 static bool
4984 inherit_reload_reg (bool def_p, int original_regno,
4985 enum reg_class cl, rtx_insn *insn, rtx next_usage_insns)
4987 if (optimize_function_for_size_p (cfun))
4988 return false;
4990 enum reg_class rclass = lra_get_allocno_class (original_regno);
4991 rtx original_reg = regno_reg_rtx[original_regno];
4992 rtx new_reg, usage_insn;
4993 rtx_insn *new_insns;
4995 lra_assert (! usage_insns[original_regno].after_p);
4996 if (lra_dump_file != NULL)
4997 fprintf (lra_dump_file,
4998 " <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<\n");
4999 if (! ira_reg_classes_intersect_p[cl][rclass])
5001 if (lra_dump_file != NULL)
5003 fprintf (lra_dump_file,
5004 " Rejecting inheritance for %d "
5005 "because of disjoint classes %s and %s\n",
5006 original_regno, reg_class_names[cl],
5007 reg_class_names[rclass]);
5008 fprintf (lra_dump_file,
5009 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5011 return false;
5013 if ((ira_class_subset_p[cl][rclass] && cl != rclass)
5014 /* We don't use a subset of two classes because it can be
5015 NO_REGS. This transformation is still profitable in most
5016 cases even if the classes are not intersected as register
5017 move is probably cheaper than a memory load. */
5018 || ira_class_hard_regs_num[cl] < ira_class_hard_regs_num[rclass])
5020 if (lra_dump_file != NULL)
5021 fprintf (lra_dump_file, " Use smallest class of %s and %s\n",
5022 reg_class_names[cl], reg_class_names[rclass]);
5024 rclass = cl;
5026 if (check_secondary_memory_needed_p (rclass, next_usage_insns))
5028 /* Reject inheritance resulting in secondary memory moves.
5029 Otherwise, there is a danger in LRA cycling. Also such
5030 transformation will be unprofitable. */
5031 if (lra_dump_file != NULL)
5033 rtx_insn *insn = skip_usage_debug_insns (next_usage_insns);
5034 rtx set = single_set (insn);
5036 lra_assert (set != NULL_RTX);
5038 rtx dest = SET_DEST (set);
5040 lra_assert (REG_P (dest));
5041 fprintf (lra_dump_file,
5042 " Rejecting inheritance for insn %d(%s)<-%d(%s) "
5043 "as secondary mem is needed\n",
5044 REGNO (dest), reg_class_names[get_reg_class (REGNO (dest))],
5045 original_regno, reg_class_names[rclass]);
5046 fprintf (lra_dump_file,
5047 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5049 return false;
5051 new_reg = lra_create_new_reg (GET_MODE (original_reg), original_reg,
5052 rclass, "inheritance");
5053 start_sequence ();
5054 if (def_p)
5055 lra_emit_move (original_reg, new_reg);
5056 else
5057 lra_emit_move (new_reg, original_reg);
5058 new_insns = get_insns ();
5059 end_sequence ();
5060 if (NEXT_INSN (new_insns) != NULL_RTX)
5062 if (lra_dump_file != NULL)
5064 fprintf (lra_dump_file,
5065 " Rejecting inheritance %d->%d "
5066 "as it results in 2 or more insns:\n",
5067 original_regno, REGNO (new_reg));
5068 dump_rtl_slim (lra_dump_file, new_insns, NULL, -1, 0);
5069 fprintf (lra_dump_file,
5070 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5072 return false;
5074 lra_substitute_pseudo_within_insn (insn, original_regno, new_reg, false);
5075 lra_update_insn_regno_info (insn);
5076 if (! def_p)
5077 /* We now have a new usage insn for original regno. */
5078 setup_next_usage_insn (original_regno, new_insns, reloads_num, false);
5079 if (lra_dump_file != NULL)
5080 fprintf (lra_dump_file, " Original reg change %d->%d (bb%d):\n",
5081 original_regno, REGNO (new_reg), BLOCK_FOR_INSN (insn)->index);
5082 lra_reg_info[REGNO (new_reg)].restore_rtx = regno_reg_rtx[original_regno];
5083 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
5084 bitmap_set_bit (&check_only_regs, original_regno);
5085 bitmap_set_bit (&lra_inheritance_pseudos, REGNO (new_reg));
5086 if (def_p)
5087 lra_process_new_insns (insn, NULL, new_insns,
5088 "Add original<-inheritance");
5089 else
5090 lra_process_new_insns (insn, new_insns, NULL,
5091 "Add inheritance<-original");
5092 while (next_usage_insns != NULL_RTX)
5094 if (GET_CODE (next_usage_insns) != INSN_LIST)
5096 usage_insn = next_usage_insns;
5097 lra_assert (NONDEBUG_INSN_P (usage_insn));
5098 next_usage_insns = NULL;
5100 else
5102 usage_insn = XEXP (next_usage_insns, 0);
5103 lra_assert (DEBUG_INSN_P (usage_insn));
5104 next_usage_insns = XEXP (next_usage_insns, 1);
5106 lra_substitute_pseudo (&usage_insn, original_regno, new_reg, false);
5107 lra_update_insn_regno_info (as_a <rtx_insn *> (usage_insn));
5108 if (lra_dump_file != NULL)
5110 fprintf (lra_dump_file,
5111 " Inheritance reuse change %d->%d (bb%d):\n",
5112 original_regno, REGNO (new_reg),
5113 BLOCK_FOR_INSN (usage_insn)->index);
5114 dump_insn_slim (lra_dump_file, as_a <rtx_insn *> (usage_insn));
5117 if (lra_dump_file != NULL)
5118 fprintf (lra_dump_file,
5119 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5120 return true;
5123 /* Return true if we need a caller save/restore for pseudo REGNO which
5124 was assigned to a hard register. */
5125 static inline bool
5126 need_for_call_save_p (int regno)
5128 lra_assert (regno >= FIRST_PSEUDO_REGISTER && reg_renumber[regno] >= 0);
5129 return (usage_insns[regno].calls_num < calls_num
5130 && (overlaps_hard_reg_set_p
5131 ((flag_ipa_ra &&
5132 ! hard_reg_set_empty_p (lra_reg_info[regno].actual_call_used_reg_set))
5133 ? lra_reg_info[regno].actual_call_used_reg_set
5134 : call_used_reg_set,
5135 PSEUDO_REGNO_MODE (regno), reg_renumber[regno])
5136 || HARD_REGNO_CALL_PART_CLOBBERED (reg_renumber[regno],
5137 PSEUDO_REGNO_MODE (regno))));
5140 /* Global registers occurring in the current EBB. */
5141 static bitmap_head ebb_global_regs;
5143 /* Return true if we need a split for hard register REGNO or pseudo
5144 REGNO which was assigned to a hard register.
5145 POTENTIAL_RELOAD_HARD_REGS contains hard registers which might be
5146 used for reloads since the EBB end. It is an approximation of the
5147 used hard registers in the split range. The exact value would
5148 require expensive calculations. If we were aggressive with
5149 splitting because of the approximation, the split pseudo will save
5150 the same hard register assignment and will be removed in the undo
5151 pass. We still need the approximation because too aggressive
5152 splitting would result in too inaccurate cost calculation in the
5153 assignment pass because of too many generated moves which will be
5154 probably removed in the undo pass. */
5155 static inline bool
5156 need_for_split_p (HARD_REG_SET potential_reload_hard_regs, int regno)
5158 int hard_regno = regno < FIRST_PSEUDO_REGISTER ? regno : reg_renumber[regno];
5160 lra_assert (hard_regno >= 0);
5161 return ((TEST_HARD_REG_BIT (potential_reload_hard_regs, hard_regno)
5162 /* Don't split eliminable hard registers, otherwise we can
5163 split hard registers like hard frame pointer, which
5164 lives on BB start/end according to DF-infrastructure,
5165 when there is a pseudo assigned to the register and
5166 living in the same BB. */
5167 && (regno >= FIRST_PSEUDO_REGISTER
5168 || ! TEST_HARD_REG_BIT (eliminable_regset, hard_regno))
5169 && ! TEST_HARD_REG_BIT (lra_no_alloc_regs, hard_regno)
5170 /* Don't split call clobbered hard regs living through
5171 calls, otherwise we might have a check problem in the
5172 assign sub-pass as in the most cases (exception is a
5173 situation when lra_risky_transformations_p value is
5174 true) the assign pass assumes that all pseudos living
5175 through calls are assigned to call saved hard regs. */
5176 && (regno >= FIRST_PSEUDO_REGISTER
5177 || ! TEST_HARD_REG_BIT (call_used_reg_set, regno)
5178 || usage_insns[regno].calls_num == calls_num)
5179 /* We need at least 2 reloads to make pseudo splitting
5180 profitable. We should provide hard regno splitting in
5181 any case to solve 1st insn scheduling problem when
5182 moving hard register definition up might result in
5183 impossibility to find hard register for reload pseudo of
5184 small register class. */
5185 && (usage_insns[regno].reloads_num
5186 + (regno < FIRST_PSEUDO_REGISTER ? 0 : 3) < reloads_num)
5187 && (regno < FIRST_PSEUDO_REGISTER
5188 /* For short living pseudos, spilling + inheritance can
5189 be considered a substitution for splitting.
5190 Therefore we do not splitting for local pseudos. It
5191 decreases also aggressiveness of splitting. The
5192 minimal number of references is chosen taking into
5193 account that for 2 references splitting has no sense
5194 as we can just spill the pseudo. */
5195 || (regno >= FIRST_PSEUDO_REGISTER
5196 && lra_reg_info[regno].nrefs > 3
5197 && bitmap_bit_p (&ebb_global_regs, regno))))
5198 || (regno >= FIRST_PSEUDO_REGISTER && need_for_call_save_p (regno)));
5201 /* Return class for the split pseudo created from original pseudo with
5202 ALLOCNO_CLASS and MODE which got a hard register HARD_REGNO. We
5203 choose subclass of ALLOCNO_CLASS which contains HARD_REGNO and
5204 results in no secondary memory movements. */
5205 static enum reg_class
5206 choose_split_class (enum reg_class allocno_class,
5207 int hard_regno ATTRIBUTE_UNUSED,
5208 machine_mode mode ATTRIBUTE_UNUSED)
5210 #ifndef SECONDARY_MEMORY_NEEDED
5211 return allocno_class;
5212 #else
5213 int i;
5214 enum reg_class cl, best_cl = NO_REGS;
5215 enum reg_class hard_reg_class ATTRIBUTE_UNUSED
5216 = REGNO_REG_CLASS (hard_regno);
5218 if (! SECONDARY_MEMORY_NEEDED (allocno_class, allocno_class, mode)
5219 && TEST_HARD_REG_BIT (reg_class_contents[allocno_class], hard_regno))
5220 return allocno_class;
5221 for (i = 0;
5222 (cl = reg_class_subclasses[allocno_class][i]) != LIM_REG_CLASSES;
5223 i++)
5224 if (! SECONDARY_MEMORY_NEEDED (cl, hard_reg_class, mode)
5225 && ! SECONDARY_MEMORY_NEEDED (hard_reg_class, cl, mode)
5226 && TEST_HARD_REG_BIT (reg_class_contents[cl], hard_regno)
5227 && (best_cl == NO_REGS
5228 || ira_class_hard_regs_num[best_cl] < ira_class_hard_regs_num[cl]))
5229 best_cl = cl;
5230 return best_cl;
5231 #endif
5234 /* Do split transformations for insn INSN, which defines or uses
5235 ORIGINAL_REGNO. NEXT_USAGE_INSNS specifies which instruction in
5236 the EBB next uses ORIGINAL_REGNO; it has the same form as the
5237 "insns" field of usage_insns.
5239 The transformations look like:
5241 p <- ... p <- ...
5242 ... s <- p (new insn -- save)
5243 ... =>
5244 ... p <- s (new insn -- restore)
5245 <- ... p ... <- ... p ...
5247 <- ... p ... <- ... p ...
5248 ... s <- p (new insn -- save)
5249 ... =>
5250 ... p <- s (new insn -- restore)
5251 <- ... p ... <- ... p ...
5253 where p is an original pseudo got a hard register or a hard
5254 register and s is a new split pseudo. The save is put before INSN
5255 if BEFORE_P is true. Return true if we succeed in such
5256 transformation. */
5257 static bool
5258 split_reg (bool before_p, int original_regno, rtx_insn *insn,
5259 rtx next_usage_insns)
5261 enum reg_class rclass;
5262 rtx original_reg;
5263 int hard_regno, nregs;
5264 rtx new_reg, usage_insn;
5265 rtx_insn *restore, *save;
5266 bool after_p;
5267 bool call_save_p;
5268 machine_mode mode;
5270 if (original_regno < FIRST_PSEUDO_REGISTER)
5272 rclass = ira_allocno_class_translate[REGNO_REG_CLASS (original_regno)];
5273 hard_regno = original_regno;
5274 call_save_p = false;
5275 nregs = 1;
5276 mode = lra_reg_info[hard_regno].biggest_mode;
5277 machine_mode reg_rtx_mode = GET_MODE (regno_reg_rtx[hard_regno]);
5278 /* A reg can have a biggest_mode of VOIDmode if it was only ever seen
5279 as part of a multi-word register. In that case, or if the biggest
5280 mode was larger than a register, just use the reg_rtx. Otherwise,
5281 limit the size to that of the biggest access in the function. */
5282 if (mode == VOIDmode
5283 || GET_MODE_SIZE (mode) > GET_MODE_SIZE (reg_rtx_mode))
5285 original_reg = regno_reg_rtx[hard_regno];
5286 mode = reg_rtx_mode;
5288 else
5289 original_reg = gen_rtx_REG (mode, hard_regno);
5291 else
5293 mode = PSEUDO_REGNO_MODE (original_regno);
5294 hard_regno = reg_renumber[original_regno];
5295 nregs = hard_regno_nregs[hard_regno][mode];
5296 rclass = lra_get_allocno_class (original_regno);
5297 original_reg = regno_reg_rtx[original_regno];
5298 call_save_p = need_for_call_save_p (original_regno);
5300 lra_assert (hard_regno >= 0);
5301 if (lra_dump_file != NULL)
5302 fprintf (lra_dump_file,
5303 " ((((((((((((((((((((((((((((((((((((((((((((((((\n");
5305 if (call_save_p)
5307 mode = HARD_REGNO_CALLER_SAVE_MODE (hard_regno,
5308 hard_regno_nregs[hard_regno][mode],
5309 mode);
5310 new_reg = lra_create_new_reg (mode, NULL_RTX, NO_REGS, "save");
5312 else
5314 rclass = choose_split_class (rclass, hard_regno, mode);
5315 if (rclass == NO_REGS)
5317 if (lra_dump_file != NULL)
5319 fprintf (lra_dump_file,
5320 " Rejecting split of %d(%s): "
5321 "no good reg class for %d(%s)\n",
5322 original_regno,
5323 reg_class_names[lra_get_allocno_class (original_regno)],
5324 hard_regno,
5325 reg_class_names[REGNO_REG_CLASS (hard_regno)]);
5326 fprintf
5327 (lra_dump_file,
5328 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5330 return false;
5332 new_reg = lra_create_new_reg (mode, original_reg, rclass, "split");
5333 reg_renumber[REGNO (new_reg)] = hard_regno;
5335 save = emit_spill_move (true, new_reg, original_reg);
5336 if (NEXT_INSN (save) != NULL_RTX && !call_save_p)
5338 if (lra_dump_file != NULL)
5340 fprintf
5341 (lra_dump_file,
5342 " Rejecting split %d->%d resulting in > 2 save insns:\n",
5343 original_regno, REGNO (new_reg));
5344 dump_rtl_slim (lra_dump_file, save, NULL, -1, 0);
5345 fprintf (lra_dump_file,
5346 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5348 return false;
5350 restore = emit_spill_move (false, new_reg, original_reg);
5351 if (NEXT_INSN (restore) != NULL_RTX && !call_save_p)
5353 if (lra_dump_file != NULL)
5355 fprintf (lra_dump_file,
5356 " Rejecting split %d->%d "
5357 "resulting in > 2 restore insns:\n",
5358 original_regno, REGNO (new_reg));
5359 dump_rtl_slim (lra_dump_file, restore, NULL, -1, 0);
5360 fprintf (lra_dump_file,
5361 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5363 return false;
5365 after_p = usage_insns[original_regno].after_p;
5366 lra_reg_info[REGNO (new_reg)].restore_rtx = regno_reg_rtx[original_regno];
5367 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
5368 bitmap_set_bit (&check_only_regs, original_regno);
5369 bitmap_set_bit (&lra_split_regs, REGNO (new_reg));
5370 for (;;)
5372 if (GET_CODE (next_usage_insns) != INSN_LIST)
5374 usage_insn = next_usage_insns;
5375 break;
5377 usage_insn = XEXP (next_usage_insns, 0);
5378 lra_assert (DEBUG_INSN_P (usage_insn));
5379 next_usage_insns = XEXP (next_usage_insns, 1);
5380 lra_substitute_pseudo (&usage_insn, original_regno, new_reg, false);
5381 lra_update_insn_regno_info (as_a <rtx_insn *> (usage_insn));
5382 if (lra_dump_file != NULL)
5384 fprintf (lra_dump_file, " Split reuse change %d->%d:\n",
5385 original_regno, REGNO (new_reg));
5386 dump_insn_slim (lra_dump_file, as_a <rtx_insn *> (usage_insn));
5389 lra_assert (NOTE_P (usage_insn) || NONDEBUG_INSN_P (usage_insn));
5390 lra_assert (usage_insn != insn || (after_p && before_p));
5391 lra_process_new_insns (as_a <rtx_insn *> (usage_insn),
5392 after_p ? NULL : restore,
5393 after_p ? restore : NULL,
5394 call_save_p
5395 ? "Add reg<-save" : "Add reg<-split");
5396 lra_process_new_insns (insn, before_p ? save : NULL,
5397 before_p ? NULL : save,
5398 call_save_p
5399 ? "Add save<-reg" : "Add split<-reg");
5400 if (nregs > 1)
5401 /* If we are trying to split multi-register. We should check
5402 conflicts on the next assignment sub-pass. IRA can allocate on
5403 sub-register levels, LRA do this on pseudos level right now and
5404 this discrepancy may create allocation conflicts after
5405 splitting. */
5406 lra_risky_transformations_p = true;
5407 if (lra_dump_file != NULL)
5408 fprintf (lra_dump_file,
5409 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5410 return true;
5413 /* Recognize that we need a split transformation for insn INSN, which
5414 defines or uses REGNO in its insn biggest MODE (we use it only if
5415 REGNO is a hard register). POTENTIAL_RELOAD_HARD_REGS contains
5416 hard registers which might be used for reloads since the EBB end.
5417 Put the save before INSN if BEFORE_P is true. MAX_UID is maximla
5418 uid before starting INSN processing. Return true if we succeed in
5419 such transformation. */
5420 static bool
5421 split_if_necessary (int regno, machine_mode mode,
5422 HARD_REG_SET potential_reload_hard_regs,
5423 bool before_p, rtx_insn *insn, int max_uid)
5425 bool res = false;
5426 int i, nregs = 1;
5427 rtx next_usage_insns;
5429 if (regno < FIRST_PSEUDO_REGISTER)
5430 nregs = hard_regno_nregs[regno][mode];
5431 for (i = 0; i < nregs; i++)
5432 if (usage_insns[regno + i].check == curr_usage_insns_check
5433 && (next_usage_insns = usage_insns[regno + i].insns) != NULL_RTX
5434 /* To avoid processing the register twice or more. */
5435 && ((GET_CODE (next_usage_insns) != INSN_LIST
5436 && INSN_UID (next_usage_insns) < max_uid)
5437 || (GET_CODE (next_usage_insns) == INSN_LIST
5438 && (INSN_UID (XEXP (next_usage_insns, 0)) < max_uid)))
5439 && need_for_split_p (potential_reload_hard_regs, regno + i)
5440 && split_reg (before_p, regno + i, insn, next_usage_insns))
5441 res = true;
5442 return res;
5445 /* Return TRUE if rtx X is considered as an invariant for
5446 inheritance. */
5447 static bool
5448 invariant_p (const_rtx x)
5450 machine_mode mode;
5451 const char *fmt;
5452 enum rtx_code code;
5453 int i, j;
5455 code = GET_CODE (x);
5456 mode = GET_MODE (x);
5457 if (code == SUBREG)
5459 x = SUBREG_REG (x);
5460 code = GET_CODE (x);
5461 if (GET_MODE_SIZE (GET_MODE (x)) > GET_MODE_SIZE (mode))
5462 mode = GET_MODE (x);
5465 if (MEM_P (x))
5466 return false;
5468 if (REG_P (x))
5470 int i, nregs, regno = REGNO (x);
5472 if (regno >= FIRST_PSEUDO_REGISTER || regno == STACK_POINTER_REGNUM
5473 || TEST_HARD_REG_BIT (eliminable_regset, regno)
5474 || GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
5475 return false;
5476 nregs = hard_regno_nregs[regno][mode];
5477 for (i = 0; i < nregs; i++)
5478 if (! fixed_regs[regno + i]
5479 /* A hard register may be clobbered in the current insn
5480 but we can ignore this case because if the hard
5481 register is used it should be set somewhere after the
5482 clobber. */
5483 || bitmap_bit_p (&invalid_invariant_regs, regno + i))
5484 return false;
5486 fmt = GET_RTX_FORMAT (code);
5487 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
5489 if (fmt[i] == 'e')
5491 if (! invariant_p (XEXP (x, i)))
5492 return false;
5494 else if (fmt[i] == 'E')
5496 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
5497 if (! invariant_p (XVECEXP (x, i, j)))
5498 return false;
5501 return true;
5504 /* We have 'dest_reg <- invariant'. Let us try to make an invariant
5505 inheritance transformation (using dest_reg instead invariant in a
5506 subsequent insn). */
5507 static bool
5508 process_invariant_for_inheritance (rtx dst_reg, rtx invariant_rtx)
5510 invariant_ptr_t invariant_ptr;
5511 rtx_insn *insn, *new_insns;
5512 rtx insn_set, insn_reg, new_reg;
5513 int insn_regno;
5514 bool succ_p = false;
5515 int dst_regno = REGNO (dst_reg);
5516 enum machine_mode dst_mode = GET_MODE (dst_reg);
5517 enum reg_class cl = lra_get_allocno_class (dst_regno), insn_reg_cl;
5519 invariant_ptr = insert_invariant (invariant_rtx);
5520 if ((insn = invariant_ptr->insn) != NULL_RTX)
5522 /* We have a subsequent insn using the invariant. */
5523 insn_set = single_set (insn);
5524 lra_assert (insn_set != NULL);
5525 insn_reg = SET_DEST (insn_set);
5526 lra_assert (REG_P (insn_reg));
5527 insn_regno = REGNO (insn_reg);
5528 insn_reg_cl = lra_get_allocno_class (insn_regno);
5530 if (dst_mode == GET_MODE (insn_reg)
5531 /* We should consider only result move reg insns which are
5532 cheap. */
5533 && targetm.register_move_cost (dst_mode, cl, insn_reg_cl) == 2
5534 && targetm.register_move_cost (dst_mode, cl, cl) == 2)
5536 if (lra_dump_file != NULL)
5537 fprintf (lra_dump_file,
5538 " [[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[\n");
5539 new_reg = lra_create_new_reg (dst_mode, dst_reg,
5540 cl, "invariant inheritance");
5541 bitmap_set_bit (&lra_inheritance_pseudos, REGNO (new_reg));
5542 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
5543 lra_reg_info[REGNO (new_reg)].restore_rtx = PATTERN (insn);
5544 start_sequence ();
5545 lra_emit_move (new_reg, dst_reg);
5546 new_insns = get_insns ();
5547 end_sequence ();
5548 lra_process_new_insns (curr_insn, NULL, new_insns,
5549 "Add invariant inheritance<-original");
5550 start_sequence ();
5551 lra_emit_move (SET_DEST (insn_set), new_reg);
5552 new_insns = get_insns ();
5553 end_sequence ();
5554 lra_process_new_insns (insn, NULL, new_insns,
5555 "Changing reload<-inheritance");
5556 lra_set_insn_deleted (insn);
5557 succ_p = true;
5558 if (lra_dump_file != NULL)
5560 fprintf (lra_dump_file,
5561 " Invariant inheritance reuse change %d (bb%d):\n",
5562 REGNO (new_reg), BLOCK_FOR_INSN (insn)->index);
5563 dump_insn_slim (lra_dump_file, insn);
5564 fprintf (lra_dump_file,
5565 " ]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]\n");
5569 invariant_ptr->insn = curr_insn;
5570 return succ_p;
5573 /* Check only registers living at the current program point in the
5574 current EBB. */
5575 static bitmap_head live_regs;
5577 /* Update live info in EBB given by its HEAD and TAIL insns after
5578 inheritance/split transformation. The function removes dead moves
5579 too. */
5580 static void
5581 update_ebb_live_info (rtx_insn *head, rtx_insn *tail)
5583 unsigned int j;
5584 int i, regno;
5585 bool live_p;
5586 rtx_insn *prev_insn;
5587 rtx set;
5588 bool remove_p;
5589 basic_block last_bb, prev_bb, curr_bb;
5590 bitmap_iterator bi;
5591 struct lra_insn_reg *reg;
5592 edge e;
5593 edge_iterator ei;
5595 last_bb = BLOCK_FOR_INSN (tail);
5596 prev_bb = NULL;
5597 for (curr_insn = tail;
5598 curr_insn != PREV_INSN (head);
5599 curr_insn = prev_insn)
5601 prev_insn = PREV_INSN (curr_insn);
5602 /* We need to process empty blocks too. They contain
5603 NOTE_INSN_BASIC_BLOCK referring for the basic block. */
5604 if (NOTE_P (curr_insn) && NOTE_KIND (curr_insn) != NOTE_INSN_BASIC_BLOCK)
5605 continue;
5606 curr_bb = BLOCK_FOR_INSN (curr_insn);
5607 if (curr_bb != prev_bb)
5609 if (prev_bb != NULL)
5611 /* Update df_get_live_in (prev_bb): */
5612 EXECUTE_IF_SET_IN_BITMAP (&check_only_regs, 0, j, bi)
5613 if (bitmap_bit_p (&live_regs, j))
5614 bitmap_set_bit (df_get_live_in (prev_bb), j);
5615 else
5616 bitmap_clear_bit (df_get_live_in (prev_bb), j);
5618 if (curr_bb != last_bb)
5620 /* Update df_get_live_out (curr_bb): */
5621 EXECUTE_IF_SET_IN_BITMAP (&check_only_regs, 0, j, bi)
5623 live_p = bitmap_bit_p (&live_regs, j);
5624 if (! live_p)
5625 FOR_EACH_EDGE (e, ei, curr_bb->succs)
5626 if (bitmap_bit_p (df_get_live_in (e->dest), j))
5628 live_p = true;
5629 break;
5631 if (live_p)
5632 bitmap_set_bit (df_get_live_out (curr_bb), j);
5633 else
5634 bitmap_clear_bit (df_get_live_out (curr_bb), j);
5637 prev_bb = curr_bb;
5638 bitmap_and (&live_regs, &check_only_regs, df_get_live_out (curr_bb));
5640 if (! NONDEBUG_INSN_P (curr_insn))
5641 continue;
5642 curr_id = lra_get_insn_recog_data (curr_insn);
5643 curr_static_id = curr_id->insn_static_data;
5644 remove_p = false;
5645 if ((set = single_set (curr_insn)) != NULL_RTX
5646 && REG_P (SET_DEST (set))
5647 && (regno = REGNO (SET_DEST (set))) >= FIRST_PSEUDO_REGISTER
5648 && SET_DEST (set) != pic_offset_table_rtx
5649 && bitmap_bit_p (&check_only_regs, regno)
5650 && ! bitmap_bit_p (&live_regs, regno))
5651 remove_p = true;
5652 /* See which defined values die here. */
5653 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5654 if (reg->type == OP_OUT && ! reg->subreg_p)
5655 bitmap_clear_bit (&live_regs, reg->regno);
5656 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
5657 if (reg->type == OP_OUT && ! reg->subreg_p)
5658 bitmap_clear_bit (&live_regs, reg->regno);
5659 if (curr_id->arg_hard_regs != NULL)
5660 /* Make clobbered argument hard registers die. */
5661 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
5662 if (regno >= FIRST_PSEUDO_REGISTER)
5663 bitmap_clear_bit (&live_regs, regno - FIRST_PSEUDO_REGISTER);
5664 /* Mark each used value as live. */
5665 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5666 if (reg->type != OP_OUT
5667 && bitmap_bit_p (&check_only_regs, reg->regno))
5668 bitmap_set_bit (&live_regs, reg->regno);
5669 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
5670 if (reg->type != OP_OUT
5671 && bitmap_bit_p (&check_only_regs, reg->regno))
5672 bitmap_set_bit (&live_regs, reg->regno);
5673 if (curr_id->arg_hard_regs != NULL)
5674 /* Make used argument hard registers live. */
5675 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
5676 if (regno < FIRST_PSEUDO_REGISTER
5677 && bitmap_bit_p (&check_only_regs, regno))
5678 bitmap_set_bit (&live_regs, regno);
5679 /* It is quite important to remove dead move insns because it
5680 means removing dead store. We don't need to process them for
5681 constraints. */
5682 if (remove_p)
5684 if (lra_dump_file != NULL)
5686 fprintf (lra_dump_file, " Removing dead insn:\n ");
5687 dump_insn_slim (lra_dump_file, curr_insn);
5689 lra_set_insn_deleted (curr_insn);
5694 /* The structure describes info to do an inheritance for the current
5695 insn. We need to collect such info first before doing the
5696 transformations because the transformations change the insn
5697 internal representation. */
5698 struct to_inherit
5700 /* Original regno. */
5701 int regno;
5702 /* Subsequent insns which can inherit original reg value. */
5703 rtx insns;
5706 /* Array containing all info for doing inheritance from the current
5707 insn. */
5708 static struct to_inherit to_inherit[LRA_MAX_INSN_RELOADS];
5710 /* Number elements in the previous array. */
5711 static int to_inherit_num;
5713 /* Add inheritance info REGNO and INSNS. Their meaning is described in
5714 structure to_inherit. */
5715 static void
5716 add_to_inherit (int regno, rtx insns)
5718 int i;
5720 for (i = 0; i < to_inherit_num; i++)
5721 if (to_inherit[i].regno == regno)
5722 return;
5723 lra_assert (to_inherit_num < LRA_MAX_INSN_RELOADS);
5724 to_inherit[to_inherit_num].regno = regno;
5725 to_inherit[to_inherit_num++].insns = insns;
5728 /* Return the last non-debug insn in basic block BB, or the block begin
5729 note if none. */
5730 static rtx_insn *
5731 get_last_insertion_point (basic_block bb)
5733 rtx_insn *insn;
5735 FOR_BB_INSNS_REVERSE (bb, insn)
5736 if (NONDEBUG_INSN_P (insn) || NOTE_INSN_BASIC_BLOCK_P (insn))
5737 return insn;
5738 gcc_unreachable ();
5741 /* Set up RES by registers living on edges FROM except the edge (FROM,
5742 TO) or by registers set up in a jump insn in BB FROM. */
5743 static void
5744 get_live_on_other_edges (basic_block from, basic_block to, bitmap res)
5746 rtx_insn *last;
5747 struct lra_insn_reg *reg;
5748 edge e;
5749 edge_iterator ei;
5751 lra_assert (to != NULL);
5752 bitmap_clear (res);
5753 FOR_EACH_EDGE (e, ei, from->succs)
5754 if (e->dest != to)
5755 bitmap_ior_into (res, df_get_live_in (e->dest));
5756 last = get_last_insertion_point (from);
5757 if (! JUMP_P (last))
5758 return;
5759 curr_id = lra_get_insn_recog_data (last);
5760 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5761 if (reg->type != OP_IN)
5762 bitmap_set_bit (res, reg->regno);
5765 /* Used as a temporary results of some bitmap calculations. */
5766 static bitmap_head temp_bitmap;
5768 /* We split for reloads of small class of hard regs. The following
5769 defines how many hard regs the class should have to be qualified as
5770 small. The code is mostly oriented to x86/x86-64 architecture
5771 where some insns need to use only specific register or pair of
5772 registers and these register can live in RTL explicitly, e.g. for
5773 parameter passing. */
5774 static const int max_small_class_regs_num = 2;
5776 /* Do inheritance/split transformations in EBB starting with HEAD and
5777 finishing on TAIL. We process EBB insns in the reverse order.
5778 Return true if we did any inheritance/split transformation in the
5779 EBB.
5781 We should avoid excessive splitting which results in worse code
5782 because of inaccurate cost calculations for spilling new split
5783 pseudos in such case. To achieve this we do splitting only if
5784 register pressure is high in given basic block and there are reload
5785 pseudos requiring hard registers. We could do more register
5786 pressure calculations at any given program point to avoid necessary
5787 splitting even more but it is to expensive and the current approach
5788 works well enough. */
5789 static bool
5790 inherit_in_ebb (rtx_insn *head, rtx_insn *tail)
5792 int i, src_regno, dst_regno, nregs;
5793 bool change_p, succ_p, update_reloads_num_p;
5794 rtx_insn *prev_insn, *last_insn;
5795 rtx next_usage_insns, curr_set;
5796 enum reg_class cl;
5797 struct lra_insn_reg *reg;
5798 basic_block last_processed_bb, curr_bb = NULL;
5799 HARD_REG_SET potential_reload_hard_regs, live_hard_regs;
5800 bitmap to_process;
5801 unsigned int j;
5802 bitmap_iterator bi;
5803 bool head_p, after_p;
5805 change_p = false;
5806 curr_usage_insns_check++;
5807 clear_invariants ();
5808 reloads_num = calls_num = 0;
5809 bitmap_clear (&check_only_regs);
5810 bitmap_clear (&invalid_invariant_regs);
5811 last_processed_bb = NULL;
5812 CLEAR_HARD_REG_SET (potential_reload_hard_regs);
5813 COPY_HARD_REG_SET (live_hard_regs, eliminable_regset);
5814 IOR_HARD_REG_SET (live_hard_regs, lra_no_alloc_regs);
5815 /* We don't process new insns generated in the loop. */
5816 for (curr_insn = tail; curr_insn != PREV_INSN (head); curr_insn = prev_insn)
5818 prev_insn = PREV_INSN (curr_insn);
5819 if (BLOCK_FOR_INSN (curr_insn) != NULL)
5820 curr_bb = BLOCK_FOR_INSN (curr_insn);
5821 if (last_processed_bb != curr_bb)
5823 /* We are at the end of BB. Add qualified living
5824 pseudos for potential splitting. */
5825 to_process = df_get_live_out (curr_bb);
5826 if (last_processed_bb != NULL)
5828 /* We are somewhere in the middle of EBB. */
5829 get_live_on_other_edges (curr_bb, last_processed_bb,
5830 &temp_bitmap);
5831 to_process = &temp_bitmap;
5833 last_processed_bb = curr_bb;
5834 last_insn = get_last_insertion_point (curr_bb);
5835 after_p = (! JUMP_P (last_insn)
5836 && (! CALL_P (last_insn)
5837 || (find_reg_note (last_insn,
5838 REG_NORETURN, NULL_RTX) == NULL_RTX
5839 && ! SIBLING_CALL_P (last_insn))));
5840 CLEAR_HARD_REG_SET (potential_reload_hard_regs);
5841 EXECUTE_IF_SET_IN_BITMAP (to_process, 0, j, bi)
5843 if ((int) j >= lra_constraint_new_regno_start)
5844 break;
5845 if (j < FIRST_PSEUDO_REGISTER || reg_renumber[j] >= 0)
5847 if (j < FIRST_PSEUDO_REGISTER)
5848 SET_HARD_REG_BIT (live_hard_regs, j);
5849 else
5850 add_to_hard_reg_set (&live_hard_regs,
5851 PSEUDO_REGNO_MODE (j),
5852 reg_renumber[j]);
5853 setup_next_usage_insn (j, last_insn, reloads_num, after_p);
5857 src_regno = dst_regno = -1;
5858 curr_set = single_set (curr_insn);
5859 if (curr_set != NULL_RTX && REG_P (SET_DEST (curr_set)))
5860 dst_regno = REGNO (SET_DEST (curr_set));
5861 if (curr_set != NULL_RTX && REG_P (SET_SRC (curr_set)))
5862 src_regno = REGNO (SET_SRC (curr_set));
5863 update_reloads_num_p = true;
5864 if (src_regno < lra_constraint_new_regno_start
5865 && src_regno >= FIRST_PSEUDO_REGISTER
5866 && reg_renumber[src_regno] < 0
5867 && dst_regno >= lra_constraint_new_regno_start
5868 && (cl = lra_get_allocno_class (dst_regno)) != NO_REGS)
5870 /* 'reload_pseudo <- original_pseudo'. */
5871 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
5872 reloads_num++;
5873 update_reloads_num_p = false;
5874 succ_p = false;
5875 if (usage_insns[src_regno].check == curr_usage_insns_check
5876 && (next_usage_insns = usage_insns[src_regno].insns) != NULL_RTX)
5877 succ_p = inherit_reload_reg (false, src_regno, cl,
5878 curr_insn, next_usage_insns);
5879 if (succ_p)
5880 change_p = true;
5881 else
5882 setup_next_usage_insn (src_regno, curr_insn, reloads_num, false);
5883 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
5884 IOR_HARD_REG_SET (potential_reload_hard_regs,
5885 reg_class_contents[cl]);
5887 else if (src_regno < 0
5888 && dst_regno >= lra_constraint_new_regno_start
5889 && invariant_p (SET_SRC (curr_set))
5890 && (cl = lra_get_allocno_class (dst_regno)) != NO_REGS
5891 && ! bitmap_bit_p (&invalid_invariant_regs, dst_regno)
5892 && ! bitmap_bit_p (&invalid_invariant_regs,
5893 ORIGINAL_REGNO(regno_reg_rtx[dst_regno])))
5895 /* 'reload_pseudo <- invariant'. */
5896 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
5897 reloads_num++;
5898 update_reloads_num_p = false;
5899 if (process_invariant_for_inheritance (SET_DEST (curr_set), SET_SRC (curr_set)))
5900 change_p = true;
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 >= lra_constraint_new_regno_start
5906 && dst_regno < lra_constraint_new_regno_start
5907 && dst_regno >= FIRST_PSEUDO_REGISTER
5908 && reg_renumber[dst_regno] < 0
5909 && (cl = lra_get_allocno_class (src_regno)) != NO_REGS
5910 && usage_insns[dst_regno].check == curr_usage_insns_check
5911 && (next_usage_insns
5912 = usage_insns[dst_regno].insns) != NULL_RTX)
5914 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
5915 reloads_num++;
5916 update_reloads_num_p = false;
5917 /* 'original_pseudo <- reload_pseudo'. */
5918 if (! JUMP_P (curr_insn)
5919 && inherit_reload_reg (true, dst_regno, cl,
5920 curr_insn, next_usage_insns))
5921 change_p = true;
5922 /* Invalidate. */
5923 usage_insns[dst_regno].check = 0;
5924 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
5925 IOR_HARD_REG_SET (potential_reload_hard_regs,
5926 reg_class_contents[cl]);
5928 else if (INSN_P (curr_insn))
5930 int iter;
5931 int max_uid = get_max_uid ();
5933 curr_id = lra_get_insn_recog_data (curr_insn);
5934 curr_static_id = curr_id->insn_static_data;
5935 to_inherit_num = 0;
5936 /* Process insn definitions. */
5937 for (iter = 0; iter < 2; iter++)
5938 for (reg = iter == 0 ? curr_id->regs : curr_static_id->hard_regs;
5939 reg != NULL;
5940 reg = reg->next)
5941 if (reg->type != OP_IN
5942 && (dst_regno = reg->regno) < lra_constraint_new_regno_start)
5944 if (dst_regno >= FIRST_PSEUDO_REGISTER && reg->type == OP_OUT
5945 && reg_renumber[dst_regno] < 0 && ! reg->subreg_p
5946 && usage_insns[dst_regno].check == curr_usage_insns_check
5947 && (next_usage_insns
5948 = usage_insns[dst_regno].insns) != NULL_RTX)
5950 struct lra_insn_reg *r;
5952 for (r = curr_id->regs; r != NULL; r = r->next)
5953 if (r->type != OP_OUT && r->regno == dst_regno)
5954 break;
5955 /* Don't do inheritance if the pseudo is also
5956 used in the insn. */
5957 if (r == NULL)
5958 /* We can not do inheritance right now
5959 because the current insn reg info (chain
5960 regs) can change after that. */
5961 add_to_inherit (dst_regno, next_usage_insns);
5963 /* We can not process one reg twice here because of
5964 usage_insns invalidation. */
5965 if ((dst_regno < FIRST_PSEUDO_REGISTER
5966 || reg_renumber[dst_regno] >= 0)
5967 && ! reg->subreg_p && reg->type != OP_IN)
5969 HARD_REG_SET s;
5971 if (split_if_necessary (dst_regno, reg->biggest_mode,
5972 potential_reload_hard_regs,
5973 false, curr_insn, max_uid))
5974 change_p = true;
5975 CLEAR_HARD_REG_SET (s);
5976 if (dst_regno < FIRST_PSEUDO_REGISTER)
5977 add_to_hard_reg_set (&s, reg->biggest_mode, dst_regno);
5978 else
5979 add_to_hard_reg_set (&s, PSEUDO_REGNO_MODE (dst_regno),
5980 reg_renumber[dst_regno]);
5981 AND_COMPL_HARD_REG_SET (live_hard_regs, s);
5983 /* We should invalidate potential inheritance or
5984 splitting for the current insn usages to the next
5985 usage insns (see code below) as the output pseudo
5986 prevents this. */
5987 if ((dst_regno >= FIRST_PSEUDO_REGISTER
5988 && reg_renumber[dst_regno] < 0)
5989 || (reg->type == OP_OUT && ! reg->subreg_p
5990 && (dst_regno < FIRST_PSEUDO_REGISTER
5991 || reg_renumber[dst_regno] >= 0)))
5993 /* Invalidate and mark definitions. */
5994 if (dst_regno >= FIRST_PSEUDO_REGISTER)
5995 usage_insns[dst_regno].check = -(int) INSN_UID (curr_insn);
5996 else
5998 nregs = hard_regno_nregs[dst_regno][reg->biggest_mode];
5999 for (i = 0; i < nregs; i++)
6000 usage_insns[dst_regno + i].check
6001 = -(int) INSN_UID (curr_insn);
6005 /* Process clobbered call regs. */
6006 if (curr_id->arg_hard_regs != NULL)
6007 for (i = 0; (dst_regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6008 if (dst_regno >= FIRST_PSEUDO_REGISTER)
6009 usage_insns[dst_regno - FIRST_PSEUDO_REGISTER].check
6010 = -(int) INSN_UID (curr_insn);
6011 if (! JUMP_P (curr_insn))
6012 for (i = 0; i < to_inherit_num; i++)
6013 if (inherit_reload_reg (true, to_inherit[i].regno,
6014 ALL_REGS, curr_insn,
6015 to_inherit[i].insns))
6016 change_p = true;
6017 if (CALL_P (curr_insn))
6019 rtx cheap, pat, dest;
6020 rtx_insn *restore;
6021 int regno, hard_regno;
6023 calls_num++;
6024 if ((cheap = find_reg_note (curr_insn,
6025 REG_RETURNED, NULL_RTX)) != NULL_RTX
6026 && ((cheap = XEXP (cheap, 0)), true)
6027 && (regno = REGNO (cheap)) >= FIRST_PSEUDO_REGISTER
6028 && (hard_regno = reg_renumber[regno]) >= 0
6029 /* If there are pending saves/restores, the
6030 optimization is not worth. */
6031 && usage_insns[regno].calls_num == calls_num - 1
6032 && TEST_HARD_REG_BIT (call_used_reg_set, hard_regno))
6034 /* Restore the pseudo from the call result as
6035 REG_RETURNED note says that the pseudo value is
6036 in the call result and the pseudo is an argument
6037 of the call. */
6038 pat = PATTERN (curr_insn);
6039 if (GET_CODE (pat) == PARALLEL)
6040 pat = XVECEXP (pat, 0, 0);
6041 dest = SET_DEST (pat);
6042 /* For multiple return values dest is PARALLEL.
6043 Currently we handle only single return value case. */
6044 if (REG_P (dest))
6046 start_sequence ();
6047 emit_move_insn (cheap, copy_rtx (dest));
6048 restore = get_insns ();
6049 end_sequence ();
6050 lra_process_new_insns (curr_insn, NULL, restore,
6051 "Inserting call parameter restore");
6052 /* We don't need to save/restore of the pseudo from
6053 this call. */
6054 usage_insns[regno].calls_num = calls_num;
6055 bitmap_set_bit (&check_only_regs, regno);
6059 to_inherit_num = 0;
6060 /* Process insn usages. */
6061 for (iter = 0; iter < 2; iter++)
6062 for (reg = iter == 0 ? curr_id->regs : curr_static_id->hard_regs;
6063 reg != NULL;
6064 reg = reg->next)
6065 if ((reg->type != OP_OUT
6066 || (reg->type == OP_OUT && reg->subreg_p))
6067 && (src_regno = reg->regno) < lra_constraint_new_regno_start)
6069 if (src_regno >= FIRST_PSEUDO_REGISTER
6070 && reg_renumber[src_regno] < 0 && reg->type == OP_IN)
6072 if (usage_insns[src_regno].check == curr_usage_insns_check
6073 && (next_usage_insns
6074 = usage_insns[src_regno].insns) != NULL_RTX
6075 && NONDEBUG_INSN_P (curr_insn))
6076 add_to_inherit (src_regno, next_usage_insns);
6077 else if (usage_insns[src_regno].check
6078 != -(int) INSN_UID (curr_insn))
6079 /* Add usages but only if the reg is not set up
6080 in the same insn. */
6081 add_next_usage_insn (src_regno, curr_insn, reloads_num);
6083 else if (src_regno < FIRST_PSEUDO_REGISTER
6084 || reg_renumber[src_regno] >= 0)
6086 bool before_p;
6087 rtx_insn *use_insn = curr_insn;
6089 before_p = (JUMP_P (curr_insn)
6090 || (CALL_P (curr_insn) && reg->type == OP_IN));
6091 if (NONDEBUG_INSN_P (curr_insn)
6092 && (! JUMP_P (curr_insn) || reg->type == OP_IN)
6093 && split_if_necessary (src_regno, reg->biggest_mode,
6094 potential_reload_hard_regs,
6095 before_p, curr_insn, max_uid))
6097 if (reg->subreg_p)
6098 lra_risky_transformations_p = true;
6099 change_p = true;
6100 /* Invalidate. */
6101 usage_insns[src_regno].check = 0;
6102 if (before_p)
6103 use_insn = PREV_INSN (curr_insn);
6105 if (NONDEBUG_INSN_P (curr_insn))
6107 if (src_regno < FIRST_PSEUDO_REGISTER)
6108 add_to_hard_reg_set (&live_hard_regs,
6109 reg->biggest_mode, src_regno);
6110 else
6111 add_to_hard_reg_set (&live_hard_regs,
6112 PSEUDO_REGNO_MODE (src_regno),
6113 reg_renumber[src_regno]);
6115 add_next_usage_insn (src_regno, use_insn, reloads_num);
6118 /* Process used call regs. */
6119 if (curr_id->arg_hard_regs != NULL)
6120 for (i = 0; (src_regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6121 if (src_regno < FIRST_PSEUDO_REGISTER)
6123 SET_HARD_REG_BIT (live_hard_regs, src_regno);
6124 add_next_usage_insn (src_regno, curr_insn, reloads_num);
6126 for (i = 0; i < to_inherit_num; i++)
6128 src_regno = to_inherit[i].regno;
6129 if (inherit_reload_reg (false, src_regno, ALL_REGS,
6130 curr_insn, to_inherit[i].insns))
6131 change_p = true;
6132 else
6133 setup_next_usage_insn (src_regno, curr_insn, reloads_num, false);
6136 if (update_reloads_num_p
6137 && NONDEBUG_INSN_P (curr_insn) && curr_set != NULL_RTX)
6139 int regno = -1;
6140 if ((REG_P (SET_DEST (curr_set))
6141 && (regno = REGNO (SET_DEST (curr_set))) >= lra_constraint_new_regno_start
6142 && reg_renumber[regno] < 0
6143 && (cl = lra_get_allocno_class (regno)) != NO_REGS)
6144 || (REG_P (SET_SRC (curr_set))
6145 && (regno = REGNO (SET_SRC (curr_set))) >= lra_constraint_new_regno_start
6146 && reg_renumber[regno] < 0
6147 && (cl = lra_get_allocno_class (regno)) != NO_REGS))
6149 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6150 reloads_num++;
6151 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6152 IOR_HARD_REG_SET (potential_reload_hard_regs,
6153 reg_class_contents[cl]);
6156 if (NONDEBUG_INSN_P (curr_insn))
6158 int regno;
6160 /* Invalidate invariants with changed regs. */
6161 curr_id = lra_get_insn_recog_data (curr_insn);
6162 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
6163 if (reg->type != OP_IN)
6165 bitmap_set_bit (&invalid_invariant_regs, reg->regno);
6166 bitmap_set_bit (&invalid_invariant_regs,
6167 ORIGINAL_REGNO (regno_reg_rtx[reg->regno]));
6169 curr_static_id = curr_id->insn_static_data;
6170 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
6171 if (reg->type != OP_IN)
6172 bitmap_set_bit (&invalid_invariant_regs, reg->regno);
6173 if (curr_id->arg_hard_regs != NULL)
6174 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6175 if (regno >= FIRST_PSEUDO_REGISTER)
6176 bitmap_set_bit (&invalid_invariant_regs,
6177 regno - FIRST_PSEUDO_REGISTER);
6179 /* We reached the start of the current basic block. */
6180 if (prev_insn == NULL_RTX || prev_insn == PREV_INSN (head)
6181 || BLOCK_FOR_INSN (prev_insn) != curr_bb)
6183 /* We reached the beginning of the current block -- do
6184 rest of spliting in the current BB. */
6185 to_process = df_get_live_in (curr_bb);
6186 if (BLOCK_FOR_INSN (head) != curr_bb)
6188 /* We are somewhere in the middle of EBB. */
6189 get_live_on_other_edges (EDGE_PRED (curr_bb, 0)->src,
6190 curr_bb, &temp_bitmap);
6191 to_process = &temp_bitmap;
6193 head_p = true;
6194 EXECUTE_IF_SET_IN_BITMAP (to_process, 0, j, bi)
6196 if ((int) j >= lra_constraint_new_regno_start)
6197 break;
6198 if (((int) j < FIRST_PSEUDO_REGISTER || reg_renumber[j] >= 0)
6199 && usage_insns[j].check == curr_usage_insns_check
6200 && (next_usage_insns = usage_insns[j].insns) != NULL_RTX)
6202 if (need_for_split_p (potential_reload_hard_regs, j))
6204 if (lra_dump_file != NULL && head_p)
6206 fprintf (lra_dump_file,
6207 " ----------------------------------\n");
6208 head_p = false;
6210 if (split_reg (false, j, bb_note (curr_bb),
6211 next_usage_insns))
6212 change_p = true;
6214 usage_insns[j].check = 0;
6219 return change_p;
6222 /* This value affects EBB forming. If probability of edge from EBB to
6223 a BB is not greater than the following value, we don't add the BB
6224 to EBB. */
6225 #define EBB_PROBABILITY_CUTOFF \
6226 ((REG_BR_PROB_BASE * LRA_INHERITANCE_EBB_PROBABILITY_CUTOFF) / 100)
6228 /* Current number of inheritance/split iteration. */
6229 int lra_inheritance_iter;
6231 /* Entry function for inheritance/split pass. */
6232 void
6233 lra_inheritance (void)
6235 int i;
6236 basic_block bb, start_bb;
6237 edge e;
6239 lra_inheritance_iter++;
6240 if (lra_inheritance_iter > LRA_MAX_INHERITANCE_PASSES)
6241 return;
6242 timevar_push (TV_LRA_INHERITANCE);
6243 if (lra_dump_file != NULL)
6244 fprintf (lra_dump_file, "\n********** Inheritance #%d: **********\n\n",
6245 lra_inheritance_iter);
6246 curr_usage_insns_check = 0;
6247 usage_insns = XNEWVEC (struct usage_insns, lra_constraint_new_regno_start);
6248 for (i = 0; i < lra_constraint_new_regno_start; i++)
6249 usage_insns[i].check = 0;
6250 bitmap_initialize (&check_only_regs, &reg_obstack);
6251 bitmap_initialize (&invalid_invariant_regs, &reg_obstack);
6252 bitmap_initialize (&live_regs, &reg_obstack);
6253 bitmap_initialize (&temp_bitmap, &reg_obstack);
6254 bitmap_initialize (&ebb_global_regs, &reg_obstack);
6255 FOR_EACH_BB_FN (bb, cfun)
6257 start_bb = bb;
6258 if (lra_dump_file != NULL)
6259 fprintf (lra_dump_file, "EBB");
6260 /* Form a EBB starting with BB. */
6261 bitmap_clear (&ebb_global_regs);
6262 bitmap_ior_into (&ebb_global_regs, df_get_live_in (bb));
6263 for (;;)
6265 if (lra_dump_file != NULL)
6266 fprintf (lra_dump_file, " %d", bb->index);
6267 if (bb->next_bb == EXIT_BLOCK_PTR_FOR_FN (cfun)
6268 || LABEL_P (BB_HEAD (bb->next_bb)))
6269 break;
6270 e = find_fallthru_edge (bb->succs);
6271 if (! e)
6272 break;
6273 if (e->probability < EBB_PROBABILITY_CUTOFF)
6274 break;
6275 bb = bb->next_bb;
6277 bitmap_ior_into (&ebb_global_regs, df_get_live_out (bb));
6278 if (lra_dump_file != NULL)
6279 fprintf (lra_dump_file, "\n");
6280 if (inherit_in_ebb (BB_HEAD (start_bb), BB_END (bb)))
6281 /* Remember that the EBB head and tail can change in
6282 inherit_in_ebb. */
6283 update_ebb_live_info (BB_HEAD (start_bb), BB_END (bb));
6285 bitmap_clear (&ebb_global_regs);
6286 bitmap_clear (&temp_bitmap);
6287 bitmap_clear (&live_regs);
6288 bitmap_clear (&invalid_invariant_regs);
6289 bitmap_clear (&check_only_regs);
6290 free (usage_insns);
6292 timevar_pop (TV_LRA_INHERITANCE);
6297 /* This page contains code to undo failed inheritance/split
6298 transformations. */
6300 /* Current number of iteration undoing inheritance/split. */
6301 int lra_undo_inheritance_iter;
6303 /* Fix BB live info LIVE after removing pseudos created on pass doing
6304 inheritance/split which are REMOVED_PSEUDOS. */
6305 static void
6306 fix_bb_live_info (bitmap live, bitmap removed_pseudos)
6308 unsigned int regno;
6309 bitmap_iterator bi;
6311 EXECUTE_IF_SET_IN_BITMAP (removed_pseudos, 0, regno, bi)
6312 if (bitmap_clear_bit (live, regno)
6313 && REG_P (lra_reg_info[regno].restore_rtx))
6314 bitmap_set_bit (live, REGNO (lra_reg_info[regno].restore_rtx));
6317 /* Return regno of the (subreg of) REG. Otherwise, return a negative
6318 number. */
6319 static int
6320 get_regno (rtx reg)
6322 if (GET_CODE (reg) == SUBREG)
6323 reg = SUBREG_REG (reg);
6324 if (REG_P (reg))
6325 return REGNO (reg);
6326 return -1;
6329 /* Delete a move INSN with destination reg DREGNO and a previous
6330 clobber insn with the same regno. The inheritance/split code can
6331 generate moves with preceding clobber and when we delete such moves
6332 we should delete the clobber insn too to keep the correct life
6333 info. */
6334 static void
6335 delete_move_and_clobber (rtx_insn *insn, int dregno)
6337 rtx_insn *prev_insn = PREV_INSN (insn);
6339 lra_set_insn_deleted (insn);
6340 lra_assert (dregno >= 0);
6341 if (prev_insn != NULL && NONDEBUG_INSN_P (prev_insn)
6342 && GET_CODE (PATTERN (prev_insn)) == CLOBBER
6343 && dregno == get_regno (XEXP (PATTERN (prev_insn), 0)))
6344 lra_set_insn_deleted (prev_insn);
6347 /* Remove inheritance/split pseudos which are in REMOVE_PSEUDOS and
6348 return true if we did any change. The undo transformations for
6349 inheritance looks like
6350 i <- i2
6351 p <- i => p <- i2
6352 or removing
6353 p <- i, i <- p, and i <- i3
6354 where p is original pseudo from which inheritance pseudo i was
6355 created, i and i3 are removed inheritance pseudos, i2 is another
6356 not removed inheritance pseudo. All split pseudos or other
6357 occurrences of removed inheritance pseudos are changed on the
6358 corresponding original pseudos.
6360 The function also schedules insns changed and created during
6361 inheritance/split pass for processing by the subsequent constraint
6362 pass. */
6363 static bool
6364 remove_inheritance_pseudos (bitmap remove_pseudos)
6366 basic_block bb;
6367 int regno, sregno, prev_sregno, dregno;
6368 rtx restore_rtx;
6369 rtx set, prev_set;
6370 rtx_insn *prev_insn;
6371 bool change_p, done_p;
6373 change_p = ! bitmap_empty_p (remove_pseudos);
6374 /* We can not finish the function right away if CHANGE_P is true
6375 because we need to marks insns affected by previous
6376 inheritance/split pass for processing by the subsequent
6377 constraint pass. */
6378 FOR_EACH_BB_FN (bb, cfun)
6380 fix_bb_live_info (df_get_live_in (bb), remove_pseudos);
6381 fix_bb_live_info (df_get_live_out (bb), remove_pseudos);
6382 FOR_BB_INSNS_REVERSE (bb, curr_insn)
6384 if (! INSN_P (curr_insn))
6385 continue;
6386 done_p = false;
6387 sregno = dregno = -1;
6388 if (change_p && NONDEBUG_INSN_P (curr_insn)
6389 && (set = single_set (curr_insn)) != NULL_RTX)
6391 dregno = get_regno (SET_DEST (set));
6392 sregno = get_regno (SET_SRC (set));
6395 if (sregno >= 0 && dregno >= 0)
6397 if (bitmap_bit_p (remove_pseudos, dregno)
6398 && ! REG_P (lra_reg_info[dregno].restore_rtx))
6400 /* invariant inheritance pseudo <- original pseudo */
6401 if (lra_dump_file != NULL)
6403 fprintf (lra_dump_file, " Removing invariant inheritance:\n");
6404 dump_insn_slim (lra_dump_file, curr_insn);
6405 fprintf (lra_dump_file, "\n");
6407 delete_move_and_clobber (curr_insn, dregno);
6408 done_p = true;
6410 else if (bitmap_bit_p (remove_pseudos, sregno)
6411 && ! REG_P (lra_reg_info[sregno].restore_rtx))
6413 /* reload pseudo <- invariant inheritance pseudo */
6414 start_sequence ();
6415 /* We can not just change the source. It might be
6416 an insn different from the move. */
6417 emit_insn (lra_reg_info[sregno].restore_rtx);
6418 rtx_insn *new_insns = get_insns ();
6419 end_sequence ();
6420 lra_assert (single_set (new_insns) != NULL
6421 && SET_DEST (set) == SET_DEST (single_set (new_insns)));
6422 lra_process_new_insns (curr_insn, NULL, new_insns,
6423 "Changing reload<-invariant inheritance");
6424 delete_move_and_clobber (curr_insn, dregno);
6425 done_p = true;
6427 else if ((bitmap_bit_p (remove_pseudos, sregno)
6428 && (get_regno (lra_reg_info[sregno].restore_rtx) == dregno
6429 || (bitmap_bit_p (remove_pseudos, dregno)
6430 && get_regno (lra_reg_info[sregno].restore_rtx) >= 0
6431 && (get_regno (lra_reg_info[sregno].restore_rtx)
6432 == get_regno (lra_reg_info[dregno].restore_rtx)))))
6433 || (bitmap_bit_p (remove_pseudos, dregno)
6434 && get_regno (lra_reg_info[dregno].restore_rtx) == sregno))
6435 /* One of the following cases:
6436 original <- removed inheritance pseudo
6437 removed inherit pseudo <- another removed inherit pseudo
6438 removed inherit pseudo <- original pseudo
6440 removed_split_pseudo <- original_reg
6441 original_reg <- removed_split_pseudo */
6443 if (lra_dump_file != NULL)
6445 fprintf (lra_dump_file, " Removing %s:\n",
6446 bitmap_bit_p (&lra_split_regs, sregno)
6447 || bitmap_bit_p (&lra_split_regs, dregno)
6448 ? "split" : "inheritance");
6449 dump_insn_slim (lra_dump_file, curr_insn);
6451 delete_move_and_clobber (curr_insn, dregno);
6452 done_p = true;
6454 else if (bitmap_bit_p (remove_pseudos, sregno)
6455 && bitmap_bit_p (&lra_inheritance_pseudos, sregno))
6457 /* Search the following pattern:
6458 inherit_or_split_pseudo1 <- inherit_or_split_pseudo2
6459 original_pseudo <- inherit_or_split_pseudo1
6460 where the 2nd insn is the current insn and
6461 inherit_or_split_pseudo2 is not removed. If it is found,
6462 change the current insn onto:
6463 original_pseudo <- inherit_or_split_pseudo2. */
6464 for (prev_insn = PREV_INSN (curr_insn);
6465 prev_insn != NULL_RTX && ! NONDEBUG_INSN_P (prev_insn);
6466 prev_insn = PREV_INSN (prev_insn))
6468 if (prev_insn != NULL_RTX && BLOCK_FOR_INSN (prev_insn) == bb
6469 && (prev_set = single_set (prev_insn)) != NULL_RTX
6470 /* There should be no subregs in insn we are
6471 searching because only the original reg might
6472 be in subreg when we changed the mode of
6473 load/store for splitting. */
6474 && REG_P (SET_DEST (prev_set))
6475 && REG_P (SET_SRC (prev_set))
6476 && (int) REGNO (SET_DEST (prev_set)) == sregno
6477 && ((prev_sregno = REGNO (SET_SRC (prev_set)))
6478 >= FIRST_PSEUDO_REGISTER)
6479 && (lra_reg_info[prev_sregno].restore_rtx == NULL_RTX
6481 /* As we consider chain of inheritance or
6482 splitting described in above comment we should
6483 check that sregno and prev_sregno were
6484 inheritance/split pseudos created from the
6485 same original regno. */
6486 (get_regno (lra_reg_info[sregno].restore_rtx) >= 0
6487 && (get_regno (lra_reg_info[sregno].restore_rtx)
6488 == get_regno (lra_reg_info[prev_sregno].restore_rtx))))
6489 && ! bitmap_bit_p (remove_pseudos, prev_sregno))
6491 lra_assert (GET_MODE (SET_SRC (prev_set))
6492 == GET_MODE (regno_reg_rtx[sregno]));
6493 if (GET_CODE (SET_SRC (set)) == SUBREG)
6494 SUBREG_REG (SET_SRC (set)) = SET_SRC (prev_set);
6495 else
6496 SET_SRC (set) = SET_SRC (prev_set);
6497 /* As we are finishing with processing the insn
6498 here, check the destination too as it might
6499 inheritance pseudo for another pseudo. */
6500 if (bitmap_bit_p (remove_pseudos, dregno)
6501 && bitmap_bit_p (&lra_inheritance_pseudos, dregno)
6502 && (restore_rtx
6503 = lra_reg_info[dregno].restore_rtx) != NULL_RTX)
6505 if (GET_CODE (SET_DEST (set)) == SUBREG)
6506 SUBREG_REG (SET_DEST (set)) = restore_rtx;
6507 else
6508 SET_DEST (set) = restore_rtx;
6510 lra_push_insn_and_update_insn_regno_info (curr_insn);
6511 lra_set_used_insn_alternative_by_uid
6512 (INSN_UID (curr_insn), -1);
6513 done_p = true;
6514 if (lra_dump_file != NULL)
6516 fprintf (lra_dump_file, " Change reload insn:\n");
6517 dump_insn_slim (lra_dump_file, curr_insn);
6522 if (! done_p)
6524 struct lra_insn_reg *reg;
6525 bool restored_regs_p = false;
6526 bool kept_regs_p = false;
6528 curr_id = lra_get_insn_recog_data (curr_insn);
6529 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
6531 regno = reg->regno;
6532 restore_rtx = lra_reg_info[regno].restore_rtx;
6533 if (restore_rtx != NULL_RTX)
6535 if (change_p && bitmap_bit_p (remove_pseudos, regno))
6537 lra_substitute_pseudo_within_insn
6538 (curr_insn, regno, restore_rtx, false);
6539 restored_regs_p = true;
6541 else
6542 kept_regs_p = true;
6545 if (NONDEBUG_INSN_P (curr_insn) && kept_regs_p)
6547 /* The instruction has changed since the previous
6548 constraints pass. */
6549 lra_push_insn_and_update_insn_regno_info (curr_insn);
6550 lra_set_used_insn_alternative_by_uid
6551 (INSN_UID (curr_insn), -1);
6553 else if (restored_regs_p)
6554 /* The instruction has been restored to the form that
6555 it had during the previous constraints pass. */
6556 lra_update_insn_regno_info (curr_insn);
6557 if (restored_regs_p && lra_dump_file != NULL)
6559 fprintf (lra_dump_file, " Insn after restoring regs:\n");
6560 dump_insn_slim (lra_dump_file, curr_insn);
6565 return change_p;
6568 /* If optional reload pseudos failed to get a hard register or was not
6569 inherited, it is better to remove optional reloads. We do this
6570 transformation after undoing inheritance to figure out necessity to
6571 remove optional reloads easier. Return true if we do any
6572 change. */
6573 static bool
6574 undo_optional_reloads (void)
6576 bool change_p, keep_p;
6577 unsigned int regno, uid;
6578 bitmap_iterator bi, bi2;
6579 rtx_insn *insn;
6580 rtx set, src, dest;
6581 bitmap_head removed_optional_reload_pseudos, insn_bitmap;
6583 bitmap_initialize (&removed_optional_reload_pseudos, &reg_obstack);
6584 bitmap_copy (&removed_optional_reload_pseudos, &lra_optional_reload_pseudos);
6585 EXECUTE_IF_SET_IN_BITMAP (&lra_optional_reload_pseudos, 0, regno, bi)
6587 keep_p = false;
6588 /* Keep optional reloads from previous subpasses. */
6589 if (lra_reg_info[regno].restore_rtx == NULL_RTX
6590 /* If the original pseudo changed its allocation, just
6591 removing the optional pseudo is dangerous as the original
6592 pseudo will have longer live range. */
6593 || reg_renumber[REGNO (lra_reg_info[regno].restore_rtx)] >= 0)
6594 keep_p = true;
6595 else if (reg_renumber[regno] >= 0)
6596 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi2)
6598 insn = lra_insn_recog_data[uid]->insn;
6599 if ((set = single_set (insn)) == NULL_RTX)
6600 continue;
6601 src = SET_SRC (set);
6602 dest = SET_DEST (set);
6603 if (! REG_P (src) || ! REG_P (dest))
6604 continue;
6605 if (REGNO (dest) == regno
6606 /* Ignore insn for optional reloads itself. */
6607 && REGNO (lra_reg_info[regno].restore_rtx) != REGNO (src)
6608 /* Check only inheritance on last inheritance pass. */
6609 && (int) REGNO (src) >= new_regno_start
6610 /* Check that the optional reload was inherited. */
6611 && bitmap_bit_p (&lra_inheritance_pseudos, REGNO (src)))
6613 keep_p = true;
6614 break;
6617 if (keep_p)
6619 bitmap_clear_bit (&removed_optional_reload_pseudos, regno);
6620 if (lra_dump_file != NULL)
6621 fprintf (lra_dump_file, "Keep optional reload reg %d\n", regno);
6624 change_p = ! bitmap_empty_p (&removed_optional_reload_pseudos);
6625 bitmap_initialize (&insn_bitmap, &reg_obstack);
6626 EXECUTE_IF_SET_IN_BITMAP (&removed_optional_reload_pseudos, 0, regno, bi)
6628 if (lra_dump_file != NULL)
6629 fprintf (lra_dump_file, "Remove optional reload reg %d\n", regno);
6630 bitmap_copy (&insn_bitmap, &lra_reg_info[regno].insn_bitmap);
6631 EXECUTE_IF_SET_IN_BITMAP (&insn_bitmap, 0, uid, bi2)
6633 insn = lra_insn_recog_data[uid]->insn;
6634 if ((set = single_set (insn)) != NULL_RTX)
6636 src = SET_SRC (set);
6637 dest = SET_DEST (set);
6638 if (REG_P (src) && REG_P (dest)
6639 && ((REGNO (src) == regno
6640 && (REGNO (lra_reg_info[regno].restore_rtx)
6641 == REGNO (dest)))
6642 || (REGNO (dest) == regno
6643 && (REGNO (lra_reg_info[regno].restore_rtx)
6644 == REGNO (src)))))
6646 if (lra_dump_file != NULL)
6648 fprintf (lra_dump_file, " Deleting move %u\n",
6649 INSN_UID (insn));
6650 dump_insn_slim (lra_dump_file, insn);
6652 delete_move_and_clobber (insn, REGNO (dest));
6653 continue;
6655 /* We should not worry about generation memory-memory
6656 moves here as if the corresponding inheritance did
6657 not work (inheritance pseudo did not get a hard reg),
6658 we remove the inheritance pseudo and the optional
6659 reload. */
6661 lra_substitute_pseudo_within_insn
6662 (insn, regno, lra_reg_info[regno].restore_rtx, false);
6663 lra_update_insn_regno_info (insn);
6664 if (lra_dump_file != NULL)
6666 fprintf (lra_dump_file,
6667 " Restoring original insn:\n");
6668 dump_insn_slim (lra_dump_file, insn);
6672 /* Clear restore_regnos. */
6673 EXECUTE_IF_SET_IN_BITMAP (&lra_optional_reload_pseudos, 0, regno, bi)
6674 lra_reg_info[regno].restore_rtx = NULL_RTX;
6675 bitmap_clear (&insn_bitmap);
6676 bitmap_clear (&removed_optional_reload_pseudos);
6677 return change_p;
6680 /* Entry function for undoing inheritance/split transformation. Return true
6681 if we did any RTL change in this pass. */
6682 bool
6683 lra_undo_inheritance (void)
6685 unsigned int regno;
6686 int hard_regno;
6687 int n_all_inherit, n_inherit, n_all_split, n_split;
6688 rtx restore_rtx;
6689 bitmap_head remove_pseudos;
6690 bitmap_iterator bi;
6691 bool change_p;
6693 lra_undo_inheritance_iter++;
6694 if (lra_undo_inheritance_iter > LRA_MAX_INHERITANCE_PASSES)
6695 return false;
6696 if (lra_dump_file != NULL)
6697 fprintf (lra_dump_file,
6698 "\n********** Undoing inheritance #%d: **********\n\n",
6699 lra_undo_inheritance_iter);
6700 bitmap_initialize (&remove_pseudos, &reg_obstack);
6701 n_inherit = n_all_inherit = 0;
6702 EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, regno, bi)
6703 if (lra_reg_info[regno].restore_rtx != NULL_RTX)
6705 n_all_inherit++;
6706 if (reg_renumber[regno] < 0
6707 /* If the original pseudo changed its allocation, just
6708 removing inheritance is dangerous as for changing
6709 allocation we used shorter live-ranges. */
6710 && (! REG_P (lra_reg_info[regno].restore_rtx)
6711 || reg_renumber[REGNO (lra_reg_info[regno].restore_rtx)] < 0))
6712 bitmap_set_bit (&remove_pseudos, regno);
6713 else
6714 n_inherit++;
6716 if (lra_dump_file != NULL && n_all_inherit != 0)
6717 fprintf (lra_dump_file, "Inherit %d out of %d (%.2f%%)\n",
6718 n_inherit, n_all_inherit,
6719 (double) n_inherit / n_all_inherit * 100);
6720 n_split = n_all_split = 0;
6721 EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, regno, bi)
6722 if ((restore_rtx = lra_reg_info[regno].restore_rtx) != NULL_RTX)
6724 int restore_regno = REGNO (restore_rtx);
6726 n_all_split++;
6727 hard_regno = (restore_regno >= FIRST_PSEUDO_REGISTER
6728 ? reg_renumber[restore_regno] : restore_regno);
6729 if (hard_regno < 0 || reg_renumber[regno] == hard_regno)
6730 bitmap_set_bit (&remove_pseudos, regno);
6731 else
6733 n_split++;
6734 if (lra_dump_file != NULL)
6735 fprintf (lra_dump_file, " Keep split r%d (orig=r%d)\n",
6736 regno, restore_regno);
6739 if (lra_dump_file != NULL && n_all_split != 0)
6740 fprintf (lra_dump_file, "Split %d out of %d (%.2f%%)\n",
6741 n_split, n_all_split,
6742 (double) n_split / n_all_split * 100);
6743 change_p = remove_inheritance_pseudos (&remove_pseudos);
6744 bitmap_clear (&remove_pseudos);
6745 /* Clear restore_regnos. */
6746 EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, regno, bi)
6747 lra_reg_info[regno].restore_rtx = NULL_RTX;
6748 EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, regno, bi)
6749 lra_reg_info[regno].restore_rtx = NULL_RTX;
6750 change_p = undo_optional_reloads () || change_p;
6751 return change_p;