* gcc-interface/trans.c (process_freeze_entity): Be prepared for a
[official-gcc.git] / gcc / lra-constraints.c
blob4adf4bfea8b02d6f7b0d5607932eb9c3f9e40223
1 /* Code for RTL transformations to satisfy insn constraints.
2 Copyright (C) 2010-2017 Free Software Foundation, Inc.
3 Contributed by Vladimir Makarov <vmakarov@redhat.com>.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
10 version.
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
22 /* This file contains code for 3 passes: constraint pass,
23 inheritance/split pass, and pass for undoing failed inheritance and
24 split.
26 The major goal of constraint pass is to transform RTL to satisfy
27 insn and address constraints by:
28 o choosing insn alternatives;
29 o generating *reload insns* (or reloads in brief) and *reload
30 pseudos* which will get necessary hard registers later;
31 o substituting pseudos with equivalent values and removing the
32 instructions that initialized those pseudos.
34 The constraint pass has biggest and most complicated code in LRA.
35 There are a lot of important details like:
36 o reuse of input reload pseudos to simplify reload pseudo
37 allocations;
38 o some heuristics to choose insn alternative to improve the
39 inheritance;
40 o early clobbers etc.
42 The pass is mimicking former reload pass in alternative choosing
43 because the reload pass is oriented to current machine description
44 model. It might be changed if the machine description model is
45 changed.
47 There is special code for preventing all LRA and this pass cycling
48 in case of bugs.
50 On the first iteration of the pass we process every instruction and
51 choose an alternative for each one. On subsequent iterations we try
52 to avoid reprocessing instructions if we can be sure that the old
53 choice is still valid.
55 The inheritance/spilt pass is to transform code to achieve
56 ineheritance and live range splitting. It is done on backward
57 traversal of EBBs.
59 The inheritance optimization goal is to reuse values in hard
60 registers. There is analogous optimization in old reload pass. The
61 inheritance is achieved by following transformation:
63 reload_p1 <- p reload_p1 <- p
64 ... new_p <- reload_p1
65 ... => ...
66 reload_p2 <- p reload_p2 <- new_p
68 where p is spilled and not changed between the insns. Reload_p1 is
69 also called *original pseudo* and new_p is called *inheritance
70 pseudo*.
72 The subsequent assignment pass will try to assign the same (or
73 another if it is not possible) hard register to new_p as to
74 reload_p1 or reload_p2.
76 If the assignment pass fails to assign a hard register to new_p,
77 this file will undo the inheritance and restore the original code.
78 This is because implementing the above sequence with a spilled
79 new_p would make the code much worse. The inheritance is done in
80 EBB scope. The above is just a simplified example to get an idea
81 of the inheritance as the inheritance is also done for non-reload
82 insns.
84 Splitting (transformation) is also done in EBB scope on the same
85 pass as the inheritance:
87 r <- ... or ... <- r r <- ... or ... <- r
88 ... s <- r (new insn -- save)
89 ... =>
90 ... r <- s (new insn -- restore)
91 ... <- r ... <- r
93 The *split pseudo* s is assigned to the hard register of the
94 original pseudo or hard register r.
96 Splitting is done:
97 o In EBBs with high register pressure for global pseudos (living
98 in at least 2 BBs) and assigned to hard registers when there
99 are more one reloads needing the hard registers;
100 o for pseudos needing save/restore code around calls.
102 If the split pseudo still has the same hard register as the
103 original pseudo after the subsequent assignment pass or the
104 original pseudo was split, the opposite transformation is done on
105 the same pass for undoing inheritance. */
107 #undef REG_OK_STRICT
109 #include "config.h"
110 #include "system.h"
111 #include "coretypes.h"
112 #include "backend.h"
113 #include "target.h"
114 #include "rtl.h"
115 #include "tree.h"
116 #include "predict.h"
117 #include "df.h"
118 #include "memmodel.h"
119 #include "tm_p.h"
120 #include "expmed.h"
121 #include "optabs.h"
122 #include "regs.h"
123 #include "ira.h"
124 #include "recog.h"
125 #include "output.h"
126 #include "addresses.h"
127 #include "expr.h"
128 #include "cfgrtl.h"
129 #include "rtl-error.h"
130 #include "params.h"
131 #include "lra.h"
132 #include "lra-int.h"
133 #include "print-rtl.h"
135 /* Value of LRA_CURR_RELOAD_NUM at the beginning of BB of the current
136 insn. Remember that LRA_CURR_RELOAD_NUM is the number of emitted
137 reload insns. */
138 static int bb_reload_num;
140 /* The current insn being processed and corresponding its single set
141 (NULL otherwise), its data (basic block, the insn data, the insn
142 static data, and the mode of each operand). */
143 static rtx_insn *curr_insn;
144 static rtx curr_insn_set;
145 static basic_block curr_bb;
146 static lra_insn_recog_data_t curr_id;
147 static struct lra_static_insn_data *curr_static_id;
148 static machine_mode curr_operand_mode[MAX_RECOG_OPERANDS];
149 /* Mode of the register substituted by its equivalence with VOIDmode
150 (e.g. constant) and whose subreg is given operand of the current
151 insn. VOIDmode in all other cases. */
152 static machine_mode original_subreg_reg_mode[MAX_RECOG_OPERANDS];
156 /* Start numbers for new registers and insns at the current constraints
157 pass start. */
158 static int new_regno_start;
159 static int new_insn_uid_start;
161 /* If LOC is nonnull, strip any outer subreg from it. */
162 static inline rtx *
163 strip_subreg (rtx *loc)
165 return loc && GET_CODE (*loc) == SUBREG ? &SUBREG_REG (*loc) : loc;
168 /* Return hard regno of REGNO or if it is was not assigned to a hard
169 register, use a hard register from its allocno class. */
170 static int
171 get_try_hard_regno (int regno)
173 int hard_regno;
174 enum reg_class rclass;
176 if ((hard_regno = regno) >= FIRST_PSEUDO_REGISTER)
177 hard_regno = lra_get_regno_hard_regno (regno);
178 if (hard_regno >= 0)
179 return hard_regno;
180 rclass = lra_get_allocno_class (regno);
181 if (rclass == NO_REGS)
182 return -1;
183 return ira_class_hard_regs[rclass][0];
186 /* Return the hard regno of X after removing its subreg. If X is not
187 a register or a subreg of a register, return -1. If X is a pseudo,
188 use its assignment. If FINAL_P return the final hard regno which will
189 be after elimination. */
190 static int
191 get_hard_regno (rtx x, bool final_p)
193 rtx reg;
194 int hard_regno;
196 reg = x;
197 if (SUBREG_P (x))
198 reg = SUBREG_REG (x);
199 if (! REG_P (reg))
200 return -1;
201 if (! HARD_REGISTER_NUM_P (hard_regno = REGNO (reg)))
202 hard_regno = lra_get_regno_hard_regno (hard_regno);
203 if (hard_regno < 0)
204 return -1;
205 if (final_p)
206 hard_regno = lra_get_elimination_hard_regno (hard_regno);
207 if (SUBREG_P (x))
208 hard_regno += subreg_regno_offset (hard_regno, GET_MODE (reg),
209 SUBREG_BYTE (x), GET_MODE (x));
210 return hard_regno;
213 /* If REGNO is a hard register or has been allocated a hard register,
214 return the class of that register. If REGNO is a reload pseudo
215 created by the current constraints pass, return its allocno class.
216 Return NO_REGS otherwise. */
217 static enum reg_class
218 get_reg_class (int regno)
220 int hard_regno;
222 if (! HARD_REGISTER_NUM_P (hard_regno = regno))
223 hard_regno = lra_get_regno_hard_regno (regno);
224 if (hard_regno >= 0)
226 hard_regno = lra_get_elimination_hard_regno (hard_regno);
227 return REGNO_REG_CLASS (hard_regno);
229 if (regno >= new_regno_start)
230 return lra_get_allocno_class (regno);
231 return NO_REGS;
234 /* Return true if REG satisfies (or will satisfy) reg class constraint
235 CL. Use elimination first if REG is a hard register. If REG is a
236 reload pseudo created by this constraints pass, assume that it will
237 be allocated a hard register from its allocno class, but allow that
238 class to be narrowed to CL if it is currently a superset of CL.
240 If NEW_CLASS is nonnull, set *NEW_CLASS to the new allocno class of
241 REGNO (reg), or NO_REGS if no change in its class was needed. */
242 static bool
243 in_class_p (rtx reg, enum reg_class cl, enum reg_class *new_class)
245 enum reg_class rclass, common_class;
246 machine_mode reg_mode;
247 int class_size, hard_regno, nregs, i, j;
248 int regno = REGNO (reg);
250 if (new_class != NULL)
251 *new_class = NO_REGS;
252 if (regno < FIRST_PSEUDO_REGISTER)
254 rtx final_reg = reg;
255 rtx *final_loc = &final_reg;
257 lra_eliminate_reg_if_possible (final_loc);
258 return TEST_HARD_REG_BIT (reg_class_contents[cl], REGNO (*final_loc));
260 reg_mode = GET_MODE (reg);
261 rclass = get_reg_class (regno);
262 if (regno < new_regno_start
263 /* Do not allow the constraints for reload instructions to
264 influence the classes of new pseudos. These reloads are
265 typically moves that have many alternatives, and restricting
266 reload pseudos for one alternative may lead to situations
267 where other reload pseudos are no longer allocatable. */
268 || (INSN_UID (curr_insn) >= new_insn_uid_start
269 && curr_insn_set != NULL
270 && ((OBJECT_P (SET_SRC (curr_insn_set))
271 && ! CONSTANT_P (SET_SRC (curr_insn_set)))
272 || (GET_CODE (SET_SRC (curr_insn_set)) == SUBREG
273 && OBJECT_P (SUBREG_REG (SET_SRC (curr_insn_set)))
274 && ! CONSTANT_P (SUBREG_REG (SET_SRC (curr_insn_set)))))))
275 /* When we don't know what class will be used finally for reload
276 pseudos, we use ALL_REGS. */
277 return ((regno >= new_regno_start && rclass == ALL_REGS)
278 || (rclass != NO_REGS && ira_class_subset_p[rclass][cl]
279 && ! hard_reg_set_subset_p (reg_class_contents[cl],
280 lra_no_alloc_regs)));
281 else
283 common_class = ira_reg_class_subset[rclass][cl];
284 if (new_class != NULL)
285 *new_class = common_class;
286 if (hard_reg_set_subset_p (reg_class_contents[common_class],
287 lra_no_alloc_regs))
288 return false;
289 /* Check that there are enough allocatable regs. */
290 class_size = ira_class_hard_regs_num[common_class];
291 for (i = 0; i < class_size; i++)
293 hard_regno = ira_class_hard_regs[common_class][i];
294 nregs = hard_regno_nregs (hard_regno, reg_mode);
295 if (nregs == 1)
296 return true;
297 for (j = 0; j < nregs; j++)
298 if (TEST_HARD_REG_BIT (lra_no_alloc_regs, hard_regno + j)
299 || ! TEST_HARD_REG_BIT (reg_class_contents[common_class],
300 hard_regno + j))
301 break;
302 if (j >= nregs)
303 return true;
305 return false;
309 /* Return true if REGNO satisfies a memory constraint. */
310 static bool
311 in_mem_p (int regno)
313 return get_reg_class (regno) == NO_REGS;
316 /* Return 1 if ADDR is a valid memory address for mode MODE in address
317 space AS, and check that each pseudo has the proper kind of hard
318 reg. */
319 static int
320 valid_address_p (machine_mode mode ATTRIBUTE_UNUSED,
321 rtx addr, addr_space_t as)
323 #ifdef GO_IF_LEGITIMATE_ADDRESS
324 lra_assert (ADDR_SPACE_GENERIC_P (as));
325 GO_IF_LEGITIMATE_ADDRESS (mode, addr, win);
326 return 0;
328 win:
329 return 1;
330 #else
331 return targetm.addr_space.legitimate_address_p (mode, addr, 0, as);
332 #endif
335 namespace {
336 /* Temporarily eliminates registers in an address (for the lifetime of
337 the object). */
338 class address_eliminator {
339 public:
340 address_eliminator (struct address_info *ad);
341 ~address_eliminator ();
343 private:
344 struct address_info *m_ad;
345 rtx *m_base_loc;
346 rtx m_base_reg;
347 rtx *m_index_loc;
348 rtx m_index_reg;
352 address_eliminator::address_eliminator (struct address_info *ad)
353 : m_ad (ad),
354 m_base_loc (strip_subreg (ad->base_term)),
355 m_base_reg (NULL_RTX),
356 m_index_loc (strip_subreg (ad->index_term)),
357 m_index_reg (NULL_RTX)
359 if (m_base_loc != NULL)
361 m_base_reg = *m_base_loc;
362 lra_eliminate_reg_if_possible (m_base_loc);
363 if (m_ad->base_term2 != NULL)
364 *m_ad->base_term2 = *m_ad->base_term;
366 if (m_index_loc != NULL)
368 m_index_reg = *m_index_loc;
369 lra_eliminate_reg_if_possible (m_index_loc);
373 address_eliminator::~address_eliminator ()
375 if (m_base_loc && *m_base_loc != m_base_reg)
377 *m_base_loc = m_base_reg;
378 if (m_ad->base_term2 != NULL)
379 *m_ad->base_term2 = *m_ad->base_term;
381 if (m_index_loc && *m_index_loc != m_index_reg)
382 *m_index_loc = m_index_reg;
385 /* Return true if the eliminated form of AD is a legitimate target address. */
386 static bool
387 valid_address_p (struct address_info *ad)
389 address_eliminator eliminator (ad);
390 return valid_address_p (ad->mode, *ad->outer, ad->as);
393 /* Return true if the eliminated form of memory reference OP satisfies
394 extra (special) memory constraint CONSTRAINT. */
395 static bool
396 satisfies_memory_constraint_p (rtx op, enum constraint_num constraint)
398 struct address_info ad;
400 decompose_mem_address (&ad, op);
401 address_eliminator eliminator (&ad);
402 return constraint_satisfied_p (op, constraint);
405 /* Return true if the eliminated form of address AD satisfies extra
406 address constraint CONSTRAINT. */
407 static bool
408 satisfies_address_constraint_p (struct address_info *ad,
409 enum constraint_num constraint)
411 address_eliminator eliminator (ad);
412 return constraint_satisfied_p (*ad->outer, constraint);
415 /* Return true if the eliminated form of address OP satisfies extra
416 address constraint CONSTRAINT. */
417 static bool
418 satisfies_address_constraint_p (rtx op, enum constraint_num constraint)
420 struct address_info ad;
422 decompose_lea_address (&ad, &op);
423 return satisfies_address_constraint_p (&ad, constraint);
426 /* Initiate equivalences for LRA. As we keep original equivalences
427 before any elimination, we need to make copies otherwise any change
428 in insns might change the equivalences. */
429 void
430 lra_init_equiv (void)
432 ira_expand_reg_equiv ();
433 for (int i = FIRST_PSEUDO_REGISTER; i < max_reg_num (); i++)
435 rtx res;
437 if ((res = ira_reg_equiv[i].memory) != NULL_RTX)
438 ira_reg_equiv[i].memory = copy_rtx (res);
439 if ((res = ira_reg_equiv[i].invariant) != NULL_RTX)
440 ira_reg_equiv[i].invariant = copy_rtx (res);
444 static rtx loc_equivalence_callback (rtx, const_rtx, void *);
446 /* Update equivalence for REGNO. We need to this as the equivalence
447 might contain other pseudos which are changed by their
448 equivalences. */
449 static void
450 update_equiv (int regno)
452 rtx x;
454 if ((x = ira_reg_equiv[regno].memory) != NULL_RTX)
455 ira_reg_equiv[regno].memory
456 = simplify_replace_fn_rtx (x, NULL_RTX, loc_equivalence_callback,
457 NULL_RTX);
458 if ((x = ira_reg_equiv[regno].invariant) != NULL_RTX)
459 ira_reg_equiv[regno].invariant
460 = simplify_replace_fn_rtx (x, NULL_RTX, loc_equivalence_callback,
461 NULL_RTX);
464 /* If we have decided to substitute X with another value, return that
465 value, otherwise return X. */
466 static rtx
467 get_equiv (rtx x)
469 int regno;
470 rtx res;
472 if (! REG_P (x) || (regno = REGNO (x)) < FIRST_PSEUDO_REGISTER
473 || ! ira_reg_equiv[regno].defined_p
474 || ! ira_reg_equiv[regno].profitable_p
475 || lra_get_regno_hard_regno (regno) >= 0)
476 return x;
477 if ((res = ira_reg_equiv[regno].memory) != NULL_RTX)
479 if (targetm.cannot_substitute_mem_equiv_p (res))
480 return x;
481 return res;
483 if ((res = ira_reg_equiv[regno].constant) != NULL_RTX)
484 return res;
485 if ((res = ira_reg_equiv[regno].invariant) != NULL_RTX)
486 return res;
487 gcc_unreachable ();
490 /* If we have decided to substitute X with the equivalent value,
491 return that value after elimination for INSN, otherwise return
492 X. */
493 static rtx
494 get_equiv_with_elimination (rtx x, rtx_insn *insn)
496 rtx res = get_equiv (x);
498 if (x == res || CONSTANT_P (res))
499 return res;
500 return lra_eliminate_regs_1 (insn, res, GET_MODE (res),
501 false, false, 0, true);
504 /* Set up curr_operand_mode. */
505 static void
506 init_curr_operand_mode (void)
508 int nop = curr_static_id->n_operands;
509 for (int i = 0; i < nop; i++)
511 machine_mode mode = GET_MODE (*curr_id->operand_loc[i]);
512 if (mode == VOIDmode)
514 /* The .md mode for address operands is the mode of the
515 addressed value rather than the mode of the address itself. */
516 if (curr_id->icode >= 0 && curr_static_id->operand[i].is_address)
517 mode = Pmode;
518 else
519 mode = curr_static_id->operand[i].mode;
521 curr_operand_mode[i] = mode;
527 /* The page contains code to reuse input reloads. */
529 /* Structure describes input reload of the current insns. */
530 struct input_reload
532 /* True for input reload of matched operands. */
533 bool match_p;
534 /* Reloaded value. */
535 rtx input;
536 /* Reload pseudo used. */
537 rtx reg;
540 /* The number of elements in the following array. */
541 static int curr_insn_input_reloads_num;
542 /* Array containing info about input reloads. It is used to find the
543 same input reload and reuse the reload pseudo in this case. */
544 static struct input_reload curr_insn_input_reloads[LRA_MAX_INSN_RELOADS];
546 /* Initiate data concerning reuse of input reloads for the current
547 insn. */
548 static void
549 init_curr_insn_input_reloads (void)
551 curr_insn_input_reloads_num = 0;
554 /* Create a new pseudo using MODE, RCLASS, ORIGINAL or reuse already
555 created input reload pseudo (only if TYPE is not OP_OUT). Don't
556 reuse pseudo if IN_SUBREG_P is true and the reused pseudo should be
557 wrapped up in SUBREG. The result pseudo is returned through
558 RESULT_REG. Return TRUE if we created a new pseudo, FALSE if we
559 reused the already created input reload pseudo. Use TITLE to
560 describe new registers for debug purposes. */
561 static bool
562 get_reload_reg (enum op_type type, machine_mode mode, rtx original,
563 enum reg_class rclass, bool in_subreg_p,
564 const char *title, rtx *result_reg)
566 int i, regno;
567 enum reg_class new_class;
568 bool unique_p = false;
570 if (type == OP_OUT)
572 *result_reg
573 = lra_create_new_reg_with_unique_value (mode, original, rclass, title);
574 return true;
576 /* Prevent reuse value of expression with side effects,
577 e.g. volatile memory. */
578 if (! side_effects_p (original))
579 for (i = 0; i < curr_insn_input_reloads_num; i++)
581 if (! curr_insn_input_reloads[i].match_p
582 && rtx_equal_p (curr_insn_input_reloads[i].input, original)
583 && in_class_p (curr_insn_input_reloads[i].reg, rclass, &new_class))
585 rtx reg = curr_insn_input_reloads[i].reg;
586 regno = REGNO (reg);
587 /* If input is equal to original and both are VOIDmode,
588 GET_MODE (reg) might be still different from mode.
589 Ensure we don't return *result_reg with wrong mode. */
590 if (GET_MODE (reg) != mode)
592 if (in_subreg_p)
593 continue;
594 if (GET_MODE_SIZE (GET_MODE (reg)) < GET_MODE_SIZE (mode))
595 continue;
596 reg = lowpart_subreg (mode, reg, GET_MODE (reg));
597 if (reg == NULL_RTX || GET_CODE (reg) != SUBREG)
598 continue;
600 *result_reg = reg;
601 if (lra_dump_file != NULL)
603 fprintf (lra_dump_file, " Reuse r%d for reload ", regno);
604 dump_value_slim (lra_dump_file, original, 1);
606 if (new_class != lra_get_allocno_class (regno))
607 lra_change_class (regno, new_class, ", change to", false);
608 if (lra_dump_file != NULL)
609 fprintf (lra_dump_file, "\n");
610 return false;
612 /* If we have an input reload with a different mode, make sure it
613 will get a different hard reg. */
614 else if (REG_P (original)
615 && REG_P (curr_insn_input_reloads[i].input)
616 && REGNO (original) == REGNO (curr_insn_input_reloads[i].input)
617 && (GET_MODE (original)
618 != GET_MODE (curr_insn_input_reloads[i].input)))
619 unique_p = true;
621 *result_reg = (unique_p
622 ? lra_create_new_reg_with_unique_value
623 : lra_create_new_reg) (mode, original, rclass, title);
624 lra_assert (curr_insn_input_reloads_num < LRA_MAX_INSN_RELOADS);
625 curr_insn_input_reloads[curr_insn_input_reloads_num].input = original;
626 curr_insn_input_reloads[curr_insn_input_reloads_num].match_p = false;
627 curr_insn_input_reloads[curr_insn_input_reloads_num++].reg = *result_reg;
628 return true;
633 /* The page contains code to extract memory address parts. */
635 /* Wrapper around REGNO_OK_FOR_INDEX_P, to allow pseudos. */
636 static inline bool
637 ok_for_index_p_nonstrict (rtx reg)
639 unsigned regno = REGNO (reg);
641 return regno >= FIRST_PSEUDO_REGISTER || REGNO_OK_FOR_INDEX_P (regno);
644 /* A version of regno_ok_for_base_p for use here, when all pseudos
645 should count as OK. Arguments as for regno_ok_for_base_p. */
646 static inline bool
647 ok_for_base_p_nonstrict (rtx reg, machine_mode mode, addr_space_t as,
648 enum rtx_code outer_code, enum rtx_code index_code)
650 unsigned regno = REGNO (reg);
652 if (regno >= FIRST_PSEUDO_REGISTER)
653 return true;
654 return ok_for_base_p_1 (regno, mode, as, outer_code, index_code);
659 /* The page contains major code to choose the current insn alternative
660 and generate reloads for it. */
662 /* Return the offset from REGNO of the least significant register
663 in (reg:MODE REGNO).
665 This function is used to tell whether two registers satisfy
666 a matching constraint. (reg:MODE1 REGNO1) matches (reg:MODE2 REGNO2) if:
668 REGNO1 + lra_constraint_offset (REGNO1, MODE1)
669 == REGNO2 + lra_constraint_offset (REGNO2, MODE2) */
671 lra_constraint_offset (int regno, machine_mode mode)
673 lra_assert (regno < FIRST_PSEUDO_REGISTER);
675 scalar_int_mode int_mode;
676 if (WORDS_BIG_ENDIAN
677 && is_a <scalar_int_mode> (mode, &int_mode)
678 && GET_MODE_SIZE (int_mode) > UNITS_PER_WORD)
679 return hard_regno_nregs (regno, mode) - 1;
680 return 0;
683 /* Like rtx_equal_p except that it allows a REG and a SUBREG to match
684 if they are the same hard reg, and has special hacks for
685 auto-increment and auto-decrement. This is specifically intended for
686 process_alt_operands to use in determining whether two operands
687 match. X is the operand whose number is the lower of the two.
689 It is supposed that X is the output operand and Y is the input
690 operand. Y_HARD_REGNO is the final hard regno of register Y or
691 register in subreg Y as we know it now. Otherwise, it is a
692 negative value. */
693 static bool
694 operands_match_p (rtx x, rtx y, int y_hard_regno)
696 int i;
697 RTX_CODE code = GET_CODE (x);
698 const char *fmt;
700 if (x == y)
701 return true;
702 if ((code == REG || (code == SUBREG && REG_P (SUBREG_REG (x))))
703 && (REG_P (y) || (GET_CODE (y) == SUBREG && REG_P (SUBREG_REG (y)))))
705 int j;
707 i = get_hard_regno (x, false);
708 if (i < 0)
709 goto slow;
711 if ((j = y_hard_regno) < 0)
712 goto slow;
714 i += lra_constraint_offset (i, GET_MODE (x));
715 j += lra_constraint_offset (j, GET_MODE (y));
717 return i == j;
720 /* If two operands must match, because they are really a single
721 operand of an assembler insn, then two post-increments are invalid
722 because the assembler insn would increment only once. On the
723 other hand, a post-increment matches ordinary indexing if the
724 post-increment is the output operand. */
725 if (code == POST_DEC || code == POST_INC || code == POST_MODIFY)
726 return operands_match_p (XEXP (x, 0), y, y_hard_regno);
728 /* Two pre-increments are invalid because the assembler insn would
729 increment only once. On the other hand, a pre-increment matches
730 ordinary indexing if the pre-increment is the input operand. */
731 if (GET_CODE (y) == PRE_DEC || GET_CODE (y) == PRE_INC
732 || GET_CODE (y) == PRE_MODIFY)
733 return operands_match_p (x, XEXP (y, 0), -1);
735 slow:
737 if (code == REG && REG_P (y))
738 return REGNO (x) == REGNO (y);
740 if (code == REG && GET_CODE (y) == SUBREG && REG_P (SUBREG_REG (y))
741 && x == SUBREG_REG (y))
742 return true;
743 if (GET_CODE (y) == REG && code == SUBREG && REG_P (SUBREG_REG (x))
744 && SUBREG_REG (x) == y)
745 return true;
747 /* Now we have disposed of all the cases in which different rtx
748 codes can match. */
749 if (code != GET_CODE (y))
750 return false;
752 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
753 if (GET_MODE (x) != GET_MODE (y))
754 return false;
756 switch (code)
758 CASE_CONST_UNIQUE:
759 return false;
761 case LABEL_REF:
762 return label_ref_label (x) == label_ref_label (y);
763 case SYMBOL_REF:
764 return XSTR (x, 0) == XSTR (y, 0);
766 default:
767 break;
770 /* Compare the elements. If any pair of corresponding elements fail
771 to match, return false for the whole things. */
773 fmt = GET_RTX_FORMAT (code);
774 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
776 int val, j;
777 switch (fmt[i])
779 case 'w':
780 if (XWINT (x, i) != XWINT (y, i))
781 return false;
782 break;
784 case 'i':
785 if (XINT (x, i) != XINT (y, i))
786 return false;
787 break;
789 case 'e':
790 val = operands_match_p (XEXP (x, i), XEXP (y, i), -1);
791 if (val == 0)
792 return false;
793 break;
795 case '0':
796 break;
798 case 'E':
799 if (XVECLEN (x, i) != XVECLEN (y, i))
800 return false;
801 for (j = XVECLEN (x, i) - 1; j >= 0; --j)
803 val = operands_match_p (XVECEXP (x, i, j), XVECEXP (y, i, j), -1);
804 if (val == 0)
805 return false;
807 break;
809 /* It is believed that rtx's at this level will never
810 contain anything but integers and other rtx's, except for
811 within LABEL_REFs and SYMBOL_REFs. */
812 default:
813 gcc_unreachable ();
816 return true;
819 /* True if X is a constant that can be forced into the constant pool.
820 MODE is the mode of the operand, or VOIDmode if not known. */
821 #define CONST_POOL_OK_P(MODE, X) \
822 ((MODE) != VOIDmode \
823 && CONSTANT_P (X) \
824 && GET_CODE (X) != HIGH \
825 && !targetm.cannot_force_const_mem (MODE, X))
827 /* True if C is a non-empty register class that has too few registers
828 to be safely used as a reload target class. */
829 #define SMALL_REGISTER_CLASS_P(C) \
830 (ira_class_hard_regs_num [(C)] == 1 \
831 || (ira_class_hard_regs_num [(C)] >= 1 \
832 && targetm.class_likely_spilled_p (C)))
834 /* If REG is a reload pseudo, try to make its class satisfying CL. */
835 static void
836 narrow_reload_pseudo_class (rtx reg, enum reg_class cl)
838 enum reg_class rclass;
840 /* Do not make more accurate class from reloads generated. They are
841 mostly moves with a lot of constraints. Making more accurate
842 class may results in very narrow class and impossibility of find
843 registers for several reloads of one insn. */
844 if (INSN_UID (curr_insn) >= new_insn_uid_start)
845 return;
846 if (GET_CODE (reg) == SUBREG)
847 reg = SUBREG_REG (reg);
848 if (! REG_P (reg) || (int) REGNO (reg) < new_regno_start)
849 return;
850 if (in_class_p (reg, cl, &rclass) && rclass != cl)
851 lra_change_class (REGNO (reg), rclass, " Change to", true);
854 /* Searches X for any reference to a reg with the same value as REGNO,
855 returning the rtx of the reference found if any. Otherwise,
856 returns NULL_RTX. */
857 static rtx
858 regno_val_use_in (unsigned int regno, rtx x)
860 const char *fmt;
861 int i, j;
862 rtx tem;
864 if (REG_P (x) && lra_reg_info[REGNO (x)].val == lra_reg_info[regno].val)
865 return x;
867 fmt = GET_RTX_FORMAT (GET_CODE (x));
868 for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
870 if (fmt[i] == 'e')
872 if ((tem = regno_val_use_in (regno, XEXP (x, i))))
873 return tem;
875 else if (fmt[i] == 'E')
876 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
877 if ((tem = regno_val_use_in (regno , XVECEXP (x, i, j))))
878 return tem;
881 return NULL_RTX;
884 /* Return true if all current insn non-output operands except INS (it
885 has a negaitve end marker) do not use pseudos with the same value
886 as REGNO. */
887 static bool
888 check_conflict_input_operands (int regno, signed char *ins)
890 int in;
891 int n_operands = curr_static_id->n_operands;
893 for (int nop = 0; nop < n_operands; nop++)
894 if (! curr_static_id->operand[nop].is_operator
895 && curr_static_id->operand[nop].type != OP_OUT)
897 for (int i = 0; (in = ins[i]) >= 0; i++)
898 if (in == nop)
899 break;
900 if (in < 0
901 && regno_val_use_in (regno, *curr_id->operand_loc[nop]) != NULL_RTX)
902 return false;
904 return true;
907 /* Generate reloads for matching OUT and INS (array of input operand
908 numbers with end marker -1) with reg class GOAL_CLASS, considering
909 output operands OUTS (similar array to INS) needing to be in different
910 registers. Add input and output reloads correspondingly to the lists
911 *BEFORE and *AFTER. OUT might be negative. In this case we generate
912 input reloads for matched input operands INS. EARLY_CLOBBER_P is a flag
913 that the output operand is early clobbered for chosen alternative. */
914 static void
915 match_reload (signed char out, signed char *ins, signed char *outs,
916 enum reg_class goal_class, rtx_insn **before,
917 rtx_insn **after, bool early_clobber_p)
919 bool out_conflict;
920 int i, in;
921 rtx new_in_reg, new_out_reg, reg;
922 machine_mode inmode, outmode;
923 rtx in_rtx = *curr_id->operand_loc[ins[0]];
924 rtx out_rtx = out < 0 ? in_rtx : *curr_id->operand_loc[out];
926 inmode = curr_operand_mode[ins[0]];
927 outmode = out < 0 ? inmode : curr_operand_mode[out];
928 push_to_sequence (*before);
929 if (inmode != outmode)
931 if (partial_subreg_p (outmode, inmode))
933 reg = new_in_reg
934 = lra_create_new_reg_with_unique_value (inmode, in_rtx,
935 goal_class, "");
936 if (SCALAR_INT_MODE_P (inmode))
937 new_out_reg = gen_lowpart_SUBREG (outmode, reg);
938 else
939 new_out_reg = gen_rtx_SUBREG (outmode, reg, 0);
940 LRA_SUBREG_P (new_out_reg) = 1;
941 /* If the input reg is dying here, we can use the same hard
942 register for REG and IN_RTX. We do it only for original
943 pseudos as reload pseudos can die although original
944 pseudos still live where reload pseudos dies. */
945 if (REG_P (in_rtx) && (int) REGNO (in_rtx) < lra_new_regno_start
946 && find_regno_note (curr_insn, REG_DEAD, REGNO (in_rtx))
947 && (!early_clobber_p
948 || check_conflict_input_operands(REGNO (in_rtx), ins)))
949 lra_assign_reg_val (REGNO (in_rtx), REGNO (reg));
951 else
953 reg = new_out_reg
954 = lra_create_new_reg_with_unique_value (outmode, out_rtx,
955 goal_class, "");
956 if (SCALAR_INT_MODE_P (outmode))
957 new_in_reg = gen_lowpart_SUBREG (inmode, reg);
958 else
959 new_in_reg = gen_rtx_SUBREG (inmode, reg, 0);
960 /* NEW_IN_REG is non-paradoxical subreg. We don't want
961 NEW_OUT_REG living above. We add clobber clause for
962 this. This is just a temporary clobber. We can remove
963 it at the end of LRA work. */
964 rtx_insn *clobber = emit_clobber (new_out_reg);
965 LRA_TEMP_CLOBBER_P (PATTERN (clobber)) = 1;
966 LRA_SUBREG_P (new_in_reg) = 1;
967 if (GET_CODE (in_rtx) == SUBREG)
969 rtx subreg_reg = SUBREG_REG (in_rtx);
971 /* If SUBREG_REG is dying here and sub-registers IN_RTX
972 and NEW_IN_REG are similar, we can use the same hard
973 register for REG and SUBREG_REG. */
974 if (REG_P (subreg_reg)
975 && (int) REGNO (subreg_reg) < lra_new_regno_start
976 && GET_MODE (subreg_reg) == outmode
977 && SUBREG_BYTE (in_rtx) == SUBREG_BYTE (new_in_reg)
978 && find_regno_note (curr_insn, REG_DEAD, REGNO (subreg_reg))
979 && (! early_clobber_p
980 || check_conflict_input_operands (REGNO (subreg_reg),
981 ins)))
982 lra_assign_reg_val (REGNO (subreg_reg), REGNO (reg));
986 else
988 /* Pseudos have values -- see comments for lra_reg_info.
989 Different pseudos with the same value do not conflict even if
990 they live in the same place. When we create a pseudo we
991 assign value of original pseudo (if any) from which we
992 created the new pseudo. If we create the pseudo from the
993 input pseudo, the new pseudo will have no conflict with the
994 input pseudo which is wrong when the input pseudo lives after
995 the insn and as the new pseudo value is changed by the insn
996 output. Therefore we create the new pseudo from the output
997 except the case when we have single matched dying input
998 pseudo.
1000 We cannot reuse the current output register because we might
1001 have a situation like "a <- a op b", where the constraints
1002 force the second input operand ("b") to match the output
1003 operand ("a"). "b" must then be copied into a new register
1004 so that it doesn't clobber the current value of "a".
1006 We can not use the same value if the output pseudo is
1007 early clobbered or the input pseudo is mentioned in the
1008 output, e.g. as an address part in memory, because
1009 output reload will actually extend the pseudo liveness.
1010 We don't care about eliminable hard regs here as we are
1011 interesting only in pseudos. */
1013 /* Matching input's register value is the same as one of the other
1014 output operand. Output operands in a parallel insn must be in
1015 different registers. */
1016 out_conflict = false;
1017 if (REG_P (in_rtx))
1019 for (i = 0; outs[i] >= 0; i++)
1021 rtx other_out_rtx = *curr_id->operand_loc[outs[i]];
1022 if (REG_P (other_out_rtx)
1023 && (regno_val_use_in (REGNO (in_rtx), other_out_rtx)
1024 != NULL_RTX))
1026 out_conflict = true;
1027 break;
1032 new_in_reg = new_out_reg
1033 = (! early_clobber_p && ins[1] < 0 && REG_P (in_rtx)
1034 && (int) REGNO (in_rtx) < lra_new_regno_start
1035 && find_regno_note (curr_insn, REG_DEAD, REGNO (in_rtx))
1036 && (! early_clobber_p
1037 || check_conflict_input_operands (REGNO (in_rtx), ins))
1038 && (out < 0
1039 || regno_val_use_in (REGNO (in_rtx), out_rtx) == NULL_RTX)
1040 && !out_conflict
1041 ? lra_create_new_reg (inmode, in_rtx, goal_class, "")
1042 : lra_create_new_reg_with_unique_value (outmode, out_rtx,
1043 goal_class, ""));
1045 /* In operand can be got from transformations before processing insn
1046 constraints. One example of such transformations is subreg
1047 reloading (see function simplify_operand_subreg). The new
1048 pseudos created by the transformations might have inaccurate
1049 class (ALL_REGS) and we should make their classes more
1050 accurate. */
1051 narrow_reload_pseudo_class (in_rtx, goal_class);
1052 lra_emit_move (copy_rtx (new_in_reg), in_rtx);
1053 *before = get_insns ();
1054 end_sequence ();
1055 /* Add the new pseudo to consider values of subsequent input reload
1056 pseudos. */
1057 lra_assert (curr_insn_input_reloads_num < LRA_MAX_INSN_RELOADS);
1058 curr_insn_input_reloads[curr_insn_input_reloads_num].input = in_rtx;
1059 curr_insn_input_reloads[curr_insn_input_reloads_num].match_p = true;
1060 curr_insn_input_reloads[curr_insn_input_reloads_num++].reg = new_in_reg;
1061 for (i = 0; (in = ins[i]) >= 0; i++)
1063 lra_assert
1064 (GET_MODE (*curr_id->operand_loc[in]) == VOIDmode
1065 || GET_MODE (new_in_reg) == GET_MODE (*curr_id->operand_loc[in]));
1066 *curr_id->operand_loc[in] = new_in_reg;
1068 lra_update_dups (curr_id, ins);
1069 if (out < 0)
1070 return;
1071 /* See a comment for the input operand above. */
1072 narrow_reload_pseudo_class (out_rtx, goal_class);
1073 if (find_reg_note (curr_insn, REG_UNUSED, out_rtx) == NULL_RTX)
1075 start_sequence ();
1076 lra_emit_move (out_rtx, copy_rtx (new_out_reg));
1077 emit_insn (*after);
1078 *after = get_insns ();
1079 end_sequence ();
1081 *curr_id->operand_loc[out] = new_out_reg;
1082 lra_update_dup (curr_id, out);
1085 /* Return register class which is union of all reg classes in insn
1086 constraint alternative string starting with P. */
1087 static enum reg_class
1088 reg_class_from_constraints (const char *p)
1090 int c, len;
1091 enum reg_class op_class = NO_REGS;
1094 switch ((c = *p, len = CONSTRAINT_LEN (c, p)), c)
1096 case '#':
1097 case ',':
1098 return op_class;
1100 case 'g':
1101 op_class = reg_class_subunion[op_class][GENERAL_REGS];
1102 break;
1104 default:
1105 enum constraint_num cn = lookup_constraint (p);
1106 enum reg_class cl = reg_class_for_constraint (cn);
1107 if (cl == NO_REGS)
1109 if (insn_extra_address_constraint (cn))
1110 op_class
1111 = (reg_class_subunion
1112 [op_class][base_reg_class (VOIDmode, ADDR_SPACE_GENERIC,
1113 ADDRESS, SCRATCH)]);
1114 break;
1117 op_class = reg_class_subunion[op_class][cl];
1118 break;
1120 while ((p += len), c);
1121 return op_class;
1124 /* If OP is a register, return the class of the register as per
1125 get_reg_class, otherwise return NO_REGS. */
1126 static inline enum reg_class
1127 get_op_class (rtx op)
1129 return REG_P (op) ? get_reg_class (REGNO (op)) : NO_REGS;
1132 /* Return generated insn mem_pseudo:=val if TO_P or val:=mem_pseudo
1133 otherwise. If modes of MEM_PSEUDO and VAL are different, use
1134 SUBREG for VAL to make them equal. */
1135 static rtx_insn *
1136 emit_spill_move (bool to_p, rtx mem_pseudo, rtx val)
1138 if (GET_MODE (mem_pseudo) != GET_MODE (val))
1140 /* Usually size of mem_pseudo is greater than val size but in
1141 rare cases it can be less as it can be defined by target
1142 dependent macro HARD_REGNO_CALLER_SAVE_MODE. */
1143 if (! MEM_P (val))
1145 val = gen_lowpart_SUBREG (GET_MODE (mem_pseudo),
1146 GET_CODE (val) == SUBREG
1147 ? SUBREG_REG (val) : val);
1148 LRA_SUBREG_P (val) = 1;
1150 else
1152 mem_pseudo = gen_lowpart_SUBREG (GET_MODE (val), mem_pseudo);
1153 LRA_SUBREG_P (mem_pseudo) = 1;
1156 return to_p ? gen_move_insn (mem_pseudo, val)
1157 : gen_move_insn (val, mem_pseudo);
1160 /* Process a special case insn (register move), return true if we
1161 don't need to process it anymore. INSN should be a single set
1162 insn. Set up that RTL was changed through CHANGE_P and that hook
1163 TARGET_SECONDARY_MEMORY_NEEDED says to use secondary memory through
1164 SEC_MEM_P. */
1165 static bool
1166 check_and_process_move (bool *change_p, bool *sec_mem_p ATTRIBUTE_UNUSED)
1168 int sregno, dregno;
1169 rtx dest, src, dreg, sreg, new_reg, scratch_reg;
1170 rtx_insn *before;
1171 enum reg_class dclass, sclass, secondary_class;
1172 secondary_reload_info sri;
1174 lra_assert (curr_insn_set != NULL_RTX);
1175 dreg = dest = SET_DEST (curr_insn_set);
1176 sreg = src = SET_SRC (curr_insn_set);
1177 if (GET_CODE (dest) == SUBREG)
1178 dreg = SUBREG_REG (dest);
1179 if (GET_CODE (src) == SUBREG)
1180 sreg = SUBREG_REG (src);
1181 if (! (REG_P (dreg) || MEM_P (dreg)) || ! (REG_P (sreg) || MEM_P (sreg)))
1182 return false;
1183 sclass = dclass = NO_REGS;
1184 if (REG_P (dreg))
1185 dclass = get_reg_class (REGNO (dreg));
1186 gcc_assert (dclass < LIM_REG_CLASSES);
1187 if (dclass == ALL_REGS)
1188 /* ALL_REGS is used for new pseudos created by transformations
1189 like reload of SUBREG_REG (see function
1190 simplify_operand_subreg). We don't know their class yet. We
1191 should figure out the class from processing the insn
1192 constraints not in this fast path function. Even if ALL_REGS
1193 were a right class for the pseudo, secondary_... hooks usually
1194 are not define for ALL_REGS. */
1195 return false;
1196 if (REG_P (sreg))
1197 sclass = get_reg_class (REGNO (sreg));
1198 gcc_assert (sclass < LIM_REG_CLASSES);
1199 if (sclass == ALL_REGS)
1200 /* See comments above. */
1201 return false;
1202 if (sclass == NO_REGS && dclass == NO_REGS)
1203 return false;
1204 if (targetm.secondary_memory_needed (GET_MODE (src), sclass, dclass)
1205 && ((sclass != NO_REGS && dclass != NO_REGS)
1206 || (GET_MODE (src)
1207 != targetm.secondary_memory_needed_mode (GET_MODE (src)))))
1209 *sec_mem_p = true;
1210 return false;
1212 if (! REG_P (dreg) || ! REG_P (sreg))
1213 return false;
1214 sri.prev_sri = NULL;
1215 sri.icode = CODE_FOR_nothing;
1216 sri.extra_cost = 0;
1217 secondary_class = NO_REGS;
1218 /* Set up hard register for a reload pseudo for hook
1219 secondary_reload because some targets just ignore unassigned
1220 pseudos in the hook. */
1221 if (dclass != NO_REGS && lra_get_regno_hard_regno (REGNO (dreg)) < 0)
1223 dregno = REGNO (dreg);
1224 reg_renumber[dregno] = ira_class_hard_regs[dclass][0];
1226 else
1227 dregno = -1;
1228 if (sclass != NO_REGS && lra_get_regno_hard_regno (REGNO (sreg)) < 0)
1230 sregno = REGNO (sreg);
1231 reg_renumber[sregno] = ira_class_hard_regs[sclass][0];
1233 else
1234 sregno = -1;
1235 if (sclass != NO_REGS)
1236 secondary_class
1237 = (enum reg_class) targetm.secondary_reload (false, dest,
1238 (reg_class_t) sclass,
1239 GET_MODE (src), &sri);
1240 if (sclass == NO_REGS
1241 || ((secondary_class != NO_REGS || sri.icode != CODE_FOR_nothing)
1242 && dclass != NO_REGS))
1244 enum reg_class old_sclass = secondary_class;
1245 secondary_reload_info old_sri = sri;
1247 sri.prev_sri = NULL;
1248 sri.icode = CODE_FOR_nothing;
1249 sri.extra_cost = 0;
1250 secondary_class
1251 = (enum reg_class) targetm.secondary_reload (true, src,
1252 (reg_class_t) dclass,
1253 GET_MODE (src), &sri);
1254 /* Check the target hook consistency. */
1255 lra_assert
1256 ((secondary_class == NO_REGS && sri.icode == CODE_FOR_nothing)
1257 || (old_sclass == NO_REGS && old_sri.icode == CODE_FOR_nothing)
1258 || (secondary_class == old_sclass && sri.icode == old_sri.icode));
1260 if (sregno >= 0)
1261 reg_renumber [sregno] = -1;
1262 if (dregno >= 0)
1263 reg_renumber [dregno] = -1;
1264 if (secondary_class == NO_REGS && sri.icode == CODE_FOR_nothing)
1265 return false;
1266 *change_p = true;
1267 new_reg = NULL_RTX;
1268 if (secondary_class != NO_REGS)
1269 new_reg = lra_create_new_reg_with_unique_value (GET_MODE (src), NULL_RTX,
1270 secondary_class,
1271 "secondary");
1272 start_sequence ();
1273 if (sri.icode == CODE_FOR_nothing)
1274 lra_emit_move (new_reg, src);
1275 else
1277 enum reg_class scratch_class;
1279 scratch_class = (reg_class_from_constraints
1280 (insn_data[sri.icode].operand[2].constraint));
1281 scratch_reg = (lra_create_new_reg_with_unique_value
1282 (insn_data[sri.icode].operand[2].mode, NULL_RTX,
1283 scratch_class, "scratch"));
1284 emit_insn (GEN_FCN (sri.icode) (new_reg != NULL_RTX ? new_reg : dest,
1285 src, scratch_reg));
1287 before = get_insns ();
1288 end_sequence ();
1289 lra_process_new_insns (curr_insn, before, NULL, "Inserting the move");
1290 if (new_reg != NULL_RTX)
1291 SET_SRC (curr_insn_set) = new_reg;
1292 else
1294 if (lra_dump_file != NULL)
1296 fprintf (lra_dump_file, "Deleting move %u\n", INSN_UID (curr_insn));
1297 dump_insn_slim (lra_dump_file, curr_insn);
1299 lra_set_insn_deleted (curr_insn);
1300 return true;
1302 return false;
1305 /* The following data describe the result of process_alt_operands.
1306 The data are used in curr_insn_transform to generate reloads. */
1308 /* The chosen reg classes which should be used for the corresponding
1309 operands. */
1310 static enum reg_class goal_alt[MAX_RECOG_OPERANDS];
1311 /* True if the operand should be the same as another operand and that
1312 other operand does not need a reload. */
1313 static bool goal_alt_match_win[MAX_RECOG_OPERANDS];
1314 /* True if the operand does not need a reload. */
1315 static bool goal_alt_win[MAX_RECOG_OPERANDS];
1316 /* True if the operand can be offsetable memory. */
1317 static bool goal_alt_offmemok[MAX_RECOG_OPERANDS];
1318 /* The number of an operand to which given operand can be matched to. */
1319 static int goal_alt_matches[MAX_RECOG_OPERANDS];
1320 /* The number of elements in the following array. */
1321 static int goal_alt_dont_inherit_ops_num;
1322 /* Numbers of operands whose reload pseudos should not be inherited. */
1323 static int goal_alt_dont_inherit_ops[MAX_RECOG_OPERANDS];
1324 /* True if the insn commutative operands should be swapped. */
1325 static bool goal_alt_swapped;
1326 /* The chosen insn alternative. */
1327 static int goal_alt_number;
1329 /* True if the corresponding operand is the result of an equivalence
1330 substitution. */
1331 static bool equiv_substition_p[MAX_RECOG_OPERANDS];
1333 /* The following five variables are used to choose the best insn
1334 alternative. They reflect final characteristics of the best
1335 alternative. */
1337 /* Number of necessary reloads and overall cost reflecting the
1338 previous value and other unpleasantness of the best alternative. */
1339 static int best_losers, best_overall;
1340 /* Overall number hard registers used for reloads. For example, on
1341 some targets we need 2 general registers to reload DFmode and only
1342 one floating point register. */
1343 static int best_reload_nregs;
1344 /* Overall number reflecting distances of previous reloading the same
1345 value. The distances are counted from the current BB start. It is
1346 used to improve inheritance chances. */
1347 static int best_reload_sum;
1349 /* True if the current insn should have no correspondingly input or
1350 output reloads. */
1351 static bool no_input_reloads_p, no_output_reloads_p;
1353 /* True if we swapped the commutative operands in the current
1354 insn. */
1355 static int curr_swapped;
1357 /* if CHECK_ONLY_P is false, arrange for address element *LOC to be a
1358 register of class CL. Add any input reloads to list BEFORE. AFTER
1359 is nonnull if *LOC is an automodified value; handle that case by
1360 adding the required output reloads to list AFTER. Return true if
1361 the RTL was changed.
1363 if CHECK_ONLY_P is true, check that the *LOC is a correct address
1364 register. Return false if the address register is correct. */
1365 static bool
1366 process_addr_reg (rtx *loc, bool check_only_p, rtx_insn **before, rtx_insn **after,
1367 enum reg_class cl)
1369 int regno;
1370 enum reg_class rclass, new_class;
1371 rtx reg;
1372 rtx new_reg;
1373 machine_mode mode;
1374 bool subreg_p, before_p = false;
1376 subreg_p = GET_CODE (*loc) == SUBREG;
1377 if (subreg_p)
1379 reg = SUBREG_REG (*loc);
1380 mode = GET_MODE (reg);
1382 /* For mode with size bigger than ptr_mode, there unlikely to be "mov"
1383 between two registers with different classes, but there normally will
1384 be "mov" which transfers element of vector register into the general
1385 register, and this normally will be a subreg which should be reloaded
1386 as a whole. This is particularly likely to be triggered when
1387 -fno-split-wide-types specified. */
1388 if (!REG_P (reg)
1389 || in_class_p (reg, cl, &new_class)
1390 || GET_MODE_SIZE (mode) <= GET_MODE_SIZE (ptr_mode))
1391 loc = &SUBREG_REG (*loc);
1394 reg = *loc;
1395 mode = GET_MODE (reg);
1396 if (! REG_P (reg))
1398 if (check_only_p)
1399 return true;
1400 /* Always reload memory in an address even if the target supports
1401 such addresses. */
1402 new_reg = lra_create_new_reg_with_unique_value (mode, reg, cl, "address");
1403 before_p = true;
1405 else
1407 regno = REGNO (reg);
1408 rclass = get_reg_class (regno);
1409 if (! check_only_p
1410 && (*loc = get_equiv_with_elimination (reg, curr_insn)) != reg)
1412 if (lra_dump_file != NULL)
1414 fprintf (lra_dump_file,
1415 "Changing pseudo %d in address of insn %u on equiv ",
1416 REGNO (reg), INSN_UID (curr_insn));
1417 dump_value_slim (lra_dump_file, *loc, 1);
1418 fprintf (lra_dump_file, "\n");
1420 *loc = copy_rtx (*loc);
1422 if (*loc != reg || ! in_class_p (reg, cl, &new_class))
1424 if (check_only_p)
1425 return true;
1426 reg = *loc;
1427 if (get_reload_reg (after == NULL ? OP_IN : OP_INOUT,
1428 mode, reg, cl, subreg_p, "address", &new_reg))
1429 before_p = true;
1431 else if (new_class != NO_REGS && rclass != new_class)
1433 if (check_only_p)
1434 return true;
1435 lra_change_class (regno, new_class, " Change to", true);
1436 return false;
1438 else
1439 return false;
1441 if (before_p)
1443 push_to_sequence (*before);
1444 lra_emit_move (new_reg, reg);
1445 *before = get_insns ();
1446 end_sequence ();
1448 *loc = new_reg;
1449 if (after != NULL)
1451 start_sequence ();
1452 lra_emit_move (before_p ? copy_rtx (reg) : reg, new_reg);
1453 emit_insn (*after);
1454 *after = get_insns ();
1455 end_sequence ();
1457 return true;
1460 /* Insert move insn in simplify_operand_subreg. BEFORE returns
1461 the insn to be inserted before curr insn. AFTER returns the
1462 the insn to be inserted after curr insn. ORIGREG and NEWREG
1463 are the original reg and new reg for reload. */
1464 static void
1465 insert_move_for_subreg (rtx_insn **before, rtx_insn **after, rtx origreg,
1466 rtx newreg)
1468 if (before)
1470 push_to_sequence (*before);
1471 lra_emit_move (newreg, origreg);
1472 *before = get_insns ();
1473 end_sequence ();
1475 if (after)
1477 start_sequence ();
1478 lra_emit_move (origreg, newreg);
1479 emit_insn (*after);
1480 *after = get_insns ();
1481 end_sequence ();
1485 static int valid_address_p (machine_mode mode, rtx addr, addr_space_t as);
1486 static bool process_address (int, bool, rtx_insn **, rtx_insn **);
1488 /* Make reloads for subreg in operand NOP with internal subreg mode
1489 REG_MODE, add new reloads for further processing. Return true if
1490 any change was done. */
1491 static bool
1492 simplify_operand_subreg (int nop, machine_mode reg_mode)
1494 int hard_regno;
1495 rtx_insn *before, *after;
1496 machine_mode mode, innermode;
1497 rtx reg, new_reg;
1498 rtx operand = *curr_id->operand_loc[nop];
1499 enum reg_class regclass;
1500 enum op_type type;
1502 before = after = NULL;
1504 if (GET_CODE (operand) != SUBREG)
1505 return false;
1507 mode = GET_MODE (operand);
1508 reg = SUBREG_REG (operand);
1509 innermode = GET_MODE (reg);
1510 type = curr_static_id->operand[nop].type;
1511 if (MEM_P (reg))
1513 const bool addr_was_valid
1514 = valid_address_p (innermode, XEXP (reg, 0), MEM_ADDR_SPACE (reg));
1515 alter_subreg (curr_id->operand_loc[nop], false);
1516 rtx subst = *curr_id->operand_loc[nop];
1517 lra_assert (MEM_P (subst));
1519 if (!addr_was_valid
1520 || valid_address_p (GET_MODE (subst), XEXP (subst, 0),
1521 MEM_ADDR_SPACE (subst))
1522 || ((get_constraint_type (lookup_constraint
1523 (curr_static_id->operand[nop].constraint))
1524 != CT_SPECIAL_MEMORY)
1525 /* We still can reload address and if the address is
1526 valid, we can remove subreg without reloading its
1527 inner memory. */
1528 && valid_address_p (GET_MODE (subst),
1529 regno_reg_rtx
1530 [ira_class_hard_regs
1531 [base_reg_class (GET_MODE (subst),
1532 MEM_ADDR_SPACE (subst),
1533 ADDRESS, SCRATCH)][0]],
1534 MEM_ADDR_SPACE (subst))))
1536 /* If we change the address for a paradoxical subreg of memory, the
1537 new address might violate the necessary alignment or the access
1538 might be slow; take this into consideration. We need not worry
1539 about accesses beyond allocated memory for paradoxical memory
1540 subregs as we don't substitute such equiv memory (see processing
1541 equivalences in function lra_constraints) and because for spilled
1542 pseudos we allocate stack memory enough for the biggest
1543 corresponding paradoxical subreg.
1545 However, do not blindly simplify a (subreg (mem ...)) for
1546 WORD_REGISTER_OPERATIONS targets as this may lead to loading junk
1547 data into a register when the inner is narrower than outer or
1548 missing important data from memory when the inner is wider than
1549 outer. This rule only applies to modes that are no wider than
1550 a word. */
1551 if (!(GET_MODE_PRECISION (mode) != GET_MODE_PRECISION (innermode)
1552 && GET_MODE_SIZE (mode) <= UNITS_PER_WORD
1553 && GET_MODE_SIZE (innermode) <= UNITS_PER_WORD
1554 && WORD_REGISTER_OPERATIONS)
1555 && (!(MEM_ALIGN (subst) < GET_MODE_ALIGNMENT (mode)
1556 && targetm.slow_unaligned_access (mode, MEM_ALIGN (subst)))
1557 || (MEM_ALIGN (reg) < GET_MODE_ALIGNMENT (innermode)
1558 && targetm.slow_unaligned_access (innermode,
1559 MEM_ALIGN (reg)))))
1560 return true;
1562 *curr_id->operand_loc[nop] = operand;
1564 /* But if the address was not valid, we cannot reload the MEM without
1565 reloading the address first. */
1566 if (!addr_was_valid)
1567 process_address (nop, false, &before, &after);
1569 /* INNERMODE is fast, MODE slow. Reload the mem in INNERMODE. */
1570 enum reg_class rclass
1571 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1572 if (get_reload_reg (curr_static_id->operand[nop].type, innermode,
1573 reg, rclass, TRUE, "slow mem", &new_reg))
1575 bool insert_before, insert_after;
1576 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1578 insert_before = (type != OP_OUT
1579 || partial_subreg_p (mode, innermode));
1580 insert_after = type != OP_IN;
1581 insert_move_for_subreg (insert_before ? &before : NULL,
1582 insert_after ? &after : NULL,
1583 reg, new_reg);
1585 SUBREG_REG (operand) = new_reg;
1587 /* Convert to MODE. */
1588 reg = operand;
1589 rclass
1590 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1591 if (get_reload_reg (curr_static_id->operand[nop].type, mode, reg,
1592 rclass, TRUE, "slow mem", &new_reg))
1594 bool insert_before, insert_after;
1595 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1597 insert_before = type != OP_OUT;
1598 insert_after = type != OP_IN;
1599 insert_move_for_subreg (insert_before ? &before : NULL,
1600 insert_after ? &after : NULL,
1601 reg, new_reg);
1603 *curr_id->operand_loc[nop] = new_reg;
1604 lra_process_new_insns (curr_insn, before, after,
1605 "Inserting slow mem reload");
1606 return true;
1609 /* If the address was valid and became invalid, prefer to reload
1610 the memory. Typical case is when the index scale should
1611 correspond the memory. */
1612 *curr_id->operand_loc[nop] = operand;
1613 /* Do not return false here as the MEM_P (reg) will be processed
1614 later in this function. */
1616 else if (REG_P (reg) && REGNO (reg) < FIRST_PSEUDO_REGISTER)
1618 alter_subreg (curr_id->operand_loc[nop], false);
1619 return true;
1621 else if (CONSTANT_P (reg))
1623 /* Try to simplify subreg of constant. It is usually result of
1624 equivalence substitution. */
1625 if (innermode == VOIDmode
1626 && (innermode = original_subreg_reg_mode[nop]) == VOIDmode)
1627 innermode = curr_static_id->operand[nop].mode;
1628 if ((new_reg = simplify_subreg (mode, reg, innermode,
1629 SUBREG_BYTE (operand))) != NULL_RTX)
1631 *curr_id->operand_loc[nop] = new_reg;
1632 return true;
1635 /* Put constant into memory when we have mixed modes. It generates
1636 a better code in most cases as it does not need a secondary
1637 reload memory. It also prevents LRA looping when LRA is using
1638 secondary reload memory again and again. */
1639 if (CONSTANT_P (reg) && CONST_POOL_OK_P (reg_mode, reg)
1640 && SCALAR_INT_MODE_P (reg_mode) != SCALAR_INT_MODE_P (mode))
1642 SUBREG_REG (operand) = force_const_mem (reg_mode, reg);
1643 alter_subreg (curr_id->operand_loc[nop], false);
1644 return true;
1646 /* Force a reload of the SUBREG_REG if this is a constant or PLUS or
1647 if there may be a problem accessing OPERAND in the outer
1648 mode. */
1649 if ((REG_P (reg)
1650 && REGNO (reg) >= FIRST_PSEUDO_REGISTER
1651 && (hard_regno = lra_get_regno_hard_regno (REGNO (reg))) >= 0
1652 /* Don't reload paradoxical subregs because we could be looping
1653 having repeatedly final regno out of hard regs range. */
1654 && (hard_regno_nregs (hard_regno, innermode)
1655 >= hard_regno_nregs (hard_regno, mode))
1656 && simplify_subreg_regno (hard_regno, innermode,
1657 SUBREG_BYTE (operand), mode) < 0
1658 /* Don't reload subreg for matching reload. It is actually
1659 valid subreg in LRA. */
1660 && ! LRA_SUBREG_P (operand))
1661 || CONSTANT_P (reg) || GET_CODE (reg) == PLUS || MEM_P (reg))
1663 enum reg_class rclass;
1665 if (REG_P (reg))
1666 /* There is a big probability that we will get the same class
1667 for the new pseudo and we will get the same insn which
1668 means infinite looping. So spill the new pseudo. */
1669 rclass = NO_REGS;
1670 else
1671 /* The class will be defined later in curr_insn_transform. */
1672 rclass
1673 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1675 if (get_reload_reg (curr_static_id->operand[nop].type, reg_mode, reg,
1676 rclass, TRUE, "subreg reg", &new_reg))
1678 bool insert_before, insert_after;
1679 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1681 insert_before = (type != OP_OUT
1682 || read_modify_subreg_p (operand));
1683 insert_after = (type != OP_IN);
1684 insert_move_for_subreg (insert_before ? &before : NULL,
1685 insert_after ? &after : NULL,
1686 reg, new_reg);
1688 SUBREG_REG (operand) = new_reg;
1689 lra_process_new_insns (curr_insn, before, after,
1690 "Inserting subreg reload");
1691 return true;
1693 /* Force a reload for a paradoxical subreg. For paradoxical subreg,
1694 IRA allocates hardreg to the inner pseudo reg according to its mode
1695 instead of the outermode, so the size of the hardreg may not be enough
1696 to contain the outermode operand, in that case we may need to insert
1697 reload for the reg. For the following two types of paradoxical subreg,
1698 we need to insert reload:
1699 1. If the op_type is OP_IN, and the hardreg could not be paired with
1700 other hardreg to contain the outermode operand
1701 (checked by in_hard_reg_set_p), we need to insert the reload.
1702 2. If the op_type is OP_OUT or OP_INOUT.
1704 Here is a paradoxical subreg example showing how the reload is generated:
1706 (insn 5 4 7 2 (set (reg:TI 106 [ __comp ])
1707 (subreg:TI (reg:DI 107 [ __comp ]) 0)) {*movti_internal_rex64}
1709 In IRA, reg107 is allocated to a DImode hardreg. We use x86-64 as example
1710 here, if reg107 is assigned to hardreg R15, because R15 is the last
1711 hardreg, compiler cannot find another hardreg to pair with R15 to
1712 contain TImode data. So we insert a TImode reload reg180 for it.
1713 After reload is inserted:
1715 (insn 283 0 0 (set (subreg:DI (reg:TI 180 [orig:107 __comp ] [107]) 0)
1716 (reg:DI 107 [ __comp ])) -1
1717 (insn 5 4 7 2 (set (reg:TI 106 [ __comp ])
1718 (subreg:TI (reg:TI 180 [orig:107 __comp ] [107]) 0)) {*movti_internal_rex64}
1720 Two reload hard registers will be allocated to reg180 to save TImode data
1721 in LRA_assign. */
1722 else if (REG_P (reg)
1723 && REGNO (reg) >= FIRST_PSEUDO_REGISTER
1724 && (hard_regno = lra_get_regno_hard_regno (REGNO (reg))) >= 0
1725 && (hard_regno_nregs (hard_regno, innermode)
1726 < hard_regno_nregs (hard_regno, mode))
1727 && (regclass = lra_get_allocno_class (REGNO (reg)))
1728 && (type != OP_IN
1729 || !in_hard_reg_set_p (reg_class_contents[regclass],
1730 mode, hard_regno)))
1732 /* The class will be defined later in curr_insn_transform. */
1733 enum reg_class rclass
1734 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1736 if (get_reload_reg (curr_static_id->operand[nop].type, mode, reg,
1737 rclass, TRUE, "paradoxical subreg", &new_reg))
1739 rtx subreg;
1740 bool insert_before, insert_after;
1742 PUT_MODE (new_reg, mode);
1743 subreg = gen_lowpart_SUBREG (innermode, new_reg);
1744 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1746 insert_before = (type != OP_OUT);
1747 insert_after = (type != OP_IN);
1748 insert_move_for_subreg (insert_before ? &before : NULL,
1749 insert_after ? &after : NULL,
1750 reg, subreg);
1752 SUBREG_REG (operand) = new_reg;
1753 lra_process_new_insns (curr_insn, before, after,
1754 "Inserting paradoxical subreg reload");
1755 return true;
1757 return false;
1760 /* Return TRUE if X refers for a hard register from SET. */
1761 static bool
1762 uses_hard_regs_p (rtx x, HARD_REG_SET set)
1764 int i, j, x_hard_regno;
1765 machine_mode mode;
1766 const char *fmt;
1767 enum rtx_code code;
1769 if (x == NULL_RTX)
1770 return false;
1771 code = GET_CODE (x);
1772 mode = GET_MODE (x);
1773 if (code == SUBREG)
1775 mode = wider_subreg_mode (x);
1776 x = SUBREG_REG (x);
1777 code = GET_CODE (x);
1780 if (REG_P (x))
1782 x_hard_regno = get_hard_regno (x, true);
1783 return (x_hard_regno >= 0
1784 && overlaps_hard_reg_set_p (set, mode, x_hard_regno));
1786 if (MEM_P (x))
1788 struct address_info ad;
1790 decompose_mem_address (&ad, x);
1791 if (ad.base_term != NULL && uses_hard_regs_p (*ad.base_term, set))
1792 return true;
1793 if (ad.index_term != NULL && uses_hard_regs_p (*ad.index_term, set))
1794 return true;
1796 fmt = GET_RTX_FORMAT (code);
1797 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1799 if (fmt[i] == 'e')
1801 if (uses_hard_regs_p (XEXP (x, i), set))
1802 return true;
1804 else if (fmt[i] == 'E')
1806 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
1807 if (uses_hard_regs_p (XVECEXP (x, i, j), set))
1808 return true;
1811 return false;
1814 /* Return true if OP is a spilled pseudo. */
1815 static inline bool
1816 spilled_pseudo_p (rtx op)
1818 return (REG_P (op)
1819 && REGNO (op) >= FIRST_PSEUDO_REGISTER && in_mem_p (REGNO (op)));
1822 /* Return true if X is a general constant. */
1823 static inline bool
1824 general_constant_p (rtx x)
1826 return CONSTANT_P (x) && (! flag_pic || LEGITIMATE_PIC_OPERAND_P (x));
1829 static bool
1830 reg_in_class_p (rtx reg, enum reg_class cl)
1832 if (cl == NO_REGS)
1833 return get_reg_class (REGNO (reg)) == NO_REGS;
1834 return in_class_p (reg, cl, NULL);
1837 /* Return true if SET of RCLASS contains no hard regs which can be
1838 used in MODE. */
1839 static bool
1840 prohibited_class_reg_set_mode_p (enum reg_class rclass,
1841 HARD_REG_SET &set,
1842 machine_mode mode)
1844 HARD_REG_SET temp;
1846 lra_assert (hard_reg_set_subset_p (reg_class_contents[rclass], set));
1847 COPY_HARD_REG_SET (temp, set);
1848 AND_COMPL_HARD_REG_SET (temp, lra_no_alloc_regs);
1849 return (hard_reg_set_subset_p
1850 (temp, ira_prohibited_class_mode_regs[rclass][mode]));
1854 /* Used to check validity info about small class input operands. It
1855 should be incremented at start of processing an insn
1856 alternative. */
1857 static unsigned int curr_small_class_check = 0;
1859 /* Update number of used inputs of class OP_CLASS for operand NOP.
1860 Return true if we have more such class operands than the number of
1861 available regs. */
1862 static bool
1863 update_and_check_small_class_inputs (int nop, enum reg_class op_class)
1865 static unsigned int small_class_check[LIM_REG_CLASSES];
1866 static int small_class_input_nums[LIM_REG_CLASSES];
1868 if (SMALL_REGISTER_CLASS_P (op_class)
1869 /* We are interesting in classes became small because of fixing
1870 some hard regs, e.g. by an user through GCC options. */
1871 && hard_reg_set_intersect_p (reg_class_contents[op_class],
1872 ira_no_alloc_regs)
1873 && (curr_static_id->operand[nop].type != OP_OUT
1874 || curr_static_id->operand[nop].early_clobber))
1876 if (small_class_check[op_class] == curr_small_class_check)
1877 small_class_input_nums[op_class]++;
1878 else
1880 small_class_check[op_class] = curr_small_class_check;
1881 small_class_input_nums[op_class] = 1;
1883 if (small_class_input_nums[op_class] > ira_class_hard_regs_num[op_class])
1884 return true;
1886 return false;
1889 /* Major function to choose the current insn alternative and what
1890 operands should be reloaded and how. If ONLY_ALTERNATIVE is not
1891 negative we should consider only this alternative. Return false if
1892 we can not choose the alternative or find how to reload the
1893 operands. */
1894 static bool
1895 process_alt_operands (int only_alternative)
1897 bool ok_p = false;
1898 int nop, overall, nalt;
1899 int n_alternatives = curr_static_id->n_alternatives;
1900 int n_operands = curr_static_id->n_operands;
1901 /* LOSERS counts the operands that don't fit this alternative and
1902 would require loading. */
1903 int losers;
1904 int addr_losers;
1905 /* REJECT is a count of how undesirable this alternative says it is
1906 if any reloading is required. If the alternative matches exactly
1907 then REJECT is ignored, but otherwise it gets this much counted
1908 against it in addition to the reloading needed. */
1909 int reject;
1910 /* This is defined by '!' or '?' alternative constraint and added to
1911 reject. But in some cases it can be ignored. */
1912 int static_reject;
1913 int op_reject;
1914 /* The number of elements in the following array. */
1915 int early_clobbered_regs_num;
1916 /* Numbers of operands which are early clobber registers. */
1917 int early_clobbered_nops[MAX_RECOG_OPERANDS];
1918 enum reg_class curr_alt[MAX_RECOG_OPERANDS];
1919 HARD_REG_SET curr_alt_set[MAX_RECOG_OPERANDS];
1920 bool curr_alt_match_win[MAX_RECOG_OPERANDS];
1921 bool curr_alt_win[MAX_RECOG_OPERANDS];
1922 bool curr_alt_offmemok[MAX_RECOG_OPERANDS];
1923 int curr_alt_matches[MAX_RECOG_OPERANDS];
1924 /* The number of elements in the following array. */
1925 int curr_alt_dont_inherit_ops_num;
1926 /* Numbers of operands whose reload pseudos should not be inherited. */
1927 int curr_alt_dont_inherit_ops[MAX_RECOG_OPERANDS];
1928 rtx op;
1929 /* The register when the operand is a subreg of register, otherwise the
1930 operand itself. */
1931 rtx no_subreg_reg_operand[MAX_RECOG_OPERANDS];
1932 /* The register if the operand is a register or subreg of register,
1933 otherwise NULL. */
1934 rtx operand_reg[MAX_RECOG_OPERANDS];
1935 int hard_regno[MAX_RECOG_OPERANDS];
1936 machine_mode biggest_mode[MAX_RECOG_OPERANDS];
1937 int reload_nregs, reload_sum;
1938 bool costly_p;
1939 enum reg_class cl;
1941 /* Calculate some data common for all alternatives to speed up the
1942 function. */
1943 for (nop = 0; nop < n_operands; nop++)
1945 rtx reg;
1947 op = no_subreg_reg_operand[nop] = *curr_id->operand_loc[nop];
1948 /* The real hard regno of the operand after the allocation. */
1949 hard_regno[nop] = get_hard_regno (op, true);
1951 operand_reg[nop] = reg = op;
1952 biggest_mode[nop] = GET_MODE (op);
1953 if (GET_CODE (op) == SUBREG)
1955 biggest_mode[nop] = wider_subreg_mode (op);
1956 operand_reg[nop] = reg = SUBREG_REG (op);
1958 if (! REG_P (reg))
1959 operand_reg[nop] = NULL_RTX;
1960 else if (REGNO (reg) >= FIRST_PSEUDO_REGISTER
1961 || ((int) REGNO (reg)
1962 == lra_get_elimination_hard_regno (REGNO (reg))))
1963 no_subreg_reg_operand[nop] = reg;
1964 else
1965 operand_reg[nop] = no_subreg_reg_operand[nop]
1966 /* Just use natural mode for elimination result. It should
1967 be enough for extra constraints hooks. */
1968 = regno_reg_rtx[hard_regno[nop]];
1971 /* The constraints are made of several alternatives. Each operand's
1972 constraint looks like foo,bar,... with commas separating the
1973 alternatives. The first alternatives for all operands go
1974 together, the second alternatives go together, etc.
1976 First loop over alternatives. */
1977 alternative_mask preferred = curr_id->preferred_alternatives;
1978 if (only_alternative >= 0)
1979 preferred &= ALTERNATIVE_BIT (only_alternative);
1981 for (nalt = 0; nalt < n_alternatives; nalt++)
1983 /* Loop over operands for one constraint alternative. */
1984 if (!TEST_BIT (preferred, nalt))
1985 continue;
1987 curr_small_class_check++;
1988 overall = losers = addr_losers = 0;
1989 static_reject = reject = reload_nregs = reload_sum = 0;
1990 for (nop = 0; nop < n_operands; nop++)
1992 int inc = (curr_static_id
1993 ->operand_alternative[nalt * n_operands + nop].reject);
1994 if (lra_dump_file != NULL && inc != 0)
1995 fprintf (lra_dump_file,
1996 " Staticly defined alt reject+=%d\n", inc);
1997 static_reject += inc;
1999 reject += static_reject;
2000 early_clobbered_regs_num = 0;
2002 for (nop = 0; nop < n_operands; nop++)
2004 const char *p;
2005 char *end;
2006 int len, c, m, i, opalt_num, this_alternative_matches;
2007 bool win, did_match, offmemok, early_clobber_p;
2008 /* false => this operand can be reloaded somehow for this
2009 alternative. */
2010 bool badop;
2011 /* true => this operand can be reloaded if the alternative
2012 allows regs. */
2013 bool winreg;
2014 /* True if a constant forced into memory would be OK for
2015 this operand. */
2016 bool constmemok;
2017 enum reg_class this_alternative, this_costly_alternative;
2018 HARD_REG_SET this_alternative_set, this_costly_alternative_set;
2019 bool this_alternative_match_win, this_alternative_win;
2020 bool this_alternative_offmemok;
2021 bool scratch_p;
2022 machine_mode mode;
2023 enum constraint_num cn;
2025 opalt_num = nalt * n_operands + nop;
2026 if (curr_static_id->operand_alternative[opalt_num].anything_ok)
2028 /* Fast track for no constraints at all. */
2029 curr_alt[nop] = NO_REGS;
2030 CLEAR_HARD_REG_SET (curr_alt_set[nop]);
2031 curr_alt_win[nop] = true;
2032 curr_alt_match_win[nop] = false;
2033 curr_alt_offmemok[nop] = false;
2034 curr_alt_matches[nop] = -1;
2035 continue;
2038 op = no_subreg_reg_operand[nop];
2039 mode = curr_operand_mode[nop];
2041 win = did_match = winreg = offmemok = constmemok = false;
2042 badop = true;
2044 early_clobber_p = false;
2045 p = curr_static_id->operand_alternative[opalt_num].constraint;
2047 this_costly_alternative = this_alternative = NO_REGS;
2048 /* We update set of possible hard regs besides its class
2049 because reg class might be inaccurate. For example,
2050 union of LO_REGS (l), HI_REGS(h), and STACK_REG(k) in ARM
2051 is translated in HI_REGS because classes are merged by
2052 pairs and there is no accurate intermediate class. */
2053 CLEAR_HARD_REG_SET (this_alternative_set);
2054 CLEAR_HARD_REG_SET (this_costly_alternative_set);
2055 this_alternative_win = false;
2056 this_alternative_match_win = false;
2057 this_alternative_offmemok = false;
2058 this_alternative_matches = -1;
2060 /* An empty constraint should be excluded by the fast
2061 track. */
2062 lra_assert (*p != 0 && *p != ',');
2064 op_reject = 0;
2065 /* Scan this alternative's specs for this operand; set WIN
2066 if the operand fits any letter in this alternative.
2067 Otherwise, clear BADOP if this operand could fit some
2068 letter after reloads, or set WINREG if this operand could
2069 fit after reloads provided the constraint allows some
2070 registers. */
2071 costly_p = false;
2074 switch ((c = *p, len = CONSTRAINT_LEN (c, p)), c)
2076 case '\0':
2077 len = 0;
2078 break;
2079 case ',':
2080 c = '\0';
2081 break;
2083 case '&':
2084 early_clobber_p = true;
2085 break;
2087 case '$':
2088 op_reject += LRA_MAX_REJECT;
2089 break;
2090 case '^':
2091 op_reject += LRA_LOSER_COST_FACTOR;
2092 break;
2094 case '#':
2095 /* Ignore rest of this alternative. */
2096 c = '\0';
2097 break;
2099 case '0': case '1': case '2': case '3': case '4':
2100 case '5': case '6': case '7': case '8': case '9':
2102 int m_hregno;
2103 bool match_p;
2105 m = strtoul (p, &end, 10);
2106 p = end;
2107 len = 0;
2108 lra_assert (nop > m);
2110 this_alternative_matches = m;
2111 m_hregno = get_hard_regno (*curr_id->operand_loc[m], false);
2112 /* We are supposed to match a previous operand.
2113 If we do, we win if that one did. If we do
2114 not, count both of the operands as losers.
2115 (This is too conservative, since most of the
2116 time only a single reload insn will be needed
2117 to make the two operands win. As a result,
2118 this alternative may be rejected when it is
2119 actually desirable.) */
2120 match_p = false;
2121 if (operands_match_p (*curr_id->operand_loc[nop],
2122 *curr_id->operand_loc[m], m_hregno))
2124 /* We should reject matching of an early
2125 clobber operand if the matching operand is
2126 not dying in the insn. */
2127 if (! curr_static_id->operand[m].early_clobber
2128 || operand_reg[nop] == NULL_RTX
2129 || (find_regno_note (curr_insn, REG_DEAD,
2130 REGNO (op))
2131 || REGNO (op) == REGNO (operand_reg[m])))
2132 match_p = true;
2134 if (match_p)
2136 /* If we are matching a non-offsettable
2137 address where an offsettable address was
2138 expected, then we must reject this
2139 combination, because we can't reload
2140 it. */
2141 if (curr_alt_offmemok[m]
2142 && MEM_P (*curr_id->operand_loc[m])
2143 && curr_alt[m] == NO_REGS && ! curr_alt_win[m])
2144 continue;
2146 else
2148 /* Operands don't match. Both operands must
2149 allow a reload register, otherwise we
2150 cannot make them match. */
2151 if (curr_alt[m] == NO_REGS)
2152 break;
2153 /* Retroactively mark the operand we had to
2154 match as a loser, if it wasn't already and
2155 it wasn't matched to a register constraint
2156 (e.g it might be matched by memory). */
2157 if (curr_alt_win[m]
2158 && (operand_reg[m] == NULL_RTX
2159 || hard_regno[m] < 0))
2161 losers++;
2162 reload_nregs
2163 += (ira_reg_class_max_nregs[curr_alt[m]]
2164 [GET_MODE (*curr_id->operand_loc[m])]);
2167 /* Prefer matching earlyclobber alternative as
2168 it results in less hard regs required for
2169 the insn than a non-matching earlyclobber
2170 alternative. */
2171 if (curr_static_id->operand[m].early_clobber)
2173 if (lra_dump_file != NULL)
2174 fprintf
2175 (lra_dump_file,
2176 " %d Matching earlyclobber alt:"
2177 " reject--\n",
2178 nop);
2179 reject--;
2181 /* Otherwise we prefer no matching
2182 alternatives because it gives more freedom
2183 in RA. */
2184 else if (operand_reg[nop] == NULL_RTX
2185 || (find_regno_note (curr_insn, REG_DEAD,
2186 REGNO (operand_reg[nop]))
2187 == NULL_RTX))
2189 if (lra_dump_file != NULL)
2190 fprintf
2191 (lra_dump_file,
2192 " %d Matching alt: reject+=2\n",
2193 nop);
2194 reject += 2;
2197 /* If we have to reload this operand and some
2198 previous operand also had to match the same
2199 thing as this operand, we don't know how to do
2200 that. */
2201 if (!match_p || !curr_alt_win[m])
2203 for (i = 0; i < nop; i++)
2204 if (curr_alt_matches[i] == m)
2205 break;
2206 if (i < nop)
2207 break;
2209 else
2210 did_match = true;
2212 /* This can be fixed with reloads if the operand
2213 we are supposed to match can be fixed with
2214 reloads. */
2215 badop = false;
2216 this_alternative = curr_alt[m];
2217 COPY_HARD_REG_SET (this_alternative_set, curr_alt_set[m]);
2218 winreg = this_alternative != NO_REGS;
2219 break;
2222 case 'g':
2223 if (MEM_P (op)
2224 || general_constant_p (op)
2225 || spilled_pseudo_p (op))
2226 win = true;
2227 cl = GENERAL_REGS;
2228 goto reg;
2230 default:
2231 cn = lookup_constraint (p);
2232 switch (get_constraint_type (cn))
2234 case CT_REGISTER:
2235 cl = reg_class_for_constraint (cn);
2236 if (cl != NO_REGS)
2237 goto reg;
2238 break;
2240 case CT_CONST_INT:
2241 if (CONST_INT_P (op)
2242 && insn_const_int_ok_for_constraint (INTVAL (op), cn))
2243 win = true;
2244 break;
2246 case CT_MEMORY:
2247 if (MEM_P (op)
2248 && satisfies_memory_constraint_p (op, cn))
2249 win = true;
2250 else if (spilled_pseudo_p (op))
2251 win = true;
2253 /* If we didn't already win, we can reload constants
2254 via force_const_mem or put the pseudo value into
2255 memory, or make other memory by reloading the
2256 address like for 'o'. */
2257 if (CONST_POOL_OK_P (mode, op)
2258 || MEM_P (op) || REG_P (op)
2259 /* We can restore the equiv insn by a
2260 reload. */
2261 || equiv_substition_p[nop])
2262 badop = false;
2263 constmemok = true;
2264 offmemok = true;
2265 break;
2267 case CT_ADDRESS:
2268 /* If we didn't already win, we can reload the address
2269 into a base register. */
2270 if (satisfies_address_constraint_p (op, cn))
2271 win = true;
2272 cl = base_reg_class (VOIDmode, ADDR_SPACE_GENERIC,
2273 ADDRESS, SCRATCH);
2274 badop = false;
2275 goto reg;
2277 case CT_FIXED_FORM:
2278 if (constraint_satisfied_p (op, cn))
2279 win = true;
2280 break;
2282 case CT_SPECIAL_MEMORY:
2283 if (MEM_P (op)
2284 && satisfies_memory_constraint_p (op, cn))
2285 win = true;
2286 else if (spilled_pseudo_p (op))
2287 win = true;
2288 break;
2290 break;
2292 reg:
2293 this_alternative = reg_class_subunion[this_alternative][cl];
2294 IOR_HARD_REG_SET (this_alternative_set,
2295 reg_class_contents[cl]);
2296 if (costly_p)
2298 this_costly_alternative
2299 = reg_class_subunion[this_costly_alternative][cl];
2300 IOR_HARD_REG_SET (this_costly_alternative_set,
2301 reg_class_contents[cl]);
2303 if (mode == BLKmode)
2304 break;
2305 winreg = true;
2306 if (REG_P (op))
2308 if (hard_regno[nop] >= 0
2309 && in_hard_reg_set_p (this_alternative_set,
2310 mode, hard_regno[nop]))
2311 win = true;
2312 else if (hard_regno[nop] < 0
2313 && in_class_p (op, this_alternative, NULL))
2314 win = true;
2316 break;
2318 if (c != ' ' && c != '\t')
2319 costly_p = c == '*';
2321 while ((p += len), c);
2323 scratch_p = (operand_reg[nop] != NULL_RTX
2324 && lra_former_scratch_p (REGNO (operand_reg[nop])));
2325 /* Record which operands fit this alternative. */
2326 if (win)
2328 this_alternative_win = true;
2329 if (operand_reg[nop] != NULL_RTX)
2331 if (hard_regno[nop] >= 0)
2333 if (in_hard_reg_set_p (this_costly_alternative_set,
2334 mode, hard_regno[nop]))
2336 if (lra_dump_file != NULL)
2337 fprintf (lra_dump_file,
2338 " %d Costly set: reject++\n",
2339 nop);
2340 reject++;
2343 else
2345 /* Prefer won reg to spilled pseudo under other
2346 equal conditions for possibe inheritance. */
2347 if (! scratch_p)
2349 if (lra_dump_file != NULL)
2350 fprintf
2351 (lra_dump_file,
2352 " %d Non pseudo reload: reject++\n",
2353 nop);
2354 reject++;
2356 if (in_class_p (operand_reg[nop],
2357 this_costly_alternative, NULL))
2359 if (lra_dump_file != NULL)
2360 fprintf
2361 (lra_dump_file,
2362 " %d Non pseudo costly reload:"
2363 " reject++\n",
2364 nop);
2365 reject++;
2368 /* We simulate the behavior of old reload here.
2369 Although scratches need hard registers and it
2370 might result in spilling other pseudos, no reload
2371 insns are generated for the scratches. So it
2372 might cost something but probably less than old
2373 reload pass believes. */
2374 if (scratch_p)
2376 if (lra_dump_file != NULL)
2377 fprintf (lra_dump_file,
2378 " %d Scratch win: reject+=2\n",
2379 nop);
2380 reject += 2;
2384 else if (did_match)
2385 this_alternative_match_win = true;
2386 else
2388 int const_to_mem = 0;
2389 bool no_regs_p;
2391 reject += op_reject;
2392 /* Never do output reload of stack pointer. It makes
2393 impossible to do elimination when SP is changed in
2394 RTL. */
2395 if (op == stack_pointer_rtx && ! frame_pointer_needed
2396 && curr_static_id->operand[nop].type != OP_IN)
2397 goto fail;
2399 /* If this alternative asks for a specific reg class, see if there
2400 is at least one allocatable register in that class. */
2401 no_regs_p
2402 = (this_alternative == NO_REGS
2403 || (hard_reg_set_subset_p
2404 (reg_class_contents[this_alternative],
2405 lra_no_alloc_regs)));
2407 /* For asms, verify that the class for this alternative is possible
2408 for the mode that is specified. */
2409 if (!no_regs_p && INSN_CODE (curr_insn) < 0)
2411 int i;
2412 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
2413 if (targetm.hard_regno_mode_ok (i, mode)
2414 && in_hard_reg_set_p (reg_class_contents[this_alternative],
2415 mode, i))
2416 break;
2417 if (i == FIRST_PSEUDO_REGISTER)
2418 winreg = false;
2421 /* If this operand accepts a register, and if the
2422 register class has at least one allocatable register,
2423 then this operand can be reloaded. */
2424 if (winreg && !no_regs_p)
2425 badop = false;
2427 if (badop)
2429 if (lra_dump_file != NULL)
2430 fprintf (lra_dump_file,
2431 " alt=%d: Bad operand -- refuse\n",
2432 nalt);
2433 goto fail;
2436 if (this_alternative != NO_REGS)
2438 HARD_REG_SET available_regs;
2440 COPY_HARD_REG_SET (available_regs,
2441 reg_class_contents[this_alternative]);
2442 AND_COMPL_HARD_REG_SET
2443 (available_regs,
2444 ira_prohibited_class_mode_regs[this_alternative][mode]);
2445 AND_COMPL_HARD_REG_SET (available_regs, lra_no_alloc_regs);
2446 if (hard_reg_set_empty_p (available_regs))
2448 /* There are no hard regs holding a value of given
2449 mode. */
2450 if (offmemok)
2452 this_alternative = NO_REGS;
2453 if (lra_dump_file != NULL)
2454 fprintf (lra_dump_file,
2455 " %d Using memory because of"
2456 " a bad mode: reject+=2\n",
2457 nop);
2458 reject += 2;
2460 else
2462 if (lra_dump_file != NULL)
2463 fprintf (lra_dump_file,
2464 " alt=%d: Wrong mode -- refuse\n",
2465 nalt);
2466 goto fail;
2471 /* If not assigned pseudo has a class which a subset of
2472 required reg class, it is a less costly alternative
2473 as the pseudo still can get a hard reg of necessary
2474 class. */
2475 if (! no_regs_p && REG_P (op) && hard_regno[nop] < 0
2476 && (cl = get_reg_class (REGNO (op))) != NO_REGS
2477 && ira_class_subset_p[this_alternative][cl])
2479 if (lra_dump_file != NULL)
2480 fprintf
2481 (lra_dump_file,
2482 " %d Super set class reg: reject-=3\n", nop);
2483 reject -= 3;
2486 this_alternative_offmemok = offmemok;
2487 if (this_costly_alternative != NO_REGS)
2489 if (lra_dump_file != NULL)
2490 fprintf (lra_dump_file,
2491 " %d Costly loser: reject++\n", nop);
2492 reject++;
2494 /* If the operand is dying, has a matching constraint,
2495 and satisfies constraints of the matched operand
2496 which failed to satisfy the own constraints, most probably
2497 the reload for this operand will be gone. */
2498 if (this_alternative_matches >= 0
2499 && !curr_alt_win[this_alternative_matches]
2500 && REG_P (op)
2501 && find_regno_note (curr_insn, REG_DEAD, REGNO (op))
2502 && (hard_regno[nop] >= 0
2503 ? in_hard_reg_set_p (this_alternative_set,
2504 mode, hard_regno[nop])
2505 : in_class_p (op, this_alternative, NULL)))
2507 if (lra_dump_file != NULL)
2508 fprintf
2509 (lra_dump_file,
2510 " %d Dying matched operand reload: reject++\n",
2511 nop);
2512 reject++;
2514 else
2516 /* Strict_low_part requires to reload the register
2517 not the sub-register. In this case we should
2518 check that a final reload hard reg can hold the
2519 value mode. */
2520 if (curr_static_id->operand[nop].strict_low
2521 && REG_P (op)
2522 && hard_regno[nop] < 0
2523 && GET_CODE (*curr_id->operand_loc[nop]) == SUBREG
2524 && ira_class_hard_regs_num[this_alternative] > 0
2525 && (!targetm.hard_regno_mode_ok
2526 (ira_class_hard_regs[this_alternative][0],
2527 GET_MODE (*curr_id->operand_loc[nop]))))
2529 if (lra_dump_file != NULL)
2530 fprintf
2531 (lra_dump_file,
2532 " alt=%d: Strict low subreg reload -- refuse\n",
2533 nalt);
2534 goto fail;
2536 losers++;
2538 if (operand_reg[nop] != NULL_RTX
2539 /* Output operands and matched input operands are
2540 not inherited. The following conditions do not
2541 exactly describe the previous statement but they
2542 are pretty close. */
2543 && curr_static_id->operand[nop].type != OP_OUT
2544 && (this_alternative_matches < 0
2545 || curr_static_id->operand[nop].type != OP_IN))
2547 int last_reload = (lra_reg_info[ORIGINAL_REGNO
2548 (operand_reg[nop])]
2549 .last_reload);
2551 /* The value of reload_sum has sense only if we
2552 process insns in their order. It happens only on
2553 the first constraints sub-pass when we do most of
2554 reload work. */
2555 if (lra_constraint_iter == 1 && last_reload > bb_reload_num)
2556 reload_sum += last_reload - bb_reload_num;
2558 /* If this is a constant that is reloaded into the
2559 desired class by copying it to memory first, count
2560 that as another reload. This is consistent with
2561 other code and is required to avoid choosing another
2562 alternative when the constant is moved into memory.
2563 Note that the test here is precisely the same as in
2564 the code below that calls force_const_mem. */
2565 if (CONST_POOL_OK_P (mode, op)
2566 && ((targetm.preferred_reload_class
2567 (op, this_alternative) == NO_REGS)
2568 || no_input_reloads_p))
2570 const_to_mem = 1;
2571 if (! no_regs_p)
2572 losers++;
2575 /* Alternative loses if it requires a type of reload not
2576 permitted for this insn. We can always reload
2577 objects with a REG_UNUSED note. */
2578 if ((curr_static_id->operand[nop].type != OP_IN
2579 && no_output_reloads_p
2580 && ! find_reg_note (curr_insn, REG_UNUSED, op))
2581 || (curr_static_id->operand[nop].type != OP_OUT
2582 && no_input_reloads_p && ! const_to_mem)
2583 || (this_alternative_matches >= 0
2584 && (no_input_reloads_p
2585 || (no_output_reloads_p
2586 && (curr_static_id->operand
2587 [this_alternative_matches].type != OP_IN)
2588 && ! find_reg_note (curr_insn, REG_UNUSED,
2589 no_subreg_reg_operand
2590 [this_alternative_matches])))))
2592 if (lra_dump_file != NULL)
2593 fprintf
2594 (lra_dump_file,
2595 " alt=%d: No input/otput reload -- refuse\n",
2596 nalt);
2597 goto fail;
2600 /* Alternative loses if it required class pseudo can not
2601 hold value of required mode. Such insns can be
2602 described by insn definitions with mode iterators. */
2603 if (GET_MODE (*curr_id->operand_loc[nop]) != VOIDmode
2604 && ! hard_reg_set_empty_p (this_alternative_set)
2605 /* It is common practice for constraints to use a
2606 class which does not have actually enough regs to
2607 hold the value (e.g. x86 AREG for mode requiring
2608 more one general reg). Therefore we have 2
2609 conditions to check that the reload pseudo can
2610 not hold the mode value. */
2611 && (!targetm.hard_regno_mode_ok
2612 (ira_class_hard_regs[this_alternative][0],
2613 GET_MODE (*curr_id->operand_loc[nop])))
2614 /* The above condition is not enough as the first
2615 reg in ira_class_hard_regs can be not aligned for
2616 multi-words mode values. */
2617 && (prohibited_class_reg_set_mode_p
2618 (this_alternative, this_alternative_set,
2619 GET_MODE (*curr_id->operand_loc[nop]))))
2621 if (lra_dump_file != NULL)
2622 fprintf (lra_dump_file,
2623 " alt=%d: reload pseudo for op %d "
2624 " can not hold the mode value -- refuse\n",
2625 nalt, nop);
2626 goto fail;
2629 /* Check strong discouragement of reload of non-constant
2630 into class THIS_ALTERNATIVE. */
2631 if (! CONSTANT_P (op) && ! no_regs_p
2632 && (targetm.preferred_reload_class
2633 (op, this_alternative) == NO_REGS
2634 || (curr_static_id->operand[nop].type == OP_OUT
2635 && (targetm.preferred_output_reload_class
2636 (op, this_alternative) == NO_REGS))))
2638 if (lra_dump_file != NULL)
2639 fprintf (lra_dump_file,
2640 " %d Non-prefered reload: reject+=%d\n",
2641 nop, LRA_MAX_REJECT);
2642 reject += LRA_MAX_REJECT;
2645 if (! (MEM_P (op) && offmemok)
2646 && ! (const_to_mem && constmemok))
2648 /* We prefer to reload pseudos over reloading other
2649 things, since such reloads may be able to be
2650 eliminated later. So bump REJECT in other cases.
2651 Don't do this in the case where we are forcing a
2652 constant into memory and it will then win since
2653 we don't want to have a different alternative
2654 match then. */
2655 if (! (REG_P (op) && REGNO (op) >= FIRST_PSEUDO_REGISTER))
2657 if (lra_dump_file != NULL)
2658 fprintf
2659 (lra_dump_file,
2660 " %d Non-pseudo reload: reject+=2\n",
2661 nop);
2662 reject += 2;
2665 if (! no_regs_p)
2666 reload_nregs
2667 += ira_reg_class_max_nregs[this_alternative][mode];
2669 if (SMALL_REGISTER_CLASS_P (this_alternative))
2671 if (lra_dump_file != NULL)
2672 fprintf
2673 (lra_dump_file,
2674 " %d Small class reload: reject+=%d\n",
2675 nop, LRA_LOSER_COST_FACTOR / 2);
2676 reject += LRA_LOSER_COST_FACTOR / 2;
2680 /* We are trying to spill pseudo into memory. It is
2681 usually more costly than moving to a hard register
2682 although it might takes the same number of
2683 reloads.
2685 Non-pseudo spill may happen also. Suppose a target allows both
2686 register and memory in the operand constraint alternatives,
2687 then it's typical that an eliminable register has a substition
2688 of "base + offset" which can either be reloaded by a simple
2689 "new_reg <= base + offset" which will match the register
2690 constraint, or a similar reg addition followed by further spill
2691 to and reload from memory which will match the memory
2692 constraint, but this memory spill will be much more costly
2693 usually.
2695 Code below increases the reject for both pseudo and non-pseudo
2696 spill. */
2697 if (no_regs_p
2698 && !(MEM_P (op) && offmemok)
2699 && !(REG_P (op) && hard_regno[nop] < 0))
2701 if (lra_dump_file != NULL)
2702 fprintf
2703 (lra_dump_file,
2704 " %d Spill %spseudo into memory: reject+=3\n",
2705 nop, REG_P (op) ? "" : "Non-");
2706 reject += 3;
2707 if (VECTOR_MODE_P (mode))
2709 /* Spilling vectors into memory is usually more
2710 costly as they contain big values. */
2711 if (lra_dump_file != NULL)
2712 fprintf
2713 (lra_dump_file,
2714 " %d Spill vector pseudo: reject+=2\n",
2715 nop);
2716 reject += 2;
2720 /* When we use an operand requiring memory in given
2721 alternative, the insn should write *and* read the
2722 value to/from memory it is costly in comparison with
2723 an insn alternative which does not use memory
2724 (e.g. register or immediate operand). We exclude
2725 memory operand for such case as we can satisfy the
2726 memory constraints by reloading address. */
2727 if (no_regs_p && offmemok && !MEM_P (op))
2729 if (lra_dump_file != NULL)
2730 fprintf
2731 (lra_dump_file,
2732 " Using memory insn operand %d: reject+=3\n",
2733 nop);
2734 reject += 3;
2737 /* If reload requires moving value through secondary
2738 memory, it will need one more insn at least. */
2739 if (this_alternative != NO_REGS
2740 && REG_P (op) && (cl = get_reg_class (REGNO (op))) != NO_REGS
2741 && ((curr_static_id->operand[nop].type != OP_OUT
2742 && targetm.secondary_memory_needed (GET_MODE (op), cl,
2743 this_alternative))
2744 || (curr_static_id->operand[nop].type != OP_IN
2745 && (targetm.secondary_memory_needed
2746 (GET_MODE (op), this_alternative, cl)))))
2747 losers++;
2749 /* Input reloads can be inherited more often than output
2750 reloads can be removed, so penalize output
2751 reloads. */
2752 if (!REG_P (op) || curr_static_id->operand[nop].type != OP_IN)
2754 if (lra_dump_file != NULL)
2755 fprintf
2756 (lra_dump_file,
2757 " %d Non input pseudo reload: reject++\n",
2758 nop);
2759 reject++;
2762 if (MEM_P (op) && offmemok)
2763 addr_losers++;
2764 else if (curr_static_id->operand[nop].type == OP_INOUT)
2766 if (lra_dump_file != NULL)
2767 fprintf
2768 (lra_dump_file,
2769 " %d Input/Output reload: reject+=%d\n",
2770 nop, LRA_LOSER_COST_FACTOR);
2771 reject += LRA_LOSER_COST_FACTOR;
2775 if (early_clobber_p && ! scratch_p)
2777 if (lra_dump_file != NULL)
2778 fprintf (lra_dump_file,
2779 " %d Early clobber: reject++\n", nop);
2780 reject++;
2782 /* ??? We check early clobbers after processing all operands
2783 (see loop below) and there we update the costs more.
2784 Should we update the cost (may be approximately) here
2785 because of early clobber register reloads or it is a rare
2786 or non-important thing to be worth to do it. */
2787 overall = (losers * LRA_LOSER_COST_FACTOR + reject
2788 - (addr_losers == losers ? static_reject : 0));
2789 if ((best_losers == 0 || losers != 0) && best_overall < overall)
2791 if (lra_dump_file != NULL)
2792 fprintf (lra_dump_file,
2793 " alt=%d,overall=%d,losers=%d -- refuse\n",
2794 nalt, overall, losers);
2795 goto fail;
2798 if (update_and_check_small_class_inputs (nop, this_alternative))
2800 if (lra_dump_file != NULL)
2801 fprintf (lra_dump_file,
2802 " alt=%d, not enough small class regs -- refuse\n",
2803 nalt);
2804 goto fail;
2806 curr_alt[nop] = this_alternative;
2807 COPY_HARD_REG_SET (curr_alt_set[nop], this_alternative_set);
2808 curr_alt_win[nop] = this_alternative_win;
2809 curr_alt_match_win[nop] = this_alternative_match_win;
2810 curr_alt_offmemok[nop] = this_alternative_offmemok;
2811 curr_alt_matches[nop] = this_alternative_matches;
2813 if (this_alternative_matches >= 0
2814 && !did_match && !this_alternative_win)
2815 curr_alt_win[this_alternative_matches] = false;
2817 if (early_clobber_p && operand_reg[nop] != NULL_RTX)
2818 early_clobbered_nops[early_clobbered_regs_num++] = nop;
2821 if (curr_insn_set != NULL_RTX && n_operands == 2
2822 /* Prevent processing non-move insns. */
2823 && (GET_CODE (SET_SRC (curr_insn_set)) == SUBREG
2824 || SET_SRC (curr_insn_set) == no_subreg_reg_operand[1])
2825 && ((! curr_alt_win[0] && ! curr_alt_win[1]
2826 && REG_P (no_subreg_reg_operand[0])
2827 && REG_P (no_subreg_reg_operand[1])
2828 && (reg_in_class_p (no_subreg_reg_operand[0], curr_alt[1])
2829 || reg_in_class_p (no_subreg_reg_operand[1], curr_alt[0])))
2830 || (! curr_alt_win[0] && curr_alt_win[1]
2831 && REG_P (no_subreg_reg_operand[1])
2832 /* Check that we reload memory not the memory
2833 address. */
2834 && ! (curr_alt_offmemok[0]
2835 && MEM_P (no_subreg_reg_operand[0]))
2836 && reg_in_class_p (no_subreg_reg_operand[1], curr_alt[0]))
2837 || (curr_alt_win[0] && ! curr_alt_win[1]
2838 && REG_P (no_subreg_reg_operand[0])
2839 /* Check that we reload memory not the memory
2840 address. */
2841 && ! (curr_alt_offmemok[1]
2842 && MEM_P (no_subreg_reg_operand[1]))
2843 && reg_in_class_p (no_subreg_reg_operand[0], curr_alt[1])
2844 && (! CONST_POOL_OK_P (curr_operand_mode[1],
2845 no_subreg_reg_operand[1])
2846 || (targetm.preferred_reload_class
2847 (no_subreg_reg_operand[1],
2848 (enum reg_class) curr_alt[1]) != NO_REGS))
2849 /* If it is a result of recent elimination in move
2850 insn we can transform it into an add still by
2851 using this alternative. */
2852 && GET_CODE (no_subreg_reg_operand[1]) != PLUS)))
2854 /* We have a move insn and a new reload insn will be similar
2855 to the current insn. We should avoid such situation as
2856 it results in LRA cycling. */
2857 if (lra_dump_file != NULL)
2858 fprintf (lra_dump_file,
2859 " Cycle danger: overall += LRA_MAX_REJECT\n");
2860 overall += LRA_MAX_REJECT;
2862 ok_p = true;
2863 curr_alt_dont_inherit_ops_num = 0;
2864 for (nop = 0; nop < early_clobbered_regs_num; nop++)
2866 int i, j, clobbered_hard_regno, first_conflict_j, last_conflict_j;
2867 HARD_REG_SET temp_set;
2869 i = early_clobbered_nops[nop];
2870 if ((! curr_alt_win[i] && ! curr_alt_match_win[i])
2871 || hard_regno[i] < 0)
2872 continue;
2873 lra_assert (operand_reg[i] != NULL_RTX);
2874 clobbered_hard_regno = hard_regno[i];
2875 CLEAR_HARD_REG_SET (temp_set);
2876 add_to_hard_reg_set (&temp_set, biggest_mode[i], clobbered_hard_regno);
2877 first_conflict_j = last_conflict_j = -1;
2878 for (j = 0; j < n_operands; j++)
2879 if (j == i
2880 /* We don't want process insides of match_operator and
2881 match_parallel because otherwise we would process
2882 their operands once again generating a wrong
2883 code. */
2884 || curr_static_id->operand[j].is_operator)
2885 continue;
2886 else if ((curr_alt_matches[j] == i && curr_alt_match_win[j])
2887 || (curr_alt_matches[i] == j && curr_alt_match_win[i]))
2888 continue;
2889 /* If we don't reload j-th operand, check conflicts. */
2890 else if ((curr_alt_win[j] || curr_alt_match_win[j])
2891 && uses_hard_regs_p (*curr_id->operand_loc[j], temp_set))
2893 if (first_conflict_j < 0)
2894 first_conflict_j = j;
2895 last_conflict_j = j;
2897 if (last_conflict_j < 0)
2898 continue;
2899 /* If earlyclobber operand conflicts with another
2900 non-matching operand which is actually the same register
2901 as the earlyclobber operand, it is better to reload the
2902 another operand as an operand matching the earlyclobber
2903 operand can be also the same. */
2904 if (first_conflict_j == last_conflict_j
2905 && operand_reg[last_conflict_j] != NULL_RTX
2906 && ! curr_alt_match_win[last_conflict_j]
2907 && REGNO (operand_reg[i]) == REGNO (operand_reg[last_conflict_j]))
2909 curr_alt_win[last_conflict_j] = false;
2910 curr_alt_dont_inherit_ops[curr_alt_dont_inherit_ops_num++]
2911 = last_conflict_j;
2912 losers++;
2913 /* Early clobber was already reflected in REJECT. */
2914 lra_assert (reject > 0);
2915 if (lra_dump_file != NULL)
2916 fprintf
2917 (lra_dump_file,
2918 " %d Conflict early clobber reload: reject--\n",
2920 reject--;
2921 overall += LRA_LOSER_COST_FACTOR - 1;
2923 else
2925 /* We need to reload early clobbered register and the
2926 matched registers. */
2927 for (j = 0; j < n_operands; j++)
2928 if (curr_alt_matches[j] == i)
2930 curr_alt_match_win[j] = false;
2931 losers++;
2932 overall += LRA_LOSER_COST_FACTOR;
2934 if (! curr_alt_match_win[i])
2935 curr_alt_dont_inherit_ops[curr_alt_dont_inherit_ops_num++] = i;
2936 else
2938 /* Remember pseudos used for match reloads are never
2939 inherited. */
2940 lra_assert (curr_alt_matches[i] >= 0);
2941 curr_alt_win[curr_alt_matches[i]] = false;
2943 curr_alt_win[i] = curr_alt_match_win[i] = false;
2944 losers++;
2945 /* Early clobber was already reflected in REJECT. */
2946 lra_assert (reject > 0);
2947 if (lra_dump_file != NULL)
2948 fprintf
2949 (lra_dump_file,
2950 " %d Matched conflict early clobber reloads: "
2951 "reject--\n",
2953 reject--;
2954 overall += LRA_LOSER_COST_FACTOR - 1;
2957 if (lra_dump_file != NULL)
2958 fprintf (lra_dump_file, " alt=%d,overall=%d,losers=%d,rld_nregs=%d\n",
2959 nalt, overall, losers, reload_nregs);
2961 /* If this alternative can be made to work by reloading, and it
2962 needs less reloading than the others checked so far, record
2963 it as the chosen goal for reloading. */
2964 if ((best_losers != 0 && losers == 0)
2965 || (((best_losers == 0 && losers == 0)
2966 || (best_losers != 0 && losers != 0))
2967 && (best_overall > overall
2968 || (best_overall == overall
2969 /* If the cost of the reloads is the same,
2970 prefer alternative which requires minimal
2971 number of reload regs. */
2972 && (reload_nregs < best_reload_nregs
2973 || (reload_nregs == best_reload_nregs
2974 && (best_reload_sum < reload_sum
2975 || (best_reload_sum == reload_sum
2976 && nalt < goal_alt_number))))))))
2978 for (nop = 0; nop < n_operands; nop++)
2980 goal_alt_win[nop] = curr_alt_win[nop];
2981 goal_alt_match_win[nop] = curr_alt_match_win[nop];
2982 goal_alt_matches[nop] = curr_alt_matches[nop];
2983 goal_alt[nop] = curr_alt[nop];
2984 goal_alt_offmemok[nop] = curr_alt_offmemok[nop];
2986 goal_alt_dont_inherit_ops_num = curr_alt_dont_inherit_ops_num;
2987 for (nop = 0; nop < curr_alt_dont_inherit_ops_num; nop++)
2988 goal_alt_dont_inherit_ops[nop] = curr_alt_dont_inherit_ops[nop];
2989 goal_alt_swapped = curr_swapped;
2990 best_overall = overall;
2991 best_losers = losers;
2992 best_reload_nregs = reload_nregs;
2993 best_reload_sum = reload_sum;
2994 goal_alt_number = nalt;
2996 if (losers == 0)
2997 /* Everything is satisfied. Do not process alternatives
2998 anymore. */
2999 break;
3000 fail:
3003 return ok_p;
3006 /* Make reload base reg from address AD. */
3007 static rtx
3008 base_to_reg (struct address_info *ad)
3010 enum reg_class cl;
3011 int code = -1;
3012 rtx new_inner = NULL_RTX;
3013 rtx new_reg = NULL_RTX;
3014 rtx_insn *insn;
3015 rtx_insn *last_insn = get_last_insn();
3017 lra_assert (ad->disp == ad->disp_term);
3018 cl = base_reg_class (ad->mode, ad->as, ad->base_outer_code,
3019 get_index_code (ad));
3020 new_reg = lra_create_new_reg (GET_MODE (*ad->base), NULL_RTX,
3021 cl, "base");
3022 new_inner = simplify_gen_binary (PLUS, GET_MODE (new_reg), new_reg,
3023 ad->disp_term == NULL
3024 ? const0_rtx
3025 : *ad->disp_term);
3026 if (!valid_address_p (ad->mode, new_inner, ad->as))
3027 return NULL_RTX;
3028 insn = emit_insn (gen_rtx_SET (new_reg, *ad->base));
3029 code = recog_memoized (insn);
3030 if (code < 0)
3032 delete_insns_since (last_insn);
3033 return NULL_RTX;
3036 return new_inner;
3039 /* Make reload base reg + disp from address AD. Return the new pseudo. */
3040 static rtx
3041 base_plus_disp_to_reg (struct address_info *ad)
3043 enum reg_class cl;
3044 rtx new_reg;
3046 lra_assert (ad->base == ad->base_term && ad->disp == ad->disp_term);
3047 cl = base_reg_class (ad->mode, ad->as, ad->base_outer_code,
3048 get_index_code (ad));
3049 new_reg = lra_create_new_reg (GET_MODE (*ad->base_term), NULL_RTX,
3050 cl, "base + disp");
3051 lra_emit_add (new_reg, *ad->base_term, *ad->disp_term);
3052 return new_reg;
3055 /* Make reload of index part of address AD. Return the new
3056 pseudo. */
3057 static rtx
3058 index_part_to_reg (struct address_info *ad)
3060 rtx new_reg;
3062 new_reg = lra_create_new_reg (GET_MODE (*ad->index), NULL_RTX,
3063 INDEX_REG_CLASS, "index term");
3064 expand_mult (GET_MODE (*ad->index), *ad->index_term,
3065 GEN_INT (get_index_scale (ad)), new_reg, 1);
3066 return new_reg;
3069 /* Return true if we can add a displacement to address AD, even if that
3070 makes the address invalid. The fix-up code requires any new address
3071 to be the sum of the BASE_TERM, INDEX and DISP_TERM fields. */
3072 static bool
3073 can_add_disp_p (struct address_info *ad)
3075 return (!ad->autoinc_p
3076 && ad->segment == NULL
3077 && ad->base == ad->base_term
3078 && ad->disp == ad->disp_term);
3081 /* Make equiv substitution in address AD. Return true if a substitution
3082 was made. */
3083 static bool
3084 equiv_address_substitution (struct address_info *ad)
3086 rtx base_reg, new_base_reg, index_reg, new_index_reg, *base_term, *index_term;
3087 HOST_WIDE_INT disp, scale;
3088 bool change_p;
3090 base_term = strip_subreg (ad->base_term);
3091 if (base_term == NULL)
3092 base_reg = new_base_reg = NULL_RTX;
3093 else
3095 base_reg = *base_term;
3096 new_base_reg = get_equiv_with_elimination (base_reg, curr_insn);
3098 index_term = strip_subreg (ad->index_term);
3099 if (index_term == NULL)
3100 index_reg = new_index_reg = NULL_RTX;
3101 else
3103 index_reg = *index_term;
3104 new_index_reg = get_equiv_with_elimination (index_reg, curr_insn);
3106 if (base_reg == new_base_reg && index_reg == new_index_reg)
3107 return false;
3108 disp = 0;
3109 change_p = false;
3110 if (lra_dump_file != NULL)
3112 fprintf (lra_dump_file, "Changing address in insn %d ",
3113 INSN_UID (curr_insn));
3114 dump_value_slim (lra_dump_file, *ad->outer, 1);
3116 if (base_reg != new_base_reg)
3118 if (REG_P (new_base_reg))
3120 *base_term = new_base_reg;
3121 change_p = true;
3123 else if (GET_CODE (new_base_reg) == PLUS
3124 && REG_P (XEXP (new_base_reg, 0))
3125 && CONST_INT_P (XEXP (new_base_reg, 1))
3126 && can_add_disp_p (ad))
3128 disp += INTVAL (XEXP (new_base_reg, 1));
3129 *base_term = XEXP (new_base_reg, 0);
3130 change_p = true;
3132 if (ad->base_term2 != NULL)
3133 *ad->base_term2 = *ad->base_term;
3135 if (index_reg != new_index_reg)
3137 if (REG_P (new_index_reg))
3139 *index_term = new_index_reg;
3140 change_p = true;
3142 else if (GET_CODE (new_index_reg) == PLUS
3143 && REG_P (XEXP (new_index_reg, 0))
3144 && CONST_INT_P (XEXP (new_index_reg, 1))
3145 && can_add_disp_p (ad)
3146 && (scale = get_index_scale (ad)))
3148 disp += INTVAL (XEXP (new_index_reg, 1)) * scale;
3149 *index_term = XEXP (new_index_reg, 0);
3150 change_p = true;
3153 if (disp != 0)
3155 if (ad->disp != NULL)
3156 *ad->disp = plus_constant (GET_MODE (*ad->inner), *ad->disp, disp);
3157 else
3159 *ad->inner = plus_constant (GET_MODE (*ad->inner), *ad->inner, disp);
3160 update_address (ad);
3162 change_p = true;
3164 if (lra_dump_file != NULL)
3166 if (! change_p)
3167 fprintf (lra_dump_file, " -- no change\n");
3168 else
3170 fprintf (lra_dump_file, " on equiv ");
3171 dump_value_slim (lra_dump_file, *ad->outer, 1);
3172 fprintf (lra_dump_file, "\n");
3175 return change_p;
3178 /* Major function to make reloads for an address in operand NOP or
3179 check its correctness (If CHECK_ONLY_P is true). The supported
3180 cases are:
3182 1) an address that existed before LRA started, at which point it
3183 must have been valid. These addresses are subject to elimination
3184 and may have become invalid due to the elimination offset being out
3185 of range.
3187 2) an address created by forcing a constant to memory
3188 (force_const_to_mem). The initial form of these addresses might
3189 not be valid, and it is this function's job to make them valid.
3191 3) a frame address formed from a register and a (possibly zero)
3192 constant offset. As above, these addresses might not be valid and
3193 this function must make them so.
3195 Add reloads to the lists *BEFORE and *AFTER. We might need to add
3196 reloads to *AFTER because of inc/dec, {pre, post} modify in the
3197 address. Return true for any RTL change.
3199 The function is a helper function which does not produce all
3200 transformations (when CHECK_ONLY_P is false) which can be
3201 necessary. It does just basic steps. To do all necessary
3202 transformations use function process_address. */
3203 static bool
3204 process_address_1 (int nop, bool check_only_p,
3205 rtx_insn **before, rtx_insn **after)
3207 struct address_info ad;
3208 rtx new_reg;
3209 HOST_WIDE_INT scale;
3210 rtx op = *curr_id->operand_loc[nop];
3211 const char *constraint = curr_static_id->operand[nop].constraint;
3212 enum constraint_num cn = lookup_constraint (constraint);
3213 bool change_p = false;
3215 if (MEM_P (op)
3216 && GET_MODE (op) == BLKmode
3217 && GET_CODE (XEXP (op, 0)) == SCRATCH)
3218 return false;
3220 if (insn_extra_address_constraint (cn))
3221 decompose_lea_address (&ad, curr_id->operand_loc[nop]);
3222 /* Do not attempt to decompose arbitrary addresses generated by combine
3223 for asm operands with loose constraints, e.g 'X'. */
3224 else if (MEM_P (op)
3225 && !(INSN_CODE (curr_insn) < 0
3226 && get_constraint_type (cn) == CT_FIXED_FORM
3227 && constraint_satisfied_p (op, cn)))
3228 decompose_mem_address (&ad, op);
3229 else if (GET_CODE (op) == SUBREG
3230 && MEM_P (SUBREG_REG (op)))
3231 decompose_mem_address (&ad, SUBREG_REG (op));
3232 else
3233 return false;
3234 /* If INDEX_REG_CLASS is assigned to base_term already and isn't to
3235 index_term, swap them so to avoid assigning INDEX_REG_CLASS to both
3236 when INDEX_REG_CLASS is a single register class. */
3237 if (ad.base_term != NULL
3238 && ad.index_term != NULL
3239 && ira_class_hard_regs_num[INDEX_REG_CLASS] == 1
3240 && REG_P (*ad.base_term)
3241 && REG_P (*ad.index_term)
3242 && in_class_p (*ad.base_term, INDEX_REG_CLASS, NULL)
3243 && ! in_class_p (*ad.index_term, INDEX_REG_CLASS, NULL))
3245 std::swap (ad.base, ad.index);
3246 std::swap (ad.base_term, ad.index_term);
3248 if (! check_only_p)
3249 change_p = equiv_address_substitution (&ad);
3250 if (ad.base_term != NULL
3251 && (process_addr_reg
3252 (ad.base_term, check_only_p, before,
3253 (ad.autoinc_p
3254 && !(REG_P (*ad.base_term)
3255 && find_regno_note (curr_insn, REG_DEAD,
3256 REGNO (*ad.base_term)) != NULL_RTX)
3257 ? after : NULL),
3258 base_reg_class (ad.mode, ad.as, ad.base_outer_code,
3259 get_index_code (&ad)))))
3261 change_p = true;
3262 if (ad.base_term2 != NULL)
3263 *ad.base_term2 = *ad.base_term;
3265 if (ad.index_term != NULL
3266 && process_addr_reg (ad.index_term, check_only_p,
3267 before, NULL, INDEX_REG_CLASS))
3268 change_p = true;
3270 /* Target hooks sometimes don't treat extra-constraint addresses as
3271 legitimate address_operands, so handle them specially. */
3272 if (insn_extra_address_constraint (cn)
3273 && satisfies_address_constraint_p (&ad, cn))
3274 return change_p;
3276 if (check_only_p)
3277 return change_p;
3279 /* There are three cases where the shape of *AD.INNER may now be invalid:
3281 1) the original address was valid, but either elimination or
3282 equiv_address_substitution was applied and that made
3283 the address invalid.
3285 2) the address is an invalid symbolic address created by
3286 force_const_to_mem.
3288 3) the address is a frame address with an invalid offset.
3290 4) the address is a frame address with an invalid base.
3292 All these cases involve a non-autoinc address, so there is no
3293 point revalidating other types. */
3294 if (ad.autoinc_p || valid_address_p (&ad))
3295 return change_p;
3297 /* Any index existed before LRA started, so we can assume that the
3298 presence and shape of the index is valid. */
3299 push_to_sequence (*before);
3300 lra_assert (ad.disp == ad.disp_term);
3301 if (ad.base == NULL)
3303 if (ad.index == NULL)
3305 rtx_insn *insn;
3306 rtx_insn *last = get_last_insn ();
3307 int code = -1;
3308 enum reg_class cl = base_reg_class (ad.mode, ad.as,
3309 SCRATCH, SCRATCH);
3310 rtx addr = *ad.inner;
3312 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "addr");
3313 if (HAVE_lo_sum)
3315 /* addr => lo_sum (new_base, addr), case (2) above. */
3316 insn = emit_insn (gen_rtx_SET
3317 (new_reg,
3318 gen_rtx_HIGH (Pmode, copy_rtx (addr))));
3319 code = recog_memoized (insn);
3320 if (code >= 0)
3322 *ad.inner = gen_rtx_LO_SUM (Pmode, new_reg, addr);
3323 if (! valid_address_p (ad.mode, *ad.outer, ad.as))
3325 /* Try to put lo_sum into register. */
3326 insn = emit_insn (gen_rtx_SET
3327 (new_reg,
3328 gen_rtx_LO_SUM (Pmode, new_reg, addr)));
3329 code = recog_memoized (insn);
3330 if (code >= 0)
3332 *ad.inner = new_reg;
3333 if (! valid_address_p (ad.mode, *ad.outer, ad.as))
3335 *ad.inner = addr;
3336 code = -1;
3342 if (code < 0)
3343 delete_insns_since (last);
3346 if (code < 0)
3348 /* addr => new_base, case (2) above. */
3349 lra_emit_move (new_reg, addr);
3351 for (insn = last == NULL_RTX ? get_insns () : NEXT_INSN (last);
3352 insn != NULL_RTX;
3353 insn = NEXT_INSN (insn))
3354 if (recog_memoized (insn) < 0)
3355 break;
3356 if (insn != NULL_RTX)
3358 /* Do nothing if we cannot generate right insns.
3359 This is analogous to reload pass behavior. */
3360 delete_insns_since (last);
3361 end_sequence ();
3362 return false;
3364 *ad.inner = new_reg;
3367 else
3369 /* index * scale + disp => new base + index * scale,
3370 case (1) above. */
3371 enum reg_class cl = base_reg_class (ad.mode, ad.as, PLUS,
3372 GET_CODE (*ad.index));
3374 lra_assert (INDEX_REG_CLASS != NO_REGS);
3375 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "disp");
3376 lra_emit_move (new_reg, *ad.disp);
3377 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3378 new_reg, *ad.index);
3381 else if (ad.index == NULL)
3383 int regno;
3384 enum reg_class cl;
3385 rtx set;
3386 rtx_insn *insns, *last_insn;
3387 /* Try to reload base into register only if the base is invalid
3388 for the address but with valid offset, case (4) above. */
3389 start_sequence ();
3390 new_reg = base_to_reg (&ad);
3392 /* base + disp => new base, cases (1) and (3) above. */
3393 /* Another option would be to reload the displacement into an
3394 index register. However, postreload has code to optimize
3395 address reloads that have the same base and different
3396 displacements, so reloading into an index register would
3397 not necessarily be a win. */
3398 if (new_reg == NULL_RTX)
3399 new_reg = base_plus_disp_to_reg (&ad);
3400 insns = get_insns ();
3401 last_insn = get_last_insn ();
3402 /* If we generated at least two insns, try last insn source as
3403 an address. If we succeed, we generate one less insn. */
3404 if (last_insn != insns && (set = single_set (last_insn)) != NULL_RTX
3405 && GET_CODE (SET_SRC (set)) == PLUS
3406 && REG_P (XEXP (SET_SRC (set), 0))
3407 && CONSTANT_P (XEXP (SET_SRC (set), 1)))
3409 *ad.inner = SET_SRC (set);
3410 if (valid_address_p (ad.mode, *ad.outer, ad.as))
3412 *ad.base_term = XEXP (SET_SRC (set), 0);
3413 *ad.disp_term = XEXP (SET_SRC (set), 1);
3414 cl = base_reg_class (ad.mode, ad.as, ad.base_outer_code,
3415 get_index_code (&ad));
3416 regno = REGNO (*ad.base_term);
3417 if (regno >= FIRST_PSEUDO_REGISTER
3418 && cl != lra_get_allocno_class (regno))
3419 lra_change_class (regno, cl, " Change to", true);
3420 new_reg = SET_SRC (set);
3421 delete_insns_since (PREV_INSN (last_insn));
3424 /* Try if target can split displacement into legitimite new disp
3425 and offset. If it's the case, we replace the last insn with
3426 insns for base + offset => new_reg and set new_reg + new disp
3427 to *ad.inner. */
3428 last_insn = get_last_insn ();
3429 if ((set = single_set (last_insn)) != NULL_RTX
3430 && GET_CODE (SET_SRC (set)) == PLUS
3431 && REG_P (XEXP (SET_SRC (set), 0))
3432 && REGNO (XEXP (SET_SRC (set), 0)) < FIRST_PSEUDO_REGISTER
3433 && CONST_INT_P (XEXP (SET_SRC (set), 1)))
3435 rtx addend, disp = XEXP (SET_SRC (set), 1);
3436 if (targetm.legitimize_address_displacement (&disp, &addend,
3437 ad.mode))
3439 rtx_insn *new_insns;
3440 start_sequence ();
3441 lra_emit_add (new_reg, XEXP (SET_SRC (set), 0), addend);
3442 new_insns = get_insns ();
3443 end_sequence ();
3444 new_reg = gen_rtx_PLUS (Pmode, new_reg, disp);
3445 delete_insns_since (PREV_INSN (last_insn));
3446 add_insn (new_insns);
3447 insns = get_insns ();
3450 end_sequence ();
3451 emit_insn (insns);
3452 *ad.inner = new_reg;
3454 else if (ad.disp_term != NULL)
3456 /* base + scale * index + disp => new base + scale * index,
3457 case (1) above. */
3458 new_reg = base_plus_disp_to_reg (&ad);
3459 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3460 new_reg, *ad.index);
3462 else if ((scale = get_index_scale (&ad)) == 1)
3464 /* The last transformation to one reg will be made in
3465 curr_insn_transform function. */
3466 end_sequence ();
3467 return false;
3469 else if (scale != 0)
3471 /* base + scale * index => base + new_reg,
3472 case (1) above.
3473 Index part of address may become invalid. For example, we
3474 changed pseudo on the equivalent memory and a subreg of the
3475 pseudo onto the memory of different mode for which the scale is
3476 prohibitted. */
3477 new_reg = index_part_to_reg (&ad);
3478 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3479 *ad.base_term, new_reg);
3481 else
3483 enum reg_class cl = base_reg_class (ad.mode, ad.as,
3484 SCRATCH, SCRATCH);
3485 rtx addr = *ad.inner;
3487 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "addr");
3488 /* addr => new_base. */
3489 lra_emit_move (new_reg, addr);
3490 *ad.inner = new_reg;
3492 *before = get_insns ();
3493 end_sequence ();
3494 return true;
3497 /* If CHECK_ONLY_P is false, do address reloads until it is necessary.
3498 Use process_address_1 as a helper function. Return true for any
3499 RTL changes.
3501 If CHECK_ONLY_P is true, just check address correctness. Return
3502 false if the address correct. */
3503 static bool
3504 process_address (int nop, bool check_only_p,
3505 rtx_insn **before, rtx_insn **after)
3507 bool res = false;
3509 while (process_address_1 (nop, check_only_p, before, after))
3511 if (check_only_p)
3512 return true;
3513 res = true;
3515 return res;
3518 /* Emit insns to reload VALUE into a new register. VALUE is an
3519 auto-increment or auto-decrement RTX whose operand is a register or
3520 memory location; so reloading involves incrementing that location.
3521 IN is either identical to VALUE, or some cheaper place to reload
3522 value being incremented/decremented from.
3524 INC_AMOUNT is the number to increment or decrement by (always
3525 positive and ignored for POST_MODIFY/PRE_MODIFY).
3527 Return pseudo containing the result. */
3528 static rtx
3529 emit_inc (enum reg_class new_rclass, rtx in, rtx value, int inc_amount)
3531 /* REG or MEM to be copied and incremented. */
3532 rtx incloc = XEXP (value, 0);
3533 /* Nonzero if increment after copying. */
3534 int post = (GET_CODE (value) == POST_DEC || GET_CODE (value) == POST_INC
3535 || GET_CODE (value) == POST_MODIFY);
3536 rtx_insn *last;
3537 rtx inc;
3538 rtx_insn *add_insn;
3539 int code;
3540 rtx real_in = in == value ? incloc : in;
3541 rtx result;
3542 bool plus_p = true;
3544 if (GET_CODE (value) == PRE_MODIFY || GET_CODE (value) == POST_MODIFY)
3546 lra_assert (GET_CODE (XEXP (value, 1)) == PLUS
3547 || GET_CODE (XEXP (value, 1)) == MINUS);
3548 lra_assert (rtx_equal_p (XEXP (XEXP (value, 1), 0), XEXP (value, 0)));
3549 plus_p = GET_CODE (XEXP (value, 1)) == PLUS;
3550 inc = XEXP (XEXP (value, 1), 1);
3552 else
3554 if (GET_CODE (value) == PRE_DEC || GET_CODE (value) == POST_DEC)
3555 inc_amount = -inc_amount;
3557 inc = GEN_INT (inc_amount);
3560 if (! post && REG_P (incloc))
3561 result = incloc;
3562 else
3563 result = lra_create_new_reg (GET_MODE (value), value, new_rclass,
3564 "INC/DEC result");
3566 if (real_in != result)
3568 /* First copy the location to the result register. */
3569 lra_assert (REG_P (result));
3570 emit_insn (gen_move_insn (result, real_in));
3573 /* We suppose that there are insns to add/sub with the constant
3574 increment permitted in {PRE/POST)_{DEC/INC/MODIFY}. At least the
3575 old reload worked with this assumption. If the assumption
3576 becomes wrong, we should use approach in function
3577 base_plus_disp_to_reg. */
3578 if (in == value)
3580 /* See if we can directly increment INCLOC. */
3581 last = get_last_insn ();
3582 add_insn = emit_insn (plus_p
3583 ? gen_add2_insn (incloc, inc)
3584 : gen_sub2_insn (incloc, inc));
3586 code = recog_memoized (add_insn);
3587 if (code >= 0)
3589 if (! post && result != incloc)
3590 emit_insn (gen_move_insn (result, incloc));
3591 return result;
3593 delete_insns_since (last);
3596 /* If couldn't do the increment directly, must increment in RESULT.
3597 The way we do this depends on whether this is pre- or
3598 post-increment. For pre-increment, copy INCLOC to the reload
3599 register, increment it there, then save back. */
3600 if (! post)
3602 if (real_in != result)
3603 emit_insn (gen_move_insn (result, real_in));
3604 if (plus_p)
3605 emit_insn (gen_add2_insn (result, inc));
3606 else
3607 emit_insn (gen_sub2_insn (result, inc));
3608 if (result != incloc)
3609 emit_insn (gen_move_insn (incloc, result));
3611 else
3613 /* Post-increment.
3615 Because this might be a jump insn or a compare, and because
3616 RESULT may not be available after the insn in an input
3617 reload, we must do the incrementing before the insn being
3618 reloaded for.
3620 We have already copied IN to RESULT. Increment the copy in
3621 RESULT, save that back, then decrement RESULT so it has
3622 the original value. */
3623 if (plus_p)
3624 emit_insn (gen_add2_insn (result, inc));
3625 else
3626 emit_insn (gen_sub2_insn (result, inc));
3627 emit_insn (gen_move_insn (incloc, result));
3628 /* Restore non-modified value for the result. We prefer this
3629 way because it does not require an additional hard
3630 register. */
3631 if (plus_p)
3633 if (CONST_INT_P (inc))
3634 emit_insn (gen_add2_insn (result,
3635 gen_int_mode (-INTVAL (inc),
3636 GET_MODE (result))));
3637 else
3638 emit_insn (gen_sub2_insn (result, inc));
3640 else
3641 emit_insn (gen_add2_insn (result, inc));
3643 return result;
3646 /* Return true if the current move insn does not need processing as we
3647 already know that it satisfies its constraints. */
3648 static bool
3649 simple_move_p (void)
3651 rtx dest, src;
3652 enum reg_class dclass, sclass;
3654 lra_assert (curr_insn_set != NULL_RTX);
3655 dest = SET_DEST (curr_insn_set);
3656 src = SET_SRC (curr_insn_set);
3658 /* If the instruction has multiple sets we need to process it even if it
3659 is single_set. This can happen if one or more of the SETs are dead.
3660 See PR73650. */
3661 if (multiple_sets (curr_insn))
3662 return false;
3664 return ((dclass = get_op_class (dest)) != NO_REGS
3665 && (sclass = get_op_class (src)) != NO_REGS
3666 /* The backend guarantees that register moves of cost 2
3667 never need reloads. */
3668 && targetm.register_move_cost (GET_MODE (src), sclass, dclass) == 2);
3671 /* Swap operands NOP and NOP + 1. */
3672 static inline void
3673 swap_operands (int nop)
3675 std::swap (curr_operand_mode[nop], curr_operand_mode[nop + 1]);
3676 std::swap (original_subreg_reg_mode[nop], original_subreg_reg_mode[nop + 1]);
3677 std::swap (*curr_id->operand_loc[nop], *curr_id->operand_loc[nop + 1]);
3678 std::swap (equiv_substition_p[nop], equiv_substition_p[nop + 1]);
3679 /* Swap the duplicates too. */
3680 lra_update_dup (curr_id, nop);
3681 lra_update_dup (curr_id, nop + 1);
3684 /* Main entry point of the constraint code: search the body of the
3685 current insn to choose the best alternative. It is mimicking insn
3686 alternative cost calculation model of former reload pass. That is
3687 because machine descriptions were written to use this model. This
3688 model can be changed in future. Make commutative operand exchange
3689 if it is chosen.
3691 if CHECK_ONLY_P is false, do RTL changes to satisfy the
3692 constraints. Return true if any change happened during function
3693 call.
3695 If CHECK_ONLY_P is true then don't do any transformation. Just
3696 check that the insn satisfies all constraints. If the insn does
3697 not satisfy any constraint, return true. */
3698 static bool
3699 curr_insn_transform (bool check_only_p)
3701 int i, j, k;
3702 int n_operands;
3703 int n_alternatives;
3704 int n_outputs;
3705 int commutative;
3706 signed char goal_alt_matched[MAX_RECOG_OPERANDS][MAX_RECOG_OPERANDS];
3707 signed char match_inputs[MAX_RECOG_OPERANDS + 1];
3708 signed char outputs[MAX_RECOG_OPERANDS + 1];
3709 rtx_insn *before, *after;
3710 bool alt_p = false;
3711 /* Flag that the insn has been changed through a transformation. */
3712 bool change_p;
3713 bool sec_mem_p;
3714 bool use_sec_mem_p;
3715 int max_regno_before;
3716 int reused_alternative_num;
3718 curr_insn_set = single_set (curr_insn);
3719 if (curr_insn_set != NULL_RTX && simple_move_p ())
3720 return false;
3722 no_input_reloads_p = no_output_reloads_p = false;
3723 goal_alt_number = -1;
3724 change_p = sec_mem_p = false;
3725 /* JUMP_INSNs and CALL_INSNs are not allowed to have any output
3726 reloads; neither are insns that SET cc0. Insns that use CC0 are
3727 not allowed to have any input reloads. */
3728 if (JUMP_P (curr_insn) || CALL_P (curr_insn))
3729 no_output_reloads_p = true;
3731 if (HAVE_cc0 && reg_referenced_p (cc0_rtx, PATTERN (curr_insn)))
3732 no_input_reloads_p = true;
3733 if (HAVE_cc0 && reg_set_p (cc0_rtx, PATTERN (curr_insn)))
3734 no_output_reloads_p = true;
3736 n_operands = curr_static_id->n_operands;
3737 n_alternatives = curr_static_id->n_alternatives;
3739 /* Just return "no reloads" if insn has no operands with
3740 constraints. */
3741 if (n_operands == 0 || n_alternatives == 0)
3742 return false;
3744 max_regno_before = max_reg_num ();
3746 for (i = 0; i < n_operands; i++)
3748 goal_alt_matched[i][0] = -1;
3749 goal_alt_matches[i] = -1;
3752 commutative = curr_static_id->commutative;
3754 /* Now see what we need for pseudos that didn't get hard regs or got
3755 the wrong kind of hard reg. For this, we must consider all the
3756 operands together against the register constraints. */
3758 best_losers = best_overall = INT_MAX;
3759 best_reload_sum = 0;
3761 curr_swapped = false;
3762 goal_alt_swapped = false;
3764 if (! check_only_p)
3765 /* Make equivalence substitution and memory subreg elimination
3766 before address processing because an address legitimacy can
3767 depend on memory mode. */
3768 for (i = 0; i < n_operands; i++)
3770 rtx op, subst, old;
3771 bool op_change_p = false;
3773 if (curr_static_id->operand[i].is_operator)
3774 continue;
3776 old = op = *curr_id->operand_loc[i];
3777 if (GET_CODE (old) == SUBREG)
3778 old = SUBREG_REG (old);
3779 subst = get_equiv_with_elimination (old, curr_insn);
3780 original_subreg_reg_mode[i] = VOIDmode;
3781 equiv_substition_p[i] = false;
3782 if (subst != old)
3784 equiv_substition_p[i] = true;
3785 subst = copy_rtx (subst);
3786 lra_assert (REG_P (old));
3787 if (GET_CODE (op) != SUBREG)
3788 *curr_id->operand_loc[i] = subst;
3789 else
3791 SUBREG_REG (op) = subst;
3792 if (GET_MODE (subst) == VOIDmode)
3793 original_subreg_reg_mode[i] = GET_MODE (old);
3795 if (lra_dump_file != NULL)
3797 fprintf (lra_dump_file,
3798 "Changing pseudo %d in operand %i of insn %u on equiv ",
3799 REGNO (old), i, INSN_UID (curr_insn));
3800 dump_value_slim (lra_dump_file, subst, 1);
3801 fprintf (lra_dump_file, "\n");
3803 op_change_p = change_p = true;
3805 if (simplify_operand_subreg (i, GET_MODE (old)) || op_change_p)
3807 change_p = true;
3808 lra_update_dup (curr_id, i);
3812 /* Reload address registers and displacements. We do it before
3813 finding an alternative because of memory constraints. */
3814 before = after = NULL;
3815 for (i = 0; i < n_operands; i++)
3816 if (! curr_static_id->operand[i].is_operator
3817 && process_address (i, check_only_p, &before, &after))
3819 if (check_only_p)
3820 return true;
3821 change_p = true;
3822 lra_update_dup (curr_id, i);
3825 if (change_p)
3826 /* If we've changed the instruction then any alternative that
3827 we chose previously may no longer be valid. */
3828 lra_set_used_insn_alternative (curr_insn, -1);
3830 if (! check_only_p && curr_insn_set != NULL_RTX
3831 && check_and_process_move (&change_p, &sec_mem_p))
3832 return change_p;
3834 try_swapped:
3836 reused_alternative_num = check_only_p ? -1 : curr_id->used_insn_alternative;
3837 if (lra_dump_file != NULL && reused_alternative_num >= 0)
3838 fprintf (lra_dump_file, "Reusing alternative %d for insn #%u\n",
3839 reused_alternative_num, INSN_UID (curr_insn));
3841 if (process_alt_operands (reused_alternative_num))
3842 alt_p = true;
3844 if (check_only_p)
3845 return ! alt_p || best_losers != 0;
3847 /* If insn is commutative (it's safe to exchange a certain pair of
3848 operands) then we need to try each alternative twice, the second
3849 time matching those two operands as if we had exchanged them. To
3850 do this, really exchange them in operands.
3852 If we have just tried the alternatives the second time, return
3853 operands to normal and drop through. */
3855 if (reused_alternative_num < 0 && commutative >= 0)
3857 curr_swapped = !curr_swapped;
3858 if (curr_swapped)
3860 swap_operands (commutative);
3861 goto try_swapped;
3863 else
3864 swap_operands (commutative);
3867 if (! alt_p && ! sec_mem_p)
3869 /* No alternative works with reloads?? */
3870 if (INSN_CODE (curr_insn) >= 0)
3871 fatal_insn ("unable to generate reloads for:", curr_insn);
3872 error_for_asm (curr_insn,
3873 "inconsistent operand constraints in an %<asm%>");
3874 /* Avoid further trouble with this insn. Don't generate use
3875 pattern here as we could use the insn SP offset. */
3876 lra_set_insn_deleted (curr_insn);
3877 return true;
3880 /* If the best alternative is with operands 1 and 2 swapped, swap
3881 them. Update the operand numbers of any reloads already
3882 pushed. */
3884 if (goal_alt_swapped)
3886 if (lra_dump_file != NULL)
3887 fprintf (lra_dump_file, " Commutative operand exchange in insn %u\n",
3888 INSN_UID (curr_insn));
3890 /* Swap the duplicates too. */
3891 swap_operands (commutative);
3892 change_p = true;
3895 /* Some targets' TARGET_SECONDARY_MEMORY_NEEDED (e.g. x86) are defined
3896 too conservatively. So we use the secondary memory only if there
3897 is no any alternative without reloads. */
3898 use_sec_mem_p = false;
3899 if (! alt_p)
3900 use_sec_mem_p = true;
3901 else if (sec_mem_p)
3903 for (i = 0; i < n_operands; i++)
3904 if (! goal_alt_win[i] && ! goal_alt_match_win[i])
3905 break;
3906 use_sec_mem_p = i < n_operands;
3909 if (use_sec_mem_p)
3911 int in = -1, out = -1;
3912 rtx new_reg, src, dest, rld;
3913 machine_mode sec_mode, rld_mode;
3915 lra_assert (curr_insn_set != NULL_RTX && sec_mem_p);
3916 dest = SET_DEST (curr_insn_set);
3917 src = SET_SRC (curr_insn_set);
3918 for (i = 0; i < n_operands; i++)
3919 if (*curr_id->operand_loc[i] == dest)
3920 out = i;
3921 else if (*curr_id->operand_loc[i] == src)
3922 in = i;
3923 for (i = 0; i < curr_static_id->n_dups; i++)
3924 if (out < 0 && *curr_id->dup_loc[i] == dest)
3925 out = curr_static_id->dup_num[i];
3926 else if (in < 0 && *curr_id->dup_loc[i] == src)
3927 in = curr_static_id->dup_num[i];
3928 lra_assert (out >= 0 && in >= 0
3929 && curr_static_id->operand[out].type == OP_OUT
3930 && curr_static_id->operand[in].type == OP_IN);
3931 rld = partial_subreg_p (GET_MODE (src), GET_MODE (dest)) ? src : dest;
3932 rld_mode = GET_MODE (rld);
3933 sec_mode = targetm.secondary_memory_needed_mode (rld_mode);
3934 new_reg = lra_create_new_reg (sec_mode, NULL_RTX,
3935 NO_REGS, "secondary");
3936 /* If the mode is changed, it should be wider. */
3937 lra_assert (!partial_subreg_p (sec_mode, rld_mode));
3938 if (sec_mode != rld_mode)
3940 /* If the target says specifically to use another mode for
3941 secondary memory moves we can not reuse the original
3942 insn. */
3943 after = emit_spill_move (false, new_reg, dest);
3944 lra_process_new_insns (curr_insn, NULL, after,
3945 "Inserting the sec. move");
3946 /* We may have non null BEFORE here (e.g. after address
3947 processing. */
3948 push_to_sequence (before);
3949 before = emit_spill_move (true, new_reg, src);
3950 emit_insn (before);
3951 before = get_insns ();
3952 end_sequence ();
3953 lra_process_new_insns (curr_insn, before, NULL, "Changing on");
3954 lra_set_insn_deleted (curr_insn);
3956 else if (dest == rld)
3958 *curr_id->operand_loc[out] = new_reg;
3959 lra_update_dup (curr_id, out);
3960 after = emit_spill_move (false, new_reg, dest);
3961 lra_process_new_insns (curr_insn, NULL, after,
3962 "Inserting the sec. move");
3964 else
3966 *curr_id->operand_loc[in] = new_reg;
3967 lra_update_dup (curr_id, in);
3968 /* See comments above. */
3969 push_to_sequence (before);
3970 before = emit_spill_move (true, new_reg, src);
3971 emit_insn (before);
3972 before = get_insns ();
3973 end_sequence ();
3974 lra_process_new_insns (curr_insn, before, NULL,
3975 "Inserting the sec. move");
3977 lra_update_insn_regno_info (curr_insn);
3978 return true;
3981 lra_assert (goal_alt_number >= 0);
3982 lra_set_used_insn_alternative (curr_insn, goal_alt_number);
3984 if (lra_dump_file != NULL)
3986 const char *p;
3988 fprintf (lra_dump_file, " Choosing alt %d in insn %u:",
3989 goal_alt_number, INSN_UID (curr_insn));
3990 for (i = 0; i < n_operands; i++)
3992 p = (curr_static_id->operand_alternative
3993 [goal_alt_number * n_operands + i].constraint);
3994 if (*p == '\0')
3995 continue;
3996 fprintf (lra_dump_file, " (%d) ", i);
3997 for (; *p != '\0' && *p != ',' && *p != '#'; p++)
3998 fputc (*p, lra_dump_file);
4000 if (INSN_CODE (curr_insn) >= 0
4001 && (p = get_insn_name (INSN_CODE (curr_insn))) != NULL)
4002 fprintf (lra_dump_file, " {%s}", p);
4003 if (curr_id->sp_offset != 0)
4004 fprintf (lra_dump_file, " (sp_off=%" HOST_WIDE_INT_PRINT "d)",
4005 curr_id->sp_offset);
4006 fprintf (lra_dump_file, "\n");
4009 /* Right now, for any pair of operands I and J that are required to
4010 match, with J < I, goal_alt_matches[I] is J. Add I to
4011 goal_alt_matched[J]. */
4013 for (i = 0; i < n_operands; i++)
4014 if ((j = goal_alt_matches[i]) >= 0)
4016 for (k = 0; goal_alt_matched[j][k] >= 0; k++)
4018 /* We allow matching one output operand and several input
4019 operands. */
4020 lra_assert (k == 0
4021 || (curr_static_id->operand[j].type == OP_OUT
4022 && curr_static_id->operand[i].type == OP_IN
4023 && (curr_static_id->operand
4024 [goal_alt_matched[j][0]].type == OP_IN)));
4025 goal_alt_matched[j][k] = i;
4026 goal_alt_matched[j][k + 1] = -1;
4029 for (i = 0; i < n_operands; i++)
4030 goal_alt_win[i] |= goal_alt_match_win[i];
4032 /* Any constants that aren't allowed and can't be reloaded into
4033 registers are here changed into memory references. */
4034 for (i = 0; i < n_operands; i++)
4035 if (goal_alt_win[i])
4037 int regno;
4038 enum reg_class new_class;
4039 rtx reg = *curr_id->operand_loc[i];
4041 if (GET_CODE (reg) == SUBREG)
4042 reg = SUBREG_REG (reg);
4044 if (REG_P (reg) && (regno = REGNO (reg)) >= FIRST_PSEUDO_REGISTER)
4046 bool ok_p = in_class_p (reg, goal_alt[i], &new_class);
4048 if (new_class != NO_REGS && get_reg_class (regno) != new_class)
4050 lra_assert (ok_p);
4051 lra_change_class (regno, new_class, " Change to", true);
4055 else
4057 const char *constraint;
4058 char c;
4059 rtx op = *curr_id->operand_loc[i];
4060 rtx subreg = NULL_RTX;
4061 machine_mode mode = curr_operand_mode[i];
4063 if (GET_CODE (op) == SUBREG)
4065 subreg = op;
4066 op = SUBREG_REG (op);
4067 mode = GET_MODE (op);
4070 if (CONST_POOL_OK_P (mode, op)
4071 && ((targetm.preferred_reload_class
4072 (op, (enum reg_class) goal_alt[i]) == NO_REGS)
4073 || no_input_reloads_p))
4075 rtx tem = force_const_mem (mode, op);
4077 change_p = true;
4078 if (subreg != NULL_RTX)
4079 tem = gen_rtx_SUBREG (mode, tem, SUBREG_BYTE (subreg));
4081 *curr_id->operand_loc[i] = tem;
4082 lra_update_dup (curr_id, i);
4083 process_address (i, false, &before, &after);
4085 /* If the alternative accepts constant pool refs directly
4086 there will be no reload needed at all. */
4087 if (subreg != NULL_RTX)
4088 continue;
4089 /* Skip alternatives before the one requested. */
4090 constraint = (curr_static_id->operand_alternative
4091 [goal_alt_number * n_operands + i].constraint);
4092 for (;
4093 (c = *constraint) && c != ',' && c != '#';
4094 constraint += CONSTRAINT_LEN (c, constraint))
4096 enum constraint_num cn = lookup_constraint (constraint);
4097 if ((insn_extra_memory_constraint (cn)
4098 || insn_extra_special_memory_constraint (cn))
4099 && satisfies_memory_constraint_p (tem, cn))
4100 break;
4102 if (c == '\0' || c == ',' || c == '#')
4103 continue;
4105 goal_alt_win[i] = true;
4109 n_outputs = 0;
4110 outputs[0] = -1;
4111 for (i = 0; i < n_operands; i++)
4113 int regno;
4114 bool optional_p = false;
4115 rtx old, new_reg;
4116 rtx op = *curr_id->operand_loc[i];
4118 if (goal_alt_win[i])
4120 if (goal_alt[i] == NO_REGS
4121 && REG_P (op)
4122 /* When we assign NO_REGS it means that we will not
4123 assign a hard register to the scratch pseudo by
4124 assigment pass and the scratch pseudo will be
4125 spilled. Spilled scratch pseudos are transformed
4126 back to scratches at the LRA end. */
4127 && lra_former_scratch_operand_p (curr_insn, i)
4128 && lra_former_scratch_p (REGNO (op)))
4130 int regno = REGNO (op);
4131 lra_change_class (regno, NO_REGS, " Change to", true);
4132 if (lra_get_regno_hard_regno (regno) >= 0)
4133 /* We don't have to mark all insn affected by the
4134 spilled pseudo as there is only one such insn, the
4135 current one. */
4136 reg_renumber[regno] = -1;
4137 lra_assert (bitmap_single_bit_set_p
4138 (&lra_reg_info[REGNO (op)].insn_bitmap));
4140 /* We can do an optional reload. If the pseudo got a hard
4141 reg, we might improve the code through inheritance. If
4142 it does not get a hard register we coalesce memory/memory
4143 moves later. Ignore move insns to avoid cycling. */
4144 if (! lra_simple_p
4145 && lra_undo_inheritance_iter < LRA_MAX_INHERITANCE_PASSES
4146 && goal_alt[i] != NO_REGS && REG_P (op)
4147 && (regno = REGNO (op)) >= FIRST_PSEUDO_REGISTER
4148 && regno < new_regno_start
4149 && ! lra_former_scratch_p (regno)
4150 && reg_renumber[regno] < 0
4151 /* Check that the optional reload pseudo will be able to
4152 hold given mode value. */
4153 && ! (prohibited_class_reg_set_mode_p
4154 (goal_alt[i], reg_class_contents[goal_alt[i]],
4155 PSEUDO_REGNO_MODE (regno)))
4156 && (curr_insn_set == NULL_RTX
4157 || !((REG_P (SET_SRC (curr_insn_set))
4158 || MEM_P (SET_SRC (curr_insn_set))
4159 || GET_CODE (SET_SRC (curr_insn_set)) == SUBREG)
4160 && (REG_P (SET_DEST (curr_insn_set))
4161 || MEM_P (SET_DEST (curr_insn_set))
4162 || GET_CODE (SET_DEST (curr_insn_set)) == SUBREG))))
4163 optional_p = true;
4164 else
4165 continue;
4168 /* Operands that match previous ones have already been handled. */
4169 if (goal_alt_matches[i] >= 0)
4170 continue;
4172 /* We should not have an operand with a non-offsettable address
4173 appearing where an offsettable address will do. It also may
4174 be a case when the address should be special in other words
4175 not a general one (e.g. it needs no index reg). */
4176 if (goal_alt_matched[i][0] == -1 && goal_alt_offmemok[i] && MEM_P (op))
4178 enum reg_class rclass;
4179 rtx *loc = &XEXP (op, 0);
4180 enum rtx_code code = GET_CODE (*loc);
4182 push_to_sequence (before);
4183 rclass = base_reg_class (GET_MODE (op), MEM_ADDR_SPACE (op),
4184 MEM, SCRATCH);
4185 if (GET_RTX_CLASS (code) == RTX_AUTOINC)
4186 new_reg = emit_inc (rclass, *loc, *loc,
4187 /* This value does not matter for MODIFY. */
4188 GET_MODE_SIZE (GET_MODE (op)));
4189 else if (get_reload_reg (OP_IN, Pmode, *loc, rclass, FALSE,
4190 "offsetable address", &new_reg))
4191 lra_emit_move (new_reg, *loc);
4192 before = get_insns ();
4193 end_sequence ();
4194 *loc = new_reg;
4195 lra_update_dup (curr_id, i);
4197 else if (goal_alt_matched[i][0] == -1)
4199 machine_mode mode;
4200 rtx reg, *loc;
4201 int hard_regno, byte;
4202 enum op_type type = curr_static_id->operand[i].type;
4204 loc = curr_id->operand_loc[i];
4205 mode = curr_operand_mode[i];
4206 if (GET_CODE (*loc) == SUBREG)
4208 reg = SUBREG_REG (*loc);
4209 byte = SUBREG_BYTE (*loc);
4210 if (REG_P (reg)
4211 /* Strict_low_part requires reloading the register and not
4212 just the subreg. Likewise for a strict subreg no wider
4213 than a word for WORD_REGISTER_OPERATIONS targets. */
4214 && (curr_static_id->operand[i].strict_low
4215 || (!paradoxical_subreg_p (mode, GET_MODE (reg))
4216 && (hard_regno
4217 = get_try_hard_regno (REGNO (reg))) >= 0
4218 && (simplify_subreg_regno
4219 (hard_regno,
4220 GET_MODE (reg), byte, mode) < 0)
4221 && (goal_alt[i] == NO_REGS
4222 || (simplify_subreg_regno
4223 (ira_class_hard_regs[goal_alt[i]][0],
4224 GET_MODE (reg), byte, mode) >= 0)))
4225 || (GET_MODE_PRECISION (mode)
4226 < GET_MODE_PRECISION (GET_MODE (reg))
4227 && GET_MODE_SIZE (GET_MODE (reg)) <= UNITS_PER_WORD
4228 && WORD_REGISTER_OPERATIONS)))
4230 /* An OP_INOUT is required when reloading a subreg of a
4231 mode wider than a word to ensure that data beyond the
4232 word being reloaded is preserved. Also automatically
4233 ensure that strict_low_part reloads are made into
4234 OP_INOUT which should already be true from the backend
4235 constraints. */
4236 if (type == OP_OUT
4237 && (curr_static_id->operand[i].strict_low
4238 || read_modify_subreg_p (*loc)))
4239 type = OP_INOUT;
4240 loc = &SUBREG_REG (*loc);
4241 mode = GET_MODE (*loc);
4244 old = *loc;
4245 if (get_reload_reg (type, mode, old, goal_alt[i],
4246 loc != curr_id->operand_loc[i], "", &new_reg)
4247 && type != OP_OUT)
4249 push_to_sequence (before);
4250 lra_emit_move (new_reg, old);
4251 before = get_insns ();
4252 end_sequence ();
4254 *loc = new_reg;
4255 if (type != OP_IN
4256 && find_reg_note (curr_insn, REG_UNUSED, old) == NULL_RTX)
4258 start_sequence ();
4259 lra_emit_move (type == OP_INOUT ? copy_rtx (old) : old, new_reg);
4260 emit_insn (after);
4261 after = get_insns ();
4262 end_sequence ();
4263 *loc = new_reg;
4265 for (j = 0; j < goal_alt_dont_inherit_ops_num; j++)
4266 if (goal_alt_dont_inherit_ops[j] == i)
4268 lra_set_regno_unique_value (REGNO (new_reg));
4269 break;
4271 lra_update_dup (curr_id, i);
4273 else if (curr_static_id->operand[i].type == OP_IN
4274 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4275 == OP_OUT
4276 || (curr_static_id->operand[goal_alt_matched[i][0]].type
4277 == OP_INOUT
4278 && (operands_match_p
4279 (*curr_id->operand_loc[i],
4280 *curr_id->operand_loc[goal_alt_matched[i][0]],
4281 -1)))))
4283 /* generate reloads for input and matched outputs. */
4284 match_inputs[0] = i;
4285 match_inputs[1] = -1;
4286 match_reload (goal_alt_matched[i][0], match_inputs, outputs,
4287 goal_alt[i], &before, &after,
4288 curr_static_id->operand_alternative
4289 [goal_alt_number * n_operands + goal_alt_matched[i][0]]
4290 .earlyclobber);
4292 else if ((curr_static_id->operand[i].type == OP_OUT
4293 || (curr_static_id->operand[i].type == OP_INOUT
4294 && (operands_match_p
4295 (*curr_id->operand_loc[i],
4296 *curr_id->operand_loc[goal_alt_matched[i][0]],
4297 -1))))
4298 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4299 == OP_IN))
4300 /* Generate reloads for output and matched inputs. */
4301 match_reload (i, goal_alt_matched[i], outputs, goal_alt[i], &before,
4302 &after, curr_static_id->operand_alternative
4303 [goal_alt_number * n_operands + i].earlyclobber);
4304 else if (curr_static_id->operand[i].type == OP_IN
4305 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4306 == OP_IN))
4308 /* Generate reloads for matched inputs. */
4309 match_inputs[0] = i;
4310 for (j = 0; (k = goal_alt_matched[i][j]) >= 0; j++)
4311 match_inputs[j + 1] = k;
4312 match_inputs[j + 1] = -1;
4313 match_reload (-1, match_inputs, outputs, goal_alt[i], &before,
4314 &after, false);
4316 else
4317 /* We must generate code in any case when function
4318 process_alt_operands decides that it is possible. */
4319 gcc_unreachable ();
4321 /* Memorise processed outputs so that output remaining to be processed
4322 can avoid using the same register value (see match_reload). */
4323 if (curr_static_id->operand[i].type == OP_OUT)
4325 outputs[n_outputs++] = i;
4326 outputs[n_outputs] = -1;
4329 if (optional_p)
4331 rtx reg = op;
4333 lra_assert (REG_P (reg));
4334 regno = REGNO (reg);
4335 op = *curr_id->operand_loc[i]; /* Substitution. */
4336 if (GET_CODE (op) == SUBREG)
4337 op = SUBREG_REG (op);
4338 gcc_assert (REG_P (op) && (int) REGNO (op) >= new_regno_start);
4339 bitmap_set_bit (&lra_optional_reload_pseudos, REGNO (op));
4340 lra_reg_info[REGNO (op)].restore_rtx = reg;
4341 if (lra_dump_file != NULL)
4342 fprintf (lra_dump_file,
4343 " Making reload reg %d for reg %d optional\n",
4344 REGNO (op), regno);
4347 if (before != NULL_RTX || after != NULL_RTX
4348 || max_regno_before != max_reg_num ())
4349 change_p = true;
4350 if (change_p)
4352 lra_update_operator_dups (curr_id);
4353 /* Something changes -- process the insn. */
4354 lra_update_insn_regno_info (curr_insn);
4356 lra_process_new_insns (curr_insn, before, after, "Inserting insn reload");
4357 return change_p;
4360 /* Return true if INSN satisfies all constraints. In other words, no
4361 reload insns are needed. */
4362 bool
4363 lra_constrain_insn (rtx_insn *insn)
4365 int saved_new_regno_start = new_regno_start;
4366 int saved_new_insn_uid_start = new_insn_uid_start;
4367 bool change_p;
4369 curr_insn = insn;
4370 curr_id = lra_get_insn_recog_data (curr_insn);
4371 curr_static_id = curr_id->insn_static_data;
4372 new_insn_uid_start = get_max_uid ();
4373 new_regno_start = max_reg_num ();
4374 change_p = curr_insn_transform (true);
4375 new_regno_start = saved_new_regno_start;
4376 new_insn_uid_start = saved_new_insn_uid_start;
4377 return ! change_p;
4380 /* Return true if X is in LIST. */
4381 static bool
4382 in_list_p (rtx x, rtx list)
4384 for (; list != NULL_RTX; list = XEXP (list, 1))
4385 if (XEXP (list, 0) == x)
4386 return true;
4387 return false;
4390 /* Return true if X contains an allocatable hard register (if
4391 HARD_REG_P) or a (spilled if SPILLED_P) pseudo. */
4392 static bool
4393 contains_reg_p (rtx x, bool hard_reg_p, bool spilled_p)
4395 int i, j;
4396 const char *fmt;
4397 enum rtx_code code;
4399 code = GET_CODE (x);
4400 if (REG_P (x))
4402 int regno = REGNO (x);
4403 HARD_REG_SET alloc_regs;
4405 if (hard_reg_p)
4407 if (regno >= FIRST_PSEUDO_REGISTER)
4408 regno = lra_get_regno_hard_regno (regno);
4409 if (regno < 0)
4410 return false;
4411 COMPL_HARD_REG_SET (alloc_regs, lra_no_alloc_regs);
4412 return overlaps_hard_reg_set_p (alloc_regs, GET_MODE (x), regno);
4414 else
4416 if (regno < FIRST_PSEUDO_REGISTER)
4417 return false;
4418 if (! spilled_p)
4419 return true;
4420 return lra_get_regno_hard_regno (regno) < 0;
4423 fmt = GET_RTX_FORMAT (code);
4424 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4426 if (fmt[i] == 'e')
4428 if (contains_reg_p (XEXP (x, i), hard_reg_p, spilled_p))
4429 return true;
4431 else if (fmt[i] == 'E')
4433 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4434 if (contains_reg_p (XVECEXP (x, i, j), hard_reg_p, spilled_p))
4435 return true;
4438 return false;
4441 /* Process all regs in location *LOC and change them on equivalent
4442 substitution. Return true if any change was done. */
4443 static bool
4444 loc_equivalence_change_p (rtx *loc)
4446 rtx subst, reg, x = *loc;
4447 bool result = false;
4448 enum rtx_code code = GET_CODE (x);
4449 const char *fmt;
4450 int i, j;
4452 if (code == SUBREG)
4454 reg = SUBREG_REG (x);
4455 if ((subst = get_equiv_with_elimination (reg, curr_insn)) != reg
4456 && GET_MODE (subst) == VOIDmode)
4458 /* We cannot reload debug location. Simplify subreg here
4459 while we know the inner mode. */
4460 *loc = simplify_gen_subreg (GET_MODE (x), subst,
4461 GET_MODE (reg), SUBREG_BYTE (x));
4462 return true;
4465 if (code == REG && (subst = get_equiv_with_elimination (x, curr_insn)) != x)
4467 *loc = subst;
4468 return true;
4471 /* Scan all the operand sub-expressions. */
4472 fmt = GET_RTX_FORMAT (code);
4473 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4475 if (fmt[i] == 'e')
4476 result = loc_equivalence_change_p (&XEXP (x, i)) || result;
4477 else if (fmt[i] == 'E')
4478 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4479 result
4480 = loc_equivalence_change_p (&XVECEXP (x, i, j)) || result;
4482 return result;
4485 /* Similar to loc_equivalence_change_p, but for use as
4486 simplify_replace_fn_rtx callback. DATA is insn for which the
4487 elimination is done. If it null we don't do the elimination. */
4488 static rtx
4489 loc_equivalence_callback (rtx loc, const_rtx, void *data)
4491 if (!REG_P (loc))
4492 return NULL_RTX;
4494 rtx subst = (data == NULL
4495 ? get_equiv (loc) : get_equiv_with_elimination (loc, (rtx_insn *) data));
4496 if (subst != loc)
4497 return subst;
4499 return NULL_RTX;
4502 /* Maximum number of generated reload insns per an insn. It is for
4503 preventing this pass cycling in a bug case. */
4504 #define MAX_RELOAD_INSNS_NUMBER LRA_MAX_INSN_RELOADS
4506 /* The current iteration number of this LRA pass. */
4507 int lra_constraint_iter;
4509 /* True if we substituted equiv which needs checking register
4510 allocation correctness because the equivalent value contains
4511 allocatable hard registers or when we restore multi-register
4512 pseudo. */
4513 bool lra_risky_transformations_p;
4515 /* Return true if REGNO is referenced in more than one block. */
4516 static bool
4517 multi_block_pseudo_p (int regno)
4519 basic_block bb = NULL;
4520 unsigned int uid;
4521 bitmap_iterator bi;
4523 if (regno < FIRST_PSEUDO_REGISTER)
4524 return false;
4526 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi)
4527 if (bb == NULL)
4528 bb = BLOCK_FOR_INSN (lra_insn_recog_data[uid]->insn);
4529 else if (BLOCK_FOR_INSN (lra_insn_recog_data[uid]->insn) != bb)
4530 return true;
4531 return false;
4534 /* Return true if LIST contains a deleted insn. */
4535 static bool
4536 contains_deleted_insn_p (rtx_insn_list *list)
4538 for (; list != NULL_RTX; list = list->next ())
4539 if (NOTE_P (list->insn ())
4540 && NOTE_KIND (list->insn ()) == NOTE_INSN_DELETED)
4541 return true;
4542 return false;
4545 /* Return true if X contains a pseudo dying in INSN. */
4546 static bool
4547 dead_pseudo_p (rtx x, rtx_insn *insn)
4549 int i, j;
4550 const char *fmt;
4551 enum rtx_code code;
4553 if (REG_P (x))
4554 return (insn != NULL_RTX
4555 && find_regno_note (insn, REG_DEAD, REGNO (x)) != NULL_RTX);
4556 code = GET_CODE (x);
4557 fmt = GET_RTX_FORMAT (code);
4558 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4560 if (fmt[i] == 'e')
4562 if (dead_pseudo_p (XEXP (x, i), insn))
4563 return true;
4565 else if (fmt[i] == 'E')
4567 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4568 if (dead_pseudo_p (XVECEXP (x, i, j), insn))
4569 return true;
4572 return false;
4575 /* Return true if INSN contains a dying pseudo in INSN right hand
4576 side. */
4577 static bool
4578 insn_rhs_dead_pseudo_p (rtx_insn *insn)
4580 rtx set = single_set (insn);
4582 gcc_assert (set != NULL);
4583 return dead_pseudo_p (SET_SRC (set), insn);
4586 /* Return true if any init insn of REGNO contains a dying pseudo in
4587 insn right hand side. */
4588 static bool
4589 init_insn_rhs_dead_pseudo_p (int regno)
4591 rtx_insn_list *insns = ira_reg_equiv[regno].init_insns;
4593 if (insns == NULL)
4594 return false;
4595 for (; insns != NULL_RTX; insns = insns->next ())
4596 if (insn_rhs_dead_pseudo_p (insns->insn ()))
4597 return true;
4598 return false;
4601 /* Return TRUE if REGNO has a reverse equivalence. The equivalence is
4602 reverse only if we have one init insn with given REGNO as a
4603 source. */
4604 static bool
4605 reverse_equiv_p (int regno)
4607 rtx_insn_list *insns = ira_reg_equiv[regno].init_insns;
4608 rtx set;
4610 if (insns == NULL)
4611 return false;
4612 if (! INSN_P (insns->insn ())
4613 || insns->next () != NULL)
4614 return false;
4615 if ((set = single_set (insns->insn ())) == NULL_RTX)
4616 return false;
4617 return REG_P (SET_SRC (set)) && (int) REGNO (SET_SRC (set)) == regno;
4620 /* Return TRUE if REGNO was reloaded in an equivalence init insn. We
4621 call this function only for non-reverse equivalence. */
4622 static bool
4623 contains_reloaded_insn_p (int regno)
4625 rtx set;
4626 rtx_insn_list *list = ira_reg_equiv[regno].init_insns;
4628 for (; list != NULL; list = list->next ())
4629 if ((set = single_set (list->insn ())) == NULL_RTX
4630 || ! REG_P (SET_DEST (set))
4631 || (int) REGNO (SET_DEST (set)) != regno)
4632 return true;
4633 return false;
4636 /* Entry function of LRA constraint pass. Return true if the
4637 constraint pass did change the code. */
4638 bool
4639 lra_constraints (bool first_p)
4641 bool changed_p;
4642 int i, hard_regno, new_insns_num;
4643 unsigned int min_len, new_min_len, uid;
4644 rtx set, x, reg, dest_reg;
4645 basic_block last_bb;
4646 bitmap_iterator bi;
4648 lra_constraint_iter++;
4649 if (lra_dump_file != NULL)
4650 fprintf (lra_dump_file, "\n********** Local #%d: **********\n\n",
4651 lra_constraint_iter);
4652 changed_p = false;
4653 if (pic_offset_table_rtx
4654 && REGNO (pic_offset_table_rtx) >= FIRST_PSEUDO_REGISTER)
4655 lra_risky_transformations_p = true;
4656 else
4657 /* On the first iteration we should check IRA assignment
4658 correctness. In rare cases, the assignments can be wrong as
4659 early clobbers operands are ignored in IRA. */
4660 lra_risky_transformations_p = first_p;
4661 new_insn_uid_start = get_max_uid ();
4662 new_regno_start = first_p ? lra_constraint_new_regno_start : max_reg_num ();
4663 /* Mark used hard regs for target stack size calulations. */
4664 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4665 if (lra_reg_info[i].nrefs != 0
4666 && (hard_regno = lra_get_regno_hard_regno (i)) >= 0)
4668 int j, nregs;
4670 nregs = hard_regno_nregs (hard_regno, lra_reg_info[i].biggest_mode);
4671 for (j = 0; j < nregs; j++)
4672 df_set_regs_ever_live (hard_regno + j, true);
4674 /* Do elimination before the equivalence processing as we can spill
4675 some pseudos during elimination. */
4676 lra_eliminate (false, first_p);
4677 auto_bitmap equiv_insn_bitmap (&reg_obstack);
4678 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4679 if (lra_reg_info[i].nrefs != 0)
4681 ira_reg_equiv[i].profitable_p = true;
4682 reg = regno_reg_rtx[i];
4683 if (lra_get_regno_hard_regno (i) < 0 && (x = get_equiv (reg)) != reg)
4685 bool pseudo_p = contains_reg_p (x, false, false);
4687 /* After RTL transformation, we can not guarantee that
4688 pseudo in the substitution was not reloaded which might
4689 make equivalence invalid. For example, in reverse
4690 equiv of p0
4692 p0 <- ...
4694 equiv_mem <- p0
4696 the memory address register was reloaded before the 2nd
4697 insn. */
4698 if ((! first_p && pseudo_p)
4699 /* We don't use DF for compilation speed sake. So it
4700 is problematic to update live info when we use an
4701 equivalence containing pseudos in more than one
4702 BB. */
4703 || (pseudo_p && multi_block_pseudo_p (i))
4704 /* If an init insn was deleted for some reason, cancel
4705 the equiv. We could update the equiv insns after
4706 transformations including an equiv insn deletion
4707 but it is not worthy as such cases are extremely
4708 rare. */
4709 || contains_deleted_insn_p (ira_reg_equiv[i].init_insns)
4710 /* If it is not a reverse equivalence, we check that a
4711 pseudo in rhs of the init insn is not dying in the
4712 insn. Otherwise, the live info at the beginning of
4713 the corresponding BB might be wrong after we
4714 removed the insn. When the equiv can be a
4715 constant, the right hand side of the init insn can
4716 be a pseudo. */
4717 || (! reverse_equiv_p (i)
4718 && (init_insn_rhs_dead_pseudo_p (i)
4719 /* If we reloaded the pseudo in an equivalence
4720 init insn, we can not remove the equiv init
4721 insns and the init insns might write into
4722 const memory in this case. */
4723 || contains_reloaded_insn_p (i)))
4724 /* Prevent access beyond equivalent memory for
4725 paradoxical subregs. */
4726 || (MEM_P (x)
4727 && (GET_MODE_SIZE (lra_reg_info[i].biggest_mode)
4728 > GET_MODE_SIZE (GET_MODE (x))))
4729 || (pic_offset_table_rtx
4730 && ((CONST_POOL_OK_P (PSEUDO_REGNO_MODE (i), x)
4731 && (targetm.preferred_reload_class
4732 (x, lra_get_allocno_class (i)) == NO_REGS))
4733 || contains_symbol_ref_p (x))))
4734 ira_reg_equiv[i].defined_p = false;
4735 if (contains_reg_p (x, false, true))
4736 ira_reg_equiv[i].profitable_p = false;
4737 if (get_equiv (reg) != reg)
4738 bitmap_ior_into (equiv_insn_bitmap, &lra_reg_info[i].insn_bitmap);
4741 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4742 update_equiv (i);
4743 /* We should add all insns containing pseudos which should be
4744 substituted by their equivalences. */
4745 EXECUTE_IF_SET_IN_BITMAP (equiv_insn_bitmap, 0, uid, bi)
4746 lra_push_insn_by_uid (uid);
4747 min_len = lra_insn_stack_length ();
4748 new_insns_num = 0;
4749 last_bb = NULL;
4750 changed_p = false;
4751 while ((new_min_len = lra_insn_stack_length ()) != 0)
4753 curr_insn = lra_pop_insn ();
4754 --new_min_len;
4755 curr_bb = BLOCK_FOR_INSN (curr_insn);
4756 if (curr_bb != last_bb)
4758 last_bb = curr_bb;
4759 bb_reload_num = lra_curr_reload_num;
4761 if (min_len > new_min_len)
4763 min_len = new_min_len;
4764 new_insns_num = 0;
4766 if (new_insns_num > MAX_RELOAD_INSNS_NUMBER)
4767 internal_error
4768 ("Max. number of generated reload insns per insn is achieved (%d)\n",
4769 MAX_RELOAD_INSNS_NUMBER);
4770 new_insns_num++;
4771 if (DEBUG_INSN_P (curr_insn))
4773 /* We need to check equivalence in debug insn and change
4774 pseudo to the equivalent value if necessary. */
4775 curr_id = lra_get_insn_recog_data (curr_insn);
4776 if (bitmap_bit_p (equiv_insn_bitmap, INSN_UID (curr_insn)))
4778 rtx old = *curr_id->operand_loc[0];
4779 *curr_id->operand_loc[0]
4780 = simplify_replace_fn_rtx (old, NULL_RTX,
4781 loc_equivalence_callback, curr_insn);
4782 if (old != *curr_id->operand_loc[0])
4784 lra_update_insn_regno_info (curr_insn);
4785 changed_p = true;
4789 else if (INSN_P (curr_insn))
4791 if ((set = single_set (curr_insn)) != NULL_RTX)
4793 dest_reg = SET_DEST (set);
4794 /* The equivalence pseudo could be set up as SUBREG in a
4795 case when it is a call restore insn in a mode
4796 different from the pseudo mode. */
4797 if (GET_CODE (dest_reg) == SUBREG)
4798 dest_reg = SUBREG_REG (dest_reg);
4799 if ((REG_P (dest_reg)
4800 && (x = get_equiv (dest_reg)) != dest_reg
4801 /* Remove insns which set up a pseudo whose value
4802 can not be changed. Such insns might be not in
4803 init_insns because we don't update equiv data
4804 during insn transformations.
4806 As an example, let suppose that a pseudo got
4807 hard register and on the 1st pass was not
4808 changed to equivalent constant. We generate an
4809 additional insn setting up the pseudo because of
4810 secondary memory movement. Then the pseudo is
4811 spilled and we use the equiv constant. In this
4812 case we should remove the additional insn and
4813 this insn is not init_insns list. */
4814 && (! MEM_P (x) || MEM_READONLY_P (x)
4815 /* Check that this is actually an insn setting
4816 up the equivalence. */
4817 || in_list_p (curr_insn,
4818 ira_reg_equiv
4819 [REGNO (dest_reg)].init_insns)))
4820 || (((x = get_equiv (SET_SRC (set))) != SET_SRC (set))
4821 && in_list_p (curr_insn,
4822 ira_reg_equiv
4823 [REGNO (SET_SRC (set))].init_insns)))
4825 /* This is equiv init insn of pseudo which did not get a
4826 hard register -- remove the insn. */
4827 if (lra_dump_file != NULL)
4829 fprintf (lra_dump_file,
4830 " Removing equiv init insn %i (freq=%d)\n",
4831 INSN_UID (curr_insn),
4832 REG_FREQ_FROM_BB (BLOCK_FOR_INSN (curr_insn)));
4833 dump_insn_slim (lra_dump_file, curr_insn);
4835 if (contains_reg_p (x, true, false))
4836 lra_risky_transformations_p = true;
4837 lra_set_insn_deleted (curr_insn);
4838 continue;
4841 curr_id = lra_get_insn_recog_data (curr_insn);
4842 curr_static_id = curr_id->insn_static_data;
4843 init_curr_insn_input_reloads ();
4844 init_curr_operand_mode ();
4845 if (curr_insn_transform (false))
4846 changed_p = true;
4847 /* Check non-transformed insns too for equiv change as USE
4848 or CLOBBER don't need reloads but can contain pseudos
4849 being changed on their equivalences. */
4850 else if (bitmap_bit_p (equiv_insn_bitmap, INSN_UID (curr_insn))
4851 && loc_equivalence_change_p (&PATTERN (curr_insn)))
4853 lra_update_insn_regno_info (curr_insn);
4854 changed_p = true;
4859 /* If we used a new hard regno, changed_p should be true because the
4860 hard reg is assigned to a new pseudo. */
4861 if (flag_checking && !changed_p)
4863 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4864 if (lra_reg_info[i].nrefs != 0
4865 && (hard_regno = lra_get_regno_hard_regno (i)) >= 0)
4867 int j, nregs = hard_regno_nregs (hard_regno,
4868 PSEUDO_REGNO_MODE (i));
4870 for (j = 0; j < nregs; j++)
4871 lra_assert (df_regs_ever_live_p (hard_regno + j));
4874 return changed_p;
4877 static void initiate_invariants (void);
4878 static void finish_invariants (void);
4880 /* Initiate the LRA constraint pass. It is done once per
4881 function. */
4882 void
4883 lra_constraints_init (void)
4885 initiate_invariants ();
4888 /* Finalize the LRA constraint pass. It is done once per
4889 function. */
4890 void
4891 lra_constraints_finish (void)
4893 finish_invariants ();
4898 /* Structure describes invariants for ineheritance. */
4899 struct lra_invariant
4901 /* The order number of the invariant. */
4902 int num;
4903 /* The invariant RTX. */
4904 rtx invariant_rtx;
4905 /* The origin insn of the invariant. */
4906 rtx_insn *insn;
4909 typedef lra_invariant invariant_t;
4910 typedef invariant_t *invariant_ptr_t;
4911 typedef const invariant_t *const_invariant_ptr_t;
4913 /* Pointer to the inheritance invariants. */
4914 static vec<invariant_ptr_t> invariants;
4916 /* Allocation pool for the invariants. */
4917 static object_allocator<lra_invariant> *invariants_pool;
4919 /* Hash table for the invariants. */
4920 static htab_t invariant_table;
4922 /* Hash function for INVARIANT. */
4923 static hashval_t
4924 invariant_hash (const void *invariant)
4926 rtx inv = ((const_invariant_ptr_t) invariant)->invariant_rtx;
4927 return lra_rtx_hash (inv);
4930 /* Equal function for invariants INVARIANT1 and INVARIANT2. */
4931 static int
4932 invariant_eq_p (const void *invariant1, const void *invariant2)
4934 rtx inv1 = ((const_invariant_ptr_t) invariant1)->invariant_rtx;
4935 rtx inv2 = ((const_invariant_ptr_t) invariant2)->invariant_rtx;
4937 return rtx_equal_p (inv1, inv2);
4940 /* Insert INVARIANT_RTX into the table if it is not there yet. Return
4941 invariant which is in the table. */
4942 static invariant_ptr_t
4943 insert_invariant (rtx invariant_rtx)
4945 void **entry_ptr;
4946 invariant_t invariant;
4947 invariant_ptr_t invariant_ptr;
4949 invariant.invariant_rtx = invariant_rtx;
4950 entry_ptr = htab_find_slot (invariant_table, &invariant, INSERT);
4951 if (*entry_ptr == NULL)
4953 invariant_ptr = invariants_pool->allocate ();
4954 invariant_ptr->invariant_rtx = invariant_rtx;
4955 invariant_ptr->insn = NULL;
4956 invariants.safe_push (invariant_ptr);
4957 *entry_ptr = (void *) invariant_ptr;
4959 return (invariant_ptr_t) *entry_ptr;
4962 /* Initiate the invariant table. */
4963 static void
4964 initiate_invariants (void)
4966 invariants.create (100);
4967 invariants_pool
4968 = new object_allocator<lra_invariant> ("Inheritance invariants");
4969 invariant_table = htab_create (100, invariant_hash, invariant_eq_p, NULL);
4972 /* Finish the invariant table. */
4973 static void
4974 finish_invariants (void)
4976 htab_delete (invariant_table);
4977 delete invariants_pool;
4978 invariants.release ();
4981 /* Make the invariant table empty. */
4982 static void
4983 clear_invariants (void)
4985 htab_empty (invariant_table);
4986 invariants_pool->release ();
4987 invariants.truncate (0);
4992 /* This page contains code to do inheritance/split
4993 transformations. */
4995 /* Number of reloads passed so far in current EBB. */
4996 static int reloads_num;
4998 /* Number of calls passed so far in current EBB. */
4999 static int calls_num;
5001 /* Current reload pseudo check for validity of elements in
5002 USAGE_INSNS. */
5003 static int curr_usage_insns_check;
5005 /* Info about last usage of registers in EBB to do inheritance/split
5006 transformation. Inheritance transformation is done from a spilled
5007 pseudo and split transformations from a hard register or a pseudo
5008 assigned to a hard register. */
5009 struct usage_insns
5011 /* If the value is equal to CURR_USAGE_INSNS_CHECK, then the member
5012 value INSNS is valid. The insns is chain of optional debug insns
5013 and a finishing non-debug insn using the corresponding reg. The
5014 value is also used to mark the registers which are set up in the
5015 current insn. The negated insn uid is used for this. */
5016 int check;
5017 /* Value of global reloads_num at the last insn in INSNS. */
5018 int reloads_num;
5019 /* Value of global reloads_nums at the last insn in INSNS. */
5020 int calls_num;
5021 /* It can be true only for splitting. And it means that the restore
5022 insn should be put after insn given by the following member. */
5023 bool after_p;
5024 /* Next insns in the current EBB which use the original reg and the
5025 original reg value is not changed between the current insn and
5026 the next insns. In order words, e.g. for inheritance, if we need
5027 to use the original reg value again in the next insns we can try
5028 to use the value in a hard register from a reload insn of the
5029 current insn. */
5030 rtx insns;
5033 /* Map: regno -> corresponding pseudo usage insns. */
5034 static struct usage_insns *usage_insns;
5036 static void
5037 setup_next_usage_insn (int regno, rtx insn, int reloads_num, bool after_p)
5039 usage_insns[regno].check = curr_usage_insns_check;
5040 usage_insns[regno].insns = insn;
5041 usage_insns[regno].reloads_num = reloads_num;
5042 usage_insns[regno].calls_num = calls_num;
5043 usage_insns[regno].after_p = after_p;
5046 /* The function is used to form list REGNO usages which consists of
5047 optional debug insns finished by a non-debug insn using REGNO.
5048 RELOADS_NUM is current number of reload insns processed so far. */
5049 static void
5050 add_next_usage_insn (int regno, rtx_insn *insn, int reloads_num)
5052 rtx next_usage_insns;
5054 if (usage_insns[regno].check == curr_usage_insns_check
5055 && (next_usage_insns = usage_insns[regno].insns) != NULL_RTX
5056 && DEBUG_INSN_P (insn))
5058 /* Check that we did not add the debug insn yet. */
5059 if (next_usage_insns != insn
5060 && (GET_CODE (next_usage_insns) != INSN_LIST
5061 || XEXP (next_usage_insns, 0) != insn))
5062 usage_insns[regno].insns = gen_rtx_INSN_LIST (VOIDmode, insn,
5063 next_usage_insns);
5065 else if (NONDEBUG_INSN_P (insn))
5066 setup_next_usage_insn (regno, insn, reloads_num, false);
5067 else
5068 usage_insns[regno].check = 0;
5071 /* Return first non-debug insn in list USAGE_INSNS. */
5072 static rtx_insn *
5073 skip_usage_debug_insns (rtx usage_insns)
5075 rtx insn;
5077 /* Skip debug insns. */
5078 for (insn = usage_insns;
5079 insn != NULL_RTX && GET_CODE (insn) == INSN_LIST;
5080 insn = XEXP (insn, 1))
5082 return safe_as_a <rtx_insn *> (insn);
5085 /* Return true if we need secondary memory moves for insn in
5086 USAGE_INSNS after inserting inherited pseudo of class INHER_CL
5087 into the insn. */
5088 static bool
5089 check_secondary_memory_needed_p (enum reg_class inher_cl ATTRIBUTE_UNUSED,
5090 rtx usage_insns ATTRIBUTE_UNUSED)
5092 rtx_insn *insn;
5093 rtx set, dest;
5094 enum reg_class cl;
5096 if (inher_cl == ALL_REGS
5097 || (insn = skip_usage_debug_insns (usage_insns)) == NULL_RTX)
5098 return false;
5099 lra_assert (INSN_P (insn));
5100 if ((set = single_set (insn)) == NULL_RTX || ! REG_P (SET_DEST (set)))
5101 return false;
5102 dest = SET_DEST (set);
5103 if (! REG_P (dest))
5104 return false;
5105 lra_assert (inher_cl != NO_REGS);
5106 cl = get_reg_class (REGNO (dest));
5107 return (cl != NO_REGS && cl != ALL_REGS
5108 && targetm.secondary_memory_needed (GET_MODE (dest), inher_cl, cl));
5111 /* Registers involved in inheritance/split in the current EBB
5112 (inheritance/split pseudos and original registers). */
5113 static bitmap_head check_only_regs;
5115 /* Reload pseudos can not be involded in invariant inheritance in the
5116 current EBB. */
5117 static bitmap_head invalid_invariant_regs;
5119 /* Do inheritance transformations for insn INSN, which defines (if
5120 DEF_P) or uses ORIGINAL_REGNO. NEXT_USAGE_INSNS specifies which
5121 instruction in the EBB next uses ORIGINAL_REGNO; it has the same
5122 form as the "insns" field of usage_insns. Return true if we
5123 succeed in such transformation.
5125 The transformations look like:
5127 p <- ... i <- ...
5128 ... p <- i (new insn)
5129 ... =>
5130 <- ... p ... <- ... i ...
5132 ... i <- p (new insn)
5133 <- ... p ... <- ... i ...
5134 ... =>
5135 <- ... p ... <- ... i ...
5136 where p is a spilled original pseudo and i is a new inheritance pseudo.
5139 The inheritance pseudo has the smallest class of two classes CL and
5140 class of ORIGINAL REGNO. */
5141 static bool
5142 inherit_reload_reg (bool def_p, int original_regno,
5143 enum reg_class cl, rtx_insn *insn, rtx next_usage_insns)
5145 if (optimize_function_for_size_p (cfun))
5146 return false;
5148 enum reg_class rclass = lra_get_allocno_class (original_regno);
5149 rtx original_reg = regno_reg_rtx[original_regno];
5150 rtx new_reg, usage_insn;
5151 rtx_insn *new_insns;
5153 lra_assert (! usage_insns[original_regno].after_p);
5154 if (lra_dump_file != NULL)
5155 fprintf (lra_dump_file,
5156 " <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<\n");
5157 if (! ira_reg_classes_intersect_p[cl][rclass])
5159 if (lra_dump_file != NULL)
5161 fprintf (lra_dump_file,
5162 " Rejecting inheritance for %d "
5163 "because of disjoint classes %s and %s\n",
5164 original_regno, reg_class_names[cl],
5165 reg_class_names[rclass]);
5166 fprintf (lra_dump_file,
5167 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5169 return false;
5171 if ((ira_class_subset_p[cl][rclass] && cl != rclass)
5172 /* We don't use a subset of two classes because it can be
5173 NO_REGS. This transformation is still profitable in most
5174 cases even if the classes are not intersected as register
5175 move is probably cheaper than a memory load. */
5176 || ira_class_hard_regs_num[cl] < ira_class_hard_regs_num[rclass])
5178 if (lra_dump_file != NULL)
5179 fprintf (lra_dump_file, " Use smallest class of %s and %s\n",
5180 reg_class_names[cl], reg_class_names[rclass]);
5182 rclass = cl;
5184 if (check_secondary_memory_needed_p (rclass, next_usage_insns))
5186 /* Reject inheritance resulting in secondary memory moves.
5187 Otherwise, there is a danger in LRA cycling. Also such
5188 transformation will be unprofitable. */
5189 if (lra_dump_file != NULL)
5191 rtx_insn *insn = skip_usage_debug_insns (next_usage_insns);
5192 rtx set = single_set (insn);
5194 lra_assert (set != NULL_RTX);
5196 rtx dest = SET_DEST (set);
5198 lra_assert (REG_P (dest));
5199 fprintf (lra_dump_file,
5200 " Rejecting inheritance for insn %d(%s)<-%d(%s) "
5201 "as secondary mem is needed\n",
5202 REGNO (dest), reg_class_names[get_reg_class (REGNO (dest))],
5203 original_regno, reg_class_names[rclass]);
5204 fprintf (lra_dump_file,
5205 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5207 return false;
5209 new_reg = lra_create_new_reg (GET_MODE (original_reg), original_reg,
5210 rclass, "inheritance");
5211 start_sequence ();
5212 if (def_p)
5213 lra_emit_move (original_reg, new_reg);
5214 else
5215 lra_emit_move (new_reg, original_reg);
5216 new_insns = get_insns ();
5217 end_sequence ();
5218 if (NEXT_INSN (new_insns) != NULL_RTX)
5220 if (lra_dump_file != NULL)
5222 fprintf (lra_dump_file,
5223 " Rejecting inheritance %d->%d "
5224 "as it results in 2 or more insns:\n",
5225 original_regno, REGNO (new_reg));
5226 dump_rtl_slim (lra_dump_file, new_insns, NULL, -1, 0);
5227 fprintf (lra_dump_file,
5228 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5230 return false;
5232 lra_substitute_pseudo_within_insn (insn, original_regno, new_reg, false);
5233 lra_update_insn_regno_info (insn);
5234 if (! def_p)
5235 /* We now have a new usage insn for original regno. */
5236 setup_next_usage_insn (original_regno, new_insns, reloads_num, false);
5237 if (lra_dump_file != NULL)
5238 fprintf (lra_dump_file, " Original reg change %d->%d (bb%d):\n",
5239 original_regno, REGNO (new_reg), BLOCK_FOR_INSN (insn)->index);
5240 lra_reg_info[REGNO (new_reg)].restore_rtx = regno_reg_rtx[original_regno];
5241 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
5242 bitmap_set_bit (&check_only_regs, original_regno);
5243 bitmap_set_bit (&lra_inheritance_pseudos, REGNO (new_reg));
5244 if (def_p)
5245 lra_process_new_insns (insn, NULL, new_insns,
5246 "Add original<-inheritance");
5247 else
5248 lra_process_new_insns (insn, new_insns, NULL,
5249 "Add inheritance<-original");
5250 while (next_usage_insns != NULL_RTX)
5252 if (GET_CODE (next_usage_insns) != INSN_LIST)
5254 usage_insn = next_usage_insns;
5255 lra_assert (NONDEBUG_INSN_P (usage_insn));
5256 next_usage_insns = NULL;
5258 else
5260 usage_insn = XEXP (next_usage_insns, 0);
5261 lra_assert (DEBUG_INSN_P (usage_insn));
5262 next_usage_insns = XEXP (next_usage_insns, 1);
5264 lra_substitute_pseudo (&usage_insn, original_regno, new_reg, false);
5265 lra_update_insn_regno_info (as_a <rtx_insn *> (usage_insn));
5266 if (lra_dump_file != NULL)
5268 basic_block bb = BLOCK_FOR_INSN (usage_insn);
5269 fprintf (lra_dump_file,
5270 " Inheritance reuse change %d->%d (bb%d):\n",
5271 original_regno, REGNO (new_reg),
5272 bb ? bb->index : -1);
5273 dump_insn_slim (lra_dump_file, as_a <rtx_insn *> (usage_insn));
5276 if (lra_dump_file != NULL)
5277 fprintf (lra_dump_file,
5278 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5279 return true;
5282 /* Return true if we need a caller save/restore for pseudo REGNO which
5283 was assigned to a hard register. */
5284 static inline bool
5285 need_for_call_save_p (int regno)
5287 lra_assert (regno >= FIRST_PSEUDO_REGISTER && reg_renumber[regno] >= 0);
5288 return (usage_insns[regno].calls_num < calls_num
5289 && (overlaps_hard_reg_set_p
5290 ((flag_ipa_ra &&
5291 ! hard_reg_set_empty_p (lra_reg_info[regno].actual_call_used_reg_set))
5292 ? lra_reg_info[regno].actual_call_used_reg_set
5293 : call_used_reg_set,
5294 PSEUDO_REGNO_MODE (regno), reg_renumber[regno])
5295 || (targetm.hard_regno_call_part_clobbered
5296 (reg_renumber[regno], PSEUDO_REGNO_MODE (regno)))));
5299 /* Global registers occurring in the current EBB. */
5300 static bitmap_head ebb_global_regs;
5302 /* Return true if we need a split for hard register REGNO or pseudo
5303 REGNO which was assigned to a hard register.
5304 POTENTIAL_RELOAD_HARD_REGS contains hard registers which might be
5305 used for reloads since the EBB end. It is an approximation of the
5306 used hard registers in the split range. The exact value would
5307 require expensive calculations. If we were aggressive with
5308 splitting because of the approximation, the split pseudo will save
5309 the same hard register assignment and will be removed in the undo
5310 pass. We still need the approximation because too aggressive
5311 splitting would result in too inaccurate cost calculation in the
5312 assignment pass because of too many generated moves which will be
5313 probably removed in the undo pass. */
5314 static inline bool
5315 need_for_split_p (HARD_REG_SET potential_reload_hard_regs, int regno)
5317 int hard_regno = regno < FIRST_PSEUDO_REGISTER ? regno : reg_renumber[regno];
5319 lra_assert (hard_regno >= 0);
5320 return ((TEST_HARD_REG_BIT (potential_reload_hard_regs, hard_regno)
5321 /* Don't split eliminable hard registers, otherwise we can
5322 split hard registers like hard frame pointer, which
5323 lives on BB start/end according to DF-infrastructure,
5324 when there is a pseudo assigned to the register and
5325 living in the same BB. */
5326 && (regno >= FIRST_PSEUDO_REGISTER
5327 || ! TEST_HARD_REG_BIT (eliminable_regset, hard_regno))
5328 && ! TEST_HARD_REG_BIT (lra_no_alloc_regs, hard_regno)
5329 /* Don't split call clobbered hard regs living through
5330 calls, otherwise we might have a check problem in the
5331 assign sub-pass as in the most cases (exception is a
5332 situation when lra_risky_transformations_p value is
5333 true) the assign pass assumes that all pseudos living
5334 through calls are assigned to call saved hard regs. */
5335 && (regno >= FIRST_PSEUDO_REGISTER
5336 || ! TEST_HARD_REG_BIT (call_used_reg_set, regno)
5337 || usage_insns[regno].calls_num == calls_num)
5338 /* We need at least 2 reloads to make pseudo splitting
5339 profitable. We should provide hard regno splitting in
5340 any case to solve 1st insn scheduling problem when
5341 moving hard register definition up might result in
5342 impossibility to find hard register for reload pseudo of
5343 small register class. */
5344 && (usage_insns[regno].reloads_num
5345 + (regno < FIRST_PSEUDO_REGISTER ? 0 : 3) < reloads_num)
5346 && (regno < FIRST_PSEUDO_REGISTER
5347 /* For short living pseudos, spilling + inheritance can
5348 be considered a substitution for splitting.
5349 Therefore we do not splitting for local pseudos. It
5350 decreases also aggressiveness of splitting. The
5351 minimal number of references is chosen taking into
5352 account that for 2 references splitting has no sense
5353 as we can just spill the pseudo. */
5354 || (regno >= FIRST_PSEUDO_REGISTER
5355 && lra_reg_info[regno].nrefs > 3
5356 && bitmap_bit_p (&ebb_global_regs, regno))))
5357 || (regno >= FIRST_PSEUDO_REGISTER && need_for_call_save_p (regno)));
5360 /* Return class for the split pseudo created from original pseudo with
5361 ALLOCNO_CLASS and MODE which got a hard register HARD_REGNO. We
5362 choose subclass of ALLOCNO_CLASS which contains HARD_REGNO and
5363 results in no secondary memory movements. */
5364 static enum reg_class
5365 choose_split_class (enum reg_class allocno_class,
5366 int hard_regno ATTRIBUTE_UNUSED,
5367 machine_mode mode ATTRIBUTE_UNUSED)
5369 int i;
5370 enum reg_class cl, best_cl = NO_REGS;
5371 enum reg_class hard_reg_class ATTRIBUTE_UNUSED
5372 = REGNO_REG_CLASS (hard_regno);
5374 if (! targetm.secondary_memory_needed (mode, allocno_class, allocno_class)
5375 && TEST_HARD_REG_BIT (reg_class_contents[allocno_class], hard_regno))
5376 return allocno_class;
5377 for (i = 0;
5378 (cl = reg_class_subclasses[allocno_class][i]) != LIM_REG_CLASSES;
5379 i++)
5380 if (! targetm.secondary_memory_needed (mode, cl, hard_reg_class)
5381 && ! targetm.secondary_memory_needed (mode, hard_reg_class, cl)
5382 && TEST_HARD_REG_BIT (reg_class_contents[cl], hard_regno)
5383 && (best_cl == NO_REGS
5384 || ira_class_hard_regs_num[best_cl] < ira_class_hard_regs_num[cl]))
5385 best_cl = cl;
5386 return best_cl;
5389 /* Copy any equivalence information from ORIGINAL_REGNO to NEW_REGNO.
5390 It only makes sense to call this function if NEW_REGNO is always
5391 equal to ORIGINAL_REGNO. */
5393 static void
5394 lra_copy_reg_equiv (unsigned int new_regno, unsigned int original_regno)
5396 if (!ira_reg_equiv[original_regno].defined_p)
5397 return;
5399 ira_expand_reg_equiv ();
5400 ira_reg_equiv[new_regno].defined_p = true;
5401 if (ira_reg_equiv[original_regno].memory)
5402 ira_reg_equiv[new_regno].memory
5403 = copy_rtx (ira_reg_equiv[original_regno].memory);
5404 if (ira_reg_equiv[original_regno].constant)
5405 ira_reg_equiv[new_regno].constant
5406 = copy_rtx (ira_reg_equiv[original_regno].constant);
5407 if (ira_reg_equiv[original_regno].invariant)
5408 ira_reg_equiv[new_regno].invariant
5409 = copy_rtx (ira_reg_equiv[original_regno].invariant);
5412 /* Do split transformations for insn INSN, which defines or uses
5413 ORIGINAL_REGNO. NEXT_USAGE_INSNS specifies which instruction in
5414 the EBB next uses ORIGINAL_REGNO; it has the same form as the
5415 "insns" field of usage_insns.
5417 The transformations look like:
5419 p <- ... p <- ...
5420 ... s <- p (new insn -- save)
5421 ... =>
5422 ... p <- s (new insn -- restore)
5423 <- ... p ... <- ... p ...
5425 <- ... p ... <- ... p ...
5426 ... s <- p (new insn -- save)
5427 ... =>
5428 ... p <- s (new insn -- restore)
5429 <- ... p ... <- ... p ...
5431 where p is an original pseudo got a hard register or a hard
5432 register and s is a new split pseudo. The save is put before INSN
5433 if BEFORE_P is true. Return true if we succeed in such
5434 transformation. */
5435 static bool
5436 split_reg (bool before_p, int original_regno, rtx_insn *insn,
5437 rtx next_usage_insns)
5439 enum reg_class rclass;
5440 rtx original_reg;
5441 int hard_regno, nregs;
5442 rtx new_reg, usage_insn;
5443 rtx_insn *restore, *save;
5444 bool after_p;
5445 bool call_save_p;
5446 machine_mode mode;
5448 if (original_regno < FIRST_PSEUDO_REGISTER)
5450 rclass = ira_allocno_class_translate[REGNO_REG_CLASS (original_regno)];
5451 hard_regno = original_regno;
5452 call_save_p = false;
5453 nregs = 1;
5454 mode = lra_reg_info[hard_regno].biggest_mode;
5455 machine_mode reg_rtx_mode = GET_MODE (regno_reg_rtx[hard_regno]);
5456 /* A reg can have a biggest_mode of VOIDmode if it was only ever seen
5457 as part of a multi-word register. In that case, or if the biggest
5458 mode was larger than a register, just use the reg_rtx. Otherwise,
5459 limit the size to that of the biggest access in the function. */
5460 if (mode == VOIDmode
5461 || paradoxical_subreg_p (mode, reg_rtx_mode))
5463 original_reg = regno_reg_rtx[hard_regno];
5464 mode = reg_rtx_mode;
5466 else
5467 original_reg = gen_rtx_REG (mode, hard_regno);
5469 else
5471 mode = PSEUDO_REGNO_MODE (original_regno);
5472 hard_regno = reg_renumber[original_regno];
5473 nregs = hard_regno_nregs (hard_regno, mode);
5474 rclass = lra_get_allocno_class (original_regno);
5475 original_reg = regno_reg_rtx[original_regno];
5476 call_save_p = need_for_call_save_p (original_regno);
5478 lra_assert (hard_regno >= 0);
5479 if (lra_dump_file != NULL)
5480 fprintf (lra_dump_file,
5481 " ((((((((((((((((((((((((((((((((((((((((((((((((\n");
5483 if (call_save_p)
5485 mode = HARD_REGNO_CALLER_SAVE_MODE (hard_regno,
5486 hard_regno_nregs (hard_regno, mode),
5487 mode);
5488 new_reg = lra_create_new_reg (mode, NULL_RTX, NO_REGS, "save");
5490 else
5492 rclass = choose_split_class (rclass, hard_regno, mode);
5493 if (rclass == NO_REGS)
5495 if (lra_dump_file != NULL)
5497 fprintf (lra_dump_file,
5498 " Rejecting split of %d(%s): "
5499 "no good reg class for %d(%s)\n",
5500 original_regno,
5501 reg_class_names[lra_get_allocno_class (original_regno)],
5502 hard_regno,
5503 reg_class_names[REGNO_REG_CLASS (hard_regno)]);
5504 fprintf
5505 (lra_dump_file,
5506 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5508 return false;
5510 /* Split_if_necessary can split hard registers used as part of a
5511 multi-register mode but splits each register individually. The
5512 mode used for each independent register may not be supported
5513 so reject the split. Splitting the wider mode should theoretically
5514 be possible but is not implemented. */
5515 if (!targetm.hard_regno_mode_ok (hard_regno, mode))
5517 if (lra_dump_file != NULL)
5519 fprintf (lra_dump_file,
5520 " Rejecting split of %d(%s): unsuitable mode %s\n",
5521 original_regno,
5522 reg_class_names[lra_get_allocno_class (original_regno)],
5523 GET_MODE_NAME (mode));
5524 fprintf
5525 (lra_dump_file,
5526 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5528 return false;
5530 new_reg = lra_create_new_reg (mode, original_reg, rclass, "split");
5531 reg_renumber[REGNO (new_reg)] = hard_regno;
5533 int new_regno = REGNO (new_reg);
5534 save = emit_spill_move (true, new_reg, original_reg);
5535 if (NEXT_INSN (save) != NULL_RTX && !call_save_p)
5537 if (lra_dump_file != NULL)
5539 fprintf
5540 (lra_dump_file,
5541 " Rejecting split %d->%d resulting in > 2 save insns:\n",
5542 original_regno, new_regno);
5543 dump_rtl_slim (lra_dump_file, save, NULL, -1, 0);
5544 fprintf (lra_dump_file,
5545 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5547 return false;
5549 restore = emit_spill_move (false, new_reg, original_reg);
5550 if (NEXT_INSN (restore) != NULL_RTX && !call_save_p)
5552 if (lra_dump_file != NULL)
5554 fprintf (lra_dump_file,
5555 " Rejecting split %d->%d "
5556 "resulting in > 2 restore insns:\n",
5557 original_regno, new_regno);
5558 dump_rtl_slim (lra_dump_file, restore, NULL, -1, 0);
5559 fprintf (lra_dump_file,
5560 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5562 return false;
5564 /* Transfer equivalence information to the spill register, so that
5565 if we fail to allocate the spill register, we have the option of
5566 rematerializing the original value instead of spilling to the stack. */
5567 if (!HARD_REGISTER_NUM_P (original_regno)
5568 && mode == PSEUDO_REGNO_MODE (original_regno))
5569 lra_copy_reg_equiv (new_regno, original_regno);
5570 after_p = usage_insns[original_regno].after_p;
5571 lra_reg_info[new_regno].restore_rtx = regno_reg_rtx[original_regno];
5572 bitmap_set_bit (&check_only_regs, new_regno);
5573 bitmap_set_bit (&check_only_regs, original_regno);
5574 bitmap_set_bit (&lra_split_regs, new_regno);
5575 for (;;)
5577 if (GET_CODE (next_usage_insns) != INSN_LIST)
5579 usage_insn = next_usage_insns;
5580 break;
5582 usage_insn = XEXP (next_usage_insns, 0);
5583 lra_assert (DEBUG_INSN_P (usage_insn));
5584 next_usage_insns = XEXP (next_usage_insns, 1);
5585 lra_substitute_pseudo (&usage_insn, original_regno, new_reg, false);
5586 lra_update_insn_regno_info (as_a <rtx_insn *> (usage_insn));
5587 if (lra_dump_file != NULL)
5589 fprintf (lra_dump_file, " Split reuse change %d->%d:\n",
5590 original_regno, new_regno);
5591 dump_insn_slim (lra_dump_file, as_a <rtx_insn *> (usage_insn));
5594 lra_assert (NOTE_P (usage_insn) || NONDEBUG_INSN_P (usage_insn));
5595 lra_assert (usage_insn != insn || (after_p && before_p));
5596 lra_process_new_insns (as_a <rtx_insn *> (usage_insn),
5597 after_p ? NULL : restore,
5598 after_p ? restore : NULL,
5599 call_save_p
5600 ? "Add reg<-save" : "Add reg<-split");
5601 lra_process_new_insns (insn, before_p ? save : NULL,
5602 before_p ? NULL : save,
5603 call_save_p
5604 ? "Add save<-reg" : "Add split<-reg");
5605 if (nregs > 1)
5606 /* If we are trying to split multi-register. We should check
5607 conflicts on the next assignment sub-pass. IRA can allocate on
5608 sub-register levels, LRA do this on pseudos level right now and
5609 this discrepancy may create allocation conflicts after
5610 splitting. */
5611 lra_risky_transformations_p = true;
5612 if (lra_dump_file != NULL)
5613 fprintf (lra_dump_file,
5614 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5615 return true;
5618 /* Recognize that we need a split transformation for insn INSN, which
5619 defines or uses REGNO in its insn biggest MODE (we use it only if
5620 REGNO is a hard register). POTENTIAL_RELOAD_HARD_REGS contains
5621 hard registers which might be used for reloads since the EBB end.
5622 Put the save before INSN if BEFORE_P is true. MAX_UID is maximla
5623 uid before starting INSN processing. Return true if we succeed in
5624 such transformation. */
5625 static bool
5626 split_if_necessary (int regno, machine_mode mode,
5627 HARD_REG_SET potential_reload_hard_regs,
5628 bool before_p, rtx_insn *insn, int max_uid)
5630 bool res = false;
5631 int i, nregs = 1;
5632 rtx next_usage_insns;
5634 if (regno < FIRST_PSEUDO_REGISTER)
5635 nregs = hard_regno_nregs (regno, mode);
5636 for (i = 0; i < nregs; i++)
5637 if (usage_insns[regno + i].check == curr_usage_insns_check
5638 && (next_usage_insns = usage_insns[regno + i].insns) != NULL_RTX
5639 /* To avoid processing the register twice or more. */
5640 && ((GET_CODE (next_usage_insns) != INSN_LIST
5641 && INSN_UID (next_usage_insns) < max_uid)
5642 || (GET_CODE (next_usage_insns) == INSN_LIST
5643 && (INSN_UID (XEXP (next_usage_insns, 0)) < max_uid)))
5644 && need_for_split_p (potential_reload_hard_regs, regno + i)
5645 && split_reg (before_p, regno + i, insn, next_usage_insns))
5646 res = true;
5647 return res;
5650 /* Return TRUE if rtx X is considered as an invariant for
5651 inheritance. */
5652 static bool
5653 invariant_p (const_rtx x)
5655 machine_mode mode;
5656 const char *fmt;
5657 enum rtx_code code;
5658 int i, j;
5660 code = GET_CODE (x);
5661 mode = GET_MODE (x);
5662 if (code == SUBREG)
5664 x = SUBREG_REG (x);
5665 code = GET_CODE (x);
5666 mode = wider_subreg_mode (mode, GET_MODE (x));
5669 if (MEM_P (x))
5670 return false;
5672 if (REG_P (x))
5674 int i, nregs, regno = REGNO (x);
5676 if (regno >= FIRST_PSEUDO_REGISTER || regno == STACK_POINTER_REGNUM
5677 || TEST_HARD_REG_BIT (eliminable_regset, regno)
5678 || GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
5679 return false;
5680 nregs = hard_regno_nregs (regno, mode);
5681 for (i = 0; i < nregs; i++)
5682 if (! fixed_regs[regno + i]
5683 /* A hard register may be clobbered in the current insn
5684 but we can ignore this case because if the hard
5685 register is used it should be set somewhere after the
5686 clobber. */
5687 || bitmap_bit_p (&invalid_invariant_regs, regno + i))
5688 return false;
5690 fmt = GET_RTX_FORMAT (code);
5691 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
5693 if (fmt[i] == 'e')
5695 if (! invariant_p (XEXP (x, i)))
5696 return false;
5698 else if (fmt[i] == 'E')
5700 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
5701 if (! invariant_p (XVECEXP (x, i, j)))
5702 return false;
5705 return true;
5708 /* We have 'dest_reg <- invariant'. Let us try to make an invariant
5709 inheritance transformation (using dest_reg instead invariant in a
5710 subsequent insn). */
5711 static bool
5712 process_invariant_for_inheritance (rtx dst_reg, rtx invariant_rtx)
5714 invariant_ptr_t invariant_ptr;
5715 rtx_insn *insn, *new_insns;
5716 rtx insn_set, insn_reg, new_reg;
5717 int insn_regno;
5718 bool succ_p = false;
5719 int dst_regno = REGNO (dst_reg);
5720 machine_mode dst_mode = GET_MODE (dst_reg);
5721 enum reg_class cl = lra_get_allocno_class (dst_regno), insn_reg_cl;
5723 invariant_ptr = insert_invariant (invariant_rtx);
5724 if ((insn = invariant_ptr->insn) != NULL_RTX)
5726 /* We have a subsequent insn using the invariant. */
5727 insn_set = single_set (insn);
5728 lra_assert (insn_set != NULL);
5729 insn_reg = SET_DEST (insn_set);
5730 lra_assert (REG_P (insn_reg));
5731 insn_regno = REGNO (insn_reg);
5732 insn_reg_cl = lra_get_allocno_class (insn_regno);
5734 if (dst_mode == GET_MODE (insn_reg)
5735 /* We should consider only result move reg insns which are
5736 cheap. */
5737 && targetm.register_move_cost (dst_mode, cl, insn_reg_cl) == 2
5738 && targetm.register_move_cost (dst_mode, cl, cl) == 2)
5740 if (lra_dump_file != NULL)
5741 fprintf (lra_dump_file,
5742 " [[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[\n");
5743 new_reg = lra_create_new_reg (dst_mode, dst_reg,
5744 cl, "invariant inheritance");
5745 bitmap_set_bit (&lra_inheritance_pseudos, REGNO (new_reg));
5746 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
5747 lra_reg_info[REGNO (new_reg)].restore_rtx = PATTERN (insn);
5748 start_sequence ();
5749 lra_emit_move (new_reg, dst_reg);
5750 new_insns = get_insns ();
5751 end_sequence ();
5752 lra_process_new_insns (curr_insn, NULL, new_insns,
5753 "Add invariant inheritance<-original");
5754 start_sequence ();
5755 lra_emit_move (SET_DEST (insn_set), new_reg);
5756 new_insns = get_insns ();
5757 end_sequence ();
5758 lra_process_new_insns (insn, NULL, new_insns,
5759 "Changing reload<-inheritance");
5760 lra_set_insn_deleted (insn);
5761 succ_p = true;
5762 if (lra_dump_file != NULL)
5764 fprintf (lra_dump_file,
5765 " Invariant inheritance reuse change %d (bb%d):\n",
5766 REGNO (new_reg), BLOCK_FOR_INSN (insn)->index);
5767 dump_insn_slim (lra_dump_file, insn);
5768 fprintf (lra_dump_file,
5769 " ]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]\n");
5773 invariant_ptr->insn = curr_insn;
5774 return succ_p;
5777 /* Check only registers living at the current program point in the
5778 current EBB. */
5779 static bitmap_head live_regs;
5781 /* Update live info in EBB given by its HEAD and TAIL insns after
5782 inheritance/split transformation. The function removes dead moves
5783 too. */
5784 static void
5785 update_ebb_live_info (rtx_insn *head, rtx_insn *tail)
5787 unsigned int j;
5788 int i, regno;
5789 bool live_p;
5790 rtx_insn *prev_insn;
5791 rtx set;
5792 bool remove_p;
5793 basic_block last_bb, prev_bb, curr_bb;
5794 bitmap_iterator bi;
5795 struct lra_insn_reg *reg;
5796 edge e;
5797 edge_iterator ei;
5799 last_bb = BLOCK_FOR_INSN (tail);
5800 prev_bb = NULL;
5801 for (curr_insn = tail;
5802 curr_insn != PREV_INSN (head);
5803 curr_insn = prev_insn)
5805 prev_insn = PREV_INSN (curr_insn);
5806 /* We need to process empty blocks too. They contain
5807 NOTE_INSN_BASIC_BLOCK referring for the basic block. */
5808 if (NOTE_P (curr_insn) && NOTE_KIND (curr_insn) != NOTE_INSN_BASIC_BLOCK)
5809 continue;
5810 curr_bb = BLOCK_FOR_INSN (curr_insn);
5811 if (!curr_bb)
5813 gcc_assert (DEBUG_INSN_P (curr_insn));
5814 if (DEBUG_MARKER_INSN_P (curr_insn))
5815 continue;
5816 curr_bb = prev_bb;
5818 if (curr_bb != prev_bb)
5820 if (prev_bb != NULL)
5822 /* Update df_get_live_in (prev_bb): */
5823 EXECUTE_IF_SET_IN_BITMAP (&check_only_regs, 0, j, bi)
5824 if (bitmap_bit_p (&live_regs, j))
5825 bitmap_set_bit (df_get_live_in (prev_bb), j);
5826 else
5827 bitmap_clear_bit (df_get_live_in (prev_bb), j);
5829 if (curr_bb != last_bb)
5831 /* Update df_get_live_out (curr_bb): */
5832 EXECUTE_IF_SET_IN_BITMAP (&check_only_regs, 0, j, bi)
5834 live_p = bitmap_bit_p (&live_regs, j);
5835 if (! live_p)
5836 FOR_EACH_EDGE (e, ei, curr_bb->succs)
5837 if (bitmap_bit_p (df_get_live_in (e->dest), j))
5839 live_p = true;
5840 break;
5842 if (live_p)
5843 bitmap_set_bit (df_get_live_out (curr_bb), j);
5844 else
5845 bitmap_clear_bit (df_get_live_out (curr_bb), j);
5848 prev_bb = curr_bb;
5849 bitmap_and (&live_regs, &check_only_regs, df_get_live_out (curr_bb));
5851 if (! NONDEBUG_INSN_P (curr_insn))
5852 continue;
5853 curr_id = lra_get_insn_recog_data (curr_insn);
5854 curr_static_id = curr_id->insn_static_data;
5855 remove_p = false;
5856 if ((set = single_set (curr_insn)) != NULL_RTX
5857 && REG_P (SET_DEST (set))
5858 && (regno = REGNO (SET_DEST (set))) >= FIRST_PSEUDO_REGISTER
5859 && SET_DEST (set) != pic_offset_table_rtx
5860 && bitmap_bit_p (&check_only_regs, regno)
5861 && ! bitmap_bit_p (&live_regs, regno))
5862 remove_p = true;
5863 /* See which defined values die here. */
5864 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5865 if (reg->type == OP_OUT && ! reg->subreg_p)
5866 bitmap_clear_bit (&live_regs, reg->regno);
5867 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
5868 if (reg->type == OP_OUT && ! reg->subreg_p)
5869 bitmap_clear_bit (&live_regs, reg->regno);
5870 if (curr_id->arg_hard_regs != NULL)
5871 /* Make clobbered argument hard registers die. */
5872 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
5873 if (regno >= FIRST_PSEUDO_REGISTER)
5874 bitmap_clear_bit (&live_regs, regno - FIRST_PSEUDO_REGISTER);
5875 /* Mark each used value as live. */
5876 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5877 if (reg->type != OP_OUT
5878 && bitmap_bit_p (&check_only_regs, reg->regno))
5879 bitmap_set_bit (&live_regs, reg->regno);
5880 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
5881 if (reg->type != OP_OUT
5882 && bitmap_bit_p (&check_only_regs, reg->regno))
5883 bitmap_set_bit (&live_regs, reg->regno);
5884 if (curr_id->arg_hard_regs != NULL)
5885 /* Make used argument hard registers live. */
5886 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
5887 if (regno < FIRST_PSEUDO_REGISTER
5888 && bitmap_bit_p (&check_only_regs, regno))
5889 bitmap_set_bit (&live_regs, regno);
5890 /* It is quite important to remove dead move insns because it
5891 means removing dead store. We don't need to process them for
5892 constraints. */
5893 if (remove_p)
5895 if (lra_dump_file != NULL)
5897 fprintf (lra_dump_file, " Removing dead insn:\n ");
5898 dump_insn_slim (lra_dump_file, curr_insn);
5900 lra_set_insn_deleted (curr_insn);
5905 /* The structure describes info to do an inheritance for the current
5906 insn. We need to collect such info first before doing the
5907 transformations because the transformations change the insn
5908 internal representation. */
5909 struct to_inherit
5911 /* Original regno. */
5912 int regno;
5913 /* Subsequent insns which can inherit original reg value. */
5914 rtx insns;
5917 /* Array containing all info for doing inheritance from the current
5918 insn. */
5919 static struct to_inherit to_inherit[LRA_MAX_INSN_RELOADS];
5921 /* Number elements in the previous array. */
5922 static int to_inherit_num;
5924 /* Add inheritance info REGNO and INSNS. Their meaning is described in
5925 structure to_inherit. */
5926 static void
5927 add_to_inherit (int regno, rtx insns)
5929 int i;
5931 for (i = 0; i < to_inherit_num; i++)
5932 if (to_inherit[i].regno == regno)
5933 return;
5934 lra_assert (to_inherit_num < LRA_MAX_INSN_RELOADS);
5935 to_inherit[to_inherit_num].regno = regno;
5936 to_inherit[to_inherit_num++].insns = insns;
5939 /* Return the last non-debug insn in basic block BB, or the block begin
5940 note if none. */
5941 static rtx_insn *
5942 get_last_insertion_point (basic_block bb)
5944 rtx_insn *insn;
5946 FOR_BB_INSNS_REVERSE (bb, insn)
5947 if (NONDEBUG_INSN_P (insn) || NOTE_INSN_BASIC_BLOCK_P (insn))
5948 return insn;
5949 gcc_unreachable ();
5952 /* Set up RES by registers living on edges FROM except the edge (FROM,
5953 TO) or by registers set up in a jump insn in BB FROM. */
5954 static void
5955 get_live_on_other_edges (basic_block from, basic_block to, bitmap res)
5957 rtx_insn *last;
5958 struct lra_insn_reg *reg;
5959 edge e;
5960 edge_iterator ei;
5962 lra_assert (to != NULL);
5963 bitmap_clear (res);
5964 FOR_EACH_EDGE (e, ei, from->succs)
5965 if (e->dest != to)
5966 bitmap_ior_into (res, df_get_live_in (e->dest));
5967 last = get_last_insertion_point (from);
5968 if (! JUMP_P (last))
5969 return;
5970 curr_id = lra_get_insn_recog_data (last);
5971 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5972 if (reg->type != OP_IN)
5973 bitmap_set_bit (res, reg->regno);
5976 /* Used as a temporary results of some bitmap calculations. */
5977 static bitmap_head temp_bitmap;
5979 /* We split for reloads of small class of hard regs. The following
5980 defines how many hard regs the class should have to be qualified as
5981 small. The code is mostly oriented to x86/x86-64 architecture
5982 where some insns need to use only specific register or pair of
5983 registers and these register can live in RTL explicitly, e.g. for
5984 parameter passing. */
5985 static const int max_small_class_regs_num = 2;
5987 /* Do inheritance/split transformations in EBB starting with HEAD and
5988 finishing on TAIL. We process EBB insns in the reverse order.
5989 Return true if we did any inheritance/split transformation in the
5990 EBB.
5992 We should avoid excessive splitting which results in worse code
5993 because of inaccurate cost calculations for spilling new split
5994 pseudos in such case. To achieve this we do splitting only if
5995 register pressure is high in given basic block and there are reload
5996 pseudos requiring hard registers. We could do more register
5997 pressure calculations at any given program point to avoid necessary
5998 splitting even more but it is to expensive and the current approach
5999 works well enough. */
6000 static bool
6001 inherit_in_ebb (rtx_insn *head, rtx_insn *tail)
6003 int i, src_regno, dst_regno, nregs;
6004 bool change_p, succ_p, update_reloads_num_p;
6005 rtx_insn *prev_insn, *last_insn;
6006 rtx next_usage_insns, curr_set;
6007 enum reg_class cl;
6008 struct lra_insn_reg *reg;
6009 basic_block last_processed_bb, curr_bb = NULL;
6010 HARD_REG_SET potential_reload_hard_regs, live_hard_regs;
6011 bitmap to_process;
6012 unsigned int j;
6013 bitmap_iterator bi;
6014 bool head_p, after_p;
6016 change_p = false;
6017 curr_usage_insns_check++;
6018 clear_invariants ();
6019 reloads_num = calls_num = 0;
6020 bitmap_clear (&check_only_regs);
6021 bitmap_clear (&invalid_invariant_regs);
6022 last_processed_bb = NULL;
6023 CLEAR_HARD_REG_SET (potential_reload_hard_regs);
6024 COPY_HARD_REG_SET (live_hard_regs, eliminable_regset);
6025 IOR_HARD_REG_SET (live_hard_regs, lra_no_alloc_regs);
6026 /* We don't process new insns generated in the loop. */
6027 for (curr_insn = tail; curr_insn != PREV_INSN (head); curr_insn = prev_insn)
6029 prev_insn = PREV_INSN (curr_insn);
6030 if (BLOCK_FOR_INSN (curr_insn) != NULL)
6031 curr_bb = BLOCK_FOR_INSN (curr_insn);
6032 if (last_processed_bb != curr_bb)
6034 /* We are at the end of BB. Add qualified living
6035 pseudos for potential splitting. */
6036 to_process = df_get_live_out (curr_bb);
6037 if (last_processed_bb != NULL)
6039 /* We are somewhere in the middle of EBB. */
6040 get_live_on_other_edges (curr_bb, last_processed_bb,
6041 &temp_bitmap);
6042 to_process = &temp_bitmap;
6044 last_processed_bb = curr_bb;
6045 last_insn = get_last_insertion_point (curr_bb);
6046 after_p = (! JUMP_P (last_insn)
6047 && (! CALL_P (last_insn)
6048 || (find_reg_note (last_insn,
6049 REG_NORETURN, NULL_RTX) == NULL_RTX
6050 && ! SIBLING_CALL_P (last_insn))));
6051 CLEAR_HARD_REG_SET (potential_reload_hard_regs);
6052 EXECUTE_IF_SET_IN_BITMAP (to_process, 0, j, bi)
6054 if ((int) j >= lra_constraint_new_regno_start)
6055 break;
6056 if (j < FIRST_PSEUDO_REGISTER || reg_renumber[j] >= 0)
6058 if (j < FIRST_PSEUDO_REGISTER)
6059 SET_HARD_REG_BIT (live_hard_regs, j);
6060 else
6061 add_to_hard_reg_set (&live_hard_regs,
6062 PSEUDO_REGNO_MODE (j),
6063 reg_renumber[j]);
6064 setup_next_usage_insn (j, last_insn, reloads_num, after_p);
6068 src_regno = dst_regno = -1;
6069 curr_set = single_set (curr_insn);
6070 if (curr_set != NULL_RTX && REG_P (SET_DEST (curr_set)))
6071 dst_regno = REGNO (SET_DEST (curr_set));
6072 if (curr_set != NULL_RTX && REG_P (SET_SRC (curr_set)))
6073 src_regno = REGNO (SET_SRC (curr_set));
6074 update_reloads_num_p = true;
6075 if (src_regno < lra_constraint_new_regno_start
6076 && src_regno >= FIRST_PSEUDO_REGISTER
6077 && reg_renumber[src_regno] < 0
6078 && dst_regno >= lra_constraint_new_regno_start
6079 && (cl = lra_get_allocno_class (dst_regno)) != NO_REGS)
6081 /* 'reload_pseudo <- original_pseudo'. */
6082 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6083 reloads_num++;
6084 update_reloads_num_p = false;
6085 succ_p = false;
6086 if (usage_insns[src_regno].check == curr_usage_insns_check
6087 && (next_usage_insns = usage_insns[src_regno].insns) != NULL_RTX)
6088 succ_p = inherit_reload_reg (false, src_regno, cl,
6089 curr_insn, next_usage_insns);
6090 if (succ_p)
6091 change_p = true;
6092 else
6093 setup_next_usage_insn (src_regno, curr_insn, reloads_num, false);
6094 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6095 IOR_HARD_REG_SET (potential_reload_hard_regs,
6096 reg_class_contents[cl]);
6098 else if (src_regno < 0
6099 && dst_regno >= lra_constraint_new_regno_start
6100 && invariant_p (SET_SRC (curr_set))
6101 && (cl = lra_get_allocno_class (dst_regno)) != NO_REGS
6102 && ! bitmap_bit_p (&invalid_invariant_regs, dst_regno)
6103 && ! bitmap_bit_p (&invalid_invariant_regs,
6104 ORIGINAL_REGNO(regno_reg_rtx[dst_regno])))
6106 /* 'reload_pseudo <- invariant'. */
6107 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6108 reloads_num++;
6109 update_reloads_num_p = false;
6110 if (process_invariant_for_inheritance (SET_DEST (curr_set), SET_SRC (curr_set)))
6111 change_p = true;
6112 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6113 IOR_HARD_REG_SET (potential_reload_hard_regs,
6114 reg_class_contents[cl]);
6116 else if (src_regno >= lra_constraint_new_regno_start
6117 && dst_regno < lra_constraint_new_regno_start
6118 && dst_regno >= FIRST_PSEUDO_REGISTER
6119 && reg_renumber[dst_regno] < 0
6120 && (cl = lra_get_allocno_class (src_regno)) != NO_REGS
6121 && usage_insns[dst_regno].check == curr_usage_insns_check
6122 && (next_usage_insns
6123 = usage_insns[dst_regno].insns) != NULL_RTX)
6125 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6126 reloads_num++;
6127 update_reloads_num_p = false;
6128 /* 'original_pseudo <- reload_pseudo'. */
6129 if (! JUMP_P (curr_insn)
6130 && inherit_reload_reg (true, dst_regno, cl,
6131 curr_insn, next_usage_insns))
6132 change_p = true;
6133 /* Invalidate. */
6134 usage_insns[dst_regno].check = 0;
6135 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6136 IOR_HARD_REG_SET (potential_reload_hard_regs,
6137 reg_class_contents[cl]);
6139 else if (INSN_P (curr_insn))
6141 int iter;
6142 int max_uid = get_max_uid ();
6144 curr_id = lra_get_insn_recog_data (curr_insn);
6145 curr_static_id = curr_id->insn_static_data;
6146 to_inherit_num = 0;
6147 /* Process insn definitions. */
6148 for (iter = 0; iter < 2; iter++)
6149 for (reg = iter == 0 ? curr_id->regs : curr_static_id->hard_regs;
6150 reg != NULL;
6151 reg = reg->next)
6152 if (reg->type != OP_IN
6153 && (dst_regno = reg->regno) < lra_constraint_new_regno_start)
6155 if (dst_regno >= FIRST_PSEUDO_REGISTER && reg->type == OP_OUT
6156 && reg_renumber[dst_regno] < 0 && ! reg->subreg_p
6157 && usage_insns[dst_regno].check == curr_usage_insns_check
6158 && (next_usage_insns
6159 = usage_insns[dst_regno].insns) != NULL_RTX)
6161 struct lra_insn_reg *r;
6163 for (r = curr_id->regs; r != NULL; r = r->next)
6164 if (r->type != OP_OUT && r->regno == dst_regno)
6165 break;
6166 /* Don't do inheritance if the pseudo is also
6167 used in the insn. */
6168 if (r == NULL)
6169 /* We can not do inheritance right now
6170 because the current insn reg info (chain
6171 regs) can change after that. */
6172 add_to_inherit (dst_regno, next_usage_insns);
6174 /* We can not process one reg twice here because of
6175 usage_insns invalidation. */
6176 if ((dst_regno < FIRST_PSEUDO_REGISTER
6177 || reg_renumber[dst_regno] >= 0)
6178 && ! reg->subreg_p && reg->type != OP_IN)
6180 HARD_REG_SET s;
6182 if (split_if_necessary (dst_regno, reg->biggest_mode,
6183 potential_reload_hard_regs,
6184 false, curr_insn, max_uid))
6185 change_p = true;
6186 CLEAR_HARD_REG_SET (s);
6187 if (dst_regno < FIRST_PSEUDO_REGISTER)
6188 add_to_hard_reg_set (&s, reg->biggest_mode, dst_regno);
6189 else
6190 add_to_hard_reg_set (&s, PSEUDO_REGNO_MODE (dst_regno),
6191 reg_renumber[dst_regno]);
6192 AND_COMPL_HARD_REG_SET (live_hard_regs, s);
6194 /* We should invalidate potential inheritance or
6195 splitting for the current insn usages to the next
6196 usage insns (see code below) as the output pseudo
6197 prevents this. */
6198 if ((dst_regno >= FIRST_PSEUDO_REGISTER
6199 && reg_renumber[dst_regno] < 0)
6200 || (reg->type == OP_OUT && ! reg->subreg_p
6201 && (dst_regno < FIRST_PSEUDO_REGISTER
6202 || reg_renumber[dst_regno] >= 0)))
6204 /* Invalidate and mark definitions. */
6205 if (dst_regno >= FIRST_PSEUDO_REGISTER)
6206 usage_insns[dst_regno].check = -(int) INSN_UID (curr_insn);
6207 else
6209 nregs = hard_regno_nregs (dst_regno,
6210 reg->biggest_mode);
6211 for (i = 0; i < nregs; i++)
6212 usage_insns[dst_regno + i].check
6213 = -(int) INSN_UID (curr_insn);
6217 /* Process clobbered call regs. */
6218 if (curr_id->arg_hard_regs != NULL)
6219 for (i = 0; (dst_regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6220 if (dst_regno >= FIRST_PSEUDO_REGISTER)
6221 usage_insns[dst_regno - FIRST_PSEUDO_REGISTER].check
6222 = -(int) INSN_UID (curr_insn);
6223 if (! JUMP_P (curr_insn))
6224 for (i = 0; i < to_inherit_num; i++)
6225 if (inherit_reload_reg (true, to_inherit[i].regno,
6226 ALL_REGS, curr_insn,
6227 to_inherit[i].insns))
6228 change_p = true;
6229 if (CALL_P (curr_insn))
6231 rtx cheap, pat, dest;
6232 rtx_insn *restore;
6233 int regno, hard_regno;
6235 calls_num++;
6236 if ((cheap = find_reg_note (curr_insn,
6237 REG_RETURNED, NULL_RTX)) != NULL_RTX
6238 && ((cheap = XEXP (cheap, 0)), true)
6239 && (regno = REGNO (cheap)) >= FIRST_PSEUDO_REGISTER
6240 && (hard_regno = reg_renumber[regno]) >= 0
6241 && usage_insns[regno].check == curr_usage_insns_check
6242 /* If there are pending saves/restores, the
6243 optimization is not worth. */
6244 && usage_insns[regno].calls_num == calls_num - 1
6245 && TEST_HARD_REG_BIT (call_used_reg_set, hard_regno))
6247 /* Restore the pseudo from the call result as
6248 REG_RETURNED note says that the pseudo value is
6249 in the call result and the pseudo is an argument
6250 of the call. */
6251 pat = PATTERN (curr_insn);
6252 if (GET_CODE (pat) == PARALLEL)
6253 pat = XVECEXP (pat, 0, 0);
6254 dest = SET_DEST (pat);
6255 /* For multiple return values dest is PARALLEL.
6256 Currently we handle only single return value case. */
6257 if (REG_P (dest))
6259 start_sequence ();
6260 emit_move_insn (cheap, copy_rtx (dest));
6261 restore = get_insns ();
6262 end_sequence ();
6263 lra_process_new_insns (curr_insn, NULL, restore,
6264 "Inserting call parameter restore");
6265 /* We don't need to save/restore of the pseudo from
6266 this call. */
6267 usage_insns[regno].calls_num = calls_num;
6268 bitmap_set_bit (&check_only_regs, regno);
6272 to_inherit_num = 0;
6273 /* Process insn usages. */
6274 for (iter = 0; iter < 2; iter++)
6275 for (reg = iter == 0 ? curr_id->regs : curr_static_id->hard_regs;
6276 reg != NULL;
6277 reg = reg->next)
6278 if ((reg->type != OP_OUT
6279 || (reg->type == OP_OUT && reg->subreg_p))
6280 && (src_regno = reg->regno) < lra_constraint_new_regno_start)
6282 if (src_regno >= FIRST_PSEUDO_REGISTER
6283 && reg_renumber[src_regno] < 0 && reg->type == OP_IN)
6285 if (usage_insns[src_regno].check == curr_usage_insns_check
6286 && (next_usage_insns
6287 = usage_insns[src_regno].insns) != NULL_RTX
6288 && NONDEBUG_INSN_P (curr_insn))
6289 add_to_inherit (src_regno, next_usage_insns);
6290 else if (usage_insns[src_regno].check
6291 != -(int) INSN_UID (curr_insn))
6292 /* Add usages but only if the reg is not set up
6293 in the same insn. */
6294 add_next_usage_insn (src_regno, curr_insn, reloads_num);
6296 else if (src_regno < FIRST_PSEUDO_REGISTER
6297 || reg_renumber[src_regno] >= 0)
6299 bool before_p;
6300 rtx_insn *use_insn = curr_insn;
6302 before_p = (JUMP_P (curr_insn)
6303 || (CALL_P (curr_insn) && reg->type == OP_IN));
6304 if (NONDEBUG_INSN_P (curr_insn)
6305 && (! JUMP_P (curr_insn) || reg->type == OP_IN)
6306 && split_if_necessary (src_regno, reg->biggest_mode,
6307 potential_reload_hard_regs,
6308 before_p, curr_insn, max_uid))
6310 if (reg->subreg_p)
6311 lra_risky_transformations_p = true;
6312 change_p = true;
6313 /* Invalidate. */
6314 usage_insns[src_regno].check = 0;
6315 if (before_p)
6316 use_insn = PREV_INSN (curr_insn);
6318 if (NONDEBUG_INSN_P (curr_insn))
6320 if (src_regno < FIRST_PSEUDO_REGISTER)
6321 add_to_hard_reg_set (&live_hard_regs,
6322 reg->biggest_mode, src_regno);
6323 else
6324 add_to_hard_reg_set (&live_hard_regs,
6325 PSEUDO_REGNO_MODE (src_regno),
6326 reg_renumber[src_regno]);
6328 add_next_usage_insn (src_regno, use_insn, reloads_num);
6331 /* Process used call regs. */
6332 if (curr_id->arg_hard_regs != NULL)
6333 for (i = 0; (src_regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6334 if (src_regno < FIRST_PSEUDO_REGISTER)
6336 SET_HARD_REG_BIT (live_hard_regs, src_regno);
6337 add_next_usage_insn (src_regno, curr_insn, reloads_num);
6339 for (i = 0; i < to_inherit_num; i++)
6341 src_regno = to_inherit[i].regno;
6342 if (inherit_reload_reg (false, src_regno, ALL_REGS,
6343 curr_insn, to_inherit[i].insns))
6344 change_p = true;
6345 else
6346 setup_next_usage_insn (src_regno, curr_insn, reloads_num, false);
6349 if (update_reloads_num_p
6350 && NONDEBUG_INSN_P (curr_insn) && curr_set != NULL_RTX)
6352 int regno = -1;
6353 if ((REG_P (SET_DEST (curr_set))
6354 && (regno = REGNO (SET_DEST (curr_set))) >= lra_constraint_new_regno_start
6355 && reg_renumber[regno] < 0
6356 && (cl = lra_get_allocno_class (regno)) != NO_REGS)
6357 || (REG_P (SET_SRC (curr_set))
6358 && (regno = REGNO (SET_SRC (curr_set))) >= lra_constraint_new_regno_start
6359 && reg_renumber[regno] < 0
6360 && (cl = lra_get_allocno_class (regno)) != NO_REGS))
6362 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6363 reloads_num++;
6364 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6365 IOR_HARD_REG_SET (potential_reload_hard_regs,
6366 reg_class_contents[cl]);
6369 if (NONDEBUG_INSN_P (curr_insn))
6371 int regno;
6373 /* Invalidate invariants with changed regs. */
6374 curr_id = lra_get_insn_recog_data (curr_insn);
6375 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
6376 if (reg->type != OP_IN)
6378 bitmap_set_bit (&invalid_invariant_regs, reg->regno);
6379 bitmap_set_bit (&invalid_invariant_regs,
6380 ORIGINAL_REGNO (regno_reg_rtx[reg->regno]));
6382 curr_static_id = curr_id->insn_static_data;
6383 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
6384 if (reg->type != OP_IN)
6385 bitmap_set_bit (&invalid_invariant_regs, reg->regno);
6386 if (curr_id->arg_hard_regs != NULL)
6387 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6388 if (regno >= FIRST_PSEUDO_REGISTER)
6389 bitmap_set_bit (&invalid_invariant_regs,
6390 regno - FIRST_PSEUDO_REGISTER);
6392 /* We reached the start of the current basic block. */
6393 if (prev_insn == NULL_RTX || prev_insn == PREV_INSN (head)
6394 || BLOCK_FOR_INSN (prev_insn) != curr_bb)
6396 /* We reached the beginning of the current block -- do
6397 rest of spliting in the current BB. */
6398 to_process = df_get_live_in (curr_bb);
6399 if (BLOCK_FOR_INSN (head) != curr_bb)
6401 /* We are somewhere in the middle of EBB. */
6402 get_live_on_other_edges (EDGE_PRED (curr_bb, 0)->src,
6403 curr_bb, &temp_bitmap);
6404 to_process = &temp_bitmap;
6406 head_p = true;
6407 EXECUTE_IF_SET_IN_BITMAP (to_process, 0, j, bi)
6409 if ((int) j >= lra_constraint_new_regno_start)
6410 break;
6411 if (((int) j < FIRST_PSEUDO_REGISTER || reg_renumber[j] >= 0)
6412 && usage_insns[j].check == curr_usage_insns_check
6413 && (next_usage_insns = usage_insns[j].insns) != NULL_RTX)
6415 if (need_for_split_p (potential_reload_hard_regs, j))
6417 if (lra_dump_file != NULL && head_p)
6419 fprintf (lra_dump_file,
6420 " ----------------------------------\n");
6421 head_p = false;
6423 if (split_reg (false, j, bb_note (curr_bb),
6424 next_usage_insns))
6425 change_p = true;
6427 usage_insns[j].check = 0;
6432 return change_p;
6435 /* This value affects EBB forming. If probability of edge from EBB to
6436 a BB is not greater than the following value, we don't add the BB
6437 to EBB. */
6438 #define EBB_PROBABILITY_CUTOFF \
6439 ((REG_BR_PROB_BASE * LRA_INHERITANCE_EBB_PROBABILITY_CUTOFF) / 100)
6441 /* Current number of inheritance/split iteration. */
6442 int lra_inheritance_iter;
6444 /* Entry function for inheritance/split pass. */
6445 void
6446 lra_inheritance (void)
6448 int i;
6449 basic_block bb, start_bb;
6450 edge e;
6452 lra_inheritance_iter++;
6453 if (lra_inheritance_iter > LRA_MAX_INHERITANCE_PASSES)
6454 return;
6455 timevar_push (TV_LRA_INHERITANCE);
6456 if (lra_dump_file != NULL)
6457 fprintf (lra_dump_file, "\n********** Inheritance #%d: **********\n\n",
6458 lra_inheritance_iter);
6459 curr_usage_insns_check = 0;
6460 usage_insns = XNEWVEC (struct usage_insns, lra_constraint_new_regno_start);
6461 for (i = 0; i < lra_constraint_new_regno_start; i++)
6462 usage_insns[i].check = 0;
6463 bitmap_initialize (&check_only_regs, &reg_obstack);
6464 bitmap_initialize (&invalid_invariant_regs, &reg_obstack);
6465 bitmap_initialize (&live_regs, &reg_obstack);
6466 bitmap_initialize (&temp_bitmap, &reg_obstack);
6467 bitmap_initialize (&ebb_global_regs, &reg_obstack);
6468 FOR_EACH_BB_FN (bb, cfun)
6470 start_bb = bb;
6471 if (lra_dump_file != NULL)
6472 fprintf (lra_dump_file, "EBB");
6473 /* Form a EBB starting with BB. */
6474 bitmap_clear (&ebb_global_regs);
6475 bitmap_ior_into (&ebb_global_regs, df_get_live_in (bb));
6476 for (;;)
6478 if (lra_dump_file != NULL)
6479 fprintf (lra_dump_file, " %d", bb->index);
6480 if (bb->next_bb == EXIT_BLOCK_PTR_FOR_FN (cfun)
6481 || LABEL_P (BB_HEAD (bb->next_bb)))
6482 break;
6483 e = find_fallthru_edge (bb->succs);
6484 if (! e)
6485 break;
6486 if (e->probability.initialized_p ()
6487 && e->probability.to_reg_br_prob_base () < EBB_PROBABILITY_CUTOFF)
6488 break;
6489 bb = bb->next_bb;
6491 bitmap_ior_into (&ebb_global_regs, df_get_live_out (bb));
6492 if (lra_dump_file != NULL)
6493 fprintf (lra_dump_file, "\n");
6494 if (inherit_in_ebb (BB_HEAD (start_bb), BB_END (bb)))
6495 /* Remember that the EBB head and tail can change in
6496 inherit_in_ebb. */
6497 update_ebb_live_info (BB_HEAD (start_bb), BB_END (bb));
6499 bitmap_clear (&ebb_global_regs);
6500 bitmap_clear (&temp_bitmap);
6501 bitmap_clear (&live_regs);
6502 bitmap_clear (&invalid_invariant_regs);
6503 bitmap_clear (&check_only_regs);
6504 free (usage_insns);
6506 timevar_pop (TV_LRA_INHERITANCE);
6511 /* This page contains code to undo failed inheritance/split
6512 transformations. */
6514 /* Current number of iteration undoing inheritance/split. */
6515 int lra_undo_inheritance_iter;
6517 /* Fix BB live info LIVE after removing pseudos created on pass doing
6518 inheritance/split which are REMOVED_PSEUDOS. */
6519 static void
6520 fix_bb_live_info (bitmap live, bitmap removed_pseudos)
6522 unsigned int regno;
6523 bitmap_iterator bi;
6525 EXECUTE_IF_SET_IN_BITMAP (removed_pseudos, 0, regno, bi)
6526 if (bitmap_clear_bit (live, regno)
6527 && REG_P (lra_reg_info[regno].restore_rtx))
6528 bitmap_set_bit (live, REGNO (lra_reg_info[regno].restore_rtx));
6531 /* Return regno of the (subreg of) REG. Otherwise, return a negative
6532 number. */
6533 static int
6534 get_regno (rtx reg)
6536 if (GET_CODE (reg) == SUBREG)
6537 reg = SUBREG_REG (reg);
6538 if (REG_P (reg))
6539 return REGNO (reg);
6540 return -1;
6543 /* Delete a move INSN with destination reg DREGNO and a previous
6544 clobber insn with the same regno. The inheritance/split code can
6545 generate moves with preceding clobber and when we delete such moves
6546 we should delete the clobber insn too to keep the correct life
6547 info. */
6548 static void
6549 delete_move_and_clobber (rtx_insn *insn, int dregno)
6551 rtx_insn *prev_insn = PREV_INSN (insn);
6553 lra_set_insn_deleted (insn);
6554 lra_assert (dregno >= 0);
6555 if (prev_insn != NULL && NONDEBUG_INSN_P (prev_insn)
6556 && GET_CODE (PATTERN (prev_insn)) == CLOBBER
6557 && dregno == get_regno (XEXP (PATTERN (prev_insn), 0)))
6558 lra_set_insn_deleted (prev_insn);
6561 /* Remove inheritance/split pseudos which are in REMOVE_PSEUDOS and
6562 return true if we did any change. The undo transformations for
6563 inheritance looks like
6564 i <- i2
6565 p <- i => p <- i2
6566 or removing
6567 p <- i, i <- p, and i <- i3
6568 where p is original pseudo from which inheritance pseudo i was
6569 created, i and i3 are removed inheritance pseudos, i2 is another
6570 not removed inheritance pseudo. All split pseudos or other
6571 occurrences of removed inheritance pseudos are changed on the
6572 corresponding original pseudos.
6574 The function also schedules insns changed and created during
6575 inheritance/split pass for processing by the subsequent constraint
6576 pass. */
6577 static bool
6578 remove_inheritance_pseudos (bitmap remove_pseudos)
6580 basic_block bb;
6581 int regno, sregno, prev_sregno, dregno;
6582 rtx restore_rtx;
6583 rtx set, prev_set;
6584 rtx_insn *prev_insn;
6585 bool change_p, done_p;
6587 change_p = ! bitmap_empty_p (remove_pseudos);
6588 /* We can not finish the function right away if CHANGE_P is true
6589 because we need to marks insns affected by previous
6590 inheritance/split pass for processing by the subsequent
6591 constraint pass. */
6592 FOR_EACH_BB_FN (bb, cfun)
6594 fix_bb_live_info (df_get_live_in (bb), remove_pseudos);
6595 fix_bb_live_info (df_get_live_out (bb), remove_pseudos);
6596 FOR_BB_INSNS_REVERSE (bb, curr_insn)
6598 if (! INSN_P (curr_insn))
6599 continue;
6600 done_p = false;
6601 sregno = dregno = -1;
6602 if (change_p && NONDEBUG_INSN_P (curr_insn)
6603 && (set = single_set (curr_insn)) != NULL_RTX)
6605 dregno = get_regno (SET_DEST (set));
6606 sregno = get_regno (SET_SRC (set));
6609 if (sregno >= 0 && dregno >= 0)
6611 if (bitmap_bit_p (remove_pseudos, dregno)
6612 && ! REG_P (lra_reg_info[dregno].restore_rtx))
6614 /* invariant inheritance pseudo <- original pseudo */
6615 if (lra_dump_file != NULL)
6617 fprintf (lra_dump_file, " Removing invariant inheritance:\n");
6618 dump_insn_slim (lra_dump_file, curr_insn);
6619 fprintf (lra_dump_file, "\n");
6621 delete_move_and_clobber (curr_insn, dregno);
6622 done_p = true;
6624 else if (bitmap_bit_p (remove_pseudos, sregno)
6625 && ! REG_P (lra_reg_info[sregno].restore_rtx))
6627 /* reload pseudo <- invariant inheritance pseudo */
6628 start_sequence ();
6629 /* We can not just change the source. It might be
6630 an insn different from the move. */
6631 emit_insn (lra_reg_info[sregno].restore_rtx);
6632 rtx_insn *new_insns = get_insns ();
6633 end_sequence ();
6634 lra_assert (single_set (new_insns) != NULL
6635 && SET_DEST (set) == SET_DEST (single_set (new_insns)));
6636 lra_process_new_insns (curr_insn, NULL, new_insns,
6637 "Changing reload<-invariant inheritance");
6638 delete_move_and_clobber (curr_insn, dregno);
6639 done_p = true;
6641 else if ((bitmap_bit_p (remove_pseudos, sregno)
6642 && (get_regno (lra_reg_info[sregno].restore_rtx) == dregno
6643 || (bitmap_bit_p (remove_pseudos, dregno)
6644 && get_regno (lra_reg_info[sregno].restore_rtx) >= 0
6645 && (get_regno (lra_reg_info[sregno].restore_rtx)
6646 == get_regno (lra_reg_info[dregno].restore_rtx)))))
6647 || (bitmap_bit_p (remove_pseudos, dregno)
6648 && get_regno (lra_reg_info[dregno].restore_rtx) == sregno))
6649 /* One of the following cases:
6650 original <- removed inheritance pseudo
6651 removed inherit pseudo <- another removed inherit pseudo
6652 removed inherit pseudo <- original pseudo
6654 removed_split_pseudo <- original_reg
6655 original_reg <- removed_split_pseudo */
6657 if (lra_dump_file != NULL)
6659 fprintf (lra_dump_file, " Removing %s:\n",
6660 bitmap_bit_p (&lra_split_regs, sregno)
6661 || bitmap_bit_p (&lra_split_regs, dregno)
6662 ? "split" : "inheritance");
6663 dump_insn_slim (lra_dump_file, curr_insn);
6665 delete_move_and_clobber (curr_insn, dregno);
6666 done_p = true;
6668 else if (bitmap_bit_p (remove_pseudos, sregno)
6669 && bitmap_bit_p (&lra_inheritance_pseudos, sregno))
6671 /* Search the following pattern:
6672 inherit_or_split_pseudo1 <- inherit_or_split_pseudo2
6673 original_pseudo <- inherit_or_split_pseudo1
6674 where the 2nd insn is the current insn and
6675 inherit_or_split_pseudo2 is not removed. If it is found,
6676 change the current insn onto:
6677 original_pseudo <- inherit_or_split_pseudo2. */
6678 for (prev_insn = PREV_INSN (curr_insn);
6679 prev_insn != NULL_RTX && ! NONDEBUG_INSN_P (prev_insn);
6680 prev_insn = PREV_INSN (prev_insn))
6682 if (prev_insn != NULL_RTX && BLOCK_FOR_INSN (prev_insn) == bb
6683 && (prev_set = single_set (prev_insn)) != NULL_RTX
6684 /* There should be no subregs in insn we are
6685 searching because only the original reg might
6686 be in subreg when we changed the mode of
6687 load/store for splitting. */
6688 && REG_P (SET_DEST (prev_set))
6689 && REG_P (SET_SRC (prev_set))
6690 && (int) REGNO (SET_DEST (prev_set)) == sregno
6691 && ((prev_sregno = REGNO (SET_SRC (prev_set)))
6692 >= FIRST_PSEUDO_REGISTER)
6693 && (lra_reg_info[prev_sregno].restore_rtx == NULL_RTX
6695 /* As we consider chain of inheritance or
6696 splitting described in above comment we should
6697 check that sregno and prev_sregno were
6698 inheritance/split pseudos created from the
6699 same original regno. */
6700 (get_regno (lra_reg_info[sregno].restore_rtx) >= 0
6701 && (get_regno (lra_reg_info[sregno].restore_rtx)
6702 == get_regno (lra_reg_info[prev_sregno].restore_rtx))))
6703 && ! bitmap_bit_p (remove_pseudos, prev_sregno))
6705 lra_assert (GET_MODE (SET_SRC (prev_set))
6706 == GET_MODE (regno_reg_rtx[sregno]));
6707 if (GET_CODE (SET_SRC (set)) == SUBREG)
6708 SUBREG_REG (SET_SRC (set)) = SET_SRC (prev_set);
6709 else
6710 SET_SRC (set) = SET_SRC (prev_set);
6711 /* As we are finishing with processing the insn
6712 here, check the destination too as it might
6713 inheritance pseudo for another pseudo. */
6714 if (bitmap_bit_p (remove_pseudos, dregno)
6715 && bitmap_bit_p (&lra_inheritance_pseudos, dregno)
6716 && (restore_rtx
6717 = lra_reg_info[dregno].restore_rtx) != NULL_RTX)
6719 if (GET_CODE (SET_DEST (set)) == SUBREG)
6720 SUBREG_REG (SET_DEST (set)) = restore_rtx;
6721 else
6722 SET_DEST (set) = restore_rtx;
6724 lra_push_insn_and_update_insn_regno_info (curr_insn);
6725 lra_set_used_insn_alternative_by_uid
6726 (INSN_UID (curr_insn), -1);
6727 done_p = true;
6728 if (lra_dump_file != NULL)
6730 fprintf (lra_dump_file, " Change reload insn:\n");
6731 dump_insn_slim (lra_dump_file, curr_insn);
6736 if (! done_p)
6738 struct lra_insn_reg *reg;
6739 bool restored_regs_p = false;
6740 bool kept_regs_p = false;
6742 curr_id = lra_get_insn_recog_data (curr_insn);
6743 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
6745 regno = reg->regno;
6746 restore_rtx = lra_reg_info[regno].restore_rtx;
6747 if (restore_rtx != NULL_RTX)
6749 if (change_p && bitmap_bit_p (remove_pseudos, regno))
6751 lra_substitute_pseudo_within_insn
6752 (curr_insn, regno, restore_rtx, false);
6753 restored_regs_p = true;
6755 else
6756 kept_regs_p = true;
6759 if (NONDEBUG_INSN_P (curr_insn) && kept_regs_p)
6761 /* The instruction has changed since the previous
6762 constraints pass. */
6763 lra_push_insn_and_update_insn_regno_info (curr_insn);
6764 lra_set_used_insn_alternative_by_uid
6765 (INSN_UID (curr_insn), -1);
6767 else if (restored_regs_p)
6768 /* The instruction has been restored to the form that
6769 it had during the previous constraints pass. */
6770 lra_update_insn_regno_info (curr_insn);
6771 if (restored_regs_p && lra_dump_file != NULL)
6773 fprintf (lra_dump_file, " Insn after restoring regs:\n");
6774 dump_insn_slim (lra_dump_file, curr_insn);
6779 return change_p;
6782 /* If optional reload pseudos failed to get a hard register or was not
6783 inherited, it is better to remove optional reloads. We do this
6784 transformation after undoing inheritance to figure out necessity to
6785 remove optional reloads easier. Return true if we do any
6786 change. */
6787 static bool
6788 undo_optional_reloads (void)
6790 bool change_p, keep_p;
6791 unsigned int regno, uid;
6792 bitmap_iterator bi, bi2;
6793 rtx_insn *insn;
6794 rtx set, src, dest;
6795 auto_bitmap removed_optional_reload_pseudos (&reg_obstack);
6797 bitmap_copy (removed_optional_reload_pseudos, &lra_optional_reload_pseudos);
6798 EXECUTE_IF_SET_IN_BITMAP (&lra_optional_reload_pseudos, 0, regno, bi)
6800 keep_p = false;
6801 /* Keep optional reloads from previous subpasses. */
6802 if (lra_reg_info[regno].restore_rtx == NULL_RTX
6803 /* If the original pseudo changed its allocation, just
6804 removing the optional pseudo is dangerous as the original
6805 pseudo will have longer live range. */
6806 || reg_renumber[REGNO (lra_reg_info[regno].restore_rtx)] >= 0)
6807 keep_p = true;
6808 else if (reg_renumber[regno] >= 0)
6809 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi2)
6811 insn = lra_insn_recog_data[uid]->insn;
6812 if ((set = single_set (insn)) == NULL_RTX)
6813 continue;
6814 src = SET_SRC (set);
6815 dest = SET_DEST (set);
6816 if (! REG_P (src) || ! REG_P (dest))
6817 continue;
6818 if (REGNO (dest) == regno
6819 /* Ignore insn for optional reloads itself. */
6820 && REGNO (lra_reg_info[regno].restore_rtx) != REGNO (src)
6821 /* Check only inheritance on last inheritance pass. */
6822 && (int) REGNO (src) >= new_regno_start
6823 /* Check that the optional reload was inherited. */
6824 && bitmap_bit_p (&lra_inheritance_pseudos, REGNO (src)))
6826 keep_p = true;
6827 break;
6830 if (keep_p)
6832 bitmap_clear_bit (removed_optional_reload_pseudos, regno);
6833 if (lra_dump_file != NULL)
6834 fprintf (lra_dump_file, "Keep optional reload reg %d\n", regno);
6837 change_p = ! bitmap_empty_p (removed_optional_reload_pseudos);
6838 auto_bitmap insn_bitmap (&reg_obstack);
6839 EXECUTE_IF_SET_IN_BITMAP (removed_optional_reload_pseudos, 0, regno, bi)
6841 if (lra_dump_file != NULL)
6842 fprintf (lra_dump_file, "Remove optional reload reg %d\n", regno);
6843 bitmap_copy (insn_bitmap, &lra_reg_info[regno].insn_bitmap);
6844 EXECUTE_IF_SET_IN_BITMAP (insn_bitmap, 0, uid, bi2)
6846 insn = lra_insn_recog_data[uid]->insn;
6847 if ((set = single_set (insn)) != NULL_RTX)
6849 src = SET_SRC (set);
6850 dest = SET_DEST (set);
6851 if (REG_P (src) && REG_P (dest)
6852 && ((REGNO (src) == regno
6853 && (REGNO (lra_reg_info[regno].restore_rtx)
6854 == REGNO (dest)))
6855 || (REGNO (dest) == regno
6856 && (REGNO (lra_reg_info[regno].restore_rtx)
6857 == REGNO (src)))))
6859 if (lra_dump_file != NULL)
6861 fprintf (lra_dump_file, " Deleting move %u\n",
6862 INSN_UID (insn));
6863 dump_insn_slim (lra_dump_file, insn);
6865 delete_move_and_clobber (insn, REGNO (dest));
6866 continue;
6868 /* We should not worry about generation memory-memory
6869 moves here as if the corresponding inheritance did
6870 not work (inheritance pseudo did not get a hard reg),
6871 we remove the inheritance pseudo and the optional
6872 reload. */
6874 lra_substitute_pseudo_within_insn
6875 (insn, regno, lra_reg_info[regno].restore_rtx, false);
6876 lra_update_insn_regno_info (insn);
6877 if (lra_dump_file != NULL)
6879 fprintf (lra_dump_file,
6880 " Restoring original insn:\n");
6881 dump_insn_slim (lra_dump_file, insn);
6885 /* Clear restore_regnos. */
6886 EXECUTE_IF_SET_IN_BITMAP (&lra_optional_reload_pseudos, 0, regno, bi)
6887 lra_reg_info[regno].restore_rtx = NULL_RTX;
6888 return change_p;
6891 /* Entry function for undoing inheritance/split transformation. Return true
6892 if we did any RTL change in this pass. */
6893 bool
6894 lra_undo_inheritance (void)
6896 unsigned int regno;
6897 int hard_regno;
6898 int n_all_inherit, n_inherit, n_all_split, n_split;
6899 rtx restore_rtx;
6900 bitmap_iterator bi;
6901 bool change_p;
6903 lra_undo_inheritance_iter++;
6904 if (lra_undo_inheritance_iter > LRA_MAX_INHERITANCE_PASSES)
6905 return false;
6906 if (lra_dump_file != NULL)
6907 fprintf (lra_dump_file,
6908 "\n********** Undoing inheritance #%d: **********\n\n",
6909 lra_undo_inheritance_iter);
6910 auto_bitmap remove_pseudos (&reg_obstack);
6911 n_inherit = n_all_inherit = 0;
6912 EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, regno, bi)
6913 if (lra_reg_info[regno].restore_rtx != NULL_RTX)
6915 n_all_inherit++;
6916 if (reg_renumber[regno] < 0
6917 /* If the original pseudo changed its allocation, just
6918 removing inheritance is dangerous as for changing
6919 allocation we used shorter live-ranges. */
6920 && (! REG_P (lra_reg_info[regno].restore_rtx)
6921 || reg_renumber[REGNO (lra_reg_info[regno].restore_rtx)] < 0))
6922 bitmap_set_bit (remove_pseudos, regno);
6923 else
6924 n_inherit++;
6926 if (lra_dump_file != NULL && n_all_inherit != 0)
6927 fprintf (lra_dump_file, "Inherit %d out of %d (%.2f%%)\n",
6928 n_inherit, n_all_inherit,
6929 (double) n_inherit / n_all_inherit * 100);
6930 n_split = n_all_split = 0;
6931 EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, regno, bi)
6932 if ((restore_rtx = lra_reg_info[regno].restore_rtx) != NULL_RTX)
6934 int restore_regno = REGNO (restore_rtx);
6936 n_all_split++;
6937 hard_regno = (restore_regno >= FIRST_PSEUDO_REGISTER
6938 ? reg_renumber[restore_regno] : restore_regno);
6939 if (hard_regno < 0 || reg_renumber[regno] == hard_regno)
6940 bitmap_set_bit (remove_pseudos, regno);
6941 else
6943 n_split++;
6944 if (lra_dump_file != NULL)
6945 fprintf (lra_dump_file, " Keep split r%d (orig=r%d)\n",
6946 regno, restore_regno);
6949 if (lra_dump_file != NULL && n_all_split != 0)
6950 fprintf (lra_dump_file, "Split %d out of %d (%.2f%%)\n",
6951 n_split, n_all_split,
6952 (double) n_split / n_all_split * 100);
6953 change_p = remove_inheritance_pseudos (remove_pseudos);
6954 /* Clear restore_regnos. */
6955 EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, regno, bi)
6956 lra_reg_info[regno].restore_rtx = NULL_RTX;
6957 EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, regno, bi)
6958 lra_reg_info[regno].restore_rtx = NULL_RTX;
6959 change_p = undo_optional_reloads () || change_p;
6960 return change_p;