* function.c (dump_stack_clash_frame_info): New function.
[official-gcc.git] / gcc / lra-constraints.c
blobd90bde2817aa0b4b29e6af27a462a79424a77cb9
1 /* Code for RTL transformations to satisfy insn constraints.
2 Copyright (C) 2010-2017 Free Software Foundation, Inc.
3 Contributed by Vladimir Makarov <vmakarov@redhat.com>.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
10 version.
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
22 /* This file contains code for 3 passes: constraint pass,
23 inheritance/split pass, and pass for undoing failed inheritance and
24 split.
26 The major goal of constraint pass is to transform RTL to satisfy
27 insn and address constraints by:
28 o choosing insn alternatives;
29 o generating *reload insns* (or reloads in brief) and *reload
30 pseudos* which will get necessary hard registers later;
31 o substituting pseudos with equivalent values and removing the
32 instructions that initialized those pseudos.
34 The constraint pass has biggest and most complicated code in LRA.
35 There are a lot of important details like:
36 o reuse of input reload pseudos to simplify reload pseudo
37 allocations;
38 o some heuristics to choose insn alternative to improve the
39 inheritance;
40 o early clobbers etc.
42 The pass is mimicking former reload pass in alternative choosing
43 because the reload pass is oriented to current machine description
44 model. It might be changed if the machine description model is
45 changed.
47 There is special code for preventing all LRA and this pass cycling
48 in case of bugs.
50 On the first iteration of the pass we process every instruction and
51 choose an alternative for each one. On subsequent iterations we try
52 to avoid reprocessing instructions if we can be sure that the old
53 choice is still valid.
55 The inheritance/spilt pass is to transform code to achieve
56 ineheritance and live range splitting. It is done on backward
57 traversal of EBBs.
59 The inheritance optimization goal is to reuse values in hard
60 registers. There is analogous optimization in old reload pass. The
61 inheritance is achieved by following transformation:
63 reload_p1 <- p reload_p1 <- p
64 ... new_p <- reload_p1
65 ... => ...
66 reload_p2 <- p reload_p2 <- new_p
68 where p is spilled and not changed between the insns. Reload_p1 is
69 also called *original pseudo* and new_p is called *inheritance
70 pseudo*.
72 The subsequent assignment pass will try to assign the same (or
73 another if it is not possible) hard register to new_p as to
74 reload_p1 or reload_p2.
76 If the assignment pass fails to assign a hard register to new_p,
77 this file will undo the inheritance and restore the original code.
78 This is because implementing the above sequence with a spilled
79 new_p would make the code much worse. The inheritance is done in
80 EBB scope. The above is just a simplified example to get an idea
81 of the inheritance as the inheritance is also done for non-reload
82 insns.
84 Splitting (transformation) is also done in EBB scope on the same
85 pass as the inheritance:
87 r <- ... or ... <- r r <- ... or ... <- r
88 ... s <- r (new insn -- save)
89 ... =>
90 ... r <- s (new insn -- restore)
91 ... <- r ... <- r
93 The *split pseudo* s is assigned to the hard register of the
94 original pseudo or hard register r.
96 Splitting is done:
97 o In EBBs with high register pressure for global pseudos (living
98 in at least 2 BBs) and assigned to hard registers when there
99 are more one reloads needing the hard registers;
100 o for pseudos needing save/restore code around calls.
102 If the split pseudo still has the same hard register as the
103 original pseudo after the subsequent assignment pass or the
104 original pseudo was split, the opposite transformation is done on
105 the same pass for undoing inheritance. */
107 #undef REG_OK_STRICT
109 #include "config.h"
110 #include "system.h"
111 #include "coretypes.h"
112 #include "backend.h"
113 #include "target.h"
114 #include "rtl.h"
115 #include "tree.h"
116 #include "predict.h"
117 #include "df.h"
118 #include "memmodel.h"
119 #include "tm_p.h"
120 #include "expmed.h"
121 #include "optabs.h"
122 #include "regs.h"
123 #include "ira.h"
124 #include "recog.h"
125 #include "output.h"
126 #include "addresses.h"
127 #include "expr.h"
128 #include "cfgrtl.h"
129 #include "rtl-error.h"
130 #include "params.h"
131 #include "lra.h"
132 #include "lra-int.h"
133 #include "print-rtl.h"
135 /* Value of LRA_CURR_RELOAD_NUM at the beginning of BB of the current
136 insn. Remember that LRA_CURR_RELOAD_NUM is the number of emitted
137 reload insns. */
138 static int bb_reload_num;
140 /* The current insn being processed and corresponding its single set
141 (NULL otherwise), its data (basic block, the insn data, the insn
142 static data, and the mode of each operand). */
143 static rtx_insn *curr_insn;
144 static rtx curr_insn_set;
145 static basic_block curr_bb;
146 static lra_insn_recog_data_t curr_id;
147 static struct lra_static_insn_data *curr_static_id;
148 static machine_mode curr_operand_mode[MAX_RECOG_OPERANDS];
149 /* Mode of the register substituted by its equivalence with VOIDmode
150 (e.g. constant) and whose subreg is given operand of the current
151 insn. VOIDmode in all other cases. */
152 static machine_mode original_subreg_reg_mode[MAX_RECOG_OPERANDS];
156 /* Start numbers for new registers and insns at the current constraints
157 pass start. */
158 static int new_regno_start;
159 static int new_insn_uid_start;
161 /* If LOC is nonnull, strip any outer subreg from it. */
162 static inline rtx *
163 strip_subreg (rtx *loc)
165 return loc && GET_CODE (*loc) == SUBREG ? &SUBREG_REG (*loc) : loc;
168 /* Return hard regno of REGNO or if it is was not assigned to a hard
169 register, use a hard register from its allocno class. */
170 static int
171 get_try_hard_regno (int regno)
173 int hard_regno;
174 enum reg_class rclass;
176 if ((hard_regno = regno) >= FIRST_PSEUDO_REGISTER)
177 hard_regno = lra_get_regno_hard_regno (regno);
178 if (hard_regno >= 0)
179 return hard_regno;
180 rclass = lra_get_allocno_class (regno);
181 if (rclass == NO_REGS)
182 return -1;
183 return ira_class_hard_regs[rclass][0];
186 /* Return the hard regno of X after removing its subreg. If X is not
187 a register or a subreg of a register, return -1. If X is a pseudo,
188 use its assignment. If FINAL_P return the final hard regno which will
189 be after elimination. */
190 static int
191 get_hard_regno (rtx x, bool final_p)
193 rtx reg;
194 int hard_regno;
196 reg = x;
197 if (SUBREG_P (x))
198 reg = SUBREG_REG (x);
199 if (! REG_P (reg))
200 return -1;
201 if (! HARD_REGISTER_NUM_P (hard_regno = REGNO (reg)))
202 hard_regno = lra_get_regno_hard_regno (hard_regno);
203 if (hard_regno < 0)
204 return -1;
205 if (final_p)
206 hard_regno = lra_get_elimination_hard_regno (hard_regno);
207 if (SUBREG_P (x))
208 hard_regno += subreg_regno_offset (hard_regno, GET_MODE (reg),
209 SUBREG_BYTE (x), GET_MODE (x));
210 return hard_regno;
213 /* If REGNO is a hard register or has been allocated a hard register,
214 return the class of that register. If REGNO is a reload pseudo
215 created by the current constraints pass, return its allocno class.
216 Return NO_REGS otherwise. */
217 static enum reg_class
218 get_reg_class (int regno)
220 int hard_regno;
222 if (! HARD_REGISTER_NUM_P (hard_regno = regno))
223 hard_regno = lra_get_regno_hard_regno (regno);
224 if (hard_regno >= 0)
226 hard_regno = lra_get_elimination_hard_regno (hard_regno);
227 return REGNO_REG_CLASS (hard_regno);
229 if (regno >= new_regno_start)
230 return lra_get_allocno_class (regno);
231 return NO_REGS;
234 /* Return true if REG satisfies (or will satisfy) reg class constraint
235 CL. Use elimination first if REG is a hard register. If REG is a
236 reload pseudo created by this constraints pass, assume that it will
237 be allocated a hard register from its allocno class, but allow that
238 class to be narrowed to CL if it is currently a superset of CL.
240 If NEW_CLASS is nonnull, set *NEW_CLASS to the new allocno class of
241 REGNO (reg), or NO_REGS if no change in its class was needed. */
242 static bool
243 in_class_p (rtx reg, enum reg_class cl, enum reg_class *new_class)
245 enum reg_class rclass, common_class;
246 machine_mode reg_mode;
247 int class_size, hard_regno, nregs, i, j;
248 int regno = REGNO (reg);
250 if (new_class != NULL)
251 *new_class = NO_REGS;
252 if (regno < FIRST_PSEUDO_REGISTER)
254 rtx final_reg = reg;
255 rtx *final_loc = &final_reg;
257 lra_eliminate_reg_if_possible (final_loc);
258 return TEST_HARD_REG_BIT (reg_class_contents[cl], REGNO (*final_loc));
260 reg_mode = GET_MODE (reg);
261 rclass = get_reg_class (regno);
262 if (regno < new_regno_start
263 /* Do not allow the constraints for reload instructions to
264 influence the classes of new pseudos. These reloads are
265 typically moves that have many alternatives, and restricting
266 reload pseudos for one alternative may lead to situations
267 where other reload pseudos are no longer allocatable. */
268 || (INSN_UID (curr_insn) >= new_insn_uid_start
269 && curr_insn_set != NULL
270 && ((OBJECT_P (SET_SRC (curr_insn_set))
271 && ! CONSTANT_P (SET_SRC (curr_insn_set)))
272 || (GET_CODE (SET_SRC (curr_insn_set)) == SUBREG
273 && OBJECT_P (SUBREG_REG (SET_SRC (curr_insn_set)))
274 && ! CONSTANT_P (SUBREG_REG (SET_SRC (curr_insn_set)))))))
275 /* When we don't know what class will be used finally for reload
276 pseudos, we use ALL_REGS. */
277 return ((regno >= new_regno_start && rclass == ALL_REGS)
278 || (rclass != NO_REGS && ira_class_subset_p[rclass][cl]
279 && ! hard_reg_set_subset_p (reg_class_contents[cl],
280 lra_no_alloc_regs)));
281 else
283 common_class = ira_reg_class_subset[rclass][cl];
284 if (new_class != NULL)
285 *new_class = common_class;
286 if (hard_reg_set_subset_p (reg_class_contents[common_class],
287 lra_no_alloc_regs))
288 return false;
289 /* Check that there are enough allocatable regs. */
290 class_size = ira_class_hard_regs_num[common_class];
291 for (i = 0; i < class_size; i++)
293 hard_regno = ira_class_hard_regs[common_class][i];
294 nregs = hard_regno_nregs (hard_regno, reg_mode);
295 if (nregs == 1)
296 return true;
297 for (j = 0; j < nregs; j++)
298 if (TEST_HARD_REG_BIT (lra_no_alloc_regs, hard_regno + j)
299 || ! TEST_HARD_REG_BIT (reg_class_contents[common_class],
300 hard_regno + j))
301 break;
302 if (j >= nregs)
303 return true;
305 return false;
309 /* Return true if REGNO satisfies a memory constraint. */
310 static bool
311 in_mem_p (int regno)
313 return get_reg_class (regno) == NO_REGS;
316 /* Return 1 if ADDR is a valid memory address for mode MODE in address
317 space AS, and check that each pseudo has the proper kind of hard
318 reg. */
319 static int
320 valid_address_p (machine_mode mode ATTRIBUTE_UNUSED,
321 rtx addr, addr_space_t as)
323 #ifdef GO_IF_LEGITIMATE_ADDRESS
324 lra_assert (ADDR_SPACE_GENERIC_P (as));
325 GO_IF_LEGITIMATE_ADDRESS (mode, addr, win);
326 return 0;
328 win:
329 return 1;
330 #else
331 return targetm.addr_space.legitimate_address_p (mode, addr, 0, as);
332 #endif
335 namespace {
336 /* Temporarily eliminates registers in an address (for the lifetime of
337 the object). */
338 class address_eliminator {
339 public:
340 address_eliminator (struct address_info *ad);
341 ~address_eliminator ();
343 private:
344 struct address_info *m_ad;
345 rtx *m_base_loc;
346 rtx m_base_reg;
347 rtx *m_index_loc;
348 rtx m_index_reg;
352 address_eliminator::address_eliminator (struct address_info *ad)
353 : m_ad (ad),
354 m_base_loc (strip_subreg (ad->base_term)),
355 m_base_reg (NULL_RTX),
356 m_index_loc (strip_subreg (ad->index_term)),
357 m_index_reg (NULL_RTX)
359 if (m_base_loc != NULL)
361 m_base_reg = *m_base_loc;
362 lra_eliminate_reg_if_possible (m_base_loc);
363 if (m_ad->base_term2 != NULL)
364 *m_ad->base_term2 = *m_ad->base_term;
366 if (m_index_loc != NULL)
368 m_index_reg = *m_index_loc;
369 lra_eliminate_reg_if_possible (m_index_loc);
373 address_eliminator::~address_eliminator ()
375 if (m_base_loc && *m_base_loc != m_base_reg)
377 *m_base_loc = m_base_reg;
378 if (m_ad->base_term2 != NULL)
379 *m_ad->base_term2 = *m_ad->base_term;
381 if (m_index_loc && *m_index_loc != m_index_reg)
382 *m_index_loc = m_index_reg;
385 /* Return true if the eliminated form of AD is a legitimate target address. */
386 static bool
387 valid_address_p (struct address_info *ad)
389 address_eliminator eliminator (ad);
390 return valid_address_p (ad->mode, *ad->outer, ad->as);
393 /* Return true if the eliminated form of memory reference OP satisfies
394 extra (special) memory constraint CONSTRAINT. */
395 static bool
396 satisfies_memory_constraint_p (rtx op, enum constraint_num constraint)
398 struct address_info ad;
400 decompose_mem_address (&ad, op);
401 address_eliminator eliminator (&ad);
402 return constraint_satisfied_p (op, constraint);
405 /* Return true if the eliminated form of address AD satisfies extra
406 address constraint CONSTRAINT. */
407 static bool
408 satisfies_address_constraint_p (struct address_info *ad,
409 enum constraint_num constraint)
411 address_eliminator eliminator (ad);
412 return constraint_satisfied_p (*ad->outer, constraint);
415 /* Return true if the eliminated form of address OP satisfies extra
416 address constraint CONSTRAINT. */
417 static bool
418 satisfies_address_constraint_p (rtx op, enum constraint_num constraint)
420 struct address_info ad;
422 decompose_lea_address (&ad, &op);
423 return satisfies_address_constraint_p (&ad, constraint);
426 /* Initiate equivalences for LRA. As we keep original equivalences
427 before any elimination, we need to make copies otherwise any change
428 in insns might change the equivalences. */
429 void
430 lra_init_equiv (void)
432 ira_expand_reg_equiv ();
433 for (int i = FIRST_PSEUDO_REGISTER; i < max_reg_num (); i++)
435 rtx res;
437 if ((res = ira_reg_equiv[i].memory) != NULL_RTX)
438 ira_reg_equiv[i].memory = copy_rtx (res);
439 if ((res = ira_reg_equiv[i].invariant) != NULL_RTX)
440 ira_reg_equiv[i].invariant = copy_rtx (res);
444 static rtx loc_equivalence_callback (rtx, const_rtx, void *);
446 /* Update equivalence for REGNO. We need to this as the equivalence
447 might contain other pseudos which are changed by their
448 equivalences. */
449 static void
450 update_equiv (int regno)
452 rtx x;
454 if ((x = ira_reg_equiv[regno].memory) != NULL_RTX)
455 ira_reg_equiv[regno].memory
456 = simplify_replace_fn_rtx (x, NULL_RTX, loc_equivalence_callback,
457 NULL_RTX);
458 if ((x = ira_reg_equiv[regno].invariant) != NULL_RTX)
459 ira_reg_equiv[regno].invariant
460 = simplify_replace_fn_rtx (x, NULL_RTX, loc_equivalence_callback,
461 NULL_RTX);
464 /* If we have decided to substitute X with another value, return that
465 value, otherwise return X. */
466 static rtx
467 get_equiv (rtx x)
469 int regno;
470 rtx res;
472 if (! REG_P (x) || (regno = REGNO (x)) < FIRST_PSEUDO_REGISTER
473 || ! ira_reg_equiv[regno].defined_p
474 || ! ira_reg_equiv[regno].profitable_p
475 || lra_get_regno_hard_regno (regno) >= 0)
476 return x;
477 if ((res = ira_reg_equiv[regno].memory) != NULL_RTX)
479 if (targetm.cannot_substitute_mem_equiv_p (res))
480 return x;
481 return res;
483 if ((res = ira_reg_equiv[regno].constant) != NULL_RTX)
484 return res;
485 if ((res = ira_reg_equiv[regno].invariant) != NULL_RTX)
486 return res;
487 gcc_unreachable ();
490 /* If we have decided to substitute X with the equivalent value,
491 return that value after elimination for INSN, otherwise return
492 X. */
493 static rtx
494 get_equiv_with_elimination (rtx x, rtx_insn *insn)
496 rtx res = get_equiv (x);
498 if (x == res || CONSTANT_P (res))
499 return res;
500 return lra_eliminate_regs_1 (insn, res, GET_MODE (res),
501 false, false, 0, true);
504 /* Set up curr_operand_mode. */
505 static void
506 init_curr_operand_mode (void)
508 int nop = curr_static_id->n_operands;
509 for (int i = 0; i < nop; i++)
511 machine_mode mode = GET_MODE (*curr_id->operand_loc[i]);
512 if (mode == VOIDmode)
514 /* The .md mode for address operands is the mode of the
515 addressed value rather than the mode of the address itself. */
516 if (curr_id->icode >= 0 && curr_static_id->operand[i].is_address)
517 mode = Pmode;
518 else
519 mode = curr_static_id->operand[i].mode;
521 curr_operand_mode[i] = mode;
527 /* The page contains code to reuse input reloads. */
529 /* Structure describes input reload of the current insns. */
530 struct input_reload
532 /* True for input reload of matched operands. */
533 bool match_p;
534 /* Reloaded value. */
535 rtx input;
536 /* Reload pseudo used. */
537 rtx reg;
540 /* The number of elements in the following array. */
541 static int curr_insn_input_reloads_num;
542 /* Array containing info about input reloads. It is used to find the
543 same input reload and reuse the reload pseudo in this case. */
544 static struct input_reload curr_insn_input_reloads[LRA_MAX_INSN_RELOADS];
546 /* Initiate data concerning reuse of input reloads for the current
547 insn. */
548 static void
549 init_curr_insn_input_reloads (void)
551 curr_insn_input_reloads_num = 0;
554 /* Create a new pseudo using MODE, RCLASS, ORIGINAL or reuse already
555 created input reload pseudo (only if TYPE is not OP_OUT). Don't
556 reuse pseudo if IN_SUBREG_P is true and the reused pseudo should be
557 wrapped up in SUBREG. The result pseudo is returned through
558 RESULT_REG. Return TRUE if we created a new pseudo, FALSE if we
559 reused the already created input reload pseudo. Use TITLE to
560 describe new registers for debug purposes. */
561 static bool
562 get_reload_reg (enum op_type type, machine_mode mode, rtx original,
563 enum reg_class rclass, bool in_subreg_p,
564 const char *title, rtx *result_reg)
566 int i, regno;
567 enum reg_class new_class;
568 bool unique_p = false;
570 if (type == OP_OUT)
572 *result_reg
573 = lra_create_new_reg_with_unique_value (mode, original, rclass, title);
574 return true;
576 /* Prevent reuse value of expression with side effects,
577 e.g. volatile memory. */
578 if (! side_effects_p (original))
579 for (i = 0; i < curr_insn_input_reloads_num; i++)
581 if (! curr_insn_input_reloads[i].match_p
582 && rtx_equal_p (curr_insn_input_reloads[i].input, original)
583 && in_class_p (curr_insn_input_reloads[i].reg, rclass, &new_class))
585 rtx reg = curr_insn_input_reloads[i].reg;
586 regno = REGNO (reg);
587 /* If input is equal to original and both are VOIDmode,
588 GET_MODE (reg) might be still different from mode.
589 Ensure we don't return *result_reg with wrong mode. */
590 if (GET_MODE (reg) != mode)
592 if (in_subreg_p)
593 continue;
594 if (GET_MODE_SIZE (GET_MODE (reg)) < GET_MODE_SIZE (mode))
595 continue;
596 reg = lowpart_subreg (mode, reg, GET_MODE (reg));
597 if (reg == NULL_RTX || GET_CODE (reg) != SUBREG)
598 continue;
600 *result_reg = reg;
601 if (lra_dump_file != NULL)
603 fprintf (lra_dump_file, " Reuse r%d for reload ", regno);
604 dump_value_slim (lra_dump_file, original, 1);
606 if (new_class != lra_get_allocno_class (regno))
607 lra_change_class (regno, new_class, ", change to", false);
608 if (lra_dump_file != NULL)
609 fprintf (lra_dump_file, "\n");
610 return false;
612 /* If we have an input reload with a different mode, make sure it
613 will get a different hard reg. */
614 else if (REG_P (original)
615 && REG_P (curr_insn_input_reloads[i].input)
616 && REGNO (original) == REGNO (curr_insn_input_reloads[i].input)
617 && (GET_MODE (original)
618 != GET_MODE (curr_insn_input_reloads[i].input)))
619 unique_p = true;
621 *result_reg = (unique_p
622 ? lra_create_new_reg_with_unique_value
623 : lra_create_new_reg) (mode, original, rclass, title);
624 lra_assert (curr_insn_input_reloads_num < LRA_MAX_INSN_RELOADS);
625 curr_insn_input_reloads[curr_insn_input_reloads_num].input = original;
626 curr_insn_input_reloads[curr_insn_input_reloads_num].match_p = false;
627 curr_insn_input_reloads[curr_insn_input_reloads_num++].reg = *result_reg;
628 return true;
633 /* The page contains code to extract memory address parts. */
635 /* Wrapper around REGNO_OK_FOR_INDEX_P, to allow pseudos. */
636 static inline bool
637 ok_for_index_p_nonstrict (rtx reg)
639 unsigned regno = REGNO (reg);
641 return regno >= FIRST_PSEUDO_REGISTER || REGNO_OK_FOR_INDEX_P (regno);
644 /* A version of regno_ok_for_base_p for use here, when all pseudos
645 should count as OK. Arguments as for regno_ok_for_base_p. */
646 static inline bool
647 ok_for_base_p_nonstrict (rtx reg, machine_mode mode, addr_space_t as,
648 enum rtx_code outer_code, enum rtx_code index_code)
650 unsigned regno = REGNO (reg);
652 if (regno >= FIRST_PSEUDO_REGISTER)
653 return true;
654 return ok_for_base_p_1 (regno, mode, as, outer_code, index_code);
659 /* The page contains major code to choose the current insn alternative
660 and generate reloads for it. */
662 /* Return the offset from REGNO of the least significant register
663 in (reg:MODE REGNO).
665 This function is used to tell whether two registers satisfy
666 a matching constraint. (reg:MODE1 REGNO1) matches (reg:MODE2 REGNO2) if:
668 REGNO1 + lra_constraint_offset (REGNO1, MODE1)
669 == REGNO2 + lra_constraint_offset (REGNO2, MODE2) */
671 lra_constraint_offset (int regno, machine_mode mode)
673 lra_assert (regno < FIRST_PSEUDO_REGISTER);
675 scalar_int_mode int_mode;
676 if (WORDS_BIG_ENDIAN
677 && is_a <scalar_int_mode> (mode, &int_mode)
678 && GET_MODE_SIZE (int_mode) > UNITS_PER_WORD)
679 return hard_regno_nregs (regno, mode) - 1;
680 return 0;
683 /* Like rtx_equal_p except that it allows a REG and a SUBREG to match
684 if they are the same hard reg, and has special hacks for
685 auto-increment and auto-decrement. This is specifically intended for
686 process_alt_operands to use in determining whether two operands
687 match. X is the operand whose number is the lower of the two.
689 It is supposed that X is the output operand and Y is the input
690 operand. Y_HARD_REGNO is the final hard regno of register Y or
691 register in subreg Y as we know it now. Otherwise, it is a
692 negative value. */
693 static bool
694 operands_match_p (rtx x, rtx y, int y_hard_regno)
696 int i;
697 RTX_CODE code = GET_CODE (x);
698 const char *fmt;
700 if (x == y)
701 return true;
702 if ((code == REG || (code == SUBREG && REG_P (SUBREG_REG (x))))
703 && (REG_P (y) || (GET_CODE (y) == SUBREG && REG_P (SUBREG_REG (y)))))
705 int j;
707 i = get_hard_regno (x, false);
708 if (i < 0)
709 goto slow;
711 if ((j = y_hard_regno) < 0)
712 goto slow;
714 i += lra_constraint_offset (i, GET_MODE (x));
715 j += lra_constraint_offset (j, GET_MODE (y));
717 return i == j;
720 /* If two operands must match, because they are really a single
721 operand of an assembler insn, then two post-increments are invalid
722 because the assembler insn would increment only once. On the
723 other hand, a post-increment matches ordinary indexing if the
724 post-increment is the output operand. */
725 if (code == POST_DEC || code == POST_INC || code == POST_MODIFY)
726 return operands_match_p (XEXP (x, 0), y, y_hard_regno);
728 /* Two pre-increments are invalid because the assembler insn would
729 increment only once. On the other hand, a pre-increment matches
730 ordinary indexing if the pre-increment is the input operand. */
731 if (GET_CODE (y) == PRE_DEC || GET_CODE (y) == PRE_INC
732 || GET_CODE (y) == PRE_MODIFY)
733 return operands_match_p (x, XEXP (y, 0), -1);
735 slow:
737 if (code == REG && REG_P (y))
738 return REGNO (x) == REGNO (y);
740 if (code == REG && GET_CODE (y) == SUBREG && REG_P (SUBREG_REG (y))
741 && x == SUBREG_REG (y))
742 return true;
743 if (GET_CODE (y) == REG && code == SUBREG && REG_P (SUBREG_REG (x))
744 && SUBREG_REG (x) == y)
745 return true;
747 /* Now we have disposed of all the cases in which different rtx
748 codes can match. */
749 if (code != GET_CODE (y))
750 return false;
752 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
753 if (GET_MODE (x) != GET_MODE (y))
754 return false;
756 switch (code)
758 CASE_CONST_UNIQUE:
759 return false;
761 case LABEL_REF:
762 return label_ref_label (x) == label_ref_label (y);
763 case SYMBOL_REF:
764 return XSTR (x, 0) == XSTR (y, 0);
766 default:
767 break;
770 /* Compare the elements. If any pair of corresponding elements fail
771 to match, return false for the whole things. */
773 fmt = GET_RTX_FORMAT (code);
774 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
776 int val, j;
777 switch (fmt[i])
779 case 'w':
780 if (XWINT (x, i) != XWINT (y, i))
781 return false;
782 break;
784 case 'i':
785 if (XINT (x, i) != XINT (y, i))
786 return false;
787 break;
789 case 'e':
790 val = operands_match_p (XEXP (x, i), XEXP (y, i), -1);
791 if (val == 0)
792 return false;
793 break;
795 case '0':
796 break;
798 case 'E':
799 if (XVECLEN (x, i) != XVECLEN (y, i))
800 return false;
801 for (j = XVECLEN (x, i) - 1; j >= 0; --j)
803 val = operands_match_p (XVECEXP (x, i, j), XVECEXP (y, i, j), -1);
804 if (val == 0)
805 return false;
807 break;
809 /* It is believed that rtx's at this level will never
810 contain anything but integers and other rtx's, except for
811 within LABEL_REFs and SYMBOL_REFs. */
812 default:
813 gcc_unreachable ();
816 return true;
819 /* True if X is a constant that can be forced into the constant pool.
820 MODE is the mode of the operand, or VOIDmode if not known. */
821 #define CONST_POOL_OK_P(MODE, X) \
822 ((MODE) != VOIDmode \
823 && CONSTANT_P (X) \
824 && GET_CODE (X) != HIGH \
825 && !targetm.cannot_force_const_mem (MODE, X))
827 /* True if C is a non-empty register class that has too few registers
828 to be safely used as a reload target class. */
829 #define SMALL_REGISTER_CLASS_P(C) \
830 (ira_class_hard_regs_num [(C)] == 1 \
831 || (ira_class_hard_regs_num [(C)] >= 1 \
832 && targetm.class_likely_spilled_p (C)))
834 /* If REG is a reload pseudo, try to make its class satisfying CL. */
835 static void
836 narrow_reload_pseudo_class (rtx reg, enum reg_class cl)
838 enum reg_class rclass;
840 /* Do not make more accurate class from reloads generated. They are
841 mostly moves with a lot of constraints. Making more accurate
842 class may results in very narrow class and impossibility of find
843 registers for several reloads of one insn. */
844 if (INSN_UID (curr_insn) >= new_insn_uid_start)
845 return;
846 if (GET_CODE (reg) == SUBREG)
847 reg = SUBREG_REG (reg);
848 if (! REG_P (reg) || (int) REGNO (reg) < new_regno_start)
849 return;
850 if (in_class_p (reg, cl, &rclass) && rclass != cl)
851 lra_change_class (REGNO (reg), rclass, " Change to", true);
854 /* Searches X for any reference to a reg with the same value as REGNO,
855 returning the rtx of the reference found if any. Otherwise,
856 returns NULL_RTX. */
857 static rtx
858 regno_val_use_in (unsigned int regno, rtx x)
860 const char *fmt;
861 int i, j;
862 rtx tem;
864 if (REG_P (x) && lra_reg_info[REGNO (x)].val == lra_reg_info[regno].val)
865 return x;
867 fmt = GET_RTX_FORMAT (GET_CODE (x));
868 for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
870 if (fmt[i] == 'e')
872 if ((tem = regno_val_use_in (regno, XEXP (x, i))))
873 return tem;
875 else if (fmt[i] == 'E')
876 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
877 if ((tem = regno_val_use_in (regno , XVECEXP (x, i, j))))
878 return tem;
881 return NULL_RTX;
884 /* Return true if all current insn non-output operands except INS (it
885 has a negaitve end marker) do not use pseudos with the same value
886 as REGNO. */
887 static bool
888 check_conflict_input_operands (int regno, signed char *ins)
890 int in;
891 int n_operands = curr_static_id->n_operands;
893 for (int nop = 0; nop < n_operands; nop++)
894 if (! curr_static_id->operand[nop].is_operator
895 && curr_static_id->operand[nop].type != OP_OUT)
897 for (int i = 0; (in = ins[i]) >= 0; i++)
898 if (in == nop)
899 break;
900 if (in < 0
901 && regno_val_use_in (regno, *curr_id->operand_loc[nop]) != NULL_RTX)
902 return false;
904 return true;
907 /* Generate reloads for matching OUT and INS (array of input operand
908 numbers with end marker -1) with reg class GOAL_CLASS, considering
909 output operands OUTS (similar array to INS) needing to be in different
910 registers. Add input and output reloads correspondingly to the lists
911 *BEFORE and *AFTER. OUT might be negative. In this case we generate
912 input reloads for matched input operands INS. EARLY_CLOBBER_P is a flag
913 that the output operand is early clobbered for chosen alternative. */
914 static void
915 match_reload (signed char out, signed char *ins, signed char *outs,
916 enum reg_class goal_class, rtx_insn **before,
917 rtx_insn **after, bool early_clobber_p)
919 bool out_conflict;
920 int i, in;
921 rtx new_in_reg, new_out_reg, reg;
922 machine_mode inmode, outmode;
923 rtx in_rtx = *curr_id->operand_loc[ins[0]];
924 rtx out_rtx = out < 0 ? in_rtx : *curr_id->operand_loc[out];
926 inmode = curr_operand_mode[ins[0]];
927 outmode = out < 0 ? inmode : curr_operand_mode[out];
928 push_to_sequence (*before);
929 if (inmode != outmode)
931 if (partial_subreg_p (outmode, inmode))
933 reg = new_in_reg
934 = lra_create_new_reg_with_unique_value (inmode, in_rtx,
935 goal_class, "");
936 if (SCALAR_INT_MODE_P (inmode))
937 new_out_reg = gen_lowpart_SUBREG (outmode, reg);
938 else
939 new_out_reg = gen_rtx_SUBREG (outmode, reg, 0);
940 LRA_SUBREG_P (new_out_reg) = 1;
941 /* If the input reg is dying here, we can use the same hard
942 register for REG and IN_RTX. We do it only for original
943 pseudos as reload pseudos can die although original
944 pseudos still live where reload pseudos dies. */
945 if (REG_P (in_rtx) && (int) REGNO (in_rtx) < lra_new_regno_start
946 && find_regno_note (curr_insn, REG_DEAD, REGNO (in_rtx))
947 && (!early_clobber_p
948 || check_conflict_input_operands(REGNO (in_rtx), ins)))
949 lra_assign_reg_val (REGNO (in_rtx), REGNO (reg));
951 else
953 reg = new_out_reg
954 = lra_create_new_reg_with_unique_value (outmode, out_rtx,
955 goal_class, "");
956 if (SCALAR_INT_MODE_P (outmode))
957 new_in_reg = gen_lowpart_SUBREG (inmode, reg);
958 else
959 new_in_reg = gen_rtx_SUBREG (inmode, reg, 0);
960 /* NEW_IN_REG is non-paradoxical subreg. We don't want
961 NEW_OUT_REG living above. We add clobber clause for
962 this. This is just a temporary clobber. We can remove
963 it at the end of LRA work. */
964 rtx_insn *clobber = emit_clobber (new_out_reg);
965 LRA_TEMP_CLOBBER_P (PATTERN (clobber)) = 1;
966 LRA_SUBREG_P (new_in_reg) = 1;
967 if (GET_CODE (in_rtx) == SUBREG)
969 rtx subreg_reg = SUBREG_REG (in_rtx);
971 /* If SUBREG_REG is dying here and sub-registers IN_RTX
972 and NEW_IN_REG are similar, we can use the same hard
973 register for REG and SUBREG_REG. */
974 if (REG_P (subreg_reg)
975 && (int) REGNO (subreg_reg) < lra_new_regno_start
976 && GET_MODE (subreg_reg) == outmode
977 && SUBREG_BYTE (in_rtx) == SUBREG_BYTE (new_in_reg)
978 && find_regno_note (curr_insn, REG_DEAD, REGNO (subreg_reg))
979 && (! early_clobber_p
980 || check_conflict_input_operands (REGNO (subreg_reg),
981 ins)))
982 lra_assign_reg_val (REGNO (subreg_reg), REGNO (reg));
986 else
988 /* Pseudos have values -- see comments for lra_reg_info.
989 Different pseudos with the same value do not conflict even if
990 they live in the same place. When we create a pseudo we
991 assign value of original pseudo (if any) from which we
992 created the new pseudo. If we create the pseudo from the
993 input pseudo, the new pseudo will have no conflict with the
994 input pseudo which is wrong when the input pseudo lives after
995 the insn and as the new pseudo value is changed by the insn
996 output. Therefore we create the new pseudo from the output
997 except the case when we have single matched dying input
998 pseudo.
1000 We cannot reuse the current output register because we might
1001 have a situation like "a <- a op b", where the constraints
1002 force the second input operand ("b") to match the output
1003 operand ("a"). "b" must then be copied into a new register
1004 so that it doesn't clobber the current value of "a".
1006 We can not use the same value if the output pseudo is
1007 early clobbered or the input pseudo is mentioned in the
1008 output, e.g. as an address part in memory, because
1009 output reload will actually extend the pseudo liveness.
1010 We don't care about eliminable hard regs here as we are
1011 interesting only in pseudos. */
1013 /* Matching input's register value is the same as one of the other
1014 output operand. Output operands in a parallel insn must be in
1015 different registers. */
1016 out_conflict = false;
1017 if (REG_P (in_rtx))
1019 for (i = 0; outs[i] >= 0; i++)
1021 rtx other_out_rtx = *curr_id->operand_loc[outs[i]];
1022 if (REG_P (other_out_rtx)
1023 && (regno_val_use_in (REGNO (in_rtx), other_out_rtx)
1024 != NULL_RTX))
1026 out_conflict = true;
1027 break;
1032 new_in_reg = new_out_reg
1033 = (! early_clobber_p && ins[1] < 0 && REG_P (in_rtx)
1034 && (int) REGNO (in_rtx) < lra_new_regno_start
1035 && find_regno_note (curr_insn, REG_DEAD, REGNO (in_rtx))
1036 && (! early_clobber_p
1037 || check_conflict_input_operands (REGNO (in_rtx), ins))
1038 && (out < 0
1039 || regno_val_use_in (REGNO (in_rtx), out_rtx) == NULL_RTX)
1040 && !out_conflict
1041 ? lra_create_new_reg (inmode, in_rtx, goal_class, "")
1042 : lra_create_new_reg_with_unique_value (outmode, out_rtx,
1043 goal_class, ""));
1045 /* In operand can be got from transformations before processing insn
1046 constraints. One example of such transformations is subreg
1047 reloading (see function simplify_operand_subreg). The new
1048 pseudos created by the transformations might have inaccurate
1049 class (ALL_REGS) and we should make their classes more
1050 accurate. */
1051 narrow_reload_pseudo_class (in_rtx, goal_class);
1052 lra_emit_move (copy_rtx (new_in_reg), in_rtx);
1053 *before = get_insns ();
1054 end_sequence ();
1055 /* Add the new pseudo to consider values of subsequent input reload
1056 pseudos. */
1057 lra_assert (curr_insn_input_reloads_num < LRA_MAX_INSN_RELOADS);
1058 curr_insn_input_reloads[curr_insn_input_reloads_num].input = in_rtx;
1059 curr_insn_input_reloads[curr_insn_input_reloads_num].match_p = true;
1060 curr_insn_input_reloads[curr_insn_input_reloads_num++].reg = new_in_reg;
1061 for (i = 0; (in = ins[i]) >= 0; i++)
1063 lra_assert
1064 (GET_MODE (*curr_id->operand_loc[in]) == VOIDmode
1065 || GET_MODE (new_in_reg) == GET_MODE (*curr_id->operand_loc[in]));
1066 *curr_id->operand_loc[in] = new_in_reg;
1068 lra_update_dups (curr_id, ins);
1069 if (out < 0)
1070 return;
1071 /* See a comment for the input operand above. */
1072 narrow_reload_pseudo_class (out_rtx, goal_class);
1073 if (find_reg_note (curr_insn, REG_UNUSED, out_rtx) == NULL_RTX)
1075 start_sequence ();
1076 lra_emit_move (out_rtx, copy_rtx (new_out_reg));
1077 emit_insn (*after);
1078 *after = get_insns ();
1079 end_sequence ();
1081 *curr_id->operand_loc[out] = new_out_reg;
1082 lra_update_dup (curr_id, out);
1085 /* Return register class which is union of all reg classes in insn
1086 constraint alternative string starting with P. */
1087 static enum reg_class
1088 reg_class_from_constraints (const char *p)
1090 int c, len;
1091 enum reg_class op_class = NO_REGS;
1094 switch ((c = *p, len = CONSTRAINT_LEN (c, p)), c)
1096 case '#':
1097 case ',':
1098 return op_class;
1100 case 'g':
1101 op_class = reg_class_subunion[op_class][GENERAL_REGS];
1102 break;
1104 default:
1105 enum constraint_num cn = lookup_constraint (p);
1106 enum reg_class cl = reg_class_for_constraint (cn);
1107 if (cl == NO_REGS)
1109 if (insn_extra_address_constraint (cn))
1110 op_class
1111 = (reg_class_subunion
1112 [op_class][base_reg_class (VOIDmode, ADDR_SPACE_GENERIC,
1113 ADDRESS, SCRATCH)]);
1114 break;
1117 op_class = reg_class_subunion[op_class][cl];
1118 break;
1120 while ((p += len), c);
1121 return op_class;
1124 /* If OP is a register, return the class of the register as per
1125 get_reg_class, otherwise return NO_REGS. */
1126 static inline enum reg_class
1127 get_op_class (rtx op)
1129 return REG_P (op) ? get_reg_class (REGNO (op)) : NO_REGS;
1132 /* Return generated insn mem_pseudo:=val if TO_P or val:=mem_pseudo
1133 otherwise. If modes of MEM_PSEUDO and VAL are different, use
1134 SUBREG for VAL to make them equal. */
1135 static rtx_insn *
1136 emit_spill_move (bool to_p, rtx mem_pseudo, rtx val)
1138 if (GET_MODE (mem_pseudo) != GET_MODE (val))
1140 /* Usually size of mem_pseudo is greater than val size but in
1141 rare cases it can be less as it can be defined by target
1142 dependent macro HARD_REGNO_CALLER_SAVE_MODE. */
1143 if (! MEM_P (val))
1145 val = gen_lowpart_SUBREG (GET_MODE (mem_pseudo),
1146 GET_CODE (val) == SUBREG
1147 ? SUBREG_REG (val) : val);
1148 LRA_SUBREG_P (val) = 1;
1150 else
1152 mem_pseudo = gen_lowpart_SUBREG (GET_MODE (val), mem_pseudo);
1153 LRA_SUBREG_P (mem_pseudo) = 1;
1156 return to_p ? gen_move_insn (mem_pseudo, val)
1157 : gen_move_insn (val, mem_pseudo);
1160 /* Process a special case insn (register move), return true if we
1161 don't need to process it anymore. INSN should be a single set
1162 insn. Set up that RTL was changed through CHANGE_P and that hook
1163 TARGET_SECONDARY_MEMORY_NEEDED says to use secondary memory through
1164 SEC_MEM_P. */
1165 static bool
1166 check_and_process_move (bool *change_p, bool *sec_mem_p ATTRIBUTE_UNUSED)
1168 int sregno, dregno;
1169 rtx dest, src, dreg, sreg, new_reg, scratch_reg;
1170 rtx_insn *before;
1171 enum reg_class dclass, sclass, secondary_class;
1172 secondary_reload_info sri;
1174 lra_assert (curr_insn_set != NULL_RTX);
1175 dreg = dest = SET_DEST (curr_insn_set);
1176 sreg = src = SET_SRC (curr_insn_set);
1177 if (GET_CODE (dest) == SUBREG)
1178 dreg = SUBREG_REG (dest);
1179 if (GET_CODE (src) == SUBREG)
1180 sreg = SUBREG_REG (src);
1181 if (! (REG_P (dreg) || MEM_P (dreg)) || ! (REG_P (sreg) || MEM_P (sreg)))
1182 return false;
1183 sclass = dclass = NO_REGS;
1184 if (REG_P (dreg))
1185 dclass = get_reg_class (REGNO (dreg));
1186 gcc_assert (dclass < LIM_REG_CLASSES);
1187 if (dclass == ALL_REGS)
1188 /* ALL_REGS is used for new pseudos created by transformations
1189 like reload of SUBREG_REG (see function
1190 simplify_operand_subreg). We don't know their class yet. We
1191 should figure out the class from processing the insn
1192 constraints not in this fast path function. Even if ALL_REGS
1193 were a right class for the pseudo, secondary_... hooks usually
1194 are not define for ALL_REGS. */
1195 return false;
1196 if (REG_P (sreg))
1197 sclass = get_reg_class (REGNO (sreg));
1198 gcc_assert (sclass < LIM_REG_CLASSES);
1199 if (sclass == ALL_REGS)
1200 /* See comments above. */
1201 return false;
1202 if (sclass == NO_REGS && dclass == NO_REGS)
1203 return false;
1204 if (targetm.secondary_memory_needed (GET_MODE (src), sclass, dclass)
1205 && ((sclass != NO_REGS && dclass != NO_REGS)
1206 || (GET_MODE (src)
1207 != targetm.secondary_memory_needed_mode (GET_MODE (src)))))
1209 *sec_mem_p = true;
1210 return false;
1212 if (! REG_P (dreg) || ! REG_P (sreg))
1213 return false;
1214 sri.prev_sri = NULL;
1215 sri.icode = CODE_FOR_nothing;
1216 sri.extra_cost = 0;
1217 secondary_class = NO_REGS;
1218 /* Set up hard register for a reload pseudo for hook
1219 secondary_reload because some targets just ignore unassigned
1220 pseudos in the hook. */
1221 if (dclass != NO_REGS && lra_get_regno_hard_regno (REGNO (dreg)) < 0)
1223 dregno = REGNO (dreg);
1224 reg_renumber[dregno] = ira_class_hard_regs[dclass][0];
1226 else
1227 dregno = -1;
1228 if (sclass != NO_REGS && lra_get_regno_hard_regno (REGNO (sreg)) < 0)
1230 sregno = REGNO (sreg);
1231 reg_renumber[sregno] = ira_class_hard_regs[sclass][0];
1233 else
1234 sregno = -1;
1235 if (sclass != NO_REGS)
1236 secondary_class
1237 = (enum reg_class) targetm.secondary_reload (false, dest,
1238 (reg_class_t) sclass,
1239 GET_MODE (src), &sri);
1240 if (sclass == NO_REGS
1241 || ((secondary_class != NO_REGS || sri.icode != CODE_FOR_nothing)
1242 && dclass != NO_REGS))
1244 enum reg_class old_sclass = secondary_class;
1245 secondary_reload_info old_sri = sri;
1247 sri.prev_sri = NULL;
1248 sri.icode = CODE_FOR_nothing;
1249 sri.extra_cost = 0;
1250 secondary_class
1251 = (enum reg_class) targetm.secondary_reload (true, src,
1252 (reg_class_t) dclass,
1253 GET_MODE (src), &sri);
1254 /* Check the target hook consistency. */
1255 lra_assert
1256 ((secondary_class == NO_REGS && sri.icode == CODE_FOR_nothing)
1257 || (old_sclass == NO_REGS && old_sri.icode == CODE_FOR_nothing)
1258 || (secondary_class == old_sclass && sri.icode == old_sri.icode));
1260 if (sregno >= 0)
1261 reg_renumber [sregno] = -1;
1262 if (dregno >= 0)
1263 reg_renumber [dregno] = -1;
1264 if (secondary_class == NO_REGS && sri.icode == CODE_FOR_nothing)
1265 return false;
1266 *change_p = true;
1267 new_reg = NULL_RTX;
1268 if (secondary_class != NO_REGS)
1269 new_reg = lra_create_new_reg_with_unique_value (GET_MODE (src), NULL_RTX,
1270 secondary_class,
1271 "secondary");
1272 start_sequence ();
1273 if (sri.icode == CODE_FOR_nothing)
1274 lra_emit_move (new_reg, src);
1275 else
1277 enum reg_class scratch_class;
1279 scratch_class = (reg_class_from_constraints
1280 (insn_data[sri.icode].operand[2].constraint));
1281 scratch_reg = (lra_create_new_reg_with_unique_value
1282 (insn_data[sri.icode].operand[2].mode, NULL_RTX,
1283 scratch_class, "scratch"));
1284 emit_insn (GEN_FCN (sri.icode) (new_reg != NULL_RTX ? new_reg : dest,
1285 src, scratch_reg));
1287 before = get_insns ();
1288 end_sequence ();
1289 lra_process_new_insns (curr_insn, before, NULL, "Inserting the move");
1290 if (new_reg != NULL_RTX)
1291 SET_SRC (curr_insn_set) = new_reg;
1292 else
1294 if (lra_dump_file != NULL)
1296 fprintf (lra_dump_file, "Deleting move %u\n", INSN_UID (curr_insn));
1297 dump_insn_slim (lra_dump_file, curr_insn);
1299 lra_set_insn_deleted (curr_insn);
1300 return true;
1302 return false;
1305 /* The following data describe the result of process_alt_operands.
1306 The data are used in curr_insn_transform to generate reloads. */
1308 /* The chosen reg classes which should be used for the corresponding
1309 operands. */
1310 static enum reg_class goal_alt[MAX_RECOG_OPERANDS];
1311 /* True if the operand should be the same as another operand and that
1312 other operand does not need a reload. */
1313 static bool goal_alt_match_win[MAX_RECOG_OPERANDS];
1314 /* True if the operand does not need a reload. */
1315 static bool goal_alt_win[MAX_RECOG_OPERANDS];
1316 /* True if the operand can be offsetable memory. */
1317 static bool goal_alt_offmemok[MAX_RECOG_OPERANDS];
1318 /* The number of an operand to which given operand can be matched to. */
1319 static int goal_alt_matches[MAX_RECOG_OPERANDS];
1320 /* The number of elements in the following array. */
1321 static int goal_alt_dont_inherit_ops_num;
1322 /* Numbers of operands whose reload pseudos should not be inherited. */
1323 static int goal_alt_dont_inherit_ops[MAX_RECOG_OPERANDS];
1324 /* True if the insn commutative operands should be swapped. */
1325 static bool goal_alt_swapped;
1326 /* The chosen insn alternative. */
1327 static int goal_alt_number;
1329 /* True if the corresponding operand is the result of an equivalence
1330 substitution. */
1331 static bool equiv_substition_p[MAX_RECOG_OPERANDS];
1333 /* The following five variables are used to choose the best insn
1334 alternative. They reflect final characteristics of the best
1335 alternative. */
1337 /* Number of necessary reloads and overall cost reflecting the
1338 previous value and other unpleasantness of the best alternative. */
1339 static int best_losers, best_overall;
1340 /* Overall number hard registers used for reloads. For example, on
1341 some targets we need 2 general registers to reload DFmode and only
1342 one floating point register. */
1343 static int best_reload_nregs;
1344 /* Overall number reflecting distances of previous reloading the same
1345 value. The distances are counted from the current BB start. It is
1346 used to improve inheritance chances. */
1347 static int best_reload_sum;
1349 /* True if the current insn should have no correspondingly input or
1350 output reloads. */
1351 static bool no_input_reloads_p, no_output_reloads_p;
1353 /* True if we swapped the commutative operands in the current
1354 insn. */
1355 static int curr_swapped;
1357 /* if CHECK_ONLY_P is false, arrange for address element *LOC to be a
1358 register of class CL. Add any input reloads to list BEFORE. AFTER
1359 is nonnull if *LOC is an automodified value; handle that case by
1360 adding the required output reloads to list AFTER. Return true if
1361 the RTL was changed.
1363 if CHECK_ONLY_P is true, check that the *LOC is a correct address
1364 register. Return false if the address register is correct. */
1365 static bool
1366 process_addr_reg (rtx *loc, bool check_only_p, rtx_insn **before, rtx_insn **after,
1367 enum reg_class cl)
1369 int regno;
1370 enum reg_class rclass, new_class;
1371 rtx reg;
1372 rtx new_reg;
1373 machine_mode mode;
1374 bool subreg_p, before_p = false;
1376 subreg_p = GET_CODE (*loc) == SUBREG;
1377 if (subreg_p)
1379 reg = SUBREG_REG (*loc);
1380 mode = GET_MODE (reg);
1382 /* For mode with size bigger than ptr_mode, there unlikely to be "mov"
1383 between two registers with different classes, but there normally will
1384 be "mov" which transfers element of vector register into the general
1385 register, and this normally will be a subreg which should be reloaded
1386 as a whole. This is particularly likely to be triggered when
1387 -fno-split-wide-types specified. */
1388 if (!REG_P (reg)
1389 || in_class_p (reg, cl, &new_class)
1390 || GET_MODE_SIZE (mode) <= GET_MODE_SIZE (ptr_mode))
1391 loc = &SUBREG_REG (*loc);
1394 reg = *loc;
1395 mode = GET_MODE (reg);
1396 if (! REG_P (reg))
1398 if (check_only_p)
1399 return true;
1400 /* Always reload memory in an address even if the target supports
1401 such addresses. */
1402 new_reg = lra_create_new_reg_with_unique_value (mode, reg, cl, "address");
1403 before_p = true;
1405 else
1407 regno = REGNO (reg);
1408 rclass = get_reg_class (regno);
1409 if (! check_only_p
1410 && (*loc = get_equiv_with_elimination (reg, curr_insn)) != reg)
1412 if (lra_dump_file != NULL)
1414 fprintf (lra_dump_file,
1415 "Changing pseudo %d in address of insn %u on equiv ",
1416 REGNO (reg), INSN_UID (curr_insn));
1417 dump_value_slim (lra_dump_file, *loc, 1);
1418 fprintf (lra_dump_file, "\n");
1420 *loc = copy_rtx (*loc);
1422 if (*loc != reg || ! in_class_p (reg, cl, &new_class))
1424 if (check_only_p)
1425 return true;
1426 reg = *loc;
1427 if (get_reload_reg (after == NULL ? OP_IN : OP_INOUT,
1428 mode, reg, cl, subreg_p, "address", &new_reg))
1429 before_p = true;
1431 else if (new_class != NO_REGS && rclass != new_class)
1433 if (check_only_p)
1434 return true;
1435 lra_change_class (regno, new_class, " Change to", true);
1436 return false;
1438 else
1439 return false;
1441 if (before_p)
1443 push_to_sequence (*before);
1444 lra_emit_move (new_reg, reg);
1445 *before = get_insns ();
1446 end_sequence ();
1448 *loc = new_reg;
1449 if (after != NULL)
1451 start_sequence ();
1452 lra_emit_move (before_p ? copy_rtx (reg) : reg, new_reg);
1453 emit_insn (*after);
1454 *after = get_insns ();
1455 end_sequence ();
1457 return true;
1460 /* Insert move insn in simplify_operand_subreg. BEFORE returns
1461 the insn to be inserted before curr insn. AFTER returns the
1462 the insn to be inserted after curr insn. ORIGREG and NEWREG
1463 are the original reg and new reg for reload. */
1464 static void
1465 insert_move_for_subreg (rtx_insn **before, rtx_insn **after, rtx origreg,
1466 rtx newreg)
1468 if (before)
1470 push_to_sequence (*before);
1471 lra_emit_move (newreg, origreg);
1472 *before = get_insns ();
1473 end_sequence ();
1475 if (after)
1477 start_sequence ();
1478 lra_emit_move (origreg, newreg);
1479 emit_insn (*after);
1480 *after = get_insns ();
1481 end_sequence ();
1485 static int valid_address_p (machine_mode mode, rtx addr, addr_space_t as);
1486 static bool process_address (int, bool, rtx_insn **, rtx_insn **);
1488 /* Make reloads for subreg in operand NOP with internal subreg mode
1489 REG_MODE, add new reloads for further processing. Return true if
1490 any change was done. */
1491 static bool
1492 simplify_operand_subreg (int nop, machine_mode reg_mode)
1494 int hard_regno;
1495 rtx_insn *before, *after;
1496 machine_mode mode, innermode;
1497 rtx reg, new_reg;
1498 rtx operand = *curr_id->operand_loc[nop];
1499 enum reg_class regclass;
1500 enum op_type type;
1502 before = after = NULL;
1504 if (GET_CODE (operand) != SUBREG)
1505 return false;
1507 mode = GET_MODE (operand);
1508 reg = SUBREG_REG (operand);
1509 innermode = GET_MODE (reg);
1510 type = curr_static_id->operand[nop].type;
1511 if (MEM_P (reg))
1513 const bool addr_was_valid
1514 = valid_address_p (innermode, XEXP (reg, 0), MEM_ADDR_SPACE (reg));
1515 alter_subreg (curr_id->operand_loc[nop], false);
1516 rtx subst = *curr_id->operand_loc[nop];
1517 lra_assert (MEM_P (subst));
1519 if (!addr_was_valid
1520 || valid_address_p (GET_MODE (subst), XEXP (subst, 0),
1521 MEM_ADDR_SPACE (subst))
1522 || ((get_constraint_type (lookup_constraint
1523 (curr_static_id->operand[nop].constraint))
1524 != CT_SPECIAL_MEMORY)
1525 /* We still can reload address and if the address is
1526 valid, we can remove subreg without reloading its
1527 inner memory. */
1528 && valid_address_p (GET_MODE (subst),
1529 regno_reg_rtx
1530 [ira_class_hard_regs
1531 [base_reg_class (GET_MODE (subst),
1532 MEM_ADDR_SPACE (subst),
1533 ADDRESS, SCRATCH)][0]],
1534 MEM_ADDR_SPACE (subst))))
1536 /* If we change the address for a paradoxical subreg of memory, the
1537 new address might violate the necessary alignment or the access
1538 might be slow; take this into consideration. We need not worry
1539 about accesses beyond allocated memory for paradoxical memory
1540 subregs as we don't substitute such equiv memory (see processing
1541 equivalences in function lra_constraints) and because for spilled
1542 pseudos we allocate stack memory enough for the biggest
1543 corresponding paradoxical subreg.
1545 However, do not blindly simplify a (subreg (mem ...)) for
1546 WORD_REGISTER_OPERATIONS targets as this may lead to loading junk
1547 data into a register when the inner is narrower than outer or
1548 missing important data from memory when the inner is wider than
1549 outer. This rule only applies to modes that are no wider than
1550 a word. */
1551 if (!(GET_MODE_PRECISION (mode) != GET_MODE_PRECISION (innermode)
1552 && GET_MODE_SIZE (mode) <= UNITS_PER_WORD
1553 && GET_MODE_SIZE (innermode) <= UNITS_PER_WORD
1554 && WORD_REGISTER_OPERATIONS)
1555 && (!(MEM_ALIGN (subst) < GET_MODE_ALIGNMENT (mode)
1556 && targetm.slow_unaligned_access (mode, MEM_ALIGN (subst)))
1557 || (MEM_ALIGN (reg) < GET_MODE_ALIGNMENT (innermode)
1558 && targetm.slow_unaligned_access (innermode,
1559 MEM_ALIGN (reg)))))
1560 return true;
1562 *curr_id->operand_loc[nop] = operand;
1564 /* But if the address was not valid, we cannot reload the MEM without
1565 reloading the address first. */
1566 if (!addr_was_valid)
1567 process_address (nop, false, &before, &after);
1569 /* INNERMODE is fast, MODE slow. Reload the mem in INNERMODE. */
1570 enum reg_class rclass
1571 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1572 if (get_reload_reg (curr_static_id->operand[nop].type, innermode,
1573 reg, rclass, TRUE, "slow mem", &new_reg))
1575 bool insert_before, insert_after;
1576 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1578 insert_before = (type != OP_OUT
1579 || partial_subreg_p (mode, innermode));
1580 insert_after = type != OP_IN;
1581 insert_move_for_subreg (insert_before ? &before : NULL,
1582 insert_after ? &after : NULL,
1583 reg, new_reg);
1585 SUBREG_REG (operand) = new_reg;
1587 /* Convert to MODE. */
1588 reg = operand;
1589 rclass
1590 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1591 if (get_reload_reg (curr_static_id->operand[nop].type, mode, reg,
1592 rclass, TRUE, "slow mem", &new_reg))
1594 bool insert_before, insert_after;
1595 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1597 insert_before = type != OP_OUT;
1598 insert_after = type != OP_IN;
1599 insert_move_for_subreg (insert_before ? &before : NULL,
1600 insert_after ? &after : NULL,
1601 reg, new_reg);
1603 *curr_id->operand_loc[nop] = new_reg;
1604 lra_process_new_insns (curr_insn, before, after,
1605 "Inserting slow mem reload");
1606 return true;
1609 /* If the address was valid and became invalid, prefer to reload
1610 the memory. Typical case is when the index scale should
1611 correspond the memory. */
1612 *curr_id->operand_loc[nop] = operand;
1613 /* Do not return false here as the MEM_P (reg) will be processed
1614 later in this function. */
1616 else if (REG_P (reg) && REGNO (reg) < FIRST_PSEUDO_REGISTER)
1618 alter_subreg (curr_id->operand_loc[nop], false);
1619 return true;
1621 else if (CONSTANT_P (reg))
1623 /* Try to simplify subreg of constant. It is usually result of
1624 equivalence substitution. */
1625 if (innermode == VOIDmode
1626 && (innermode = original_subreg_reg_mode[nop]) == VOIDmode)
1627 innermode = curr_static_id->operand[nop].mode;
1628 if ((new_reg = simplify_subreg (mode, reg, innermode,
1629 SUBREG_BYTE (operand))) != NULL_RTX)
1631 *curr_id->operand_loc[nop] = new_reg;
1632 return true;
1635 /* Put constant into memory when we have mixed modes. It generates
1636 a better code in most cases as it does not need a secondary
1637 reload memory. It also prevents LRA looping when LRA is using
1638 secondary reload memory again and again. */
1639 if (CONSTANT_P (reg) && CONST_POOL_OK_P (reg_mode, reg)
1640 && SCALAR_INT_MODE_P (reg_mode) != SCALAR_INT_MODE_P (mode))
1642 SUBREG_REG (operand) = force_const_mem (reg_mode, reg);
1643 alter_subreg (curr_id->operand_loc[nop], false);
1644 return true;
1646 /* Force a reload of the SUBREG_REG if this is a constant or PLUS or
1647 if there may be a problem accessing OPERAND in the outer
1648 mode. */
1649 if ((REG_P (reg)
1650 && REGNO (reg) >= FIRST_PSEUDO_REGISTER
1651 && (hard_regno = lra_get_regno_hard_regno (REGNO (reg))) >= 0
1652 /* Don't reload paradoxical subregs because we could be looping
1653 having repeatedly final regno out of hard regs range. */
1654 && (hard_regno_nregs (hard_regno, innermode)
1655 >= hard_regno_nregs (hard_regno, mode))
1656 && simplify_subreg_regno (hard_regno, innermode,
1657 SUBREG_BYTE (operand), mode) < 0
1658 /* Don't reload subreg for matching reload. It is actually
1659 valid subreg in LRA. */
1660 && ! LRA_SUBREG_P (operand))
1661 || CONSTANT_P (reg) || GET_CODE (reg) == PLUS || MEM_P (reg))
1663 enum reg_class rclass;
1665 if (REG_P (reg))
1666 /* There is a big probability that we will get the same class
1667 for the new pseudo and we will get the same insn which
1668 means infinite looping. So spill the new pseudo. */
1669 rclass = NO_REGS;
1670 else
1671 /* The class will be defined later in curr_insn_transform. */
1672 rclass
1673 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1675 if (get_reload_reg (curr_static_id->operand[nop].type, reg_mode, reg,
1676 rclass, TRUE, "subreg reg", &new_reg))
1678 bool insert_before, insert_after;
1679 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1681 insert_before = (type != OP_OUT
1682 || GET_MODE_SIZE (innermode) > GET_MODE_SIZE (mode));
1683 insert_after = (type != OP_IN);
1684 insert_move_for_subreg (insert_before ? &before : NULL,
1685 insert_after ? &after : NULL,
1686 reg, new_reg);
1688 SUBREG_REG (operand) = new_reg;
1689 lra_process_new_insns (curr_insn, before, after,
1690 "Inserting subreg reload");
1691 return true;
1693 /* Force a reload for a paradoxical subreg. For paradoxical subreg,
1694 IRA allocates hardreg to the inner pseudo reg according to its mode
1695 instead of the outermode, so the size of the hardreg may not be enough
1696 to contain the outermode operand, in that case we may need to insert
1697 reload for the reg. For the following two types of paradoxical subreg,
1698 we need to insert reload:
1699 1. If the op_type is OP_IN, and the hardreg could not be paired with
1700 other hardreg to contain the outermode operand
1701 (checked by in_hard_reg_set_p), we need to insert the reload.
1702 2. If the op_type is OP_OUT or OP_INOUT.
1704 Here is a paradoxical subreg example showing how the reload is generated:
1706 (insn 5 4 7 2 (set (reg:TI 106 [ __comp ])
1707 (subreg:TI (reg:DI 107 [ __comp ]) 0)) {*movti_internal_rex64}
1709 In IRA, reg107 is allocated to a DImode hardreg. We use x86-64 as example
1710 here, if reg107 is assigned to hardreg R15, because R15 is the last
1711 hardreg, compiler cannot find another hardreg to pair with R15 to
1712 contain TImode data. So we insert a TImode reload reg180 for it.
1713 After reload is inserted:
1715 (insn 283 0 0 (set (subreg:DI (reg:TI 180 [orig:107 __comp ] [107]) 0)
1716 (reg:DI 107 [ __comp ])) -1
1717 (insn 5 4 7 2 (set (reg:TI 106 [ __comp ])
1718 (subreg:TI (reg:TI 180 [orig:107 __comp ] [107]) 0)) {*movti_internal_rex64}
1720 Two reload hard registers will be allocated to reg180 to save TImode data
1721 in LRA_assign. */
1722 else if (REG_P (reg)
1723 && REGNO (reg) >= FIRST_PSEUDO_REGISTER
1724 && (hard_regno = lra_get_regno_hard_regno (REGNO (reg))) >= 0
1725 && (hard_regno_nregs (hard_regno, innermode)
1726 < hard_regno_nregs (hard_regno, mode))
1727 && (regclass = lra_get_allocno_class (REGNO (reg)))
1728 && (type != OP_IN
1729 || !in_hard_reg_set_p (reg_class_contents[regclass],
1730 mode, hard_regno)))
1732 /* The class will be defined later in curr_insn_transform. */
1733 enum reg_class rclass
1734 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1736 if (get_reload_reg (curr_static_id->operand[nop].type, mode, reg,
1737 rclass, TRUE, "paradoxical subreg", &new_reg))
1739 rtx subreg;
1740 bool insert_before, insert_after;
1742 PUT_MODE (new_reg, mode);
1743 subreg = gen_lowpart_SUBREG (innermode, new_reg);
1744 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1746 insert_before = (type != OP_OUT);
1747 insert_after = (type != OP_IN);
1748 insert_move_for_subreg (insert_before ? &before : NULL,
1749 insert_after ? &after : NULL,
1750 reg, subreg);
1752 SUBREG_REG (operand) = new_reg;
1753 lra_process_new_insns (curr_insn, before, after,
1754 "Inserting paradoxical subreg reload");
1755 return true;
1757 return false;
1760 /* Return TRUE if X refers for a hard register from SET. */
1761 static bool
1762 uses_hard_regs_p (rtx x, HARD_REG_SET set)
1764 int i, j, x_hard_regno;
1765 machine_mode mode;
1766 const char *fmt;
1767 enum rtx_code code;
1769 if (x == NULL_RTX)
1770 return false;
1771 code = GET_CODE (x);
1772 mode = GET_MODE (x);
1773 if (code == SUBREG)
1775 x = SUBREG_REG (x);
1776 code = GET_CODE (x);
1777 if (GET_MODE_SIZE (GET_MODE (x)) > GET_MODE_SIZE (mode))
1778 mode = GET_MODE (x);
1781 if (REG_P (x))
1783 x_hard_regno = get_hard_regno (x, true);
1784 return (x_hard_regno >= 0
1785 && overlaps_hard_reg_set_p (set, mode, x_hard_regno));
1787 if (MEM_P (x))
1789 struct address_info ad;
1791 decompose_mem_address (&ad, x);
1792 if (ad.base_term != NULL && uses_hard_regs_p (*ad.base_term, set))
1793 return true;
1794 if (ad.index_term != NULL && uses_hard_regs_p (*ad.index_term, set))
1795 return true;
1797 fmt = GET_RTX_FORMAT (code);
1798 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1800 if (fmt[i] == 'e')
1802 if (uses_hard_regs_p (XEXP (x, i), set))
1803 return true;
1805 else if (fmt[i] == 'E')
1807 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
1808 if (uses_hard_regs_p (XVECEXP (x, i, j), set))
1809 return true;
1812 return false;
1815 /* Return true if OP is a spilled pseudo. */
1816 static inline bool
1817 spilled_pseudo_p (rtx op)
1819 return (REG_P (op)
1820 && REGNO (op) >= FIRST_PSEUDO_REGISTER && in_mem_p (REGNO (op)));
1823 /* Return true if X is a general constant. */
1824 static inline bool
1825 general_constant_p (rtx x)
1827 return CONSTANT_P (x) && (! flag_pic || LEGITIMATE_PIC_OPERAND_P (x));
1830 static bool
1831 reg_in_class_p (rtx reg, enum reg_class cl)
1833 if (cl == NO_REGS)
1834 return get_reg_class (REGNO (reg)) == NO_REGS;
1835 return in_class_p (reg, cl, NULL);
1838 /* Return true if SET of RCLASS contains no hard regs which can be
1839 used in MODE. */
1840 static bool
1841 prohibited_class_reg_set_mode_p (enum reg_class rclass,
1842 HARD_REG_SET &set,
1843 machine_mode mode)
1845 HARD_REG_SET temp;
1847 lra_assert (hard_reg_set_subset_p (reg_class_contents[rclass], set));
1848 COPY_HARD_REG_SET (temp, set);
1849 AND_COMPL_HARD_REG_SET (temp, lra_no_alloc_regs);
1850 return (hard_reg_set_subset_p
1851 (temp, ira_prohibited_class_mode_regs[rclass][mode]));
1855 /* Used to check validity info about small class input operands. It
1856 should be incremented at start of processing an insn
1857 alternative. */
1858 static unsigned int curr_small_class_check = 0;
1860 /* Update number of used inputs of class OP_CLASS for operand NOP.
1861 Return true if we have more such class operands than the number of
1862 available regs. */
1863 static bool
1864 update_and_check_small_class_inputs (int nop, enum reg_class op_class)
1866 static unsigned int small_class_check[LIM_REG_CLASSES];
1867 static int small_class_input_nums[LIM_REG_CLASSES];
1869 if (SMALL_REGISTER_CLASS_P (op_class)
1870 /* We are interesting in classes became small because of fixing
1871 some hard regs, e.g. by an user through GCC options. */
1872 && hard_reg_set_intersect_p (reg_class_contents[op_class],
1873 ira_no_alloc_regs)
1874 && (curr_static_id->operand[nop].type != OP_OUT
1875 || curr_static_id->operand[nop].early_clobber))
1877 if (small_class_check[op_class] == curr_small_class_check)
1878 small_class_input_nums[op_class]++;
1879 else
1881 small_class_check[op_class] = curr_small_class_check;
1882 small_class_input_nums[op_class] = 1;
1884 if (small_class_input_nums[op_class] > ira_class_hard_regs_num[op_class])
1885 return true;
1887 return false;
1890 /* Major function to choose the current insn alternative and what
1891 operands should be reloaded and how. If ONLY_ALTERNATIVE is not
1892 negative we should consider only this alternative. Return false if
1893 we can not choose the alternative or find how to reload the
1894 operands. */
1895 static bool
1896 process_alt_operands (int only_alternative)
1898 bool ok_p = false;
1899 int nop, overall, nalt;
1900 int n_alternatives = curr_static_id->n_alternatives;
1901 int n_operands = curr_static_id->n_operands;
1902 /* LOSERS counts the operands that don't fit this alternative and
1903 would require loading. */
1904 int losers;
1905 int addr_losers;
1906 /* REJECT is a count of how undesirable this alternative says it is
1907 if any reloading is required. If the alternative matches exactly
1908 then REJECT is ignored, but otherwise it gets this much counted
1909 against it in addition to the reloading needed. */
1910 int reject;
1911 /* This is defined by '!' or '?' alternative constraint and added to
1912 reject. But in some cases it can be ignored. */
1913 int static_reject;
1914 int op_reject;
1915 /* The number of elements in the following array. */
1916 int early_clobbered_regs_num;
1917 /* Numbers of operands which are early clobber registers. */
1918 int early_clobbered_nops[MAX_RECOG_OPERANDS];
1919 enum reg_class curr_alt[MAX_RECOG_OPERANDS];
1920 HARD_REG_SET curr_alt_set[MAX_RECOG_OPERANDS];
1921 bool curr_alt_match_win[MAX_RECOG_OPERANDS];
1922 bool curr_alt_win[MAX_RECOG_OPERANDS];
1923 bool curr_alt_offmemok[MAX_RECOG_OPERANDS];
1924 int curr_alt_matches[MAX_RECOG_OPERANDS];
1925 /* The number of elements in the following array. */
1926 int curr_alt_dont_inherit_ops_num;
1927 /* Numbers of operands whose reload pseudos should not be inherited. */
1928 int curr_alt_dont_inherit_ops[MAX_RECOG_OPERANDS];
1929 rtx op;
1930 /* The register when the operand is a subreg of register, otherwise the
1931 operand itself. */
1932 rtx no_subreg_reg_operand[MAX_RECOG_OPERANDS];
1933 /* The register if the operand is a register or subreg of register,
1934 otherwise NULL. */
1935 rtx operand_reg[MAX_RECOG_OPERANDS];
1936 int hard_regno[MAX_RECOG_OPERANDS];
1937 machine_mode biggest_mode[MAX_RECOG_OPERANDS];
1938 int reload_nregs, reload_sum;
1939 bool costly_p;
1940 enum reg_class cl;
1942 /* Calculate some data common for all alternatives to speed up the
1943 function. */
1944 for (nop = 0; nop < n_operands; nop++)
1946 rtx reg;
1948 op = no_subreg_reg_operand[nop] = *curr_id->operand_loc[nop];
1949 /* The real hard regno of the operand after the allocation. */
1950 hard_regno[nop] = get_hard_regno (op, true);
1952 operand_reg[nop] = reg = op;
1953 biggest_mode[nop] = GET_MODE (op);
1954 if (GET_CODE (op) == SUBREG)
1956 operand_reg[nop] = reg = SUBREG_REG (op);
1957 if (GET_MODE_SIZE (biggest_mode[nop])
1958 < GET_MODE_SIZE (GET_MODE (reg)))
1959 biggest_mode[nop] = GET_MODE (reg);
1961 if (! REG_P (reg))
1962 operand_reg[nop] = NULL_RTX;
1963 else if (REGNO (reg) >= FIRST_PSEUDO_REGISTER
1964 || ((int) REGNO (reg)
1965 == lra_get_elimination_hard_regno (REGNO (reg))))
1966 no_subreg_reg_operand[nop] = reg;
1967 else
1968 operand_reg[nop] = no_subreg_reg_operand[nop]
1969 /* Just use natural mode for elimination result. It should
1970 be enough for extra constraints hooks. */
1971 = regno_reg_rtx[hard_regno[nop]];
1974 /* The constraints are made of several alternatives. Each operand's
1975 constraint looks like foo,bar,... with commas separating the
1976 alternatives. The first alternatives for all operands go
1977 together, the second alternatives go together, etc.
1979 First loop over alternatives. */
1980 alternative_mask preferred = curr_id->preferred_alternatives;
1981 if (only_alternative >= 0)
1982 preferred &= ALTERNATIVE_BIT (only_alternative);
1984 for (nalt = 0; nalt < n_alternatives; nalt++)
1986 /* Loop over operands for one constraint alternative. */
1987 if (!TEST_BIT (preferred, nalt))
1988 continue;
1990 curr_small_class_check++;
1991 overall = losers = addr_losers = 0;
1992 static_reject = reject = reload_nregs = reload_sum = 0;
1993 for (nop = 0; nop < n_operands; nop++)
1995 int inc = (curr_static_id
1996 ->operand_alternative[nalt * n_operands + nop].reject);
1997 if (lra_dump_file != NULL && inc != 0)
1998 fprintf (lra_dump_file,
1999 " Staticly defined alt reject+=%d\n", inc);
2000 static_reject += inc;
2002 reject += static_reject;
2003 early_clobbered_regs_num = 0;
2005 for (nop = 0; nop < n_operands; nop++)
2007 const char *p;
2008 char *end;
2009 int len, c, m, i, opalt_num, this_alternative_matches;
2010 bool win, did_match, offmemok, early_clobber_p;
2011 /* false => this operand can be reloaded somehow for this
2012 alternative. */
2013 bool badop;
2014 /* true => this operand can be reloaded if the alternative
2015 allows regs. */
2016 bool winreg;
2017 /* True if a constant forced into memory would be OK for
2018 this operand. */
2019 bool constmemok;
2020 enum reg_class this_alternative, this_costly_alternative;
2021 HARD_REG_SET this_alternative_set, this_costly_alternative_set;
2022 bool this_alternative_match_win, this_alternative_win;
2023 bool this_alternative_offmemok;
2024 bool scratch_p;
2025 machine_mode mode;
2026 enum constraint_num cn;
2028 opalt_num = nalt * n_operands + nop;
2029 if (curr_static_id->operand_alternative[opalt_num].anything_ok)
2031 /* Fast track for no constraints at all. */
2032 curr_alt[nop] = NO_REGS;
2033 CLEAR_HARD_REG_SET (curr_alt_set[nop]);
2034 curr_alt_win[nop] = true;
2035 curr_alt_match_win[nop] = false;
2036 curr_alt_offmemok[nop] = false;
2037 curr_alt_matches[nop] = -1;
2038 continue;
2041 op = no_subreg_reg_operand[nop];
2042 mode = curr_operand_mode[nop];
2044 win = did_match = winreg = offmemok = constmemok = false;
2045 badop = true;
2047 early_clobber_p = false;
2048 p = curr_static_id->operand_alternative[opalt_num].constraint;
2050 this_costly_alternative = this_alternative = NO_REGS;
2051 /* We update set of possible hard regs besides its class
2052 because reg class might be inaccurate. For example,
2053 union of LO_REGS (l), HI_REGS(h), and STACK_REG(k) in ARM
2054 is translated in HI_REGS because classes are merged by
2055 pairs and there is no accurate intermediate class. */
2056 CLEAR_HARD_REG_SET (this_alternative_set);
2057 CLEAR_HARD_REG_SET (this_costly_alternative_set);
2058 this_alternative_win = false;
2059 this_alternative_match_win = false;
2060 this_alternative_offmemok = false;
2061 this_alternative_matches = -1;
2063 /* An empty constraint should be excluded by the fast
2064 track. */
2065 lra_assert (*p != 0 && *p != ',');
2067 op_reject = 0;
2068 /* Scan this alternative's specs for this operand; set WIN
2069 if the operand fits any letter in this alternative.
2070 Otherwise, clear BADOP if this operand could fit some
2071 letter after reloads, or set WINREG if this operand could
2072 fit after reloads provided the constraint allows some
2073 registers. */
2074 costly_p = false;
2077 switch ((c = *p, len = CONSTRAINT_LEN (c, p)), c)
2079 case '\0':
2080 len = 0;
2081 break;
2082 case ',':
2083 c = '\0';
2084 break;
2086 case '&':
2087 early_clobber_p = true;
2088 break;
2090 case '$':
2091 op_reject += LRA_MAX_REJECT;
2092 break;
2093 case '^':
2094 op_reject += LRA_LOSER_COST_FACTOR;
2095 break;
2097 case '#':
2098 /* Ignore rest of this alternative. */
2099 c = '\0';
2100 break;
2102 case '0': case '1': case '2': case '3': case '4':
2103 case '5': case '6': case '7': case '8': case '9':
2105 int m_hregno;
2106 bool match_p;
2108 m = strtoul (p, &end, 10);
2109 p = end;
2110 len = 0;
2111 lra_assert (nop > m);
2113 this_alternative_matches = m;
2114 m_hregno = get_hard_regno (*curr_id->operand_loc[m], false);
2115 /* We are supposed to match a previous operand.
2116 If we do, we win if that one did. If we do
2117 not, count both of the operands as losers.
2118 (This is too conservative, since most of the
2119 time only a single reload insn will be needed
2120 to make the two operands win. As a result,
2121 this alternative may be rejected when it is
2122 actually desirable.) */
2123 match_p = false;
2124 if (operands_match_p (*curr_id->operand_loc[nop],
2125 *curr_id->operand_loc[m], m_hregno))
2127 /* We should reject matching of an early
2128 clobber operand if the matching operand is
2129 not dying in the insn. */
2130 if (! curr_static_id->operand[m].early_clobber
2131 || operand_reg[nop] == NULL_RTX
2132 || (find_regno_note (curr_insn, REG_DEAD,
2133 REGNO (op))
2134 || REGNO (op) == REGNO (operand_reg[m])))
2135 match_p = true;
2137 if (match_p)
2139 /* If we are matching a non-offsettable
2140 address where an offsettable address was
2141 expected, then we must reject this
2142 combination, because we can't reload
2143 it. */
2144 if (curr_alt_offmemok[m]
2145 && MEM_P (*curr_id->operand_loc[m])
2146 && curr_alt[m] == NO_REGS && ! curr_alt_win[m])
2147 continue;
2149 else
2151 /* Operands don't match. Both operands must
2152 allow a reload register, otherwise we
2153 cannot make them match. */
2154 if (curr_alt[m] == NO_REGS)
2155 break;
2156 /* Retroactively mark the operand we had to
2157 match as a loser, if it wasn't already and
2158 it wasn't matched to a register constraint
2159 (e.g it might be matched by memory). */
2160 if (curr_alt_win[m]
2161 && (operand_reg[m] == NULL_RTX
2162 || hard_regno[m] < 0))
2164 losers++;
2165 reload_nregs
2166 += (ira_reg_class_max_nregs[curr_alt[m]]
2167 [GET_MODE (*curr_id->operand_loc[m])]);
2170 /* Prefer matching earlyclobber alternative as
2171 it results in less hard regs required for
2172 the insn than a non-matching earlyclobber
2173 alternative. */
2174 if (curr_static_id->operand[m].early_clobber)
2176 if (lra_dump_file != NULL)
2177 fprintf
2178 (lra_dump_file,
2179 " %d Matching earlyclobber alt:"
2180 " reject--\n",
2181 nop);
2182 reject--;
2184 /* Otherwise we prefer no matching
2185 alternatives because it gives more freedom
2186 in RA. */
2187 else if (operand_reg[nop] == NULL_RTX
2188 || (find_regno_note (curr_insn, REG_DEAD,
2189 REGNO (operand_reg[nop]))
2190 == NULL_RTX))
2192 if (lra_dump_file != NULL)
2193 fprintf
2194 (lra_dump_file,
2195 " %d Matching alt: reject+=2\n",
2196 nop);
2197 reject += 2;
2200 /* If we have to reload this operand and some
2201 previous operand also had to match the same
2202 thing as this operand, we don't know how to do
2203 that. */
2204 if (!match_p || !curr_alt_win[m])
2206 for (i = 0; i < nop; i++)
2207 if (curr_alt_matches[i] == m)
2208 break;
2209 if (i < nop)
2210 break;
2212 else
2213 did_match = true;
2215 /* This can be fixed with reloads if the operand
2216 we are supposed to match can be fixed with
2217 reloads. */
2218 badop = false;
2219 this_alternative = curr_alt[m];
2220 COPY_HARD_REG_SET (this_alternative_set, curr_alt_set[m]);
2221 winreg = this_alternative != NO_REGS;
2222 break;
2225 case 'g':
2226 if (MEM_P (op)
2227 || general_constant_p (op)
2228 || spilled_pseudo_p (op))
2229 win = true;
2230 cl = GENERAL_REGS;
2231 goto reg;
2233 default:
2234 cn = lookup_constraint (p);
2235 switch (get_constraint_type (cn))
2237 case CT_REGISTER:
2238 cl = reg_class_for_constraint (cn);
2239 if (cl != NO_REGS)
2240 goto reg;
2241 break;
2243 case CT_CONST_INT:
2244 if (CONST_INT_P (op)
2245 && insn_const_int_ok_for_constraint (INTVAL (op), cn))
2246 win = true;
2247 break;
2249 case CT_MEMORY:
2250 if (MEM_P (op)
2251 && satisfies_memory_constraint_p (op, cn))
2252 win = true;
2253 else if (spilled_pseudo_p (op))
2254 win = true;
2256 /* If we didn't already win, we can reload constants
2257 via force_const_mem or put the pseudo value into
2258 memory, or make other memory by reloading the
2259 address like for 'o'. */
2260 if (CONST_POOL_OK_P (mode, op)
2261 || MEM_P (op) || REG_P (op)
2262 /* We can restore the equiv insn by a
2263 reload. */
2264 || equiv_substition_p[nop])
2265 badop = false;
2266 constmemok = true;
2267 offmemok = true;
2268 break;
2270 case CT_ADDRESS:
2271 /* If we didn't already win, we can reload the address
2272 into a base register. */
2273 if (satisfies_address_constraint_p (op, cn))
2274 win = true;
2275 cl = base_reg_class (VOIDmode, ADDR_SPACE_GENERIC,
2276 ADDRESS, SCRATCH);
2277 badop = false;
2278 goto reg;
2280 case CT_FIXED_FORM:
2281 if (constraint_satisfied_p (op, cn))
2282 win = true;
2283 break;
2285 case CT_SPECIAL_MEMORY:
2286 if (MEM_P (op)
2287 && satisfies_memory_constraint_p (op, cn))
2288 win = true;
2289 else if (spilled_pseudo_p (op))
2290 win = true;
2291 break;
2293 break;
2295 reg:
2296 this_alternative = reg_class_subunion[this_alternative][cl];
2297 IOR_HARD_REG_SET (this_alternative_set,
2298 reg_class_contents[cl]);
2299 if (costly_p)
2301 this_costly_alternative
2302 = reg_class_subunion[this_costly_alternative][cl];
2303 IOR_HARD_REG_SET (this_costly_alternative_set,
2304 reg_class_contents[cl]);
2306 if (mode == BLKmode)
2307 break;
2308 winreg = true;
2309 if (REG_P (op))
2311 if (hard_regno[nop] >= 0
2312 && in_hard_reg_set_p (this_alternative_set,
2313 mode, hard_regno[nop]))
2314 win = true;
2315 else if (hard_regno[nop] < 0
2316 && in_class_p (op, this_alternative, NULL))
2317 win = true;
2319 break;
2321 if (c != ' ' && c != '\t')
2322 costly_p = c == '*';
2324 while ((p += len), c);
2326 scratch_p = (operand_reg[nop] != NULL_RTX
2327 && lra_former_scratch_p (REGNO (operand_reg[nop])));
2328 /* Record which operands fit this alternative. */
2329 if (win)
2331 this_alternative_win = true;
2332 if (operand_reg[nop] != NULL_RTX)
2334 if (hard_regno[nop] >= 0)
2336 if (in_hard_reg_set_p (this_costly_alternative_set,
2337 mode, hard_regno[nop]))
2339 if (lra_dump_file != NULL)
2340 fprintf (lra_dump_file,
2341 " %d Costly set: reject++\n",
2342 nop);
2343 reject++;
2346 else
2348 /* Prefer won reg to spilled pseudo under other
2349 equal conditions for possibe inheritance. */
2350 if (! scratch_p)
2352 if (lra_dump_file != NULL)
2353 fprintf
2354 (lra_dump_file,
2355 " %d Non pseudo reload: reject++\n",
2356 nop);
2357 reject++;
2359 if (in_class_p (operand_reg[nop],
2360 this_costly_alternative, NULL))
2362 if (lra_dump_file != NULL)
2363 fprintf
2364 (lra_dump_file,
2365 " %d Non pseudo costly reload:"
2366 " reject++\n",
2367 nop);
2368 reject++;
2371 /* We simulate the behavior of old reload here.
2372 Although scratches need hard registers and it
2373 might result in spilling other pseudos, no reload
2374 insns are generated for the scratches. So it
2375 might cost something but probably less than old
2376 reload pass believes. */
2377 if (scratch_p)
2379 if (lra_dump_file != NULL)
2380 fprintf (lra_dump_file,
2381 " %d Scratch win: reject+=2\n",
2382 nop);
2383 reject += 2;
2387 else if (did_match)
2388 this_alternative_match_win = true;
2389 else
2391 int const_to_mem = 0;
2392 bool no_regs_p;
2394 reject += op_reject;
2395 /* Never do output reload of stack pointer. It makes
2396 impossible to do elimination when SP is changed in
2397 RTL. */
2398 if (op == stack_pointer_rtx && ! frame_pointer_needed
2399 && curr_static_id->operand[nop].type != OP_IN)
2400 goto fail;
2402 /* If this alternative asks for a specific reg class, see if there
2403 is at least one allocatable register in that class. */
2404 no_regs_p
2405 = (this_alternative == NO_REGS
2406 || (hard_reg_set_subset_p
2407 (reg_class_contents[this_alternative],
2408 lra_no_alloc_regs)));
2410 /* For asms, verify that the class for this alternative is possible
2411 for the mode that is specified. */
2412 if (!no_regs_p && INSN_CODE (curr_insn) < 0)
2414 int i;
2415 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
2416 if (targetm.hard_regno_mode_ok (i, mode)
2417 && in_hard_reg_set_p (reg_class_contents[this_alternative],
2418 mode, i))
2419 break;
2420 if (i == FIRST_PSEUDO_REGISTER)
2421 winreg = false;
2424 /* If this operand accepts a register, and if the
2425 register class has at least one allocatable register,
2426 then this operand can be reloaded. */
2427 if (winreg && !no_regs_p)
2428 badop = false;
2430 if (badop)
2432 if (lra_dump_file != NULL)
2433 fprintf (lra_dump_file,
2434 " alt=%d: Bad operand -- refuse\n",
2435 nalt);
2436 goto fail;
2439 if (this_alternative != NO_REGS)
2441 HARD_REG_SET available_regs;
2443 COPY_HARD_REG_SET (available_regs,
2444 reg_class_contents[this_alternative]);
2445 AND_COMPL_HARD_REG_SET
2446 (available_regs,
2447 ira_prohibited_class_mode_regs[this_alternative][mode]);
2448 AND_COMPL_HARD_REG_SET (available_regs, lra_no_alloc_regs);
2449 if (hard_reg_set_empty_p (available_regs))
2451 /* There are no hard regs holding a value of given
2452 mode. */
2453 if (offmemok)
2455 this_alternative = NO_REGS;
2456 if (lra_dump_file != NULL)
2457 fprintf (lra_dump_file,
2458 " %d Using memory because of"
2459 " a bad mode: reject+=2\n",
2460 nop);
2461 reject += 2;
2463 else
2465 if (lra_dump_file != NULL)
2466 fprintf (lra_dump_file,
2467 " alt=%d: Wrong mode -- refuse\n",
2468 nalt);
2469 goto fail;
2474 /* If not assigned pseudo has a class which a subset of
2475 required reg class, it is a less costly alternative
2476 as the pseudo still can get a hard reg of necessary
2477 class. */
2478 if (! no_regs_p && REG_P (op) && hard_regno[nop] < 0
2479 && (cl = get_reg_class (REGNO (op))) != NO_REGS
2480 && ira_class_subset_p[this_alternative][cl])
2482 if (lra_dump_file != NULL)
2483 fprintf
2484 (lra_dump_file,
2485 " %d Super set class reg: reject-=3\n", nop);
2486 reject -= 3;
2489 this_alternative_offmemok = offmemok;
2490 if (this_costly_alternative != NO_REGS)
2492 if (lra_dump_file != NULL)
2493 fprintf (lra_dump_file,
2494 " %d Costly loser: reject++\n", nop);
2495 reject++;
2497 /* If the operand is dying, has a matching constraint,
2498 and satisfies constraints of the matched operand
2499 which failed to satisfy the own constraints, most probably
2500 the reload for this operand will be gone. */
2501 if (this_alternative_matches >= 0
2502 && !curr_alt_win[this_alternative_matches]
2503 && REG_P (op)
2504 && find_regno_note (curr_insn, REG_DEAD, REGNO (op))
2505 && (hard_regno[nop] >= 0
2506 ? in_hard_reg_set_p (this_alternative_set,
2507 mode, hard_regno[nop])
2508 : in_class_p (op, this_alternative, NULL)))
2510 if (lra_dump_file != NULL)
2511 fprintf
2512 (lra_dump_file,
2513 " %d Dying matched operand reload: reject++\n",
2514 nop);
2515 reject++;
2517 else
2519 /* Strict_low_part requires to reload the register
2520 not the sub-register. In this case we should
2521 check that a final reload hard reg can hold the
2522 value mode. */
2523 if (curr_static_id->operand[nop].strict_low
2524 && REG_P (op)
2525 && hard_regno[nop] < 0
2526 && GET_CODE (*curr_id->operand_loc[nop]) == SUBREG
2527 && ira_class_hard_regs_num[this_alternative] > 0
2528 && (!targetm.hard_regno_mode_ok
2529 (ira_class_hard_regs[this_alternative][0],
2530 GET_MODE (*curr_id->operand_loc[nop]))))
2532 if (lra_dump_file != NULL)
2533 fprintf
2534 (lra_dump_file,
2535 " alt=%d: Strict low subreg reload -- refuse\n",
2536 nalt);
2537 goto fail;
2539 losers++;
2541 if (operand_reg[nop] != NULL_RTX
2542 /* Output operands and matched input operands are
2543 not inherited. The following conditions do not
2544 exactly describe the previous statement but they
2545 are pretty close. */
2546 && curr_static_id->operand[nop].type != OP_OUT
2547 && (this_alternative_matches < 0
2548 || curr_static_id->operand[nop].type != OP_IN))
2550 int last_reload = (lra_reg_info[ORIGINAL_REGNO
2551 (operand_reg[nop])]
2552 .last_reload);
2554 /* The value of reload_sum has sense only if we
2555 process insns in their order. It happens only on
2556 the first constraints sub-pass when we do most of
2557 reload work. */
2558 if (lra_constraint_iter == 1 && last_reload > bb_reload_num)
2559 reload_sum += last_reload - bb_reload_num;
2561 /* If this is a constant that is reloaded into the
2562 desired class by copying it to memory first, count
2563 that as another reload. This is consistent with
2564 other code and is required to avoid choosing another
2565 alternative when the constant is moved into memory.
2566 Note that the test here is precisely the same as in
2567 the code below that calls force_const_mem. */
2568 if (CONST_POOL_OK_P (mode, op)
2569 && ((targetm.preferred_reload_class
2570 (op, this_alternative) == NO_REGS)
2571 || no_input_reloads_p))
2573 const_to_mem = 1;
2574 if (! no_regs_p)
2575 losers++;
2578 /* Alternative loses if it requires a type of reload not
2579 permitted for this insn. We can always reload
2580 objects with a REG_UNUSED note. */
2581 if ((curr_static_id->operand[nop].type != OP_IN
2582 && no_output_reloads_p
2583 && ! find_reg_note (curr_insn, REG_UNUSED, op))
2584 || (curr_static_id->operand[nop].type != OP_OUT
2585 && no_input_reloads_p && ! const_to_mem)
2586 || (this_alternative_matches >= 0
2587 && (no_input_reloads_p
2588 || (no_output_reloads_p
2589 && (curr_static_id->operand
2590 [this_alternative_matches].type != OP_IN)
2591 && ! find_reg_note (curr_insn, REG_UNUSED,
2592 no_subreg_reg_operand
2593 [this_alternative_matches])))))
2595 if (lra_dump_file != NULL)
2596 fprintf
2597 (lra_dump_file,
2598 " alt=%d: No input/otput reload -- refuse\n",
2599 nalt);
2600 goto fail;
2603 /* Alternative loses if it required class pseudo can not
2604 hold value of required mode. Such insns can be
2605 described by insn definitions with mode iterators. */
2606 if (GET_MODE (*curr_id->operand_loc[nop]) != VOIDmode
2607 && ! hard_reg_set_empty_p (this_alternative_set)
2608 /* It is common practice for constraints to use a
2609 class which does not have actually enough regs to
2610 hold the value (e.g. x86 AREG for mode requiring
2611 more one general reg). Therefore we have 2
2612 conditions to check that the reload pseudo can
2613 not hold the mode value. */
2614 && (!targetm.hard_regno_mode_ok
2615 (ira_class_hard_regs[this_alternative][0],
2616 GET_MODE (*curr_id->operand_loc[nop])))
2617 /* The above condition is not enough as the first
2618 reg in ira_class_hard_regs can be not aligned for
2619 multi-words mode values. */
2620 && (prohibited_class_reg_set_mode_p
2621 (this_alternative, this_alternative_set,
2622 GET_MODE (*curr_id->operand_loc[nop]))))
2624 if (lra_dump_file != NULL)
2625 fprintf (lra_dump_file,
2626 " alt=%d: reload pseudo for op %d "
2627 " can not hold the mode value -- refuse\n",
2628 nalt, nop);
2629 goto fail;
2632 /* Check strong discouragement of reload of non-constant
2633 into class THIS_ALTERNATIVE. */
2634 if (! CONSTANT_P (op) && ! no_regs_p
2635 && (targetm.preferred_reload_class
2636 (op, this_alternative) == NO_REGS
2637 || (curr_static_id->operand[nop].type == OP_OUT
2638 && (targetm.preferred_output_reload_class
2639 (op, this_alternative) == NO_REGS))))
2641 if (lra_dump_file != NULL)
2642 fprintf (lra_dump_file,
2643 " %d Non-prefered reload: reject+=%d\n",
2644 nop, LRA_MAX_REJECT);
2645 reject += LRA_MAX_REJECT;
2648 if (! (MEM_P (op) && offmemok)
2649 && ! (const_to_mem && constmemok))
2651 /* We prefer to reload pseudos over reloading other
2652 things, since such reloads may be able to be
2653 eliminated later. So bump REJECT in other cases.
2654 Don't do this in the case where we are forcing a
2655 constant into memory and it will then win since
2656 we don't want to have a different alternative
2657 match then. */
2658 if (! (REG_P (op) && REGNO (op) >= FIRST_PSEUDO_REGISTER))
2660 if (lra_dump_file != NULL)
2661 fprintf
2662 (lra_dump_file,
2663 " %d Non-pseudo reload: reject+=2\n",
2664 nop);
2665 reject += 2;
2668 if (! no_regs_p)
2669 reload_nregs
2670 += ira_reg_class_max_nregs[this_alternative][mode];
2672 if (SMALL_REGISTER_CLASS_P (this_alternative))
2674 if (lra_dump_file != NULL)
2675 fprintf
2676 (lra_dump_file,
2677 " %d Small class reload: reject+=%d\n",
2678 nop, LRA_LOSER_COST_FACTOR / 2);
2679 reject += LRA_LOSER_COST_FACTOR / 2;
2683 /* We are trying to spill pseudo into memory. It is
2684 usually more costly than moving to a hard register
2685 although it might takes the same number of
2686 reloads.
2688 Non-pseudo spill may happen also. Suppose a target allows both
2689 register and memory in the operand constraint alternatives,
2690 then it's typical that an eliminable register has a substition
2691 of "base + offset" which can either be reloaded by a simple
2692 "new_reg <= base + offset" which will match the register
2693 constraint, or a similar reg addition followed by further spill
2694 to and reload from memory which will match the memory
2695 constraint, but this memory spill will be much more costly
2696 usually.
2698 Code below increases the reject for both pseudo and non-pseudo
2699 spill. */
2700 if (no_regs_p
2701 && !(MEM_P (op) && offmemok)
2702 && !(REG_P (op) && hard_regno[nop] < 0))
2704 if (lra_dump_file != NULL)
2705 fprintf
2706 (lra_dump_file,
2707 " %d Spill %spseudo into memory: reject+=3\n",
2708 nop, REG_P (op) ? "" : "Non-");
2709 reject += 3;
2710 if (VECTOR_MODE_P (mode))
2712 /* Spilling vectors into memory is usually more
2713 costly as they contain big values. */
2714 if (lra_dump_file != NULL)
2715 fprintf
2716 (lra_dump_file,
2717 " %d Spill vector pseudo: reject+=2\n",
2718 nop);
2719 reject += 2;
2723 /* When we use an operand requiring memory in given
2724 alternative, the insn should write *and* read the
2725 value to/from memory it is costly in comparison with
2726 an insn alternative which does not use memory
2727 (e.g. register or immediate operand). We exclude
2728 memory operand for such case as we can satisfy the
2729 memory constraints by reloading address. */
2730 if (no_regs_p && offmemok && !MEM_P (op))
2732 if (lra_dump_file != NULL)
2733 fprintf
2734 (lra_dump_file,
2735 " Using memory insn operand %d: reject+=3\n",
2736 nop);
2737 reject += 3;
2740 /* If reload requires moving value through secondary
2741 memory, it will need one more insn at least. */
2742 if (this_alternative != NO_REGS
2743 && REG_P (op) && (cl = get_reg_class (REGNO (op))) != NO_REGS
2744 && ((curr_static_id->operand[nop].type != OP_OUT
2745 && targetm.secondary_memory_needed (GET_MODE (op), cl,
2746 this_alternative))
2747 || (curr_static_id->operand[nop].type != OP_IN
2748 && (targetm.secondary_memory_needed
2749 (GET_MODE (op), this_alternative, cl)))))
2750 losers++;
2752 /* Input reloads can be inherited more often than output
2753 reloads can be removed, so penalize output
2754 reloads. */
2755 if (!REG_P (op) || curr_static_id->operand[nop].type != OP_IN)
2757 if (lra_dump_file != NULL)
2758 fprintf
2759 (lra_dump_file,
2760 " %d Non input pseudo reload: reject++\n",
2761 nop);
2762 reject++;
2765 if (MEM_P (op) && offmemok)
2766 addr_losers++;
2767 else if (curr_static_id->operand[nop].type == OP_INOUT)
2769 if (lra_dump_file != NULL)
2770 fprintf
2771 (lra_dump_file,
2772 " %d Input/Output reload: reject+=%d\n",
2773 nop, LRA_LOSER_COST_FACTOR);
2774 reject += LRA_LOSER_COST_FACTOR;
2778 if (early_clobber_p && ! scratch_p)
2780 if (lra_dump_file != NULL)
2781 fprintf (lra_dump_file,
2782 " %d Early clobber: reject++\n", nop);
2783 reject++;
2785 /* ??? We check early clobbers after processing all operands
2786 (see loop below) and there we update the costs more.
2787 Should we update the cost (may be approximately) here
2788 because of early clobber register reloads or it is a rare
2789 or non-important thing to be worth to do it. */
2790 overall = (losers * LRA_LOSER_COST_FACTOR + reject
2791 - (addr_losers == losers ? static_reject : 0));
2792 if ((best_losers == 0 || losers != 0) && best_overall < overall)
2794 if (lra_dump_file != NULL)
2795 fprintf (lra_dump_file,
2796 " alt=%d,overall=%d,losers=%d -- refuse\n",
2797 nalt, overall, losers);
2798 goto fail;
2801 if (update_and_check_small_class_inputs (nop, this_alternative))
2803 if (lra_dump_file != NULL)
2804 fprintf (lra_dump_file,
2805 " alt=%d, not enough small class regs -- refuse\n",
2806 nalt);
2807 goto fail;
2809 curr_alt[nop] = this_alternative;
2810 COPY_HARD_REG_SET (curr_alt_set[nop], this_alternative_set);
2811 curr_alt_win[nop] = this_alternative_win;
2812 curr_alt_match_win[nop] = this_alternative_match_win;
2813 curr_alt_offmemok[nop] = this_alternative_offmemok;
2814 curr_alt_matches[nop] = this_alternative_matches;
2816 if (this_alternative_matches >= 0
2817 && !did_match && !this_alternative_win)
2818 curr_alt_win[this_alternative_matches] = false;
2820 if (early_clobber_p && operand_reg[nop] != NULL_RTX)
2821 early_clobbered_nops[early_clobbered_regs_num++] = nop;
2824 if (curr_insn_set != NULL_RTX && n_operands == 2
2825 /* Prevent processing non-move insns. */
2826 && (GET_CODE (SET_SRC (curr_insn_set)) == SUBREG
2827 || SET_SRC (curr_insn_set) == no_subreg_reg_operand[1])
2828 && ((! curr_alt_win[0] && ! curr_alt_win[1]
2829 && REG_P (no_subreg_reg_operand[0])
2830 && REG_P (no_subreg_reg_operand[1])
2831 && (reg_in_class_p (no_subreg_reg_operand[0], curr_alt[1])
2832 || reg_in_class_p (no_subreg_reg_operand[1], curr_alt[0])))
2833 || (! curr_alt_win[0] && curr_alt_win[1]
2834 && REG_P (no_subreg_reg_operand[1])
2835 /* Check that we reload memory not the memory
2836 address. */
2837 && ! (curr_alt_offmemok[0]
2838 && MEM_P (no_subreg_reg_operand[0]))
2839 && reg_in_class_p (no_subreg_reg_operand[1], curr_alt[0]))
2840 || (curr_alt_win[0] && ! curr_alt_win[1]
2841 && REG_P (no_subreg_reg_operand[0])
2842 /* Check that we reload memory not the memory
2843 address. */
2844 && ! (curr_alt_offmemok[1]
2845 && MEM_P (no_subreg_reg_operand[1]))
2846 && reg_in_class_p (no_subreg_reg_operand[0], curr_alt[1])
2847 && (! CONST_POOL_OK_P (curr_operand_mode[1],
2848 no_subreg_reg_operand[1])
2849 || (targetm.preferred_reload_class
2850 (no_subreg_reg_operand[1],
2851 (enum reg_class) curr_alt[1]) != NO_REGS))
2852 /* If it is a result of recent elimination in move
2853 insn we can transform it into an add still by
2854 using this alternative. */
2855 && GET_CODE (no_subreg_reg_operand[1]) != PLUS)))
2857 /* We have a move insn and a new reload insn will be similar
2858 to the current insn. We should avoid such situation as
2859 it results in LRA cycling. */
2860 if (lra_dump_file != NULL)
2861 fprintf (lra_dump_file,
2862 " Cycle danger: overall += LRA_MAX_REJECT\n");
2863 overall += LRA_MAX_REJECT;
2865 ok_p = true;
2866 curr_alt_dont_inherit_ops_num = 0;
2867 for (nop = 0; nop < early_clobbered_regs_num; nop++)
2869 int i, j, clobbered_hard_regno, first_conflict_j, last_conflict_j;
2870 HARD_REG_SET temp_set;
2872 i = early_clobbered_nops[nop];
2873 if ((! curr_alt_win[i] && ! curr_alt_match_win[i])
2874 || hard_regno[i] < 0)
2875 continue;
2876 lra_assert (operand_reg[i] != NULL_RTX);
2877 clobbered_hard_regno = hard_regno[i];
2878 CLEAR_HARD_REG_SET (temp_set);
2879 add_to_hard_reg_set (&temp_set, biggest_mode[i], clobbered_hard_regno);
2880 first_conflict_j = last_conflict_j = -1;
2881 for (j = 0; j < n_operands; j++)
2882 if (j == i
2883 /* We don't want process insides of match_operator and
2884 match_parallel because otherwise we would process
2885 their operands once again generating a wrong
2886 code. */
2887 || curr_static_id->operand[j].is_operator)
2888 continue;
2889 else if ((curr_alt_matches[j] == i && curr_alt_match_win[j])
2890 || (curr_alt_matches[i] == j && curr_alt_match_win[i]))
2891 continue;
2892 /* If we don't reload j-th operand, check conflicts. */
2893 else if ((curr_alt_win[j] || curr_alt_match_win[j])
2894 && uses_hard_regs_p (*curr_id->operand_loc[j], temp_set))
2896 if (first_conflict_j < 0)
2897 first_conflict_j = j;
2898 last_conflict_j = j;
2900 if (last_conflict_j < 0)
2901 continue;
2902 /* If earlyclobber operand conflicts with another
2903 non-matching operand which is actually the same register
2904 as the earlyclobber operand, it is better to reload the
2905 another operand as an operand matching the earlyclobber
2906 operand can be also the same. */
2907 if (first_conflict_j == last_conflict_j
2908 && operand_reg[last_conflict_j] != NULL_RTX
2909 && ! curr_alt_match_win[last_conflict_j]
2910 && REGNO (operand_reg[i]) == REGNO (operand_reg[last_conflict_j]))
2912 curr_alt_win[last_conflict_j] = false;
2913 curr_alt_dont_inherit_ops[curr_alt_dont_inherit_ops_num++]
2914 = last_conflict_j;
2915 losers++;
2916 /* Early clobber was already reflected in REJECT. */
2917 lra_assert (reject > 0);
2918 if (lra_dump_file != NULL)
2919 fprintf
2920 (lra_dump_file,
2921 " %d Conflict early clobber reload: reject--\n",
2923 reject--;
2924 overall += LRA_LOSER_COST_FACTOR - 1;
2926 else
2928 /* We need to reload early clobbered register and the
2929 matched registers. */
2930 for (j = 0; j < n_operands; j++)
2931 if (curr_alt_matches[j] == i)
2933 curr_alt_match_win[j] = false;
2934 losers++;
2935 overall += LRA_LOSER_COST_FACTOR;
2937 if (! curr_alt_match_win[i])
2938 curr_alt_dont_inherit_ops[curr_alt_dont_inherit_ops_num++] = i;
2939 else
2941 /* Remember pseudos used for match reloads are never
2942 inherited. */
2943 lra_assert (curr_alt_matches[i] >= 0);
2944 curr_alt_win[curr_alt_matches[i]] = false;
2946 curr_alt_win[i] = curr_alt_match_win[i] = false;
2947 losers++;
2948 /* Early clobber was already reflected in REJECT. */
2949 lra_assert (reject > 0);
2950 if (lra_dump_file != NULL)
2951 fprintf
2952 (lra_dump_file,
2953 " %d Matched conflict early clobber reloads: "
2954 "reject--\n",
2956 reject--;
2957 overall += LRA_LOSER_COST_FACTOR - 1;
2960 if (lra_dump_file != NULL)
2961 fprintf (lra_dump_file, " alt=%d,overall=%d,losers=%d,rld_nregs=%d\n",
2962 nalt, overall, losers, reload_nregs);
2964 /* If this alternative can be made to work by reloading, and it
2965 needs less reloading than the others checked so far, record
2966 it as the chosen goal for reloading. */
2967 if ((best_losers != 0 && losers == 0)
2968 || (((best_losers == 0 && losers == 0)
2969 || (best_losers != 0 && losers != 0))
2970 && (best_overall > overall
2971 || (best_overall == overall
2972 /* If the cost of the reloads is the same,
2973 prefer alternative which requires minimal
2974 number of reload regs. */
2975 && (reload_nregs < best_reload_nregs
2976 || (reload_nregs == best_reload_nregs
2977 && (best_reload_sum < reload_sum
2978 || (best_reload_sum == reload_sum
2979 && nalt < goal_alt_number))))))))
2981 for (nop = 0; nop < n_operands; nop++)
2983 goal_alt_win[nop] = curr_alt_win[nop];
2984 goal_alt_match_win[nop] = curr_alt_match_win[nop];
2985 goal_alt_matches[nop] = curr_alt_matches[nop];
2986 goal_alt[nop] = curr_alt[nop];
2987 goal_alt_offmemok[nop] = curr_alt_offmemok[nop];
2989 goal_alt_dont_inherit_ops_num = curr_alt_dont_inherit_ops_num;
2990 for (nop = 0; nop < curr_alt_dont_inherit_ops_num; nop++)
2991 goal_alt_dont_inherit_ops[nop] = curr_alt_dont_inherit_ops[nop];
2992 goal_alt_swapped = curr_swapped;
2993 best_overall = overall;
2994 best_losers = losers;
2995 best_reload_nregs = reload_nregs;
2996 best_reload_sum = reload_sum;
2997 goal_alt_number = nalt;
2999 if (losers == 0)
3000 /* Everything is satisfied. Do not process alternatives
3001 anymore. */
3002 break;
3003 fail:
3006 return ok_p;
3009 /* Make reload base reg from address AD. */
3010 static rtx
3011 base_to_reg (struct address_info *ad)
3013 enum reg_class cl;
3014 int code = -1;
3015 rtx new_inner = NULL_RTX;
3016 rtx new_reg = NULL_RTX;
3017 rtx_insn *insn;
3018 rtx_insn *last_insn = get_last_insn();
3020 lra_assert (ad->disp == ad->disp_term);
3021 cl = base_reg_class (ad->mode, ad->as, ad->base_outer_code,
3022 get_index_code (ad));
3023 new_reg = lra_create_new_reg (GET_MODE (*ad->base), NULL_RTX,
3024 cl, "base");
3025 new_inner = simplify_gen_binary (PLUS, GET_MODE (new_reg), new_reg,
3026 ad->disp_term == NULL
3027 ? const0_rtx
3028 : *ad->disp_term);
3029 if (!valid_address_p (ad->mode, new_inner, ad->as))
3030 return NULL_RTX;
3031 insn = emit_insn (gen_rtx_SET (new_reg, *ad->base));
3032 code = recog_memoized (insn);
3033 if (code < 0)
3035 delete_insns_since (last_insn);
3036 return NULL_RTX;
3039 return new_inner;
3042 /* Make reload base reg + disp from address AD. Return the new pseudo. */
3043 static rtx
3044 base_plus_disp_to_reg (struct address_info *ad)
3046 enum reg_class cl;
3047 rtx new_reg;
3049 lra_assert (ad->base == ad->base_term && ad->disp == ad->disp_term);
3050 cl = base_reg_class (ad->mode, ad->as, ad->base_outer_code,
3051 get_index_code (ad));
3052 new_reg = lra_create_new_reg (GET_MODE (*ad->base_term), NULL_RTX,
3053 cl, "base + disp");
3054 lra_emit_add (new_reg, *ad->base_term, *ad->disp_term);
3055 return new_reg;
3058 /* Make reload of index part of address AD. Return the new
3059 pseudo. */
3060 static rtx
3061 index_part_to_reg (struct address_info *ad)
3063 rtx new_reg;
3065 new_reg = lra_create_new_reg (GET_MODE (*ad->index), NULL_RTX,
3066 INDEX_REG_CLASS, "index term");
3067 expand_mult (GET_MODE (*ad->index), *ad->index_term,
3068 GEN_INT (get_index_scale (ad)), new_reg, 1);
3069 return new_reg;
3072 /* Return true if we can add a displacement to address AD, even if that
3073 makes the address invalid. The fix-up code requires any new address
3074 to be the sum of the BASE_TERM, INDEX and DISP_TERM fields. */
3075 static bool
3076 can_add_disp_p (struct address_info *ad)
3078 return (!ad->autoinc_p
3079 && ad->segment == NULL
3080 && ad->base == ad->base_term
3081 && ad->disp == ad->disp_term);
3084 /* Make equiv substitution in address AD. Return true if a substitution
3085 was made. */
3086 static bool
3087 equiv_address_substitution (struct address_info *ad)
3089 rtx base_reg, new_base_reg, index_reg, new_index_reg, *base_term, *index_term;
3090 HOST_WIDE_INT disp, scale;
3091 bool change_p;
3093 base_term = strip_subreg (ad->base_term);
3094 if (base_term == NULL)
3095 base_reg = new_base_reg = NULL_RTX;
3096 else
3098 base_reg = *base_term;
3099 new_base_reg = get_equiv_with_elimination (base_reg, curr_insn);
3101 index_term = strip_subreg (ad->index_term);
3102 if (index_term == NULL)
3103 index_reg = new_index_reg = NULL_RTX;
3104 else
3106 index_reg = *index_term;
3107 new_index_reg = get_equiv_with_elimination (index_reg, curr_insn);
3109 if (base_reg == new_base_reg && index_reg == new_index_reg)
3110 return false;
3111 disp = 0;
3112 change_p = false;
3113 if (lra_dump_file != NULL)
3115 fprintf (lra_dump_file, "Changing address in insn %d ",
3116 INSN_UID (curr_insn));
3117 dump_value_slim (lra_dump_file, *ad->outer, 1);
3119 if (base_reg != new_base_reg)
3121 if (REG_P (new_base_reg))
3123 *base_term = new_base_reg;
3124 change_p = true;
3126 else if (GET_CODE (new_base_reg) == PLUS
3127 && REG_P (XEXP (new_base_reg, 0))
3128 && CONST_INT_P (XEXP (new_base_reg, 1))
3129 && can_add_disp_p (ad))
3131 disp += INTVAL (XEXP (new_base_reg, 1));
3132 *base_term = XEXP (new_base_reg, 0);
3133 change_p = true;
3135 if (ad->base_term2 != NULL)
3136 *ad->base_term2 = *ad->base_term;
3138 if (index_reg != new_index_reg)
3140 if (REG_P (new_index_reg))
3142 *index_term = new_index_reg;
3143 change_p = true;
3145 else if (GET_CODE (new_index_reg) == PLUS
3146 && REG_P (XEXP (new_index_reg, 0))
3147 && CONST_INT_P (XEXP (new_index_reg, 1))
3148 && can_add_disp_p (ad)
3149 && (scale = get_index_scale (ad)))
3151 disp += INTVAL (XEXP (new_index_reg, 1)) * scale;
3152 *index_term = XEXP (new_index_reg, 0);
3153 change_p = true;
3156 if (disp != 0)
3158 if (ad->disp != NULL)
3159 *ad->disp = plus_constant (GET_MODE (*ad->inner), *ad->disp, disp);
3160 else
3162 *ad->inner = plus_constant (GET_MODE (*ad->inner), *ad->inner, disp);
3163 update_address (ad);
3165 change_p = true;
3167 if (lra_dump_file != NULL)
3169 if (! change_p)
3170 fprintf (lra_dump_file, " -- no change\n");
3171 else
3173 fprintf (lra_dump_file, " on equiv ");
3174 dump_value_slim (lra_dump_file, *ad->outer, 1);
3175 fprintf (lra_dump_file, "\n");
3178 return change_p;
3181 /* Major function to make reloads for an address in operand NOP or
3182 check its correctness (If CHECK_ONLY_P is true). The supported
3183 cases are:
3185 1) an address that existed before LRA started, at which point it
3186 must have been valid. These addresses are subject to elimination
3187 and may have become invalid due to the elimination offset being out
3188 of range.
3190 2) an address created by forcing a constant to memory
3191 (force_const_to_mem). The initial form of these addresses might
3192 not be valid, and it is this function's job to make them valid.
3194 3) a frame address formed from a register and a (possibly zero)
3195 constant offset. As above, these addresses might not be valid and
3196 this function must make them so.
3198 Add reloads to the lists *BEFORE and *AFTER. We might need to add
3199 reloads to *AFTER because of inc/dec, {pre, post} modify in the
3200 address. Return true for any RTL change.
3202 The function is a helper function which does not produce all
3203 transformations (when CHECK_ONLY_P is false) which can be
3204 necessary. It does just basic steps. To do all necessary
3205 transformations use function process_address. */
3206 static bool
3207 process_address_1 (int nop, bool check_only_p,
3208 rtx_insn **before, rtx_insn **after)
3210 struct address_info ad;
3211 rtx new_reg;
3212 HOST_WIDE_INT scale;
3213 rtx op = *curr_id->operand_loc[nop];
3214 const char *constraint = curr_static_id->operand[nop].constraint;
3215 enum constraint_num cn = lookup_constraint (constraint);
3216 bool change_p = false;
3218 if (MEM_P (op)
3219 && GET_MODE (op) == BLKmode
3220 && GET_CODE (XEXP (op, 0)) == SCRATCH)
3221 return false;
3223 if (insn_extra_address_constraint (cn))
3224 decompose_lea_address (&ad, curr_id->operand_loc[nop]);
3225 /* Do not attempt to decompose arbitrary addresses generated by combine
3226 for asm operands with loose constraints, e.g 'X'. */
3227 else if (MEM_P (op)
3228 && !(get_constraint_type (cn) == CT_FIXED_FORM
3229 && constraint_satisfied_p (op, cn)))
3230 decompose_mem_address (&ad, op);
3231 else if (GET_CODE (op) == SUBREG
3232 && MEM_P (SUBREG_REG (op)))
3233 decompose_mem_address (&ad, SUBREG_REG (op));
3234 else
3235 return false;
3236 /* If INDEX_REG_CLASS is assigned to base_term already and isn't to
3237 index_term, swap them so to avoid assigning INDEX_REG_CLASS to both
3238 when INDEX_REG_CLASS is a single register class. */
3239 if (ad.base_term != NULL
3240 && ad.index_term != NULL
3241 && ira_class_hard_regs_num[INDEX_REG_CLASS] == 1
3242 && REG_P (*ad.base_term)
3243 && REG_P (*ad.index_term)
3244 && in_class_p (*ad.base_term, INDEX_REG_CLASS, NULL)
3245 && ! in_class_p (*ad.index_term, INDEX_REG_CLASS, NULL))
3247 std::swap (ad.base, ad.index);
3248 std::swap (ad.base_term, ad.index_term);
3250 if (! check_only_p)
3251 change_p = equiv_address_substitution (&ad);
3252 if (ad.base_term != NULL
3253 && (process_addr_reg
3254 (ad.base_term, check_only_p, before,
3255 (ad.autoinc_p
3256 && !(REG_P (*ad.base_term)
3257 && find_regno_note (curr_insn, REG_DEAD,
3258 REGNO (*ad.base_term)) != NULL_RTX)
3259 ? after : NULL),
3260 base_reg_class (ad.mode, ad.as, ad.base_outer_code,
3261 get_index_code (&ad)))))
3263 change_p = true;
3264 if (ad.base_term2 != NULL)
3265 *ad.base_term2 = *ad.base_term;
3267 if (ad.index_term != NULL
3268 && process_addr_reg (ad.index_term, check_only_p,
3269 before, NULL, INDEX_REG_CLASS))
3270 change_p = true;
3272 /* Target hooks sometimes don't treat extra-constraint addresses as
3273 legitimate address_operands, so handle them specially. */
3274 if (insn_extra_address_constraint (cn)
3275 && satisfies_address_constraint_p (&ad, cn))
3276 return change_p;
3278 if (check_only_p)
3279 return change_p;
3281 /* There are three cases where the shape of *AD.INNER may now be invalid:
3283 1) the original address was valid, but either elimination or
3284 equiv_address_substitution was applied and that made
3285 the address invalid.
3287 2) the address is an invalid symbolic address created by
3288 force_const_to_mem.
3290 3) the address is a frame address with an invalid offset.
3292 4) the address is a frame address with an invalid base.
3294 All these cases involve a non-autoinc address, so there is no
3295 point revalidating other types. */
3296 if (ad.autoinc_p || valid_address_p (&ad))
3297 return change_p;
3299 /* Any index existed before LRA started, so we can assume that the
3300 presence and shape of the index is valid. */
3301 push_to_sequence (*before);
3302 lra_assert (ad.disp == ad.disp_term);
3303 if (ad.base == NULL)
3305 if (ad.index == NULL)
3307 rtx_insn *insn;
3308 rtx_insn *last = get_last_insn ();
3309 int code = -1;
3310 enum reg_class cl = base_reg_class (ad.mode, ad.as,
3311 SCRATCH, SCRATCH);
3312 rtx addr = *ad.inner;
3314 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "addr");
3315 if (HAVE_lo_sum)
3317 /* addr => lo_sum (new_base, addr), case (2) above. */
3318 insn = emit_insn (gen_rtx_SET
3319 (new_reg,
3320 gen_rtx_HIGH (Pmode, copy_rtx (addr))));
3321 code = recog_memoized (insn);
3322 if (code >= 0)
3324 *ad.inner = gen_rtx_LO_SUM (Pmode, new_reg, addr);
3325 if (! valid_address_p (ad.mode, *ad.outer, ad.as))
3327 /* Try to put lo_sum into register. */
3328 insn = emit_insn (gen_rtx_SET
3329 (new_reg,
3330 gen_rtx_LO_SUM (Pmode, new_reg, addr)));
3331 code = recog_memoized (insn);
3332 if (code >= 0)
3334 *ad.inner = new_reg;
3335 if (! valid_address_p (ad.mode, *ad.outer, ad.as))
3337 *ad.inner = addr;
3338 code = -1;
3344 if (code < 0)
3345 delete_insns_since (last);
3348 if (code < 0)
3350 /* addr => new_base, case (2) above. */
3351 lra_emit_move (new_reg, addr);
3353 for (insn = last == NULL_RTX ? get_insns () : NEXT_INSN (last);
3354 insn != NULL_RTX;
3355 insn = NEXT_INSN (insn))
3356 if (recog_memoized (insn) < 0)
3357 break;
3358 if (insn != NULL_RTX)
3360 /* Do nothing if we cannot generate right insns.
3361 This is analogous to reload pass behavior. */
3362 delete_insns_since (last);
3363 end_sequence ();
3364 return false;
3366 *ad.inner = new_reg;
3369 else
3371 /* index * scale + disp => new base + index * scale,
3372 case (1) above. */
3373 enum reg_class cl = base_reg_class (ad.mode, ad.as, PLUS,
3374 GET_CODE (*ad.index));
3376 lra_assert (INDEX_REG_CLASS != NO_REGS);
3377 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "disp");
3378 lra_emit_move (new_reg, *ad.disp);
3379 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3380 new_reg, *ad.index);
3383 else if (ad.index == NULL)
3385 int regno;
3386 enum reg_class cl;
3387 rtx set;
3388 rtx_insn *insns, *last_insn;
3389 /* Try to reload base into register only if the base is invalid
3390 for the address but with valid offset, case (4) above. */
3391 start_sequence ();
3392 new_reg = base_to_reg (&ad);
3394 /* base + disp => new base, cases (1) and (3) above. */
3395 /* Another option would be to reload the displacement into an
3396 index register. However, postreload has code to optimize
3397 address reloads that have the same base and different
3398 displacements, so reloading into an index register would
3399 not necessarily be a win. */
3400 if (new_reg == NULL_RTX)
3401 new_reg = base_plus_disp_to_reg (&ad);
3402 insns = get_insns ();
3403 last_insn = get_last_insn ();
3404 /* If we generated at least two insns, try last insn source as
3405 an address. If we succeed, we generate one less insn. */
3406 if (last_insn != insns && (set = single_set (last_insn)) != NULL_RTX
3407 && GET_CODE (SET_SRC (set)) == PLUS
3408 && REG_P (XEXP (SET_SRC (set), 0))
3409 && CONSTANT_P (XEXP (SET_SRC (set), 1)))
3411 *ad.inner = SET_SRC (set);
3412 if (valid_address_p (ad.mode, *ad.outer, ad.as))
3414 *ad.base_term = XEXP (SET_SRC (set), 0);
3415 *ad.disp_term = XEXP (SET_SRC (set), 1);
3416 cl = base_reg_class (ad.mode, ad.as, ad.base_outer_code,
3417 get_index_code (&ad));
3418 regno = REGNO (*ad.base_term);
3419 if (regno >= FIRST_PSEUDO_REGISTER
3420 && cl != lra_get_allocno_class (regno))
3421 lra_change_class (regno, cl, " Change to", true);
3422 new_reg = SET_SRC (set);
3423 delete_insns_since (PREV_INSN (last_insn));
3426 /* Try if target can split displacement into legitimite new disp
3427 and offset. If it's the case, we replace the last insn with
3428 insns for base + offset => new_reg and set new_reg + new disp
3429 to *ad.inner. */
3430 last_insn = get_last_insn ();
3431 if ((set = single_set (last_insn)) != NULL_RTX
3432 && GET_CODE (SET_SRC (set)) == PLUS
3433 && REG_P (XEXP (SET_SRC (set), 0))
3434 && REGNO (XEXP (SET_SRC (set), 0)) < FIRST_PSEUDO_REGISTER
3435 && CONST_INT_P (XEXP (SET_SRC (set), 1)))
3437 rtx addend, disp = XEXP (SET_SRC (set), 1);
3438 if (targetm.legitimize_address_displacement (&disp, &addend,
3439 ad.mode))
3441 rtx_insn *new_insns;
3442 start_sequence ();
3443 lra_emit_add (new_reg, XEXP (SET_SRC (set), 0), addend);
3444 new_insns = get_insns ();
3445 end_sequence ();
3446 new_reg = gen_rtx_PLUS (Pmode, new_reg, disp);
3447 delete_insns_since (PREV_INSN (last_insn));
3448 add_insn (new_insns);
3449 insns = get_insns ();
3452 end_sequence ();
3453 emit_insn (insns);
3454 *ad.inner = new_reg;
3456 else if (ad.disp_term != NULL)
3458 /* base + scale * index + disp => new base + scale * index,
3459 case (1) above. */
3460 new_reg = base_plus_disp_to_reg (&ad);
3461 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3462 new_reg, *ad.index);
3464 else if ((scale = get_index_scale (&ad)) == 1)
3466 /* The last transformation to one reg will be made in
3467 curr_insn_transform function. */
3468 end_sequence ();
3469 return false;
3471 else if (scale != 0)
3473 /* base + scale * index => base + new_reg,
3474 case (1) above.
3475 Index part of address may become invalid. For example, we
3476 changed pseudo on the equivalent memory and a subreg of the
3477 pseudo onto the memory of different mode for which the scale is
3478 prohibitted. */
3479 new_reg = index_part_to_reg (&ad);
3480 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3481 *ad.base_term, new_reg);
3483 else
3485 enum reg_class cl = base_reg_class (ad.mode, ad.as,
3486 SCRATCH, SCRATCH);
3487 rtx addr = *ad.inner;
3489 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "addr");
3490 /* addr => new_base. */
3491 lra_emit_move (new_reg, addr);
3492 *ad.inner = new_reg;
3494 *before = get_insns ();
3495 end_sequence ();
3496 return true;
3499 /* If CHECK_ONLY_P is false, do address reloads until it is necessary.
3500 Use process_address_1 as a helper function. Return true for any
3501 RTL changes.
3503 If CHECK_ONLY_P is true, just check address correctness. Return
3504 false if the address correct. */
3505 static bool
3506 process_address (int nop, bool check_only_p,
3507 rtx_insn **before, rtx_insn **after)
3509 bool res = false;
3511 while (process_address_1 (nop, check_only_p, before, after))
3513 if (check_only_p)
3514 return true;
3515 res = true;
3517 return res;
3520 /* Emit insns to reload VALUE into a new register. VALUE is an
3521 auto-increment or auto-decrement RTX whose operand is a register or
3522 memory location; so reloading involves incrementing that location.
3523 IN is either identical to VALUE, or some cheaper place to reload
3524 value being incremented/decremented from.
3526 INC_AMOUNT is the number to increment or decrement by (always
3527 positive and ignored for POST_MODIFY/PRE_MODIFY).
3529 Return pseudo containing the result. */
3530 static rtx
3531 emit_inc (enum reg_class new_rclass, rtx in, rtx value, int inc_amount)
3533 /* REG or MEM to be copied and incremented. */
3534 rtx incloc = XEXP (value, 0);
3535 /* Nonzero if increment after copying. */
3536 int post = (GET_CODE (value) == POST_DEC || GET_CODE (value) == POST_INC
3537 || GET_CODE (value) == POST_MODIFY);
3538 rtx_insn *last;
3539 rtx inc;
3540 rtx_insn *add_insn;
3541 int code;
3542 rtx real_in = in == value ? incloc : in;
3543 rtx result;
3544 bool plus_p = true;
3546 if (GET_CODE (value) == PRE_MODIFY || GET_CODE (value) == POST_MODIFY)
3548 lra_assert (GET_CODE (XEXP (value, 1)) == PLUS
3549 || GET_CODE (XEXP (value, 1)) == MINUS);
3550 lra_assert (rtx_equal_p (XEXP (XEXP (value, 1), 0), XEXP (value, 0)));
3551 plus_p = GET_CODE (XEXP (value, 1)) == PLUS;
3552 inc = XEXP (XEXP (value, 1), 1);
3554 else
3556 if (GET_CODE (value) == PRE_DEC || GET_CODE (value) == POST_DEC)
3557 inc_amount = -inc_amount;
3559 inc = GEN_INT (inc_amount);
3562 if (! post && REG_P (incloc))
3563 result = incloc;
3564 else
3565 result = lra_create_new_reg (GET_MODE (value), value, new_rclass,
3566 "INC/DEC result");
3568 if (real_in != result)
3570 /* First copy the location to the result register. */
3571 lra_assert (REG_P (result));
3572 emit_insn (gen_move_insn (result, real_in));
3575 /* We suppose that there are insns to add/sub with the constant
3576 increment permitted in {PRE/POST)_{DEC/INC/MODIFY}. At least the
3577 old reload worked with this assumption. If the assumption
3578 becomes wrong, we should use approach in function
3579 base_plus_disp_to_reg. */
3580 if (in == value)
3582 /* See if we can directly increment INCLOC. */
3583 last = get_last_insn ();
3584 add_insn = emit_insn (plus_p
3585 ? gen_add2_insn (incloc, inc)
3586 : gen_sub2_insn (incloc, inc));
3588 code = recog_memoized (add_insn);
3589 if (code >= 0)
3591 if (! post && result != incloc)
3592 emit_insn (gen_move_insn (result, incloc));
3593 return result;
3595 delete_insns_since (last);
3598 /* If couldn't do the increment directly, must increment in RESULT.
3599 The way we do this depends on whether this is pre- or
3600 post-increment. For pre-increment, copy INCLOC to the reload
3601 register, increment it there, then save back. */
3602 if (! post)
3604 if (real_in != result)
3605 emit_insn (gen_move_insn (result, real_in));
3606 if (plus_p)
3607 emit_insn (gen_add2_insn (result, inc));
3608 else
3609 emit_insn (gen_sub2_insn (result, inc));
3610 if (result != incloc)
3611 emit_insn (gen_move_insn (incloc, result));
3613 else
3615 /* Post-increment.
3617 Because this might be a jump insn or a compare, and because
3618 RESULT may not be available after the insn in an input
3619 reload, we must do the incrementing before the insn being
3620 reloaded for.
3622 We have already copied IN to RESULT. Increment the copy in
3623 RESULT, save that back, then decrement RESULT so it has
3624 the original value. */
3625 if (plus_p)
3626 emit_insn (gen_add2_insn (result, inc));
3627 else
3628 emit_insn (gen_sub2_insn (result, inc));
3629 emit_insn (gen_move_insn (incloc, result));
3630 /* Restore non-modified value for the result. We prefer this
3631 way because it does not require an additional hard
3632 register. */
3633 if (plus_p)
3635 if (CONST_INT_P (inc))
3636 emit_insn (gen_add2_insn (result,
3637 gen_int_mode (-INTVAL (inc),
3638 GET_MODE (result))));
3639 else
3640 emit_insn (gen_sub2_insn (result, inc));
3642 else
3643 emit_insn (gen_add2_insn (result, inc));
3645 return result;
3648 /* Return true if the current move insn does not need processing as we
3649 already know that it satisfies its constraints. */
3650 static bool
3651 simple_move_p (void)
3653 rtx dest, src;
3654 enum reg_class dclass, sclass;
3656 lra_assert (curr_insn_set != NULL_RTX);
3657 dest = SET_DEST (curr_insn_set);
3658 src = SET_SRC (curr_insn_set);
3660 /* If the instruction has multiple sets we need to process it even if it
3661 is single_set. This can happen if one or more of the SETs are dead.
3662 See PR73650. */
3663 if (multiple_sets (curr_insn))
3664 return false;
3666 return ((dclass = get_op_class (dest)) != NO_REGS
3667 && (sclass = get_op_class (src)) != NO_REGS
3668 /* The backend guarantees that register moves of cost 2
3669 never need reloads. */
3670 && targetm.register_move_cost (GET_MODE (src), sclass, dclass) == 2);
3673 /* Swap operands NOP and NOP + 1. */
3674 static inline void
3675 swap_operands (int nop)
3677 std::swap (curr_operand_mode[nop], curr_operand_mode[nop + 1]);
3678 std::swap (original_subreg_reg_mode[nop], original_subreg_reg_mode[nop + 1]);
3679 std::swap (*curr_id->operand_loc[nop], *curr_id->operand_loc[nop + 1]);
3680 std::swap (equiv_substition_p[nop], equiv_substition_p[nop + 1]);
3681 /* Swap the duplicates too. */
3682 lra_update_dup (curr_id, nop);
3683 lra_update_dup (curr_id, nop + 1);
3686 /* Main entry point of the constraint code: search the body of the
3687 current insn to choose the best alternative. It is mimicking insn
3688 alternative cost calculation model of former reload pass. That is
3689 because machine descriptions were written to use this model. This
3690 model can be changed in future. Make commutative operand exchange
3691 if it is chosen.
3693 if CHECK_ONLY_P is false, do RTL changes to satisfy the
3694 constraints. Return true if any change happened during function
3695 call.
3697 If CHECK_ONLY_P is true then don't do any transformation. Just
3698 check that the insn satisfies all constraints. If the insn does
3699 not satisfy any constraint, return true. */
3700 static bool
3701 curr_insn_transform (bool check_only_p)
3703 int i, j, k;
3704 int n_operands;
3705 int n_alternatives;
3706 int n_outputs;
3707 int commutative;
3708 signed char goal_alt_matched[MAX_RECOG_OPERANDS][MAX_RECOG_OPERANDS];
3709 signed char match_inputs[MAX_RECOG_OPERANDS + 1];
3710 signed char outputs[MAX_RECOG_OPERANDS + 1];
3711 rtx_insn *before, *after;
3712 bool alt_p = false;
3713 /* Flag that the insn has been changed through a transformation. */
3714 bool change_p;
3715 bool sec_mem_p;
3716 bool use_sec_mem_p;
3717 int max_regno_before;
3718 int reused_alternative_num;
3720 curr_insn_set = single_set (curr_insn);
3721 if (curr_insn_set != NULL_RTX && simple_move_p ())
3722 return false;
3724 no_input_reloads_p = no_output_reloads_p = false;
3725 goal_alt_number = -1;
3726 change_p = sec_mem_p = false;
3727 /* JUMP_INSNs and CALL_INSNs are not allowed to have any output
3728 reloads; neither are insns that SET cc0. Insns that use CC0 are
3729 not allowed to have any input reloads. */
3730 if (JUMP_P (curr_insn) || CALL_P (curr_insn))
3731 no_output_reloads_p = true;
3733 if (HAVE_cc0 && reg_referenced_p (cc0_rtx, PATTERN (curr_insn)))
3734 no_input_reloads_p = true;
3735 if (HAVE_cc0 && reg_set_p (cc0_rtx, PATTERN (curr_insn)))
3736 no_output_reloads_p = true;
3738 n_operands = curr_static_id->n_operands;
3739 n_alternatives = curr_static_id->n_alternatives;
3741 /* Just return "no reloads" if insn has no operands with
3742 constraints. */
3743 if (n_operands == 0 || n_alternatives == 0)
3744 return false;
3746 max_regno_before = max_reg_num ();
3748 for (i = 0; i < n_operands; i++)
3750 goal_alt_matched[i][0] = -1;
3751 goal_alt_matches[i] = -1;
3754 commutative = curr_static_id->commutative;
3756 /* Now see what we need for pseudos that didn't get hard regs or got
3757 the wrong kind of hard reg. For this, we must consider all the
3758 operands together against the register constraints. */
3760 best_losers = best_overall = INT_MAX;
3761 best_reload_sum = 0;
3763 curr_swapped = false;
3764 goal_alt_swapped = false;
3766 if (! check_only_p)
3767 /* Make equivalence substitution and memory subreg elimination
3768 before address processing because an address legitimacy can
3769 depend on memory mode. */
3770 for (i = 0; i < n_operands; i++)
3772 rtx op, subst, old;
3773 bool op_change_p = false;
3775 if (curr_static_id->operand[i].is_operator)
3776 continue;
3778 old = op = *curr_id->operand_loc[i];
3779 if (GET_CODE (old) == SUBREG)
3780 old = SUBREG_REG (old);
3781 subst = get_equiv_with_elimination (old, curr_insn);
3782 original_subreg_reg_mode[i] = VOIDmode;
3783 equiv_substition_p[i] = false;
3784 if (subst != old)
3786 equiv_substition_p[i] = true;
3787 subst = copy_rtx (subst);
3788 lra_assert (REG_P (old));
3789 if (GET_CODE (op) != SUBREG)
3790 *curr_id->operand_loc[i] = subst;
3791 else
3793 SUBREG_REG (op) = subst;
3794 if (GET_MODE (subst) == VOIDmode)
3795 original_subreg_reg_mode[i] = GET_MODE (old);
3797 if (lra_dump_file != NULL)
3799 fprintf (lra_dump_file,
3800 "Changing pseudo %d in operand %i of insn %u on equiv ",
3801 REGNO (old), i, INSN_UID (curr_insn));
3802 dump_value_slim (lra_dump_file, subst, 1);
3803 fprintf (lra_dump_file, "\n");
3805 op_change_p = change_p = true;
3807 if (simplify_operand_subreg (i, GET_MODE (old)) || op_change_p)
3809 change_p = true;
3810 lra_update_dup (curr_id, i);
3814 /* Reload address registers and displacements. We do it before
3815 finding an alternative because of memory constraints. */
3816 before = after = NULL;
3817 for (i = 0; i < n_operands; i++)
3818 if (! curr_static_id->operand[i].is_operator
3819 && process_address (i, check_only_p, &before, &after))
3821 if (check_only_p)
3822 return true;
3823 change_p = true;
3824 lra_update_dup (curr_id, i);
3827 if (change_p)
3828 /* If we've changed the instruction then any alternative that
3829 we chose previously may no longer be valid. */
3830 lra_set_used_insn_alternative (curr_insn, -1);
3832 if (! check_only_p && curr_insn_set != NULL_RTX
3833 && check_and_process_move (&change_p, &sec_mem_p))
3834 return change_p;
3836 try_swapped:
3838 reused_alternative_num = check_only_p ? -1 : curr_id->used_insn_alternative;
3839 if (lra_dump_file != NULL && reused_alternative_num >= 0)
3840 fprintf (lra_dump_file, "Reusing alternative %d for insn #%u\n",
3841 reused_alternative_num, INSN_UID (curr_insn));
3843 if (process_alt_operands (reused_alternative_num))
3844 alt_p = true;
3846 if (check_only_p)
3847 return ! alt_p || best_losers != 0;
3849 /* If insn is commutative (it's safe to exchange a certain pair of
3850 operands) then we need to try each alternative twice, the second
3851 time matching those two operands as if we had exchanged them. To
3852 do this, really exchange them in operands.
3854 If we have just tried the alternatives the second time, return
3855 operands to normal and drop through. */
3857 if (reused_alternative_num < 0 && commutative >= 0)
3859 curr_swapped = !curr_swapped;
3860 if (curr_swapped)
3862 swap_operands (commutative);
3863 goto try_swapped;
3865 else
3866 swap_operands (commutative);
3869 if (! alt_p && ! sec_mem_p)
3871 /* No alternative works with reloads?? */
3872 if (INSN_CODE (curr_insn) >= 0)
3873 fatal_insn ("unable to generate reloads for:", curr_insn);
3874 error_for_asm (curr_insn,
3875 "inconsistent operand constraints in an %<asm%>");
3876 /* Avoid further trouble with this insn. Don't generate use
3877 pattern here as we could use the insn SP offset. */
3878 lra_set_insn_deleted (curr_insn);
3879 return true;
3882 /* If the best alternative is with operands 1 and 2 swapped, swap
3883 them. Update the operand numbers of any reloads already
3884 pushed. */
3886 if (goal_alt_swapped)
3888 if (lra_dump_file != NULL)
3889 fprintf (lra_dump_file, " Commutative operand exchange in insn %u\n",
3890 INSN_UID (curr_insn));
3892 /* Swap the duplicates too. */
3893 swap_operands (commutative);
3894 change_p = true;
3897 /* Some targets' TARGET_SECONDARY_MEMORY_NEEDED (e.g. x86) are defined
3898 too conservatively. So we use the secondary memory only if there
3899 is no any alternative without reloads. */
3900 use_sec_mem_p = false;
3901 if (! alt_p)
3902 use_sec_mem_p = true;
3903 else if (sec_mem_p)
3905 for (i = 0; i < n_operands; i++)
3906 if (! goal_alt_win[i] && ! goal_alt_match_win[i])
3907 break;
3908 use_sec_mem_p = i < n_operands;
3911 if (use_sec_mem_p)
3913 int in = -1, out = -1;
3914 rtx new_reg, src, dest, rld;
3915 machine_mode sec_mode, rld_mode;
3917 lra_assert (curr_insn_set != NULL_RTX && sec_mem_p);
3918 dest = SET_DEST (curr_insn_set);
3919 src = SET_SRC (curr_insn_set);
3920 for (i = 0; i < n_operands; i++)
3921 if (*curr_id->operand_loc[i] == dest)
3922 out = i;
3923 else if (*curr_id->operand_loc[i] == src)
3924 in = i;
3925 for (i = 0; i < curr_static_id->n_dups; i++)
3926 if (out < 0 && *curr_id->dup_loc[i] == dest)
3927 out = curr_static_id->dup_num[i];
3928 else if (in < 0 && *curr_id->dup_loc[i] == src)
3929 in = curr_static_id->dup_num[i];
3930 lra_assert (out >= 0 && in >= 0
3931 && curr_static_id->operand[out].type == OP_OUT
3932 && curr_static_id->operand[in].type == OP_IN);
3933 rld = partial_subreg_p (GET_MODE (src), GET_MODE (dest)) ? src : dest;
3934 rld_mode = GET_MODE (rld);
3935 sec_mode = targetm.secondary_memory_needed_mode (rld_mode);
3936 new_reg = lra_create_new_reg (sec_mode, NULL_RTX,
3937 NO_REGS, "secondary");
3938 /* If the mode is changed, it should be wider. */
3939 lra_assert (!partial_subreg_p (sec_mode, rld_mode));
3940 if (sec_mode != rld_mode)
3942 /* If the target says specifically to use another mode for
3943 secondary memory moves we can not reuse the original
3944 insn. */
3945 after = emit_spill_move (false, new_reg, dest);
3946 lra_process_new_insns (curr_insn, NULL, after,
3947 "Inserting the sec. move");
3948 /* We may have non null BEFORE here (e.g. after address
3949 processing. */
3950 push_to_sequence (before);
3951 before = emit_spill_move (true, new_reg, src);
3952 emit_insn (before);
3953 before = get_insns ();
3954 end_sequence ();
3955 lra_process_new_insns (curr_insn, before, NULL, "Changing on");
3956 lra_set_insn_deleted (curr_insn);
3958 else if (dest == rld)
3960 *curr_id->operand_loc[out] = new_reg;
3961 lra_update_dup (curr_id, out);
3962 after = emit_spill_move (false, new_reg, dest);
3963 lra_process_new_insns (curr_insn, NULL, after,
3964 "Inserting the sec. move");
3966 else
3968 *curr_id->operand_loc[in] = new_reg;
3969 lra_update_dup (curr_id, in);
3970 /* See comments above. */
3971 push_to_sequence (before);
3972 before = emit_spill_move (true, new_reg, src);
3973 emit_insn (before);
3974 before = get_insns ();
3975 end_sequence ();
3976 lra_process_new_insns (curr_insn, before, NULL,
3977 "Inserting the sec. move");
3979 lra_update_insn_regno_info (curr_insn);
3980 return true;
3983 lra_assert (goal_alt_number >= 0);
3984 lra_set_used_insn_alternative (curr_insn, goal_alt_number);
3986 if (lra_dump_file != NULL)
3988 const char *p;
3990 fprintf (lra_dump_file, " Choosing alt %d in insn %u:",
3991 goal_alt_number, INSN_UID (curr_insn));
3992 for (i = 0; i < n_operands; i++)
3994 p = (curr_static_id->operand_alternative
3995 [goal_alt_number * n_operands + i].constraint);
3996 if (*p == '\0')
3997 continue;
3998 fprintf (lra_dump_file, " (%d) ", i);
3999 for (; *p != '\0' && *p != ',' && *p != '#'; p++)
4000 fputc (*p, lra_dump_file);
4002 if (INSN_CODE (curr_insn) >= 0
4003 && (p = get_insn_name (INSN_CODE (curr_insn))) != NULL)
4004 fprintf (lra_dump_file, " {%s}", p);
4005 if (curr_id->sp_offset != 0)
4006 fprintf (lra_dump_file, " (sp_off=%" HOST_WIDE_INT_PRINT "d)",
4007 curr_id->sp_offset);
4008 fprintf (lra_dump_file, "\n");
4011 /* Right now, for any pair of operands I and J that are required to
4012 match, with J < I, goal_alt_matches[I] is J. Add I to
4013 goal_alt_matched[J]. */
4015 for (i = 0; i < n_operands; i++)
4016 if ((j = goal_alt_matches[i]) >= 0)
4018 for (k = 0; goal_alt_matched[j][k] >= 0; k++)
4020 /* We allow matching one output operand and several input
4021 operands. */
4022 lra_assert (k == 0
4023 || (curr_static_id->operand[j].type == OP_OUT
4024 && curr_static_id->operand[i].type == OP_IN
4025 && (curr_static_id->operand
4026 [goal_alt_matched[j][0]].type == OP_IN)));
4027 goal_alt_matched[j][k] = i;
4028 goal_alt_matched[j][k + 1] = -1;
4031 for (i = 0; i < n_operands; i++)
4032 goal_alt_win[i] |= goal_alt_match_win[i];
4034 /* Any constants that aren't allowed and can't be reloaded into
4035 registers are here changed into memory references. */
4036 for (i = 0; i < n_operands; i++)
4037 if (goal_alt_win[i])
4039 int regno;
4040 enum reg_class new_class;
4041 rtx reg = *curr_id->operand_loc[i];
4043 if (GET_CODE (reg) == SUBREG)
4044 reg = SUBREG_REG (reg);
4046 if (REG_P (reg) && (regno = REGNO (reg)) >= FIRST_PSEUDO_REGISTER)
4048 bool ok_p = in_class_p (reg, goal_alt[i], &new_class);
4050 if (new_class != NO_REGS && get_reg_class (regno) != new_class)
4052 lra_assert (ok_p);
4053 lra_change_class (regno, new_class, " Change to", true);
4057 else
4059 const char *constraint;
4060 char c;
4061 rtx op = *curr_id->operand_loc[i];
4062 rtx subreg = NULL_RTX;
4063 machine_mode mode = curr_operand_mode[i];
4065 if (GET_CODE (op) == SUBREG)
4067 subreg = op;
4068 op = SUBREG_REG (op);
4069 mode = GET_MODE (op);
4072 if (CONST_POOL_OK_P (mode, op)
4073 && ((targetm.preferred_reload_class
4074 (op, (enum reg_class) goal_alt[i]) == NO_REGS)
4075 || no_input_reloads_p))
4077 rtx tem = force_const_mem (mode, op);
4079 change_p = true;
4080 if (subreg != NULL_RTX)
4081 tem = gen_rtx_SUBREG (mode, tem, SUBREG_BYTE (subreg));
4083 *curr_id->operand_loc[i] = tem;
4084 lra_update_dup (curr_id, i);
4085 process_address (i, false, &before, &after);
4087 /* If the alternative accepts constant pool refs directly
4088 there will be no reload needed at all. */
4089 if (subreg != NULL_RTX)
4090 continue;
4091 /* Skip alternatives before the one requested. */
4092 constraint = (curr_static_id->operand_alternative
4093 [goal_alt_number * n_operands + i].constraint);
4094 for (;
4095 (c = *constraint) && c != ',' && c != '#';
4096 constraint += CONSTRAINT_LEN (c, constraint))
4098 enum constraint_num cn = lookup_constraint (constraint);
4099 if ((insn_extra_memory_constraint (cn)
4100 || insn_extra_special_memory_constraint (cn))
4101 && satisfies_memory_constraint_p (tem, cn))
4102 break;
4104 if (c == '\0' || c == ',' || c == '#')
4105 continue;
4107 goal_alt_win[i] = true;
4111 n_outputs = 0;
4112 outputs[0] = -1;
4113 for (i = 0; i < n_operands; i++)
4115 int regno;
4116 bool optional_p = false;
4117 rtx old, new_reg;
4118 rtx op = *curr_id->operand_loc[i];
4120 if (goal_alt_win[i])
4122 if (goal_alt[i] == NO_REGS
4123 && REG_P (op)
4124 /* When we assign NO_REGS it means that we will not
4125 assign a hard register to the scratch pseudo by
4126 assigment pass and the scratch pseudo will be
4127 spilled. Spilled scratch pseudos are transformed
4128 back to scratches at the LRA end. */
4129 && lra_former_scratch_operand_p (curr_insn, i)
4130 && lra_former_scratch_p (REGNO (op)))
4132 int regno = REGNO (op);
4133 lra_change_class (regno, NO_REGS, " Change to", true);
4134 if (lra_get_regno_hard_regno (regno) >= 0)
4135 /* We don't have to mark all insn affected by the
4136 spilled pseudo as there is only one such insn, the
4137 current one. */
4138 reg_renumber[regno] = -1;
4139 lra_assert (bitmap_single_bit_set_p
4140 (&lra_reg_info[REGNO (op)].insn_bitmap));
4142 /* We can do an optional reload. If the pseudo got a hard
4143 reg, we might improve the code through inheritance. If
4144 it does not get a hard register we coalesce memory/memory
4145 moves later. Ignore move insns to avoid cycling. */
4146 if (! lra_simple_p
4147 && lra_undo_inheritance_iter < LRA_MAX_INHERITANCE_PASSES
4148 && goal_alt[i] != NO_REGS && REG_P (op)
4149 && (regno = REGNO (op)) >= FIRST_PSEUDO_REGISTER
4150 && regno < new_regno_start
4151 && ! lra_former_scratch_p (regno)
4152 && reg_renumber[regno] < 0
4153 /* Check that the optional reload pseudo will be able to
4154 hold given mode value. */
4155 && ! (prohibited_class_reg_set_mode_p
4156 (goal_alt[i], reg_class_contents[goal_alt[i]],
4157 PSEUDO_REGNO_MODE (regno)))
4158 && (curr_insn_set == NULL_RTX
4159 || !((REG_P (SET_SRC (curr_insn_set))
4160 || MEM_P (SET_SRC (curr_insn_set))
4161 || GET_CODE (SET_SRC (curr_insn_set)) == SUBREG)
4162 && (REG_P (SET_DEST (curr_insn_set))
4163 || MEM_P (SET_DEST (curr_insn_set))
4164 || GET_CODE (SET_DEST (curr_insn_set)) == SUBREG))))
4165 optional_p = true;
4166 else
4167 continue;
4170 /* Operands that match previous ones have already been handled. */
4171 if (goal_alt_matches[i] >= 0)
4172 continue;
4174 /* We should not have an operand with a non-offsettable address
4175 appearing where an offsettable address will do. It also may
4176 be a case when the address should be special in other words
4177 not a general one (e.g. it needs no index reg). */
4178 if (goal_alt_matched[i][0] == -1 && goal_alt_offmemok[i] && MEM_P (op))
4180 enum reg_class rclass;
4181 rtx *loc = &XEXP (op, 0);
4182 enum rtx_code code = GET_CODE (*loc);
4184 push_to_sequence (before);
4185 rclass = base_reg_class (GET_MODE (op), MEM_ADDR_SPACE (op),
4186 MEM, SCRATCH);
4187 if (GET_RTX_CLASS (code) == RTX_AUTOINC)
4188 new_reg = emit_inc (rclass, *loc, *loc,
4189 /* This value does not matter for MODIFY. */
4190 GET_MODE_SIZE (GET_MODE (op)));
4191 else if (get_reload_reg (OP_IN, Pmode, *loc, rclass, FALSE,
4192 "offsetable address", &new_reg))
4193 lra_emit_move (new_reg, *loc);
4194 before = get_insns ();
4195 end_sequence ();
4196 *loc = new_reg;
4197 lra_update_dup (curr_id, i);
4199 else if (goal_alt_matched[i][0] == -1)
4201 machine_mode mode;
4202 rtx reg, *loc;
4203 int hard_regno, byte;
4204 enum op_type type = curr_static_id->operand[i].type;
4206 loc = curr_id->operand_loc[i];
4207 mode = curr_operand_mode[i];
4208 if (GET_CODE (*loc) == SUBREG)
4210 reg = SUBREG_REG (*loc);
4211 byte = SUBREG_BYTE (*loc);
4212 if (REG_P (reg)
4213 /* Strict_low_part requires reload the register not
4214 the sub-register. */
4215 && (curr_static_id->operand[i].strict_low
4216 || (!paradoxical_subreg_p (mode, GET_MODE (reg))
4217 && (hard_regno
4218 = get_try_hard_regno (REGNO (reg))) >= 0
4219 && (simplify_subreg_regno
4220 (hard_regno,
4221 GET_MODE (reg), byte, mode) < 0)
4222 && (goal_alt[i] == NO_REGS
4223 || (simplify_subreg_regno
4224 (ira_class_hard_regs[goal_alt[i]][0],
4225 GET_MODE (reg), byte, mode) >= 0)))))
4227 /* An OP_INOUT is required when reloading a subreg of a
4228 mode wider than a word to ensure that data beyond the
4229 word being reloaded is preserved. Also automatically
4230 ensure that strict_low_part reloads are made into
4231 OP_INOUT which should already be true from the backend
4232 constraints. */
4233 if (type == OP_OUT
4234 && (curr_static_id->operand[i].strict_low
4235 || (GET_MODE_SIZE (GET_MODE (reg)) > UNITS_PER_WORD
4236 && (GET_MODE_SIZE (mode)
4237 < GET_MODE_SIZE (GET_MODE (reg))))))
4238 type = OP_INOUT;
4239 loc = &SUBREG_REG (*loc);
4240 mode = GET_MODE (*loc);
4243 old = *loc;
4244 if (get_reload_reg (type, mode, old, goal_alt[i],
4245 loc != curr_id->operand_loc[i], "", &new_reg)
4246 && type != OP_OUT)
4248 push_to_sequence (before);
4249 lra_emit_move (new_reg, old);
4250 before = get_insns ();
4251 end_sequence ();
4253 *loc = new_reg;
4254 if (type != OP_IN
4255 && find_reg_note (curr_insn, REG_UNUSED, old) == NULL_RTX)
4257 start_sequence ();
4258 lra_emit_move (type == OP_INOUT ? copy_rtx (old) : old, new_reg);
4259 emit_insn (after);
4260 after = get_insns ();
4261 end_sequence ();
4262 *loc = new_reg;
4264 for (j = 0; j < goal_alt_dont_inherit_ops_num; j++)
4265 if (goal_alt_dont_inherit_ops[j] == i)
4267 lra_set_regno_unique_value (REGNO (new_reg));
4268 break;
4270 lra_update_dup (curr_id, i);
4272 else if (curr_static_id->operand[i].type == OP_IN
4273 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4274 == OP_OUT))
4276 /* generate reloads for input and matched outputs. */
4277 match_inputs[0] = i;
4278 match_inputs[1] = -1;
4279 match_reload (goal_alt_matched[i][0], match_inputs, outputs,
4280 goal_alt[i], &before, &after,
4281 curr_static_id->operand_alternative
4282 [goal_alt_number * n_operands + goal_alt_matched[i][0]]
4283 .earlyclobber);
4285 else if (curr_static_id->operand[i].type == OP_OUT
4286 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4287 == OP_IN))
4288 /* Generate reloads for output and matched inputs. */
4289 match_reload (i, goal_alt_matched[i], outputs, goal_alt[i], &before,
4290 &after, curr_static_id->operand_alternative
4291 [goal_alt_number * n_operands + i].earlyclobber);
4292 else if (curr_static_id->operand[i].type == OP_IN
4293 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4294 == OP_IN))
4296 /* Generate reloads for matched inputs. */
4297 match_inputs[0] = i;
4298 for (j = 0; (k = goal_alt_matched[i][j]) >= 0; j++)
4299 match_inputs[j + 1] = k;
4300 match_inputs[j + 1] = -1;
4301 match_reload (-1, match_inputs, outputs, goal_alt[i], &before,
4302 &after, false);
4304 else
4305 /* We must generate code in any case when function
4306 process_alt_operands decides that it is possible. */
4307 gcc_unreachable ();
4309 /* Memorise processed outputs so that output remaining to be processed
4310 can avoid using the same register value (see match_reload). */
4311 if (curr_static_id->operand[i].type == OP_OUT)
4313 outputs[n_outputs++] = i;
4314 outputs[n_outputs] = -1;
4317 if (optional_p)
4319 rtx reg = op;
4321 lra_assert (REG_P (reg));
4322 regno = REGNO (reg);
4323 op = *curr_id->operand_loc[i]; /* Substitution. */
4324 if (GET_CODE (op) == SUBREG)
4325 op = SUBREG_REG (op);
4326 gcc_assert (REG_P (op) && (int) REGNO (op) >= new_regno_start);
4327 bitmap_set_bit (&lra_optional_reload_pseudos, REGNO (op));
4328 lra_reg_info[REGNO (op)].restore_rtx = reg;
4329 if (lra_dump_file != NULL)
4330 fprintf (lra_dump_file,
4331 " Making reload reg %d for reg %d optional\n",
4332 REGNO (op), regno);
4335 if (before != NULL_RTX || after != NULL_RTX
4336 || max_regno_before != max_reg_num ())
4337 change_p = true;
4338 if (change_p)
4340 lra_update_operator_dups (curr_id);
4341 /* Something changes -- process the insn. */
4342 lra_update_insn_regno_info (curr_insn);
4344 lra_process_new_insns (curr_insn, before, after, "Inserting insn reload");
4345 return change_p;
4348 /* Return true if INSN satisfies all constraints. In other words, no
4349 reload insns are needed. */
4350 bool
4351 lra_constrain_insn (rtx_insn *insn)
4353 int saved_new_regno_start = new_regno_start;
4354 int saved_new_insn_uid_start = new_insn_uid_start;
4355 bool change_p;
4357 curr_insn = insn;
4358 curr_id = lra_get_insn_recog_data (curr_insn);
4359 curr_static_id = curr_id->insn_static_data;
4360 new_insn_uid_start = get_max_uid ();
4361 new_regno_start = max_reg_num ();
4362 change_p = curr_insn_transform (true);
4363 new_regno_start = saved_new_regno_start;
4364 new_insn_uid_start = saved_new_insn_uid_start;
4365 return ! change_p;
4368 /* Return true if X is in LIST. */
4369 static bool
4370 in_list_p (rtx x, rtx list)
4372 for (; list != NULL_RTX; list = XEXP (list, 1))
4373 if (XEXP (list, 0) == x)
4374 return true;
4375 return false;
4378 /* Return true if X contains an allocatable hard register (if
4379 HARD_REG_P) or a (spilled if SPILLED_P) pseudo. */
4380 static bool
4381 contains_reg_p (rtx x, bool hard_reg_p, bool spilled_p)
4383 int i, j;
4384 const char *fmt;
4385 enum rtx_code code;
4387 code = GET_CODE (x);
4388 if (REG_P (x))
4390 int regno = REGNO (x);
4391 HARD_REG_SET alloc_regs;
4393 if (hard_reg_p)
4395 if (regno >= FIRST_PSEUDO_REGISTER)
4396 regno = lra_get_regno_hard_regno (regno);
4397 if (regno < 0)
4398 return false;
4399 COMPL_HARD_REG_SET (alloc_regs, lra_no_alloc_regs);
4400 return overlaps_hard_reg_set_p (alloc_regs, GET_MODE (x), regno);
4402 else
4404 if (regno < FIRST_PSEUDO_REGISTER)
4405 return false;
4406 if (! spilled_p)
4407 return true;
4408 return lra_get_regno_hard_regno (regno) < 0;
4411 fmt = GET_RTX_FORMAT (code);
4412 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4414 if (fmt[i] == 'e')
4416 if (contains_reg_p (XEXP (x, i), hard_reg_p, spilled_p))
4417 return true;
4419 else if (fmt[i] == 'E')
4421 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4422 if (contains_reg_p (XVECEXP (x, i, j), hard_reg_p, spilled_p))
4423 return true;
4426 return false;
4429 /* Process all regs in location *LOC and change them on equivalent
4430 substitution. Return true if any change was done. */
4431 static bool
4432 loc_equivalence_change_p (rtx *loc)
4434 rtx subst, reg, x = *loc;
4435 bool result = false;
4436 enum rtx_code code = GET_CODE (x);
4437 const char *fmt;
4438 int i, j;
4440 if (code == SUBREG)
4442 reg = SUBREG_REG (x);
4443 if ((subst = get_equiv_with_elimination (reg, curr_insn)) != reg
4444 && GET_MODE (subst) == VOIDmode)
4446 /* We cannot reload debug location. Simplify subreg here
4447 while we know the inner mode. */
4448 *loc = simplify_gen_subreg (GET_MODE (x), subst,
4449 GET_MODE (reg), SUBREG_BYTE (x));
4450 return true;
4453 if (code == REG && (subst = get_equiv_with_elimination (x, curr_insn)) != x)
4455 *loc = subst;
4456 return true;
4459 /* Scan all the operand sub-expressions. */
4460 fmt = GET_RTX_FORMAT (code);
4461 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4463 if (fmt[i] == 'e')
4464 result = loc_equivalence_change_p (&XEXP (x, i)) || result;
4465 else if (fmt[i] == 'E')
4466 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4467 result
4468 = loc_equivalence_change_p (&XVECEXP (x, i, j)) || result;
4470 return result;
4473 /* Similar to loc_equivalence_change_p, but for use as
4474 simplify_replace_fn_rtx callback. DATA is insn for which the
4475 elimination is done. If it null we don't do the elimination. */
4476 static rtx
4477 loc_equivalence_callback (rtx loc, const_rtx, void *data)
4479 if (!REG_P (loc))
4480 return NULL_RTX;
4482 rtx subst = (data == NULL
4483 ? get_equiv (loc) : get_equiv_with_elimination (loc, (rtx_insn *) data));
4484 if (subst != loc)
4485 return subst;
4487 return NULL_RTX;
4490 /* Maximum number of generated reload insns per an insn. It is for
4491 preventing this pass cycling in a bug case. */
4492 #define MAX_RELOAD_INSNS_NUMBER LRA_MAX_INSN_RELOADS
4494 /* The current iteration number of this LRA pass. */
4495 int lra_constraint_iter;
4497 /* True if we substituted equiv which needs checking register
4498 allocation correctness because the equivalent value contains
4499 allocatable hard registers or when we restore multi-register
4500 pseudo. */
4501 bool lra_risky_transformations_p;
4503 /* Return true if REGNO is referenced in more than one block. */
4504 static bool
4505 multi_block_pseudo_p (int regno)
4507 basic_block bb = NULL;
4508 unsigned int uid;
4509 bitmap_iterator bi;
4511 if (regno < FIRST_PSEUDO_REGISTER)
4512 return false;
4514 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi)
4515 if (bb == NULL)
4516 bb = BLOCK_FOR_INSN (lra_insn_recog_data[uid]->insn);
4517 else if (BLOCK_FOR_INSN (lra_insn_recog_data[uid]->insn) != bb)
4518 return true;
4519 return false;
4522 /* Return true if LIST contains a deleted insn. */
4523 static bool
4524 contains_deleted_insn_p (rtx_insn_list *list)
4526 for (; list != NULL_RTX; list = list->next ())
4527 if (NOTE_P (list->insn ())
4528 && NOTE_KIND (list->insn ()) == NOTE_INSN_DELETED)
4529 return true;
4530 return false;
4533 /* Return true if X contains a pseudo dying in INSN. */
4534 static bool
4535 dead_pseudo_p (rtx x, rtx_insn *insn)
4537 int i, j;
4538 const char *fmt;
4539 enum rtx_code code;
4541 if (REG_P (x))
4542 return (insn != NULL_RTX
4543 && find_regno_note (insn, REG_DEAD, REGNO (x)) != NULL_RTX);
4544 code = GET_CODE (x);
4545 fmt = GET_RTX_FORMAT (code);
4546 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4548 if (fmt[i] == 'e')
4550 if (dead_pseudo_p (XEXP (x, i), insn))
4551 return true;
4553 else if (fmt[i] == 'E')
4555 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4556 if (dead_pseudo_p (XVECEXP (x, i, j), insn))
4557 return true;
4560 return false;
4563 /* Return true if INSN contains a dying pseudo in INSN right hand
4564 side. */
4565 static bool
4566 insn_rhs_dead_pseudo_p (rtx_insn *insn)
4568 rtx set = single_set (insn);
4570 gcc_assert (set != NULL);
4571 return dead_pseudo_p (SET_SRC (set), insn);
4574 /* Return true if any init insn of REGNO contains a dying pseudo in
4575 insn right hand side. */
4576 static bool
4577 init_insn_rhs_dead_pseudo_p (int regno)
4579 rtx_insn_list *insns = ira_reg_equiv[regno].init_insns;
4581 if (insns == NULL)
4582 return false;
4583 for (; insns != NULL_RTX; insns = insns->next ())
4584 if (insn_rhs_dead_pseudo_p (insns->insn ()))
4585 return true;
4586 return false;
4589 /* Return TRUE if REGNO has a reverse equivalence. The equivalence is
4590 reverse only if we have one init insn with given REGNO as a
4591 source. */
4592 static bool
4593 reverse_equiv_p (int regno)
4595 rtx_insn_list *insns = ira_reg_equiv[regno].init_insns;
4596 rtx set;
4598 if (insns == NULL)
4599 return false;
4600 if (! INSN_P (insns->insn ())
4601 || insns->next () != NULL)
4602 return false;
4603 if ((set = single_set (insns->insn ())) == NULL_RTX)
4604 return false;
4605 return REG_P (SET_SRC (set)) && (int) REGNO (SET_SRC (set)) == regno;
4608 /* Return TRUE if REGNO was reloaded in an equivalence init insn. We
4609 call this function only for non-reverse equivalence. */
4610 static bool
4611 contains_reloaded_insn_p (int regno)
4613 rtx set;
4614 rtx_insn_list *list = ira_reg_equiv[regno].init_insns;
4616 for (; list != NULL; list = list->next ())
4617 if ((set = single_set (list->insn ())) == NULL_RTX
4618 || ! REG_P (SET_DEST (set))
4619 || (int) REGNO (SET_DEST (set)) != regno)
4620 return true;
4621 return false;
4624 /* Entry function of LRA constraint pass. Return true if the
4625 constraint pass did change the code. */
4626 bool
4627 lra_constraints (bool first_p)
4629 bool changed_p;
4630 int i, hard_regno, new_insns_num;
4631 unsigned int min_len, new_min_len, uid;
4632 rtx set, x, reg, dest_reg;
4633 basic_block last_bb;
4634 bitmap_iterator bi;
4636 lra_constraint_iter++;
4637 if (lra_dump_file != NULL)
4638 fprintf (lra_dump_file, "\n********** Local #%d: **********\n\n",
4639 lra_constraint_iter);
4640 changed_p = false;
4641 if (pic_offset_table_rtx
4642 && REGNO (pic_offset_table_rtx) >= FIRST_PSEUDO_REGISTER)
4643 lra_risky_transformations_p = true;
4644 else
4645 /* On the first iteration we should check IRA assignment
4646 correctness. In rare cases, the assignments can be wrong as
4647 early clobbers operands are ignored in IRA. */
4648 lra_risky_transformations_p = first_p;
4649 new_insn_uid_start = get_max_uid ();
4650 new_regno_start = first_p ? lra_constraint_new_regno_start : max_reg_num ();
4651 /* Mark used hard regs for target stack size calulations. */
4652 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4653 if (lra_reg_info[i].nrefs != 0
4654 && (hard_regno = lra_get_regno_hard_regno (i)) >= 0)
4656 int j, nregs;
4658 nregs = hard_regno_nregs (hard_regno, lra_reg_info[i].biggest_mode);
4659 for (j = 0; j < nregs; j++)
4660 df_set_regs_ever_live (hard_regno + j, true);
4662 /* Do elimination before the equivalence processing as we can spill
4663 some pseudos during elimination. */
4664 lra_eliminate (false, first_p);
4665 auto_bitmap equiv_insn_bitmap (&reg_obstack);
4666 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4667 if (lra_reg_info[i].nrefs != 0)
4669 ira_reg_equiv[i].profitable_p = true;
4670 reg = regno_reg_rtx[i];
4671 if (lra_get_regno_hard_regno (i) < 0 && (x = get_equiv (reg)) != reg)
4673 bool pseudo_p = contains_reg_p (x, false, false);
4675 /* After RTL transformation, we can not guarantee that
4676 pseudo in the substitution was not reloaded which might
4677 make equivalence invalid. For example, in reverse
4678 equiv of p0
4680 p0 <- ...
4682 equiv_mem <- p0
4684 the memory address register was reloaded before the 2nd
4685 insn. */
4686 if ((! first_p && pseudo_p)
4687 /* We don't use DF for compilation speed sake. So it
4688 is problematic to update live info when we use an
4689 equivalence containing pseudos in more than one
4690 BB. */
4691 || (pseudo_p && multi_block_pseudo_p (i))
4692 /* If an init insn was deleted for some reason, cancel
4693 the equiv. We could update the equiv insns after
4694 transformations including an equiv insn deletion
4695 but it is not worthy as such cases are extremely
4696 rare. */
4697 || contains_deleted_insn_p (ira_reg_equiv[i].init_insns)
4698 /* If it is not a reverse equivalence, we check that a
4699 pseudo in rhs of the init insn is not dying in the
4700 insn. Otherwise, the live info at the beginning of
4701 the corresponding BB might be wrong after we
4702 removed the insn. When the equiv can be a
4703 constant, the right hand side of the init insn can
4704 be a pseudo. */
4705 || (! reverse_equiv_p (i)
4706 && (init_insn_rhs_dead_pseudo_p (i)
4707 /* If we reloaded the pseudo in an equivalence
4708 init insn, we can not remove the equiv init
4709 insns and the init insns might write into
4710 const memory in this case. */
4711 || contains_reloaded_insn_p (i)))
4712 /* Prevent access beyond equivalent memory for
4713 paradoxical subregs. */
4714 || (MEM_P (x)
4715 && (GET_MODE_SIZE (lra_reg_info[i].biggest_mode)
4716 > GET_MODE_SIZE (GET_MODE (x))))
4717 || (pic_offset_table_rtx
4718 && ((CONST_POOL_OK_P (PSEUDO_REGNO_MODE (i), x)
4719 && (targetm.preferred_reload_class
4720 (x, lra_get_allocno_class (i)) == NO_REGS))
4721 || contains_symbol_ref_p (x))))
4722 ira_reg_equiv[i].defined_p = false;
4723 if (contains_reg_p (x, false, true))
4724 ira_reg_equiv[i].profitable_p = false;
4725 if (get_equiv (reg) != reg)
4726 bitmap_ior_into (equiv_insn_bitmap, &lra_reg_info[i].insn_bitmap);
4729 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4730 update_equiv (i);
4731 /* We should add all insns containing pseudos which should be
4732 substituted by their equivalences. */
4733 EXECUTE_IF_SET_IN_BITMAP (equiv_insn_bitmap, 0, uid, bi)
4734 lra_push_insn_by_uid (uid);
4735 min_len = lra_insn_stack_length ();
4736 new_insns_num = 0;
4737 last_bb = NULL;
4738 changed_p = false;
4739 while ((new_min_len = lra_insn_stack_length ()) != 0)
4741 curr_insn = lra_pop_insn ();
4742 --new_min_len;
4743 curr_bb = BLOCK_FOR_INSN (curr_insn);
4744 if (curr_bb != last_bb)
4746 last_bb = curr_bb;
4747 bb_reload_num = lra_curr_reload_num;
4749 if (min_len > new_min_len)
4751 min_len = new_min_len;
4752 new_insns_num = 0;
4754 if (new_insns_num > MAX_RELOAD_INSNS_NUMBER)
4755 internal_error
4756 ("Max. number of generated reload insns per insn is achieved (%d)\n",
4757 MAX_RELOAD_INSNS_NUMBER);
4758 new_insns_num++;
4759 if (DEBUG_INSN_P (curr_insn))
4761 /* We need to check equivalence in debug insn and change
4762 pseudo to the equivalent value if necessary. */
4763 curr_id = lra_get_insn_recog_data (curr_insn);
4764 if (bitmap_bit_p (equiv_insn_bitmap, INSN_UID (curr_insn)))
4766 rtx old = *curr_id->operand_loc[0];
4767 *curr_id->operand_loc[0]
4768 = simplify_replace_fn_rtx (old, NULL_RTX,
4769 loc_equivalence_callback, curr_insn);
4770 if (old != *curr_id->operand_loc[0])
4772 lra_update_insn_regno_info (curr_insn);
4773 changed_p = true;
4777 else if (INSN_P (curr_insn))
4779 if ((set = single_set (curr_insn)) != NULL_RTX)
4781 dest_reg = SET_DEST (set);
4782 /* The equivalence pseudo could be set up as SUBREG in a
4783 case when it is a call restore insn in a mode
4784 different from the pseudo mode. */
4785 if (GET_CODE (dest_reg) == SUBREG)
4786 dest_reg = SUBREG_REG (dest_reg);
4787 if ((REG_P (dest_reg)
4788 && (x = get_equiv (dest_reg)) != dest_reg
4789 /* Remove insns which set up a pseudo whose value
4790 can not be changed. Such insns might be not in
4791 init_insns because we don't update equiv data
4792 during insn transformations.
4794 As an example, let suppose that a pseudo got
4795 hard register and on the 1st pass was not
4796 changed to equivalent constant. We generate an
4797 additional insn setting up the pseudo because of
4798 secondary memory movement. Then the pseudo is
4799 spilled and we use the equiv constant. In this
4800 case we should remove the additional insn and
4801 this insn is not init_insns list. */
4802 && (! MEM_P (x) || MEM_READONLY_P (x)
4803 /* Check that this is actually an insn setting
4804 up the equivalence. */
4805 || in_list_p (curr_insn,
4806 ira_reg_equiv
4807 [REGNO (dest_reg)].init_insns)))
4808 || (((x = get_equiv (SET_SRC (set))) != SET_SRC (set))
4809 && in_list_p (curr_insn,
4810 ira_reg_equiv
4811 [REGNO (SET_SRC (set))].init_insns)))
4813 /* This is equiv init insn of pseudo which did not get a
4814 hard register -- remove the insn. */
4815 if (lra_dump_file != NULL)
4817 fprintf (lra_dump_file,
4818 " Removing equiv init insn %i (freq=%d)\n",
4819 INSN_UID (curr_insn),
4820 REG_FREQ_FROM_BB (BLOCK_FOR_INSN (curr_insn)));
4821 dump_insn_slim (lra_dump_file, curr_insn);
4823 if (contains_reg_p (x, true, false))
4824 lra_risky_transformations_p = true;
4825 lra_set_insn_deleted (curr_insn);
4826 continue;
4829 curr_id = lra_get_insn_recog_data (curr_insn);
4830 curr_static_id = curr_id->insn_static_data;
4831 init_curr_insn_input_reloads ();
4832 init_curr_operand_mode ();
4833 if (curr_insn_transform (false))
4834 changed_p = true;
4835 /* Check non-transformed insns too for equiv change as USE
4836 or CLOBBER don't need reloads but can contain pseudos
4837 being changed on their equivalences. */
4838 else if (bitmap_bit_p (equiv_insn_bitmap, INSN_UID (curr_insn))
4839 && loc_equivalence_change_p (&PATTERN (curr_insn)))
4841 lra_update_insn_regno_info (curr_insn);
4842 changed_p = true;
4847 /* If we used a new hard regno, changed_p should be true because the
4848 hard reg is assigned to a new pseudo. */
4849 if (flag_checking && !changed_p)
4851 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4852 if (lra_reg_info[i].nrefs != 0
4853 && (hard_regno = lra_get_regno_hard_regno (i)) >= 0)
4855 int j, nregs = hard_regno_nregs (hard_regno,
4856 PSEUDO_REGNO_MODE (i));
4858 for (j = 0; j < nregs; j++)
4859 lra_assert (df_regs_ever_live_p (hard_regno + j));
4862 return changed_p;
4865 static void initiate_invariants (void);
4866 static void finish_invariants (void);
4868 /* Initiate the LRA constraint pass. It is done once per
4869 function. */
4870 void
4871 lra_constraints_init (void)
4873 initiate_invariants ();
4876 /* Finalize the LRA constraint pass. It is done once per
4877 function. */
4878 void
4879 lra_constraints_finish (void)
4881 finish_invariants ();
4886 /* Structure describes invariants for ineheritance. */
4887 struct lra_invariant
4889 /* The order number of the invariant. */
4890 int num;
4891 /* The invariant RTX. */
4892 rtx invariant_rtx;
4893 /* The origin insn of the invariant. */
4894 rtx_insn *insn;
4897 typedef lra_invariant invariant_t;
4898 typedef invariant_t *invariant_ptr_t;
4899 typedef const invariant_t *const_invariant_ptr_t;
4901 /* Pointer to the inheritance invariants. */
4902 static vec<invariant_ptr_t> invariants;
4904 /* Allocation pool for the invariants. */
4905 static object_allocator<lra_invariant> *invariants_pool;
4907 /* Hash table for the invariants. */
4908 static htab_t invariant_table;
4910 /* Hash function for INVARIANT. */
4911 static hashval_t
4912 invariant_hash (const void *invariant)
4914 rtx inv = ((const_invariant_ptr_t) invariant)->invariant_rtx;
4915 return lra_rtx_hash (inv);
4918 /* Equal function for invariants INVARIANT1 and INVARIANT2. */
4919 static int
4920 invariant_eq_p (const void *invariant1, const void *invariant2)
4922 rtx inv1 = ((const_invariant_ptr_t) invariant1)->invariant_rtx;
4923 rtx inv2 = ((const_invariant_ptr_t) invariant2)->invariant_rtx;
4925 return rtx_equal_p (inv1, inv2);
4928 /* Insert INVARIANT_RTX into the table if it is not there yet. Return
4929 invariant which is in the table. */
4930 static invariant_ptr_t
4931 insert_invariant (rtx invariant_rtx)
4933 void **entry_ptr;
4934 invariant_t invariant;
4935 invariant_ptr_t invariant_ptr;
4937 invariant.invariant_rtx = invariant_rtx;
4938 entry_ptr = htab_find_slot (invariant_table, &invariant, INSERT);
4939 if (*entry_ptr == NULL)
4941 invariant_ptr = invariants_pool->allocate ();
4942 invariant_ptr->invariant_rtx = invariant_rtx;
4943 invariant_ptr->insn = NULL;
4944 invariants.safe_push (invariant_ptr);
4945 *entry_ptr = (void *) invariant_ptr;
4947 return (invariant_ptr_t) *entry_ptr;
4950 /* Initiate the invariant table. */
4951 static void
4952 initiate_invariants (void)
4954 invariants.create (100);
4955 invariants_pool
4956 = new object_allocator<lra_invariant> ("Inheritance invariants");
4957 invariant_table = htab_create (100, invariant_hash, invariant_eq_p, NULL);
4960 /* Finish the invariant table. */
4961 static void
4962 finish_invariants (void)
4964 htab_delete (invariant_table);
4965 delete invariants_pool;
4966 invariants.release ();
4969 /* Make the invariant table empty. */
4970 static void
4971 clear_invariants (void)
4973 htab_empty (invariant_table);
4974 invariants_pool->release ();
4975 invariants.truncate (0);
4980 /* This page contains code to do inheritance/split
4981 transformations. */
4983 /* Number of reloads passed so far in current EBB. */
4984 static int reloads_num;
4986 /* Number of calls passed so far in current EBB. */
4987 static int calls_num;
4989 /* Current reload pseudo check for validity of elements in
4990 USAGE_INSNS. */
4991 static int curr_usage_insns_check;
4993 /* Info about last usage of registers in EBB to do inheritance/split
4994 transformation. Inheritance transformation is done from a spilled
4995 pseudo and split transformations from a hard register or a pseudo
4996 assigned to a hard register. */
4997 struct usage_insns
4999 /* If the value is equal to CURR_USAGE_INSNS_CHECK, then the member
5000 value INSNS is valid. The insns is chain of optional debug insns
5001 and a finishing non-debug insn using the corresponding reg. The
5002 value is also used to mark the registers which are set up in the
5003 current insn. The negated insn uid is used for this. */
5004 int check;
5005 /* Value of global reloads_num at the last insn in INSNS. */
5006 int reloads_num;
5007 /* Value of global reloads_nums at the last insn in INSNS. */
5008 int calls_num;
5009 /* It can be true only for splitting. And it means that the restore
5010 insn should be put after insn given by the following member. */
5011 bool after_p;
5012 /* Next insns in the current EBB which use the original reg and the
5013 original reg value is not changed between the current insn and
5014 the next insns. In order words, e.g. for inheritance, if we need
5015 to use the original reg value again in the next insns we can try
5016 to use the value in a hard register from a reload insn of the
5017 current insn. */
5018 rtx insns;
5021 /* Map: regno -> corresponding pseudo usage insns. */
5022 static struct usage_insns *usage_insns;
5024 static void
5025 setup_next_usage_insn (int regno, rtx insn, int reloads_num, bool after_p)
5027 usage_insns[regno].check = curr_usage_insns_check;
5028 usage_insns[regno].insns = insn;
5029 usage_insns[regno].reloads_num = reloads_num;
5030 usage_insns[regno].calls_num = calls_num;
5031 usage_insns[regno].after_p = after_p;
5034 /* The function is used to form list REGNO usages which consists of
5035 optional debug insns finished by a non-debug insn using REGNO.
5036 RELOADS_NUM is current number of reload insns processed so far. */
5037 static void
5038 add_next_usage_insn (int regno, rtx_insn *insn, int reloads_num)
5040 rtx next_usage_insns;
5042 if (usage_insns[regno].check == curr_usage_insns_check
5043 && (next_usage_insns = usage_insns[regno].insns) != NULL_RTX
5044 && DEBUG_INSN_P (insn))
5046 /* Check that we did not add the debug insn yet. */
5047 if (next_usage_insns != insn
5048 && (GET_CODE (next_usage_insns) != INSN_LIST
5049 || XEXP (next_usage_insns, 0) != insn))
5050 usage_insns[regno].insns = gen_rtx_INSN_LIST (VOIDmode, insn,
5051 next_usage_insns);
5053 else if (NONDEBUG_INSN_P (insn))
5054 setup_next_usage_insn (regno, insn, reloads_num, false);
5055 else
5056 usage_insns[regno].check = 0;
5059 /* Return first non-debug insn in list USAGE_INSNS. */
5060 static rtx_insn *
5061 skip_usage_debug_insns (rtx usage_insns)
5063 rtx insn;
5065 /* Skip debug insns. */
5066 for (insn = usage_insns;
5067 insn != NULL_RTX && GET_CODE (insn) == INSN_LIST;
5068 insn = XEXP (insn, 1))
5070 return safe_as_a <rtx_insn *> (insn);
5073 /* Return true if we need secondary memory moves for insn in
5074 USAGE_INSNS after inserting inherited pseudo of class INHER_CL
5075 into the insn. */
5076 static bool
5077 check_secondary_memory_needed_p (enum reg_class inher_cl ATTRIBUTE_UNUSED,
5078 rtx usage_insns ATTRIBUTE_UNUSED)
5080 rtx_insn *insn;
5081 rtx set, dest;
5082 enum reg_class cl;
5084 if (inher_cl == ALL_REGS
5085 || (insn = skip_usage_debug_insns (usage_insns)) == NULL_RTX)
5086 return false;
5087 lra_assert (INSN_P (insn));
5088 if ((set = single_set (insn)) == NULL_RTX || ! REG_P (SET_DEST (set)))
5089 return false;
5090 dest = SET_DEST (set);
5091 if (! REG_P (dest))
5092 return false;
5093 lra_assert (inher_cl != NO_REGS);
5094 cl = get_reg_class (REGNO (dest));
5095 return (cl != NO_REGS && cl != ALL_REGS
5096 && targetm.secondary_memory_needed (GET_MODE (dest), inher_cl, cl));
5099 /* Registers involved in inheritance/split in the current EBB
5100 (inheritance/split pseudos and original registers). */
5101 static bitmap_head check_only_regs;
5103 /* Reload pseudos can not be involded in invariant inheritance in the
5104 current EBB. */
5105 static bitmap_head invalid_invariant_regs;
5107 /* Do inheritance transformations for insn INSN, which defines (if
5108 DEF_P) or uses ORIGINAL_REGNO. NEXT_USAGE_INSNS specifies which
5109 instruction in the EBB next uses ORIGINAL_REGNO; it has the same
5110 form as the "insns" field of usage_insns. Return true if we
5111 succeed in such transformation.
5113 The transformations look like:
5115 p <- ... i <- ...
5116 ... p <- i (new insn)
5117 ... =>
5118 <- ... p ... <- ... i ...
5120 ... i <- p (new insn)
5121 <- ... p ... <- ... i ...
5122 ... =>
5123 <- ... p ... <- ... i ...
5124 where p is a spilled original pseudo and i is a new inheritance pseudo.
5127 The inheritance pseudo has the smallest class of two classes CL and
5128 class of ORIGINAL REGNO. */
5129 static bool
5130 inherit_reload_reg (bool def_p, int original_regno,
5131 enum reg_class cl, rtx_insn *insn, rtx next_usage_insns)
5133 if (optimize_function_for_size_p (cfun))
5134 return false;
5136 enum reg_class rclass = lra_get_allocno_class (original_regno);
5137 rtx original_reg = regno_reg_rtx[original_regno];
5138 rtx new_reg, usage_insn;
5139 rtx_insn *new_insns;
5141 lra_assert (! usage_insns[original_regno].after_p);
5142 if (lra_dump_file != NULL)
5143 fprintf (lra_dump_file,
5144 " <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<\n");
5145 if (! ira_reg_classes_intersect_p[cl][rclass])
5147 if (lra_dump_file != NULL)
5149 fprintf (lra_dump_file,
5150 " Rejecting inheritance for %d "
5151 "because of disjoint classes %s and %s\n",
5152 original_regno, reg_class_names[cl],
5153 reg_class_names[rclass]);
5154 fprintf (lra_dump_file,
5155 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5157 return false;
5159 if ((ira_class_subset_p[cl][rclass] && cl != rclass)
5160 /* We don't use a subset of two classes because it can be
5161 NO_REGS. This transformation is still profitable in most
5162 cases even if the classes are not intersected as register
5163 move is probably cheaper than a memory load. */
5164 || ira_class_hard_regs_num[cl] < ira_class_hard_regs_num[rclass])
5166 if (lra_dump_file != NULL)
5167 fprintf (lra_dump_file, " Use smallest class of %s and %s\n",
5168 reg_class_names[cl], reg_class_names[rclass]);
5170 rclass = cl;
5172 if (check_secondary_memory_needed_p (rclass, next_usage_insns))
5174 /* Reject inheritance resulting in secondary memory moves.
5175 Otherwise, there is a danger in LRA cycling. Also such
5176 transformation will be unprofitable. */
5177 if (lra_dump_file != NULL)
5179 rtx_insn *insn = skip_usage_debug_insns (next_usage_insns);
5180 rtx set = single_set (insn);
5182 lra_assert (set != NULL_RTX);
5184 rtx dest = SET_DEST (set);
5186 lra_assert (REG_P (dest));
5187 fprintf (lra_dump_file,
5188 " Rejecting inheritance for insn %d(%s)<-%d(%s) "
5189 "as secondary mem is needed\n",
5190 REGNO (dest), reg_class_names[get_reg_class (REGNO (dest))],
5191 original_regno, reg_class_names[rclass]);
5192 fprintf (lra_dump_file,
5193 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5195 return false;
5197 new_reg = lra_create_new_reg (GET_MODE (original_reg), original_reg,
5198 rclass, "inheritance");
5199 start_sequence ();
5200 if (def_p)
5201 lra_emit_move (original_reg, new_reg);
5202 else
5203 lra_emit_move (new_reg, original_reg);
5204 new_insns = get_insns ();
5205 end_sequence ();
5206 if (NEXT_INSN (new_insns) != NULL_RTX)
5208 if (lra_dump_file != NULL)
5210 fprintf (lra_dump_file,
5211 " Rejecting inheritance %d->%d "
5212 "as it results in 2 or more insns:\n",
5213 original_regno, REGNO (new_reg));
5214 dump_rtl_slim (lra_dump_file, new_insns, NULL, -1, 0);
5215 fprintf (lra_dump_file,
5216 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5218 return false;
5220 lra_substitute_pseudo_within_insn (insn, original_regno, new_reg, false);
5221 lra_update_insn_regno_info (insn);
5222 if (! def_p)
5223 /* We now have a new usage insn for original regno. */
5224 setup_next_usage_insn (original_regno, new_insns, reloads_num, false);
5225 if (lra_dump_file != NULL)
5226 fprintf (lra_dump_file, " Original reg change %d->%d (bb%d):\n",
5227 original_regno, REGNO (new_reg), BLOCK_FOR_INSN (insn)->index);
5228 lra_reg_info[REGNO (new_reg)].restore_rtx = regno_reg_rtx[original_regno];
5229 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
5230 bitmap_set_bit (&check_only_regs, original_regno);
5231 bitmap_set_bit (&lra_inheritance_pseudos, REGNO (new_reg));
5232 if (def_p)
5233 lra_process_new_insns (insn, NULL, new_insns,
5234 "Add original<-inheritance");
5235 else
5236 lra_process_new_insns (insn, new_insns, NULL,
5237 "Add inheritance<-original");
5238 while (next_usage_insns != NULL_RTX)
5240 if (GET_CODE (next_usage_insns) != INSN_LIST)
5242 usage_insn = next_usage_insns;
5243 lra_assert (NONDEBUG_INSN_P (usage_insn));
5244 next_usage_insns = NULL;
5246 else
5248 usage_insn = XEXP (next_usage_insns, 0);
5249 lra_assert (DEBUG_INSN_P (usage_insn));
5250 next_usage_insns = XEXP (next_usage_insns, 1);
5252 lra_substitute_pseudo (&usage_insn, original_regno, new_reg, false);
5253 lra_update_insn_regno_info (as_a <rtx_insn *> (usage_insn));
5254 if (lra_dump_file != NULL)
5256 fprintf (lra_dump_file,
5257 " Inheritance reuse change %d->%d (bb%d):\n",
5258 original_regno, REGNO (new_reg),
5259 BLOCK_FOR_INSN (usage_insn)->index);
5260 dump_insn_slim (lra_dump_file, as_a <rtx_insn *> (usage_insn));
5263 if (lra_dump_file != NULL)
5264 fprintf (lra_dump_file,
5265 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5266 return true;
5269 /* Return true if we need a caller save/restore for pseudo REGNO which
5270 was assigned to a hard register. */
5271 static inline bool
5272 need_for_call_save_p (int regno)
5274 lra_assert (regno >= FIRST_PSEUDO_REGISTER && reg_renumber[regno] >= 0);
5275 return (usage_insns[regno].calls_num < calls_num
5276 && (overlaps_hard_reg_set_p
5277 ((flag_ipa_ra &&
5278 ! hard_reg_set_empty_p (lra_reg_info[regno].actual_call_used_reg_set))
5279 ? lra_reg_info[regno].actual_call_used_reg_set
5280 : call_used_reg_set,
5281 PSEUDO_REGNO_MODE (regno), reg_renumber[regno])
5282 || (targetm.hard_regno_call_part_clobbered
5283 (reg_renumber[regno], PSEUDO_REGNO_MODE (regno)))));
5286 /* Global registers occurring in the current EBB. */
5287 static bitmap_head ebb_global_regs;
5289 /* Return true if we need a split for hard register REGNO or pseudo
5290 REGNO which was assigned to a hard register.
5291 POTENTIAL_RELOAD_HARD_REGS contains hard registers which might be
5292 used for reloads since the EBB end. It is an approximation of the
5293 used hard registers in the split range. The exact value would
5294 require expensive calculations. If we were aggressive with
5295 splitting because of the approximation, the split pseudo will save
5296 the same hard register assignment and will be removed in the undo
5297 pass. We still need the approximation because too aggressive
5298 splitting would result in too inaccurate cost calculation in the
5299 assignment pass because of too many generated moves which will be
5300 probably removed in the undo pass. */
5301 static inline bool
5302 need_for_split_p (HARD_REG_SET potential_reload_hard_regs, int regno)
5304 int hard_regno = regno < FIRST_PSEUDO_REGISTER ? regno : reg_renumber[regno];
5306 lra_assert (hard_regno >= 0);
5307 return ((TEST_HARD_REG_BIT (potential_reload_hard_regs, hard_regno)
5308 /* Don't split eliminable hard registers, otherwise we can
5309 split hard registers like hard frame pointer, which
5310 lives on BB start/end according to DF-infrastructure,
5311 when there is a pseudo assigned to the register and
5312 living in the same BB. */
5313 && (regno >= FIRST_PSEUDO_REGISTER
5314 || ! TEST_HARD_REG_BIT (eliminable_regset, hard_regno))
5315 && ! TEST_HARD_REG_BIT (lra_no_alloc_regs, hard_regno)
5316 /* Don't split call clobbered hard regs living through
5317 calls, otherwise we might have a check problem in the
5318 assign sub-pass as in the most cases (exception is a
5319 situation when lra_risky_transformations_p value is
5320 true) the assign pass assumes that all pseudos living
5321 through calls are assigned to call saved hard regs. */
5322 && (regno >= FIRST_PSEUDO_REGISTER
5323 || ! TEST_HARD_REG_BIT (call_used_reg_set, regno)
5324 || usage_insns[regno].calls_num == calls_num)
5325 /* We need at least 2 reloads to make pseudo splitting
5326 profitable. We should provide hard regno splitting in
5327 any case to solve 1st insn scheduling problem when
5328 moving hard register definition up might result in
5329 impossibility to find hard register for reload pseudo of
5330 small register class. */
5331 && (usage_insns[regno].reloads_num
5332 + (regno < FIRST_PSEUDO_REGISTER ? 0 : 3) < reloads_num)
5333 && (regno < FIRST_PSEUDO_REGISTER
5334 /* For short living pseudos, spilling + inheritance can
5335 be considered a substitution for splitting.
5336 Therefore we do not splitting for local pseudos. It
5337 decreases also aggressiveness of splitting. The
5338 minimal number of references is chosen taking into
5339 account that for 2 references splitting has no sense
5340 as we can just spill the pseudo. */
5341 || (regno >= FIRST_PSEUDO_REGISTER
5342 && lra_reg_info[regno].nrefs > 3
5343 && bitmap_bit_p (&ebb_global_regs, regno))))
5344 || (regno >= FIRST_PSEUDO_REGISTER && need_for_call_save_p (regno)));
5347 /* Return class for the split pseudo created from original pseudo with
5348 ALLOCNO_CLASS and MODE which got a hard register HARD_REGNO. We
5349 choose subclass of ALLOCNO_CLASS which contains HARD_REGNO and
5350 results in no secondary memory movements. */
5351 static enum reg_class
5352 choose_split_class (enum reg_class allocno_class,
5353 int hard_regno ATTRIBUTE_UNUSED,
5354 machine_mode mode ATTRIBUTE_UNUSED)
5356 int i;
5357 enum reg_class cl, best_cl = NO_REGS;
5358 enum reg_class hard_reg_class ATTRIBUTE_UNUSED
5359 = REGNO_REG_CLASS (hard_regno);
5361 if (! targetm.secondary_memory_needed (mode, allocno_class, allocno_class)
5362 && TEST_HARD_REG_BIT (reg_class_contents[allocno_class], hard_regno))
5363 return allocno_class;
5364 for (i = 0;
5365 (cl = reg_class_subclasses[allocno_class][i]) != LIM_REG_CLASSES;
5366 i++)
5367 if (! targetm.secondary_memory_needed (mode, cl, hard_reg_class)
5368 && ! targetm.secondary_memory_needed (mode, hard_reg_class, cl)
5369 && TEST_HARD_REG_BIT (reg_class_contents[cl], hard_regno)
5370 && (best_cl == NO_REGS
5371 || ira_class_hard_regs_num[best_cl] < ira_class_hard_regs_num[cl]))
5372 best_cl = cl;
5373 return best_cl;
5376 /* Copy any equivalence information from ORIGINAL_REGNO to NEW_REGNO.
5377 It only makes sense to call this function if NEW_REGNO is always
5378 equal to ORIGINAL_REGNO. */
5380 static void
5381 lra_copy_reg_equiv (unsigned int new_regno, unsigned int original_regno)
5383 if (!ira_reg_equiv[original_regno].defined_p)
5384 return;
5386 ira_expand_reg_equiv ();
5387 ira_reg_equiv[new_regno].defined_p = true;
5388 if (ira_reg_equiv[original_regno].memory)
5389 ira_reg_equiv[new_regno].memory
5390 = copy_rtx (ira_reg_equiv[original_regno].memory);
5391 if (ira_reg_equiv[original_regno].constant)
5392 ira_reg_equiv[new_regno].constant
5393 = copy_rtx (ira_reg_equiv[original_regno].constant);
5394 if (ira_reg_equiv[original_regno].invariant)
5395 ira_reg_equiv[new_regno].invariant
5396 = copy_rtx (ira_reg_equiv[original_regno].invariant);
5399 /* Do split transformations for insn INSN, which defines or uses
5400 ORIGINAL_REGNO. NEXT_USAGE_INSNS specifies which instruction in
5401 the EBB next uses ORIGINAL_REGNO; it has the same form as the
5402 "insns" field of usage_insns.
5404 The transformations look like:
5406 p <- ... p <- ...
5407 ... s <- p (new insn -- save)
5408 ... =>
5409 ... p <- s (new insn -- restore)
5410 <- ... p ... <- ... p ...
5412 <- ... p ... <- ... p ...
5413 ... s <- p (new insn -- save)
5414 ... =>
5415 ... p <- s (new insn -- restore)
5416 <- ... p ... <- ... p ...
5418 where p is an original pseudo got a hard register or a hard
5419 register and s is a new split pseudo. The save is put before INSN
5420 if BEFORE_P is true. Return true if we succeed in such
5421 transformation. */
5422 static bool
5423 split_reg (bool before_p, int original_regno, rtx_insn *insn,
5424 rtx next_usage_insns)
5426 enum reg_class rclass;
5427 rtx original_reg;
5428 int hard_regno, nregs;
5429 rtx new_reg, usage_insn;
5430 rtx_insn *restore, *save;
5431 bool after_p;
5432 bool call_save_p;
5433 machine_mode mode;
5435 if (original_regno < FIRST_PSEUDO_REGISTER)
5437 rclass = ira_allocno_class_translate[REGNO_REG_CLASS (original_regno)];
5438 hard_regno = original_regno;
5439 call_save_p = false;
5440 nregs = 1;
5441 mode = lra_reg_info[hard_regno].biggest_mode;
5442 machine_mode reg_rtx_mode = GET_MODE (regno_reg_rtx[hard_regno]);
5443 /* A reg can have a biggest_mode of VOIDmode if it was only ever seen
5444 as part of a multi-word register. In that case, or if the biggest
5445 mode was larger than a register, just use the reg_rtx. Otherwise,
5446 limit the size to that of the biggest access in the function. */
5447 if (mode == VOIDmode
5448 || paradoxical_subreg_p (mode, reg_rtx_mode))
5450 original_reg = regno_reg_rtx[hard_regno];
5451 mode = reg_rtx_mode;
5453 else
5454 original_reg = gen_rtx_REG (mode, hard_regno);
5456 else
5458 mode = PSEUDO_REGNO_MODE (original_regno);
5459 hard_regno = reg_renumber[original_regno];
5460 nregs = hard_regno_nregs (hard_regno, mode);
5461 rclass = lra_get_allocno_class (original_regno);
5462 original_reg = regno_reg_rtx[original_regno];
5463 call_save_p = need_for_call_save_p (original_regno);
5465 lra_assert (hard_regno >= 0);
5466 if (lra_dump_file != NULL)
5467 fprintf (lra_dump_file,
5468 " ((((((((((((((((((((((((((((((((((((((((((((((((\n");
5470 if (call_save_p)
5472 mode = HARD_REGNO_CALLER_SAVE_MODE (hard_regno,
5473 hard_regno_nregs (hard_regno, mode),
5474 mode);
5475 new_reg = lra_create_new_reg (mode, NULL_RTX, NO_REGS, "save");
5477 else
5479 rclass = choose_split_class (rclass, hard_regno, mode);
5480 if (rclass == NO_REGS)
5482 if (lra_dump_file != NULL)
5484 fprintf (lra_dump_file,
5485 " Rejecting split of %d(%s): "
5486 "no good reg class for %d(%s)\n",
5487 original_regno,
5488 reg_class_names[lra_get_allocno_class (original_regno)],
5489 hard_regno,
5490 reg_class_names[REGNO_REG_CLASS (hard_regno)]);
5491 fprintf
5492 (lra_dump_file,
5493 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5495 return false;
5497 /* Split_if_necessary can split hard registers used as part of a
5498 multi-register mode but splits each register individually. The
5499 mode used for each independent register may not be supported
5500 so reject the split. Splitting the wider mode should theoretically
5501 be possible but is not implemented. */
5502 if (!targetm.hard_regno_mode_ok (hard_regno, mode))
5504 if (lra_dump_file != NULL)
5506 fprintf (lra_dump_file,
5507 " Rejecting split of %d(%s): unsuitable mode %s\n",
5508 original_regno,
5509 reg_class_names[lra_get_allocno_class (original_regno)],
5510 GET_MODE_NAME (mode));
5511 fprintf
5512 (lra_dump_file,
5513 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5515 return false;
5517 new_reg = lra_create_new_reg (mode, original_reg, rclass, "split");
5518 reg_renumber[REGNO (new_reg)] = hard_regno;
5520 int new_regno = REGNO (new_reg);
5521 save = emit_spill_move (true, new_reg, original_reg);
5522 if (NEXT_INSN (save) != NULL_RTX && !call_save_p)
5524 if (lra_dump_file != NULL)
5526 fprintf
5527 (lra_dump_file,
5528 " Rejecting split %d->%d resulting in > 2 save insns:\n",
5529 original_regno, new_regno);
5530 dump_rtl_slim (lra_dump_file, save, NULL, -1, 0);
5531 fprintf (lra_dump_file,
5532 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5534 return false;
5536 restore = emit_spill_move (false, new_reg, original_reg);
5537 if (NEXT_INSN (restore) != NULL_RTX && !call_save_p)
5539 if (lra_dump_file != NULL)
5541 fprintf (lra_dump_file,
5542 " Rejecting split %d->%d "
5543 "resulting in > 2 restore insns:\n",
5544 original_regno, new_regno);
5545 dump_rtl_slim (lra_dump_file, restore, NULL, -1, 0);
5546 fprintf (lra_dump_file,
5547 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5549 return false;
5551 /* Transfer equivalence information to the spill register, so that
5552 if we fail to allocate the spill register, we have the option of
5553 rematerializing the original value instead of spilling to the stack. */
5554 if (!HARD_REGISTER_NUM_P (original_regno)
5555 && mode == PSEUDO_REGNO_MODE (original_regno))
5556 lra_copy_reg_equiv (new_regno, original_regno);
5557 after_p = usage_insns[original_regno].after_p;
5558 lra_reg_info[new_regno].restore_rtx = regno_reg_rtx[original_regno];
5559 bitmap_set_bit (&check_only_regs, new_regno);
5560 bitmap_set_bit (&check_only_regs, original_regno);
5561 bitmap_set_bit (&lra_split_regs, new_regno);
5562 for (;;)
5564 if (GET_CODE (next_usage_insns) != INSN_LIST)
5566 usage_insn = next_usage_insns;
5567 break;
5569 usage_insn = XEXP (next_usage_insns, 0);
5570 lra_assert (DEBUG_INSN_P (usage_insn));
5571 next_usage_insns = XEXP (next_usage_insns, 1);
5572 lra_substitute_pseudo (&usage_insn, original_regno, new_reg, false);
5573 lra_update_insn_regno_info (as_a <rtx_insn *> (usage_insn));
5574 if (lra_dump_file != NULL)
5576 fprintf (lra_dump_file, " Split reuse change %d->%d:\n",
5577 original_regno, new_regno);
5578 dump_insn_slim (lra_dump_file, as_a <rtx_insn *> (usage_insn));
5581 lra_assert (NOTE_P (usage_insn) || NONDEBUG_INSN_P (usage_insn));
5582 lra_assert (usage_insn != insn || (after_p && before_p));
5583 lra_process_new_insns (as_a <rtx_insn *> (usage_insn),
5584 after_p ? NULL : restore,
5585 after_p ? restore : NULL,
5586 call_save_p
5587 ? "Add reg<-save" : "Add reg<-split");
5588 lra_process_new_insns (insn, before_p ? save : NULL,
5589 before_p ? NULL : save,
5590 call_save_p
5591 ? "Add save<-reg" : "Add split<-reg");
5592 if (nregs > 1)
5593 /* If we are trying to split multi-register. We should check
5594 conflicts on the next assignment sub-pass. IRA can allocate on
5595 sub-register levels, LRA do this on pseudos level right now and
5596 this discrepancy may create allocation conflicts after
5597 splitting. */
5598 lra_risky_transformations_p = true;
5599 if (lra_dump_file != NULL)
5600 fprintf (lra_dump_file,
5601 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5602 return true;
5605 /* Recognize that we need a split transformation for insn INSN, which
5606 defines or uses REGNO in its insn biggest MODE (we use it only if
5607 REGNO is a hard register). POTENTIAL_RELOAD_HARD_REGS contains
5608 hard registers which might be used for reloads since the EBB end.
5609 Put the save before INSN if BEFORE_P is true. MAX_UID is maximla
5610 uid before starting INSN processing. Return true if we succeed in
5611 such transformation. */
5612 static bool
5613 split_if_necessary (int regno, machine_mode mode,
5614 HARD_REG_SET potential_reload_hard_regs,
5615 bool before_p, rtx_insn *insn, int max_uid)
5617 bool res = false;
5618 int i, nregs = 1;
5619 rtx next_usage_insns;
5621 if (regno < FIRST_PSEUDO_REGISTER)
5622 nregs = hard_regno_nregs (regno, mode);
5623 for (i = 0; i < nregs; i++)
5624 if (usage_insns[regno + i].check == curr_usage_insns_check
5625 && (next_usage_insns = usage_insns[regno + i].insns) != NULL_RTX
5626 /* To avoid processing the register twice or more. */
5627 && ((GET_CODE (next_usage_insns) != INSN_LIST
5628 && INSN_UID (next_usage_insns) < max_uid)
5629 || (GET_CODE (next_usage_insns) == INSN_LIST
5630 && (INSN_UID (XEXP (next_usage_insns, 0)) < max_uid)))
5631 && need_for_split_p (potential_reload_hard_regs, regno + i)
5632 && split_reg (before_p, regno + i, insn, next_usage_insns))
5633 res = true;
5634 return res;
5637 /* Return TRUE if rtx X is considered as an invariant for
5638 inheritance. */
5639 static bool
5640 invariant_p (const_rtx x)
5642 machine_mode mode;
5643 const char *fmt;
5644 enum rtx_code code;
5645 int i, j;
5647 code = GET_CODE (x);
5648 mode = GET_MODE (x);
5649 if (code == SUBREG)
5651 x = SUBREG_REG (x);
5652 code = GET_CODE (x);
5653 if (GET_MODE_SIZE (GET_MODE (x)) > GET_MODE_SIZE (mode))
5654 mode = GET_MODE (x);
5657 if (MEM_P (x))
5658 return false;
5660 if (REG_P (x))
5662 int i, nregs, regno = REGNO (x);
5664 if (regno >= FIRST_PSEUDO_REGISTER || regno == STACK_POINTER_REGNUM
5665 || TEST_HARD_REG_BIT (eliminable_regset, regno)
5666 || GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
5667 return false;
5668 nregs = hard_regno_nregs (regno, mode);
5669 for (i = 0; i < nregs; i++)
5670 if (! fixed_regs[regno + i]
5671 /* A hard register may be clobbered in the current insn
5672 but we can ignore this case because if the hard
5673 register is used it should be set somewhere after the
5674 clobber. */
5675 || bitmap_bit_p (&invalid_invariant_regs, regno + i))
5676 return false;
5678 fmt = GET_RTX_FORMAT (code);
5679 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
5681 if (fmt[i] == 'e')
5683 if (! invariant_p (XEXP (x, i)))
5684 return false;
5686 else if (fmt[i] == 'E')
5688 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
5689 if (! invariant_p (XVECEXP (x, i, j)))
5690 return false;
5693 return true;
5696 /* We have 'dest_reg <- invariant'. Let us try to make an invariant
5697 inheritance transformation (using dest_reg instead invariant in a
5698 subsequent insn). */
5699 static bool
5700 process_invariant_for_inheritance (rtx dst_reg, rtx invariant_rtx)
5702 invariant_ptr_t invariant_ptr;
5703 rtx_insn *insn, *new_insns;
5704 rtx insn_set, insn_reg, new_reg;
5705 int insn_regno;
5706 bool succ_p = false;
5707 int dst_regno = REGNO (dst_reg);
5708 machine_mode dst_mode = GET_MODE (dst_reg);
5709 enum reg_class cl = lra_get_allocno_class (dst_regno), insn_reg_cl;
5711 invariant_ptr = insert_invariant (invariant_rtx);
5712 if ((insn = invariant_ptr->insn) != NULL_RTX)
5714 /* We have a subsequent insn using the invariant. */
5715 insn_set = single_set (insn);
5716 lra_assert (insn_set != NULL);
5717 insn_reg = SET_DEST (insn_set);
5718 lra_assert (REG_P (insn_reg));
5719 insn_regno = REGNO (insn_reg);
5720 insn_reg_cl = lra_get_allocno_class (insn_regno);
5722 if (dst_mode == GET_MODE (insn_reg)
5723 /* We should consider only result move reg insns which are
5724 cheap. */
5725 && targetm.register_move_cost (dst_mode, cl, insn_reg_cl) == 2
5726 && targetm.register_move_cost (dst_mode, cl, cl) == 2)
5728 if (lra_dump_file != NULL)
5729 fprintf (lra_dump_file,
5730 " [[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[\n");
5731 new_reg = lra_create_new_reg (dst_mode, dst_reg,
5732 cl, "invariant inheritance");
5733 bitmap_set_bit (&lra_inheritance_pseudos, REGNO (new_reg));
5734 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
5735 lra_reg_info[REGNO (new_reg)].restore_rtx = PATTERN (insn);
5736 start_sequence ();
5737 lra_emit_move (new_reg, dst_reg);
5738 new_insns = get_insns ();
5739 end_sequence ();
5740 lra_process_new_insns (curr_insn, NULL, new_insns,
5741 "Add invariant inheritance<-original");
5742 start_sequence ();
5743 lra_emit_move (SET_DEST (insn_set), new_reg);
5744 new_insns = get_insns ();
5745 end_sequence ();
5746 lra_process_new_insns (insn, NULL, new_insns,
5747 "Changing reload<-inheritance");
5748 lra_set_insn_deleted (insn);
5749 succ_p = true;
5750 if (lra_dump_file != NULL)
5752 fprintf (lra_dump_file,
5753 " Invariant inheritance reuse change %d (bb%d):\n",
5754 REGNO (new_reg), BLOCK_FOR_INSN (insn)->index);
5755 dump_insn_slim (lra_dump_file, insn);
5756 fprintf (lra_dump_file,
5757 " ]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]\n");
5761 invariant_ptr->insn = curr_insn;
5762 return succ_p;
5765 /* Check only registers living at the current program point in the
5766 current EBB. */
5767 static bitmap_head live_regs;
5769 /* Update live info in EBB given by its HEAD and TAIL insns after
5770 inheritance/split transformation. The function removes dead moves
5771 too. */
5772 static void
5773 update_ebb_live_info (rtx_insn *head, rtx_insn *tail)
5775 unsigned int j;
5776 int i, regno;
5777 bool live_p;
5778 rtx_insn *prev_insn;
5779 rtx set;
5780 bool remove_p;
5781 basic_block last_bb, prev_bb, curr_bb;
5782 bitmap_iterator bi;
5783 struct lra_insn_reg *reg;
5784 edge e;
5785 edge_iterator ei;
5787 last_bb = BLOCK_FOR_INSN (tail);
5788 prev_bb = NULL;
5789 for (curr_insn = tail;
5790 curr_insn != PREV_INSN (head);
5791 curr_insn = prev_insn)
5793 prev_insn = PREV_INSN (curr_insn);
5794 /* We need to process empty blocks too. They contain
5795 NOTE_INSN_BASIC_BLOCK referring for the basic block. */
5796 if (NOTE_P (curr_insn) && NOTE_KIND (curr_insn) != NOTE_INSN_BASIC_BLOCK)
5797 continue;
5798 curr_bb = BLOCK_FOR_INSN (curr_insn);
5799 if (curr_bb != prev_bb)
5801 if (prev_bb != NULL)
5803 /* Update df_get_live_in (prev_bb): */
5804 EXECUTE_IF_SET_IN_BITMAP (&check_only_regs, 0, j, bi)
5805 if (bitmap_bit_p (&live_regs, j))
5806 bitmap_set_bit (df_get_live_in (prev_bb), j);
5807 else
5808 bitmap_clear_bit (df_get_live_in (prev_bb), j);
5810 if (curr_bb != last_bb)
5812 /* Update df_get_live_out (curr_bb): */
5813 EXECUTE_IF_SET_IN_BITMAP (&check_only_regs, 0, j, bi)
5815 live_p = bitmap_bit_p (&live_regs, j);
5816 if (! live_p)
5817 FOR_EACH_EDGE (e, ei, curr_bb->succs)
5818 if (bitmap_bit_p (df_get_live_in (e->dest), j))
5820 live_p = true;
5821 break;
5823 if (live_p)
5824 bitmap_set_bit (df_get_live_out (curr_bb), j);
5825 else
5826 bitmap_clear_bit (df_get_live_out (curr_bb), j);
5829 prev_bb = curr_bb;
5830 bitmap_and (&live_regs, &check_only_regs, df_get_live_out (curr_bb));
5832 if (! NONDEBUG_INSN_P (curr_insn))
5833 continue;
5834 curr_id = lra_get_insn_recog_data (curr_insn);
5835 curr_static_id = curr_id->insn_static_data;
5836 remove_p = false;
5837 if ((set = single_set (curr_insn)) != NULL_RTX
5838 && REG_P (SET_DEST (set))
5839 && (regno = REGNO (SET_DEST (set))) >= FIRST_PSEUDO_REGISTER
5840 && SET_DEST (set) != pic_offset_table_rtx
5841 && bitmap_bit_p (&check_only_regs, regno)
5842 && ! bitmap_bit_p (&live_regs, regno))
5843 remove_p = true;
5844 /* See which defined values die here. */
5845 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5846 if (reg->type == OP_OUT && ! reg->subreg_p)
5847 bitmap_clear_bit (&live_regs, reg->regno);
5848 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
5849 if (reg->type == OP_OUT && ! reg->subreg_p)
5850 bitmap_clear_bit (&live_regs, reg->regno);
5851 if (curr_id->arg_hard_regs != NULL)
5852 /* Make clobbered argument hard registers die. */
5853 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
5854 if (regno >= FIRST_PSEUDO_REGISTER)
5855 bitmap_clear_bit (&live_regs, regno - FIRST_PSEUDO_REGISTER);
5856 /* Mark each used value as live. */
5857 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5858 if (reg->type != OP_OUT
5859 && bitmap_bit_p (&check_only_regs, reg->regno))
5860 bitmap_set_bit (&live_regs, reg->regno);
5861 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
5862 if (reg->type != OP_OUT
5863 && bitmap_bit_p (&check_only_regs, reg->regno))
5864 bitmap_set_bit (&live_regs, reg->regno);
5865 if (curr_id->arg_hard_regs != NULL)
5866 /* Make used argument hard registers live. */
5867 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
5868 if (regno < FIRST_PSEUDO_REGISTER
5869 && bitmap_bit_p (&check_only_regs, regno))
5870 bitmap_set_bit (&live_regs, regno);
5871 /* It is quite important to remove dead move insns because it
5872 means removing dead store. We don't need to process them for
5873 constraints. */
5874 if (remove_p)
5876 if (lra_dump_file != NULL)
5878 fprintf (lra_dump_file, " Removing dead insn:\n ");
5879 dump_insn_slim (lra_dump_file, curr_insn);
5881 lra_set_insn_deleted (curr_insn);
5886 /* The structure describes info to do an inheritance for the current
5887 insn. We need to collect such info first before doing the
5888 transformations because the transformations change the insn
5889 internal representation. */
5890 struct to_inherit
5892 /* Original regno. */
5893 int regno;
5894 /* Subsequent insns which can inherit original reg value. */
5895 rtx insns;
5898 /* Array containing all info for doing inheritance from the current
5899 insn. */
5900 static struct to_inherit to_inherit[LRA_MAX_INSN_RELOADS];
5902 /* Number elements in the previous array. */
5903 static int to_inherit_num;
5905 /* Add inheritance info REGNO and INSNS. Their meaning is described in
5906 structure to_inherit. */
5907 static void
5908 add_to_inherit (int regno, rtx insns)
5910 int i;
5912 for (i = 0; i < to_inherit_num; i++)
5913 if (to_inherit[i].regno == regno)
5914 return;
5915 lra_assert (to_inherit_num < LRA_MAX_INSN_RELOADS);
5916 to_inherit[to_inherit_num].regno = regno;
5917 to_inherit[to_inherit_num++].insns = insns;
5920 /* Return the last non-debug insn in basic block BB, or the block begin
5921 note if none. */
5922 static rtx_insn *
5923 get_last_insertion_point (basic_block bb)
5925 rtx_insn *insn;
5927 FOR_BB_INSNS_REVERSE (bb, insn)
5928 if (NONDEBUG_INSN_P (insn) || NOTE_INSN_BASIC_BLOCK_P (insn))
5929 return insn;
5930 gcc_unreachable ();
5933 /* Set up RES by registers living on edges FROM except the edge (FROM,
5934 TO) or by registers set up in a jump insn in BB FROM. */
5935 static void
5936 get_live_on_other_edges (basic_block from, basic_block to, bitmap res)
5938 rtx_insn *last;
5939 struct lra_insn_reg *reg;
5940 edge e;
5941 edge_iterator ei;
5943 lra_assert (to != NULL);
5944 bitmap_clear (res);
5945 FOR_EACH_EDGE (e, ei, from->succs)
5946 if (e->dest != to)
5947 bitmap_ior_into (res, df_get_live_in (e->dest));
5948 last = get_last_insertion_point (from);
5949 if (! JUMP_P (last))
5950 return;
5951 curr_id = lra_get_insn_recog_data (last);
5952 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5953 if (reg->type != OP_IN)
5954 bitmap_set_bit (res, reg->regno);
5957 /* Used as a temporary results of some bitmap calculations. */
5958 static bitmap_head temp_bitmap;
5960 /* We split for reloads of small class of hard regs. The following
5961 defines how many hard regs the class should have to be qualified as
5962 small. The code is mostly oriented to x86/x86-64 architecture
5963 where some insns need to use only specific register or pair of
5964 registers and these register can live in RTL explicitly, e.g. for
5965 parameter passing. */
5966 static const int max_small_class_regs_num = 2;
5968 /* Do inheritance/split transformations in EBB starting with HEAD and
5969 finishing on TAIL. We process EBB insns in the reverse order.
5970 Return true if we did any inheritance/split transformation in the
5971 EBB.
5973 We should avoid excessive splitting which results in worse code
5974 because of inaccurate cost calculations for spilling new split
5975 pseudos in such case. To achieve this we do splitting only if
5976 register pressure is high in given basic block and there are reload
5977 pseudos requiring hard registers. We could do more register
5978 pressure calculations at any given program point to avoid necessary
5979 splitting even more but it is to expensive and the current approach
5980 works well enough. */
5981 static bool
5982 inherit_in_ebb (rtx_insn *head, rtx_insn *tail)
5984 int i, src_regno, dst_regno, nregs;
5985 bool change_p, succ_p, update_reloads_num_p;
5986 rtx_insn *prev_insn, *last_insn;
5987 rtx next_usage_insns, curr_set;
5988 enum reg_class cl;
5989 struct lra_insn_reg *reg;
5990 basic_block last_processed_bb, curr_bb = NULL;
5991 HARD_REG_SET potential_reload_hard_regs, live_hard_regs;
5992 bitmap to_process;
5993 unsigned int j;
5994 bitmap_iterator bi;
5995 bool head_p, after_p;
5997 change_p = false;
5998 curr_usage_insns_check++;
5999 clear_invariants ();
6000 reloads_num = calls_num = 0;
6001 bitmap_clear (&check_only_regs);
6002 bitmap_clear (&invalid_invariant_regs);
6003 last_processed_bb = NULL;
6004 CLEAR_HARD_REG_SET (potential_reload_hard_regs);
6005 COPY_HARD_REG_SET (live_hard_regs, eliminable_regset);
6006 IOR_HARD_REG_SET (live_hard_regs, lra_no_alloc_regs);
6007 /* We don't process new insns generated in the loop. */
6008 for (curr_insn = tail; curr_insn != PREV_INSN (head); curr_insn = prev_insn)
6010 prev_insn = PREV_INSN (curr_insn);
6011 if (BLOCK_FOR_INSN (curr_insn) != NULL)
6012 curr_bb = BLOCK_FOR_INSN (curr_insn);
6013 if (last_processed_bb != curr_bb)
6015 /* We are at the end of BB. Add qualified living
6016 pseudos for potential splitting. */
6017 to_process = df_get_live_out (curr_bb);
6018 if (last_processed_bb != NULL)
6020 /* We are somewhere in the middle of EBB. */
6021 get_live_on_other_edges (curr_bb, last_processed_bb,
6022 &temp_bitmap);
6023 to_process = &temp_bitmap;
6025 last_processed_bb = curr_bb;
6026 last_insn = get_last_insertion_point (curr_bb);
6027 after_p = (! JUMP_P (last_insn)
6028 && (! CALL_P (last_insn)
6029 || (find_reg_note (last_insn,
6030 REG_NORETURN, NULL_RTX) == NULL_RTX
6031 && ! SIBLING_CALL_P (last_insn))));
6032 CLEAR_HARD_REG_SET (potential_reload_hard_regs);
6033 EXECUTE_IF_SET_IN_BITMAP (to_process, 0, j, bi)
6035 if ((int) j >= lra_constraint_new_regno_start)
6036 break;
6037 if (j < FIRST_PSEUDO_REGISTER || reg_renumber[j] >= 0)
6039 if (j < FIRST_PSEUDO_REGISTER)
6040 SET_HARD_REG_BIT (live_hard_regs, j);
6041 else
6042 add_to_hard_reg_set (&live_hard_regs,
6043 PSEUDO_REGNO_MODE (j),
6044 reg_renumber[j]);
6045 setup_next_usage_insn (j, last_insn, reloads_num, after_p);
6049 src_regno = dst_regno = -1;
6050 curr_set = single_set (curr_insn);
6051 if (curr_set != NULL_RTX && REG_P (SET_DEST (curr_set)))
6052 dst_regno = REGNO (SET_DEST (curr_set));
6053 if (curr_set != NULL_RTX && REG_P (SET_SRC (curr_set)))
6054 src_regno = REGNO (SET_SRC (curr_set));
6055 update_reloads_num_p = true;
6056 if (src_regno < lra_constraint_new_regno_start
6057 && src_regno >= FIRST_PSEUDO_REGISTER
6058 && reg_renumber[src_regno] < 0
6059 && dst_regno >= lra_constraint_new_regno_start
6060 && (cl = lra_get_allocno_class (dst_regno)) != NO_REGS)
6062 /* 'reload_pseudo <- original_pseudo'. */
6063 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6064 reloads_num++;
6065 update_reloads_num_p = false;
6066 succ_p = false;
6067 if (usage_insns[src_regno].check == curr_usage_insns_check
6068 && (next_usage_insns = usage_insns[src_regno].insns) != NULL_RTX)
6069 succ_p = inherit_reload_reg (false, src_regno, cl,
6070 curr_insn, next_usage_insns);
6071 if (succ_p)
6072 change_p = true;
6073 else
6074 setup_next_usage_insn (src_regno, curr_insn, reloads_num, false);
6075 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6076 IOR_HARD_REG_SET (potential_reload_hard_regs,
6077 reg_class_contents[cl]);
6079 else if (src_regno < 0
6080 && dst_regno >= lra_constraint_new_regno_start
6081 && invariant_p (SET_SRC (curr_set))
6082 && (cl = lra_get_allocno_class (dst_regno)) != NO_REGS
6083 && ! bitmap_bit_p (&invalid_invariant_regs, dst_regno)
6084 && ! bitmap_bit_p (&invalid_invariant_regs,
6085 ORIGINAL_REGNO(regno_reg_rtx[dst_regno])))
6087 /* 'reload_pseudo <- invariant'. */
6088 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6089 reloads_num++;
6090 update_reloads_num_p = false;
6091 if (process_invariant_for_inheritance (SET_DEST (curr_set), SET_SRC (curr_set)))
6092 change_p = true;
6093 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6094 IOR_HARD_REG_SET (potential_reload_hard_regs,
6095 reg_class_contents[cl]);
6097 else if (src_regno >= lra_constraint_new_regno_start
6098 && dst_regno < lra_constraint_new_regno_start
6099 && dst_regno >= FIRST_PSEUDO_REGISTER
6100 && reg_renumber[dst_regno] < 0
6101 && (cl = lra_get_allocno_class (src_regno)) != NO_REGS
6102 && usage_insns[dst_regno].check == curr_usage_insns_check
6103 && (next_usage_insns
6104 = usage_insns[dst_regno].insns) != NULL_RTX)
6106 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6107 reloads_num++;
6108 update_reloads_num_p = false;
6109 /* 'original_pseudo <- reload_pseudo'. */
6110 if (! JUMP_P (curr_insn)
6111 && inherit_reload_reg (true, dst_regno, cl,
6112 curr_insn, next_usage_insns))
6113 change_p = true;
6114 /* Invalidate. */
6115 usage_insns[dst_regno].check = 0;
6116 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6117 IOR_HARD_REG_SET (potential_reload_hard_regs,
6118 reg_class_contents[cl]);
6120 else if (INSN_P (curr_insn))
6122 int iter;
6123 int max_uid = get_max_uid ();
6125 curr_id = lra_get_insn_recog_data (curr_insn);
6126 curr_static_id = curr_id->insn_static_data;
6127 to_inherit_num = 0;
6128 /* Process insn definitions. */
6129 for (iter = 0; iter < 2; iter++)
6130 for (reg = iter == 0 ? curr_id->regs : curr_static_id->hard_regs;
6131 reg != NULL;
6132 reg = reg->next)
6133 if (reg->type != OP_IN
6134 && (dst_regno = reg->regno) < lra_constraint_new_regno_start)
6136 if (dst_regno >= FIRST_PSEUDO_REGISTER && reg->type == OP_OUT
6137 && reg_renumber[dst_regno] < 0 && ! reg->subreg_p
6138 && usage_insns[dst_regno].check == curr_usage_insns_check
6139 && (next_usage_insns
6140 = usage_insns[dst_regno].insns) != NULL_RTX)
6142 struct lra_insn_reg *r;
6144 for (r = curr_id->regs; r != NULL; r = r->next)
6145 if (r->type != OP_OUT && r->regno == dst_regno)
6146 break;
6147 /* Don't do inheritance if the pseudo is also
6148 used in the insn. */
6149 if (r == NULL)
6150 /* We can not do inheritance right now
6151 because the current insn reg info (chain
6152 regs) can change after that. */
6153 add_to_inherit (dst_regno, next_usage_insns);
6155 /* We can not process one reg twice here because of
6156 usage_insns invalidation. */
6157 if ((dst_regno < FIRST_PSEUDO_REGISTER
6158 || reg_renumber[dst_regno] >= 0)
6159 && ! reg->subreg_p && reg->type != OP_IN)
6161 HARD_REG_SET s;
6163 if (split_if_necessary (dst_regno, reg->biggest_mode,
6164 potential_reload_hard_regs,
6165 false, curr_insn, max_uid))
6166 change_p = true;
6167 CLEAR_HARD_REG_SET (s);
6168 if (dst_regno < FIRST_PSEUDO_REGISTER)
6169 add_to_hard_reg_set (&s, reg->biggest_mode, dst_regno);
6170 else
6171 add_to_hard_reg_set (&s, PSEUDO_REGNO_MODE (dst_regno),
6172 reg_renumber[dst_regno]);
6173 AND_COMPL_HARD_REG_SET (live_hard_regs, s);
6175 /* We should invalidate potential inheritance or
6176 splitting for the current insn usages to the next
6177 usage insns (see code below) as the output pseudo
6178 prevents this. */
6179 if ((dst_regno >= FIRST_PSEUDO_REGISTER
6180 && reg_renumber[dst_regno] < 0)
6181 || (reg->type == OP_OUT && ! reg->subreg_p
6182 && (dst_regno < FIRST_PSEUDO_REGISTER
6183 || reg_renumber[dst_regno] >= 0)))
6185 /* Invalidate and mark definitions. */
6186 if (dst_regno >= FIRST_PSEUDO_REGISTER)
6187 usage_insns[dst_regno].check = -(int) INSN_UID (curr_insn);
6188 else
6190 nregs = hard_regno_nregs (dst_regno,
6191 reg->biggest_mode);
6192 for (i = 0; i < nregs; i++)
6193 usage_insns[dst_regno + i].check
6194 = -(int) INSN_UID (curr_insn);
6198 /* Process clobbered call regs. */
6199 if (curr_id->arg_hard_regs != NULL)
6200 for (i = 0; (dst_regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6201 if (dst_regno >= FIRST_PSEUDO_REGISTER)
6202 usage_insns[dst_regno - FIRST_PSEUDO_REGISTER].check
6203 = -(int) INSN_UID (curr_insn);
6204 if (! JUMP_P (curr_insn))
6205 for (i = 0; i < to_inherit_num; i++)
6206 if (inherit_reload_reg (true, to_inherit[i].regno,
6207 ALL_REGS, curr_insn,
6208 to_inherit[i].insns))
6209 change_p = true;
6210 if (CALL_P (curr_insn))
6212 rtx cheap, pat, dest;
6213 rtx_insn *restore;
6214 int regno, hard_regno;
6216 calls_num++;
6217 if ((cheap = find_reg_note (curr_insn,
6218 REG_RETURNED, NULL_RTX)) != NULL_RTX
6219 && ((cheap = XEXP (cheap, 0)), true)
6220 && (regno = REGNO (cheap)) >= FIRST_PSEUDO_REGISTER
6221 && (hard_regno = reg_renumber[regno]) >= 0
6222 /* If there are pending saves/restores, the
6223 optimization is not worth. */
6224 && usage_insns[regno].calls_num == calls_num - 1
6225 && TEST_HARD_REG_BIT (call_used_reg_set, hard_regno))
6227 /* Restore the pseudo from the call result as
6228 REG_RETURNED note says that the pseudo value is
6229 in the call result and the pseudo is an argument
6230 of the call. */
6231 pat = PATTERN (curr_insn);
6232 if (GET_CODE (pat) == PARALLEL)
6233 pat = XVECEXP (pat, 0, 0);
6234 dest = SET_DEST (pat);
6235 /* For multiple return values dest is PARALLEL.
6236 Currently we handle only single return value case. */
6237 if (REG_P (dest))
6239 start_sequence ();
6240 emit_move_insn (cheap, copy_rtx (dest));
6241 restore = get_insns ();
6242 end_sequence ();
6243 lra_process_new_insns (curr_insn, NULL, restore,
6244 "Inserting call parameter restore");
6245 /* We don't need to save/restore of the pseudo from
6246 this call. */
6247 usage_insns[regno].calls_num = calls_num;
6248 bitmap_set_bit (&check_only_regs, regno);
6252 to_inherit_num = 0;
6253 /* Process insn usages. */
6254 for (iter = 0; iter < 2; iter++)
6255 for (reg = iter == 0 ? curr_id->regs : curr_static_id->hard_regs;
6256 reg != NULL;
6257 reg = reg->next)
6258 if ((reg->type != OP_OUT
6259 || (reg->type == OP_OUT && reg->subreg_p))
6260 && (src_regno = reg->regno) < lra_constraint_new_regno_start)
6262 if (src_regno >= FIRST_PSEUDO_REGISTER
6263 && reg_renumber[src_regno] < 0 && reg->type == OP_IN)
6265 if (usage_insns[src_regno].check == curr_usage_insns_check
6266 && (next_usage_insns
6267 = usage_insns[src_regno].insns) != NULL_RTX
6268 && NONDEBUG_INSN_P (curr_insn))
6269 add_to_inherit (src_regno, next_usage_insns);
6270 else if (usage_insns[src_regno].check
6271 != -(int) INSN_UID (curr_insn))
6272 /* Add usages but only if the reg is not set up
6273 in the same insn. */
6274 add_next_usage_insn (src_regno, curr_insn, reloads_num);
6276 else if (src_regno < FIRST_PSEUDO_REGISTER
6277 || reg_renumber[src_regno] >= 0)
6279 bool before_p;
6280 rtx_insn *use_insn = curr_insn;
6282 before_p = (JUMP_P (curr_insn)
6283 || (CALL_P (curr_insn) && reg->type == OP_IN));
6284 if (NONDEBUG_INSN_P (curr_insn)
6285 && (! JUMP_P (curr_insn) || reg->type == OP_IN)
6286 && split_if_necessary (src_regno, reg->biggest_mode,
6287 potential_reload_hard_regs,
6288 before_p, curr_insn, max_uid))
6290 if (reg->subreg_p)
6291 lra_risky_transformations_p = true;
6292 change_p = true;
6293 /* Invalidate. */
6294 usage_insns[src_regno].check = 0;
6295 if (before_p)
6296 use_insn = PREV_INSN (curr_insn);
6298 if (NONDEBUG_INSN_P (curr_insn))
6300 if (src_regno < FIRST_PSEUDO_REGISTER)
6301 add_to_hard_reg_set (&live_hard_regs,
6302 reg->biggest_mode, src_regno);
6303 else
6304 add_to_hard_reg_set (&live_hard_regs,
6305 PSEUDO_REGNO_MODE (src_regno),
6306 reg_renumber[src_regno]);
6308 add_next_usage_insn (src_regno, use_insn, reloads_num);
6311 /* Process used call regs. */
6312 if (curr_id->arg_hard_regs != NULL)
6313 for (i = 0; (src_regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6314 if (src_regno < FIRST_PSEUDO_REGISTER)
6316 SET_HARD_REG_BIT (live_hard_regs, src_regno);
6317 add_next_usage_insn (src_regno, curr_insn, reloads_num);
6319 for (i = 0; i < to_inherit_num; i++)
6321 src_regno = to_inherit[i].regno;
6322 if (inherit_reload_reg (false, src_regno, ALL_REGS,
6323 curr_insn, to_inherit[i].insns))
6324 change_p = true;
6325 else
6326 setup_next_usage_insn (src_regno, curr_insn, reloads_num, false);
6329 if (update_reloads_num_p
6330 && NONDEBUG_INSN_P (curr_insn) && curr_set != NULL_RTX)
6332 int regno = -1;
6333 if ((REG_P (SET_DEST (curr_set))
6334 && (regno = REGNO (SET_DEST (curr_set))) >= lra_constraint_new_regno_start
6335 && reg_renumber[regno] < 0
6336 && (cl = lra_get_allocno_class (regno)) != NO_REGS)
6337 || (REG_P (SET_SRC (curr_set))
6338 && (regno = REGNO (SET_SRC (curr_set))) >= lra_constraint_new_regno_start
6339 && reg_renumber[regno] < 0
6340 && (cl = lra_get_allocno_class (regno)) != NO_REGS))
6342 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6343 reloads_num++;
6344 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6345 IOR_HARD_REG_SET (potential_reload_hard_regs,
6346 reg_class_contents[cl]);
6349 if (NONDEBUG_INSN_P (curr_insn))
6351 int regno;
6353 /* Invalidate invariants with changed regs. */
6354 curr_id = lra_get_insn_recog_data (curr_insn);
6355 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
6356 if (reg->type != OP_IN)
6358 bitmap_set_bit (&invalid_invariant_regs, reg->regno);
6359 bitmap_set_bit (&invalid_invariant_regs,
6360 ORIGINAL_REGNO (regno_reg_rtx[reg->regno]));
6362 curr_static_id = curr_id->insn_static_data;
6363 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
6364 if (reg->type != OP_IN)
6365 bitmap_set_bit (&invalid_invariant_regs, reg->regno);
6366 if (curr_id->arg_hard_regs != NULL)
6367 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6368 if (regno >= FIRST_PSEUDO_REGISTER)
6369 bitmap_set_bit (&invalid_invariant_regs,
6370 regno - FIRST_PSEUDO_REGISTER);
6372 /* We reached the start of the current basic block. */
6373 if (prev_insn == NULL_RTX || prev_insn == PREV_INSN (head)
6374 || BLOCK_FOR_INSN (prev_insn) != curr_bb)
6376 /* We reached the beginning of the current block -- do
6377 rest of spliting in the current BB. */
6378 to_process = df_get_live_in (curr_bb);
6379 if (BLOCK_FOR_INSN (head) != curr_bb)
6381 /* We are somewhere in the middle of EBB. */
6382 get_live_on_other_edges (EDGE_PRED (curr_bb, 0)->src,
6383 curr_bb, &temp_bitmap);
6384 to_process = &temp_bitmap;
6386 head_p = true;
6387 EXECUTE_IF_SET_IN_BITMAP (to_process, 0, j, bi)
6389 if ((int) j >= lra_constraint_new_regno_start)
6390 break;
6391 if (((int) j < FIRST_PSEUDO_REGISTER || reg_renumber[j] >= 0)
6392 && usage_insns[j].check == curr_usage_insns_check
6393 && (next_usage_insns = usage_insns[j].insns) != NULL_RTX)
6395 if (need_for_split_p (potential_reload_hard_regs, j))
6397 if (lra_dump_file != NULL && head_p)
6399 fprintf (lra_dump_file,
6400 " ----------------------------------\n");
6401 head_p = false;
6403 if (split_reg (false, j, bb_note (curr_bb),
6404 next_usage_insns))
6405 change_p = true;
6407 usage_insns[j].check = 0;
6412 return change_p;
6415 /* This value affects EBB forming. If probability of edge from EBB to
6416 a BB is not greater than the following value, we don't add the BB
6417 to EBB. */
6418 #define EBB_PROBABILITY_CUTOFF \
6419 ((REG_BR_PROB_BASE * LRA_INHERITANCE_EBB_PROBABILITY_CUTOFF) / 100)
6421 /* Current number of inheritance/split iteration. */
6422 int lra_inheritance_iter;
6424 /* Entry function for inheritance/split pass. */
6425 void
6426 lra_inheritance (void)
6428 int i;
6429 basic_block bb, start_bb;
6430 edge e;
6432 lra_inheritance_iter++;
6433 if (lra_inheritance_iter > LRA_MAX_INHERITANCE_PASSES)
6434 return;
6435 timevar_push (TV_LRA_INHERITANCE);
6436 if (lra_dump_file != NULL)
6437 fprintf (lra_dump_file, "\n********** Inheritance #%d: **********\n\n",
6438 lra_inheritance_iter);
6439 curr_usage_insns_check = 0;
6440 usage_insns = XNEWVEC (struct usage_insns, lra_constraint_new_regno_start);
6441 for (i = 0; i < lra_constraint_new_regno_start; i++)
6442 usage_insns[i].check = 0;
6443 bitmap_initialize (&check_only_regs, &reg_obstack);
6444 bitmap_initialize (&invalid_invariant_regs, &reg_obstack);
6445 bitmap_initialize (&live_regs, &reg_obstack);
6446 bitmap_initialize (&temp_bitmap, &reg_obstack);
6447 bitmap_initialize (&ebb_global_regs, &reg_obstack);
6448 FOR_EACH_BB_FN (bb, cfun)
6450 start_bb = bb;
6451 if (lra_dump_file != NULL)
6452 fprintf (lra_dump_file, "EBB");
6453 /* Form a EBB starting with BB. */
6454 bitmap_clear (&ebb_global_regs);
6455 bitmap_ior_into (&ebb_global_regs, df_get_live_in (bb));
6456 for (;;)
6458 if (lra_dump_file != NULL)
6459 fprintf (lra_dump_file, " %d", bb->index);
6460 if (bb->next_bb == EXIT_BLOCK_PTR_FOR_FN (cfun)
6461 || LABEL_P (BB_HEAD (bb->next_bb)))
6462 break;
6463 e = find_fallthru_edge (bb->succs);
6464 if (! e)
6465 break;
6466 if (e->probability.initialized_p ()
6467 && e->probability.to_reg_br_prob_base () < EBB_PROBABILITY_CUTOFF)
6468 break;
6469 bb = bb->next_bb;
6471 bitmap_ior_into (&ebb_global_regs, df_get_live_out (bb));
6472 if (lra_dump_file != NULL)
6473 fprintf (lra_dump_file, "\n");
6474 if (inherit_in_ebb (BB_HEAD (start_bb), BB_END (bb)))
6475 /* Remember that the EBB head and tail can change in
6476 inherit_in_ebb. */
6477 update_ebb_live_info (BB_HEAD (start_bb), BB_END (bb));
6479 bitmap_clear (&ebb_global_regs);
6480 bitmap_clear (&temp_bitmap);
6481 bitmap_clear (&live_regs);
6482 bitmap_clear (&invalid_invariant_regs);
6483 bitmap_clear (&check_only_regs);
6484 free (usage_insns);
6486 timevar_pop (TV_LRA_INHERITANCE);
6491 /* This page contains code to undo failed inheritance/split
6492 transformations. */
6494 /* Current number of iteration undoing inheritance/split. */
6495 int lra_undo_inheritance_iter;
6497 /* Fix BB live info LIVE after removing pseudos created on pass doing
6498 inheritance/split which are REMOVED_PSEUDOS. */
6499 static void
6500 fix_bb_live_info (bitmap live, bitmap removed_pseudos)
6502 unsigned int regno;
6503 bitmap_iterator bi;
6505 EXECUTE_IF_SET_IN_BITMAP (removed_pseudos, 0, regno, bi)
6506 if (bitmap_clear_bit (live, regno)
6507 && REG_P (lra_reg_info[regno].restore_rtx))
6508 bitmap_set_bit (live, REGNO (lra_reg_info[regno].restore_rtx));
6511 /* Return regno of the (subreg of) REG. Otherwise, return a negative
6512 number. */
6513 static int
6514 get_regno (rtx reg)
6516 if (GET_CODE (reg) == SUBREG)
6517 reg = SUBREG_REG (reg);
6518 if (REG_P (reg))
6519 return REGNO (reg);
6520 return -1;
6523 /* Delete a move INSN with destination reg DREGNO and a previous
6524 clobber insn with the same regno. The inheritance/split code can
6525 generate moves with preceding clobber and when we delete such moves
6526 we should delete the clobber insn too to keep the correct life
6527 info. */
6528 static void
6529 delete_move_and_clobber (rtx_insn *insn, int dregno)
6531 rtx_insn *prev_insn = PREV_INSN (insn);
6533 lra_set_insn_deleted (insn);
6534 lra_assert (dregno >= 0);
6535 if (prev_insn != NULL && NONDEBUG_INSN_P (prev_insn)
6536 && GET_CODE (PATTERN (prev_insn)) == CLOBBER
6537 && dregno == get_regno (XEXP (PATTERN (prev_insn), 0)))
6538 lra_set_insn_deleted (prev_insn);
6541 /* Remove inheritance/split pseudos which are in REMOVE_PSEUDOS and
6542 return true if we did any change. The undo transformations for
6543 inheritance looks like
6544 i <- i2
6545 p <- i => p <- i2
6546 or removing
6547 p <- i, i <- p, and i <- i3
6548 where p is original pseudo from which inheritance pseudo i was
6549 created, i and i3 are removed inheritance pseudos, i2 is another
6550 not removed inheritance pseudo. All split pseudos or other
6551 occurrences of removed inheritance pseudos are changed on the
6552 corresponding original pseudos.
6554 The function also schedules insns changed and created during
6555 inheritance/split pass for processing by the subsequent constraint
6556 pass. */
6557 static bool
6558 remove_inheritance_pseudos (bitmap remove_pseudos)
6560 basic_block bb;
6561 int regno, sregno, prev_sregno, dregno;
6562 rtx restore_rtx;
6563 rtx set, prev_set;
6564 rtx_insn *prev_insn;
6565 bool change_p, done_p;
6567 change_p = ! bitmap_empty_p (remove_pseudos);
6568 /* We can not finish the function right away if CHANGE_P is true
6569 because we need to marks insns affected by previous
6570 inheritance/split pass for processing by the subsequent
6571 constraint pass. */
6572 FOR_EACH_BB_FN (bb, cfun)
6574 fix_bb_live_info (df_get_live_in (bb), remove_pseudos);
6575 fix_bb_live_info (df_get_live_out (bb), remove_pseudos);
6576 FOR_BB_INSNS_REVERSE (bb, curr_insn)
6578 if (! INSN_P (curr_insn))
6579 continue;
6580 done_p = false;
6581 sregno = dregno = -1;
6582 if (change_p && NONDEBUG_INSN_P (curr_insn)
6583 && (set = single_set (curr_insn)) != NULL_RTX)
6585 dregno = get_regno (SET_DEST (set));
6586 sregno = get_regno (SET_SRC (set));
6589 if (sregno >= 0 && dregno >= 0)
6591 if (bitmap_bit_p (remove_pseudos, dregno)
6592 && ! REG_P (lra_reg_info[dregno].restore_rtx))
6594 /* invariant inheritance pseudo <- original pseudo */
6595 if (lra_dump_file != NULL)
6597 fprintf (lra_dump_file, " Removing invariant inheritance:\n");
6598 dump_insn_slim (lra_dump_file, curr_insn);
6599 fprintf (lra_dump_file, "\n");
6601 delete_move_and_clobber (curr_insn, dregno);
6602 done_p = true;
6604 else if (bitmap_bit_p (remove_pseudos, sregno)
6605 && ! REG_P (lra_reg_info[sregno].restore_rtx))
6607 /* reload pseudo <- invariant inheritance pseudo */
6608 start_sequence ();
6609 /* We can not just change the source. It might be
6610 an insn different from the move. */
6611 emit_insn (lra_reg_info[sregno].restore_rtx);
6612 rtx_insn *new_insns = get_insns ();
6613 end_sequence ();
6614 lra_assert (single_set (new_insns) != NULL
6615 && SET_DEST (set) == SET_DEST (single_set (new_insns)));
6616 lra_process_new_insns (curr_insn, NULL, new_insns,
6617 "Changing reload<-invariant inheritance");
6618 delete_move_and_clobber (curr_insn, dregno);
6619 done_p = true;
6621 else if ((bitmap_bit_p (remove_pseudos, sregno)
6622 && (get_regno (lra_reg_info[sregno].restore_rtx) == dregno
6623 || (bitmap_bit_p (remove_pseudos, dregno)
6624 && get_regno (lra_reg_info[sregno].restore_rtx) >= 0
6625 && (get_regno (lra_reg_info[sregno].restore_rtx)
6626 == get_regno (lra_reg_info[dregno].restore_rtx)))))
6627 || (bitmap_bit_p (remove_pseudos, dregno)
6628 && get_regno (lra_reg_info[dregno].restore_rtx) == sregno))
6629 /* One of the following cases:
6630 original <- removed inheritance pseudo
6631 removed inherit pseudo <- another removed inherit pseudo
6632 removed inherit pseudo <- original pseudo
6634 removed_split_pseudo <- original_reg
6635 original_reg <- removed_split_pseudo */
6637 if (lra_dump_file != NULL)
6639 fprintf (lra_dump_file, " Removing %s:\n",
6640 bitmap_bit_p (&lra_split_regs, sregno)
6641 || bitmap_bit_p (&lra_split_regs, dregno)
6642 ? "split" : "inheritance");
6643 dump_insn_slim (lra_dump_file, curr_insn);
6645 delete_move_and_clobber (curr_insn, dregno);
6646 done_p = true;
6648 else if (bitmap_bit_p (remove_pseudos, sregno)
6649 && bitmap_bit_p (&lra_inheritance_pseudos, sregno))
6651 /* Search the following pattern:
6652 inherit_or_split_pseudo1 <- inherit_or_split_pseudo2
6653 original_pseudo <- inherit_or_split_pseudo1
6654 where the 2nd insn is the current insn and
6655 inherit_or_split_pseudo2 is not removed. If it is found,
6656 change the current insn onto:
6657 original_pseudo <- inherit_or_split_pseudo2. */
6658 for (prev_insn = PREV_INSN (curr_insn);
6659 prev_insn != NULL_RTX && ! NONDEBUG_INSN_P (prev_insn);
6660 prev_insn = PREV_INSN (prev_insn))
6662 if (prev_insn != NULL_RTX && BLOCK_FOR_INSN (prev_insn) == bb
6663 && (prev_set = single_set (prev_insn)) != NULL_RTX
6664 /* There should be no subregs in insn we are
6665 searching because only the original reg might
6666 be in subreg when we changed the mode of
6667 load/store for splitting. */
6668 && REG_P (SET_DEST (prev_set))
6669 && REG_P (SET_SRC (prev_set))
6670 && (int) REGNO (SET_DEST (prev_set)) == sregno
6671 && ((prev_sregno = REGNO (SET_SRC (prev_set)))
6672 >= FIRST_PSEUDO_REGISTER)
6673 && (lra_reg_info[prev_sregno].restore_rtx == NULL_RTX
6675 /* As we consider chain of inheritance or
6676 splitting described in above comment we should
6677 check that sregno and prev_sregno were
6678 inheritance/split pseudos created from the
6679 same original regno. */
6680 (get_regno (lra_reg_info[sregno].restore_rtx) >= 0
6681 && (get_regno (lra_reg_info[sregno].restore_rtx)
6682 == get_regno (lra_reg_info[prev_sregno].restore_rtx))))
6683 && ! bitmap_bit_p (remove_pseudos, prev_sregno))
6685 lra_assert (GET_MODE (SET_SRC (prev_set))
6686 == GET_MODE (regno_reg_rtx[sregno]));
6687 if (GET_CODE (SET_SRC (set)) == SUBREG)
6688 SUBREG_REG (SET_SRC (set)) = SET_SRC (prev_set);
6689 else
6690 SET_SRC (set) = SET_SRC (prev_set);
6691 /* As we are finishing with processing the insn
6692 here, check the destination too as it might
6693 inheritance pseudo for another pseudo. */
6694 if (bitmap_bit_p (remove_pseudos, dregno)
6695 && bitmap_bit_p (&lra_inheritance_pseudos, dregno)
6696 && (restore_rtx
6697 = lra_reg_info[dregno].restore_rtx) != NULL_RTX)
6699 if (GET_CODE (SET_DEST (set)) == SUBREG)
6700 SUBREG_REG (SET_DEST (set)) = restore_rtx;
6701 else
6702 SET_DEST (set) = restore_rtx;
6704 lra_push_insn_and_update_insn_regno_info (curr_insn);
6705 lra_set_used_insn_alternative_by_uid
6706 (INSN_UID (curr_insn), -1);
6707 done_p = true;
6708 if (lra_dump_file != NULL)
6710 fprintf (lra_dump_file, " Change reload insn:\n");
6711 dump_insn_slim (lra_dump_file, curr_insn);
6716 if (! done_p)
6718 struct lra_insn_reg *reg;
6719 bool restored_regs_p = false;
6720 bool kept_regs_p = false;
6722 curr_id = lra_get_insn_recog_data (curr_insn);
6723 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
6725 regno = reg->regno;
6726 restore_rtx = lra_reg_info[regno].restore_rtx;
6727 if (restore_rtx != NULL_RTX)
6729 if (change_p && bitmap_bit_p (remove_pseudos, regno))
6731 lra_substitute_pseudo_within_insn
6732 (curr_insn, regno, restore_rtx, false);
6733 restored_regs_p = true;
6735 else
6736 kept_regs_p = true;
6739 if (NONDEBUG_INSN_P (curr_insn) && kept_regs_p)
6741 /* The instruction has changed since the previous
6742 constraints pass. */
6743 lra_push_insn_and_update_insn_regno_info (curr_insn);
6744 lra_set_used_insn_alternative_by_uid
6745 (INSN_UID (curr_insn), -1);
6747 else if (restored_regs_p)
6748 /* The instruction has been restored to the form that
6749 it had during the previous constraints pass. */
6750 lra_update_insn_regno_info (curr_insn);
6751 if (restored_regs_p && lra_dump_file != NULL)
6753 fprintf (lra_dump_file, " Insn after restoring regs:\n");
6754 dump_insn_slim (lra_dump_file, curr_insn);
6759 return change_p;
6762 /* If optional reload pseudos failed to get a hard register or was not
6763 inherited, it is better to remove optional reloads. We do this
6764 transformation after undoing inheritance to figure out necessity to
6765 remove optional reloads easier. Return true if we do any
6766 change. */
6767 static bool
6768 undo_optional_reloads (void)
6770 bool change_p, keep_p;
6771 unsigned int regno, uid;
6772 bitmap_iterator bi, bi2;
6773 rtx_insn *insn;
6774 rtx set, src, dest;
6775 auto_bitmap removed_optional_reload_pseudos (&reg_obstack);
6777 bitmap_copy (removed_optional_reload_pseudos, &lra_optional_reload_pseudos);
6778 EXECUTE_IF_SET_IN_BITMAP (&lra_optional_reload_pseudos, 0, regno, bi)
6780 keep_p = false;
6781 /* Keep optional reloads from previous subpasses. */
6782 if (lra_reg_info[regno].restore_rtx == NULL_RTX
6783 /* If the original pseudo changed its allocation, just
6784 removing the optional pseudo is dangerous as the original
6785 pseudo will have longer live range. */
6786 || reg_renumber[REGNO (lra_reg_info[regno].restore_rtx)] >= 0)
6787 keep_p = true;
6788 else if (reg_renumber[regno] >= 0)
6789 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi2)
6791 insn = lra_insn_recog_data[uid]->insn;
6792 if ((set = single_set (insn)) == NULL_RTX)
6793 continue;
6794 src = SET_SRC (set);
6795 dest = SET_DEST (set);
6796 if (! REG_P (src) || ! REG_P (dest))
6797 continue;
6798 if (REGNO (dest) == regno
6799 /* Ignore insn for optional reloads itself. */
6800 && REGNO (lra_reg_info[regno].restore_rtx) != REGNO (src)
6801 /* Check only inheritance on last inheritance pass. */
6802 && (int) REGNO (src) >= new_regno_start
6803 /* Check that the optional reload was inherited. */
6804 && bitmap_bit_p (&lra_inheritance_pseudos, REGNO (src)))
6806 keep_p = true;
6807 break;
6810 if (keep_p)
6812 bitmap_clear_bit (removed_optional_reload_pseudos, regno);
6813 if (lra_dump_file != NULL)
6814 fprintf (lra_dump_file, "Keep optional reload reg %d\n", regno);
6817 change_p = ! bitmap_empty_p (removed_optional_reload_pseudos);
6818 auto_bitmap insn_bitmap (&reg_obstack);
6819 EXECUTE_IF_SET_IN_BITMAP (removed_optional_reload_pseudos, 0, regno, bi)
6821 if (lra_dump_file != NULL)
6822 fprintf (lra_dump_file, "Remove optional reload reg %d\n", regno);
6823 bitmap_copy (insn_bitmap, &lra_reg_info[regno].insn_bitmap);
6824 EXECUTE_IF_SET_IN_BITMAP (insn_bitmap, 0, uid, bi2)
6826 insn = lra_insn_recog_data[uid]->insn;
6827 if ((set = single_set (insn)) != NULL_RTX)
6829 src = SET_SRC (set);
6830 dest = SET_DEST (set);
6831 if (REG_P (src) && REG_P (dest)
6832 && ((REGNO (src) == regno
6833 && (REGNO (lra_reg_info[regno].restore_rtx)
6834 == REGNO (dest)))
6835 || (REGNO (dest) == regno
6836 && (REGNO (lra_reg_info[regno].restore_rtx)
6837 == REGNO (src)))))
6839 if (lra_dump_file != NULL)
6841 fprintf (lra_dump_file, " Deleting move %u\n",
6842 INSN_UID (insn));
6843 dump_insn_slim (lra_dump_file, insn);
6845 delete_move_and_clobber (insn, REGNO (dest));
6846 continue;
6848 /* We should not worry about generation memory-memory
6849 moves here as if the corresponding inheritance did
6850 not work (inheritance pseudo did not get a hard reg),
6851 we remove the inheritance pseudo and the optional
6852 reload. */
6854 lra_substitute_pseudo_within_insn
6855 (insn, regno, lra_reg_info[regno].restore_rtx, false);
6856 lra_update_insn_regno_info (insn);
6857 if (lra_dump_file != NULL)
6859 fprintf (lra_dump_file,
6860 " Restoring original insn:\n");
6861 dump_insn_slim (lra_dump_file, insn);
6865 /* Clear restore_regnos. */
6866 EXECUTE_IF_SET_IN_BITMAP (&lra_optional_reload_pseudos, 0, regno, bi)
6867 lra_reg_info[regno].restore_rtx = NULL_RTX;
6868 return change_p;
6871 /* Entry function for undoing inheritance/split transformation. Return true
6872 if we did any RTL change in this pass. */
6873 bool
6874 lra_undo_inheritance (void)
6876 unsigned int regno;
6877 int hard_regno;
6878 int n_all_inherit, n_inherit, n_all_split, n_split;
6879 rtx restore_rtx;
6880 bitmap_iterator bi;
6881 bool change_p;
6883 lra_undo_inheritance_iter++;
6884 if (lra_undo_inheritance_iter > LRA_MAX_INHERITANCE_PASSES)
6885 return false;
6886 if (lra_dump_file != NULL)
6887 fprintf (lra_dump_file,
6888 "\n********** Undoing inheritance #%d: **********\n\n",
6889 lra_undo_inheritance_iter);
6890 auto_bitmap remove_pseudos (&reg_obstack);
6891 n_inherit = n_all_inherit = 0;
6892 EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, regno, bi)
6893 if (lra_reg_info[regno].restore_rtx != NULL_RTX)
6895 n_all_inherit++;
6896 if (reg_renumber[regno] < 0
6897 /* If the original pseudo changed its allocation, just
6898 removing inheritance is dangerous as for changing
6899 allocation we used shorter live-ranges. */
6900 && (! REG_P (lra_reg_info[regno].restore_rtx)
6901 || reg_renumber[REGNO (lra_reg_info[regno].restore_rtx)] < 0))
6902 bitmap_set_bit (remove_pseudos, regno);
6903 else
6904 n_inherit++;
6906 if (lra_dump_file != NULL && n_all_inherit != 0)
6907 fprintf (lra_dump_file, "Inherit %d out of %d (%.2f%%)\n",
6908 n_inherit, n_all_inherit,
6909 (double) n_inherit / n_all_inherit * 100);
6910 n_split = n_all_split = 0;
6911 EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, regno, bi)
6912 if ((restore_rtx = lra_reg_info[regno].restore_rtx) != NULL_RTX)
6914 int restore_regno = REGNO (restore_rtx);
6916 n_all_split++;
6917 hard_regno = (restore_regno >= FIRST_PSEUDO_REGISTER
6918 ? reg_renumber[restore_regno] : restore_regno);
6919 if (hard_regno < 0 || reg_renumber[regno] == hard_regno)
6920 bitmap_set_bit (remove_pseudos, regno);
6921 else
6923 n_split++;
6924 if (lra_dump_file != NULL)
6925 fprintf (lra_dump_file, " Keep split r%d (orig=r%d)\n",
6926 regno, restore_regno);
6929 if (lra_dump_file != NULL && n_all_split != 0)
6930 fprintf (lra_dump_file, "Split %d out of %d (%.2f%%)\n",
6931 n_split, n_all_split,
6932 (double) n_split / n_all_split * 100);
6933 change_p = remove_inheritance_pseudos (remove_pseudos);
6934 /* Clear restore_regnos. */
6935 EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, regno, bi)
6936 lra_reg_info[regno].restore_rtx = NULL_RTX;
6937 EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, regno, bi)
6938 lra_reg_info[regno].restore_rtx = NULL_RTX;
6939 change_p = undo_optional_reloads () || change_p;
6940 return change_p;