pdp11.md (movmemhi): Adjust operand numbers to match_scratch operand is highest.
[official-gcc.git] / gcc / lra-constraints.c
blobe0ce8072ab053cb2d527e28ff5b8b1c26d275eb8
1 /* Code for RTL transformations to satisfy insn constraints.
2 Copyright (C) 2010-2017 Free Software Foundation, Inc.
3 Contributed by Vladimir Makarov <vmakarov@redhat.com>.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
10 version.
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
22 /* This file contains code for 3 passes: constraint pass,
23 inheritance/split pass, and pass for undoing failed inheritance and
24 split.
26 The major goal of constraint pass is to transform RTL to satisfy
27 insn and address constraints by:
28 o choosing insn alternatives;
29 o generating *reload insns* (or reloads in brief) and *reload
30 pseudos* which will get necessary hard registers later;
31 o substituting pseudos with equivalent values and removing the
32 instructions that initialized those pseudos.
34 The constraint pass has biggest and most complicated code in LRA.
35 There are a lot of important details like:
36 o reuse of input reload pseudos to simplify reload pseudo
37 allocations;
38 o some heuristics to choose insn alternative to improve the
39 inheritance;
40 o early clobbers etc.
42 The pass is mimicking former reload pass in alternative choosing
43 because the reload pass is oriented to current machine description
44 model. It might be changed if the machine description model is
45 changed.
47 There is special code for preventing all LRA and this pass cycling
48 in case of bugs.
50 On the first iteration of the pass we process every instruction and
51 choose an alternative for each one. On subsequent iterations we try
52 to avoid reprocessing instructions if we can be sure that the old
53 choice is still valid.
55 The inheritance/spilt pass is to transform code to achieve
56 ineheritance and live range splitting. It is done on backward
57 traversal of EBBs.
59 The inheritance optimization goal is to reuse values in hard
60 registers. There is analogous optimization in old reload pass. The
61 inheritance is achieved by following transformation:
63 reload_p1 <- p reload_p1 <- p
64 ... new_p <- reload_p1
65 ... => ...
66 reload_p2 <- p reload_p2 <- new_p
68 where p is spilled and not changed between the insns. Reload_p1 is
69 also called *original pseudo* and new_p is called *inheritance
70 pseudo*.
72 The subsequent assignment pass will try to assign the same (or
73 another if it is not possible) hard register to new_p as to
74 reload_p1 or reload_p2.
76 If the assignment pass fails to assign a hard register to new_p,
77 this file will undo the inheritance and restore the original code.
78 This is because implementing the above sequence with a spilled
79 new_p would make the code much worse. The inheritance is done in
80 EBB scope. The above is just a simplified example to get an idea
81 of the inheritance as the inheritance is also done for non-reload
82 insns.
84 Splitting (transformation) is also done in EBB scope on the same
85 pass as the inheritance:
87 r <- ... or ... <- r r <- ... or ... <- r
88 ... s <- r (new insn -- save)
89 ... =>
90 ... r <- s (new insn -- restore)
91 ... <- r ... <- r
93 The *split pseudo* s is assigned to the hard register of the
94 original pseudo or hard register r.
96 Splitting is done:
97 o In EBBs with high register pressure for global pseudos (living
98 in at least 2 BBs) and assigned to hard registers when there
99 are more one reloads needing the hard registers;
100 o for pseudos needing save/restore code around calls.
102 If the split pseudo still has the same hard register as the
103 original pseudo after the subsequent assignment pass or the
104 original pseudo was split, the opposite transformation is done on
105 the same pass for undoing inheritance. */
107 #undef REG_OK_STRICT
109 #include "config.h"
110 #include "system.h"
111 #include "coretypes.h"
112 #include "backend.h"
113 #include "target.h"
114 #include "rtl.h"
115 #include "tree.h"
116 #include "predict.h"
117 #include "df.h"
118 #include "memmodel.h"
119 #include "tm_p.h"
120 #include "expmed.h"
121 #include "optabs.h"
122 #include "regs.h"
123 #include "ira.h"
124 #include "recog.h"
125 #include "output.h"
126 #include "addresses.h"
127 #include "expr.h"
128 #include "cfgrtl.h"
129 #include "rtl-error.h"
130 #include "params.h"
131 #include "lra.h"
132 #include "lra-int.h"
133 #include "print-rtl.h"
135 /* Value of LRA_CURR_RELOAD_NUM at the beginning of BB of the current
136 insn. Remember that LRA_CURR_RELOAD_NUM is the number of emitted
137 reload insns. */
138 static int bb_reload_num;
140 /* The current insn being processed and corresponding its single set
141 (NULL otherwise), its data (basic block, the insn data, the insn
142 static data, and the mode of each operand). */
143 static rtx_insn *curr_insn;
144 static rtx curr_insn_set;
145 static basic_block curr_bb;
146 static lra_insn_recog_data_t curr_id;
147 static struct lra_static_insn_data *curr_static_id;
148 static machine_mode curr_operand_mode[MAX_RECOG_OPERANDS];
149 /* Mode of the register substituted by its equivalence with VOIDmode
150 (e.g. constant) and whose subreg is given operand of the current
151 insn. VOIDmode in all other cases. */
152 static machine_mode original_subreg_reg_mode[MAX_RECOG_OPERANDS];
156 /* Start numbers for new registers and insns at the current constraints
157 pass start. */
158 static int new_regno_start;
159 static int new_insn_uid_start;
161 /* If LOC is nonnull, strip any outer subreg from it. */
162 static inline rtx *
163 strip_subreg (rtx *loc)
165 return loc && GET_CODE (*loc) == SUBREG ? &SUBREG_REG (*loc) : loc;
168 /* Return hard regno of REGNO or if it is was not assigned to a hard
169 register, use a hard register from its allocno class. */
170 static int
171 get_try_hard_regno (int regno)
173 int hard_regno;
174 enum reg_class rclass;
176 if ((hard_regno = regno) >= FIRST_PSEUDO_REGISTER)
177 hard_regno = lra_get_regno_hard_regno (regno);
178 if (hard_regno >= 0)
179 return hard_regno;
180 rclass = lra_get_allocno_class (regno);
181 if (rclass == NO_REGS)
182 return -1;
183 return ira_class_hard_regs[rclass][0];
186 /* Return the hard regno of X after removing its subreg. If X is not
187 a register or a subreg of a register, return -1. If X is a pseudo,
188 use its assignment. If FINAL_P return the final hard regno which will
189 be after elimination. */
190 static int
191 get_hard_regno (rtx x, bool final_p)
193 rtx reg;
194 int hard_regno;
196 reg = x;
197 if (SUBREG_P (x))
198 reg = SUBREG_REG (x);
199 if (! REG_P (reg))
200 return -1;
201 if (! HARD_REGISTER_NUM_P (hard_regno = REGNO (reg)))
202 hard_regno = lra_get_regno_hard_regno (hard_regno);
203 if (hard_regno < 0)
204 return -1;
205 if (final_p)
206 hard_regno = lra_get_elimination_hard_regno (hard_regno);
207 if (SUBREG_P (x))
208 hard_regno += subreg_regno_offset (hard_regno, GET_MODE (reg),
209 SUBREG_BYTE (x), GET_MODE (x));
210 return hard_regno;
213 /* If REGNO is a hard register or has been allocated a hard register,
214 return the class of that register. If REGNO is a reload pseudo
215 created by the current constraints pass, return its allocno class.
216 Return NO_REGS otherwise. */
217 static enum reg_class
218 get_reg_class (int regno)
220 int hard_regno;
222 if (! HARD_REGISTER_NUM_P (hard_regno = regno))
223 hard_regno = lra_get_regno_hard_regno (regno);
224 if (hard_regno >= 0)
226 hard_regno = lra_get_elimination_hard_regno (hard_regno);
227 return REGNO_REG_CLASS (hard_regno);
229 if (regno >= new_regno_start)
230 return lra_get_allocno_class (regno);
231 return NO_REGS;
234 /* Return true if REG satisfies (or will satisfy) reg class constraint
235 CL. Use elimination first if REG is a hard register. If REG is a
236 reload pseudo created by this constraints pass, assume that it will
237 be allocated a hard register from its allocno class, but allow that
238 class to be narrowed to CL if it is currently a superset of CL.
240 If NEW_CLASS is nonnull, set *NEW_CLASS to the new allocno class of
241 REGNO (reg), or NO_REGS if no change in its class was needed. */
242 static bool
243 in_class_p (rtx reg, enum reg_class cl, enum reg_class *new_class)
245 enum reg_class rclass, common_class;
246 machine_mode reg_mode;
247 int class_size, hard_regno, nregs, i, j;
248 int regno = REGNO (reg);
250 if (new_class != NULL)
251 *new_class = NO_REGS;
252 if (regno < FIRST_PSEUDO_REGISTER)
254 rtx final_reg = reg;
255 rtx *final_loc = &final_reg;
257 lra_eliminate_reg_if_possible (final_loc);
258 return TEST_HARD_REG_BIT (reg_class_contents[cl], REGNO (*final_loc));
260 reg_mode = GET_MODE (reg);
261 rclass = get_reg_class (regno);
262 if (regno < new_regno_start
263 /* Do not allow the constraints for reload instructions to
264 influence the classes of new pseudos. These reloads are
265 typically moves that have many alternatives, and restricting
266 reload pseudos for one alternative may lead to situations
267 where other reload pseudos are no longer allocatable. */
268 || (INSN_UID (curr_insn) >= new_insn_uid_start
269 && curr_insn_set != NULL
270 && ((OBJECT_P (SET_SRC (curr_insn_set))
271 && ! CONSTANT_P (SET_SRC (curr_insn_set)))
272 || (GET_CODE (SET_SRC (curr_insn_set)) == SUBREG
273 && OBJECT_P (SUBREG_REG (SET_SRC (curr_insn_set)))
274 && ! CONSTANT_P (SUBREG_REG (SET_SRC (curr_insn_set)))))))
275 /* When we don't know what class will be used finally for reload
276 pseudos, we use ALL_REGS. */
277 return ((regno >= new_regno_start && rclass == ALL_REGS)
278 || (rclass != NO_REGS && ira_class_subset_p[rclass][cl]
279 && ! hard_reg_set_subset_p (reg_class_contents[cl],
280 lra_no_alloc_regs)));
281 else
283 common_class = ira_reg_class_subset[rclass][cl];
284 if (new_class != NULL)
285 *new_class = common_class;
286 if (hard_reg_set_subset_p (reg_class_contents[common_class],
287 lra_no_alloc_regs))
288 return false;
289 /* Check that there are enough allocatable regs. */
290 class_size = ira_class_hard_regs_num[common_class];
291 for (i = 0; i < class_size; i++)
293 hard_regno = ira_class_hard_regs[common_class][i];
294 nregs = hard_regno_nregs[hard_regno][reg_mode];
295 if (nregs == 1)
296 return true;
297 for (j = 0; j < nregs; j++)
298 if (TEST_HARD_REG_BIT (lra_no_alloc_regs, hard_regno + j)
299 || ! TEST_HARD_REG_BIT (reg_class_contents[common_class],
300 hard_regno + j))
301 break;
302 if (j >= nregs)
303 return true;
305 return false;
309 /* Return true if REGNO satisfies a memory constraint. */
310 static bool
311 in_mem_p (int regno)
313 return get_reg_class (regno) == NO_REGS;
316 /* Return 1 if ADDR is a valid memory address for mode MODE in address
317 space AS, and check that each pseudo has the proper kind of hard
318 reg. */
319 static int
320 valid_address_p (machine_mode mode ATTRIBUTE_UNUSED,
321 rtx addr, addr_space_t as)
323 #ifdef GO_IF_LEGITIMATE_ADDRESS
324 lra_assert (ADDR_SPACE_GENERIC_P (as));
325 GO_IF_LEGITIMATE_ADDRESS (mode, addr, win);
326 return 0;
328 win:
329 return 1;
330 #else
331 return targetm.addr_space.legitimate_address_p (mode, addr, 0, as);
332 #endif
335 namespace {
336 /* Temporarily eliminates registers in an address (for the lifetime of
337 the object). */
338 class address_eliminator {
339 public:
340 address_eliminator (struct address_info *ad);
341 ~address_eliminator ();
343 private:
344 struct address_info *m_ad;
345 rtx *m_base_loc;
346 rtx m_base_reg;
347 rtx *m_index_loc;
348 rtx m_index_reg;
352 address_eliminator::address_eliminator (struct address_info *ad)
353 : m_ad (ad),
354 m_base_loc (strip_subreg (ad->base_term)),
355 m_base_reg (NULL_RTX),
356 m_index_loc (strip_subreg (ad->index_term)),
357 m_index_reg (NULL_RTX)
359 if (m_base_loc != NULL)
361 m_base_reg = *m_base_loc;
362 lra_eliminate_reg_if_possible (m_base_loc);
363 if (m_ad->base_term2 != NULL)
364 *m_ad->base_term2 = *m_ad->base_term;
366 if (m_index_loc != NULL)
368 m_index_reg = *m_index_loc;
369 lra_eliminate_reg_if_possible (m_index_loc);
373 address_eliminator::~address_eliminator ()
375 if (m_base_loc && *m_base_loc != m_base_reg)
377 *m_base_loc = m_base_reg;
378 if (m_ad->base_term2 != NULL)
379 *m_ad->base_term2 = *m_ad->base_term;
381 if (m_index_loc && *m_index_loc != m_index_reg)
382 *m_index_loc = m_index_reg;
385 /* Return true if the eliminated form of AD is a legitimate target address. */
386 static bool
387 valid_address_p (struct address_info *ad)
389 address_eliminator eliminator (ad);
390 return valid_address_p (ad->mode, *ad->outer, ad->as);
393 /* Return true if the eliminated form of memory reference OP satisfies
394 extra (special) memory constraint CONSTRAINT. */
395 static bool
396 satisfies_memory_constraint_p (rtx op, enum constraint_num constraint)
398 struct address_info ad;
400 decompose_mem_address (&ad, op);
401 address_eliminator eliminator (&ad);
402 return constraint_satisfied_p (op, constraint);
405 /* Return true if the eliminated form of address AD satisfies extra
406 address constraint CONSTRAINT. */
407 static bool
408 satisfies_address_constraint_p (struct address_info *ad,
409 enum constraint_num constraint)
411 address_eliminator eliminator (ad);
412 return constraint_satisfied_p (*ad->outer, constraint);
415 /* Return true if the eliminated form of address OP satisfies extra
416 address constraint CONSTRAINT. */
417 static bool
418 satisfies_address_constraint_p (rtx op, enum constraint_num constraint)
420 struct address_info ad;
422 decompose_lea_address (&ad, &op);
423 return satisfies_address_constraint_p (&ad, constraint);
426 /* Initiate equivalences for LRA. As we keep original equivalences
427 before any elimination, we need to make copies otherwise any change
428 in insns might change the equivalences. */
429 void
430 lra_init_equiv (void)
432 ira_expand_reg_equiv ();
433 for (int i = FIRST_PSEUDO_REGISTER; i < max_reg_num (); i++)
435 rtx res;
437 if ((res = ira_reg_equiv[i].memory) != NULL_RTX)
438 ira_reg_equiv[i].memory = copy_rtx (res);
439 if ((res = ira_reg_equiv[i].invariant) != NULL_RTX)
440 ira_reg_equiv[i].invariant = copy_rtx (res);
444 static rtx loc_equivalence_callback (rtx, const_rtx, void *);
446 /* Update equivalence for REGNO. We need to this as the equivalence
447 might contain other pseudos which are changed by their
448 equivalences. */
449 static void
450 update_equiv (int regno)
452 rtx x;
454 if ((x = ira_reg_equiv[regno].memory) != NULL_RTX)
455 ira_reg_equiv[regno].memory
456 = simplify_replace_fn_rtx (x, NULL_RTX, loc_equivalence_callback,
457 NULL_RTX);
458 if ((x = ira_reg_equiv[regno].invariant) != NULL_RTX)
459 ira_reg_equiv[regno].invariant
460 = simplify_replace_fn_rtx (x, NULL_RTX, loc_equivalence_callback,
461 NULL_RTX);
464 /* If we have decided to substitute X with another value, return that
465 value, otherwise return X. */
466 static rtx
467 get_equiv (rtx x)
469 int regno;
470 rtx res;
472 if (! REG_P (x) || (regno = REGNO (x)) < FIRST_PSEUDO_REGISTER
473 || ! ira_reg_equiv[regno].defined_p
474 || ! ira_reg_equiv[regno].profitable_p
475 || lra_get_regno_hard_regno (regno) >= 0)
476 return x;
477 if ((res = ira_reg_equiv[regno].memory) != NULL_RTX)
479 if (targetm.cannot_substitute_mem_equiv_p (res))
480 return x;
481 return res;
483 if ((res = ira_reg_equiv[regno].constant) != NULL_RTX)
484 return res;
485 if ((res = ira_reg_equiv[regno].invariant) != NULL_RTX)
486 return res;
487 gcc_unreachable ();
490 /* If we have decided to substitute X with the equivalent value,
491 return that value after elimination for INSN, otherwise return
492 X. */
493 static rtx
494 get_equiv_with_elimination (rtx x, rtx_insn *insn)
496 rtx res = get_equiv (x);
498 if (x == res || CONSTANT_P (res))
499 return res;
500 return lra_eliminate_regs_1 (insn, res, GET_MODE (res),
501 false, false, 0, true);
504 /* Set up curr_operand_mode. */
505 static void
506 init_curr_operand_mode (void)
508 int nop = curr_static_id->n_operands;
509 for (int i = 0; i < nop; i++)
511 machine_mode mode = GET_MODE (*curr_id->operand_loc[i]);
512 if (mode == VOIDmode)
514 /* The .md mode for address operands is the mode of the
515 addressed value rather than the mode of the address itself. */
516 if (curr_id->icode >= 0 && curr_static_id->operand[i].is_address)
517 mode = Pmode;
518 else
519 mode = curr_static_id->operand[i].mode;
521 curr_operand_mode[i] = mode;
527 /* The page contains code to reuse input reloads. */
529 /* Structure describes input reload of the current insns. */
530 struct input_reload
532 /* True for input reload of matched operands. */
533 bool match_p;
534 /* Reloaded value. */
535 rtx input;
536 /* Reload pseudo used. */
537 rtx reg;
540 /* The number of elements in the following array. */
541 static int curr_insn_input_reloads_num;
542 /* Array containing info about input reloads. It is used to find the
543 same input reload and reuse the reload pseudo in this case. */
544 static struct input_reload curr_insn_input_reloads[LRA_MAX_INSN_RELOADS];
546 /* Initiate data concerning reuse of input reloads for the current
547 insn. */
548 static void
549 init_curr_insn_input_reloads (void)
551 curr_insn_input_reloads_num = 0;
554 /* Create a new pseudo using MODE, RCLASS, ORIGINAL or reuse already
555 created input reload pseudo (only if TYPE is not OP_OUT). Don't
556 reuse pseudo if IN_SUBREG_P is true and the reused pseudo should be
557 wrapped up in SUBREG. The result pseudo is returned through
558 RESULT_REG. Return TRUE if we created a new pseudo, FALSE if we
559 reused the already created input reload pseudo. Use TITLE to
560 describe new registers for debug purposes. */
561 static bool
562 get_reload_reg (enum op_type type, machine_mode mode, rtx original,
563 enum reg_class rclass, bool in_subreg_p,
564 const char *title, rtx *result_reg)
566 int i, regno;
567 enum reg_class new_class;
568 bool unique_p = false;
570 if (type == OP_OUT)
572 *result_reg
573 = lra_create_new_reg_with_unique_value (mode, original, rclass, title);
574 return true;
576 /* Prevent reuse value of expression with side effects,
577 e.g. volatile memory. */
578 if (! side_effects_p (original))
579 for (i = 0; i < curr_insn_input_reloads_num; i++)
581 if (! curr_insn_input_reloads[i].match_p
582 && rtx_equal_p (curr_insn_input_reloads[i].input, original)
583 && in_class_p (curr_insn_input_reloads[i].reg, rclass, &new_class))
585 rtx reg = curr_insn_input_reloads[i].reg;
586 regno = REGNO (reg);
587 /* If input is equal to original and both are VOIDmode,
588 GET_MODE (reg) might be still different from mode.
589 Ensure we don't return *result_reg with wrong mode. */
590 if (GET_MODE (reg) != mode)
592 if (in_subreg_p)
593 continue;
594 if (GET_MODE_SIZE (GET_MODE (reg)) < GET_MODE_SIZE (mode))
595 continue;
596 reg = lowpart_subreg (mode, reg, GET_MODE (reg));
597 if (reg == NULL_RTX || GET_CODE (reg) != SUBREG)
598 continue;
600 *result_reg = reg;
601 if (lra_dump_file != NULL)
603 fprintf (lra_dump_file, " Reuse r%d for reload ", regno);
604 dump_value_slim (lra_dump_file, original, 1);
606 if (new_class != lra_get_allocno_class (regno))
607 lra_change_class (regno, new_class, ", change to", false);
608 if (lra_dump_file != NULL)
609 fprintf (lra_dump_file, "\n");
610 return false;
612 /* If we have an input reload with a different mode, make sure it
613 will get a different hard reg. */
614 else if (REG_P (original)
615 && REG_P (curr_insn_input_reloads[i].input)
616 && REGNO (original) == REGNO (curr_insn_input_reloads[i].input)
617 && (GET_MODE (original)
618 != GET_MODE (curr_insn_input_reloads[i].input)))
619 unique_p = true;
621 *result_reg = (unique_p
622 ? lra_create_new_reg_with_unique_value
623 : lra_create_new_reg) (mode, original, rclass, title);
624 lra_assert (curr_insn_input_reloads_num < LRA_MAX_INSN_RELOADS);
625 curr_insn_input_reloads[curr_insn_input_reloads_num].input = original;
626 curr_insn_input_reloads[curr_insn_input_reloads_num].match_p = false;
627 curr_insn_input_reloads[curr_insn_input_reloads_num++].reg = *result_reg;
628 return true;
633 /* The page contains code to extract memory address parts. */
635 /* Wrapper around REGNO_OK_FOR_INDEX_P, to allow pseudos. */
636 static inline bool
637 ok_for_index_p_nonstrict (rtx reg)
639 unsigned regno = REGNO (reg);
641 return regno >= FIRST_PSEUDO_REGISTER || REGNO_OK_FOR_INDEX_P (regno);
644 /* A version of regno_ok_for_base_p for use here, when all pseudos
645 should count as OK. Arguments as for regno_ok_for_base_p. */
646 static inline bool
647 ok_for_base_p_nonstrict (rtx reg, machine_mode mode, addr_space_t as,
648 enum rtx_code outer_code, enum rtx_code index_code)
650 unsigned regno = REGNO (reg);
652 if (regno >= FIRST_PSEUDO_REGISTER)
653 return true;
654 return ok_for_base_p_1 (regno, mode, as, outer_code, index_code);
659 /* The page contains major code to choose the current insn alternative
660 and generate reloads for it. */
662 /* Return the offset from REGNO of the least significant register
663 in (reg:MODE REGNO).
665 This function is used to tell whether two registers satisfy
666 a matching constraint. (reg:MODE1 REGNO1) matches (reg:MODE2 REGNO2) if:
668 REGNO1 + lra_constraint_offset (REGNO1, MODE1)
669 == REGNO2 + lra_constraint_offset (REGNO2, MODE2) */
671 lra_constraint_offset (int regno, machine_mode mode)
673 lra_assert (regno < FIRST_PSEUDO_REGISTER);
674 if (WORDS_BIG_ENDIAN && GET_MODE_SIZE (mode) > UNITS_PER_WORD
675 && SCALAR_INT_MODE_P (mode))
676 return hard_regno_nregs[regno][mode] - 1;
677 return 0;
680 /* Like rtx_equal_p except that it allows a REG and a SUBREG to match
681 if they are the same hard reg, and has special hacks for
682 auto-increment and auto-decrement. This is specifically intended for
683 process_alt_operands to use in determining whether two operands
684 match. X is the operand whose number is the lower of the two.
686 It is supposed that X is the output operand and Y is the input
687 operand. Y_HARD_REGNO is the final hard regno of register Y or
688 register in subreg Y as we know it now. Otherwise, it is a
689 negative value. */
690 static bool
691 operands_match_p (rtx x, rtx y, int y_hard_regno)
693 int i;
694 RTX_CODE code = GET_CODE (x);
695 const char *fmt;
697 if (x == y)
698 return true;
699 if ((code == REG || (code == SUBREG && REG_P (SUBREG_REG (x))))
700 && (REG_P (y) || (GET_CODE (y) == SUBREG && REG_P (SUBREG_REG (y)))))
702 int j;
704 i = get_hard_regno (x, false);
705 if (i < 0)
706 goto slow;
708 if ((j = y_hard_regno) < 0)
709 goto slow;
711 i += lra_constraint_offset (i, GET_MODE (x));
712 j += lra_constraint_offset (j, GET_MODE (y));
714 return i == j;
717 /* If two operands must match, because they are really a single
718 operand of an assembler insn, then two post-increments are invalid
719 because the assembler insn would increment only once. On the
720 other hand, a post-increment matches ordinary indexing if the
721 post-increment is the output operand. */
722 if (code == POST_DEC || code == POST_INC || code == POST_MODIFY)
723 return operands_match_p (XEXP (x, 0), y, y_hard_regno);
725 /* Two pre-increments are invalid because the assembler insn would
726 increment only once. On the other hand, a pre-increment matches
727 ordinary indexing if the pre-increment is the input operand. */
728 if (GET_CODE (y) == PRE_DEC || GET_CODE (y) == PRE_INC
729 || GET_CODE (y) == PRE_MODIFY)
730 return operands_match_p (x, XEXP (y, 0), -1);
732 slow:
734 if (code == REG && REG_P (y))
735 return REGNO (x) == REGNO (y);
737 if (code == REG && GET_CODE (y) == SUBREG && REG_P (SUBREG_REG (y))
738 && x == SUBREG_REG (y))
739 return true;
740 if (GET_CODE (y) == REG && code == SUBREG && REG_P (SUBREG_REG (x))
741 && SUBREG_REG (x) == y)
742 return true;
744 /* Now we have disposed of all the cases in which different rtx
745 codes can match. */
746 if (code != GET_CODE (y))
747 return false;
749 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
750 if (GET_MODE (x) != GET_MODE (y))
751 return false;
753 switch (code)
755 CASE_CONST_UNIQUE:
756 return false;
758 case LABEL_REF:
759 return label_ref_label (x) == label_ref_label (y);
760 case SYMBOL_REF:
761 return XSTR (x, 0) == XSTR (y, 0);
763 default:
764 break;
767 /* Compare the elements. If any pair of corresponding elements fail
768 to match, return false for the whole things. */
770 fmt = GET_RTX_FORMAT (code);
771 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
773 int val, j;
774 switch (fmt[i])
776 case 'w':
777 if (XWINT (x, i) != XWINT (y, i))
778 return false;
779 break;
781 case 'i':
782 if (XINT (x, i) != XINT (y, i))
783 return false;
784 break;
786 case 'e':
787 val = operands_match_p (XEXP (x, i), XEXP (y, i), -1);
788 if (val == 0)
789 return false;
790 break;
792 case '0':
793 break;
795 case 'E':
796 if (XVECLEN (x, i) != XVECLEN (y, i))
797 return false;
798 for (j = XVECLEN (x, i) - 1; j >= 0; --j)
800 val = operands_match_p (XVECEXP (x, i, j), XVECEXP (y, i, j), -1);
801 if (val == 0)
802 return false;
804 break;
806 /* It is believed that rtx's at this level will never
807 contain anything but integers and other rtx's, except for
808 within LABEL_REFs and SYMBOL_REFs. */
809 default:
810 gcc_unreachable ();
813 return true;
816 /* True if X is a constant that can be forced into the constant pool.
817 MODE is the mode of the operand, or VOIDmode if not known. */
818 #define CONST_POOL_OK_P(MODE, X) \
819 ((MODE) != VOIDmode \
820 && CONSTANT_P (X) \
821 && GET_CODE (X) != HIGH \
822 && !targetm.cannot_force_const_mem (MODE, X))
824 /* True if C is a non-empty register class that has too few registers
825 to be safely used as a reload target class. */
826 #define SMALL_REGISTER_CLASS_P(C) \
827 (ira_class_hard_regs_num [(C)] == 1 \
828 || (ira_class_hard_regs_num [(C)] >= 1 \
829 && targetm.class_likely_spilled_p (C)))
831 /* If REG is a reload pseudo, try to make its class satisfying CL. */
832 static void
833 narrow_reload_pseudo_class (rtx reg, enum reg_class cl)
835 enum reg_class rclass;
837 /* Do not make more accurate class from reloads generated. They are
838 mostly moves with a lot of constraints. Making more accurate
839 class may results in very narrow class and impossibility of find
840 registers for several reloads of one insn. */
841 if (INSN_UID (curr_insn) >= new_insn_uid_start)
842 return;
843 if (GET_CODE (reg) == SUBREG)
844 reg = SUBREG_REG (reg);
845 if (! REG_P (reg) || (int) REGNO (reg) < new_regno_start)
846 return;
847 if (in_class_p (reg, cl, &rclass) && rclass != cl)
848 lra_change_class (REGNO (reg), rclass, " Change to", true);
851 /* Searches X for any reference to a reg with the same value as REGNO,
852 returning the rtx of the reference found if any. Otherwise,
853 returns NULL_RTX. */
854 static rtx
855 regno_val_use_in (unsigned int regno, rtx x)
857 const char *fmt;
858 int i, j;
859 rtx tem;
861 if (REG_P (x) && lra_reg_info[REGNO (x)].val == lra_reg_info[regno].val)
862 return x;
864 fmt = GET_RTX_FORMAT (GET_CODE (x));
865 for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
867 if (fmt[i] == 'e')
869 if ((tem = regno_val_use_in (regno, XEXP (x, i))))
870 return tem;
872 else if (fmt[i] == 'E')
873 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
874 if ((tem = regno_val_use_in (regno , XVECEXP (x, i, j))))
875 return tem;
878 return NULL_RTX;
881 /* Return true if all current insn non-output operands except INS (it
882 has a negaitve end marker) do not use pseudos with the same value
883 as REGNO. */
884 static bool
885 check_conflict_input_operands (int regno, signed char *ins)
887 int in;
888 int n_operands = curr_static_id->n_operands;
890 for (int nop = 0; nop < n_operands; nop++)
891 if (! curr_static_id->operand[nop].is_operator
892 && curr_static_id->operand[nop].type != OP_OUT)
894 for (int i = 0; (in = ins[i]) >= 0; i++)
895 if (in == nop)
896 break;
897 if (in < 0
898 && regno_val_use_in (regno, *curr_id->operand_loc[nop]) != NULL_RTX)
899 return false;
901 return true;
904 /* Generate reloads for matching OUT and INS (array of input operand
905 numbers with end marker -1) with reg class GOAL_CLASS, considering
906 output operands OUTS (similar array to INS) needing to be in different
907 registers. Add input and output reloads correspondingly to the lists
908 *BEFORE and *AFTER. OUT might be negative. In this case we generate
909 input reloads for matched input operands INS. EARLY_CLOBBER_P is a flag
910 that the output operand is early clobbered for chosen alternative. */
911 static void
912 match_reload (signed char out, signed char *ins, signed char *outs,
913 enum reg_class goal_class, rtx_insn **before,
914 rtx_insn **after, bool early_clobber_p)
916 bool out_conflict;
917 int i, in;
918 rtx new_in_reg, new_out_reg, reg;
919 machine_mode inmode, outmode;
920 rtx in_rtx = *curr_id->operand_loc[ins[0]];
921 rtx out_rtx = out < 0 ? in_rtx : *curr_id->operand_loc[out];
923 inmode = curr_operand_mode[ins[0]];
924 outmode = out < 0 ? inmode : curr_operand_mode[out];
925 push_to_sequence (*before);
926 if (inmode != outmode)
928 if (GET_MODE_SIZE (inmode) > GET_MODE_SIZE (outmode))
930 reg = new_in_reg
931 = lra_create_new_reg_with_unique_value (inmode, in_rtx,
932 goal_class, "");
933 if (SCALAR_INT_MODE_P (inmode))
934 new_out_reg = gen_lowpart_SUBREG (outmode, reg);
935 else
936 new_out_reg = gen_rtx_SUBREG (outmode, reg, 0);
937 LRA_SUBREG_P (new_out_reg) = 1;
938 /* If the input reg is dying here, we can use the same hard
939 register for REG and IN_RTX. We do it only for original
940 pseudos as reload pseudos can die although original
941 pseudos still live where reload pseudos dies. */
942 if (REG_P (in_rtx) && (int) REGNO (in_rtx) < lra_new_regno_start
943 && find_regno_note (curr_insn, REG_DEAD, REGNO (in_rtx))
944 && (!early_clobber_p
945 || check_conflict_input_operands(REGNO (in_rtx), ins)))
946 lra_assign_reg_val (REGNO (in_rtx), REGNO (reg));
948 else
950 reg = new_out_reg
951 = lra_create_new_reg_with_unique_value (outmode, out_rtx,
952 goal_class, "");
953 if (SCALAR_INT_MODE_P (outmode))
954 new_in_reg = gen_lowpart_SUBREG (inmode, reg);
955 else
956 new_in_reg = gen_rtx_SUBREG (inmode, reg, 0);
957 /* NEW_IN_REG is non-paradoxical subreg. We don't want
958 NEW_OUT_REG living above. We add clobber clause for
959 this. This is just a temporary clobber. We can remove
960 it at the end of LRA work. */
961 rtx_insn *clobber = emit_clobber (new_out_reg);
962 LRA_TEMP_CLOBBER_P (PATTERN (clobber)) = 1;
963 LRA_SUBREG_P (new_in_reg) = 1;
964 if (GET_CODE (in_rtx) == SUBREG)
966 rtx subreg_reg = SUBREG_REG (in_rtx);
968 /* If SUBREG_REG is dying here and sub-registers IN_RTX
969 and NEW_IN_REG are similar, we can use the same hard
970 register for REG and SUBREG_REG. */
971 if (REG_P (subreg_reg)
972 && (int) REGNO (subreg_reg) < lra_new_regno_start
973 && GET_MODE (subreg_reg) == outmode
974 && SUBREG_BYTE (in_rtx) == SUBREG_BYTE (new_in_reg)
975 && find_regno_note (curr_insn, REG_DEAD, REGNO (subreg_reg))
976 && (! early_clobber_p
977 || check_conflict_input_operands (REGNO (subreg_reg),
978 ins)))
979 lra_assign_reg_val (REGNO (subreg_reg), REGNO (reg));
983 else
985 /* Pseudos have values -- see comments for lra_reg_info.
986 Different pseudos with the same value do not conflict even if
987 they live in the same place. When we create a pseudo we
988 assign value of original pseudo (if any) from which we
989 created the new pseudo. If we create the pseudo from the
990 input pseudo, the new pseudo will have no conflict with the
991 input pseudo which is wrong when the input pseudo lives after
992 the insn and as the new pseudo value is changed by the insn
993 output. Therefore we create the new pseudo from the output
994 except the case when we have single matched dying input
995 pseudo.
997 We cannot reuse the current output register because we might
998 have a situation like "a <- a op b", where the constraints
999 force the second input operand ("b") to match the output
1000 operand ("a"). "b" must then be copied into a new register
1001 so that it doesn't clobber the current value of "a".
1003 We can not use the same value if the output pseudo is
1004 early clobbered or the input pseudo is mentioned in the
1005 output, e.g. as an address part in memory, because
1006 output reload will actually extend the pseudo liveness.
1007 We don't care about eliminable hard regs here as we are
1008 interesting only in pseudos. */
1010 /* Matching input's register value is the same as one of the other
1011 output operand. Output operands in a parallel insn must be in
1012 different registers. */
1013 out_conflict = false;
1014 if (REG_P (in_rtx))
1016 for (i = 0; outs[i] >= 0; i++)
1018 rtx other_out_rtx = *curr_id->operand_loc[outs[i]];
1019 if (REG_P (other_out_rtx)
1020 && (regno_val_use_in (REGNO (in_rtx), other_out_rtx)
1021 != NULL_RTX))
1023 out_conflict = true;
1024 break;
1029 new_in_reg = new_out_reg
1030 = (! early_clobber_p && ins[1] < 0 && REG_P (in_rtx)
1031 && (int) REGNO (in_rtx) < lra_new_regno_start
1032 && find_regno_note (curr_insn, REG_DEAD, REGNO (in_rtx))
1033 && (! early_clobber_p
1034 || check_conflict_input_operands (REGNO (in_rtx), ins))
1035 && (out < 0
1036 || regno_val_use_in (REGNO (in_rtx), out_rtx) == NULL_RTX)
1037 && !out_conflict
1038 ? lra_create_new_reg (inmode, in_rtx, goal_class, "")
1039 : lra_create_new_reg_with_unique_value (outmode, out_rtx,
1040 goal_class, ""));
1042 /* In operand can be got from transformations before processing insn
1043 constraints. One example of such transformations is subreg
1044 reloading (see function simplify_operand_subreg). The new
1045 pseudos created by the transformations might have inaccurate
1046 class (ALL_REGS) and we should make their classes more
1047 accurate. */
1048 narrow_reload_pseudo_class (in_rtx, goal_class);
1049 lra_emit_move (copy_rtx (new_in_reg), in_rtx);
1050 *before = get_insns ();
1051 end_sequence ();
1052 /* Add the new pseudo to consider values of subsequent input reload
1053 pseudos. */
1054 lra_assert (curr_insn_input_reloads_num < LRA_MAX_INSN_RELOADS);
1055 curr_insn_input_reloads[curr_insn_input_reloads_num].input = in_rtx;
1056 curr_insn_input_reloads[curr_insn_input_reloads_num].match_p = true;
1057 curr_insn_input_reloads[curr_insn_input_reloads_num++].reg = new_in_reg;
1058 for (i = 0; (in = ins[i]) >= 0; i++)
1060 lra_assert
1061 (GET_MODE (*curr_id->operand_loc[in]) == VOIDmode
1062 || GET_MODE (new_in_reg) == GET_MODE (*curr_id->operand_loc[in]));
1063 *curr_id->operand_loc[in] = new_in_reg;
1065 lra_update_dups (curr_id, ins);
1066 if (out < 0)
1067 return;
1068 /* See a comment for the input operand above. */
1069 narrow_reload_pseudo_class (out_rtx, goal_class);
1070 if (find_reg_note (curr_insn, REG_UNUSED, out_rtx) == NULL_RTX)
1072 start_sequence ();
1073 lra_emit_move (out_rtx, copy_rtx (new_out_reg));
1074 emit_insn (*after);
1075 *after = get_insns ();
1076 end_sequence ();
1078 *curr_id->operand_loc[out] = new_out_reg;
1079 lra_update_dup (curr_id, out);
1082 /* Return register class which is union of all reg classes in insn
1083 constraint alternative string starting with P. */
1084 static enum reg_class
1085 reg_class_from_constraints (const char *p)
1087 int c, len;
1088 enum reg_class op_class = NO_REGS;
1091 switch ((c = *p, len = CONSTRAINT_LEN (c, p)), c)
1093 case '#':
1094 case ',':
1095 return op_class;
1097 case 'g':
1098 op_class = reg_class_subunion[op_class][GENERAL_REGS];
1099 break;
1101 default:
1102 enum constraint_num cn = lookup_constraint (p);
1103 enum reg_class cl = reg_class_for_constraint (cn);
1104 if (cl == NO_REGS)
1106 if (insn_extra_address_constraint (cn))
1107 op_class
1108 = (reg_class_subunion
1109 [op_class][base_reg_class (VOIDmode, ADDR_SPACE_GENERIC,
1110 ADDRESS, SCRATCH)]);
1111 break;
1114 op_class = reg_class_subunion[op_class][cl];
1115 break;
1117 while ((p += len), c);
1118 return op_class;
1121 /* If OP is a register, return the class of the register as per
1122 get_reg_class, otherwise return NO_REGS. */
1123 static inline enum reg_class
1124 get_op_class (rtx op)
1126 return REG_P (op) ? get_reg_class (REGNO (op)) : NO_REGS;
1129 /* Return generated insn mem_pseudo:=val if TO_P or val:=mem_pseudo
1130 otherwise. If modes of MEM_PSEUDO and VAL are different, use
1131 SUBREG for VAL to make them equal. */
1132 static rtx_insn *
1133 emit_spill_move (bool to_p, rtx mem_pseudo, rtx val)
1135 if (GET_MODE (mem_pseudo) != GET_MODE (val))
1137 /* Usually size of mem_pseudo is greater than val size but in
1138 rare cases it can be less as it can be defined by target
1139 dependent macro HARD_REGNO_CALLER_SAVE_MODE. */
1140 if (! MEM_P (val))
1142 val = gen_lowpart_SUBREG (GET_MODE (mem_pseudo),
1143 GET_CODE (val) == SUBREG
1144 ? SUBREG_REG (val) : val);
1145 LRA_SUBREG_P (val) = 1;
1147 else
1149 mem_pseudo = gen_lowpart_SUBREG (GET_MODE (val), mem_pseudo);
1150 LRA_SUBREG_P (mem_pseudo) = 1;
1153 return to_p ? gen_move_insn (mem_pseudo, val)
1154 : gen_move_insn (val, mem_pseudo);
1157 /* Process a special case insn (register move), return true if we
1158 don't need to process it anymore. INSN should be a single set
1159 insn. Set up that RTL was changed through CHANGE_P and macro
1160 SECONDARY_MEMORY_NEEDED says to use secondary memory through
1161 SEC_MEM_P. */
1162 static bool
1163 check_and_process_move (bool *change_p, bool *sec_mem_p ATTRIBUTE_UNUSED)
1165 int sregno, dregno;
1166 rtx dest, src, dreg, sreg, new_reg, scratch_reg;
1167 rtx_insn *before;
1168 enum reg_class dclass, sclass, secondary_class;
1169 secondary_reload_info sri;
1171 lra_assert (curr_insn_set != NULL_RTX);
1172 dreg = dest = SET_DEST (curr_insn_set);
1173 sreg = src = SET_SRC (curr_insn_set);
1174 if (GET_CODE (dest) == SUBREG)
1175 dreg = SUBREG_REG (dest);
1176 if (GET_CODE (src) == SUBREG)
1177 sreg = SUBREG_REG (src);
1178 if (! (REG_P (dreg) || MEM_P (dreg)) || ! (REG_P (sreg) || MEM_P (sreg)))
1179 return false;
1180 sclass = dclass = NO_REGS;
1181 if (REG_P (dreg))
1182 dclass = get_reg_class (REGNO (dreg));
1183 gcc_assert (dclass < LIM_REG_CLASSES);
1184 if (dclass == ALL_REGS)
1185 /* ALL_REGS is used for new pseudos created by transformations
1186 like reload of SUBREG_REG (see function
1187 simplify_operand_subreg). We don't know their class yet. We
1188 should figure out the class from processing the insn
1189 constraints not in this fast path function. Even if ALL_REGS
1190 were a right class for the pseudo, secondary_... hooks usually
1191 are not define for ALL_REGS. */
1192 return false;
1193 if (REG_P (sreg))
1194 sclass = get_reg_class (REGNO (sreg));
1195 gcc_assert (sclass < LIM_REG_CLASSES);
1196 if (sclass == ALL_REGS)
1197 /* See comments above. */
1198 return false;
1199 if (sclass == NO_REGS && dclass == NO_REGS)
1200 return false;
1201 #ifdef SECONDARY_MEMORY_NEEDED
1202 if (SECONDARY_MEMORY_NEEDED (sclass, dclass, GET_MODE (src))
1203 #ifdef SECONDARY_MEMORY_NEEDED_MODE
1204 && ((sclass != NO_REGS && dclass != NO_REGS)
1205 || GET_MODE (src) != SECONDARY_MEMORY_NEEDED_MODE (GET_MODE (src)))
1206 #endif
1209 *sec_mem_p = true;
1210 return false;
1212 #endif
1213 if (! REG_P (dreg) || ! REG_P (sreg))
1214 return false;
1215 sri.prev_sri = NULL;
1216 sri.icode = CODE_FOR_nothing;
1217 sri.extra_cost = 0;
1218 secondary_class = NO_REGS;
1219 /* Set up hard register for a reload pseudo for hook
1220 secondary_reload because some targets just ignore unassigned
1221 pseudos in the hook. */
1222 if (dclass != NO_REGS && lra_get_regno_hard_regno (REGNO (dreg)) < 0)
1224 dregno = REGNO (dreg);
1225 reg_renumber[dregno] = ira_class_hard_regs[dclass][0];
1227 else
1228 dregno = -1;
1229 if (sclass != NO_REGS && lra_get_regno_hard_regno (REGNO (sreg)) < 0)
1231 sregno = REGNO (sreg);
1232 reg_renumber[sregno] = ira_class_hard_regs[sclass][0];
1234 else
1235 sregno = -1;
1236 if (sclass != NO_REGS)
1237 secondary_class
1238 = (enum reg_class) targetm.secondary_reload (false, dest,
1239 (reg_class_t) sclass,
1240 GET_MODE (src), &sri);
1241 if (sclass == NO_REGS
1242 || ((secondary_class != NO_REGS || sri.icode != CODE_FOR_nothing)
1243 && dclass != NO_REGS))
1245 enum reg_class old_sclass = secondary_class;
1246 secondary_reload_info old_sri = sri;
1248 sri.prev_sri = NULL;
1249 sri.icode = CODE_FOR_nothing;
1250 sri.extra_cost = 0;
1251 secondary_class
1252 = (enum reg_class) targetm.secondary_reload (true, src,
1253 (reg_class_t) dclass,
1254 GET_MODE (src), &sri);
1255 /* Check the target hook consistency. */
1256 lra_assert
1257 ((secondary_class == NO_REGS && sri.icode == CODE_FOR_nothing)
1258 || (old_sclass == NO_REGS && old_sri.icode == CODE_FOR_nothing)
1259 || (secondary_class == old_sclass && sri.icode == old_sri.icode));
1261 if (sregno >= 0)
1262 reg_renumber [sregno] = -1;
1263 if (dregno >= 0)
1264 reg_renumber [dregno] = -1;
1265 if (secondary_class == NO_REGS && sri.icode == CODE_FOR_nothing)
1266 return false;
1267 *change_p = true;
1268 new_reg = NULL_RTX;
1269 if (secondary_class != NO_REGS)
1270 new_reg = lra_create_new_reg_with_unique_value (GET_MODE (src), NULL_RTX,
1271 secondary_class,
1272 "secondary");
1273 start_sequence ();
1274 if (sri.icode == CODE_FOR_nothing)
1275 lra_emit_move (new_reg, src);
1276 else
1278 enum reg_class scratch_class;
1280 scratch_class = (reg_class_from_constraints
1281 (insn_data[sri.icode].operand[2].constraint));
1282 scratch_reg = (lra_create_new_reg_with_unique_value
1283 (insn_data[sri.icode].operand[2].mode, NULL_RTX,
1284 scratch_class, "scratch"));
1285 emit_insn (GEN_FCN (sri.icode) (new_reg != NULL_RTX ? new_reg : dest,
1286 src, scratch_reg));
1288 before = get_insns ();
1289 end_sequence ();
1290 lra_process_new_insns (curr_insn, before, NULL, "Inserting the move");
1291 if (new_reg != NULL_RTX)
1292 SET_SRC (curr_insn_set) = new_reg;
1293 else
1295 if (lra_dump_file != NULL)
1297 fprintf (lra_dump_file, "Deleting move %u\n", INSN_UID (curr_insn));
1298 dump_insn_slim (lra_dump_file, curr_insn);
1300 lra_set_insn_deleted (curr_insn);
1301 return true;
1303 return false;
1306 /* The following data describe the result of process_alt_operands.
1307 The data are used in curr_insn_transform to generate reloads. */
1309 /* The chosen reg classes which should be used for the corresponding
1310 operands. */
1311 static enum reg_class goal_alt[MAX_RECOG_OPERANDS];
1312 /* True if the operand should be the same as another operand and that
1313 other operand does not need a reload. */
1314 static bool goal_alt_match_win[MAX_RECOG_OPERANDS];
1315 /* True if the operand does not need a reload. */
1316 static bool goal_alt_win[MAX_RECOG_OPERANDS];
1317 /* True if the operand can be offsetable memory. */
1318 static bool goal_alt_offmemok[MAX_RECOG_OPERANDS];
1319 /* The number of an operand to which given operand can be matched to. */
1320 static int goal_alt_matches[MAX_RECOG_OPERANDS];
1321 /* The number of elements in the following array. */
1322 static int goal_alt_dont_inherit_ops_num;
1323 /* Numbers of operands whose reload pseudos should not be inherited. */
1324 static int goal_alt_dont_inherit_ops[MAX_RECOG_OPERANDS];
1325 /* True if the insn commutative operands should be swapped. */
1326 static bool goal_alt_swapped;
1327 /* The chosen insn alternative. */
1328 static int goal_alt_number;
1330 /* True if the corresponding operand is the result of an equivalence
1331 substitution. */
1332 static bool equiv_substition_p[MAX_RECOG_OPERANDS];
1334 /* The following five variables are used to choose the best insn
1335 alternative. They reflect final characteristics of the best
1336 alternative. */
1338 /* Number of necessary reloads and overall cost reflecting the
1339 previous value and other unpleasantness of the best alternative. */
1340 static int best_losers, best_overall;
1341 /* Overall number hard registers used for reloads. For example, on
1342 some targets we need 2 general registers to reload DFmode and only
1343 one floating point register. */
1344 static int best_reload_nregs;
1345 /* Overall number reflecting distances of previous reloading the same
1346 value. The distances are counted from the current BB start. It is
1347 used to improve inheritance chances. */
1348 static int best_reload_sum;
1350 /* True if the current insn should have no correspondingly input or
1351 output reloads. */
1352 static bool no_input_reloads_p, no_output_reloads_p;
1354 /* True if we swapped the commutative operands in the current
1355 insn. */
1356 static int curr_swapped;
1358 /* if CHECK_ONLY_P is false, arrange for address element *LOC to be a
1359 register of class CL. Add any input reloads to list BEFORE. AFTER
1360 is nonnull if *LOC is an automodified value; handle that case by
1361 adding the required output reloads to list AFTER. Return true if
1362 the RTL was changed.
1364 if CHECK_ONLY_P is true, check that the *LOC is a correct address
1365 register. Return false if the address register is correct. */
1366 static bool
1367 process_addr_reg (rtx *loc, bool check_only_p, rtx_insn **before, rtx_insn **after,
1368 enum reg_class cl)
1370 int regno;
1371 enum reg_class rclass, new_class;
1372 rtx reg;
1373 rtx new_reg;
1374 machine_mode mode;
1375 bool subreg_p, before_p = false;
1377 subreg_p = GET_CODE (*loc) == SUBREG;
1378 if (subreg_p)
1380 reg = SUBREG_REG (*loc);
1381 mode = GET_MODE (reg);
1383 /* For mode with size bigger than ptr_mode, there unlikely to be "mov"
1384 between two registers with different classes, but there normally will
1385 be "mov" which transfers element of vector register into the general
1386 register, and this normally will be a subreg which should be reloaded
1387 as a whole. This is particularly likely to be triggered when
1388 -fno-split-wide-types specified. */
1389 if (!REG_P (reg)
1390 || in_class_p (reg, cl, &new_class)
1391 || GET_MODE_SIZE (mode) <= GET_MODE_SIZE (ptr_mode))
1392 loc = &SUBREG_REG (*loc);
1395 reg = *loc;
1396 mode = GET_MODE (reg);
1397 if (! REG_P (reg))
1399 if (check_only_p)
1400 return true;
1401 /* Always reload memory in an address even if the target supports
1402 such addresses. */
1403 new_reg = lra_create_new_reg_with_unique_value (mode, reg, cl, "address");
1404 before_p = true;
1406 else
1408 regno = REGNO (reg);
1409 rclass = get_reg_class (regno);
1410 if (! check_only_p
1411 && (*loc = get_equiv_with_elimination (reg, curr_insn)) != reg)
1413 if (lra_dump_file != NULL)
1415 fprintf (lra_dump_file,
1416 "Changing pseudo %d in address of insn %u on equiv ",
1417 REGNO (reg), INSN_UID (curr_insn));
1418 dump_value_slim (lra_dump_file, *loc, 1);
1419 fprintf (lra_dump_file, "\n");
1421 *loc = copy_rtx (*loc);
1423 if (*loc != reg || ! in_class_p (reg, cl, &new_class))
1425 if (check_only_p)
1426 return true;
1427 reg = *loc;
1428 if (get_reload_reg (after == NULL ? OP_IN : OP_INOUT,
1429 mode, reg, cl, subreg_p, "address", &new_reg))
1430 before_p = true;
1432 else if (new_class != NO_REGS && rclass != new_class)
1434 if (check_only_p)
1435 return true;
1436 lra_change_class (regno, new_class, " Change to", true);
1437 return false;
1439 else
1440 return false;
1442 if (before_p)
1444 push_to_sequence (*before);
1445 lra_emit_move (new_reg, reg);
1446 *before = get_insns ();
1447 end_sequence ();
1449 *loc = new_reg;
1450 if (after != NULL)
1452 start_sequence ();
1453 lra_emit_move (before_p ? copy_rtx (reg) : reg, new_reg);
1454 emit_insn (*after);
1455 *after = get_insns ();
1456 end_sequence ();
1458 return true;
1461 /* Insert move insn in simplify_operand_subreg. BEFORE returns
1462 the insn to be inserted before curr insn. AFTER returns the
1463 the insn to be inserted after curr insn. ORIGREG and NEWREG
1464 are the original reg and new reg for reload. */
1465 static void
1466 insert_move_for_subreg (rtx_insn **before, rtx_insn **after, rtx origreg,
1467 rtx newreg)
1469 if (before)
1471 push_to_sequence (*before);
1472 lra_emit_move (newreg, origreg);
1473 *before = get_insns ();
1474 end_sequence ();
1476 if (after)
1478 start_sequence ();
1479 lra_emit_move (origreg, newreg);
1480 emit_insn (*after);
1481 *after = get_insns ();
1482 end_sequence ();
1486 static int valid_address_p (machine_mode mode, rtx addr, addr_space_t as);
1487 static bool process_address (int, bool, rtx_insn **, rtx_insn **);
1489 /* Make reloads for subreg in operand NOP with internal subreg mode
1490 REG_MODE, add new reloads for further processing. Return true if
1491 any change was done. */
1492 static bool
1493 simplify_operand_subreg (int nop, machine_mode reg_mode)
1495 int hard_regno;
1496 rtx_insn *before, *after;
1497 machine_mode mode, innermode;
1498 rtx reg, new_reg;
1499 rtx operand = *curr_id->operand_loc[nop];
1500 enum reg_class regclass;
1501 enum op_type type;
1503 before = after = NULL;
1505 if (GET_CODE (operand) != SUBREG)
1506 return false;
1508 mode = GET_MODE (operand);
1509 reg = SUBREG_REG (operand);
1510 innermode = GET_MODE (reg);
1511 type = curr_static_id->operand[nop].type;
1512 if (MEM_P (reg))
1514 const bool addr_was_valid
1515 = valid_address_p (innermode, XEXP (reg, 0), MEM_ADDR_SPACE (reg));
1516 alter_subreg (curr_id->operand_loc[nop], false);
1517 rtx subst = *curr_id->operand_loc[nop];
1518 lra_assert (MEM_P (subst));
1520 if (!addr_was_valid
1521 || valid_address_p (GET_MODE (subst), XEXP (subst, 0),
1522 MEM_ADDR_SPACE (subst))
1523 || ((get_constraint_type (lookup_constraint
1524 (curr_static_id->operand[nop].constraint))
1525 != CT_SPECIAL_MEMORY)
1526 /* We still can reload address and if the address is
1527 valid, we can remove subreg without reloading its
1528 inner memory. */
1529 && valid_address_p (GET_MODE (subst),
1530 regno_reg_rtx
1531 [ira_class_hard_regs
1532 [base_reg_class (GET_MODE (subst),
1533 MEM_ADDR_SPACE (subst),
1534 ADDRESS, SCRATCH)][0]],
1535 MEM_ADDR_SPACE (subst))))
1537 /* If we change the address for a paradoxical subreg of memory, the
1538 new address might violate the necessary alignment or the access
1539 might be slow; take this into consideration. We need not worry
1540 about accesses beyond allocated memory for paradoxical memory
1541 subregs as we don't substitute such equiv memory (see processing
1542 equivalences in function lra_constraints) and because for spilled
1543 pseudos we allocate stack memory enough for the biggest
1544 corresponding paradoxical subreg.
1546 However, do not blindly simplify a (subreg (mem ...)) for
1547 WORD_REGISTER_OPERATIONS targets as this may lead to loading junk
1548 data into a register when the inner is narrower than outer or
1549 missing important data from memory when the inner is wider than
1550 outer. This rule only applies to modes that are no wider than
1551 a word. */
1552 if (!(GET_MODE_PRECISION (mode) != GET_MODE_PRECISION (innermode)
1553 && GET_MODE_SIZE (mode) <= UNITS_PER_WORD
1554 && GET_MODE_SIZE (innermode) <= UNITS_PER_WORD
1555 && WORD_REGISTER_OPERATIONS)
1556 && (!(MEM_ALIGN (subst) < GET_MODE_ALIGNMENT (mode)
1557 && SLOW_UNALIGNED_ACCESS (mode, MEM_ALIGN (subst)))
1558 || (MEM_ALIGN (reg) < GET_MODE_ALIGNMENT (innermode)
1559 && SLOW_UNALIGNED_ACCESS (innermode, MEM_ALIGN (reg)))))
1560 return true;
1562 *curr_id->operand_loc[nop] = operand;
1564 /* But if the address was not valid, we cannot reload the MEM without
1565 reloading the address first. */
1566 if (!addr_was_valid)
1567 process_address (nop, false, &before, &after);
1569 /* INNERMODE is fast, MODE slow. Reload the mem in INNERMODE. */
1570 enum reg_class rclass
1571 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1572 if (get_reload_reg (curr_static_id->operand[nop].type, innermode,
1573 reg, rclass, TRUE, "slow mem", &new_reg))
1575 bool insert_before, insert_after;
1576 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1578 insert_before = (type != OP_OUT
1579 || GET_MODE_SIZE (innermode)
1580 > GET_MODE_SIZE (mode));
1581 insert_after = type != OP_IN;
1582 insert_move_for_subreg (insert_before ? &before : NULL,
1583 insert_after ? &after : NULL,
1584 reg, new_reg);
1586 SUBREG_REG (operand) = new_reg;
1588 /* Convert to MODE. */
1589 reg = operand;
1590 rclass
1591 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1592 if (get_reload_reg (curr_static_id->operand[nop].type, mode, reg,
1593 rclass, TRUE, "slow mem", &new_reg))
1595 bool insert_before, insert_after;
1596 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1598 insert_before = type != OP_OUT;
1599 insert_after = type != OP_IN;
1600 insert_move_for_subreg (insert_before ? &before : NULL,
1601 insert_after ? &after : NULL,
1602 reg, new_reg);
1604 *curr_id->operand_loc[nop] = new_reg;
1605 lra_process_new_insns (curr_insn, before, after,
1606 "Inserting slow mem reload");
1607 return true;
1610 /* If the address was valid and became invalid, prefer to reload
1611 the memory. Typical case is when the index scale should
1612 correspond the memory. */
1613 *curr_id->operand_loc[nop] = operand;
1614 /* Do not return false here as the MEM_P (reg) will be processed
1615 later in this function. */
1617 else if (REG_P (reg) && REGNO (reg) < FIRST_PSEUDO_REGISTER)
1619 alter_subreg (curr_id->operand_loc[nop], false);
1620 return true;
1622 else if (CONSTANT_P (reg))
1624 /* Try to simplify subreg of constant. It is usually result of
1625 equivalence substitution. */
1626 if (innermode == VOIDmode
1627 && (innermode = original_subreg_reg_mode[nop]) == VOIDmode)
1628 innermode = curr_static_id->operand[nop].mode;
1629 if ((new_reg = simplify_subreg (mode, reg, innermode,
1630 SUBREG_BYTE (operand))) != NULL_RTX)
1632 *curr_id->operand_loc[nop] = new_reg;
1633 return true;
1636 /* Put constant into memory when we have mixed modes. It generates
1637 a better code in most cases as it does not need a secondary
1638 reload memory. It also prevents LRA looping when LRA is using
1639 secondary reload memory again and again. */
1640 if (CONSTANT_P (reg) && CONST_POOL_OK_P (reg_mode, reg)
1641 && SCALAR_INT_MODE_P (reg_mode) != SCALAR_INT_MODE_P (mode))
1643 SUBREG_REG (operand) = force_const_mem (reg_mode, reg);
1644 alter_subreg (curr_id->operand_loc[nop], false);
1645 return true;
1647 /* Force a reload of the SUBREG_REG if this is a constant or PLUS or
1648 if there may be a problem accessing OPERAND in the outer
1649 mode. */
1650 if ((REG_P (reg)
1651 && REGNO (reg) >= FIRST_PSEUDO_REGISTER
1652 && (hard_regno = lra_get_regno_hard_regno (REGNO (reg))) >= 0
1653 /* Don't reload paradoxical subregs because we could be looping
1654 having repeatedly final regno out of hard regs range. */
1655 && (hard_regno_nregs[hard_regno][innermode]
1656 >= hard_regno_nregs[hard_regno][mode])
1657 && simplify_subreg_regno (hard_regno, innermode,
1658 SUBREG_BYTE (operand), mode) < 0
1659 /* Don't reload subreg for matching reload. It is actually
1660 valid subreg in LRA. */
1661 && ! LRA_SUBREG_P (operand))
1662 || CONSTANT_P (reg) || GET_CODE (reg) == PLUS || MEM_P (reg))
1664 enum reg_class rclass;
1666 if (REG_P (reg))
1667 /* There is a big probability that we will get the same class
1668 for the new pseudo and we will get the same insn which
1669 means infinite looping. So spill the new pseudo. */
1670 rclass = NO_REGS;
1671 else
1672 /* The class will be defined later in curr_insn_transform. */
1673 rclass
1674 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1676 if (get_reload_reg (curr_static_id->operand[nop].type, reg_mode, reg,
1677 rclass, TRUE, "subreg reg", &new_reg))
1679 bool insert_before, insert_after;
1680 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1682 insert_before = (type != OP_OUT
1683 || GET_MODE_SIZE (innermode) > GET_MODE_SIZE (mode));
1684 insert_after = (type != OP_IN);
1685 insert_move_for_subreg (insert_before ? &before : NULL,
1686 insert_after ? &after : NULL,
1687 reg, new_reg);
1689 SUBREG_REG (operand) = new_reg;
1690 lra_process_new_insns (curr_insn, before, after,
1691 "Inserting subreg reload");
1692 return true;
1694 /* Force a reload for a paradoxical subreg. For paradoxical subreg,
1695 IRA allocates hardreg to the inner pseudo reg according to its mode
1696 instead of the outermode, so the size of the hardreg may not be enough
1697 to contain the outermode operand, in that case we may need to insert
1698 reload for the reg. For the following two types of paradoxical subreg,
1699 we need to insert reload:
1700 1. If the op_type is OP_IN, and the hardreg could not be paired with
1701 other hardreg to contain the outermode operand
1702 (checked by in_hard_reg_set_p), we need to insert the reload.
1703 2. If the op_type is OP_OUT or OP_INOUT.
1705 Here is a paradoxical subreg example showing how the reload is generated:
1707 (insn 5 4 7 2 (set (reg:TI 106 [ __comp ])
1708 (subreg:TI (reg:DI 107 [ __comp ]) 0)) {*movti_internal_rex64}
1710 In IRA, reg107 is allocated to a DImode hardreg. We use x86-64 as example
1711 here, if reg107 is assigned to hardreg R15, because R15 is the last
1712 hardreg, compiler cannot find another hardreg to pair with R15 to
1713 contain TImode data. So we insert a TImode reload reg180 for it.
1714 After reload is inserted:
1716 (insn 283 0 0 (set (subreg:DI (reg:TI 180 [orig:107 __comp ] [107]) 0)
1717 (reg:DI 107 [ __comp ])) -1
1718 (insn 5 4 7 2 (set (reg:TI 106 [ __comp ])
1719 (subreg:TI (reg:TI 180 [orig:107 __comp ] [107]) 0)) {*movti_internal_rex64}
1721 Two reload hard registers will be allocated to reg180 to save TImode data
1722 in LRA_assign. */
1723 else if (REG_P (reg)
1724 && REGNO (reg) >= FIRST_PSEUDO_REGISTER
1725 && (hard_regno = lra_get_regno_hard_regno (REGNO (reg))) >= 0
1726 && (hard_regno_nregs[hard_regno][innermode]
1727 < hard_regno_nregs[hard_regno][mode])
1728 && (regclass = lra_get_allocno_class (REGNO (reg)))
1729 && (type != OP_IN
1730 || !in_hard_reg_set_p (reg_class_contents[regclass],
1731 mode, hard_regno)))
1733 /* The class will be defined later in curr_insn_transform. */
1734 enum reg_class rclass
1735 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1737 if (get_reload_reg (curr_static_id->operand[nop].type, mode, reg,
1738 rclass, TRUE, "paradoxical subreg", &new_reg))
1740 rtx subreg;
1741 bool insert_before, insert_after;
1743 PUT_MODE (new_reg, mode);
1744 subreg = gen_lowpart_SUBREG (innermode, new_reg);
1745 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1747 insert_before = (type != OP_OUT);
1748 insert_after = (type != OP_IN);
1749 insert_move_for_subreg (insert_before ? &before : NULL,
1750 insert_after ? &after : NULL,
1751 reg, subreg);
1753 SUBREG_REG (operand) = new_reg;
1754 lra_process_new_insns (curr_insn, before, after,
1755 "Inserting paradoxical subreg reload");
1756 return true;
1758 return false;
1761 /* Return TRUE if X refers for a hard register from SET. */
1762 static bool
1763 uses_hard_regs_p (rtx x, HARD_REG_SET set)
1765 int i, j, x_hard_regno;
1766 machine_mode mode;
1767 const char *fmt;
1768 enum rtx_code code;
1770 if (x == NULL_RTX)
1771 return false;
1772 code = GET_CODE (x);
1773 mode = GET_MODE (x);
1774 if (code == SUBREG)
1776 x = SUBREG_REG (x);
1777 code = GET_CODE (x);
1778 if (GET_MODE_SIZE (GET_MODE (x)) > GET_MODE_SIZE (mode))
1779 mode = GET_MODE (x);
1782 if (REG_P (x))
1784 x_hard_regno = get_hard_regno (x, true);
1785 return (x_hard_regno >= 0
1786 && overlaps_hard_reg_set_p (set, mode, x_hard_regno));
1788 if (MEM_P (x))
1790 struct address_info ad;
1792 decompose_mem_address (&ad, x);
1793 if (ad.base_term != NULL && uses_hard_regs_p (*ad.base_term, set))
1794 return true;
1795 if (ad.index_term != NULL && uses_hard_regs_p (*ad.index_term, set))
1796 return true;
1798 fmt = GET_RTX_FORMAT (code);
1799 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1801 if (fmt[i] == 'e')
1803 if (uses_hard_regs_p (XEXP (x, i), set))
1804 return true;
1806 else if (fmt[i] == 'E')
1808 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
1809 if (uses_hard_regs_p (XVECEXP (x, i, j), set))
1810 return true;
1813 return false;
1816 /* Return true if OP is a spilled pseudo. */
1817 static inline bool
1818 spilled_pseudo_p (rtx op)
1820 return (REG_P (op)
1821 && REGNO (op) >= FIRST_PSEUDO_REGISTER && in_mem_p (REGNO (op)));
1824 /* Return true if X is a general constant. */
1825 static inline bool
1826 general_constant_p (rtx x)
1828 return CONSTANT_P (x) && (! flag_pic || LEGITIMATE_PIC_OPERAND_P (x));
1831 static bool
1832 reg_in_class_p (rtx reg, enum reg_class cl)
1834 if (cl == NO_REGS)
1835 return get_reg_class (REGNO (reg)) == NO_REGS;
1836 return in_class_p (reg, cl, NULL);
1839 /* Return true if SET of RCLASS contains no hard regs which can be
1840 used in MODE. */
1841 static bool
1842 prohibited_class_reg_set_mode_p (enum reg_class rclass,
1843 HARD_REG_SET &set,
1844 enum machine_mode mode)
1846 HARD_REG_SET temp;
1848 lra_assert (hard_reg_set_subset_p (reg_class_contents[rclass], set));
1849 COPY_HARD_REG_SET (temp, set);
1850 AND_COMPL_HARD_REG_SET (temp, lra_no_alloc_regs);
1851 return (hard_reg_set_subset_p
1852 (temp, ira_prohibited_class_mode_regs[rclass][mode]));
1855 /* Major function to choose the current insn alternative and what
1856 operands should be reloaded and how. If ONLY_ALTERNATIVE is not
1857 negative we should consider only this alternative. Return false if
1858 we can not choose the alternative or find how to reload the
1859 operands. */
1860 static bool
1861 process_alt_operands (int only_alternative)
1863 bool ok_p = false;
1864 int nop, overall, nalt;
1865 int n_alternatives = curr_static_id->n_alternatives;
1866 int n_operands = curr_static_id->n_operands;
1867 /* LOSERS counts the operands that don't fit this alternative and
1868 would require loading. */
1869 int losers;
1870 int addr_losers;
1871 /* REJECT is a count of how undesirable this alternative says it is
1872 if any reloading is required. If the alternative matches exactly
1873 then REJECT is ignored, but otherwise it gets this much counted
1874 against it in addition to the reloading needed. */
1875 int reject;
1876 /* This is defined by '!' or '?' alternative constraint and added to
1877 reject. But in some cases it can be ignored. */
1878 int static_reject;
1879 int op_reject;
1880 /* The number of elements in the following array. */
1881 int early_clobbered_regs_num;
1882 /* Numbers of operands which are early clobber registers. */
1883 int early_clobbered_nops[MAX_RECOG_OPERANDS];
1884 enum reg_class curr_alt[MAX_RECOG_OPERANDS];
1885 HARD_REG_SET curr_alt_set[MAX_RECOG_OPERANDS];
1886 bool curr_alt_match_win[MAX_RECOG_OPERANDS];
1887 bool curr_alt_win[MAX_RECOG_OPERANDS];
1888 bool curr_alt_offmemok[MAX_RECOG_OPERANDS];
1889 int curr_alt_matches[MAX_RECOG_OPERANDS];
1890 /* The number of elements in the following array. */
1891 int curr_alt_dont_inherit_ops_num;
1892 /* Numbers of operands whose reload pseudos should not be inherited. */
1893 int curr_alt_dont_inherit_ops[MAX_RECOG_OPERANDS];
1894 rtx op;
1895 /* The register when the operand is a subreg of register, otherwise the
1896 operand itself. */
1897 rtx no_subreg_reg_operand[MAX_RECOG_OPERANDS];
1898 /* The register if the operand is a register or subreg of register,
1899 otherwise NULL. */
1900 rtx operand_reg[MAX_RECOG_OPERANDS];
1901 int hard_regno[MAX_RECOG_OPERANDS];
1902 machine_mode biggest_mode[MAX_RECOG_OPERANDS];
1903 int reload_nregs, reload_sum;
1904 bool costly_p;
1905 enum reg_class cl;
1907 /* Calculate some data common for all alternatives to speed up the
1908 function. */
1909 for (nop = 0; nop < n_operands; nop++)
1911 rtx reg;
1913 op = no_subreg_reg_operand[nop] = *curr_id->operand_loc[nop];
1914 /* The real hard regno of the operand after the allocation. */
1915 hard_regno[nop] = get_hard_regno (op, true);
1917 operand_reg[nop] = reg = op;
1918 biggest_mode[nop] = GET_MODE (op);
1919 if (GET_CODE (op) == SUBREG)
1921 operand_reg[nop] = reg = SUBREG_REG (op);
1922 if (GET_MODE_SIZE (biggest_mode[nop])
1923 < GET_MODE_SIZE (GET_MODE (reg)))
1924 biggest_mode[nop] = GET_MODE (reg);
1926 if (! REG_P (reg))
1927 operand_reg[nop] = NULL_RTX;
1928 else if (REGNO (reg) >= FIRST_PSEUDO_REGISTER
1929 || ((int) REGNO (reg)
1930 == lra_get_elimination_hard_regno (REGNO (reg))))
1931 no_subreg_reg_operand[nop] = reg;
1932 else
1933 operand_reg[nop] = no_subreg_reg_operand[nop]
1934 /* Just use natural mode for elimination result. It should
1935 be enough for extra constraints hooks. */
1936 = regno_reg_rtx[hard_regno[nop]];
1939 /* The constraints are made of several alternatives. Each operand's
1940 constraint looks like foo,bar,... with commas separating the
1941 alternatives. The first alternatives for all operands go
1942 together, the second alternatives go together, etc.
1944 First loop over alternatives. */
1945 alternative_mask preferred = curr_id->preferred_alternatives;
1946 if (only_alternative >= 0)
1947 preferred &= ALTERNATIVE_BIT (only_alternative);
1949 for (nalt = 0; nalt < n_alternatives; nalt++)
1951 /* Loop over operands for one constraint alternative. */
1952 if (!TEST_BIT (preferred, nalt))
1953 continue;
1955 overall = losers = addr_losers = 0;
1956 static_reject = reject = reload_nregs = reload_sum = 0;
1957 for (nop = 0; nop < n_operands; nop++)
1959 int inc = (curr_static_id
1960 ->operand_alternative[nalt * n_operands + nop].reject);
1961 if (lra_dump_file != NULL && inc != 0)
1962 fprintf (lra_dump_file,
1963 " Staticly defined alt reject+=%d\n", inc);
1964 static_reject += inc;
1966 reject += static_reject;
1967 early_clobbered_regs_num = 0;
1969 for (nop = 0; nop < n_operands; nop++)
1971 const char *p;
1972 char *end;
1973 int len, c, m, i, opalt_num, this_alternative_matches;
1974 bool win, did_match, offmemok, early_clobber_p;
1975 /* false => this operand can be reloaded somehow for this
1976 alternative. */
1977 bool badop;
1978 /* true => this operand can be reloaded if the alternative
1979 allows regs. */
1980 bool winreg;
1981 /* True if a constant forced into memory would be OK for
1982 this operand. */
1983 bool constmemok;
1984 enum reg_class this_alternative, this_costly_alternative;
1985 HARD_REG_SET this_alternative_set, this_costly_alternative_set;
1986 bool this_alternative_match_win, this_alternative_win;
1987 bool this_alternative_offmemok;
1988 bool scratch_p;
1989 machine_mode mode;
1990 enum constraint_num cn;
1992 opalt_num = nalt * n_operands + nop;
1993 if (curr_static_id->operand_alternative[opalt_num].anything_ok)
1995 /* Fast track for no constraints at all. */
1996 curr_alt[nop] = NO_REGS;
1997 CLEAR_HARD_REG_SET (curr_alt_set[nop]);
1998 curr_alt_win[nop] = true;
1999 curr_alt_match_win[nop] = false;
2000 curr_alt_offmemok[nop] = false;
2001 curr_alt_matches[nop] = -1;
2002 continue;
2005 op = no_subreg_reg_operand[nop];
2006 mode = curr_operand_mode[nop];
2008 win = did_match = winreg = offmemok = constmemok = false;
2009 badop = true;
2011 early_clobber_p = false;
2012 p = curr_static_id->operand_alternative[opalt_num].constraint;
2014 this_costly_alternative = this_alternative = NO_REGS;
2015 /* We update set of possible hard regs besides its class
2016 because reg class might be inaccurate. For example,
2017 union of LO_REGS (l), HI_REGS(h), and STACK_REG(k) in ARM
2018 is translated in HI_REGS because classes are merged by
2019 pairs and there is no accurate intermediate class. */
2020 CLEAR_HARD_REG_SET (this_alternative_set);
2021 CLEAR_HARD_REG_SET (this_costly_alternative_set);
2022 this_alternative_win = false;
2023 this_alternative_match_win = false;
2024 this_alternative_offmemok = false;
2025 this_alternative_matches = -1;
2027 /* An empty constraint should be excluded by the fast
2028 track. */
2029 lra_assert (*p != 0 && *p != ',');
2031 op_reject = 0;
2032 /* Scan this alternative's specs for this operand; set WIN
2033 if the operand fits any letter in this alternative.
2034 Otherwise, clear BADOP if this operand could fit some
2035 letter after reloads, or set WINREG if this operand could
2036 fit after reloads provided the constraint allows some
2037 registers. */
2038 costly_p = false;
2041 switch ((c = *p, len = CONSTRAINT_LEN (c, p)), c)
2043 case '\0':
2044 len = 0;
2045 break;
2046 case ',':
2047 c = '\0';
2048 break;
2050 case '&':
2051 early_clobber_p = true;
2052 break;
2054 case '$':
2055 op_reject += LRA_MAX_REJECT;
2056 break;
2057 case '^':
2058 op_reject += LRA_LOSER_COST_FACTOR;
2059 break;
2061 case '#':
2062 /* Ignore rest of this alternative. */
2063 c = '\0';
2064 break;
2066 case '0': case '1': case '2': case '3': case '4':
2067 case '5': case '6': case '7': case '8': case '9':
2069 int m_hregno;
2070 bool match_p;
2072 m = strtoul (p, &end, 10);
2073 p = end;
2074 len = 0;
2075 lra_assert (nop > m);
2077 this_alternative_matches = m;
2078 m_hregno = get_hard_regno (*curr_id->operand_loc[m], false);
2079 /* We are supposed to match a previous operand.
2080 If we do, we win if that one did. If we do
2081 not, count both of the operands as losers.
2082 (This is too conservative, since most of the
2083 time only a single reload insn will be needed
2084 to make the two operands win. As a result,
2085 this alternative may be rejected when it is
2086 actually desirable.) */
2087 match_p = false;
2088 if (operands_match_p (*curr_id->operand_loc[nop],
2089 *curr_id->operand_loc[m], m_hregno))
2091 /* We should reject matching of an early
2092 clobber operand if the matching operand is
2093 not dying in the insn. */
2094 if (! curr_static_id->operand[m].early_clobber
2095 || operand_reg[nop] == NULL_RTX
2096 || (find_regno_note (curr_insn, REG_DEAD,
2097 REGNO (op))
2098 || REGNO (op) == REGNO (operand_reg[m])))
2099 match_p = true;
2101 if (match_p)
2103 /* If we are matching a non-offsettable
2104 address where an offsettable address was
2105 expected, then we must reject this
2106 combination, because we can't reload
2107 it. */
2108 if (curr_alt_offmemok[m]
2109 && MEM_P (*curr_id->operand_loc[m])
2110 && curr_alt[m] == NO_REGS && ! curr_alt_win[m])
2111 continue;
2113 else
2115 /* Operands don't match. Both operands must
2116 allow a reload register, otherwise we
2117 cannot make them match. */
2118 if (curr_alt[m] == NO_REGS)
2119 break;
2120 /* Retroactively mark the operand we had to
2121 match as a loser, if it wasn't already and
2122 it wasn't matched to a register constraint
2123 (e.g it might be matched by memory). */
2124 if (curr_alt_win[m]
2125 && (operand_reg[m] == NULL_RTX
2126 || hard_regno[m] < 0))
2128 losers++;
2129 reload_nregs
2130 += (ira_reg_class_max_nregs[curr_alt[m]]
2131 [GET_MODE (*curr_id->operand_loc[m])]);
2134 /* Prefer matching earlyclobber alternative as
2135 it results in less hard regs required for
2136 the insn than a non-matching earlyclobber
2137 alternative. */
2138 if (curr_static_id->operand[m].early_clobber)
2140 if (lra_dump_file != NULL)
2141 fprintf
2142 (lra_dump_file,
2143 " %d Matching earlyclobber alt:"
2144 " reject--\n",
2145 nop);
2146 reject--;
2148 /* Otherwise we prefer no matching
2149 alternatives because it gives more freedom
2150 in RA. */
2151 else if (operand_reg[nop] == NULL_RTX
2152 || (find_regno_note (curr_insn, REG_DEAD,
2153 REGNO (operand_reg[nop]))
2154 == NULL_RTX))
2156 if (lra_dump_file != NULL)
2157 fprintf
2158 (lra_dump_file,
2159 " %d Matching alt: reject+=2\n",
2160 nop);
2161 reject += 2;
2164 /* If we have to reload this operand and some
2165 previous operand also had to match the same
2166 thing as this operand, we don't know how to do
2167 that. */
2168 if (!match_p || !curr_alt_win[m])
2170 for (i = 0; i < nop; i++)
2171 if (curr_alt_matches[i] == m)
2172 break;
2173 if (i < nop)
2174 break;
2176 else
2177 did_match = true;
2179 /* This can be fixed with reloads if the operand
2180 we are supposed to match can be fixed with
2181 reloads. */
2182 badop = false;
2183 this_alternative = curr_alt[m];
2184 COPY_HARD_REG_SET (this_alternative_set, curr_alt_set[m]);
2185 winreg = this_alternative != NO_REGS;
2186 break;
2189 case 'g':
2190 if (MEM_P (op)
2191 || general_constant_p (op)
2192 || spilled_pseudo_p (op))
2193 win = true;
2194 cl = GENERAL_REGS;
2195 goto reg;
2197 default:
2198 cn = lookup_constraint (p);
2199 switch (get_constraint_type (cn))
2201 case CT_REGISTER:
2202 cl = reg_class_for_constraint (cn);
2203 if (cl != NO_REGS)
2204 goto reg;
2205 break;
2207 case CT_CONST_INT:
2208 if (CONST_INT_P (op)
2209 && insn_const_int_ok_for_constraint (INTVAL (op), cn))
2210 win = true;
2211 break;
2213 case CT_MEMORY:
2214 if (MEM_P (op)
2215 && satisfies_memory_constraint_p (op, cn))
2216 win = true;
2217 else if (spilled_pseudo_p (op))
2218 win = true;
2220 /* If we didn't already win, we can reload constants
2221 via force_const_mem or put the pseudo value into
2222 memory, or make other memory by reloading the
2223 address like for 'o'. */
2224 if (CONST_POOL_OK_P (mode, op)
2225 || MEM_P (op) || REG_P (op)
2226 /* We can restore the equiv insn by a
2227 reload. */
2228 || equiv_substition_p[nop])
2229 badop = false;
2230 constmemok = true;
2231 offmemok = true;
2232 break;
2234 case CT_ADDRESS:
2235 /* If we didn't already win, we can reload the address
2236 into a base register. */
2237 if (satisfies_address_constraint_p (op, cn))
2238 win = true;
2239 cl = base_reg_class (VOIDmode, ADDR_SPACE_GENERIC,
2240 ADDRESS, SCRATCH);
2241 badop = false;
2242 goto reg;
2244 case CT_FIXED_FORM:
2245 if (constraint_satisfied_p (op, cn))
2246 win = true;
2247 break;
2249 case CT_SPECIAL_MEMORY:
2250 if (MEM_P (op)
2251 && satisfies_memory_constraint_p (op, cn))
2252 win = true;
2253 else if (spilled_pseudo_p (op))
2254 win = true;
2255 break;
2257 break;
2259 reg:
2260 this_alternative = reg_class_subunion[this_alternative][cl];
2261 IOR_HARD_REG_SET (this_alternative_set,
2262 reg_class_contents[cl]);
2263 if (costly_p)
2265 this_costly_alternative
2266 = reg_class_subunion[this_costly_alternative][cl];
2267 IOR_HARD_REG_SET (this_costly_alternative_set,
2268 reg_class_contents[cl]);
2270 if (mode == BLKmode)
2271 break;
2272 winreg = true;
2273 if (REG_P (op))
2275 if (hard_regno[nop] >= 0
2276 && in_hard_reg_set_p (this_alternative_set,
2277 mode, hard_regno[nop]))
2278 win = true;
2279 else if (hard_regno[nop] < 0
2280 && in_class_p (op, this_alternative, NULL))
2281 win = true;
2283 break;
2285 if (c != ' ' && c != '\t')
2286 costly_p = c == '*';
2288 while ((p += len), c);
2290 scratch_p = (operand_reg[nop] != NULL_RTX
2291 && lra_former_scratch_p (REGNO (operand_reg[nop])));
2292 /* Record which operands fit this alternative. */
2293 if (win)
2295 this_alternative_win = true;
2296 if (operand_reg[nop] != NULL_RTX)
2298 if (hard_regno[nop] >= 0)
2300 if (in_hard_reg_set_p (this_costly_alternative_set,
2301 mode, hard_regno[nop]))
2303 if (lra_dump_file != NULL)
2304 fprintf (lra_dump_file,
2305 " %d Costly set: reject++\n",
2306 nop);
2307 reject++;
2310 else
2312 /* Prefer won reg to spilled pseudo under other
2313 equal conditions for possibe inheritance. */
2314 if (! scratch_p)
2316 if (lra_dump_file != NULL)
2317 fprintf
2318 (lra_dump_file,
2319 " %d Non pseudo reload: reject++\n",
2320 nop);
2321 reject++;
2323 if (in_class_p (operand_reg[nop],
2324 this_costly_alternative, NULL))
2326 if (lra_dump_file != NULL)
2327 fprintf
2328 (lra_dump_file,
2329 " %d Non pseudo costly reload:"
2330 " reject++\n",
2331 nop);
2332 reject++;
2335 /* We simulate the behavior of old reload here.
2336 Although scratches need hard registers and it
2337 might result in spilling other pseudos, no reload
2338 insns are generated for the scratches. So it
2339 might cost something but probably less than old
2340 reload pass believes. */
2341 if (scratch_p)
2343 if (lra_dump_file != NULL)
2344 fprintf (lra_dump_file,
2345 " %d Scratch win: reject+=2\n",
2346 nop);
2347 reject += 2;
2351 else if (did_match)
2352 this_alternative_match_win = true;
2353 else
2355 int const_to_mem = 0;
2356 bool no_regs_p;
2358 reject += op_reject;
2359 /* Never do output reload of stack pointer. It makes
2360 impossible to do elimination when SP is changed in
2361 RTL. */
2362 if (op == stack_pointer_rtx && ! frame_pointer_needed
2363 && curr_static_id->operand[nop].type != OP_IN)
2364 goto fail;
2366 /* If this alternative asks for a specific reg class, see if there
2367 is at least one allocatable register in that class. */
2368 no_regs_p
2369 = (this_alternative == NO_REGS
2370 || (hard_reg_set_subset_p
2371 (reg_class_contents[this_alternative],
2372 lra_no_alloc_regs)));
2374 /* For asms, verify that the class for this alternative is possible
2375 for the mode that is specified. */
2376 if (!no_regs_p && INSN_CODE (curr_insn) < 0)
2378 int i;
2379 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
2380 if (HARD_REGNO_MODE_OK (i, mode)
2381 && in_hard_reg_set_p (reg_class_contents[this_alternative],
2382 mode, i))
2383 break;
2384 if (i == FIRST_PSEUDO_REGISTER)
2385 winreg = false;
2388 /* If this operand accepts a register, and if the
2389 register class has at least one allocatable register,
2390 then this operand can be reloaded. */
2391 if (winreg && !no_regs_p)
2392 badop = false;
2394 if (badop)
2396 if (lra_dump_file != NULL)
2397 fprintf (lra_dump_file,
2398 " alt=%d: Bad operand -- refuse\n",
2399 nalt);
2400 goto fail;
2403 if (this_alternative != NO_REGS)
2405 HARD_REG_SET available_regs;
2407 COPY_HARD_REG_SET (available_regs,
2408 reg_class_contents[this_alternative]);
2409 AND_COMPL_HARD_REG_SET
2410 (available_regs,
2411 ira_prohibited_class_mode_regs[this_alternative][mode]);
2412 AND_COMPL_HARD_REG_SET (available_regs, lra_no_alloc_regs);
2413 if (hard_reg_set_empty_p (available_regs))
2415 /* There are no hard regs holding a value of given
2416 mode. */
2417 if (offmemok)
2419 this_alternative = NO_REGS;
2420 if (lra_dump_file != NULL)
2421 fprintf (lra_dump_file,
2422 " %d Using memory because of"
2423 " a bad mode: reject+=2\n",
2424 nop);
2425 reject += 2;
2427 else
2429 if (lra_dump_file != NULL)
2430 fprintf (lra_dump_file,
2431 " alt=%d: Wrong mode -- refuse\n",
2432 nalt);
2433 goto fail;
2438 /* If not assigned pseudo has a class which a subset of
2439 required reg class, it is a less costly alternative
2440 as the pseudo still can get a hard reg of necessary
2441 class. */
2442 if (! no_regs_p && REG_P (op) && hard_regno[nop] < 0
2443 && (cl = get_reg_class (REGNO (op))) != NO_REGS
2444 && ira_class_subset_p[this_alternative][cl])
2446 if (lra_dump_file != NULL)
2447 fprintf
2448 (lra_dump_file,
2449 " %d Super set class reg: reject-=3\n", nop);
2450 reject -= 3;
2453 this_alternative_offmemok = offmemok;
2454 if (this_costly_alternative != NO_REGS)
2456 if (lra_dump_file != NULL)
2457 fprintf (lra_dump_file,
2458 " %d Costly loser: reject++\n", nop);
2459 reject++;
2461 /* If the operand is dying, has a matching constraint,
2462 and satisfies constraints of the matched operand
2463 which failed to satisfy the own constraints, most probably
2464 the reload for this operand will be gone. */
2465 if (this_alternative_matches >= 0
2466 && !curr_alt_win[this_alternative_matches]
2467 && REG_P (op)
2468 && find_regno_note (curr_insn, REG_DEAD, REGNO (op))
2469 && (hard_regno[nop] >= 0
2470 ? in_hard_reg_set_p (this_alternative_set,
2471 mode, hard_regno[nop])
2472 : in_class_p (op, this_alternative, NULL)))
2474 if (lra_dump_file != NULL)
2475 fprintf
2476 (lra_dump_file,
2477 " %d Dying matched operand reload: reject++\n",
2478 nop);
2479 reject++;
2481 else
2483 /* Strict_low_part requires to reload the register
2484 not the sub-register. In this case we should
2485 check that a final reload hard reg can hold the
2486 value mode. */
2487 if (curr_static_id->operand[nop].strict_low
2488 && REG_P (op)
2489 && hard_regno[nop] < 0
2490 && GET_CODE (*curr_id->operand_loc[nop]) == SUBREG
2491 && ira_class_hard_regs_num[this_alternative] > 0
2492 && ! HARD_REGNO_MODE_OK (ira_class_hard_regs
2493 [this_alternative][0],
2494 GET_MODE
2495 (*curr_id->operand_loc[nop])))
2497 if (lra_dump_file != NULL)
2498 fprintf
2499 (lra_dump_file,
2500 " alt=%d: Strict low subreg reload -- refuse\n",
2501 nalt);
2502 goto fail;
2504 losers++;
2506 if (operand_reg[nop] != NULL_RTX
2507 /* Output operands and matched input operands are
2508 not inherited. The following conditions do not
2509 exactly describe the previous statement but they
2510 are pretty close. */
2511 && curr_static_id->operand[nop].type != OP_OUT
2512 && (this_alternative_matches < 0
2513 || curr_static_id->operand[nop].type != OP_IN))
2515 int last_reload = (lra_reg_info[ORIGINAL_REGNO
2516 (operand_reg[nop])]
2517 .last_reload);
2519 /* The value of reload_sum has sense only if we
2520 process insns in their order. It happens only on
2521 the first constraints sub-pass when we do most of
2522 reload work. */
2523 if (lra_constraint_iter == 1 && last_reload > bb_reload_num)
2524 reload_sum += last_reload - bb_reload_num;
2526 /* If this is a constant that is reloaded into the
2527 desired class by copying it to memory first, count
2528 that as another reload. This is consistent with
2529 other code and is required to avoid choosing another
2530 alternative when the constant is moved into memory.
2531 Note that the test here is precisely the same as in
2532 the code below that calls force_const_mem. */
2533 if (CONST_POOL_OK_P (mode, op)
2534 && ((targetm.preferred_reload_class
2535 (op, this_alternative) == NO_REGS)
2536 || no_input_reloads_p))
2538 const_to_mem = 1;
2539 if (! no_regs_p)
2540 losers++;
2543 /* Alternative loses if it requires a type of reload not
2544 permitted for this insn. We can always reload
2545 objects with a REG_UNUSED note. */
2546 if ((curr_static_id->operand[nop].type != OP_IN
2547 && no_output_reloads_p
2548 && ! find_reg_note (curr_insn, REG_UNUSED, op))
2549 || (curr_static_id->operand[nop].type != OP_OUT
2550 && no_input_reloads_p && ! const_to_mem)
2551 || (this_alternative_matches >= 0
2552 && (no_input_reloads_p
2553 || (no_output_reloads_p
2554 && (curr_static_id->operand
2555 [this_alternative_matches].type != OP_IN)
2556 && ! find_reg_note (curr_insn, REG_UNUSED,
2557 no_subreg_reg_operand
2558 [this_alternative_matches])))))
2560 if (lra_dump_file != NULL)
2561 fprintf
2562 (lra_dump_file,
2563 " alt=%d: No input/otput reload -- refuse\n",
2564 nalt);
2565 goto fail;
2568 /* Alternative loses if it required class pseudo can not
2569 hold value of required mode. Such insns can be
2570 described by insn definitions with mode iterators. */
2571 if (GET_MODE (*curr_id->operand_loc[nop]) != VOIDmode
2572 && ! hard_reg_set_empty_p (this_alternative_set)
2573 /* It is common practice for constraints to use a
2574 class which does not have actually enough regs to
2575 hold the value (e.g. x86 AREG for mode requiring
2576 more one general reg). Therefore we have 2
2577 conditions to check that the reload pseudo can
2578 not hold the mode value. */
2579 && ! HARD_REGNO_MODE_OK (ira_class_hard_regs
2580 [this_alternative][0],
2581 GET_MODE (*curr_id->operand_loc[nop]))
2582 /* The above condition is not enough as the first
2583 reg in ira_class_hard_regs can be not aligned for
2584 multi-words mode values. */
2585 && (prohibited_class_reg_set_mode_p
2586 (this_alternative, this_alternative_set,
2587 GET_MODE (*curr_id->operand_loc[nop]))))
2589 if (lra_dump_file != NULL)
2590 fprintf (lra_dump_file,
2591 " alt=%d: reload pseudo for op %d "
2592 " can not hold the mode value -- refuse\n",
2593 nalt, nop);
2594 goto fail;
2597 /* Check strong discouragement of reload of non-constant
2598 into class THIS_ALTERNATIVE. */
2599 if (! CONSTANT_P (op) && ! no_regs_p
2600 && (targetm.preferred_reload_class
2601 (op, this_alternative) == NO_REGS
2602 || (curr_static_id->operand[nop].type == OP_OUT
2603 && (targetm.preferred_output_reload_class
2604 (op, this_alternative) == NO_REGS))))
2606 if (lra_dump_file != NULL)
2607 fprintf (lra_dump_file,
2608 " %d Non-prefered reload: reject+=%d\n",
2609 nop, LRA_MAX_REJECT);
2610 reject += LRA_MAX_REJECT;
2613 if (! (MEM_P (op) && offmemok)
2614 && ! (const_to_mem && constmemok))
2616 /* We prefer to reload pseudos over reloading other
2617 things, since such reloads may be able to be
2618 eliminated later. So bump REJECT in other cases.
2619 Don't do this in the case where we are forcing a
2620 constant into memory and it will then win since
2621 we don't want to have a different alternative
2622 match then. */
2623 if (! (REG_P (op) && REGNO (op) >= FIRST_PSEUDO_REGISTER))
2625 if (lra_dump_file != NULL)
2626 fprintf
2627 (lra_dump_file,
2628 " %d Non-pseudo reload: reject+=2\n",
2629 nop);
2630 reject += 2;
2633 if (! no_regs_p)
2634 reload_nregs
2635 += ira_reg_class_max_nregs[this_alternative][mode];
2637 if (SMALL_REGISTER_CLASS_P (this_alternative))
2639 if (lra_dump_file != NULL)
2640 fprintf
2641 (lra_dump_file,
2642 " %d Small class reload: reject+=%d\n",
2643 nop, LRA_LOSER_COST_FACTOR / 2);
2644 reject += LRA_LOSER_COST_FACTOR / 2;
2648 /* We are trying to spill pseudo into memory. It is
2649 usually more costly than moving to a hard register
2650 although it might takes the same number of
2651 reloads.
2653 Non-pseudo spill may happen also. Suppose a target allows both
2654 register and memory in the operand constraint alternatives,
2655 then it's typical that an eliminable register has a substition
2656 of "base + offset" which can either be reloaded by a simple
2657 "new_reg <= base + offset" which will match the register
2658 constraint, or a similar reg addition followed by further spill
2659 to and reload from memory which will match the memory
2660 constraint, but this memory spill will be much more costly
2661 usually.
2663 Code below increases the reject for both pseudo and non-pseudo
2664 spill. */
2665 if (no_regs_p
2666 && !(MEM_P (op) && offmemok)
2667 && !(REG_P (op) && hard_regno[nop] < 0))
2669 if (lra_dump_file != NULL)
2670 fprintf
2671 (lra_dump_file,
2672 " %d Spill %spseudo into memory: reject+=3\n",
2673 nop, REG_P (op) ? "" : "Non-");
2674 reject += 3;
2675 if (VECTOR_MODE_P (mode))
2677 /* Spilling vectors into memory is usually more
2678 costly as they contain big values. */
2679 if (lra_dump_file != NULL)
2680 fprintf
2681 (lra_dump_file,
2682 " %d Spill vector pseudo: reject+=2\n",
2683 nop);
2684 reject += 2;
2688 #ifdef SECONDARY_MEMORY_NEEDED
2689 /* If reload requires moving value through secondary
2690 memory, it will need one more insn at least. */
2691 if (this_alternative != NO_REGS
2692 && REG_P (op) && (cl = get_reg_class (REGNO (op))) != NO_REGS
2693 && ((curr_static_id->operand[nop].type != OP_OUT
2694 && SECONDARY_MEMORY_NEEDED (cl, this_alternative,
2695 GET_MODE (op)))
2696 || (curr_static_id->operand[nop].type != OP_IN
2697 && SECONDARY_MEMORY_NEEDED (this_alternative, cl,
2698 GET_MODE (op)))))
2699 losers++;
2700 #endif
2701 /* Input reloads can be inherited more often than output
2702 reloads can be removed, so penalize output
2703 reloads. */
2704 if (!REG_P (op) || curr_static_id->operand[nop].type != OP_IN)
2706 if (lra_dump_file != NULL)
2707 fprintf
2708 (lra_dump_file,
2709 " %d Non input pseudo reload: reject++\n",
2710 nop);
2711 reject++;
2714 if (MEM_P (op) && offmemok)
2715 addr_losers++;
2718 if (early_clobber_p && ! scratch_p)
2720 if (lra_dump_file != NULL)
2721 fprintf (lra_dump_file,
2722 " %d Early clobber: reject++\n", nop);
2723 reject++;
2725 /* ??? We check early clobbers after processing all operands
2726 (see loop below) and there we update the costs more.
2727 Should we update the cost (may be approximately) here
2728 because of early clobber register reloads or it is a rare
2729 or non-important thing to be worth to do it. */
2730 overall = (losers * LRA_LOSER_COST_FACTOR + reject
2731 - (addr_losers == losers ? static_reject : 0));
2732 if ((best_losers == 0 || losers != 0) && best_overall < overall)
2734 if (lra_dump_file != NULL)
2735 fprintf (lra_dump_file,
2736 " alt=%d,overall=%d,losers=%d -- refuse\n",
2737 nalt, overall, losers);
2738 goto fail;
2741 curr_alt[nop] = this_alternative;
2742 COPY_HARD_REG_SET (curr_alt_set[nop], this_alternative_set);
2743 curr_alt_win[nop] = this_alternative_win;
2744 curr_alt_match_win[nop] = this_alternative_match_win;
2745 curr_alt_offmemok[nop] = this_alternative_offmemok;
2746 curr_alt_matches[nop] = this_alternative_matches;
2748 if (this_alternative_matches >= 0
2749 && !did_match && !this_alternative_win)
2750 curr_alt_win[this_alternative_matches] = false;
2752 if (early_clobber_p && operand_reg[nop] != NULL_RTX)
2753 early_clobbered_nops[early_clobbered_regs_num++] = nop;
2756 if (curr_insn_set != NULL_RTX && n_operands == 2
2757 /* Prevent processing non-move insns. */
2758 && (GET_CODE (SET_SRC (curr_insn_set)) == SUBREG
2759 || SET_SRC (curr_insn_set) == no_subreg_reg_operand[1])
2760 && ((! curr_alt_win[0] && ! curr_alt_win[1]
2761 && REG_P (no_subreg_reg_operand[0])
2762 && REG_P (no_subreg_reg_operand[1])
2763 && (reg_in_class_p (no_subreg_reg_operand[0], curr_alt[1])
2764 || reg_in_class_p (no_subreg_reg_operand[1], curr_alt[0])))
2765 || (! curr_alt_win[0] && curr_alt_win[1]
2766 && REG_P (no_subreg_reg_operand[1])
2767 /* Check that we reload memory not the memory
2768 address. */
2769 && ! (curr_alt_offmemok[0]
2770 && MEM_P (no_subreg_reg_operand[0]))
2771 && reg_in_class_p (no_subreg_reg_operand[1], curr_alt[0]))
2772 || (curr_alt_win[0] && ! curr_alt_win[1]
2773 && REG_P (no_subreg_reg_operand[0])
2774 /* Check that we reload memory not the memory
2775 address. */
2776 && ! (curr_alt_offmemok[1]
2777 && MEM_P (no_subreg_reg_operand[1]))
2778 && reg_in_class_p (no_subreg_reg_operand[0], curr_alt[1])
2779 && (! CONST_POOL_OK_P (curr_operand_mode[1],
2780 no_subreg_reg_operand[1])
2781 || (targetm.preferred_reload_class
2782 (no_subreg_reg_operand[1],
2783 (enum reg_class) curr_alt[1]) != NO_REGS))
2784 /* If it is a result of recent elimination in move
2785 insn we can transform it into an add still by
2786 using this alternative. */
2787 && GET_CODE (no_subreg_reg_operand[1]) != PLUS)))
2789 /* We have a move insn and a new reload insn will be similar
2790 to the current insn. We should avoid such situation as
2791 it results in LRA cycling. */
2792 if (lra_dump_file != NULL)
2793 fprintf (lra_dump_file,
2794 " Cycle danger: overall += LRA_MAX_REJECT\n");
2795 overall += LRA_MAX_REJECT;
2797 ok_p = true;
2798 curr_alt_dont_inherit_ops_num = 0;
2799 for (nop = 0; nop < early_clobbered_regs_num; nop++)
2801 int i, j, clobbered_hard_regno, first_conflict_j, last_conflict_j;
2802 HARD_REG_SET temp_set;
2804 i = early_clobbered_nops[nop];
2805 if ((! curr_alt_win[i] && ! curr_alt_match_win[i])
2806 || hard_regno[i] < 0)
2807 continue;
2808 lra_assert (operand_reg[i] != NULL_RTX);
2809 clobbered_hard_regno = hard_regno[i];
2810 CLEAR_HARD_REG_SET (temp_set);
2811 add_to_hard_reg_set (&temp_set, biggest_mode[i], clobbered_hard_regno);
2812 first_conflict_j = last_conflict_j = -1;
2813 for (j = 0; j < n_operands; j++)
2814 if (j == i
2815 /* We don't want process insides of match_operator and
2816 match_parallel because otherwise we would process
2817 their operands once again generating a wrong
2818 code. */
2819 || curr_static_id->operand[j].is_operator)
2820 continue;
2821 else if ((curr_alt_matches[j] == i && curr_alt_match_win[j])
2822 || (curr_alt_matches[i] == j && curr_alt_match_win[i]))
2823 continue;
2824 /* If we don't reload j-th operand, check conflicts. */
2825 else if ((curr_alt_win[j] || curr_alt_match_win[j])
2826 && uses_hard_regs_p (*curr_id->operand_loc[j], temp_set))
2828 if (first_conflict_j < 0)
2829 first_conflict_j = j;
2830 last_conflict_j = j;
2832 if (last_conflict_j < 0)
2833 continue;
2834 /* If earlyclobber operand conflicts with another
2835 non-matching operand which is actually the same register
2836 as the earlyclobber operand, it is better to reload the
2837 another operand as an operand matching the earlyclobber
2838 operand can be also the same. */
2839 if (first_conflict_j == last_conflict_j
2840 && operand_reg[last_conflict_j] != NULL_RTX
2841 && ! curr_alt_match_win[last_conflict_j]
2842 && REGNO (operand_reg[i]) == REGNO (operand_reg[last_conflict_j]))
2844 curr_alt_win[last_conflict_j] = false;
2845 curr_alt_dont_inherit_ops[curr_alt_dont_inherit_ops_num++]
2846 = last_conflict_j;
2847 losers++;
2848 /* Early clobber was already reflected in REJECT. */
2849 lra_assert (reject > 0);
2850 if (lra_dump_file != NULL)
2851 fprintf
2852 (lra_dump_file,
2853 " %d Conflict early clobber reload: reject--\n",
2855 reject--;
2856 overall += LRA_LOSER_COST_FACTOR - 1;
2858 else
2860 /* We need to reload early clobbered register and the
2861 matched registers. */
2862 for (j = 0; j < n_operands; j++)
2863 if (curr_alt_matches[j] == i)
2865 curr_alt_match_win[j] = false;
2866 losers++;
2867 overall += LRA_LOSER_COST_FACTOR;
2869 if (! curr_alt_match_win[i])
2870 curr_alt_dont_inherit_ops[curr_alt_dont_inherit_ops_num++] = i;
2871 else
2873 /* Remember pseudos used for match reloads are never
2874 inherited. */
2875 lra_assert (curr_alt_matches[i] >= 0);
2876 curr_alt_win[curr_alt_matches[i]] = false;
2878 curr_alt_win[i] = curr_alt_match_win[i] = false;
2879 losers++;
2880 /* Early clobber was already reflected in REJECT. */
2881 lra_assert (reject > 0);
2882 if (lra_dump_file != NULL)
2883 fprintf
2884 (lra_dump_file,
2885 " %d Matched conflict early clobber reloads: "
2886 "reject--\n",
2888 reject--;
2889 overall += LRA_LOSER_COST_FACTOR - 1;
2892 if (lra_dump_file != NULL)
2893 fprintf (lra_dump_file, " alt=%d,overall=%d,losers=%d,rld_nregs=%d\n",
2894 nalt, overall, losers, reload_nregs);
2896 /* If this alternative can be made to work by reloading, and it
2897 needs less reloading than the others checked so far, record
2898 it as the chosen goal for reloading. */
2899 if ((best_losers != 0 && losers == 0)
2900 || (((best_losers == 0 && losers == 0)
2901 || (best_losers != 0 && losers != 0))
2902 && (best_overall > overall
2903 || (best_overall == overall
2904 /* If the cost of the reloads is the same,
2905 prefer alternative which requires minimal
2906 number of reload regs. */
2907 && (reload_nregs < best_reload_nregs
2908 || (reload_nregs == best_reload_nregs
2909 && (best_reload_sum < reload_sum
2910 || (best_reload_sum == reload_sum
2911 && nalt < goal_alt_number))))))))
2913 for (nop = 0; nop < n_operands; nop++)
2915 goal_alt_win[nop] = curr_alt_win[nop];
2916 goal_alt_match_win[nop] = curr_alt_match_win[nop];
2917 goal_alt_matches[nop] = curr_alt_matches[nop];
2918 goal_alt[nop] = curr_alt[nop];
2919 goal_alt_offmemok[nop] = curr_alt_offmemok[nop];
2921 goal_alt_dont_inherit_ops_num = curr_alt_dont_inherit_ops_num;
2922 for (nop = 0; nop < curr_alt_dont_inherit_ops_num; nop++)
2923 goal_alt_dont_inherit_ops[nop] = curr_alt_dont_inherit_ops[nop];
2924 goal_alt_swapped = curr_swapped;
2925 best_overall = overall;
2926 best_losers = losers;
2927 best_reload_nregs = reload_nregs;
2928 best_reload_sum = reload_sum;
2929 goal_alt_number = nalt;
2931 if (losers == 0)
2932 /* Everything is satisfied. Do not process alternatives
2933 anymore. */
2934 break;
2935 fail:
2938 return ok_p;
2941 /* Make reload base reg from address AD. */
2942 static rtx
2943 base_to_reg (struct address_info *ad)
2945 enum reg_class cl;
2946 int code = -1;
2947 rtx new_inner = NULL_RTX;
2948 rtx new_reg = NULL_RTX;
2949 rtx_insn *insn;
2950 rtx_insn *last_insn = get_last_insn();
2952 lra_assert (ad->disp == ad->disp_term);
2953 cl = base_reg_class (ad->mode, ad->as, ad->base_outer_code,
2954 get_index_code (ad));
2955 new_reg = lra_create_new_reg (GET_MODE (*ad->base), NULL_RTX,
2956 cl, "base");
2957 new_inner = simplify_gen_binary (PLUS, GET_MODE (new_reg), new_reg,
2958 ad->disp_term == NULL
2959 ? const0_rtx
2960 : *ad->disp_term);
2961 if (!valid_address_p (ad->mode, new_inner, ad->as))
2962 return NULL_RTX;
2963 insn = emit_insn (gen_rtx_SET (new_reg, *ad->base));
2964 code = recog_memoized (insn);
2965 if (code < 0)
2967 delete_insns_since (last_insn);
2968 return NULL_RTX;
2971 return new_inner;
2974 /* Make reload base reg + disp from address AD. Return the new pseudo. */
2975 static rtx
2976 base_plus_disp_to_reg (struct address_info *ad)
2978 enum reg_class cl;
2979 rtx new_reg;
2981 lra_assert (ad->base == ad->base_term && ad->disp == ad->disp_term);
2982 cl = base_reg_class (ad->mode, ad->as, ad->base_outer_code,
2983 get_index_code (ad));
2984 new_reg = lra_create_new_reg (GET_MODE (*ad->base_term), NULL_RTX,
2985 cl, "base + disp");
2986 lra_emit_add (new_reg, *ad->base_term, *ad->disp_term);
2987 return new_reg;
2990 /* Make reload of index part of address AD. Return the new
2991 pseudo. */
2992 static rtx
2993 index_part_to_reg (struct address_info *ad)
2995 rtx new_reg;
2997 new_reg = lra_create_new_reg (GET_MODE (*ad->index), NULL_RTX,
2998 INDEX_REG_CLASS, "index term");
2999 expand_mult (GET_MODE (*ad->index), *ad->index_term,
3000 GEN_INT (get_index_scale (ad)), new_reg, 1);
3001 return new_reg;
3004 /* Return true if we can add a displacement to address AD, even if that
3005 makes the address invalid. The fix-up code requires any new address
3006 to be the sum of the BASE_TERM, INDEX and DISP_TERM fields. */
3007 static bool
3008 can_add_disp_p (struct address_info *ad)
3010 return (!ad->autoinc_p
3011 && ad->segment == NULL
3012 && ad->base == ad->base_term
3013 && ad->disp == ad->disp_term);
3016 /* Make equiv substitution in address AD. Return true if a substitution
3017 was made. */
3018 static bool
3019 equiv_address_substitution (struct address_info *ad)
3021 rtx base_reg, new_base_reg, index_reg, new_index_reg, *base_term, *index_term;
3022 HOST_WIDE_INT disp, scale;
3023 bool change_p;
3025 base_term = strip_subreg (ad->base_term);
3026 if (base_term == NULL)
3027 base_reg = new_base_reg = NULL_RTX;
3028 else
3030 base_reg = *base_term;
3031 new_base_reg = get_equiv_with_elimination (base_reg, curr_insn);
3033 index_term = strip_subreg (ad->index_term);
3034 if (index_term == NULL)
3035 index_reg = new_index_reg = NULL_RTX;
3036 else
3038 index_reg = *index_term;
3039 new_index_reg = get_equiv_with_elimination (index_reg, curr_insn);
3041 if (base_reg == new_base_reg && index_reg == new_index_reg)
3042 return false;
3043 disp = 0;
3044 change_p = false;
3045 if (lra_dump_file != NULL)
3047 fprintf (lra_dump_file, "Changing address in insn %d ",
3048 INSN_UID (curr_insn));
3049 dump_value_slim (lra_dump_file, *ad->outer, 1);
3051 if (base_reg != new_base_reg)
3053 if (REG_P (new_base_reg))
3055 *base_term = new_base_reg;
3056 change_p = true;
3058 else if (GET_CODE (new_base_reg) == PLUS
3059 && REG_P (XEXP (new_base_reg, 0))
3060 && CONST_INT_P (XEXP (new_base_reg, 1))
3061 && can_add_disp_p (ad))
3063 disp += INTVAL (XEXP (new_base_reg, 1));
3064 *base_term = XEXP (new_base_reg, 0);
3065 change_p = true;
3067 if (ad->base_term2 != NULL)
3068 *ad->base_term2 = *ad->base_term;
3070 if (index_reg != new_index_reg)
3072 if (REG_P (new_index_reg))
3074 *index_term = new_index_reg;
3075 change_p = true;
3077 else if (GET_CODE (new_index_reg) == PLUS
3078 && REG_P (XEXP (new_index_reg, 0))
3079 && CONST_INT_P (XEXP (new_index_reg, 1))
3080 && can_add_disp_p (ad)
3081 && (scale = get_index_scale (ad)))
3083 disp += INTVAL (XEXP (new_index_reg, 1)) * scale;
3084 *index_term = XEXP (new_index_reg, 0);
3085 change_p = true;
3088 if (disp != 0)
3090 if (ad->disp != NULL)
3091 *ad->disp = plus_constant (GET_MODE (*ad->inner), *ad->disp, disp);
3092 else
3094 *ad->inner = plus_constant (GET_MODE (*ad->inner), *ad->inner, disp);
3095 update_address (ad);
3097 change_p = true;
3099 if (lra_dump_file != NULL)
3101 if (! change_p)
3102 fprintf (lra_dump_file, " -- no change\n");
3103 else
3105 fprintf (lra_dump_file, " on equiv ");
3106 dump_value_slim (lra_dump_file, *ad->outer, 1);
3107 fprintf (lra_dump_file, "\n");
3110 return change_p;
3113 /* Major function to make reloads for an address in operand NOP or
3114 check its correctness (If CHECK_ONLY_P is true). The supported
3115 cases are:
3117 1) an address that existed before LRA started, at which point it
3118 must have been valid. These addresses are subject to elimination
3119 and may have become invalid due to the elimination offset being out
3120 of range.
3122 2) an address created by forcing a constant to memory
3123 (force_const_to_mem). The initial form of these addresses might
3124 not be valid, and it is this function's job to make them valid.
3126 3) a frame address formed from a register and a (possibly zero)
3127 constant offset. As above, these addresses might not be valid and
3128 this function must make them so.
3130 Add reloads to the lists *BEFORE and *AFTER. We might need to add
3131 reloads to *AFTER because of inc/dec, {pre, post} modify in the
3132 address. Return true for any RTL change.
3134 The function is a helper function which does not produce all
3135 transformations (when CHECK_ONLY_P is false) which can be
3136 necessary. It does just basic steps. To do all necessary
3137 transformations use function process_address. */
3138 static bool
3139 process_address_1 (int nop, bool check_only_p,
3140 rtx_insn **before, rtx_insn **after)
3142 struct address_info ad;
3143 rtx new_reg;
3144 HOST_WIDE_INT scale;
3145 rtx op = *curr_id->operand_loc[nop];
3146 const char *constraint = curr_static_id->operand[nop].constraint;
3147 enum constraint_num cn = lookup_constraint (constraint);
3148 bool change_p = false;
3150 if (MEM_P (op)
3151 && GET_MODE (op) == BLKmode
3152 && GET_CODE (XEXP (op, 0)) == SCRATCH)
3153 return false;
3155 if (insn_extra_address_constraint (cn))
3156 decompose_lea_address (&ad, curr_id->operand_loc[nop]);
3157 /* Do not attempt to decompose arbitrary addresses generated by combine
3158 for asm operands with loose constraints, e.g 'X'. */
3159 else if (MEM_P (op)
3160 && !(get_constraint_type (cn) == CT_FIXED_FORM
3161 && constraint_satisfied_p (op, cn)))
3162 decompose_mem_address (&ad, op);
3163 else if (GET_CODE (op) == SUBREG
3164 && MEM_P (SUBREG_REG (op)))
3165 decompose_mem_address (&ad, SUBREG_REG (op));
3166 else
3167 return false;
3168 /* If INDEX_REG_CLASS is assigned to base_term already and isn't to
3169 index_term, swap them so to avoid assigning INDEX_REG_CLASS to both
3170 when INDEX_REG_CLASS is a single register class. */
3171 if (ad.base_term != NULL
3172 && ad.index_term != NULL
3173 && ira_class_hard_regs_num[INDEX_REG_CLASS] == 1
3174 && REG_P (*ad.base_term)
3175 && REG_P (*ad.index_term)
3176 && in_class_p (*ad.base_term, INDEX_REG_CLASS, NULL)
3177 && ! in_class_p (*ad.index_term, INDEX_REG_CLASS, NULL))
3179 std::swap (ad.base, ad.index);
3180 std::swap (ad.base_term, ad.index_term);
3182 if (! check_only_p)
3183 change_p = equiv_address_substitution (&ad);
3184 if (ad.base_term != NULL
3185 && (process_addr_reg
3186 (ad.base_term, check_only_p, before,
3187 (ad.autoinc_p
3188 && !(REG_P (*ad.base_term)
3189 && find_regno_note (curr_insn, REG_DEAD,
3190 REGNO (*ad.base_term)) != NULL_RTX)
3191 ? after : NULL),
3192 base_reg_class (ad.mode, ad.as, ad.base_outer_code,
3193 get_index_code (&ad)))))
3195 change_p = true;
3196 if (ad.base_term2 != NULL)
3197 *ad.base_term2 = *ad.base_term;
3199 if (ad.index_term != NULL
3200 && process_addr_reg (ad.index_term, check_only_p,
3201 before, NULL, INDEX_REG_CLASS))
3202 change_p = true;
3204 /* Target hooks sometimes don't treat extra-constraint addresses as
3205 legitimate address_operands, so handle them specially. */
3206 if (insn_extra_address_constraint (cn)
3207 && satisfies_address_constraint_p (&ad, cn))
3208 return change_p;
3210 if (check_only_p)
3211 return change_p;
3213 /* There are three cases where the shape of *AD.INNER may now be invalid:
3215 1) the original address was valid, but either elimination or
3216 equiv_address_substitution was applied and that made
3217 the address invalid.
3219 2) the address is an invalid symbolic address created by
3220 force_const_to_mem.
3222 3) the address is a frame address with an invalid offset.
3224 4) the address is a frame address with an invalid base.
3226 All these cases involve a non-autoinc address, so there is no
3227 point revalidating other types. */
3228 if (ad.autoinc_p || valid_address_p (&ad))
3229 return change_p;
3231 /* Any index existed before LRA started, so we can assume that the
3232 presence and shape of the index is valid. */
3233 push_to_sequence (*before);
3234 lra_assert (ad.disp == ad.disp_term);
3235 if (ad.base == NULL)
3237 if (ad.index == NULL)
3239 rtx_insn *insn;
3240 rtx_insn *last = get_last_insn ();
3241 int code = -1;
3242 enum reg_class cl = base_reg_class (ad.mode, ad.as,
3243 SCRATCH, SCRATCH);
3244 rtx addr = *ad.inner;
3246 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "addr");
3247 if (HAVE_lo_sum)
3249 /* addr => lo_sum (new_base, addr), case (2) above. */
3250 insn = emit_insn (gen_rtx_SET
3251 (new_reg,
3252 gen_rtx_HIGH (Pmode, copy_rtx (addr))));
3253 code = recog_memoized (insn);
3254 if (code >= 0)
3256 *ad.inner = gen_rtx_LO_SUM (Pmode, new_reg, addr);
3257 if (! valid_address_p (ad.mode, *ad.outer, ad.as))
3259 /* Try to put lo_sum into register. */
3260 insn = emit_insn (gen_rtx_SET
3261 (new_reg,
3262 gen_rtx_LO_SUM (Pmode, new_reg, addr)));
3263 code = recog_memoized (insn);
3264 if (code >= 0)
3266 *ad.inner = new_reg;
3267 if (! valid_address_p (ad.mode, *ad.outer, ad.as))
3269 *ad.inner = addr;
3270 code = -1;
3276 if (code < 0)
3277 delete_insns_since (last);
3280 if (code < 0)
3282 /* addr => new_base, case (2) above. */
3283 lra_emit_move (new_reg, addr);
3285 for (insn = last == NULL_RTX ? get_insns () : NEXT_INSN (last);
3286 insn != NULL_RTX;
3287 insn = NEXT_INSN (insn))
3288 if (recog_memoized (insn) < 0)
3289 break;
3290 if (insn != NULL_RTX)
3292 /* Do nothing if we cannot generate right insns.
3293 This is analogous to reload pass behavior. */
3294 delete_insns_since (last);
3295 end_sequence ();
3296 return false;
3298 *ad.inner = new_reg;
3301 else
3303 /* index * scale + disp => new base + index * scale,
3304 case (1) above. */
3305 enum reg_class cl = base_reg_class (ad.mode, ad.as, PLUS,
3306 GET_CODE (*ad.index));
3308 lra_assert (INDEX_REG_CLASS != NO_REGS);
3309 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "disp");
3310 lra_emit_move (new_reg, *ad.disp);
3311 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3312 new_reg, *ad.index);
3315 else if (ad.index == NULL)
3317 int regno;
3318 enum reg_class cl;
3319 rtx set;
3320 rtx_insn *insns, *last_insn;
3321 /* Try to reload base into register only if the base is invalid
3322 for the address but with valid offset, case (4) above. */
3323 start_sequence ();
3324 new_reg = base_to_reg (&ad);
3326 /* base + disp => new base, cases (1) and (3) above. */
3327 /* Another option would be to reload the displacement into an
3328 index register. However, postreload has code to optimize
3329 address reloads that have the same base and different
3330 displacements, so reloading into an index register would
3331 not necessarily be a win. */
3332 if (new_reg == NULL_RTX)
3333 new_reg = base_plus_disp_to_reg (&ad);
3334 insns = get_insns ();
3335 last_insn = get_last_insn ();
3336 /* If we generated at least two insns, try last insn source as
3337 an address. If we succeed, we generate one less insn. */
3338 if (last_insn != insns && (set = single_set (last_insn)) != NULL_RTX
3339 && GET_CODE (SET_SRC (set)) == PLUS
3340 && REG_P (XEXP (SET_SRC (set), 0))
3341 && CONSTANT_P (XEXP (SET_SRC (set), 1)))
3343 *ad.inner = SET_SRC (set);
3344 if (valid_address_p (ad.mode, *ad.outer, ad.as))
3346 *ad.base_term = XEXP (SET_SRC (set), 0);
3347 *ad.disp_term = XEXP (SET_SRC (set), 1);
3348 cl = base_reg_class (ad.mode, ad.as, ad.base_outer_code,
3349 get_index_code (&ad));
3350 regno = REGNO (*ad.base_term);
3351 if (regno >= FIRST_PSEUDO_REGISTER
3352 && cl != lra_get_allocno_class (regno))
3353 lra_change_class (regno, cl, " Change to", true);
3354 new_reg = SET_SRC (set);
3355 delete_insns_since (PREV_INSN (last_insn));
3358 /* Try if target can split displacement into legitimite new disp
3359 and offset. If it's the case, we replace the last insn with
3360 insns for base + offset => new_reg and set new_reg + new disp
3361 to *ad.inner. */
3362 last_insn = get_last_insn ();
3363 if ((set = single_set (last_insn)) != NULL_RTX
3364 && GET_CODE (SET_SRC (set)) == PLUS
3365 && REG_P (XEXP (SET_SRC (set), 0))
3366 && REGNO (XEXP (SET_SRC (set), 0)) < FIRST_PSEUDO_REGISTER
3367 && CONST_INT_P (XEXP (SET_SRC (set), 1)))
3369 rtx addend, disp = XEXP (SET_SRC (set), 1);
3370 if (targetm.legitimize_address_displacement (&disp, &addend,
3371 ad.mode))
3373 rtx_insn *new_insns;
3374 start_sequence ();
3375 lra_emit_add (new_reg, XEXP (SET_SRC (set), 0), addend);
3376 new_insns = get_insns ();
3377 end_sequence ();
3378 new_reg = gen_rtx_PLUS (Pmode, new_reg, disp);
3379 delete_insns_since (PREV_INSN (last_insn));
3380 add_insn (new_insns);
3381 insns = get_insns ();
3384 end_sequence ();
3385 emit_insn (insns);
3386 *ad.inner = new_reg;
3388 else if (ad.disp_term != NULL)
3390 /* base + scale * index + disp => new base + scale * index,
3391 case (1) above. */
3392 new_reg = base_plus_disp_to_reg (&ad);
3393 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3394 new_reg, *ad.index);
3396 else if ((scale = get_index_scale (&ad)) == 1)
3398 /* The last transformation to one reg will be made in
3399 curr_insn_transform function. */
3400 end_sequence ();
3401 return false;
3403 else if (scale != 0)
3405 /* base + scale * index => base + new_reg,
3406 case (1) above.
3407 Index part of address may become invalid. For example, we
3408 changed pseudo on the equivalent memory and a subreg of the
3409 pseudo onto the memory of different mode for which the scale is
3410 prohibitted. */
3411 new_reg = index_part_to_reg (&ad);
3412 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3413 *ad.base_term, new_reg);
3415 else
3417 enum reg_class cl = base_reg_class (ad.mode, ad.as,
3418 SCRATCH, SCRATCH);
3419 rtx addr = *ad.inner;
3421 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "addr");
3422 /* addr => new_base. */
3423 lra_emit_move (new_reg, addr);
3424 *ad.inner = new_reg;
3426 *before = get_insns ();
3427 end_sequence ();
3428 return true;
3431 /* If CHECK_ONLY_P is false, do address reloads until it is necessary.
3432 Use process_address_1 as a helper function. Return true for any
3433 RTL changes.
3435 If CHECK_ONLY_P is true, just check address correctness. Return
3436 false if the address correct. */
3437 static bool
3438 process_address (int nop, bool check_only_p,
3439 rtx_insn **before, rtx_insn **after)
3441 bool res = false;
3443 while (process_address_1 (nop, check_only_p, before, after))
3445 if (check_only_p)
3446 return true;
3447 res = true;
3449 return res;
3452 /* Emit insns to reload VALUE into a new register. VALUE is an
3453 auto-increment or auto-decrement RTX whose operand is a register or
3454 memory location; so reloading involves incrementing that location.
3455 IN is either identical to VALUE, or some cheaper place to reload
3456 value being incremented/decremented from.
3458 INC_AMOUNT is the number to increment or decrement by (always
3459 positive and ignored for POST_MODIFY/PRE_MODIFY).
3461 Return pseudo containing the result. */
3462 static rtx
3463 emit_inc (enum reg_class new_rclass, rtx in, rtx value, int inc_amount)
3465 /* REG or MEM to be copied and incremented. */
3466 rtx incloc = XEXP (value, 0);
3467 /* Nonzero if increment after copying. */
3468 int post = (GET_CODE (value) == POST_DEC || GET_CODE (value) == POST_INC
3469 || GET_CODE (value) == POST_MODIFY);
3470 rtx_insn *last;
3471 rtx inc;
3472 rtx_insn *add_insn;
3473 int code;
3474 rtx real_in = in == value ? incloc : in;
3475 rtx result;
3476 bool plus_p = true;
3478 if (GET_CODE (value) == PRE_MODIFY || GET_CODE (value) == POST_MODIFY)
3480 lra_assert (GET_CODE (XEXP (value, 1)) == PLUS
3481 || GET_CODE (XEXP (value, 1)) == MINUS);
3482 lra_assert (rtx_equal_p (XEXP (XEXP (value, 1), 0), XEXP (value, 0)));
3483 plus_p = GET_CODE (XEXP (value, 1)) == PLUS;
3484 inc = XEXP (XEXP (value, 1), 1);
3486 else
3488 if (GET_CODE (value) == PRE_DEC || GET_CODE (value) == POST_DEC)
3489 inc_amount = -inc_amount;
3491 inc = GEN_INT (inc_amount);
3494 if (! post && REG_P (incloc))
3495 result = incloc;
3496 else
3497 result = lra_create_new_reg (GET_MODE (value), value, new_rclass,
3498 "INC/DEC result");
3500 if (real_in != result)
3502 /* First copy the location to the result register. */
3503 lra_assert (REG_P (result));
3504 emit_insn (gen_move_insn (result, real_in));
3507 /* We suppose that there are insns to add/sub with the constant
3508 increment permitted in {PRE/POST)_{DEC/INC/MODIFY}. At least the
3509 old reload worked with this assumption. If the assumption
3510 becomes wrong, we should use approach in function
3511 base_plus_disp_to_reg. */
3512 if (in == value)
3514 /* See if we can directly increment INCLOC. */
3515 last = get_last_insn ();
3516 add_insn = emit_insn (plus_p
3517 ? gen_add2_insn (incloc, inc)
3518 : gen_sub2_insn (incloc, inc));
3520 code = recog_memoized (add_insn);
3521 if (code >= 0)
3523 if (! post && result != incloc)
3524 emit_insn (gen_move_insn (result, incloc));
3525 return result;
3527 delete_insns_since (last);
3530 /* If couldn't do the increment directly, must increment in RESULT.
3531 The way we do this depends on whether this is pre- or
3532 post-increment. For pre-increment, copy INCLOC to the reload
3533 register, increment it there, then save back. */
3534 if (! post)
3536 if (real_in != result)
3537 emit_insn (gen_move_insn (result, real_in));
3538 if (plus_p)
3539 emit_insn (gen_add2_insn (result, inc));
3540 else
3541 emit_insn (gen_sub2_insn (result, inc));
3542 if (result != incloc)
3543 emit_insn (gen_move_insn (incloc, result));
3545 else
3547 /* Post-increment.
3549 Because this might be a jump insn or a compare, and because
3550 RESULT may not be available after the insn in an input
3551 reload, we must do the incrementing before the insn being
3552 reloaded for.
3554 We have already copied IN to RESULT. Increment the copy in
3555 RESULT, save that back, then decrement RESULT so it has
3556 the original value. */
3557 if (plus_p)
3558 emit_insn (gen_add2_insn (result, inc));
3559 else
3560 emit_insn (gen_sub2_insn (result, inc));
3561 emit_insn (gen_move_insn (incloc, result));
3562 /* Restore non-modified value for the result. We prefer this
3563 way because it does not require an additional hard
3564 register. */
3565 if (plus_p)
3567 if (CONST_INT_P (inc))
3568 emit_insn (gen_add2_insn (result,
3569 gen_int_mode (-INTVAL (inc),
3570 GET_MODE (result))));
3571 else
3572 emit_insn (gen_sub2_insn (result, inc));
3574 else
3575 emit_insn (gen_add2_insn (result, inc));
3577 return result;
3580 /* Return true if the current move insn does not need processing as we
3581 already know that it satisfies its constraints. */
3582 static bool
3583 simple_move_p (void)
3585 rtx dest, src;
3586 enum reg_class dclass, sclass;
3588 lra_assert (curr_insn_set != NULL_RTX);
3589 dest = SET_DEST (curr_insn_set);
3590 src = SET_SRC (curr_insn_set);
3592 /* If the instruction has multiple sets we need to process it even if it
3593 is single_set. This can happen if one or more of the SETs are dead.
3594 See PR73650. */
3595 if (multiple_sets (curr_insn))
3596 return false;
3598 return ((dclass = get_op_class (dest)) != NO_REGS
3599 && (sclass = get_op_class (src)) != NO_REGS
3600 /* The backend guarantees that register moves of cost 2
3601 never need reloads. */
3602 && targetm.register_move_cost (GET_MODE (src), sclass, dclass) == 2);
3605 /* Swap operands NOP and NOP + 1. */
3606 static inline void
3607 swap_operands (int nop)
3609 std::swap (curr_operand_mode[nop], curr_operand_mode[nop + 1]);
3610 std::swap (original_subreg_reg_mode[nop], original_subreg_reg_mode[nop + 1]);
3611 std::swap (*curr_id->operand_loc[nop], *curr_id->operand_loc[nop + 1]);
3612 std::swap (equiv_substition_p[nop], equiv_substition_p[nop + 1]);
3613 /* Swap the duplicates too. */
3614 lra_update_dup (curr_id, nop);
3615 lra_update_dup (curr_id, nop + 1);
3618 /* Main entry point of the constraint code: search the body of the
3619 current insn to choose the best alternative. It is mimicking insn
3620 alternative cost calculation model of former reload pass. That is
3621 because machine descriptions were written to use this model. This
3622 model can be changed in future. Make commutative operand exchange
3623 if it is chosen.
3625 if CHECK_ONLY_P is false, do RTL changes to satisfy the
3626 constraints. Return true if any change happened during function
3627 call.
3629 If CHECK_ONLY_P is true then don't do any transformation. Just
3630 check that the insn satisfies all constraints. If the insn does
3631 not satisfy any constraint, return true. */
3632 static bool
3633 curr_insn_transform (bool check_only_p)
3635 int i, j, k;
3636 int n_operands;
3637 int n_alternatives;
3638 int n_outputs;
3639 int commutative;
3640 signed char goal_alt_matched[MAX_RECOG_OPERANDS][MAX_RECOG_OPERANDS];
3641 signed char match_inputs[MAX_RECOG_OPERANDS + 1];
3642 signed char outputs[MAX_RECOG_OPERANDS + 1];
3643 rtx_insn *before, *after;
3644 bool alt_p = false;
3645 /* Flag that the insn has been changed through a transformation. */
3646 bool change_p;
3647 bool sec_mem_p;
3648 #ifdef SECONDARY_MEMORY_NEEDED
3649 bool use_sec_mem_p;
3650 #endif
3651 int max_regno_before;
3652 int reused_alternative_num;
3654 curr_insn_set = single_set (curr_insn);
3655 if (curr_insn_set != NULL_RTX && simple_move_p ())
3656 return false;
3658 no_input_reloads_p = no_output_reloads_p = false;
3659 goal_alt_number = -1;
3660 change_p = sec_mem_p = false;
3661 /* JUMP_INSNs and CALL_INSNs are not allowed to have any output
3662 reloads; neither are insns that SET cc0. Insns that use CC0 are
3663 not allowed to have any input reloads. */
3664 if (JUMP_P (curr_insn) || CALL_P (curr_insn))
3665 no_output_reloads_p = true;
3667 if (HAVE_cc0 && reg_referenced_p (cc0_rtx, PATTERN (curr_insn)))
3668 no_input_reloads_p = true;
3669 if (HAVE_cc0 && reg_set_p (cc0_rtx, PATTERN (curr_insn)))
3670 no_output_reloads_p = true;
3672 n_operands = curr_static_id->n_operands;
3673 n_alternatives = curr_static_id->n_alternatives;
3675 /* Just return "no reloads" if insn has no operands with
3676 constraints. */
3677 if (n_operands == 0 || n_alternatives == 0)
3678 return false;
3680 max_regno_before = max_reg_num ();
3682 for (i = 0; i < n_operands; i++)
3684 goal_alt_matched[i][0] = -1;
3685 goal_alt_matches[i] = -1;
3688 commutative = curr_static_id->commutative;
3690 /* Now see what we need for pseudos that didn't get hard regs or got
3691 the wrong kind of hard reg. For this, we must consider all the
3692 operands together against the register constraints. */
3694 best_losers = best_overall = INT_MAX;
3695 best_reload_sum = 0;
3697 curr_swapped = false;
3698 goal_alt_swapped = false;
3700 if (! check_only_p)
3701 /* Make equivalence substitution and memory subreg elimination
3702 before address processing because an address legitimacy can
3703 depend on memory mode. */
3704 for (i = 0; i < n_operands; i++)
3706 rtx op, subst, old;
3707 bool op_change_p = false;
3709 if (curr_static_id->operand[i].is_operator)
3710 continue;
3712 old = op = *curr_id->operand_loc[i];
3713 if (GET_CODE (old) == SUBREG)
3714 old = SUBREG_REG (old);
3715 subst = get_equiv_with_elimination (old, curr_insn);
3716 original_subreg_reg_mode[i] = VOIDmode;
3717 equiv_substition_p[i] = false;
3718 if (subst != old)
3720 equiv_substition_p[i] = true;
3721 subst = copy_rtx (subst);
3722 lra_assert (REG_P (old));
3723 if (GET_CODE (op) != SUBREG)
3724 *curr_id->operand_loc[i] = subst;
3725 else
3727 SUBREG_REG (op) = subst;
3728 if (GET_MODE (subst) == VOIDmode)
3729 original_subreg_reg_mode[i] = GET_MODE (old);
3731 if (lra_dump_file != NULL)
3733 fprintf (lra_dump_file,
3734 "Changing pseudo %d in operand %i of insn %u on equiv ",
3735 REGNO (old), i, INSN_UID (curr_insn));
3736 dump_value_slim (lra_dump_file, subst, 1);
3737 fprintf (lra_dump_file, "\n");
3739 op_change_p = change_p = true;
3741 if (simplify_operand_subreg (i, GET_MODE (old)) || op_change_p)
3743 change_p = true;
3744 lra_update_dup (curr_id, i);
3748 /* Reload address registers and displacements. We do it before
3749 finding an alternative because of memory constraints. */
3750 before = after = NULL;
3751 for (i = 0; i < n_operands; i++)
3752 if (! curr_static_id->operand[i].is_operator
3753 && process_address (i, check_only_p, &before, &after))
3755 if (check_only_p)
3756 return true;
3757 change_p = true;
3758 lra_update_dup (curr_id, i);
3761 if (change_p)
3762 /* If we've changed the instruction then any alternative that
3763 we chose previously may no longer be valid. */
3764 lra_set_used_insn_alternative (curr_insn, -1);
3766 if (! check_only_p && curr_insn_set != NULL_RTX
3767 && check_and_process_move (&change_p, &sec_mem_p))
3768 return change_p;
3770 try_swapped:
3772 reused_alternative_num = check_only_p ? -1 : curr_id->used_insn_alternative;
3773 if (lra_dump_file != NULL && reused_alternative_num >= 0)
3774 fprintf (lra_dump_file, "Reusing alternative %d for insn #%u\n",
3775 reused_alternative_num, INSN_UID (curr_insn));
3777 if (process_alt_operands (reused_alternative_num))
3778 alt_p = true;
3780 if (check_only_p)
3781 return ! alt_p || best_losers != 0;
3783 /* If insn is commutative (it's safe to exchange a certain pair of
3784 operands) then we need to try each alternative twice, the second
3785 time matching those two operands as if we had exchanged them. To
3786 do this, really exchange them in operands.
3788 If we have just tried the alternatives the second time, return
3789 operands to normal and drop through. */
3791 if (reused_alternative_num < 0 && commutative >= 0)
3793 curr_swapped = !curr_swapped;
3794 if (curr_swapped)
3796 swap_operands (commutative);
3797 goto try_swapped;
3799 else
3800 swap_operands (commutative);
3803 if (! alt_p && ! sec_mem_p)
3805 /* No alternative works with reloads?? */
3806 if (INSN_CODE (curr_insn) >= 0)
3807 fatal_insn ("unable to generate reloads for:", curr_insn);
3808 error_for_asm (curr_insn,
3809 "inconsistent operand constraints in an %<asm%>");
3810 /* Avoid further trouble with this insn. Don't generate use
3811 pattern here as we could use the insn SP offset. */
3812 lra_set_insn_deleted (curr_insn);
3813 return true;
3816 /* If the best alternative is with operands 1 and 2 swapped, swap
3817 them. Update the operand numbers of any reloads already
3818 pushed. */
3820 if (goal_alt_swapped)
3822 if (lra_dump_file != NULL)
3823 fprintf (lra_dump_file, " Commutative operand exchange in insn %u\n",
3824 INSN_UID (curr_insn));
3826 /* Swap the duplicates too. */
3827 swap_operands (commutative);
3828 change_p = true;
3831 #ifdef SECONDARY_MEMORY_NEEDED
3832 /* Some target macros SECONDARY_MEMORY_NEEDED (e.g. x86) are defined
3833 too conservatively. So we use the secondary memory only if there
3834 is no any alternative without reloads. */
3835 use_sec_mem_p = false;
3836 if (! alt_p)
3837 use_sec_mem_p = true;
3838 else if (sec_mem_p)
3840 for (i = 0; i < n_operands; i++)
3841 if (! goal_alt_win[i] && ! goal_alt_match_win[i])
3842 break;
3843 use_sec_mem_p = i < n_operands;
3846 if (use_sec_mem_p)
3848 int in = -1, out = -1;
3849 rtx new_reg, src, dest, rld;
3850 machine_mode sec_mode, rld_mode;
3852 lra_assert (curr_insn_set != NULL_RTX && sec_mem_p);
3853 dest = SET_DEST (curr_insn_set);
3854 src = SET_SRC (curr_insn_set);
3855 for (i = 0; i < n_operands; i++)
3856 if (*curr_id->operand_loc[i] == dest)
3857 out = i;
3858 else if (*curr_id->operand_loc[i] == src)
3859 in = i;
3860 for (i = 0; i < curr_static_id->n_dups; i++)
3861 if (out < 0 && *curr_id->dup_loc[i] == dest)
3862 out = curr_static_id->dup_num[i];
3863 else if (in < 0 && *curr_id->dup_loc[i] == src)
3864 in = curr_static_id->dup_num[i];
3865 lra_assert (out >= 0 && in >= 0
3866 && curr_static_id->operand[out].type == OP_OUT
3867 && curr_static_id->operand[in].type == OP_IN);
3868 rld = (GET_MODE_SIZE (GET_MODE (dest)) <= GET_MODE_SIZE (GET_MODE (src))
3869 ? dest : src);
3870 rld_mode = GET_MODE (rld);
3871 #ifdef SECONDARY_MEMORY_NEEDED_MODE
3872 sec_mode = SECONDARY_MEMORY_NEEDED_MODE (rld_mode);
3873 #else
3874 sec_mode = rld_mode;
3875 #endif
3876 new_reg = lra_create_new_reg (sec_mode, NULL_RTX,
3877 NO_REGS, "secondary");
3878 /* If the mode is changed, it should be wider. */
3879 lra_assert (GET_MODE_SIZE (sec_mode) >= GET_MODE_SIZE (rld_mode));
3880 if (sec_mode != rld_mode)
3882 /* If the target says specifically to use another mode for
3883 secondary memory moves we can not reuse the original
3884 insn. */
3885 after = emit_spill_move (false, new_reg, dest);
3886 lra_process_new_insns (curr_insn, NULL, after,
3887 "Inserting the sec. move");
3888 /* We may have non null BEFORE here (e.g. after address
3889 processing. */
3890 push_to_sequence (before);
3891 before = emit_spill_move (true, new_reg, src);
3892 emit_insn (before);
3893 before = get_insns ();
3894 end_sequence ();
3895 lra_process_new_insns (curr_insn, before, NULL, "Changing on");
3896 lra_set_insn_deleted (curr_insn);
3898 else if (dest == rld)
3900 *curr_id->operand_loc[out] = new_reg;
3901 lra_update_dup (curr_id, out);
3902 after = emit_spill_move (false, new_reg, dest);
3903 lra_process_new_insns (curr_insn, NULL, after,
3904 "Inserting the sec. move");
3906 else
3908 *curr_id->operand_loc[in] = new_reg;
3909 lra_update_dup (curr_id, in);
3910 /* See comments above. */
3911 push_to_sequence (before);
3912 before = emit_spill_move (true, new_reg, src);
3913 emit_insn (before);
3914 before = get_insns ();
3915 end_sequence ();
3916 lra_process_new_insns (curr_insn, before, NULL,
3917 "Inserting the sec. move");
3919 lra_update_insn_regno_info (curr_insn);
3920 return true;
3922 #endif
3924 lra_assert (goal_alt_number >= 0);
3925 lra_set_used_insn_alternative (curr_insn, goal_alt_number);
3927 if (lra_dump_file != NULL)
3929 const char *p;
3931 fprintf (lra_dump_file, " Choosing alt %d in insn %u:",
3932 goal_alt_number, INSN_UID (curr_insn));
3933 for (i = 0; i < n_operands; i++)
3935 p = (curr_static_id->operand_alternative
3936 [goal_alt_number * n_operands + i].constraint);
3937 if (*p == '\0')
3938 continue;
3939 fprintf (lra_dump_file, " (%d) ", i);
3940 for (; *p != '\0' && *p != ',' && *p != '#'; p++)
3941 fputc (*p, lra_dump_file);
3943 if (INSN_CODE (curr_insn) >= 0
3944 && (p = get_insn_name (INSN_CODE (curr_insn))) != NULL)
3945 fprintf (lra_dump_file, " {%s}", p);
3946 if (curr_id->sp_offset != 0)
3947 fprintf (lra_dump_file, " (sp_off=%" HOST_WIDE_INT_PRINT "d)",
3948 curr_id->sp_offset);
3949 fprintf (lra_dump_file, "\n");
3952 /* Right now, for any pair of operands I and J that are required to
3953 match, with J < I, goal_alt_matches[I] is J. Add I to
3954 goal_alt_matched[J]. */
3956 for (i = 0; i < n_operands; i++)
3957 if ((j = goal_alt_matches[i]) >= 0)
3959 for (k = 0; goal_alt_matched[j][k] >= 0; k++)
3961 /* We allow matching one output operand and several input
3962 operands. */
3963 lra_assert (k == 0
3964 || (curr_static_id->operand[j].type == OP_OUT
3965 && curr_static_id->operand[i].type == OP_IN
3966 && (curr_static_id->operand
3967 [goal_alt_matched[j][0]].type == OP_IN)));
3968 goal_alt_matched[j][k] = i;
3969 goal_alt_matched[j][k + 1] = -1;
3972 for (i = 0; i < n_operands; i++)
3973 goal_alt_win[i] |= goal_alt_match_win[i];
3975 /* Any constants that aren't allowed and can't be reloaded into
3976 registers are here changed into memory references. */
3977 for (i = 0; i < n_operands; i++)
3978 if (goal_alt_win[i])
3980 int regno;
3981 enum reg_class new_class;
3982 rtx reg = *curr_id->operand_loc[i];
3984 if (GET_CODE (reg) == SUBREG)
3985 reg = SUBREG_REG (reg);
3987 if (REG_P (reg) && (regno = REGNO (reg)) >= FIRST_PSEUDO_REGISTER)
3989 bool ok_p = in_class_p (reg, goal_alt[i], &new_class);
3991 if (new_class != NO_REGS && get_reg_class (regno) != new_class)
3993 lra_assert (ok_p);
3994 lra_change_class (regno, new_class, " Change to", true);
3998 else
4000 const char *constraint;
4001 char c;
4002 rtx op = *curr_id->operand_loc[i];
4003 rtx subreg = NULL_RTX;
4004 machine_mode mode = curr_operand_mode[i];
4006 if (GET_CODE (op) == SUBREG)
4008 subreg = op;
4009 op = SUBREG_REG (op);
4010 mode = GET_MODE (op);
4013 if (CONST_POOL_OK_P (mode, op)
4014 && ((targetm.preferred_reload_class
4015 (op, (enum reg_class) goal_alt[i]) == NO_REGS)
4016 || no_input_reloads_p))
4018 rtx tem = force_const_mem (mode, op);
4020 change_p = true;
4021 if (subreg != NULL_RTX)
4022 tem = gen_rtx_SUBREG (mode, tem, SUBREG_BYTE (subreg));
4024 *curr_id->operand_loc[i] = tem;
4025 lra_update_dup (curr_id, i);
4026 process_address (i, false, &before, &after);
4028 /* If the alternative accepts constant pool refs directly
4029 there will be no reload needed at all. */
4030 if (subreg != NULL_RTX)
4031 continue;
4032 /* Skip alternatives before the one requested. */
4033 constraint = (curr_static_id->operand_alternative
4034 [goal_alt_number * n_operands + i].constraint);
4035 for (;
4036 (c = *constraint) && c != ',' && c != '#';
4037 constraint += CONSTRAINT_LEN (c, constraint))
4039 enum constraint_num cn = lookup_constraint (constraint);
4040 if ((insn_extra_memory_constraint (cn)
4041 || insn_extra_special_memory_constraint (cn))
4042 && satisfies_memory_constraint_p (tem, cn))
4043 break;
4045 if (c == '\0' || c == ',' || c == '#')
4046 continue;
4048 goal_alt_win[i] = true;
4052 n_outputs = 0;
4053 outputs[0] = -1;
4054 for (i = 0; i < n_operands; i++)
4056 int regno;
4057 bool optional_p = false;
4058 rtx old, new_reg;
4059 rtx op = *curr_id->operand_loc[i];
4061 if (goal_alt_win[i])
4063 if (goal_alt[i] == NO_REGS
4064 && REG_P (op)
4065 /* When we assign NO_REGS it means that we will not
4066 assign a hard register to the scratch pseudo by
4067 assigment pass and the scratch pseudo will be
4068 spilled. Spilled scratch pseudos are transformed
4069 back to scratches at the LRA end. */
4070 && lra_former_scratch_operand_p (curr_insn, i)
4071 && lra_former_scratch_p (REGNO (op)))
4073 int regno = REGNO (op);
4074 lra_change_class (regno, NO_REGS, " Change to", true);
4075 if (lra_get_regno_hard_regno (regno) >= 0)
4076 /* We don't have to mark all insn affected by the
4077 spilled pseudo as there is only one such insn, the
4078 current one. */
4079 reg_renumber[regno] = -1;
4080 lra_assert (bitmap_single_bit_set_p
4081 (&lra_reg_info[REGNO (op)].insn_bitmap));
4083 /* We can do an optional reload. If the pseudo got a hard
4084 reg, we might improve the code through inheritance. If
4085 it does not get a hard register we coalesce memory/memory
4086 moves later. Ignore move insns to avoid cycling. */
4087 if (! lra_simple_p
4088 && lra_undo_inheritance_iter < LRA_MAX_INHERITANCE_PASSES
4089 && goal_alt[i] != NO_REGS && REG_P (op)
4090 && (regno = REGNO (op)) >= FIRST_PSEUDO_REGISTER
4091 && regno < new_regno_start
4092 && ! lra_former_scratch_p (regno)
4093 && reg_renumber[regno] < 0
4094 /* Check that the optional reload pseudo will be able to
4095 hold given mode value. */
4096 && ! (prohibited_class_reg_set_mode_p
4097 (goal_alt[i], reg_class_contents[goal_alt[i]],
4098 PSEUDO_REGNO_MODE (regno)))
4099 && (curr_insn_set == NULL_RTX
4100 || !((REG_P (SET_SRC (curr_insn_set))
4101 || MEM_P (SET_SRC (curr_insn_set))
4102 || GET_CODE (SET_SRC (curr_insn_set)) == SUBREG)
4103 && (REG_P (SET_DEST (curr_insn_set))
4104 || MEM_P (SET_DEST (curr_insn_set))
4105 || GET_CODE (SET_DEST (curr_insn_set)) == SUBREG))))
4106 optional_p = true;
4107 else
4108 continue;
4111 /* Operands that match previous ones have already been handled. */
4112 if (goal_alt_matches[i] >= 0)
4113 continue;
4115 /* We should not have an operand with a non-offsettable address
4116 appearing where an offsettable address will do. It also may
4117 be a case when the address should be special in other words
4118 not a general one (e.g. it needs no index reg). */
4119 if (goal_alt_matched[i][0] == -1 && goal_alt_offmemok[i] && MEM_P (op))
4121 enum reg_class rclass;
4122 rtx *loc = &XEXP (op, 0);
4123 enum rtx_code code = GET_CODE (*loc);
4125 push_to_sequence (before);
4126 rclass = base_reg_class (GET_MODE (op), MEM_ADDR_SPACE (op),
4127 MEM, SCRATCH);
4128 if (GET_RTX_CLASS (code) == RTX_AUTOINC)
4129 new_reg = emit_inc (rclass, *loc, *loc,
4130 /* This value does not matter for MODIFY. */
4131 GET_MODE_SIZE (GET_MODE (op)));
4132 else if (get_reload_reg (OP_IN, Pmode, *loc, rclass, FALSE,
4133 "offsetable address", &new_reg))
4134 lra_emit_move (new_reg, *loc);
4135 before = get_insns ();
4136 end_sequence ();
4137 *loc = new_reg;
4138 lra_update_dup (curr_id, i);
4140 else if (goal_alt_matched[i][0] == -1)
4142 machine_mode mode;
4143 rtx reg, *loc;
4144 int hard_regno, byte;
4145 enum op_type type = curr_static_id->operand[i].type;
4147 loc = curr_id->operand_loc[i];
4148 mode = curr_operand_mode[i];
4149 if (GET_CODE (*loc) == SUBREG)
4151 reg = SUBREG_REG (*loc);
4152 byte = SUBREG_BYTE (*loc);
4153 if (REG_P (reg)
4154 /* Strict_low_part requires reload the register not
4155 the sub-register. */
4156 && (curr_static_id->operand[i].strict_low
4157 || (GET_MODE_SIZE (mode)
4158 <= GET_MODE_SIZE (GET_MODE (reg))
4159 && (hard_regno
4160 = get_try_hard_regno (REGNO (reg))) >= 0
4161 && (simplify_subreg_regno
4162 (hard_regno,
4163 GET_MODE (reg), byte, mode) < 0)
4164 && (goal_alt[i] == NO_REGS
4165 || (simplify_subreg_regno
4166 (ira_class_hard_regs[goal_alt[i]][0],
4167 GET_MODE (reg), byte, mode) >= 0)))))
4169 /* An OP_INOUT is required when reloading a subreg of a
4170 mode wider than a word to ensure that data beyond the
4171 word being reloaded is preserved. Also automatically
4172 ensure that strict_low_part reloads are made into
4173 OP_INOUT which should already be true from the backend
4174 constraints. */
4175 if (type == OP_OUT
4176 && (curr_static_id->operand[i].strict_low
4177 || (GET_MODE_SIZE (GET_MODE (reg)) > UNITS_PER_WORD
4178 && (GET_MODE_SIZE (mode)
4179 < GET_MODE_SIZE (GET_MODE (reg))))))
4180 type = OP_INOUT;
4181 loc = &SUBREG_REG (*loc);
4182 mode = GET_MODE (*loc);
4185 old = *loc;
4186 if (get_reload_reg (type, mode, old, goal_alt[i],
4187 loc != curr_id->operand_loc[i], "", &new_reg)
4188 && type != OP_OUT)
4190 push_to_sequence (before);
4191 lra_emit_move (new_reg, old);
4192 before = get_insns ();
4193 end_sequence ();
4195 *loc = new_reg;
4196 if (type != OP_IN
4197 && find_reg_note (curr_insn, REG_UNUSED, old) == NULL_RTX)
4199 start_sequence ();
4200 lra_emit_move (type == OP_INOUT ? copy_rtx (old) : old, new_reg);
4201 emit_insn (after);
4202 after = get_insns ();
4203 end_sequence ();
4204 *loc = new_reg;
4206 for (j = 0; j < goal_alt_dont_inherit_ops_num; j++)
4207 if (goal_alt_dont_inherit_ops[j] == i)
4209 lra_set_regno_unique_value (REGNO (new_reg));
4210 break;
4212 lra_update_dup (curr_id, i);
4214 else if (curr_static_id->operand[i].type == OP_IN
4215 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4216 == OP_OUT))
4218 /* generate reloads for input and matched outputs. */
4219 match_inputs[0] = i;
4220 match_inputs[1] = -1;
4221 match_reload (goal_alt_matched[i][0], match_inputs, outputs,
4222 goal_alt[i], &before, &after,
4223 curr_static_id->operand_alternative
4224 [goal_alt_number * n_operands + goal_alt_matched[i][0]]
4225 .earlyclobber);
4227 else if (curr_static_id->operand[i].type == OP_OUT
4228 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4229 == OP_IN))
4230 /* Generate reloads for output and matched inputs. */
4231 match_reload (i, goal_alt_matched[i], outputs, goal_alt[i], &before,
4232 &after, curr_static_id->operand_alternative
4233 [goal_alt_number * n_operands + i].earlyclobber);
4234 else if (curr_static_id->operand[i].type == OP_IN
4235 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4236 == OP_IN))
4238 /* Generate reloads for matched inputs. */
4239 match_inputs[0] = i;
4240 for (j = 0; (k = goal_alt_matched[i][j]) >= 0; j++)
4241 match_inputs[j + 1] = k;
4242 match_inputs[j + 1] = -1;
4243 match_reload (-1, match_inputs, outputs, goal_alt[i], &before,
4244 &after, false);
4246 else
4247 /* We must generate code in any case when function
4248 process_alt_operands decides that it is possible. */
4249 gcc_unreachable ();
4251 /* Memorise processed outputs so that output remaining to be processed
4252 can avoid using the same register value (see match_reload). */
4253 if (curr_static_id->operand[i].type == OP_OUT)
4255 outputs[n_outputs++] = i;
4256 outputs[n_outputs] = -1;
4259 if (optional_p)
4261 rtx reg = op;
4263 lra_assert (REG_P (reg));
4264 regno = REGNO (reg);
4265 op = *curr_id->operand_loc[i]; /* Substitution. */
4266 if (GET_CODE (op) == SUBREG)
4267 op = SUBREG_REG (op);
4268 gcc_assert (REG_P (op) && (int) REGNO (op) >= new_regno_start);
4269 bitmap_set_bit (&lra_optional_reload_pseudos, REGNO (op));
4270 lra_reg_info[REGNO (op)].restore_rtx = reg;
4271 if (lra_dump_file != NULL)
4272 fprintf (lra_dump_file,
4273 " Making reload reg %d for reg %d optional\n",
4274 REGNO (op), regno);
4277 if (before != NULL_RTX || after != NULL_RTX
4278 || max_regno_before != max_reg_num ())
4279 change_p = true;
4280 if (change_p)
4282 lra_update_operator_dups (curr_id);
4283 /* Something changes -- process the insn. */
4284 lra_update_insn_regno_info (curr_insn);
4286 lra_process_new_insns (curr_insn, before, after, "Inserting insn reload");
4287 return change_p;
4290 /* Return true if INSN satisfies all constraints. In other words, no
4291 reload insns are needed. */
4292 bool
4293 lra_constrain_insn (rtx_insn *insn)
4295 int saved_new_regno_start = new_regno_start;
4296 int saved_new_insn_uid_start = new_insn_uid_start;
4297 bool change_p;
4299 curr_insn = insn;
4300 curr_id = lra_get_insn_recog_data (curr_insn);
4301 curr_static_id = curr_id->insn_static_data;
4302 new_insn_uid_start = get_max_uid ();
4303 new_regno_start = max_reg_num ();
4304 change_p = curr_insn_transform (true);
4305 new_regno_start = saved_new_regno_start;
4306 new_insn_uid_start = saved_new_insn_uid_start;
4307 return ! change_p;
4310 /* Return true if X is in LIST. */
4311 static bool
4312 in_list_p (rtx x, rtx list)
4314 for (; list != NULL_RTX; list = XEXP (list, 1))
4315 if (XEXP (list, 0) == x)
4316 return true;
4317 return false;
4320 /* Return true if X contains an allocatable hard register (if
4321 HARD_REG_P) or a (spilled if SPILLED_P) pseudo. */
4322 static bool
4323 contains_reg_p (rtx x, bool hard_reg_p, bool spilled_p)
4325 int i, j;
4326 const char *fmt;
4327 enum rtx_code code;
4329 code = GET_CODE (x);
4330 if (REG_P (x))
4332 int regno = REGNO (x);
4333 HARD_REG_SET alloc_regs;
4335 if (hard_reg_p)
4337 if (regno >= FIRST_PSEUDO_REGISTER)
4338 regno = lra_get_regno_hard_regno (regno);
4339 if (regno < 0)
4340 return false;
4341 COMPL_HARD_REG_SET (alloc_regs, lra_no_alloc_regs);
4342 return overlaps_hard_reg_set_p (alloc_regs, GET_MODE (x), regno);
4344 else
4346 if (regno < FIRST_PSEUDO_REGISTER)
4347 return false;
4348 if (! spilled_p)
4349 return true;
4350 return lra_get_regno_hard_regno (regno) < 0;
4353 fmt = GET_RTX_FORMAT (code);
4354 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4356 if (fmt[i] == 'e')
4358 if (contains_reg_p (XEXP (x, i), hard_reg_p, spilled_p))
4359 return true;
4361 else if (fmt[i] == 'E')
4363 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4364 if (contains_reg_p (XVECEXP (x, i, j), hard_reg_p, spilled_p))
4365 return true;
4368 return false;
4371 /* Process all regs in location *LOC and change them on equivalent
4372 substitution. Return true if any change was done. */
4373 static bool
4374 loc_equivalence_change_p (rtx *loc)
4376 rtx subst, reg, x = *loc;
4377 bool result = false;
4378 enum rtx_code code = GET_CODE (x);
4379 const char *fmt;
4380 int i, j;
4382 if (code == SUBREG)
4384 reg = SUBREG_REG (x);
4385 if ((subst = get_equiv_with_elimination (reg, curr_insn)) != reg
4386 && GET_MODE (subst) == VOIDmode)
4388 /* We cannot reload debug location. Simplify subreg here
4389 while we know the inner mode. */
4390 *loc = simplify_gen_subreg (GET_MODE (x), subst,
4391 GET_MODE (reg), SUBREG_BYTE (x));
4392 return true;
4395 if (code == REG && (subst = get_equiv_with_elimination (x, curr_insn)) != x)
4397 *loc = subst;
4398 return true;
4401 /* Scan all the operand sub-expressions. */
4402 fmt = GET_RTX_FORMAT (code);
4403 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4405 if (fmt[i] == 'e')
4406 result = loc_equivalence_change_p (&XEXP (x, i)) || result;
4407 else if (fmt[i] == 'E')
4408 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4409 result
4410 = loc_equivalence_change_p (&XVECEXP (x, i, j)) || result;
4412 return result;
4415 /* Similar to loc_equivalence_change_p, but for use as
4416 simplify_replace_fn_rtx callback. DATA is insn for which the
4417 elimination is done. If it null we don't do the elimination. */
4418 static rtx
4419 loc_equivalence_callback (rtx loc, const_rtx, void *data)
4421 if (!REG_P (loc))
4422 return NULL_RTX;
4424 rtx subst = (data == NULL
4425 ? get_equiv (loc) : get_equiv_with_elimination (loc, (rtx_insn *) data));
4426 if (subst != loc)
4427 return subst;
4429 return NULL_RTX;
4432 /* Maximum number of generated reload insns per an insn. It is for
4433 preventing this pass cycling in a bug case. */
4434 #define MAX_RELOAD_INSNS_NUMBER LRA_MAX_INSN_RELOADS
4436 /* The current iteration number of this LRA pass. */
4437 int lra_constraint_iter;
4439 /* True if we substituted equiv which needs checking register
4440 allocation correctness because the equivalent value contains
4441 allocatable hard registers or when we restore multi-register
4442 pseudo. */
4443 bool lra_risky_transformations_p;
4445 /* Return true if REGNO is referenced in more than one block. */
4446 static bool
4447 multi_block_pseudo_p (int regno)
4449 basic_block bb = NULL;
4450 unsigned int uid;
4451 bitmap_iterator bi;
4453 if (regno < FIRST_PSEUDO_REGISTER)
4454 return false;
4456 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi)
4457 if (bb == NULL)
4458 bb = BLOCK_FOR_INSN (lra_insn_recog_data[uid]->insn);
4459 else if (BLOCK_FOR_INSN (lra_insn_recog_data[uid]->insn) != bb)
4460 return true;
4461 return false;
4464 /* Return true if LIST contains a deleted insn. */
4465 static bool
4466 contains_deleted_insn_p (rtx_insn_list *list)
4468 for (; list != NULL_RTX; list = list->next ())
4469 if (NOTE_P (list->insn ())
4470 && NOTE_KIND (list->insn ()) == NOTE_INSN_DELETED)
4471 return true;
4472 return false;
4475 /* Return true if X contains a pseudo dying in INSN. */
4476 static bool
4477 dead_pseudo_p (rtx x, rtx_insn *insn)
4479 int i, j;
4480 const char *fmt;
4481 enum rtx_code code;
4483 if (REG_P (x))
4484 return (insn != NULL_RTX
4485 && find_regno_note (insn, REG_DEAD, REGNO (x)) != NULL_RTX);
4486 code = GET_CODE (x);
4487 fmt = GET_RTX_FORMAT (code);
4488 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4490 if (fmt[i] == 'e')
4492 if (dead_pseudo_p (XEXP (x, i), insn))
4493 return true;
4495 else if (fmt[i] == 'E')
4497 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4498 if (dead_pseudo_p (XVECEXP (x, i, j), insn))
4499 return true;
4502 return false;
4505 /* Return true if INSN contains a dying pseudo in INSN right hand
4506 side. */
4507 static bool
4508 insn_rhs_dead_pseudo_p (rtx_insn *insn)
4510 rtx set = single_set (insn);
4512 gcc_assert (set != NULL);
4513 return dead_pseudo_p (SET_SRC (set), insn);
4516 /* Return true if any init insn of REGNO contains a dying pseudo in
4517 insn right hand side. */
4518 static bool
4519 init_insn_rhs_dead_pseudo_p (int regno)
4521 rtx_insn_list *insns = ira_reg_equiv[regno].init_insns;
4523 if (insns == NULL)
4524 return false;
4525 for (; insns != NULL_RTX; insns = insns->next ())
4526 if (insn_rhs_dead_pseudo_p (insns->insn ()))
4527 return true;
4528 return false;
4531 /* Return TRUE if REGNO has a reverse equivalence. The equivalence is
4532 reverse only if we have one init insn with given REGNO as a
4533 source. */
4534 static bool
4535 reverse_equiv_p (int regno)
4537 rtx_insn_list *insns = ira_reg_equiv[regno].init_insns;
4538 rtx set;
4540 if (insns == NULL)
4541 return false;
4542 if (! INSN_P (insns->insn ())
4543 || insns->next () != NULL)
4544 return false;
4545 if ((set = single_set (insns->insn ())) == NULL_RTX)
4546 return false;
4547 return REG_P (SET_SRC (set)) && (int) REGNO (SET_SRC (set)) == regno;
4550 /* Return TRUE if REGNO was reloaded in an equivalence init insn. We
4551 call this function only for non-reverse equivalence. */
4552 static bool
4553 contains_reloaded_insn_p (int regno)
4555 rtx set;
4556 rtx_insn_list *list = ira_reg_equiv[regno].init_insns;
4558 for (; list != NULL; list = list->next ())
4559 if ((set = single_set (list->insn ())) == NULL_RTX
4560 || ! REG_P (SET_DEST (set))
4561 || (int) REGNO (SET_DEST (set)) != regno)
4562 return true;
4563 return false;
4566 /* Entry function of LRA constraint pass. Return true if the
4567 constraint pass did change the code. */
4568 bool
4569 lra_constraints (bool first_p)
4571 bool changed_p;
4572 int i, hard_regno, new_insns_num;
4573 unsigned int min_len, new_min_len, uid;
4574 rtx set, x, reg, dest_reg;
4575 basic_block last_bb;
4576 bitmap_head equiv_insn_bitmap;
4577 bitmap_iterator bi;
4579 lra_constraint_iter++;
4580 if (lra_dump_file != NULL)
4581 fprintf (lra_dump_file, "\n********** Local #%d: **********\n\n",
4582 lra_constraint_iter);
4583 changed_p = false;
4584 if (pic_offset_table_rtx
4585 && REGNO (pic_offset_table_rtx) >= FIRST_PSEUDO_REGISTER)
4586 lra_risky_transformations_p = true;
4587 else
4588 /* On the first iteration we should check IRA assignment
4589 correctness. In rare cases, the assignments can be wrong as
4590 early clobbers operands are ignored in IRA. */
4591 lra_risky_transformations_p = first_p;
4592 new_insn_uid_start = get_max_uid ();
4593 new_regno_start = first_p ? lra_constraint_new_regno_start : max_reg_num ();
4594 /* Mark used hard regs for target stack size calulations. */
4595 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4596 if (lra_reg_info[i].nrefs != 0
4597 && (hard_regno = lra_get_regno_hard_regno (i)) >= 0)
4599 int j, nregs;
4601 nregs = hard_regno_nregs[hard_regno][lra_reg_info[i].biggest_mode];
4602 for (j = 0; j < nregs; j++)
4603 df_set_regs_ever_live (hard_regno + j, true);
4605 /* Do elimination before the equivalence processing as we can spill
4606 some pseudos during elimination. */
4607 lra_eliminate (false, first_p);
4608 bitmap_initialize (&equiv_insn_bitmap, &reg_obstack);
4609 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4610 if (lra_reg_info[i].nrefs != 0)
4612 ira_reg_equiv[i].profitable_p = true;
4613 reg = regno_reg_rtx[i];
4614 if (lra_get_regno_hard_regno (i) < 0 && (x = get_equiv (reg)) != reg)
4616 bool pseudo_p = contains_reg_p (x, false, false);
4618 /* After RTL transformation, we can not guarantee that
4619 pseudo in the substitution was not reloaded which might
4620 make equivalence invalid. For example, in reverse
4621 equiv of p0
4623 p0 <- ...
4625 equiv_mem <- p0
4627 the memory address register was reloaded before the 2nd
4628 insn. */
4629 if ((! first_p && pseudo_p)
4630 /* We don't use DF for compilation speed sake. So it
4631 is problematic to update live info when we use an
4632 equivalence containing pseudos in more than one
4633 BB. */
4634 || (pseudo_p && multi_block_pseudo_p (i))
4635 /* If an init insn was deleted for some reason, cancel
4636 the equiv. We could update the equiv insns after
4637 transformations including an equiv insn deletion
4638 but it is not worthy as such cases are extremely
4639 rare. */
4640 || contains_deleted_insn_p (ira_reg_equiv[i].init_insns)
4641 /* If it is not a reverse equivalence, we check that a
4642 pseudo in rhs of the init insn is not dying in the
4643 insn. Otherwise, the live info at the beginning of
4644 the corresponding BB might be wrong after we
4645 removed the insn. When the equiv can be a
4646 constant, the right hand side of the init insn can
4647 be a pseudo. */
4648 || (! reverse_equiv_p (i)
4649 && (init_insn_rhs_dead_pseudo_p (i)
4650 /* If we reloaded the pseudo in an equivalence
4651 init insn, we can not remove the equiv init
4652 insns and the init insns might write into
4653 const memory in this case. */
4654 || contains_reloaded_insn_p (i)))
4655 /* Prevent access beyond equivalent memory for
4656 paradoxical subregs. */
4657 || (MEM_P (x)
4658 && (GET_MODE_SIZE (lra_reg_info[i].biggest_mode)
4659 > GET_MODE_SIZE (GET_MODE (x))))
4660 || (pic_offset_table_rtx
4661 && ((CONST_POOL_OK_P (PSEUDO_REGNO_MODE (i), x)
4662 && (targetm.preferred_reload_class
4663 (x, lra_get_allocno_class (i)) == NO_REGS))
4664 || contains_symbol_ref_p (x))))
4665 ira_reg_equiv[i].defined_p = false;
4666 if (contains_reg_p (x, false, true))
4667 ira_reg_equiv[i].profitable_p = false;
4668 if (get_equiv (reg) != reg)
4669 bitmap_ior_into (&equiv_insn_bitmap, &lra_reg_info[i].insn_bitmap);
4672 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4673 update_equiv (i);
4674 /* We should add all insns containing pseudos which should be
4675 substituted by their equivalences. */
4676 EXECUTE_IF_SET_IN_BITMAP (&equiv_insn_bitmap, 0, uid, bi)
4677 lra_push_insn_by_uid (uid);
4678 min_len = lra_insn_stack_length ();
4679 new_insns_num = 0;
4680 last_bb = NULL;
4681 changed_p = false;
4682 while ((new_min_len = lra_insn_stack_length ()) != 0)
4684 curr_insn = lra_pop_insn ();
4685 --new_min_len;
4686 curr_bb = BLOCK_FOR_INSN (curr_insn);
4687 if (curr_bb != last_bb)
4689 last_bb = curr_bb;
4690 bb_reload_num = lra_curr_reload_num;
4692 if (min_len > new_min_len)
4694 min_len = new_min_len;
4695 new_insns_num = 0;
4697 if (new_insns_num > MAX_RELOAD_INSNS_NUMBER)
4698 internal_error
4699 ("Max. number of generated reload insns per insn is achieved (%d)\n",
4700 MAX_RELOAD_INSNS_NUMBER);
4701 new_insns_num++;
4702 if (DEBUG_INSN_P (curr_insn))
4704 /* We need to check equivalence in debug insn and change
4705 pseudo to the equivalent value if necessary. */
4706 curr_id = lra_get_insn_recog_data (curr_insn);
4707 if (bitmap_bit_p (&equiv_insn_bitmap, INSN_UID (curr_insn)))
4709 rtx old = *curr_id->operand_loc[0];
4710 *curr_id->operand_loc[0]
4711 = simplify_replace_fn_rtx (old, NULL_RTX,
4712 loc_equivalence_callback, curr_insn);
4713 if (old != *curr_id->operand_loc[0])
4715 lra_update_insn_regno_info (curr_insn);
4716 changed_p = true;
4720 else if (INSN_P (curr_insn))
4722 if ((set = single_set (curr_insn)) != NULL_RTX)
4724 dest_reg = SET_DEST (set);
4725 /* The equivalence pseudo could be set up as SUBREG in a
4726 case when it is a call restore insn in a mode
4727 different from the pseudo mode. */
4728 if (GET_CODE (dest_reg) == SUBREG)
4729 dest_reg = SUBREG_REG (dest_reg);
4730 if ((REG_P (dest_reg)
4731 && (x = get_equiv (dest_reg)) != dest_reg
4732 /* Remove insns which set up a pseudo whose value
4733 can not be changed. Such insns might be not in
4734 init_insns because we don't update equiv data
4735 during insn transformations.
4737 As an example, let suppose that a pseudo got
4738 hard register and on the 1st pass was not
4739 changed to equivalent constant. We generate an
4740 additional insn setting up the pseudo because of
4741 secondary memory movement. Then the pseudo is
4742 spilled and we use the equiv constant. In this
4743 case we should remove the additional insn and
4744 this insn is not init_insns list. */
4745 && (! MEM_P (x) || MEM_READONLY_P (x)
4746 /* Check that this is actually an insn setting
4747 up the equivalence. */
4748 || in_list_p (curr_insn,
4749 ira_reg_equiv
4750 [REGNO (dest_reg)].init_insns)))
4751 || (((x = get_equiv (SET_SRC (set))) != SET_SRC (set))
4752 && in_list_p (curr_insn,
4753 ira_reg_equiv
4754 [REGNO (SET_SRC (set))].init_insns)))
4756 /* This is equiv init insn of pseudo which did not get a
4757 hard register -- remove the insn. */
4758 if (lra_dump_file != NULL)
4760 fprintf (lra_dump_file,
4761 " Removing equiv init insn %i (freq=%d)\n",
4762 INSN_UID (curr_insn),
4763 REG_FREQ_FROM_BB (BLOCK_FOR_INSN (curr_insn)));
4764 dump_insn_slim (lra_dump_file, curr_insn);
4766 if (contains_reg_p (x, true, false))
4767 lra_risky_transformations_p = true;
4768 lra_set_insn_deleted (curr_insn);
4769 continue;
4772 curr_id = lra_get_insn_recog_data (curr_insn);
4773 curr_static_id = curr_id->insn_static_data;
4774 init_curr_insn_input_reloads ();
4775 init_curr_operand_mode ();
4776 if (curr_insn_transform (false))
4777 changed_p = true;
4778 /* Check non-transformed insns too for equiv change as USE
4779 or CLOBBER don't need reloads but can contain pseudos
4780 being changed on their equivalences. */
4781 else if (bitmap_bit_p (&equiv_insn_bitmap, INSN_UID (curr_insn))
4782 && loc_equivalence_change_p (&PATTERN (curr_insn)))
4784 lra_update_insn_regno_info (curr_insn);
4785 changed_p = true;
4789 bitmap_clear (&equiv_insn_bitmap);
4790 /* If we used a new hard regno, changed_p should be true because the
4791 hard reg is assigned to a new pseudo. */
4792 if (flag_checking && !changed_p)
4794 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4795 if (lra_reg_info[i].nrefs != 0
4796 && (hard_regno = lra_get_regno_hard_regno (i)) >= 0)
4798 int j, nregs = hard_regno_nregs[hard_regno][PSEUDO_REGNO_MODE (i)];
4800 for (j = 0; j < nregs; j++)
4801 lra_assert (df_regs_ever_live_p (hard_regno + j));
4804 return changed_p;
4807 static void initiate_invariants (void);
4808 static void finish_invariants (void);
4810 /* Initiate the LRA constraint pass. It is done once per
4811 function. */
4812 void
4813 lra_constraints_init (void)
4815 initiate_invariants ();
4818 /* Finalize the LRA constraint pass. It is done once per
4819 function. */
4820 void
4821 lra_constraints_finish (void)
4823 finish_invariants ();
4828 /* Structure describes invariants for ineheritance. */
4829 struct lra_invariant
4831 /* The order number of the invariant. */
4832 int num;
4833 /* The invariant RTX. */
4834 rtx invariant_rtx;
4835 /* The origin insn of the invariant. */
4836 rtx_insn *insn;
4839 typedef lra_invariant invariant_t;
4840 typedef invariant_t *invariant_ptr_t;
4841 typedef const invariant_t *const_invariant_ptr_t;
4843 /* Pointer to the inheritance invariants. */
4844 static vec<invariant_ptr_t> invariants;
4846 /* Allocation pool for the invariants. */
4847 static object_allocator<lra_invariant> *invariants_pool;
4849 /* Hash table for the invariants. */
4850 static htab_t invariant_table;
4852 /* Hash function for INVARIANT. */
4853 static hashval_t
4854 invariant_hash (const void *invariant)
4856 rtx inv = ((const_invariant_ptr_t) invariant)->invariant_rtx;
4857 return lra_rtx_hash (inv);
4860 /* Equal function for invariants INVARIANT1 and INVARIANT2. */
4861 static int
4862 invariant_eq_p (const void *invariant1, const void *invariant2)
4864 rtx inv1 = ((const_invariant_ptr_t) invariant1)->invariant_rtx;
4865 rtx inv2 = ((const_invariant_ptr_t) invariant2)->invariant_rtx;
4867 return rtx_equal_p (inv1, inv2);
4870 /* Insert INVARIANT_RTX into the table if it is not there yet. Return
4871 invariant which is in the table. */
4872 static invariant_ptr_t
4873 insert_invariant (rtx invariant_rtx)
4875 void **entry_ptr;
4876 invariant_t invariant;
4877 invariant_ptr_t invariant_ptr;
4879 invariant.invariant_rtx = invariant_rtx;
4880 entry_ptr = htab_find_slot (invariant_table, &invariant, INSERT);
4881 if (*entry_ptr == NULL)
4883 invariant_ptr = invariants_pool->allocate ();
4884 invariant_ptr->invariant_rtx = invariant_rtx;
4885 invariant_ptr->insn = NULL;
4886 invariants.safe_push (invariant_ptr);
4887 *entry_ptr = (void *) invariant_ptr;
4889 return (invariant_ptr_t) *entry_ptr;
4892 /* Initiate the invariant table. */
4893 static void
4894 initiate_invariants (void)
4896 invariants.create (100);
4897 invariants_pool
4898 = new object_allocator<lra_invariant> ("Inheritance invariants");
4899 invariant_table = htab_create (100, invariant_hash, invariant_eq_p, NULL);
4902 /* Finish the invariant table. */
4903 static void
4904 finish_invariants (void)
4906 htab_delete (invariant_table);
4907 delete invariants_pool;
4908 invariants.release ();
4911 /* Make the invariant table empty. */
4912 static void
4913 clear_invariants (void)
4915 htab_empty (invariant_table);
4916 invariants_pool->release ();
4917 invariants.truncate (0);
4922 /* This page contains code to do inheritance/split
4923 transformations. */
4925 /* Number of reloads passed so far in current EBB. */
4926 static int reloads_num;
4928 /* Number of calls passed so far in current EBB. */
4929 static int calls_num;
4931 /* Current reload pseudo check for validity of elements in
4932 USAGE_INSNS. */
4933 static int curr_usage_insns_check;
4935 /* Info about last usage of registers in EBB to do inheritance/split
4936 transformation. Inheritance transformation is done from a spilled
4937 pseudo and split transformations from a hard register or a pseudo
4938 assigned to a hard register. */
4939 struct usage_insns
4941 /* If the value is equal to CURR_USAGE_INSNS_CHECK, then the member
4942 value INSNS is valid. The insns is chain of optional debug insns
4943 and a finishing non-debug insn using the corresponding reg. The
4944 value is also used to mark the registers which are set up in the
4945 current insn. The negated insn uid is used for this. */
4946 int check;
4947 /* Value of global reloads_num at the last insn in INSNS. */
4948 int reloads_num;
4949 /* Value of global reloads_nums at the last insn in INSNS. */
4950 int calls_num;
4951 /* It can be true only for splitting. And it means that the restore
4952 insn should be put after insn given by the following member. */
4953 bool after_p;
4954 /* Next insns in the current EBB which use the original reg and the
4955 original reg value is not changed between the current insn and
4956 the next insns. In order words, e.g. for inheritance, if we need
4957 to use the original reg value again in the next insns we can try
4958 to use the value in a hard register from a reload insn of the
4959 current insn. */
4960 rtx insns;
4963 /* Map: regno -> corresponding pseudo usage insns. */
4964 static struct usage_insns *usage_insns;
4966 static void
4967 setup_next_usage_insn (int regno, rtx insn, int reloads_num, bool after_p)
4969 usage_insns[regno].check = curr_usage_insns_check;
4970 usage_insns[regno].insns = insn;
4971 usage_insns[regno].reloads_num = reloads_num;
4972 usage_insns[regno].calls_num = calls_num;
4973 usage_insns[regno].after_p = after_p;
4976 /* The function is used to form list REGNO usages which consists of
4977 optional debug insns finished by a non-debug insn using REGNO.
4978 RELOADS_NUM is current number of reload insns processed so far. */
4979 static void
4980 add_next_usage_insn (int regno, rtx_insn *insn, int reloads_num)
4982 rtx next_usage_insns;
4984 if (usage_insns[regno].check == curr_usage_insns_check
4985 && (next_usage_insns = usage_insns[regno].insns) != NULL_RTX
4986 && DEBUG_INSN_P (insn))
4988 /* Check that we did not add the debug insn yet. */
4989 if (next_usage_insns != insn
4990 && (GET_CODE (next_usage_insns) != INSN_LIST
4991 || XEXP (next_usage_insns, 0) != insn))
4992 usage_insns[regno].insns = gen_rtx_INSN_LIST (VOIDmode, insn,
4993 next_usage_insns);
4995 else if (NONDEBUG_INSN_P (insn))
4996 setup_next_usage_insn (regno, insn, reloads_num, false);
4997 else
4998 usage_insns[regno].check = 0;
5001 /* Return first non-debug insn in list USAGE_INSNS. */
5002 static rtx_insn *
5003 skip_usage_debug_insns (rtx usage_insns)
5005 rtx insn;
5007 /* Skip debug insns. */
5008 for (insn = usage_insns;
5009 insn != NULL_RTX && GET_CODE (insn) == INSN_LIST;
5010 insn = XEXP (insn, 1))
5012 return safe_as_a <rtx_insn *> (insn);
5015 /* Return true if we need secondary memory moves for insn in
5016 USAGE_INSNS after inserting inherited pseudo of class INHER_CL
5017 into the insn. */
5018 static bool
5019 check_secondary_memory_needed_p (enum reg_class inher_cl ATTRIBUTE_UNUSED,
5020 rtx usage_insns ATTRIBUTE_UNUSED)
5022 #ifndef SECONDARY_MEMORY_NEEDED
5023 return false;
5024 #else
5025 rtx_insn *insn;
5026 rtx set, dest;
5027 enum reg_class cl;
5029 if (inher_cl == ALL_REGS
5030 || (insn = skip_usage_debug_insns (usage_insns)) == NULL_RTX)
5031 return false;
5032 lra_assert (INSN_P (insn));
5033 if ((set = single_set (insn)) == NULL_RTX || ! REG_P (SET_DEST (set)))
5034 return false;
5035 dest = SET_DEST (set);
5036 if (! REG_P (dest))
5037 return false;
5038 lra_assert (inher_cl != NO_REGS);
5039 cl = get_reg_class (REGNO (dest));
5040 return (cl != NO_REGS && cl != ALL_REGS
5041 && SECONDARY_MEMORY_NEEDED (inher_cl, cl, GET_MODE (dest)));
5042 #endif
5045 /* Registers involved in inheritance/split in the current EBB
5046 (inheritance/split pseudos and original registers). */
5047 static bitmap_head check_only_regs;
5049 /* Reload pseudos can not be involded in invariant inheritance in the
5050 current EBB. */
5051 static bitmap_head invalid_invariant_regs;
5053 /* Do inheritance transformations for insn INSN, which defines (if
5054 DEF_P) or uses ORIGINAL_REGNO. NEXT_USAGE_INSNS specifies which
5055 instruction in the EBB next uses ORIGINAL_REGNO; it has the same
5056 form as the "insns" field of usage_insns. Return true if we
5057 succeed in such transformation.
5059 The transformations look like:
5061 p <- ... i <- ...
5062 ... p <- i (new insn)
5063 ... =>
5064 <- ... p ... <- ... i ...
5066 ... i <- p (new insn)
5067 <- ... p ... <- ... i ...
5068 ... =>
5069 <- ... p ... <- ... i ...
5070 where p is a spilled original pseudo and i is a new inheritance pseudo.
5073 The inheritance pseudo has the smallest class of two classes CL and
5074 class of ORIGINAL REGNO. */
5075 static bool
5076 inherit_reload_reg (bool def_p, int original_regno,
5077 enum reg_class cl, rtx_insn *insn, rtx next_usage_insns)
5079 if (optimize_function_for_size_p (cfun))
5080 return false;
5082 enum reg_class rclass = lra_get_allocno_class (original_regno);
5083 rtx original_reg = regno_reg_rtx[original_regno];
5084 rtx new_reg, usage_insn;
5085 rtx_insn *new_insns;
5087 lra_assert (! usage_insns[original_regno].after_p);
5088 if (lra_dump_file != NULL)
5089 fprintf (lra_dump_file,
5090 " <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<\n");
5091 if (! ira_reg_classes_intersect_p[cl][rclass])
5093 if (lra_dump_file != NULL)
5095 fprintf (lra_dump_file,
5096 " Rejecting inheritance for %d "
5097 "because of disjoint classes %s and %s\n",
5098 original_regno, reg_class_names[cl],
5099 reg_class_names[rclass]);
5100 fprintf (lra_dump_file,
5101 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5103 return false;
5105 if ((ira_class_subset_p[cl][rclass] && cl != rclass)
5106 /* We don't use a subset of two classes because it can be
5107 NO_REGS. This transformation is still profitable in most
5108 cases even if the classes are not intersected as register
5109 move is probably cheaper than a memory load. */
5110 || ira_class_hard_regs_num[cl] < ira_class_hard_regs_num[rclass])
5112 if (lra_dump_file != NULL)
5113 fprintf (lra_dump_file, " Use smallest class of %s and %s\n",
5114 reg_class_names[cl], reg_class_names[rclass]);
5116 rclass = cl;
5118 if (check_secondary_memory_needed_p (rclass, next_usage_insns))
5120 /* Reject inheritance resulting in secondary memory moves.
5121 Otherwise, there is a danger in LRA cycling. Also such
5122 transformation will be unprofitable. */
5123 if (lra_dump_file != NULL)
5125 rtx_insn *insn = skip_usage_debug_insns (next_usage_insns);
5126 rtx set = single_set (insn);
5128 lra_assert (set != NULL_RTX);
5130 rtx dest = SET_DEST (set);
5132 lra_assert (REG_P (dest));
5133 fprintf (lra_dump_file,
5134 " Rejecting inheritance for insn %d(%s)<-%d(%s) "
5135 "as secondary mem is needed\n",
5136 REGNO (dest), reg_class_names[get_reg_class (REGNO (dest))],
5137 original_regno, reg_class_names[rclass]);
5138 fprintf (lra_dump_file,
5139 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5141 return false;
5143 new_reg = lra_create_new_reg (GET_MODE (original_reg), original_reg,
5144 rclass, "inheritance");
5145 start_sequence ();
5146 if (def_p)
5147 lra_emit_move (original_reg, new_reg);
5148 else
5149 lra_emit_move (new_reg, original_reg);
5150 new_insns = get_insns ();
5151 end_sequence ();
5152 if (NEXT_INSN (new_insns) != NULL_RTX)
5154 if (lra_dump_file != NULL)
5156 fprintf (lra_dump_file,
5157 " Rejecting inheritance %d->%d "
5158 "as it results in 2 or more insns:\n",
5159 original_regno, REGNO (new_reg));
5160 dump_rtl_slim (lra_dump_file, new_insns, NULL, -1, 0);
5161 fprintf (lra_dump_file,
5162 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5164 return false;
5166 lra_substitute_pseudo_within_insn (insn, original_regno, new_reg, false);
5167 lra_update_insn_regno_info (insn);
5168 if (! def_p)
5169 /* We now have a new usage insn for original regno. */
5170 setup_next_usage_insn (original_regno, new_insns, reloads_num, false);
5171 if (lra_dump_file != NULL)
5172 fprintf (lra_dump_file, " Original reg change %d->%d (bb%d):\n",
5173 original_regno, REGNO (new_reg), BLOCK_FOR_INSN (insn)->index);
5174 lra_reg_info[REGNO (new_reg)].restore_rtx = regno_reg_rtx[original_regno];
5175 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
5176 bitmap_set_bit (&check_only_regs, original_regno);
5177 bitmap_set_bit (&lra_inheritance_pseudos, REGNO (new_reg));
5178 if (def_p)
5179 lra_process_new_insns (insn, NULL, new_insns,
5180 "Add original<-inheritance");
5181 else
5182 lra_process_new_insns (insn, new_insns, NULL,
5183 "Add inheritance<-original");
5184 while (next_usage_insns != NULL_RTX)
5186 if (GET_CODE (next_usage_insns) != INSN_LIST)
5188 usage_insn = next_usage_insns;
5189 lra_assert (NONDEBUG_INSN_P (usage_insn));
5190 next_usage_insns = NULL;
5192 else
5194 usage_insn = XEXP (next_usage_insns, 0);
5195 lra_assert (DEBUG_INSN_P (usage_insn));
5196 next_usage_insns = XEXP (next_usage_insns, 1);
5198 lra_substitute_pseudo (&usage_insn, original_regno, new_reg, false);
5199 lra_update_insn_regno_info (as_a <rtx_insn *> (usage_insn));
5200 if (lra_dump_file != NULL)
5202 fprintf (lra_dump_file,
5203 " Inheritance reuse change %d->%d (bb%d):\n",
5204 original_regno, REGNO (new_reg),
5205 BLOCK_FOR_INSN (usage_insn)->index);
5206 dump_insn_slim (lra_dump_file, as_a <rtx_insn *> (usage_insn));
5209 if (lra_dump_file != NULL)
5210 fprintf (lra_dump_file,
5211 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5212 return true;
5215 /* Return true if we need a caller save/restore for pseudo REGNO which
5216 was assigned to a hard register. */
5217 static inline bool
5218 need_for_call_save_p (int regno)
5220 lra_assert (regno >= FIRST_PSEUDO_REGISTER && reg_renumber[regno] >= 0);
5221 return (usage_insns[regno].calls_num < calls_num
5222 && (overlaps_hard_reg_set_p
5223 ((flag_ipa_ra &&
5224 ! hard_reg_set_empty_p (lra_reg_info[regno].actual_call_used_reg_set))
5225 ? lra_reg_info[regno].actual_call_used_reg_set
5226 : call_used_reg_set,
5227 PSEUDO_REGNO_MODE (regno), reg_renumber[regno])
5228 || HARD_REGNO_CALL_PART_CLOBBERED (reg_renumber[regno],
5229 PSEUDO_REGNO_MODE (regno))));
5232 /* Global registers occurring in the current EBB. */
5233 static bitmap_head ebb_global_regs;
5235 /* Return true if we need a split for hard register REGNO or pseudo
5236 REGNO which was assigned to a hard register.
5237 POTENTIAL_RELOAD_HARD_REGS contains hard registers which might be
5238 used for reloads since the EBB end. It is an approximation of the
5239 used hard registers in the split range. The exact value would
5240 require expensive calculations. If we were aggressive with
5241 splitting because of the approximation, the split pseudo will save
5242 the same hard register assignment and will be removed in the undo
5243 pass. We still need the approximation because too aggressive
5244 splitting would result in too inaccurate cost calculation in the
5245 assignment pass because of too many generated moves which will be
5246 probably removed in the undo pass. */
5247 static inline bool
5248 need_for_split_p (HARD_REG_SET potential_reload_hard_regs, int regno)
5250 int hard_regno = regno < FIRST_PSEUDO_REGISTER ? regno : reg_renumber[regno];
5252 lra_assert (hard_regno >= 0);
5253 return ((TEST_HARD_REG_BIT (potential_reload_hard_regs, hard_regno)
5254 /* Don't split eliminable hard registers, otherwise we can
5255 split hard registers like hard frame pointer, which
5256 lives on BB start/end according to DF-infrastructure,
5257 when there is a pseudo assigned to the register and
5258 living in the same BB. */
5259 && (regno >= FIRST_PSEUDO_REGISTER
5260 || ! TEST_HARD_REG_BIT (eliminable_regset, hard_regno))
5261 && ! TEST_HARD_REG_BIT (lra_no_alloc_regs, hard_regno)
5262 /* Don't split call clobbered hard regs living through
5263 calls, otherwise we might have a check problem in the
5264 assign sub-pass as in the most cases (exception is a
5265 situation when lra_risky_transformations_p value is
5266 true) the assign pass assumes that all pseudos living
5267 through calls are assigned to call saved hard regs. */
5268 && (regno >= FIRST_PSEUDO_REGISTER
5269 || ! TEST_HARD_REG_BIT (call_used_reg_set, regno)
5270 || usage_insns[regno].calls_num == calls_num)
5271 /* We need at least 2 reloads to make pseudo splitting
5272 profitable. We should provide hard regno splitting in
5273 any case to solve 1st insn scheduling problem when
5274 moving hard register definition up might result in
5275 impossibility to find hard register for reload pseudo of
5276 small register class. */
5277 && (usage_insns[regno].reloads_num
5278 + (regno < FIRST_PSEUDO_REGISTER ? 0 : 3) < reloads_num)
5279 && (regno < FIRST_PSEUDO_REGISTER
5280 /* For short living pseudos, spilling + inheritance can
5281 be considered a substitution for splitting.
5282 Therefore we do not splitting for local pseudos. It
5283 decreases also aggressiveness of splitting. The
5284 minimal number of references is chosen taking into
5285 account that for 2 references splitting has no sense
5286 as we can just spill the pseudo. */
5287 || (regno >= FIRST_PSEUDO_REGISTER
5288 && lra_reg_info[regno].nrefs > 3
5289 && bitmap_bit_p (&ebb_global_regs, regno))))
5290 || (regno >= FIRST_PSEUDO_REGISTER && need_for_call_save_p (regno)));
5293 /* Return class for the split pseudo created from original pseudo with
5294 ALLOCNO_CLASS and MODE which got a hard register HARD_REGNO. We
5295 choose subclass of ALLOCNO_CLASS which contains HARD_REGNO and
5296 results in no secondary memory movements. */
5297 static enum reg_class
5298 choose_split_class (enum reg_class allocno_class,
5299 int hard_regno ATTRIBUTE_UNUSED,
5300 machine_mode mode ATTRIBUTE_UNUSED)
5302 #ifndef SECONDARY_MEMORY_NEEDED
5303 return allocno_class;
5304 #else
5305 int i;
5306 enum reg_class cl, best_cl = NO_REGS;
5307 enum reg_class hard_reg_class ATTRIBUTE_UNUSED
5308 = REGNO_REG_CLASS (hard_regno);
5310 if (! SECONDARY_MEMORY_NEEDED (allocno_class, allocno_class, mode)
5311 && TEST_HARD_REG_BIT (reg_class_contents[allocno_class], hard_regno))
5312 return allocno_class;
5313 for (i = 0;
5314 (cl = reg_class_subclasses[allocno_class][i]) != LIM_REG_CLASSES;
5315 i++)
5316 if (! SECONDARY_MEMORY_NEEDED (cl, hard_reg_class, mode)
5317 && ! SECONDARY_MEMORY_NEEDED (hard_reg_class, cl, mode)
5318 && TEST_HARD_REG_BIT (reg_class_contents[cl], hard_regno)
5319 && (best_cl == NO_REGS
5320 || ira_class_hard_regs_num[best_cl] < ira_class_hard_regs_num[cl]))
5321 best_cl = cl;
5322 return best_cl;
5323 #endif
5326 /* Do split transformations for insn INSN, which defines or uses
5327 ORIGINAL_REGNO. NEXT_USAGE_INSNS specifies which instruction in
5328 the EBB next uses ORIGINAL_REGNO; it has the same form as the
5329 "insns" field of usage_insns.
5331 The transformations look like:
5333 p <- ... p <- ...
5334 ... s <- p (new insn -- save)
5335 ... =>
5336 ... p <- s (new insn -- restore)
5337 <- ... p ... <- ... p ...
5339 <- ... p ... <- ... p ...
5340 ... s <- p (new insn -- save)
5341 ... =>
5342 ... p <- s (new insn -- restore)
5343 <- ... p ... <- ... p ...
5345 where p is an original pseudo got a hard register or a hard
5346 register and s is a new split pseudo. The save is put before INSN
5347 if BEFORE_P is true. Return true if we succeed in such
5348 transformation. */
5349 static bool
5350 split_reg (bool before_p, int original_regno, rtx_insn *insn,
5351 rtx next_usage_insns)
5353 enum reg_class rclass;
5354 rtx original_reg;
5355 int hard_regno, nregs;
5356 rtx new_reg, usage_insn;
5357 rtx_insn *restore, *save;
5358 bool after_p;
5359 bool call_save_p;
5360 machine_mode mode;
5362 if (original_regno < FIRST_PSEUDO_REGISTER)
5364 rclass = ira_allocno_class_translate[REGNO_REG_CLASS (original_regno)];
5365 hard_regno = original_regno;
5366 call_save_p = false;
5367 nregs = 1;
5368 mode = lra_reg_info[hard_regno].biggest_mode;
5369 machine_mode reg_rtx_mode = GET_MODE (regno_reg_rtx[hard_regno]);
5370 /* A reg can have a biggest_mode of VOIDmode if it was only ever seen
5371 as part of a multi-word register. In that case, or if the biggest
5372 mode was larger than a register, just use the reg_rtx. Otherwise,
5373 limit the size to that of the biggest access in the function. */
5374 if (mode == VOIDmode
5375 || GET_MODE_SIZE (mode) > GET_MODE_SIZE (reg_rtx_mode))
5377 original_reg = regno_reg_rtx[hard_regno];
5378 mode = reg_rtx_mode;
5380 else
5381 original_reg = gen_rtx_REG (mode, hard_regno);
5383 else
5385 mode = PSEUDO_REGNO_MODE (original_regno);
5386 hard_regno = reg_renumber[original_regno];
5387 nregs = hard_regno_nregs[hard_regno][mode];
5388 rclass = lra_get_allocno_class (original_regno);
5389 original_reg = regno_reg_rtx[original_regno];
5390 call_save_p = need_for_call_save_p (original_regno);
5392 lra_assert (hard_regno >= 0);
5393 if (lra_dump_file != NULL)
5394 fprintf (lra_dump_file,
5395 " ((((((((((((((((((((((((((((((((((((((((((((((((\n");
5397 if (call_save_p)
5399 mode = HARD_REGNO_CALLER_SAVE_MODE (hard_regno,
5400 hard_regno_nregs[hard_regno][mode],
5401 mode);
5402 new_reg = lra_create_new_reg (mode, NULL_RTX, NO_REGS, "save");
5404 else
5406 rclass = choose_split_class (rclass, hard_regno, mode);
5407 if (rclass == NO_REGS)
5409 if (lra_dump_file != NULL)
5411 fprintf (lra_dump_file,
5412 " Rejecting split of %d(%s): "
5413 "no good reg class for %d(%s)\n",
5414 original_regno,
5415 reg_class_names[lra_get_allocno_class (original_regno)],
5416 hard_regno,
5417 reg_class_names[REGNO_REG_CLASS (hard_regno)]);
5418 fprintf
5419 (lra_dump_file,
5420 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5422 return false;
5424 /* Split_if_necessary can split hard registers used as part of a
5425 multi-register mode but splits each register individually. The
5426 mode used for each independent register may not be supported
5427 so reject the split. Splitting the wider mode should theoretically
5428 be possible but is not implemented. */
5429 if (! HARD_REGNO_MODE_OK (hard_regno, mode))
5431 if (lra_dump_file != NULL)
5433 fprintf (lra_dump_file,
5434 " Rejecting split of %d(%s): unsuitable mode %s\n",
5435 original_regno,
5436 reg_class_names[lra_get_allocno_class (original_regno)],
5437 GET_MODE_NAME (mode));
5438 fprintf
5439 (lra_dump_file,
5440 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5442 return false;
5444 new_reg = lra_create_new_reg (mode, original_reg, rclass, "split");
5445 reg_renumber[REGNO (new_reg)] = hard_regno;
5447 save = emit_spill_move (true, new_reg, original_reg);
5448 if (NEXT_INSN (save) != NULL_RTX && !call_save_p)
5450 if (lra_dump_file != NULL)
5452 fprintf
5453 (lra_dump_file,
5454 " Rejecting split %d->%d resulting in > 2 save insns:\n",
5455 original_regno, REGNO (new_reg));
5456 dump_rtl_slim (lra_dump_file, save, NULL, -1, 0);
5457 fprintf (lra_dump_file,
5458 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5460 return false;
5462 restore = emit_spill_move (false, new_reg, original_reg);
5463 if (NEXT_INSN (restore) != NULL_RTX && !call_save_p)
5465 if (lra_dump_file != NULL)
5467 fprintf (lra_dump_file,
5468 " Rejecting split %d->%d "
5469 "resulting in > 2 restore insns:\n",
5470 original_regno, REGNO (new_reg));
5471 dump_rtl_slim (lra_dump_file, restore, NULL, -1, 0);
5472 fprintf (lra_dump_file,
5473 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5475 return false;
5477 after_p = usage_insns[original_regno].after_p;
5478 lra_reg_info[REGNO (new_reg)].restore_rtx = regno_reg_rtx[original_regno];
5479 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
5480 bitmap_set_bit (&check_only_regs, original_regno);
5481 bitmap_set_bit (&lra_split_regs, REGNO (new_reg));
5482 for (;;)
5484 if (GET_CODE (next_usage_insns) != INSN_LIST)
5486 usage_insn = next_usage_insns;
5487 break;
5489 usage_insn = XEXP (next_usage_insns, 0);
5490 lra_assert (DEBUG_INSN_P (usage_insn));
5491 next_usage_insns = XEXP (next_usage_insns, 1);
5492 lra_substitute_pseudo (&usage_insn, original_regno, new_reg, false);
5493 lra_update_insn_regno_info (as_a <rtx_insn *> (usage_insn));
5494 if (lra_dump_file != NULL)
5496 fprintf (lra_dump_file, " Split reuse change %d->%d:\n",
5497 original_regno, REGNO (new_reg));
5498 dump_insn_slim (lra_dump_file, as_a <rtx_insn *> (usage_insn));
5501 lra_assert (NOTE_P (usage_insn) || NONDEBUG_INSN_P (usage_insn));
5502 lra_assert (usage_insn != insn || (after_p && before_p));
5503 lra_process_new_insns (as_a <rtx_insn *> (usage_insn),
5504 after_p ? NULL : restore,
5505 after_p ? restore : NULL,
5506 call_save_p
5507 ? "Add reg<-save" : "Add reg<-split");
5508 lra_process_new_insns (insn, before_p ? save : NULL,
5509 before_p ? NULL : save,
5510 call_save_p
5511 ? "Add save<-reg" : "Add split<-reg");
5512 if (nregs > 1)
5513 /* If we are trying to split multi-register. We should check
5514 conflicts on the next assignment sub-pass. IRA can allocate on
5515 sub-register levels, LRA do this on pseudos level right now and
5516 this discrepancy may create allocation conflicts after
5517 splitting. */
5518 lra_risky_transformations_p = true;
5519 if (lra_dump_file != NULL)
5520 fprintf (lra_dump_file,
5521 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5522 return true;
5525 /* Recognize that we need a split transformation for insn INSN, which
5526 defines or uses REGNO in its insn biggest MODE (we use it only if
5527 REGNO is a hard register). POTENTIAL_RELOAD_HARD_REGS contains
5528 hard registers which might be used for reloads since the EBB end.
5529 Put the save before INSN if BEFORE_P is true. MAX_UID is maximla
5530 uid before starting INSN processing. Return true if we succeed in
5531 such transformation. */
5532 static bool
5533 split_if_necessary (int regno, machine_mode mode,
5534 HARD_REG_SET potential_reload_hard_regs,
5535 bool before_p, rtx_insn *insn, int max_uid)
5537 bool res = false;
5538 int i, nregs = 1;
5539 rtx next_usage_insns;
5541 if (regno < FIRST_PSEUDO_REGISTER)
5542 nregs = hard_regno_nregs[regno][mode];
5543 for (i = 0; i < nregs; i++)
5544 if (usage_insns[regno + i].check == curr_usage_insns_check
5545 && (next_usage_insns = usage_insns[regno + i].insns) != NULL_RTX
5546 /* To avoid processing the register twice or more. */
5547 && ((GET_CODE (next_usage_insns) != INSN_LIST
5548 && INSN_UID (next_usage_insns) < max_uid)
5549 || (GET_CODE (next_usage_insns) == INSN_LIST
5550 && (INSN_UID (XEXP (next_usage_insns, 0)) < max_uid)))
5551 && need_for_split_p (potential_reload_hard_regs, regno + i)
5552 && split_reg (before_p, regno + i, insn, next_usage_insns))
5553 res = true;
5554 return res;
5557 /* Return TRUE if rtx X is considered as an invariant for
5558 inheritance. */
5559 static bool
5560 invariant_p (const_rtx x)
5562 machine_mode mode;
5563 const char *fmt;
5564 enum rtx_code code;
5565 int i, j;
5567 code = GET_CODE (x);
5568 mode = GET_MODE (x);
5569 if (code == SUBREG)
5571 x = SUBREG_REG (x);
5572 code = GET_CODE (x);
5573 if (GET_MODE_SIZE (GET_MODE (x)) > GET_MODE_SIZE (mode))
5574 mode = GET_MODE (x);
5577 if (MEM_P (x))
5578 return false;
5580 if (REG_P (x))
5582 int i, nregs, regno = REGNO (x);
5584 if (regno >= FIRST_PSEUDO_REGISTER || regno == STACK_POINTER_REGNUM
5585 || TEST_HARD_REG_BIT (eliminable_regset, regno)
5586 || GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
5587 return false;
5588 nregs = hard_regno_nregs[regno][mode];
5589 for (i = 0; i < nregs; i++)
5590 if (! fixed_regs[regno + i]
5591 /* A hard register may be clobbered in the current insn
5592 but we can ignore this case because if the hard
5593 register is used it should be set somewhere after the
5594 clobber. */
5595 || bitmap_bit_p (&invalid_invariant_regs, regno + i))
5596 return false;
5598 fmt = GET_RTX_FORMAT (code);
5599 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
5601 if (fmt[i] == 'e')
5603 if (! invariant_p (XEXP (x, i)))
5604 return false;
5606 else if (fmt[i] == 'E')
5608 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
5609 if (! invariant_p (XVECEXP (x, i, j)))
5610 return false;
5613 return true;
5616 /* We have 'dest_reg <- invariant'. Let us try to make an invariant
5617 inheritance transformation (using dest_reg instead invariant in a
5618 subsequent insn). */
5619 static bool
5620 process_invariant_for_inheritance (rtx dst_reg, rtx invariant_rtx)
5622 invariant_ptr_t invariant_ptr;
5623 rtx_insn *insn, *new_insns;
5624 rtx insn_set, insn_reg, new_reg;
5625 int insn_regno;
5626 bool succ_p = false;
5627 int dst_regno = REGNO (dst_reg);
5628 enum machine_mode dst_mode = GET_MODE (dst_reg);
5629 enum reg_class cl = lra_get_allocno_class (dst_regno), insn_reg_cl;
5631 invariant_ptr = insert_invariant (invariant_rtx);
5632 if ((insn = invariant_ptr->insn) != NULL_RTX)
5634 /* We have a subsequent insn using the invariant. */
5635 insn_set = single_set (insn);
5636 lra_assert (insn_set != NULL);
5637 insn_reg = SET_DEST (insn_set);
5638 lra_assert (REG_P (insn_reg));
5639 insn_regno = REGNO (insn_reg);
5640 insn_reg_cl = lra_get_allocno_class (insn_regno);
5642 if (dst_mode == GET_MODE (insn_reg)
5643 /* We should consider only result move reg insns which are
5644 cheap. */
5645 && targetm.register_move_cost (dst_mode, cl, insn_reg_cl) == 2
5646 && targetm.register_move_cost (dst_mode, cl, cl) == 2)
5648 if (lra_dump_file != NULL)
5649 fprintf (lra_dump_file,
5650 " [[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[\n");
5651 new_reg = lra_create_new_reg (dst_mode, dst_reg,
5652 cl, "invariant inheritance");
5653 bitmap_set_bit (&lra_inheritance_pseudos, REGNO (new_reg));
5654 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
5655 lra_reg_info[REGNO (new_reg)].restore_rtx = PATTERN (insn);
5656 start_sequence ();
5657 lra_emit_move (new_reg, dst_reg);
5658 new_insns = get_insns ();
5659 end_sequence ();
5660 lra_process_new_insns (curr_insn, NULL, new_insns,
5661 "Add invariant inheritance<-original");
5662 start_sequence ();
5663 lra_emit_move (SET_DEST (insn_set), new_reg);
5664 new_insns = get_insns ();
5665 end_sequence ();
5666 lra_process_new_insns (insn, NULL, new_insns,
5667 "Changing reload<-inheritance");
5668 lra_set_insn_deleted (insn);
5669 succ_p = true;
5670 if (lra_dump_file != NULL)
5672 fprintf (lra_dump_file,
5673 " Invariant inheritance reuse change %d (bb%d):\n",
5674 REGNO (new_reg), BLOCK_FOR_INSN (insn)->index);
5675 dump_insn_slim (lra_dump_file, insn);
5676 fprintf (lra_dump_file,
5677 " ]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]\n");
5681 invariant_ptr->insn = curr_insn;
5682 return succ_p;
5685 /* Check only registers living at the current program point in the
5686 current EBB. */
5687 static bitmap_head live_regs;
5689 /* Update live info in EBB given by its HEAD and TAIL insns after
5690 inheritance/split transformation. The function removes dead moves
5691 too. */
5692 static void
5693 update_ebb_live_info (rtx_insn *head, rtx_insn *tail)
5695 unsigned int j;
5696 int i, regno;
5697 bool live_p;
5698 rtx_insn *prev_insn;
5699 rtx set;
5700 bool remove_p;
5701 basic_block last_bb, prev_bb, curr_bb;
5702 bitmap_iterator bi;
5703 struct lra_insn_reg *reg;
5704 edge e;
5705 edge_iterator ei;
5707 last_bb = BLOCK_FOR_INSN (tail);
5708 prev_bb = NULL;
5709 for (curr_insn = tail;
5710 curr_insn != PREV_INSN (head);
5711 curr_insn = prev_insn)
5713 prev_insn = PREV_INSN (curr_insn);
5714 /* We need to process empty blocks too. They contain
5715 NOTE_INSN_BASIC_BLOCK referring for the basic block. */
5716 if (NOTE_P (curr_insn) && NOTE_KIND (curr_insn) != NOTE_INSN_BASIC_BLOCK)
5717 continue;
5718 curr_bb = BLOCK_FOR_INSN (curr_insn);
5719 if (curr_bb != prev_bb)
5721 if (prev_bb != NULL)
5723 /* Update df_get_live_in (prev_bb): */
5724 EXECUTE_IF_SET_IN_BITMAP (&check_only_regs, 0, j, bi)
5725 if (bitmap_bit_p (&live_regs, j))
5726 bitmap_set_bit (df_get_live_in (prev_bb), j);
5727 else
5728 bitmap_clear_bit (df_get_live_in (prev_bb), j);
5730 if (curr_bb != last_bb)
5732 /* Update df_get_live_out (curr_bb): */
5733 EXECUTE_IF_SET_IN_BITMAP (&check_only_regs, 0, j, bi)
5735 live_p = bitmap_bit_p (&live_regs, j);
5736 if (! live_p)
5737 FOR_EACH_EDGE (e, ei, curr_bb->succs)
5738 if (bitmap_bit_p (df_get_live_in (e->dest), j))
5740 live_p = true;
5741 break;
5743 if (live_p)
5744 bitmap_set_bit (df_get_live_out (curr_bb), j);
5745 else
5746 bitmap_clear_bit (df_get_live_out (curr_bb), j);
5749 prev_bb = curr_bb;
5750 bitmap_and (&live_regs, &check_only_regs, df_get_live_out (curr_bb));
5752 if (! NONDEBUG_INSN_P (curr_insn))
5753 continue;
5754 curr_id = lra_get_insn_recog_data (curr_insn);
5755 curr_static_id = curr_id->insn_static_data;
5756 remove_p = false;
5757 if ((set = single_set (curr_insn)) != NULL_RTX
5758 && REG_P (SET_DEST (set))
5759 && (regno = REGNO (SET_DEST (set))) >= FIRST_PSEUDO_REGISTER
5760 && SET_DEST (set) != pic_offset_table_rtx
5761 && bitmap_bit_p (&check_only_regs, regno)
5762 && ! bitmap_bit_p (&live_regs, regno))
5763 remove_p = true;
5764 /* See which defined values die here. */
5765 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5766 if (reg->type == OP_OUT && ! reg->subreg_p)
5767 bitmap_clear_bit (&live_regs, reg->regno);
5768 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
5769 if (reg->type == OP_OUT && ! reg->subreg_p)
5770 bitmap_clear_bit (&live_regs, reg->regno);
5771 if (curr_id->arg_hard_regs != NULL)
5772 /* Make clobbered argument hard registers die. */
5773 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
5774 if (regno >= FIRST_PSEUDO_REGISTER)
5775 bitmap_clear_bit (&live_regs, regno - FIRST_PSEUDO_REGISTER);
5776 /* Mark each used value as live. */
5777 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5778 if (reg->type != OP_OUT
5779 && bitmap_bit_p (&check_only_regs, reg->regno))
5780 bitmap_set_bit (&live_regs, reg->regno);
5781 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
5782 if (reg->type != OP_OUT
5783 && bitmap_bit_p (&check_only_regs, reg->regno))
5784 bitmap_set_bit (&live_regs, reg->regno);
5785 if (curr_id->arg_hard_regs != NULL)
5786 /* Make used argument hard registers live. */
5787 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
5788 if (regno < FIRST_PSEUDO_REGISTER
5789 && bitmap_bit_p (&check_only_regs, regno))
5790 bitmap_set_bit (&live_regs, regno);
5791 /* It is quite important to remove dead move insns because it
5792 means removing dead store. We don't need to process them for
5793 constraints. */
5794 if (remove_p)
5796 if (lra_dump_file != NULL)
5798 fprintf (lra_dump_file, " Removing dead insn:\n ");
5799 dump_insn_slim (lra_dump_file, curr_insn);
5801 lra_set_insn_deleted (curr_insn);
5806 /* The structure describes info to do an inheritance for the current
5807 insn. We need to collect such info first before doing the
5808 transformations because the transformations change the insn
5809 internal representation. */
5810 struct to_inherit
5812 /* Original regno. */
5813 int regno;
5814 /* Subsequent insns which can inherit original reg value. */
5815 rtx insns;
5818 /* Array containing all info for doing inheritance from the current
5819 insn. */
5820 static struct to_inherit to_inherit[LRA_MAX_INSN_RELOADS];
5822 /* Number elements in the previous array. */
5823 static int to_inherit_num;
5825 /* Add inheritance info REGNO and INSNS. Their meaning is described in
5826 structure to_inherit. */
5827 static void
5828 add_to_inherit (int regno, rtx insns)
5830 int i;
5832 for (i = 0; i < to_inherit_num; i++)
5833 if (to_inherit[i].regno == regno)
5834 return;
5835 lra_assert (to_inherit_num < LRA_MAX_INSN_RELOADS);
5836 to_inherit[to_inherit_num].regno = regno;
5837 to_inherit[to_inherit_num++].insns = insns;
5840 /* Return the last non-debug insn in basic block BB, or the block begin
5841 note if none. */
5842 static rtx_insn *
5843 get_last_insertion_point (basic_block bb)
5845 rtx_insn *insn;
5847 FOR_BB_INSNS_REVERSE (bb, insn)
5848 if (NONDEBUG_INSN_P (insn) || NOTE_INSN_BASIC_BLOCK_P (insn))
5849 return insn;
5850 gcc_unreachable ();
5853 /* Set up RES by registers living on edges FROM except the edge (FROM,
5854 TO) or by registers set up in a jump insn in BB FROM. */
5855 static void
5856 get_live_on_other_edges (basic_block from, basic_block to, bitmap res)
5858 rtx_insn *last;
5859 struct lra_insn_reg *reg;
5860 edge e;
5861 edge_iterator ei;
5863 lra_assert (to != NULL);
5864 bitmap_clear (res);
5865 FOR_EACH_EDGE (e, ei, from->succs)
5866 if (e->dest != to)
5867 bitmap_ior_into (res, df_get_live_in (e->dest));
5868 last = get_last_insertion_point (from);
5869 if (! JUMP_P (last))
5870 return;
5871 curr_id = lra_get_insn_recog_data (last);
5872 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5873 if (reg->type != OP_IN)
5874 bitmap_set_bit (res, reg->regno);
5877 /* Used as a temporary results of some bitmap calculations. */
5878 static bitmap_head temp_bitmap;
5880 /* We split for reloads of small class of hard regs. The following
5881 defines how many hard regs the class should have to be qualified as
5882 small. The code is mostly oriented to x86/x86-64 architecture
5883 where some insns need to use only specific register or pair of
5884 registers and these register can live in RTL explicitly, e.g. for
5885 parameter passing. */
5886 static const int max_small_class_regs_num = 2;
5888 /* Do inheritance/split transformations in EBB starting with HEAD and
5889 finishing on TAIL. We process EBB insns in the reverse order.
5890 Return true if we did any inheritance/split transformation in the
5891 EBB.
5893 We should avoid excessive splitting which results in worse code
5894 because of inaccurate cost calculations for spilling new split
5895 pseudos in such case. To achieve this we do splitting only if
5896 register pressure is high in given basic block and there are reload
5897 pseudos requiring hard registers. We could do more register
5898 pressure calculations at any given program point to avoid necessary
5899 splitting even more but it is to expensive and the current approach
5900 works well enough. */
5901 static bool
5902 inherit_in_ebb (rtx_insn *head, rtx_insn *tail)
5904 int i, src_regno, dst_regno, nregs;
5905 bool change_p, succ_p, update_reloads_num_p;
5906 rtx_insn *prev_insn, *last_insn;
5907 rtx next_usage_insns, curr_set;
5908 enum reg_class cl;
5909 struct lra_insn_reg *reg;
5910 basic_block last_processed_bb, curr_bb = NULL;
5911 HARD_REG_SET potential_reload_hard_regs, live_hard_regs;
5912 bitmap to_process;
5913 unsigned int j;
5914 bitmap_iterator bi;
5915 bool head_p, after_p;
5917 change_p = false;
5918 curr_usage_insns_check++;
5919 clear_invariants ();
5920 reloads_num = calls_num = 0;
5921 bitmap_clear (&check_only_regs);
5922 bitmap_clear (&invalid_invariant_regs);
5923 last_processed_bb = NULL;
5924 CLEAR_HARD_REG_SET (potential_reload_hard_regs);
5925 COPY_HARD_REG_SET (live_hard_regs, eliminable_regset);
5926 IOR_HARD_REG_SET (live_hard_regs, lra_no_alloc_regs);
5927 /* We don't process new insns generated in the loop. */
5928 for (curr_insn = tail; curr_insn != PREV_INSN (head); curr_insn = prev_insn)
5930 prev_insn = PREV_INSN (curr_insn);
5931 if (BLOCK_FOR_INSN (curr_insn) != NULL)
5932 curr_bb = BLOCK_FOR_INSN (curr_insn);
5933 if (last_processed_bb != curr_bb)
5935 /* We are at the end of BB. Add qualified living
5936 pseudos for potential splitting. */
5937 to_process = df_get_live_out (curr_bb);
5938 if (last_processed_bb != NULL)
5940 /* We are somewhere in the middle of EBB. */
5941 get_live_on_other_edges (curr_bb, last_processed_bb,
5942 &temp_bitmap);
5943 to_process = &temp_bitmap;
5945 last_processed_bb = curr_bb;
5946 last_insn = get_last_insertion_point (curr_bb);
5947 after_p = (! JUMP_P (last_insn)
5948 && (! CALL_P (last_insn)
5949 || (find_reg_note (last_insn,
5950 REG_NORETURN, NULL_RTX) == NULL_RTX
5951 && ! SIBLING_CALL_P (last_insn))));
5952 CLEAR_HARD_REG_SET (potential_reload_hard_regs);
5953 EXECUTE_IF_SET_IN_BITMAP (to_process, 0, j, bi)
5955 if ((int) j >= lra_constraint_new_regno_start)
5956 break;
5957 if (j < FIRST_PSEUDO_REGISTER || reg_renumber[j] >= 0)
5959 if (j < FIRST_PSEUDO_REGISTER)
5960 SET_HARD_REG_BIT (live_hard_regs, j);
5961 else
5962 add_to_hard_reg_set (&live_hard_regs,
5963 PSEUDO_REGNO_MODE (j),
5964 reg_renumber[j]);
5965 setup_next_usage_insn (j, last_insn, reloads_num, after_p);
5969 src_regno = dst_regno = -1;
5970 curr_set = single_set (curr_insn);
5971 if (curr_set != NULL_RTX && REG_P (SET_DEST (curr_set)))
5972 dst_regno = REGNO (SET_DEST (curr_set));
5973 if (curr_set != NULL_RTX && REG_P (SET_SRC (curr_set)))
5974 src_regno = REGNO (SET_SRC (curr_set));
5975 update_reloads_num_p = true;
5976 if (src_regno < lra_constraint_new_regno_start
5977 && src_regno >= FIRST_PSEUDO_REGISTER
5978 && reg_renumber[src_regno] < 0
5979 && dst_regno >= lra_constraint_new_regno_start
5980 && (cl = lra_get_allocno_class (dst_regno)) != NO_REGS)
5982 /* 'reload_pseudo <- original_pseudo'. */
5983 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
5984 reloads_num++;
5985 update_reloads_num_p = false;
5986 succ_p = false;
5987 if (usage_insns[src_regno].check == curr_usage_insns_check
5988 && (next_usage_insns = usage_insns[src_regno].insns) != NULL_RTX)
5989 succ_p = inherit_reload_reg (false, src_regno, cl,
5990 curr_insn, next_usage_insns);
5991 if (succ_p)
5992 change_p = true;
5993 else
5994 setup_next_usage_insn (src_regno, curr_insn, reloads_num, false);
5995 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
5996 IOR_HARD_REG_SET (potential_reload_hard_regs,
5997 reg_class_contents[cl]);
5999 else if (src_regno < 0
6000 && dst_regno >= lra_constraint_new_regno_start
6001 && invariant_p (SET_SRC (curr_set))
6002 && (cl = lra_get_allocno_class (dst_regno)) != NO_REGS
6003 && ! bitmap_bit_p (&invalid_invariant_regs, dst_regno)
6004 && ! bitmap_bit_p (&invalid_invariant_regs,
6005 ORIGINAL_REGNO(regno_reg_rtx[dst_regno])))
6007 /* 'reload_pseudo <- invariant'. */
6008 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6009 reloads_num++;
6010 update_reloads_num_p = false;
6011 if (process_invariant_for_inheritance (SET_DEST (curr_set), SET_SRC (curr_set)))
6012 change_p = true;
6013 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6014 IOR_HARD_REG_SET (potential_reload_hard_regs,
6015 reg_class_contents[cl]);
6017 else if (src_regno >= lra_constraint_new_regno_start
6018 && dst_regno < lra_constraint_new_regno_start
6019 && dst_regno >= FIRST_PSEUDO_REGISTER
6020 && reg_renumber[dst_regno] < 0
6021 && (cl = lra_get_allocno_class (src_regno)) != NO_REGS
6022 && usage_insns[dst_regno].check == curr_usage_insns_check
6023 && (next_usage_insns
6024 = usage_insns[dst_regno].insns) != NULL_RTX)
6026 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6027 reloads_num++;
6028 update_reloads_num_p = false;
6029 /* 'original_pseudo <- reload_pseudo'. */
6030 if (! JUMP_P (curr_insn)
6031 && inherit_reload_reg (true, dst_regno, cl,
6032 curr_insn, next_usage_insns))
6033 change_p = true;
6034 /* Invalidate. */
6035 usage_insns[dst_regno].check = 0;
6036 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6037 IOR_HARD_REG_SET (potential_reload_hard_regs,
6038 reg_class_contents[cl]);
6040 else if (INSN_P (curr_insn))
6042 int iter;
6043 int max_uid = get_max_uid ();
6045 curr_id = lra_get_insn_recog_data (curr_insn);
6046 curr_static_id = curr_id->insn_static_data;
6047 to_inherit_num = 0;
6048 /* Process insn definitions. */
6049 for (iter = 0; iter < 2; iter++)
6050 for (reg = iter == 0 ? curr_id->regs : curr_static_id->hard_regs;
6051 reg != NULL;
6052 reg = reg->next)
6053 if (reg->type != OP_IN
6054 && (dst_regno = reg->regno) < lra_constraint_new_regno_start)
6056 if (dst_regno >= FIRST_PSEUDO_REGISTER && reg->type == OP_OUT
6057 && reg_renumber[dst_regno] < 0 && ! reg->subreg_p
6058 && usage_insns[dst_regno].check == curr_usage_insns_check
6059 && (next_usage_insns
6060 = usage_insns[dst_regno].insns) != NULL_RTX)
6062 struct lra_insn_reg *r;
6064 for (r = curr_id->regs; r != NULL; r = r->next)
6065 if (r->type != OP_OUT && r->regno == dst_regno)
6066 break;
6067 /* Don't do inheritance if the pseudo is also
6068 used in the insn. */
6069 if (r == NULL)
6070 /* We can not do inheritance right now
6071 because the current insn reg info (chain
6072 regs) can change after that. */
6073 add_to_inherit (dst_regno, next_usage_insns);
6075 /* We can not process one reg twice here because of
6076 usage_insns invalidation. */
6077 if ((dst_regno < FIRST_PSEUDO_REGISTER
6078 || reg_renumber[dst_regno] >= 0)
6079 && ! reg->subreg_p && reg->type != OP_IN)
6081 HARD_REG_SET s;
6083 if (split_if_necessary (dst_regno, reg->biggest_mode,
6084 potential_reload_hard_regs,
6085 false, curr_insn, max_uid))
6086 change_p = true;
6087 CLEAR_HARD_REG_SET (s);
6088 if (dst_regno < FIRST_PSEUDO_REGISTER)
6089 add_to_hard_reg_set (&s, reg->biggest_mode, dst_regno);
6090 else
6091 add_to_hard_reg_set (&s, PSEUDO_REGNO_MODE (dst_regno),
6092 reg_renumber[dst_regno]);
6093 AND_COMPL_HARD_REG_SET (live_hard_regs, s);
6095 /* We should invalidate potential inheritance or
6096 splitting for the current insn usages to the next
6097 usage insns (see code below) as the output pseudo
6098 prevents this. */
6099 if ((dst_regno >= FIRST_PSEUDO_REGISTER
6100 && reg_renumber[dst_regno] < 0)
6101 || (reg->type == OP_OUT && ! reg->subreg_p
6102 && (dst_regno < FIRST_PSEUDO_REGISTER
6103 || reg_renumber[dst_regno] >= 0)))
6105 /* Invalidate and mark definitions. */
6106 if (dst_regno >= FIRST_PSEUDO_REGISTER)
6107 usage_insns[dst_regno].check = -(int) INSN_UID (curr_insn);
6108 else
6110 nregs = hard_regno_nregs[dst_regno][reg->biggest_mode];
6111 for (i = 0; i < nregs; i++)
6112 usage_insns[dst_regno + i].check
6113 = -(int) INSN_UID (curr_insn);
6117 /* Process clobbered call regs. */
6118 if (curr_id->arg_hard_regs != NULL)
6119 for (i = 0; (dst_regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6120 if (dst_regno >= FIRST_PSEUDO_REGISTER)
6121 usage_insns[dst_regno - FIRST_PSEUDO_REGISTER].check
6122 = -(int) INSN_UID (curr_insn);
6123 if (! JUMP_P (curr_insn))
6124 for (i = 0; i < to_inherit_num; i++)
6125 if (inherit_reload_reg (true, to_inherit[i].regno,
6126 ALL_REGS, curr_insn,
6127 to_inherit[i].insns))
6128 change_p = true;
6129 if (CALL_P (curr_insn))
6131 rtx cheap, pat, dest;
6132 rtx_insn *restore;
6133 int regno, hard_regno;
6135 calls_num++;
6136 if ((cheap = find_reg_note (curr_insn,
6137 REG_RETURNED, NULL_RTX)) != NULL_RTX
6138 && ((cheap = XEXP (cheap, 0)), true)
6139 && (regno = REGNO (cheap)) >= FIRST_PSEUDO_REGISTER
6140 && (hard_regno = reg_renumber[regno]) >= 0
6141 /* If there are pending saves/restores, the
6142 optimization is not worth. */
6143 && usage_insns[regno].calls_num == calls_num - 1
6144 && TEST_HARD_REG_BIT (call_used_reg_set, hard_regno))
6146 /* Restore the pseudo from the call result as
6147 REG_RETURNED note says that the pseudo value is
6148 in the call result and the pseudo is an argument
6149 of the call. */
6150 pat = PATTERN (curr_insn);
6151 if (GET_CODE (pat) == PARALLEL)
6152 pat = XVECEXP (pat, 0, 0);
6153 dest = SET_DEST (pat);
6154 /* For multiple return values dest is PARALLEL.
6155 Currently we handle only single return value case. */
6156 if (REG_P (dest))
6158 start_sequence ();
6159 emit_move_insn (cheap, copy_rtx (dest));
6160 restore = get_insns ();
6161 end_sequence ();
6162 lra_process_new_insns (curr_insn, NULL, restore,
6163 "Inserting call parameter restore");
6164 /* We don't need to save/restore of the pseudo from
6165 this call. */
6166 usage_insns[regno].calls_num = calls_num;
6167 bitmap_set_bit (&check_only_regs, regno);
6171 to_inherit_num = 0;
6172 /* Process insn usages. */
6173 for (iter = 0; iter < 2; iter++)
6174 for (reg = iter == 0 ? curr_id->regs : curr_static_id->hard_regs;
6175 reg != NULL;
6176 reg = reg->next)
6177 if ((reg->type != OP_OUT
6178 || (reg->type == OP_OUT && reg->subreg_p))
6179 && (src_regno = reg->regno) < lra_constraint_new_regno_start)
6181 if (src_regno >= FIRST_PSEUDO_REGISTER
6182 && reg_renumber[src_regno] < 0 && reg->type == OP_IN)
6184 if (usage_insns[src_regno].check == curr_usage_insns_check
6185 && (next_usage_insns
6186 = usage_insns[src_regno].insns) != NULL_RTX
6187 && NONDEBUG_INSN_P (curr_insn))
6188 add_to_inherit (src_regno, next_usage_insns);
6189 else if (usage_insns[src_regno].check
6190 != -(int) INSN_UID (curr_insn))
6191 /* Add usages but only if the reg is not set up
6192 in the same insn. */
6193 add_next_usage_insn (src_regno, curr_insn, reloads_num);
6195 else if (src_regno < FIRST_PSEUDO_REGISTER
6196 || reg_renumber[src_regno] >= 0)
6198 bool before_p;
6199 rtx_insn *use_insn = curr_insn;
6201 before_p = (JUMP_P (curr_insn)
6202 || (CALL_P (curr_insn) && reg->type == OP_IN));
6203 if (NONDEBUG_INSN_P (curr_insn)
6204 && (! JUMP_P (curr_insn) || reg->type == OP_IN)
6205 && split_if_necessary (src_regno, reg->biggest_mode,
6206 potential_reload_hard_regs,
6207 before_p, curr_insn, max_uid))
6209 if (reg->subreg_p)
6210 lra_risky_transformations_p = true;
6211 change_p = true;
6212 /* Invalidate. */
6213 usage_insns[src_regno].check = 0;
6214 if (before_p)
6215 use_insn = PREV_INSN (curr_insn);
6217 if (NONDEBUG_INSN_P (curr_insn))
6219 if (src_regno < FIRST_PSEUDO_REGISTER)
6220 add_to_hard_reg_set (&live_hard_regs,
6221 reg->biggest_mode, src_regno);
6222 else
6223 add_to_hard_reg_set (&live_hard_regs,
6224 PSEUDO_REGNO_MODE (src_regno),
6225 reg_renumber[src_regno]);
6227 add_next_usage_insn (src_regno, use_insn, reloads_num);
6230 /* Process used call regs. */
6231 if (curr_id->arg_hard_regs != NULL)
6232 for (i = 0; (src_regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6233 if (src_regno < FIRST_PSEUDO_REGISTER)
6235 SET_HARD_REG_BIT (live_hard_regs, src_regno);
6236 add_next_usage_insn (src_regno, curr_insn, reloads_num);
6238 for (i = 0; i < to_inherit_num; i++)
6240 src_regno = to_inherit[i].regno;
6241 if (inherit_reload_reg (false, src_regno, ALL_REGS,
6242 curr_insn, to_inherit[i].insns))
6243 change_p = true;
6244 else
6245 setup_next_usage_insn (src_regno, curr_insn, reloads_num, false);
6248 if (update_reloads_num_p
6249 && NONDEBUG_INSN_P (curr_insn) && curr_set != NULL_RTX)
6251 int regno = -1;
6252 if ((REG_P (SET_DEST (curr_set))
6253 && (regno = REGNO (SET_DEST (curr_set))) >= lra_constraint_new_regno_start
6254 && reg_renumber[regno] < 0
6255 && (cl = lra_get_allocno_class (regno)) != NO_REGS)
6256 || (REG_P (SET_SRC (curr_set))
6257 && (regno = REGNO (SET_SRC (curr_set))) >= lra_constraint_new_regno_start
6258 && reg_renumber[regno] < 0
6259 && (cl = lra_get_allocno_class (regno)) != NO_REGS))
6261 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6262 reloads_num++;
6263 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6264 IOR_HARD_REG_SET (potential_reload_hard_regs,
6265 reg_class_contents[cl]);
6268 if (NONDEBUG_INSN_P (curr_insn))
6270 int regno;
6272 /* Invalidate invariants with changed regs. */
6273 curr_id = lra_get_insn_recog_data (curr_insn);
6274 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
6275 if (reg->type != OP_IN)
6277 bitmap_set_bit (&invalid_invariant_regs, reg->regno);
6278 bitmap_set_bit (&invalid_invariant_regs,
6279 ORIGINAL_REGNO (regno_reg_rtx[reg->regno]));
6281 curr_static_id = curr_id->insn_static_data;
6282 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
6283 if (reg->type != OP_IN)
6284 bitmap_set_bit (&invalid_invariant_regs, reg->regno);
6285 if (curr_id->arg_hard_regs != NULL)
6286 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6287 if (regno >= FIRST_PSEUDO_REGISTER)
6288 bitmap_set_bit (&invalid_invariant_regs,
6289 regno - FIRST_PSEUDO_REGISTER);
6291 /* We reached the start of the current basic block. */
6292 if (prev_insn == NULL_RTX || prev_insn == PREV_INSN (head)
6293 || BLOCK_FOR_INSN (prev_insn) != curr_bb)
6295 /* We reached the beginning of the current block -- do
6296 rest of spliting in the current BB. */
6297 to_process = df_get_live_in (curr_bb);
6298 if (BLOCK_FOR_INSN (head) != curr_bb)
6300 /* We are somewhere in the middle of EBB. */
6301 get_live_on_other_edges (EDGE_PRED (curr_bb, 0)->src,
6302 curr_bb, &temp_bitmap);
6303 to_process = &temp_bitmap;
6305 head_p = true;
6306 EXECUTE_IF_SET_IN_BITMAP (to_process, 0, j, bi)
6308 if ((int) j >= lra_constraint_new_regno_start)
6309 break;
6310 if (((int) j < FIRST_PSEUDO_REGISTER || reg_renumber[j] >= 0)
6311 && usage_insns[j].check == curr_usage_insns_check
6312 && (next_usage_insns = usage_insns[j].insns) != NULL_RTX)
6314 if (need_for_split_p (potential_reload_hard_regs, j))
6316 if (lra_dump_file != NULL && head_p)
6318 fprintf (lra_dump_file,
6319 " ----------------------------------\n");
6320 head_p = false;
6322 if (split_reg (false, j, bb_note (curr_bb),
6323 next_usage_insns))
6324 change_p = true;
6326 usage_insns[j].check = 0;
6331 return change_p;
6334 /* This value affects EBB forming. If probability of edge from EBB to
6335 a BB is not greater than the following value, we don't add the BB
6336 to EBB. */
6337 #define EBB_PROBABILITY_CUTOFF \
6338 ((REG_BR_PROB_BASE * LRA_INHERITANCE_EBB_PROBABILITY_CUTOFF) / 100)
6340 /* Current number of inheritance/split iteration. */
6341 int lra_inheritance_iter;
6343 /* Entry function for inheritance/split pass. */
6344 void
6345 lra_inheritance (void)
6347 int i;
6348 basic_block bb, start_bb;
6349 edge e;
6351 lra_inheritance_iter++;
6352 if (lra_inheritance_iter > LRA_MAX_INHERITANCE_PASSES)
6353 return;
6354 timevar_push (TV_LRA_INHERITANCE);
6355 if (lra_dump_file != NULL)
6356 fprintf (lra_dump_file, "\n********** Inheritance #%d: **********\n\n",
6357 lra_inheritance_iter);
6358 curr_usage_insns_check = 0;
6359 usage_insns = XNEWVEC (struct usage_insns, lra_constraint_new_regno_start);
6360 for (i = 0; i < lra_constraint_new_regno_start; i++)
6361 usage_insns[i].check = 0;
6362 bitmap_initialize (&check_only_regs, &reg_obstack);
6363 bitmap_initialize (&invalid_invariant_regs, &reg_obstack);
6364 bitmap_initialize (&live_regs, &reg_obstack);
6365 bitmap_initialize (&temp_bitmap, &reg_obstack);
6366 bitmap_initialize (&ebb_global_regs, &reg_obstack);
6367 FOR_EACH_BB_FN (bb, cfun)
6369 start_bb = bb;
6370 if (lra_dump_file != NULL)
6371 fprintf (lra_dump_file, "EBB");
6372 /* Form a EBB starting with BB. */
6373 bitmap_clear (&ebb_global_regs);
6374 bitmap_ior_into (&ebb_global_regs, df_get_live_in (bb));
6375 for (;;)
6377 if (lra_dump_file != NULL)
6378 fprintf (lra_dump_file, " %d", bb->index);
6379 if (bb->next_bb == EXIT_BLOCK_PTR_FOR_FN (cfun)
6380 || LABEL_P (BB_HEAD (bb->next_bb)))
6381 break;
6382 e = find_fallthru_edge (bb->succs);
6383 if (! e)
6384 break;
6385 if (e->probability < EBB_PROBABILITY_CUTOFF)
6386 break;
6387 bb = bb->next_bb;
6389 bitmap_ior_into (&ebb_global_regs, df_get_live_out (bb));
6390 if (lra_dump_file != NULL)
6391 fprintf (lra_dump_file, "\n");
6392 if (inherit_in_ebb (BB_HEAD (start_bb), BB_END (bb)))
6393 /* Remember that the EBB head and tail can change in
6394 inherit_in_ebb. */
6395 update_ebb_live_info (BB_HEAD (start_bb), BB_END (bb));
6397 bitmap_clear (&ebb_global_regs);
6398 bitmap_clear (&temp_bitmap);
6399 bitmap_clear (&live_regs);
6400 bitmap_clear (&invalid_invariant_regs);
6401 bitmap_clear (&check_only_regs);
6402 free (usage_insns);
6404 timevar_pop (TV_LRA_INHERITANCE);
6409 /* This page contains code to undo failed inheritance/split
6410 transformations. */
6412 /* Current number of iteration undoing inheritance/split. */
6413 int lra_undo_inheritance_iter;
6415 /* Fix BB live info LIVE after removing pseudos created on pass doing
6416 inheritance/split which are REMOVED_PSEUDOS. */
6417 static void
6418 fix_bb_live_info (bitmap live, bitmap removed_pseudos)
6420 unsigned int regno;
6421 bitmap_iterator bi;
6423 EXECUTE_IF_SET_IN_BITMAP (removed_pseudos, 0, regno, bi)
6424 if (bitmap_clear_bit (live, regno)
6425 && REG_P (lra_reg_info[regno].restore_rtx))
6426 bitmap_set_bit (live, REGNO (lra_reg_info[regno].restore_rtx));
6429 /* Return regno of the (subreg of) REG. Otherwise, return a negative
6430 number. */
6431 static int
6432 get_regno (rtx reg)
6434 if (GET_CODE (reg) == SUBREG)
6435 reg = SUBREG_REG (reg);
6436 if (REG_P (reg))
6437 return REGNO (reg);
6438 return -1;
6441 /* Delete a move INSN with destination reg DREGNO and a previous
6442 clobber insn with the same regno. The inheritance/split code can
6443 generate moves with preceding clobber and when we delete such moves
6444 we should delete the clobber insn too to keep the correct life
6445 info. */
6446 static void
6447 delete_move_and_clobber (rtx_insn *insn, int dregno)
6449 rtx_insn *prev_insn = PREV_INSN (insn);
6451 lra_set_insn_deleted (insn);
6452 lra_assert (dregno >= 0);
6453 if (prev_insn != NULL && NONDEBUG_INSN_P (prev_insn)
6454 && GET_CODE (PATTERN (prev_insn)) == CLOBBER
6455 && dregno == get_regno (XEXP (PATTERN (prev_insn), 0)))
6456 lra_set_insn_deleted (prev_insn);
6459 /* Remove inheritance/split pseudos which are in REMOVE_PSEUDOS and
6460 return true if we did any change. The undo transformations for
6461 inheritance looks like
6462 i <- i2
6463 p <- i => p <- i2
6464 or removing
6465 p <- i, i <- p, and i <- i3
6466 where p is original pseudo from which inheritance pseudo i was
6467 created, i and i3 are removed inheritance pseudos, i2 is another
6468 not removed inheritance pseudo. All split pseudos or other
6469 occurrences of removed inheritance pseudos are changed on the
6470 corresponding original pseudos.
6472 The function also schedules insns changed and created during
6473 inheritance/split pass for processing by the subsequent constraint
6474 pass. */
6475 static bool
6476 remove_inheritance_pseudos (bitmap remove_pseudos)
6478 basic_block bb;
6479 int regno, sregno, prev_sregno, dregno;
6480 rtx restore_rtx;
6481 rtx set, prev_set;
6482 rtx_insn *prev_insn;
6483 bool change_p, done_p;
6485 change_p = ! bitmap_empty_p (remove_pseudos);
6486 /* We can not finish the function right away if CHANGE_P is true
6487 because we need to marks insns affected by previous
6488 inheritance/split pass for processing by the subsequent
6489 constraint pass. */
6490 FOR_EACH_BB_FN (bb, cfun)
6492 fix_bb_live_info (df_get_live_in (bb), remove_pseudos);
6493 fix_bb_live_info (df_get_live_out (bb), remove_pseudos);
6494 FOR_BB_INSNS_REVERSE (bb, curr_insn)
6496 if (! INSN_P (curr_insn))
6497 continue;
6498 done_p = false;
6499 sregno = dregno = -1;
6500 if (change_p && NONDEBUG_INSN_P (curr_insn)
6501 && (set = single_set (curr_insn)) != NULL_RTX)
6503 dregno = get_regno (SET_DEST (set));
6504 sregno = get_regno (SET_SRC (set));
6507 if (sregno >= 0 && dregno >= 0)
6509 if (bitmap_bit_p (remove_pseudos, dregno)
6510 && ! REG_P (lra_reg_info[dregno].restore_rtx))
6512 /* invariant inheritance pseudo <- original pseudo */
6513 if (lra_dump_file != NULL)
6515 fprintf (lra_dump_file, " Removing invariant inheritance:\n");
6516 dump_insn_slim (lra_dump_file, curr_insn);
6517 fprintf (lra_dump_file, "\n");
6519 delete_move_and_clobber (curr_insn, dregno);
6520 done_p = true;
6522 else if (bitmap_bit_p (remove_pseudos, sregno)
6523 && ! REG_P (lra_reg_info[sregno].restore_rtx))
6525 /* reload pseudo <- invariant inheritance pseudo */
6526 start_sequence ();
6527 /* We can not just change the source. It might be
6528 an insn different from the move. */
6529 emit_insn (lra_reg_info[sregno].restore_rtx);
6530 rtx_insn *new_insns = get_insns ();
6531 end_sequence ();
6532 lra_assert (single_set (new_insns) != NULL
6533 && SET_DEST (set) == SET_DEST (single_set (new_insns)));
6534 lra_process_new_insns (curr_insn, NULL, new_insns,
6535 "Changing reload<-invariant inheritance");
6536 delete_move_and_clobber (curr_insn, dregno);
6537 done_p = true;
6539 else if ((bitmap_bit_p (remove_pseudos, sregno)
6540 && (get_regno (lra_reg_info[sregno].restore_rtx) == dregno
6541 || (bitmap_bit_p (remove_pseudos, dregno)
6542 && get_regno (lra_reg_info[sregno].restore_rtx) >= 0
6543 && (get_regno (lra_reg_info[sregno].restore_rtx)
6544 == get_regno (lra_reg_info[dregno].restore_rtx)))))
6545 || (bitmap_bit_p (remove_pseudos, dregno)
6546 && get_regno (lra_reg_info[dregno].restore_rtx) == sregno))
6547 /* One of the following cases:
6548 original <- removed inheritance pseudo
6549 removed inherit pseudo <- another removed inherit pseudo
6550 removed inherit pseudo <- original pseudo
6552 removed_split_pseudo <- original_reg
6553 original_reg <- removed_split_pseudo */
6555 if (lra_dump_file != NULL)
6557 fprintf (lra_dump_file, " Removing %s:\n",
6558 bitmap_bit_p (&lra_split_regs, sregno)
6559 || bitmap_bit_p (&lra_split_regs, dregno)
6560 ? "split" : "inheritance");
6561 dump_insn_slim (lra_dump_file, curr_insn);
6563 delete_move_and_clobber (curr_insn, dregno);
6564 done_p = true;
6566 else if (bitmap_bit_p (remove_pseudos, sregno)
6567 && bitmap_bit_p (&lra_inheritance_pseudos, sregno))
6569 /* Search the following pattern:
6570 inherit_or_split_pseudo1 <- inherit_or_split_pseudo2
6571 original_pseudo <- inherit_or_split_pseudo1
6572 where the 2nd insn is the current insn and
6573 inherit_or_split_pseudo2 is not removed. If it is found,
6574 change the current insn onto:
6575 original_pseudo <- inherit_or_split_pseudo2. */
6576 for (prev_insn = PREV_INSN (curr_insn);
6577 prev_insn != NULL_RTX && ! NONDEBUG_INSN_P (prev_insn);
6578 prev_insn = PREV_INSN (prev_insn))
6580 if (prev_insn != NULL_RTX && BLOCK_FOR_INSN (prev_insn) == bb
6581 && (prev_set = single_set (prev_insn)) != NULL_RTX
6582 /* There should be no subregs in insn we are
6583 searching because only the original reg might
6584 be in subreg when we changed the mode of
6585 load/store for splitting. */
6586 && REG_P (SET_DEST (prev_set))
6587 && REG_P (SET_SRC (prev_set))
6588 && (int) REGNO (SET_DEST (prev_set)) == sregno
6589 && ((prev_sregno = REGNO (SET_SRC (prev_set)))
6590 >= FIRST_PSEUDO_REGISTER)
6591 && (lra_reg_info[prev_sregno].restore_rtx == NULL_RTX
6593 /* As we consider chain of inheritance or
6594 splitting described in above comment we should
6595 check that sregno and prev_sregno were
6596 inheritance/split pseudos created from the
6597 same original regno. */
6598 (get_regno (lra_reg_info[sregno].restore_rtx) >= 0
6599 && (get_regno (lra_reg_info[sregno].restore_rtx)
6600 == get_regno (lra_reg_info[prev_sregno].restore_rtx))))
6601 && ! bitmap_bit_p (remove_pseudos, prev_sregno))
6603 lra_assert (GET_MODE (SET_SRC (prev_set))
6604 == GET_MODE (regno_reg_rtx[sregno]));
6605 if (GET_CODE (SET_SRC (set)) == SUBREG)
6606 SUBREG_REG (SET_SRC (set)) = SET_SRC (prev_set);
6607 else
6608 SET_SRC (set) = SET_SRC (prev_set);
6609 /* As we are finishing with processing the insn
6610 here, check the destination too as it might
6611 inheritance pseudo for another pseudo. */
6612 if (bitmap_bit_p (remove_pseudos, dregno)
6613 && bitmap_bit_p (&lra_inheritance_pseudos, dregno)
6614 && (restore_rtx
6615 = lra_reg_info[dregno].restore_rtx) != NULL_RTX)
6617 if (GET_CODE (SET_DEST (set)) == SUBREG)
6618 SUBREG_REG (SET_DEST (set)) = restore_rtx;
6619 else
6620 SET_DEST (set) = restore_rtx;
6622 lra_push_insn_and_update_insn_regno_info (curr_insn);
6623 lra_set_used_insn_alternative_by_uid
6624 (INSN_UID (curr_insn), -1);
6625 done_p = true;
6626 if (lra_dump_file != NULL)
6628 fprintf (lra_dump_file, " Change reload insn:\n");
6629 dump_insn_slim (lra_dump_file, curr_insn);
6634 if (! done_p)
6636 struct lra_insn_reg *reg;
6637 bool restored_regs_p = false;
6638 bool kept_regs_p = false;
6640 curr_id = lra_get_insn_recog_data (curr_insn);
6641 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
6643 regno = reg->regno;
6644 restore_rtx = lra_reg_info[regno].restore_rtx;
6645 if (restore_rtx != NULL_RTX)
6647 if (change_p && bitmap_bit_p (remove_pseudos, regno))
6649 lra_substitute_pseudo_within_insn
6650 (curr_insn, regno, restore_rtx, false);
6651 restored_regs_p = true;
6653 else
6654 kept_regs_p = true;
6657 if (NONDEBUG_INSN_P (curr_insn) && kept_regs_p)
6659 /* The instruction has changed since the previous
6660 constraints pass. */
6661 lra_push_insn_and_update_insn_regno_info (curr_insn);
6662 lra_set_used_insn_alternative_by_uid
6663 (INSN_UID (curr_insn), -1);
6665 else if (restored_regs_p)
6666 /* The instruction has been restored to the form that
6667 it had during the previous constraints pass. */
6668 lra_update_insn_regno_info (curr_insn);
6669 if (restored_regs_p && lra_dump_file != NULL)
6671 fprintf (lra_dump_file, " Insn after restoring regs:\n");
6672 dump_insn_slim (lra_dump_file, curr_insn);
6677 return change_p;
6680 /* If optional reload pseudos failed to get a hard register or was not
6681 inherited, it is better to remove optional reloads. We do this
6682 transformation after undoing inheritance to figure out necessity to
6683 remove optional reloads easier. Return true if we do any
6684 change. */
6685 static bool
6686 undo_optional_reloads (void)
6688 bool change_p, keep_p;
6689 unsigned int regno, uid;
6690 bitmap_iterator bi, bi2;
6691 rtx_insn *insn;
6692 rtx set, src, dest;
6693 bitmap_head removed_optional_reload_pseudos, insn_bitmap;
6695 bitmap_initialize (&removed_optional_reload_pseudos, &reg_obstack);
6696 bitmap_copy (&removed_optional_reload_pseudos, &lra_optional_reload_pseudos);
6697 EXECUTE_IF_SET_IN_BITMAP (&lra_optional_reload_pseudos, 0, regno, bi)
6699 keep_p = false;
6700 /* Keep optional reloads from previous subpasses. */
6701 if (lra_reg_info[regno].restore_rtx == NULL_RTX
6702 /* If the original pseudo changed its allocation, just
6703 removing the optional pseudo is dangerous as the original
6704 pseudo will have longer live range. */
6705 || reg_renumber[REGNO (lra_reg_info[regno].restore_rtx)] >= 0)
6706 keep_p = true;
6707 else if (reg_renumber[regno] >= 0)
6708 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi2)
6710 insn = lra_insn_recog_data[uid]->insn;
6711 if ((set = single_set (insn)) == NULL_RTX)
6712 continue;
6713 src = SET_SRC (set);
6714 dest = SET_DEST (set);
6715 if (! REG_P (src) || ! REG_P (dest))
6716 continue;
6717 if (REGNO (dest) == regno
6718 /* Ignore insn for optional reloads itself. */
6719 && REGNO (lra_reg_info[regno].restore_rtx) != REGNO (src)
6720 /* Check only inheritance on last inheritance pass. */
6721 && (int) REGNO (src) >= new_regno_start
6722 /* Check that the optional reload was inherited. */
6723 && bitmap_bit_p (&lra_inheritance_pseudos, REGNO (src)))
6725 keep_p = true;
6726 break;
6729 if (keep_p)
6731 bitmap_clear_bit (&removed_optional_reload_pseudos, regno);
6732 if (lra_dump_file != NULL)
6733 fprintf (lra_dump_file, "Keep optional reload reg %d\n", regno);
6736 change_p = ! bitmap_empty_p (&removed_optional_reload_pseudos);
6737 bitmap_initialize (&insn_bitmap, &reg_obstack);
6738 EXECUTE_IF_SET_IN_BITMAP (&removed_optional_reload_pseudos, 0, regno, bi)
6740 if (lra_dump_file != NULL)
6741 fprintf (lra_dump_file, "Remove optional reload reg %d\n", regno);
6742 bitmap_copy (&insn_bitmap, &lra_reg_info[regno].insn_bitmap);
6743 EXECUTE_IF_SET_IN_BITMAP (&insn_bitmap, 0, uid, bi2)
6745 insn = lra_insn_recog_data[uid]->insn;
6746 if ((set = single_set (insn)) != NULL_RTX)
6748 src = SET_SRC (set);
6749 dest = SET_DEST (set);
6750 if (REG_P (src) && REG_P (dest)
6751 && ((REGNO (src) == regno
6752 && (REGNO (lra_reg_info[regno].restore_rtx)
6753 == REGNO (dest)))
6754 || (REGNO (dest) == regno
6755 && (REGNO (lra_reg_info[regno].restore_rtx)
6756 == REGNO (src)))))
6758 if (lra_dump_file != NULL)
6760 fprintf (lra_dump_file, " Deleting move %u\n",
6761 INSN_UID (insn));
6762 dump_insn_slim (lra_dump_file, insn);
6764 delete_move_and_clobber (insn, REGNO (dest));
6765 continue;
6767 /* We should not worry about generation memory-memory
6768 moves here as if the corresponding inheritance did
6769 not work (inheritance pseudo did not get a hard reg),
6770 we remove the inheritance pseudo and the optional
6771 reload. */
6773 lra_substitute_pseudo_within_insn
6774 (insn, regno, lra_reg_info[regno].restore_rtx, false);
6775 lra_update_insn_regno_info (insn);
6776 if (lra_dump_file != NULL)
6778 fprintf (lra_dump_file,
6779 " Restoring original insn:\n");
6780 dump_insn_slim (lra_dump_file, insn);
6784 /* Clear restore_regnos. */
6785 EXECUTE_IF_SET_IN_BITMAP (&lra_optional_reload_pseudos, 0, regno, bi)
6786 lra_reg_info[regno].restore_rtx = NULL_RTX;
6787 bitmap_clear (&insn_bitmap);
6788 bitmap_clear (&removed_optional_reload_pseudos);
6789 return change_p;
6792 /* Entry function for undoing inheritance/split transformation. Return true
6793 if we did any RTL change in this pass. */
6794 bool
6795 lra_undo_inheritance (void)
6797 unsigned int regno;
6798 int hard_regno;
6799 int n_all_inherit, n_inherit, n_all_split, n_split;
6800 rtx restore_rtx;
6801 bitmap_head remove_pseudos;
6802 bitmap_iterator bi;
6803 bool change_p;
6805 lra_undo_inheritance_iter++;
6806 if (lra_undo_inheritance_iter > LRA_MAX_INHERITANCE_PASSES)
6807 return false;
6808 if (lra_dump_file != NULL)
6809 fprintf (lra_dump_file,
6810 "\n********** Undoing inheritance #%d: **********\n\n",
6811 lra_undo_inheritance_iter);
6812 bitmap_initialize (&remove_pseudos, &reg_obstack);
6813 n_inherit = n_all_inherit = 0;
6814 EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, regno, bi)
6815 if (lra_reg_info[regno].restore_rtx != NULL_RTX)
6817 n_all_inherit++;
6818 if (reg_renumber[regno] < 0
6819 /* If the original pseudo changed its allocation, just
6820 removing inheritance is dangerous as for changing
6821 allocation we used shorter live-ranges. */
6822 && (! REG_P (lra_reg_info[regno].restore_rtx)
6823 || reg_renumber[REGNO (lra_reg_info[regno].restore_rtx)] < 0))
6824 bitmap_set_bit (&remove_pseudos, regno);
6825 else
6826 n_inherit++;
6828 if (lra_dump_file != NULL && n_all_inherit != 0)
6829 fprintf (lra_dump_file, "Inherit %d out of %d (%.2f%%)\n",
6830 n_inherit, n_all_inherit,
6831 (double) n_inherit / n_all_inherit * 100);
6832 n_split = n_all_split = 0;
6833 EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, regno, bi)
6834 if ((restore_rtx = lra_reg_info[regno].restore_rtx) != NULL_RTX)
6836 int restore_regno = REGNO (restore_rtx);
6838 n_all_split++;
6839 hard_regno = (restore_regno >= FIRST_PSEUDO_REGISTER
6840 ? reg_renumber[restore_regno] : restore_regno);
6841 if (hard_regno < 0 || reg_renumber[regno] == hard_regno)
6842 bitmap_set_bit (&remove_pseudos, regno);
6843 else
6845 n_split++;
6846 if (lra_dump_file != NULL)
6847 fprintf (lra_dump_file, " Keep split r%d (orig=r%d)\n",
6848 regno, restore_regno);
6851 if (lra_dump_file != NULL && n_all_split != 0)
6852 fprintf (lra_dump_file, "Split %d out of %d (%.2f%%)\n",
6853 n_split, n_all_split,
6854 (double) n_split / n_all_split * 100);
6855 change_p = remove_inheritance_pseudos (&remove_pseudos);
6856 bitmap_clear (&remove_pseudos);
6857 /* Clear restore_regnos. */
6858 EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, regno, bi)
6859 lra_reg_info[regno].restore_rtx = NULL_RTX;
6860 EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, regno, bi)
6861 lra_reg_info[regno].restore_rtx = NULL_RTX;
6862 change_p = undo_optional_reloads () || change_p;
6863 return change_p;