2017-04-11 Vladimir Makarov <vmakarov@redhat.com>
[official-gcc.git] / gcc / lra-constraints.c
blobc8bc9b9a66f3c47aa6b7cf34e1c8b96092490cf2
1 /* Code for RTL transformations to satisfy insn constraints.
2 Copyright (C) 2010-2017 Free Software Foundation, Inc.
3 Contributed by Vladimir Makarov <vmakarov@redhat.com>.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
10 version.
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
22 /* This file contains code for 3 passes: constraint pass,
23 inheritance/split pass, and pass for undoing failed inheritance and
24 split.
26 The major goal of constraint pass is to transform RTL to satisfy
27 insn and address constraints by:
28 o choosing insn alternatives;
29 o generating *reload insns* (or reloads in brief) and *reload
30 pseudos* which will get necessary hard registers later;
31 o substituting pseudos with equivalent values and removing the
32 instructions that initialized those pseudos.
34 The constraint pass has biggest and most complicated code in LRA.
35 There are a lot of important details like:
36 o reuse of input reload pseudos to simplify reload pseudo
37 allocations;
38 o some heuristics to choose insn alternative to improve the
39 inheritance;
40 o early clobbers etc.
42 The pass is mimicking former reload pass in alternative choosing
43 because the reload pass is oriented to current machine description
44 model. It might be changed if the machine description model is
45 changed.
47 There is special code for preventing all LRA and this pass cycling
48 in case of bugs.
50 On the first iteration of the pass we process every instruction and
51 choose an alternative for each one. On subsequent iterations we try
52 to avoid reprocessing instructions if we can be sure that the old
53 choice is still valid.
55 The inheritance/spilt pass is to transform code to achieve
56 ineheritance and live range splitting. It is done on backward
57 traversal of EBBs.
59 The inheritance optimization goal is to reuse values in hard
60 registers. There is analogous optimization in old reload pass. The
61 inheritance is achieved by following transformation:
63 reload_p1 <- p reload_p1 <- p
64 ... new_p <- reload_p1
65 ... => ...
66 reload_p2 <- p reload_p2 <- new_p
68 where p is spilled and not changed between the insns. Reload_p1 is
69 also called *original pseudo* and new_p is called *inheritance
70 pseudo*.
72 The subsequent assignment pass will try to assign the same (or
73 another if it is not possible) hard register to new_p as to
74 reload_p1 or reload_p2.
76 If the assignment pass fails to assign a hard register to new_p,
77 this file will undo the inheritance and restore the original code.
78 This is because implementing the above sequence with a spilled
79 new_p would make the code much worse. The inheritance is done in
80 EBB scope. The above is just a simplified example to get an idea
81 of the inheritance as the inheritance is also done for non-reload
82 insns.
84 Splitting (transformation) is also done in EBB scope on the same
85 pass as the inheritance:
87 r <- ... or ... <- r r <- ... or ... <- r
88 ... s <- r (new insn -- save)
89 ... =>
90 ... r <- s (new insn -- restore)
91 ... <- r ... <- r
93 The *split pseudo* s is assigned to the hard register of the
94 original pseudo or hard register r.
96 Splitting is done:
97 o In EBBs with high register pressure for global pseudos (living
98 in at least 2 BBs) and assigned to hard registers when there
99 are more one reloads needing the hard registers;
100 o for pseudos needing save/restore code around calls.
102 If the split pseudo still has the same hard register as the
103 original pseudo after the subsequent assignment pass or the
104 original pseudo was split, the opposite transformation is done on
105 the same pass for undoing inheritance. */
107 #undef REG_OK_STRICT
109 #include "config.h"
110 #include "system.h"
111 #include "coretypes.h"
112 #include "backend.h"
113 #include "target.h"
114 #include "rtl.h"
115 #include "tree.h"
116 #include "predict.h"
117 #include "df.h"
118 #include "memmodel.h"
119 #include "tm_p.h"
120 #include "expmed.h"
121 #include "optabs.h"
122 #include "regs.h"
123 #include "ira.h"
124 #include "recog.h"
125 #include "output.h"
126 #include "addresses.h"
127 #include "expr.h"
128 #include "cfgrtl.h"
129 #include "rtl-error.h"
130 #include "params.h"
131 #include "lra.h"
132 #include "lra-int.h"
133 #include "print-rtl.h"
135 /* Value of LRA_CURR_RELOAD_NUM at the beginning of BB of the current
136 insn. Remember that LRA_CURR_RELOAD_NUM is the number of emitted
137 reload insns. */
138 static int bb_reload_num;
140 /* The current insn being processed and corresponding its single set
141 (NULL otherwise), its data (basic block, the insn data, the insn
142 static data, and the mode of each operand). */
143 static rtx_insn *curr_insn;
144 static rtx curr_insn_set;
145 static basic_block curr_bb;
146 static lra_insn_recog_data_t curr_id;
147 static struct lra_static_insn_data *curr_static_id;
148 static machine_mode curr_operand_mode[MAX_RECOG_OPERANDS];
149 /* Mode of the register substituted by its equivalence with VOIDmode
150 (e.g. constant) and whose subreg is given operand of the current
151 insn. VOIDmode in all other cases. */
152 static machine_mode original_subreg_reg_mode[MAX_RECOG_OPERANDS];
156 /* Start numbers for new registers and insns at the current constraints
157 pass start. */
158 static int new_regno_start;
159 static int new_insn_uid_start;
161 /* If LOC is nonnull, strip any outer subreg from it. */
162 static inline rtx *
163 strip_subreg (rtx *loc)
165 return loc && GET_CODE (*loc) == SUBREG ? &SUBREG_REG (*loc) : loc;
168 /* Return hard regno of REGNO or if it is was not assigned to a hard
169 register, use a hard register from its allocno class. */
170 static int
171 get_try_hard_regno (int regno)
173 int hard_regno;
174 enum reg_class rclass;
176 if ((hard_regno = regno) >= FIRST_PSEUDO_REGISTER)
177 hard_regno = lra_get_regno_hard_regno (regno);
178 if (hard_regno >= 0)
179 return hard_regno;
180 rclass = lra_get_allocno_class (regno);
181 if (rclass == NO_REGS)
182 return -1;
183 return ira_class_hard_regs[rclass][0];
186 /* Return the hard regno of X after removing its subreg. If X is not
187 a register or a subreg of a register, return -1. If X is a pseudo,
188 use its assignment. If FINAL_P return the final hard regno which will
189 be after elimination. */
190 static int
191 get_hard_regno (rtx x, bool final_p)
193 rtx reg;
194 int hard_regno;
196 reg = x;
197 if (SUBREG_P (x))
198 reg = SUBREG_REG (x);
199 if (! REG_P (reg))
200 return -1;
201 if (! HARD_REGISTER_NUM_P (hard_regno = REGNO (reg)))
202 hard_regno = lra_get_regno_hard_regno (hard_regno);
203 if (hard_regno < 0)
204 return -1;
205 if (final_p)
206 hard_regno = lra_get_elimination_hard_regno (hard_regno);
207 if (SUBREG_P (x))
208 hard_regno += subreg_regno_offset (hard_regno, GET_MODE (reg),
209 SUBREG_BYTE (x), GET_MODE (x));
210 return hard_regno;
213 /* If REGNO is a hard register or has been allocated a hard register,
214 return the class of that register. If REGNO is a reload pseudo
215 created by the current constraints pass, return its allocno class.
216 Return NO_REGS otherwise. */
217 static enum reg_class
218 get_reg_class (int regno)
220 int hard_regno;
222 if (! HARD_REGISTER_NUM_P (hard_regno = regno))
223 hard_regno = lra_get_regno_hard_regno (regno);
224 if (hard_regno >= 0)
226 hard_regno = lra_get_elimination_hard_regno (hard_regno);
227 return REGNO_REG_CLASS (hard_regno);
229 if (regno >= new_regno_start)
230 return lra_get_allocno_class (regno);
231 return NO_REGS;
234 /* Return true if REG satisfies (or will satisfy) reg class constraint
235 CL. Use elimination first if REG is a hard register. If REG is a
236 reload pseudo created by this constraints pass, assume that it will
237 be allocated a hard register from its allocno class, but allow that
238 class to be narrowed to CL if it is currently a superset of CL.
240 If NEW_CLASS is nonnull, set *NEW_CLASS to the new allocno class of
241 REGNO (reg), or NO_REGS if no change in its class was needed. */
242 static bool
243 in_class_p (rtx reg, enum reg_class cl, enum reg_class *new_class)
245 enum reg_class rclass, common_class;
246 machine_mode reg_mode;
247 int class_size, hard_regno, nregs, i, j;
248 int regno = REGNO (reg);
250 if (new_class != NULL)
251 *new_class = NO_REGS;
252 if (regno < FIRST_PSEUDO_REGISTER)
254 rtx final_reg = reg;
255 rtx *final_loc = &final_reg;
257 lra_eliminate_reg_if_possible (final_loc);
258 return TEST_HARD_REG_BIT (reg_class_contents[cl], REGNO (*final_loc));
260 reg_mode = GET_MODE (reg);
261 rclass = get_reg_class (regno);
262 if (regno < new_regno_start
263 /* Do not allow the constraints for reload instructions to
264 influence the classes of new pseudos. These reloads are
265 typically moves that have many alternatives, and restricting
266 reload pseudos for one alternative may lead to situations
267 where other reload pseudos are no longer allocatable. */
268 || (INSN_UID (curr_insn) >= new_insn_uid_start
269 && curr_insn_set != NULL
270 && ((OBJECT_P (SET_SRC (curr_insn_set))
271 && ! CONSTANT_P (SET_SRC (curr_insn_set)))
272 || (GET_CODE (SET_SRC (curr_insn_set)) == SUBREG
273 && OBJECT_P (SUBREG_REG (SET_SRC (curr_insn_set)))
274 && ! CONSTANT_P (SUBREG_REG (SET_SRC (curr_insn_set)))))))
275 /* When we don't know what class will be used finally for reload
276 pseudos, we use ALL_REGS. */
277 return ((regno >= new_regno_start && rclass == ALL_REGS)
278 || (rclass != NO_REGS && ira_class_subset_p[rclass][cl]
279 && ! hard_reg_set_subset_p (reg_class_contents[cl],
280 lra_no_alloc_regs)));
281 else
283 common_class = ira_reg_class_subset[rclass][cl];
284 if (new_class != NULL)
285 *new_class = common_class;
286 if (hard_reg_set_subset_p (reg_class_contents[common_class],
287 lra_no_alloc_regs))
288 return false;
289 /* Check that there are enough allocatable regs. */
290 class_size = ira_class_hard_regs_num[common_class];
291 for (i = 0; i < class_size; i++)
293 hard_regno = ira_class_hard_regs[common_class][i];
294 nregs = hard_regno_nregs[hard_regno][reg_mode];
295 if (nregs == 1)
296 return true;
297 for (j = 0; j < nregs; j++)
298 if (TEST_HARD_REG_BIT (lra_no_alloc_regs, hard_regno + j)
299 || ! TEST_HARD_REG_BIT (reg_class_contents[common_class],
300 hard_regno + j))
301 break;
302 if (j >= nregs)
303 return true;
305 return false;
309 /* Return true if REGNO satisfies a memory constraint. */
310 static bool
311 in_mem_p (int regno)
313 return get_reg_class (regno) == NO_REGS;
316 /* Return 1 if ADDR is a valid memory address for mode MODE in address
317 space AS, and check that each pseudo has the proper kind of hard
318 reg. */
319 static int
320 valid_address_p (machine_mode mode ATTRIBUTE_UNUSED,
321 rtx addr, addr_space_t as)
323 #ifdef GO_IF_LEGITIMATE_ADDRESS
324 lra_assert (ADDR_SPACE_GENERIC_P (as));
325 GO_IF_LEGITIMATE_ADDRESS (mode, addr, win);
326 return 0;
328 win:
329 return 1;
330 #else
331 return targetm.addr_space.legitimate_address_p (mode, addr, 0, as);
332 #endif
335 namespace {
336 /* Temporarily eliminates registers in an address (for the lifetime of
337 the object). */
338 class address_eliminator {
339 public:
340 address_eliminator (struct address_info *ad);
341 ~address_eliminator ();
343 private:
344 struct address_info *m_ad;
345 rtx *m_base_loc;
346 rtx m_base_reg;
347 rtx *m_index_loc;
348 rtx m_index_reg;
352 address_eliminator::address_eliminator (struct address_info *ad)
353 : m_ad (ad),
354 m_base_loc (strip_subreg (ad->base_term)),
355 m_base_reg (NULL_RTX),
356 m_index_loc (strip_subreg (ad->index_term)),
357 m_index_reg (NULL_RTX)
359 if (m_base_loc != NULL)
361 m_base_reg = *m_base_loc;
362 lra_eliminate_reg_if_possible (m_base_loc);
363 if (m_ad->base_term2 != NULL)
364 *m_ad->base_term2 = *m_ad->base_term;
366 if (m_index_loc != NULL)
368 m_index_reg = *m_index_loc;
369 lra_eliminate_reg_if_possible (m_index_loc);
373 address_eliminator::~address_eliminator ()
375 if (m_base_loc && *m_base_loc != m_base_reg)
377 *m_base_loc = m_base_reg;
378 if (m_ad->base_term2 != NULL)
379 *m_ad->base_term2 = *m_ad->base_term;
381 if (m_index_loc && *m_index_loc != m_index_reg)
382 *m_index_loc = m_index_reg;
385 /* Return true if the eliminated form of AD is a legitimate target address. */
386 static bool
387 valid_address_p (struct address_info *ad)
389 address_eliminator eliminator (ad);
390 return valid_address_p (ad->mode, *ad->outer, ad->as);
393 /* Return true if the eliminated form of memory reference OP satisfies
394 extra (special) memory constraint CONSTRAINT. */
395 static bool
396 satisfies_memory_constraint_p (rtx op, enum constraint_num constraint)
398 struct address_info ad;
400 decompose_mem_address (&ad, op);
401 address_eliminator eliminator (&ad);
402 return constraint_satisfied_p (op, constraint);
405 /* Return true if the eliminated form of address AD satisfies extra
406 address constraint CONSTRAINT. */
407 static bool
408 satisfies_address_constraint_p (struct address_info *ad,
409 enum constraint_num constraint)
411 address_eliminator eliminator (ad);
412 return constraint_satisfied_p (*ad->outer, constraint);
415 /* Return true if the eliminated form of address OP satisfies extra
416 address constraint CONSTRAINT. */
417 static bool
418 satisfies_address_constraint_p (rtx op, enum constraint_num constraint)
420 struct address_info ad;
422 decompose_lea_address (&ad, &op);
423 return satisfies_address_constraint_p (&ad, constraint);
426 /* Initiate equivalences for LRA. As we keep original equivalences
427 before any elimination, we need to make copies otherwise any change
428 in insns might change the equivalences. */
429 void
430 lra_init_equiv (void)
432 ira_expand_reg_equiv ();
433 for (int i = FIRST_PSEUDO_REGISTER; i < max_reg_num (); i++)
435 rtx res;
437 if ((res = ira_reg_equiv[i].memory) != NULL_RTX)
438 ira_reg_equiv[i].memory = copy_rtx (res);
439 if ((res = ira_reg_equiv[i].invariant) != NULL_RTX)
440 ira_reg_equiv[i].invariant = copy_rtx (res);
444 static rtx loc_equivalence_callback (rtx, const_rtx, void *);
446 /* Update equivalence for REGNO. We need to this as the equivalence
447 might contain other pseudos which are changed by their
448 equivalences. */
449 static void
450 update_equiv (int regno)
452 rtx x;
454 if ((x = ira_reg_equiv[regno].memory) != NULL_RTX)
455 ira_reg_equiv[regno].memory
456 = simplify_replace_fn_rtx (x, NULL_RTX, loc_equivalence_callback,
457 NULL_RTX);
458 if ((x = ira_reg_equiv[regno].invariant) != NULL_RTX)
459 ira_reg_equiv[regno].invariant
460 = simplify_replace_fn_rtx (x, NULL_RTX, loc_equivalence_callback,
461 NULL_RTX);
464 /* If we have decided to substitute X with another value, return that
465 value, otherwise return X. */
466 static rtx
467 get_equiv (rtx x)
469 int regno;
470 rtx res;
472 if (! REG_P (x) || (regno = REGNO (x)) < FIRST_PSEUDO_REGISTER
473 || ! ira_reg_equiv[regno].defined_p
474 || ! ira_reg_equiv[regno].profitable_p
475 || lra_get_regno_hard_regno (regno) >= 0)
476 return x;
477 if ((res = ira_reg_equiv[regno].memory) != NULL_RTX)
479 if (targetm.cannot_substitute_mem_equiv_p (res))
480 return x;
481 return res;
483 if ((res = ira_reg_equiv[regno].constant) != NULL_RTX)
484 return res;
485 if ((res = ira_reg_equiv[regno].invariant) != NULL_RTX)
486 return res;
487 gcc_unreachable ();
490 /* If we have decided to substitute X with the equivalent value,
491 return that value after elimination for INSN, otherwise return
492 X. */
493 static rtx
494 get_equiv_with_elimination (rtx x, rtx_insn *insn)
496 rtx res = get_equiv (x);
498 if (x == res || CONSTANT_P (res))
499 return res;
500 return lra_eliminate_regs_1 (insn, res, GET_MODE (res),
501 false, false, 0, true);
504 /* Set up curr_operand_mode. */
505 static void
506 init_curr_operand_mode (void)
508 int nop = curr_static_id->n_operands;
509 for (int i = 0; i < nop; i++)
511 machine_mode mode = GET_MODE (*curr_id->operand_loc[i]);
512 if (mode == VOIDmode)
514 /* The .md mode for address operands is the mode of the
515 addressed value rather than the mode of the address itself. */
516 if (curr_id->icode >= 0 && curr_static_id->operand[i].is_address)
517 mode = Pmode;
518 else
519 mode = curr_static_id->operand[i].mode;
521 curr_operand_mode[i] = mode;
527 /* The page contains code to reuse input reloads. */
529 /* Structure describes input reload of the current insns. */
530 struct input_reload
532 /* True for input reload of matched operands. */
533 bool match_p;
534 /* Reloaded value. */
535 rtx input;
536 /* Reload pseudo used. */
537 rtx reg;
540 /* The number of elements in the following array. */
541 static int curr_insn_input_reloads_num;
542 /* Array containing info about input reloads. It is used to find the
543 same input reload and reuse the reload pseudo in this case. */
544 static struct input_reload curr_insn_input_reloads[LRA_MAX_INSN_RELOADS];
546 /* Initiate data concerning reuse of input reloads for the current
547 insn. */
548 static void
549 init_curr_insn_input_reloads (void)
551 curr_insn_input_reloads_num = 0;
554 /* Create a new pseudo using MODE, RCLASS, ORIGINAL or reuse already
555 created input reload pseudo (only if TYPE is not OP_OUT). Don't
556 reuse pseudo if IN_SUBREG_P is true and the reused pseudo should be
557 wrapped up in SUBREG. The result pseudo is returned through
558 RESULT_REG. Return TRUE if we created a new pseudo, FALSE if we
559 reused the already created input reload pseudo. Use TITLE to
560 describe new registers for debug purposes. */
561 static bool
562 get_reload_reg (enum op_type type, machine_mode mode, rtx original,
563 enum reg_class rclass, bool in_subreg_p,
564 const char *title, rtx *result_reg)
566 int i, regno;
567 enum reg_class new_class;
568 bool unique_p = false;
570 if (type == OP_OUT)
572 *result_reg
573 = lra_create_new_reg_with_unique_value (mode, original, rclass, title);
574 return true;
576 /* Prevent reuse value of expression with side effects,
577 e.g. volatile memory. */
578 if (! side_effects_p (original))
579 for (i = 0; i < curr_insn_input_reloads_num; i++)
581 if (! curr_insn_input_reloads[i].match_p
582 && rtx_equal_p (curr_insn_input_reloads[i].input, original)
583 && in_class_p (curr_insn_input_reloads[i].reg, rclass, &new_class))
585 rtx reg = curr_insn_input_reloads[i].reg;
586 regno = REGNO (reg);
587 /* If input is equal to original and both are VOIDmode,
588 GET_MODE (reg) might be still different from mode.
589 Ensure we don't return *result_reg with wrong mode. */
590 if (GET_MODE (reg) != mode)
592 if (in_subreg_p)
593 continue;
594 if (GET_MODE_SIZE (GET_MODE (reg)) < GET_MODE_SIZE (mode))
595 continue;
596 reg = lowpart_subreg (mode, reg, GET_MODE (reg));
597 if (reg == NULL_RTX || GET_CODE (reg) != SUBREG)
598 continue;
600 *result_reg = reg;
601 if (lra_dump_file != NULL)
603 fprintf (lra_dump_file, " Reuse r%d for reload ", regno);
604 dump_value_slim (lra_dump_file, original, 1);
606 if (new_class != lra_get_allocno_class (regno))
607 lra_change_class (regno, new_class, ", change to", false);
608 if (lra_dump_file != NULL)
609 fprintf (lra_dump_file, "\n");
610 return false;
612 /* If we have an input reload with a different mode, make sure it
613 will get a different hard reg. */
614 else if (REG_P (original)
615 && REG_P (curr_insn_input_reloads[i].input)
616 && REGNO (original) == REGNO (curr_insn_input_reloads[i].input)
617 && (GET_MODE (original)
618 != GET_MODE (curr_insn_input_reloads[i].input)))
619 unique_p = true;
621 *result_reg = (unique_p
622 ? lra_create_new_reg_with_unique_value
623 : lra_create_new_reg) (mode, original, rclass, title);
624 lra_assert (curr_insn_input_reloads_num < LRA_MAX_INSN_RELOADS);
625 curr_insn_input_reloads[curr_insn_input_reloads_num].input = original;
626 curr_insn_input_reloads[curr_insn_input_reloads_num].match_p = false;
627 curr_insn_input_reloads[curr_insn_input_reloads_num++].reg = *result_reg;
628 return true;
633 /* The page contains code to extract memory address parts. */
635 /* Wrapper around REGNO_OK_FOR_INDEX_P, to allow pseudos. */
636 static inline bool
637 ok_for_index_p_nonstrict (rtx reg)
639 unsigned regno = REGNO (reg);
641 return regno >= FIRST_PSEUDO_REGISTER || REGNO_OK_FOR_INDEX_P (regno);
644 /* A version of regno_ok_for_base_p for use here, when all pseudos
645 should count as OK. Arguments as for regno_ok_for_base_p. */
646 static inline bool
647 ok_for_base_p_nonstrict (rtx reg, machine_mode mode, addr_space_t as,
648 enum rtx_code outer_code, enum rtx_code index_code)
650 unsigned regno = REGNO (reg);
652 if (regno >= FIRST_PSEUDO_REGISTER)
653 return true;
654 return ok_for_base_p_1 (regno, mode, as, outer_code, index_code);
659 /* The page contains major code to choose the current insn alternative
660 and generate reloads for it. */
662 /* Return the offset from REGNO of the least significant register
663 in (reg:MODE REGNO).
665 This function is used to tell whether two registers satisfy
666 a matching constraint. (reg:MODE1 REGNO1) matches (reg:MODE2 REGNO2) if:
668 REGNO1 + lra_constraint_offset (REGNO1, MODE1)
669 == REGNO2 + lra_constraint_offset (REGNO2, MODE2) */
671 lra_constraint_offset (int regno, machine_mode mode)
673 lra_assert (regno < FIRST_PSEUDO_REGISTER);
674 if (WORDS_BIG_ENDIAN && GET_MODE_SIZE (mode) > UNITS_PER_WORD
675 && SCALAR_INT_MODE_P (mode))
676 return hard_regno_nregs[regno][mode] - 1;
677 return 0;
680 /* Like rtx_equal_p except that it allows a REG and a SUBREG to match
681 if they are the same hard reg, and has special hacks for
682 auto-increment and auto-decrement. This is specifically intended for
683 process_alt_operands to use in determining whether two operands
684 match. X is the operand whose number is the lower of the two.
686 It is supposed that X is the output operand and Y is the input
687 operand. Y_HARD_REGNO is the final hard regno of register Y or
688 register in subreg Y as we know it now. Otherwise, it is a
689 negative value. */
690 static bool
691 operands_match_p (rtx x, rtx y, int y_hard_regno)
693 int i;
694 RTX_CODE code = GET_CODE (x);
695 const char *fmt;
697 if (x == y)
698 return true;
699 if ((code == REG || (code == SUBREG && REG_P (SUBREG_REG (x))))
700 && (REG_P (y) || (GET_CODE (y) == SUBREG && REG_P (SUBREG_REG (y)))))
702 int j;
704 i = get_hard_regno (x, false);
705 if (i < 0)
706 goto slow;
708 if ((j = y_hard_regno) < 0)
709 goto slow;
711 i += lra_constraint_offset (i, GET_MODE (x));
712 j += lra_constraint_offset (j, GET_MODE (y));
714 return i == j;
717 /* If two operands must match, because they are really a single
718 operand of an assembler insn, then two post-increments are invalid
719 because the assembler insn would increment only once. On the
720 other hand, a post-increment matches ordinary indexing if the
721 post-increment is the output operand. */
722 if (code == POST_DEC || code == POST_INC || code == POST_MODIFY)
723 return operands_match_p (XEXP (x, 0), y, y_hard_regno);
725 /* Two pre-increments are invalid because the assembler insn would
726 increment only once. On the other hand, a pre-increment matches
727 ordinary indexing if the pre-increment is the input operand. */
728 if (GET_CODE (y) == PRE_DEC || GET_CODE (y) == PRE_INC
729 || GET_CODE (y) == PRE_MODIFY)
730 return operands_match_p (x, XEXP (y, 0), -1);
732 slow:
734 if (code == REG && REG_P (y))
735 return REGNO (x) == REGNO (y);
737 if (code == REG && GET_CODE (y) == SUBREG && REG_P (SUBREG_REG (y))
738 && x == SUBREG_REG (y))
739 return true;
740 if (GET_CODE (y) == REG && code == SUBREG && REG_P (SUBREG_REG (x))
741 && SUBREG_REG (x) == y)
742 return true;
744 /* Now we have disposed of all the cases in which different rtx
745 codes can match. */
746 if (code != GET_CODE (y))
747 return false;
749 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
750 if (GET_MODE (x) != GET_MODE (y))
751 return false;
753 switch (code)
755 CASE_CONST_UNIQUE:
756 return false;
758 case LABEL_REF:
759 return label_ref_label (x) == label_ref_label (y);
760 case SYMBOL_REF:
761 return XSTR (x, 0) == XSTR (y, 0);
763 default:
764 break;
767 /* Compare the elements. If any pair of corresponding elements fail
768 to match, return false for the whole things. */
770 fmt = GET_RTX_FORMAT (code);
771 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
773 int val, j;
774 switch (fmt[i])
776 case 'w':
777 if (XWINT (x, i) != XWINT (y, i))
778 return false;
779 break;
781 case 'i':
782 if (XINT (x, i) != XINT (y, i))
783 return false;
784 break;
786 case 'e':
787 val = operands_match_p (XEXP (x, i), XEXP (y, i), -1);
788 if (val == 0)
789 return false;
790 break;
792 case '0':
793 break;
795 case 'E':
796 if (XVECLEN (x, i) != XVECLEN (y, i))
797 return false;
798 for (j = XVECLEN (x, i) - 1; j >= 0; --j)
800 val = operands_match_p (XVECEXP (x, i, j), XVECEXP (y, i, j), -1);
801 if (val == 0)
802 return false;
804 break;
806 /* It is believed that rtx's at this level will never
807 contain anything but integers and other rtx's, except for
808 within LABEL_REFs and SYMBOL_REFs. */
809 default:
810 gcc_unreachable ();
813 return true;
816 /* True if X is a constant that can be forced into the constant pool.
817 MODE is the mode of the operand, or VOIDmode if not known. */
818 #define CONST_POOL_OK_P(MODE, X) \
819 ((MODE) != VOIDmode \
820 && CONSTANT_P (X) \
821 && GET_CODE (X) != HIGH \
822 && !targetm.cannot_force_const_mem (MODE, X))
824 /* True if C is a non-empty register class that has too few registers
825 to be safely used as a reload target class. */
826 #define SMALL_REGISTER_CLASS_P(C) \
827 (ira_class_hard_regs_num [(C)] == 1 \
828 || (ira_class_hard_regs_num [(C)] >= 1 \
829 && targetm.class_likely_spilled_p (C)))
831 /* If REG is a reload pseudo, try to make its class satisfying CL. */
832 static void
833 narrow_reload_pseudo_class (rtx reg, enum reg_class cl)
835 enum reg_class rclass;
837 /* Do not make more accurate class from reloads generated. They are
838 mostly moves with a lot of constraints. Making more accurate
839 class may results in very narrow class and impossibility of find
840 registers for several reloads of one insn. */
841 if (INSN_UID (curr_insn) >= new_insn_uid_start)
842 return;
843 if (GET_CODE (reg) == SUBREG)
844 reg = SUBREG_REG (reg);
845 if (! REG_P (reg) || (int) REGNO (reg) < new_regno_start)
846 return;
847 if (in_class_p (reg, cl, &rclass) && rclass != cl)
848 lra_change_class (REGNO (reg), rclass, " Change to", true);
851 /* Searches X for any reference to a reg with the same value as REGNO,
852 returning the rtx of the reference found if any. Otherwise,
853 returns NULL_RTX. */
854 static rtx
855 regno_val_use_in (unsigned int regno, rtx x)
857 const char *fmt;
858 int i, j;
859 rtx tem;
861 if (REG_P (x) && lra_reg_info[REGNO (x)].val == lra_reg_info[regno].val)
862 return x;
864 fmt = GET_RTX_FORMAT (GET_CODE (x));
865 for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
867 if (fmt[i] == 'e')
869 if ((tem = regno_val_use_in (regno, XEXP (x, i))))
870 return tem;
872 else if (fmt[i] == 'E')
873 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
874 if ((tem = regno_val_use_in (regno , XVECEXP (x, i, j))))
875 return tem;
878 return NULL_RTX;
881 /* Return true if all current insn non-output operands except INS (it
882 has a negaitve end marker) do not use pseudos with the same value
883 as REGNO. */
884 static bool
885 check_conflict_input_operands (int regno, signed char *ins)
887 int in;
888 int n_operands = curr_static_id->n_operands;
890 for (int nop = 0; nop < n_operands; nop++)
891 if (! curr_static_id->operand[nop].is_operator
892 && curr_static_id->operand[nop].type != OP_OUT)
894 for (int i = 0; (in = ins[i]) >= 0; i++)
895 if (in == nop)
896 break;
897 if (in < 0
898 && regno_val_use_in (regno, *curr_id->operand_loc[nop]) != NULL_RTX)
899 return false;
901 return true;
904 /* Generate reloads for matching OUT and INS (array of input operand
905 numbers with end marker -1) with reg class GOAL_CLASS, considering
906 output operands OUTS (similar array to INS) needing to be in different
907 registers. Add input and output reloads correspondingly to the lists
908 *BEFORE and *AFTER. OUT might be negative. In this case we generate
909 input reloads for matched input operands INS. EARLY_CLOBBER_P is a flag
910 that the output operand is early clobbered for chosen alternative. */
911 static void
912 match_reload (signed char out, signed char *ins, signed char *outs,
913 enum reg_class goal_class, rtx_insn **before,
914 rtx_insn **after, bool early_clobber_p)
916 bool out_conflict;
917 int i, in;
918 rtx new_in_reg, new_out_reg, reg;
919 machine_mode inmode, outmode;
920 rtx in_rtx = *curr_id->operand_loc[ins[0]];
921 rtx out_rtx = out < 0 ? in_rtx : *curr_id->operand_loc[out];
923 inmode = curr_operand_mode[ins[0]];
924 outmode = out < 0 ? inmode : curr_operand_mode[out];
925 push_to_sequence (*before);
926 if (inmode != outmode)
928 if (GET_MODE_SIZE (inmode) > GET_MODE_SIZE (outmode))
930 reg = new_in_reg
931 = lra_create_new_reg_with_unique_value (inmode, in_rtx,
932 goal_class, "");
933 if (SCALAR_INT_MODE_P (inmode))
934 new_out_reg = gen_lowpart_SUBREG (outmode, reg);
935 else
936 new_out_reg = gen_rtx_SUBREG (outmode, reg, 0);
937 LRA_SUBREG_P (new_out_reg) = 1;
938 /* If the input reg is dying here, we can use the same hard
939 register for REG and IN_RTX. We do it only for original
940 pseudos as reload pseudos can die although original
941 pseudos still live where reload pseudos dies. */
942 if (REG_P (in_rtx) && (int) REGNO (in_rtx) < lra_new_regno_start
943 && find_regno_note (curr_insn, REG_DEAD, REGNO (in_rtx))
944 && (!early_clobber_p
945 || check_conflict_input_operands(REGNO (in_rtx), ins)))
946 lra_assign_reg_val (REGNO (in_rtx), REGNO (reg));
948 else
950 reg = new_out_reg
951 = lra_create_new_reg_with_unique_value (outmode, out_rtx,
952 goal_class, "");
953 if (SCALAR_INT_MODE_P (outmode))
954 new_in_reg = gen_lowpart_SUBREG (inmode, reg);
955 else
956 new_in_reg = gen_rtx_SUBREG (inmode, reg, 0);
957 /* NEW_IN_REG is non-paradoxical subreg. We don't want
958 NEW_OUT_REG living above. We add clobber clause for
959 this. This is just a temporary clobber. We can remove
960 it at the end of LRA work. */
961 rtx_insn *clobber = emit_clobber (new_out_reg);
962 LRA_TEMP_CLOBBER_P (PATTERN (clobber)) = 1;
963 LRA_SUBREG_P (new_in_reg) = 1;
964 if (GET_CODE (in_rtx) == SUBREG)
966 rtx subreg_reg = SUBREG_REG (in_rtx);
968 /* If SUBREG_REG is dying here and sub-registers IN_RTX
969 and NEW_IN_REG are similar, we can use the same hard
970 register for REG and SUBREG_REG. */
971 if (REG_P (subreg_reg)
972 && (int) REGNO (subreg_reg) < lra_new_regno_start
973 && GET_MODE (subreg_reg) == outmode
974 && SUBREG_BYTE (in_rtx) == SUBREG_BYTE (new_in_reg)
975 && find_regno_note (curr_insn, REG_DEAD, REGNO (subreg_reg))
976 && (! early_clobber_p
977 || check_conflict_input_operands (REGNO (subreg_reg),
978 ins)))
979 lra_assign_reg_val (REGNO (subreg_reg), REGNO (reg));
983 else
985 /* Pseudos have values -- see comments for lra_reg_info.
986 Different pseudos with the same value do not conflict even if
987 they live in the same place. When we create a pseudo we
988 assign value of original pseudo (if any) from which we
989 created the new pseudo. If we create the pseudo from the
990 input pseudo, the new pseudo will have no conflict with the
991 input pseudo which is wrong when the input pseudo lives after
992 the insn and as the new pseudo value is changed by the insn
993 output. Therefore we create the new pseudo from the output
994 except the case when we have single matched dying input
995 pseudo.
997 We cannot reuse the current output register because we might
998 have a situation like "a <- a op b", where the constraints
999 force the second input operand ("b") to match the output
1000 operand ("a"). "b" must then be copied into a new register
1001 so that it doesn't clobber the current value of "a".
1003 We can not use the same value if the output pseudo is
1004 early clobbered or the input pseudo is mentioned in the
1005 output, e.g. as an address part in memory, because
1006 output reload will actually extend the pseudo liveness.
1007 We don't care about eliminable hard regs here as we are
1008 interesting only in pseudos. */
1010 /* Matching input's register value is the same as one of the other
1011 output operand. Output operands in a parallel insn must be in
1012 different registers. */
1013 out_conflict = false;
1014 if (REG_P (in_rtx))
1016 for (i = 0; outs[i] >= 0; i++)
1018 rtx other_out_rtx = *curr_id->operand_loc[outs[i]];
1019 if (REG_P (other_out_rtx)
1020 && (regno_val_use_in (REGNO (in_rtx), other_out_rtx)
1021 != NULL_RTX))
1023 out_conflict = true;
1024 break;
1029 new_in_reg = new_out_reg
1030 = (! early_clobber_p && ins[1] < 0 && REG_P (in_rtx)
1031 && (int) REGNO (in_rtx) < lra_new_regno_start
1032 && find_regno_note (curr_insn, REG_DEAD, REGNO (in_rtx))
1033 && (! early_clobber_p
1034 || check_conflict_input_operands (REGNO (in_rtx), ins))
1035 && (out < 0
1036 || regno_val_use_in (REGNO (in_rtx), out_rtx) == NULL_RTX)
1037 && !out_conflict
1038 ? lra_create_new_reg (inmode, in_rtx, goal_class, "")
1039 : lra_create_new_reg_with_unique_value (outmode, out_rtx,
1040 goal_class, ""));
1042 /* In operand can be got from transformations before processing insn
1043 constraints. One example of such transformations is subreg
1044 reloading (see function simplify_operand_subreg). The new
1045 pseudos created by the transformations might have inaccurate
1046 class (ALL_REGS) and we should make their classes more
1047 accurate. */
1048 narrow_reload_pseudo_class (in_rtx, goal_class);
1049 lra_emit_move (copy_rtx (new_in_reg), in_rtx);
1050 *before = get_insns ();
1051 end_sequence ();
1052 /* Add the new pseudo to consider values of subsequent input reload
1053 pseudos. */
1054 lra_assert (curr_insn_input_reloads_num < LRA_MAX_INSN_RELOADS);
1055 curr_insn_input_reloads[curr_insn_input_reloads_num].input = in_rtx;
1056 curr_insn_input_reloads[curr_insn_input_reloads_num].match_p = true;
1057 curr_insn_input_reloads[curr_insn_input_reloads_num++].reg = new_in_reg;
1058 for (i = 0; (in = ins[i]) >= 0; i++)
1060 lra_assert
1061 (GET_MODE (*curr_id->operand_loc[in]) == VOIDmode
1062 || GET_MODE (new_in_reg) == GET_MODE (*curr_id->operand_loc[in]));
1063 *curr_id->operand_loc[in] = new_in_reg;
1065 lra_update_dups (curr_id, ins);
1066 if (out < 0)
1067 return;
1068 /* See a comment for the input operand above. */
1069 narrow_reload_pseudo_class (out_rtx, goal_class);
1070 if (find_reg_note (curr_insn, REG_UNUSED, out_rtx) == NULL_RTX)
1072 start_sequence ();
1073 lra_emit_move (out_rtx, copy_rtx (new_out_reg));
1074 emit_insn (*after);
1075 *after = get_insns ();
1076 end_sequence ();
1078 *curr_id->operand_loc[out] = new_out_reg;
1079 lra_update_dup (curr_id, out);
1082 /* Return register class which is union of all reg classes in insn
1083 constraint alternative string starting with P. */
1084 static enum reg_class
1085 reg_class_from_constraints (const char *p)
1087 int c, len;
1088 enum reg_class op_class = NO_REGS;
1091 switch ((c = *p, len = CONSTRAINT_LEN (c, p)), c)
1093 case '#':
1094 case ',':
1095 return op_class;
1097 case 'g':
1098 op_class = reg_class_subunion[op_class][GENERAL_REGS];
1099 break;
1101 default:
1102 enum constraint_num cn = lookup_constraint (p);
1103 enum reg_class cl = reg_class_for_constraint (cn);
1104 if (cl == NO_REGS)
1106 if (insn_extra_address_constraint (cn))
1107 op_class
1108 = (reg_class_subunion
1109 [op_class][base_reg_class (VOIDmode, ADDR_SPACE_GENERIC,
1110 ADDRESS, SCRATCH)]);
1111 break;
1114 op_class = reg_class_subunion[op_class][cl];
1115 break;
1117 while ((p += len), c);
1118 return op_class;
1121 /* If OP is a register, return the class of the register as per
1122 get_reg_class, otherwise return NO_REGS. */
1123 static inline enum reg_class
1124 get_op_class (rtx op)
1126 return REG_P (op) ? get_reg_class (REGNO (op)) : NO_REGS;
1129 /* Return generated insn mem_pseudo:=val if TO_P or val:=mem_pseudo
1130 otherwise. If modes of MEM_PSEUDO and VAL are different, use
1131 SUBREG for VAL to make them equal. */
1132 static rtx_insn *
1133 emit_spill_move (bool to_p, rtx mem_pseudo, rtx val)
1135 if (GET_MODE (mem_pseudo) != GET_MODE (val))
1137 /* Usually size of mem_pseudo is greater than val size but in
1138 rare cases it can be less as it can be defined by target
1139 dependent macro HARD_REGNO_CALLER_SAVE_MODE. */
1140 if (! MEM_P (val))
1142 val = gen_lowpart_SUBREG (GET_MODE (mem_pseudo),
1143 GET_CODE (val) == SUBREG
1144 ? SUBREG_REG (val) : val);
1145 LRA_SUBREG_P (val) = 1;
1147 else
1149 mem_pseudo = gen_lowpart_SUBREG (GET_MODE (val), mem_pseudo);
1150 LRA_SUBREG_P (mem_pseudo) = 1;
1153 return to_p ? gen_move_insn (mem_pseudo, val)
1154 : gen_move_insn (val, mem_pseudo);
1157 /* Process a special case insn (register move), return true if we
1158 don't need to process it anymore. INSN should be a single set
1159 insn. Set up that RTL was changed through CHANGE_P and macro
1160 SECONDARY_MEMORY_NEEDED says to use secondary memory through
1161 SEC_MEM_P. */
1162 static bool
1163 check_and_process_move (bool *change_p, bool *sec_mem_p ATTRIBUTE_UNUSED)
1165 int sregno, dregno;
1166 rtx dest, src, dreg, sreg, new_reg, scratch_reg;
1167 rtx_insn *before;
1168 enum reg_class dclass, sclass, secondary_class;
1169 secondary_reload_info sri;
1171 lra_assert (curr_insn_set != NULL_RTX);
1172 dreg = dest = SET_DEST (curr_insn_set);
1173 sreg = src = SET_SRC (curr_insn_set);
1174 if (GET_CODE (dest) == SUBREG)
1175 dreg = SUBREG_REG (dest);
1176 if (GET_CODE (src) == SUBREG)
1177 sreg = SUBREG_REG (src);
1178 if (! (REG_P (dreg) || MEM_P (dreg)) || ! (REG_P (sreg) || MEM_P (sreg)))
1179 return false;
1180 sclass = dclass = NO_REGS;
1181 if (REG_P (dreg))
1182 dclass = get_reg_class (REGNO (dreg));
1183 gcc_assert (dclass < LIM_REG_CLASSES);
1184 if (dclass == ALL_REGS)
1185 /* ALL_REGS is used for new pseudos created by transformations
1186 like reload of SUBREG_REG (see function
1187 simplify_operand_subreg). We don't know their class yet. We
1188 should figure out the class from processing the insn
1189 constraints not in this fast path function. Even if ALL_REGS
1190 were a right class for the pseudo, secondary_... hooks usually
1191 are not define for ALL_REGS. */
1192 return false;
1193 if (REG_P (sreg))
1194 sclass = get_reg_class (REGNO (sreg));
1195 gcc_assert (sclass < LIM_REG_CLASSES);
1196 if (sclass == ALL_REGS)
1197 /* See comments above. */
1198 return false;
1199 if (sclass == NO_REGS && dclass == NO_REGS)
1200 return false;
1201 #ifdef SECONDARY_MEMORY_NEEDED
1202 if (SECONDARY_MEMORY_NEEDED (sclass, dclass, GET_MODE (src))
1203 #ifdef SECONDARY_MEMORY_NEEDED_MODE
1204 && ((sclass != NO_REGS && dclass != NO_REGS)
1205 || GET_MODE (src) != SECONDARY_MEMORY_NEEDED_MODE (GET_MODE (src)))
1206 #endif
1209 *sec_mem_p = true;
1210 return false;
1212 #endif
1213 if (! REG_P (dreg) || ! REG_P (sreg))
1214 return false;
1215 sri.prev_sri = NULL;
1216 sri.icode = CODE_FOR_nothing;
1217 sri.extra_cost = 0;
1218 secondary_class = NO_REGS;
1219 /* Set up hard register for a reload pseudo for hook
1220 secondary_reload because some targets just ignore unassigned
1221 pseudos in the hook. */
1222 if (dclass != NO_REGS && lra_get_regno_hard_regno (REGNO (dreg)) < 0)
1224 dregno = REGNO (dreg);
1225 reg_renumber[dregno] = ira_class_hard_regs[dclass][0];
1227 else
1228 dregno = -1;
1229 if (sclass != NO_REGS && lra_get_regno_hard_regno (REGNO (sreg)) < 0)
1231 sregno = REGNO (sreg);
1232 reg_renumber[sregno] = ira_class_hard_regs[sclass][0];
1234 else
1235 sregno = -1;
1236 if (sclass != NO_REGS)
1237 secondary_class
1238 = (enum reg_class) targetm.secondary_reload (false, dest,
1239 (reg_class_t) sclass,
1240 GET_MODE (src), &sri);
1241 if (sclass == NO_REGS
1242 || ((secondary_class != NO_REGS || sri.icode != CODE_FOR_nothing)
1243 && dclass != NO_REGS))
1245 enum reg_class old_sclass = secondary_class;
1246 secondary_reload_info old_sri = sri;
1248 sri.prev_sri = NULL;
1249 sri.icode = CODE_FOR_nothing;
1250 sri.extra_cost = 0;
1251 secondary_class
1252 = (enum reg_class) targetm.secondary_reload (true, src,
1253 (reg_class_t) dclass,
1254 GET_MODE (src), &sri);
1255 /* Check the target hook consistency. */
1256 lra_assert
1257 ((secondary_class == NO_REGS && sri.icode == CODE_FOR_nothing)
1258 || (old_sclass == NO_REGS && old_sri.icode == CODE_FOR_nothing)
1259 || (secondary_class == old_sclass && sri.icode == old_sri.icode));
1261 if (sregno >= 0)
1262 reg_renumber [sregno] = -1;
1263 if (dregno >= 0)
1264 reg_renumber [dregno] = -1;
1265 if (secondary_class == NO_REGS && sri.icode == CODE_FOR_nothing)
1266 return false;
1267 *change_p = true;
1268 new_reg = NULL_RTX;
1269 if (secondary_class != NO_REGS)
1270 new_reg = lra_create_new_reg_with_unique_value (GET_MODE (src), NULL_RTX,
1271 secondary_class,
1272 "secondary");
1273 start_sequence ();
1274 if (sri.icode == CODE_FOR_nothing)
1275 lra_emit_move (new_reg, src);
1276 else
1278 enum reg_class scratch_class;
1280 scratch_class = (reg_class_from_constraints
1281 (insn_data[sri.icode].operand[2].constraint));
1282 scratch_reg = (lra_create_new_reg_with_unique_value
1283 (insn_data[sri.icode].operand[2].mode, NULL_RTX,
1284 scratch_class, "scratch"));
1285 emit_insn (GEN_FCN (sri.icode) (new_reg != NULL_RTX ? new_reg : dest,
1286 src, scratch_reg));
1288 before = get_insns ();
1289 end_sequence ();
1290 lra_process_new_insns (curr_insn, before, NULL, "Inserting the move");
1291 if (new_reg != NULL_RTX)
1292 SET_SRC (curr_insn_set) = new_reg;
1293 else
1295 if (lra_dump_file != NULL)
1297 fprintf (lra_dump_file, "Deleting move %u\n", INSN_UID (curr_insn));
1298 dump_insn_slim (lra_dump_file, curr_insn);
1300 lra_set_insn_deleted (curr_insn);
1301 return true;
1303 return false;
1306 /* The following data describe the result of process_alt_operands.
1307 The data are used in curr_insn_transform to generate reloads. */
1309 /* The chosen reg classes which should be used for the corresponding
1310 operands. */
1311 static enum reg_class goal_alt[MAX_RECOG_OPERANDS];
1312 /* True if the operand should be the same as another operand and that
1313 other operand does not need a reload. */
1314 static bool goal_alt_match_win[MAX_RECOG_OPERANDS];
1315 /* True if the operand does not need a reload. */
1316 static bool goal_alt_win[MAX_RECOG_OPERANDS];
1317 /* True if the operand can be offsetable memory. */
1318 static bool goal_alt_offmemok[MAX_RECOG_OPERANDS];
1319 /* The number of an operand to which given operand can be matched to. */
1320 static int goal_alt_matches[MAX_RECOG_OPERANDS];
1321 /* The number of elements in the following array. */
1322 static int goal_alt_dont_inherit_ops_num;
1323 /* Numbers of operands whose reload pseudos should not be inherited. */
1324 static int goal_alt_dont_inherit_ops[MAX_RECOG_OPERANDS];
1325 /* True if the insn commutative operands should be swapped. */
1326 static bool goal_alt_swapped;
1327 /* The chosen insn alternative. */
1328 static int goal_alt_number;
1330 /* True if the corresponding operand is the result of an equivalence
1331 substitution. */
1332 static bool equiv_substition_p[MAX_RECOG_OPERANDS];
1334 /* The following five variables are used to choose the best insn
1335 alternative. They reflect final characteristics of the best
1336 alternative. */
1338 /* Number of necessary reloads and overall cost reflecting the
1339 previous value and other unpleasantness of the best alternative. */
1340 static int best_losers, best_overall;
1341 /* Overall number hard registers used for reloads. For example, on
1342 some targets we need 2 general registers to reload DFmode and only
1343 one floating point register. */
1344 static int best_reload_nregs;
1345 /* Overall number reflecting distances of previous reloading the same
1346 value. The distances are counted from the current BB start. It is
1347 used to improve inheritance chances. */
1348 static int best_reload_sum;
1350 /* True if the current insn should have no correspondingly input or
1351 output reloads. */
1352 static bool no_input_reloads_p, no_output_reloads_p;
1354 /* True if we swapped the commutative operands in the current
1355 insn. */
1356 static int curr_swapped;
1358 /* if CHECK_ONLY_P is false, arrange for address element *LOC to be a
1359 register of class CL. Add any input reloads to list BEFORE. AFTER
1360 is nonnull if *LOC is an automodified value; handle that case by
1361 adding the required output reloads to list AFTER. Return true if
1362 the RTL was changed.
1364 if CHECK_ONLY_P is true, check that the *LOC is a correct address
1365 register. Return false if the address register is correct. */
1366 static bool
1367 process_addr_reg (rtx *loc, bool check_only_p, rtx_insn **before, rtx_insn **after,
1368 enum reg_class cl)
1370 int regno;
1371 enum reg_class rclass, new_class;
1372 rtx reg;
1373 rtx new_reg;
1374 machine_mode mode;
1375 bool subreg_p, before_p = false;
1377 subreg_p = GET_CODE (*loc) == SUBREG;
1378 if (subreg_p)
1380 reg = SUBREG_REG (*loc);
1381 mode = GET_MODE (reg);
1383 /* For mode with size bigger than ptr_mode, there unlikely to be "mov"
1384 between two registers with different classes, but there normally will
1385 be "mov" which transfers element of vector register into the general
1386 register, and this normally will be a subreg which should be reloaded
1387 as a whole. This is particularly likely to be triggered when
1388 -fno-split-wide-types specified. */
1389 if (!REG_P (reg)
1390 || in_class_p (reg, cl, &new_class)
1391 || GET_MODE_SIZE (mode) <= GET_MODE_SIZE (ptr_mode))
1392 loc = &SUBREG_REG (*loc);
1395 reg = *loc;
1396 mode = GET_MODE (reg);
1397 if (! REG_P (reg))
1399 if (check_only_p)
1400 return true;
1401 /* Always reload memory in an address even if the target supports
1402 such addresses. */
1403 new_reg = lra_create_new_reg_with_unique_value (mode, reg, cl, "address");
1404 before_p = true;
1406 else
1408 regno = REGNO (reg);
1409 rclass = get_reg_class (regno);
1410 if (! check_only_p
1411 && (*loc = get_equiv_with_elimination (reg, curr_insn)) != reg)
1413 if (lra_dump_file != NULL)
1415 fprintf (lra_dump_file,
1416 "Changing pseudo %d in address of insn %u on equiv ",
1417 REGNO (reg), INSN_UID (curr_insn));
1418 dump_value_slim (lra_dump_file, *loc, 1);
1419 fprintf (lra_dump_file, "\n");
1421 *loc = copy_rtx (*loc);
1423 if (*loc != reg || ! in_class_p (reg, cl, &new_class))
1425 if (check_only_p)
1426 return true;
1427 reg = *loc;
1428 if (get_reload_reg (after == NULL ? OP_IN : OP_INOUT,
1429 mode, reg, cl, subreg_p, "address", &new_reg))
1430 before_p = true;
1432 else if (new_class != NO_REGS && rclass != new_class)
1434 if (check_only_p)
1435 return true;
1436 lra_change_class (regno, new_class, " Change to", true);
1437 return false;
1439 else
1440 return false;
1442 if (before_p)
1444 push_to_sequence (*before);
1445 lra_emit_move (new_reg, reg);
1446 *before = get_insns ();
1447 end_sequence ();
1449 *loc = new_reg;
1450 if (after != NULL)
1452 start_sequence ();
1453 lra_emit_move (before_p ? copy_rtx (reg) : reg, new_reg);
1454 emit_insn (*after);
1455 *after = get_insns ();
1456 end_sequence ();
1458 return true;
1461 /* Insert move insn in simplify_operand_subreg. BEFORE returns
1462 the insn to be inserted before curr insn. AFTER returns the
1463 the insn to be inserted after curr insn. ORIGREG and NEWREG
1464 are the original reg and new reg for reload. */
1465 static void
1466 insert_move_for_subreg (rtx_insn **before, rtx_insn **after, rtx origreg,
1467 rtx newreg)
1469 if (before)
1471 push_to_sequence (*before);
1472 lra_emit_move (newreg, origreg);
1473 *before = get_insns ();
1474 end_sequence ();
1476 if (after)
1478 start_sequence ();
1479 lra_emit_move (origreg, newreg);
1480 emit_insn (*after);
1481 *after = get_insns ();
1482 end_sequence ();
1486 static int valid_address_p (machine_mode mode, rtx addr, addr_space_t as);
1487 static bool process_address (int, bool, rtx_insn **, rtx_insn **);
1489 /* Make reloads for subreg in operand NOP with internal subreg mode
1490 REG_MODE, add new reloads for further processing. Return true if
1491 any change was done. */
1492 static bool
1493 simplify_operand_subreg (int nop, machine_mode reg_mode)
1495 int hard_regno;
1496 rtx_insn *before, *after;
1497 machine_mode mode, innermode;
1498 rtx reg, new_reg;
1499 rtx operand = *curr_id->operand_loc[nop];
1500 enum reg_class regclass;
1501 enum op_type type;
1503 before = after = NULL;
1505 if (GET_CODE (operand) != SUBREG)
1506 return false;
1508 mode = GET_MODE (operand);
1509 reg = SUBREG_REG (operand);
1510 innermode = GET_MODE (reg);
1511 type = curr_static_id->operand[nop].type;
1512 if (MEM_P (reg))
1514 const bool addr_was_valid
1515 = valid_address_p (innermode, XEXP (reg, 0), MEM_ADDR_SPACE (reg));
1516 alter_subreg (curr_id->operand_loc[nop], false);
1517 rtx subst = *curr_id->operand_loc[nop];
1518 lra_assert (MEM_P (subst));
1520 if (!addr_was_valid
1521 || valid_address_p (GET_MODE (subst), XEXP (subst, 0),
1522 MEM_ADDR_SPACE (subst))
1523 || ((get_constraint_type (lookup_constraint
1524 (curr_static_id->operand[nop].constraint))
1525 != CT_SPECIAL_MEMORY)
1526 /* We still can reload address and if the address is
1527 valid, we can remove subreg without reloading its
1528 inner memory. */
1529 && valid_address_p (GET_MODE (subst),
1530 regno_reg_rtx
1531 [ira_class_hard_regs
1532 [base_reg_class (GET_MODE (subst),
1533 MEM_ADDR_SPACE (subst),
1534 ADDRESS, SCRATCH)][0]],
1535 MEM_ADDR_SPACE (subst))))
1537 /* If we change the address for a paradoxical subreg of memory, the
1538 new address might violate the necessary alignment or the access
1539 might be slow; take this into consideration. We need not worry
1540 about accesses beyond allocated memory for paradoxical memory
1541 subregs as we don't substitute such equiv memory (see processing
1542 equivalences in function lra_constraints) and because for spilled
1543 pseudos we allocate stack memory enough for the biggest
1544 corresponding paradoxical subreg.
1546 However, do not blindly simplify a (subreg (mem ...)) for
1547 WORD_REGISTER_OPERATIONS targets as this may lead to loading junk
1548 data into a register when the inner is narrower than outer or
1549 missing important data from memory when the inner is wider than
1550 outer. This rule only applies to modes that are no wider than
1551 a word. */
1552 if (!(GET_MODE_PRECISION (mode) != GET_MODE_PRECISION (innermode)
1553 && GET_MODE_SIZE (mode) <= UNITS_PER_WORD
1554 && GET_MODE_SIZE (innermode) <= UNITS_PER_WORD
1555 && WORD_REGISTER_OPERATIONS)
1556 && (!(MEM_ALIGN (subst) < GET_MODE_ALIGNMENT (mode)
1557 && SLOW_UNALIGNED_ACCESS (mode, MEM_ALIGN (subst)))
1558 || (MEM_ALIGN (reg) < GET_MODE_ALIGNMENT (innermode)
1559 && SLOW_UNALIGNED_ACCESS (innermode, MEM_ALIGN (reg)))))
1560 return true;
1562 *curr_id->operand_loc[nop] = operand;
1564 /* But if the address was not valid, we cannot reload the MEM without
1565 reloading the address first. */
1566 if (!addr_was_valid)
1567 process_address (nop, false, &before, &after);
1569 /* INNERMODE is fast, MODE slow. Reload the mem in INNERMODE. */
1570 enum reg_class rclass
1571 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1572 if (get_reload_reg (curr_static_id->operand[nop].type, innermode,
1573 reg, rclass, TRUE, "slow mem", &new_reg))
1575 bool insert_before, insert_after;
1576 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1578 insert_before = (type != OP_OUT
1579 || GET_MODE_SIZE (innermode)
1580 > GET_MODE_SIZE (mode));
1581 insert_after = type != OP_IN;
1582 insert_move_for_subreg (insert_before ? &before : NULL,
1583 insert_after ? &after : NULL,
1584 reg, new_reg);
1586 SUBREG_REG (operand) = new_reg;
1588 /* Convert to MODE. */
1589 reg = operand;
1590 rclass
1591 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1592 if (get_reload_reg (curr_static_id->operand[nop].type, mode, reg,
1593 rclass, TRUE, "slow mem", &new_reg))
1595 bool insert_before, insert_after;
1596 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1598 insert_before = type != OP_OUT;
1599 insert_after = type != OP_IN;
1600 insert_move_for_subreg (insert_before ? &before : NULL,
1601 insert_after ? &after : NULL,
1602 reg, new_reg);
1604 *curr_id->operand_loc[nop] = new_reg;
1605 lra_process_new_insns (curr_insn, before, after,
1606 "Inserting slow mem reload");
1607 return true;
1610 /* If the address was valid and became invalid, prefer to reload
1611 the memory. Typical case is when the index scale should
1612 correspond the memory. */
1613 *curr_id->operand_loc[nop] = operand;
1614 /* Do not return false here as the MEM_P (reg) will be processed
1615 later in this function. */
1617 else if (REG_P (reg) && REGNO (reg) < FIRST_PSEUDO_REGISTER)
1619 alter_subreg (curr_id->operand_loc[nop], false);
1620 return true;
1622 else if (CONSTANT_P (reg))
1624 /* Try to simplify subreg of constant. It is usually result of
1625 equivalence substitution. */
1626 if (innermode == VOIDmode
1627 && (innermode = original_subreg_reg_mode[nop]) == VOIDmode)
1628 innermode = curr_static_id->operand[nop].mode;
1629 if ((new_reg = simplify_subreg (mode, reg, innermode,
1630 SUBREG_BYTE (operand))) != NULL_RTX)
1632 *curr_id->operand_loc[nop] = new_reg;
1633 return true;
1636 /* Put constant into memory when we have mixed modes. It generates
1637 a better code in most cases as it does not need a secondary
1638 reload memory. It also prevents LRA looping when LRA is using
1639 secondary reload memory again and again. */
1640 if (CONSTANT_P (reg) && CONST_POOL_OK_P (reg_mode, reg)
1641 && SCALAR_INT_MODE_P (reg_mode) != SCALAR_INT_MODE_P (mode))
1643 SUBREG_REG (operand) = force_const_mem (reg_mode, reg);
1644 alter_subreg (curr_id->operand_loc[nop], false);
1645 return true;
1647 /* Force a reload of the SUBREG_REG if this is a constant or PLUS or
1648 if there may be a problem accessing OPERAND in the outer
1649 mode. */
1650 if ((REG_P (reg)
1651 && REGNO (reg) >= FIRST_PSEUDO_REGISTER
1652 && (hard_regno = lra_get_regno_hard_regno (REGNO (reg))) >= 0
1653 /* Don't reload paradoxical subregs because we could be looping
1654 having repeatedly final regno out of hard regs range. */
1655 && (hard_regno_nregs[hard_regno][innermode]
1656 >= hard_regno_nregs[hard_regno][mode])
1657 && simplify_subreg_regno (hard_regno, innermode,
1658 SUBREG_BYTE (operand), mode) < 0
1659 /* Don't reload subreg for matching reload. It is actually
1660 valid subreg in LRA. */
1661 && ! LRA_SUBREG_P (operand))
1662 || CONSTANT_P (reg) || GET_CODE (reg) == PLUS || MEM_P (reg))
1664 enum reg_class rclass;
1666 if (REG_P (reg))
1667 /* There is a big probability that we will get the same class
1668 for the new pseudo and we will get the same insn which
1669 means infinite looping. So spill the new pseudo. */
1670 rclass = NO_REGS;
1671 else
1672 /* The class will be defined later in curr_insn_transform. */
1673 rclass
1674 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1676 if (get_reload_reg (curr_static_id->operand[nop].type, reg_mode, reg,
1677 rclass, TRUE, "subreg reg", &new_reg))
1679 bool insert_before, insert_after;
1680 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1682 insert_before = (type != OP_OUT
1683 || GET_MODE_SIZE (innermode) > GET_MODE_SIZE (mode));
1684 insert_after = (type != OP_IN);
1685 insert_move_for_subreg (insert_before ? &before : NULL,
1686 insert_after ? &after : NULL,
1687 reg, new_reg);
1689 SUBREG_REG (operand) = new_reg;
1690 lra_process_new_insns (curr_insn, before, after,
1691 "Inserting subreg reload");
1692 return true;
1694 /* Force a reload for a paradoxical subreg. For paradoxical subreg,
1695 IRA allocates hardreg to the inner pseudo reg according to its mode
1696 instead of the outermode, so the size of the hardreg may not be enough
1697 to contain the outermode operand, in that case we may need to insert
1698 reload for the reg. For the following two types of paradoxical subreg,
1699 we need to insert reload:
1700 1. If the op_type is OP_IN, and the hardreg could not be paired with
1701 other hardreg to contain the outermode operand
1702 (checked by in_hard_reg_set_p), we need to insert the reload.
1703 2. If the op_type is OP_OUT or OP_INOUT.
1705 Here is a paradoxical subreg example showing how the reload is generated:
1707 (insn 5 4 7 2 (set (reg:TI 106 [ __comp ])
1708 (subreg:TI (reg:DI 107 [ __comp ]) 0)) {*movti_internal_rex64}
1710 In IRA, reg107 is allocated to a DImode hardreg. We use x86-64 as example
1711 here, if reg107 is assigned to hardreg R15, because R15 is the last
1712 hardreg, compiler cannot find another hardreg to pair with R15 to
1713 contain TImode data. So we insert a TImode reload reg180 for it.
1714 After reload is inserted:
1716 (insn 283 0 0 (set (subreg:DI (reg:TI 180 [orig:107 __comp ] [107]) 0)
1717 (reg:DI 107 [ __comp ])) -1
1718 (insn 5 4 7 2 (set (reg:TI 106 [ __comp ])
1719 (subreg:TI (reg:TI 180 [orig:107 __comp ] [107]) 0)) {*movti_internal_rex64}
1721 Two reload hard registers will be allocated to reg180 to save TImode data
1722 in LRA_assign. */
1723 else if (REG_P (reg)
1724 && REGNO (reg) >= FIRST_PSEUDO_REGISTER
1725 && (hard_regno = lra_get_regno_hard_regno (REGNO (reg))) >= 0
1726 && (hard_regno_nregs[hard_regno][innermode]
1727 < hard_regno_nregs[hard_regno][mode])
1728 && (regclass = lra_get_allocno_class (REGNO (reg)))
1729 && (type != OP_IN
1730 || !in_hard_reg_set_p (reg_class_contents[regclass],
1731 mode, hard_regno)))
1733 /* The class will be defined later in curr_insn_transform. */
1734 enum reg_class rclass
1735 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1737 if (get_reload_reg (curr_static_id->operand[nop].type, mode, reg,
1738 rclass, TRUE, "paradoxical subreg", &new_reg))
1740 rtx subreg;
1741 bool insert_before, insert_after;
1743 PUT_MODE (new_reg, mode);
1744 subreg = gen_lowpart_SUBREG (innermode, new_reg);
1745 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1747 insert_before = (type != OP_OUT);
1748 insert_after = (type != OP_IN);
1749 insert_move_for_subreg (insert_before ? &before : NULL,
1750 insert_after ? &after : NULL,
1751 reg, subreg);
1753 SUBREG_REG (operand) = new_reg;
1754 lra_process_new_insns (curr_insn, before, after,
1755 "Inserting paradoxical subreg reload");
1756 return true;
1758 return false;
1761 /* Return TRUE if X refers for a hard register from SET. */
1762 static bool
1763 uses_hard_regs_p (rtx x, HARD_REG_SET set)
1765 int i, j, x_hard_regno;
1766 machine_mode mode;
1767 const char *fmt;
1768 enum rtx_code code;
1770 if (x == NULL_RTX)
1771 return false;
1772 code = GET_CODE (x);
1773 mode = GET_MODE (x);
1774 if (code == SUBREG)
1776 x = SUBREG_REG (x);
1777 code = GET_CODE (x);
1778 if (GET_MODE_SIZE (GET_MODE (x)) > GET_MODE_SIZE (mode))
1779 mode = GET_MODE (x);
1782 if (REG_P (x))
1784 x_hard_regno = get_hard_regno (x, true);
1785 return (x_hard_regno >= 0
1786 && overlaps_hard_reg_set_p (set, mode, x_hard_regno));
1788 if (MEM_P (x))
1790 struct address_info ad;
1792 decompose_mem_address (&ad, x);
1793 if (ad.base_term != NULL && uses_hard_regs_p (*ad.base_term, set))
1794 return true;
1795 if (ad.index_term != NULL && uses_hard_regs_p (*ad.index_term, set))
1796 return true;
1798 fmt = GET_RTX_FORMAT (code);
1799 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1801 if (fmt[i] == 'e')
1803 if (uses_hard_regs_p (XEXP (x, i), set))
1804 return true;
1806 else if (fmt[i] == 'E')
1808 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
1809 if (uses_hard_regs_p (XVECEXP (x, i, j), set))
1810 return true;
1813 return false;
1816 /* Return true if OP is a spilled pseudo. */
1817 static inline bool
1818 spilled_pseudo_p (rtx op)
1820 return (REG_P (op)
1821 && REGNO (op) >= FIRST_PSEUDO_REGISTER && in_mem_p (REGNO (op)));
1824 /* Return true if X is a general constant. */
1825 static inline bool
1826 general_constant_p (rtx x)
1828 return CONSTANT_P (x) && (! flag_pic || LEGITIMATE_PIC_OPERAND_P (x));
1831 static bool
1832 reg_in_class_p (rtx reg, enum reg_class cl)
1834 if (cl == NO_REGS)
1835 return get_reg_class (REGNO (reg)) == NO_REGS;
1836 return in_class_p (reg, cl, NULL);
1839 /* Return true if SET of RCLASS contains no hard regs which can be
1840 used in MODE. */
1841 static bool
1842 prohibited_class_reg_set_mode_p (enum reg_class rclass,
1843 HARD_REG_SET &set,
1844 enum machine_mode mode)
1846 HARD_REG_SET temp;
1848 lra_assert (hard_reg_set_subset_p (reg_class_contents[rclass], set));
1849 COPY_HARD_REG_SET (temp, set);
1850 AND_COMPL_HARD_REG_SET (temp, lra_no_alloc_regs);
1851 return (hard_reg_set_subset_p
1852 (temp, ira_prohibited_class_mode_regs[rclass][mode]));
1856 /* Used to check validity info about small class input operands. It
1857 should be incremented at start of processing an insn
1858 alternative. */
1859 static unsigned int curr_small_class_check = 0;
1861 /* Update number of used inputs of class OP_CLASS for operand NOP.
1862 Return true if we have more such class operands than the number of
1863 available regs. */
1864 static bool
1865 update_and_check_small_class_inputs (int nop, enum reg_class op_class)
1867 static unsigned int small_class_check[LIM_REG_CLASSES];
1868 static int small_class_input_nums[LIM_REG_CLASSES];
1870 if (SMALL_REGISTER_CLASS_P (op_class)
1871 /* We are interesting in classes became small because of fixing
1872 some hard regs, e.g. by an user through GCC options. */
1873 && hard_reg_set_intersect_p (reg_class_contents[op_class],
1874 ira_no_alloc_regs)
1875 && (curr_static_id->operand[nop].type != OP_OUT
1876 || curr_static_id->operand[nop].early_clobber))
1878 if (small_class_check[op_class] == curr_small_class_check)
1879 small_class_input_nums[op_class]++;
1880 else
1882 small_class_check[op_class] = curr_small_class_check;
1883 small_class_input_nums[op_class] = 1;
1885 if (small_class_input_nums[op_class] > ira_class_hard_regs_num[op_class])
1886 return true;
1888 return false;
1891 /* Major function to choose the current insn alternative and what
1892 operands should be reloaded and how. If ONLY_ALTERNATIVE is not
1893 negative we should consider only this alternative. Return false if
1894 we can not choose the alternative or find how to reload the
1895 operands. */
1896 static bool
1897 process_alt_operands (int only_alternative)
1899 bool ok_p = false;
1900 int nop, overall, nalt;
1901 int n_alternatives = curr_static_id->n_alternatives;
1902 int n_operands = curr_static_id->n_operands;
1903 /* LOSERS counts the operands that don't fit this alternative and
1904 would require loading. */
1905 int losers;
1906 int addr_losers;
1907 /* REJECT is a count of how undesirable this alternative says it is
1908 if any reloading is required. If the alternative matches exactly
1909 then REJECT is ignored, but otherwise it gets this much counted
1910 against it in addition to the reloading needed. */
1911 int reject;
1912 /* This is defined by '!' or '?' alternative constraint and added to
1913 reject. But in some cases it can be ignored. */
1914 int static_reject;
1915 int op_reject;
1916 /* The number of elements in the following array. */
1917 int early_clobbered_regs_num;
1918 /* Numbers of operands which are early clobber registers. */
1919 int early_clobbered_nops[MAX_RECOG_OPERANDS];
1920 enum reg_class curr_alt[MAX_RECOG_OPERANDS];
1921 HARD_REG_SET curr_alt_set[MAX_RECOG_OPERANDS];
1922 bool curr_alt_match_win[MAX_RECOG_OPERANDS];
1923 bool curr_alt_win[MAX_RECOG_OPERANDS];
1924 bool curr_alt_offmemok[MAX_RECOG_OPERANDS];
1925 int curr_alt_matches[MAX_RECOG_OPERANDS];
1926 /* The number of elements in the following array. */
1927 int curr_alt_dont_inherit_ops_num;
1928 /* Numbers of operands whose reload pseudos should not be inherited. */
1929 int curr_alt_dont_inherit_ops[MAX_RECOG_OPERANDS];
1930 rtx op;
1931 /* The register when the operand is a subreg of register, otherwise the
1932 operand itself. */
1933 rtx no_subreg_reg_operand[MAX_RECOG_OPERANDS];
1934 /* The register if the operand is a register or subreg of register,
1935 otherwise NULL. */
1936 rtx operand_reg[MAX_RECOG_OPERANDS];
1937 int hard_regno[MAX_RECOG_OPERANDS];
1938 machine_mode biggest_mode[MAX_RECOG_OPERANDS];
1939 int reload_nregs, reload_sum;
1940 bool costly_p;
1941 enum reg_class cl;
1943 /* Calculate some data common for all alternatives to speed up the
1944 function. */
1945 for (nop = 0; nop < n_operands; nop++)
1947 rtx reg;
1949 op = no_subreg_reg_operand[nop] = *curr_id->operand_loc[nop];
1950 /* The real hard regno of the operand after the allocation. */
1951 hard_regno[nop] = get_hard_regno (op, true);
1953 operand_reg[nop] = reg = op;
1954 biggest_mode[nop] = GET_MODE (op);
1955 if (GET_CODE (op) == SUBREG)
1957 operand_reg[nop] = reg = SUBREG_REG (op);
1958 if (GET_MODE_SIZE (biggest_mode[nop])
1959 < GET_MODE_SIZE (GET_MODE (reg)))
1960 biggest_mode[nop] = GET_MODE (reg);
1962 if (! REG_P (reg))
1963 operand_reg[nop] = NULL_RTX;
1964 else if (REGNO (reg) >= FIRST_PSEUDO_REGISTER
1965 || ((int) REGNO (reg)
1966 == lra_get_elimination_hard_regno (REGNO (reg))))
1967 no_subreg_reg_operand[nop] = reg;
1968 else
1969 operand_reg[nop] = no_subreg_reg_operand[nop]
1970 /* Just use natural mode for elimination result. It should
1971 be enough for extra constraints hooks. */
1972 = regno_reg_rtx[hard_regno[nop]];
1975 /* The constraints are made of several alternatives. Each operand's
1976 constraint looks like foo,bar,... with commas separating the
1977 alternatives. The first alternatives for all operands go
1978 together, the second alternatives go together, etc.
1980 First loop over alternatives. */
1981 alternative_mask preferred = curr_id->preferred_alternatives;
1982 if (only_alternative >= 0)
1983 preferred &= ALTERNATIVE_BIT (only_alternative);
1985 for (nalt = 0; nalt < n_alternatives; nalt++)
1987 /* Loop over operands for one constraint alternative. */
1988 if (!TEST_BIT (preferred, nalt))
1989 continue;
1991 curr_small_class_check++;
1992 overall = losers = addr_losers = 0;
1993 static_reject = reject = reload_nregs = reload_sum = 0;
1994 for (nop = 0; nop < n_operands; nop++)
1996 int inc = (curr_static_id
1997 ->operand_alternative[nalt * n_operands + nop].reject);
1998 if (lra_dump_file != NULL && inc != 0)
1999 fprintf (lra_dump_file,
2000 " Staticly defined alt reject+=%d\n", inc);
2001 static_reject += inc;
2003 reject += static_reject;
2004 early_clobbered_regs_num = 0;
2006 for (nop = 0; nop < n_operands; nop++)
2008 const char *p;
2009 char *end;
2010 int len, c, m, i, opalt_num, this_alternative_matches;
2011 bool win, did_match, offmemok, early_clobber_p;
2012 /* false => this operand can be reloaded somehow for this
2013 alternative. */
2014 bool badop;
2015 /* true => this operand can be reloaded if the alternative
2016 allows regs. */
2017 bool winreg;
2018 /* True if a constant forced into memory would be OK for
2019 this operand. */
2020 bool constmemok;
2021 enum reg_class this_alternative, this_costly_alternative;
2022 HARD_REG_SET this_alternative_set, this_costly_alternative_set;
2023 bool this_alternative_match_win, this_alternative_win;
2024 bool this_alternative_offmemok;
2025 bool scratch_p;
2026 machine_mode mode;
2027 enum constraint_num cn;
2029 opalt_num = nalt * n_operands + nop;
2030 if (curr_static_id->operand_alternative[opalt_num].anything_ok)
2032 /* Fast track for no constraints at all. */
2033 curr_alt[nop] = NO_REGS;
2034 CLEAR_HARD_REG_SET (curr_alt_set[nop]);
2035 curr_alt_win[nop] = true;
2036 curr_alt_match_win[nop] = false;
2037 curr_alt_offmemok[nop] = false;
2038 curr_alt_matches[nop] = -1;
2039 continue;
2042 op = no_subreg_reg_operand[nop];
2043 mode = curr_operand_mode[nop];
2045 win = did_match = winreg = offmemok = constmemok = false;
2046 badop = true;
2048 early_clobber_p = false;
2049 p = curr_static_id->operand_alternative[opalt_num].constraint;
2051 this_costly_alternative = this_alternative = NO_REGS;
2052 /* We update set of possible hard regs besides its class
2053 because reg class might be inaccurate. For example,
2054 union of LO_REGS (l), HI_REGS(h), and STACK_REG(k) in ARM
2055 is translated in HI_REGS because classes are merged by
2056 pairs and there is no accurate intermediate class. */
2057 CLEAR_HARD_REG_SET (this_alternative_set);
2058 CLEAR_HARD_REG_SET (this_costly_alternative_set);
2059 this_alternative_win = false;
2060 this_alternative_match_win = false;
2061 this_alternative_offmemok = false;
2062 this_alternative_matches = -1;
2064 /* An empty constraint should be excluded by the fast
2065 track. */
2066 lra_assert (*p != 0 && *p != ',');
2068 op_reject = 0;
2069 /* Scan this alternative's specs for this operand; set WIN
2070 if the operand fits any letter in this alternative.
2071 Otherwise, clear BADOP if this operand could fit some
2072 letter after reloads, or set WINREG if this operand could
2073 fit after reloads provided the constraint allows some
2074 registers. */
2075 costly_p = false;
2078 switch ((c = *p, len = CONSTRAINT_LEN (c, p)), c)
2080 case '\0':
2081 len = 0;
2082 break;
2083 case ',':
2084 c = '\0';
2085 break;
2087 case '&':
2088 early_clobber_p = true;
2089 break;
2091 case '$':
2092 op_reject += LRA_MAX_REJECT;
2093 break;
2094 case '^':
2095 op_reject += LRA_LOSER_COST_FACTOR;
2096 break;
2098 case '#':
2099 /* Ignore rest of this alternative. */
2100 c = '\0';
2101 break;
2103 case '0': case '1': case '2': case '3': case '4':
2104 case '5': case '6': case '7': case '8': case '9':
2106 int m_hregno;
2107 bool match_p;
2109 m = strtoul (p, &end, 10);
2110 p = end;
2111 len = 0;
2112 lra_assert (nop > m);
2114 this_alternative_matches = m;
2115 m_hregno = get_hard_regno (*curr_id->operand_loc[m], false);
2116 /* We are supposed to match a previous operand.
2117 If we do, we win if that one did. If we do
2118 not, count both of the operands as losers.
2119 (This is too conservative, since most of the
2120 time only a single reload insn will be needed
2121 to make the two operands win. As a result,
2122 this alternative may be rejected when it is
2123 actually desirable.) */
2124 match_p = false;
2125 if (operands_match_p (*curr_id->operand_loc[nop],
2126 *curr_id->operand_loc[m], m_hregno))
2128 /* We should reject matching of an early
2129 clobber operand if the matching operand is
2130 not dying in the insn. */
2131 if (! curr_static_id->operand[m].early_clobber
2132 || operand_reg[nop] == NULL_RTX
2133 || (find_regno_note (curr_insn, REG_DEAD,
2134 REGNO (op))
2135 || REGNO (op) == REGNO (operand_reg[m])))
2136 match_p = true;
2138 if (match_p)
2140 /* If we are matching a non-offsettable
2141 address where an offsettable address was
2142 expected, then we must reject this
2143 combination, because we can't reload
2144 it. */
2145 if (curr_alt_offmemok[m]
2146 && MEM_P (*curr_id->operand_loc[m])
2147 && curr_alt[m] == NO_REGS && ! curr_alt_win[m])
2148 continue;
2150 else
2152 /* Operands don't match. Both operands must
2153 allow a reload register, otherwise we
2154 cannot make them match. */
2155 if (curr_alt[m] == NO_REGS)
2156 break;
2157 /* Retroactively mark the operand we had to
2158 match as a loser, if it wasn't already and
2159 it wasn't matched to a register constraint
2160 (e.g it might be matched by memory). */
2161 if (curr_alt_win[m]
2162 && (operand_reg[m] == NULL_RTX
2163 || hard_regno[m] < 0))
2165 losers++;
2166 reload_nregs
2167 += (ira_reg_class_max_nregs[curr_alt[m]]
2168 [GET_MODE (*curr_id->operand_loc[m])]);
2171 /* Prefer matching earlyclobber alternative as
2172 it results in less hard regs required for
2173 the insn than a non-matching earlyclobber
2174 alternative. */
2175 if (curr_static_id->operand[m].early_clobber)
2177 if (lra_dump_file != NULL)
2178 fprintf
2179 (lra_dump_file,
2180 " %d Matching earlyclobber alt:"
2181 " reject--\n",
2182 nop);
2183 reject--;
2185 /* Otherwise we prefer no matching
2186 alternatives because it gives more freedom
2187 in RA. */
2188 else if (operand_reg[nop] == NULL_RTX
2189 || (find_regno_note (curr_insn, REG_DEAD,
2190 REGNO (operand_reg[nop]))
2191 == NULL_RTX))
2193 if (lra_dump_file != NULL)
2194 fprintf
2195 (lra_dump_file,
2196 " %d Matching alt: reject+=2\n",
2197 nop);
2198 reject += 2;
2201 /* If we have to reload this operand and some
2202 previous operand also had to match the same
2203 thing as this operand, we don't know how to do
2204 that. */
2205 if (!match_p || !curr_alt_win[m])
2207 for (i = 0; i < nop; i++)
2208 if (curr_alt_matches[i] == m)
2209 break;
2210 if (i < nop)
2211 break;
2213 else
2214 did_match = true;
2216 /* This can be fixed with reloads if the operand
2217 we are supposed to match can be fixed with
2218 reloads. */
2219 badop = false;
2220 this_alternative = curr_alt[m];
2221 COPY_HARD_REG_SET (this_alternative_set, curr_alt_set[m]);
2222 winreg = this_alternative != NO_REGS;
2223 break;
2226 case 'g':
2227 if (MEM_P (op)
2228 || general_constant_p (op)
2229 || spilled_pseudo_p (op))
2230 win = true;
2231 cl = GENERAL_REGS;
2232 goto reg;
2234 default:
2235 cn = lookup_constraint (p);
2236 switch (get_constraint_type (cn))
2238 case CT_REGISTER:
2239 cl = reg_class_for_constraint (cn);
2240 if (cl != NO_REGS)
2241 goto reg;
2242 break;
2244 case CT_CONST_INT:
2245 if (CONST_INT_P (op)
2246 && insn_const_int_ok_for_constraint (INTVAL (op), cn))
2247 win = true;
2248 break;
2250 case CT_MEMORY:
2251 if (MEM_P (op)
2252 && satisfies_memory_constraint_p (op, cn))
2253 win = true;
2254 else if (spilled_pseudo_p (op))
2255 win = true;
2257 /* If we didn't already win, we can reload constants
2258 via force_const_mem or put the pseudo value into
2259 memory, or make other memory by reloading the
2260 address like for 'o'. */
2261 if (CONST_POOL_OK_P (mode, op)
2262 || MEM_P (op) || REG_P (op)
2263 /* We can restore the equiv insn by a
2264 reload. */
2265 || equiv_substition_p[nop])
2266 badop = false;
2267 constmemok = true;
2268 offmemok = true;
2269 break;
2271 case CT_ADDRESS:
2272 /* If we didn't already win, we can reload the address
2273 into a base register. */
2274 if (satisfies_address_constraint_p (op, cn))
2275 win = true;
2276 cl = base_reg_class (VOIDmode, ADDR_SPACE_GENERIC,
2277 ADDRESS, SCRATCH);
2278 badop = false;
2279 goto reg;
2281 case CT_FIXED_FORM:
2282 if (constraint_satisfied_p (op, cn))
2283 win = true;
2284 break;
2286 case CT_SPECIAL_MEMORY:
2287 if (MEM_P (op)
2288 && satisfies_memory_constraint_p (op, cn))
2289 win = true;
2290 else if (spilled_pseudo_p (op))
2291 win = true;
2292 break;
2294 break;
2296 reg:
2297 this_alternative = reg_class_subunion[this_alternative][cl];
2298 IOR_HARD_REG_SET (this_alternative_set,
2299 reg_class_contents[cl]);
2300 if (costly_p)
2302 this_costly_alternative
2303 = reg_class_subunion[this_costly_alternative][cl];
2304 IOR_HARD_REG_SET (this_costly_alternative_set,
2305 reg_class_contents[cl]);
2307 if (mode == BLKmode)
2308 break;
2309 winreg = true;
2310 if (REG_P (op))
2312 if (hard_regno[nop] >= 0
2313 && in_hard_reg_set_p (this_alternative_set,
2314 mode, hard_regno[nop]))
2315 win = true;
2316 else if (hard_regno[nop] < 0
2317 && in_class_p (op, this_alternative, NULL))
2318 win = true;
2320 break;
2322 if (c != ' ' && c != '\t')
2323 costly_p = c == '*';
2325 while ((p += len), c);
2327 scratch_p = (operand_reg[nop] != NULL_RTX
2328 && lra_former_scratch_p (REGNO (operand_reg[nop])));
2329 /* Record which operands fit this alternative. */
2330 if (win)
2332 this_alternative_win = true;
2333 if (operand_reg[nop] != NULL_RTX)
2335 if (hard_regno[nop] >= 0)
2337 if (in_hard_reg_set_p (this_costly_alternative_set,
2338 mode, hard_regno[nop]))
2340 if (lra_dump_file != NULL)
2341 fprintf (lra_dump_file,
2342 " %d Costly set: reject++\n",
2343 nop);
2344 reject++;
2347 else
2349 /* Prefer won reg to spilled pseudo under other
2350 equal conditions for possibe inheritance. */
2351 if (! scratch_p)
2353 if (lra_dump_file != NULL)
2354 fprintf
2355 (lra_dump_file,
2356 " %d Non pseudo reload: reject++\n",
2357 nop);
2358 reject++;
2360 if (in_class_p (operand_reg[nop],
2361 this_costly_alternative, NULL))
2363 if (lra_dump_file != NULL)
2364 fprintf
2365 (lra_dump_file,
2366 " %d Non pseudo costly reload:"
2367 " reject++\n",
2368 nop);
2369 reject++;
2372 /* We simulate the behavior of old reload here.
2373 Although scratches need hard registers and it
2374 might result in spilling other pseudos, no reload
2375 insns are generated for the scratches. So it
2376 might cost something but probably less than old
2377 reload pass believes. */
2378 if (scratch_p)
2380 if (lra_dump_file != NULL)
2381 fprintf (lra_dump_file,
2382 " %d Scratch win: reject+=2\n",
2383 nop);
2384 reject += 2;
2388 else if (did_match)
2389 this_alternative_match_win = true;
2390 else
2392 int const_to_mem = 0;
2393 bool no_regs_p;
2395 reject += op_reject;
2396 /* Never do output reload of stack pointer. It makes
2397 impossible to do elimination when SP is changed in
2398 RTL. */
2399 if (op == stack_pointer_rtx && ! frame_pointer_needed
2400 && curr_static_id->operand[nop].type != OP_IN)
2401 goto fail;
2403 /* If this alternative asks for a specific reg class, see if there
2404 is at least one allocatable register in that class. */
2405 no_regs_p
2406 = (this_alternative == NO_REGS
2407 || (hard_reg_set_subset_p
2408 (reg_class_contents[this_alternative],
2409 lra_no_alloc_regs)));
2411 /* For asms, verify that the class for this alternative is possible
2412 for the mode that is specified. */
2413 if (!no_regs_p && INSN_CODE (curr_insn) < 0)
2415 int i;
2416 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
2417 if (HARD_REGNO_MODE_OK (i, mode)
2418 && in_hard_reg_set_p (reg_class_contents[this_alternative],
2419 mode, i))
2420 break;
2421 if (i == FIRST_PSEUDO_REGISTER)
2422 winreg = false;
2425 /* If this operand accepts a register, and if the
2426 register class has at least one allocatable register,
2427 then this operand can be reloaded. */
2428 if (winreg && !no_regs_p)
2429 badop = false;
2431 if (badop)
2433 if (lra_dump_file != NULL)
2434 fprintf (lra_dump_file,
2435 " alt=%d: Bad operand -- refuse\n",
2436 nalt);
2437 goto fail;
2440 if (this_alternative != NO_REGS)
2442 HARD_REG_SET available_regs;
2444 COPY_HARD_REG_SET (available_regs,
2445 reg_class_contents[this_alternative]);
2446 AND_COMPL_HARD_REG_SET
2447 (available_regs,
2448 ira_prohibited_class_mode_regs[this_alternative][mode]);
2449 AND_COMPL_HARD_REG_SET (available_regs, lra_no_alloc_regs);
2450 if (hard_reg_set_empty_p (available_regs))
2452 /* There are no hard regs holding a value of given
2453 mode. */
2454 if (offmemok)
2456 this_alternative = NO_REGS;
2457 if (lra_dump_file != NULL)
2458 fprintf (lra_dump_file,
2459 " %d Using memory because of"
2460 " a bad mode: reject+=2\n",
2461 nop);
2462 reject += 2;
2464 else
2466 if (lra_dump_file != NULL)
2467 fprintf (lra_dump_file,
2468 " alt=%d: Wrong mode -- refuse\n",
2469 nalt);
2470 goto fail;
2475 /* If not assigned pseudo has a class which a subset of
2476 required reg class, it is a less costly alternative
2477 as the pseudo still can get a hard reg of necessary
2478 class. */
2479 if (! no_regs_p && REG_P (op) && hard_regno[nop] < 0
2480 && (cl = get_reg_class (REGNO (op))) != NO_REGS
2481 && ira_class_subset_p[this_alternative][cl])
2483 if (lra_dump_file != NULL)
2484 fprintf
2485 (lra_dump_file,
2486 " %d Super set class reg: reject-=3\n", nop);
2487 reject -= 3;
2490 this_alternative_offmemok = offmemok;
2491 if (this_costly_alternative != NO_REGS)
2493 if (lra_dump_file != NULL)
2494 fprintf (lra_dump_file,
2495 " %d Costly loser: reject++\n", nop);
2496 reject++;
2498 /* If the operand is dying, has a matching constraint,
2499 and satisfies constraints of the matched operand
2500 which failed to satisfy the own constraints, most probably
2501 the reload for this operand will be gone. */
2502 if (this_alternative_matches >= 0
2503 && !curr_alt_win[this_alternative_matches]
2504 && REG_P (op)
2505 && find_regno_note (curr_insn, REG_DEAD, REGNO (op))
2506 && (hard_regno[nop] >= 0
2507 ? in_hard_reg_set_p (this_alternative_set,
2508 mode, hard_regno[nop])
2509 : in_class_p (op, this_alternative, NULL)))
2511 if (lra_dump_file != NULL)
2512 fprintf
2513 (lra_dump_file,
2514 " %d Dying matched operand reload: reject++\n",
2515 nop);
2516 reject++;
2518 else
2520 /* Strict_low_part requires to reload the register
2521 not the sub-register. In this case we should
2522 check that a final reload hard reg can hold the
2523 value mode. */
2524 if (curr_static_id->operand[nop].strict_low
2525 && REG_P (op)
2526 && hard_regno[nop] < 0
2527 && GET_CODE (*curr_id->operand_loc[nop]) == SUBREG
2528 && ira_class_hard_regs_num[this_alternative] > 0
2529 && ! HARD_REGNO_MODE_OK (ira_class_hard_regs
2530 [this_alternative][0],
2531 GET_MODE
2532 (*curr_id->operand_loc[nop])))
2534 if (lra_dump_file != NULL)
2535 fprintf
2536 (lra_dump_file,
2537 " alt=%d: Strict low subreg reload -- refuse\n",
2538 nalt);
2539 goto fail;
2541 losers++;
2543 if (operand_reg[nop] != NULL_RTX
2544 /* Output operands and matched input operands are
2545 not inherited. The following conditions do not
2546 exactly describe the previous statement but they
2547 are pretty close. */
2548 && curr_static_id->operand[nop].type != OP_OUT
2549 && (this_alternative_matches < 0
2550 || curr_static_id->operand[nop].type != OP_IN))
2552 int last_reload = (lra_reg_info[ORIGINAL_REGNO
2553 (operand_reg[nop])]
2554 .last_reload);
2556 /* The value of reload_sum has sense only if we
2557 process insns in their order. It happens only on
2558 the first constraints sub-pass when we do most of
2559 reload work. */
2560 if (lra_constraint_iter == 1 && last_reload > bb_reload_num)
2561 reload_sum += last_reload - bb_reload_num;
2563 /* If this is a constant that is reloaded into the
2564 desired class by copying it to memory first, count
2565 that as another reload. This is consistent with
2566 other code and is required to avoid choosing another
2567 alternative when the constant is moved into memory.
2568 Note that the test here is precisely the same as in
2569 the code below that calls force_const_mem. */
2570 if (CONST_POOL_OK_P (mode, op)
2571 && ((targetm.preferred_reload_class
2572 (op, this_alternative) == NO_REGS)
2573 || no_input_reloads_p))
2575 const_to_mem = 1;
2576 if (! no_regs_p)
2577 losers++;
2580 /* Alternative loses if it requires a type of reload not
2581 permitted for this insn. We can always reload
2582 objects with a REG_UNUSED note. */
2583 if ((curr_static_id->operand[nop].type != OP_IN
2584 && no_output_reloads_p
2585 && ! find_reg_note (curr_insn, REG_UNUSED, op))
2586 || (curr_static_id->operand[nop].type != OP_OUT
2587 && no_input_reloads_p && ! const_to_mem)
2588 || (this_alternative_matches >= 0
2589 && (no_input_reloads_p
2590 || (no_output_reloads_p
2591 && (curr_static_id->operand
2592 [this_alternative_matches].type != OP_IN)
2593 && ! find_reg_note (curr_insn, REG_UNUSED,
2594 no_subreg_reg_operand
2595 [this_alternative_matches])))))
2597 if (lra_dump_file != NULL)
2598 fprintf
2599 (lra_dump_file,
2600 " alt=%d: No input/otput reload -- refuse\n",
2601 nalt);
2602 goto fail;
2605 /* Alternative loses if it required class pseudo can not
2606 hold value of required mode. Such insns can be
2607 described by insn definitions with mode iterators. */
2608 if (GET_MODE (*curr_id->operand_loc[nop]) != VOIDmode
2609 && ! hard_reg_set_empty_p (this_alternative_set)
2610 /* It is common practice for constraints to use a
2611 class which does not have actually enough regs to
2612 hold the value (e.g. x86 AREG for mode requiring
2613 more one general reg). Therefore we have 2
2614 conditions to check that the reload pseudo can
2615 not hold the mode value. */
2616 && ! HARD_REGNO_MODE_OK (ira_class_hard_regs
2617 [this_alternative][0],
2618 GET_MODE (*curr_id->operand_loc[nop]))
2619 /* The above condition is not enough as the first
2620 reg in ira_class_hard_regs can be not aligned for
2621 multi-words mode values. */
2622 && (prohibited_class_reg_set_mode_p
2623 (this_alternative, this_alternative_set,
2624 GET_MODE (*curr_id->operand_loc[nop]))))
2626 if (lra_dump_file != NULL)
2627 fprintf (lra_dump_file,
2628 " alt=%d: reload pseudo for op %d "
2629 " can not hold the mode value -- refuse\n",
2630 nalt, nop);
2631 goto fail;
2634 /* Check strong discouragement of reload of non-constant
2635 into class THIS_ALTERNATIVE. */
2636 if (! CONSTANT_P (op) && ! no_regs_p
2637 && (targetm.preferred_reload_class
2638 (op, this_alternative) == NO_REGS
2639 || (curr_static_id->operand[nop].type == OP_OUT
2640 && (targetm.preferred_output_reload_class
2641 (op, this_alternative) == NO_REGS))))
2643 if (lra_dump_file != NULL)
2644 fprintf (lra_dump_file,
2645 " %d Non-prefered reload: reject+=%d\n",
2646 nop, LRA_MAX_REJECT);
2647 reject += LRA_MAX_REJECT;
2650 if (! (MEM_P (op) && offmemok)
2651 && ! (const_to_mem && constmemok))
2653 /* We prefer to reload pseudos over reloading other
2654 things, since such reloads may be able to be
2655 eliminated later. So bump REJECT in other cases.
2656 Don't do this in the case where we are forcing a
2657 constant into memory and it will then win since
2658 we don't want to have a different alternative
2659 match then. */
2660 if (! (REG_P (op) && REGNO (op) >= FIRST_PSEUDO_REGISTER))
2662 if (lra_dump_file != NULL)
2663 fprintf
2664 (lra_dump_file,
2665 " %d Non-pseudo reload: reject+=2\n",
2666 nop);
2667 reject += 2;
2670 if (! no_regs_p)
2671 reload_nregs
2672 += ira_reg_class_max_nregs[this_alternative][mode];
2674 if (SMALL_REGISTER_CLASS_P (this_alternative))
2676 if (lra_dump_file != NULL)
2677 fprintf
2678 (lra_dump_file,
2679 " %d Small class reload: reject+=%d\n",
2680 nop, LRA_LOSER_COST_FACTOR / 2);
2681 reject += LRA_LOSER_COST_FACTOR / 2;
2685 /* We are trying to spill pseudo into memory. It is
2686 usually more costly than moving to a hard register
2687 although it might takes the same number of
2688 reloads.
2690 Non-pseudo spill may happen also. Suppose a target allows both
2691 register and memory in the operand constraint alternatives,
2692 then it's typical that an eliminable register has a substition
2693 of "base + offset" which can either be reloaded by a simple
2694 "new_reg <= base + offset" which will match the register
2695 constraint, or a similar reg addition followed by further spill
2696 to and reload from memory which will match the memory
2697 constraint, but this memory spill will be much more costly
2698 usually.
2700 Code below increases the reject for both pseudo and non-pseudo
2701 spill. */
2702 if (no_regs_p
2703 && !(MEM_P (op) && offmemok)
2704 && !(REG_P (op) && hard_regno[nop] < 0))
2706 if (lra_dump_file != NULL)
2707 fprintf
2708 (lra_dump_file,
2709 " %d Spill %spseudo into memory: reject+=3\n",
2710 nop, REG_P (op) ? "" : "Non-");
2711 reject += 3;
2712 if (VECTOR_MODE_P (mode))
2714 /* Spilling vectors into memory is usually more
2715 costly as they contain big values. */
2716 if (lra_dump_file != NULL)
2717 fprintf
2718 (lra_dump_file,
2719 " %d Spill vector pseudo: reject+=2\n",
2720 nop);
2721 reject += 2;
2725 /* When we use an operand requiring memory in given
2726 alternative, the insn should write *and* read the
2727 value to/from memory it is costly in comparison with
2728 an insn alternative which does not use memory
2729 (e.g. register or immediate operand). We exclude
2730 memory operand for such case as we can satisfy the
2731 memory constraints by reloading address. */
2732 if (no_regs_p && offmemok && !MEM_P (op))
2734 if (lra_dump_file != NULL)
2735 fprintf
2736 (lra_dump_file,
2737 " Using memory insn operand %d: reject+=3\n",
2738 nop);
2739 reject += 3;
2742 #ifdef SECONDARY_MEMORY_NEEDED
2743 /* If reload requires moving value through secondary
2744 memory, it will need one more insn at least. */
2745 if (this_alternative != NO_REGS
2746 && REG_P (op) && (cl = get_reg_class (REGNO (op))) != NO_REGS
2747 && ((curr_static_id->operand[nop].type != OP_OUT
2748 && SECONDARY_MEMORY_NEEDED (cl, this_alternative,
2749 GET_MODE (op)))
2750 || (curr_static_id->operand[nop].type != OP_IN
2751 && SECONDARY_MEMORY_NEEDED (this_alternative, cl,
2752 GET_MODE (op)))))
2753 losers++;
2754 #endif
2755 /* Input reloads can be inherited more often than output
2756 reloads can be removed, so penalize output
2757 reloads. */
2758 if (!REG_P (op) || curr_static_id->operand[nop].type != OP_IN)
2760 if (lra_dump_file != NULL)
2761 fprintf
2762 (lra_dump_file,
2763 " %d Non input pseudo reload: reject++\n",
2764 nop);
2765 reject++;
2768 if (MEM_P (op) && offmemok)
2769 addr_losers++;
2770 else if (curr_static_id->operand[nop].type == OP_INOUT)
2772 if (lra_dump_file != NULL)
2773 fprintf
2774 (lra_dump_file,
2775 " %d Input/Output reload: reject+=%d\n",
2776 nop, LRA_LOSER_COST_FACTOR);
2777 reject += LRA_LOSER_COST_FACTOR;
2781 if (early_clobber_p && ! scratch_p)
2783 if (lra_dump_file != NULL)
2784 fprintf (lra_dump_file,
2785 " %d Early clobber: reject++\n", nop);
2786 reject++;
2788 /* ??? We check early clobbers after processing all operands
2789 (see loop below) and there we update the costs more.
2790 Should we update the cost (may be approximately) here
2791 because of early clobber register reloads or it is a rare
2792 or non-important thing to be worth to do it. */
2793 overall = (losers * LRA_LOSER_COST_FACTOR + reject
2794 - (addr_losers == losers ? static_reject : 0));
2795 if ((best_losers == 0 || losers != 0) && best_overall < overall)
2797 if (lra_dump_file != NULL)
2798 fprintf (lra_dump_file,
2799 " alt=%d,overall=%d,losers=%d -- refuse\n",
2800 nalt, overall, losers);
2801 goto fail;
2804 if (update_and_check_small_class_inputs (nop, this_alternative))
2806 if (lra_dump_file != NULL)
2807 fprintf (lra_dump_file,
2808 " alt=%d, not enough small class regs -- refuse\n",
2809 nalt);
2810 goto fail;
2812 curr_alt[nop] = this_alternative;
2813 COPY_HARD_REG_SET (curr_alt_set[nop], this_alternative_set);
2814 curr_alt_win[nop] = this_alternative_win;
2815 curr_alt_match_win[nop] = this_alternative_match_win;
2816 curr_alt_offmemok[nop] = this_alternative_offmemok;
2817 curr_alt_matches[nop] = this_alternative_matches;
2819 if (this_alternative_matches >= 0
2820 && !did_match && !this_alternative_win)
2821 curr_alt_win[this_alternative_matches] = false;
2823 if (early_clobber_p && operand_reg[nop] != NULL_RTX)
2824 early_clobbered_nops[early_clobbered_regs_num++] = nop;
2827 if (curr_insn_set != NULL_RTX && n_operands == 2
2828 /* Prevent processing non-move insns. */
2829 && (GET_CODE (SET_SRC (curr_insn_set)) == SUBREG
2830 || SET_SRC (curr_insn_set) == no_subreg_reg_operand[1])
2831 && ((! curr_alt_win[0] && ! curr_alt_win[1]
2832 && REG_P (no_subreg_reg_operand[0])
2833 && REG_P (no_subreg_reg_operand[1])
2834 && (reg_in_class_p (no_subreg_reg_operand[0], curr_alt[1])
2835 || reg_in_class_p (no_subreg_reg_operand[1], curr_alt[0])))
2836 || (! curr_alt_win[0] && curr_alt_win[1]
2837 && REG_P (no_subreg_reg_operand[1])
2838 /* Check that we reload memory not the memory
2839 address. */
2840 && ! (curr_alt_offmemok[0]
2841 && MEM_P (no_subreg_reg_operand[0]))
2842 && reg_in_class_p (no_subreg_reg_operand[1], curr_alt[0]))
2843 || (curr_alt_win[0] && ! curr_alt_win[1]
2844 && REG_P (no_subreg_reg_operand[0])
2845 /* Check that we reload memory not the memory
2846 address. */
2847 && ! (curr_alt_offmemok[1]
2848 && MEM_P (no_subreg_reg_operand[1]))
2849 && reg_in_class_p (no_subreg_reg_operand[0], curr_alt[1])
2850 && (! CONST_POOL_OK_P (curr_operand_mode[1],
2851 no_subreg_reg_operand[1])
2852 || (targetm.preferred_reload_class
2853 (no_subreg_reg_operand[1],
2854 (enum reg_class) curr_alt[1]) != NO_REGS))
2855 /* If it is a result of recent elimination in move
2856 insn we can transform it into an add still by
2857 using this alternative. */
2858 && GET_CODE (no_subreg_reg_operand[1]) != PLUS)))
2860 /* We have a move insn and a new reload insn will be similar
2861 to the current insn. We should avoid such situation as
2862 it results in LRA cycling. */
2863 if (lra_dump_file != NULL)
2864 fprintf (lra_dump_file,
2865 " Cycle danger: overall += LRA_MAX_REJECT\n");
2866 overall += LRA_MAX_REJECT;
2868 ok_p = true;
2869 curr_alt_dont_inherit_ops_num = 0;
2870 for (nop = 0; nop < early_clobbered_regs_num; nop++)
2872 int i, j, clobbered_hard_regno, first_conflict_j, last_conflict_j;
2873 HARD_REG_SET temp_set;
2875 i = early_clobbered_nops[nop];
2876 if ((! curr_alt_win[i] && ! curr_alt_match_win[i])
2877 || hard_regno[i] < 0)
2878 continue;
2879 lra_assert (operand_reg[i] != NULL_RTX);
2880 clobbered_hard_regno = hard_regno[i];
2881 CLEAR_HARD_REG_SET (temp_set);
2882 add_to_hard_reg_set (&temp_set, biggest_mode[i], clobbered_hard_regno);
2883 first_conflict_j = last_conflict_j = -1;
2884 for (j = 0; j < n_operands; j++)
2885 if (j == i
2886 /* We don't want process insides of match_operator and
2887 match_parallel because otherwise we would process
2888 their operands once again generating a wrong
2889 code. */
2890 || curr_static_id->operand[j].is_operator)
2891 continue;
2892 else if ((curr_alt_matches[j] == i && curr_alt_match_win[j])
2893 || (curr_alt_matches[i] == j && curr_alt_match_win[i]))
2894 continue;
2895 /* If we don't reload j-th operand, check conflicts. */
2896 else if ((curr_alt_win[j] || curr_alt_match_win[j])
2897 && uses_hard_regs_p (*curr_id->operand_loc[j], temp_set))
2899 if (first_conflict_j < 0)
2900 first_conflict_j = j;
2901 last_conflict_j = j;
2903 if (last_conflict_j < 0)
2904 continue;
2905 /* If earlyclobber operand conflicts with another
2906 non-matching operand which is actually the same register
2907 as the earlyclobber operand, it is better to reload the
2908 another operand as an operand matching the earlyclobber
2909 operand can be also the same. */
2910 if (first_conflict_j == last_conflict_j
2911 && operand_reg[last_conflict_j] != NULL_RTX
2912 && ! curr_alt_match_win[last_conflict_j]
2913 && REGNO (operand_reg[i]) == REGNO (operand_reg[last_conflict_j]))
2915 curr_alt_win[last_conflict_j] = false;
2916 curr_alt_dont_inherit_ops[curr_alt_dont_inherit_ops_num++]
2917 = last_conflict_j;
2918 losers++;
2919 /* Early clobber was already reflected in REJECT. */
2920 lra_assert (reject > 0);
2921 if (lra_dump_file != NULL)
2922 fprintf
2923 (lra_dump_file,
2924 " %d Conflict early clobber reload: reject--\n",
2926 reject--;
2927 overall += LRA_LOSER_COST_FACTOR - 1;
2929 else
2931 /* We need to reload early clobbered register and the
2932 matched registers. */
2933 for (j = 0; j < n_operands; j++)
2934 if (curr_alt_matches[j] == i)
2936 curr_alt_match_win[j] = false;
2937 losers++;
2938 overall += LRA_LOSER_COST_FACTOR;
2940 if (! curr_alt_match_win[i])
2941 curr_alt_dont_inherit_ops[curr_alt_dont_inherit_ops_num++] = i;
2942 else
2944 /* Remember pseudos used for match reloads are never
2945 inherited. */
2946 lra_assert (curr_alt_matches[i] >= 0);
2947 curr_alt_win[curr_alt_matches[i]] = false;
2949 curr_alt_win[i] = curr_alt_match_win[i] = false;
2950 losers++;
2951 /* Early clobber was already reflected in REJECT. */
2952 lra_assert (reject > 0);
2953 if (lra_dump_file != NULL)
2954 fprintf
2955 (lra_dump_file,
2956 " %d Matched conflict early clobber reloads: "
2957 "reject--\n",
2959 reject--;
2960 overall += LRA_LOSER_COST_FACTOR - 1;
2963 if (lra_dump_file != NULL)
2964 fprintf (lra_dump_file, " alt=%d,overall=%d,losers=%d,rld_nregs=%d\n",
2965 nalt, overall, losers, reload_nregs);
2967 /* If this alternative can be made to work by reloading, and it
2968 needs less reloading than the others checked so far, record
2969 it as the chosen goal for reloading. */
2970 if ((best_losers != 0 && losers == 0)
2971 || (((best_losers == 0 && losers == 0)
2972 || (best_losers != 0 && losers != 0))
2973 && (best_overall > overall
2974 || (best_overall == overall
2975 /* If the cost of the reloads is the same,
2976 prefer alternative which requires minimal
2977 number of reload regs. */
2978 && (reload_nregs < best_reload_nregs
2979 || (reload_nregs == best_reload_nregs
2980 && (best_reload_sum < reload_sum
2981 || (best_reload_sum == reload_sum
2982 && nalt < goal_alt_number))))))))
2984 for (nop = 0; nop < n_operands; nop++)
2986 goal_alt_win[nop] = curr_alt_win[nop];
2987 goal_alt_match_win[nop] = curr_alt_match_win[nop];
2988 goal_alt_matches[nop] = curr_alt_matches[nop];
2989 goal_alt[nop] = curr_alt[nop];
2990 goal_alt_offmemok[nop] = curr_alt_offmemok[nop];
2992 goal_alt_dont_inherit_ops_num = curr_alt_dont_inherit_ops_num;
2993 for (nop = 0; nop < curr_alt_dont_inherit_ops_num; nop++)
2994 goal_alt_dont_inherit_ops[nop] = curr_alt_dont_inherit_ops[nop];
2995 goal_alt_swapped = curr_swapped;
2996 best_overall = overall;
2997 best_losers = losers;
2998 best_reload_nregs = reload_nregs;
2999 best_reload_sum = reload_sum;
3000 goal_alt_number = nalt;
3002 if (losers == 0)
3003 /* Everything is satisfied. Do not process alternatives
3004 anymore. */
3005 break;
3006 fail:
3009 return ok_p;
3012 /* Make reload base reg from address AD. */
3013 static rtx
3014 base_to_reg (struct address_info *ad)
3016 enum reg_class cl;
3017 int code = -1;
3018 rtx new_inner = NULL_RTX;
3019 rtx new_reg = NULL_RTX;
3020 rtx_insn *insn;
3021 rtx_insn *last_insn = get_last_insn();
3023 lra_assert (ad->disp == ad->disp_term);
3024 cl = base_reg_class (ad->mode, ad->as, ad->base_outer_code,
3025 get_index_code (ad));
3026 new_reg = lra_create_new_reg (GET_MODE (*ad->base), NULL_RTX,
3027 cl, "base");
3028 new_inner = simplify_gen_binary (PLUS, GET_MODE (new_reg), new_reg,
3029 ad->disp_term == NULL
3030 ? const0_rtx
3031 : *ad->disp_term);
3032 if (!valid_address_p (ad->mode, new_inner, ad->as))
3033 return NULL_RTX;
3034 insn = emit_insn (gen_rtx_SET (new_reg, *ad->base));
3035 code = recog_memoized (insn);
3036 if (code < 0)
3038 delete_insns_since (last_insn);
3039 return NULL_RTX;
3042 return new_inner;
3045 /* Make reload base reg + disp from address AD. Return the new pseudo. */
3046 static rtx
3047 base_plus_disp_to_reg (struct address_info *ad)
3049 enum reg_class cl;
3050 rtx new_reg;
3052 lra_assert (ad->base == ad->base_term && ad->disp == ad->disp_term);
3053 cl = base_reg_class (ad->mode, ad->as, ad->base_outer_code,
3054 get_index_code (ad));
3055 new_reg = lra_create_new_reg (GET_MODE (*ad->base_term), NULL_RTX,
3056 cl, "base + disp");
3057 lra_emit_add (new_reg, *ad->base_term, *ad->disp_term);
3058 return new_reg;
3061 /* Make reload of index part of address AD. Return the new
3062 pseudo. */
3063 static rtx
3064 index_part_to_reg (struct address_info *ad)
3066 rtx new_reg;
3068 new_reg = lra_create_new_reg (GET_MODE (*ad->index), NULL_RTX,
3069 INDEX_REG_CLASS, "index term");
3070 expand_mult (GET_MODE (*ad->index), *ad->index_term,
3071 GEN_INT (get_index_scale (ad)), new_reg, 1);
3072 return new_reg;
3075 /* Return true if we can add a displacement to address AD, even if that
3076 makes the address invalid. The fix-up code requires any new address
3077 to be the sum of the BASE_TERM, INDEX and DISP_TERM fields. */
3078 static bool
3079 can_add_disp_p (struct address_info *ad)
3081 return (!ad->autoinc_p
3082 && ad->segment == NULL
3083 && ad->base == ad->base_term
3084 && ad->disp == ad->disp_term);
3087 /* Make equiv substitution in address AD. Return true if a substitution
3088 was made. */
3089 static bool
3090 equiv_address_substitution (struct address_info *ad)
3092 rtx base_reg, new_base_reg, index_reg, new_index_reg, *base_term, *index_term;
3093 HOST_WIDE_INT disp, scale;
3094 bool change_p;
3096 base_term = strip_subreg (ad->base_term);
3097 if (base_term == NULL)
3098 base_reg = new_base_reg = NULL_RTX;
3099 else
3101 base_reg = *base_term;
3102 new_base_reg = get_equiv_with_elimination (base_reg, curr_insn);
3104 index_term = strip_subreg (ad->index_term);
3105 if (index_term == NULL)
3106 index_reg = new_index_reg = NULL_RTX;
3107 else
3109 index_reg = *index_term;
3110 new_index_reg = get_equiv_with_elimination (index_reg, curr_insn);
3112 if (base_reg == new_base_reg && index_reg == new_index_reg)
3113 return false;
3114 disp = 0;
3115 change_p = false;
3116 if (lra_dump_file != NULL)
3118 fprintf (lra_dump_file, "Changing address in insn %d ",
3119 INSN_UID (curr_insn));
3120 dump_value_slim (lra_dump_file, *ad->outer, 1);
3122 if (base_reg != new_base_reg)
3124 if (REG_P (new_base_reg))
3126 *base_term = new_base_reg;
3127 change_p = true;
3129 else if (GET_CODE (new_base_reg) == PLUS
3130 && REG_P (XEXP (new_base_reg, 0))
3131 && CONST_INT_P (XEXP (new_base_reg, 1))
3132 && can_add_disp_p (ad))
3134 disp += INTVAL (XEXP (new_base_reg, 1));
3135 *base_term = XEXP (new_base_reg, 0);
3136 change_p = true;
3138 if (ad->base_term2 != NULL)
3139 *ad->base_term2 = *ad->base_term;
3141 if (index_reg != new_index_reg)
3143 if (REG_P (new_index_reg))
3145 *index_term = new_index_reg;
3146 change_p = true;
3148 else if (GET_CODE (new_index_reg) == PLUS
3149 && REG_P (XEXP (new_index_reg, 0))
3150 && CONST_INT_P (XEXP (new_index_reg, 1))
3151 && can_add_disp_p (ad)
3152 && (scale = get_index_scale (ad)))
3154 disp += INTVAL (XEXP (new_index_reg, 1)) * scale;
3155 *index_term = XEXP (new_index_reg, 0);
3156 change_p = true;
3159 if (disp != 0)
3161 if (ad->disp != NULL)
3162 *ad->disp = plus_constant (GET_MODE (*ad->inner), *ad->disp, disp);
3163 else
3165 *ad->inner = plus_constant (GET_MODE (*ad->inner), *ad->inner, disp);
3166 update_address (ad);
3168 change_p = true;
3170 if (lra_dump_file != NULL)
3172 if (! change_p)
3173 fprintf (lra_dump_file, " -- no change\n");
3174 else
3176 fprintf (lra_dump_file, " on equiv ");
3177 dump_value_slim (lra_dump_file, *ad->outer, 1);
3178 fprintf (lra_dump_file, "\n");
3181 return change_p;
3184 /* Major function to make reloads for an address in operand NOP or
3185 check its correctness (If CHECK_ONLY_P is true). The supported
3186 cases are:
3188 1) an address that existed before LRA started, at which point it
3189 must have been valid. These addresses are subject to elimination
3190 and may have become invalid due to the elimination offset being out
3191 of range.
3193 2) an address created by forcing a constant to memory
3194 (force_const_to_mem). The initial form of these addresses might
3195 not be valid, and it is this function's job to make them valid.
3197 3) a frame address formed from a register and a (possibly zero)
3198 constant offset. As above, these addresses might not be valid and
3199 this function must make them so.
3201 Add reloads to the lists *BEFORE and *AFTER. We might need to add
3202 reloads to *AFTER because of inc/dec, {pre, post} modify in the
3203 address. Return true for any RTL change.
3205 The function is a helper function which does not produce all
3206 transformations (when CHECK_ONLY_P is false) which can be
3207 necessary. It does just basic steps. To do all necessary
3208 transformations use function process_address. */
3209 static bool
3210 process_address_1 (int nop, bool check_only_p,
3211 rtx_insn **before, rtx_insn **after)
3213 struct address_info ad;
3214 rtx new_reg;
3215 HOST_WIDE_INT scale;
3216 rtx op = *curr_id->operand_loc[nop];
3217 const char *constraint = curr_static_id->operand[nop].constraint;
3218 enum constraint_num cn = lookup_constraint (constraint);
3219 bool change_p = false;
3221 if (MEM_P (op)
3222 && GET_MODE (op) == BLKmode
3223 && GET_CODE (XEXP (op, 0)) == SCRATCH)
3224 return false;
3226 if (insn_extra_address_constraint (cn))
3227 decompose_lea_address (&ad, curr_id->operand_loc[nop]);
3228 /* Do not attempt to decompose arbitrary addresses generated by combine
3229 for asm operands with loose constraints, e.g 'X'. */
3230 else if (MEM_P (op)
3231 && !(get_constraint_type (cn) == CT_FIXED_FORM
3232 && constraint_satisfied_p (op, cn)))
3233 decompose_mem_address (&ad, op);
3234 else if (GET_CODE (op) == SUBREG
3235 && MEM_P (SUBREG_REG (op)))
3236 decompose_mem_address (&ad, SUBREG_REG (op));
3237 else
3238 return false;
3239 /* If INDEX_REG_CLASS is assigned to base_term already and isn't to
3240 index_term, swap them so to avoid assigning INDEX_REG_CLASS to both
3241 when INDEX_REG_CLASS is a single register class. */
3242 if (ad.base_term != NULL
3243 && ad.index_term != NULL
3244 && ira_class_hard_regs_num[INDEX_REG_CLASS] == 1
3245 && REG_P (*ad.base_term)
3246 && REG_P (*ad.index_term)
3247 && in_class_p (*ad.base_term, INDEX_REG_CLASS, NULL)
3248 && ! in_class_p (*ad.index_term, INDEX_REG_CLASS, NULL))
3250 std::swap (ad.base, ad.index);
3251 std::swap (ad.base_term, ad.index_term);
3253 if (! check_only_p)
3254 change_p = equiv_address_substitution (&ad);
3255 if (ad.base_term != NULL
3256 && (process_addr_reg
3257 (ad.base_term, check_only_p, before,
3258 (ad.autoinc_p
3259 && !(REG_P (*ad.base_term)
3260 && find_regno_note (curr_insn, REG_DEAD,
3261 REGNO (*ad.base_term)) != NULL_RTX)
3262 ? after : NULL),
3263 base_reg_class (ad.mode, ad.as, ad.base_outer_code,
3264 get_index_code (&ad)))))
3266 change_p = true;
3267 if (ad.base_term2 != NULL)
3268 *ad.base_term2 = *ad.base_term;
3270 if (ad.index_term != NULL
3271 && process_addr_reg (ad.index_term, check_only_p,
3272 before, NULL, INDEX_REG_CLASS))
3273 change_p = true;
3275 /* Target hooks sometimes don't treat extra-constraint addresses as
3276 legitimate address_operands, so handle them specially. */
3277 if (insn_extra_address_constraint (cn)
3278 && satisfies_address_constraint_p (&ad, cn))
3279 return change_p;
3281 if (check_only_p)
3282 return change_p;
3284 /* There are three cases where the shape of *AD.INNER may now be invalid:
3286 1) the original address was valid, but either elimination or
3287 equiv_address_substitution was applied and that made
3288 the address invalid.
3290 2) the address is an invalid symbolic address created by
3291 force_const_to_mem.
3293 3) the address is a frame address with an invalid offset.
3295 4) the address is a frame address with an invalid base.
3297 All these cases involve a non-autoinc address, so there is no
3298 point revalidating other types. */
3299 if (ad.autoinc_p || valid_address_p (&ad))
3300 return change_p;
3302 /* Any index existed before LRA started, so we can assume that the
3303 presence and shape of the index is valid. */
3304 push_to_sequence (*before);
3305 lra_assert (ad.disp == ad.disp_term);
3306 if (ad.base == NULL)
3308 if (ad.index == NULL)
3310 rtx_insn *insn;
3311 rtx_insn *last = get_last_insn ();
3312 int code = -1;
3313 enum reg_class cl = base_reg_class (ad.mode, ad.as,
3314 SCRATCH, SCRATCH);
3315 rtx addr = *ad.inner;
3317 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "addr");
3318 if (HAVE_lo_sum)
3320 /* addr => lo_sum (new_base, addr), case (2) above. */
3321 insn = emit_insn (gen_rtx_SET
3322 (new_reg,
3323 gen_rtx_HIGH (Pmode, copy_rtx (addr))));
3324 code = recog_memoized (insn);
3325 if (code >= 0)
3327 *ad.inner = gen_rtx_LO_SUM (Pmode, new_reg, addr);
3328 if (! valid_address_p (ad.mode, *ad.outer, ad.as))
3330 /* Try to put lo_sum into register. */
3331 insn = emit_insn (gen_rtx_SET
3332 (new_reg,
3333 gen_rtx_LO_SUM (Pmode, new_reg, addr)));
3334 code = recog_memoized (insn);
3335 if (code >= 0)
3337 *ad.inner = new_reg;
3338 if (! valid_address_p (ad.mode, *ad.outer, ad.as))
3340 *ad.inner = addr;
3341 code = -1;
3347 if (code < 0)
3348 delete_insns_since (last);
3351 if (code < 0)
3353 /* addr => new_base, case (2) above. */
3354 lra_emit_move (new_reg, addr);
3356 for (insn = last == NULL_RTX ? get_insns () : NEXT_INSN (last);
3357 insn != NULL_RTX;
3358 insn = NEXT_INSN (insn))
3359 if (recog_memoized (insn) < 0)
3360 break;
3361 if (insn != NULL_RTX)
3363 /* Do nothing if we cannot generate right insns.
3364 This is analogous to reload pass behavior. */
3365 delete_insns_since (last);
3366 end_sequence ();
3367 return false;
3369 *ad.inner = new_reg;
3372 else
3374 /* index * scale + disp => new base + index * scale,
3375 case (1) above. */
3376 enum reg_class cl = base_reg_class (ad.mode, ad.as, PLUS,
3377 GET_CODE (*ad.index));
3379 lra_assert (INDEX_REG_CLASS != NO_REGS);
3380 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "disp");
3381 lra_emit_move (new_reg, *ad.disp);
3382 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3383 new_reg, *ad.index);
3386 else if (ad.index == NULL)
3388 int regno;
3389 enum reg_class cl;
3390 rtx set;
3391 rtx_insn *insns, *last_insn;
3392 /* Try to reload base into register only if the base is invalid
3393 for the address but with valid offset, case (4) above. */
3394 start_sequence ();
3395 new_reg = base_to_reg (&ad);
3397 /* base + disp => new base, cases (1) and (3) above. */
3398 /* Another option would be to reload the displacement into an
3399 index register. However, postreload has code to optimize
3400 address reloads that have the same base and different
3401 displacements, so reloading into an index register would
3402 not necessarily be a win. */
3403 if (new_reg == NULL_RTX)
3404 new_reg = base_plus_disp_to_reg (&ad);
3405 insns = get_insns ();
3406 last_insn = get_last_insn ();
3407 /* If we generated at least two insns, try last insn source as
3408 an address. If we succeed, we generate one less insn. */
3409 if (last_insn != insns && (set = single_set (last_insn)) != NULL_RTX
3410 && GET_CODE (SET_SRC (set)) == PLUS
3411 && REG_P (XEXP (SET_SRC (set), 0))
3412 && CONSTANT_P (XEXP (SET_SRC (set), 1)))
3414 *ad.inner = SET_SRC (set);
3415 if (valid_address_p (ad.mode, *ad.outer, ad.as))
3417 *ad.base_term = XEXP (SET_SRC (set), 0);
3418 *ad.disp_term = XEXP (SET_SRC (set), 1);
3419 cl = base_reg_class (ad.mode, ad.as, ad.base_outer_code,
3420 get_index_code (&ad));
3421 regno = REGNO (*ad.base_term);
3422 if (regno >= FIRST_PSEUDO_REGISTER
3423 && cl != lra_get_allocno_class (regno))
3424 lra_change_class (regno, cl, " Change to", true);
3425 new_reg = SET_SRC (set);
3426 delete_insns_since (PREV_INSN (last_insn));
3429 /* Try if target can split displacement into legitimite new disp
3430 and offset. If it's the case, we replace the last insn with
3431 insns for base + offset => new_reg and set new_reg + new disp
3432 to *ad.inner. */
3433 last_insn = get_last_insn ();
3434 if ((set = single_set (last_insn)) != NULL_RTX
3435 && GET_CODE (SET_SRC (set)) == PLUS
3436 && REG_P (XEXP (SET_SRC (set), 0))
3437 && REGNO (XEXP (SET_SRC (set), 0)) < FIRST_PSEUDO_REGISTER
3438 && CONST_INT_P (XEXP (SET_SRC (set), 1)))
3440 rtx addend, disp = XEXP (SET_SRC (set), 1);
3441 if (targetm.legitimize_address_displacement (&disp, &addend,
3442 ad.mode))
3444 rtx_insn *new_insns;
3445 start_sequence ();
3446 lra_emit_add (new_reg, XEXP (SET_SRC (set), 0), addend);
3447 new_insns = get_insns ();
3448 end_sequence ();
3449 new_reg = gen_rtx_PLUS (Pmode, new_reg, disp);
3450 delete_insns_since (PREV_INSN (last_insn));
3451 add_insn (new_insns);
3452 insns = get_insns ();
3455 end_sequence ();
3456 emit_insn (insns);
3457 *ad.inner = new_reg;
3459 else if (ad.disp_term != NULL)
3461 /* base + scale * index + disp => new base + scale * index,
3462 case (1) above. */
3463 new_reg = base_plus_disp_to_reg (&ad);
3464 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3465 new_reg, *ad.index);
3467 else if ((scale = get_index_scale (&ad)) == 1)
3469 /* The last transformation to one reg will be made in
3470 curr_insn_transform function. */
3471 end_sequence ();
3472 return false;
3474 else if (scale != 0)
3476 /* base + scale * index => base + new_reg,
3477 case (1) above.
3478 Index part of address may become invalid. For example, we
3479 changed pseudo on the equivalent memory and a subreg of the
3480 pseudo onto the memory of different mode for which the scale is
3481 prohibitted. */
3482 new_reg = index_part_to_reg (&ad);
3483 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3484 *ad.base_term, new_reg);
3486 else
3488 enum reg_class cl = base_reg_class (ad.mode, ad.as,
3489 SCRATCH, SCRATCH);
3490 rtx addr = *ad.inner;
3492 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "addr");
3493 /* addr => new_base. */
3494 lra_emit_move (new_reg, addr);
3495 *ad.inner = new_reg;
3497 *before = get_insns ();
3498 end_sequence ();
3499 return true;
3502 /* If CHECK_ONLY_P is false, do address reloads until it is necessary.
3503 Use process_address_1 as a helper function. Return true for any
3504 RTL changes.
3506 If CHECK_ONLY_P is true, just check address correctness. Return
3507 false if the address correct. */
3508 static bool
3509 process_address (int nop, bool check_only_p,
3510 rtx_insn **before, rtx_insn **after)
3512 bool res = false;
3514 while (process_address_1 (nop, check_only_p, before, after))
3516 if (check_only_p)
3517 return true;
3518 res = true;
3520 return res;
3523 /* Emit insns to reload VALUE into a new register. VALUE is an
3524 auto-increment or auto-decrement RTX whose operand is a register or
3525 memory location; so reloading involves incrementing that location.
3526 IN is either identical to VALUE, or some cheaper place to reload
3527 value being incremented/decremented from.
3529 INC_AMOUNT is the number to increment or decrement by (always
3530 positive and ignored for POST_MODIFY/PRE_MODIFY).
3532 Return pseudo containing the result. */
3533 static rtx
3534 emit_inc (enum reg_class new_rclass, rtx in, rtx value, int inc_amount)
3536 /* REG or MEM to be copied and incremented. */
3537 rtx incloc = XEXP (value, 0);
3538 /* Nonzero if increment after copying. */
3539 int post = (GET_CODE (value) == POST_DEC || GET_CODE (value) == POST_INC
3540 || GET_CODE (value) == POST_MODIFY);
3541 rtx_insn *last;
3542 rtx inc;
3543 rtx_insn *add_insn;
3544 int code;
3545 rtx real_in = in == value ? incloc : in;
3546 rtx result;
3547 bool plus_p = true;
3549 if (GET_CODE (value) == PRE_MODIFY || GET_CODE (value) == POST_MODIFY)
3551 lra_assert (GET_CODE (XEXP (value, 1)) == PLUS
3552 || GET_CODE (XEXP (value, 1)) == MINUS);
3553 lra_assert (rtx_equal_p (XEXP (XEXP (value, 1), 0), XEXP (value, 0)));
3554 plus_p = GET_CODE (XEXP (value, 1)) == PLUS;
3555 inc = XEXP (XEXP (value, 1), 1);
3557 else
3559 if (GET_CODE (value) == PRE_DEC || GET_CODE (value) == POST_DEC)
3560 inc_amount = -inc_amount;
3562 inc = GEN_INT (inc_amount);
3565 if (! post && REG_P (incloc))
3566 result = incloc;
3567 else
3568 result = lra_create_new_reg (GET_MODE (value), value, new_rclass,
3569 "INC/DEC result");
3571 if (real_in != result)
3573 /* First copy the location to the result register. */
3574 lra_assert (REG_P (result));
3575 emit_insn (gen_move_insn (result, real_in));
3578 /* We suppose that there are insns to add/sub with the constant
3579 increment permitted in {PRE/POST)_{DEC/INC/MODIFY}. At least the
3580 old reload worked with this assumption. If the assumption
3581 becomes wrong, we should use approach in function
3582 base_plus_disp_to_reg. */
3583 if (in == value)
3585 /* See if we can directly increment INCLOC. */
3586 last = get_last_insn ();
3587 add_insn = emit_insn (plus_p
3588 ? gen_add2_insn (incloc, inc)
3589 : gen_sub2_insn (incloc, inc));
3591 code = recog_memoized (add_insn);
3592 if (code >= 0)
3594 if (! post && result != incloc)
3595 emit_insn (gen_move_insn (result, incloc));
3596 return result;
3598 delete_insns_since (last);
3601 /* If couldn't do the increment directly, must increment in RESULT.
3602 The way we do this depends on whether this is pre- or
3603 post-increment. For pre-increment, copy INCLOC to the reload
3604 register, increment it there, then save back. */
3605 if (! post)
3607 if (real_in != result)
3608 emit_insn (gen_move_insn (result, real_in));
3609 if (plus_p)
3610 emit_insn (gen_add2_insn (result, inc));
3611 else
3612 emit_insn (gen_sub2_insn (result, inc));
3613 if (result != incloc)
3614 emit_insn (gen_move_insn (incloc, result));
3616 else
3618 /* Post-increment.
3620 Because this might be a jump insn or a compare, and because
3621 RESULT may not be available after the insn in an input
3622 reload, we must do the incrementing before the insn being
3623 reloaded for.
3625 We have already copied IN to RESULT. Increment the copy in
3626 RESULT, save that back, then decrement RESULT so it has
3627 the original value. */
3628 if (plus_p)
3629 emit_insn (gen_add2_insn (result, inc));
3630 else
3631 emit_insn (gen_sub2_insn (result, inc));
3632 emit_insn (gen_move_insn (incloc, result));
3633 /* Restore non-modified value for the result. We prefer this
3634 way because it does not require an additional hard
3635 register. */
3636 if (plus_p)
3638 if (CONST_INT_P (inc))
3639 emit_insn (gen_add2_insn (result,
3640 gen_int_mode (-INTVAL (inc),
3641 GET_MODE (result))));
3642 else
3643 emit_insn (gen_sub2_insn (result, inc));
3645 else
3646 emit_insn (gen_add2_insn (result, inc));
3648 return result;
3651 /* Return true if the current move insn does not need processing as we
3652 already know that it satisfies its constraints. */
3653 static bool
3654 simple_move_p (void)
3656 rtx dest, src;
3657 enum reg_class dclass, sclass;
3659 lra_assert (curr_insn_set != NULL_RTX);
3660 dest = SET_DEST (curr_insn_set);
3661 src = SET_SRC (curr_insn_set);
3663 /* If the instruction has multiple sets we need to process it even if it
3664 is single_set. This can happen if one or more of the SETs are dead.
3665 See PR73650. */
3666 if (multiple_sets (curr_insn))
3667 return false;
3669 return ((dclass = get_op_class (dest)) != NO_REGS
3670 && (sclass = get_op_class (src)) != NO_REGS
3671 /* The backend guarantees that register moves of cost 2
3672 never need reloads. */
3673 && targetm.register_move_cost (GET_MODE (src), sclass, dclass) == 2);
3676 /* Swap operands NOP and NOP + 1. */
3677 static inline void
3678 swap_operands (int nop)
3680 std::swap (curr_operand_mode[nop], curr_operand_mode[nop + 1]);
3681 std::swap (original_subreg_reg_mode[nop], original_subreg_reg_mode[nop + 1]);
3682 std::swap (*curr_id->operand_loc[nop], *curr_id->operand_loc[nop + 1]);
3683 std::swap (equiv_substition_p[nop], equiv_substition_p[nop + 1]);
3684 /* Swap the duplicates too. */
3685 lra_update_dup (curr_id, nop);
3686 lra_update_dup (curr_id, nop + 1);
3689 /* Main entry point of the constraint code: search the body of the
3690 current insn to choose the best alternative. It is mimicking insn
3691 alternative cost calculation model of former reload pass. That is
3692 because machine descriptions were written to use this model. This
3693 model can be changed in future. Make commutative operand exchange
3694 if it is chosen.
3696 if CHECK_ONLY_P is false, do RTL changes to satisfy the
3697 constraints. Return true if any change happened during function
3698 call.
3700 If CHECK_ONLY_P is true then don't do any transformation. Just
3701 check that the insn satisfies all constraints. If the insn does
3702 not satisfy any constraint, return true. */
3703 static bool
3704 curr_insn_transform (bool check_only_p)
3706 int i, j, k;
3707 int n_operands;
3708 int n_alternatives;
3709 int n_outputs;
3710 int commutative;
3711 signed char goal_alt_matched[MAX_RECOG_OPERANDS][MAX_RECOG_OPERANDS];
3712 signed char match_inputs[MAX_RECOG_OPERANDS + 1];
3713 signed char outputs[MAX_RECOG_OPERANDS + 1];
3714 rtx_insn *before, *after;
3715 bool alt_p = false;
3716 /* Flag that the insn has been changed through a transformation. */
3717 bool change_p;
3718 bool sec_mem_p;
3719 #ifdef SECONDARY_MEMORY_NEEDED
3720 bool use_sec_mem_p;
3721 #endif
3722 int max_regno_before;
3723 int reused_alternative_num;
3725 curr_insn_set = single_set (curr_insn);
3726 if (curr_insn_set != NULL_RTX && simple_move_p ())
3727 return false;
3729 no_input_reloads_p = no_output_reloads_p = false;
3730 goal_alt_number = -1;
3731 change_p = sec_mem_p = false;
3732 /* JUMP_INSNs and CALL_INSNs are not allowed to have any output
3733 reloads; neither are insns that SET cc0. Insns that use CC0 are
3734 not allowed to have any input reloads. */
3735 if (JUMP_P (curr_insn) || CALL_P (curr_insn))
3736 no_output_reloads_p = true;
3738 if (HAVE_cc0 && reg_referenced_p (cc0_rtx, PATTERN (curr_insn)))
3739 no_input_reloads_p = true;
3740 if (HAVE_cc0 && reg_set_p (cc0_rtx, PATTERN (curr_insn)))
3741 no_output_reloads_p = true;
3743 n_operands = curr_static_id->n_operands;
3744 n_alternatives = curr_static_id->n_alternatives;
3746 /* Just return "no reloads" if insn has no operands with
3747 constraints. */
3748 if (n_operands == 0 || n_alternatives == 0)
3749 return false;
3751 max_regno_before = max_reg_num ();
3753 for (i = 0; i < n_operands; i++)
3755 goal_alt_matched[i][0] = -1;
3756 goal_alt_matches[i] = -1;
3759 commutative = curr_static_id->commutative;
3761 /* Now see what we need for pseudos that didn't get hard regs or got
3762 the wrong kind of hard reg. For this, we must consider all the
3763 operands together against the register constraints. */
3765 best_losers = best_overall = INT_MAX;
3766 best_reload_sum = 0;
3768 curr_swapped = false;
3769 goal_alt_swapped = false;
3771 if (! check_only_p)
3772 /* Make equivalence substitution and memory subreg elimination
3773 before address processing because an address legitimacy can
3774 depend on memory mode. */
3775 for (i = 0; i < n_operands; i++)
3777 rtx op, subst, old;
3778 bool op_change_p = false;
3780 if (curr_static_id->operand[i].is_operator)
3781 continue;
3783 old = op = *curr_id->operand_loc[i];
3784 if (GET_CODE (old) == SUBREG)
3785 old = SUBREG_REG (old);
3786 subst = get_equiv_with_elimination (old, curr_insn);
3787 original_subreg_reg_mode[i] = VOIDmode;
3788 equiv_substition_p[i] = false;
3789 if (subst != old)
3791 equiv_substition_p[i] = true;
3792 subst = copy_rtx (subst);
3793 lra_assert (REG_P (old));
3794 if (GET_CODE (op) != SUBREG)
3795 *curr_id->operand_loc[i] = subst;
3796 else
3798 SUBREG_REG (op) = subst;
3799 if (GET_MODE (subst) == VOIDmode)
3800 original_subreg_reg_mode[i] = GET_MODE (old);
3802 if (lra_dump_file != NULL)
3804 fprintf (lra_dump_file,
3805 "Changing pseudo %d in operand %i of insn %u on equiv ",
3806 REGNO (old), i, INSN_UID (curr_insn));
3807 dump_value_slim (lra_dump_file, subst, 1);
3808 fprintf (lra_dump_file, "\n");
3810 op_change_p = change_p = true;
3812 if (simplify_operand_subreg (i, GET_MODE (old)) || op_change_p)
3814 change_p = true;
3815 lra_update_dup (curr_id, i);
3819 /* Reload address registers and displacements. We do it before
3820 finding an alternative because of memory constraints. */
3821 before = after = NULL;
3822 for (i = 0; i < n_operands; i++)
3823 if (! curr_static_id->operand[i].is_operator
3824 && process_address (i, check_only_p, &before, &after))
3826 if (check_only_p)
3827 return true;
3828 change_p = true;
3829 lra_update_dup (curr_id, i);
3832 if (change_p)
3833 /* If we've changed the instruction then any alternative that
3834 we chose previously may no longer be valid. */
3835 lra_set_used_insn_alternative (curr_insn, -1);
3837 if (! check_only_p && curr_insn_set != NULL_RTX
3838 && check_and_process_move (&change_p, &sec_mem_p))
3839 return change_p;
3841 try_swapped:
3843 reused_alternative_num = check_only_p ? -1 : curr_id->used_insn_alternative;
3844 if (lra_dump_file != NULL && reused_alternative_num >= 0)
3845 fprintf (lra_dump_file, "Reusing alternative %d for insn #%u\n",
3846 reused_alternative_num, INSN_UID (curr_insn));
3848 if (process_alt_operands (reused_alternative_num))
3849 alt_p = true;
3851 if (check_only_p)
3852 return ! alt_p || best_losers != 0;
3854 /* If insn is commutative (it's safe to exchange a certain pair of
3855 operands) then we need to try each alternative twice, the second
3856 time matching those two operands as if we had exchanged them. To
3857 do this, really exchange them in operands.
3859 If we have just tried the alternatives the second time, return
3860 operands to normal and drop through. */
3862 if (reused_alternative_num < 0 && commutative >= 0)
3864 curr_swapped = !curr_swapped;
3865 if (curr_swapped)
3867 swap_operands (commutative);
3868 goto try_swapped;
3870 else
3871 swap_operands (commutative);
3874 if (! alt_p && ! sec_mem_p)
3876 /* No alternative works with reloads?? */
3877 if (INSN_CODE (curr_insn) >= 0)
3878 fatal_insn ("unable to generate reloads for:", curr_insn);
3879 error_for_asm (curr_insn,
3880 "inconsistent operand constraints in an %<asm%>");
3881 /* Avoid further trouble with this insn. Don't generate use
3882 pattern here as we could use the insn SP offset. */
3883 lra_set_insn_deleted (curr_insn);
3884 return true;
3887 /* If the best alternative is with operands 1 and 2 swapped, swap
3888 them. Update the operand numbers of any reloads already
3889 pushed. */
3891 if (goal_alt_swapped)
3893 if (lra_dump_file != NULL)
3894 fprintf (lra_dump_file, " Commutative operand exchange in insn %u\n",
3895 INSN_UID (curr_insn));
3897 /* Swap the duplicates too. */
3898 swap_operands (commutative);
3899 change_p = true;
3902 #ifdef SECONDARY_MEMORY_NEEDED
3903 /* Some target macros SECONDARY_MEMORY_NEEDED (e.g. x86) are defined
3904 too conservatively. So we use the secondary memory only if there
3905 is no any alternative without reloads. */
3906 use_sec_mem_p = false;
3907 if (! alt_p)
3908 use_sec_mem_p = true;
3909 else if (sec_mem_p)
3911 for (i = 0; i < n_operands; i++)
3912 if (! goal_alt_win[i] && ! goal_alt_match_win[i])
3913 break;
3914 use_sec_mem_p = i < n_operands;
3917 if (use_sec_mem_p)
3919 int in = -1, out = -1;
3920 rtx new_reg, src, dest, rld;
3921 machine_mode sec_mode, rld_mode;
3923 lra_assert (curr_insn_set != NULL_RTX && sec_mem_p);
3924 dest = SET_DEST (curr_insn_set);
3925 src = SET_SRC (curr_insn_set);
3926 for (i = 0; i < n_operands; i++)
3927 if (*curr_id->operand_loc[i] == dest)
3928 out = i;
3929 else if (*curr_id->operand_loc[i] == src)
3930 in = i;
3931 for (i = 0; i < curr_static_id->n_dups; i++)
3932 if (out < 0 && *curr_id->dup_loc[i] == dest)
3933 out = curr_static_id->dup_num[i];
3934 else if (in < 0 && *curr_id->dup_loc[i] == src)
3935 in = curr_static_id->dup_num[i];
3936 lra_assert (out >= 0 && in >= 0
3937 && curr_static_id->operand[out].type == OP_OUT
3938 && curr_static_id->operand[in].type == OP_IN);
3939 rld = (GET_MODE_SIZE (GET_MODE (dest)) <= GET_MODE_SIZE (GET_MODE (src))
3940 ? dest : src);
3941 rld_mode = GET_MODE (rld);
3942 #ifdef SECONDARY_MEMORY_NEEDED_MODE
3943 sec_mode = SECONDARY_MEMORY_NEEDED_MODE (rld_mode);
3944 #else
3945 sec_mode = rld_mode;
3946 #endif
3947 new_reg = lra_create_new_reg (sec_mode, NULL_RTX,
3948 NO_REGS, "secondary");
3949 /* If the mode is changed, it should be wider. */
3950 lra_assert (GET_MODE_SIZE (sec_mode) >= GET_MODE_SIZE (rld_mode));
3951 if (sec_mode != rld_mode)
3953 /* If the target says specifically to use another mode for
3954 secondary memory moves we can not reuse the original
3955 insn. */
3956 after = emit_spill_move (false, new_reg, dest);
3957 lra_process_new_insns (curr_insn, NULL, after,
3958 "Inserting the sec. move");
3959 /* We may have non null BEFORE here (e.g. after address
3960 processing. */
3961 push_to_sequence (before);
3962 before = emit_spill_move (true, new_reg, src);
3963 emit_insn (before);
3964 before = get_insns ();
3965 end_sequence ();
3966 lra_process_new_insns (curr_insn, before, NULL, "Changing on");
3967 lra_set_insn_deleted (curr_insn);
3969 else if (dest == rld)
3971 *curr_id->operand_loc[out] = new_reg;
3972 lra_update_dup (curr_id, out);
3973 after = emit_spill_move (false, new_reg, dest);
3974 lra_process_new_insns (curr_insn, NULL, after,
3975 "Inserting the sec. move");
3977 else
3979 *curr_id->operand_loc[in] = new_reg;
3980 lra_update_dup (curr_id, in);
3981 /* See comments above. */
3982 push_to_sequence (before);
3983 before = emit_spill_move (true, new_reg, src);
3984 emit_insn (before);
3985 before = get_insns ();
3986 end_sequence ();
3987 lra_process_new_insns (curr_insn, before, NULL,
3988 "Inserting the sec. move");
3990 lra_update_insn_regno_info (curr_insn);
3991 return true;
3993 #endif
3995 lra_assert (goal_alt_number >= 0);
3996 lra_set_used_insn_alternative (curr_insn, goal_alt_number);
3998 if (lra_dump_file != NULL)
4000 const char *p;
4002 fprintf (lra_dump_file, " Choosing alt %d in insn %u:",
4003 goal_alt_number, INSN_UID (curr_insn));
4004 for (i = 0; i < n_operands; i++)
4006 p = (curr_static_id->operand_alternative
4007 [goal_alt_number * n_operands + i].constraint);
4008 if (*p == '\0')
4009 continue;
4010 fprintf (lra_dump_file, " (%d) ", i);
4011 for (; *p != '\0' && *p != ',' && *p != '#'; p++)
4012 fputc (*p, lra_dump_file);
4014 if (INSN_CODE (curr_insn) >= 0
4015 && (p = get_insn_name (INSN_CODE (curr_insn))) != NULL)
4016 fprintf (lra_dump_file, " {%s}", p);
4017 if (curr_id->sp_offset != 0)
4018 fprintf (lra_dump_file, " (sp_off=%" HOST_WIDE_INT_PRINT "d)",
4019 curr_id->sp_offset);
4020 fprintf (lra_dump_file, "\n");
4023 /* Right now, for any pair of operands I and J that are required to
4024 match, with J < I, goal_alt_matches[I] is J. Add I to
4025 goal_alt_matched[J]. */
4027 for (i = 0; i < n_operands; i++)
4028 if ((j = goal_alt_matches[i]) >= 0)
4030 for (k = 0; goal_alt_matched[j][k] >= 0; k++)
4032 /* We allow matching one output operand and several input
4033 operands. */
4034 lra_assert (k == 0
4035 || (curr_static_id->operand[j].type == OP_OUT
4036 && curr_static_id->operand[i].type == OP_IN
4037 && (curr_static_id->operand
4038 [goal_alt_matched[j][0]].type == OP_IN)));
4039 goal_alt_matched[j][k] = i;
4040 goal_alt_matched[j][k + 1] = -1;
4043 for (i = 0; i < n_operands; i++)
4044 goal_alt_win[i] |= goal_alt_match_win[i];
4046 /* Any constants that aren't allowed and can't be reloaded into
4047 registers are here changed into memory references. */
4048 for (i = 0; i < n_operands; i++)
4049 if (goal_alt_win[i])
4051 int regno;
4052 enum reg_class new_class;
4053 rtx reg = *curr_id->operand_loc[i];
4055 if (GET_CODE (reg) == SUBREG)
4056 reg = SUBREG_REG (reg);
4058 if (REG_P (reg) && (regno = REGNO (reg)) >= FIRST_PSEUDO_REGISTER)
4060 bool ok_p = in_class_p (reg, goal_alt[i], &new_class);
4062 if (new_class != NO_REGS && get_reg_class (regno) != new_class)
4064 lra_assert (ok_p);
4065 lra_change_class (regno, new_class, " Change to", true);
4069 else
4071 const char *constraint;
4072 char c;
4073 rtx op = *curr_id->operand_loc[i];
4074 rtx subreg = NULL_RTX;
4075 machine_mode mode = curr_operand_mode[i];
4077 if (GET_CODE (op) == SUBREG)
4079 subreg = op;
4080 op = SUBREG_REG (op);
4081 mode = GET_MODE (op);
4084 if (CONST_POOL_OK_P (mode, op)
4085 && ((targetm.preferred_reload_class
4086 (op, (enum reg_class) goal_alt[i]) == NO_REGS)
4087 || no_input_reloads_p))
4089 rtx tem = force_const_mem (mode, op);
4091 change_p = true;
4092 if (subreg != NULL_RTX)
4093 tem = gen_rtx_SUBREG (mode, tem, SUBREG_BYTE (subreg));
4095 *curr_id->operand_loc[i] = tem;
4096 lra_update_dup (curr_id, i);
4097 process_address (i, false, &before, &after);
4099 /* If the alternative accepts constant pool refs directly
4100 there will be no reload needed at all. */
4101 if (subreg != NULL_RTX)
4102 continue;
4103 /* Skip alternatives before the one requested. */
4104 constraint = (curr_static_id->operand_alternative
4105 [goal_alt_number * n_operands + i].constraint);
4106 for (;
4107 (c = *constraint) && c != ',' && c != '#';
4108 constraint += CONSTRAINT_LEN (c, constraint))
4110 enum constraint_num cn = lookup_constraint (constraint);
4111 if ((insn_extra_memory_constraint (cn)
4112 || insn_extra_special_memory_constraint (cn))
4113 && satisfies_memory_constraint_p (tem, cn))
4114 break;
4116 if (c == '\0' || c == ',' || c == '#')
4117 continue;
4119 goal_alt_win[i] = true;
4123 n_outputs = 0;
4124 outputs[0] = -1;
4125 for (i = 0; i < n_operands; i++)
4127 int regno;
4128 bool optional_p = false;
4129 rtx old, new_reg;
4130 rtx op = *curr_id->operand_loc[i];
4132 if (goal_alt_win[i])
4134 if (goal_alt[i] == NO_REGS
4135 && REG_P (op)
4136 /* When we assign NO_REGS it means that we will not
4137 assign a hard register to the scratch pseudo by
4138 assigment pass and the scratch pseudo will be
4139 spilled. Spilled scratch pseudos are transformed
4140 back to scratches at the LRA end. */
4141 && lra_former_scratch_operand_p (curr_insn, i)
4142 && lra_former_scratch_p (REGNO (op)))
4144 int regno = REGNO (op);
4145 lra_change_class (regno, NO_REGS, " Change to", true);
4146 if (lra_get_regno_hard_regno (regno) >= 0)
4147 /* We don't have to mark all insn affected by the
4148 spilled pseudo as there is only one such insn, the
4149 current one. */
4150 reg_renumber[regno] = -1;
4151 lra_assert (bitmap_single_bit_set_p
4152 (&lra_reg_info[REGNO (op)].insn_bitmap));
4154 /* We can do an optional reload. If the pseudo got a hard
4155 reg, we might improve the code through inheritance. If
4156 it does not get a hard register we coalesce memory/memory
4157 moves later. Ignore move insns to avoid cycling. */
4158 if (! lra_simple_p
4159 && lra_undo_inheritance_iter < LRA_MAX_INHERITANCE_PASSES
4160 && goal_alt[i] != NO_REGS && REG_P (op)
4161 && (regno = REGNO (op)) >= FIRST_PSEUDO_REGISTER
4162 && regno < new_regno_start
4163 && ! lra_former_scratch_p (regno)
4164 && reg_renumber[regno] < 0
4165 /* Check that the optional reload pseudo will be able to
4166 hold given mode value. */
4167 && ! (prohibited_class_reg_set_mode_p
4168 (goal_alt[i], reg_class_contents[goal_alt[i]],
4169 PSEUDO_REGNO_MODE (regno)))
4170 && (curr_insn_set == NULL_RTX
4171 || !((REG_P (SET_SRC (curr_insn_set))
4172 || MEM_P (SET_SRC (curr_insn_set))
4173 || GET_CODE (SET_SRC (curr_insn_set)) == SUBREG)
4174 && (REG_P (SET_DEST (curr_insn_set))
4175 || MEM_P (SET_DEST (curr_insn_set))
4176 || GET_CODE (SET_DEST (curr_insn_set)) == SUBREG))))
4177 optional_p = true;
4178 else
4179 continue;
4182 /* Operands that match previous ones have already been handled. */
4183 if (goal_alt_matches[i] >= 0)
4184 continue;
4186 /* We should not have an operand with a non-offsettable address
4187 appearing where an offsettable address will do. It also may
4188 be a case when the address should be special in other words
4189 not a general one (e.g. it needs no index reg). */
4190 if (goal_alt_matched[i][0] == -1 && goal_alt_offmemok[i] && MEM_P (op))
4192 enum reg_class rclass;
4193 rtx *loc = &XEXP (op, 0);
4194 enum rtx_code code = GET_CODE (*loc);
4196 push_to_sequence (before);
4197 rclass = base_reg_class (GET_MODE (op), MEM_ADDR_SPACE (op),
4198 MEM, SCRATCH);
4199 if (GET_RTX_CLASS (code) == RTX_AUTOINC)
4200 new_reg = emit_inc (rclass, *loc, *loc,
4201 /* This value does not matter for MODIFY. */
4202 GET_MODE_SIZE (GET_MODE (op)));
4203 else if (get_reload_reg (OP_IN, Pmode, *loc, rclass, FALSE,
4204 "offsetable address", &new_reg))
4205 lra_emit_move (new_reg, *loc);
4206 before = get_insns ();
4207 end_sequence ();
4208 *loc = new_reg;
4209 lra_update_dup (curr_id, i);
4211 else if (goal_alt_matched[i][0] == -1)
4213 machine_mode mode;
4214 rtx reg, *loc;
4215 int hard_regno, byte;
4216 enum op_type type = curr_static_id->operand[i].type;
4218 loc = curr_id->operand_loc[i];
4219 mode = curr_operand_mode[i];
4220 if (GET_CODE (*loc) == SUBREG)
4222 reg = SUBREG_REG (*loc);
4223 byte = SUBREG_BYTE (*loc);
4224 if (REG_P (reg)
4225 /* Strict_low_part requires reload the register not
4226 the sub-register. */
4227 && (curr_static_id->operand[i].strict_low
4228 || (GET_MODE_SIZE (mode)
4229 <= GET_MODE_SIZE (GET_MODE (reg))
4230 && (hard_regno
4231 = get_try_hard_regno (REGNO (reg))) >= 0
4232 && (simplify_subreg_regno
4233 (hard_regno,
4234 GET_MODE (reg), byte, mode) < 0)
4235 && (goal_alt[i] == NO_REGS
4236 || (simplify_subreg_regno
4237 (ira_class_hard_regs[goal_alt[i]][0],
4238 GET_MODE (reg), byte, mode) >= 0)))))
4240 /* An OP_INOUT is required when reloading a subreg of a
4241 mode wider than a word to ensure that data beyond the
4242 word being reloaded is preserved. Also automatically
4243 ensure that strict_low_part reloads are made into
4244 OP_INOUT which should already be true from the backend
4245 constraints. */
4246 if (type == OP_OUT
4247 && (curr_static_id->operand[i].strict_low
4248 || (GET_MODE_SIZE (GET_MODE (reg)) > UNITS_PER_WORD
4249 && (GET_MODE_SIZE (mode)
4250 < GET_MODE_SIZE (GET_MODE (reg))))))
4251 type = OP_INOUT;
4252 loc = &SUBREG_REG (*loc);
4253 mode = GET_MODE (*loc);
4256 old = *loc;
4257 if (get_reload_reg (type, mode, old, goal_alt[i],
4258 loc != curr_id->operand_loc[i], "", &new_reg)
4259 && type != OP_OUT)
4261 push_to_sequence (before);
4262 lra_emit_move (new_reg, old);
4263 before = get_insns ();
4264 end_sequence ();
4266 *loc = new_reg;
4267 if (type != OP_IN
4268 && find_reg_note (curr_insn, REG_UNUSED, old) == NULL_RTX)
4270 start_sequence ();
4271 lra_emit_move (type == OP_INOUT ? copy_rtx (old) : old, new_reg);
4272 emit_insn (after);
4273 after = get_insns ();
4274 end_sequence ();
4275 *loc = new_reg;
4277 for (j = 0; j < goal_alt_dont_inherit_ops_num; j++)
4278 if (goal_alt_dont_inherit_ops[j] == i)
4280 lra_set_regno_unique_value (REGNO (new_reg));
4281 break;
4283 lra_update_dup (curr_id, i);
4285 else if (curr_static_id->operand[i].type == OP_IN
4286 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4287 == OP_OUT))
4289 /* generate reloads for input and matched outputs. */
4290 match_inputs[0] = i;
4291 match_inputs[1] = -1;
4292 match_reload (goal_alt_matched[i][0], match_inputs, outputs,
4293 goal_alt[i], &before, &after,
4294 curr_static_id->operand_alternative
4295 [goal_alt_number * n_operands + goal_alt_matched[i][0]]
4296 .earlyclobber);
4298 else if (curr_static_id->operand[i].type == OP_OUT
4299 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4300 == OP_IN))
4301 /* Generate reloads for output and matched inputs. */
4302 match_reload (i, goal_alt_matched[i], outputs, goal_alt[i], &before,
4303 &after, curr_static_id->operand_alternative
4304 [goal_alt_number * n_operands + i].earlyclobber);
4305 else if (curr_static_id->operand[i].type == OP_IN
4306 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4307 == OP_IN))
4309 /* Generate reloads for matched inputs. */
4310 match_inputs[0] = i;
4311 for (j = 0; (k = goal_alt_matched[i][j]) >= 0; j++)
4312 match_inputs[j + 1] = k;
4313 match_inputs[j + 1] = -1;
4314 match_reload (-1, match_inputs, outputs, goal_alt[i], &before,
4315 &after, false);
4317 else
4318 /* We must generate code in any case when function
4319 process_alt_operands decides that it is possible. */
4320 gcc_unreachable ();
4322 /* Memorise processed outputs so that output remaining to be processed
4323 can avoid using the same register value (see match_reload). */
4324 if (curr_static_id->operand[i].type == OP_OUT)
4326 outputs[n_outputs++] = i;
4327 outputs[n_outputs] = -1;
4330 if (optional_p)
4332 rtx reg = op;
4334 lra_assert (REG_P (reg));
4335 regno = REGNO (reg);
4336 op = *curr_id->operand_loc[i]; /* Substitution. */
4337 if (GET_CODE (op) == SUBREG)
4338 op = SUBREG_REG (op);
4339 gcc_assert (REG_P (op) && (int) REGNO (op) >= new_regno_start);
4340 bitmap_set_bit (&lra_optional_reload_pseudos, REGNO (op));
4341 lra_reg_info[REGNO (op)].restore_rtx = reg;
4342 if (lra_dump_file != NULL)
4343 fprintf (lra_dump_file,
4344 " Making reload reg %d for reg %d optional\n",
4345 REGNO (op), regno);
4348 if (before != NULL_RTX || after != NULL_RTX
4349 || max_regno_before != max_reg_num ())
4350 change_p = true;
4351 if (change_p)
4353 lra_update_operator_dups (curr_id);
4354 /* Something changes -- process the insn. */
4355 lra_update_insn_regno_info (curr_insn);
4357 lra_process_new_insns (curr_insn, before, after, "Inserting insn reload");
4358 return change_p;
4361 /* Return true if INSN satisfies all constraints. In other words, no
4362 reload insns are needed. */
4363 bool
4364 lra_constrain_insn (rtx_insn *insn)
4366 int saved_new_regno_start = new_regno_start;
4367 int saved_new_insn_uid_start = new_insn_uid_start;
4368 bool change_p;
4370 curr_insn = insn;
4371 curr_id = lra_get_insn_recog_data (curr_insn);
4372 curr_static_id = curr_id->insn_static_data;
4373 new_insn_uid_start = get_max_uid ();
4374 new_regno_start = max_reg_num ();
4375 change_p = curr_insn_transform (true);
4376 new_regno_start = saved_new_regno_start;
4377 new_insn_uid_start = saved_new_insn_uid_start;
4378 return ! change_p;
4381 /* Return true if X is in LIST. */
4382 static bool
4383 in_list_p (rtx x, rtx list)
4385 for (; list != NULL_RTX; list = XEXP (list, 1))
4386 if (XEXP (list, 0) == x)
4387 return true;
4388 return false;
4391 /* Return true if X contains an allocatable hard register (if
4392 HARD_REG_P) or a (spilled if SPILLED_P) pseudo. */
4393 static bool
4394 contains_reg_p (rtx x, bool hard_reg_p, bool spilled_p)
4396 int i, j;
4397 const char *fmt;
4398 enum rtx_code code;
4400 code = GET_CODE (x);
4401 if (REG_P (x))
4403 int regno = REGNO (x);
4404 HARD_REG_SET alloc_regs;
4406 if (hard_reg_p)
4408 if (regno >= FIRST_PSEUDO_REGISTER)
4409 regno = lra_get_regno_hard_regno (regno);
4410 if (regno < 0)
4411 return false;
4412 COMPL_HARD_REG_SET (alloc_regs, lra_no_alloc_regs);
4413 return overlaps_hard_reg_set_p (alloc_regs, GET_MODE (x), regno);
4415 else
4417 if (regno < FIRST_PSEUDO_REGISTER)
4418 return false;
4419 if (! spilled_p)
4420 return true;
4421 return lra_get_regno_hard_regno (regno) < 0;
4424 fmt = GET_RTX_FORMAT (code);
4425 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4427 if (fmt[i] == 'e')
4429 if (contains_reg_p (XEXP (x, i), hard_reg_p, spilled_p))
4430 return true;
4432 else if (fmt[i] == 'E')
4434 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4435 if (contains_reg_p (XVECEXP (x, i, j), hard_reg_p, spilled_p))
4436 return true;
4439 return false;
4442 /* Process all regs in location *LOC and change them on equivalent
4443 substitution. Return true if any change was done. */
4444 static bool
4445 loc_equivalence_change_p (rtx *loc)
4447 rtx subst, reg, x = *loc;
4448 bool result = false;
4449 enum rtx_code code = GET_CODE (x);
4450 const char *fmt;
4451 int i, j;
4453 if (code == SUBREG)
4455 reg = SUBREG_REG (x);
4456 if ((subst = get_equiv_with_elimination (reg, curr_insn)) != reg
4457 && GET_MODE (subst) == VOIDmode)
4459 /* We cannot reload debug location. Simplify subreg here
4460 while we know the inner mode. */
4461 *loc = simplify_gen_subreg (GET_MODE (x), subst,
4462 GET_MODE (reg), SUBREG_BYTE (x));
4463 return true;
4466 if (code == REG && (subst = get_equiv_with_elimination (x, curr_insn)) != x)
4468 *loc = subst;
4469 return true;
4472 /* Scan all the operand sub-expressions. */
4473 fmt = GET_RTX_FORMAT (code);
4474 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4476 if (fmt[i] == 'e')
4477 result = loc_equivalence_change_p (&XEXP (x, i)) || result;
4478 else if (fmt[i] == 'E')
4479 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4480 result
4481 = loc_equivalence_change_p (&XVECEXP (x, i, j)) || result;
4483 return result;
4486 /* Similar to loc_equivalence_change_p, but for use as
4487 simplify_replace_fn_rtx callback. DATA is insn for which the
4488 elimination is done. If it null we don't do the elimination. */
4489 static rtx
4490 loc_equivalence_callback (rtx loc, const_rtx, void *data)
4492 if (!REG_P (loc))
4493 return NULL_RTX;
4495 rtx subst = (data == NULL
4496 ? get_equiv (loc) : get_equiv_with_elimination (loc, (rtx_insn *) data));
4497 if (subst != loc)
4498 return subst;
4500 return NULL_RTX;
4503 /* Maximum number of generated reload insns per an insn. It is for
4504 preventing this pass cycling in a bug case. */
4505 #define MAX_RELOAD_INSNS_NUMBER LRA_MAX_INSN_RELOADS
4507 /* The current iteration number of this LRA pass. */
4508 int lra_constraint_iter;
4510 /* True if we substituted equiv which needs checking register
4511 allocation correctness because the equivalent value contains
4512 allocatable hard registers or when we restore multi-register
4513 pseudo. */
4514 bool lra_risky_transformations_p;
4516 /* Return true if REGNO is referenced in more than one block. */
4517 static bool
4518 multi_block_pseudo_p (int regno)
4520 basic_block bb = NULL;
4521 unsigned int uid;
4522 bitmap_iterator bi;
4524 if (regno < FIRST_PSEUDO_REGISTER)
4525 return false;
4527 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi)
4528 if (bb == NULL)
4529 bb = BLOCK_FOR_INSN (lra_insn_recog_data[uid]->insn);
4530 else if (BLOCK_FOR_INSN (lra_insn_recog_data[uid]->insn) != bb)
4531 return true;
4532 return false;
4535 /* Return true if LIST contains a deleted insn. */
4536 static bool
4537 contains_deleted_insn_p (rtx_insn_list *list)
4539 for (; list != NULL_RTX; list = list->next ())
4540 if (NOTE_P (list->insn ())
4541 && NOTE_KIND (list->insn ()) == NOTE_INSN_DELETED)
4542 return true;
4543 return false;
4546 /* Return true if X contains a pseudo dying in INSN. */
4547 static bool
4548 dead_pseudo_p (rtx x, rtx_insn *insn)
4550 int i, j;
4551 const char *fmt;
4552 enum rtx_code code;
4554 if (REG_P (x))
4555 return (insn != NULL_RTX
4556 && find_regno_note (insn, REG_DEAD, REGNO (x)) != NULL_RTX);
4557 code = GET_CODE (x);
4558 fmt = GET_RTX_FORMAT (code);
4559 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4561 if (fmt[i] == 'e')
4563 if (dead_pseudo_p (XEXP (x, i), insn))
4564 return true;
4566 else if (fmt[i] == 'E')
4568 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4569 if (dead_pseudo_p (XVECEXP (x, i, j), insn))
4570 return true;
4573 return false;
4576 /* Return true if INSN contains a dying pseudo in INSN right hand
4577 side. */
4578 static bool
4579 insn_rhs_dead_pseudo_p (rtx_insn *insn)
4581 rtx set = single_set (insn);
4583 gcc_assert (set != NULL);
4584 return dead_pseudo_p (SET_SRC (set), insn);
4587 /* Return true if any init insn of REGNO contains a dying pseudo in
4588 insn right hand side. */
4589 static bool
4590 init_insn_rhs_dead_pseudo_p (int regno)
4592 rtx_insn_list *insns = ira_reg_equiv[regno].init_insns;
4594 if (insns == NULL)
4595 return false;
4596 for (; insns != NULL_RTX; insns = insns->next ())
4597 if (insn_rhs_dead_pseudo_p (insns->insn ()))
4598 return true;
4599 return false;
4602 /* Return TRUE if REGNO has a reverse equivalence. The equivalence is
4603 reverse only if we have one init insn with given REGNO as a
4604 source. */
4605 static bool
4606 reverse_equiv_p (int regno)
4608 rtx_insn_list *insns = ira_reg_equiv[regno].init_insns;
4609 rtx set;
4611 if (insns == NULL)
4612 return false;
4613 if (! INSN_P (insns->insn ())
4614 || insns->next () != NULL)
4615 return false;
4616 if ((set = single_set (insns->insn ())) == NULL_RTX)
4617 return false;
4618 return REG_P (SET_SRC (set)) && (int) REGNO (SET_SRC (set)) == regno;
4621 /* Return TRUE if REGNO was reloaded in an equivalence init insn. We
4622 call this function only for non-reverse equivalence. */
4623 static bool
4624 contains_reloaded_insn_p (int regno)
4626 rtx set;
4627 rtx_insn_list *list = ira_reg_equiv[regno].init_insns;
4629 for (; list != NULL; list = list->next ())
4630 if ((set = single_set (list->insn ())) == NULL_RTX
4631 || ! REG_P (SET_DEST (set))
4632 || (int) REGNO (SET_DEST (set)) != regno)
4633 return true;
4634 return false;
4637 /* Entry function of LRA constraint pass. Return true if the
4638 constraint pass did change the code. */
4639 bool
4640 lra_constraints (bool first_p)
4642 bool changed_p;
4643 int i, hard_regno, new_insns_num;
4644 unsigned int min_len, new_min_len, uid;
4645 rtx set, x, reg, dest_reg;
4646 basic_block last_bb;
4647 bitmap_head equiv_insn_bitmap;
4648 bitmap_iterator bi;
4650 lra_constraint_iter++;
4651 if (lra_dump_file != NULL)
4652 fprintf (lra_dump_file, "\n********** Local #%d: **********\n\n",
4653 lra_constraint_iter);
4654 changed_p = false;
4655 if (pic_offset_table_rtx
4656 && REGNO (pic_offset_table_rtx) >= FIRST_PSEUDO_REGISTER)
4657 lra_risky_transformations_p = true;
4658 else
4659 /* On the first iteration we should check IRA assignment
4660 correctness. In rare cases, the assignments can be wrong as
4661 early clobbers operands are ignored in IRA. */
4662 lra_risky_transformations_p = first_p;
4663 new_insn_uid_start = get_max_uid ();
4664 new_regno_start = first_p ? lra_constraint_new_regno_start : max_reg_num ();
4665 /* Mark used hard regs for target stack size calulations. */
4666 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4667 if (lra_reg_info[i].nrefs != 0
4668 && (hard_regno = lra_get_regno_hard_regno (i)) >= 0)
4670 int j, nregs;
4672 nregs = hard_regno_nregs[hard_regno][lra_reg_info[i].biggest_mode];
4673 for (j = 0; j < nregs; j++)
4674 df_set_regs_ever_live (hard_regno + j, true);
4676 /* Do elimination before the equivalence processing as we can spill
4677 some pseudos during elimination. */
4678 lra_eliminate (false, first_p);
4679 bitmap_initialize (&equiv_insn_bitmap, &reg_obstack);
4680 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4681 if (lra_reg_info[i].nrefs != 0)
4683 ira_reg_equiv[i].profitable_p = true;
4684 reg = regno_reg_rtx[i];
4685 if (lra_get_regno_hard_regno (i) < 0 && (x = get_equiv (reg)) != reg)
4687 bool pseudo_p = contains_reg_p (x, false, false);
4689 /* After RTL transformation, we can not guarantee that
4690 pseudo in the substitution was not reloaded which might
4691 make equivalence invalid. For example, in reverse
4692 equiv of p0
4694 p0 <- ...
4696 equiv_mem <- p0
4698 the memory address register was reloaded before the 2nd
4699 insn. */
4700 if ((! first_p && pseudo_p)
4701 /* We don't use DF for compilation speed sake. So it
4702 is problematic to update live info when we use an
4703 equivalence containing pseudos in more than one
4704 BB. */
4705 || (pseudo_p && multi_block_pseudo_p (i))
4706 /* If an init insn was deleted for some reason, cancel
4707 the equiv. We could update the equiv insns after
4708 transformations including an equiv insn deletion
4709 but it is not worthy as such cases are extremely
4710 rare. */
4711 || contains_deleted_insn_p (ira_reg_equiv[i].init_insns)
4712 /* If it is not a reverse equivalence, we check that a
4713 pseudo in rhs of the init insn is not dying in the
4714 insn. Otherwise, the live info at the beginning of
4715 the corresponding BB might be wrong after we
4716 removed the insn. When the equiv can be a
4717 constant, the right hand side of the init insn can
4718 be a pseudo. */
4719 || (! reverse_equiv_p (i)
4720 && (init_insn_rhs_dead_pseudo_p (i)
4721 /* If we reloaded the pseudo in an equivalence
4722 init insn, we can not remove the equiv init
4723 insns and the init insns might write into
4724 const memory in this case. */
4725 || contains_reloaded_insn_p (i)))
4726 /* Prevent access beyond equivalent memory for
4727 paradoxical subregs. */
4728 || (MEM_P (x)
4729 && (GET_MODE_SIZE (lra_reg_info[i].biggest_mode)
4730 > GET_MODE_SIZE (GET_MODE (x))))
4731 || (pic_offset_table_rtx
4732 && ((CONST_POOL_OK_P (PSEUDO_REGNO_MODE (i), x)
4733 && (targetm.preferred_reload_class
4734 (x, lra_get_allocno_class (i)) == NO_REGS))
4735 || contains_symbol_ref_p (x))))
4736 ira_reg_equiv[i].defined_p = false;
4737 if (contains_reg_p (x, false, true))
4738 ira_reg_equiv[i].profitable_p = false;
4739 if (get_equiv (reg) != reg)
4740 bitmap_ior_into (&equiv_insn_bitmap, &lra_reg_info[i].insn_bitmap);
4743 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4744 update_equiv (i);
4745 /* We should add all insns containing pseudos which should be
4746 substituted by their equivalences. */
4747 EXECUTE_IF_SET_IN_BITMAP (&equiv_insn_bitmap, 0, uid, bi)
4748 lra_push_insn_by_uid (uid);
4749 min_len = lra_insn_stack_length ();
4750 new_insns_num = 0;
4751 last_bb = NULL;
4752 changed_p = false;
4753 while ((new_min_len = lra_insn_stack_length ()) != 0)
4755 curr_insn = lra_pop_insn ();
4756 --new_min_len;
4757 curr_bb = BLOCK_FOR_INSN (curr_insn);
4758 if (curr_bb != last_bb)
4760 last_bb = curr_bb;
4761 bb_reload_num = lra_curr_reload_num;
4763 if (min_len > new_min_len)
4765 min_len = new_min_len;
4766 new_insns_num = 0;
4768 if (new_insns_num > MAX_RELOAD_INSNS_NUMBER)
4769 internal_error
4770 ("Max. number of generated reload insns per insn is achieved (%d)\n",
4771 MAX_RELOAD_INSNS_NUMBER);
4772 new_insns_num++;
4773 if (DEBUG_INSN_P (curr_insn))
4775 /* We need to check equivalence in debug insn and change
4776 pseudo to the equivalent value if necessary. */
4777 curr_id = lra_get_insn_recog_data (curr_insn);
4778 if (bitmap_bit_p (&equiv_insn_bitmap, INSN_UID (curr_insn)))
4780 rtx old = *curr_id->operand_loc[0];
4781 *curr_id->operand_loc[0]
4782 = simplify_replace_fn_rtx (old, NULL_RTX,
4783 loc_equivalence_callback, curr_insn);
4784 if (old != *curr_id->operand_loc[0])
4786 lra_update_insn_regno_info (curr_insn);
4787 changed_p = true;
4791 else if (INSN_P (curr_insn))
4793 if ((set = single_set (curr_insn)) != NULL_RTX)
4795 dest_reg = SET_DEST (set);
4796 /* The equivalence pseudo could be set up as SUBREG in a
4797 case when it is a call restore insn in a mode
4798 different from the pseudo mode. */
4799 if (GET_CODE (dest_reg) == SUBREG)
4800 dest_reg = SUBREG_REG (dest_reg);
4801 if ((REG_P (dest_reg)
4802 && (x = get_equiv (dest_reg)) != dest_reg
4803 /* Remove insns which set up a pseudo whose value
4804 can not be changed. Such insns might be not in
4805 init_insns because we don't update equiv data
4806 during insn transformations.
4808 As an example, let suppose that a pseudo got
4809 hard register and on the 1st pass was not
4810 changed to equivalent constant. We generate an
4811 additional insn setting up the pseudo because of
4812 secondary memory movement. Then the pseudo is
4813 spilled and we use the equiv constant. In this
4814 case we should remove the additional insn and
4815 this insn is not init_insns list. */
4816 && (! MEM_P (x) || MEM_READONLY_P (x)
4817 /* Check that this is actually an insn setting
4818 up the equivalence. */
4819 || in_list_p (curr_insn,
4820 ira_reg_equiv
4821 [REGNO (dest_reg)].init_insns)))
4822 || (((x = get_equiv (SET_SRC (set))) != SET_SRC (set))
4823 && in_list_p (curr_insn,
4824 ira_reg_equiv
4825 [REGNO (SET_SRC (set))].init_insns)))
4827 /* This is equiv init insn of pseudo which did not get a
4828 hard register -- remove the insn. */
4829 if (lra_dump_file != NULL)
4831 fprintf (lra_dump_file,
4832 " Removing equiv init insn %i (freq=%d)\n",
4833 INSN_UID (curr_insn),
4834 REG_FREQ_FROM_BB (BLOCK_FOR_INSN (curr_insn)));
4835 dump_insn_slim (lra_dump_file, curr_insn);
4837 if (contains_reg_p (x, true, false))
4838 lra_risky_transformations_p = true;
4839 lra_set_insn_deleted (curr_insn);
4840 continue;
4843 curr_id = lra_get_insn_recog_data (curr_insn);
4844 curr_static_id = curr_id->insn_static_data;
4845 init_curr_insn_input_reloads ();
4846 init_curr_operand_mode ();
4847 if (curr_insn_transform (false))
4848 changed_p = true;
4849 /* Check non-transformed insns too for equiv change as USE
4850 or CLOBBER don't need reloads but can contain pseudos
4851 being changed on their equivalences. */
4852 else if (bitmap_bit_p (&equiv_insn_bitmap, INSN_UID (curr_insn))
4853 && loc_equivalence_change_p (&PATTERN (curr_insn)))
4855 lra_update_insn_regno_info (curr_insn);
4856 changed_p = true;
4860 bitmap_clear (&equiv_insn_bitmap);
4861 /* If we used a new hard regno, changed_p should be true because the
4862 hard reg is assigned to a new pseudo. */
4863 if (flag_checking && !changed_p)
4865 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4866 if (lra_reg_info[i].nrefs != 0
4867 && (hard_regno = lra_get_regno_hard_regno (i)) >= 0)
4869 int j, nregs = hard_regno_nregs[hard_regno][PSEUDO_REGNO_MODE (i)];
4871 for (j = 0; j < nregs; j++)
4872 lra_assert (df_regs_ever_live_p (hard_regno + j));
4875 return changed_p;
4878 static void initiate_invariants (void);
4879 static void finish_invariants (void);
4881 /* Initiate the LRA constraint pass. It is done once per
4882 function. */
4883 void
4884 lra_constraints_init (void)
4886 initiate_invariants ();
4889 /* Finalize the LRA constraint pass. It is done once per
4890 function. */
4891 void
4892 lra_constraints_finish (void)
4894 finish_invariants ();
4899 /* Structure describes invariants for ineheritance. */
4900 struct lra_invariant
4902 /* The order number of the invariant. */
4903 int num;
4904 /* The invariant RTX. */
4905 rtx invariant_rtx;
4906 /* The origin insn of the invariant. */
4907 rtx_insn *insn;
4910 typedef lra_invariant invariant_t;
4911 typedef invariant_t *invariant_ptr_t;
4912 typedef const invariant_t *const_invariant_ptr_t;
4914 /* Pointer to the inheritance invariants. */
4915 static vec<invariant_ptr_t> invariants;
4917 /* Allocation pool for the invariants. */
4918 static object_allocator<lra_invariant> *invariants_pool;
4920 /* Hash table for the invariants. */
4921 static htab_t invariant_table;
4923 /* Hash function for INVARIANT. */
4924 static hashval_t
4925 invariant_hash (const void *invariant)
4927 rtx inv = ((const_invariant_ptr_t) invariant)->invariant_rtx;
4928 return lra_rtx_hash (inv);
4931 /* Equal function for invariants INVARIANT1 and INVARIANT2. */
4932 static int
4933 invariant_eq_p (const void *invariant1, const void *invariant2)
4935 rtx inv1 = ((const_invariant_ptr_t) invariant1)->invariant_rtx;
4936 rtx inv2 = ((const_invariant_ptr_t) invariant2)->invariant_rtx;
4938 return rtx_equal_p (inv1, inv2);
4941 /* Insert INVARIANT_RTX into the table if it is not there yet. Return
4942 invariant which is in the table. */
4943 static invariant_ptr_t
4944 insert_invariant (rtx invariant_rtx)
4946 void **entry_ptr;
4947 invariant_t invariant;
4948 invariant_ptr_t invariant_ptr;
4950 invariant.invariant_rtx = invariant_rtx;
4951 entry_ptr = htab_find_slot (invariant_table, &invariant, INSERT);
4952 if (*entry_ptr == NULL)
4954 invariant_ptr = invariants_pool->allocate ();
4955 invariant_ptr->invariant_rtx = invariant_rtx;
4956 invariant_ptr->insn = NULL;
4957 invariants.safe_push (invariant_ptr);
4958 *entry_ptr = (void *) invariant_ptr;
4960 return (invariant_ptr_t) *entry_ptr;
4963 /* Initiate the invariant table. */
4964 static void
4965 initiate_invariants (void)
4967 invariants.create (100);
4968 invariants_pool
4969 = new object_allocator<lra_invariant> ("Inheritance invariants");
4970 invariant_table = htab_create (100, invariant_hash, invariant_eq_p, NULL);
4973 /* Finish the invariant table. */
4974 static void
4975 finish_invariants (void)
4977 htab_delete (invariant_table);
4978 delete invariants_pool;
4979 invariants.release ();
4982 /* Make the invariant table empty. */
4983 static void
4984 clear_invariants (void)
4986 htab_empty (invariant_table);
4987 invariants_pool->release ();
4988 invariants.truncate (0);
4993 /* This page contains code to do inheritance/split
4994 transformations. */
4996 /* Number of reloads passed so far in current EBB. */
4997 static int reloads_num;
4999 /* Number of calls passed so far in current EBB. */
5000 static int calls_num;
5002 /* Current reload pseudo check for validity of elements in
5003 USAGE_INSNS. */
5004 static int curr_usage_insns_check;
5006 /* Info about last usage of registers in EBB to do inheritance/split
5007 transformation. Inheritance transformation is done from a spilled
5008 pseudo and split transformations from a hard register or a pseudo
5009 assigned to a hard register. */
5010 struct usage_insns
5012 /* If the value is equal to CURR_USAGE_INSNS_CHECK, then the member
5013 value INSNS is valid. The insns is chain of optional debug insns
5014 and a finishing non-debug insn using the corresponding reg. The
5015 value is also used to mark the registers which are set up in the
5016 current insn. The negated insn uid is used for this. */
5017 int check;
5018 /* Value of global reloads_num at the last insn in INSNS. */
5019 int reloads_num;
5020 /* Value of global reloads_nums at the last insn in INSNS. */
5021 int calls_num;
5022 /* It can be true only for splitting. And it means that the restore
5023 insn should be put after insn given by the following member. */
5024 bool after_p;
5025 /* Next insns in the current EBB which use the original reg and the
5026 original reg value is not changed between the current insn and
5027 the next insns. In order words, e.g. for inheritance, if we need
5028 to use the original reg value again in the next insns we can try
5029 to use the value in a hard register from a reload insn of the
5030 current insn. */
5031 rtx insns;
5034 /* Map: regno -> corresponding pseudo usage insns. */
5035 static struct usage_insns *usage_insns;
5037 static void
5038 setup_next_usage_insn (int regno, rtx insn, int reloads_num, bool after_p)
5040 usage_insns[regno].check = curr_usage_insns_check;
5041 usage_insns[regno].insns = insn;
5042 usage_insns[regno].reloads_num = reloads_num;
5043 usage_insns[regno].calls_num = calls_num;
5044 usage_insns[regno].after_p = after_p;
5047 /* The function is used to form list REGNO usages which consists of
5048 optional debug insns finished by a non-debug insn using REGNO.
5049 RELOADS_NUM is current number of reload insns processed so far. */
5050 static void
5051 add_next_usage_insn (int regno, rtx_insn *insn, int reloads_num)
5053 rtx next_usage_insns;
5055 if (usage_insns[regno].check == curr_usage_insns_check
5056 && (next_usage_insns = usage_insns[regno].insns) != NULL_RTX
5057 && DEBUG_INSN_P (insn))
5059 /* Check that we did not add the debug insn yet. */
5060 if (next_usage_insns != insn
5061 && (GET_CODE (next_usage_insns) != INSN_LIST
5062 || XEXP (next_usage_insns, 0) != insn))
5063 usage_insns[regno].insns = gen_rtx_INSN_LIST (VOIDmode, insn,
5064 next_usage_insns);
5066 else if (NONDEBUG_INSN_P (insn))
5067 setup_next_usage_insn (regno, insn, reloads_num, false);
5068 else
5069 usage_insns[regno].check = 0;
5072 /* Return first non-debug insn in list USAGE_INSNS. */
5073 static rtx_insn *
5074 skip_usage_debug_insns (rtx usage_insns)
5076 rtx insn;
5078 /* Skip debug insns. */
5079 for (insn = usage_insns;
5080 insn != NULL_RTX && GET_CODE (insn) == INSN_LIST;
5081 insn = XEXP (insn, 1))
5083 return safe_as_a <rtx_insn *> (insn);
5086 /* Return true if we need secondary memory moves for insn in
5087 USAGE_INSNS after inserting inherited pseudo of class INHER_CL
5088 into the insn. */
5089 static bool
5090 check_secondary_memory_needed_p (enum reg_class inher_cl ATTRIBUTE_UNUSED,
5091 rtx usage_insns ATTRIBUTE_UNUSED)
5093 #ifndef SECONDARY_MEMORY_NEEDED
5094 return false;
5095 #else
5096 rtx_insn *insn;
5097 rtx set, dest;
5098 enum reg_class cl;
5100 if (inher_cl == ALL_REGS
5101 || (insn = skip_usage_debug_insns (usage_insns)) == NULL_RTX)
5102 return false;
5103 lra_assert (INSN_P (insn));
5104 if ((set = single_set (insn)) == NULL_RTX || ! REG_P (SET_DEST (set)))
5105 return false;
5106 dest = SET_DEST (set);
5107 if (! REG_P (dest))
5108 return false;
5109 lra_assert (inher_cl != NO_REGS);
5110 cl = get_reg_class (REGNO (dest));
5111 return (cl != NO_REGS && cl != ALL_REGS
5112 && SECONDARY_MEMORY_NEEDED (inher_cl, cl, GET_MODE (dest)));
5113 #endif
5116 /* Registers involved in inheritance/split in the current EBB
5117 (inheritance/split pseudos and original registers). */
5118 static bitmap_head check_only_regs;
5120 /* Reload pseudos can not be involded in invariant inheritance in the
5121 current EBB. */
5122 static bitmap_head invalid_invariant_regs;
5124 /* Do inheritance transformations for insn INSN, which defines (if
5125 DEF_P) or uses ORIGINAL_REGNO. NEXT_USAGE_INSNS specifies which
5126 instruction in the EBB next uses ORIGINAL_REGNO; it has the same
5127 form as the "insns" field of usage_insns. Return true if we
5128 succeed in such transformation.
5130 The transformations look like:
5132 p <- ... i <- ...
5133 ... p <- i (new insn)
5134 ... =>
5135 <- ... p ... <- ... i ...
5137 ... i <- p (new insn)
5138 <- ... p ... <- ... i ...
5139 ... =>
5140 <- ... p ... <- ... i ...
5141 where p is a spilled original pseudo and i is a new inheritance pseudo.
5144 The inheritance pseudo has the smallest class of two classes CL and
5145 class of ORIGINAL REGNO. */
5146 static bool
5147 inherit_reload_reg (bool def_p, int original_regno,
5148 enum reg_class cl, rtx_insn *insn, rtx next_usage_insns)
5150 if (optimize_function_for_size_p (cfun))
5151 return false;
5153 enum reg_class rclass = lra_get_allocno_class (original_regno);
5154 rtx original_reg = regno_reg_rtx[original_regno];
5155 rtx new_reg, usage_insn;
5156 rtx_insn *new_insns;
5158 lra_assert (! usage_insns[original_regno].after_p);
5159 if (lra_dump_file != NULL)
5160 fprintf (lra_dump_file,
5161 " <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<\n");
5162 if (! ira_reg_classes_intersect_p[cl][rclass])
5164 if (lra_dump_file != NULL)
5166 fprintf (lra_dump_file,
5167 " Rejecting inheritance for %d "
5168 "because of disjoint classes %s and %s\n",
5169 original_regno, reg_class_names[cl],
5170 reg_class_names[rclass]);
5171 fprintf (lra_dump_file,
5172 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5174 return false;
5176 if ((ira_class_subset_p[cl][rclass] && cl != rclass)
5177 /* We don't use a subset of two classes because it can be
5178 NO_REGS. This transformation is still profitable in most
5179 cases even if the classes are not intersected as register
5180 move is probably cheaper than a memory load. */
5181 || ira_class_hard_regs_num[cl] < ira_class_hard_regs_num[rclass])
5183 if (lra_dump_file != NULL)
5184 fprintf (lra_dump_file, " Use smallest class of %s and %s\n",
5185 reg_class_names[cl], reg_class_names[rclass]);
5187 rclass = cl;
5189 if (check_secondary_memory_needed_p (rclass, next_usage_insns))
5191 /* Reject inheritance resulting in secondary memory moves.
5192 Otherwise, there is a danger in LRA cycling. Also such
5193 transformation will be unprofitable. */
5194 if (lra_dump_file != NULL)
5196 rtx_insn *insn = skip_usage_debug_insns (next_usage_insns);
5197 rtx set = single_set (insn);
5199 lra_assert (set != NULL_RTX);
5201 rtx dest = SET_DEST (set);
5203 lra_assert (REG_P (dest));
5204 fprintf (lra_dump_file,
5205 " Rejecting inheritance for insn %d(%s)<-%d(%s) "
5206 "as secondary mem is needed\n",
5207 REGNO (dest), reg_class_names[get_reg_class (REGNO (dest))],
5208 original_regno, reg_class_names[rclass]);
5209 fprintf (lra_dump_file,
5210 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5212 return false;
5214 new_reg = lra_create_new_reg (GET_MODE (original_reg), original_reg,
5215 rclass, "inheritance");
5216 start_sequence ();
5217 if (def_p)
5218 lra_emit_move (original_reg, new_reg);
5219 else
5220 lra_emit_move (new_reg, original_reg);
5221 new_insns = get_insns ();
5222 end_sequence ();
5223 if (NEXT_INSN (new_insns) != NULL_RTX)
5225 if (lra_dump_file != NULL)
5227 fprintf (lra_dump_file,
5228 " Rejecting inheritance %d->%d "
5229 "as it results in 2 or more insns:\n",
5230 original_regno, REGNO (new_reg));
5231 dump_rtl_slim (lra_dump_file, new_insns, NULL, -1, 0);
5232 fprintf (lra_dump_file,
5233 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5235 return false;
5237 lra_substitute_pseudo_within_insn (insn, original_regno, new_reg, false);
5238 lra_update_insn_regno_info (insn);
5239 if (! def_p)
5240 /* We now have a new usage insn for original regno. */
5241 setup_next_usage_insn (original_regno, new_insns, reloads_num, false);
5242 if (lra_dump_file != NULL)
5243 fprintf (lra_dump_file, " Original reg change %d->%d (bb%d):\n",
5244 original_regno, REGNO (new_reg), BLOCK_FOR_INSN (insn)->index);
5245 lra_reg_info[REGNO (new_reg)].restore_rtx = regno_reg_rtx[original_regno];
5246 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
5247 bitmap_set_bit (&check_only_regs, original_regno);
5248 bitmap_set_bit (&lra_inheritance_pseudos, REGNO (new_reg));
5249 if (def_p)
5250 lra_process_new_insns (insn, NULL, new_insns,
5251 "Add original<-inheritance");
5252 else
5253 lra_process_new_insns (insn, new_insns, NULL,
5254 "Add inheritance<-original");
5255 while (next_usage_insns != NULL_RTX)
5257 if (GET_CODE (next_usage_insns) != INSN_LIST)
5259 usage_insn = next_usage_insns;
5260 lra_assert (NONDEBUG_INSN_P (usage_insn));
5261 next_usage_insns = NULL;
5263 else
5265 usage_insn = XEXP (next_usage_insns, 0);
5266 lra_assert (DEBUG_INSN_P (usage_insn));
5267 next_usage_insns = XEXP (next_usage_insns, 1);
5269 lra_substitute_pseudo (&usage_insn, original_regno, new_reg, false);
5270 lra_update_insn_regno_info (as_a <rtx_insn *> (usage_insn));
5271 if (lra_dump_file != NULL)
5273 fprintf (lra_dump_file,
5274 " Inheritance reuse change %d->%d (bb%d):\n",
5275 original_regno, REGNO (new_reg),
5276 BLOCK_FOR_INSN (usage_insn)->index);
5277 dump_insn_slim (lra_dump_file, as_a <rtx_insn *> (usage_insn));
5280 if (lra_dump_file != NULL)
5281 fprintf (lra_dump_file,
5282 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5283 return true;
5286 /* Return true if we need a caller save/restore for pseudo REGNO which
5287 was assigned to a hard register. */
5288 static inline bool
5289 need_for_call_save_p (int regno)
5291 lra_assert (regno >= FIRST_PSEUDO_REGISTER && reg_renumber[regno] >= 0);
5292 return (usage_insns[regno].calls_num < calls_num
5293 && (overlaps_hard_reg_set_p
5294 ((flag_ipa_ra &&
5295 ! hard_reg_set_empty_p (lra_reg_info[regno].actual_call_used_reg_set))
5296 ? lra_reg_info[regno].actual_call_used_reg_set
5297 : call_used_reg_set,
5298 PSEUDO_REGNO_MODE (regno), reg_renumber[regno])
5299 || HARD_REGNO_CALL_PART_CLOBBERED (reg_renumber[regno],
5300 PSEUDO_REGNO_MODE (regno))));
5303 /* Global registers occurring in the current EBB. */
5304 static bitmap_head ebb_global_regs;
5306 /* Return true if we need a split for hard register REGNO or pseudo
5307 REGNO which was assigned to a hard register.
5308 POTENTIAL_RELOAD_HARD_REGS contains hard registers which might be
5309 used for reloads since the EBB end. It is an approximation of the
5310 used hard registers in the split range. The exact value would
5311 require expensive calculations. If we were aggressive with
5312 splitting because of the approximation, the split pseudo will save
5313 the same hard register assignment and will be removed in the undo
5314 pass. We still need the approximation because too aggressive
5315 splitting would result in too inaccurate cost calculation in the
5316 assignment pass because of too many generated moves which will be
5317 probably removed in the undo pass. */
5318 static inline bool
5319 need_for_split_p (HARD_REG_SET potential_reload_hard_regs, int regno)
5321 int hard_regno = regno < FIRST_PSEUDO_REGISTER ? regno : reg_renumber[regno];
5323 lra_assert (hard_regno >= 0);
5324 return ((TEST_HARD_REG_BIT (potential_reload_hard_regs, hard_regno)
5325 /* Don't split eliminable hard registers, otherwise we can
5326 split hard registers like hard frame pointer, which
5327 lives on BB start/end according to DF-infrastructure,
5328 when there is a pseudo assigned to the register and
5329 living in the same BB. */
5330 && (regno >= FIRST_PSEUDO_REGISTER
5331 || ! TEST_HARD_REG_BIT (eliminable_regset, hard_regno))
5332 && ! TEST_HARD_REG_BIT (lra_no_alloc_regs, hard_regno)
5333 /* Don't split call clobbered hard regs living through
5334 calls, otherwise we might have a check problem in the
5335 assign sub-pass as in the most cases (exception is a
5336 situation when lra_risky_transformations_p value is
5337 true) the assign pass assumes that all pseudos living
5338 through calls are assigned to call saved hard regs. */
5339 && (regno >= FIRST_PSEUDO_REGISTER
5340 || ! TEST_HARD_REG_BIT (call_used_reg_set, regno)
5341 || usage_insns[regno].calls_num == calls_num)
5342 /* We need at least 2 reloads to make pseudo splitting
5343 profitable. We should provide hard regno splitting in
5344 any case to solve 1st insn scheduling problem when
5345 moving hard register definition up might result in
5346 impossibility to find hard register for reload pseudo of
5347 small register class. */
5348 && (usage_insns[regno].reloads_num
5349 + (regno < FIRST_PSEUDO_REGISTER ? 0 : 3) < reloads_num)
5350 && (regno < FIRST_PSEUDO_REGISTER
5351 /* For short living pseudos, spilling + inheritance can
5352 be considered a substitution for splitting.
5353 Therefore we do not splitting for local pseudos. It
5354 decreases also aggressiveness of splitting. The
5355 minimal number of references is chosen taking into
5356 account that for 2 references splitting has no sense
5357 as we can just spill the pseudo. */
5358 || (regno >= FIRST_PSEUDO_REGISTER
5359 && lra_reg_info[regno].nrefs > 3
5360 && bitmap_bit_p (&ebb_global_regs, regno))))
5361 || (regno >= FIRST_PSEUDO_REGISTER && need_for_call_save_p (regno)));
5364 /* Return class for the split pseudo created from original pseudo with
5365 ALLOCNO_CLASS and MODE which got a hard register HARD_REGNO. We
5366 choose subclass of ALLOCNO_CLASS which contains HARD_REGNO and
5367 results in no secondary memory movements. */
5368 static enum reg_class
5369 choose_split_class (enum reg_class allocno_class,
5370 int hard_regno ATTRIBUTE_UNUSED,
5371 machine_mode mode ATTRIBUTE_UNUSED)
5373 #ifndef SECONDARY_MEMORY_NEEDED
5374 return allocno_class;
5375 #else
5376 int i;
5377 enum reg_class cl, best_cl = NO_REGS;
5378 enum reg_class hard_reg_class ATTRIBUTE_UNUSED
5379 = REGNO_REG_CLASS (hard_regno);
5381 if (! SECONDARY_MEMORY_NEEDED (allocno_class, allocno_class, mode)
5382 && TEST_HARD_REG_BIT (reg_class_contents[allocno_class], hard_regno))
5383 return allocno_class;
5384 for (i = 0;
5385 (cl = reg_class_subclasses[allocno_class][i]) != LIM_REG_CLASSES;
5386 i++)
5387 if (! SECONDARY_MEMORY_NEEDED (cl, hard_reg_class, mode)
5388 && ! SECONDARY_MEMORY_NEEDED (hard_reg_class, cl, mode)
5389 && TEST_HARD_REG_BIT (reg_class_contents[cl], hard_regno)
5390 && (best_cl == NO_REGS
5391 || ira_class_hard_regs_num[best_cl] < ira_class_hard_regs_num[cl]))
5392 best_cl = cl;
5393 return best_cl;
5394 #endif
5397 /* Do split transformations for insn INSN, which defines or uses
5398 ORIGINAL_REGNO. NEXT_USAGE_INSNS specifies which instruction in
5399 the EBB next uses ORIGINAL_REGNO; it has the same form as the
5400 "insns" field of usage_insns.
5402 The transformations look like:
5404 p <- ... p <- ...
5405 ... s <- p (new insn -- save)
5406 ... =>
5407 ... p <- s (new insn -- restore)
5408 <- ... p ... <- ... p ...
5410 <- ... p ... <- ... p ...
5411 ... s <- p (new insn -- save)
5412 ... =>
5413 ... p <- s (new insn -- restore)
5414 <- ... p ... <- ... p ...
5416 where p is an original pseudo got a hard register or a hard
5417 register and s is a new split pseudo. The save is put before INSN
5418 if BEFORE_P is true. Return true if we succeed in such
5419 transformation. */
5420 static bool
5421 split_reg (bool before_p, int original_regno, rtx_insn *insn,
5422 rtx next_usage_insns)
5424 enum reg_class rclass;
5425 rtx original_reg;
5426 int hard_regno, nregs;
5427 rtx new_reg, usage_insn;
5428 rtx_insn *restore, *save;
5429 bool after_p;
5430 bool call_save_p;
5431 machine_mode mode;
5433 if (original_regno < FIRST_PSEUDO_REGISTER)
5435 rclass = ira_allocno_class_translate[REGNO_REG_CLASS (original_regno)];
5436 hard_regno = original_regno;
5437 call_save_p = false;
5438 nregs = 1;
5439 mode = lra_reg_info[hard_regno].biggest_mode;
5440 machine_mode reg_rtx_mode = GET_MODE (regno_reg_rtx[hard_regno]);
5441 /* A reg can have a biggest_mode of VOIDmode if it was only ever seen
5442 as part of a multi-word register. In that case, or if the biggest
5443 mode was larger than a register, just use the reg_rtx. Otherwise,
5444 limit the size to that of the biggest access in the function. */
5445 if (mode == VOIDmode
5446 || GET_MODE_SIZE (mode) > GET_MODE_SIZE (reg_rtx_mode))
5448 original_reg = regno_reg_rtx[hard_regno];
5449 mode = reg_rtx_mode;
5451 else
5452 original_reg = gen_rtx_REG (mode, hard_regno);
5454 else
5456 mode = PSEUDO_REGNO_MODE (original_regno);
5457 hard_regno = reg_renumber[original_regno];
5458 nregs = hard_regno_nregs[hard_regno][mode];
5459 rclass = lra_get_allocno_class (original_regno);
5460 original_reg = regno_reg_rtx[original_regno];
5461 call_save_p = need_for_call_save_p (original_regno);
5463 lra_assert (hard_regno >= 0);
5464 if (lra_dump_file != NULL)
5465 fprintf (lra_dump_file,
5466 " ((((((((((((((((((((((((((((((((((((((((((((((((\n");
5468 if (call_save_p)
5470 mode = HARD_REGNO_CALLER_SAVE_MODE (hard_regno,
5471 hard_regno_nregs[hard_regno][mode],
5472 mode);
5473 new_reg = lra_create_new_reg (mode, NULL_RTX, NO_REGS, "save");
5475 else
5477 rclass = choose_split_class (rclass, hard_regno, mode);
5478 if (rclass == NO_REGS)
5480 if (lra_dump_file != NULL)
5482 fprintf (lra_dump_file,
5483 " Rejecting split of %d(%s): "
5484 "no good reg class for %d(%s)\n",
5485 original_regno,
5486 reg_class_names[lra_get_allocno_class (original_regno)],
5487 hard_regno,
5488 reg_class_names[REGNO_REG_CLASS (hard_regno)]);
5489 fprintf
5490 (lra_dump_file,
5491 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5493 return false;
5495 /* Split_if_necessary can split hard registers used as part of a
5496 multi-register mode but splits each register individually. The
5497 mode used for each independent register may not be supported
5498 so reject the split. Splitting the wider mode should theoretically
5499 be possible but is not implemented. */
5500 if (! HARD_REGNO_MODE_OK (hard_regno, mode))
5502 if (lra_dump_file != NULL)
5504 fprintf (lra_dump_file,
5505 " Rejecting split of %d(%s): unsuitable mode %s\n",
5506 original_regno,
5507 reg_class_names[lra_get_allocno_class (original_regno)],
5508 GET_MODE_NAME (mode));
5509 fprintf
5510 (lra_dump_file,
5511 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5513 return false;
5515 new_reg = lra_create_new_reg (mode, original_reg, rclass, "split");
5516 reg_renumber[REGNO (new_reg)] = hard_regno;
5518 save = emit_spill_move (true, new_reg, original_reg);
5519 if (NEXT_INSN (save) != NULL_RTX && !call_save_p)
5521 if (lra_dump_file != NULL)
5523 fprintf
5524 (lra_dump_file,
5525 " Rejecting split %d->%d resulting in > 2 save insns:\n",
5526 original_regno, REGNO (new_reg));
5527 dump_rtl_slim (lra_dump_file, save, NULL, -1, 0);
5528 fprintf (lra_dump_file,
5529 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5531 return false;
5533 restore = emit_spill_move (false, new_reg, original_reg);
5534 if (NEXT_INSN (restore) != NULL_RTX && !call_save_p)
5536 if (lra_dump_file != NULL)
5538 fprintf (lra_dump_file,
5539 " Rejecting split %d->%d "
5540 "resulting in > 2 restore insns:\n",
5541 original_regno, REGNO (new_reg));
5542 dump_rtl_slim (lra_dump_file, restore, NULL, -1, 0);
5543 fprintf (lra_dump_file,
5544 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5546 return false;
5548 after_p = usage_insns[original_regno].after_p;
5549 lra_reg_info[REGNO (new_reg)].restore_rtx = regno_reg_rtx[original_regno];
5550 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
5551 bitmap_set_bit (&check_only_regs, original_regno);
5552 bitmap_set_bit (&lra_split_regs, REGNO (new_reg));
5553 for (;;)
5555 if (GET_CODE (next_usage_insns) != INSN_LIST)
5557 usage_insn = next_usage_insns;
5558 break;
5560 usage_insn = XEXP (next_usage_insns, 0);
5561 lra_assert (DEBUG_INSN_P (usage_insn));
5562 next_usage_insns = XEXP (next_usage_insns, 1);
5563 lra_substitute_pseudo (&usage_insn, original_regno, new_reg, false);
5564 lra_update_insn_regno_info (as_a <rtx_insn *> (usage_insn));
5565 if (lra_dump_file != NULL)
5567 fprintf (lra_dump_file, " Split reuse change %d->%d:\n",
5568 original_regno, REGNO (new_reg));
5569 dump_insn_slim (lra_dump_file, as_a <rtx_insn *> (usage_insn));
5572 lra_assert (NOTE_P (usage_insn) || NONDEBUG_INSN_P (usage_insn));
5573 lra_assert (usage_insn != insn || (after_p && before_p));
5574 lra_process_new_insns (as_a <rtx_insn *> (usage_insn),
5575 after_p ? NULL : restore,
5576 after_p ? restore : NULL,
5577 call_save_p
5578 ? "Add reg<-save" : "Add reg<-split");
5579 lra_process_new_insns (insn, before_p ? save : NULL,
5580 before_p ? NULL : save,
5581 call_save_p
5582 ? "Add save<-reg" : "Add split<-reg");
5583 if (nregs > 1)
5584 /* If we are trying to split multi-register. We should check
5585 conflicts on the next assignment sub-pass. IRA can allocate on
5586 sub-register levels, LRA do this on pseudos level right now and
5587 this discrepancy may create allocation conflicts after
5588 splitting. */
5589 lra_risky_transformations_p = true;
5590 if (lra_dump_file != NULL)
5591 fprintf (lra_dump_file,
5592 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5593 return true;
5596 /* Recognize that we need a split transformation for insn INSN, which
5597 defines or uses REGNO in its insn biggest MODE (we use it only if
5598 REGNO is a hard register). POTENTIAL_RELOAD_HARD_REGS contains
5599 hard registers which might be used for reloads since the EBB end.
5600 Put the save before INSN if BEFORE_P is true. MAX_UID is maximla
5601 uid before starting INSN processing. Return true if we succeed in
5602 such transformation. */
5603 static bool
5604 split_if_necessary (int regno, machine_mode mode,
5605 HARD_REG_SET potential_reload_hard_regs,
5606 bool before_p, rtx_insn *insn, int max_uid)
5608 bool res = false;
5609 int i, nregs = 1;
5610 rtx next_usage_insns;
5612 if (regno < FIRST_PSEUDO_REGISTER)
5613 nregs = hard_regno_nregs[regno][mode];
5614 for (i = 0; i < nregs; i++)
5615 if (usage_insns[regno + i].check == curr_usage_insns_check
5616 && (next_usage_insns = usage_insns[regno + i].insns) != NULL_RTX
5617 /* To avoid processing the register twice or more. */
5618 && ((GET_CODE (next_usage_insns) != INSN_LIST
5619 && INSN_UID (next_usage_insns) < max_uid)
5620 || (GET_CODE (next_usage_insns) == INSN_LIST
5621 && (INSN_UID (XEXP (next_usage_insns, 0)) < max_uid)))
5622 && need_for_split_p (potential_reload_hard_regs, regno + i)
5623 && split_reg (before_p, regno + i, insn, next_usage_insns))
5624 res = true;
5625 return res;
5628 /* Return TRUE if rtx X is considered as an invariant for
5629 inheritance. */
5630 static bool
5631 invariant_p (const_rtx x)
5633 machine_mode mode;
5634 const char *fmt;
5635 enum rtx_code code;
5636 int i, j;
5638 code = GET_CODE (x);
5639 mode = GET_MODE (x);
5640 if (code == SUBREG)
5642 x = SUBREG_REG (x);
5643 code = GET_CODE (x);
5644 if (GET_MODE_SIZE (GET_MODE (x)) > GET_MODE_SIZE (mode))
5645 mode = GET_MODE (x);
5648 if (MEM_P (x))
5649 return false;
5651 if (REG_P (x))
5653 int i, nregs, regno = REGNO (x);
5655 if (regno >= FIRST_PSEUDO_REGISTER || regno == STACK_POINTER_REGNUM
5656 || TEST_HARD_REG_BIT (eliminable_regset, regno)
5657 || GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
5658 return false;
5659 nregs = hard_regno_nregs[regno][mode];
5660 for (i = 0; i < nregs; i++)
5661 if (! fixed_regs[regno + i]
5662 /* A hard register may be clobbered in the current insn
5663 but we can ignore this case because if the hard
5664 register is used it should be set somewhere after the
5665 clobber. */
5666 || bitmap_bit_p (&invalid_invariant_regs, regno + i))
5667 return false;
5669 fmt = GET_RTX_FORMAT (code);
5670 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
5672 if (fmt[i] == 'e')
5674 if (! invariant_p (XEXP (x, i)))
5675 return false;
5677 else if (fmt[i] == 'E')
5679 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
5680 if (! invariant_p (XVECEXP (x, i, j)))
5681 return false;
5684 return true;
5687 /* We have 'dest_reg <- invariant'. Let us try to make an invariant
5688 inheritance transformation (using dest_reg instead invariant in a
5689 subsequent insn). */
5690 static bool
5691 process_invariant_for_inheritance (rtx dst_reg, rtx invariant_rtx)
5693 invariant_ptr_t invariant_ptr;
5694 rtx_insn *insn, *new_insns;
5695 rtx insn_set, insn_reg, new_reg;
5696 int insn_regno;
5697 bool succ_p = false;
5698 int dst_regno = REGNO (dst_reg);
5699 enum machine_mode dst_mode = GET_MODE (dst_reg);
5700 enum reg_class cl = lra_get_allocno_class (dst_regno), insn_reg_cl;
5702 invariant_ptr = insert_invariant (invariant_rtx);
5703 if ((insn = invariant_ptr->insn) != NULL_RTX)
5705 /* We have a subsequent insn using the invariant. */
5706 insn_set = single_set (insn);
5707 lra_assert (insn_set != NULL);
5708 insn_reg = SET_DEST (insn_set);
5709 lra_assert (REG_P (insn_reg));
5710 insn_regno = REGNO (insn_reg);
5711 insn_reg_cl = lra_get_allocno_class (insn_regno);
5713 if (dst_mode == GET_MODE (insn_reg)
5714 /* We should consider only result move reg insns which are
5715 cheap. */
5716 && targetm.register_move_cost (dst_mode, cl, insn_reg_cl) == 2
5717 && targetm.register_move_cost (dst_mode, cl, cl) == 2)
5719 if (lra_dump_file != NULL)
5720 fprintf (lra_dump_file,
5721 " [[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[\n");
5722 new_reg = lra_create_new_reg (dst_mode, dst_reg,
5723 cl, "invariant inheritance");
5724 bitmap_set_bit (&lra_inheritance_pseudos, REGNO (new_reg));
5725 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
5726 lra_reg_info[REGNO (new_reg)].restore_rtx = PATTERN (insn);
5727 start_sequence ();
5728 lra_emit_move (new_reg, dst_reg);
5729 new_insns = get_insns ();
5730 end_sequence ();
5731 lra_process_new_insns (curr_insn, NULL, new_insns,
5732 "Add invariant inheritance<-original");
5733 start_sequence ();
5734 lra_emit_move (SET_DEST (insn_set), new_reg);
5735 new_insns = get_insns ();
5736 end_sequence ();
5737 lra_process_new_insns (insn, NULL, new_insns,
5738 "Changing reload<-inheritance");
5739 lra_set_insn_deleted (insn);
5740 succ_p = true;
5741 if (lra_dump_file != NULL)
5743 fprintf (lra_dump_file,
5744 " Invariant inheritance reuse change %d (bb%d):\n",
5745 REGNO (new_reg), BLOCK_FOR_INSN (insn)->index);
5746 dump_insn_slim (lra_dump_file, insn);
5747 fprintf (lra_dump_file,
5748 " ]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]\n");
5752 invariant_ptr->insn = curr_insn;
5753 return succ_p;
5756 /* Check only registers living at the current program point in the
5757 current EBB. */
5758 static bitmap_head live_regs;
5760 /* Update live info in EBB given by its HEAD and TAIL insns after
5761 inheritance/split transformation. The function removes dead moves
5762 too. */
5763 static void
5764 update_ebb_live_info (rtx_insn *head, rtx_insn *tail)
5766 unsigned int j;
5767 int i, regno;
5768 bool live_p;
5769 rtx_insn *prev_insn;
5770 rtx set;
5771 bool remove_p;
5772 basic_block last_bb, prev_bb, curr_bb;
5773 bitmap_iterator bi;
5774 struct lra_insn_reg *reg;
5775 edge e;
5776 edge_iterator ei;
5778 last_bb = BLOCK_FOR_INSN (tail);
5779 prev_bb = NULL;
5780 for (curr_insn = tail;
5781 curr_insn != PREV_INSN (head);
5782 curr_insn = prev_insn)
5784 prev_insn = PREV_INSN (curr_insn);
5785 /* We need to process empty blocks too. They contain
5786 NOTE_INSN_BASIC_BLOCK referring for the basic block. */
5787 if (NOTE_P (curr_insn) && NOTE_KIND (curr_insn) != NOTE_INSN_BASIC_BLOCK)
5788 continue;
5789 curr_bb = BLOCK_FOR_INSN (curr_insn);
5790 if (curr_bb != prev_bb)
5792 if (prev_bb != NULL)
5794 /* Update df_get_live_in (prev_bb): */
5795 EXECUTE_IF_SET_IN_BITMAP (&check_only_regs, 0, j, bi)
5796 if (bitmap_bit_p (&live_regs, j))
5797 bitmap_set_bit (df_get_live_in (prev_bb), j);
5798 else
5799 bitmap_clear_bit (df_get_live_in (prev_bb), j);
5801 if (curr_bb != last_bb)
5803 /* Update df_get_live_out (curr_bb): */
5804 EXECUTE_IF_SET_IN_BITMAP (&check_only_regs, 0, j, bi)
5806 live_p = bitmap_bit_p (&live_regs, j);
5807 if (! live_p)
5808 FOR_EACH_EDGE (e, ei, curr_bb->succs)
5809 if (bitmap_bit_p (df_get_live_in (e->dest), j))
5811 live_p = true;
5812 break;
5814 if (live_p)
5815 bitmap_set_bit (df_get_live_out (curr_bb), j);
5816 else
5817 bitmap_clear_bit (df_get_live_out (curr_bb), j);
5820 prev_bb = curr_bb;
5821 bitmap_and (&live_regs, &check_only_regs, df_get_live_out (curr_bb));
5823 if (! NONDEBUG_INSN_P (curr_insn))
5824 continue;
5825 curr_id = lra_get_insn_recog_data (curr_insn);
5826 curr_static_id = curr_id->insn_static_data;
5827 remove_p = false;
5828 if ((set = single_set (curr_insn)) != NULL_RTX
5829 && REG_P (SET_DEST (set))
5830 && (regno = REGNO (SET_DEST (set))) >= FIRST_PSEUDO_REGISTER
5831 && SET_DEST (set) != pic_offset_table_rtx
5832 && bitmap_bit_p (&check_only_regs, regno)
5833 && ! bitmap_bit_p (&live_regs, regno))
5834 remove_p = true;
5835 /* See which defined values die here. */
5836 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5837 if (reg->type == OP_OUT && ! reg->subreg_p)
5838 bitmap_clear_bit (&live_regs, reg->regno);
5839 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
5840 if (reg->type == OP_OUT && ! reg->subreg_p)
5841 bitmap_clear_bit (&live_regs, reg->regno);
5842 if (curr_id->arg_hard_regs != NULL)
5843 /* Make clobbered argument hard registers die. */
5844 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
5845 if (regno >= FIRST_PSEUDO_REGISTER)
5846 bitmap_clear_bit (&live_regs, regno - FIRST_PSEUDO_REGISTER);
5847 /* Mark each used value as live. */
5848 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5849 if (reg->type != OP_OUT
5850 && bitmap_bit_p (&check_only_regs, reg->regno))
5851 bitmap_set_bit (&live_regs, reg->regno);
5852 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
5853 if (reg->type != OP_OUT
5854 && bitmap_bit_p (&check_only_regs, reg->regno))
5855 bitmap_set_bit (&live_regs, reg->regno);
5856 if (curr_id->arg_hard_regs != NULL)
5857 /* Make used argument hard registers live. */
5858 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
5859 if (regno < FIRST_PSEUDO_REGISTER
5860 && bitmap_bit_p (&check_only_regs, regno))
5861 bitmap_set_bit (&live_regs, regno);
5862 /* It is quite important to remove dead move insns because it
5863 means removing dead store. We don't need to process them for
5864 constraints. */
5865 if (remove_p)
5867 if (lra_dump_file != NULL)
5869 fprintf (lra_dump_file, " Removing dead insn:\n ");
5870 dump_insn_slim (lra_dump_file, curr_insn);
5872 lra_set_insn_deleted (curr_insn);
5877 /* The structure describes info to do an inheritance for the current
5878 insn. We need to collect such info first before doing the
5879 transformations because the transformations change the insn
5880 internal representation. */
5881 struct to_inherit
5883 /* Original regno. */
5884 int regno;
5885 /* Subsequent insns which can inherit original reg value. */
5886 rtx insns;
5889 /* Array containing all info for doing inheritance from the current
5890 insn. */
5891 static struct to_inherit to_inherit[LRA_MAX_INSN_RELOADS];
5893 /* Number elements in the previous array. */
5894 static int to_inherit_num;
5896 /* Add inheritance info REGNO and INSNS. Their meaning is described in
5897 structure to_inherit. */
5898 static void
5899 add_to_inherit (int regno, rtx insns)
5901 int i;
5903 for (i = 0; i < to_inherit_num; i++)
5904 if (to_inherit[i].regno == regno)
5905 return;
5906 lra_assert (to_inherit_num < LRA_MAX_INSN_RELOADS);
5907 to_inherit[to_inherit_num].regno = regno;
5908 to_inherit[to_inherit_num++].insns = insns;
5911 /* Return the last non-debug insn in basic block BB, or the block begin
5912 note if none. */
5913 static rtx_insn *
5914 get_last_insertion_point (basic_block bb)
5916 rtx_insn *insn;
5918 FOR_BB_INSNS_REVERSE (bb, insn)
5919 if (NONDEBUG_INSN_P (insn) || NOTE_INSN_BASIC_BLOCK_P (insn))
5920 return insn;
5921 gcc_unreachable ();
5924 /* Set up RES by registers living on edges FROM except the edge (FROM,
5925 TO) or by registers set up in a jump insn in BB FROM. */
5926 static void
5927 get_live_on_other_edges (basic_block from, basic_block to, bitmap res)
5929 rtx_insn *last;
5930 struct lra_insn_reg *reg;
5931 edge e;
5932 edge_iterator ei;
5934 lra_assert (to != NULL);
5935 bitmap_clear (res);
5936 FOR_EACH_EDGE (e, ei, from->succs)
5937 if (e->dest != to)
5938 bitmap_ior_into (res, df_get_live_in (e->dest));
5939 last = get_last_insertion_point (from);
5940 if (! JUMP_P (last))
5941 return;
5942 curr_id = lra_get_insn_recog_data (last);
5943 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5944 if (reg->type != OP_IN)
5945 bitmap_set_bit (res, reg->regno);
5948 /* Used as a temporary results of some bitmap calculations. */
5949 static bitmap_head temp_bitmap;
5951 /* We split for reloads of small class of hard regs. The following
5952 defines how many hard regs the class should have to be qualified as
5953 small. The code is mostly oriented to x86/x86-64 architecture
5954 where some insns need to use only specific register or pair of
5955 registers and these register can live in RTL explicitly, e.g. for
5956 parameter passing. */
5957 static const int max_small_class_regs_num = 2;
5959 /* Do inheritance/split transformations in EBB starting with HEAD and
5960 finishing on TAIL. We process EBB insns in the reverse order.
5961 Return true if we did any inheritance/split transformation in the
5962 EBB.
5964 We should avoid excessive splitting which results in worse code
5965 because of inaccurate cost calculations for spilling new split
5966 pseudos in such case. To achieve this we do splitting only if
5967 register pressure is high in given basic block and there are reload
5968 pseudos requiring hard registers. We could do more register
5969 pressure calculations at any given program point to avoid necessary
5970 splitting even more but it is to expensive and the current approach
5971 works well enough. */
5972 static bool
5973 inherit_in_ebb (rtx_insn *head, rtx_insn *tail)
5975 int i, src_regno, dst_regno, nregs;
5976 bool change_p, succ_p, update_reloads_num_p;
5977 rtx_insn *prev_insn, *last_insn;
5978 rtx next_usage_insns, curr_set;
5979 enum reg_class cl;
5980 struct lra_insn_reg *reg;
5981 basic_block last_processed_bb, curr_bb = NULL;
5982 HARD_REG_SET potential_reload_hard_regs, live_hard_regs;
5983 bitmap to_process;
5984 unsigned int j;
5985 bitmap_iterator bi;
5986 bool head_p, after_p;
5988 change_p = false;
5989 curr_usage_insns_check++;
5990 clear_invariants ();
5991 reloads_num = calls_num = 0;
5992 bitmap_clear (&check_only_regs);
5993 bitmap_clear (&invalid_invariant_regs);
5994 last_processed_bb = NULL;
5995 CLEAR_HARD_REG_SET (potential_reload_hard_regs);
5996 COPY_HARD_REG_SET (live_hard_regs, eliminable_regset);
5997 IOR_HARD_REG_SET (live_hard_regs, lra_no_alloc_regs);
5998 /* We don't process new insns generated in the loop. */
5999 for (curr_insn = tail; curr_insn != PREV_INSN (head); curr_insn = prev_insn)
6001 prev_insn = PREV_INSN (curr_insn);
6002 if (BLOCK_FOR_INSN (curr_insn) != NULL)
6003 curr_bb = BLOCK_FOR_INSN (curr_insn);
6004 if (last_processed_bb != curr_bb)
6006 /* We are at the end of BB. Add qualified living
6007 pseudos for potential splitting. */
6008 to_process = df_get_live_out (curr_bb);
6009 if (last_processed_bb != NULL)
6011 /* We are somewhere in the middle of EBB. */
6012 get_live_on_other_edges (curr_bb, last_processed_bb,
6013 &temp_bitmap);
6014 to_process = &temp_bitmap;
6016 last_processed_bb = curr_bb;
6017 last_insn = get_last_insertion_point (curr_bb);
6018 after_p = (! JUMP_P (last_insn)
6019 && (! CALL_P (last_insn)
6020 || (find_reg_note (last_insn,
6021 REG_NORETURN, NULL_RTX) == NULL_RTX
6022 && ! SIBLING_CALL_P (last_insn))));
6023 CLEAR_HARD_REG_SET (potential_reload_hard_regs);
6024 EXECUTE_IF_SET_IN_BITMAP (to_process, 0, j, bi)
6026 if ((int) j >= lra_constraint_new_regno_start)
6027 break;
6028 if (j < FIRST_PSEUDO_REGISTER || reg_renumber[j] >= 0)
6030 if (j < FIRST_PSEUDO_REGISTER)
6031 SET_HARD_REG_BIT (live_hard_regs, j);
6032 else
6033 add_to_hard_reg_set (&live_hard_regs,
6034 PSEUDO_REGNO_MODE (j),
6035 reg_renumber[j]);
6036 setup_next_usage_insn (j, last_insn, reloads_num, after_p);
6040 src_regno = dst_regno = -1;
6041 curr_set = single_set (curr_insn);
6042 if (curr_set != NULL_RTX && REG_P (SET_DEST (curr_set)))
6043 dst_regno = REGNO (SET_DEST (curr_set));
6044 if (curr_set != NULL_RTX && REG_P (SET_SRC (curr_set)))
6045 src_regno = REGNO (SET_SRC (curr_set));
6046 update_reloads_num_p = true;
6047 if (src_regno < lra_constraint_new_regno_start
6048 && src_regno >= FIRST_PSEUDO_REGISTER
6049 && reg_renumber[src_regno] < 0
6050 && dst_regno >= lra_constraint_new_regno_start
6051 && (cl = lra_get_allocno_class (dst_regno)) != NO_REGS)
6053 /* 'reload_pseudo <- original_pseudo'. */
6054 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6055 reloads_num++;
6056 update_reloads_num_p = false;
6057 succ_p = false;
6058 if (usage_insns[src_regno].check == curr_usage_insns_check
6059 && (next_usage_insns = usage_insns[src_regno].insns) != NULL_RTX)
6060 succ_p = inherit_reload_reg (false, src_regno, cl,
6061 curr_insn, next_usage_insns);
6062 if (succ_p)
6063 change_p = true;
6064 else
6065 setup_next_usage_insn (src_regno, curr_insn, reloads_num, false);
6066 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6067 IOR_HARD_REG_SET (potential_reload_hard_regs,
6068 reg_class_contents[cl]);
6070 else if (src_regno < 0
6071 && dst_regno >= lra_constraint_new_regno_start
6072 && invariant_p (SET_SRC (curr_set))
6073 && (cl = lra_get_allocno_class (dst_regno)) != NO_REGS
6074 && ! bitmap_bit_p (&invalid_invariant_regs, dst_regno)
6075 && ! bitmap_bit_p (&invalid_invariant_regs,
6076 ORIGINAL_REGNO(regno_reg_rtx[dst_regno])))
6078 /* 'reload_pseudo <- invariant'. */
6079 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6080 reloads_num++;
6081 update_reloads_num_p = false;
6082 if (process_invariant_for_inheritance (SET_DEST (curr_set), SET_SRC (curr_set)))
6083 change_p = true;
6084 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6085 IOR_HARD_REG_SET (potential_reload_hard_regs,
6086 reg_class_contents[cl]);
6088 else if (src_regno >= lra_constraint_new_regno_start
6089 && dst_regno < lra_constraint_new_regno_start
6090 && dst_regno >= FIRST_PSEUDO_REGISTER
6091 && reg_renumber[dst_regno] < 0
6092 && (cl = lra_get_allocno_class (src_regno)) != NO_REGS
6093 && usage_insns[dst_regno].check == curr_usage_insns_check
6094 && (next_usage_insns
6095 = usage_insns[dst_regno].insns) != NULL_RTX)
6097 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6098 reloads_num++;
6099 update_reloads_num_p = false;
6100 /* 'original_pseudo <- reload_pseudo'. */
6101 if (! JUMP_P (curr_insn)
6102 && inherit_reload_reg (true, dst_regno, cl,
6103 curr_insn, next_usage_insns))
6104 change_p = true;
6105 /* Invalidate. */
6106 usage_insns[dst_regno].check = 0;
6107 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6108 IOR_HARD_REG_SET (potential_reload_hard_regs,
6109 reg_class_contents[cl]);
6111 else if (INSN_P (curr_insn))
6113 int iter;
6114 int max_uid = get_max_uid ();
6116 curr_id = lra_get_insn_recog_data (curr_insn);
6117 curr_static_id = curr_id->insn_static_data;
6118 to_inherit_num = 0;
6119 /* Process insn definitions. */
6120 for (iter = 0; iter < 2; iter++)
6121 for (reg = iter == 0 ? curr_id->regs : curr_static_id->hard_regs;
6122 reg != NULL;
6123 reg = reg->next)
6124 if (reg->type != OP_IN
6125 && (dst_regno = reg->regno) < lra_constraint_new_regno_start)
6127 if (dst_regno >= FIRST_PSEUDO_REGISTER && reg->type == OP_OUT
6128 && reg_renumber[dst_regno] < 0 && ! reg->subreg_p
6129 && usage_insns[dst_regno].check == curr_usage_insns_check
6130 && (next_usage_insns
6131 = usage_insns[dst_regno].insns) != NULL_RTX)
6133 struct lra_insn_reg *r;
6135 for (r = curr_id->regs; r != NULL; r = r->next)
6136 if (r->type != OP_OUT && r->regno == dst_regno)
6137 break;
6138 /* Don't do inheritance if the pseudo is also
6139 used in the insn. */
6140 if (r == NULL)
6141 /* We can not do inheritance right now
6142 because the current insn reg info (chain
6143 regs) can change after that. */
6144 add_to_inherit (dst_regno, next_usage_insns);
6146 /* We can not process one reg twice here because of
6147 usage_insns invalidation. */
6148 if ((dst_regno < FIRST_PSEUDO_REGISTER
6149 || reg_renumber[dst_regno] >= 0)
6150 && ! reg->subreg_p && reg->type != OP_IN)
6152 HARD_REG_SET s;
6154 if (split_if_necessary (dst_regno, reg->biggest_mode,
6155 potential_reload_hard_regs,
6156 false, curr_insn, max_uid))
6157 change_p = true;
6158 CLEAR_HARD_REG_SET (s);
6159 if (dst_regno < FIRST_PSEUDO_REGISTER)
6160 add_to_hard_reg_set (&s, reg->biggest_mode, dst_regno);
6161 else
6162 add_to_hard_reg_set (&s, PSEUDO_REGNO_MODE (dst_regno),
6163 reg_renumber[dst_regno]);
6164 AND_COMPL_HARD_REG_SET (live_hard_regs, s);
6166 /* We should invalidate potential inheritance or
6167 splitting for the current insn usages to the next
6168 usage insns (see code below) as the output pseudo
6169 prevents this. */
6170 if ((dst_regno >= FIRST_PSEUDO_REGISTER
6171 && reg_renumber[dst_regno] < 0)
6172 || (reg->type == OP_OUT && ! reg->subreg_p
6173 && (dst_regno < FIRST_PSEUDO_REGISTER
6174 || reg_renumber[dst_regno] >= 0)))
6176 /* Invalidate and mark definitions. */
6177 if (dst_regno >= FIRST_PSEUDO_REGISTER)
6178 usage_insns[dst_regno].check = -(int) INSN_UID (curr_insn);
6179 else
6181 nregs = hard_regno_nregs[dst_regno][reg->biggest_mode];
6182 for (i = 0; i < nregs; i++)
6183 usage_insns[dst_regno + i].check
6184 = -(int) INSN_UID (curr_insn);
6188 /* Process clobbered call regs. */
6189 if (curr_id->arg_hard_regs != NULL)
6190 for (i = 0; (dst_regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6191 if (dst_regno >= FIRST_PSEUDO_REGISTER)
6192 usage_insns[dst_regno - FIRST_PSEUDO_REGISTER].check
6193 = -(int) INSN_UID (curr_insn);
6194 if (! JUMP_P (curr_insn))
6195 for (i = 0; i < to_inherit_num; i++)
6196 if (inherit_reload_reg (true, to_inherit[i].regno,
6197 ALL_REGS, curr_insn,
6198 to_inherit[i].insns))
6199 change_p = true;
6200 if (CALL_P (curr_insn))
6202 rtx cheap, pat, dest;
6203 rtx_insn *restore;
6204 int regno, hard_regno;
6206 calls_num++;
6207 if ((cheap = find_reg_note (curr_insn,
6208 REG_RETURNED, NULL_RTX)) != NULL_RTX
6209 && ((cheap = XEXP (cheap, 0)), true)
6210 && (regno = REGNO (cheap)) >= FIRST_PSEUDO_REGISTER
6211 && (hard_regno = reg_renumber[regno]) >= 0
6212 /* If there are pending saves/restores, the
6213 optimization is not worth. */
6214 && usage_insns[regno].calls_num == calls_num - 1
6215 && TEST_HARD_REG_BIT (call_used_reg_set, hard_regno))
6217 /* Restore the pseudo from the call result as
6218 REG_RETURNED note says that the pseudo value is
6219 in the call result and the pseudo is an argument
6220 of the call. */
6221 pat = PATTERN (curr_insn);
6222 if (GET_CODE (pat) == PARALLEL)
6223 pat = XVECEXP (pat, 0, 0);
6224 dest = SET_DEST (pat);
6225 /* For multiple return values dest is PARALLEL.
6226 Currently we handle only single return value case. */
6227 if (REG_P (dest))
6229 start_sequence ();
6230 emit_move_insn (cheap, copy_rtx (dest));
6231 restore = get_insns ();
6232 end_sequence ();
6233 lra_process_new_insns (curr_insn, NULL, restore,
6234 "Inserting call parameter restore");
6235 /* We don't need to save/restore of the pseudo from
6236 this call. */
6237 usage_insns[regno].calls_num = calls_num;
6238 bitmap_set_bit (&check_only_regs, regno);
6242 to_inherit_num = 0;
6243 /* Process insn usages. */
6244 for (iter = 0; iter < 2; iter++)
6245 for (reg = iter == 0 ? curr_id->regs : curr_static_id->hard_regs;
6246 reg != NULL;
6247 reg = reg->next)
6248 if ((reg->type != OP_OUT
6249 || (reg->type == OP_OUT && reg->subreg_p))
6250 && (src_regno = reg->regno) < lra_constraint_new_regno_start)
6252 if (src_regno >= FIRST_PSEUDO_REGISTER
6253 && reg_renumber[src_regno] < 0 && reg->type == OP_IN)
6255 if (usage_insns[src_regno].check == curr_usage_insns_check
6256 && (next_usage_insns
6257 = usage_insns[src_regno].insns) != NULL_RTX
6258 && NONDEBUG_INSN_P (curr_insn))
6259 add_to_inherit (src_regno, next_usage_insns);
6260 else if (usage_insns[src_regno].check
6261 != -(int) INSN_UID (curr_insn))
6262 /* Add usages but only if the reg is not set up
6263 in the same insn. */
6264 add_next_usage_insn (src_regno, curr_insn, reloads_num);
6266 else if (src_regno < FIRST_PSEUDO_REGISTER
6267 || reg_renumber[src_regno] >= 0)
6269 bool before_p;
6270 rtx_insn *use_insn = curr_insn;
6272 before_p = (JUMP_P (curr_insn)
6273 || (CALL_P (curr_insn) && reg->type == OP_IN));
6274 if (NONDEBUG_INSN_P (curr_insn)
6275 && (! JUMP_P (curr_insn) || reg->type == OP_IN)
6276 && split_if_necessary (src_regno, reg->biggest_mode,
6277 potential_reload_hard_regs,
6278 before_p, curr_insn, max_uid))
6280 if (reg->subreg_p)
6281 lra_risky_transformations_p = true;
6282 change_p = true;
6283 /* Invalidate. */
6284 usage_insns[src_regno].check = 0;
6285 if (before_p)
6286 use_insn = PREV_INSN (curr_insn);
6288 if (NONDEBUG_INSN_P (curr_insn))
6290 if (src_regno < FIRST_PSEUDO_REGISTER)
6291 add_to_hard_reg_set (&live_hard_regs,
6292 reg->biggest_mode, src_regno);
6293 else
6294 add_to_hard_reg_set (&live_hard_regs,
6295 PSEUDO_REGNO_MODE (src_regno),
6296 reg_renumber[src_regno]);
6298 add_next_usage_insn (src_regno, use_insn, reloads_num);
6301 /* Process used call regs. */
6302 if (curr_id->arg_hard_regs != NULL)
6303 for (i = 0; (src_regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6304 if (src_regno < FIRST_PSEUDO_REGISTER)
6306 SET_HARD_REG_BIT (live_hard_regs, src_regno);
6307 add_next_usage_insn (src_regno, curr_insn, reloads_num);
6309 for (i = 0; i < to_inherit_num; i++)
6311 src_regno = to_inherit[i].regno;
6312 if (inherit_reload_reg (false, src_regno, ALL_REGS,
6313 curr_insn, to_inherit[i].insns))
6314 change_p = true;
6315 else
6316 setup_next_usage_insn (src_regno, curr_insn, reloads_num, false);
6319 if (update_reloads_num_p
6320 && NONDEBUG_INSN_P (curr_insn) && curr_set != NULL_RTX)
6322 int regno = -1;
6323 if ((REG_P (SET_DEST (curr_set))
6324 && (regno = REGNO (SET_DEST (curr_set))) >= lra_constraint_new_regno_start
6325 && reg_renumber[regno] < 0
6326 && (cl = lra_get_allocno_class (regno)) != NO_REGS)
6327 || (REG_P (SET_SRC (curr_set))
6328 && (regno = REGNO (SET_SRC (curr_set))) >= lra_constraint_new_regno_start
6329 && reg_renumber[regno] < 0
6330 && (cl = lra_get_allocno_class (regno)) != NO_REGS))
6332 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6333 reloads_num++;
6334 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6335 IOR_HARD_REG_SET (potential_reload_hard_regs,
6336 reg_class_contents[cl]);
6339 if (NONDEBUG_INSN_P (curr_insn))
6341 int regno;
6343 /* Invalidate invariants with changed regs. */
6344 curr_id = lra_get_insn_recog_data (curr_insn);
6345 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
6346 if (reg->type != OP_IN)
6348 bitmap_set_bit (&invalid_invariant_regs, reg->regno);
6349 bitmap_set_bit (&invalid_invariant_regs,
6350 ORIGINAL_REGNO (regno_reg_rtx[reg->regno]));
6352 curr_static_id = curr_id->insn_static_data;
6353 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
6354 if (reg->type != OP_IN)
6355 bitmap_set_bit (&invalid_invariant_regs, reg->regno);
6356 if (curr_id->arg_hard_regs != NULL)
6357 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6358 if (regno >= FIRST_PSEUDO_REGISTER)
6359 bitmap_set_bit (&invalid_invariant_regs,
6360 regno - FIRST_PSEUDO_REGISTER);
6362 /* We reached the start of the current basic block. */
6363 if (prev_insn == NULL_RTX || prev_insn == PREV_INSN (head)
6364 || BLOCK_FOR_INSN (prev_insn) != curr_bb)
6366 /* We reached the beginning of the current block -- do
6367 rest of spliting in the current BB. */
6368 to_process = df_get_live_in (curr_bb);
6369 if (BLOCK_FOR_INSN (head) != curr_bb)
6371 /* We are somewhere in the middle of EBB. */
6372 get_live_on_other_edges (EDGE_PRED (curr_bb, 0)->src,
6373 curr_bb, &temp_bitmap);
6374 to_process = &temp_bitmap;
6376 head_p = true;
6377 EXECUTE_IF_SET_IN_BITMAP (to_process, 0, j, bi)
6379 if ((int) j >= lra_constraint_new_regno_start)
6380 break;
6381 if (((int) j < FIRST_PSEUDO_REGISTER || reg_renumber[j] >= 0)
6382 && usage_insns[j].check == curr_usage_insns_check
6383 && (next_usage_insns = usage_insns[j].insns) != NULL_RTX)
6385 if (need_for_split_p (potential_reload_hard_regs, j))
6387 if (lra_dump_file != NULL && head_p)
6389 fprintf (lra_dump_file,
6390 " ----------------------------------\n");
6391 head_p = false;
6393 if (split_reg (false, j, bb_note (curr_bb),
6394 next_usage_insns))
6395 change_p = true;
6397 usage_insns[j].check = 0;
6402 return change_p;
6405 /* This value affects EBB forming. If probability of edge from EBB to
6406 a BB is not greater than the following value, we don't add the BB
6407 to EBB. */
6408 #define EBB_PROBABILITY_CUTOFF \
6409 ((REG_BR_PROB_BASE * LRA_INHERITANCE_EBB_PROBABILITY_CUTOFF) / 100)
6411 /* Current number of inheritance/split iteration. */
6412 int lra_inheritance_iter;
6414 /* Entry function for inheritance/split pass. */
6415 void
6416 lra_inheritance (void)
6418 int i;
6419 basic_block bb, start_bb;
6420 edge e;
6422 lra_inheritance_iter++;
6423 if (lra_inheritance_iter > LRA_MAX_INHERITANCE_PASSES)
6424 return;
6425 timevar_push (TV_LRA_INHERITANCE);
6426 if (lra_dump_file != NULL)
6427 fprintf (lra_dump_file, "\n********** Inheritance #%d: **********\n\n",
6428 lra_inheritance_iter);
6429 curr_usage_insns_check = 0;
6430 usage_insns = XNEWVEC (struct usage_insns, lra_constraint_new_regno_start);
6431 for (i = 0; i < lra_constraint_new_regno_start; i++)
6432 usage_insns[i].check = 0;
6433 bitmap_initialize (&check_only_regs, &reg_obstack);
6434 bitmap_initialize (&invalid_invariant_regs, &reg_obstack);
6435 bitmap_initialize (&live_regs, &reg_obstack);
6436 bitmap_initialize (&temp_bitmap, &reg_obstack);
6437 bitmap_initialize (&ebb_global_regs, &reg_obstack);
6438 FOR_EACH_BB_FN (bb, cfun)
6440 start_bb = bb;
6441 if (lra_dump_file != NULL)
6442 fprintf (lra_dump_file, "EBB");
6443 /* Form a EBB starting with BB. */
6444 bitmap_clear (&ebb_global_regs);
6445 bitmap_ior_into (&ebb_global_regs, df_get_live_in (bb));
6446 for (;;)
6448 if (lra_dump_file != NULL)
6449 fprintf (lra_dump_file, " %d", bb->index);
6450 if (bb->next_bb == EXIT_BLOCK_PTR_FOR_FN (cfun)
6451 || LABEL_P (BB_HEAD (bb->next_bb)))
6452 break;
6453 e = find_fallthru_edge (bb->succs);
6454 if (! e)
6455 break;
6456 if (e->probability < EBB_PROBABILITY_CUTOFF)
6457 break;
6458 bb = bb->next_bb;
6460 bitmap_ior_into (&ebb_global_regs, df_get_live_out (bb));
6461 if (lra_dump_file != NULL)
6462 fprintf (lra_dump_file, "\n");
6463 if (inherit_in_ebb (BB_HEAD (start_bb), BB_END (bb)))
6464 /* Remember that the EBB head and tail can change in
6465 inherit_in_ebb. */
6466 update_ebb_live_info (BB_HEAD (start_bb), BB_END (bb));
6468 bitmap_clear (&ebb_global_regs);
6469 bitmap_clear (&temp_bitmap);
6470 bitmap_clear (&live_regs);
6471 bitmap_clear (&invalid_invariant_regs);
6472 bitmap_clear (&check_only_regs);
6473 free (usage_insns);
6475 timevar_pop (TV_LRA_INHERITANCE);
6480 /* This page contains code to undo failed inheritance/split
6481 transformations. */
6483 /* Current number of iteration undoing inheritance/split. */
6484 int lra_undo_inheritance_iter;
6486 /* Fix BB live info LIVE after removing pseudos created on pass doing
6487 inheritance/split which are REMOVED_PSEUDOS. */
6488 static void
6489 fix_bb_live_info (bitmap live, bitmap removed_pseudos)
6491 unsigned int regno;
6492 bitmap_iterator bi;
6494 EXECUTE_IF_SET_IN_BITMAP (removed_pseudos, 0, regno, bi)
6495 if (bitmap_clear_bit (live, regno)
6496 && REG_P (lra_reg_info[regno].restore_rtx))
6497 bitmap_set_bit (live, REGNO (lra_reg_info[regno].restore_rtx));
6500 /* Return regno of the (subreg of) REG. Otherwise, return a negative
6501 number. */
6502 static int
6503 get_regno (rtx reg)
6505 if (GET_CODE (reg) == SUBREG)
6506 reg = SUBREG_REG (reg);
6507 if (REG_P (reg))
6508 return REGNO (reg);
6509 return -1;
6512 /* Delete a move INSN with destination reg DREGNO and a previous
6513 clobber insn with the same regno. The inheritance/split code can
6514 generate moves with preceding clobber and when we delete such moves
6515 we should delete the clobber insn too to keep the correct life
6516 info. */
6517 static void
6518 delete_move_and_clobber (rtx_insn *insn, int dregno)
6520 rtx_insn *prev_insn = PREV_INSN (insn);
6522 lra_set_insn_deleted (insn);
6523 lra_assert (dregno >= 0);
6524 if (prev_insn != NULL && NONDEBUG_INSN_P (prev_insn)
6525 && GET_CODE (PATTERN (prev_insn)) == CLOBBER
6526 && dregno == get_regno (XEXP (PATTERN (prev_insn), 0)))
6527 lra_set_insn_deleted (prev_insn);
6530 /* Remove inheritance/split pseudos which are in REMOVE_PSEUDOS and
6531 return true if we did any change. The undo transformations for
6532 inheritance looks like
6533 i <- i2
6534 p <- i => p <- i2
6535 or removing
6536 p <- i, i <- p, and i <- i3
6537 where p is original pseudo from which inheritance pseudo i was
6538 created, i and i3 are removed inheritance pseudos, i2 is another
6539 not removed inheritance pseudo. All split pseudos or other
6540 occurrences of removed inheritance pseudos are changed on the
6541 corresponding original pseudos.
6543 The function also schedules insns changed and created during
6544 inheritance/split pass for processing by the subsequent constraint
6545 pass. */
6546 static bool
6547 remove_inheritance_pseudos (bitmap remove_pseudos)
6549 basic_block bb;
6550 int regno, sregno, prev_sregno, dregno;
6551 rtx restore_rtx;
6552 rtx set, prev_set;
6553 rtx_insn *prev_insn;
6554 bool change_p, done_p;
6556 change_p = ! bitmap_empty_p (remove_pseudos);
6557 /* We can not finish the function right away if CHANGE_P is true
6558 because we need to marks insns affected by previous
6559 inheritance/split pass for processing by the subsequent
6560 constraint pass. */
6561 FOR_EACH_BB_FN (bb, cfun)
6563 fix_bb_live_info (df_get_live_in (bb), remove_pseudos);
6564 fix_bb_live_info (df_get_live_out (bb), remove_pseudos);
6565 FOR_BB_INSNS_REVERSE (bb, curr_insn)
6567 if (! INSN_P (curr_insn))
6568 continue;
6569 done_p = false;
6570 sregno = dregno = -1;
6571 if (change_p && NONDEBUG_INSN_P (curr_insn)
6572 && (set = single_set (curr_insn)) != NULL_RTX)
6574 dregno = get_regno (SET_DEST (set));
6575 sregno = get_regno (SET_SRC (set));
6578 if (sregno >= 0 && dregno >= 0)
6580 if (bitmap_bit_p (remove_pseudos, dregno)
6581 && ! REG_P (lra_reg_info[dregno].restore_rtx))
6583 /* invariant inheritance pseudo <- original pseudo */
6584 if (lra_dump_file != NULL)
6586 fprintf (lra_dump_file, " Removing invariant inheritance:\n");
6587 dump_insn_slim (lra_dump_file, curr_insn);
6588 fprintf (lra_dump_file, "\n");
6590 delete_move_and_clobber (curr_insn, dregno);
6591 done_p = true;
6593 else if (bitmap_bit_p (remove_pseudos, sregno)
6594 && ! REG_P (lra_reg_info[sregno].restore_rtx))
6596 /* reload pseudo <- invariant inheritance pseudo */
6597 start_sequence ();
6598 /* We can not just change the source. It might be
6599 an insn different from the move. */
6600 emit_insn (lra_reg_info[sregno].restore_rtx);
6601 rtx_insn *new_insns = get_insns ();
6602 end_sequence ();
6603 lra_assert (single_set (new_insns) != NULL
6604 && SET_DEST (set) == SET_DEST (single_set (new_insns)));
6605 lra_process_new_insns (curr_insn, NULL, new_insns,
6606 "Changing reload<-invariant inheritance");
6607 delete_move_and_clobber (curr_insn, dregno);
6608 done_p = true;
6610 else if ((bitmap_bit_p (remove_pseudos, sregno)
6611 && (get_regno (lra_reg_info[sregno].restore_rtx) == dregno
6612 || (bitmap_bit_p (remove_pseudos, dregno)
6613 && get_regno (lra_reg_info[sregno].restore_rtx) >= 0
6614 && (get_regno (lra_reg_info[sregno].restore_rtx)
6615 == get_regno (lra_reg_info[dregno].restore_rtx)))))
6616 || (bitmap_bit_p (remove_pseudos, dregno)
6617 && get_regno (lra_reg_info[dregno].restore_rtx) == sregno))
6618 /* One of the following cases:
6619 original <- removed inheritance pseudo
6620 removed inherit pseudo <- another removed inherit pseudo
6621 removed inherit pseudo <- original pseudo
6623 removed_split_pseudo <- original_reg
6624 original_reg <- removed_split_pseudo */
6626 if (lra_dump_file != NULL)
6628 fprintf (lra_dump_file, " Removing %s:\n",
6629 bitmap_bit_p (&lra_split_regs, sregno)
6630 || bitmap_bit_p (&lra_split_regs, dregno)
6631 ? "split" : "inheritance");
6632 dump_insn_slim (lra_dump_file, curr_insn);
6634 delete_move_and_clobber (curr_insn, dregno);
6635 done_p = true;
6637 else if (bitmap_bit_p (remove_pseudos, sregno)
6638 && bitmap_bit_p (&lra_inheritance_pseudos, sregno))
6640 /* Search the following pattern:
6641 inherit_or_split_pseudo1 <- inherit_or_split_pseudo2
6642 original_pseudo <- inherit_or_split_pseudo1
6643 where the 2nd insn is the current insn and
6644 inherit_or_split_pseudo2 is not removed. If it is found,
6645 change the current insn onto:
6646 original_pseudo <- inherit_or_split_pseudo2. */
6647 for (prev_insn = PREV_INSN (curr_insn);
6648 prev_insn != NULL_RTX && ! NONDEBUG_INSN_P (prev_insn);
6649 prev_insn = PREV_INSN (prev_insn))
6651 if (prev_insn != NULL_RTX && BLOCK_FOR_INSN (prev_insn) == bb
6652 && (prev_set = single_set (prev_insn)) != NULL_RTX
6653 /* There should be no subregs in insn we are
6654 searching because only the original reg might
6655 be in subreg when we changed the mode of
6656 load/store for splitting. */
6657 && REG_P (SET_DEST (prev_set))
6658 && REG_P (SET_SRC (prev_set))
6659 && (int) REGNO (SET_DEST (prev_set)) == sregno
6660 && ((prev_sregno = REGNO (SET_SRC (prev_set)))
6661 >= FIRST_PSEUDO_REGISTER)
6662 && (lra_reg_info[prev_sregno].restore_rtx == NULL_RTX
6664 /* As we consider chain of inheritance or
6665 splitting described in above comment we should
6666 check that sregno and prev_sregno were
6667 inheritance/split pseudos created from the
6668 same original regno. */
6669 (get_regno (lra_reg_info[sregno].restore_rtx) >= 0
6670 && (get_regno (lra_reg_info[sregno].restore_rtx)
6671 == get_regno (lra_reg_info[prev_sregno].restore_rtx))))
6672 && ! bitmap_bit_p (remove_pseudos, prev_sregno))
6674 lra_assert (GET_MODE (SET_SRC (prev_set))
6675 == GET_MODE (regno_reg_rtx[sregno]));
6676 if (GET_CODE (SET_SRC (set)) == SUBREG)
6677 SUBREG_REG (SET_SRC (set)) = SET_SRC (prev_set);
6678 else
6679 SET_SRC (set) = SET_SRC (prev_set);
6680 /* As we are finishing with processing the insn
6681 here, check the destination too as it might
6682 inheritance pseudo for another pseudo. */
6683 if (bitmap_bit_p (remove_pseudos, dregno)
6684 && bitmap_bit_p (&lra_inheritance_pseudos, dregno)
6685 && (restore_rtx
6686 = lra_reg_info[dregno].restore_rtx) != NULL_RTX)
6688 if (GET_CODE (SET_DEST (set)) == SUBREG)
6689 SUBREG_REG (SET_DEST (set)) = restore_rtx;
6690 else
6691 SET_DEST (set) = restore_rtx;
6693 lra_push_insn_and_update_insn_regno_info (curr_insn);
6694 lra_set_used_insn_alternative_by_uid
6695 (INSN_UID (curr_insn), -1);
6696 done_p = true;
6697 if (lra_dump_file != NULL)
6699 fprintf (lra_dump_file, " Change reload insn:\n");
6700 dump_insn_slim (lra_dump_file, curr_insn);
6705 if (! done_p)
6707 struct lra_insn_reg *reg;
6708 bool restored_regs_p = false;
6709 bool kept_regs_p = false;
6711 curr_id = lra_get_insn_recog_data (curr_insn);
6712 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
6714 regno = reg->regno;
6715 restore_rtx = lra_reg_info[regno].restore_rtx;
6716 if (restore_rtx != NULL_RTX)
6718 if (change_p && bitmap_bit_p (remove_pseudos, regno))
6720 lra_substitute_pseudo_within_insn
6721 (curr_insn, regno, restore_rtx, false);
6722 restored_regs_p = true;
6724 else
6725 kept_regs_p = true;
6728 if (NONDEBUG_INSN_P (curr_insn) && kept_regs_p)
6730 /* The instruction has changed since the previous
6731 constraints pass. */
6732 lra_push_insn_and_update_insn_regno_info (curr_insn);
6733 lra_set_used_insn_alternative_by_uid
6734 (INSN_UID (curr_insn), -1);
6736 else if (restored_regs_p)
6737 /* The instruction has been restored to the form that
6738 it had during the previous constraints pass. */
6739 lra_update_insn_regno_info (curr_insn);
6740 if (restored_regs_p && lra_dump_file != NULL)
6742 fprintf (lra_dump_file, " Insn after restoring regs:\n");
6743 dump_insn_slim (lra_dump_file, curr_insn);
6748 return change_p;
6751 /* If optional reload pseudos failed to get a hard register or was not
6752 inherited, it is better to remove optional reloads. We do this
6753 transformation after undoing inheritance to figure out necessity to
6754 remove optional reloads easier. Return true if we do any
6755 change. */
6756 static bool
6757 undo_optional_reloads (void)
6759 bool change_p, keep_p;
6760 unsigned int regno, uid;
6761 bitmap_iterator bi, bi2;
6762 rtx_insn *insn;
6763 rtx set, src, dest;
6764 bitmap_head removed_optional_reload_pseudos, insn_bitmap;
6766 bitmap_initialize (&removed_optional_reload_pseudos, &reg_obstack);
6767 bitmap_copy (&removed_optional_reload_pseudos, &lra_optional_reload_pseudos);
6768 EXECUTE_IF_SET_IN_BITMAP (&lra_optional_reload_pseudos, 0, regno, bi)
6770 keep_p = false;
6771 /* Keep optional reloads from previous subpasses. */
6772 if (lra_reg_info[regno].restore_rtx == NULL_RTX
6773 /* If the original pseudo changed its allocation, just
6774 removing the optional pseudo is dangerous as the original
6775 pseudo will have longer live range. */
6776 || reg_renumber[REGNO (lra_reg_info[regno].restore_rtx)] >= 0)
6777 keep_p = true;
6778 else if (reg_renumber[regno] >= 0)
6779 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi2)
6781 insn = lra_insn_recog_data[uid]->insn;
6782 if ((set = single_set (insn)) == NULL_RTX)
6783 continue;
6784 src = SET_SRC (set);
6785 dest = SET_DEST (set);
6786 if (! REG_P (src) || ! REG_P (dest))
6787 continue;
6788 if (REGNO (dest) == regno
6789 /* Ignore insn for optional reloads itself. */
6790 && REGNO (lra_reg_info[regno].restore_rtx) != REGNO (src)
6791 /* Check only inheritance on last inheritance pass. */
6792 && (int) REGNO (src) >= new_regno_start
6793 /* Check that the optional reload was inherited. */
6794 && bitmap_bit_p (&lra_inheritance_pseudos, REGNO (src)))
6796 keep_p = true;
6797 break;
6800 if (keep_p)
6802 bitmap_clear_bit (&removed_optional_reload_pseudos, regno);
6803 if (lra_dump_file != NULL)
6804 fprintf (lra_dump_file, "Keep optional reload reg %d\n", regno);
6807 change_p = ! bitmap_empty_p (&removed_optional_reload_pseudos);
6808 bitmap_initialize (&insn_bitmap, &reg_obstack);
6809 EXECUTE_IF_SET_IN_BITMAP (&removed_optional_reload_pseudos, 0, regno, bi)
6811 if (lra_dump_file != NULL)
6812 fprintf (lra_dump_file, "Remove optional reload reg %d\n", regno);
6813 bitmap_copy (&insn_bitmap, &lra_reg_info[regno].insn_bitmap);
6814 EXECUTE_IF_SET_IN_BITMAP (&insn_bitmap, 0, uid, bi2)
6816 insn = lra_insn_recog_data[uid]->insn;
6817 if ((set = single_set (insn)) != NULL_RTX)
6819 src = SET_SRC (set);
6820 dest = SET_DEST (set);
6821 if (REG_P (src) && REG_P (dest)
6822 && ((REGNO (src) == regno
6823 && (REGNO (lra_reg_info[regno].restore_rtx)
6824 == REGNO (dest)))
6825 || (REGNO (dest) == regno
6826 && (REGNO (lra_reg_info[regno].restore_rtx)
6827 == REGNO (src)))))
6829 if (lra_dump_file != NULL)
6831 fprintf (lra_dump_file, " Deleting move %u\n",
6832 INSN_UID (insn));
6833 dump_insn_slim (lra_dump_file, insn);
6835 delete_move_and_clobber (insn, REGNO (dest));
6836 continue;
6838 /* We should not worry about generation memory-memory
6839 moves here as if the corresponding inheritance did
6840 not work (inheritance pseudo did not get a hard reg),
6841 we remove the inheritance pseudo and the optional
6842 reload. */
6844 lra_substitute_pseudo_within_insn
6845 (insn, regno, lra_reg_info[regno].restore_rtx, false);
6846 lra_update_insn_regno_info (insn);
6847 if (lra_dump_file != NULL)
6849 fprintf (lra_dump_file,
6850 " Restoring original insn:\n");
6851 dump_insn_slim (lra_dump_file, insn);
6855 /* Clear restore_regnos. */
6856 EXECUTE_IF_SET_IN_BITMAP (&lra_optional_reload_pseudos, 0, regno, bi)
6857 lra_reg_info[regno].restore_rtx = NULL_RTX;
6858 bitmap_clear (&insn_bitmap);
6859 bitmap_clear (&removed_optional_reload_pseudos);
6860 return change_p;
6863 /* Entry function for undoing inheritance/split transformation. Return true
6864 if we did any RTL change in this pass. */
6865 bool
6866 lra_undo_inheritance (void)
6868 unsigned int regno;
6869 int hard_regno;
6870 int n_all_inherit, n_inherit, n_all_split, n_split;
6871 rtx restore_rtx;
6872 bitmap_head remove_pseudos;
6873 bitmap_iterator bi;
6874 bool change_p;
6876 lra_undo_inheritance_iter++;
6877 if (lra_undo_inheritance_iter > LRA_MAX_INHERITANCE_PASSES)
6878 return false;
6879 if (lra_dump_file != NULL)
6880 fprintf (lra_dump_file,
6881 "\n********** Undoing inheritance #%d: **********\n\n",
6882 lra_undo_inheritance_iter);
6883 bitmap_initialize (&remove_pseudos, &reg_obstack);
6884 n_inherit = n_all_inherit = 0;
6885 EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, regno, bi)
6886 if (lra_reg_info[regno].restore_rtx != NULL_RTX)
6888 n_all_inherit++;
6889 if (reg_renumber[regno] < 0
6890 /* If the original pseudo changed its allocation, just
6891 removing inheritance is dangerous as for changing
6892 allocation we used shorter live-ranges. */
6893 && (! REG_P (lra_reg_info[regno].restore_rtx)
6894 || reg_renumber[REGNO (lra_reg_info[regno].restore_rtx)] < 0))
6895 bitmap_set_bit (&remove_pseudos, regno);
6896 else
6897 n_inherit++;
6899 if (lra_dump_file != NULL && n_all_inherit != 0)
6900 fprintf (lra_dump_file, "Inherit %d out of %d (%.2f%%)\n",
6901 n_inherit, n_all_inherit,
6902 (double) n_inherit / n_all_inherit * 100);
6903 n_split = n_all_split = 0;
6904 EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, regno, bi)
6905 if ((restore_rtx = lra_reg_info[regno].restore_rtx) != NULL_RTX)
6907 int restore_regno = REGNO (restore_rtx);
6909 n_all_split++;
6910 hard_regno = (restore_regno >= FIRST_PSEUDO_REGISTER
6911 ? reg_renumber[restore_regno] : restore_regno);
6912 if (hard_regno < 0 || reg_renumber[regno] == hard_regno)
6913 bitmap_set_bit (&remove_pseudos, regno);
6914 else
6916 n_split++;
6917 if (lra_dump_file != NULL)
6918 fprintf (lra_dump_file, " Keep split r%d (orig=r%d)\n",
6919 regno, restore_regno);
6922 if (lra_dump_file != NULL && n_all_split != 0)
6923 fprintf (lra_dump_file, "Split %d out of %d (%.2f%%)\n",
6924 n_split, n_all_split,
6925 (double) n_split / n_all_split * 100);
6926 change_p = remove_inheritance_pseudos (&remove_pseudos);
6927 bitmap_clear (&remove_pseudos);
6928 /* Clear restore_regnos. */
6929 EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, regno, bi)
6930 lra_reg_info[regno].restore_rtx = NULL_RTX;
6931 EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, regno, bi)
6932 lra_reg_info[regno].restore_rtx = NULL_RTX;
6933 change_p = undo_optional_reloads () || change_p;
6934 return change_p;