Make vect_model_store_cost take a vec_load_store_type
[official-gcc.git] / gcc / lra-constraints.c
blob4d307af701138c0c4282296ad493722bc6145a23
1 /* Code for RTL transformations to satisfy insn constraints.
2 Copyright (C) 2010-2018 Free Software Foundation, Inc.
3 Contributed by Vladimir Makarov <vmakarov@redhat.com>.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
10 version.
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
22 /* This file contains code for 3 passes: constraint pass,
23 inheritance/split pass, and pass for undoing failed inheritance and
24 split.
26 The major goal of constraint pass is to transform RTL to satisfy
27 insn and address constraints by:
28 o choosing insn alternatives;
29 o generating *reload insns* (or reloads in brief) and *reload
30 pseudos* which will get necessary hard registers later;
31 o substituting pseudos with equivalent values and removing the
32 instructions that initialized those pseudos.
34 The constraint pass has biggest and most complicated code in LRA.
35 There are a lot of important details like:
36 o reuse of input reload pseudos to simplify reload pseudo
37 allocations;
38 o some heuristics to choose insn alternative to improve the
39 inheritance;
40 o early clobbers etc.
42 The pass is mimicking former reload pass in alternative choosing
43 because the reload pass is oriented to current machine description
44 model. It might be changed if the machine description model is
45 changed.
47 There is special code for preventing all LRA and this pass cycling
48 in case of bugs.
50 On the first iteration of the pass we process every instruction and
51 choose an alternative for each one. On subsequent iterations we try
52 to avoid reprocessing instructions if we can be sure that the old
53 choice is still valid.
55 The inheritance/spilt pass is to transform code to achieve
56 ineheritance and live range splitting. It is done on backward
57 traversal of EBBs.
59 The inheritance optimization goal is to reuse values in hard
60 registers. There is analogous optimization in old reload pass. The
61 inheritance is achieved by following transformation:
63 reload_p1 <- p reload_p1 <- p
64 ... new_p <- reload_p1
65 ... => ...
66 reload_p2 <- p reload_p2 <- new_p
68 where p is spilled and not changed between the insns. Reload_p1 is
69 also called *original pseudo* and new_p is called *inheritance
70 pseudo*.
72 The subsequent assignment pass will try to assign the same (or
73 another if it is not possible) hard register to new_p as to
74 reload_p1 or reload_p2.
76 If the assignment pass fails to assign a hard register to new_p,
77 this file will undo the inheritance and restore the original code.
78 This is because implementing the above sequence with a spilled
79 new_p would make the code much worse. The inheritance is done in
80 EBB scope. The above is just a simplified example to get an idea
81 of the inheritance as the inheritance is also done for non-reload
82 insns.
84 Splitting (transformation) is also done in EBB scope on the same
85 pass as the inheritance:
87 r <- ... or ... <- r r <- ... or ... <- r
88 ... s <- r (new insn -- save)
89 ... =>
90 ... r <- s (new insn -- restore)
91 ... <- r ... <- r
93 The *split pseudo* s is assigned to the hard register of the
94 original pseudo or hard register r.
96 Splitting is done:
97 o In EBBs with high register pressure for global pseudos (living
98 in at least 2 BBs) and assigned to hard registers when there
99 are more one reloads needing the hard registers;
100 o for pseudos needing save/restore code around calls.
102 If the split pseudo still has the same hard register as the
103 original pseudo after the subsequent assignment pass or the
104 original pseudo was split, the opposite transformation is done on
105 the same pass for undoing inheritance. */
107 #undef REG_OK_STRICT
109 #include "config.h"
110 #include "system.h"
111 #include "coretypes.h"
112 #include "backend.h"
113 #include "target.h"
114 #include "rtl.h"
115 #include "tree.h"
116 #include "predict.h"
117 #include "df.h"
118 #include "memmodel.h"
119 #include "tm_p.h"
120 #include "expmed.h"
121 #include "optabs.h"
122 #include "regs.h"
123 #include "ira.h"
124 #include "recog.h"
125 #include "output.h"
126 #include "addresses.h"
127 #include "expr.h"
128 #include "cfgrtl.h"
129 #include "rtl-error.h"
130 #include "params.h"
131 #include "lra.h"
132 #include "lra-int.h"
133 #include "print-rtl.h"
135 /* Value of LRA_CURR_RELOAD_NUM at the beginning of BB of the current
136 insn. Remember that LRA_CURR_RELOAD_NUM is the number of emitted
137 reload insns. */
138 static int bb_reload_num;
140 /* The current insn being processed and corresponding its single set
141 (NULL otherwise), its data (basic block, the insn data, the insn
142 static data, and the mode of each operand). */
143 static rtx_insn *curr_insn;
144 static rtx curr_insn_set;
145 static basic_block curr_bb;
146 static lra_insn_recog_data_t curr_id;
147 static struct lra_static_insn_data *curr_static_id;
148 static machine_mode curr_operand_mode[MAX_RECOG_OPERANDS];
149 /* Mode of the register substituted by its equivalence with VOIDmode
150 (e.g. constant) and whose subreg is given operand of the current
151 insn. VOIDmode in all other cases. */
152 static machine_mode original_subreg_reg_mode[MAX_RECOG_OPERANDS];
156 /* Start numbers for new registers and insns at the current constraints
157 pass start. */
158 static int new_regno_start;
159 static int new_insn_uid_start;
161 /* If LOC is nonnull, strip any outer subreg from it. */
162 static inline rtx *
163 strip_subreg (rtx *loc)
165 return loc && GET_CODE (*loc) == SUBREG ? &SUBREG_REG (*loc) : loc;
168 /* Return hard regno of REGNO or if it is was not assigned to a hard
169 register, use a hard register from its allocno class. */
170 static int
171 get_try_hard_regno (int regno)
173 int hard_regno;
174 enum reg_class rclass;
176 if ((hard_regno = regno) >= FIRST_PSEUDO_REGISTER)
177 hard_regno = lra_get_regno_hard_regno (regno);
178 if (hard_regno >= 0)
179 return hard_regno;
180 rclass = lra_get_allocno_class (regno);
181 if (rclass == NO_REGS)
182 return -1;
183 return ira_class_hard_regs[rclass][0];
186 /* Return the hard regno of X after removing its subreg. If X is not
187 a register or a subreg of a register, return -1. If X is a pseudo,
188 use its assignment. If FINAL_P return the final hard regno which will
189 be after elimination. */
190 static int
191 get_hard_regno (rtx x, bool final_p)
193 rtx reg;
194 int hard_regno;
196 reg = x;
197 if (SUBREG_P (x))
198 reg = SUBREG_REG (x);
199 if (! REG_P (reg))
200 return -1;
201 if (! HARD_REGISTER_NUM_P (hard_regno = REGNO (reg)))
202 hard_regno = lra_get_regno_hard_regno (hard_regno);
203 if (hard_regno < 0)
204 return -1;
205 if (final_p)
206 hard_regno = lra_get_elimination_hard_regno (hard_regno);
207 if (SUBREG_P (x))
208 hard_regno += subreg_regno_offset (hard_regno, GET_MODE (reg),
209 SUBREG_BYTE (x), GET_MODE (x));
210 return hard_regno;
213 /* If REGNO is a hard register or has been allocated a hard register,
214 return the class of that register. If REGNO is a reload pseudo
215 created by the current constraints pass, return its allocno class.
216 Return NO_REGS otherwise. */
217 static enum reg_class
218 get_reg_class (int regno)
220 int hard_regno;
222 if (! HARD_REGISTER_NUM_P (hard_regno = regno))
223 hard_regno = lra_get_regno_hard_regno (regno);
224 if (hard_regno >= 0)
226 hard_regno = lra_get_elimination_hard_regno (hard_regno);
227 return REGNO_REG_CLASS (hard_regno);
229 if (regno >= new_regno_start)
230 return lra_get_allocno_class (regno);
231 return NO_REGS;
234 /* Return true if REG satisfies (or will satisfy) reg class constraint
235 CL. Use elimination first if REG is a hard register. If REG is a
236 reload pseudo created by this constraints pass, assume that it will
237 be allocated a hard register from its allocno class, but allow that
238 class to be narrowed to CL if it is currently a superset of CL.
240 If NEW_CLASS is nonnull, set *NEW_CLASS to the new allocno class of
241 REGNO (reg), or NO_REGS if no change in its class was needed. */
242 static bool
243 in_class_p (rtx reg, enum reg_class cl, enum reg_class *new_class)
245 enum reg_class rclass, common_class;
246 machine_mode reg_mode;
247 int class_size, hard_regno, nregs, i, j;
248 int regno = REGNO (reg);
250 if (new_class != NULL)
251 *new_class = NO_REGS;
252 if (regno < FIRST_PSEUDO_REGISTER)
254 rtx final_reg = reg;
255 rtx *final_loc = &final_reg;
257 lra_eliminate_reg_if_possible (final_loc);
258 return TEST_HARD_REG_BIT (reg_class_contents[cl], REGNO (*final_loc));
260 reg_mode = GET_MODE (reg);
261 rclass = get_reg_class (regno);
262 if (regno < new_regno_start
263 /* Do not allow the constraints for reload instructions to
264 influence the classes of new pseudos. These reloads are
265 typically moves that have many alternatives, and restricting
266 reload pseudos for one alternative may lead to situations
267 where other reload pseudos are no longer allocatable. */
268 || (INSN_UID (curr_insn) >= new_insn_uid_start
269 && curr_insn_set != NULL
270 && ((OBJECT_P (SET_SRC (curr_insn_set))
271 && ! CONSTANT_P (SET_SRC (curr_insn_set)))
272 || (GET_CODE (SET_SRC (curr_insn_set)) == SUBREG
273 && OBJECT_P (SUBREG_REG (SET_SRC (curr_insn_set)))
274 && ! CONSTANT_P (SUBREG_REG (SET_SRC (curr_insn_set)))))))
275 /* When we don't know what class will be used finally for reload
276 pseudos, we use ALL_REGS. */
277 return ((regno >= new_regno_start && rclass == ALL_REGS)
278 || (rclass != NO_REGS && ira_class_subset_p[rclass][cl]
279 && ! hard_reg_set_subset_p (reg_class_contents[cl],
280 lra_no_alloc_regs)));
281 else
283 common_class = ira_reg_class_subset[rclass][cl];
284 if (new_class != NULL)
285 *new_class = common_class;
286 if (hard_reg_set_subset_p (reg_class_contents[common_class],
287 lra_no_alloc_regs))
288 return false;
289 /* Check that there are enough allocatable regs. */
290 class_size = ira_class_hard_regs_num[common_class];
291 for (i = 0; i < class_size; i++)
293 hard_regno = ira_class_hard_regs[common_class][i];
294 nregs = hard_regno_nregs (hard_regno, reg_mode);
295 if (nregs == 1)
296 return true;
297 for (j = 0; j < nregs; j++)
298 if (TEST_HARD_REG_BIT (lra_no_alloc_regs, hard_regno + j)
299 || ! TEST_HARD_REG_BIT (reg_class_contents[common_class],
300 hard_regno + j))
301 break;
302 if (j >= nregs)
303 return true;
305 return false;
309 /* Return true if REGNO satisfies a memory constraint. */
310 static bool
311 in_mem_p (int regno)
313 return get_reg_class (regno) == NO_REGS;
316 /* Return 1 if ADDR is a valid memory address for mode MODE in address
317 space AS, and check that each pseudo has the proper kind of hard
318 reg. */
319 static int
320 valid_address_p (machine_mode mode ATTRIBUTE_UNUSED,
321 rtx addr, addr_space_t as)
323 #ifdef GO_IF_LEGITIMATE_ADDRESS
324 lra_assert (ADDR_SPACE_GENERIC_P (as));
325 GO_IF_LEGITIMATE_ADDRESS (mode, addr, win);
326 return 0;
328 win:
329 return 1;
330 #else
331 return targetm.addr_space.legitimate_address_p (mode, addr, 0, as);
332 #endif
335 namespace {
336 /* Temporarily eliminates registers in an address (for the lifetime of
337 the object). */
338 class address_eliminator {
339 public:
340 address_eliminator (struct address_info *ad);
341 ~address_eliminator ();
343 private:
344 struct address_info *m_ad;
345 rtx *m_base_loc;
346 rtx m_base_reg;
347 rtx *m_index_loc;
348 rtx m_index_reg;
352 address_eliminator::address_eliminator (struct address_info *ad)
353 : m_ad (ad),
354 m_base_loc (strip_subreg (ad->base_term)),
355 m_base_reg (NULL_RTX),
356 m_index_loc (strip_subreg (ad->index_term)),
357 m_index_reg (NULL_RTX)
359 if (m_base_loc != NULL)
361 m_base_reg = *m_base_loc;
362 lra_eliminate_reg_if_possible (m_base_loc);
363 if (m_ad->base_term2 != NULL)
364 *m_ad->base_term2 = *m_ad->base_term;
366 if (m_index_loc != NULL)
368 m_index_reg = *m_index_loc;
369 lra_eliminate_reg_if_possible (m_index_loc);
373 address_eliminator::~address_eliminator ()
375 if (m_base_loc && *m_base_loc != m_base_reg)
377 *m_base_loc = m_base_reg;
378 if (m_ad->base_term2 != NULL)
379 *m_ad->base_term2 = *m_ad->base_term;
381 if (m_index_loc && *m_index_loc != m_index_reg)
382 *m_index_loc = m_index_reg;
385 /* Return true if the eliminated form of AD is a legitimate target address. */
386 static bool
387 valid_address_p (struct address_info *ad)
389 address_eliminator eliminator (ad);
390 return valid_address_p (ad->mode, *ad->outer, ad->as);
393 /* Return true if the eliminated form of memory reference OP satisfies
394 extra (special) memory constraint CONSTRAINT. */
395 static bool
396 satisfies_memory_constraint_p (rtx op, enum constraint_num constraint)
398 struct address_info ad;
400 decompose_mem_address (&ad, op);
401 address_eliminator eliminator (&ad);
402 return constraint_satisfied_p (op, constraint);
405 /* Return true if the eliminated form of address AD satisfies extra
406 address constraint CONSTRAINT. */
407 static bool
408 satisfies_address_constraint_p (struct address_info *ad,
409 enum constraint_num constraint)
411 address_eliminator eliminator (ad);
412 return constraint_satisfied_p (*ad->outer, constraint);
415 /* Return true if the eliminated form of address OP satisfies extra
416 address constraint CONSTRAINT. */
417 static bool
418 satisfies_address_constraint_p (rtx op, enum constraint_num constraint)
420 struct address_info ad;
422 decompose_lea_address (&ad, &op);
423 return satisfies_address_constraint_p (&ad, constraint);
426 /* Initiate equivalences for LRA. As we keep original equivalences
427 before any elimination, we need to make copies otherwise any change
428 in insns might change the equivalences. */
429 void
430 lra_init_equiv (void)
432 ira_expand_reg_equiv ();
433 for (int i = FIRST_PSEUDO_REGISTER; i < max_reg_num (); i++)
435 rtx res;
437 if ((res = ira_reg_equiv[i].memory) != NULL_RTX)
438 ira_reg_equiv[i].memory = copy_rtx (res);
439 if ((res = ira_reg_equiv[i].invariant) != NULL_RTX)
440 ira_reg_equiv[i].invariant = copy_rtx (res);
444 static rtx loc_equivalence_callback (rtx, const_rtx, void *);
446 /* Update equivalence for REGNO. We need to this as the equivalence
447 might contain other pseudos which are changed by their
448 equivalences. */
449 static void
450 update_equiv (int regno)
452 rtx x;
454 if ((x = ira_reg_equiv[regno].memory) != NULL_RTX)
455 ira_reg_equiv[regno].memory
456 = simplify_replace_fn_rtx (x, NULL_RTX, loc_equivalence_callback,
457 NULL_RTX);
458 if ((x = ira_reg_equiv[regno].invariant) != NULL_RTX)
459 ira_reg_equiv[regno].invariant
460 = simplify_replace_fn_rtx (x, NULL_RTX, loc_equivalence_callback,
461 NULL_RTX);
464 /* If we have decided to substitute X with another value, return that
465 value, otherwise return X. */
466 static rtx
467 get_equiv (rtx x)
469 int regno;
470 rtx res;
472 if (! REG_P (x) || (regno = REGNO (x)) < FIRST_PSEUDO_REGISTER
473 || ! ira_reg_equiv[regno].defined_p
474 || ! ira_reg_equiv[regno].profitable_p
475 || lra_get_regno_hard_regno (regno) >= 0)
476 return x;
477 if ((res = ira_reg_equiv[regno].memory) != NULL_RTX)
479 if (targetm.cannot_substitute_mem_equiv_p (res))
480 return x;
481 return res;
483 if ((res = ira_reg_equiv[regno].constant) != NULL_RTX)
484 return res;
485 if ((res = ira_reg_equiv[regno].invariant) != NULL_RTX)
486 return res;
487 gcc_unreachable ();
490 /* If we have decided to substitute X with the equivalent value,
491 return that value after elimination for INSN, otherwise return
492 X. */
493 static rtx
494 get_equiv_with_elimination (rtx x, rtx_insn *insn)
496 rtx res = get_equiv (x);
498 if (x == res || CONSTANT_P (res))
499 return res;
500 return lra_eliminate_regs_1 (insn, res, GET_MODE (res),
501 false, false, 0, true);
504 /* Set up curr_operand_mode. */
505 static void
506 init_curr_operand_mode (void)
508 int nop = curr_static_id->n_operands;
509 for (int i = 0; i < nop; i++)
511 machine_mode mode = GET_MODE (*curr_id->operand_loc[i]);
512 if (mode == VOIDmode)
514 /* The .md mode for address operands is the mode of the
515 addressed value rather than the mode of the address itself. */
516 if (curr_id->icode >= 0 && curr_static_id->operand[i].is_address)
517 mode = Pmode;
518 else
519 mode = curr_static_id->operand[i].mode;
521 curr_operand_mode[i] = mode;
527 /* The page contains code to reuse input reloads. */
529 /* Structure describes input reload of the current insns. */
530 struct input_reload
532 /* True for input reload of matched operands. */
533 bool match_p;
534 /* Reloaded value. */
535 rtx input;
536 /* Reload pseudo used. */
537 rtx reg;
540 /* The number of elements in the following array. */
541 static int curr_insn_input_reloads_num;
542 /* Array containing info about input reloads. It is used to find the
543 same input reload and reuse the reload pseudo in this case. */
544 static struct input_reload curr_insn_input_reloads[LRA_MAX_INSN_RELOADS];
546 /* Initiate data concerning reuse of input reloads for the current
547 insn. */
548 static void
549 init_curr_insn_input_reloads (void)
551 curr_insn_input_reloads_num = 0;
554 /* Create a new pseudo using MODE, RCLASS, ORIGINAL or reuse already
555 created input reload pseudo (only if TYPE is not OP_OUT). Don't
556 reuse pseudo if IN_SUBREG_P is true and the reused pseudo should be
557 wrapped up in SUBREG. The result pseudo is returned through
558 RESULT_REG. Return TRUE if we created a new pseudo, FALSE if we
559 reused the already created input reload pseudo. Use TITLE to
560 describe new registers for debug purposes. */
561 static bool
562 get_reload_reg (enum op_type type, machine_mode mode, rtx original,
563 enum reg_class rclass, bool in_subreg_p,
564 const char *title, rtx *result_reg)
566 int i, regno;
567 enum reg_class new_class;
568 bool unique_p = false;
570 if (type == OP_OUT)
572 *result_reg
573 = lra_create_new_reg_with_unique_value (mode, original, rclass, title);
574 return true;
576 /* Prevent reuse value of expression with side effects,
577 e.g. volatile memory. */
578 if (! side_effects_p (original))
579 for (i = 0; i < curr_insn_input_reloads_num; i++)
581 if (! curr_insn_input_reloads[i].match_p
582 && rtx_equal_p (curr_insn_input_reloads[i].input, original)
583 && in_class_p (curr_insn_input_reloads[i].reg, rclass, &new_class))
585 rtx reg = curr_insn_input_reloads[i].reg;
586 regno = REGNO (reg);
587 /* If input is equal to original and both are VOIDmode,
588 GET_MODE (reg) might be still different from mode.
589 Ensure we don't return *result_reg with wrong mode. */
590 if (GET_MODE (reg) != mode)
592 if (in_subreg_p)
593 continue;
594 if (maybe_lt (GET_MODE_SIZE (GET_MODE (reg)),
595 GET_MODE_SIZE (mode)))
596 continue;
597 reg = lowpart_subreg (mode, reg, GET_MODE (reg));
598 if (reg == NULL_RTX || GET_CODE (reg) != SUBREG)
599 continue;
601 *result_reg = reg;
602 if (lra_dump_file != NULL)
604 fprintf (lra_dump_file, " Reuse r%d for reload ", regno);
605 dump_value_slim (lra_dump_file, original, 1);
607 if (new_class != lra_get_allocno_class (regno))
608 lra_change_class (regno, new_class, ", change to", false);
609 if (lra_dump_file != NULL)
610 fprintf (lra_dump_file, "\n");
611 return false;
613 /* If we have an input reload with a different mode, make sure it
614 will get a different hard reg. */
615 else if (REG_P (original)
616 && REG_P (curr_insn_input_reloads[i].input)
617 && REGNO (original) == REGNO (curr_insn_input_reloads[i].input)
618 && (GET_MODE (original)
619 != GET_MODE (curr_insn_input_reloads[i].input)))
620 unique_p = true;
622 *result_reg = (unique_p
623 ? lra_create_new_reg_with_unique_value
624 : lra_create_new_reg) (mode, original, rclass, title);
625 lra_assert (curr_insn_input_reloads_num < LRA_MAX_INSN_RELOADS);
626 curr_insn_input_reloads[curr_insn_input_reloads_num].input = original;
627 curr_insn_input_reloads[curr_insn_input_reloads_num].match_p = false;
628 curr_insn_input_reloads[curr_insn_input_reloads_num++].reg = *result_reg;
629 return true;
634 /* The page contains code to extract memory address parts. */
636 /* Wrapper around REGNO_OK_FOR_INDEX_P, to allow pseudos. */
637 static inline bool
638 ok_for_index_p_nonstrict (rtx reg)
640 unsigned regno = REGNO (reg);
642 return regno >= FIRST_PSEUDO_REGISTER || REGNO_OK_FOR_INDEX_P (regno);
645 /* A version of regno_ok_for_base_p for use here, when all pseudos
646 should count as OK. Arguments as for regno_ok_for_base_p. */
647 static inline bool
648 ok_for_base_p_nonstrict (rtx reg, machine_mode mode, addr_space_t as,
649 enum rtx_code outer_code, enum rtx_code index_code)
651 unsigned regno = REGNO (reg);
653 if (regno >= FIRST_PSEUDO_REGISTER)
654 return true;
655 return ok_for_base_p_1 (regno, mode, as, outer_code, index_code);
660 /* The page contains major code to choose the current insn alternative
661 and generate reloads for it. */
663 /* Return the offset from REGNO of the least significant register
664 in (reg:MODE REGNO).
666 This function is used to tell whether two registers satisfy
667 a matching constraint. (reg:MODE1 REGNO1) matches (reg:MODE2 REGNO2) if:
669 REGNO1 + lra_constraint_offset (REGNO1, MODE1)
670 == REGNO2 + lra_constraint_offset (REGNO2, MODE2) */
672 lra_constraint_offset (int regno, machine_mode mode)
674 lra_assert (regno < FIRST_PSEUDO_REGISTER);
676 scalar_int_mode int_mode;
677 if (WORDS_BIG_ENDIAN
678 && is_a <scalar_int_mode> (mode, &int_mode)
679 && GET_MODE_SIZE (int_mode) > UNITS_PER_WORD)
680 return hard_regno_nregs (regno, mode) - 1;
681 return 0;
684 /* Like rtx_equal_p except that it allows a REG and a SUBREG to match
685 if they are the same hard reg, and has special hacks for
686 auto-increment and auto-decrement. This is specifically intended for
687 process_alt_operands to use in determining whether two operands
688 match. X is the operand whose number is the lower of the two.
690 It is supposed that X is the output operand and Y is the input
691 operand. Y_HARD_REGNO is the final hard regno of register Y or
692 register in subreg Y as we know it now. Otherwise, it is a
693 negative value. */
694 static bool
695 operands_match_p (rtx x, rtx y, int y_hard_regno)
697 int i;
698 RTX_CODE code = GET_CODE (x);
699 const char *fmt;
701 if (x == y)
702 return true;
703 if ((code == REG || (code == SUBREG && REG_P (SUBREG_REG (x))))
704 && (REG_P (y) || (GET_CODE (y) == SUBREG && REG_P (SUBREG_REG (y)))))
706 int j;
708 i = get_hard_regno (x, false);
709 if (i < 0)
710 goto slow;
712 if ((j = y_hard_regno) < 0)
713 goto slow;
715 i += lra_constraint_offset (i, GET_MODE (x));
716 j += lra_constraint_offset (j, GET_MODE (y));
718 return i == j;
721 /* If two operands must match, because they are really a single
722 operand of an assembler insn, then two post-increments are invalid
723 because the assembler insn would increment only once. On the
724 other hand, a post-increment matches ordinary indexing if the
725 post-increment is the output operand. */
726 if (code == POST_DEC || code == POST_INC || code == POST_MODIFY)
727 return operands_match_p (XEXP (x, 0), y, y_hard_regno);
729 /* Two pre-increments are invalid because the assembler insn would
730 increment only once. On the other hand, a pre-increment matches
731 ordinary indexing if the pre-increment is the input operand. */
732 if (GET_CODE (y) == PRE_DEC || GET_CODE (y) == PRE_INC
733 || GET_CODE (y) == PRE_MODIFY)
734 return operands_match_p (x, XEXP (y, 0), -1);
736 slow:
738 if (code == REG && REG_P (y))
739 return REGNO (x) == REGNO (y);
741 if (code == REG && GET_CODE (y) == SUBREG && REG_P (SUBREG_REG (y))
742 && x == SUBREG_REG (y))
743 return true;
744 if (GET_CODE (y) == REG && code == SUBREG && REG_P (SUBREG_REG (x))
745 && SUBREG_REG (x) == y)
746 return true;
748 /* Now we have disposed of all the cases in which different rtx
749 codes can match. */
750 if (code != GET_CODE (y))
751 return false;
753 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
754 if (GET_MODE (x) != GET_MODE (y))
755 return false;
757 switch (code)
759 CASE_CONST_UNIQUE:
760 return false;
762 case LABEL_REF:
763 return label_ref_label (x) == label_ref_label (y);
764 case SYMBOL_REF:
765 return XSTR (x, 0) == XSTR (y, 0);
767 default:
768 break;
771 /* Compare the elements. If any pair of corresponding elements fail
772 to match, return false for the whole things. */
774 fmt = GET_RTX_FORMAT (code);
775 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
777 int val, j;
778 switch (fmt[i])
780 case 'w':
781 if (XWINT (x, i) != XWINT (y, i))
782 return false;
783 break;
785 case 'i':
786 if (XINT (x, i) != XINT (y, i))
787 return false;
788 break;
790 case 'p':
791 if (maybe_ne (SUBREG_BYTE (x), SUBREG_BYTE (y)))
792 return false;
793 break;
795 case 'e':
796 val = operands_match_p (XEXP (x, i), XEXP (y, i), -1);
797 if (val == 0)
798 return false;
799 break;
801 case '0':
802 break;
804 case 'E':
805 if (XVECLEN (x, i) != XVECLEN (y, i))
806 return false;
807 for (j = XVECLEN (x, i) - 1; j >= 0; --j)
809 val = operands_match_p (XVECEXP (x, i, j), XVECEXP (y, i, j), -1);
810 if (val == 0)
811 return false;
813 break;
815 /* It is believed that rtx's at this level will never
816 contain anything but integers and other rtx's, except for
817 within LABEL_REFs and SYMBOL_REFs. */
818 default:
819 gcc_unreachable ();
822 return true;
825 /* True if X is a constant that can be forced into the constant pool.
826 MODE is the mode of the operand, or VOIDmode if not known. */
827 #define CONST_POOL_OK_P(MODE, X) \
828 ((MODE) != VOIDmode \
829 && CONSTANT_P (X) \
830 && GET_CODE (X) != HIGH \
831 && GET_MODE_SIZE (MODE).is_constant () \
832 && !targetm.cannot_force_const_mem (MODE, X))
834 /* True if C is a non-empty register class that has too few registers
835 to be safely used as a reload target class. */
836 #define SMALL_REGISTER_CLASS_P(C) \
837 (ira_class_hard_regs_num [(C)] == 1 \
838 || (ira_class_hard_regs_num [(C)] >= 1 \
839 && targetm.class_likely_spilled_p (C)))
841 /* If REG is a reload pseudo, try to make its class satisfying CL. */
842 static void
843 narrow_reload_pseudo_class (rtx reg, enum reg_class cl)
845 enum reg_class rclass;
847 /* Do not make more accurate class from reloads generated. They are
848 mostly moves with a lot of constraints. Making more accurate
849 class may results in very narrow class and impossibility of find
850 registers for several reloads of one insn. */
851 if (INSN_UID (curr_insn) >= new_insn_uid_start)
852 return;
853 if (GET_CODE (reg) == SUBREG)
854 reg = SUBREG_REG (reg);
855 if (! REG_P (reg) || (int) REGNO (reg) < new_regno_start)
856 return;
857 if (in_class_p (reg, cl, &rclass) && rclass != cl)
858 lra_change_class (REGNO (reg), rclass, " Change to", true);
861 /* Searches X for any reference to a reg with the same value as REGNO,
862 returning the rtx of the reference found if any. Otherwise,
863 returns NULL_RTX. */
864 static rtx
865 regno_val_use_in (unsigned int regno, rtx x)
867 const char *fmt;
868 int i, j;
869 rtx tem;
871 if (REG_P (x) && lra_reg_info[REGNO (x)].val == lra_reg_info[regno].val)
872 return x;
874 fmt = GET_RTX_FORMAT (GET_CODE (x));
875 for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
877 if (fmt[i] == 'e')
879 if ((tem = regno_val_use_in (regno, XEXP (x, i))))
880 return tem;
882 else if (fmt[i] == 'E')
883 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
884 if ((tem = regno_val_use_in (regno , XVECEXP (x, i, j))))
885 return tem;
888 return NULL_RTX;
891 /* Return true if all current insn non-output operands except INS (it
892 has a negaitve end marker) do not use pseudos with the same value
893 as REGNO. */
894 static bool
895 check_conflict_input_operands (int regno, signed char *ins)
897 int in;
898 int n_operands = curr_static_id->n_operands;
900 for (int nop = 0; nop < n_operands; nop++)
901 if (! curr_static_id->operand[nop].is_operator
902 && curr_static_id->operand[nop].type != OP_OUT)
904 for (int i = 0; (in = ins[i]) >= 0; i++)
905 if (in == nop)
906 break;
907 if (in < 0
908 && regno_val_use_in (regno, *curr_id->operand_loc[nop]) != NULL_RTX)
909 return false;
911 return true;
914 /* Generate reloads for matching OUT and INS (array of input operand
915 numbers with end marker -1) with reg class GOAL_CLASS, considering
916 output operands OUTS (similar array to INS) needing to be in different
917 registers. Add input and output reloads correspondingly to the lists
918 *BEFORE and *AFTER. OUT might be negative. In this case we generate
919 input reloads for matched input operands INS. EARLY_CLOBBER_P is a flag
920 that the output operand is early clobbered for chosen alternative. */
921 static void
922 match_reload (signed char out, signed char *ins, signed char *outs,
923 enum reg_class goal_class, rtx_insn **before,
924 rtx_insn **after, bool early_clobber_p)
926 bool out_conflict;
927 int i, in;
928 rtx new_in_reg, new_out_reg, reg;
929 machine_mode inmode, outmode;
930 rtx in_rtx = *curr_id->operand_loc[ins[0]];
931 rtx out_rtx = out < 0 ? in_rtx : *curr_id->operand_loc[out];
933 inmode = curr_operand_mode[ins[0]];
934 outmode = out < 0 ? inmode : curr_operand_mode[out];
935 push_to_sequence (*before);
936 if (inmode != outmode)
938 /* process_alt_operands has already checked that the mode sizes
939 are ordered. */
940 if (partial_subreg_p (outmode, inmode))
942 reg = new_in_reg
943 = lra_create_new_reg_with_unique_value (inmode, in_rtx,
944 goal_class, "");
945 if (SCALAR_INT_MODE_P (inmode))
946 new_out_reg = gen_lowpart_SUBREG (outmode, reg);
947 else
948 new_out_reg = gen_rtx_SUBREG (outmode, reg, 0);
949 LRA_SUBREG_P (new_out_reg) = 1;
950 /* If the input reg is dying here, we can use the same hard
951 register for REG and IN_RTX. We do it only for original
952 pseudos as reload pseudos can die although original
953 pseudos still live where reload pseudos dies. */
954 if (REG_P (in_rtx) && (int) REGNO (in_rtx) < lra_new_regno_start
955 && find_regno_note (curr_insn, REG_DEAD, REGNO (in_rtx))
956 && (!early_clobber_p
957 || check_conflict_input_operands(REGNO (in_rtx), ins)))
958 lra_assign_reg_val (REGNO (in_rtx), REGNO (reg));
960 else
962 reg = new_out_reg
963 = lra_create_new_reg_with_unique_value (outmode, out_rtx,
964 goal_class, "");
965 if (SCALAR_INT_MODE_P (outmode))
966 new_in_reg = gen_lowpart_SUBREG (inmode, reg);
967 else
968 new_in_reg = gen_rtx_SUBREG (inmode, reg, 0);
969 /* NEW_IN_REG is non-paradoxical subreg. We don't want
970 NEW_OUT_REG living above. We add clobber clause for
971 this. This is just a temporary clobber. We can remove
972 it at the end of LRA work. */
973 rtx_insn *clobber = emit_clobber (new_out_reg);
974 LRA_TEMP_CLOBBER_P (PATTERN (clobber)) = 1;
975 LRA_SUBREG_P (new_in_reg) = 1;
976 if (GET_CODE (in_rtx) == SUBREG)
978 rtx subreg_reg = SUBREG_REG (in_rtx);
980 /* If SUBREG_REG is dying here and sub-registers IN_RTX
981 and NEW_IN_REG are similar, we can use the same hard
982 register for REG and SUBREG_REG. */
983 if (REG_P (subreg_reg)
984 && (int) REGNO (subreg_reg) < lra_new_regno_start
985 && GET_MODE (subreg_reg) == outmode
986 && known_eq (SUBREG_BYTE (in_rtx), SUBREG_BYTE (new_in_reg))
987 && find_regno_note (curr_insn, REG_DEAD, REGNO (subreg_reg))
988 && (! early_clobber_p
989 || check_conflict_input_operands (REGNO (subreg_reg),
990 ins)))
991 lra_assign_reg_val (REGNO (subreg_reg), REGNO (reg));
995 else
997 /* Pseudos have values -- see comments for lra_reg_info.
998 Different pseudos with the same value do not conflict even if
999 they live in the same place. When we create a pseudo we
1000 assign value of original pseudo (if any) from which we
1001 created the new pseudo. If we create the pseudo from the
1002 input pseudo, the new pseudo will have no conflict with the
1003 input pseudo which is wrong when the input pseudo lives after
1004 the insn and as the new pseudo value is changed by the insn
1005 output. Therefore we create the new pseudo from the output
1006 except the case when we have single matched dying input
1007 pseudo.
1009 We cannot reuse the current output register because we might
1010 have a situation like "a <- a op b", where the constraints
1011 force the second input operand ("b") to match the output
1012 operand ("a"). "b" must then be copied into a new register
1013 so that it doesn't clobber the current value of "a".
1015 We can not use the same value if the output pseudo is
1016 early clobbered or the input pseudo is mentioned in the
1017 output, e.g. as an address part in memory, because
1018 output reload will actually extend the pseudo liveness.
1019 We don't care about eliminable hard regs here as we are
1020 interesting only in pseudos. */
1022 /* Matching input's register value is the same as one of the other
1023 output operand. Output operands in a parallel insn must be in
1024 different registers. */
1025 out_conflict = false;
1026 if (REG_P (in_rtx))
1028 for (i = 0; outs[i] >= 0; i++)
1030 rtx other_out_rtx = *curr_id->operand_loc[outs[i]];
1031 if (REG_P (other_out_rtx)
1032 && (regno_val_use_in (REGNO (in_rtx), other_out_rtx)
1033 != NULL_RTX))
1035 out_conflict = true;
1036 break;
1041 new_in_reg = new_out_reg
1042 = (! early_clobber_p && ins[1] < 0 && REG_P (in_rtx)
1043 && (int) REGNO (in_rtx) < lra_new_regno_start
1044 && find_regno_note (curr_insn, REG_DEAD, REGNO (in_rtx))
1045 && (! early_clobber_p
1046 || check_conflict_input_operands (REGNO (in_rtx), ins))
1047 && (out < 0
1048 || regno_val_use_in (REGNO (in_rtx), out_rtx) == NULL_RTX)
1049 && !out_conflict
1050 ? lra_create_new_reg (inmode, in_rtx, goal_class, "")
1051 : lra_create_new_reg_with_unique_value (outmode, out_rtx,
1052 goal_class, ""));
1054 /* In operand can be got from transformations before processing insn
1055 constraints. One example of such transformations is subreg
1056 reloading (see function simplify_operand_subreg). The new
1057 pseudos created by the transformations might have inaccurate
1058 class (ALL_REGS) and we should make their classes more
1059 accurate. */
1060 narrow_reload_pseudo_class (in_rtx, goal_class);
1061 lra_emit_move (copy_rtx (new_in_reg), in_rtx);
1062 *before = get_insns ();
1063 end_sequence ();
1064 /* Add the new pseudo to consider values of subsequent input reload
1065 pseudos. */
1066 lra_assert (curr_insn_input_reloads_num < LRA_MAX_INSN_RELOADS);
1067 curr_insn_input_reloads[curr_insn_input_reloads_num].input = in_rtx;
1068 curr_insn_input_reloads[curr_insn_input_reloads_num].match_p = true;
1069 curr_insn_input_reloads[curr_insn_input_reloads_num++].reg = new_in_reg;
1070 for (i = 0; (in = ins[i]) >= 0; i++)
1072 lra_assert
1073 (GET_MODE (*curr_id->operand_loc[in]) == VOIDmode
1074 || GET_MODE (new_in_reg) == GET_MODE (*curr_id->operand_loc[in]));
1075 *curr_id->operand_loc[in] = new_in_reg;
1077 lra_update_dups (curr_id, ins);
1078 if (out < 0)
1079 return;
1080 /* See a comment for the input operand above. */
1081 narrow_reload_pseudo_class (out_rtx, goal_class);
1082 if (find_reg_note (curr_insn, REG_UNUSED, out_rtx) == NULL_RTX)
1084 start_sequence ();
1085 lra_emit_move (out_rtx, copy_rtx (new_out_reg));
1086 emit_insn (*after);
1087 *after = get_insns ();
1088 end_sequence ();
1090 *curr_id->operand_loc[out] = new_out_reg;
1091 lra_update_dup (curr_id, out);
1094 /* Return register class which is union of all reg classes in insn
1095 constraint alternative string starting with P. */
1096 static enum reg_class
1097 reg_class_from_constraints (const char *p)
1099 int c, len;
1100 enum reg_class op_class = NO_REGS;
1103 switch ((c = *p, len = CONSTRAINT_LEN (c, p)), c)
1105 case '#':
1106 case ',':
1107 return op_class;
1109 case 'g':
1110 op_class = reg_class_subunion[op_class][GENERAL_REGS];
1111 break;
1113 default:
1114 enum constraint_num cn = lookup_constraint (p);
1115 enum reg_class cl = reg_class_for_constraint (cn);
1116 if (cl == NO_REGS)
1118 if (insn_extra_address_constraint (cn))
1119 op_class
1120 = (reg_class_subunion
1121 [op_class][base_reg_class (VOIDmode, ADDR_SPACE_GENERIC,
1122 ADDRESS, SCRATCH)]);
1123 break;
1126 op_class = reg_class_subunion[op_class][cl];
1127 break;
1129 while ((p += len), c);
1130 return op_class;
1133 /* If OP is a register, return the class of the register as per
1134 get_reg_class, otherwise return NO_REGS. */
1135 static inline enum reg_class
1136 get_op_class (rtx op)
1138 return REG_P (op) ? get_reg_class (REGNO (op)) : NO_REGS;
1141 /* Return generated insn mem_pseudo:=val if TO_P or val:=mem_pseudo
1142 otherwise. If modes of MEM_PSEUDO and VAL are different, use
1143 SUBREG for VAL to make them equal. */
1144 static rtx_insn *
1145 emit_spill_move (bool to_p, rtx mem_pseudo, rtx val)
1147 if (GET_MODE (mem_pseudo) != GET_MODE (val))
1149 /* Usually size of mem_pseudo is greater than val size but in
1150 rare cases it can be less as it can be defined by target
1151 dependent macro HARD_REGNO_CALLER_SAVE_MODE. */
1152 if (! MEM_P (val))
1154 val = gen_lowpart_SUBREG (GET_MODE (mem_pseudo),
1155 GET_CODE (val) == SUBREG
1156 ? SUBREG_REG (val) : val);
1157 LRA_SUBREG_P (val) = 1;
1159 else
1161 mem_pseudo = gen_lowpart_SUBREG (GET_MODE (val), mem_pseudo);
1162 LRA_SUBREG_P (mem_pseudo) = 1;
1165 return to_p ? gen_move_insn (mem_pseudo, val)
1166 : gen_move_insn (val, mem_pseudo);
1169 /* Process a special case insn (register move), return true if we
1170 don't need to process it anymore. INSN should be a single set
1171 insn. Set up that RTL was changed through CHANGE_P and that hook
1172 TARGET_SECONDARY_MEMORY_NEEDED says to use secondary memory through
1173 SEC_MEM_P. */
1174 static bool
1175 check_and_process_move (bool *change_p, bool *sec_mem_p ATTRIBUTE_UNUSED)
1177 int sregno, dregno;
1178 rtx dest, src, dreg, sreg, new_reg, scratch_reg;
1179 rtx_insn *before;
1180 enum reg_class dclass, sclass, secondary_class;
1181 secondary_reload_info sri;
1183 lra_assert (curr_insn_set != NULL_RTX);
1184 dreg = dest = SET_DEST (curr_insn_set);
1185 sreg = src = SET_SRC (curr_insn_set);
1186 if (GET_CODE (dest) == SUBREG)
1187 dreg = SUBREG_REG (dest);
1188 if (GET_CODE (src) == SUBREG)
1189 sreg = SUBREG_REG (src);
1190 if (! (REG_P (dreg) || MEM_P (dreg)) || ! (REG_P (sreg) || MEM_P (sreg)))
1191 return false;
1192 sclass = dclass = NO_REGS;
1193 if (REG_P (dreg))
1194 dclass = get_reg_class (REGNO (dreg));
1195 gcc_assert (dclass < LIM_REG_CLASSES);
1196 if (dclass == ALL_REGS)
1197 /* ALL_REGS is used for new pseudos created by transformations
1198 like reload of SUBREG_REG (see function
1199 simplify_operand_subreg). We don't know their class yet. We
1200 should figure out the class from processing the insn
1201 constraints not in this fast path function. Even if ALL_REGS
1202 were a right class for the pseudo, secondary_... hooks usually
1203 are not define for ALL_REGS. */
1204 return false;
1205 if (REG_P (sreg))
1206 sclass = get_reg_class (REGNO (sreg));
1207 gcc_assert (sclass < LIM_REG_CLASSES);
1208 if (sclass == ALL_REGS)
1209 /* See comments above. */
1210 return false;
1211 if (sclass == NO_REGS && dclass == NO_REGS)
1212 return false;
1213 if (targetm.secondary_memory_needed (GET_MODE (src), sclass, dclass)
1214 && ((sclass != NO_REGS && dclass != NO_REGS)
1215 || (GET_MODE (src)
1216 != targetm.secondary_memory_needed_mode (GET_MODE (src)))))
1218 *sec_mem_p = true;
1219 return false;
1221 if (! REG_P (dreg) || ! REG_P (sreg))
1222 return false;
1223 sri.prev_sri = NULL;
1224 sri.icode = CODE_FOR_nothing;
1225 sri.extra_cost = 0;
1226 secondary_class = NO_REGS;
1227 /* Set up hard register for a reload pseudo for hook
1228 secondary_reload because some targets just ignore unassigned
1229 pseudos in the hook. */
1230 if (dclass != NO_REGS && lra_get_regno_hard_regno (REGNO (dreg)) < 0)
1232 dregno = REGNO (dreg);
1233 reg_renumber[dregno] = ira_class_hard_regs[dclass][0];
1235 else
1236 dregno = -1;
1237 if (sclass != NO_REGS && lra_get_regno_hard_regno (REGNO (sreg)) < 0)
1239 sregno = REGNO (sreg);
1240 reg_renumber[sregno] = ira_class_hard_regs[sclass][0];
1242 else
1243 sregno = -1;
1244 if (sclass != NO_REGS)
1245 secondary_class
1246 = (enum reg_class) targetm.secondary_reload (false, dest,
1247 (reg_class_t) sclass,
1248 GET_MODE (src), &sri);
1249 if (sclass == NO_REGS
1250 || ((secondary_class != NO_REGS || sri.icode != CODE_FOR_nothing)
1251 && dclass != NO_REGS))
1253 enum reg_class old_sclass = secondary_class;
1254 secondary_reload_info old_sri = sri;
1256 sri.prev_sri = NULL;
1257 sri.icode = CODE_FOR_nothing;
1258 sri.extra_cost = 0;
1259 secondary_class
1260 = (enum reg_class) targetm.secondary_reload (true, src,
1261 (reg_class_t) dclass,
1262 GET_MODE (src), &sri);
1263 /* Check the target hook consistency. */
1264 lra_assert
1265 ((secondary_class == NO_REGS && sri.icode == CODE_FOR_nothing)
1266 || (old_sclass == NO_REGS && old_sri.icode == CODE_FOR_nothing)
1267 || (secondary_class == old_sclass && sri.icode == old_sri.icode));
1269 if (sregno >= 0)
1270 reg_renumber [sregno] = -1;
1271 if (dregno >= 0)
1272 reg_renumber [dregno] = -1;
1273 if (secondary_class == NO_REGS && sri.icode == CODE_FOR_nothing)
1274 return false;
1275 *change_p = true;
1276 new_reg = NULL_RTX;
1277 if (secondary_class != NO_REGS)
1278 new_reg = lra_create_new_reg_with_unique_value (GET_MODE (src), NULL_RTX,
1279 secondary_class,
1280 "secondary");
1281 start_sequence ();
1282 if (sri.icode == CODE_FOR_nothing)
1283 lra_emit_move (new_reg, src);
1284 else
1286 enum reg_class scratch_class;
1288 scratch_class = (reg_class_from_constraints
1289 (insn_data[sri.icode].operand[2].constraint));
1290 scratch_reg = (lra_create_new_reg_with_unique_value
1291 (insn_data[sri.icode].operand[2].mode, NULL_RTX,
1292 scratch_class, "scratch"));
1293 emit_insn (GEN_FCN (sri.icode) (new_reg != NULL_RTX ? new_reg : dest,
1294 src, scratch_reg));
1296 before = get_insns ();
1297 end_sequence ();
1298 lra_process_new_insns (curr_insn, before, NULL, "Inserting the move");
1299 if (new_reg != NULL_RTX)
1300 SET_SRC (curr_insn_set) = new_reg;
1301 else
1303 if (lra_dump_file != NULL)
1305 fprintf (lra_dump_file, "Deleting move %u\n", INSN_UID (curr_insn));
1306 dump_insn_slim (lra_dump_file, curr_insn);
1308 lra_set_insn_deleted (curr_insn);
1309 return true;
1311 return false;
1314 /* The following data describe the result of process_alt_operands.
1315 The data are used in curr_insn_transform to generate reloads. */
1317 /* The chosen reg classes which should be used for the corresponding
1318 operands. */
1319 static enum reg_class goal_alt[MAX_RECOG_OPERANDS];
1320 /* True if the operand should be the same as another operand and that
1321 other operand does not need a reload. */
1322 static bool goal_alt_match_win[MAX_RECOG_OPERANDS];
1323 /* True if the operand does not need a reload. */
1324 static bool goal_alt_win[MAX_RECOG_OPERANDS];
1325 /* True if the operand can be offsetable memory. */
1326 static bool goal_alt_offmemok[MAX_RECOG_OPERANDS];
1327 /* The number of an operand to which given operand can be matched to. */
1328 static int goal_alt_matches[MAX_RECOG_OPERANDS];
1329 /* The number of elements in the following array. */
1330 static int goal_alt_dont_inherit_ops_num;
1331 /* Numbers of operands whose reload pseudos should not be inherited. */
1332 static int goal_alt_dont_inherit_ops[MAX_RECOG_OPERANDS];
1333 /* True if the insn commutative operands should be swapped. */
1334 static bool goal_alt_swapped;
1335 /* The chosen insn alternative. */
1336 static int goal_alt_number;
1338 /* True if the corresponding operand is the result of an equivalence
1339 substitution. */
1340 static bool equiv_substition_p[MAX_RECOG_OPERANDS];
1342 /* The following five variables are used to choose the best insn
1343 alternative. They reflect final characteristics of the best
1344 alternative. */
1346 /* Number of necessary reloads and overall cost reflecting the
1347 previous value and other unpleasantness of the best alternative. */
1348 static int best_losers, best_overall;
1349 /* Overall number hard registers used for reloads. For example, on
1350 some targets we need 2 general registers to reload DFmode and only
1351 one floating point register. */
1352 static int best_reload_nregs;
1353 /* Overall number reflecting distances of previous reloading the same
1354 value. The distances are counted from the current BB start. It is
1355 used to improve inheritance chances. */
1356 static int best_reload_sum;
1358 /* True if the current insn should have no correspondingly input or
1359 output reloads. */
1360 static bool no_input_reloads_p, no_output_reloads_p;
1362 /* True if we swapped the commutative operands in the current
1363 insn. */
1364 static int curr_swapped;
1366 /* if CHECK_ONLY_P is false, arrange for address element *LOC to be a
1367 register of class CL. Add any input reloads to list BEFORE. AFTER
1368 is nonnull if *LOC is an automodified value; handle that case by
1369 adding the required output reloads to list AFTER. Return true if
1370 the RTL was changed.
1372 if CHECK_ONLY_P is true, check that the *LOC is a correct address
1373 register. Return false if the address register is correct. */
1374 static bool
1375 process_addr_reg (rtx *loc, bool check_only_p, rtx_insn **before, rtx_insn **after,
1376 enum reg_class cl)
1378 int regno;
1379 enum reg_class rclass, new_class;
1380 rtx reg;
1381 rtx new_reg;
1382 machine_mode mode;
1383 bool subreg_p, before_p = false;
1385 subreg_p = GET_CODE (*loc) == SUBREG;
1386 if (subreg_p)
1388 reg = SUBREG_REG (*loc);
1389 mode = GET_MODE (reg);
1391 /* For mode with size bigger than ptr_mode, there unlikely to be "mov"
1392 between two registers with different classes, but there normally will
1393 be "mov" which transfers element of vector register into the general
1394 register, and this normally will be a subreg which should be reloaded
1395 as a whole. This is particularly likely to be triggered when
1396 -fno-split-wide-types specified. */
1397 if (!REG_P (reg)
1398 || in_class_p (reg, cl, &new_class)
1399 || known_le (GET_MODE_SIZE (mode), GET_MODE_SIZE (ptr_mode)))
1400 loc = &SUBREG_REG (*loc);
1403 reg = *loc;
1404 mode = GET_MODE (reg);
1405 if (! REG_P (reg))
1407 if (check_only_p)
1408 return true;
1409 /* Always reload memory in an address even if the target supports
1410 such addresses. */
1411 new_reg = lra_create_new_reg_with_unique_value (mode, reg, cl, "address");
1412 before_p = true;
1414 else
1416 regno = REGNO (reg);
1417 rclass = get_reg_class (regno);
1418 if (! check_only_p
1419 && (*loc = get_equiv_with_elimination (reg, curr_insn)) != reg)
1421 if (lra_dump_file != NULL)
1423 fprintf (lra_dump_file,
1424 "Changing pseudo %d in address of insn %u on equiv ",
1425 REGNO (reg), INSN_UID (curr_insn));
1426 dump_value_slim (lra_dump_file, *loc, 1);
1427 fprintf (lra_dump_file, "\n");
1429 *loc = copy_rtx (*loc);
1431 if (*loc != reg || ! in_class_p (reg, cl, &new_class))
1433 if (check_only_p)
1434 return true;
1435 reg = *loc;
1436 if (get_reload_reg (after == NULL ? OP_IN : OP_INOUT,
1437 mode, reg, cl, subreg_p, "address", &new_reg))
1438 before_p = true;
1440 else if (new_class != NO_REGS && rclass != new_class)
1442 if (check_only_p)
1443 return true;
1444 lra_change_class (regno, new_class, " Change to", true);
1445 return false;
1447 else
1448 return false;
1450 if (before_p)
1452 push_to_sequence (*before);
1453 lra_emit_move (new_reg, reg);
1454 *before = get_insns ();
1455 end_sequence ();
1457 *loc = new_reg;
1458 if (after != NULL)
1460 start_sequence ();
1461 lra_emit_move (before_p ? copy_rtx (reg) : reg, new_reg);
1462 emit_insn (*after);
1463 *after = get_insns ();
1464 end_sequence ();
1466 return true;
1469 /* Insert move insn in simplify_operand_subreg. BEFORE returns
1470 the insn to be inserted before curr insn. AFTER returns the
1471 the insn to be inserted after curr insn. ORIGREG and NEWREG
1472 are the original reg and new reg for reload. */
1473 static void
1474 insert_move_for_subreg (rtx_insn **before, rtx_insn **after, rtx origreg,
1475 rtx newreg)
1477 if (before)
1479 push_to_sequence (*before);
1480 lra_emit_move (newreg, origreg);
1481 *before = get_insns ();
1482 end_sequence ();
1484 if (after)
1486 start_sequence ();
1487 lra_emit_move (origreg, newreg);
1488 emit_insn (*after);
1489 *after = get_insns ();
1490 end_sequence ();
1494 static int valid_address_p (machine_mode mode, rtx addr, addr_space_t as);
1495 static bool process_address (int, bool, rtx_insn **, rtx_insn **);
1497 /* Make reloads for subreg in operand NOP with internal subreg mode
1498 REG_MODE, add new reloads for further processing. Return true if
1499 any change was done. */
1500 static bool
1501 simplify_operand_subreg (int nop, machine_mode reg_mode)
1503 int hard_regno;
1504 rtx_insn *before, *after;
1505 machine_mode mode, innermode;
1506 rtx reg, new_reg;
1507 rtx operand = *curr_id->operand_loc[nop];
1508 enum reg_class regclass;
1509 enum op_type type;
1511 before = after = NULL;
1513 if (GET_CODE (operand) != SUBREG)
1514 return false;
1516 mode = GET_MODE (operand);
1517 reg = SUBREG_REG (operand);
1518 innermode = GET_MODE (reg);
1519 type = curr_static_id->operand[nop].type;
1520 if (MEM_P (reg))
1522 const bool addr_was_valid
1523 = valid_address_p (innermode, XEXP (reg, 0), MEM_ADDR_SPACE (reg));
1524 alter_subreg (curr_id->operand_loc[nop], false);
1525 rtx subst = *curr_id->operand_loc[nop];
1526 lra_assert (MEM_P (subst));
1528 if (!addr_was_valid
1529 || valid_address_p (GET_MODE (subst), XEXP (subst, 0),
1530 MEM_ADDR_SPACE (subst))
1531 || ((get_constraint_type (lookup_constraint
1532 (curr_static_id->operand[nop].constraint))
1533 != CT_SPECIAL_MEMORY)
1534 /* We still can reload address and if the address is
1535 valid, we can remove subreg without reloading its
1536 inner memory. */
1537 && valid_address_p (GET_MODE (subst),
1538 regno_reg_rtx
1539 [ira_class_hard_regs
1540 [base_reg_class (GET_MODE (subst),
1541 MEM_ADDR_SPACE (subst),
1542 ADDRESS, SCRATCH)][0]],
1543 MEM_ADDR_SPACE (subst))))
1545 /* If we change the address for a paradoxical subreg of memory, the
1546 new address might violate the necessary alignment or the access
1547 might be slow; take this into consideration. We need not worry
1548 about accesses beyond allocated memory for paradoxical memory
1549 subregs as we don't substitute such equiv memory (see processing
1550 equivalences in function lra_constraints) and because for spilled
1551 pseudos we allocate stack memory enough for the biggest
1552 corresponding paradoxical subreg.
1554 However, do not blindly simplify a (subreg (mem ...)) for
1555 WORD_REGISTER_OPERATIONS targets as this may lead to loading junk
1556 data into a register when the inner is narrower than outer or
1557 missing important data from memory when the inner is wider than
1558 outer. This rule only applies to modes that are no wider than
1559 a word. */
1560 if (!(maybe_ne (GET_MODE_PRECISION (mode),
1561 GET_MODE_PRECISION (innermode))
1562 && known_le (GET_MODE_SIZE (mode), UNITS_PER_WORD)
1563 && known_le (GET_MODE_SIZE (innermode), UNITS_PER_WORD)
1564 && WORD_REGISTER_OPERATIONS)
1565 && (!(MEM_ALIGN (subst) < GET_MODE_ALIGNMENT (mode)
1566 && targetm.slow_unaligned_access (mode, MEM_ALIGN (subst)))
1567 || (MEM_ALIGN (reg) < GET_MODE_ALIGNMENT (innermode)
1568 && targetm.slow_unaligned_access (innermode,
1569 MEM_ALIGN (reg)))))
1570 return true;
1572 *curr_id->operand_loc[nop] = operand;
1574 /* But if the address was not valid, we cannot reload the MEM without
1575 reloading the address first. */
1576 if (!addr_was_valid)
1577 process_address (nop, false, &before, &after);
1579 /* INNERMODE is fast, MODE slow. Reload the mem in INNERMODE. */
1580 enum reg_class rclass
1581 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1582 if (get_reload_reg (curr_static_id->operand[nop].type, innermode,
1583 reg, rclass, TRUE, "slow mem", &new_reg))
1585 bool insert_before, insert_after;
1586 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1588 insert_before = (type != OP_OUT
1589 || partial_subreg_p (mode, innermode));
1590 insert_after = type != OP_IN;
1591 insert_move_for_subreg (insert_before ? &before : NULL,
1592 insert_after ? &after : NULL,
1593 reg, new_reg);
1595 SUBREG_REG (operand) = new_reg;
1597 /* Convert to MODE. */
1598 reg = operand;
1599 rclass
1600 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1601 if (get_reload_reg (curr_static_id->operand[nop].type, mode, reg,
1602 rclass, TRUE, "slow mem", &new_reg))
1604 bool insert_before, insert_after;
1605 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1607 insert_before = type != OP_OUT;
1608 insert_after = type != OP_IN;
1609 insert_move_for_subreg (insert_before ? &before : NULL,
1610 insert_after ? &after : NULL,
1611 reg, new_reg);
1613 *curr_id->operand_loc[nop] = new_reg;
1614 lra_process_new_insns (curr_insn, before, after,
1615 "Inserting slow mem reload");
1616 return true;
1619 /* If the address was valid and became invalid, prefer to reload
1620 the memory. Typical case is when the index scale should
1621 correspond the memory. */
1622 *curr_id->operand_loc[nop] = operand;
1623 /* Do not return false here as the MEM_P (reg) will be processed
1624 later in this function. */
1626 else if (REG_P (reg) && REGNO (reg) < FIRST_PSEUDO_REGISTER)
1628 alter_subreg (curr_id->operand_loc[nop], false);
1629 return true;
1631 else if (CONSTANT_P (reg))
1633 /* Try to simplify subreg of constant. It is usually result of
1634 equivalence substitution. */
1635 if (innermode == VOIDmode
1636 && (innermode = original_subreg_reg_mode[nop]) == VOIDmode)
1637 innermode = curr_static_id->operand[nop].mode;
1638 if ((new_reg = simplify_subreg (mode, reg, innermode,
1639 SUBREG_BYTE (operand))) != NULL_RTX)
1641 *curr_id->operand_loc[nop] = new_reg;
1642 return true;
1645 /* Put constant into memory when we have mixed modes. It generates
1646 a better code in most cases as it does not need a secondary
1647 reload memory. It also prevents LRA looping when LRA is using
1648 secondary reload memory again and again. */
1649 if (CONSTANT_P (reg) && CONST_POOL_OK_P (reg_mode, reg)
1650 && SCALAR_INT_MODE_P (reg_mode) != SCALAR_INT_MODE_P (mode))
1652 SUBREG_REG (operand) = force_const_mem (reg_mode, reg);
1653 alter_subreg (curr_id->operand_loc[nop], false);
1654 return true;
1656 /* Force a reload of the SUBREG_REG if this is a constant or PLUS or
1657 if there may be a problem accessing OPERAND in the outer
1658 mode. */
1659 if ((REG_P (reg)
1660 && REGNO (reg) >= FIRST_PSEUDO_REGISTER
1661 && (hard_regno = lra_get_regno_hard_regno (REGNO (reg))) >= 0
1662 /* Don't reload paradoxical subregs because we could be looping
1663 having repeatedly final regno out of hard regs range. */
1664 && (hard_regno_nregs (hard_regno, innermode)
1665 >= hard_regno_nregs (hard_regno, mode))
1666 && simplify_subreg_regno (hard_regno, innermode,
1667 SUBREG_BYTE (operand), mode) < 0
1668 /* Don't reload subreg for matching reload. It is actually
1669 valid subreg in LRA. */
1670 && ! LRA_SUBREG_P (operand))
1671 || CONSTANT_P (reg) || GET_CODE (reg) == PLUS || MEM_P (reg))
1673 enum reg_class rclass;
1675 if (REG_P (reg))
1676 /* There is a big probability that we will get the same class
1677 for the new pseudo and we will get the same insn which
1678 means infinite looping. So spill the new pseudo. */
1679 rclass = NO_REGS;
1680 else
1681 /* The class will be defined later in curr_insn_transform. */
1682 rclass
1683 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1685 if (get_reload_reg (curr_static_id->operand[nop].type, reg_mode, reg,
1686 rclass, TRUE, "subreg reg", &new_reg))
1688 bool insert_before, insert_after;
1689 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1691 insert_before = (type != OP_OUT
1692 || read_modify_subreg_p (operand));
1693 insert_after = (type != OP_IN);
1694 insert_move_for_subreg (insert_before ? &before : NULL,
1695 insert_after ? &after : NULL,
1696 reg, new_reg);
1698 SUBREG_REG (operand) = new_reg;
1699 lra_process_new_insns (curr_insn, before, after,
1700 "Inserting subreg reload");
1701 return true;
1703 /* Force a reload for a paradoxical subreg. For paradoxical subreg,
1704 IRA allocates hardreg to the inner pseudo reg according to its mode
1705 instead of the outermode, so the size of the hardreg may not be enough
1706 to contain the outermode operand, in that case we may need to insert
1707 reload for the reg. For the following two types of paradoxical subreg,
1708 we need to insert reload:
1709 1. If the op_type is OP_IN, and the hardreg could not be paired with
1710 other hardreg to contain the outermode operand
1711 (checked by in_hard_reg_set_p), we need to insert the reload.
1712 2. If the op_type is OP_OUT or OP_INOUT.
1714 Here is a paradoxical subreg example showing how the reload is generated:
1716 (insn 5 4 7 2 (set (reg:TI 106 [ __comp ])
1717 (subreg:TI (reg:DI 107 [ __comp ]) 0)) {*movti_internal_rex64}
1719 In IRA, reg107 is allocated to a DImode hardreg. We use x86-64 as example
1720 here, if reg107 is assigned to hardreg R15, because R15 is the last
1721 hardreg, compiler cannot find another hardreg to pair with R15 to
1722 contain TImode data. So we insert a TImode reload reg180 for it.
1723 After reload is inserted:
1725 (insn 283 0 0 (set (subreg:DI (reg:TI 180 [orig:107 __comp ] [107]) 0)
1726 (reg:DI 107 [ __comp ])) -1
1727 (insn 5 4 7 2 (set (reg:TI 106 [ __comp ])
1728 (subreg:TI (reg:TI 180 [orig:107 __comp ] [107]) 0)) {*movti_internal_rex64}
1730 Two reload hard registers will be allocated to reg180 to save TImode data
1731 in LRA_assign. */
1732 else if (REG_P (reg)
1733 && REGNO (reg) >= FIRST_PSEUDO_REGISTER
1734 && (hard_regno = lra_get_regno_hard_regno (REGNO (reg))) >= 0
1735 && (hard_regno_nregs (hard_regno, innermode)
1736 < hard_regno_nregs (hard_regno, mode))
1737 && (regclass = lra_get_allocno_class (REGNO (reg)))
1738 && (type != OP_IN
1739 || !in_hard_reg_set_p (reg_class_contents[regclass],
1740 mode, hard_regno)))
1742 /* The class will be defined later in curr_insn_transform. */
1743 enum reg_class rclass
1744 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1746 if (get_reload_reg (curr_static_id->operand[nop].type, mode, reg,
1747 rclass, TRUE, "paradoxical subreg", &new_reg))
1749 rtx subreg;
1750 bool insert_before, insert_after;
1752 PUT_MODE (new_reg, mode);
1753 subreg = gen_lowpart_SUBREG (innermode, new_reg);
1754 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1756 insert_before = (type != OP_OUT);
1757 insert_after = (type != OP_IN);
1758 insert_move_for_subreg (insert_before ? &before : NULL,
1759 insert_after ? &after : NULL,
1760 reg, subreg);
1762 SUBREG_REG (operand) = new_reg;
1763 lra_process_new_insns (curr_insn, before, after,
1764 "Inserting paradoxical subreg reload");
1765 return true;
1767 return false;
1770 /* Return TRUE if X refers for a hard register from SET. */
1771 static bool
1772 uses_hard_regs_p (rtx x, HARD_REG_SET set)
1774 int i, j, x_hard_regno;
1775 machine_mode mode;
1776 const char *fmt;
1777 enum rtx_code code;
1779 if (x == NULL_RTX)
1780 return false;
1781 code = GET_CODE (x);
1782 mode = GET_MODE (x);
1783 if (code == SUBREG)
1785 mode = wider_subreg_mode (x);
1786 x = SUBREG_REG (x);
1787 code = GET_CODE (x);
1790 if (REG_P (x))
1792 x_hard_regno = get_hard_regno (x, true);
1793 return (x_hard_regno >= 0
1794 && overlaps_hard_reg_set_p (set, mode, x_hard_regno));
1796 if (MEM_P (x))
1798 struct address_info ad;
1800 decompose_mem_address (&ad, x);
1801 if (ad.base_term != NULL && uses_hard_regs_p (*ad.base_term, set))
1802 return true;
1803 if (ad.index_term != NULL && uses_hard_regs_p (*ad.index_term, set))
1804 return true;
1806 fmt = GET_RTX_FORMAT (code);
1807 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1809 if (fmt[i] == 'e')
1811 if (uses_hard_regs_p (XEXP (x, i), set))
1812 return true;
1814 else if (fmt[i] == 'E')
1816 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
1817 if (uses_hard_regs_p (XVECEXP (x, i, j), set))
1818 return true;
1821 return false;
1824 /* Return true if OP is a spilled pseudo. */
1825 static inline bool
1826 spilled_pseudo_p (rtx op)
1828 return (REG_P (op)
1829 && REGNO (op) >= FIRST_PSEUDO_REGISTER && in_mem_p (REGNO (op)));
1832 /* Return true if X is a general constant. */
1833 static inline bool
1834 general_constant_p (rtx x)
1836 return CONSTANT_P (x) && (! flag_pic || LEGITIMATE_PIC_OPERAND_P (x));
1839 static bool
1840 reg_in_class_p (rtx reg, enum reg_class cl)
1842 if (cl == NO_REGS)
1843 return get_reg_class (REGNO (reg)) == NO_REGS;
1844 return in_class_p (reg, cl, NULL);
1847 /* Return true if SET of RCLASS contains no hard regs which can be
1848 used in MODE. */
1849 static bool
1850 prohibited_class_reg_set_mode_p (enum reg_class rclass,
1851 HARD_REG_SET &set,
1852 machine_mode mode)
1854 HARD_REG_SET temp;
1856 lra_assert (hard_reg_set_subset_p (reg_class_contents[rclass], set));
1857 COPY_HARD_REG_SET (temp, set);
1858 AND_COMPL_HARD_REG_SET (temp, lra_no_alloc_regs);
1859 return (hard_reg_set_subset_p
1860 (temp, ira_prohibited_class_mode_regs[rclass][mode]));
1864 /* Used to check validity info about small class input operands. It
1865 should be incremented at start of processing an insn
1866 alternative. */
1867 static unsigned int curr_small_class_check = 0;
1869 /* Update number of used inputs of class OP_CLASS for operand NOP.
1870 Return true if we have more such class operands than the number of
1871 available regs. */
1872 static bool
1873 update_and_check_small_class_inputs (int nop, enum reg_class op_class)
1875 static unsigned int small_class_check[LIM_REG_CLASSES];
1876 static int small_class_input_nums[LIM_REG_CLASSES];
1878 if (SMALL_REGISTER_CLASS_P (op_class)
1879 /* We are interesting in classes became small because of fixing
1880 some hard regs, e.g. by an user through GCC options. */
1881 && hard_reg_set_intersect_p (reg_class_contents[op_class],
1882 ira_no_alloc_regs)
1883 && (curr_static_id->operand[nop].type != OP_OUT
1884 || curr_static_id->operand[nop].early_clobber))
1886 if (small_class_check[op_class] == curr_small_class_check)
1887 small_class_input_nums[op_class]++;
1888 else
1890 small_class_check[op_class] = curr_small_class_check;
1891 small_class_input_nums[op_class] = 1;
1893 if (small_class_input_nums[op_class] > ira_class_hard_regs_num[op_class])
1894 return true;
1896 return false;
1899 /* Major function to choose the current insn alternative and what
1900 operands should be reloaded and how. If ONLY_ALTERNATIVE is not
1901 negative we should consider only this alternative. Return false if
1902 we can not choose the alternative or find how to reload the
1903 operands. */
1904 static bool
1905 process_alt_operands (int only_alternative)
1907 bool ok_p = false;
1908 int nop, overall, nalt;
1909 int n_alternatives = curr_static_id->n_alternatives;
1910 int n_operands = curr_static_id->n_operands;
1911 /* LOSERS counts the operands that don't fit this alternative and
1912 would require loading. */
1913 int losers;
1914 int addr_losers;
1915 /* REJECT is a count of how undesirable this alternative says it is
1916 if any reloading is required. If the alternative matches exactly
1917 then REJECT is ignored, but otherwise it gets this much counted
1918 against it in addition to the reloading needed. */
1919 int reject;
1920 /* This is defined by '!' or '?' alternative constraint and added to
1921 reject. But in some cases it can be ignored. */
1922 int static_reject;
1923 int op_reject;
1924 /* The number of elements in the following array. */
1925 int early_clobbered_regs_num;
1926 /* Numbers of operands which are early clobber registers. */
1927 int early_clobbered_nops[MAX_RECOG_OPERANDS];
1928 enum reg_class curr_alt[MAX_RECOG_OPERANDS];
1929 HARD_REG_SET curr_alt_set[MAX_RECOG_OPERANDS];
1930 bool curr_alt_match_win[MAX_RECOG_OPERANDS];
1931 bool curr_alt_win[MAX_RECOG_OPERANDS];
1932 bool curr_alt_offmemok[MAX_RECOG_OPERANDS];
1933 int curr_alt_matches[MAX_RECOG_OPERANDS];
1934 /* The number of elements in the following array. */
1935 int curr_alt_dont_inherit_ops_num;
1936 /* Numbers of operands whose reload pseudos should not be inherited. */
1937 int curr_alt_dont_inherit_ops[MAX_RECOG_OPERANDS];
1938 rtx op;
1939 /* The register when the operand is a subreg of register, otherwise the
1940 operand itself. */
1941 rtx no_subreg_reg_operand[MAX_RECOG_OPERANDS];
1942 /* The register if the operand is a register or subreg of register,
1943 otherwise NULL. */
1944 rtx operand_reg[MAX_RECOG_OPERANDS];
1945 int hard_regno[MAX_RECOG_OPERANDS];
1946 machine_mode biggest_mode[MAX_RECOG_OPERANDS];
1947 int reload_nregs, reload_sum;
1948 bool costly_p;
1949 enum reg_class cl;
1951 /* Calculate some data common for all alternatives to speed up the
1952 function. */
1953 for (nop = 0; nop < n_operands; nop++)
1955 rtx reg;
1957 op = no_subreg_reg_operand[nop] = *curr_id->operand_loc[nop];
1958 /* The real hard regno of the operand after the allocation. */
1959 hard_regno[nop] = get_hard_regno (op, true);
1961 operand_reg[nop] = reg = op;
1962 biggest_mode[nop] = GET_MODE (op);
1963 if (GET_CODE (op) == SUBREG)
1965 biggest_mode[nop] = wider_subreg_mode (op);
1966 operand_reg[nop] = reg = SUBREG_REG (op);
1968 if (! REG_P (reg))
1969 operand_reg[nop] = NULL_RTX;
1970 else if (REGNO (reg) >= FIRST_PSEUDO_REGISTER
1971 || ((int) REGNO (reg)
1972 == lra_get_elimination_hard_regno (REGNO (reg))))
1973 no_subreg_reg_operand[nop] = reg;
1974 else
1975 operand_reg[nop] = no_subreg_reg_operand[nop]
1976 /* Just use natural mode for elimination result. It should
1977 be enough for extra constraints hooks. */
1978 = regno_reg_rtx[hard_regno[nop]];
1981 /* The constraints are made of several alternatives. Each operand's
1982 constraint looks like foo,bar,... with commas separating the
1983 alternatives. The first alternatives for all operands go
1984 together, the second alternatives go together, etc.
1986 First loop over alternatives. */
1987 alternative_mask preferred = curr_id->preferred_alternatives;
1988 if (only_alternative >= 0)
1989 preferred &= ALTERNATIVE_BIT (only_alternative);
1991 for (nalt = 0; nalt < n_alternatives; nalt++)
1993 /* Loop over operands for one constraint alternative. */
1994 if (!TEST_BIT (preferred, nalt))
1995 continue;
1997 curr_small_class_check++;
1998 overall = losers = addr_losers = 0;
1999 static_reject = reject = reload_nregs = reload_sum = 0;
2000 for (nop = 0; nop < n_operands; nop++)
2002 int inc = (curr_static_id
2003 ->operand_alternative[nalt * n_operands + nop].reject);
2004 if (lra_dump_file != NULL && inc != 0)
2005 fprintf (lra_dump_file,
2006 " Staticly defined alt reject+=%d\n", inc);
2007 static_reject += inc;
2009 reject += static_reject;
2010 early_clobbered_regs_num = 0;
2012 for (nop = 0; nop < n_operands; nop++)
2014 const char *p;
2015 char *end;
2016 int len, c, m, i, opalt_num, this_alternative_matches;
2017 bool win, did_match, offmemok, early_clobber_p;
2018 /* false => this operand can be reloaded somehow for this
2019 alternative. */
2020 bool badop;
2021 /* true => this operand can be reloaded if the alternative
2022 allows regs. */
2023 bool winreg;
2024 /* True if a constant forced into memory would be OK for
2025 this operand. */
2026 bool constmemok;
2027 enum reg_class this_alternative, this_costly_alternative;
2028 HARD_REG_SET this_alternative_set, this_costly_alternative_set;
2029 bool this_alternative_match_win, this_alternative_win;
2030 bool this_alternative_offmemok;
2031 bool scratch_p;
2032 machine_mode mode;
2033 enum constraint_num cn;
2035 opalt_num = nalt * n_operands + nop;
2036 if (curr_static_id->operand_alternative[opalt_num].anything_ok)
2038 /* Fast track for no constraints at all. */
2039 curr_alt[nop] = NO_REGS;
2040 CLEAR_HARD_REG_SET (curr_alt_set[nop]);
2041 curr_alt_win[nop] = true;
2042 curr_alt_match_win[nop] = false;
2043 curr_alt_offmemok[nop] = false;
2044 curr_alt_matches[nop] = -1;
2045 continue;
2048 op = no_subreg_reg_operand[nop];
2049 mode = curr_operand_mode[nop];
2051 win = did_match = winreg = offmemok = constmemok = false;
2052 badop = true;
2054 early_clobber_p = false;
2055 p = curr_static_id->operand_alternative[opalt_num].constraint;
2057 this_costly_alternative = this_alternative = NO_REGS;
2058 /* We update set of possible hard regs besides its class
2059 because reg class might be inaccurate. For example,
2060 union of LO_REGS (l), HI_REGS(h), and STACK_REG(k) in ARM
2061 is translated in HI_REGS because classes are merged by
2062 pairs and there is no accurate intermediate class. */
2063 CLEAR_HARD_REG_SET (this_alternative_set);
2064 CLEAR_HARD_REG_SET (this_costly_alternative_set);
2065 this_alternative_win = false;
2066 this_alternative_match_win = false;
2067 this_alternative_offmemok = false;
2068 this_alternative_matches = -1;
2070 /* An empty constraint should be excluded by the fast
2071 track. */
2072 lra_assert (*p != 0 && *p != ',');
2074 op_reject = 0;
2075 /* Scan this alternative's specs for this operand; set WIN
2076 if the operand fits any letter in this alternative.
2077 Otherwise, clear BADOP if this operand could fit some
2078 letter after reloads, or set WINREG if this operand could
2079 fit after reloads provided the constraint allows some
2080 registers. */
2081 costly_p = false;
2084 switch ((c = *p, len = CONSTRAINT_LEN (c, p)), c)
2086 case '\0':
2087 len = 0;
2088 break;
2089 case ',':
2090 c = '\0';
2091 break;
2093 case '&':
2094 early_clobber_p = true;
2095 break;
2097 case '$':
2098 op_reject += LRA_MAX_REJECT;
2099 break;
2100 case '^':
2101 op_reject += LRA_LOSER_COST_FACTOR;
2102 break;
2104 case '#':
2105 /* Ignore rest of this alternative. */
2106 c = '\0';
2107 break;
2109 case '0': case '1': case '2': case '3': case '4':
2110 case '5': case '6': case '7': case '8': case '9':
2112 int m_hregno;
2113 bool match_p;
2115 m = strtoul (p, &end, 10);
2116 p = end;
2117 len = 0;
2118 lra_assert (nop > m);
2120 /* Reject matches if we don't know which operand is
2121 bigger. This situation would arguably be a bug in
2122 an .md pattern, but could also occur in a user asm. */
2123 if (!ordered_p (GET_MODE_SIZE (biggest_mode[m]),
2124 GET_MODE_SIZE (biggest_mode[nop])))
2125 break;
2127 this_alternative_matches = m;
2128 m_hregno = get_hard_regno (*curr_id->operand_loc[m], false);
2129 /* We are supposed to match a previous operand.
2130 If we do, we win if that one did. If we do
2131 not, count both of the operands as losers.
2132 (This is too conservative, since most of the
2133 time only a single reload insn will be needed
2134 to make the two operands win. As a result,
2135 this alternative may be rejected when it is
2136 actually desirable.) */
2137 match_p = false;
2138 if (operands_match_p (*curr_id->operand_loc[nop],
2139 *curr_id->operand_loc[m], m_hregno))
2141 /* We should reject matching of an early
2142 clobber operand if the matching operand is
2143 not dying in the insn. */
2144 if (! curr_static_id->operand[m].early_clobber
2145 || operand_reg[nop] == NULL_RTX
2146 || (find_regno_note (curr_insn, REG_DEAD,
2147 REGNO (op))
2148 || REGNO (op) == REGNO (operand_reg[m])))
2149 match_p = true;
2151 if (match_p)
2153 /* If we are matching a non-offsettable
2154 address where an offsettable address was
2155 expected, then we must reject this
2156 combination, because we can't reload
2157 it. */
2158 if (curr_alt_offmemok[m]
2159 && MEM_P (*curr_id->operand_loc[m])
2160 && curr_alt[m] == NO_REGS && ! curr_alt_win[m])
2161 continue;
2163 else
2165 /* Operands don't match. Both operands must
2166 allow a reload register, otherwise we
2167 cannot make them match. */
2168 if (curr_alt[m] == NO_REGS)
2169 break;
2170 /* Retroactively mark the operand we had to
2171 match as a loser, if it wasn't already and
2172 it wasn't matched to a register constraint
2173 (e.g it might be matched by memory). */
2174 if (curr_alt_win[m]
2175 && (operand_reg[m] == NULL_RTX
2176 || hard_regno[m] < 0))
2178 losers++;
2179 reload_nregs
2180 += (ira_reg_class_max_nregs[curr_alt[m]]
2181 [GET_MODE (*curr_id->operand_loc[m])]);
2184 /* Prefer matching earlyclobber alternative as
2185 it results in less hard regs required for
2186 the insn than a non-matching earlyclobber
2187 alternative. */
2188 if (curr_static_id->operand[m].early_clobber)
2190 if (lra_dump_file != NULL)
2191 fprintf
2192 (lra_dump_file,
2193 " %d Matching earlyclobber alt:"
2194 " reject--\n",
2195 nop);
2196 reject--;
2198 /* Otherwise we prefer no matching
2199 alternatives because it gives more freedom
2200 in RA. */
2201 else if (operand_reg[nop] == NULL_RTX
2202 || (find_regno_note (curr_insn, REG_DEAD,
2203 REGNO (operand_reg[nop]))
2204 == NULL_RTX))
2206 if (lra_dump_file != NULL)
2207 fprintf
2208 (lra_dump_file,
2209 " %d Matching alt: reject+=2\n",
2210 nop);
2211 reject += 2;
2214 /* If we have to reload this operand and some
2215 previous operand also had to match the same
2216 thing as this operand, we don't know how to do
2217 that. */
2218 if (!match_p || !curr_alt_win[m])
2220 for (i = 0; i < nop; i++)
2221 if (curr_alt_matches[i] == m)
2222 break;
2223 if (i < nop)
2224 break;
2226 else
2227 did_match = true;
2229 /* This can be fixed with reloads if the operand
2230 we are supposed to match can be fixed with
2231 reloads. */
2232 badop = false;
2233 this_alternative = curr_alt[m];
2234 COPY_HARD_REG_SET (this_alternative_set, curr_alt_set[m]);
2235 winreg = this_alternative != NO_REGS;
2236 break;
2239 case 'g':
2240 if (MEM_P (op)
2241 || general_constant_p (op)
2242 || spilled_pseudo_p (op))
2243 win = true;
2244 cl = GENERAL_REGS;
2245 goto reg;
2247 default:
2248 cn = lookup_constraint (p);
2249 switch (get_constraint_type (cn))
2251 case CT_REGISTER:
2252 cl = reg_class_for_constraint (cn);
2253 if (cl != NO_REGS)
2254 goto reg;
2255 break;
2257 case CT_CONST_INT:
2258 if (CONST_INT_P (op)
2259 && insn_const_int_ok_for_constraint (INTVAL (op), cn))
2260 win = true;
2261 break;
2263 case CT_MEMORY:
2264 if (MEM_P (op)
2265 && satisfies_memory_constraint_p (op, cn))
2266 win = true;
2267 else if (spilled_pseudo_p (op))
2268 win = true;
2270 /* If we didn't already win, we can reload constants
2271 via force_const_mem or put the pseudo value into
2272 memory, or make other memory by reloading the
2273 address like for 'o'. */
2274 if (CONST_POOL_OK_P (mode, op)
2275 || MEM_P (op) || REG_P (op)
2276 /* We can restore the equiv insn by a
2277 reload. */
2278 || equiv_substition_p[nop])
2279 badop = false;
2280 constmemok = true;
2281 offmemok = true;
2282 break;
2284 case CT_ADDRESS:
2285 /* If we didn't already win, we can reload the address
2286 into a base register. */
2287 if (satisfies_address_constraint_p (op, cn))
2288 win = true;
2289 cl = base_reg_class (VOIDmode, ADDR_SPACE_GENERIC,
2290 ADDRESS, SCRATCH);
2291 badop = false;
2292 goto reg;
2294 case CT_FIXED_FORM:
2295 if (constraint_satisfied_p (op, cn))
2296 win = true;
2297 break;
2299 case CT_SPECIAL_MEMORY:
2300 if (MEM_P (op)
2301 && satisfies_memory_constraint_p (op, cn))
2302 win = true;
2303 else if (spilled_pseudo_p (op))
2304 win = true;
2305 break;
2307 break;
2309 reg:
2310 this_alternative = reg_class_subunion[this_alternative][cl];
2311 IOR_HARD_REG_SET (this_alternative_set,
2312 reg_class_contents[cl]);
2313 if (costly_p)
2315 this_costly_alternative
2316 = reg_class_subunion[this_costly_alternative][cl];
2317 IOR_HARD_REG_SET (this_costly_alternative_set,
2318 reg_class_contents[cl]);
2320 if (mode == BLKmode)
2321 break;
2322 winreg = true;
2323 if (REG_P (op))
2325 if (hard_regno[nop] >= 0
2326 && in_hard_reg_set_p (this_alternative_set,
2327 mode, hard_regno[nop]))
2328 win = true;
2329 else if (hard_regno[nop] < 0
2330 && in_class_p (op, this_alternative, NULL))
2331 win = true;
2333 break;
2335 if (c != ' ' && c != '\t')
2336 costly_p = c == '*';
2338 while ((p += len), c);
2340 scratch_p = (operand_reg[nop] != NULL_RTX
2341 && lra_former_scratch_p (REGNO (operand_reg[nop])));
2342 /* Record which operands fit this alternative. */
2343 if (win)
2345 this_alternative_win = true;
2346 if (operand_reg[nop] != NULL_RTX)
2348 if (hard_regno[nop] >= 0)
2350 if (in_hard_reg_set_p (this_costly_alternative_set,
2351 mode, hard_regno[nop]))
2353 if (lra_dump_file != NULL)
2354 fprintf (lra_dump_file,
2355 " %d Costly set: reject++\n",
2356 nop);
2357 reject++;
2360 else
2362 /* Prefer won reg to spilled pseudo under other
2363 equal conditions for possibe inheritance. */
2364 if (! scratch_p)
2366 if (lra_dump_file != NULL)
2367 fprintf
2368 (lra_dump_file,
2369 " %d Non pseudo reload: reject++\n",
2370 nop);
2371 reject++;
2373 if (in_class_p (operand_reg[nop],
2374 this_costly_alternative, NULL))
2376 if (lra_dump_file != NULL)
2377 fprintf
2378 (lra_dump_file,
2379 " %d Non pseudo costly reload:"
2380 " reject++\n",
2381 nop);
2382 reject++;
2385 /* We simulate the behavior of old reload here.
2386 Although scratches need hard registers and it
2387 might result in spilling other pseudos, no reload
2388 insns are generated for the scratches. So it
2389 might cost something but probably less than old
2390 reload pass believes. */
2391 if (scratch_p)
2393 if (lra_dump_file != NULL)
2394 fprintf (lra_dump_file,
2395 " %d Scratch win: reject+=2\n",
2396 nop);
2397 reject += 2;
2401 else if (did_match)
2402 this_alternative_match_win = true;
2403 else
2405 int const_to_mem = 0;
2406 bool no_regs_p;
2408 reject += op_reject;
2409 /* Never do output reload of stack pointer. It makes
2410 impossible to do elimination when SP is changed in
2411 RTL. */
2412 if (op == stack_pointer_rtx && ! frame_pointer_needed
2413 && curr_static_id->operand[nop].type != OP_IN)
2414 goto fail;
2416 /* If this alternative asks for a specific reg class, see if there
2417 is at least one allocatable register in that class. */
2418 no_regs_p
2419 = (this_alternative == NO_REGS
2420 || (hard_reg_set_subset_p
2421 (reg_class_contents[this_alternative],
2422 lra_no_alloc_regs)));
2424 /* For asms, verify that the class for this alternative is possible
2425 for the mode that is specified. */
2426 if (!no_regs_p && INSN_CODE (curr_insn) < 0)
2428 int i;
2429 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
2430 if (targetm.hard_regno_mode_ok (i, mode)
2431 && in_hard_reg_set_p (reg_class_contents[this_alternative],
2432 mode, i))
2433 break;
2434 if (i == FIRST_PSEUDO_REGISTER)
2435 winreg = false;
2438 /* If this operand accepts a register, and if the
2439 register class has at least one allocatable register,
2440 then this operand can be reloaded. */
2441 if (winreg && !no_regs_p)
2442 badop = false;
2444 if (badop)
2446 if (lra_dump_file != NULL)
2447 fprintf (lra_dump_file,
2448 " alt=%d: Bad operand -- refuse\n",
2449 nalt);
2450 goto fail;
2453 if (this_alternative != NO_REGS)
2455 HARD_REG_SET available_regs;
2457 COPY_HARD_REG_SET (available_regs,
2458 reg_class_contents[this_alternative]);
2459 AND_COMPL_HARD_REG_SET
2460 (available_regs,
2461 ira_prohibited_class_mode_regs[this_alternative][mode]);
2462 AND_COMPL_HARD_REG_SET (available_regs, lra_no_alloc_regs);
2463 if (hard_reg_set_empty_p (available_regs))
2465 /* There are no hard regs holding a value of given
2466 mode. */
2467 if (offmemok)
2469 this_alternative = NO_REGS;
2470 if (lra_dump_file != NULL)
2471 fprintf (lra_dump_file,
2472 " %d Using memory because of"
2473 " a bad mode: reject+=2\n",
2474 nop);
2475 reject += 2;
2477 else
2479 if (lra_dump_file != NULL)
2480 fprintf (lra_dump_file,
2481 " alt=%d: Wrong mode -- refuse\n",
2482 nalt);
2483 goto fail;
2488 /* If not assigned pseudo has a class which a subset of
2489 required reg class, it is a less costly alternative
2490 as the pseudo still can get a hard reg of necessary
2491 class. */
2492 if (! no_regs_p && REG_P (op) && hard_regno[nop] < 0
2493 && (cl = get_reg_class (REGNO (op))) != NO_REGS
2494 && ira_class_subset_p[this_alternative][cl])
2496 if (lra_dump_file != NULL)
2497 fprintf
2498 (lra_dump_file,
2499 " %d Super set class reg: reject-=3\n", nop);
2500 reject -= 3;
2503 this_alternative_offmemok = offmemok;
2504 if (this_costly_alternative != NO_REGS)
2506 if (lra_dump_file != NULL)
2507 fprintf (lra_dump_file,
2508 " %d Costly loser: reject++\n", nop);
2509 reject++;
2511 /* If the operand is dying, has a matching constraint,
2512 and satisfies constraints of the matched operand
2513 which failed to satisfy the own constraints, most probably
2514 the reload for this operand will be gone. */
2515 if (this_alternative_matches >= 0
2516 && !curr_alt_win[this_alternative_matches]
2517 && REG_P (op)
2518 && find_regno_note (curr_insn, REG_DEAD, REGNO (op))
2519 && (hard_regno[nop] >= 0
2520 ? in_hard_reg_set_p (this_alternative_set,
2521 mode, hard_regno[nop])
2522 : in_class_p (op, this_alternative, NULL)))
2524 if (lra_dump_file != NULL)
2525 fprintf
2526 (lra_dump_file,
2527 " %d Dying matched operand reload: reject++\n",
2528 nop);
2529 reject++;
2531 else
2533 /* Strict_low_part requires to reload the register
2534 not the sub-register. In this case we should
2535 check that a final reload hard reg can hold the
2536 value mode. */
2537 if (curr_static_id->operand[nop].strict_low
2538 && REG_P (op)
2539 && hard_regno[nop] < 0
2540 && GET_CODE (*curr_id->operand_loc[nop]) == SUBREG
2541 && ira_class_hard_regs_num[this_alternative] > 0
2542 && (!targetm.hard_regno_mode_ok
2543 (ira_class_hard_regs[this_alternative][0],
2544 GET_MODE (*curr_id->operand_loc[nop]))))
2546 if (lra_dump_file != NULL)
2547 fprintf
2548 (lra_dump_file,
2549 " alt=%d: Strict low subreg reload -- refuse\n",
2550 nalt);
2551 goto fail;
2553 losers++;
2555 if (operand_reg[nop] != NULL_RTX
2556 /* Output operands and matched input operands are
2557 not inherited. The following conditions do not
2558 exactly describe the previous statement but they
2559 are pretty close. */
2560 && curr_static_id->operand[nop].type != OP_OUT
2561 && (this_alternative_matches < 0
2562 || curr_static_id->operand[nop].type != OP_IN))
2564 int last_reload = (lra_reg_info[ORIGINAL_REGNO
2565 (operand_reg[nop])]
2566 .last_reload);
2568 /* The value of reload_sum has sense only if we
2569 process insns in their order. It happens only on
2570 the first constraints sub-pass when we do most of
2571 reload work. */
2572 if (lra_constraint_iter == 1 && last_reload > bb_reload_num)
2573 reload_sum += last_reload - bb_reload_num;
2575 /* If this is a constant that is reloaded into the
2576 desired class by copying it to memory first, count
2577 that as another reload. This is consistent with
2578 other code and is required to avoid choosing another
2579 alternative when the constant is moved into memory.
2580 Note that the test here is precisely the same as in
2581 the code below that calls force_const_mem. */
2582 if (CONST_POOL_OK_P (mode, op)
2583 && ((targetm.preferred_reload_class
2584 (op, this_alternative) == NO_REGS)
2585 || no_input_reloads_p))
2587 const_to_mem = 1;
2588 if (! no_regs_p)
2589 losers++;
2592 /* Alternative loses if it requires a type of reload not
2593 permitted for this insn. We can always reload
2594 objects with a REG_UNUSED note. */
2595 if ((curr_static_id->operand[nop].type != OP_IN
2596 && no_output_reloads_p
2597 && ! find_reg_note (curr_insn, REG_UNUSED, op))
2598 || (curr_static_id->operand[nop].type != OP_OUT
2599 && no_input_reloads_p && ! const_to_mem)
2600 || (this_alternative_matches >= 0
2601 && (no_input_reloads_p
2602 || (no_output_reloads_p
2603 && (curr_static_id->operand
2604 [this_alternative_matches].type != OP_IN)
2605 && ! find_reg_note (curr_insn, REG_UNUSED,
2606 no_subreg_reg_operand
2607 [this_alternative_matches])))))
2609 if (lra_dump_file != NULL)
2610 fprintf
2611 (lra_dump_file,
2612 " alt=%d: No input/otput reload -- refuse\n",
2613 nalt);
2614 goto fail;
2617 /* Alternative loses if it required class pseudo can not
2618 hold value of required mode. Such insns can be
2619 described by insn definitions with mode iterators. */
2620 if (GET_MODE (*curr_id->operand_loc[nop]) != VOIDmode
2621 && ! hard_reg_set_empty_p (this_alternative_set)
2622 /* It is common practice for constraints to use a
2623 class which does not have actually enough regs to
2624 hold the value (e.g. x86 AREG for mode requiring
2625 more one general reg). Therefore we have 2
2626 conditions to check that the reload pseudo can
2627 not hold the mode value. */
2628 && (!targetm.hard_regno_mode_ok
2629 (ira_class_hard_regs[this_alternative][0],
2630 GET_MODE (*curr_id->operand_loc[nop])))
2631 /* The above condition is not enough as the first
2632 reg in ira_class_hard_regs can be not aligned for
2633 multi-words mode values. */
2634 && (prohibited_class_reg_set_mode_p
2635 (this_alternative, this_alternative_set,
2636 GET_MODE (*curr_id->operand_loc[nop]))))
2638 if (lra_dump_file != NULL)
2639 fprintf (lra_dump_file,
2640 " alt=%d: reload pseudo for op %d "
2641 " can not hold the mode value -- refuse\n",
2642 nalt, nop);
2643 goto fail;
2646 /* Check strong discouragement of reload of non-constant
2647 into class THIS_ALTERNATIVE. */
2648 if (! CONSTANT_P (op) && ! no_regs_p
2649 && (targetm.preferred_reload_class
2650 (op, this_alternative) == NO_REGS
2651 || (curr_static_id->operand[nop].type == OP_OUT
2652 && (targetm.preferred_output_reload_class
2653 (op, this_alternative) == NO_REGS))))
2655 if (lra_dump_file != NULL)
2656 fprintf (lra_dump_file,
2657 " %d Non-prefered reload: reject+=%d\n",
2658 nop, LRA_MAX_REJECT);
2659 reject += LRA_MAX_REJECT;
2662 if (! (MEM_P (op) && offmemok)
2663 && ! (const_to_mem && constmemok))
2665 /* We prefer to reload pseudos over reloading other
2666 things, since such reloads may be able to be
2667 eliminated later. So bump REJECT in other cases.
2668 Don't do this in the case where we are forcing a
2669 constant into memory and it will then win since
2670 we don't want to have a different alternative
2671 match then. */
2672 if (! (REG_P (op) && REGNO (op) >= FIRST_PSEUDO_REGISTER))
2674 if (lra_dump_file != NULL)
2675 fprintf
2676 (lra_dump_file,
2677 " %d Non-pseudo reload: reject+=2\n",
2678 nop);
2679 reject += 2;
2682 if (! no_regs_p)
2683 reload_nregs
2684 += ira_reg_class_max_nregs[this_alternative][mode];
2686 if (SMALL_REGISTER_CLASS_P (this_alternative))
2688 if (lra_dump_file != NULL)
2689 fprintf
2690 (lra_dump_file,
2691 " %d Small class reload: reject+=%d\n",
2692 nop, LRA_LOSER_COST_FACTOR / 2);
2693 reject += LRA_LOSER_COST_FACTOR / 2;
2697 /* We are trying to spill pseudo into memory. It is
2698 usually more costly than moving to a hard register
2699 although it might takes the same number of
2700 reloads.
2702 Non-pseudo spill may happen also. Suppose a target allows both
2703 register and memory in the operand constraint alternatives,
2704 then it's typical that an eliminable register has a substition
2705 of "base + offset" which can either be reloaded by a simple
2706 "new_reg <= base + offset" which will match the register
2707 constraint, or a similar reg addition followed by further spill
2708 to and reload from memory which will match the memory
2709 constraint, but this memory spill will be much more costly
2710 usually.
2712 Code below increases the reject for both pseudo and non-pseudo
2713 spill. */
2714 if (no_regs_p
2715 && !(MEM_P (op) && offmemok)
2716 && !(REG_P (op) && hard_regno[nop] < 0))
2718 if (lra_dump_file != NULL)
2719 fprintf
2720 (lra_dump_file,
2721 " %d Spill %spseudo into memory: reject+=3\n",
2722 nop, REG_P (op) ? "" : "Non-");
2723 reject += 3;
2724 if (VECTOR_MODE_P (mode))
2726 /* Spilling vectors into memory is usually more
2727 costly as they contain big values. */
2728 if (lra_dump_file != NULL)
2729 fprintf
2730 (lra_dump_file,
2731 " %d Spill vector pseudo: reject+=2\n",
2732 nop);
2733 reject += 2;
2737 /* When we use an operand requiring memory in given
2738 alternative, the insn should write *and* read the
2739 value to/from memory it is costly in comparison with
2740 an insn alternative which does not use memory
2741 (e.g. register or immediate operand). We exclude
2742 memory operand for such case as we can satisfy the
2743 memory constraints by reloading address. */
2744 if (no_regs_p && offmemok && !MEM_P (op))
2746 if (lra_dump_file != NULL)
2747 fprintf
2748 (lra_dump_file,
2749 " Using memory insn operand %d: reject+=3\n",
2750 nop);
2751 reject += 3;
2754 /* If reload requires moving value through secondary
2755 memory, it will need one more insn at least. */
2756 if (this_alternative != NO_REGS
2757 && REG_P (op) && (cl = get_reg_class (REGNO (op))) != NO_REGS
2758 && ((curr_static_id->operand[nop].type != OP_OUT
2759 && targetm.secondary_memory_needed (GET_MODE (op), cl,
2760 this_alternative))
2761 || (curr_static_id->operand[nop].type != OP_IN
2762 && (targetm.secondary_memory_needed
2763 (GET_MODE (op), this_alternative, cl)))))
2764 losers++;
2766 /* Input reloads can be inherited more often than output
2767 reloads can be removed, so penalize output
2768 reloads. */
2769 if (!REG_P (op) || curr_static_id->operand[nop].type != OP_IN)
2771 if (lra_dump_file != NULL)
2772 fprintf
2773 (lra_dump_file,
2774 " %d Non input pseudo reload: reject++\n",
2775 nop);
2776 reject++;
2779 if (MEM_P (op) && offmemok)
2780 addr_losers++;
2781 else if (curr_static_id->operand[nop].type == OP_INOUT)
2783 if (lra_dump_file != NULL)
2784 fprintf
2785 (lra_dump_file,
2786 " %d Input/Output reload: reject+=%d\n",
2787 nop, LRA_LOSER_COST_FACTOR);
2788 reject += LRA_LOSER_COST_FACTOR;
2792 if (early_clobber_p && ! scratch_p)
2794 if (lra_dump_file != NULL)
2795 fprintf (lra_dump_file,
2796 " %d Early clobber: reject++\n", nop);
2797 reject++;
2799 /* ??? We check early clobbers after processing all operands
2800 (see loop below) and there we update the costs more.
2801 Should we update the cost (may be approximately) here
2802 because of early clobber register reloads or it is a rare
2803 or non-important thing to be worth to do it. */
2804 overall = (losers * LRA_LOSER_COST_FACTOR + reject
2805 - (addr_losers == losers ? static_reject : 0));
2806 if ((best_losers == 0 || losers != 0) && best_overall < overall)
2808 if (lra_dump_file != NULL)
2809 fprintf (lra_dump_file,
2810 " alt=%d,overall=%d,losers=%d -- refuse\n",
2811 nalt, overall, losers);
2812 goto fail;
2815 if (update_and_check_small_class_inputs (nop, this_alternative))
2817 if (lra_dump_file != NULL)
2818 fprintf (lra_dump_file,
2819 " alt=%d, not enough small class regs -- refuse\n",
2820 nalt);
2821 goto fail;
2823 curr_alt[nop] = this_alternative;
2824 COPY_HARD_REG_SET (curr_alt_set[nop], this_alternative_set);
2825 curr_alt_win[nop] = this_alternative_win;
2826 curr_alt_match_win[nop] = this_alternative_match_win;
2827 curr_alt_offmemok[nop] = this_alternative_offmemok;
2828 curr_alt_matches[nop] = this_alternative_matches;
2830 if (this_alternative_matches >= 0
2831 && !did_match && !this_alternative_win)
2832 curr_alt_win[this_alternative_matches] = false;
2834 if (early_clobber_p && operand_reg[nop] != NULL_RTX)
2835 early_clobbered_nops[early_clobbered_regs_num++] = nop;
2838 if (curr_insn_set != NULL_RTX && n_operands == 2
2839 /* Prevent processing non-move insns. */
2840 && (GET_CODE (SET_SRC (curr_insn_set)) == SUBREG
2841 || SET_SRC (curr_insn_set) == no_subreg_reg_operand[1])
2842 && ((! curr_alt_win[0] && ! curr_alt_win[1]
2843 && REG_P (no_subreg_reg_operand[0])
2844 && REG_P (no_subreg_reg_operand[1])
2845 && (reg_in_class_p (no_subreg_reg_operand[0], curr_alt[1])
2846 || reg_in_class_p (no_subreg_reg_operand[1], curr_alt[0])))
2847 || (! curr_alt_win[0] && curr_alt_win[1]
2848 && REG_P (no_subreg_reg_operand[1])
2849 /* Check that we reload memory not the memory
2850 address. */
2851 && ! (curr_alt_offmemok[0]
2852 && MEM_P (no_subreg_reg_operand[0]))
2853 && reg_in_class_p (no_subreg_reg_operand[1], curr_alt[0]))
2854 || (curr_alt_win[0] && ! curr_alt_win[1]
2855 && REG_P (no_subreg_reg_operand[0])
2856 /* Check that we reload memory not the memory
2857 address. */
2858 && ! (curr_alt_offmemok[1]
2859 && MEM_P (no_subreg_reg_operand[1]))
2860 && reg_in_class_p (no_subreg_reg_operand[0], curr_alt[1])
2861 && (! CONST_POOL_OK_P (curr_operand_mode[1],
2862 no_subreg_reg_operand[1])
2863 || (targetm.preferred_reload_class
2864 (no_subreg_reg_operand[1],
2865 (enum reg_class) curr_alt[1]) != NO_REGS))
2866 /* If it is a result of recent elimination in move
2867 insn we can transform it into an add still by
2868 using this alternative. */
2869 && GET_CODE (no_subreg_reg_operand[1]) != PLUS)))
2871 /* We have a move insn and a new reload insn will be similar
2872 to the current insn. We should avoid such situation as
2873 it results in LRA cycling. */
2874 if (lra_dump_file != NULL)
2875 fprintf (lra_dump_file,
2876 " Cycle danger: overall += LRA_MAX_REJECT\n");
2877 overall += LRA_MAX_REJECT;
2879 ok_p = true;
2880 curr_alt_dont_inherit_ops_num = 0;
2881 for (nop = 0; nop < early_clobbered_regs_num; nop++)
2883 int i, j, clobbered_hard_regno, first_conflict_j, last_conflict_j;
2884 HARD_REG_SET temp_set;
2886 i = early_clobbered_nops[nop];
2887 if ((! curr_alt_win[i] && ! curr_alt_match_win[i])
2888 || hard_regno[i] < 0)
2889 continue;
2890 lra_assert (operand_reg[i] != NULL_RTX);
2891 clobbered_hard_regno = hard_regno[i];
2892 CLEAR_HARD_REG_SET (temp_set);
2893 add_to_hard_reg_set (&temp_set, biggest_mode[i], clobbered_hard_regno);
2894 first_conflict_j = last_conflict_j = -1;
2895 for (j = 0; j < n_operands; j++)
2896 if (j == i
2897 /* We don't want process insides of match_operator and
2898 match_parallel because otherwise we would process
2899 their operands once again generating a wrong
2900 code. */
2901 || curr_static_id->operand[j].is_operator)
2902 continue;
2903 else if ((curr_alt_matches[j] == i && curr_alt_match_win[j])
2904 || (curr_alt_matches[i] == j && curr_alt_match_win[i]))
2905 continue;
2906 /* If we don't reload j-th operand, check conflicts. */
2907 else if ((curr_alt_win[j] || curr_alt_match_win[j])
2908 && uses_hard_regs_p (*curr_id->operand_loc[j], temp_set))
2910 if (first_conflict_j < 0)
2911 first_conflict_j = j;
2912 last_conflict_j = j;
2914 if (last_conflict_j < 0)
2915 continue;
2916 /* If earlyclobber operand conflicts with another
2917 non-matching operand which is actually the same register
2918 as the earlyclobber operand, it is better to reload the
2919 another operand as an operand matching the earlyclobber
2920 operand can be also the same. */
2921 if (first_conflict_j == last_conflict_j
2922 && operand_reg[last_conflict_j] != NULL_RTX
2923 && ! curr_alt_match_win[last_conflict_j]
2924 && REGNO (operand_reg[i]) == REGNO (operand_reg[last_conflict_j]))
2926 curr_alt_win[last_conflict_j] = false;
2927 curr_alt_dont_inherit_ops[curr_alt_dont_inherit_ops_num++]
2928 = last_conflict_j;
2929 losers++;
2930 /* Early clobber was already reflected in REJECT. */
2931 lra_assert (reject > 0);
2932 if (lra_dump_file != NULL)
2933 fprintf
2934 (lra_dump_file,
2935 " %d Conflict early clobber reload: reject--\n",
2937 reject--;
2938 overall += LRA_LOSER_COST_FACTOR - 1;
2940 else
2942 /* We need to reload early clobbered register and the
2943 matched registers. */
2944 for (j = 0; j < n_operands; j++)
2945 if (curr_alt_matches[j] == i)
2947 curr_alt_match_win[j] = false;
2948 losers++;
2949 overall += LRA_LOSER_COST_FACTOR;
2951 if (! curr_alt_match_win[i])
2952 curr_alt_dont_inherit_ops[curr_alt_dont_inherit_ops_num++] = i;
2953 else
2955 /* Remember pseudos used for match reloads are never
2956 inherited. */
2957 lra_assert (curr_alt_matches[i] >= 0);
2958 curr_alt_win[curr_alt_matches[i]] = false;
2960 curr_alt_win[i] = curr_alt_match_win[i] = false;
2961 losers++;
2962 /* Early clobber was already reflected in REJECT. */
2963 lra_assert (reject > 0);
2964 if (lra_dump_file != NULL)
2965 fprintf
2966 (lra_dump_file,
2967 " %d Matched conflict early clobber reloads: "
2968 "reject--\n",
2970 reject--;
2971 overall += LRA_LOSER_COST_FACTOR - 1;
2974 if (lra_dump_file != NULL)
2975 fprintf (lra_dump_file, " alt=%d,overall=%d,losers=%d,rld_nregs=%d\n",
2976 nalt, overall, losers, reload_nregs);
2978 /* If this alternative can be made to work by reloading, and it
2979 needs less reloading than the others checked so far, record
2980 it as the chosen goal for reloading. */
2981 if ((best_losers != 0 && losers == 0)
2982 || (((best_losers == 0 && losers == 0)
2983 || (best_losers != 0 && losers != 0))
2984 && (best_overall > overall
2985 || (best_overall == overall
2986 /* If the cost of the reloads is the same,
2987 prefer alternative which requires minimal
2988 number of reload regs. */
2989 && (reload_nregs < best_reload_nregs
2990 || (reload_nregs == best_reload_nregs
2991 && (best_reload_sum < reload_sum
2992 || (best_reload_sum == reload_sum
2993 && nalt < goal_alt_number))))))))
2995 for (nop = 0; nop < n_operands; nop++)
2997 goal_alt_win[nop] = curr_alt_win[nop];
2998 goal_alt_match_win[nop] = curr_alt_match_win[nop];
2999 goal_alt_matches[nop] = curr_alt_matches[nop];
3000 goal_alt[nop] = curr_alt[nop];
3001 goal_alt_offmemok[nop] = curr_alt_offmemok[nop];
3003 goal_alt_dont_inherit_ops_num = curr_alt_dont_inherit_ops_num;
3004 for (nop = 0; nop < curr_alt_dont_inherit_ops_num; nop++)
3005 goal_alt_dont_inherit_ops[nop] = curr_alt_dont_inherit_ops[nop];
3006 goal_alt_swapped = curr_swapped;
3007 best_overall = overall;
3008 best_losers = losers;
3009 best_reload_nregs = reload_nregs;
3010 best_reload_sum = reload_sum;
3011 goal_alt_number = nalt;
3013 if (losers == 0)
3014 /* Everything is satisfied. Do not process alternatives
3015 anymore. */
3016 break;
3017 fail:
3020 return ok_p;
3023 /* Make reload base reg from address AD. */
3024 static rtx
3025 base_to_reg (struct address_info *ad)
3027 enum reg_class cl;
3028 int code = -1;
3029 rtx new_inner = NULL_RTX;
3030 rtx new_reg = NULL_RTX;
3031 rtx_insn *insn;
3032 rtx_insn *last_insn = get_last_insn();
3034 lra_assert (ad->disp == ad->disp_term);
3035 cl = base_reg_class (ad->mode, ad->as, ad->base_outer_code,
3036 get_index_code (ad));
3037 new_reg = lra_create_new_reg (GET_MODE (*ad->base), NULL_RTX,
3038 cl, "base");
3039 new_inner = simplify_gen_binary (PLUS, GET_MODE (new_reg), new_reg,
3040 ad->disp_term == NULL
3041 ? const0_rtx
3042 : *ad->disp_term);
3043 if (!valid_address_p (ad->mode, new_inner, ad->as))
3044 return NULL_RTX;
3045 insn = emit_insn (gen_rtx_SET (new_reg, *ad->base));
3046 code = recog_memoized (insn);
3047 if (code < 0)
3049 delete_insns_since (last_insn);
3050 return NULL_RTX;
3053 return new_inner;
3056 /* Make reload base reg + disp from address AD. Return the new pseudo. */
3057 static rtx
3058 base_plus_disp_to_reg (struct address_info *ad)
3060 enum reg_class cl;
3061 rtx new_reg;
3063 lra_assert (ad->base == ad->base_term && ad->disp == ad->disp_term);
3064 cl = base_reg_class (ad->mode, ad->as, ad->base_outer_code,
3065 get_index_code (ad));
3066 new_reg = lra_create_new_reg (GET_MODE (*ad->base_term), NULL_RTX,
3067 cl, "base + disp");
3068 lra_emit_add (new_reg, *ad->base_term, *ad->disp_term);
3069 return new_reg;
3072 /* Make reload of index part of address AD. Return the new
3073 pseudo. */
3074 static rtx
3075 index_part_to_reg (struct address_info *ad)
3077 rtx new_reg;
3079 new_reg = lra_create_new_reg (GET_MODE (*ad->index), NULL_RTX,
3080 INDEX_REG_CLASS, "index term");
3081 expand_mult (GET_MODE (*ad->index), *ad->index_term,
3082 GEN_INT (get_index_scale (ad)), new_reg, 1);
3083 return new_reg;
3086 /* Return true if we can add a displacement to address AD, even if that
3087 makes the address invalid. The fix-up code requires any new address
3088 to be the sum of the BASE_TERM, INDEX and DISP_TERM fields. */
3089 static bool
3090 can_add_disp_p (struct address_info *ad)
3092 return (!ad->autoinc_p
3093 && ad->segment == NULL
3094 && ad->base == ad->base_term
3095 && ad->disp == ad->disp_term);
3098 /* Make equiv substitution in address AD. Return true if a substitution
3099 was made. */
3100 static bool
3101 equiv_address_substitution (struct address_info *ad)
3103 rtx base_reg, new_base_reg, index_reg, new_index_reg, *base_term, *index_term;
3104 poly_int64 disp;
3105 HOST_WIDE_INT scale;
3106 bool change_p;
3108 base_term = strip_subreg (ad->base_term);
3109 if (base_term == NULL)
3110 base_reg = new_base_reg = NULL_RTX;
3111 else
3113 base_reg = *base_term;
3114 new_base_reg = get_equiv_with_elimination (base_reg, curr_insn);
3116 index_term = strip_subreg (ad->index_term);
3117 if (index_term == NULL)
3118 index_reg = new_index_reg = NULL_RTX;
3119 else
3121 index_reg = *index_term;
3122 new_index_reg = get_equiv_with_elimination (index_reg, curr_insn);
3124 if (base_reg == new_base_reg && index_reg == new_index_reg)
3125 return false;
3126 disp = 0;
3127 change_p = false;
3128 if (lra_dump_file != NULL)
3130 fprintf (lra_dump_file, "Changing address in insn %d ",
3131 INSN_UID (curr_insn));
3132 dump_value_slim (lra_dump_file, *ad->outer, 1);
3134 if (base_reg != new_base_reg)
3136 poly_int64 offset;
3137 if (REG_P (new_base_reg))
3139 *base_term = new_base_reg;
3140 change_p = true;
3142 else if (GET_CODE (new_base_reg) == PLUS
3143 && REG_P (XEXP (new_base_reg, 0))
3144 && poly_int_rtx_p (XEXP (new_base_reg, 1), &offset)
3145 && can_add_disp_p (ad))
3147 disp += offset;
3148 *base_term = XEXP (new_base_reg, 0);
3149 change_p = true;
3151 if (ad->base_term2 != NULL)
3152 *ad->base_term2 = *ad->base_term;
3154 if (index_reg != new_index_reg)
3156 poly_int64 offset;
3157 if (REG_P (new_index_reg))
3159 *index_term = new_index_reg;
3160 change_p = true;
3162 else if (GET_CODE (new_index_reg) == PLUS
3163 && REG_P (XEXP (new_index_reg, 0))
3164 && poly_int_rtx_p (XEXP (new_index_reg, 1), &offset)
3165 && can_add_disp_p (ad)
3166 && (scale = get_index_scale (ad)))
3168 disp += offset * scale;
3169 *index_term = XEXP (new_index_reg, 0);
3170 change_p = true;
3173 if (maybe_ne (disp, 0))
3175 if (ad->disp != NULL)
3176 *ad->disp = plus_constant (GET_MODE (*ad->inner), *ad->disp, disp);
3177 else
3179 *ad->inner = plus_constant (GET_MODE (*ad->inner), *ad->inner, disp);
3180 update_address (ad);
3182 change_p = true;
3184 if (lra_dump_file != NULL)
3186 if (! change_p)
3187 fprintf (lra_dump_file, " -- no change\n");
3188 else
3190 fprintf (lra_dump_file, " on equiv ");
3191 dump_value_slim (lra_dump_file, *ad->outer, 1);
3192 fprintf (lra_dump_file, "\n");
3195 return change_p;
3198 /* Major function to make reloads for an address in operand NOP or
3199 check its correctness (If CHECK_ONLY_P is true). The supported
3200 cases are:
3202 1) an address that existed before LRA started, at which point it
3203 must have been valid. These addresses are subject to elimination
3204 and may have become invalid due to the elimination offset being out
3205 of range.
3207 2) an address created by forcing a constant to memory
3208 (force_const_to_mem). The initial form of these addresses might
3209 not be valid, and it is this function's job to make them valid.
3211 3) a frame address formed from a register and a (possibly zero)
3212 constant offset. As above, these addresses might not be valid and
3213 this function must make them so.
3215 Add reloads to the lists *BEFORE and *AFTER. We might need to add
3216 reloads to *AFTER because of inc/dec, {pre, post} modify in the
3217 address. Return true for any RTL change.
3219 The function is a helper function which does not produce all
3220 transformations (when CHECK_ONLY_P is false) which can be
3221 necessary. It does just basic steps. To do all necessary
3222 transformations use function process_address. */
3223 static bool
3224 process_address_1 (int nop, bool check_only_p,
3225 rtx_insn **before, rtx_insn **after)
3227 struct address_info ad;
3228 rtx new_reg;
3229 HOST_WIDE_INT scale;
3230 rtx op = *curr_id->operand_loc[nop];
3231 const char *constraint = curr_static_id->operand[nop].constraint;
3232 enum constraint_num cn = lookup_constraint (constraint);
3233 bool change_p = false;
3235 if (MEM_P (op)
3236 && GET_MODE (op) == BLKmode
3237 && GET_CODE (XEXP (op, 0)) == SCRATCH)
3238 return false;
3240 if (insn_extra_address_constraint (cn))
3241 decompose_lea_address (&ad, curr_id->operand_loc[nop]);
3242 /* Do not attempt to decompose arbitrary addresses generated by combine
3243 for asm operands with loose constraints, e.g 'X'. */
3244 else if (MEM_P (op)
3245 && !(INSN_CODE (curr_insn) < 0
3246 && get_constraint_type (cn) == CT_FIXED_FORM
3247 && constraint_satisfied_p (op, cn)))
3248 decompose_mem_address (&ad, op);
3249 else if (GET_CODE (op) == SUBREG
3250 && MEM_P (SUBREG_REG (op)))
3251 decompose_mem_address (&ad, SUBREG_REG (op));
3252 else
3253 return false;
3254 /* If INDEX_REG_CLASS is assigned to base_term already and isn't to
3255 index_term, swap them so to avoid assigning INDEX_REG_CLASS to both
3256 when INDEX_REG_CLASS is a single register class. */
3257 if (ad.base_term != NULL
3258 && ad.index_term != NULL
3259 && ira_class_hard_regs_num[INDEX_REG_CLASS] == 1
3260 && REG_P (*ad.base_term)
3261 && REG_P (*ad.index_term)
3262 && in_class_p (*ad.base_term, INDEX_REG_CLASS, NULL)
3263 && ! in_class_p (*ad.index_term, INDEX_REG_CLASS, NULL))
3265 std::swap (ad.base, ad.index);
3266 std::swap (ad.base_term, ad.index_term);
3268 if (! check_only_p)
3269 change_p = equiv_address_substitution (&ad);
3270 if (ad.base_term != NULL
3271 && (process_addr_reg
3272 (ad.base_term, check_only_p, before,
3273 (ad.autoinc_p
3274 && !(REG_P (*ad.base_term)
3275 && find_regno_note (curr_insn, REG_DEAD,
3276 REGNO (*ad.base_term)) != NULL_RTX)
3277 ? after : NULL),
3278 base_reg_class (ad.mode, ad.as, ad.base_outer_code,
3279 get_index_code (&ad)))))
3281 change_p = true;
3282 if (ad.base_term2 != NULL)
3283 *ad.base_term2 = *ad.base_term;
3285 if (ad.index_term != NULL
3286 && process_addr_reg (ad.index_term, check_only_p,
3287 before, NULL, INDEX_REG_CLASS))
3288 change_p = true;
3290 /* Target hooks sometimes don't treat extra-constraint addresses as
3291 legitimate address_operands, so handle them specially. */
3292 if (insn_extra_address_constraint (cn)
3293 && satisfies_address_constraint_p (&ad, cn))
3294 return change_p;
3296 if (check_only_p)
3297 return change_p;
3299 /* There are three cases where the shape of *AD.INNER may now be invalid:
3301 1) the original address was valid, but either elimination or
3302 equiv_address_substitution was applied and that made
3303 the address invalid.
3305 2) the address is an invalid symbolic address created by
3306 force_const_to_mem.
3308 3) the address is a frame address with an invalid offset.
3310 4) the address is a frame address with an invalid base.
3312 All these cases involve a non-autoinc address, so there is no
3313 point revalidating other types. */
3314 if (ad.autoinc_p || valid_address_p (&ad))
3315 return change_p;
3317 /* Any index existed before LRA started, so we can assume that the
3318 presence and shape of the index is valid. */
3319 push_to_sequence (*before);
3320 lra_assert (ad.disp == ad.disp_term);
3321 if (ad.base == NULL)
3323 if (ad.index == NULL)
3325 rtx_insn *insn;
3326 rtx_insn *last = get_last_insn ();
3327 int code = -1;
3328 enum reg_class cl = base_reg_class (ad.mode, ad.as,
3329 SCRATCH, SCRATCH);
3330 rtx addr = *ad.inner;
3332 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "addr");
3333 if (HAVE_lo_sum)
3335 /* addr => lo_sum (new_base, addr), case (2) above. */
3336 insn = emit_insn (gen_rtx_SET
3337 (new_reg,
3338 gen_rtx_HIGH (Pmode, copy_rtx (addr))));
3339 code = recog_memoized (insn);
3340 if (code >= 0)
3342 *ad.inner = gen_rtx_LO_SUM (Pmode, new_reg, addr);
3343 if (! valid_address_p (ad.mode, *ad.outer, ad.as))
3345 /* Try to put lo_sum into register. */
3346 insn = emit_insn (gen_rtx_SET
3347 (new_reg,
3348 gen_rtx_LO_SUM (Pmode, new_reg, addr)));
3349 code = recog_memoized (insn);
3350 if (code >= 0)
3352 *ad.inner = new_reg;
3353 if (! valid_address_p (ad.mode, *ad.outer, ad.as))
3355 *ad.inner = addr;
3356 code = -1;
3362 if (code < 0)
3363 delete_insns_since (last);
3366 if (code < 0)
3368 /* addr => new_base, case (2) above. */
3369 lra_emit_move (new_reg, addr);
3371 for (insn = last == NULL_RTX ? get_insns () : NEXT_INSN (last);
3372 insn != NULL_RTX;
3373 insn = NEXT_INSN (insn))
3374 if (recog_memoized (insn) < 0)
3375 break;
3376 if (insn != NULL_RTX)
3378 /* Do nothing if we cannot generate right insns.
3379 This is analogous to reload pass behavior. */
3380 delete_insns_since (last);
3381 end_sequence ();
3382 return false;
3384 *ad.inner = new_reg;
3387 else
3389 /* index * scale + disp => new base + index * scale,
3390 case (1) above. */
3391 enum reg_class cl = base_reg_class (ad.mode, ad.as, PLUS,
3392 GET_CODE (*ad.index));
3394 lra_assert (INDEX_REG_CLASS != NO_REGS);
3395 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "disp");
3396 lra_emit_move (new_reg, *ad.disp);
3397 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3398 new_reg, *ad.index);
3401 else if (ad.index == NULL)
3403 int regno;
3404 enum reg_class cl;
3405 rtx set;
3406 rtx_insn *insns, *last_insn;
3407 /* Try to reload base into register only if the base is invalid
3408 for the address but with valid offset, case (4) above. */
3409 start_sequence ();
3410 new_reg = base_to_reg (&ad);
3412 /* base + disp => new base, cases (1) and (3) above. */
3413 /* Another option would be to reload the displacement into an
3414 index register. However, postreload has code to optimize
3415 address reloads that have the same base and different
3416 displacements, so reloading into an index register would
3417 not necessarily be a win. */
3418 if (new_reg == NULL_RTX)
3419 new_reg = base_plus_disp_to_reg (&ad);
3420 insns = get_insns ();
3421 last_insn = get_last_insn ();
3422 /* If we generated at least two insns, try last insn source as
3423 an address. If we succeed, we generate one less insn. */
3424 if (last_insn != insns && (set = single_set (last_insn)) != NULL_RTX
3425 && GET_CODE (SET_SRC (set)) == PLUS
3426 && REG_P (XEXP (SET_SRC (set), 0))
3427 && CONSTANT_P (XEXP (SET_SRC (set), 1)))
3429 *ad.inner = SET_SRC (set);
3430 if (valid_address_p (ad.mode, *ad.outer, ad.as))
3432 *ad.base_term = XEXP (SET_SRC (set), 0);
3433 *ad.disp_term = XEXP (SET_SRC (set), 1);
3434 cl = base_reg_class (ad.mode, ad.as, ad.base_outer_code,
3435 get_index_code (&ad));
3436 regno = REGNO (*ad.base_term);
3437 if (regno >= FIRST_PSEUDO_REGISTER
3438 && cl != lra_get_allocno_class (regno))
3439 lra_change_class (regno, cl, " Change to", true);
3440 new_reg = SET_SRC (set);
3441 delete_insns_since (PREV_INSN (last_insn));
3444 /* Try if target can split displacement into legitimite new disp
3445 and offset. If it's the case, we replace the last insn with
3446 insns for base + offset => new_reg and set new_reg + new disp
3447 to *ad.inner. */
3448 last_insn = get_last_insn ();
3449 if ((set = single_set (last_insn)) != NULL_RTX
3450 && GET_CODE (SET_SRC (set)) == PLUS
3451 && REG_P (XEXP (SET_SRC (set), 0))
3452 && REGNO (XEXP (SET_SRC (set), 0)) < FIRST_PSEUDO_REGISTER
3453 && CONST_INT_P (XEXP (SET_SRC (set), 1)))
3455 rtx addend, disp = XEXP (SET_SRC (set), 1);
3456 if (targetm.legitimize_address_displacement (&disp, &addend,
3457 ad.mode))
3459 rtx_insn *new_insns;
3460 start_sequence ();
3461 lra_emit_add (new_reg, XEXP (SET_SRC (set), 0), addend);
3462 new_insns = get_insns ();
3463 end_sequence ();
3464 new_reg = gen_rtx_PLUS (Pmode, new_reg, disp);
3465 delete_insns_since (PREV_INSN (last_insn));
3466 add_insn (new_insns);
3467 insns = get_insns ();
3470 end_sequence ();
3471 emit_insn (insns);
3472 *ad.inner = new_reg;
3474 else if (ad.disp_term != NULL)
3476 /* base + scale * index + disp => new base + scale * index,
3477 case (1) above. */
3478 new_reg = base_plus_disp_to_reg (&ad);
3479 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3480 new_reg, *ad.index);
3482 else if ((scale = get_index_scale (&ad)) == 1)
3484 /* The last transformation to one reg will be made in
3485 curr_insn_transform function. */
3486 end_sequence ();
3487 return false;
3489 else if (scale != 0)
3491 /* base + scale * index => base + new_reg,
3492 case (1) above.
3493 Index part of address may become invalid. For example, we
3494 changed pseudo on the equivalent memory and a subreg of the
3495 pseudo onto the memory of different mode for which the scale is
3496 prohibitted. */
3497 new_reg = index_part_to_reg (&ad);
3498 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3499 *ad.base_term, new_reg);
3501 else
3503 enum reg_class cl = base_reg_class (ad.mode, ad.as,
3504 SCRATCH, SCRATCH);
3505 rtx addr = *ad.inner;
3507 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "addr");
3508 /* addr => new_base. */
3509 lra_emit_move (new_reg, addr);
3510 *ad.inner = new_reg;
3512 *before = get_insns ();
3513 end_sequence ();
3514 return true;
3517 /* If CHECK_ONLY_P is false, do address reloads until it is necessary.
3518 Use process_address_1 as a helper function. Return true for any
3519 RTL changes.
3521 If CHECK_ONLY_P is true, just check address correctness. Return
3522 false if the address correct. */
3523 static bool
3524 process_address (int nop, bool check_only_p,
3525 rtx_insn **before, rtx_insn **after)
3527 bool res = false;
3529 while (process_address_1 (nop, check_only_p, before, after))
3531 if (check_only_p)
3532 return true;
3533 res = true;
3535 return res;
3538 /* Emit insns to reload VALUE into a new register. VALUE is an
3539 auto-increment or auto-decrement RTX whose operand is a register or
3540 memory location; so reloading involves incrementing that location.
3541 IN is either identical to VALUE, or some cheaper place to reload
3542 value being incremented/decremented from.
3544 INC_AMOUNT is the number to increment or decrement by (always
3545 positive and ignored for POST_MODIFY/PRE_MODIFY).
3547 Return pseudo containing the result. */
3548 static rtx
3549 emit_inc (enum reg_class new_rclass, rtx in, rtx value, poly_int64 inc_amount)
3551 /* REG or MEM to be copied and incremented. */
3552 rtx incloc = XEXP (value, 0);
3553 /* Nonzero if increment after copying. */
3554 int post = (GET_CODE (value) == POST_DEC || GET_CODE (value) == POST_INC
3555 || GET_CODE (value) == POST_MODIFY);
3556 rtx_insn *last;
3557 rtx inc;
3558 rtx_insn *add_insn;
3559 int code;
3560 rtx real_in = in == value ? incloc : in;
3561 rtx result;
3562 bool plus_p = true;
3564 if (GET_CODE (value) == PRE_MODIFY || GET_CODE (value) == POST_MODIFY)
3566 lra_assert (GET_CODE (XEXP (value, 1)) == PLUS
3567 || GET_CODE (XEXP (value, 1)) == MINUS);
3568 lra_assert (rtx_equal_p (XEXP (XEXP (value, 1), 0), XEXP (value, 0)));
3569 plus_p = GET_CODE (XEXP (value, 1)) == PLUS;
3570 inc = XEXP (XEXP (value, 1), 1);
3572 else
3574 if (GET_CODE (value) == PRE_DEC || GET_CODE (value) == POST_DEC)
3575 inc_amount = -inc_amount;
3577 inc = gen_int_mode (inc_amount, GET_MODE (value));
3580 if (! post && REG_P (incloc))
3581 result = incloc;
3582 else
3583 result = lra_create_new_reg (GET_MODE (value), value, new_rclass,
3584 "INC/DEC result");
3586 if (real_in != result)
3588 /* First copy the location to the result register. */
3589 lra_assert (REG_P (result));
3590 emit_insn (gen_move_insn (result, real_in));
3593 /* We suppose that there are insns to add/sub with the constant
3594 increment permitted in {PRE/POST)_{DEC/INC/MODIFY}. At least the
3595 old reload worked with this assumption. If the assumption
3596 becomes wrong, we should use approach in function
3597 base_plus_disp_to_reg. */
3598 if (in == value)
3600 /* See if we can directly increment INCLOC. */
3601 last = get_last_insn ();
3602 add_insn = emit_insn (plus_p
3603 ? gen_add2_insn (incloc, inc)
3604 : gen_sub2_insn (incloc, inc));
3606 code = recog_memoized (add_insn);
3607 if (code >= 0)
3609 if (! post && result != incloc)
3610 emit_insn (gen_move_insn (result, incloc));
3611 return result;
3613 delete_insns_since (last);
3616 /* If couldn't do the increment directly, must increment in RESULT.
3617 The way we do this depends on whether this is pre- or
3618 post-increment. For pre-increment, copy INCLOC to the reload
3619 register, increment it there, then save back. */
3620 if (! post)
3622 if (real_in != result)
3623 emit_insn (gen_move_insn (result, real_in));
3624 if (plus_p)
3625 emit_insn (gen_add2_insn (result, inc));
3626 else
3627 emit_insn (gen_sub2_insn (result, inc));
3628 if (result != incloc)
3629 emit_insn (gen_move_insn (incloc, result));
3631 else
3633 /* Post-increment.
3635 Because this might be a jump insn or a compare, and because
3636 RESULT may not be available after the insn in an input
3637 reload, we must do the incrementing before the insn being
3638 reloaded for.
3640 We have already copied IN to RESULT. Increment the copy in
3641 RESULT, save that back, then decrement RESULT so it has
3642 the original value. */
3643 if (plus_p)
3644 emit_insn (gen_add2_insn (result, inc));
3645 else
3646 emit_insn (gen_sub2_insn (result, inc));
3647 emit_insn (gen_move_insn (incloc, result));
3648 /* Restore non-modified value for the result. We prefer this
3649 way because it does not require an additional hard
3650 register. */
3651 if (plus_p)
3653 poly_int64 offset;
3654 if (poly_int_rtx_p (inc, &offset))
3655 emit_insn (gen_add2_insn (result,
3656 gen_int_mode (-offset,
3657 GET_MODE (result))));
3658 else
3659 emit_insn (gen_sub2_insn (result, inc));
3661 else
3662 emit_insn (gen_add2_insn (result, inc));
3664 return result;
3667 /* Return true if the current move insn does not need processing as we
3668 already know that it satisfies its constraints. */
3669 static bool
3670 simple_move_p (void)
3672 rtx dest, src;
3673 enum reg_class dclass, sclass;
3675 lra_assert (curr_insn_set != NULL_RTX);
3676 dest = SET_DEST (curr_insn_set);
3677 src = SET_SRC (curr_insn_set);
3679 /* If the instruction has multiple sets we need to process it even if it
3680 is single_set. This can happen if one or more of the SETs are dead.
3681 See PR73650. */
3682 if (multiple_sets (curr_insn))
3683 return false;
3685 return ((dclass = get_op_class (dest)) != NO_REGS
3686 && (sclass = get_op_class (src)) != NO_REGS
3687 /* The backend guarantees that register moves of cost 2
3688 never need reloads. */
3689 && targetm.register_move_cost (GET_MODE (src), sclass, dclass) == 2);
3692 /* Swap operands NOP and NOP + 1. */
3693 static inline void
3694 swap_operands (int nop)
3696 std::swap (curr_operand_mode[nop], curr_operand_mode[nop + 1]);
3697 std::swap (original_subreg_reg_mode[nop], original_subreg_reg_mode[nop + 1]);
3698 std::swap (*curr_id->operand_loc[nop], *curr_id->operand_loc[nop + 1]);
3699 std::swap (equiv_substition_p[nop], equiv_substition_p[nop + 1]);
3700 /* Swap the duplicates too. */
3701 lra_update_dup (curr_id, nop);
3702 lra_update_dup (curr_id, nop + 1);
3705 /* Main entry point of the constraint code: search the body of the
3706 current insn to choose the best alternative. It is mimicking insn
3707 alternative cost calculation model of former reload pass. That is
3708 because machine descriptions were written to use this model. This
3709 model can be changed in future. Make commutative operand exchange
3710 if it is chosen.
3712 if CHECK_ONLY_P is false, do RTL changes to satisfy the
3713 constraints. Return true if any change happened during function
3714 call.
3716 If CHECK_ONLY_P is true then don't do any transformation. Just
3717 check that the insn satisfies all constraints. If the insn does
3718 not satisfy any constraint, return true. */
3719 static bool
3720 curr_insn_transform (bool check_only_p)
3722 int i, j, k;
3723 int n_operands;
3724 int n_alternatives;
3725 int n_outputs;
3726 int commutative;
3727 signed char goal_alt_matched[MAX_RECOG_OPERANDS][MAX_RECOG_OPERANDS];
3728 signed char match_inputs[MAX_RECOG_OPERANDS + 1];
3729 signed char outputs[MAX_RECOG_OPERANDS + 1];
3730 rtx_insn *before, *after;
3731 bool alt_p = false;
3732 /* Flag that the insn has been changed through a transformation. */
3733 bool change_p;
3734 bool sec_mem_p;
3735 bool use_sec_mem_p;
3736 int max_regno_before;
3737 int reused_alternative_num;
3739 curr_insn_set = single_set (curr_insn);
3740 if (curr_insn_set != NULL_RTX && simple_move_p ())
3741 return false;
3743 no_input_reloads_p = no_output_reloads_p = false;
3744 goal_alt_number = -1;
3745 change_p = sec_mem_p = false;
3746 /* JUMP_INSNs and CALL_INSNs are not allowed to have any output
3747 reloads; neither are insns that SET cc0. Insns that use CC0 are
3748 not allowed to have any input reloads. */
3749 if (JUMP_P (curr_insn) || CALL_P (curr_insn))
3750 no_output_reloads_p = true;
3752 if (HAVE_cc0 && reg_referenced_p (cc0_rtx, PATTERN (curr_insn)))
3753 no_input_reloads_p = true;
3754 if (HAVE_cc0 && reg_set_p (cc0_rtx, PATTERN (curr_insn)))
3755 no_output_reloads_p = true;
3757 n_operands = curr_static_id->n_operands;
3758 n_alternatives = curr_static_id->n_alternatives;
3760 /* Just return "no reloads" if insn has no operands with
3761 constraints. */
3762 if (n_operands == 0 || n_alternatives == 0)
3763 return false;
3765 max_regno_before = max_reg_num ();
3767 for (i = 0; i < n_operands; i++)
3769 goal_alt_matched[i][0] = -1;
3770 goal_alt_matches[i] = -1;
3773 commutative = curr_static_id->commutative;
3775 /* Now see what we need for pseudos that didn't get hard regs or got
3776 the wrong kind of hard reg. For this, we must consider all the
3777 operands together against the register constraints. */
3779 best_losers = best_overall = INT_MAX;
3780 best_reload_sum = 0;
3782 curr_swapped = false;
3783 goal_alt_swapped = false;
3785 if (! check_only_p)
3786 /* Make equivalence substitution and memory subreg elimination
3787 before address processing because an address legitimacy can
3788 depend on memory mode. */
3789 for (i = 0; i < n_operands; i++)
3791 rtx op, subst, old;
3792 bool op_change_p = false;
3794 if (curr_static_id->operand[i].is_operator)
3795 continue;
3797 old = op = *curr_id->operand_loc[i];
3798 if (GET_CODE (old) == SUBREG)
3799 old = SUBREG_REG (old);
3800 subst = get_equiv_with_elimination (old, curr_insn);
3801 original_subreg_reg_mode[i] = VOIDmode;
3802 equiv_substition_p[i] = false;
3803 if (subst != old)
3805 equiv_substition_p[i] = true;
3806 subst = copy_rtx (subst);
3807 lra_assert (REG_P (old));
3808 if (GET_CODE (op) != SUBREG)
3809 *curr_id->operand_loc[i] = subst;
3810 else
3812 SUBREG_REG (op) = subst;
3813 if (GET_MODE (subst) == VOIDmode)
3814 original_subreg_reg_mode[i] = GET_MODE (old);
3816 if (lra_dump_file != NULL)
3818 fprintf (lra_dump_file,
3819 "Changing pseudo %d in operand %i of insn %u on equiv ",
3820 REGNO (old), i, INSN_UID (curr_insn));
3821 dump_value_slim (lra_dump_file, subst, 1);
3822 fprintf (lra_dump_file, "\n");
3824 op_change_p = change_p = true;
3826 if (simplify_operand_subreg (i, GET_MODE (old)) || op_change_p)
3828 change_p = true;
3829 lra_update_dup (curr_id, i);
3833 /* Reload address registers and displacements. We do it before
3834 finding an alternative because of memory constraints. */
3835 before = after = NULL;
3836 for (i = 0; i < n_operands; i++)
3837 if (! curr_static_id->operand[i].is_operator
3838 && process_address (i, check_only_p, &before, &after))
3840 if (check_only_p)
3841 return true;
3842 change_p = true;
3843 lra_update_dup (curr_id, i);
3846 if (change_p)
3847 /* If we've changed the instruction then any alternative that
3848 we chose previously may no longer be valid. */
3849 lra_set_used_insn_alternative (curr_insn, -1);
3851 if (! check_only_p && curr_insn_set != NULL_RTX
3852 && check_and_process_move (&change_p, &sec_mem_p))
3853 return change_p;
3855 try_swapped:
3857 reused_alternative_num = check_only_p ? -1 : curr_id->used_insn_alternative;
3858 if (lra_dump_file != NULL && reused_alternative_num >= 0)
3859 fprintf (lra_dump_file, "Reusing alternative %d for insn #%u\n",
3860 reused_alternative_num, INSN_UID (curr_insn));
3862 if (process_alt_operands (reused_alternative_num))
3863 alt_p = true;
3865 if (check_only_p)
3866 return ! alt_p || best_losers != 0;
3868 /* If insn is commutative (it's safe to exchange a certain pair of
3869 operands) then we need to try each alternative twice, the second
3870 time matching those two operands as if we had exchanged them. To
3871 do this, really exchange them in operands.
3873 If we have just tried the alternatives the second time, return
3874 operands to normal and drop through. */
3876 if (reused_alternative_num < 0 && commutative >= 0)
3878 curr_swapped = !curr_swapped;
3879 if (curr_swapped)
3881 swap_operands (commutative);
3882 goto try_swapped;
3884 else
3885 swap_operands (commutative);
3888 if (! alt_p && ! sec_mem_p)
3890 /* No alternative works with reloads?? */
3891 if (INSN_CODE (curr_insn) >= 0)
3892 fatal_insn ("unable to generate reloads for:", curr_insn);
3893 error_for_asm (curr_insn,
3894 "inconsistent operand constraints in an %<asm%>");
3895 /* Avoid further trouble with this insn. Don't generate use
3896 pattern here as we could use the insn SP offset. */
3897 lra_set_insn_deleted (curr_insn);
3898 return true;
3901 /* If the best alternative is with operands 1 and 2 swapped, swap
3902 them. Update the operand numbers of any reloads already
3903 pushed. */
3905 if (goal_alt_swapped)
3907 if (lra_dump_file != NULL)
3908 fprintf (lra_dump_file, " Commutative operand exchange in insn %u\n",
3909 INSN_UID (curr_insn));
3911 /* Swap the duplicates too. */
3912 swap_operands (commutative);
3913 change_p = true;
3916 /* Some targets' TARGET_SECONDARY_MEMORY_NEEDED (e.g. x86) are defined
3917 too conservatively. So we use the secondary memory only if there
3918 is no any alternative without reloads. */
3919 use_sec_mem_p = false;
3920 if (! alt_p)
3921 use_sec_mem_p = true;
3922 else if (sec_mem_p)
3924 for (i = 0; i < n_operands; i++)
3925 if (! goal_alt_win[i] && ! goal_alt_match_win[i])
3926 break;
3927 use_sec_mem_p = i < n_operands;
3930 if (use_sec_mem_p)
3932 int in = -1, out = -1;
3933 rtx new_reg, src, dest, rld;
3934 machine_mode sec_mode, rld_mode;
3936 lra_assert (curr_insn_set != NULL_RTX && sec_mem_p);
3937 dest = SET_DEST (curr_insn_set);
3938 src = SET_SRC (curr_insn_set);
3939 for (i = 0; i < n_operands; i++)
3940 if (*curr_id->operand_loc[i] == dest)
3941 out = i;
3942 else if (*curr_id->operand_loc[i] == src)
3943 in = i;
3944 for (i = 0; i < curr_static_id->n_dups; i++)
3945 if (out < 0 && *curr_id->dup_loc[i] == dest)
3946 out = curr_static_id->dup_num[i];
3947 else if (in < 0 && *curr_id->dup_loc[i] == src)
3948 in = curr_static_id->dup_num[i];
3949 lra_assert (out >= 0 && in >= 0
3950 && curr_static_id->operand[out].type == OP_OUT
3951 && curr_static_id->operand[in].type == OP_IN);
3952 rld = partial_subreg_p (GET_MODE (src), GET_MODE (dest)) ? src : dest;
3953 rld_mode = GET_MODE (rld);
3954 sec_mode = targetm.secondary_memory_needed_mode (rld_mode);
3955 new_reg = lra_create_new_reg (sec_mode, NULL_RTX,
3956 NO_REGS, "secondary");
3957 /* If the mode is changed, it should be wider. */
3958 lra_assert (!partial_subreg_p (sec_mode, rld_mode));
3959 if (sec_mode != rld_mode)
3961 /* If the target says specifically to use another mode for
3962 secondary memory moves we can not reuse the original
3963 insn. */
3964 after = emit_spill_move (false, new_reg, dest);
3965 lra_process_new_insns (curr_insn, NULL, after,
3966 "Inserting the sec. move");
3967 /* We may have non null BEFORE here (e.g. after address
3968 processing. */
3969 push_to_sequence (before);
3970 before = emit_spill_move (true, new_reg, src);
3971 emit_insn (before);
3972 before = get_insns ();
3973 end_sequence ();
3974 lra_process_new_insns (curr_insn, before, NULL, "Changing on");
3975 lra_set_insn_deleted (curr_insn);
3977 else if (dest == rld)
3979 *curr_id->operand_loc[out] = new_reg;
3980 lra_update_dup (curr_id, out);
3981 after = emit_spill_move (false, new_reg, dest);
3982 lra_process_new_insns (curr_insn, NULL, after,
3983 "Inserting the sec. move");
3985 else
3987 *curr_id->operand_loc[in] = new_reg;
3988 lra_update_dup (curr_id, in);
3989 /* See comments above. */
3990 push_to_sequence (before);
3991 before = emit_spill_move (true, new_reg, src);
3992 emit_insn (before);
3993 before = get_insns ();
3994 end_sequence ();
3995 lra_process_new_insns (curr_insn, before, NULL,
3996 "Inserting the sec. move");
3998 lra_update_insn_regno_info (curr_insn);
3999 return true;
4002 lra_assert (goal_alt_number >= 0);
4003 lra_set_used_insn_alternative (curr_insn, goal_alt_number);
4005 if (lra_dump_file != NULL)
4007 const char *p;
4009 fprintf (lra_dump_file, " Choosing alt %d in insn %u:",
4010 goal_alt_number, INSN_UID (curr_insn));
4011 for (i = 0; i < n_operands; i++)
4013 p = (curr_static_id->operand_alternative
4014 [goal_alt_number * n_operands + i].constraint);
4015 if (*p == '\0')
4016 continue;
4017 fprintf (lra_dump_file, " (%d) ", i);
4018 for (; *p != '\0' && *p != ',' && *p != '#'; p++)
4019 fputc (*p, lra_dump_file);
4021 if (INSN_CODE (curr_insn) >= 0
4022 && (p = get_insn_name (INSN_CODE (curr_insn))) != NULL)
4023 fprintf (lra_dump_file, " {%s}", p);
4024 if (maybe_ne (curr_id->sp_offset, 0))
4026 fprintf (lra_dump_file, " (sp_off=");
4027 print_dec (curr_id->sp_offset, lra_dump_file);
4028 fprintf (lra_dump_file, ")");
4030 fprintf (lra_dump_file, "\n");
4033 /* Right now, for any pair of operands I and J that are required to
4034 match, with J < I, goal_alt_matches[I] is J. Add I to
4035 goal_alt_matched[J]. */
4037 for (i = 0; i < n_operands; i++)
4038 if ((j = goal_alt_matches[i]) >= 0)
4040 for (k = 0; goal_alt_matched[j][k] >= 0; k++)
4042 /* We allow matching one output operand and several input
4043 operands. */
4044 lra_assert (k == 0
4045 || (curr_static_id->operand[j].type == OP_OUT
4046 && curr_static_id->operand[i].type == OP_IN
4047 && (curr_static_id->operand
4048 [goal_alt_matched[j][0]].type == OP_IN)));
4049 goal_alt_matched[j][k] = i;
4050 goal_alt_matched[j][k + 1] = -1;
4053 for (i = 0; i < n_operands; i++)
4054 goal_alt_win[i] |= goal_alt_match_win[i];
4056 /* Any constants that aren't allowed and can't be reloaded into
4057 registers are here changed into memory references. */
4058 for (i = 0; i < n_operands; i++)
4059 if (goal_alt_win[i])
4061 int regno;
4062 enum reg_class new_class;
4063 rtx reg = *curr_id->operand_loc[i];
4065 if (GET_CODE (reg) == SUBREG)
4066 reg = SUBREG_REG (reg);
4068 if (REG_P (reg) && (regno = REGNO (reg)) >= FIRST_PSEUDO_REGISTER)
4070 bool ok_p = in_class_p (reg, goal_alt[i], &new_class);
4072 if (new_class != NO_REGS && get_reg_class (regno) != new_class)
4074 lra_assert (ok_p);
4075 lra_change_class (regno, new_class, " Change to", true);
4079 else
4081 const char *constraint;
4082 char c;
4083 rtx op = *curr_id->operand_loc[i];
4084 rtx subreg = NULL_RTX;
4085 machine_mode mode = curr_operand_mode[i];
4087 if (GET_CODE (op) == SUBREG)
4089 subreg = op;
4090 op = SUBREG_REG (op);
4091 mode = GET_MODE (op);
4094 if (CONST_POOL_OK_P (mode, op)
4095 && ((targetm.preferred_reload_class
4096 (op, (enum reg_class) goal_alt[i]) == NO_REGS)
4097 || no_input_reloads_p))
4099 rtx tem = force_const_mem (mode, op);
4101 change_p = true;
4102 if (subreg != NULL_RTX)
4103 tem = gen_rtx_SUBREG (mode, tem, SUBREG_BYTE (subreg));
4105 *curr_id->operand_loc[i] = tem;
4106 lra_update_dup (curr_id, i);
4107 process_address (i, false, &before, &after);
4109 /* If the alternative accepts constant pool refs directly
4110 there will be no reload needed at all. */
4111 if (subreg != NULL_RTX)
4112 continue;
4113 /* Skip alternatives before the one requested. */
4114 constraint = (curr_static_id->operand_alternative
4115 [goal_alt_number * n_operands + i].constraint);
4116 for (;
4117 (c = *constraint) && c != ',' && c != '#';
4118 constraint += CONSTRAINT_LEN (c, constraint))
4120 enum constraint_num cn = lookup_constraint (constraint);
4121 if ((insn_extra_memory_constraint (cn)
4122 || insn_extra_special_memory_constraint (cn))
4123 && satisfies_memory_constraint_p (tem, cn))
4124 break;
4126 if (c == '\0' || c == ',' || c == '#')
4127 continue;
4129 goal_alt_win[i] = true;
4133 n_outputs = 0;
4134 outputs[0] = -1;
4135 for (i = 0; i < n_operands; i++)
4137 int regno;
4138 bool optional_p = false;
4139 rtx old, new_reg;
4140 rtx op = *curr_id->operand_loc[i];
4142 if (goal_alt_win[i])
4144 if (goal_alt[i] == NO_REGS
4145 && REG_P (op)
4146 /* When we assign NO_REGS it means that we will not
4147 assign a hard register to the scratch pseudo by
4148 assigment pass and the scratch pseudo will be
4149 spilled. Spilled scratch pseudos are transformed
4150 back to scratches at the LRA end. */
4151 && lra_former_scratch_operand_p (curr_insn, i)
4152 && lra_former_scratch_p (REGNO (op)))
4154 int regno = REGNO (op);
4155 lra_change_class (regno, NO_REGS, " Change to", true);
4156 if (lra_get_regno_hard_regno (regno) >= 0)
4157 /* We don't have to mark all insn affected by the
4158 spilled pseudo as there is only one such insn, the
4159 current one. */
4160 reg_renumber[regno] = -1;
4161 lra_assert (bitmap_single_bit_set_p
4162 (&lra_reg_info[REGNO (op)].insn_bitmap));
4164 /* We can do an optional reload. If the pseudo got a hard
4165 reg, we might improve the code through inheritance. If
4166 it does not get a hard register we coalesce memory/memory
4167 moves later. Ignore move insns to avoid cycling. */
4168 if (! lra_simple_p
4169 && lra_undo_inheritance_iter < LRA_MAX_INHERITANCE_PASSES
4170 && goal_alt[i] != NO_REGS && REG_P (op)
4171 && (regno = REGNO (op)) >= FIRST_PSEUDO_REGISTER
4172 && regno < new_regno_start
4173 && ! lra_former_scratch_p (regno)
4174 && reg_renumber[regno] < 0
4175 /* Check that the optional reload pseudo will be able to
4176 hold given mode value. */
4177 && ! (prohibited_class_reg_set_mode_p
4178 (goal_alt[i], reg_class_contents[goal_alt[i]],
4179 PSEUDO_REGNO_MODE (regno)))
4180 && (curr_insn_set == NULL_RTX
4181 || !((REG_P (SET_SRC (curr_insn_set))
4182 || MEM_P (SET_SRC (curr_insn_set))
4183 || GET_CODE (SET_SRC (curr_insn_set)) == SUBREG)
4184 && (REG_P (SET_DEST (curr_insn_set))
4185 || MEM_P (SET_DEST (curr_insn_set))
4186 || GET_CODE (SET_DEST (curr_insn_set)) == SUBREG))))
4187 optional_p = true;
4188 else
4189 continue;
4192 /* Operands that match previous ones have already been handled. */
4193 if (goal_alt_matches[i] >= 0)
4194 continue;
4196 /* We should not have an operand with a non-offsettable address
4197 appearing where an offsettable address will do. It also may
4198 be a case when the address should be special in other words
4199 not a general one (e.g. it needs no index reg). */
4200 if (goal_alt_matched[i][0] == -1 && goal_alt_offmemok[i] && MEM_P (op))
4202 enum reg_class rclass;
4203 rtx *loc = &XEXP (op, 0);
4204 enum rtx_code code = GET_CODE (*loc);
4206 push_to_sequence (before);
4207 rclass = base_reg_class (GET_MODE (op), MEM_ADDR_SPACE (op),
4208 MEM, SCRATCH);
4209 if (GET_RTX_CLASS (code) == RTX_AUTOINC)
4210 new_reg = emit_inc (rclass, *loc, *loc,
4211 /* This value does not matter for MODIFY. */
4212 GET_MODE_SIZE (GET_MODE (op)));
4213 else if (get_reload_reg (OP_IN, Pmode, *loc, rclass, FALSE,
4214 "offsetable address", &new_reg))
4215 lra_emit_move (new_reg, *loc);
4216 before = get_insns ();
4217 end_sequence ();
4218 *loc = new_reg;
4219 lra_update_dup (curr_id, i);
4221 else if (goal_alt_matched[i][0] == -1)
4223 machine_mode mode;
4224 rtx reg, *loc;
4225 int hard_regno;
4226 enum op_type type = curr_static_id->operand[i].type;
4228 loc = curr_id->operand_loc[i];
4229 mode = curr_operand_mode[i];
4230 if (GET_CODE (*loc) == SUBREG)
4232 reg = SUBREG_REG (*loc);
4233 poly_int64 byte = SUBREG_BYTE (*loc);
4234 if (REG_P (reg)
4235 /* Strict_low_part requires reloading the register and not
4236 just the subreg. Likewise for a strict subreg no wider
4237 than a word for WORD_REGISTER_OPERATIONS targets. */
4238 && (curr_static_id->operand[i].strict_low
4239 || (!paradoxical_subreg_p (mode, GET_MODE (reg))
4240 && (hard_regno
4241 = get_try_hard_regno (REGNO (reg))) >= 0
4242 && (simplify_subreg_regno
4243 (hard_regno,
4244 GET_MODE (reg), byte, mode) < 0)
4245 && (goal_alt[i] == NO_REGS
4246 || (simplify_subreg_regno
4247 (ira_class_hard_regs[goal_alt[i]][0],
4248 GET_MODE (reg), byte, mode) >= 0)))
4249 || (partial_subreg_p (mode, GET_MODE (reg))
4250 && known_le (GET_MODE_SIZE (GET_MODE (reg)),
4251 UNITS_PER_WORD)
4252 && WORD_REGISTER_OPERATIONS)))
4254 /* An OP_INOUT is required when reloading a subreg of a
4255 mode wider than a word to ensure that data beyond the
4256 word being reloaded is preserved. Also automatically
4257 ensure that strict_low_part reloads are made into
4258 OP_INOUT which should already be true from the backend
4259 constraints. */
4260 if (type == OP_OUT
4261 && (curr_static_id->operand[i].strict_low
4262 || read_modify_subreg_p (*loc)))
4263 type = OP_INOUT;
4264 loc = &SUBREG_REG (*loc);
4265 mode = GET_MODE (*loc);
4268 old = *loc;
4269 if (get_reload_reg (type, mode, old, goal_alt[i],
4270 loc != curr_id->operand_loc[i], "", &new_reg)
4271 && type != OP_OUT)
4273 push_to_sequence (before);
4274 lra_emit_move (new_reg, old);
4275 before = get_insns ();
4276 end_sequence ();
4278 *loc = new_reg;
4279 if (type != OP_IN
4280 && find_reg_note (curr_insn, REG_UNUSED, old) == NULL_RTX)
4282 start_sequence ();
4283 lra_emit_move (type == OP_INOUT ? copy_rtx (old) : old, new_reg);
4284 emit_insn (after);
4285 after = get_insns ();
4286 end_sequence ();
4287 *loc = new_reg;
4289 for (j = 0; j < goal_alt_dont_inherit_ops_num; j++)
4290 if (goal_alt_dont_inherit_ops[j] == i)
4292 lra_set_regno_unique_value (REGNO (new_reg));
4293 break;
4295 lra_update_dup (curr_id, i);
4297 else if (curr_static_id->operand[i].type == OP_IN
4298 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4299 == OP_OUT
4300 || (curr_static_id->operand[goal_alt_matched[i][0]].type
4301 == OP_INOUT
4302 && (operands_match_p
4303 (*curr_id->operand_loc[i],
4304 *curr_id->operand_loc[goal_alt_matched[i][0]],
4305 -1)))))
4307 /* generate reloads for input and matched outputs. */
4308 match_inputs[0] = i;
4309 match_inputs[1] = -1;
4310 match_reload (goal_alt_matched[i][0], match_inputs, outputs,
4311 goal_alt[i], &before, &after,
4312 curr_static_id->operand_alternative
4313 [goal_alt_number * n_operands + goal_alt_matched[i][0]]
4314 .earlyclobber);
4316 else if ((curr_static_id->operand[i].type == OP_OUT
4317 || (curr_static_id->operand[i].type == OP_INOUT
4318 && (operands_match_p
4319 (*curr_id->operand_loc[i],
4320 *curr_id->operand_loc[goal_alt_matched[i][0]],
4321 -1))))
4322 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4323 == OP_IN))
4324 /* Generate reloads for output and matched inputs. */
4325 match_reload (i, goal_alt_matched[i], outputs, goal_alt[i], &before,
4326 &after, curr_static_id->operand_alternative
4327 [goal_alt_number * n_operands + i].earlyclobber);
4328 else if (curr_static_id->operand[i].type == OP_IN
4329 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4330 == OP_IN))
4332 /* Generate reloads for matched inputs. */
4333 match_inputs[0] = i;
4334 for (j = 0; (k = goal_alt_matched[i][j]) >= 0; j++)
4335 match_inputs[j + 1] = k;
4336 match_inputs[j + 1] = -1;
4337 match_reload (-1, match_inputs, outputs, goal_alt[i], &before,
4338 &after, false);
4340 else
4341 /* We must generate code in any case when function
4342 process_alt_operands decides that it is possible. */
4343 gcc_unreachable ();
4345 /* Memorise processed outputs so that output remaining to be processed
4346 can avoid using the same register value (see match_reload). */
4347 if (curr_static_id->operand[i].type == OP_OUT)
4349 outputs[n_outputs++] = i;
4350 outputs[n_outputs] = -1;
4353 if (optional_p)
4355 rtx reg = op;
4357 lra_assert (REG_P (reg));
4358 regno = REGNO (reg);
4359 op = *curr_id->operand_loc[i]; /* Substitution. */
4360 if (GET_CODE (op) == SUBREG)
4361 op = SUBREG_REG (op);
4362 gcc_assert (REG_P (op) && (int) REGNO (op) >= new_regno_start);
4363 bitmap_set_bit (&lra_optional_reload_pseudos, REGNO (op));
4364 lra_reg_info[REGNO (op)].restore_rtx = reg;
4365 if (lra_dump_file != NULL)
4366 fprintf (lra_dump_file,
4367 " Making reload reg %d for reg %d optional\n",
4368 REGNO (op), regno);
4371 if (before != NULL_RTX || after != NULL_RTX
4372 || max_regno_before != max_reg_num ())
4373 change_p = true;
4374 if (change_p)
4376 lra_update_operator_dups (curr_id);
4377 /* Something changes -- process the insn. */
4378 lra_update_insn_regno_info (curr_insn);
4380 lra_process_new_insns (curr_insn, before, after, "Inserting insn reload");
4381 return change_p;
4384 /* Return true if INSN satisfies all constraints. In other words, no
4385 reload insns are needed. */
4386 bool
4387 lra_constrain_insn (rtx_insn *insn)
4389 int saved_new_regno_start = new_regno_start;
4390 int saved_new_insn_uid_start = new_insn_uid_start;
4391 bool change_p;
4393 curr_insn = insn;
4394 curr_id = lra_get_insn_recog_data (curr_insn);
4395 curr_static_id = curr_id->insn_static_data;
4396 new_insn_uid_start = get_max_uid ();
4397 new_regno_start = max_reg_num ();
4398 change_p = curr_insn_transform (true);
4399 new_regno_start = saved_new_regno_start;
4400 new_insn_uid_start = saved_new_insn_uid_start;
4401 return ! change_p;
4404 /* Return true if X is in LIST. */
4405 static bool
4406 in_list_p (rtx x, rtx list)
4408 for (; list != NULL_RTX; list = XEXP (list, 1))
4409 if (XEXP (list, 0) == x)
4410 return true;
4411 return false;
4414 /* Return true if X contains an allocatable hard register (if
4415 HARD_REG_P) or a (spilled if SPILLED_P) pseudo. */
4416 static bool
4417 contains_reg_p (rtx x, bool hard_reg_p, bool spilled_p)
4419 int i, j;
4420 const char *fmt;
4421 enum rtx_code code;
4423 code = GET_CODE (x);
4424 if (REG_P (x))
4426 int regno = REGNO (x);
4427 HARD_REG_SET alloc_regs;
4429 if (hard_reg_p)
4431 if (regno >= FIRST_PSEUDO_REGISTER)
4432 regno = lra_get_regno_hard_regno (regno);
4433 if (regno < 0)
4434 return false;
4435 COMPL_HARD_REG_SET (alloc_regs, lra_no_alloc_regs);
4436 return overlaps_hard_reg_set_p (alloc_regs, GET_MODE (x), regno);
4438 else
4440 if (regno < FIRST_PSEUDO_REGISTER)
4441 return false;
4442 if (! spilled_p)
4443 return true;
4444 return lra_get_regno_hard_regno (regno) < 0;
4447 fmt = GET_RTX_FORMAT (code);
4448 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4450 if (fmt[i] == 'e')
4452 if (contains_reg_p (XEXP (x, i), hard_reg_p, spilled_p))
4453 return true;
4455 else if (fmt[i] == 'E')
4457 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4458 if (contains_reg_p (XVECEXP (x, i, j), hard_reg_p, spilled_p))
4459 return true;
4462 return false;
4465 /* Process all regs in location *LOC and change them on equivalent
4466 substitution. Return true if any change was done. */
4467 static bool
4468 loc_equivalence_change_p (rtx *loc)
4470 rtx subst, reg, x = *loc;
4471 bool result = false;
4472 enum rtx_code code = GET_CODE (x);
4473 const char *fmt;
4474 int i, j;
4476 if (code == SUBREG)
4478 reg = SUBREG_REG (x);
4479 if ((subst = get_equiv_with_elimination (reg, curr_insn)) != reg
4480 && GET_MODE (subst) == VOIDmode)
4482 /* We cannot reload debug location. Simplify subreg here
4483 while we know the inner mode. */
4484 *loc = simplify_gen_subreg (GET_MODE (x), subst,
4485 GET_MODE (reg), SUBREG_BYTE (x));
4486 return true;
4489 if (code == REG && (subst = get_equiv_with_elimination (x, curr_insn)) != x)
4491 *loc = subst;
4492 return true;
4495 /* Scan all the operand sub-expressions. */
4496 fmt = GET_RTX_FORMAT (code);
4497 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4499 if (fmt[i] == 'e')
4500 result = loc_equivalence_change_p (&XEXP (x, i)) || result;
4501 else if (fmt[i] == 'E')
4502 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4503 result
4504 = loc_equivalence_change_p (&XVECEXP (x, i, j)) || result;
4506 return result;
4509 /* Similar to loc_equivalence_change_p, but for use as
4510 simplify_replace_fn_rtx callback. DATA is insn for which the
4511 elimination is done. If it null we don't do the elimination. */
4512 static rtx
4513 loc_equivalence_callback (rtx loc, const_rtx, void *data)
4515 if (!REG_P (loc))
4516 return NULL_RTX;
4518 rtx subst = (data == NULL
4519 ? get_equiv (loc) : get_equiv_with_elimination (loc, (rtx_insn *) data));
4520 if (subst != loc)
4521 return subst;
4523 return NULL_RTX;
4526 /* Maximum number of generated reload insns per an insn. It is for
4527 preventing this pass cycling in a bug case. */
4528 #define MAX_RELOAD_INSNS_NUMBER LRA_MAX_INSN_RELOADS
4530 /* The current iteration number of this LRA pass. */
4531 int lra_constraint_iter;
4533 /* True if we substituted equiv which needs checking register
4534 allocation correctness because the equivalent value contains
4535 allocatable hard registers or when we restore multi-register
4536 pseudo. */
4537 bool lra_risky_transformations_p;
4539 /* Return true if REGNO is referenced in more than one block. */
4540 static bool
4541 multi_block_pseudo_p (int regno)
4543 basic_block bb = NULL;
4544 unsigned int uid;
4545 bitmap_iterator bi;
4547 if (regno < FIRST_PSEUDO_REGISTER)
4548 return false;
4550 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi)
4551 if (bb == NULL)
4552 bb = BLOCK_FOR_INSN (lra_insn_recog_data[uid]->insn);
4553 else if (BLOCK_FOR_INSN (lra_insn_recog_data[uid]->insn) != bb)
4554 return true;
4555 return false;
4558 /* Return true if LIST contains a deleted insn. */
4559 static bool
4560 contains_deleted_insn_p (rtx_insn_list *list)
4562 for (; list != NULL_RTX; list = list->next ())
4563 if (NOTE_P (list->insn ())
4564 && NOTE_KIND (list->insn ()) == NOTE_INSN_DELETED)
4565 return true;
4566 return false;
4569 /* Return true if X contains a pseudo dying in INSN. */
4570 static bool
4571 dead_pseudo_p (rtx x, rtx_insn *insn)
4573 int i, j;
4574 const char *fmt;
4575 enum rtx_code code;
4577 if (REG_P (x))
4578 return (insn != NULL_RTX
4579 && find_regno_note (insn, REG_DEAD, REGNO (x)) != NULL_RTX);
4580 code = GET_CODE (x);
4581 fmt = GET_RTX_FORMAT (code);
4582 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4584 if (fmt[i] == 'e')
4586 if (dead_pseudo_p (XEXP (x, i), insn))
4587 return true;
4589 else if (fmt[i] == 'E')
4591 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4592 if (dead_pseudo_p (XVECEXP (x, i, j), insn))
4593 return true;
4596 return false;
4599 /* Return true if INSN contains a dying pseudo in INSN right hand
4600 side. */
4601 static bool
4602 insn_rhs_dead_pseudo_p (rtx_insn *insn)
4604 rtx set = single_set (insn);
4606 gcc_assert (set != NULL);
4607 return dead_pseudo_p (SET_SRC (set), insn);
4610 /* Return true if any init insn of REGNO contains a dying pseudo in
4611 insn right hand side. */
4612 static bool
4613 init_insn_rhs_dead_pseudo_p (int regno)
4615 rtx_insn_list *insns = ira_reg_equiv[regno].init_insns;
4617 if (insns == NULL)
4618 return false;
4619 for (; insns != NULL_RTX; insns = insns->next ())
4620 if (insn_rhs_dead_pseudo_p (insns->insn ()))
4621 return true;
4622 return false;
4625 /* Return TRUE if REGNO has a reverse equivalence. The equivalence is
4626 reverse only if we have one init insn with given REGNO as a
4627 source. */
4628 static bool
4629 reverse_equiv_p (int regno)
4631 rtx_insn_list *insns = ira_reg_equiv[regno].init_insns;
4632 rtx set;
4634 if (insns == NULL)
4635 return false;
4636 if (! INSN_P (insns->insn ())
4637 || insns->next () != NULL)
4638 return false;
4639 if ((set = single_set (insns->insn ())) == NULL_RTX)
4640 return false;
4641 return REG_P (SET_SRC (set)) && (int) REGNO (SET_SRC (set)) == regno;
4644 /* Return TRUE if REGNO was reloaded in an equivalence init insn. We
4645 call this function only for non-reverse equivalence. */
4646 static bool
4647 contains_reloaded_insn_p (int regno)
4649 rtx set;
4650 rtx_insn_list *list = ira_reg_equiv[regno].init_insns;
4652 for (; list != NULL; list = list->next ())
4653 if ((set = single_set (list->insn ())) == NULL_RTX
4654 || ! REG_P (SET_DEST (set))
4655 || (int) REGNO (SET_DEST (set)) != regno)
4656 return true;
4657 return false;
4660 /* Entry function of LRA constraint pass. Return true if the
4661 constraint pass did change the code. */
4662 bool
4663 lra_constraints (bool first_p)
4665 bool changed_p;
4666 int i, hard_regno, new_insns_num;
4667 unsigned int min_len, new_min_len, uid;
4668 rtx set, x, reg, dest_reg;
4669 basic_block last_bb;
4670 bitmap_iterator bi;
4672 lra_constraint_iter++;
4673 if (lra_dump_file != NULL)
4674 fprintf (lra_dump_file, "\n********** Local #%d: **********\n\n",
4675 lra_constraint_iter);
4676 changed_p = false;
4677 if (pic_offset_table_rtx
4678 && REGNO (pic_offset_table_rtx) >= FIRST_PSEUDO_REGISTER)
4679 lra_risky_transformations_p = true;
4680 else
4681 /* On the first iteration we should check IRA assignment
4682 correctness. In rare cases, the assignments can be wrong as
4683 early clobbers operands are ignored in IRA. */
4684 lra_risky_transformations_p = first_p;
4685 new_insn_uid_start = get_max_uid ();
4686 new_regno_start = first_p ? lra_constraint_new_regno_start : max_reg_num ();
4687 /* Mark used hard regs for target stack size calulations. */
4688 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4689 if (lra_reg_info[i].nrefs != 0
4690 && (hard_regno = lra_get_regno_hard_regno (i)) >= 0)
4692 int j, nregs;
4694 nregs = hard_regno_nregs (hard_regno, lra_reg_info[i].biggest_mode);
4695 for (j = 0; j < nregs; j++)
4696 df_set_regs_ever_live (hard_regno + j, true);
4698 /* Do elimination before the equivalence processing as we can spill
4699 some pseudos during elimination. */
4700 lra_eliminate (false, first_p);
4701 auto_bitmap equiv_insn_bitmap (&reg_obstack);
4702 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4703 if (lra_reg_info[i].nrefs != 0)
4705 ira_reg_equiv[i].profitable_p = true;
4706 reg = regno_reg_rtx[i];
4707 if (lra_get_regno_hard_regno (i) < 0 && (x = get_equiv (reg)) != reg)
4709 bool pseudo_p = contains_reg_p (x, false, false);
4711 /* After RTL transformation, we can not guarantee that
4712 pseudo in the substitution was not reloaded which might
4713 make equivalence invalid. For example, in reverse
4714 equiv of p0
4716 p0 <- ...
4718 equiv_mem <- p0
4720 the memory address register was reloaded before the 2nd
4721 insn. */
4722 if ((! first_p && pseudo_p)
4723 /* We don't use DF for compilation speed sake. So it
4724 is problematic to update live info when we use an
4725 equivalence containing pseudos in more than one
4726 BB. */
4727 || (pseudo_p && multi_block_pseudo_p (i))
4728 /* If an init insn was deleted for some reason, cancel
4729 the equiv. We could update the equiv insns after
4730 transformations including an equiv insn deletion
4731 but it is not worthy as such cases are extremely
4732 rare. */
4733 || contains_deleted_insn_p (ira_reg_equiv[i].init_insns)
4734 /* If it is not a reverse equivalence, we check that a
4735 pseudo in rhs of the init insn is not dying in the
4736 insn. Otherwise, the live info at the beginning of
4737 the corresponding BB might be wrong after we
4738 removed the insn. When the equiv can be a
4739 constant, the right hand side of the init insn can
4740 be a pseudo. */
4741 || (! reverse_equiv_p (i)
4742 && (init_insn_rhs_dead_pseudo_p (i)
4743 /* If we reloaded the pseudo in an equivalence
4744 init insn, we can not remove the equiv init
4745 insns and the init insns might write into
4746 const memory in this case. */
4747 || contains_reloaded_insn_p (i)))
4748 /* Prevent access beyond equivalent memory for
4749 paradoxical subregs. */
4750 || (MEM_P (x)
4751 && maybe_gt (GET_MODE_SIZE (lra_reg_info[i].biggest_mode),
4752 GET_MODE_SIZE (GET_MODE (x))))
4753 || (pic_offset_table_rtx
4754 && ((CONST_POOL_OK_P (PSEUDO_REGNO_MODE (i), x)
4755 && (targetm.preferred_reload_class
4756 (x, lra_get_allocno_class (i)) == NO_REGS))
4757 || contains_symbol_ref_p (x))))
4758 ira_reg_equiv[i].defined_p = false;
4759 if (contains_reg_p (x, false, true))
4760 ira_reg_equiv[i].profitable_p = false;
4761 if (get_equiv (reg) != reg)
4762 bitmap_ior_into (equiv_insn_bitmap, &lra_reg_info[i].insn_bitmap);
4765 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4766 update_equiv (i);
4767 /* We should add all insns containing pseudos which should be
4768 substituted by their equivalences. */
4769 EXECUTE_IF_SET_IN_BITMAP (equiv_insn_bitmap, 0, uid, bi)
4770 lra_push_insn_by_uid (uid);
4771 min_len = lra_insn_stack_length ();
4772 new_insns_num = 0;
4773 last_bb = NULL;
4774 changed_p = false;
4775 while ((new_min_len = lra_insn_stack_length ()) != 0)
4777 curr_insn = lra_pop_insn ();
4778 --new_min_len;
4779 curr_bb = BLOCK_FOR_INSN (curr_insn);
4780 if (curr_bb != last_bb)
4782 last_bb = curr_bb;
4783 bb_reload_num = lra_curr_reload_num;
4785 if (min_len > new_min_len)
4787 min_len = new_min_len;
4788 new_insns_num = 0;
4790 if (new_insns_num > MAX_RELOAD_INSNS_NUMBER)
4791 internal_error
4792 ("Max. number of generated reload insns per insn is achieved (%d)\n",
4793 MAX_RELOAD_INSNS_NUMBER);
4794 new_insns_num++;
4795 if (DEBUG_INSN_P (curr_insn))
4797 /* We need to check equivalence in debug insn and change
4798 pseudo to the equivalent value if necessary. */
4799 curr_id = lra_get_insn_recog_data (curr_insn);
4800 if (bitmap_bit_p (equiv_insn_bitmap, INSN_UID (curr_insn)))
4802 rtx old = *curr_id->operand_loc[0];
4803 *curr_id->operand_loc[0]
4804 = simplify_replace_fn_rtx (old, NULL_RTX,
4805 loc_equivalence_callback, curr_insn);
4806 if (old != *curr_id->operand_loc[0])
4808 lra_update_insn_regno_info (curr_insn);
4809 changed_p = true;
4813 else if (INSN_P (curr_insn))
4815 if ((set = single_set (curr_insn)) != NULL_RTX)
4817 dest_reg = SET_DEST (set);
4818 /* The equivalence pseudo could be set up as SUBREG in a
4819 case when it is a call restore insn in a mode
4820 different from the pseudo mode. */
4821 if (GET_CODE (dest_reg) == SUBREG)
4822 dest_reg = SUBREG_REG (dest_reg);
4823 if ((REG_P (dest_reg)
4824 && (x = get_equiv (dest_reg)) != dest_reg
4825 /* Remove insns which set up a pseudo whose value
4826 can not be changed. Such insns might be not in
4827 init_insns because we don't update equiv data
4828 during insn transformations.
4830 As an example, let suppose that a pseudo got
4831 hard register and on the 1st pass was not
4832 changed to equivalent constant. We generate an
4833 additional insn setting up the pseudo because of
4834 secondary memory movement. Then the pseudo is
4835 spilled and we use the equiv constant. In this
4836 case we should remove the additional insn and
4837 this insn is not init_insns list. */
4838 && (! MEM_P (x) || MEM_READONLY_P (x)
4839 /* Check that this is actually an insn setting
4840 up the equivalence. */
4841 || in_list_p (curr_insn,
4842 ira_reg_equiv
4843 [REGNO (dest_reg)].init_insns)))
4844 || (((x = get_equiv (SET_SRC (set))) != SET_SRC (set))
4845 && in_list_p (curr_insn,
4846 ira_reg_equiv
4847 [REGNO (SET_SRC (set))].init_insns)))
4849 /* This is equiv init insn of pseudo which did not get a
4850 hard register -- remove the insn. */
4851 if (lra_dump_file != NULL)
4853 fprintf (lra_dump_file,
4854 " Removing equiv init insn %i (freq=%d)\n",
4855 INSN_UID (curr_insn),
4856 REG_FREQ_FROM_BB (BLOCK_FOR_INSN (curr_insn)));
4857 dump_insn_slim (lra_dump_file, curr_insn);
4859 if (contains_reg_p (x, true, false))
4860 lra_risky_transformations_p = true;
4861 lra_set_insn_deleted (curr_insn);
4862 continue;
4865 curr_id = lra_get_insn_recog_data (curr_insn);
4866 curr_static_id = curr_id->insn_static_data;
4867 init_curr_insn_input_reloads ();
4868 init_curr_operand_mode ();
4869 if (curr_insn_transform (false))
4870 changed_p = true;
4871 /* Check non-transformed insns too for equiv change as USE
4872 or CLOBBER don't need reloads but can contain pseudos
4873 being changed on their equivalences. */
4874 else if (bitmap_bit_p (equiv_insn_bitmap, INSN_UID (curr_insn))
4875 && loc_equivalence_change_p (&PATTERN (curr_insn)))
4877 lra_update_insn_regno_info (curr_insn);
4878 changed_p = true;
4883 /* If we used a new hard regno, changed_p should be true because the
4884 hard reg is assigned to a new pseudo. */
4885 if (flag_checking && !changed_p)
4887 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4888 if (lra_reg_info[i].nrefs != 0
4889 && (hard_regno = lra_get_regno_hard_regno (i)) >= 0)
4891 int j, nregs = hard_regno_nregs (hard_regno,
4892 PSEUDO_REGNO_MODE (i));
4894 for (j = 0; j < nregs; j++)
4895 lra_assert (df_regs_ever_live_p (hard_regno + j));
4898 return changed_p;
4901 static void initiate_invariants (void);
4902 static void finish_invariants (void);
4904 /* Initiate the LRA constraint pass. It is done once per
4905 function. */
4906 void
4907 lra_constraints_init (void)
4909 initiate_invariants ();
4912 /* Finalize the LRA constraint pass. It is done once per
4913 function. */
4914 void
4915 lra_constraints_finish (void)
4917 finish_invariants ();
4922 /* Structure describes invariants for ineheritance. */
4923 struct lra_invariant
4925 /* The order number of the invariant. */
4926 int num;
4927 /* The invariant RTX. */
4928 rtx invariant_rtx;
4929 /* The origin insn of the invariant. */
4930 rtx_insn *insn;
4933 typedef lra_invariant invariant_t;
4934 typedef invariant_t *invariant_ptr_t;
4935 typedef const invariant_t *const_invariant_ptr_t;
4937 /* Pointer to the inheritance invariants. */
4938 static vec<invariant_ptr_t> invariants;
4940 /* Allocation pool for the invariants. */
4941 static object_allocator<lra_invariant> *invariants_pool;
4943 /* Hash table for the invariants. */
4944 static htab_t invariant_table;
4946 /* Hash function for INVARIANT. */
4947 static hashval_t
4948 invariant_hash (const void *invariant)
4950 rtx inv = ((const_invariant_ptr_t) invariant)->invariant_rtx;
4951 return lra_rtx_hash (inv);
4954 /* Equal function for invariants INVARIANT1 and INVARIANT2. */
4955 static int
4956 invariant_eq_p (const void *invariant1, const void *invariant2)
4958 rtx inv1 = ((const_invariant_ptr_t) invariant1)->invariant_rtx;
4959 rtx inv2 = ((const_invariant_ptr_t) invariant2)->invariant_rtx;
4961 return rtx_equal_p (inv1, inv2);
4964 /* Insert INVARIANT_RTX into the table if it is not there yet. Return
4965 invariant which is in the table. */
4966 static invariant_ptr_t
4967 insert_invariant (rtx invariant_rtx)
4969 void **entry_ptr;
4970 invariant_t invariant;
4971 invariant_ptr_t invariant_ptr;
4973 invariant.invariant_rtx = invariant_rtx;
4974 entry_ptr = htab_find_slot (invariant_table, &invariant, INSERT);
4975 if (*entry_ptr == NULL)
4977 invariant_ptr = invariants_pool->allocate ();
4978 invariant_ptr->invariant_rtx = invariant_rtx;
4979 invariant_ptr->insn = NULL;
4980 invariants.safe_push (invariant_ptr);
4981 *entry_ptr = (void *) invariant_ptr;
4983 return (invariant_ptr_t) *entry_ptr;
4986 /* Initiate the invariant table. */
4987 static void
4988 initiate_invariants (void)
4990 invariants.create (100);
4991 invariants_pool
4992 = new object_allocator<lra_invariant> ("Inheritance invariants");
4993 invariant_table = htab_create (100, invariant_hash, invariant_eq_p, NULL);
4996 /* Finish the invariant table. */
4997 static void
4998 finish_invariants (void)
5000 htab_delete (invariant_table);
5001 delete invariants_pool;
5002 invariants.release ();
5005 /* Make the invariant table empty. */
5006 static void
5007 clear_invariants (void)
5009 htab_empty (invariant_table);
5010 invariants_pool->release ();
5011 invariants.truncate (0);
5016 /* This page contains code to do inheritance/split
5017 transformations. */
5019 /* Number of reloads passed so far in current EBB. */
5020 static int reloads_num;
5022 /* Number of calls passed so far in current EBB. */
5023 static int calls_num;
5025 /* Current reload pseudo check for validity of elements in
5026 USAGE_INSNS. */
5027 static int curr_usage_insns_check;
5029 /* Info about last usage of registers in EBB to do inheritance/split
5030 transformation. Inheritance transformation is done from a spilled
5031 pseudo and split transformations from a hard register or a pseudo
5032 assigned to a hard register. */
5033 struct usage_insns
5035 /* If the value is equal to CURR_USAGE_INSNS_CHECK, then the member
5036 value INSNS is valid. The insns is chain of optional debug insns
5037 and a finishing non-debug insn using the corresponding reg. The
5038 value is also used to mark the registers which are set up in the
5039 current insn. The negated insn uid is used for this. */
5040 int check;
5041 /* Value of global reloads_num at the last insn in INSNS. */
5042 int reloads_num;
5043 /* Value of global reloads_nums at the last insn in INSNS. */
5044 int calls_num;
5045 /* It can be true only for splitting. And it means that the restore
5046 insn should be put after insn given by the following member. */
5047 bool after_p;
5048 /* Next insns in the current EBB which use the original reg and the
5049 original reg value is not changed between the current insn and
5050 the next insns. In order words, e.g. for inheritance, if we need
5051 to use the original reg value again in the next insns we can try
5052 to use the value in a hard register from a reload insn of the
5053 current insn. */
5054 rtx insns;
5057 /* Map: regno -> corresponding pseudo usage insns. */
5058 static struct usage_insns *usage_insns;
5060 static void
5061 setup_next_usage_insn (int regno, rtx insn, int reloads_num, bool after_p)
5063 usage_insns[regno].check = curr_usage_insns_check;
5064 usage_insns[regno].insns = insn;
5065 usage_insns[regno].reloads_num = reloads_num;
5066 usage_insns[regno].calls_num = calls_num;
5067 usage_insns[regno].after_p = after_p;
5070 /* The function is used to form list REGNO usages which consists of
5071 optional debug insns finished by a non-debug insn using REGNO.
5072 RELOADS_NUM is current number of reload insns processed so far. */
5073 static void
5074 add_next_usage_insn (int regno, rtx_insn *insn, int reloads_num)
5076 rtx next_usage_insns;
5078 if (usage_insns[regno].check == curr_usage_insns_check
5079 && (next_usage_insns = usage_insns[regno].insns) != NULL_RTX
5080 && DEBUG_INSN_P (insn))
5082 /* Check that we did not add the debug insn yet. */
5083 if (next_usage_insns != insn
5084 && (GET_CODE (next_usage_insns) != INSN_LIST
5085 || XEXP (next_usage_insns, 0) != insn))
5086 usage_insns[regno].insns = gen_rtx_INSN_LIST (VOIDmode, insn,
5087 next_usage_insns);
5089 else if (NONDEBUG_INSN_P (insn))
5090 setup_next_usage_insn (regno, insn, reloads_num, false);
5091 else
5092 usage_insns[regno].check = 0;
5095 /* Return first non-debug insn in list USAGE_INSNS. */
5096 static rtx_insn *
5097 skip_usage_debug_insns (rtx usage_insns)
5099 rtx insn;
5101 /* Skip debug insns. */
5102 for (insn = usage_insns;
5103 insn != NULL_RTX && GET_CODE (insn) == INSN_LIST;
5104 insn = XEXP (insn, 1))
5106 return safe_as_a <rtx_insn *> (insn);
5109 /* Return true if we need secondary memory moves for insn in
5110 USAGE_INSNS after inserting inherited pseudo of class INHER_CL
5111 into the insn. */
5112 static bool
5113 check_secondary_memory_needed_p (enum reg_class inher_cl ATTRIBUTE_UNUSED,
5114 rtx usage_insns ATTRIBUTE_UNUSED)
5116 rtx_insn *insn;
5117 rtx set, dest;
5118 enum reg_class cl;
5120 if (inher_cl == ALL_REGS
5121 || (insn = skip_usage_debug_insns (usage_insns)) == NULL_RTX)
5122 return false;
5123 lra_assert (INSN_P (insn));
5124 if ((set = single_set (insn)) == NULL_RTX || ! REG_P (SET_DEST (set)))
5125 return false;
5126 dest = SET_DEST (set);
5127 if (! REG_P (dest))
5128 return false;
5129 lra_assert (inher_cl != NO_REGS);
5130 cl = get_reg_class (REGNO (dest));
5131 return (cl != NO_REGS && cl != ALL_REGS
5132 && targetm.secondary_memory_needed (GET_MODE (dest), inher_cl, cl));
5135 /* Registers involved in inheritance/split in the current EBB
5136 (inheritance/split pseudos and original registers). */
5137 static bitmap_head check_only_regs;
5139 /* Reload pseudos can not be involded in invariant inheritance in the
5140 current EBB. */
5141 static bitmap_head invalid_invariant_regs;
5143 /* Do inheritance transformations for insn INSN, which defines (if
5144 DEF_P) or uses ORIGINAL_REGNO. NEXT_USAGE_INSNS specifies which
5145 instruction in the EBB next uses ORIGINAL_REGNO; it has the same
5146 form as the "insns" field of usage_insns. Return true if we
5147 succeed in such transformation.
5149 The transformations look like:
5151 p <- ... i <- ...
5152 ... p <- i (new insn)
5153 ... =>
5154 <- ... p ... <- ... i ...
5156 ... i <- p (new insn)
5157 <- ... p ... <- ... i ...
5158 ... =>
5159 <- ... p ... <- ... i ...
5160 where p is a spilled original pseudo and i is a new inheritance pseudo.
5163 The inheritance pseudo has the smallest class of two classes CL and
5164 class of ORIGINAL REGNO. */
5165 static bool
5166 inherit_reload_reg (bool def_p, int original_regno,
5167 enum reg_class cl, rtx_insn *insn, rtx next_usage_insns)
5169 if (optimize_function_for_size_p (cfun))
5170 return false;
5172 enum reg_class rclass = lra_get_allocno_class (original_regno);
5173 rtx original_reg = regno_reg_rtx[original_regno];
5174 rtx new_reg, usage_insn;
5175 rtx_insn *new_insns;
5177 lra_assert (! usage_insns[original_regno].after_p);
5178 if (lra_dump_file != NULL)
5179 fprintf (lra_dump_file,
5180 " <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<\n");
5181 if (! ira_reg_classes_intersect_p[cl][rclass])
5183 if (lra_dump_file != NULL)
5185 fprintf (lra_dump_file,
5186 " Rejecting inheritance for %d "
5187 "because of disjoint classes %s and %s\n",
5188 original_regno, reg_class_names[cl],
5189 reg_class_names[rclass]);
5190 fprintf (lra_dump_file,
5191 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5193 return false;
5195 if ((ira_class_subset_p[cl][rclass] && cl != rclass)
5196 /* We don't use a subset of two classes because it can be
5197 NO_REGS. This transformation is still profitable in most
5198 cases even if the classes are not intersected as register
5199 move is probably cheaper than a memory load. */
5200 || ira_class_hard_regs_num[cl] < ira_class_hard_regs_num[rclass])
5202 if (lra_dump_file != NULL)
5203 fprintf (lra_dump_file, " Use smallest class of %s and %s\n",
5204 reg_class_names[cl], reg_class_names[rclass]);
5206 rclass = cl;
5208 if (check_secondary_memory_needed_p (rclass, next_usage_insns))
5210 /* Reject inheritance resulting in secondary memory moves.
5211 Otherwise, there is a danger in LRA cycling. Also such
5212 transformation will be unprofitable. */
5213 if (lra_dump_file != NULL)
5215 rtx_insn *insn = skip_usage_debug_insns (next_usage_insns);
5216 rtx set = single_set (insn);
5218 lra_assert (set != NULL_RTX);
5220 rtx dest = SET_DEST (set);
5222 lra_assert (REG_P (dest));
5223 fprintf (lra_dump_file,
5224 " Rejecting inheritance for insn %d(%s)<-%d(%s) "
5225 "as secondary mem is needed\n",
5226 REGNO (dest), reg_class_names[get_reg_class (REGNO (dest))],
5227 original_regno, reg_class_names[rclass]);
5228 fprintf (lra_dump_file,
5229 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5231 return false;
5233 new_reg = lra_create_new_reg (GET_MODE (original_reg), original_reg,
5234 rclass, "inheritance");
5235 start_sequence ();
5236 if (def_p)
5237 lra_emit_move (original_reg, new_reg);
5238 else
5239 lra_emit_move (new_reg, original_reg);
5240 new_insns = get_insns ();
5241 end_sequence ();
5242 if (NEXT_INSN (new_insns) != NULL_RTX)
5244 if (lra_dump_file != NULL)
5246 fprintf (lra_dump_file,
5247 " Rejecting inheritance %d->%d "
5248 "as it results in 2 or more insns:\n",
5249 original_regno, REGNO (new_reg));
5250 dump_rtl_slim (lra_dump_file, new_insns, NULL, -1, 0);
5251 fprintf (lra_dump_file,
5252 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5254 return false;
5256 lra_substitute_pseudo_within_insn (insn, original_regno, new_reg, false);
5257 lra_update_insn_regno_info (insn);
5258 if (! def_p)
5259 /* We now have a new usage insn for original regno. */
5260 setup_next_usage_insn (original_regno, new_insns, reloads_num, false);
5261 if (lra_dump_file != NULL)
5262 fprintf (lra_dump_file, " Original reg change %d->%d (bb%d):\n",
5263 original_regno, REGNO (new_reg), BLOCK_FOR_INSN (insn)->index);
5264 lra_reg_info[REGNO (new_reg)].restore_rtx = regno_reg_rtx[original_regno];
5265 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
5266 bitmap_set_bit (&check_only_regs, original_regno);
5267 bitmap_set_bit (&lra_inheritance_pseudos, REGNO (new_reg));
5268 if (def_p)
5269 lra_process_new_insns (insn, NULL, new_insns,
5270 "Add original<-inheritance");
5271 else
5272 lra_process_new_insns (insn, new_insns, NULL,
5273 "Add inheritance<-original");
5274 while (next_usage_insns != NULL_RTX)
5276 if (GET_CODE (next_usage_insns) != INSN_LIST)
5278 usage_insn = next_usage_insns;
5279 lra_assert (NONDEBUG_INSN_P (usage_insn));
5280 next_usage_insns = NULL;
5282 else
5284 usage_insn = XEXP (next_usage_insns, 0);
5285 lra_assert (DEBUG_INSN_P (usage_insn));
5286 next_usage_insns = XEXP (next_usage_insns, 1);
5288 lra_substitute_pseudo (&usage_insn, original_regno, new_reg, false);
5289 lra_update_insn_regno_info (as_a <rtx_insn *> (usage_insn));
5290 if (lra_dump_file != NULL)
5292 basic_block bb = BLOCK_FOR_INSN (usage_insn);
5293 fprintf (lra_dump_file,
5294 " Inheritance reuse change %d->%d (bb%d):\n",
5295 original_regno, REGNO (new_reg),
5296 bb ? bb->index : -1);
5297 dump_insn_slim (lra_dump_file, as_a <rtx_insn *> (usage_insn));
5300 if (lra_dump_file != NULL)
5301 fprintf (lra_dump_file,
5302 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5303 return true;
5306 /* Return true if we need a caller save/restore for pseudo REGNO which
5307 was assigned to a hard register. */
5308 static inline bool
5309 need_for_call_save_p (int regno)
5311 lra_assert (regno >= FIRST_PSEUDO_REGISTER && reg_renumber[regno] >= 0);
5312 return (usage_insns[regno].calls_num < calls_num
5313 && (overlaps_hard_reg_set_p
5314 ((flag_ipa_ra &&
5315 ! hard_reg_set_empty_p (lra_reg_info[regno].actual_call_used_reg_set))
5316 ? lra_reg_info[regno].actual_call_used_reg_set
5317 : call_used_reg_set,
5318 PSEUDO_REGNO_MODE (regno), reg_renumber[regno])
5319 || (targetm.hard_regno_call_part_clobbered
5320 (reg_renumber[regno], PSEUDO_REGNO_MODE (regno)))));
5323 /* Global registers occurring in the current EBB. */
5324 static bitmap_head ebb_global_regs;
5326 /* Return true if we need a split for hard register REGNO or pseudo
5327 REGNO which was assigned to a hard register.
5328 POTENTIAL_RELOAD_HARD_REGS contains hard registers which might be
5329 used for reloads since the EBB end. It is an approximation of the
5330 used hard registers in the split range. The exact value would
5331 require expensive calculations. If we were aggressive with
5332 splitting because of the approximation, the split pseudo will save
5333 the same hard register assignment and will be removed in the undo
5334 pass. We still need the approximation because too aggressive
5335 splitting would result in too inaccurate cost calculation in the
5336 assignment pass because of too many generated moves which will be
5337 probably removed in the undo pass. */
5338 static inline bool
5339 need_for_split_p (HARD_REG_SET potential_reload_hard_regs, int regno)
5341 int hard_regno = regno < FIRST_PSEUDO_REGISTER ? regno : reg_renumber[regno];
5343 lra_assert (hard_regno >= 0);
5344 return ((TEST_HARD_REG_BIT (potential_reload_hard_regs, hard_regno)
5345 /* Don't split eliminable hard registers, otherwise we can
5346 split hard registers like hard frame pointer, which
5347 lives on BB start/end according to DF-infrastructure,
5348 when there is a pseudo assigned to the register and
5349 living in the same BB. */
5350 && (regno >= FIRST_PSEUDO_REGISTER
5351 || ! TEST_HARD_REG_BIT (eliminable_regset, hard_regno))
5352 && ! TEST_HARD_REG_BIT (lra_no_alloc_regs, hard_regno)
5353 /* Don't split call clobbered hard regs living through
5354 calls, otherwise we might have a check problem in the
5355 assign sub-pass as in the most cases (exception is a
5356 situation when lra_risky_transformations_p value is
5357 true) the assign pass assumes that all pseudos living
5358 through calls are assigned to call saved hard regs. */
5359 && (regno >= FIRST_PSEUDO_REGISTER
5360 || ! TEST_HARD_REG_BIT (call_used_reg_set, regno)
5361 || usage_insns[regno].calls_num == calls_num)
5362 /* We need at least 2 reloads to make pseudo splitting
5363 profitable. We should provide hard regno splitting in
5364 any case to solve 1st insn scheduling problem when
5365 moving hard register definition up might result in
5366 impossibility to find hard register for reload pseudo of
5367 small register class. */
5368 && (usage_insns[regno].reloads_num
5369 + (regno < FIRST_PSEUDO_REGISTER ? 0 : 3) < reloads_num)
5370 && (regno < FIRST_PSEUDO_REGISTER
5371 /* For short living pseudos, spilling + inheritance can
5372 be considered a substitution for splitting.
5373 Therefore we do not splitting for local pseudos. It
5374 decreases also aggressiveness of splitting. The
5375 minimal number of references is chosen taking into
5376 account that for 2 references splitting has no sense
5377 as we can just spill the pseudo. */
5378 || (regno >= FIRST_PSEUDO_REGISTER
5379 && lra_reg_info[regno].nrefs > 3
5380 && bitmap_bit_p (&ebb_global_regs, regno))))
5381 || (regno >= FIRST_PSEUDO_REGISTER && need_for_call_save_p (regno)));
5384 /* Return class for the split pseudo created from original pseudo with
5385 ALLOCNO_CLASS and MODE which got a hard register HARD_REGNO. We
5386 choose subclass of ALLOCNO_CLASS which contains HARD_REGNO and
5387 results in no secondary memory movements. */
5388 static enum reg_class
5389 choose_split_class (enum reg_class allocno_class,
5390 int hard_regno ATTRIBUTE_UNUSED,
5391 machine_mode mode ATTRIBUTE_UNUSED)
5393 int i;
5394 enum reg_class cl, best_cl = NO_REGS;
5395 enum reg_class hard_reg_class ATTRIBUTE_UNUSED
5396 = REGNO_REG_CLASS (hard_regno);
5398 if (! targetm.secondary_memory_needed (mode, allocno_class, allocno_class)
5399 && TEST_HARD_REG_BIT (reg_class_contents[allocno_class], hard_regno))
5400 return allocno_class;
5401 for (i = 0;
5402 (cl = reg_class_subclasses[allocno_class][i]) != LIM_REG_CLASSES;
5403 i++)
5404 if (! targetm.secondary_memory_needed (mode, cl, hard_reg_class)
5405 && ! targetm.secondary_memory_needed (mode, hard_reg_class, cl)
5406 && TEST_HARD_REG_BIT (reg_class_contents[cl], hard_regno)
5407 && (best_cl == NO_REGS
5408 || ira_class_hard_regs_num[best_cl] < ira_class_hard_regs_num[cl]))
5409 best_cl = cl;
5410 return best_cl;
5413 /* Copy any equivalence information from ORIGINAL_REGNO to NEW_REGNO.
5414 It only makes sense to call this function if NEW_REGNO is always
5415 equal to ORIGINAL_REGNO. */
5417 static void
5418 lra_copy_reg_equiv (unsigned int new_regno, unsigned int original_regno)
5420 if (!ira_reg_equiv[original_regno].defined_p)
5421 return;
5423 ira_expand_reg_equiv ();
5424 ira_reg_equiv[new_regno].defined_p = true;
5425 if (ira_reg_equiv[original_regno].memory)
5426 ira_reg_equiv[new_regno].memory
5427 = copy_rtx (ira_reg_equiv[original_regno].memory);
5428 if (ira_reg_equiv[original_regno].constant)
5429 ira_reg_equiv[new_regno].constant
5430 = copy_rtx (ira_reg_equiv[original_regno].constant);
5431 if (ira_reg_equiv[original_regno].invariant)
5432 ira_reg_equiv[new_regno].invariant
5433 = copy_rtx (ira_reg_equiv[original_regno].invariant);
5436 /* Do split transformations for insn INSN, which defines or uses
5437 ORIGINAL_REGNO. NEXT_USAGE_INSNS specifies which instruction in
5438 the EBB next uses ORIGINAL_REGNO; it has the same form as the
5439 "insns" field of usage_insns.
5441 The transformations look like:
5443 p <- ... p <- ...
5444 ... s <- p (new insn -- save)
5445 ... =>
5446 ... p <- s (new insn -- restore)
5447 <- ... p ... <- ... p ...
5449 <- ... p ... <- ... p ...
5450 ... s <- p (new insn -- save)
5451 ... =>
5452 ... p <- s (new insn -- restore)
5453 <- ... p ... <- ... p ...
5455 where p is an original pseudo got a hard register or a hard
5456 register and s is a new split pseudo. The save is put before INSN
5457 if BEFORE_P is true. Return true if we succeed in such
5458 transformation. */
5459 static bool
5460 split_reg (bool before_p, int original_regno, rtx_insn *insn,
5461 rtx next_usage_insns)
5463 enum reg_class rclass;
5464 rtx original_reg;
5465 int hard_regno, nregs;
5466 rtx new_reg, usage_insn;
5467 rtx_insn *restore, *save;
5468 bool after_p;
5469 bool call_save_p;
5470 machine_mode mode;
5472 if (original_regno < FIRST_PSEUDO_REGISTER)
5474 rclass = ira_allocno_class_translate[REGNO_REG_CLASS (original_regno)];
5475 hard_regno = original_regno;
5476 call_save_p = false;
5477 nregs = 1;
5478 mode = lra_reg_info[hard_regno].biggest_mode;
5479 machine_mode reg_rtx_mode = GET_MODE (regno_reg_rtx[hard_regno]);
5480 /* A reg can have a biggest_mode of VOIDmode if it was only ever seen
5481 as part of a multi-word register. In that case, or if the biggest
5482 mode was larger than a register, just use the reg_rtx. Otherwise,
5483 limit the size to that of the biggest access in the function. */
5484 if (mode == VOIDmode
5485 || paradoxical_subreg_p (mode, reg_rtx_mode))
5487 original_reg = regno_reg_rtx[hard_regno];
5488 mode = reg_rtx_mode;
5490 else
5491 original_reg = gen_rtx_REG (mode, hard_regno);
5493 else
5495 mode = PSEUDO_REGNO_MODE (original_regno);
5496 hard_regno = reg_renumber[original_regno];
5497 nregs = hard_regno_nregs (hard_regno, mode);
5498 rclass = lra_get_allocno_class (original_regno);
5499 original_reg = regno_reg_rtx[original_regno];
5500 call_save_p = need_for_call_save_p (original_regno);
5502 lra_assert (hard_regno >= 0);
5503 if (lra_dump_file != NULL)
5504 fprintf (lra_dump_file,
5505 " ((((((((((((((((((((((((((((((((((((((((((((((((\n");
5507 if (call_save_p)
5509 mode = HARD_REGNO_CALLER_SAVE_MODE (hard_regno,
5510 hard_regno_nregs (hard_regno, mode),
5511 mode);
5512 new_reg = lra_create_new_reg (mode, NULL_RTX, NO_REGS, "save");
5514 else
5516 rclass = choose_split_class (rclass, hard_regno, mode);
5517 if (rclass == NO_REGS)
5519 if (lra_dump_file != NULL)
5521 fprintf (lra_dump_file,
5522 " Rejecting split of %d(%s): "
5523 "no good reg class for %d(%s)\n",
5524 original_regno,
5525 reg_class_names[lra_get_allocno_class (original_regno)],
5526 hard_regno,
5527 reg_class_names[REGNO_REG_CLASS (hard_regno)]);
5528 fprintf
5529 (lra_dump_file,
5530 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5532 return false;
5534 /* Split_if_necessary can split hard registers used as part of a
5535 multi-register mode but splits each register individually. The
5536 mode used for each independent register may not be supported
5537 so reject the split. Splitting the wider mode should theoretically
5538 be possible but is not implemented. */
5539 if (!targetm.hard_regno_mode_ok (hard_regno, mode))
5541 if (lra_dump_file != NULL)
5543 fprintf (lra_dump_file,
5544 " Rejecting split of %d(%s): unsuitable mode %s\n",
5545 original_regno,
5546 reg_class_names[lra_get_allocno_class (original_regno)],
5547 GET_MODE_NAME (mode));
5548 fprintf
5549 (lra_dump_file,
5550 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5552 return false;
5554 new_reg = lra_create_new_reg (mode, original_reg, rclass, "split");
5555 reg_renumber[REGNO (new_reg)] = hard_regno;
5557 int new_regno = REGNO (new_reg);
5558 save = emit_spill_move (true, new_reg, original_reg);
5559 if (NEXT_INSN (save) != NULL_RTX && !call_save_p)
5561 if (lra_dump_file != NULL)
5563 fprintf
5564 (lra_dump_file,
5565 " Rejecting split %d->%d resulting in > 2 save insns:\n",
5566 original_regno, new_regno);
5567 dump_rtl_slim (lra_dump_file, save, NULL, -1, 0);
5568 fprintf (lra_dump_file,
5569 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5571 return false;
5573 restore = emit_spill_move (false, new_reg, original_reg);
5574 if (NEXT_INSN (restore) != NULL_RTX && !call_save_p)
5576 if (lra_dump_file != NULL)
5578 fprintf (lra_dump_file,
5579 " Rejecting split %d->%d "
5580 "resulting in > 2 restore insns:\n",
5581 original_regno, new_regno);
5582 dump_rtl_slim (lra_dump_file, restore, NULL, -1, 0);
5583 fprintf (lra_dump_file,
5584 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5586 return false;
5588 /* Transfer equivalence information to the spill register, so that
5589 if we fail to allocate the spill register, we have the option of
5590 rematerializing the original value instead of spilling to the stack. */
5591 if (!HARD_REGISTER_NUM_P (original_regno)
5592 && mode == PSEUDO_REGNO_MODE (original_regno))
5593 lra_copy_reg_equiv (new_regno, original_regno);
5594 after_p = usage_insns[original_regno].after_p;
5595 lra_reg_info[new_regno].restore_rtx = regno_reg_rtx[original_regno];
5596 bitmap_set_bit (&check_only_regs, new_regno);
5597 bitmap_set_bit (&check_only_regs, original_regno);
5598 bitmap_set_bit (&lra_split_regs, new_regno);
5599 for (;;)
5601 if (GET_CODE (next_usage_insns) != INSN_LIST)
5603 usage_insn = next_usage_insns;
5604 break;
5606 usage_insn = XEXP (next_usage_insns, 0);
5607 lra_assert (DEBUG_INSN_P (usage_insn));
5608 next_usage_insns = XEXP (next_usage_insns, 1);
5609 lra_substitute_pseudo (&usage_insn, original_regno, new_reg, false);
5610 lra_update_insn_regno_info (as_a <rtx_insn *> (usage_insn));
5611 if (lra_dump_file != NULL)
5613 fprintf (lra_dump_file, " Split reuse change %d->%d:\n",
5614 original_regno, new_regno);
5615 dump_insn_slim (lra_dump_file, as_a <rtx_insn *> (usage_insn));
5618 lra_assert (NOTE_P (usage_insn) || NONDEBUG_INSN_P (usage_insn));
5619 lra_assert (usage_insn != insn || (after_p && before_p));
5620 lra_process_new_insns (as_a <rtx_insn *> (usage_insn),
5621 after_p ? NULL : restore,
5622 after_p ? restore : NULL,
5623 call_save_p
5624 ? "Add reg<-save" : "Add reg<-split");
5625 lra_process_new_insns (insn, before_p ? save : NULL,
5626 before_p ? NULL : save,
5627 call_save_p
5628 ? "Add save<-reg" : "Add split<-reg");
5629 if (nregs > 1)
5630 /* If we are trying to split multi-register. We should check
5631 conflicts on the next assignment sub-pass. IRA can allocate on
5632 sub-register levels, LRA do this on pseudos level right now and
5633 this discrepancy may create allocation conflicts after
5634 splitting. */
5635 lra_risky_transformations_p = true;
5636 if (lra_dump_file != NULL)
5637 fprintf (lra_dump_file,
5638 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5639 return true;
5642 /* Recognize that we need a split transformation for insn INSN, which
5643 defines or uses REGNO in its insn biggest MODE (we use it only if
5644 REGNO is a hard register). POTENTIAL_RELOAD_HARD_REGS contains
5645 hard registers which might be used for reloads since the EBB end.
5646 Put the save before INSN if BEFORE_P is true. MAX_UID is maximla
5647 uid before starting INSN processing. Return true if we succeed in
5648 such transformation. */
5649 static bool
5650 split_if_necessary (int regno, machine_mode mode,
5651 HARD_REG_SET potential_reload_hard_regs,
5652 bool before_p, rtx_insn *insn, int max_uid)
5654 bool res = false;
5655 int i, nregs = 1;
5656 rtx next_usage_insns;
5658 if (regno < FIRST_PSEUDO_REGISTER)
5659 nregs = hard_regno_nregs (regno, mode);
5660 for (i = 0; i < nregs; i++)
5661 if (usage_insns[regno + i].check == curr_usage_insns_check
5662 && (next_usage_insns = usage_insns[regno + i].insns) != NULL_RTX
5663 /* To avoid processing the register twice or more. */
5664 && ((GET_CODE (next_usage_insns) != INSN_LIST
5665 && INSN_UID (next_usage_insns) < max_uid)
5666 || (GET_CODE (next_usage_insns) == INSN_LIST
5667 && (INSN_UID (XEXP (next_usage_insns, 0)) < max_uid)))
5668 && need_for_split_p (potential_reload_hard_regs, regno + i)
5669 && split_reg (before_p, regno + i, insn, next_usage_insns))
5670 res = true;
5671 return res;
5674 /* Return TRUE if rtx X is considered as an invariant for
5675 inheritance. */
5676 static bool
5677 invariant_p (const_rtx x)
5679 machine_mode mode;
5680 const char *fmt;
5681 enum rtx_code code;
5682 int i, j;
5684 code = GET_CODE (x);
5685 mode = GET_MODE (x);
5686 if (code == SUBREG)
5688 x = SUBREG_REG (x);
5689 code = GET_CODE (x);
5690 mode = wider_subreg_mode (mode, GET_MODE (x));
5693 if (MEM_P (x))
5694 return false;
5696 if (REG_P (x))
5698 int i, nregs, regno = REGNO (x);
5700 if (regno >= FIRST_PSEUDO_REGISTER || regno == STACK_POINTER_REGNUM
5701 || TEST_HARD_REG_BIT (eliminable_regset, regno)
5702 || GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
5703 return false;
5704 nregs = hard_regno_nregs (regno, mode);
5705 for (i = 0; i < nregs; i++)
5706 if (! fixed_regs[regno + i]
5707 /* A hard register may be clobbered in the current insn
5708 but we can ignore this case because if the hard
5709 register is used it should be set somewhere after the
5710 clobber. */
5711 || bitmap_bit_p (&invalid_invariant_regs, regno + i))
5712 return false;
5714 fmt = GET_RTX_FORMAT (code);
5715 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
5717 if (fmt[i] == 'e')
5719 if (! invariant_p (XEXP (x, i)))
5720 return false;
5722 else if (fmt[i] == 'E')
5724 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
5725 if (! invariant_p (XVECEXP (x, i, j)))
5726 return false;
5729 return true;
5732 /* We have 'dest_reg <- invariant'. Let us try to make an invariant
5733 inheritance transformation (using dest_reg instead invariant in a
5734 subsequent insn). */
5735 static bool
5736 process_invariant_for_inheritance (rtx dst_reg, rtx invariant_rtx)
5738 invariant_ptr_t invariant_ptr;
5739 rtx_insn *insn, *new_insns;
5740 rtx insn_set, insn_reg, new_reg;
5741 int insn_regno;
5742 bool succ_p = false;
5743 int dst_regno = REGNO (dst_reg);
5744 machine_mode dst_mode = GET_MODE (dst_reg);
5745 enum reg_class cl = lra_get_allocno_class (dst_regno), insn_reg_cl;
5747 invariant_ptr = insert_invariant (invariant_rtx);
5748 if ((insn = invariant_ptr->insn) != NULL_RTX)
5750 /* We have a subsequent insn using the invariant. */
5751 insn_set = single_set (insn);
5752 lra_assert (insn_set != NULL);
5753 insn_reg = SET_DEST (insn_set);
5754 lra_assert (REG_P (insn_reg));
5755 insn_regno = REGNO (insn_reg);
5756 insn_reg_cl = lra_get_allocno_class (insn_regno);
5758 if (dst_mode == GET_MODE (insn_reg)
5759 /* We should consider only result move reg insns which are
5760 cheap. */
5761 && targetm.register_move_cost (dst_mode, cl, insn_reg_cl) == 2
5762 && targetm.register_move_cost (dst_mode, cl, cl) == 2)
5764 if (lra_dump_file != NULL)
5765 fprintf (lra_dump_file,
5766 " [[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[\n");
5767 new_reg = lra_create_new_reg (dst_mode, dst_reg,
5768 cl, "invariant inheritance");
5769 bitmap_set_bit (&lra_inheritance_pseudos, REGNO (new_reg));
5770 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
5771 lra_reg_info[REGNO (new_reg)].restore_rtx = PATTERN (insn);
5772 start_sequence ();
5773 lra_emit_move (new_reg, dst_reg);
5774 new_insns = get_insns ();
5775 end_sequence ();
5776 lra_process_new_insns (curr_insn, NULL, new_insns,
5777 "Add invariant inheritance<-original");
5778 start_sequence ();
5779 lra_emit_move (SET_DEST (insn_set), new_reg);
5780 new_insns = get_insns ();
5781 end_sequence ();
5782 lra_process_new_insns (insn, NULL, new_insns,
5783 "Changing reload<-inheritance");
5784 lra_set_insn_deleted (insn);
5785 succ_p = true;
5786 if (lra_dump_file != NULL)
5788 fprintf (lra_dump_file,
5789 " Invariant inheritance reuse change %d (bb%d):\n",
5790 REGNO (new_reg), BLOCK_FOR_INSN (insn)->index);
5791 dump_insn_slim (lra_dump_file, insn);
5792 fprintf (lra_dump_file,
5793 " ]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]\n");
5797 invariant_ptr->insn = curr_insn;
5798 return succ_p;
5801 /* Check only registers living at the current program point in the
5802 current EBB. */
5803 static bitmap_head live_regs;
5805 /* Update live info in EBB given by its HEAD and TAIL insns after
5806 inheritance/split transformation. The function removes dead moves
5807 too. */
5808 static void
5809 update_ebb_live_info (rtx_insn *head, rtx_insn *tail)
5811 unsigned int j;
5812 int i, regno;
5813 bool live_p;
5814 rtx_insn *prev_insn;
5815 rtx set;
5816 bool remove_p;
5817 basic_block last_bb, prev_bb, curr_bb;
5818 bitmap_iterator bi;
5819 struct lra_insn_reg *reg;
5820 edge e;
5821 edge_iterator ei;
5823 last_bb = BLOCK_FOR_INSN (tail);
5824 prev_bb = NULL;
5825 for (curr_insn = tail;
5826 curr_insn != PREV_INSN (head);
5827 curr_insn = prev_insn)
5829 prev_insn = PREV_INSN (curr_insn);
5830 /* We need to process empty blocks too. They contain
5831 NOTE_INSN_BASIC_BLOCK referring for the basic block. */
5832 if (NOTE_P (curr_insn) && NOTE_KIND (curr_insn) != NOTE_INSN_BASIC_BLOCK)
5833 continue;
5834 curr_bb = BLOCK_FOR_INSN (curr_insn);
5835 if (curr_bb != prev_bb)
5837 if (prev_bb != NULL)
5839 /* Update df_get_live_in (prev_bb): */
5840 EXECUTE_IF_SET_IN_BITMAP (&check_only_regs, 0, j, bi)
5841 if (bitmap_bit_p (&live_regs, j))
5842 bitmap_set_bit (df_get_live_in (prev_bb), j);
5843 else
5844 bitmap_clear_bit (df_get_live_in (prev_bb), j);
5846 if (curr_bb != last_bb)
5848 /* Update df_get_live_out (curr_bb): */
5849 EXECUTE_IF_SET_IN_BITMAP (&check_only_regs, 0, j, bi)
5851 live_p = bitmap_bit_p (&live_regs, j);
5852 if (! live_p)
5853 FOR_EACH_EDGE (e, ei, curr_bb->succs)
5854 if (bitmap_bit_p (df_get_live_in (e->dest), j))
5856 live_p = true;
5857 break;
5859 if (live_p)
5860 bitmap_set_bit (df_get_live_out (curr_bb), j);
5861 else
5862 bitmap_clear_bit (df_get_live_out (curr_bb), j);
5865 prev_bb = curr_bb;
5866 bitmap_and (&live_regs, &check_only_regs, df_get_live_out (curr_bb));
5868 if (! NONDEBUG_INSN_P (curr_insn))
5869 continue;
5870 curr_id = lra_get_insn_recog_data (curr_insn);
5871 curr_static_id = curr_id->insn_static_data;
5872 remove_p = false;
5873 if ((set = single_set (curr_insn)) != NULL_RTX
5874 && REG_P (SET_DEST (set))
5875 && (regno = REGNO (SET_DEST (set))) >= FIRST_PSEUDO_REGISTER
5876 && SET_DEST (set) != pic_offset_table_rtx
5877 && bitmap_bit_p (&check_only_regs, regno)
5878 && ! bitmap_bit_p (&live_regs, regno))
5879 remove_p = true;
5880 /* See which defined values die here. */
5881 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5882 if (reg->type == OP_OUT && ! reg->subreg_p)
5883 bitmap_clear_bit (&live_regs, reg->regno);
5884 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
5885 if (reg->type == OP_OUT && ! reg->subreg_p)
5886 bitmap_clear_bit (&live_regs, reg->regno);
5887 if (curr_id->arg_hard_regs != NULL)
5888 /* Make clobbered argument hard registers die. */
5889 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
5890 if (regno >= FIRST_PSEUDO_REGISTER)
5891 bitmap_clear_bit (&live_regs, regno - FIRST_PSEUDO_REGISTER);
5892 /* Mark each used value as live. */
5893 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5894 if (reg->type != OP_OUT
5895 && bitmap_bit_p (&check_only_regs, reg->regno))
5896 bitmap_set_bit (&live_regs, reg->regno);
5897 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
5898 if (reg->type != OP_OUT
5899 && bitmap_bit_p (&check_only_regs, reg->regno))
5900 bitmap_set_bit (&live_regs, reg->regno);
5901 if (curr_id->arg_hard_regs != NULL)
5902 /* Make used argument hard registers live. */
5903 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
5904 if (regno < FIRST_PSEUDO_REGISTER
5905 && bitmap_bit_p (&check_only_regs, regno))
5906 bitmap_set_bit (&live_regs, regno);
5907 /* It is quite important to remove dead move insns because it
5908 means removing dead store. We don't need to process them for
5909 constraints. */
5910 if (remove_p)
5912 if (lra_dump_file != NULL)
5914 fprintf (lra_dump_file, " Removing dead insn:\n ");
5915 dump_insn_slim (lra_dump_file, curr_insn);
5917 lra_set_insn_deleted (curr_insn);
5922 /* The structure describes info to do an inheritance for the current
5923 insn. We need to collect such info first before doing the
5924 transformations because the transformations change the insn
5925 internal representation. */
5926 struct to_inherit
5928 /* Original regno. */
5929 int regno;
5930 /* Subsequent insns which can inherit original reg value. */
5931 rtx insns;
5934 /* Array containing all info for doing inheritance from the current
5935 insn. */
5936 static struct to_inherit to_inherit[LRA_MAX_INSN_RELOADS];
5938 /* Number elements in the previous array. */
5939 static int to_inherit_num;
5941 /* Add inheritance info REGNO and INSNS. Their meaning is described in
5942 structure to_inherit. */
5943 static void
5944 add_to_inherit (int regno, rtx insns)
5946 int i;
5948 for (i = 0; i < to_inherit_num; i++)
5949 if (to_inherit[i].regno == regno)
5950 return;
5951 lra_assert (to_inherit_num < LRA_MAX_INSN_RELOADS);
5952 to_inherit[to_inherit_num].regno = regno;
5953 to_inherit[to_inherit_num++].insns = insns;
5956 /* Return the last non-debug insn in basic block BB, or the block begin
5957 note if none. */
5958 static rtx_insn *
5959 get_last_insertion_point (basic_block bb)
5961 rtx_insn *insn;
5963 FOR_BB_INSNS_REVERSE (bb, insn)
5964 if (NONDEBUG_INSN_P (insn) || NOTE_INSN_BASIC_BLOCK_P (insn))
5965 return insn;
5966 gcc_unreachable ();
5969 /* Set up RES by registers living on edges FROM except the edge (FROM,
5970 TO) or by registers set up in a jump insn in BB FROM. */
5971 static void
5972 get_live_on_other_edges (basic_block from, basic_block to, bitmap res)
5974 rtx_insn *last;
5975 struct lra_insn_reg *reg;
5976 edge e;
5977 edge_iterator ei;
5979 lra_assert (to != NULL);
5980 bitmap_clear (res);
5981 FOR_EACH_EDGE (e, ei, from->succs)
5982 if (e->dest != to)
5983 bitmap_ior_into (res, df_get_live_in (e->dest));
5984 last = get_last_insertion_point (from);
5985 if (! JUMP_P (last))
5986 return;
5987 curr_id = lra_get_insn_recog_data (last);
5988 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5989 if (reg->type != OP_IN)
5990 bitmap_set_bit (res, reg->regno);
5993 /* Used as a temporary results of some bitmap calculations. */
5994 static bitmap_head temp_bitmap;
5996 /* We split for reloads of small class of hard regs. The following
5997 defines how many hard regs the class should have to be qualified as
5998 small. The code is mostly oriented to x86/x86-64 architecture
5999 where some insns need to use only specific register or pair of
6000 registers and these register can live in RTL explicitly, e.g. for
6001 parameter passing. */
6002 static const int max_small_class_regs_num = 2;
6004 /* Do inheritance/split transformations in EBB starting with HEAD and
6005 finishing on TAIL. We process EBB insns in the reverse order.
6006 Return true if we did any inheritance/split transformation in the
6007 EBB.
6009 We should avoid excessive splitting which results in worse code
6010 because of inaccurate cost calculations for spilling new split
6011 pseudos in such case. To achieve this we do splitting only if
6012 register pressure is high in given basic block and there are reload
6013 pseudos requiring hard registers. We could do more register
6014 pressure calculations at any given program point to avoid necessary
6015 splitting even more but it is to expensive and the current approach
6016 works well enough. */
6017 static bool
6018 inherit_in_ebb (rtx_insn *head, rtx_insn *tail)
6020 int i, src_regno, dst_regno, nregs;
6021 bool change_p, succ_p, update_reloads_num_p;
6022 rtx_insn *prev_insn, *last_insn;
6023 rtx next_usage_insns, curr_set;
6024 enum reg_class cl;
6025 struct lra_insn_reg *reg;
6026 basic_block last_processed_bb, curr_bb = NULL;
6027 HARD_REG_SET potential_reload_hard_regs, live_hard_regs;
6028 bitmap to_process;
6029 unsigned int j;
6030 bitmap_iterator bi;
6031 bool head_p, after_p;
6033 change_p = false;
6034 curr_usage_insns_check++;
6035 clear_invariants ();
6036 reloads_num = calls_num = 0;
6037 bitmap_clear (&check_only_regs);
6038 bitmap_clear (&invalid_invariant_regs);
6039 last_processed_bb = NULL;
6040 CLEAR_HARD_REG_SET (potential_reload_hard_regs);
6041 COPY_HARD_REG_SET (live_hard_regs, eliminable_regset);
6042 IOR_HARD_REG_SET (live_hard_regs, lra_no_alloc_regs);
6043 /* We don't process new insns generated in the loop. */
6044 for (curr_insn = tail; curr_insn != PREV_INSN (head); curr_insn = prev_insn)
6046 prev_insn = PREV_INSN (curr_insn);
6047 if (BLOCK_FOR_INSN (curr_insn) != NULL)
6048 curr_bb = BLOCK_FOR_INSN (curr_insn);
6049 if (last_processed_bb != curr_bb)
6051 /* We are at the end of BB. Add qualified living
6052 pseudos for potential splitting. */
6053 to_process = df_get_live_out (curr_bb);
6054 if (last_processed_bb != NULL)
6056 /* We are somewhere in the middle of EBB. */
6057 get_live_on_other_edges (curr_bb, last_processed_bb,
6058 &temp_bitmap);
6059 to_process = &temp_bitmap;
6061 last_processed_bb = curr_bb;
6062 last_insn = get_last_insertion_point (curr_bb);
6063 after_p = (! JUMP_P (last_insn)
6064 && (! CALL_P (last_insn)
6065 || (find_reg_note (last_insn,
6066 REG_NORETURN, NULL_RTX) == NULL_RTX
6067 && ! SIBLING_CALL_P (last_insn))));
6068 CLEAR_HARD_REG_SET (potential_reload_hard_regs);
6069 EXECUTE_IF_SET_IN_BITMAP (to_process, 0, j, bi)
6071 if ((int) j >= lra_constraint_new_regno_start)
6072 break;
6073 if (j < FIRST_PSEUDO_REGISTER || reg_renumber[j] >= 0)
6075 if (j < FIRST_PSEUDO_REGISTER)
6076 SET_HARD_REG_BIT (live_hard_regs, j);
6077 else
6078 add_to_hard_reg_set (&live_hard_regs,
6079 PSEUDO_REGNO_MODE (j),
6080 reg_renumber[j]);
6081 setup_next_usage_insn (j, last_insn, reloads_num, after_p);
6085 src_regno = dst_regno = -1;
6086 curr_set = single_set (curr_insn);
6087 if (curr_set != NULL_RTX && REG_P (SET_DEST (curr_set)))
6088 dst_regno = REGNO (SET_DEST (curr_set));
6089 if (curr_set != NULL_RTX && REG_P (SET_SRC (curr_set)))
6090 src_regno = REGNO (SET_SRC (curr_set));
6091 update_reloads_num_p = true;
6092 if (src_regno < lra_constraint_new_regno_start
6093 && src_regno >= FIRST_PSEUDO_REGISTER
6094 && reg_renumber[src_regno] < 0
6095 && dst_regno >= lra_constraint_new_regno_start
6096 && (cl = lra_get_allocno_class (dst_regno)) != NO_REGS)
6098 /* 'reload_pseudo <- original_pseudo'. */
6099 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6100 reloads_num++;
6101 update_reloads_num_p = false;
6102 succ_p = false;
6103 if (usage_insns[src_regno].check == curr_usage_insns_check
6104 && (next_usage_insns = usage_insns[src_regno].insns) != NULL_RTX)
6105 succ_p = inherit_reload_reg (false, src_regno, cl,
6106 curr_insn, next_usage_insns);
6107 if (succ_p)
6108 change_p = true;
6109 else
6110 setup_next_usage_insn (src_regno, curr_insn, reloads_num, false);
6111 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6112 IOR_HARD_REG_SET (potential_reload_hard_regs,
6113 reg_class_contents[cl]);
6115 else if (src_regno < 0
6116 && dst_regno >= lra_constraint_new_regno_start
6117 && invariant_p (SET_SRC (curr_set))
6118 && (cl = lra_get_allocno_class (dst_regno)) != NO_REGS
6119 && ! bitmap_bit_p (&invalid_invariant_regs, dst_regno)
6120 && ! bitmap_bit_p (&invalid_invariant_regs,
6121 ORIGINAL_REGNO(regno_reg_rtx[dst_regno])))
6123 /* 'reload_pseudo <- invariant'. */
6124 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6125 reloads_num++;
6126 update_reloads_num_p = false;
6127 if (process_invariant_for_inheritance (SET_DEST (curr_set), SET_SRC (curr_set)))
6128 change_p = true;
6129 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6130 IOR_HARD_REG_SET (potential_reload_hard_regs,
6131 reg_class_contents[cl]);
6133 else if (src_regno >= lra_constraint_new_regno_start
6134 && dst_regno < lra_constraint_new_regno_start
6135 && dst_regno >= FIRST_PSEUDO_REGISTER
6136 && reg_renumber[dst_regno] < 0
6137 && (cl = lra_get_allocno_class (src_regno)) != NO_REGS
6138 && usage_insns[dst_regno].check == curr_usage_insns_check
6139 && (next_usage_insns
6140 = usage_insns[dst_regno].insns) != NULL_RTX)
6142 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6143 reloads_num++;
6144 update_reloads_num_p = false;
6145 /* 'original_pseudo <- reload_pseudo'. */
6146 if (! JUMP_P (curr_insn)
6147 && inherit_reload_reg (true, dst_regno, cl,
6148 curr_insn, next_usage_insns))
6149 change_p = true;
6150 /* Invalidate. */
6151 usage_insns[dst_regno].check = 0;
6152 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6153 IOR_HARD_REG_SET (potential_reload_hard_regs,
6154 reg_class_contents[cl]);
6156 else if (INSN_P (curr_insn))
6158 int iter;
6159 int max_uid = get_max_uid ();
6161 curr_id = lra_get_insn_recog_data (curr_insn);
6162 curr_static_id = curr_id->insn_static_data;
6163 to_inherit_num = 0;
6164 /* Process insn definitions. */
6165 for (iter = 0; iter < 2; iter++)
6166 for (reg = iter == 0 ? curr_id->regs : curr_static_id->hard_regs;
6167 reg != NULL;
6168 reg = reg->next)
6169 if (reg->type != OP_IN
6170 && (dst_regno = reg->regno) < lra_constraint_new_regno_start)
6172 if (dst_regno >= FIRST_PSEUDO_REGISTER && reg->type == OP_OUT
6173 && reg_renumber[dst_regno] < 0 && ! reg->subreg_p
6174 && usage_insns[dst_regno].check == curr_usage_insns_check
6175 && (next_usage_insns
6176 = usage_insns[dst_regno].insns) != NULL_RTX)
6178 struct lra_insn_reg *r;
6180 for (r = curr_id->regs; r != NULL; r = r->next)
6181 if (r->type != OP_OUT && r->regno == dst_regno)
6182 break;
6183 /* Don't do inheritance if the pseudo is also
6184 used in the insn. */
6185 if (r == NULL)
6186 /* We can not do inheritance right now
6187 because the current insn reg info (chain
6188 regs) can change after that. */
6189 add_to_inherit (dst_regno, next_usage_insns);
6191 /* We can not process one reg twice here because of
6192 usage_insns invalidation. */
6193 if ((dst_regno < FIRST_PSEUDO_REGISTER
6194 || reg_renumber[dst_regno] >= 0)
6195 && ! reg->subreg_p && reg->type != OP_IN)
6197 HARD_REG_SET s;
6199 if (split_if_necessary (dst_regno, reg->biggest_mode,
6200 potential_reload_hard_regs,
6201 false, curr_insn, max_uid))
6202 change_p = true;
6203 CLEAR_HARD_REG_SET (s);
6204 if (dst_regno < FIRST_PSEUDO_REGISTER)
6205 add_to_hard_reg_set (&s, reg->biggest_mode, dst_regno);
6206 else
6207 add_to_hard_reg_set (&s, PSEUDO_REGNO_MODE (dst_regno),
6208 reg_renumber[dst_regno]);
6209 AND_COMPL_HARD_REG_SET (live_hard_regs, s);
6211 /* We should invalidate potential inheritance or
6212 splitting for the current insn usages to the next
6213 usage insns (see code below) as the output pseudo
6214 prevents this. */
6215 if ((dst_regno >= FIRST_PSEUDO_REGISTER
6216 && reg_renumber[dst_regno] < 0)
6217 || (reg->type == OP_OUT && ! reg->subreg_p
6218 && (dst_regno < FIRST_PSEUDO_REGISTER
6219 || reg_renumber[dst_regno] >= 0)))
6221 /* Invalidate and mark definitions. */
6222 if (dst_regno >= FIRST_PSEUDO_REGISTER)
6223 usage_insns[dst_regno].check = -(int) INSN_UID (curr_insn);
6224 else
6226 nregs = hard_regno_nregs (dst_regno,
6227 reg->biggest_mode);
6228 for (i = 0; i < nregs; i++)
6229 usage_insns[dst_regno + i].check
6230 = -(int) INSN_UID (curr_insn);
6234 /* Process clobbered call regs. */
6235 if (curr_id->arg_hard_regs != NULL)
6236 for (i = 0; (dst_regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6237 if (dst_regno >= FIRST_PSEUDO_REGISTER)
6238 usage_insns[dst_regno - FIRST_PSEUDO_REGISTER].check
6239 = -(int) INSN_UID (curr_insn);
6240 if (! JUMP_P (curr_insn))
6241 for (i = 0; i < to_inherit_num; i++)
6242 if (inherit_reload_reg (true, to_inherit[i].regno,
6243 ALL_REGS, curr_insn,
6244 to_inherit[i].insns))
6245 change_p = true;
6246 if (CALL_P (curr_insn))
6248 rtx cheap, pat, dest;
6249 rtx_insn *restore;
6250 int regno, hard_regno;
6252 calls_num++;
6253 if ((cheap = find_reg_note (curr_insn,
6254 REG_RETURNED, NULL_RTX)) != NULL_RTX
6255 && ((cheap = XEXP (cheap, 0)), true)
6256 && (regno = REGNO (cheap)) >= FIRST_PSEUDO_REGISTER
6257 && (hard_regno = reg_renumber[regno]) >= 0
6258 && usage_insns[regno].check == curr_usage_insns_check
6259 /* If there are pending saves/restores, the
6260 optimization is not worth. */
6261 && usage_insns[regno].calls_num == calls_num - 1
6262 && TEST_HARD_REG_BIT (call_used_reg_set, hard_regno))
6264 /* Restore the pseudo from the call result as
6265 REG_RETURNED note says that the pseudo value is
6266 in the call result and the pseudo is an argument
6267 of the call. */
6268 pat = PATTERN (curr_insn);
6269 if (GET_CODE (pat) == PARALLEL)
6270 pat = XVECEXP (pat, 0, 0);
6271 dest = SET_DEST (pat);
6272 /* For multiple return values dest is PARALLEL.
6273 Currently we handle only single return value case. */
6274 if (REG_P (dest))
6276 start_sequence ();
6277 emit_move_insn (cheap, copy_rtx (dest));
6278 restore = get_insns ();
6279 end_sequence ();
6280 lra_process_new_insns (curr_insn, NULL, restore,
6281 "Inserting call parameter restore");
6282 /* We don't need to save/restore of the pseudo from
6283 this call. */
6284 usage_insns[regno].calls_num = calls_num;
6285 bitmap_set_bit (&check_only_regs, regno);
6289 to_inherit_num = 0;
6290 /* Process insn usages. */
6291 for (iter = 0; iter < 2; iter++)
6292 for (reg = iter == 0 ? curr_id->regs : curr_static_id->hard_regs;
6293 reg != NULL;
6294 reg = reg->next)
6295 if ((reg->type != OP_OUT
6296 || (reg->type == OP_OUT && reg->subreg_p))
6297 && (src_regno = reg->regno) < lra_constraint_new_regno_start)
6299 if (src_regno >= FIRST_PSEUDO_REGISTER
6300 && reg_renumber[src_regno] < 0 && reg->type == OP_IN)
6302 if (usage_insns[src_regno].check == curr_usage_insns_check
6303 && (next_usage_insns
6304 = usage_insns[src_regno].insns) != NULL_RTX
6305 && NONDEBUG_INSN_P (curr_insn))
6306 add_to_inherit (src_regno, next_usage_insns);
6307 else if (usage_insns[src_regno].check
6308 != -(int) INSN_UID (curr_insn))
6309 /* Add usages but only if the reg is not set up
6310 in the same insn. */
6311 add_next_usage_insn (src_regno, curr_insn, reloads_num);
6313 else if (src_regno < FIRST_PSEUDO_REGISTER
6314 || reg_renumber[src_regno] >= 0)
6316 bool before_p;
6317 rtx_insn *use_insn = curr_insn;
6319 before_p = (JUMP_P (curr_insn)
6320 || (CALL_P (curr_insn) && reg->type == OP_IN));
6321 if (NONDEBUG_INSN_P (curr_insn)
6322 && (! JUMP_P (curr_insn) || reg->type == OP_IN)
6323 && split_if_necessary (src_regno, reg->biggest_mode,
6324 potential_reload_hard_regs,
6325 before_p, curr_insn, max_uid))
6327 if (reg->subreg_p)
6328 lra_risky_transformations_p = true;
6329 change_p = true;
6330 /* Invalidate. */
6331 usage_insns[src_regno].check = 0;
6332 if (before_p)
6333 use_insn = PREV_INSN (curr_insn);
6335 if (NONDEBUG_INSN_P (curr_insn))
6337 if (src_regno < FIRST_PSEUDO_REGISTER)
6338 add_to_hard_reg_set (&live_hard_regs,
6339 reg->biggest_mode, src_regno);
6340 else
6341 add_to_hard_reg_set (&live_hard_regs,
6342 PSEUDO_REGNO_MODE (src_regno),
6343 reg_renumber[src_regno]);
6345 add_next_usage_insn (src_regno, use_insn, reloads_num);
6348 /* Process used call regs. */
6349 if (curr_id->arg_hard_regs != NULL)
6350 for (i = 0; (src_regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6351 if (src_regno < FIRST_PSEUDO_REGISTER)
6353 SET_HARD_REG_BIT (live_hard_regs, src_regno);
6354 add_next_usage_insn (src_regno, curr_insn, reloads_num);
6356 for (i = 0; i < to_inherit_num; i++)
6358 src_regno = to_inherit[i].regno;
6359 if (inherit_reload_reg (false, src_regno, ALL_REGS,
6360 curr_insn, to_inherit[i].insns))
6361 change_p = true;
6362 else
6363 setup_next_usage_insn (src_regno, curr_insn, reloads_num, false);
6366 if (update_reloads_num_p
6367 && NONDEBUG_INSN_P (curr_insn) && curr_set != NULL_RTX)
6369 int regno = -1;
6370 if ((REG_P (SET_DEST (curr_set))
6371 && (regno = REGNO (SET_DEST (curr_set))) >= lra_constraint_new_regno_start
6372 && reg_renumber[regno] < 0
6373 && (cl = lra_get_allocno_class (regno)) != NO_REGS)
6374 || (REG_P (SET_SRC (curr_set))
6375 && (regno = REGNO (SET_SRC (curr_set))) >= lra_constraint_new_regno_start
6376 && reg_renumber[regno] < 0
6377 && (cl = lra_get_allocno_class (regno)) != NO_REGS))
6379 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6380 reloads_num++;
6381 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6382 IOR_HARD_REG_SET (potential_reload_hard_regs,
6383 reg_class_contents[cl]);
6386 if (NONDEBUG_INSN_P (curr_insn))
6388 int regno;
6390 /* Invalidate invariants with changed regs. */
6391 curr_id = lra_get_insn_recog_data (curr_insn);
6392 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
6393 if (reg->type != OP_IN)
6395 bitmap_set_bit (&invalid_invariant_regs, reg->regno);
6396 bitmap_set_bit (&invalid_invariant_regs,
6397 ORIGINAL_REGNO (regno_reg_rtx[reg->regno]));
6399 curr_static_id = curr_id->insn_static_data;
6400 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
6401 if (reg->type != OP_IN)
6402 bitmap_set_bit (&invalid_invariant_regs, reg->regno);
6403 if (curr_id->arg_hard_regs != NULL)
6404 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6405 if (regno >= FIRST_PSEUDO_REGISTER)
6406 bitmap_set_bit (&invalid_invariant_regs,
6407 regno - FIRST_PSEUDO_REGISTER);
6409 /* We reached the start of the current basic block. */
6410 if (prev_insn == NULL_RTX || prev_insn == PREV_INSN (head)
6411 || BLOCK_FOR_INSN (prev_insn) != curr_bb)
6413 /* We reached the beginning of the current block -- do
6414 rest of spliting in the current BB. */
6415 to_process = df_get_live_in (curr_bb);
6416 if (BLOCK_FOR_INSN (head) != curr_bb)
6418 /* We are somewhere in the middle of EBB. */
6419 get_live_on_other_edges (EDGE_PRED (curr_bb, 0)->src,
6420 curr_bb, &temp_bitmap);
6421 to_process = &temp_bitmap;
6423 head_p = true;
6424 EXECUTE_IF_SET_IN_BITMAP (to_process, 0, j, bi)
6426 if ((int) j >= lra_constraint_new_regno_start)
6427 break;
6428 if (((int) j < FIRST_PSEUDO_REGISTER || reg_renumber[j] >= 0)
6429 && usage_insns[j].check == curr_usage_insns_check
6430 && (next_usage_insns = usage_insns[j].insns) != NULL_RTX)
6432 if (need_for_split_p (potential_reload_hard_regs, j))
6434 if (lra_dump_file != NULL && head_p)
6436 fprintf (lra_dump_file,
6437 " ----------------------------------\n");
6438 head_p = false;
6440 if (split_reg (false, j, bb_note (curr_bb),
6441 next_usage_insns))
6442 change_p = true;
6444 usage_insns[j].check = 0;
6449 return change_p;
6452 /* This value affects EBB forming. If probability of edge from EBB to
6453 a BB is not greater than the following value, we don't add the BB
6454 to EBB. */
6455 #define EBB_PROBABILITY_CUTOFF \
6456 ((REG_BR_PROB_BASE * LRA_INHERITANCE_EBB_PROBABILITY_CUTOFF) / 100)
6458 /* Current number of inheritance/split iteration. */
6459 int lra_inheritance_iter;
6461 /* Entry function for inheritance/split pass. */
6462 void
6463 lra_inheritance (void)
6465 int i;
6466 basic_block bb, start_bb;
6467 edge e;
6469 lra_inheritance_iter++;
6470 if (lra_inheritance_iter > LRA_MAX_INHERITANCE_PASSES)
6471 return;
6472 timevar_push (TV_LRA_INHERITANCE);
6473 if (lra_dump_file != NULL)
6474 fprintf (lra_dump_file, "\n********** Inheritance #%d: **********\n\n",
6475 lra_inheritance_iter);
6476 curr_usage_insns_check = 0;
6477 usage_insns = XNEWVEC (struct usage_insns, lra_constraint_new_regno_start);
6478 for (i = 0; i < lra_constraint_new_regno_start; i++)
6479 usage_insns[i].check = 0;
6480 bitmap_initialize (&check_only_regs, &reg_obstack);
6481 bitmap_initialize (&invalid_invariant_regs, &reg_obstack);
6482 bitmap_initialize (&live_regs, &reg_obstack);
6483 bitmap_initialize (&temp_bitmap, &reg_obstack);
6484 bitmap_initialize (&ebb_global_regs, &reg_obstack);
6485 FOR_EACH_BB_FN (bb, cfun)
6487 start_bb = bb;
6488 if (lra_dump_file != NULL)
6489 fprintf (lra_dump_file, "EBB");
6490 /* Form a EBB starting with BB. */
6491 bitmap_clear (&ebb_global_regs);
6492 bitmap_ior_into (&ebb_global_regs, df_get_live_in (bb));
6493 for (;;)
6495 if (lra_dump_file != NULL)
6496 fprintf (lra_dump_file, " %d", bb->index);
6497 if (bb->next_bb == EXIT_BLOCK_PTR_FOR_FN (cfun)
6498 || LABEL_P (BB_HEAD (bb->next_bb)))
6499 break;
6500 e = find_fallthru_edge (bb->succs);
6501 if (! e)
6502 break;
6503 if (e->probability.initialized_p ()
6504 && e->probability.to_reg_br_prob_base () < EBB_PROBABILITY_CUTOFF)
6505 break;
6506 bb = bb->next_bb;
6508 bitmap_ior_into (&ebb_global_regs, df_get_live_out (bb));
6509 if (lra_dump_file != NULL)
6510 fprintf (lra_dump_file, "\n");
6511 if (inherit_in_ebb (BB_HEAD (start_bb), BB_END (bb)))
6512 /* Remember that the EBB head and tail can change in
6513 inherit_in_ebb. */
6514 update_ebb_live_info (BB_HEAD (start_bb), BB_END (bb));
6516 bitmap_clear (&ebb_global_regs);
6517 bitmap_clear (&temp_bitmap);
6518 bitmap_clear (&live_regs);
6519 bitmap_clear (&invalid_invariant_regs);
6520 bitmap_clear (&check_only_regs);
6521 free (usage_insns);
6523 timevar_pop (TV_LRA_INHERITANCE);
6528 /* This page contains code to undo failed inheritance/split
6529 transformations. */
6531 /* Current number of iteration undoing inheritance/split. */
6532 int lra_undo_inheritance_iter;
6534 /* Fix BB live info LIVE after removing pseudos created on pass doing
6535 inheritance/split which are REMOVED_PSEUDOS. */
6536 static void
6537 fix_bb_live_info (bitmap live, bitmap removed_pseudos)
6539 unsigned int regno;
6540 bitmap_iterator bi;
6542 EXECUTE_IF_SET_IN_BITMAP (removed_pseudos, 0, regno, bi)
6543 if (bitmap_clear_bit (live, regno)
6544 && REG_P (lra_reg_info[regno].restore_rtx))
6545 bitmap_set_bit (live, REGNO (lra_reg_info[regno].restore_rtx));
6548 /* Return regno of the (subreg of) REG. Otherwise, return a negative
6549 number. */
6550 static int
6551 get_regno (rtx reg)
6553 if (GET_CODE (reg) == SUBREG)
6554 reg = SUBREG_REG (reg);
6555 if (REG_P (reg))
6556 return REGNO (reg);
6557 return -1;
6560 /* Delete a move INSN with destination reg DREGNO and a previous
6561 clobber insn with the same regno. The inheritance/split code can
6562 generate moves with preceding clobber and when we delete such moves
6563 we should delete the clobber insn too to keep the correct life
6564 info. */
6565 static void
6566 delete_move_and_clobber (rtx_insn *insn, int dregno)
6568 rtx_insn *prev_insn = PREV_INSN (insn);
6570 lra_set_insn_deleted (insn);
6571 lra_assert (dregno >= 0);
6572 if (prev_insn != NULL && NONDEBUG_INSN_P (prev_insn)
6573 && GET_CODE (PATTERN (prev_insn)) == CLOBBER
6574 && dregno == get_regno (XEXP (PATTERN (prev_insn), 0)))
6575 lra_set_insn_deleted (prev_insn);
6578 /* Remove inheritance/split pseudos which are in REMOVE_PSEUDOS and
6579 return true if we did any change. The undo transformations for
6580 inheritance looks like
6581 i <- i2
6582 p <- i => p <- i2
6583 or removing
6584 p <- i, i <- p, and i <- i3
6585 where p is original pseudo from which inheritance pseudo i was
6586 created, i and i3 are removed inheritance pseudos, i2 is another
6587 not removed inheritance pseudo. All split pseudos or other
6588 occurrences of removed inheritance pseudos are changed on the
6589 corresponding original pseudos.
6591 The function also schedules insns changed and created during
6592 inheritance/split pass for processing by the subsequent constraint
6593 pass. */
6594 static bool
6595 remove_inheritance_pseudos (bitmap remove_pseudos)
6597 basic_block bb;
6598 int regno, sregno, prev_sregno, dregno;
6599 rtx restore_rtx;
6600 rtx set, prev_set;
6601 rtx_insn *prev_insn;
6602 bool change_p, done_p;
6604 change_p = ! bitmap_empty_p (remove_pseudos);
6605 /* We can not finish the function right away if CHANGE_P is true
6606 because we need to marks insns affected by previous
6607 inheritance/split pass for processing by the subsequent
6608 constraint pass. */
6609 FOR_EACH_BB_FN (bb, cfun)
6611 fix_bb_live_info (df_get_live_in (bb), remove_pseudos);
6612 fix_bb_live_info (df_get_live_out (bb), remove_pseudos);
6613 FOR_BB_INSNS_REVERSE (bb, curr_insn)
6615 if (! INSN_P (curr_insn))
6616 continue;
6617 done_p = false;
6618 sregno = dregno = -1;
6619 if (change_p && NONDEBUG_INSN_P (curr_insn)
6620 && (set = single_set (curr_insn)) != NULL_RTX)
6622 dregno = get_regno (SET_DEST (set));
6623 sregno = get_regno (SET_SRC (set));
6626 if (sregno >= 0 && dregno >= 0)
6628 if (bitmap_bit_p (remove_pseudos, dregno)
6629 && ! REG_P (lra_reg_info[dregno].restore_rtx))
6631 /* invariant inheritance pseudo <- original pseudo */
6632 if (lra_dump_file != NULL)
6634 fprintf (lra_dump_file, " Removing invariant inheritance:\n");
6635 dump_insn_slim (lra_dump_file, curr_insn);
6636 fprintf (lra_dump_file, "\n");
6638 delete_move_and_clobber (curr_insn, dregno);
6639 done_p = true;
6641 else if (bitmap_bit_p (remove_pseudos, sregno)
6642 && ! REG_P (lra_reg_info[sregno].restore_rtx))
6644 /* reload pseudo <- invariant inheritance pseudo */
6645 start_sequence ();
6646 /* We can not just change the source. It might be
6647 an insn different from the move. */
6648 emit_insn (lra_reg_info[sregno].restore_rtx);
6649 rtx_insn *new_insns = get_insns ();
6650 end_sequence ();
6651 lra_assert (single_set (new_insns) != NULL
6652 && SET_DEST (set) == SET_DEST (single_set (new_insns)));
6653 lra_process_new_insns (curr_insn, NULL, new_insns,
6654 "Changing reload<-invariant inheritance");
6655 delete_move_and_clobber (curr_insn, dregno);
6656 done_p = true;
6658 else if ((bitmap_bit_p (remove_pseudos, sregno)
6659 && (get_regno (lra_reg_info[sregno].restore_rtx) == dregno
6660 || (bitmap_bit_p (remove_pseudos, dregno)
6661 && get_regno (lra_reg_info[sregno].restore_rtx) >= 0
6662 && (get_regno (lra_reg_info[sregno].restore_rtx)
6663 == get_regno (lra_reg_info[dregno].restore_rtx)))))
6664 || (bitmap_bit_p (remove_pseudos, dregno)
6665 && get_regno (lra_reg_info[dregno].restore_rtx) == sregno))
6666 /* One of the following cases:
6667 original <- removed inheritance pseudo
6668 removed inherit pseudo <- another removed inherit pseudo
6669 removed inherit pseudo <- original pseudo
6671 removed_split_pseudo <- original_reg
6672 original_reg <- removed_split_pseudo */
6674 if (lra_dump_file != NULL)
6676 fprintf (lra_dump_file, " Removing %s:\n",
6677 bitmap_bit_p (&lra_split_regs, sregno)
6678 || bitmap_bit_p (&lra_split_regs, dregno)
6679 ? "split" : "inheritance");
6680 dump_insn_slim (lra_dump_file, curr_insn);
6682 delete_move_and_clobber (curr_insn, dregno);
6683 done_p = true;
6685 else if (bitmap_bit_p (remove_pseudos, sregno)
6686 && bitmap_bit_p (&lra_inheritance_pseudos, sregno))
6688 /* Search the following pattern:
6689 inherit_or_split_pseudo1 <- inherit_or_split_pseudo2
6690 original_pseudo <- inherit_or_split_pseudo1
6691 where the 2nd insn is the current insn and
6692 inherit_or_split_pseudo2 is not removed. If it is found,
6693 change the current insn onto:
6694 original_pseudo <- inherit_or_split_pseudo2. */
6695 for (prev_insn = PREV_INSN (curr_insn);
6696 prev_insn != NULL_RTX && ! NONDEBUG_INSN_P (prev_insn);
6697 prev_insn = PREV_INSN (prev_insn))
6699 if (prev_insn != NULL_RTX && BLOCK_FOR_INSN (prev_insn) == bb
6700 && (prev_set = single_set (prev_insn)) != NULL_RTX
6701 /* There should be no subregs in insn we are
6702 searching because only the original reg might
6703 be in subreg when we changed the mode of
6704 load/store for splitting. */
6705 && REG_P (SET_DEST (prev_set))
6706 && REG_P (SET_SRC (prev_set))
6707 && (int) REGNO (SET_DEST (prev_set)) == sregno
6708 && ((prev_sregno = REGNO (SET_SRC (prev_set)))
6709 >= FIRST_PSEUDO_REGISTER)
6710 && (lra_reg_info[prev_sregno].restore_rtx == NULL_RTX
6712 /* As we consider chain of inheritance or
6713 splitting described in above comment we should
6714 check that sregno and prev_sregno were
6715 inheritance/split pseudos created from the
6716 same original regno. */
6717 (get_regno (lra_reg_info[sregno].restore_rtx) >= 0
6718 && (get_regno (lra_reg_info[sregno].restore_rtx)
6719 == get_regno (lra_reg_info[prev_sregno].restore_rtx))))
6720 && ! bitmap_bit_p (remove_pseudos, prev_sregno))
6722 lra_assert (GET_MODE (SET_SRC (prev_set))
6723 == GET_MODE (regno_reg_rtx[sregno]));
6724 if (GET_CODE (SET_SRC (set)) == SUBREG)
6725 SUBREG_REG (SET_SRC (set)) = SET_SRC (prev_set);
6726 else
6727 SET_SRC (set) = SET_SRC (prev_set);
6728 /* As we are finishing with processing the insn
6729 here, check the destination too as it might
6730 inheritance pseudo for another pseudo. */
6731 if (bitmap_bit_p (remove_pseudos, dregno)
6732 && bitmap_bit_p (&lra_inheritance_pseudos, dregno)
6733 && (restore_rtx
6734 = lra_reg_info[dregno].restore_rtx) != NULL_RTX)
6736 if (GET_CODE (SET_DEST (set)) == SUBREG)
6737 SUBREG_REG (SET_DEST (set)) = restore_rtx;
6738 else
6739 SET_DEST (set) = restore_rtx;
6741 lra_push_insn_and_update_insn_regno_info (curr_insn);
6742 lra_set_used_insn_alternative_by_uid
6743 (INSN_UID (curr_insn), -1);
6744 done_p = true;
6745 if (lra_dump_file != NULL)
6747 fprintf (lra_dump_file, " Change reload insn:\n");
6748 dump_insn_slim (lra_dump_file, curr_insn);
6753 if (! done_p)
6755 struct lra_insn_reg *reg;
6756 bool restored_regs_p = false;
6757 bool kept_regs_p = false;
6759 curr_id = lra_get_insn_recog_data (curr_insn);
6760 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
6762 regno = reg->regno;
6763 restore_rtx = lra_reg_info[regno].restore_rtx;
6764 if (restore_rtx != NULL_RTX)
6766 if (change_p && bitmap_bit_p (remove_pseudos, regno))
6768 lra_substitute_pseudo_within_insn
6769 (curr_insn, regno, restore_rtx, false);
6770 restored_regs_p = true;
6772 else
6773 kept_regs_p = true;
6776 if (NONDEBUG_INSN_P (curr_insn) && kept_regs_p)
6778 /* The instruction has changed since the previous
6779 constraints pass. */
6780 lra_push_insn_and_update_insn_regno_info (curr_insn);
6781 lra_set_used_insn_alternative_by_uid
6782 (INSN_UID (curr_insn), -1);
6784 else if (restored_regs_p)
6785 /* The instruction has been restored to the form that
6786 it had during the previous constraints pass. */
6787 lra_update_insn_regno_info (curr_insn);
6788 if (restored_regs_p && lra_dump_file != NULL)
6790 fprintf (lra_dump_file, " Insn after restoring regs:\n");
6791 dump_insn_slim (lra_dump_file, curr_insn);
6796 return change_p;
6799 /* If optional reload pseudos failed to get a hard register or was not
6800 inherited, it is better to remove optional reloads. We do this
6801 transformation after undoing inheritance to figure out necessity to
6802 remove optional reloads easier. Return true if we do any
6803 change. */
6804 static bool
6805 undo_optional_reloads (void)
6807 bool change_p, keep_p;
6808 unsigned int regno, uid;
6809 bitmap_iterator bi, bi2;
6810 rtx_insn *insn;
6811 rtx set, src, dest;
6812 auto_bitmap removed_optional_reload_pseudos (&reg_obstack);
6814 bitmap_copy (removed_optional_reload_pseudos, &lra_optional_reload_pseudos);
6815 EXECUTE_IF_SET_IN_BITMAP (&lra_optional_reload_pseudos, 0, regno, bi)
6817 keep_p = false;
6818 /* Keep optional reloads from previous subpasses. */
6819 if (lra_reg_info[regno].restore_rtx == NULL_RTX
6820 /* If the original pseudo changed its allocation, just
6821 removing the optional pseudo is dangerous as the original
6822 pseudo will have longer live range. */
6823 || reg_renumber[REGNO (lra_reg_info[regno].restore_rtx)] >= 0)
6824 keep_p = true;
6825 else if (reg_renumber[regno] >= 0)
6826 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi2)
6828 insn = lra_insn_recog_data[uid]->insn;
6829 if ((set = single_set (insn)) == NULL_RTX)
6830 continue;
6831 src = SET_SRC (set);
6832 dest = SET_DEST (set);
6833 if (! REG_P (src) || ! REG_P (dest))
6834 continue;
6835 if (REGNO (dest) == regno
6836 /* Ignore insn for optional reloads itself. */
6837 && REGNO (lra_reg_info[regno].restore_rtx) != REGNO (src)
6838 /* Check only inheritance on last inheritance pass. */
6839 && (int) REGNO (src) >= new_regno_start
6840 /* Check that the optional reload was inherited. */
6841 && bitmap_bit_p (&lra_inheritance_pseudos, REGNO (src)))
6843 keep_p = true;
6844 break;
6847 if (keep_p)
6849 bitmap_clear_bit (removed_optional_reload_pseudos, regno);
6850 if (lra_dump_file != NULL)
6851 fprintf (lra_dump_file, "Keep optional reload reg %d\n", regno);
6854 change_p = ! bitmap_empty_p (removed_optional_reload_pseudos);
6855 auto_bitmap insn_bitmap (&reg_obstack);
6856 EXECUTE_IF_SET_IN_BITMAP (removed_optional_reload_pseudos, 0, regno, bi)
6858 if (lra_dump_file != NULL)
6859 fprintf (lra_dump_file, "Remove optional reload reg %d\n", regno);
6860 bitmap_copy (insn_bitmap, &lra_reg_info[regno].insn_bitmap);
6861 EXECUTE_IF_SET_IN_BITMAP (insn_bitmap, 0, uid, bi2)
6863 insn = lra_insn_recog_data[uid]->insn;
6864 if ((set = single_set (insn)) != NULL_RTX)
6866 src = SET_SRC (set);
6867 dest = SET_DEST (set);
6868 if (REG_P (src) && REG_P (dest)
6869 && ((REGNO (src) == regno
6870 && (REGNO (lra_reg_info[regno].restore_rtx)
6871 == REGNO (dest)))
6872 || (REGNO (dest) == regno
6873 && (REGNO (lra_reg_info[regno].restore_rtx)
6874 == REGNO (src)))))
6876 if (lra_dump_file != NULL)
6878 fprintf (lra_dump_file, " Deleting move %u\n",
6879 INSN_UID (insn));
6880 dump_insn_slim (lra_dump_file, insn);
6882 delete_move_and_clobber (insn, REGNO (dest));
6883 continue;
6885 /* We should not worry about generation memory-memory
6886 moves here as if the corresponding inheritance did
6887 not work (inheritance pseudo did not get a hard reg),
6888 we remove the inheritance pseudo and the optional
6889 reload. */
6891 lra_substitute_pseudo_within_insn
6892 (insn, regno, lra_reg_info[regno].restore_rtx, false);
6893 lra_update_insn_regno_info (insn);
6894 if (lra_dump_file != NULL)
6896 fprintf (lra_dump_file,
6897 " Restoring original insn:\n");
6898 dump_insn_slim (lra_dump_file, insn);
6902 /* Clear restore_regnos. */
6903 EXECUTE_IF_SET_IN_BITMAP (&lra_optional_reload_pseudos, 0, regno, bi)
6904 lra_reg_info[regno].restore_rtx = NULL_RTX;
6905 return change_p;
6908 /* Entry function for undoing inheritance/split transformation. Return true
6909 if we did any RTL change in this pass. */
6910 bool
6911 lra_undo_inheritance (void)
6913 unsigned int regno;
6914 int hard_regno;
6915 int n_all_inherit, n_inherit, n_all_split, n_split;
6916 rtx restore_rtx;
6917 bitmap_iterator bi;
6918 bool change_p;
6920 lra_undo_inheritance_iter++;
6921 if (lra_undo_inheritance_iter > LRA_MAX_INHERITANCE_PASSES)
6922 return false;
6923 if (lra_dump_file != NULL)
6924 fprintf (lra_dump_file,
6925 "\n********** Undoing inheritance #%d: **********\n\n",
6926 lra_undo_inheritance_iter);
6927 auto_bitmap remove_pseudos (&reg_obstack);
6928 n_inherit = n_all_inherit = 0;
6929 EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, regno, bi)
6930 if (lra_reg_info[regno].restore_rtx != NULL_RTX)
6932 n_all_inherit++;
6933 if (reg_renumber[regno] < 0
6934 /* If the original pseudo changed its allocation, just
6935 removing inheritance is dangerous as for changing
6936 allocation we used shorter live-ranges. */
6937 && (! REG_P (lra_reg_info[regno].restore_rtx)
6938 || reg_renumber[REGNO (lra_reg_info[regno].restore_rtx)] < 0))
6939 bitmap_set_bit (remove_pseudos, regno);
6940 else
6941 n_inherit++;
6943 if (lra_dump_file != NULL && n_all_inherit != 0)
6944 fprintf (lra_dump_file, "Inherit %d out of %d (%.2f%%)\n",
6945 n_inherit, n_all_inherit,
6946 (double) n_inherit / n_all_inherit * 100);
6947 n_split = n_all_split = 0;
6948 EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, regno, bi)
6949 if ((restore_rtx = lra_reg_info[regno].restore_rtx) != NULL_RTX)
6951 int restore_regno = REGNO (restore_rtx);
6953 n_all_split++;
6954 hard_regno = (restore_regno >= FIRST_PSEUDO_REGISTER
6955 ? reg_renumber[restore_regno] : restore_regno);
6956 if (hard_regno < 0 || reg_renumber[regno] == hard_regno)
6957 bitmap_set_bit (remove_pseudos, regno);
6958 else
6960 n_split++;
6961 if (lra_dump_file != NULL)
6962 fprintf (lra_dump_file, " Keep split r%d (orig=r%d)\n",
6963 regno, restore_regno);
6966 if (lra_dump_file != NULL && n_all_split != 0)
6967 fprintf (lra_dump_file, "Split %d out of %d (%.2f%%)\n",
6968 n_split, n_all_split,
6969 (double) n_split / n_all_split * 100);
6970 change_p = remove_inheritance_pseudos (remove_pseudos);
6971 /* Clear restore_regnos. */
6972 EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, regno, bi)
6973 lra_reg_info[regno].restore_rtx = NULL_RTX;
6974 EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, regno, bi)
6975 lra_reg_info[regno].restore_rtx = NULL_RTX;
6976 change_p = undo_optional_reloads () || change_p;
6977 return change_p;