2018-02-19 Sebastian Perta <sebastian.perta@renesas.com>
[official-gcc.git] / gcc / lra-constraints.c
blobd730f36fba0fac47ba3c17e01c9f3ffe0a9883c6
1 /* Code for RTL transformations to satisfy insn constraints.
2 Copyright (C) 2010-2018 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 (maybe_lt (GET_MODE_SIZE (GET_MODE (reg)),
595 GET_MODE_SIZE (mode)))
596 continue;
597 reg = lowpart_subreg (mode, reg, GET_MODE (reg));
598 if (reg == NULL_RTX || GET_CODE (reg) != SUBREG)
599 continue;
601 *result_reg = reg;
602 if (lra_dump_file != NULL)
604 fprintf (lra_dump_file, " Reuse r%d for reload ", regno);
605 dump_value_slim (lra_dump_file, original, 1);
607 if (new_class != lra_get_allocno_class (regno))
608 lra_change_class (regno, new_class, ", change to", false);
609 if (lra_dump_file != NULL)
610 fprintf (lra_dump_file, "\n");
611 return false;
613 /* If we have an input reload with a different mode, make sure it
614 will get a different hard reg. */
615 else if (REG_P (original)
616 && REG_P (curr_insn_input_reloads[i].input)
617 && REGNO (original) == REGNO (curr_insn_input_reloads[i].input)
618 && (GET_MODE (original)
619 != GET_MODE (curr_insn_input_reloads[i].input)))
620 unique_p = true;
622 *result_reg = (unique_p
623 ? lra_create_new_reg_with_unique_value
624 : lra_create_new_reg) (mode, original, rclass, title);
625 lra_assert (curr_insn_input_reloads_num < LRA_MAX_INSN_RELOADS);
626 curr_insn_input_reloads[curr_insn_input_reloads_num].input = original;
627 curr_insn_input_reloads[curr_insn_input_reloads_num].match_p = false;
628 curr_insn_input_reloads[curr_insn_input_reloads_num++].reg = *result_reg;
629 return true;
634 /* The page contains code to extract memory address parts. */
636 /* Wrapper around REGNO_OK_FOR_INDEX_P, to allow pseudos. */
637 static inline bool
638 ok_for_index_p_nonstrict (rtx reg)
640 unsigned regno = REGNO (reg);
642 return regno >= FIRST_PSEUDO_REGISTER || REGNO_OK_FOR_INDEX_P (regno);
645 /* A version of regno_ok_for_base_p for use here, when all pseudos
646 should count as OK. Arguments as for regno_ok_for_base_p. */
647 static inline bool
648 ok_for_base_p_nonstrict (rtx reg, machine_mode mode, addr_space_t as,
649 enum rtx_code outer_code, enum rtx_code index_code)
651 unsigned regno = REGNO (reg);
653 if (regno >= FIRST_PSEUDO_REGISTER)
654 return true;
655 return ok_for_base_p_1 (regno, mode, as, outer_code, index_code);
660 /* The page contains major code to choose the current insn alternative
661 and generate reloads for it. */
663 /* Return the offset from REGNO of the least significant register
664 in (reg:MODE REGNO).
666 This function is used to tell whether two registers satisfy
667 a matching constraint. (reg:MODE1 REGNO1) matches (reg:MODE2 REGNO2) if:
669 REGNO1 + lra_constraint_offset (REGNO1, MODE1)
670 == REGNO2 + lra_constraint_offset (REGNO2, MODE2) */
672 lra_constraint_offset (int regno, machine_mode mode)
674 lra_assert (regno < FIRST_PSEUDO_REGISTER);
676 scalar_int_mode int_mode;
677 if (WORDS_BIG_ENDIAN
678 && is_a <scalar_int_mode> (mode, &int_mode)
679 && GET_MODE_SIZE (int_mode) > UNITS_PER_WORD)
680 return hard_regno_nregs (regno, mode) - 1;
681 return 0;
684 /* Like rtx_equal_p except that it allows a REG and a SUBREG to match
685 if they are the same hard reg, and has special hacks for
686 auto-increment and auto-decrement. This is specifically intended for
687 process_alt_operands to use in determining whether two operands
688 match. X is the operand whose number is the lower of the two.
690 It is supposed that X is the output operand and Y is the input
691 operand. Y_HARD_REGNO is the final hard regno of register Y or
692 register in subreg Y as we know it now. Otherwise, it is a
693 negative value. */
694 static bool
695 operands_match_p (rtx x, rtx y, int y_hard_regno)
697 int i;
698 RTX_CODE code = GET_CODE (x);
699 const char *fmt;
701 if (x == y)
702 return true;
703 if ((code == REG || (code == SUBREG && REG_P (SUBREG_REG (x))))
704 && (REG_P (y) || (GET_CODE (y) == SUBREG && REG_P (SUBREG_REG (y)))))
706 int j;
708 i = get_hard_regno (x, false);
709 if (i < 0)
710 goto slow;
712 if ((j = y_hard_regno) < 0)
713 goto slow;
715 i += lra_constraint_offset (i, GET_MODE (x));
716 j += lra_constraint_offset (j, GET_MODE (y));
718 return i == j;
721 /* If two operands must match, because they are really a single
722 operand of an assembler insn, then two post-increments are invalid
723 because the assembler insn would increment only once. On the
724 other hand, a post-increment matches ordinary indexing if the
725 post-increment is the output operand. */
726 if (code == POST_DEC || code == POST_INC || code == POST_MODIFY)
727 return operands_match_p (XEXP (x, 0), y, y_hard_regno);
729 /* Two pre-increments are invalid because the assembler insn would
730 increment only once. On the other hand, a pre-increment matches
731 ordinary indexing if the pre-increment is the input operand. */
732 if (GET_CODE (y) == PRE_DEC || GET_CODE (y) == PRE_INC
733 || GET_CODE (y) == PRE_MODIFY)
734 return operands_match_p (x, XEXP (y, 0), -1);
736 slow:
738 if (code == REG && REG_P (y))
739 return REGNO (x) == REGNO (y);
741 if (code == REG && GET_CODE (y) == SUBREG && REG_P (SUBREG_REG (y))
742 && x == SUBREG_REG (y))
743 return true;
744 if (GET_CODE (y) == REG && code == SUBREG && REG_P (SUBREG_REG (x))
745 && SUBREG_REG (x) == y)
746 return true;
748 /* Now we have disposed of all the cases in which different rtx
749 codes can match. */
750 if (code != GET_CODE (y))
751 return false;
753 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
754 if (GET_MODE (x) != GET_MODE (y))
755 return false;
757 switch (code)
759 CASE_CONST_UNIQUE:
760 return false;
762 case LABEL_REF:
763 return label_ref_label (x) == label_ref_label (y);
764 case SYMBOL_REF:
765 return XSTR (x, 0) == XSTR (y, 0);
767 default:
768 break;
771 /* Compare the elements. If any pair of corresponding elements fail
772 to match, return false for the whole things. */
774 fmt = GET_RTX_FORMAT (code);
775 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
777 int val, j;
778 switch (fmt[i])
780 case 'w':
781 if (XWINT (x, i) != XWINT (y, i))
782 return false;
783 break;
785 case 'i':
786 if (XINT (x, i) != XINT (y, i))
787 return false;
788 break;
790 case 'p':
791 if (maybe_ne (SUBREG_BYTE (x), SUBREG_BYTE (y)))
792 return false;
793 break;
795 case 'e':
796 val = operands_match_p (XEXP (x, i), XEXP (y, i), -1);
797 if (val == 0)
798 return false;
799 break;
801 case '0':
802 break;
804 case 'E':
805 if (XVECLEN (x, i) != XVECLEN (y, i))
806 return false;
807 for (j = XVECLEN (x, i) - 1; j >= 0; --j)
809 val = operands_match_p (XVECEXP (x, i, j), XVECEXP (y, i, j), -1);
810 if (val == 0)
811 return false;
813 break;
815 /* It is believed that rtx's at this level will never
816 contain anything but integers and other rtx's, except for
817 within LABEL_REFs and SYMBOL_REFs. */
818 default:
819 gcc_unreachable ();
822 return true;
825 /* True if X is a constant that can be forced into the constant pool.
826 MODE is the mode of the operand, or VOIDmode if not known. */
827 #define CONST_POOL_OK_P(MODE, X) \
828 ((MODE) != VOIDmode \
829 && CONSTANT_P (X) \
830 && GET_CODE (X) != HIGH \
831 && GET_MODE_SIZE (MODE).is_constant () \
832 && !targetm.cannot_force_const_mem (MODE, X))
834 /* True if C is a non-empty register class that has too few registers
835 to be safely used as a reload target class. */
836 #define SMALL_REGISTER_CLASS_P(C) \
837 (ira_class_hard_regs_num [(C)] == 1 \
838 || (ira_class_hard_regs_num [(C)] >= 1 \
839 && targetm.class_likely_spilled_p (C)))
841 /* If REG is a reload pseudo, try to make its class satisfying CL. */
842 static void
843 narrow_reload_pseudo_class (rtx reg, enum reg_class cl)
845 enum reg_class rclass;
847 /* Do not make more accurate class from reloads generated. They are
848 mostly moves with a lot of constraints. Making more accurate
849 class may results in very narrow class and impossibility of find
850 registers for several reloads of one insn. */
851 if (INSN_UID (curr_insn) >= new_insn_uid_start)
852 return;
853 if (GET_CODE (reg) == SUBREG)
854 reg = SUBREG_REG (reg);
855 if (! REG_P (reg) || (int) REGNO (reg) < new_regno_start)
856 return;
857 if (in_class_p (reg, cl, &rclass) && rclass != cl)
858 lra_change_class (REGNO (reg), rclass, " Change to", true);
861 /* Searches X for any reference to a reg with the same value as REGNO,
862 returning the rtx of the reference found if any. Otherwise,
863 returns NULL_RTX. */
864 static rtx
865 regno_val_use_in (unsigned int regno, rtx x)
867 const char *fmt;
868 int i, j;
869 rtx tem;
871 if (REG_P (x) && lra_reg_info[REGNO (x)].val == lra_reg_info[regno].val)
872 return x;
874 fmt = GET_RTX_FORMAT (GET_CODE (x));
875 for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
877 if (fmt[i] == 'e')
879 if ((tem = regno_val_use_in (regno, XEXP (x, i))))
880 return tem;
882 else if (fmt[i] == 'E')
883 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
884 if ((tem = regno_val_use_in (regno , XVECEXP (x, i, j))))
885 return tem;
888 return NULL_RTX;
891 /* Return true if all current insn non-output operands except INS (it
892 has a negaitve end marker) do not use pseudos with the same value
893 as REGNO. */
894 static bool
895 check_conflict_input_operands (int regno, signed char *ins)
897 int in;
898 int n_operands = curr_static_id->n_operands;
900 for (int nop = 0; nop < n_operands; nop++)
901 if (! curr_static_id->operand[nop].is_operator
902 && curr_static_id->operand[nop].type != OP_OUT)
904 for (int i = 0; (in = ins[i]) >= 0; i++)
905 if (in == nop)
906 break;
907 if (in < 0
908 && regno_val_use_in (regno, *curr_id->operand_loc[nop]) != NULL_RTX)
909 return false;
911 return true;
914 /* Generate reloads for matching OUT and INS (array of input operand
915 numbers with end marker -1) with reg class GOAL_CLASS, considering
916 output operands OUTS (similar array to INS) needing to be in different
917 registers. Add input and output reloads correspondingly to the lists
918 *BEFORE and *AFTER. OUT might be negative. In this case we generate
919 input reloads for matched input operands INS. EARLY_CLOBBER_P is a flag
920 that the output operand is early clobbered for chosen alternative. */
921 static void
922 match_reload (signed char out, signed char *ins, signed char *outs,
923 enum reg_class goal_class, rtx_insn **before,
924 rtx_insn **after, bool early_clobber_p)
926 bool out_conflict;
927 int i, in;
928 rtx new_in_reg, new_out_reg, reg;
929 machine_mode inmode, outmode;
930 rtx in_rtx = *curr_id->operand_loc[ins[0]];
931 rtx out_rtx = out < 0 ? in_rtx : *curr_id->operand_loc[out];
933 inmode = curr_operand_mode[ins[0]];
934 outmode = out < 0 ? inmode : curr_operand_mode[out];
935 push_to_sequence (*before);
936 if (inmode != outmode)
938 /* process_alt_operands has already checked that the mode sizes
939 are ordered. */
940 if (partial_subreg_p (outmode, inmode))
942 reg = new_in_reg
943 = lra_create_new_reg_with_unique_value (inmode, in_rtx,
944 goal_class, "");
945 new_out_reg = gen_lowpart_SUBREG (outmode, reg);
946 LRA_SUBREG_P (new_out_reg) = 1;
947 /* If the input reg is dying here, we can use the same hard
948 register for REG and IN_RTX. We do it only for original
949 pseudos as reload pseudos can die although original
950 pseudos still live where reload pseudos dies. */
951 if (REG_P (in_rtx) && (int) REGNO (in_rtx) < lra_new_regno_start
952 && find_regno_note (curr_insn, REG_DEAD, REGNO (in_rtx))
953 && (!early_clobber_p
954 || check_conflict_input_operands(REGNO (in_rtx), ins)))
955 lra_assign_reg_val (REGNO (in_rtx), REGNO (reg));
957 else
959 reg = new_out_reg
960 = lra_create_new_reg_with_unique_value (outmode, out_rtx,
961 goal_class, "");
962 new_in_reg = gen_lowpart_SUBREG (inmode, reg);
963 /* NEW_IN_REG is non-paradoxical subreg. We don't want
964 NEW_OUT_REG living above. We add clobber clause for
965 this. This is just a temporary clobber. We can remove
966 it at the end of LRA work. */
967 rtx_insn *clobber = emit_clobber (new_out_reg);
968 LRA_TEMP_CLOBBER_P (PATTERN (clobber)) = 1;
969 LRA_SUBREG_P (new_in_reg) = 1;
970 if (GET_CODE (in_rtx) == SUBREG)
972 rtx subreg_reg = SUBREG_REG (in_rtx);
974 /* If SUBREG_REG is dying here and sub-registers IN_RTX
975 and NEW_IN_REG are similar, we can use the same hard
976 register for REG and SUBREG_REG. */
977 if (REG_P (subreg_reg)
978 && (int) REGNO (subreg_reg) < lra_new_regno_start
979 && GET_MODE (subreg_reg) == outmode
980 && known_eq (SUBREG_BYTE (in_rtx), SUBREG_BYTE (new_in_reg))
981 && find_regno_note (curr_insn, REG_DEAD, REGNO (subreg_reg))
982 && (! early_clobber_p
983 || check_conflict_input_operands (REGNO (subreg_reg),
984 ins)))
985 lra_assign_reg_val (REGNO (subreg_reg), REGNO (reg));
989 else
991 /* Pseudos have values -- see comments for lra_reg_info.
992 Different pseudos with the same value do not conflict even if
993 they live in the same place. When we create a pseudo we
994 assign value of original pseudo (if any) from which we
995 created the new pseudo. If we create the pseudo from the
996 input pseudo, the new pseudo will have no conflict with the
997 input pseudo which is wrong when the input pseudo lives after
998 the insn and as the new pseudo value is changed by the insn
999 output. Therefore we create the new pseudo from the output
1000 except the case when we have single matched dying input
1001 pseudo.
1003 We cannot reuse the current output register because we might
1004 have a situation like "a <- a op b", where the constraints
1005 force the second input operand ("b") to match the output
1006 operand ("a"). "b" must then be copied into a new register
1007 so that it doesn't clobber the current value of "a".
1009 We can not use the same value if the output pseudo is
1010 early clobbered or the input pseudo is mentioned in the
1011 output, e.g. as an address part in memory, because
1012 output reload will actually extend the pseudo liveness.
1013 We don't care about eliminable hard regs here as we are
1014 interesting only in pseudos. */
1016 /* Matching input's register value is the same as one of the other
1017 output operand. Output operands in a parallel insn must be in
1018 different registers. */
1019 out_conflict = false;
1020 if (REG_P (in_rtx))
1022 for (i = 0; outs[i] >= 0; i++)
1024 rtx other_out_rtx = *curr_id->operand_loc[outs[i]];
1025 if (REG_P (other_out_rtx)
1026 && (regno_val_use_in (REGNO (in_rtx), other_out_rtx)
1027 != NULL_RTX))
1029 out_conflict = true;
1030 break;
1035 new_in_reg = new_out_reg
1036 = (! early_clobber_p && ins[1] < 0 && REG_P (in_rtx)
1037 && (int) REGNO (in_rtx) < lra_new_regno_start
1038 && find_regno_note (curr_insn, REG_DEAD, REGNO (in_rtx))
1039 && (! early_clobber_p
1040 || check_conflict_input_operands (REGNO (in_rtx), ins))
1041 && (out < 0
1042 || regno_val_use_in (REGNO (in_rtx), out_rtx) == NULL_RTX)
1043 && !out_conflict
1044 ? lra_create_new_reg (inmode, in_rtx, goal_class, "")
1045 : lra_create_new_reg_with_unique_value (outmode, out_rtx,
1046 goal_class, ""));
1048 /* In operand can be got from transformations before processing insn
1049 constraints. One example of such transformations is subreg
1050 reloading (see function simplify_operand_subreg). The new
1051 pseudos created by the transformations might have inaccurate
1052 class (ALL_REGS) and we should make their classes more
1053 accurate. */
1054 narrow_reload_pseudo_class (in_rtx, goal_class);
1055 lra_emit_move (copy_rtx (new_in_reg), in_rtx);
1056 *before = get_insns ();
1057 end_sequence ();
1058 /* Add the new pseudo to consider values of subsequent input reload
1059 pseudos. */
1060 lra_assert (curr_insn_input_reloads_num < LRA_MAX_INSN_RELOADS);
1061 curr_insn_input_reloads[curr_insn_input_reloads_num].input = in_rtx;
1062 curr_insn_input_reloads[curr_insn_input_reloads_num].match_p = true;
1063 curr_insn_input_reloads[curr_insn_input_reloads_num++].reg = new_in_reg;
1064 for (i = 0; (in = ins[i]) >= 0; i++)
1066 lra_assert
1067 (GET_MODE (*curr_id->operand_loc[in]) == VOIDmode
1068 || GET_MODE (new_in_reg) == GET_MODE (*curr_id->operand_loc[in]));
1069 *curr_id->operand_loc[in] = new_in_reg;
1071 lra_update_dups (curr_id, ins);
1072 if (out < 0)
1073 return;
1074 /* See a comment for the input operand above. */
1075 narrow_reload_pseudo_class (out_rtx, goal_class);
1076 if (find_reg_note (curr_insn, REG_UNUSED, out_rtx) == NULL_RTX)
1078 start_sequence ();
1079 lra_emit_move (out_rtx, copy_rtx (new_out_reg));
1080 emit_insn (*after);
1081 *after = get_insns ();
1082 end_sequence ();
1084 *curr_id->operand_loc[out] = new_out_reg;
1085 lra_update_dup (curr_id, out);
1088 /* Return register class which is union of all reg classes in insn
1089 constraint alternative string starting with P. */
1090 static enum reg_class
1091 reg_class_from_constraints (const char *p)
1093 int c, len;
1094 enum reg_class op_class = NO_REGS;
1097 switch ((c = *p, len = CONSTRAINT_LEN (c, p)), c)
1099 case '#':
1100 case ',':
1101 return op_class;
1103 case 'g':
1104 op_class = reg_class_subunion[op_class][GENERAL_REGS];
1105 break;
1107 default:
1108 enum constraint_num cn = lookup_constraint (p);
1109 enum reg_class cl = reg_class_for_constraint (cn);
1110 if (cl == NO_REGS)
1112 if (insn_extra_address_constraint (cn))
1113 op_class
1114 = (reg_class_subunion
1115 [op_class][base_reg_class (VOIDmode, ADDR_SPACE_GENERIC,
1116 ADDRESS, SCRATCH)]);
1117 break;
1120 op_class = reg_class_subunion[op_class][cl];
1121 break;
1123 while ((p += len), c);
1124 return op_class;
1127 /* If OP is a register, return the class of the register as per
1128 get_reg_class, otherwise return NO_REGS. */
1129 static inline enum reg_class
1130 get_op_class (rtx op)
1132 return REG_P (op) ? get_reg_class (REGNO (op)) : NO_REGS;
1135 /* Return generated insn mem_pseudo:=val if TO_P or val:=mem_pseudo
1136 otherwise. If modes of MEM_PSEUDO and VAL are different, use
1137 SUBREG for VAL to make them equal. */
1138 static rtx_insn *
1139 emit_spill_move (bool to_p, rtx mem_pseudo, rtx val)
1141 if (GET_MODE (mem_pseudo) != GET_MODE (val))
1143 /* Usually size of mem_pseudo is greater than val size but in
1144 rare cases it can be less as it can be defined by target
1145 dependent macro HARD_REGNO_CALLER_SAVE_MODE. */
1146 if (! MEM_P (val))
1148 val = gen_lowpart_SUBREG (GET_MODE (mem_pseudo),
1149 GET_CODE (val) == SUBREG
1150 ? SUBREG_REG (val) : val);
1151 LRA_SUBREG_P (val) = 1;
1153 else
1155 mem_pseudo = gen_lowpart_SUBREG (GET_MODE (val), mem_pseudo);
1156 LRA_SUBREG_P (mem_pseudo) = 1;
1159 return to_p ? gen_move_insn (mem_pseudo, val)
1160 : gen_move_insn (val, mem_pseudo);
1163 /* Process a special case insn (register move), return true if we
1164 don't need to process it anymore. INSN should be a single set
1165 insn. Set up that RTL was changed through CHANGE_P and that hook
1166 TARGET_SECONDARY_MEMORY_NEEDED says to use secondary memory through
1167 SEC_MEM_P. */
1168 static bool
1169 check_and_process_move (bool *change_p, bool *sec_mem_p ATTRIBUTE_UNUSED)
1171 int sregno, dregno;
1172 rtx dest, src, dreg, sreg, new_reg, scratch_reg;
1173 rtx_insn *before;
1174 enum reg_class dclass, sclass, secondary_class;
1175 secondary_reload_info sri;
1177 lra_assert (curr_insn_set != NULL_RTX);
1178 dreg = dest = SET_DEST (curr_insn_set);
1179 sreg = src = SET_SRC (curr_insn_set);
1180 if (GET_CODE (dest) == SUBREG)
1181 dreg = SUBREG_REG (dest);
1182 if (GET_CODE (src) == SUBREG)
1183 sreg = SUBREG_REG (src);
1184 if (! (REG_P (dreg) || MEM_P (dreg)) || ! (REG_P (sreg) || MEM_P (sreg)))
1185 return false;
1186 sclass = dclass = NO_REGS;
1187 if (REG_P (dreg))
1188 dclass = get_reg_class (REGNO (dreg));
1189 gcc_assert (dclass < LIM_REG_CLASSES);
1190 if (dclass == ALL_REGS)
1191 /* ALL_REGS is used for new pseudos created by transformations
1192 like reload of SUBREG_REG (see function
1193 simplify_operand_subreg). We don't know their class yet. We
1194 should figure out the class from processing the insn
1195 constraints not in this fast path function. Even if ALL_REGS
1196 were a right class for the pseudo, secondary_... hooks usually
1197 are not define for ALL_REGS. */
1198 return false;
1199 if (REG_P (sreg))
1200 sclass = get_reg_class (REGNO (sreg));
1201 gcc_assert (sclass < LIM_REG_CLASSES);
1202 if (sclass == ALL_REGS)
1203 /* See comments above. */
1204 return false;
1205 if (sclass == NO_REGS && dclass == NO_REGS)
1206 return false;
1207 if (targetm.secondary_memory_needed (GET_MODE (src), sclass, dclass)
1208 && ((sclass != NO_REGS && dclass != NO_REGS)
1209 || (GET_MODE (src)
1210 != targetm.secondary_memory_needed_mode (GET_MODE (src)))))
1212 *sec_mem_p = true;
1213 return false;
1215 if (! REG_P (dreg) || ! REG_P (sreg))
1216 return false;
1217 sri.prev_sri = NULL;
1218 sri.icode = CODE_FOR_nothing;
1219 sri.extra_cost = 0;
1220 secondary_class = NO_REGS;
1221 /* Set up hard register for a reload pseudo for hook
1222 secondary_reload because some targets just ignore unassigned
1223 pseudos in the hook. */
1224 if (dclass != NO_REGS && lra_get_regno_hard_regno (REGNO (dreg)) < 0)
1226 dregno = REGNO (dreg);
1227 reg_renumber[dregno] = ira_class_hard_regs[dclass][0];
1229 else
1230 dregno = -1;
1231 if (sclass != NO_REGS && lra_get_regno_hard_regno (REGNO (sreg)) < 0)
1233 sregno = REGNO (sreg);
1234 reg_renumber[sregno] = ira_class_hard_regs[sclass][0];
1236 else
1237 sregno = -1;
1238 if (sclass != NO_REGS)
1239 secondary_class
1240 = (enum reg_class) targetm.secondary_reload (false, dest,
1241 (reg_class_t) sclass,
1242 GET_MODE (src), &sri);
1243 if (sclass == NO_REGS
1244 || ((secondary_class != NO_REGS || sri.icode != CODE_FOR_nothing)
1245 && dclass != NO_REGS))
1247 enum reg_class old_sclass = secondary_class;
1248 secondary_reload_info old_sri = sri;
1250 sri.prev_sri = NULL;
1251 sri.icode = CODE_FOR_nothing;
1252 sri.extra_cost = 0;
1253 secondary_class
1254 = (enum reg_class) targetm.secondary_reload (true, src,
1255 (reg_class_t) dclass,
1256 GET_MODE (src), &sri);
1257 /* Check the target hook consistency. */
1258 lra_assert
1259 ((secondary_class == NO_REGS && sri.icode == CODE_FOR_nothing)
1260 || (old_sclass == NO_REGS && old_sri.icode == CODE_FOR_nothing)
1261 || (secondary_class == old_sclass && sri.icode == old_sri.icode));
1263 if (sregno >= 0)
1264 reg_renumber [sregno] = -1;
1265 if (dregno >= 0)
1266 reg_renumber [dregno] = -1;
1267 if (secondary_class == NO_REGS && sri.icode == CODE_FOR_nothing)
1268 return false;
1269 *change_p = true;
1270 new_reg = NULL_RTX;
1271 if (secondary_class != NO_REGS)
1272 new_reg = lra_create_new_reg_with_unique_value (GET_MODE (src), NULL_RTX,
1273 secondary_class,
1274 "secondary");
1275 start_sequence ();
1276 if (sri.icode == CODE_FOR_nothing)
1277 lra_emit_move (new_reg, src);
1278 else
1280 enum reg_class scratch_class;
1282 scratch_class = (reg_class_from_constraints
1283 (insn_data[sri.icode].operand[2].constraint));
1284 scratch_reg = (lra_create_new_reg_with_unique_value
1285 (insn_data[sri.icode].operand[2].mode, NULL_RTX,
1286 scratch_class, "scratch"));
1287 emit_insn (GEN_FCN (sri.icode) (new_reg != NULL_RTX ? new_reg : dest,
1288 src, scratch_reg));
1290 before = get_insns ();
1291 end_sequence ();
1292 lra_process_new_insns (curr_insn, before, NULL, "Inserting the move");
1293 if (new_reg != NULL_RTX)
1294 SET_SRC (curr_insn_set) = new_reg;
1295 else
1297 if (lra_dump_file != NULL)
1299 fprintf (lra_dump_file, "Deleting move %u\n", INSN_UID (curr_insn));
1300 dump_insn_slim (lra_dump_file, curr_insn);
1302 lra_set_insn_deleted (curr_insn);
1303 return true;
1305 return false;
1308 /* The following data describe the result of process_alt_operands.
1309 The data are used in curr_insn_transform to generate reloads. */
1311 /* The chosen reg classes which should be used for the corresponding
1312 operands. */
1313 static enum reg_class goal_alt[MAX_RECOG_OPERANDS];
1314 /* True if the operand should be the same as another operand and that
1315 other operand does not need a reload. */
1316 static bool goal_alt_match_win[MAX_RECOG_OPERANDS];
1317 /* True if the operand does not need a reload. */
1318 static bool goal_alt_win[MAX_RECOG_OPERANDS];
1319 /* True if the operand can be offsetable memory. */
1320 static bool goal_alt_offmemok[MAX_RECOG_OPERANDS];
1321 /* The number of an operand to which given operand can be matched to. */
1322 static int goal_alt_matches[MAX_RECOG_OPERANDS];
1323 /* The number of elements in the following array. */
1324 static int goal_alt_dont_inherit_ops_num;
1325 /* Numbers of operands whose reload pseudos should not be inherited. */
1326 static int goal_alt_dont_inherit_ops[MAX_RECOG_OPERANDS];
1327 /* True if the insn commutative operands should be swapped. */
1328 static bool goal_alt_swapped;
1329 /* The chosen insn alternative. */
1330 static int goal_alt_number;
1332 /* True if the corresponding operand is the result of an equivalence
1333 substitution. */
1334 static bool equiv_substition_p[MAX_RECOG_OPERANDS];
1336 /* The following five variables are used to choose the best insn
1337 alternative. They reflect final characteristics of the best
1338 alternative. */
1340 /* Number of necessary reloads and overall cost reflecting the
1341 previous value and other unpleasantness of the best alternative. */
1342 static int best_losers, best_overall;
1343 /* Overall number hard registers used for reloads. For example, on
1344 some targets we need 2 general registers to reload DFmode and only
1345 one floating point register. */
1346 static int best_reload_nregs;
1347 /* Overall number reflecting distances of previous reloading the same
1348 value. The distances are counted from the current BB start. It is
1349 used to improve inheritance chances. */
1350 static int best_reload_sum;
1352 /* True if the current insn should have no correspondingly input or
1353 output reloads. */
1354 static bool no_input_reloads_p, no_output_reloads_p;
1356 /* True if we swapped the commutative operands in the current
1357 insn. */
1358 static int curr_swapped;
1360 /* if CHECK_ONLY_P is false, arrange for address element *LOC to be a
1361 register of class CL. Add any input reloads to list BEFORE. AFTER
1362 is nonnull if *LOC is an automodified value; handle that case by
1363 adding the required output reloads to list AFTER. Return true if
1364 the RTL was changed.
1366 if CHECK_ONLY_P is true, check that the *LOC is a correct address
1367 register. Return false if the address register is correct. */
1368 static bool
1369 process_addr_reg (rtx *loc, bool check_only_p, rtx_insn **before, rtx_insn **after,
1370 enum reg_class cl)
1372 int regno;
1373 enum reg_class rclass, new_class;
1374 rtx reg;
1375 rtx new_reg;
1376 machine_mode mode;
1377 bool subreg_p, before_p = false;
1379 subreg_p = GET_CODE (*loc) == SUBREG;
1380 if (subreg_p)
1382 reg = SUBREG_REG (*loc);
1383 mode = GET_MODE (reg);
1385 /* For mode with size bigger than ptr_mode, there unlikely to be "mov"
1386 between two registers with different classes, but there normally will
1387 be "mov" which transfers element of vector register into the general
1388 register, and this normally will be a subreg which should be reloaded
1389 as a whole. This is particularly likely to be triggered when
1390 -fno-split-wide-types specified. */
1391 if (!REG_P (reg)
1392 || in_class_p (reg, cl, &new_class)
1393 || known_le (GET_MODE_SIZE (mode), GET_MODE_SIZE (ptr_mode)))
1394 loc = &SUBREG_REG (*loc);
1397 reg = *loc;
1398 mode = GET_MODE (reg);
1399 if (! REG_P (reg))
1401 if (check_only_p)
1402 return true;
1403 /* Always reload memory in an address even if the target supports
1404 such addresses. */
1405 new_reg = lra_create_new_reg_with_unique_value (mode, reg, cl, "address");
1406 before_p = true;
1408 else
1410 regno = REGNO (reg);
1411 rclass = get_reg_class (regno);
1412 if (! check_only_p
1413 && (*loc = get_equiv_with_elimination (reg, curr_insn)) != reg)
1415 if (lra_dump_file != NULL)
1417 fprintf (lra_dump_file,
1418 "Changing pseudo %d in address of insn %u on equiv ",
1419 REGNO (reg), INSN_UID (curr_insn));
1420 dump_value_slim (lra_dump_file, *loc, 1);
1421 fprintf (lra_dump_file, "\n");
1423 *loc = copy_rtx (*loc);
1425 if (*loc != reg || ! in_class_p (reg, cl, &new_class))
1427 if (check_only_p)
1428 return true;
1429 reg = *loc;
1430 if (get_reload_reg (after == NULL ? OP_IN : OP_INOUT,
1431 mode, reg, cl, subreg_p, "address", &new_reg))
1432 before_p = true;
1434 else if (new_class != NO_REGS && rclass != new_class)
1436 if (check_only_p)
1437 return true;
1438 lra_change_class (regno, new_class, " Change to", true);
1439 return false;
1441 else
1442 return false;
1444 if (before_p)
1446 push_to_sequence (*before);
1447 lra_emit_move (new_reg, reg);
1448 *before = get_insns ();
1449 end_sequence ();
1451 *loc = new_reg;
1452 if (after != NULL)
1454 start_sequence ();
1455 lra_emit_move (before_p ? copy_rtx (reg) : reg, new_reg);
1456 emit_insn (*after);
1457 *after = get_insns ();
1458 end_sequence ();
1460 return true;
1463 /* Insert move insn in simplify_operand_subreg. BEFORE returns
1464 the insn to be inserted before curr insn. AFTER returns the
1465 the insn to be inserted after curr insn. ORIGREG and NEWREG
1466 are the original reg and new reg for reload. */
1467 static void
1468 insert_move_for_subreg (rtx_insn **before, rtx_insn **after, rtx origreg,
1469 rtx newreg)
1471 if (before)
1473 push_to_sequence (*before);
1474 lra_emit_move (newreg, origreg);
1475 *before = get_insns ();
1476 end_sequence ();
1478 if (after)
1480 start_sequence ();
1481 lra_emit_move (origreg, newreg);
1482 emit_insn (*after);
1483 *after = get_insns ();
1484 end_sequence ();
1488 static int valid_address_p (machine_mode mode, rtx addr, addr_space_t as);
1489 static bool process_address (int, bool, rtx_insn **, rtx_insn **);
1491 /* Make reloads for subreg in operand NOP with internal subreg mode
1492 REG_MODE, add new reloads for further processing. Return true if
1493 any change was done. */
1494 static bool
1495 simplify_operand_subreg (int nop, machine_mode reg_mode)
1497 int hard_regno;
1498 rtx_insn *before, *after;
1499 machine_mode mode, innermode;
1500 rtx reg, new_reg;
1501 rtx operand = *curr_id->operand_loc[nop];
1502 enum reg_class regclass;
1503 enum op_type type;
1505 before = after = NULL;
1507 if (GET_CODE (operand) != SUBREG)
1508 return false;
1510 mode = GET_MODE (operand);
1511 reg = SUBREG_REG (operand);
1512 innermode = GET_MODE (reg);
1513 type = curr_static_id->operand[nop].type;
1514 if (MEM_P (reg))
1516 const bool addr_was_valid
1517 = valid_address_p (innermode, XEXP (reg, 0), MEM_ADDR_SPACE (reg));
1518 alter_subreg (curr_id->operand_loc[nop], false);
1519 rtx subst = *curr_id->operand_loc[nop];
1520 lra_assert (MEM_P (subst));
1522 if (!addr_was_valid
1523 || valid_address_p (GET_MODE (subst), XEXP (subst, 0),
1524 MEM_ADDR_SPACE (subst))
1525 || ((get_constraint_type (lookup_constraint
1526 (curr_static_id->operand[nop].constraint))
1527 != CT_SPECIAL_MEMORY)
1528 /* We still can reload address and if the address is
1529 valid, we can remove subreg without reloading its
1530 inner memory. */
1531 && valid_address_p (GET_MODE (subst),
1532 regno_reg_rtx
1533 [ira_class_hard_regs
1534 [base_reg_class (GET_MODE (subst),
1535 MEM_ADDR_SPACE (subst),
1536 ADDRESS, SCRATCH)][0]],
1537 MEM_ADDR_SPACE (subst))))
1539 /* If we change the address for a paradoxical subreg of memory, the
1540 new address might violate the necessary alignment or the access
1541 might be slow; take this into consideration. We need not worry
1542 about accesses beyond allocated memory for paradoxical memory
1543 subregs as we don't substitute such equiv memory (see processing
1544 equivalences in function lra_constraints) and because for spilled
1545 pseudos we allocate stack memory enough for the biggest
1546 corresponding paradoxical subreg.
1548 However, do not blindly simplify a (subreg (mem ...)) for
1549 WORD_REGISTER_OPERATIONS targets as this may lead to loading junk
1550 data into a register when the inner is narrower than outer or
1551 missing important data from memory when the inner is wider than
1552 outer. This rule only applies to modes that are no wider than
1553 a word. */
1554 if (!(maybe_ne (GET_MODE_PRECISION (mode),
1555 GET_MODE_PRECISION (innermode))
1556 && known_le (GET_MODE_SIZE (mode), UNITS_PER_WORD)
1557 && known_le (GET_MODE_SIZE (innermode), UNITS_PER_WORD)
1558 && WORD_REGISTER_OPERATIONS)
1559 && (!(MEM_ALIGN (subst) < GET_MODE_ALIGNMENT (mode)
1560 && targetm.slow_unaligned_access (mode, MEM_ALIGN (subst)))
1561 || (MEM_ALIGN (reg) < GET_MODE_ALIGNMENT (innermode)
1562 && targetm.slow_unaligned_access (innermode,
1563 MEM_ALIGN (reg)))))
1564 return true;
1566 *curr_id->operand_loc[nop] = operand;
1568 /* But if the address was not valid, we cannot reload the MEM without
1569 reloading the address first. */
1570 if (!addr_was_valid)
1571 process_address (nop, false, &before, &after);
1573 /* INNERMODE is fast, MODE slow. Reload the mem in INNERMODE. */
1574 enum reg_class rclass
1575 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1576 if (get_reload_reg (curr_static_id->operand[nop].type, innermode,
1577 reg, rclass, TRUE, "slow mem", &new_reg))
1579 bool insert_before, insert_after;
1580 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1582 insert_before = (type != OP_OUT
1583 || partial_subreg_p (mode, innermode));
1584 insert_after = type != OP_IN;
1585 insert_move_for_subreg (insert_before ? &before : NULL,
1586 insert_after ? &after : NULL,
1587 reg, new_reg);
1589 SUBREG_REG (operand) = new_reg;
1591 /* Convert to MODE. */
1592 reg = operand;
1593 rclass
1594 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1595 if (get_reload_reg (curr_static_id->operand[nop].type, mode, reg,
1596 rclass, TRUE, "slow mem", &new_reg))
1598 bool insert_before, insert_after;
1599 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1601 insert_before = type != OP_OUT;
1602 insert_after = type != OP_IN;
1603 insert_move_for_subreg (insert_before ? &before : NULL,
1604 insert_after ? &after : NULL,
1605 reg, new_reg);
1607 *curr_id->operand_loc[nop] = new_reg;
1608 lra_process_new_insns (curr_insn, before, after,
1609 "Inserting slow mem reload");
1610 return true;
1613 /* If the address was valid and became invalid, prefer to reload
1614 the memory. Typical case is when the index scale should
1615 correspond the memory. */
1616 *curr_id->operand_loc[nop] = operand;
1617 /* Do not return false here as the MEM_P (reg) will be processed
1618 later in this function. */
1620 else if (REG_P (reg) && REGNO (reg) < FIRST_PSEUDO_REGISTER)
1622 alter_subreg (curr_id->operand_loc[nop], false);
1623 return true;
1625 else if (CONSTANT_P (reg))
1627 /* Try to simplify subreg of constant. It is usually result of
1628 equivalence substitution. */
1629 if (innermode == VOIDmode
1630 && (innermode = original_subreg_reg_mode[nop]) == VOIDmode)
1631 innermode = curr_static_id->operand[nop].mode;
1632 if ((new_reg = simplify_subreg (mode, reg, innermode,
1633 SUBREG_BYTE (operand))) != NULL_RTX)
1635 *curr_id->operand_loc[nop] = new_reg;
1636 return true;
1639 /* Put constant into memory when we have mixed modes. It generates
1640 a better code in most cases as it does not need a secondary
1641 reload memory. It also prevents LRA looping when LRA is using
1642 secondary reload memory again and again. */
1643 if (CONSTANT_P (reg) && CONST_POOL_OK_P (reg_mode, reg)
1644 && SCALAR_INT_MODE_P (reg_mode) != SCALAR_INT_MODE_P (mode))
1646 SUBREG_REG (operand) = force_const_mem (reg_mode, reg);
1647 alter_subreg (curr_id->operand_loc[nop], false);
1648 return true;
1650 /* Force a reload of the SUBREG_REG if this is a constant or PLUS or
1651 if there may be a problem accessing OPERAND in the outer
1652 mode. */
1653 if ((REG_P (reg)
1654 && REGNO (reg) >= FIRST_PSEUDO_REGISTER
1655 && (hard_regno = lra_get_regno_hard_regno (REGNO (reg))) >= 0
1656 /* Don't reload paradoxical subregs because we could be looping
1657 having repeatedly final regno out of hard regs range. */
1658 && (hard_regno_nregs (hard_regno, innermode)
1659 >= hard_regno_nregs (hard_regno, mode))
1660 && simplify_subreg_regno (hard_regno, innermode,
1661 SUBREG_BYTE (operand), mode) < 0
1662 /* Don't reload subreg for matching reload. It is actually
1663 valid subreg in LRA. */
1664 && ! LRA_SUBREG_P (operand))
1665 || CONSTANT_P (reg) || GET_CODE (reg) == PLUS || MEM_P (reg))
1667 enum reg_class rclass;
1669 if (REG_P (reg))
1670 /* There is a big probability that we will get the same class
1671 for the new pseudo and we will get the same insn which
1672 means infinite looping. So spill the new pseudo. */
1673 rclass = NO_REGS;
1674 else
1675 /* The class will be defined later in curr_insn_transform. */
1676 rclass
1677 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1679 if (get_reload_reg (curr_static_id->operand[nop].type, reg_mode, reg,
1680 rclass, TRUE, "subreg reg", &new_reg))
1682 bool insert_before, insert_after;
1683 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1685 insert_before = (type != OP_OUT
1686 || read_modify_subreg_p (operand));
1687 insert_after = (type != OP_IN);
1688 insert_move_for_subreg (insert_before ? &before : NULL,
1689 insert_after ? &after : NULL,
1690 reg, new_reg);
1692 SUBREG_REG (operand) = new_reg;
1693 lra_process_new_insns (curr_insn, before, after,
1694 "Inserting subreg reload");
1695 return true;
1697 /* Force a reload for a paradoxical subreg. For paradoxical subreg,
1698 IRA allocates hardreg to the inner pseudo reg according to its mode
1699 instead of the outermode, so the size of the hardreg may not be enough
1700 to contain the outermode operand, in that case we may need to insert
1701 reload for the reg. For the following two types of paradoxical subreg,
1702 we need to insert reload:
1703 1. If the op_type is OP_IN, and the hardreg could not be paired with
1704 other hardreg to contain the outermode operand
1705 (checked by in_hard_reg_set_p), we need to insert the reload.
1706 2. If the op_type is OP_OUT or OP_INOUT.
1708 Here is a paradoxical subreg example showing how the reload is generated:
1710 (insn 5 4 7 2 (set (reg:TI 106 [ __comp ])
1711 (subreg:TI (reg:DI 107 [ __comp ]) 0)) {*movti_internal_rex64}
1713 In IRA, reg107 is allocated to a DImode hardreg. We use x86-64 as example
1714 here, if reg107 is assigned to hardreg R15, because R15 is the last
1715 hardreg, compiler cannot find another hardreg to pair with R15 to
1716 contain TImode data. So we insert a TImode reload reg180 for it.
1717 After reload is inserted:
1719 (insn 283 0 0 (set (subreg:DI (reg:TI 180 [orig:107 __comp ] [107]) 0)
1720 (reg:DI 107 [ __comp ])) -1
1721 (insn 5 4 7 2 (set (reg:TI 106 [ __comp ])
1722 (subreg:TI (reg:TI 180 [orig:107 __comp ] [107]) 0)) {*movti_internal_rex64}
1724 Two reload hard registers will be allocated to reg180 to save TImode data
1725 in LRA_assign. */
1726 else if (REG_P (reg)
1727 && REGNO (reg) >= FIRST_PSEUDO_REGISTER
1728 && (hard_regno = lra_get_regno_hard_regno (REGNO (reg))) >= 0
1729 && (hard_regno_nregs (hard_regno, innermode)
1730 < hard_regno_nregs (hard_regno, mode))
1731 && (regclass = lra_get_allocno_class (REGNO (reg)))
1732 && (type != OP_IN
1733 || !in_hard_reg_set_p (reg_class_contents[regclass],
1734 mode, hard_regno)))
1736 /* The class will be defined later in curr_insn_transform. */
1737 enum reg_class rclass
1738 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1740 if (get_reload_reg (curr_static_id->operand[nop].type, mode, reg,
1741 rclass, TRUE, "paradoxical subreg", &new_reg))
1743 rtx subreg;
1744 bool insert_before, insert_after;
1746 PUT_MODE (new_reg, mode);
1747 subreg = gen_lowpart_SUBREG (innermode, new_reg);
1748 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1750 insert_before = (type != OP_OUT);
1751 insert_after = (type != OP_IN);
1752 insert_move_for_subreg (insert_before ? &before : NULL,
1753 insert_after ? &after : NULL,
1754 reg, subreg);
1756 SUBREG_REG (operand) = new_reg;
1757 lra_process_new_insns (curr_insn, before, after,
1758 "Inserting paradoxical subreg reload");
1759 return true;
1761 return false;
1764 /* Return TRUE if X refers for a hard register from SET. */
1765 static bool
1766 uses_hard_regs_p (rtx x, HARD_REG_SET set)
1768 int i, j, x_hard_regno;
1769 machine_mode mode;
1770 const char *fmt;
1771 enum rtx_code code;
1773 if (x == NULL_RTX)
1774 return false;
1775 code = GET_CODE (x);
1776 mode = GET_MODE (x);
1777 if (code == SUBREG)
1779 mode = wider_subreg_mode (x);
1780 x = SUBREG_REG (x);
1781 code = GET_CODE (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 biggest_mode[nop] = wider_subreg_mode (op);
1960 operand_reg[nop] = reg = SUBREG_REG (op);
1962 if (! REG_P (reg))
1963 operand_reg[nop] = NULL_RTX;
1964 else if (REGNO (reg) >= FIRST_PSEUDO_REGISTER
1965 || ((int) REGNO (reg)
1966 == lra_get_elimination_hard_regno (REGNO (reg))))
1967 no_subreg_reg_operand[nop] = reg;
1968 else
1969 operand_reg[nop] = no_subreg_reg_operand[nop]
1970 /* Just use natural mode for elimination result. It should
1971 be enough for extra constraints hooks. */
1972 = regno_reg_rtx[hard_regno[nop]];
1975 /* The constraints are made of several alternatives. Each operand's
1976 constraint looks like foo,bar,... with commas separating the
1977 alternatives. The first alternatives for all operands go
1978 together, the second alternatives go together, etc.
1980 First loop over alternatives. */
1981 alternative_mask preferred = curr_id->preferred_alternatives;
1982 if (only_alternative >= 0)
1983 preferred &= ALTERNATIVE_BIT (only_alternative);
1985 for (nalt = 0; nalt < n_alternatives; nalt++)
1987 /* Loop over operands for one constraint alternative. */
1988 if (!TEST_BIT (preferred, nalt))
1989 continue;
1991 curr_small_class_check++;
1992 overall = losers = addr_losers = 0;
1993 static_reject = reject = reload_nregs = reload_sum = 0;
1994 for (nop = 0; nop < n_operands; nop++)
1996 int inc = (curr_static_id
1997 ->operand_alternative[nalt * n_operands + nop].reject);
1998 if (lra_dump_file != NULL && inc != 0)
1999 fprintf (lra_dump_file,
2000 " Staticly defined alt reject+=%d\n", inc);
2001 static_reject += inc;
2003 reject += static_reject;
2004 early_clobbered_regs_num = 0;
2006 for (nop = 0; nop < n_operands; nop++)
2008 const char *p;
2009 char *end;
2010 int len, c, m, i, opalt_num, this_alternative_matches;
2011 bool win, did_match, offmemok, early_clobber_p;
2012 /* false => this operand can be reloaded somehow for this
2013 alternative. */
2014 bool badop;
2015 /* true => this operand can be reloaded if the alternative
2016 allows regs. */
2017 bool winreg;
2018 /* True if a constant forced into memory would be OK for
2019 this operand. */
2020 bool constmemok;
2021 enum reg_class this_alternative, this_costly_alternative;
2022 HARD_REG_SET this_alternative_set, this_costly_alternative_set;
2023 bool this_alternative_match_win, this_alternative_win;
2024 bool this_alternative_offmemok;
2025 bool scratch_p;
2026 machine_mode mode;
2027 enum constraint_num cn;
2029 opalt_num = nalt * n_operands + nop;
2030 if (curr_static_id->operand_alternative[opalt_num].anything_ok)
2032 /* Fast track for no constraints at all. */
2033 curr_alt[nop] = NO_REGS;
2034 CLEAR_HARD_REG_SET (curr_alt_set[nop]);
2035 curr_alt_win[nop] = true;
2036 curr_alt_match_win[nop] = false;
2037 curr_alt_offmemok[nop] = false;
2038 curr_alt_matches[nop] = -1;
2039 continue;
2042 op = no_subreg_reg_operand[nop];
2043 mode = curr_operand_mode[nop];
2045 win = did_match = winreg = offmemok = constmemok = false;
2046 badop = true;
2048 early_clobber_p = false;
2049 p = curr_static_id->operand_alternative[opalt_num].constraint;
2051 this_costly_alternative = this_alternative = NO_REGS;
2052 /* We update set of possible hard regs besides its class
2053 because reg class might be inaccurate. For example,
2054 union of LO_REGS (l), HI_REGS(h), and STACK_REG(k) in ARM
2055 is translated in HI_REGS because classes are merged by
2056 pairs and there is no accurate intermediate class. */
2057 CLEAR_HARD_REG_SET (this_alternative_set);
2058 CLEAR_HARD_REG_SET (this_costly_alternative_set);
2059 this_alternative_win = false;
2060 this_alternative_match_win = false;
2061 this_alternative_offmemok = false;
2062 this_alternative_matches = -1;
2064 /* An empty constraint should be excluded by the fast
2065 track. */
2066 lra_assert (*p != 0 && *p != ',');
2068 op_reject = 0;
2069 /* Scan this alternative's specs for this operand; set WIN
2070 if the operand fits any letter in this alternative.
2071 Otherwise, clear BADOP if this operand could fit some
2072 letter after reloads, or set WINREG if this operand could
2073 fit after reloads provided the constraint allows some
2074 registers. */
2075 costly_p = false;
2078 switch ((c = *p, len = CONSTRAINT_LEN (c, p)), c)
2080 case '\0':
2081 len = 0;
2082 break;
2083 case ',':
2084 c = '\0';
2085 break;
2087 case '&':
2088 early_clobber_p = true;
2089 break;
2091 case '$':
2092 op_reject += LRA_MAX_REJECT;
2093 break;
2094 case '^':
2095 op_reject += LRA_LOSER_COST_FACTOR;
2096 break;
2098 case '#':
2099 /* Ignore rest of this alternative. */
2100 c = '\0';
2101 break;
2103 case '0': case '1': case '2': case '3': case '4':
2104 case '5': case '6': case '7': case '8': case '9':
2106 int m_hregno;
2107 bool match_p;
2109 m = strtoul (p, &end, 10);
2110 p = end;
2111 len = 0;
2112 lra_assert (nop > m);
2114 /* Reject matches if we don't know which operand is
2115 bigger. This situation would arguably be a bug in
2116 an .md pattern, but could also occur in a user asm. */
2117 if (!ordered_p (GET_MODE_SIZE (biggest_mode[m]),
2118 GET_MODE_SIZE (biggest_mode[nop])))
2119 break;
2121 this_alternative_matches = m;
2122 m_hregno = get_hard_regno (*curr_id->operand_loc[m], false);
2123 /* We are supposed to match a previous operand.
2124 If we do, we win if that one did. If we do
2125 not, count both of the operands as losers.
2126 (This is too conservative, since most of the
2127 time only a single reload insn will be needed
2128 to make the two operands win. As a result,
2129 this alternative may be rejected when it is
2130 actually desirable.) */
2131 match_p = false;
2132 if (operands_match_p (*curr_id->operand_loc[nop],
2133 *curr_id->operand_loc[m], m_hregno))
2135 /* We should reject matching of an early
2136 clobber operand if the matching operand is
2137 not dying in the insn. */
2138 if (! curr_static_id->operand[m].early_clobber
2139 || operand_reg[nop] == NULL_RTX
2140 || (find_regno_note (curr_insn, REG_DEAD,
2141 REGNO (op))
2142 || REGNO (op) == REGNO (operand_reg[m])))
2143 match_p = true;
2145 if (match_p)
2147 /* If we are matching a non-offsettable
2148 address where an offsettable address was
2149 expected, then we must reject this
2150 combination, because we can't reload
2151 it. */
2152 if (curr_alt_offmemok[m]
2153 && MEM_P (*curr_id->operand_loc[m])
2154 && curr_alt[m] == NO_REGS && ! curr_alt_win[m])
2155 continue;
2157 else
2159 /* Operands don't match. Both operands must
2160 allow a reload register, otherwise we
2161 cannot make them match. */
2162 if (curr_alt[m] == NO_REGS)
2163 break;
2164 /* Retroactively mark the operand we had to
2165 match as a loser, if it wasn't already and
2166 it wasn't matched to a register constraint
2167 (e.g it might be matched by memory). */
2168 if (curr_alt_win[m]
2169 && (operand_reg[m] == NULL_RTX
2170 || hard_regno[m] < 0))
2172 losers++;
2173 reload_nregs
2174 += (ira_reg_class_max_nregs[curr_alt[m]]
2175 [GET_MODE (*curr_id->operand_loc[m])]);
2178 /* Prefer matching earlyclobber alternative as
2179 it results in less hard regs required for
2180 the insn than a non-matching earlyclobber
2181 alternative. */
2182 if (curr_static_id->operand[m].early_clobber)
2184 if (lra_dump_file != NULL)
2185 fprintf
2186 (lra_dump_file,
2187 " %d Matching earlyclobber alt:"
2188 " reject--\n",
2189 nop);
2190 reject--;
2192 /* Otherwise we prefer no matching
2193 alternatives because it gives more freedom
2194 in RA. */
2195 else if (operand_reg[nop] == NULL_RTX
2196 || (find_regno_note (curr_insn, REG_DEAD,
2197 REGNO (operand_reg[nop]))
2198 == NULL_RTX))
2200 if (lra_dump_file != NULL)
2201 fprintf
2202 (lra_dump_file,
2203 " %d Matching alt: reject+=2\n",
2204 nop);
2205 reject += 2;
2208 /* If we have to reload this operand and some
2209 previous operand also had to match the same
2210 thing as this operand, we don't know how to do
2211 that. */
2212 if (!match_p || !curr_alt_win[m])
2214 for (i = 0; i < nop; i++)
2215 if (curr_alt_matches[i] == m)
2216 break;
2217 if (i < nop)
2218 break;
2220 else
2221 did_match = true;
2223 /* This can be fixed with reloads if the operand
2224 we are supposed to match can be fixed with
2225 reloads. */
2226 badop = false;
2227 this_alternative = curr_alt[m];
2228 COPY_HARD_REG_SET (this_alternative_set, curr_alt_set[m]);
2229 winreg = this_alternative != NO_REGS;
2230 break;
2233 case 'g':
2234 if (MEM_P (op)
2235 || general_constant_p (op)
2236 || spilled_pseudo_p (op))
2237 win = true;
2238 cl = GENERAL_REGS;
2239 goto reg;
2241 default:
2242 cn = lookup_constraint (p);
2243 switch (get_constraint_type (cn))
2245 case CT_REGISTER:
2246 cl = reg_class_for_constraint (cn);
2247 if (cl != NO_REGS)
2248 goto reg;
2249 break;
2251 case CT_CONST_INT:
2252 if (CONST_INT_P (op)
2253 && insn_const_int_ok_for_constraint (INTVAL (op), cn))
2254 win = true;
2255 break;
2257 case CT_MEMORY:
2258 if (MEM_P (op)
2259 && satisfies_memory_constraint_p (op, cn))
2260 win = true;
2261 else if (spilled_pseudo_p (op))
2262 win = true;
2264 /* If we didn't already win, we can reload constants
2265 via force_const_mem or put the pseudo value into
2266 memory, or make other memory by reloading the
2267 address like for 'o'. */
2268 if (CONST_POOL_OK_P (mode, op)
2269 || MEM_P (op) || REG_P (op)
2270 /* We can restore the equiv insn by a
2271 reload. */
2272 || equiv_substition_p[nop])
2273 badop = false;
2274 constmemok = true;
2275 offmemok = true;
2276 break;
2278 case CT_ADDRESS:
2279 /* If we didn't already win, we can reload the address
2280 into a base register. */
2281 if (satisfies_address_constraint_p (op, cn))
2282 win = true;
2283 cl = base_reg_class (VOIDmode, ADDR_SPACE_GENERIC,
2284 ADDRESS, SCRATCH);
2285 badop = false;
2286 goto reg;
2288 case CT_FIXED_FORM:
2289 if (constraint_satisfied_p (op, cn))
2290 win = true;
2291 break;
2293 case CT_SPECIAL_MEMORY:
2294 if (MEM_P (op)
2295 && satisfies_memory_constraint_p (op, cn))
2296 win = true;
2297 else if (spilled_pseudo_p (op))
2298 win = true;
2299 break;
2301 break;
2303 reg:
2304 this_alternative = reg_class_subunion[this_alternative][cl];
2305 IOR_HARD_REG_SET (this_alternative_set,
2306 reg_class_contents[cl]);
2307 if (costly_p)
2309 this_costly_alternative
2310 = reg_class_subunion[this_costly_alternative][cl];
2311 IOR_HARD_REG_SET (this_costly_alternative_set,
2312 reg_class_contents[cl]);
2314 if (mode == BLKmode)
2315 break;
2316 winreg = true;
2317 if (REG_P (op))
2319 if (hard_regno[nop] >= 0
2320 && in_hard_reg_set_p (this_alternative_set,
2321 mode, hard_regno[nop]))
2322 win = true;
2323 else if (hard_regno[nop] < 0
2324 && in_class_p (op, this_alternative, NULL))
2325 win = true;
2327 break;
2329 if (c != ' ' && c != '\t')
2330 costly_p = c == '*';
2332 while ((p += len), c);
2334 scratch_p = (operand_reg[nop] != NULL_RTX
2335 && lra_former_scratch_p (REGNO (operand_reg[nop])));
2336 /* Record which operands fit this alternative. */
2337 if (win)
2339 this_alternative_win = true;
2340 if (operand_reg[nop] != NULL_RTX)
2342 if (hard_regno[nop] >= 0)
2344 if (in_hard_reg_set_p (this_costly_alternative_set,
2345 mode, hard_regno[nop]))
2347 if (lra_dump_file != NULL)
2348 fprintf (lra_dump_file,
2349 " %d Costly set: reject++\n",
2350 nop);
2351 reject++;
2354 else
2356 /* Prefer won reg to spilled pseudo under other
2357 equal conditions for possibe inheritance. */
2358 if (! scratch_p)
2360 if (lra_dump_file != NULL)
2361 fprintf
2362 (lra_dump_file,
2363 " %d Non pseudo reload: reject++\n",
2364 nop);
2365 reject++;
2367 if (in_class_p (operand_reg[nop],
2368 this_costly_alternative, NULL))
2370 if (lra_dump_file != NULL)
2371 fprintf
2372 (lra_dump_file,
2373 " %d Non pseudo costly reload:"
2374 " reject++\n",
2375 nop);
2376 reject++;
2379 /* We simulate the behavior of old reload here.
2380 Although scratches need hard registers and it
2381 might result in spilling other pseudos, no reload
2382 insns are generated for the scratches. So it
2383 might cost something but probably less than old
2384 reload pass believes. */
2385 if (scratch_p)
2387 if (lra_dump_file != NULL)
2388 fprintf (lra_dump_file,
2389 " %d Scratch win: reject+=2\n",
2390 nop);
2391 reject += 2;
2395 else if (did_match)
2396 this_alternative_match_win = true;
2397 else
2399 int const_to_mem = 0;
2400 bool no_regs_p;
2402 reject += op_reject;
2403 /* Never do output reload of stack pointer. It makes
2404 impossible to do elimination when SP is changed in
2405 RTL. */
2406 if (op == stack_pointer_rtx && ! frame_pointer_needed
2407 && curr_static_id->operand[nop].type != OP_IN)
2408 goto fail;
2410 /* If this alternative asks for a specific reg class, see if there
2411 is at least one allocatable register in that class. */
2412 no_regs_p
2413 = (this_alternative == NO_REGS
2414 || (hard_reg_set_subset_p
2415 (reg_class_contents[this_alternative],
2416 lra_no_alloc_regs)));
2418 /* For asms, verify that the class for this alternative is possible
2419 for the mode that is specified. */
2420 if (!no_regs_p && INSN_CODE (curr_insn) < 0)
2422 int i;
2423 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
2424 if (targetm.hard_regno_mode_ok (i, mode)
2425 && in_hard_reg_set_p (reg_class_contents[this_alternative],
2426 mode, i))
2427 break;
2428 if (i == FIRST_PSEUDO_REGISTER)
2429 winreg = false;
2432 /* If this operand accepts a register, and if the
2433 register class has at least one allocatable register,
2434 then this operand can be reloaded. */
2435 if (winreg && !no_regs_p)
2436 badop = false;
2438 if (badop)
2440 if (lra_dump_file != NULL)
2441 fprintf (lra_dump_file,
2442 " alt=%d: Bad operand -- refuse\n",
2443 nalt);
2444 goto fail;
2447 if (this_alternative != NO_REGS)
2449 HARD_REG_SET available_regs;
2451 COPY_HARD_REG_SET (available_regs,
2452 reg_class_contents[this_alternative]);
2453 AND_COMPL_HARD_REG_SET
2454 (available_regs,
2455 ira_prohibited_class_mode_regs[this_alternative][mode]);
2456 AND_COMPL_HARD_REG_SET (available_regs, lra_no_alloc_regs);
2457 if (hard_reg_set_empty_p (available_regs))
2459 /* There are no hard regs holding a value of given
2460 mode. */
2461 if (offmemok)
2463 this_alternative = NO_REGS;
2464 if (lra_dump_file != NULL)
2465 fprintf (lra_dump_file,
2466 " %d Using memory because of"
2467 " a bad mode: reject+=2\n",
2468 nop);
2469 reject += 2;
2471 else
2473 if (lra_dump_file != NULL)
2474 fprintf (lra_dump_file,
2475 " alt=%d: Wrong mode -- refuse\n",
2476 nalt);
2477 goto fail;
2482 /* If not assigned pseudo has a class which a subset of
2483 required reg class, it is a less costly alternative
2484 as the pseudo still can get a hard reg of necessary
2485 class. */
2486 if (! no_regs_p && REG_P (op) && hard_regno[nop] < 0
2487 && (cl = get_reg_class (REGNO (op))) != NO_REGS
2488 && ira_class_subset_p[this_alternative][cl])
2490 if (lra_dump_file != NULL)
2491 fprintf
2492 (lra_dump_file,
2493 " %d Super set class reg: reject-=3\n", nop);
2494 reject -= 3;
2497 this_alternative_offmemok = offmemok;
2498 if (this_costly_alternative != NO_REGS)
2500 if (lra_dump_file != NULL)
2501 fprintf (lra_dump_file,
2502 " %d Costly loser: reject++\n", nop);
2503 reject++;
2505 /* If the operand is dying, has a matching constraint,
2506 and satisfies constraints of the matched operand
2507 which failed to satisfy the own constraints, most probably
2508 the reload for this operand will be gone. */
2509 if (this_alternative_matches >= 0
2510 && !curr_alt_win[this_alternative_matches]
2511 && REG_P (op)
2512 && find_regno_note (curr_insn, REG_DEAD, REGNO (op))
2513 && (hard_regno[nop] >= 0
2514 ? in_hard_reg_set_p (this_alternative_set,
2515 mode, hard_regno[nop])
2516 : in_class_p (op, this_alternative, NULL)))
2518 if (lra_dump_file != NULL)
2519 fprintf
2520 (lra_dump_file,
2521 " %d Dying matched operand reload: reject++\n",
2522 nop);
2523 reject++;
2525 else
2527 /* Strict_low_part requires to reload the register
2528 not the sub-register. In this case we should
2529 check that a final reload hard reg can hold the
2530 value mode. */
2531 if (curr_static_id->operand[nop].strict_low
2532 && REG_P (op)
2533 && hard_regno[nop] < 0
2534 && GET_CODE (*curr_id->operand_loc[nop]) == SUBREG
2535 && ira_class_hard_regs_num[this_alternative] > 0
2536 && (!targetm.hard_regno_mode_ok
2537 (ira_class_hard_regs[this_alternative][0],
2538 GET_MODE (*curr_id->operand_loc[nop]))))
2540 if (lra_dump_file != NULL)
2541 fprintf
2542 (lra_dump_file,
2543 " alt=%d: Strict low subreg reload -- refuse\n",
2544 nalt);
2545 goto fail;
2547 losers++;
2549 if (operand_reg[nop] != NULL_RTX
2550 /* Output operands and matched input operands are
2551 not inherited. The following conditions do not
2552 exactly describe the previous statement but they
2553 are pretty close. */
2554 && curr_static_id->operand[nop].type != OP_OUT
2555 && (this_alternative_matches < 0
2556 || curr_static_id->operand[nop].type != OP_IN))
2558 int last_reload = (lra_reg_info[ORIGINAL_REGNO
2559 (operand_reg[nop])]
2560 .last_reload);
2562 /* The value of reload_sum has sense only if we
2563 process insns in their order. It happens only on
2564 the first constraints sub-pass when we do most of
2565 reload work. */
2566 if (lra_constraint_iter == 1 && last_reload > bb_reload_num)
2567 reload_sum += last_reload - bb_reload_num;
2569 /* If this is a constant that is reloaded into the
2570 desired class by copying it to memory first, count
2571 that as another reload. This is consistent with
2572 other code and is required to avoid choosing another
2573 alternative when the constant is moved into memory.
2574 Note that the test here is precisely the same as in
2575 the code below that calls force_const_mem. */
2576 if (CONST_POOL_OK_P (mode, op)
2577 && ((targetm.preferred_reload_class
2578 (op, this_alternative) == NO_REGS)
2579 || no_input_reloads_p))
2581 const_to_mem = 1;
2582 if (! no_regs_p)
2583 losers++;
2586 /* Alternative loses if it requires a type of reload not
2587 permitted for this insn. We can always reload
2588 objects with a REG_UNUSED note. */
2589 if ((curr_static_id->operand[nop].type != OP_IN
2590 && no_output_reloads_p
2591 && ! find_reg_note (curr_insn, REG_UNUSED, op))
2592 || (curr_static_id->operand[nop].type != OP_OUT
2593 && no_input_reloads_p && ! const_to_mem)
2594 || (this_alternative_matches >= 0
2595 && (no_input_reloads_p
2596 || (no_output_reloads_p
2597 && (curr_static_id->operand
2598 [this_alternative_matches].type != OP_IN)
2599 && ! find_reg_note (curr_insn, REG_UNUSED,
2600 no_subreg_reg_operand
2601 [this_alternative_matches])))))
2603 if (lra_dump_file != NULL)
2604 fprintf
2605 (lra_dump_file,
2606 " alt=%d: No input/otput reload -- refuse\n",
2607 nalt);
2608 goto fail;
2611 /* Alternative loses if it required class pseudo can not
2612 hold value of required mode. Such insns can be
2613 described by insn definitions with mode iterators. */
2614 if (GET_MODE (*curr_id->operand_loc[nop]) != VOIDmode
2615 && ! hard_reg_set_empty_p (this_alternative_set)
2616 /* It is common practice for constraints to use a
2617 class which does not have actually enough regs to
2618 hold the value (e.g. x86 AREG for mode requiring
2619 more one general reg). Therefore we have 2
2620 conditions to check that the reload pseudo can
2621 not hold the mode value. */
2622 && (!targetm.hard_regno_mode_ok
2623 (ira_class_hard_regs[this_alternative][0],
2624 GET_MODE (*curr_id->operand_loc[nop])))
2625 /* The above condition is not enough as the first
2626 reg in ira_class_hard_regs can be not aligned for
2627 multi-words mode values. */
2628 && (prohibited_class_reg_set_mode_p
2629 (this_alternative, this_alternative_set,
2630 GET_MODE (*curr_id->operand_loc[nop]))))
2632 if (lra_dump_file != NULL)
2633 fprintf (lra_dump_file,
2634 " alt=%d: reload pseudo for op %d "
2635 " can not hold the mode value -- refuse\n",
2636 nalt, nop);
2637 goto fail;
2640 /* Check strong discouragement of reload of non-constant
2641 into class THIS_ALTERNATIVE. */
2642 if (! CONSTANT_P (op) && ! no_regs_p
2643 && (targetm.preferred_reload_class
2644 (op, this_alternative) == NO_REGS
2645 || (curr_static_id->operand[nop].type == OP_OUT
2646 && (targetm.preferred_output_reload_class
2647 (op, this_alternative) == NO_REGS))))
2649 if (lra_dump_file != NULL)
2650 fprintf (lra_dump_file,
2651 " %d Non-prefered reload: reject+=%d\n",
2652 nop, LRA_MAX_REJECT);
2653 reject += LRA_MAX_REJECT;
2656 if (! (MEM_P (op) && offmemok)
2657 && ! (const_to_mem && constmemok))
2659 /* We prefer to reload pseudos over reloading other
2660 things, since such reloads may be able to be
2661 eliminated later. So bump REJECT in other cases.
2662 Don't do this in the case where we are forcing a
2663 constant into memory and it will then win since
2664 we don't want to have a different alternative
2665 match then. */
2666 if (! (REG_P (op) && REGNO (op) >= FIRST_PSEUDO_REGISTER))
2668 if (lra_dump_file != NULL)
2669 fprintf
2670 (lra_dump_file,
2671 " %d Non-pseudo reload: reject+=2\n",
2672 nop);
2673 reject += 2;
2676 if (! no_regs_p)
2677 reload_nregs
2678 += ira_reg_class_max_nregs[this_alternative][mode];
2680 if (SMALL_REGISTER_CLASS_P (this_alternative))
2682 if (lra_dump_file != NULL)
2683 fprintf
2684 (lra_dump_file,
2685 " %d Small class reload: reject+=%d\n",
2686 nop, LRA_LOSER_COST_FACTOR / 2);
2687 reject += LRA_LOSER_COST_FACTOR / 2;
2691 /* We are trying to spill pseudo into memory. It is
2692 usually more costly than moving to a hard register
2693 although it might takes the same number of
2694 reloads.
2696 Non-pseudo spill may happen also. Suppose a target allows both
2697 register and memory in the operand constraint alternatives,
2698 then it's typical that an eliminable register has a substition
2699 of "base + offset" which can either be reloaded by a simple
2700 "new_reg <= base + offset" which will match the register
2701 constraint, or a similar reg addition followed by further spill
2702 to and reload from memory which will match the memory
2703 constraint, but this memory spill will be much more costly
2704 usually.
2706 Code below increases the reject for both pseudo and non-pseudo
2707 spill. */
2708 if (no_regs_p
2709 && !(MEM_P (op) && offmemok)
2710 && !(REG_P (op) && hard_regno[nop] < 0))
2712 if (lra_dump_file != NULL)
2713 fprintf
2714 (lra_dump_file,
2715 " %d Spill %spseudo into memory: reject+=3\n",
2716 nop, REG_P (op) ? "" : "Non-");
2717 reject += 3;
2718 if (VECTOR_MODE_P (mode))
2720 /* Spilling vectors into memory is usually more
2721 costly as they contain big values. */
2722 if (lra_dump_file != NULL)
2723 fprintf
2724 (lra_dump_file,
2725 " %d Spill vector pseudo: reject+=2\n",
2726 nop);
2727 reject += 2;
2731 /* When we use an operand requiring memory in given
2732 alternative, the insn should write *and* read the
2733 value to/from memory it is costly in comparison with
2734 an insn alternative which does not use memory
2735 (e.g. register or immediate operand). We exclude
2736 memory operand for such case as we can satisfy the
2737 memory constraints by reloading address. */
2738 if (no_regs_p && offmemok && !MEM_P (op))
2740 if (lra_dump_file != NULL)
2741 fprintf
2742 (lra_dump_file,
2743 " Using memory insn operand %d: reject+=3\n",
2744 nop);
2745 reject += 3;
2748 /* If reload requires moving value through secondary
2749 memory, it will need one more insn at least. */
2750 if (this_alternative != NO_REGS
2751 && REG_P (op) && (cl = get_reg_class (REGNO (op))) != NO_REGS
2752 && ((curr_static_id->operand[nop].type != OP_OUT
2753 && targetm.secondary_memory_needed (GET_MODE (op), cl,
2754 this_alternative))
2755 || (curr_static_id->operand[nop].type != OP_IN
2756 && (targetm.secondary_memory_needed
2757 (GET_MODE (op), this_alternative, cl)))))
2758 losers++;
2760 /* Input reloads can be inherited more often than output
2761 reloads can be removed, so penalize output
2762 reloads. */
2763 if (!REG_P (op) || curr_static_id->operand[nop].type != OP_IN)
2765 if (lra_dump_file != NULL)
2766 fprintf
2767 (lra_dump_file,
2768 " %d Non input pseudo reload: reject++\n",
2769 nop);
2770 reject++;
2773 if (MEM_P (op) && offmemok)
2774 addr_losers++;
2775 else if (curr_static_id->operand[nop].type == OP_INOUT)
2777 if (lra_dump_file != NULL)
2778 fprintf
2779 (lra_dump_file,
2780 " %d Input/Output reload: reject+=%d\n",
2781 nop, LRA_LOSER_COST_FACTOR);
2782 reject += LRA_LOSER_COST_FACTOR;
2786 if (early_clobber_p && ! scratch_p)
2788 if (lra_dump_file != NULL)
2789 fprintf (lra_dump_file,
2790 " %d Early clobber: reject++\n", nop);
2791 reject++;
2793 /* ??? We check early clobbers after processing all operands
2794 (see loop below) and there we update the costs more.
2795 Should we update the cost (may be approximately) here
2796 because of early clobber register reloads or it is a rare
2797 or non-important thing to be worth to do it. */
2798 overall = (losers * LRA_LOSER_COST_FACTOR + reject
2799 - (addr_losers == losers ? static_reject : 0));
2800 if ((best_losers == 0 || losers != 0) && best_overall < overall)
2802 if (lra_dump_file != NULL)
2803 fprintf (lra_dump_file,
2804 " alt=%d,overall=%d,losers=%d -- refuse\n",
2805 nalt, overall, losers);
2806 goto fail;
2809 if (update_and_check_small_class_inputs (nop, this_alternative))
2811 if (lra_dump_file != NULL)
2812 fprintf (lra_dump_file,
2813 " alt=%d, not enough small class regs -- refuse\n",
2814 nalt);
2815 goto fail;
2817 curr_alt[nop] = this_alternative;
2818 COPY_HARD_REG_SET (curr_alt_set[nop], this_alternative_set);
2819 curr_alt_win[nop] = this_alternative_win;
2820 curr_alt_match_win[nop] = this_alternative_match_win;
2821 curr_alt_offmemok[nop] = this_alternative_offmemok;
2822 curr_alt_matches[nop] = this_alternative_matches;
2824 if (this_alternative_matches >= 0
2825 && !did_match && !this_alternative_win)
2826 curr_alt_win[this_alternative_matches] = false;
2828 if (early_clobber_p && operand_reg[nop] != NULL_RTX)
2829 early_clobbered_nops[early_clobbered_regs_num++] = nop;
2832 if (curr_insn_set != NULL_RTX && n_operands == 2
2833 /* Prevent processing non-move insns. */
2834 && (GET_CODE (SET_SRC (curr_insn_set)) == SUBREG
2835 || SET_SRC (curr_insn_set) == no_subreg_reg_operand[1])
2836 && ((! curr_alt_win[0] && ! curr_alt_win[1]
2837 && REG_P (no_subreg_reg_operand[0])
2838 && REG_P (no_subreg_reg_operand[1])
2839 && (reg_in_class_p (no_subreg_reg_operand[0], curr_alt[1])
2840 || reg_in_class_p (no_subreg_reg_operand[1], curr_alt[0])))
2841 || (! curr_alt_win[0] && curr_alt_win[1]
2842 && REG_P (no_subreg_reg_operand[1])
2843 /* Check that we reload memory not the memory
2844 address. */
2845 && ! (curr_alt_offmemok[0]
2846 && MEM_P (no_subreg_reg_operand[0]))
2847 && reg_in_class_p (no_subreg_reg_operand[1], curr_alt[0]))
2848 || (curr_alt_win[0] && ! curr_alt_win[1]
2849 && REG_P (no_subreg_reg_operand[0])
2850 /* Check that we reload memory not the memory
2851 address. */
2852 && ! (curr_alt_offmemok[1]
2853 && MEM_P (no_subreg_reg_operand[1]))
2854 && reg_in_class_p (no_subreg_reg_operand[0], curr_alt[1])
2855 && (! CONST_POOL_OK_P (curr_operand_mode[1],
2856 no_subreg_reg_operand[1])
2857 || (targetm.preferred_reload_class
2858 (no_subreg_reg_operand[1],
2859 (enum reg_class) curr_alt[1]) != NO_REGS))
2860 /* If it is a result of recent elimination in move
2861 insn we can transform it into an add still by
2862 using this alternative. */
2863 && GET_CODE (no_subreg_reg_operand[1]) != PLUS
2864 /* Likewise if the source has been replaced with an
2865 equivalent value. This only happens once -- the reload
2866 will use the equivalent value instead of the register it
2867 replaces -- so there should be no danger of cycling. */
2868 && !equiv_substition_p[1])))
2870 /* We have a move insn and a new reload insn will be similar
2871 to the current insn. We should avoid such situation as
2872 it results in LRA cycling. */
2873 if (lra_dump_file != NULL)
2874 fprintf (lra_dump_file,
2875 " Cycle danger: overall += LRA_MAX_REJECT\n");
2876 overall += LRA_MAX_REJECT;
2878 ok_p = true;
2879 curr_alt_dont_inherit_ops_num = 0;
2880 for (nop = 0; nop < early_clobbered_regs_num; nop++)
2882 int i, j, clobbered_hard_regno, first_conflict_j, last_conflict_j;
2883 HARD_REG_SET temp_set;
2885 i = early_clobbered_nops[nop];
2886 if ((! curr_alt_win[i] && ! curr_alt_match_win[i])
2887 || hard_regno[i] < 0)
2888 continue;
2889 lra_assert (operand_reg[i] != NULL_RTX);
2890 clobbered_hard_regno = hard_regno[i];
2891 CLEAR_HARD_REG_SET (temp_set);
2892 add_to_hard_reg_set (&temp_set, biggest_mode[i], clobbered_hard_regno);
2893 first_conflict_j = last_conflict_j = -1;
2894 for (j = 0; j < n_operands; j++)
2895 if (j == i
2896 /* We don't want process insides of match_operator and
2897 match_parallel because otherwise we would process
2898 their operands once again generating a wrong
2899 code. */
2900 || curr_static_id->operand[j].is_operator)
2901 continue;
2902 else if ((curr_alt_matches[j] == i && curr_alt_match_win[j])
2903 || (curr_alt_matches[i] == j && curr_alt_match_win[i]))
2904 continue;
2905 /* If we don't reload j-th operand, check conflicts. */
2906 else if ((curr_alt_win[j] || curr_alt_match_win[j])
2907 && uses_hard_regs_p (*curr_id->operand_loc[j], temp_set))
2909 if (first_conflict_j < 0)
2910 first_conflict_j = j;
2911 last_conflict_j = j;
2913 if (last_conflict_j < 0)
2914 continue;
2915 /* If earlyclobber operand conflicts with another
2916 non-matching operand which is actually the same register
2917 as the earlyclobber operand, it is better to reload the
2918 another operand as an operand matching the earlyclobber
2919 operand can be also the same. */
2920 if (first_conflict_j == last_conflict_j
2921 && operand_reg[last_conflict_j] != NULL_RTX
2922 && ! curr_alt_match_win[last_conflict_j]
2923 && REGNO (operand_reg[i]) == REGNO (operand_reg[last_conflict_j]))
2925 curr_alt_win[last_conflict_j] = false;
2926 curr_alt_dont_inherit_ops[curr_alt_dont_inherit_ops_num++]
2927 = last_conflict_j;
2928 losers++;
2929 /* Early clobber was already reflected in REJECT. */
2930 lra_assert (reject > 0);
2931 if (lra_dump_file != NULL)
2932 fprintf
2933 (lra_dump_file,
2934 " %d Conflict early clobber reload: reject--\n",
2936 reject--;
2937 overall += LRA_LOSER_COST_FACTOR - 1;
2939 else
2941 /* We need to reload early clobbered register and the
2942 matched registers. */
2943 for (j = 0; j < n_operands; j++)
2944 if (curr_alt_matches[j] == i)
2946 curr_alt_match_win[j] = false;
2947 losers++;
2948 overall += LRA_LOSER_COST_FACTOR;
2950 if (! curr_alt_match_win[i])
2951 curr_alt_dont_inherit_ops[curr_alt_dont_inherit_ops_num++] = i;
2952 else
2954 /* Remember pseudos used for match reloads are never
2955 inherited. */
2956 lra_assert (curr_alt_matches[i] >= 0);
2957 curr_alt_win[curr_alt_matches[i]] = false;
2959 curr_alt_win[i] = curr_alt_match_win[i] = false;
2960 losers++;
2961 /* Early clobber was already reflected in REJECT. */
2962 lra_assert (reject > 0);
2963 if (lra_dump_file != NULL)
2964 fprintf
2965 (lra_dump_file,
2966 " %d Matched conflict early clobber reloads: "
2967 "reject--\n",
2969 reject--;
2970 overall += LRA_LOSER_COST_FACTOR - 1;
2973 if (lra_dump_file != NULL)
2974 fprintf (lra_dump_file, " alt=%d,overall=%d,losers=%d,rld_nregs=%d\n",
2975 nalt, overall, losers, reload_nregs);
2977 /* If this alternative can be made to work by reloading, and it
2978 needs less reloading than the others checked so far, record
2979 it as the chosen goal for reloading. */
2980 if ((best_losers != 0 && losers == 0)
2981 || (((best_losers == 0 && losers == 0)
2982 || (best_losers != 0 && losers != 0))
2983 && (best_overall > overall
2984 || (best_overall == overall
2985 /* If the cost of the reloads is the same,
2986 prefer alternative which requires minimal
2987 number of reload regs. */
2988 && (reload_nregs < best_reload_nregs
2989 || (reload_nregs == best_reload_nregs
2990 && (best_reload_sum < reload_sum
2991 || (best_reload_sum == reload_sum
2992 && nalt < goal_alt_number))))))))
2994 for (nop = 0; nop < n_operands; nop++)
2996 goal_alt_win[nop] = curr_alt_win[nop];
2997 goal_alt_match_win[nop] = curr_alt_match_win[nop];
2998 goal_alt_matches[nop] = curr_alt_matches[nop];
2999 goal_alt[nop] = curr_alt[nop];
3000 goal_alt_offmemok[nop] = curr_alt_offmemok[nop];
3002 goal_alt_dont_inherit_ops_num = curr_alt_dont_inherit_ops_num;
3003 for (nop = 0; nop < curr_alt_dont_inherit_ops_num; nop++)
3004 goal_alt_dont_inherit_ops[nop] = curr_alt_dont_inherit_ops[nop];
3005 goal_alt_swapped = curr_swapped;
3006 best_overall = overall;
3007 best_losers = losers;
3008 best_reload_nregs = reload_nregs;
3009 best_reload_sum = reload_sum;
3010 goal_alt_number = nalt;
3012 if (losers == 0)
3013 /* Everything is satisfied. Do not process alternatives
3014 anymore. */
3015 break;
3016 fail:
3019 return ok_p;
3022 /* Make reload base reg from address AD. */
3023 static rtx
3024 base_to_reg (struct address_info *ad)
3026 enum reg_class cl;
3027 int code = -1;
3028 rtx new_inner = NULL_RTX;
3029 rtx new_reg = NULL_RTX;
3030 rtx_insn *insn;
3031 rtx_insn *last_insn = get_last_insn();
3033 lra_assert (ad->disp == ad->disp_term);
3034 cl = base_reg_class (ad->mode, ad->as, ad->base_outer_code,
3035 get_index_code (ad));
3036 new_reg = lra_create_new_reg (GET_MODE (*ad->base), NULL_RTX,
3037 cl, "base");
3038 new_inner = simplify_gen_binary (PLUS, GET_MODE (new_reg), new_reg,
3039 ad->disp_term == NULL
3040 ? const0_rtx
3041 : *ad->disp_term);
3042 if (!valid_address_p (ad->mode, new_inner, ad->as))
3043 return NULL_RTX;
3044 insn = emit_insn (gen_rtx_SET (new_reg, *ad->base));
3045 code = recog_memoized (insn);
3046 if (code < 0)
3048 delete_insns_since (last_insn);
3049 return NULL_RTX;
3052 return new_inner;
3055 /* Make reload base reg + DISP from address AD. Return the new pseudo. */
3056 static rtx
3057 base_plus_disp_to_reg (struct address_info *ad, rtx disp)
3059 enum reg_class cl;
3060 rtx new_reg;
3062 lra_assert (ad->base == ad->base_term);
3063 cl = base_reg_class (ad->mode, ad->as, ad->base_outer_code,
3064 get_index_code (ad));
3065 new_reg = lra_create_new_reg (GET_MODE (*ad->base_term), NULL_RTX,
3066 cl, "base + disp");
3067 lra_emit_add (new_reg, *ad->base_term, disp);
3068 return new_reg;
3071 /* Make reload of index part of address AD. Return the new
3072 pseudo. */
3073 static rtx
3074 index_part_to_reg (struct address_info *ad)
3076 rtx new_reg;
3078 new_reg = lra_create_new_reg (GET_MODE (*ad->index), NULL_RTX,
3079 INDEX_REG_CLASS, "index term");
3080 expand_mult (GET_MODE (*ad->index), *ad->index_term,
3081 GEN_INT (get_index_scale (ad)), new_reg, 1);
3082 return new_reg;
3085 /* Return true if we can add a displacement to address AD, even if that
3086 makes the address invalid. The fix-up code requires any new address
3087 to be the sum of the BASE_TERM, INDEX and DISP_TERM fields. */
3088 static bool
3089 can_add_disp_p (struct address_info *ad)
3091 return (!ad->autoinc_p
3092 && ad->segment == NULL
3093 && ad->base == ad->base_term
3094 && ad->disp == ad->disp_term);
3097 /* Make equiv substitution in address AD. Return true if a substitution
3098 was made. */
3099 static bool
3100 equiv_address_substitution (struct address_info *ad)
3102 rtx base_reg, new_base_reg, index_reg, new_index_reg, *base_term, *index_term;
3103 poly_int64 disp;
3104 HOST_WIDE_INT scale;
3105 bool change_p;
3107 base_term = strip_subreg (ad->base_term);
3108 if (base_term == NULL)
3109 base_reg = new_base_reg = NULL_RTX;
3110 else
3112 base_reg = *base_term;
3113 new_base_reg = get_equiv_with_elimination (base_reg, curr_insn);
3115 index_term = strip_subreg (ad->index_term);
3116 if (index_term == NULL)
3117 index_reg = new_index_reg = NULL_RTX;
3118 else
3120 index_reg = *index_term;
3121 new_index_reg = get_equiv_with_elimination (index_reg, curr_insn);
3123 if (base_reg == new_base_reg && index_reg == new_index_reg)
3124 return false;
3125 disp = 0;
3126 change_p = false;
3127 if (lra_dump_file != NULL)
3129 fprintf (lra_dump_file, "Changing address in insn %d ",
3130 INSN_UID (curr_insn));
3131 dump_value_slim (lra_dump_file, *ad->outer, 1);
3133 if (base_reg != new_base_reg)
3135 poly_int64 offset;
3136 if (REG_P (new_base_reg))
3138 *base_term = new_base_reg;
3139 change_p = true;
3141 else if (GET_CODE (new_base_reg) == PLUS
3142 && REG_P (XEXP (new_base_reg, 0))
3143 && poly_int_rtx_p (XEXP (new_base_reg, 1), &offset)
3144 && can_add_disp_p (ad))
3146 disp += offset;
3147 *base_term = XEXP (new_base_reg, 0);
3148 change_p = true;
3150 if (ad->base_term2 != NULL)
3151 *ad->base_term2 = *ad->base_term;
3153 if (index_reg != new_index_reg)
3155 poly_int64 offset;
3156 if (REG_P (new_index_reg))
3158 *index_term = new_index_reg;
3159 change_p = true;
3161 else if (GET_CODE (new_index_reg) == PLUS
3162 && REG_P (XEXP (new_index_reg, 0))
3163 && poly_int_rtx_p (XEXP (new_index_reg, 1), &offset)
3164 && can_add_disp_p (ad)
3165 && (scale = get_index_scale (ad)))
3167 disp += offset * scale;
3168 *index_term = XEXP (new_index_reg, 0);
3169 change_p = true;
3172 if (maybe_ne (disp, 0))
3174 if (ad->disp != NULL)
3175 *ad->disp = plus_constant (GET_MODE (*ad->inner), *ad->disp, disp);
3176 else
3178 *ad->inner = plus_constant (GET_MODE (*ad->inner), *ad->inner, disp);
3179 update_address (ad);
3181 change_p = true;
3183 if (lra_dump_file != NULL)
3185 if (! change_p)
3186 fprintf (lra_dump_file, " -- no change\n");
3187 else
3189 fprintf (lra_dump_file, " on equiv ");
3190 dump_value_slim (lra_dump_file, *ad->outer, 1);
3191 fprintf (lra_dump_file, "\n");
3194 return change_p;
3197 /* Major function to make reloads for an address in operand NOP or
3198 check its correctness (If CHECK_ONLY_P is true). The supported
3199 cases are:
3201 1) an address that existed before LRA started, at which point it
3202 must have been valid. These addresses are subject to elimination
3203 and may have become invalid due to the elimination offset being out
3204 of range.
3206 2) an address created by forcing a constant to memory
3207 (force_const_to_mem). The initial form of these addresses might
3208 not be valid, and it is this function's job to make them valid.
3210 3) a frame address formed from a register and a (possibly zero)
3211 constant offset. As above, these addresses might not be valid and
3212 this function must make them so.
3214 Add reloads to the lists *BEFORE and *AFTER. We might need to add
3215 reloads to *AFTER because of inc/dec, {pre, post} modify in the
3216 address. Return true for any RTL change.
3218 The function is a helper function which does not produce all
3219 transformations (when CHECK_ONLY_P is false) which can be
3220 necessary. It does just basic steps. To do all necessary
3221 transformations use function process_address. */
3222 static bool
3223 process_address_1 (int nop, bool check_only_p,
3224 rtx_insn **before, rtx_insn **after)
3226 struct address_info ad;
3227 rtx new_reg;
3228 HOST_WIDE_INT scale;
3229 rtx op = *curr_id->operand_loc[nop];
3230 const char *constraint = curr_static_id->operand[nop].constraint;
3231 enum constraint_num cn = lookup_constraint (constraint);
3232 bool change_p = false;
3234 if (MEM_P (op)
3235 && GET_MODE (op) == BLKmode
3236 && GET_CODE (XEXP (op, 0)) == SCRATCH)
3237 return false;
3239 if (insn_extra_address_constraint (cn))
3240 decompose_lea_address (&ad, curr_id->operand_loc[nop]);
3241 /* Do not attempt to decompose arbitrary addresses generated by combine
3242 for asm operands with loose constraints, e.g 'X'. */
3243 else if (MEM_P (op)
3244 && !(INSN_CODE (curr_insn) < 0
3245 && get_constraint_type (cn) == CT_FIXED_FORM
3246 && constraint_satisfied_p (op, cn)))
3247 decompose_mem_address (&ad, op);
3248 else if (GET_CODE (op) == SUBREG
3249 && MEM_P (SUBREG_REG (op)))
3250 decompose_mem_address (&ad, SUBREG_REG (op));
3251 else
3252 return false;
3253 /* If INDEX_REG_CLASS is assigned to base_term already and isn't to
3254 index_term, swap them so to avoid assigning INDEX_REG_CLASS to both
3255 when INDEX_REG_CLASS is a single register class. */
3256 if (ad.base_term != NULL
3257 && ad.index_term != NULL
3258 && ira_class_hard_regs_num[INDEX_REG_CLASS] == 1
3259 && REG_P (*ad.base_term)
3260 && REG_P (*ad.index_term)
3261 && in_class_p (*ad.base_term, INDEX_REG_CLASS, NULL)
3262 && ! in_class_p (*ad.index_term, INDEX_REG_CLASS, NULL))
3264 std::swap (ad.base, ad.index);
3265 std::swap (ad.base_term, ad.index_term);
3267 if (! check_only_p)
3268 change_p = equiv_address_substitution (&ad);
3269 if (ad.base_term != NULL
3270 && (process_addr_reg
3271 (ad.base_term, check_only_p, before,
3272 (ad.autoinc_p
3273 && !(REG_P (*ad.base_term)
3274 && find_regno_note (curr_insn, REG_DEAD,
3275 REGNO (*ad.base_term)) != NULL_RTX)
3276 ? after : NULL),
3277 base_reg_class (ad.mode, ad.as, ad.base_outer_code,
3278 get_index_code (&ad)))))
3280 change_p = true;
3281 if (ad.base_term2 != NULL)
3282 *ad.base_term2 = *ad.base_term;
3284 if (ad.index_term != NULL
3285 && process_addr_reg (ad.index_term, check_only_p,
3286 before, NULL, INDEX_REG_CLASS))
3287 change_p = true;
3289 /* Target hooks sometimes don't treat extra-constraint addresses as
3290 legitimate address_operands, so handle them specially. */
3291 if (insn_extra_address_constraint (cn)
3292 && satisfies_address_constraint_p (&ad, cn))
3293 return change_p;
3295 if (check_only_p)
3296 return change_p;
3298 /* There are three cases where the shape of *AD.INNER may now be invalid:
3300 1) the original address was valid, but either elimination or
3301 equiv_address_substitution was applied and that made
3302 the address invalid.
3304 2) the address is an invalid symbolic address created by
3305 force_const_to_mem.
3307 3) the address is a frame address with an invalid offset.
3309 4) the address is a frame address with an invalid base.
3311 All these cases involve a non-autoinc address, so there is no
3312 point revalidating other types. */
3313 if (ad.autoinc_p || valid_address_p (&ad))
3314 return change_p;
3316 /* Any index existed before LRA started, so we can assume that the
3317 presence and shape of the index is valid. */
3318 push_to_sequence (*before);
3319 lra_assert (ad.disp == ad.disp_term);
3320 if (ad.base == NULL)
3322 if (ad.index == NULL)
3324 rtx_insn *insn;
3325 rtx_insn *last = get_last_insn ();
3326 int code = -1;
3327 enum reg_class cl = base_reg_class (ad.mode, ad.as,
3328 SCRATCH, SCRATCH);
3329 rtx addr = *ad.inner;
3331 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "addr");
3332 if (HAVE_lo_sum)
3334 /* addr => lo_sum (new_base, addr), case (2) above. */
3335 insn = emit_insn (gen_rtx_SET
3336 (new_reg,
3337 gen_rtx_HIGH (Pmode, copy_rtx (addr))));
3338 code = recog_memoized (insn);
3339 if (code >= 0)
3341 *ad.inner = gen_rtx_LO_SUM (Pmode, new_reg, addr);
3342 if (! valid_address_p (ad.mode, *ad.outer, ad.as))
3344 /* Try to put lo_sum into register. */
3345 insn = emit_insn (gen_rtx_SET
3346 (new_reg,
3347 gen_rtx_LO_SUM (Pmode, new_reg, addr)));
3348 code = recog_memoized (insn);
3349 if (code >= 0)
3351 *ad.inner = new_reg;
3352 if (! valid_address_p (ad.mode, *ad.outer, ad.as))
3354 *ad.inner = addr;
3355 code = -1;
3361 if (code < 0)
3362 delete_insns_since (last);
3365 if (code < 0)
3367 /* addr => new_base, case (2) above. */
3368 lra_emit_move (new_reg, addr);
3370 for (insn = last == NULL_RTX ? get_insns () : NEXT_INSN (last);
3371 insn != NULL_RTX;
3372 insn = NEXT_INSN (insn))
3373 if (recog_memoized (insn) < 0)
3374 break;
3375 if (insn != NULL_RTX)
3377 /* Do nothing if we cannot generate right insns.
3378 This is analogous to reload pass behavior. */
3379 delete_insns_since (last);
3380 end_sequence ();
3381 return false;
3383 *ad.inner = new_reg;
3386 else
3388 /* index * scale + disp => new base + index * scale,
3389 case (1) above. */
3390 enum reg_class cl = base_reg_class (ad.mode, ad.as, PLUS,
3391 GET_CODE (*ad.index));
3393 lra_assert (INDEX_REG_CLASS != NO_REGS);
3394 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "disp");
3395 lra_emit_move (new_reg, *ad.disp);
3396 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3397 new_reg, *ad.index);
3400 else if (ad.index == NULL)
3402 int regno;
3403 enum reg_class cl;
3404 rtx set;
3405 rtx_insn *insns, *last_insn;
3406 /* Try to reload base into register only if the base is invalid
3407 for the address but with valid offset, case (4) above. */
3408 start_sequence ();
3409 new_reg = base_to_reg (&ad);
3411 /* base + disp => new base, cases (1) and (3) above. */
3412 /* Another option would be to reload the displacement into an
3413 index register. However, postreload has code to optimize
3414 address reloads that have the same base and different
3415 displacements, so reloading into an index register would
3416 not necessarily be a win. */
3417 if (new_reg == NULL_RTX)
3419 /* See if the target can split the displacement into a
3420 legitimate new displacement from a local anchor. */
3421 gcc_assert (ad.disp == ad.disp_term);
3422 poly_int64 orig_offset;
3423 rtx offset1, offset2;
3424 if (poly_int_rtx_p (*ad.disp, &orig_offset)
3425 && targetm.legitimize_address_displacement (&offset1, &offset2,
3426 orig_offset,
3427 ad.mode))
3429 new_reg = base_plus_disp_to_reg (&ad, offset1);
3430 new_reg = gen_rtx_PLUS (GET_MODE (new_reg), new_reg, offset2);
3432 else
3433 new_reg = base_plus_disp_to_reg (&ad, *ad.disp);
3435 insns = get_insns ();
3436 last_insn = get_last_insn ();
3437 /* If we generated at least two insns, try last insn source as
3438 an address. If we succeed, we generate one less insn. */
3439 if (REG_P (new_reg)
3440 && last_insn != insns
3441 && (set = single_set (last_insn)) != NULL_RTX
3442 && GET_CODE (SET_SRC (set)) == PLUS
3443 && REG_P (XEXP (SET_SRC (set), 0))
3444 && CONSTANT_P (XEXP (SET_SRC (set), 1)))
3446 *ad.inner = SET_SRC (set);
3447 if (valid_address_p (ad.mode, *ad.outer, ad.as))
3449 *ad.base_term = XEXP (SET_SRC (set), 0);
3450 *ad.disp_term = XEXP (SET_SRC (set), 1);
3451 cl = base_reg_class (ad.mode, ad.as, ad.base_outer_code,
3452 get_index_code (&ad));
3453 regno = REGNO (*ad.base_term);
3454 if (regno >= FIRST_PSEUDO_REGISTER
3455 && cl != lra_get_allocno_class (regno))
3456 lra_change_class (regno, cl, " Change to", true);
3457 new_reg = SET_SRC (set);
3458 delete_insns_since (PREV_INSN (last_insn));
3461 end_sequence ();
3462 emit_insn (insns);
3463 *ad.inner = new_reg;
3465 else if (ad.disp_term != NULL)
3467 /* base + scale * index + disp => new base + scale * index,
3468 case (1) above. */
3469 gcc_assert (ad.disp == ad.disp_term);
3470 new_reg = base_plus_disp_to_reg (&ad, *ad.disp);
3471 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3472 new_reg, *ad.index);
3474 else if ((scale = get_index_scale (&ad)) == 1)
3476 /* The last transformation to one reg will be made in
3477 curr_insn_transform function. */
3478 end_sequence ();
3479 return false;
3481 else if (scale != 0)
3483 /* base + scale * index => base + new_reg,
3484 case (1) above.
3485 Index part of address may become invalid. For example, we
3486 changed pseudo on the equivalent memory and a subreg of the
3487 pseudo onto the memory of different mode for which the scale is
3488 prohibitted. */
3489 new_reg = index_part_to_reg (&ad);
3490 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3491 *ad.base_term, new_reg);
3493 else
3495 enum reg_class cl = base_reg_class (ad.mode, ad.as,
3496 SCRATCH, SCRATCH);
3497 rtx addr = *ad.inner;
3499 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "addr");
3500 /* addr => new_base. */
3501 lra_emit_move (new_reg, addr);
3502 *ad.inner = new_reg;
3504 *before = get_insns ();
3505 end_sequence ();
3506 return true;
3509 /* If CHECK_ONLY_P is false, do address reloads until it is necessary.
3510 Use process_address_1 as a helper function. Return true for any
3511 RTL changes.
3513 If CHECK_ONLY_P is true, just check address correctness. Return
3514 false if the address correct. */
3515 static bool
3516 process_address (int nop, bool check_only_p,
3517 rtx_insn **before, rtx_insn **after)
3519 bool res = false;
3521 while (process_address_1 (nop, check_only_p, before, after))
3523 if (check_only_p)
3524 return true;
3525 res = true;
3527 return res;
3530 /* Emit insns to reload VALUE into a new register. VALUE is an
3531 auto-increment or auto-decrement RTX whose operand is a register or
3532 memory location; so reloading involves incrementing that location.
3533 IN is either identical to VALUE, or some cheaper place to reload
3534 value being incremented/decremented from.
3536 INC_AMOUNT is the number to increment or decrement by (always
3537 positive and ignored for POST_MODIFY/PRE_MODIFY).
3539 Return pseudo containing the result. */
3540 static rtx
3541 emit_inc (enum reg_class new_rclass, rtx in, rtx value, poly_int64 inc_amount)
3543 /* REG or MEM to be copied and incremented. */
3544 rtx incloc = XEXP (value, 0);
3545 /* Nonzero if increment after copying. */
3546 int post = (GET_CODE (value) == POST_DEC || GET_CODE (value) == POST_INC
3547 || GET_CODE (value) == POST_MODIFY);
3548 rtx_insn *last;
3549 rtx inc;
3550 rtx_insn *add_insn;
3551 int code;
3552 rtx real_in = in == value ? incloc : in;
3553 rtx result;
3554 bool plus_p = true;
3556 if (GET_CODE (value) == PRE_MODIFY || GET_CODE (value) == POST_MODIFY)
3558 lra_assert (GET_CODE (XEXP (value, 1)) == PLUS
3559 || GET_CODE (XEXP (value, 1)) == MINUS);
3560 lra_assert (rtx_equal_p (XEXP (XEXP (value, 1), 0), XEXP (value, 0)));
3561 plus_p = GET_CODE (XEXP (value, 1)) == PLUS;
3562 inc = XEXP (XEXP (value, 1), 1);
3564 else
3566 if (GET_CODE (value) == PRE_DEC || GET_CODE (value) == POST_DEC)
3567 inc_amount = -inc_amount;
3569 inc = gen_int_mode (inc_amount, GET_MODE (value));
3572 if (! post && REG_P (incloc))
3573 result = incloc;
3574 else
3575 result = lra_create_new_reg (GET_MODE (value), value, new_rclass,
3576 "INC/DEC result");
3578 if (real_in != result)
3580 /* First copy the location to the result register. */
3581 lra_assert (REG_P (result));
3582 emit_insn (gen_move_insn (result, real_in));
3585 /* We suppose that there are insns to add/sub with the constant
3586 increment permitted in {PRE/POST)_{DEC/INC/MODIFY}. At least the
3587 old reload worked with this assumption. If the assumption
3588 becomes wrong, we should use approach in function
3589 base_plus_disp_to_reg. */
3590 if (in == value)
3592 /* See if we can directly increment INCLOC. */
3593 last = get_last_insn ();
3594 add_insn = emit_insn (plus_p
3595 ? gen_add2_insn (incloc, inc)
3596 : gen_sub2_insn (incloc, inc));
3598 code = recog_memoized (add_insn);
3599 if (code >= 0)
3601 if (! post && result != incloc)
3602 emit_insn (gen_move_insn (result, incloc));
3603 return result;
3605 delete_insns_since (last);
3608 /* If couldn't do the increment directly, must increment in RESULT.
3609 The way we do this depends on whether this is pre- or
3610 post-increment. For pre-increment, copy INCLOC to the reload
3611 register, increment it there, then save back. */
3612 if (! post)
3614 if (real_in != result)
3615 emit_insn (gen_move_insn (result, real_in));
3616 if (plus_p)
3617 emit_insn (gen_add2_insn (result, inc));
3618 else
3619 emit_insn (gen_sub2_insn (result, inc));
3620 if (result != incloc)
3621 emit_insn (gen_move_insn (incloc, result));
3623 else
3625 /* Post-increment.
3627 Because this might be a jump insn or a compare, and because
3628 RESULT may not be available after the insn in an input
3629 reload, we must do the incrementing before the insn being
3630 reloaded for.
3632 We have already copied IN to RESULT. Increment the copy in
3633 RESULT, save that back, then decrement RESULT so it has
3634 the original value. */
3635 if (plus_p)
3636 emit_insn (gen_add2_insn (result, inc));
3637 else
3638 emit_insn (gen_sub2_insn (result, inc));
3639 emit_insn (gen_move_insn (incloc, result));
3640 /* Restore non-modified value for the result. We prefer this
3641 way because it does not require an additional hard
3642 register. */
3643 if (plus_p)
3645 poly_int64 offset;
3646 if (poly_int_rtx_p (inc, &offset))
3647 emit_insn (gen_add2_insn (result,
3648 gen_int_mode (-offset,
3649 GET_MODE (result))));
3650 else
3651 emit_insn (gen_sub2_insn (result, inc));
3653 else
3654 emit_insn (gen_add2_insn (result, inc));
3656 return result;
3659 /* Return true if the current move insn does not need processing as we
3660 already know that it satisfies its constraints. */
3661 static bool
3662 simple_move_p (void)
3664 rtx dest, src;
3665 enum reg_class dclass, sclass;
3667 lra_assert (curr_insn_set != NULL_RTX);
3668 dest = SET_DEST (curr_insn_set);
3669 src = SET_SRC (curr_insn_set);
3671 /* If the instruction has multiple sets we need to process it even if it
3672 is single_set. This can happen if one or more of the SETs are dead.
3673 See PR73650. */
3674 if (multiple_sets (curr_insn))
3675 return false;
3677 return ((dclass = get_op_class (dest)) != NO_REGS
3678 && (sclass = get_op_class (src)) != NO_REGS
3679 /* The backend guarantees that register moves of cost 2
3680 never need reloads. */
3681 && targetm.register_move_cost (GET_MODE (src), sclass, dclass) == 2);
3684 /* Swap operands NOP and NOP + 1. */
3685 static inline void
3686 swap_operands (int nop)
3688 std::swap (curr_operand_mode[nop], curr_operand_mode[nop + 1]);
3689 std::swap (original_subreg_reg_mode[nop], original_subreg_reg_mode[nop + 1]);
3690 std::swap (*curr_id->operand_loc[nop], *curr_id->operand_loc[nop + 1]);
3691 std::swap (equiv_substition_p[nop], equiv_substition_p[nop + 1]);
3692 /* Swap the duplicates too. */
3693 lra_update_dup (curr_id, nop);
3694 lra_update_dup (curr_id, nop + 1);
3697 /* Main entry point of the constraint code: search the body of the
3698 current insn to choose the best alternative. It is mimicking insn
3699 alternative cost calculation model of former reload pass. That is
3700 because machine descriptions were written to use this model. This
3701 model can be changed in future. Make commutative operand exchange
3702 if it is chosen.
3704 if CHECK_ONLY_P is false, do RTL changes to satisfy the
3705 constraints. Return true if any change happened during function
3706 call.
3708 If CHECK_ONLY_P is true then don't do any transformation. Just
3709 check that the insn satisfies all constraints. If the insn does
3710 not satisfy any constraint, return true. */
3711 static bool
3712 curr_insn_transform (bool check_only_p)
3714 int i, j, k;
3715 int n_operands;
3716 int n_alternatives;
3717 int n_outputs;
3718 int commutative;
3719 signed char goal_alt_matched[MAX_RECOG_OPERANDS][MAX_RECOG_OPERANDS];
3720 signed char match_inputs[MAX_RECOG_OPERANDS + 1];
3721 signed char outputs[MAX_RECOG_OPERANDS + 1];
3722 rtx_insn *before, *after;
3723 bool alt_p = false;
3724 /* Flag that the insn has been changed through a transformation. */
3725 bool change_p;
3726 bool sec_mem_p;
3727 bool use_sec_mem_p;
3728 int max_regno_before;
3729 int reused_alternative_num;
3731 curr_insn_set = single_set (curr_insn);
3732 if (curr_insn_set != NULL_RTX && simple_move_p ())
3733 return false;
3735 no_input_reloads_p = no_output_reloads_p = false;
3736 goal_alt_number = -1;
3737 change_p = sec_mem_p = false;
3738 /* JUMP_INSNs and CALL_INSNs are not allowed to have any output
3739 reloads; neither are insns that SET cc0. Insns that use CC0 are
3740 not allowed to have any input reloads. */
3741 if (JUMP_P (curr_insn) || CALL_P (curr_insn))
3742 no_output_reloads_p = true;
3744 if (HAVE_cc0 && reg_referenced_p (cc0_rtx, PATTERN (curr_insn)))
3745 no_input_reloads_p = true;
3746 if (HAVE_cc0 && reg_set_p (cc0_rtx, PATTERN (curr_insn)))
3747 no_output_reloads_p = true;
3749 n_operands = curr_static_id->n_operands;
3750 n_alternatives = curr_static_id->n_alternatives;
3752 /* Just return "no reloads" if insn has no operands with
3753 constraints. */
3754 if (n_operands == 0 || n_alternatives == 0)
3755 return false;
3757 max_regno_before = max_reg_num ();
3759 for (i = 0; i < n_operands; i++)
3761 goal_alt_matched[i][0] = -1;
3762 goal_alt_matches[i] = -1;
3765 commutative = curr_static_id->commutative;
3767 /* Now see what we need for pseudos that didn't get hard regs or got
3768 the wrong kind of hard reg. For this, we must consider all the
3769 operands together against the register constraints. */
3771 best_losers = best_overall = INT_MAX;
3772 best_reload_sum = 0;
3774 curr_swapped = false;
3775 goal_alt_swapped = false;
3777 if (! check_only_p)
3778 /* Make equivalence substitution and memory subreg elimination
3779 before address processing because an address legitimacy can
3780 depend on memory mode. */
3781 for (i = 0; i < n_operands; i++)
3783 rtx op, subst, old;
3784 bool op_change_p = false;
3786 if (curr_static_id->operand[i].is_operator)
3787 continue;
3789 old = op = *curr_id->operand_loc[i];
3790 if (GET_CODE (old) == SUBREG)
3791 old = SUBREG_REG (old);
3792 subst = get_equiv_with_elimination (old, curr_insn);
3793 original_subreg_reg_mode[i] = VOIDmode;
3794 equiv_substition_p[i] = false;
3795 if (subst != old)
3797 equiv_substition_p[i] = true;
3798 subst = copy_rtx (subst);
3799 lra_assert (REG_P (old));
3800 if (GET_CODE (op) != SUBREG)
3801 *curr_id->operand_loc[i] = subst;
3802 else
3804 SUBREG_REG (op) = subst;
3805 if (GET_MODE (subst) == VOIDmode)
3806 original_subreg_reg_mode[i] = GET_MODE (old);
3808 if (lra_dump_file != NULL)
3810 fprintf (lra_dump_file,
3811 "Changing pseudo %d in operand %i of insn %u on equiv ",
3812 REGNO (old), i, INSN_UID (curr_insn));
3813 dump_value_slim (lra_dump_file, subst, 1);
3814 fprintf (lra_dump_file, "\n");
3816 op_change_p = change_p = true;
3818 if (simplify_operand_subreg (i, GET_MODE (old)) || op_change_p)
3820 change_p = true;
3821 lra_update_dup (curr_id, i);
3825 /* Reload address registers and displacements. We do it before
3826 finding an alternative because of memory constraints. */
3827 before = after = NULL;
3828 for (i = 0; i < n_operands; i++)
3829 if (! curr_static_id->operand[i].is_operator
3830 && process_address (i, check_only_p, &before, &after))
3832 if (check_only_p)
3833 return true;
3834 change_p = true;
3835 lra_update_dup (curr_id, i);
3838 if (change_p)
3839 /* If we've changed the instruction then any alternative that
3840 we chose previously may no longer be valid. */
3841 lra_set_used_insn_alternative (curr_insn, -1);
3843 if (! check_only_p && curr_insn_set != NULL_RTX
3844 && check_and_process_move (&change_p, &sec_mem_p))
3845 return change_p;
3847 try_swapped:
3849 reused_alternative_num = check_only_p ? -1 : curr_id->used_insn_alternative;
3850 if (lra_dump_file != NULL && reused_alternative_num >= 0)
3851 fprintf (lra_dump_file, "Reusing alternative %d for insn #%u\n",
3852 reused_alternative_num, INSN_UID (curr_insn));
3854 if (process_alt_operands (reused_alternative_num))
3855 alt_p = true;
3857 if (check_only_p)
3858 return ! alt_p || best_losers != 0;
3860 /* If insn is commutative (it's safe to exchange a certain pair of
3861 operands) then we need to try each alternative twice, the second
3862 time matching those two operands as if we had exchanged them. To
3863 do this, really exchange them in operands.
3865 If we have just tried the alternatives the second time, return
3866 operands to normal and drop through. */
3868 if (reused_alternative_num < 0 && commutative >= 0)
3870 curr_swapped = !curr_swapped;
3871 if (curr_swapped)
3873 swap_operands (commutative);
3874 goto try_swapped;
3876 else
3877 swap_operands (commutative);
3880 if (! alt_p && ! sec_mem_p)
3882 /* No alternative works with reloads?? */
3883 if (INSN_CODE (curr_insn) >= 0)
3884 fatal_insn ("unable to generate reloads for:", curr_insn);
3885 error_for_asm (curr_insn,
3886 "inconsistent operand constraints in an %<asm%>");
3887 /* Avoid further trouble with this insn. Don't generate use
3888 pattern here as we could use the insn SP offset. */
3889 lra_set_insn_deleted (curr_insn);
3890 return true;
3893 /* If the best alternative is with operands 1 and 2 swapped, swap
3894 them. Update the operand numbers of any reloads already
3895 pushed. */
3897 if (goal_alt_swapped)
3899 if (lra_dump_file != NULL)
3900 fprintf (lra_dump_file, " Commutative operand exchange in insn %u\n",
3901 INSN_UID (curr_insn));
3903 /* Swap the duplicates too. */
3904 swap_operands (commutative);
3905 change_p = true;
3908 /* Some targets' TARGET_SECONDARY_MEMORY_NEEDED (e.g. x86) are defined
3909 too conservatively. So we use the secondary memory only if there
3910 is no any alternative without reloads. */
3911 use_sec_mem_p = false;
3912 if (! alt_p)
3913 use_sec_mem_p = true;
3914 else if (sec_mem_p)
3916 for (i = 0; i < n_operands; i++)
3917 if (! goal_alt_win[i] && ! goal_alt_match_win[i])
3918 break;
3919 use_sec_mem_p = i < n_operands;
3922 if (use_sec_mem_p)
3924 int in = -1, out = -1;
3925 rtx new_reg, src, dest, rld;
3926 machine_mode sec_mode, rld_mode;
3928 lra_assert (curr_insn_set != NULL_RTX && sec_mem_p);
3929 dest = SET_DEST (curr_insn_set);
3930 src = SET_SRC (curr_insn_set);
3931 for (i = 0; i < n_operands; i++)
3932 if (*curr_id->operand_loc[i] == dest)
3933 out = i;
3934 else if (*curr_id->operand_loc[i] == src)
3935 in = i;
3936 for (i = 0; i < curr_static_id->n_dups; i++)
3937 if (out < 0 && *curr_id->dup_loc[i] == dest)
3938 out = curr_static_id->dup_num[i];
3939 else if (in < 0 && *curr_id->dup_loc[i] == src)
3940 in = curr_static_id->dup_num[i];
3941 lra_assert (out >= 0 && in >= 0
3942 && curr_static_id->operand[out].type == OP_OUT
3943 && curr_static_id->operand[in].type == OP_IN);
3944 rld = partial_subreg_p (GET_MODE (src), GET_MODE (dest)) ? src : dest;
3945 rld_mode = GET_MODE (rld);
3946 sec_mode = targetm.secondary_memory_needed_mode (rld_mode);
3947 new_reg = lra_create_new_reg (sec_mode, NULL_RTX,
3948 NO_REGS, "secondary");
3949 /* If the mode is changed, it should be wider. */
3950 lra_assert (!partial_subreg_p (sec_mode, rld_mode));
3951 if (sec_mode != rld_mode)
3953 /* If the target says specifically to use another mode for
3954 secondary memory moves we can not reuse the original
3955 insn. */
3956 after = emit_spill_move (false, new_reg, dest);
3957 lra_process_new_insns (curr_insn, NULL, after,
3958 "Inserting the sec. move");
3959 /* We may have non null BEFORE here (e.g. after address
3960 processing. */
3961 push_to_sequence (before);
3962 before = emit_spill_move (true, new_reg, src);
3963 emit_insn (before);
3964 before = get_insns ();
3965 end_sequence ();
3966 lra_process_new_insns (curr_insn, before, NULL, "Changing on");
3967 lra_set_insn_deleted (curr_insn);
3969 else if (dest == rld)
3971 *curr_id->operand_loc[out] = new_reg;
3972 lra_update_dup (curr_id, out);
3973 after = emit_spill_move (false, new_reg, dest);
3974 lra_process_new_insns (curr_insn, NULL, after,
3975 "Inserting the sec. move");
3977 else
3979 *curr_id->operand_loc[in] = new_reg;
3980 lra_update_dup (curr_id, in);
3981 /* See comments above. */
3982 push_to_sequence (before);
3983 before = emit_spill_move (true, new_reg, src);
3984 emit_insn (before);
3985 before = get_insns ();
3986 end_sequence ();
3987 lra_process_new_insns (curr_insn, before, NULL,
3988 "Inserting the sec. move");
3990 lra_update_insn_regno_info (curr_insn);
3991 return true;
3994 lra_assert (goal_alt_number >= 0);
3995 lra_set_used_insn_alternative (curr_insn, goal_alt_number);
3997 if (lra_dump_file != NULL)
3999 const char *p;
4001 fprintf (lra_dump_file, " Choosing alt %d in insn %u:",
4002 goal_alt_number, INSN_UID (curr_insn));
4003 for (i = 0; i < n_operands; i++)
4005 p = (curr_static_id->operand_alternative
4006 [goal_alt_number * n_operands + i].constraint);
4007 if (*p == '\0')
4008 continue;
4009 fprintf (lra_dump_file, " (%d) ", i);
4010 for (; *p != '\0' && *p != ',' && *p != '#'; p++)
4011 fputc (*p, lra_dump_file);
4013 if (INSN_CODE (curr_insn) >= 0
4014 && (p = get_insn_name (INSN_CODE (curr_insn))) != NULL)
4015 fprintf (lra_dump_file, " {%s}", p);
4016 if (maybe_ne (curr_id->sp_offset, 0))
4018 fprintf (lra_dump_file, " (sp_off=");
4019 print_dec (curr_id->sp_offset, lra_dump_file);
4020 fprintf (lra_dump_file, ")");
4022 fprintf (lra_dump_file, "\n");
4025 /* Right now, for any pair of operands I and J that are required to
4026 match, with J < I, goal_alt_matches[I] is J. Add I to
4027 goal_alt_matched[J]. */
4029 for (i = 0; i < n_operands; i++)
4030 if ((j = goal_alt_matches[i]) >= 0)
4032 for (k = 0; goal_alt_matched[j][k] >= 0; k++)
4034 /* We allow matching one output operand and several input
4035 operands. */
4036 lra_assert (k == 0
4037 || (curr_static_id->operand[j].type == OP_OUT
4038 && curr_static_id->operand[i].type == OP_IN
4039 && (curr_static_id->operand
4040 [goal_alt_matched[j][0]].type == OP_IN)));
4041 goal_alt_matched[j][k] = i;
4042 goal_alt_matched[j][k + 1] = -1;
4045 for (i = 0; i < n_operands; i++)
4046 goal_alt_win[i] |= goal_alt_match_win[i];
4048 /* Any constants that aren't allowed and can't be reloaded into
4049 registers are here changed into memory references. */
4050 for (i = 0; i < n_operands; i++)
4051 if (goal_alt_win[i])
4053 int regno;
4054 enum reg_class new_class;
4055 rtx reg = *curr_id->operand_loc[i];
4057 if (GET_CODE (reg) == SUBREG)
4058 reg = SUBREG_REG (reg);
4060 if (REG_P (reg) && (regno = REGNO (reg)) >= FIRST_PSEUDO_REGISTER)
4062 bool ok_p = in_class_p (reg, goal_alt[i], &new_class);
4064 if (new_class != NO_REGS && get_reg_class (regno) != new_class)
4066 lra_assert (ok_p);
4067 lra_change_class (regno, new_class, " Change to", true);
4071 else
4073 const char *constraint;
4074 char c;
4075 rtx op = *curr_id->operand_loc[i];
4076 rtx subreg = NULL_RTX;
4077 machine_mode mode = curr_operand_mode[i];
4079 if (GET_CODE (op) == SUBREG)
4081 subreg = op;
4082 op = SUBREG_REG (op);
4083 mode = GET_MODE (op);
4086 if (CONST_POOL_OK_P (mode, op)
4087 && ((targetm.preferred_reload_class
4088 (op, (enum reg_class) goal_alt[i]) == NO_REGS)
4089 || no_input_reloads_p))
4091 rtx tem = force_const_mem (mode, op);
4093 change_p = true;
4094 if (subreg != NULL_RTX)
4095 tem = gen_rtx_SUBREG (mode, tem, SUBREG_BYTE (subreg));
4097 *curr_id->operand_loc[i] = tem;
4098 lra_update_dup (curr_id, i);
4099 process_address (i, false, &before, &after);
4101 /* If the alternative accepts constant pool refs directly
4102 there will be no reload needed at all. */
4103 if (subreg != NULL_RTX)
4104 continue;
4105 /* Skip alternatives before the one requested. */
4106 constraint = (curr_static_id->operand_alternative
4107 [goal_alt_number * n_operands + i].constraint);
4108 for (;
4109 (c = *constraint) && c != ',' && c != '#';
4110 constraint += CONSTRAINT_LEN (c, constraint))
4112 enum constraint_num cn = lookup_constraint (constraint);
4113 if ((insn_extra_memory_constraint (cn)
4114 || insn_extra_special_memory_constraint (cn))
4115 && satisfies_memory_constraint_p (tem, cn))
4116 break;
4118 if (c == '\0' || c == ',' || c == '#')
4119 continue;
4121 goal_alt_win[i] = true;
4125 n_outputs = 0;
4126 outputs[0] = -1;
4127 for (i = 0; i < n_operands; i++)
4129 int regno;
4130 bool optional_p = false;
4131 rtx old, new_reg;
4132 rtx op = *curr_id->operand_loc[i];
4134 if (goal_alt_win[i])
4136 if (goal_alt[i] == NO_REGS
4137 && REG_P (op)
4138 /* When we assign NO_REGS it means that we will not
4139 assign a hard register to the scratch pseudo by
4140 assigment pass and the scratch pseudo will be
4141 spilled. Spilled scratch pseudos are transformed
4142 back to scratches at the LRA end. */
4143 && lra_former_scratch_operand_p (curr_insn, i)
4144 && lra_former_scratch_p (REGNO (op)))
4146 int regno = REGNO (op);
4147 lra_change_class (regno, NO_REGS, " Change to", true);
4148 if (lra_get_regno_hard_regno (regno) >= 0)
4149 /* We don't have to mark all insn affected by the
4150 spilled pseudo as there is only one such insn, the
4151 current one. */
4152 reg_renumber[regno] = -1;
4153 lra_assert (bitmap_single_bit_set_p
4154 (&lra_reg_info[REGNO (op)].insn_bitmap));
4156 /* We can do an optional reload. If the pseudo got a hard
4157 reg, we might improve the code through inheritance. If
4158 it does not get a hard register we coalesce memory/memory
4159 moves later. Ignore move insns to avoid cycling. */
4160 if (! lra_simple_p
4161 && lra_undo_inheritance_iter < LRA_MAX_INHERITANCE_PASSES
4162 && goal_alt[i] != NO_REGS && REG_P (op)
4163 && (regno = REGNO (op)) >= FIRST_PSEUDO_REGISTER
4164 && regno < new_regno_start
4165 && ! lra_former_scratch_p (regno)
4166 && reg_renumber[regno] < 0
4167 /* Check that the optional reload pseudo will be able to
4168 hold given mode value. */
4169 && ! (prohibited_class_reg_set_mode_p
4170 (goal_alt[i], reg_class_contents[goal_alt[i]],
4171 PSEUDO_REGNO_MODE (regno)))
4172 && (curr_insn_set == NULL_RTX
4173 || !((REG_P (SET_SRC (curr_insn_set))
4174 || MEM_P (SET_SRC (curr_insn_set))
4175 || GET_CODE (SET_SRC (curr_insn_set)) == SUBREG)
4176 && (REG_P (SET_DEST (curr_insn_set))
4177 || MEM_P (SET_DEST (curr_insn_set))
4178 || GET_CODE (SET_DEST (curr_insn_set)) == SUBREG))))
4179 optional_p = true;
4180 else
4181 continue;
4184 /* Operands that match previous ones have already been handled. */
4185 if (goal_alt_matches[i] >= 0)
4186 continue;
4188 /* We should not have an operand with a non-offsettable address
4189 appearing where an offsettable address will do. It also may
4190 be a case when the address should be special in other words
4191 not a general one (e.g. it needs no index reg). */
4192 if (goal_alt_matched[i][0] == -1 && goal_alt_offmemok[i] && MEM_P (op))
4194 enum reg_class rclass;
4195 rtx *loc = &XEXP (op, 0);
4196 enum rtx_code code = GET_CODE (*loc);
4198 push_to_sequence (before);
4199 rclass = base_reg_class (GET_MODE (op), MEM_ADDR_SPACE (op),
4200 MEM, SCRATCH);
4201 if (GET_RTX_CLASS (code) == RTX_AUTOINC)
4202 new_reg = emit_inc (rclass, *loc, *loc,
4203 /* This value does not matter for MODIFY. */
4204 GET_MODE_SIZE (GET_MODE (op)));
4205 else if (get_reload_reg (OP_IN, Pmode, *loc, rclass, FALSE,
4206 "offsetable address", &new_reg))
4208 rtx addr = *loc;
4209 enum rtx_code code = GET_CODE (addr);
4211 if (code == AND && CONST_INT_P (XEXP (addr, 1)))
4212 /* (and ... (const_int -X)) is used to align to X bytes. */
4213 addr = XEXP (*loc, 0);
4214 lra_emit_move (new_reg, addr);
4215 if (addr != *loc)
4216 emit_move_insn (new_reg, gen_rtx_AND (GET_MODE (new_reg), new_reg, XEXP (*loc, 1)));
4218 before = get_insns ();
4219 end_sequence ();
4220 *loc = new_reg;
4221 lra_update_dup (curr_id, i);
4223 else if (goal_alt_matched[i][0] == -1)
4225 machine_mode mode;
4226 rtx reg, *loc;
4227 int hard_regno;
4228 enum op_type type = curr_static_id->operand[i].type;
4230 loc = curr_id->operand_loc[i];
4231 mode = curr_operand_mode[i];
4232 if (GET_CODE (*loc) == SUBREG)
4234 reg = SUBREG_REG (*loc);
4235 poly_int64 byte = SUBREG_BYTE (*loc);
4236 if (REG_P (reg)
4237 /* Strict_low_part requires reloading the register and not
4238 just the subreg. Likewise for a strict subreg no wider
4239 than a word for WORD_REGISTER_OPERATIONS targets. */
4240 && (curr_static_id->operand[i].strict_low
4241 || (!paradoxical_subreg_p (mode, GET_MODE (reg))
4242 && (hard_regno
4243 = get_try_hard_regno (REGNO (reg))) >= 0
4244 && (simplify_subreg_regno
4245 (hard_regno,
4246 GET_MODE (reg), byte, mode) < 0)
4247 && (goal_alt[i] == NO_REGS
4248 || (simplify_subreg_regno
4249 (ira_class_hard_regs[goal_alt[i]][0],
4250 GET_MODE (reg), byte, mode) >= 0)))
4251 || (partial_subreg_p (mode, GET_MODE (reg))
4252 && known_le (GET_MODE_SIZE (GET_MODE (reg)),
4253 UNITS_PER_WORD)
4254 && WORD_REGISTER_OPERATIONS)))
4256 /* An OP_INOUT is required when reloading a subreg of a
4257 mode wider than a word to ensure that data beyond the
4258 word being reloaded is preserved. Also automatically
4259 ensure that strict_low_part reloads are made into
4260 OP_INOUT which should already be true from the backend
4261 constraints. */
4262 if (type == OP_OUT
4263 && (curr_static_id->operand[i].strict_low
4264 || read_modify_subreg_p (*loc)))
4265 type = OP_INOUT;
4266 loc = &SUBREG_REG (*loc);
4267 mode = GET_MODE (*loc);
4270 old = *loc;
4271 if (get_reload_reg (type, mode, old, goal_alt[i],
4272 loc != curr_id->operand_loc[i], "", &new_reg)
4273 && type != OP_OUT)
4275 push_to_sequence (before);
4276 lra_emit_move (new_reg, old);
4277 before = get_insns ();
4278 end_sequence ();
4280 *loc = new_reg;
4281 if (type != OP_IN
4282 && find_reg_note (curr_insn, REG_UNUSED, old) == NULL_RTX)
4284 start_sequence ();
4285 lra_emit_move (type == OP_INOUT ? copy_rtx (old) : old, new_reg);
4286 emit_insn (after);
4287 after = get_insns ();
4288 end_sequence ();
4289 *loc = new_reg;
4291 for (j = 0; j < goal_alt_dont_inherit_ops_num; j++)
4292 if (goal_alt_dont_inherit_ops[j] == i)
4294 lra_set_regno_unique_value (REGNO (new_reg));
4295 break;
4297 lra_update_dup (curr_id, i);
4299 else if (curr_static_id->operand[i].type == OP_IN
4300 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4301 == OP_OUT
4302 || (curr_static_id->operand[goal_alt_matched[i][0]].type
4303 == OP_INOUT
4304 && (operands_match_p
4305 (*curr_id->operand_loc[i],
4306 *curr_id->operand_loc[goal_alt_matched[i][0]],
4307 -1)))))
4309 /* generate reloads for input and matched outputs. */
4310 match_inputs[0] = i;
4311 match_inputs[1] = -1;
4312 match_reload (goal_alt_matched[i][0], match_inputs, outputs,
4313 goal_alt[i], &before, &after,
4314 curr_static_id->operand_alternative
4315 [goal_alt_number * n_operands + goal_alt_matched[i][0]]
4316 .earlyclobber);
4318 else if ((curr_static_id->operand[i].type == OP_OUT
4319 || (curr_static_id->operand[i].type == OP_INOUT
4320 && (operands_match_p
4321 (*curr_id->operand_loc[i],
4322 *curr_id->operand_loc[goal_alt_matched[i][0]],
4323 -1))))
4324 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4325 == OP_IN))
4326 /* Generate reloads for output and matched inputs. */
4327 match_reload (i, goal_alt_matched[i], outputs, goal_alt[i], &before,
4328 &after, curr_static_id->operand_alternative
4329 [goal_alt_number * n_operands + i].earlyclobber);
4330 else if (curr_static_id->operand[i].type == OP_IN
4331 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4332 == OP_IN))
4334 /* Generate reloads for matched inputs. */
4335 match_inputs[0] = i;
4336 for (j = 0; (k = goal_alt_matched[i][j]) >= 0; j++)
4337 match_inputs[j + 1] = k;
4338 match_inputs[j + 1] = -1;
4339 match_reload (-1, match_inputs, outputs, goal_alt[i], &before,
4340 &after, false);
4342 else
4343 /* We must generate code in any case when function
4344 process_alt_operands decides that it is possible. */
4345 gcc_unreachable ();
4347 /* Memorise processed outputs so that output remaining to be processed
4348 can avoid using the same register value (see match_reload). */
4349 if (curr_static_id->operand[i].type == OP_OUT)
4351 outputs[n_outputs++] = i;
4352 outputs[n_outputs] = -1;
4355 if (optional_p)
4357 rtx reg = op;
4359 lra_assert (REG_P (reg));
4360 regno = REGNO (reg);
4361 op = *curr_id->operand_loc[i]; /* Substitution. */
4362 if (GET_CODE (op) == SUBREG)
4363 op = SUBREG_REG (op);
4364 gcc_assert (REG_P (op) && (int) REGNO (op) >= new_regno_start);
4365 bitmap_set_bit (&lra_optional_reload_pseudos, REGNO (op));
4366 lra_reg_info[REGNO (op)].restore_rtx = reg;
4367 if (lra_dump_file != NULL)
4368 fprintf (lra_dump_file,
4369 " Making reload reg %d for reg %d optional\n",
4370 REGNO (op), regno);
4373 if (before != NULL_RTX || after != NULL_RTX
4374 || max_regno_before != max_reg_num ())
4375 change_p = true;
4376 if (change_p)
4378 lra_update_operator_dups (curr_id);
4379 /* Something changes -- process the insn. */
4380 lra_update_insn_regno_info (curr_insn);
4382 lra_process_new_insns (curr_insn, before, after, "Inserting insn reload");
4383 return change_p;
4386 /* Return true if INSN satisfies all constraints. In other words, no
4387 reload insns are needed. */
4388 bool
4389 lra_constrain_insn (rtx_insn *insn)
4391 int saved_new_regno_start = new_regno_start;
4392 int saved_new_insn_uid_start = new_insn_uid_start;
4393 bool change_p;
4395 curr_insn = insn;
4396 curr_id = lra_get_insn_recog_data (curr_insn);
4397 curr_static_id = curr_id->insn_static_data;
4398 new_insn_uid_start = get_max_uid ();
4399 new_regno_start = max_reg_num ();
4400 change_p = curr_insn_transform (true);
4401 new_regno_start = saved_new_regno_start;
4402 new_insn_uid_start = saved_new_insn_uid_start;
4403 return ! change_p;
4406 /* Return true if X is in LIST. */
4407 static bool
4408 in_list_p (rtx x, rtx list)
4410 for (; list != NULL_RTX; list = XEXP (list, 1))
4411 if (XEXP (list, 0) == x)
4412 return true;
4413 return false;
4416 /* Return true if X contains an allocatable hard register (if
4417 HARD_REG_P) or a (spilled if SPILLED_P) pseudo. */
4418 static bool
4419 contains_reg_p (rtx x, bool hard_reg_p, bool spilled_p)
4421 int i, j;
4422 const char *fmt;
4423 enum rtx_code code;
4425 code = GET_CODE (x);
4426 if (REG_P (x))
4428 int regno = REGNO (x);
4429 HARD_REG_SET alloc_regs;
4431 if (hard_reg_p)
4433 if (regno >= FIRST_PSEUDO_REGISTER)
4434 regno = lra_get_regno_hard_regno (regno);
4435 if (regno < 0)
4436 return false;
4437 COMPL_HARD_REG_SET (alloc_regs, lra_no_alloc_regs);
4438 return overlaps_hard_reg_set_p (alloc_regs, GET_MODE (x), regno);
4440 else
4442 if (regno < FIRST_PSEUDO_REGISTER)
4443 return false;
4444 if (! spilled_p)
4445 return true;
4446 return lra_get_regno_hard_regno (regno) < 0;
4449 fmt = GET_RTX_FORMAT (code);
4450 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4452 if (fmt[i] == 'e')
4454 if (contains_reg_p (XEXP (x, i), hard_reg_p, spilled_p))
4455 return true;
4457 else if (fmt[i] == 'E')
4459 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4460 if (contains_reg_p (XVECEXP (x, i, j), hard_reg_p, spilled_p))
4461 return true;
4464 return false;
4467 /* Process all regs in location *LOC and change them on equivalent
4468 substitution. Return true if any change was done. */
4469 static bool
4470 loc_equivalence_change_p (rtx *loc)
4472 rtx subst, reg, x = *loc;
4473 bool result = false;
4474 enum rtx_code code = GET_CODE (x);
4475 const char *fmt;
4476 int i, j;
4478 if (code == SUBREG)
4480 reg = SUBREG_REG (x);
4481 if ((subst = get_equiv_with_elimination (reg, curr_insn)) != reg
4482 && GET_MODE (subst) == VOIDmode)
4484 /* We cannot reload debug location. Simplify subreg here
4485 while we know the inner mode. */
4486 *loc = simplify_gen_subreg (GET_MODE (x), subst,
4487 GET_MODE (reg), SUBREG_BYTE (x));
4488 return true;
4491 if (code == REG && (subst = get_equiv_with_elimination (x, curr_insn)) != x)
4493 *loc = subst;
4494 return true;
4497 /* Scan all the operand sub-expressions. */
4498 fmt = GET_RTX_FORMAT (code);
4499 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4501 if (fmt[i] == 'e')
4502 result = loc_equivalence_change_p (&XEXP (x, i)) || result;
4503 else if (fmt[i] == 'E')
4504 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4505 result
4506 = loc_equivalence_change_p (&XVECEXP (x, i, j)) || result;
4508 return result;
4511 /* Similar to loc_equivalence_change_p, but for use as
4512 simplify_replace_fn_rtx callback. DATA is insn for which the
4513 elimination is done. If it null we don't do the elimination. */
4514 static rtx
4515 loc_equivalence_callback (rtx loc, const_rtx, void *data)
4517 if (!REG_P (loc))
4518 return NULL_RTX;
4520 rtx subst = (data == NULL
4521 ? get_equiv (loc) : get_equiv_with_elimination (loc, (rtx_insn *) data));
4522 if (subst != loc)
4523 return subst;
4525 return NULL_RTX;
4528 /* Maximum number of generated reload insns per an insn. It is for
4529 preventing this pass cycling in a bug case. */
4530 #define MAX_RELOAD_INSNS_NUMBER LRA_MAX_INSN_RELOADS
4532 /* The current iteration number of this LRA pass. */
4533 int lra_constraint_iter;
4535 /* True if we substituted equiv which needs checking register
4536 allocation correctness because the equivalent value contains
4537 allocatable hard registers or when we restore multi-register
4538 pseudo. */
4539 bool lra_risky_transformations_p;
4541 /* Return true if REGNO is referenced in more than one block. */
4542 static bool
4543 multi_block_pseudo_p (int regno)
4545 basic_block bb = NULL;
4546 unsigned int uid;
4547 bitmap_iterator bi;
4549 if (regno < FIRST_PSEUDO_REGISTER)
4550 return false;
4552 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi)
4553 if (bb == NULL)
4554 bb = BLOCK_FOR_INSN (lra_insn_recog_data[uid]->insn);
4555 else if (BLOCK_FOR_INSN (lra_insn_recog_data[uid]->insn) != bb)
4556 return true;
4557 return false;
4560 /* Return true if LIST contains a deleted insn. */
4561 static bool
4562 contains_deleted_insn_p (rtx_insn_list *list)
4564 for (; list != NULL_RTX; list = list->next ())
4565 if (NOTE_P (list->insn ())
4566 && NOTE_KIND (list->insn ()) == NOTE_INSN_DELETED)
4567 return true;
4568 return false;
4571 /* Return true if X contains a pseudo dying in INSN. */
4572 static bool
4573 dead_pseudo_p (rtx x, rtx_insn *insn)
4575 int i, j;
4576 const char *fmt;
4577 enum rtx_code code;
4579 if (REG_P (x))
4580 return (insn != NULL_RTX
4581 && find_regno_note (insn, REG_DEAD, REGNO (x)) != NULL_RTX);
4582 code = GET_CODE (x);
4583 fmt = GET_RTX_FORMAT (code);
4584 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4586 if (fmt[i] == 'e')
4588 if (dead_pseudo_p (XEXP (x, i), insn))
4589 return true;
4591 else if (fmt[i] == 'E')
4593 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4594 if (dead_pseudo_p (XVECEXP (x, i, j), insn))
4595 return true;
4598 return false;
4601 /* Return true if INSN contains a dying pseudo in INSN right hand
4602 side. */
4603 static bool
4604 insn_rhs_dead_pseudo_p (rtx_insn *insn)
4606 rtx set = single_set (insn);
4608 gcc_assert (set != NULL);
4609 return dead_pseudo_p (SET_SRC (set), insn);
4612 /* Return true if any init insn of REGNO contains a dying pseudo in
4613 insn right hand side. */
4614 static bool
4615 init_insn_rhs_dead_pseudo_p (int regno)
4617 rtx_insn_list *insns = ira_reg_equiv[regno].init_insns;
4619 if (insns == NULL)
4620 return false;
4621 for (; insns != NULL_RTX; insns = insns->next ())
4622 if (insn_rhs_dead_pseudo_p (insns->insn ()))
4623 return true;
4624 return false;
4627 /* Return TRUE if REGNO has a reverse equivalence. The equivalence is
4628 reverse only if we have one init insn with given REGNO as a
4629 source. */
4630 static bool
4631 reverse_equiv_p (int regno)
4633 rtx_insn_list *insns = ira_reg_equiv[regno].init_insns;
4634 rtx set;
4636 if (insns == NULL)
4637 return false;
4638 if (! INSN_P (insns->insn ())
4639 || insns->next () != NULL)
4640 return false;
4641 if ((set = single_set (insns->insn ())) == NULL_RTX)
4642 return false;
4643 return REG_P (SET_SRC (set)) && (int) REGNO (SET_SRC (set)) == regno;
4646 /* Return TRUE if REGNO was reloaded in an equivalence init insn. We
4647 call this function only for non-reverse equivalence. */
4648 static bool
4649 contains_reloaded_insn_p (int regno)
4651 rtx set;
4652 rtx_insn_list *list = ira_reg_equiv[regno].init_insns;
4654 for (; list != NULL; list = list->next ())
4655 if ((set = single_set (list->insn ())) == NULL_RTX
4656 || ! REG_P (SET_DEST (set))
4657 || (int) REGNO (SET_DEST (set)) != regno)
4658 return true;
4659 return false;
4662 /* Entry function of LRA constraint pass. Return true if the
4663 constraint pass did change the code. */
4664 bool
4665 lra_constraints (bool first_p)
4667 bool changed_p;
4668 int i, hard_regno, new_insns_num;
4669 unsigned int min_len, new_min_len, uid;
4670 rtx set, x, reg, dest_reg;
4671 basic_block last_bb;
4672 bitmap_iterator bi;
4674 lra_constraint_iter++;
4675 if (lra_dump_file != NULL)
4676 fprintf (lra_dump_file, "\n********** Local #%d: **********\n\n",
4677 lra_constraint_iter);
4678 changed_p = false;
4679 if (pic_offset_table_rtx
4680 && REGNO (pic_offset_table_rtx) >= FIRST_PSEUDO_REGISTER)
4681 lra_risky_transformations_p = true;
4682 else
4683 /* On the first iteration we should check IRA assignment
4684 correctness. In rare cases, the assignments can be wrong as
4685 early clobbers operands are ignored in IRA. */
4686 lra_risky_transformations_p = first_p;
4687 new_insn_uid_start = get_max_uid ();
4688 new_regno_start = first_p ? lra_constraint_new_regno_start : max_reg_num ();
4689 /* Mark used hard regs for target stack size calulations. */
4690 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4691 if (lra_reg_info[i].nrefs != 0
4692 && (hard_regno = lra_get_regno_hard_regno (i)) >= 0)
4694 int j, nregs;
4696 nregs = hard_regno_nregs (hard_regno, lra_reg_info[i].biggest_mode);
4697 for (j = 0; j < nregs; j++)
4698 df_set_regs_ever_live (hard_regno + j, true);
4700 /* Do elimination before the equivalence processing as we can spill
4701 some pseudos during elimination. */
4702 lra_eliminate (false, first_p);
4703 auto_bitmap equiv_insn_bitmap (&reg_obstack);
4704 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4705 if (lra_reg_info[i].nrefs != 0)
4707 ira_reg_equiv[i].profitable_p = true;
4708 reg = regno_reg_rtx[i];
4709 if (lra_get_regno_hard_regno (i) < 0 && (x = get_equiv (reg)) != reg)
4711 bool pseudo_p = contains_reg_p (x, false, false);
4713 /* After RTL transformation, we can not guarantee that
4714 pseudo in the substitution was not reloaded which might
4715 make equivalence invalid. For example, in reverse
4716 equiv of p0
4718 p0 <- ...
4720 equiv_mem <- p0
4722 the memory address register was reloaded before the 2nd
4723 insn. */
4724 if ((! first_p && pseudo_p)
4725 /* We don't use DF for compilation speed sake. So it
4726 is problematic to update live info when we use an
4727 equivalence containing pseudos in more than one
4728 BB. */
4729 || (pseudo_p && multi_block_pseudo_p (i))
4730 /* If an init insn was deleted for some reason, cancel
4731 the equiv. We could update the equiv insns after
4732 transformations including an equiv insn deletion
4733 but it is not worthy as such cases are extremely
4734 rare. */
4735 || contains_deleted_insn_p (ira_reg_equiv[i].init_insns)
4736 /* If it is not a reverse equivalence, we check that a
4737 pseudo in rhs of the init insn is not dying in the
4738 insn. Otherwise, the live info at the beginning of
4739 the corresponding BB might be wrong after we
4740 removed the insn. When the equiv can be a
4741 constant, the right hand side of the init insn can
4742 be a pseudo. */
4743 || (! reverse_equiv_p (i)
4744 && (init_insn_rhs_dead_pseudo_p (i)
4745 /* If we reloaded the pseudo in an equivalence
4746 init insn, we can not remove the equiv init
4747 insns and the init insns might write into
4748 const memory in this case. */
4749 || contains_reloaded_insn_p (i)))
4750 /* Prevent access beyond equivalent memory for
4751 paradoxical subregs. */
4752 || (MEM_P (x)
4753 && maybe_gt (GET_MODE_SIZE (lra_reg_info[i].biggest_mode),
4754 GET_MODE_SIZE (GET_MODE (x))))
4755 || (pic_offset_table_rtx
4756 && ((CONST_POOL_OK_P (PSEUDO_REGNO_MODE (i), x)
4757 && (targetm.preferred_reload_class
4758 (x, lra_get_allocno_class (i)) == NO_REGS))
4759 || contains_symbol_ref_p (x))))
4760 ira_reg_equiv[i].defined_p = false;
4761 if (contains_reg_p (x, false, true))
4762 ira_reg_equiv[i].profitable_p = false;
4763 if (get_equiv (reg) != reg)
4764 bitmap_ior_into (equiv_insn_bitmap, &lra_reg_info[i].insn_bitmap);
4767 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4768 update_equiv (i);
4769 /* We should add all insns containing pseudos which should be
4770 substituted by their equivalences. */
4771 EXECUTE_IF_SET_IN_BITMAP (equiv_insn_bitmap, 0, uid, bi)
4772 lra_push_insn_by_uid (uid);
4773 min_len = lra_insn_stack_length ();
4774 new_insns_num = 0;
4775 last_bb = NULL;
4776 changed_p = false;
4777 while ((new_min_len = lra_insn_stack_length ()) != 0)
4779 curr_insn = lra_pop_insn ();
4780 --new_min_len;
4781 curr_bb = BLOCK_FOR_INSN (curr_insn);
4782 if (curr_bb != last_bb)
4784 last_bb = curr_bb;
4785 bb_reload_num = lra_curr_reload_num;
4787 if (min_len > new_min_len)
4789 min_len = new_min_len;
4790 new_insns_num = 0;
4792 if (new_insns_num > MAX_RELOAD_INSNS_NUMBER)
4793 internal_error
4794 ("Max. number of generated reload insns per insn is achieved (%d)\n",
4795 MAX_RELOAD_INSNS_NUMBER);
4796 new_insns_num++;
4797 if (DEBUG_INSN_P (curr_insn))
4799 /* We need to check equivalence in debug insn and change
4800 pseudo to the equivalent value if necessary. */
4801 curr_id = lra_get_insn_recog_data (curr_insn);
4802 if (bitmap_bit_p (equiv_insn_bitmap, INSN_UID (curr_insn)))
4804 rtx old = *curr_id->operand_loc[0];
4805 *curr_id->operand_loc[0]
4806 = simplify_replace_fn_rtx (old, NULL_RTX,
4807 loc_equivalence_callback, curr_insn);
4808 if (old != *curr_id->operand_loc[0])
4810 lra_update_insn_regno_info (curr_insn);
4811 changed_p = true;
4815 else if (INSN_P (curr_insn))
4817 if ((set = single_set (curr_insn)) != NULL_RTX)
4819 dest_reg = SET_DEST (set);
4820 /* The equivalence pseudo could be set up as SUBREG in a
4821 case when it is a call restore insn in a mode
4822 different from the pseudo mode. */
4823 if (GET_CODE (dest_reg) == SUBREG)
4824 dest_reg = SUBREG_REG (dest_reg);
4825 if ((REG_P (dest_reg)
4826 && (x = get_equiv (dest_reg)) != dest_reg
4827 /* Remove insns which set up a pseudo whose value
4828 can not be changed. Such insns might be not in
4829 init_insns because we don't update equiv data
4830 during insn transformations.
4832 As an example, let suppose that a pseudo got
4833 hard register and on the 1st pass was not
4834 changed to equivalent constant. We generate an
4835 additional insn setting up the pseudo because of
4836 secondary memory movement. Then the pseudo is
4837 spilled and we use the equiv constant. In this
4838 case we should remove the additional insn and
4839 this insn is not init_insns list. */
4840 && (! MEM_P (x) || MEM_READONLY_P (x)
4841 /* Check that this is actually an insn setting
4842 up the equivalence. */
4843 || in_list_p (curr_insn,
4844 ira_reg_equiv
4845 [REGNO (dest_reg)].init_insns)))
4846 || (((x = get_equiv (SET_SRC (set))) != SET_SRC (set))
4847 && in_list_p (curr_insn,
4848 ira_reg_equiv
4849 [REGNO (SET_SRC (set))].init_insns)))
4851 /* This is equiv init insn of pseudo which did not get a
4852 hard register -- remove the insn. */
4853 if (lra_dump_file != NULL)
4855 fprintf (lra_dump_file,
4856 " Removing equiv init insn %i (freq=%d)\n",
4857 INSN_UID (curr_insn),
4858 REG_FREQ_FROM_BB (BLOCK_FOR_INSN (curr_insn)));
4859 dump_insn_slim (lra_dump_file, curr_insn);
4861 if (contains_reg_p (x, true, false))
4862 lra_risky_transformations_p = true;
4863 lra_set_insn_deleted (curr_insn);
4864 continue;
4867 curr_id = lra_get_insn_recog_data (curr_insn);
4868 curr_static_id = curr_id->insn_static_data;
4869 init_curr_insn_input_reloads ();
4870 init_curr_operand_mode ();
4871 if (curr_insn_transform (false))
4872 changed_p = true;
4873 /* Check non-transformed insns too for equiv change as USE
4874 or CLOBBER don't need reloads but can contain pseudos
4875 being changed on their equivalences. */
4876 else if (bitmap_bit_p (equiv_insn_bitmap, INSN_UID (curr_insn))
4877 && loc_equivalence_change_p (&PATTERN (curr_insn)))
4879 lra_update_insn_regno_info (curr_insn);
4880 changed_p = true;
4885 /* If we used a new hard regno, changed_p should be true because the
4886 hard reg is assigned to a new pseudo. */
4887 if (flag_checking && !changed_p)
4889 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4890 if (lra_reg_info[i].nrefs != 0
4891 && (hard_regno = lra_get_regno_hard_regno (i)) >= 0)
4893 int j, nregs = hard_regno_nregs (hard_regno,
4894 PSEUDO_REGNO_MODE (i));
4896 for (j = 0; j < nregs; j++)
4897 lra_assert (df_regs_ever_live_p (hard_regno + j));
4900 return changed_p;
4903 static void initiate_invariants (void);
4904 static void finish_invariants (void);
4906 /* Initiate the LRA constraint pass. It is done once per
4907 function. */
4908 void
4909 lra_constraints_init (void)
4911 initiate_invariants ();
4914 /* Finalize the LRA constraint pass. It is done once per
4915 function. */
4916 void
4917 lra_constraints_finish (void)
4919 finish_invariants ();
4924 /* Structure describes invariants for ineheritance. */
4925 struct lra_invariant
4927 /* The order number of the invariant. */
4928 int num;
4929 /* The invariant RTX. */
4930 rtx invariant_rtx;
4931 /* The origin insn of the invariant. */
4932 rtx_insn *insn;
4935 typedef lra_invariant invariant_t;
4936 typedef invariant_t *invariant_ptr_t;
4937 typedef const invariant_t *const_invariant_ptr_t;
4939 /* Pointer to the inheritance invariants. */
4940 static vec<invariant_ptr_t> invariants;
4942 /* Allocation pool for the invariants. */
4943 static object_allocator<lra_invariant> *invariants_pool;
4945 /* Hash table for the invariants. */
4946 static htab_t invariant_table;
4948 /* Hash function for INVARIANT. */
4949 static hashval_t
4950 invariant_hash (const void *invariant)
4952 rtx inv = ((const_invariant_ptr_t) invariant)->invariant_rtx;
4953 return lra_rtx_hash (inv);
4956 /* Equal function for invariants INVARIANT1 and INVARIANT2. */
4957 static int
4958 invariant_eq_p (const void *invariant1, const void *invariant2)
4960 rtx inv1 = ((const_invariant_ptr_t) invariant1)->invariant_rtx;
4961 rtx inv2 = ((const_invariant_ptr_t) invariant2)->invariant_rtx;
4963 return rtx_equal_p (inv1, inv2);
4966 /* Insert INVARIANT_RTX into the table if it is not there yet. Return
4967 invariant which is in the table. */
4968 static invariant_ptr_t
4969 insert_invariant (rtx invariant_rtx)
4971 void **entry_ptr;
4972 invariant_t invariant;
4973 invariant_ptr_t invariant_ptr;
4975 invariant.invariant_rtx = invariant_rtx;
4976 entry_ptr = htab_find_slot (invariant_table, &invariant, INSERT);
4977 if (*entry_ptr == NULL)
4979 invariant_ptr = invariants_pool->allocate ();
4980 invariant_ptr->invariant_rtx = invariant_rtx;
4981 invariant_ptr->insn = NULL;
4982 invariants.safe_push (invariant_ptr);
4983 *entry_ptr = (void *) invariant_ptr;
4985 return (invariant_ptr_t) *entry_ptr;
4988 /* Initiate the invariant table. */
4989 static void
4990 initiate_invariants (void)
4992 invariants.create (100);
4993 invariants_pool
4994 = new object_allocator<lra_invariant> ("Inheritance invariants");
4995 invariant_table = htab_create (100, invariant_hash, invariant_eq_p, NULL);
4998 /* Finish the invariant table. */
4999 static void
5000 finish_invariants (void)
5002 htab_delete (invariant_table);
5003 delete invariants_pool;
5004 invariants.release ();
5007 /* Make the invariant table empty. */
5008 static void
5009 clear_invariants (void)
5011 htab_empty (invariant_table);
5012 invariants_pool->release ();
5013 invariants.truncate (0);
5018 /* This page contains code to do inheritance/split
5019 transformations. */
5021 /* Number of reloads passed so far in current EBB. */
5022 static int reloads_num;
5024 /* Number of calls passed so far in current EBB. */
5025 static int calls_num;
5027 /* Current reload pseudo check for validity of elements in
5028 USAGE_INSNS. */
5029 static int curr_usage_insns_check;
5031 /* Info about last usage of registers in EBB to do inheritance/split
5032 transformation. Inheritance transformation is done from a spilled
5033 pseudo and split transformations from a hard register or a pseudo
5034 assigned to a hard register. */
5035 struct usage_insns
5037 /* If the value is equal to CURR_USAGE_INSNS_CHECK, then the member
5038 value INSNS is valid. The insns is chain of optional debug insns
5039 and a finishing non-debug insn using the corresponding reg. The
5040 value is also used to mark the registers which are set up in the
5041 current insn. The negated insn uid is used for this. */
5042 int check;
5043 /* Value of global reloads_num at the last insn in INSNS. */
5044 int reloads_num;
5045 /* Value of global reloads_nums at the last insn in INSNS. */
5046 int calls_num;
5047 /* It can be true only for splitting. And it means that the restore
5048 insn should be put after insn given by the following member. */
5049 bool after_p;
5050 /* Next insns in the current EBB which use the original reg and the
5051 original reg value is not changed between the current insn and
5052 the next insns. In order words, e.g. for inheritance, if we need
5053 to use the original reg value again in the next insns we can try
5054 to use the value in a hard register from a reload insn of the
5055 current insn. */
5056 rtx insns;
5059 /* Map: regno -> corresponding pseudo usage insns. */
5060 static struct usage_insns *usage_insns;
5062 static void
5063 setup_next_usage_insn (int regno, rtx insn, int reloads_num, bool after_p)
5065 usage_insns[regno].check = curr_usage_insns_check;
5066 usage_insns[regno].insns = insn;
5067 usage_insns[regno].reloads_num = reloads_num;
5068 usage_insns[regno].calls_num = calls_num;
5069 usage_insns[regno].after_p = after_p;
5072 /* The function is used to form list REGNO usages which consists of
5073 optional debug insns finished by a non-debug insn using REGNO.
5074 RELOADS_NUM is current number of reload insns processed so far. */
5075 static void
5076 add_next_usage_insn (int regno, rtx_insn *insn, int reloads_num)
5078 rtx next_usage_insns;
5080 if (usage_insns[regno].check == curr_usage_insns_check
5081 && (next_usage_insns = usage_insns[regno].insns) != NULL_RTX
5082 && DEBUG_INSN_P (insn))
5084 /* Check that we did not add the debug insn yet. */
5085 if (next_usage_insns != insn
5086 && (GET_CODE (next_usage_insns) != INSN_LIST
5087 || XEXP (next_usage_insns, 0) != insn))
5088 usage_insns[regno].insns = gen_rtx_INSN_LIST (VOIDmode, insn,
5089 next_usage_insns);
5091 else if (NONDEBUG_INSN_P (insn))
5092 setup_next_usage_insn (regno, insn, reloads_num, false);
5093 else
5094 usage_insns[regno].check = 0;
5097 /* Return first non-debug insn in list USAGE_INSNS. */
5098 static rtx_insn *
5099 skip_usage_debug_insns (rtx usage_insns)
5101 rtx insn;
5103 /* Skip debug insns. */
5104 for (insn = usage_insns;
5105 insn != NULL_RTX && GET_CODE (insn) == INSN_LIST;
5106 insn = XEXP (insn, 1))
5108 return safe_as_a <rtx_insn *> (insn);
5111 /* Return true if we need secondary memory moves for insn in
5112 USAGE_INSNS after inserting inherited pseudo of class INHER_CL
5113 into the insn. */
5114 static bool
5115 check_secondary_memory_needed_p (enum reg_class inher_cl ATTRIBUTE_UNUSED,
5116 rtx usage_insns ATTRIBUTE_UNUSED)
5118 rtx_insn *insn;
5119 rtx set, dest;
5120 enum reg_class cl;
5122 if (inher_cl == ALL_REGS
5123 || (insn = skip_usage_debug_insns (usage_insns)) == NULL_RTX)
5124 return false;
5125 lra_assert (INSN_P (insn));
5126 if ((set = single_set (insn)) == NULL_RTX || ! REG_P (SET_DEST (set)))
5127 return false;
5128 dest = SET_DEST (set);
5129 if (! REG_P (dest))
5130 return false;
5131 lra_assert (inher_cl != NO_REGS);
5132 cl = get_reg_class (REGNO (dest));
5133 return (cl != NO_REGS && cl != ALL_REGS
5134 && targetm.secondary_memory_needed (GET_MODE (dest), inher_cl, cl));
5137 /* Registers involved in inheritance/split in the current EBB
5138 (inheritance/split pseudos and original registers). */
5139 static bitmap_head check_only_regs;
5141 /* Reload pseudos can not be involded in invariant inheritance in the
5142 current EBB. */
5143 static bitmap_head invalid_invariant_regs;
5145 /* Do inheritance transformations for insn INSN, which defines (if
5146 DEF_P) or uses ORIGINAL_REGNO. NEXT_USAGE_INSNS specifies which
5147 instruction in the EBB next uses ORIGINAL_REGNO; it has the same
5148 form as the "insns" field of usage_insns. Return true if we
5149 succeed in such transformation.
5151 The transformations look like:
5153 p <- ... i <- ...
5154 ... p <- i (new insn)
5155 ... =>
5156 <- ... p ... <- ... i ...
5158 ... i <- p (new insn)
5159 <- ... p ... <- ... i ...
5160 ... =>
5161 <- ... p ... <- ... i ...
5162 where p is a spilled original pseudo and i is a new inheritance pseudo.
5165 The inheritance pseudo has the smallest class of two classes CL and
5166 class of ORIGINAL REGNO. */
5167 static bool
5168 inherit_reload_reg (bool def_p, int original_regno,
5169 enum reg_class cl, rtx_insn *insn, rtx next_usage_insns)
5171 if (optimize_function_for_size_p (cfun))
5172 return false;
5174 enum reg_class rclass = lra_get_allocno_class (original_regno);
5175 rtx original_reg = regno_reg_rtx[original_regno];
5176 rtx new_reg, usage_insn;
5177 rtx_insn *new_insns;
5179 lra_assert (! usage_insns[original_regno].after_p);
5180 if (lra_dump_file != NULL)
5181 fprintf (lra_dump_file,
5182 " <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<\n");
5183 if (! ira_reg_classes_intersect_p[cl][rclass])
5185 if (lra_dump_file != NULL)
5187 fprintf (lra_dump_file,
5188 " Rejecting inheritance for %d "
5189 "because of disjoint classes %s and %s\n",
5190 original_regno, reg_class_names[cl],
5191 reg_class_names[rclass]);
5192 fprintf (lra_dump_file,
5193 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5195 return false;
5197 if ((ira_class_subset_p[cl][rclass] && cl != rclass)
5198 /* We don't use a subset of two classes because it can be
5199 NO_REGS. This transformation is still profitable in most
5200 cases even if the classes are not intersected as register
5201 move is probably cheaper than a memory load. */
5202 || ira_class_hard_regs_num[cl] < ira_class_hard_regs_num[rclass])
5204 if (lra_dump_file != NULL)
5205 fprintf (lra_dump_file, " Use smallest class of %s and %s\n",
5206 reg_class_names[cl], reg_class_names[rclass]);
5208 rclass = cl;
5210 if (check_secondary_memory_needed_p (rclass, next_usage_insns))
5212 /* Reject inheritance resulting in secondary memory moves.
5213 Otherwise, there is a danger in LRA cycling. Also such
5214 transformation will be unprofitable. */
5215 if (lra_dump_file != NULL)
5217 rtx_insn *insn = skip_usage_debug_insns (next_usage_insns);
5218 rtx set = single_set (insn);
5220 lra_assert (set != NULL_RTX);
5222 rtx dest = SET_DEST (set);
5224 lra_assert (REG_P (dest));
5225 fprintf (lra_dump_file,
5226 " Rejecting inheritance for insn %d(%s)<-%d(%s) "
5227 "as secondary mem is needed\n",
5228 REGNO (dest), reg_class_names[get_reg_class (REGNO (dest))],
5229 original_regno, reg_class_names[rclass]);
5230 fprintf (lra_dump_file,
5231 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5233 return false;
5235 new_reg = lra_create_new_reg (GET_MODE (original_reg), original_reg,
5236 rclass, "inheritance");
5237 start_sequence ();
5238 if (def_p)
5239 lra_emit_move (original_reg, new_reg);
5240 else
5241 lra_emit_move (new_reg, original_reg);
5242 new_insns = get_insns ();
5243 end_sequence ();
5244 if (NEXT_INSN (new_insns) != NULL_RTX)
5246 if (lra_dump_file != NULL)
5248 fprintf (lra_dump_file,
5249 " Rejecting inheritance %d->%d "
5250 "as it results in 2 or more insns:\n",
5251 original_regno, REGNO (new_reg));
5252 dump_rtl_slim (lra_dump_file, new_insns, NULL, -1, 0);
5253 fprintf (lra_dump_file,
5254 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5256 return false;
5258 lra_substitute_pseudo_within_insn (insn, original_regno, new_reg, false);
5259 lra_update_insn_regno_info (insn);
5260 if (! def_p)
5261 /* We now have a new usage insn for original regno. */
5262 setup_next_usage_insn (original_regno, new_insns, reloads_num, false);
5263 if (lra_dump_file != NULL)
5264 fprintf (lra_dump_file, " Original reg change %d->%d (bb%d):\n",
5265 original_regno, REGNO (new_reg), BLOCK_FOR_INSN (insn)->index);
5266 lra_reg_info[REGNO (new_reg)].restore_rtx = regno_reg_rtx[original_regno];
5267 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
5268 bitmap_set_bit (&check_only_regs, original_regno);
5269 bitmap_set_bit (&lra_inheritance_pseudos, REGNO (new_reg));
5270 if (def_p)
5271 lra_process_new_insns (insn, NULL, new_insns,
5272 "Add original<-inheritance");
5273 else
5274 lra_process_new_insns (insn, new_insns, NULL,
5275 "Add inheritance<-original");
5276 while (next_usage_insns != NULL_RTX)
5278 if (GET_CODE (next_usage_insns) != INSN_LIST)
5280 usage_insn = next_usage_insns;
5281 lra_assert (NONDEBUG_INSN_P (usage_insn));
5282 next_usage_insns = NULL;
5284 else
5286 usage_insn = XEXP (next_usage_insns, 0);
5287 lra_assert (DEBUG_INSN_P (usage_insn));
5288 next_usage_insns = XEXP (next_usage_insns, 1);
5290 lra_substitute_pseudo (&usage_insn, original_regno, new_reg, false,
5291 DEBUG_INSN_P (usage_insn));
5292 lra_update_insn_regno_info (as_a <rtx_insn *> (usage_insn));
5293 if (lra_dump_file != NULL)
5295 basic_block bb = BLOCK_FOR_INSN (usage_insn);
5296 fprintf (lra_dump_file,
5297 " Inheritance reuse change %d->%d (bb%d):\n",
5298 original_regno, REGNO (new_reg),
5299 bb ? bb->index : -1);
5300 dump_insn_slim (lra_dump_file, as_a <rtx_insn *> (usage_insn));
5303 if (lra_dump_file != NULL)
5304 fprintf (lra_dump_file,
5305 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5306 return true;
5309 /* Return true if we need a caller save/restore for pseudo REGNO which
5310 was assigned to a hard register. */
5311 static inline bool
5312 need_for_call_save_p (int regno)
5314 lra_assert (regno >= FIRST_PSEUDO_REGISTER && reg_renumber[regno] >= 0);
5315 return (usage_insns[regno].calls_num < calls_num
5316 && (overlaps_hard_reg_set_p
5317 ((flag_ipa_ra &&
5318 ! hard_reg_set_empty_p (lra_reg_info[regno].actual_call_used_reg_set))
5319 ? lra_reg_info[regno].actual_call_used_reg_set
5320 : call_used_reg_set,
5321 PSEUDO_REGNO_MODE (regno), reg_renumber[regno])
5322 || (targetm.hard_regno_call_part_clobbered
5323 (reg_renumber[regno], PSEUDO_REGNO_MODE (regno)))));
5326 /* Global registers occurring in the current EBB. */
5327 static bitmap_head ebb_global_regs;
5329 /* Return true if we need a split for hard register REGNO or pseudo
5330 REGNO which was assigned to a hard register.
5331 POTENTIAL_RELOAD_HARD_REGS contains hard registers which might be
5332 used for reloads since the EBB end. It is an approximation of the
5333 used hard registers in the split range. The exact value would
5334 require expensive calculations. If we were aggressive with
5335 splitting because of the approximation, the split pseudo will save
5336 the same hard register assignment and will be removed in the undo
5337 pass. We still need the approximation because too aggressive
5338 splitting would result in too inaccurate cost calculation in the
5339 assignment pass because of too many generated moves which will be
5340 probably removed in the undo pass. */
5341 static inline bool
5342 need_for_split_p (HARD_REG_SET potential_reload_hard_regs, int regno)
5344 int hard_regno = regno < FIRST_PSEUDO_REGISTER ? regno : reg_renumber[regno];
5346 lra_assert (hard_regno >= 0);
5347 return ((TEST_HARD_REG_BIT (potential_reload_hard_regs, hard_regno)
5348 /* Don't split eliminable hard registers, otherwise we can
5349 split hard registers like hard frame pointer, which
5350 lives on BB start/end according to DF-infrastructure,
5351 when there is a pseudo assigned to the register and
5352 living in the same BB. */
5353 && (regno >= FIRST_PSEUDO_REGISTER
5354 || ! TEST_HARD_REG_BIT (eliminable_regset, hard_regno))
5355 && ! TEST_HARD_REG_BIT (lra_no_alloc_regs, hard_regno)
5356 /* Don't split call clobbered hard regs living through
5357 calls, otherwise we might have a check problem in the
5358 assign sub-pass as in the most cases (exception is a
5359 situation when lra_risky_transformations_p value is
5360 true) the assign pass assumes that all pseudos living
5361 through calls are assigned to call saved hard regs. */
5362 && (regno >= FIRST_PSEUDO_REGISTER
5363 || ! TEST_HARD_REG_BIT (call_used_reg_set, regno)
5364 || usage_insns[regno].calls_num == calls_num)
5365 /* We need at least 2 reloads to make pseudo splitting
5366 profitable. We should provide hard regno splitting in
5367 any case to solve 1st insn scheduling problem when
5368 moving hard register definition up might result in
5369 impossibility to find hard register for reload pseudo of
5370 small register class. */
5371 && (usage_insns[regno].reloads_num
5372 + (regno < FIRST_PSEUDO_REGISTER ? 0 : 3) < reloads_num)
5373 && (regno < FIRST_PSEUDO_REGISTER
5374 /* For short living pseudos, spilling + inheritance can
5375 be considered a substitution for splitting.
5376 Therefore we do not splitting for local pseudos. It
5377 decreases also aggressiveness of splitting. The
5378 minimal number of references is chosen taking into
5379 account that for 2 references splitting has no sense
5380 as we can just spill the pseudo. */
5381 || (regno >= FIRST_PSEUDO_REGISTER
5382 && lra_reg_info[regno].nrefs > 3
5383 && bitmap_bit_p (&ebb_global_regs, regno))))
5384 || (regno >= FIRST_PSEUDO_REGISTER && need_for_call_save_p (regno)));
5387 /* Return class for the split pseudo created from original pseudo with
5388 ALLOCNO_CLASS and MODE which got a hard register HARD_REGNO. We
5389 choose subclass of ALLOCNO_CLASS which contains HARD_REGNO and
5390 results in no secondary memory movements. */
5391 static enum reg_class
5392 choose_split_class (enum reg_class allocno_class,
5393 int hard_regno ATTRIBUTE_UNUSED,
5394 machine_mode mode ATTRIBUTE_UNUSED)
5396 int i;
5397 enum reg_class cl, best_cl = NO_REGS;
5398 enum reg_class hard_reg_class ATTRIBUTE_UNUSED
5399 = REGNO_REG_CLASS (hard_regno);
5401 if (! targetm.secondary_memory_needed (mode, allocno_class, allocno_class)
5402 && TEST_HARD_REG_BIT (reg_class_contents[allocno_class], hard_regno))
5403 return allocno_class;
5404 for (i = 0;
5405 (cl = reg_class_subclasses[allocno_class][i]) != LIM_REG_CLASSES;
5406 i++)
5407 if (! targetm.secondary_memory_needed (mode, cl, hard_reg_class)
5408 && ! targetm.secondary_memory_needed (mode, hard_reg_class, cl)
5409 && TEST_HARD_REG_BIT (reg_class_contents[cl], hard_regno)
5410 && (best_cl == NO_REGS
5411 || ira_class_hard_regs_num[best_cl] < ira_class_hard_regs_num[cl]))
5412 best_cl = cl;
5413 return best_cl;
5416 /* Copy any equivalence information from ORIGINAL_REGNO to NEW_REGNO.
5417 It only makes sense to call this function if NEW_REGNO is always
5418 equal to ORIGINAL_REGNO. */
5420 static void
5421 lra_copy_reg_equiv (unsigned int new_regno, unsigned int original_regno)
5423 if (!ira_reg_equiv[original_regno].defined_p)
5424 return;
5426 ira_expand_reg_equiv ();
5427 ira_reg_equiv[new_regno].defined_p = true;
5428 if (ira_reg_equiv[original_regno].memory)
5429 ira_reg_equiv[new_regno].memory
5430 = copy_rtx (ira_reg_equiv[original_regno].memory);
5431 if (ira_reg_equiv[original_regno].constant)
5432 ira_reg_equiv[new_regno].constant
5433 = copy_rtx (ira_reg_equiv[original_regno].constant);
5434 if (ira_reg_equiv[original_regno].invariant)
5435 ira_reg_equiv[new_regno].invariant
5436 = copy_rtx (ira_reg_equiv[original_regno].invariant);
5439 /* Do split transformations for insn INSN, which defines or uses
5440 ORIGINAL_REGNO. NEXT_USAGE_INSNS specifies which instruction in
5441 the EBB next uses ORIGINAL_REGNO; it has the same form as the
5442 "insns" field of usage_insns.
5444 The transformations look like:
5446 p <- ... p <- ...
5447 ... s <- p (new insn -- save)
5448 ... =>
5449 ... p <- s (new insn -- restore)
5450 <- ... p ... <- ... p ...
5452 <- ... p ... <- ... p ...
5453 ... s <- p (new insn -- save)
5454 ... =>
5455 ... p <- s (new insn -- restore)
5456 <- ... p ... <- ... p ...
5458 where p is an original pseudo got a hard register or a hard
5459 register and s is a new split pseudo. The save is put before INSN
5460 if BEFORE_P is true. Return true if we succeed in such
5461 transformation. */
5462 static bool
5463 split_reg (bool before_p, int original_regno, rtx_insn *insn,
5464 rtx next_usage_insns)
5466 enum reg_class rclass;
5467 rtx original_reg;
5468 int hard_regno, nregs;
5469 rtx new_reg, usage_insn;
5470 rtx_insn *restore, *save;
5471 bool after_p;
5472 bool call_save_p;
5473 machine_mode mode;
5475 if (original_regno < FIRST_PSEUDO_REGISTER)
5477 rclass = ira_allocno_class_translate[REGNO_REG_CLASS (original_regno)];
5478 hard_regno = original_regno;
5479 call_save_p = false;
5480 nregs = 1;
5481 mode = lra_reg_info[hard_regno].biggest_mode;
5482 machine_mode reg_rtx_mode = GET_MODE (regno_reg_rtx[hard_regno]);
5483 /* A reg can have a biggest_mode of VOIDmode if it was only ever seen
5484 as part of a multi-word register. In that case, or if the biggest
5485 mode was larger than a register, just use the reg_rtx. Otherwise,
5486 limit the size to that of the biggest access in the function. */
5487 if (mode == VOIDmode
5488 || paradoxical_subreg_p (mode, reg_rtx_mode))
5490 original_reg = regno_reg_rtx[hard_regno];
5491 mode = reg_rtx_mode;
5493 else
5494 original_reg = gen_rtx_REG (mode, hard_regno);
5496 else
5498 mode = PSEUDO_REGNO_MODE (original_regno);
5499 hard_regno = reg_renumber[original_regno];
5500 nregs = hard_regno_nregs (hard_regno, mode);
5501 rclass = lra_get_allocno_class (original_regno);
5502 original_reg = regno_reg_rtx[original_regno];
5503 call_save_p = need_for_call_save_p (original_regno);
5505 lra_assert (hard_regno >= 0);
5506 if (lra_dump_file != NULL)
5507 fprintf (lra_dump_file,
5508 " ((((((((((((((((((((((((((((((((((((((((((((((((\n");
5510 if (call_save_p)
5512 mode = HARD_REGNO_CALLER_SAVE_MODE (hard_regno,
5513 hard_regno_nregs (hard_regno, mode),
5514 mode);
5515 new_reg = lra_create_new_reg (mode, NULL_RTX, NO_REGS, "save");
5517 else
5519 rclass = choose_split_class (rclass, hard_regno, mode);
5520 if (rclass == NO_REGS)
5522 if (lra_dump_file != NULL)
5524 fprintf (lra_dump_file,
5525 " Rejecting split of %d(%s): "
5526 "no good reg class for %d(%s)\n",
5527 original_regno,
5528 reg_class_names[lra_get_allocno_class (original_regno)],
5529 hard_regno,
5530 reg_class_names[REGNO_REG_CLASS (hard_regno)]);
5531 fprintf
5532 (lra_dump_file,
5533 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5535 return false;
5537 /* Split_if_necessary can split hard registers used as part of a
5538 multi-register mode but splits each register individually. The
5539 mode used for each independent register may not be supported
5540 so reject the split. Splitting the wider mode should theoretically
5541 be possible but is not implemented. */
5542 if (!targetm.hard_regno_mode_ok (hard_regno, mode))
5544 if (lra_dump_file != NULL)
5546 fprintf (lra_dump_file,
5547 " Rejecting split of %d(%s): unsuitable mode %s\n",
5548 original_regno,
5549 reg_class_names[lra_get_allocno_class (original_regno)],
5550 GET_MODE_NAME (mode));
5551 fprintf
5552 (lra_dump_file,
5553 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5555 return false;
5557 new_reg = lra_create_new_reg (mode, original_reg, rclass, "split");
5558 reg_renumber[REGNO (new_reg)] = hard_regno;
5560 int new_regno = REGNO (new_reg);
5561 save = emit_spill_move (true, new_reg, original_reg);
5562 if (NEXT_INSN (save) != NULL_RTX && !call_save_p)
5564 if (lra_dump_file != NULL)
5566 fprintf
5567 (lra_dump_file,
5568 " Rejecting split %d->%d resulting in > 2 save insns:\n",
5569 original_regno, new_regno);
5570 dump_rtl_slim (lra_dump_file, save, NULL, -1, 0);
5571 fprintf (lra_dump_file,
5572 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5574 return false;
5576 restore = emit_spill_move (false, new_reg, original_reg);
5577 if (NEXT_INSN (restore) != NULL_RTX && !call_save_p)
5579 if (lra_dump_file != NULL)
5581 fprintf (lra_dump_file,
5582 " Rejecting split %d->%d "
5583 "resulting in > 2 restore insns:\n",
5584 original_regno, new_regno);
5585 dump_rtl_slim (lra_dump_file, restore, NULL, -1, 0);
5586 fprintf (lra_dump_file,
5587 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5589 return false;
5591 /* Transfer equivalence information to the spill register, so that
5592 if we fail to allocate the spill register, we have the option of
5593 rematerializing the original value instead of spilling to the stack. */
5594 if (!HARD_REGISTER_NUM_P (original_regno)
5595 && mode == PSEUDO_REGNO_MODE (original_regno))
5596 lra_copy_reg_equiv (new_regno, original_regno);
5597 after_p = usage_insns[original_regno].after_p;
5598 lra_reg_info[new_regno].restore_rtx = regno_reg_rtx[original_regno];
5599 bitmap_set_bit (&check_only_regs, new_regno);
5600 bitmap_set_bit (&check_only_regs, original_regno);
5601 bitmap_set_bit (&lra_split_regs, new_regno);
5602 for (;;)
5604 if (GET_CODE (next_usage_insns) != INSN_LIST)
5606 usage_insn = next_usage_insns;
5607 break;
5609 usage_insn = XEXP (next_usage_insns, 0);
5610 lra_assert (DEBUG_INSN_P (usage_insn));
5611 next_usage_insns = XEXP (next_usage_insns, 1);
5612 lra_substitute_pseudo (&usage_insn, original_regno, new_reg, false,
5613 true);
5614 lra_update_insn_regno_info (as_a <rtx_insn *> (usage_insn));
5615 if (lra_dump_file != NULL)
5617 fprintf (lra_dump_file, " Split reuse change %d->%d:\n",
5618 original_regno, new_regno);
5619 dump_insn_slim (lra_dump_file, as_a <rtx_insn *> (usage_insn));
5622 lra_assert (NOTE_P (usage_insn) || NONDEBUG_INSN_P (usage_insn));
5623 lra_assert (usage_insn != insn || (after_p && before_p));
5624 lra_process_new_insns (as_a <rtx_insn *> (usage_insn),
5625 after_p ? NULL : restore,
5626 after_p ? restore : NULL,
5627 call_save_p
5628 ? "Add reg<-save" : "Add reg<-split");
5629 lra_process_new_insns (insn, before_p ? save : NULL,
5630 before_p ? NULL : save,
5631 call_save_p
5632 ? "Add save<-reg" : "Add split<-reg");
5633 if (nregs > 1)
5634 /* If we are trying to split multi-register. We should check
5635 conflicts on the next assignment sub-pass. IRA can allocate on
5636 sub-register levels, LRA do this on pseudos level right now and
5637 this discrepancy may create allocation conflicts after
5638 splitting. */
5639 lra_risky_transformations_p = true;
5640 if (lra_dump_file != NULL)
5641 fprintf (lra_dump_file,
5642 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5643 return true;
5646 /* Recognize that we need a split transformation for insn INSN, which
5647 defines or uses REGNO in its insn biggest MODE (we use it only if
5648 REGNO is a hard register). POTENTIAL_RELOAD_HARD_REGS contains
5649 hard registers which might be used for reloads since the EBB end.
5650 Put the save before INSN if BEFORE_P is true. MAX_UID is maximla
5651 uid before starting INSN processing. Return true if we succeed in
5652 such transformation. */
5653 static bool
5654 split_if_necessary (int regno, machine_mode mode,
5655 HARD_REG_SET potential_reload_hard_regs,
5656 bool before_p, rtx_insn *insn, int max_uid)
5658 bool res = false;
5659 int i, nregs = 1;
5660 rtx next_usage_insns;
5662 if (regno < FIRST_PSEUDO_REGISTER)
5663 nregs = hard_regno_nregs (regno, mode);
5664 for (i = 0; i < nregs; i++)
5665 if (usage_insns[regno + i].check == curr_usage_insns_check
5666 && (next_usage_insns = usage_insns[regno + i].insns) != NULL_RTX
5667 /* To avoid processing the register twice or more. */
5668 && ((GET_CODE (next_usage_insns) != INSN_LIST
5669 && INSN_UID (next_usage_insns) < max_uid)
5670 || (GET_CODE (next_usage_insns) == INSN_LIST
5671 && (INSN_UID (XEXP (next_usage_insns, 0)) < max_uid)))
5672 && need_for_split_p (potential_reload_hard_regs, regno + i)
5673 && split_reg (before_p, regno + i, insn, next_usage_insns))
5674 res = true;
5675 return res;
5678 /* Return TRUE if rtx X is considered as an invariant for
5679 inheritance. */
5680 static bool
5681 invariant_p (const_rtx x)
5683 machine_mode mode;
5684 const char *fmt;
5685 enum rtx_code code;
5686 int i, j;
5688 code = GET_CODE (x);
5689 mode = GET_MODE (x);
5690 if (code == SUBREG)
5692 x = SUBREG_REG (x);
5693 code = GET_CODE (x);
5694 mode = wider_subreg_mode (mode, GET_MODE (x));
5697 if (MEM_P (x))
5698 return false;
5700 if (REG_P (x))
5702 int i, nregs, regno = REGNO (x);
5704 if (regno >= FIRST_PSEUDO_REGISTER || regno == STACK_POINTER_REGNUM
5705 || TEST_HARD_REG_BIT (eliminable_regset, regno)
5706 || GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
5707 return false;
5708 nregs = hard_regno_nregs (regno, mode);
5709 for (i = 0; i < nregs; i++)
5710 if (! fixed_regs[regno + i]
5711 /* A hard register may be clobbered in the current insn
5712 but we can ignore this case because if the hard
5713 register is used it should be set somewhere after the
5714 clobber. */
5715 || bitmap_bit_p (&invalid_invariant_regs, regno + i))
5716 return false;
5718 fmt = GET_RTX_FORMAT (code);
5719 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
5721 if (fmt[i] == 'e')
5723 if (! invariant_p (XEXP (x, i)))
5724 return false;
5726 else if (fmt[i] == 'E')
5728 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
5729 if (! invariant_p (XVECEXP (x, i, j)))
5730 return false;
5733 return true;
5736 /* We have 'dest_reg <- invariant'. Let us try to make an invariant
5737 inheritance transformation (using dest_reg instead invariant in a
5738 subsequent insn). */
5739 static bool
5740 process_invariant_for_inheritance (rtx dst_reg, rtx invariant_rtx)
5742 invariant_ptr_t invariant_ptr;
5743 rtx_insn *insn, *new_insns;
5744 rtx insn_set, insn_reg, new_reg;
5745 int insn_regno;
5746 bool succ_p = false;
5747 int dst_regno = REGNO (dst_reg);
5748 machine_mode dst_mode = GET_MODE (dst_reg);
5749 enum reg_class cl = lra_get_allocno_class (dst_regno), insn_reg_cl;
5751 invariant_ptr = insert_invariant (invariant_rtx);
5752 if ((insn = invariant_ptr->insn) != NULL_RTX)
5754 /* We have a subsequent insn using the invariant. */
5755 insn_set = single_set (insn);
5756 lra_assert (insn_set != NULL);
5757 insn_reg = SET_DEST (insn_set);
5758 lra_assert (REG_P (insn_reg));
5759 insn_regno = REGNO (insn_reg);
5760 insn_reg_cl = lra_get_allocno_class (insn_regno);
5762 if (dst_mode == GET_MODE (insn_reg)
5763 /* We should consider only result move reg insns which are
5764 cheap. */
5765 && targetm.register_move_cost (dst_mode, cl, insn_reg_cl) == 2
5766 && targetm.register_move_cost (dst_mode, cl, cl) == 2)
5768 if (lra_dump_file != NULL)
5769 fprintf (lra_dump_file,
5770 " [[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[\n");
5771 new_reg = lra_create_new_reg (dst_mode, dst_reg,
5772 cl, "invariant inheritance");
5773 bitmap_set_bit (&lra_inheritance_pseudos, REGNO (new_reg));
5774 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
5775 lra_reg_info[REGNO (new_reg)].restore_rtx = PATTERN (insn);
5776 start_sequence ();
5777 lra_emit_move (new_reg, dst_reg);
5778 new_insns = get_insns ();
5779 end_sequence ();
5780 lra_process_new_insns (curr_insn, NULL, new_insns,
5781 "Add invariant inheritance<-original");
5782 start_sequence ();
5783 lra_emit_move (SET_DEST (insn_set), new_reg);
5784 new_insns = get_insns ();
5785 end_sequence ();
5786 lra_process_new_insns (insn, NULL, new_insns,
5787 "Changing reload<-inheritance");
5788 lra_set_insn_deleted (insn);
5789 succ_p = true;
5790 if (lra_dump_file != NULL)
5792 fprintf (lra_dump_file,
5793 " Invariant inheritance reuse change %d (bb%d):\n",
5794 REGNO (new_reg), BLOCK_FOR_INSN (insn)->index);
5795 dump_insn_slim (lra_dump_file, insn);
5796 fprintf (lra_dump_file,
5797 " ]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]\n");
5801 invariant_ptr->insn = curr_insn;
5802 return succ_p;
5805 /* Check only registers living at the current program point in the
5806 current EBB. */
5807 static bitmap_head live_regs;
5809 /* Update live info in EBB given by its HEAD and TAIL insns after
5810 inheritance/split transformation. The function removes dead moves
5811 too. */
5812 static void
5813 update_ebb_live_info (rtx_insn *head, rtx_insn *tail)
5815 unsigned int j;
5816 int i, regno;
5817 bool live_p;
5818 rtx_insn *prev_insn;
5819 rtx set;
5820 bool remove_p;
5821 basic_block last_bb, prev_bb, curr_bb;
5822 bitmap_iterator bi;
5823 struct lra_insn_reg *reg;
5824 edge e;
5825 edge_iterator ei;
5827 last_bb = BLOCK_FOR_INSN (tail);
5828 prev_bb = NULL;
5829 for (curr_insn = tail;
5830 curr_insn != PREV_INSN (head);
5831 curr_insn = prev_insn)
5833 prev_insn = PREV_INSN (curr_insn);
5834 /* We need to process empty blocks too. They contain
5835 NOTE_INSN_BASIC_BLOCK referring for the basic block. */
5836 if (NOTE_P (curr_insn) && NOTE_KIND (curr_insn) != NOTE_INSN_BASIC_BLOCK)
5837 continue;
5838 curr_bb = BLOCK_FOR_INSN (curr_insn);
5839 if (curr_bb != prev_bb)
5841 if (prev_bb != NULL)
5843 /* Update df_get_live_in (prev_bb): */
5844 EXECUTE_IF_SET_IN_BITMAP (&check_only_regs, 0, j, bi)
5845 if (bitmap_bit_p (&live_regs, j))
5846 bitmap_set_bit (df_get_live_in (prev_bb), j);
5847 else
5848 bitmap_clear_bit (df_get_live_in (prev_bb), j);
5850 if (curr_bb != last_bb)
5852 /* Update df_get_live_out (curr_bb): */
5853 EXECUTE_IF_SET_IN_BITMAP (&check_only_regs, 0, j, bi)
5855 live_p = bitmap_bit_p (&live_regs, j);
5856 if (! live_p)
5857 FOR_EACH_EDGE (e, ei, curr_bb->succs)
5858 if (bitmap_bit_p (df_get_live_in (e->dest), j))
5860 live_p = true;
5861 break;
5863 if (live_p)
5864 bitmap_set_bit (df_get_live_out (curr_bb), j);
5865 else
5866 bitmap_clear_bit (df_get_live_out (curr_bb), j);
5869 prev_bb = curr_bb;
5870 bitmap_and (&live_regs, &check_only_regs, df_get_live_out (curr_bb));
5872 if (! NONDEBUG_INSN_P (curr_insn))
5873 continue;
5874 curr_id = lra_get_insn_recog_data (curr_insn);
5875 curr_static_id = curr_id->insn_static_data;
5876 remove_p = false;
5877 if ((set = single_set (curr_insn)) != NULL_RTX
5878 && REG_P (SET_DEST (set))
5879 && (regno = REGNO (SET_DEST (set))) >= FIRST_PSEUDO_REGISTER
5880 && SET_DEST (set) != pic_offset_table_rtx
5881 && bitmap_bit_p (&check_only_regs, regno)
5882 && ! bitmap_bit_p (&live_regs, regno))
5883 remove_p = true;
5884 /* See which defined values die here. */
5885 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5886 if (reg->type == OP_OUT && ! reg->subreg_p)
5887 bitmap_clear_bit (&live_regs, reg->regno);
5888 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
5889 if (reg->type == OP_OUT && ! reg->subreg_p)
5890 bitmap_clear_bit (&live_regs, reg->regno);
5891 if (curr_id->arg_hard_regs != NULL)
5892 /* Make clobbered argument hard registers die. */
5893 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
5894 if (regno >= FIRST_PSEUDO_REGISTER)
5895 bitmap_clear_bit (&live_regs, regno - FIRST_PSEUDO_REGISTER);
5896 /* Mark each used value as live. */
5897 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5898 if (reg->type != OP_OUT
5899 && bitmap_bit_p (&check_only_regs, reg->regno))
5900 bitmap_set_bit (&live_regs, reg->regno);
5901 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
5902 if (reg->type != OP_OUT
5903 && bitmap_bit_p (&check_only_regs, reg->regno))
5904 bitmap_set_bit (&live_regs, reg->regno);
5905 if (curr_id->arg_hard_regs != NULL)
5906 /* Make used argument hard registers live. */
5907 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
5908 if (regno < FIRST_PSEUDO_REGISTER
5909 && bitmap_bit_p (&check_only_regs, regno))
5910 bitmap_set_bit (&live_regs, regno);
5911 /* It is quite important to remove dead move insns because it
5912 means removing dead store. We don't need to process them for
5913 constraints. */
5914 if (remove_p)
5916 if (lra_dump_file != NULL)
5918 fprintf (lra_dump_file, " Removing dead insn:\n ");
5919 dump_insn_slim (lra_dump_file, curr_insn);
5921 lra_set_insn_deleted (curr_insn);
5926 /* The structure describes info to do an inheritance for the current
5927 insn. We need to collect such info first before doing the
5928 transformations because the transformations change the insn
5929 internal representation. */
5930 struct to_inherit
5932 /* Original regno. */
5933 int regno;
5934 /* Subsequent insns which can inherit original reg value. */
5935 rtx insns;
5938 /* Array containing all info for doing inheritance from the current
5939 insn. */
5940 static struct to_inherit to_inherit[LRA_MAX_INSN_RELOADS];
5942 /* Number elements in the previous array. */
5943 static int to_inherit_num;
5945 /* Add inheritance info REGNO and INSNS. Their meaning is described in
5946 structure to_inherit. */
5947 static void
5948 add_to_inherit (int regno, rtx insns)
5950 int i;
5952 for (i = 0; i < to_inherit_num; i++)
5953 if (to_inherit[i].regno == regno)
5954 return;
5955 lra_assert (to_inherit_num < LRA_MAX_INSN_RELOADS);
5956 to_inherit[to_inherit_num].regno = regno;
5957 to_inherit[to_inherit_num++].insns = insns;
5960 /* Return the last non-debug insn in basic block BB, or the block begin
5961 note if none. */
5962 static rtx_insn *
5963 get_last_insertion_point (basic_block bb)
5965 rtx_insn *insn;
5967 FOR_BB_INSNS_REVERSE (bb, insn)
5968 if (NONDEBUG_INSN_P (insn) || NOTE_INSN_BASIC_BLOCK_P (insn))
5969 return insn;
5970 gcc_unreachable ();
5973 /* Set up RES by registers living on edges FROM except the edge (FROM,
5974 TO) or by registers set up in a jump insn in BB FROM. */
5975 static void
5976 get_live_on_other_edges (basic_block from, basic_block to, bitmap res)
5978 rtx_insn *last;
5979 struct lra_insn_reg *reg;
5980 edge e;
5981 edge_iterator ei;
5983 lra_assert (to != NULL);
5984 bitmap_clear (res);
5985 FOR_EACH_EDGE (e, ei, from->succs)
5986 if (e->dest != to)
5987 bitmap_ior_into (res, df_get_live_in (e->dest));
5988 last = get_last_insertion_point (from);
5989 if (! JUMP_P (last))
5990 return;
5991 curr_id = lra_get_insn_recog_data (last);
5992 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5993 if (reg->type != OP_IN)
5994 bitmap_set_bit (res, reg->regno);
5997 /* Used as a temporary results of some bitmap calculations. */
5998 static bitmap_head temp_bitmap;
6000 /* We split for reloads of small class of hard regs. The following
6001 defines how many hard regs the class should have to be qualified as
6002 small. The code is mostly oriented to x86/x86-64 architecture
6003 where some insns need to use only specific register or pair of
6004 registers and these register can live in RTL explicitly, e.g. for
6005 parameter passing. */
6006 static const int max_small_class_regs_num = 2;
6008 /* Do inheritance/split transformations in EBB starting with HEAD and
6009 finishing on TAIL. We process EBB insns in the reverse order.
6010 Return true if we did any inheritance/split transformation in the
6011 EBB.
6013 We should avoid excessive splitting which results in worse code
6014 because of inaccurate cost calculations for spilling new split
6015 pseudos in such case. To achieve this we do splitting only if
6016 register pressure is high in given basic block and there are reload
6017 pseudos requiring hard registers. We could do more register
6018 pressure calculations at any given program point to avoid necessary
6019 splitting even more but it is to expensive and the current approach
6020 works well enough. */
6021 static bool
6022 inherit_in_ebb (rtx_insn *head, rtx_insn *tail)
6024 int i, src_regno, dst_regno, nregs;
6025 bool change_p, succ_p, update_reloads_num_p;
6026 rtx_insn *prev_insn, *last_insn;
6027 rtx next_usage_insns, curr_set;
6028 enum reg_class cl;
6029 struct lra_insn_reg *reg;
6030 basic_block last_processed_bb, curr_bb = NULL;
6031 HARD_REG_SET potential_reload_hard_regs, live_hard_regs;
6032 bitmap to_process;
6033 unsigned int j;
6034 bitmap_iterator bi;
6035 bool head_p, after_p;
6037 change_p = false;
6038 curr_usage_insns_check++;
6039 clear_invariants ();
6040 reloads_num = calls_num = 0;
6041 bitmap_clear (&check_only_regs);
6042 bitmap_clear (&invalid_invariant_regs);
6043 last_processed_bb = NULL;
6044 CLEAR_HARD_REG_SET (potential_reload_hard_regs);
6045 COPY_HARD_REG_SET (live_hard_regs, eliminable_regset);
6046 IOR_HARD_REG_SET (live_hard_regs, lra_no_alloc_regs);
6047 /* We don't process new insns generated in the loop. */
6048 for (curr_insn = tail; curr_insn != PREV_INSN (head); curr_insn = prev_insn)
6050 prev_insn = PREV_INSN (curr_insn);
6051 if (BLOCK_FOR_INSN (curr_insn) != NULL)
6052 curr_bb = BLOCK_FOR_INSN (curr_insn);
6053 if (last_processed_bb != curr_bb)
6055 /* We are at the end of BB. Add qualified living
6056 pseudos for potential splitting. */
6057 to_process = df_get_live_out (curr_bb);
6058 if (last_processed_bb != NULL)
6060 /* We are somewhere in the middle of EBB. */
6061 get_live_on_other_edges (curr_bb, last_processed_bb,
6062 &temp_bitmap);
6063 to_process = &temp_bitmap;
6065 last_processed_bb = curr_bb;
6066 last_insn = get_last_insertion_point (curr_bb);
6067 after_p = (! JUMP_P (last_insn)
6068 && (! CALL_P (last_insn)
6069 || (find_reg_note (last_insn,
6070 REG_NORETURN, NULL_RTX) == NULL_RTX
6071 && ! SIBLING_CALL_P (last_insn))));
6072 CLEAR_HARD_REG_SET (potential_reload_hard_regs);
6073 EXECUTE_IF_SET_IN_BITMAP (to_process, 0, j, bi)
6075 if ((int) j >= lra_constraint_new_regno_start)
6076 break;
6077 if (j < FIRST_PSEUDO_REGISTER || reg_renumber[j] >= 0)
6079 if (j < FIRST_PSEUDO_REGISTER)
6080 SET_HARD_REG_BIT (live_hard_regs, j);
6081 else
6082 add_to_hard_reg_set (&live_hard_regs,
6083 PSEUDO_REGNO_MODE (j),
6084 reg_renumber[j]);
6085 setup_next_usage_insn (j, last_insn, reloads_num, after_p);
6089 src_regno = dst_regno = -1;
6090 curr_set = single_set (curr_insn);
6091 if (curr_set != NULL_RTX && REG_P (SET_DEST (curr_set)))
6092 dst_regno = REGNO (SET_DEST (curr_set));
6093 if (curr_set != NULL_RTX && REG_P (SET_SRC (curr_set)))
6094 src_regno = REGNO (SET_SRC (curr_set));
6095 update_reloads_num_p = true;
6096 if (src_regno < lra_constraint_new_regno_start
6097 && src_regno >= FIRST_PSEUDO_REGISTER
6098 && reg_renumber[src_regno] < 0
6099 && dst_regno >= lra_constraint_new_regno_start
6100 && (cl = lra_get_allocno_class (dst_regno)) != NO_REGS)
6102 /* 'reload_pseudo <- original_pseudo'. */
6103 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6104 reloads_num++;
6105 update_reloads_num_p = false;
6106 succ_p = false;
6107 if (usage_insns[src_regno].check == curr_usage_insns_check
6108 && (next_usage_insns = usage_insns[src_regno].insns) != NULL_RTX)
6109 succ_p = inherit_reload_reg (false, src_regno, cl,
6110 curr_insn, next_usage_insns);
6111 if (succ_p)
6112 change_p = true;
6113 else
6114 setup_next_usage_insn (src_regno, curr_insn, reloads_num, false);
6115 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6116 IOR_HARD_REG_SET (potential_reload_hard_regs,
6117 reg_class_contents[cl]);
6119 else if (src_regno < 0
6120 && dst_regno >= lra_constraint_new_regno_start
6121 && invariant_p (SET_SRC (curr_set))
6122 && (cl = lra_get_allocno_class (dst_regno)) != NO_REGS
6123 && ! bitmap_bit_p (&invalid_invariant_regs, dst_regno)
6124 && ! bitmap_bit_p (&invalid_invariant_regs,
6125 ORIGINAL_REGNO(regno_reg_rtx[dst_regno])))
6127 /* 'reload_pseudo <- invariant'. */
6128 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6129 reloads_num++;
6130 update_reloads_num_p = false;
6131 if (process_invariant_for_inheritance (SET_DEST (curr_set), SET_SRC (curr_set)))
6132 change_p = true;
6133 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6134 IOR_HARD_REG_SET (potential_reload_hard_regs,
6135 reg_class_contents[cl]);
6137 else if (src_regno >= lra_constraint_new_regno_start
6138 && dst_regno < lra_constraint_new_regno_start
6139 && dst_regno >= FIRST_PSEUDO_REGISTER
6140 && reg_renumber[dst_regno] < 0
6141 && (cl = lra_get_allocno_class (src_regno)) != NO_REGS
6142 && usage_insns[dst_regno].check == curr_usage_insns_check
6143 && (next_usage_insns
6144 = usage_insns[dst_regno].insns) != NULL_RTX)
6146 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6147 reloads_num++;
6148 update_reloads_num_p = false;
6149 /* 'original_pseudo <- reload_pseudo'. */
6150 if (! JUMP_P (curr_insn)
6151 && inherit_reload_reg (true, dst_regno, cl,
6152 curr_insn, next_usage_insns))
6153 change_p = true;
6154 /* Invalidate. */
6155 usage_insns[dst_regno].check = 0;
6156 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6157 IOR_HARD_REG_SET (potential_reload_hard_regs,
6158 reg_class_contents[cl]);
6160 else if (INSN_P (curr_insn))
6162 int iter;
6163 int max_uid = get_max_uid ();
6165 curr_id = lra_get_insn_recog_data (curr_insn);
6166 curr_static_id = curr_id->insn_static_data;
6167 to_inherit_num = 0;
6168 /* Process insn definitions. */
6169 for (iter = 0; iter < 2; iter++)
6170 for (reg = iter == 0 ? curr_id->regs : curr_static_id->hard_regs;
6171 reg != NULL;
6172 reg = reg->next)
6173 if (reg->type != OP_IN
6174 && (dst_regno = reg->regno) < lra_constraint_new_regno_start)
6176 if (dst_regno >= FIRST_PSEUDO_REGISTER && reg->type == OP_OUT
6177 && reg_renumber[dst_regno] < 0 && ! reg->subreg_p
6178 && usage_insns[dst_regno].check == curr_usage_insns_check
6179 && (next_usage_insns
6180 = usage_insns[dst_regno].insns) != NULL_RTX)
6182 struct lra_insn_reg *r;
6184 for (r = curr_id->regs; r != NULL; r = r->next)
6185 if (r->type != OP_OUT && r->regno == dst_regno)
6186 break;
6187 /* Don't do inheritance if the pseudo is also
6188 used in the insn. */
6189 if (r == NULL)
6190 /* We can not do inheritance right now
6191 because the current insn reg info (chain
6192 regs) can change after that. */
6193 add_to_inherit (dst_regno, next_usage_insns);
6195 /* We can not process one reg twice here because of
6196 usage_insns invalidation. */
6197 if ((dst_regno < FIRST_PSEUDO_REGISTER
6198 || reg_renumber[dst_regno] >= 0)
6199 && ! reg->subreg_p && reg->type != OP_IN)
6201 HARD_REG_SET s;
6203 if (split_if_necessary (dst_regno, reg->biggest_mode,
6204 potential_reload_hard_regs,
6205 false, curr_insn, max_uid))
6206 change_p = true;
6207 CLEAR_HARD_REG_SET (s);
6208 if (dst_regno < FIRST_PSEUDO_REGISTER)
6209 add_to_hard_reg_set (&s, reg->biggest_mode, dst_regno);
6210 else
6211 add_to_hard_reg_set (&s, PSEUDO_REGNO_MODE (dst_regno),
6212 reg_renumber[dst_regno]);
6213 AND_COMPL_HARD_REG_SET (live_hard_regs, s);
6215 /* We should invalidate potential inheritance or
6216 splitting for the current insn usages to the next
6217 usage insns (see code below) as the output pseudo
6218 prevents this. */
6219 if ((dst_regno >= FIRST_PSEUDO_REGISTER
6220 && reg_renumber[dst_regno] < 0)
6221 || (reg->type == OP_OUT && ! reg->subreg_p
6222 && (dst_regno < FIRST_PSEUDO_REGISTER
6223 || reg_renumber[dst_regno] >= 0)))
6225 /* Invalidate and mark definitions. */
6226 if (dst_regno >= FIRST_PSEUDO_REGISTER)
6227 usage_insns[dst_regno].check = -(int) INSN_UID (curr_insn);
6228 else
6230 nregs = hard_regno_nregs (dst_regno,
6231 reg->biggest_mode);
6232 for (i = 0; i < nregs; i++)
6233 usage_insns[dst_regno + i].check
6234 = -(int) INSN_UID (curr_insn);
6238 /* Process clobbered call regs. */
6239 if (curr_id->arg_hard_regs != NULL)
6240 for (i = 0; (dst_regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6241 if (dst_regno >= FIRST_PSEUDO_REGISTER)
6242 usage_insns[dst_regno - FIRST_PSEUDO_REGISTER].check
6243 = -(int) INSN_UID (curr_insn);
6244 if (! JUMP_P (curr_insn))
6245 for (i = 0; i < to_inherit_num; i++)
6246 if (inherit_reload_reg (true, to_inherit[i].regno,
6247 ALL_REGS, curr_insn,
6248 to_inherit[i].insns))
6249 change_p = true;
6250 if (CALL_P (curr_insn))
6252 rtx cheap, pat, dest;
6253 rtx_insn *restore;
6254 int regno, hard_regno;
6256 calls_num++;
6257 if ((cheap = find_reg_note (curr_insn,
6258 REG_RETURNED, NULL_RTX)) != NULL_RTX
6259 && ((cheap = XEXP (cheap, 0)), true)
6260 && (regno = REGNO (cheap)) >= FIRST_PSEUDO_REGISTER
6261 && (hard_regno = reg_renumber[regno]) >= 0
6262 && usage_insns[regno].check == curr_usage_insns_check
6263 /* If there are pending saves/restores, the
6264 optimization is not worth. */
6265 && usage_insns[regno].calls_num == calls_num - 1
6266 && TEST_HARD_REG_BIT (call_used_reg_set, hard_regno))
6268 /* Restore the pseudo from the call result as
6269 REG_RETURNED note says that the pseudo value is
6270 in the call result and the pseudo is an argument
6271 of the call. */
6272 pat = PATTERN (curr_insn);
6273 if (GET_CODE (pat) == PARALLEL)
6274 pat = XVECEXP (pat, 0, 0);
6275 dest = SET_DEST (pat);
6276 /* For multiple return values dest is PARALLEL.
6277 Currently we handle only single return value case. */
6278 if (REG_P (dest))
6280 start_sequence ();
6281 emit_move_insn (cheap, copy_rtx (dest));
6282 restore = get_insns ();
6283 end_sequence ();
6284 lra_process_new_insns (curr_insn, NULL, restore,
6285 "Inserting call parameter restore");
6286 /* We don't need to save/restore of the pseudo from
6287 this call. */
6288 usage_insns[regno].calls_num = calls_num;
6289 bitmap_set_bit (&check_only_regs, regno);
6293 to_inherit_num = 0;
6294 /* Process insn usages. */
6295 for (iter = 0; iter < 2; iter++)
6296 for (reg = iter == 0 ? curr_id->regs : curr_static_id->hard_regs;
6297 reg != NULL;
6298 reg = reg->next)
6299 if ((reg->type != OP_OUT
6300 || (reg->type == OP_OUT && reg->subreg_p))
6301 && (src_regno = reg->regno) < lra_constraint_new_regno_start)
6303 if (src_regno >= FIRST_PSEUDO_REGISTER
6304 && reg_renumber[src_regno] < 0 && reg->type == OP_IN)
6306 if (usage_insns[src_regno].check == curr_usage_insns_check
6307 && (next_usage_insns
6308 = usage_insns[src_regno].insns) != NULL_RTX
6309 && NONDEBUG_INSN_P (curr_insn))
6310 add_to_inherit (src_regno, next_usage_insns);
6311 else if (usage_insns[src_regno].check
6312 != -(int) INSN_UID (curr_insn))
6313 /* Add usages but only if the reg is not set up
6314 in the same insn. */
6315 add_next_usage_insn (src_regno, curr_insn, reloads_num);
6317 else if (src_regno < FIRST_PSEUDO_REGISTER
6318 || reg_renumber[src_regno] >= 0)
6320 bool before_p;
6321 rtx_insn *use_insn = curr_insn;
6323 before_p = (JUMP_P (curr_insn)
6324 || (CALL_P (curr_insn) && reg->type == OP_IN));
6325 if (NONDEBUG_INSN_P (curr_insn)
6326 && (! JUMP_P (curr_insn) || reg->type == OP_IN)
6327 && split_if_necessary (src_regno, reg->biggest_mode,
6328 potential_reload_hard_regs,
6329 before_p, curr_insn, max_uid))
6331 if (reg->subreg_p)
6332 lra_risky_transformations_p = true;
6333 change_p = true;
6334 /* Invalidate. */
6335 usage_insns[src_regno].check = 0;
6336 if (before_p)
6337 use_insn = PREV_INSN (curr_insn);
6339 if (NONDEBUG_INSN_P (curr_insn))
6341 if (src_regno < FIRST_PSEUDO_REGISTER)
6342 add_to_hard_reg_set (&live_hard_regs,
6343 reg->biggest_mode, src_regno);
6344 else
6345 add_to_hard_reg_set (&live_hard_regs,
6346 PSEUDO_REGNO_MODE (src_regno),
6347 reg_renumber[src_regno]);
6349 if (src_regno >= FIRST_PSEUDO_REGISTER)
6350 add_next_usage_insn (src_regno, use_insn, reloads_num);
6351 else
6353 for (i = 0; i < hard_regno_nregs (src_regno, reg->biggest_mode); i++)
6354 add_next_usage_insn (src_regno + i, use_insn, reloads_num);
6358 /* Process used call regs. */
6359 if (curr_id->arg_hard_regs != NULL)
6360 for (i = 0; (src_regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6361 if (src_regno < FIRST_PSEUDO_REGISTER)
6363 SET_HARD_REG_BIT (live_hard_regs, src_regno);
6364 add_next_usage_insn (src_regno, curr_insn, reloads_num);
6366 for (i = 0; i < to_inherit_num; i++)
6368 src_regno = to_inherit[i].regno;
6369 if (inherit_reload_reg (false, src_regno, ALL_REGS,
6370 curr_insn, to_inherit[i].insns))
6371 change_p = true;
6372 else
6373 setup_next_usage_insn (src_regno, curr_insn, reloads_num, false);
6376 if (update_reloads_num_p
6377 && NONDEBUG_INSN_P (curr_insn) && curr_set != NULL_RTX)
6379 int regno = -1;
6380 if ((REG_P (SET_DEST (curr_set))
6381 && (regno = REGNO (SET_DEST (curr_set))) >= lra_constraint_new_regno_start
6382 && reg_renumber[regno] < 0
6383 && (cl = lra_get_allocno_class (regno)) != NO_REGS)
6384 || (REG_P (SET_SRC (curr_set))
6385 && (regno = REGNO (SET_SRC (curr_set))) >= lra_constraint_new_regno_start
6386 && reg_renumber[regno] < 0
6387 && (cl = lra_get_allocno_class (regno)) != NO_REGS))
6389 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6390 reloads_num++;
6391 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6392 IOR_HARD_REG_SET (potential_reload_hard_regs,
6393 reg_class_contents[cl]);
6396 if (NONDEBUG_INSN_P (curr_insn))
6398 int regno;
6400 /* Invalidate invariants with changed regs. */
6401 curr_id = lra_get_insn_recog_data (curr_insn);
6402 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
6403 if (reg->type != OP_IN)
6405 bitmap_set_bit (&invalid_invariant_regs, reg->regno);
6406 bitmap_set_bit (&invalid_invariant_regs,
6407 ORIGINAL_REGNO (regno_reg_rtx[reg->regno]));
6409 curr_static_id = curr_id->insn_static_data;
6410 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
6411 if (reg->type != OP_IN)
6412 bitmap_set_bit (&invalid_invariant_regs, reg->regno);
6413 if (curr_id->arg_hard_regs != NULL)
6414 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6415 if (regno >= FIRST_PSEUDO_REGISTER)
6416 bitmap_set_bit (&invalid_invariant_regs,
6417 regno - FIRST_PSEUDO_REGISTER);
6419 /* We reached the start of the current basic block. */
6420 if (prev_insn == NULL_RTX || prev_insn == PREV_INSN (head)
6421 || BLOCK_FOR_INSN (prev_insn) != curr_bb)
6423 /* We reached the beginning of the current block -- do
6424 rest of spliting in the current BB. */
6425 to_process = df_get_live_in (curr_bb);
6426 if (BLOCK_FOR_INSN (head) != curr_bb)
6428 /* We are somewhere in the middle of EBB. */
6429 get_live_on_other_edges (EDGE_PRED (curr_bb, 0)->src,
6430 curr_bb, &temp_bitmap);
6431 to_process = &temp_bitmap;
6433 head_p = true;
6434 EXECUTE_IF_SET_IN_BITMAP (to_process, 0, j, bi)
6436 if ((int) j >= lra_constraint_new_regno_start)
6437 break;
6438 if (((int) j < FIRST_PSEUDO_REGISTER || reg_renumber[j] >= 0)
6439 && usage_insns[j].check == curr_usage_insns_check
6440 && (next_usage_insns = usage_insns[j].insns) != NULL_RTX)
6442 if (need_for_split_p (potential_reload_hard_regs, j))
6444 if (lra_dump_file != NULL && head_p)
6446 fprintf (lra_dump_file,
6447 " ----------------------------------\n");
6448 head_p = false;
6450 if (split_reg (false, j, bb_note (curr_bb),
6451 next_usage_insns))
6452 change_p = true;
6454 usage_insns[j].check = 0;
6459 return change_p;
6462 /* This value affects EBB forming. If probability of edge from EBB to
6463 a BB is not greater than the following value, we don't add the BB
6464 to EBB. */
6465 #define EBB_PROBABILITY_CUTOFF \
6466 ((REG_BR_PROB_BASE * LRA_INHERITANCE_EBB_PROBABILITY_CUTOFF) / 100)
6468 /* Current number of inheritance/split iteration. */
6469 int lra_inheritance_iter;
6471 /* Entry function for inheritance/split pass. */
6472 void
6473 lra_inheritance (void)
6475 int i;
6476 basic_block bb, start_bb;
6477 edge e;
6479 lra_inheritance_iter++;
6480 if (lra_inheritance_iter > LRA_MAX_INHERITANCE_PASSES)
6481 return;
6482 timevar_push (TV_LRA_INHERITANCE);
6483 if (lra_dump_file != NULL)
6484 fprintf (lra_dump_file, "\n********** Inheritance #%d: **********\n\n",
6485 lra_inheritance_iter);
6486 curr_usage_insns_check = 0;
6487 usage_insns = XNEWVEC (struct usage_insns, lra_constraint_new_regno_start);
6488 for (i = 0; i < lra_constraint_new_regno_start; i++)
6489 usage_insns[i].check = 0;
6490 bitmap_initialize (&check_only_regs, &reg_obstack);
6491 bitmap_initialize (&invalid_invariant_regs, &reg_obstack);
6492 bitmap_initialize (&live_regs, &reg_obstack);
6493 bitmap_initialize (&temp_bitmap, &reg_obstack);
6494 bitmap_initialize (&ebb_global_regs, &reg_obstack);
6495 FOR_EACH_BB_FN (bb, cfun)
6497 start_bb = bb;
6498 if (lra_dump_file != NULL)
6499 fprintf (lra_dump_file, "EBB");
6500 /* Form a EBB starting with BB. */
6501 bitmap_clear (&ebb_global_regs);
6502 bitmap_ior_into (&ebb_global_regs, df_get_live_in (bb));
6503 for (;;)
6505 if (lra_dump_file != NULL)
6506 fprintf (lra_dump_file, " %d", bb->index);
6507 if (bb->next_bb == EXIT_BLOCK_PTR_FOR_FN (cfun)
6508 || LABEL_P (BB_HEAD (bb->next_bb)))
6509 break;
6510 e = find_fallthru_edge (bb->succs);
6511 if (! e)
6512 break;
6513 if (e->probability.initialized_p ()
6514 && e->probability.to_reg_br_prob_base () < EBB_PROBABILITY_CUTOFF)
6515 break;
6516 bb = bb->next_bb;
6518 bitmap_ior_into (&ebb_global_regs, df_get_live_out (bb));
6519 if (lra_dump_file != NULL)
6520 fprintf (lra_dump_file, "\n");
6521 if (inherit_in_ebb (BB_HEAD (start_bb), BB_END (bb)))
6522 /* Remember that the EBB head and tail can change in
6523 inherit_in_ebb. */
6524 update_ebb_live_info (BB_HEAD (start_bb), BB_END (bb));
6526 bitmap_clear (&ebb_global_regs);
6527 bitmap_clear (&temp_bitmap);
6528 bitmap_clear (&live_regs);
6529 bitmap_clear (&invalid_invariant_regs);
6530 bitmap_clear (&check_only_regs);
6531 free (usage_insns);
6533 timevar_pop (TV_LRA_INHERITANCE);
6538 /* This page contains code to undo failed inheritance/split
6539 transformations. */
6541 /* Current number of iteration undoing inheritance/split. */
6542 int lra_undo_inheritance_iter;
6544 /* Fix BB live info LIVE after removing pseudos created on pass doing
6545 inheritance/split which are REMOVED_PSEUDOS. */
6546 static void
6547 fix_bb_live_info (bitmap live, bitmap removed_pseudos)
6549 unsigned int regno;
6550 bitmap_iterator bi;
6552 EXECUTE_IF_SET_IN_BITMAP (removed_pseudos, 0, regno, bi)
6553 if (bitmap_clear_bit (live, regno)
6554 && REG_P (lra_reg_info[regno].restore_rtx))
6555 bitmap_set_bit (live, REGNO (lra_reg_info[regno].restore_rtx));
6558 /* Return regno of the (subreg of) REG. Otherwise, return a negative
6559 number. */
6560 static int
6561 get_regno (rtx reg)
6563 if (GET_CODE (reg) == SUBREG)
6564 reg = SUBREG_REG (reg);
6565 if (REG_P (reg))
6566 return REGNO (reg);
6567 return -1;
6570 /* Delete a move INSN with destination reg DREGNO and a previous
6571 clobber insn with the same regno. The inheritance/split code can
6572 generate moves with preceding clobber and when we delete such moves
6573 we should delete the clobber insn too to keep the correct life
6574 info. */
6575 static void
6576 delete_move_and_clobber (rtx_insn *insn, int dregno)
6578 rtx_insn *prev_insn = PREV_INSN (insn);
6580 lra_set_insn_deleted (insn);
6581 lra_assert (dregno >= 0);
6582 if (prev_insn != NULL && NONDEBUG_INSN_P (prev_insn)
6583 && GET_CODE (PATTERN (prev_insn)) == CLOBBER
6584 && dregno == get_regno (XEXP (PATTERN (prev_insn), 0)))
6585 lra_set_insn_deleted (prev_insn);
6588 /* Remove inheritance/split pseudos which are in REMOVE_PSEUDOS and
6589 return true if we did any change. The undo transformations for
6590 inheritance looks like
6591 i <- i2
6592 p <- i => p <- i2
6593 or removing
6594 p <- i, i <- p, and i <- i3
6595 where p is original pseudo from which inheritance pseudo i was
6596 created, i and i3 are removed inheritance pseudos, i2 is another
6597 not removed inheritance pseudo. All split pseudos or other
6598 occurrences of removed inheritance pseudos are changed on the
6599 corresponding original pseudos.
6601 The function also schedules insns changed and created during
6602 inheritance/split pass for processing by the subsequent constraint
6603 pass. */
6604 static bool
6605 remove_inheritance_pseudos (bitmap remove_pseudos)
6607 basic_block bb;
6608 int regno, sregno, prev_sregno, dregno;
6609 rtx restore_rtx;
6610 rtx set, prev_set;
6611 rtx_insn *prev_insn;
6612 bool change_p, done_p;
6614 change_p = ! bitmap_empty_p (remove_pseudos);
6615 /* We can not finish the function right away if CHANGE_P is true
6616 because we need to marks insns affected by previous
6617 inheritance/split pass for processing by the subsequent
6618 constraint pass. */
6619 FOR_EACH_BB_FN (bb, cfun)
6621 fix_bb_live_info (df_get_live_in (bb), remove_pseudos);
6622 fix_bb_live_info (df_get_live_out (bb), remove_pseudos);
6623 FOR_BB_INSNS_REVERSE (bb, curr_insn)
6625 if (! INSN_P (curr_insn))
6626 continue;
6627 done_p = false;
6628 sregno = dregno = -1;
6629 if (change_p && NONDEBUG_INSN_P (curr_insn)
6630 && (set = single_set (curr_insn)) != NULL_RTX)
6632 dregno = get_regno (SET_DEST (set));
6633 sregno = get_regno (SET_SRC (set));
6636 if (sregno >= 0 && dregno >= 0)
6638 if (bitmap_bit_p (remove_pseudos, dregno)
6639 && ! REG_P (lra_reg_info[dregno].restore_rtx))
6641 /* invariant inheritance pseudo <- original pseudo */
6642 if (lra_dump_file != NULL)
6644 fprintf (lra_dump_file, " Removing invariant inheritance:\n");
6645 dump_insn_slim (lra_dump_file, curr_insn);
6646 fprintf (lra_dump_file, "\n");
6648 delete_move_and_clobber (curr_insn, dregno);
6649 done_p = true;
6651 else if (bitmap_bit_p (remove_pseudos, sregno)
6652 && ! REG_P (lra_reg_info[sregno].restore_rtx))
6654 /* reload pseudo <- invariant inheritance pseudo */
6655 start_sequence ();
6656 /* We can not just change the source. It might be
6657 an insn different from the move. */
6658 emit_insn (lra_reg_info[sregno].restore_rtx);
6659 rtx_insn *new_insns = get_insns ();
6660 end_sequence ();
6661 lra_assert (single_set (new_insns) != NULL
6662 && SET_DEST (set) == SET_DEST (single_set (new_insns)));
6663 lra_process_new_insns (curr_insn, NULL, new_insns,
6664 "Changing reload<-invariant inheritance");
6665 delete_move_and_clobber (curr_insn, dregno);
6666 done_p = true;
6668 else if ((bitmap_bit_p (remove_pseudos, sregno)
6669 && (get_regno (lra_reg_info[sregno].restore_rtx) == dregno
6670 || (bitmap_bit_p (remove_pseudos, dregno)
6671 && get_regno (lra_reg_info[sregno].restore_rtx) >= 0
6672 && (get_regno (lra_reg_info[sregno].restore_rtx)
6673 == get_regno (lra_reg_info[dregno].restore_rtx)))))
6674 || (bitmap_bit_p (remove_pseudos, dregno)
6675 && get_regno (lra_reg_info[dregno].restore_rtx) == sregno))
6676 /* One of the following cases:
6677 original <- removed inheritance pseudo
6678 removed inherit pseudo <- another removed inherit pseudo
6679 removed inherit pseudo <- original pseudo
6681 removed_split_pseudo <- original_reg
6682 original_reg <- removed_split_pseudo */
6684 if (lra_dump_file != NULL)
6686 fprintf (lra_dump_file, " Removing %s:\n",
6687 bitmap_bit_p (&lra_split_regs, sregno)
6688 || bitmap_bit_p (&lra_split_regs, dregno)
6689 ? "split" : "inheritance");
6690 dump_insn_slim (lra_dump_file, curr_insn);
6692 delete_move_and_clobber (curr_insn, dregno);
6693 done_p = true;
6695 else if (bitmap_bit_p (remove_pseudos, sregno)
6696 && bitmap_bit_p (&lra_inheritance_pseudos, sregno))
6698 /* Search the following pattern:
6699 inherit_or_split_pseudo1 <- inherit_or_split_pseudo2
6700 original_pseudo <- inherit_or_split_pseudo1
6701 where the 2nd insn is the current insn and
6702 inherit_or_split_pseudo2 is not removed. If it is found,
6703 change the current insn onto:
6704 original_pseudo <- inherit_or_split_pseudo2. */
6705 for (prev_insn = PREV_INSN (curr_insn);
6706 prev_insn != NULL_RTX && ! NONDEBUG_INSN_P (prev_insn);
6707 prev_insn = PREV_INSN (prev_insn))
6709 if (prev_insn != NULL_RTX && BLOCK_FOR_INSN (prev_insn) == bb
6710 && (prev_set = single_set (prev_insn)) != NULL_RTX
6711 /* There should be no subregs in insn we are
6712 searching because only the original reg might
6713 be in subreg when we changed the mode of
6714 load/store for splitting. */
6715 && REG_P (SET_DEST (prev_set))
6716 && REG_P (SET_SRC (prev_set))
6717 && (int) REGNO (SET_DEST (prev_set)) == sregno
6718 && ((prev_sregno = REGNO (SET_SRC (prev_set)))
6719 >= FIRST_PSEUDO_REGISTER)
6720 && (lra_reg_info[prev_sregno].restore_rtx == NULL_RTX
6722 /* As we consider chain of inheritance or
6723 splitting described in above comment we should
6724 check that sregno and prev_sregno were
6725 inheritance/split pseudos created from the
6726 same original regno. */
6727 (get_regno (lra_reg_info[sregno].restore_rtx) >= 0
6728 && (get_regno (lra_reg_info[sregno].restore_rtx)
6729 == get_regno (lra_reg_info[prev_sregno].restore_rtx))))
6730 && ! bitmap_bit_p (remove_pseudos, prev_sregno))
6732 lra_assert (GET_MODE (SET_SRC (prev_set))
6733 == GET_MODE (regno_reg_rtx[sregno]));
6734 /* Although we have a single set, the insn can
6735 contain more one sregno register occurrence
6736 as a source. Change all occurrences. */
6737 lra_substitute_pseudo_within_insn (curr_insn, sregno,
6738 SET_SRC (prev_set),
6739 false);
6740 /* As we are finishing with processing the insn
6741 here, check the destination too as it might
6742 inheritance pseudo for another pseudo. */
6743 if (bitmap_bit_p (remove_pseudos, dregno)
6744 && bitmap_bit_p (&lra_inheritance_pseudos, dregno)
6745 && (restore_rtx
6746 = lra_reg_info[dregno].restore_rtx) != NULL_RTX)
6748 if (GET_CODE (SET_DEST (set)) == SUBREG)
6749 SUBREG_REG (SET_DEST (set)) = restore_rtx;
6750 else
6751 SET_DEST (set) = restore_rtx;
6753 lra_push_insn_and_update_insn_regno_info (curr_insn);
6754 lra_set_used_insn_alternative_by_uid
6755 (INSN_UID (curr_insn), -1);
6756 done_p = true;
6757 if (lra_dump_file != NULL)
6759 fprintf (lra_dump_file, " Change reload insn:\n");
6760 dump_insn_slim (lra_dump_file, curr_insn);
6765 if (! done_p)
6767 struct lra_insn_reg *reg;
6768 bool restored_regs_p = false;
6769 bool kept_regs_p = false;
6771 curr_id = lra_get_insn_recog_data (curr_insn);
6772 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
6774 regno = reg->regno;
6775 restore_rtx = lra_reg_info[regno].restore_rtx;
6776 if (restore_rtx != NULL_RTX)
6778 if (change_p && bitmap_bit_p (remove_pseudos, regno))
6780 lra_substitute_pseudo_within_insn
6781 (curr_insn, regno, restore_rtx, false);
6782 restored_regs_p = true;
6784 else
6785 kept_regs_p = true;
6788 if (NONDEBUG_INSN_P (curr_insn) && kept_regs_p)
6790 /* The instruction has changed since the previous
6791 constraints pass. */
6792 lra_push_insn_and_update_insn_regno_info (curr_insn);
6793 lra_set_used_insn_alternative_by_uid
6794 (INSN_UID (curr_insn), -1);
6796 else if (restored_regs_p)
6797 /* The instruction has been restored to the form that
6798 it had during the previous constraints pass. */
6799 lra_update_insn_regno_info (curr_insn);
6800 if (restored_regs_p && lra_dump_file != NULL)
6802 fprintf (lra_dump_file, " Insn after restoring regs:\n");
6803 dump_insn_slim (lra_dump_file, curr_insn);
6808 return change_p;
6811 /* If optional reload pseudos failed to get a hard register or was not
6812 inherited, it is better to remove optional reloads. We do this
6813 transformation after undoing inheritance to figure out necessity to
6814 remove optional reloads easier. Return true if we do any
6815 change. */
6816 static bool
6817 undo_optional_reloads (void)
6819 bool change_p, keep_p;
6820 unsigned int regno, uid;
6821 bitmap_iterator bi, bi2;
6822 rtx_insn *insn;
6823 rtx set, src, dest;
6824 auto_bitmap removed_optional_reload_pseudos (&reg_obstack);
6826 bitmap_copy (removed_optional_reload_pseudos, &lra_optional_reload_pseudos);
6827 EXECUTE_IF_SET_IN_BITMAP (&lra_optional_reload_pseudos, 0, regno, bi)
6829 keep_p = false;
6830 /* Keep optional reloads from previous subpasses. */
6831 if (lra_reg_info[regno].restore_rtx == NULL_RTX
6832 /* If the original pseudo changed its allocation, just
6833 removing the optional pseudo is dangerous as the original
6834 pseudo will have longer live range. */
6835 || reg_renumber[REGNO (lra_reg_info[regno].restore_rtx)] >= 0)
6836 keep_p = true;
6837 else if (reg_renumber[regno] >= 0)
6838 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi2)
6840 insn = lra_insn_recog_data[uid]->insn;
6841 if ((set = single_set (insn)) == NULL_RTX)
6842 continue;
6843 src = SET_SRC (set);
6844 dest = SET_DEST (set);
6845 if (! REG_P (src) || ! REG_P (dest))
6846 continue;
6847 if (REGNO (dest) == regno
6848 /* Ignore insn for optional reloads itself. */
6849 && REGNO (lra_reg_info[regno].restore_rtx) != REGNO (src)
6850 /* Check only inheritance on last inheritance pass. */
6851 && (int) REGNO (src) >= new_regno_start
6852 /* Check that the optional reload was inherited. */
6853 && bitmap_bit_p (&lra_inheritance_pseudos, REGNO (src)))
6855 keep_p = true;
6856 break;
6859 if (keep_p)
6861 bitmap_clear_bit (removed_optional_reload_pseudos, regno);
6862 if (lra_dump_file != NULL)
6863 fprintf (lra_dump_file, "Keep optional reload reg %d\n", regno);
6866 change_p = ! bitmap_empty_p (removed_optional_reload_pseudos);
6867 auto_bitmap insn_bitmap (&reg_obstack);
6868 EXECUTE_IF_SET_IN_BITMAP (removed_optional_reload_pseudos, 0, regno, bi)
6870 if (lra_dump_file != NULL)
6871 fprintf (lra_dump_file, "Remove optional reload reg %d\n", regno);
6872 bitmap_copy (insn_bitmap, &lra_reg_info[regno].insn_bitmap);
6873 EXECUTE_IF_SET_IN_BITMAP (insn_bitmap, 0, uid, bi2)
6875 insn = lra_insn_recog_data[uid]->insn;
6876 if ((set = single_set (insn)) != NULL_RTX)
6878 src = SET_SRC (set);
6879 dest = SET_DEST (set);
6880 if (REG_P (src) && REG_P (dest)
6881 && ((REGNO (src) == regno
6882 && (REGNO (lra_reg_info[regno].restore_rtx)
6883 == REGNO (dest)))
6884 || (REGNO (dest) == regno
6885 && (REGNO (lra_reg_info[regno].restore_rtx)
6886 == REGNO (src)))))
6888 if (lra_dump_file != NULL)
6890 fprintf (lra_dump_file, " Deleting move %u\n",
6891 INSN_UID (insn));
6892 dump_insn_slim (lra_dump_file, insn);
6894 delete_move_and_clobber (insn, REGNO (dest));
6895 continue;
6897 /* We should not worry about generation memory-memory
6898 moves here as if the corresponding inheritance did
6899 not work (inheritance pseudo did not get a hard reg),
6900 we remove the inheritance pseudo and the optional
6901 reload. */
6903 lra_substitute_pseudo_within_insn
6904 (insn, regno, lra_reg_info[regno].restore_rtx, false);
6905 lra_update_insn_regno_info (insn);
6906 if (lra_dump_file != NULL)
6908 fprintf (lra_dump_file,
6909 " Restoring original insn:\n");
6910 dump_insn_slim (lra_dump_file, insn);
6914 /* Clear restore_regnos. */
6915 EXECUTE_IF_SET_IN_BITMAP (&lra_optional_reload_pseudos, 0, regno, bi)
6916 lra_reg_info[regno].restore_rtx = NULL_RTX;
6917 return change_p;
6920 /* Entry function for undoing inheritance/split transformation. Return true
6921 if we did any RTL change in this pass. */
6922 bool
6923 lra_undo_inheritance (void)
6925 unsigned int regno;
6926 int hard_regno;
6927 int n_all_inherit, n_inherit, n_all_split, n_split;
6928 rtx restore_rtx;
6929 bitmap_iterator bi;
6930 bool change_p;
6932 lra_undo_inheritance_iter++;
6933 if (lra_undo_inheritance_iter > LRA_MAX_INHERITANCE_PASSES)
6934 return false;
6935 if (lra_dump_file != NULL)
6936 fprintf (lra_dump_file,
6937 "\n********** Undoing inheritance #%d: **********\n\n",
6938 lra_undo_inheritance_iter);
6939 auto_bitmap remove_pseudos (&reg_obstack);
6940 n_inherit = n_all_inherit = 0;
6941 EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, regno, bi)
6942 if (lra_reg_info[regno].restore_rtx != NULL_RTX)
6944 n_all_inherit++;
6945 if (reg_renumber[regno] < 0
6946 /* If the original pseudo changed its allocation, just
6947 removing inheritance is dangerous as for changing
6948 allocation we used shorter live-ranges. */
6949 && (! REG_P (lra_reg_info[regno].restore_rtx)
6950 || reg_renumber[REGNO (lra_reg_info[regno].restore_rtx)] < 0))
6951 bitmap_set_bit (remove_pseudos, regno);
6952 else
6953 n_inherit++;
6955 if (lra_dump_file != NULL && n_all_inherit != 0)
6956 fprintf (lra_dump_file, "Inherit %d out of %d (%.2f%%)\n",
6957 n_inherit, n_all_inherit,
6958 (double) n_inherit / n_all_inherit * 100);
6959 n_split = n_all_split = 0;
6960 EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, regno, bi)
6961 if ((restore_rtx = lra_reg_info[regno].restore_rtx) != NULL_RTX)
6963 int restore_regno = REGNO (restore_rtx);
6965 n_all_split++;
6966 hard_regno = (restore_regno >= FIRST_PSEUDO_REGISTER
6967 ? reg_renumber[restore_regno] : restore_regno);
6968 if (hard_regno < 0 || reg_renumber[regno] == hard_regno)
6969 bitmap_set_bit (remove_pseudos, regno);
6970 else
6972 n_split++;
6973 if (lra_dump_file != NULL)
6974 fprintf (lra_dump_file, " Keep split r%d (orig=r%d)\n",
6975 regno, restore_regno);
6978 if (lra_dump_file != NULL && n_all_split != 0)
6979 fprintf (lra_dump_file, "Split %d out of %d (%.2f%%)\n",
6980 n_split, n_all_split,
6981 (double) n_split / n_all_split * 100);
6982 change_p = remove_inheritance_pseudos (remove_pseudos);
6983 /* Clear restore_regnos. */
6984 EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, regno, bi)
6985 lra_reg_info[regno].restore_rtx = NULL_RTX;
6986 EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, regno, bi)
6987 lra_reg_info[regno].restore_rtx = NULL_RTX;
6988 change_p = undo_optional_reloads () || change_p;
6989 return change_p;