re PR target/82015 (PowerPC should check if 2nd argument to __builtin_unpackv1ti...
[official-gcc.git] / gcc / lra-constraints.c
blob6da910e33a8720c023f6827b0f1ba6f7f30e8f24
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);
675 scalar_int_mode int_mode;
676 if (WORDS_BIG_ENDIAN
677 && is_a <scalar_int_mode> (mode, &int_mode)
678 && GET_MODE_SIZE (int_mode) > UNITS_PER_WORD)
679 return hard_regno_nregs[regno][mode] - 1;
680 return 0;
683 /* Like rtx_equal_p except that it allows a REG and a SUBREG to match
684 if they are the same hard reg, and has special hacks for
685 auto-increment and auto-decrement. This is specifically intended for
686 process_alt_operands to use in determining whether two operands
687 match. X is the operand whose number is the lower of the two.
689 It is supposed that X is the output operand and Y is the input
690 operand. Y_HARD_REGNO is the final hard regno of register Y or
691 register in subreg Y as we know it now. Otherwise, it is a
692 negative value. */
693 static bool
694 operands_match_p (rtx x, rtx y, int y_hard_regno)
696 int i;
697 RTX_CODE code = GET_CODE (x);
698 const char *fmt;
700 if (x == y)
701 return true;
702 if ((code == REG || (code == SUBREG && REG_P (SUBREG_REG (x))))
703 && (REG_P (y) || (GET_CODE (y) == SUBREG && REG_P (SUBREG_REG (y)))))
705 int j;
707 i = get_hard_regno (x, false);
708 if (i < 0)
709 goto slow;
711 if ((j = y_hard_regno) < 0)
712 goto slow;
714 i += lra_constraint_offset (i, GET_MODE (x));
715 j += lra_constraint_offset (j, GET_MODE (y));
717 return i == j;
720 /* If two operands must match, because they are really a single
721 operand of an assembler insn, then two post-increments are invalid
722 because the assembler insn would increment only once. On the
723 other hand, a post-increment matches ordinary indexing if the
724 post-increment is the output operand. */
725 if (code == POST_DEC || code == POST_INC || code == POST_MODIFY)
726 return operands_match_p (XEXP (x, 0), y, y_hard_regno);
728 /* Two pre-increments are invalid because the assembler insn would
729 increment only once. On the other hand, a pre-increment matches
730 ordinary indexing if the pre-increment is the input operand. */
731 if (GET_CODE (y) == PRE_DEC || GET_CODE (y) == PRE_INC
732 || GET_CODE (y) == PRE_MODIFY)
733 return operands_match_p (x, XEXP (y, 0), -1);
735 slow:
737 if (code == REG && REG_P (y))
738 return REGNO (x) == REGNO (y);
740 if (code == REG && GET_CODE (y) == SUBREG && REG_P (SUBREG_REG (y))
741 && x == SUBREG_REG (y))
742 return true;
743 if (GET_CODE (y) == REG && code == SUBREG && REG_P (SUBREG_REG (x))
744 && SUBREG_REG (x) == y)
745 return true;
747 /* Now we have disposed of all the cases in which different rtx
748 codes can match. */
749 if (code != GET_CODE (y))
750 return false;
752 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
753 if (GET_MODE (x) != GET_MODE (y))
754 return false;
756 switch (code)
758 CASE_CONST_UNIQUE:
759 return false;
761 case LABEL_REF:
762 return label_ref_label (x) == label_ref_label (y);
763 case SYMBOL_REF:
764 return XSTR (x, 0) == XSTR (y, 0);
766 default:
767 break;
770 /* Compare the elements. If any pair of corresponding elements fail
771 to match, return false for the whole things. */
773 fmt = GET_RTX_FORMAT (code);
774 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
776 int val, j;
777 switch (fmt[i])
779 case 'w':
780 if (XWINT (x, i) != XWINT (y, i))
781 return false;
782 break;
784 case 'i':
785 if (XINT (x, i) != XINT (y, i))
786 return false;
787 break;
789 case 'e':
790 val = operands_match_p (XEXP (x, i), XEXP (y, i), -1);
791 if (val == 0)
792 return false;
793 break;
795 case '0':
796 break;
798 case 'E':
799 if (XVECLEN (x, i) != XVECLEN (y, i))
800 return false;
801 for (j = XVECLEN (x, i) - 1; j >= 0; --j)
803 val = operands_match_p (XVECEXP (x, i, j), XVECEXP (y, i, j), -1);
804 if (val == 0)
805 return false;
807 break;
809 /* It is believed that rtx's at this level will never
810 contain anything but integers and other rtx's, except for
811 within LABEL_REFs and SYMBOL_REFs. */
812 default:
813 gcc_unreachable ();
816 return true;
819 /* True if X is a constant that can be forced into the constant pool.
820 MODE is the mode of the operand, or VOIDmode if not known. */
821 #define CONST_POOL_OK_P(MODE, X) \
822 ((MODE) != VOIDmode \
823 && CONSTANT_P (X) \
824 && GET_CODE (X) != HIGH \
825 && !targetm.cannot_force_const_mem (MODE, X))
827 /* True if C is a non-empty register class that has too few registers
828 to be safely used as a reload target class. */
829 #define SMALL_REGISTER_CLASS_P(C) \
830 (ira_class_hard_regs_num [(C)] == 1 \
831 || (ira_class_hard_regs_num [(C)] >= 1 \
832 && targetm.class_likely_spilled_p (C)))
834 /* If REG is a reload pseudo, try to make its class satisfying CL. */
835 static void
836 narrow_reload_pseudo_class (rtx reg, enum reg_class cl)
838 enum reg_class rclass;
840 /* Do not make more accurate class from reloads generated. They are
841 mostly moves with a lot of constraints. Making more accurate
842 class may results in very narrow class and impossibility of find
843 registers for several reloads of one insn. */
844 if (INSN_UID (curr_insn) >= new_insn_uid_start)
845 return;
846 if (GET_CODE (reg) == SUBREG)
847 reg = SUBREG_REG (reg);
848 if (! REG_P (reg) || (int) REGNO (reg) < new_regno_start)
849 return;
850 if (in_class_p (reg, cl, &rclass) && rclass != cl)
851 lra_change_class (REGNO (reg), rclass, " Change to", true);
854 /* Searches X for any reference to a reg with the same value as REGNO,
855 returning the rtx of the reference found if any. Otherwise,
856 returns NULL_RTX. */
857 static rtx
858 regno_val_use_in (unsigned int regno, rtx x)
860 const char *fmt;
861 int i, j;
862 rtx tem;
864 if (REG_P (x) && lra_reg_info[REGNO (x)].val == lra_reg_info[regno].val)
865 return x;
867 fmt = GET_RTX_FORMAT (GET_CODE (x));
868 for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
870 if (fmt[i] == 'e')
872 if ((tem = regno_val_use_in (regno, XEXP (x, i))))
873 return tem;
875 else if (fmt[i] == 'E')
876 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
877 if ((tem = regno_val_use_in (regno , XVECEXP (x, i, j))))
878 return tem;
881 return NULL_RTX;
884 /* Return true if all current insn non-output operands except INS (it
885 has a negaitve end marker) do not use pseudos with the same value
886 as REGNO. */
887 static bool
888 check_conflict_input_operands (int regno, signed char *ins)
890 int in;
891 int n_operands = curr_static_id->n_operands;
893 for (int nop = 0; nop < n_operands; nop++)
894 if (! curr_static_id->operand[nop].is_operator
895 && curr_static_id->operand[nop].type != OP_OUT)
897 for (int i = 0; (in = ins[i]) >= 0; i++)
898 if (in == nop)
899 break;
900 if (in < 0
901 && regno_val_use_in (regno, *curr_id->operand_loc[nop]) != NULL_RTX)
902 return false;
904 return true;
907 /* Generate reloads for matching OUT and INS (array of input operand
908 numbers with end marker -1) with reg class GOAL_CLASS, considering
909 output operands OUTS (similar array to INS) needing to be in different
910 registers. Add input and output reloads correspondingly to the lists
911 *BEFORE and *AFTER. OUT might be negative. In this case we generate
912 input reloads for matched input operands INS. EARLY_CLOBBER_P is a flag
913 that the output operand is early clobbered for chosen alternative. */
914 static void
915 match_reload (signed char out, signed char *ins, signed char *outs,
916 enum reg_class goal_class, rtx_insn **before,
917 rtx_insn **after, bool early_clobber_p)
919 bool out_conflict;
920 int i, in;
921 rtx new_in_reg, new_out_reg, reg;
922 machine_mode inmode, outmode;
923 rtx in_rtx = *curr_id->operand_loc[ins[0]];
924 rtx out_rtx = out < 0 ? in_rtx : *curr_id->operand_loc[out];
926 inmode = curr_operand_mode[ins[0]];
927 outmode = out < 0 ? inmode : curr_operand_mode[out];
928 push_to_sequence (*before);
929 if (inmode != outmode)
931 if (partial_subreg_p (outmode, inmode))
933 reg = new_in_reg
934 = lra_create_new_reg_with_unique_value (inmode, in_rtx,
935 goal_class, "");
936 if (SCALAR_INT_MODE_P (inmode))
937 new_out_reg = gen_lowpart_SUBREG (outmode, reg);
938 else
939 new_out_reg = gen_rtx_SUBREG (outmode, reg, 0);
940 LRA_SUBREG_P (new_out_reg) = 1;
941 /* If the input reg is dying here, we can use the same hard
942 register for REG and IN_RTX. We do it only for original
943 pseudos as reload pseudos can die although original
944 pseudos still live where reload pseudos dies. */
945 if (REG_P (in_rtx) && (int) REGNO (in_rtx) < lra_new_regno_start
946 && find_regno_note (curr_insn, REG_DEAD, REGNO (in_rtx))
947 && (!early_clobber_p
948 || check_conflict_input_operands(REGNO (in_rtx), ins)))
949 lra_assign_reg_val (REGNO (in_rtx), REGNO (reg));
951 else
953 reg = new_out_reg
954 = lra_create_new_reg_with_unique_value (outmode, out_rtx,
955 goal_class, "");
956 if (SCALAR_INT_MODE_P (outmode))
957 new_in_reg = gen_lowpart_SUBREG (inmode, reg);
958 else
959 new_in_reg = gen_rtx_SUBREG (inmode, reg, 0);
960 /* NEW_IN_REG is non-paradoxical subreg. We don't want
961 NEW_OUT_REG living above. We add clobber clause for
962 this. This is just a temporary clobber. We can remove
963 it at the end of LRA work. */
964 rtx_insn *clobber = emit_clobber (new_out_reg);
965 LRA_TEMP_CLOBBER_P (PATTERN (clobber)) = 1;
966 LRA_SUBREG_P (new_in_reg) = 1;
967 if (GET_CODE (in_rtx) == SUBREG)
969 rtx subreg_reg = SUBREG_REG (in_rtx);
971 /* If SUBREG_REG is dying here and sub-registers IN_RTX
972 and NEW_IN_REG are similar, we can use the same hard
973 register for REG and SUBREG_REG. */
974 if (REG_P (subreg_reg)
975 && (int) REGNO (subreg_reg) < lra_new_regno_start
976 && GET_MODE (subreg_reg) == outmode
977 && SUBREG_BYTE (in_rtx) == SUBREG_BYTE (new_in_reg)
978 && find_regno_note (curr_insn, REG_DEAD, REGNO (subreg_reg))
979 && (! early_clobber_p
980 || check_conflict_input_operands (REGNO (subreg_reg),
981 ins)))
982 lra_assign_reg_val (REGNO (subreg_reg), REGNO (reg));
986 else
988 /* Pseudos have values -- see comments for lra_reg_info.
989 Different pseudos with the same value do not conflict even if
990 they live in the same place. When we create a pseudo we
991 assign value of original pseudo (if any) from which we
992 created the new pseudo. If we create the pseudo from the
993 input pseudo, the new pseudo will have no conflict with the
994 input pseudo which is wrong when the input pseudo lives after
995 the insn and as the new pseudo value is changed by the insn
996 output. Therefore we create the new pseudo from the output
997 except the case when we have single matched dying input
998 pseudo.
1000 We cannot reuse the current output register because we might
1001 have a situation like "a <- a op b", where the constraints
1002 force the second input operand ("b") to match the output
1003 operand ("a"). "b" must then be copied into a new register
1004 so that it doesn't clobber the current value of "a".
1006 We can not use the same value if the output pseudo is
1007 early clobbered or the input pseudo is mentioned in the
1008 output, e.g. as an address part in memory, because
1009 output reload will actually extend the pseudo liveness.
1010 We don't care about eliminable hard regs here as we are
1011 interesting only in pseudos. */
1013 /* Matching input's register value is the same as one of the other
1014 output operand. Output operands in a parallel insn must be in
1015 different registers. */
1016 out_conflict = false;
1017 if (REG_P (in_rtx))
1019 for (i = 0; outs[i] >= 0; i++)
1021 rtx other_out_rtx = *curr_id->operand_loc[outs[i]];
1022 if (REG_P (other_out_rtx)
1023 && (regno_val_use_in (REGNO (in_rtx), other_out_rtx)
1024 != NULL_RTX))
1026 out_conflict = true;
1027 break;
1032 new_in_reg = new_out_reg
1033 = (! early_clobber_p && ins[1] < 0 && REG_P (in_rtx)
1034 && (int) REGNO (in_rtx) < lra_new_regno_start
1035 && find_regno_note (curr_insn, REG_DEAD, REGNO (in_rtx))
1036 && (! early_clobber_p
1037 || check_conflict_input_operands (REGNO (in_rtx), ins))
1038 && (out < 0
1039 || regno_val_use_in (REGNO (in_rtx), out_rtx) == NULL_RTX)
1040 && !out_conflict
1041 ? lra_create_new_reg (inmode, in_rtx, goal_class, "")
1042 : lra_create_new_reg_with_unique_value (outmode, out_rtx,
1043 goal_class, ""));
1045 /* In operand can be got from transformations before processing insn
1046 constraints. One example of such transformations is subreg
1047 reloading (see function simplify_operand_subreg). The new
1048 pseudos created by the transformations might have inaccurate
1049 class (ALL_REGS) and we should make their classes more
1050 accurate. */
1051 narrow_reload_pseudo_class (in_rtx, goal_class);
1052 lra_emit_move (copy_rtx (new_in_reg), in_rtx);
1053 *before = get_insns ();
1054 end_sequence ();
1055 /* Add the new pseudo to consider values of subsequent input reload
1056 pseudos. */
1057 lra_assert (curr_insn_input_reloads_num < LRA_MAX_INSN_RELOADS);
1058 curr_insn_input_reloads[curr_insn_input_reloads_num].input = in_rtx;
1059 curr_insn_input_reloads[curr_insn_input_reloads_num].match_p = true;
1060 curr_insn_input_reloads[curr_insn_input_reloads_num++].reg = new_in_reg;
1061 for (i = 0; (in = ins[i]) >= 0; i++)
1063 lra_assert
1064 (GET_MODE (*curr_id->operand_loc[in]) == VOIDmode
1065 || GET_MODE (new_in_reg) == GET_MODE (*curr_id->operand_loc[in]));
1066 *curr_id->operand_loc[in] = new_in_reg;
1068 lra_update_dups (curr_id, ins);
1069 if (out < 0)
1070 return;
1071 /* See a comment for the input operand above. */
1072 narrow_reload_pseudo_class (out_rtx, goal_class);
1073 if (find_reg_note (curr_insn, REG_UNUSED, out_rtx) == NULL_RTX)
1075 start_sequence ();
1076 lra_emit_move (out_rtx, copy_rtx (new_out_reg));
1077 emit_insn (*after);
1078 *after = get_insns ();
1079 end_sequence ();
1081 *curr_id->operand_loc[out] = new_out_reg;
1082 lra_update_dup (curr_id, out);
1085 /* Return register class which is union of all reg classes in insn
1086 constraint alternative string starting with P. */
1087 static enum reg_class
1088 reg_class_from_constraints (const char *p)
1090 int c, len;
1091 enum reg_class op_class = NO_REGS;
1094 switch ((c = *p, len = CONSTRAINT_LEN (c, p)), c)
1096 case '#':
1097 case ',':
1098 return op_class;
1100 case 'g':
1101 op_class = reg_class_subunion[op_class][GENERAL_REGS];
1102 break;
1104 default:
1105 enum constraint_num cn = lookup_constraint (p);
1106 enum reg_class cl = reg_class_for_constraint (cn);
1107 if (cl == NO_REGS)
1109 if (insn_extra_address_constraint (cn))
1110 op_class
1111 = (reg_class_subunion
1112 [op_class][base_reg_class (VOIDmode, ADDR_SPACE_GENERIC,
1113 ADDRESS, SCRATCH)]);
1114 break;
1117 op_class = reg_class_subunion[op_class][cl];
1118 break;
1120 while ((p += len), c);
1121 return op_class;
1124 /* If OP is a register, return the class of the register as per
1125 get_reg_class, otherwise return NO_REGS. */
1126 static inline enum reg_class
1127 get_op_class (rtx op)
1129 return REG_P (op) ? get_reg_class (REGNO (op)) : NO_REGS;
1132 /* Return generated insn mem_pseudo:=val if TO_P or val:=mem_pseudo
1133 otherwise. If modes of MEM_PSEUDO and VAL are different, use
1134 SUBREG for VAL to make them equal. */
1135 static rtx_insn *
1136 emit_spill_move (bool to_p, rtx mem_pseudo, rtx val)
1138 if (GET_MODE (mem_pseudo) != GET_MODE (val))
1140 /* Usually size of mem_pseudo is greater than val size but in
1141 rare cases it can be less as it can be defined by target
1142 dependent macro HARD_REGNO_CALLER_SAVE_MODE. */
1143 if (! MEM_P (val))
1145 val = gen_lowpart_SUBREG (GET_MODE (mem_pseudo),
1146 GET_CODE (val) == SUBREG
1147 ? SUBREG_REG (val) : val);
1148 LRA_SUBREG_P (val) = 1;
1150 else
1152 mem_pseudo = gen_lowpart_SUBREG (GET_MODE (val), mem_pseudo);
1153 LRA_SUBREG_P (mem_pseudo) = 1;
1156 return to_p ? gen_move_insn (mem_pseudo, val)
1157 : gen_move_insn (val, mem_pseudo);
1160 /* Process a special case insn (register move), return true if we
1161 don't need to process it anymore. INSN should be a single set
1162 insn. Set up that RTL was changed through CHANGE_P and macro
1163 SECONDARY_MEMORY_NEEDED says to use secondary memory through
1164 SEC_MEM_P. */
1165 static bool
1166 check_and_process_move (bool *change_p, bool *sec_mem_p ATTRIBUTE_UNUSED)
1168 int sregno, dregno;
1169 rtx dest, src, dreg, sreg, new_reg, scratch_reg;
1170 rtx_insn *before;
1171 enum reg_class dclass, sclass, secondary_class;
1172 secondary_reload_info sri;
1174 lra_assert (curr_insn_set != NULL_RTX);
1175 dreg = dest = SET_DEST (curr_insn_set);
1176 sreg = src = SET_SRC (curr_insn_set);
1177 if (GET_CODE (dest) == SUBREG)
1178 dreg = SUBREG_REG (dest);
1179 if (GET_CODE (src) == SUBREG)
1180 sreg = SUBREG_REG (src);
1181 if (! (REG_P (dreg) || MEM_P (dreg)) || ! (REG_P (sreg) || MEM_P (sreg)))
1182 return false;
1183 sclass = dclass = NO_REGS;
1184 if (REG_P (dreg))
1185 dclass = get_reg_class (REGNO (dreg));
1186 gcc_assert (dclass < LIM_REG_CLASSES);
1187 if (dclass == ALL_REGS)
1188 /* ALL_REGS is used for new pseudos created by transformations
1189 like reload of SUBREG_REG (see function
1190 simplify_operand_subreg). We don't know their class yet. We
1191 should figure out the class from processing the insn
1192 constraints not in this fast path function. Even if ALL_REGS
1193 were a right class for the pseudo, secondary_... hooks usually
1194 are not define for ALL_REGS. */
1195 return false;
1196 if (REG_P (sreg))
1197 sclass = get_reg_class (REGNO (sreg));
1198 gcc_assert (sclass < LIM_REG_CLASSES);
1199 if (sclass == ALL_REGS)
1200 /* See comments above. */
1201 return false;
1202 if (sclass == NO_REGS && dclass == NO_REGS)
1203 return false;
1204 #ifdef SECONDARY_MEMORY_NEEDED
1205 if (SECONDARY_MEMORY_NEEDED (sclass, dclass, GET_MODE (src))
1206 #ifdef SECONDARY_MEMORY_NEEDED_MODE
1207 && ((sclass != NO_REGS && dclass != NO_REGS)
1208 || GET_MODE (src) != SECONDARY_MEMORY_NEEDED_MODE (GET_MODE (src)))
1209 #endif
1212 *sec_mem_p = true;
1213 return false;
1215 #endif
1216 if (! REG_P (dreg) || ! REG_P (sreg))
1217 return false;
1218 sri.prev_sri = NULL;
1219 sri.icode = CODE_FOR_nothing;
1220 sri.extra_cost = 0;
1221 secondary_class = NO_REGS;
1222 /* Set up hard register for a reload pseudo for hook
1223 secondary_reload because some targets just ignore unassigned
1224 pseudos in the hook. */
1225 if (dclass != NO_REGS && lra_get_regno_hard_regno (REGNO (dreg)) < 0)
1227 dregno = REGNO (dreg);
1228 reg_renumber[dregno] = ira_class_hard_regs[dclass][0];
1230 else
1231 dregno = -1;
1232 if (sclass != NO_REGS && lra_get_regno_hard_regno (REGNO (sreg)) < 0)
1234 sregno = REGNO (sreg);
1235 reg_renumber[sregno] = ira_class_hard_regs[sclass][0];
1237 else
1238 sregno = -1;
1239 if (sclass != NO_REGS)
1240 secondary_class
1241 = (enum reg_class) targetm.secondary_reload (false, dest,
1242 (reg_class_t) sclass,
1243 GET_MODE (src), &sri);
1244 if (sclass == NO_REGS
1245 || ((secondary_class != NO_REGS || sri.icode != CODE_FOR_nothing)
1246 && dclass != NO_REGS))
1248 enum reg_class old_sclass = secondary_class;
1249 secondary_reload_info old_sri = sri;
1251 sri.prev_sri = NULL;
1252 sri.icode = CODE_FOR_nothing;
1253 sri.extra_cost = 0;
1254 secondary_class
1255 = (enum reg_class) targetm.secondary_reload (true, src,
1256 (reg_class_t) dclass,
1257 GET_MODE (src), &sri);
1258 /* Check the target hook consistency. */
1259 lra_assert
1260 ((secondary_class == NO_REGS && sri.icode == CODE_FOR_nothing)
1261 || (old_sclass == NO_REGS && old_sri.icode == CODE_FOR_nothing)
1262 || (secondary_class == old_sclass && sri.icode == old_sri.icode));
1264 if (sregno >= 0)
1265 reg_renumber [sregno] = -1;
1266 if (dregno >= 0)
1267 reg_renumber [dregno] = -1;
1268 if (secondary_class == NO_REGS && sri.icode == CODE_FOR_nothing)
1269 return false;
1270 *change_p = true;
1271 new_reg = NULL_RTX;
1272 if (secondary_class != NO_REGS)
1273 new_reg = lra_create_new_reg_with_unique_value (GET_MODE (src), NULL_RTX,
1274 secondary_class,
1275 "secondary");
1276 start_sequence ();
1277 if (sri.icode == CODE_FOR_nothing)
1278 lra_emit_move (new_reg, src);
1279 else
1281 enum reg_class scratch_class;
1283 scratch_class = (reg_class_from_constraints
1284 (insn_data[sri.icode].operand[2].constraint));
1285 scratch_reg = (lra_create_new_reg_with_unique_value
1286 (insn_data[sri.icode].operand[2].mode, NULL_RTX,
1287 scratch_class, "scratch"));
1288 emit_insn (GEN_FCN (sri.icode) (new_reg != NULL_RTX ? new_reg : dest,
1289 src, scratch_reg));
1291 before = get_insns ();
1292 end_sequence ();
1293 lra_process_new_insns (curr_insn, before, NULL, "Inserting the move");
1294 if (new_reg != NULL_RTX)
1295 SET_SRC (curr_insn_set) = new_reg;
1296 else
1298 if (lra_dump_file != NULL)
1300 fprintf (lra_dump_file, "Deleting move %u\n", INSN_UID (curr_insn));
1301 dump_insn_slim (lra_dump_file, curr_insn);
1303 lra_set_insn_deleted (curr_insn);
1304 return true;
1306 return false;
1309 /* The following data describe the result of process_alt_operands.
1310 The data are used in curr_insn_transform to generate reloads. */
1312 /* The chosen reg classes which should be used for the corresponding
1313 operands. */
1314 static enum reg_class goal_alt[MAX_RECOG_OPERANDS];
1315 /* True if the operand should be the same as another operand and that
1316 other operand does not need a reload. */
1317 static bool goal_alt_match_win[MAX_RECOG_OPERANDS];
1318 /* True if the operand does not need a reload. */
1319 static bool goal_alt_win[MAX_RECOG_OPERANDS];
1320 /* True if the operand can be offsetable memory. */
1321 static bool goal_alt_offmemok[MAX_RECOG_OPERANDS];
1322 /* The number of an operand to which given operand can be matched to. */
1323 static int goal_alt_matches[MAX_RECOG_OPERANDS];
1324 /* The number of elements in the following array. */
1325 static int goal_alt_dont_inherit_ops_num;
1326 /* Numbers of operands whose reload pseudos should not be inherited. */
1327 static int goal_alt_dont_inherit_ops[MAX_RECOG_OPERANDS];
1328 /* True if the insn commutative operands should be swapped. */
1329 static bool goal_alt_swapped;
1330 /* The chosen insn alternative. */
1331 static int goal_alt_number;
1333 /* True if the corresponding operand is the result of an equivalence
1334 substitution. */
1335 static bool equiv_substition_p[MAX_RECOG_OPERANDS];
1337 /* The following five variables are used to choose the best insn
1338 alternative. They reflect final characteristics of the best
1339 alternative. */
1341 /* Number of necessary reloads and overall cost reflecting the
1342 previous value and other unpleasantness of the best alternative. */
1343 static int best_losers, best_overall;
1344 /* Overall number hard registers used for reloads. For example, on
1345 some targets we need 2 general registers to reload DFmode and only
1346 one floating point register. */
1347 static int best_reload_nregs;
1348 /* Overall number reflecting distances of previous reloading the same
1349 value. The distances are counted from the current BB start. It is
1350 used to improve inheritance chances. */
1351 static int best_reload_sum;
1353 /* True if the current insn should have no correspondingly input or
1354 output reloads. */
1355 static bool no_input_reloads_p, no_output_reloads_p;
1357 /* True if we swapped the commutative operands in the current
1358 insn. */
1359 static int curr_swapped;
1361 /* if CHECK_ONLY_P is false, arrange for address element *LOC to be a
1362 register of class CL. Add any input reloads to list BEFORE. AFTER
1363 is nonnull if *LOC is an automodified value; handle that case by
1364 adding the required output reloads to list AFTER. Return true if
1365 the RTL was changed.
1367 if CHECK_ONLY_P is true, check that the *LOC is a correct address
1368 register. Return false if the address register is correct. */
1369 static bool
1370 process_addr_reg (rtx *loc, bool check_only_p, rtx_insn **before, rtx_insn **after,
1371 enum reg_class cl)
1373 int regno;
1374 enum reg_class rclass, new_class;
1375 rtx reg;
1376 rtx new_reg;
1377 machine_mode mode;
1378 bool subreg_p, before_p = false;
1380 subreg_p = GET_CODE (*loc) == SUBREG;
1381 if (subreg_p)
1383 reg = SUBREG_REG (*loc);
1384 mode = GET_MODE (reg);
1386 /* For mode with size bigger than ptr_mode, there unlikely to be "mov"
1387 between two registers with different classes, but there normally will
1388 be "mov" which transfers element of vector register into the general
1389 register, and this normally will be a subreg which should be reloaded
1390 as a whole. This is particularly likely to be triggered when
1391 -fno-split-wide-types specified. */
1392 if (!REG_P (reg)
1393 || in_class_p (reg, cl, &new_class)
1394 || GET_MODE_SIZE (mode) <= GET_MODE_SIZE (ptr_mode))
1395 loc = &SUBREG_REG (*loc);
1398 reg = *loc;
1399 mode = GET_MODE (reg);
1400 if (! REG_P (reg))
1402 if (check_only_p)
1403 return true;
1404 /* Always reload memory in an address even if the target supports
1405 such addresses. */
1406 new_reg = lra_create_new_reg_with_unique_value (mode, reg, cl, "address");
1407 before_p = true;
1409 else
1411 regno = REGNO (reg);
1412 rclass = get_reg_class (regno);
1413 if (! check_only_p
1414 && (*loc = get_equiv_with_elimination (reg, curr_insn)) != reg)
1416 if (lra_dump_file != NULL)
1418 fprintf (lra_dump_file,
1419 "Changing pseudo %d in address of insn %u on equiv ",
1420 REGNO (reg), INSN_UID (curr_insn));
1421 dump_value_slim (lra_dump_file, *loc, 1);
1422 fprintf (lra_dump_file, "\n");
1424 *loc = copy_rtx (*loc);
1426 if (*loc != reg || ! in_class_p (reg, cl, &new_class))
1428 if (check_only_p)
1429 return true;
1430 reg = *loc;
1431 if (get_reload_reg (after == NULL ? OP_IN : OP_INOUT,
1432 mode, reg, cl, subreg_p, "address", &new_reg))
1433 before_p = true;
1435 else if (new_class != NO_REGS && rclass != new_class)
1437 if (check_only_p)
1438 return true;
1439 lra_change_class (regno, new_class, " Change to", true);
1440 return false;
1442 else
1443 return false;
1445 if (before_p)
1447 push_to_sequence (*before);
1448 lra_emit_move (new_reg, reg);
1449 *before = get_insns ();
1450 end_sequence ();
1452 *loc = new_reg;
1453 if (after != NULL)
1455 start_sequence ();
1456 lra_emit_move (before_p ? copy_rtx (reg) : reg, new_reg);
1457 emit_insn (*after);
1458 *after = get_insns ();
1459 end_sequence ();
1461 return true;
1464 /* Insert move insn in simplify_operand_subreg. BEFORE returns
1465 the insn to be inserted before curr insn. AFTER returns the
1466 the insn to be inserted after curr insn. ORIGREG and NEWREG
1467 are the original reg and new reg for reload. */
1468 static void
1469 insert_move_for_subreg (rtx_insn **before, rtx_insn **after, rtx origreg,
1470 rtx newreg)
1472 if (before)
1474 push_to_sequence (*before);
1475 lra_emit_move (newreg, origreg);
1476 *before = get_insns ();
1477 end_sequence ();
1479 if (after)
1481 start_sequence ();
1482 lra_emit_move (origreg, newreg);
1483 emit_insn (*after);
1484 *after = get_insns ();
1485 end_sequence ();
1489 static int valid_address_p (machine_mode mode, rtx addr, addr_space_t as);
1490 static bool process_address (int, bool, rtx_insn **, rtx_insn **);
1492 /* Make reloads for subreg in operand NOP with internal subreg mode
1493 REG_MODE, add new reloads for further processing. Return true if
1494 any change was done. */
1495 static bool
1496 simplify_operand_subreg (int nop, machine_mode reg_mode)
1498 int hard_regno;
1499 rtx_insn *before, *after;
1500 machine_mode mode, innermode;
1501 rtx reg, new_reg;
1502 rtx operand = *curr_id->operand_loc[nop];
1503 enum reg_class regclass;
1504 enum op_type type;
1506 before = after = NULL;
1508 if (GET_CODE (operand) != SUBREG)
1509 return false;
1511 mode = GET_MODE (operand);
1512 reg = SUBREG_REG (operand);
1513 innermode = GET_MODE (reg);
1514 type = curr_static_id->operand[nop].type;
1515 if (MEM_P (reg))
1517 const bool addr_was_valid
1518 = valid_address_p (innermode, XEXP (reg, 0), MEM_ADDR_SPACE (reg));
1519 alter_subreg (curr_id->operand_loc[nop], false);
1520 rtx subst = *curr_id->operand_loc[nop];
1521 lra_assert (MEM_P (subst));
1523 if (!addr_was_valid
1524 || valid_address_p (GET_MODE (subst), XEXP (subst, 0),
1525 MEM_ADDR_SPACE (subst))
1526 || ((get_constraint_type (lookup_constraint
1527 (curr_static_id->operand[nop].constraint))
1528 != CT_SPECIAL_MEMORY)
1529 /* We still can reload address and if the address is
1530 valid, we can remove subreg without reloading its
1531 inner memory. */
1532 && valid_address_p (GET_MODE (subst),
1533 regno_reg_rtx
1534 [ira_class_hard_regs
1535 [base_reg_class (GET_MODE (subst),
1536 MEM_ADDR_SPACE (subst),
1537 ADDRESS, SCRATCH)][0]],
1538 MEM_ADDR_SPACE (subst))))
1540 /* If we change the address for a paradoxical subreg of memory, the
1541 new address might violate the necessary alignment or the access
1542 might be slow; take this into consideration. We need not worry
1543 about accesses beyond allocated memory for paradoxical memory
1544 subregs as we don't substitute such equiv memory (see processing
1545 equivalences in function lra_constraints) and because for spilled
1546 pseudos we allocate stack memory enough for the biggest
1547 corresponding paradoxical subreg.
1549 However, do not blindly simplify a (subreg (mem ...)) for
1550 WORD_REGISTER_OPERATIONS targets as this may lead to loading junk
1551 data into a register when the inner is narrower than outer or
1552 missing important data from memory when the inner is wider than
1553 outer. This rule only applies to modes that are no wider than
1554 a word. */
1555 if (!(GET_MODE_PRECISION (mode) != GET_MODE_PRECISION (innermode)
1556 && GET_MODE_SIZE (mode) <= UNITS_PER_WORD
1557 && GET_MODE_SIZE (innermode) <= UNITS_PER_WORD
1558 && WORD_REGISTER_OPERATIONS)
1559 && (!(MEM_ALIGN (subst) < GET_MODE_ALIGNMENT (mode)
1560 && SLOW_UNALIGNED_ACCESS (mode, MEM_ALIGN (subst)))
1561 || (MEM_ALIGN (reg) < GET_MODE_ALIGNMENT (innermode)
1562 && SLOW_UNALIGNED_ACCESS (innermode, MEM_ALIGN (reg)))))
1563 return true;
1565 *curr_id->operand_loc[nop] = operand;
1567 /* But if the address was not valid, we cannot reload the MEM without
1568 reloading the address first. */
1569 if (!addr_was_valid)
1570 process_address (nop, false, &before, &after);
1572 /* INNERMODE is fast, MODE slow. Reload the mem in INNERMODE. */
1573 enum reg_class rclass
1574 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1575 if (get_reload_reg (curr_static_id->operand[nop].type, innermode,
1576 reg, rclass, TRUE, "slow mem", &new_reg))
1578 bool insert_before, insert_after;
1579 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1581 insert_before = (type != OP_OUT
1582 || partial_subreg_p (mode, innermode));
1583 insert_after = type != OP_IN;
1584 insert_move_for_subreg (insert_before ? &before : NULL,
1585 insert_after ? &after : NULL,
1586 reg, new_reg);
1588 SUBREG_REG (operand) = new_reg;
1590 /* Convert to MODE. */
1591 reg = operand;
1592 rclass
1593 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1594 if (get_reload_reg (curr_static_id->operand[nop].type, mode, reg,
1595 rclass, TRUE, "slow mem", &new_reg))
1597 bool insert_before, insert_after;
1598 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1600 insert_before = type != OP_OUT;
1601 insert_after = type != OP_IN;
1602 insert_move_for_subreg (insert_before ? &before : NULL,
1603 insert_after ? &after : NULL,
1604 reg, new_reg);
1606 *curr_id->operand_loc[nop] = new_reg;
1607 lra_process_new_insns (curr_insn, before, after,
1608 "Inserting slow mem reload");
1609 return true;
1612 /* If the address was valid and became invalid, prefer to reload
1613 the memory. Typical case is when the index scale should
1614 correspond the memory. */
1615 *curr_id->operand_loc[nop] = operand;
1616 /* Do not return false here as the MEM_P (reg) will be processed
1617 later in this function. */
1619 else if (REG_P (reg) && REGNO (reg) < FIRST_PSEUDO_REGISTER)
1621 alter_subreg (curr_id->operand_loc[nop], false);
1622 return true;
1624 else if (CONSTANT_P (reg))
1626 /* Try to simplify subreg of constant. It is usually result of
1627 equivalence substitution. */
1628 if (innermode == VOIDmode
1629 && (innermode = original_subreg_reg_mode[nop]) == VOIDmode)
1630 innermode = curr_static_id->operand[nop].mode;
1631 if ((new_reg = simplify_subreg (mode, reg, innermode,
1632 SUBREG_BYTE (operand))) != NULL_RTX)
1634 *curr_id->operand_loc[nop] = new_reg;
1635 return true;
1638 /* Put constant into memory when we have mixed modes. It generates
1639 a better code in most cases as it does not need a secondary
1640 reload memory. It also prevents LRA looping when LRA is using
1641 secondary reload memory again and again. */
1642 if (CONSTANT_P (reg) && CONST_POOL_OK_P (reg_mode, reg)
1643 && SCALAR_INT_MODE_P (reg_mode) != SCALAR_INT_MODE_P (mode))
1645 SUBREG_REG (operand) = force_const_mem (reg_mode, reg);
1646 alter_subreg (curr_id->operand_loc[nop], false);
1647 return true;
1649 /* Force a reload of the SUBREG_REG if this is a constant or PLUS or
1650 if there may be a problem accessing OPERAND in the outer
1651 mode. */
1652 if ((REG_P (reg)
1653 && REGNO (reg) >= FIRST_PSEUDO_REGISTER
1654 && (hard_regno = lra_get_regno_hard_regno (REGNO (reg))) >= 0
1655 /* Don't reload paradoxical subregs because we could be looping
1656 having repeatedly final regno out of hard regs range. */
1657 && (hard_regno_nregs[hard_regno][innermode]
1658 >= hard_regno_nregs[hard_regno][mode])
1659 && simplify_subreg_regno (hard_regno, innermode,
1660 SUBREG_BYTE (operand), mode) < 0
1661 /* Don't reload subreg for matching reload. It is actually
1662 valid subreg in LRA. */
1663 && ! LRA_SUBREG_P (operand))
1664 || CONSTANT_P (reg) || GET_CODE (reg) == PLUS || MEM_P (reg))
1666 enum reg_class rclass;
1668 if (REG_P (reg))
1669 /* There is a big probability that we will get the same class
1670 for the new pseudo and we will get the same insn which
1671 means infinite looping. So spill the new pseudo. */
1672 rclass = NO_REGS;
1673 else
1674 /* The class will be defined later in curr_insn_transform. */
1675 rclass
1676 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1678 if (get_reload_reg (curr_static_id->operand[nop].type, reg_mode, reg,
1679 rclass, TRUE, "subreg reg", &new_reg))
1681 bool insert_before, insert_after;
1682 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1684 insert_before = (type != OP_OUT
1685 || GET_MODE_SIZE (innermode) > GET_MODE_SIZE (mode));
1686 insert_after = (type != OP_IN);
1687 insert_move_for_subreg (insert_before ? &before : NULL,
1688 insert_after ? &after : NULL,
1689 reg, new_reg);
1691 SUBREG_REG (operand) = new_reg;
1692 lra_process_new_insns (curr_insn, before, after,
1693 "Inserting subreg reload");
1694 return true;
1696 /* Force a reload for a paradoxical subreg. For paradoxical subreg,
1697 IRA allocates hardreg to the inner pseudo reg according to its mode
1698 instead of the outermode, so the size of the hardreg may not be enough
1699 to contain the outermode operand, in that case we may need to insert
1700 reload for the reg. For the following two types of paradoxical subreg,
1701 we need to insert reload:
1702 1. If the op_type is OP_IN, and the hardreg could not be paired with
1703 other hardreg to contain the outermode operand
1704 (checked by in_hard_reg_set_p), we need to insert the reload.
1705 2. If the op_type is OP_OUT or OP_INOUT.
1707 Here is a paradoxical subreg example showing how the reload is generated:
1709 (insn 5 4 7 2 (set (reg:TI 106 [ __comp ])
1710 (subreg:TI (reg:DI 107 [ __comp ]) 0)) {*movti_internal_rex64}
1712 In IRA, reg107 is allocated to a DImode hardreg. We use x86-64 as example
1713 here, if reg107 is assigned to hardreg R15, because R15 is the last
1714 hardreg, compiler cannot find another hardreg to pair with R15 to
1715 contain TImode data. So we insert a TImode reload reg180 for it.
1716 After reload is inserted:
1718 (insn 283 0 0 (set (subreg:DI (reg:TI 180 [orig:107 __comp ] [107]) 0)
1719 (reg:DI 107 [ __comp ])) -1
1720 (insn 5 4 7 2 (set (reg:TI 106 [ __comp ])
1721 (subreg:TI (reg:TI 180 [orig:107 __comp ] [107]) 0)) {*movti_internal_rex64}
1723 Two reload hard registers will be allocated to reg180 to save TImode data
1724 in LRA_assign. */
1725 else if (REG_P (reg)
1726 && REGNO (reg) >= FIRST_PSEUDO_REGISTER
1727 && (hard_regno = lra_get_regno_hard_regno (REGNO (reg))) >= 0
1728 && (hard_regno_nregs[hard_regno][innermode]
1729 < hard_regno_nregs[hard_regno][mode])
1730 && (regclass = lra_get_allocno_class (REGNO (reg)))
1731 && (type != OP_IN
1732 || !in_hard_reg_set_p (reg_class_contents[regclass],
1733 mode, hard_regno)))
1735 /* The class will be defined later in curr_insn_transform. */
1736 enum reg_class rclass
1737 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1739 if (get_reload_reg (curr_static_id->operand[nop].type, mode, reg,
1740 rclass, TRUE, "paradoxical subreg", &new_reg))
1742 rtx subreg;
1743 bool insert_before, insert_after;
1745 PUT_MODE (new_reg, mode);
1746 subreg = gen_lowpart_SUBREG (innermode, new_reg);
1747 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1749 insert_before = (type != OP_OUT);
1750 insert_after = (type != OP_IN);
1751 insert_move_for_subreg (insert_before ? &before : NULL,
1752 insert_after ? &after : NULL,
1753 reg, subreg);
1755 SUBREG_REG (operand) = new_reg;
1756 lra_process_new_insns (curr_insn, before, after,
1757 "Inserting paradoxical subreg reload");
1758 return true;
1760 return false;
1763 /* Return TRUE if X refers for a hard register from SET. */
1764 static bool
1765 uses_hard_regs_p (rtx x, HARD_REG_SET set)
1767 int i, j, x_hard_regno;
1768 machine_mode mode;
1769 const char *fmt;
1770 enum rtx_code code;
1772 if (x == NULL_RTX)
1773 return false;
1774 code = GET_CODE (x);
1775 mode = GET_MODE (x);
1776 if (code == SUBREG)
1778 x = SUBREG_REG (x);
1779 code = GET_CODE (x);
1780 if (GET_MODE_SIZE (GET_MODE (x)) > GET_MODE_SIZE (mode))
1781 mode = GET_MODE (x);
1784 if (REG_P (x))
1786 x_hard_regno = get_hard_regno (x, true);
1787 return (x_hard_regno >= 0
1788 && overlaps_hard_reg_set_p (set, mode, x_hard_regno));
1790 if (MEM_P (x))
1792 struct address_info ad;
1794 decompose_mem_address (&ad, x);
1795 if (ad.base_term != NULL && uses_hard_regs_p (*ad.base_term, set))
1796 return true;
1797 if (ad.index_term != NULL && uses_hard_regs_p (*ad.index_term, set))
1798 return true;
1800 fmt = GET_RTX_FORMAT (code);
1801 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1803 if (fmt[i] == 'e')
1805 if (uses_hard_regs_p (XEXP (x, i), set))
1806 return true;
1808 else if (fmt[i] == 'E')
1810 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
1811 if (uses_hard_regs_p (XVECEXP (x, i, j), set))
1812 return true;
1815 return false;
1818 /* Return true if OP is a spilled pseudo. */
1819 static inline bool
1820 spilled_pseudo_p (rtx op)
1822 return (REG_P (op)
1823 && REGNO (op) >= FIRST_PSEUDO_REGISTER && in_mem_p (REGNO (op)));
1826 /* Return true if X is a general constant. */
1827 static inline bool
1828 general_constant_p (rtx x)
1830 return CONSTANT_P (x) && (! flag_pic || LEGITIMATE_PIC_OPERAND_P (x));
1833 static bool
1834 reg_in_class_p (rtx reg, enum reg_class cl)
1836 if (cl == NO_REGS)
1837 return get_reg_class (REGNO (reg)) == NO_REGS;
1838 return in_class_p (reg, cl, NULL);
1841 /* Return true if SET of RCLASS contains no hard regs which can be
1842 used in MODE. */
1843 static bool
1844 prohibited_class_reg_set_mode_p (enum reg_class rclass,
1845 HARD_REG_SET &set,
1846 machine_mode mode)
1848 HARD_REG_SET temp;
1850 lra_assert (hard_reg_set_subset_p (reg_class_contents[rclass], set));
1851 COPY_HARD_REG_SET (temp, set);
1852 AND_COMPL_HARD_REG_SET (temp, lra_no_alloc_regs);
1853 return (hard_reg_set_subset_p
1854 (temp, ira_prohibited_class_mode_regs[rclass][mode]));
1858 /* Used to check validity info about small class input operands. It
1859 should be incremented at start of processing an insn
1860 alternative. */
1861 static unsigned int curr_small_class_check = 0;
1863 /* Update number of used inputs of class OP_CLASS for operand NOP.
1864 Return true if we have more such class operands than the number of
1865 available regs. */
1866 static bool
1867 update_and_check_small_class_inputs (int nop, enum reg_class op_class)
1869 static unsigned int small_class_check[LIM_REG_CLASSES];
1870 static int small_class_input_nums[LIM_REG_CLASSES];
1872 if (SMALL_REGISTER_CLASS_P (op_class)
1873 /* We are interesting in classes became small because of fixing
1874 some hard regs, e.g. by an user through GCC options. */
1875 && hard_reg_set_intersect_p (reg_class_contents[op_class],
1876 ira_no_alloc_regs)
1877 && (curr_static_id->operand[nop].type != OP_OUT
1878 || curr_static_id->operand[nop].early_clobber))
1880 if (small_class_check[op_class] == curr_small_class_check)
1881 small_class_input_nums[op_class]++;
1882 else
1884 small_class_check[op_class] = curr_small_class_check;
1885 small_class_input_nums[op_class] = 1;
1887 if (small_class_input_nums[op_class] > ira_class_hard_regs_num[op_class])
1888 return true;
1890 return false;
1893 /* Major function to choose the current insn alternative and what
1894 operands should be reloaded and how. If ONLY_ALTERNATIVE is not
1895 negative we should consider only this alternative. Return false if
1896 we can not choose the alternative or find how to reload the
1897 operands. */
1898 static bool
1899 process_alt_operands (int only_alternative)
1901 bool ok_p = false;
1902 int nop, overall, nalt;
1903 int n_alternatives = curr_static_id->n_alternatives;
1904 int n_operands = curr_static_id->n_operands;
1905 /* LOSERS counts the operands that don't fit this alternative and
1906 would require loading. */
1907 int losers;
1908 int addr_losers;
1909 /* REJECT is a count of how undesirable this alternative says it is
1910 if any reloading is required. If the alternative matches exactly
1911 then REJECT is ignored, but otherwise it gets this much counted
1912 against it in addition to the reloading needed. */
1913 int reject;
1914 /* This is defined by '!' or '?' alternative constraint and added to
1915 reject. But in some cases it can be ignored. */
1916 int static_reject;
1917 int op_reject;
1918 /* The number of elements in the following array. */
1919 int early_clobbered_regs_num;
1920 /* Numbers of operands which are early clobber registers. */
1921 int early_clobbered_nops[MAX_RECOG_OPERANDS];
1922 enum reg_class curr_alt[MAX_RECOG_OPERANDS];
1923 HARD_REG_SET curr_alt_set[MAX_RECOG_OPERANDS];
1924 bool curr_alt_match_win[MAX_RECOG_OPERANDS];
1925 bool curr_alt_win[MAX_RECOG_OPERANDS];
1926 bool curr_alt_offmemok[MAX_RECOG_OPERANDS];
1927 int curr_alt_matches[MAX_RECOG_OPERANDS];
1928 /* The number of elements in the following array. */
1929 int curr_alt_dont_inherit_ops_num;
1930 /* Numbers of operands whose reload pseudos should not be inherited. */
1931 int curr_alt_dont_inherit_ops[MAX_RECOG_OPERANDS];
1932 rtx op;
1933 /* The register when the operand is a subreg of register, otherwise the
1934 operand itself. */
1935 rtx no_subreg_reg_operand[MAX_RECOG_OPERANDS];
1936 /* The register if the operand is a register or subreg of register,
1937 otherwise NULL. */
1938 rtx operand_reg[MAX_RECOG_OPERANDS];
1939 int hard_regno[MAX_RECOG_OPERANDS];
1940 machine_mode biggest_mode[MAX_RECOG_OPERANDS];
1941 int reload_nregs, reload_sum;
1942 bool costly_p;
1943 enum reg_class cl;
1945 /* Calculate some data common for all alternatives to speed up the
1946 function. */
1947 for (nop = 0; nop < n_operands; nop++)
1949 rtx reg;
1951 op = no_subreg_reg_operand[nop] = *curr_id->operand_loc[nop];
1952 /* The real hard regno of the operand after the allocation. */
1953 hard_regno[nop] = get_hard_regno (op, true);
1955 operand_reg[nop] = reg = op;
1956 biggest_mode[nop] = GET_MODE (op);
1957 if (GET_CODE (op) == SUBREG)
1959 operand_reg[nop] = reg = SUBREG_REG (op);
1960 if (GET_MODE_SIZE (biggest_mode[nop])
1961 < GET_MODE_SIZE (GET_MODE (reg)))
1962 biggest_mode[nop] = GET_MODE (reg);
1964 if (! REG_P (reg))
1965 operand_reg[nop] = NULL_RTX;
1966 else if (REGNO (reg) >= FIRST_PSEUDO_REGISTER
1967 || ((int) REGNO (reg)
1968 == lra_get_elimination_hard_regno (REGNO (reg))))
1969 no_subreg_reg_operand[nop] = reg;
1970 else
1971 operand_reg[nop] = no_subreg_reg_operand[nop]
1972 /* Just use natural mode for elimination result. It should
1973 be enough for extra constraints hooks. */
1974 = regno_reg_rtx[hard_regno[nop]];
1977 /* The constraints are made of several alternatives. Each operand's
1978 constraint looks like foo,bar,... with commas separating the
1979 alternatives. The first alternatives for all operands go
1980 together, the second alternatives go together, etc.
1982 First loop over alternatives. */
1983 alternative_mask preferred = curr_id->preferred_alternatives;
1984 if (only_alternative >= 0)
1985 preferred &= ALTERNATIVE_BIT (only_alternative);
1987 for (nalt = 0; nalt < n_alternatives; nalt++)
1989 /* Loop over operands for one constraint alternative. */
1990 if (!TEST_BIT (preferred, nalt))
1991 continue;
1993 curr_small_class_check++;
1994 overall = losers = addr_losers = 0;
1995 static_reject = reject = reload_nregs = reload_sum = 0;
1996 for (nop = 0; nop < n_operands; nop++)
1998 int inc = (curr_static_id
1999 ->operand_alternative[nalt * n_operands + nop].reject);
2000 if (lra_dump_file != NULL && inc != 0)
2001 fprintf (lra_dump_file,
2002 " Staticly defined alt reject+=%d\n", inc);
2003 static_reject += inc;
2005 reject += static_reject;
2006 early_clobbered_regs_num = 0;
2008 for (nop = 0; nop < n_operands; nop++)
2010 const char *p;
2011 char *end;
2012 int len, c, m, i, opalt_num, this_alternative_matches;
2013 bool win, did_match, offmemok, early_clobber_p;
2014 /* false => this operand can be reloaded somehow for this
2015 alternative. */
2016 bool badop;
2017 /* true => this operand can be reloaded if the alternative
2018 allows regs. */
2019 bool winreg;
2020 /* True if a constant forced into memory would be OK for
2021 this operand. */
2022 bool constmemok;
2023 enum reg_class this_alternative, this_costly_alternative;
2024 HARD_REG_SET this_alternative_set, this_costly_alternative_set;
2025 bool this_alternative_match_win, this_alternative_win;
2026 bool this_alternative_offmemok;
2027 bool scratch_p;
2028 machine_mode mode;
2029 enum constraint_num cn;
2031 opalt_num = nalt * n_operands + nop;
2032 if (curr_static_id->operand_alternative[opalt_num].anything_ok)
2034 /* Fast track for no constraints at all. */
2035 curr_alt[nop] = NO_REGS;
2036 CLEAR_HARD_REG_SET (curr_alt_set[nop]);
2037 curr_alt_win[nop] = true;
2038 curr_alt_match_win[nop] = false;
2039 curr_alt_offmemok[nop] = false;
2040 curr_alt_matches[nop] = -1;
2041 continue;
2044 op = no_subreg_reg_operand[nop];
2045 mode = curr_operand_mode[nop];
2047 win = did_match = winreg = offmemok = constmemok = false;
2048 badop = true;
2050 early_clobber_p = false;
2051 p = curr_static_id->operand_alternative[opalt_num].constraint;
2053 this_costly_alternative = this_alternative = NO_REGS;
2054 /* We update set of possible hard regs besides its class
2055 because reg class might be inaccurate. For example,
2056 union of LO_REGS (l), HI_REGS(h), and STACK_REG(k) in ARM
2057 is translated in HI_REGS because classes are merged by
2058 pairs and there is no accurate intermediate class. */
2059 CLEAR_HARD_REG_SET (this_alternative_set);
2060 CLEAR_HARD_REG_SET (this_costly_alternative_set);
2061 this_alternative_win = false;
2062 this_alternative_match_win = false;
2063 this_alternative_offmemok = false;
2064 this_alternative_matches = -1;
2066 /* An empty constraint should be excluded by the fast
2067 track. */
2068 lra_assert (*p != 0 && *p != ',');
2070 op_reject = 0;
2071 /* Scan this alternative's specs for this operand; set WIN
2072 if the operand fits any letter in this alternative.
2073 Otherwise, clear BADOP if this operand could fit some
2074 letter after reloads, or set WINREG if this operand could
2075 fit after reloads provided the constraint allows some
2076 registers. */
2077 costly_p = false;
2080 switch ((c = *p, len = CONSTRAINT_LEN (c, p)), c)
2082 case '\0':
2083 len = 0;
2084 break;
2085 case ',':
2086 c = '\0';
2087 break;
2089 case '&':
2090 early_clobber_p = true;
2091 break;
2093 case '$':
2094 op_reject += LRA_MAX_REJECT;
2095 break;
2096 case '^':
2097 op_reject += LRA_LOSER_COST_FACTOR;
2098 break;
2100 case '#':
2101 /* Ignore rest of this alternative. */
2102 c = '\0';
2103 break;
2105 case '0': case '1': case '2': case '3': case '4':
2106 case '5': case '6': case '7': case '8': case '9':
2108 int m_hregno;
2109 bool match_p;
2111 m = strtoul (p, &end, 10);
2112 p = end;
2113 len = 0;
2114 lra_assert (nop > m);
2116 this_alternative_matches = m;
2117 m_hregno = get_hard_regno (*curr_id->operand_loc[m], false);
2118 /* We are supposed to match a previous operand.
2119 If we do, we win if that one did. If we do
2120 not, count both of the operands as losers.
2121 (This is too conservative, since most of the
2122 time only a single reload insn will be needed
2123 to make the two operands win. As a result,
2124 this alternative may be rejected when it is
2125 actually desirable.) */
2126 match_p = false;
2127 if (operands_match_p (*curr_id->operand_loc[nop],
2128 *curr_id->operand_loc[m], m_hregno))
2130 /* We should reject matching of an early
2131 clobber operand if the matching operand is
2132 not dying in the insn. */
2133 if (! curr_static_id->operand[m].early_clobber
2134 || operand_reg[nop] == NULL_RTX
2135 || (find_regno_note (curr_insn, REG_DEAD,
2136 REGNO (op))
2137 || REGNO (op) == REGNO (operand_reg[m])))
2138 match_p = true;
2140 if (match_p)
2142 /* If we are matching a non-offsettable
2143 address where an offsettable address was
2144 expected, then we must reject this
2145 combination, because we can't reload
2146 it. */
2147 if (curr_alt_offmemok[m]
2148 && MEM_P (*curr_id->operand_loc[m])
2149 && curr_alt[m] == NO_REGS && ! curr_alt_win[m])
2150 continue;
2152 else
2154 /* Operands don't match. Both operands must
2155 allow a reload register, otherwise we
2156 cannot make them match. */
2157 if (curr_alt[m] == NO_REGS)
2158 break;
2159 /* Retroactively mark the operand we had to
2160 match as a loser, if it wasn't already and
2161 it wasn't matched to a register constraint
2162 (e.g it might be matched by memory). */
2163 if (curr_alt_win[m]
2164 && (operand_reg[m] == NULL_RTX
2165 || hard_regno[m] < 0))
2167 losers++;
2168 reload_nregs
2169 += (ira_reg_class_max_nregs[curr_alt[m]]
2170 [GET_MODE (*curr_id->operand_loc[m])]);
2173 /* Prefer matching earlyclobber alternative as
2174 it results in less hard regs required for
2175 the insn than a non-matching earlyclobber
2176 alternative. */
2177 if (curr_static_id->operand[m].early_clobber)
2179 if (lra_dump_file != NULL)
2180 fprintf
2181 (lra_dump_file,
2182 " %d Matching earlyclobber alt:"
2183 " reject--\n",
2184 nop);
2185 reject--;
2187 /* Otherwise we prefer no matching
2188 alternatives because it gives more freedom
2189 in RA. */
2190 else if (operand_reg[nop] == NULL_RTX
2191 || (find_regno_note (curr_insn, REG_DEAD,
2192 REGNO (operand_reg[nop]))
2193 == NULL_RTX))
2195 if (lra_dump_file != NULL)
2196 fprintf
2197 (lra_dump_file,
2198 " %d Matching alt: reject+=2\n",
2199 nop);
2200 reject += 2;
2203 /* If we have to reload this operand and some
2204 previous operand also had to match the same
2205 thing as this operand, we don't know how to do
2206 that. */
2207 if (!match_p || !curr_alt_win[m])
2209 for (i = 0; i < nop; i++)
2210 if (curr_alt_matches[i] == m)
2211 break;
2212 if (i < nop)
2213 break;
2215 else
2216 did_match = true;
2218 /* This can be fixed with reloads if the operand
2219 we are supposed to match can be fixed with
2220 reloads. */
2221 badop = false;
2222 this_alternative = curr_alt[m];
2223 COPY_HARD_REG_SET (this_alternative_set, curr_alt_set[m]);
2224 winreg = this_alternative != NO_REGS;
2225 break;
2228 case 'g':
2229 if (MEM_P (op)
2230 || general_constant_p (op)
2231 || spilled_pseudo_p (op))
2232 win = true;
2233 cl = GENERAL_REGS;
2234 goto reg;
2236 default:
2237 cn = lookup_constraint (p);
2238 switch (get_constraint_type (cn))
2240 case CT_REGISTER:
2241 cl = reg_class_for_constraint (cn);
2242 if (cl != NO_REGS)
2243 goto reg;
2244 break;
2246 case CT_CONST_INT:
2247 if (CONST_INT_P (op)
2248 && insn_const_int_ok_for_constraint (INTVAL (op), cn))
2249 win = true;
2250 break;
2252 case CT_MEMORY:
2253 if (MEM_P (op)
2254 && satisfies_memory_constraint_p (op, cn))
2255 win = true;
2256 else if (spilled_pseudo_p (op))
2257 win = true;
2259 /* If we didn't already win, we can reload constants
2260 via force_const_mem or put the pseudo value into
2261 memory, or make other memory by reloading the
2262 address like for 'o'. */
2263 if (CONST_POOL_OK_P (mode, op)
2264 || MEM_P (op) || REG_P (op)
2265 /* We can restore the equiv insn by a
2266 reload. */
2267 || equiv_substition_p[nop])
2268 badop = false;
2269 constmemok = true;
2270 offmemok = true;
2271 break;
2273 case CT_ADDRESS:
2274 /* If we didn't already win, we can reload the address
2275 into a base register. */
2276 if (satisfies_address_constraint_p (op, cn))
2277 win = true;
2278 cl = base_reg_class (VOIDmode, ADDR_SPACE_GENERIC,
2279 ADDRESS, SCRATCH);
2280 badop = false;
2281 goto reg;
2283 case CT_FIXED_FORM:
2284 if (constraint_satisfied_p (op, cn))
2285 win = true;
2286 break;
2288 case CT_SPECIAL_MEMORY:
2289 if (MEM_P (op)
2290 && satisfies_memory_constraint_p (op, cn))
2291 win = true;
2292 else if (spilled_pseudo_p (op))
2293 win = true;
2294 break;
2296 break;
2298 reg:
2299 this_alternative = reg_class_subunion[this_alternative][cl];
2300 IOR_HARD_REG_SET (this_alternative_set,
2301 reg_class_contents[cl]);
2302 if (costly_p)
2304 this_costly_alternative
2305 = reg_class_subunion[this_costly_alternative][cl];
2306 IOR_HARD_REG_SET (this_costly_alternative_set,
2307 reg_class_contents[cl]);
2309 if (mode == BLKmode)
2310 break;
2311 winreg = true;
2312 if (REG_P (op))
2314 if (hard_regno[nop] >= 0
2315 && in_hard_reg_set_p (this_alternative_set,
2316 mode, hard_regno[nop]))
2317 win = true;
2318 else if (hard_regno[nop] < 0
2319 && in_class_p (op, this_alternative, NULL))
2320 win = true;
2322 break;
2324 if (c != ' ' && c != '\t')
2325 costly_p = c == '*';
2327 while ((p += len), c);
2329 scratch_p = (operand_reg[nop] != NULL_RTX
2330 && lra_former_scratch_p (REGNO (operand_reg[nop])));
2331 /* Record which operands fit this alternative. */
2332 if (win)
2334 this_alternative_win = true;
2335 if (operand_reg[nop] != NULL_RTX)
2337 if (hard_regno[nop] >= 0)
2339 if (in_hard_reg_set_p (this_costly_alternative_set,
2340 mode, hard_regno[nop]))
2342 if (lra_dump_file != NULL)
2343 fprintf (lra_dump_file,
2344 " %d Costly set: reject++\n",
2345 nop);
2346 reject++;
2349 else
2351 /* Prefer won reg to spilled pseudo under other
2352 equal conditions for possibe inheritance. */
2353 if (! scratch_p)
2355 if (lra_dump_file != NULL)
2356 fprintf
2357 (lra_dump_file,
2358 " %d Non pseudo reload: reject++\n",
2359 nop);
2360 reject++;
2362 if (in_class_p (operand_reg[nop],
2363 this_costly_alternative, NULL))
2365 if (lra_dump_file != NULL)
2366 fprintf
2367 (lra_dump_file,
2368 " %d Non pseudo costly reload:"
2369 " reject++\n",
2370 nop);
2371 reject++;
2374 /* We simulate the behavior of old reload here.
2375 Although scratches need hard registers and it
2376 might result in spilling other pseudos, no reload
2377 insns are generated for the scratches. So it
2378 might cost something but probably less than old
2379 reload pass believes. */
2380 if (scratch_p)
2382 if (lra_dump_file != NULL)
2383 fprintf (lra_dump_file,
2384 " %d Scratch win: reject+=2\n",
2385 nop);
2386 reject += 2;
2390 else if (did_match)
2391 this_alternative_match_win = true;
2392 else
2394 int const_to_mem = 0;
2395 bool no_regs_p;
2397 reject += op_reject;
2398 /* Never do output reload of stack pointer. It makes
2399 impossible to do elimination when SP is changed in
2400 RTL. */
2401 if (op == stack_pointer_rtx && ! frame_pointer_needed
2402 && curr_static_id->operand[nop].type != OP_IN)
2403 goto fail;
2405 /* If this alternative asks for a specific reg class, see if there
2406 is at least one allocatable register in that class. */
2407 no_regs_p
2408 = (this_alternative == NO_REGS
2409 || (hard_reg_set_subset_p
2410 (reg_class_contents[this_alternative],
2411 lra_no_alloc_regs)));
2413 /* For asms, verify that the class for this alternative is possible
2414 for the mode that is specified. */
2415 if (!no_regs_p && INSN_CODE (curr_insn) < 0)
2417 int i;
2418 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
2419 if (HARD_REGNO_MODE_OK (i, mode)
2420 && in_hard_reg_set_p (reg_class_contents[this_alternative],
2421 mode, i))
2422 break;
2423 if (i == FIRST_PSEUDO_REGISTER)
2424 winreg = false;
2427 /* If this operand accepts a register, and if the
2428 register class has at least one allocatable register,
2429 then this operand can be reloaded. */
2430 if (winreg && !no_regs_p)
2431 badop = false;
2433 if (badop)
2435 if (lra_dump_file != NULL)
2436 fprintf (lra_dump_file,
2437 " alt=%d: Bad operand -- refuse\n",
2438 nalt);
2439 goto fail;
2442 if (this_alternative != NO_REGS)
2444 HARD_REG_SET available_regs;
2446 COPY_HARD_REG_SET (available_regs,
2447 reg_class_contents[this_alternative]);
2448 AND_COMPL_HARD_REG_SET
2449 (available_regs,
2450 ira_prohibited_class_mode_regs[this_alternative][mode]);
2451 AND_COMPL_HARD_REG_SET (available_regs, lra_no_alloc_regs);
2452 if (hard_reg_set_empty_p (available_regs))
2454 /* There are no hard regs holding a value of given
2455 mode. */
2456 if (offmemok)
2458 this_alternative = NO_REGS;
2459 if (lra_dump_file != NULL)
2460 fprintf (lra_dump_file,
2461 " %d Using memory because of"
2462 " a bad mode: reject+=2\n",
2463 nop);
2464 reject += 2;
2466 else
2468 if (lra_dump_file != NULL)
2469 fprintf (lra_dump_file,
2470 " alt=%d: Wrong mode -- refuse\n",
2471 nalt);
2472 goto fail;
2477 /* If not assigned pseudo has a class which a subset of
2478 required reg class, it is a less costly alternative
2479 as the pseudo still can get a hard reg of necessary
2480 class. */
2481 if (! no_regs_p && REG_P (op) && hard_regno[nop] < 0
2482 && (cl = get_reg_class (REGNO (op))) != NO_REGS
2483 && ira_class_subset_p[this_alternative][cl])
2485 if (lra_dump_file != NULL)
2486 fprintf
2487 (lra_dump_file,
2488 " %d Super set class reg: reject-=3\n", nop);
2489 reject -= 3;
2492 this_alternative_offmemok = offmemok;
2493 if (this_costly_alternative != NO_REGS)
2495 if (lra_dump_file != NULL)
2496 fprintf (lra_dump_file,
2497 " %d Costly loser: reject++\n", nop);
2498 reject++;
2500 /* If the operand is dying, has a matching constraint,
2501 and satisfies constraints of the matched operand
2502 which failed to satisfy the own constraints, most probably
2503 the reload for this operand will be gone. */
2504 if (this_alternative_matches >= 0
2505 && !curr_alt_win[this_alternative_matches]
2506 && REG_P (op)
2507 && find_regno_note (curr_insn, REG_DEAD, REGNO (op))
2508 && (hard_regno[nop] >= 0
2509 ? in_hard_reg_set_p (this_alternative_set,
2510 mode, hard_regno[nop])
2511 : in_class_p (op, this_alternative, NULL)))
2513 if (lra_dump_file != NULL)
2514 fprintf
2515 (lra_dump_file,
2516 " %d Dying matched operand reload: reject++\n",
2517 nop);
2518 reject++;
2520 else
2522 /* Strict_low_part requires to reload the register
2523 not the sub-register. In this case we should
2524 check that a final reload hard reg can hold the
2525 value mode. */
2526 if (curr_static_id->operand[nop].strict_low
2527 && REG_P (op)
2528 && hard_regno[nop] < 0
2529 && GET_CODE (*curr_id->operand_loc[nop]) == SUBREG
2530 && ira_class_hard_regs_num[this_alternative] > 0
2531 && ! HARD_REGNO_MODE_OK (ira_class_hard_regs
2532 [this_alternative][0],
2533 GET_MODE
2534 (*curr_id->operand_loc[nop])))
2536 if (lra_dump_file != NULL)
2537 fprintf
2538 (lra_dump_file,
2539 " alt=%d: Strict low subreg reload -- refuse\n",
2540 nalt);
2541 goto fail;
2543 losers++;
2545 if (operand_reg[nop] != NULL_RTX
2546 /* Output operands and matched input operands are
2547 not inherited. The following conditions do not
2548 exactly describe the previous statement but they
2549 are pretty close. */
2550 && curr_static_id->operand[nop].type != OP_OUT
2551 && (this_alternative_matches < 0
2552 || curr_static_id->operand[nop].type != OP_IN))
2554 int last_reload = (lra_reg_info[ORIGINAL_REGNO
2555 (operand_reg[nop])]
2556 .last_reload);
2558 /* The value of reload_sum has sense only if we
2559 process insns in their order. It happens only on
2560 the first constraints sub-pass when we do most of
2561 reload work. */
2562 if (lra_constraint_iter == 1 && last_reload > bb_reload_num)
2563 reload_sum += last_reload - bb_reload_num;
2565 /* If this is a constant that is reloaded into the
2566 desired class by copying it to memory first, count
2567 that as another reload. This is consistent with
2568 other code and is required to avoid choosing another
2569 alternative when the constant is moved into memory.
2570 Note that the test here is precisely the same as in
2571 the code below that calls force_const_mem. */
2572 if (CONST_POOL_OK_P (mode, op)
2573 && ((targetm.preferred_reload_class
2574 (op, this_alternative) == NO_REGS)
2575 || no_input_reloads_p))
2577 const_to_mem = 1;
2578 if (! no_regs_p)
2579 losers++;
2582 /* Alternative loses if it requires a type of reload not
2583 permitted for this insn. We can always reload
2584 objects with a REG_UNUSED note. */
2585 if ((curr_static_id->operand[nop].type != OP_IN
2586 && no_output_reloads_p
2587 && ! find_reg_note (curr_insn, REG_UNUSED, op))
2588 || (curr_static_id->operand[nop].type != OP_OUT
2589 && no_input_reloads_p && ! const_to_mem)
2590 || (this_alternative_matches >= 0
2591 && (no_input_reloads_p
2592 || (no_output_reloads_p
2593 && (curr_static_id->operand
2594 [this_alternative_matches].type != OP_IN)
2595 && ! find_reg_note (curr_insn, REG_UNUSED,
2596 no_subreg_reg_operand
2597 [this_alternative_matches])))))
2599 if (lra_dump_file != NULL)
2600 fprintf
2601 (lra_dump_file,
2602 " alt=%d: No input/otput reload -- refuse\n",
2603 nalt);
2604 goto fail;
2607 /* Alternative loses if it required class pseudo can not
2608 hold value of required mode. Such insns can be
2609 described by insn definitions with mode iterators. */
2610 if (GET_MODE (*curr_id->operand_loc[nop]) != VOIDmode
2611 && ! hard_reg_set_empty_p (this_alternative_set)
2612 /* It is common practice for constraints to use a
2613 class which does not have actually enough regs to
2614 hold the value (e.g. x86 AREG for mode requiring
2615 more one general reg). Therefore we have 2
2616 conditions to check that the reload pseudo can
2617 not hold the mode value. */
2618 && ! HARD_REGNO_MODE_OK (ira_class_hard_regs
2619 [this_alternative][0],
2620 GET_MODE (*curr_id->operand_loc[nop]))
2621 /* The above condition is not enough as the first
2622 reg in ira_class_hard_regs can be not aligned for
2623 multi-words mode values. */
2624 && (prohibited_class_reg_set_mode_p
2625 (this_alternative, this_alternative_set,
2626 GET_MODE (*curr_id->operand_loc[nop]))))
2628 if (lra_dump_file != NULL)
2629 fprintf (lra_dump_file,
2630 " alt=%d: reload pseudo for op %d "
2631 " can not hold the mode value -- refuse\n",
2632 nalt, nop);
2633 goto fail;
2636 /* Check strong discouragement of reload of non-constant
2637 into class THIS_ALTERNATIVE. */
2638 if (! CONSTANT_P (op) && ! no_regs_p
2639 && (targetm.preferred_reload_class
2640 (op, this_alternative) == NO_REGS
2641 || (curr_static_id->operand[nop].type == OP_OUT
2642 && (targetm.preferred_output_reload_class
2643 (op, this_alternative) == NO_REGS))))
2645 if (lra_dump_file != NULL)
2646 fprintf (lra_dump_file,
2647 " %d Non-prefered reload: reject+=%d\n",
2648 nop, LRA_MAX_REJECT);
2649 reject += LRA_MAX_REJECT;
2652 if (! (MEM_P (op) && offmemok)
2653 && ! (const_to_mem && constmemok))
2655 /* We prefer to reload pseudos over reloading other
2656 things, since such reloads may be able to be
2657 eliminated later. So bump REJECT in other cases.
2658 Don't do this in the case where we are forcing a
2659 constant into memory and it will then win since
2660 we don't want to have a different alternative
2661 match then. */
2662 if (! (REG_P (op) && REGNO (op) >= FIRST_PSEUDO_REGISTER))
2664 if (lra_dump_file != NULL)
2665 fprintf
2666 (lra_dump_file,
2667 " %d Non-pseudo reload: reject+=2\n",
2668 nop);
2669 reject += 2;
2672 if (! no_regs_p)
2673 reload_nregs
2674 += ira_reg_class_max_nregs[this_alternative][mode];
2676 if (SMALL_REGISTER_CLASS_P (this_alternative))
2678 if (lra_dump_file != NULL)
2679 fprintf
2680 (lra_dump_file,
2681 " %d Small class reload: reject+=%d\n",
2682 nop, LRA_LOSER_COST_FACTOR / 2);
2683 reject += LRA_LOSER_COST_FACTOR / 2;
2687 /* We are trying to spill pseudo into memory. It is
2688 usually more costly than moving to a hard register
2689 although it might takes the same number of
2690 reloads.
2692 Non-pseudo spill may happen also. Suppose a target allows both
2693 register and memory in the operand constraint alternatives,
2694 then it's typical that an eliminable register has a substition
2695 of "base + offset" which can either be reloaded by a simple
2696 "new_reg <= base + offset" which will match the register
2697 constraint, or a similar reg addition followed by further spill
2698 to and reload from memory which will match the memory
2699 constraint, but this memory spill will be much more costly
2700 usually.
2702 Code below increases the reject for both pseudo and non-pseudo
2703 spill. */
2704 if (no_regs_p
2705 && !(MEM_P (op) && offmemok)
2706 && !(REG_P (op) && hard_regno[nop] < 0))
2708 if (lra_dump_file != NULL)
2709 fprintf
2710 (lra_dump_file,
2711 " %d Spill %spseudo into memory: reject+=3\n",
2712 nop, REG_P (op) ? "" : "Non-");
2713 reject += 3;
2714 if (VECTOR_MODE_P (mode))
2716 /* Spilling vectors into memory is usually more
2717 costly as they contain big values. */
2718 if (lra_dump_file != NULL)
2719 fprintf
2720 (lra_dump_file,
2721 " %d Spill vector pseudo: reject+=2\n",
2722 nop);
2723 reject += 2;
2727 /* When we use an operand requiring memory in given
2728 alternative, the insn should write *and* read the
2729 value to/from memory it is costly in comparison with
2730 an insn alternative which does not use memory
2731 (e.g. register or immediate operand). We exclude
2732 memory operand for such case as we can satisfy the
2733 memory constraints by reloading address. */
2734 if (no_regs_p && offmemok && !MEM_P (op))
2736 if (lra_dump_file != NULL)
2737 fprintf
2738 (lra_dump_file,
2739 " Using memory insn operand %d: reject+=3\n",
2740 nop);
2741 reject += 3;
2744 #ifdef SECONDARY_MEMORY_NEEDED
2745 /* If reload requires moving value through secondary
2746 memory, it will need one more insn at least. */
2747 if (this_alternative != NO_REGS
2748 && REG_P (op) && (cl = get_reg_class (REGNO (op))) != NO_REGS
2749 && ((curr_static_id->operand[nop].type != OP_OUT
2750 && SECONDARY_MEMORY_NEEDED (cl, this_alternative,
2751 GET_MODE (op)))
2752 || (curr_static_id->operand[nop].type != OP_IN
2753 && SECONDARY_MEMORY_NEEDED (this_alternative, cl,
2754 GET_MODE (op)))))
2755 losers++;
2756 #endif
2757 /* Input reloads can be inherited more often than output
2758 reloads can be removed, so penalize output
2759 reloads. */
2760 if (!REG_P (op) || curr_static_id->operand[nop].type != OP_IN)
2762 if (lra_dump_file != NULL)
2763 fprintf
2764 (lra_dump_file,
2765 " %d Non input pseudo reload: reject++\n",
2766 nop);
2767 reject++;
2770 if (MEM_P (op) && offmemok)
2771 addr_losers++;
2772 else if (curr_static_id->operand[nop].type == OP_INOUT)
2774 if (lra_dump_file != NULL)
2775 fprintf
2776 (lra_dump_file,
2777 " %d Input/Output reload: reject+=%d\n",
2778 nop, LRA_LOSER_COST_FACTOR);
2779 reject += LRA_LOSER_COST_FACTOR;
2783 if (early_clobber_p && ! scratch_p)
2785 if (lra_dump_file != NULL)
2786 fprintf (lra_dump_file,
2787 " %d Early clobber: reject++\n", nop);
2788 reject++;
2790 /* ??? We check early clobbers after processing all operands
2791 (see loop below) and there we update the costs more.
2792 Should we update the cost (may be approximately) here
2793 because of early clobber register reloads or it is a rare
2794 or non-important thing to be worth to do it. */
2795 overall = (losers * LRA_LOSER_COST_FACTOR + reject
2796 - (addr_losers == losers ? static_reject : 0));
2797 if ((best_losers == 0 || losers != 0) && best_overall < overall)
2799 if (lra_dump_file != NULL)
2800 fprintf (lra_dump_file,
2801 " alt=%d,overall=%d,losers=%d -- refuse\n",
2802 nalt, overall, losers);
2803 goto fail;
2806 if (update_and_check_small_class_inputs (nop, this_alternative))
2808 if (lra_dump_file != NULL)
2809 fprintf (lra_dump_file,
2810 " alt=%d, not enough small class regs -- refuse\n",
2811 nalt);
2812 goto fail;
2814 curr_alt[nop] = this_alternative;
2815 COPY_HARD_REG_SET (curr_alt_set[nop], this_alternative_set);
2816 curr_alt_win[nop] = this_alternative_win;
2817 curr_alt_match_win[nop] = this_alternative_match_win;
2818 curr_alt_offmemok[nop] = this_alternative_offmemok;
2819 curr_alt_matches[nop] = this_alternative_matches;
2821 if (this_alternative_matches >= 0
2822 && !did_match && !this_alternative_win)
2823 curr_alt_win[this_alternative_matches] = false;
2825 if (early_clobber_p && operand_reg[nop] != NULL_RTX)
2826 early_clobbered_nops[early_clobbered_regs_num++] = nop;
2829 if (curr_insn_set != NULL_RTX && n_operands == 2
2830 /* Prevent processing non-move insns. */
2831 && (GET_CODE (SET_SRC (curr_insn_set)) == SUBREG
2832 || SET_SRC (curr_insn_set) == no_subreg_reg_operand[1])
2833 && ((! curr_alt_win[0] && ! curr_alt_win[1]
2834 && REG_P (no_subreg_reg_operand[0])
2835 && REG_P (no_subreg_reg_operand[1])
2836 && (reg_in_class_p (no_subreg_reg_operand[0], curr_alt[1])
2837 || reg_in_class_p (no_subreg_reg_operand[1], curr_alt[0])))
2838 || (! curr_alt_win[0] && curr_alt_win[1]
2839 && REG_P (no_subreg_reg_operand[1])
2840 /* Check that we reload memory not the memory
2841 address. */
2842 && ! (curr_alt_offmemok[0]
2843 && MEM_P (no_subreg_reg_operand[0]))
2844 && reg_in_class_p (no_subreg_reg_operand[1], curr_alt[0]))
2845 || (curr_alt_win[0] && ! curr_alt_win[1]
2846 && REG_P (no_subreg_reg_operand[0])
2847 /* Check that we reload memory not the memory
2848 address. */
2849 && ! (curr_alt_offmemok[1]
2850 && MEM_P (no_subreg_reg_operand[1]))
2851 && reg_in_class_p (no_subreg_reg_operand[0], curr_alt[1])
2852 && (! CONST_POOL_OK_P (curr_operand_mode[1],
2853 no_subreg_reg_operand[1])
2854 || (targetm.preferred_reload_class
2855 (no_subreg_reg_operand[1],
2856 (enum reg_class) curr_alt[1]) != NO_REGS))
2857 /* If it is a result of recent elimination in move
2858 insn we can transform it into an add still by
2859 using this alternative. */
2860 && GET_CODE (no_subreg_reg_operand[1]) != PLUS)))
2862 /* We have a move insn and a new reload insn will be similar
2863 to the current insn. We should avoid such situation as
2864 it results in LRA cycling. */
2865 if (lra_dump_file != NULL)
2866 fprintf (lra_dump_file,
2867 " Cycle danger: overall += LRA_MAX_REJECT\n");
2868 overall += LRA_MAX_REJECT;
2870 ok_p = true;
2871 curr_alt_dont_inherit_ops_num = 0;
2872 for (nop = 0; nop < early_clobbered_regs_num; nop++)
2874 int i, j, clobbered_hard_regno, first_conflict_j, last_conflict_j;
2875 HARD_REG_SET temp_set;
2877 i = early_clobbered_nops[nop];
2878 if ((! curr_alt_win[i] && ! curr_alt_match_win[i])
2879 || hard_regno[i] < 0)
2880 continue;
2881 lra_assert (operand_reg[i] != NULL_RTX);
2882 clobbered_hard_regno = hard_regno[i];
2883 CLEAR_HARD_REG_SET (temp_set);
2884 add_to_hard_reg_set (&temp_set, biggest_mode[i], clobbered_hard_regno);
2885 first_conflict_j = last_conflict_j = -1;
2886 for (j = 0; j < n_operands; j++)
2887 if (j == i
2888 /* We don't want process insides of match_operator and
2889 match_parallel because otherwise we would process
2890 their operands once again generating a wrong
2891 code. */
2892 || curr_static_id->operand[j].is_operator)
2893 continue;
2894 else if ((curr_alt_matches[j] == i && curr_alt_match_win[j])
2895 || (curr_alt_matches[i] == j && curr_alt_match_win[i]))
2896 continue;
2897 /* If we don't reload j-th operand, check conflicts. */
2898 else if ((curr_alt_win[j] || curr_alt_match_win[j])
2899 && uses_hard_regs_p (*curr_id->operand_loc[j], temp_set))
2901 if (first_conflict_j < 0)
2902 first_conflict_j = j;
2903 last_conflict_j = j;
2905 if (last_conflict_j < 0)
2906 continue;
2907 /* If earlyclobber operand conflicts with another
2908 non-matching operand which is actually the same register
2909 as the earlyclobber operand, it is better to reload the
2910 another operand as an operand matching the earlyclobber
2911 operand can be also the same. */
2912 if (first_conflict_j == last_conflict_j
2913 && operand_reg[last_conflict_j] != NULL_RTX
2914 && ! curr_alt_match_win[last_conflict_j]
2915 && REGNO (operand_reg[i]) == REGNO (operand_reg[last_conflict_j]))
2917 curr_alt_win[last_conflict_j] = false;
2918 curr_alt_dont_inherit_ops[curr_alt_dont_inherit_ops_num++]
2919 = last_conflict_j;
2920 losers++;
2921 /* Early clobber was already reflected in REJECT. */
2922 lra_assert (reject > 0);
2923 if (lra_dump_file != NULL)
2924 fprintf
2925 (lra_dump_file,
2926 " %d Conflict early clobber reload: reject--\n",
2928 reject--;
2929 overall += LRA_LOSER_COST_FACTOR - 1;
2931 else
2933 /* We need to reload early clobbered register and the
2934 matched registers. */
2935 for (j = 0; j < n_operands; j++)
2936 if (curr_alt_matches[j] == i)
2938 curr_alt_match_win[j] = false;
2939 losers++;
2940 overall += LRA_LOSER_COST_FACTOR;
2942 if (! curr_alt_match_win[i])
2943 curr_alt_dont_inherit_ops[curr_alt_dont_inherit_ops_num++] = i;
2944 else
2946 /* Remember pseudos used for match reloads are never
2947 inherited. */
2948 lra_assert (curr_alt_matches[i] >= 0);
2949 curr_alt_win[curr_alt_matches[i]] = false;
2951 curr_alt_win[i] = curr_alt_match_win[i] = false;
2952 losers++;
2953 /* Early clobber was already reflected in REJECT. */
2954 lra_assert (reject > 0);
2955 if (lra_dump_file != NULL)
2956 fprintf
2957 (lra_dump_file,
2958 " %d Matched conflict early clobber reloads: "
2959 "reject--\n",
2961 reject--;
2962 overall += LRA_LOSER_COST_FACTOR - 1;
2965 if (lra_dump_file != NULL)
2966 fprintf (lra_dump_file, " alt=%d,overall=%d,losers=%d,rld_nregs=%d\n",
2967 nalt, overall, losers, reload_nregs);
2969 /* If this alternative can be made to work by reloading, and it
2970 needs less reloading than the others checked so far, record
2971 it as the chosen goal for reloading. */
2972 if ((best_losers != 0 && losers == 0)
2973 || (((best_losers == 0 && losers == 0)
2974 || (best_losers != 0 && losers != 0))
2975 && (best_overall > overall
2976 || (best_overall == overall
2977 /* If the cost of the reloads is the same,
2978 prefer alternative which requires minimal
2979 number of reload regs. */
2980 && (reload_nregs < best_reload_nregs
2981 || (reload_nregs == best_reload_nregs
2982 && (best_reload_sum < reload_sum
2983 || (best_reload_sum == reload_sum
2984 && nalt < goal_alt_number))))))))
2986 for (nop = 0; nop < n_operands; nop++)
2988 goal_alt_win[nop] = curr_alt_win[nop];
2989 goal_alt_match_win[nop] = curr_alt_match_win[nop];
2990 goal_alt_matches[nop] = curr_alt_matches[nop];
2991 goal_alt[nop] = curr_alt[nop];
2992 goal_alt_offmemok[nop] = curr_alt_offmemok[nop];
2994 goal_alt_dont_inherit_ops_num = curr_alt_dont_inherit_ops_num;
2995 for (nop = 0; nop < curr_alt_dont_inherit_ops_num; nop++)
2996 goal_alt_dont_inherit_ops[nop] = curr_alt_dont_inherit_ops[nop];
2997 goal_alt_swapped = curr_swapped;
2998 best_overall = overall;
2999 best_losers = losers;
3000 best_reload_nregs = reload_nregs;
3001 best_reload_sum = reload_sum;
3002 goal_alt_number = nalt;
3004 if (losers == 0)
3005 /* Everything is satisfied. Do not process alternatives
3006 anymore. */
3007 break;
3008 fail:
3011 return ok_p;
3014 /* Make reload base reg from address AD. */
3015 static rtx
3016 base_to_reg (struct address_info *ad)
3018 enum reg_class cl;
3019 int code = -1;
3020 rtx new_inner = NULL_RTX;
3021 rtx new_reg = NULL_RTX;
3022 rtx_insn *insn;
3023 rtx_insn *last_insn = get_last_insn();
3025 lra_assert (ad->disp == ad->disp_term);
3026 cl = base_reg_class (ad->mode, ad->as, ad->base_outer_code,
3027 get_index_code (ad));
3028 new_reg = lra_create_new_reg (GET_MODE (*ad->base), NULL_RTX,
3029 cl, "base");
3030 new_inner = simplify_gen_binary (PLUS, GET_MODE (new_reg), new_reg,
3031 ad->disp_term == NULL
3032 ? const0_rtx
3033 : *ad->disp_term);
3034 if (!valid_address_p (ad->mode, new_inner, ad->as))
3035 return NULL_RTX;
3036 insn = emit_insn (gen_rtx_SET (new_reg, *ad->base));
3037 code = recog_memoized (insn);
3038 if (code < 0)
3040 delete_insns_since (last_insn);
3041 return NULL_RTX;
3044 return new_inner;
3047 /* Make reload base reg + disp from address AD. Return the new pseudo. */
3048 static rtx
3049 base_plus_disp_to_reg (struct address_info *ad)
3051 enum reg_class cl;
3052 rtx new_reg;
3054 lra_assert (ad->base == ad->base_term && ad->disp == ad->disp_term);
3055 cl = base_reg_class (ad->mode, ad->as, ad->base_outer_code,
3056 get_index_code (ad));
3057 new_reg = lra_create_new_reg (GET_MODE (*ad->base_term), NULL_RTX,
3058 cl, "base + disp");
3059 lra_emit_add (new_reg, *ad->base_term, *ad->disp_term);
3060 return new_reg;
3063 /* Make reload of index part of address AD. Return the new
3064 pseudo. */
3065 static rtx
3066 index_part_to_reg (struct address_info *ad)
3068 rtx new_reg;
3070 new_reg = lra_create_new_reg (GET_MODE (*ad->index), NULL_RTX,
3071 INDEX_REG_CLASS, "index term");
3072 expand_mult (GET_MODE (*ad->index), *ad->index_term,
3073 GEN_INT (get_index_scale (ad)), new_reg, 1);
3074 return new_reg;
3077 /* Return true if we can add a displacement to address AD, even if that
3078 makes the address invalid. The fix-up code requires any new address
3079 to be the sum of the BASE_TERM, INDEX and DISP_TERM fields. */
3080 static bool
3081 can_add_disp_p (struct address_info *ad)
3083 return (!ad->autoinc_p
3084 && ad->segment == NULL
3085 && ad->base == ad->base_term
3086 && ad->disp == ad->disp_term);
3089 /* Make equiv substitution in address AD. Return true if a substitution
3090 was made. */
3091 static bool
3092 equiv_address_substitution (struct address_info *ad)
3094 rtx base_reg, new_base_reg, index_reg, new_index_reg, *base_term, *index_term;
3095 HOST_WIDE_INT disp, scale;
3096 bool change_p;
3098 base_term = strip_subreg (ad->base_term);
3099 if (base_term == NULL)
3100 base_reg = new_base_reg = NULL_RTX;
3101 else
3103 base_reg = *base_term;
3104 new_base_reg = get_equiv_with_elimination (base_reg, curr_insn);
3106 index_term = strip_subreg (ad->index_term);
3107 if (index_term == NULL)
3108 index_reg = new_index_reg = NULL_RTX;
3109 else
3111 index_reg = *index_term;
3112 new_index_reg = get_equiv_with_elimination (index_reg, curr_insn);
3114 if (base_reg == new_base_reg && index_reg == new_index_reg)
3115 return false;
3116 disp = 0;
3117 change_p = false;
3118 if (lra_dump_file != NULL)
3120 fprintf (lra_dump_file, "Changing address in insn %d ",
3121 INSN_UID (curr_insn));
3122 dump_value_slim (lra_dump_file, *ad->outer, 1);
3124 if (base_reg != new_base_reg)
3126 if (REG_P (new_base_reg))
3128 *base_term = new_base_reg;
3129 change_p = true;
3131 else if (GET_CODE (new_base_reg) == PLUS
3132 && REG_P (XEXP (new_base_reg, 0))
3133 && CONST_INT_P (XEXP (new_base_reg, 1))
3134 && can_add_disp_p (ad))
3136 disp += INTVAL (XEXP (new_base_reg, 1));
3137 *base_term = XEXP (new_base_reg, 0);
3138 change_p = true;
3140 if (ad->base_term2 != NULL)
3141 *ad->base_term2 = *ad->base_term;
3143 if (index_reg != new_index_reg)
3145 if (REG_P (new_index_reg))
3147 *index_term = new_index_reg;
3148 change_p = true;
3150 else if (GET_CODE (new_index_reg) == PLUS
3151 && REG_P (XEXP (new_index_reg, 0))
3152 && CONST_INT_P (XEXP (new_index_reg, 1))
3153 && can_add_disp_p (ad)
3154 && (scale = get_index_scale (ad)))
3156 disp += INTVAL (XEXP (new_index_reg, 1)) * scale;
3157 *index_term = XEXP (new_index_reg, 0);
3158 change_p = true;
3161 if (disp != 0)
3163 if (ad->disp != NULL)
3164 *ad->disp = plus_constant (GET_MODE (*ad->inner), *ad->disp, disp);
3165 else
3167 *ad->inner = plus_constant (GET_MODE (*ad->inner), *ad->inner, disp);
3168 update_address (ad);
3170 change_p = true;
3172 if (lra_dump_file != NULL)
3174 if (! change_p)
3175 fprintf (lra_dump_file, " -- no change\n");
3176 else
3178 fprintf (lra_dump_file, " on equiv ");
3179 dump_value_slim (lra_dump_file, *ad->outer, 1);
3180 fprintf (lra_dump_file, "\n");
3183 return change_p;
3186 /* Major function to make reloads for an address in operand NOP or
3187 check its correctness (If CHECK_ONLY_P is true). The supported
3188 cases are:
3190 1) an address that existed before LRA started, at which point it
3191 must have been valid. These addresses are subject to elimination
3192 and may have become invalid due to the elimination offset being out
3193 of range.
3195 2) an address created by forcing a constant to memory
3196 (force_const_to_mem). The initial form of these addresses might
3197 not be valid, and it is this function's job to make them valid.
3199 3) a frame address formed from a register and a (possibly zero)
3200 constant offset. As above, these addresses might not be valid and
3201 this function must make them so.
3203 Add reloads to the lists *BEFORE and *AFTER. We might need to add
3204 reloads to *AFTER because of inc/dec, {pre, post} modify in the
3205 address. Return true for any RTL change.
3207 The function is a helper function which does not produce all
3208 transformations (when CHECK_ONLY_P is false) which can be
3209 necessary. It does just basic steps. To do all necessary
3210 transformations use function process_address. */
3211 static bool
3212 process_address_1 (int nop, bool check_only_p,
3213 rtx_insn **before, rtx_insn **after)
3215 struct address_info ad;
3216 rtx new_reg;
3217 HOST_WIDE_INT scale;
3218 rtx op = *curr_id->operand_loc[nop];
3219 const char *constraint = curr_static_id->operand[nop].constraint;
3220 enum constraint_num cn = lookup_constraint (constraint);
3221 bool change_p = false;
3223 if (MEM_P (op)
3224 && GET_MODE (op) == BLKmode
3225 && GET_CODE (XEXP (op, 0)) == SCRATCH)
3226 return false;
3228 if (insn_extra_address_constraint (cn))
3229 decompose_lea_address (&ad, curr_id->operand_loc[nop]);
3230 /* Do not attempt to decompose arbitrary addresses generated by combine
3231 for asm operands with loose constraints, e.g 'X'. */
3232 else if (MEM_P (op)
3233 && !(get_constraint_type (cn) == CT_FIXED_FORM
3234 && constraint_satisfied_p (op, cn)))
3235 decompose_mem_address (&ad, op);
3236 else if (GET_CODE (op) == SUBREG
3237 && MEM_P (SUBREG_REG (op)))
3238 decompose_mem_address (&ad, SUBREG_REG (op));
3239 else
3240 return false;
3241 /* If INDEX_REG_CLASS is assigned to base_term already and isn't to
3242 index_term, swap them so to avoid assigning INDEX_REG_CLASS to both
3243 when INDEX_REG_CLASS is a single register class. */
3244 if (ad.base_term != NULL
3245 && ad.index_term != NULL
3246 && ira_class_hard_regs_num[INDEX_REG_CLASS] == 1
3247 && REG_P (*ad.base_term)
3248 && REG_P (*ad.index_term)
3249 && in_class_p (*ad.base_term, INDEX_REG_CLASS, NULL)
3250 && ! in_class_p (*ad.index_term, INDEX_REG_CLASS, NULL))
3252 std::swap (ad.base, ad.index);
3253 std::swap (ad.base_term, ad.index_term);
3255 if (! check_only_p)
3256 change_p = equiv_address_substitution (&ad);
3257 if (ad.base_term != NULL
3258 && (process_addr_reg
3259 (ad.base_term, check_only_p, before,
3260 (ad.autoinc_p
3261 && !(REG_P (*ad.base_term)
3262 && find_regno_note (curr_insn, REG_DEAD,
3263 REGNO (*ad.base_term)) != NULL_RTX)
3264 ? after : NULL),
3265 base_reg_class (ad.mode, ad.as, ad.base_outer_code,
3266 get_index_code (&ad)))))
3268 change_p = true;
3269 if (ad.base_term2 != NULL)
3270 *ad.base_term2 = *ad.base_term;
3272 if (ad.index_term != NULL
3273 && process_addr_reg (ad.index_term, check_only_p,
3274 before, NULL, INDEX_REG_CLASS))
3275 change_p = true;
3277 /* Target hooks sometimes don't treat extra-constraint addresses as
3278 legitimate address_operands, so handle them specially. */
3279 if (insn_extra_address_constraint (cn)
3280 && satisfies_address_constraint_p (&ad, cn))
3281 return change_p;
3283 if (check_only_p)
3284 return change_p;
3286 /* There are three cases where the shape of *AD.INNER may now be invalid:
3288 1) the original address was valid, but either elimination or
3289 equiv_address_substitution was applied and that made
3290 the address invalid.
3292 2) the address is an invalid symbolic address created by
3293 force_const_to_mem.
3295 3) the address is a frame address with an invalid offset.
3297 4) the address is a frame address with an invalid base.
3299 All these cases involve a non-autoinc address, so there is no
3300 point revalidating other types. */
3301 if (ad.autoinc_p || valid_address_p (&ad))
3302 return change_p;
3304 /* Any index existed before LRA started, so we can assume that the
3305 presence and shape of the index is valid. */
3306 push_to_sequence (*before);
3307 lra_assert (ad.disp == ad.disp_term);
3308 if (ad.base == NULL)
3310 if (ad.index == NULL)
3312 rtx_insn *insn;
3313 rtx_insn *last = get_last_insn ();
3314 int code = -1;
3315 enum reg_class cl = base_reg_class (ad.mode, ad.as,
3316 SCRATCH, SCRATCH);
3317 rtx addr = *ad.inner;
3319 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "addr");
3320 if (HAVE_lo_sum)
3322 /* addr => lo_sum (new_base, addr), case (2) above. */
3323 insn = emit_insn (gen_rtx_SET
3324 (new_reg,
3325 gen_rtx_HIGH (Pmode, copy_rtx (addr))));
3326 code = recog_memoized (insn);
3327 if (code >= 0)
3329 *ad.inner = gen_rtx_LO_SUM (Pmode, new_reg, addr);
3330 if (! valid_address_p (ad.mode, *ad.outer, ad.as))
3332 /* Try to put lo_sum into register. */
3333 insn = emit_insn (gen_rtx_SET
3334 (new_reg,
3335 gen_rtx_LO_SUM (Pmode, new_reg, addr)));
3336 code = recog_memoized (insn);
3337 if (code >= 0)
3339 *ad.inner = new_reg;
3340 if (! valid_address_p (ad.mode, *ad.outer, ad.as))
3342 *ad.inner = addr;
3343 code = -1;
3349 if (code < 0)
3350 delete_insns_since (last);
3353 if (code < 0)
3355 /* addr => new_base, case (2) above. */
3356 lra_emit_move (new_reg, addr);
3358 for (insn = last == NULL_RTX ? get_insns () : NEXT_INSN (last);
3359 insn != NULL_RTX;
3360 insn = NEXT_INSN (insn))
3361 if (recog_memoized (insn) < 0)
3362 break;
3363 if (insn != NULL_RTX)
3365 /* Do nothing if we cannot generate right insns.
3366 This is analogous to reload pass behavior. */
3367 delete_insns_since (last);
3368 end_sequence ();
3369 return false;
3371 *ad.inner = new_reg;
3374 else
3376 /* index * scale + disp => new base + index * scale,
3377 case (1) above. */
3378 enum reg_class cl = base_reg_class (ad.mode, ad.as, PLUS,
3379 GET_CODE (*ad.index));
3381 lra_assert (INDEX_REG_CLASS != NO_REGS);
3382 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "disp");
3383 lra_emit_move (new_reg, *ad.disp);
3384 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3385 new_reg, *ad.index);
3388 else if (ad.index == NULL)
3390 int regno;
3391 enum reg_class cl;
3392 rtx set;
3393 rtx_insn *insns, *last_insn;
3394 /* Try to reload base into register only if the base is invalid
3395 for the address but with valid offset, case (4) above. */
3396 start_sequence ();
3397 new_reg = base_to_reg (&ad);
3399 /* base + disp => new base, cases (1) and (3) above. */
3400 /* Another option would be to reload the displacement into an
3401 index register. However, postreload has code to optimize
3402 address reloads that have the same base and different
3403 displacements, so reloading into an index register would
3404 not necessarily be a win. */
3405 if (new_reg == NULL_RTX)
3406 new_reg = base_plus_disp_to_reg (&ad);
3407 insns = get_insns ();
3408 last_insn = get_last_insn ();
3409 /* If we generated at least two insns, try last insn source as
3410 an address. If we succeed, we generate one less insn. */
3411 if (last_insn != insns && (set = single_set (last_insn)) != NULL_RTX
3412 && GET_CODE (SET_SRC (set)) == PLUS
3413 && REG_P (XEXP (SET_SRC (set), 0))
3414 && CONSTANT_P (XEXP (SET_SRC (set), 1)))
3416 *ad.inner = SET_SRC (set);
3417 if (valid_address_p (ad.mode, *ad.outer, ad.as))
3419 *ad.base_term = XEXP (SET_SRC (set), 0);
3420 *ad.disp_term = XEXP (SET_SRC (set), 1);
3421 cl = base_reg_class (ad.mode, ad.as, ad.base_outer_code,
3422 get_index_code (&ad));
3423 regno = REGNO (*ad.base_term);
3424 if (regno >= FIRST_PSEUDO_REGISTER
3425 && cl != lra_get_allocno_class (regno))
3426 lra_change_class (regno, cl, " Change to", true);
3427 new_reg = SET_SRC (set);
3428 delete_insns_since (PREV_INSN (last_insn));
3431 /* Try if target can split displacement into legitimite new disp
3432 and offset. If it's the case, we replace the last insn with
3433 insns for base + offset => new_reg and set new_reg + new disp
3434 to *ad.inner. */
3435 last_insn = get_last_insn ();
3436 if ((set = single_set (last_insn)) != NULL_RTX
3437 && GET_CODE (SET_SRC (set)) == PLUS
3438 && REG_P (XEXP (SET_SRC (set), 0))
3439 && REGNO (XEXP (SET_SRC (set), 0)) < FIRST_PSEUDO_REGISTER
3440 && CONST_INT_P (XEXP (SET_SRC (set), 1)))
3442 rtx addend, disp = XEXP (SET_SRC (set), 1);
3443 if (targetm.legitimize_address_displacement (&disp, &addend,
3444 ad.mode))
3446 rtx_insn *new_insns;
3447 start_sequence ();
3448 lra_emit_add (new_reg, XEXP (SET_SRC (set), 0), addend);
3449 new_insns = get_insns ();
3450 end_sequence ();
3451 new_reg = gen_rtx_PLUS (Pmode, new_reg, disp);
3452 delete_insns_since (PREV_INSN (last_insn));
3453 add_insn (new_insns);
3454 insns = get_insns ();
3457 end_sequence ();
3458 emit_insn (insns);
3459 *ad.inner = new_reg;
3461 else if (ad.disp_term != NULL)
3463 /* base + scale * index + disp => new base + scale * index,
3464 case (1) above. */
3465 new_reg = base_plus_disp_to_reg (&ad);
3466 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3467 new_reg, *ad.index);
3469 else if ((scale = get_index_scale (&ad)) == 1)
3471 /* The last transformation to one reg will be made in
3472 curr_insn_transform function. */
3473 end_sequence ();
3474 return false;
3476 else if (scale != 0)
3478 /* base + scale * index => base + new_reg,
3479 case (1) above.
3480 Index part of address may become invalid. For example, we
3481 changed pseudo on the equivalent memory and a subreg of the
3482 pseudo onto the memory of different mode for which the scale is
3483 prohibitted. */
3484 new_reg = index_part_to_reg (&ad);
3485 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3486 *ad.base_term, new_reg);
3488 else
3490 enum reg_class cl = base_reg_class (ad.mode, ad.as,
3491 SCRATCH, SCRATCH);
3492 rtx addr = *ad.inner;
3494 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "addr");
3495 /* addr => new_base. */
3496 lra_emit_move (new_reg, addr);
3497 *ad.inner = new_reg;
3499 *before = get_insns ();
3500 end_sequence ();
3501 return true;
3504 /* If CHECK_ONLY_P is false, do address reloads until it is necessary.
3505 Use process_address_1 as a helper function. Return true for any
3506 RTL changes.
3508 If CHECK_ONLY_P is true, just check address correctness. Return
3509 false if the address correct. */
3510 static bool
3511 process_address (int nop, bool check_only_p,
3512 rtx_insn **before, rtx_insn **after)
3514 bool res = false;
3516 while (process_address_1 (nop, check_only_p, before, after))
3518 if (check_only_p)
3519 return true;
3520 res = true;
3522 return res;
3525 /* Emit insns to reload VALUE into a new register. VALUE is an
3526 auto-increment or auto-decrement RTX whose operand is a register or
3527 memory location; so reloading involves incrementing that location.
3528 IN is either identical to VALUE, or some cheaper place to reload
3529 value being incremented/decremented from.
3531 INC_AMOUNT is the number to increment or decrement by (always
3532 positive and ignored for POST_MODIFY/PRE_MODIFY).
3534 Return pseudo containing the result. */
3535 static rtx
3536 emit_inc (enum reg_class new_rclass, rtx in, rtx value, int inc_amount)
3538 /* REG or MEM to be copied and incremented. */
3539 rtx incloc = XEXP (value, 0);
3540 /* Nonzero if increment after copying. */
3541 int post = (GET_CODE (value) == POST_DEC || GET_CODE (value) == POST_INC
3542 || GET_CODE (value) == POST_MODIFY);
3543 rtx_insn *last;
3544 rtx inc;
3545 rtx_insn *add_insn;
3546 int code;
3547 rtx real_in = in == value ? incloc : in;
3548 rtx result;
3549 bool plus_p = true;
3551 if (GET_CODE (value) == PRE_MODIFY || GET_CODE (value) == POST_MODIFY)
3553 lra_assert (GET_CODE (XEXP (value, 1)) == PLUS
3554 || GET_CODE (XEXP (value, 1)) == MINUS);
3555 lra_assert (rtx_equal_p (XEXP (XEXP (value, 1), 0), XEXP (value, 0)));
3556 plus_p = GET_CODE (XEXP (value, 1)) == PLUS;
3557 inc = XEXP (XEXP (value, 1), 1);
3559 else
3561 if (GET_CODE (value) == PRE_DEC || GET_CODE (value) == POST_DEC)
3562 inc_amount = -inc_amount;
3564 inc = GEN_INT (inc_amount);
3567 if (! post && REG_P (incloc))
3568 result = incloc;
3569 else
3570 result = lra_create_new_reg (GET_MODE (value), value, new_rclass,
3571 "INC/DEC result");
3573 if (real_in != result)
3575 /* First copy the location to the result register. */
3576 lra_assert (REG_P (result));
3577 emit_insn (gen_move_insn (result, real_in));
3580 /* We suppose that there are insns to add/sub with the constant
3581 increment permitted in {PRE/POST)_{DEC/INC/MODIFY}. At least the
3582 old reload worked with this assumption. If the assumption
3583 becomes wrong, we should use approach in function
3584 base_plus_disp_to_reg. */
3585 if (in == value)
3587 /* See if we can directly increment INCLOC. */
3588 last = get_last_insn ();
3589 add_insn = emit_insn (plus_p
3590 ? gen_add2_insn (incloc, inc)
3591 : gen_sub2_insn (incloc, inc));
3593 code = recog_memoized (add_insn);
3594 if (code >= 0)
3596 if (! post && result != incloc)
3597 emit_insn (gen_move_insn (result, incloc));
3598 return result;
3600 delete_insns_since (last);
3603 /* If couldn't do the increment directly, must increment in RESULT.
3604 The way we do this depends on whether this is pre- or
3605 post-increment. For pre-increment, copy INCLOC to the reload
3606 register, increment it there, then save back. */
3607 if (! post)
3609 if (real_in != result)
3610 emit_insn (gen_move_insn (result, real_in));
3611 if (plus_p)
3612 emit_insn (gen_add2_insn (result, inc));
3613 else
3614 emit_insn (gen_sub2_insn (result, inc));
3615 if (result != incloc)
3616 emit_insn (gen_move_insn (incloc, result));
3618 else
3620 /* Post-increment.
3622 Because this might be a jump insn or a compare, and because
3623 RESULT may not be available after the insn in an input
3624 reload, we must do the incrementing before the insn being
3625 reloaded for.
3627 We have already copied IN to RESULT. Increment the copy in
3628 RESULT, save that back, then decrement RESULT so it has
3629 the original value. */
3630 if (plus_p)
3631 emit_insn (gen_add2_insn (result, inc));
3632 else
3633 emit_insn (gen_sub2_insn (result, inc));
3634 emit_insn (gen_move_insn (incloc, result));
3635 /* Restore non-modified value for the result. We prefer this
3636 way because it does not require an additional hard
3637 register. */
3638 if (plus_p)
3640 if (CONST_INT_P (inc))
3641 emit_insn (gen_add2_insn (result,
3642 gen_int_mode (-INTVAL (inc),
3643 GET_MODE (result))));
3644 else
3645 emit_insn (gen_sub2_insn (result, inc));
3647 else
3648 emit_insn (gen_add2_insn (result, inc));
3650 return result;
3653 /* Return true if the current move insn does not need processing as we
3654 already know that it satisfies its constraints. */
3655 static bool
3656 simple_move_p (void)
3658 rtx dest, src;
3659 enum reg_class dclass, sclass;
3661 lra_assert (curr_insn_set != NULL_RTX);
3662 dest = SET_DEST (curr_insn_set);
3663 src = SET_SRC (curr_insn_set);
3665 /* If the instruction has multiple sets we need to process it even if it
3666 is single_set. This can happen if one or more of the SETs are dead.
3667 See PR73650. */
3668 if (multiple_sets (curr_insn))
3669 return false;
3671 return ((dclass = get_op_class (dest)) != NO_REGS
3672 && (sclass = get_op_class (src)) != NO_REGS
3673 /* The backend guarantees that register moves of cost 2
3674 never need reloads. */
3675 && targetm.register_move_cost (GET_MODE (src), sclass, dclass) == 2);
3678 /* Swap operands NOP and NOP + 1. */
3679 static inline void
3680 swap_operands (int nop)
3682 std::swap (curr_operand_mode[nop], curr_operand_mode[nop + 1]);
3683 std::swap (original_subreg_reg_mode[nop], original_subreg_reg_mode[nop + 1]);
3684 std::swap (*curr_id->operand_loc[nop], *curr_id->operand_loc[nop + 1]);
3685 std::swap (equiv_substition_p[nop], equiv_substition_p[nop + 1]);
3686 /* Swap the duplicates too. */
3687 lra_update_dup (curr_id, nop);
3688 lra_update_dup (curr_id, nop + 1);
3691 /* Main entry point of the constraint code: search the body of the
3692 current insn to choose the best alternative. It is mimicking insn
3693 alternative cost calculation model of former reload pass. That is
3694 because machine descriptions were written to use this model. This
3695 model can be changed in future. Make commutative operand exchange
3696 if it is chosen.
3698 if CHECK_ONLY_P is false, do RTL changes to satisfy the
3699 constraints. Return true if any change happened during function
3700 call.
3702 If CHECK_ONLY_P is true then don't do any transformation. Just
3703 check that the insn satisfies all constraints. If the insn does
3704 not satisfy any constraint, return true. */
3705 static bool
3706 curr_insn_transform (bool check_only_p)
3708 int i, j, k;
3709 int n_operands;
3710 int n_alternatives;
3711 int n_outputs;
3712 int commutative;
3713 signed char goal_alt_matched[MAX_RECOG_OPERANDS][MAX_RECOG_OPERANDS];
3714 signed char match_inputs[MAX_RECOG_OPERANDS + 1];
3715 signed char outputs[MAX_RECOG_OPERANDS + 1];
3716 rtx_insn *before, *after;
3717 bool alt_p = false;
3718 /* Flag that the insn has been changed through a transformation. */
3719 bool change_p;
3720 bool sec_mem_p;
3721 #ifdef SECONDARY_MEMORY_NEEDED
3722 bool use_sec_mem_p;
3723 #endif
3724 int max_regno_before;
3725 int reused_alternative_num;
3727 curr_insn_set = single_set (curr_insn);
3728 if (curr_insn_set != NULL_RTX && simple_move_p ())
3729 return false;
3731 no_input_reloads_p = no_output_reloads_p = false;
3732 goal_alt_number = -1;
3733 change_p = sec_mem_p = false;
3734 /* JUMP_INSNs and CALL_INSNs are not allowed to have any output
3735 reloads; neither are insns that SET cc0. Insns that use CC0 are
3736 not allowed to have any input reloads. */
3737 if (JUMP_P (curr_insn) || CALL_P (curr_insn))
3738 no_output_reloads_p = true;
3740 if (HAVE_cc0 && reg_referenced_p (cc0_rtx, PATTERN (curr_insn)))
3741 no_input_reloads_p = true;
3742 if (HAVE_cc0 && reg_set_p (cc0_rtx, PATTERN (curr_insn)))
3743 no_output_reloads_p = true;
3745 n_operands = curr_static_id->n_operands;
3746 n_alternatives = curr_static_id->n_alternatives;
3748 /* Just return "no reloads" if insn has no operands with
3749 constraints. */
3750 if (n_operands == 0 || n_alternatives == 0)
3751 return false;
3753 max_regno_before = max_reg_num ();
3755 for (i = 0; i < n_operands; i++)
3757 goal_alt_matched[i][0] = -1;
3758 goal_alt_matches[i] = -1;
3761 commutative = curr_static_id->commutative;
3763 /* Now see what we need for pseudos that didn't get hard regs or got
3764 the wrong kind of hard reg. For this, we must consider all the
3765 operands together against the register constraints. */
3767 best_losers = best_overall = INT_MAX;
3768 best_reload_sum = 0;
3770 curr_swapped = false;
3771 goal_alt_swapped = false;
3773 if (! check_only_p)
3774 /* Make equivalence substitution and memory subreg elimination
3775 before address processing because an address legitimacy can
3776 depend on memory mode. */
3777 for (i = 0; i < n_operands; i++)
3779 rtx op, subst, old;
3780 bool op_change_p = false;
3782 if (curr_static_id->operand[i].is_operator)
3783 continue;
3785 old = op = *curr_id->operand_loc[i];
3786 if (GET_CODE (old) == SUBREG)
3787 old = SUBREG_REG (old);
3788 subst = get_equiv_with_elimination (old, curr_insn);
3789 original_subreg_reg_mode[i] = VOIDmode;
3790 equiv_substition_p[i] = false;
3791 if (subst != old)
3793 equiv_substition_p[i] = true;
3794 subst = copy_rtx (subst);
3795 lra_assert (REG_P (old));
3796 if (GET_CODE (op) != SUBREG)
3797 *curr_id->operand_loc[i] = subst;
3798 else
3800 SUBREG_REG (op) = subst;
3801 if (GET_MODE (subst) == VOIDmode)
3802 original_subreg_reg_mode[i] = GET_MODE (old);
3804 if (lra_dump_file != NULL)
3806 fprintf (lra_dump_file,
3807 "Changing pseudo %d in operand %i of insn %u on equiv ",
3808 REGNO (old), i, INSN_UID (curr_insn));
3809 dump_value_slim (lra_dump_file, subst, 1);
3810 fprintf (lra_dump_file, "\n");
3812 op_change_p = change_p = true;
3814 if (simplify_operand_subreg (i, GET_MODE (old)) || op_change_p)
3816 change_p = true;
3817 lra_update_dup (curr_id, i);
3821 /* Reload address registers and displacements. We do it before
3822 finding an alternative because of memory constraints. */
3823 before = after = NULL;
3824 for (i = 0; i < n_operands; i++)
3825 if (! curr_static_id->operand[i].is_operator
3826 && process_address (i, check_only_p, &before, &after))
3828 if (check_only_p)
3829 return true;
3830 change_p = true;
3831 lra_update_dup (curr_id, i);
3834 if (change_p)
3835 /* If we've changed the instruction then any alternative that
3836 we chose previously may no longer be valid. */
3837 lra_set_used_insn_alternative (curr_insn, -1);
3839 if (! check_only_p && curr_insn_set != NULL_RTX
3840 && check_and_process_move (&change_p, &sec_mem_p))
3841 return change_p;
3843 try_swapped:
3845 reused_alternative_num = check_only_p ? -1 : curr_id->used_insn_alternative;
3846 if (lra_dump_file != NULL && reused_alternative_num >= 0)
3847 fprintf (lra_dump_file, "Reusing alternative %d for insn #%u\n",
3848 reused_alternative_num, INSN_UID (curr_insn));
3850 if (process_alt_operands (reused_alternative_num))
3851 alt_p = true;
3853 if (check_only_p)
3854 return ! alt_p || best_losers != 0;
3856 /* If insn is commutative (it's safe to exchange a certain pair of
3857 operands) then we need to try each alternative twice, the second
3858 time matching those two operands as if we had exchanged them. To
3859 do this, really exchange them in operands.
3861 If we have just tried the alternatives the second time, return
3862 operands to normal and drop through. */
3864 if (reused_alternative_num < 0 && commutative >= 0)
3866 curr_swapped = !curr_swapped;
3867 if (curr_swapped)
3869 swap_operands (commutative);
3870 goto try_swapped;
3872 else
3873 swap_operands (commutative);
3876 if (! alt_p && ! sec_mem_p)
3878 /* No alternative works with reloads?? */
3879 if (INSN_CODE (curr_insn) >= 0)
3880 fatal_insn ("unable to generate reloads for:", curr_insn);
3881 error_for_asm (curr_insn,
3882 "inconsistent operand constraints in an %<asm%>");
3883 /* Avoid further trouble with this insn. Don't generate use
3884 pattern here as we could use the insn SP offset. */
3885 lra_set_insn_deleted (curr_insn);
3886 return true;
3889 /* If the best alternative is with operands 1 and 2 swapped, swap
3890 them. Update the operand numbers of any reloads already
3891 pushed. */
3893 if (goal_alt_swapped)
3895 if (lra_dump_file != NULL)
3896 fprintf (lra_dump_file, " Commutative operand exchange in insn %u\n",
3897 INSN_UID (curr_insn));
3899 /* Swap the duplicates too. */
3900 swap_operands (commutative);
3901 change_p = true;
3904 #ifdef SECONDARY_MEMORY_NEEDED
3905 /* Some target macros SECONDARY_MEMORY_NEEDED (e.g. x86) are defined
3906 too conservatively. So we use the secondary memory only if there
3907 is no any alternative without reloads. */
3908 use_sec_mem_p = false;
3909 if (! alt_p)
3910 use_sec_mem_p = true;
3911 else if (sec_mem_p)
3913 for (i = 0; i < n_operands; i++)
3914 if (! goal_alt_win[i] && ! goal_alt_match_win[i])
3915 break;
3916 use_sec_mem_p = i < n_operands;
3919 if (use_sec_mem_p)
3921 int in = -1, out = -1;
3922 rtx new_reg, src, dest, rld;
3923 machine_mode sec_mode, rld_mode;
3925 lra_assert (curr_insn_set != NULL_RTX && sec_mem_p);
3926 dest = SET_DEST (curr_insn_set);
3927 src = SET_SRC (curr_insn_set);
3928 for (i = 0; i < n_operands; i++)
3929 if (*curr_id->operand_loc[i] == dest)
3930 out = i;
3931 else if (*curr_id->operand_loc[i] == src)
3932 in = i;
3933 for (i = 0; i < curr_static_id->n_dups; i++)
3934 if (out < 0 && *curr_id->dup_loc[i] == dest)
3935 out = curr_static_id->dup_num[i];
3936 else if (in < 0 && *curr_id->dup_loc[i] == src)
3937 in = curr_static_id->dup_num[i];
3938 lra_assert (out >= 0 && in >= 0
3939 && curr_static_id->operand[out].type == OP_OUT
3940 && curr_static_id->operand[in].type == OP_IN);
3941 rld = partial_subreg_p (GET_MODE (src), GET_MODE (dest)) ? src : dest;
3942 rld_mode = GET_MODE (rld);
3943 #ifdef SECONDARY_MEMORY_NEEDED_MODE
3944 sec_mode = SECONDARY_MEMORY_NEEDED_MODE (rld_mode);
3945 #else
3946 sec_mode = rld_mode;
3947 #endif
3948 new_reg = lra_create_new_reg (sec_mode, NULL_RTX,
3949 NO_REGS, "secondary");
3950 /* If the mode is changed, it should be wider. */
3951 lra_assert (!partial_subreg_p (sec_mode, rld_mode));
3952 if (sec_mode != rld_mode)
3954 /* If the target says specifically to use another mode for
3955 secondary memory moves we can not reuse the original
3956 insn. */
3957 after = emit_spill_move (false, new_reg, dest);
3958 lra_process_new_insns (curr_insn, NULL, after,
3959 "Inserting the sec. move");
3960 /* We may have non null BEFORE here (e.g. after address
3961 processing. */
3962 push_to_sequence (before);
3963 before = emit_spill_move (true, new_reg, src);
3964 emit_insn (before);
3965 before = get_insns ();
3966 end_sequence ();
3967 lra_process_new_insns (curr_insn, before, NULL, "Changing on");
3968 lra_set_insn_deleted (curr_insn);
3970 else if (dest == rld)
3972 *curr_id->operand_loc[out] = new_reg;
3973 lra_update_dup (curr_id, out);
3974 after = emit_spill_move (false, new_reg, dest);
3975 lra_process_new_insns (curr_insn, NULL, after,
3976 "Inserting the sec. move");
3978 else
3980 *curr_id->operand_loc[in] = new_reg;
3981 lra_update_dup (curr_id, in);
3982 /* See comments above. */
3983 push_to_sequence (before);
3984 before = emit_spill_move (true, new_reg, src);
3985 emit_insn (before);
3986 before = get_insns ();
3987 end_sequence ();
3988 lra_process_new_insns (curr_insn, before, NULL,
3989 "Inserting the sec. move");
3991 lra_update_insn_regno_info (curr_insn);
3992 return true;
3994 #endif
3996 lra_assert (goal_alt_number >= 0);
3997 lra_set_used_insn_alternative (curr_insn, goal_alt_number);
3999 if (lra_dump_file != NULL)
4001 const char *p;
4003 fprintf (lra_dump_file, " Choosing alt %d in insn %u:",
4004 goal_alt_number, INSN_UID (curr_insn));
4005 for (i = 0; i < n_operands; i++)
4007 p = (curr_static_id->operand_alternative
4008 [goal_alt_number * n_operands + i].constraint);
4009 if (*p == '\0')
4010 continue;
4011 fprintf (lra_dump_file, " (%d) ", i);
4012 for (; *p != '\0' && *p != ',' && *p != '#'; p++)
4013 fputc (*p, lra_dump_file);
4015 if (INSN_CODE (curr_insn) >= 0
4016 && (p = get_insn_name (INSN_CODE (curr_insn))) != NULL)
4017 fprintf (lra_dump_file, " {%s}", p);
4018 if (curr_id->sp_offset != 0)
4019 fprintf (lra_dump_file, " (sp_off=%" HOST_WIDE_INT_PRINT "d)",
4020 curr_id->sp_offset);
4021 fprintf (lra_dump_file, "\n");
4024 /* Right now, for any pair of operands I and J that are required to
4025 match, with J < I, goal_alt_matches[I] is J. Add I to
4026 goal_alt_matched[J]. */
4028 for (i = 0; i < n_operands; i++)
4029 if ((j = goal_alt_matches[i]) >= 0)
4031 for (k = 0; goal_alt_matched[j][k] >= 0; k++)
4033 /* We allow matching one output operand and several input
4034 operands. */
4035 lra_assert (k == 0
4036 || (curr_static_id->operand[j].type == OP_OUT
4037 && curr_static_id->operand[i].type == OP_IN
4038 && (curr_static_id->operand
4039 [goal_alt_matched[j][0]].type == OP_IN)));
4040 goal_alt_matched[j][k] = i;
4041 goal_alt_matched[j][k + 1] = -1;
4044 for (i = 0; i < n_operands; i++)
4045 goal_alt_win[i] |= goal_alt_match_win[i];
4047 /* Any constants that aren't allowed and can't be reloaded into
4048 registers are here changed into memory references. */
4049 for (i = 0; i < n_operands; i++)
4050 if (goal_alt_win[i])
4052 int regno;
4053 enum reg_class new_class;
4054 rtx reg = *curr_id->operand_loc[i];
4056 if (GET_CODE (reg) == SUBREG)
4057 reg = SUBREG_REG (reg);
4059 if (REG_P (reg) && (regno = REGNO (reg)) >= FIRST_PSEUDO_REGISTER)
4061 bool ok_p = in_class_p (reg, goal_alt[i], &new_class);
4063 if (new_class != NO_REGS && get_reg_class (regno) != new_class)
4065 lra_assert (ok_p);
4066 lra_change_class (regno, new_class, " Change to", true);
4070 else
4072 const char *constraint;
4073 char c;
4074 rtx op = *curr_id->operand_loc[i];
4075 rtx subreg = NULL_RTX;
4076 machine_mode mode = curr_operand_mode[i];
4078 if (GET_CODE (op) == SUBREG)
4080 subreg = op;
4081 op = SUBREG_REG (op);
4082 mode = GET_MODE (op);
4085 if (CONST_POOL_OK_P (mode, op)
4086 && ((targetm.preferred_reload_class
4087 (op, (enum reg_class) goal_alt[i]) == NO_REGS)
4088 || no_input_reloads_p))
4090 rtx tem = force_const_mem (mode, op);
4092 change_p = true;
4093 if (subreg != NULL_RTX)
4094 tem = gen_rtx_SUBREG (mode, tem, SUBREG_BYTE (subreg));
4096 *curr_id->operand_loc[i] = tem;
4097 lra_update_dup (curr_id, i);
4098 process_address (i, false, &before, &after);
4100 /* If the alternative accepts constant pool refs directly
4101 there will be no reload needed at all. */
4102 if (subreg != NULL_RTX)
4103 continue;
4104 /* Skip alternatives before the one requested. */
4105 constraint = (curr_static_id->operand_alternative
4106 [goal_alt_number * n_operands + i].constraint);
4107 for (;
4108 (c = *constraint) && c != ',' && c != '#';
4109 constraint += CONSTRAINT_LEN (c, constraint))
4111 enum constraint_num cn = lookup_constraint (constraint);
4112 if ((insn_extra_memory_constraint (cn)
4113 || insn_extra_special_memory_constraint (cn))
4114 && satisfies_memory_constraint_p (tem, cn))
4115 break;
4117 if (c == '\0' || c == ',' || c == '#')
4118 continue;
4120 goal_alt_win[i] = true;
4124 n_outputs = 0;
4125 outputs[0] = -1;
4126 for (i = 0; i < n_operands; i++)
4128 int regno;
4129 bool optional_p = false;
4130 rtx old, new_reg;
4131 rtx op = *curr_id->operand_loc[i];
4133 if (goal_alt_win[i])
4135 if (goal_alt[i] == NO_REGS
4136 && REG_P (op)
4137 /* When we assign NO_REGS it means that we will not
4138 assign a hard register to the scratch pseudo by
4139 assigment pass and the scratch pseudo will be
4140 spilled. Spilled scratch pseudos are transformed
4141 back to scratches at the LRA end. */
4142 && lra_former_scratch_operand_p (curr_insn, i)
4143 && lra_former_scratch_p (REGNO (op)))
4145 int regno = REGNO (op);
4146 lra_change_class (regno, NO_REGS, " Change to", true);
4147 if (lra_get_regno_hard_regno (regno) >= 0)
4148 /* We don't have to mark all insn affected by the
4149 spilled pseudo as there is only one such insn, the
4150 current one. */
4151 reg_renumber[regno] = -1;
4152 lra_assert (bitmap_single_bit_set_p
4153 (&lra_reg_info[REGNO (op)].insn_bitmap));
4155 /* We can do an optional reload. If the pseudo got a hard
4156 reg, we might improve the code through inheritance. If
4157 it does not get a hard register we coalesce memory/memory
4158 moves later. Ignore move insns to avoid cycling. */
4159 if (! lra_simple_p
4160 && lra_undo_inheritance_iter < LRA_MAX_INHERITANCE_PASSES
4161 && goal_alt[i] != NO_REGS && REG_P (op)
4162 && (regno = REGNO (op)) >= FIRST_PSEUDO_REGISTER
4163 && regno < new_regno_start
4164 && ! lra_former_scratch_p (regno)
4165 && reg_renumber[regno] < 0
4166 /* Check that the optional reload pseudo will be able to
4167 hold given mode value. */
4168 && ! (prohibited_class_reg_set_mode_p
4169 (goal_alt[i], reg_class_contents[goal_alt[i]],
4170 PSEUDO_REGNO_MODE (regno)))
4171 && (curr_insn_set == NULL_RTX
4172 || !((REG_P (SET_SRC (curr_insn_set))
4173 || MEM_P (SET_SRC (curr_insn_set))
4174 || GET_CODE (SET_SRC (curr_insn_set)) == SUBREG)
4175 && (REG_P (SET_DEST (curr_insn_set))
4176 || MEM_P (SET_DEST (curr_insn_set))
4177 || GET_CODE (SET_DEST (curr_insn_set)) == SUBREG))))
4178 optional_p = true;
4179 else
4180 continue;
4183 /* Operands that match previous ones have already been handled. */
4184 if (goal_alt_matches[i] >= 0)
4185 continue;
4187 /* We should not have an operand with a non-offsettable address
4188 appearing where an offsettable address will do. It also may
4189 be a case when the address should be special in other words
4190 not a general one (e.g. it needs no index reg). */
4191 if (goal_alt_matched[i][0] == -1 && goal_alt_offmemok[i] && MEM_P (op))
4193 enum reg_class rclass;
4194 rtx *loc = &XEXP (op, 0);
4195 enum rtx_code code = GET_CODE (*loc);
4197 push_to_sequence (before);
4198 rclass = base_reg_class (GET_MODE (op), MEM_ADDR_SPACE (op),
4199 MEM, SCRATCH);
4200 if (GET_RTX_CLASS (code) == RTX_AUTOINC)
4201 new_reg = emit_inc (rclass, *loc, *loc,
4202 /* This value does not matter for MODIFY. */
4203 GET_MODE_SIZE (GET_MODE (op)));
4204 else if (get_reload_reg (OP_IN, Pmode, *loc, rclass, FALSE,
4205 "offsetable address", &new_reg))
4206 lra_emit_move (new_reg, *loc);
4207 before = get_insns ();
4208 end_sequence ();
4209 *loc = new_reg;
4210 lra_update_dup (curr_id, i);
4212 else if (goal_alt_matched[i][0] == -1)
4214 machine_mode mode;
4215 rtx reg, *loc;
4216 int hard_regno, byte;
4217 enum op_type type = curr_static_id->operand[i].type;
4219 loc = curr_id->operand_loc[i];
4220 mode = curr_operand_mode[i];
4221 if (GET_CODE (*loc) == SUBREG)
4223 reg = SUBREG_REG (*loc);
4224 byte = SUBREG_BYTE (*loc);
4225 if (REG_P (reg)
4226 /* Strict_low_part requires reload the register not
4227 the sub-register. */
4228 && (curr_static_id->operand[i].strict_low
4229 || (!paradoxical_subreg_p (mode, GET_MODE (reg))
4230 && (hard_regno
4231 = get_try_hard_regno (REGNO (reg))) >= 0
4232 && (simplify_subreg_regno
4233 (hard_regno,
4234 GET_MODE (reg), byte, mode) < 0)
4235 && (goal_alt[i] == NO_REGS
4236 || (simplify_subreg_regno
4237 (ira_class_hard_regs[goal_alt[i]][0],
4238 GET_MODE (reg), byte, mode) >= 0)))))
4240 /* An OP_INOUT is required when reloading a subreg of a
4241 mode wider than a word to ensure that data beyond the
4242 word being reloaded is preserved. Also automatically
4243 ensure that strict_low_part reloads are made into
4244 OP_INOUT which should already be true from the backend
4245 constraints. */
4246 if (type == OP_OUT
4247 && (curr_static_id->operand[i].strict_low
4248 || (GET_MODE_SIZE (GET_MODE (reg)) > UNITS_PER_WORD
4249 && (GET_MODE_SIZE (mode)
4250 < GET_MODE_SIZE (GET_MODE (reg))))))
4251 type = OP_INOUT;
4252 loc = &SUBREG_REG (*loc);
4253 mode = GET_MODE (*loc);
4256 old = *loc;
4257 if (get_reload_reg (type, mode, old, goal_alt[i],
4258 loc != curr_id->operand_loc[i], "", &new_reg)
4259 && type != OP_OUT)
4261 push_to_sequence (before);
4262 lra_emit_move (new_reg, old);
4263 before = get_insns ();
4264 end_sequence ();
4266 *loc = new_reg;
4267 if (type != OP_IN
4268 && find_reg_note (curr_insn, REG_UNUSED, old) == NULL_RTX)
4270 start_sequence ();
4271 lra_emit_move (type == OP_INOUT ? copy_rtx (old) : old, new_reg);
4272 emit_insn (after);
4273 after = get_insns ();
4274 end_sequence ();
4275 *loc = new_reg;
4277 for (j = 0; j < goal_alt_dont_inherit_ops_num; j++)
4278 if (goal_alt_dont_inherit_ops[j] == i)
4280 lra_set_regno_unique_value (REGNO (new_reg));
4281 break;
4283 lra_update_dup (curr_id, i);
4285 else if (curr_static_id->operand[i].type == OP_IN
4286 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4287 == OP_OUT))
4289 /* generate reloads for input and matched outputs. */
4290 match_inputs[0] = i;
4291 match_inputs[1] = -1;
4292 match_reload (goal_alt_matched[i][0], match_inputs, outputs,
4293 goal_alt[i], &before, &after,
4294 curr_static_id->operand_alternative
4295 [goal_alt_number * n_operands + goal_alt_matched[i][0]]
4296 .earlyclobber);
4298 else if (curr_static_id->operand[i].type == OP_OUT
4299 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4300 == OP_IN))
4301 /* Generate reloads for output and matched inputs. */
4302 match_reload (i, goal_alt_matched[i], outputs, goal_alt[i], &before,
4303 &after, curr_static_id->operand_alternative
4304 [goal_alt_number * n_operands + i].earlyclobber);
4305 else if (curr_static_id->operand[i].type == OP_IN
4306 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4307 == OP_IN))
4309 /* Generate reloads for matched inputs. */
4310 match_inputs[0] = i;
4311 for (j = 0; (k = goal_alt_matched[i][j]) >= 0; j++)
4312 match_inputs[j + 1] = k;
4313 match_inputs[j + 1] = -1;
4314 match_reload (-1, match_inputs, outputs, goal_alt[i], &before,
4315 &after, false);
4317 else
4318 /* We must generate code in any case when function
4319 process_alt_operands decides that it is possible. */
4320 gcc_unreachable ();
4322 /* Memorise processed outputs so that output remaining to be processed
4323 can avoid using the same register value (see match_reload). */
4324 if (curr_static_id->operand[i].type == OP_OUT)
4326 outputs[n_outputs++] = i;
4327 outputs[n_outputs] = -1;
4330 if (optional_p)
4332 rtx reg = op;
4334 lra_assert (REG_P (reg));
4335 regno = REGNO (reg);
4336 op = *curr_id->operand_loc[i]; /* Substitution. */
4337 if (GET_CODE (op) == SUBREG)
4338 op = SUBREG_REG (op);
4339 gcc_assert (REG_P (op) && (int) REGNO (op) >= new_regno_start);
4340 bitmap_set_bit (&lra_optional_reload_pseudos, REGNO (op));
4341 lra_reg_info[REGNO (op)].restore_rtx = reg;
4342 if (lra_dump_file != NULL)
4343 fprintf (lra_dump_file,
4344 " Making reload reg %d for reg %d optional\n",
4345 REGNO (op), regno);
4348 if (before != NULL_RTX || after != NULL_RTX
4349 || max_regno_before != max_reg_num ())
4350 change_p = true;
4351 if (change_p)
4353 lra_update_operator_dups (curr_id);
4354 /* Something changes -- process the insn. */
4355 lra_update_insn_regno_info (curr_insn);
4357 lra_process_new_insns (curr_insn, before, after, "Inserting insn reload");
4358 return change_p;
4361 /* Return true if INSN satisfies all constraints. In other words, no
4362 reload insns are needed. */
4363 bool
4364 lra_constrain_insn (rtx_insn *insn)
4366 int saved_new_regno_start = new_regno_start;
4367 int saved_new_insn_uid_start = new_insn_uid_start;
4368 bool change_p;
4370 curr_insn = insn;
4371 curr_id = lra_get_insn_recog_data (curr_insn);
4372 curr_static_id = curr_id->insn_static_data;
4373 new_insn_uid_start = get_max_uid ();
4374 new_regno_start = max_reg_num ();
4375 change_p = curr_insn_transform (true);
4376 new_regno_start = saved_new_regno_start;
4377 new_insn_uid_start = saved_new_insn_uid_start;
4378 return ! change_p;
4381 /* Return true if X is in LIST. */
4382 static bool
4383 in_list_p (rtx x, rtx list)
4385 for (; list != NULL_RTX; list = XEXP (list, 1))
4386 if (XEXP (list, 0) == x)
4387 return true;
4388 return false;
4391 /* Return true if X contains an allocatable hard register (if
4392 HARD_REG_P) or a (spilled if SPILLED_P) pseudo. */
4393 static bool
4394 contains_reg_p (rtx x, bool hard_reg_p, bool spilled_p)
4396 int i, j;
4397 const char *fmt;
4398 enum rtx_code code;
4400 code = GET_CODE (x);
4401 if (REG_P (x))
4403 int regno = REGNO (x);
4404 HARD_REG_SET alloc_regs;
4406 if (hard_reg_p)
4408 if (regno >= FIRST_PSEUDO_REGISTER)
4409 regno = lra_get_regno_hard_regno (regno);
4410 if (regno < 0)
4411 return false;
4412 COMPL_HARD_REG_SET (alloc_regs, lra_no_alloc_regs);
4413 return overlaps_hard_reg_set_p (alloc_regs, GET_MODE (x), regno);
4415 else
4417 if (regno < FIRST_PSEUDO_REGISTER)
4418 return false;
4419 if (! spilled_p)
4420 return true;
4421 return lra_get_regno_hard_regno (regno) < 0;
4424 fmt = GET_RTX_FORMAT (code);
4425 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4427 if (fmt[i] == 'e')
4429 if (contains_reg_p (XEXP (x, i), hard_reg_p, spilled_p))
4430 return true;
4432 else if (fmt[i] == 'E')
4434 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4435 if (contains_reg_p (XVECEXP (x, i, j), hard_reg_p, spilled_p))
4436 return true;
4439 return false;
4442 /* Process all regs in location *LOC and change them on equivalent
4443 substitution. Return true if any change was done. */
4444 static bool
4445 loc_equivalence_change_p (rtx *loc)
4447 rtx subst, reg, x = *loc;
4448 bool result = false;
4449 enum rtx_code code = GET_CODE (x);
4450 const char *fmt;
4451 int i, j;
4453 if (code == SUBREG)
4455 reg = SUBREG_REG (x);
4456 if ((subst = get_equiv_with_elimination (reg, curr_insn)) != reg
4457 && GET_MODE (subst) == VOIDmode)
4459 /* We cannot reload debug location. Simplify subreg here
4460 while we know the inner mode. */
4461 *loc = simplify_gen_subreg (GET_MODE (x), subst,
4462 GET_MODE (reg), SUBREG_BYTE (x));
4463 return true;
4466 if (code == REG && (subst = get_equiv_with_elimination (x, curr_insn)) != x)
4468 *loc = subst;
4469 return true;
4472 /* Scan all the operand sub-expressions. */
4473 fmt = GET_RTX_FORMAT (code);
4474 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4476 if (fmt[i] == 'e')
4477 result = loc_equivalence_change_p (&XEXP (x, i)) || result;
4478 else if (fmt[i] == 'E')
4479 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4480 result
4481 = loc_equivalence_change_p (&XVECEXP (x, i, j)) || result;
4483 return result;
4486 /* Similar to loc_equivalence_change_p, but for use as
4487 simplify_replace_fn_rtx callback. DATA is insn for which the
4488 elimination is done. If it null we don't do the elimination. */
4489 static rtx
4490 loc_equivalence_callback (rtx loc, const_rtx, void *data)
4492 if (!REG_P (loc))
4493 return NULL_RTX;
4495 rtx subst = (data == NULL
4496 ? get_equiv (loc) : get_equiv_with_elimination (loc, (rtx_insn *) data));
4497 if (subst != loc)
4498 return subst;
4500 return NULL_RTX;
4503 /* Maximum number of generated reload insns per an insn. It is for
4504 preventing this pass cycling in a bug case. */
4505 #define MAX_RELOAD_INSNS_NUMBER LRA_MAX_INSN_RELOADS
4507 /* The current iteration number of this LRA pass. */
4508 int lra_constraint_iter;
4510 /* True if we substituted equiv which needs checking register
4511 allocation correctness because the equivalent value contains
4512 allocatable hard registers or when we restore multi-register
4513 pseudo. */
4514 bool lra_risky_transformations_p;
4516 /* Return true if REGNO is referenced in more than one block. */
4517 static bool
4518 multi_block_pseudo_p (int regno)
4520 basic_block bb = NULL;
4521 unsigned int uid;
4522 bitmap_iterator bi;
4524 if (regno < FIRST_PSEUDO_REGISTER)
4525 return false;
4527 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi)
4528 if (bb == NULL)
4529 bb = BLOCK_FOR_INSN (lra_insn_recog_data[uid]->insn);
4530 else if (BLOCK_FOR_INSN (lra_insn_recog_data[uid]->insn) != bb)
4531 return true;
4532 return false;
4535 /* Return true if LIST contains a deleted insn. */
4536 static bool
4537 contains_deleted_insn_p (rtx_insn_list *list)
4539 for (; list != NULL_RTX; list = list->next ())
4540 if (NOTE_P (list->insn ())
4541 && NOTE_KIND (list->insn ()) == NOTE_INSN_DELETED)
4542 return true;
4543 return false;
4546 /* Return true if X contains a pseudo dying in INSN. */
4547 static bool
4548 dead_pseudo_p (rtx x, rtx_insn *insn)
4550 int i, j;
4551 const char *fmt;
4552 enum rtx_code code;
4554 if (REG_P (x))
4555 return (insn != NULL_RTX
4556 && find_regno_note (insn, REG_DEAD, REGNO (x)) != NULL_RTX);
4557 code = GET_CODE (x);
4558 fmt = GET_RTX_FORMAT (code);
4559 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4561 if (fmt[i] == 'e')
4563 if (dead_pseudo_p (XEXP (x, i), insn))
4564 return true;
4566 else if (fmt[i] == 'E')
4568 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4569 if (dead_pseudo_p (XVECEXP (x, i, j), insn))
4570 return true;
4573 return false;
4576 /* Return true if INSN contains a dying pseudo in INSN right hand
4577 side. */
4578 static bool
4579 insn_rhs_dead_pseudo_p (rtx_insn *insn)
4581 rtx set = single_set (insn);
4583 gcc_assert (set != NULL);
4584 return dead_pseudo_p (SET_SRC (set), insn);
4587 /* Return true if any init insn of REGNO contains a dying pseudo in
4588 insn right hand side. */
4589 static bool
4590 init_insn_rhs_dead_pseudo_p (int regno)
4592 rtx_insn_list *insns = ira_reg_equiv[regno].init_insns;
4594 if (insns == NULL)
4595 return false;
4596 for (; insns != NULL_RTX; insns = insns->next ())
4597 if (insn_rhs_dead_pseudo_p (insns->insn ()))
4598 return true;
4599 return false;
4602 /* Return TRUE if REGNO has a reverse equivalence. The equivalence is
4603 reverse only if we have one init insn with given REGNO as a
4604 source. */
4605 static bool
4606 reverse_equiv_p (int regno)
4608 rtx_insn_list *insns = ira_reg_equiv[regno].init_insns;
4609 rtx set;
4611 if (insns == NULL)
4612 return false;
4613 if (! INSN_P (insns->insn ())
4614 || insns->next () != NULL)
4615 return false;
4616 if ((set = single_set (insns->insn ())) == NULL_RTX)
4617 return false;
4618 return REG_P (SET_SRC (set)) && (int) REGNO (SET_SRC (set)) == regno;
4621 /* Return TRUE if REGNO was reloaded in an equivalence init insn. We
4622 call this function only for non-reverse equivalence. */
4623 static bool
4624 contains_reloaded_insn_p (int regno)
4626 rtx set;
4627 rtx_insn_list *list = ira_reg_equiv[regno].init_insns;
4629 for (; list != NULL; list = list->next ())
4630 if ((set = single_set (list->insn ())) == NULL_RTX
4631 || ! REG_P (SET_DEST (set))
4632 || (int) REGNO (SET_DEST (set)) != regno)
4633 return true;
4634 return false;
4637 /* Entry function of LRA constraint pass. Return true if the
4638 constraint pass did change the code. */
4639 bool
4640 lra_constraints (bool first_p)
4642 bool changed_p;
4643 int i, hard_regno, new_insns_num;
4644 unsigned int min_len, new_min_len, uid;
4645 rtx set, x, reg, dest_reg;
4646 basic_block last_bb;
4647 bitmap_iterator bi;
4649 lra_constraint_iter++;
4650 if (lra_dump_file != NULL)
4651 fprintf (lra_dump_file, "\n********** Local #%d: **********\n\n",
4652 lra_constraint_iter);
4653 changed_p = false;
4654 if (pic_offset_table_rtx
4655 && REGNO (pic_offset_table_rtx) >= FIRST_PSEUDO_REGISTER)
4656 lra_risky_transformations_p = true;
4657 else
4658 /* On the first iteration we should check IRA assignment
4659 correctness. In rare cases, the assignments can be wrong as
4660 early clobbers operands are ignored in IRA. */
4661 lra_risky_transformations_p = first_p;
4662 new_insn_uid_start = get_max_uid ();
4663 new_regno_start = first_p ? lra_constraint_new_regno_start : max_reg_num ();
4664 /* Mark used hard regs for target stack size calulations. */
4665 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4666 if (lra_reg_info[i].nrefs != 0
4667 && (hard_regno = lra_get_regno_hard_regno (i)) >= 0)
4669 int j, nregs;
4671 nregs = hard_regno_nregs[hard_regno][lra_reg_info[i].biggest_mode];
4672 for (j = 0; j < nregs; j++)
4673 df_set_regs_ever_live (hard_regno + j, true);
4675 /* Do elimination before the equivalence processing as we can spill
4676 some pseudos during elimination. */
4677 lra_eliminate (false, first_p);
4678 auto_bitmap equiv_insn_bitmap (&reg_obstack);
4679 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4680 if (lra_reg_info[i].nrefs != 0)
4682 ira_reg_equiv[i].profitable_p = true;
4683 reg = regno_reg_rtx[i];
4684 if (lra_get_regno_hard_regno (i) < 0 && (x = get_equiv (reg)) != reg)
4686 bool pseudo_p = contains_reg_p (x, false, false);
4688 /* After RTL transformation, we can not guarantee that
4689 pseudo in the substitution was not reloaded which might
4690 make equivalence invalid. For example, in reverse
4691 equiv of p0
4693 p0 <- ...
4695 equiv_mem <- p0
4697 the memory address register was reloaded before the 2nd
4698 insn. */
4699 if ((! first_p && pseudo_p)
4700 /* We don't use DF for compilation speed sake. So it
4701 is problematic to update live info when we use an
4702 equivalence containing pseudos in more than one
4703 BB. */
4704 || (pseudo_p && multi_block_pseudo_p (i))
4705 /* If an init insn was deleted for some reason, cancel
4706 the equiv. We could update the equiv insns after
4707 transformations including an equiv insn deletion
4708 but it is not worthy as such cases are extremely
4709 rare. */
4710 || contains_deleted_insn_p (ira_reg_equiv[i].init_insns)
4711 /* If it is not a reverse equivalence, we check that a
4712 pseudo in rhs of the init insn is not dying in the
4713 insn. Otherwise, the live info at the beginning of
4714 the corresponding BB might be wrong after we
4715 removed the insn. When the equiv can be a
4716 constant, the right hand side of the init insn can
4717 be a pseudo. */
4718 || (! reverse_equiv_p (i)
4719 && (init_insn_rhs_dead_pseudo_p (i)
4720 /* If we reloaded the pseudo in an equivalence
4721 init insn, we can not remove the equiv init
4722 insns and the init insns might write into
4723 const memory in this case. */
4724 || contains_reloaded_insn_p (i)))
4725 /* Prevent access beyond equivalent memory for
4726 paradoxical subregs. */
4727 || (MEM_P (x)
4728 && (GET_MODE_SIZE (lra_reg_info[i].biggest_mode)
4729 > GET_MODE_SIZE (GET_MODE (x))))
4730 || (pic_offset_table_rtx
4731 && ((CONST_POOL_OK_P (PSEUDO_REGNO_MODE (i), x)
4732 && (targetm.preferred_reload_class
4733 (x, lra_get_allocno_class (i)) == NO_REGS))
4734 || contains_symbol_ref_p (x))))
4735 ira_reg_equiv[i].defined_p = false;
4736 if (contains_reg_p (x, false, true))
4737 ira_reg_equiv[i].profitable_p = false;
4738 if (get_equiv (reg) != reg)
4739 bitmap_ior_into (equiv_insn_bitmap, &lra_reg_info[i].insn_bitmap);
4742 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4743 update_equiv (i);
4744 /* We should add all insns containing pseudos which should be
4745 substituted by their equivalences. */
4746 EXECUTE_IF_SET_IN_BITMAP (equiv_insn_bitmap, 0, uid, bi)
4747 lra_push_insn_by_uid (uid);
4748 min_len = lra_insn_stack_length ();
4749 new_insns_num = 0;
4750 last_bb = NULL;
4751 changed_p = false;
4752 while ((new_min_len = lra_insn_stack_length ()) != 0)
4754 curr_insn = lra_pop_insn ();
4755 --new_min_len;
4756 curr_bb = BLOCK_FOR_INSN (curr_insn);
4757 if (curr_bb != last_bb)
4759 last_bb = curr_bb;
4760 bb_reload_num = lra_curr_reload_num;
4762 if (min_len > new_min_len)
4764 min_len = new_min_len;
4765 new_insns_num = 0;
4767 if (new_insns_num > MAX_RELOAD_INSNS_NUMBER)
4768 internal_error
4769 ("Max. number of generated reload insns per insn is achieved (%d)\n",
4770 MAX_RELOAD_INSNS_NUMBER);
4771 new_insns_num++;
4772 if (DEBUG_INSN_P (curr_insn))
4774 /* We need to check equivalence in debug insn and change
4775 pseudo to the equivalent value if necessary. */
4776 curr_id = lra_get_insn_recog_data (curr_insn);
4777 if (bitmap_bit_p (equiv_insn_bitmap, INSN_UID (curr_insn)))
4779 rtx old = *curr_id->operand_loc[0];
4780 *curr_id->operand_loc[0]
4781 = simplify_replace_fn_rtx (old, NULL_RTX,
4782 loc_equivalence_callback, curr_insn);
4783 if (old != *curr_id->operand_loc[0])
4785 lra_update_insn_regno_info (curr_insn);
4786 changed_p = true;
4790 else if (INSN_P (curr_insn))
4792 if ((set = single_set (curr_insn)) != NULL_RTX)
4794 dest_reg = SET_DEST (set);
4795 /* The equivalence pseudo could be set up as SUBREG in a
4796 case when it is a call restore insn in a mode
4797 different from the pseudo mode. */
4798 if (GET_CODE (dest_reg) == SUBREG)
4799 dest_reg = SUBREG_REG (dest_reg);
4800 if ((REG_P (dest_reg)
4801 && (x = get_equiv (dest_reg)) != dest_reg
4802 /* Remove insns which set up a pseudo whose value
4803 can not be changed. Such insns might be not in
4804 init_insns because we don't update equiv data
4805 during insn transformations.
4807 As an example, let suppose that a pseudo got
4808 hard register and on the 1st pass was not
4809 changed to equivalent constant. We generate an
4810 additional insn setting up the pseudo because of
4811 secondary memory movement. Then the pseudo is
4812 spilled and we use the equiv constant. In this
4813 case we should remove the additional insn and
4814 this insn is not init_insns list. */
4815 && (! MEM_P (x) || MEM_READONLY_P (x)
4816 /* Check that this is actually an insn setting
4817 up the equivalence. */
4818 || in_list_p (curr_insn,
4819 ira_reg_equiv
4820 [REGNO (dest_reg)].init_insns)))
4821 || (((x = get_equiv (SET_SRC (set))) != SET_SRC (set))
4822 && in_list_p (curr_insn,
4823 ira_reg_equiv
4824 [REGNO (SET_SRC (set))].init_insns)))
4826 /* This is equiv init insn of pseudo which did not get a
4827 hard register -- remove the insn. */
4828 if (lra_dump_file != NULL)
4830 fprintf (lra_dump_file,
4831 " Removing equiv init insn %i (freq=%d)\n",
4832 INSN_UID (curr_insn),
4833 REG_FREQ_FROM_BB (BLOCK_FOR_INSN (curr_insn)));
4834 dump_insn_slim (lra_dump_file, curr_insn);
4836 if (contains_reg_p (x, true, false))
4837 lra_risky_transformations_p = true;
4838 lra_set_insn_deleted (curr_insn);
4839 continue;
4842 curr_id = lra_get_insn_recog_data (curr_insn);
4843 curr_static_id = curr_id->insn_static_data;
4844 init_curr_insn_input_reloads ();
4845 init_curr_operand_mode ();
4846 if (curr_insn_transform (false))
4847 changed_p = true;
4848 /* Check non-transformed insns too for equiv change as USE
4849 or CLOBBER don't need reloads but can contain pseudos
4850 being changed on their equivalences. */
4851 else if (bitmap_bit_p (equiv_insn_bitmap, INSN_UID (curr_insn))
4852 && loc_equivalence_change_p (&PATTERN (curr_insn)))
4854 lra_update_insn_regno_info (curr_insn);
4855 changed_p = true;
4860 /* If we used a new hard regno, changed_p should be true because the
4861 hard reg is assigned to a new pseudo. */
4862 if (flag_checking && !changed_p)
4864 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4865 if (lra_reg_info[i].nrefs != 0
4866 && (hard_regno = lra_get_regno_hard_regno (i)) >= 0)
4868 int j, nregs = hard_regno_nregs[hard_regno][PSEUDO_REGNO_MODE (i)];
4870 for (j = 0; j < nregs; j++)
4871 lra_assert (df_regs_ever_live_p (hard_regno + j));
4874 return changed_p;
4877 static void initiate_invariants (void);
4878 static void finish_invariants (void);
4880 /* Initiate the LRA constraint pass. It is done once per
4881 function. */
4882 void
4883 lra_constraints_init (void)
4885 initiate_invariants ();
4888 /* Finalize the LRA constraint pass. It is done once per
4889 function. */
4890 void
4891 lra_constraints_finish (void)
4893 finish_invariants ();
4898 /* Structure describes invariants for ineheritance. */
4899 struct lra_invariant
4901 /* The order number of the invariant. */
4902 int num;
4903 /* The invariant RTX. */
4904 rtx invariant_rtx;
4905 /* The origin insn of the invariant. */
4906 rtx_insn *insn;
4909 typedef lra_invariant invariant_t;
4910 typedef invariant_t *invariant_ptr_t;
4911 typedef const invariant_t *const_invariant_ptr_t;
4913 /* Pointer to the inheritance invariants. */
4914 static vec<invariant_ptr_t> invariants;
4916 /* Allocation pool for the invariants. */
4917 static object_allocator<lra_invariant> *invariants_pool;
4919 /* Hash table for the invariants. */
4920 static htab_t invariant_table;
4922 /* Hash function for INVARIANT. */
4923 static hashval_t
4924 invariant_hash (const void *invariant)
4926 rtx inv = ((const_invariant_ptr_t) invariant)->invariant_rtx;
4927 return lra_rtx_hash (inv);
4930 /* Equal function for invariants INVARIANT1 and INVARIANT2. */
4931 static int
4932 invariant_eq_p (const void *invariant1, const void *invariant2)
4934 rtx inv1 = ((const_invariant_ptr_t) invariant1)->invariant_rtx;
4935 rtx inv2 = ((const_invariant_ptr_t) invariant2)->invariant_rtx;
4937 return rtx_equal_p (inv1, inv2);
4940 /* Insert INVARIANT_RTX into the table if it is not there yet. Return
4941 invariant which is in the table. */
4942 static invariant_ptr_t
4943 insert_invariant (rtx invariant_rtx)
4945 void **entry_ptr;
4946 invariant_t invariant;
4947 invariant_ptr_t invariant_ptr;
4949 invariant.invariant_rtx = invariant_rtx;
4950 entry_ptr = htab_find_slot (invariant_table, &invariant, INSERT);
4951 if (*entry_ptr == NULL)
4953 invariant_ptr = invariants_pool->allocate ();
4954 invariant_ptr->invariant_rtx = invariant_rtx;
4955 invariant_ptr->insn = NULL;
4956 invariants.safe_push (invariant_ptr);
4957 *entry_ptr = (void *) invariant_ptr;
4959 return (invariant_ptr_t) *entry_ptr;
4962 /* Initiate the invariant table. */
4963 static void
4964 initiate_invariants (void)
4966 invariants.create (100);
4967 invariants_pool
4968 = new object_allocator<lra_invariant> ("Inheritance invariants");
4969 invariant_table = htab_create (100, invariant_hash, invariant_eq_p, NULL);
4972 /* Finish the invariant table. */
4973 static void
4974 finish_invariants (void)
4976 htab_delete (invariant_table);
4977 delete invariants_pool;
4978 invariants.release ();
4981 /* Make the invariant table empty. */
4982 static void
4983 clear_invariants (void)
4985 htab_empty (invariant_table);
4986 invariants_pool->release ();
4987 invariants.truncate (0);
4992 /* This page contains code to do inheritance/split
4993 transformations. */
4995 /* Number of reloads passed so far in current EBB. */
4996 static int reloads_num;
4998 /* Number of calls passed so far in current EBB. */
4999 static int calls_num;
5001 /* Current reload pseudo check for validity of elements in
5002 USAGE_INSNS. */
5003 static int curr_usage_insns_check;
5005 /* Info about last usage of registers in EBB to do inheritance/split
5006 transformation. Inheritance transformation is done from a spilled
5007 pseudo and split transformations from a hard register or a pseudo
5008 assigned to a hard register. */
5009 struct usage_insns
5011 /* If the value is equal to CURR_USAGE_INSNS_CHECK, then the member
5012 value INSNS is valid. The insns is chain of optional debug insns
5013 and a finishing non-debug insn using the corresponding reg. The
5014 value is also used to mark the registers which are set up in the
5015 current insn. The negated insn uid is used for this. */
5016 int check;
5017 /* Value of global reloads_num at the last insn in INSNS. */
5018 int reloads_num;
5019 /* Value of global reloads_nums at the last insn in INSNS. */
5020 int calls_num;
5021 /* It can be true only for splitting. And it means that the restore
5022 insn should be put after insn given by the following member. */
5023 bool after_p;
5024 /* Next insns in the current EBB which use the original reg and the
5025 original reg value is not changed between the current insn and
5026 the next insns. In order words, e.g. for inheritance, if we need
5027 to use the original reg value again in the next insns we can try
5028 to use the value in a hard register from a reload insn of the
5029 current insn. */
5030 rtx insns;
5033 /* Map: regno -> corresponding pseudo usage insns. */
5034 static struct usage_insns *usage_insns;
5036 static void
5037 setup_next_usage_insn (int regno, rtx insn, int reloads_num, bool after_p)
5039 usage_insns[regno].check = curr_usage_insns_check;
5040 usage_insns[regno].insns = insn;
5041 usage_insns[regno].reloads_num = reloads_num;
5042 usage_insns[regno].calls_num = calls_num;
5043 usage_insns[regno].after_p = after_p;
5046 /* The function is used to form list REGNO usages which consists of
5047 optional debug insns finished by a non-debug insn using REGNO.
5048 RELOADS_NUM is current number of reload insns processed so far. */
5049 static void
5050 add_next_usage_insn (int regno, rtx_insn *insn, int reloads_num)
5052 rtx next_usage_insns;
5054 if (usage_insns[regno].check == curr_usage_insns_check
5055 && (next_usage_insns = usage_insns[regno].insns) != NULL_RTX
5056 && DEBUG_INSN_P (insn))
5058 /* Check that we did not add the debug insn yet. */
5059 if (next_usage_insns != insn
5060 && (GET_CODE (next_usage_insns) != INSN_LIST
5061 || XEXP (next_usage_insns, 0) != insn))
5062 usage_insns[regno].insns = gen_rtx_INSN_LIST (VOIDmode, insn,
5063 next_usage_insns);
5065 else if (NONDEBUG_INSN_P (insn))
5066 setup_next_usage_insn (regno, insn, reloads_num, false);
5067 else
5068 usage_insns[regno].check = 0;
5071 /* Return first non-debug insn in list USAGE_INSNS. */
5072 static rtx_insn *
5073 skip_usage_debug_insns (rtx usage_insns)
5075 rtx insn;
5077 /* Skip debug insns. */
5078 for (insn = usage_insns;
5079 insn != NULL_RTX && GET_CODE (insn) == INSN_LIST;
5080 insn = XEXP (insn, 1))
5082 return safe_as_a <rtx_insn *> (insn);
5085 /* Return true if we need secondary memory moves for insn in
5086 USAGE_INSNS after inserting inherited pseudo of class INHER_CL
5087 into the insn. */
5088 static bool
5089 check_secondary_memory_needed_p (enum reg_class inher_cl ATTRIBUTE_UNUSED,
5090 rtx usage_insns ATTRIBUTE_UNUSED)
5092 #ifndef SECONDARY_MEMORY_NEEDED
5093 return false;
5094 #else
5095 rtx_insn *insn;
5096 rtx set, dest;
5097 enum reg_class cl;
5099 if (inher_cl == ALL_REGS
5100 || (insn = skip_usage_debug_insns (usage_insns)) == NULL_RTX)
5101 return false;
5102 lra_assert (INSN_P (insn));
5103 if ((set = single_set (insn)) == NULL_RTX || ! REG_P (SET_DEST (set)))
5104 return false;
5105 dest = SET_DEST (set);
5106 if (! REG_P (dest))
5107 return false;
5108 lra_assert (inher_cl != NO_REGS);
5109 cl = get_reg_class (REGNO (dest));
5110 return (cl != NO_REGS && cl != ALL_REGS
5111 && SECONDARY_MEMORY_NEEDED (inher_cl, cl, GET_MODE (dest)));
5112 #endif
5115 /* Registers involved in inheritance/split in the current EBB
5116 (inheritance/split pseudos and original registers). */
5117 static bitmap_head check_only_regs;
5119 /* Reload pseudos can not be involded in invariant inheritance in the
5120 current EBB. */
5121 static bitmap_head invalid_invariant_regs;
5123 /* Do inheritance transformations for insn INSN, which defines (if
5124 DEF_P) or uses ORIGINAL_REGNO. NEXT_USAGE_INSNS specifies which
5125 instruction in the EBB next uses ORIGINAL_REGNO; it has the same
5126 form as the "insns" field of usage_insns. Return true if we
5127 succeed in such transformation.
5129 The transformations look like:
5131 p <- ... i <- ...
5132 ... p <- i (new insn)
5133 ... =>
5134 <- ... p ... <- ... i ...
5136 ... i <- p (new insn)
5137 <- ... p ... <- ... i ...
5138 ... =>
5139 <- ... p ... <- ... i ...
5140 where p is a spilled original pseudo and i is a new inheritance pseudo.
5143 The inheritance pseudo has the smallest class of two classes CL and
5144 class of ORIGINAL REGNO. */
5145 static bool
5146 inherit_reload_reg (bool def_p, int original_regno,
5147 enum reg_class cl, rtx_insn *insn, rtx next_usage_insns)
5149 if (optimize_function_for_size_p (cfun))
5150 return false;
5152 enum reg_class rclass = lra_get_allocno_class (original_regno);
5153 rtx original_reg = regno_reg_rtx[original_regno];
5154 rtx new_reg, usage_insn;
5155 rtx_insn *new_insns;
5157 lra_assert (! usage_insns[original_regno].after_p);
5158 if (lra_dump_file != NULL)
5159 fprintf (lra_dump_file,
5160 " <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<\n");
5161 if (! ira_reg_classes_intersect_p[cl][rclass])
5163 if (lra_dump_file != NULL)
5165 fprintf (lra_dump_file,
5166 " Rejecting inheritance for %d "
5167 "because of disjoint classes %s and %s\n",
5168 original_regno, reg_class_names[cl],
5169 reg_class_names[rclass]);
5170 fprintf (lra_dump_file,
5171 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5173 return false;
5175 if ((ira_class_subset_p[cl][rclass] && cl != rclass)
5176 /* We don't use a subset of two classes because it can be
5177 NO_REGS. This transformation is still profitable in most
5178 cases even if the classes are not intersected as register
5179 move is probably cheaper than a memory load. */
5180 || ira_class_hard_regs_num[cl] < ira_class_hard_regs_num[rclass])
5182 if (lra_dump_file != NULL)
5183 fprintf (lra_dump_file, " Use smallest class of %s and %s\n",
5184 reg_class_names[cl], reg_class_names[rclass]);
5186 rclass = cl;
5188 if (check_secondary_memory_needed_p (rclass, next_usage_insns))
5190 /* Reject inheritance resulting in secondary memory moves.
5191 Otherwise, there is a danger in LRA cycling. Also such
5192 transformation will be unprofitable. */
5193 if (lra_dump_file != NULL)
5195 rtx_insn *insn = skip_usage_debug_insns (next_usage_insns);
5196 rtx set = single_set (insn);
5198 lra_assert (set != NULL_RTX);
5200 rtx dest = SET_DEST (set);
5202 lra_assert (REG_P (dest));
5203 fprintf (lra_dump_file,
5204 " Rejecting inheritance for insn %d(%s)<-%d(%s) "
5205 "as secondary mem is needed\n",
5206 REGNO (dest), reg_class_names[get_reg_class (REGNO (dest))],
5207 original_regno, reg_class_names[rclass]);
5208 fprintf (lra_dump_file,
5209 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5211 return false;
5213 new_reg = lra_create_new_reg (GET_MODE (original_reg), original_reg,
5214 rclass, "inheritance");
5215 start_sequence ();
5216 if (def_p)
5217 lra_emit_move (original_reg, new_reg);
5218 else
5219 lra_emit_move (new_reg, original_reg);
5220 new_insns = get_insns ();
5221 end_sequence ();
5222 if (NEXT_INSN (new_insns) != NULL_RTX)
5224 if (lra_dump_file != NULL)
5226 fprintf (lra_dump_file,
5227 " Rejecting inheritance %d->%d "
5228 "as it results in 2 or more insns:\n",
5229 original_regno, REGNO (new_reg));
5230 dump_rtl_slim (lra_dump_file, new_insns, NULL, -1, 0);
5231 fprintf (lra_dump_file,
5232 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5234 return false;
5236 lra_substitute_pseudo_within_insn (insn, original_regno, new_reg, false);
5237 lra_update_insn_regno_info (insn);
5238 if (! def_p)
5239 /* We now have a new usage insn for original regno. */
5240 setup_next_usage_insn (original_regno, new_insns, reloads_num, false);
5241 if (lra_dump_file != NULL)
5242 fprintf (lra_dump_file, " Original reg change %d->%d (bb%d):\n",
5243 original_regno, REGNO (new_reg), BLOCK_FOR_INSN (insn)->index);
5244 lra_reg_info[REGNO (new_reg)].restore_rtx = regno_reg_rtx[original_regno];
5245 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
5246 bitmap_set_bit (&check_only_regs, original_regno);
5247 bitmap_set_bit (&lra_inheritance_pseudos, REGNO (new_reg));
5248 if (def_p)
5249 lra_process_new_insns (insn, NULL, new_insns,
5250 "Add original<-inheritance");
5251 else
5252 lra_process_new_insns (insn, new_insns, NULL,
5253 "Add inheritance<-original");
5254 while (next_usage_insns != NULL_RTX)
5256 if (GET_CODE (next_usage_insns) != INSN_LIST)
5258 usage_insn = next_usage_insns;
5259 lra_assert (NONDEBUG_INSN_P (usage_insn));
5260 next_usage_insns = NULL;
5262 else
5264 usage_insn = XEXP (next_usage_insns, 0);
5265 lra_assert (DEBUG_INSN_P (usage_insn));
5266 next_usage_insns = XEXP (next_usage_insns, 1);
5268 lra_substitute_pseudo (&usage_insn, original_regno, new_reg, false);
5269 lra_update_insn_regno_info (as_a <rtx_insn *> (usage_insn));
5270 if (lra_dump_file != NULL)
5272 fprintf (lra_dump_file,
5273 " Inheritance reuse change %d->%d (bb%d):\n",
5274 original_regno, REGNO (new_reg),
5275 BLOCK_FOR_INSN (usage_insn)->index);
5276 dump_insn_slim (lra_dump_file, as_a <rtx_insn *> (usage_insn));
5279 if (lra_dump_file != NULL)
5280 fprintf (lra_dump_file,
5281 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5282 return true;
5285 /* Return true if we need a caller save/restore for pseudo REGNO which
5286 was assigned to a hard register. */
5287 static inline bool
5288 need_for_call_save_p (int regno)
5290 lra_assert (regno >= FIRST_PSEUDO_REGISTER && reg_renumber[regno] >= 0);
5291 return (usage_insns[regno].calls_num < calls_num
5292 && (overlaps_hard_reg_set_p
5293 ((flag_ipa_ra &&
5294 ! hard_reg_set_empty_p (lra_reg_info[regno].actual_call_used_reg_set))
5295 ? lra_reg_info[regno].actual_call_used_reg_set
5296 : call_used_reg_set,
5297 PSEUDO_REGNO_MODE (regno), reg_renumber[regno])
5298 || HARD_REGNO_CALL_PART_CLOBBERED (reg_renumber[regno],
5299 PSEUDO_REGNO_MODE (regno))));
5302 /* Global registers occurring in the current EBB. */
5303 static bitmap_head ebb_global_regs;
5305 /* Return true if we need a split for hard register REGNO or pseudo
5306 REGNO which was assigned to a hard register.
5307 POTENTIAL_RELOAD_HARD_REGS contains hard registers which might be
5308 used for reloads since the EBB end. It is an approximation of the
5309 used hard registers in the split range. The exact value would
5310 require expensive calculations. If we were aggressive with
5311 splitting because of the approximation, the split pseudo will save
5312 the same hard register assignment and will be removed in the undo
5313 pass. We still need the approximation because too aggressive
5314 splitting would result in too inaccurate cost calculation in the
5315 assignment pass because of too many generated moves which will be
5316 probably removed in the undo pass. */
5317 static inline bool
5318 need_for_split_p (HARD_REG_SET potential_reload_hard_regs, int regno)
5320 int hard_regno = regno < FIRST_PSEUDO_REGISTER ? regno : reg_renumber[regno];
5322 lra_assert (hard_regno >= 0);
5323 return ((TEST_HARD_REG_BIT (potential_reload_hard_regs, hard_regno)
5324 /* Don't split eliminable hard registers, otherwise we can
5325 split hard registers like hard frame pointer, which
5326 lives on BB start/end according to DF-infrastructure,
5327 when there is a pseudo assigned to the register and
5328 living in the same BB. */
5329 && (regno >= FIRST_PSEUDO_REGISTER
5330 || ! TEST_HARD_REG_BIT (eliminable_regset, hard_regno))
5331 && ! TEST_HARD_REG_BIT (lra_no_alloc_regs, hard_regno)
5332 /* Don't split call clobbered hard regs living through
5333 calls, otherwise we might have a check problem in the
5334 assign sub-pass as in the most cases (exception is a
5335 situation when lra_risky_transformations_p value is
5336 true) the assign pass assumes that all pseudos living
5337 through calls are assigned to call saved hard regs. */
5338 && (regno >= FIRST_PSEUDO_REGISTER
5339 || ! TEST_HARD_REG_BIT (call_used_reg_set, regno)
5340 || usage_insns[regno].calls_num == calls_num)
5341 /* We need at least 2 reloads to make pseudo splitting
5342 profitable. We should provide hard regno splitting in
5343 any case to solve 1st insn scheduling problem when
5344 moving hard register definition up might result in
5345 impossibility to find hard register for reload pseudo of
5346 small register class. */
5347 && (usage_insns[regno].reloads_num
5348 + (regno < FIRST_PSEUDO_REGISTER ? 0 : 3) < reloads_num)
5349 && (regno < FIRST_PSEUDO_REGISTER
5350 /* For short living pseudos, spilling + inheritance can
5351 be considered a substitution for splitting.
5352 Therefore we do not splitting for local pseudos. It
5353 decreases also aggressiveness of splitting. The
5354 minimal number of references is chosen taking into
5355 account that for 2 references splitting has no sense
5356 as we can just spill the pseudo. */
5357 || (regno >= FIRST_PSEUDO_REGISTER
5358 && lra_reg_info[regno].nrefs > 3
5359 && bitmap_bit_p (&ebb_global_regs, regno))))
5360 || (regno >= FIRST_PSEUDO_REGISTER && need_for_call_save_p (regno)));
5363 /* Return class for the split pseudo created from original pseudo with
5364 ALLOCNO_CLASS and MODE which got a hard register HARD_REGNO. We
5365 choose subclass of ALLOCNO_CLASS which contains HARD_REGNO and
5366 results in no secondary memory movements. */
5367 static enum reg_class
5368 choose_split_class (enum reg_class allocno_class,
5369 int hard_regno ATTRIBUTE_UNUSED,
5370 machine_mode mode ATTRIBUTE_UNUSED)
5372 #ifndef SECONDARY_MEMORY_NEEDED
5373 return allocno_class;
5374 #else
5375 int i;
5376 enum reg_class cl, best_cl = NO_REGS;
5377 enum reg_class hard_reg_class ATTRIBUTE_UNUSED
5378 = REGNO_REG_CLASS (hard_regno);
5380 if (! SECONDARY_MEMORY_NEEDED (allocno_class, allocno_class, mode)
5381 && TEST_HARD_REG_BIT (reg_class_contents[allocno_class], hard_regno))
5382 return allocno_class;
5383 for (i = 0;
5384 (cl = reg_class_subclasses[allocno_class][i]) != LIM_REG_CLASSES;
5385 i++)
5386 if (! SECONDARY_MEMORY_NEEDED (cl, hard_reg_class, mode)
5387 && ! SECONDARY_MEMORY_NEEDED (hard_reg_class, cl, mode)
5388 && TEST_HARD_REG_BIT (reg_class_contents[cl], hard_regno)
5389 && (best_cl == NO_REGS
5390 || ira_class_hard_regs_num[best_cl] < ira_class_hard_regs_num[cl]))
5391 best_cl = cl;
5392 return best_cl;
5393 #endif
5396 /* Copy any equivalence information from ORIGINAL_REGNO to NEW_REGNO.
5397 It only makes sense to call this function if NEW_REGNO is always
5398 equal to ORIGINAL_REGNO. */
5400 static void
5401 lra_copy_reg_equiv (unsigned int new_regno, unsigned int original_regno)
5403 if (!ira_reg_equiv[original_regno].defined_p)
5404 return;
5406 ira_expand_reg_equiv ();
5407 ira_reg_equiv[new_regno].defined_p = true;
5408 if (ira_reg_equiv[original_regno].memory)
5409 ira_reg_equiv[new_regno].memory
5410 = copy_rtx (ira_reg_equiv[original_regno].memory);
5411 if (ira_reg_equiv[original_regno].constant)
5412 ira_reg_equiv[new_regno].constant
5413 = copy_rtx (ira_reg_equiv[original_regno].constant);
5414 if (ira_reg_equiv[original_regno].invariant)
5415 ira_reg_equiv[new_regno].invariant
5416 = copy_rtx (ira_reg_equiv[original_regno].invariant);
5419 /* Do split transformations for insn INSN, which defines or uses
5420 ORIGINAL_REGNO. NEXT_USAGE_INSNS specifies which instruction in
5421 the EBB next uses ORIGINAL_REGNO; it has the same form as the
5422 "insns" field of usage_insns.
5424 The transformations look like:
5426 p <- ... p <- ...
5427 ... s <- p (new insn -- save)
5428 ... =>
5429 ... p <- s (new insn -- restore)
5430 <- ... p ... <- ... p ...
5432 <- ... p ... <- ... p ...
5433 ... s <- p (new insn -- save)
5434 ... =>
5435 ... p <- s (new insn -- restore)
5436 <- ... p ... <- ... p ...
5438 where p is an original pseudo got a hard register or a hard
5439 register and s is a new split pseudo. The save is put before INSN
5440 if BEFORE_P is true. Return true if we succeed in such
5441 transformation. */
5442 static bool
5443 split_reg (bool before_p, int original_regno, rtx_insn *insn,
5444 rtx next_usage_insns)
5446 enum reg_class rclass;
5447 rtx original_reg;
5448 int hard_regno, nregs;
5449 rtx new_reg, usage_insn;
5450 rtx_insn *restore, *save;
5451 bool after_p;
5452 bool call_save_p;
5453 machine_mode mode;
5455 if (original_regno < FIRST_PSEUDO_REGISTER)
5457 rclass = ira_allocno_class_translate[REGNO_REG_CLASS (original_regno)];
5458 hard_regno = original_regno;
5459 call_save_p = false;
5460 nregs = 1;
5461 mode = lra_reg_info[hard_regno].biggest_mode;
5462 machine_mode reg_rtx_mode = GET_MODE (regno_reg_rtx[hard_regno]);
5463 /* A reg can have a biggest_mode of VOIDmode if it was only ever seen
5464 as part of a multi-word register. In that case, or if the biggest
5465 mode was larger than a register, just use the reg_rtx. Otherwise,
5466 limit the size to that of the biggest access in the function. */
5467 if (mode == VOIDmode
5468 || paradoxical_subreg_p (mode, reg_rtx_mode))
5470 original_reg = regno_reg_rtx[hard_regno];
5471 mode = reg_rtx_mode;
5473 else
5474 original_reg = gen_rtx_REG (mode, hard_regno);
5476 else
5478 mode = PSEUDO_REGNO_MODE (original_regno);
5479 hard_regno = reg_renumber[original_regno];
5480 nregs = hard_regno_nregs[hard_regno][mode];
5481 rclass = lra_get_allocno_class (original_regno);
5482 original_reg = regno_reg_rtx[original_regno];
5483 call_save_p = need_for_call_save_p (original_regno);
5485 lra_assert (hard_regno >= 0);
5486 if (lra_dump_file != NULL)
5487 fprintf (lra_dump_file,
5488 " ((((((((((((((((((((((((((((((((((((((((((((((((\n");
5490 if (call_save_p)
5492 mode = HARD_REGNO_CALLER_SAVE_MODE (hard_regno,
5493 hard_regno_nregs[hard_regno][mode],
5494 mode);
5495 new_reg = lra_create_new_reg (mode, NULL_RTX, NO_REGS, "save");
5497 else
5499 rclass = choose_split_class (rclass, hard_regno, mode);
5500 if (rclass == NO_REGS)
5502 if (lra_dump_file != NULL)
5504 fprintf (lra_dump_file,
5505 " Rejecting split of %d(%s): "
5506 "no good reg class for %d(%s)\n",
5507 original_regno,
5508 reg_class_names[lra_get_allocno_class (original_regno)],
5509 hard_regno,
5510 reg_class_names[REGNO_REG_CLASS (hard_regno)]);
5511 fprintf
5512 (lra_dump_file,
5513 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5515 return false;
5517 /* Split_if_necessary can split hard registers used as part of a
5518 multi-register mode but splits each register individually. The
5519 mode used for each independent register may not be supported
5520 so reject the split. Splitting the wider mode should theoretically
5521 be possible but is not implemented. */
5522 if (! HARD_REGNO_MODE_OK (hard_regno, mode))
5524 if (lra_dump_file != NULL)
5526 fprintf (lra_dump_file,
5527 " Rejecting split of %d(%s): unsuitable mode %s\n",
5528 original_regno,
5529 reg_class_names[lra_get_allocno_class (original_regno)],
5530 GET_MODE_NAME (mode));
5531 fprintf
5532 (lra_dump_file,
5533 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5535 return false;
5537 new_reg = lra_create_new_reg (mode, original_reg, rclass, "split");
5538 reg_renumber[REGNO (new_reg)] = hard_regno;
5540 int new_regno = REGNO (new_reg);
5541 save = emit_spill_move (true, new_reg, original_reg);
5542 if (NEXT_INSN (save) != NULL_RTX && !call_save_p)
5544 if (lra_dump_file != NULL)
5546 fprintf
5547 (lra_dump_file,
5548 " Rejecting split %d->%d resulting in > 2 save insns:\n",
5549 original_regno, new_regno);
5550 dump_rtl_slim (lra_dump_file, save, NULL, -1, 0);
5551 fprintf (lra_dump_file,
5552 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5554 return false;
5556 restore = emit_spill_move (false, new_reg, original_reg);
5557 if (NEXT_INSN (restore) != NULL_RTX && !call_save_p)
5559 if (lra_dump_file != NULL)
5561 fprintf (lra_dump_file,
5562 " Rejecting split %d->%d "
5563 "resulting in > 2 restore insns:\n",
5564 original_regno, new_regno);
5565 dump_rtl_slim (lra_dump_file, restore, NULL, -1, 0);
5566 fprintf (lra_dump_file,
5567 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5569 return false;
5571 /* Transfer equivalence information to the spill register, so that
5572 if we fail to allocate the spill register, we have the option of
5573 rematerializing the original value instead of spilling to the stack. */
5574 if (!HARD_REGISTER_NUM_P (original_regno)
5575 && mode == PSEUDO_REGNO_MODE (original_regno))
5576 lra_copy_reg_equiv (new_regno, original_regno);
5577 after_p = usage_insns[original_regno].after_p;
5578 lra_reg_info[new_regno].restore_rtx = regno_reg_rtx[original_regno];
5579 bitmap_set_bit (&check_only_regs, new_regno);
5580 bitmap_set_bit (&check_only_regs, original_regno);
5581 bitmap_set_bit (&lra_split_regs, new_regno);
5582 for (;;)
5584 if (GET_CODE (next_usage_insns) != INSN_LIST)
5586 usage_insn = next_usage_insns;
5587 break;
5589 usage_insn = XEXP (next_usage_insns, 0);
5590 lra_assert (DEBUG_INSN_P (usage_insn));
5591 next_usage_insns = XEXP (next_usage_insns, 1);
5592 lra_substitute_pseudo (&usage_insn, original_regno, new_reg, false);
5593 lra_update_insn_regno_info (as_a <rtx_insn *> (usage_insn));
5594 if (lra_dump_file != NULL)
5596 fprintf (lra_dump_file, " Split reuse change %d->%d:\n",
5597 original_regno, new_regno);
5598 dump_insn_slim (lra_dump_file, as_a <rtx_insn *> (usage_insn));
5601 lra_assert (NOTE_P (usage_insn) || NONDEBUG_INSN_P (usage_insn));
5602 lra_assert (usage_insn != insn || (after_p && before_p));
5603 lra_process_new_insns (as_a <rtx_insn *> (usage_insn),
5604 after_p ? NULL : restore,
5605 after_p ? restore : NULL,
5606 call_save_p
5607 ? "Add reg<-save" : "Add reg<-split");
5608 lra_process_new_insns (insn, before_p ? save : NULL,
5609 before_p ? NULL : save,
5610 call_save_p
5611 ? "Add save<-reg" : "Add split<-reg");
5612 if (nregs > 1)
5613 /* If we are trying to split multi-register. We should check
5614 conflicts on the next assignment sub-pass. IRA can allocate on
5615 sub-register levels, LRA do this on pseudos level right now and
5616 this discrepancy may create allocation conflicts after
5617 splitting. */
5618 lra_risky_transformations_p = true;
5619 if (lra_dump_file != NULL)
5620 fprintf (lra_dump_file,
5621 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5622 return true;
5625 /* Recognize that we need a split transformation for insn INSN, which
5626 defines or uses REGNO in its insn biggest MODE (we use it only if
5627 REGNO is a hard register). POTENTIAL_RELOAD_HARD_REGS contains
5628 hard registers which might be used for reloads since the EBB end.
5629 Put the save before INSN if BEFORE_P is true. MAX_UID is maximla
5630 uid before starting INSN processing. Return true if we succeed in
5631 such transformation. */
5632 static bool
5633 split_if_necessary (int regno, machine_mode mode,
5634 HARD_REG_SET potential_reload_hard_regs,
5635 bool before_p, rtx_insn *insn, int max_uid)
5637 bool res = false;
5638 int i, nregs = 1;
5639 rtx next_usage_insns;
5641 if (regno < FIRST_PSEUDO_REGISTER)
5642 nregs = hard_regno_nregs[regno][mode];
5643 for (i = 0; i < nregs; i++)
5644 if (usage_insns[regno + i].check == curr_usage_insns_check
5645 && (next_usage_insns = usage_insns[regno + i].insns) != NULL_RTX
5646 /* To avoid processing the register twice or more. */
5647 && ((GET_CODE (next_usage_insns) != INSN_LIST
5648 && INSN_UID (next_usage_insns) < max_uid)
5649 || (GET_CODE (next_usage_insns) == INSN_LIST
5650 && (INSN_UID (XEXP (next_usage_insns, 0)) < max_uid)))
5651 && need_for_split_p (potential_reload_hard_regs, regno + i)
5652 && split_reg (before_p, regno + i, insn, next_usage_insns))
5653 res = true;
5654 return res;
5657 /* Return TRUE if rtx X is considered as an invariant for
5658 inheritance. */
5659 static bool
5660 invariant_p (const_rtx x)
5662 machine_mode mode;
5663 const char *fmt;
5664 enum rtx_code code;
5665 int i, j;
5667 code = GET_CODE (x);
5668 mode = GET_MODE (x);
5669 if (code == SUBREG)
5671 x = SUBREG_REG (x);
5672 code = GET_CODE (x);
5673 if (GET_MODE_SIZE (GET_MODE (x)) > GET_MODE_SIZE (mode))
5674 mode = GET_MODE (x);
5677 if (MEM_P (x))
5678 return false;
5680 if (REG_P (x))
5682 int i, nregs, regno = REGNO (x);
5684 if (regno >= FIRST_PSEUDO_REGISTER || regno == STACK_POINTER_REGNUM
5685 || TEST_HARD_REG_BIT (eliminable_regset, regno)
5686 || GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
5687 return false;
5688 nregs = hard_regno_nregs[regno][mode];
5689 for (i = 0; i < nregs; i++)
5690 if (! fixed_regs[regno + i]
5691 /* A hard register may be clobbered in the current insn
5692 but we can ignore this case because if the hard
5693 register is used it should be set somewhere after the
5694 clobber. */
5695 || bitmap_bit_p (&invalid_invariant_regs, regno + i))
5696 return false;
5698 fmt = GET_RTX_FORMAT (code);
5699 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
5701 if (fmt[i] == 'e')
5703 if (! invariant_p (XEXP (x, i)))
5704 return false;
5706 else if (fmt[i] == 'E')
5708 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
5709 if (! invariant_p (XVECEXP (x, i, j)))
5710 return false;
5713 return true;
5716 /* We have 'dest_reg <- invariant'. Let us try to make an invariant
5717 inheritance transformation (using dest_reg instead invariant in a
5718 subsequent insn). */
5719 static bool
5720 process_invariant_for_inheritance (rtx dst_reg, rtx invariant_rtx)
5722 invariant_ptr_t invariant_ptr;
5723 rtx_insn *insn, *new_insns;
5724 rtx insn_set, insn_reg, new_reg;
5725 int insn_regno;
5726 bool succ_p = false;
5727 int dst_regno = REGNO (dst_reg);
5728 machine_mode dst_mode = GET_MODE (dst_reg);
5729 enum reg_class cl = lra_get_allocno_class (dst_regno), insn_reg_cl;
5731 invariant_ptr = insert_invariant (invariant_rtx);
5732 if ((insn = invariant_ptr->insn) != NULL_RTX)
5734 /* We have a subsequent insn using the invariant. */
5735 insn_set = single_set (insn);
5736 lra_assert (insn_set != NULL);
5737 insn_reg = SET_DEST (insn_set);
5738 lra_assert (REG_P (insn_reg));
5739 insn_regno = REGNO (insn_reg);
5740 insn_reg_cl = lra_get_allocno_class (insn_regno);
5742 if (dst_mode == GET_MODE (insn_reg)
5743 /* We should consider only result move reg insns which are
5744 cheap. */
5745 && targetm.register_move_cost (dst_mode, cl, insn_reg_cl) == 2
5746 && targetm.register_move_cost (dst_mode, cl, cl) == 2)
5748 if (lra_dump_file != NULL)
5749 fprintf (lra_dump_file,
5750 " [[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[\n");
5751 new_reg = lra_create_new_reg (dst_mode, dst_reg,
5752 cl, "invariant inheritance");
5753 bitmap_set_bit (&lra_inheritance_pseudos, REGNO (new_reg));
5754 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
5755 lra_reg_info[REGNO (new_reg)].restore_rtx = PATTERN (insn);
5756 start_sequence ();
5757 lra_emit_move (new_reg, dst_reg);
5758 new_insns = get_insns ();
5759 end_sequence ();
5760 lra_process_new_insns (curr_insn, NULL, new_insns,
5761 "Add invariant inheritance<-original");
5762 start_sequence ();
5763 lra_emit_move (SET_DEST (insn_set), new_reg);
5764 new_insns = get_insns ();
5765 end_sequence ();
5766 lra_process_new_insns (insn, NULL, new_insns,
5767 "Changing reload<-inheritance");
5768 lra_set_insn_deleted (insn);
5769 succ_p = true;
5770 if (lra_dump_file != NULL)
5772 fprintf (lra_dump_file,
5773 " Invariant inheritance reuse change %d (bb%d):\n",
5774 REGNO (new_reg), BLOCK_FOR_INSN (insn)->index);
5775 dump_insn_slim (lra_dump_file, insn);
5776 fprintf (lra_dump_file,
5777 " ]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]\n");
5781 invariant_ptr->insn = curr_insn;
5782 return succ_p;
5785 /* Check only registers living at the current program point in the
5786 current EBB. */
5787 static bitmap_head live_regs;
5789 /* Update live info in EBB given by its HEAD and TAIL insns after
5790 inheritance/split transformation. The function removes dead moves
5791 too. */
5792 static void
5793 update_ebb_live_info (rtx_insn *head, rtx_insn *tail)
5795 unsigned int j;
5796 int i, regno;
5797 bool live_p;
5798 rtx_insn *prev_insn;
5799 rtx set;
5800 bool remove_p;
5801 basic_block last_bb, prev_bb, curr_bb;
5802 bitmap_iterator bi;
5803 struct lra_insn_reg *reg;
5804 edge e;
5805 edge_iterator ei;
5807 last_bb = BLOCK_FOR_INSN (tail);
5808 prev_bb = NULL;
5809 for (curr_insn = tail;
5810 curr_insn != PREV_INSN (head);
5811 curr_insn = prev_insn)
5813 prev_insn = PREV_INSN (curr_insn);
5814 /* We need to process empty blocks too. They contain
5815 NOTE_INSN_BASIC_BLOCK referring for the basic block. */
5816 if (NOTE_P (curr_insn) && NOTE_KIND (curr_insn) != NOTE_INSN_BASIC_BLOCK)
5817 continue;
5818 curr_bb = BLOCK_FOR_INSN (curr_insn);
5819 if (curr_bb != prev_bb)
5821 if (prev_bb != NULL)
5823 /* Update df_get_live_in (prev_bb): */
5824 EXECUTE_IF_SET_IN_BITMAP (&check_only_regs, 0, j, bi)
5825 if (bitmap_bit_p (&live_regs, j))
5826 bitmap_set_bit (df_get_live_in (prev_bb), j);
5827 else
5828 bitmap_clear_bit (df_get_live_in (prev_bb), j);
5830 if (curr_bb != last_bb)
5832 /* Update df_get_live_out (curr_bb): */
5833 EXECUTE_IF_SET_IN_BITMAP (&check_only_regs, 0, j, bi)
5835 live_p = bitmap_bit_p (&live_regs, j);
5836 if (! live_p)
5837 FOR_EACH_EDGE (e, ei, curr_bb->succs)
5838 if (bitmap_bit_p (df_get_live_in (e->dest), j))
5840 live_p = true;
5841 break;
5843 if (live_p)
5844 bitmap_set_bit (df_get_live_out (curr_bb), j);
5845 else
5846 bitmap_clear_bit (df_get_live_out (curr_bb), j);
5849 prev_bb = curr_bb;
5850 bitmap_and (&live_regs, &check_only_regs, df_get_live_out (curr_bb));
5852 if (! NONDEBUG_INSN_P (curr_insn))
5853 continue;
5854 curr_id = lra_get_insn_recog_data (curr_insn);
5855 curr_static_id = curr_id->insn_static_data;
5856 remove_p = false;
5857 if ((set = single_set (curr_insn)) != NULL_RTX
5858 && REG_P (SET_DEST (set))
5859 && (regno = REGNO (SET_DEST (set))) >= FIRST_PSEUDO_REGISTER
5860 && SET_DEST (set) != pic_offset_table_rtx
5861 && bitmap_bit_p (&check_only_regs, regno)
5862 && ! bitmap_bit_p (&live_regs, regno))
5863 remove_p = true;
5864 /* See which defined values die here. */
5865 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5866 if (reg->type == OP_OUT && ! reg->subreg_p)
5867 bitmap_clear_bit (&live_regs, reg->regno);
5868 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
5869 if (reg->type == OP_OUT && ! reg->subreg_p)
5870 bitmap_clear_bit (&live_regs, reg->regno);
5871 if (curr_id->arg_hard_regs != NULL)
5872 /* Make clobbered argument hard registers die. */
5873 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
5874 if (regno >= FIRST_PSEUDO_REGISTER)
5875 bitmap_clear_bit (&live_regs, regno - FIRST_PSEUDO_REGISTER);
5876 /* Mark each used value as live. */
5877 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5878 if (reg->type != OP_OUT
5879 && bitmap_bit_p (&check_only_regs, reg->regno))
5880 bitmap_set_bit (&live_regs, reg->regno);
5881 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
5882 if (reg->type != OP_OUT
5883 && bitmap_bit_p (&check_only_regs, reg->regno))
5884 bitmap_set_bit (&live_regs, reg->regno);
5885 if (curr_id->arg_hard_regs != NULL)
5886 /* Make used argument hard registers live. */
5887 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
5888 if (regno < FIRST_PSEUDO_REGISTER
5889 && bitmap_bit_p (&check_only_regs, regno))
5890 bitmap_set_bit (&live_regs, regno);
5891 /* It is quite important to remove dead move insns because it
5892 means removing dead store. We don't need to process them for
5893 constraints. */
5894 if (remove_p)
5896 if (lra_dump_file != NULL)
5898 fprintf (lra_dump_file, " Removing dead insn:\n ");
5899 dump_insn_slim (lra_dump_file, curr_insn);
5901 lra_set_insn_deleted (curr_insn);
5906 /* The structure describes info to do an inheritance for the current
5907 insn. We need to collect such info first before doing the
5908 transformations because the transformations change the insn
5909 internal representation. */
5910 struct to_inherit
5912 /* Original regno. */
5913 int regno;
5914 /* Subsequent insns which can inherit original reg value. */
5915 rtx insns;
5918 /* Array containing all info for doing inheritance from the current
5919 insn. */
5920 static struct to_inherit to_inherit[LRA_MAX_INSN_RELOADS];
5922 /* Number elements in the previous array. */
5923 static int to_inherit_num;
5925 /* Add inheritance info REGNO and INSNS. Their meaning is described in
5926 structure to_inherit. */
5927 static void
5928 add_to_inherit (int regno, rtx insns)
5930 int i;
5932 for (i = 0; i < to_inherit_num; i++)
5933 if (to_inherit[i].regno == regno)
5934 return;
5935 lra_assert (to_inherit_num < LRA_MAX_INSN_RELOADS);
5936 to_inherit[to_inherit_num].regno = regno;
5937 to_inherit[to_inherit_num++].insns = insns;
5940 /* Return the last non-debug insn in basic block BB, or the block begin
5941 note if none. */
5942 static rtx_insn *
5943 get_last_insertion_point (basic_block bb)
5945 rtx_insn *insn;
5947 FOR_BB_INSNS_REVERSE (bb, insn)
5948 if (NONDEBUG_INSN_P (insn) || NOTE_INSN_BASIC_BLOCK_P (insn))
5949 return insn;
5950 gcc_unreachable ();
5953 /* Set up RES by registers living on edges FROM except the edge (FROM,
5954 TO) or by registers set up in a jump insn in BB FROM. */
5955 static void
5956 get_live_on_other_edges (basic_block from, basic_block to, bitmap res)
5958 rtx_insn *last;
5959 struct lra_insn_reg *reg;
5960 edge e;
5961 edge_iterator ei;
5963 lra_assert (to != NULL);
5964 bitmap_clear (res);
5965 FOR_EACH_EDGE (e, ei, from->succs)
5966 if (e->dest != to)
5967 bitmap_ior_into (res, df_get_live_in (e->dest));
5968 last = get_last_insertion_point (from);
5969 if (! JUMP_P (last))
5970 return;
5971 curr_id = lra_get_insn_recog_data (last);
5972 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5973 if (reg->type != OP_IN)
5974 bitmap_set_bit (res, reg->regno);
5977 /* Used as a temporary results of some bitmap calculations. */
5978 static bitmap_head temp_bitmap;
5980 /* We split for reloads of small class of hard regs. The following
5981 defines how many hard regs the class should have to be qualified as
5982 small. The code is mostly oriented to x86/x86-64 architecture
5983 where some insns need to use only specific register or pair of
5984 registers and these register can live in RTL explicitly, e.g. for
5985 parameter passing. */
5986 static const int max_small_class_regs_num = 2;
5988 /* Do inheritance/split transformations in EBB starting with HEAD and
5989 finishing on TAIL. We process EBB insns in the reverse order.
5990 Return true if we did any inheritance/split transformation in the
5991 EBB.
5993 We should avoid excessive splitting which results in worse code
5994 because of inaccurate cost calculations for spilling new split
5995 pseudos in such case. To achieve this we do splitting only if
5996 register pressure is high in given basic block and there are reload
5997 pseudos requiring hard registers. We could do more register
5998 pressure calculations at any given program point to avoid necessary
5999 splitting even more but it is to expensive and the current approach
6000 works well enough. */
6001 static bool
6002 inherit_in_ebb (rtx_insn *head, rtx_insn *tail)
6004 int i, src_regno, dst_regno, nregs;
6005 bool change_p, succ_p, update_reloads_num_p;
6006 rtx_insn *prev_insn, *last_insn;
6007 rtx next_usage_insns, curr_set;
6008 enum reg_class cl;
6009 struct lra_insn_reg *reg;
6010 basic_block last_processed_bb, curr_bb = NULL;
6011 HARD_REG_SET potential_reload_hard_regs, live_hard_regs;
6012 bitmap to_process;
6013 unsigned int j;
6014 bitmap_iterator bi;
6015 bool head_p, after_p;
6017 change_p = false;
6018 curr_usage_insns_check++;
6019 clear_invariants ();
6020 reloads_num = calls_num = 0;
6021 bitmap_clear (&check_only_regs);
6022 bitmap_clear (&invalid_invariant_regs);
6023 last_processed_bb = NULL;
6024 CLEAR_HARD_REG_SET (potential_reload_hard_regs);
6025 COPY_HARD_REG_SET (live_hard_regs, eliminable_regset);
6026 IOR_HARD_REG_SET (live_hard_regs, lra_no_alloc_regs);
6027 /* We don't process new insns generated in the loop. */
6028 for (curr_insn = tail; curr_insn != PREV_INSN (head); curr_insn = prev_insn)
6030 prev_insn = PREV_INSN (curr_insn);
6031 if (BLOCK_FOR_INSN (curr_insn) != NULL)
6032 curr_bb = BLOCK_FOR_INSN (curr_insn);
6033 if (last_processed_bb != curr_bb)
6035 /* We are at the end of BB. Add qualified living
6036 pseudos for potential splitting. */
6037 to_process = df_get_live_out (curr_bb);
6038 if (last_processed_bb != NULL)
6040 /* We are somewhere in the middle of EBB. */
6041 get_live_on_other_edges (curr_bb, last_processed_bb,
6042 &temp_bitmap);
6043 to_process = &temp_bitmap;
6045 last_processed_bb = curr_bb;
6046 last_insn = get_last_insertion_point (curr_bb);
6047 after_p = (! JUMP_P (last_insn)
6048 && (! CALL_P (last_insn)
6049 || (find_reg_note (last_insn,
6050 REG_NORETURN, NULL_RTX) == NULL_RTX
6051 && ! SIBLING_CALL_P (last_insn))));
6052 CLEAR_HARD_REG_SET (potential_reload_hard_regs);
6053 EXECUTE_IF_SET_IN_BITMAP (to_process, 0, j, bi)
6055 if ((int) j >= lra_constraint_new_regno_start)
6056 break;
6057 if (j < FIRST_PSEUDO_REGISTER || reg_renumber[j] >= 0)
6059 if (j < FIRST_PSEUDO_REGISTER)
6060 SET_HARD_REG_BIT (live_hard_regs, j);
6061 else
6062 add_to_hard_reg_set (&live_hard_regs,
6063 PSEUDO_REGNO_MODE (j),
6064 reg_renumber[j]);
6065 setup_next_usage_insn (j, last_insn, reloads_num, after_p);
6069 src_regno = dst_regno = -1;
6070 curr_set = single_set (curr_insn);
6071 if (curr_set != NULL_RTX && REG_P (SET_DEST (curr_set)))
6072 dst_regno = REGNO (SET_DEST (curr_set));
6073 if (curr_set != NULL_RTX && REG_P (SET_SRC (curr_set)))
6074 src_regno = REGNO (SET_SRC (curr_set));
6075 update_reloads_num_p = true;
6076 if (src_regno < lra_constraint_new_regno_start
6077 && src_regno >= FIRST_PSEUDO_REGISTER
6078 && reg_renumber[src_regno] < 0
6079 && dst_regno >= lra_constraint_new_regno_start
6080 && (cl = lra_get_allocno_class (dst_regno)) != NO_REGS)
6082 /* 'reload_pseudo <- original_pseudo'. */
6083 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6084 reloads_num++;
6085 update_reloads_num_p = false;
6086 succ_p = false;
6087 if (usage_insns[src_regno].check == curr_usage_insns_check
6088 && (next_usage_insns = usage_insns[src_regno].insns) != NULL_RTX)
6089 succ_p = inherit_reload_reg (false, src_regno, cl,
6090 curr_insn, next_usage_insns);
6091 if (succ_p)
6092 change_p = true;
6093 else
6094 setup_next_usage_insn (src_regno, curr_insn, reloads_num, false);
6095 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6096 IOR_HARD_REG_SET (potential_reload_hard_regs,
6097 reg_class_contents[cl]);
6099 else if (src_regno < 0
6100 && dst_regno >= lra_constraint_new_regno_start
6101 && invariant_p (SET_SRC (curr_set))
6102 && (cl = lra_get_allocno_class (dst_regno)) != NO_REGS
6103 && ! bitmap_bit_p (&invalid_invariant_regs, dst_regno)
6104 && ! bitmap_bit_p (&invalid_invariant_regs,
6105 ORIGINAL_REGNO(regno_reg_rtx[dst_regno])))
6107 /* 'reload_pseudo <- invariant'. */
6108 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6109 reloads_num++;
6110 update_reloads_num_p = false;
6111 if (process_invariant_for_inheritance (SET_DEST (curr_set), SET_SRC (curr_set)))
6112 change_p = true;
6113 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6114 IOR_HARD_REG_SET (potential_reload_hard_regs,
6115 reg_class_contents[cl]);
6117 else if (src_regno >= lra_constraint_new_regno_start
6118 && dst_regno < lra_constraint_new_regno_start
6119 && dst_regno >= FIRST_PSEUDO_REGISTER
6120 && reg_renumber[dst_regno] < 0
6121 && (cl = lra_get_allocno_class (src_regno)) != NO_REGS
6122 && usage_insns[dst_regno].check == curr_usage_insns_check
6123 && (next_usage_insns
6124 = usage_insns[dst_regno].insns) != NULL_RTX)
6126 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6127 reloads_num++;
6128 update_reloads_num_p = false;
6129 /* 'original_pseudo <- reload_pseudo'. */
6130 if (! JUMP_P (curr_insn)
6131 && inherit_reload_reg (true, dst_regno, cl,
6132 curr_insn, next_usage_insns))
6133 change_p = true;
6134 /* Invalidate. */
6135 usage_insns[dst_regno].check = 0;
6136 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6137 IOR_HARD_REG_SET (potential_reload_hard_regs,
6138 reg_class_contents[cl]);
6140 else if (INSN_P (curr_insn))
6142 int iter;
6143 int max_uid = get_max_uid ();
6145 curr_id = lra_get_insn_recog_data (curr_insn);
6146 curr_static_id = curr_id->insn_static_data;
6147 to_inherit_num = 0;
6148 /* Process insn definitions. */
6149 for (iter = 0; iter < 2; iter++)
6150 for (reg = iter == 0 ? curr_id->regs : curr_static_id->hard_regs;
6151 reg != NULL;
6152 reg = reg->next)
6153 if (reg->type != OP_IN
6154 && (dst_regno = reg->regno) < lra_constraint_new_regno_start)
6156 if (dst_regno >= FIRST_PSEUDO_REGISTER && reg->type == OP_OUT
6157 && reg_renumber[dst_regno] < 0 && ! reg->subreg_p
6158 && usage_insns[dst_regno].check == curr_usage_insns_check
6159 && (next_usage_insns
6160 = usage_insns[dst_regno].insns) != NULL_RTX)
6162 struct lra_insn_reg *r;
6164 for (r = curr_id->regs; r != NULL; r = r->next)
6165 if (r->type != OP_OUT && r->regno == dst_regno)
6166 break;
6167 /* Don't do inheritance if the pseudo is also
6168 used in the insn. */
6169 if (r == NULL)
6170 /* We can not do inheritance right now
6171 because the current insn reg info (chain
6172 regs) can change after that. */
6173 add_to_inherit (dst_regno, next_usage_insns);
6175 /* We can not process one reg twice here because of
6176 usage_insns invalidation. */
6177 if ((dst_regno < FIRST_PSEUDO_REGISTER
6178 || reg_renumber[dst_regno] >= 0)
6179 && ! reg->subreg_p && reg->type != OP_IN)
6181 HARD_REG_SET s;
6183 if (split_if_necessary (dst_regno, reg->biggest_mode,
6184 potential_reload_hard_regs,
6185 false, curr_insn, max_uid))
6186 change_p = true;
6187 CLEAR_HARD_REG_SET (s);
6188 if (dst_regno < FIRST_PSEUDO_REGISTER)
6189 add_to_hard_reg_set (&s, reg->biggest_mode, dst_regno);
6190 else
6191 add_to_hard_reg_set (&s, PSEUDO_REGNO_MODE (dst_regno),
6192 reg_renumber[dst_regno]);
6193 AND_COMPL_HARD_REG_SET (live_hard_regs, s);
6195 /* We should invalidate potential inheritance or
6196 splitting for the current insn usages to the next
6197 usage insns (see code below) as the output pseudo
6198 prevents this. */
6199 if ((dst_regno >= FIRST_PSEUDO_REGISTER
6200 && reg_renumber[dst_regno] < 0)
6201 || (reg->type == OP_OUT && ! reg->subreg_p
6202 && (dst_regno < FIRST_PSEUDO_REGISTER
6203 || reg_renumber[dst_regno] >= 0)))
6205 /* Invalidate and mark definitions. */
6206 if (dst_regno >= FIRST_PSEUDO_REGISTER)
6207 usage_insns[dst_regno].check = -(int) INSN_UID (curr_insn);
6208 else
6210 nregs = hard_regno_nregs[dst_regno][reg->biggest_mode];
6211 for (i = 0; i < nregs; i++)
6212 usage_insns[dst_regno + i].check
6213 = -(int) INSN_UID (curr_insn);
6217 /* Process clobbered call regs. */
6218 if (curr_id->arg_hard_regs != NULL)
6219 for (i = 0; (dst_regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6220 if (dst_regno >= FIRST_PSEUDO_REGISTER)
6221 usage_insns[dst_regno - FIRST_PSEUDO_REGISTER].check
6222 = -(int) INSN_UID (curr_insn);
6223 if (! JUMP_P (curr_insn))
6224 for (i = 0; i < to_inherit_num; i++)
6225 if (inherit_reload_reg (true, to_inherit[i].regno,
6226 ALL_REGS, curr_insn,
6227 to_inherit[i].insns))
6228 change_p = true;
6229 if (CALL_P (curr_insn))
6231 rtx cheap, pat, dest;
6232 rtx_insn *restore;
6233 int regno, hard_regno;
6235 calls_num++;
6236 if ((cheap = find_reg_note (curr_insn,
6237 REG_RETURNED, NULL_RTX)) != NULL_RTX
6238 && ((cheap = XEXP (cheap, 0)), true)
6239 && (regno = REGNO (cheap)) >= FIRST_PSEUDO_REGISTER
6240 && (hard_regno = reg_renumber[regno]) >= 0
6241 /* If there are pending saves/restores, the
6242 optimization is not worth. */
6243 && usage_insns[regno].calls_num == calls_num - 1
6244 && TEST_HARD_REG_BIT (call_used_reg_set, hard_regno))
6246 /* Restore the pseudo from the call result as
6247 REG_RETURNED note says that the pseudo value is
6248 in the call result and the pseudo is an argument
6249 of the call. */
6250 pat = PATTERN (curr_insn);
6251 if (GET_CODE (pat) == PARALLEL)
6252 pat = XVECEXP (pat, 0, 0);
6253 dest = SET_DEST (pat);
6254 /* For multiple return values dest is PARALLEL.
6255 Currently we handle only single return value case. */
6256 if (REG_P (dest))
6258 start_sequence ();
6259 emit_move_insn (cheap, copy_rtx (dest));
6260 restore = get_insns ();
6261 end_sequence ();
6262 lra_process_new_insns (curr_insn, NULL, restore,
6263 "Inserting call parameter restore");
6264 /* We don't need to save/restore of the pseudo from
6265 this call. */
6266 usage_insns[regno].calls_num = calls_num;
6267 bitmap_set_bit (&check_only_regs, regno);
6271 to_inherit_num = 0;
6272 /* Process insn usages. */
6273 for (iter = 0; iter < 2; iter++)
6274 for (reg = iter == 0 ? curr_id->regs : curr_static_id->hard_regs;
6275 reg != NULL;
6276 reg = reg->next)
6277 if ((reg->type != OP_OUT
6278 || (reg->type == OP_OUT && reg->subreg_p))
6279 && (src_regno = reg->regno) < lra_constraint_new_regno_start)
6281 if (src_regno >= FIRST_PSEUDO_REGISTER
6282 && reg_renumber[src_regno] < 0 && reg->type == OP_IN)
6284 if (usage_insns[src_regno].check == curr_usage_insns_check
6285 && (next_usage_insns
6286 = usage_insns[src_regno].insns) != NULL_RTX
6287 && NONDEBUG_INSN_P (curr_insn))
6288 add_to_inherit (src_regno, next_usage_insns);
6289 else if (usage_insns[src_regno].check
6290 != -(int) INSN_UID (curr_insn))
6291 /* Add usages but only if the reg is not set up
6292 in the same insn. */
6293 add_next_usage_insn (src_regno, curr_insn, reloads_num);
6295 else if (src_regno < FIRST_PSEUDO_REGISTER
6296 || reg_renumber[src_regno] >= 0)
6298 bool before_p;
6299 rtx_insn *use_insn = curr_insn;
6301 before_p = (JUMP_P (curr_insn)
6302 || (CALL_P (curr_insn) && reg->type == OP_IN));
6303 if (NONDEBUG_INSN_P (curr_insn)
6304 && (! JUMP_P (curr_insn) || reg->type == OP_IN)
6305 && split_if_necessary (src_regno, reg->biggest_mode,
6306 potential_reload_hard_regs,
6307 before_p, curr_insn, max_uid))
6309 if (reg->subreg_p)
6310 lra_risky_transformations_p = true;
6311 change_p = true;
6312 /* Invalidate. */
6313 usage_insns[src_regno].check = 0;
6314 if (before_p)
6315 use_insn = PREV_INSN (curr_insn);
6317 if (NONDEBUG_INSN_P (curr_insn))
6319 if (src_regno < FIRST_PSEUDO_REGISTER)
6320 add_to_hard_reg_set (&live_hard_regs,
6321 reg->biggest_mode, src_regno);
6322 else
6323 add_to_hard_reg_set (&live_hard_regs,
6324 PSEUDO_REGNO_MODE (src_regno),
6325 reg_renumber[src_regno]);
6327 add_next_usage_insn (src_regno, use_insn, reloads_num);
6330 /* Process used call regs. */
6331 if (curr_id->arg_hard_regs != NULL)
6332 for (i = 0; (src_regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6333 if (src_regno < FIRST_PSEUDO_REGISTER)
6335 SET_HARD_REG_BIT (live_hard_regs, src_regno);
6336 add_next_usage_insn (src_regno, curr_insn, reloads_num);
6338 for (i = 0; i < to_inherit_num; i++)
6340 src_regno = to_inherit[i].regno;
6341 if (inherit_reload_reg (false, src_regno, ALL_REGS,
6342 curr_insn, to_inherit[i].insns))
6343 change_p = true;
6344 else
6345 setup_next_usage_insn (src_regno, curr_insn, reloads_num, false);
6348 if (update_reloads_num_p
6349 && NONDEBUG_INSN_P (curr_insn) && curr_set != NULL_RTX)
6351 int regno = -1;
6352 if ((REG_P (SET_DEST (curr_set))
6353 && (regno = REGNO (SET_DEST (curr_set))) >= lra_constraint_new_regno_start
6354 && reg_renumber[regno] < 0
6355 && (cl = lra_get_allocno_class (regno)) != NO_REGS)
6356 || (REG_P (SET_SRC (curr_set))
6357 && (regno = REGNO (SET_SRC (curr_set))) >= lra_constraint_new_regno_start
6358 && reg_renumber[regno] < 0
6359 && (cl = lra_get_allocno_class (regno)) != NO_REGS))
6361 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6362 reloads_num++;
6363 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6364 IOR_HARD_REG_SET (potential_reload_hard_regs,
6365 reg_class_contents[cl]);
6368 if (NONDEBUG_INSN_P (curr_insn))
6370 int regno;
6372 /* Invalidate invariants with changed regs. */
6373 curr_id = lra_get_insn_recog_data (curr_insn);
6374 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
6375 if (reg->type != OP_IN)
6377 bitmap_set_bit (&invalid_invariant_regs, reg->regno);
6378 bitmap_set_bit (&invalid_invariant_regs,
6379 ORIGINAL_REGNO (regno_reg_rtx[reg->regno]));
6381 curr_static_id = curr_id->insn_static_data;
6382 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
6383 if (reg->type != OP_IN)
6384 bitmap_set_bit (&invalid_invariant_regs, reg->regno);
6385 if (curr_id->arg_hard_regs != NULL)
6386 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6387 if (regno >= FIRST_PSEUDO_REGISTER)
6388 bitmap_set_bit (&invalid_invariant_regs,
6389 regno - FIRST_PSEUDO_REGISTER);
6391 /* We reached the start of the current basic block. */
6392 if (prev_insn == NULL_RTX || prev_insn == PREV_INSN (head)
6393 || BLOCK_FOR_INSN (prev_insn) != curr_bb)
6395 /* We reached the beginning of the current block -- do
6396 rest of spliting in the current BB. */
6397 to_process = df_get_live_in (curr_bb);
6398 if (BLOCK_FOR_INSN (head) != curr_bb)
6400 /* We are somewhere in the middle of EBB. */
6401 get_live_on_other_edges (EDGE_PRED (curr_bb, 0)->src,
6402 curr_bb, &temp_bitmap);
6403 to_process = &temp_bitmap;
6405 head_p = true;
6406 EXECUTE_IF_SET_IN_BITMAP (to_process, 0, j, bi)
6408 if ((int) j >= lra_constraint_new_regno_start)
6409 break;
6410 if (((int) j < FIRST_PSEUDO_REGISTER || reg_renumber[j] >= 0)
6411 && usage_insns[j].check == curr_usage_insns_check
6412 && (next_usage_insns = usage_insns[j].insns) != NULL_RTX)
6414 if (need_for_split_p (potential_reload_hard_regs, j))
6416 if (lra_dump_file != NULL && head_p)
6418 fprintf (lra_dump_file,
6419 " ----------------------------------\n");
6420 head_p = false;
6422 if (split_reg (false, j, bb_note (curr_bb),
6423 next_usage_insns))
6424 change_p = true;
6426 usage_insns[j].check = 0;
6431 return change_p;
6434 /* This value affects EBB forming. If probability of edge from EBB to
6435 a BB is not greater than the following value, we don't add the BB
6436 to EBB. */
6437 #define EBB_PROBABILITY_CUTOFF \
6438 ((REG_BR_PROB_BASE * LRA_INHERITANCE_EBB_PROBABILITY_CUTOFF) / 100)
6440 /* Current number of inheritance/split iteration. */
6441 int lra_inheritance_iter;
6443 /* Entry function for inheritance/split pass. */
6444 void
6445 lra_inheritance (void)
6447 int i;
6448 basic_block bb, start_bb;
6449 edge e;
6451 lra_inheritance_iter++;
6452 if (lra_inheritance_iter > LRA_MAX_INHERITANCE_PASSES)
6453 return;
6454 timevar_push (TV_LRA_INHERITANCE);
6455 if (lra_dump_file != NULL)
6456 fprintf (lra_dump_file, "\n********** Inheritance #%d: **********\n\n",
6457 lra_inheritance_iter);
6458 curr_usage_insns_check = 0;
6459 usage_insns = XNEWVEC (struct usage_insns, lra_constraint_new_regno_start);
6460 for (i = 0; i < lra_constraint_new_regno_start; i++)
6461 usage_insns[i].check = 0;
6462 bitmap_initialize (&check_only_regs, &reg_obstack);
6463 bitmap_initialize (&invalid_invariant_regs, &reg_obstack);
6464 bitmap_initialize (&live_regs, &reg_obstack);
6465 bitmap_initialize (&temp_bitmap, &reg_obstack);
6466 bitmap_initialize (&ebb_global_regs, &reg_obstack);
6467 FOR_EACH_BB_FN (bb, cfun)
6469 start_bb = bb;
6470 if (lra_dump_file != NULL)
6471 fprintf (lra_dump_file, "EBB");
6472 /* Form a EBB starting with BB. */
6473 bitmap_clear (&ebb_global_regs);
6474 bitmap_ior_into (&ebb_global_regs, df_get_live_in (bb));
6475 for (;;)
6477 if (lra_dump_file != NULL)
6478 fprintf (lra_dump_file, " %d", bb->index);
6479 if (bb->next_bb == EXIT_BLOCK_PTR_FOR_FN (cfun)
6480 || LABEL_P (BB_HEAD (bb->next_bb)))
6481 break;
6482 e = find_fallthru_edge (bb->succs);
6483 if (! e)
6484 break;
6485 if (e->probability.initialized_p ()
6486 && e->probability.to_reg_br_prob_base () < EBB_PROBABILITY_CUTOFF)
6487 break;
6488 bb = bb->next_bb;
6490 bitmap_ior_into (&ebb_global_regs, df_get_live_out (bb));
6491 if (lra_dump_file != NULL)
6492 fprintf (lra_dump_file, "\n");
6493 if (inherit_in_ebb (BB_HEAD (start_bb), BB_END (bb)))
6494 /* Remember that the EBB head and tail can change in
6495 inherit_in_ebb. */
6496 update_ebb_live_info (BB_HEAD (start_bb), BB_END (bb));
6498 bitmap_clear (&ebb_global_regs);
6499 bitmap_clear (&temp_bitmap);
6500 bitmap_clear (&live_regs);
6501 bitmap_clear (&invalid_invariant_regs);
6502 bitmap_clear (&check_only_regs);
6503 free (usage_insns);
6505 timevar_pop (TV_LRA_INHERITANCE);
6510 /* This page contains code to undo failed inheritance/split
6511 transformations. */
6513 /* Current number of iteration undoing inheritance/split. */
6514 int lra_undo_inheritance_iter;
6516 /* Fix BB live info LIVE after removing pseudos created on pass doing
6517 inheritance/split which are REMOVED_PSEUDOS. */
6518 static void
6519 fix_bb_live_info (bitmap live, bitmap removed_pseudos)
6521 unsigned int regno;
6522 bitmap_iterator bi;
6524 EXECUTE_IF_SET_IN_BITMAP (removed_pseudos, 0, regno, bi)
6525 if (bitmap_clear_bit (live, regno)
6526 && REG_P (lra_reg_info[regno].restore_rtx))
6527 bitmap_set_bit (live, REGNO (lra_reg_info[regno].restore_rtx));
6530 /* Return regno of the (subreg of) REG. Otherwise, return a negative
6531 number. */
6532 static int
6533 get_regno (rtx reg)
6535 if (GET_CODE (reg) == SUBREG)
6536 reg = SUBREG_REG (reg);
6537 if (REG_P (reg))
6538 return REGNO (reg);
6539 return -1;
6542 /* Delete a move INSN with destination reg DREGNO and a previous
6543 clobber insn with the same regno. The inheritance/split code can
6544 generate moves with preceding clobber and when we delete such moves
6545 we should delete the clobber insn too to keep the correct life
6546 info. */
6547 static void
6548 delete_move_and_clobber (rtx_insn *insn, int dregno)
6550 rtx_insn *prev_insn = PREV_INSN (insn);
6552 lra_set_insn_deleted (insn);
6553 lra_assert (dregno >= 0);
6554 if (prev_insn != NULL && NONDEBUG_INSN_P (prev_insn)
6555 && GET_CODE (PATTERN (prev_insn)) == CLOBBER
6556 && dregno == get_regno (XEXP (PATTERN (prev_insn), 0)))
6557 lra_set_insn_deleted (prev_insn);
6560 /* Remove inheritance/split pseudos which are in REMOVE_PSEUDOS and
6561 return true if we did any change. The undo transformations for
6562 inheritance looks like
6563 i <- i2
6564 p <- i => p <- i2
6565 or removing
6566 p <- i, i <- p, and i <- i3
6567 where p is original pseudo from which inheritance pseudo i was
6568 created, i and i3 are removed inheritance pseudos, i2 is another
6569 not removed inheritance pseudo. All split pseudos or other
6570 occurrences of removed inheritance pseudos are changed on the
6571 corresponding original pseudos.
6573 The function also schedules insns changed and created during
6574 inheritance/split pass for processing by the subsequent constraint
6575 pass. */
6576 static bool
6577 remove_inheritance_pseudos (bitmap remove_pseudos)
6579 basic_block bb;
6580 int regno, sregno, prev_sregno, dregno;
6581 rtx restore_rtx;
6582 rtx set, prev_set;
6583 rtx_insn *prev_insn;
6584 bool change_p, done_p;
6586 change_p = ! bitmap_empty_p (remove_pseudos);
6587 /* We can not finish the function right away if CHANGE_P is true
6588 because we need to marks insns affected by previous
6589 inheritance/split pass for processing by the subsequent
6590 constraint pass. */
6591 FOR_EACH_BB_FN (bb, cfun)
6593 fix_bb_live_info (df_get_live_in (bb), remove_pseudos);
6594 fix_bb_live_info (df_get_live_out (bb), remove_pseudos);
6595 FOR_BB_INSNS_REVERSE (bb, curr_insn)
6597 if (! INSN_P (curr_insn))
6598 continue;
6599 done_p = false;
6600 sregno = dregno = -1;
6601 if (change_p && NONDEBUG_INSN_P (curr_insn)
6602 && (set = single_set (curr_insn)) != NULL_RTX)
6604 dregno = get_regno (SET_DEST (set));
6605 sregno = get_regno (SET_SRC (set));
6608 if (sregno >= 0 && dregno >= 0)
6610 if (bitmap_bit_p (remove_pseudos, dregno)
6611 && ! REG_P (lra_reg_info[dregno].restore_rtx))
6613 /* invariant inheritance pseudo <- original pseudo */
6614 if (lra_dump_file != NULL)
6616 fprintf (lra_dump_file, " Removing invariant inheritance:\n");
6617 dump_insn_slim (lra_dump_file, curr_insn);
6618 fprintf (lra_dump_file, "\n");
6620 delete_move_and_clobber (curr_insn, dregno);
6621 done_p = true;
6623 else if (bitmap_bit_p (remove_pseudos, sregno)
6624 && ! REG_P (lra_reg_info[sregno].restore_rtx))
6626 /* reload pseudo <- invariant inheritance pseudo */
6627 start_sequence ();
6628 /* We can not just change the source. It might be
6629 an insn different from the move. */
6630 emit_insn (lra_reg_info[sregno].restore_rtx);
6631 rtx_insn *new_insns = get_insns ();
6632 end_sequence ();
6633 lra_assert (single_set (new_insns) != NULL
6634 && SET_DEST (set) == SET_DEST (single_set (new_insns)));
6635 lra_process_new_insns (curr_insn, NULL, new_insns,
6636 "Changing reload<-invariant inheritance");
6637 delete_move_and_clobber (curr_insn, dregno);
6638 done_p = true;
6640 else if ((bitmap_bit_p (remove_pseudos, sregno)
6641 && (get_regno (lra_reg_info[sregno].restore_rtx) == dregno
6642 || (bitmap_bit_p (remove_pseudos, dregno)
6643 && get_regno (lra_reg_info[sregno].restore_rtx) >= 0
6644 && (get_regno (lra_reg_info[sregno].restore_rtx)
6645 == get_regno (lra_reg_info[dregno].restore_rtx)))))
6646 || (bitmap_bit_p (remove_pseudos, dregno)
6647 && get_regno (lra_reg_info[dregno].restore_rtx) == sregno))
6648 /* One of the following cases:
6649 original <- removed inheritance pseudo
6650 removed inherit pseudo <- another removed inherit pseudo
6651 removed inherit pseudo <- original pseudo
6653 removed_split_pseudo <- original_reg
6654 original_reg <- removed_split_pseudo */
6656 if (lra_dump_file != NULL)
6658 fprintf (lra_dump_file, " Removing %s:\n",
6659 bitmap_bit_p (&lra_split_regs, sregno)
6660 || bitmap_bit_p (&lra_split_regs, dregno)
6661 ? "split" : "inheritance");
6662 dump_insn_slim (lra_dump_file, curr_insn);
6664 delete_move_and_clobber (curr_insn, dregno);
6665 done_p = true;
6667 else if (bitmap_bit_p (remove_pseudos, sregno)
6668 && bitmap_bit_p (&lra_inheritance_pseudos, sregno))
6670 /* Search the following pattern:
6671 inherit_or_split_pseudo1 <- inherit_or_split_pseudo2
6672 original_pseudo <- inherit_or_split_pseudo1
6673 where the 2nd insn is the current insn and
6674 inherit_or_split_pseudo2 is not removed. If it is found,
6675 change the current insn onto:
6676 original_pseudo <- inherit_or_split_pseudo2. */
6677 for (prev_insn = PREV_INSN (curr_insn);
6678 prev_insn != NULL_RTX && ! NONDEBUG_INSN_P (prev_insn);
6679 prev_insn = PREV_INSN (prev_insn))
6681 if (prev_insn != NULL_RTX && BLOCK_FOR_INSN (prev_insn) == bb
6682 && (prev_set = single_set (prev_insn)) != NULL_RTX
6683 /* There should be no subregs in insn we are
6684 searching because only the original reg might
6685 be in subreg when we changed the mode of
6686 load/store for splitting. */
6687 && REG_P (SET_DEST (prev_set))
6688 && REG_P (SET_SRC (prev_set))
6689 && (int) REGNO (SET_DEST (prev_set)) == sregno
6690 && ((prev_sregno = REGNO (SET_SRC (prev_set)))
6691 >= FIRST_PSEUDO_REGISTER)
6692 && (lra_reg_info[prev_sregno].restore_rtx == NULL_RTX
6694 /* As we consider chain of inheritance or
6695 splitting described in above comment we should
6696 check that sregno and prev_sregno were
6697 inheritance/split pseudos created from the
6698 same original regno. */
6699 (get_regno (lra_reg_info[sregno].restore_rtx) >= 0
6700 && (get_regno (lra_reg_info[sregno].restore_rtx)
6701 == get_regno (lra_reg_info[prev_sregno].restore_rtx))))
6702 && ! bitmap_bit_p (remove_pseudos, prev_sregno))
6704 lra_assert (GET_MODE (SET_SRC (prev_set))
6705 == GET_MODE (regno_reg_rtx[sregno]));
6706 if (GET_CODE (SET_SRC (set)) == SUBREG)
6707 SUBREG_REG (SET_SRC (set)) = SET_SRC (prev_set);
6708 else
6709 SET_SRC (set) = SET_SRC (prev_set);
6710 /* As we are finishing with processing the insn
6711 here, check the destination too as it might
6712 inheritance pseudo for another pseudo. */
6713 if (bitmap_bit_p (remove_pseudos, dregno)
6714 && bitmap_bit_p (&lra_inheritance_pseudos, dregno)
6715 && (restore_rtx
6716 = lra_reg_info[dregno].restore_rtx) != NULL_RTX)
6718 if (GET_CODE (SET_DEST (set)) == SUBREG)
6719 SUBREG_REG (SET_DEST (set)) = restore_rtx;
6720 else
6721 SET_DEST (set) = restore_rtx;
6723 lra_push_insn_and_update_insn_regno_info (curr_insn);
6724 lra_set_used_insn_alternative_by_uid
6725 (INSN_UID (curr_insn), -1);
6726 done_p = true;
6727 if (lra_dump_file != NULL)
6729 fprintf (lra_dump_file, " Change reload insn:\n");
6730 dump_insn_slim (lra_dump_file, curr_insn);
6735 if (! done_p)
6737 struct lra_insn_reg *reg;
6738 bool restored_regs_p = false;
6739 bool kept_regs_p = false;
6741 curr_id = lra_get_insn_recog_data (curr_insn);
6742 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
6744 regno = reg->regno;
6745 restore_rtx = lra_reg_info[regno].restore_rtx;
6746 if (restore_rtx != NULL_RTX)
6748 if (change_p && bitmap_bit_p (remove_pseudos, regno))
6750 lra_substitute_pseudo_within_insn
6751 (curr_insn, regno, restore_rtx, false);
6752 restored_regs_p = true;
6754 else
6755 kept_regs_p = true;
6758 if (NONDEBUG_INSN_P (curr_insn) && kept_regs_p)
6760 /* The instruction has changed since the previous
6761 constraints pass. */
6762 lra_push_insn_and_update_insn_regno_info (curr_insn);
6763 lra_set_used_insn_alternative_by_uid
6764 (INSN_UID (curr_insn), -1);
6766 else if (restored_regs_p)
6767 /* The instruction has been restored to the form that
6768 it had during the previous constraints pass. */
6769 lra_update_insn_regno_info (curr_insn);
6770 if (restored_regs_p && lra_dump_file != NULL)
6772 fprintf (lra_dump_file, " Insn after restoring regs:\n");
6773 dump_insn_slim (lra_dump_file, curr_insn);
6778 return change_p;
6781 /* If optional reload pseudos failed to get a hard register or was not
6782 inherited, it is better to remove optional reloads. We do this
6783 transformation after undoing inheritance to figure out necessity to
6784 remove optional reloads easier. Return true if we do any
6785 change. */
6786 static bool
6787 undo_optional_reloads (void)
6789 bool change_p, keep_p;
6790 unsigned int regno, uid;
6791 bitmap_iterator bi, bi2;
6792 rtx_insn *insn;
6793 rtx set, src, dest;
6794 auto_bitmap removed_optional_reload_pseudos (&reg_obstack);
6796 bitmap_copy (removed_optional_reload_pseudos, &lra_optional_reload_pseudos);
6797 EXECUTE_IF_SET_IN_BITMAP (&lra_optional_reload_pseudos, 0, regno, bi)
6799 keep_p = false;
6800 /* Keep optional reloads from previous subpasses. */
6801 if (lra_reg_info[regno].restore_rtx == NULL_RTX
6802 /* If the original pseudo changed its allocation, just
6803 removing the optional pseudo is dangerous as the original
6804 pseudo will have longer live range. */
6805 || reg_renumber[REGNO (lra_reg_info[regno].restore_rtx)] >= 0)
6806 keep_p = true;
6807 else if (reg_renumber[regno] >= 0)
6808 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi2)
6810 insn = lra_insn_recog_data[uid]->insn;
6811 if ((set = single_set (insn)) == NULL_RTX)
6812 continue;
6813 src = SET_SRC (set);
6814 dest = SET_DEST (set);
6815 if (! REG_P (src) || ! REG_P (dest))
6816 continue;
6817 if (REGNO (dest) == regno
6818 /* Ignore insn for optional reloads itself. */
6819 && REGNO (lra_reg_info[regno].restore_rtx) != REGNO (src)
6820 /* Check only inheritance on last inheritance pass. */
6821 && (int) REGNO (src) >= new_regno_start
6822 /* Check that the optional reload was inherited. */
6823 && bitmap_bit_p (&lra_inheritance_pseudos, REGNO (src)))
6825 keep_p = true;
6826 break;
6829 if (keep_p)
6831 bitmap_clear_bit (removed_optional_reload_pseudos, regno);
6832 if (lra_dump_file != NULL)
6833 fprintf (lra_dump_file, "Keep optional reload reg %d\n", regno);
6836 change_p = ! bitmap_empty_p (removed_optional_reload_pseudos);
6837 auto_bitmap insn_bitmap (&reg_obstack);
6838 EXECUTE_IF_SET_IN_BITMAP (removed_optional_reload_pseudos, 0, regno, bi)
6840 if (lra_dump_file != NULL)
6841 fprintf (lra_dump_file, "Remove optional reload reg %d\n", regno);
6842 bitmap_copy (insn_bitmap, &lra_reg_info[regno].insn_bitmap);
6843 EXECUTE_IF_SET_IN_BITMAP (insn_bitmap, 0, uid, bi2)
6845 insn = lra_insn_recog_data[uid]->insn;
6846 if ((set = single_set (insn)) != NULL_RTX)
6848 src = SET_SRC (set);
6849 dest = SET_DEST (set);
6850 if (REG_P (src) && REG_P (dest)
6851 && ((REGNO (src) == regno
6852 && (REGNO (lra_reg_info[regno].restore_rtx)
6853 == REGNO (dest)))
6854 || (REGNO (dest) == regno
6855 && (REGNO (lra_reg_info[regno].restore_rtx)
6856 == REGNO (src)))))
6858 if (lra_dump_file != NULL)
6860 fprintf (lra_dump_file, " Deleting move %u\n",
6861 INSN_UID (insn));
6862 dump_insn_slim (lra_dump_file, insn);
6864 delete_move_and_clobber (insn, REGNO (dest));
6865 continue;
6867 /* We should not worry about generation memory-memory
6868 moves here as if the corresponding inheritance did
6869 not work (inheritance pseudo did not get a hard reg),
6870 we remove the inheritance pseudo and the optional
6871 reload. */
6873 lra_substitute_pseudo_within_insn
6874 (insn, regno, lra_reg_info[regno].restore_rtx, false);
6875 lra_update_insn_regno_info (insn);
6876 if (lra_dump_file != NULL)
6878 fprintf (lra_dump_file,
6879 " Restoring original insn:\n");
6880 dump_insn_slim (lra_dump_file, insn);
6884 /* Clear restore_regnos. */
6885 EXECUTE_IF_SET_IN_BITMAP (&lra_optional_reload_pseudos, 0, regno, bi)
6886 lra_reg_info[regno].restore_rtx = NULL_RTX;
6887 return change_p;
6890 /* Entry function for undoing inheritance/split transformation. Return true
6891 if we did any RTL change in this pass. */
6892 bool
6893 lra_undo_inheritance (void)
6895 unsigned int regno;
6896 int hard_regno;
6897 int n_all_inherit, n_inherit, n_all_split, n_split;
6898 rtx restore_rtx;
6899 bitmap_iterator bi;
6900 bool change_p;
6902 lra_undo_inheritance_iter++;
6903 if (lra_undo_inheritance_iter > LRA_MAX_INHERITANCE_PASSES)
6904 return false;
6905 if (lra_dump_file != NULL)
6906 fprintf (lra_dump_file,
6907 "\n********** Undoing inheritance #%d: **********\n\n",
6908 lra_undo_inheritance_iter);
6909 auto_bitmap remove_pseudos (&reg_obstack);
6910 n_inherit = n_all_inherit = 0;
6911 EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, regno, bi)
6912 if (lra_reg_info[regno].restore_rtx != NULL_RTX)
6914 n_all_inherit++;
6915 if (reg_renumber[regno] < 0
6916 /* If the original pseudo changed its allocation, just
6917 removing inheritance is dangerous as for changing
6918 allocation we used shorter live-ranges. */
6919 && (! REG_P (lra_reg_info[regno].restore_rtx)
6920 || reg_renumber[REGNO (lra_reg_info[regno].restore_rtx)] < 0))
6921 bitmap_set_bit (remove_pseudos, regno);
6922 else
6923 n_inherit++;
6925 if (lra_dump_file != NULL && n_all_inherit != 0)
6926 fprintf (lra_dump_file, "Inherit %d out of %d (%.2f%%)\n",
6927 n_inherit, n_all_inherit,
6928 (double) n_inherit / n_all_inherit * 100);
6929 n_split = n_all_split = 0;
6930 EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, regno, bi)
6931 if ((restore_rtx = lra_reg_info[regno].restore_rtx) != NULL_RTX)
6933 int restore_regno = REGNO (restore_rtx);
6935 n_all_split++;
6936 hard_regno = (restore_regno >= FIRST_PSEUDO_REGISTER
6937 ? reg_renumber[restore_regno] : restore_regno);
6938 if (hard_regno < 0 || reg_renumber[regno] == hard_regno)
6939 bitmap_set_bit (remove_pseudos, regno);
6940 else
6942 n_split++;
6943 if (lra_dump_file != NULL)
6944 fprintf (lra_dump_file, " Keep split r%d (orig=r%d)\n",
6945 regno, restore_regno);
6948 if (lra_dump_file != NULL && n_all_split != 0)
6949 fprintf (lra_dump_file, "Split %d out of %d (%.2f%%)\n",
6950 n_split, n_all_split,
6951 (double) n_split / n_all_split * 100);
6952 change_p = remove_inheritance_pseudos (remove_pseudos);
6953 /* Clear restore_regnos. */
6954 EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, regno, bi)
6955 lra_reg_info[regno].restore_rtx = NULL_RTX;
6956 EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, regno, bi)
6957 lra_reg_info[regno].restore_rtx = NULL_RTX;
6958 change_p = undo_optional_reloads () || change_p;
6959 return change_p;