Add run tests for recent sibcall patches
[official-gcc.git] / gcc / lra-constraints.c
blob133b55ce0ba48f7924af722039ff8380c8934b2e
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 if (dclass == ALL_REGS)
1154 /* ALL_REGS is used for new pseudos created by transformations
1155 like reload of SUBREG_REG (see function
1156 simplify_operand_subreg). We don't know their class yet. We
1157 should figure out the class from processing the insn
1158 constraints not in this fast path function. Even if ALL_REGS
1159 were a right class for the pseudo, secondary_... hooks usually
1160 are not define for ALL_REGS. */
1161 return false;
1162 if (REG_P (sreg))
1163 sclass = get_reg_class (REGNO (sreg));
1164 if (sclass == ALL_REGS)
1165 /* See comments above. */
1166 return false;
1167 if (sclass == NO_REGS && dclass == NO_REGS)
1168 return false;
1169 #ifdef SECONDARY_MEMORY_NEEDED
1170 if (SECONDARY_MEMORY_NEEDED (sclass, dclass, GET_MODE (src))
1171 #ifdef SECONDARY_MEMORY_NEEDED_MODE
1172 && ((sclass != NO_REGS && dclass != NO_REGS)
1173 || GET_MODE (src) != SECONDARY_MEMORY_NEEDED_MODE (GET_MODE (src)))
1174 #endif
1177 *sec_mem_p = true;
1178 return false;
1180 #endif
1181 if (! REG_P (dreg) || ! REG_P (sreg))
1182 return false;
1183 sri.prev_sri = NULL;
1184 sri.icode = CODE_FOR_nothing;
1185 sri.extra_cost = 0;
1186 secondary_class = NO_REGS;
1187 /* Set up hard register for a reload pseudo for hook
1188 secondary_reload because some targets just ignore unassigned
1189 pseudos in the hook. */
1190 if (dclass != NO_REGS && lra_get_regno_hard_regno (REGNO (dreg)) < 0)
1192 dregno = REGNO (dreg);
1193 reg_renumber[dregno] = ira_class_hard_regs[dclass][0];
1195 else
1196 dregno = -1;
1197 if (sclass != NO_REGS && lra_get_regno_hard_regno (REGNO (sreg)) < 0)
1199 sregno = REGNO (sreg);
1200 reg_renumber[sregno] = ira_class_hard_regs[sclass][0];
1202 else
1203 sregno = -1;
1204 if (sclass != NO_REGS)
1205 secondary_class
1206 = (enum reg_class) targetm.secondary_reload (false, dest,
1207 (reg_class_t) sclass,
1208 GET_MODE (src), &sri);
1209 if (sclass == NO_REGS
1210 || ((secondary_class != NO_REGS || sri.icode != CODE_FOR_nothing)
1211 && dclass != NO_REGS))
1213 enum reg_class old_sclass = secondary_class;
1214 secondary_reload_info old_sri = sri;
1216 sri.prev_sri = NULL;
1217 sri.icode = CODE_FOR_nothing;
1218 sri.extra_cost = 0;
1219 secondary_class
1220 = (enum reg_class) targetm.secondary_reload (true, src,
1221 (reg_class_t) dclass,
1222 GET_MODE (src), &sri);
1223 /* Check the target hook consistency. */
1224 lra_assert
1225 ((secondary_class == NO_REGS && sri.icode == CODE_FOR_nothing)
1226 || (old_sclass == NO_REGS && old_sri.icode == CODE_FOR_nothing)
1227 || (secondary_class == old_sclass && sri.icode == old_sri.icode));
1229 if (sregno >= 0)
1230 reg_renumber [sregno] = -1;
1231 if (dregno >= 0)
1232 reg_renumber [dregno] = -1;
1233 if (secondary_class == NO_REGS && sri.icode == CODE_FOR_nothing)
1234 return false;
1235 *change_p = true;
1236 new_reg = NULL_RTX;
1237 if (secondary_class != NO_REGS)
1238 new_reg = lra_create_new_reg_with_unique_value (GET_MODE (src), NULL_RTX,
1239 secondary_class,
1240 "secondary");
1241 start_sequence ();
1242 if (sri.icode == CODE_FOR_nothing)
1243 lra_emit_move (new_reg, src);
1244 else
1246 enum reg_class scratch_class;
1248 scratch_class = (reg_class_from_constraints
1249 (insn_data[sri.icode].operand[2].constraint));
1250 scratch_reg = (lra_create_new_reg_with_unique_value
1251 (insn_data[sri.icode].operand[2].mode, NULL_RTX,
1252 scratch_class, "scratch"));
1253 emit_insn (GEN_FCN (sri.icode) (new_reg != NULL_RTX ? new_reg : dest,
1254 src, scratch_reg));
1256 before = get_insns ();
1257 end_sequence ();
1258 lra_process_new_insns (curr_insn, before, NULL, "Inserting the move");
1259 if (new_reg != NULL_RTX)
1260 SET_SRC (curr_insn_set) = new_reg;
1261 else
1263 if (lra_dump_file != NULL)
1265 fprintf (lra_dump_file, "Deleting move %u\n", INSN_UID (curr_insn));
1266 dump_insn_slim (lra_dump_file, curr_insn);
1268 lra_set_insn_deleted (curr_insn);
1269 return true;
1271 return false;
1274 /* The following data describe the result of process_alt_operands.
1275 The data are used in curr_insn_transform to generate reloads. */
1277 /* The chosen reg classes which should be used for the corresponding
1278 operands. */
1279 static enum reg_class goal_alt[MAX_RECOG_OPERANDS];
1280 /* True if the operand should be the same as another operand and that
1281 other operand does not need a reload. */
1282 static bool goal_alt_match_win[MAX_RECOG_OPERANDS];
1283 /* True if the operand does not need a reload. */
1284 static bool goal_alt_win[MAX_RECOG_OPERANDS];
1285 /* True if the operand can be offsetable memory. */
1286 static bool goal_alt_offmemok[MAX_RECOG_OPERANDS];
1287 /* The number of an operand to which given operand can be matched to. */
1288 static int goal_alt_matches[MAX_RECOG_OPERANDS];
1289 /* The number of elements in the following array. */
1290 static int goal_alt_dont_inherit_ops_num;
1291 /* Numbers of operands whose reload pseudos should not be inherited. */
1292 static int goal_alt_dont_inherit_ops[MAX_RECOG_OPERANDS];
1293 /* True if the insn commutative operands should be swapped. */
1294 static bool goal_alt_swapped;
1295 /* The chosen insn alternative. */
1296 static int goal_alt_number;
1298 /* True if the corresponding operand is the result of an equivalence
1299 substitution. */
1300 static bool equiv_substition_p[MAX_RECOG_OPERANDS];
1302 /* The following five variables are used to choose the best insn
1303 alternative. They reflect final characteristics of the best
1304 alternative. */
1306 /* Number of necessary reloads and overall cost reflecting the
1307 previous value and other unpleasantness of the best alternative. */
1308 static int best_losers, best_overall;
1309 /* Overall number hard registers used for reloads. For example, on
1310 some targets we need 2 general registers to reload DFmode and only
1311 one floating point register. */
1312 static int best_reload_nregs;
1313 /* Overall number reflecting distances of previous reloading the same
1314 value. The distances are counted from the current BB start. It is
1315 used to improve inheritance chances. */
1316 static int best_reload_sum;
1318 /* True if the current insn should have no correspondingly input or
1319 output reloads. */
1320 static bool no_input_reloads_p, no_output_reloads_p;
1322 /* True if we swapped the commutative operands in the current
1323 insn. */
1324 static int curr_swapped;
1326 /* if CHECK_ONLY_P is false, arrange for address element *LOC to be a
1327 register of class CL. Add any input reloads to list BEFORE. AFTER
1328 is nonnull if *LOC is an automodified value; handle that case by
1329 adding the required output reloads to list AFTER. Return true if
1330 the RTL was changed.
1332 if CHECK_ONLY_P is true, check that the *LOC is a correct address
1333 register. Return false if the address register is correct. */
1334 static bool
1335 process_addr_reg (rtx *loc, bool check_only_p, rtx_insn **before, rtx_insn **after,
1336 enum reg_class cl)
1338 int regno;
1339 enum reg_class rclass, new_class;
1340 rtx reg;
1341 rtx new_reg;
1342 machine_mode mode;
1343 bool subreg_p, before_p = false;
1345 subreg_p = GET_CODE (*loc) == SUBREG;
1346 if (subreg_p)
1348 reg = SUBREG_REG (*loc);
1349 mode = GET_MODE (reg);
1351 /* For mode with size bigger than ptr_mode, there unlikely to be "mov"
1352 between two registers with different classes, but there normally will
1353 be "mov" which transfers element of vector register into the general
1354 register, and this normally will be a subreg which should be reloaded
1355 as a whole. This is particularly likely to be triggered when
1356 -fno-split-wide-types specified. */
1357 if (!REG_P (reg)
1358 || in_class_p (reg, cl, &new_class)
1359 || GET_MODE_SIZE (mode) <= GET_MODE_SIZE (ptr_mode))
1360 loc = &SUBREG_REG (*loc);
1363 reg = *loc;
1364 mode = GET_MODE (reg);
1365 if (! REG_P (reg))
1367 if (check_only_p)
1368 return true;
1369 /* Always reload memory in an address even if the target supports
1370 such addresses. */
1371 new_reg = lra_create_new_reg_with_unique_value (mode, reg, cl, "address");
1372 before_p = true;
1374 else
1376 regno = REGNO (reg);
1377 rclass = get_reg_class (regno);
1378 if (! check_only_p
1379 && (*loc = get_equiv_with_elimination (reg, curr_insn)) != reg)
1381 if (lra_dump_file != NULL)
1383 fprintf (lra_dump_file,
1384 "Changing pseudo %d in address of insn %u on equiv ",
1385 REGNO (reg), INSN_UID (curr_insn));
1386 dump_value_slim (lra_dump_file, *loc, 1);
1387 fprintf (lra_dump_file, "\n");
1389 *loc = copy_rtx (*loc);
1391 if (*loc != reg || ! in_class_p (reg, cl, &new_class))
1393 if (check_only_p)
1394 return true;
1395 reg = *loc;
1396 if (get_reload_reg (after == NULL ? OP_IN : OP_INOUT,
1397 mode, reg, cl, subreg_p, "address", &new_reg))
1398 before_p = true;
1400 else if (new_class != NO_REGS && rclass != new_class)
1402 if (check_only_p)
1403 return true;
1404 lra_change_class (regno, new_class, " Change to", true);
1405 return false;
1407 else
1408 return false;
1410 if (before_p)
1412 push_to_sequence (*before);
1413 lra_emit_move (new_reg, reg);
1414 *before = get_insns ();
1415 end_sequence ();
1417 *loc = new_reg;
1418 if (after != NULL)
1420 start_sequence ();
1421 lra_emit_move (before_p ? copy_rtx (reg) : reg, new_reg);
1422 emit_insn (*after);
1423 *after = get_insns ();
1424 end_sequence ();
1426 return true;
1429 /* Insert move insn in simplify_operand_subreg. BEFORE returns
1430 the insn to be inserted before curr insn. AFTER returns the
1431 the insn to be inserted after curr insn. ORIGREG and NEWREG
1432 are the original reg and new reg for reload. */
1433 static void
1434 insert_move_for_subreg (rtx_insn **before, rtx_insn **after, rtx origreg,
1435 rtx newreg)
1437 if (before)
1439 push_to_sequence (*before);
1440 lra_emit_move (newreg, origreg);
1441 *before = get_insns ();
1442 end_sequence ();
1444 if (after)
1446 start_sequence ();
1447 lra_emit_move (origreg, newreg);
1448 emit_insn (*after);
1449 *after = get_insns ();
1450 end_sequence ();
1454 static int valid_address_p (machine_mode mode, rtx addr, addr_space_t as);
1456 /* Make reloads for subreg in operand NOP with internal subreg mode
1457 REG_MODE, add new reloads for further processing. Return true if
1458 any change was done. */
1459 static bool
1460 simplify_operand_subreg (int nop, machine_mode reg_mode)
1462 int hard_regno;
1463 rtx_insn *before, *after;
1464 machine_mode mode, innermode;
1465 rtx reg, new_reg;
1466 rtx operand = *curr_id->operand_loc[nop];
1467 enum reg_class regclass;
1468 enum op_type type;
1470 before = after = NULL;
1472 if (GET_CODE (operand) != SUBREG)
1473 return false;
1475 mode = GET_MODE (operand);
1476 reg = SUBREG_REG (operand);
1477 innermode = GET_MODE (reg);
1478 type = curr_static_id->operand[nop].type;
1479 if (MEM_P (reg))
1481 rtx subst;
1483 alter_subreg (curr_id->operand_loc[nop], false);
1484 subst = *curr_id->operand_loc[nop];
1485 lra_assert (MEM_P (subst));
1486 if (! valid_address_p (innermode, XEXP (reg, 0),
1487 MEM_ADDR_SPACE (reg))
1488 || valid_address_p (GET_MODE (subst), XEXP (subst, 0),
1489 MEM_ADDR_SPACE (subst))
1490 || ((get_constraint_type (lookup_constraint
1491 (curr_static_id->operand[nop].constraint))
1492 != CT_SPECIAL_MEMORY)
1493 /* We still can reload address and if the address is
1494 valid, we can remove subreg without reloading its
1495 inner memory. */
1496 && valid_address_p (GET_MODE (subst),
1497 regno_reg_rtx
1498 [ira_class_hard_regs
1499 [base_reg_class (GET_MODE (subst),
1500 MEM_ADDR_SPACE (subst),
1501 ADDRESS, SCRATCH)][0]],
1502 MEM_ADDR_SPACE (subst))))
1504 /* If we change address for paradoxical subreg of memory, the
1505 address might violate the necessary alignment or the access might
1506 be slow. So take this into consideration. We should not worry
1507 about access beyond allocated memory for paradoxical memory
1508 subregs as we don't substitute such equiv memory (see processing
1509 equivalences in function lra_constraints) and because for spilled
1510 pseudos we allocate stack memory enough for the biggest
1511 corresponding paradoxical subreg. */
1512 if (!(MEM_ALIGN (reg) < GET_MODE_ALIGNMENT (mode)
1513 && SLOW_UNALIGNED_ACCESS (mode, MEM_ALIGN (reg)))
1514 || (MEM_ALIGN (reg) < GET_MODE_ALIGNMENT (innermode)
1515 && SLOW_UNALIGNED_ACCESS (innermode, MEM_ALIGN (reg))))
1516 return true;
1518 /* INNERMODE is fast, MODE slow. Reload the mem in INNERMODE. */
1519 enum reg_class rclass
1520 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1521 if (get_reload_reg (curr_static_id->operand[nop].type, innermode, reg,
1522 rclass, TRUE, "slow mem", &new_reg))
1524 bool insert_before, insert_after;
1525 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1527 insert_before = (type != OP_OUT
1528 || GET_MODE_SIZE (innermode) > GET_MODE_SIZE (mode));
1529 insert_after = type != OP_IN;
1530 insert_move_for_subreg (insert_before ? &before : NULL,
1531 insert_after ? &after : NULL,
1532 reg, new_reg);
1534 *curr_id->operand_loc[nop] = operand;
1535 SUBREG_REG (operand) = new_reg;
1537 /* Convert to MODE. */
1538 reg = operand;
1539 rclass = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1540 if (get_reload_reg (curr_static_id->operand[nop].type, mode, reg,
1541 rclass, TRUE, "slow mem", &new_reg))
1543 bool insert_before, insert_after;
1544 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1546 insert_before = type != OP_OUT;
1547 insert_after = type != OP_IN;
1548 insert_move_for_subreg (insert_before ? &before : NULL,
1549 insert_after ? &after : NULL,
1550 reg, new_reg);
1552 *curr_id->operand_loc[nop] = new_reg;
1553 lra_process_new_insns (curr_insn, before, after,
1554 "Inserting slow mem reload");
1555 return true;
1558 /* If the address was valid and became invalid, prefer to reload
1559 the memory. Typical case is when the index scale should
1560 correspond the memory. */
1561 *curr_id->operand_loc[nop] = operand;
1563 else if (REG_P (reg) && REGNO (reg) < FIRST_PSEUDO_REGISTER)
1565 alter_subreg (curr_id->operand_loc[nop], false);
1566 return true;
1568 else if (CONSTANT_P (reg))
1570 /* Try to simplify subreg of constant. It is usually result of
1571 equivalence substitution. */
1572 if (innermode == VOIDmode
1573 && (innermode = original_subreg_reg_mode[nop]) == VOIDmode)
1574 innermode = curr_static_id->operand[nop].mode;
1575 if ((new_reg = simplify_subreg (mode, reg, innermode,
1576 SUBREG_BYTE (operand))) != NULL_RTX)
1578 *curr_id->operand_loc[nop] = new_reg;
1579 return true;
1582 /* Put constant into memory when we have mixed modes. It generates
1583 a better code in most cases as it does not need a secondary
1584 reload memory. It also prevents LRA looping when LRA is using
1585 secondary reload memory again and again. */
1586 if (CONSTANT_P (reg) && CONST_POOL_OK_P (reg_mode, reg)
1587 && SCALAR_INT_MODE_P (reg_mode) != SCALAR_INT_MODE_P (mode))
1589 SUBREG_REG (operand) = force_const_mem (reg_mode, reg);
1590 alter_subreg (curr_id->operand_loc[nop], false);
1591 return true;
1593 /* Force a reload of the SUBREG_REG if this is a constant or PLUS or
1594 if there may be a problem accessing OPERAND in the outer
1595 mode. */
1596 if ((REG_P (reg)
1597 && REGNO (reg) >= FIRST_PSEUDO_REGISTER
1598 && (hard_regno = lra_get_regno_hard_regno (REGNO (reg))) >= 0
1599 /* Don't reload paradoxical subregs because we could be looping
1600 having repeatedly final regno out of hard regs range. */
1601 && (hard_regno_nregs[hard_regno][innermode]
1602 >= hard_regno_nregs[hard_regno][mode])
1603 && simplify_subreg_regno (hard_regno, innermode,
1604 SUBREG_BYTE (operand), mode) < 0
1605 /* Don't reload subreg for matching reload. It is actually
1606 valid subreg in LRA. */
1607 && ! LRA_SUBREG_P (operand))
1608 || CONSTANT_P (reg) || GET_CODE (reg) == PLUS || MEM_P (reg))
1610 enum reg_class rclass;
1612 if (REG_P (reg))
1613 /* There is a big probability that we will get the same class
1614 for the new pseudo and we will get the same insn which
1615 means infinite looping. So spill the new pseudo. */
1616 rclass = NO_REGS;
1617 else
1618 /* The class will be defined later in curr_insn_transform. */
1619 rclass
1620 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1622 if (get_reload_reg (curr_static_id->operand[nop].type, reg_mode, reg,
1623 rclass, TRUE, "subreg reg", &new_reg))
1625 bool insert_before, insert_after;
1626 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1628 insert_before = (type != OP_OUT
1629 || GET_MODE_SIZE (innermode) > GET_MODE_SIZE (mode));
1630 insert_after = (type != OP_IN);
1631 insert_move_for_subreg (insert_before ? &before : NULL,
1632 insert_after ? &after : NULL,
1633 reg, new_reg);
1635 SUBREG_REG (operand) = new_reg;
1636 lra_process_new_insns (curr_insn, before, after,
1637 "Inserting subreg reload");
1638 return true;
1640 /* Force a reload for a paradoxical subreg. For paradoxical subreg,
1641 IRA allocates hardreg to the inner pseudo reg according to its mode
1642 instead of the outermode, so the size of the hardreg may not be enough
1643 to contain the outermode operand, in that case we may need to insert
1644 reload for the reg. For the following two types of paradoxical subreg,
1645 we need to insert reload:
1646 1. If the op_type is OP_IN, and the hardreg could not be paired with
1647 other hardreg to contain the outermode operand
1648 (checked by in_hard_reg_set_p), we need to insert the reload.
1649 2. If the op_type is OP_OUT or OP_INOUT.
1651 Here is a paradoxical subreg example showing how the reload is generated:
1653 (insn 5 4 7 2 (set (reg:TI 106 [ __comp ])
1654 (subreg:TI (reg:DI 107 [ __comp ]) 0)) {*movti_internal_rex64}
1656 In IRA, reg107 is allocated to a DImode hardreg. We use x86-64 as example
1657 here, if reg107 is assigned to hardreg R15, because R15 is the last
1658 hardreg, compiler cannot find another hardreg to pair with R15 to
1659 contain TImode data. So we insert a TImode reload reg180 for it.
1660 After reload is inserted:
1662 (insn 283 0 0 (set (subreg:DI (reg:TI 180 [orig:107 __comp ] [107]) 0)
1663 (reg:DI 107 [ __comp ])) -1
1664 (insn 5 4 7 2 (set (reg:TI 106 [ __comp ])
1665 (subreg:TI (reg:TI 180 [orig:107 __comp ] [107]) 0)) {*movti_internal_rex64}
1667 Two reload hard registers will be allocated to reg180 to save TImode data
1668 in LRA_assign. */
1669 else if (REG_P (reg)
1670 && REGNO (reg) >= FIRST_PSEUDO_REGISTER
1671 && (hard_regno = lra_get_regno_hard_regno (REGNO (reg))) >= 0
1672 && (hard_regno_nregs[hard_regno][innermode]
1673 < hard_regno_nregs[hard_regno][mode])
1674 && (regclass = lra_get_allocno_class (REGNO (reg)))
1675 && (type != OP_IN
1676 || !in_hard_reg_set_p (reg_class_contents[regclass],
1677 mode, hard_regno)))
1679 /* The class will be defined later in curr_insn_transform. */
1680 enum reg_class rclass
1681 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1683 if (get_reload_reg (curr_static_id->operand[nop].type, mode, reg,
1684 rclass, TRUE, "paradoxical subreg", &new_reg))
1686 rtx subreg;
1687 bool insert_before, insert_after;
1689 PUT_MODE (new_reg, mode);
1690 subreg = gen_lowpart_SUBREG (innermode, new_reg);
1691 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1693 insert_before = (type != OP_OUT);
1694 insert_after = (type != OP_IN);
1695 insert_move_for_subreg (insert_before ? &before : NULL,
1696 insert_after ? &after : NULL,
1697 reg, subreg);
1699 SUBREG_REG (operand) = new_reg;
1700 lra_process_new_insns (curr_insn, before, after,
1701 "Inserting paradoxical subreg reload");
1702 return true;
1704 return false;
1707 /* Return TRUE if X refers for a hard register from SET. */
1708 static bool
1709 uses_hard_regs_p (rtx x, HARD_REG_SET set)
1711 int i, j, x_hard_regno;
1712 machine_mode mode;
1713 const char *fmt;
1714 enum rtx_code code;
1716 if (x == NULL_RTX)
1717 return false;
1718 code = GET_CODE (x);
1719 mode = GET_MODE (x);
1720 if (code == SUBREG)
1722 x = SUBREG_REG (x);
1723 code = GET_CODE (x);
1724 if (GET_MODE_SIZE (GET_MODE (x)) > GET_MODE_SIZE (mode))
1725 mode = GET_MODE (x);
1728 if (REG_P (x))
1730 x_hard_regno = get_hard_regno (x, true);
1731 return (x_hard_regno >= 0
1732 && overlaps_hard_reg_set_p (set, mode, x_hard_regno));
1734 if (MEM_P (x))
1736 struct address_info ad;
1738 decompose_mem_address (&ad, x);
1739 if (ad.base_term != NULL && uses_hard_regs_p (*ad.base_term, set))
1740 return true;
1741 if (ad.index_term != NULL && uses_hard_regs_p (*ad.index_term, set))
1742 return true;
1744 fmt = GET_RTX_FORMAT (code);
1745 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1747 if (fmt[i] == 'e')
1749 if (uses_hard_regs_p (XEXP (x, i), set))
1750 return true;
1752 else if (fmt[i] == 'E')
1754 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
1755 if (uses_hard_regs_p (XVECEXP (x, i, j), set))
1756 return true;
1759 return false;
1762 /* Return true if OP is a spilled pseudo. */
1763 static inline bool
1764 spilled_pseudo_p (rtx op)
1766 return (REG_P (op)
1767 && REGNO (op) >= FIRST_PSEUDO_REGISTER && in_mem_p (REGNO (op)));
1770 /* Return true if X is a general constant. */
1771 static inline bool
1772 general_constant_p (rtx x)
1774 return CONSTANT_P (x) && (! flag_pic || LEGITIMATE_PIC_OPERAND_P (x));
1777 static bool
1778 reg_in_class_p (rtx reg, enum reg_class cl)
1780 if (cl == NO_REGS)
1781 return get_reg_class (REGNO (reg)) == NO_REGS;
1782 return in_class_p (reg, cl, NULL);
1785 /* Return true if SET of RCLASS contains no hard regs which can be
1786 used in MODE. */
1787 static bool
1788 prohibited_class_reg_set_mode_p (enum reg_class rclass,
1789 HARD_REG_SET &set,
1790 enum machine_mode mode)
1792 HARD_REG_SET temp;
1794 lra_assert (hard_reg_set_subset_p (reg_class_contents[rclass], set));
1795 COPY_HARD_REG_SET (temp, set);
1796 AND_COMPL_HARD_REG_SET (temp, lra_no_alloc_regs);
1797 return (hard_reg_set_subset_p
1798 (temp, ira_prohibited_class_mode_regs[rclass][mode]));
1801 /* Major function to choose the current insn alternative and what
1802 operands should be reloaded and how. If ONLY_ALTERNATIVE is not
1803 negative we should consider only this alternative. Return false if
1804 we can not choose the alternative or find how to reload the
1805 operands. */
1806 static bool
1807 process_alt_operands (int only_alternative)
1809 bool ok_p = false;
1810 int nop, overall, nalt;
1811 int n_alternatives = curr_static_id->n_alternatives;
1812 int n_operands = curr_static_id->n_operands;
1813 /* LOSERS counts the operands that don't fit this alternative and
1814 would require loading. */
1815 int losers;
1816 /* REJECT is a count of how undesirable this alternative says it is
1817 if any reloading is required. If the alternative matches exactly
1818 then REJECT is ignored, but otherwise it gets this much counted
1819 against it in addition to the reloading needed. */
1820 int reject;
1821 int op_reject;
1822 /* The number of elements in the following array. */
1823 int early_clobbered_regs_num;
1824 /* Numbers of operands which are early clobber registers. */
1825 int early_clobbered_nops[MAX_RECOG_OPERANDS];
1826 enum reg_class curr_alt[MAX_RECOG_OPERANDS];
1827 HARD_REG_SET curr_alt_set[MAX_RECOG_OPERANDS];
1828 bool curr_alt_match_win[MAX_RECOG_OPERANDS];
1829 bool curr_alt_win[MAX_RECOG_OPERANDS];
1830 bool curr_alt_offmemok[MAX_RECOG_OPERANDS];
1831 int curr_alt_matches[MAX_RECOG_OPERANDS];
1832 /* The number of elements in the following array. */
1833 int curr_alt_dont_inherit_ops_num;
1834 /* Numbers of operands whose reload pseudos should not be inherited. */
1835 int curr_alt_dont_inherit_ops[MAX_RECOG_OPERANDS];
1836 rtx op;
1837 /* The register when the operand is a subreg of register, otherwise the
1838 operand itself. */
1839 rtx no_subreg_reg_operand[MAX_RECOG_OPERANDS];
1840 /* The register if the operand is a register or subreg of register,
1841 otherwise NULL. */
1842 rtx operand_reg[MAX_RECOG_OPERANDS];
1843 int hard_regno[MAX_RECOG_OPERANDS];
1844 machine_mode biggest_mode[MAX_RECOG_OPERANDS];
1845 int reload_nregs, reload_sum;
1846 bool costly_p;
1847 enum reg_class cl;
1849 /* Calculate some data common for all alternatives to speed up the
1850 function. */
1851 for (nop = 0; nop < n_operands; nop++)
1853 rtx reg;
1855 op = no_subreg_reg_operand[nop] = *curr_id->operand_loc[nop];
1856 /* The real hard regno of the operand after the allocation. */
1857 hard_regno[nop] = get_hard_regno (op, true);
1859 operand_reg[nop] = reg = op;
1860 biggest_mode[nop] = GET_MODE (op);
1861 if (GET_CODE (op) == SUBREG)
1863 operand_reg[nop] = reg = SUBREG_REG (op);
1864 if (GET_MODE_SIZE (biggest_mode[nop])
1865 < GET_MODE_SIZE (GET_MODE (reg)))
1866 biggest_mode[nop] = GET_MODE (reg);
1868 if (! REG_P (reg))
1869 operand_reg[nop] = NULL_RTX;
1870 else if (REGNO (reg) >= FIRST_PSEUDO_REGISTER
1871 || ((int) REGNO (reg)
1872 == lra_get_elimination_hard_regno (REGNO (reg))))
1873 no_subreg_reg_operand[nop] = reg;
1874 else
1875 operand_reg[nop] = no_subreg_reg_operand[nop]
1876 /* Just use natural mode for elimination result. It should
1877 be enough for extra constraints hooks. */
1878 = regno_reg_rtx[hard_regno[nop]];
1881 /* The constraints are made of several alternatives. Each operand's
1882 constraint looks like foo,bar,... with commas separating the
1883 alternatives. The first alternatives for all operands go
1884 together, the second alternatives go together, etc.
1886 First loop over alternatives. */
1887 alternative_mask preferred = curr_id->preferred_alternatives;
1888 if (only_alternative >= 0)
1889 preferred &= ALTERNATIVE_BIT (only_alternative);
1891 for (nalt = 0; nalt < n_alternatives; nalt++)
1893 /* Loop over operands for one constraint alternative. */
1894 if (!TEST_BIT (preferred, nalt))
1895 continue;
1897 overall = losers = reject = reload_nregs = reload_sum = 0;
1898 for (nop = 0; nop < n_operands; nop++)
1900 int inc = (curr_static_id
1901 ->operand_alternative[nalt * n_operands + nop].reject);
1902 if (lra_dump_file != NULL && inc != 0)
1903 fprintf (lra_dump_file,
1904 " Staticly defined alt reject+=%d\n", inc);
1905 reject += inc;
1907 early_clobbered_regs_num = 0;
1909 for (nop = 0; nop < n_operands; nop++)
1911 const char *p;
1912 char *end;
1913 int len, c, m, i, opalt_num, this_alternative_matches;
1914 bool win, did_match, offmemok, early_clobber_p;
1915 /* false => this operand can be reloaded somehow for this
1916 alternative. */
1917 bool badop;
1918 /* true => this operand can be reloaded if the alternative
1919 allows regs. */
1920 bool winreg;
1921 /* True if a constant forced into memory would be OK for
1922 this operand. */
1923 bool constmemok;
1924 enum reg_class this_alternative, this_costly_alternative;
1925 HARD_REG_SET this_alternative_set, this_costly_alternative_set;
1926 bool this_alternative_match_win, this_alternative_win;
1927 bool this_alternative_offmemok;
1928 bool scratch_p;
1929 machine_mode mode;
1930 enum constraint_num cn;
1932 opalt_num = nalt * n_operands + nop;
1933 if (curr_static_id->operand_alternative[opalt_num].anything_ok)
1935 /* Fast track for no constraints at all. */
1936 curr_alt[nop] = NO_REGS;
1937 CLEAR_HARD_REG_SET (curr_alt_set[nop]);
1938 curr_alt_win[nop] = true;
1939 curr_alt_match_win[nop] = false;
1940 curr_alt_offmemok[nop] = false;
1941 curr_alt_matches[nop] = -1;
1942 continue;
1945 op = no_subreg_reg_operand[nop];
1946 mode = curr_operand_mode[nop];
1948 win = did_match = winreg = offmemok = constmemok = false;
1949 badop = true;
1951 early_clobber_p = false;
1952 p = curr_static_id->operand_alternative[opalt_num].constraint;
1954 this_costly_alternative = this_alternative = NO_REGS;
1955 /* We update set of possible hard regs besides its class
1956 because reg class might be inaccurate. For example,
1957 union of LO_REGS (l), HI_REGS(h), and STACK_REG(k) in ARM
1958 is translated in HI_REGS because classes are merged by
1959 pairs and there is no accurate intermediate class. */
1960 CLEAR_HARD_REG_SET (this_alternative_set);
1961 CLEAR_HARD_REG_SET (this_costly_alternative_set);
1962 this_alternative_win = false;
1963 this_alternative_match_win = false;
1964 this_alternative_offmemok = false;
1965 this_alternative_matches = -1;
1967 /* An empty constraint should be excluded by the fast
1968 track. */
1969 lra_assert (*p != 0 && *p != ',');
1971 op_reject = 0;
1972 /* Scan this alternative's specs for this operand; set WIN
1973 if the operand fits any letter in this alternative.
1974 Otherwise, clear BADOP if this operand could fit some
1975 letter after reloads, or set WINREG if this operand could
1976 fit after reloads provided the constraint allows some
1977 registers. */
1978 costly_p = false;
1981 switch ((c = *p, len = CONSTRAINT_LEN (c, p)), c)
1983 case '\0':
1984 len = 0;
1985 break;
1986 case ',':
1987 c = '\0';
1988 break;
1990 case '&':
1991 early_clobber_p = true;
1992 break;
1994 case '$':
1995 op_reject += LRA_MAX_REJECT;
1996 break;
1997 case '^':
1998 op_reject += LRA_LOSER_COST_FACTOR;
1999 break;
2001 case '#':
2002 /* Ignore rest of this alternative. */
2003 c = '\0';
2004 break;
2006 case '0': case '1': case '2': case '3': case '4':
2007 case '5': case '6': case '7': case '8': case '9':
2009 int m_hregno;
2010 bool match_p;
2012 m = strtoul (p, &end, 10);
2013 p = end;
2014 len = 0;
2015 lra_assert (nop > m);
2017 this_alternative_matches = m;
2018 m_hregno = get_hard_regno (*curr_id->operand_loc[m], false);
2019 /* We are supposed to match a previous operand.
2020 If we do, we win if that one did. If we do
2021 not, count both of the operands as losers.
2022 (This is too conservative, since most of the
2023 time only a single reload insn will be needed
2024 to make the two operands win. As a result,
2025 this alternative may be rejected when it is
2026 actually desirable.) */
2027 match_p = false;
2028 if (operands_match_p (*curr_id->operand_loc[nop],
2029 *curr_id->operand_loc[m], m_hregno))
2031 /* We should reject matching of an early
2032 clobber operand if the matching operand is
2033 not dying in the insn. */
2034 if (! curr_static_id->operand[m].early_clobber
2035 || operand_reg[nop] == NULL_RTX
2036 || (find_regno_note (curr_insn, REG_DEAD,
2037 REGNO (op))
2038 || REGNO (op) == REGNO (operand_reg[m])))
2039 match_p = true;
2041 if (match_p)
2043 /* If we are matching a non-offsettable
2044 address where an offsettable address was
2045 expected, then we must reject this
2046 combination, because we can't reload
2047 it. */
2048 if (curr_alt_offmemok[m]
2049 && MEM_P (*curr_id->operand_loc[m])
2050 && curr_alt[m] == NO_REGS && ! curr_alt_win[m])
2051 continue;
2053 else
2055 /* Operands don't match. Both operands must
2056 allow a reload register, otherwise we
2057 cannot make them match. */
2058 if (curr_alt[m] == NO_REGS)
2059 break;
2060 /* Retroactively mark the operand we had to
2061 match as a loser, if it wasn't already and
2062 it wasn't matched to a register constraint
2063 (e.g it might be matched by memory). */
2064 if (curr_alt_win[m]
2065 && (operand_reg[m] == NULL_RTX
2066 || hard_regno[m] < 0))
2068 losers++;
2069 reload_nregs
2070 += (ira_reg_class_max_nregs[curr_alt[m]]
2071 [GET_MODE (*curr_id->operand_loc[m])]);
2074 /* Prefer matching earlyclobber alternative as
2075 it results in less hard regs required for
2076 the insn than a non-matching earlyclobber
2077 alternative. */
2078 if (curr_static_id->operand[m].early_clobber)
2080 if (lra_dump_file != NULL)
2081 fprintf
2082 (lra_dump_file,
2083 " %d Matching earlyclobber alt:"
2084 " reject--\n",
2085 nop);
2086 reject--;
2088 /* Otherwise we prefer no matching
2089 alternatives because it gives more freedom
2090 in RA. */
2091 else if (operand_reg[nop] == NULL_RTX
2092 || (find_regno_note (curr_insn, REG_DEAD,
2093 REGNO (operand_reg[nop]))
2094 == NULL_RTX))
2096 if (lra_dump_file != NULL)
2097 fprintf
2098 (lra_dump_file,
2099 " %d Matching alt: reject+=2\n",
2100 nop);
2101 reject += 2;
2104 /* If we have to reload this operand and some
2105 previous operand also had to match the same
2106 thing as this operand, we don't know how to do
2107 that. */
2108 if (!match_p || !curr_alt_win[m])
2110 for (i = 0; i < nop; i++)
2111 if (curr_alt_matches[i] == m)
2112 break;
2113 if (i < nop)
2114 break;
2116 else
2117 did_match = true;
2119 /* This can be fixed with reloads if the operand
2120 we are supposed to match can be fixed with
2121 reloads. */
2122 badop = false;
2123 this_alternative = curr_alt[m];
2124 COPY_HARD_REG_SET (this_alternative_set, curr_alt_set[m]);
2125 winreg = this_alternative != NO_REGS;
2126 break;
2129 case 'g':
2130 if (MEM_P (op)
2131 || general_constant_p (op)
2132 || spilled_pseudo_p (op))
2133 win = true;
2134 cl = GENERAL_REGS;
2135 goto reg;
2137 default:
2138 cn = lookup_constraint (p);
2139 switch (get_constraint_type (cn))
2141 case CT_REGISTER:
2142 cl = reg_class_for_constraint (cn);
2143 if (cl != NO_REGS)
2144 goto reg;
2145 break;
2147 case CT_CONST_INT:
2148 if (CONST_INT_P (op)
2149 && insn_const_int_ok_for_constraint (INTVAL (op), cn))
2150 win = true;
2151 break;
2153 case CT_MEMORY:
2154 if (MEM_P (op)
2155 && satisfies_memory_constraint_p (op, cn))
2156 win = true;
2157 else if (spilled_pseudo_p (op))
2158 win = true;
2160 /* If we didn't already win, we can reload constants
2161 via force_const_mem or put the pseudo value into
2162 memory, or make other memory by reloading the
2163 address like for 'o'. */
2164 if (CONST_POOL_OK_P (mode, op)
2165 || MEM_P (op) || REG_P (op)
2166 /* We can restore the equiv insn by a
2167 reload. */
2168 || equiv_substition_p[nop])
2169 badop = false;
2170 constmemok = true;
2171 offmemok = true;
2172 break;
2174 case CT_ADDRESS:
2175 /* If we didn't already win, we can reload the address
2176 into a base register. */
2177 if (satisfies_address_constraint_p (op, cn))
2178 win = true;
2179 cl = base_reg_class (VOIDmode, ADDR_SPACE_GENERIC,
2180 ADDRESS, SCRATCH);
2181 badop = false;
2182 goto reg;
2184 case CT_FIXED_FORM:
2185 if (constraint_satisfied_p (op, cn))
2186 win = true;
2187 break;
2189 case CT_SPECIAL_MEMORY:
2190 if (MEM_P (op)
2191 && satisfies_memory_constraint_p (op, cn))
2192 win = true;
2193 else if (spilled_pseudo_p (op))
2194 win = true;
2195 break;
2197 break;
2199 reg:
2200 this_alternative = reg_class_subunion[this_alternative][cl];
2201 IOR_HARD_REG_SET (this_alternative_set,
2202 reg_class_contents[cl]);
2203 if (costly_p)
2205 this_costly_alternative
2206 = reg_class_subunion[this_costly_alternative][cl];
2207 IOR_HARD_REG_SET (this_costly_alternative_set,
2208 reg_class_contents[cl]);
2210 if (mode == BLKmode)
2211 break;
2212 winreg = true;
2213 if (REG_P (op))
2215 if (hard_regno[nop] >= 0
2216 && in_hard_reg_set_p (this_alternative_set,
2217 mode, hard_regno[nop]))
2218 win = true;
2219 else if (hard_regno[nop] < 0
2220 && in_class_p (op, this_alternative, NULL))
2221 win = true;
2223 break;
2225 if (c != ' ' && c != '\t')
2226 costly_p = c == '*';
2228 while ((p += len), c);
2230 scratch_p = (operand_reg[nop] != NULL_RTX
2231 && lra_former_scratch_p (REGNO (operand_reg[nop])));
2232 /* Record which operands fit this alternative. */
2233 if (win)
2235 this_alternative_win = true;
2236 if (operand_reg[nop] != NULL_RTX)
2238 if (hard_regno[nop] >= 0)
2240 if (in_hard_reg_set_p (this_costly_alternative_set,
2241 mode, hard_regno[nop]))
2243 if (lra_dump_file != NULL)
2244 fprintf (lra_dump_file,
2245 " %d Costly set: reject++\n",
2246 nop);
2247 reject++;
2250 else
2252 /* Prefer won reg to spilled pseudo under other
2253 equal conditions for possibe inheritance. */
2254 if (! scratch_p)
2256 if (lra_dump_file != NULL)
2257 fprintf
2258 (lra_dump_file,
2259 " %d Non pseudo reload: reject++\n",
2260 nop);
2261 reject++;
2263 if (in_class_p (operand_reg[nop],
2264 this_costly_alternative, NULL))
2266 if (lra_dump_file != NULL)
2267 fprintf
2268 (lra_dump_file,
2269 " %d Non pseudo costly reload:"
2270 " reject++\n",
2271 nop);
2272 reject++;
2275 /* We simulate the behavior of old reload here.
2276 Although scratches need hard registers and it
2277 might result in spilling other pseudos, no reload
2278 insns are generated for the scratches. So it
2279 might cost something but probably less than old
2280 reload pass believes. */
2281 if (scratch_p)
2283 if (lra_dump_file != NULL)
2284 fprintf (lra_dump_file,
2285 " %d Scratch win: reject+=2\n",
2286 nop);
2287 reject += 2;
2291 else if (did_match)
2292 this_alternative_match_win = true;
2293 else
2295 int const_to_mem = 0;
2296 bool no_regs_p;
2298 reject += op_reject;
2299 /* Never do output reload of stack pointer. It makes
2300 impossible to do elimination when SP is changed in
2301 RTL. */
2302 if (op == stack_pointer_rtx && ! frame_pointer_needed
2303 && curr_static_id->operand[nop].type != OP_IN)
2304 goto fail;
2306 /* If this alternative asks for a specific reg class, see if there
2307 is at least one allocatable register in that class. */
2308 no_regs_p
2309 = (this_alternative == NO_REGS
2310 || (hard_reg_set_subset_p
2311 (reg_class_contents[this_alternative],
2312 lra_no_alloc_regs)));
2314 /* For asms, verify that the class for this alternative is possible
2315 for the mode that is specified. */
2316 if (!no_regs_p && INSN_CODE (curr_insn) < 0)
2318 int i;
2319 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
2320 if (HARD_REGNO_MODE_OK (i, mode)
2321 && in_hard_reg_set_p (reg_class_contents[this_alternative],
2322 mode, i))
2323 break;
2324 if (i == FIRST_PSEUDO_REGISTER)
2325 winreg = false;
2328 /* If this operand accepts a register, and if the
2329 register class has at least one allocatable register,
2330 then this operand can be reloaded. */
2331 if (winreg && !no_regs_p)
2332 badop = false;
2334 if (badop)
2336 if (lra_dump_file != NULL)
2337 fprintf (lra_dump_file,
2338 " alt=%d: Bad operand -- refuse\n",
2339 nalt);
2340 goto fail;
2343 if (this_alternative != NO_REGS)
2345 HARD_REG_SET available_regs;
2347 COPY_HARD_REG_SET (available_regs,
2348 reg_class_contents[this_alternative]);
2349 AND_COMPL_HARD_REG_SET
2350 (available_regs,
2351 ira_prohibited_class_mode_regs[this_alternative][mode]);
2352 AND_COMPL_HARD_REG_SET (available_regs, lra_no_alloc_regs);
2353 if (hard_reg_set_empty_p (available_regs))
2355 /* There are no hard regs holding a value of given
2356 mode. */
2357 if (offmemok)
2359 this_alternative = NO_REGS;
2360 if (lra_dump_file != NULL)
2361 fprintf (lra_dump_file,
2362 " %d Using memory because of"
2363 " a bad mode: reject+=2\n",
2364 nop);
2365 reject += 2;
2367 else
2369 if (lra_dump_file != NULL)
2370 fprintf (lra_dump_file,
2371 " alt=%d: Wrong mode -- refuse\n",
2372 nalt);
2373 goto fail;
2378 /* If not assigned pseudo has a class which a subset of
2379 required reg class, it is a less costly alternative
2380 as the pseudo still can get a hard reg of necessary
2381 class. */
2382 if (! no_regs_p && REG_P (op) && hard_regno[nop] < 0
2383 && (cl = get_reg_class (REGNO (op))) != NO_REGS
2384 && ira_class_subset_p[this_alternative][cl])
2386 if (lra_dump_file != NULL)
2387 fprintf
2388 (lra_dump_file,
2389 " %d Super set class reg: reject-=3\n", nop);
2390 reject -= 3;
2393 this_alternative_offmemok = offmemok;
2394 if (this_costly_alternative != NO_REGS)
2396 if (lra_dump_file != NULL)
2397 fprintf (lra_dump_file,
2398 " %d Costly loser: reject++\n", nop);
2399 reject++;
2401 /* If the operand is dying, has a matching constraint,
2402 and satisfies constraints of the matched operand
2403 which failed to satisfy the own constraints, most probably
2404 the reload for this operand will be gone. */
2405 if (this_alternative_matches >= 0
2406 && !curr_alt_win[this_alternative_matches]
2407 && REG_P (op)
2408 && find_regno_note (curr_insn, REG_DEAD, REGNO (op))
2409 && (hard_regno[nop] >= 0
2410 ? in_hard_reg_set_p (this_alternative_set,
2411 mode, hard_regno[nop])
2412 : in_class_p (op, this_alternative, NULL)))
2414 if (lra_dump_file != NULL)
2415 fprintf
2416 (lra_dump_file,
2417 " %d Dying matched operand reload: reject++\n",
2418 nop);
2419 reject++;
2421 else
2423 /* Strict_low_part requires to reload the register
2424 not the sub-register. In this case we should
2425 check that a final reload hard reg can hold the
2426 value mode. */
2427 if (curr_static_id->operand[nop].strict_low
2428 && REG_P (op)
2429 && hard_regno[nop] < 0
2430 && GET_CODE (*curr_id->operand_loc[nop]) == SUBREG
2431 && ira_class_hard_regs_num[this_alternative] > 0
2432 && ! HARD_REGNO_MODE_OK (ira_class_hard_regs
2433 [this_alternative][0],
2434 GET_MODE
2435 (*curr_id->operand_loc[nop])))
2437 if (lra_dump_file != NULL)
2438 fprintf
2439 (lra_dump_file,
2440 " alt=%d: Strict low subreg reload -- refuse\n",
2441 nalt);
2442 goto fail;
2444 losers++;
2446 if (operand_reg[nop] != NULL_RTX
2447 /* Output operands and matched input operands are
2448 not inherited. The following conditions do not
2449 exactly describe the previous statement but they
2450 are pretty close. */
2451 && curr_static_id->operand[nop].type != OP_OUT
2452 && (this_alternative_matches < 0
2453 || curr_static_id->operand[nop].type != OP_IN))
2455 int last_reload = (lra_reg_info[ORIGINAL_REGNO
2456 (operand_reg[nop])]
2457 .last_reload);
2459 /* The value of reload_sum has sense only if we
2460 process insns in their order. It happens only on
2461 the first constraints sub-pass when we do most of
2462 reload work. */
2463 if (lra_constraint_iter == 1 && last_reload > bb_reload_num)
2464 reload_sum += last_reload - bb_reload_num;
2466 /* If this is a constant that is reloaded into the
2467 desired class by copying it to memory first, count
2468 that as another reload. This is consistent with
2469 other code and is required to avoid choosing another
2470 alternative when the constant is moved into memory.
2471 Note that the test here is precisely the same as in
2472 the code below that calls force_const_mem. */
2473 if (CONST_POOL_OK_P (mode, op)
2474 && ((targetm.preferred_reload_class
2475 (op, this_alternative) == NO_REGS)
2476 || no_input_reloads_p))
2478 const_to_mem = 1;
2479 if (! no_regs_p)
2480 losers++;
2483 /* Alternative loses if it requires a type of reload not
2484 permitted for this insn. We can always reload
2485 objects with a REG_UNUSED note. */
2486 if ((curr_static_id->operand[nop].type != OP_IN
2487 && no_output_reloads_p
2488 && ! find_reg_note (curr_insn, REG_UNUSED, op))
2489 || (curr_static_id->operand[nop].type != OP_OUT
2490 && no_input_reloads_p && ! const_to_mem)
2491 || (this_alternative_matches >= 0
2492 && (no_input_reloads_p
2493 || (no_output_reloads_p
2494 && (curr_static_id->operand
2495 [this_alternative_matches].type != OP_IN)
2496 && ! find_reg_note (curr_insn, REG_UNUSED,
2497 no_subreg_reg_operand
2498 [this_alternative_matches])))))
2500 if (lra_dump_file != NULL)
2501 fprintf
2502 (lra_dump_file,
2503 " alt=%d: No input/otput reload -- refuse\n",
2504 nalt);
2505 goto fail;
2508 /* Alternative loses if it required class pseudo can not
2509 hold value of required mode. Such insns can be
2510 described by insn definitions with mode iterators. */
2511 if (GET_MODE (*curr_id->operand_loc[nop]) != VOIDmode
2512 && ! hard_reg_set_empty_p (this_alternative_set)
2513 /* It is common practice for constraints to use a
2514 class which does not have actually enough regs to
2515 hold the value (e.g. x86 AREG for mode requiring
2516 more one general reg). Therefore we have 2
2517 conditions to check that the reload pseudo can
2518 not hold the mode value. */
2519 && ! HARD_REGNO_MODE_OK (ira_class_hard_regs
2520 [this_alternative][0],
2521 GET_MODE (*curr_id->operand_loc[nop]))
2522 /* The above condition is not enough as the first
2523 reg in ira_class_hard_regs can be not aligned for
2524 multi-words mode values. */
2525 && (prohibited_class_reg_set_mode_p
2526 (this_alternative, this_alternative_set,
2527 GET_MODE (*curr_id->operand_loc[nop]))))
2529 if (lra_dump_file != NULL)
2530 fprintf (lra_dump_file,
2531 " alt=%d: reload pseudo for op %d "
2532 " can not hold the mode value -- refuse\n",
2533 nalt, nop);
2534 goto fail;
2537 /* Check strong discouragement of reload of non-constant
2538 into class THIS_ALTERNATIVE. */
2539 if (! CONSTANT_P (op) && ! no_regs_p
2540 && (targetm.preferred_reload_class
2541 (op, this_alternative) == NO_REGS
2542 || (curr_static_id->operand[nop].type == OP_OUT
2543 && (targetm.preferred_output_reload_class
2544 (op, this_alternative) == NO_REGS))))
2546 if (lra_dump_file != NULL)
2547 fprintf (lra_dump_file,
2548 " %d Non-prefered reload: reject+=%d\n",
2549 nop, LRA_MAX_REJECT);
2550 reject += LRA_MAX_REJECT;
2553 if (! (MEM_P (op) && offmemok)
2554 && ! (const_to_mem && constmemok))
2556 /* We prefer to reload pseudos over reloading other
2557 things, since such reloads may be able to be
2558 eliminated later. So bump REJECT in other cases.
2559 Don't do this in the case where we are forcing a
2560 constant into memory and it will then win since
2561 we don't want to have a different alternative
2562 match then. */
2563 if (! (REG_P (op) && REGNO (op) >= FIRST_PSEUDO_REGISTER))
2565 if (lra_dump_file != NULL)
2566 fprintf
2567 (lra_dump_file,
2568 " %d Non-pseudo reload: reject+=2\n",
2569 nop);
2570 reject += 2;
2573 if (! no_regs_p)
2574 reload_nregs
2575 += ira_reg_class_max_nregs[this_alternative][mode];
2577 if (SMALL_REGISTER_CLASS_P (this_alternative))
2579 if (lra_dump_file != NULL)
2580 fprintf
2581 (lra_dump_file,
2582 " %d Small class reload: reject+=%d\n",
2583 nop, LRA_LOSER_COST_FACTOR / 2);
2584 reject += LRA_LOSER_COST_FACTOR / 2;
2588 /* We are trying to spill pseudo into memory. It is
2589 usually more costly than moving to a hard register
2590 although it might takes the same number of
2591 reloads.
2593 Non-pseudo spill may happen also. Suppose a target allows both
2594 register and memory in the operand constraint alternatives,
2595 then it's typical that an eliminable register has a substition
2596 of "base + offset" which can either be reloaded by a simple
2597 "new_reg <= base + offset" which will match the register
2598 constraint, or a similar reg addition followed by further spill
2599 to and reload from memory which will match the memory
2600 constraint, but this memory spill will be much more costly
2601 usually.
2603 Code below increases the reject for both pseudo and non-pseudo
2604 spill. */
2605 if (no_regs_p
2606 && !(MEM_P (op) && offmemok)
2607 && !(REG_P (op) && hard_regno[nop] < 0))
2609 if (lra_dump_file != NULL)
2610 fprintf
2611 (lra_dump_file,
2612 " %d Spill %spseudo into memory: reject+=3\n",
2613 nop, REG_P (op) ? "" : "Non-");
2614 reject += 3;
2615 if (VECTOR_MODE_P (mode))
2617 /* Spilling vectors into memory is usually more
2618 costly as they contain big values. */
2619 if (lra_dump_file != NULL)
2620 fprintf
2621 (lra_dump_file,
2622 " %d Spill vector pseudo: reject+=2\n",
2623 nop);
2624 reject += 2;
2628 #ifdef SECONDARY_MEMORY_NEEDED
2629 /* If reload requires moving value through secondary
2630 memory, it will need one more insn at least. */
2631 if (this_alternative != NO_REGS
2632 && REG_P (op) && (cl = get_reg_class (REGNO (op))) != NO_REGS
2633 && ((curr_static_id->operand[nop].type != OP_OUT
2634 && SECONDARY_MEMORY_NEEDED (cl, this_alternative,
2635 GET_MODE (op)))
2636 || (curr_static_id->operand[nop].type != OP_IN
2637 && SECONDARY_MEMORY_NEEDED (this_alternative, cl,
2638 GET_MODE (op)))))
2639 losers++;
2640 #endif
2641 /* Input reloads can be inherited more often than output
2642 reloads can be removed, so penalize output
2643 reloads. */
2644 if (!REG_P (op) || curr_static_id->operand[nop].type != OP_IN)
2646 if (lra_dump_file != NULL)
2647 fprintf
2648 (lra_dump_file,
2649 " %d Non input pseudo reload: reject++\n",
2650 nop);
2651 reject++;
2655 if (early_clobber_p && ! scratch_p)
2657 if (lra_dump_file != NULL)
2658 fprintf (lra_dump_file,
2659 " %d Early clobber: reject++\n", nop);
2660 reject++;
2662 /* ??? We check early clobbers after processing all operands
2663 (see loop below) and there we update the costs more.
2664 Should we update the cost (may be approximately) here
2665 because of early clobber register reloads or it is a rare
2666 or non-important thing to be worth to do it. */
2667 overall = losers * LRA_LOSER_COST_FACTOR + reject;
2668 if ((best_losers == 0 || losers != 0) && best_overall < overall)
2670 if (lra_dump_file != NULL)
2671 fprintf (lra_dump_file,
2672 " alt=%d,overall=%d,losers=%d -- refuse\n",
2673 nalt, overall, losers);
2674 goto fail;
2677 curr_alt[nop] = this_alternative;
2678 COPY_HARD_REG_SET (curr_alt_set[nop], this_alternative_set);
2679 curr_alt_win[nop] = this_alternative_win;
2680 curr_alt_match_win[nop] = this_alternative_match_win;
2681 curr_alt_offmemok[nop] = this_alternative_offmemok;
2682 curr_alt_matches[nop] = this_alternative_matches;
2684 if (this_alternative_matches >= 0
2685 && !did_match && !this_alternative_win)
2686 curr_alt_win[this_alternative_matches] = false;
2688 if (early_clobber_p && operand_reg[nop] != NULL_RTX)
2689 early_clobbered_nops[early_clobbered_regs_num++] = nop;
2691 if (curr_insn_set != NULL_RTX && n_operands == 2
2692 /* Prevent processing non-move insns. */
2693 && (GET_CODE (SET_SRC (curr_insn_set)) == SUBREG
2694 || SET_SRC (curr_insn_set) == no_subreg_reg_operand[1])
2695 && ((! curr_alt_win[0] && ! curr_alt_win[1]
2696 && REG_P (no_subreg_reg_operand[0])
2697 && REG_P (no_subreg_reg_operand[1])
2698 && (reg_in_class_p (no_subreg_reg_operand[0], curr_alt[1])
2699 || reg_in_class_p (no_subreg_reg_operand[1], curr_alt[0])))
2700 || (! curr_alt_win[0] && curr_alt_win[1]
2701 && REG_P (no_subreg_reg_operand[1])
2702 && reg_in_class_p (no_subreg_reg_operand[1], curr_alt[0]))
2703 || (curr_alt_win[0] && ! curr_alt_win[1]
2704 && REG_P (no_subreg_reg_operand[0])
2705 && reg_in_class_p (no_subreg_reg_operand[0], curr_alt[1])
2706 && (! CONST_POOL_OK_P (curr_operand_mode[1],
2707 no_subreg_reg_operand[1])
2708 || (targetm.preferred_reload_class
2709 (no_subreg_reg_operand[1],
2710 (enum reg_class) curr_alt[1]) != NO_REGS))
2711 /* If it is a result of recent elimination in move
2712 insn we can transform it into an add still by
2713 using this alternative. */
2714 && GET_CODE (no_subreg_reg_operand[1]) != PLUS)))
2716 /* We have a move insn and a new reload insn will be similar
2717 to the current insn. We should avoid such situation as it
2718 results in LRA cycling. */
2719 overall += LRA_MAX_REJECT;
2721 ok_p = true;
2722 curr_alt_dont_inherit_ops_num = 0;
2723 for (nop = 0; nop < early_clobbered_regs_num; nop++)
2725 int i, j, clobbered_hard_regno, first_conflict_j, last_conflict_j;
2726 HARD_REG_SET temp_set;
2728 i = early_clobbered_nops[nop];
2729 if ((! curr_alt_win[i] && ! curr_alt_match_win[i])
2730 || hard_regno[i] < 0)
2731 continue;
2732 lra_assert (operand_reg[i] != NULL_RTX);
2733 clobbered_hard_regno = hard_regno[i];
2734 CLEAR_HARD_REG_SET (temp_set);
2735 add_to_hard_reg_set (&temp_set, biggest_mode[i], clobbered_hard_regno);
2736 first_conflict_j = last_conflict_j = -1;
2737 for (j = 0; j < n_operands; j++)
2738 if (j == i
2739 /* We don't want process insides of match_operator and
2740 match_parallel because otherwise we would process
2741 their operands once again generating a wrong
2742 code. */
2743 || curr_static_id->operand[j].is_operator)
2744 continue;
2745 else if ((curr_alt_matches[j] == i && curr_alt_match_win[j])
2746 || (curr_alt_matches[i] == j && curr_alt_match_win[i]))
2747 continue;
2748 /* If we don't reload j-th operand, check conflicts. */
2749 else if ((curr_alt_win[j] || curr_alt_match_win[j])
2750 && uses_hard_regs_p (*curr_id->operand_loc[j], temp_set))
2752 if (first_conflict_j < 0)
2753 first_conflict_j = j;
2754 last_conflict_j = j;
2756 if (last_conflict_j < 0)
2757 continue;
2758 /* If earlyclobber operand conflicts with another
2759 non-matching operand which is actually the same register
2760 as the earlyclobber operand, it is better to reload the
2761 another operand as an operand matching the earlyclobber
2762 operand can be also the same. */
2763 if (first_conflict_j == last_conflict_j
2764 && operand_reg[last_conflict_j] != NULL_RTX
2765 && ! curr_alt_match_win[last_conflict_j]
2766 && REGNO (operand_reg[i]) == REGNO (operand_reg[last_conflict_j]))
2768 curr_alt_win[last_conflict_j] = false;
2769 curr_alt_dont_inherit_ops[curr_alt_dont_inherit_ops_num++]
2770 = last_conflict_j;
2771 losers++;
2772 /* Early clobber was already reflected in REJECT. */
2773 lra_assert (reject > 0);
2774 if (lra_dump_file != NULL)
2775 fprintf
2776 (lra_dump_file,
2777 " %d Conflict early clobber reload: reject--\n",
2779 reject--;
2780 overall += LRA_LOSER_COST_FACTOR - 1;
2782 else
2784 /* We need to reload early clobbered register and the
2785 matched registers. */
2786 for (j = 0; j < n_operands; j++)
2787 if (curr_alt_matches[j] == i)
2789 curr_alt_match_win[j] = false;
2790 losers++;
2791 overall += LRA_LOSER_COST_FACTOR;
2793 if (! curr_alt_match_win[i])
2794 curr_alt_dont_inherit_ops[curr_alt_dont_inherit_ops_num++] = i;
2795 else
2797 /* Remember pseudos used for match reloads are never
2798 inherited. */
2799 lra_assert (curr_alt_matches[i] >= 0);
2800 curr_alt_win[curr_alt_matches[i]] = false;
2802 curr_alt_win[i] = curr_alt_match_win[i] = false;
2803 losers++;
2804 /* Early clobber was already reflected in REJECT. */
2805 lra_assert (reject > 0);
2806 if (lra_dump_file != NULL)
2807 fprintf
2808 (lra_dump_file,
2809 " %d Matched conflict early clobber reloads:"
2810 "reject--\n",
2812 reject--;
2813 overall += LRA_LOSER_COST_FACTOR - 1;
2816 if (lra_dump_file != NULL)
2817 fprintf (lra_dump_file, " alt=%d,overall=%d,losers=%d,rld_nregs=%d\n",
2818 nalt, overall, losers, reload_nregs);
2820 /* If this alternative can be made to work by reloading, and it
2821 needs less reloading than the others checked so far, record
2822 it as the chosen goal for reloading. */
2823 if ((best_losers != 0 && losers == 0)
2824 || (((best_losers == 0 && losers == 0)
2825 || (best_losers != 0 && losers != 0))
2826 && (best_overall > overall
2827 || (best_overall == overall
2828 /* If the cost of the reloads is the same,
2829 prefer alternative which requires minimal
2830 number of reload regs. */
2831 && (reload_nregs < best_reload_nregs
2832 || (reload_nregs == best_reload_nregs
2833 && (best_reload_sum < reload_sum
2834 || (best_reload_sum == reload_sum
2835 && nalt < goal_alt_number))))))))
2837 for (nop = 0; nop < n_operands; nop++)
2839 goal_alt_win[nop] = curr_alt_win[nop];
2840 goal_alt_match_win[nop] = curr_alt_match_win[nop];
2841 goal_alt_matches[nop] = curr_alt_matches[nop];
2842 goal_alt[nop] = curr_alt[nop];
2843 goal_alt_offmemok[nop] = curr_alt_offmemok[nop];
2845 goal_alt_dont_inherit_ops_num = curr_alt_dont_inherit_ops_num;
2846 for (nop = 0; nop < curr_alt_dont_inherit_ops_num; nop++)
2847 goal_alt_dont_inherit_ops[nop] = curr_alt_dont_inherit_ops[nop];
2848 goal_alt_swapped = curr_swapped;
2849 best_overall = overall;
2850 best_losers = losers;
2851 best_reload_nregs = reload_nregs;
2852 best_reload_sum = reload_sum;
2853 goal_alt_number = nalt;
2855 if (losers == 0)
2856 /* Everything is satisfied. Do not process alternatives
2857 anymore. */
2858 break;
2859 fail:
2862 return ok_p;
2865 /* Make reload base reg from address AD. */
2866 static rtx
2867 base_to_reg (struct address_info *ad)
2869 enum reg_class cl;
2870 int code = -1;
2871 rtx new_inner = NULL_RTX;
2872 rtx new_reg = NULL_RTX;
2873 rtx_insn *insn;
2874 rtx_insn *last_insn = get_last_insn();
2876 lra_assert (ad->base == ad->base_term && ad->disp == ad->disp_term);
2877 cl = base_reg_class (ad->mode, ad->as, ad->base_outer_code,
2878 get_index_code (ad));
2879 new_reg = lra_create_new_reg (GET_MODE (*ad->base_term), NULL_RTX,
2880 cl, "base");
2881 new_inner = simplify_gen_binary (PLUS, GET_MODE (new_reg), new_reg,
2882 ad->disp_term == NULL
2883 ? gen_int_mode (0, ad->mode)
2884 : *ad->disp_term);
2885 if (!valid_address_p (ad->mode, new_inner, ad->as))
2886 return NULL_RTX;
2887 insn = emit_insn (gen_rtx_SET (new_reg, *ad->base_term));
2888 code = recog_memoized (insn);
2889 if (code < 0)
2891 delete_insns_since (last_insn);
2892 return NULL_RTX;
2895 return new_inner;
2898 /* Make reload base reg + disp from address AD. Return the new pseudo. */
2899 static rtx
2900 base_plus_disp_to_reg (struct address_info *ad)
2902 enum reg_class cl;
2903 rtx new_reg;
2905 lra_assert (ad->base == ad->base_term && ad->disp == ad->disp_term);
2906 cl = base_reg_class (ad->mode, ad->as, ad->base_outer_code,
2907 get_index_code (ad));
2908 new_reg = lra_create_new_reg (GET_MODE (*ad->base_term), NULL_RTX,
2909 cl, "base + disp");
2910 lra_emit_add (new_reg, *ad->base_term, *ad->disp_term);
2911 return new_reg;
2914 /* Make reload of index part of address AD. Return the new
2915 pseudo. */
2916 static rtx
2917 index_part_to_reg (struct address_info *ad)
2919 rtx new_reg;
2921 new_reg = lra_create_new_reg (GET_MODE (*ad->index), NULL_RTX,
2922 INDEX_REG_CLASS, "index term");
2923 expand_mult (GET_MODE (*ad->index), *ad->index_term,
2924 GEN_INT (get_index_scale (ad)), new_reg, 1);
2925 return new_reg;
2928 /* Return true if we can add a displacement to address AD, even if that
2929 makes the address invalid. The fix-up code requires any new address
2930 to be the sum of the BASE_TERM, INDEX and DISP_TERM fields. */
2931 static bool
2932 can_add_disp_p (struct address_info *ad)
2934 return (!ad->autoinc_p
2935 && ad->segment == NULL
2936 && ad->base == ad->base_term
2937 && ad->disp == ad->disp_term);
2940 /* Make equiv substitution in address AD. Return true if a substitution
2941 was made. */
2942 static bool
2943 equiv_address_substitution (struct address_info *ad)
2945 rtx base_reg, new_base_reg, index_reg, new_index_reg, *base_term, *index_term;
2946 HOST_WIDE_INT disp, scale;
2947 bool change_p;
2949 base_term = strip_subreg (ad->base_term);
2950 if (base_term == NULL)
2951 base_reg = new_base_reg = NULL_RTX;
2952 else
2954 base_reg = *base_term;
2955 new_base_reg = get_equiv_with_elimination (base_reg, curr_insn);
2957 index_term = strip_subreg (ad->index_term);
2958 if (index_term == NULL)
2959 index_reg = new_index_reg = NULL_RTX;
2960 else
2962 index_reg = *index_term;
2963 new_index_reg = get_equiv_with_elimination (index_reg, curr_insn);
2965 if (base_reg == new_base_reg && index_reg == new_index_reg)
2966 return false;
2967 disp = 0;
2968 change_p = false;
2969 if (lra_dump_file != NULL)
2971 fprintf (lra_dump_file, "Changing address in insn %d ",
2972 INSN_UID (curr_insn));
2973 dump_value_slim (lra_dump_file, *ad->outer, 1);
2975 if (base_reg != new_base_reg)
2977 if (REG_P (new_base_reg))
2979 *base_term = new_base_reg;
2980 change_p = true;
2982 else if (GET_CODE (new_base_reg) == PLUS
2983 && REG_P (XEXP (new_base_reg, 0))
2984 && CONST_INT_P (XEXP (new_base_reg, 1))
2985 && can_add_disp_p (ad))
2987 disp += INTVAL (XEXP (new_base_reg, 1));
2988 *base_term = XEXP (new_base_reg, 0);
2989 change_p = true;
2991 if (ad->base_term2 != NULL)
2992 *ad->base_term2 = *ad->base_term;
2994 if (index_reg != new_index_reg)
2996 if (REG_P (new_index_reg))
2998 *index_term = new_index_reg;
2999 change_p = true;
3001 else if (GET_CODE (new_index_reg) == PLUS
3002 && REG_P (XEXP (new_index_reg, 0))
3003 && CONST_INT_P (XEXP (new_index_reg, 1))
3004 && can_add_disp_p (ad)
3005 && (scale = get_index_scale (ad)))
3007 disp += INTVAL (XEXP (new_index_reg, 1)) * scale;
3008 *index_term = XEXP (new_index_reg, 0);
3009 change_p = true;
3012 if (disp != 0)
3014 if (ad->disp != NULL)
3015 *ad->disp = plus_constant (GET_MODE (*ad->inner), *ad->disp, disp);
3016 else
3018 *ad->inner = plus_constant (GET_MODE (*ad->inner), *ad->inner, disp);
3019 update_address (ad);
3021 change_p = true;
3023 if (lra_dump_file != NULL)
3025 if (! change_p)
3026 fprintf (lra_dump_file, " -- no change\n");
3027 else
3029 fprintf (lra_dump_file, " on equiv ");
3030 dump_value_slim (lra_dump_file, *ad->outer, 1);
3031 fprintf (lra_dump_file, "\n");
3034 return change_p;
3037 /* Major function to make reloads for an address in operand NOP or
3038 check its correctness (If CHECK_ONLY_P is true). The supported
3039 cases are:
3041 1) an address that existed before LRA started, at which point it
3042 must have been valid. These addresses are subject to elimination
3043 and may have become invalid due to the elimination offset being out
3044 of range.
3046 2) an address created by forcing a constant to memory
3047 (force_const_to_mem). The initial form of these addresses might
3048 not be valid, and it is this function's job to make them valid.
3050 3) a frame address formed from a register and a (possibly zero)
3051 constant offset. As above, these addresses might not be valid and
3052 this function must make them so.
3054 Add reloads to the lists *BEFORE and *AFTER. We might need to add
3055 reloads to *AFTER because of inc/dec, {pre, post} modify in the
3056 address. Return true for any RTL change.
3058 The function is a helper function which does not produce all
3059 transformations (when CHECK_ONLY_P is false) which can be
3060 necessary. It does just basic steps. To do all necessary
3061 transformations use function process_address. */
3062 static bool
3063 process_address_1 (int nop, bool check_only_p,
3064 rtx_insn **before, rtx_insn **after)
3066 struct address_info ad;
3067 rtx new_reg;
3068 HOST_WIDE_INT scale;
3069 rtx op = *curr_id->operand_loc[nop];
3070 const char *constraint = curr_static_id->operand[nop].constraint;
3071 enum constraint_num cn = lookup_constraint (constraint);
3072 bool change_p = false;
3074 if (MEM_P (op)
3075 && GET_MODE (op) == BLKmode
3076 && GET_CODE (XEXP (op, 0)) == SCRATCH)
3077 return false;
3079 if (insn_extra_address_constraint (cn))
3080 decompose_lea_address (&ad, curr_id->operand_loc[nop]);
3081 else if (MEM_P (op))
3082 decompose_mem_address (&ad, op);
3083 else if (GET_CODE (op) == SUBREG
3084 && MEM_P (SUBREG_REG (op)))
3085 decompose_mem_address (&ad, SUBREG_REG (op));
3086 else
3087 return false;
3088 /* If INDEX_REG_CLASS is assigned to base_term already and isn't to
3089 index_term, swap them so to avoid assigning INDEX_REG_CLASS to both
3090 when INDEX_REG_CLASS is a single register class. */
3091 if (ad.base_term != NULL
3092 && ad.index_term != NULL
3093 && ira_class_hard_regs_num[INDEX_REG_CLASS] == 1
3094 && REG_P (*ad.base_term)
3095 && REG_P (*ad.index_term)
3096 && in_class_p (*ad.base_term, INDEX_REG_CLASS, NULL)
3097 && ! in_class_p (*ad.index_term, INDEX_REG_CLASS, NULL))
3099 std::swap (ad.base, ad.index);
3100 std::swap (ad.base_term, ad.index_term);
3102 if (! check_only_p)
3103 change_p = equiv_address_substitution (&ad);
3104 if (ad.base_term != NULL
3105 && (process_addr_reg
3106 (ad.base_term, check_only_p, before,
3107 (ad.autoinc_p
3108 && !(REG_P (*ad.base_term)
3109 && find_regno_note (curr_insn, REG_DEAD,
3110 REGNO (*ad.base_term)) != NULL_RTX)
3111 ? after : NULL),
3112 base_reg_class (ad.mode, ad.as, ad.base_outer_code,
3113 get_index_code (&ad)))))
3115 change_p = true;
3116 if (ad.base_term2 != NULL)
3117 *ad.base_term2 = *ad.base_term;
3119 if (ad.index_term != NULL
3120 && process_addr_reg (ad.index_term, check_only_p,
3121 before, NULL, INDEX_REG_CLASS))
3122 change_p = true;
3124 /* Target hooks sometimes don't treat extra-constraint addresses as
3125 legitimate address_operands, so handle them specially. */
3126 if (insn_extra_address_constraint (cn)
3127 && satisfies_address_constraint_p (&ad, cn))
3128 return change_p;
3130 if (check_only_p)
3131 return change_p;
3133 /* There are three cases where the shape of *AD.INNER may now be invalid:
3135 1) the original address was valid, but either elimination or
3136 equiv_address_substitution was applied and that made
3137 the address invalid.
3139 2) the address is an invalid symbolic address created by
3140 force_const_to_mem.
3142 3) the address is a frame address with an invalid offset.
3144 4) the address is a frame address with an invalid base.
3146 All these cases involve a non-autoinc address, so there is no
3147 point revalidating other types. */
3148 if (ad.autoinc_p || valid_address_p (&ad))
3149 return change_p;
3151 /* Any index existed before LRA started, so we can assume that the
3152 presence and shape of the index is valid. */
3153 push_to_sequence (*before);
3154 lra_assert (ad.disp == ad.disp_term);
3155 if (ad.base == NULL)
3157 if (ad.index == NULL)
3159 rtx_insn *insn;
3160 rtx_insn *last = get_last_insn ();
3161 int code = -1;
3162 enum reg_class cl = base_reg_class (ad.mode, ad.as,
3163 SCRATCH, SCRATCH);
3164 rtx addr = *ad.inner;
3166 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "addr");
3167 if (HAVE_lo_sum)
3169 /* addr => lo_sum (new_base, addr), case (2) above. */
3170 insn = emit_insn (gen_rtx_SET
3171 (new_reg,
3172 gen_rtx_HIGH (Pmode, copy_rtx (addr))));
3173 code = recog_memoized (insn);
3174 if (code >= 0)
3176 *ad.inner = gen_rtx_LO_SUM (Pmode, new_reg, addr);
3177 if (! valid_address_p (ad.mode, *ad.outer, ad.as))
3179 /* Try to put lo_sum into register. */
3180 insn = emit_insn (gen_rtx_SET
3181 (new_reg,
3182 gen_rtx_LO_SUM (Pmode, new_reg, addr)));
3183 code = recog_memoized (insn);
3184 if (code >= 0)
3186 *ad.inner = new_reg;
3187 if (! valid_address_p (ad.mode, *ad.outer, ad.as))
3189 *ad.inner = addr;
3190 code = -1;
3196 if (code < 0)
3197 delete_insns_since (last);
3200 if (code < 0)
3202 /* addr => new_base, case (2) above. */
3203 lra_emit_move (new_reg, addr);
3205 for (insn = last == NULL_RTX ? get_insns () : NEXT_INSN (last);
3206 insn != NULL_RTX;
3207 insn = NEXT_INSN (insn))
3208 if (recog_memoized (insn) < 0)
3209 break;
3210 if (insn != NULL_RTX)
3212 /* Do nothing if we cannot generate right insns.
3213 This is analogous to reload pass behavior. */
3214 delete_insns_since (last);
3215 end_sequence ();
3216 return false;
3218 *ad.inner = new_reg;
3221 else
3223 /* index * scale + disp => new base + index * scale,
3224 case (1) above. */
3225 enum reg_class cl = base_reg_class (ad.mode, ad.as, PLUS,
3226 GET_CODE (*ad.index));
3228 lra_assert (INDEX_REG_CLASS != NO_REGS);
3229 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "disp");
3230 lra_emit_move (new_reg, *ad.disp);
3231 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3232 new_reg, *ad.index);
3235 else if (ad.index == NULL)
3237 int regno;
3238 enum reg_class cl;
3239 rtx set;
3240 rtx_insn *insns, *last_insn;
3241 /* Try to reload base into register only if the base is invalid
3242 for the address but with valid offset, case (4) above. */
3243 start_sequence ();
3244 new_reg = base_to_reg (&ad);
3246 /* base + disp => new base, cases (1) and (3) above. */
3247 /* Another option would be to reload the displacement into an
3248 index register. However, postreload has code to optimize
3249 address reloads that have the same base and different
3250 displacements, so reloading into an index register would
3251 not necessarily be a win. */
3252 if (new_reg == NULL_RTX)
3253 new_reg = base_plus_disp_to_reg (&ad);
3254 insns = get_insns ();
3255 last_insn = get_last_insn ();
3256 /* If we generated at least two insns, try last insn source as
3257 an address. If we succeed, we generate one less insn. */
3258 if (last_insn != insns && (set = single_set (last_insn)) != NULL_RTX
3259 && GET_CODE (SET_SRC (set)) == PLUS
3260 && REG_P (XEXP (SET_SRC (set), 0))
3261 && CONSTANT_P (XEXP (SET_SRC (set), 1)))
3263 *ad.inner = SET_SRC (set);
3264 if (valid_address_p (ad.mode, *ad.outer, ad.as))
3266 *ad.base_term = XEXP (SET_SRC (set), 0);
3267 *ad.disp_term = XEXP (SET_SRC (set), 1);
3268 cl = base_reg_class (ad.mode, ad.as, ad.base_outer_code,
3269 get_index_code (&ad));
3270 regno = REGNO (*ad.base_term);
3271 if (regno >= FIRST_PSEUDO_REGISTER
3272 && cl != lra_get_allocno_class (regno))
3273 lra_change_class (regno, cl, " Change to", true);
3274 new_reg = SET_SRC (set);
3275 delete_insns_since (PREV_INSN (last_insn));
3278 /* Try if target can split displacement into legitimite new disp
3279 and offset. If it's the case, we replace the last insn with
3280 insns for base + offset => new_reg and set new_reg + new disp
3281 to *ad.inner. */
3282 last_insn = get_last_insn ();
3283 if ((set = single_set (last_insn)) != NULL_RTX
3284 && GET_CODE (SET_SRC (set)) == PLUS
3285 && REG_P (XEXP (SET_SRC (set), 0))
3286 && REGNO (XEXP (SET_SRC (set), 0)) < FIRST_PSEUDO_REGISTER
3287 && CONST_INT_P (XEXP (SET_SRC (set), 1)))
3289 rtx addend, disp = XEXP (SET_SRC (set), 1);
3290 if (targetm.legitimize_address_displacement (&disp, &addend,
3291 ad.mode))
3293 rtx_insn *new_insns;
3294 start_sequence ();
3295 lra_emit_add (new_reg, XEXP (SET_SRC (set), 0), addend);
3296 new_insns = get_insns ();
3297 end_sequence ();
3298 new_reg = gen_rtx_PLUS (Pmode, new_reg, disp);
3299 delete_insns_since (PREV_INSN (last_insn));
3300 add_insn (new_insns);
3301 insns = get_insns ();
3304 end_sequence ();
3305 emit_insn (insns);
3306 *ad.inner = new_reg;
3308 else if (ad.disp_term != NULL)
3310 /* base + scale * index + disp => new base + scale * index,
3311 case (1) above. */
3312 new_reg = base_plus_disp_to_reg (&ad);
3313 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3314 new_reg, *ad.index);
3316 else if ((scale = get_index_scale (&ad)) == 1)
3318 /* The last transformation to one reg will be made in
3319 curr_insn_transform function. */
3320 end_sequence ();
3321 return false;
3323 else if (scale != 0)
3325 /* base + scale * index => base + new_reg,
3326 case (1) above.
3327 Index part of address may become invalid. For example, we
3328 changed pseudo on the equivalent memory and a subreg of the
3329 pseudo onto the memory of different mode for which the scale is
3330 prohibitted. */
3331 new_reg = index_part_to_reg (&ad);
3332 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3333 *ad.base_term, new_reg);
3335 else
3337 enum reg_class cl = base_reg_class (ad.mode, ad.as,
3338 SCRATCH, SCRATCH);
3339 rtx addr = *ad.inner;
3341 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "addr");
3342 /* addr => new_base. */
3343 lra_emit_move (new_reg, addr);
3344 *ad.inner = new_reg;
3346 *before = get_insns ();
3347 end_sequence ();
3348 return true;
3351 /* If CHECK_ONLY_P is false, do address reloads until it is necessary.
3352 Use process_address_1 as a helper function. Return true for any
3353 RTL changes.
3355 If CHECK_ONLY_P is true, just check address correctness. Return
3356 false if the address correct. */
3357 static bool
3358 process_address (int nop, bool check_only_p,
3359 rtx_insn **before, rtx_insn **after)
3361 bool res = false;
3363 while (process_address_1 (nop, check_only_p, before, after))
3365 if (check_only_p)
3366 return true;
3367 res = true;
3369 return res;
3372 /* Emit insns to reload VALUE into a new register. VALUE is an
3373 auto-increment or auto-decrement RTX whose operand is a register or
3374 memory location; so reloading involves incrementing that location.
3375 IN is either identical to VALUE, or some cheaper place to reload
3376 value being incremented/decremented from.
3378 INC_AMOUNT is the number to increment or decrement by (always
3379 positive and ignored for POST_MODIFY/PRE_MODIFY).
3381 Return pseudo containing the result. */
3382 static rtx
3383 emit_inc (enum reg_class new_rclass, rtx in, rtx value, int inc_amount)
3385 /* REG or MEM to be copied and incremented. */
3386 rtx incloc = XEXP (value, 0);
3387 /* Nonzero if increment after copying. */
3388 int post = (GET_CODE (value) == POST_DEC || GET_CODE (value) == POST_INC
3389 || GET_CODE (value) == POST_MODIFY);
3390 rtx_insn *last;
3391 rtx inc;
3392 rtx_insn *add_insn;
3393 int code;
3394 rtx real_in = in == value ? incloc : in;
3395 rtx result;
3396 bool plus_p = true;
3398 if (GET_CODE (value) == PRE_MODIFY || GET_CODE (value) == POST_MODIFY)
3400 lra_assert (GET_CODE (XEXP (value, 1)) == PLUS
3401 || GET_CODE (XEXP (value, 1)) == MINUS);
3402 lra_assert (rtx_equal_p (XEXP (XEXP (value, 1), 0), XEXP (value, 0)));
3403 plus_p = GET_CODE (XEXP (value, 1)) == PLUS;
3404 inc = XEXP (XEXP (value, 1), 1);
3406 else
3408 if (GET_CODE (value) == PRE_DEC || GET_CODE (value) == POST_DEC)
3409 inc_amount = -inc_amount;
3411 inc = GEN_INT (inc_amount);
3414 if (! post && REG_P (incloc))
3415 result = incloc;
3416 else
3417 result = lra_create_new_reg (GET_MODE (value), value, new_rclass,
3418 "INC/DEC result");
3420 if (real_in != result)
3422 /* First copy the location to the result register. */
3423 lra_assert (REG_P (result));
3424 emit_insn (gen_move_insn (result, real_in));
3427 /* We suppose that there are insns to add/sub with the constant
3428 increment permitted in {PRE/POST)_{DEC/INC/MODIFY}. At least the
3429 old reload worked with this assumption. If the assumption
3430 becomes wrong, we should use approach in function
3431 base_plus_disp_to_reg. */
3432 if (in == value)
3434 /* See if we can directly increment INCLOC. */
3435 last = get_last_insn ();
3436 add_insn = emit_insn (plus_p
3437 ? gen_add2_insn (incloc, inc)
3438 : gen_sub2_insn (incloc, inc));
3440 code = recog_memoized (add_insn);
3441 if (code >= 0)
3443 if (! post && result != incloc)
3444 emit_insn (gen_move_insn (result, incloc));
3445 return result;
3447 delete_insns_since (last);
3450 /* If couldn't do the increment directly, must increment in RESULT.
3451 The way we do this depends on whether this is pre- or
3452 post-increment. For pre-increment, copy INCLOC to the reload
3453 register, increment it there, then save back. */
3454 if (! post)
3456 if (real_in != result)
3457 emit_insn (gen_move_insn (result, real_in));
3458 if (plus_p)
3459 emit_insn (gen_add2_insn (result, inc));
3460 else
3461 emit_insn (gen_sub2_insn (result, inc));
3462 if (result != incloc)
3463 emit_insn (gen_move_insn (incloc, result));
3465 else
3467 /* Post-increment.
3469 Because this might be a jump insn or a compare, and because
3470 RESULT may not be available after the insn in an input
3471 reload, we must do the incrementing before the insn being
3472 reloaded for.
3474 We have already copied IN to RESULT. Increment the copy in
3475 RESULT, save that back, then decrement RESULT so it has
3476 the original value. */
3477 if (plus_p)
3478 emit_insn (gen_add2_insn (result, inc));
3479 else
3480 emit_insn (gen_sub2_insn (result, inc));
3481 emit_insn (gen_move_insn (incloc, result));
3482 /* Restore non-modified value for the result. We prefer this
3483 way because it does not require an additional hard
3484 register. */
3485 if (plus_p)
3487 if (CONST_INT_P (inc))
3488 emit_insn (gen_add2_insn (result,
3489 gen_int_mode (-INTVAL (inc),
3490 GET_MODE (result))));
3491 else
3492 emit_insn (gen_sub2_insn (result, inc));
3494 else
3495 emit_insn (gen_add2_insn (result, inc));
3497 return result;
3500 /* Return true if the current move insn does not need processing as we
3501 already know that it satisfies its constraints. */
3502 static bool
3503 simple_move_p (void)
3505 rtx dest, src;
3506 enum reg_class dclass, sclass;
3508 lra_assert (curr_insn_set != NULL_RTX);
3509 dest = SET_DEST (curr_insn_set);
3510 src = SET_SRC (curr_insn_set);
3512 /* If the instruction has multiple sets we need to process it even if it
3513 is single_set. This can happen if one or more of the SETs are dead.
3514 See PR73650. */
3515 if (multiple_sets (curr_insn))
3516 return false;
3518 return ((dclass = get_op_class (dest)) != NO_REGS
3519 && (sclass = get_op_class (src)) != NO_REGS
3520 /* The backend guarantees that register moves of cost 2
3521 never need reloads. */
3522 && targetm.register_move_cost (GET_MODE (src), sclass, dclass) == 2);
3525 /* Swap operands NOP and NOP + 1. */
3526 static inline void
3527 swap_operands (int nop)
3529 std::swap (curr_operand_mode[nop], curr_operand_mode[nop + 1]);
3530 std::swap (original_subreg_reg_mode[nop], original_subreg_reg_mode[nop + 1]);
3531 std::swap (*curr_id->operand_loc[nop], *curr_id->operand_loc[nop + 1]);
3532 std::swap (equiv_substition_p[nop], equiv_substition_p[nop + 1]);
3533 /* Swap the duplicates too. */
3534 lra_update_dup (curr_id, nop);
3535 lra_update_dup (curr_id, nop + 1);
3538 /* Main entry point of the constraint code: search the body of the
3539 current insn to choose the best alternative. It is mimicking insn
3540 alternative cost calculation model of former reload pass. That is
3541 because machine descriptions were written to use this model. This
3542 model can be changed in future. Make commutative operand exchange
3543 if it is chosen.
3545 if CHECK_ONLY_P is false, do RTL changes to satisfy the
3546 constraints. Return true if any change happened during function
3547 call.
3549 If CHECK_ONLY_P is true then don't do any transformation. Just
3550 check that the insn satisfies all constraints. If the insn does
3551 not satisfy any constraint, return true. */
3552 static bool
3553 curr_insn_transform (bool check_only_p)
3555 int i, j, k;
3556 int n_operands;
3557 int n_alternatives;
3558 int n_outputs;
3559 int commutative;
3560 signed char goal_alt_matched[MAX_RECOG_OPERANDS][MAX_RECOG_OPERANDS];
3561 signed char match_inputs[MAX_RECOG_OPERANDS + 1];
3562 signed char outputs[MAX_RECOG_OPERANDS + 1];
3563 rtx_insn *before, *after;
3564 bool alt_p = false;
3565 /* Flag that the insn has been changed through a transformation. */
3566 bool change_p;
3567 bool sec_mem_p;
3568 #ifdef SECONDARY_MEMORY_NEEDED
3569 bool use_sec_mem_p;
3570 #endif
3571 int max_regno_before;
3572 int reused_alternative_num;
3574 curr_insn_set = single_set (curr_insn);
3575 if (curr_insn_set != NULL_RTX && simple_move_p ())
3576 return false;
3578 no_input_reloads_p = no_output_reloads_p = false;
3579 goal_alt_number = -1;
3580 change_p = sec_mem_p = false;
3581 /* JUMP_INSNs and CALL_INSNs are not allowed to have any output
3582 reloads; neither are insns that SET cc0. Insns that use CC0 are
3583 not allowed to have any input reloads. */
3584 if (JUMP_P (curr_insn) || CALL_P (curr_insn))
3585 no_output_reloads_p = true;
3587 if (HAVE_cc0 && reg_referenced_p (cc0_rtx, PATTERN (curr_insn)))
3588 no_input_reloads_p = true;
3589 if (HAVE_cc0 && reg_set_p (cc0_rtx, PATTERN (curr_insn)))
3590 no_output_reloads_p = true;
3592 n_operands = curr_static_id->n_operands;
3593 n_alternatives = curr_static_id->n_alternatives;
3595 /* Just return "no reloads" if insn has no operands with
3596 constraints. */
3597 if (n_operands == 0 || n_alternatives == 0)
3598 return false;
3600 max_regno_before = max_reg_num ();
3602 for (i = 0; i < n_operands; i++)
3604 goal_alt_matched[i][0] = -1;
3605 goal_alt_matches[i] = -1;
3608 commutative = curr_static_id->commutative;
3610 /* Now see what we need for pseudos that didn't get hard regs or got
3611 the wrong kind of hard reg. For this, we must consider all the
3612 operands together against the register constraints. */
3614 best_losers = best_overall = INT_MAX;
3615 best_reload_sum = 0;
3617 curr_swapped = false;
3618 goal_alt_swapped = false;
3620 if (! check_only_p)
3621 /* Make equivalence substitution and memory subreg elimination
3622 before address processing because an address legitimacy can
3623 depend on memory mode. */
3624 for (i = 0; i < n_operands; i++)
3626 rtx op, subst, old;
3627 bool op_change_p = false;
3629 if (curr_static_id->operand[i].is_operator)
3630 continue;
3632 old = op = *curr_id->operand_loc[i];
3633 if (GET_CODE (old) == SUBREG)
3634 old = SUBREG_REG (old);
3635 subst = get_equiv_with_elimination (old, curr_insn);
3636 original_subreg_reg_mode[i] = VOIDmode;
3637 equiv_substition_p[i] = false;
3638 if (subst != old)
3640 equiv_substition_p[i] = true;
3641 subst = copy_rtx (subst);
3642 lra_assert (REG_P (old));
3643 if (GET_CODE (op) != SUBREG)
3644 *curr_id->operand_loc[i] = subst;
3645 else
3647 SUBREG_REG (op) = subst;
3648 if (GET_MODE (subst) == VOIDmode)
3649 original_subreg_reg_mode[i] = GET_MODE (old);
3651 if (lra_dump_file != NULL)
3653 fprintf (lra_dump_file,
3654 "Changing pseudo %d in operand %i of insn %u on equiv ",
3655 REGNO (old), i, INSN_UID (curr_insn));
3656 dump_value_slim (lra_dump_file, subst, 1);
3657 fprintf (lra_dump_file, "\n");
3659 op_change_p = change_p = true;
3661 if (simplify_operand_subreg (i, GET_MODE (old)) || op_change_p)
3663 change_p = true;
3664 lra_update_dup (curr_id, i);
3668 /* Reload address registers and displacements. We do it before
3669 finding an alternative because of memory constraints. */
3670 before = after = NULL;
3671 for (i = 0; i < n_operands; i++)
3672 if (! curr_static_id->operand[i].is_operator
3673 && process_address (i, check_only_p, &before, &after))
3675 if (check_only_p)
3676 return true;
3677 change_p = true;
3678 lra_update_dup (curr_id, i);
3681 if (change_p)
3682 /* If we've changed the instruction then any alternative that
3683 we chose previously may no longer be valid. */
3684 lra_set_used_insn_alternative (curr_insn, -1);
3686 if (! check_only_p && curr_insn_set != NULL_RTX
3687 && check_and_process_move (&change_p, &sec_mem_p))
3688 return change_p;
3690 try_swapped:
3692 reused_alternative_num = check_only_p ? -1 : curr_id->used_insn_alternative;
3693 if (lra_dump_file != NULL && reused_alternative_num >= 0)
3694 fprintf (lra_dump_file, "Reusing alternative %d for insn #%u\n",
3695 reused_alternative_num, INSN_UID (curr_insn));
3697 if (process_alt_operands (reused_alternative_num))
3698 alt_p = true;
3700 if (check_only_p)
3701 return ! alt_p || best_losers != 0;
3703 /* If insn is commutative (it's safe to exchange a certain pair of
3704 operands) then we need to try each alternative twice, the second
3705 time matching those two operands as if we had exchanged them. To
3706 do this, really exchange them in operands.
3708 If we have just tried the alternatives the second time, return
3709 operands to normal and drop through. */
3711 if (reused_alternative_num < 0 && commutative >= 0)
3713 curr_swapped = !curr_swapped;
3714 if (curr_swapped)
3716 swap_operands (commutative);
3717 goto try_swapped;
3719 else
3720 swap_operands (commutative);
3723 if (! alt_p && ! sec_mem_p)
3725 /* No alternative works with reloads?? */
3726 if (INSN_CODE (curr_insn) >= 0)
3727 fatal_insn ("unable to generate reloads for:", curr_insn);
3728 error_for_asm (curr_insn,
3729 "inconsistent operand constraints in an %<asm%>");
3730 /* Avoid further trouble with this insn. */
3731 PATTERN (curr_insn) = gen_rtx_USE (VOIDmode, const0_rtx);
3732 lra_invalidate_insn_data (curr_insn);
3733 return true;
3736 /* If the best alternative is with operands 1 and 2 swapped, swap
3737 them. Update the operand numbers of any reloads already
3738 pushed. */
3740 if (goal_alt_swapped)
3742 if (lra_dump_file != NULL)
3743 fprintf (lra_dump_file, " Commutative operand exchange in insn %u\n",
3744 INSN_UID (curr_insn));
3746 /* Swap the duplicates too. */
3747 swap_operands (commutative);
3748 change_p = true;
3751 #ifdef SECONDARY_MEMORY_NEEDED
3752 /* Some target macros SECONDARY_MEMORY_NEEDED (e.g. x86) are defined
3753 too conservatively. So we use the secondary memory only if there
3754 is no any alternative without reloads. */
3755 use_sec_mem_p = false;
3756 if (! alt_p)
3757 use_sec_mem_p = true;
3758 else if (sec_mem_p)
3760 for (i = 0; i < n_operands; i++)
3761 if (! goal_alt_win[i] && ! goal_alt_match_win[i])
3762 break;
3763 use_sec_mem_p = i < n_operands;
3766 if (use_sec_mem_p)
3768 int in = -1, out = -1;
3769 rtx new_reg, src, dest, rld;
3770 machine_mode sec_mode, rld_mode;
3772 lra_assert (curr_insn_set != NULL_RTX && sec_mem_p);
3773 dest = SET_DEST (curr_insn_set);
3774 src = SET_SRC (curr_insn_set);
3775 for (i = 0; i < n_operands; i++)
3776 if (*curr_id->operand_loc[i] == dest)
3777 out = i;
3778 else if (*curr_id->operand_loc[i] == src)
3779 in = i;
3780 for (i = 0; i < curr_static_id->n_dups; i++)
3781 if (out < 0 && *curr_id->dup_loc[i] == dest)
3782 out = curr_static_id->dup_num[i];
3783 else if (in < 0 && *curr_id->dup_loc[i] == src)
3784 in = curr_static_id->dup_num[i];
3785 lra_assert (out >= 0 && in >= 0
3786 && curr_static_id->operand[out].type == OP_OUT
3787 && curr_static_id->operand[in].type == OP_IN);
3788 rld = (GET_MODE_SIZE (GET_MODE (dest)) <= GET_MODE_SIZE (GET_MODE (src))
3789 ? dest : src);
3790 rld_mode = GET_MODE (rld);
3791 #ifdef SECONDARY_MEMORY_NEEDED_MODE
3792 sec_mode = SECONDARY_MEMORY_NEEDED_MODE (rld_mode);
3793 #else
3794 sec_mode = rld_mode;
3795 #endif
3796 new_reg = lra_create_new_reg (sec_mode, NULL_RTX,
3797 NO_REGS, "secondary");
3798 /* If the mode is changed, it should be wider. */
3799 lra_assert (GET_MODE_SIZE (sec_mode) >= GET_MODE_SIZE (rld_mode));
3800 if (sec_mode != rld_mode)
3802 /* If the target says specifically to use another mode for
3803 secondary memory moves we can not reuse the original
3804 insn. */
3805 after = emit_spill_move (false, new_reg, dest);
3806 lra_process_new_insns (curr_insn, NULL, after,
3807 "Inserting the sec. move");
3808 /* We may have non null BEFORE here (e.g. after address
3809 processing. */
3810 push_to_sequence (before);
3811 before = emit_spill_move (true, new_reg, src);
3812 emit_insn (before);
3813 before = get_insns ();
3814 end_sequence ();
3815 lra_process_new_insns (curr_insn, before, NULL, "Changing on");
3816 lra_set_insn_deleted (curr_insn);
3818 else if (dest == rld)
3820 *curr_id->operand_loc[out] = new_reg;
3821 lra_update_dup (curr_id, out);
3822 after = emit_spill_move (false, new_reg, dest);
3823 lra_process_new_insns (curr_insn, NULL, after,
3824 "Inserting the sec. move");
3826 else
3828 *curr_id->operand_loc[in] = new_reg;
3829 lra_update_dup (curr_id, in);
3830 /* See comments above. */
3831 push_to_sequence (before);
3832 before = emit_spill_move (true, new_reg, src);
3833 emit_insn (before);
3834 before = get_insns ();
3835 end_sequence ();
3836 lra_process_new_insns (curr_insn, before, NULL,
3837 "Inserting the sec. move");
3839 lra_update_insn_regno_info (curr_insn);
3840 return true;
3842 #endif
3844 lra_assert (goal_alt_number >= 0);
3845 lra_set_used_insn_alternative (curr_insn, goal_alt_number);
3847 if (lra_dump_file != NULL)
3849 const char *p;
3851 fprintf (lra_dump_file, " Choosing alt %d in insn %u:",
3852 goal_alt_number, INSN_UID (curr_insn));
3853 for (i = 0; i < n_operands; i++)
3855 p = (curr_static_id->operand_alternative
3856 [goal_alt_number * n_operands + i].constraint);
3857 if (*p == '\0')
3858 continue;
3859 fprintf (lra_dump_file, " (%d) ", i);
3860 for (; *p != '\0' && *p != ',' && *p != '#'; p++)
3861 fputc (*p, lra_dump_file);
3863 if (INSN_CODE (curr_insn) >= 0
3864 && (p = get_insn_name (INSN_CODE (curr_insn))) != NULL)
3865 fprintf (lra_dump_file, " {%s}", p);
3866 if (curr_id->sp_offset != 0)
3867 fprintf (lra_dump_file, " (sp_off=%" HOST_WIDE_INT_PRINT "d)",
3868 curr_id->sp_offset);
3869 fprintf (lra_dump_file, "\n");
3872 /* Right now, for any pair of operands I and J that are required to
3873 match, with J < I, goal_alt_matches[I] is J. Add I to
3874 goal_alt_matched[J]. */
3876 for (i = 0; i < n_operands; i++)
3877 if ((j = goal_alt_matches[i]) >= 0)
3879 for (k = 0; goal_alt_matched[j][k] >= 0; k++)
3881 /* We allow matching one output operand and several input
3882 operands. */
3883 lra_assert (k == 0
3884 || (curr_static_id->operand[j].type == OP_OUT
3885 && curr_static_id->operand[i].type == OP_IN
3886 && (curr_static_id->operand
3887 [goal_alt_matched[j][0]].type == OP_IN)));
3888 goal_alt_matched[j][k] = i;
3889 goal_alt_matched[j][k + 1] = -1;
3892 for (i = 0; i < n_operands; i++)
3893 goal_alt_win[i] |= goal_alt_match_win[i];
3895 /* Any constants that aren't allowed and can't be reloaded into
3896 registers are here changed into memory references. */
3897 for (i = 0; i < n_operands; i++)
3898 if (goal_alt_win[i])
3900 int regno;
3901 enum reg_class new_class;
3902 rtx reg = *curr_id->operand_loc[i];
3904 if (GET_CODE (reg) == SUBREG)
3905 reg = SUBREG_REG (reg);
3907 if (REG_P (reg) && (regno = REGNO (reg)) >= FIRST_PSEUDO_REGISTER)
3909 bool ok_p = in_class_p (reg, goal_alt[i], &new_class);
3911 if (new_class != NO_REGS && get_reg_class (regno) != new_class)
3913 lra_assert (ok_p);
3914 lra_change_class (regno, new_class, " Change to", true);
3918 else
3920 const char *constraint;
3921 char c;
3922 rtx op = *curr_id->operand_loc[i];
3923 rtx subreg = NULL_RTX;
3924 machine_mode mode = curr_operand_mode[i];
3926 if (GET_CODE (op) == SUBREG)
3928 subreg = op;
3929 op = SUBREG_REG (op);
3930 mode = GET_MODE (op);
3933 if (CONST_POOL_OK_P (mode, op)
3934 && ((targetm.preferred_reload_class
3935 (op, (enum reg_class) goal_alt[i]) == NO_REGS)
3936 || no_input_reloads_p))
3938 rtx tem = force_const_mem (mode, op);
3940 change_p = true;
3941 if (subreg != NULL_RTX)
3942 tem = gen_rtx_SUBREG (mode, tem, SUBREG_BYTE (subreg));
3944 *curr_id->operand_loc[i] = tem;
3945 lra_update_dup (curr_id, i);
3946 process_address (i, false, &before, &after);
3948 /* If the alternative accepts constant pool refs directly
3949 there will be no reload needed at all. */
3950 if (subreg != NULL_RTX)
3951 continue;
3952 /* Skip alternatives before the one requested. */
3953 constraint = (curr_static_id->operand_alternative
3954 [goal_alt_number * n_operands + i].constraint);
3955 for (;
3956 (c = *constraint) && c != ',' && c != '#';
3957 constraint += CONSTRAINT_LEN (c, constraint))
3959 enum constraint_num cn = lookup_constraint (constraint);
3960 if ((insn_extra_memory_constraint (cn)
3961 || insn_extra_special_memory_constraint (cn))
3962 && satisfies_memory_constraint_p (tem, cn))
3963 break;
3965 if (c == '\0' || c == ',' || c == '#')
3966 continue;
3968 goal_alt_win[i] = true;
3972 n_outputs = 0;
3973 outputs[0] = -1;
3974 for (i = 0; i < n_operands; i++)
3976 int regno;
3977 bool optional_p = false;
3978 rtx old, new_reg;
3979 rtx op = *curr_id->operand_loc[i];
3981 if (goal_alt_win[i])
3983 if (goal_alt[i] == NO_REGS
3984 && REG_P (op)
3985 /* When we assign NO_REGS it means that we will not
3986 assign a hard register to the scratch pseudo by
3987 assigment pass and the scratch pseudo will be
3988 spilled. Spilled scratch pseudos are transformed
3989 back to scratches at the LRA end. */
3990 && lra_former_scratch_operand_p (curr_insn, i)
3991 && lra_former_scratch_p (REGNO (op)))
3993 int regno = REGNO (op);
3994 lra_change_class (regno, NO_REGS, " Change to", true);
3995 if (lra_get_regno_hard_regno (regno) >= 0)
3996 /* We don't have to mark all insn affected by the
3997 spilled pseudo as there is only one such insn, the
3998 current one. */
3999 reg_renumber[regno] = -1;
4000 lra_assert (bitmap_single_bit_set_p
4001 (&lra_reg_info[REGNO (op)].insn_bitmap));
4003 /* We can do an optional reload. If the pseudo got a hard
4004 reg, we might improve the code through inheritance. If
4005 it does not get a hard register we coalesce memory/memory
4006 moves later. Ignore move insns to avoid cycling. */
4007 if (! lra_simple_p
4008 && lra_undo_inheritance_iter < LRA_MAX_INHERITANCE_PASSES
4009 && goal_alt[i] != NO_REGS && REG_P (op)
4010 && (regno = REGNO (op)) >= FIRST_PSEUDO_REGISTER
4011 && regno < new_regno_start
4012 && ! lra_former_scratch_p (regno)
4013 && reg_renumber[regno] < 0
4014 /* Check that the optional reload pseudo will be able to
4015 hold given mode value. */
4016 && ! (prohibited_class_reg_set_mode_p
4017 (goal_alt[i], reg_class_contents[goal_alt[i]],
4018 PSEUDO_REGNO_MODE (regno)))
4019 && (curr_insn_set == NULL_RTX
4020 || !((REG_P (SET_SRC (curr_insn_set))
4021 || MEM_P (SET_SRC (curr_insn_set))
4022 || GET_CODE (SET_SRC (curr_insn_set)) == SUBREG)
4023 && (REG_P (SET_DEST (curr_insn_set))
4024 || MEM_P (SET_DEST (curr_insn_set))
4025 || GET_CODE (SET_DEST (curr_insn_set)) == SUBREG))))
4026 optional_p = true;
4027 else
4028 continue;
4031 /* Operands that match previous ones have already been handled. */
4032 if (goal_alt_matches[i] >= 0)
4033 continue;
4035 /* We should not have an operand with a non-offsettable address
4036 appearing where an offsettable address will do. It also may
4037 be a case when the address should be special in other words
4038 not a general one (e.g. it needs no index reg). */
4039 if (goal_alt_matched[i][0] == -1 && goal_alt_offmemok[i] && MEM_P (op))
4041 enum reg_class rclass;
4042 rtx *loc = &XEXP (op, 0);
4043 enum rtx_code code = GET_CODE (*loc);
4045 push_to_sequence (before);
4046 rclass = base_reg_class (GET_MODE (op), MEM_ADDR_SPACE (op),
4047 MEM, SCRATCH);
4048 if (GET_RTX_CLASS (code) == RTX_AUTOINC)
4049 new_reg = emit_inc (rclass, *loc, *loc,
4050 /* This value does not matter for MODIFY. */
4051 GET_MODE_SIZE (GET_MODE (op)));
4052 else if (get_reload_reg (OP_IN, Pmode, *loc, rclass, FALSE,
4053 "offsetable address", &new_reg))
4054 lra_emit_move (new_reg, *loc);
4055 before = get_insns ();
4056 end_sequence ();
4057 *loc = new_reg;
4058 lra_update_dup (curr_id, i);
4060 else if (goal_alt_matched[i][0] == -1)
4062 machine_mode mode;
4063 rtx reg, *loc;
4064 int hard_regno, byte;
4065 enum op_type type = curr_static_id->operand[i].type;
4067 loc = curr_id->operand_loc[i];
4068 mode = curr_operand_mode[i];
4069 if (GET_CODE (*loc) == SUBREG)
4071 reg = SUBREG_REG (*loc);
4072 byte = SUBREG_BYTE (*loc);
4073 if (REG_P (reg)
4074 /* Strict_low_part requires reload the register not
4075 the sub-register. */
4076 && (curr_static_id->operand[i].strict_low
4077 || (GET_MODE_SIZE (mode)
4078 <= GET_MODE_SIZE (GET_MODE (reg))
4079 && (hard_regno
4080 = get_try_hard_regno (REGNO (reg))) >= 0
4081 && (simplify_subreg_regno
4082 (hard_regno,
4083 GET_MODE (reg), byte, mode) < 0)
4084 && (goal_alt[i] == NO_REGS
4085 || (simplify_subreg_regno
4086 (ira_class_hard_regs[goal_alt[i]][0],
4087 GET_MODE (reg), byte, mode) >= 0)))))
4089 if (type == OP_OUT)
4090 type = OP_INOUT;
4091 loc = &SUBREG_REG (*loc);
4092 mode = GET_MODE (*loc);
4095 old = *loc;
4096 if (get_reload_reg (type, mode, old, goal_alt[i],
4097 loc != curr_id->operand_loc[i], "", &new_reg)
4098 && type != OP_OUT)
4100 push_to_sequence (before);
4101 lra_emit_move (new_reg, old);
4102 before = get_insns ();
4103 end_sequence ();
4105 *loc = new_reg;
4106 if (type != OP_IN
4107 && find_reg_note (curr_insn, REG_UNUSED, old) == NULL_RTX)
4109 start_sequence ();
4110 lra_emit_move (type == OP_INOUT ? copy_rtx (old) : old, new_reg);
4111 emit_insn (after);
4112 after = get_insns ();
4113 end_sequence ();
4114 *loc = new_reg;
4116 for (j = 0; j < goal_alt_dont_inherit_ops_num; j++)
4117 if (goal_alt_dont_inherit_ops[j] == i)
4119 lra_set_regno_unique_value (REGNO (new_reg));
4120 break;
4122 lra_update_dup (curr_id, i);
4124 else if (curr_static_id->operand[i].type == OP_IN
4125 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4126 == OP_OUT))
4128 /* generate reloads for input and matched outputs. */
4129 match_inputs[0] = i;
4130 match_inputs[1] = -1;
4131 match_reload (goal_alt_matched[i][0], match_inputs, outputs,
4132 goal_alt[i], &before, &after,
4133 curr_static_id->operand_alternative
4134 [goal_alt_number * n_operands + goal_alt_matched[i][0]]
4135 .earlyclobber);
4137 else if (curr_static_id->operand[i].type == OP_OUT
4138 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4139 == OP_IN))
4140 /* Generate reloads for output and matched inputs. */
4141 match_reload (i, goal_alt_matched[i], outputs, goal_alt[i], &before,
4142 &after, curr_static_id->operand_alternative
4143 [goal_alt_number * n_operands + i].earlyclobber);
4144 else if (curr_static_id->operand[i].type == OP_IN
4145 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4146 == OP_IN))
4148 /* Generate reloads for matched inputs. */
4149 match_inputs[0] = i;
4150 for (j = 0; (k = goal_alt_matched[i][j]) >= 0; j++)
4151 match_inputs[j + 1] = k;
4152 match_inputs[j + 1] = -1;
4153 match_reload (-1, match_inputs, outputs, goal_alt[i], &before,
4154 &after, false);
4156 else
4157 /* We must generate code in any case when function
4158 process_alt_operands decides that it is possible. */
4159 gcc_unreachable ();
4161 /* Memorise processed outputs so that output remaining to be processed
4162 can avoid using the same register value (see match_reload). */
4163 if (curr_static_id->operand[i].type == OP_OUT)
4165 outputs[n_outputs++] = i;
4166 outputs[n_outputs] = -1;
4169 if (optional_p)
4171 rtx reg = op;
4173 lra_assert (REG_P (reg));
4174 regno = REGNO (reg);
4175 op = *curr_id->operand_loc[i]; /* Substitution. */
4176 if (GET_CODE (op) == SUBREG)
4177 op = SUBREG_REG (op);
4178 gcc_assert (REG_P (op) && (int) REGNO (op) >= new_regno_start);
4179 bitmap_set_bit (&lra_optional_reload_pseudos, REGNO (op));
4180 lra_reg_info[REGNO (op)].restore_rtx = reg;
4181 if (lra_dump_file != NULL)
4182 fprintf (lra_dump_file,
4183 " Making reload reg %d for reg %d optional\n",
4184 REGNO (op), regno);
4187 if (before != NULL_RTX || after != NULL_RTX
4188 || max_regno_before != max_reg_num ())
4189 change_p = true;
4190 if (change_p)
4192 lra_update_operator_dups (curr_id);
4193 /* Something changes -- process the insn. */
4194 lra_update_insn_regno_info (curr_insn);
4196 lra_process_new_insns (curr_insn, before, after, "Inserting insn reload");
4197 return change_p;
4200 /* Return true if INSN satisfies all constraints. In other words, no
4201 reload insns are needed. */
4202 bool
4203 lra_constrain_insn (rtx_insn *insn)
4205 int saved_new_regno_start = new_regno_start;
4206 int saved_new_insn_uid_start = new_insn_uid_start;
4207 bool change_p;
4209 curr_insn = insn;
4210 curr_id = lra_get_insn_recog_data (curr_insn);
4211 curr_static_id = curr_id->insn_static_data;
4212 new_insn_uid_start = get_max_uid ();
4213 new_regno_start = max_reg_num ();
4214 change_p = curr_insn_transform (true);
4215 new_regno_start = saved_new_regno_start;
4216 new_insn_uid_start = saved_new_insn_uid_start;
4217 return ! change_p;
4220 /* Return true if X is in LIST. */
4221 static bool
4222 in_list_p (rtx x, rtx list)
4224 for (; list != NULL_RTX; list = XEXP (list, 1))
4225 if (XEXP (list, 0) == x)
4226 return true;
4227 return false;
4230 /* Return true if X contains an allocatable hard register (if
4231 HARD_REG_P) or a (spilled if SPILLED_P) pseudo. */
4232 static bool
4233 contains_reg_p (rtx x, bool hard_reg_p, bool spilled_p)
4235 int i, j;
4236 const char *fmt;
4237 enum rtx_code code;
4239 code = GET_CODE (x);
4240 if (REG_P (x))
4242 int regno = REGNO (x);
4243 HARD_REG_SET alloc_regs;
4245 if (hard_reg_p)
4247 if (regno >= FIRST_PSEUDO_REGISTER)
4248 regno = lra_get_regno_hard_regno (regno);
4249 if (regno < 0)
4250 return false;
4251 COMPL_HARD_REG_SET (alloc_regs, lra_no_alloc_regs);
4252 return overlaps_hard_reg_set_p (alloc_regs, GET_MODE (x), regno);
4254 else
4256 if (regno < FIRST_PSEUDO_REGISTER)
4257 return false;
4258 if (! spilled_p)
4259 return true;
4260 return lra_get_regno_hard_regno (regno) < 0;
4263 fmt = GET_RTX_FORMAT (code);
4264 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4266 if (fmt[i] == 'e')
4268 if (contains_reg_p (XEXP (x, i), hard_reg_p, spilled_p))
4269 return true;
4271 else if (fmt[i] == 'E')
4273 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4274 if (contains_reg_p (XVECEXP (x, i, j), hard_reg_p, spilled_p))
4275 return true;
4278 return false;
4281 /* Process all regs in location *LOC and change them on equivalent
4282 substitution. Return true if any change was done. */
4283 static bool
4284 loc_equivalence_change_p (rtx *loc)
4286 rtx subst, reg, x = *loc;
4287 bool result = false;
4288 enum rtx_code code = GET_CODE (x);
4289 const char *fmt;
4290 int i, j;
4292 if (code == SUBREG)
4294 reg = SUBREG_REG (x);
4295 if ((subst = get_equiv_with_elimination (reg, curr_insn)) != reg
4296 && GET_MODE (subst) == VOIDmode)
4298 /* We cannot reload debug location. Simplify subreg here
4299 while we know the inner mode. */
4300 *loc = simplify_gen_subreg (GET_MODE (x), subst,
4301 GET_MODE (reg), SUBREG_BYTE (x));
4302 return true;
4305 if (code == REG && (subst = get_equiv_with_elimination (x, curr_insn)) != x)
4307 *loc = subst;
4308 return true;
4311 /* Scan all the operand sub-expressions. */
4312 fmt = GET_RTX_FORMAT (code);
4313 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4315 if (fmt[i] == 'e')
4316 result = loc_equivalence_change_p (&XEXP (x, i)) || result;
4317 else if (fmt[i] == 'E')
4318 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4319 result
4320 = loc_equivalence_change_p (&XVECEXP (x, i, j)) || result;
4322 return result;
4325 /* Similar to loc_equivalence_change_p, but for use as
4326 simplify_replace_fn_rtx callback. DATA is insn for which the
4327 elimination is done. If it null we don't do the elimination. */
4328 static rtx
4329 loc_equivalence_callback (rtx loc, const_rtx, void *data)
4331 if (!REG_P (loc))
4332 return NULL_RTX;
4334 rtx subst = (data == NULL
4335 ? get_equiv (loc) : get_equiv_with_elimination (loc, (rtx_insn *) data));
4336 if (subst != loc)
4337 return subst;
4339 return NULL_RTX;
4342 /* Maximum number of generated reload insns per an insn. It is for
4343 preventing this pass cycling in a bug case. */
4344 #define MAX_RELOAD_INSNS_NUMBER LRA_MAX_INSN_RELOADS
4346 /* The current iteration number of this LRA pass. */
4347 int lra_constraint_iter;
4349 /* True if we substituted equiv which needs checking register
4350 allocation correctness because the equivalent value contains
4351 allocatable hard registers or when we restore multi-register
4352 pseudo. */
4353 bool lra_risky_transformations_p;
4355 /* Return true if REGNO is referenced in more than one block. */
4356 static bool
4357 multi_block_pseudo_p (int regno)
4359 basic_block bb = NULL;
4360 unsigned int uid;
4361 bitmap_iterator bi;
4363 if (regno < FIRST_PSEUDO_REGISTER)
4364 return false;
4366 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi)
4367 if (bb == NULL)
4368 bb = BLOCK_FOR_INSN (lra_insn_recog_data[uid]->insn);
4369 else if (BLOCK_FOR_INSN (lra_insn_recog_data[uid]->insn) != bb)
4370 return true;
4371 return false;
4374 /* Return true if LIST contains a deleted insn. */
4375 static bool
4376 contains_deleted_insn_p (rtx_insn_list *list)
4378 for (; list != NULL_RTX; list = list->next ())
4379 if (NOTE_P (list->insn ())
4380 && NOTE_KIND (list->insn ()) == NOTE_INSN_DELETED)
4381 return true;
4382 return false;
4385 /* Return true if X contains a pseudo dying in INSN. */
4386 static bool
4387 dead_pseudo_p (rtx x, rtx_insn *insn)
4389 int i, j;
4390 const char *fmt;
4391 enum rtx_code code;
4393 if (REG_P (x))
4394 return (insn != NULL_RTX
4395 && find_regno_note (insn, REG_DEAD, REGNO (x)) != NULL_RTX);
4396 code = GET_CODE (x);
4397 fmt = GET_RTX_FORMAT (code);
4398 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4400 if (fmt[i] == 'e')
4402 if (dead_pseudo_p (XEXP (x, i), insn))
4403 return true;
4405 else if (fmt[i] == 'E')
4407 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4408 if (dead_pseudo_p (XVECEXP (x, i, j), insn))
4409 return true;
4412 return false;
4415 /* Return true if INSN contains a dying pseudo in INSN right hand
4416 side. */
4417 static bool
4418 insn_rhs_dead_pseudo_p (rtx_insn *insn)
4420 rtx set = single_set (insn);
4422 gcc_assert (set != NULL);
4423 return dead_pseudo_p (SET_SRC (set), insn);
4426 /* Return true if any init insn of REGNO contains a dying pseudo in
4427 insn right hand side. */
4428 static bool
4429 init_insn_rhs_dead_pseudo_p (int regno)
4431 rtx_insn_list *insns = ira_reg_equiv[regno].init_insns;
4433 if (insns == NULL)
4434 return false;
4435 for (; insns != NULL_RTX; insns = insns->next ())
4436 if (insn_rhs_dead_pseudo_p (insns->insn ()))
4437 return true;
4438 return false;
4441 /* Return TRUE if REGNO has a reverse equivalence. The equivalence is
4442 reverse only if we have one init insn with given REGNO as a
4443 source. */
4444 static bool
4445 reverse_equiv_p (int regno)
4447 rtx_insn_list *insns = ira_reg_equiv[regno].init_insns;
4448 rtx set;
4450 if (insns == NULL)
4451 return false;
4452 if (! INSN_P (insns->insn ())
4453 || insns->next () != NULL)
4454 return false;
4455 if ((set = single_set (insns->insn ())) == NULL_RTX)
4456 return false;
4457 return REG_P (SET_SRC (set)) && (int) REGNO (SET_SRC (set)) == regno;
4460 /* Return TRUE if REGNO was reloaded in an equivalence init insn. We
4461 call this function only for non-reverse equivalence. */
4462 static bool
4463 contains_reloaded_insn_p (int regno)
4465 rtx set;
4466 rtx_insn_list *list = ira_reg_equiv[regno].init_insns;
4468 for (; list != NULL; list = list->next ())
4469 if ((set = single_set (list->insn ())) == NULL_RTX
4470 || ! REG_P (SET_DEST (set))
4471 || (int) REGNO (SET_DEST (set)) != regno)
4472 return true;
4473 return false;
4476 /* Entry function of LRA constraint pass. Return true if the
4477 constraint pass did change the code. */
4478 bool
4479 lra_constraints (bool first_p)
4481 bool changed_p;
4482 int i, hard_regno, new_insns_num;
4483 unsigned int min_len, new_min_len, uid;
4484 rtx set, x, reg, dest_reg;
4485 basic_block last_bb;
4486 bitmap_head equiv_insn_bitmap;
4487 bitmap_iterator bi;
4489 lra_constraint_iter++;
4490 if (lra_dump_file != NULL)
4491 fprintf (lra_dump_file, "\n********** Local #%d: **********\n\n",
4492 lra_constraint_iter);
4493 changed_p = false;
4494 if (pic_offset_table_rtx
4495 && REGNO (pic_offset_table_rtx) >= FIRST_PSEUDO_REGISTER)
4496 lra_risky_transformations_p = true;
4497 else
4498 lra_risky_transformations_p = false;
4499 new_insn_uid_start = get_max_uid ();
4500 new_regno_start = first_p ? lra_constraint_new_regno_start : max_reg_num ();
4501 /* Mark used hard regs for target stack size calulations. */
4502 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4503 if (lra_reg_info[i].nrefs != 0
4504 && (hard_regno = lra_get_regno_hard_regno (i)) >= 0)
4506 int j, nregs;
4508 nregs = hard_regno_nregs[hard_regno][lra_reg_info[i].biggest_mode];
4509 for (j = 0; j < nregs; j++)
4510 df_set_regs_ever_live (hard_regno + j, true);
4512 /* Do elimination before the equivalence processing as we can spill
4513 some pseudos during elimination. */
4514 lra_eliminate (false, first_p);
4515 bitmap_initialize (&equiv_insn_bitmap, &reg_obstack);
4516 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4517 if (lra_reg_info[i].nrefs != 0)
4519 ira_reg_equiv[i].profitable_p = true;
4520 reg = regno_reg_rtx[i];
4521 if (lra_get_regno_hard_regno (i) < 0 && (x = get_equiv (reg)) != reg)
4523 bool pseudo_p = contains_reg_p (x, false, false);
4525 /* After RTL transformation, we can not guarantee that
4526 pseudo in the substitution was not reloaded which might
4527 make equivalence invalid. For example, in reverse
4528 equiv of p0
4530 p0 <- ...
4532 equiv_mem <- p0
4534 the memory address register was reloaded before the 2nd
4535 insn. */
4536 if ((! first_p && pseudo_p)
4537 /* We don't use DF for compilation speed sake. So it
4538 is problematic to update live info when we use an
4539 equivalence containing pseudos in more than one
4540 BB. */
4541 || (pseudo_p && multi_block_pseudo_p (i))
4542 /* If an init insn was deleted for some reason, cancel
4543 the equiv. We could update the equiv insns after
4544 transformations including an equiv insn deletion
4545 but it is not worthy as such cases are extremely
4546 rare. */
4547 || contains_deleted_insn_p (ira_reg_equiv[i].init_insns)
4548 /* If it is not a reverse equivalence, we check that a
4549 pseudo in rhs of the init insn is not dying in the
4550 insn. Otherwise, the live info at the beginning of
4551 the corresponding BB might be wrong after we
4552 removed the insn. When the equiv can be a
4553 constant, the right hand side of the init insn can
4554 be a pseudo. */
4555 || (! reverse_equiv_p (i)
4556 && (init_insn_rhs_dead_pseudo_p (i)
4557 /* If we reloaded the pseudo in an equivalence
4558 init insn, we can not remove the equiv init
4559 insns and the init insns might write into
4560 const memory in this case. */
4561 || contains_reloaded_insn_p (i)))
4562 /* Prevent access beyond equivalent memory for
4563 paradoxical subregs. */
4564 || (MEM_P (x)
4565 && (GET_MODE_SIZE (lra_reg_info[i].biggest_mode)
4566 > GET_MODE_SIZE (GET_MODE (x))))
4567 || (pic_offset_table_rtx
4568 && ((CONST_POOL_OK_P (PSEUDO_REGNO_MODE (i), x)
4569 && (targetm.preferred_reload_class
4570 (x, lra_get_allocno_class (i)) == NO_REGS))
4571 || contains_symbol_ref_p (x))))
4572 ira_reg_equiv[i].defined_p = false;
4573 if (contains_reg_p (x, false, true))
4574 ira_reg_equiv[i].profitable_p = false;
4575 if (get_equiv (reg) != reg)
4576 bitmap_ior_into (&equiv_insn_bitmap, &lra_reg_info[i].insn_bitmap);
4579 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4580 update_equiv (i);
4581 /* We should add all insns containing pseudos which should be
4582 substituted by their equivalences. */
4583 EXECUTE_IF_SET_IN_BITMAP (&equiv_insn_bitmap, 0, uid, bi)
4584 lra_push_insn_by_uid (uid);
4585 min_len = lra_insn_stack_length ();
4586 new_insns_num = 0;
4587 last_bb = NULL;
4588 changed_p = false;
4589 while ((new_min_len = lra_insn_stack_length ()) != 0)
4591 curr_insn = lra_pop_insn ();
4592 --new_min_len;
4593 curr_bb = BLOCK_FOR_INSN (curr_insn);
4594 if (curr_bb != last_bb)
4596 last_bb = curr_bb;
4597 bb_reload_num = lra_curr_reload_num;
4599 if (min_len > new_min_len)
4601 min_len = new_min_len;
4602 new_insns_num = 0;
4604 if (new_insns_num > MAX_RELOAD_INSNS_NUMBER)
4605 internal_error
4606 ("Max. number of generated reload insns per insn is achieved (%d)\n",
4607 MAX_RELOAD_INSNS_NUMBER);
4608 new_insns_num++;
4609 if (DEBUG_INSN_P (curr_insn))
4611 /* We need to check equivalence in debug insn and change
4612 pseudo to the equivalent value if necessary. */
4613 curr_id = lra_get_insn_recog_data (curr_insn);
4614 if (bitmap_bit_p (&equiv_insn_bitmap, INSN_UID (curr_insn)))
4616 rtx old = *curr_id->operand_loc[0];
4617 *curr_id->operand_loc[0]
4618 = simplify_replace_fn_rtx (old, NULL_RTX,
4619 loc_equivalence_callback, curr_insn);
4620 if (old != *curr_id->operand_loc[0])
4622 lra_update_insn_regno_info (curr_insn);
4623 changed_p = true;
4627 else if (INSN_P (curr_insn))
4629 if ((set = single_set (curr_insn)) != NULL_RTX)
4631 dest_reg = SET_DEST (set);
4632 /* The equivalence pseudo could be set up as SUBREG in a
4633 case when it is a call restore insn in a mode
4634 different from the pseudo mode. */
4635 if (GET_CODE (dest_reg) == SUBREG)
4636 dest_reg = SUBREG_REG (dest_reg);
4637 if ((REG_P (dest_reg)
4638 && (x = get_equiv (dest_reg)) != dest_reg
4639 /* Remove insns which set up a pseudo whose value
4640 can not be changed. Such insns might be not in
4641 init_insns because we don't update equiv data
4642 during insn transformations.
4644 As an example, let suppose that a pseudo got
4645 hard register and on the 1st pass was not
4646 changed to equivalent constant. We generate an
4647 additional insn setting up the pseudo because of
4648 secondary memory movement. Then the pseudo is
4649 spilled and we use the equiv constant. In this
4650 case we should remove the additional insn and
4651 this insn is not init_insns list. */
4652 && (! MEM_P (x) || MEM_READONLY_P (x)
4653 /* Check that this is actually an insn setting
4654 up the equivalence. */
4655 || in_list_p (curr_insn,
4656 ira_reg_equiv
4657 [REGNO (dest_reg)].init_insns)))
4658 || (((x = get_equiv (SET_SRC (set))) != SET_SRC (set))
4659 && in_list_p (curr_insn,
4660 ira_reg_equiv
4661 [REGNO (SET_SRC (set))].init_insns)))
4663 /* This is equiv init insn of pseudo which did not get a
4664 hard register -- remove the insn. */
4665 if (lra_dump_file != NULL)
4667 fprintf (lra_dump_file,
4668 " Removing equiv init insn %i (freq=%d)\n",
4669 INSN_UID (curr_insn),
4670 REG_FREQ_FROM_BB (BLOCK_FOR_INSN (curr_insn)));
4671 dump_insn_slim (lra_dump_file, curr_insn);
4673 if (contains_reg_p (x, true, false))
4674 lra_risky_transformations_p = true;
4675 lra_set_insn_deleted (curr_insn);
4676 continue;
4679 curr_id = lra_get_insn_recog_data (curr_insn);
4680 curr_static_id = curr_id->insn_static_data;
4681 init_curr_insn_input_reloads ();
4682 init_curr_operand_mode ();
4683 if (curr_insn_transform (false))
4684 changed_p = true;
4685 /* Check non-transformed insns too for equiv change as USE
4686 or CLOBBER don't need reloads but can contain pseudos
4687 being changed on their equivalences. */
4688 else if (bitmap_bit_p (&equiv_insn_bitmap, INSN_UID (curr_insn))
4689 && loc_equivalence_change_p (&PATTERN (curr_insn)))
4691 lra_update_insn_regno_info (curr_insn);
4692 changed_p = true;
4696 bitmap_clear (&equiv_insn_bitmap);
4697 /* If we used a new hard regno, changed_p should be true because the
4698 hard reg is assigned to a new pseudo. */
4699 if (flag_checking && !changed_p)
4701 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4702 if (lra_reg_info[i].nrefs != 0
4703 && (hard_regno = lra_get_regno_hard_regno (i)) >= 0)
4705 int j, nregs = hard_regno_nregs[hard_regno][PSEUDO_REGNO_MODE (i)];
4707 for (j = 0; j < nregs; j++)
4708 lra_assert (df_regs_ever_live_p (hard_regno + j));
4711 return changed_p;
4714 static void initiate_invariants (void);
4715 static void finish_invariants (void);
4717 /* Initiate the LRA constraint pass. It is done once per
4718 function. */
4719 void
4720 lra_constraints_init (void)
4722 initiate_invariants ();
4725 /* Finalize the LRA constraint pass. It is done once per
4726 function. */
4727 void
4728 lra_constraints_finish (void)
4730 finish_invariants ();
4735 /* Structure describes invariants for ineheritance. */
4736 struct invariant
4738 /* The order number of the invariant. */
4739 int num;
4740 /* The invariant RTX. */
4741 rtx invariant_rtx;
4742 /* The origin insn of the invariant. */
4743 rtx_insn *insn;
4746 typedef struct invariant invariant_t;
4747 typedef invariant_t *invariant_ptr_t;
4748 typedef const invariant_t *const_invariant_ptr_t;
4750 /* Pointer to the inheritance invariants. */
4751 static vec<invariant_ptr_t> invariants;
4753 /* Allocation pool for the invariants. */
4754 static object_allocator<struct invariant> *invariants_pool;
4756 /* Hash table for the invariants. */
4757 static htab_t invariant_table;
4759 /* Hash function for INVARIANT. */
4760 static hashval_t
4761 invariant_hash (const void *invariant)
4763 rtx inv = ((const_invariant_ptr_t) invariant)->invariant_rtx;
4764 return lra_rtx_hash (inv);
4767 /* Equal function for invariants INVARIANT1 and INVARIANT2. */
4768 static int
4769 invariant_eq_p (const void *invariant1, const void *invariant2)
4771 rtx inv1 = ((const_invariant_ptr_t) invariant1)->invariant_rtx;
4772 rtx inv2 = ((const_invariant_ptr_t) invariant2)->invariant_rtx;
4774 return rtx_equal_p (inv1, inv2);
4777 /* Insert INVARIANT_RTX into the table if it is not there yet. Return
4778 invariant which is in the table. */
4779 static invariant_ptr_t
4780 insert_invariant (rtx invariant_rtx)
4782 void **entry_ptr;
4783 invariant_t invariant;
4784 invariant_ptr_t invariant_ptr;
4786 invariant.invariant_rtx = invariant_rtx;
4787 entry_ptr = htab_find_slot (invariant_table, &invariant, INSERT);
4788 if (*entry_ptr == NULL)
4790 invariant_ptr = invariants_pool->allocate ();
4791 invariant_ptr->invariant_rtx = invariant_rtx;
4792 invariant_ptr->insn = NULL;
4793 invariants.safe_push (invariant_ptr);
4794 *entry_ptr = (void *) invariant_ptr;
4796 return (invariant_ptr_t) *entry_ptr;
4799 /* Initiate the invariant table. */
4800 static void
4801 initiate_invariants (void)
4803 invariants.create (100);
4804 invariants_pool = new object_allocator<struct invariant> ("Inheritance invariants");
4805 invariant_table = htab_create (100, invariant_hash, invariant_eq_p, NULL);
4808 /* Finish the invariant table. */
4809 static void
4810 finish_invariants (void)
4812 htab_delete (invariant_table);
4813 delete invariants_pool;
4814 invariants.release ();
4817 /* Make the invariant table empty. */
4818 static void
4819 clear_invariants (void)
4821 htab_empty (invariant_table);
4822 invariants_pool->release ();
4823 invariants.truncate (0);
4828 /* This page contains code to do inheritance/split
4829 transformations. */
4831 /* Number of reloads passed so far in current EBB. */
4832 static int reloads_num;
4834 /* Number of calls passed so far in current EBB. */
4835 static int calls_num;
4837 /* Current reload pseudo check for validity of elements in
4838 USAGE_INSNS. */
4839 static int curr_usage_insns_check;
4841 /* Info about last usage of registers in EBB to do inheritance/split
4842 transformation. Inheritance transformation is done from a spilled
4843 pseudo and split transformations from a hard register or a pseudo
4844 assigned to a hard register. */
4845 struct usage_insns
4847 /* If the value is equal to CURR_USAGE_INSNS_CHECK, then the member
4848 value INSNS is valid. The insns is chain of optional debug insns
4849 and a finishing non-debug insn using the corresponding reg. The
4850 value is also used to mark the registers which are set up in the
4851 current insn. The negated insn uid is used for this. */
4852 int check;
4853 /* Value of global reloads_num at the last insn in INSNS. */
4854 int reloads_num;
4855 /* Value of global reloads_nums at the last insn in INSNS. */
4856 int calls_num;
4857 /* It can be true only for splitting. And it means that the restore
4858 insn should be put after insn given by the following member. */
4859 bool after_p;
4860 /* Next insns in the current EBB which use the original reg and the
4861 original reg value is not changed between the current insn and
4862 the next insns. In order words, e.g. for inheritance, if we need
4863 to use the original reg value again in the next insns we can try
4864 to use the value in a hard register from a reload insn of the
4865 current insn. */
4866 rtx insns;
4869 /* Map: regno -> corresponding pseudo usage insns. */
4870 static struct usage_insns *usage_insns;
4872 static void
4873 setup_next_usage_insn (int regno, rtx insn, int reloads_num, bool after_p)
4875 usage_insns[regno].check = curr_usage_insns_check;
4876 usage_insns[regno].insns = insn;
4877 usage_insns[regno].reloads_num = reloads_num;
4878 usage_insns[regno].calls_num = calls_num;
4879 usage_insns[regno].after_p = after_p;
4882 /* The function is used to form list REGNO usages which consists of
4883 optional debug insns finished by a non-debug insn using REGNO.
4884 RELOADS_NUM is current number of reload insns processed so far. */
4885 static void
4886 add_next_usage_insn (int regno, rtx_insn *insn, int reloads_num)
4888 rtx next_usage_insns;
4890 if (usage_insns[regno].check == curr_usage_insns_check
4891 && (next_usage_insns = usage_insns[regno].insns) != NULL_RTX
4892 && DEBUG_INSN_P (insn))
4894 /* Check that we did not add the debug insn yet. */
4895 if (next_usage_insns != insn
4896 && (GET_CODE (next_usage_insns) != INSN_LIST
4897 || XEXP (next_usage_insns, 0) != insn))
4898 usage_insns[regno].insns = gen_rtx_INSN_LIST (VOIDmode, insn,
4899 next_usage_insns);
4901 else if (NONDEBUG_INSN_P (insn))
4902 setup_next_usage_insn (regno, insn, reloads_num, false);
4903 else
4904 usage_insns[regno].check = 0;
4907 /* Return first non-debug insn in list USAGE_INSNS. */
4908 static rtx_insn *
4909 skip_usage_debug_insns (rtx usage_insns)
4911 rtx insn;
4913 /* Skip debug insns. */
4914 for (insn = usage_insns;
4915 insn != NULL_RTX && GET_CODE (insn) == INSN_LIST;
4916 insn = XEXP (insn, 1))
4918 return safe_as_a <rtx_insn *> (insn);
4921 /* Return true if we need secondary memory moves for insn in
4922 USAGE_INSNS after inserting inherited pseudo of class INHER_CL
4923 into the insn. */
4924 static bool
4925 check_secondary_memory_needed_p (enum reg_class inher_cl ATTRIBUTE_UNUSED,
4926 rtx usage_insns ATTRIBUTE_UNUSED)
4928 #ifndef SECONDARY_MEMORY_NEEDED
4929 return false;
4930 #else
4931 rtx_insn *insn;
4932 rtx set, dest;
4933 enum reg_class cl;
4935 if (inher_cl == ALL_REGS
4936 || (insn = skip_usage_debug_insns (usage_insns)) == NULL_RTX)
4937 return false;
4938 lra_assert (INSN_P (insn));
4939 if ((set = single_set (insn)) == NULL_RTX || ! REG_P (SET_DEST (set)))
4940 return false;
4941 dest = SET_DEST (set);
4942 if (! REG_P (dest))
4943 return false;
4944 lra_assert (inher_cl != NO_REGS);
4945 cl = get_reg_class (REGNO (dest));
4946 return (cl != NO_REGS && cl != ALL_REGS
4947 && SECONDARY_MEMORY_NEEDED (inher_cl, cl, GET_MODE (dest)));
4948 #endif
4951 /* Registers involved in inheritance/split in the current EBB
4952 (inheritance/split pseudos and original registers). */
4953 static bitmap_head check_only_regs;
4955 /* Reload pseudos can not be involded in invariant inheritance in the
4956 current EBB. */
4957 static bitmap_head invalid_invariant_regs;
4959 /* Do inheritance transformations for insn INSN, which defines (if
4960 DEF_P) or uses ORIGINAL_REGNO. NEXT_USAGE_INSNS specifies which
4961 instruction in the EBB next uses ORIGINAL_REGNO; it has the same
4962 form as the "insns" field of usage_insns. Return true if we
4963 succeed in such transformation.
4965 The transformations look like:
4967 p <- ... i <- ...
4968 ... p <- i (new insn)
4969 ... =>
4970 <- ... p ... <- ... i ...
4972 ... i <- p (new insn)
4973 <- ... p ... <- ... i ...
4974 ... =>
4975 <- ... p ... <- ... i ...
4976 where p is a spilled original pseudo and i is a new inheritance pseudo.
4979 The inheritance pseudo has the smallest class of two classes CL and
4980 class of ORIGINAL REGNO. */
4981 static bool
4982 inherit_reload_reg (bool def_p, int original_regno,
4983 enum reg_class cl, rtx_insn *insn, rtx next_usage_insns)
4985 if (optimize_function_for_size_p (cfun))
4986 return false;
4988 enum reg_class rclass = lra_get_allocno_class (original_regno);
4989 rtx original_reg = regno_reg_rtx[original_regno];
4990 rtx new_reg, usage_insn;
4991 rtx_insn *new_insns;
4993 lra_assert (! usage_insns[original_regno].after_p);
4994 if (lra_dump_file != NULL)
4995 fprintf (lra_dump_file,
4996 " <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<\n");
4997 if (! ira_reg_classes_intersect_p[cl][rclass])
4999 if (lra_dump_file != NULL)
5001 fprintf (lra_dump_file,
5002 " Rejecting inheritance for %d "
5003 "because of disjoint classes %s and %s\n",
5004 original_regno, reg_class_names[cl],
5005 reg_class_names[rclass]);
5006 fprintf (lra_dump_file,
5007 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5009 return false;
5011 if ((ira_class_subset_p[cl][rclass] && cl != rclass)
5012 /* We don't use a subset of two classes because it can be
5013 NO_REGS. This transformation is still profitable in most
5014 cases even if the classes are not intersected as register
5015 move is probably cheaper than a memory load. */
5016 || ira_class_hard_regs_num[cl] < ira_class_hard_regs_num[rclass])
5018 if (lra_dump_file != NULL)
5019 fprintf (lra_dump_file, " Use smallest class of %s and %s\n",
5020 reg_class_names[cl], reg_class_names[rclass]);
5022 rclass = cl;
5024 if (check_secondary_memory_needed_p (rclass, next_usage_insns))
5026 /* Reject inheritance resulting in secondary memory moves.
5027 Otherwise, there is a danger in LRA cycling. Also such
5028 transformation will be unprofitable. */
5029 if (lra_dump_file != NULL)
5031 rtx_insn *insn = skip_usage_debug_insns (next_usage_insns);
5032 rtx set = single_set (insn);
5034 lra_assert (set != NULL_RTX);
5036 rtx dest = SET_DEST (set);
5038 lra_assert (REG_P (dest));
5039 fprintf (lra_dump_file,
5040 " Rejecting inheritance for insn %d(%s)<-%d(%s) "
5041 "as secondary mem is needed\n",
5042 REGNO (dest), reg_class_names[get_reg_class (REGNO (dest))],
5043 original_regno, reg_class_names[rclass]);
5044 fprintf (lra_dump_file,
5045 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5047 return false;
5049 new_reg = lra_create_new_reg (GET_MODE (original_reg), original_reg,
5050 rclass, "inheritance");
5051 start_sequence ();
5052 if (def_p)
5053 lra_emit_move (original_reg, new_reg);
5054 else
5055 lra_emit_move (new_reg, original_reg);
5056 new_insns = get_insns ();
5057 end_sequence ();
5058 if (NEXT_INSN (new_insns) != NULL_RTX)
5060 if (lra_dump_file != NULL)
5062 fprintf (lra_dump_file,
5063 " Rejecting inheritance %d->%d "
5064 "as it results in 2 or more insns:\n",
5065 original_regno, REGNO (new_reg));
5066 dump_rtl_slim (lra_dump_file, new_insns, NULL, -1, 0);
5067 fprintf (lra_dump_file,
5068 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5070 return false;
5072 lra_substitute_pseudo_within_insn (insn, original_regno, new_reg, false);
5073 lra_update_insn_regno_info (insn);
5074 if (! def_p)
5075 /* We now have a new usage insn for original regno. */
5076 setup_next_usage_insn (original_regno, new_insns, reloads_num, false);
5077 if (lra_dump_file != NULL)
5078 fprintf (lra_dump_file, " Original reg change %d->%d (bb%d):\n",
5079 original_regno, REGNO (new_reg), BLOCK_FOR_INSN (insn)->index);
5080 lra_reg_info[REGNO (new_reg)].restore_rtx = regno_reg_rtx[original_regno];
5081 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
5082 bitmap_set_bit (&check_only_regs, original_regno);
5083 bitmap_set_bit (&lra_inheritance_pseudos, REGNO (new_reg));
5084 if (def_p)
5085 lra_process_new_insns (insn, NULL, new_insns,
5086 "Add original<-inheritance");
5087 else
5088 lra_process_new_insns (insn, new_insns, NULL,
5089 "Add inheritance<-original");
5090 while (next_usage_insns != NULL_RTX)
5092 if (GET_CODE (next_usage_insns) != INSN_LIST)
5094 usage_insn = next_usage_insns;
5095 lra_assert (NONDEBUG_INSN_P (usage_insn));
5096 next_usage_insns = NULL;
5098 else
5100 usage_insn = XEXP (next_usage_insns, 0);
5101 lra_assert (DEBUG_INSN_P (usage_insn));
5102 next_usage_insns = XEXP (next_usage_insns, 1);
5104 lra_substitute_pseudo (&usage_insn, original_regno, new_reg, false);
5105 lra_update_insn_regno_info (as_a <rtx_insn *> (usage_insn));
5106 if (lra_dump_file != NULL)
5108 fprintf (lra_dump_file,
5109 " Inheritance reuse change %d->%d (bb%d):\n",
5110 original_regno, REGNO (new_reg),
5111 BLOCK_FOR_INSN (usage_insn)->index);
5112 dump_insn_slim (lra_dump_file, as_a <rtx_insn *> (usage_insn));
5115 if (lra_dump_file != NULL)
5116 fprintf (lra_dump_file,
5117 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5118 return true;
5121 /* Return true if we need a caller save/restore for pseudo REGNO which
5122 was assigned to a hard register. */
5123 static inline bool
5124 need_for_call_save_p (int regno)
5126 lra_assert (regno >= FIRST_PSEUDO_REGISTER && reg_renumber[regno] >= 0);
5127 return (usage_insns[regno].calls_num < calls_num
5128 && (overlaps_hard_reg_set_p
5129 ((flag_ipa_ra &&
5130 ! hard_reg_set_empty_p (lra_reg_info[regno].actual_call_used_reg_set))
5131 ? lra_reg_info[regno].actual_call_used_reg_set
5132 : call_used_reg_set,
5133 PSEUDO_REGNO_MODE (regno), reg_renumber[regno])
5134 || HARD_REGNO_CALL_PART_CLOBBERED (reg_renumber[regno],
5135 PSEUDO_REGNO_MODE (regno))));
5138 /* Global registers occurring in the current EBB. */
5139 static bitmap_head ebb_global_regs;
5141 /* Return true if we need a split for hard register REGNO or pseudo
5142 REGNO which was assigned to a hard register.
5143 POTENTIAL_RELOAD_HARD_REGS contains hard registers which might be
5144 used for reloads since the EBB end. It is an approximation of the
5145 used hard registers in the split range. The exact value would
5146 require expensive calculations. If we were aggressive with
5147 splitting because of the approximation, the split pseudo will save
5148 the same hard register assignment and will be removed in the undo
5149 pass. We still need the approximation because too aggressive
5150 splitting would result in too inaccurate cost calculation in the
5151 assignment pass because of too many generated moves which will be
5152 probably removed in the undo pass. */
5153 static inline bool
5154 need_for_split_p (HARD_REG_SET potential_reload_hard_regs, int regno)
5156 int hard_regno = regno < FIRST_PSEUDO_REGISTER ? regno : reg_renumber[regno];
5158 lra_assert (hard_regno >= 0);
5159 return ((TEST_HARD_REG_BIT (potential_reload_hard_regs, hard_regno)
5160 /* Don't split eliminable hard registers, otherwise we can
5161 split hard registers like hard frame pointer, which
5162 lives on BB start/end according to DF-infrastructure,
5163 when there is a pseudo assigned to the register and
5164 living in the same BB. */
5165 && (regno >= FIRST_PSEUDO_REGISTER
5166 || ! TEST_HARD_REG_BIT (eliminable_regset, hard_regno))
5167 && ! TEST_HARD_REG_BIT (lra_no_alloc_regs, hard_regno)
5168 /* Don't split call clobbered hard regs living through
5169 calls, otherwise we might have a check problem in the
5170 assign sub-pass as in the most cases (exception is a
5171 situation when lra_risky_transformations_p value is
5172 true) the assign pass assumes that all pseudos living
5173 through calls are assigned to call saved hard regs. */
5174 && (regno >= FIRST_PSEUDO_REGISTER
5175 || ! TEST_HARD_REG_BIT (call_used_reg_set, regno)
5176 || usage_insns[regno].calls_num == calls_num)
5177 /* We need at least 2 reloads to make pseudo splitting
5178 profitable. We should provide hard regno splitting in
5179 any case to solve 1st insn scheduling problem when
5180 moving hard register definition up might result in
5181 impossibility to find hard register for reload pseudo of
5182 small register class. */
5183 && (usage_insns[regno].reloads_num
5184 + (regno < FIRST_PSEUDO_REGISTER ? 0 : 3) < reloads_num)
5185 && (regno < FIRST_PSEUDO_REGISTER
5186 /* For short living pseudos, spilling + inheritance can
5187 be considered a substitution for splitting.
5188 Therefore we do not splitting for local pseudos. It
5189 decreases also aggressiveness of splitting. The
5190 minimal number of references is chosen taking into
5191 account that for 2 references splitting has no sense
5192 as we can just spill the pseudo. */
5193 || (regno >= FIRST_PSEUDO_REGISTER
5194 && lra_reg_info[regno].nrefs > 3
5195 && bitmap_bit_p (&ebb_global_regs, regno))))
5196 || (regno >= FIRST_PSEUDO_REGISTER && need_for_call_save_p (regno)));
5199 /* Return class for the split pseudo created from original pseudo with
5200 ALLOCNO_CLASS and MODE which got a hard register HARD_REGNO. We
5201 choose subclass of ALLOCNO_CLASS which contains HARD_REGNO and
5202 results in no secondary memory movements. */
5203 static enum reg_class
5204 choose_split_class (enum reg_class allocno_class,
5205 int hard_regno ATTRIBUTE_UNUSED,
5206 machine_mode mode ATTRIBUTE_UNUSED)
5208 #ifndef SECONDARY_MEMORY_NEEDED
5209 return allocno_class;
5210 #else
5211 int i;
5212 enum reg_class cl, best_cl = NO_REGS;
5213 enum reg_class hard_reg_class ATTRIBUTE_UNUSED
5214 = REGNO_REG_CLASS (hard_regno);
5216 if (! SECONDARY_MEMORY_NEEDED (allocno_class, allocno_class, mode)
5217 && TEST_HARD_REG_BIT (reg_class_contents[allocno_class], hard_regno))
5218 return allocno_class;
5219 for (i = 0;
5220 (cl = reg_class_subclasses[allocno_class][i]) != LIM_REG_CLASSES;
5221 i++)
5222 if (! SECONDARY_MEMORY_NEEDED (cl, hard_reg_class, mode)
5223 && ! SECONDARY_MEMORY_NEEDED (hard_reg_class, cl, mode)
5224 && TEST_HARD_REG_BIT (reg_class_contents[cl], hard_regno)
5225 && (best_cl == NO_REGS
5226 || ira_class_hard_regs_num[best_cl] < ira_class_hard_regs_num[cl]))
5227 best_cl = cl;
5228 return best_cl;
5229 #endif
5232 /* Do split transformations for insn INSN, which defines or uses
5233 ORIGINAL_REGNO. NEXT_USAGE_INSNS specifies which instruction in
5234 the EBB next uses ORIGINAL_REGNO; it has the same form as the
5235 "insns" field of usage_insns.
5237 The transformations look like:
5239 p <- ... p <- ...
5240 ... s <- p (new insn -- save)
5241 ... =>
5242 ... p <- s (new insn -- restore)
5243 <- ... p ... <- ... p ...
5245 <- ... p ... <- ... p ...
5246 ... s <- p (new insn -- save)
5247 ... =>
5248 ... p <- s (new insn -- restore)
5249 <- ... p ... <- ... p ...
5251 where p is an original pseudo got a hard register or a hard
5252 register and s is a new split pseudo. The save is put before INSN
5253 if BEFORE_P is true. Return true if we succeed in such
5254 transformation. */
5255 static bool
5256 split_reg (bool before_p, int original_regno, rtx_insn *insn,
5257 rtx next_usage_insns)
5259 enum reg_class rclass;
5260 rtx original_reg;
5261 int hard_regno, nregs;
5262 rtx new_reg, usage_insn;
5263 rtx_insn *restore, *save;
5264 bool after_p;
5265 bool call_save_p;
5266 machine_mode mode;
5268 if (original_regno < FIRST_PSEUDO_REGISTER)
5270 rclass = ira_allocno_class_translate[REGNO_REG_CLASS (original_regno)];
5271 hard_regno = original_regno;
5272 call_save_p = false;
5273 nregs = 1;
5274 mode = lra_reg_info[hard_regno].biggest_mode;
5275 machine_mode reg_rtx_mode = GET_MODE (regno_reg_rtx[hard_regno]);
5276 /* A reg can have a biggest_mode of VOIDmode if it was only ever seen
5277 as part of a multi-word register. In that case, or if the biggest
5278 mode was larger than a register, just use the reg_rtx. Otherwise,
5279 limit the size to that of the biggest access in the function. */
5280 if (mode == VOIDmode
5281 || GET_MODE_SIZE (mode) > GET_MODE_SIZE (reg_rtx_mode))
5283 original_reg = regno_reg_rtx[hard_regno];
5284 mode = reg_rtx_mode;
5286 else
5287 original_reg = gen_rtx_REG (mode, hard_regno);
5289 else
5291 mode = PSEUDO_REGNO_MODE (original_regno);
5292 hard_regno = reg_renumber[original_regno];
5293 nregs = hard_regno_nregs[hard_regno][mode];
5294 rclass = lra_get_allocno_class (original_regno);
5295 original_reg = regno_reg_rtx[original_regno];
5296 call_save_p = need_for_call_save_p (original_regno);
5298 lra_assert (hard_regno >= 0);
5299 if (lra_dump_file != NULL)
5300 fprintf (lra_dump_file,
5301 " ((((((((((((((((((((((((((((((((((((((((((((((((\n");
5303 if (call_save_p)
5305 mode = HARD_REGNO_CALLER_SAVE_MODE (hard_regno,
5306 hard_regno_nregs[hard_regno][mode],
5307 mode);
5308 new_reg = lra_create_new_reg (mode, NULL_RTX, NO_REGS, "save");
5310 else
5312 rclass = choose_split_class (rclass, hard_regno, mode);
5313 if (rclass == NO_REGS)
5315 if (lra_dump_file != NULL)
5317 fprintf (lra_dump_file,
5318 " Rejecting split of %d(%s): "
5319 "no good reg class for %d(%s)\n",
5320 original_regno,
5321 reg_class_names[lra_get_allocno_class (original_regno)],
5322 hard_regno,
5323 reg_class_names[REGNO_REG_CLASS (hard_regno)]);
5324 fprintf
5325 (lra_dump_file,
5326 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5328 return false;
5330 new_reg = lra_create_new_reg (mode, original_reg, rclass, "split");
5331 reg_renumber[REGNO (new_reg)] = hard_regno;
5333 save = emit_spill_move (true, new_reg, original_reg);
5334 if (NEXT_INSN (save) != NULL_RTX && !call_save_p)
5336 if (lra_dump_file != NULL)
5338 fprintf
5339 (lra_dump_file,
5340 " Rejecting split %d->%d resulting in > 2 save insns:\n",
5341 original_regno, REGNO (new_reg));
5342 dump_rtl_slim (lra_dump_file, save, NULL, -1, 0);
5343 fprintf (lra_dump_file,
5344 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5346 return false;
5348 restore = emit_spill_move (false, new_reg, original_reg);
5349 if (NEXT_INSN (restore) != NULL_RTX && !call_save_p)
5351 if (lra_dump_file != NULL)
5353 fprintf (lra_dump_file,
5354 " Rejecting split %d->%d "
5355 "resulting in > 2 restore insns:\n",
5356 original_regno, REGNO (new_reg));
5357 dump_rtl_slim (lra_dump_file, restore, NULL, -1, 0);
5358 fprintf (lra_dump_file,
5359 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5361 return false;
5363 after_p = usage_insns[original_regno].after_p;
5364 lra_reg_info[REGNO (new_reg)].restore_rtx = regno_reg_rtx[original_regno];
5365 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
5366 bitmap_set_bit (&check_only_regs, original_regno);
5367 bitmap_set_bit (&lra_split_regs, REGNO (new_reg));
5368 for (;;)
5370 if (GET_CODE (next_usage_insns) != INSN_LIST)
5372 usage_insn = next_usage_insns;
5373 break;
5375 usage_insn = XEXP (next_usage_insns, 0);
5376 lra_assert (DEBUG_INSN_P (usage_insn));
5377 next_usage_insns = XEXP (next_usage_insns, 1);
5378 lra_substitute_pseudo (&usage_insn, original_regno, new_reg, false);
5379 lra_update_insn_regno_info (as_a <rtx_insn *> (usage_insn));
5380 if (lra_dump_file != NULL)
5382 fprintf (lra_dump_file, " Split reuse change %d->%d:\n",
5383 original_regno, REGNO (new_reg));
5384 dump_insn_slim (lra_dump_file, as_a <rtx_insn *> (usage_insn));
5387 lra_assert (NOTE_P (usage_insn) || NONDEBUG_INSN_P (usage_insn));
5388 lra_assert (usage_insn != insn || (after_p && before_p));
5389 lra_process_new_insns (as_a <rtx_insn *> (usage_insn),
5390 after_p ? NULL : restore,
5391 after_p ? restore : NULL,
5392 call_save_p
5393 ? "Add reg<-save" : "Add reg<-split");
5394 lra_process_new_insns (insn, before_p ? save : NULL,
5395 before_p ? NULL : save,
5396 call_save_p
5397 ? "Add save<-reg" : "Add split<-reg");
5398 if (nregs > 1)
5399 /* If we are trying to split multi-register. We should check
5400 conflicts on the next assignment sub-pass. IRA can allocate on
5401 sub-register levels, LRA do this on pseudos level right now and
5402 this discrepancy may create allocation conflicts after
5403 splitting. */
5404 lra_risky_transformations_p = true;
5405 if (lra_dump_file != NULL)
5406 fprintf (lra_dump_file,
5407 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5408 return true;
5411 /* Recognize that we need a split transformation for insn INSN, which
5412 defines or uses REGNO in its insn biggest MODE (we use it only if
5413 REGNO is a hard register). POTENTIAL_RELOAD_HARD_REGS contains
5414 hard registers which might be used for reloads since the EBB end.
5415 Put the save before INSN if BEFORE_P is true. MAX_UID is maximla
5416 uid before starting INSN processing. Return true if we succeed in
5417 such transformation. */
5418 static bool
5419 split_if_necessary (int regno, machine_mode mode,
5420 HARD_REG_SET potential_reload_hard_regs,
5421 bool before_p, rtx_insn *insn, int max_uid)
5423 bool res = false;
5424 int i, nregs = 1;
5425 rtx next_usage_insns;
5427 if (regno < FIRST_PSEUDO_REGISTER)
5428 nregs = hard_regno_nregs[regno][mode];
5429 for (i = 0; i < nregs; i++)
5430 if (usage_insns[regno + i].check == curr_usage_insns_check
5431 && (next_usage_insns = usage_insns[regno + i].insns) != NULL_RTX
5432 /* To avoid processing the register twice or more. */
5433 && ((GET_CODE (next_usage_insns) != INSN_LIST
5434 && INSN_UID (next_usage_insns) < max_uid)
5435 || (GET_CODE (next_usage_insns) == INSN_LIST
5436 && (INSN_UID (XEXP (next_usage_insns, 0)) < max_uid)))
5437 && need_for_split_p (potential_reload_hard_regs, regno + i)
5438 && split_reg (before_p, regno + i, insn, next_usage_insns))
5439 res = true;
5440 return res;
5443 /* Return TRUE if rtx X is considered as an invariant for
5444 inheritance. */
5445 static bool
5446 invariant_p (const_rtx x)
5448 machine_mode mode;
5449 const char *fmt;
5450 enum rtx_code code;
5451 int i, j;
5453 code = GET_CODE (x);
5454 mode = GET_MODE (x);
5455 if (code == SUBREG)
5457 x = SUBREG_REG (x);
5458 code = GET_CODE (x);
5459 if (GET_MODE_SIZE (GET_MODE (x)) > GET_MODE_SIZE (mode))
5460 mode = GET_MODE (x);
5463 if (MEM_P (x))
5464 return false;
5466 if (REG_P (x))
5468 int i, nregs, regno = REGNO (x);
5470 if (regno >= FIRST_PSEUDO_REGISTER || regno == STACK_POINTER_REGNUM
5471 || TEST_HARD_REG_BIT (eliminable_regset, regno)
5472 || GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
5473 return false;
5474 nregs = hard_regno_nregs[regno][mode];
5475 for (i = 0; i < nregs; i++)
5476 if (! fixed_regs[regno + i]
5477 /* A hard register may be clobbered in the current insn
5478 but we can ignore this case because if the hard
5479 register is used it should be set somewhere after the
5480 clobber. */
5481 || bitmap_bit_p (&invalid_invariant_regs, regno + i))
5482 return false;
5484 fmt = GET_RTX_FORMAT (code);
5485 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
5487 if (fmt[i] == 'e')
5489 if (! invariant_p (XEXP (x, i)))
5490 return false;
5492 else if (fmt[i] == 'E')
5494 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
5495 if (! invariant_p (XVECEXP (x, i, j)))
5496 return false;
5499 return true;
5502 /* We have 'dest_reg <- invariant'. Let us try to make an invariant
5503 inheritance transformation (using dest_reg instead invariant in a
5504 subsequent insn). */
5505 static bool
5506 process_invariant_for_inheritance (rtx dst_reg, rtx invariant_rtx)
5508 invariant_ptr_t invariant_ptr;
5509 rtx_insn *insn, *new_insns;
5510 rtx insn_set, insn_reg, new_reg;
5511 int insn_regno;
5512 bool succ_p = false;
5513 int dst_regno = REGNO (dst_reg);
5514 enum machine_mode dst_mode = GET_MODE (dst_reg);
5515 enum reg_class cl = lra_get_allocno_class (dst_regno), insn_reg_cl;
5517 invariant_ptr = insert_invariant (invariant_rtx);
5518 if ((insn = invariant_ptr->insn) != NULL_RTX)
5520 /* We have a subsequent insn using the invariant. */
5521 insn_set = single_set (insn);
5522 lra_assert (insn_set != NULL);
5523 insn_reg = SET_DEST (insn_set);
5524 lra_assert (REG_P (insn_reg));
5525 insn_regno = REGNO (insn_reg);
5526 insn_reg_cl = lra_get_allocno_class (insn_regno);
5528 if (dst_mode == GET_MODE (insn_reg)
5529 /* We should consider only result move reg insns which are
5530 cheap. */
5531 && targetm.register_move_cost (dst_mode, cl, insn_reg_cl) == 2
5532 && targetm.register_move_cost (dst_mode, cl, cl) == 2)
5534 if (lra_dump_file != NULL)
5535 fprintf (lra_dump_file,
5536 " [[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[\n");
5537 new_reg = lra_create_new_reg (dst_mode, dst_reg,
5538 cl, "invariant inheritance");
5539 bitmap_set_bit (&lra_inheritance_pseudos, REGNO (new_reg));
5540 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
5541 lra_reg_info[REGNO (new_reg)].restore_rtx = PATTERN (insn);
5542 start_sequence ();
5543 lra_emit_move (new_reg, dst_reg);
5544 new_insns = get_insns ();
5545 end_sequence ();
5546 lra_process_new_insns (curr_insn, NULL, new_insns,
5547 "Add invariant inheritance<-original");
5548 start_sequence ();
5549 lra_emit_move (SET_DEST (insn_set), new_reg);
5550 new_insns = get_insns ();
5551 end_sequence ();
5552 lra_process_new_insns (insn, NULL, new_insns,
5553 "Changing reload<-inheritance");
5554 lra_set_insn_deleted (insn);
5555 succ_p = true;
5556 if (lra_dump_file != NULL)
5558 fprintf (lra_dump_file,
5559 " Invariant inheritance reuse change %d (bb%d):\n",
5560 REGNO (new_reg), BLOCK_FOR_INSN (insn)->index);
5561 dump_insn_slim (lra_dump_file, insn);
5562 fprintf (lra_dump_file,
5563 " ]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]\n");
5567 invariant_ptr->insn = curr_insn;
5568 return succ_p;
5571 /* Check only registers living at the current program point in the
5572 current EBB. */
5573 static bitmap_head live_regs;
5575 /* Update live info in EBB given by its HEAD and TAIL insns after
5576 inheritance/split transformation. The function removes dead moves
5577 too. */
5578 static void
5579 update_ebb_live_info (rtx_insn *head, rtx_insn *tail)
5581 unsigned int j;
5582 int i, regno;
5583 bool live_p;
5584 rtx_insn *prev_insn;
5585 rtx set;
5586 bool remove_p;
5587 basic_block last_bb, prev_bb, curr_bb;
5588 bitmap_iterator bi;
5589 struct lra_insn_reg *reg;
5590 edge e;
5591 edge_iterator ei;
5593 last_bb = BLOCK_FOR_INSN (tail);
5594 prev_bb = NULL;
5595 for (curr_insn = tail;
5596 curr_insn != PREV_INSN (head);
5597 curr_insn = prev_insn)
5599 prev_insn = PREV_INSN (curr_insn);
5600 /* We need to process empty blocks too. They contain
5601 NOTE_INSN_BASIC_BLOCK referring for the basic block. */
5602 if (NOTE_P (curr_insn) && NOTE_KIND (curr_insn) != NOTE_INSN_BASIC_BLOCK)
5603 continue;
5604 curr_bb = BLOCK_FOR_INSN (curr_insn);
5605 if (curr_bb != prev_bb)
5607 if (prev_bb != NULL)
5609 /* Update df_get_live_in (prev_bb): */
5610 EXECUTE_IF_SET_IN_BITMAP (&check_only_regs, 0, j, bi)
5611 if (bitmap_bit_p (&live_regs, j))
5612 bitmap_set_bit (df_get_live_in (prev_bb), j);
5613 else
5614 bitmap_clear_bit (df_get_live_in (prev_bb), j);
5616 if (curr_bb != last_bb)
5618 /* Update df_get_live_out (curr_bb): */
5619 EXECUTE_IF_SET_IN_BITMAP (&check_only_regs, 0, j, bi)
5621 live_p = bitmap_bit_p (&live_regs, j);
5622 if (! live_p)
5623 FOR_EACH_EDGE (e, ei, curr_bb->succs)
5624 if (bitmap_bit_p (df_get_live_in (e->dest), j))
5626 live_p = true;
5627 break;
5629 if (live_p)
5630 bitmap_set_bit (df_get_live_out (curr_bb), j);
5631 else
5632 bitmap_clear_bit (df_get_live_out (curr_bb), j);
5635 prev_bb = curr_bb;
5636 bitmap_and (&live_regs, &check_only_regs, df_get_live_out (curr_bb));
5638 if (! NONDEBUG_INSN_P (curr_insn))
5639 continue;
5640 curr_id = lra_get_insn_recog_data (curr_insn);
5641 curr_static_id = curr_id->insn_static_data;
5642 remove_p = false;
5643 if ((set = single_set (curr_insn)) != NULL_RTX
5644 && REG_P (SET_DEST (set))
5645 && (regno = REGNO (SET_DEST (set))) >= FIRST_PSEUDO_REGISTER
5646 && SET_DEST (set) != pic_offset_table_rtx
5647 && bitmap_bit_p (&check_only_regs, regno)
5648 && ! bitmap_bit_p (&live_regs, regno))
5649 remove_p = true;
5650 /* See which defined values die here. */
5651 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5652 if (reg->type == OP_OUT && ! reg->subreg_p)
5653 bitmap_clear_bit (&live_regs, reg->regno);
5654 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
5655 if (reg->type == OP_OUT && ! reg->subreg_p)
5656 bitmap_clear_bit (&live_regs, reg->regno);
5657 if (curr_id->arg_hard_regs != NULL)
5658 /* Make clobbered argument hard registers die. */
5659 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
5660 if (regno >= FIRST_PSEUDO_REGISTER)
5661 bitmap_clear_bit (&live_regs, regno - FIRST_PSEUDO_REGISTER);
5662 /* Mark each used value as live. */
5663 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5664 if (reg->type != OP_OUT
5665 && bitmap_bit_p (&check_only_regs, reg->regno))
5666 bitmap_set_bit (&live_regs, reg->regno);
5667 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
5668 if (reg->type != OP_OUT
5669 && bitmap_bit_p (&check_only_regs, reg->regno))
5670 bitmap_set_bit (&live_regs, reg->regno);
5671 if (curr_id->arg_hard_regs != NULL)
5672 /* Make used argument hard registers live. */
5673 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
5674 if (regno < FIRST_PSEUDO_REGISTER
5675 && bitmap_bit_p (&check_only_regs, regno))
5676 bitmap_set_bit (&live_regs, regno);
5677 /* It is quite important to remove dead move insns because it
5678 means removing dead store. We don't need to process them for
5679 constraints. */
5680 if (remove_p)
5682 if (lra_dump_file != NULL)
5684 fprintf (lra_dump_file, " Removing dead insn:\n ");
5685 dump_insn_slim (lra_dump_file, curr_insn);
5687 lra_set_insn_deleted (curr_insn);
5692 /* The structure describes info to do an inheritance for the current
5693 insn. We need to collect such info first before doing the
5694 transformations because the transformations change the insn
5695 internal representation. */
5696 struct to_inherit
5698 /* Original regno. */
5699 int regno;
5700 /* Subsequent insns which can inherit original reg value. */
5701 rtx insns;
5704 /* Array containing all info for doing inheritance from the current
5705 insn. */
5706 static struct to_inherit to_inherit[LRA_MAX_INSN_RELOADS];
5708 /* Number elements in the previous array. */
5709 static int to_inherit_num;
5711 /* Add inheritance info REGNO and INSNS. Their meaning is described in
5712 structure to_inherit. */
5713 static void
5714 add_to_inherit (int regno, rtx insns)
5716 int i;
5718 for (i = 0; i < to_inherit_num; i++)
5719 if (to_inherit[i].regno == regno)
5720 return;
5721 lra_assert (to_inherit_num < LRA_MAX_INSN_RELOADS);
5722 to_inherit[to_inherit_num].regno = regno;
5723 to_inherit[to_inherit_num++].insns = insns;
5726 /* Return the last non-debug insn in basic block BB, or the block begin
5727 note if none. */
5728 static rtx_insn *
5729 get_last_insertion_point (basic_block bb)
5731 rtx_insn *insn;
5733 FOR_BB_INSNS_REVERSE (bb, insn)
5734 if (NONDEBUG_INSN_P (insn) || NOTE_INSN_BASIC_BLOCK_P (insn))
5735 return insn;
5736 gcc_unreachable ();
5739 /* Set up RES by registers living on edges FROM except the edge (FROM,
5740 TO) or by registers set up in a jump insn in BB FROM. */
5741 static void
5742 get_live_on_other_edges (basic_block from, basic_block to, bitmap res)
5744 rtx_insn *last;
5745 struct lra_insn_reg *reg;
5746 edge e;
5747 edge_iterator ei;
5749 lra_assert (to != NULL);
5750 bitmap_clear (res);
5751 FOR_EACH_EDGE (e, ei, from->succs)
5752 if (e->dest != to)
5753 bitmap_ior_into (res, df_get_live_in (e->dest));
5754 last = get_last_insertion_point (from);
5755 if (! JUMP_P (last))
5756 return;
5757 curr_id = lra_get_insn_recog_data (last);
5758 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5759 if (reg->type != OP_IN)
5760 bitmap_set_bit (res, reg->regno);
5763 /* Used as a temporary results of some bitmap calculations. */
5764 static bitmap_head temp_bitmap;
5766 /* We split for reloads of small class of hard regs. The following
5767 defines how many hard regs the class should have to be qualified as
5768 small. The code is mostly oriented to x86/x86-64 architecture
5769 where some insns need to use only specific register or pair of
5770 registers and these register can live in RTL explicitly, e.g. for
5771 parameter passing. */
5772 static const int max_small_class_regs_num = 2;
5774 /* Do inheritance/split transformations in EBB starting with HEAD and
5775 finishing on TAIL. We process EBB insns in the reverse order.
5776 Return true if we did any inheritance/split transformation in the
5777 EBB.
5779 We should avoid excessive splitting which results in worse code
5780 because of inaccurate cost calculations for spilling new split
5781 pseudos in such case. To achieve this we do splitting only if
5782 register pressure is high in given basic block and there are reload
5783 pseudos requiring hard registers. We could do more register
5784 pressure calculations at any given program point to avoid necessary
5785 splitting even more but it is to expensive and the current approach
5786 works well enough. */
5787 static bool
5788 inherit_in_ebb (rtx_insn *head, rtx_insn *tail)
5790 int i, src_regno, dst_regno, nregs;
5791 bool change_p, succ_p, update_reloads_num_p;
5792 rtx_insn *prev_insn, *last_insn;
5793 rtx next_usage_insns, curr_set;
5794 enum reg_class cl;
5795 struct lra_insn_reg *reg;
5796 basic_block last_processed_bb, curr_bb = NULL;
5797 HARD_REG_SET potential_reload_hard_regs, live_hard_regs;
5798 bitmap to_process;
5799 unsigned int j;
5800 bitmap_iterator bi;
5801 bool head_p, after_p;
5803 change_p = false;
5804 curr_usage_insns_check++;
5805 clear_invariants ();
5806 reloads_num = calls_num = 0;
5807 bitmap_clear (&check_only_regs);
5808 bitmap_clear (&invalid_invariant_regs);
5809 last_processed_bb = NULL;
5810 CLEAR_HARD_REG_SET (potential_reload_hard_regs);
5811 COPY_HARD_REG_SET (live_hard_regs, eliminable_regset);
5812 IOR_HARD_REG_SET (live_hard_regs, lra_no_alloc_regs);
5813 /* We don't process new insns generated in the loop. */
5814 for (curr_insn = tail; curr_insn != PREV_INSN (head); curr_insn = prev_insn)
5816 prev_insn = PREV_INSN (curr_insn);
5817 if (BLOCK_FOR_INSN (curr_insn) != NULL)
5818 curr_bb = BLOCK_FOR_INSN (curr_insn);
5819 if (last_processed_bb != curr_bb)
5821 /* We are at the end of BB. Add qualified living
5822 pseudos for potential splitting. */
5823 to_process = df_get_live_out (curr_bb);
5824 if (last_processed_bb != NULL)
5826 /* We are somewhere in the middle of EBB. */
5827 get_live_on_other_edges (curr_bb, last_processed_bb,
5828 &temp_bitmap);
5829 to_process = &temp_bitmap;
5831 last_processed_bb = curr_bb;
5832 last_insn = get_last_insertion_point (curr_bb);
5833 after_p = (! JUMP_P (last_insn)
5834 && (! CALL_P (last_insn)
5835 || (find_reg_note (last_insn,
5836 REG_NORETURN, NULL_RTX) == NULL_RTX
5837 && ! SIBLING_CALL_P (last_insn))));
5838 CLEAR_HARD_REG_SET (potential_reload_hard_regs);
5839 EXECUTE_IF_SET_IN_BITMAP (to_process, 0, j, bi)
5841 if ((int) j >= lra_constraint_new_regno_start)
5842 break;
5843 if (j < FIRST_PSEUDO_REGISTER || reg_renumber[j] >= 0)
5845 if (j < FIRST_PSEUDO_REGISTER)
5846 SET_HARD_REG_BIT (live_hard_regs, j);
5847 else
5848 add_to_hard_reg_set (&live_hard_regs,
5849 PSEUDO_REGNO_MODE (j),
5850 reg_renumber[j]);
5851 setup_next_usage_insn (j, last_insn, reloads_num, after_p);
5855 src_regno = dst_regno = -1;
5856 curr_set = single_set (curr_insn);
5857 if (curr_set != NULL_RTX && REG_P (SET_DEST (curr_set)))
5858 dst_regno = REGNO (SET_DEST (curr_set));
5859 if (curr_set != NULL_RTX && REG_P (SET_SRC (curr_set)))
5860 src_regno = REGNO (SET_SRC (curr_set));
5861 update_reloads_num_p = true;
5862 if (src_regno < lra_constraint_new_regno_start
5863 && src_regno >= FIRST_PSEUDO_REGISTER
5864 && reg_renumber[src_regno] < 0
5865 && dst_regno >= lra_constraint_new_regno_start
5866 && (cl = lra_get_allocno_class (dst_regno)) != NO_REGS)
5868 /* 'reload_pseudo <- original_pseudo'. */
5869 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
5870 reloads_num++;
5871 update_reloads_num_p = false;
5872 succ_p = false;
5873 if (usage_insns[src_regno].check == curr_usage_insns_check
5874 && (next_usage_insns = usage_insns[src_regno].insns) != NULL_RTX)
5875 succ_p = inherit_reload_reg (false, src_regno, cl,
5876 curr_insn, next_usage_insns);
5877 if (succ_p)
5878 change_p = true;
5879 else
5880 setup_next_usage_insn (src_regno, curr_insn, reloads_num, false);
5881 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
5882 IOR_HARD_REG_SET (potential_reload_hard_regs,
5883 reg_class_contents[cl]);
5885 else if (src_regno < 0
5886 && dst_regno >= lra_constraint_new_regno_start
5887 && invariant_p (SET_SRC (curr_set))
5888 && (cl = lra_get_allocno_class (dst_regno)) != NO_REGS
5889 && ! bitmap_bit_p (&invalid_invariant_regs, dst_regno))
5891 /* 'reload_pseudo <- invariant'. */
5892 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
5893 reloads_num++;
5894 update_reloads_num_p = false;
5895 if (process_invariant_for_inheritance (SET_DEST (curr_set), SET_SRC (curr_set)))
5896 change_p = true;
5897 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
5898 IOR_HARD_REG_SET (potential_reload_hard_regs,
5899 reg_class_contents[cl]);
5901 else if (src_regno >= lra_constraint_new_regno_start
5902 && dst_regno < lra_constraint_new_regno_start
5903 && dst_regno >= FIRST_PSEUDO_REGISTER
5904 && reg_renumber[dst_regno] < 0
5905 && (cl = lra_get_allocno_class (src_regno)) != NO_REGS
5906 && usage_insns[dst_regno].check == curr_usage_insns_check
5907 && (next_usage_insns
5908 = usage_insns[dst_regno].insns) != NULL_RTX)
5910 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
5911 reloads_num++;
5912 update_reloads_num_p = false;
5913 /* 'original_pseudo <- reload_pseudo'. */
5914 if (! JUMP_P (curr_insn)
5915 && inherit_reload_reg (true, dst_regno, cl,
5916 curr_insn, next_usage_insns))
5917 change_p = true;
5918 /* Invalidate. */
5919 usage_insns[dst_regno].check = 0;
5920 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
5921 IOR_HARD_REG_SET (potential_reload_hard_regs,
5922 reg_class_contents[cl]);
5924 else if (INSN_P (curr_insn))
5926 int iter;
5927 int max_uid = get_max_uid ();
5929 curr_id = lra_get_insn_recog_data (curr_insn);
5930 curr_static_id = curr_id->insn_static_data;
5931 to_inherit_num = 0;
5932 /* Process insn definitions. */
5933 for (iter = 0; iter < 2; iter++)
5934 for (reg = iter == 0 ? curr_id->regs : curr_static_id->hard_regs;
5935 reg != NULL;
5936 reg = reg->next)
5937 if (reg->type != OP_IN
5938 && (dst_regno = reg->regno) < lra_constraint_new_regno_start)
5940 if (dst_regno >= FIRST_PSEUDO_REGISTER && reg->type == OP_OUT
5941 && reg_renumber[dst_regno] < 0 && ! reg->subreg_p
5942 && usage_insns[dst_regno].check == curr_usage_insns_check
5943 && (next_usage_insns
5944 = usage_insns[dst_regno].insns) != NULL_RTX)
5946 struct lra_insn_reg *r;
5948 for (r = curr_id->regs; r != NULL; r = r->next)
5949 if (r->type != OP_OUT && r->regno == dst_regno)
5950 break;
5951 /* Don't do inheritance if the pseudo is also
5952 used in the insn. */
5953 if (r == NULL)
5954 /* We can not do inheritance right now
5955 because the current insn reg info (chain
5956 regs) can change after that. */
5957 add_to_inherit (dst_regno, next_usage_insns);
5959 /* We can not process one reg twice here because of
5960 usage_insns invalidation. */
5961 if ((dst_regno < FIRST_PSEUDO_REGISTER
5962 || reg_renumber[dst_regno] >= 0)
5963 && ! reg->subreg_p && reg->type != OP_IN)
5965 HARD_REG_SET s;
5967 if (split_if_necessary (dst_regno, reg->biggest_mode,
5968 potential_reload_hard_regs,
5969 false, curr_insn, max_uid))
5970 change_p = true;
5971 CLEAR_HARD_REG_SET (s);
5972 if (dst_regno < FIRST_PSEUDO_REGISTER)
5973 add_to_hard_reg_set (&s, reg->biggest_mode, dst_regno);
5974 else
5975 add_to_hard_reg_set (&s, PSEUDO_REGNO_MODE (dst_regno),
5976 reg_renumber[dst_regno]);
5977 AND_COMPL_HARD_REG_SET (live_hard_regs, s);
5979 /* We should invalidate potential inheritance or
5980 splitting for the current insn usages to the next
5981 usage insns (see code below) as the output pseudo
5982 prevents this. */
5983 if ((dst_regno >= FIRST_PSEUDO_REGISTER
5984 && reg_renumber[dst_regno] < 0)
5985 || (reg->type == OP_OUT && ! reg->subreg_p
5986 && (dst_regno < FIRST_PSEUDO_REGISTER
5987 || reg_renumber[dst_regno] >= 0)))
5989 /* Invalidate and mark definitions. */
5990 if (dst_regno >= FIRST_PSEUDO_REGISTER)
5991 usage_insns[dst_regno].check = -(int) INSN_UID (curr_insn);
5992 else
5994 nregs = hard_regno_nregs[dst_regno][reg->biggest_mode];
5995 for (i = 0; i < nregs; i++)
5996 usage_insns[dst_regno + i].check
5997 = -(int) INSN_UID (curr_insn);
6001 /* Process clobbered call regs. */
6002 if (curr_id->arg_hard_regs != NULL)
6003 for (i = 0; (dst_regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6004 if (dst_regno >= FIRST_PSEUDO_REGISTER)
6005 usage_insns[dst_regno - FIRST_PSEUDO_REGISTER].check
6006 = -(int) INSN_UID (curr_insn);
6007 if (! JUMP_P (curr_insn))
6008 for (i = 0; i < to_inherit_num; i++)
6009 if (inherit_reload_reg (true, to_inherit[i].regno,
6010 ALL_REGS, curr_insn,
6011 to_inherit[i].insns))
6012 change_p = true;
6013 if (CALL_P (curr_insn))
6015 rtx cheap, pat, dest;
6016 rtx_insn *restore;
6017 int regno, hard_regno;
6019 calls_num++;
6020 if ((cheap = find_reg_note (curr_insn,
6021 REG_RETURNED, NULL_RTX)) != NULL_RTX
6022 && ((cheap = XEXP (cheap, 0)), true)
6023 && (regno = REGNO (cheap)) >= FIRST_PSEUDO_REGISTER
6024 && (hard_regno = reg_renumber[regno]) >= 0
6025 /* If there are pending saves/restores, the
6026 optimization is not worth. */
6027 && usage_insns[regno].calls_num == calls_num - 1
6028 && TEST_HARD_REG_BIT (call_used_reg_set, hard_regno))
6030 /* Restore the pseudo from the call result as
6031 REG_RETURNED note says that the pseudo value is
6032 in the call result and the pseudo is an argument
6033 of the call. */
6034 pat = PATTERN (curr_insn);
6035 if (GET_CODE (pat) == PARALLEL)
6036 pat = XVECEXP (pat, 0, 0);
6037 dest = SET_DEST (pat);
6038 /* For multiple return values dest is PARALLEL.
6039 Currently we handle only single return value case. */
6040 if (REG_P (dest))
6042 start_sequence ();
6043 emit_move_insn (cheap, copy_rtx (dest));
6044 restore = get_insns ();
6045 end_sequence ();
6046 lra_process_new_insns (curr_insn, NULL, restore,
6047 "Inserting call parameter restore");
6048 /* We don't need to save/restore of the pseudo from
6049 this call. */
6050 usage_insns[regno].calls_num = calls_num;
6051 bitmap_set_bit (&check_only_regs, regno);
6055 to_inherit_num = 0;
6056 /* Process insn usages. */
6057 for (iter = 0; iter < 2; iter++)
6058 for (reg = iter == 0 ? curr_id->regs : curr_static_id->hard_regs;
6059 reg != NULL;
6060 reg = reg->next)
6061 if ((reg->type != OP_OUT
6062 || (reg->type == OP_OUT && reg->subreg_p))
6063 && (src_regno = reg->regno) < lra_constraint_new_regno_start)
6065 if (src_regno >= FIRST_PSEUDO_REGISTER
6066 && reg_renumber[src_regno] < 0 && reg->type == OP_IN)
6068 if (usage_insns[src_regno].check == curr_usage_insns_check
6069 && (next_usage_insns
6070 = usage_insns[src_regno].insns) != NULL_RTX
6071 && NONDEBUG_INSN_P (curr_insn))
6072 add_to_inherit (src_regno, next_usage_insns);
6073 else if (usage_insns[src_regno].check
6074 != -(int) INSN_UID (curr_insn))
6075 /* Add usages but only if the reg is not set up
6076 in the same insn. */
6077 add_next_usage_insn (src_regno, curr_insn, reloads_num);
6079 else if (src_regno < FIRST_PSEUDO_REGISTER
6080 || reg_renumber[src_regno] >= 0)
6082 bool before_p;
6083 rtx_insn *use_insn = curr_insn;
6085 before_p = (JUMP_P (curr_insn)
6086 || (CALL_P (curr_insn) && reg->type == OP_IN));
6087 if (NONDEBUG_INSN_P (curr_insn)
6088 && (! JUMP_P (curr_insn) || reg->type == OP_IN)
6089 && split_if_necessary (src_regno, reg->biggest_mode,
6090 potential_reload_hard_regs,
6091 before_p, curr_insn, max_uid))
6093 if (reg->subreg_p)
6094 lra_risky_transformations_p = true;
6095 change_p = true;
6096 /* Invalidate. */
6097 usage_insns[src_regno].check = 0;
6098 if (before_p)
6099 use_insn = PREV_INSN (curr_insn);
6101 if (NONDEBUG_INSN_P (curr_insn))
6103 if (src_regno < FIRST_PSEUDO_REGISTER)
6104 add_to_hard_reg_set (&live_hard_regs,
6105 reg->biggest_mode, src_regno);
6106 else
6107 add_to_hard_reg_set (&live_hard_regs,
6108 PSEUDO_REGNO_MODE (src_regno),
6109 reg_renumber[src_regno]);
6111 add_next_usage_insn (src_regno, use_insn, reloads_num);
6114 /* Process used call regs. */
6115 if (curr_id->arg_hard_regs != NULL)
6116 for (i = 0; (src_regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6117 if (src_regno < FIRST_PSEUDO_REGISTER)
6119 SET_HARD_REG_BIT (live_hard_regs, src_regno);
6120 add_next_usage_insn (src_regno, curr_insn, reloads_num);
6122 for (i = 0; i < to_inherit_num; i++)
6124 src_regno = to_inherit[i].regno;
6125 if (inherit_reload_reg (false, src_regno, ALL_REGS,
6126 curr_insn, to_inherit[i].insns))
6127 change_p = true;
6128 else
6129 setup_next_usage_insn (src_regno, curr_insn, reloads_num, false);
6132 if (update_reloads_num_p
6133 && NONDEBUG_INSN_P (curr_insn) && curr_set != NULL_RTX)
6135 int regno = -1;
6136 if ((REG_P (SET_DEST (curr_set))
6137 && (regno = REGNO (SET_DEST (curr_set))) >= lra_constraint_new_regno_start
6138 && reg_renumber[regno] < 0
6139 && (cl = lra_get_allocno_class (regno)) != NO_REGS)
6140 || (REG_P (SET_SRC (curr_set))
6141 && (regno = REGNO (SET_SRC (curr_set))) >= lra_constraint_new_regno_start
6142 && reg_renumber[regno] < 0
6143 && (cl = lra_get_allocno_class (regno)) != NO_REGS))
6145 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6146 reloads_num++;
6147 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6148 IOR_HARD_REG_SET (potential_reload_hard_regs,
6149 reg_class_contents[cl]);
6152 if (NONDEBUG_INSN_P (curr_insn))
6154 int regno;
6156 /* Invalidate invariants with changed regs. */
6157 curr_id = lra_get_insn_recog_data (curr_insn);
6158 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
6159 if (reg->type != OP_IN)
6160 bitmap_set_bit (&invalid_invariant_regs, reg->regno);
6161 curr_static_id = curr_id->insn_static_data;
6162 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
6163 if (reg->type != OP_IN)
6164 bitmap_set_bit (&invalid_invariant_regs, reg->regno);
6165 if (curr_id->arg_hard_regs != NULL)
6166 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6167 bitmap_set_bit (&invalid_invariant_regs,
6168 regno >= FIRST_PSEUDO_REGISTER
6169 ? regno : regno - FIRST_PSEUDO_REGISTER);
6171 /* We reached the start of the current basic block. */
6172 if (prev_insn == NULL_RTX || prev_insn == PREV_INSN (head)
6173 || BLOCK_FOR_INSN (prev_insn) != curr_bb)
6175 /* We reached the beginning of the current block -- do
6176 rest of spliting in the current BB. */
6177 to_process = df_get_live_in (curr_bb);
6178 if (BLOCK_FOR_INSN (head) != curr_bb)
6180 /* We are somewhere in the middle of EBB. */
6181 get_live_on_other_edges (EDGE_PRED (curr_bb, 0)->src,
6182 curr_bb, &temp_bitmap);
6183 to_process = &temp_bitmap;
6185 head_p = true;
6186 EXECUTE_IF_SET_IN_BITMAP (to_process, 0, j, bi)
6188 if ((int) j >= lra_constraint_new_regno_start)
6189 break;
6190 if (((int) j < FIRST_PSEUDO_REGISTER || reg_renumber[j] >= 0)
6191 && usage_insns[j].check == curr_usage_insns_check
6192 && (next_usage_insns = usage_insns[j].insns) != NULL_RTX)
6194 if (need_for_split_p (potential_reload_hard_regs, j))
6196 if (lra_dump_file != NULL && head_p)
6198 fprintf (lra_dump_file,
6199 " ----------------------------------\n");
6200 head_p = false;
6202 if (split_reg (false, j, bb_note (curr_bb),
6203 next_usage_insns))
6204 change_p = true;
6206 usage_insns[j].check = 0;
6211 return change_p;
6214 /* This value affects EBB forming. If probability of edge from EBB to
6215 a BB is not greater than the following value, we don't add the BB
6216 to EBB. */
6217 #define EBB_PROBABILITY_CUTOFF \
6218 ((REG_BR_PROB_BASE * LRA_INHERITANCE_EBB_PROBABILITY_CUTOFF) / 100)
6220 /* Current number of inheritance/split iteration. */
6221 int lra_inheritance_iter;
6223 /* Entry function for inheritance/split pass. */
6224 void
6225 lra_inheritance (void)
6227 int i;
6228 basic_block bb, start_bb;
6229 edge e;
6231 lra_inheritance_iter++;
6232 if (lra_inheritance_iter > LRA_MAX_INHERITANCE_PASSES)
6233 return;
6234 timevar_push (TV_LRA_INHERITANCE);
6235 if (lra_dump_file != NULL)
6236 fprintf (lra_dump_file, "\n********** Inheritance #%d: **********\n\n",
6237 lra_inheritance_iter);
6238 curr_usage_insns_check = 0;
6239 usage_insns = XNEWVEC (struct usage_insns, lra_constraint_new_regno_start);
6240 for (i = 0; i < lra_constraint_new_regno_start; i++)
6241 usage_insns[i].check = 0;
6242 bitmap_initialize (&check_only_regs, &reg_obstack);
6243 bitmap_initialize (&invalid_invariant_regs, &reg_obstack);
6244 bitmap_initialize (&live_regs, &reg_obstack);
6245 bitmap_initialize (&temp_bitmap, &reg_obstack);
6246 bitmap_initialize (&ebb_global_regs, &reg_obstack);
6247 FOR_EACH_BB_FN (bb, cfun)
6249 start_bb = bb;
6250 if (lra_dump_file != NULL)
6251 fprintf (lra_dump_file, "EBB");
6252 /* Form a EBB starting with BB. */
6253 bitmap_clear (&ebb_global_regs);
6254 bitmap_ior_into (&ebb_global_regs, df_get_live_in (bb));
6255 for (;;)
6257 if (lra_dump_file != NULL)
6258 fprintf (lra_dump_file, " %d", bb->index);
6259 if (bb->next_bb == EXIT_BLOCK_PTR_FOR_FN (cfun)
6260 || LABEL_P (BB_HEAD (bb->next_bb)))
6261 break;
6262 e = find_fallthru_edge (bb->succs);
6263 if (! e)
6264 break;
6265 if (e->probability < EBB_PROBABILITY_CUTOFF)
6266 break;
6267 bb = bb->next_bb;
6269 bitmap_ior_into (&ebb_global_regs, df_get_live_out (bb));
6270 if (lra_dump_file != NULL)
6271 fprintf (lra_dump_file, "\n");
6272 if (inherit_in_ebb (BB_HEAD (start_bb), BB_END (bb)))
6273 /* Remember that the EBB head and tail can change in
6274 inherit_in_ebb. */
6275 update_ebb_live_info (BB_HEAD (start_bb), BB_END (bb));
6277 bitmap_clear (&ebb_global_regs);
6278 bitmap_clear (&temp_bitmap);
6279 bitmap_clear (&live_regs);
6280 bitmap_clear (&invalid_invariant_regs);
6281 bitmap_clear (&check_only_regs);
6282 free (usage_insns);
6284 timevar_pop (TV_LRA_INHERITANCE);
6289 /* This page contains code to undo failed inheritance/split
6290 transformations. */
6292 /* Current number of iteration undoing inheritance/split. */
6293 int lra_undo_inheritance_iter;
6295 /* Fix BB live info LIVE after removing pseudos created on pass doing
6296 inheritance/split which are REMOVED_PSEUDOS. */
6297 static void
6298 fix_bb_live_info (bitmap live, bitmap removed_pseudos)
6300 unsigned int regno;
6301 bitmap_iterator bi;
6303 EXECUTE_IF_SET_IN_BITMAP (removed_pseudos, 0, regno, bi)
6304 if (bitmap_clear_bit (live, regno)
6305 && REG_P (lra_reg_info[regno].restore_rtx))
6306 bitmap_set_bit (live, REGNO (lra_reg_info[regno].restore_rtx));
6309 /* Return regno of the (subreg of) REG. Otherwise, return a negative
6310 number. */
6311 static int
6312 get_regno (rtx reg)
6314 if (GET_CODE (reg) == SUBREG)
6315 reg = SUBREG_REG (reg);
6316 if (REG_P (reg))
6317 return REGNO (reg);
6318 return -1;
6321 /* Delete a move INSN with destination reg DREGNO and a previous
6322 clobber insn with the same regno. The inheritance/split code can
6323 generate moves with preceding clobber and when we delete such moves
6324 we should delete the clobber insn too to keep the correct life
6325 info. */
6326 static void
6327 delete_move_and_clobber (rtx_insn *insn, int dregno)
6329 rtx_insn *prev_insn = PREV_INSN (insn);
6331 lra_set_insn_deleted (insn);
6332 lra_assert (dregno >= 0);
6333 if (prev_insn != NULL && NONDEBUG_INSN_P (prev_insn)
6334 && GET_CODE (PATTERN (prev_insn)) == CLOBBER
6335 && dregno == get_regno (XEXP (PATTERN (prev_insn), 0)))
6336 lra_set_insn_deleted (prev_insn);
6339 /* Remove inheritance/split pseudos which are in REMOVE_PSEUDOS and
6340 return true if we did any change. The undo transformations for
6341 inheritance looks like
6342 i <- i2
6343 p <- i => p <- i2
6344 or removing
6345 p <- i, i <- p, and i <- i3
6346 where p is original pseudo from which inheritance pseudo i was
6347 created, i and i3 are removed inheritance pseudos, i2 is another
6348 not removed inheritance pseudo. All split pseudos or other
6349 occurrences of removed inheritance pseudos are changed on the
6350 corresponding original pseudos.
6352 The function also schedules insns changed and created during
6353 inheritance/split pass for processing by the subsequent constraint
6354 pass. */
6355 static bool
6356 remove_inheritance_pseudos (bitmap remove_pseudos)
6358 basic_block bb;
6359 int regno, sregno, prev_sregno, dregno;
6360 rtx restore_rtx;
6361 rtx set, prev_set;
6362 rtx_insn *prev_insn;
6363 bool change_p, done_p;
6365 change_p = ! bitmap_empty_p (remove_pseudos);
6366 /* We can not finish the function right away if CHANGE_P is true
6367 because we need to marks insns affected by previous
6368 inheritance/split pass for processing by the subsequent
6369 constraint pass. */
6370 FOR_EACH_BB_FN (bb, cfun)
6372 fix_bb_live_info (df_get_live_in (bb), remove_pseudos);
6373 fix_bb_live_info (df_get_live_out (bb), remove_pseudos);
6374 FOR_BB_INSNS_REVERSE (bb, curr_insn)
6376 if (! INSN_P (curr_insn))
6377 continue;
6378 done_p = false;
6379 sregno = dregno = -1;
6380 if (change_p && NONDEBUG_INSN_P (curr_insn)
6381 && (set = single_set (curr_insn)) != NULL_RTX)
6383 dregno = get_regno (SET_DEST (set));
6384 sregno = get_regno (SET_SRC (set));
6387 if (sregno >= 0 && dregno >= 0)
6389 if (bitmap_bit_p (remove_pseudos, dregno)
6390 && ! REG_P (lra_reg_info[dregno].restore_rtx))
6392 /* invariant inheritance pseudo <- original pseudo */
6393 if (lra_dump_file != NULL)
6395 fprintf (lra_dump_file, " Removing invariant inheritance:\n");
6396 dump_insn_slim (lra_dump_file, curr_insn);
6397 fprintf (lra_dump_file, "\n");
6399 delete_move_and_clobber (curr_insn, dregno);
6400 done_p = true;
6402 else if (bitmap_bit_p (remove_pseudos, sregno)
6403 && ! REG_P (lra_reg_info[sregno].restore_rtx))
6405 /* reload pseudo <- invariant inheritance pseudo */
6406 start_sequence ();
6407 /* We can not just change the source. It might be
6408 an insn different from the move. */
6409 emit_insn (lra_reg_info[sregno].restore_rtx);
6410 rtx_insn *new_insns = get_insns ();
6411 end_sequence ();
6412 lra_assert (single_set (new_insns) != NULL
6413 && SET_DEST (set) == SET_DEST (single_set (new_insns)));
6414 lra_process_new_insns (curr_insn, NULL, new_insns,
6415 "Changing reload<-invariant inheritance");
6416 delete_move_and_clobber (curr_insn, dregno);
6417 done_p = true;
6419 else if ((bitmap_bit_p (remove_pseudos, sregno)
6420 && (get_regno (lra_reg_info[sregno].restore_rtx) == dregno
6421 || (bitmap_bit_p (remove_pseudos, dregno)
6422 && get_regno (lra_reg_info[sregno].restore_rtx) >= 0
6423 && (get_regno (lra_reg_info[sregno].restore_rtx)
6424 == get_regno (lra_reg_info[dregno].restore_rtx)))))
6425 || (bitmap_bit_p (remove_pseudos, dregno)
6426 && get_regno (lra_reg_info[dregno].restore_rtx) == sregno))
6427 /* One of the following cases:
6428 original <- removed inheritance pseudo
6429 removed inherit pseudo <- another removed inherit pseudo
6430 removed inherit pseudo <- original pseudo
6432 removed_split_pseudo <- original_reg
6433 original_reg <- removed_split_pseudo */
6435 if (lra_dump_file != NULL)
6437 fprintf (lra_dump_file, " Removing %s:\n",
6438 bitmap_bit_p (&lra_split_regs, sregno)
6439 || bitmap_bit_p (&lra_split_regs, dregno)
6440 ? "split" : "inheritance");
6441 dump_insn_slim (lra_dump_file, curr_insn);
6443 delete_move_and_clobber (curr_insn, dregno);
6444 done_p = true;
6446 else if (bitmap_bit_p (remove_pseudos, sregno)
6447 && bitmap_bit_p (&lra_inheritance_pseudos, sregno))
6449 /* Search the following pattern:
6450 inherit_or_split_pseudo1 <- inherit_or_split_pseudo2
6451 original_pseudo <- inherit_or_split_pseudo1
6452 where the 2nd insn is the current insn and
6453 inherit_or_split_pseudo2 is not removed. If it is found,
6454 change the current insn onto:
6455 original_pseudo <- inherit_or_split_pseudo2. */
6456 for (prev_insn = PREV_INSN (curr_insn);
6457 prev_insn != NULL_RTX && ! NONDEBUG_INSN_P (prev_insn);
6458 prev_insn = PREV_INSN (prev_insn))
6460 if (prev_insn != NULL_RTX && BLOCK_FOR_INSN (prev_insn) == bb
6461 && (prev_set = single_set (prev_insn)) != NULL_RTX
6462 /* There should be no subregs in insn we are
6463 searching because only the original reg might
6464 be in subreg when we changed the mode of
6465 load/store for splitting. */
6466 && REG_P (SET_DEST (prev_set))
6467 && REG_P (SET_SRC (prev_set))
6468 && (int) REGNO (SET_DEST (prev_set)) == sregno
6469 && ((prev_sregno = REGNO (SET_SRC (prev_set)))
6470 >= FIRST_PSEUDO_REGISTER)
6471 && (lra_reg_info[prev_sregno].restore_rtx == NULL_RTX
6473 /* As we consider chain of inheritance or
6474 splitting described in above comment we should
6475 check that sregno and prev_sregno were
6476 inheritance/split pseudos created from the
6477 same original regno. */
6478 (get_regno (lra_reg_info[sregno].restore_rtx) >= 0
6479 && (get_regno (lra_reg_info[sregno].restore_rtx)
6480 == get_regno (lra_reg_info[prev_sregno].restore_rtx))))
6481 && ! bitmap_bit_p (remove_pseudos, prev_sregno))
6483 lra_assert (GET_MODE (SET_SRC (prev_set))
6484 == GET_MODE (regno_reg_rtx[sregno]));
6485 if (GET_CODE (SET_SRC (set)) == SUBREG)
6486 SUBREG_REG (SET_SRC (set)) = SET_SRC (prev_set);
6487 else
6488 SET_SRC (set) = SET_SRC (prev_set);
6489 /* As we are finishing with processing the insn
6490 here, check the destination too as it might
6491 inheritance pseudo for another pseudo. */
6492 if (bitmap_bit_p (remove_pseudos, dregno)
6493 && bitmap_bit_p (&lra_inheritance_pseudos, dregno)
6494 && (restore_rtx
6495 = lra_reg_info[dregno].restore_rtx) != NULL_RTX)
6497 if (GET_CODE (SET_DEST (set)) == SUBREG)
6498 SUBREG_REG (SET_DEST (set)) = restore_rtx;
6499 else
6500 SET_DEST (set) = restore_rtx;
6502 lra_push_insn_and_update_insn_regno_info (curr_insn);
6503 lra_set_used_insn_alternative_by_uid
6504 (INSN_UID (curr_insn), -1);
6505 done_p = true;
6506 if (lra_dump_file != NULL)
6508 fprintf (lra_dump_file, " Change reload insn:\n");
6509 dump_insn_slim (lra_dump_file, curr_insn);
6514 if (! done_p)
6516 struct lra_insn_reg *reg;
6517 bool restored_regs_p = false;
6518 bool kept_regs_p = false;
6520 curr_id = lra_get_insn_recog_data (curr_insn);
6521 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
6523 regno = reg->regno;
6524 restore_rtx = lra_reg_info[regno].restore_rtx;
6525 if (restore_rtx != NULL_RTX)
6527 if (change_p && bitmap_bit_p (remove_pseudos, regno))
6529 lra_substitute_pseudo_within_insn
6530 (curr_insn, regno, restore_rtx, false);
6531 restored_regs_p = true;
6533 else
6534 kept_regs_p = true;
6537 if (NONDEBUG_INSN_P (curr_insn) && kept_regs_p)
6539 /* The instruction has changed since the previous
6540 constraints pass. */
6541 lra_push_insn_and_update_insn_regno_info (curr_insn);
6542 lra_set_used_insn_alternative_by_uid
6543 (INSN_UID (curr_insn), -1);
6545 else if (restored_regs_p)
6546 /* The instruction has been restored to the form that
6547 it had during the previous constraints pass. */
6548 lra_update_insn_regno_info (curr_insn);
6549 if (restored_regs_p && lra_dump_file != NULL)
6551 fprintf (lra_dump_file, " Insn after restoring regs:\n");
6552 dump_insn_slim (lra_dump_file, curr_insn);
6557 return change_p;
6560 /* If optional reload pseudos failed to get a hard register or was not
6561 inherited, it is better to remove optional reloads. We do this
6562 transformation after undoing inheritance to figure out necessity to
6563 remove optional reloads easier. Return true if we do any
6564 change. */
6565 static bool
6566 undo_optional_reloads (void)
6568 bool change_p, keep_p;
6569 unsigned int regno, uid;
6570 bitmap_iterator bi, bi2;
6571 rtx_insn *insn;
6572 rtx set, src, dest;
6573 bitmap_head removed_optional_reload_pseudos, insn_bitmap;
6575 bitmap_initialize (&removed_optional_reload_pseudos, &reg_obstack);
6576 bitmap_copy (&removed_optional_reload_pseudos, &lra_optional_reload_pseudos);
6577 EXECUTE_IF_SET_IN_BITMAP (&lra_optional_reload_pseudos, 0, regno, bi)
6579 keep_p = false;
6580 /* Keep optional reloads from previous subpasses. */
6581 if (lra_reg_info[regno].restore_rtx == NULL_RTX
6582 /* If the original pseudo changed its allocation, just
6583 removing the optional pseudo is dangerous as the original
6584 pseudo will have longer live range. */
6585 || reg_renumber[REGNO (lra_reg_info[regno].restore_rtx)] >= 0)
6586 keep_p = true;
6587 else if (reg_renumber[regno] >= 0)
6588 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi2)
6590 insn = lra_insn_recog_data[uid]->insn;
6591 if ((set = single_set (insn)) == NULL_RTX)
6592 continue;
6593 src = SET_SRC (set);
6594 dest = SET_DEST (set);
6595 if (! REG_P (src) || ! REG_P (dest))
6596 continue;
6597 if (REGNO (dest) == regno
6598 /* Ignore insn for optional reloads itself. */
6599 && REGNO (lra_reg_info[regno].restore_rtx) != REGNO (src)
6600 /* Check only inheritance on last inheritance pass. */
6601 && (int) REGNO (src) >= new_regno_start
6602 /* Check that the optional reload was inherited. */
6603 && bitmap_bit_p (&lra_inheritance_pseudos, REGNO (src)))
6605 keep_p = true;
6606 break;
6609 if (keep_p)
6611 bitmap_clear_bit (&removed_optional_reload_pseudos, regno);
6612 if (lra_dump_file != NULL)
6613 fprintf (lra_dump_file, "Keep optional reload reg %d\n", regno);
6616 change_p = ! bitmap_empty_p (&removed_optional_reload_pseudos);
6617 bitmap_initialize (&insn_bitmap, &reg_obstack);
6618 EXECUTE_IF_SET_IN_BITMAP (&removed_optional_reload_pseudos, 0, regno, bi)
6620 if (lra_dump_file != NULL)
6621 fprintf (lra_dump_file, "Remove optional reload reg %d\n", regno);
6622 bitmap_copy (&insn_bitmap, &lra_reg_info[regno].insn_bitmap);
6623 EXECUTE_IF_SET_IN_BITMAP (&insn_bitmap, 0, uid, bi2)
6625 insn = lra_insn_recog_data[uid]->insn;
6626 if ((set = single_set (insn)) != NULL_RTX)
6628 src = SET_SRC (set);
6629 dest = SET_DEST (set);
6630 if (REG_P (src) && REG_P (dest)
6631 && ((REGNO (src) == regno
6632 && (REGNO (lra_reg_info[regno].restore_rtx)
6633 == REGNO (dest)))
6634 || (REGNO (dest) == regno
6635 && (REGNO (lra_reg_info[regno].restore_rtx)
6636 == REGNO (src)))))
6638 if (lra_dump_file != NULL)
6640 fprintf (lra_dump_file, " Deleting move %u\n",
6641 INSN_UID (insn));
6642 dump_insn_slim (lra_dump_file, insn);
6644 delete_move_and_clobber (insn, REGNO (dest));
6645 continue;
6647 /* We should not worry about generation memory-memory
6648 moves here as if the corresponding inheritance did
6649 not work (inheritance pseudo did not get a hard reg),
6650 we remove the inheritance pseudo and the optional
6651 reload. */
6653 lra_substitute_pseudo_within_insn
6654 (insn, regno, lra_reg_info[regno].restore_rtx, false);
6655 lra_update_insn_regno_info (insn);
6656 if (lra_dump_file != NULL)
6658 fprintf (lra_dump_file,
6659 " Restoring original insn:\n");
6660 dump_insn_slim (lra_dump_file, insn);
6664 /* Clear restore_regnos. */
6665 EXECUTE_IF_SET_IN_BITMAP (&lra_optional_reload_pseudos, 0, regno, bi)
6666 lra_reg_info[regno].restore_rtx = NULL_RTX;
6667 bitmap_clear (&insn_bitmap);
6668 bitmap_clear (&removed_optional_reload_pseudos);
6669 return change_p;
6672 /* Entry function for undoing inheritance/split transformation. Return true
6673 if we did any RTL change in this pass. */
6674 bool
6675 lra_undo_inheritance (void)
6677 unsigned int regno;
6678 int hard_regno;
6679 int n_all_inherit, n_inherit, n_all_split, n_split;
6680 rtx restore_rtx;
6681 bitmap_head remove_pseudos;
6682 bitmap_iterator bi;
6683 bool change_p;
6685 lra_undo_inheritance_iter++;
6686 if (lra_undo_inheritance_iter > LRA_MAX_INHERITANCE_PASSES)
6687 return false;
6688 if (lra_dump_file != NULL)
6689 fprintf (lra_dump_file,
6690 "\n********** Undoing inheritance #%d: **********\n\n",
6691 lra_undo_inheritance_iter);
6692 bitmap_initialize (&remove_pseudos, &reg_obstack);
6693 n_inherit = n_all_inherit = 0;
6694 EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, regno, bi)
6695 if (lra_reg_info[regno].restore_rtx != NULL_RTX)
6697 n_all_inherit++;
6698 if (reg_renumber[regno] < 0
6699 /* If the original pseudo changed its allocation, just
6700 removing inheritance is dangerous as for changing
6701 allocation we used shorter live-ranges. */
6702 && (! REG_P (lra_reg_info[regno].restore_rtx)
6703 || reg_renumber[REGNO (lra_reg_info[regno].restore_rtx)] < 0))
6704 bitmap_set_bit (&remove_pseudos, regno);
6705 else
6706 n_inherit++;
6708 if (lra_dump_file != NULL && n_all_inherit != 0)
6709 fprintf (lra_dump_file, "Inherit %d out of %d (%.2f%%)\n",
6710 n_inherit, n_all_inherit,
6711 (double) n_inherit / n_all_inherit * 100);
6712 n_split = n_all_split = 0;
6713 EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, regno, bi)
6714 if ((restore_rtx = lra_reg_info[regno].restore_rtx) != NULL_RTX)
6716 int restore_regno = REGNO (restore_rtx);
6718 n_all_split++;
6719 hard_regno = (restore_regno >= FIRST_PSEUDO_REGISTER
6720 ? reg_renumber[restore_regno] : restore_regno);
6721 if (hard_regno < 0 || reg_renumber[regno] == hard_regno)
6722 bitmap_set_bit (&remove_pseudos, regno);
6723 else
6725 n_split++;
6726 if (lra_dump_file != NULL)
6727 fprintf (lra_dump_file, " Keep split r%d (orig=r%d)\n",
6728 regno, restore_regno);
6731 if (lra_dump_file != NULL && n_all_split != 0)
6732 fprintf (lra_dump_file, "Split %d out of %d (%.2f%%)\n",
6733 n_split, n_all_split,
6734 (double) n_split / n_all_split * 100);
6735 change_p = remove_inheritance_pseudos (&remove_pseudos);
6736 bitmap_clear (&remove_pseudos);
6737 /* Clear restore_regnos. */
6738 EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, regno, bi)
6739 lra_reg_info[regno].restore_rtx = NULL_RTX;
6740 EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, regno, bi)
6741 lra_reg_info[regno].restore_rtx = NULL_RTX;
6742 change_p = undo_optional_reloads () || change_p;
6743 return change_p;