poly_int: lra frame offsets
[official-gcc.git] / gcc / lra-constraints.c
blob0681a4c188d3958e6519bc7708a0b2eafe7c9b18
1 /* Code for RTL transformations to satisfy insn constraints.
2 Copyright (C) 2010-2017 Free Software Foundation, Inc.
3 Contributed by Vladimir Makarov <vmakarov@redhat.com>.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
10 version.
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
22 /* This file contains code for 3 passes: constraint pass,
23 inheritance/split pass, and pass for undoing failed inheritance and
24 split.
26 The major goal of constraint pass is to transform RTL to satisfy
27 insn and address constraints by:
28 o choosing insn alternatives;
29 o generating *reload insns* (or reloads in brief) and *reload
30 pseudos* which will get necessary hard registers later;
31 o substituting pseudos with equivalent values and removing the
32 instructions that initialized those pseudos.
34 The constraint pass has biggest and most complicated code in LRA.
35 There are a lot of important details like:
36 o reuse of input reload pseudos to simplify reload pseudo
37 allocations;
38 o some heuristics to choose insn alternative to improve the
39 inheritance;
40 o early clobbers etc.
42 The pass is mimicking former reload pass in alternative choosing
43 because the reload pass is oriented to current machine description
44 model. It might be changed if the machine description model is
45 changed.
47 There is special code for preventing all LRA and this pass cycling
48 in case of bugs.
50 On the first iteration of the pass we process every instruction and
51 choose an alternative for each one. On subsequent iterations we try
52 to avoid reprocessing instructions if we can be sure that the old
53 choice is still valid.
55 The inheritance/spilt pass is to transform code to achieve
56 ineheritance and live range splitting. It is done on backward
57 traversal of EBBs.
59 The inheritance optimization goal is to reuse values in hard
60 registers. There is analogous optimization in old reload pass. The
61 inheritance is achieved by following transformation:
63 reload_p1 <- p reload_p1 <- p
64 ... new_p <- reload_p1
65 ... => ...
66 reload_p2 <- p reload_p2 <- new_p
68 where p is spilled and not changed between the insns. Reload_p1 is
69 also called *original pseudo* and new_p is called *inheritance
70 pseudo*.
72 The subsequent assignment pass will try to assign the same (or
73 another if it is not possible) hard register to new_p as to
74 reload_p1 or reload_p2.
76 If the assignment pass fails to assign a hard register to new_p,
77 this file will undo the inheritance and restore the original code.
78 This is because implementing the above sequence with a spilled
79 new_p would make the code much worse. The inheritance is done in
80 EBB scope. The above is just a simplified example to get an idea
81 of the inheritance as the inheritance is also done for non-reload
82 insns.
84 Splitting (transformation) is also done in EBB scope on the same
85 pass as the inheritance:
87 r <- ... or ... <- r r <- ... or ... <- r
88 ... s <- r (new insn -- save)
89 ... =>
90 ... r <- s (new insn -- restore)
91 ... <- r ... <- r
93 The *split pseudo* s is assigned to the hard register of the
94 original pseudo or hard register r.
96 Splitting is done:
97 o In EBBs with high register pressure for global pseudos (living
98 in at least 2 BBs) and assigned to hard registers when there
99 are more one reloads needing the hard registers;
100 o for pseudos needing save/restore code around calls.
102 If the split pseudo still has the same hard register as the
103 original pseudo after the subsequent assignment pass or the
104 original pseudo was split, the opposite transformation is done on
105 the same pass for undoing inheritance. */
107 #undef REG_OK_STRICT
109 #include "config.h"
110 #include "system.h"
111 #include "coretypes.h"
112 #include "backend.h"
113 #include "target.h"
114 #include "rtl.h"
115 #include "tree.h"
116 #include "predict.h"
117 #include "df.h"
118 #include "memmodel.h"
119 #include "tm_p.h"
120 #include "expmed.h"
121 #include "optabs.h"
122 #include "regs.h"
123 #include "ira.h"
124 #include "recog.h"
125 #include "output.h"
126 #include "addresses.h"
127 #include "expr.h"
128 #include "cfgrtl.h"
129 #include "rtl-error.h"
130 #include "params.h"
131 #include "lra.h"
132 #include "lra-int.h"
133 #include "print-rtl.h"
135 /* Value of LRA_CURR_RELOAD_NUM at the beginning of BB of the current
136 insn. Remember that LRA_CURR_RELOAD_NUM is the number of emitted
137 reload insns. */
138 static int bb_reload_num;
140 /* The current insn being processed and corresponding its single set
141 (NULL otherwise), its data (basic block, the insn data, the insn
142 static data, and the mode of each operand). */
143 static rtx_insn *curr_insn;
144 static rtx curr_insn_set;
145 static basic_block curr_bb;
146 static lra_insn_recog_data_t curr_id;
147 static struct lra_static_insn_data *curr_static_id;
148 static machine_mode curr_operand_mode[MAX_RECOG_OPERANDS];
149 /* Mode of the register substituted by its equivalence with VOIDmode
150 (e.g. constant) and whose subreg is given operand of the current
151 insn. VOIDmode in all other cases. */
152 static machine_mode original_subreg_reg_mode[MAX_RECOG_OPERANDS];
156 /* Start numbers for new registers and insns at the current constraints
157 pass start. */
158 static int new_regno_start;
159 static int new_insn_uid_start;
161 /* If LOC is nonnull, strip any outer subreg from it. */
162 static inline rtx *
163 strip_subreg (rtx *loc)
165 return loc && GET_CODE (*loc) == SUBREG ? &SUBREG_REG (*loc) : loc;
168 /* Return hard regno of REGNO or if it is was not assigned to a hard
169 register, use a hard register from its allocno class. */
170 static int
171 get_try_hard_regno (int regno)
173 int hard_regno;
174 enum reg_class rclass;
176 if ((hard_regno = regno) >= FIRST_PSEUDO_REGISTER)
177 hard_regno = lra_get_regno_hard_regno (regno);
178 if (hard_regno >= 0)
179 return hard_regno;
180 rclass = lra_get_allocno_class (regno);
181 if (rclass == NO_REGS)
182 return -1;
183 return ira_class_hard_regs[rclass][0];
186 /* Return the hard regno of X after removing its subreg. If X is not
187 a register or a subreg of a register, return -1. If X is a pseudo,
188 use its assignment. If FINAL_P return the final hard regno which will
189 be after elimination. */
190 static int
191 get_hard_regno (rtx x, bool final_p)
193 rtx reg;
194 int hard_regno;
196 reg = x;
197 if (SUBREG_P (x))
198 reg = SUBREG_REG (x);
199 if (! REG_P (reg))
200 return -1;
201 if (! HARD_REGISTER_NUM_P (hard_regno = REGNO (reg)))
202 hard_regno = lra_get_regno_hard_regno (hard_regno);
203 if (hard_regno < 0)
204 return -1;
205 if (final_p)
206 hard_regno = lra_get_elimination_hard_regno (hard_regno);
207 if (SUBREG_P (x))
208 hard_regno += subreg_regno_offset (hard_regno, GET_MODE (reg),
209 SUBREG_BYTE (x), GET_MODE (x));
210 return hard_regno;
213 /* If REGNO is a hard register or has been allocated a hard register,
214 return the class of that register. If REGNO is a reload pseudo
215 created by the current constraints pass, return its allocno class.
216 Return NO_REGS otherwise. */
217 static enum reg_class
218 get_reg_class (int regno)
220 int hard_regno;
222 if (! HARD_REGISTER_NUM_P (hard_regno = regno))
223 hard_regno = lra_get_regno_hard_regno (regno);
224 if (hard_regno >= 0)
226 hard_regno = lra_get_elimination_hard_regno (hard_regno);
227 return REGNO_REG_CLASS (hard_regno);
229 if (regno >= new_regno_start)
230 return lra_get_allocno_class (regno);
231 return NO_REGS;
234 /* Return true if REG satisfies (or will satisfy) reg class constraint
235 CL. Use elimination first if REG is a hard register. If REG is a
236 reload pseudo created by this constraints pass, assume that it will
237 be allocated a hard register from its allocno class, but allow that
238 class to be narrowed to CL if it is currently a superset of CL.
240 If NEW_CLASS is nonnull, set *NEW_CLASS to the new allocno class of
241 REGNO (reg), or NO_REGS if no change in its class was needed. */
242 static bool
243 in_class_p (rtx reg, enum reg_class cl, enum reg_class *new_class)
245 enum reg_class rclass, common_class;
246 machine_mode reg_mode;
247 int class_size, hard_regno, nregs, i, j;
248 int regno = REGNO (reg);
250 if (new_class != NULL)
251 *new_class = NO_REGS;
252 if (regno < FIRST_PSEUDO_REGISTER)
254 rtx final_reg = reg;
255 rtx *final_loc = &final_reg;
257 lra_eliminate_reg_if_possible (final_loc);
258 return TEST_HARD_REG_BIT (reg_class_contents[cl], REGNO (*final_loc));
260 reg_mode = GET_MODE (reg);
261 rclass = get_reg_class (regno);
262 if (regno < new_regno_start
263 /* Do not allow the constraints for reload instructions to
264 influence the classes of new pseudos. These reloads are
265 typically moves that have many alternatives, and restricting
266 reload pseudos for one alternative may lead to situations
267 where other reload pseudos are no longer allocatable. */
268 || (INSN_UID (curr_insn) >= new_insn_uid_start
269 && curr_insn_set != NULL
270 && ((OBJECT_P (SET_SRC (curr_insn_set))
271 && ! CONSTANT_P (SET_SRC (curr_insn_set)))
272 || (GET_CODE (SET_SRC (curr_insn_set)) == SUBREG
273 && OBJECT_P (SUBREG_REG (SET_SRC (curr_insn_set)))
274 && ! CONSTANT_P (SUBREG_REG (SET_SRC (curr_insn_set)))))))
275 /* When we don't know what class will be used finally for reload
276 pseudos, we use ALL_REGS. */
277 return ((regno >= new_regno_start && rclass == ALL_REGS)
278 || (rclass != NO_REGS && ira_class_subset_p[rclass][cl]
279 && ! hard_reg_set_subset_p (reg_class_contents[cl],
280 lra_no_alloc_regs)));
281 else
283 common_class = ira_reg_class_subset[rclass][cl];
284 if (new_class != NULL)
285 *new_class = common_class;
286 if (hard_reg_set_subset_p (reg_class_contents[common_class],
287 lra_no_alloc_regs))
288 return false;
289 /* Check that there are enough allocatable regs. */
290 class_size = ira_class_hard_regs_num[common_class];
291 for (i = 0; i < class_size; i++)
293 hard_regno = ira_class_hard_regs[common_class][i];
294 nregs = hard_regno_nregs (hard_regno, reg_mode);
295 if (nregs == 1)
296 return true;
297 for (j = 0; j < nregs; j++)
298 if (TEST_HARD_REG_BIT (lra_no_alloc_regs, hard_regno + j)
299 || ! TEST_HARD_REG_BIT (reg_class_contents[common_class],
300 hard_regno + j))
301 break;
302 if (j >= nregs)
303 return true;
305 return false;
309 /* Return true if REGNO satisfies a memory constraint. */
310 static bool
311 in_mem_p (int regno)
313 return get_reg_class (regno) == NO_REGS;
316 /* Return 1 if ADDR is a valid memory address for mode MODE in address
317 space AS, and check that each pseudo has the proper kind of hard
318 reg. */
319 static int
320 valid_address_p (machine_mode mode ATTRIBUTE_UNUSED,
321 rtx addr, addr_space_t as)
323 #ifdef GO_IF_LEGITIMATE_ADDRESS
324 lra_assert (ADDR_SPACE_GENERIC_P (as));
325 GO_IF_LEGITIMATE_ADDRESS (mode, addr, win);
326 return 0;
328 win:
329 return 1;
330 #else
331 return targetm.addr_space.legitimate_address_p (mode, addr, 0, as);
332 #endif
335 namespace {
336 /* Temporarily eliminates registers in an address (for the lifetime of
337 the object). */
338 class address_eliminator {
339 public:
340 address_eliminator (struct address_info *ad);
341 ~address_eliminator ();
343 private:
344 struct address_info *m_ad;
345 rtx *m_base_loc;
346 rtx m_base_reg;
347 rtx *m_index_loc;
348 rtx m_index_reg;
352 address_eliminator::address_eliminator (struct address_info *ad)
353 : m_ad (ad),
354 m_base_loc (strip_subreg (ad->base_term)),
355 m_base_reg (NULL_RTX),
356 m_index_loc (strip_subreg (ad->index_term)),
357 m_index_reg (NULL_RTX)
359 if (m_base_loc != NULL)
361 m_base_reg = *m_base_loc;
362 lra_eliminate_reg_if_possible (m_base_loc);
363 if (m_ad->base_term2 != NULL)
364 *m_ad->base_term2 = *m_ad->base_term;
366 if (m_index_loc != NULL)
368 m_index_reg = *m_index_loc;
369 lra_eliminate_reg_if_possible (m_index_loc);
373 address_eliminator::~address_eliminator ()
375 if (m_base_loc && *m_base_loc != m_base_reg)
377 *m_base_loc = m_base_reg;
378 if (m_ad->base_term2 != NULL)
379 *m_ad->base_term2 = *m_ad->base_term;
381 if (m_index_loc && *m_index_loc != m_index_reg)
382 *m_index_loc = m_index_reg;
385 /* Return true if the eliminated form of AD is a legitimate target address. */
386 static bool
387 valid_address_p (struct address_info *ad)
389 address_eliminator eliminator (ad);
390 return valid_address_p (ad->mode, *ad->outer, ad->as);
393 /* Return true if the eliminated form of memory reference OP satisfies
394 extra (special) memory constraint CONSTRAINT. */
395 static bool
396 satisfies_memory_constraint_p (rtx op, enum constraint_num constraint)
398 struct address_info ad;
400 decompose_mem_address (&ad, op);
401 address_eliminator eliminator (&ad);
402 return constraint_satisfied_p (op, constraint);
405 /* Return true if the eliminated form of address AD satisfies extra
406 address constraint CONSTRAINT. */
407 static bool
408 satisfies_address_constraint_p (struct address_info *ad,
409 enum constraint_num constraint)
411 address_eliminator eliminator (ad);
412 return constraint_satisfied_p (*ad->outer, constraint);
415 /* Return true if the eliminated form of address OP satisfies extra
416 address constraint CONSTRAINT. */
417 static bool
418 satisfies_address_constraint_p (rtx op, enum constraint_num constraint)
420 struct address_info ad;
422 decompose_lea_address (&ad, &op);
423 return satisfies_address_constraint_p (&ad, constraint);
426 /* Initiate equivalences for LRA. As we keep original equivalences
427 before any elimination, we need to make copies otherwise any change
428 in insns might change the equivalences. */
429 void
430 lra_init_equiv (void)
432 ira_expand_reg_equiv ();
433 for (int i = FIRST_PSEUDO_REGISTER; i < max_reg_num (); i++)
435 rtx res;
437 if ((res = ira_reg_equiv[i].memory) != NULL_RTX)
438 ira_reg_equiv[i].memory = copy_rtx (res);
439 if ((res = ira_reg_equiv[i].invariant) != NULL_RTX)
440 ira_reg_equiv[i].invariant = copy_rtx (res);
444 static rtx loc_equivalence_callback (rtx, const_rtx, void *);
446 /* Update equivalence for REGNO. We need to this as the equivalence
447 might contain other pseudos which are changed by their
448 equivalences. */
449 static void
450 update_equiv (int regno)
452 rtx x;
454 if ((x = ira_reg_equiv[regno].memory) != NULL_RTX)
455 ira_reg_equiv[regno].memory
456 = simplify_replace_fn_rtx (x, NULL_RTX, loc_equivalence_callback,
457 NULL_RTX);
458 if ((x = ira_reg_equiv[regno].invariant) != NULL_RTX)
459 ira_reg_equiv[regno].invariant
460 = simplify_replace_fn_rtx (x, NULL_RTX, loc_equivalence_callback,
461 NULL_RTX);
464 /* If we have decided to substitute X with another value, return that
465 value, otherwise return X. */
466 static rtx
467 get_equiv (rtx x)
469 int regno;
470 rtx res;
472 if (! REG_P (x) || (regno = REGNO (x)) < FIRST_PSEUDO_REGISTER
473 || ! ira_reg_equiv[regno].defined_p
474 || ! ira_reg_equiv[regno].profitable_p
475 || lra_get_regno_hard_regno (regno) >= 0)
476 return x;
477 if ((res = ira_reg_equiv[regno].memory) != NULL_RTX)
479 if (targetm.cannot_substitute_mem_equiv_p (res))
480 return x;
481 return res;
483 if ((res = ira_reg_equiv[regno].constant) != NULL_RTX)
484 return res;
485 if ((res = ira_reg_equiv[regno].invariant) != NULL_RTX)
486 return res;
487 gcc_unreachable ();
490 /* If we have decided to substitute X with the equivalent value,
491 return that value after elimination for INSN, otherwise return
492 X. */
493 static rtx
494 get_equiv_with_elimination (rtx x, rtx_insn *insn)
496 rtx res = get_equiv (x);
498 if (x == res || CONSTANT_P (res))
499 return res;
500 return lra_eliminate_regs_1 (insn, res, GET_MODE (res),
501 false, false, 0, true);
504 /* Set up curr_operand_mode. */
505 static void
506 init_curr_operand_mode (void)
508 int nop = curr_static_id->n_operands;
509 for (int i = 0; i < nop; i++)
511 machine_mode mode = GET_MODE (*curr_id->operand_loc[i]);
512 if (mode == VOIDmode)
514 /* The .md mode for address operands is the mode of the
515 addressed value rather than the mode of the address itself. */
516 if (curr_id->icode >= 0 && curr_static_id->operand[i].is_address)
517 mode = Pmode;
518 else
519 mode = curr_static_id->operand[i].mode;
521 curr_operand_mode[i] = mode;
527 /* The page contains code to reuse input reloads. */
529 /* Structure describes input reload of the current insns. */
530 struct input_reload
532 /* True for input reload of matched operands. */
533 bool match_p;
534 /* Reloaded value. */
535 rtx input;
536 /* Reload pseudo used. */
537 rtx reg;
540 /* The number of elements in the following array. */
541 static int curr_insn_input_reloads_num;
542 /* Array containing info about input reloads. It is used to find the
543 same input reload and reuse the reload pseudo in this case. */
544 static struct input_reload curr_insn_input_reloads[LRA_MAX_INSN_RELOADS];
546 /* Initiate data concerning reuse of input reloads for the current
547 insn. */
548 static void
549 init_curr_insn_input_reloads (void)
551 curr_insn_input_reloads_num = 0;
554 /* Create a new pseudo using MODE, RCLASS, ORIGINAL or reuse already
555 created input reload pseudo (only if TYPE is not OP_OUT). Don't
556 reuse pseudo if IN_SUBREG_P is true and the reused pseudo should be
557 wrapped up in SUBREG. The result pseudo is returned through
558 RESULT_REG. Return TRUE if we created a new pseudo, FALSE if we
559 reused the already created input reload pseudo. Use TITLE to
560 describe new registers for debug purposes. */
561 static bool
562 get_reload_reg (enum op_type type, machine_mode mode, rtx original,
563 enum reg_class rclass, bool in_subreg_p,
564 const char *title, rtx *result_reg)
566 int i, regno;
567 enum reg_class new_class;
568 bool unique_p = false;
570 if (type == OP_OUT)
572 *result_reg
573 = lra_create_new_reg_with_unique_value (mode, original, rclass, title);
574 return true;
576 /* Prevent reuse value of expression with side effects,
577 e.g. volatile memory. */
578 if (! side_effects_p (original))
579 for (i = 0; i < curr_insn_input_reloads_num; i++)
581 if (! curr_insn_input_reloads[i].match_p
582 && rtx_equal_p (curr_insn_input_reloads[i].input, original)
583 && in_class_p (curr_insn_input_reloads[i].reg, rclass, &new_class))
585 rtx reg = curr_insn_input_reloads[i].reg;
586 regno = REGNO (reg);
587 /* If input is equal to original and both are VOIDmode,
588 GET_MODE (reg) might be still different from mode.
589 Ensure we don't return *result_reg with wrong mode. */
590 if (GET_MODE (reg) != mode)
592 if (in_subreg_p)
593 continue;
594 if (GET_MODE_SIZE (GET_MODE (reg)) < GET_MODE_SIZE (mode))
595 continue;
596 reg = lowpart_subreg (mode, reg, GET_MODE (reg));
597 if (reg == NULL_RTX || GET_CODE (reg) != SUBREG)
598 continue;
600 *result_reg = reg;
601 if (lra_dump_file != NULL)
603 fprintf (lra_dump_file, " Reuse r%d for reload ", regno);
604 dump_value_slim (lra_dump_file, original, 1);
606 if (new_class != lra_get_allocno_class (regno))
607 lra_change_class (regno, new_class, ", change to", false);
608 if (lra_dump_file != NULL)
609 fprintf (lra_dump_file, "\n");
610 return false;
612 /* If we have an input reload with a different mode, make sure it
613 will get a different hard reg. */
614 else if (REG_P (original)
615 && REG_P (curr_insn_input_reloads[i].input)
616 && REGNO (original) == REGNO (curr_insn_input_reloads[i].input)
617 && (GET_MODE (original)
618 != GET_MODE (curr_insn_input_reloads[i].input)))
619 unique_p = true;
621 *result_reg = (unique_p
622 ? lra_create_new_reg_with_unique_value
623 : lra_create_new_reg) (mode, original, rclass, title);
624 lra_assert (curr_insn_input_reloads_num < LRA_MAX_INSN_RELOADS);
625 curr_insn_input_reloads[curr_insn_input_reloads_num].input = original;
626 curr_insn_input_reloads[curr_insn_input_reloads_num].match_p = false;
627 curr_insn_input_reloads[curr_insn_input_reloads_num++].reg = *result_reg;
628 return true;
633 /* The page contains code to extract memory address parts. */
635 /* Wrapper around REGNO_OK_FOR_INDEX_P, to allow pseudos. */
636 static inline bool
637 ok_for_index_p_nonstrict (rtx reg)
639 unsigned regno = REGNO (reg);
641 return regno >= FIRST_PSEUDO_REGISTER || REGNO_OK_FOR_INDEX_P (regno);
644 /* A version of regno_ok_for_base_p for use here, when all pseudos
645 should count as OK. Arguments as for regno_ok_for_base_p. */
646 static inline bool
647 ok_for_base_p_nonstrict (rtx reg, machine_mode mode, addr_space_t as,
648 enum rtx_code outer_code, enum rtx_code index_code)
650 unsigned regno = REGNO (reg);
652 if (regno >= FIRST_PSEUDO_REGISTER)
653 return true;
654 return ok_for_base_p_1 (regno, mode, as, outer_code, index_code);
659 /* The page contains major code to choose the current insn alternative
660 and generate reloads for it. */
662 /* Return the offset from REGNO of the least significant register
663 in (reg:MODE REGNO).
665 This function is used to tell whether two registers satisfy
666 a matching constraint. (reg:MODE1 REGNO1) matches (reg:MODE2 REGNO2) if:
668 REGNO1 + lra_constraint_offset (REGNO1, MODE1)
669 == REGNO2 + lra_constraint_offset (REGNO2, MODE2) */
671 lra_constraint_offset (int regno, machine_mode mode)
673 lra_assert (regno < FIRST_PSEUDO_REGISTER);
675 scalar_int_mode int_mode;
676 if (WORDS_BIG_ENDIAN
677 && is_a <scalar_int_mode> (mode, &int_mode)
678 && GET_MODE_SIZE (int_mode) > UNITS_PER_WORD)
679 return hard_regno_nregs (regno, mode) - 1;
680 return 0;
683 /* Like rtx_equal_p except that it allows a REG and a SUBREG to match
684 if they are the same hard reg, and has special hacks for
685 auto-increment and auto-decrement. This is specifically intended for
686 process_alt_operands to use in determining whether two operands
687 match. X is the operand whose number is the lower of the two.
689 It is supposed that X is the output operand and Y is the input
690 operand. Y_HARD_REGNO is the final hard regno of register Y or
691 register in subreg Y as we know it now. Otherwise, it is a
692 negative value. */
693 static bool
694 operands_match_p (rtx x, rtx y, int y_hard_regno)
696 int i;
697 RTX_CODE code = GET_CODE (x);
698 const char *fmt;
700 if (x == y)
701 return true;
702 if ((code == REG || (code == SUBREG && REG_P (SUBREG_REG (x))))
703 && (REG_P (y) || (GET_CODE (y) == SUBREG && REG_P (SUBREG_REG (y)))))
705 int j;
707 i = get_hard_regno (x, false);
708 if (i < 0)
709 goto slow;
711 if ((j = y_hard_regno) < 0)
712 goto slow;
714 i += lra_constraint_offset (i, GET_MODE (x));
715 j += lra_constraint_offset (j, GET_MODE (y));
717 return i == j;
720 /* If two operands must match, because they are really a single
721 operand of an assembler insn, then two post-increments are invalid
722 because the assembler insn would increment only once. On the
723 other hand, a post-increment matches ordinary indexing if the
724 post-increment is the output operand. */
725 if (code == POST_DEC || code == POST_INC || code == POST_MODIFY)
726 return operands_match_p (XEXP (x, 0), y, y_hard_regno);
728 /* Two pre-increments are invalid because the assembler insn would
729 increment only once. On the other hand, a pre-increment matches
730 ordinary indexing if the pre-increment is the input operand. */
731 if (GET_CODE (y) == PRE_DEC || GET_CODE (y) == PRE_INC
732 || GET_CODE (y) == PRE_MODIFY)
733 return operands_match_p (x, XEXP (y, 0), -1);
735 slow:
737 if (code == REG && REG_P (y))
738 return REGNO (x) == REGNO (y);
740 if (code == REG && GET_CODE (y) == SUBREG && REG_P (SUBREG_REG (y))
741 && x == SUBREG_REG (y))
742 return true;
743 if (GET_CODE (y) == REG && code == SUBREG && REG_P (SUBREG_REG (x))
744 && SUBREG_REG (x) == y)
745 return true;
747 /* Now we have disposed of all the cases in which different rtx
748 codes can match. */
749 if (code != GET_CODE (y))
750 return false;
752 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
753 if (GET_MODE (x) != GET_MODE (y))
754 return false;
756 switch (code)
758 CASE_CONST_UNIQUE:
759 return false;
761 case LABEL_REF:
762 return label_ref_label (x) == label_ref_label (y);
763 case SYMBOL_REF:
764 return XSTR (x, 0) == XSTR (y, 0);
766 default:
767 break;
770 /* Compare the elements. If any pair of corresponding elements fail
771 to match, return false for the whole things. */
773 fmt = GET_RTX_FORMAT (code);
774 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
776 int val, j;
777 switch (fmt[i])
779 case 'w':
780 if (XWINT (x, i) != XWINT (y, i))
781 return false;
782 break;
784 case 'i':
785 if (XINT (x, i) != XINT (y, i))
786 return false;
787 break;
789 case 'e':
790 val = operands_match_p (XEXP (x, i), XEXP (y, i), -1);
791 if (val == 0)
792 return false;
793 break;
795 case '0':
796 break;
798 case 'E':
799 if (XVECLEN (x, i) != XVECLEN (y, i))
800 return false;
801 for (j = XVECLEN (x, i) - 1; j >= 0; --j)
803 val = operands_match_p (XVECEXP (x, i, j), XVECEXP (y, i, j), -1);
804 if (val == 0)
805 return false;
807 break;
809 /* It is believed that rtx's at this level will never
810 contain anything but integers and other rtx's, except for
811 within LABEL_REFs and SYMBOL_REFs. */
812 default:
813 gcc_unreachable ();
816 return true;
819 /* True if X is a constant that can be forced into the constant pool.
820 MODE is the mode of the operand, or VOIDmode if not known. */
821 #define CONST_POOL_OK_P(MODE, X) \
822 ((MODE) != VOIDmode \
823 && CONSTANT_P (X) \
824 && GET_CODE (X) != HIGH \
825 && !targetm.cannot_force_const_mem (MODE, X))
827 /* True if C is a non-empty register class that has too few registers
828 to be safely used as a reload target class. */
829 #define SMALL_REGISTER_CLASS_P(C) \
830 (ira_class_hard_regs_num [(C)] == 1 \
831 || (ira_class_hard_regs_num [(C)] >= 1 \
832 && targetm.class_likely_spilled_p (C)))
834 /* If REG is a reload pseudo, try to make its class satisfying CL. */
835 static void
836 narrow_reload_pseudo_class (rtx reg, enum reg_class cl)
838 enum reg_class rclass;
840 /* Do not make more accurate class from reloads generated. They are
841 mostly moves with a lot of constraints. Making more accurate
842 class may results in very narrow class and impossibility of find
843 registers for several reloads of one insn. */
844 if (INSN_UID (curr_insn) >= new_insn_uid_start)
845 return;
846 if (GET_CODE (reg) == SUBREG)
847 reg = SUBREG_REG (reg);
848 if (! REG_P (reg) || (int) REGNO (reg) < new_regno_start)
849 return;
850 if (in_class_p (reg, cl, &rclass) && rclass != cl)
851 lra_change_class (REGNO (reg), rclass, " Change to", true);
854 /* Searches X for any reference to a reg with the same value as REGNO,
855 returning the rtx of the reference found if any. Otherwise,
856 returns NULL_RTX. */
857 static rtx
858 regno_val_use_in (unsigned int regno, rtx x)
860 const char *fmt;
861 int i, j;
862 rtx tem;
864 if (REG_P (x) && lra_reg_info[REGNO (x)].val == lra_reg_info[regno].val)
865 return x;
867 fmt = GET_RTX_FORMAT (GET_CODE (x));
868 for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
870 if (fmt[i] == 'e')
872 if ((tem = regno_val_use_in (regno, XEXP (x, i))))
873 return tem;
875 else if (fmt[i] == 'E')
876 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
877 if ((tem = regno_val_use_in (regno , XVECEXP (x, i, j))))
878 return tem;
881 return NULL_RTX;
884 /* Return true if all current insn non-output operands except INS (it
885 has a negaitve end marker) do not use pseudos with the same value
886 as REGNO. */
887 static bool
888 check_conflict_input_operands (int regno, signed char *ins)
890 int in;
891 int n_operands = curr_static_id->n_operands;
893 for (int nop = 0; nop < n_operands; nop++)
894 if (! curr_static_id->operand[nop].is_operator
895 && curr_static_id->operand[nop].type != OP_OUT)
897 for (int i = 0; (in = ins[i]) >= 0; i++)
898 if (in == nop)
899 break;
900 if (in < 0
901 && regno_val_use_in (regno, *curr_id->operand_loc[nop]) != NULL_RTX)
902 return false;
904 return true;
907 /* Generate reloads for matching OUT and INS (array of input operand
908 numbers with end marker -1) with reg class GOAL_CLASS, considering
909 output operands OUTS (similar array to INS) needing to be in different
910 registers. Add input and output reloads correspondingly to the lists
911 *BEFORE and *AFTER. OUT might be negative. In this case we generate
912 input reloads for matched input operands INS. EARLY_CLOBBER_P is a flag
913 that the output operand is early clobbered for chosen alternative. */
914 static void
915 match_reload (signed char out, signed char *ins, signed char *outs,
916 enum reg_class goal_class, rtx_insn **before,
917 rtx_insn **after, bool early_clobber_p)
919 bool out_conflict;
920 int i, in;
921 rtx new_in_reg, new_out_reg, reg;
922 machine_mode inmode, outmode;
923 rtx in_rtx = *curr_id->operand_loc[ins[0]];
924 rtx out_rtx = out < 0 ? in_rtx : *curr_id->operand_loc[out];
926 inmode = curr_operand_mode[ins[0]];
927 outmode = out < 0 ? inmode : curr_operand_mode[out];
928 push_to_sequence (*before);
929 if (inmode != outmode)
931 if (partial_subreg_p (outmode, inmode))
933 reg = new_in_reg
934 = lra_create_new_reg_with_unique_value (inmode, in_rtx,
935 goal_class, "");
936 if (SCALAR_INT_MODE_P (inmode))
937 new_out_reg = gen_lowpart_SUBREG (outmode, reg);
938 else
939 new_out_reg = gen_rtx_SUBREG (outmode, reg, 0);
940 LRA_SUBREG_P (new_out_reg) = 1;
941 /* If the input reg is dying here, we can use the same hard
942 register for REG and IN_RTX. We do it only for original
943 pseudos as reload pseudos can die although original
944 pseudos still live where reload pseudos dies. */
945 if (REG_P (in_rtx) && (int) REGNO (in_rtx) < lra_new_regno_start
946 && find_regno_note (curr_insn, REG_DEAD, REGNO (in_rtx))
947 && (!early_clobber_p
948 || check_conflict_input_operands(REGNO (in_rtx), ins)))
949 lra_assign_reg_val (REGNO (in_rtx), REGNO (reg));
951 else
953 reg = new_out_reg
954 = lra_create_new_reg_with_unique_value (outmode, out_rtx,
955 goal_class, "");
956 if (SCALAR_INT_MODE_P (outmode))
957 new_in_reg = gen_lowpart_SUBREG (inmode, reg);
958 else
959 new_in_reg = gen_rtx_SUBREG (inmode, reg, 0);
960 /* NEW_IN_REG is non-paradoxical subreg. We don't want
961 NEW_OUT_REG living above. We add clobber clause for
962 this. This is just a temporary clobber. We can remove
963 it at the end of LRA work. */
964 rtx_insn *clobber = emit_clobber (new_out_reg);
965 LRA_TEMP_CLOBBER_P (PATTERN (clobber)) = 1;
966 LRA_SUBREG_P (new_in_reg) = 1;
967 if (GET_CODE (in_rtx) == SUBREG)
969 rtx subreg_reg = SUBREG_REG (in_rtx);
971 /* If SUBREG_REG is dying here and sub-registers IN_RTX
972 and NEW_IN_REG are similar, we can use the same hard
973 register for REG and SUBREG_REG. */
974 if (REG_P (subreg_reg)
975 && (int) REGNO (subreg_reg) < lra_new_regno_start
976 && GET_MODE (subreg_reg) == outmode
977 && SUBREG_BYTE (in_rtx) == SUBREG_BYTE (new_in_reg)
978 && find_regno_note (curr_insn, REG_DEAD, REGNO (subreg_reg))
979 && (! early_clobber_p
980 || check_conflict_input_operands (REGNO (subreg_reg),
981 ins)))
982 lra_assign_reg_val (REGNO (subreg_reg), REGNO (reg));
986 else
988 /* Pseudos have values -- see comments for lra_reg_info.
989 Different pseudos with the same value do not conflict even if
990 they live in the same place. When we create a pseudo we
991 assign value of original pseudo (if any) from which we
992 created the new pseudo. If we create the pseudo from the
993 input pseudo, the new pseudo will have no conflict with the
994 input pseudo which is wrong when the input pseudo lives after
995 the insn and as the new pseudo value is changed by the insn
996 output. Therefore we create the new pseudo from the output
997 except the case when we have single matched dying input
998 pseudo.
1000 We cannot reuse the current output register because we might
1001 have a situation like "a <- a op b", where the constraints
1002 force the second input operand ("b") to match the output
1003 operand ("a"). "b" must then be copied into a new register
1004 so that it doesn't clobber the current value of "a".
1006 We can not use the same value if the output pseudo is
1007 early clobbered or the input pseudo is mentioned in the
1008 output, e.g. as an address part in memory, because
1009 output reload will actually extend the pseudo liveness.
1010 We don't care about eliminable hard regs here as we are
1011 interesting only in pseudos. */
1013 /* Matching input's register value is the same as one of the other
1014 output operand. Output operands in a parallel insn must be in
1015 different registers. */
1016 out_conflict = false;
1017 if (REG_P (in_rtx))
1019 for (i = 0; outs[i] >= 0; i++)
1021 rtx other_out_rtx = *curr_id->operand_loc[outs[i]];
1022 if (REG_P (other_out_rtx)
1023 && (regno_val_use_in (REGNO (in_rtx), other_out_rtx)
1024 != NULL_RTX))
1026 out_conflict = true;
1027 break;
1032 new_in_reg = new_out_reg
1033 = (! early_clobber_p && ins[1] < 0 && REG_P (in_rtx)
1034 && (int) REGNO (in_rtx) < lra_new_regno_start
1035 && find_regno_note (curr_insn, REG_DEAD, REGNO (in_rtx))
1036 && (! early_clobber_p
1037 || check_conflict_input_operands (REGNO (in_rtx), ins))
1038 && (out < 0
1039 || regno_val_use_in (REGNO (in_rtx), out_rtx) == NULL_RTX)
1040 && !out_conflict
1041 ? lra_create_new_reg (inmode, in_rtx, goal_class, "")
1042 : lra_create_new_reg_with_unique_value (outmode, out_rtx,
1043 goal_class, ""));
1045 /* In operand can be got from transformations before processing insn
1046 constraints. One example of such transformations is subreg
1047 reloading (see function simplify_operand_subreg). The new
1048 pseudos created by the transformations might have inaccurate
1049 class (ALL_REGS) and we should make their classes more
1050 accurate. */
1051 narrow_reload_pseudo_class (in_rtx, goal_class);
1052 lra_emit_move (copy_rtx (new_in_reg), in_rtx);
1053 *before = get_insns ();
1054 end_sequence ();
1055 /* Add the new pseudo to consider values of subsequent input reload
1056 pseudos. */
1057 lra_assert (curr_insn_input_reloads_num < LRA_MAX_INSN_RELOADS);
1058 curr_insn_input_reloads[curr_insn_input_reloads_num].input = in_rtx;
1059 curr_insn_input_reloads[curr_insn_input_reloads_num].match_p = true;
1060 curr_insn_input_reloads[curr_insn_input_reloads_num++].reg = new_in_reg;
1061 for (i = 0; (in = ins[i]) >= 0; i++)
1063 lra_assert
1064 (GET_MODE (*curr_id->operand_loc[in]) == VOIDmode
1065 || GET_MODE (new_in_reg) == GET_MODE (*curr_id->operand_loc[in]));
1066 *curr_id->operand_loc[in] = new_in_reg;
1068 lra_update_dups (curr_id, ins);
1069 if (out < 0)
1070 return;
1071 /* See a comment for the input operand above. */
1072 narrow_reload_pseudo_class (out_rtx, goal_class);
1073 if (find_reg_note (curr_insn, REG_UNUSED, out_rtx) == NULL_RTX)
1075 start_sequence ();
1076 lra_emit_move (out_rtx, copy_rtx (new_out_reg));
1077 emit_insn (*after);
1078 *after = get_insns ();
1079 end_sequence ();
1081 *curr_id->operand_loc[out] = new_out_reg;
1082 lra_update_dup (curr_id, out);
1085 /* Return register class which is union of all reg classes in insn
1086 constraint alternative string starting with P. */
1087 static enum reg_class
1088 reg_class_from_constraints (const char *p)
1090 int c, len;
1091 enum reg_class op_class = NO_REGS;
1094 switch ((c = *p, len = CONSTRAINT_LEN (c, p)), c)
1096 case '#':
1097 case ',':
1098 return op_class;
1100 case 'g':
1101 op_class = reg_class_subunion[op_class][GENERAL_REGS];
1102 break;
1104 default:
1105 enum constraint_num cn = lookup_constraint (p);
1106 enum reg_class cl = reg_class_for_constraint (cn);
1107 if (cl == NO_REGS)
1109 if (insn_extra_address_constraint (cn))
1110 op_class
1111 = (reg_class_subunion
1112 [op_class][base_reg_class (VOIDmode, ADDR_SPACE_GENERIC,
1113 ADDRESS, SCRATCH)]);
1114 break;
1117 op_class = reg_class_subunion[op_class][cl];
1118 break;
1120 while ((p += len), c);
1121 return op_class;
1124 /* If OP is a register, return the class of the register as per
1125 get_reg_class, otherwise return NO_REGS. */
1126 static inline enum reg_class
1127 get_op_class (rtx op)
1129 return REG_P (op) ? get_reg_class (REGNO (op)) : NO_REGS;
1132 /* Return generated insn mem_pseudo:=val if TO_P or val:=mem_pseudo
1133 otherwise. If modes of MEM_PSEUDO and VAL are different, use
1134 SUBREG for VAL to make them equal. */
1135 static rtx_insn *
1136 emit_spill_move (bool to_p, rtx mem_pseudo, rtx val)
1138 if (GET_MODE (mem_pseudo) != GET_MODE (val))
1140 /* Usually size of mem_pseudo is greater than val size but in
1141 rare cases it can be less as it can be defined by target
1142 dependent macro HARD_REGNO_CALLER_SAVE_MODE. */
1143 if (! MEM_P (val))
1145 val = gen_lowpart_SUBREG (GET_MODE (mem_pseudo),
1146 GET_CODE (val) == SUBREG
1147 ? SUBREG_REG (val) : val);
1148 LRA_SUBREG_P (val) = 1;
1150 else
1152 mem_pseudo = gen_lowpart_SUBREG (GET_MODE (val), mem_pseudo);
1153 LRA_SUBREG_P (mem_pseudo) = 1;
1156 return to_p ? gen_move_insn (mem_pseudo, val)
1157 : gen_move_insn (val, mem_pseudo);
1160 /* Process a special case insn (register move), return true if we
1161 don't need to process it anymore. INSN should be a single set
1162 insn. Set up that RTL was changed through CHANGE_P and that hook
1163 TARGET_SECONDARY_MEMORY_NEEDED says to use secondary memory through
1164 SEC_MEM_P. */
1165 static bool
1166 check_and_process_move (bool *change_p, bool *sec_mem_p ATTRIBUTE_UNUSED)
1168 int sregno, dregno;
1169 rtx dest, src, dreg, sreg, new_reg, scratch_reg;
1170 rtx_insn *before;
1171 enum reg_class dclass, sclass, secondary_class;
1172 secondary_reload_info sri;
1174 lra_assert (curr_insn_set != NULL_RTX);
1175 dreg = dest = SET_DEST (curr_insn_set);
1176 sreg = src = SET_SRC (curr_insn_set);
1177 if (GET_CODE (dest) == SUBREG)
1178 dreg = SUBREG_REG (dest);
1179 if (GET_CODE (src) == SUBREG)
1180 sreg = SUBREG_REG (src);
1181 if (! (REG_P (dreg) || MEM_P (dreg)) || ! (REG_P (sreg) || MEM_P (sreg)))
1182 return false;
1183 sclass = dclass = NO_REGS;
1184 if (REG_P (dreg))
1185 dclass = get_reg_class (REGNO (dreg));
1186 gcc_assert (dclass < LIM_REG_CLASSES);
1187 if (dclass == ALL_REGS)
1188 /* ALL_REGS is used for new pseudos created by transformations
1189 like reload of SUBREG_REG (see function
1190 simplify_operand_subreg). We don't know their class yet. We
1191 should figure out the class from processing the insn
1192 constraints not in this fast path function. Even if ALL_REGS
1193 were a right class for the pseudo, secondary_... hooks usually
1194 are not define for ALL_REGS. */
1195 return false;
1196 if (REG_P (sreg))
1197 sclass = get_reg_class (REGNO (sreg));
1198 gcc_assert (sclass < LIM_REG_CLASSES);
1199 if (sclass == ALL_REGS)
1200 /* See comments above. */
1201 return false;
1202 if (sclass == NO_REGS && dclass == NO_REGS)
1203 return false;
1204 if (targetm.secondary_memory_needed (GET_MODE (src), sclass, dclass)
1205 && ((sclass != NO_REGS && dclass != NO_REGS)
1206 || (GET_MODE (src)
1207 != targetm.secondary_memory_needed_mode (GET_MODE (src)))))
1209 *sec_mem_p = true;
1210 return false;
1212 if (! REG_P (dreg) || ! REG_P (sreg))
1213 return false;
1214 sri.prev_sri = NULL;
1215 sri.icode = CODE_FOR_nothing;
1216 sri.extra_cost = 0;
1217 secondary_class = NO_REGS;
1218 /* Set up hard register for a reload pseudo for hook
1219 secondary_reload because some targets just ignore unassigned
1220 pseudos in the hook. */
1221 if (dclass != NO_REGS && lra_get_regno_hard_regno (REGNO (dreg)) < 0)
1223 dregno = REGNO (dreg);
1224 reg_renumber[dregno] = ira_class_hard_regs[dclass][0];
1226 else
1227 dregno = -1;
1228 if (sclass != NO_REGS && lra_get_regno_hard_regno (REGNO (sreg)) < 0)
1230 sregno = REGNO (sreg);
1231 reg_renumber[sregno] = ira_class_hard_regs[sclass][0];
1233 else
1234 sregno = -1;
1235 if (sclass != NO_REGS)
1236 secondary_class
1237 = (enum reg_class) targetm.secondary_reload (false, dest,
1238 (reg_class_t) sclass,
1239 GET_MODE (src), &sri);
1240 if (sclass == NO_REGS
1241 || ((secondary_class != NO_REGS || sri.icode != CODE_FOR_nothing)
1242 && dclass != NO_REGS))
1244 enum reg_class old_sclass = secondary_class;
1245 secondary_reload_info old_sri = sri;
1247 sri.prev_sri = NULL;
1248 sri.icode = CODE_FOR_nothing;
1249 sri.extra_cost = 0;
1250 secondary_class
1251 = (enum reg_class) targetm.secondary_reload (true, src,
1252 (reg_class_t) dclass,
1253 GET_MODE (src), &sri);
1254 /* Check the target hook consistency. */
1255 lra_assert
1256 ((secondary_class == NO_REGS && sri.icode == CODE_FOR_nothing)
1257 || (old_sclass == NO_REGS && old_sri.icode == CODE_FOR_nothing)
1258 || (secondary_class == old_sclass && sri.icode == old_sri.icode));
1260 if (sregno >= 0)
1261 reg_renumber [sregno] = -1;
1262 if (dregno >= 0)
1263 reg_renumber [dregno] = -1;
1264 if (secondary_class == NO_REGS && sri.icode == CODE_FOR_nothing)
1265 return false;
1266 *change_p = true;
1267 new_reg = NULL_RTX;
1268 if (secondary_class != NO_REGS)
1269 new_reg = lra_create_new_reg_with_unique_value (GET_MODE (src), NULL_RTX,
1270 secondary_class,
1271 "secondary");
1272 start_sequence ();
1273 if (sri.icode == CODE_FOR_nothing)
1274 lra_emit_move (new_reg, src);
1275 else
1277 enum reg_class scratch_class;
1279 scratch_class = (reg_class_from_constraints
1280 (insn_data[sri.icode].operand[2].constraint));
1281 scratch_reg = (lra_create_new_reg_with_unique_value
1282 (insn_data[sri.icode].operand[2].mode, NULL_RTX,
1283 scratch_class, "scratch"));
1284 emit_insn (GEN_FCN (sri.icode) (new_reg != NULL_RTX ? new_reg : dest,
1285 src, scratch_reg));
1287 before = get_insns ();
1288 end_sequence ();
1289 lra_process_new_insns (curr_insn, before, NULL, "Inserting the move");
1290 if (new_reg != NULL_RTX)
1291 SET_SRC (curr_insn_set) = new_reg;
1292 else
1294 if (lra_dump_file != NULL)
1296 fprintf (lra_dump_file, "Deleting move %u\n", INSN_UID (curr_insn));
1297 dump_insn_slim (lra_dump_file, curr_insn);
1299 lra_set_insn_deleted (curr_insn);
1300 return true;
1302 return false;
1305 /* The following data describe the result of process_alt_operands.
1306 The data are used in curr_insn_transform to generate reloads. */
1308 /* The chosen reg classes which should be used for the corresponding
1309 operands. */
1310 static enum reg_class goal_alt[MAX_RECOG_OPERANDS];
1311 /* True if the operand should be the same as another operand and that
1312 other operand does not need a reload. */
1313 static bool goal_alt_match_win[MAX_RECOG_OPERANDS];
1314 /* True if the operand does not need a reload. */
1315 static bool goal_alt_win[MAX_RECOG_OPERANDS];
1316 /* True if the operand can be offsetable memory. */
1317 static bool goal_alt_offmemok[MAX_RECOG_OPERANDS];
1318 /* The number of an operand to which given operand can be matched to. */
1319 static int goal_alt_matches[MAX_RECOG_OPERANDS];
1320 /* The number of elements in the following array. */
1321 static int goal_alt_dont_inherit_ops_num;
1322 /* Numbers of operands whose reload pseudos should not be inherited. */
1323 static int goal_alt_dont_inherit_ops[MAX_RECOG_OPERANDS];
1324 /* True if the insn commutative operands should be swapped. */
1325 static bool goal_alt_swapped;
1326 /* The chosen insn alternative. */
1327 static int goal_alt_number;
1329 /* True if the corresponding operand is the result of an equivalence
1330 substitution. */
1331 static bool equiv_substition_p[MAX_RECOG_OPERANDS];
1333 /* The following five variables are used to choose the best insn
1334 alternative. They reflect final characteristics of the best
1335 alternative. */
1337 /* Number of necessary reloads and overall cost reflecting the
1338 previous value and other unpleasantness of the best alternative. */
1339 static int best_losers, best_overall;
1340 /* Overall number hard registers used for reloads. For example, on
1341 some targets we need 2 general registers to reload DFmode and only
1342 one floating point register. */
1343 static int best_reload_nregs;
1344 /* Overall number reflecting distances of previous reloading the same
1345 value. The distances are counted from the current BB start. It is
1346 used to improve inheritance chances. */
1347 static int best_reload_sum;
1349 /* True if the current insn should have no correspondingly input or
1350 output reloads. */
1351 static bool no_input_reloads_p, no_output_reloads_p;
1353 /* True if we swapped the commutative operands in the current
1354 insn. */
1355 static int curr_swapped;
1357 /* if CHECK_ONLY_P is false, arrange for address element *LOC to be a
1358 register of class CL. Add any input reloads to list BEFORE. AFTER
1359 is nonnull if *LOC is an automodified value; handle that case by
1360 adding the required output reloads to list AFTER. Return true if
1361 the RTL was changed.
1363 if CHECK_ONLY_P is true, check that the *LOC is a correct address
1364 register. Return false if the address register is correct. */
1365 static bool
1366 process_addr_reg (rtx *loc, bool check_only_p, rtx_insn **before, rtx_insn **after,
1367 enum reg_class cl)
1369 int regno;
1370 enum reg_class rclass, new_class;
1371 rtx reg;
1372 rtx new_reg;
1373 machine_mode mode;
1374 bool subreg_p, before_p = false;
1376 subreg_p = GET_CODE (*loc) == SUBREG;
1377 if (subreg_p)
1379 reg = SUBREG_REG (*loc);
1380 mode = GET_MODE (reg);
1382 /* For mode with size bigger than ptr_mode, there unlikely to be "mov"
1383 between two registers with different classes, but there normally will
1384 be "mov" which transfers element of vector register into the general
1385 register, and this normally will be a subreg which should be reloaded
1386 as a whole. This is particularly likely to be triggered when
1387 -fno-split-wide-types specified. */
1388 if (!REG_P (reg)
1389 || in_class_p (reg, cl, &new_class)
1390 || GET_MODE_SIZE (mode) <= GET_MODE_SIZE (ptr_mode))
1391 loc = &SUBREG_REG (*loc);
1394 reg = *loc;
1395 mode = GET_MODE (reg);
1396 if (! REG_P (reg))
1398 if (check_only_p)
1399 return true;
1400 /* Always reload memory in an address even if the target supports
1401 such addresses. */
1402 new_reg = lra_create_new_reg_with_unique_value (mode, reg, cl, "address");
1403 before_p = true;
1405 else
1407 regno = REGNO (reg);
1408 rclass = get_reg_class (regno);
1409 if (! check_only_p
1410 && (*loc = get_equiv_with_elimination (reg, curr_insn)) != reg)
1412 if (lra_dump_file != NULL)
1414 fprintf (lra_dump_file,
1415 "Changing pseudo %d in address of insn %u on equiv ",
1416 REGNO (reg), INSN_UID (curr_insn));
1417 dump_value_slim (lra_dump_file, *loc, 1);
1418 fprintf (lra_dump_file, "\n");
1420 *loc = copy_rtx (*loc);
1422 if (*loc != reg || ! in_class_p (reg, cl, &new_class))
1424 if (check_only_p)
1425 return true;
1426 reg = *loc;
1427 if (get_reload_reg (after == NULL ? OP_IN : OP_INOUT,
1428 mode, reg, cl, subreg_p, "address", &new_reg))
1429 before_p = true;
1431 else if (new_class != NO_REGS && rclass != new_class)
1433 if (check_only_p)
1434 return true;
1435 lra_change_class (regno, new_class, " Change to", true);
1436 return false;
1438 else
1439 return false;
1441 if (before_p)
1443 push_to_sequence (*before);
1444 lra_emit_move (new_reg, reg);
1445 *before = get_insns ();
1446 end_sequence ();
1448 *loc = new_reg;
1449 if (after != NULL)
1451 start_sequence ();
1452 lra_emit_move (before_p ? copy_rtx (reg) : reg, new_reg);
1453 emit_insn (*after);
1454 *after = get_insns ();
1455 end_sequence ();
1457 return true;
1460 /* Insert move insn in simplify_operand_subreg. BEFORE returns
1461 the insn to be inserted before curr insn. AFTER returns the
1462 the insn to be inserted after curr insn. ORIGREG and NEWREG
1463 are the original reg and new reg for reload. */
1464 static void
1465 insert_move_for_subreg (rtx_insn **before, rtx_insn **after, rtx origreg,
1466 rtx newreg)
1468 if (before)
1470 push_to_sequence (*before);
1471 lra_emit_move (newreg, origreg);
1472 *before = get_insns ();
1473 end_sequence ();
1475 if (after)
1477 start_sequence ();
1478 lra_emit_move (origreg, newreg);
1479 emit_insn (*after);
1480 *after = get_insns ();
1481 end_sequence ();
1485 static int valid_address_p (machine_mode mode, rtx addr, addr_space_t as);
1486 static bool process_address (int, bool, rtx_insn **, rtx_insn **);
1488 /* Make reloads for subreg in operand NOP with internal subreg mode
1489 REG_MODE, add new reloads for further processing. Return true if
1490 any change was done. */
1491 static bool
1492 simplify_operand_subreg (int nop, machine_mode reg_mode)
1494 int hard_regno;
1495 rtx_insn *before, *after;
1496 machine_mode mode, innermode;
1497 rtx reg, new_reg;
1498 rtx operand = *curr_id->operand_loc[nop];
1499 enum reg_class regclass;
1500 enum op_type type;
1502 before = after = NULL;
1504 if (GET_CODE (operand) != SUBREG)
1505 return false;
1507 mode = GET_MODE (operand);
1508 reg = SUBREG_REG (operand);
1509 innermode = GET_MODE (reg);
1510 type = curr_static_id->operand[nop].type;
1511 if (MEM_P (reg))
1513 const bool addr_was_valid
1514 = valid_address_p (innermode, XEXP (reg, 0), MEM_ADDR_SPACE (reg));
1515 alter_subreg (curr_id->operand_loc[nop], false);
1516 rtx subst = *curr_id->operand_loc[nop];
1517 lra_assert (MEM_P (subst));
1519 if (!addr_was_valid
1520 || valid_address_p (GET_MODE (subst), XEXP (subst, 0),
1521 MEM_ADDR_SPACE (subst))
1522 || ((get_constraint_type (lookup_constraint
1523 (curr_static_id->operand[nop].constraint))
1524 != CT_SPECIAL_MEMORY)
1525 /* We still can reload address and if the address is
1526 valid, we can remove subreg without reloading its
1527 inner memory. */
1528 && valid_address_p (GET_MODE (subst),
1529 regno_reg_rtx
1530 [ira_class_hard_regs
1531 [base_reg_class (GET_MODE (subst),
1532 MEM_ADDR_SPACE (subst),
1533 ADDRESS, SCRATCH)][0]],
1534 MEM_ADDR_SPACE (subst))))
1536 /* If we change the address for a paradoxical subreg of memory, the
1537 new address might violate the necessary alignment or the access
1538 might be slow; take this into consideration. We need not worry
1539 about accesses beyond allocated memory for paradoxical memory
1540 subregs as we don't substitute such equiv memory (see processing
1541 equivalences in function lra_constraints) and because for spilled
1542 pseudos we allocate stack memory enough for the biggest
1543 corresponding paradoxical subreg.
1545 However, do not blindly simplify a (subreg (mem ...)) for
1546 WORD_REGISTER_OPERATIONS targets as this may lead to loading junk
1547 data into a register when the inner is narrower than outer or
1548 missing important data from memory when the inner is wider than
1549 outer. This rule only applies to modes that are no wider than
1550 a word. */
1551 if (!(GET_MODE_PRECISION (mode) != GET_MODE_PRECISION (innermode)
1552 && GET_MODE_SIZE (mode) <= UNITS_PER_WORD
1553 && GET_MODE_SIZE (innermode) <= UNITS_PER_WORD
1554 && WORD_REGISTER_OPERATIONS)
1555 && (!(MEM_ALIGN (subst) < GET_MODE_ALIGNMENT (mode)
1556 && targetm.slow_unaligned_access (mode, MEM_ALIGN (subst)))
1557 || (MEM_ALIGN (reg) < GET_MODE_ALIGNMENT (innermode)
1558 && targetm.slow_unaligned_access (innermode,
1559 MEM_ALIGN (reg)))))
1560 return true;
1562 *curr_id->operand_loc[nop] = operand;
1564 /* But if the address was not valid, we cannot reload the MEM without
1565 reloading the address first. */
1566 if (!addr_was_valid)
1567 process_address (nop, false, &before, &after);
1569 /* INNERMODE is fast, MODE slow. Reload the mem in INNERMODE. */
1570 enum reg_class rclass
1571 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1572 if (get_reload_reg (curr_static_id->operand[nop].type, innermode,
1573 reg, rclass, TRUE, "slow mem", &new_reg))
1575 bool insert_before, insert_after;
1576 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1578 insert_before = (type != OP_OUT
1579 || partial_subreg_p (mode, innermode));
1580 insert_after = type != OP_IN;
1581 insert_move_for_subreg (insert_before ? &before : NULL,
1582 insert_after ? &after : NULL,
1583 reg, new_reg);
1585 SUBREG_REG (operand) = new_reg;
1587 /* Convert to MODE. */
1588 reg = operand;
1589 rclass
1590 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1591 if (get_reload_reg (curr_static_id->operand[nop].type, mode, reg,
1592 rclass, TRUE, "slow mem", &new_reg))
1594 bool insert_before, insert_after;
1595 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1597 insert_before = type != OP_OUT;
1598 insert_after = type != OP_IN;
1599 insert_move_for_subreg (insert_before ? &before : NULL,
1600 insert_after ? &after : NULL,
1601 reg, new_reg);
1603 *curr_id->operand_loc[nop] = new_reg;
1604 lra_process_new_insns (curr_insn, before, after,
1605 "Inserting slow mem reload");
1606 return true;
1609 /* If the address was valid and became invalid, prefer to reload
1610 the memory. Typical case is when the index scale should
1611 correspond the memory. */
1612 *curr_id->operand_loc[nop] = operand;
1613 /* Do not return false here as the MEM_P (reg) will be processed
1614 later in this function. */
1616 else if (REG_P (reg) && REGNO (reg) < FIRST_PSEUDO_REGISTER)
1618 alter_subreg (curr_id->operand_loc[nop], false);
1619 return true;
1621 else if (CONSTANT_P (reg))
1623 /* Try to simplify subreg of constant. It is usually result of
1624 equivalence substitution. */
1625 if (innermode == VOIDmode
1626 && (innermode = original_subreg_reg_mode[nop]) == VOIDmode)
1627 innermode = curr_static_id->operand[nop].mode;
1628 if ((new_reg = simplify_subreg (mode, reg, innermode,
1629 SUBREG_BYTE (operand))) != NULL_RTX)
1631 *curr_id->operand_loc[nop] = new_reg;
1632 return true;
1635 /* Put constant into memory when we have mixed modes. It generates
1636 a better code in most cases as it does not need a secondary
1637 reload memory. It also prevents LRA looping when LRA is using
1638 secondary reload memory again and again. */
1639 if (CONSTANT_P (reg) && CONST_POOL_OK_P (reg_mode, reg)
1640 && SCALAR_INT_MODE_P (reg_mode) != SCALAR_INT_MODE_P (mode))
1642 SUBREG_REG (operand) = force_const_mem (reg_mode, reg);
1643 alter_subreg (curr_id->operand_loc[nop], false);
1644 return true;
1646 /* Force a reload of the SUBREG_REG if this is a constant or PLUS or
1647 if there may be a problem accessing OPERAND in the outer
1648 mode. */
1649 if ((REG_P (reg)
1650 && REGNO (reg) >= FIRST_PSEUDO_REGISTER
1651 && (hard_regno = lra_get_regno_hard_regno (REGNO (reg))) >= 0
1652 /* Don't reload paradoxical subregs because we could be looping
1653 having repeatedly final regno out of hard regs range. */
1654 && (hard_regno_nregs (hard_regno, innermode)
1655 >= hard_regno_nregs (hard_regno, mode))
1656 && simplify_subreg_regno (hard_regno, innermode,
1657 SUBREG_BYTE (operand), mode) < 0
1658 /* Don't reload subreg for matching reload. It is actually
1659 valid subreg in LRA. */
1660 && ! LRA_SUBREG_P (operand))
1661 || CONSTANT_P (reg) || GET_CODE (reg) == PLUS || MEM_P (reg))
1663 enum reg_class rclass;
1665 if (REG_P (reg))
1666 /* There is a big probability that we will get the same class
1667 for the new pseudo and we will get the same insn which
1668 means infinite looping. So spill the new pseudo. */
1669 rclass = NO_REGS;
1670 else
1671 /* The class will be defined later in curr_insn_transform. */
1672 rclass
1673 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1675 if (get_reload_reg (curr_static_id->operand[nop].type, reg_mode, reg,
1676 rclass, TRUE, "subreg reg", &new_reg))
1678 bool insert_before, insert_after;
1679 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1681 insert_before = (type != OP_OUT
1682 || read_modify_subreg_p (operand));
1683 insert_after = (type != OP_IN);
1684 insert_move_for_subreg (insert_before ? &before : NULL,
1685 insert_after ? &after : NULL,
1686 reg, new_reg);
1688 SUBREG_REG (operand) = new_reg;
1689 lra_process_new_insns (curr_insn, before, after,
1690 "Inserting subreg reload");
1691 return true;
1693 /* Force a reload for a paradoxical subreg. For paradoxical subreg,
1694 IRA allocates hardreg to the inner pseudo reg according to its mode
1695 instead of the outermode, so the size of the hardreg may not be enough
1696 to contain the outermode operand, in that case we may need to insert
1697 reload for the reg. For the following two types of paradoxical subreg,
1698 we need to insert reload:
1699 1. If the op_type is OP_IN, and the hardreg could not be paired with
1700 other hardreg to contain the outermode operand
1701 (checked by in_hard_reg_set_p), we need to insert the reload.
1702 2. If the op_type is OP_OUT or OP_INOUT.
1704 Here is a paradoxical subreg example showing how the reload is generated:
1706 (insn 5 4 7 2 (set (reg:TI 106 [ __comp ])
1707 (subreg:TI (reg:DI 107 [ __comp ]) 0)) {*movti_internal_rex64}
1709 In IRA, reg107 is allocated to a DImode hardreg. We use x86-64 as example
1710 here, if reg107 is assigned to hardreg R15, because R15 is the last
1711 hardreg, compiler cannot find another hardreg to pair with R15 to
1712 contain TImode data. So we insert a TImode reload reg180 for it.
1713 After reload is inserted:
1715 (insn 283 0 0 (set (subreg:DI (reg:TI 180 [orig:107 __comp ] [107]) 0)
1716 (reg:DI 107 [ __comp ])) -1
1717 (insn 5 4 7 2 (set (reg:TI 106 [ __comp ])
1718 (subreg:TI (reg:TI 180 [orig:107 __comp ] [107]) 0)) {*movti_internal_rex64}
1720 Two reload hard registers will be allocated to reg180 to save TImode data
1721 in LRA_assign. */
1722 else if (REG_P (reg)
1723 && REGNO (reg) >= FIRST_PSEUDO_REGISTER
1724 && (hard_regno = lra_get_regno_hard_regno (REGNO (reg))) >= 0
1725 && (hard_regno_nregs (hard_regno, innermode)
1726 < hard_regno_nregs (hard_regno, mode))
1727 && (regclass = lra_get_allocno_class (REGNO (reg)))
1728 && (type != OP_IN
1729 || !in_hard_reg_set_p (reg_class_contents[regclass],
1730 mode, hard_regno)))
1732 /* The class will be defined later in curr_insn_transform. */
1733 enum reg_class rclass
1734 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1736 if (get_reload_reg (curr_static_id->operand[nop].type, mode, reg,
1737 rclass, TRUE, "paradoxical subreg", &new_reg))
1739 rtx subreg;
1740 bool insert_before, insert_after;
1742 PUT_MODE (new_reg, mode);
1743 subreg = gen_lowpart_SUBREG (innermode, new_reg);
1744 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1746 insert_before = (type != OP_OUT);
1747 insert_after = (type != OP_IN);
1748 insert_move_for_subreg (insert_before ? &before : NULL,
1749 insert_after ? &after : NULL,
1750 reg, subreg);
1752 SUBREG_REG (operand) = new_reg;
1753 lra_process_new_insns (curr_insn, before, after,
1754 "Inserting paradoxical subreg reload");
1755 return true;
1757 return false;
1760 /* Return TRUE if X refers for a hard register from SET. */
1761 static bool
1762 uses_hard_regs_p (rtx x, HARD_REG_SET set)
1764 int i, j, x_hard_regno;
1765 machine_mode mode;
1766 const char *fmt;
1767 enum rtx_code code;
1769 if (x == NULL_RTX)
1770 return false;
1771 code = GET_CODE (x);
1772 mode = GET_MODE (x);
1773 if (code == SUBREG)
1775 mode = wider_subreg_mode (x);
1776 x = SUBREG_REG (x);
1777 code = GET_CODE (x);
1780 if (REG_P (x))
1782 x_hard_regno = get_hard_regno (x, true);
1783 return (x_hard_regno >= 0
1784 && overlaps_hard_reg_set_p (set, mode, x_hard_regno));
1786 if (MEM_P (x))
1788 struct address_info ad;
1790 decompose_mem_address (&ad, x);
1791 if (ad.base_term != NULL && uses_hard_regs_p (*ad.base_term, set))
1792 return true;
1793 if (ad.index_term != NULL && uses_hard_regs_p (*ad.index_term, set))
1794 return true;
1796 fmt = GET_RTX_FORMAT (code);
1797 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1799 if (fmt[i] == 'e')
1801 if (uses_hard_regs_p (XEXP (x, i), set))
1802 return true;
1804 else if (fmt[i] == 'E')
1806 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
1807 if (uses_hard_regs_p (XVECEXP (x, i, j), set))
1808 return true;
1811 return false;
1814 /* Return true if OP is a spilled pseudo. */
1815 static inline bool
1816 spilled_pseudo_p (rtx op)
1818 return (REG_P (op)
1819 && REGNO (op) >= FIRST_PSEUDO_REGISTER && in_mem_p (REGNO (op)));
1822 /* Return true if X is a general constant. */
1823 static inline bool
1824 general_constant_p (rtx x)
1826 return CONSTANT_P (x) && (! flag_pic || LEGITIMATE_PIC_OPERAND_P (x));
1829 static bool
1830 reg_in_class_p (rtx reg, enum reg_class cl)
1832 if (cl == NO_REGS)
1833 return get_reg_class (REGNO (reg)) == NO_REGS;
1834 return in_class_p (reg, cl, NULL);
1837 /* Return true if SET of RCLASS contains no hard regs which can be
1838 used in MODE. */
1839 static bool
1840 prohibited_class_reg_set_mode_p (enum reg_class rclass,
1841 HARD_REG_SET &set,
1842 machine_mode mode)
1844 HARD_REG_SET temp;
1846 lra_assert (hard_reg_set_subset_p (reg_class_contents[rclass], set));
1847 COPY_HARD_REG_SET (temp, set);
1848 AND_COMPL_HARD_REG_SET (temp, lra_no_alloc_regs);
1849 return (hard_reg_set_subset_p
1850 (temp, ira_prohibited_class_mode_regs[rclass][mode]));
1854 /* Used to check validity info about small class input operands. It
1855 should be incremented at start of processing an insn
1856 alternative. */
1857 static unsigned int curr_small_class_check = 0;
1859 /* Update number of used inputs of class OP_CLASS for operand NOP.
1860 Return true if we have more such class operands than the number of
1861 available regs. */
1862 static bool
1863 update_and_check_small_class_inputs (int nop, enum reg_class op_class)
1865 static unsigned int small_class_check[LIM_REG_CLASSES];
1866 static int small_class_input_nums[LIM_REG_CLASSES];
1868 if (SMALL_REGISTER_CLASS_P (op_class)
1869 /* We are interesting in classes became small because of fixing
1870 some hard regs, e.g. by an user through GCC options. */
1871 && hard_reg_set_intersect_p (reg_class_contents[op_class],
1872 ira_no_alloc_regs)
1873 && (curr_static_id->operand[nop].type != OP_OUT
1874 || curr_static_id->operand[nop].early_clobber))
1876 if (small_class_check[op_class] == curr_small_class_check)
1877 small_class_input_nums[op_class]++;
1878 else
1880 small_class_check[op_class] = curr_small_class_check;
1881 small_class_input_nums[op_class] = 1;
1883 if (small_class_input_nums[op_class] > ira_class_hard_regs_num[op_class])
1884 return true;
1886 return false;
1889 /* Major function to choose the current insn alternative and what
1890 operands should be reloaded and how. If ONLY_ALTERNATIVE is not
1891 negative we should consider only this alternative. Return false if
1892 we can not choose the alternative or find how to reload the
1893 operands. */
1894 static bool
1895 process_alt_operands (int only_alternative)
1897 bool ok_p = false;
1898 int nop, overall, nalt;
1899 int n_alternatives = curr_static_id->n_alternatives;
1900 int n_operands = curr_static_id->n_operands;
1901 /* LOSERS counts the operands that don't fit this alternative and
1902 would require loading. */
1903 int losers;
1904 int addr_losers;
1905 /* REJECT is a count of how undesirable this alternative says it is
1906 if any reloading is required. If the alternative matches exactly
1907 then REJECT is ignored, but otherwise it gets this much counted
1908 against it in addition to the reloading needed. */
1909 int reject;
1910 /* This is defined by '!' or '?' alternative constraint and added to
1911 reject. But in some cases it can be ignored. */
1912 int static_reject;
1913 int op_reject;
1914 /* The number of elements in the following array. */
1915 int early_clobbered_regs_num;
1916 /* Numbers of operands which are early clobber registers. */
1917 int early_clobbered_nops[MAX_RECOG_OPERANDS];
1918 enum reg_class curr_alt[MAX_RECOG_OPERANDS];
1919 HARD_REG_SET curr_alt_set[MAX_RECOG_OPERANDS];
1920 bool curr_alt_match_win[MAX_RECOG_OPERANDS];
1921 bool curr_alt_win[MAX_RECOG_OPERANDS];
1922 bool curr_alt_offmemok[MAX_RECOG_OPERANDS];
1923 int curr_alt_matches[MAX_RECOG_OPERANDS];
1924 /* The number of elements in the following array. */
1925 int curr_alt_dont_inherit_ops_num;
1926 /* Numbers of operands whose reload pseudos should not be inherited. */
1927 int curr_alt_dont_inherit_ops[MAX_RECOG_OPERANDS];
1928 rtx op;
1929 /* The register when the operand is a subreg of register, otherwise the
1930 operand itself. */
1931 rtx no_subreg_reg_operand[MAX_RECOG_OPERANDS];
1932 /* The register if the operand is a register or subreg of register,
1933 otherwise NULL. */
1934 rtx operand_reg[MAX_RECOG_OPERANDS];
1935 int hard_regno[MAX_RECOG_OPERANDS];
1936 machine_mode biggest_mode[MAX_RECOG_OPERANDS];
1937 int reload_nregs, reload_sum;
1938 bool costly_p;
1939 enum reg_class cl;
1941 /* Calculate some data common for all alternatives to speed up the
1942 function. */
1943 for (nop = 0; nop < n_operands; nop++)
1945 rtx reg;
1947 op = no_subreg_reg_operand[nop] = *curr_id->operand_loc[nop];
1948 /* The real hard regno of the operand after the allocation. */
1949 hard_regno[nop] = get_hard_regno (op, true);
1951 operand_reg[nop] = reg = op;
1952 biggest_mode[nop] = GET_MODE (op);
1953 if (GET_CODE (op) == SUBREG)
1955 biggest_mode[nop] = wider_subreg_mode (op);
1956 operand_reg[nop] = reg = SUBREG_REG (op);
1958 if (! REG_P (reg))
1959 operand_reg[nop] = NULL_RTX;
1960 else if (REGNO (reg) >= FIRST_PSEUDO_REGISTER
1961 || ((int) REGNO (reg)
1962 == lra_get_elimination_hard_regno (REGNO (reg))))
1963 no_subreg_reg_operand[nop] = reg;
1964 else
1965 operand_reg[nop] = no_subreg_reg_operand[nop]
1966 /* Just use natural mode for elimination result. It should
1967 be enough for extra constraints hooks. */
1968 = regno_reg_rtx[hard_regno[nop]];
1971 /* The constraints are made of several alternatives. Each operand's
1972 constraint looks like foo,bar,... with commas separating the
1973 alternatives. The first alternatives for all operands go
1974 together, the second alternatives go together, etc.
1976 First loop over alternatives. */
1977 alternative_mask preferred = curr_id->preferred_alternatives;
1978 if (only_alternative >= 0)
1979 preferred &= ALTERNATIVE_BIT (only_alternative);
1981 for (nalt = 0; nalt < n_alternatives; nalt++)
1983 /* Loop over operands for one constraint alternative. */
1984 if (!TEST_BIT (preferred, nalt))
1985 continue;
1987 curr_small_class_check++;
1988 overall = losers = addr_losers = 0;
1989 static_reject = reject = reload_nregs = reload_sum = 0;
1990 for (nop = 0; nop < n_operands; nop++)
1992 int inc = (curr_static_id
1993 ->operand_alternative[nalt * n_operands + nop].reject);
1994 if (lra_dump_file != NULL && inc != 0)
1995 fprintf (lra_dump_file,
1996 " Staticly defined alt reject+=%d\n", inc);
1997 static_reject += inc;
1999 reject += static_reject;
2000 early_clobbered_regs_num = 0;
2002 for (nop = 0; nop < n_operands; nop++)
2004 const char *p;
2005 char *end;
2006 int len, c, m, i, opalt_num, this_alternative_matches;
2007 bool win, did_match, offmemok, early_clobber_p;
2008 /* false => this operand can be reloaded somehow for this
2009 alternative. */
2010 bool badop;
2011 /* true => this operand can be reloaded if the alternative
2012 allows regs. */
2013 bool winreg;
2014 /* True if a constant forced into memory would be OK for
2015 this operand. */
2016 bool constmemok;
2017 enum reg_class this_alternative, this_costly_alternative;
2018 HARD_REG_SET this_alternative_set, this_costly_alternative_set;
2019 bool this_alternative_match_win, this_alternative_win;
2020 bool this_alternative_offmemok;
2021 bool scratch_p;
2022 machine_mode mode;
2023 enum constraint_num cn;
2025 opalt_num = nalt * n_operands + nop;
2026 if (curr_static_id->operand_alternative[opalt_num].anything_ok)
2028 /* Fast track for no constraints at all. */
2029 curr_alt[nop] = NO_REGS;
2030 CLEAR_HARD_REG_SET (curr_alt_set[nop]);
2031 curr_alt_win[nop] = true;
2032 curr_alt_match_win[nop] = false;
2033 curr_alt_offmemok[nop] = false;
2034 curr_alt_matches[nop] = -1;
2035 continue;
2038 op = no_subreg_reg_operand[nop];
2039 mode = curr_operand_mode[nop];
2041 win = did_match = winreg = offmemok = constmemok = false;
2042 badop = true;
2044 early_clobber_p = false;
2045 p = curr_static_id->operand_alternative[opalt_num].constraint;
2047 this_costly_alternative = this_alternative = NO_REGS;
2048 /* We update set of possible hard regs besides its class
2049 because reg class might be inaccurate. For example,
2050 union of LO_REGS (l), HI_REGS(h), and STACK_REG(k) in ARM
2051 is translated in HI_REGS because classes are merged by
2052 pairs and there is no accurate intermediate class. */
2053 CLEAR_HARD_REG_SET (this_alternative_set);
2054 CLEAR_HARD_REG_SET (this_costly_alternative_set);
2055 this_alternative_win = false;
2056 this_alternative_match_win = false;
2057 this_alternative_offmemok = false;
2058 this_alternative_matches = -1;
2060 /* An empty constraint should be excluded by the fast
2061 track. */
2062 lra_assert (*p != 0 && *p != ',');
2064 op_reject = 0;
2065 /* Scan this alternative's specs for this operand; set WIN
2066 if the operand fits any letter in this alternative.
2067 Otherwise, clear BADOP if this operand could fit some
2068 letter after reloads, or set WINREG if this operand could
2069 fit after reloads provided the constraint allows some
2070 registers. */
2071 costly_p = false;
2074 switch ((c = *p, len = CONSTRAINT_LEN (c, p)), c)
2076 case '\0':
2077 len = 0;
2078 break;
2079 case ',':
2080 c = '\0';
2081 break;
2083 case '&':
2084 early_clobber_p = true;
2085 break;
2087 case '$':
2088 op_reject += LRA_MAX_REJECT;
2089 break;
2090 case '^':
2091 op_reject += LRA_LOSER_COST_FACTOR;
2092 break;
2094 case '#':
2095 /* Ignore rest of this alternative. */
2096 c = '\0';
2097 break;
2099 case '0': case '1': case '2': case '3': case '4':
2100 case '5': case '6': case '7': case '8': case '9':
2102 int m_hregno;
2103 bool match_p;
2105 m = strtoul (p, &end, 10);
2106 p = end;
2107 len = 0;
2108 lra_assert (nop > m);
2110 this_alternative_matches = m;
2111 m_hregno = get_hard_regno (*curr_id->operand_loc[m], false);
2112 /* We are supposed to match a previous operand.
2113 If we do, we win if that one did. If we do
2114 not, count both of the operands as losers.
2115 (This is too conservative, since most of the
2116 time only a single reload insn will be needed
2117 to make the two operands win. As a result,
2118 this alternative may be rejected when it is
2119 actually desirable.) */
2120 match_p = false;
2121 if (operands_match_p (*curr_id->operand_loc[nop],
2122 *curr_id->operand_loc[m], m_hregno))
2124 /* We should reject matching of an early
2125 clobber operand if the matching operand is
2126 not dying in the insn. */
2127 if (! curr_static_id->operand[m].early_clobber
2128 || operand_reg[nop] == NULL_RTX
2129 || (find_regno_note (curr_insn, REG_DEAD,
2130 REGNO (op))
2131 || REGNO (op) == REGNO (operand_reg[m])))
2132 match_p = true;
2134 if (match_p)
2136 /* If we are matching a non-offsettable
2137 address where an offsettable address was
2138 expected, then we must reject this
2139 combination, because we can't reload
2140 it. */
2141 if (curr_alt_offmemok[m]
2142 && MEM_P (*curr_id->operand_loc[m])
2143 && curr_alt[m] == NO_REGS && ! curr_alt_win[m])
2144 continue;
2146 else
2148 /* Operands don't match. Both operands must
2149 allow a reload register, otherwise we
2150 cannot make them match. */
2151 if (curr_alt[m] == NO_REGS)
2152 break;
2153 /* Retroactively mark the operand we had to
2154 match as a loser, if it wasn't already and
2155 it wasn't matched to a register constraint
2156 (e.g it might be matched by memory). */
2157 if (curr_alt_win[m]
2158 && (operand_reg[m] == NULL_RTX
2159 || hard_regno[m] < 0))
2161 losers++;
2162 reload_nregs
2163 += (ira_reg_class_max_nregs[curr_alt[m]]
2164 [GET_MODE (*curr_id->operand_loc[m])]);
2167 /* Prefer matching earlyclobber alternative as
2168 it results in less hard regs required for
2169 the insn than a non-matching earlyclobber
2170 alternative. */
2171 if (curr_static_id->operand[m].early_clobber)
2173 if (lra_dump_file != NULL)
2174 fprintf
2175 (lra_dump_file,
2176 " %d Matching earlyclobber alt:"
2177 " reject--\n",
2178 nop);
2179 reject--;
2181 /* Otherwise we prefer no matching
2182 alternatives because it gives more freedom
2183 in RA. */
2184 else if (operand_reg[nop] == NULL_RTX
2185 || (find_regno_note (curr_insn, REG_DEAD,
2186 REGNO (operand_reg[nop]))
2187 == NULL_RTX))
2189 if (lra_dump_file != NULL)
2190 fprintf
2191 (lra_dump_file,
2192 " %d Matching alt: reject+=2\n",
2193 nop);
2194 reject += 2;
2197 /* If we have to reload this operand and some
2198 previous operand also had to match the same
2199 thing as this operand, we don't know how to do
2200 that. */
2201 if (!match_p || !curr_alt_win[m])
2203 for (i = 0; i < nop; i++)
2204 if (curr_alt_matches[i] == m)
2205 break;
2206 if (i < nop)
2207 break;
2209 else
2210 did_match = true;
2212 /* This can be fixed with reloads if the operand
2213 we are supposed to match can be fixed with
2214 reloads. */
2215 badop = false;
2216 this_alternative = curr_alt[m];
2217 COPY_HARD_REG_SET (this_alternative_set, curr_alt_set[m]);
2218 winreg = this_alternative != NO_REGS;
2219 break;
2222 case 'g':
2223 if (MEM_P (op)
2224 || general_constant_p (op)
2225 || spilled_pseudo_p (op))
2226 win = true;
2227 cl = GENERAL_REGS;
2228 goto reg;
2230 default:
2231 cn = lookup_constraint (p);
2232 switch (get_constraint_type (cn))
2234 case CT_REGISTER:
2235 cl = reg_class_for_constraint (cn);
2236 if (cl != NO_REGS)
2237 goto reg;
2238 break;
2240 case CT_CONST_INT:
2241 if (CONST_INT_P (op)
2242 && insn_const_int_ok_for_constraint (INTVAL (op), cn))
2243 win = true;
2244 break;
2246 case CT_MEMORY:
2247 if (MEM_P (op)
2248 && satisfies_memory_constraint_p (op, cn))
2249 win = true;
2250 else if (spilled_pseudo_p (op))
2251 win = true;
2253 /* If we didn't already win, we can reload constants
2254 via force_const_mem or put the pseudo value into
2255 memory, or make other memory by reloading the
2256 address like for 'o'. */
2257 if (CONST_POOL_OK_P (mode, op)
2258 || MEM_P (op) || REG_P (op)
2259 /* We can restore the equiv insn by a
2260 reload. */
2261 || equiv_substition_p[nop])
2262 badop = false;
2263 constmemok = true;
2264 offmemok = true;
2265 break;
2267 case CT_ADDRESS:
2268 /* If we didn't already win, we can reload the address
2269 into a base register. */
2270 if (satisfies_address_constraint_p (op, cn))
2271 win = true;
2272 cl = base_reg_class (VOIDmode, ADDR_SPACE_GENERIC,
2273 ADDRESS, SCRATCH);
2274 badop = false;
2275 goto reg;
2277 case CT_FIXED_FORM:
2278 if (constraint_satisfied_p (op, cn))
2279 win = true;
2280 break;
2282 case CT_SPECIAL_MEMORY:
2283 if (MEM_P (op)
2284 && satisfies_memory_constraint_p (op, cn))
2285 win = true;
2286 else if (spilled_pseudo_p (op))
2287 win = true;
2288 break;
2290 break;
2292 reg:
2293 this_alternative = reg_class_subunion[this_alternative][cl];
2294 IOR_HARD_REG_SET (this_alternative_set,
2295 reg_class_contents[cl]);
2296 if (costly_p)
2298 this_costly_alternative
2299 = reg_class_subunion[this_costly_alternative][cl];
2300 IOR_HARD_REG_SET (this_costly_alternative_set,
2301 reg_class_contents[cl]);
2303 if (mode == BLKmode)
2304 break;
2305 winreg = true;
2306 if (REG_P (op))
2308 if (hard_regno[nop] >= 0
2309 && in_hard_reg_set_p (this_alternative_set,
2310 mode, hard_regno[nop]))
2311 win = true;
2312 else if (hard_regno[nop] < 0
2313 && in_class_p (op, this_alternative, NULL))
2314 win = true;
2316 break;
2318 if (c != ' ' && c != '\t')
2319 costly_p = c == '*';
2321 while ((p += len), c);
2323 scratch_p = (operand_reg[nop] != NULL_RTX
2324 && lra_former_scratch_p (REGNO (operand_reg[nop])));
2325 /* Record which operands fit this alternative. */
2326 if (win)
2328 this_alternative_win = true;
2329 if (operand_reg[nop] != NULL_RTX)
2331 if (hard_regno[nop] >= 0)
2333 if (in_hard_reg_set_p (this_costly_alternative_set,
2334 mode, hard_regno[nop]))
2336 if (lra_dump_file != NULL)
2337 fprintf (lra_dump_file,
2338 " %d Costly set: reject++\n",
2339 nop);
2340 reject++;
2343 else
2345 /* Prefer won reg to spilled pseudo under other
2346 equal conditions for possibe inheritance. */
2347 if (! scratch_p)
2349 if (lra_dump_file != NULL)
2350 fprintf
2351 (lra_dump_file,
2352 " %d Non pseudo reload: reject++\n",
2353 nop);
2354 reject++;
2356 if (in_class_p (operand_reg[nop],
2357 this_costly_alternative, NULL))
2359 if (lra_dump_file != NULL)
2360 fprintf
2361 (lra_dump_file,
2362 " %d Non pseudo costly reload:"
2363 " reject++\n",
2364 nop);
2365 reject++;
2368 /* We simulate the behavior of old reload here.
2369 Although scratches need hard registers and it
2370 might result in spilling other pseudos, no reload
2371 insns are generated for the scratches. So it
2372 might cost something but probably less than old
2373 reload pass believes. */
2374 if (scratch_p)
2376 if (lra_dump_file != NULL)
2377 fprintf (lra_dump_file,
2378 " %d Scratch win: reject+=2\n",
2379 nop);
2380 reject += 2;
2384 else if (did_match)
2385 this_alternative_match_win = true;
2386 else
2388 int const_to_mem = 0;
2389 bool no_regs_p;
2391 reject += op_reject;
2392 /* Never do output reload of stack pointer. It makes
2393 impossible to do elimination when SP is changed in
2394 RTL. */
2395 if (op == stack_pointer_rtx && ! frame_pointer_needed
2396 && curr_static_id->operand[nop].type != OP_IN)
2397 goto fail;
2399 /* If this alternative asks for a specific reg class, see if there
2400 is at least one allocatable register in that class. */
2401 no_regs_p
2402 = (this_alternative == NO_REGS
2403 || (hard_reg_set_subset_p
2404 (reg_class_contents[this_alternative],
2405 lra_no_alloc_regs)));
2407 /* For asms, verify that the class for this alternative is possible
2408 for the mode that is specified. */
2409 if (!no_regs_p && INSN_CODE (curr_insn) < 0)
2411 int i;
2412 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
2413 if (targetm.hard_regno_mode_ok (i, mode)
2414 && in_hard_reg_set_p (reg_class_contents[this_alternative],
2415 mode, i))
2416 break;
2417 if (i == FIRST_PSEUDO_REGISTER)
2418 winreg = false;
2421 /* If this operand accepts a register, and if the
2422 register class has at least one allocatable register,
2423 then this operand can be reloaded. */
2424 if (winreg && !no_regs_p)
2425 badop = false;
2427 if (badop)
2429 if (lra_dump_file != NULL)
2430 fprintf (lra_dump_file,
2431 " alt=%d: Bad operand -- refuse\n",
2432 nalt);
2433 goto fail;
2436 if (this_alternative != NO_REGS)
2438 HARD_REG_SET available_regs;
2440 COPY_HARD_REG_SET (available_regs,
2441 reg_class_contents[this_alternative]);
2442 AND_COMPL_HARD_REG_SET
2443 (available_regs,
2444 ira_prohibited_class_mode_regs[this_alternative][mode]);
2445 AND_COMPL_HARD_REG_SET (available_regs, lra_no_alloc_regs);
2446 if (hard_reg_set_empty_p (available_regs))
2448 /* There are no hard regs holding a value of given
2449 mode. */
2450 if (offmemok)
2452 this_alternative = NO_REGS;
2453 if (lra_dump_file != NULL)
2454 fprintf (lra_dump_file,
2455 " %d Using memory because of"
2456 " a bad mode: reject+=2\n",
2457 nop);
2458 reject += 2;
2460 else
2462 if (lra_dump_file != NULL)
2463 fprintf (lra_dump_file,
2464 " alt=%d: Wrong mode -- refuse\n",
2465 nalt);
2466 goto fail;
2471 /* If not assigned pseudo has a class which a subset of
2472 required reg class, it is a less costly alternative
2473 as the pseudo still can get a hard reg of necessary
2474 class. */
2475 if (! no_regs_p && REG_P (op) && hard_regno[nop] < 0
2476 && (cl = get_reg_class (REGNO (op))) != NO_REGS
2477 && ira_class_subset_p[this_alternative][cl])
2479 if (lra_dump_file != NULL)
2480 fprintf
2481 (lra_dump_file,
2482 " %d Super set class reg: reject-=3\n", nop);
2483 reject -= 3;
2486 this_alternative_offmemok = offmemok;
2487 if (this_costly_alternative != NO_REGS)
2489 if (lra_dump_file != NULL)
2490 fprintf (lra_dump_file,
2491 " %d Costly loser: reject++\n", nop);
2492 reject++;
2494 /* If the operand is dying, has a matching constraint,
2495 and satisfies constraints of the matched operand
2496 which failed to satisfy the own constraints, most probably
2497 the reload for this operand will be gone. */
2498 if (this_alternative_matches >= 0
2499 && !curr_alt_win[this_alternative_matches]
2500 && REG_P (op)
2501 && find_regno_note (curr_insn, REG_DEAD, REGNO (op))
2502 && (hard_regno[nop] >= 0
2503 ? in_hard_reg_set_p (this_alternative_set,
2504 mode, hard_regno[nop])
2505 : in_class_p (op, this_alternative, NULL)))
2507 if (lra_dump_file != NULL)
2508 fprintf
2509 (lra_dump_file,
2510 " %d Dying matched operand reload: reject++\n",
2511 nop);
2512 reject++;
2514 else
2516 /* Strict_low_part requires to reload the register
2517 not the sub-register. In this case we should
2518 check that a final reload hard reg can hold the
2519 value mode. */
2520 if (curr_static_id->operand[nop].strict_low
2521 && REG_P (op)
2522 && hard_regno[nop] < 0
2523 && GET_CODE (*curr_id->operand_loc[nop]) == SUBREG
2524 && ira_class_hard_regs_num[this_alternative] > 0
2525 && (!targetm.hard_regno_mode_ok
2526 (ira_class_hard_regs[this_alternative][0],
2527 GET_MODE (*curr_id->operand_loc[nop]))))
2529 if (lra_dump_file != NULL)
2530 fprintf
2531 (lra_dump_file,
2532 " alt=%d: Strict low subreg reload -- refuse\n",
2533 nalt);
2534 goto fail;
2536 losers++;
2538 if (operand_reg[nop] != NULL_RTX
2539 /* Output operands and matched input operands are
2540 not inherited. The following conditions do not
2541 exactly describe the previous statement but they
2542 are pretty close. */
2543 && curr_static_id->operand[nop].type != OP_OUT
2544 && (this_alternative_matches < 0
2545 || curr_static_id->operand[nop].type != OP_IN))
2547 int last_reload = (lra_reg_info[ORIGINAL_REGNO
2548 (operand_reg[nop])]
2549 .last_reload);
2551 /* The value of reload_sum has sense only if we
2552 process insns in their order. It happens only on
2553 the first constraints sub-pass when we do most of
2554 reload work. */
2555 if (lra_constraint_iter == 1 && last_reload > bb_reload_num)
2556 reload_sum += last_reload - bb_reload_num;
2558 /* If this is a constant that is reloaded into the
2559 desired class by copying it to memory first, count
2560 that as another reload. This is consistent with
2561 other code and is required to avoid choosing another
2562 alternative when the constant is moved into memory.
2563 Note that the test here is precisely the same as in
2564 the code below that calls force_const_mem. */
2565 if (CONST_POOL_OK_P (mode, op)
2566 && ((targetm.preferred_reload_class
2567 (op, this_alternative) == NO_REGS)
2568 || no_input_reloads_p))
2570 const_to_mem = 1;
2571 if (! no_regs_p)
2572 losers++;
2575 /* Alternative loses if it requires a type of reload not
2576 permitted for this insn. We can always reload
2577 objects with a REG_UNUSED note. */
2578 if ((curr_static_id->operand[nop].type != OP_IN
2579 && no_output_reloads_p
2580 && ! find_reg_note (curr_insn, REG_UNUSED, op))
2581 || (curr_static_id->operand[nop].type != OP_OUT
2582 && no_input_reloads_p && ! const_to_mem)
2583 || (this_alternative_matches >= 0
2584 && (no_input_reloads_p
2585 || (no_output_reloads_p
2586 && (curr_static_id->operand
2587 [this_alternative_matches].type != OP_IN)
2588 && ! find_reg_note (curr_insn, REG_UNUSED,
2589 no_subreg_reg_operand
2590 [this_alternative_matches])))))
2592 if (lra_dump_file != NULL)
2593 fprintf
2594 (lra_dump_file,
2595 " alt=%d: No input/otput reload -- refuse\n",
2596 nalt);
2597 goto fail;
2600 /* Alternative loses if it required class pseudo can not
2601 hold value of required mode. Such insns can be
2602 described by insn definitions with mode iterators. */
2603 if (GET_MODE (*curr_id->operand_loc[nop]) != VOIDmode
2604 && ! hard_reg_set_empty_p (this_alternative_set)
2605 /* It is common practice for constraints to use a
2606 class which does not have actually enough regs to
2607 hold the value (e.g. x86 AREG for mode requiring
2608 more one general reg). Therefore we have 2
2609 conditions to check that the reload pseudo can
2610 not hold the mode value. */
2611 && (!targetm.hard_regno_mode_ok
2612 (ira_class_hard_regs[this_alternative][0],
2613 GET_MODE (*curr_id->operand_loc[nop])))
2614 /* The above condition is not enough as the first
2615 reg in ira_class_hard_regs can be not aligned for
2616 multi-words mode values. */
2617 && (prohibited_class_reg_set_mode_p
2618 (this_alternative, this_alternative_set,
2619 GET_MODE (*curr_id->operand_loc[nop]))))
2621 if (lra_dump_file != NULL)
2622 fprintf (lra_dump_file,
2623 " alt=%d: reload pseudo for op %d "
2624 " can not hold the mode value -- refuse\n",
2625 nalt, nop);
2626 goto fail;
2629 /* Check strong discouragement of reload of non-constant
2630 into class THIS_ALTERNATIVE. */
2631 if (! CONSTANT_P (op) && ! no_regs_p
2632 && (targetm.preferred_reload_class
2633 (op, this_alternative) == NO_REGS
2634 || (curr_static_id->operand[nop].type == OP_OUT
2635 && (targetm.preferred_output_reload_class
2636 (op, this_alternative) == NO_REGS))))
2638 if (lra_dump_file != NULL)
2639 fprintf (lra_dump_file,
2640 " %d Non-prefered reload: reject+=%d\n",
2641 nop, LRA_MAX_REJECT);
2642 reject += LRA_MAX_REJECT;
2645 if (! (MEM_P (op) && offmemok)
2646 && ! (const_to_mem && constmemok))
2648 /* We prefer to reload pseudos over reloading other
2649 things, since such reloads may be able to be
2650 eliminated later. So bump REJECT in other cases.
2651 Don't do this in the case where we are forcing a
2652 constant into memory and it will then win since
2653 we don't want to have a different alternative
2654 match then. */
2655 if (! (REG_P (op) && REGNO (op) >= FIRST_PSEUDO_REGISTER))
2657 if (lra_dump_file != NULL)
2658 fprintf
2659 (lra_dump_file,
2660 " %d Non-pseudo reload: reject+=2\n",
2661 nop);
2662 reject += 2;
2665 if (! no_regs_p)
2666 reload_nregs
2667 += ira_reg_class_max_nregs[this_alternative][mode];
2669 if (SMALL_REGISTER_CLASS_P (this_alternative))
2671 if (lra_dump_file != NULL)
2672 fprintf
2673 (lra_dump_file,
2674 " %d Small class reload: reject+=%d\n",
2675 nop, LRA_LOSER_COST_FACTOR / 2);
2676 reject += LRA_LOSER_COST_FACTOR / 2;
2680 /* We are trying to spill pseudo into memory. It is
2681 usually more costly than moving to a hard register
2682 although it might takes the same number of
2683 reloads.
2685 Non-pseudo spill may happen also. Suppose a target allows both
2686 register and memory in the operand constraint alternatives,
2687 then it's typical that an eliminable register has a substition
2688 of "base + offset" which can either be reloaded by a simple
2689 "new_reg <= base + offset" which will match the register
2690 constraint, or a similar reg addition followed by further spill
2691 to and reload from memory which will match the memory
2692 constraint, but this memory spill will be much more costly
2693 usually.
2695 Code below increases the reject for both pseudo and non-pseudo
2696 spill. */
2697 if (no_regs_p
2698 && !(MEM_P (op) && offmemok)
2699 && !(REG_P (op) && hard_regno[nop] < 0))
2701 if (lra_dump_file != NULL)
2702 fprintf
2703 (lra_dump_file,
2704 " %d Spill %spseudo into memory: reject+=3\n",
2705 nop, REG_P (op) ? "" : "Non-");
2706 reject += 3;
2707 if (VECTOR_MODE_P (mode))
2709 /* Spilling vectors into memory is usually more
2710 costly as they contain big values. */
2711 if (lra_dump_file != NULL)
2712 fprintf
2713 (lra_dump_file,
2714 " %d Spill vector pseudo: reject+=2\n",
2715 nop);
2716 reject += 2;
2720 /* When we use an operand requiring memory in given
2721 alternative, the insn should write *and* read the
2722 value to/from memory it is costly in comparison with
2723 an insn alternative which does not use memory
2724 (e.g. register or immediate operand). We exclude
2725 memory operand for such case as we can satisfy the
2726 memory constraints by reloading address. */
2727 if (no_regs_p && offmemok && !MEM_P (op))
2729 if (lra_dump_file != NULL)
2730 fprintf
2731 (lra_dump_file,
2732 " Using memory insn operand %d: reject+=3\n",
2733 nop);
2734 reject += 3;
2737 /* If reload requires moving value through secondary
2738 memory, it will need one more insn at least. */
2739 if (this_alternative != NO_REGS
2740 && REG_P (op) && (cl = get_reg_class (REGNO (op))) != NO_REGS
2741 && ((curr_static_id->operand[nop].type != OP_OUT
2742 && targetm.secondary_memory_needed (GET_MODE (op), cl,
2743 this_alternative))
2744 || (curr_static_id->operand[nop].type != OP_IN
2745 && (targetm.secondary_memory_needed
2746 (GET_MODE (op), this_alternative, cl)))))
2747 losers++;
2749 /* Input reloads can be inherited more often than output
2750 reloads can be removed, so penalize output
2751 reloads. */
2752 if (!REG_P (op) || curr_static_id->operand[nop].type != OP_IN)
2754 if (lra_dump_file != NULL)
2755 fprintf
2756 (lra_dump_file,
2757 " %d Non input pseudo reload: reject++\n",
2758 nop);
2759 reject++;
2762 if (MEM_P (op) && offmemok)
2763 addr_losers++;
2764 else if (curr_static_id->operand[nop].type == OP_INOUT)
2766 if (lra_dump_file != NULL)
2767 fprintf
2768 (lra_dump_file,
2769 " %d Input/Output reload: reject+=%d\n",
2770 nop, LRA_LOSER_COST_FACTOR);
2771 reject += LRA_LOSER_COST_FACTOR;
2775 if (early_clobber_p && ! scratch_p)
2777 if (lra_dump_file != NULL)
2778 fprintf (lra_dump_file,
2779 " %d Early clobber: reject++\n", nop);
2780 reject++;
2782 /* ??? We check early clobbers after processing all operands
2783 (see loop below) and there we update the costs more.
2784 Should we update the cost (may be approximately) here
2785 because of early clobber register reloads or it is a rare
2786 or non-important thing to be worth to do it. */
2787 overall = (losers * LRA_LOSER_COST_FACTOR + reject
2788 - (addr_losers == losers ? static_reject : 0));
2789 if ((best_losers == 0 || losers != 0) && best_overall < overall)
2791 if (lra_dump_file != NULL)
2792 fprintf (lra_dump_file,
2793 " alt=%d,overall=%d,losers=%d -- refuse\n",
2794 nalt, overall, losers);
2795 goto fail;
2798 if (update_and_check_small_class_inputs (nop, this_alternative))
2800 if (lra_dump_file != NULL)
2801 fprintf (lra_dump_file,
2802 " alt=%d, not enough small class regs -- refuse\n",
2803 nalt);
2804 goto fail;
2806 curr_alt[nop] = this_alternative;
2807 COPY_HARD_REG_SET (curr_alt_set[nop], this_alternative_set);
2808 curr_alt_win[nop] = this_alternative_win;
2809 curr_alt_match_win[nop] = this_alternative_match_win;
2810 curr_alt_offmemok[nop] = this_alternative_offmemok;
2811 curr_alt_matches[nop] = this_alternative_matches;
2813 if (this_alternative_matches >= 0
2814 && !did_match && !this_alternative_win)
2815 curr_alt_win[this_alternative_matches] = false;
2817 if (early_clobber_p && operand_reg[nop] != NULL_RTX)
2818 early_clobbered_nops[early_clobbered_regs_num++] = nop;
2821 if (curr_insn_set != NULL_RTX && n_operands == 2
2822 /* Prevent processing non-move insns. */
2823 && (GET_CODE (SET_SRC (curr_insn_set)) == SUBREG
2824 || SET_SRC (curr_insn_set) == no_subreg_reg_operand[1])
2825 && ((! curr_alt_win[0] && ! curr_alt_win[1]
2826 && REG_P (no_subreg_reg_operand[0])
2827 && REG_P (no_subreg_reg_operand[1])
2828 && (reg_in_class_p (no_subreg_reg_operand[0], curr_alt[1])
2829 || reg_in_class_p (no_subreg_reg_operand[1], curr_alt[0])))
2830 || (! curr_alt_win[0] && curr_alt_win[1]
2831 && REG_P (no_subreg_reg_operand[1])
2832 /* Check that we reload memory not the memory
2833 address. */
2834 && ! (curr_alt_offmemok[0]
2835 && MEM_P (no_subreg_reg_operand[0]))
2836 && reg_in_class_p (no_subreg_reg_operand[1], curr_alt[0]))
2837 || (curr_alt_win[0] && ! curr_alt_win[1]
2838 && REG_P (no_subreg_reg_operand[0])
2839 /* Check that we reload memory not the memory
2840 address. */
2841 && ! (curr_alt_offmemok[1]
2842 && MEM_P (no_subreg_reg_operand[1]))
2843 && reg_in_class_p (no_subreg_reg_operand[0], curr_alt[1])
2844 && (! CONST_POOL_OK_P (curr_operand_mode[1],
2845 no_subreg_reg_operand[1])
2846 || (targetm.preferred_reload_class
2847 (no_subreg_reg_operand[1],
2848 (enum reg_class) curr_alt[1]) != NO_REGS))
2849 /* If it is a result of recent elimination in move
2850 insn we can transform it into an add still by
2851 using this alternative. */
2852 && GET_CODE (no_subreg_reg_operand[1]) != PLUS)))
2854 /* We have a move insn and a new reload insn will be similar
2855 to the current insn. We should avoid such situation as
2856 it results in LRA cycling. */
2857 if (lra_dump_file != NULL)
2858 fprintf (lra_dump_file,
2859 " Cycle danger: overall += LRA_MAX_REJECT\n");
2860 overall += LRA_MAX_REJECT;
2862 ok_p = true;
2863 curr_alt_dont_inherit_ops_num = 0;
2864 for (nop = 0; nop < early_clobbered_regs_num; nop++)
2866 int i, j, clobbered_hard_regno, first_conflict_j, last_conflict_j;
2867 HARD_REG_SET temp_set;
2869 i = early_clobbered_nops[nop];
2870 if ((! curr_alt_win[i] && ! curr_alt_match_win[i])
2871 || hard_regno[i] < 0)
2872 continue;
2873 lra_assert (operand_reg[i] != NULL_RTX);
2874 clobbered_hard_regno = hard_regno[i];
2875 CLEAR_HARD_REG_SET (temp_set);
2876 add_to_hard_reg_set (&temp_set, biggest_mode[i], clobbered_hard_regno);
2877 first_conflict_j = last_conflict_j = -1;
2878 for (j = 0; j < n_operands; j++)
2879 if (j == i
2880 /* We don't want process insides of match_operator and
2881 match_parallel because otherwise we would process
2882 their operands once again generating a wrong
2883 code. */
2884 || curr_static_id->operand[j].is_operator)
2885 continue;
2886 else if ((curr_alt_matches[j] == i && curr_alt_match_win[j])
2887 || (curr_alt_matches[i] == j && curr_alt_match_win[i]))
2888 continue;
2889 /* If we don't reload j-th operand, check conflicts. */
2890 else if ((curr_alt_win[j] || curr_alt_match_win[j])
2891 && uses_hard_regs_p (*curr_id->operand_loc[j], temp_set))
2893 if (first_conflict_j < 0)
2894 first_conflict_j = j;
2895 last_conflict_j = j;
2897 if (last_conflict_j < 0)
2898 continue;
2899 /* If earlyclobber operand conflicts with another
2900 non-matching operand which is actually the same register
2901 as the earlyclobber operand, it is better to reload the
2902 another operand as an operand matching the earlyclobber
2903 operand can be also the same. */
2904 if (first_conflict_j == last_conflict_j
2905 && operand_reg[last_conflict_j] != NULL_RTX
2906 && ! curr_alt_match_win[last_conflict_j]
2907 && REGNO (operand_reg[i]) == REGNO (operand_reg[last_conflict_j]))
2909 curr_alt_win[last_conflict_j] = false;
2910 curr_alt_dont_inherit_ops[curr_alt_dont_inherit_ops_num++]
2911 = last_conflict_j;
2912 losers++;
2913 /* Early clobber was already reflected in REJECT. */
2914 lra_assert (reject > 0);
2915 if (lra_dump_file != NULL)
2916 fprintf
2917 (lra_dump_file,
2918 " %d Conflict early clobber reload: reject--\n",
2920 reject--;
2921 overall += LRA_LOSER_COST_FACTOR - 1;
2923 else
2925 /* We need to reload early clobbered register and the
2926 matched registers. */
2927 for (j = 0; j < n_operands; j++)
2928 if (curr_alt_matches[j] == i)
2930 curr_alt_match_win[j] = false;
2931 losers++;
2932 overall += LRA_LOSER_COST_FACTOR;
2934 if (! curr_alt_match_win[i])
2935 curr_alt_dont_inherit_ops[curr_alt_dont_inherit_ops_num++] = i;
2936 else
2938 /* Remember pseudos used for match reloads are never
2939 inherited. */
2940 lra_assert (curr_alt_matches[i] >= 0);
2941 curr_alt_win[curr_alt_matches[i]] = false;
2943 curr_alt_win[i] = curr_alt_match_win[i] = false;
2944 losers++;
2945 /* Early clobber was already reflected in REJECT. */
2946 lra_assert (reject > 0);
2947 if (lra_dump_file != NULL)
2948 fprintf
2949 (lra_dump_file,
2950 " %d Matched conflict early clobber reloads: "
2951 "reject--\n",
2953 reject--;
2954 overall += LRA_LOSER_COST_FACTOR - 1;
2957 if (lra_dump_file != NULL)
2958 fprintf (lra_dump_file, " alt=%d,overall=%d,losers=%d,rld_nregs=%d\n",
2959 nalt, overall, losers, reload_nregs);
2961 /* If this alternative can be made to work by reloading, and it
2962 needs less reloading than the others checked so far, record
2963 it as the chosen goal for reloading. */
2964 if ((best_losers != 0 && losers == 0)
2965 || (((best_losers == 0 && losers == 0)
2966 || (best_losers != 0 && losers != 0))
2967 && (best_overall > overall
2968 || (best_overall == overall
2969 /* If the cost of the reloads is the same,
2970 prefer alternative which requires minimal
2971 number of reload regs. */
2972 && (reload_nregs < best_reload_nregs
2973 || (reload_nregs == best_reload_nregs
2974 && (best_reload_sum < reload_sum
2975 || (best_reload_sum == reload_sum
2976 && nalt < goal_alt_number))))))))
2978 for (nop = 0; nop < n_operands; nop++)
2980 goal_alt_win[nop] = curr_alt_win[nop];
2981 goal_alt_match_win[nop] = curr_alt_match_win[nop];
2982 goal_alt_matches[nop] = curr_alt_matches[nop];
2983 goal_alt[nop] = curr_alt[nop];
2984 goal_alt_offmemok[nop] = curr_alt_offmemok[nop];
2986 goal_alt_dont_inherit_ops_num = curr_alt_dont_inherit_ops_num;
2987 for (nop = 0; nop < curr_alt_dont_inherit_ops_num; nop++)
2988 goal_alt_dont_inherit_ops[nop] = curr_alt_dont_inherit_ops[nop];
2989 goal_alt_swapped = curr_swapped;
2990 best_overall = overall;
2991 best_losers = losers;
2992 best_reload_nregs = reload_nregs;
2993 best_reload_sum = reload_sum;
2994 goal_alt_number = nalt;
2996 if (losers == 0)
2997 /* Everything is satisfied. Do not process alternatives
2998 anymore. */
2999 break;
3000 fail:
3003 return ok_p;
3006 /* Make reload base reg from address AD. */
3007 static rtx
3008 base_to_reg (struct address_info *ad)
3010 enum reg_class cl;
3011 int code = -1;
3012 rtx new_inner = NULL_RTX;
3013 rtx new_reg = NULL_RTX;
3014 rtx_insn *insn;
3015 rtx_insn *last_insn = get_last_insn();
3017 lra_assert (ad->disp == ad->disp_term);
3018 cl = base_reg_class (ad->mode, ad->as, ad->base_outer_code,
3019 get_index_code (ad));
3020 new_reg = lra_create_new_reg (GET_MODE (*ad->base), NULL_RTX,
3021 cl, "base");
3022 new_inner = simplify_gen_binary (PLUS, GET_MODE (new_reg), new_reg,
3023 ad->disp_term == NULL
3024 ? const0_rtx
3025 : *ad->disp_term);
3026 if (!valid_address_p (ad->mode, new_inner, ad->as))
3027 return NULL_RTX;
3028 insn = emit_insn (gen_rtx_SET (new_reg, *ad->base));
3029 code = recog_memoized (insn);
3030 if (code < 0)
3032 delete_insns_since (last_insn);
3033 return NULL_RTX;
3036 return new_inner;
3039 /* Make reload base reg + disp from address AD. Return the new pseudo. */
3040 static rtx
3041 base_plus_disp_to_reg (struct address_info *ad)
3043 enum reg_class cl;
3044 rtx new_reg;
3046 lra_assert (ad->base == ad->base_term && ad->disp == ad->disp_term);
3047 cl = base_reg_class (ad->mode, ad->as, ad->base_outer_code,
3048 get_index_code (ad));
3049 new_reg = lra_create_new_reg (GET_MODE (*ad->base_term), NULL_RTX,
3050 cl, "base + disp");
3051 lra_emit_add (new_reg, *ad->base_term, *ad->disp_term);
3052 return new_reg;
3055 /* Make reload of index part of address AD. Return the new
3056 pseudo. */
3057 static rtx
3058 index_part_to_reg (struct address_info *ad)
3060 rtx new_reg;
3062 new_reg = lra_create_new_reg (GET_MODE (*ad->index), NULL_RTX,
3063 INDEX_REG_CLASS, "index term");
3064 expand_mult (GET_MODE (*ad->index), *ad->index_term,
3065 GEN_INT (get_index_scale (ad)), new_reg, 1);
3066 return new_reg;
3069 /* Return true if we can add a displacement to address AD, even if that
3070 makes the address invalid. The fix-up code requires any new address
3071 to be the sum of the BASE_TERM, INDEX and DISP_TERM fields. */
3072 static bool
3073 can_add_disp_p (struct address_info *ad)
3075 return (!ad->autoinc_p
3076 && ad->segment == NULL
3077 && ad->base == ad->base_term
3078 && ad->disp == ad->disp_term);
3081 /* Make equiv substitution in address AD. Return true if a substitution
3082 was made. */
3083 static bool
3084 equiv_address_substitution (struct address_info *ad)
3086 rtx base_reg, new_base_reg, index_reg, new_index_reg, *base_term, *index_term;
3087 poly_int64 disp;
3088 HOST_WIDE_INT scale;
3089 bool change_p;
3091 base_term = strip_subreg (ad->base_term);
3092 if (base_term == NULL)
3093 base_reg = new_base_reg = NULL_RTX;
3094 else
3096 base_reg = *base_term;
3097 new_base_reg = get_equiv_with_elimination (base_reg, curr_insn);
3099 index_term = strip_subreg (ad->index_term);
3100 if (index_term == NULL)
3101 index_reg = new_index_reg = NULL_RTX;
3102 else
3104 index_reg = *index_term;
3105 new_index_reg = get_equiv_with_elimination (index_reg, curr_insn);
3107 if (base_reg == new_base_reg && index_reg == new_index_reg)
3108 return false;
3109 disp = 0;
3110 change_p = false;
3111 if (lra_dump_file != NULL)
3113 fprintf (lra_dump_file, "Changing address in insn %d ",
3114 INSN_UID (curr_insn));
3115 dump_value_slim (lra_dump_file, *ad->outer, 1);
3117 if (base_reg != new_base_reg)
3119 poly_int64 offset;
3120 if (REG_P (new_base_reg))
3122 *base_term = new_base_reg;
3123 change_p = true;
3125 else if (GET_CODE (new_base_reg) == PLUS
3126 && REG_P (XEXP (new_base_reg, 0))
3127 && poly_int_rtx_p (XEXP (new_base_reg, 1), &offset)
3128 && can_add_disp_p (ad))
3130 disp += offset;
3131 *base_term = XEXP (new_base_reg, 0);
3132 change_p = true;
3134 if (ad->base_term2 != NULL)
3135 *ad->base_term2 = *ad->base_term;
3137 if (index_reg != new_index_reg)
3139 poly_int64 offset;
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 && poly_int_rtx_p (XEXP (new_index_reg, 1), &offset)
3148 && can_add_disp_p (ad)
3149 && (scale = get_index_scale (ad)))
3151 disp += offset * scale;
3152 *index_term = XEXP (new_index_reg, 0);
3153 change_p = true;
3156 if (maybe_ne (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 && !(INSN_CODE (curr_insn) < 0
3229 && get_constraint_type (cn) == CT_FIXED_FORM
3230 && constraint_satisfied_p (op, cn)))
3231 decompose_mem_address (&ad, op);
3232 else if (GET_CODE (op) == SUBREG
3233 && MEM_P (SUBREG_REG (op)))
3234 decompose_mem_address (&ad, SUBREG_REG (op));
3235 else
3236 return false;
3237 /* If INDEX_REG_CLASS is assigned to base_term already and isn't to
3238 index_term, swap them so to avoid assigning INDEX_REG_CLASS to both
3239 when INDEX_REG_CLASS is a single register class. */
3240 if (ad.base_term != NULL
3241 && ad.index_term != NULL
3242 && ira_class_hard_regs_num[INDEX_REG_CLASS] == 1
3243 && REG_P (*ad.base_term)
3244 && REG_P (*ad.index_term)
3245 && in_class_p (*ad.base_term, INDEX_REG_CLASS, NULL)
3246 && ! in_class_p (*ad.index_term, INDEX_REG_CLASS, NULL))
3248 std::swap (ad.base, ad.index);
3249 std::swap (ad.base_term, ad.index_term);
3251 if (! check_only_p)
3252 change_p = equiv_address_substitution (&ad);
3253 if (ad.base_term != NULL
3254 && (process_addr_reg
3255 (ad.base_term, check_only_p, before,
3256 (ad.autoinc_p
3257 && !(REG_P (*ad.base_term)
3258 && find_regno_note (curr_insn, REG_DEAD,
3259 REGNO (*ad.base_term)) != NULL_RTX)
3260 ? after : NULL),
3261 base_reg_class (ad.mode, ad.as, ad.base_outer_code,
3262 get_index_code (&ad)))))
3264 change_p = true;
3265 if (ad.base_term2 != NULL)
3266 *ad.base_term2 = *ad.base_term;
3268 if (ad.index_term != NULL
3269 && process_addr_reg (ad.index_term, check_only_p,
3270 before, NULL, INDEX_REG_CLASS))
3271 change_p = true;
3273 /* Target hooks sometimes don't treat extra-constraint addresses as
3274 legitimate address_operands, so handle them specially. */
3275 if (insn_extra_address_constraint (cn)
3276 && satisfies_address_constraint_p (&ad, cn))
3277 return change_p;
3279 if (check_only_p)
3280 return change_p;
3282 /* There are three cases where the shape of *AD.INNER may now be invalid:
3284 1) the original address was valid, but either elimination or
3285 equiv_address_substitution was applied and that made
3286 the address invalid.
3288 2) the address is an invalid symbolic address created by
3289 force_const_to_mem.
3291 3) the address is a frame address with an invalid offset.
3293 4) the address is a frame address with an invalid base.
3295 All these cases involve a non-autoinc address, so there is no
3296 point revalidating other types. */
3297 if (ad.autoinc_p || valid_address_p (&ad))
3298 return change_p;
3300 /* Any index existed before LRA started, so we can assume that the
3301 presence and shape of the index is valid. */
3302 push_to_sequence (*before);
3303 lra_assert (ad.disp == ad.disp_term);
3304 if (ad.base == NULL)
3306 if (ad.index == NULL)
3308 rtx_insn *insn;
3309 rtx_insn *last = get_last_insn ();
3310 int code = -1;
3311 enum reg_class cl = base_reg_class (ad.mode, ad.as,
3312 SCRATCH, SCRATCH);
3313 rtx addr = *ad.inner;
3315 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "addr");
3316 if (HAVE_lo_sum)
3318 /* addr => lo_sum (new_base, addr), case (2) above. */
3319 insn = emit_insn (gen_rtx_SET
3320 (new_reg,
3321 gen_rtx_HIGH (Pmode, copy_rtx (addr))));
3322 code = recog_memoized (insn);
3323 if (code >= 0)
3325 *ad.inner = gen_rtx_LO_SUM (Pmode, new_reg, addr);
3326 if (! valid_address_p (ad.mode, *ad.outer, ad.as))
3328 /* Try to put lo_sum into register. */
3329 insn = emit_insn (gen_rtx_SET
3330 (new_reg,
3331 gen_rtx_LO_SUM (Pmode, new_reg, addr)));
3332 code = recog_memoized (insn);
3333 if (code >= 0)
3335 *ad.inner = new_reg;
3336 if (! valid_address_p (ad.mode, *ad.outer, ad.as))
3338 *ad.inner = addr;
3339 code = -1;
3345 if (code < 0)
3346 delete_insns_since (last);
3349 if (code < 0)
3351 /* addr => new_base, case (2) above. */
3352 lra_emit_move (new_reg, addr);
3354 for (insn = last == NULL_RTX ? get_insns () : NEXT_INSN (last);
3355 insn != NULL_RTX;
3356 insn = NEXT_INSN (insn))
3357 if (recog_memoized (insn) < 0)
3358 break;
3359 if (insn != NULL_RTX)
3361 /* Do nothing if we cannot generate right insns.
3362 This is analogous to reload pass behavior. */
3363 delete_insns_since (last);
3364 end_sequence ();
3365 return false;
3367 *ad.inner = new_reg;
3370 else
3372 /* index * scale + disp => new base + index * scale,
3373 case (1) above. */
3374 enum reg_class cl = base_reg_class (ad.mode, ad.as, PLUS,
3375 GET_CODE (*ad.index));
3377 lra_assert (INDEX_REG_CLASS != NO_REGS);
3378 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "disp");
3379 lra_emit_move (new_reg, *ad.disp);
3380 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3381 new_reg, *ad.index);
3384 else if (ad.index == NULL)
3386 int regno;
3387 enum reg_class cl;
3388 rtx set;
3389 rtx_insn *insns, *last_insn;
3390 /* Try to reload base into register only if the base is invalid
3391 for the address but with valid offset, case (4) above. */
3392 start_sequence ();
3393 new_reg = base_to_reg (&ad);
3395 /* base + disp => new base, cases (1) and (3) above. */
3396 /* Another option would be to reload the displacement into an
3397 index register. However, postreload has code to optimize
3398 address reloads that have the same base and different
3399 displacements, so reloading into an index register would
3400 not necessarily be a win. */
3401 if (new_reg == NULL_RTX)
3402 new_reg = base_plus_disp_to_reg (&ad);
3403 insns = get_insns ();
3404 last_insn = get_last_insn ();
3405 /* If we generated at least two insns, try last insn source as
3406 an address. If we succeed, we generate one less insn. */
3407 if (last_insn != insns && (set = single_set (last_insn)) != NULL_RTX
3408 && GET_CODE (SET_SRC (set)) == PLUS
3409 && REG_P (XEXP (SET_SRC (set), 0))
3410 && CONSTANT_P (XEXP (SET_SRC (set), 1)))
3412 *ad.inner = SET_SRC (set);
3413 if (valid_address_p (ad.mode, *ad.outer, ad.as))
3415 *ad.base_term = XEXP (SET_SRC (set), 0);
3416 *ad.disp_term = XEXP (SET_SRC (set), 1);
3417 cl = base_reg_class (ad.mode, ad.as, ad.base_outer_code,
3418 get_index_code (&ad));
3419 regno = REGNO (*ad.base_term);
3420 if (regno >= FIRST_PSEUDO_REGISTER
3421 && cl != lra_get_allocno_class (regno))
3422 lra_change_class (regno, cl, " Change to", true);
3423 new_reg = SET_SRC (set);
3424 delete_insns_since (PREV_INSN (last_insn));
3427 /* Try if target can split displacement into legitimite new disp
3428 and offset. If it's the case, we replace the last insn with
3429 insns for base + offset => new_reg and set new_reg + new disp
3430 to *ad.inner. */
3431 last_insn = get_last_insn ();
3432 if ((set = single_set (last_insn)) != NULL_RTX
3433 && GET_CODE (SET_SRC (set)) == PLUS
3434 && REG_P (XEXP (SET_SRC (set), 0))
3435 && REGNO (XEXP (SET_SRC (set), 0)) < FIRST_PSEUDO_REGISTER
3436 && CONST_INT_P (XEXP (SET_SRC (set), 1)))
3438 rtx addend, disp = XEXP (SET_SRC (set), 1);
3439 if (targetm.legitimize_address_displacement (&disp, &addend,
3440 ad.mode))
3442 rtx_insn *new_insns;
3443 start_sequence ();
3444 lra_emit_add (new_reg, XEXP (SET_SRC (set), 0), addend);
3445 new_insns = get_insns ();
3446 end_sequence ();
3447 new_reg = gen_rtx_PLUS (Pmode, new_reg, disp);
3448 delete_insns_since (PREV_INSN (last_insn));
3449 add_insn (new_insns);
3450 insns = get_insns ();
3453 end_sequence ();
3454 emit_insn (insns);
3455 *ad.inner = new_reg;
3457 else if (ad.disp_term != NULL)
3459 /* base + scale * index + disp => new base + scale * index,
3460 case (1) above. */
3461 new_reg = base_plus_disp_to_reg (&ad);
3462 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3463 new_reg, *ad.index);
3465 else if ((scale = get_index_scale (&ad)) == 1)
3467 /* The last transformation to one reg will be made in
3468 curr_insn_transform function. */
3469 end_sequence ();
3470 return false;
3472 else if (scale != 0)
3474 /* base + scale * index => base + new_reg,
3475 case (1) above.
3476 Index part of address may become invalid. For example, we
3477 changed pseudo on the equivalent memory and a subreg of the
3478 pseudo onto the memory of different mode for which the scale is
3479 prohibitted. */
3480 new_reg = index_part_to_reg (&ad);
3481 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3482 *ad.base_term, new_reg);
3484 else
3486 enum reg_class cl = base_reg_class (ad.mode, ad.as,
3487 SCRATCH, SCRATCH);
3488 rtx addr = *ad.inner;
3490 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "addr");
3491 /* addr => new_base. */
3492 lra_emit_move (new_reg, addr);
3493 *ad.inner = new_reg;
3495 *before = get_insns ();
3496 end_sequence ();
3497 return true;
3500 /* If CHECK_ONLY_P is false, do address reloads until it is necessary.
3501 Use process_address_1 as a helper function. Return true for any
3502 RTL changes.
3504 If CHECK_ONLY_P is true, just check address correctness. Return
3505 false if the address correct. */
3506 static bool
3507 process_address (int nop, bool check_only_p,
3508 rtx_insn **before, rtx_insn **after)
3510 bool res = false;
3512 while (process_address_1 (nop, check_only_p, before, after))
3514 if (check_only_p)
3515 return true;
3516 res = true;
3518 return res;
3521 /* Emit insns to reload VALUE into a new register. VALUE is an
3522 auto-increment or auto-decrement RTX whose operand is a register or
3523 memory location; so reloading involves incrementing that location.
3524 IN is either identical to VALUE, or some cheaper place to reload
3525 value being incremented/decremented from.
3527 INC_AMOUNT is the number to increment or decrement by (always
3528 positive and ignored for POST_MODIFY/PRE_MODIFY).
3530 Return pseudo containing the result. */
3531 static rtx
3532 emit_inc (enum reg_class new_rclass, rtx in, rtx value, int inc_amount)
3534 /* REG or MEM to be copied and incremented. */
3535 rtx incloc = XEXP (value, 0);
3536 /* Nonzero if increment after copying. */
3537 int post = (GET_CODE (value) == POST_DEC || GET_CODE (value) == POST_INC
3538 || GET_CODE (value) == POST_MODIFY);
3539 rtx_insn *last;
3540 rtx inc;
3541 rtx_insn *add_insn;
3542 int code;
3543 rtx real_in = in == value ? incloc : in;
3544 rtx result;
3545 bool plus_p = true;
3547 if (GET_CODE (value) == PRE_MODIFY || GET_CODE (value) == POST_MODIFY)
3549 lra_assert (GET_CODE (XEXP (value, 1)) == PLUS
3550 || GET_CODE (XEXP (value, 1)) == MINUS);
3551 lra_assert (rtx_equal_p (XEXP (XEXP (value, 1), 0), XEXP (value, 0)));
3552 plus_p = GET_CODE (XEXP (value, 1)) == PLUS;
3553 inc = XEXP (XEXP (value, 1), 1);
3555 else
3557 if (GET_CODE (value) == PRE_DEC || GET_CODE (value) == POST_DEC)
3558 inc_amount = -inc_amount;
3560 inc = GEN_INT (inc_amount);
3563 if (! post && REG_P (incloc))
3564 result = incloc;
3565 else
3566 result = lra_create_new_reg (GET_MODE (value), value, new_rclass,
3567 "INC/DEC result");
3569 if (real_in != result)
3571 /* First copy the location to the result register. */
3572 lra_assert (REG_P (result));
3573 emit_insn (gen_move_insn (result, real_in));
3576 /* We suppose that there are insns to add/sub with the constant
3577 increment permitted in {PRE/POST)_{DEC/INC/MODIFY}. At least the
3578 old reload worked with this assumption. If the assumption
3579 becomes wrong, we should use approach in function
3580 base_plus_disp_to_reg. */
3581 if (in == value)
3583 /* See if we can directly increment INCLOC. */
3584 last = get_last_insn ();
3585 add_insn = emit_insn (plus_p
3586 ? gen_add2_insn (incloc, inc)
3587 : gen_sub2_insn (incloc, inc));
3589 code = recog_memoized (add_insn);
3590 if (code >= 0)
3592 if (! post && result != incloc)
3593 emit_insn (gen_move_insn (result, incloc));
3594 return result;
3596 delete_insns_since (last);
3599 /* If couldn't do the increment directly, must increment in RESULT.
3600 The way we do this depends on whether this is pre- or
3601 post-increment. For pre-increment, copy INCLOC to the reload
3602 register, increment it there, then save back. */
3603 if (! post)
3605 if (real_in != result)
3606 emit_insn (gen_move_insn (result, real_in));
3607 if (plus_p)
3608 emit_insn (gen_add2_insn (result, inc));
3609 else
3610 emit_insn (gen_sub2_insn (result, inc));
3611 if (result != incloc)
3612 emit_insn (gen_move_insn (incloc, result));
3614 else
3616 /* Post-increment.
3618 Because this might be a jump insn or a compare, and because
3619 RESULT may not be available after the insn in an input
3620 reload, we must do the incrementing before the insn being
3621 reloaded for.
3623 We have already copied IN to RESULT. Increment the copy in
3624 RESULT, save that back, then decrement RESULT so it has
3625 the original value. */
3626 if (plus_p)
3627 emit_insn (gen_add2_insn (result, inc));
3628 else
3629 emit_insn (gen_sub2_insn (result, inc));
3630 emit_insn (gen_move_insn (incloc, result));
3631 /* Restore non-modified value for the result. We prefer this
3632 way because it does not require an additional hard
3633 register. */
3634 if (plus_p)
3636 poly_int64 offset;
3637 if (poly_int_rtx_p (inc, &offset))
3638 emit_insn (gen_add2_insn (result,
3639 gen_int_mode (-offset,
3640 GET_MODE (result))));
3641 else
3642 emit_insn (gen_sub2_insn (result, inc));
3644 else
3645 emit_insn (gen_add2_insn (result, inc));
3647 return result;
3650 /* Return true if the current move insn does not need processing as we
3651 already know that it satisfies its constraints. */
3652 static bool
3653 simple_move_p (void)
3655 rtx dest, src;
3656 enum reg_class dclass, sclass;
3658 lra_assert (curr_insn_set != NULL_RTX);
3659 dest = SET_DEST (curr_insn_set);
3660 src = SET_SRC (curr_insn_set);
3662 /* If the instruction has multiple sets we need to process it even if it
3663 is single_set. This can happen if one or more of the SETs are dead.
3664 See PR73650. */
3665 if (multiple_sets (curr_insn))
3666 return false;
3668 return ((dclass = get_op_class (dest)) != NO_REGS
3669 && (sclass = get_op_class (src)) != NO_REGS
3670 /* The backend guarantees that register moves of cost 2
3671 never need reloads. */
3672 && targetm.register_move_cost (GET_MODE (src), sclass, dclass) == 2);
3675 /* Swap operands NOP and NOP + 1. */
3676 static inline void
3677 swap_operands (int nop)
3679 std::swap (curr_operand_mode[nop], curr_operand_mode[nop + 1]);
3680 std::swap (original_subreg_reg_mode[nop], original_subreg_reg_mode[nop + 1]);
3681 std::swap (*curr_id->operand_loc[nop], *curr_id->operand_loc[nop + 1]);
3682 std::swap (equiv_substition_p[nop], equiv_substition_p[nop + 1]);
3683 /* Swap the duplicates too. */
3684 lra_update_dup (curr_id, nop);
3685 lra_update_dup (curr_id, nop + 1);
3688 /* Main entry point of the constraint code: search the body of the
3689 current insn to choose the best alternative. It is mimicking insn
3690 alternative cost calculation model of former reload pass. That is
3691 because machine descriptions were written to use this model. This
3692 model can be changed in future. Make commutative operand exchange
3693 if it is chosen.
3695 if CHECK_ONLY_P is false, do RTL changes to satisfy the
3696 constraints. Return true if any change happened during function
3697 call.
3699 If CHECK_ONLY_P is true then don't do any transformation. Just
3700 check that the insn satisfies all constraints. If the insn does
3701 not satisfy any constraint, return true. */
3702 static bool
3703 curr_insn_transform (bool check_only_p)
3705 int i, j, k;
3706 int n_operands;
3707 int n_alternatives;
3708 int n_outputs;
3709 int commutative;
3710 signed char goal_alt_matched[MAX_RECOG_OPERANDS][MAX_RECOG_OPERANDS];
3711 signed char match_inputs[MAX_RECOG_OPERANDS + 1];
3712 signed char outputs[MAX_RECOG_OPERANDS + 1];
3713 rtx_insn *before, *after;
3714 bool alt_p = false;
3715 /* Flag that the insn has been changed through a transformation. */
3716 bool change_p;
3717 bool sec_mem_p;
3718 bool use_sec_mem_p;
3719 int max_regno_before;
3720 int reused_alternative_num;
3722 curr_insn_set = single_set (curr_insn);
3723 if (curr_insn_set != NULL_RTX && simple_move_p ())
3724 return false;
3726 no_input_reloads_p = no_output_reloads_p = false;
3727 goal_alt_number = -1;
3728 change_p = sec_mem_p = false;
3729 /* JUMP_INSNs and CALL_INSNs are not allowed to have any output
3730 reloads; neither are insns that SET cc0. Insns that use CC0 are
3731 not allowed to have any input reloads. */
3732 if (JUMP_P (curr_insn) || CALL_P (curr_insn))
3733 no_output_reloads_p = true;
3735 if (HAVE_cc0 && reg_referenced_p (cc0_rtx, PATTERN (curr_insn)))
3736 no_input_reloads_p = true;
3737 if (HAVE_cc0 && reg_set_p (cc0_rtx, PATTERN (curr_insn)))
3738 no_output_reloads_p = true;
3740 n_operands = curr_static_id->n_operands;
3741 n_alternatives = curr_static_id->n_alternatives;
3743 /* Just return "no reloads" if insn has no operands with
3744 constraints. */
3745 if (n_operands == 0 || n_alternatives == 0)
3746 return false;
3748 max_regno_before = max_reg_num ();
3750 for (i = 0; i < n_operands; i++)
3752 goal_alt_matched[i][0] = -1;
3753 goal_alt_matches[i] = -1;
3756 commutative = curr_static_id->commutative;
3758 /* Now see what we need for pseudos that didn't get hard regs or got
3759 the wrong kind of hard reg. For this, we must consider all the
3760 operands together against the register constraints. */
3762 best_losers = best_overall = INT_MAX;
3763 best_reload_sum = 0;
3765 curr_swapped = false;
3766 goal_alt_swapped = false;
3768 if (! check_only_p)
3769 /* Make equivalence substitution and memory subreg elimination
3770 before address processing because an address legitimacy can
3771 depend on memory mode. */
3772 for (i = 0; i < n_operands; i++)
3774 rtx op, subst, old;
3775 bool op_change_p = false;
3777 if (curr_static_id->operand[i].is_operator)
3778 continue;
3780 old = op = *curr_id->operand_loc[i];
3781 if (GET_CODE (old) == SUBREG)
3782 old = SUBREG_REG (old);
3783 subst = get_equiv_with_elimination (old, curr_insn);
3784 original_subreg_reg_mode[i] = VOIDmode;
3785 equiv_substition_p[i] = false;
3786 if (subst != old)
3788 equiv_substition_p[i] = true;
3789 subst = copy_rtx (subst);
3790 lra_assert (REG_P (old));
3791 if (GET_CODE (op) != SUBREG)
3792 *curr_id->operand_loc[i] = subst;
3793 else
3795 SUBREG_REG (op) = subst;
3796 if (GET_MODE (subst) == VOIDmode)
3797 original_subreg_reg_mode[i] = GET_MODE (old);
3799 if (lra_dump_file != NULL)
3801 fprintf (lra_dump_file,
3802 "Changing pseudo %d in operand %i of insn %u on equiv ",
3803 REGNO (old), i, INSN_UID (curr_insn));
3804 dump_value_slim (lra_dump_file, subst, 1);
3805 fprintf (lra_dump_file, "\n");
3807 op_change_p = change_p = true;
3809 if (simplify_operand_subreg (i, GET_MODE (old)) || op_change_p)
3811 change_p = true;
3812 lra_update_dup (curr_id, i);
3816 /* Reload address registers and displacements. We do it before
3817 finding an alternative because of memory constraints. */
3818 before = after = NULL;
3819 for (i = 0; i < n_operands; i++)
3820 if (! curr_static_id->operand[i].is_operator
3821 && process_address (i, check_only_p, &before, &after))
3823 if (check_only_p)
3824 return true;
3825 change_p = true;
3826 lra_update_dup (curr_id, i);
3829 if (change_p)
3830 /* If we've changed the instruction then any alternative that
3831 we chose previously may no longer be valid. */
3832 lra_set_used_insn_alternative (curr_insn, -1);
3834 if (! check_only_p && curr_insn_set != NULL_RTX
3835 && check_and_process_move (&change_p, &sec_mem_p))
3836 return change_p;
3838 try_swapped:
3840 reused_alternative_num = check_only_p ? -1 : curr_id->used_insn_alternative;
3841 if (lra_dump_file != NULL && reused_alternative_num >= 0)
3842 fprintf (lra_dump_file, "Reusing alternative %d for insn #%u\n",
3843 reused_alternative_num, INSN_UID (curr_insn));
3845 if (process_alt_operands (reused_alternative_num))
3846 alt_p = true;
3848 if (check_only_p)
3849 return ! alt_p || best_losers != 0;
3851 /* If insn is commutative (it's safe to exchange a certain pair of
3852 operands) then we need to try each alternative twice, the second
3853 time matching those two operands as if we had exchanged them. To
3854 do this, really exchange them in operands.
3856 If we have just tried the alternatives the second time, return
3857 operands to normal and drop through. */
3859 if (reused_alternative_num < 0 && commutative >= 0)
3861 curr_swapped = !curr_swapped;
3862 if (curr_swapped)
3864 swap_operands (commutative);
3865 goto try_swapped;
3867 else
3868 swap_operands (commutative);
3871 if (! alt_p && ! sec_mem_p)
3873 /* No alternative works with reloads?? */
3874 if (INSN_CODE (curr_insn) >= 0)
3875 fatal_insn ("unable to generate reloads for:", curr_insn);
3876 error_for_asm (curr_insn,
3877 "inconsistent operand constraints in an %<asm%>");
3878 /* Avoid further trouble with this insn. Don't generate use
3879 pattern here as we could use the insn SP offset. */
3880 lra_set_insn_deleted (curr_insn);
3881 return true;
3884 /* If the best alternative is with operands 1 and 2 swapped, swap
3885 them. Update the operand numbers of any reloads already
3886 pushed. */
3888 if (goal_alt_swapped)
3890 if (lra_dump_file != NULL)
3891 fprintf (lra_dump_file, " Commutative operand exchange in insn %u\n",
3892 INSN_UID (curr_insn));
3894 /* Swap the duplicates too. */
3895 swap_operands (commutative);
3896 change_p = true;
3899 /* Some targets' TARGET_SECONDARY_MEMORY_NEEDED (e.g. x86) are defined
3900 too conservatively. So we use the secondary memory only if there
3901 is no any alternative without reloads. */
3902 use_sec_mem_p = false;
3903 if (! alt_p)
3904 use_sec_mem_p = true;
3905 else if (sec_mem_p)
3907 for (i = 0; i < n_operands; i++)
3908 if (! goal_alt_win[i] && ! goal_alt_match_win[i])
3909 break;
3910 use_sec_mem_p = i < n_operands;
3913 if (use_sec_mem_p)
3915 int in = -1, out = -1;
3916 rtx new_reg, src, dest, rld;
3917 machine_mode sec_mode, rld_mode;
3919 lra_assert (curr_insn_set != NULL_RTX && sec_mem_p);
3920 dest = SET_DEST (curr_insn_set);
3921 src = SET_SRC (curr_insn_set);
3922 for (i = 0; i < n_operands; i++)
3923 if (*curr_id->operand_loc[i] == dest)
3924 out = i;
3925 else if (*curr_id->operand_loc[i] == src)
3926 in = i;
3927 for (i = 0; i < curr_static_id->n_dups; i++)
3928 if (out < 0 && *curr_id->dup_loc[i] == dest)
3929 out = curr_static_id->dup_num[i];
3930 else if (in < 0 && *curr_id->dup_loc[i] == src)
3931 in = curr_static_id->dup_num[i];
3932 lra_assert (out >= 0 && in >= 0
3933 && curr_static_id->operand[out].type == OP_OUT
3934 && curr_static_id->operand[in].type == OP_IN);
3935 rld = partial_subreg_p (GET_MODE (src), GET_MODE (dest)) ? src : dest;
3936 rld_mode = GET_MODE (rld);
3937 sec_mode = targetm.secondary_memory_needed_mode (rld_mode);
3938 new_reg = lra_create_new_reg (sec_mode, NULL_RTX,
3939 NO_REGS, "secondary");
3940 /* If the mode is changed, it should be wider. */
3941 lra_assert (!partial_subreg_p (sec_mode, rld_mode));
3942 if (sec_mode != rld_mode)
3944 /* If the target says specifically to use another mode for
3945 secondary memory moves we can not reuse the original
3946 insn. */
3947 after = emit_spill_move (false, new_reg, dest);
3948 lra_process_new_insns (curr_insn, NULL, after,
3949 "Inserting the sec. move");
3950 /* We may have non null BEFORE here (e.g. after address
3951 processing. */
3952 push_to_sequence (before);
3953 before = emit_spill_move (true, new_reg, src);
3954 emit_insn (before);
3955 before = get_insns ();
3956 end_sequence ();
3957 lra_process_new_insns (curr_insn, before, NULL, "Changing on");
3958 lra_set_insn_deleted (curr_insn);
3960 else if (dest == rld)
3962 *curr_id->operand_loc[out] = new_reg;
3963 lra_update_dup (curr_id, out);
3964 after = emit_spill_move (false, new_reg, dest);
3965 lra_process_new_insns (curr_insn, NULL, after,
3966 "Inserting the sec. move");
3968 else
3970 *curr_id->operand_loc[in] = new_reg;
3971 lra_update_dup (curr_id, in);
3972 /* See comments above. */
3973 push_to_sequence (before);
3974 before = emit_spill_move (true, new_reg, src);
3975 emit_insn (before);
3976 before = get_insns ();
3977 end_sequence ();
3978 lra_process_new_insns (curr_insn, before, NULL,
3979 "Inserting the sec. move");
3981 lra_update_insn_regno_info (curr_insn);
3982 return true;
3985 lra_assert (goal_alt_number >= 0);
3986 lra_set_used_insn_alternative (curr_insn, goal_alt_number);
3988 if (lra_dump_file != NULL)
3990 const char *p;
3992 fprintf (lra_dump_file, " Choosing alt %d in insn %u:",
3993 goal_alt_number, INSN_UID (curr_insn));
3994 for (i = 0; i < n_operands; i++)
3996 p = (curr_static_id->operand_alternative
3997 [goal_alt_number * n_operands + i].constraint);
3998 if (*p == '\0')
3999 continue;
4000 fprintf (lra_dump_file, " (%d) ", i);
4001 for (; *p != '\0' && *p != ',' && *p != '#'; p++)
4002 fputc (*p, lra_dump_file);
4004 if (INSN_CODE (curr_insn) >= 0
4005 && (p = get_insn_name (INSN_CODE (curr_insn))) != NULL)
4006 fprintf (lra_dump_file, " {%s}", p);
4007 if (maybe_ne (curr_id->sp_offset, 0))
4009 fprintf (lra_dump_file, " (sp_off=");
4010 print_dec (curr_id->sp_offset, lra_dump_file);
4011 fprintf (lra_dump_file, ")");
4013 fprintf (lra_dump_file, "\n");
4016 /* Right now, for any pair of operands I and J that are required to
4017 match, with J < I, goal_alt_matches[I] is J. Add I to
4018 goal_alt_matched[J]. */
4020 for (i = 0; i < n_operands; i++)
4021 if ((j = goal_alt_matches[i]) >= 0)
4023 for (k = 0; goal_alt_matched[j][k] >= 0; k++)
4025 /* We allow matching one output operand and several input
4026 operands. */
4027 lra_assert (k == 0
4028 || (curr_static_id->operand[j].type == OP_OUT
4029 && curr_static_id->operand[i].type == OP_IN
4030 && (curr_static_id->operand
4031 [goal_alt_matched[j][0]].type == OP_IN)));
4032 goal_alt_matched[j][k] = i;
4033 goal_alt_matched[j][k + 1] = -1;
4036 for (i = 0; i < n_operands; i++)
4037 goal_alt_win[i] |= goal_alt_match_win[i];
4039 /* Any constants that aren't allowed and can't be reloaded into
4040 registers are here changed into memory references. */
4041 for (i = 0; i < n_operands; i++)
4042 if (goal_alt_win[i])
4044 int regno;
4045 enum reg_class new_class;
4046 rtx reg = *curr_id->operand_loc[i];
4048 if (GET_CODE (reg) == SUBREG)
4049 reg = SUBREG_REG (reg);
4051 if (REG_P (reg) && (regno = REGNO (reg)) >= FIRST_PSEUDO_REGISTER)
4053 bool ok_p = in_class_p (reg, goal_alt[i], &new_class);
4055 if (new_class != NO_REGS && get_reg_class (regno) != new_class)
4057 lra_assert (ok_p);
4058 lra_change_class (regno, new_class, " Change to", true);
4062 else
4064 const char *constraint;
4065 char c;
4066 rtx op = *curr_id->operand_loc[i];
4067 rtx subreg = NULL_RTX;
4068 machine_mode mode = curr_operand_mode[i];
4070 if (GET_CODE (op) == SUBREG)
4072 subreg = op;
4073 op = SUBREG_REG (op);
4074 mode = GET_MODE (op);
4077 if (CONST_POOL_OK_P (mode, op)
4078 && ((targetm.preferred_reload_class
4079 (op, (enum reg_class) goal_alt[i]) == NO_REGS)
4080 || no_input_reloads_p))
4082 rtx tem = force_const_mem (mode, op);
4084 change_p = true;
4085 if (subreg != NULL_RTX)
4086 tem = gen_rtx_SUBREG (mode, tem, SUBREG_BYTE (subreg));
4088 *curr_id->operand_loc[i] = tem;
4089 lra_update_dup (curr_id, i);
4090 process_address (i, false, &before, &after);
4092 /* If the alternative accepts constant pool refs directly
4093 there will be no reload needed at all. */
4094 if (subreg != NULL_RTX)
4095 continue;
4096 /* Skip alternatives before the one requested. */
4097 constraint = (curr_static_id->operand_alternative
4098 [goal_alt_number * n_operands + i].constraint);
4099 for (;
4100 (c = *constraint) && c != ',' && c != '#';
4101 constraint += CONSTRAINT_LEN (c, constraint))
4103 enum constraint_num cn = lookup_constraint (constraint);
4104 if ((insn_extra_memory_constraint (cn)
4105 || insn_extra_special_memory_constraint (cn))
4106 && satisfies_memory_constraint_p (tem, cn))
4107 break;
4109 if (c == '\0' || c == ',' || c == '#')
4110 continue;
4112 goal_alt_win[i] = true;
4116 n_outputs = 0;
4117 outputs[0] = -1;
4118 for (i = 0; i < n_operands; i++)
4120 int regno;
4121 bool optional_p = false;
4122 rtx old, new_reg;
4123 rtx op = *curr_id->operand_loc[i];
4125 if (goal_alt_win[i])
4127 if (goal_alt[i] == NO_REGS
4128 && REG_P (op)
4129 /* When we assign NO_REGS it means that we will not
4130 assign a hard register to the scratch pseudo by
4131 assigment pass and the scratch pseudo will be
4132 spilled. Spilled scratch pseudos are transformed
4133 back to scratches at the LRA end. */
4134 && lra_former_scratch_operand_p (curr_insn, i)
4135 && lra_former_scratch_p (REGNO (op)))
4137 int regno = REGNO (op);
4138 lra_change_class (regno, NO_REGS, " Change to", true);
4139 if (lra_get_regno_hard_regno (regno) >= 0)
4140 /* We don't have to mark all insn affected by the
4141 spilled pseudo as there is only one such insn, the
4142 current one. */
4143 reg_renumber[regno] = -1;
4144 lra_assert (bitmap_single_bit_set_p
4145 (&lra_reg_info[REGNO (op)].insn_bitmap));
4147 /* We can do an optional reload. If the pseudo got a hard
4148 reg, we might improve the code through inheritance. If
4149 it does not get a hard register we coalesce memory/memory
4150 moves later. Ignore move insns to avoid cycling. */
4151 if (! lra_simple_p
4152 && lra_undo_inheritance_iter < LRA_MAX_INHERITANCE_PASSES
4153 && goal_alt[i] != NO_REGS && REG_P (op)
4154 && (regno = REGNO (op)) >= FIRST_PSEUDO_REGISTER
4155 && regno < new_regno_start
4156 && ! lra_former_scratch_p (regno)
4157 && reg_renumber[regno] < 0
4158 /* Check that the optional reload pseudo will be able to
4159 hold given mode value. */
4160 && ! (prohibited_class_reg_set_mode_p
4161 (goal_alt[i], reg_class_contents[goal_alt[i]],
4162 PSEUDO_REGNO_MODE (regno)))
4163 && (curr_insn_set == NULL_RTX
4164 || !((REG_P (SET_SRC (curr_insn_set))
4165 || MEM_P (SET_SRC (curr_insn_set))
4166 || GET_CODE (SET_SRC (curr_insn_set)) == SUBREG)
4167 && (REG_P (SET_DEST (curr_insn_set))
4168 || MEM_P (SET_DEST (curr_insn_set))
4169 || GET_CODE (SET_DEST (curr_insn_set)) == SUBREG))))
4170 optional_p = true;
4171 else
4172 continue;
4175 /* Operands that match previous ones have already been handled. */
4176 if (goal_alt_matches[i] >= 0)
4177 continue;
4179 /* We should not have an operand with a non-offsettable address
4180 appearing where an offsettable address will do. It also may
4181 be a case when the address should be special in other words
4182 not a general one (e.g. it needs no index reg). */
4183 if (goal_alt_matched[i][0] == -1 && goal_alt_offmemok[i] && MEM_P (op))
4185 enum reg_class rclass;
4186 rtx *loc = &XEXP (op, 0);
4187 enum rtx_code code = GET_CODE (*loc);
4189 push_to_sequence (before);
4190 rclass = base_reg_class (GET_MODE (op), MEM_ADDR_SPACE (op),
4191 MEM, SCRATCH);
4192 if (GET_RTX_CLASS (code) == RTX_AUTOINC)
4193 new_reg = emit_inc (rclass, *loc, *loc,
4194 /* This value does not matter for MODIFY. */
4195 GET_MODE_SIZE (GET_MODE (op)));
4196 else if (get_reload_reg (OP_IN, Pmode, *loc, rclass, FALSE,
4197 "offsetable address", &new_reg))
4198 lra_emit_move (new_reg, *loc);
4199 before = get_insns ();
4200 end_sequence ();
4201 *loc = new_reg;
4202 lra_update_dup (curr_id, i);
4204 else if (goal_alt_matched[i][0] == -1)
4206 machine_mode mode;
4207 rtx reg, *loc;
4208 int hard_regno, byte;
4209 enum op_type type = curr_static_id->operand[i].type;
4211 loc = curr_id->operand_loc[i];
4212 mode = curr_operand_mode[i];
4213 if (GET_CODE (*loc) == SUBREG)
4215 reg = SUBREG_REG (*loc);
4216 byte = SUBREG_BYTE (*loc);
4217 if (REG_P (reg)
4218 /* Strict_low_part requires reloading the register and not
4219 just the subreg. Likewise for a strict subreg no wider
4220 than a word for WORD_REGISTER_OPERATIONS targets. */
4221 && (curr_static_id->operand[i].strict_low
4222 || (!paradoxical_subreg_p (mode, GET_MODE (reg))
4223 && (hard_regno
4224 = get_try_hard_regno (REGNO (reg))) >= 0
4225 && (simplify_subreg_regno
4226 (hard_regno,
4227 GET_MODE (reg), byte, mode) < 0)
4228 && (goal_alt[i] == NO_REGS
4229 || (simplify_subreg_regno
4230 (ira_class_hard_regs[goal_alt[i]][0],
4231 GET_MODE (reg), byte, mode) >= 0)))
4232 || (GET_MODE_PRECISION (mode)
4233 < GET_MODE_PRECISION (GET_MODE (reg))
4234 && GET_MODE_SIZE (GET_MODE (reg)) <= UNITS_PER_WORD
4235 && WORD_REGISTER_OPERATIONS)))
4237 /* An OP_INOUT is required when reloading a subreg of a
4238 mode wider than a word to ensure that data beyond the
4239 word being reloaded is preserved. Also automatically
4240 ensure that strict_low_part reloads are made into
4241 OP_INOUT which should already be true from the backend
4242 constraints. */
4243 if (type == OP_OUT
4244 && (curr_static_id->operand[i].strict_low
4245 || read_modify_subreg_p (*loc)))
4246 type = OP_INOUT;
4247 loc = &SUBREG_REG (*loc);
4248 mode = GET_MODE (*loc);
4251 old = *loc;
4252 if (get_reload_reg (type, mode, old, goal_alt[i],
4253 loc != curr_id->operand_loc[i], "", &new_reg)
4254 && type != OP_OUT)
4256 push_to_sequence (before);
4257 lra_emit_move (new_reg, old);
4258 before = get_insns ();
4259 end_sequence ();
4261 *loc = new_reg;
4262 if (type != OP_IN
4263 && find_reg_note (curr_insn, REG_UNUSED, old) == NULL_RTX)
4265 start_sequence ();
4266 lra_emit_move (type == OP_INOUT ? copy_rtx (old) : old, new_reg);
4267 emit_insn (after);
4268 after = get_insns ();
4269 end_sequence ();
4270 *loc = new_reg;
4272 for (j = 0; j < goal_alt_dont_inherit_ops_num; j++)
4273 if (goal_alt_dont_inherit_ops[j] == i)
4275 lra_set_regno_unique_value (REGNO (new_reg));
4276 break;
4278 lra_update_dup (curr_id, i);
4280 else if (curr_static_id->operand[i].type == OP_IN
4281 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4282 == OP_OUT
4283 || (curr_static_id->operand[goal_alt_matched[i][0]].type
4284 == OP_INOUT
4285 && (operands_match_p
4286 (*curr_id->operand_loc[i],
4287 *curr_id->operand_loc[goal_alt_matched[i][0]],
4288 -1)))))
4290 /* generate reloads for input and matched outputs. */
4291 match_inputs[0] = i;
4292 match_inputs[1] = -1;
4293 match_reload (goal_alt_matched[i][0], match_inputs, outputs,
4294 goal_alt[i], &before, &after,
4295 curr_static_id->operand_alternative
4296 [goal_alt_number * n_operands + goal_alt_matched[i][0]]
4297 .earlyclobber);
4299 else if ((curr_static_id->operand[i].type == OP_OUT
4300 || (curr_static_id->operand[i].type == OP_INOUT
4301 && (operands_match_p
4302 (*curr_id->operand_loc[i],
4303 *curr_id->operand_loc[goal_alt_matched[i][0]],
4304 -1))))
4305 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4306 == OP_IN))
4307 /* Generate reloads for output and matched inputs. */
4308 match_reload (i, goal_alt_matched[i], outputs, goal_alt[i], &before,
4309 &after, curr_static_id->operand_alternative
4310 [goal_alt_number * n_operands + i].earlyclobber);
4311 else if (curr_static_id->operand[i].type == OP_IN
4312 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4313 == OP_IN))
4315 /* Generate reloads for matched inputs. */
4316 match_inputs[0] = i;
4317 for (j = 0; (k = goal_alt_matched[i][j]) >= 0; j++)
4318 match_inputs[j + 1] = k;
4319 match_inputs[j + 1] = -1;
4320 match_reload (-1, match_inputs, outputs, goal_alt[i], &before,
4321 &after, false);
4323 else
4324 /* We must generate code in any case when function
4325 process_alt_operands decides that it is possible. */
4326 gcc_unreachable ();
4328 /* Memorise processed outputs so that output remaining to be processed
4329 can avoid using the same register value (see match_reload). */
4330 if (curr_static_id->operand[i].type == OP_OUT)
4332 outputs[n_outputs++] = i;
4333 outputs[n_outputs] = -1;
4336 if (optional_p)
4338 rtx reg = op;
4340 lra_assert (REG_P (reg));
4341 regno = REGNO (reg);
4342 op = *curr_id->operand_loc[i]; /* Substitution. */
4343 if (GET_CODE (op) == SUBREG)
4344 op = SUBREG_REG (op);
4345 gcc_assert (REG_P (op) && (int) REGNO (op) >= new_regno_start);
4346 bitmap_set_bit (&lra_optional_reload_pseudos, REGNO (op));
4347 lra_reg_info[REGNO (op)].restore_rtx = reg;
4348 if (lra_dump_file != NULL)
4349 fprintf (lra_dump_file,
4350 " Making reload reg %d for reg %d optional\n",
4351 REGNO (op), regno);
4354 if (before != NULL_RTX || after != NULL_RTX
4355 || max_regno_before != max_reg_num ())
4356 change_p = true;
4357 if (change_p)
4359 lra_update_operator_dups (curr_id);
4360 /* Something changes -- process the insn. */
4361 lra_update_insn_regno_info (curr_insn);
4363 lra_process_new_insns (curr_insn, before, after, "Inserting insn reload");
4364 return change_p;
4367 /* Return true if INSN satisfies all constraints. In other words, no
4368 reload insns are needed. */
4369 bool
4370 lra_constrain_insn (rtx_insn *insn)
4372 int saved_new_regno_start = new_regno_start;
4373 int saved_new_insn_uid_start = new_insn_uid_start;
4374 bool change_p;
4376 curr_insn = insn;
4377 curr_id = lra_get_insn_recog_data (curr_insn);
4378 curr_static_id = curr_id->insn_static_data;
4379 new_insn_uid_start = get_max_uid ();
4380 new_regno_start = max_reg_num ();
4381 change_p = curr_insn_transform (true);
4382 new_regno_start = saved_new_regno_start;
4383 new_insn_uid_start = saved_new_insn_uid_start;
4384 return ! change_p;
4387 /* Return true if X is in LIST. */
4388 static bool
4389 in_list_p (rtx x, rtx list)
4391 for (; list != NULL_RTX; list = XEXP (list, 1))
4392 if (XEXP (list, 0) == x)
4393 return true;
4394 return false;
4397 /* Return true if X contains an allocatable hard register (if
4398 HARD_REG_P) or a (spilled if SPILLED_P) pseudo. */
4399 static bool
4400 contains_reg_p (rtx x, bool hard_reg_p, bool spilled_p)
4402 int i, j;
4403 const char *fmt;
4404 enum rtx_code code;
4406 code = GET_CODE (x);
4407 if (REG_P (x))
4409 int regno = REGNO (x);
4410 HARD_REG_SET alloc_regs;
4412 if (hard_reg_p)
4414 if (regno >= FIRST_PSEUDO_REGISTER)
4415 regno = lra_get_regno_hard_regno (regno);
4416 if (regno < 0)
4417 return false;
4418 COMPL_HARD_REG_SET (alloc_regs, lra_no_alloc_regs);
4419 return overlaps_hard_reg_set_p (alloc_regs, GET_MODE (x), regno);
4421 else
4423 if (regno < FIRST_PSEUDO_REGISTER)
4424 return false;
4425 if (! spilled_p)
4426 return true;
4427 return lra_get_regno_hard_regno (regno) < 0;
4430 fmt = GET_RTX_FORMAT (code);
4431 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4433 if (fmt[i] == 'e')
4435 if (contains_reg_p (XEXP (x, i), hard_reg_p, spilled_p))
4436 return true;
4438 else if (fmt[i] == 'E')
4440 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4441 if (contains_reg_p (XVECEXP (x, i, j), hard_reg_p, spilled_p))
4442 return true;
4445 return false;
4448 /* Process all regs in location *LOC and change them on equivalent
4449 substitution. Return true if any change was done. */
4450 static bool
4451 loc_equivalence_change_p (rtx *loc)
4453 rtx subst, reg, x = *loc;
4454 bool result = false;
4455 enum rtx_code code = GET_CODE (x);
4456 const char *fmt;
4457 int i, j;
4459 if (code == SUBREG)
4461 reg = SUBREG_REG (x);
4462 if ((subst = get_equiv_with_elimination (reg, curr_insn)) != reg
4463 && GET_MODE (subst) == VOIDmode)
4465 /* We cannot reload debug location. Simplify subreg here
4466 while we know the inner mode. */
4467 *loc = simplify_gen_subreg (GET_MODE (x), subst,
4468 GET_MODE (reg), SUBREG_BYTE (x));
4469 return true;
4472 if (code == REG && (subst = get_equiv_with_elimination (x, curr_insn)) != x)
4474 *loc = subst;
4475 return true;
4478 /* Scan all the operand sub-expressions. */
4479 fmt = GET_RTX_FORMAT (code);
4480 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4482 if (fmt[i] == 'e')
4483 result = loc_equivalence_change_p (&XEXP (x, i)) || result;
4484 else if (fmt[i] == 'E')
4485 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4486 result
4487 = loc_equivalence_change_p (&XVECEXP (x, i, j)) || result;
4489 return result;
4492 /* Similar to loc_equivalence_change_p, but for use as
4493 simplify_replace_fn_rtx callback. DATA is insn for which the
4494 elimination is done. If it null we don't do the elimination. */
4495 static rtx
4496 loc_equivalence_callback (rtx loc, const_rtx, void *data)
4498 if (!REG_P (loc))
4499 return NULL_RTX;
4501 rtx subst = (data == NULL
4502 ? get_equiv (loc) : get_equiv_with_elimination (loc, (rtx_insn *) data));
4503 if (subst != loc)
4504 return subst;
4506 return NULL_RTX;
4509 /* Maximum number of generated reload insns per an insn. It is for
4510 preventing this pass cycling in a bug case. */
4511 #define MAX_RELOAD_INSNS_NUMBER LRA_MAX_INSN_RELOADS
4513 /* The current iteration number of this LRA pass. */
4514 int lra_constraint_iter;
4516 /* True if we substituted equiv which needs checking register
4517 allocation correctness because the equivalent value contains
4518 allocatable hard registers or when we restore multi-register
4519 pseudo. */
4520 bool lra_risky_transformations_p;
4522 /* Return true if REGNO is referenced in more than one block. */
4523 static bool
4524 multi_block_pseudo_p (int regno)
4526 basic_block bb = NULL;
4527 unsigned int uid;
4528 bitmap_iterator bi;
4530 if (regno < FIRST_PSEUDO_REGISTER)
4531 return false;
4533 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi)
4534 if (bb == NULL)
4535 bb = BLOCK_FOR_INSN (lra_insn_recog_data[uid]->insn);
4536 else if (BLOCK_FOR_INSN (lra_insn_recog_data[uid]->insn) != bb)
4537 return true;
4538 return false;
4541 /* Return true if LIST contains a deleted insn. */
4542 static bool
4543 contains_deleted_insn_p (rtx_insn_list *list)
4545 for (; list != NULL_RTX; list = list->next ())
4546 if (NOTE_P (list->insn ())
4547 && NOTE_KIND (list->insn ()) == NOTE_INSN_DELETED)
4548 return true;
4549 return false;
4552 /* Return true if X contains a pseudo dying in INSN. */
4553 static bool
4554 dead_pseudo_p (rtx x, rtx_insn *insn)
4556 int i, j;
4557 const char *fmt;
4558 enum rtx_code code;
4560 if (REG_P (x))
4561 return (insn != NULL_RTX
4562 && find_regno_note (insn, REG_DEAD, REGNO (x)) != NULL_RTX);
4563 code = GET_CODE (x);
4564 fmt = GET_RTX_FORMAT (code);
4565 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4567 if (fmt[i] == 'e')
4569 if (dead_pseudo_p (XEXP (x, i), insn))
4570 return true;
4572 else if (fmt[i] == 'E')
4574 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4575 if (dead_pseudo_p (XVECEXP (x, i, j), insn))
4576 return true;
4579 return false;
4582 /* Return true if INSN contains a dying pseudo in INSN right hand
4583 side. */
4584 static bool
4585 insn_rhs_dead_pseudo_p (rtx_insn *insn)
4587 rtx set = single_set (insn);
4589 gcc_assert (set != NULL);
4590 return dead_pseudo_p (SET_SRC (set), insn);
4593 /* Return true if any init insn of REGNO contains a dying pseudo in
4594 insn right hand side. */
4595 static bool
4596 init_insn_rhs_dead_pseudo_p (int regno)
4598 rtx_insn_list *insns = ira_reg_equiv[regno].init_insns;
4600 if (insns == NULL)
4601 return false;
4602 for (; insns != NULL_RTX; insns = insns->next ())
4603 if (insn_rhs_dead_pseudo_p (insns->insn ()))
4604 return true;
4605 return false;
4608 /* Return TRUE if REGNO has a reverse equivalence. The equivalence is
4609 reverse only if we have one init insn with given REGNO as a
4610 source. */
4611 static bool
4612 reverse_equiv_p (int regno)
4614 rtx_insn_list *insns = ira_reg_equiv[regno].init_insns;
4615 rtx set;
4617 if (insns == NULL)
4618 return false;
4619 if (! INSN_P (insns->insn ())
4620 || insns->next () != NULL)
4621 return false;
4622 if ((set = single_set (insns->insn ())) == NULL_RTX)
4623 return false;
4624 return REG_P (SET_SRC (set)) && (int) REGNO (SET_SRC (set)) == regno;
4627 /* Return TRUE if REGNO was reloaded in an equivalence init insn. We
4628 call this function only for non-reverse equivalence. */
4629 static bool
4630 contains_reloaded_insn_p (int regno)
4632 rtx set;
4633 rtx_insn_list *list = ira_reg_equiv[regno].init_insns;
4635 for (; list != NULL; list = list->next ())
4636 if ((set = single_set (list->insn ())) == NULL_RTX
4637 || ! REG_P (SET_DEST (set))
4638 || (int) REGNO (SET_DEST (set)) != regno)
4639 return true;
4640 return false;
4643 /* Entry function of LRA constraint pass. Return true if the
4644 constraint pass did change the code. */
4645 bool
4646 lra_constraints (bool first_p)
4648 bool changed_p;
4649 int i, hard_regno, new_insns_num;
4650 unsigned int min_len, new_min_len, uid;
4651 rtx set, x, reg, dest_reg;
4652 basic_block last_bb;
4653 bitmap_iterator bi;
4655 lra_constraint_iter++;
4656 if (lra_dump_file != NULL)
4657 fprintf (lra_dump_file, "\n********** Local #%d: **********\n\n",
4658 lra_constraint_iter);
4659 changed_p = false;
4660 if (pic_offset_table_rtx
4661 && REGNO (pic_offset_table_rtx) >= FIRST_PSEUDO_REGISTER)
4662 lra_risky_transformations_p = true;
4663 else
4664 /* On the first iteration we should check IRA assignment
4665 correctness. In rare cases, the assignments can be wrong as
4666 early clobbers operands are ignored in IRA. */
4667 lra_risky_transformations_p = first_p;
4668 new_insn_uid_start = get_max_uid ();
4669 new_regno_start = first_p ? lra_constraint_new_regno_start : max_reg_num ();
4670 /* Mark used hard regs for target stack size calulations. */
4671 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4672 if (lra_reg_info[i].nrefs != 0
4673 && (hard_regno = lra_get_regno_hard_regno (i)) >= 0)
4675 int j, nregs;
4677 nregs = hard_regno_nregs (hard_regno, lra_reg_info[i].biggest_mode);
4678 for (j = 0; j < nregs; j++)
4679 df_set_regs_ever_live (hard_regno + j, true);
4681 /* Do elimination before the equivalence processing as we can spill
4682 some pseudos during elimination. */
4683 lra_eliminate (false, first_p);
4684 auto_bitmap equiv_insn_bitmap (&reg_obstack);
4685 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4686 if (lra_reg_info[i].nrefs != 0)
4688 ira_reg_equiv[i].profitable_p = true;
4689 reg = regno_reg_rtx[i];
4690 if (lra_get_regno_hard_regno (i) < 0 && (x = get_equiv (reg)) != reg)
4692 bool pseudo_p = contains_reg_p (x, false, false);
4694 /* After RTL transformation, we can not guarantee that
4695 pseudo in the substitution was not reloaded which might
4696 make equivalence invalid. For example, in reverse
4697 equiv of p0
4699 p0 <- ...
4701 equiv_mem <- p0
4703 the memory address register was reloaded before the 2nd
4704 insn. */
4705 if ((! first_p && pseudo_p)
4706 /* We don't use DF for compilation speed sake. So it
4707 is problematic to update live info when we use an
4708 equivalence containing pseudos in more than one
4709 BB. */
4710 || (pseudo_p && multi_block_pseudo_p (i))
4711 /* If an init insn was deleted for some reason, cancel
4712 the equiv. We could update the equiv insns after
4713 transformations including an equiv insn deletion
4714 but it is not worthy as such cases are extremely
4715 rare. */
4716 || contains_deleted_insn_p (ira_reg_equiv[i].init_insns)
4717 /* If it is not a reverse equivalence, we check that a
4718 pseudo in rhs of the init insn is not dying in the
4719 insn. Otherwise, the live info at the beginning of
4720 the corresponding BB might be wrong after we
4721 removed the insn. When the equiv can be a
4722 constant, the right hand side of the init insn can
4723 be a pseudo. */
4724 || (! reverse_equiv_p (i)
4725 && (init_insn_rhs_dead_pseudo_p (i)
4726 /* If we reloaded the pseudo in an equivalence
4727 init insn, we can not remove the equiv init
4728 insns and the init insns might write into
4729 const memory in this case. */
4730 || contains_reloaded_insn_p (i)))
4731 /* Prevent access beyond equivalent memory for
4732 paradoxical subregs. */
4733 || (MEM_P (x)
4734 && (GET_MODE_SIZE (lra_reg_info[i].biggest_mode)
4735 > GET_MODE_SIZE (GET_MODE (x))))
4736 || (pic_offset_table_rtx
4737 && ((CONST_POOL_OK_P (PSEUDO_REGNO_MODE (i), x)
4738 && (targetm.preferred_reload_class
4739 (x, lra_get_allocno_class (i)) == NO_REGS))
4740 || contains_symbol_ref_p (x))))
4741 ira_reg_equiv[i].defined_p = false;
4742 if (contains_reg_p (x, false, true))
4743 ira_reg_equiv[i].profitable_p = false;
4744 if (get_equiv (reg) != reg)
4745 bitmap_ior_into (equiv_insn_bitmap, &lra_reg_info[i].insn_bitmap);
4748 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4749 update_equiv (i);
4750 /* We should add all insns containing pseudos which should be
4751 substituted by their equivalences. */
4752 EXECUTE_IF_SET_IN_BITMAP (equiv_insn_bitmap, 0, uid, bi)
4753 lra_push_insn_by_uid (uid);
4754 min_len = lra_insn_stack_length ();
4755 new_insns_num = 0;
4756 last_bb = NULL;
4757 changed_p = false;
4758 while ((new_min_len = lra_insn_stack_length ()) != 0)
4760 curr_insn = lra_pop_insn ();
4761 --new_min_len;
4762 curr_bb = BLOCK_FOR_INSN (curr_insn);
4763 if (curr_bb != last_bb)
4765 last_bb = curr_bb;
4766 bb_reload_num = lra_curr_reload_num;
4768 if (min_len > new_min_len)
4770 min_len = new_min_len;
4771 new_insns_num = 0;
4773 if (new_insns_num > MAX_RELOAD_INSNS_NUMBER)
4774 internal_error
4775 ("Max. number of generated reload insns per insn is achieved (%d)\n",
4776 MAX_RELOAD_INSNS_NUMBER);
4777 new_insns_num++;
4778 if (DEBUG_INSN_P (curr_insn))
4780 /* We need to check equivalence in debug insn and change
4781 pseudo to the equivalent value if necessary. */
4782 curr_id = lra_get_insn_recog_data (curr_insn);
4783 if (bitmap_bit_p (equiv_insn_bitmap, INSN_UID (curr_insn)))
4785 rtx old = *curr_id->operand_loc[0];
4786 *curr_id->operand_loc[0]
4787 = simplify_replace_fn_rtx (old, NULL_RTX,
4788 loc_equivalence_callback, curr_insn);
4789 if (old != *curr_id->operand_loc[0])
4791 lra_update_insn_regno_info (curr_insn);
4792 changed_p = true;
4796 else if (INSN_P (curr_insn))
4798 if ((set = single_set (curr_insn)) != NULL_RTX)
4800 dest_reg = SET_DEST (set);
4801 /* The equivalence pseudo could be set up as SUBREG in a
4802 case when it is a call restore insn in a mode
4803 different from the pseudo mode. */
4804 if (GET_CODE (dest_reg) == SUBREG)
4805 dest_reg = SUBREG_REG (dest_reg);
4806 if ((REG_P (dest_reg)
4807 && (x = get_equiv (dest_reg)) != dest_reg
4808 /* Remove insns which set up a pseudo whose value
4809 can not be changed. Such insns might be not in
4810 init_insns because we don't update equiv data
4811 during insn transformations.
4813 As an example, let suppose that a pseudo got
4814 hard register and on the 1st pass was not
4815 changed to equivalent constant. We generate an
4816 additional insn setting up the pseudo because of
4817 secondary memory movement. Then the pseudo is
4818 spilled and we use the equiv constant. In this
4819 case we should remove the additional insn and
4820 this insn is not init_insns list. */
4821 && (! MEM_P (x) || MEM_READONLY_P (x)
4822 /* Check that this is actually an insn setting
4823 up the equivalence. */
4824 || in_list_p (curr_insn,
4825 ira_reg_equiv
4826 [REGNO (dest_reg)].init_insns)))
4827 || (((x = get_equiv (SET_SRC (set))) != SET_SRC (set))
4828 && in_list_p (curr_insn,
4829 ira_reg_equiv
4830 [REGNO (SET_SRC (set))].init_insns)))
4832 /* This is equiv init insn of pseudo which did not get a
4833 hard register -- remove the insn. */
4834 if (lra_dump_file != NULL)
4836 fprintf (lra_dump_file,
4837 " Removing equiv init insn %i (freq=%d)\n",
4838 INSN_UID (curr_insn),
4839 REG_FREQ_FROM_BB (BLOCK_FOR_INSN (curr_insn)));
4840 dump_insn_slim (lra_dump_file, curr_insn);
4842 if (contains_reg_p (x, true, false))
4843 lra_risky_transformations_p = true;
4844 lra_set_insn_deleted (curr_insn);
4845 continue;
4848 curr_id = lra_get_insn_recog_data (curr_insn);
4849 curr_static_id = curr_id->insn_static_data;
4850 init_curr_insn_input_reloads ();
4851 init_curr_operand_mode ();
4852 if (curr_insn_transform (false))
4853 changed_p = true;
4854 /* Check non-transformed insns too for equiv change as USE
4855 or CLOBBER don't need reloads but can contain pseudos
4856 being changed on their equivalences. */
4857 else if (bitmap_bit_p (equiv_insn_bitmap, INSN_UID (curr_insn))
4858 && loc_equivalence_change_p (&PATTERN (curr_insn)))
4860 lra_update_insn_regno_info (curr_insn);
4861 changed_p = true;
4866 /* If we used a new hard regno, changed_p should be true because the
4867 hard reg is assigned to a new pseudo. */
4868 if (flag_checking && !changed_p)
4870 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4871 if (lra_reg_info[i].nrefs != 0
4872 && (hard_regno = lra_get_regno_hard_regno (i)) >= 0)
4874 int j, nregs = hard_regno_nregs (hard_regno,
4875 PSEUDO_REGNO_MODE (i));
4877 for (j = 0; j < nregs; j++)
4878 lra_assert (df_regs_ever_live_p (hard_regno + j));
4881 return changed_p;
4884 static void initiate_invariants (void);
4885 static void finish_invariants (void);
4887 /* Initiate the LRA constraint pass. It is done once per
4888 function. */
4889 void
4890 lra_constraints_init (void)
4892 initiate_invariants ();
4895 /* Finalize the LRA constraint pass. It is done once per
4896 function. */
4897 void
4898 lra_constraints_finish (void)
4900 finish_invariants ();
4905 /* Structure describes invariants for ineheritance. */
4906 struct lra_invariant
4908 /* The order number of the invariant. */
4909 int num;
4910 /* The invariant RTX. */
4911 rtx invariant_rtx;
4912 /* The origin insn of the invariant. */
4913 rtx_insn *insn;
4916 typedef lra_invariant invariant_t;
4917 typedef invariant_t *invariant_ptr_t;
4918 typedef const invariant_t *const_invariant_ptr_t;
4920 /* Pointer to the inheritance invariants. */
4921 static vec<invariant_ptr_t> invariants;
4923 /* Allocation pool for the invariants. */
4924 static object_allocator<lra_invariant> *invariants_pool;
4926 /* Hash table for the invariants. */
4927 static htab_t invariant_table;
4929 /* Hash function for INVARIANT. */
4930 static hashval_t
4931 invariant_hash (const void *invariant)
4933 rtx inv = ((const_invariant_ptr_t) invariant)->invariant_rtx;
4934 return lra_rtx_hash (inv);
4937 /* Equal function for invariants INVARIANT1 and INVARIANT2. */
4938 static int
4939 invariant_eq_p (const void *invariant1, const void *invariant2)
4941 rtx inv1 = ((const_invariant_ptr_t) invariant1)->invariant_rtx;
4942 rtx inv2 = ((const_invariant_ptr_t) invariant2)->invariant_rtx;
4944 return rtx_equal_p (inv1, inv2);
4947 /* Insert INVARIANT_RTX into the table if it is not there yet. Return
4948 invariant which is in the table. */
4949 static invariant_ptr_t
4950 insert_invariant (rtx invariant_rtx)
4952 void **entry_ptr;
4953 invariant_t invariant;
4954 invariant_ptr_t invariant_ptr;
4956 invariant.invariant_rtx = invariant_rtx;
4957 entry_ptr = htab_find_slot (invariant_table, &invariant, INSERT);
4958 if (*entry_ptr == NULL)
4960 invariant_ptr = invariants_pool->allocate ();
4961 invariant_ptr->invariant_rtx = invariant_rtx;
4962 invariant_ptr->insn = NULL;
4963 invariants.safe_push (invariant_ptr);
4964 *entry_ptr = (void *) invariant_ptr;
4966 return (invariant_ptr_t) *entry_ptr;
4969 /* Initiate the invariant table. */
4970 static void
4971 initiate_invariants (void)
4973 invariants.create (100);
4974 invariants_pool
4975 = new object_allocator<lra_invariant> ("Inheritance invariants");
4976 invariant_table = htab_create (100, invariant_hash, invariant_eq_p, NULL);
4979 /* Finish the invariant table. */
4980 static void
4981 finish_invariants (void)
4983 htab_delete (invariant_table);
4984 delete invariants_pool;
4985 invariants.release ();
4988 /* Make the invariant table empty. */
4989 static void
4990 clear_invariants (void)
4992 htab_empty (invariant_table);
4993 invariants_pool->release ();
4994 invariants.truncate (0);
4999 /* This page contains code to do inheritance/split
5000 transformations. */
5002 /* Number of reloads passed so far in current EBB. */
5003 static int reloads_num;
5005 /* Number of calls passed so far in current EBB. */
5006 static int calls_num;
5008 /* Current reload pseudo check for validity of elements in
5009 USAGE_INSNS. */
5010 static int curr_usage_insns_check;
5012 /* Info about last usage of registers in EBB to do inheritance/split
5013 transformation. Inheritance transformation is done from a spilled
5014 pseudo and split transformations from a hard register or a pseudo
5015 assigned to a hard register. */
5016 struct usage_insns
5018 /* If the value is equal to CURR_USAGE_INSNS_CHECK, then the member
5019 value INSNS is valid. The insns is chain of optional debug insns
5020 and a finishing non-debug insn using the corresponding reg. The
5021 value is also used to mark the registers which are set up in the
5022 current insn. The negated insn uid is used for this. */
5023 int check;
5024 /* Value of global reloads_num at the last insn in INSNS. */
5025 int reloads_num;
5026 /* Value of global reloads_nums at the last insn in INSNS. */
5027 int calls_num;
5028 /* It can be true only for splitting. And it means that the restore
5029 insn should be put after insn given by the following member. */
5030 bool after_p;
5031 /* Next insns in the current EBB which use the original reg and the
5032 original reg value is not changed between the current insn and
5033 the next insns. In order words, e.g. for inheritance, if we need
5034 to use the original reg value again in the next insns we can try
5035 to use the value in a hard register from a reload insn of the
5036 current insn. */
5037 rtx insns;
5040 /* Map: regno -> corresponding pseudo usage insns. */
5041 static struct usage_insns *usage_insns;
5043 static void
5044 setup_next_usage_insn (int regno, rtx insn, int reloads_num, bool after_p)
5046 usage_insns[regno].check = curr_usage_insns_check;
5047 usage_insns[regno].insns = insn;
5048 usage_insns[regno].reloads_num = reloads_num;
5049 usage_insns[regno].calls_num = calls_num;
5050 usage_insns[regno].after_p = after_p;
5053 /* The function is used to form list REGNO usages which consists of
5054 optional debug insns finished by a non-debug insn using REGNO.
5055 RELOADS_NUM is current number of reload insns processed so far. */
5056 static void
5057 add_next_usage_insn (int regno, rtx_insn *insn, int reloads_num)
5059 rtx next_usage_insns;
5061 if (usage_insns[regno].check == curr_usage_insns_check
5062 && (next_usage_insns = usage_insns[regno].insns) != NULL_RTX
5063 && DEBUG_INSN_P (insn))
5065 /* Check that we did not add the debug insn yet. */
5066 if (next_usage_insns != insn
5067 && (GET_CODE (next_usage_insns) != INSN_LIST
5068 || XEXP (next_usage_insns, 0) != insn))
5069 usage_insns[regno].insns = gen_rtx_INSN_LIST (VOIDmode, insn,
5070 next_usage_insns);
5072 else if (NONDEBUG_INSN_P (insn))
5073 setup_next_usage_insn (regno, insn, reloads_num, false);
5074 else
5075 usage_insns[regno].check = 0;
5078 /* Return first non-debug insn in list USAGE_INSNS. */
5079 static rtx_insn *
5080 skip_usage_debug_insns (rtx usage_insns)
5082 rtx insn;
5084 /* Skip debug insns. */
5085 for (insn = usage_insns;
5086 insn != NULL_RTX && GET_CODE (insn) == INSN_LIST;
5087 insn = XEXP (insn, 1))
5089 return safe_as_a <rtx_insn *> (insn);
5092 /* Return true if we need secondary memory moves for insn in
5093 USAGE_INSNS after inserting inherited pseudo of class INHER_CL
5094 into the insn. */
5095 static bool
5096 check_secondary_memory_needed_p (enum reg_class inher_cl ATTRIBUTE_UNUSED,
5097 rtx usage_insns ATTRIBUTE_UNUSED)
5099 rtx_insn *insn;
5100 rtx set, dest;
5101 enum reg_class cl;
5103 if (inher_cl == ALL_REGS
5104 || (insn = skip_usage_debug_insns (usage_insns)) == NULL_RTX)
5105 return false;
5106 lra_assert (INSN_P (insn));
5107 if ((set = single_set (insn)) == NULL_RTX || ! REG_P (SET_DEST (set)))
5108 return false;
5109 dest = SET_DEST (set);
5110 if (! REG_P (dest))
5111 return false;
5112 lra_assert (inher_cl != NO_REGS);
5113 cl = get_reg_class (REGNO (dest));
5114 return (cl != NO_REGS && cl != ALL_REGS
5115 && targetm.secondary_memory_needed (GET_MODE (dest), inher_cl, cl));
5118 /* Registers involved in inheritance/split in the current EBB
5119 (inheritance/split pseudos and original registers). */
5120 static bitmap_head check_only_regs;
5122 /* Reload pseudos can not be involded in invariant inheritance in the
5123 current EBB. */
5124 static bitmap_head invalid_invariant_regs;
5126 /* Do inheritance transformations for insn INSN, which defines (if
5127 DEF_P) or uses ORIGINAL_REGNO. NEXT_USAGE_INSNS specifies which
5128 instruction in the EBB next uses ORIGINAL_REGNO; it has the same
5129 form as the "insns" field of usage_insns. Return true if we
5130 succeed in such transformation.
5132 The transformations look like:
5134 p <- ... i <- ...
5135 ... p <- i (new insn)
5136 ... =>
5137 <- ... p ... <- ... i ...
5139 ... i <- p (new insn)
5140 <- ... p ... <- ... i ...
5141 ... =>
5142 <- ... p ... <- ... i ...
5143 where p is a spilled original pseudo and i is a new inheritance pseudo.
5146 The inheritance pseudo has the smallest class of two classes CL and
5147 class of ORIGINAL REGNO. */
5148 static bool
5149 inherit_reload_reg (bool def_p, int original_regno,
5150 enum reg_class cl, rtx_insn *insn, rtx next_usage_insns)
5152 if (optimize_function_for_size_p (cfun))
5153 return false;
5155 enum reg_class rclass = lra_get_allocno_class (original_regno);
5156 rtx original_reg = regno_reg_rtx[original_regno];
5157 rtx new_reg, usage_insn;
5158 rtx_insn *new_insns;
5160 lra_assert (! usage_insns[original_regno].after_p);
5161 if (lra_dump_file != NULL)
5162 fprintf (lra_dump_file,
5163 " <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<\n");
5164 if (! ira_reg_classes_intersect_p[cl][rclass])
5166 if (lra_dump_file != NULL)
5168 fprintf (lra_dump_file,
5169 " Rejecting inheritance for %d "
5170 "because of disjoint classes %s and %s\n",
5171 original_regno, reg_class_names[cl],
5172 reg_class_names[rclass]);
5173 fprintf (lra_dump_file,
5174 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5176 return false;
5178 if ((ira_class_subset_p[cl][rclass] && cl != rclass)
5179 /* We don't use a subset of two classes because it can be
5180 NO_REGS. This transformation is still profitable in most
5181 cases even if the classes are not intersected as register
5182 move is probably cheaper than a memory load. */
5183 || ira_class_hard_regs_num[cl] < ira_class_hard_regs_num[rclass])
5185 if (lra_dump_file != NULL)
5186 fprintf (lra_dump_file, " Use smallest class of %s and %s\n",
5187 reg_class_names[cl], reg_class_names[rclass]);
5189 rclass = cl;
5191 if (check_secondary_memory_needed_p (rclass, next_usage_insns))
5193 /* Reject inheritance resulting in secondary memory moves.
5194 Otherwise, there is a danger in LRA cycling. Also such
5195 transformation will be unprofitable. */
5196 if (lra_dump_file != NULL)
5198 rtx_insn *insn = skip_usage_debug_insns (next_usage_insns);
5199 rtx set = single_set (insn);
5201 lra_assert (set != NULL_RTX);
5203 rtx dest = SET_DEST (set);
5205 lra_assert (REG_P (dest));
5206 fprintf (lra_dump_file,
5207 " Rejecting inheritance for insn %d(%s)<-%d(%s) "
5208 "as secondary mem is needed\n",
5209 REGNO (dest), reg_class_names[get_reg_class (REGNO (dest))],
5210 original_regno, reg_class_names[rclass]);
5211 fprintf (lra_dump_file,
5212 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5214 return false;
5216 new_reg = lra_create_new_reg (GET_MODE (original_reg), original_reg,
5217 rclass, "inheritance");
5218 start_sequence ();
5219 if (def_p)
5220 lra_emit_move (original_reg, new_reg);
5221 else
5222 lra_emit_move (new_reg, original_reg);
5223 new_insns = get_insns ();
5224 end_sequence ();
5225 if (NEXT_INSN (new_insns) != NULL_RTX)
5227 if (lra_dump_file != NULL)
5229 fprintf (lra_dump_file,
5230 " Rejecting inheritance %d->%d "
5231 "as it results in 2 or more insns:\n",
5232 original_regno, REGNO (new_reg));
5233 dump_rtl_slim (lra_dump_file, new_insns, NULL, -1, 0);
5234 fprintf (lra_dump_file,
5235 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5237 return false;
5239 lra_substitute_pseudo_within_insn (insn, original_regno, new_reg, false);
5240 lra_update_insn_regno_info (insn);
5241 if (! def_p)
5242 /* We now have a new usage insn for original regno. */
5243 setup_next_usage_insn (original_regno, new_insns, reloads_num, false);
5244 if (lra_dump_file != NULL)
5245 fprintf (lra_dump_file, " Original reg change %d->%d (bb%d):\n",
5246 original_regno, REGNO (new_reg), BLOCK_FOR_INSN (insn)->index);
5247 lra_reg_info[REGNO (new_reg)].restore_rtx = regno_reg_rtx[original_regno];
5248 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
5249 bitmap_set_bit (&check_only_regs, original_regno);
5250 bitmap_set_bit (&lra_inheritance_pseudos, REGNO (new_reg));
5251 if (def_p)
5252 lra_process_new_insns (insn, NULL, new_insns,
5253 "Add original<-inheritance");
5254 else
5255 lra_process_new_insns (insn, new_insns, NULL,
5256 "Add inheritance<-original");
5257 while (next_usage_insns != NULL_RTX)
5259 if (GET_CODE (next_usage_insns) != INSN_LIST)
5261 usage_insn = next_usage_insns;
5262 lra_assert (NONDEBUG_INSN_P (usage_insn));
5263 next_usage_insns = NULL;
5265 else
5267 usage_insn = XEXP (next_usage_insns, 0);
5268 lra_assert (DEBUG_INSN_P (usage_insn));
5269 next_usage_insns = XEXP (next_usage_insns, 1);
5271 lra_substitute_pseudo (&usage_insn, original_regno, new_reg, false);
5272 lra_update_insn_regno_info (as_a <rtx_insn *> (usage_insn));
5273 if (lra_dump_file != NULL)
5275 basic_block bb = BLOCK_FOR_INSN (usage_insn);
5276 fprintf (lra_dump_file,
5277 " Inheritance reuse change %d->%d (bb%d):\n",
5278 original_regno, REGNO (new_reg),
5279 bb ? bb->index : -1);
5280 dump_insn_slim (lra_dump_file, as_a <rtx_insn *> (usage_insn));
5283 if (lra_dump_file != NULL)
5284 fprintf (lra_dump_file,
5285 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5286 return true;
5289 /* Return true if we need a caller save/restore for pseudo REGNO which
5290 was assigned to a hard register. */
5291 static inline bool
5292 need_for_call_save_p (int regno)
5294 lra_assert (regno >= FIRST_PSEUDO_REGISTER && reg_renumber[regno] >= 0);
5295 return (usage_insns[regno].calls_num < calls_num
5296 && (overlaps_hard_reg_set_p
5297 ((flag_ipa_ra &&
5298 ! hard_reg_set_empty_p (lra_reg_info[regno].actual_call_used_reg_set))
5299 ? lra_reg_info[regno].actual_call_used_reg_set
5300 : call_used_reg_set,
5301 PSEUDO_REGNO_MODE (regno), reg_renumber[regno])
5302 || (targetm.hard_regno_call_part_clobbered
5303 (reg_renumber[regno], PSEUDO_REGNO_MODE (regno)))));
5306 /* Global registers occurring in the current EBB. */
5307 static bitmap_head ebb_global_regs;
5309 /* Return true if we need a split for hard register REGNO or pseudo
5310 REGNO which was assigned to a hard register.
5311 POTENTIAL_RELOAD_HARD_REGS contains hard registers which might be
5312 used for reloads since the EBB end. It is an approximation of the
5313 used hard registers in the split range. The exact value would
5314 require expensive calculations. If we were aggressive with
5315 splitting because of the approximation, the split pseudo will save
5316 the same hard register assignment and will be removed in the undo
5317 pass. We still need the approximation because too aggressive
5318 splitting would result in too inaccurate cost calculation in the
5319 assignment pass because of too many generated moves which will be
5320 probably removed in the undo pass. */
5321 static inline bool
5322 need_for_split_p (HARD_REG_SET potential_reload_hard_regs, int regno)
5324 int hard_regno = regno < FIRST_PSEUDO_REGISTER ? regno : reg_renumber[regno];
5326 lra_assert (hard_regno >= 0);
5327 return ((TEST_HARD_REG_BIT (potential_reload_hard_regs, hard_regno)
5328 /* Don't split eliminable hard registers, otherwise we can
5329 split hard registers like hard frame pointer, which
5330 lives on BB start/end according to DF-infrastructure,
5331 when there is a pseudo assigned to the register and
5332 living in the same BB. */
5333 && (regno >= FIRST_PSEUDO_REGISTER
5334 || ! TEST_HARD_REG_BIT (eliminable_regset, hard_regno))
5335 && ! TEST_HARD_REG_BIT (lra_no_alloc_regs, hard_regno)
5336 /* Don't split call clobbered hard regs living through
5337 calls, otherwise we might have a check problem in the
5338 assign sub-pass as in the most cases (exception is a
5339 situation when lra_risky_transformations_p value is
5340 true) the assign pass assumes that all pseudos living
5341 through calls are assigned to call saved hard regs. */
5342 && (regno >= FIRST_PSEUDO_REGISTER
5343 || ! TEST_HARD_REG_BIT (call_used_reg_set, regno)
5344 || usage_insns[regno].calls_num == calls_num)
5345 /* We need at least 2 reloads to make pseudo splitting
5346 profitable. We should provide hard regno splitting in
5347 any case to solve 1st insn scheduling problem when
5348 moving hard register definition up might result in
5349 impossibility to find hard register for reload pseudo of
5350 small register class. */
5351 && (usage_insns[regno].reloads_num
5352 + (regno < FIRST_PSEUDO_REGISTER ? 0 : 3) < reloads_num)
5353 && (regno < FIRST_PSEUDO_REGISTER
5354 /* For short living pseudos, spilling + inheritance can
5355 be considered a substitution for splitting.
5356 Therefore we do not splitting for local pseudos. It
5357 decreases also aggressiveness of splitting. The
5358 minimal number of references is chosen taking into
5359 account that for 2 references splitting has no sense
5360 as we can just spill the pseudo. */
5361 || (regno >= FIRST_PSEUDO_REGISTER
5362 && lra_reg_info[regno].nrefs > 3
5363 && bitmap_bit_p (&ebb_global_regs, regno))))
5364 || (regno >= FIRST_PSEUDO_REGISTER && need_for_call_save_p (regno)));
5367 /* Return class for the split pseudo created from original pseudo with
5368 ALLOCNO_CLASS and MODE which got a hard register HARD_REGNO. We
5369 choose subclass of ALLOCNO_CLASS which contains HARD_REGNO and
5370 results in no secondary memory movements. */
5371 static enum reg_class
5372 choose_split_class (enum reg_class allocno_class,
5373 int hard_regno ATTRIBUTE_UNUSED,
5374 machine_mode mode ATTRIBUTE_UNUSED)
5376 int i;
5377 enum reg_class cl, best_cl = NO_REGS;
5378 enum reg_class hard_reg_class ATTRIBUTE_UNUSED
5379 = REGNO_REG_CLASS (hard_regno);
5381 if (! targetm.secondary_memory_needed (mode, allocno_class, allocno_class)
5382 && TEST_HARD_REG_BIT (reg_class_contents[allocno_class], hard_regno))
5383 return allocno_class;
5384 for (i = 0;
5385 (cl = reg_class_subclasses[allocno_class][i]) != LIM_REG_CLASSES;
5386 i++)
5387 if (! targetm.secondary_memory_needed (mode, cl, hard_reg_class)
5388 && ! targetm.secondary_memory_needed (mode, hard_reg_class, cl)
5389 && TEST_HARD_REG_BIT (reg_class_contents[cl], hard_regno)
5390 && (best_cl == NO_REGS
5391 || ira_class_hard_regs_num[best_cl] < ira_class_hard_regs_num[cl]))
5392 best_cl = cl;
5393 return best_cl;
5396 /* Copy any equivalence information from ORIGINAL_REGNO to NEW_REGNO.
5397 It only makes sense to call this function if NEW_REGNO is always
5398 equal to ORIGINAL_REGNO. */
5400 static void
5401 lra_copy_reg_equiv (unsigned int new_regno, unsigned int original_regno)
5403 if (!ira_reg_equiv[original_regno].defined_p)
5404 return;
5406 ira_expand_reg_equiv ();
5407 ira_reg_equiv[new_regno].defined_p = true;
5408 if (ira_reg_equiv[original_regno].memory)
5409 ira_reg_equiv[new_regno].memory
5410 = copy_rtx (ira_reg_equiv[original_regno].memory);
5411 if (ira_reg_equiv[original_regno].constant)
5412 ira_reg_equiv[new_regno].constant
5413 = copy_rtx (ira_reg_equiv[original_regno].constant);
5414 if (ira_reg_equiv[original_regno].invariant)
5415 ira_reg_equiv[new_regno].invariant
5416 = copy_rtx (ira_reg_equiv[original_regno].invariant);
5419 /* Do split transformations for insn INSN, which defines or uses
5420 ORIGINAL_REGNO. NEXT_USAGE_INSNS specifies which instruction in
5421 the EBB next uses ORIGINAL_REGNO; it has the same form as the
5422 "insns" field of usage_insns.
5424 The transformations look like:
5426 p <- ... p <- ...
5427 ... s <- p (new insn -- save)
5428 ... =>
5429 ... p <- s (new insn -- restore)
5430 <- ... p ... <- ... p ...
5432 <- ... p ... <- ... p ...
5433 ... s <- p (new insn -- save)
5434 ... =>
5435 ... p <- s (new insn -- restore)
5436 <- ... p ... <- ... p ...
5438 where p is an original pseudo got a hard register or a hard
5439 register and s is a new split pseudo. The save is put before INSN
5440 if BEFORE_P is true. Return true if we succeed in such
5441 transformation. */
5442 static bool
5443 split_reg (bool before_p, int original_regno, rtx_insn *insn,
5444 rtx next_usage_insns)
5446 enum reg_class rclass;
5447 rtx original_reg;
5448 int hard_regno, nregs;
5449 rtx new_reg, usage_insn;
5450 rtx_insn *restore, *save;
5451 bool after_p;
5452 bool call_save_p;
5453 machine_mode mode;
5455 if (original_regno < FIRST_PSEUDO_REGISTER)
5457 rclass = ira_allocno_class_translate[REGNO_REG_CLASS (original_regno)];
5458 hard_regno = original_regno;
5459 call_save_p = false;
5460 nregs = 1;
5461 mode = lra_reg_info[hard_regno].biggest_mode;
5462 machine_mode reg_rtx_mode = GET_MODE (regno_reg_rtx[hard_regno]);
5463 /* A reg can have a biggest_mode of VOIDmode if it was only ever seen
5464 as part of a multi-word register. In that case, or if the biggest
5465 mode was larger than a register, just use the reg_rtx. Otherwise,
5466 limit the size to that of the biggest access in the function. */
5467 if (mode == VOIDmode
5468 || paradoxical_subreg_p (mode, reg_rtx_mode))
5470 original_reg = regno_reg_rtx[hard_regno];
5471 mode = reg_rtx_mode;
5473 else
5474 original_reg = gen_rtx_REG (mode, hard_regno);
5476 else
5478 mode = PSEUDO_REGNO_MODE (original_regno);
5479 hard_regno = reg_renumber[original_regno];
5480 nregs = hard_regno_nregs (hard_regno, mode);
5481 rclass = lra_get_allocno_class (original_regno);
5482 original_reg = regno_reg_rtx[original_regno];
5483 call_save_p = need_for_call_save_p (original_regno);
5485 lra_assert (hard_regno >= 0);
5486 if (lra_dump_file != NULL)
5487 fprintf (lra_dump_file,
5488 " ((((((((((((((((((((((((((((((((((((((((((((((((\n");
5490 if (call_save_p)
5492 mode = HARD_REGNO_CALLER_SAVE_MODE (hard_regno,
5493 hard_regno_nregs (hard_regno, mode),
5494 mode);
5495 new_reg = lra_create_new_reg (mode, NULL_RTX, NO_REGS, "save");
5497 else
5499 rclass = choose_split_class (rclass, hard_regno, mode);
5500 if (rclass == NO_REGS)
5502 if (lra_dump_file != NULL)
5504 fprintf (lra_dump_file,
5505 " Rejecting split of %d(%s): "
5506 "no good reg class for %d(%s)\n",
5507 original_regno,
5508 reg_class_names[lra_get_allocno_class (original_regno)],
5509 hard_regno,
5510 reg_class_names[REGNO_REG_CLASS (hard_regno)]);
5511 fprintf
5512 (lra_dump_file,
5513 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5515 return false;
5517 /* Split_if_necessary can split hard registers used as part of a
5518 multi-register mode but splits each register individually. The
5519 mode used for each independent register may not be supported
5520 so reject the split. Splitting the wider mode should theoretically
5521 be possible but is not implemented. */
5522 if (!targetm.hard_regno_mode_ok (hard_regno, mode))
5524 if (lra_dump_file != NULL)
5526 fprintf (lra_dump_file,
5527 " Rejecting split of %d(%s): unsuitable mode %s\n",
5528 original_regno,
5529 reg_class_names[lra_get_allocno_class (original_regno)],
5530 GET_MODE_NAME (mode));
5531 fprintf
5532 (lra_dump_file,
5533 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5535 return false;
5537 new_reg = lra_create_new_reg (mode, original_reg, rclass, "split");
5538 reg_renumber[REGNO (new_reg)] = hard_regno;
5540 int new_regno = REGNO (new_reg);
5541 save = emit_spill_move (true, new_reg, original_reg);
5542 if (NEXT_INSN (save) != NULL_RTX && !call_save_p)
5544 if (lra_dump_file != NULL)
5546 fprintf
5547 (lra_dump_file,
5548 " Rejecting split %d->%d resulting in > 2 save insns:\n",
5549 original_regno, new_regno);
5550 dump_rtl_slim (lra_dump_file, save, NULL, -1, 0);
5551 fprintf (lra_dump_file,
5552 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5554 return false;
5556 restore = emit_spill_move (false, new_reg, original_reg);
5557 if (NEXT_INSN (restore) != NULL_RTX && !call_save_p)
5559 if (lra_dump_file != NULL)
5561 fprintf (lra_dump_file,
5562 " Rejecting split %d->%d "
5563 "resulting in > 2 restore insns:\n",
5564 original_regno, new_regno);
5565 dump_rtl_slim (lra_dump_file, restore, NULL, -1, 0);
5566 fprintf (lra_dump_file,
5567 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5569 return false;
5571 /* Transfer equivalence information to the spill register, so that
5572 if we fail to allocate the spill register, we have the option of
5573 rematerializing the original value instead of spilling to the stack. */
5574 if (!HARD_REGISTER_NUM_P (original_regno)
5575 && mode == PSEUDO_REGNO_MODE (original_regno))
5576 lra_copy_reg_equiv (new_regno, original_regno);
5577 after_p = usage_insns[original_regno].after_p;
5578 lra_reg_info[new_regno].restore_rtx = regno_reg_rtx[original_regno];
5579 bitmap_set_bit (&check_only_regs, new_regno);
5580 bitmap_set_bit (&check_only_regs, original_regno);
5581 bitmap_set_bit (&lra_split_regs, new_regno);
5582 for (;;)
5584 if (GET_CODE (next_usage_insns) != INSN_LIST)
5586 usage_insn = next_usage_insns;
5587 break;
5589 usage_insn = XEXP (next_usage_insns, 0);
5590 lra_assert (DEBUG_INSN_P (usage_insn));
5591 next_usage_insns = XEXP (next_usage_insns, 1);
5592 lra_substitute_pseudo (&usage_insn, original_regno, new_reg, false);
5593 lra_update_insn_regno_info (as_a <rtx_insn *> (usage_insn));
5594 if (lra_dump_file != NULL)
5596 fprintf (lra_dump_file, " Split reuse change %d->%d:\n",
5597 original_regno, new_regno);
5598 dump_insn_slim (lra_dump_file, as_a <rtx_insn *> (usage_insn));
5601 lra_assert (NOTE_P (usage_insn) || NONDEBUG_INSN_P (usage_insn));
5602 lra_assert (usage_insn != insn || (after_p && before_p));
5603 lra_process_new_insns (as_a <rtx_insn *> (usage_insn),
5604 after_p ? NULL : restore,
5605 after_p ? restore : NULL,
5606 call_save_p
5607 ? "Add reg<-save" : "Add reg<-split");
5608 lra_process_new_insns (insn, before_p ? save : NULL,
5609 before_p ? NULL : save,
5610 call_save_p
5611 ? "Add save<-reg" : "Add split<-reg");
5612 if (nregs > 1)
5613 /* If we are trying to split multi-register. We should check
5614 conflicts on the next assignment sub-pass. IRA can allocate on
5615 sub-register levels, LRA do this on pseudos level right now and
5616 this discrepancy may create allocation conflicts after
5617 splitting. */
5618 lra_risky_transformations_p = true;
5619 if (lra_dump_file != NULL)
5620 fprintf (lra_dump_file,
5621 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5622 return true;
5625 /* Recognize that we need a split transformation for insn INSN, which
5626 defines or uses REGNO in its insn biggest MODE (we use it only if
5627 REGNO is a hard register). POTENTIAL_RELOAD_HARD_REGS contains
5628 hard registers which might be used for reloads since the EBB end.
5629 Put the save before INSN if BEFORE_P is true. MAX_UID is maximla
5630 uid before starting INSN processing. Return true if we succeed in
5631 such transformation. */
5632 static bool
5633 split_if_necessary (int regno, machine_mode mode,
5634 HARD_REG_SET potential_reload_hard_regs,
5635 bool before_p, rtx_insn *insn, int max_uid)
5637 bool res = false;
5638 int i, nregs = 1;
5639 rtx next_usage_insns;
5641 if (regno < FIRST_PSEUDO_REGISTER)
5642 nregs = hard_regno_nregs (regno, mode);
5643 for (i = 0; i < nregs; i++)
5644 if (usage_insns[regno + i].check == curr_usage_insns_check
5645 && (next_usage_insns = usage_insns[regno + i].insns) != NULL_RTX
5646 /* To avoid processing the register twice or more. */
5647 && ((GET_CODE (next_usage_insns) != INSN_LIST
5648 && INSN_UID (next_usage_insns) < max_uid)
5649 || (GET_CODE (next_usage_insns) == INSN_LIST
5650 && (INSN_UID (XEXP (next_usage_insns, 0)) < max_uid)))
5651 && need_for_split_p (potential_reload_hard_regs, regno + i)
5652 && split_reg (before_p, regno + i, insn, next_usage_insns))
5653 res = true;
5654 return res;
5657 /* Return TRUE if rtx X is considered as an invariant for
5658 inheritance. */
5659 static bool
5660 invariant_p (const_rtx x)
5662 machine_mode mode;
5663 const char *fmt;
5664 enum rtx_code code;
5665 int i, j;
5667 code = GET_CODE (x);
5668 mode = GET_MODE (x);
5669 if (code == SUBREG)
5671 x = SUBREG_REG (x);
5672 code = GET_CODE (x);
5673 mode = wider_subreg_mode (mode, GET_MODE (x));
5676 if (MEM_P (x))
5677 return false;
5679 if (REG_P (x))
5681 int i, nregs, regno = REGNO (x);
5683 if (regno >= FIRST_PSEUDO_REGISTER || regno == STACK_POINTER_REGNUM
5684 || TEST_HARD_REG_BIT (eliminable_regset, regno)
5685 || GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
5686 return false;
5687 nregs = hard_regno_nregs (regno, mode);
5688 for (i = 0; i < nregs; i++)
5689 if (! fixed_regs[regno + i]
5690 /* A hard register may be clobbered in the current insn
5691 but we can ignore this case because if the hard
5692 register is used it should be set somewhere after the
5693 clobber. */
5694 || bitmap_bit_p (&invalid_invariant_regs, regno + i))
5695 return false;
5697 fmt = GET_RTX_FORMAT (code);
5698 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
5700 if (fmt[i] == 'e')
5702 if (! invariant_p (XEXP (x, i)))
5703 return false;
5705 else if (fmt[i] == 'E')
5707 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
5708 if (! invariant_p (XVECEXP (x, i, j)))
5709 return false;
5712 return true;
5715 /* We have 'dest_reg <- invariant'. Let us try to make an invariant
5716 inheritance transformation (using dest_reg instead invariant in a
5717 subsequent insn). */
5718 static bool
5719 process_invariant_for_inheritance (rtx dst_reg, rtx invariant_rtx)
5721 invariant_ptr_t invariant_ptr;
5722 rtx_insn *insn, *new_insns;
5723 rtx insn_set, insn_reg, new_reg;
5724 int insn_regno;
5725 bool succ_p = false;
5726 int dst_regno = REGNO (dst_reg);
5727 machine_mode dst_mode = GET_MODE (dst_reg);
5728 enum reg_class cl = lra_get_allocno_class (dst_regno), insn_reg_cl;
5730 invariant_ptr = insert_invariant (invariant_rtx);
5731 if ((insn = invariant_ptr->insn) != NULL_RTX)
5733 /* We have a subsequent insn using the invariant. */
5734 insn_set = single_set (insn);
5735 lra_assert (insn_set != NULL);
5736 insn_reg = SET_DEST (insn_set);
5737 lra_assert (REG_P (insn_reg));
5738 insn_regno = REGNO (insn_reg);
5739 insn_reg_cl = lra_get_allocno_class (insn_regno);
5741 if (dst_mode == GET_MODE (insn_reg)
5742 /* We should consider only result move reg insns which are
5743 cheap. */
5744 && targetm.register_move_cost (dst_mode, cl, insn_reg_cl) == 2
5745 && targetm.register_move_cost (dst_mode, cl, cl) == 2)
5747 if (lra_dump_file != NULL)
5748 fprintf (lra_dump_file,
5749 " [[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[\n");
5750 new_reg = lra_create_new_reg (dst_mode, dst_reg,
5751 cl, "invariant inheritance");
5752 bitmap_set_bit (&lra_inheritance_pseudos, REGNO (new_reg));
5753 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
5754 lra_reg_info[REGNO (new_reg)].restore_rtx = PATTERN (insn);
5755 start_sequence ();
5756 lra_emit_move (new_reg, dst_reg);
5757 new_insns = get_insns ();
5758 end_sequence ();
5759 lra_process_new_insns (curr_insn, NULL, new_insns,
5760 "Add invariant inheritance<-original");
5761 start_sequence ();
5762 lra_emit_move (SET_DEST (insn_set), new_reg);
5763 new_insns = get_insns ();
5764 end_sequence ();
5765 lra_process_new_insns (insn, NULL, new_insns,
5766 "Changing reload<-inheritance");
5767 lra_set_insn_deleted (insn);
5768 succ_p = true;
5769 if (lra_dump_file != NULL)
5771 fprintf (lra_dump_file,
5772 " Invariant inheritance reuse change %d (bb%d):\n",
5773 REGNO (new_reg), BLOCK_FOR_INSN (insn)->index);
5774 dump_insn_slim (lra_dump_file, insn);
5775 fprintf (lra_dump_file,
5776 " ]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]\n");
5780 invariant_ptr->insn = curr_insn;
5781 return succ_p;
5784 /* Check only registers living at the current program point in the
5785 current EBB. */
5786 static bitmap_head live_regs;
5788 /* Update live info in EBB given by its HEAD and TAIL insns after
5789 inheritance/split transformation. The function removes dead moves
5790 too. */
5791 static void
5792 update_ebb_live_info (rtx_insn *head, rtx_insn *tail)
5794 unsigned int j;
5795 int i, regno;
5796 bool live_p;
5797 rtx_insn *prev_insn;
5798 rtx set;
5799 bool remove_p;
5800 basic_block last_bb, prev_bb, curr_bb;
5801 bitmap_iterator bi;
5802 struct lra_insn_reg *reg;
5803 edge e;
5804 edge_iterator ei;
5806 last_bb = BLOCK_FOR_INSN (tail);
5807 prev_bb = NULL;
5808 for (curr_insn = tail;
5809 curr_insn != PREV_INSN (head);
5810 curr_insn = prev_insn)
5812 prev_insn = PREV_INSN (curr_insn);
5813 /* We need to process empty blocks too. They contain
5814 NOTE_INSN_BASIC_BLOCK referring for the basic block. */
5815 if (NOTE_P (curr_insn) && NOTE_KIND (curr_insn) != NOTE_INSN_BASIC_BLOCK)
5816 continue;
5817 curr_bb = BLOCK_FOR_INSN (curr_insn);
5818 if (!curr_bb)
5820 gcc_assert (DEBUG_INSN_P (curr_insn));
5821 if (DEBUG_MARKER_INSN_P (curr_insn))
5822 continue;
5823 curr_bb = prev_bb;
5825 if (curr_bb != prev_bb)
5827 if (prev_bb != NULL)
5829 /* Update df_get_live_in (prev_bb): */
5830 EXECUTE_IF_SET_IN_BITMAP (&check_only_regs, 0, j, bi)
5831 if (bitmap_bit_p (&live_regs, j))
5832 bitmap_set_bit (df_get_live_in (prev_bb), j);
5833 else
5834 bitmap_clear_bit (df_get_live_in (prev_bb), j);
5836 if (curr_bb != last_bb)
5838 /* Update df_get_live_out (curr_bb): */
5839 EXECUTE_IF_SET_IN_BITMAP (&check_only_regs, 0, j, bi)
5841 live_p = bitmap_bit_p (&live_regs, j);
5842 if (! live_p)
5843 FOR_EACH_EDGE (e, ei, curr_bb->succs)
5844 if (bitmap_bit_p (df_get_live_in (e->dest), j))
5846 live_p = true;
5847 break;
5849 if (live_p)
5850 bitmap_set_bit (df_get_live_out (curr_bb), j);
5851 else
5852 bitmap_clear_bit (df_get_live_out (curr_bb), j);
5855 prev_bb = curr_bb;
5856 bitmap_and (&live_regs, &check_only_regs, df_get_live_out (curr_bb));
5858 if (! NONDEBUG_INSN_P (curr_insn))
5859 continue;
5860 curr_id = lra_get_insn_recog_data (curr_insn);
5861 curr_static_id = curr_id->insn_static_data;
5862 remove_p = false;
5863 if ((set = single_set (curr_insn)) != NULL_RTX
5864 && REG_P (SET_DEST (set))
5865 && (regno = REGNO (SET_DEST (set))) >= FIRST_PSEUDO_REGISTER
5866 && SET_DEST (set) != pic_offset_table_rtx
5867 && bitmap_bit_p (&check_only_regs, regno)
5868 && ! bitmap_bit_p (&live_regs, regno))
5869 remove_p = true;
5870 /* See which defined values die here. */
5871 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5872 if (reg->type == OP_OUT && ! reg->subreg_p)
5873 bitmap_clear_bit (&live_regs, reg->regno);
5874 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
5875 if (reg->type == OP_OUT && ! reg->subreg_p)
5876 bitmap_clear_bit (&live_regs, reg->regno);
5877 if (curr_id->arg_hard_regs != NULL)
5878 /* Make clobbered argument hard registers die. */
5879 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
5880 if (regno >= FIRST_PSEUDO_REGISTER)
5881 bitmap_clear_bit (&live_regs, regno - FIRST_PSEUDO_REGISTER);
5882 /* Mark each used value as live. */
5883 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5884 if (reg->type != OP_OUT
5885 && bitmap_bit_p (&check_only_regs, reg->regno))
5886 bitmap_set_bit (&live_regs, reg->regno);
5887 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
5888 if (reg->type != OP_OUT
5889 && bitmap_bit_p (&check_only_regs, reg->regno))
5890 bitmap_set_bit (&live_regs, reg->regno);
5891 if (curr_id->arg_hard_regs != NULL)
5892 /* Make used argument hard registers live. */
5893 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
5894 if (regno < FIRST_PSEUDO_REGISTER
5895 && bitmap_bit_p (&check_only_regs, regno))
5896 bitmap_set_bit (&live_regs, regno);
5897 /* It is quite important to remove dead move insns because it
5898 means removing dead store. We don't need to process them for
5899 constraints. */
5900 if (remove_p)
5902 if (lra_dump_file != NULL)
5904 fprintf (lra_dump_file, " Removing dead insn:\n ");
5905 dump_insn_slim (lra_dump_file, curr_insn);
5907 lra_set_insn_deleted (curr_insn);
5912 /* The structure describes info to do an inheritance for the current
5913 insn. We need to collect such info first before doing the
5914 transformations because the transformations change the insn
5915 internal representation. */
5916 struct to_inherit
5918 /* Original regno. */
5919 int regno;
5920 /* Subsequent insns which can inherit original reg value. */
5921 rtx insns;
5924 /* Array containing all info for doing inheritance from the current
5925 insn. */
5926 static struct to_inherit to_inherit[LRA_MAX_INSN_RELOADS];
5928 /* Number elements in the previous array. */
5929 static int to_inherit_num;
5931 /* Add inheritance info REGNO and INSNS. Their meaning is described in
5932 structure to_inherit. */
5933 static void
5934 add_to_inherit (int regno, rtx insns)
5936 int i;
5938 for (i = 0; i < to_inherit_num; i++)
5939 if (to_inherit[i].regno == regno)
5940 return;
5941 lra_assert (to_inherit_num < LRA_MAX_INSN_RELOADS);
5942 to_inherit[to_inherit_num].regno = regno;
5943 to_inherit[to_inherit_num++].insns = insns;
5946 /* Return the last non-debug insn in basic block BB, or the block begin
5947 note if none. */
5948 static rtx_insn *
5949 get_last_insertion_point (basic_block bb)
5951 rtx_insn *insn;
5953 FOR_BB_INSNS_REVERSE (bb, insn)
5954 if (NONDEBUG_INSN_P (insn) || NOTE_INSN_BASIC_BLOCK_P (insn))
5955 return insn;
5956 gcc_unreachable ();
5959 /* Set up RES by registers living on edges FROM except the edge (FROM,
5960 TO) or by registers set up in a jump insn in BB FROM. */
5961 static void
5962 get_live_on_other_edges (basic_block from, basic_block to, bitmap res)
5964 rtx_insn *last;
5965 struct lra_insn_reg *reg;
5966 edge e;
5967 edge_iterator ei;
5969 lra_assert (to != NULL);
5970 bitmap_clear (res);
5971 FOR_EACH_EDGE (e, ei, from->succs)
5972 if (e->dest != to)
5973 bitmap_ior_into (res, df_get_live_in (e->dest));
5974 last = get_last_insertion_point (from);
5975 if (! JUMP_P (last))
5976 return;
5977 curr_id = lra_get_insn_recog_data (last);
5978 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5979 if (reg->type != OP_IN)
5980 bitmap_set_bit (res, reg->regno);
5983 /* Used as a temporary results of some bitmap calculations. */
5984 static bitmap_head temp_bitmap;
5986 /* We split for reloads of small class of hard regs. The following
5987 defines how many hard regs the class should have to be qualified as
5988 small. The code is mostly oriented to x86/x86-64 architecture
5989 where some insns need to use only specific register or pair of
5990 registers and these register can live in RTL explicitly, e.g. for
5991 parameter passing. */
5992 static const int max_small_class_regs_num = 2;
5994 /* Do inheritance/split transformations in EBB starting with HEAD and
5995 finishing on TAIL. We process EBB insns in the reverse order.
5996 Return true if we did any inheritance/split transformation in the
5997 EBB.
5999 We should avoid excessive splitting which results in worse code
6000 because of inaccurate cost calculations for spilling new split
6001 pseudos in such case. To achieve this we do splitting only if
6002 register pressure is high in given basic block and there are reload
6003 pseudos requiring hard registers. We could do more register
6004 pressure calculations at any given program point to avoid necessary
6005 splitting even more but it is to expensive and the current approach
6006 works well enough. */
6007 static bool
6008 inherit_in_ebb (rtx_insn *head, rtx_insn *tail)
6010 int i, src_regno, dst_regno, nregs;
6011 bool change_p, succ_p, update_reloads_num_p;
6012 rtx_insn *prev_insn, *last_insn;
6013 rtx next_usage_insns, curr_set;
6014 enum reg_class cl;
6015 struct lra_insn_reg *reg;
6016 basic_block last_processed_bb, curr_bb = NULL;
6017 HARD_REG_SET potential_reload_hard_regs, live_hard_regs;
6018 bitmap to_process;
6019 unsigned int j;
6020 bitmap_iterator bi;
6021 bool head_p, after_p;
6023 change_p = false;
6024 curr_usage_insns_check++;
6025 clear_invariants ();
6026 reloads_num = calls_num = 0;
6027 bitmap_clear (&check_only_regs);
6028 bitmap_clear (&invalid_invariant_regs);
6029 last_processed_bb = NULL;
6030 CLEAR_HARD_REG_SET (potential_reload_hard_regs);
6031 COPY_HARD_REG_SET (live_hard_regs, eliminable_regset);
6032 IOR_HARD_REG_SET (live_hard_regs, lra_no_alloc_regs);
6033 /* We don't process new insns generated in the loop. */
6034 for (curr_insn = tail; curr_insn != PREV_INSN (head); curr_insn = prev_insn)
6036 prev_insn = PREV_INSN (curr_insn);
6037 if (BLOCK_FOR_INSN (curr_insn) != NULL)
6038 curr_bb = BLOCK_FOR_INSN (curr_insn);
6039 if (last_processed_bb != curr_bb)
6041 /* We are at the end of BB. Add qualified living
6042 pseudos for potential splitting. */
6043 to_process = df_get_live_out (curr_bb);
6044 if (last_processed_bb != NULL)
6046 /* We are somewhere in the middle of EBB. */
6047 get_live_on_other_edges (curr_bb, last_processed_bb,
6048 &temp_bitmap);
6049 to_process = &temp_bitmap;
6051 last_processed_bb = curr_bb;
6052 last_insn = get_last_insertion_point (curr_bb);
6053 after_p = (! JUMP_P (last_insn)
6054 && (! CALL_P (last_insn)
6055 || (find_reg_note (last_insn,
6056 REG_NORETURN, NULL_RTX) == NULL_RTX
6057 && ! SIBLING_CALL_P (last_insn))));
6058 CLEAR_HARD_REG_SET (potential_reload_hard_regs);
6059 EXECUTE_IF_SET_IN_BITMAP (to_process, 0, j, bi)
6061 if ((int) j >= lra_constraint_new_regno_start)
6062 break;
6063 if (j < FIRST_PSEUDO_REGISTER || reg_renumber[j] >= 0)
6065 if (j < FIRST_PSEUDO_REGISTER)
6066 SET_HARD_REG_BIT (live_hard_regs, j);
6067 else
6068 add_to_hard_reg_set (&live_hard_regs,
6069 PSEUDO_REGNO_MODE (j),
6070 reg_renumber[j]);
6071 setup_next_usage_insn (j, last_insn, reloads_num, after_p);
6075 src_regno = dst_regno = -1;
6076 curr_set = single_set (curr_insn);
6077 if (curr_set != NULL_RTX && REG_P (SET_DEST (curr_set)))
6078 dst_regno = REGNO (SET_DEST (curr_set));
6079 if (curr_set != NULL_RTX && REG_P (SET_SRC (curr_set)))
6080 src_regno = REGNO (SET_SRC (curr_set));
6081 update_reloads_num_p = true;
6082 if (src_regno < lra_constraint_new_regno_start
6083 && src_regno >= FIRST_PSEUDO_REGISTER
6084 && reg_renumber[src_regno] < 0
6085 && dst_regno >= lra_constraint_new_regno_start
6086 && (cl = lra_get_allocno_class (dst_regno)) != NO_REGS)
6088 /* 'reload_pseudo <- original_pseudo'. */
6089 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6090 reloads_num++;
6091 update_reloads_num_p = false;
6092 succ_p = false;
6093 if (usage_insns[src_regno].check == curr_usage_insns_check
6094 && (next_usage_insns = usage_insns[src_regno].insns) != NULL_RTX)
6095 succ_p = inherit_reload_reg (false, src_regno, cl,
6096 curr_insn, next_usage_insns);
6097 if (succ_p)
6098 change_p = true;
6099 else
6100 setup_next_usage_insn (src_regno, curr_insn, reloads_num, false);
6101 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6102 IOR_HARD_REG_SET (potential_reload_hard_regs,
6103 reg_class_contents[cl]);
6105 else if (src_regno < 0
6106 && dst_regno >= lra_constraint_new_regno_start
6107 && invariant_p (SET_SRC (curr_set))
6108 && (cl = lra_get_allocno_class (dst_regno)) != NO_REGS
6109 && ! bitmap_bit_p (&invalid_invariant_regs, dst_regno)
6110 && ! bitmap_bit_p (&invalid_invariant_regs,
6111 ORIGINAL_REGNO(regno_reg_rtx[dst_regno])))
6113 /* 'reload_pseudo <- invariant'. */
6114 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6115 reloads_num++;
6116 update_reloads_num_p = false;
6117 if (process_invariant_for_inheritance (SET_DEST (curr_set), SET_SRC (curr_set)))
6118 change_p = true;
6119 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6120 IOR_HARD_REG_SET (potential_reload_hard_regs,
6121 reg_class_contents[cl]);
6123 else if (src_regno >= lra_constraint_new_regno_start
6124 && dst_regno < lra_constraint_new_regno_start
6125 && dst_regno >= FIRST_PSEUDO_REGISTER
6126 && reg_renumber[dst_regno] < 0
6127 && (cl = lra_get_allocno_class (src_regno)) != NO_REGS
6128 && usage_insns[dst_regno].check == curr_usage_insns_check
6129 && (next_usage_insns
6130 = usage_insns[dst_regno].insns) != NULL_RTX)
6132 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6133 reloads_num++;
6134 update_reloads_num_p = false;
6135 /* 'original_pseudo <- reload_pseudo'. */
6136 if (! JUMP_P (curr_insn)
6137 && inherit_reload_reg (true, dst_regno, cl,
6138 curr_insn, next_usage_insns))
6139 change_p = true;
6140 /* Invalidate. */
6141 usage_insns[dst_regno].check = 0;
6142 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6143 IOR_HARD_REG_SET (potential_reload_hard_regs,
6144 reg_class_contents[cl]);
6146 else if (INSN_P (curr_insn))
6148 int iter;
6149 int max_uid = get_max_uid ();
6151 curr_id = lra_get_insn_recog_data (curr_insn);
6152 curr_static_id = curr_id->insn_static_data;
6153 to_inherit_num = 0;
6154 /* Process insn definitions. */
6155 for (iter = 0; iter < 2; iter++)
6156 for (reg = iter == 0 ? curr_id->regs : curr_static_id->hard_regs;
6157 reg != NULL;
6158 reg = reg->next)
6159 if (reg->type != OP_IN
6160 && (dst_regno = reg->regno) < lra_constraint_new_regno_start)
6162 if (dst_regno >= FIRST_PSEUDO_REGISTER && reg->type == OP_OUT
6163 && reg_renumber[dst_regno] < 0 && ! reg->subreg_p
6164 && usage_insns[dst_regno].check == curr_usage_insns_check
6165 && (next_usage_insns
6166 = usage_insns[dst_regno].insns) != NULL_RTX)
6168 struct lra_insn_reg *r;
6170 for (r = curr_id->regs; r != NULL; r = r->next)
6171 if (r->type != OP_OUT && r->regno == dst_regno)
6172 break;
6173 /* Don't do inheritance if the pseudo is also
6174 used in the insn. */
6175 if (r == NULL)
6176 /* We can not do inheritance right now
6177 because the current insn reg info (chain
6178 regs) can change after that. */
6179 add_to_inherit (dst_regno, next_usage_insns);
6181 /* We can not process one reg twice here because of
6182 usage_insns invalidation. */
6183 if ((dst_regno < FIRST_PSEUDO_REGISTER
6184 || reg_renumber[dst_regno] >= 0)
6185 && ! reg->subreg_p && reg->type != OP_IN)
6187 HARD_REG_SET s;
6189 if (split_if_necessary (dst_regno, reg->biggest_mode,
6190 potential_reload_hard_regs,
6191 false, curr_insn, max_uid))
6192 change_p = true;
6193 CLEAR_HARD_REG_SET (s);
6194 if (dst_regno < FIRST_PSEUDO_REGISTER)
6195 add_to_hard_reg_set (&s, reg->biggest_mode, dst_regno);
6196 else
6197 add_to_hard_reg_set (&s, PSEUDO_REGNO_MODE (dst_regno),
6198 reg_renumber[dst_regno]);
6199 AND_COMPL_HARD_REG_SET (live_hard_regs, s);
6201 /* We should invalidate potential inheritance or
6202 splitting for the current insn usages to the next
6203 usage insns (see code below) as the output pseudo
6204 prevents this. */
6205 if ((dst_regno >= FIRST_PSEUDO_REGISTER
6206 && reg_renumber[dst_regno] < 0)
6207 || (reg->type == OP_OUT && ! reg->subreg_p
6208 && (dst_regno < FIRST_PSEUDO_REGISTER
6209 || reg_renumber[dst_regno] >= 0)))
6211 /* Invalidate and mark definitions. */
6212 if (dst_regno >= FIRST_PSEUDO_REGISTER)
6213 usage_insns[dst_regno].check = -(int) INSN_UID (curr_insn);
6214 else
6216 nregs = hard_regno_nregs (dst_regno,
6217 reg->biggest_mode);
6218 for (i = 0; i < nregs; i++)
6219 usage_insns[dst_regno + i].check
6220 = -(int) INSN_UID (curr_insn);
6224 /* Process clobbered call regs. */
6225 if (curr_id->arg_hard_regs != NULL)
6226 for (i = 0; (dst_regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6227 if (dst_regno >= FIRST_PSEUDO_REGISTER)
6228 usage_insns[dst_regno - FIRST_PSEUDO_REGISTER].check
6229 = -(int) INSN_UID (curr_insn);
6230 if (! JUMP_P (curr_insn))
6231 for (i = 0; i < to_inherit_num; i++)
6232 if (inherit_reload_reg (true, to_inherit[i].regno,
6233 ALL_REGS, curr_insn,
6234 to_inherit[i].insns))
6235 change_p = true;
6236 if (CALL_P (curr_insn))
6238 rtx cheap, pat, dest;
6239 rtx_insn *restore;
6240 int regno, hard_regno;
6242 calls_num++;
6243 if ((cheap = find_reg_note (curr_insn,
6244 REG_RETURNED, NULL_RTX)) != NULL_RTX
6245 && ((cheap = XEXP (cheap, 0)), true)
6246 && (regno = REGNO (cheap)) >= FIRST_PSEUDO_REGISTER
6247 && (hard_regno = reg_renumber[regno]) >= 0
6248 && usage_insns[regno].check == curr_usage_insns_check
6249 /* If there are pending saves/restores, the
6250 optimization is not worth. */
6251 && usage_insns[regno].calls_num == calls_num - 1
6252 && TEST_HARD_REG_BIT (call_used_reg_set, hard_regno))
6254 /* Restore the pseudo from the call result as
6255 REG_RETURNED note says that the pseudo value is
6256 in the call result and the pseudo is an argument
6257 of the call. */
6258 pat = PATTERN (curr_insn);
6259 if (GET_CODE (pat) == PARALLEL)
6260 pat = XVECEXP (pat, 0, 0);
6261 dest = SET_DEST (pat);
6262 /* For multiple return values dest is PARALLEL.
6263 Currently we handle only single return value case. */
6264 if (REG_P (dest))
6266 start_sequence ();
6267 emit_move_insn (cheap, copy_rtx (dest));
6268 restore = get_insns ();
6269 end_sequence ();
6270 lra_process_new_insns (curr_insn, NULL, restore,
6271 "Inserting call parameter restore");
6272 /* We don't need to save/restore of the pseudo from
6273 this call. */
6274 usage_insns[regno].calls_num = calls_num;
6275 bitmap_set_bit (&check_only_regs, regno);
6279 to_inherit_num = 0;
6280 /* Process insn usages. */
6281 for (iter = 0; iter < 2; iter++)
6282 for (reg = iter == 0 ? curr_id->regs : curr_static_id->hard_regs;
6283 reg != NULL;
6284 reg = reg->next)
6285 if ((reg->type != OP_OUT
6286 || (reg->type == OP_OUT && reg->subreg_p))
6287 && (src_regno = reg->regno) < lra_constraint_new_regno_start)
6289 if (src_regno >= FIRST_PSEUDO_REGISTER
6290 && reg_renumber[src_regno] < 0 && reg->type == OP_IN)
6292 if (usage_insns[src_regno].check == curr_usage_insns_check
6293 && (next_usage_insns
6294 = usage_insns[src_regno].insns) != NULL_RTX
6295 && NONDEBUG_INSN_P (curr_insn))
6296 add_to_inherit (src_regno, next_usage_insns);
6297 else if (usage_insns[src_regno].check
6298 != -(int) INSN_UID (curr_insn))
6299 /* Add usages but only if the reg is not set up
6300 in the same insn. */
6301 add_next_usage_insn (src_regno, curr_insn, reloads_num);
6303 else if (src_regno < FIRST_PSEUDO_REGISTER
6304 || reg_renumber[src_regno] >= 0)
6306 bool before_p;
6307 rtx_insn *use_insn = curr_insn;
6309 before_p = (JUMP_P (curr_insn)
6310 || (CALL_P (curr_insn) && reg->type == OP_IN));
6311 if (NONDEBUG_INSN_P (curr_insn)
6312 && (! JUMP_P (curr_insn) || reg->type == OP_IN)
6313 && split_if_necessary (src_regno, reg->biggest_mode,
6314 potential_reload_hard_regs,
6315 before_p, curr_insn, max_uid))
6317 if (reg->subreg_p)
6318 lra_risky_transformations_p = true;
6319 change_p = true;
6320 /* Invalidate. */
6321 usage_insns[src_regno].check = 0;
6322 if (before_p)
6323 use_insn = PREV_INSN (curr_insn);
6325 if (NONDEBUG_INSN_P (curr_insn))
6327 if (src_regno < FIRST_PSEUDO_REGISTER)
6328 add_to_hard_reg_set (&live_hard_regs,
6329 reg->biggest_mode, src_regno);
6330 else
6331 add_to_hard_reg_set (&live_hard_regs,
6332 PSEUDO_REGNO_MODE (src_regno),
6333 reg_renumber[src_regno]);
6335 add_next_usage_insn (src_regno, use_insn, reloads_num);
6338 /* Process used call regs. */
6339 if (curr_id->arg_hard_regs != NULL)
6340 for (i = 0; (src_regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6341 if (src_regno < FIRST_PSEUDO_REGISTER)
6343 SET_HARD_REG_BIT (live_hard_regs, src_regno);
6344 add_next_usage_insn (src_regno, curr_insn, reloads_num);
6346 for (i = 0; i < to_inherit_num; i++)
6348 src_regno = to_inherit[i].regno;
6349 if (inherit_reload_reg (false, src_regno, ALL_REGS,
6350 curr_insn, to_inherit[i].insns))
6351 change_p = true;
6352 else
6353 setup_next_usage_insn (src_regno, curr_insn, reloads_num, false);
6356 if (update_reloads_num_p
6357 && NONDEBUG_INSN_P (curr_insn) && curr_set != NULL_RTX)
6359 int regno = -1;
6360 if ((REG_P (SET_DEST (curr_set))
6361 && (regno = REGNO (SET_DEST (curr_set))) >= lra_constraint_new_regno_start
6362 && reg_renumber[regno] < 0
6363 && (cl = lra_get_allocno_class (regno)) != NO_REGS)
6364 || (REG_P (SET_SRC (curr_set))
6365 && (regno = REGNO (SET_SRC (curr_set))) >= lra_constraint_new_regno_start
6366 && reg_renumber[regno] < 0
6367 && (cl = lra_get_allocno_class (regno)) != NO_REGS))
6369 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6370 reloads_num++;
6371 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6372 IOR_HARD_REG_SET (potential_reload_hard_regs,
6373 reg_class_contents[cl]);
6376 if (NONDEBUG_INSN_P (curr_insn))
6378 int regno;
6380 /* Invalidate invariants with changed regs. */
6381 curr_id = lra_get_insn_recog_data (curr_insn);
6382 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
6383 if (reg->type != OP_IN)
6385 bitmap_set_bit (&invalid_invariant_regs, reg->regno);
6386 bitmap_set_bit (&invalid_invariant_regs,
6387 ORIGINAL_REGNO (regno_reg_rtx[reg->regno]));
6389 curr_static_id = curr_id->insn_static_data;
6390 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
6391 if (reg->type != OP_IN)
6392 bitmap_set_bit (&invalid_invariant_regs, reg->regno);
6393 if (curr_id->arg_hard_regs != NULL)
6394 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6395 if (regno >= FIRST_PSEUDO_REGISTER)
6396 bitmap_set_bit (&invalid_invariant_regs,
6397 regno - FIRST_PSEUDO_REGISTER);
6399 /* We reached the start of the current basic block. */
6400 if (prev_insn == NULL_RTX || prev_insn == PREV_INSN (head)
6401 || BLOCK_FOR_INSN (prev_insn) != curr_bb)
6403 /* We reached the beginning of the current block -- do
6404 rest of spliting in the current BB. */
6405 to_process = df_get_live_in (curr_bb);
6406 if (BLOCK_FOR_INSN (head) != curr_bb)
6408 /* We are somewhere in the middle of EBB. */
6409 get_live_on_other_edges (EDGE_PRED (curr_bb, 0)->src,
6410 curr_bb, &temp_bitmap);
6411 to_process = &temp_bitmap;
6413 head_p = true;
6414 EXECUTE_IF_SET_IN_BITMAP (to_process, 0, j, bi)
6416 if ((int) j >= lra_constraint_new_regno_start)
6417 break;
6418 if (((int) j < FIRST_PSEUDO_REGISTER || reg_renumber[j] >= 0)
6419 && usage_insns[j].check == curr_usage_insns_check
6420 && (next_usage_insns = usage_insns[j].insns) != NULL_RTX)
6422 if (need_for_split_p (potential_reload_hard_regs, j))
6424 if (lra_dump_file != NULL && head_p)
6426 fprintf (lra_dump_file,
6427 " ----------------------------------\n");
6428 head_p = false;
6430 if (split_reg (false, j, bb_note (curr_bb),
6431 next_usage_insns))
6432 change_p = true;
6434 usage_insns[j].check = 0;
6439 return change_p;
6442 /* This value affects EBB forming. If probability of edge from EBB to
6443 a BB is not greater than the following value, we don't add the BB
6444 to EBB. */
6445 #define EBB_PROBABILITY_CUTOFF \
6446 ((REG_BR_PROB_BASE * LRA_INHERITANCE_EBB_PROBABILITY_CUTOFF) / 100)
6448 /* Current number of inheritance/split iteration. */
6449 int lra_inheritance_iter;
6451 /* Entry function for inheritance/split pass. */
6452 void
6453 lra_inheritance (void)
6455 int i;
6456 basic_block bb, start_bb;
6457 edge e;
6459 lra_inheritance_iter++;
6460 if (lra_inheritance_iter > LRA_MAX_INHERITANCE_PASSES)
6461 return;
6462 timevar_push (TV_LRA_INHERITANCE);
6463 if (lra_dump_file != NULL)
6464 fprintf (lra_dump_file, "\n********** Inheritance #%d: **********\n\n",
6465 lra_inheritance_iter);
6466 curr_usage_insns_check = 0;
6467 usage_insns = XNEWVEC (struct usage_insns, lra_constraint_new_regno_start);
6468 for (i = 0; i < lra_constraint_new_regno_start; i++)
6469 usage_insns[i].check = 0;
6470 bitmap_initialize (&check_only_regs, &reg_obstack);
6471 bitmap_initialize (&invalid_invariant_regs, &reg_obstack);
6472 bitmap_initialize (&live_regs, &reg_obstack);
6473 bitmap_initialize (&temp_bitmap, &reg_obstack);
6474 bitmap_initialize (&ebb_global_regs, &reg_obstack);
6475 FOR_EACH_BB_FN (bb, cfun)
6477 start_bb = bb;
6478 if (lra_dump_file != NULL)
6479 fprintf (lra_dump_file, "EBB");
6480 /* Form a EBB starting with BB. */
6481 bitmap_clear (&ebb_global_regs);
6482 bitmap_ior_into (&ebb_global_regs, df_get_live_in (bb));
6483 for (;;)
6485 if (lra_dump_file != NULL)
6486 fprintf (lra_dump_file, " %d", bb->index);
6487 if (bb->next_bb == EXIT_BLOCK_PTR_FOR_FN (cfun)
6488 || LABEL_P (BB_HEAD (bb->next_bb)))
6489 break;
6490 e = find_fallthru_edge (bb->succs);
6491 if (! e)
6492 break;
6493 if (e->probability.initialized_p ()
6494 && e->probability.to_reg_br_prob_base () < EBB_PROBABILITY_CUTOFF)
6495 break;
6496 bb = bb->next_bb;
6498 bitmap_ior_into (&ebb_global_regs, df_get_live_out (bb));
6499 if (lra_dump_file != NULL)
6500 fprintf (lra_dump_file, "\n");
6501 if (inherit_in_ebb (BB_HEAD (start_bb), BB_END (bb)))
6502 /* Remember that the EBB head and tail can change in
6503 inherit_in_ebb. */
6504 update_ebb_live_info (BB_HEAD (start_bb), BB_END (bb));
6506 bitmap_clear (&ebb_global_regs);
6507 bitmap_clear (&temp_bitmap);
6508 bitmap_clear (&live_regs);
6509 bitmap_clear (&invalid_invariant_regs);
6510 bitmap_clear (&check_only_regs);
6511 free (usage_insns);
6513 timevar_pop (TV_LRA_INHERITANCE);
6518 /* This page contains code to undo failed inheritance/split
6519 transformations. */
6521 /* Current number of iteration undoing inheritance/split. */
6522 int lra_undo_inheritance_iter;
6524 /* Fix BB live info LIVE after removing pseudos created on pass doing
6525 inheritance/split which are REMOVED_PSEUDOS. */
6526 static void
6527 fix_bb_live_info (bitmap live, bitmap removed_pseudos)
6529 unsigned int regno;
6530 bitmap_iterator bi;
6532 EXECUTE_IF_SET_IN_BITMAP (removed_pseudos, 0, regno, bi)
6533 if (bitmap_clear_bit (live, regno)
6534 && REG_P (lra_reg_info[regno].restore_rtx))
6535 bitmap_set_bit (live, REGNO (lra_reg_info[regno].restore_rtx));
6538 /* Return regno of the (subreg of) REG. Otherwise, return a negative
6539 number. */
6540 static int
6541 get_regno (rtx reg)
6543 if (GET_CODE (reg) == SUBREG)
6544 reg = SUBREG_REG (reg);
6545 if (REG_P (reg))
6546 return REGNO (reg);
6547 return -1;
6550 /* Delete a move INSN with destination reg DREGNO and a previous
6551 clobber insn with the same regno. The inheritance/split code can
6552 generate moves with preceding clobber and when we delete such moves
6553 we should delete the clobber insn too to keep the correct life
6554 info. */
6555 static void
6556 delete_move_and_clobber (rtx_insn *insn, int dregno)
6558 rtx_insn *prev_insn = PREV_INSN (insn);
6560 lra_set_insn_deleted (insn);
6561 lra_assert (dregno >= 0);
6562 if (prev_insn != NULL && NONDEBUG_INSN_P (prev_insn)
6563 && GET_CODE (PATTERN (prev_insn)) == CLOBBER
6564 && dregno == get_regno (XEXP (PATTERN (prev_insn), 0)))
6565 lra_set_insn_deleted (prev_insn);
6568 /* Remove inheritance/split pseudos which are in REMOVE_PSEUDOS and
6569 return true if we did any change. The undo transformations for
6570 inheritance looks like
6571 i <- i2
6572 p <- i => p <- i2
6573 or removing
6574 p <- i, i <- p, and i <- i3
6575 where p is original pseudo from which inheritance pseudo i was
6576 created, i and i3 are removed inheritance pseudos, i2 is another
6577 not removed inheritance pseudo. All split pseudos or other
6578 occurrences of removed inheritance pseudos are changed on the
6579 corresponding original pseudos.
6581 The function also schedules insns changed and created during
6582 inheritance/split pass for processing by the subsequent constraint
6583 pass. */
6584 static bool
6585 remove_inheritance_pseudos (bitmap remove_pseudos)
6587 basic_block bb;
6588 int regno, sregno, prev_sregno, dregno;
6589 rtx restore_rtx;
6590 rtx set, prev_set;
6591 rtx_insn *prev_insn;
6592 bool change_p, done_p;
6594 change_p = ! bitmap_empty_p (remove_pseudos);
6595 /* We can not finish the function right away if CHANGE_P is true
6596 because we need to marks insns affected by previous
6597 inheritance/split pass for processing by the subsequent
6598 constraint pass. */
6599 FOR_EACH_BB_FN (bb, cfun)
6601 fix_bb_live_info (df_get_live_in (bb), remove_pseudos);
6602 fix_bb_live_info (df_get_live_out (bb), remove_pseudos);
6603 FOR_BB_INSNS_REVERSE (bb, curr_insn)
6605 if (! INSN_P (curr_insn))
6606 continue;
6607 done_p = false;
6608 sregno = dregno = -1;
6609 if (change_p && NONDEBUG_INSN_P (curr_insn)
6610 && (set = single_set (curr_insn)) != NULL_RTX)
6612 dregno = get_regno (SET_DEST (set));
6613 sregno = get_regno (SET_SRC (set));
6616 if (sregno >= 0 && dregno >= 0)
6618 if (bitmap_bit_p (remove_pseudos, dregno)
6619 && ! REG_P (lra_reg_info[dregno].restore_rtx))
6621 /* invariant inheritance pseudo <- original pseudo */
6622 if (lra_dump_file != NULL)
6624 fprintf (lra_dump_file, " Removing invariant inheritance:\n");
6625 dump_insn_slim (lra_dump_file, curr_insn);
6626 fprintf (lra_dump_file, "\n");
6628 delete_move_and_clobber (curr_insn, dregno);
6629 done_p = true;
6631 else if (bitmap_bit_p (remove_pseudos, sregno)
6632 && ! REG_P (lra_reg_info[sregno].restore_rtx))
6634 /* reload pseudo <- invariant inheritance pseudo */
6635 start_sequence ();
6636 /* We can not just change the source. It might be
6637 an insn different from the move. */
6638 emit_insn (lra_reg_info[sregno].restore_rtx);
6639 rtx_insn *new_insns = get_insns ();
6640 end_sequence ();
6641 lra_assert (single_set (new_insns) != NULL
6642 && SET_DEST (set) == SET_DEST (single_set (new_insns)));
6643 lra_process_new_insns (curr_insn, NULL, new_insns,
6644 "Changing reload<-invariant inheritance");
6645 delete_move_and_clobber (curr_insn, dregno);
6646 done_p = true;
6648 else if ((bitmap_bit_p (remove_pseudos, sregno)
6649 && (get_regno (lra_reg_info[sregno].restore_rtx) == dregno
6650 || (bitmap_bit_p (remove_pseudos, dregno)
6651 && get_regno (lra_reg_info[sregno].restore_rtx) >= 0
6652 && (get_regno (lra_reg_info[sregno].restore_rtx)
6653 == get_regno (lra_reg_info[dregno].restore_rtx)))))
6654 || (bitmap_bit_p (remove_pseudos, dregno)
6655 && get_regno (lra_reg_info[dregno].restore_rtx) == sregno))
6656 /* One of the following cases:
6657 original <- removed inheritance pseudo
6658 removed inherit pseudo <- another removed inherit pseudo
6659 removed inherit pseudo <- original pseudo
6661 removed_split_pseudo <- original_reg
6662 original_reg <- removed_split_pseudo */
6664 if (lra_dump_file != NULL)
6666 fprintf (lra_dump_file, " Removing %s:\n",
6667 bitmap_bit_p (&lra_split_regs, sregno)
6668 || bitmap_bit_p (&lra_split_regs, dregno)
6669 ? "split" : "inheritance");
6670 dump_insn_slim (lra_dump_file, curr_insn);
6672 delete_move_and_clobber (curr_insn, dregno);
6673 done_p = true;
6675 else if (bitmap_bit_p (remove_pseudos, sregno)
6676 && bitmap_bit_p (&lra_inheritance_pseudos, sregno))
6678 /* Search the following pattern:
6679 inherit_or_split_pseudo1 <- inherit_or_split_pseudo2
6680 original_pseudo <- inherit_or_split_pseudo1
6681 where the 2nd insn is the current insn and
6682 inherit_or_split_pseudo2 is not removed. If it is found,
6683 change the current insn onto:
6684 original_pseudo <- inherit_or_split_pseudo2. */
6685 for (prev_insn = PREV_INSN (curr_insn);
6686 prev_insn != NULL_RTX && ! NONDEBUG_INSN_P (prev_insn);
6687 prev_insn = PREV_INSN (prev_insn))
6689 if (prev_insn != NULL_RTX && BLOCK_FOR_INSN (prev_insn) == bb
6690 && (prev_set = single_set (prev_insn)) != NULL_RTX
6691 /* There should be no subregs in insn we are
6692 searching because only the original reg might
6693 be in subreg when we changed the mode of
6694 load/store for splitting. */
6695 && REG_P (SET_DEST (prev_set))
6696 && REG_P (SET_SRC (prev_set))
6697 && (int) REGNO (SET_DEST (prev_set)) == sregno
6698 && ((prev_sregno = REGNO (SET_SRC (prev_set)))
6699 >= FIRST_PSEUDO_REGISTER)
6700 && (lra_reg_info[prev_sregno].restore_rtx == NULL_RTX
6702 /* As we consider chain of inheritance or
6703 splitting described in above comment we should
6704 check that sregno and prev_sregno were
6705 inheritance/split pseudos created from the
6706 same original regno. */
6707 (get_regno (lra_reg_info[sregno].restore_rtx) >= 0
6708 && (get_regno (lra_reg_info[sregno].restore_rtx)
6709 == get_regno (lra_reg_info[prev_sregno].restore_rtx))))
6710 && ! bitmap_bit_p (remove_pseudos, prev_sregno))
6712 lra_assert (GET_MODE (SET_SRC (prev_set))
6713 == GET_MODE (regno_reg_rtx[sregno]));
6714 if (GET_CODE (SET_SRC (set)) == SUBREG)
6715 SUBREG_REG (SET_SRC (set)) = SET_SRC (prev_set);
6716 else
6717 SET_SRC (set) = SET_SRC (prev_set);
6718 /* As we are finishing with processing the insn
6719 here, check the destination too as it might
6720 inheritance pseudo for another pseudo. */
6721 if (bitmap_bit_p (remove_pseudos, dregno)
6722 && bitmap_bit_p (&lra_inheritance_pseudos, dregno)
6723 && (restore_rtx
6724 = lra_reg_info[dregno].restore_rtx) != NULL_RTX)
6726 if (GET_CODE (SET_DEST (set)) == SUBREG)
6727 SUBREG_REG (SET_DEST (set)) = restore_rtx;
6728 else
6729 SET_DEST (set) = restore_rtx;
6731 lra_push_insn_and_update_insn_regno_info (curr_insn);
6732 lra_set_used_insn_alternative_by_uid
6733 (INSN_UID (curr_insn), -1);
6734 done_p = true;
6735 if (lra_dump_file != NULL)
6737 fprintf (lra_dump_file, " Change reload insn:\n");
6738 dump_insn_slim (lra_dump_file, curr_insn);
6743 if (! done_p)
6745 struct lra_insn_reg *reg;
6746 bool restored_regs_p = false;
6747 bool kept_regs_p = false;
6749 curr_id = lra_get_insn_recog_data (curr_insn);
6750 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
6752 regno = reg->regno;
6753 restore_rtx = lra_reg_info[regno].restore_rtx;
6754 if (restore_rtx != NULL_RTX)
6756 if (change_p && bitmap_bit_p (remove_pseudos, regno))
6758 lra_substitute_pseudo_within_insn
6759 (curr_insn, regno, restore_rtx, false);
6760 restored_regs_p = true;
6762 else
6763 kept_regs_p = true;
6766 if (NONDEBUG_INSN_P (curr_insn) && kept_regs_p)
6768 /* The instruction has changed since the previous
6769 constraints pass. */
6770 lra_push_insn_and_update_insn_regno_info (curr_insn);
6771 lra_set_used_insn_alternative_by_uid
6772 (INSN_UID (curr_insn), -1);
6774 else if (restored_regs_p)
6775 /* The instruction has been restored to the form that
6776 it had during the previous constraints pass. */
6777 lra_update_insn_regno_info (curr_insn);
6778 if (restored_regs_p && lra_dump_file != NULL)
6780 fprintf (lra_dump_file, " Insn after restoring regs:\n");
6781 dump_insn_slim (lra_dump_file, curr_insn);
6786 return change_p;
6789 /* If optional reload pseudos failed to get a hard register or was not
6790 inherited, it is better to remove optional reloads. We do this
6791 transformation after undoing inheritance to figure out necessity to
6792 remove optional reloads easier. Return true if we do any
6793 change. */
6794 static bool
6795 undo_optional_reloads (void)
6797 bool change_p, keep_p;
6798 unsigned int regno, uid;
6799 bitmap_iterator bi, bi2;
6800 rtx_insn *insn;
6801 rtx set, src, dest;
6802 auto_bitmap removed_optional_reload_pseudos (&reg_obstack);
6804 bitmap_copy (removed_optional_reload_pseudos, &lra_optional_reload_pseudos);
6805 EXECUTE_IF_SET_IN_BITMAP (&lra_optional_reload_pseudos, 0, regno, bi)
6807 keep_p = false;
6808 /* Keep optional reloads from previous subpasses. */
6809 if (lra_reg_info[regno].restore_rtx == NULL_RTX
6810 /* If the original pseudo changed its allocation, just
6811 removing the optional pseudo is dangerous as the original
6812 pseudo will have longer live range. */
6813 || reg_renumber[REGNO (lra_reg_info[regno].restore_rtx)] >= 0)
6814 keep_p = true;
6815 else if (reg_renumber[regno] >= 0)
6816 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi2)
6818 insn = lra_insn_recog_data[uid]->insn;
6819 if ((set = single_set (insn)) == NULL_RTX)
6820 continue;
6821 src = SET_SRC (set);
6822 dest = SET_DEST (set);
6823 if (! REG_P (src) || ! REG_P (dest))
6824 continue;
6825 if (REGNO (dest) == regno
6826 /* Ignore insn for optional reloads itself. */
6827 && REGNO (lra_reg_info[regno].restore_rtx) != REGNO (src)
6828 /* Check only inheritance on last inheritance pass. */
6829 && (int) REGNO (src) >= new_regno_start
6830 /* Check that the optional reload was inherited. */
6831 && bitmap_bit_p (&lra_inheritance_pseudos, REGNO (src)))
6833 keep_p = true;
6834 break;
6837 if (keep_p)
6839 bitmap_clear_bit (removed_optional_reload_pseudos, regno);
6840 if (lra_dump_file != NULL)
6841 fprintf (lra_dump_file, "Keep optional reload reg %d\n", regno);
6844 change_p = ! bitmap_empty_p (removed_optional_reload_pseudos);
6845 auto_bitmap insn_bitmap (&reg_obstack);
6846 EXECUTE_IF_SET_IN_BITMAP (removed_optional_reload_pseudos, 0, regno, bi)
6848 if (lra_dump_file != NULL)
6849 fprintf (lra_dump_file, "Remove optional reload reg %d\n", regno);
6850 bitmap_copy (insn_bitmap, &lra_reg_info[regno].insn_bitmap);
6851 EXECUTE_IF_SET_IN_BITMAP (insn_bitmap, 0, uid, bi2)
6853 insn = lra_insn_recog_data[uid]->insn;
6854 if ((set = single_set (insn)) != NULL_RTX)
6856 src = SET_SRC (set);
6857 dest = SET_DEST (set);
6858 if (REG_P (src) && REG_P (dest)
6859 && ((REGNO (src) == regno
6860 && (REGNO (lra_reg_info[regno].restore_rtx)
6861 == REGNO (dest)))
6862 || (REGNO (dest) == regno
6863 && (REGNO (lra_reg_info[regno].restore_rtx)
6864 == REGNO (src)))))
6866 if (lra_dump_file != NULL)
6868 fprintf (lra_dump_file, " Deleting move %u\n",
6869 INSN_UID (insn));
6870 dump_insn_slim (lra_dump_file, insn);
6872 delete_move_and_clobber (insn, REGNO (dest));
6873 continue;
6875 /* We should not worry about generation memory-memory
6876 moves here as if the corresponding inheritance did
6877 not work (inheritance pseudo did not get a hard reg),
6878 we remove the inheritance pseudo and the optional
6879 reload. */
6881 lra_substitute_pseudo_within_insn
6882 (insn, regno, lra_reg_info[regno].restore_rtx, false);
6883 lra_update_insn_regno_info (insn);
6884 if (lra_dump_file != NULL)
6886 fprintf (lra_dump_file,
6887 " Restoring original insn:\n");
6888 dump_insn_slim (lra_dump_file, insn);
6892 /* Clear restore_regnos. */
6893 EXECUTE_IF_SET_IN_BITMAP (&lra_optional_reload_pseudos, 0, regno, bi)
6894 lra_reg_info[regno].restore_rtx = NULL_RTX;
6895 return change_p;
6898 /* Entry function for undoing inheritance/split transformation. Return true
6899 if we did any RTL change in this pass. */
6900 bool
6901 lra_undo_inheritance (void)
6903 unsigned int regno;
6904 int hard_regno;
6905 int n_all_inherit, n_inherit, n_all_split, n_split;
6906 rtx restore_rtx;
6907 bitmap_iterator bi;
6908 bool change_p;
6910 lra_undo_inheritance_iter++;
6911 if (lra_undo_inheritance_iter > LRA_MAX_INHERITANCE_PASSES)
6912 return false;
6913 if (lra_dump_file != NULL)
6914 fprintf (lra_dump_file,
6915 "\n********** Undoing inheritance #%d: **********\n\n",
6916 lra_undo_inheritance_iter);
6917 auto_bitmap remove_pseudos (&reg_obstack);
6918 n_inherit = n_all_inherit = 0;
6919 EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, regno, bi)
6920 if (lra_reg_info[regno].restore_rtx != NULL_RTX)
6922 n_all_inherit++;
6923 if (reg_renumber[regno] < 0
6924 /* If the original pseudo changed its allocation, just
6925 removing inheritance is dangerous as for changing
6926 allocation we used shorter live-ranges. */
6927 && (! REG_P (lra_reg_info[regno].restore_rtx)
6928 || reg_renumber[REGNO (lra_reg_info[regno].restore_rtx)] < 0))
6929 bitmap_set_bit (remove_pseudos, regno);
6930 else
6931 n_inherit++;
6933 if (lra_dump_file != NULL && n_all_inherit != 0)
6934 fprintf (lra_dump_file, "Inherit %d out of %d (%.2f%%)\n",
6935 n_inherit, n_all_inherit,
6936 (double) n_inherit / n_all_inherit * 100);
6937 n_split = n_all_split = 0;
6938 EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, regno, bi)
6939 if ((restore_rtx = lra_reg_info[regno].restore_rtx) != NULL_RTX)
6941 int restore_regno = REGNO (restore_rtx);
6943 n_all_split++;
6944 hard_regno = (restore_regno >= FIRST_PSEUDO_REGISTER
6945 ? reg_renumber[restore_regno] : restore_regno);
6946 if (hard_regno < 0 || reg_renumber[regno] == hard_regno)
6947 bitmap_set_bit (remove_pseudos, regno);
6948 else
6950 n_split++;
6951 if (lra_dump_file != NULL)
6952 fprintf (lra_dump_file, " Keep split r%d (orig=r%d)\n",
6953 regno, restore_regno);
6956 if (lra_dump_file != NULL && n_all_split != 0)
6957 fprintf (lra_dump_file, "Split %d out of %d (%.2f%%)\n",
6958 n_split, n_all_split,
6959 (double) n_split / n_all_split * 100);
6960 change_p = remove_inheritance_pseudos (remove_pseudos);
6961 /* Clear restore_regnos. */
6962 EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, regno, bi)
6963 lra_reg_info[regno].restore_rtx = NULL_RTX;
6964 EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, regno, bi)
6965 lra_reg_info[regno].restore_rtx = NULL_RTX;
6966 change_p = undo_optional_reloads () || change_p;
6967 return change_p;