PR c++/79653
[official-gcc.git] / gcc / lra-constraints.c
blob224a9560c5650dad29a65eda99a013fbbb28c84c
1 /* Code for RTL transformations to satisfy insn constraints.
2 Copyright (C) 2010-2017 Free Software Foundation, Inc.
3 Contributed by Vladimir Makarov <vmakarov@redhat.com>.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
10 version.
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
22 /* This file contains code for 3 passes: constraint pass,
23 inheritance/split pass, and pass for undoing failed inheritance and
24 split.
26 The major goal of constraint pass is to transform RTL to satisfy
27 insn and address constraints by:
28 o choosing insn alternatives;
29 o generating *reload insns* (or reloads in brief) and *reload
30 pseudos* which will get necessary hard registers later;
31 o substituting pseudos with equivalent values and removing the
32 instructions that initialized those pseudos.
34 The constraint pass has biggest and most complicated code in LRA.
35 There are a lot of important details like:
36 o reuse of input reload pseudos to simplify reload pseudo
37 allocations;
38 o some heuristics to choose insn alternative to improve the
39 inheritance;
40 o early clobbers etc.
42 The pass is mimicking former reload pass in alternative choosing
43 because the reload pass is oriented to current machine description
44 model. It might be changed if the machine description model is
45 changed.
47 There is special code for preventing all LRA and this pass cycling
48 in case of bugs.
50 On the first iteration of the pass we process every instruction and
51 choose an alternative for each one. On subsequent iterations we try
52 to avoid reprocessing instructions if we can be sure that the old
53 choice is still valid.
55 The inheritance/spilt pass is to transform code to achieve
56 ineheritance and live range splitting. It is done on backward
57 traversal of EBBs.
59 The inheritance optimization goal is to reuse values in hard
60 registers. There is analogous optimization in old reload pass. The
61 inheritance is achieved by following transformation:
63 reload_p1 <- p reload_p1 <- p
64 ... new_p <- reload_p1
65 ... => ...
66 reload_p2 <- p reload_p2 <- new_p
68 where p is spilled and not changed between the insns. Reload_p1 is
69 also called *original pseudo* and new_p is called *inheritance
70 pseudo*.
72 The subsequent assignment pass will try to assign the same (or
73 another if it is not possible) hard register to new_p as to
74 reload_p1 or reload_p2.
76 If the assignment pass fails to assign a hard register to new_p,
77 this file will undo the inheritance and restore the original code.
78 This is because implementing the above sequence with a spilled
79 new_p would make the code much worse. The inheritance is done in
80 EBB scope. The above is just a simplified example to get an idea
81 of the inheritance as the inheritance is also done for non-reload
82 insns.
84 Splitting (transformation) is also done in EBB scope on the same
85 pass as the inheritance:
87 r <- ... or ... <- r r <- ... or ... <- r
88 ... s <- r (new insn -- save)
89 ... =>
90 ... r <- s (new insn -- restore)
91 ... <- r ... <- r
93 The *split pseudo* s is assigned to the hard register of the
94 original pseudo or hard register r.
96 Splitting is done:
97 o In EBBs with high register pressure for global pseudos (living
98 in at least 2 BBs) and assigned to hard registers when there
99 are more one reloads needing the hard registers;
100 o for pseudos needing save/restore code around calls.
102 If the split pseudo still has the same hard register as the
103 original pseudo after the subsequent assignment pass or the
104 original pseudo was split, the opposite transformation is done on
105 the same pass for undoing inheritance. */
107 #undef REG_OK_STRICT
109 #include "config.h"
110 #include "system.h"
111 #include "coretypes.h"
112 #include "backend.h"
113 #include "target.h"
114 #include "rtl.h"
115 #include "tree.h"
116 #include "predict.h"
117 #include "df.h"
118 #include "memmodel.h"
119 #include "tm_p.h"
120 #include "expmed.h"
121 #include "optabs.h"
122 #include "regs.h"
123 #include "ira.h"
124 #include "recog.h"
125 #include "output.h"
126 #include "addresses.h"
127 #include "expr.h"
128 #include "cfgrtl.h"
129 #include "rtl-error.h"
130 #include "params.h"
131 #include "lra.h"
132 #include "lra-int.h"
133 #include "print-rtl.h"
135 /* Value of LRA_CURR_RELOAD_NUM at the beginning of BB of the current
136 insn. Remember that LRA_CURR_RELOAD_NUM is the number of emitted
137 reload insns. */
138 static int bb_reload_num;
140 /* The current insn being processed and corresponding its single set
141 (NULL otherwise), its data (basic block, the insn data, the insn
142 static data, and the mode of each operand). */
143 static rtx_insn *curr_insn;
144 static rtx curr_insn_set;
145 static basic_block curr_bb;
146 static lra_insn_recog_data_t curr_id;
147 static struct lra_static_insn_data *curr_static_id;
148 static machine_mode curr_operand_mode[MAX_RECOG_OPERANDS];
149 /* Mode of the register substituted by its equivalence with VOIDmode
150 (e.g. constant) and whose subreg is given operand of the current
151 insn. VOIDmode in all other cases. */
152 static machine_mode original_subreg_reg_mode[MAX_RECOG_OPERANDS];
156 /* Start numbers for new registers and insns at the current constraints
157 pass start. */
158 static int new_regno_start;
159 static int new_insn_uid_start;
161 /* If LOC is nonnull, strip any outer subreg from it. */
162 static inline rtx *
163 strip_subreg (rtx *loc)
165 return loc && GET_CODE (*loc) == SUBREG ? &SUBREG_REG (*loc) : loc;
168 /* Return hard regno of REGNO or if it is was not assigned to a hard
169 register, use a hard register from its allocno class. */
170 static int
171 get_try_hard_regno (int regno)
173 int hard_regno;
174 enum reg_class rclass;
176 if ((hard_regno = regno) >= FIRST_PSEUDO_REGISTER)
177 hard_regno = lra_get_regno_hard_regno (regno);
178 if (hard_regno >= 0)
179 return hard_regno;
180 rclass = lra_get_allocno_class (regno);
181 if (rclass == NO_REGS)
182 return -1;
183 return ira_class_hard_regs[rclass][0];
186 /* Return the hard regno of X after removing its subreg. If X is not
187 a register or a subreg of a register, return -1. If X is a pseudo,
188 use its assignment. If FINAL_P return the final hard regno which will
189 be after elimination. */
190 static int
191 get_hard_regno (rtx x, bool final_p)
193 rtx reg;
194 int hard_regno;
196 reg = x;
197 if (SUBREG_P (x))
198 reg = SUBREG_REG (x);
199 if (! REG_P (reg))
200 return -1;
201 if (! HARD_REGISTER_NUM_P (hard_regno = REGNO (reg)))
202 hard_regno = lra_get_regno_hard_regno (hard_regno);
203 if (hard_regno < 0)
204 return -1;
205 if (final_p)
206 hard_regno = lra_get_elimination_hard_regno (hard_regno);
207 if (SUBREG_P (x))
208 hard_regno += subreg_regno_offset (hard_regno, GET_MODE (reg),
209 SUBREG_BYTE (x), GET_MODE (x));
210 return hard_regno;
213 /* If REGNO is a hard register or has been allocated a hard register,
214 return the class of that register. If REGNO is a reload pseudo
215 created by the current constraints pass, return its allocno class.
216 Return NO_REGS otherwise. */
217 static enum reg_class
218 get_reg_class (int regno)
220 int hard_regno;
222 if (! HARD_REGISTER_NUM_P (hard_regno = regno))
223 hard_regno = lra_get_regno_hard_regno (regno);
224 if (hard_regno >= 0)
226 hard_regno = lra_get_elimination_hard_regno (hard_regno);
227 return REGNO_REG_CLASS (hard_regno);
229 if (regno >= new_regno_start)
230 return lra_get_allocno_class (regno);
231 return NO_REGS;
234 /* Return true if REG satisfies (or will satisfy) reg class constraint
235 CL. Use elimination first if REG is a hard register. If REG is a
236 reload pseudo created by this constraints pass, assume that it will
237 be allocated a hard register from its allocno class, but allow that
238 class to be narrowed to CL if it is currently a superset of CL.
240 If NEW_CLASS is nonnull, set *NEW_CLASS to the new allocno class of
241 REGNO (reg), or NO_REGS if no change in its class was needed. */
242 static bool
243 in_class_p (rtx reg, enum reg_class cl, enum reg_class *new_class)
245 enum reg_class rclass, common_class;
246 machine_mode reg_mode;
247 int class_size, hard_regno, nregs, i, j;
248 int regno = REGNO (reg);
250 if (new_class != NULL)
251 *new_class = NO_REGS;
252 if (regno < FIRST_PSEUDO_REGISTER)
254 rtx final_reg = reg;
255 rtx *final_loc = &final_reg;
257 lra_eliminate_reg_if_possible (final_loc);
258 return TEST_HARD_REG_BIT (reg_class_contents[cl], REGNO (*final_loc));
260 reg_mode = GET_MODE (reg);
261 rclass = get_reg_class (regno);
262 if (regno < new_regno_start
263 /* Do not allow the constraints for reload instructions to
264 influence the classes of new pseudos. These reloads are
265 typically moves that have many alternatives, and restricting
266 reload pseudos for one alternative may lead to situations
267 where other reload pseudos are no longer allocatable. */
268 || (INSN_UID (curr_insn) >= new_insn_uid_start
269 && curr_insn_set != NULL
270 && ((OBJECT_P (SET_SRC (curr_insn_set))
271 && ! CONSTANT_P (SET_SRC (curr_insn_set)))
272 || (GET_CODE (SET_SRC (curr_insn_set)) == SUBREG
273 && OBJECT_P (SUBREG_REG (SET_SRC (curr_insn_set)))
274 && ! CONSTANT_P (SUBREG_REG (SET_SRC (curr_insn_set)))))))
275 /* When we don't know what class will be used finally for reload
276 pseudos, we use ALL_REGS. */
277 return ((regno >= new_regno_start && rclass == ALL_REGS)
278 || (rclass != NO_REGS && ira_class_subset_p[rclass][cl]
279 && ! hard_reg_set_subset_p (reg_class_contents[cl],
280 lra_no_alloc_regs)));
281 else
283 common_class = ira_reg_class_subset[rclass][cl];
284 if (new_class != NULL)
285 *new_class = common_class;
286 if (hard_reg_set_subset_p (reg_class_contents[common_class],
287 lra_no_alloc_regs))
288 return false;
289 /* Check that there are enough allocatable regs. */
290 class_size = ira_class_hard_regs_num[common_class];
291 for (i = 0; i < class_size; i++)
293 hard_regno = ira_class_hard_regs[common_class][i];
294 nregs = hard_regno_nregs[hard_regno][reg_mode];
295 if (nregs == 1)
296 return true;
297 for (j = 0; j < nregs; j++)
298 if (TEST_HARD_REG_BIT (lra_no_alloc_regs, hard_regno + j)
299 || ! TEST_HARD_REG_BIT (reg_class_contents[common_class],
300 hard_regno + j))
301 break;
302 if (j >= nregs)
303 return true;
305 return false;
309 /* Return true if REGNO satisfies a memory constraint. */
310 static bool
311 in_mem_p (int regno)
313 return get_reg_class (regno) == NO_REGS;
316 /* Return 1 if ADDR is a valid memory address for mode MODE in address
317 space AS, and check that each pseudo has the proper kind of hard
318 reg. */
319 static int
320 valid_address_p (machine_mode mode ATTRIBUTE_UNUSED,
321 rtx addr, addr_space_t as)
323 #ifdef GO_IF_LEGITIMATE_ADDRESS
324 lra_assert (ADDR_SPACE_GENERIC_P (as));
325 GO_IF_LEGITIMATE_ADDRESS (mode, addr, win);
326 return 0;
328 win:
329 return 1;
330 #else
331 return targetm.addr_space.legitimate_address_p (mode, addr, 0, as);
332 #endif
335 namespace {
336 /* Temporarily eliminates registers in an address (for the lifetime of
337 the object). */
338 class address_eliminator {
339 public:
340 address_eliminator (struct address_info *ad);
341 ~address_eliminator ();
343 private:
344 struct address_info *m_ad;
345 rtx *m_base_loc;
346 rtx m_base_reg;
347 rtx *m_index_loc;
348 rtx m_index_reg;
352 address_eliminator::address_eliminator (struct address_info *ad)
353 : m_ad (ad),
354 m_base_loc (strip_subreg (ad->base_term)),
355 m_base_reg (NULL_RTX),
356 m_index_loc (strip_subreg (ad->index_term)),
357 m_index_reg (NULL_RTX)
359 if (m_base_loc != NULL)
361 m_base_reg = *m_base_loc;
362 lra_eliminate_reg_if_possible (m_base_loc);
363 if (m_ad->base_term2 != NULL)
364 *m_ad->base_term2 = *m_ad->base_term;
366 if (m_index_loc != NULL)
368 m_index_reg = *m_index_loc;
369 lra_eliminate_reg_if_possible (m_index_loc);
373 address_eliminator::~address_eliminator ()
375 if (m_base_loc && *m_base_loc != m_base_reg)
377 *m_base_loc = m_base_reg;
378 if (m_ad->base_term2 != NULL)
379 *m_ad->base_term2 = *m_ad->base_term;
381 if (m_index_loc && *m_index_loc != m_index_reg)
382 *m_index_loc = m_index_reg;
385 /* Return true if the eliminated form of AD is a legitimate target address. */
386 static bool
387 valid_address_p (struct address_info *ad)
389 address_eliminator eliminator (ad);
390 return valid_address_p (ad->mode, *ad->outer, ad->as);
393 /* Return true if the eliminated form of memory reference OP satisfies
394 extra (special) memory constraint CONSTRAINT. */
395 static bool
396 satisfies_memory_constraint_p (rtx op, enum constraint_num constraint)
398 struct address_info ad;
400 decompose_mem_address (&ad, op);
401 address_eliminator eliminator (&ad);
402 return constraint_satisfied_p (op, constraint);
405 /* Return true if the eliminated form of address AD satisfies extra
406 address constraint CONSTRAINT. */
407 static bool
408 satisfies_address_constraint_p (struct address_info *ad,
409 enum constraint_num constraint)
411 address_eliminator eliminator (ad);
412 return constraint_satisfied_p (*ad->outer, constraint);
415 /* Return true if the eliminated form of address OP satisfies extra
416 address constraint CONSTRAINT. */
417 static bool
418 satisfies_address_constraint_p (rtx op, enum constraint_num constraint)
420 struct address_info ad;
422 decompose_lea_address (&ad, &op);
423 return satisfies_address_constraint_p (&ad, constraint);
426 /* Initiate equivalences for LRA. As we keep original equivalences
427 before any elimination, we need to make copies otherwise any change
428 in insns might change the equivalences. */
429 void
430 lra_init_equiv (void)
432 ira_expand_reg_equiv ();
433 for (int i = FIRST_PSEUDO_REGISTER; i < max_reg_num (); i++)
435 rtx res;
437 if ((res = ira_reg_equiv[i].memory) != NULL_RTX)
438 ira_reg_equiv[i].memory = copy_rtx (res);
439 if ((res = ira_reg_equiv[i].invariant) != NULL_RTX)
440 ira_reg_equiv[i].invariant = copy_rtx (res);
444 static rtx loc_equivalence_callback (rtx, const_rtx, void *);
446 /* Update equivalence for REGNO. We need to this as the equivalence
447 might contain other pseudos which are changed by their
448 equivalences. */
449 static void
450 update_equiv (int regno)
452 rtx x;
454 if ((x = ira_reg_equiv[regno].memory) != NULL_RTX)
455 ira_reg_equiv[regno].memory
456 = simplify_replace_fn_rtx (x, NULL_RTX, loc_equivalence_callback,
457 NULL_RTX);
458 if ((x = ira_reg_equiv[regno].invariant) != NULL_RTX)
459 ira_reg_equiv[regno].invariant
460 = simplify_replace_fn_rtx (x, NULL_RTX, loc_equivalence_callback,
461 NULL_RTX);
464 /* If we have decided to substitute X with another value, return that
465 value, otherwise return X. */
466 static rtx
467 get_equiv (rtx x)
469 int regno;
470 rtx res;
472 if (! REG_P (x) || (regno = REGNO (x)) < FIRST_PSEUDO_REGISTER
473 || ! ira_reg_equiv[regno].defined_p
474 || ! ira_reg_equiv[regno].profitable_p
475 || lra_get_regno_hard_regno (regno) >= 0)
476 return x;
477 if ((res = ira_reg_equiv[regno].memory) != NULL_RTX)
479 if (targetm.cannot_substitute_mem_equiv_p (res))
480 return x;
481 return res;
483 if ((res = ira_reg_equiv[regno].constant) != NULL_RTX)
484 return res;
485 if ((res = ira_reg_equiv[regno].invariant) != NULL_RTX)
486 return res;
487 gcc_unreachable ();
490 /* If we have decided to substitute X with the equivalent value,
491 return that value after elimination for INSN, otherwise return
492 X. */
493 static rtx
494 get_equiv_with_elimination (rtx x, rtx_insn *insn)
496 rtx res = get_equiv (x);
498 if (x == res || CONSTANT_P (res))
499 return res;
500 return lra_eliminate_regs_1 (insn, res, GET_MODE (res),
501 false, false, 0, true);
504 /* Set up curr_operand_mode. */
505 static void
506 init_curr_operand_mode (void)
508 int nop = curr_static_id->n_operands;
509 for (int i = 0; i < nop; i++)
511 machine_mode mode = GET_MODE (*curr_id->operand_loc[i]);
512 if (mode == VOIDmode)
514 /* The .md mode for address operands is the mode of the
515 addressed value rather than the mode of the address itself. */
516 if (curr_id->icode >= 0 && curr_static_id->operand[i].is_address)
517 mode = Pmode;
518 else
519 mode = curr_static_id->operand[i].mode;
521 curr_operand_mode[i] = mode;
527 /* The page contains code to reuse input reloads. */
529 /* Structure describes input reload of the current insns. */
530 struct input_reload
532 /* True for input reload of matched operands. */
533 bool match_p;
534 /* Reloaded value. */
535 rtx input;
536 /* Reload pseudo used. */
537 rtx reg;
540 /* The number of elements in the following array. */
541 static int curr_insn_input_reloads_num;
542 /* Array containing info about input reloads. It is used to find the
543 same input reload and reuse the reload pseudo in this case. */
544 static struct input_reload curr_insn_input_reloads[LRA_MAX_INSN_RELOADS];
546 /* Initiate data concerning reuse of input reloads for the current
547 insn. */
548 static void
549 init_curr_insn_input_reloads (void)
551 curr_insn_input_reloads_num = 0;
554 /* Create a new pseudo using MODE, RCLASS, ORIGINAL or reuse already
555 created input reload pseudo (only if TYPE is not OP_OUT). Don't
556 reuse pseudo if IN_SUBREG_P is true and the reused pseudo should be
557 wrapped up in SUBREG. The result pseudo is returned through
558 RESULT_REG. Return TRUE if we created a new pseudo, FALSE if we
559 reused the already created input reload pseudo. Use TITLE to
560 describe new registers for debug purposes. */
561 static bool
562 get_reload_reg (enum op_type type, machine_mode mode, rtx original,
563 enum reg_class rclass, bool in_subreg_p,
564 const char *title, rtx *result_reg)
566 int i, regno;
567 enum reg_class new_class;
568 bool unique_p = false;
570 if (type == OP_OUT)
572 *result_reg
573 = lra_create_new_reg_with_unique_value (mode, original, rclass, title);
574 return true;
576 /* Prevent reuse value of expression with side effects,
577 e.g. volatile memory. */
578 if (! side_effects_p (original))
579 for (i = 0; i < curr_insn_input_reloads_num; i++)
581 if (! curr_insn_input_reloads[i].match_p
582 && rtx_equal_p (curr_insn_input_reloads[i].input, original)
583 && in_class_p (curr_insn_input_reloads[i].reg, rclass, &new_class))
585 rtx reg = curr_insn_input_reloads[i].reg;
586 regno = REGNO (reg);
587 /* If input is equal to original and both are VOIDmode,
588 GET_MODE (reg) might be still different from mode.
589 Ensure we don't return *result_reg with wrong mode. */
590 if (GET_MODE (reg) != mode)
592 if (in_subreg_p)
593 continue;
594 if (GET_MODE_SIZE (GET_MODE (reg)) < GET_MODE_SIZE (mode))
595 continue;
596 reg = lowpart_subreg (mode, reg, GET_MODE (reg));
597 if (reg == NULL_RTX || GET_CODE (reg) != SUBREG)
598 continue;
600 *result_reg = reg;
601 if (lra_dump_file != NULL)
603 fprintf (lra_dump_file, " Reuse r%d for reload ", regno);
604 dump_value_slim (lra_dump_file, original, 1);
606 if (new_class != lra_get_allocno_class (regno))
607 lra_change_class (regno, new_class, ", change to", false);
608 if (lra_dump_file != NULL)
609 fprintf (lra_dump_file, "\n");
610 return false;
612 /* If we have an input reload with a different mode, make sure it
613 will get a different hard reg. */
614 else if (REG_P (original)
615 && REG_P (curr_insn_input_reloads[i].input)
616 && REGNO (original) == REGNO (curr_insn_input_reloads[i].input)
617 && (GET_MODE (original)
618 != GET_MODE (curr_insn_input_reloads[i].input)))
619 unique_p = true;
621 *result_reg = (unique_p
622 ? lra_create_new_reg_with_unique_value
623 : lra_create_new_reg) (mode, original, rclass, title);
624 lra_assert (curr_insn_input_reloads_num < LRA_MAX_INSN_RELOADS);
625 curr_insn_input_reloads[curr_insn_input_reloads_num].input = original;
626 curr_insn_input_reloads[curr_insn_input_reloads_num].match_p = false;
627 curr_insn_input_reloads[curr_insn_input_reloads_num++].reg = *result_reg;
628 return true;
633 /* The page contains code to extract memory address parts. */
635 /* Wrapper around REGNO_OK_FOR_INDEX_P, to allow pseudos. */
636 static inline bool
637 ok_for_index_p_nonstrict (rtx reg)
639 unsigned regno = REGNO (reg);
641 return regno >= FIRST_PSEUDO_REGISTER || REGNO_OK_FOR_INDEX_P (regno);
644 /* A version of regno_ok_for_base_p for use here, when all pseudos
645 should count as OK. Arguments as for regno_ok_for_base_p. */
646 static inline bool
647 ok_for_base_p_nonstrict (rtx reg, machine_mode mode, addr_space_t as,
648 enum rtx_code outer_code, enum rtx_code index_code)
650 unsigned regno = REGNO (reg);
652 if (regno >= FIRST_PSEUDO_REGISTER)
653 return true;
654 return ok_for_base_p_1 (regno, mode, as, outer_code, index_code);
659 /* The page contains major code to choose the current insn alternative
660 and generate reloads for it. */
662 /* Return the offset from REGNO of the least significant register
663 in (reg:MODE REGNO).
665 This function is used to tell whether two registers satisfy
666 a matching constraint. (reg:MODE1 REGNO1) matches (reg:MODE2 REGNO2) if:
668 REGNO1 + lra_constraint_offset (REGNO1, MODE1)
669 == REGNO2 + lra_constraint_offset (REGNO2, MODE2) */
671 lra_constraint_offset (int regno, machine_mode mode)
673 lra_assert (regno < FIRST_PSEUDO_REGISTER);
674 if (WORDS_BIG_ENDIAN && GET_MODE_SIZE (mode) > UNITS_PER_WORD
675 && SCALAR_INT_MODE_P (mode))
676 return hard_regno_nregs[regno][mode] - 1;
677 return 0;
680 /* Like rtx_equal_p except that it allows a REG and a SUBREG to match
681 if they are the same hard reg, and has special hacks for
682 auto-increment and auto-decrement. This is specifically intended for
683 process_alt_operands to use in determining whether two operands
684 match. X is the operand whose number is the lower of the two.
686 It is supposed that X is the output operand and Y is the input
687 operand. Y_HARD_REGNO is the final hard regno of register Y or
688 register in subreg Y as we know it now. Otherwise, it is a
689 negative value. */
690 static bool
691 operands_match_p (rtx x, rtx y, int y_hard_regno)
693 int i;
694 RTX_CODE code = GET_CODE (x);
695 const char *fmt;
697 if (x == y)
698 return true;
699 if ((code == REG || (code == SUBREG && REG_P (SUBREG_REG (x))))
700 && (REG_P (y) || (GET_CODE (y) == SUBREG && REG_P (SUBREG_REG (y)))))
702 int j;
704 i = get_hard_regno (x, false);
705 if (i < 0)
706 goto slow;
708 if ((j = y_hard_regno) < 0)
709 goto slow;
711 i += lra_constraint_offset (i, GET_MODE (x));
712 j += lra_constraint_offset (j, GET_MODE (y));
714 return i == j;
717 /* If two operands must match, because they are really a single
718 operand of an assembler insn, then two post-increments are invalid
719 because the assembler insn would increment only once. On the
720 other hand, a post-increment matches ordinary indexing if the
721 post-increment is the output operand. */
722 if (code == POST_DEC || code == POST_INC || code == POST_MODIFY)
723 return operands_match_p (XEXP (x, 0), y, y_hard_regno);
725 /* Two pre-increments are invalid because the assembler insn would
726 increment only once. On the other hand, a pre-increment matches
727 ordinary indexing if the pre-increment is the input operand. */
728 if (GET_CODE (y) == PRE_DEC || GET_CODE (y) == PRE_INC
729 || GET_CODE (y) == PRE_MODIFY)
730 return operands_match_p (x, XEXP (y, 0), -1);
732 slow:
734 if (code == REG && REG_P (y))
735 return REGNO (x) == REGNO (y);
737 if (code == REG && GET_CODE (y) == SUBREG && REG_P (SUBREG_REG (y))
738 && x == SUBREG_REG (y))
739 return true;
740 if (GET_CODE (y) == REG && code == SUBREG && REG_P (SUBREG_REG (x))
741 && SUBREG_REG (x) == y)
742 return true;
744 /* Now we have disposed of all the cases in which different rtx
745 codes can match. */
746 if (code != GET_CODE (y))
747 return false;
749 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
750 if (GET_MODE (x) != GET_MODE (y))
751 return false;
753 switch (code)
755 CASE_CONST_UNIQUE:
756 return false;
758 case LABEL_REF:
759 return label_ref_label (x) == label_ref_label (y);
760 case SYMBOL_REF:
761 return XSTR (x, 0) == XSTR (y, 0);
763 default:
764 break;
767 /* Compare the elements. If any pair of corresponding elements fail
768 to match, return false for the whole things. */
770 fmt = GET_RTX_FORMAT (code);
771 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
773 int val, j;
774 switch (fmt[i])
776 case 'w':
777 if (XWINT (x, i) != XWINT (y, i))
778 return false;
779 break;
781 case 'i':
782 if (XINT (x, i) != XINT (y, i))
783 return false;
784 break;
786 case 'e':
787 val = operands_match_p (XEXP (x, i), XEXP (y, i), -1);
788 if (val == 0)
789 return false;
790 break;
792 case '0':
793 break;
795 case 'E':
796 if (XVECLEN (x, i) != XVECLEN (y, i))
797 return false;
798 for (j = XVECLEN (x, i) - 1; j >= 0; --j)
800 val = operands_match_p (XVECEXP (x, i, j), XVECEXP (y, i, j), -1);
801 if (val == 0)
802 return false;
804 break;
806 /* It is believed that rtx's at this level will never
807 contain anything but integers and other rtx's, except for
808 within LABEL_REFs and SYMBOL_REFs. */
809 default:
810 gcc_unreachable ();
813 return true;
816 /* True if X is a constant that can be forced into the constant pool.
817 MODE is the mode of the operand, or VOIDmode if not known. */
818 #define CONST_POOL_OK_P(MODE, X) \
819 ((MODE) != VOIDmode \
820 && CONSTANT_P (X) \
821 && GET_CODE (X) != HIGH \
822 && !targetm.cannot_force_const_mem (MODE, X))
824 /* True if C is a non-empty register class that has too few registers
825 to be safely used as a reload target class. */
826 #define SMALL_REGISTER_CLASS_P(C) \
827 (ira_class_hard_regs_num [(C)] == 1 \
828 || (ira_class_hard_regs_num [(C)] >= 1 \
829 && targetm.class_likely_spilled_p (C)))
831 /* If REG is a reload pseudo, try to make its class satisfying CL. */
832 static void
833 narrow_reload_pseudo_class (rtx reg, enum reg_class cl)
835 enum reg_class rclass;
837 /* Do not make more accurate class from reloads generated. They are
838 mostly moves with a lot of constraints. Making more accurate
839 class may results in very narrow class and impossibility of find
840 registers for several reloads of one insn. */
841 if (INSN_UID (curr_insn) >= new_insn_uid_start)
842 return;
843 if (GET_CODE (reg) == SUBREG)
844 reg = SUBREG_REG (reg);
845 if (! REG_P (reg) || (int) REGNO (reg) < new_regno_start)
846 return;
847 if (in_class_p (reg, cl, &rclass) && rclass != cl)
848 lra_change_class (REGNO (reg), rclass, " Change to", true);
851 /* Searches X for any reference to a reg with the same value as REGNO,
852 returning the rtx of the reference found if any. Otherwise,
853 returns NULL_RTX. */
854 static rtx
855 regno_val_use_in (unsigned int regno, rtx x)
857 const char *fmt;
858 int i, j;
859 rtx tem;
861 if (REG_P (x) && lra_reg_info[REGNO (x)].val == lra_reg_info[regno].val)
862 return x;
864 fmt = GET_RTX_FORMAT (GET_CODE (x));
865 for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
867 if (fmt[i] == 'e')
869 if ((tem = regno_val_use_in (regno, XEXP (x, i))))
870 return tem;
872 else if (fmt[i] == 'E')
873 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
874 if ((tem = regno_val_use_in (regno , XVECEXP (x, i, j))))
875 return tem;
878 return NULL_RTX;
881 /* Return true if all current insn non-output operands except INS (it
882 has a negaitve end marker) do not use pseudos with the same value
883 as REGNO. */
884 static bool
885 check_conflict_input_operands (int regno, signed char *ins)
887 int in;
888 int n_operands = curr_static_id->n_operands;
890 for (int nop = 0; nop < n_operands; nop++)
891 if (! curr_static_id->operand[nop].is_operator
892 && curr_static_id->operand[nop].type != OP_OUT)
894 for (int i = 0; (in = ins[i]) >= 0; i++)
895 if (in == nop)
896 break;
897 if (in < 0
898 && regno_val_use_in (regno, *curr_id->operand_loc[nop]) != NULL_RTX)
899 return false;
901 return true;
904 /* Generate reloads for matching OUT and INS (array of input operand
905 numbers with end marker -1) with reg class GOAL_CLASS, considering
906 output operands OUTS (similar array to INS) needing to be in different
907 registers. Add input and output reloads correspondingly to the lists
908 *BEFORE and *AFTER. OUT might be negative. In this case we generate
909 input reloads for matched input operands INS. EARLY_CLOBBER_P is a flag
910 that the output operand is early clobbered for chosen alternative. */
911 static void
912 match_reload (signed char out, signed char *ins, signed char *outs,
913 enum reg_class goal_class, rtx_insn **before,
914 rtx_insn **after, bool early_clobber_p)
916 bool out_conflict;
917 int i, in;
918 rtx new_in_reg, new_out_reg, reg;
919 machine_mode inmode, outmode;
920 rtx in_rtx = *curr_id->operand_loc[ins[0]];
921 rtx out_rtx = out < 0 ? in_rtx : *curr_id->operand_loc[out];
923 inmode = curr_operand_mode[ins[0]];
924 outmode = out < 0 ? inmode : curr_operand_mode[out];
925 push_to_sequence (*before);
926 if (inmode != outmode)
928 if (GET_MODE_SIZE (inmode) > GET_MODE_SIZE (outmode))
930 reg = new_in_reg
931 = lra_create_new_reg_with_unique_value (inmode, in_rtx,
932 goal_class, "");
933 if (SCALAR_INT_MODE_P (inmode))
934 new_out_reg = gen_lowpart_SUBREG (outmode, reg);
935 else
936 new_out_reg = gen_rtx_SUBREG (outmode, reg, 0);
937 LRA_SUBREG_P (new_out_reg) = 1;
938 /* If the input reg is dying here, we can use the same hard
939 register for REG and IN_RTX. We do it only for original
940 pseudos as reload pseudos can die although original
941 pseudos still live where reload pseudos dies. */
942 if (REG_P (in_rtx) && (int) REGNO (in_rtx) < lra_new_regno_start
943 && find_regno_note (curr_insn, REG_DEAD, REGNO (in_rtx))
944 && (!early_clobber_p
945 || check_conflict_input_operands(REGNO (in_rtx), ins)))
946 lra_assign_reg_val (REGNO (in_rtx), REGNO (reg));
948 else
950 reg = new_out_reg
951 = lra_create_new_reg_with_unique_value (outmode, out_rtx,
952 goal_class, "");
953 if (SCALAR_INT_MODE_P (outmode))
954 new_in_reg = gen_lowpart_SUBREG (inmode, reg);
955 else
956 new_in_reg = gen_rtx_SUBREG (inmode, reg, 0);
957 /* NEW_IN_REG is non-paradoxical subreg. We don't want
958 NEW_OUT_REG living above. We add clobber clause for
959 this. This is just a temporary clobber. We can remove
960 it at the end of LRA work. */
961 rtx_insn *clobber = emit_clobber (new_out_reg);
962 LRA_TEMP_CLOBBER_P (PATTERN (clobber)) = 1;
963 LRA_SUBREG_P (new_in_reg) = 1;
964 if (GET_CODE (in_rtx) == SUBREG)
966 rtx subreg_reg = SUBREG_REG (in_rtx);
968 /* If SUBREG_REG is dying here and sub-registers IN_RTX
969 and NEW_IN_REG are similar, we can use the same hard
970 register for REG and SUBREG_REG. */
971 if (REG_P (subreg_reg)
972 && (int) REGNO (subreg_reg) < lra_new_regno_start
973 && GET_MODE (subreg_reg) == outmode
974 && SUBREG_BYTE (in_rtx) == SUBREG_BYTE (new_in_reg)
975 && find_regno_note (curr_insn, REG_DEAD, REGNO (subreg_reg))
976 && (! early_clobber_p
977 || check_conflict_input_operands (REGNO (subreg_reg),
978 ins)))
979 lra_assign_reg_val (REGNO (subreg_reg), REGNO (reg));
983 else
985 /* Pseudos have values -- see comments for lra_reg_info.
986 Different pseudos with the same value do not conflict even if
987 they live in the same place. When we create a pseudo we
988 assign value of original pseudo (if any) from which we
989 created the new pseudo. If we create the pseudo from the
990 input pseudo, the new pseudo will have no conflict with the
991 input pseudo which is wrong when the input pseudo lives after
992 the insn and as the new pseudo value is changed by the insn
993 output. Therefore we create the new pseudo from the output
994 except the case when we have single matched dying input
995 pseudo.
997 We cannot reuse the current output register because we might
998 have a situation like "a <- a op b", where the constraints
999 force the second input operand ("b") to match the output
1000 operand ("a"). "b" must then be copied into a new register
1001 so that it doesn't clobber the current value of "a".
1003 We can not use the same value if the output pseudo is
1004 early clobbered or the input pseudo is mentioned in the
1005 output, e.g. as an address part in memory, because
1006 output reload will actually extend the pseudo liveness.
1007 We don't care about eliminable hard regs here as we are
1008 interesting only in pseudos. */
1010 /* Matching input's register value is the same as one of the other
1011 output operand. Output operands in a parallel insn must be in
1012 different registers. */
1013 out_conflict = false;
1014 if (REG_P (in_rtx))
1016 for (i = 0; outs[i] >= 0; i++)
1018 rtx other_out_rtx = *curr_id->operand_loc[outs[i]];
1019 if (REG_P (other_out_rtx)
1020 && (regno_val_use_in (REGNO (in_rtx), other_out_rtx)
1021 != NULL_RTX))
1023 out_conflict = true;
1024 break;
1029 new_in_reg = new_out_reg
1030 = (! early_clobber_p && ins[1] < 0 && REG_P (in_rtx)
1031 && (int) REGNO (in_rtx) < lra_new_regno_start
1032 && find_regno_note (curr_insn, REG_DEAD, REGNO (in_rtx))
1033 && (! early_clobber_p
1034 || check_conflict_input_operands (REGNO (in_rtx), ins))
1035 && (out < 0
1036 || regno_val_use_in (REGNO (in_rtx), out_rtx) == NULL_RTX)
1037 && !out_conflict
1038 ? lra_create_new_reg (inmode, in_rtx, goal_class, "")
1039 : lra_create_new_reg_with_unique_value (outmode, out_rtx,
1040 goal_class, ""));
1042 /* In operand can be got from transformations before processing insn
1043 constraints. One example of such transformations is subreg
1044 reloading (see function simplify_operand_subreg). The new
1045 pseudos created by the transformations might have inaccurate
1046 class (ALL_REGS) and we should make their classes more
1047 accurate. */
1048 narrow_reload_pseudo_class (in_rtx, goal_class);
1049 lra_emit_move (copy_rtx (new_in_reg), in_rtx);
1050 *before = get_insns ();
1051 end_sequence ();
1052 /* Add the new pseudo to consider values of subsequent input reload
1053 pseudos. */
1054 lra_assert (curr_insn_input_reloads_num < LRA_MAX_INSN_RELOADS);
1055 curr_insn_input_reloads[curr_insn_input_reloads_num].input = in_rtx;
1056 curr_insn_input_reloads[curr_insn_input_reloads_num].match_p = true;
1057 curr_insn_input_reloads[curr_insn_input_reloads_num++].reg = new_in_reg;
1058 for (i = 0; (in = ins[i]) >= 0; i++)
1060 lra_assert
1061 (GET_MODE (*curr_id->operand_loc[in]) == VOIDmode
1062 || GET_MODE (new_in_reg) == GET_MODE (*curr_id->operand_loc[in]));
1063 *curr_id->operand_loc[in] = new_in_reg;
1065 lra_update_dups (curr_id, ins);
1066 if (out < 0)
1067 return;
1068 /* See a comment for the input operand above. */
1069 narrow_reload_pseudo_class (out_rtx, goal_class);
1070 if (find_reg_note (curr_insn, REG_UNUSED, out_rtx) == NULL_RTX)
1072 start_sequence ();
1073 lra_emit_move (out_rtx, copy_rtx (new_out_reg));
1074 emit_insn (*after);
1075 *after = get_insns ();
1076 end_sequence ();
1078 *curr_id->operand_loc[out] = new_out_reg;
1079 lra_update_dup (curr_id, out);
1082 /* Return register class which is union of all reg classes in insn
1083 constraint alternative string starting with P. */
1084 static enum reg_class
1085 reg_class_from_constraints (const char *p)
1087 int c, len;
1088 enum reg_class op_class = NO_REGS;
1091 switch ((c = *p, len = CONSTRAINT_LEN (c, p)), c)
1093 case '#':
1094 case ',':
1095 return op_class;
1097 case 'g':
1098 op_class = reg_class_subunion[op_class][GENERAL_REGS];
1099 break;
1101 default:
1102 enum constraint_num cn = lookup_constraint (p);
1103 enum reg_class cl = reg_class_for_constraint (cn);
1104 if (cl == NO_REGS)
1106 if (insn_extra_address_constraint (cn))
1107 op_class
1108 = (reg_class_subunion
1109 [op_class][base_reg_class (VOIDmode, ADDR_SPACE_GENERIC,
1110 ADDRESS, SCRATCH)]);
1111 break;
1114 op_class = reg_class_subunion[op_class][cl];
1115 break;
1117 while ((p += len), c);
1118 return op_class;
1121 /* If OP is a register, return the class of the register as per
1122 get_reg_class, otherwise return NO_REGS. */
1123 static inline enum reg_class
1124 get_op_class (rtx op)
1126 return REG_P (op) ? get_reg_class (REGNO (op)) : NO_REGS;
1129 /* Return generated insn mem_pseudo:=val if TO_P or val:=mem_pseudo
1130 otherwise. If modes of MEM_PSEUDO and VAL are different, use
1131 SUBREG for VAL to make them equal. */
1132 static rtx_insn *
1133 emit_spill_move (bool to_p, rtx mem_pseudo, rtx val)
1135 if (GET_MODE (mem_pseudo) != GET_MODE (val))
1137 /* Usually size of mem_pseudo is greater than val size but in
1138 rare cases it can be less as it can be defined by target
1139 dependent macro HARD_REGNO_CALLER_SAVE_MODE. */
1140 if (! MEM_P (val))
1142 val = gen_lowpart_SUBREG (GET_MODE (mem_pseudo),
1143 GET_CODE (val) == SUBREG
1144 ? SUBREG_REG (val) : val);
1145 LRA_SUBREG_P (val) = 1;
1147 else
1149 mem_pseudo = gen_lowpart_SUBREG (GET_MODE (val), mem_pseudo);
1150 LRA_SUBREG_P (mem_pseudo) = 1;
1153 return to_p ? gen_move_insn (mem_pseudo, val)
1154 : gen_move_insn (val, mem_pseudo);
1157 /* Process a special case insn (register move), return true if we
1158 don't need to process it anymore. INSN should be a single set
1159 insn. Set up that RTL was changed through CHANGE_P and macro
1160 SECONDARY_MEMORY_NEEDED says to use secondary memory through
1161 SEC_MEM_P. */
1162 static bool
1163 check_and_process_move (bool *change_p, bool *sec_mem_p ATTRIBUTE_UNUSED)
1165 int sregno, dregno;
1166 rtx dest, src, dreg, sreg, new_reg, scratch_reg;
1167 rtx_insn *before;
1168 enum reg_class dclass, sclass, secondary_class;
1169 secondary_reload_info sri;
1171 lra_assert (curr_insn_set != NULL_RTX);
1172 dreg = dest = SET_DEST (curr_insn_set);
1173 sreg = src = SET_SRC (curr_insn_set);
1174 if (GET_CODE (dest) == SUBREG)
1175 dreg = SUBREG_REG (dest);
1176 if (GET_CODE (src) == SUBREG)
1177 sreg = SUBREG_REG (src);
1178 if (! (REG_P (dreg) || MEM_P (dreg)) || ! (REG_P (sreg) || MEM_P (sreg)))
1179 return false;
1180 sclass = dclass = NO_REGS;
1181 if (REG_P (dreg))
1182 dclass = get_reg_class (REGNO (dreg));
1183 gcc_assert (dclass < LIM_REG_CLASSES);
1184 if (dclass == ALL_REGS)
1185 /* ALL_REGS is used for new pseudos created by transformations
1186 like reload of SUBREG_REG (see function
1187 simplify_operand_subreg). We don't know their class yet. We
1188 should figure out the class from processing the insn
1189 constraints not in this fast path function. Even if ALL_REGS
1190 were a right class for the pseudo, secondary_... hooks usually
1191 are not define for ALL_REGS. */
1192 return false;
1193 if (REG_P (sreg))
1194 sclass = get_reg_class (REGNO (sreg));
1195 gcc_assert (sclass < LIM_REG_CLASSES);
1196 if (sclass == ALL_REGS)
1197 /* See comments above. */
1198 return false;
1199 if (sclass == NO_REGS && dclass == NO_REGS)
1200 return false;
1201 #ifdef SECONDARY_MEMORY_NEEDED
1202 if (SECONDARY_MEMORY_NEEDED (sclass, dclass, GET_MODE (src))
1203 #ifdef SECONDARY_MEMORY_NEEDED_MODE
1204 && ((sclass != NO_REGS && dclass != NO_REGS)
1205 || GET_MODE (src) != SECONDARY_MEMORY_NEEDED_MODE (GET_MODE (src)))
1206 #endif
1209 *sec_mem_p = true;
1210 return false;
1212 #endif
1213 if (! REG_P (dreg) || ! REG_P (sreg))
1214 return false;
1215 sri.prev_sri = NULL;
1216 sri.icode = CODE_FOR_nothing;
1217 sri.extra_cost = 0;
1218 secondary_class = NO_REGS;
1219 /* Set up hard register for a reload pseudo for hook
1220 secondary_reload because some targets just ignore unassigned
1221 pseudos in the hook. */
1222 if (dclass != NO_REGS && lra_get_regno_hard_regno (REGNO (dreg)) < 0)
1224 dregno = REGNO (dreg);
1225 reg_renumber[dregno] = ira_class_hard_regs[dclass][0];
1227 else
1228 dregno = -1;
1229 if (sclass != NO_REGS && lra_get_regno_hard_regno (REGNO (sreg)) < 0)
1231 sregno = REGNO (sreg);
1232 reg_renumber[sregno] = ira_class_hard_regs[sclass][0];
1234 else
1235 sregno = -1;
1236 if (sclass != NO_REGS)
1237 secondary_class
1238 = (enum reg_class) targetm.secondary_reload (false, dest,
1239 (reg_class_t) sclass,
1240 GET_MODE (src), &sri);
1241 if (sclass == NO_REGS
1242 || ((secondary_class != NO_REGS || sri.icode != CODE_FOR_nothing)
1243 && dclass != NO_REGS))
1245 enum reg_class old_sclass = secondary_class;
1246 secondary_reload_info old_sri = sri;
1248 sri.prev_sri = NULL;
1249 sri.icode = CODE_FOR_nothing;
1250 sri.extra_cost = 0;
1251 secondary_class
1252 = (enum reg_class) targetm.secondary_reload (true, src,
1253 (reg_class_t) dclass,
1254 GET_MODE (src), &sri);
1255 /* Check the target hook consistency. */
1256 lra_assert
1257 ((secondary_class == NO_REGS && sri.icode == CODE_FOR_nothing)
1258 || (old_sclass == NO_REGS && old_sri.icode == CODE_FOR_nothing)
1259 || (secondary_class == old_sclass && sri.icode == old_sri.icode));
1261 if (sregno >= 0)
1262 reg_renumber [sregno] = -1;
1263 if (dregno >= 0)
1264 reg_renumber [dregno] = -1;
1265 if (secondary_class == NO_REGS && sri.icode == CODE_FOR_nothing)
1266 return false;
1267 *change_p = true;
1268 new_reg = NULL_RTX;
1269 if (secondary_class != NO_REGS)
1270 new_reg = lra_create_new_reg_with_unique_value (GET_MODE (src), NULL_RTX,
1271 secondary_class,
1272 "secondary");
1273 start_sequence ();
1274 if (sri.icode == CODE_FOR_nothing)
1275 lra_emit_move (new_reg, src);
1276 else
1278 enum reg_class scratch_class;
1280 scratch_class = (reg_class_from_constraints
1281 (insn_data[sri.icode].operand[2].constraint));
1282 scratch_reg = (lra_create_new_reg_with_unique_value
1283 (insn_data[sri.icode].operand[2].mode, NULL_RTX,
1284 scratch_class, "scratch"));
1285 emit_insn (GEN_FCN (sri.icode) (new_reg != NULL_RTX ? new_reg : dest,
1286 src, scratch_reg));
1288 before = get_insns ();
1289 end_sequence ();
1290 lra_process_new_insns (curr_insn, before, NULL, "Inserting the move");
1291 if (new_reg != NULL_RTX)
1292 SET_SRC (curr_insn_set) = new_reg;
1293 else
1295 if (lra_dump_file != NULL)
1297 fprintf (lra_dump_file, "Deleting move %u\n", INSN_UID (curr_insn));
1298 dump_insn_slim (lra_dump_file, curr_insn);
1300 lra_set_insn_deleted (curr_insn);
1301 return true;
1303 return false;
1306 /* The following data describe the result of process_alt_operands.
1307 The data are used in curr_insn_transform to generate reloads. */
1309 /* The chosen reg classes which should be used for the corresponding
1310 operands. */
1311 static enum reg_class goal_alt[MAX_RECOG_OPERANDS];
1312 /* True if the operand should be the same as another operand and that
1313 other operand does not need a reload. */
1314 static bool goal_alt_match_win[MAX_RECOG_OPERANDS];
1315 /* True if the operand does not need a reload. */
1316 static bool goal_alt_win[MAX_RECOG_OPERANDS];
1317 /* True if the operand can be offsetable memory. */
1318 static bool goal_alt_offmemok[MAX_RECOG_OPERANDS];
1319 /* The number of an operand to which given operand can be matched to. */
1320 static int goal_alt_matches[MAX_RECOG_OPERANDS];
1321 /* The number of elements in the following array. */
1322 static int goal_alt_dont_inherit_ops_num;
1323 /* Numbers of operands whose reload pseudos should not be inherited. */
1324 static int goal_alt_dont_inherit_ops[MAX_RECOG_OPERANDS];
1325 /* True if the insn commutative operands should be swapped. */
1326 static bool goal_alt_swapped;
1327 /* The chosen insn alternative. */
1328 static int goal_alt_number;
1330 /* True if the corresponding operand is the result of an equivalence
1331 substitution. */
1332 static bool equiv_substition_p[MAX_RECOG_OPERANDS];
1334 /* The following five variables are used to choose the best insn
1335 alternative. They reflect final characteristics of the best
1336 alternative. */
1338 /* Number of necessary reloads and overall cost reflecting the
1339 previous value and other unpleasantness of the best alternative. */
1340 static int best_losers, best_overall;
1341 /* Overall number hard registers used for reloads. For example, on
1342 some targets we need 2 general registers to reload DFmode and only
1343 one floating point register. */
1344 static int best_reload_nregs;
1345 /* Overall number reflecting distances of previous reloading the same
1346 value. The distances are counted from the current BB start. It is
1347 used to improve inheritance chances. */
1348 static int best_reload_sum;
1350 /* True if the current insn should have no correspondingly input or
1351 output reloads. */
1352 static bool no_input_reloads_p, no_output_reloads_p;
1354 /* True if we swapped the commutative operands in the current
1355 insn. */
1356 static int curr_swapped;
1358 /* if CHECK_ONLY_P is false, arrange for address element *LOC to be a
1359 register of class CL. Add any input reloads to list BEFORE. AFTER
1360 is nonnull if *LOC is an automodified value; handle that case by
1361 adding the required output reloads to list AFTER. Return true if
1362 the RTL was changed.
1364 if CHECK_ONLY_P is true, check that the *LOC is a correct address
1365 register. Return false if the address register is correct. */
1366 static bool
1367 process_addr_reg (rtx *loc, bool check_only_p, rtx_insn **before, rtx_insn **after,
1368 enum reg_class cl)
1370 int regno;
1371 enum reg_class rclass, new_class;
1372 rtx reg;
1373 rtx new_reg;
1374 machine_mode mode;
1375 bool subreg_p, before_p = false;
1377 subreg_p = GET_CODE (*loc) == SUBREG;
1378 if (subreg_p)
1380 reg = SUBREG_REG (*loc);
1381 mode = GET_MODE (reg);
1383 /* For mode with size bigger than ptr_mode, there unlikely to be "mov"
1384 between two registers with different classes, but there normally will
1385 be "mov" which transfers element of vector register into the general
1386 register, and this normally will be a subreg which should be reloaded
1387 as a whole. This is particularly likely to be triggered when
1388 -fno-split-wide-types specified. */
1389 if (!REG_P (reg)
1390 || in_class_p (reg, cl, &new_class)
1391 || GET_MODE_SIZE (mode) <= GET_MODE_SIZE (ptr_mode))
1392 loc = &SUBREG_REG (*loc);
1395 reg = *loc;
1396 mode = GET_MODE (reg);
1397 if (! REG_P (reg))
1399 if (check_only_p)
1400 return true;
1401 /* Always reload memory in an address even if the target supports
1402 such addresses. */
1403 new_reg = lra_create_new_reg_with_unique_value (mode, reg, cl, "address");
1404 before_p = true;
1406 else
1408 regno = REGNO (reg);
1409 rclass = get_reg_class (regno);
1410 if (! check_only_p
1411 && (*loc = get_equiv_with_elimination (reg, curr_insn)) != reg)
1413 if (lra_dump_file != NULL)
1415 fprintf (lra_dump_file,
1416 "Changing pseudo %d in address of insn %u on equiv ",
1417 REGNO (reg), INSN_UID (curr_insn));
1418 dump_value_slim (lra_dump_file, *loc, 1);
1419 fprintf (lra_dump_file, "\n");
1421 *loc = copy_rtx (*loc);
1423 if (*loc != reg || ! in_class_p (reg, cl, &new_class))
1425 if (check_only_p)
1426 return true;
1427 reg = *loc;
1428 if (get_reload_reg (after == NULL ? OP_IN : OP_INOUT,
1429 mode, reg, cl, subreg_p, "address", &new_reg))
1430 before_p = true;
1432 else if (new_class != NO_REGS && rclass != new_class)
1434 if (check_only_p)
1435 return true;
1436 lra_change_class (regno, new_class, " Change to", true);
1437 return false;
1439 else
1440 return false;
1442 if (before_p)
1444 push_to_sequence (*before);
1445 lra_emit_move (new_reg, reg);
1446 *before = get_insns ();
1447 end_sequence ();
1449 *loc = new_reg;
1450 if (after != NULL)
1452 start_sequence ();
1453 lra_emit_move (before_p ? copy_rtx (reg) : reg, new_reg);
1454 emit_insn (*after);
1455 *after = get_insns ();
1456 end_sequence ();
1458 return true;
1461 /* Insert move insn in simplify_operand_subreg. BEFORE returns
1462 the insn to be inserted before curr insn. AFTER returns the
1463 the insn to be inserted after curr insn. ORIGREG and NEWREG
1464 are the original reg and new reg for reload. */
1465 static void
1466 insert_move_for_subreg (rtx_insn **before, rtx_insn **after, rtx origreg,
1467 rtx newreg)
1469 if (before)
1471 push_to_sequence (*before);
1472 lra_emit_move (newreg, origreg);
1473 *before = get_insns ();
1474 end_sequence ();
1476 if (after)
1478 start_sequence ();
1479 lra_emit_move (origreg, newreg);
1480 emit_insn (*after);
1481 *after = get_insns ();
1482 end_sequence ();
1486 static int valid_address_p (machine_mode mode, rtx addr, addr_space_t as);
1487 static bool process_address (int, bool, rtx_insn **, rtx_insn **);
1489 /* Make reloads for subreg in operand NOP with internal subreg mode
1490 REG_MODE, add new reloads for further processing. Return true if
1491 any change was done. */
1492 static bool
1493 simplify_operand_subreg (int nop, machine_mode reg_mode)
1495 int hard_regno;
1496 rtx_insn *before, *after;
1497 machine_mode mode, innermode;
1498 rtx reg, new_reg;
1499 rtx operand = *curr_id->operand_loc[nop];
1500 enum reg_class regclass;
1501 enum op_type type;
1503 before = after = NULL;
1505 if (GET_CODE (operand) != SUBREG)
1506 return false;
1508 mode = GET_MODE (operand);
1509 reg = SUBREG_REG (operand);
1510 innermode = GET_MODE (reg);
1511 type = curr_static_id->operand[nop].type;
1512 if (MEM_P (reg))
1514 const bool addr_was_valid
1515 = valid_address_p (innermode, XEXP (reg, 0), MEM_ADDR_SPACE (reg));
1516 alter_subreg (curr_id->operand_loc[nop], false);
1517 rtx subst = *curr_id->operand_loc[nop];
1518 lra_assert (MEM_P (subst));
1520 if (!addr_was_valid
1521 || valid_address_p (GET_MODE (subst), XEXP (subst, 0),
1522 MEM_ADDR_SPACE (subst))
1523 || ((get_constraint_type (lookup_constraint
1524 (curr_static_id->operand[nop].constraint))
1525 != CT_SPECIAL_MEMORY)
1526 /* We still can reload address and if the address is
1527 valid, we can remove subreg without reloading its
1528 inner memory. */
1529 && valid_address_p (GET_MODE (subst),
1530 regno_reg_rtx
1531 [ira_class_hard_regs
1532 [base_reg_class (GET_MODE (subst),
1533 MEM_ADDR_SPACE (subst),
1534 ADDRESS, SCRATCH)][0]],
1535 MEM_ADDR_SPACE (subst))))
1537 /* If we change the address for a paradoxical subreg of memory, the
1538 new address might violate the necessary alignment or the access
1539 might be slow; take this into consideration. We need not worry
1540 about accesses beyond allocated memory for paradoxical memory
1541 subregs as we don't substitute such equiv memory (see processing
1542 equivalences in function lra_constraints) and because for spilled
1543 pseudos we allocate stack memory enough for the biggest
1544 corresponding paradoxical subreg.
1546 However, do not blindly simplify a (subreg (mem ...)) for
1547 WORD_REGISTER_OPERATIONS targets as this may lead to loading junk
1548 data into a register when the inner is narrower than outer or
1549 missing important data from memory when the inner is wider than
1550 outer. This rule only applies to modes that are no wider than
1551 a word. */
1552 if (!(GET_MODE_PRECISION (mode) != GET_MODE_PRECISION (innermode)
1553 && GET_MODE_SIZE (mode) <= UNITS_PER_WORD
1554 && GET_MODE_SIZE (innermode) <= UNITS_PER_WORD
1555 && WORD_REGISTER_OPERATIONS)
1556 && (!(MEM_ALIGN (subst) < GET_MODE_ALIGNMENT (mode)
1557 && SLOW_UNALIGNED_ACCESS (mode, MEM_ALIGN (subst)))
1558 || (MEM_ALIGN (reg) < GET_MODE_ALIGNMENT (innermode)
1559 && SLOW_UNALIGNED_ACCESS (innermode, MEM_ALIGN (reg)))))
1560 return true;
1562 *curr_id->operand_loc[nop] = operand;
1564 /* But if the address was not valid, we cannot reload the MEM without
1565 reloading the address first. */
1566 if (!addr_was_valid)
1567 process_address (nop, false, &before, &after);
1569 /* INNERMODE is fast, MODE slow. Reload the mem in INNERMODE. */
1570 enum reg_class rclass
1571 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1572 if (get_reload_reg (curr_static_id->operand[nop].type, innermode,
1573 reg, rclass, TRUE, "slow mem", &new_reg))
1575 bool insert_before, insert_after;
1576 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1578 insert_before = (type != OP_OUT
1579 || GET_MODE_SIZE (innermode)
1580 > GET_MODE_SIZE (mode));
1581 insert_after = type != OP_IN;
1582 insert_move_for_subreg (insert_before ? &before : NULL,
1583 insert_after ? &after : NULL,
1584 reg, new_reg);
1586 SUBREG_REG (operand) = new_reg;
1588 /* Convert to MODE. */
1589 reg = operand;
1590 rclass
1591 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1592 if (get_reload_reg (curr_static_id->operand[nop].type, mode, reg,
1593 rclass, TRUE, "slow mem", &new_reg))
1595 bool insert_before, insert_after;
1596 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1598 insert_before = type != OP_OUT;
1599 insert_after = type != OP_IN;
1600 insert_move_for_subreg (insert_before ? &before : NULL,
1601 insert_after ? &after : NULL,
1602 reg, new_reg);
1604 *curr_id->operand_loc[nop] = new_reg;
1605 lra_process_new_insns (curr_insn, before, after,
1606 "Inserting slow mem reload");
1607 return true;
1610 /* If the address was valid and became invalid, prefer to reload
1611 the memory. Typical case is when the index scale should
1612 correspond the memory. */
1613 *curr_id->operand_loc[nop] = operand;
1614 /* Do not return false here as the MEM_P (reg) will be processed
1615 later in this function. */
1617 else if (REG_P (reg) && REGNO (reg) < FIRST_PSEUDO_REGISTER)
1619 alter_subreg (curr_id->operand_loc[nop], false);
1620 return true;
1622 else if (CONSTANT_P (reg))
1624 /* Try to simplify subreg of constant. It is usually result of
1625 equivalence substitution. */
1626 if (innermode == VOIDmode
1627 && (innermode = original_subreg_reg_mode[nop]) == VOIDmode)
1628 innermode = curr_static_id->operand[nop].mode;
1629 if ((new_reg = simplify_subreg (mode, reg, innermode,
1630 SUBREG_BYTE (operand))) != NULL_RTX)
1632 *curr_id->operand_loc[nop] = new_reg;
1633 return true;
1636 /* Put constant into memory when we have mixed modes. It generates
1637 a better code in most cases as it does not need a secondary
1638 reload memory. It also prevents LRA looping when LRA is using
1639 secondary reload memory again and again. */
1640 if (CONSTANT_P (reg) && CONST_POOL_OK_P (reg_mode, reg)
1641 && SCALAR_INT_MODE_P (reg_mode) != SCALAR_INT_MODE_P (mode))
1643 SUBREG_REG (operand) = force_const_mem (reg_mode, reg);
1644 alter_subreg (curr_id->operand_loc[nop], false);
1645 return true;
1647 /* Force a reload of the SUBREG_REG if this is a constant or PLUS or
1648 if there may be a problem accessing OPERAND in the outer
1649 mode. */
1650 if ((REG_P (reg)
1651 && REGNO (reg) >= FIRST_PSEUDO_REGISTER
1652 && (hard_regno = lra_get_regno_hard_regno (REGNO (reg))) >= 0
1653 /* Don't reload paradoxical subregs because we could be looping
1654 having repeatedly final regno out of hard regs range. */
1655 && (hard_regno_nregs[hard_regno][innermode]
1656 >= hard_regno_nregs[hard_regno][mode])
1657 && simplify_subreg_regno (hard_regno, innermode,
1658 SUBREG_BYTE (operand), mode) < 0
1659 /* Don't reload subreg for matching reload. It is actually
1660 valid subreg in LRA. */
1661 && ! LRA_SUBREG_P (operand))
1662 || CONSTANT_P (reg) || GET_CODE (reg) == PLUS || MEM_P (reg))
1664 enum reg_class rclass;
1666 if (REG_P (reg))
1667 /* There is a big probability that we will get the same class
1668 for the new pseudo and we will get the same insn which
1669 means infinite looping. So spill the new pseudo. */
1670 rclass = NO_REGS;
1671 else
1672 /* The class will be defined later in curr_insn_transform. */
1673 rclass
1674 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1676 if (get_reload_reg (curr_static_id->operand[nop].type, reg_mode, reg,
1677 rclass, TRUE, "subreg reg", &new_reg))
1679 bool insert_before, insert_after;
1680 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1682 insert_before = (type != OP_OUT
1683 || GET_MODE_SIZE (innermode) > GET_MODE_SIZE (mode));
1684 insert_after = (type != OP_IN);
1685 insert_move_for_subreg (insert_before ? &before : NULL,
1686 insert_after ? &after : NULL,
1687 reg, new_reg);
1689 SUBREG_REG (operand) = new_reg;
1690 lra_process_new_insns (curr_insn, before, after,
1691 "Inserting subreg reload");
1692 return true;
1694 /* Force a reload for a paradoxical subreg. For paradoxical subreg,
1695 IRA allocates hardreg to the inner pseudo reg according to its mode
1696 instead of the outermode, so the size of the hardreg may not be enough
1697 to contain the outermode operand, in that case we may need to insert
1698 reload for the reg. For the following two types of paradoxical subreg,
1699 we need to insert reload:
1700 1. If the op_type is OP_IN, and the hardreg could not be paired with
1701 other hardreg to contain the outermode operand
1702 (checked by in_hard_reg_set_p), we need to insert the reload.
1703 2. If the op_type is OP_OUT or OP_INOUT.
1705 Here is a paradoxical subreg example showing how the reload is generated:
1707 (insn 5 4 7 2 (set (reg:TI 106 [ __comp ])
1708 (subreg:TI (reg:DI 107 [ __comp ]) 0)) {*movti_internal_rex64}
1710 In IRA, reg107 is allocated to a DImode hardreg. We use x86-64 as example
1711 here, if reg107 is assigned to hardreg R15, because R15 is the last
1712 hardreg, compiler cannot find another hardreg to pair with R15 to
1713 contain TImode data. So we insert a TImode reload reg180 for it.
1714 After reload is inserted:
1716 (insn 283 0 0 (set (subreg:DI (reg:TI 180 [orig:107 __comp ] [107]) 0)
1717 (reg:DI 107 [ __comp ])) -1
1718 (insn 5 4 7 2 (set (reg:TI 106 [ __comp ])
1719 (subreg:TI (reg:TI 180 [orig:107 __comp ] [107]) 0)) {*movti_internal_rex64}
1721 Two reload hard registers will be allocated to reg180 to save TImode data
1722 in LRA_assign. */
1723 else if (REG_P (reg)
1724 && REGNO (reg) >= FIRST_PSEUDO_REGISTER
1725 && (hard_regno = lra_get_regno_hard_regno (REGNO (reg))) >= 0
1726 && (hard_regno_nregs[hard_regno][innermode]
1727 < hard_regno_nregs[hard_regno][mode])
1728 && (regclass = lra_get_allocno_class (REGNO (reg)))
1729 && (type != OP_IN
1730 || !in_hard_reg_set_p (reg_class_contents[regclass],
1731 mode, hard_regno)))
1733 /* The class will be defined later in curr_insn_transform. */
1734 enum reg_class rclass
1735 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1737 if (get_reload_reg (curr_static_id->operand[nop].type, mode, reg,
1738 rclass, TRUE, "paradoxical subreg", &new_reg))
1740 rtx subreg;
1741 bool insert_before, insert_after;
1743 PUT_MODE (new_reg, mode);
1744 subreg = gen_lowpart_SUBREG (innermode, new_reg);
1745 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1747 insert_before = (type != OP_OUT);
1748 insert_after = (type != OP_IN);
1749 insert_move_for_subreg (insert_before ? &before : NULL,
1750 insert_after ? &after : NULL,
1751 reg, subreg);
1753 SUBREG_REG (operand) = new_reg;
1754 lra_process_new_insns (curr_insn, before, after,
1755 "Inserting paradoxical subreg reload");
1756 return true;
1758 return false;
1761 /* Return TRUE if X refers for a hard register from SET. */
1762 static bool
1763 uses_hard_regs_p (rtx x, HARD_REG_SET set)
1765 int i, j, x_hard_regno;
1766 machine_mode mode;
1767 const char *fmt;
1768 enum rtx_code code;
1770 if (x == NULL_RTX)
1771 return false;
1772 code = GET_CODE (x);
1773 mode = GET_MODE (x);
1774 if (code == SUBREG)
1776 x = SUBREG_REG (x);
1777 code = GET_CODE (x);
1778 if (GET_MODE_SIZE (GET_MODE (x)) > GET_MODE_SIZE (mode))
1779 mode = GET_MODE (x);
1782 if (REG_P (x))
1784 x_hard_regno = get_hard_regno (x, true);
1785 return (x_hard_regno >= 0
1786 && overlaps_hard_reg_set_p (set, mode, x_hard_regno));
1788 if (MEM_P (x))
1790 struct address_info ad;
1792 decompose_mem_address (&ad, x);
1793 if (ad.base_term != NULL && uses_hard_regs_p (*ad.base_term, set))
1794 return true;
1795 if (ad.index_term != NULL && uses_hard_regs_p (*ad.index_term, set))
1796 return true;
1798 fmt = GET_RTX_FORMAT (code);
1799 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1801 if (fmt[i] == 'e')
1803 if (uses_hard_regs_p (XEXP (x, i), set))
1804 return true;
1806 else if (fmt[i] == 'E')
1808 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
1809 if (uses_hard_regs_p (XVECEXP (x, i, j), set))
1810 return true;
1813 return false;
1816 /* Return true if OP is a spilled pseudo. */
1817 static inline bool
1818 spilled_pseudo_p (rtx op)
1820 return (REG_P (op)
1821 && REGNO (op) >= FIRST_PSEUDO_REGISTER && in_mem_p (REGNO (op)));
1824 /* Return true if X is a general constant. */
1825 static inline bool
1826 general_constant_p (rtx x)
1828 return CONSTANT_P (x) && (! flag_pic || LEGITIMATE_PIC_OPERAND_P (x));
1831 static bool
1832 reg_in_class_p (rtx reg, enum reg_class cl)
1834 if (cl == NO_REGS)
1835 return get_reg_class (REGNO (reg)) == NO_REGS;
1836 return in_class_p (reg, cl, NULL);
1839 /* Return true if SET of RCLASS contains no hard regs which can be
1840 used in MODE. */
1841 static bool
1842 prohibited_class_reg_set_mode_p (enum reg_class rclass,
1843 HARD_REG_SET &set,
1844 enum machine_mode mode)
1846 HARD_REG_SET temp;
1848 lra_assert (hard_reg_set_subset_p (reg_class_contents[rclass], set));
1849 COPY_HARD_REG_SET (temp, set);
1850 AND_COMPL_HARD_REG_SET (temp, lra_no_alloc_regs);
1851 return (hard_reg_set_subset_p
1852 (temp, ira_prohibited_class_mode_regs[rclass][mode]));
1855 /* Major function to choose the current insn alternative and what
1856 operands should be reloaded and how. If ONLY_ALTERNATIVE is not
1857 negative we should consider only this alternative. Return false if
1858 we can not choose the alternative or find how to reload the
1859 operands. */
1860 static bool
1861 process_alt_operands (int only_alternative)
1863 bool ok_p = false;
1864 int nop, overall, nalt;
1865 int n_alternatives = curr_static_id->n_alternatives;
1866 int n_operands = curr_static_id->n_operands;
1867 /* LOSERS counts the operands that don't fit this alternative and
1868 would require loading. */
1869 int losers;
1870 /* REJECT is a count of how undesirable this alternative says it is
1871 if any reloading is required. If the alternative matches exactly
1872 then REJECT is ignored, but otherwise it gets this much counted
1873 against it in addition to the reloading needed. */
1874 int reject;
1875 int op_reject;
1876 /* The number of elements in the following array. */
1877 int early_clobbered_regs_num;
1878 /* Numbers of operands which are early clobber registers. */
1879 int early_clobbered_nops[MAX_RECOG_OPERANDS];
1880 enum reg_class curr_alt[MAX_RECOG_OPERANDS];
1881 HARD_REG_SET curr_alt_set[MAX_RECOG_OPERANDS];
1882 bool curr_alt_match_win[MAX_RECOG_OPERANDS];
1883 bool curr_alt_win[MAX_RECOG_OPERANDS];
1884 bool curr_alt_offmemok[MAX_RECOG_OPERANDS];
1885 int curr_alt_matches[MAX_RECOG_OPERANDS];
1886 /* The number of elements in the following array. */
1887 int curr_alt_dont_inherit_ops_num;
1888 /* Numbers of operands whose reload pseudos should not be inherited. */
1889 int curr_alt_dont_inherit_ops[MAX_RECOG_OPERANDS];
1890 rtx op;
1891 /* The register when the operand is a subreg of register, otherwise the
1892 operand itself. */
1893 rtx no_subreg_reg_operand[MAX_RECOG_OPERANDS];
1894 /* The register if the operand is a register or subreg of register,
1895 otherwise NULL. */
1896 rtx operand_reg[MAX_RECOG_OPERANDS];
1897 int hard_regno[MAX_RECOG_OPERANDS];
1898 machine_mode biggest_mode[MAX_RECOG_OPERANDS];
1899 int reload_nregs, reload_sum;
1900 bool costly_p;
1901 enum reg_class cl;
1903 /* Calculate some data common for all alternatives to speed up the
1904 function. */
1905 for (nop = 0; nop < n_operands; nop++)
1907 rtx reg;
1909 op = no_subreg_reg_operand[nop] = *curr_id->operand_loc[nop];
1910 /* The real hard regno of the operand after the allocation. */
1911 hard_regno[nop] = get_hard_regno (op, true);
1913 operand_reg[nop] = reg = op;
1914 biggest_mode[nop] = GET_MODE (op);
1915 if (GET_CODE (op) == SUBREG)
1917 operand_reg[nop] = reg = SUBREG_REG (op);
1918 if (GET_MODE_SIZE (biggest_mode[nop])
1919 < GET_MODE_SIZE (GET_MODE (reg)))
1920 biggest_mode[nop] = GET_MODE (reg);
1922 if (! REG_P (reg))
1923 operand_reg[nop] = NULL_RTX;
1924 else if (REGNO (reg) >= FIRST_PSEUDO_REGISTER
1925 || ((int) REGNO (reg)
1926 == lra_get_elimination_hard_regno (REGNO (reg))))
1927 no_subreg_reg_operand[nop] = reg;
1928 else
1929 operand_reg[nop] = no_subreg_reg_operand[nop]
1930 /* Just use natural mode for elimination result. It should
1931 be enough for extra constraints hooks. */
1932 = regno_reg_rtx[hard_regno[nop]];
1935 /* The constraints are made of several alternatives. Each operand's
1936 constraint looks like foo,bar,... with commas separating the
1937 alternatives. The first alternatives for all operands go
1938 together, the second alternatives go together, etc.
1940 First loop over alternatives. */
1941 alternative_mask preferred = curr_id->preferred_alternatives;
1942 if (only_alternative >= 0)
1943 preferred &= ALTERNATIVE_BIT (only_alternative);
1945 for (nalt = 0; nalt < n_alternatives; nalt++)
1947 /* Loop over operands for one constraint alternative. */
1948 if (!TEST_BIT (preferred, nalt))
1949 continue;
1951 overall = losers = reject = reload_nregs = reload_sum = 0;
1952 for (nop = 0; nop < n_operands; nop++)
1954 int inc = (curr_static_id
1955 ->operand_alternative[nalt * n_operands + nop].reject);
1956 if (lra_dump_file != NULL && inc != 0)
1957 fprintf (lra_dump_file,
1958 " Staticly defined alt reject+=%d\n", inc);
1959 reject += inc;
1961 early_clobbered_regs_num = 0;
1963 for (nop = 0; nop < n_operands; nop++)
1965 const char *p;
1966 char *end;
1967 int len, c, m, i, opalt_num, this_alternative_matches;
1968 bool win, did_match, offmemok, early_clobber_p;
1969 /* false => this operand can be reloaded somehow for this
1970 alternative. */
1971 bool badop;
1972 /* true => this operand can be reloaded if the alternative
1973 allows regs. */
1974 bool winreg;
1975 /* True if a constant forced into memory would be OK for
1976 this operand. */
1977 bool constmemok;
1978 enum reg_class this_alternative, this_costly_alternative;
1979 HARD_REG_SET this_alternative_set, this_costly_alternative_set;
1980 bool this_alternative_match_win, this_alternative_win;
1981 bool this_alternative_offmemok;
1982 bool scratch_p;
1983 machine_mode mode;
1984 enum constraint_num cn;
1986 opalt_num = nalt * n_operands + nop;
1987 if (curr_static_id->operand_alternative[opalt_num].anything_ok)
1989 /* Fast track for no constraints at all. */
1990 curr_alt[nop] = NO_REGS;
1991 CLEAR_HARD_REG_SET (curr_alt_set[nop]);
1992 curr_alt_win[nop] = true;
1993 curr_alt_match_win[nop] = false;
1994 curr_alt_offmemok[nop] = false;
1995 curr_alt_matches[nop] = -1;
1996 continue;
1999 op = no_subreg_reg_operand[nop];
2000 mode = curr_operand_mode[nop];
2002 win = did_match = winreg = offmemok = constmemok = false;
2003 badop = true;
2005 early_clobber_p = false;
2006 p = curr_static_id->operand_alternative[opalt_num].constraint;
2008 this_costly_alternative = this_alternative = NO_REGS;
2009 /* We update set of possible hard regs besides its class
2010 because reg class might be inaccurate. For example,
2011 union of LO_REGS (l), HI_REGS(h), and STACK_REG(k) in ARM
2012 is translated in HI_REGS because classes are merged by
2013 pairs and there is no accurate intermediate class. */
2014 CLEAR_HARD_REG_SET (this_alternative_set);
2015 CLEAR_HARD_REG_SET (this_costly_alternative_set);
2016 this_alternative_win = false;
2017 this_alternative_match_win = false;
2018 this_alternative_offmemok = false;
2019 this_alternative_matches = -1;
2021 /* An empty constraint should be excluded by the fast
2022 track. */
2023 lra_assert (*p != 0 && *p != ',');
2025 op_reject = 0;
2026 /* Scan this alternative's specs for this operand; set WIN
2027 if the operand fits any letter in this alternative.
2028 Otherwise, clear BADOP if this operand could fit some
2029 letter after reloads, or set WINREG if this operand could
2030 fit after reloads provided the constraint allows some
2031 registers. */
2032 costly_p = false;
2035 switch ((c = *p, len = CONSTRAINT_LEN (c, p)), c)
2037 case '\0':
2038 len = 0;
2039 break;
2040 case ',':
2041 c = '\0';
2042 break;
2044 case '&':
2045 early_clobber_p = true;
2046 break;
2048 case '$':
2049 op_reject += LRA_MAX_REJECT;
2050 break;
2051 case '^':
2052 op_reject += LRA_LOSER_COST_FACTOR;
2053 break;
2055 case '#':
2056 /* Ignore rest of this alternative. */
2057 c = '\0';
2058 break;
2060 case '0': case '1': case '2': case '3': case '4':
2061 case '5': case '6': case '7': case '8': case '9':
2063 int m_hregno;
2064 bool match_p;
2066 m = strtoul (p, &end, 10);
2067 p = end;
2068 len = 0;
2069 lra_assert (nop > m);
2071 this_alternative_matches = m;
2072 m_hregno = get_hard_regno (*curr_id->operand_loc[m], false);
2073 /* We are supposed to match a previous operand.
2074 If we do, we win if that one did. If we do
2075 not, count both of the operands as losers.
2076 (This is too conservative, since most of the
2077 time only a single reload insn will be needed
2078 to make the two operands win. As a result,
2079 this alternative may be rejected when it is
2080 actually desirable.) */
2081 match_p = false;
2082 if (operands_match_p (*curr_id->operand_loc[nop],
2083 *curr_id->operand_loc[m], m_hregno))
2085 /* We should reject matching of an early
2086 clobber operand if the matching operand is
2087 not dying in the insn. */
2088 if (! curr_static_id->operand[m].early_clobber
2089 || operand_reg[nop] == NULL_RTX
2090 || (find_regno_note (curr_insn, REG_DEAD,
2091 REGNO (op))
2092 || REGNO (op) == REGNO (operand_reg[m])))
2093 match_p = true;
2095 if (match_p)
2097 /* If we are matching a non-offsettable
2098 address where an offsettable address was
2099 expected, then we must reject this
2100 combination, because we can't reload
2101 it. */
2102 if (curr_alt_offmemok[m]
2103 && MEM_P (*curr_id->operand_loc[m])
2104 && curr_alt[m] == NO_REGS && ! curr_alt_win[m])
2105 continue;
2107 else
2109 /* Operands don't match. Both operands must
2110 allow a reload register, otherwise we
2111 cannot make them match. */
2112 if (curr_alt[m] == NO_REGS)
2113 break;
2114 /* Retroactively mark the operand we had to
2115 match as a loser, if it wasn't already and
2116 it wasn't matched to a register constraint
2117 (e.g it might be matched by memory). */
2118 if (curr_alt_win[m]
2119 && (operand_reg[m] == NULL_RTX
2120 || hard_regno[m] < 0))
2122 losers++;
2123 reload_nregs
2124 += (ira_reg_class_max_nregs[curr_alt[m]]
2125 [GET_MODE (*curr_id->operand_loc[m])]);
2128 /* Prefer matching earlyclobber alternative as
2129 it results in less hard regs required for
2130 the insn than a non-matching earlyclobber
2131 alternative. */
2132 if (curr_static_id->operand[m].early_clobber)
2134 if (lra_dump_file != NULL)
2135 fprintf
2136 (lra_dump_file,
2137 " %d Matching earlyclobber alt:"
2138 " reject--\n",
2139 nop);
2140 reject--;
2142 /* Otherwise we prefer no matching
2143 alternatives because it gives more freedom
2144 in RA. */
2145 else if (operand_reg[nop] == NULL_RTX
2146 || (find_regno_note (curr_insn, REG_DEAD,
2147 REGNO (operand_reg[nop]))
2148 == NULL_RTX))
2150 if (lra_dump_file != NULL)
2151 fprintf
2152 (lra_dump_file,
2153 " %d Matching alt: reject+=2\n",
2154 nop);
2155 reject += 2;
2158 /* If we have to reload this operand and some
2159 previous operand also had to match the same
2160 thing as this operand, we don't know how to do
2161 that. */
2162 if (!match_p || !curr_alt_win[m])
2164 for (i = 0; i < nop; i++)
2165 if (curr_alt_matches[i] == m)
2166 break;
2167 if (i < nop)
2168 break;
2170 else
2171 did_match = true;
2173 /* This can be fixed with reloads if the operand
2174 we are supposed to match can be fixed with
2175 reloads. */
2176 badop = false;
2177 this_alternative = curr_alt[m];
2178 COPY_HARD_REG_SET (this_alternative_set, curr_alt_set[m]);
2179 winreg = this_alternative != NO_REGS;
2180 break;
2183 case 'g':
2184 if (MEM_P (op)
2185 || general_constant_p (op)
2186 || spilled_pseudo_p (op))
2187 win = true;
2188 cl = GENERAL_REGS;
2189 goto reg;
2191 default:
2192 cn = lookup_constraint (p);
2193 switch (get_constraint_type (cn))
2195 case CT_REGISTER:
2196 cl = reg_class_for_constraint (cn);
2197 if (cl != NO_REGS)
2198 goto reg;
2199 break;
2201 case CT_CONST_INT:
2202 if (CONST_INT_P (op)
2203 && insn_const_int_ok_for_constraint (INTVAL (op), cn))
2204 win = true;
2205 break;
2207 case CT_MEMORY:
2208 if (MEM_P (op)
2209 && satisfies_memory_constraint_p (op, cn))
2210 win = true;
2211 else if (spilled_pseudo_p (op))
2212 win = true;
2214 /* If we didn't already win, we can reload constants
2215 via force_const_mem or put the pseudo value into
2216 memory, or make other memory by reloading the
2217 address like for 'o'. */
2218 if (CONST_POOL_OK_P (mode, op)
2219 || MEM_P (op) || REG_P (op)
2220 /* We can restore the equiv insn by a
2221 reload. */
2222 || equiv_substition_p[nop])
2223 badop = false;
2224 constmemok = true;
2225 offmemok = true;
2226 break;
2228 case CT_ADDRESS:
2229 /* If we didn't already win, we can reload the address
2230 into a base register. */
2231 if (satisfies_address_constraint_p (op, cn))
2232 win = true;
2233 cl = base_reg_class (VOIDmode, ADDR_SPACE_GENERIC,
2234 ADDRESS, SCRATCH);
2235 badop = false;
2236 goto reg;
2238 case CT_FIXED_FORM:
2239 if (constraint_satisfied_p (op, cn))
2240 win = true;
2241 break;
2243 case CT_SPECIAL_MEMORY:
2244 if (MEM_P (op)
2245 && satisfies_memory_constraint_p (op, cn))
2246 win = true;
2247 else if (spilled_pseudo_p (op))
2248 win = true;
2249 break;
2251 break;
2253 reg:
2254 this_alternative = reg_class_subunion[this_alternative][cl];
2255 IOR_HARD_REG_SET (this_alternative_set,
2256 reg_class_contents[cl]);
2257 if (costly_p)
2259 this_costly_alternative
2260 = reg_class_subunion[this_costly_alternative][cl];
2261 IOR_HARD_REG_SET (this_costly_alternative_set,
2262 reg_class_contents[cl]);
2264 if (mode == BLKmode)
2265 break;
2266 winreg = true;
2267 if (REG_P (op))
2269 if (hard_regno[nop] >= 0
2270 && in_hard_reg_set_p (this_alternative_set,
2271 mode, hard_regno[nop]))
2272 win = true;
2273 else if (hard_regno[nop] < 0
2274 && in_class_p (op, this_alternative, NULL))
2275 win = true;
2277 break;
2279 if (c != ' ' && c != '\t')
2280 costly_p = c == '*';
2282 while ((p += len), c);
2284 scratch_p = (operand_reg[nop] != NULL_RTX
2285 && lra_former_scratch_p (REGNO (operand_reg[nop])));
2286 /* Record which operands fit this alternative. */
2287 if (win)
2289 this_alternative_win = true;
2290 if (operand_reg[nop] != NULL_RTX)
2292 if (hard_regno[nop] >= 0)
2294 if (in_hard_reg_set_p (this_costly_alternative_set,
2295 mode, hard_regno[nop]))
2297 if (lra_dump_file != NULL)
2298 fprintf (lra_dump_file,
2299 " %d Costly set: reject++\n",
2300 nop);
2301 reject++;
2304 else
2306 /* Prefer won reg to spilled pseudo under other
2307 equal conditions for possibe inheritance. */
2308 if (! scratch_p)
2310 if (lra_dump_file != NULL)
2311 fprintf
2312 (lra_dump_file,
2313 " %d Non pseudo reload: reject++\n",
2314 nop);
2315 reject++;
2317 if (in_class_p (operand_reg[nop],
2318 this_costly_alternative, NULL))
2320 if (lra_dump_file != NULL)
2321 fprintf
2322 (lra_dump_file,
2323 " %d Non pseudo costly reload:"
2324 " reject++\n",
2325 nop);
2326 reject++;
2329 /* We simulate the behavior of old reload here.
2330 Although scratches need hard registers and it
2331 might result in spilling other pseudos, no reload
2332 insns are generated for the scratches. So it
2333 might cost something but probably less than old
2334 reload pass believes. */
2335 if (scratch_p)
2337 if (lra_dump_file != NULL)
2338 fprintf (lra_dump_file,
2339 " %d Scratch win: reject+=2\n",
2340 nop);
2341 reject += 2;
2345 else if (did_match)
2346 this_alternative_match_win = true;
2347 else
2349 int const_to_mem = 0;
2350 bool no_regs_p;
2352 reject += op_reject;
2353 /* Never do output reload of stack pointer. It makes
2354 impossible to do elimination when SP is changed in
2355 RTL. */
2356 if (op == stack_pointer_rtx && ! frame_pointer_needed
2357 && curr_static_id->operand[nop].type != OP_IN)
2358 goto fail;
2360 /* If this alternative asks for a specific reg class, see if there
2361 is at least one allocatable register in that class. */
2362 no_regs_p
2363 = (this_alternative == NO_REGS
2364 || (hard_reg_set_subset_p
2365 (reg_class_contents[this_alternative],
2366 lra_no_alloc_regs)));
2368 /* For asms, verify that the class for this alternative is possible
2369 for the mode that is specified. */
2370 if (!no_regs_p && INSN_CODE (curr_insn) < 0)
2372 int i;
2373 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
2374 if (HARD_REGNO_MODE_OK (i, mode)
2375 && in_hard_reg_set_p (reg_class_contents[this_alternative],
2376 mode, i))
2377 break;
2378 if (i == FIRST_PSEUDO_REGISTER)
2379 winreg = false;
2382 /* If this operand accepts a register, and if the
2383 register class has at least one allocatable register,
2384 then this operand can be reloaded. */
2385 if (winreg && !no_regs_p)
2386 badop = false;
2388 if (badop)
2390 if (lra_dump_file != NULL)
2391 fprintf (lra_dump_file,
2392 " alt=%d: Bad operand -- refuse\n",
2393 nalt);
2394 goto fail;
2397 if (this_alternative != NO_REGS)
2399 HARD_REG_SET available_regs;
2401 COPY_HARD_REG_SET (available_regs,
2402 reg_class_contents[this_alternative]);
2403 AND_COMPL_HARD_REG_SET
2404 (available_regs,
2405 ira_prohibited_class_mode_regs[this_alternative][mode]);
2406 AND_COMPL_HARD_REG_SET (available_regs, lra_no_alloc_regs);
2407 if (hard_reg_set_empty_p (available_regs))
2409 /* There are no hard regs holding a value of given
2410 mode. */
2411 if (offmemok)
2413 this_alternative = NO_REGS;
2414 if (lra_dump_file != NULL)
2415 fprintf (lra_dump_file,
2416 " %d Using memory because of"
2417 " a bad mode: reject+=2\n",
2418 nop);
2419 reject += 2;
2421 else
2423 if (lra_dump_file != NULL)
2424 fprintf (lra_dump_file,
2425 " alt=%d: Wrong mode -- refuse\n",
2426 nalt);
2427 goto fail;
2432 /* If not assigned pseudo has a class which a subset of
2433 required reg class, it is a less costly alternative
2434 as the pseudo still can get a hard reg of necessary
2435 class. */
2436 if (! no_regs_p && REG_P (op) && hard_regno[nop] < 0
2437 && (cl = get_reg_class (REGNO (op))) != NO_REGS
2438 && ira_class_subset_p[this_alternative][cl])
2440 if (lra_dump_file != NULL)
2441 fprintf
2442 (lra_dump_file,
2443 " %d Super set class reg: reject-=3\n", nop);
2444 reject -= 3;
2447 this_alternative_offmemok = offmemok;
2448 if (this_costly_alternative != NO_REGS)
2450 if (lra_dump_file != NULL)
2451 fprintf (lra_dump_file,
2452 " %d Costly loser: reject++\n", nop);
2453 reject++;
2455 /* If the operand is dying, has a matching constraint,
2456 and satisfies constraints of the matched operand
2457 which failed to satisfy the own constraints, most probably
2458 the reload for this operand will be gone. */
2459 if (this_alternative_matches >= 0
2460 && !curr_alt_win[this_alternative_matches]
2461 && REG_P (op)
2462 && find_regno_note (curr_insn, REG_DEAD, REGNO (op))
2463 && (hard_regno[nop] >= 0
2464 ? in_hard_reg_set_p (this_alternative_set,
2465 mode, hard_regno[nop])
2466 : in_class_p (op, this_alternative, NULL)))
2468 if (lra_dump_file != NULL)
2469 fprintf
2470 (lra_dump_file,
2471 " %d Dying matched operand reload: reject++\n",
2472 nop);
2473 reject++;
2475 else
2477 /* Strict_low_part requires to reload the register
2478 not the sub-register. In this case we should
2479 check that a final reload hard reg can hold the
2480 value mode. */
2481 if (curr_static_id->operand[nop].strict_low
2482 && REG_P (op)
2483 && hard_regno[nop] < 0
2484 && GET_CODE (*curr_id->operand_loc[nop]) == SUBREG
2485 && ira_class_hard_regs_num[this_alternative] > 0
2486 && ! HARD_REGNO_MODE_OK (ira_class_hard_regs
2487 [this_alternative][0],
2488 GET_MODE
2489 (*curr_id->operand_loc[nop])))
2491 if (lra_dump_file != NULL)
2492 fprintf
2493 (lra_dump_file,
2494 " alt=%d: Strict low subreg reload -- refuse\n",
2495 nalt);
2496 goto fail;
2498 losers++;
2500 if (operand_reg[nop] != NULL_RTX
2501 /* Output operands and matched input operands are
2502 not inherited. The following conditions do not
2503 exactly describe the previous statement but they
2504 are pretty close. */
2505 && curr_static_id->operand[nop].type != OP_OUT
2506 && (this_alternative_matches < 0
2507 || curr_static_id->operand[nop].type != OP_IN))
2509 int last_reload = (lra_reg_info[ORIGINAL_REGNO
2510 (operand_reg[nop])]
2511 .last_reload);
2513 /* The value of reload_sum has sense only if we
2514 process insns in their order. It happens only on
2515 the first constraints sub-pass when we do most of
2516 reload work. */
2517 if (lra_constraint_iter == 1 && last_reload > bb_reload_num)
2518 reload_sum += last_reload - bb_reload_num;
2520 /* If this is a constant that is reloaded into the
2521 desired class by copying it to memory first, count
2522 that as another reload. This is consistent with
2523 other code and is required to avoid choosing another
2524 alternative when the constant is moved into memory.
2525 Note that the test here is precisely the same as in
2526 the code below that calls force_const_mem. */
2527 if (CONST_POOL_OK_P (mode, op)
2528 && ((targetm.preferred_reload_class
2529 (op, this_alternative) == NO_REGS)
2530 || no_input_reloads_p))
2532 const_to_mem = 1;
2533 if (! no_regs_p)
2534 losers++;
2537 /* Alternative loses if it requires a type of reload not
2538 permitted for this insn. We can always reload
2539 objects with a REG_UNUSED note. */
2540 if ((curr_static_id->operand[nop].type != OP_IN
2541 && no_output_reloads_p
2542 && ! find_reg_note (curr_insn, REG_UNUSED, op))
2543 || (curr_static_id->operand[nop].type != OP_OUT
2544 && no_input_reloads_p && ! const_to_mem)
2545 || (this_alternative_matches >= 0
2546 && (no_input_reloads_p
2547 || (no_output_reloads_p
2548 && (curr_static_id->operand
2549 [this_alternative_matches].type != OP_IN)
2550 && ! find_reg_note (curr_insn, REG_UNUSED,
2551 no_subreg_reg_operand
2552 [this_alternative_matches])))))
2554 if (lra_dump_file != NULL)
2555 fprintf
2556 (lra_dump_file,
2557 " alt=%d: No input/otput reload -- refuse\n",
2558 nalt);
2559 goto fail;
2562 /* Alternative loses if it required class pseudo can not
2563 hold value of required mode. Such insns can be
2564 described by insn definitions with mode iterators. */
2565 if (GET_MODE (*curr_id->operand_loc[nop]) != VOIDmode
2566 && ! hard_reg_set_empty_p (this_alternative_set)
2567 /* It is common practice for constraints to use a
2568 class which does not have actually enough regs to
2569 hold the value (e.g. x86 AREG for mode requiring
2570 more one general reg). Therefore we have 2
2571 conditions to check that the reload pseudo can
2572 not hold the mode value. */
2573 && ! HARD_REGNO_MODE_OK (ira_class_hard_regs
2574 [this_alternative][0],
2575 GET_MODE (*curr_id->operand_loc[nop]))
2576 /* The above condition is not enough as the first
2577 reg in ira_class_hard_regs can be not aligned for
2578 multi-words mode values. */
2579 && (prohibited_class_reg_set_mode_p
2580 (this_alternative, this_alternative_set,
2581 GET_MODE (*curr_id->operand_loc[nop]))))
2583 if (lra_dump_file != NULL)
2584 fprintf (lra_dump_file,
2585 " alt=%d: reload pseudo for op %d "
2586 " can not hold the mode value -- refuse\n",
2587 nalt, nop);
2588 goto fail;
2591 /* Check strong discouragement of reload of non-constant
2592 into class THIS_ALTERNATIVE. */
2593 if (! CONSTANT_P (op) && ! no_regs_p
2594 && (targetm.preferred_reload_class
2595 (op, this_alternative) == NO_REGS
2596 || (curr_static_id->operand[nop].type == OP_OUT
2597 && (targetm.preferred_output_reload_class
2598 (op, this_alternative) == NO_REGS))))
2600 if (lra_dump_file != NULL)
2601 fprintf (lra_dump_file,
2602 " %d Non-prefered reload: reject+=%d\n",
2603 nop, LRA_MAX_REJECT);
2604 reject += LRA_MAX_REJECT;
2607 if (! (MEM_P (op) && offmemok)
2608 && ! (const_to_mem && constmemok))
2610 /* We prefer to reload pseudos over reloading other
2611 things, since such reloads may be able to be
2612 eliminated later. So bump REJECT in other cases.
2613 Don't do this in the case where we are forcing a
2614 constant into memory and it will then win since
2615 we don't want to have a different alternative
2616 match then. */
2617 if (! (REG_P (op) && REGNO (op) >= FIRST_PSEUDO_REGISTER))
2619 if (lra_dump_file != NULL)
2620 fprintf
2621 (lra_dump_file,
2622 " %d Non-pseudo reload: reject+=2\n",
2623 nop);
2624 reject += 2;
2627 if (! no_regs_p)
2628 reload_nregs
2629 += ira_reg_class_max_nregs[this_alternative][mode];
2631 if (SMALL_REGISTER_CLASS_P (this_alternative))
2633 if (lra_dump_file != NULL)
2634 fprintf
2635 (lra_dump_file,
2636 " %d Small class reload: reject+=%d\n",
2637 nop, LRA_LOSER_COST_FACTOR / 2);
2638 reject += LRA_LOSER_COST_FACTOR / 2;
2642 /* We are trying to spill pseudo into memory. It is
2643 usually more costly than moving to a hard register
2644 although it might takes the same number of
2645 reloads.
2647 Non-pseudo spill may happen also. Suppose a target allows both
2648 register and memory in the operand constraint alternatives,
2649 then it's typical that an eliminable register has a substition
2650 of "base + offset" which can either be reloaded by a simple
2651 "new_reg <= base + offset" which will match the register
2652 constraint, or a similar reg addition followed by further spill
2653 to and reload from memory which will match the memory
2654 constraint, but this memory spill will be much more costly
2655 usually.
2657 Code below increases the reject for both pseudo and non-pseudo
2658 spill. */
2659 if (no_regs_p
2660 && !(MEM_P (op) && offmemok)
2661 && !(REG_P (op) && hard_regno[nop] < 0))
2663 if (lra_dump_file != NULL)
2664 fprintf
2665 (lra_dump_file,
2666 " %d Spill %spseudo into memory: reject+=3\n",
2667 nop, REG_P (op) ? "" : "Non-");
2668 reject += 3;
2669 if (VECTOR_MODE_P (mode))
2671 /* Spilling vectors into memory is usually more
2672 costly as they contain big values. */
2673 if (lra_dump_file != NULL)
2674 fprintf
2675 (lra_dump_file,
2676 " %d Spill vector pseudo: reject+=2\n",
2677 nop);
2678 reject += 2;
2682 #ifdef SECONDARY_MEMORY_NEEDED
2683 /* If reload requires moving value through secondary
2684 memory, it will need one more insn at least. */
2685 if (this_alternative != NO_REGS
2686 && REG_P (op) && (cl = get_reg_class (REGNO (op))) != NO_REGS
2687 && ((curr_static_id->operand[nop].type != OP_OUT
2688 && SECONDARY_MEMORY_NEEDED (cl, this_alternative,
2689 GET_MODE (op)))
2690 || (curr_static_id->operand[nop].type != OP_IN
2691 && SECONDARY_MEMORY_NEEDED (this_alternative, cl,
2692 GET_MODE (op)))))
2693 losers++;
2694 #endif
2695 /* Input reloads can be inherited more often than output
2696 reloads can be removed, so penalize output
2697 reloads. */
2698 if (!REG_P (op) || curr_static_id->operand[nop].type != OP_IN)
2700 if (lra_dump_file != NULL)
2701 fprintf
2702 (lra_dump_file,
2703 " %d Non input pseudo reload: reject++\n",
2704 nop);
2705 reject++;
2709 if (early_clobber_p && ! scratch_p)
2711 if (lra_dump_file != NULL)
2712 fprintf (lra_dump_file,
2713 " %d Early clobber: reject++\n", nop);
2714 reject++;
2716 /* ??? We check early clobbers after processing all operands
2717 (see loop below) and there we update the costs more.
2718 Should we update the cost (may be approximately) here
2719 because of early clobber register reloads or it is a rare
2720 or non-important thing to be worth to do it. */
2721 overall = losers * LRA_LOSER_COST_FACTOR + reject;
2722 if ((best_losers == 0 || losers != 0) && best_overall < overall)
2724 if (lra_dump_file != NULL)
2725 fprintf (lra_dump_file,
2726 " alt=%d,overall=%d,losers=%d -- refuse\n",
2727 nalt, overall, losers);
2728 goto fail;
2731 curr_alt[nop] = this_alternative;
2732 COPY_HARD_REG_SET (curr_alt_set[nop], this_alternative_set);
2733 curr_alt_win[nop] = this_alternative_win;
2734 curr_alt_match_win[nop] = this_alternative_match_win;
2735 curr_alt_offmemok[nop] = this_alternative_offmemok;
2736 curr_alt_matches[nop] = this_alternative_matches;
2738 if (this_alternative_matches >= 0
2739 && !did_match && !this_alternative_win)
2740 curr_alt_win[this_alternative_matches] = false;
2742 if (early_clobber_p && operand_reg[nop] != NULL_RTX)
2743 early_clobbered_nops[early_clobbered_regs_num++] = nop;
2745 if (curr_insn_set != NULL_RTX && n_operands == 2
2746 /* Prevent processing non-move insns. */
2747 && (GET_CODE (SET_SRC (curr_insn_set)) == SUBREG
2748 || SET_SRC (curr_insn_set) == no_subreg_reg_operand[1])
2749 && ((! curr_alt_win[0] && ! curr_alt_win[1]
2750 && REG_P (no_subreg_reg_operand[0])
2751 && REG_P (no_subreg_reg_operand[1])
2752 && (reg_in_class_p (no_subreg_reg_operand[0], curr_alt[1])
2753 || reg_in_class_p (no_subreg_reg_operand[1], curr_alt[0])))
2754 || (! curr_alt_win[0] && curr_alt_win[1]
2755 && REG_P (no_subreg_reg_operand[1])
2756 && reg_in_class_p (no_subreg_reg_operand[1], curr_alt[0]))
2757 || (curr_alt_win[0] && ! curr_alt_win[1]
2758 && REG_P (no_subreg_reg_operand[0])
2759 && reg_in_class_p (no_subreg_reg_operand[0], curr_alt[1])
2760 && (! CONST_POOL_OK_P (curr_operand_mode[1],
2761 no_subreg_reg_operand[1])
2762 || (targetm.preferred_reload_class
2763 (no_subreg_reg_operand[1],
2764 (enum reg_class) curr_alt[1]) != NO_REGS))
2765 /* If it is a result of recent elimination in move
2766 insn we can transform it into an add still by
2767 using this alternative. */
2768 && GET_CODE (no_subreg_reg_operand[1]) != PLUS)))
2770 /* We have a move insn and a new reload insn will be similar
2771 to the current insn. We should avoid such situation as it
2772 results in LRA cycling. */
2773 overall += LRA_MAX_REJECT;
2775 ok_p = true;
2776 curr_alt_dont_inherit_ops_num = 0;
2777 for (nop = 0; nop < early_clobbered_regs_num; nop++)
2779 int i, j, clobbered_hard_regno, first_conflict_j, last_conflict_j;
2780 HARD_REG_SET temp_set;
2782 i = early_clobbered_nops[nop];
2783 if ((! curr_alt_win[i] && ! curr_alt_match_win[i])
2784 || hard_regno[i] < 0)
2785 continue;
2786 lra_assert (operand_reg[i] != NULL_RTX);
2787 clobbered_hard_regno = hard_regno[i];
2788 CLEAR_HARD_REG_SET (temp_set);
2789 add_to_hard_reg_set (&temp_set, biggest_mode[i], clobbered_hard_regno);
2790 first_conflict_j = last_conflict_j = -1;
2791 for (j = 0; j < n_operands; j++)
2792 if (j == i
2793 /* We don't want process insides of match_operator and
2794 match_parallel because otherwise we would process
2795 their operands once again generating a wrong
2796 code. */
2797 || curr_static_id->operand[j].is_operator)
2798 continue;
2799 else if ((curr_alt_matches[j] == i && curr_alt_match_win[j])
2800 || (curr_alt_matches[i] == j && curr_alt_match_win[i]))
2801 continue;
2802 /* If we don't reload j-th operand, check conflicts. */
2803 else if ((curr_alt_win[j] || curr_alt_match_win[j])
2804 && uses_hard_regs_p (*curr_id->operand_loc[j], temp_set))
2806 if (first_conflict_j < 0)
2807 first_conflict_j = j;
2808 last_conflict_j = j;
2810 if (last_conflict_j < 0)
2811 continue;
2812 /* If earlyclobber operand conflicts with another
2813 non-matching operand which is actually the same register
2814 as the earlyclobber operand, it is better to reload the
2815 another operand as an operand matching the earlyclobber
2816 operand can be also the same. */
2817 if (first_conflict_j == last_conflict_j
2818 && operand_reg[last_conflict_j] != NULL_RTX
2819 && ! curr_alt_match_win[last_conflict_j]
2820 && REGNO (operand_reg[i]) == REGNO (operand_reg[last_conflict_j]))
2822 curr_alt_win[last_conflict_j] = false;
2823 curr_alt_dont_inherit_ops[curr_alt_dont_inherit_ops_num++]
2824 = last_conflict_j;
2825 losers++;
2826 /* Early clobber was already reflected in REJECT. */
2827 lra_assert (reject > 0);
2828 if (lra_dump_file != NULL)
2829 fprintf
2830 (lra_dump_file,
2831 " %d Conflict early clobber reload: reject--\n",
2833 reject--;
2834 overall += LRA_LOSER_COST_FACTOR - 1;
2836 else
2838 /* We need to reload early clobbered register and the
2839 matched registers. */
2840 for (j = 0; j < n_operands; j++)
2841 if (curr_alt_matches[j] == i)
2843 curr_alt_match_win[j] = false;
2844 losers++;
2845 overall += LRA_LOSER_COST_FACTOR;
2847 if (! curr_alt_match_win[i])
2848 curr_alt_dont_inherit_ops[curr_alt_dont_inherit_ops_num++] = i;
2849 else
2851 /* Remember pseudos used for match reloads are never
2852 inherited. */
2853 lra_assert (curr_alt_matches[i] >= 0);
2854 curr_alt_win[curr_alt_matches[i]] = false;
2856 curr_alt_win[i] = curr_alt_match_win[i] = false;
2857 losers++;
2858 /* Early clobber was already reflected in REJECT. */
2859 lra_assert (reject > 0);
2860 if (lra_dump_file != NULL)
2861 fprintf
2862 (lra_dump_file,
2863 " %d Matched conflict early clobber reloads: "
2864 "reject--\n",
2866 reject--;
2867 overall += LRA_LOSER_COST_FACTOR - 1;
2870 if (lra_dump_file != NULL)
2871 fprintf (lra_dump_file, " alt=%d,overall=%d,losers=%d,rld_nregs=%d\n",
2872 nalt, overall, losers, reload_nregs);
2874 /* If this alternative can be made to work by reloading, and it
2875 needs less reloading than the others checked so far, record
2876 it as the chosen goal for reloading. */
2877 if ((best_losers != 0 && losers == 0)
2878 || (((best_losers == 0 && losers == 0)
2879 || (best_losers != 0 && losers != 0))
2880 && (best_overall > overall
2881 || (best_overall == overall
2882 /* If the cost of the reloads is the same,
2883 prefer alternative which requires minimal
2884 number of reload regs. */
2885 && (reload_nregs < best_reload_nregs
2886 || (reload_nregs == best_reload_nregs
2887 && (best_reload_sum < reload_sum
2888 || (best_reload_sum == reload_sum
2889 && nalt < goal_alt_number))))))))
2891 for (nop = 0; nop < n_operands; nop++)
2893 goal_alt_win[nop] = curr_alt_win[nop];
2894 goal_alt_match_win[nop] = curr_alt_match_win[nop];
2895 goal_alt_matches[nop] = curr_alt_matches[nop];
2896 goal_alt[nop] = curr_alt[nop];
2897 goal_alt_offmemok[nop] = curr_alt_offmemok[nop];
2899 goal_alt_dont_inherit_ops_num = curr_alt_dont_inherit_ops_num;
2900 for (nop = 0; nop < curr_alt_dont_inherit_ops_num; nop++)
2901 goal_alt_dont_inherit_ops[nop] = curr_alt_dont_inherit_ops[nop];
2902 goal_alt_swapped = curr_swapped;
2903 best_overall = overall;
2904 best_losers = losers;
2905 best_reload_nregs = reload_nregs;
2906 best_reload_sum = reload_sum;
2907 goal_alt_number = nalt;
2909 if (losers == 0)
2910 /* Everything is satisfied. Do not process alternatives
2911 anymore. */
2912 break;
2913 fail:
2916 return ok_p;
2919 /* Make reload base reg from address AD. */
2920 static rtx
2921 base_to_reg (struct address_info *ad)
2923 enum reg_class cl;
2924 int code = -1;
2925 rtx new_inner = NULL_RTX;
2926 rtx new_reg = NULL_RTX;
2927 rtx_insn *insn;
2928 rtx_insn *last_insn = get_last_insn();
2930 lra_assert (ad->base == ad->base_term && ad->disp == ad->disp_term);
2931 cl = base_reg_class (ad->mode, ad->as, ad->base_outer_code,
2932 get_index_code (ad));
2933 new_reg = lra_create_new_reg (GET_MODE (*ad->base_term), NULL_RTX,
2934 cl, "base");
2935 new_inner = simplify_gen_binary (PLUS, GET_MODE (new_reg), new_reg,
2936 ad->disp_term == NULL
2937 ? gen_int_mode (0, ad->mode)
2938 : *ad->disp_term);
2939 if (!valid_address_p (ad->mode, new_inner, ad->as))
2940 return NULL_RTX;
2941 insn = emit_insn (gen_rtx_SET (new_reg, *ad->base_term));
2942 code = recog_memoized (insn);
2943 if (code < 0)
2945 delete_insns_since (last_insn);
2946 return NULL_RTX;
2949 return new_inner;
2952 /* Make reload base reg + disp from address AD. Return the new pseudo. */
2953 static rtx
2954 base_plus_disp_to_reg (struct address_info *ad)
2956 enum reg_class cl;
2957 rtx new_reg;
2959 lra_assert (ad->base == ad->base_term && ad->disp == ad->disp_term);
2960 cl = base_reg_class (ad->mode, ad->as, ad->base_outer_code,
2961 get_index_code (ad));
2962 new_reg = lra_create_new_reg (GET_MODE (*ad->base_term), NULL_RTX,
2963 cl, "base + disp");
2964 lra_emit_add (new_reg, *ad->base_term, *ad->disp_term);
2965 return new_reg;
2968 /* Make reload of index part of address AD. Return the new
2969 pseudo. */
2970 static rtx
2971 index_part_to_reg (struct address_info *ad)
2973 rtx new_reg;
2975 new_reg = lra_create_new_reg (GET_MODE (*ad->index), NULL_RTX,
2976 INDEX_REG_CLASS, "index term");
2977 expand_mult (GET_MODE (*ad->index), *ad->index_term,
2978 GEN_INT (get_index_scale (ad)), new_reg, 1);
2979 return new_reg;
2982 /* Return true if we can add a displacement to address AD, even if that
2983 makes the address invalid. The fix-up code requires any new address
2984 to be the sum of the BASE_TERM, INDEX and DISP_TERM fields. */
2985 static bool
2986 can_add_disp_p (struct address_info *ad)
2988 return (!ad->autoinc_p
2989 && ad->segment == NULL
2990 && ad->base == ad->base_term
2991 && ad->disp == ad->disp_term);
2994 /* Make equiv substitution in address AD. Return true if a substitution
2995 was made. */
2996 static bool
2997 equiv_address_substitution (struct address_info *ad)
2999 rtx base_reg, new_base_reg, index_reg, new_index_reg, *base_term, *index_term;
3000 HOST_WIDE_INT disp, scale;
3001 bool change_p;
3003 base_term = strip_subreg (ad->base_term);
3004 if (base_term == NULL)
3005 base_reg = new_base_reg = NULL_RTX;
3006 else
3008 base_reg = *base_term;
3009 new_base_reg = get_equiv_with_elimination (base_reg, curr_insn);
3011 index_term = strip_subreg (ad->index_term);
3012 if (index_term == NULL)
3013 index_reg = new_index_reg = NULL_RTX;
3014 else
3016 index_reg = *index_term;
3017 new_index_reg = get_equiv_with_elimination (index_reg, curr_insn);
3019 if (base_reg == new_base_reg && index_reg == new_index_reg)
3020 return false;
3021 disp = 0;
3022 change_p = false;
3023 if (lra_dump_file != NULL)
3025 fprintf (lra_dump_file, "Changing address in insn %d ",
3026 INSN_UID (curr_insn));
3027 dump_value_slim (lra_dump_file, *ad->outer, 1);
3029 if (base_reg != new_base_reg)
3031 if (REG_P (new_base_reg))
3033 *base_term = new_base_reg;
3034 change_p = true;
3036 else if (GET_CODE (new_base_reg) == PLUS
3037 && REG_P (XEXP (new_base_reg, 0))
3038 && CONST_INT_P (XEXP (new_base_reg, 1))
3039 && can_add_disp_p (ad))
3041 disp += INTVAL (XEXP (new_base_reg, 1));
3042 *base_term = XEXP (new_base_reg, 0);
3043 change_p = true;
3045 if (ad->base_term2 != NULL)
3046 *ad->base_term2 = *ad->base_term;
3048 if (index_reg != new_index_reg)
3050 if (REG_P (new_index_reg))
3052 *index_term = new_index_reg;
3053 change_p = true;
3055 else if (GET_CODE (new_index_reg) == PLUS
3056 && REG_P (XEXP (new_index_reg, 0))
3057 && CONST_INT_P (XEXP (new_index_reg, 1))
3058 && can_add_disp_p (ad)
3059 && (scale = get_index_scale (ad)))
3061 disp += INTVAL (XEXP (new_index_reg, 1)) * scale;
3062 *index_term = XEXP (new_index_reg, 0);
3063 change_p = true;
3066 if (disp != 0)
3068 if (ad->disp != NULL)
3069 *ad->disp = plus_constant (GET_MODE (*ad->inner), *ad->disp, disp);
3070 else
3072 *ad->inner = plus_constant (GET_MODE (*ad->inner), *ad->inner, disp);
3073 update_address (ad);
3075 change_p = true;
3077 if (lra_dump_file != NULL)
3079 if (! change_p)
3080 fprintf (lra_dump_file, " -- no change\n");
3081 else
3083 fprintf (lra_dump_file, " on equiv ");
3084 dump_value_slim (lra_dump_file, *ad->outer, 1);
3085 fprintf (lra_dump_file, "\n");
3088 return change_p;
3091 /* Major function to make reloads for an address in operand NOP or
3092 check its correctness (If CHECK_ONLY_P is true). The supported
3093 cases are:
3095 1) an address that existed before LRA started, at which point it
3096 must have been valid. These addresses are subject to elimination
3097 and may have become invalid due to the elimination offset being out
3098 of range.
3100 2) an address created by forcing a constant to memory
3101 (force_const_to_mem). The initial form of these addresses might
3102 not be valid, and it is this function's job to make them valid.
3104 3) a frame address formed from a register and a (possibly zero)
3105 constant offset. As above, these addresses might not be valid and
3106 this function must make them so.
3108 Add reloads to the lists *BEFORE and *AFTER. We might need to add
3109 reloads to *AFTER because of inc/dec, {pre, post} modify in the
3110 address. Return true for any RTL change.
3112 The function is a helper function which does not produce all
3113 transformations (when CHECK_ONLY_P is false) which can be
3114 necessary. It does just basic steps. To do all necessary
3115 transformations use function process_address. */
3116 static bool
3117 process_address_1 (int nop, bool check_only_p,
3118 rtx_insn **before, rtx_insn **after)
3120 struct address_info ad;
3121 rtx new_reg;
3122 HOST_WIDE_INT scale;
3123 rtx op = *curr_id->operand_loc[nop];
3124 const char *constraint = curr_static_id->operand[nop].constraint;
3125 enum constraint_num cn = lookup_constraint (constraint);
3126 bool change_p = false;
3128 if (MEM_P (op)
3129 && GET_MODE (op) == BLKmode
3130 && GET_CODE (XEXP (op, 0)) == SCRATCH)
3131 return false;
3133 if (insn_extra_address_constraint (cn))
3134 decompose_lea_address (&ad, curr_id->operand_loc[nop]);
3135 /* Do not attempt to decompose arbitrary addresses generated by combine
3136 for asm operands with loose constraints, e.g 'X'. */
3137 else if (MEM_P (op)
3138 && !(get_constraint_type (cn) == CT_FIXED_FORM
3139 && constraint_satisfied_p (op, cn)))
3140 decompose_mem_address (&ad, op);
3141 else if (GET_CODE (op) == SUBREG
3142 && MEM_P (SUBREG_REG (op)))
3143 decompose_mem_address (&ad, SUBREG_REG (op));
3144 else
3145 return false;
3146 /* If INDEX_REG_CLASS is assigned to base_term already and isn't to
3147 index_term, swap them so to avoid assigning INDEX_REG_CLASS to both
3148 when INDEX_REG_CLASS is a single register class. */
3149 if (ad.base_term != NULL
3150 && ad.index_term != NULL
3151 && ira_class_hard_regs_num[INDEX_REG_CLASS] == 1
3152 && REG_P (*ad.base_term)
3153 && REG_P (*ad.index_term)
3154 && in_class_p (*ad.base_term, INDEX_REG_CLASS, NULL)
3155 && ! in_class_p (*ad.index_term, INDEX_REG_CLASS, NULL))
3157 std::swap (ad.base, ad.index);
3158 std::swap (ad.base_term, ad.index_term);
3160 if (! check_only_p)
3161 change_p = equiv_address_substitution (&ad);
3162 if (ad.base_term != NULL
3163 && (process_addr_reg
3164 (ad.base_term, check_only_p, before,
3165 (ad.autoinc_p
3166 && !(REG_P (*ad.base_term)
3167 && find_regno_note (curr_insn, REG_DEAD,
3168 REGNO (*ad.base_term)) != NULL_RTX)
3169 ? after : NULL),
3170 base_reg_class (ad.mode, ad.as, ad.base_outer_code,
3171 get_index_code (&ad)))))
3173 change_p = true;
3174 if (ad.base_term2 != NULL)
3175 *ad.base_term2 = *ad.base_term;
3177 if (ad.index_term != NULL
3178 && process_addr_reg (ad.index_term, check_only_p,
3179 before, NULL, INDEX_REG_CLASS))
3180 change_p = true;
3182 /* Target hooks sometimes don't treat extra-constraint addresses as
3183 legitimate address_operands, so handle them specially. */
3184 if (insn_extra_address_constraint (cn)
3185 && satisfies_address_constraint_p (&ad, cn))
3186 return change_p;
3188 if (check_only_p)
3189 return change_p;
3191 /* There are three cases where the shape of *AD.INNER may now be invalid:
3193 1) the original address was valid, but either elimination or
3194 equiv_address_substitution was applied and that made
3195 the address invalid.
3197 2) the address is an invalid symbolic address created by
3198 force_const_to_mem.
3200 3) the address is a frame address with an invalid offset.
3202 4) the address is a frame address with an invalid base.
3204 All these cases involve a non-autoinc address, so there is no
3205 point revalidating other types. */
3206 if (ad.autoinc_p || valid_address_p (&ad))
3207 return change_p;
3209 /* Any index existed before LRA started, so we can assume that the
3210 presence and shape of the index is valid. */
3211 push_to_sequence (*before);
3212 lra_assert (ad.disp == ad.disp_term);
3213 if (ad.base == NULL)
3215 if (ad.index == NULL)
3217 rtx_insn *insn;
3218 rtx_insn *last = get_last_insn ();
3219 int code = -1;
3220 enum reg_class cl = base_reg_class (ad.mode, ad.as,
3221 SCRATCH, SCRATCH);
3222 rtx addr = *ad.inner;
3224 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "addr");
3225 if (HAVE_lo_sum)
3227 /* addr => lo_sum (new_base, addr), case (2) above. */
3228 insn = emit_insn (gen_rtx_SET
3229 (new_reg,
3230 gen_rtx_HIGH (Pmode, copy_rtx (addr))));
3231 code = recog_memoized (insn);
3232 if (code >= 0)
3234 *ad.inner = gen_rtx_LO_SUM (Pmode, new_reg, addr);
3235 if (! valid_address_p (ad.mode, *ad.outer, ad.as))
3237 /* Try to put lo_sum into register. */
3238 insn = emit_insn (gen_rtx_SET
3239 (new_reg,
3240 gen_rtx_LO_SUM (Pmode, new_reg, addr)));
3241 code = recog_memoized (insn);
3242 if (code >= 0)
3244 *ad.inner = new_reg;
3245 if (! valid_address_p (ad.mode, *ad.outer, ad.as))
3247 *ad.inner = addr;
3248 code = -1;
3254 if (code < 0)
3255 delete_insns_since (last);
3258 if (code < 0)
3260 /* addr => new_base, case (2) above. */
3261 lra_emit_move (new_reg, addr);
3263 for (insn = last == NULL_RTX ? get_insns () : NEXT_INSN (last);
3264 insn != NULL_RTX;
3265 insn = NEXT_INSN (insn))
3266 if (recog_memoized (insn) < 0)
3267 break;
3268 if (insn != NULL_RTX)
3270 /* Do nothing if we cannot generate right insns.
3271 This is analogous to reload pass behavior. */
3272 delete_insns_since (last);
3273 end_sequence ();
3274 return false;
3276 *ad.inner = new_reg;
3279 else
3281 /* index * scale + disp => new base + index * scale,
3282 case (1) above. */
3283 enum reg_class cl = base_reg_class (ad.mode, ad.as, PLUS,
3284 GET_CODE (*ad.index));
3286 lra_assert (INDEX_REG_CLASS != NO_REGS);
3287 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "disp");
3288 lra_emit_move (new_reg, *ad.disp);
3289 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3290 new_reg, *ad.index);
3293 else if (ad.index == NULL)
3295 int regno;
3296 enum reg_class cl;
3297 rtx set;
3298 rtx_insn *insns, *last_insn;
3299 /* Try to reload base into register only if the base is invalid
3300 for the address but with valid offset, case (4) above. */
3301 start_sequence ();
3302 new_reg = base_to_reg (&ad);
3304 /* base + disp => new base, cases (1) and (3) above. */
3305 /* Another option would be to reload the displacement into an
3306 index register. However, postreload has code to optimize
3307 address reloads that have the same base and different
3308 displacements, so reloading into an index register would
3309 not necessarily be a win. */
3310 if (new_reg == NULL_RTX)
3311 new_reg = base_plus_disp_to_reg (&ad);
3312 insns = get_insns ();
3313 last_insn = get_last_insn ();
3314 /* If we generated at least two insns, try last insn source as
3315 an address. If we succeed, we generate one less insn. */
3316 if (last_insn != insns && (set = single_set (last_insn)) != NULL_RTX
3317 && GET_CODE (SET_SRC (set)) == PLUS
3318 && REG_P (XEXP (SET_SRC (set), 0))
3319 && CONSTANT_P (XEXP (SET_SRC (set), 1)))
3321 *ad.inner = SET_SRC (set);
3322 if (valid_address_p (ad.mode, *ad.outer, ad.as))
3324 *ad.base_term = XEXP (SET_SRC (set), 0);
3325 *ad.disp_term = XEXP (SET_SRC (set), 1);
3326 cl = base_reg_class (ad.mode, ad.as, ad.base_outer_code,
3327 get_index_code (&ad));
3328 regno = REGNO (*ad.base_term);
3329 if (regno >= FIRST_PSEUDO_REGISTER
3330 && cl != lra_get_allocno_class (regno))
3331 lra_change_class (regno, cl, " Change to", true);
3332 new_reg = SET_SRC (set);
3333 delete_insns_since (PREV_INSN (last_insn));
3336 /* Try if target can split displacement into legitimite new disp
3337 and offset. If it's the case, we replace the last insn with
3338 insns for base + offset => new_reg and set new_reg + new disp
3339 to *ad.inner. */
3340 last_insn = get_last_insn ();
3341 if ((set = single_set (last_insn)) != NULL_RTX
3342 && GET_CODE (SET_SRC (set)) == PLUS
3343 && REG_P (XEXP (SET_SRC (set), 0))
3344 && REGNO (XEXP (SET_SRC (set), 0)) < FIRST_PSEUDO_REGISTER
3345 && CONST_INT_P (XEXP (SET_SRC (set), 1)))
3347 rtx addend, disp = XEXP (SET_SRC (set), 1);
3348 if (targetm.legitimize_address_displacement (&disp, &addend,
3349 ad.mode))
3351 rtx_insn *new_insns;
3352 start_sequence ();
3353 lra_emit_add (new_reg, XEXP (SET_SRC (set), 0), addend);
3354 new_insns = get_insns ();
3355 end_sequence ();
3356 new_reg = gen_rtx_PLUS (Pmode, new_reg, disp);
3357 delete_insns_since (PREV_INSN (last_insn));
3358 add_insn (new_insns);
3359 insns = get_insns ();
3362 end_sequence ();
3363 emit_insn (insns);
3364 *ad.inner = new_reg;
3366 else if (ad.disp_term != NULL)
3368 /* base + scale * index + disp => new base + scale * index,
3369 case (1) above. */
3370 new_reg = base_plus_disp_to_reg (&ad);
3371 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3372 new_reg, *ad.index);
3374 else if ((scale = get_index_scale (&ad)) == 1)
3376 /* The last transformation to one reg will be made in
3377 curr_insn_transform function. */
3378 end_sequence ();
3379 return false;
3381 else if (scale != 0)
3383 /* base + scale * index => base + new_reg,
3384 case (1) above.
3385 Index part of address may become invalid. For example, we
3386 changed pseudo on the equivalent memory and a subreg of the
3387 pseudo onto the memory of different mode for which the scale is
3388 prohibitted. */
3389 new_reg = index_part_to_reg (&ad);
3390 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3391 *ad.base_term, new_reg);
3393 else
3395 enum reg_class cl = base_reg_class (ad.mode, ad.as,
3396 SCRATCH, SCRATCH);
3397 rtx addr = *ad.inner;
3399 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "addr");
3400 /* addr => new_base. */
3401 lra_emit_move (new_reg, addr);
3402 *ad.inner = new_reg;
3404 *before = get_insns ();
3405 end_sequence ();
3406 return true;
3409 /* If CHECK_ONLY_P is false, do address reloads until it is necessary.
3410 Use process_address_1 as a helper function. Return true for any
3411 RTL changes.
3413 If CHECK_ONLY_P is true, just check address correctness. Return
3414 false if the address correct. */
3415 static bool
3416 process_address (int nop, bool check_only_p,
3417 rtx_insn **before, rtx_insn **after)
3419 bool res = false;
3421 while (process_address_1 (nop, check_only_p, before, after))
3423 if (check_only_p)
3424 return true;
3425 res = true;
3427 return res;
3430 /* Emit insns to reload VALUE into a new register. VALUE is an
3431 auto-increment or auto-decrement RTX whose operand is a register or
3432 memory location; so reloading involves incrementing that location.
3433 IN is either identical to VALUE, or some cheaper place to reload
3434 value being incremented/decremented from.
3436 INC_AMOUNT is the number to increment or decrement by (always
3437 positive and ignored for POST_MODIFY/PRE_MODIFY).
3439 Return pseudo containing the result. */
3440 static rtx
3441 emit_inc (enum reg_class new_rclass, rtx in, rtx value, int inc_amount)
3443 /* REG or MEM to be copied and incremented. */
3444 rtx incloc = XEXP (value, 0);
3445 /* Nonzero if increment after copying. */
3446 int post = (GET_CODE (value) == POST_DEC || GET_CODE (value) == POST_INC
3447 || GET_CODE (value) == POST_MODIFY);
3448 rtx_insn *last;
3449 rtx inc;
3450 rtx_insn *add_insn;
3451 int code;
3452 rtx real_in = in == value ? incloc : in;
3453 rtx result;
3454 bool plus_p = true;
3456 if (GET_CODE (value) == PRE_MODIFY || GET_CODE (value) == POST_MODIFY)
3458 lra_assert (GET_CODE (XEXP (value, 1)) == PLUS
3459 || GET_CODE (XEXP (value, 1)) == MINUS);
3460 lra_assert (rtx_equal_p (XEXP (XEXP (value, 1), 0), XEXP (value, 0)));
3461 plus_p = GET_CODE (XEXP (value, 1)) == PLUS;
3462 inc = XEXP (XEXP (value, 1), 1);
3464 else
3466 if (GET_CODE (value) == PRE_DEC || GET_CODE (value) == POST_DEC)
3467 inc_amount = -inc_amount;
3469 inc = GEN_INT (inc_amount);
3472 if (! post && REG_P (incloc))
3473 result = incloc;
3474 else
3475 result = lra_create_new_reg (GET_MODE (value), value, new_rclass,
3476 "INC/DEC result");
3478 if (real_in != result)
3480 /* First copy the location to the result register. */
3481 lra_assert (REG_P (result));
3482 emit_insn (gen_move_insn (result, real_in));
3485 /* We suppose that there are insns to add/sub with the constant
3486 increment permitted in {PRE/POST)_{DEC/INC/MODIFY}. At least the
3487 old reload worked with this assumption. If the assumption
3488 becomes wrong, we should use approach in function
3489 base_plus_disp_to_reg. */
3490 if (in == value)
3492 /* See if we can directly increment INCLOC. */
3493 last = get_last_insn ();
3494 add_insn = emit_insn (plus_p
3495 ? gen_add2_insn (incloc, inc)
3496 : gen_sub2_insn (incloc, inc));
3498 code = recog_memoized (add_insn);
3499 if (code >= 0)
3501 if (! post && result != incloc)
3502 emit_insn (gen_move_insn (result, incloc));
3503 return result;
3505 delete_insns_since (last);
3508 /* If couldn't do the increment directly, must increment in RESULT.
3509 The way we do this depends on whether this is pre- or
3510 post-increment. For pre-increment, copy INCLOC to the reload
3511 register, increment it there, then save back. */
3512 if (! post)
3514 if (real_in != result)
3515 emit_insn (gen_move_insn (result, real_in));
3516 if (plus_p)
3517 emit_insn (gen_add2_insn (result, inc));
3518 else
3519 emit_insn (gen_sub2_insn (result, inc));
3520 if (result != incloc)
3521 emit_insn (gen_move_insn (incloc, result));
3523 else
3525 /* Post-increment.
3527 Because this might be a jump insn or a compare, and because
3528 RESULT may not be available after the insn in an input
3529 reload, we must do the incrementing before the insn being
3530 reloaded for.
3532 We have already copied IN to RESULT. Increment the copy in
3533 RESULT, save that back, then decrement RESULT so it has
3534 the original value. */
3535 if (plus_p)
3536 emit_insn (gen_add2_insn (result, inc));
3537 else
3538 emit_insn (gen_sub2_insn (result, inc));
3539 emit_insn (gen_move_insn (incloc, result));
3540 /* Restore non-modified value for the result. We prefer this
3541 way because it does not require an additional hard
3542 register. */
3543 if (plus_p)
3545 if (CONST_INT_P (inc))
3546 emit_insn (gen_add2_insn (result,
3547 gen_int_mode (-INTVAL (inc),
3548 GET_MODE (result))));
3549 else
3550 emit_insn (gen_sub2_insn (result, inc));
3552 else
3553 emit_insn (gen_add2_insn (result, inc));
3555 return result;
3558 /* Return true if the current move insn does not need processing as we
3559 already know that it satisfies its constraints. */
3560 static bool
3561 simple_move_p (void)
3563 rtx dest, src;
3564 enum reg_class dclass, sclass;
3566 lra_assert (curr_insn_set != NULL_RTX);
3567 dest = SET_DEST (curr_insn_set);
3568 src = SET_SRC (curr_insn_set);
3570 /* If the instruction has multiple sets we need to process it even if it
3571 is single_set. This can happen if one or more of the SETs are dead.
3572 See PR73650. */
3573 if (multiple_sets (curr_insn))
3574 return false;
3576 return ((dclass = get_op_class (dest)) != NO_REGS
3577 && (sclass = get_op_class (src)) != NO_REGS
3578 /* The backend guarantees that register moves of cost 2
3579 never need reloads. */
3580 && targetm.register_move_cost (GET_MODE (src), sclass, dclass) == 2);
3583 /* Swap operands NOP and NOP + 1. */
3584 static inline void
3585 swap_operands (int nop)
3587 std::swap (curr_operand_mode[nop], curr_operand_mode[nop + 1]);
3588 std::swap (original_subreg_reg_mode[nop], original_subreg_reg_mode[nop + 1]);
3589 std::swap (*curr_id->operand_loc[nop], *curr_id->operand_loc[nop + 1]);
3590 std::swap (equiv_substition_p[nop], equiv_substition_p[nop + 1]);
3591 /* Swap the duplicates too. */
3592 lra_update_dup (curr_id, nop);
3593 lra_update_dup (curr_id, nop + 1);
3596 /* Main entry point of the constraint code: search the body of the
3597 current insn to choose the best alternative. It is mimicking insn
3598 alternative cost calculation model of former reload pass. That is
3599 because machine descriptions were written to use this model. This
3600 model can be changed in future. Make commutative operand exchange
3601 if it is chosen.
3603 if CHECK_ONLY_P is false, do RTL changes to satisfy the
3604 constraints. Return true if any change happened during function
3605 call.
3607 If CHECK_ONLY_P is true then don't do any transformation. Just
3608 check that the insn satisfies all constraints. If the insn does
3609 not satisfy any constraint, return true. */
3610 static bool
3611 curr_insn_transform (bool check_only_p)
3613 int i, j, k;
3614 int n_operands;
3615 int n_alternatives;
3616 int n_outputs;
3617 int commutative;
3618 signed char goal_alt_matched[MAX_RECOG_OPERANDS][MAX_RECOG_OPERANDS];
3619 signed char match_inputs[MAX_RECOG_OPERANDS + 1];
3620 signed char outputs[MAX_RECOG_OPERANDS + 1];
3621 rtx_insn *before, *after;
3622 bool alt_p = false;
3623 /* Flag that the insn has been changed through a transformation. */
3624 bool change_p;
3625 bool sec_mem_p;
3626 #ifdef SECONDARY_MEMORY_NEEDED
3627 bool use_sec_mem_p;
3628 #endif
3629 int max_regno_before;
3630 int reused_alternative_num;
3632 curr_insn_set = single_set (curr_insn);
3633 if (curr_insn_set != NULL_RTX && simple_move_p ())
3634 return false;
3636 no_input_reloads_p = no_output_reloads_p = false;
3637 goal_alt_number = -1;
3638 change_p = sec_mem_p = false;
3639 /* JUMP_INSNs and CALL_INSNs are not allowed to have any output
3640 reloads; neither are insns that SET cc0. Insns that use CC0 are
3641 not allowed to have any input reloads. */
3642 if (JUMP_P (curr_insn) || CALL_P (curr_insn))
3643 no_output_reloads_p = true;
3645 if (HAVE_cc0 && reg_referenced_p (cc0_rtx, PATTERN (curr_insn)))
3646 no_input_reloads_p = true;
3647 if (HAVE_cc0 && reg_set_p (cc0_rtx, PATTERN (curr_insn)))
3648 no_output_reloads_p = true;
3650 n_operands = curr_static_id->n_operands;
3651 n_alternatives = curr_static_id->n_alternatives;
3653 /* Just return "no reloads" if insn has no operands with
3654 constraints. */
3655 if (n_operands == 0 || n_alternatives == 0)
3656 return false;
3658 max_regno_before = max_reg_num ();
3660 for (i = 0; i < n_operands; i++)
3662 goal_alt_matched[i][0] = -1;
3663 goal_alt_matches[i] = -1;
3666 commutative = curr_static_id->commutative;
3668 /* Now see what we need for pseudos that didn't get hard regs or got
3669 the wrong kind of hard reg. For this, we must consider all the
3670 operands together against the register constraints. */
3672 best_losers = best_overall = INT_MAX;
3673 best_reload_sum = 0;
3675 curr_swapped = false;
3676 goal_alt_swapped = false;
3678 if (! check_only_p)
3679 /* Make equivalence substitution and memory subreg elimination
3680 before address processing because an address legitimacy can
3681 depend on memory mode. */
3682 for (i = 0; i < n_operands; i++)
3684 rtx op, subst, old;
3685 bool op_change_p = false;
3687 if (curr_static_id->operand[i].is_operator)
3688 continue;
3690 old = op = *curr_id->operand_loc[i];
3691 if (GET_CODE (old) == SUBREG)
3692 old = SUBREG_REG (old);
3693 subst = get_equiv_with_elimination (old, curr_insn);
3694 original_subreg_reg_mode[i] = VOIDmode;
3695 equiv_substition_p[i] = false;
3696 if (subst != old)
3698 equiv_substition_p[i] = true;
3699 subst = copy_rtx (subst);
3700 lra_assert (REG_P (old));
3701 if (GET_CODE (op) != SUBREG)
3702 *curr_id->operand_loc[i] = subst;
3703 else
3705 SUBREG_REG (op) = subst;
3706 if (GET_MODE (subst) == VOIDmode)
3707 original_subreg_reg_mode[i] = GET_MODE (old);
3709 if (lra_dump_file != NULL)
3711 fprintf (lra_dump_file,
3712 "Changing pseudo %d in operand %i of insn %u on equiv ",
3713 REGNO (old), i, INSN_UID (curr_insn));
3714 dump_value_slim (lra_dump_file, subst, 1);
3715 fprintf (lra_dump_file, "\n");
3717 op_change_p = change_p = true;
3719 if (simplify_operand_subreg (i, GET_MODE (old)) || op_change_p)
3721 change_p = true;
3722 lra_update_dup (curr_id, i);
3726 /* Reload address registers and displacements. We do it before
3727 finding an alternative because of memory constraints. */
3728 before = after = NULL;
3729 for (i = 0; i < n_operands; i++)
3730 if (! curr_static_id->operand[i].is_operator
3731 && process_address (i, check_only_p, &before, &after))
3733 if (check_only_p)
3734 return true;
3735 change_p = true;
3736 lra_update_dup (curr_id, i);
3739 if (change_p)
3740 /* If we've changed the instruction then any alternative that
3741 we chose previously may no longer be valid. */
3742 lra_set_used_insn_alternative (curr_insn, -1);
3744 if (! check_only_p && curr_insn_set != NULL_RTX
3745 && check_and_process_move (&change_p, &sec_mem_p))
3746 return change_p;
3748 try_swapped:
3750 reused_alternative_num = check_only_p ? -1 : curr_id->used_insn_alternative;
3751 if (lra_dump_file != NULL && reused_alternative_num >= 0)
3752 fprintf (lra_dump_file, "Reusing alternative %d for insn #%u\n",
3753 reused_alternative_num, INSN_UID (curr_insn));
3755 if (process_alt_operands (reused_alternative_num))
3756 alt_p = true;
3758 if (check_only_p)
3759 return ! alt_p || best_losers != 0;
3761 /* If insn is commutative (it's safe to exchange a certain pair of
3762 operands) then we need to try each alternative twice, the second
3763 time matching those two operands as if we had exchanged them. To
3764 do this, really exchange them in operands.
3766 If we have just tried the alternatives the second time, return
3767 operands to normal and drop through. */
3769 if (reused_alternative_num < 0 && commutative >= 0)
3771 curr_swapped = !curr_swapped;
3772 if (curr_swapped)
3774 swap_operands (commutative);
3775 goto try_swapped;
3777 else
3778 swap_operands (commutative);
3781 if (! alt_p && ! sec_mem_p)
3783 /* No alternative works with reloads?? */
3784 if (INSN_CODE (curr_insn) >= 0)
3785 fatal_insn ("unable to generate reloads for:", curr_insn);
3786 error_for_asm (curr_insn,
3787 "inconsistent operand constraints in an %<asm%>");
3788 /* Avoid further trouble with this insn. Don't generate use
3789 pattern here as we could use the insn SP offset. */
3790 lra_set_insn_deleted (curr_insn);
3791 return true;
3794 /* If the best alternative is with operands 1 and 2 swapped, swap
3795 them. Update the operand numbers of any reloads already
3796 pushed. */
3798 if (goal_alt_swapped)
3800 if (lra_dump_file != NULL)
3801 fprintf (lra_dump_file, " Commutative operand exchange in insn %u\n",
3802 INSN_UID (curr_insn));
3804 /* Swap the duplicates too. */
3805 swap_operands (commutative);
3806 change_p = true;
3809 #ifdef SECONDARY_MEMORY_NEEDED
3810 /* Some target macros SECONDARY_MEMORY_NEEDED (e.g. x86) are defined
3811 too conservatively. So we use the secondary memory only if there
3812 is no any alternative without reloads. */
3813 use_sec_mem_p = false;
3814 if (! alt_p)
3815 use_sec_mem_p = true;
3816 else if (sec_mem_p)
3818 for (i = 0; i < n_operands; i++)
3819 if (! goal_alt_win[i] && ! goal_alt_match_win[i])
3820 break;
3821 use_sec_mem_p = i < n_operands;
3824 if (use_sec_mem_p)
3826 int in = -1, out = -1;
3827 rtx new_reg, src, dest, rld;
3828 machine_mode sec_mode, rld_mode;
3830 lra_assert (curr_insn_set != NULL_RTX && sec_mem_p);
3831 dest = SET_DEST (curr_insn_set);
3832 src = SET_SRC (curr_insn_set);
3833 for (i = 0; i < n_operands; i++)
3834 if (*curr_id->operand_loc[i] == dest)
3835 out = i;
3836 else if (*curr_id->operand_loc[i] == src)
3837 in = i;
3838 for (i = 0; i < curr_static_id->n_dups; i++)
3839 if (out < 0 && *curr_id->dup_loc[i] == dest)
3840 out = curr_static_id->dup_num[i];
3841 else if (in < 0 && *curr_id->dup_loc[i] == src)
3842 in = curr_static_id->dup_num[i];
3843 lra_assert (out >= 0 && in >= 0
3844 && curr_static_id->operand[out].type == OP_OUT
3845 && curr_static_id->operand[in].type == OP_IN);
3846 rld = (GET_MODE_SIZE (GET_MODE (dest)) <= GET_MODE_SIZE (GET_MODE (src))
3847 ? dest : src);
3848 rld_mode = GET_MODE (rld);
3849 #ifdef SECONDARY_MEMORY_NEEDED_MODE
3850 sec_mode = SECONDARY_MEMORY_NEEDED_MODE (rld_mode);
3851 #else
3852 sec_mode = rld_mode;
3853 #endif
3854 new_reg = lra_create_new_reg (sec_mode, NULL_RTX,
3855 NO_REGS, "secondary");
3856 /* If the mode is changed, it should be wider. */
3857 lra_assert (GET_MODE_SIZE (sec_mode) >= GET_MODE_SIZE (rld_mode));
3858 if (sec_mode != rld_mode)
3860 /* If the target says specifically to use another mode for
3861 secondary memory moves we can not reuse the original
3862 insn. */
3863 after = emit_spill_move (false, new_reg, dest);
3864 lra_process_new_insns (curr_insn, NULL, after,
3865 "Inserting the sec. move");
3866 /* We may have non null BEFORE here (e.g. after address
3867 processing. */
3868 push_to_sequence (before);
3869 before = emit_spill_move (true, new_reg, src);
3870 emit_insn (before);
3871 before = get_insns ();
3872 end_sequence ();
3873 lra_process_new_insns (curr_insn, before, NULL, "Changing on");
3874 lra_set_insn_deleted (curr_insn);
3876 else if (dest == rld)
3878 *curr_id->operand_loc[out] = new_reg;
3879 lra_update_dup (curr_id, out);
3880 after = emit_spill_move (false, new_reg, dest);
3881 lra_process_new_insns (curr_insn, NULL, after,
3882 "Inserting the sec. move");
3884 else
3886 *curr_id->operand_loc[in] = new_reg;
3887 lra_update_dup (curr_id, in);
3888 /* See comments above. */
3889 push_to_sequence (before);
3890 before = emit_spill_move (true, new_reg, src);
3891 emit_insn (before);
3892 before = get_insns ();
3893 end_sequence ();
3894 lra_process_new_insns (curr_insn, before, NULL,
3895 "Inserting the sec. move");
3897 lra_update_insn_regno_info (curr_insn);
3898 return true;
3900 #endif
3902 lra_assert (goal_alt_number >= 0);
3903 lra_set_used_insn_alternative (curr_insn, goal_alt_number);
3905 if (lra_dump_file != NULL)
3907 const char *p;
3909 fprintf (lra_dump_file, " Choosing alt %d in insn %u:",
3910 goal_alt_number, INSN_UID (curr_insn));
3911 for (i = 0; i < n_operands; i++)
3913 p = (curr_static_id->operand_alternative
3914 [goal_alt_number * n_operands + i].constraint);
3915 if (*p == '\0')
3916 continue;
3917 fprintf (lra_dump_file, " (%d) ", i);
3918 for (; *p != '\0' && *p != ',' && *p != '#'; p++)
3919 fputc (*p, lra_dump_file);
3921 if (INSN_CODE (curr_insn) >= 0
3922 && (p = get_insn_name (INSN_CODE (curr_insn))) != NULL)
3923 fprintf (lra_dump_file, " {%s}", p);
3924 if (curr_id->sp_offset != 0)
3925 fprintf (lra_dump_file, " (sp_off=%" HOST_WIDE_INT_PRINT "d)",
3926 curr_id->sp_offset);
3927 fprintf (lra_dump_file, "\n");
3930 /* Right now, for any pair of operands I and J that are required to
3931 match, with J < I, goal_alt_matches[I] is J. Add I to
3932 goal_alt_matched[J]. */
3934 for (i = 0; i < n_operands; i++)
3935 if ((j = goal_alt_matches[i]) >= 0)
3937 for (k = 0; goal_alt_matched[j][k] >= 0; k++)
3939 /* We allow matching one output operand and several input
3940 operands. */
3941 lra_assert (k == 0
3942 || (curr_static_id->operand[j].type == OP_OUT
3943 && curr_static_id->operand[i].type == OP_IN
3944 && (curr_static_id->operand
3945 [goal_alt_matched[j][0]].type == OP_IN)));
3946 goal_alt_matched[j][k] = i;
3947 goal_alt_matched[j][k + 1] = -1;
3950 for (i = 0; i < n_operands; i++)
3951 goal_alt_win[i] |= goal_alt_match_win[i];
3953 /* Any constants that aren't allowed and can't be reloaded into
3954 registers are here changed into memory references. */
3955 for (i = 0; i < n_operands; i++)
3956 if (goal_alt_win[i])
3958 int regno;
3959 enum reg_class new_class;
3960 rtx reg = *curr_id->operand_loc[i];
3962 if (GET_CODE (reg) == SUBREG)
3963 reg = SUBREG_REG (reg);
3965 if (REG_P (reg) && (regno = REGNO (reg)) >= FIRST_PSEUDO_REGISTER)
3967 bool ok_p = in_class_p (reg, goal_alt[i], &new_class);
3969 if (new_class != NO_REGS && get_reg_class (regno) != new_class)
3971 lra_assert (ok_p);
3972 lra_change_class (regno, new_class, " Change to", true);
3976 else
3978 const char *constraint;
3979 char c;
3980 rtx op = *curr_id->operand_loc[i];
3981 rtx subreg = NULL_RTX;
3982 machine_mode mode = curr_operand_mode[i];
3984 if (GET_CODE (op) == SUBREG)
3986 subreg = op;
3987 op = SUBREG_REG (op);
3988 mode = GET_MODE (op);
3991 if (CONST_POOL_OK_P (mode, op)
3992 && ((targetm.preferred_reload_class
3993 (op, (enum reg_class) goal_alt[i]) == NO_REGS)
3994 || no_input_reloads_p))
3996 rtx tem = force_const_mem (mode, op);
3998 change_p = true;
3999 if (subreg != NULL_RTX)
4000 tem = gen_rtx_SUBREG (mode, tem, SUBREG_BYTE (subreg));
4002 *curr_id->operand_loc[i] = tem;
4003 lra_update_dup (curr_id, i);
4004 process_address (i, false, &before, &after);
4006 /* If the alternative accepts constant pool refs directly
4007 there will be no reload needed at all. */
4008 if (subreg != NULL_RTX)
4009 continue;
4010 /* Skip alternatives before the one requested. */
4011 constraint = (curr_static_id->operand_alternative
4012 [goal_alt_number * n_operands + i].constraint);
4013 for (;
4014 (c = *constraint) && c != ',' && c != '#';
4015 constraint += CONSTRAINT_LEN (c, constraint))
4017 enum constraint_num cn = lookup_constraint (constraint);
4018 if ((insn_extra_memory_constraint (cn)
4019 || insn_extra_special_memory_constraint (cn))
4020 && satisfies_memory_constraint_p (tem, cn))
4021 break;
4023 if (c == '\0' || c == ',' || c == '#')
4024 continue;
4026 goal_alt_win[i] = true;
4030 n_outputs = 0;
4031 outputs[0] = -1;
4032 for (i = 0; i < n_operands; i++)
4034 int regno;
4035 bool optional_p = false;
4036 rtx old, new_reg;
4037 rtx op = *curr_id->operand_loc[i];
4039 if (goal_alt_win[i])
4041 if (goal_alt[i] == NO_REGS
4042 && REG_P (op)
4043 /* When we assign NO_REGS it means that we will not
4044 assign a hard register to the scratch pseudo by
4045 assigment pass and the scratch pseudo will be
4046 spilled. Spilled scratch pseudos are transformed
4047 back to scratches at the LRA end. */
4048 && lra_former_scratch_operand_p (curr_insn, i)
4049 && lra_former_scratch_p (REGNO (op)))
4051 int regno = REGNO (op);
4052 lra_change_class (regno, NO_REGS, " Change to", true);
4053 if (lra_get_regno_hard_regno (regno) >= 0)
4054 /* We don't have to mark all insn affected by the
4055 spilled pseudo as there is only one such insn, the
4056 current one. */
4057 reg_renumber[regno] = -1;
4058 lra_assert (bitmap_single_bit_set_p
4059 (&lra_reg_info[REGNO (op)].insn_bitmap));
4061 /* We can do an optional reload. If the pseudo got a hard
4062 reg, we might improve the code through inheritance. If
4063 it does not get a hard register we coalesce memory/memory
4064 moves later. Ignore move insns to avoid cycling. */
4065 if (! lra_simple_p
4066 && lra_undo_inheritance_iter < LRA_MAX_INHERITANCE_PASSES
4067 && goal_alt[i] != NO_REGS && REG_P (op)
4068 && (regno = REGNO (op)) >= FIRST_PSEUDO_REGISTER
4069 && regno < new_regno_start
4070 && ! lra_former_scratch_p (regno)
4071 && reg_renumber[regno] < 0
4072 /* Check that the optional reload pseudo will be able to
4073 hold given mode value. */
4074 && ! (prohibited_class_reg_set_mode_p
4075 (goal_alt[i], reg_class_contents[goal_alt[i]],
4076 PSEUDO_REGNO_MODE (regno)))
4077 && (curr_insn_set == NULL_RTX
4078 || !((REG_P (SET_SRC (curr_insn_set))
4079 || MEM_P (SET_SRC (curr_insn_set))
4080 || GET_CODE (SET_SRC (curr_insn_set)) == SUBREG)
4081 && (REG_P (SET_DEST (curr_insn_set))
4082 || MEM_P (SET_DEST (curr_insn_set))
4083 || GET_CODE (SET_DEST (curr_insn_set)) == SUBREG))))
4084 optional_p = true;
4085 else
4086 continue;
4089 /* Operands that match previous ones have already been handled. */
4090 if (goal_alt_matches[i] >= 0)
4091 continue;
4093 /* We should not have an operand with a non-offsettable address
4094 appearing where an offsettable address will do. It also may
4095 be a case when the address should be special in other words
4096 not a general one (e.g. it needs no index reg). */
4097 if (goal_alt_matched[i][0] == -1 && goal_alt_offmemok[i] && MEM_P (op))
4099 enum reg_class rclass;
4100 rtx *loc = &XEXP (op, 0);
4101 enum rtx_code code = GET_CODE (*loc);
4103 push_to_sequence (before);
4104 rclass = base_reg_class (GET_MODE (op), MEM_ADDR_SPACE (op),
4105 MEM, SCRATCH);
4106 if (GET_RTX_CLASS (code) == RTX_AUTOINC)
4107 new_reg = emit_inc (rclass, *loc, *loc,
4108 /* This value does not matter for MODIFY. */
4109 GET_MODE_SIZE (GET_MODE (op)));
4110 else if (get_reload_reg (OP_IN, Pmode, *loc, rclass, FALSE,
4111 "offsetable address", &new_reg))
4112 lra_emit_move (new_reg, *loc);
4113 before = get_insns ();
4114 end_sequence ();
4115 *loc = new_reg;
4116 lra_update_dup (curr_id, i);
4118 else if (goal_alt_matched[i][0] == -1)
4120 machine_mode mode;
4121 rtx reg, *loc;
4122 int hard_regno, byte;
4123 enum op_type type = curr_static_id->operand[i].type;
4125 loc = curr_id->operand_loc[i];
4126 mode = curr_operand_mode[i];
4127 if (GET_CODE (*loc) == SUBREG)
4129 reg = SUBREG_REG (*loc);
4130 byte = SUBREG_BYTE (*loc);
4131 if (REG_P (reg)
4132 /* Strict_low_part requires reload the register not
4133 the sub-register. */
4134 && (curr_static_id->operand[i].strict_low
4135 || (GET_MODE_SIZE (mode)
4136 <= GET_MODE_SIZE (GET_MODE (reg))
4137 && (hard_regno
4138 = get_try_hard_regno (REGNO (reg))) >= 0
4139 && (simplify_subreg_regno
4140 (hard_regno,
4141 GET_MODE (reg), byte, mode) < 0)
4142 && (goal_alt[i] == NO_REGS
4143 || (simplify_subreg_regno
4144 (ira_class_hard_regs[goal_alt[i]][0],
4145 GET_MODE (reg), byte, mode) >= 0)))))
4147 /* An OP_INOUT is required when reloading a subreg of a
4148 mode wider than a word to ensure that data beyond the
4149 word being reloaded is preserved. Also automatically
4150 ensure that strict_low_part reloads are made into
4151 OP_INOUT which should already be true from the backend
4152 constraints. */
4153 if (type == OP_OUT
4154 && (curr_static_id->operand[i].strict_low
4155 || (GET_MODE_SIZE (GET_MODE (reg)) > UNITS_PER_WORD
4156 && (GET_MODE_SIZE (mode)
4157 < GET_MODE_SIZE (GET_MODE (reg))))))
4158 type = OP_INOUT;
4159 loc = &SUBREG_REG (*loc);
4160 mode = GET_MODE (*loc);
4163 old = *loc;
4164 if (get_reload_reg (type, mode, old, goal_alt[i],
4165 loc != curr_id->operand_loc[i], "", &new_reg)
4166 && type != OP_OUT)
4168 push_to_sequence (before);
4169 lra_emit_move (new_reg, old);
4170 before = get_insns ();
4171 end_sequence ();
4173 *loc = new_reg;
4174 if (type != OP_IN
4175 && find_reg_note (curr_insn, REG_UNUSED, old) == NULL_RTX)
4177 start_sequence ();
4178 lra_emit_move (type == OP_INOUT ? copy_rtx (old) : old, new_reg);
4179 emit_insn (after);
4180 after = get_insns ();
4181 end_sequence ();
4182 *loc = new_reg;
4184 for (j = 0; j < goal_alt_dont_inherit_ops_num; j++)
4185 if (goal_alt_dont_inherit_ops[j] == i)
4187 lra_set_regno_unique_value (REGNO (new_reg));
4188 break;
4190 lra_update_dup (curr_id, i);
4192 else if (curr_static_id->operand[i].type == OP_IN
4193 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4194 == OP_OUT))
4196 /* generate reloads for input and matched outputs. */
4197 match_inputs[0] = i;
4198 match_inputs[1] = -1;
4199 match_reload (goal_alt_matched[i][0], match_inputs, outputs,
4200 goal_alt[i], &before, &after,
4201 curr_static_id->operand_alternative
4202 [goal_alt_number * n_operands + goal_alt_matched[i][0]]
4203 .earlyclobber);
4205 else if (curr_static_id->operand[i].type == OP_OUT
4206 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4207 == OP_IN))
4208 /* Generate reloads for output and matched inputs. */
4209 match_reload (i, goal_alt_matched[i], outputs, goal_alt[i], &before,
4210 &after, curr_static_id->operand_alternative
4211 [goal_alt_number * n_operands + i].earlyclobber);
4212 else if (curr_static_id->operand[i].type == OP_IN
4213 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4214 == OP_IN))
4216 /* Generate reloads for matched inputs. */
4217 match_inputs[0] = i;
4218 for (j = 0; (k = goal_alt_matched[i][j]) >= 0; j++)
4219 match_inputs[j + 1] = k;
4220 match_inputs[j + 1] = -1;
4221 match_reload (-1, match_inputs, outputs, goal_alt[i], &before,
4222 &after, false);
4224 else
4225 /* We must generate code in any case when function
4226 process_alt_operands decides that it is possible. */
4227 gcc_unreachable ();
4229 /* Memorise processed outputs so that output remaining to be processed
4230 can avoid using the same register value (see match_reload). */
4231 if (curr_static_id->operand[i].type == OP_OUT)
4233 outputs[n_outputs++] = i;
4234 outputs[n_outputs] = -1;
4237 if (optional_p)
4239 rtx reg = op;
4241 lra_assert (REG_P (reg));
4242 regno = REGNO (reg);
4243 op = *curr_id->operand_loc[i]; /* Substitution. */
4244 if (GET_CODE (op) == SUBREG)
4245 op = SUBREG_REG (op);
4246 gcc_assert (REG_P (op) && (int) REGNO (op) >= new_regno_start);
4247 bitmap_set_bit (&lra_optional_reload_pseudos, REGNO (op));
4248 lra_reg_info[REGNO (op)].restore_rtx = reg;
4249 if (lra_dump_file != NULL)
4250 fprintf (lra_dump_file,
4251 " Making reload reg %d for reg %d optional\n",
4252 REGNO (op), regno);
4255 if (before != NULL_RTX || after != NULL_RTX
4256 || max_regno_before != max_reg_num ())
4257 change_p = true;
4258 if (change_p)
4260 lra_update_operator_dups (curr_id);
4261 /* Something changes -- process the insn. */
4262 lra_update_insn_regno_info (curr_insn);
4264 lra_process_new_insns (curr_insn, before, after, "Inserting insn reload");
4265 return change_p;
4268 /* Return true if INSN satisfies all constraints. In other words, no
4269 reload insns are needed. */
4270 bool
4271 lra_constrain_insn (rtx_insn *insn)
4273 int saved_new_regno_start = new_regno_start;
4274 int saved_new_insn_uid_start = new_insn_uid_start;
4275 bool change_p;
4277 curr_insn = insn;
4278 curr_id = lra_get_insn_recog_data (curr_insn);
4279 curr_static_id = curr_id->insn_static_data;
4280 new_insn_uid_start = get_max_uid ();
4281 new_regno_start = max_reg_num ();
4282 change_p = curr_insn_transform (true);
4283 new_regno_start = saved_new_regno_start;
4284 new_insn_uid_start = saved_new_insn_uid_start;
4285 return ! change_p;
4288 /* Return true if X is in LIST. */
4289 static bool
4290 in_list_p (rtx x, rtx list)
4292 for (; list != NULL_RTX; list = XEXP (list, 1))
4293 if (XEXP (list, 0) == x)
4294 return true;
4295 return false;
4298 /* Return true if X contains an allocatable hard register (if
4299 HARD_REG_P) or a (spilled if SPILLED_P) pseudo. */
4300 static bool
4301 contains_reg_p (rtx x, bool hard_reg_p, bool spilled_p)
4303 int i, j;
4304 const char *fmt;
4305 enum rtx_code code;
4307 code = GET_CODE (x);
4308 if (REG_P (x))
4310 int regno = REGNO (x);
4311 HARD_REG_SET alloc_regs;
4313 if (hard_reg_p)
4315 if (regno >= FIRST_PSEUDO_REGISTER)
4316 regno = lra_get_regno_hard_regno (regno);
4317 if (regno < 0)
4318 return false;
4319 COMPL_HARD_REG_SET (alloc_regs, lra_no_alloc_regs);
4320 return overlaps_hard_reg_set_p (alloc_regs, GET_MODE (x), regno);
4322 else
4324 if (regno < FIRST_PSEUDO_REGISTER)
4325 return false;
4326 if (! spilled_p)
4327 return true;
4328 return lra_get_regno_hard_regno (regno) < 0;
4331 fmt = GET_RTX_FORMAT (code);
4332 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4334 if (fmt[i] == 'e')
4336 if (contains_reg_p (XEXP (x, i), hard_reg_p, spilled_p))
4337 return true;
4339 else if (fmt[i] == 'E')
4341 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4342 if (contains_reg_p (XVECEXP (x, i, j), hard_reg_p, spilled_p))
4343 return true;
4346 return false;
4349 /* Process all regs in location *LOC and change them on equivalent
4350 substitution. Return true if any change was done. */
4351 static bool
4352 loc_equivalence_change_p (rtx *loc)
4354 rtx subst, reg, x = *loc;
4355 bool result = false;
4356 enum rtx_code code = GET_CODE (x);
4357 const char *fmt;
4358 int i, j;
4360 if (code == SUBREG)
4362 reg = SUBREG_REG (x);
4363 if ((subst = get_equiv_with_elimination (reg, curr_insn)) != reg
4364 && GET_MODE (subst) == VOIDmode)
4366 /* We cannot reload debug location. Simplify subreg here
4367 while we know the inner mode. */
4368 *loc = simplify_gen_subreg (GET_MODE (x), subst,
4369 GET_MODE (reg), SUBREG_BYTE (x));
4370 return true;
4373 if (code == REG && (subst = get_equiv_with_elimination (x, curr_insn)) != x)
4375 *loc = subst;
4376 return true;
4379 /* Scan all the operand sub-expressions. */
4380 fmt = GET_RTX_FORMAT (code);
4381 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4383 if (fmt[i] == 'e')
4384 result = loc_equivalence_change_p (&XEXP (x, i)) || result;
4385 else if (fmt[i] == 'E')
4386 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4387 result
4388 = loc_equivalence_change_p (&XVECEXP (x, i, j)) || result;
4390 return result;
4393 /* Similar to loc_equivalence_change_p, but for use as
4394 simplify_replace_fn_rtx callback. DATA is insn for which the
4395 elimination is done. If it null we don't do the elimination. */
4396 static rtx
4397 loc_equivalence_callback (rtx loc, const_rtx, void *data)
4399 if (!REG_P (loc))
4400 return NULL_RTX;
4402 rtx subst = (data == NULL
4403 ? get_equiv (loc) : get_equiv_with_elimination (loc, (rtx_insn *) data));
4404 if (subst != loc)
4405 return subst;
4407 return NULL_RTX;
4410 /* Maximum number of generated reload insns per an insn. It is for
4411 preventing this pass cycling in a bug case. */
4412 #define MAX_RELOAD_INSNS_NUMBER LRA_MAX_INSN_RELOADS
4414 /* The current iteration number of this LRA pass. */
4415 int lra_constraint_iter;
4417 /* True if we substituted equiv which needs checking register
4418 allocation correctness because the equivalent value contains
4419 allocatable hard registers or when we restore multi-register
4420 pseudo. */
4421 bool lra_risky_transformations_p;
4423 /* Return true if REGNO is referenced in more than one block. */
4424 static bool
4425 multi_block_pseudo_p (int regno)
4427 basic_block bb = NULL;
4428 unsigned int uid;
4429 bitmap_iterator bi;
4431 if (regno < FIRST_PSEUDO_REGISTER)
4432 return false;
4434 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi)
4435 if (bb == NULL)
4436 bb = BLOCK_FOR_INSN (lra_insn_recog_data[uid]->insn);
4437 else if (BLOCK_FOR_INSN (lra_insn_recog_data[uid]->insn) != bb)
4438 return true;
4439 return false;
4442 /* Return true if LIST contains a deleted insn. */
4443 static bool
4444 contains_deleted_insn_p (rtx_insn_list *list)
4446 for (; list != NULL_RTX; list = list->next ())
4447 if (NOTE_P (list->insn ())
4448 && NOTE_KIND (list->insn ()) == NOTE_INSN_DELETED)
4449 return true;
4450 return false;
4453 /* Return true if X contains a pseudo dying in INSN. */
4454 static bool
4455 dead_pseudo_p (rtx x, rtx_insn *insn)
4457 int i, j;
4458 const char *fmt;
4459 enum rtx_code code;
4461 if (REG_P (x))
4462 return (insn != NULL_RTX
4463 && find_regno_note (insn, REG_DEAD, REGNO (x)) != NULL_RTX);
4464 code = GET_CODE (x);
4465 fmt = GET_RTX_FORMAT (code);
4466 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4468 if (fmt[i] == 'e')
4470 if (dead_pseudo_p (XEXP (x, i), insn))
4471 return true;
4473 else if (fmt[i] == 'E')
4475 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4476 if (dead_pseudo_p (XVECEXP (x, i, j), insn))
4477 return true;
4480 return false;
4483 /* Return true if INSN contains a dying pseudo in INSN right hand
4484 side. */
4485 static bool
4486 insn_rhs_dead_pseudo_p (rtx_insn *insn)
4488 rtx set = single_set (insn);
4490 gcc_assert (set != NULL);
4491 return dead_pseudo_p (SET_SRC (set), insn);
4494 /* Return true if any init insn of REGNO contains a dying pseudo in
4495 insn right hand side. */
4496 static bool
4497 init_insn_rhs_dead_pseudo_p (int regno)
4499 rtx_insn_list *insns = ira_reg_equiv[regno].init_insns;
4501 if (insns == NULL)
4502 return false;
4503 for (; insns != NULL_RTX; insns = insns->next ())
4504 if (insn_rhs_dead_pseudo_p (insns->insn ()))
4505 return true;
4506 return false;
4509 /* Return TRUE if REGNO has a reverse equivalence. The equivalence is
4510 reverse only if we have one init insn with given REGNO as a
4511 source. */
4512 static bool
4513 reverse_equiv_p (int regno)
4515 rtx_insn_list *insns = ira_reg_equiv[regno].init_insns;
4516 rtx set;
4518 if (insns == NULL)
4519 return false;
4520 if (! INSN_P (insns->insn ())
4521 || insns->next () != NULL)
4522 return false;
4523 if ((set = single_set (insns->insn ())) == NULL_RTX)
4524 return false;
4525 return REG_P (SET_SRC (set)) && (int) REGNO (SET_SRC (set)) == regno;
4528 /* Return TRUE if REGNO was reloaded in an equivalence init insn. We
4529 call this function only for non-reverse equivalence. */
4530 static bool
4531 contains_reloaded_insn_p (int regno)
4533 rtx set;
4534 rtx_insn_list *list = ira_reg_equiv[regno].init_insns;
4536 for (; list != NULL; list = list->next ())
4537 if ((set = single_set (list->insn ())) == NULL_RTX
4538 || ! REG_P (SET_DEST (set))
4539 || (int) REGNO (SET_DEST (set)) != regno)
4540 return true;
4541 return false;
4544 /* Entry function of LRA constraint pass. Return true if the
4545 constraint pass did change the code. */
4546 bool
4547 lra_constraints (bool first_p)
4549 bool changed_p;
4550 int i, hard_regno, new_insns_num;
4551 unsigned int min_len, new_min_len, uid;
4552 rtx set, x, reg, dest_reg;
4553 basic_block last_bb;
4554 bitmap_head equiv_insn_bitmap;
4555 bitmap_iterator bi;
4557 lra_constraint_iter++;
4558 if (lra_dump_file != NULL)
4559 fprintf (lra_dump_file, "\n********** Local #%d: **********\n\n",
4560 lra_constraint_iter);
4561 changed_p = false;
4562 if (pic_offset_table_rtx
4563 && REGNO (pic_offset_table_rtx) >= FIRST_PSEUDO_REGISTER)
4564 lra_risky_transformations_p = true;
4565 else
4566 /* On the first iteration we should check IRA assignment
4567 correctness. In rare cases, the assignments can be wrong as
4568 early clobbers operands are ignored in IRA. */
4569 lra_risky_transformations_p = first_p;
4570 new_insn_uid_start = get_max_uid ();
4571 new_regno_start = first_p ? lra_constraint_new_regno_start : max_reg_num ();
4572 /* Mark used hard regs for target stack size calulations. */
4573 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4574 if (lra_reg_info[i].nrefs != 0
4575 && (hard_regno = lra_get_regno_hard_regno (i)) >= 0)
4577 int j, nregs;
4579 nregs = hard_regno_nregs[hard_regno][lra_reg_info[i].biggest_mode];
4580 for (j = 0; j < nregs; j++)
4581 df_set_regs_ever_live (hard_regno + j, true);
4583 /* Do elimination before the equivalence processing as we can spill
4584 some pseudos during elimination. */
4585 lra_eliminate (false, first_p);
4586 bitmap_initialize (&equiv_insn_bitmap, &reg_obstack);
4587 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4588 if (lra_reg_info[i].nrefs != 0)
4590 ira_reg_equiv[i].profitable_p = true;
4591 reg = regno_reg_rtx[i];
4592 if (lra_get_regno_hard_regno (i) < 0 && (x = get_equiv (reg)) != reg)
4594 bool pseudo_p = contains_reg_p (x, false, false);
4596 /* After RTL transformation, we can not guarantee that
4597 pseudo in the substitution was not reloaded which might
4598 make equivalence invalid. For example, in reverse
4599 equiv of p0
4601 p0 <- ...
4603 equiv_mem <- p0
4605 the memory address register was reloaded before the 2nd
4606 insn. */
4607 if ((! first_p && pseudo_p)
4608 /* We don't use DF for compilation speed sake. So it
4609 is problematic to update live info when we use an
4610 equivalence containing pseudos in more than one
4611 BB. */
4612 || (pseudo_p && multi_block_pseudo_p (i))
4613 /* If an init insn was deleted for some reason, cancel
4614 the equiv. We could update the equiv insns after
4615 transformations including an equiv insn deletion
4616 but it is not worthy as such cases are extremely
4617 rare. */
4618 || contains_deleted_insn_p (ira_reg_equiv[i].init_insns)
4619 /* If it is not a reverse equivalence, we check that a
4620 pseudo in rhs of the init insn is not dying in the
4621 insn. Otherwise, the live info at the beginning of
4622 the corresponding BB might be wrong after we
4623 removed the insn. When the equiv can be a
4624 constant, the right hand side of the init insn can
4625 be a pseudo. */
4626 || (! reverse_equiv_p (i)
4627 && (init_insn_rhs_dead_pseudo_p (i)
4628 /* If we reloaded the pseudo in an equivalence
4629 init insn, we can not remove the equiv init
4630 insns and the init insns might write into
4631 const memory in this case. */
4632 || contains_reloaded_insn_p (i)))
4633 /* Prevent access beyond equivalent memory for
4634 paradoxical subregs. */
4635 || (MEM_P (x)
4636 && (GET_MODE_SIZE (lra_reg_info[i].biggest_mode)
4637 > GET_MODE_SIZE (GET_MODE (x))))
4638 || (pic_offset_table_rtx
4639 && ((CONST_POOL_OK_P (PSEUDO_REGNO_MODE (i), x)
4640 && (targetm.preferred_reload_class
4641 (x, lra_get_allocno_class (i)) == NO_REGS))
4642 || contains_symbol_ref_p (x))))
4643 ira_reg_equiv[i].defined_p = false;
4644 if (contains_reg_p (x, false, true))
4645 ira_reg_equiv[i].profitable_p = false;
4646 if (get_equiv (reg) != reg)
4647 bitmap_ior_into (&equiv_insn_bitmap, &lra_reg_info[i].insn_bitmap);
4650 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4651 update_equiv (i);
4652 /* We should add all insns containing pseudos which should be
4653 substituted by their equivalences. */
4654 EXECUTE_IF_SET_IN_BITMAP (&equiv_insn_bitmap, 0, uid, bi)
4655 lra_push_insn_by_uid (uid);
4656 min_len = lra_insn_stack_length ();
4657 new_insns_num = 0;
4658 last_bb = NULL;
4659 changed_p = false;
4660 while ((new_min_len = lra_insn_stack_length ()) != 0)
4662 curr_insn = lra_pop_insn ();
4663 --new_min_len;
4664 curr_bb = BLOCK_FOR_INSN (curr_insn);
4665 if (curr_bb != last_bb)
4667 last_bb = curr_bb;
4668 bb_reload_num = lra_curr_reload_num;
4670 if (min_len > new_min_len)
4672 min_len = new_min_len;
4673 new_insns_num = 0;
4675 if (new_insns_num > MAX_RELOAD_INSNS_NUMBER)
4676 internal_error
4677 ("Max. number of generated reload insns per insn is achieved (%d)\n",
4678 MAX_RELOAD_INSNS_NUMBER);
4679 new_insns_num++;
4680 if (DEBUG_INSN_P (curr_insn))
4682 /* We need to check equivalence in debug insn and change
4683 pseudo to the equivalent value if necessary. */
4684 curr_id = lra_get_insn_recog_data (curr_insn);
4685 if (bitmap_bit_p (&equiv_insn_bitmap, INSN_UID (curr_insn)))
4687 rtx old = *curr_id->operand_loc[0];
4688 *curr_id->operand_loc[0]
4689 = simplify_replace_fn_rtx (old, NULL_RTX,
4690 loc_equivalence_callback, curr_insn);
4691 if (old != *curr_id->operand_loc[0])
4693 lra_update_insn_regno_info (curr_insn);
4694 changed_p = true;
4698 else if (INSN_P (curr_insn))
4700 if ((set = single_set (curr_insn)) != NULL_RTX)
4702 dest_reg = SET_DEST (set);
4703 /* The equivalence pseudo could be set up as SUBREG in a
4704 case when it is a call restore insn in a mode
4705 different from the pseudo mode. */
4706 if (GET_CODE (dest_reg) == SUBREG)
4707 dest_reg = SUBREG_REG (dest_reg);
4708 if ((REG_P (dest_reg)
4709 && (x = get_equiv (dest_reg)) != dest_reg
4710 /* Remove insns which set up a pseudo whose value
4711 can not be changed. Such insns might be not in
4712 init_insns because we don't update equiv data
4713 during insn transformations.
4715 As an example, let suppose that a pseudo got
4716 hard register and on the 1st pass was not
4717 changed to equivalent constant. We generate an
4718 additional insn setting up the pseudo because of
4719 secondary memory movement. Then the pseudo is
4720 spilled and we use the equiv constant. In this
4721 case we should remove the additional insn and
4722 this insn is not init_insns list. */
4723 && (! MEM_P (x) || MEM_READONLY_P (x)
4724 /* Check that this is actually an insn setting
4725 up the equivalence. */
4726 || in_list_p (curr_insn,
4727 ira_reg_equiv
4728 [REGNO (dest_reg)].init_insns)))
4729 || (((x = get_equiv (SET_SRC (set))) != SET_SRC (set))
4730 && in_list_p (curr_insn,
4731 ira_reg_equiv
4732 [REGNO (SET_SRC (set))].init_insns)))
4734 /* This is equiv init insn of pseudo which did not get a
4735 hard register -- remove the insn. */
4736 if (lra_dump_file != NULL)
4738 fprintf (lra_dump_file,
4739 " Removing equiv init insn %i (freq=%d)\n",
4740 INSN_UID (curr_insn),
4741 REG_FREQ_FROM_BB (BLOCK_FOR_INSN (curr_insn)));
4742 dump_insn_slim (lra_dump_file, curr_insn);
4744 if (contains_reg_p (x, true, false))
4745 lra_risky_transformations_p = true;
4746 lra_set_insn_deleted (curr_insn);
4747 continue;
4750 curr_id = lra_get_insn_recog_data (curr_insn);
4751 curr_static_id = curr_id->insn_static_data;
4752 init_curr_insn_input_reloads ();
4753 init_curr_operand_mode ();
4754 if (curr_insn_transform (false))
4755 changed_p = true;
4756 /* Check non-transformed insns too for equiv change as USE
4757 or CLOBBER don't need reloads but can contain pseudos
4758 being changed on their equivalences. */
4759 else if (bitmap_bit_p (&equiv_insn_bitmap, INSN_UID (curr_insn))
4760 && loc_equivalence_change_p (&PATTERN (curr_insn)))
4762 lra_update_insn_regno_info (curr_insn);
4763 changed_p = true;
4767 bitmap_clear (&equiv_insn_bitmap);
4768 /* If we used a new hard regno, changed_p should be true because the
4769 hard reg is assigned to a new pseudo. */
4770 if (flag_checking && !changed_p)
4772 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4773 if (lra_reg_info[i].nrefs != 0
4774 && (hard_regno = lra_get_regno_hard_regno (i)) >= 0)
4776 int j, nregs = hard_regno_nregs[hard_regno][PSEUDO_REGNO_MODE (i)];
4778 for (j = 0; j < nregs; j++)
4779 lra_assert (df_regs_ever_live_p (hard_regno + j));
4782 return changed_p;
4785 static void initiate_invariants (void);
4786 static void finish_invariants (void);
4788 /* Initiate the LRA constraint pass. It is done once per
4789 function. */
4790 void
4791 lra_constraints_init (void)
4793 initiate_invariants ();
4796 /* Finalize the LRA constraint pass. It is done once per
4797 function. */
4798 void
4799 lra_constraints_finish (void)
4801 finish_invariants ();
4806 /* Structure describes invariants for ineheritance. */
4807 struct lra_invariant
4809 /* The order number of the invariant. */
4810 int num;
4811 /* The invariant RTX. */
4812 rtx invariant_rtx;
4813 /* The origin insn of the invariant. */
4814 rtx_insn *insn;
4817 typedef lra_invariant invariant_t;
4818 typedef invariant_t *invariant_ptr_t;
4819 typedef const invariant_t *const_invariant_ptr_t;
4821 /* Pointer to the inheritance invariants. */
4822 static vec<invariant_ptr_t> invariants;
4824 /* Allocation pool for the invariants. */
4825 static object_allocator<lra_invariant> *invariants_pool;
4827 /* Hash table for the invariants. */
4828 static htab_t invariant_table;
4830 /* Hash function for INVARIANT. */
4831 static hashval_t
4832 invariant_hash (const void *invariant)
4834 rtx inv = ((const_invariant_ptr_t) invariant)->invariant_rtx;
4835 return lra_rtx_hash (inv);
4838 /* Equal function for invariants INVARIANT1 and INVARIANT2. */
4839 static int
4840 invariant_eq_p (const void *invariant1, const void *invariant2)
4842 rtx inv1 = ((const_invariant_ptr_t) invariant1)->invariant_rtx;
4843 rtx inv2 = ((const_invariant_ptr_t) invariant2)->invariant_rtx;
4845 return rtx_equal_p (inv1, inv2);
4848 /* Insert INVARIANT_RTX into the table if it is not there yet. Return
4849 invariant which is in the table. */
4850 static invariant_ptr_t
4851 insert_invariant (rtx invariant_rtx)
4853 void **entry_ptr;
4854 invariant_t invariant;
4855 invariant_ptr_t invariant_ptr;
4857 invariant.invariant_rtx = invariant_rtx;
4858 entry_ptr = htab_find_slot (invariant_table, &invariant, INSERT);
4859 if (*entry_ptr == NULL)
4861 invariant_ptr = invariants_pool->allocate ();
4862 invariant_ptr->invariant_rtx = invariant_rtx;
4863 invariant_ptr->insn = NULL;
4864 invariants.safe_push (invariant_ptr);
4865 *entry_ptr = (void *) invariant_ptr;
4867 return (invariant_ptr_t) *entry_ptr;
4870 /* Initiate the invariant table. */
4871 static void
4872 initiate_invariants (void)
4874 invariants.create (100);
4875 invariants_pool
4876 = new object_allocator<lra_invariant> ("Inheritance invariants");
4877 invariant_table = htab_create (100, invariant_hash, invariant_eq_p, NULL);
4880 /* Finish the invariant table. */
4881 static void
4882 finish_invariants (void)
4884 htab_delete (invariant_table);
4885 delete invariants_pool;
4886 invariants.release ();
4889 /* Make the invariant table empty. */
4890 static void
4891 clear_invariants (void)
4893 htab_empty (invariant_table);
4894 invariants_pool->release ();
4895 invariants.truncate (0);
4900 /* This page contains code to do inheritance/split
4901 transformations. */
4903 /* Number of reloads passed so far in current EBB. */
4904 static int reloads_num;
4906 /* Number of calls passed so far in current EBB. */
4907 static int calls_num;
4909 /* Current reload pseudo check for validity of elements in
4910 USAGE_INSNS. */
4911 static int curr_usage_insns_check;
4913 /* Info about last usage of registers in EBB to do inheritance/split
4914 transformation. Inheritance transformation is done from a spilled
4915 pseudo and split transformations from a hard register or a pseudo
4916 assigned to a hard register. */
4917 struct usage_insns
4919 /* If the value is equal to CURR_USAGE_INSNS_CHECK, then the member
4920 value INSNS is valid. The insns is chain of optional debug insns
4921 and a finishing non-debug insn using the corresponding reg. The
4922 value is also used to mark the registers which are set up in the
4923 current insn. The negated insn uid is used for this. */
4924 int check;
4925 /* Value of global reloads_num at the last insn in INSNS. */
4926 int reloads_num;
4927 /* Value of global reloads_nums at the last insn in INSNS. */
4928 int calls_num;
4929 /* It can be true only for splitting. And it means that the restore
4930 insn should be put after insn given by the following member. */
4931 bool after_p;
4932 /* Next insns in the current EBB which use the original reg and the
4933 original reg value is not changed between the current insn and
4934 the next insns. In order words, e.g. for inheritance, if we need
4935 to use the original reg value again in the next insns we can try
4936 to use the value in a hard register from a reload insn of the
4937 current insn. */
4938 rtx insns;
4941 /* Map: regno -> corresponding pseudo usage insns. */
4942 static struct usage_insns *usage_insns;
4944 static void
4945 setup_next_usage_insn (int regno, rtx insn, int reloads_num, bool after_p)
4947 usage_insns[regno].check = curr_usage_insns_check;
4948 usage_insns[regno].insns = insn;
4949 usage_insns[regno].reloads_num = reloads_num;
4950 usage_insns[regno].calls_num = calls_num;
4951 usage_insns[regno].after_p = after_p;
4954 /* The function is used to form list REGNO usages which consists of
4955 optional debug insns finished by a non-debug insn using REGNO.
4956 RELOADS_NUM is current number of reload insns processed so far. */
4957 static void
4958 add_next_usage_insn (int regno, rtx_insn *insn, int reloads_num)
4960 rtx next_usage_insns;
4962 if (usage_insns[regno].check == curr_usage_insns_check
4963 && (next_usage_insns = usage_insns[regno].insns) != NULL_RTX
4964 && DEBUG_INSN_P (insn))
4966 /* Check that we did not add the debug insn yet. */
4967 if (next_usage_insns != insn
4968 && (GET_CODE (next_usage_insns) != INSN_LIST
4969 || XEXP (next_usage_insns, 0) != insn))
4970 usage_insns[regno].insns = gen_rtx_INSN_LIST (VOIDmode, insn,
4971 next_usage_insns);
4973 else if (NONDEBUG_INSN_P (insn))
4974 setup_next_usage_insn (regno, insn, reloads_num, false);
4975 else
4976 usage_insns[regno].check = 0;
4979 /* Return first non-debug insn in list USAGE_INSNS. */
4980 static rtx_insn *
4981 skip_usage_debug_insns (rtx usage_insns)
4983 rtx insn;
4985 /* Skip debug insns. */
4986 for (insn = usage_insns;
4987 insn != NULL_RTX && GET_CODE (insn) == INSN_LIST;
4988 insn = XEXP (insn, 1))
4990 return safe_as_a <rtx_insn *> (insn);
4993 /* Return true if we need secondary memory moves for insn in
4994 USAGE_INSNS after inserting inherited pseudo of class INHER_CL
4995 into the insn. */
4996 static bool
4997 check_secondary_memory_needed_p (enum reg_class inher_cl ATTRIBUTE_UNUSED,
4998 rtx usage_insns ATTRIBUTE_UNUSED)
5000 #ifndef SECONDARY_MEMORY_NEEDED
5001 return false;
5002 #else
5003 rtx_insn *insn;
5004 rtx set, dest;
5005 enum reg_class cl;
5007 if (inher_cl == ALL_REGS
5008 || (insn = skip_usage_debug_insns (usage_insns)) == NULL_RTX)
5009 return false;
5010 lra_assert (INSN_P (insn));
5011 if ((set = single_set (insn)) == NULL_RTX || ! REG_P (SET_DEST (set)))
5012 return false;
5013 dest = SET_DEST (set);
5014 if (! REG_P (dest))
5015 return false;
5016 lra_assert (inher_cl != NO_REGS);
5017 cl = get_reg_class (REGNO (dest));
5018 return (cl != NO_REGS && cl != ALL_REGS
5019 && SECONDARY_MEMORY_NEEDED (inher_cl, cl, GET_MODE (dest)));
5020 #endif
5023 /* Registers involved in inheritance/split in the current EBB
5024 (inheritance/split pseudos and original registers). */
5025 static bitmap_head check_only_regs;
5027 /* Reload pseudos can not be involded in invariant inheritance in the
5028 current EBB. */
5029 static bitmap_head invalid_invariant_regs;
5031 /* Do inheritance transformations for insn INSN, which defines (if
5032 DEF_P) or uses ORIGINAL_REGNO. NEXT_USAGE_INSNS specifies which
5033 instruction in the EBB next uses ORIGINAL_REGNO; it has the same
5034 form as the "insns" field of usage_insns. Return true if we
5035 succeed in such transformation.
5037 The transformations look like:
5039 p <- ... i <- ...
5040 ... p <- i (new insn)
5041 ... =>
5042 <- ... p ... <- ... i ...
5044 ... i <- p (new insn)
5045 <- ... p ... <- ... i ...
5046 ... =>
5047 <- ... p ... <- ... i ...
5048 where p is a spilled original pseudo and i is a new inheritance pseudo.
5051 The inheritance pseudo has the smallest class of two classes CL and
5052 class of ORIGINAL REGNO. */
5053 static bool
5054 inherit_reload_reg (bool def_p, int original_regno,
5055 enum reg_class cl, rtx_insn *insn, rtx next_usage_insns)
5057 if (optimize_function_for_size_p (cfun))
5058 return false;
5060 enum reg_class rclass = lra_get_allocno_class (original_regno);
5061 rtx original_reg = regno_reg_rtx[original_regno];
5062 rtx new_reg, usage_insn;
5063 rtx_insn *new_insns;
5065 lra_assert (! usage_insns[original_regno].after_p);
5066 if (lra_dump_file != NULL)
5067 fprintf (lra_dump_file,
5068 " <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<\n");
5069 if (! ira_reg_classes_intersect_p[cl][rclass])
5071 if (lra_dump_file != NULL)
5073 fprintf (lra_dump_file,
5074 " Rejecting inheritance for %d "
5075 "because of disjoint classes %s and %s\n",
5076 original_regno, reg_class_names[cl],
5077 reg_class_names[rclass]);
5078 fprintf (lra_dump_file,
5079 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5081 return false;
5083 if ((ira_class_subset_p[cl][rclass] && cl != rclass)
5084 /* We don't use a subset of two classes because it can be
5085 NO_REGS. This transformation is still profitable in most
5086 cases even if the classes are not intersected as register
5087 move is probably cheaper than a memory load. */
5088 || ira_class_hard_regs_num[cl] < ira_class_hard_regs_num[rclass])
5090 if (lra_dump_file != NULL)
5091 fprintf (lra_dump_file, " Use smallest class of %s and %s\n",
5092 reg_class_names[cl], reg_class_names[rclass]);
5094 rclass = cl;
5096 if (check_secondary_memory_needed_p (rclass, next_usage_insns))
5098 /* Reject inheritance resulting in secondary memory moves.
5099 Otherwise, there is a danger in LRA cycling. Also such
5100 transformation will be unprofitable. */
5101 if (lra_dump_file != NULL)
5103 rtx_insn *insn = skip_usage_debug_insns (next_usage_insns);
5104 rtx set = single_set (insn);
5106 lra_assert (set != NULL_RTX);
5108 rtx dest = SET_DEST (set);
5110 lra_assert (REG_P (dest));
5111 fprintf (lra_dump_file,
5112 " Rejecting inheritance for insn %d(%s)<-%d(%s) "
5113 "as secondary mem is needed\n",
5114 REGNO (dest), reg_class_names[get_reg_class (REGNO (dest))],
5115 original_regno, reg_class_names[rclass]);
5116 fprintf (lra_dump_file,
5117 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5119 return false;
5121 new_reg = lra_create_new_reg (GET_MODE (original_reg), original_reg,
5122 rclass, "inheritance");
5123 start_sequence ();
5124 if (def_p)
5125 lra_emit_move (original_reg, new_reg);
5126 else
5127 lra_emit_move (new_reg, original_reg);
5128 new_insns = get_insns ();
5129 end_sequence ();
5130 if (NEXT_INSN (new_insns) != NULL_RTX)
5132 if (lra_dump_file != NULL)
5134 fprintf (lra_dump_file,
5135 " Rejecting inheritance %d->%d "
5136 "as it results in 2 or more insns:\n",
5137 original_regno, REGNO (new_reg));
5138 dump_rtl_slim (lra_dump_file, new_insns, NULL, -1, 0);
5139 fprintf (lra_dump_file,
5140 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5142 return false;
5144 lra_substitute_pseudo_within_insn (insn, original_regno, new_reg, false);
5145 lra_update_insn_regno_info (insn);
5146 if (! def_p)
5147 /* We now have a new usage insn for original regno. */
5148 setup_next_usage_insn (original_regno, new_insns, reloads_num, false);
5149 if (lra_dump_file != NULL)
5150 fprintf (lra_dump_file, " Original reg change %d->%d (bb%d):\n",
5151 original_regno, REGNO (new_reg), BLOCK_FOR_INSN (insn)->index);
5152 lra_reg_info[REGNO (new_reg)].restore_rtx = regno_reg_rtx[original_regno];
5153 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
5154 bitmap_set_bit (&check_only_regs, original_regno);
5155 bitmap_set_bit (&lra_inheritance_pseudos, REGNO (new_reg));
5156 if (def_p)
5157 lra_process_new_insns (insn, NULL, new_insns,
5158 "Add original<-inheritance");
5159 else
5160 lra_process_new_insns (insn, new_insns, NULL,
5161 "Add inheritance<-original");
5162 while (next_usage_insns != NULL_RTX)
5164 if (GET_CODE (next_usage_insns) != INSN_LIST)
5166 usage_insn = next_usage_insns;
5167 lra_assert (NONDEBUG_INSN_P (usage_insn));
5168 next_usage_insns = NULL;
5170 else
5172 usage_insn = XEXP (next_usage_insns, 0);
5173 lra_assert (DEBUG_INSN_P (usage_insn));
5174 next_usage_insns = XEXP (next_usage_insns, 1);
5176 lra_substitute_pseudo (&usage_insn, original_regno, new_reg, false);
5177 lra_update_insn_regno_info (as_a <rtx_insn *> (usage_insn));
5178 if (lra_dump_file != NULL)
5180 fprintf (lra_dump_file,
5181 " Inheritance reuse change %d->%d (bb%d):\n",
5182 original_regno, REGNO (new_reg),
5183 BLOCK_FOR_INSN (usage_insn)->index);
5184 dump_insn_slim (lra_dump_file, as_a <rtx_insn *> (usage_insn));
5187 if (lra_dump_file != NULL)
5188 fprintf (lra_dump_file,
5189 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5190 return true;
5193 /* Return true if we need a caller save/restore for pseudo REGNO which
5194 was assigned to a hard register. */
5195 static inline bool
5196 need_for_call_save_p (int regno)
5198 lra_assert (regno >= FIRST_PSEUDO_REGISTER && reg_renumber[regno] >= 0);
5199 return (usage_insns[regno].calls_num < calls_num
5200 && (overlaps_hard_reg_set_p
5201 ((flag_ipa_ra &&
5202 ! hard_reg_set_empty_p (lra_reg_info[regno].actual_call_used_reg_set))
5203 ? lra_reg_info[regno].actual_call_used_reg_set
5204 : call_used_reg_set,
5205 PSEUDO_REGNO_MODE (regno), reg_renumber[regno])
5206 || HARD_REGNO_CALL_PART_CLOBBERED (reg_renumber[regno],
5207 PSEUDO_REGNO_MODE (regno))));
5210 /* Global registers occurring in the current EBB. */
5211 static bitmap_head ebb_global_regs;
5213 /* Return true if we need a split for hard register REGNO or pseudo
5214 REGNO which was assigned to a hard register.
5215 POTENTIAL_RELOAD_HARD_REGS contains hard registers which might be
5216 used for reloads since the EBB end. It is an approximation of the
5217 used hard registers in the split range. The exact value would
5218 require expensive calculations. If we were aggressive with
5219 splitting because of the approximation, the split pseudo will save
5220 the same hard register assignment and will be removed in the undo
5221 pass. We still need the approximation because too aggressive
5222 splitting would result in too inaccurate cost calculation in the
5223 assignment pass because of too many generated moves which will be
5224 probably removed in the undo pass. */
5225 static inline bool
5226 need_for_split_p (HARD_REG_SET potential_reload_hard_regs, int regno)
5228 int hard_regno = regno < FIRST_PSEUDO_REGISTER ? regno : reg_renumber[regno];
5230 lra_assert (hard_regno >= 0);
5231 return ((TEST_HARD_REG_BIT (potential_reload_hard_regs, hard_regno)
5232 /* Don't split eliminable hard registers, otherwise we can
5233 split hard registers like hard frame pointer, which
5234 lives on BB start/end according to DF-infrastructure,
5235 when there is a pseudo assigned to the register and
5236 living in the same BB. */
5237 && (regno >= FIRST_PSEUDO_REGISTER
5238 || ! TEST_HARD_REG_BIT (eliminable_regset, hard_regno))
5239 && ! TEST_HARD_REG_BIT (lra_no_alloc_regs, hard_regno)
5240 /* Don't split call clobbered hard regs living through
5241 calls, otherwise we might have a check problem in the
5242 assign sub-pass as in the most cases (exception is a
5243 situation when lra_risky_transformations_p value is
5244 true) the assign pass assumes that all pseudos living
5245 through calls are assigned to call saved hard regs. */
5246 && (regno >= FIRST_PSEUDO_REGISTER
5247 || ! TEST_HARD_REG_BIT (call_used_reg_set, regno)
5248 || usage_insns[regno].calls_num == calls_num)
5249 /* We need at least 2 reloads to make pseudo splitting
5250 profitable. We should provide hard regno splitting in
5251 any case to solve 1st insn scheduling problem when
5252 moving hard register definition up might result in
5253 impossibility to find hard register for reload pseudo of
5254 small register class. */
5255 && (usage_insns[regno].reloads_num
5256 + (regno < FIRST_PSEUDO_REGISTER ? 0 : 3) < reloads_num)
5257 && (regno < FIRST_PSEUDO_REGISTER
5258 /* For short living pseudos, spilling + inheritance can
5259 be considered a substitution for splitting.
5260 Therefore we do not splitting for local pseudos. It
5261 decreases also aggressiveness of splitting. The
5262 minimal number of references is chosen taking into
5263 account that for 2 references splitting has no sense
5264 as we can just spill the pseudo. */
5265 || (regno >= FIRST_PSEUDO_REGISTER
5266 && lra_reg_info[regno].nrefs > 3
5267 && bitmap_bit_p (&ebb_global_regs, regno))))
5268 || (regno >= FIRST_PSEUDO_REGISTER && need_for_call_save_p (regno)));
5271 /* Return class for the split pseudo created from original pseudo with
5272 ALLOCNO_CLASS and MODE which got a hard register HARD_REGNO. We
5273 choose subclass of ALLOCNO_CLASS which contains HARD_REGNO and
5274 results in no secondary memory movements. */
5275 static enum reg_class
5276 choose_split_class (enum reg_class allocno_class,
5277 int hard_regno ATTRIBUTE_UNUSED,
5278 machine_mode mode ATTRIBUTE_UNUSED)
5280 #ifndef SECONDARY_MEMORY_NEEDED
5281 return allocno_class;
5282 #else
5283 int i;
5284 enum reg_class cl, best_cl = NO_REGS;
5285 enum reg_class hard_reg_class ATTRIBUTE_UNUSED
5286 = REGNO_REG_CLASS (hard_regno);
5288 if (! SECONDARY_MEMORY_NEEDED (allocno_class, allocno_class, mode)
5289 && TEST_HARD_REG_BIT (reg_class_contents[allocno_class], hard_regno))
5290 return allocno_class;
5291 for (i = 0;
5292 (cl = reg_class_subclasses[allocno_class][i]) != LIM_REG_CLASSES;
5293 i++)
5294 if (! SECONDARY_MEMORY_NEEDED (cl, hard_reg_class, mode)
5295 && ! SECONDARY_MEMORY_NEEDED (hard_reg_class, cl, mode)
5296 && TEST_HARD_REG_BIT (reg_class_contents[cl], hard_regno)
5297 && (best_cl == NO_REGS
5298 || ira_class_hard_regs_num[best_cl] < ira_class_hard_regs_num[cl]))
5299 best_cl = cl;
5300 return best_cl;
5301 #endif
5304 /* Do split transformations for insn INSN, which defines or uses
5305 ORIGINAL_REGNO. NEXT_USAGE_INSNS specifies which instruction in
5306 the EBB next uses ORIGINAL_REGNO; it has the same form as the
5307 "insns" field of usage_insns.
5309 The transformations look like:
5311 p <- ... p <- ...
5312 ... s <- p (new insn -- save)
5313 ... =>
5314 ... p <- s (new insn -- restore)
5315 <- ... p ... <- ... p ...
5317 <- ... p ... <- ... p ...
5318 ... s <- p (new insn -- save)
5319 ... =>
5320 ... p <- s (new insn -- restore)
5321 <- ... p ... <- ... p ...
5323 where p is an original pseudo got a hard register or a hard
5324 register and s is a new split pseudo. The save is put before INSN
5325 if BEFORE_P is true. Return true if we succeed in such
5326 transformation. */
5327 static bool
5328 split_reg (bool before_p, int original_regno, rtx_insn *insn,
5329 rtx next_usage_insns)
5331 enum reg_class rclass;
5332 rtx original_reg;
5333 int hard_regno, nregs;
5334 rtx new_reg, usage_insn;
5335 rtx_insn *restore, *save;
5336 bool after_p;
5337 bool call_save_p;
5338 machine_mode mode;
5340 if (original_regno < FIRST_PSEUDO_REGISTER)
5342 rclass = ira_allocno_class_translate[REGNO_REG_CLASS (original_regno)];
5343 hard_regno = original_regno;
5344 call_save_p = false;
5345 nregs = 1;
5346 mode = lra_reg_info[hard_regno].biggest_mode;
5347 machine_mode reg_rtx_mode = GET_MODE (regno_reg_rtx[hard_regno]);
5348 /* A reg can have a biggest_mode of VOIDmode if it was only ever seen
5349 as part of a multi-word register. In that case, or if the biggest
5350 mode was larger than a register, just use the reg_rtx. Otherwise,
5351 limit the size to that of the biggest access in the function. */
5352 if (mode == VOIDmode
5353 || GET_MODE_SIZE (mode) > GET_MODE_SIZE (reg_rtx_mode))
5355 original_reg = regno_reg_rtx[hard_regno];
5356 mode = reg_rtx_mode;
5358 else
5359 original_reg = gen_rtx_REG (mode, hard_regno);
5361 else
5363 mode = PSEUDO_REGNO_MODE (original_regno);
5364 hard_regno = reg_renumber[original_regno];
5365 nregs = hard_regno_nregs[hard_regno][mode];
5366 rclass = lra_get_allocno_class (original_regno);
5367 original_reg = regno_reg_rtx[original_regno];
5368 call_save_p = need_for_call_save_p (original_regno);
5370 lra_assert (hard_regno >= 0);
5371 if (lra_dump_file != NULL)
5372 fprintf (lra_dump_file,
5373 " ((((((((((((((((((((((((((((((((((((((((((((((((\n");
5375 if (call_save_p)
5377 mode = HARD_REGNO_CALLER_SAVE_MODE (hard_regno,
5378 hard_regno_nregs[hard_regno][mode],
5379 mode);
5380 new_reg = lra_create_new_reg (mode, NULL_RTX, NO_REGS, "save");
5382 else
5384 rclass = choose_split_class (rclass, hard_regno, mode);
5385 if (rclass == NO_REGS)
5387 if (lra_dump_file != NULL)
5389 fprintf (lra_dump_file,
5390 " Rejecting split of %d(%s): "
5391 "no good reg class for %d(%s)\n",
5392 original_regno,
5393 reg_class_names[lra_get_allocno_class (original_regno)],
5394 hard_regno,
5395 reg_class_names[REGNO_REG_CLASS (hard_regno)]);
5396 fprintf
5397 (lra_dump_file,
5398 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5400 return false;
5402 /* Split_if_necessary can split hard registers used as part of a
5403 multi-register mode but splits each register individually. The
5404 mode used for each independent register may not be supported
5405 so reject the split. Splitting the wider mode should theoretically
5406 be possible but is not implemented. */
5407 if (! HARD_REGNO_MODE_OK (hard_regno, mode))
5409 if (lra_dump_file != NULL)
5411 fprintf (lra_dump_file,
5412 " Rejecting split of %d(%s): unsuitable mode %s\n",
5413 original_regno,
5414 reg_class_names[lra_get_allocno_class (original_regno)],
5415 GET_MODE_NAME (mode));
5416 fprintf
5417 (lra_dump_file,
5418 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5420 return false;
5422 new_reg = lra_create_new_reg (mode, original_reg, rclass, "split");
5423 reg_renumber[REGNO (new_reg)] = hard_regno;
5425 save = emit_spill_move (true, new_reg, original_reg);
5426 if (NEXT_INSN (save) != NULL_RTX && !call_save_p)
5428 if (lra_dump_file != NULL)
5430 fprintf
5431 (lra_dump_file,
5432 " Rejecting split %d->%d resulting in > 2 save insns:\n",
5433 original_regno, REGNO (new_reg));
5434 dump_rtl_slim (lra_dump_file, save, NULL, -1, 0);
5435 fprintf (lra_dump_file,
5436 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5438 return false;
5440 restore = emit_spill_move (false, new_reg, original_reg);
5441 if (NEXT_INSN (restore) != NULL_RTX && !call_save_p)
5443 if (lra_dump_file != NULL)
5445 fprintf (lra_dump_file,
5446 " Rejecting split %d->%d "
5447 "resulting in > 2 restore insns:\n",
5448 original_regno, REGNO (new_reg));
5449 dump_rtl_slim (lra_dump_file, restore, NULL, -1, 0);
5450 fprintf (lra_dump_file,
5451 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5453 return false;
5455 after_p = usage_insns[original_regno].after_p;
5456 lra_reg_info[REGNO (new_reg)].restore_rtx = regno_reg_rtx[original_regno];
5457 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
5458 bitmap_set_bit (&check_only_regs, original_regno);
5459 bitmap_set_bit (&lra_split_regs, REGNO (new_reg));
5460 for (;;)
5462 if (GET_CODE (next_usage_insns) != INSN_LIST)
5464 usage_insn = next_usage_insns;
5465 break;
5467 usage_insn = XEXP (next_usage_insns, 0);
5468 lra_assert (DEBUG_INSN_P (usage_insn));
5469 next_usage_insns = XEXP (next_usage_insns, 1);
5470 lra_substitute_pseudo (&usage_insn, original_regno, new_reg, false);
5471 lra_update_insn_regno_info (as_a <rtx_insn *> (usage_insn));
5472 if (lra_dump_file != NULL)
5474 fprintf (lra_dump_file, " Split reuse change %d->%d:\n",
5475 original_regno, REGNO (new_reg));
5476 dump_insn_slim (lra_dump_file, as_a <rtx_insn *> (usage_insn));
5479 lra_assert (NOTE_P (usage_insn) || NONDEBUG_INSN_P (usage_insn));
5480 lra_assert (usage_insn != insn || (after_p && before_p));
5481 lra_process_new_insns (as_a <rtx_insn *> (usage_insn),
5482 after_p ? NULL : restore,
5483 after_p ? restore : NULL,
5484 call_save_p
5485 ? "Add reg<-save" : "Add reg<-split");
5486 lra_process_new_insns (insn, before_p ? save : NULL,
5487 before_p ? NULL : save,
5488 call_save_p
5489 ? "Add save<-reg" : "Add split<-reg");
5490 if (nregs > 1)
5491 /* If we are trying to split multi-register. We should check
5492 conflicts on the next assignment sub-pass. IRA can allocate on
5493 sub-register levels, LRA do this on pseudos level right now and
5494 this discrepancy may create allocation conflicts after
5495 splitting. */
5496 lra_risky_transformations_p = true;
5497 if (lra_dump_file != NULL)
5498 fprintf (lra_dump_file,
5499 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5500 return true;
5503 /* Recognize that we need a split transformation for insn INSN, which
5504 defines or uses REGNO in its insn biggest MODE (we use it only if
5505 REGNO is a hard register). POTENTIAL_RELOAD_HARD_REGS contains
5506 hard registers which might be used for reloads since the EBB end.
5507 Put the save before INSN if BEFORE_P is true. MAX_UID is maximla
5508 uid before starting INSN processing. Return true if we succeed in
5509 such transformation. */
5510 static bool
5511 split_if_necessary (int regno, machine_mode mode,
5512 HARD_REG_SET potential_reload_hard_regs,
5513 bool before_p, rtx_insn *insn, int max_uid)
5515 bool res = false;
5516 int i, nregs = 1;
5517 rtx next_usage_insns;
5519 if (regno < FIRST_PSEUDO_REGISTER)
5520 nregs = hard_regno_nregs[regno][mode];
5521 for (i = 0; i < nregs; i++)
5522 if (usage_insns[regno + i].check == curr_usage_insns_check
5523 && (next_usage_insns = usage_insns[regno + i].insns) != NULL_RTX
5524 /* To avoid processing the register twice or more. */
5525 && ((GET_CODE (next_usage_insns) != INSN_LIST
5526 && INSN_UID (next_usage_insns) < max_uid)
5527 || (GET_CODE (next_usage_insns) == INSN_LIST
5528 && (INSN_UID (XEXP (next_usage_insns, 0)) < max_uid)))
5529 && need_for_split_p (potential_reload_hard_regs, regno + i)
5530 && split_reg (before_p, regno + i, insn, next_usage_insns))
5531 res = true;
5532 return res;
5535 /* Return TRUE if rtx X is considered as an invariant for
5536 inheritance. */
5537 static bool
5538 invariant_p (const_rtx x)
5540 machine_mode mode;
5541 const char *fmt;
5542 enum rtx_code code;
5543 int i, j;
5545 code = GET_CODE (x);
5546 mode = GET_MODE (x);
5547 if (code == SUBREG)
5549 x = SUBREG_REG (x);
5550 code = GET_CODE (x);
5551 if (GET_MODE_SIZE (GET_MODE (x)) > GET_MODE_SIZE (mode))
5552 mode = GET_MODE (x);
5555 if (MEM_P (x))
5556 return false;
5558 if (REG_P (x))
5560 int i, nregs, regno = REGNO (x);
5562 if (regno >= FIRST_PSEUDO_REGISTER || regno == STACK_POINTER_REGNUM
5563 || TEST_HARD_REG_BIT (eliminable_regset, regno)
5564 || GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
5565 return false;
5566 nregs = hard_regno_nregs[regno][mode];
5567 for (i = 0; i < nregs; i++)
5568 if (! fixed_regs[regno + i]
5569 /* A hard register may be clobbered in the current insn
5570 but we can ignore this case because if the hard
5571 register is used it should be set somewhere after the
5572 clobber. */
5573 || bitmap_bit_p (&invalid_invariant_regs, regno + i))
5574 return false;
5576 fmt = GET_RTX_FORMAT (code);
5577 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
5579 if (fmt[i] == 'e')
5581 if (! invariant_p (XEXP (x, i)))
5582 return false;
5584 else if (fmt[i] == 'E')
5586 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
5587 if (! invariant_p (XVECEXP (x, i, j)))
5588 return false;
5591 return true;
5594 /* We have 'dest_reg <- invariant'. Let us try to make an invariant
5595 inheritance transformation (using dest_reg instead invariant in a
5596 subsequent insn). */
5597 static bool
5598 process_invariant_for_inheritance (rtx dst_reg, rtx invariant_rtx)
5600 invariant_ptr_t invariant_ptr;
5601 rtx_insn *insn, *new_insns;
5602 rtx insn_set, insn_reg, new_reg;
5603 int insn_regno;
5604 bool succ_p = false;
5605 int dst_regno = REGNO (dst_reg);
5606 enum machine_mode dst_mode = GET_MODE (dst_reg);
5607 enum reg_class cl = lra_get_allocno_class (dst_regno), insn_reg_cl;
5609 invariant_ptr = insert_invariant (invariant_rtx);
5610 if ((insn = invariant_ptr->insn) != NULL_RTX)
5612 /* We have a subsequent insn using the invariant. */
5613 insn_set = single_set (insn);
5614 lra_assert (insn_set != NULL);
5615 insn_reg = SET_DEST (insn_set);
5616 lra_assert (REG_P (insn_reg));
5617 insn_regno = REGNO (insn_reg);
5618 insn_reg_cl = lra_get_allocno_class (insn_regno);
5620 if (dst_mode == GET_MODE (insn_reg)
5621 /* We should consider only result move reg insns which are
5622 cheap. */
5623 && targetm.register_move_cost (dst_mode, cl, insn_reg_cl) == 2
5624 && targetm.register_move_cost (dst_mode, cl, cl) == 2)
5626 if (lra_dump_file != NULL)
5627 fprintf (lra_dump_file,
5628 " [[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[\n");
5629 new_reg = lra_create_new_reg (dst_mode, dst_reg,
5630 cl, "invariant inheritance");
5631 bitmap_set_bit (&lra_inheritance_pseudos, REGNO (new_reg));
5632 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
5633 lra_reg_info[REGNO (new_reg)].restore_rtx = PATTERN (insn);
5634 start_sequence ();
5635 lra_emit_move (new_reg, dst_reg);
5636 new_insns = get_insns ();
5637 end_sequence ();
5638 lra_process_new_insns (curr_insn, NULL, new_insns,
5639 "Add invariant inheritance<-original");
5640 start_sequence ();
5641 lra_emit_move (SET_DEST (insn_set), new_reg);
5642 new_insns = get_insns ();
5643 end_sequence ();
5644 lra_process_new_insns (insn, NULL, new_insns,
5645 "Changing reload<-inheritance");
5646 lra_set_insn_deleted (insn);
5647 succ_p = true;
5648 if (lra_dump_file != NULL)
5650 fprintf (lra_dump_file,
5651 " Invariant inheritance reuse change %d (bb%d):\n",
5652 REGNO (new_reg), BLOCK_FOR_INSN (insn)->index);
5653 dump_insn_slim (lra_dump_file, insn);
5654 fprintf (lra_dump_file,
5655 " ]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]\n");
5659 invariant_ptr->insn = curr_insn;
5660 return succ_p;
5663 /* Check only registers living at the current program point in the
5664 current EBB. */
5665 static bitmap_head live_regs;
5667 /* Update live info in EBB given by its HEAD and TAIL insns after
5668 inheritance/split transformation. The function removes dead moves
5669 too. */
5670 static void
5671 update_ebb_live_info (rtx_insn *head, rtx_insn *tail)
5673 unsigned int j;
5674 int i, regno;
5675 bool live_p;
5676 rtx_insn *prev_insn;
5677 rtx set;
5678 bool remove_p;
5679 basic_block last_bb, prev_bb, curr_bb;
5680 bitmap_iterator bi;
5681 struct lra_insn_reg *reg;
5682 edge e;
5683 edge_iterator ei;
5685 last_bb = BLOCK_FOR_INSN (tail);
5686 prev_bb = NULL;
5687 for (curr_insn = tail;
5688 curr_insn != PREV_INSN (head);
5689 curr_insn = prev_insn)
5691 prev_insn = PREV_INSN (curr_insn);
5692 /* We need to process empty blocks too. They contain
5693 NOTE_INSN_BASIC_BLOCK referring for the basic block. */
5694 if (NOTE_P (curr_insn) && NOTE_KIND (curr_insn) != NOTE_INSN_BASIC_BLOCK)
5695 continue;
5696 curr_bb = BLOCK_FOR_INSN (curr_insn);
5697 if (curr_bb != prev_bb)
5699 if (prev_bb != NULL)
5701 /* Update df_get_live_in (prev_bb): */
5702 EXECUTE_IF_SET_IN_BITMAP (&check_only_regs, 0, j, bi)
5703 if (bitmap_bit_p (&live_regs, j))
5704 bitmap_set_bit (df_get_live_in (prev_bb), j);
5705 else
5706 bitmap_clear_bit (df_get_live_in (prev_bb), j);
5708 if (curr_bb != last_bb)
5710 /* Update df_get_live_out (curr_bb): */
5711 EXECUTE_IF_SET_IN_BITMAP (&check_only_regs, 0, j, bi)
5713 live_p = bitmap_bit_p (&live_regs, j);
5714 if (! live_p)
5715 FOR_EACH_EDGE (e, ei, curr_bb->succs)
5716 if (bitmap_bit_p (df_get_live_in (e->dest), j))
5718 live_p = true;
5719 break;
5721 if (live_p)
5722 bitmap_set_bit (df_get_live_out (curr_bb), j);
5723 else
5724 bitmap_clear_bit (df_get_live_out (curr_bb), j);
5727 prev_bb = curr_bb;
5728 bitmap_and (&live_regs, &check_only_regs, df_get_live_out (curr_bb));
5730 if (! NONDEBUG_INSN_P (curr_insn))
5731 continue;
5732 curr_id = lra_get_insn_recog_data (curr_insn);
5733 curr_static_id = curr_id->insn_static_data;
5734 remove_p = false;
5735 if ((set = single_set (curr_insn)) != NULL_RTX
5736 && REG_P (SET_DEST (set))
5737 && (regno = REGNO (SET_DEST (set))) >= FIRST_PSEUDO_REGISTER
5738 && SET_DEST (set) != pic_offset_table_rtx
5739 && bitmap_bit_p (&check_only_regs, regno)
5740 && ! bitmap_bit_p (&live_regs, regno))
5741 remove_p = true;
5742 /* See which defined values die here. */
5743 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5744 if (reg->type == OP_OUT && ! reg->subreg_p)
5745 bitmap_clear_bit (&live_regs, reg->regno);
5746 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
5747 if (reg->type == OP_OUT && ! reg->subreg_p)
5748 bitmap_clear_bit (&live_regs, reg->regno);
5749 if (curr_id->arg_hard_regs != NULL)
5750 /* Make clobbered argument hard registers die. */
5751 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
5752 if (regno >= FIRST_PSEUDO_REGISTER)
5753 bitmap_clear_bit (&live_regs, regno - FIRST_PSEUDO_REGISTER);
5754 /* Mark each used value as live. */
5755 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5756 if (reg->type != OP_OUT
5757 && bitmap_bit_p (&check_only_regs, reg->regno))
5758 bitmap_set_bit (&live_regs, reg->regno);
5759 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
5760 if (reg->type != OP_OUT
5761 && bitmap_bit_p (&check_only_regs, reg->regno))
5762 bitmap_set_bit (&live_regs, reg->regno);
5763 if (curr_id->arg_hard_regs != NULL)
5764 /* Make used argument hard registers live. */
5765 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
5766 if (regno < FIRST_PSEUDO_REGISTER
5767 && bitmap_bit_p (&check_only_regs, regno))
5768 bitmap_set_bit (&live_regs, regno);
5769 /* It is quite important to remove dead move insns because it
5770 means removing dead store. We don't need to process them for
5771 constraints. */
5772 if (remove_p)
5774 if (lra_dump_file != NULL)
5776 fprintf (lra_dump_file, " Removing dead insn:\n ");
5777 dump_insn_slim (lra_dump_file, curr_insn);
5779 lra_set_insn_deleted (curr_insn);
5784 /* The structure describes info to do an inheritance for the current
5785 insn. We need to collect such info first before doing the
5786 transformations because the transformations change the insn
5787 internal representation. */
5788 struct to_inherit
5790 /* Original regno. */
5791 int regno;
5792 /* Subsequent insns which can inherit original reg value. */
5793 rtx insns;
5796 /* Array containing all info for doing inheritance from the current
5797 insn. */
5798 static struct to_inherit to_inherit[LRA_MAX_INSN_RELOADS];
5800 /* Number elements in the previous array. */
5801 static int to_inherit_num;
5803 /* Add inheritance info REGNO and INSNS. Their meaning is described in
5804 structure to_inherit. */
5805 static void
5806 add_to_inherit (int regno, rtx insns)
5808 int i;
5810 for (i = 0; i < to_inherit_num; i++)
5811 if (to_inherit[i].regno == regno)
5812 return;
5813 lra_assert (to_inherit_num < LRA_MAX_INSN_RELOADS);
5814 to_inherit[to_inherit_num].regno = regno;
5815 to_inherit[to_inherit_num++].insns = insns;
5818 /* Return the last non-debug insn in basic block BB, or the block begin
5819 note if none. */
5820 static rtx_insn *
5821 get_last_insertion_point (basic_block bb)
5823 rtx_insn *insn;
5825 FOR_BB_INSNS_REVERSE (bb, insn)
5826 if (NONDEBUG_INSN_P (insn) || NOTE_INSN_BASIC_BLOCK_P (insn))
5827 return insn;
5828 gcc_unreachable ();
5831 /* Set up RES by registers living on edges FROM except the edge (FROM,
5832 TO) or by registers set up in a jump insn in BB FROM. */
5833 static void
5834 get_live_on_other_edges (basic_block from, basic_block to, bitmap res)
5836 rtx_insn *last;
5837 struct lra_insn_reg *reg;
5838 edge e;
5839 edge_iterator ei;
5841 lra_assert (to != NULL);
5842 bitmap_clear (res);
5843 FOR_EACH_EDGE (e, ei, from->succs)
5844 if (e->dest != to)
5845 bitmap_ior_into (res, df_get_live_in (e->dest));
5846 last = get_last_insertion_point (from);
5847 if (! JUMP_P (last))
5848 return;
5849 curr_id = lra_get_insn_recog_data (last);
5850 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5851 if (reg->type != OP_IN)
5852 bitmap_set_bit (res, reg->regno);
5855 /* Used as a temporary results of some bitmap calculations. */
5856 static bitmap_head temp_bitmap;
5858 /* We split for reloads of small class of hard regs. The following
5859 defines how many hard regs the class should have to be qualified as
5860 small. The code is mostly oriented to x86/x86-64 architecture
5861 where some insns need to use only specific register or pair of
5862 registers and these register can live in RTL explicitly, e.g. for
5863 parameter passing. */
5864 static const int max_small_class_regs_num = 2;
5866 /* Do inheritance/split transformations in EBB starting with HEAD and
5867 finishing on TAIL. We process EBB insns in the reverse order.
5868 Return true if we did any inheritance/split transformation in the
5869 EBB.
5871 We should avoid excessive splitting which results in worse code
5872 because of inaccurate cost calculations for spilling new split
5873 pseudos in such case. To achieve this we do splitting only if
5874 register pressure is high in given basic block and there are reload
5875 pseudos requiring hard registers. We could do more register
5876 pressure calculations at any given program point to avoid necessary
5877 splitting even more but it is to expensive and the current approach
5878 works well enough. */
5879 static bool
5880 inherit_in_ebb (rtx_insn *head, rtx_insn *tail)
5882 int i, src_regno, dst_regno, nregs;
5883 bool change_p, succ_p, update_reloads_num_p;
5884 rtx_insn *prev_insn, *last_insn;
5885 rtx next_usage_insns, curr_set;
5886 enum reg_class cl;
5887 struct lra_insn_reg *reg;
5888 basic_block last_processed_bb, curr_bb = NULL;
5889 HARD_REG_SET potential_reload_hard_regs, live_hard_regs;
5890 bitmap to_process;
5891 unsigned int j;
5892 bitmap_iterator bi;
5893 bool head_p, after_p;
5895 change_p = false;
5896 curr_usage_insns_check++;
5897 clear_invariants ();
5898 reloads_num = calls_num = 0;
5899 bitmap_clear (&check_only_regs);
5900 bitmap_clear (&invalid_invariant_regs);
5901 last_processed_bb = NULL;
5902 CLEAR_HARD_REG_SET (potential_reload_hard_regs);
5903 COPY_HARD_REG_SET (live_hard_regs, eliminable_regset);
5904 IOR_HARD_REG_SET (live_hard_regs, lra_no_alloc_regs);
5905 /* We don't process new insns generated in the loop. */
5906 for (curr_insn = tail; curr_insn != PREV_INSN (head); curr_insn = prev_insn)
5908 prev_insn = PREV_INSN (curr_insn);
5909 if (BLOCK_FOR_INSN (curr_insn) != NULL)
5910 curr_bb = BLOCK_FOR_INSN (curr_insn);
5911 if (last_processed_bb != curr_bb)
5913 /* We are at the end of BB. Add qualified living
5914 pseudos for potential splitting. */
5915 to_process = df_get_live_out (curr_bb);
5916 if (last_processed_bb != NULL)
5918 /* We are somewhere in the middle of EBB. */
5919 get_live_on_other_edges (curr_bb, last_processed_bb,
5920 &temp_bitmap);
5921 to_process = &temp_bitmap;
5923 last_processed_bb = curr_bb;
5924 last_insn = get_last_insertion_point (curr_bb);
5925 after_p = (! JUMP_P (last_insn)
5926 && (! CALL_P (last_insn)
5927 || (find_reg_note (last_insn,
5928 REG_NORETURN, NULL_RTX) == NULL_RTX
5929 && ! SIBLING_CALL_P (last_insn))));
5930 CLEAR_HARD_REG_SET (potential_reload_hard_regs);
5931 EXECUTE_IF_SET_IN_BITMAP (to_process, 0, j, bi)
5933 if ((int) j >= lra_constraint_new_regno_start)
5934 break;
5935 if (j < FIRST_PSEUDO_REGISTER || reg_renumber[j] >= 0)
5937 if (j < FIRST_PSEUDO_REGISTER)
5938 SET_HARD_REG_BIT (live_hard_regs, j);
5939 else
5940 add_to_hard_reg_set (&live_hard_regs,
5941 PSEUDO_REGNO_MODE (j),
5942 reg_renumber[j]);
5943 setup_next_usage_insn (j, last_insn, reloads_num, after_p);
5947 src_regno = dst_regno = -1;
5948 curr_set = single_set (curr_insn);
5949 if (curr_set != NULL_RTX && REG_P (SET_DEST (curr_set)))
5950 dst_regno = REGNO (SET_DEST (curr_set));
5951 if (curr_set != NULL_RTX && REG_P (SET_SRC (curr_set)))
5952 src_regno = REGNO (SET_SRC (curr_set));
5953 update_reloads_num_p = true;
5954 if (src_regno < lra_constraint_new_regno_start
5955 && src_regno >= FIRST_PSEUDO_REGISTER
5956 && reg_renumber[src_regno] < 0
5957 && dst_regno >= lra_constraint_new_regno_start
5958 && (cl = lra_get_allocno_class (dst_regno)) != NO_REGS)
5960 /* 'reload_pseudo <- original_pseudo'. */
5961 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
5962 reloads_num++;
5963 update_reloads_num_p = false;
5964 succ_p = false;
5965 if (usage_insns[src_regno].check == curr_usage_insns_check
5966 && (next_usage_insns = usage_insns[src_regno].insns) != NULL_RTX)
5967 succ_p = inherit_reload_reg (false, src_regno, cl,
5968 curr_insn, next_usage_insns);
5969 if (succ_p)
5970 change_p = true;
5971 else
5972 setup_next_usage_insn (src_regno, curr_insn, reloads_num, false);
5973 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
5974 IOR_HARD_REG_SET (potential_reload_hard_regs,
5975 reg_class_contents[cl]);
5977 else if (src_regno < 0
5978 && dst_regno >= lra_constraint_new_regno_start
5979 && invariant_p (SET_SRC (curr_set))
5980 && (cl = lra_get_allocno_class (dst_regno)) != NO_REGS
5981 && ! bitmap_bit_p (&invalid_invariant_regs, dst_regno)
5982 && ! bitmap_bit_p (&invalid_invariant_regs,
5983 ORIGINAL_REGNO(regno_reg_rtx[dst_regno])))
5985 /* 'reload_pseudo <- invariant'. */
5986 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
5987 reloads_num++;
5988 update_reloads_num_p = false;
5989 if (process_invariant_for_inheritance (SET_DEST (curr_set), SET_SRC (curr_set)))
5990 change_p = true;
5991 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
5992 IOR_HARD_REG_SET (potential_reload_hard_regs,
5993 reg_class_contents[cl]);
5995 else if (src_regno >= lra_constraint_new_regno_start
5996 && dst_regno < lra_constraint_new_regno_start
5997 && dst_regno >= FIRST_PSEUDO_REGISTER
5998 && reg_renumber[dst_regno] < 0
5999 && (cl = lra_get_allocno_class (src_regno)) != NO_REGS
6000 && usage_insns[dst_regno].check == curr_usage_insns_check
6001 && (next_usage_insns
6002 = usage_insns[dst_regno].insns) != NULL_RTX)
6004 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6005 reloads_num++;
6006 update_reloads_num_p = false;
6007 /* 'original_pseudo <- reload_pseudo'. */
6008 if (! JUMP_P (curr_insn)
6009 && inherit_reload_reg (true, dst_regno, cl,
6010 curr_insn, next_usage_insns))
6011 change_p = true;
6012 /* Invalidate. */
6013 usage_insns[dst_regno].check = 0;
6014 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6015 IOR_HARD_REG_SET (potential_reload_hard_regs,
6016 reg_class_contents[cl]);
6018 else if (INSN_P (curr_insn))
6020 int iter;
6021 int max_uid = get_max_uid ();
6023 curr_id = lra_get_insn_recog_data (curr_insn);
6024 curr_static_id = curr_id->insn_static_data;
6025 to_inherit_num = 0;
6026 /* Process insn definitions. */
6027 for (iter = 0; iter < 2; iter++)
6028 for (reg = iter == 0 ? curr_id->regs : curr_static_id->hard_regs;
6029 reg != NULL;
6030 reg = reg->next)
6031 if (reg->type != OP_IN
6032 && (dst_regno = reg->regno) < lra_constraint_new_regno_start)
6034 if (dst_regno >= FIRST_PSEUDO_REGISTER && reg->type == OP_OUT
6035 && reg_renumber[dst_regno] < 0 && ! reg->subreg_p
6036 && usage_insns[dst_regno].check == curr_usage_insns_check
6037 && (next_usage_insns
6038 = usage_insns[dst_regno].insns) != NULL_RTX)
6040 struct lra_insn_reg *r;
6042 for (r = curr_id->regs; r != NULL; r = r->next)
6043 if (r->type != OP_OUT && r->regno == dst_regno)
6044 break;
6045 /* Don't do inheritance if the pseudo is also
6046 used in the insn. */
6047 if (r == NULL)
6048 /* We can not do inheritance right now
6049 because the current insn reg info (chain
6050 regs) can change after that. */
6051 add_to_inherit (dst_regno, next_usage_insns);
6053 /* We can not process one reg twice here because of
6054 usage_insns invalidation. */
6055 if ((dst_regno < FIRST_PSEUDO_REGISTER
6056 || reg_renumber[dst_regno] >= 0)
6057 && ! reg->subreg_p && reg->type != OP_IN)
6059 HARD_REG_SET s;
6061 if (split_if_necessary (dst_regno, reg->biggest_mode,
6062 potential_reload_hard_regs,
6063 false, curr_insn, max_uid))
6064 change_p = true;
6065 CLEAR_HARD_REG_SET (s);
6066 if (dst_regno < FIRST_PSEUDO_REGISTER)
6067 add_to_hard_reg_set (&s, reg->biggest_mode, dst_regno);
6068 else
6069 add_to_hard_reg_set (&s, PSEUDO_REGNO_MODE (dst_regno),
6070 reg_renumber[dst_regno]);
6071 AND_COMPL_HARD_REG_SET (live_hard_regs, s);
6073 /* We should invalidate potential inheritance or
6074 splitting for the current insn usages to the next
6075 usage insns (see code below) as the output pseudo
6076 prevents this. */
6077 if ((dst_regno >= FIRST_PSEUDO_REGISTER
6078 && reg_renumber[dst_regno] < 0)
6079 || (reg->type == OP_OUT && ! reg->subreg_p
6080 && (dst_regno < FIRST_PSEUDO_REGISTER
6081 || reg_renumber[dst_regno] >= 0)))
6083 /* Invalidate and mark definitions. */
6084 if (dst_regno >= FIRST_PSEUDO_REGISTER)
6085 usage_insns[dst_regno].check = -(int) INSN_UID (curr_insn);
6086 else
6088 nregs = hard_regno_nregs[dst_regno][reg->biggest_mode];
6089 for (i = 0; i < nregs; i++)
6090 usage_insns[dst_regno + i].check
6091 = -(int) INSN_UID (curr_insn);
6095 /* Process clobbered call regs. */
6096 if (curr_id->arg_hard_regs != NULL)
6097 for (i = 0; (dst_regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6098 if (dst_regno >= FIRST_PSEUDO_REGISTER)
6099 usage_insns[dst_regno - FIRST_PSEUDO_REGISTER].check
6100 = -(int) INSN_UID (curr_insn);
6101 if (! JUMP_P (curr_insn))
6102 for (i = 0; i < to_inherit_num; i++)
6103 if (inherit_reload_reg (true, to_inherit[i].regno,
6104 ALL_REGS, curr_insn,
6105 to_inherit[i].insns))
6106 change_p = true;
6107 if (CALL_P (curr_insn))
6109 rtx cheap, pat, dest;
6110 rtx_insn *restore;
6111 int regno, hard_regno;
6113 calls_num++;
6114 if ((cheap = find_reg_note (curr_insn,
6115 REG_RETURNED, NULL_RTX)) != NULL_RTX
6116 && ((cheap = XEXP (cheap, 0)), true)
6117 && (regno = REGNO (cheap)) >= FIRST_PSEUDO_REGISTER
6118 && (hard_regno = reg_renumber[regno]) >= 0
6119 /* If there are pending saves/restores, the
6120 optimization is not worth. */
6121 && usage_insns[regno].calls_num == calls_num - 1
6122 && TEST_HARD_REG_BIT (call_used_reg_set, hard_regno))
6124 /* Restore the pseudo from the call result as
6125 REG_RETURNED note says that the pseudo value is
6126 in the call result and the pseudo is an argument
6127 of the call. */
6128 pat = PATTERN (curr_insn);
6129 if (GET_CODE (pat) == PARALLEL)
6130 pat = XVECEXP (pat, 0, 0);
6131 dest = SET_DEST (pat);
6132 /* For multiple return values dest is PARALLEL.
6133 Currently we handle only single return value case. */
6134 if (REG_P (dest))
6136 start_sequence ();
6137 emit_move_insn (cheap, copy_rtx (dest));
6138 restore = get_insns ();
6139 end_sequence ();
6140 lra_process_new_insns (curr_insn, NULL, restore,
6141 "Inserting call parameter restore");
6142 /* We don't need to save/restore of the pseudo from
6143 this call. */
6144 usage_insns[regno].calls_num = calls_num;
6145 bitmap_set_bit (&check_only_regs, regno);
6149 to_inherit_num = 0;
6150 /* Process insn usages. */
6151 for (iter = 0; iter < 2; iter++)
6152 for (reg = iter == 0 ? curr_id->regs : curr_static_id->hard_regs;
6153 reg != NULL;
6154 reg = reg->next)
6155 if ((reg->type != OP_OUT
6156 || (reg->type == OP_OUT && reg->subreg_p))
6157 && (src_regno = reg->regno) < lra_constraint_new_regno_start)
6159 if (src_regno >= FIRST_PSEUDO_REGISTER
6160 && reg_renumber[src_regno] < 0 && reg->type == OP_IN)
6162 if (usage_insns[src_regno].check == curr_usage_insns_check
6163 && (next_usage_insns
6164 = usage_insns[src_regno].insns) != NULL_RTX
6165 && NONDEBUG_INSN_P (curr_insn))
6166 add_to_inherit (src_regno, next_usage_insns);
6167 else if (usage_insns[src_regno].check
6168 != -(int) INSN_UID (curr_insn))
6169 /* Add usages but only if the reg is not set up
6170 in the same insn. */
6171 add_next_usage_insn (src_regno, curr_insn, reloads_num);
6173 else if (src_regno < FIRST_PSEUDO_REGISTER
6174 || reg_renumber[src_regno] >= 0)
6176 bool before_p;
6177 rtx_insn *use_insn = curr_insn;
6179 before_p = (JUMP_P (curr_insn)
6180 || (CALL_P (curr_insn) && reg->type == OP_IN));
6181 if (NONDEBUG_INSN_P (curr_insn)
6182 && (! JUMP_P (curr_insn) || reg->type == OP_IN)
6183 && split_if_necessary (src_regno, reg->biggest_mode,
6184 potential_reload_hard_regs,
6185 before_p, curr_insn, max_uid))
6187 if (reg->subreg_p)
6188 lra_risky_transformations_p = true;
6189 change_p = true;
6190 /* Invalidate. */
6191 usage_insns[src_regno].check = 0;
6192 if (before_p)
6193 use_insn = PREV_INSN (curr_insn);
6195 if (NONDEBUG_INSN_P (curr_insn))
6197 if (src_regno < FIRST_PSEUDO_REGISTER)
6198 add_to_hard_reg_set (&live_hard_regs,
6199 reg->biggest_mode, src_regno);
6200 else
6201 add_to_hard_reg_set (&live_hard_regs,
6202 PSEUDO_REGNO_MODE (src_regno),
6203 reg_renumber[src_regno]);
6205 add_next_usage_insn (src_regno, use_insn, reloads_num);
6208 /* Process used call regs. */
6209 if (curr_id->arg_hard_regs != NULL)
6210 for (i = 0; (src_regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6211 if (src_regno < FIRST_PSEUDO_REGISTER)
6213 SET_HARD_REG_BIT (live_hard_regs, src_regno);
6214 add_next_usage_insn (src_regno, curr_insn, reloads_num);
6216 for (i = 0; i < to_inherit_num; i++)
6218 src_regno = to_inherit[i].regno;
6219 if (inherit_reload_reg (false, src_regno, ALL_REGS,
6220 curr_insn, to_inherit[i].insns))
6221 change_p = true;
6222 else
6223 setup_next_usage_insn (src_regno, curr_insn, reloads_num, false);
6226 if (update_reloads_num_p
6227 && NONDEBUG_INSN_P (curr_insn) && curr_set != NULL_RTX)
6229 int regno = -1;
6230 if ((REG_P (SET_DEST (curr_set))
6231 && (regno = REGNO (SET_DEST (curr_set))) >= lra_constraint_new_regno_start
6232 && reg_renumber[regno] < 0
6233 && (cl = lra_get_allocno_class (regno)) != NO_REGS)
6234 || (REG_P (SET_SRC (curr_set))
6235 && (regno = REGNO (SET_SRC (curr_set))) >= lra_constraint_new_regno_start
6236 && reg_renumber[regno] < 0
6237 && (cl = lra_get_allocno_class (regno)) != NO_REGS))
6239 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6240 reloads_num++;
6241 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6242 IOR_HARD_REG_SET (potential_reload_hard_regs,
6243 reg_class_contents[cl]);
6246 if (NONDEBUG_INSN_P (curr_insn))
6248 int regno;
6250 /* Invalidate invariants with changed regs. */
6251 curr_id = lra_get_insn_recog_data (curr_insn);
6252 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
6253 if (reg->type != OP_IN)
6255 bitmap_set_bit (&invalid_invariant_regs, reg->regno);
6256 bitmap_set_bit (&invalid_invariant_regs,
6257 ORIGINAL_REGNO (regno_reg_rtx[reg->regno]));
6259 curr_static_id = curr_id->insn_static_data;
6260 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
6261 if (reg->type != OP_IN)
6262 bitmap_set_bit (&invalid_invariant_regs, reg->regno);
6263 if (curr_id->arg_hard_regs != NULL)
6264 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6265 if (regno >= FIRST_PSEUDO_REGISTER)
6266 bitmap_set_bit (&invalid_invariant_regs,
6267 regno - FIRST_PSEUDO_REGISTER);
6269 /* We reached the start of the current basic block. */
6270 if (prev_insn == NULL_RTX || prev_insn == PREV_INSN (head)
6271 || BLOCK_FOR_INSN (prev_insn) != curr_bb)
6273 /* We reached the beginning of the current block -- do
6274 rest of spliting in the current BB. */
6275 to_process = df_get_live_in (curr_bb);
6276 if (BLOCK_FOR_INSN (head) != curr_bb)
6278 /* We are somewhere in the middle of EBB. */
6279 get_live_on_other_edges (EDGE_PRED (curr_bb, 0)->src,
6280 curr_bb, &temp_bitmap);
6281 to_process = &temp_bitmap;
6283 head_p = true;
6284 EXECUTE_IF_SET_IN_BITMAP (to_process, 0, j, bi)
6286 if ((int) j >= lra_constraint_new_regno_start)
6287 break;
6288 if (((int) j < FIRST_PSEUDO_REGISTER || reg_renumber[j] >= 0)
6289 && usage_insns[j].check == curr_usage_insns_check
6290 && (next_usage_insns = usage_insns[j].insns) != NULL_RTX)
6292 if (need_for_split_p (potential_reload_hard_regs, j))
6294 if (lra_dump_file != NULL && head_p)
6296 fprintf (lra_dump_file,
6297 " ----------------------------------\n");
6298 head_p = false;
6300 if (split_reg (false, j, bb_note (curr_bb),
6301 next_usage_insns))
6302 change_p = true;
6304 usage_insns[j].check = 0;
6309 return change_p;
6312 /* This value affects EBB forming. If probability of edge from EBB to
6313 a BB is not greater than the following value, we don't add the BB
6314 to EBB. */
6315 #define EBB_PROBABILITY_CUTOFF \
6316 ((REG_BR_PROB_BASE * LRA_INHERITANCE_EBB_PROBABILITY_CUTOFF) / 100)
6318 /* Current number of inheritance/split iteration. */
6319 int lra_inheritance_iter;
6321 /* Entry function for inheritance/split pass. */
6322 void
6323 lra_inheritance (void)
6325 int i;
6326 basic_block bb, start_bb;
6327 edge e;
6329 lra_inheritance_iter++;
6330 if (lra_inheritance_iter > LRA_MAX_INHERITANCE_PASSES)
6331 return;
6332 timevar_push (TV_LRA_INHERITANCE);
6333 if (lra_dump_file != NULL)
6334 fprintf (lra_dump_file, "\n********** Inheritance #%d: **********\n\n",
6335 lra_inheritance_iter);
6336 curr_usage_insns_check = 0;
6337 usage_insns = XNEWVEC (struct usage_insns, lra_constraint_new_regno_start);
6338 for (i = 0; i < lra_constraint_new_regno_start; i++)
6339 usage_insns[i].check = 0;
6340 bitmap_initialize (&check_only_regs, &reg_obstack);
6341 bitmap_initialize (&invalid_invariant_regs, &reg_obstack);
6342 bitmap_initialize (&live_regs, &reg_obstack);
6343 bitmap_initialize (&temp_bitmap, &reg_obstack);
6344 bitmap_initialize (&ebb_global_regs, &reg_obstack);
6345 FOR_EACH_BB_FN (bb, cfun)
6347 start_bb = bb;
6348 if (lra_dump_file != NULL)
6349 fprintf (lra_dump_file, "EBB");
6350 /* Form a EBB starting with BB. */
6351 bitmap_clear (&ebb_global_regs);
6352 bitmap_ior_into (&ebb_global_regs, df_get_live_in (bb));
6353 for (;;)
6355 if (lra_dump_file != NULL)
6356 fprintf (lra_dump_file, " %d", bb->index);
6357 if (bb->next_bb == EXIT_BLOCK_PTR_FOR_FN (cfun)
6358 || LABEL_P (BB_HEAD (bb->next_bb)))
6359 break;
6360 e = find_fallthru_edge (bb->succs);
6361 if (! e)
6362 break;
6363 if (e->probability < EBB_PROBABILITY_CUTOFF)
6364 break;
6365 bb = bb->next_bb;
6367 bitmap_ior_into (&ebb_global_regs, df_get_live_out (bb));
6368 if (lra_dump_file != NULL)
6369 fprintf (lra_dump_file, "\n");
6370 if (inherit_in_ebb (BB_HEAD (start_bb), BB_END (bb)))
6371 /* Remember that the EBB head and tail can change in
6372 inherit_in_ebb. */
6373 update_ebb_live_info (BB_HEAD (start_bb), BB_END (bb));
6375 bitmap_clear (&ebb_global_regs);
6376 bitmap_clear (&temp_bitmap);
6377 bitmap_clear (&live_regs);
6378 bitmap_clear (&invalid_invariant_regs);
6379 bitmap_clear (&check_only_regs);
6380 free (usage_insns);
6382 timevar_pop (TV_LRA_INHERITANCE);
6387 /* This page contains code to undo failed inheritance/split
6388 transformations. */
6390 /* Current number of iteration undoing inheritance/split. */
6391 int lra_undo_inheritance_iter;
6393 /* Fix BB live info LIVE after removing pseudos created on pass doing
6394 inheritance/split which are REMOVED_PSEUDOS. */
6395 static void
6396 fix_bb_live_info (bitmap live, bitmap removed_pseudos)
6398 unsigned int regno;
6399 bitmap_iterator bi;
6401 EXECUTE_IF_SET_IN_BITMAP (removed_pseudos, 0, regno, bi)
6402 if (bitmap_clear_bit (live, regno)
6403 && REG_P (lra_reg_info[regno].restore_rtx))
6404 bitmap_set_bit (live, REGNO (lra_reg_info[regno].restore_rtx));
6407 /* Return regno of the (subreg of) REG. Otherwise, return a negative
6408 number. */
6409 static int
6410 get_regno (rtx reg)
6412 if (GET_CODE (reg) == SUBREG)
6413 reg = SUBREG_REG (reg);
6414 if (REG_P (reg))
6415 return REGNO (reg);
6416 return -1;
6419 /* Delete a move INSN with destination reg DREGNO and a previous
6420 clobber insn with the same regno. The inheritance/split code can
6421 generate moves with preceding clobber and when we delete such moves
6422 we should delete the clobber insn too to keep the correct life
6423 info. */
6424 static void
6425 delete_move_and_clobber (rtx_insn *insn, int dregno)
6427 rtx_insn *prev_insn = PREV_INSN (insn);
6429 lra_set_insn_deleted (insn);
6430 lra_assert (dregno >= 0);
6431 if (prev_insn != NULL && NONDEBUG_INSN_P (prev_insn)
6432 && GET_CODE (PATTERN (prev_insn)) == CLOBBER
6433 && dregno == get_regno (XEXP (PATTERN (prev_insn), 0)))
6434 lra_set_insn_deleted (prev_insn);
6437 /* Remove inheritance/split pseudos which are in REMOVE_PSEUDOS and
6438 return true if we did any change. The undo transformations for
6439 inheritance looks like
6440 i <- i2
6441 p <- i => p <- i2
6442 or removing
6443 p <- i, i <- p, and i <- i3
6444 where p is original pseudo from which inheritance pseudo i was
6445 created, i and i3 are removed inheritance pseudos, i2 is another
6446 not removed inheritance pseudo. All split pseudos or other
6447 occurrences of removed inheritance pseudos are changed on the
6448 corresponding original pseudos.
6450 The function also schedules insns changed and created during
6451 inheritance/split pass for processing by the subsequent constraint
6452 pass. */
6453 static bool
6454 remove_inheritance_pseudos (bitmap remove_pseudos)
6456 basic_block bb;
6457 int regno, sregno, prev_sregno, dregno;
6458 rtx restore_rtx;
6459 rtx set, prev_set;
6460 rtx_insn *prev_insn;
6461 bool change_p, done_p;
6463 change_p = ! bitmap_empty_p (remove_pseudos);
6464 /* We can not finish the function right away if CHANGE_P is true
6465 because we need to marks insns affected by previous
6466 inheritance/split pass for processing by the subsequent
6467 constraint pass. */
6468 FOR_EACH_BB_FN (bb, cfun)
6470 fix_bb_live_info (df_get_live_in (bb), remove_pseudos);
6471 fix_bb_live_info (df_get_live_out (bb), remove_pseudos);
6472 FOR_BB_INSNS_REVERSE (bb, curr_insn)
6474 if (! INSN_P (curr_insn))
6475 continue;
6476 done_p = false;
6477 sregno = dregno = -1;
6478 if (change_p && NONDEBUG_INSN_P (curr_insn)
6479 && (set = single_set (curr_insn)) != NULL_RTX)
6481 dregno = get_regno (SET_DEST (set));
6482 sregno = get_regno (SET_SRC (set));
6485 if (sregno >= 0 && dregno >= 0)
6487 if (bitmap_bit_p (remove_pseudos, dregno)
6488 && ! REG_P (lra_reg_info[dregno].restore_rtx))
6490 /* invariant inheritance pseudo <- original pseudo */
6491 if (lra_dump_file != NULL)
6493 fprintf (lra_dump_file, " Removing invariant inheritance:\n");
6494 dump_insn_slim (lra_dump_file, curr_insn);
6495 fprintf (lra_dump_file, "\n");
6497 delete_move_and_clobber (curr_insn, dregno);
6498 done_p = true;
6500 else if (bitmap_bit_p (remove_pseudos, sregno)
6501 && ! REG_P (lra_reg_info[sregno].restore_rtx))
6503 /* reload pseudo <- invariant inheritance pseudo */
6504 start_sequence ();
6505 /* We can not just change the source. It might be
6506 an insn different from the move. */
6507 emit_insn (lra_reg_info[sregno].restore_rtx);
6508 rtx_insn *new_insns = get_insns ();
6509 end_sequence ();
6510 lra_assert (single_set (new_insns) != NULL
6511 && SET_DEST (set) == SET_DEST (single_set (new_insns)));
6512 lra_process_new_insns (curr_insn, NULL, new_insns,
6513 "Changing reload<-invariant inheritance");
6514 delete_move_and_clobber (curr_insn, dregno);
6515 done_p = true;
6517 else if ((bitmap_bit_p (remove_pseudos, sregno)
6518 && (get_regno (lra_reg_info[sregno].restore_rtx) == dregno
6519 || (bitmap_bit_p (remove_pseudos, dregno)
6520 && get_regno (lra_reg_info[sregno].restore_rtx) >= 0
6521 && (get_regno (lra_reg_info[sregno].restore_rtx)
6522 == get_regno (lra_reg_info[dregno].restore_rtx)))))
6523 || (bitmap_bit_p (remove_pseudos, dregno)
6524 && get_regno (lra_reg_info[dregno].restore_rtx) == sregno))
6525 /* One of the following cases:
6526 original <- removed inheritance pseudo
6527 removed inherit pseudo <- another removed inherit pseudo
6528 removed inherit pseudo <- original pseudo
6530 removed_split_pseudo <- original_reg
6531 original_reg <- removed_split_pseudo */
6533 if (lra_dump_file != NULL)
6535 fprintf (lra_dump_file, " Removing %s:\n",
6536 bitmap_bit_p (&lra_split_regs, sregno)
6537 || bitmap_bit_p (&lra_split_regs, dregno)
6538 ? "split" : "inheritance");
6539 dump_insn_slim (lra_dump_file, curr_insn);
6541 delete_move_and_clobber (curr_insn, dregno);
6542 done_p = true;
6544 else if (bitmap_bit_p (remove_pseudos, sregno)
6545 && bitmap_bit_p (&lra_inheritance_pseudos, sregno))
6547 /* Search the following pattern:
6548 inherit_or_split_pseudo1 <- inherit_or_split_pseudo2
6549 original_pseudo <- inherit_or_split_pseudo1
6550 where the 2nd insn is the current insn and
6551 inherit_or_split_pseudo2 is not removed. If it is found,
6552 change the current insn onto:
6553 original_pseudo <- inherit_or_split_pseudo2. */
6554 for (prev_insn = PREV_INSN (curr_insn);
6555 prev_insn != NULL_RTX && ! NONDEBUG_INSN_P (prev_insn);
6556 prev_insn = PREV_INSN (prev_insn))
6558 if (prev_insn != NULL_RTX && BLOCK_FOR_INSN (prev_insn) == bb
6559 && (prev_set = single_set (prev_insn)) != NULL_RTX
6560 /* There should be no subregs in insn we are
6561 searching because only the original reg might
6562 be in subreg when we changed the mode of
6563 load/store for splitting. */
6564 && REG_P (SET_DEST (prev_set))
6565 && REG_P (SET_SRC (prev_set))
6566 && (int) REGNO (SET_DEST (prev_set)) == sregno
6567 && ((prev_sregno = REGNO (SET_SRC (prev_set)))
6568 >= FIRST_PSEUDO_REGISTER)
6569 && (lra_reg_info[prev_sregno].restore_rtx == NULL_RTX
6571 /* As we consider chain of inheritance or
6572 splitting described in above comment we should
6573 check that sregno and prev_sregno were
6574 inheritance/split pseudos created from the
6575 same original regno. */
6576 (get_regno (lra_reg_info[sregno].restore_rtx) >= 0
6577 && (get_regno (lra_reg_info[sregno].restore_rtx)
6578 == get_regno (lra_reg_info[prev_sregno].restore_rtx))))
6579 && ! bitmap_bit_p (remove_pseudos, prev_sregno))
6581 lra_assert (GET_MODE (SET_SRC (prev_set))
6582 == GET_MODE (regno_reg_rtx[sregno]));
6583 if (GET_CODE (SET_SRC (set)) == SUBREG)
6584 SUBREG_REG (SET_SRC (set)) = SET_SRC (prev_set);
6585 else
6586 SET_SRC (set) = SET_SRC (prev_set);
6587 /* As we are finishing with processing the insn
6588 here, check the destination too as it might
6589 inheritance pseudo for another pseudo. */
6590 if (bitmap_bit_p (remove_pseudos, dregno)
6591 && bitmap_bit_p (&lra_inheritance_pseudos, dregno)
6592 && (restore_rtx
6593 = lra_reg_info[dregno].restore_rtx) != NULL_RTX)
6595 if (GET_CODE (SET_DEST (set)) == SUBREG)
6596 SUBREG_REG (SET_DEST (set)) = restore_rtx;
6597 else
6598 SET_DEST (set) = restore_rtx;
6600 lra_push_insn_and_update_insn_regno_info (curr_insn);
6601 lra_set_used_insn_alternative_by_uid
6602 (INSN_UID (curr_insn), -1);
6603 done_p = true;
6604 if (lra_dump_file != NULL)
6606 fprintf (lra_dump_file, " Change reload insn:\n");
6607 dump_insn_slim (lra_dump_file, curr_insn);
6612 if (! done_p)
6614 struct lra_insn_reg *reg;
6615 bool restored_regs_p = false;
6616 bool kept_regs_p = false;
6618 curr_id = lra_get_insn_recog_data (curr_insn);
6619 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
6621 regno = reg->regno;
6622 restore_rtx = lra_reg_info[regno].restore_rtx;
6623 if (restore_rtx != NULL_RTX)
6625 if (change_p && bitmap_bit_p (remove_pseudos, regno))
6627 lra_substitute_pseudo_within_insn
6628 (curr_insn, regno, restore_rtx, false);
6629 restored_regs_p = true;
6631 else
6632 kept_regs_p = true;
6635 if (NONDEBUG_INSN_P (curr_insn) && kept_regs_p)
6637 /* The instruction has changed since the previous
6638 constraints pass. */
6639 lra_push_insn_and_update_insn_regno_info (curr_insn);
6640 lra_set_used_insn_alternative_by_uid
6641 (INSN_UID (curr_insn), -1);
6643 else if (restored_regs_p)
6644 /* The instruction has been restored to the form that
6645 it had during the previous constraints pass. */
6646 lra_update_insn_regno_info (curr_insn);
6647 if (restored_regs_p && lra_dump_file != NULL)
6649 fprintf (lra_dump_file, " Insn after restoring regs:\n");
6650 dump_insn_slim (lra_dump_file, curr_insn);
6655 return change_p;
6658 /* If optional reload pseudos failed to get a hard register or was not
6659 inherited, it is better to remove optional reloads. We do this
6660 transformation after undoing inheritance to figure out necessity to
6661 remove optional reloads easier. Return true if we do any
6662 change. */
6663 static bool
6664 undo_optional_reloads (void)
6666 bool change_p, keep_p;
6667 unsigned int regno, uid;
6668 bitmap_iterator bi, bi2;
6669 rtx_insn *insn;
6670 rtx set, src, dest;
6671 bitmap_head removed_optional_reload_pseudos, insn_bitmap;
6673 bitmap_initialize (&removed_optional_reload_pseudos, &reg_obstack);
6674 bitmap_copy (&removed_optional_reload_pseudos, &lra_optional_reload_pseudos);
6675 EXECUTE_IF_SET_IN_BITMAP (&lra_optional_reload_pseudos, 0, regno, bi)
6677 keep_p = false;
6678 /* Keep optional reloads from previous subpasses. */
6679 if (lra_reg_info[regno].restore_rtx == NULL_RTX
6680 /* If the original pseudo changed its allocation, just
6681 removing the optional pseudo is dangerous as the original
6682 pseudo will have longer live range. */
6683 || reg_renumber[REGNO (lra_reg_info[regno].restore_rtx)] >= 0)
6684 keep_p = true;
6685 else if (reg_renumber[regno] >= 0)
6686 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi2)
6688 insn = lra_insn_recog_data[uid]->insn;
6689 if ((set = single_set (insn)) == NULL_RTX)
6690 continue;
6691 src = SET_SRC (set);
6692 dest = SET_DEST (set);
6693 if (! REG_P (src) || ! REG_P (dest))
6694 continue;
6695 if (REGNO (dest) == regno
6696 /* Ignore insn for optional reloads itself. */
6697 && REGNO (lra_reg_info[regno].restore_rtx) != REGNO (src)
6698 /* Check only inheritance on last inheritance pass. */
6699 && (int) REGNO (src) >= new_regno_start
6700 /* Check that the optional reload was inherited. */
6701 && bitmap_bit_p (&lra_inheritance_pseudos, REGNO (src)))
6703 keep_p = true;
6704 break;
6707 if (keep_p)
6709 bitmap_clear_bit (&removed_optional_reload_pseudos, regno);
6710 if (lra_dump_file != NULL)
6711 fprintf (lra_dump_file, "Keep optional reload reg %d\n", regno);
6714 change_p = ! bitmap_empty_p (&removed_optional_reload_pseudos);
6715 bitmap_initialize (&insn_bitmap, &reg_obstack);
6716 EXECUTE_IF_SET_IN_BITMAP (&removed_optional_reload_pseudos, 0, regno, bi)
6718 if (lra_dump_file != NULL)
6719 fprintf (lra_dump_file, "Remove optional reload reg %d\n", regno);
6720 bitmap_copy (&insn_bitmap, &lra_reg_info[regno].insn_bitmap);
6721 EXECUTE_IF_SET_IN_BITMAP (&insn_bitmap, 0, uid, bi2)
6723 insn = lra_insn_recog_data[uid]->insn;
6724 if ((set = single_set (insn)) != NULL_RTX)
6726 src = SET_SRC (set);
6727 dest = SET_DEST (set);
6728 if (REG_P (src) && REG_P (dest)
6729 && ((REGNO (src) == regno
6730 && (REGNO (lra_reg_info[regno].restore_rtx)
6731 == REGNO (dest)))
6732 || (REGNO (dest) == regno
6733 && (REGNO (lra_reg_info[regno].restore_rtx)
6734 == REGNO (src)))))
6736 if (lra_dump_file != NULL)
6738 fprintf (lra_dump_file, " Deleting move %u\n",
6739 INSN_UID (insn));
6740 dump_insn_slim (lra_dump_file, insn);
6742 delete_move_and_clobber (insn, REGNO (dest));
6743 continue;
6745 /* We should not worry about generation memory-memory
6746 moves here as if the corresponding inheritance did
6747 not work (inheritance pseudo did not get a hard reg),
6748 we remove the inheritance pseudo and the optional
6749 reload. */
6751 lra_substitute_pseudo_within_insn
6752 (insn, regno, lra_reg_info[regno].restore_rtx, false);
6753 lra_update_insn_regno_info (insn);
6754 if (lra_dump_file != NULL)
6756 fprintf (lra_dump_file,
6757 " Restoring original insn:\n");
6758 dump_insn_slim (lra_dump_file, insn);
6762 /* Clear restore_regnos. */
6763 EXECUTE_IF_SET_IN_BITMAP (&lra_optional_reload_pseudos, 0, regno, bi)
6764 lra_reg_info[regno].restore_rtx = NULL_RTX;
6765 bitmap_clear (&insn_bitmap);
6766 bitmap_clear (&removed_optional_reload_pseudos);
6767 return change_p;
6770 /* Entry function for undoing inheritance/split transformation. Return true
6771 if we did any RTL change in this pass. */
6772 bool
6773 lra_undo_inheritance (void)
6775 unsigned int regno;
6776 int hard_regno;
6777 int n_all_inherit, n_inherit, n_all_split, n_split;
6778 rtx restore_rtx;
6779 bitmap_head remove_pseudos;
6780 bitmap_iterator bi;
6781 bool change_p;
6783 lra_undo_inheritance_iter++;
6784 if (lra_undo_inheritance_iter > LRA_MAX_INHERITANCE_PASSES)
6785 return false;
6786 if (lra_dump_file != NULL)
6787 fprintf (lra_dump_file,
6788 "\n********** Undoing inheritance #%d: **********\n\n",
6789 lra_undo_inheritance_iter);
6790 bitmap_initialize (&remove_pseudos, &reg_obstack);
6791 n_inherit = n_all_inherit = 0;
6792 EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, regno, bi)
6793 if (lra_reg_info[regno].restore_rtx != NULL_RTX)
6795 n_all_inherit++;
6796 if (reg_renumber[regno] < 0
6797 /* If the original pseudo changed its allocation, just
6798 removing inheritance is dangerous as for changing
6799 allocation we used shorter live-ranges. */
6800 && (! REG_P (lra_reg_info[regno].restore_rtx)
6801 || reg_renumber[REGNO (lra_reg_info[regno].restore_rtx)] < 0))
6802 bitmap_set_bit (&remove_pseudos, regno);
6803 else
6804 n_inherit++;
6806 if (lra_dump_file != NULL && n_all_inherit != 0)
6807 fprintf (lra_dump_file, "Inherit %d out of %d (%.2f%%)\n",
6808 n_inherit, n_all_inherit,
6809 (double) n_inherit / n_all_inherit * 100);
6810 n_split = n_all_split = 0;
6811 EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, regno, bi)
6812 if ((restore_rtx = lra_reg_info[regno].restore_rtx) != NULL_RTX)
6814 int restore_regno = REGNO (restore_rtx);
6816 n_all_split++;
6817 hard_regno = (restore_regno >= FIRST_PSEUDO_REGISTER
6818 ? reg_renumber[restore_regno] : restore_regno);
6819 if (hard_regno < 0 || reg_renumber[regno] == hard_regno)
6820 bitmap_set_bit (&remove_pseudos, regno);
6821 else
6823 n_split++;
6824 if (lra_dump_file != NULL)
6825 fprintf (lra_dump_file, " Keep split r%d (orig=r%d)\n",
6826 regno, restore_regno);
6829 if (lra_dump_file != NULL && n_all_split != 0)
6830 fprintf (lra_dump_file, "Split %d out of %d (%.2f%%)\n",
6831 n_split, n_all_split,
6832 (double) n_split / n_all_split * 100);
6833 change_p = remove_inheritance_pseudos (&remove_pseudos);
6834 bitmap_clear (&remove_pseudos);
6835 /* Clear restore_regnos. */
6836 EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, regno, bi)
6837 lra_reg_info[regno].restore_rtx = NULL_RTX;
6838 EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, regno, bi)
6839 lra_reg_info[regno].restore_rtx = NULL_RTX;
6840 change_p = undo_optional_reloads () || change_p;
6841 return change_p;