Default to dwarf version 4 on hppa64-hpux
[official-gcc.git] / gcc / lra-constraints.c
blob4d734548c3879076dd158ffec736b336f0e83847
1 /* Code for RTL transformations to satisfy insn constraints.
2 Copyright (C) 2010-2021 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 "lra.h"
131 #include "lra-int.h"
132 #include "print-rtl.h"
133 #include "function-abi.h"
134 #include "rtl-iter.h"
136 /* Value of LRA_CURR_RELOAD_NUM at the beginning of BB of the current
137 insn. Remember that LRA_CURR_RELOAD_NUM is the number of emitted
138 reload insns. */
139 static int bb_reload_num;
141 /* The current insn being processed and corresponding its single set
142 (NULL otherwise), its data (basic block, the insn data, the insn
143 static data, and the mode of each operand). */
144 static rtx_insn *curr_insn;
145 static rtx curr_insn_set;
146 static basic_block curr_bb;
147 static lra_insn_recog_data_t curr_id;
148 static struct lra_static_insn_data *curr_static_id;
149 static machine_mode curr_operand_mode[MAX_RECOG_OPERANDS];
150 /* Mode of the register substituted by its equivalence with VOIDmode
151 (e.g. constant) and whose subreg is given operand of the current
152 insn. VOIDmode in all other cases. */
153 static machine_mode original_subreg_reg_mode[MAX_RECOG_OPERANDS];
157 /* Start numbers for new registers and insns at the current constraints
158 pass start. */
159 static int new_regno_start;
160 static int new_insn_uid_start;
162 /* If LOC is nonnull, strip any outer subreg from it. */
163 static inline rtx *
164 strip_subreg (rtx *loc)
166 return loc && GET_CODE (*loc) == SUBREG ? &SUBREG_REG (*loc) : loc;
169 /* Return hard regno of REGNO or if it is was not assigned to a hard
170 register, use a hard register from its allocno class. */
171 static int
172 get_try_hard_regno (int regno)
174 int hard_regno;
175 enum reg_class rclass;
177 if ((hard_regno = regno) >= FIRST_PSEUDO_REGISTER)
178 hard_regno = lra_get_regno_hard_regno (regno);
179 if (hard_regno >= 0)
180 return hard_regno;
181 rclass = lra_get_allocno_class (regno);
182 if (rclass == NO_REGS)
183 return -1;
184 return ira_class_hard_regs[rclass][0];
187 /* Return the hard regno of X after removing its subreg. If X is not
188 a register or a subreg of a register, return -1. If X is a pseudo,
189 use its assignment. If FINAL_P return the final hard regno which will
190 be after elimination. */
191 static int
192 get_hard_regno (rtx x, bool final_p)
194 rtx reg;
195 int hard_regno;
197 reg = x;
198 if (SUBREG_P (x))
199 reg = SUBREG_REG (x);
200 if (! REG_P (reg))
201 return -1;
202 if (! HARD_REGISTER_NUM_P (hard_regno = REGNO (reg)))
203 hard_regno = lra_get_regno_hard_regno (hard_regno);
204 if (hard_regno < 0)
205 return -1;
206 if (final_p)
207 hard_regno = lra_get_elimination_hard_regno (hard_regno);
208 if (SUBREG_P (x))
209 hard_regno += subreg_regno_offset (hard_regno, GET_MODE (reg),
210 SUBREG_BYTE (x), GET_MODE (x));
211 return hard_regno;
214 /* If REGNO is a hard register or has been allocated a hard register,
215 return the class of that register. If REGNO is a reload pseudo
216 created by the current constraints pass, return its allocno class.
217 Return NO_REGS otherwise. */
218 static enum reg_class
219 get_reg_class (int regno)
221 int hard_regno;
223 if (! HARD_REGISTER_NUM_P (hard_regno = regno))
224 hard_regno = lra_get_regno_hard_regno (regno);
225 if (hard_regno >= 0)
227 hard_regno = lra_get_elimination_hard_regno (hard_regno);
228 return REGNO_REG_CLASS (hard_regno);
230 if (regno >= new_regno_start)
231 return lra_get_allocno_class (regno);
232 return NO_REGS;
235 /* Return true if REG satisfies (or will satisfy) reg class constraint
236 CL. Use elimination first if REG is a hard register. If REG is a
237 reload pseudo created by this constraints pass, assume that it will
238 be allocated a hard register from its allocno class, but allow that
239 class to be narrowed to CL if it is currently a superset of CL and
240 if either:
242 - ALLOW_ALL_RELOAD_CLASS_CHANGES_P is true or
243 - the instruction we're processing is not a reload move.
245 If NEW_CLASS is nonnull, set *NEW_CLASS to the new allocno class of
246 REGNO (reg), or NO_REGS if no change in its class was needed. */
247 static bool
248 in_class_p (rtx reg, enum reg_class cl, enum reg_class *new_class,
249 bool allow_all_reload_class_changes_p = false)
251 enum reg_class rclass, common_class;
252 machine_mode reg_mode;
253 rtx src;
254 int class_size, hard_regno, nregs, i, j;
255 int regno = REGNO (reg);
257 if (new_class != NULL)
258 *new_class = NO_REGS;
259 if (regno < FIRST_PSEUDO_REGISTER)
261 rtx final_reg = reg;
262 rtx *final_loc = &final_reg;
264 lra_eliminate_reg_if_possible (final_loc);
265 return TEST_HARD_REG_BIT (reg_class_contents[cl], REGNO (*final_loc));
267 reg_mode = GET_MODE (reg);
268 rclass = get_reg_class (regno);
269 src = curr_insn_set != NULL ? SET_SRC (curr_insn_set) : NULL;
270 if (regno < new_regno_start
271 /* Do not allow the constraints for reload instructions to
272 influence the classes of new pseudos. These reloads are
273 typically moves that have many alternatives, and restricting
274 reload pseudos for one alternative may lead to situations
275 where other reload pseudos are no longer allocatable. */
276 || (!allow_all_reload_class_changes_p
277 && INSN_UID (curr_insn) >= new_insn_uid_start
278 && src != NULL
279 && ((REG_P (src) || MEM_P (src))
280 || (GET_CODE (src) == SUBREG
281 && (REG_P (SUBREG_REG (src)) || MEM_P (SUBREG_REG (src)))))))
282 /* When we don't know what class will be used finally for reload
283 pseudos, we use ALL_REGS. */
284 return ((regno >= new_regno_start && rclass == ALL_REGS)
285 || (rclass != NO_REGS && ira_class_subset_p[rclass][cl]
286 && ! hard_reg_set_subset_p (reg_class_contents[cl],
287 lra_no_alloc_regs)));
288 else
290 common_class = ira_reg_class_subset[rclass][cl];
291 if (new_class != NULL)
292 *new_class = common_class;
293 if (hard_reg_set_subset_p (reg_class_contents[common_class],
294 lra_no_alloc_regs))
295 return false;
296 /* Check that there are enough allocatable regs. */
297 class_size = ira_class_hard_regs_num[common_class];
298 for (i = 0; i < class_size; i++)
300 hard_regno = ira_class_hard_regs[common_class][i];
301 nregs = hard_regno_nregs (hard_regno, reg_mode);
302 if (nregs == 1)
303 return true;
304 for (j = 0; j < nregs; j++)
305 if (TEST_HARD_REG_BIT (lra_no_alloc_regs, hard_regno + j)
306 || ! TEST_HARD_REG_BIT (reg_class_contents[common_class],
307 hard_regno + j))
308 break;
309 if (j >= nregs)
310 return true;
312 return false;
316 /* Return true if REGNO satisfies a memory constraint. */
317 static bool
318 in_mem_p (int regno)
320 return get_reg_class (regno) == NO_REGS;
323 /* Return 1 if ADDR is a valid memory address for mode MODE in address
324 space AS, and check that each pseudo has the proper kind of hard
325 reg. */
326 static int
327 valid_address_p (machine_mode mode ATTRIBUTE_UNUSED,
328 rtx addr, addr_space_t as)
330 #ifdef GO_IF_LEGITIMATE_ADDRESS
331 lra_assert (ADDR_SPACE_GENERIC_P (as));
332 GO_IF_LEGITIMATE_ADDRESS (mode, addr, win);
333 return 0;
335 win:
336 return 1;
337 #else
338 return targetm.addr_space.legitimate_address_p (mode, addr, 0, as);
339 #endif
342 namespace {
343 /* Temporarily eliminates registers in an address (for the lifetime of
344 the object). */
345 class address_eliminator {
346 public:
347 address_eliminator (struct address_info *ad);
348 ~address_eliminator ();
350 private:
351 struct address_info *m_ad;
352 rtx *m_base_loc;
353 rtx m_base_reg;
354 rtx *m_index_loc;
355 rtx m_index_reg;
359 address_eliminator::address_eliminator (struct address_info *ad)
360 : m_ad (ad),
361 m_base_loc (strip_subreg (ad->base_term)),
362 m_base_reg (NULL_RTX),
363 m_index_loc (strip_subreg (ad->index_term)),
364 m_index_reg (NULL_RTX)
366 if (m_base_loc != NULL)
368 m_base_reg = *m_base_loc;
369 /* If we have non-legitimate address which is decomposed not in
370 the way we expected, don't do elimination here. In such case
371 the address will be reloaded and elimination will be done in
372 reload insn finally. */
373 if (REG_P (m_base_reg))
374 lra_eliminate_reg_if_possible (m_base_loc);
375 if (m_ad->base_term2 != NULL)
376 *m_ad->base_term2 = *m_ad->base_term;
378 if (m_index_loc != NULL)
380 m_index_reg = *m_index_loc;
381 if (REG_P (m_index_reg))
382 lra_eliminate_reg_if_possible (m_index_loc);
386 address_eliminator::~address_eliminator ()
388 if (m_base_loc && *m_base_loc != m_base_reg)
390 *m_base_loc = m_base_reg;
391 if (m_ad->base_term2 != NULL)
392 *m_ad->base_term2 = *m_ad->base_term;
394 if (m_index_loc && *m_index_loc != m_index_reg)
395 *m_index_loc = m_index_reg;
398 /* Return true if the eliminated form of AD is a legitimate target address.
399 If OP is a MEM, AD is the address within OP, otherwise OP should be
400 ignored. CONSTRAINT is one constraint that the operand may need
401 to meet. */
402 static bool
403 valid_address_p (rtx op, struct address_info *ad,
404 enum constraint_num constraint)
406 address_eliminator eliminator (ad);
408 /* Allow a memory OP if it matches CONSTRAINT, even if CONSTRAINT is more
409 forgiving than "m".
410 Need to extract memory from op for special memory constraint,
411 i.e. bcst_mem_operand in i386 backend. */
412 if (MEM_P (extract_mem_from_operand (op))
413 && insn_extra_relaxed_memory_constraint (constraint)
414 && constraint_satisfied_p (op, constraint))
415 return true;
417 return valid_address_p (ad->mode, *ad->outer, ad->as);
420 /* For special_memory_operand, it could be false for MEM_P (op),
421 i.e. bcst_mem_operand in i386 backend.
422 Extract and return real memory operand or op. */
424 extract_mem_from_operand (rtx op)
426 for (rtx x = op;; x = XEXP (x, 0))
428 if (MEM_P (x))
429 return x;
430 if (GET_RTX_LENGTH (GET_CODE (x)) != 1
431 || GET_RTX_FORMAT (GET_CODE (x))[0] != 'e')
432 break;
434 return op;
437 /* Return true if the eliminated form of memory reference OP satisfies
438 extra (special) memory constraint CONSTRAINT. */
439 static bool
440 satisfies_memory_constraint_p (rtx op, enum constraint_num constraint)
442 struct address_info ad;
443 rtx mem = extract_mem_from_operand (op);
444 if (!MEM_P (mem))
445 return false;
447 decompose_mem_address (&ad, mem);
448 address_eliminator eliminator (&ad);
449 return constraint_satisfied_p (op, constraint);
452 /* Return true if the eliminated form of address AD satisfies extra
453 address constraint CONSTRAINT. */
454 static bool
455 satisfies_address_constraint_p (struct address_info *ad,
456 enum constraint_num constraint)
458 address_eliminator eliminator (ad);
459 return constraint_satisfied_p (*ad->outer, constraint);
462 /* Return true if the eliminated form of address OP satisfies extra
463 address constraint CONSTRAINT. */
464 static bool
465 satisfies_address_constraint_p (rtx op, enum constraint_num constraint)
467 struct address_info ad;
469 decompose_lea_address (&ad, &op);
470 return satisfies_address_constraint_p (&ad, constraint);
473 /* Initiate equivalences for LRA. As we keep original equivalences
474 before any elimination, we need to make copies otherwise any change
475 in insns might change the equivalences. */
476 void
477 lra_init_equiv (void)
479 ira_expand_reg_equiv ();
480 for (int i = FIRST_PSEUDO_REGISTER; i < max_reg_num (); i++)
482 rtx res;
484 if ((res = ira_reg_equiv[i].memory) != NULL_RTX)
485 ira_reg_equiv[i].memory = copy_rtx (res);
486 if ((res = ira_reg_equiv[i].invariant) != NULL_RTX)
487 ira_reg_equiv[i].invariant = copy_rtx (res);
491 static rtx loc_equivalence_callback (rtx, const_rtx, void *);
493 /* Update equivalence for REGNO. We need to this as the equivalence
494 might contain other pseudos which are changed by their
495 equivalences. */
496 static void
497 update_equiv (int regno)
499 rtx x;
501 if ((x = ira_reg_equiv[regno].memory) != NULL_RTX)
502 ira_reg_equiv[regno].memory
503 = simplify_replace_fn_rtx (x, NULL_RTX, loc_equivalence_callback,
504 NULL_RTX);
505 if ((x = ira_reg_equiv[regno].invariant) != NULL_RTX)
506 ira_reg_equiv[regno].invariant
507 = simplify_replace_fn_rtx (x, NULL_RTX, loc_equivalence_callback,
508 NULL_RTX);
511 /* If we have decided to substitute X with another value, return that
512 value, otherwise return X. */
513 static rtx
514 get_equiv (rtx x)
516 int regno;
517 rtx res;
519 if (! REG_P (x) || (regno = REGNO (x)) < FIRST_PSEUDO_REGISTER
520 || ! ira_reg_equiv[regno].defined_p
521 || ! ira_reg_equiv[regno].profitable_p
522 || lra_get_regno_hard_regno (regno) >= 0)
523 return x;
524 if ((res = ira_reg_equiv[regno].memory) != NULL_RTX)
526 if (targetm.cannot_substitute_mem_equiv_p (res))
527 return x;
528 return res;
530 if ((res = ira_reg_equiv[regno].constant) != NULL_RTX)
531 return res;
532 if ((res = ira_reg_equiv[regno].invariant) != NULL_RTX)
533 return res;
534 gcc_unreachable ();
537 /* If we have decided to substitute X with the equivalent value,
538 return that value after elimination for INSN, otherwise return
539 X. */
540 static rtx
541 get_equiv_with_elimination (rtx x, rtx_insn *insn)
543 rtx res = get_equiv (x);
545 if (x == res || CONSTANT_P (res))
546 return res;
547 return lra_eliminate_regs_1 (insn, res, GET_MODE (res),
548 false, false, 0, true);
551 /* Set up curr_operand_mode. */
552 static void
553 init_curr_operand_mode (void)
555 int nop = curr_static_id->n_operands;
556 for (int i = 0; i < nop; i++)
558 machine_mode mode = GET_MODE (*curr_id->operand_loc[i]);
559 if (mode == VOIDmode)
561 /* The .md mode for address operands is the mode of the
562 addressed value rather than the mode of the address itself. */
563 if (curr_id->icode >= 0 && curr_static_id->operand[i].is_address)
564 mode = Pmode;
565 else
566 mode = curr_static_id->operand[i].mode;
568 curr_operand_mode[i] = mode;
574 /* The page contains code to reuse input reloads. */
576 /* Structure describes input reload of the current insns. */
577 struct input_reload
579 /* True for input reload of matched operands. */
580 bool match_p;
581 /* Reloaded value. */
582 rtx input;
583 /* Reload pseudo used. */
584 rtx reg;
587 /* The number of elements in the following array. */
588 static int curr_insn_input_reloads_num;
589 /* Array containing info about input reloads. It is used to find the
590 same input reload and reuse the reload pseudo in this case. */
591 static struct input_reload curr_insn_input_reloads[LRA_MAX_INSN_RELOADS];
593 /* Initiate data concerning reuse of input reloads for the current
594 insn. */
595 static void
596 init_curr_insn_input_reloads (void)
598 curr_insn_input_reloads_num = 0;
601 /* The canonical form of an rtx inside a MEM is not necessarily the same as the
602 canonical form of the rtx outside the MEM. Fix this up in the case that
603 we're reloading an address (and therefore pulling it outside a MEM). */
604 static rtx
605 canonicalize_reload_addr (rtx addr)
607 subrtx_var_iterator::array_type array;
608 FOR_EACH_SUBRTX_VAR (iter, array, addr, NONCONST)
610 rtx x = *iter;
611 if (GET_CODE (x) == MULT && CONST_INT_P (XEXP (x, 1)))
613 const HOST_WIDE_INT ci = INTVAL (XEXP (x, 1));
614 const int pwr2 = exact_log2 (ci);
615 if (pwr2 > 0)
617 /* Rewrite this to use a shift instead, which is canonical when
618 outside of a MEM. */
619 PUT_CODE (x, ASHIFT);
620 XEXP (x, 1) = GEN_INT (pwr2);
625 return addr;
628 /* Create a new pseudo using MODE, RCLASS, ORIGINAL or reuse an existing
629 reload pseudo. Don't reuse an existing reload pseudo if IN_SUBREG_P
630 is true and the reused pseudo should be wrapped up in a SUBREG.
631 The result pseudo is returned through RESULT_REG. Return TRUE if we
632 created a new pseudo, FALSE if we reused an existing reload pseudo.
633 Use TITLE to describe new registers for debug purposes. */
634 static bool
635 get_reload_reg (enum op_type type, machine_mode mode, rtx original,
636 enum reg_class rclass, bool in_subreg_p,
637 const char *title, rtx *result_reg)
639 int i, regno;
640 enum reg_class new_class;
641 bool unique_p = false;
643 if (type == OP_OUT)
645 /* Output reload registers tend to start out with a conservative
646 choice of register class. Usually this is ALL_REGS, although
647 a target might narrow it (for performance reasons) through
648 targetm.preferred_reload_class. It's therefore quite common
649 for a reload instruction to require a more restrictive class
650 than the class that was originally assigned to the reload register.
652 In these situations, it's more efficient to refine the choice
653 of register class rather than create a second reload register.
654 This also helps to avoid cycling for registers that are only
655 used by reload instructions. */
656 if (REG_P (original)
657 && (int) REGNO (original) >= new_regno_start
658 && INSN_UID (curr_insn) >= new_insn_uid_start
659 && in_class_p (original, rclass, &new_class, true))
661 unsigned int regno = REGNO (original);
662 if (lra_dump_file != NULL)
664 fprintf (lra_dump_file, " Reuse r%d for output ", regno);
665 dump_value_slim (lra_dump_file, original, 1);
667 if (new_class != lra_get_allocno_class (regno))
668 lra_change_class (regno, new_class, ", change to", false);
669 if (lra_dump_file != NULL)
670 fprintf (lra_dump_file, "\n");
671 *result_reg = original;
672 return false;
674 *result_reg
675 = lra_create_new_reg_with_unique_value (mode, original, rclass, title);
676 return true;
678 /* Prevent reuse value of expression with side effects,
679 e.g. volatile memory. */
680 if (! side_effects_p (original))
681 for (i = 0; i < curr_insn_input_reloads_num; i++)
683 if (! curr_insn_input_reloads[i].match_p
684 && rtx_equal_p (curr_insn_input_reloads[i].input, original)
685 && in_class_p (curr_insn_input_reloads[i].reg, rclass, &new_class))
687 rtx reg = curr_insn_input_reloads[i].reg;
688 regno = REGNO (reg);
689 /* If input is equal to original and both are VOIDmode,
690 GET_MODE (reg) might be still different from mode.
691 Ensure we don't return *result_reg with wrong mode. */
692 if (GET_MODE (reg) != mode)
694 if (in_subreg_p)
695 continue;
696 if (maybe_lt (GET_MODE_SIZE (GET_MODE (reg)),
697 GET_MODE_SIZE (mode)))
698 continue;
699 reg = lowpart_subreg (mode, reg, GET_MODE (reg));
700 if (reg == NULL_RTX || GET_CODE (reg) != SUBREG)
701 continue;
703 *result_reg = reg;
704 if (lra_dump_file != NULL)
706 fprintf (lra_dump_file, " Reuse r%d for reload ", regno);
707 dump_value_slim (lra_dump_file, original, 1);
709 if (new_class != lra_get_allocno_class (regno))
710 lra_change_class (regno, new_class, ", change to", false);
711 if (lra_dump_file != NULL)
712 fprintf (lra_dump_file, "\n");
713 return false;
715 /* If we have an input reload with a different mode, make sure it
716 will get a different hard reg. */
717 else if (REG_P (original)
718 && REG_P (curr_insn_input_reloads[i].input)
719 && REGNO (original) == REGNO (curr_insn_input_reloads[i].input)
720 && (GET_MODE (original)
721 != GET_MODE (curr_insn_input_reloads[i].input)))
722 unique_p = true;
724 *result_reg = (unique_p
725 ? lra_create_new_reg_with_unique_value
726 : lra_create_new_reg) (mode, original, rclass, title);
727 lra_assert (curr_insn_input_reloads_num < LRA_MAX_INSN_RELOADS);
728 curr_insn_input_reloads[curr_insn_input_reloads_num].input = original;
729 curr_insn_input_reloads[curr_insn_input_reloads_num].match_p = false;
730 curr_insn_input_reloads[curr_insn_input_reloads_num++].reg = *result_reg;
731 return true;
735 /* The page contains major code to choose the current insn alternative
736 and generate reloads for it. */
738 /* Return the offset from REGNO of the least significant register
739 in (reg:MODE REGNO).
741 This function is used to tell whether two registers satisfy
742 a matching constraint. (reg:MODE1 REGNO1) matches (reg:MODE2 REGNO2) if:
744 REGNO1 + lra_constraint_offset (REGNO1, MODE1)
745 == REGNO2 + lra_constraint_offset (REGNO2, MODE2) */
747 lra_constraint_offset (int regno, machine_mode mode)
749 lra_assert (regno < FIRST_PSEUDO_REGISTER);
751 scalar_int_mode int_mode;
752 if (WORDS_BIG_ENDIAN
753 && is_a <scalar_int_mode> (mode, &int_mode)
754 && GET_MODE_SIZE (int_mode) > UNITS_PER_WORD)
755 return hard_regno_nregs (regno, mode) - 1;
756 return 0;
759 /* Like rtx_equal_p except that it allows a REG and a SUBREG to match
760 if they are the same hard reg, and has special hacks for
761 auto-increment and auto-decrement. This is specifically intended for
762 process_alt_operands to use in determining whether two operands
763 match. X is the operand whose number is the lower of the two.
765 It is supposed that X is the output operand and Y is the input
766 operand. Y_HARD_REGNO is the final hard regno of register Y or
767 register in subreg Y as we know it now. Otherwise, it is a
768 negative value. */
769 static bool
770 operands_match_p (rtx x, rtx y, int y_hard_regno)
772 int i;
773 RTX_CODE code = GET_CODE (x);
774 const char *fmt;
776 if (x == y)
777 return true;
778 if ((code == REG || (code == SUBREG && REG_P (SUBREG_REG (x))))
779 && (REG_P (y) || (GET_CODE (y) == SUBREG && REG_P (SUBREG_REG (y)))))
781 int j;
783 i = get_hard_regno (x, false);
784 if (i < 0)
785 goto slow;
787 if ((j = y_hard_regno) < 0)
788 goto slow;
790 i += lra_constraint_offset (i, GET_MODE (x));
791 j += lra_constraint_offset (j, GET_MODE (y));
793 return i == j;
796 /* If two operands must match, because they are really a single
797 operand of an assembler insn, then two post-increments are invalid
798 because the assembler insn would increment only once. On the
799 other hand, a post-increment matches ordinary indexing if the
800 post-increment is the output operand. */
801 if (code == POST_DEC || code == POST_INC || code == POST_MODIFY)
802 return operands_match_p (XEXP (x, 0), y, y_hard_regno);
804 /* Two pre-increments are invalid because the assembler insn would
805 increment only once. On the other hand, a pre-increment matches
806 ordinary indexing if the pre-increment is the input operand. */
807 if (GET_CODE (y) == PRE_DEC || GET_CODE (y) == PRE_INC
808 || GET_CODE (y) == PRE_MODIFY)
809 return operands_match_p (x, XEXP (y, 0), -1);
811 slow:
813 if (code == REG && REG_P (y))
814 return REGNO (x) == REGNO (y);
816 if (code == REG && GET_CODE (y) == SUBREG && REG_P (SUBREG_REG (y))
817 && x == SUBREG_REG (y))
818 return true;
819 if (GET_CODE (y) == REG && code == SUBREG && REG_P (SUBREG_REG (x))
820 && SUBREG_REG (x) == y)
821 return true;
823 /* Now we have disposed of all the cases in which different rtx
824 codes can match. */
825 if (code != GET_CODE (y))
826 return false;
828 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
829 if (GET_MODE (x) != GET_MODE (y))
830 return false;
832 switch (code)
834 CASE_CONST_UNIQUE:
835 return false;
837 case CONST_VECTOR:
838 if (!same_vector_encodings_p (x, y))
839 return false;
840 break;
842 case LABEL_REF:
843 return label_ref_label (x) == label_ref_label (y);
844 case SYMBOL_REF:
845 return XSTR (x, 0) == XSTR (y, 0);
847 default:
848 break;
851 /* Compare the elements. If any pair of corresponding elements fail
852 to match, return false for the whole things. */
854 fmt = GET_RTX_FORMAT (code);
855 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
857 int val, j;
858 switch (fmt[i])
860 case 'w':
861 if (XWINT (x, i) != XWINT (y, i))
862 return false;
863 break;
865 case 'i':
866 if (XINT (x, i) != XINT (y, i))
867 return false;
868 break;
870 case 'p':
871 if (maybe_ne (SUBREG_BYTE (x), SUBREG_BYTE (y)))
872 return false;
873 break;
875 case 'e':
876 val = operands_match_p (XEXP (x, i), XEXP (y, i), -1);
877 if (val == 0)
878 return false;
879 break;
881 case '0':
882 break;
884 case 'E':
885 if (XVECLEN (x, i) != XVECLEN (y, i))
886 return false;
887 for (j = XVECLEN (x, i) - 1; j >= 0; --j)
889 val = operands_match_p (XVECEXP (x, i, j), XVECEXP (y, i, j), -1);
890 if (val == 0)
891 return false;
893 break;
895 /* It is believed that rtx's at this level will never
896 contain anything but integers and other rtx's, except for
897 within LABEL_REFs and SYMBOL_REFs. */
898 default:
899 gcc_unreachable ();
902 return true;
905 /* True if X is a constant that can be forced into the constant pool.
906 MODE is the mode of the operand, or VOIDmode if not known. */
907 #define CONST_POOL_OK_P(MODE, X) \
908 ((MODE) != VOIDmode \
909 && CONSTANT_P (X) \
910 && GET_CODE (X) != HIGH \
911 && GET_MODE_SIZE (MODE).is_constant () \
912 && !targetm.cannot_force_const_mem (MODE, X))
914 /* True if C is a non-empty register class that has too few registers
915 to be safely used as a reload target class. */
916 #define SMALL_REGISTER_CLASS_P(C) \
917 (ira_class_hard_regs_num [(C)] == 1 \
918 || (ira_class_hard_regs_num [(C)] >= 1 \
919 && targetm.class_likely_spilled_p (C)))
921 /* If REG is a reload pseudo, try to make its class satisfying CL. */
922 static void
923 narrow_reload_pseudo_class (rtx reg, enum reg_class cl)
925 enum reg_class rclass;
927 /* Do not make more accurate class from reloads generated. They are
928 mostly moves with a lot of constraints. Making more accurate
929 class may results in very narrow class and impossibility of find
930 registers for several reloads of one insn. */
931 if (INSN_UID (curr_insn) >= new_insn_uid_start)
932 return;
933 if (GET_CODE (reg) == SUBREG)
934 reg = SUBREG_REG (reg);
935 if (! REG_P (reg) || (int) REGNO (reg) < new_regno_start)
936 return;
937 if (in_class_p (reg, cl, &rclass) && rclass != cl)
938 lra_change_class (REGNO (reg), rclass, " Change to", true);
941 /* Searches X for any reference to a reg with the same value as REGNO,
942 returning the rtx of the reference found if any. Otherwise,
943 returns NULL_RTX. */
944 static rtx
945 regno_val_use_in (unsigned int regno, rtx x)
947 const char *fmt;
948 int i, j;
949 rtx tem;
951 if (REG_P (x) && lra_reg_info[REGNO (x)].val == lra_reg_info[regno].val)
952 return x;
954 fmt = GET_RTX_FORMAT (GET_CODE (x));
955 for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
957 if (fmt[i] == 'e')
959 if ((tem = regno_val_use_in (regno, XEXP (x, i))))
960 return tem;
962 else if (fmt[i] == 'E')
963 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
964 if ((tem = regno_val_use_in (regno , XVECEXP (x, i, j))))
965 return tem;
968 return NULL_RTX;
971 /* Return true if all current insn non-output operands except INS (it
972 has a negaitve end marker) do not use pseudos with the same value
973 as REGNO. */
974 static bool
975 check_conflict_input_operands (int regno, signed char *ins)
977 int in;
978 int n_operands = curr_static_id->n_operands;
980 for (int nop = 0; nop < n_operands; nop++)
981 if (! curr_static_id->operand[nop].is_operator
982 && curr_static_id->operand[nop].type != OP_OUT)
984 for (int i = 0; (in = ins[i]) >= 0; i++)
985 if (in == nop)
986 break;
987 if (in < 0
988 && regno_val_use_in (regno, *curr_id->operand_loc[nop]) != NULL_RTX)
989 return false;
991 return true;
994 /* Generate reloads for matching OUT and INS (array of input operand
995 numbers with end marker -1) with reg class GOAL_CLASS, considering
996 output operands OUTS (similar array to INS) needing to be in different
997 registers. Add input and output reloads correspondingly to the lists
998 *BEFORE and *AFTER. OUT might be negative. In this case we generate
999 input reloads for matched input operands INS. EARLY_CLOBBER_P is a flag
1000 that the output operand is early clobbered for chosen alternative. */
1001 static void
1002 match_reload (signed char out, signed char *ins, signed char *outs,
1003 enum reg_class goal_class, rtx_insn **before,
1004 rtx_insn **after, bool early_clobber_p)
1006 bool out_conflict;
1007 int i, in;
1008 rtx new_in_reg, new_out_reg, reg;
1009 machine_mode inmode, outmode;
1010 rtx in_rtx = *curr_id->operand_loc[ins[0]];
1011 rtx out_rtx = out < 0 ? in_rtx : *curr_id->operand_loc[out];
1013 inmode = curr_operand_mode[ins[0]];
1014 outmode = out < 0 ? inmode : curr_operand_mode[out];
1015 push_to_sequence (*before);
1016 if (inmode != outmode)
1018 /* process_alt_operands has already checked that the mode sizes
1019 are ordered. */
1020 if (partial_subreg_p (outmode, inmode))
1022 reg = new_in_reg
1023 = lra_create_new_reg_with_unique_value (inmode, in_rtx,
1024 goal_class, "");
1025 new_out_reg = gen_lowpart_SUBREG (outmode, reg);
1026 LRA_SUBREG_P (new_out_reg) = 1;
1027 /* If the input reg is dying here, we can use the same hard
1028 register for REG and IN_RTX. We do it only for original
1029 pseudos as reload pseudos can die although original
1030 pseudos still live where reload pseudos dies. */
1031 if (REG_P (in_rtx) && (int) REGNO (in_rtx) < lra_new_regno_start
1032 && find_regno_note (curr_insn, REG_DEAD, REGNO (in_rtx))
1033 && (!early_clobber_p
1034 || check_conflict_input_operands(REGNO (in_rtx), ins)))
1035 lra_assign_reg_val (REGNO (in_rtx), REGNO (reg));
1037 else
1039 reg = new_out_reg
1040 = lra_create_new_reg_with_unique_value (outmode, out_rtx,
1041 goal_class, "");
1042 new_in_reg = gen_lowpart_SUBREG (inmode, reg);
1043 /* NEW_IN_REG is non-paradoxical subreg. We don't want
1044 NEW_OUT_REG living above. We add clobber clause for
1045 this. This is just a temporary clobber. We can remove
1046 it at the end of LRA work. */
1047 rtx_insn *clobber = emit_clobber (new_out_reg);
1048 LRA_TEMP_CLOBBER_P (PATTERN (clobber)) = 1;
1049 LRA_SUBREG_P (new_in_reg) = 1;
1050 if (GET_CODE (in_rtx) == SUBREG)
1052 rtx subreg_reg = SUBREG_REG (in_rtx);
1054 /* If SUBREG_REG is dying here and sub-registers IN_RTX
1055 and NEW_IN_REG are similar, we can use the same hard
1056 register for REG and SUBREG_REG. */
1057 if (REG_P (subreg_reg)
1058 && (int) REGNO (subreg_reg) < lra_new_regno_start
1059 && GET_MODE (subreg_reg) == outmode
1060 && known_eq (SUBREG_BYTE (in_rtx), SUBREG_BYTE (new_in_reg))
1061 && find_regno_note (curr_insn, REG_DEAD, REGNO (subreg_reg))
1062 && (! early_clobber_p
1063 || check_conflict_input_operands (REGNO (subreg_reg),
1064 ins)))
1065 lra_assign_reg_val (REGNO (subreg_reg), REGNO (reg));
1069 else
1071 /* Pseudos have values -- see comments for lra_reg_info.
1072 Different pseudos with the same value do not conflict even if
1073 they live in the same place. When we create a pseudo we
1074 assign value of original pseudo (if any) from which we
1075 created the new pseudo. If we create the pseudo from the
1076 input pseudo, the new pseudo will have no conflict with the
1077 input pseudo which is wrong when the input pseudo lives after
1078 the insn and as the new pseudo value is changed by the insn
1079 output. Therefore we create the new pseudo from the output
1080 except the case when we have single matched dying input
1081 pseudo.
1083 We cannot reuse the current output register because we might
1084 have a situation like "a <- a op b", where the constraints
1085 force the second input operand ("b") to match the output
1086 operand ("a"). "b" must then be copied into a new register
1087 so that it doesn't clobber the current value of "a".
1089 We cannot use the same value if the output pseudo is
1090 early clobbered or the input pseudo is mentioned in the
1091 output, e.g. as an address part in memory, because
1092 output reload will actually extend the pseudo liveness.
1093 We don't care about eliminable hard regs here as we are
1094 interesting only in pseudos. */
1096 /* Matching input's register value is the same as one of the other
1097 output operand. Output operands in a parallel insn must be in
1098 different registers. */
1099 out_conflict = false;
1100 if (REG_P (in_rtx))
1102 for (i = 0; outs[i] >= 0; i++)
1104 rtx other_out_rtx = *curr_id->operand_loc[outs[i]];
1105 if (REG_P (other_out_rtx)
1106 && (regno_val_use_in (REGNO (in_rtx), other_out_rtx)
1107 != NULL_RTX))
1109 out_conflict = true;
1110 break;
1115 new_in_reg = new_out_reg
1116 = (! early_clobber_p && ins[1] < 0 && REG_P (in_rtx)
1117 && (int) REGNO (in_rtx) < lra_new_regno_start
1118 && find_regno_note (curr_insn, REG_DEAD, REGNO (in_rtx))
1119 && (! early_clobber_p
1120 || check_conflict_input_operands (REGNO (in_rtx), ins))
1121 && (out < 0
1122 || regno_val_use_in (REGNO (in_rtx), out_rtx) == NULL_RTX)
1123 && !out_conflict
1124 ? lra_create_new_reg (inmode, in_rtx, goal_class, "")
1125 : lra_create_new_reg_with_unique_value (outmode, out_rtx,
1126 goal_class, ""));
1128 /* In operand can be got from transformations before processing insn
1129 constraints. One example of such transformations is subreg
1130 reloading (see function simplify_operand_subreg). The new
1131 pseudos created by the transformations might have inaccurate
1132 class (ALL_REGS) and we should make their classes more
1133 accurate. */
1134 narrow_reload_pseudo_class (in_rtx, goal_class);
1135 lra_emit_move (copy_rtx (new_in_reg), in_rtx);
1136 *before = get_insns ();
1137 end_sequence ();
1138 /* Add the new pseudo to consider values of subsequent input reload
1139 pseudos. */
1140 lra_assert (curr_insn_input_reloads_num < LRA_MAX_INSN_RELOADS);
1141 curr_insn_input_reloads[curr_insn_input_reloads_num].input = in_rtx;
1142 curr_insn_input_reloads[curr_insn_input_reloads_num].match_p = true;
1143 curr_insn_input_reloads[curr_insn_input_reloads_num++].reg = new_in_reg;
1144 for (i = 0; (in = ins[i]) >= 0; i++)
1145 if (GET_MODE (*curr_id->operand_loc[in]) == VOIDmode
1146 || GET_MODE (new_in_reg) == GET_MODE (*curr_id->operand_loc[in]))
1147 *curr_id->operand_loc[in] = new_in_reg;
1148 else
1150 lra_assert
1151 (GET_MODE (new_out_reg) == GET_MODE (*curr_id->operand_loc[in]));
1152 *curr_id->operand_loc[in] = new_out_reg;
1154 lra_update_dups (curr_id, ins);
1155 if (out < 0)
1156 return;
1157 /* See a comment for the input operand above. */
1158 narrow_reload_pseudo_class (out_rtx, goal_class);
1159 if (find_reg_note (curr_insn, REG_UNUSED, out_rtx) == NULL_RTX)
1161 reg = SUBREG_P (out_rtx) ? SUBREG_REG (out_rtx) : out_rtx;
1162 start_sequence ();
1163 /* If we had strict_low_part, use it also in reload to keep other
1164 parts unchanged but do it only for regs as strict_low_part
1165 has no sense for memory and probably there is no insn pattern
1166 to match the reload insn in memory case. */
1167 if (out >= 0 && curr_static_id->operand[out].strict_low && REG_P (reg))
1168 out_rtx = gen_rtx_STRICT_LOW_PART (VOIDmode, out_rtx);
1169 lra_emit_move (out_rtx, copy_rtx (new_out_reg));
1170 emit_insn (*after);
1171 *after = get_insns ();
1172 end_sequence ();
1174 *curr_id->operand_loc[out] = new_out_reg;
1175 lra_update_dup (curr_id, out);
1178 /* Return register class which is union of all reg classes in insn
1179 constraint alternative string starting with P. */
1180 static enum reg_class
1181 reg_class_from_constraints (const char *p)
1183 int c, len;
1184 enum reg_class op_class = NO_REGS;
1187 switch ((c = *p, len = CONSTRAINT_LEN (c, p)), c)
1189 case '#':
1190 case ',':
1191 return op_class;
1193 case 'g':
1194 op_class = reg_class_subunion[op_class][GENERAL_REGS];
1195 break;
1197 default:
1198 enum constraint_num cn = lookup_constraint (p);
1199 enum reg_class cl = reg_class_for_constraint (cn);
1200 if (cl == NO_REGS)
1202 if (insn_extra_address_constraint (cn))
1203 op_class
1204 = (reg_class_subunion
1205 [op_class][base_reg_class (VOIDmode, ADDR_SPACE_GENERIC,
1206 ADDRESS, SCRATCH)]);
1207 break;
1210 op_class = reg_class_subunion[op_class][cl];
1211 break;
1213 while ((p += len), c);
1214 return op_class;
1217 /* If OP is a register, return the class of the register as per
1218 get_reg_class, otherwise return NO_REGS. */
1219 static inline enum reg_class
1220 get_op_class (rtx op)
1222 return REG_P (op) ? get_reg_class (REGNO (op)) : NO_REGS;
1225 /* Return generated insn mem_pseudo:=val if TO_P or val:=mem_pseudo
1226 otherwise. If modes of MEM_PSEUDO and VAL are different, use
1227 SUBREG for VAL to make them equal. */
1228 static rtx_insn *
1229 emit_spill_move (bool to_p, rtx mem_pseudo, rtx val)
1231 if (GET_MODE (mem_pseudo) != GET_MODE (val))
1233 /* Usually size of mem_pseudo is greater than val size but in
1234 rare cases it can be less as it can be defined by target
1235 dependent macro HARD_REGNO_CALLER_SAVE_MODE. */
1236 if (! MEM_P (val))
1238 val = gen_lowpart_SUBREG (GET_MODE (mem_pseudo),
1239 GET_CODE (val) == SUBREG
1240 ? SUBREG_REG (val) : val);
1241 LRA_SUBREG_P (val) = 1;
1243 else
1245 mem_pseudo = gen_lowpart_SUBREG (GET_MODE (val), mem_pseudo);
1246 LRA_SUBREG_P (mem_pseudo) = 1;
1249 return to_p ? gen_move_insn (mem_pseudo, val)
1250 : gen_move_insn (val, mem_pseudo);
1253 /* Process a special case insn (register move), return true if we
1254 don't need to process it anymore. INSN should be a single set
1255 insn. Set up that RTL was changed through CHANGE_P and that hook
1256 TARGET_SECONDARY_MEMORY_NEEDED says to use secondary memory through
1257 SEC_MEM_P. */
1258 static bool
1259 check_and_process_move (bool *change_p, bool *sec_mem_p ATTRIBUTE_UNUSED)
1261 int sregno, dregno;
1262 rtx dest, src, dreg, sreg, new_reg, scratch_reg;
1263 rtx_insn *before;
1264 enum reg_class dclass, sclass, secondary_class;
1265 secondary_reload_info sri;
1267 lra_assert (curr_insn_set != NULL_RTX);
1268 dreg = dest = SET_DEST (curr_insn_set);
1269 sreg = src = SET_SRC (curr_insn_set);
1270 if (GET_CODE (dest) == SUBREG)
1271 dreg = SUBREG_REG (dest);
1272 if (GET_CODE (src) == SUBREG)
1273 sreg = SUBREG_REG (src);
1274 if (! (REG_P (dreg) || MEM_P (dreg)) || ! (REG_P (sreg) || MEM_P (sreg)))
1275 return false;
1276 sclass = dclass = NO_REGS;
1277 if (REG_P (dreg))
1278 dclass = get_reg_class (REGNO (dreg));
1279 gcc_assert (dclass < LIM_REG_CLASSES && dclass >= NO_REGS);
1280 if (dclass == ALL_REGS)
1281 /* ALL_REGS is used for new pseudos created by transformations
1282 like reload of SUBREG_REG (see function
1283 simplify_operand_subreg). We don't know their class yet. We
1284 should figure out the class from processing the insn
1285 constraints not in this fast path function. Even if ALL_REGS
1286 were a right class for the pseudo, secondary_... hooks usually
1287 are not define for ALL_REGS. */
1288 return false;
1289 if (REG_P (sreg))
1290 sclass = get_reg_class (REGNO (sreg));
1291 gcc_assert (sclass < LIM_REG_CLASSES && sclass >= NO_REGS);
1292 if (sclass == ALL_REGS)
1293 /* See comments above. */
1294 return false;
1295 if (sclass == NO_REGS && dclass == NO_REGS)
1296 return false;
1297 if (targetm.secondary_memory_needed (GET_MODE (src), sclass, dclass)
1298 && ((sclass != NO_REGS && dclass != NO_REGS)
1299 || (GET_MODE (src)
1300 != targetm.secondary_memory_needed_mode (GET_MODE (src)))))
1302 *sec_mem_p = true;
1303 return false;
1305 if (! REG_P (dreg) || ! REG_P (sreg))
1306 return false;
1307 sri.prev_sri = NULL;
1308 sri.icode = CODE_FOR_nothing;
1309 sri.extra_cost = 0;
1310 secondary_class = NO_REGS;
1311 /* Set up hard register for a reload pseudo for hook
1312 secondary_reload because some targets just ignore unassigned
1313 pseudos in the hook. */
1314 if (dclass != NO_REGS && lra_get_regno_hard_regno (REGNO (dreg)) < 0)
1316 dregno = REGNO (dreg);
1317 reg_renumber[dregno] = ira_class_hard_regs[dclass][0];
1319 else
1320 dregno = -1;
1321 if (sclass != NO_REGS && lra_get_regno_hard_regno (REGNO (sreg)) < 0)
1323 sregno = REGNO (sreg);
1324 reg_renumber[sregno] = ira_class_hard_regs[sclass][0];
1326 else
1327 sregno = -1;
1328 if (sclass != NO_REGS)
1329 secondary_class
1330 = (enum reg_class) targetm.secondary_reload (false, dest,
1331 (reg_class_t) sclass,
1332 GET_MODE (src), &sri);
1333 if (sclass == NO_REGS
1334 || ((secondary_class != NO_REGS || sri.icode != CODE_FOR_nothing)
1335 && dclass != NO_REGS))
1337 enum reg_class old_sclass = secondary_class;
1338 secondary_reload_info old_sri = sri;
1340 sri.prev_sri = NULL;
1341 sri.icode = CODE_FOR_nothing;
1342 sri.extra_cost = 0;
1343 secondary_class
1344 = (enum reg_class) targetm.secondary_reload (true, src,
1345 (reg_class_t) dclass,
1346 GET_MODE (src), &sri);
1347 /* Check the target hook consistency. */
1348 lra_assert
1349 ((secondary_class == NO_REGS && sri.icode == CODE_FOR_nothing)
1350 || (old_sclass == NO_REGS && old_sri.icode == CODE_FOR_nothing)
1351 || (secondary_class == old_sclass && sri.icode == old_sri.icode));
1353 if (sregno >= 0)
1354 reg_renumber [sregno] = -1;
1355 if (dregno >= 0)
1356 reg_renumber [dregno] = -1;
1357 if (secondary_class == NO_REGS && sri.icode == CODE_FOR_nothing)
1358 return false;
1359 *change_p = true;
1360 new_reg = NULL_RTX;
1361 if (secondary_class != NO_REGS)
1362 new_reg = lra_create_new_reg_with_unique_value (GET_MODE (src), NULL_RTX,
1363 secondary_class,
1364 "secondary");
1365 start_sequence ();
1366 if (sri.icode == CODE_FOR_nothing)
1367 lra_emit_move (new_reg, src);
1368 else
1370 enum reg_class scratch_class;
1372 scratch_class = (reg_class_from_constraints
1373 (insn_data[sri.icode].operand[2].constraint));
1374 scratch_reg = (lra_create_new_reg_with_unique_value
1375 (insn_data[sri.icode].operand[2].mode, NULL_RTX,
1376 scratch_class, "scratch"));
1377 emit_insn (GEN_FCN (sri.icode) (new_reg != NULL_RTX ? new_reg : dest,
1378 src, scratch_reg));
1380 before = get_insns ();
1381 end_sequence ();
1382 lra_process_new_insns (curr_insn, before, NULL, "Inserting the move");
1383 if (new_reg != NULL_RTX)
1384 SET_SRC (curr_insn_set) = new_reg;
1385 else
1387 if (lra_dump_file != NULL)
1389 fprintf (lra_dump_file, "Deleting move %u\n", INSN_UID (curr_insn));
1390 dump_insn_slim (lra_dump_file, curr_insn);
1392 lra_set_insn_deleted (curr_insn);
1393 return true;
1395 return false;
1398 /* The following data describe the result of process_alt_operands.
1399 The data are used in curr_insn_transform to generate reloads. */
1401 /* The chosen reg classes which should be used for the corresponding
1402 operands. */
1403 static enum reg_class goal_alt[MAX_RECOG_OPERANDS];
1404 /* True if the operand should be the same as another operand and that
1405 other operand does not need a reload. */
1406 static bool goal_alt_match_win[MAX_RECOG_OPERANDS];
1407 /* True if the operand does not need a reload. */
1408 static bool goal_alt_win[MAX_RECOG_OPERANDS];
1409 /* True if the operand can be offsetable memory. */
1410 static bool goal_alt_offmemok[MAX_RECOG_OPERANDS];
1411 /* The number of an operand to which given operand can be matched to. */
1412 static int goal_alt_matches[MAX_RECOG_OPERANDS];
1413 /* The number of elements in the following array. */
1414 static int goal_alt_dont_inherit_ops_num;
1415 /* Numbers of operands whose reload pseudos should not be inherited. */
1416 static int goal_alt_dont_inherit_ops[MAX_RECOG_OPERANDS];
1417 /* True if the insn commutative operands should be swapped. */
1418 static bool goal_alt_swapped;
1419 /* The chosen insn alternative. */
1420 static int goal_alt_number;
1422 /* True if the corresponding operand is the result of an equivalence
1423 substitution. */
1424 static bool equiv_substition_p[MAX_RECOG_OPERANDS];
1426 /* The following five variables are used to choose the best insn
1427 alternative. They reflect final characteristics of the best
1428 alternative. */
1430 /* Number of necessary reloads and overall cost reflecting the
1431 previous value and other unpleasantness of the best alternative. */
1432 static int best_losers, best_overall;
1433 /* Overall number hard registers used for reloads. For example, on
1434 some targets we need 2 general registers to reload DFmode and only
1435 one floating point register. */
1436 static int best_reload_nregs;
1437 /* Overall number reflecting distances of previous reloading the same
1438 value. The distances are counted from the current BB start. It is
1439 used to improve inheritance chances. */
1440 static int best_reload_sum;
1442 /* True if the current insn should have no correspondingly input or
1443 output reloads. */
1444 static bool no_input_reloads_p, no_output_reloads_p;
1446 /* True if we swapped the commutative operands in the current
1447 insn. */
1448 static int curr_swapped;
1450 /* if CHECK_ONLY_P is false, arrange for address element *LOC to be a
1451 register of class CL. Add any input reloads to list BEFORE. AFTER
1452 is nonnull if *LOC is an automodified value; handle that case by
1453 adding the required output reloads to list AFTER. Return true if
1454 the RTL was changed.
1456 if CHECK_ONLY_P is true, check that the *LOC is a correct address
1457 register. Return false if the address register is correct. */
1458 static bool
1459 process_addr_reg (rtx *loc, bool check_only_p, rtx_insn **before, rtx_insn **after,
1460 enum reg_class cl)
1462 int regno;
1463 enum reg_class rclass, new_class;
1464 rtx reg;
1465 rtx new_reg;
1466 machine_mode mode;
1467 bool subreg_p, before_p = false;
1469 subreg_p = GET_CODE (*loc) == SUBREG;
1470 if (subreg_p)
1472 reg = SUBREG_REG (*loc);
1473 mode = GET_MODE (reg);
1475 /* For mode with size bigger than ptr_mode, there unlikely to be "mov"
1476 between two registers with different classes, but there normally will
1477 be "mov" which transfers element of vector register into the general
1478 register, and this normally will be a subreg which should be reloaded
1479 as a whole. This is particularly likely to be triggered when
1480 -fno-split-wide-types specified. */
1481 if (!REG_P (reg)
1482 || in_class_p (reg, cl, &new_class)
1483 || known_le (GET_MODE_SIZE (mode), GET_MODE_SIZE (ptr_mode)))
1484 loc = &SUBREG_REG (*loc);
1487 reg = *loc;
1488 mode = GET_MODE (reg);
1489 if (! REG_P (reg))
1491 if (check_only_p)
1492 return true;
1493 /* Always reload memory in an address even if the target supports
1494 such addresses. */
1495 new_reg = lra_create_new_reg_with_unique_value (mode, reg, cl, "address");
1496 before_p = true;
1498 else
1500 regno = REGNO (reg);
1501 rclass = get_reg_class (regno);
1502 if (! check_only_p
1503 && (*loc = get_equiv_with_elimination (reg, curr_insn)) != reg)
1505 if (lra_dump_file != NULL)
1507 fprintf (lra_dump_file,
1508 "Changing pseudo %d in address of insn %u on equiv ",
1509 REGNO (reg), INSN_UID (curr_insn));
1510 dump_value_slim (lra_dump_file, *loc, 1);
1511 fprintf (lra_dump_file, "\n");
1513 *loc = copy_rtx (*loc);
1515 if (*loc != reg || ! in_class_p (reg, cl, &new_class))
1517 if (check_only_p)
1518 return true;
1519 reg = *loc;
1520 if (get_reload_reg (after == NULL ? OP_IN : OP_INOUT,
1521 mode, reg, cl, subreg_p, "address", &new_reg))
1522 before_p = true;
1524 else if (new_class != NO_REGS && rclass != new_class)
1526 if (check_only_p)
1527 return true;
1528 lra_change_class (regno, new_class, " Change to", true);
1529 return false;
1531 else
1532 return false;
1534 if (before_p)
1536 push_to_sequence (*before);
1537 lra_emit_move (new_reg, reg);
1538 *before = get_insns ();
1539 end_sequence ();
1541 *loc = new_reg;
1542 if (after != NULL)
1544 start_sequence ();
1545 lra_emit_move (before_p ? copy_rtx (reg) : reg, new_reg);
1546 emit_insn (*after);
1547 *after = get_insns ();
1548 end_sequence ();
1550 return true;
1553 /* Insert move insn in simplify_operand_subreg. BEFORE returns
1554 the insn to be inserted before curr insn. AFTER returns the
1555 the insn to be inserted after curr insn. ORIGREG and NEWREG
1556 are the original reg and new reg for reload. */
1557 static void
1558 insert_move_for_subreg (rtx_insn **before, rtx_insn **after, rtx origreg,
1559 rtx newreg)
1561 if (before)
1563 push_to_sequence (*before);
1564 lra_emit_move (newreg, origreg);
1565 *before = get_insns ();
1566 end_sequence ();
1568 if (after)
1570 start_sequence ();
1571 lra_emit_move (origreg, newreg);
1572 emit_insn (*after);
1573 *after = get_insns ();
1574 end_sequence ();
1578 static int valid_address_p (machine_mode mode, rtx addr, addr_space_t as);
1579 static bool process_address (int, bool, rtx_insn **, rtx_insn **);
1581 /* Make reloads for subreg in operand NOP with internal subreg mode
1582 REG_MODE, add new reloads for further processing. Return true if
1583 any change was done. */
1584 static bool
1585 simplify_operand_subreg (int nop, machine_mode reg_mode)
1587 int hard_regno, inner_hard_regno;
1588 rtx_insn *before, *after;
1589 machine_mode mode, innermode;
1590 rtx reg, new_reg;
1591 rtx operand = *curr_id->operand_loc[nop];
1592 enum reg_class regclass;
1593 enum op_type type;
1595 before = after = NULL;
1597 if (GET_CODE (operand) != SUBREG)
1598 return false;
1600 mode = GET_MODE (operand);
1601 reg = SUBREG_REG (operand);
1602 innermode = GET_MODE (reg);
1603 type = curr_static_id->operand[nop].type;
1604 if (MEM_P (reg))
1606 const bool addr_was_valid
1607 = valid_address_p (innermode, XEXP (reg, 0), MEM_ADDR_SPACE (reg));
1608 alter_subreg (curr_id->operand_loc[nop], false);
1609 rtx subst = *curr_id->operand_loc[nop];
1610 lra_assert (MEM_P (subst));
1611 const bool addr_is_valid = valid_address_p (GET_MODE (subst),
1612 XEXP (subst, 0),
1613 MEM_ADDR_SPACE (subst));
1614 if (!addr_was_valid
1615 || addr_is_valid
1616 || ((get_constraint_type (lookup_constraint
1617 (curr_static_id->operand[nop].constraint))
1618 != CT_SPECIAL_MEMORY)
1619 /* We still can reload address and if the address is
1620 valid, we can remove subreg without reloading its
1621 inner memory. */
1622 && valid_address_p (GET_MODE (subst),
1623 regno_reg_rtx
1624 [ira_class_hard_regs
1625 [base_reg_class (GET_MODE (subst),
1626 MEM_ADDR_SPACE (subst),
1627 ADDRESS, SCRATCH)][0]],
1628 MEM_ADDR_SPACE (subst))))
1630 /* If we change the address for a paradoxical subreg of memory, the
1631 new address might violate the necessary alignment or the access
1632 might be slow; take this into consideration. We need not worry
1633 about accesses beyond allocated memory for paradoxical memory
1634 subregs as we don't substitute such equiv memory (see processing
1635 equivalences in function lra_constraints) and because for spilled
1636 pseudos we allocate stack memory enough for the biggest
1637 corresponding paradoxical subreg.
1639 However, do not blindly simplify a (subreg (mem ...)) for
1640 WORD_REGISTER_OPERATIONS targets as this may lead to loading junk
1641 data into a register when the inner is narrower than outer or
1642 missing important data from memory when the inner is wider than
1643 outer. This rule only applies to modes that are no wider than
1644 a word.
1646 If valid memory becomes invalid after subreg elimination
1647 and address might be different we still have to reload
1648 memory.
1650 if ((! addr_was_valid
1651 || addr_is_valid
1652 || known_eq (GET_MODE_SIZE (mode), GET_MODE_SIZE (innermode)))
1653 && !(maybe_ne (GET_MODE_PRECISION (mode),
1654 GET_MODE_PRECISION (innermode))
1655 && known_le (GET_MODE_SIZE (mode), UNITS_PER_WORD)
1656 && known_le (GET_MODE_SIZE (innermode), UNITS_PER_WORD)
1657 && WORD_REGISTER_OPERATIONS)
1658 && (!(MEM_ALIGN (subst) < GET_MODE_ALIGNMENT (mode)
1659 && targetm.slow_unaligned_access (mode, MEM_ALIGN (subst)))
1660 || (MEM_ALIGN (reg) < GET_MODE_ALIGNMENT (innermode)
1661 && targetm.slow_unaligned_access (innermode,
1662 MEM_ALIGN (reg)))))
1663 return true;
1665 *curr_id->operand_loc[nop] = operand;
1667 /* But if the address was not valid, we cannot reload the MEM without
1668 reloading the address first. */
1669 if (!addr_was_valid)
1670 process_address (nop, false, &before, &after);
1672 /* INNERMODE is fast, MODE slow. Reload the mem in INNERMODE. */
1673 enum reg_class rclass
1674 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1675 if (get_reload_reg (curr_static_id->operand[nop].type, innermode,
1676 reg, rclass, TRUE, "slow/invalid mem", &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 || partial_subreg_p (mode, innermode));
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;
1690 /* Convert to MODE. */
1691 reg = operand;
1692 rclass
1693 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1694 if (get_reload_reg (curr_static_id->operand[nop].type, mode, reg,
1695 rclass, TRUE, "slow/invalid mem", &new_reg))
1697 bool insert_before, insert_after;
1698 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1700 insert_before = type != OP_OUT;
1701 insert_after = type != OP_IN;
1702 insert_move_for_subreg (insert_before ? &before : NULL,
1703 insert_after ? &after : NULL,
1704 reg, new_reg);
1706 *curr_id->operand_loc[nop] = new_reg;
1707 lra_process_new_insns (curr_insn, before, after,
1708 "Inserting slow/invalid mem reload");
1709 return true;
1712 /* If the address was valid and became invalid, prefer to reload
1713 the memory. Typical case is when the index scale should
1714 correspond the memory. */
1715 *curr_id->operand_loc[nop] = operand;
1716 /* Do not return false here as the MEM_P (reg) will be processed
1717 later in this function. */
1719 else if (REG_P (reg) && REGNO (reg) < FIRST_PSEUDO_REGISTER)
1721 alter_subreg (curr_id->operand_loc[nop], false);
1722 return true;
1724 else if (CONSTANT_P (reg))
1726 /* Try to simplify subreg of constant. It is usually result of
1727 equivalence substitution. */
1728 if (innermode == VOIDmode
1729 && (innermode = original_subreg_reg_mode[nop]) == VOIDmode)
1730 innermode = curr_static_id->operand[nop].mode;
1731 if ((new_reg = simplify_subreg (mode, reg, innermode,
1732 SUBREG_BYTE (operand))) != NULL_RTX)
1734 *curr_id->operand_loc[nop] = new_reg;
1735 return true;
1738 /* Put constant into memory when we have mixed modes. It generates
1739 a better code in most cases as it does not need a secondary
1740 reload memory. It also prevents LRA looping when LRA is using
1741 secondary reload memory again and again. */
1742 if (CONSTANT_P (reg) && CONST_POOL_OK_P (reg_mode, reg)
1743 && SCALAR_INT_MODE_P (reg_mode) != SCALAR_INT_MODE_P (mode))
1745 SUBREG_REG (operand) = force_const_mem (reg_mode, reg);
1746 alter_subreg (curr_id->operand_loc[nop], false);
1747 return true;
1749 /* Force a reload of the SUBREG_REG if this is a constant or PLUS or
1750 if there may be a problem accessing OPERAND in the outer
1751 mode. */
1752 if ((REG_P (reg)
1753 && REGNO (reg) >= FIRST_PSEUDO_REGISTER
1754 && (hard_regno = lra_get_regno_hard_regno (REGNO (reg))) >= 0
1755 /* Don't reload paradoxical subregs because we could be looping
1756 having repeatedly final regno out of hard regs range. */
1757 && (hard_regno_nregs (hard_regno, innermode)
1758 >= hard_regno_nregs (hard_regno, mode))
1759 && simplify_subreg_regno (hard_regno, innermode,
1760 SUBREG_BYTE (operand), mode) < 0
1761 /* Don't reload subreg for matching reload. It is actually
1762 valid subreg in LRA. */
1763 && ! LRA_SUBREG_P (operand))
1764 || CONSTANT_P (reg) || GET_CODE (reg) == PLUS || MEM_P (reg))
1766 enum reg_class rclass;
1768 if (REG_P (reg))
1769 /* There is a big probability that we will get the same class
1770 for the new pseudo and we will get the same insn which
1771 means infinite looping. So spill the new pseudo. */
1772 rclass = NO_REGS;
1773 else
1774 /* The class will be defined later in curr_insn_transform. */
1775 rclass
1776 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1778 if (get_reload_reg (curr_static_id->operand[nop].type, reg_mode, reg,
1779 rclass, TRUE, "subreg reg", &new_reg))
1781 bool insert_before, insert_after;
1782 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1784 insert_before = (type != OP_OUT
1785 || read_modify_subreg_p (operand));
1786 insert_after = (type != OP_IN);
1787 insert_move_for_subreg (insert_before ? &before : NULL,
1788 insert_after ? &after : NULL,
1789 reg, new_reg);
1791 SUBREG_REG (operand) = new_reg;
1792 lra_process_new_insns (curr_insn, before, after,
1793 "Inserting subreg reload");
1794 return true;
1796 /* Force a reload for a paradoxical subreg. For paradoxical subreg,
1797 IRA allocates hardreg to the inner pseudo reg according to its mode
1798 instead of the outermode, so the size of the hardreg may not be enough
1799 to contain the outermode operand, in that case we may need to insert
1800 reload for the reg. For the following two types of paradoxical subreg,
1801 we need to insert reload:
1802 1. If the op_type is OP_IN, and the hardreg could not be paired with
1803 other hardreg to contain the outermode operand
1804 (checked by in_hard_reg_set_p), we need to insert the reload.
1805 2. If the op_type is OP_OUT or OP_INOUT.
1807 Here is a paradoxical subreg example showing how the reload is generated:
1809 (insn 5 4 7 2 (set (reg:TI 106 [ __comp ])
1810 (subreg:TI (reg:DI 107 [ __comp ]) 0)) {*movti_internal_rex64}
1812 In IRA, reg107 is allocated to a DImode hardreg. We use x86-64 as example
1813 here, if reg107 is assigned to hardreg R15, because R15 is the last
1814 hardreg, compiler cannot find another hardreg to pair with R15 to
1815 contain TImode data. So we insert a TImode reload reg180 for it.
1816 After reload is inserted:
1818 (insn 283 0 0 (set (subreg:DI (reg:TI 180 [orig:107 __comp ] [107]) 0)
1819 (reg:DI 107 [ __comp ])) -1
1820 (insn 5 4 7 2 (set (reg:TI 106 [ __comp ])
1821 (subreg:TI (reg:TI 180 [orig:107 __comp ] [107]) 0)) {*movti_internal_rex64}
1823 Two reload hard registers will be allocated to reg180 to save TImode data
1824 in LRA_assign.
1826 For LRA pseudos this should normally be handled by the biggest_mode
1827 mechanism. However, it's possible for new uses of an LRA pseudo
1828 to be introduced after we've allocated it, such as when undoing
1829 inheritance, and the allocated register might not then be appropriate
1830 for the new uses. */
1831 else if (REG_P (reg)
1832 && REGNO (reg) >= FIRST_PSEUDO_REGISTER
1833 && paradoxical_subreg_p (operand)
1834 && (inner_hard_regno = lra_get_regno_hard_regno (REGNO (reg))) >= 0
1835 && ((hard_regno
1836 = simplify_subreg_regno (inner_hard_regno, innermode,
1837 SUBREG_BYTE (operand), mode)) < 0
1838 || ((hard_regno_nregs (inner_hard_regno, innermode)
1839 < hard_regno_nregs (hard_regno, mode))
1840 && (regclass = lra_get_allocno_class (REGNO (reg)))
1841 && (type != OP_IN
1842 || !in_hard_reg_set_p (reg_class_contents[regclass],
1843 mode, hard_regno)
1844 || overlaps_hard_reg_set_p (lra_no_alloc_regs,
1845 mode, hard_regno)))))
1847 /* The class will be defined later in curr_insn_transform. */
1848 enum reg_class rclass
1849 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1851 if (get_reload_reg (curr_static_id->operand[nop].type, mode, reg,
1852 rclass, TRUE, "paradoxical subreg", &new_reg))
1854 rtx subreg;
1855 bool insert_before, insert_after;
1857 PUT_MODE (new_reg, mode);
1858 subreg = gen_lowpart_SUBREG (innermode, new_reg);
1859 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1861 insert_before = (type != OP_OUT);
1862 insert_after = (type != OP_IN);
1863 insert_move_for_subreg (insert_before ? &before : NULL,
1864 insert_after ? &after : NULL,
1865 reg, subreg);
1867 SUBREG_REG (operand) = new_reg;
1868 lra_process_new_insns (curr_insn, before, after,
1869 "Inserting paradoxical subreg reload");
1870 return true;
1872 return false;
1875 /* Return TRUE if X refers for a hard register from SET. */
1876 static bool
1877 uses_hard_regs_p (rtx x, HARD_REG_SET set)
1879 int i, j, x_hard_regno;
1880 machine_mode mode;
1881 const char *fmt;
1882 enum rtx_code code;
1884 if (x == NULL_RTX)
1885 return false;
1886 code = GET_CODE (x);
1887 mode = GET_MODE (x);
1889 if (code == SUBREG)
1891 /* For all SUBREGs we want to check whether the full multi-register
1892 overlaps the set. For normal SUBREGs this means 'get_hard_regno' of
1893 the inner register, for paradoxical SUBREGs this means the
1894 'get_hard_regno' of the full SUBREG and for complete SUBREGs either is
1895 fine. Use the wider mode for all cases. */
1896 rtx subreg = SUBREG_REG (x);
1897 mode = wider_subreg_mode (x);
1898 if (mode == GET_MODE (subreg))
1900 x = subreg;
1901 code = GET_CODE (x);
1905 if (REG_P (x) || SUBREG_P (x))
1907 x_hard_regno = get_hard_regno (x, true);
1908 return (x_hard_regno >= 0
1909 && overlaps_hard_reg_set_p (set, mode, x_hard_regno));
1911 fmt = GET_RTX_FORMAT (code);
1912 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1914 if (fmt[i] == 'e')
1916 if (uses_hard_regs_p (XEXP (x, i), set))
1917 return true;
1919 else if (fmt[i] == 'E')
1921 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
1922 if (uses_hard_regs_p (XVECEXP (x, i, j), set))
1923 return true;
1926 return false;
1929 /* Return true if OP is a spilled pseudo. */
1930 static inline bool
1931 spilled_pseudo_p (rtx op)
1933 return (REG_P (op)
1934 && REGNO (op) >= FIRST_PSEUDO_REGISTER && in_mem_p (REGNO (op)));
1937 /* Return true if X is a general constant. */
1938 static inline bool
1939 general_constant_p (rtx x)
1941 return CONSTANT_P (x) && (! flag_pic || LEGITIMATE_PIC_OPERAND_P (x));
1944 static bool
1945 reg_in_class_p (rtx reg, enum reg_class cl)
1947 if (cl == NO_REGS)
1948 return get_reg_class (REGNO (reg)) == NO_REGS;
1949 return in_class_p (reg, cl, NULL);
1952 /* Return true if SET of RCLASS contains no hard regs which can be
1953 used in MODE. */
1954 static bool
1955 prohibited_class_reg_set_mode_p (enum reg_class rclass,
1956 HARD_REG_SET &set,
1957 machine_mode mode)
1959 HARD_REG_SET temp;
1961 lra_assert (hard_reg_set_subset_p (reg_class_contents[rclass], set));
1962 temp = set & ~lra_no_alloc_regs;
1963 return (hard_reg_set_subset_p
1964 (temp, ira_prohibited_class_mode_regs[rclass][mode]));
1968 /* Used to check validity info about small class input operands. It
1969 should be incremented at start of processing an insn
1970 alternative. */
1971 static unsigned int curr_small_class_check = 0;
1973 /* Update number of used inputs of class OP_CLASS for operand NOP
1974 of alternative NALT. Return true if we have more such class operands
1975 than the number of available regs. */
1976 static bool
1977 update_and_check_small_class_inputs (int nop, int nalt,
1978 enum reg_class op_class)
1980 static unsigned int small_class_check[LIM_REG_CLASSES];
1981 static int small_class_input_nums[LIM_REG_CLASSES];
1983 if (SMALL_REGISTER_CLASS_P (op_class)
1984 /* We are interesting in classes became small because of fixing
1985 some hard regs, e.g. by an user through GCC options. */
1986 && hard_reg_set_intersect_p (reg_class_contents[op_class],
1987 ira_no_alloc_regs)
1988 && (curr_static_id->operand[nop].type != OP_OUT
1989 || TEST_BIT (curr_static_id->operand[nop].early_clobber_alts, nalt)))
1991 if (small_class_check[op_class] == curr_small_class_check)
1992 small_class_input_nums[op_class]++;
1993 else
1995 small_class_check[op_class] = curr_small_class_check;
1996 small_class_input_nums[op_class] = 1;
1998 if (small_class_input_nums[op_class] > ira_class_hard_regs_num[op_class])
1999 return true;
2001 return false;
2004 /* Major function to choose the current insn alternative and what
2005 operands should be reloaded and how. If ONLY_ALTERNATIVE is not
2006 negative we should consider only this alternative. Return false if
2007 we cannot choose the alternative or find how to reload the
2008 operands. */
2009 static bool
2010 process_alt_operands (int only_alternative)
2012 bool ok_p = false;
2013 int nop, overall, nalt;
2014 int n_alternatives = curr_static_id->n_alternatives;
2015 int n_operands = curr_static_id->n_operands;
2016 /* LOSERS counts the operands that don't fit this alternative and
2017 would require loading. */
2018 int losers;
2019 int addr_losers;
2020 /* REJECT is a count of how undesirable this alternative says it is
2021 if any reloading is required. If the alternative matches exactly
2022 then REJECT is ignored, but otherwise it gets this much counted
2023 against it in addition to the reloading needed. */
2024 int reject;
2025 /* This is defined by '!' or '?' alternative constraint and added to
2026 reject. But in some cases it can be ignored. */
2027 int static_reject;
2028 int op_reject;
2029 /* The number of elements in the following array. */
2030 int early_clobbered_regs_num;
2031 /* Numbers of operands which are early clobber registers. */
2032 int early_clobbered_nops[MAX_RECOG_OPERANDS];
2033 enum reg_class curr_alt[MAX_RECOG_OPERANDS];
2034 HARD_REG_SET curr_alt_set[MAX_RECOG_OPERANDS];
2035 bool curr_alt_match_win[MAX_RECOG_OPERANDS];
2036 bool curr_alt_win[MAX_RECOG_OPERANDS];
2037 bool curr_alt_offmemok[MAX_RECOG_OPERANDS];
2038 int curr_alt_matches[MAX_RECOG_OPERANDS];
2039 /* The number of elements in the following array. */
2040 int curr_alt_dont_inherit_ops_num;
2041 /* Numbers of operands whose reload pseudos should not be inherited. */
2042 int curr_alt_dont_inherit_ops[MAX_RECOG_OPERANDS];
2043 rtx op;
2044 /* The register when the operand is a subreg of register, otherwise the
2045 operand itself. */
2046 rtx no_subreg_reg_operand[MAX_RECOG_OPERANDS];
2047 /* The register if the operand is a register or subreg of register,
2048 otherwise NULL. */
2049 rtx operand_reg[MAX_RECOG_OPERANDS];
2050 int hard_regno[MAX_RECOG_OPERANDS];
2051 machine_mode biggest_mode[MAX_RECOG_OPERANDS];
2052 int reload_nregs, reload_sum;
2053 bool costly_p;
2054 enum reg_class cl;
2056 /* Calculate some data common for all alternatives to speed up the
2057 function. */
2058 for (nop = 0; nop < n_operands; nop++)
2060 rtx reg;
2062 op = no_subreg_reg_operand[nop] = *curr_id->operand_loc[nop];
2063 /* The real hard regno of the operand after the allocation. */
2064 hard_regno[nop] = get_hard_regno (op, true);
2066 operand_reg[nop] = reg = op;
2067 biggest_mode[nop] = GET_MODE (op);
2068 if (GET_CODE (op) == SUBREG)
2070 biggest_mode[nop] = wider_subreg_mode (op);
2071 operand_reg[nop] = reg = SUBREG_REG (op);
2073 if (! REG_P (reg))
2074 operand_reg[nop] = NULL_RTX;
2075 else if (REGNO (reg) >= FIRST_PSEUDO_REGISTER
2076 || ((int) REGNO (reg)
2077 == lra_get_elimination_hard_regno (REGNO (reg))))
2078 no_subreg_reg_operand[nop] = reg;
2079 else
2080 operand_reg[nop] = no_subreg_reg_operand[nop]
2081 /* Just use natural mode for elimination result. It should
2082 be enough for extra constraints hooks. */
2083 = regno_reg_rtx[hard_regno[nop]];
2086 /* The constraints are made of several alternatives. Each operand's
2087 constraint looks like foo,bar,... with commas separating the
2088 alternatives. The first alternatives for all operands go
2089 together, the second alternatives go together, etc.
2091 First loop over alternatives. */
2092 alternative_mask preferred = curr_id->preferred_alternatives;
2093 if (only_alternative >= 0)
2094 preferred &= ALTERNATIVE_BIT (only_alternative);
2096 for (nalt = 0; nalt < n_alternatives; nalt++)
2098 /* Loop over operands for one constraint alternative. */
2099 if (!TEST_BIT (preferred, nalt))
2100 continue;
2102 bool matching_early_clobber[MAX_RECOG_OPERANDS];
2103 curr_small_class_check++;
2104 overall = losers = addr_losers = 0;
2105 static_reject = reject = reload_nregs = reload_sum = 0;
2106 for (nop = 0; nop < n_operands; nop++)
2108 int inc = (curr_static_id
2109 ->operand_alternative[nalt * n_operands + nop].reject);
2110 if (lra_dump_file != NULL && inc != 0)
2111 fprintf (lra_dump_file,
2112 " Staticly defined alt reject+=%d\n", inc);
2113 static_reject += inc;
2114 matching_early_clobber[nop] = 0;
2116 reject += static_reject;
2117 early_clobbered_regs_num = 0;
2119 for (nop = 0; nop < n_operands; nop++)
2121 const char *p;
2122 char *end;
2123 int len, c, m, i, opalt_num, this_alternative_matches;
2124 bool win, did_match, offmemok, early_clobber_p;
2125 /* false => this operand can be reloaded somehow for this
2126 alternative. */
2127 bool badop;
2128 /* true => this operand can be reloaded if the alternative
2129 allows regs. */
2130 bool winreg;
2131 /* True if a constant forced into memory would be OK for
2132 this operand. */
2133 bool constmemok;
2134 enum reg_class this_alternative, this_costly_alternative;
2135 HARD_REG_SET this_alternative_set, this_costly_alternative_set;
2136 bool this_alternative_match_win, this_alternative_win;
2137 bool this_alternative_offmemok;
2138 bool scratch_p;
2139 machine_mode mode;
2140 enum constraint_num cn;
2142 opalt_num = nalt * n_operands + nop;
2143 if (curr_static_id->operand_alternative[opalt_num].anything_ok)
2145 /* Fast track for no constraints at all. */
2146 curr_alt[nop] = NO_REGS;
2147 CLEAR_HARD_REG_SET (curr_alt_set[nop]);
2148 curr_alt_win[nop] = true;
2149 curr_alt_match_win[nop] = false;
2150 curr_alt_offmemok[nop] = false;
2151 curr_alt_matches[nop] = -1;
2152 continue;
2155 op = no_subreg_reg_operand[nop];
2156 mode = curr_operand_mode[nop];
2158 win = did_match = winreg = offmemok = constmemok = false;
2159 badop = true;
2161 early_clobber_p = false;
2162 p = curr_static_id->operand_alternative[opalt_num].constraint;
2164 this_costly_alternative = this_alternative = NO_REGS;
2165 /* We update set of possible hard regs besides its class
2166 because reg class might be inaccurate. For example,
2167 union of LO_REGS (l), HI_REGS(h), and STACK_REG(k) in ARM
2168 is translated in HI_REGS because classes are merged by
2169 pairs and there is no accurate intermediate class. */
2170 CLEAR_HARD_REG_SET (this_alternative_set);
2171 CLEAR_HARD_REG_SET (this_costly_alternative_set);
2172 this_alternative_win = false;
2173 this_alternative_match_win = false;
2174 this_alternative_offmemok = false;
2175 this_alternative_matches = -1;
2177 /* An empty constraint should be excluded by the fast
2178 track. */
2179 lra_assert (*p != 0 && *p != ',');
2181 op_reject = 0;
2182 /* Scan this alternative's specs for this operand; set WIN
2183 if the operand fits any letter in this alternative.
2184 Otherwise, clear BADOP if this operand could fit some
2185 letter after reloads, or set WINREG if this operand could
2186 fit after reloads provided the constraint allows some
2187 registers. */
2188 costly_p = false;
2191 switch ((c = *p, len = CONSTRAINT_LEN (c, p)), c)
2193 case '\0':
2194 len = 0;
2195 break;
2196 case ',':
2197 c = '\0';
2198 break;
2200 case '&':
2201 early_clobber_p = true;
2202 break;
2204 case '$':
2205 op_reject += LRA_MAX_REJECT;
2206 break;
2207 case '^':
2208 op_reject += LRA_LOSER_COST_FACTOR;
2209 break;
2211 case '#':
2212 /* Ignore rest of this alternative. */
2213 c = '\0';
2214 break;
2216 case '0': case '1': case '2': case '3': case '4':
2217 case '5': case '6': case '7': case '8': case '9':
2219 int m_hregno;
2220 bool match_p;
2222 m = strtoul (p, &end, 10);
2223 p = end;
2224 len = 0;
2225 lra_assert (nop > m);
2227 /* Reject matches if we don't know which operand is
2228 bigger. This situation would arguably be a bug in
2229 an .md pattern, but could also occur in a user asm. */
2230 if (!ordered_p (GET_MODE_SIZE (biggest_mode[m]),
2231 GET_MODE_SIZE (biggest_mode[nop])))
2232 break;
2234 /* Don't match wrong asm insn operands for proper
2235 diagnostic later. */
2236 if (INSN_CODE (curr_insn) < 0
2237 && (curr_operand_mode[m] == BLKmode
2238 || curr_operand_mode[nop] == BLKmode)
2239 && curr_operand_mode[m] != curr_operand_mode[nop])
2240 break;
2242 m_hregno = get_hard_regno (*curr_id->operand_loc[m], false);
2243 /* We are supposed to match a previous operand.
2244 If we do, we win if that one did. If we do
2245 not, count both of the operands as losers.
2246 (This is too conservative, since most of the
2247 time only a single reload insn will be needed
2248 to make the two operands win. As a result,
2249 this alternative may be rejected when it is
2250 actually desirable.) */
2251 match_p = false;
2252 if (operands_match_p (*curr_id->operand_loc[nop],
2253 *curr_id->operand_loc[m], m_hregno))
2255 /* We should reject matching of an early
2256 clobber operand if the matching operand is
2257 not dying in the insn. */
2258 if (!TEST_BIT (curr_static_id->operand[m]
2259 .early_clobber_alts, nalt)
2260 || operand_reg[nop] == NULL_RTX
2261 || (find_regno_note (curr_insn, REG_DEAD,
2262 REGNO (op))
2263 || REGNO (op) == REGNO (operand_reg[m])))
2264 match_p = true;
2266 if (match_p)
2268 /* If we are matching a non-offsettable
2269 address where an offsettable address was
2270 expected, then we must reject this
2271 combination, because we can't reload
2272 it. */
2273 if (curr_alt_offmemok[m]
2274 && MEM_P (*curr_id->operand_loc[m])
2275 && curr_alt[m] == NO_REGS && ! curr_alt_win[m])
2276 continue;
2278 else
2280 /* If the operands do not match and one
2281 operand is INOUT, we can not match them.
2282 Try other possibilities, e.g. other
2283 alternatives or commutative operand
2284 exchange. */
2285 if (curr_static_id->operand[nop].type == OP_INOUT
2286 || curr_static_id->operand[m].type == OP_INOUT)
2287 break;
2288 /* Operands don't match. If the operands are
2289 different user defined explicit hard
2290 registers, then we cannot make them match
2291 when one is early clobber operand. */
2292 if ((REG_P (*curr_id->operand_loc[nop])
2293 || SUBREG_P (*curr_id->operand_loc[nop]))
2294 && (REG_P (*curr_id->operand_loc[m])
2295 || SUBREG_P (*curr_id->operand_loc[m])))
2297 rtx nop_reg = *curr_id->operand_loc[nop];
2298 if (SUBREG_P (nop_reg))
2299 nop_reg = SUBREG_REG (nop_reg);
2300 rtx m_reg = *curr_id->operand_loc[m];
2301 if (SUBREG_P (m_reg))
2302 m_reg = SUBREG_REG (m_reg);
2304 if (REG_P (nop_reg)
2305 && HARD_REGISTER_P (nop_reg)
2306 && REG_USERVAR_P (nop_reg)
2307 && REG_P (m_reg)
2308 && HARD_REGISTER_P (m_reg)
2309 && REG_USERVAR_P (m_reg))
2311 int i;
2313 for (i = 0; i < early_clobbered_regs_num; i++)
2314 if (m == early_clobbered_nops[i])
2315 break;
2316 if (i < early_clobbered_regs_num
2317 || early_clobber_p)
2318 break;
2321 /* Both operands must allow a reload register,
2322 otherwise we cannot make them match. */
2323 if (curr_alt[m] == NO_REGS)
2324 break;
2325 /* Retroactively mark the operand we had to
2326 match as a loser, if it wasn't already and
2327 it wasn't matched to a register constraint
2328 (e.g it might be matched by memory). */
2329 if (curr_alt_win[m]
2330 && (operand_reg[m] == NULL_RTX
2331 || hard_regno[m] < 0))
2333 losers++;
2334 reload_nregs
2335 += (ira_reg_class_max_nregs[curr_alt[m]]
2336 [GET_MODE (*curr_id->operand_loc[m])]);
2339 /* Prefer matching earlyclobber alternative as
2340 it results in less hard regs required for
2341 the insn than a non-matching earlyclobber
2342 alternative. */
2343 if (TEST_BIT (curr_static_id->operand[m]
2344 .early_clobber_alts, nalt))
2346 if (lra_dump_file != NULL)
2347 fprintf
2348 (lra_dump_file,
2349 " %d Matching earlyclobber alt:"
2350 " reject--\n",
2351 nop);
2352 if (!matching_early_clobber[m])
2354 reject--;
2355 matching_early_clobber[m] = 1;
2358 /* Otherwise we prefer no matching
2359 alternatives because it gives more freedom
2360 in RA. */
2361 else if (operand_reg[nop] == NULL_RTX
2362 || (find_regno_note (curr_insn, REG_DEAD,
2363 REGNO (operand_reg[nop]))
2364 == NULL_RTX))
2366 if (lra_dump_file != NULL)
2367 fprintf
2368 (lra_dump_file,
2369 " %d Matching alt: reject+=2\n",
2370 nop);
2371 reject += 2;
2374 /* If we have to reload this operand and some
2375 previous operand also had to match the same
2376 thing as this operand, we don't know how to do
2377 that. */
2378 if (!match_p || !curr_alt_win[m])
2380 for (i = 0; i < nop; i++)
2381 if (curr_alt_matches[i] == m)
2382 break;
2383 if (i < nop)
2384 break;
2386 else
2387 did_match = true;
2389 this_alternative_matches = m;
2390 /* This can be fixed with reloads if the operand
2391 we are supposed to match can be fixed with
2392 reloads. */
2393 badop = false;
2394 this_alternative = curr_alt[m];
2395 this_alternative_set = curr_alt_set[m];
2396 winreg = this_alternative != NO_REGS;
2397 break;
2400 case 'g':
2401 if (MEM_P (op)
2402 || general_constant_p (op)
2403 || spilled_pseudo_p (op))
2404 win = true;
2405 cl = GENERAL_REGS;
2406 goto reg;
2408 default:
2409 cn = lookup_constraint (p);
2410 switch (get_constraint_type (cn))
2412 case CT_REGISTER:
2413 cl = reg_class_for_constraint (cn);
2414 if (cl != NO_REGS)
2415 goto reg;
2416 break;
2418 case CT_CONST_INT:
2419 if (CONST_INT_P (op)
2420 && insn_const_int_ok_for_constraint (INTVAL (op), cn))
2421 win = true;
2422 break;
2424 case CT_MEMORY:
2425 case CT_RELAXED_MEMORY:
2426 if (MEM_P (op)
2427 && satisfies_memory_constraint_p (op, cn))
2428 win = true;
2429 else if (spilled_pseudo_p (op))
2430 win = true;
2432 /* If we didn't already win, we can reload constants
2433 via force_const_mem or put the pseudo value into
2434 memory, or make other memory by reloading the
2435 address like for 'o'. */
2436 if (CONST_POOL_OK_P (mode, op)
2437 || MEM_P (op) || REG_P (op)
2438 /* We can restore the equiv insn by a
2439 reload. */
2440 || equiv_substition_p[nop])
2441 badop = false;
2442 constmemok = true;
2443 offmemok = true;
2444 break;
2446 case CT_ADDRESS:
2447 /* An asm operand with an address constraint
2448 that doesn't satisfy address_operand has
2449 is_address cleared, so that we don't try to
2450 make a non-address fit. */
2451 if (!curr_static_id->operand[nop].is_address)
2452 break;
2453 /* If we didn't already win, we can reload the address
2454 into a base register. */
2455 if (satisfies_address_constraint_p (op, cn))
2456 win = true;
2457 cl = base_reg_class (VOIDmode, ADDR_SPACE_GENERIC,
2458 ADDRESS, SCRATCH);
2459 badop = false;
2460 goto reg;
2462 case CT_FIXED_FORM:
2463 if (constraint_satisfied_p (op, cn))
2464 win = true;
2465 break;
2467 case CT_SPECIAL_MEMORY:
2468 if (satisfies_memory_constraint_p (op, cn))
2469 win = true;
2470 else if (spilled_pseudo_p (op))
2471 win = true;
2472 break;
2474 break;
2476 reg:
2477 if (mode == BLKmode)
2478 break;
2479 this_alternative = reg_class_subunion[this_alternative][cl];
2480 this_alternative_set |= reg_class_contents[cl];
2481 if (costly_p)
2483 this_costly_alternative
2484 = reg_class_subunion[this_costly_alternative][cl];
2485 this_costly_alternative_set |= reg_class_contents[cl];
2487 winreg = true;
2488 if (REG_P (op))
2490 if (hard_regno[nop] >= 0
2491 && in_hard_reg_set_p (this_alternative_set,
2492 mode, hard_regno[nop]))
2493 win = true;
2494 else if (hard_regno[nop] < 0
2495 && in_class_p (op, this_alternative, NULL))
2496 win = true;
2498 break;
2500 if (c != ' ' && c != '\t')
2501 costly_p = c == '*';
2503 while ((p += len), c);
2505 scratch_p = (operand_reg[nop] != NULL_RTX
2506 && ira_former_scratch_p (REGNO (operand_reg[nop])));
2507 /* Record which operands fit this alternative. */
2508 if (win)
2510 this_alternative_win = true;
2511 if (operand_reg[nop] != NULL_RTX)
2513 if (hard_regno[nop] >= 0)
2515 if (in_hard_reg_set_p (this_costly_alternative_set,
2516 mode, hard_regno[nop]))
2518 if (lra_dump_file != NULL)
2519 fprintf (lra_dump_file,
2520 " %d Costly set: reject++\n",
2521 nop);
2522 reject++;
2525 else
2527 /* Prefer won reg to spilled pseudo under other
2528 equal conditions for possibe inheritance. */
2529 if (! scratch_p)
2531 if (lra_dump_file != NULL)
2532 fprintf
2533 (lra_dump_file,
2534 " %d Non pseudo reload: reject++\n",
2535 nop);
2536 reject++;
2538 if (in_class_p (operand_reg[nop],
2539 this_costly_alternative, NULL))
2541 if (lra_dump_file != NULL)
2542 fprintf
2543 (lra_dump_file,
2544 " %d Non pseudo costly reload:"
2545 " reject++\n",
2546 nop);
2547 reject++;
2550 /* We simulate the behavior of old reload here.
2551 Although scratches need hard registers and it
2552 might result in spilling other pseudos, no reload
2553 insns are generated for the scratches. So it
2554 might cost something but probably less than old
2555 reload pass believes. */
2556 if (scratch_p)
2558 if (lra_dump_file != NULL)
2559 fprintf (lra_dump_file,
2560 " %d Scratch win: reject+=2\n",
2561 nop);
2562 reject += 2;
2566 else if (did_match)
2567 this_alternative_match_win = true;
2568 else
2570 int const_to_mem = 0;
2571 bool no_regs_p;
2573 reject += op_reject;
2574 /* Never do output reload of stack pointer. It makes
2575 impossible to do elimination when SP is changed in
2576 RTL. */
2577 if (op == stack_pointer_rtx && ! frame_pointer_needed
2578 && curr_static_id->operand[nop].type != OP_IN)
2579 goto fail;
2581 /* If this alternative asks for a specific reg class, see if there
2582 is at least one allocatable register in that class. */
2583 no_regs_p
2584 = (this_alternative == NO_REGS
2585 || (hard_reg_set_subset_p
2586 (reg_class_contents[this_alternative],
2587 lra_no_alloc_regs)));
2589 /* For asms, verify that the class for this alternative is possible
2590 for the mode that is specified. */
2591 if (!no_regs_p && INSN_CODE (curr_insn) < 0)
2593 int i;
2594 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
2595 if (targetm.hard_regno_mode_ok (i, mode)
2596 && in_hard_reg_set_p (reg_class_contents[this_alternative],
2597 mode, i))
2598 break;
2599 if (i == FIRST_PSEUDO_REGISTER)
2600 winreg = false;
2603 /* If this operand accepts a register, and if the
2604 register class has at least one allocatable register,
2605 then this operand can be reloaded. */
2606 if (winreg && !no_regs_p)
2607 badop = false;
2609 if (badop)
2611 if (lra_dump_file != NULL)
2612 fprintf (lra_dump_file,
2613 " alt=%d: Bad operand -- refuse\n",
2614 nalt);
2615 goto fail;
2618 if (this_alternative != NO_REGS)
2620 HARD_REG_SET available_regs
2621 = (reg_class_contents[this_alternative]
2622 & ~((ira_prohibited_class_mode_regs
2623 [this_alternative][mode])
2624 | lra_no_alloc_regs));
2625 if (hard_reg_set_empty_p (available_regs))
2627 /* There are no hard regs holding a value of given
2628 mode. */
2629 if (offmemok)
2631 this_alternative = NO_REGS;
2632 if (lra_dump_file != NULL)
2633 fprintf (lra_dump_file,
2634 " %d Using memory because of"
2635 " a bad mode: reject+=2\n",
2636 nop);
2637 reject += 2;
2639 else
2641 if (lra_dump_file != NULL)
2642 fprintf (lra_dump_file,
2643 " alt=%d: Wrong mode -- refuse\n",
2644 nalt);
2645 goto fail;
2650 /* If not assigned pseudo has a class which a subset of
2651 required reg class, it is a less costly alternative
2652 as the pseudo still can get a hard reg of necessary
2653 class. */
2654 if (! no_regs_p && REG_P (op) && hard_regno[nop] < 0
2655 && (cl = get_reg_class (REGNO (op))) != NO_REGS
2656 && ira_class_subset_p[this_alternative][cl])
2658 if (lra_dump_file != NULL)
2659 fprintf
2660 (lra_dump_file,
2661 " %d Super set class reg: reject-=3\n", nop);
2662 reject -= 3;
2665 this_alternative_offmemok = offmemok;
2666 if (this_costly_alternative != NO_REGS)
2668 if (lra_dump_file != NULL)
2669 fprintf (lra_dump_file,
2670 " %d Costly loser: reject++\n", nop);
2671 reject++;
2673 /* If the operand is dying, has a matching constraint,
2674 and satisfies constraints of the matched operand
2675 which failed to satisfy the own constraints, most probably
2676 the reload for this operand will be gone. */
2677 if (this_alternative_matches >= 0
2678 && !curr_alt_win[this_alternative_matches]
2679 && REG_P (op)
2680 && find_regno_note (curr_insn, REG_DEAD, REGNO (op))
2681 && (hard_regno[nop] >= 0
2682 ? in_hard_reg_set_p (this_alternative_set,
2683 mode, hard_regno[nop])
2684 : in_class_p (op, this_alternative, NULL)))
2686 if (lra_dump_file != NULL)
2687 fprintf
2688 (lra_dump_file,
2689 " %d Dying matched operand reload: reject++\n",
2690 nop);
2691 reject++;
2693 else
2695 /* Strict_low_part requires to reload the register
2696 not the sub-register. In this case we should
2697 check that a final reload hard reg can hold the
2698 value mode. */
2699 if (curr_static_id->operand[nop].strict_low
2700 && REG_P (op)
2701 && hard_regno[nop] < 0
2702 && GET_CODE (*curr_id->operand_loc[nop]) == SUBREG
2703 && ira_class_hard_regs_num[this_alternative] > 0
2704 && (!targetm.hard_regno_mode_ok
2705 (ira_class_hard_regs[this_alternative][0],
2706 GET_MODE (*curr_id->operand_loc[nop]))))
2708 if (lra_dump_file != NULL)
2709 fprintf
2710 (lra_dump_file,
2711 " alt=%d: Strict low subreg reload -- refuse\n",
2712 nalt);
2713 goto fail;
2715 losers++;
2717 if (operand_reg[nop] != NULL_RTX
2718 /* Output operands and matched input operands are
2719 not inherited. The following conditions do not
2720 exactly describe the previous statement but they
2721 are pretty close. */
2722 && curr_static_id->operand[nop].type != OP_OUT
2723 && (this_alternative_matches < 0
2724 || curr_static_id->operand[nop].type != OP_IN))
2726 int last_reload = (lra_reg_info[ORIGINAL_REGNO
2727 (operand_reg[nop])]
2728 .last_reload);
2730 /* The value of reload_sum has sense only if we
2731 process insns in their order. It happens only on
2732 the first constraints sub-pass when we do most of
2733 reload work. */
2734 if (lra_constraint_iter == 1 && last_reload > bb_reload_num)
2735 reload_sum += last_reload - bb_reload_num;
2737 /* If this is a constant that is reloaded into the
2738 desired class by copying it to memory first, count
2739 that as another reload. This is consistent with
2740 other code and is required to avoid choosing another
2741 alternative when the constant is moved into memory.
2742 Note that the test here is precisely the same as in
2743 the code below that calls force_const_mem. */
2744 if (CONST_POOL_OK_P (mode, op)
2745 && ((targetm.preferred_reload_class
2746 (op, this_alternative) == NO_REGS)
2747 || no_input_reloads_p))
2749 const_to_mem = 1;
2750 if (! no_regs_p)
2751 losers++;
2754 /* Alternative loses if it requires a type of reload not
2755 permitted for this insn. We can always reload
2756 objects with a REG_UNUSED note. */
2757 if ((curr_static_id->operand[nop].type != OP_IN
2758 && no_output_reloads_p
2759 && ! find_reg_note (curr_insn, REG_UNUSED, op))
2760 || (curr_static_id->operand[nop].type != OP_OUT
2761 && no_input_reloads_p && ! const_to_mem)
2762 || (this_alternative_matches >= 0
2763 && (no_input_reloads_p
2764 || (no_output_reloads_p
2765 && (curr_static_id->operand
2766 [this_alternative_matches].type != OP_IN)
2767 && ! find_reg_note (curr_insn, REG_UNUSED,
2768 no_subreg_reg_operand
2769 [this_alternative_matches])))))
2771 if (lra_dump_file != NULL)
2772 fprintf
2773 (lra_dump_file,
2774 " alt=%d: No input/output reload -- refuse\n",
2775 nalt);
2776 goto fail;
2779 /* Alternative loses if it required class pseudo cannot
2780 hold value of required mode. Such insns can be
2781 described by insn definitions with mode iterators. */
2782 if (GET_MODE (*curr_id->operand_loc[nop]) != VOIDmode
2783 && ! hard_reg_set_empty_p (this_alternative_set)
2784 /* It is common practice for constraints to use a
2785 class which does not have actually enough regs to
2786 hold the value (e.g. x86 AREG for mode requiring
2787 more one general reg). Therefore we have 2
2788 conditions to check that the reload pseudo cannot
2789 hold the mode value. */
2790 && (!targetm.hard_regno_mode_ok
2791 (ira_class_hard_regs[this_alternative][0],
2792 GET_MODE (*curr_id->operand_loc[nop])))
2793 /* The above condition is not enough as the first
2794 reg in ira_class_hard_regs can be not aligned for
2795 multi-words mode values. */
2796 && (prohibited_class_reg_set_mode_p
2797 (this_alternative, this_alternative_set,
2798 GET_MODE (*curr_id->operand_loc[nop]))))
2800 if (lra_dump_file != NULL)
2801 fprintf (lra_dump_file,
2802 " alt=%d: reload pseudo for op %d "
2803 "cannot hold the mode value -- refuse\n",
2804 nalt, nop);
2805 goto fail;
2808 /* Check strong discouragement of reload of non-constant
2809 into class THIS_ALTERNATIVE. */
2810 if (! CONSTANT_P (op) && ! no_regs_p
2811 && (targetm.preferred_reload_class
2812 (op, this_alternative) == NO_REGS
2813 || (curr_static_id->operand[nop].type == OP_OUT
2814 && (targetm.preferred_output_reload_class
2815 (op, this_alternative) == NO_REGS))))
2817 if (offmemok && REG_P (op))
2819 if (lra_dump_file != NULL)
2820 fprintf
2821 (lra_dump_file,
2822 " %d Spill pseudo into memory: reject+=3\n",
2823 nop);
2824 reject += 3;
2826 else
2828 if (lra_dump_file != NULL)
2829 fprintf
2830 (lra_dump_file,
2831 " %d Non-prefered reload: reject+=%d\n",
2832 nop, LRA_MAX_REJECT);
2833 reject += LRA_MAX_REJECT;
2837 if (! (MEM_P (op) && offmemok)
2838 && ! (const_to_mem && constmemok))
2840 /* We prefer to reload pseudos over reloading other
2841 things, since such reloads may be able to be
2842 eliminated later. So bump REJECT in other cases.
2843 Don't do this in the case where we are forcing a
2844 constant into memory and it will then win since
2845 we don't want to have a different alternative
2846 match then. */
2847 if (! (REG_P (op) && REGNO (op) >= FIRST_PSEUDO_REGISTER))
2849 if (lra_dump_file != NULL)
2850 fprintf
2851 (lra_dump_file,
2852 " %d Non-pseudo reload: reject+=2\n",
2853 nop);
2854 reject += 2;
2857 if (! no_regs_p)
2858 reload_nregs
2859 += ira_reg_class_max_nregs[this_alternative][mode];
2861 if (SMALL_REGISTER_CLASS_P (this_alternative))
2863 if (lra_dump_file != NULL)
2864 fprintf
2865 (lra_dump_file,
2866 " %d Small class reload: reject+=%d\n",
2867 nop, LRA_LOSER_COST_FACTOR / 2);
2868 reject += LRA_LOSER_COST_FACTOR / 2;
2872 /* We are trying to spill pseudo into memory. It is
2873 usually more costly than moving to a hard register
2874 although it might takes the same number of
2875 reloads.
2877 Non-pseudo spill may happen also. Suppose a target allows both
2878 register and memory in the operand constraint alternatives,
2879 then it's typical that an eliminable register has a substition
2880 of "base + offset" which can either be reloaded by a simple
2881 "new_reg <= base + offset" which will match the register
2882 constraint, or a similar reg addition followed by further spill
2883 to and reload from memory which will match the memory
2884 constraint, but this memory spill will be much more costly
2885 usually.
2887 Code below increases the reject for both pseudo and non-pseudo
2888 spill. */
2889 if (no_regs_p
2890 && !(MEM_P (op) && offmemok)
2891 && !(REG_P (op) && hard_regno[nop] < 0))
2893 if (lra_dump_file != NULL)
2894 fprintf
2895 (lra_dump_file,
2896 " %d Spill %spseudo into memory: reject+=3\n",
2897 nop, REG_P (op) ? "" : "Non-");
2898 reject += 3;
2899 if (VECTOR_MODE_P (mode))
2901 /* Spilling vectors into memory is usually more
2902 costly as they contain big values. */
2903 if (lra_dump_file != NULL)
2904 fprintf
2905 (lra_dump_file,
2906 " %d Spill vector pseudo: reject+=2\n",
2907 nop);
2908 reject += 2;
2912 /* When we use an operand requiring memory in given
2913 alternative, the insn should write *and* read the
2914 value to/from memory it is costly in comparison with
2915 an insn alternative which does not use memory
2916 (e.g. register or immediate operand). We exclude
2917 memory operand for such case as we can satisfy the
2918 memory constraints by reloading address. */
2919 if (no_regs_p && offmemok && !MEM_P (op))
2921 if (lra_dump_file != NULL)
2922 fprintf
2923 (lra_dump_file,
2924 " Using memory insn operand %d: reject+=3\n",
2925 nop);
2926 reject += 3;
2929 /* If reload requires moving value through secondary
2930 memory, it will need one more insn at least. */
2931 if (this_alternative != NO_REGS
2932 && REG_P (op) && (cl = get_reg_class (REGNO (op))) != NO_REGS
2933 && ((curr_static_id->operand[nop].type != OP_OUT
2934 && targetm.secondary_memory_needed (GET_MODE (op), cl,
2935 this_alternative))
2936 || (curr_static_id->operand[nop].type != OP_IN
2937 && (targetm.secondary_memory_needed
2938 (GET_MODE (op), this_alternative, cl)))))
2939 losers++;
2941 if (MEM_P (op) && offmemok)
2942 addr_losers++;
2943 else
2945 /* Input reloads can be inherited more often than
2946 output reloads can be removed, so penalize output
2947 reloads. */
2948 if (!REG_P (op) || curr_static_id->operand[nop].type != OP_IN)
2950 if (lra_dump_file != NULL)
2951 fprintf
2952 (lra_dump_file,
2953 " %d Non input pseudo reload: reject++\n",
2954 nop);
2955 reject++;
2958 if (curr_static_id->operand[nop].type == OP_INOUT)
2960 if (lra_dump_file != NULL)
2961 fprintf
2962 (lra_dump_file,
2963 " %d Input/Output reload: reject+=%d\n",
2964 nop, LRA_LOSER_COST_FACTOR);
2965 reject += LRA_LOSER_COST_FACTOR;
2970 if (early_clobber_p && ! scratch_p)
2972 if (lra_dump_file != NULL)
2973 fprintf (lra_dump_file,
2974 " %d Early clobber: reject++\n", nop);
2975 reject++;
2977 /* ??? We check early clobbers after processing all operands
2978 (see loop below) and there we update the costs more.
2979 Should we update the cost (may be approximately) here
2980 because of early clobber register reloads or it is a rare
2981 or non-important thing to be worth to do it. */
2982 overall = (losers * LRA_LOSER_COST_FACTOR + reject
2983 - (addr_losers == losers ? static_reject : 0));
2984 if ((best_losers == 0 || losers != 0) && best_overall < overall)
2986 if (lra_dump_file != NULL)
2987 fprintf (lra_dump_file,
2988 " alt=%d,overall=%d,losers=%d -- refuse\n",
2989 nalt, overall, losers);
2990 goto fail;
2993 if (update_and_check_small_class_inputs (nop, nalt,
2994 this_alternative))
2996 if (lra_dump_file != NULL)
2997 fprintf (lra_dump_file,
2998 " alt=%d, not enough small class regs -- refuse\n",
2999 nalt);
3000 goto fail;
3002 curr_alt[nop] = this_alternative;
3003 curr_alt_set[nop] = this_alternative_set;
3004 curr_alt_win[nop] = this_alternative_win;
3005 curr_alt_match_win[nop] = this_alternative_match_win;
3006 curr_alt_offmemok[nop] = this_alternative_offmemok;
3007 curr_alt_matches[nop] = this_alternative_matches;
3009 if (this_alternative_matches >= 0
3010 && !did_match && !this_alternative_win)
3011 curr_alt_win[this_alternative_matches] = false;
3013 if (early_clobber_p && operand_reg[nop] != NULL_RTX)
3014 early_clobbered_nops[early_clobbered_regs_num++] = nop;
3017 if (curr_insn_set != NULL_RTX && n_operands == 2
3018 /* Prevent processing non-move insns. */
3019 && (GET_CODE (SET_SRC (curr_insn_set)) == SUBREG
3020 || SET_SRC (curr_insn_set) == no_subreg_reg_operand[1])
3021 && ((! curr_alt_win[0] && ! curr_alt_win[1]
3022 && REG_P (no_subreg_reg_operand[0])
3023 && REG_P (no_subreg_reg_operand[1])
3024 && (reg_in_class_p (no_subreg_reg_operand[0], curr_alt[1])
3025 || reg_in_class_p (no_subreg_reg_operand[1], curr_alt[0])))
3026 || (! curr_alt_win[0] && curr_alt_win[1]
3027 && REG_P (no_subreg_reg_operand[1])
3028 /* Check that we reload memory not the memory
3029 address. */
3030 && ! (curr_alt_offmemok[0]
3031 && MEM_P (no_subreg_reg_operand[0]))
3032 && reg_in_class_p (no_subreg_reg_operand[1], curr_alt[0]))
3033 || (curr_alt_win[0] && ! curr_alt_win[1]
3034 && REG_P (no_subreg_reg_operand[0])
3035 /* Check that we reload memory not the memory
3036 address. */
3037 && ! (curr_alt_offmemok[1]
3038 && MEM_P (no_subreg_reg_operand[1]))
3039 && reg_in_class_p (no_subreg_reg_operand[0], curr_alt[1])
3040 && (! CONST_POOL_OK_P (curr_operand_mode[1],
3041 no_subreg_reg_operand[1])
3042 || (targetm.preferred_reload_class
3043 (no_subreg_reg_operand[1],
3044 (enum reg_class) curr_alt[1]) != NO_REGS))
3045 /* If it is a result of recent elimination in move
3046 insn we can transform it into an add still by
3047 using this alternative. */
3048 && GET_CODE (no_subreg_reg_operand[1]) != PLUS
3049 /* Likewise if the source has been replaced with an
3050 equivalent value. This only happens once -- the reload
3051 will use the equivalent value instead of the register it
3052 replaces -- so there should be no danger of cycling. */
3053 && !equiv_substition_p[1])))
3055 /* We have a move insn and a new reload insn will be similar
3056 to the current insn. We should avoid such situation as
3057 it results in LRA cycling. */
3058 if (lra_dump_file != NULL)
3059 fprintf (lra_dump_file,
3060 " Cycle danger: overall += LRA_MAX_REJECT\n");
3061 overall += LRA_MAX_REJECT;
3063 ok_p = true;
3064 curr_alt_dont_inherit_ops_num = 0;
3065 for (nop = 0; nop < early_clobbered_regs_num; nop++)
3067 int i, j, clobbered_hard_regno, first_conflict_j, last_conflict_j;
3068 HARD_REG_SET temp_set;
3070 i = early_clobbered_nops[nop];
3071 if ((! curr_alt_win[i] && ! curr_alt_match_win[i])
3072 || hard_regno[i] < 0)
3073 continue;
3074 lra_assert (operand_reg[i] != NULL_RTX);
3075 clobbered_hard_regno = hard_regno[i];
3076 CLEAR_HARD_REG_SET (temp_set);
3077 add_to_hard_reg_set (&temp_set, biggest_mode[i], clobbered_hard_regno);
3078 first_conflict_j = last_conflict_j = -1;
3079 for (j = 0; j < n_operands; j++)
3080 if (j == i
3081 /* We don't want process insides of match_operator and
3082 match_parallel because otherwise we would process
3083 their operands once again generating a wrong
3084 code. */
3085 || curr_static_id->operand[j].is_operator)
3086 continue;
3087 else if ((curr_alt_matches[j] == i && curr_alt_match_win[j])
3088 || (curr_alt_matches[i] == j && curr_alt_match_win[i]))
3089 continue;
3090 /* If we don't reload j-th operand, check conflicts. */
3091 else if ((curr_alt_win[j] || curr_alt_match_win[j])
3092 && uses_hard_regs_p (*curr_id->operand_loc[j], temp_set))
3094 if (first_conflict_j < 0)
3095 first_conflict_j = j;
3096 last_conflict_j = j;
3097 /* Both the earlyclobber operand and conflicting operand
3098 cannot both be user defined hard registers. */
3099 if (HARD_REGISTER_P (operand_reg[i])
3100 && REG_USERVAR_P (operand_reg[i])
3101 && operand_reg[j] != NULL_RTX
3102 && HARD_REGISTER_P (operand_reg[j])
3103 && REG_USERVAR_P (operand_reg[j]))
3105 /* For asm, let curr_insn_transform diagnose it. */
3106 if (INSN_CODE (curr_insn) < 0)
3107 return false;
3108 fatal_insn ("unable to generate reloads for "
3109 "impossible constraints:", curr_insn);
3112 if (last_conflict_j < 0)
3113 continue;
3115 /* If an earlyclobber operand conflicts with another non-matching
3116 operand (ie, they have been assigned the same hard register),
3117 then it is better to reload the other operand, as there may
3118 exist yet another operand with a matching constraint associated
3119 with the earlyclobber operand. However, if one of the operands
3120 is an explicit use of a hard register, then we must reload the
3121 other non-hard register operand. */
3122 if (HARD_REGISTER_P (operand_reg[i])
3123 || (first_conflict_j == last_conflict_j
3124 && operand_reg[last_conflict_j] != NULL_RTX
3125 && !curr_alt_match_win[last_conflict_j]
3126 && !HARD_REGISTER_P (operand_reg[last_conflict_j])))
3128 curr_alt_win[last_conflict_j] = false;
3129 curr_alt_dont_inherit_ops[curr_alt_dont_inherit_ops_num++]
3130 = last_conflict_j;
3131 losers++;
3132 if (lra_dump_file != NULL)
3133 fprintf
3134 (lra_dump_file,
3135 " %d Conflict early clobber reload: reject--\n",
3138 else
3140 /* We need to reload early clobbered register and the
3141 matched registers. */
3142 for (j = 0; j < n_operands; j++)
3143 if (curr_alt_matches[j] == i)
3145 curr_alt_match_win[j] = false;
3146 losers++;
3147 overall += LRA_LOSER_COST_FACTOR;
3149 if (! curr_alt_match_win[i])
3150 curr_alt_dont_inherit_ops[curr_alt_dont_inherit_ops_num++] = i;
3151 else
3153 /* Remember pseudos used for match reloads are never
3154 inherited. */
3155 lra_assert (curr_alt_matches[i] >= 0);
3156 curr_alt_win[curr_alt_matches[i]] = false;
3158 curr_alt_win[i] = curr_alt_match_win[i] = false;
3159 losers++;
3160 if (lra_dump_file != NULL)
3161 fprintf
3162 (lra_dump_file,
3163 " %d Matched conflict early clobber reloads: "
3164 "reject--\n",
3167 /* Early clobber was already reflected in REJECT. */
3168 if (!matching_early_clobber[i])
3170 lra_assert (reject > 0);
3171 reject--;
3172 matching_early_clobber[i] = 1;
3174 overall += LRA_LOSER_COST_FACTOR - 1;
3176 if (lra_dump_file != NULL)
3177 fprintf (lra_dump_file, " alt=%d,overall=%d,losers=%d,rld_nregs=%d\n",
3178 nalt, overall, losers, reload_nregs);
3180 /* If this alternative can be made to work by reloading, and it
3181 needs less reloading than the others checked so far, record
3182 it as the chosen goal for reloading. */
3183 if ((best_losers != 0 && losers == 0)
3184 || (((best_losers == 0 && losers == 0)
3185 || (best_losers != 0 && losers != 0))
3186 && (best_overall > overall
3187 || (best_overall == overall
3188 /* If the cost of the reloads is the same,
3189 prefer alternative which requires minimal
3190 number of reload regs. */
3191 && (reload_nregs < best_reload_nregs
3192 || (reload_nregs == best_reload_nregs
3193 && (best_reload_sum < reload_sum
3194 || (best_reload_sum == reload_sum
3195 && nalt < goal_alt_number))))))))
3197 for (nop = 0; nop < n_operands; nop++)
3199 goal_alt_win[nop] = curr_alt_win[nop];
3200 goal_alt_match_win[nop] = curr_alt_match_win[nop];
3201 goal_alt_matches[nop] = curr_alt_matches[nop];
3202 goal_alt[nop] = curr_alt[nop];
3203 goal_alt_offmemok[nop] = curr_alt_offmemok[nop];
3205 goal_alt_dont_inherit_ops_num = curr_alt_dont_inherit_ops_num;
3206 for (nop = 0; nop < curr_alt_dont_inherit_ops_num; nop++)
3207 goal_alt_dont_inherit_ops[nop] = curr_alt_dont_inherit_ops[nop];
3208 goal_alt_swapped = curr_swapped;
3209 best_overall = overall;
3210 best_losers = losers;
3211 best_reload_nregs = reload_nregs;
3212 best_reload_sum = reload_sum;
3213 goal_alt_number = nalt;
3215 if (losers == 0)
3216 /* Everything is satisfied. Do not process alternatives
3217 anymore. */
3218 break;
3219 fail:
3222 return ok_p;
3225 /* Make reload base reg from address AD. */
3226 static rtx
3227 base_to_reg (struct address_info *ad)
3229 enum reg_class cl;
3230 int code = -1;
3231 rtx new_inner = NULL_RTX;
3232 rtx new_reg = NULL_RTX;
3233 rtx_insn *insn;
3234 rtx_insn *last_insn = get_last_insn();
3236 lra_assert (ad->disp == ad->disp_term);
3237 cl = base_reg_class (ad->mode, ad->as, ad->base_outer_code,
3238 get_index_code (ad));
3239 new_reg = lra_create_new_reg (GET_MODE (*ad->base), NULL_RTX,
3240 cl, "base");
3241 new_inner = simplify_gen_binary (PLUS, GET_MODE (new_reg), new_reg,
3242 ad->disp_term == NULL
3243 ? const0_rtx
3244 : *ad->disp_term);
3245 if (!valid_address_p (ad->mode, new_inner, ad->as))
3246 return NULL_RTX;
3247 insn = emit_insn (gen_rtx_SET (new_reg, *ad->base));
3248 code = recog_memoized (insn);
3249 if (code < 0)
3251 delete_insns_since (last_insn);
3252 return NULL_RTX;
3255 return new_inner;
3258 /* Make reload base reg + DISP from address AD. Return the new pseudo. */
3259 static rtx
3260 base_plus_disp_to_reg (struct address_info *ad, rtx disp)
3262 enum reg_class cl;
3263 rtx new_reg;
3265 lra_assert (ad->base == ad->base_term);
3266 cl = base_reg_class (ad->mode, ad->as, ad->base_outer_code,
3267 get_index_code (ad));
3268 new_reg = lra_create_new_reg (GET_MODE (*ad->base_term), NULL_RTX,
3269 cl, "base + disp");
3270 lra_emit_add (new_reg, *ad->base_term, disp);
3271 return new_reg;
3274 /* Make reload of index part of address AD. Return the new
3275 pseudo. */
3276 static rtx
3277 index_part_to_reg (struct address_info *ad)
3279 rtx new_reg;
3281 new_reg = lra_create_new_reg (GET_MODE (*ad->index), NULL_RTX,
3282 INDEX_REG_CLASS, "index term");
3283 expand_mult (GET_MODE (*ad->index), *ad->index_term,
3284 GEN_INT (get_index_scale (ad)), new_reg, 1);
3285 return new_reg;
3288 /* Return true if we can add a displacement to address AD, even if that
3289 makes the address invalid. The fix-up code requires any new address
3290 to be the sum of the BASE_TERM, INDEX and DISP_TERM fields. */
3291 static bool
3292 can_add_disp_p (struct address_info *ad)
3294 return (!ad->autoinc_p
3295 && ad->segment == NULL
3296 && ad->base == ad->base_term
3297 && ad->disp == ad->disp_term);
3300 /* Make equiv substitution in address AD. Return true if a substitution
3301 was made. */
3302 static bool
3303 equiv_address_substitution (struct address_info *ad)
3305 rtx base_reg, new_base_reg, index_reg, new_index_reg, *base_term, *index_term;
3306 poly_int64 disp;
3307 HOST_WIDE_INT scale;
3308 bool change_p;
3310 base_term = strip_subreg (ad->base_term);
3311 if (base_term == NULL)
3312 base_reg = new_base_reg = NULL_RTX;
3313 else
3315 base_reg = *base_term;
3316 new_base_reg = get_equiv_with_elimination (base_reg, curr_insn);
3318 index_term = strip_subreg (ad->index_term);
3319 if (index_term == NULL)
3320 index_reg = new_index_reg = NULL_RTX;
3321 else
3323 index_reg = *index_term;
3324 new_index_reg = get_equiv_with_elimination (index_reg, curr_insn);
3326 if (base_reg == new_base_reg && index_reg == new_index_reg)
3327 return false;
3328 disp = 0;
3329 change_p = false;
3330 if (lra_dump_file != NULL)
3332 fprintf (lra_dump_file, "Changing address in insn %d ",
3333 INSN_UID (curr_insn));
3334 dump_value_slim (lra_dump_file, *ad->outer, 1);
3336 if (base_reg != new_base_reg)
3338 poly_int64 offset;
3339 if (REG_P (new_base_reg))
3341 *base_term = new_base_reg;
3342 change_p = true;
3344 else if (GET_CODE (new_base_reg) == PLUS
3345 && REG_P (XEXP (new_base_reg, 0))
3346 && poly_int_rtx_p (XEXP (new_base_reg, 1), &offset)
3347 && can_add_disp_p (ad))
3349 disp += offset;
3350 *base_term = XEXP (new_base_reg, 0);
3351 change_p = true;
3353 if (ad->base_term2 != NULL)
3354 *ad->base_term2 = *ad->base_term;
3356 if (index_reg != new_index_reg)
3358 poly_int64 offset;
3359 if (REG_P (new_index_reg))
3361 *index_term = new_index_reg;
3362 change_p = true;
3364 else if (GET_CODE (new_index_reg) == PLUS
3365 && REG_P (XEXP (new_index_reg, 0))
3366 && poly_int_rtx_p (XEXP (new_index_reg, 1), &offset)
3367 && can_add_disp_p (ad)
3368 && (scale = get_index_scale (ad)))
3370 disp += offset * scale;
3371 *index_term = XEXP (new_index_reg, 0);
3372 change_p = true;
3375 if (maybe_ne (disp, 0))
3377 if (ad->disp != NULL)
3378 *ad->disp = plus_constant (GET_MODE (*ad->inner), *ad->disp, disp);
3379 else
3381 *ad->inner = plus_constant (GET_MODE (*ad->inner), *ad->inner, disp);
3382 update_address (ad);
3384 change_p = true;
3386 if (lra_dump_file != NULL)
3388 if (! change_p)
3389 fprintf (lra_dump_file, " -- no change\n");
3390 else
3392 fprintf (lra_dump_file, " on equiv ");
3393 dump_value_slim (lra_dump_file, *ad->outer, 1);
3394 fprintf (lra_dump_file, "\n");
3397 return change_p;
3400 /* Skip all modifiers and whitespaces in constraint STR and return the
3401 result. */
3402 static const char *
3403 skip_constraint_modifiers (const char *str)
3405 for (;;str++)
3406 switch (*str)
3408 case '+': case '&' : case '=': case '*': case ' ': case '\t':
3409 case '$': case '^' : case '%': case '?': case '!':
3410 break;
3411 default: return str;
3415 /* Major function to make reloads for an address in operand NOP or
3416 check its correctness (If CHECK_ONLY_P is true). The supported
3417 cases are:
3419 1) an address that existed before LRA started, at which point it
3420 must have been valid. These addresses are subject to elimination
3421 and may have become invalid due to the elimination offset being out
3422 of range.
3424 2) an address created by forcing a constant to memory
3425 (force_const_to_mem). The initial form of these addresses might
3426 not be valid, and it is this function's job to make them valid.
3428 3) a frame address formed from a register and a (possibly zero)
3429 constant offset. As above, these addresses might not be valid and
3430 this function must make them so.
3432 Add reloads to the lists *BEFORE and *AFTER. We might need to add
3433 reloads to *AFTER because of inc/dec, {pre, post} modify in the
3434 address. Return true for any RTL change.
3436 The function is a helper function which does not produce all
3437 transformations (when CHECK_ONLY_P is false) which can be
3438 necessary. It does just basic steps. To do all necessary
3439 transformations use function process_address. */
3440 static bool
3441 process_address_1 (int nop, bool check_only_p,
3442 rtx_insn **before, rtx_insn **after)
3444 struct address_info ad;
3445 rtx new_reg;
3446 HOST_WIDE_INT scale;
3447 rtx op = *curr_id->operand_loc[nop];
3448 rtx mem = extract_mem_from_operand (op);
3449 const char *constraint;
3450 enum constraint_num cn;
3451 bool change_p = false;
3453 if (MEM_P (mem)
3454 && GET_MODE (mem) == BLKmode
3455 && GET_CODE (XEXP (mem, 0)) == SCRATCH)
3456 return false;
3458 constraint
3459 = skip_constraint_modifiers (curr_static_id->operand[nop].constraint);
3460 if (IN_RANGE (constraint[0], '0', '9'))
3462 char *end;
3463 unsigned long dup = strtoul (constraint, &end, 10);
3464 constraint
3465 = skip_constraint_modifiers (curr_static_id->operand[dup].constraint);
3467 cn = lookup_constraint (*constraint == '\0' ? "X" : constraint);
3468 /* If we have several alternatives or/and several constraints in an
3469 alternative and we can not say at this stage what constraint will be used,
3470 use unknown constraint. The exception is an address constraint. If
3471 operand has one address constraint, probably all others constraints are
3472 address ones. */
3473 if (constraint[0] != '\0' && get_constraint_type (cn) != CT_ADDRESS
3474 && *skip_constraint_modifiers (constraint
3475 + CONSTRAINT_LEN (constraint[0],
3476 constraint)) != '\0')
3477 cn = CONSTRAINT__UNKNOWN;
3478 if (insn_extra_address_constraint (cn)
3479 /* When we find an asm operand with an address constraint that
3480 doesn't satisfy address_operand to begin with, we clear
3481 is_address, so that we don't try to make a non-address fit.
3482 If the asm statement got this far, it's because other
3483 constraints are available, and we'll use them, disregarding
3484 the unsatisfiable address ones. */
3485 && curr_static_id->operand[nop].is_address)
3486 decompose_lea_address (&ad, curr_id->operand_loc[nop]);
3487 /* Do not attempt to decompose arbitrary addresses generated by combine
3488 for asm operands with loose constraints, e.g 'X'.
3489 Need to extract memory from op for special memory constraint,
3490 i.e. bcst_mem_operand in i386 backend. */
3491 else if (MEM_P (mem)
3492 && !(INSN_CODE (curr_insn) < 0
3493 && get_constraint_type (cn) == CT_FIXED_FORM
3494 && constraint_satisfied_p (op, cn)))
3495 decompose_mem_address (&ad, mem);
3496 else if (GET_CODE (op) == SUBREG
3497 && MEM_P (SUBREG_REG (op)))
3498 decompose_mem_address (&ad, SUBREG_REG (op));
3499 else
3500 return false;
3501 /* If INDEX_REG_CLASS is assigned to base_term already and isn't to
3502 index_term, swap them so to avoid assigning INDEX_REG_CLASS to both
3503 when INDEX_REG_CLASS is a single register class. */
3504 if (ad.base_term != NULL
3505 && ad.index_term != NULL
3506 && ira_class_hard_regs_num[INDEX_REG_CLASS] == 1
3507 && REG_P (*ad.base_term)
3508 && REG_P (*ad.index_term)
3509 && in_class_p (*ad.base_term, INDEX_REG_CLASS, NULL)
3510 && ! in_class_p (*ad.index_term, INDEX_REG_CLASS, NULL))
3512 std::swap (ad.base, ad.index);
3513 std::swap (ad.base_term, ad.index_term);
3515 if (! check_only_p)
3516 change_p = equiv_address_substitution (&ad);
3517 if (ad.base_term != NULL
3518 && (process_addr_reg
3519 (ad.base_term, check_only_p, before,
3520 (ad.autoinc_p
3521 && !(REG_P (*ad.base_term)
3522 && find_regno_note (curr_insn, REG_DEAD,
3523 REGNO (*ad.base_term)) != NULL_RTX)
3524 ? after : NULL),
3525 base_reg_class (ad.mode, ad.as, ad.base_outer_code,
3526 get_index_code (&ad)))))
3528 change_p = true;
3529 if (ad.base_term2 != NULL)
3530 *ad.base_term2 = *ad.base_term;
3532 if (ad.index_term != NULL
3533 && process_addr_reg (ad.index_term, check_only_p,
3534 before, NULL, INDEX_REG_CLASS))
3535 change_p = true;
3537 /* Target hooks sometimes don't treat extra-constraint addresses as
3538 legitimate address_operands, so handle them specially. */
3539 if (insn_extra_address_constraint (cn)
3540 && satisfies_address_constraint_p (&ad, cn))
3541 return change_p;
3543 if (check_only_p)
3544 return change_p;
3546 /* There are three cases where the shape of *AD.INNER may now be invalid:
3548 1) the original address was valid, but either elimination or
3549 equiv_address_substitution was applied and that made
3550 the address invalid.
3552 2) the address is an invalid symbolic address created by
3553 force_const_to_mem.
3555 3) the address is a frame address with an invalid offset.
3557 4) the address is a frame address with an invalid base.
3559 All these cases involve a non-autoinc address, so there is no
3560 point revalidating other types. */
3561 if (ad.autoinc_p || valid_address_p (op, &ad, cn))
3562 return change_p;
3564 /* Any index existed before LRA started, so we can assume that the
3565 presence and shape of the index is valid. */
3566 push_to_sequence (*before);
3567 lra_assert (ad.disp == ad.disp_term);
3568 if (ad.base == NULL)
3570 if (ad.index == NULL)
3572 rtx_insn *insn;
3573 rtx_insn *last = get_last_insn ();
3574 int code = -1;
3575 enum reg_class cl = base_reg_class (ad.mode, ad.as,
3576 SCRATCH, SCRATCH);
3577 rtx addr = *ad.inner;
3579 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "addr");
3580 if (HAVE_lo_sum)
3582 /* addr => lo_sum (new_base, addr), case (2) above. */
3583 insn = emit_insn (gen_rtx_SET
3584 (new_reg,
3585 gen_rtx_HIGH (Pmode, copy_rtx (addr))));
3586 code = recog_memoized (insn);
3587 if (code >= 0)
3589 *ad.inner = gen_rtx_LO_SUM (Pmode, new_reg, addr);
3590 if (!valid_address_p (op, &ad, cn))
3592 /* Try to put lo_sum into register. */
3593 insn = emit_insn (gen_rtx_SET
3594 (new_reg,
3595 gen_rtx_LO_SUM (Pmode, new_reg, addr)));
3596 code = recog_memoized (insn);
3597 if (code >= 0)
3599 *ad.inner = new_reg;
3600 if (!valid_address_p (op, &ad, cn))
3602 *ad.inner = addr;
3603 code = -1;
3609 if (code < 0)
3610 delete_insns_since (last);
3613 if (code < 0)
3615 /* addr => new_base, case (2) above. */
3616 lra_emit_move (new_reg, addr);
3618 for (insn = last == NULL_RTX ? get_insns () : NEXT_INSN (last);
3619 insn != NULL_RTX;
3620 insn = NEXT_INSN (insn))
3621 if (recog_memoized (insn) < 0)
3622 break;
3623 if (insn != NULL_RTX)
3625 /* Do nothing if we cannot generate right insns.
3626 This is analogous to reload pass behavior. */
3627 delete_insns_since (last);
3628 end_sequence ();
3629 return false;
3631 *ad.inner = new_reg;
3634 else
3636 /* index * scale + disp => new base + index * scale,
3637 case (1) above. */
3638 enum reg_class cl = base_reg_class (ad.mode, ad.as, PLUS,
3639 GET_CODE (*ad.index));
3641 lra_assert (INDEX_REG_CLASS != NO_REGS);
3642 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "disp");
3643 lra_emit_move (new_reg, *ad.disp);
3644 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3645 new_reg, *ad.index);
3648 else if (ad.index == NULL)
3650 int regno;
3651 enum reg_class cl;
3652 rtx set;
3653 rtx_insn *insns, *last_insn;
3654 /* Try to reload base into register only if the base is invalid
3655 for the address but with valid offset, case (4) above. */
3656 start_sequence ();
3657 new_reg = base_to_reg (&ad);
3659 /* base + disp => new base, cases (1) and (3) above. */
3660 /* Another option would be to reload the displacement into an
3661 index register. However, postreload has code to optimize
3662 address reloads that have the same base and different
3663 displacements, so reloading into an index register would
3664 not necessarily be a win. */
3665 if (new_reg == NULL_RTX)
3667 /* See if the target can split the displacement into a
3668 legitimate new displacement from a local anchor. */
3669 gcc_assert (ad.disp == ad.disp_term);
3670 poly_int64 orig_offset;
3671 rtx offset1, offset2;
3672 if (poly_int_rtx_p (*ad.disp, &orig_offset)
3673 && targetm.legitimize_address_displacement (&offset1, &offset2,
3674 orig_offset,
3675 ad.mode))
3677 new_reg = base_plus_disp_to_reg (&ad, offset1);
3678 new_reg = gen_rtx_PLUS (GET_MODE (new_reg), new_reg, offset2);
3680 else
3681 new_reg = base_plus_disp_to_reg (&ad, *ad.disp);
3683 insns = get_insns ();
3684 last_insn = get_last_insn ();
3685 /* If we generated at least two insns, try last insn source as
3686 an address. If we succeed, we generate one less insn. */
3687 if (REG_P (new_reg)
3688 && last_insn != insns
3689 && (set = single_set (last_insn)) != NULL_RTX
3690 && GET_CODE (SET_SRC (set)) == PLUS
3691 && REG_P (XEXP (SET_SRC (set), 0))
3692 && CONSTANT_P (XEXP (SET_SRC (set), 1)))
3694 *ad.inner = SET_SRC (set);
3695 if (valid_address_p (op, &ad, cn))
3697 *ad.base_term = XEXP (SET_SRC (set), 0);
3698 *ad.disp_term = XEXP (SET_SRC (set), 1);
3699 cl = base_reg_class (ad.mode, ad.as, ad.base_outer_code,
3700 get_index_code (&ad));
3701 regno = REGNO (*ad.base_term);
3702 if (regno >= FIRST_PSEUDO_REGISTER
3703 && cl != lra_get_allocno_class (regno))
3704 lra_change_class (regno, cl, " Change to", true);
3705 new_reg = SET_SRC (set);
3706 delete_insns_since (PREV_INSN (last_insn));
3709 end_sequence ();
3710 emit_insn (insns);
3711 *ad.inner = new_reg;
3713 else if (ad.disp_term != NULL)
3715 /* base + scale * index + disp => new base + scale * index,
3716 case (1) above. */
3717 gcc_assert (ad.disp == ad.disp_term);
3718 new_reg = base_plus_disp_to_reg (&ad, *ad.disp);
3719 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3720 new_reg, *ad.index);
3722 else if ((scale = get_index_scale (&ad)) == 1)
3724 /* The last transformation to one reg will be made in
3725 curr_insn_transform function. */
3726 end_sequence ();
3727 return false;
3729 else if (scale != 0)
3731 /* base + scale * index => base + new_reg,
3732 case (1) above.
3733 Index part of address may become invalid. For example, we
3734 changed pseudo on the equivalent memory and a subreg of the
3735 pseudo onto the memory of different mode for which the scale is
3736 prohibitted. */
3737 new_reg = index_part_to_reg (&ad);
3738 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3739 *ad.base_term, new_reg);
3741 else
3743 enum reg_class cl = base_reg_class (ad.mode, ad.as,
3744 SCRATCH, SCRATCH);
3745 rtx addr = *ad.inner;
3747 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "addr");
3748 /* addr => new_base. */
3749 lra_emit_move (new_reg, addr);
3750 *ad.inner = new_reg;
3752 *before = get_insns ();
3753 end_sequence ();
3754 return true;
3757 /* If CHECK_ONLY_P is false, do address reloads until it is necessary.
3758 Use process_address_1 as a helper function. Return true for any
3759 RTL changes.
3761 If CHECK_ONLY_P is true, just check address correctness. Return
3762 false if the address correct. */
3763 static bool
3764 process_address (int nop, bool check_only_p,
3765 rtx_insn **before, rtx_insn **after)
3767 bool res = false;
3769 while (process_address_1 (nop, check_only_p, before, after))
3771 if (check_only_p)
3772 return true;
3773 res = true;
3775 return res;
3778 /* Emit insns to reload VALUE into a new register. VALUE is an
3779 auto-increment or auto-decrement RTX whose operand is a register or
3780 memory location; so reloading involves incrementing that location.
3781 IN is either identical to VALUE, or some cheaper place to reload
3782 value being incremented/decremented from.
3784 INC_AMOUNT is the number to increment or decrement by (always
3785 positive and ignored for POST_MODIFY/PRE_MODIFY).
3787 Return pseudo containing the result. */
3788 static rtx
3789 emit_inc (enum reg_class new_rclass, rtx in, rtx value, poly_int64 inc_amount)
3791 /* REG or MEM to be copied and incremented. */
3792 rtx incloc = XEXP (value, 0);
3793 /* Nonzero if increment after copying. */
3794 int post = (GET_CODE (value) == POST_DEC || GET_CODE (value) == POST_INC
3795 || GET_CODE (value) == POST_MODIFY);
3796 rtx_insn *last;
3797 rtx inc;
3798 rtx_insn *add_insn;
3799 int code;
3800 rtx real_in = in == value ? incloc : in;
3801 rtx result;
3802 bool plus_p = true;
3804 if (GET_CODE (value) == PRE_MODIFY || GET_CODE (value) == POST_MODIFY)
3806 lra_assert (GET_CODE (XEXP (value, 1)) == PLUS
3807 || GET_CODE (XEXP (value, 1)) == MINUS);
3808 lra_assert (rtx_equal_p (XEXP (XEXP (value, 1), 0), XEXP (value, 0)));
3809 plus_p = GET_CODE (XEXP (value, 1)) == PLUS;
3810 inc = XEXP (XEXP (value, 1), 1);
3812 else
3814 if (GET_CODE (value) == PRE_DEC || GET_CODE (value) == POST_DEC)
3815 inc_amount = -inc_amount;
3817 inc = gen_int_mode (inc_amount, GET_MODE (value));
3820 if (! post && REG_P (incloc))
3821 result = incloc;
3822 else
3823 result = lra_create_new_reg (GET_MODE (value), value, new_rclass,
3824 "INC/DEC result");
3826 if (real_in != result)
3828 /* First copy the location to the result register. */
3829 lra_assert (REG_P (result));
3830 emit_insn (gen_move_insn (result, real_in));
3833 /* We suppose that there are insns to add/sub with the constant
3834 increment permitted in {PRE/POST)_{DEC/INC/MODIFY}. At least the
3835 old reload worked with this assumption. If the assumption
3836 becomes wrong, we should use approach in function
3837 base_plus_disp_to_reg. */
3838 if (in == value)
3840 /* See if we can directly increment INCLOC. */
3841 last = get_last_insn ();
3842 add_insn = emit_insn (plus_p
3843 ? gen_add2_insn (incloc, inc)
3844 : gen_sub2_insn (incloc, inc));
3846 code = recog_memoized (add_insn);
3847 if (code >= 0)
3849 if (! post && result != incloc)
3850 emit_insn (gen_move_insn (result, incloc));
3851 return result;
3853 delete_insns_since (last);
3856 /* If couldn't do the increment directly, must increment in RESULT.
3857 The way we do this depends on whether this is pre- or
3858 post-increment. For pre-increment, copy INCLOC to the reload
3859 register, increment it there, then save back. */
3860 if (! post)
3862 if (real_in != result)
3863 emit_insn (gen_move_insn (result, real_in));
3864 if (plus_p)
3865 emit_insn (gen_add2_insn (result, inc));
3866 else
3867 emit_insn (gen_sub2_insn (result, inc));
3868 if (result != incloc)
3869 emit_insn (gen_move_insn (incloc, result));
3871 else
3873 /* Post-increment.
3875 Because this might be a jump insn or a compare, and because
3876 RESULT may not be available after the insn in an input
3877 reload, we must do the incrementing before the insn being
3878 reloaded for.
3880 We have already copied IN to RESULT. Increment the copy in
3881 RESULT, save that back, then decrement RESULT so it has
3882 the original value. */
3883 if (plus_p)
3884 emit_insn (gen_add2_insn (result, inc));
3885 else
3886 emit_insn (gen_sub2_insn (result, inc));
3887 emit_insn (gen_move_insn (incloc, result));
3888 /* Restore non-modified value for the result. We prefer this
3889 way because it does not require an additional hard
3890 register. */
3891 if (plus_p)
3893 poly_int64 offset;
3894 if (poly_int_rtx_p (inc, &offset))
3895 emit_insn (gen_add2_insn (result,
3896 gen_int_mode (-offset,
3897 GET_MODE (result))));
3898 else
3899 emit_insn (gen_sub2_insn (result, inc));
3901 else
3902 emit_insn (gen_add2_insn (result, inc));
3904 return result;
3907 /* Return true if the current move insn does not need processing as we
3908 already know that it satisfies its constraints. */
3909 static bool
3910 simple_move_p (void)
3912 rtx dest, src;
3913 enum reg_class dclass, sclass;
3915 lra_assert (curr_insn_set != NULL_RTX);
3916 dest = SET_DEST (curr_insn_set);
3917 src = SET_SRC (curr_insn_set);
3919 /* If the instruction has multiple sets we need to process it even if it
3920 is single_set. This can happen if one or more of the SETs are dead.
3921 See PR73650. */
3922 if (multiple_sets (curr_insn))
3923 return false;
3925 return ((dclass = get_op_class (dest)) != NO_REGS
3926 && (sclass = get_op_class (src)) != NO_REGS
3927 /* The backend guarantees that register moves of cost 2
3928 never need reloads. */
3929 && targetm.register_move_cost (GET_MODE (src), sclass, dclass) == 2);
3932 /* Swap operands NOP and NOP + 1. */
3933 static inline void
3934 swap_operands (int nop)
3936 std::swap (curr_operand_mode[nop], curr_operand_mode[nop + 1]);
3937 std::swap (original_subreg_reg_mode[nop], original_subreg_reg_mode[nop + 1]);
3938 std::swap (*curr_id->operand_loc[nop], *curr_id->operand_loc[nop + 1]);
3939 std::swap (equiv_substition_p[nop], equiv_substition_p[nop + 1]);
3940 /* Swap the duplicates too. */
3941 lra_update_dup (curr_id, nop);
3942 lra_update_dup (curr_id, nop + 1);
3945 /* Main entry point of the constraint code: search the body of the
3946 current insn to choose the best alternative. It is mimicking insn
3947 alternative cost calculation model of former reload pass. That is
3948 because machine descriptions were written to use this model. This
3949 model can be changed in future. Make commutative operand exchange
3950 if it is chosen.
3952 if CHECK_ONLY_P is false, do RTL changes to satisfy the
3953 constraints. Return true if any change happened during function
3954 call.
3956 If CHECK_ONLY_P is true then don't do any transformation. Just
3957 check that the insn satisfies all constraints. If the insn does
3958 not satisfy any constraint, return true. */
3959 static bool
3960 curr_insn_transform (bool check_only_p)
3962 int i, j, k;
3963 int n_operands;
3964 int n_alternatives;
3965 int n_outputs;
3966 int commutative;
3967 signed char goal_alt_matched[MAX_RECOG_OPERANDS][MAX_RECOG_OPERANDS];
3968 signed char match_inputs[MAX_RECOG_OPERANDS + 1];
3969 signed char outputs[MAX_RECOG_OPERANDS + 1];
3970 rtx_insn *before, *after;
3971 bool alt_p = false;
3972 /* Flag that the insn has been changed through a transformation. */
3973 bool change_p;
3974 bool sec_mem_p;
3975 bool use_sec_mem_p;
3976 int max_regno_before;
3977 int reused_alternative_num;
3979 curr_insn_set = single_set (curr_insn);
3980 if (curr_insn_set != NULL_RTX && simple_move_p ())
3982 /* We assume that the corresponding insn alternative has no
3983 earlier clobbers. If it is not the case, don't define move
3984 cost equal to 2 for the corresponding register classes. */
3985 lra_set_used_insn_alternative (curr_insn, LRA_NON_CLOBBERED_ALT);
3986 return false;
3989 no_input_reloads_p = no_output_reloads_p = false;
3990 goal_alt_number = -1;
3991 change_p = sec_mem_p = false;
3993 /* CALL_INSNs are not allowed to have any output reloads. */
3994 if (CALL_P (curr_insn))
3995 no_output_reloads_p = true;
3997 n_operands = curr_static_id->n_operands;
3998 n_alternatives = curr_static_id->n_alternatives;
4000 /* Just return "no reloads" if insn has no operands with
4001 constraints. */
4002 if (n_operands == 0 || n_alternatives == 0)
4003 return false;
4005 max_regno_before = max_reg_num ();
4007 for (i = 0; i < n_operands; i++)
4009 goal_alt_matched[i][0] = -1;
4010 goal_alt_matches[i] = -1;
4013 commutative = curr_static_id->commutative;
4015 /* Now see what we need for pseudos that didn't get hard regs or got
4016 the wrong kind of hard reg. For this, we must consider all the
4017 operands together against the register constraints. */
4019 best_losers = best_overall = INT_MAX;
4020 best_reload_sum = 0;
4022 curr_swapped = false;
4023 goal_alt_swapped = false;
4025 if (! check_only_p)
4026 /* Make equivalence substitution and memory subreg elimination
4027 before address processing because an address legitimacy can
4028 depend on memory mode. */
4029 for (i = 0; i < n_operands; i++)
4031 rtx op, subst, old;
4032 bool op_change_p = false;
4034 if (curr_static_id->operand[i].is_operator)
4035 continue;
4037 old = op = *curr_id->operand_loc[i];
4038 if (GET_CODE (old) == SUBREG)
4039 old = SUBREG_REG (old);
4040 subst = get_equiv_with_elimination (old, curr_insn);
4041 original_subreg_reg_mode[i] = VOIDmode;
4042 equiv_substition_p[i] = false;
4043 if (subst != old)
4045 equiv_substition_p[i] = true;
4046 subst = copy_rtx (subst);
4047 lra_assert (REG_P (old));
4048 if (GET_CODE (op) != SUBREG)
4049 *curr_id->operand_loc[i] = subst;
4050 else
4052 SUBREG_REG (op) = subst;
4053 if (GET_MODE (subst) == VOIDmode)
4054 original_subreg_reg_mode[i] = GET_MODE (old);
4056 if (lra_dump_file != NULL)
4058 fprintf (lra_dump_file,
4059 "Changing pseudo %d in operand %i of insn %u on equiv ",
4060 REGNO (old), i, INSN_UID (curr_insn));
4061 dump_value_slim (lra_dump_file, subst, 1);
4062 fprintf (lra_dump_file, "\n");
4064 op_change_p = change_p = true;
4066 if (simplify_operand_subreg (i, GET_MODE (old)) || op_change_p)
4068 change_p = true;
4069 lra_update_dup (curr_id, i);
4073 /* Reload address registers and displacements. We do it before
4074 finding an alternative because of memory constraints. */
4075 before = after = NULL;
4076 for (i = 0; i < n_operands; i++)
4077 if (! curr_static_id->operand[i].is_operator
4078 && process_address (i, check_only_p, &before, &after))
4080 if (check_only_p)
4081 return true;
4082 change_p = true;
4083 lra_update_dup (curr_id, i);
4086 if (change_p)
4087 /* If we've changed the instruction then any alternative that
4088 we chose previously may no longer be valid. */
4089 lra_set_used_insn_alternative (curr_insn, LRA_UNKNOWN_ALT);
4091 if (! check_only_p && curr_insn_set != NULL_RTX
4092 && check_and_process_move (&change_p, &sec_mem_p))
4093 return change_p;
4095 try_swapped:
4097 reused_alternative_num = check_only_p ? LRA_UNKNOWN_ALT : curr_id->used_insn_alternative;
4098 if (lra_dump_file != NULL && reused_alternative_num >= 0)
4099 fprintf (lra_dump_file, "Reusing alternative %d for insn #%u\n",
4100 reused_alternative_num, INSN_UID (curr_insn));
4102 if (process_alt_operands (reused_alternative_num))
4103 alt_p = true;
4105 if (check_only_p)
4106 return ! alt_p || best_losers != 0;
4108 /* If insn is commutative (it's safe to exchange a certain pair of
4109 operands) then we need to try each alternative twice, the second
4110 time matching those two operands as if we had exchanged them. To
4111 do this, really exchange them in operands.
4113 If we have just tried the alternatives the second time, return
4114 operands to normal and drop through. */
4116 if (reused_alternative_num < 0 && commutative >= 0)
4118 curr_swapped = !curr_swapped;
4119 if (curr_swapped)
4121 swap_operands (commutative);
4122 goto try_swapped;
4124 else
4125 swap_operands (commutative);
4128 if (! alt_p && ! sec_mem_p)
4130 /* No alternative works with reloads?? */
4131 if (INSN_CODE (curr_insn) >= 0)
4132 fatal_insn ("unable to generate reloads for:", curr_insn);
4133 error_for_asm (curr_insn,
4134 "inconsistent operand constraints in an %<asm%>");
4135 lra_asm_error_p = true;
4136 if (! JUMP_P (curr_insn))
4138 /* Avoid further trouble with this insn. Don't generate use
4139 pattern here as we could use the insn SP offset. */
4140 lra_set_insn_deleted (curr_insn);
4142 else
4144 lra_invalidate_insn_data (curr_insn);
4145 ira_nullify_asm_goto (curr_insn);
4146 lra_update_insn_regno_info (curr_insn);
4148 return true;
4151 /* If the best alternative is with operands 1 and 2 swapped, swap
4152 them. Update the operand numbers of any reloads already
4153 pushed. */
4155 if (goal_alt_swapped)
4157 if (lra_dump_file != NULL)
4158 fprintf (lra_dump_file, " Commutative operand exchange in insn %u\n",
4159 INSN_UID (curr_insn));
4161 /* Swap the duplicates too. */
4162 swap_operands (commutative);
4163 change_p = true;
4166 /* Some targets' TARGET_SECONDARY_MEMORY_NEEDED (e.g. x86) are defined
4167 too conservatively. So we use the secondary memory only if there
4168 is no any alternative without reloads. */
4169 use_sec_mem_p = false;
4170 if (! alt_p)
4171 use_sec_mem_p = true;
4172 else if (sec_mem_p)
4174 for (i = 0; i < n_operands; i++)
4175 if (! goal_alt_win[i] && ! goal_alt_match_win[i])
4176 break;
4177 use_sec_mem_p = i < n_operands;
4180 if (use_sec_mem_p)
4182 int in = -1, out = -1;
4183 rtx new_reg, src, dest, rld;
4184 machine_mode sec_mode, rld_mode;
4186 lra_assert (curr_insn_set != NULL_RTX && sec_mem_p);
4187 dest = SET_DEST (curr_insn_set);
4188 src = SET_SRC (curr_insn_set);
4189 for (i = 0; i < n_operands; i++)
4190 if (*curr_id->operand_loc[i] == dest)
4191 out = i;
4192 else if (*curr_id->operand_loc[i] == src)
4193 in = i;
4194 for (i = 0; i < curr_static_id->n_dups; i++)
4195 if (out < 0 && *curr_id->dup_loc[i] == dest)
4196 out = curr_static_id->dup_num[i];
4197 else if (in < 0 && *curr_id->dup_loc[i] == src)
4198 in = curr_static_id->dup_num[i];
4199 lra_assert (out >= 0 && in >= 0
4200 && curr_static_id->operand[out].type == OP_OUT
4201 && curr_static_id->operand[in].type == OP_IN);
4202 rld = partial_subreg_p (GET_MODE (src), GET_MODE (dest)) ? src : dest;
4203 rld_mode = GET_MODE (rld);
4204 sec_mode = targetm.secondary_memory_needed_mode (rld_mode);
4205 new_reg = lra_create_new_reg (sec_mode, NULL_RTX,
4206 NO_REGS, "secondary");
4207 /* If the mode is changed, it should be wider. */
4208 lra_assert (!partial_subreg_p (sec_mode, rld_mode));
4209 if (sec_mode != rld_mode)
4211 /* If the target says specifically to use another mode for
4212 secondary memory moves we cannot reuse the original
4213 insn. */
4214 after = emit_spill_move (false, new_reg, dest);
4215 lra_process_new_insns (curr_insn, NULL, after,
4216 "Inserting the sec. move");
4217 /* We may have non null BEFORE here (e.g. after address
4218 processing. */
4219 push_to_sequence (before);
4220 before = emit_spill_move (true, new_reg, src);
4221 emit_insn (before);
4222 before = get_insns ();
4223 end_sequence ();
4224 lra_process_new_insns (curr_insn, before, NULL, "Changing on");
4225 lra_set_insn_deleted (curr_insn);
4227 else if (dest == rld)
4229 *curr_id->operand_loc[out] = new_reg;
4230 lra_update_dup (curr_id, out);
4231 after = emit_spill_move (false, new_reg, dest);
4232 lra_process_new_insns (curr_insn, NULL, after,
4233 "Inserting the sec. move");
4235 else
4237 *curr_id->operand_loc[in] = new_reg;
4238 lra_update_dup (curr_id, in);
4239 /* See comments above. */
4240 push_to_sequence (before);
4241 before = emit_spill_move (true, new_reg, src);
4242 emit_insn (before);
4243 before = get_insns ();
4244 end_sequence ();
4245 lra_process_new_insns (curr_insn, before, NULL,
4246 "Inserting the sec. move");
4248 lra_update_insn_regno_info (curr_insn);
4249 return true;
4252 lra_assert (goal_alt_number >= 0);
4253 lra_set_used_insn_alternative (curr_insn, goal_alt_number);
4255 if (lra_dump_file != NULL)
4257 const char *p;
4259 fprintf (lra_dump_file, " Choosing alt %d in insn %u:",
4260 goal_alt_number, INSN_UID (curr_insn));
4261 for (i = 0; i < n_operands; i++)
4263 p = (curr_static_id->operand_alternative
4264 [goal_alt_number * n_operands + i].constraint);
4265 if (*p == '\0')
4266 continue;
4267 fprintf (lra_dump_file, " (%d) ", i);
4268 for (; *p != '\0' && *p != ',' && *p != '#'; p++)
4269 fputc (*p, lra_dump_file);
4271 if (INSN_CODE (curr_insn) >= 0
4272 && (p = get_insn_name (INSN_CODE (curr_insn))) != NULL)
4273 fprintf (lra_dump_file, " {%s}", p);
4274 if (maybe_ne (curr_id->sp_offset, 0))
4276 fprintf (lra_dump_file, " (sp_off=");
4277 print_dec (curr_id->sp_offset, lra_dump_file);
4278 fprintf (lra_dump_file, ")");
4280 fprintf (lra_dump_file, "\n");
4283 /* Right now, for any pair of operands I and J that are required to
4284 match, with J < I, goal_alt_matches[I] is J. Add I to
4285 goal_alt_matched[J]. */
4287 for (i = 0; i < n_operands; i++)
4288 if ((j = goal_alt_matches[i]) >= 0)
4290 for (k = 0; goal_alt_matched[j][k] >= 0; k++)
4292 /* We allow matching one output operand and several input
4293 operands. */
4294 lra_assert (k == 0
4295 || (curr_static_id->operand[j].type == OP_OUT
4296 && curr_static_id->operand[i].type == OP_IN
4297 && (curr_static_id->operand
4298 [goal_alt_matched[j][0]].type == OP_IN)));
4299 goal_alt_matched[j][k] = i;
4300 goal_alt_matched[j][k + 1] = -1;
4303 for (i = 0; i < n_operands; i++)
4304 goal_alt_win[i] |= goal_alt_match_win[i];
4306 /* Any constants that aren't allowed and can't be reloaded into
4307 registers are here changed into memory references. */
4308 for (i = 0; i < n_operands; i++)
4309 if (goal_alt_win[i])
4311 int regno;
4312 enum reg_class new_class;
4313 rtx reg = *curr_id->operand_loc[i];
4315 if (GET_CODE (reg) == SUBREG)
4316 reg = SUBREG_REG (reg);
4318 if (REG_P (reg) && (regno = REGNO (reg)) >= FIRST_PSEUDO_REGISTER)
4320 bool ok_p = in_class_p (reg, goal_alt[i], &new_class);
4322 if (new_class != NO_REGS && get_reg_class (regno) != new_class)
4324 lra_assert (ok_p);
4325 lra_change_class (regno, new_class, " Change to", true);
4329 else
4331 const char *constraint;
4332 char c;
4333 rtx op = *curr_id->operand_loc[i];
4334 rtx subreg = NULL_RTX;
4335 machine_mode mode = curr_operand_mode[i];
4337 if (GET_CODE (op) == SUBREG)
4339 subreg = op;
4340 op = SUBREG_REG (op);
4341 mode = GET_MODE (op);
4344 if (CONST_POOL_OK_P (mode, op)
4345 && ((targetm.preferred_reload_class
4346 (op, (enum reg_class) goal_alt[i]) == NO_REGS)
4347 || no_input_reloads_p))
4349 rtx tem = force_const_mem (mode, op);
4351 change_p = true;
4352 if (subreg != NULL_RTX)
4353 tem = gen_rtx_SUBREG (mode, tem, SUBREG_BYTE (subreg));
4355 *curr_id->operand_loc[i] = tem;
4356 lra_update_dup (curr_id, i);
4357 process_address (i, false, &before, &after);
4359 /* If the alternative accepts constant pool refs directly
4360 there will be no reload needed at all. */
4361 if (subreg != NULL_RTX)
4362 continue;
4363 /* Skip alternatives before the one requested. */
4364 constraint = (curr_static_id->operand_alternative
4365 [goal_alt_number * n_operands + i].constraint);
4366 for (;
4367 (c = *constraint) && c != ',' && c != '#';
4368 constraint += CONSTRAINT_LEN (c, constraint))
4370 enum constraint_num cn = lookup_constraint (constraint);
4371 if ((insn_extra_memory_constraint (cn)
4372 || insn_extra_special_memory_constraint (cn)
4373 || insn_extra_relaxed_memory_constraint (cn))
4374 && satisfies_memory_constraint_p (tem, cn))
4375 break;
4377 if (c == '\0' || c == ',' || c == '#')
4378 continue;
4380 goal_alt_win[i] = true;
4384 n_outputs = 0;
4385 outputs[0] = -1;
4386 for (i = 0; i < n_operands; i++)
4388 int regno;
4389 bool optional_p = false;
4390 rtx old, new_reg;
4391 rtx op = *curr_id->operand_loc[i];
4393 if (goal_alt_win[i])
4395 if (goal_alt[i] == NO_REGS
4396 && REG_P (op)
4397 /* When we assign NO_REGS it means that we will not
4398 assign a hard register to the scratch pseudo by
4399 assigment pass and the scratch pseudo will be
4400 spilled. Spilled scratch pseudos are transformed
4401 back to scratches at the LRA end. */
4402 && ira_former_scratch_operand_p (curr_insn, i)
4403 && ira_former_scratch_p (REGNO (op)))
4405 int regno = REGNO (op);
4406 lra_change_class (regno, NO_REGS, " Change to", true);
4407 if (lra_get_regno_hard_regno (regno) >= 0)
4408 /* We don't have to mark all insn affected by the
4409 spilled pseudo as there is only one such insn, the
4410 current one. */
4411 reg_renumber[regno] = -1;
4412 lra_assert (bitmap_single_bit_set_p
4413 (&lra_reg_info[REGNO (op)].insn_bitmap));
4415 /* We can do an optional reload. If the pseudo got a hard
4416 reg, we might improve the code through inheritance. If
4417 it does not get a hard register we coalesce memory/memory
4418 moves later. Ignore move insns to avoid cycling. */
4419 if (! lra_simple_p
4420 && lra_undo_inheritance_iter < LRA_MAX_INHERITANCE_PASSES
4421 && goal_alt[i] != NO_REGS && REG_P (op)
4422 && (regno = REGNO (op)) >= FIRST_PSEUDO_REGISTER
4423 && regno < new_regno_start
4424 && ! ira_former_scratch_p (regno)
4425 && reg_renumber[regno] < 0
4426 /* Check that the optional reload pseudo will be able to
4427 hold given mode value. */
4428 && ! (prohibited_class_reg_set_mode_p
4429 (goal_alt[i], reg_class_contents[goal_alt[i]],
4430 PSEUDO_REGNO_MODE (regno)))
4431 && (curr_insn_set == NULL_RTX
4432 || !((REG_P (SET_SRC (curr_insn_set))
4433 || MEM_P (SET_SRC (curr_insn_set))
4434 || GET_CODE (SET_SRC (curr_insn_set)) == SUBREG)
4435 && (REG_P (SET_DEST (curr_insn_set))
4436 || MEM_P (SET_DEST (curr_insn_set))
4437 || GET_CODE (SET_DEST (curr_insn_set)) == SUBREG))))
4438 optional_p = true;
4439 else if (goal_alt_matched[i][0] != -1
4440 && curr_static_id->operand[i].type == OP_OUT
4441 && (curr_static_id->operand_alternative
4442 [goal_alt_number * n_operands + i].earlyclobber)
4443 && REG_P (op))
4445 for (j = 0; goal_alt_matched[i][j] != -1; j++)
4447 rtx op2 = *curr_id->operand_loc[goal_alt_matched[i][j]];
4449 if (REG_P (op2) && REGNO (op) != REGNO (op2))
4450 break;
4452 if (goal_alt_matched[i][j] != -1)
4454 /* Generate reloads for different output and matched
4455 input registers. This is the easiest way to avoid
4456 creation of non-existing register conflicts in
4457 lra-lives.c. */
4458 match_reload (i, goal_alt_matched[i], outputs, goal_alt[i], &before,
4459 &after, TRUE);
4460 outputs[n_outputs++] = i;
4461 outputs[n_outputs] = -1;
4463 continue;
4465 else
4466 continue;
4469 /* Operands that match previous ones have already been handled. */
4470 if (goal_alt_matches[i] >= 0)
4471 continue;
4473 /* We should not have an operand with a non-offsettable address
4474 appearing where an offsettable address will do. It also may
4475 be a case when the address should be special in other words
4476 not a general one (e.g. it needs no index reg). */
4477 if (goal_alt_matched[i][0] == -1 && goal_alt_offmemok[i] && MEM_P (op))
4479 enum reg_class rclass;
4480 rtx *loc = &XEXP (op, 0);
4481 enum rtx_code code = GET_CODE (*loc);
4483 push_to_sequence (before);
4484 rclass = base_reg_class (GET_MODE (op), MEM_ADDR_SPACE (op),
4485 MEM, SCRATCH);
4486 if (GET_RTX_CLASS (code) == RTX_AUTOINC)
4487 new_reg = emit_inc (rclass, *loc, *loc,
4488 /* This value does not matter for MODIFY. */
4489 GET_MODE_SIZE (GET_MODE (op)));
4490 else if (get_reload_reg (OP_IN, Pmode, *loc, rclass, FALSE,
4491 "offsetable address", &new_reg))
4493 rtx addr = *loc;
4494 enum rtx_code code = GET_CODE (addr);
4495 bool align_p = false;
4497 if (code == AND && CONST_INT_P (XEXP (addr, 1)))
4499 /* (and ... (const_int -X)) is used to align to X bytes. */
4500 align_p = true;
4501 addr = XEXP (*loc, 0);
4503 else
4504 addr = canonicalize_reload_addr (addr);
4506 lra_emit_move (new_reg, addr);
4507 if (align_p)
4508 emit_move_insn (new_reg, gen_rtx_AND (GET_MODE (new_reg), new_reg, XEXP (*loc, 1)));
4510 before = get_insns ();
4511 end_sequence ();
4512 *loc = new_reg;
4513 lra_update_dup (curr_id, i);
4515 else if (goal_alt_matched[i][0] == -1)
4517 machine_mode mode;
4518 rtx reg, *loc;
4519 int hard_regno;
4520 enum op_type type = curr_static_id->operand[i].type;
4522 loc = curr_id->operand_loc[i];
4523 mode = curr_operand_mode[i];
4524 if (GET_CODE (*loc) == SUBREG)
4526 reg = SUBREG_REG (*loc);
4527 poly_int64 byte = SUBREG_BYTE (*loc);
4528 if (REG_P (reg)
4529 /* Strict_low_part requires reloading the register and not
4530 just the subreg. Likewise for a strict subreg no wider
4531 than a word for WORD_REGISTER_OPERATIONS targets. */
4532 && (curr_static_id->operand[i].strict_low
4533 || (!paradoxical_subreg_p (mode, GET_MODE (reg))
4534 && (hard_regno
4535 = get_try_hard_regno (REGNO (reg))) >= 0
4536 && (simplify_subreg_regno
4537 (hard_regno,
4538 GET_MODE (reg), byte, mode) < 0)
4539 && (goal_alt[i] == NO_REGS
4540 || (simplify_subreg_regno
4541 (ira_class_hard_regs[goal_alt[i]][0],
4542 GET_MODE (reg), byte, mode) >= 0)))
4543 || (partial_subreg_p (mode, GET_MODE (reg))
4544 && known_le (GET_MODE_SIZE (GET_MODE (reg)),
4545 UNITS_PER_WORD)
4546 && WORD_REGISTER_OPERATIONS)))
4548 /* An OP_INOUT is required when reloading a subreg of a
4549 mode wider than a word to ensure that data beyond the
4550 word being reloaded is preserved. Also automatically
4551 ensure that strict_low_part reloads are made into
4552 OP_INOUT which should already be true from the backend
4553 constraints. */
4554 if (type == OP_OUT
4555 && (curr_static_id->operand[i].strict_low
4556 || read_modify_subreg_p (*loc)))
4557 type = OP_INOUT;
4558 loc = &SUBREG_REG (*loc);
4559 mode = GET_MODE (*loc);
4562 old = *loc;
4563 if (get_reload_reg (type, mode, old, goal_alt[i],
4564 loc != curr_id->operand_loc[i], "", &new_reg)
4565 && type != OP_OUT)
4567 push_to_sequence (before);
4568 lra_emit_move (new_reg, old);
4569 before = get_insns ();
4570 end_sequence ();
4572 *loc = new_reg;
4573 if (type != OP_IN
4574 && find_reg_note (curr_insn, REG_UNUSED, old) == NULL_RTX)
4576 start_sequence ();
4577 lra_emit_move (type == OP_INOUT ? copy_rtx (old) : old, new_reg);
4578 emit_insn (after);
4579 after = get_insns ();
4580 end_sequence ();
4581 *loc = new_reg;
4583 for (j = 0; j < goal_alt_dont_inherit_ops_num; j++)
4584 if (goal_alt_dont_inherit_ops[j] == i)
4586 lra_set_regno_unique_value (REGNO (new_reg));
4587 break;
4589 lra_update_dup (curr_id, i);
4591 else if (curr_static_id->operand[i].type == OP_IN
4592 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4593 == OP_OUT
4594 || (curr_static_id->operand[goal_alt_matched[i][0]].type
4595 == OP_INOUT
4596 && (operands_match_p
4597 (*curr_id->operand_loc[i],
4598 *curr_id->operand_loc[goal_alt_matched[i][0]],
4599 -1)))))
4601 /* generate reloads for input and matched outputs. */
4602 match_inputs[0] = i;
4603 match_inputs[1] = -1;
4604 match_reload (goal_alt_matched[i][0], match_inputs, outputs,
4605 goal_alt[i], &before, &after,
4606 curr_static_id->operand_alternative
4607 [goal_alt_number * n_operands + goal_alt_matched[i][0]]
4608 .earlyclobber);
4610 else if ((curr_static_id->operand[i].type == OP_OUT
4611 || (curr_static_id->operand[i].type == OP_INOUT
4612 && (operands_match_p
4613 (*curr_id->operand_loc[i],
4614 *curr_id->operand_loc[goal_alt_matched[i][0]],
4615 -1))))
4616 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4617 == OP_IN))
4618 /* Generate reloads for output and matched inputs. */
4619 match_reload (i, goal_alt_matched[i], outputs, goal_alt[i], &before,
4620 &after, curr_static_id->operand_alternative
4621 [goal_alt_number * n_operands + i].earlyclobber);
4622 else if (curr_static_id->operand[i].type == OP_IN
4623 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4624 == OP_IN))
4626 /* Generate reloads for matched inputs. */
4627 match_inputs[0] = i;
4628 for (j = 0; (k = goal_alt_matched[i][j]) >= 0; j++)
4629 match_inputs[j + 1] = k;
4630 match_inputs[j + 1] = -1;
4631 match_reload (-1, match_inputs, outputs, goal_alt[i], &before,
4632 &after, false);
4634 else
4635 /* We must generate code in any case when function
4636 process_alt_operands decides that it is possible. */
4637 gcc_unreachable ();
4639 /* Memorise processed outputs so that output remaining to be processed
4640 can avoid using the same register value (see match_reload). */
4641 if (curr_static_id->operand[i].type == OP_OUT)
4643 outputs[n_outputs++] = i;
4644 outputs[n_outputs] = -1;
4647 if (optional_p)
4649 rtx reg = op;
4651 lra_assert (REG_P (reg));
4652 regno = REGNO (reg);
4653 op = *curr_id->operand_loc[i]; /* Substitution. */
4654 if (GET_CODE (op) == SUBREG)
4655 op = SUBREG_REG (op);
4656 gcc_assert (REG_P (op) && (int) REGNO (op) >= new_regno_start);
4657 bitmap_set_bit (&lra_optional_reload_pseudos, REGNO (op));
4658 lra_reg_info[REGNO (op)].restore_rtx = reg;
4659 if (lra_dump_file != NULL)
4660 fprintf (lra_dump_file,
4661 " Making reload reg %d for reg %d optional\n",
4662 REGNO (op), regno);
4665 if (before != NULL_RTX || after != NULL_RTX
4666 || max_regno_before != max_reg_num ())
4667 change_p = true;
4668 if (change_p)
4670 lra_update_operator_dups (curr_id);
4671 /* Something changes -- process the insn. */
4672 lra_update_insn_regno_info (curr_insn);
4674 lra_process_new_insns (curr_insn, before, after, "Inserting insn reload");
4675 return change_p;
4678 /* Return true if INSN satisfies all constraints. In other words, no
4679 reload insns are needed. */
4680 bool
4681 lra_constrain_insn (rtx_insn *insn)
4683 int saved_new_regno_start = new_regno_start;
4684 int saved_new_insn_uid_start = new_insn_uid_start;
4685 bool change_p;
4687 curr_insn = insn;
4688 curr_id = lra_get_insn_recog_data (curr_insn);
4689 curr_static_id = curr_id->insn_static_data;
4690 new_insn_uid_start = get_max_uid ();
4691 new_regno_start = max_reg_num ();
4692 change_p = curr_insn_transform (true);
4693 new_regno_start = saved_new_regno_start;
4694 new_insn_uid_start = saved_new_insn_uid_start;
4695 return ! change_p;
4698 /* Return true if X is in LIST. */
4699 static bool
4700 in_list_p (rtx x, rtx list)
4702 for (; list != NULL_RTX; list = XEXP (list, 1))
4703 if (XEXP (list, 0) == x)
4704 return true;
4705 return false;
4708 /* Return true if X contains an allocatable hard register (if
4709 HARD_REG_P) or a (spilled if SPILLED_P) pseudo. */
4710 static bool
4711 contains_reg_p (rtx x, bool hard_reg_p, bool spilled_p)
4713 int i, j;
4714 const char *fmt;
4715 enum rtx_code code;
4717 code = GET_CODE (x);
4718 if (REG_P (x))
4720 int regno = REGNO (x);
4721 HARD_REG_SET alloc_regs;
4723 if (hard_reg_p)
4725 if (regno >= FIRST_PSEUDO_REGISTER)
4726 regno = lra_get_regno_hard_regno (regno);
4727 if (regno < 0)
4728 return false;
4729 alloc_regs = ~lra_no_alloc_regs;
4730 return overlaps_hard_reg_set_p (alloc_regs, GET_MODE (x), regno);
4732 else
4734 if (regno < FIRST_PSEUDO_REGISTER)
4735 return false;
4736 if (! spilled_p)
4737 return true;
4738 return lra_get_regno_hard_regno (regno) < 0;
4741 fmt = GET_RTX_FORMAT (code);
4742 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4744 if (fmt[i] == 'e')
4746 if (contains_reg_p (XEXP (x, i), hard_reg_p, spilled_p))
4747 return true;
4749 else if (fmt[i] == 'E')
4751 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4752 if (contains_reg_p (XVECEXP (x, i, j), hard_reg_p, spilled_p))
4753 return true;
4756 return false;
4759 /* Process all regs in location *LOC and change them on equivalent
4760 substitution. Return true if any change was done. */
4761 static bool
4762 loc_equivalence_change_p (rtx *loc)
4764 rtx subst, reg, x = *loc;
4765 bool result = false;
4766 enum rtx_code code = GET_CODE (x);
4767 const char *fmt;
4768 int i, j;
4770 if (code == SUBREG)
4772 reg = SUBREG_REG (x);
4773 if ((subst = get_equiv_with_elimination (reg, curr_insn)) != reg
4774 && GET_MODE (subst) == VOIDmode)
4776 /* We cannot reload debug location. Simplify subreg here
4777 while we know the inner mode. */
4778 *loc = simplify_gen_subreg (GET_MODE (x), subst,
4779 GET_MODE (reg), SUBREG_BYTE (x));
4780 return true;
4783 if (code == REG && (subst = get_equiv_with_elimination (x, curr_insn)) != x)
4785 *loc = subst;
4786 return true;
4789 /* Scan all the operand sub-expressions. */
4790 fmt = GET_RTX_FORMAT (code);
4791 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4793 if (fmt[i] == 'e')
4794 result = loc_equivalence_change_p (&XEXP (x, i)) || result;
4795 else if (fmt[i] == 'E')
4796 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4797 result
4798 = loc_equivalence_change_p (&XVECEXP (x, i, j)) || result;
4800 return result;
4803 /* Similar to loc_equivalence_change_p, but for use as
4804 simplify_replace_fn_rtx callback. DATA is insn for which the
4805 elimination is done. If it null we don't do the elimination. */
4806 static rtx
4807 loc_equivalence_callback (rtx loc, const_rtx, void *data)
4809 if (!REG_P (loc))
4810 return NULL_RTX;
4812 rtx subst = (data == NULL
4813 ? get_equiv (loc) : get_equiv_with_elimination (loc, (rtx_insn *) data));
4814 if (subst != loc)
4815 return subst;
4817 return NULL_RTX;
4820 /* Maximum number of generated reload insns per an insn. It is for
4821 preventing this pass cycling in a bug case. */
4822 #define MAX_RELOAD_INSNS_NUMBER LRA_MAX_INSN_RELOADS
4824 /* The current iteration number of this LRA pass. */
4825 int lra_constraint_iter;
4827 /* True if we should during assignment sub-pass check assignment
4828 correctness for all pseudos and spill some of them to correct
4829 conflicts. It can be necessary when we substitute equiv which
4830 needs checking register allocation correctness because the
4831 equivalent value contains allocatable hard registers, or when we
4832 restore multi-register pseudo, or when we change the insn code and
4833 its operand became INOUT operand when it was IN one before. */
4834 bool check_and_force_assignment_correctness_p;
4836 /* Return true if REGNO is referenced in more than one block. */
4837 static bool
4838 multi_block_pseudo_p (int regno)
4840 basic_block bb = NULL;
4841 unsigned int uid;
4842 bitmap_iterator bi;
4844 if (regno < FIRST_PSEUDO_REGISTER)
4845 return false;
4847 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi)
4848 if (bb == NULL)
4849 bb = BLOCK_FOR_INSN (lra_insn_recog_data[uid]->insn);
4850 else if (BLOCK_FOR_INSN (lra_insn_recog_data[uid]->insn) != bb)
4851 return true;
4852 return false;
4855 /* Return true if LIST contains a deleted insn. */
4856 static bool
4857 contains_deleted_insn_p (rtx_insn_list *list)
4859 for (; list != NULL_RTX; list = list->next ())
4860 if (NOTE_P (list->insn ())
4861 && NOTE_KIND (list->insn ()) == NOTE_INSN_DELETED)
4862 return true;
4863 return false;
4866 /* Return true if X contains a pseudo dying in INSN. */
4867 static bool
4868 dead_pseudo_p (rtx x, rtx_insn *insn)
4870 int i, j;
4871 const char *fmt;
4872 enum rtx_code code;
4874 if (REG_P (x))
4875 return (insn != NULL_RTX
4876 && find_regno_note (insn, REG_DEAD, REGNO (x)) != NULL_RTX);
4877 code = GET_CODE (x);
4878 fmt = GET_RTX_FORMAT (code);
4879 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4881 if (fmt[i] == 'e')
4883 if (dead_pseudo_p (XEXP (x, i), insn))
4884 return true;
4886 else if (fmt[i] == 'E')
4888 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4889 if (dead_pseudo_p (XVECEXP (x, i, j), insn))
4890 return true;
4893 return false;
4896 /* Return true if INSN contains a dying pseudo in INSN right hand
4897 side. */
4898 static bool
4899 insn_rhs_dead_pseudo_p (rtx_insn *insn)
4901 rtx set = single_set (insn);
4903 gcc_assert (set != NULL);
4904 return dead_pseudo_p (SET_SRC (set), insn);
4907 /* Return true if any init insn of REGNO contains a dying pseudo in
4908 insn right hand side. */
4909 static bool
4910 init_insn_rhs_dead_pseudo_p (int regno)
4912 rtx_insn_list *insns = ira_reg_equiv[regno].init_insns;
4914 if (insns == NULL)
4915 return false;
4916 for (; insns != NULL_RTX; insns = insns->next ())
4917 if (insn_rhs_dead_pseudo_p (insns->insn ()))
4918 return true;
4919 return false;
4922 /* Return TRUE if REGNO has a reverse equivalence. The equivalence is
4923 reverse only if we have one init insn with given REGNO as a
4924 source. */
4925 static bool
4926 reverse_equiv_p (int regno)
4928 rtx_insn_list *insns = ira_reg_equiv[regno].init_insns;
4929 rtx set;
4931 if (insns == NULL)
4932 return false;
4933 if (! INSN_P (insns->insn ())
4934 || insns->next () != NULL)
4935 return false;
4936 if ((set = single_set (insns->insn ())) == NULL_RTX)
4937 return false;
4938 return REG_P (SET_SRC (set)) && (int) REGNO (SET_SRC (set)) == regno;
4941 /* Return TRUE if REGNO was reloaded in an equivalence init insn. We
4942 call this function only for non-reverse equivalence. */
4943 static bool
4944 contains_reloaded_insn_p (int regno)
4946 rtx set;
4947 rtx_insn_list *list = ira_reg_equiv[regno].init_insns;
4949 for (; list != NULL; list = list->next ())
4950 if ((set = single_set (list->insn ())) == NULL_RTX
4951 || ! REG_P (SET_DEST (set))
4952 || (int) REGNO (SET_DEST (set)) != regno)
4953 return true;
4954 return false;
4957 /* Entry function of LRA constraint pass. Return true if the
4958 constraint pass did change the code. */
4959 bool
4960 lra_constraints (bool first_p)
4962 bool changed_p;
4963 int i, hard_regno, new_insns_num;
4964 unsigned int min_len, new_min_len, uid;
4965 rtx set, x, reg, dest_reg;
4966 basic_block last_bb;
4967 bitmap_iterator bi;
4969 lra_constraint_iter++;
4970 if (lra_dump_file != NULL)
4971 fprintf (lra_dump_file, "\n********** Local #%d: **********\n\n",
4972 lra_constraint_iter);
4973 changed_p = false;
4974 if (pic_offset_table_rtx
4975 && REGNO (pic_offset_table_rtx) >= FIRST_PSEUDO_REGISTER)
4976 check_and_force_assignment_correctness_p = true;
4977 else if (first_p)
4978 /* On the first iteration we should check IRA assignment
4979 correctness. In rare cases, the assignments can be wrong as
4980 early clobbers operands are ignored in IRA or usages of
4981 paradoxical sub-registers are not taken into account by
4982 IRA. */
4983 check_and_force_assignment_correctness_p = true;
4984 new_insn_uid_start = get_max_uid ();
4985 new_regno_start = first_p ? lra_constraint_new_regno_start : max_reg_num ();
4986 /* Mark used hard regs for target stack size calulations. */
4987 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4988 if (lra_reg_info[i].nrefs != 0
4989 && (hard_regno = lra_get_regno_hard_regno (i)) >= 0)
4991 int j, nregs;
4993 nregs = hard_regno_nregs (hard_regno, lra_reg_info[i].biggest_mode);
4994 for (j = 0; j < nregs; j++)
4995 df_set_regs_ever_live (hard_regno + j, true);
4997 /* Do elimination before the equivalence processing as we can spill
4998 some pseudos during elimination. */
4999 lra_eliminate (false, first_p);
5000 auto_bitmap equiv_insn_bitmap (&reg_obstack);
5001 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
5002 if (lra_reg_info[i].nrefs != 0)
5004 ira_reg_equiv[i].profitable_p = true;
5005 reg = regno_reg_rtx[i];
5006 if (lra_get_regno_hard_regno (i) < 0 && (x = get_equiv (reg)) != reg)
5008 bool pseudo_p = contains_reg_p (x, false, false);
5010 /* After RTL transformation, we cannot guarantee that
5011 pseudo in the substitution was not reloaded which might
5012 make equivalence invalid. For example, in reverse
5013 equiv of p0
5015 p0 <- ...
5017 equiv_mem <- p0
5019 the memory address register was reloaded before the 2nd
5020 insn. */
5021 if ((! first_p && pseudo_p)
5022 /* We don't use DF for compilation speed sake. So it
5023 is problematic to update live info when we use an
5024 equivalence containing pseudos in more than one
5025 BB. */
5026 || (pseudo_p && multi_block_pseudo_p (i))
5027 /* If an init insn was deleted for some reason, cancel
5028 the equiv. We could update the equiv insns after
5029 transformations including an equiv insn deletion
5030 but it is not worthy as such cases are extremely
5031 rare. */
5032 || contains_deleted_insn_p (ira_reg_equiv[i].init_insns)
5033 /* If it is not a reverse equivalence, we check that a
5034 pseudo in rhs of the init insn is not dying in the
5035 insn. Otherwise, the live info at the beginning of
5036 the corresponding BB might be wrong after we
5037 removed the insn. When the equiv can be a
5038 constant, the right hand side of the init insn can
5039 be a pseudo. */
5040 || (! reverse_equiv_p (i)
5041 && (init_insn_rhs_dead_pseudo_p (i)
5042 /* If we reloaded the pseudo in an equivalence
5043 init insn, we cannot remove the equiv init
5044 insns and the init insns might write into
5045 const memory in this case. */
5046 || contains_reloaded_insn_p (i)))
5047 /* Prevent access beyond equivalent memory for
5048 paradoxical subregs. */
5049 || (MEM_P (x)
5050 && maybe_gt (GET_MODE_SIZE (lra_reg_info[i].biggest_mode),
5051 GET_MODE_SIZE (GET_MODE (x))))
5052 || (pic_offset_table_rtx
5053 && ((CONST_POOL_OK_P (PSEUDO_REGNO_MODE (i), x)
5054 && (targetm.preferred_reload_class
5055 (x, lra_get_allocno_class (i)) == NO_REGS))
5056 || contains_symbol_ref_p (x))))
5057 ira_reg_equiv[i].defined_p = false;
5058 if (contains_reg_p (x, false, true))
5059 ira_reg_equiv[i].profitable_p = false;
5060 if (get_equiv (reg) != reg)
5061 bitmap_ior_into (equiv_insn_bitmap, &lra_reg_info[i].insn_bitmap);
5064 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
5065 update_equiv (i);
5066 /* We should add all insns containing pseudos which should be
5067 substituted by their equivalences. */
5068 EXECUTE_IF_SET_IN_BITMAP (equiv_insn_bitmap, 0, uid, bi)
5069 lra_push_insn_by_uid (uid);
5070 min_len = lra_insn_stack_length ();
5071 new_insns_num = 0;
5072 last_bb = NULL;
5073 changed_p = false;
5074 while ((new_min_len = lra_insn_stack_length ()) != 0)
5076 curr_insn = lra_pop_insn ();
5077 --new_min_len;
5078 curr_bb = BLOCK_FOR_INSN (curr_insn);
5079 if (curr_bb != last_bb)
5081 last_bb = curr_bb;
5082 bb_reload_num = lra_curr_reload_num;
5084 if (min_len > new_min_len)
5086 min_len = new_min_len;
5087 new_insns_num = 0;
5089 if (new_insns_num > MAX_RELOAD_INSNS_NUMBER)
5090 internal_error
5091 ("maximum number of generated reload insns per insn achieved (%d)",
5092 MAX_RELOAD_INSNS_NUMBER);
5093 new_insns_num++;
5094 if (DEBUG_INSN_P (curr_insn))
5096 /* We need to check equivalence in debug insn and change
5097 pseudo to the equivalent value if necessary. */
5098 curr_id = lra_get_insn_recog_data (curr_insn);
5099 if (bitmap_bit_p (equiv_insn_bitmap, INSN_UID (curr_insn)))
5101 rtx old = *curr_id->operand_loc[0];
5102 *curr_id->operand_loc[0]
5103 = simplify_replace_fn_rtx (old, NULL_RTX,
5104 loc_equivalence_callback, curr_insn);
5105 if (old != *curr_id->operand_loc[0])
5107 lra_update_insn_regno_info (curr_insn);
5108 changed_p = true;
5112 else if (INSN_P (curr_insn))
5114 if ((set = single_set (curr_insn)) != NULL_RTX)
5116 dest_reg = SET_DEST (set);
5117 /* The equivalence pseudo could be set up as SUBREG in a
5118 case when it is a call restore insn in a mode
5119 different from the pseudo mode. */
5120 if (GET_CODE (dest_reg) == SUBREG)
5121 dest_reg = SUBREG_REG (dest_reg);
5122 if ((REG_P (dest_reg)
5123 && (x = get_equiv (dest_reg)) != dest_reg
5124 /* Remove insns which set up a pseudo whose value
5125 cannot be changed. Such insns might be not in
5126 init_insns because we don't update equiv data
5127 during insn transformations.
5129 As an example, let suppose that a pseudo got
5130 hard register and on the 1st pass was not
5131 changed to equivalent constant. We generate an
5132 additional insn setting up the pseudo because of
5133 secondary memory movement. Then the pseudo is
5134 spilled and we use the equiv constant. In this
5135 case we should remove the additional insn and
5136 this insn is not init_insns list. */
5137 && (! MEM_P (x) || MEM_READONLY_P (x)
5138 /* Check that this is actually an insn setting
5139 up the equivalence. */
5140 || in_list_p (curr_insn,
5141 ira_reg_equiv
5142 [REGNO (dest_reg)].init_insns)))
5143 || (((x = get_equiv (SET_SRC (set))) != SET_SRC (set))
5144 && in_list_p (curr_insn,
5145 ira_reg_equiv
5146 [REGNO (SET_SRC (set))].init_insns)))
5148 /* This is equiv init insn of pseudo which did not get a
5149 hard register -- remove the insn. */
5150 if (lra_dump_file != NULL)
5152 fprintf (lra_dump_file,
5153 " Removing equiv init insn %i (freq=%d)\n",
5154 INSN_UID (curr_insn),
5155 REG_FREQ_FROM_BB (BLOCK_FOR_INSN (curr_insn)));
5156 dump_insn_slim (lra_dump_file, curr_insn);
5158 if (contains_reg_p (x, true, false))
5159 check_and_force_assignment_correctness_p = true;
5160 lra_set_insn_deleted (curr_insn);
5161 continue;
5164 curr_id = lra_get_insn_recog_data (curr_insn);
5165 curr_static_id = curr_id->insn_static_data;
5166 init_curr_insn_input_reloads ();
5167 init_curr_operand_mode ();
5168 if (curr_insn_transform (false))
5169 changed_p = true;
5170 /* Check non-transformed insns too for equiv change as USE
5171 or CLOBBER don't need reloads but can contain pseudos
5172 being changed on their equivalences. */
5173 else if (bitmap_bit_p (equiv_insn_bitmap, INSN_UID (curr_insn))
5174 && loc_equivalence_change_p (&PATTERN (curr_insn)))
5176 lra_update_insn_regno_info (curr_insn);
5177 changed_p = true;
5182 /* If we used a new hard regno, changed_p should be true because the
5183 hard reg is assigned to a new pseudo. */
5184 if (flag_checking && !changed_p)
5186 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
5187 if (lra_reg_info[i].nrefs != 0
5188 && (hard_regno = lra_get_regno_hard_regno (i)) >= 0)
5190 int j, nregs = hard_regno_nregs (hard_regno,
5191 PSEUDO_REGNO_MODE (i));
5193 for (j = 0; j < nregs; j++)
5194 lra_assert (df_regs_ever_live_p (hard_regno + j));
5197 return changed_p;
5200 static void initiate_invariants (void);
5201 static void finish_invariants (void);
5203 /* Initiate the LRA constraint pass. It is done once per
5204 function. */
5205 void
5206 lra_constraints_init (void)
5208 initiate_invariants ();
5211 /* Finalize the LRA constraint pass. It is done once per
5212 function. */
5213 void
5214 lra_constraints_finish (void)
5216 finish_invariants ();
5221 /* Structure describes invariants for ineheritance. */
5222 struct lra_invariant
5224 /* The order number of the invariant. */
5225 int num;
5226 /* The invariant RTX. */
5227 rtx invariant_rtx;
5228 /* The origin insn of the invariant. */
5229 rtx_insn *insn;
5232 typedef lra_invariant invariant_t;
5233 typedef invariant_t *invariant_ptr_t;
5234 typedef const invariant_t *const_invariant_ptr_t;
5236 /* Pointer to the inheritance invariants. */
5237 static vec<invariant_ptr_t> invariants;
5239 /* Allocation pool for the invariants. */
5240 static object_allocator<lra_invariant> *invariants_pool;
5242 /* Hash table for the invariants. */
5243 static htab_t invariant_table;
5245 /* Hash function for INVARIANT. */
5246 static hashval_t
5247 invariant_hash (const void *invariant)
5249 rtx inv = ((const_invariant_ptr_t) invariant)->invariant_rtx;
5250 return lra_rtx_hash (inv);
5253 /* Equal function for invariants INVARIANT1 and INVARIANT2. */
5254 static int
5255 invariant_eq_p (const void *invariant1, const void *invariant2)
5257 rtx inv1 = ((const_invariant_ptr_t) invariant1)->invariant_rtx;
5258 rtx inv2 = ((const_invariant_ptr_t) invariant2)->invariant_rtx;
5260 return rtx_equal_p (inv1, inv2);
5263 /* Insert INVARIANT_RTX into the table if it is not there yet. Return
5264 invariant which is in the table. */
5265 static invariant_ptr_t
5266 insert_invariant (rtx invariant_rtx)
5268 void **entry_ptr;
5269 invariant_t invariant;
5270 invariant_ptr_t invariant_ptr;
5272 invariant.invariant_rtx = invariant_rtx;
5273 entry_ptr = htab_find_slot (invariant_table, &invariant, INSERT);
5274 if (*entry_ptr == NULL)
5276 invariant_ptr = invariants_pool->allocate ();
5277 invariant_ptr->invariant_rtx = invariant_rtx;
5278 invariant_ptr->insn = NULL;
5279 invariants.safe_push (invariant_ptr);
5280 *entry_ptr = (void *) invariant_ptr;
5282 return (invariant_ptr_t) *entry_ptr;
5285 /* Initiate the invariant table. */
5286 static void
5287 initiate_invariants (void)
5289 invariants.create (100);
5290 invariants_pool
5291 = new object_allocator<lra_invariant> ("Inheritance invariants");
5292 invariant_table = htab_create (100, invariant_hash, invariant_eq_p, NULL);
5295 /* Finish the invariant table. */
5296 static void
5297 finish_invariants (void)
5299 htab_delete (invariant_table);
5300 delete invariants_pool;
5301 invariants.release ();
5304 /* Make the invariant table empty. */
5305 static void
5306 clear_invariants (void)
5308 htab_empty (invariant_table);
5309 invariants_pool->release ();
5310 invariants.truncate (0);
5315 /* This page contains code to do inheritance/split
5316 transformations. */
5318 /* Number of reloads passed so far in current EBB. */
5319 static int reloads_num;
5321 /* Number of calls passed so far in current EBB. */
5322 static int calls_num;
5324 /* Index ID is the CALLS_NUM associated the last call we saw with
5325 ABI identifier ID. */
5326 static int last_call_for_abi[NUM_ABI_IDS];
5328 /* Which registers have been fully or partially clobbered by a call
5329 since they were last used. */
5330 static HARD_REG_SET full_and_partial_call_clobbers;
5332 /* Current reload pseudo check for validity of elements in
5333 USAGE_INSNS. */
5334 static int curr_usage_insns_check;
5336 /* Info about last usage of registers in EBB to do inheritance/split
5337 transformation. Inheritance transformation is done from a spilled
5338 pseudo and split transformations from a hard register or a pseudo
5339 assigned to a hard register. */
5340 struct usage_insns
5342 /* If the value is equal to CURR_USAGE_INSNS_CHECK, then the member
5343 value INSNS is valid. The insns is chain of optional debug insns
5344 and a finishing non-debug insn using the corresponding reg. The
5345 value is also used to mark the registers which are set up in the
5346 current insn. The negated insn uid is used for this. */
5347 int check;
5348 /* Value of global reloads_num at the last insn in INSNS. */
5349 int reloads_num;
5350 /* Value of global reloads_nums at the last insn in INSNS. */
5351 int calls_num;
5352 /* It can be true only for splitting. And it means that the restore
5353 insn should be put after insn given by the following member. */
5354 bool after_p;
5355 /* Next insns in the current EBB which use the original reg and the
5356 original reg value is not changed between the current insn and
5357 the next insns. In order words, e.g. for inheritance, if we need
5358 to use the original reg value again in the next insns we can try
5359 to use the value in a hard register from a reload insn of the
5360 current insn. */
5361 rtx insns;
5364 /* Map: regno -> corresponding pseudo usage insns. */
5365 static struct usage_insns *usage_insns;
5367 static void
5368 setup_next_usage_insn (int regno, rtx insn, int reloads_num, bool after_p)
5370 usage_insns[regno].check = curr_usage_insns_check;
5371 usage_insns[regno].insns = insn;
5372 usage_insns[regno].reloads_num = reloads_num;
5373 usage_insns[regno].calls_num = calls_num;
5374 usage_insns[regno].after_p = after_p;
5375 if (regno >= FIRST_PSEUDO_REGISTER && reg_renumber[regno] >= 0)
5376 remove_from_hard_reg_set (&full_and_partial_call_clobbers,
5377 PSEUDO_REGNO_MODE (regno),
5378 reg_renumber[regno]);
5381 /* The function is used to form list REGNO usages which consists of
5382 optional debug insns finished by a non-debug insn using REGNO.
5383 RELOADS_NUM is current number of reload insns processed so far. */
5384 static void
5385 add_next_usage_insn (int regno, rtx_insn *insn, int reloads_num)
5387 rtx next_usage_insns;
5389 if (usage_insns[regno].check == curr_usage_insns_check
5390 && (next_usage_insns = usage_insns[regno].insns) != NULL_RTX
5391 && DEBUG_INSN_P (insn))
5393 /* Check that we did not add the debug insn yet. */
5394 if (next_usage_insns != insn
5395 && (GET_CODE (next_usage_insns) != INSN_LIST
5396 || XEXP (next_usage_insns, 0) != insn))
5397 usage_insns[regno].insns = gen_rtx_INSN_LIST (VOIDmode, insn,
5398 next_usage_insns);
5400 else if (NONDEBUG_INSN_P (insn))
5401 setup_next_usage_insn (regno, insn, reloads_num, false);
5402 else
5403 usage_insns[regno].check = 0;
5406 /* Return first non-debug insn in list USAGE_INSNS. */
5407 static rtx_insn *
5408 skip_usage_debug_insns (rtx usage_insns)
5410 rtx insn;
5412 /* Skip debug insns. */
5413 for (insn = usage_insns;
5414 insn != NULL_RTX && GET_CODE (insn) == INSN_LIST;
5415 insn = XEXP (insn, 1))
5417 return safe_as_a <rtx_insn *> (insn);
5420 /* Return true if we need secondary memory moves for insn in
5421 USAGE_INSNS after inserting inherited pseudo of class INHER_CL
5422 into the insn. */
5423 static bool
5424 check_secondary_memory_needed_p (enum reg_class inher_cl ATTRIBUTE_UNUSED,
5425 rtx usage_insns ATTRIBUTE_UNUSED)
5427 rtx_insn *insn;
5428 rtx set, dest;
5429 enum reg_class cl;
5431 if (inher_cl == ALL_REGS
5432 || (insn = skip_usage_debug_insns (usage_insns)) == NULL_RTX)
5433 return false;
5434 lra_assert (INSN_P (insn));
5435 if ((set = single_set (insn)) == NULL_RTX || ! REG_P (SET_DEST (set)))
5436 return false;
5437 dest = SET_DEST (set);
5438 if (! REG_P (dest))
5439 return false;
5440 lra_assert (inher_cl != NO_REGS);
5441 cl = get_reg_class (REGNO (dest));
5442 return (cl != NO_REGS && cl != ALL_REGS
5443 && targetm.secondary_memory_needed (GET_MODE (dest), inher_cl, cl));
5446 /* Registers involved in inheritance/split in the current EBB
5447 (inheritance/split pseudos and original registers). */
5448 static bitmap_head check_only_regs;
5450 /* Reload pseudos cannot be involded in invariant inheritance in the
5451 current EBB. */
5452 static bitmap_head invalid_invariant_regs;
5454 /* Do inheritance transformations for insn INSN, which defines (if
5455 DEF_P) or uses ORIGINAL_REGNO. NEXT_USAGE_INSNS specifies which
5456 instruction in the EBB next uses ORIGINAL_REGNO; it has the same
5457 form as the "insns" field of usage_insns. Return true if we
5458 succeed in such transformation.
5460 The transformations look like:
5462 p <- ... i <- ...
5463 ... p <- i (new insn)
5464 ... =>
5465 <- ... p ... <- ... i ...
5467 ... i <- p (new insn)
5468 <- ... p ... <- ... i ...
5469 ... =>
5470 <- ... p ... <- ... i ...
5471 where p is a spilled original pseudo and i is a new inheritance pseudo.
5474 The inheritance pseudo has the smallest class of two classes CL and
5475 class of ORIGINAL REGNO. */
5476 static bool
5477 inherit_reload_reg (bool def_p, int original_regno,
5478 enum reg_class cl, rtx_insn *insn, rtx next_usage_insns)
5480 if (optimize_function_for_size_p (cfun))
5481 return false;
5483 enum reg_class rclass = lra_get_allocno_class (original_regno);
5484 rtx original_reg = regno_reg_rtx[original_regno];
5485 rtx new_reg, usage_insn;
5486 rtx_insn *new_insns;
5488 lra_assert (! usage_insns[original_regno].after_p);
5489 if (lra_dump_file != NULL)
5490 fprintf (lra_dump_file,
5491 " <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<\n");
5492 if (! ira_reg_classes_intersect_p[cl][rclass])
5494 if (lra_dump_file != NULL)
5496 fprintf (lra_dump_file,
5497 " Rejecting inheritance for %d "
5498 "because of disjoint classes %s and %s\n",
5499 original_regno, reg_class_names[cl],
5500 reg_class_names[rclass]);
5501 fprintf (lra_dump_file,
5502 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5504 return false;
5506 if ((ira_class_subset_p[cl][rclass] && cl != rclass)
5507 /* We don't use a subset of two classes because it can be
5508 NO_REGS. This transformation is still profitable in most
5509 cases even if the classes are not intersected as register
5510 move is probably cheaper than a memory load. */
5511 || ira_class_hard_regs_num[cl] < ira_class_hard_regs_num[rclass])
5513 if (lra_dump_file != NULL)
5514 fprintf (lra_dump_file, " Use smallest class of %s and %s\n",
5515 reg_class_names[cl], reg_class_names[rclass]);
5517 rclass = cl;
5519 if (check_secondary_memory_needed_p (rclass, next_usage_insns))
5521 /* Reject inheritance resulting in secondary memory moves.
5522 Otherwise, there is a danger in LRA cycling. Also such
5523 transformation will be unprofitable. */
5524 if (lra_dump_file != NULL)
5526 rtx_insn *insn = skip_usage_debug_insns (next_usage_insns);
5527 rtx set = single_set (insn);
5529 lra_assert (set != NULL_RTX);
5531 rtx dest = SET_DEST (set);
5533 lra_assert (REG_P (dest));
5534 fprintf (lra_dump_file,
5535 " Rejecting inheritance for insn %d(%s)<-%d(%s) "
5536 "as secondary mem is needed\n",
5537 REGNO (dest), reg_class_names[get_reg_class (REGNO (dest))],
5538 original_regno, reg_class_names[rclass]);
5539 fprintf (lra_dump_file,
5540 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5542 return false;
5544 new_reg = lra_create_new_reg (GET_MODE (original_reg), original_reg,
5545 rclass, "inheritance");
5546 start_sequence ();
5547 if (def_p)
5548 lra_emit_move (original_reg, new_reg);
5549 else
5550 lra_emit_move (new_reg, original_reg);
5551 new_insns = get_insns ();
5552 end_sequence ();
5553 if (NEXT_INSN (new_insns) != NULL_RTX)
5555 if (lra_dump_file != NULL)
5557 fprintf (lra_dump_file,
5558 " Rejecting inheritance %d->%d "
5559 "as it results in 2 or more insns:\n",
5560 original_regno, REGNO (new_reg));
5561 dump_rtl_slim (lra_dump_file, new_insns, NULL, -1, 0);
5562 fprintf (lra_dump_file,
5563 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5565 return false;
5567 lra_substitute_pseudo_within_insn (insn, original_regno, new_reg, false);
5568 lra_update_insn_regno_info (insn);
5569 if (! def_p)
5570 /* We now have a new usage insn for original regno. */
5571 setup_next_usage_insn (original_regno, new_insns, reloads_num, false);
5572 if (lra_dump_file != NULL)
5573 fprintf (lra_dump_file, " Original reg change %d->%d (bb%d):\n",
5574 original_regno, REGNO (new_reg), BLOCK_FOR_INSN (insn)->index);
5575 lra_reg_info[REGNO (new_reg)].restore_rtx = regno_reg_rtx[original_regno];
5576 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
5577 bitmap_set_bit (&check_only_regs, original_regno);
5578 bitmap_set_bit (&lra_inheritance_pseudos, REGNO (new_reg));
5579 if (def_p)
5580 lra_process_new_insns (insn, NULL, new_insns,
5581 "Add original<-inheritance");
5582 else
5583 lra_process_new_insns (insn, new_insns, NULL,
5584 "Add inheritance<-original");
5585 while (next_usage_insns != NULL_RTX)
5587 if (GET_CODE (next_usage_insns) != INSN_LIST)
5589 usage_insn = next_usage_insns;
5590 lra_assert (NONDEBUG_INSN_P (usage_insn));
5591 next_usage_insns = NULL;
5593 else
5595 usage_insn = XEXP (next_usage_insns, 0);
5596 lra_assert (DEBUG_INSN_P (usage_insn));
5597 next_usage_insns = XEXP (next_usage_insns, 1);
5599 lra_substitute_pseudo (&usage_insn, original_regno, new_reg, false,
5600 DEBUG_INSN_P (usage_insn));
5601 lra_update_insn_regno_info (as_a <rtx_insn *> (usage_insn));
5602 if (lra_dump_file != NULL)
5604 basic_block bb = BLOCK_FOR_INSN (usage_insn);
5605 fprintf (lra_dump_file,
5606 " Inheritance reuse change %d->%d (bb%d):\n",
5607 original_regno, REGNO (new_reg),
5608 bb ? bb->index : -1);
5609 dump_insn_slim (lra_dump_file, as_a <rtx_insn *> (usage_insn));
5612 if (lra_dump_file != NULL)
5613 fprintf (lra_dump_file,
5614 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5615 return true;
5618 /* Return true if we need a caller save/restore for pseudo REGNO which
5619 was assigned to a hard register. */
5620 static inline bool
5621 need_for_call_save_p (int regno)
5623 lra_assert (regno >= FIRST_PSEUDO_REGISTER && reg_renumber[regno] >= 0);
5624 if (usage_insns[regno].calls_num < calls_num)
5626 unsigned int abis = 0;
5627 for (unsigned int i = 0; i < NUM_ABI_IDS; ++i)
5628 if (last_call_for_abi[i] > usage_insns[regno].calls_num)
5629 abis |= 1 << i;
5630 gcc_assert (abis);
5631 if (call_clobbered_in_region_p (abis, full_and_partial_call_clobbers,
5632 PSEUDO_REGNO_MODE (regno),
5633 reg_renumber[regno]))
5634 return true;
5636 return false;
5639 /* Global registers occurring in the current EBB. */
5640 static bitmap_head ebb_global_regs;
5642 /* Return true if we need a split for hard register REGNO or pseudo
5643 REGNO which was assigned to a hard register.
5644 POTENTIAL_RELOAD_HARD_REGS contains hard registers which might be
5645 used for reloads since the EBB end. It is an approximation of the
5646 used hard registers in the split range. The exact value would
5647 require expensive calculations. If we were aggressive with
5648 splitting because of the approximation, the split pseudo will save
5649 the same hard register assignment and will be removed in the undo
5650 pass. We still need the approximation because too aggressive
5651 splitting would result in too inaccurate cost calculation in the
5652 assignment pass because of too many generated moves which will be
5653 probably removed in the undo pass. */
5654 static inline bool
5655 need_for_split_p (HARD_REG_SET potential_reload_hard_regs, int regno)
5657 int hard_regno = regno < FIRST_PSEUDO_REGISTER ? regno : reg_renumber[regno];
5659 lra_assert (hard_regno >= 0);
5660 return ((TEST_HARD_REG_BIT (potential_reload_hard_regs, hard_regno)
5661 /* Don't split eliminable hard registers, otherwise we can
5662 split hard registers like hard frame pointer, which
5663 lives on BB start/end according to DF-infrastructure,
5664 when there is a pseudo assigned to the register and
5665 living in the same BB. */
5666 && (regno >= FIRST_PSEUDO_REGISTER
5667 || ! TEST_HARD_REG_BIT (eliminable_regset, hard_regno))
5668 && ! TEST_HARD_REG_BIT (lra_no_alloc_regs, hard_regno)
5669 /* Don't split call clobbered hard regs living through
5670 calls, otherwise we might have a check problem in the
5671 assign sub-pass as in the most cases (exception is a
5672 situation when check_and_force_assignment_correctness_p value is
5673 true) the assign pass assumes that all pseudos living
5674 through calls are assigned to call saved hard regs. */
5675 && (regno >= FIRST_PSEUDO_REGISTER
5676 || !TEST_HARD_REG_BIT (full_and_partial_call_clobbers, regno))
5677 /* We need at least 2 reloads to make pseudo splitting
5678 profitable. We should provide hard regno splitting in
5679 any case to solve 1st insn scheduling problem when
5680 moving hard register definition up might result in
5681 impossibility to find hard register for reload pseudo of
5682 small register class. */
5683 && (usage_insns[regno].reloads_num
5684 + (regno < FIRST_PSEUDO_REGISTER ? 0 : 3) < reloads_num)
5685 && (regno < FIRST_PSEUDO_REGISTER
5686 /* For short living pseudos, spilling + inheritance can
5687 be considered a substitution for splitting.
5688 Therefore we do not splitting for local pseudos. It
5689 decreases also aggressiveness of splitting. The
5690 minimal number of references is chosen taking into
5691 account that for 2 references splitting has no sense
5692 as we can just spill the pseudo. */
5693 || (regno >= FIRST_PSEUDO_REGISTER
5694 && lra_reg_info[regno].nrefs > 3
5695 && bitmap_bit_p (&ebb_global_regs, regno))))
5696 || (regno >= FIRST_PSEUDO_REGISTER && need_for_call_save_p (regno)));
5699 /* Return class for the split pseudo created from original pseudo with
5700 ALLOCNO_CLASS and MODE which got a hard register HARD_REGNO. We
5701 choose subclass of ALLOCNO_CLASS which contains HARD_REGNO and
5702 results in no secondary memory movements. */
5703 static enum reg_class
5704 choose_split_class (enum reg_class allocno_class,
5705 int hard_regno ATTRIBUTE_UNUSED,
5706 machine_mode mode ATTRIBUTE_UNUSED)
5708 int i;
5709 enum reg_class cl, best_cl = NO_REGS;
5710 enum reg_class hard_reg_class ATTRIBUTE_UNUSED
5711 = REGNO_REG_CLASS (hard_regno);
5713 if (! targetm.secondary_memory_needed (mode, allocno_class, allocno_class)
5714 && TEST_HARD_REG_BIT (reg_class_contents[allocno_class], hard_regno))
5715 return allocno_class;
5716 for (i = 0;
5717 (cl = reg_class_subclasses[allocno_class][i]) != LIM_REG_CLASSES;
5718 i++)
5719 if (! targetm.secondary_memory_needed (mode, cl, hard_reg_class)
5720 && ! targetm.secondary_memory_needed (mode, hard_reg_class, cl)
5721 && TEST_HARD_REG_BIT (reg_class_contents[cl], hard_regno)
5722 && (best_cl == NO_REGS
5723 || ira_class_hard_regs_num[best_cl] < ira_class_hard_regs_num[cl]))
5724 best_cl = cl;
5725 return best_cl;
5728 /* Copy any equivalence information from ORIGINAL_REGNO to NEW_REGNO.
5729 It only makes sense to call this function if NEW_REGNO is always
5730 equal to ORIGINAL_REGNO. */
5732 static void
5733 lra_copy_reg_equiv (unsigned int new_regno, unsigned int original_regno)
5735 if (!ira_reg_equiv[original_regno].defined_p)
5736 return;
5738 ira_expand_reg_equiv ();
5739 ira_reg_equiv[new_regno].defined_p = true;
5740 if (ira_reg_equiv[original_regno].memory)
5741 ira_reg_equiv[new_regno].memory
5742 = copy_rtx (ira_reg_equiv[original_regno].memory);
5743 if (ira_reg_equiv[original_regno].constant)
5744 ira_reg_equiv[new_regno].constant
5745 = copy_rtx (ira_reg_equiv[original_regno].constant);
5746 if (ira_reg_equiv[original_regno].invariant)
5747 ira_reg_equiv[new_regno].invariant
5748 = copy_rtx (ira_reg_equiv[original_regno].invariant);
5751 /* Do split transformations for insn INSN, which defines or uses
5752 ORIGINAL_REGNO. NEXT_USAGE_INSNS specifies which instruction in
5753 the EBB next uses ORIGINAL_REGNO; it has the same form as the
5754 "insns" field of usage_insns. If TO is not NULL, we don't use
5755 usage_insns, we put restore insns after TO insn. It is a case when
5756 we call it from lra_split_hard_reg_for, outside the inheritance
5757 pass.
5759 The transformations look like:
5761 p <- ... p <- ...
5762 ... s <- p (new insn -- save)
5763 ... =>
5764 ... p <- s (new insn -- restore)
5765 <- ... p ... <- ... p ...
5767 <- ... p ... <- ... p ...
5768 ... s <- p (new insn -- save)
5769 ... =>
5770 ... p <- s (new insn -- restore)
5771 <- ... p ... <- ... p ...
5773 where p is an original pseudo got a hard register or a hard
5774 register and s is a new split pseudo. The save is put before INSN
5775 if BEFORE_P is true. Return true if we succeed in such
5776 transformation. */
5777 static bool
5778 split_reg (bool before_p, int original_regno, rtx_insn *insn,
5779 rtx next_usage_insns, rtx_insn *to)
5781 enum reg_class rclass;
5782 rtx original_reg;
5783 int hard_regno, nregs;
5784 rtx new_reg, usage_insn;
5785 rtx_insn *restore, *save;
5786 bool after_p;
5787 bool call_save_p;
5788 machine_mode mode;
5790 if (original_regno < FIRST_PSEUDO_REGISTER)
5792 rclass = ira_allocno_class_translate[REGNO_REG_CLASS (original_regno)];
5793 hard_regno = original_regno;
5794 call_save_p = false;
5795 nregs = 1;
5796 mode = lra_reg_info[hard_regno].biggest_mode;
5797 machine_mode reg_rtx_mode = GET_MODE (regno_reg_rtx[hard_regno]);
5798 /* A reg can have a biggest_mode of VOIDmode if it was only ever seen as
5799 part of a multi-word register. In that case, just use the reg_rtx
5800 mode. Do the same also if the biggest mode was larger than a register
5801 or we can not compare the modes. Otherwise, limit the size to that of
5802 the biggest access in the function. */
5803 if (mode == VOIDmode
5804 || !ordered_p (GET_MODE_PRECISION (mode),
5805 GET_MODE_PRECISION (reg_rtx_mode))
5806 || paradoxical_subreg_p (mode, reg_rtx_mode))
5808 original_reg = regno_reg_rtx[hard_regno];
5809 mode = reg_rtx_mode;
5811 else
5812 original_reg = gen_rtx_REG (mode, hard_regno);
5814 else
5816 mode = PSEUDO_REGNO_MODE (original_regno);
5817 hard_regno = reg_renumber[original_regno];
5818 nregs = hard_regno_nregs (hard_regno, mode);
5819 rclass = lra_get_allocno_class (original_regno);
5820 original_reg = regno_reg_rtx[original_regno];
5821 call_save_p = need_for_call_save_p (original_regno);
5823 lra_assert (hard_regno >= 0);
5824 if (lra_dump_file != NULL)
5825 fprintf (lra_dump_file,
5826 " ((((((((((((((((((((((((((((((((((((((((((((((((\n");
5828 if (call_save_p)
5830 mode = HARD_REGNO_CALLER_SAVE_MODE (hard_regno,
5831 hard_regno_nregs (hard_regno, mode),
5832 mode);
5833 new_reg = lra_create_new_reg (mode, NULL_RTX, NO_REGS, "save");
5835 else
5837 rclass = choose_split_class (rclass, hard_regno, mode);
5838 if (rclass == NO_REGS)
5840 if (lra_dump_file != NULL)
5842 fprintf (lra_dump_file,
5843 " Rejecting split of %d(%s): "
5844 "no good reg class for %d(%s)\n",
5845 original_regno,
5846 reg_class_names[lra_get_allocno_class (original_regno)],
5847 hard_regno,
5848 reg_class_names[REGNO_REG_CLASS (hard_regno)]);
5849 fprintf
5850 (lra_dump_file,
5851 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5853 return false;
5855 /* Split_if_necessary can split hard registers used as part of a
5856 multi-register mode but splits each register individually. The
5857 mode used for each independent register may not be supported
5858 so reject the split. Splitting the wider mode should theoretically
5859 be possible but is not implemented. */
5860 if (!targetm.hard_regno_mode_ok (hard_regno, mode))
5862 if (lra_dump_file != NULL)
5864 fprintf (lra_dump_file,
5865 " Rejecting split of %d(%s): unsuitable mode %s\n",
5866 original_regno,
5867 reg_class_names[lra_get_allocno_class (original_regno)],
5868 GET_MODE_NAME (mode));
5869 fprintf
5870 (lra_dump_file,
5871 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5873 return false;
5875 new_reg = lra_create_new_reg (mode, original_reg, rclass, "split");
5876 reg_renumber[REGNO (new_reg)] = hard_regno;
5878 int new_regno = REGNO (new_reg);
5879 save = emit_spill_move (true, new_reg, original_reg);
5880 if (NEXT_INSN (save) != NULL_RTX && !call_save_p)
5882 if (lra_dump_file != NULL)
5884 fprintf
5885 (lra_dump_file,
5886 " Rejecting split %d->%d resulting in > 2 save insns:\n",
5887 original_regno, new_regno);
5888 dump_rtl_slim (lra_dump_file, save, NULL, -1, 0);
5889 fprintf (lra_dump_file,
5890 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5892 return false;
5894 restore = emit_spill_move (false, new_reg, original_reg);
5895 if (NEXT_INSN (restore) != NULL_RTX && !call_save_p)
5897 if (lra_dump_file != NULL)
5899 fprintf (lra_dump_file,
5900 " Rejecting split %d->%d "
5901 "resulting in > 2 restore insns:\n",
5902 original_regno, new_regno);
5903 dump_rtl_slim (lra_dump_file, restore, NULL, -1, 0);
5904 fprintf (lra_dump_file,
5905 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5907 return false;
5909 /* Transfer equivalence information to the spill register, so that
5910 if we fail to allocate the spill register, we have the option of
5911 rematerializing the original value instead of spilling to the stack. */
5912 if (!HARD_REGISTER_NUM_P (original_regno)
5913 && mode == PSEUDO_REGNO_MODE (original_regno))
5914 lra_copy_reg_equiv (new_regno, original_regno);
5915 lra_reg_info[new_regno].restore_rtx = regno_reg_rtx[original_regno];
5916 bitmap_set_bit (&lra_split_regs, new_regno);
5917 if (to != NULL)
5919 lra_assert (next_usage_insns == NULL);
5920 usage_insn = to;
5921 after_p = TRUE;
5923 else
5925 /* We need check_only_regs only inside the inheritance pass. */
5926 bitmap_set_bit (&check_only_regs, new_regno);
5927 bitmap_set_bit (&check_only_regs, original_regno);
5928 after_p = usage_insns[original_regno].after_p;
5929 for (;;)
5931 if (GET_CODE (next_usage_insns) != INSN_LIST)
5933 usage_insn = next_usage_insns;
5934 break;
5936 usage_insn = XEXP (next_usage_insns, 0);
5937 lra_assert (DEBUG_INSN_P (usage_insn));
5938 next_usage_insns = XEXP (next_usage_insns, 1);
5939 lra_substitute_pseudo (&usage_insn, original_regno, new_reg, false,
5940 true);
5941 lra_update_insn_regno_info (as_a <rtx_insn *> (usage_insn));
5942 if (lra_dump_file != NULL)
5944 fprintf (lra_dump_file, " Split reuse change %d->%d:\n",
5945 original_regno, new_regno);
5946 dump_insn_slim (lra_dump_file, as_a <rtx_insn *> (usage_insn));
5950 lra_assert (NOTE_P (usage_insn) || NONDEBUG_INSN_P (usage_insn));
5951 lra_assert (usage_insn != insn || (after_p && before_p));
5952 lra_process_new_insns (as_a <rtx_insn *> (usage_insn),
5953 after_p ? NULL : restore,
5954 after_p ? restore : NULL,
5955 call_save_p
5956 ? "Add reg<-save" : "Add reg<-split");
5957 lra_process_new_insns (insn, before_p ? save : NULL,
5958 before_p ? NULL : save,
5959 call_save_p
5960 ? "Add save<-reg" : "Add split<-reg");
5961 if (nregs > 1)
5962 /* If we are trying to split multi-register. We should check
5963 conflicts on the next assignment sub-pass. IRA can allocate on
5964 sub-register levels, LRA do this on pseudos level right now and
5965 this discrepancy may create allocation conflicts after
5966 splitting. */
5967 check_and_force_assignment_correctness_p = true;
5968 if (lra_dump_file != NULL)
5969 fprintf (lra_dump_file,
5970 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5971 return true;
5974 /* Split a hard reg for reload pseudo REGNO having RCLASS and living
5975 in the range [FROM, TO]. Return true if did a split. Otherwise,
5976 return false. */
5977 bool
5978 spill_hard_reg_in_range (int regno, enum reg_class rclass, rtx_insn *from, rtx_insn *to)
5980 int i, hard_regno;
5981 int rclass_size;
5982 rtx_insn *insn;
5983 unsigned int uid;
5984 bitmap_iterator bi;
5985 HARD_REG_SET ignore;
5987 lra_assert (from != NULL && to != NULL);
5988 CLEAR_HARD_REG_SET (ignore);
5989 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi)
5991 lra_insn_recog_data_t id = lra_insn_recog_data[uid];
5992 struct lra_static_insn_data *static_id = id->insn_static_data;
5993 struct lra_insn_reg *reg;
5995 for (reg = id->regs; reg != NULL; reg = reg->next)
5996 if (reg->regno < FIRST_PSEUDO_REGISTER)
5997 SET_HARD_REG_BIT (ignore, reg->regno);
5998 for (reg = static_id->hard_regs; reg != NULL; reg = reg->next)
5999 SET_HARD_REG_BIT (ignore, reg->regno);
6001 rclass_size = ira_class_hard_regs_num[rclass];
6002 for (i = 0; i < rclass_size; i++)
6004 hard_regno = ira_class_hard_regs[rclass][i];
6005 if (! TEST_HARD_REG_BIT (lra_reg_info[regno].conflict_hard_regs, hard_regno)
6006 || TEST_HARD_REG_BIT (ignore, hard_regno))
6007 continue;
6008 for (insn = from; insn != NEXT_INSN (to); insn = NEXT_INSN (insn))
6010 struct lra_static_insn_data *static_id;
6011 struct lra_insn_reg *reg;
6013 if (!INSN_P (insn))
6014 continue;
6015 if (bitmap_bit_p (&lra_reg_info[hard_regno].insn_bitmap,
6016 INSN_UID (insn)))
6017 break;
6018 static_id = lra_get_insn_recog_data (insn)->insn_static_data;
6019 for (reg = static_id->hard_regs; reg != NULL; reg = reg->next)
6020 if (reg->regno == hard_regno)
6021 break;
6022 if (reg != NULL)
6023 break;
6025 if (insn != NEXT_INSN (to))
6026 continue;
6027 if (split_reg (TRUE, hard_regno, from, NULL, to))
6028 return true;
6030 return false;
6033 /* Recognize that we need a split transformation for insn INSN, which
6034 defines or uses REGNO in its insn biggest MODE (we use it only if
6035 REGNO is a hard register). POTENTIAL_RELOAD_HARD_REGS contains
6036 hard registers which might be used for reloads since the EBB end.
6037 Put the save before INSN if BEFORE_P is true. MAX_UID is maximla
6038 uid before starting INSN processing. Return true if we succeed in
6039 such transformation. */
6040 static bool
6041 split_if_necessary (int regno, machine_mode mode,
6042 HARD_REG_SET potential_reload_hard_regs,
6043 bool before_p, rtx_insn *insn, int max_uid)
6045 bool res = false;
6046 int i, nregs = 1;
6047 rtx next_usage_insns;
6049 if (regno < FIRST_PSEUDO_REGISTER)
6050 nregs = hard_regno_nregs (regno, mode);
6051 for (i = 0; i < nregs; i++)
6052 if (usage_insns[regno + i].check == curr_usage_insns_check
6053 && (next_usage_insns = usage_insns[regno + i].insns) != NULL_RTX
6054 /* To avoid processing the register twice or more. */
6055 && ((GET_CODE (next_usage_insns) != INSN_LIST
6056 && INSN_UID (next_usage_insns) < max_uid)
6057 || (GET_CODE (next_usage_insns) == INSN_LIST
6058 && (INSN_UID (XEXP (next_usage_insns, 0)) < max_uid)))
6059 && need_for_split_p (potential_reload_hard_regs, regno + i)
6060 && split_reg (before_p, regno + i, insn, next_usage_insns, NULL))
6061 res = true;
6062 return res;
6065 /* Return TRUE if rtx X is considered as an invariant for
6066 inheritance. */
6067 static bool
6068 invariant_p (const_rtx x)
6070 machine_mode mode;
6071 const char *fmt;
6072 enum rtx_code code;
6073 int i, j;
6075 if (side_effects_p (x))
6076 return false;
6078 code = GET_CODE (x);
6079 mode = GET_MODE (x);
6080 if (code == SUBREG)
6082 x = SUBREG_REG (x);
6083 code = GET_CODE (x);
6084 mode = wider_subreg_mode (mode, GET_MODE (x));
6087 if (MEM_P (x))
6088 return false;
6090 if (REG_P (x))
6092 int i, nregs, regno = REGNO (x);
6094 if (regno >= FIRST_PSEUDO_REGISTER || regno == STACK_POINTER_REGNUM
6095 || TEST_HARD_REG_BIT (eliminable_regset, regno)
6096 || GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
6097 return false;
6098 nregs = hard_regno_nregs (regno, mode);
6099 for (i = 0; i < nregs; i++)
6100 if (! fixed_regs[regno + i]
6101 /* A hard register may be clobbered in the current insn
6102 but we can ignore this case because if the hard
6103 register is used it should be set somewhere after the
6104 clobber. */
6105 || bitmap_bit_p (&invalid_invariant_regs, regno + i))
6106 return false;
6108 fmt = GET_RTX_FORMAT (code);
6109 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
6111 if (fmt[i] == 'e')
6113 if (! invariant_p (XEXP (x, i)))
6114 return false;
6116 else if (fmt[i] == 'E')
6118 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
6119 if (! invariant_p (XVECEXP (x, i, j)))
6120 return false;
6123 return true;
6126 /* We have 'dest_reg <- invariant'. Let us try to make an invariant
6127 inheritance transformation (using dest_reg instead invariant in a
6128 subsequent insn). */
6129 static bool
6130 process_invariant_for_inheritance (rtx dst_reg, rtx invariant_rtx)
6132 invariant_ptr_t invariant_ptr;
6133 rtx_insn *insn, *new_insns;
6134 rtx insn_set, insn_reg, new_reg;
6135 int insn_regno;
6136 bool succ_p = false;
6137 int dst_regno = REGNO (dst_reg);
6138 machine_mode dst_mode = GET_MODE (dst_reg);
6139 enum reg_class cl = lra_get_allocno_class (dst_regno), insn_reg_cl;
6141 invariant_ptr = insert_invariant (invariant_rtx);
6142 if ((insn = invariant_ptr->insn) != NULL_RTX)
6144 /* We have a subsequent insn using the invariant. */
6145 insn_set = single_set (insn);
6146 lra_assert (insn_set != NULL);
6147 insn_reg = SET_DEST (insn_set);
6148 lra_assert (REG_P (insn_reg));
6149 insn_regno = REGNO (insn_reg);
6150 insn_reg_cl = lra_get_allocno_class (insn_regno);
6152 if (dst_mode == GET_MODE (insn_reg)
6153 /* We should consider only result move reg insns which are
6154 cheap. */
6155 && targetm.register_move_cost (dst_mode, cl, insn_reg_cl) == 2
6156 && targetm.register_move_cost (dst_mode, cl, cl) == 2)
6158 if (lra_dump_file != NULL)
6159 fprintf (lra_dump_file,
6160 " [[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[\n");
6161 new_reg = lra_create_new_reg (dst_mode, dst_reg,
6162 cl, "invariant inheritance");
6163 bitmap_set_bit (&lra_inheritance_pseudos, REGNO (new_reg));
6164 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
6165 lra_reg_info[REGNO (new_reg)].restore_rtx = PATTERN (insn);
6166 start_sequence ();
6167 lra_emit_move (new_reg, dst_reg);
6168 new_insns = get_insns ();
6169 end_sequence ();
6170 lra_process_new_insns (curr_insn, NULL, new_insns,
6171 "Add invariant inheritance<-original");
6172 start_sequence ();
6173 lra_emit_move (SET_DEST (insn_set), new_reg);
6174 new_insns = get_insns ();
6175 end_sequence ();
6176 lra_process_new_insns (insn, NULL, new_insns,
6177 "Changing reload<-inheritance");
6178 lra_set_insn_deleted (insn);
6179 succ_p = true;
6180 if (lra_dump_file != NULL)
6182 fprintf (lra_dump_file,
6183 " Invariant inheritance reuse change %d (bb%d):\n",
6184 REGNO (new_reg), BLOCK_FOR_INSN (insn)->index);
6185 dump_insn_slim (lra_dump_file, insn);
6186 fprintf (lra_dump_file,
6187 " ]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]\n");
6191 invariant_ptr->insn = curr_insn;
6192 return succ_p;
6195 /* Check only registers living at the current program point in the
6196 current EBB. */
6197 static bitmap_head live_regs;
6199 /* Update live info in EBB given by its HEAD and TAIL insns after
6200 inheritance/split transformation. The function removes dead moves
6201 too. */
6202 static void
6203 update_ebb_live_info (rtx_insn *head, rtx_insn *tail)
6205 unsigned int j;
6206 int i, regno;
6207 bool live_p;
6208 rtx_insn *prev_insn;
6209 rtx set;
6210 bool remove_p;
6211 basic_block last_bb, prev_bb, curr_bb;
6212 bitmap_iterator bi;
6213 struct lra_insn_reg *reg;
6214 edge e;
6215 edge_iterator ei;
6217 last_bb = BLOCK_FOR_INSN (tail);
6218 prev_bb = NULL;
6219 for (curr_insn = tail;
6220 curr_insn != PREV_INSN (head);
6221 curr_insn = prev_insn)
6223 prev_insn = PREV_INSN (curr_insn);
6224 /* We need to process empty blocks too. They contain
6225 NOTE_INSN_BASIC_BLOCK referring for the basic block. */
6226 if (NOTE_P (curr_insn) && NOTE_KIND (curr_insn) != NOTE_INSN_BASIC_BLOCK)
6227 continue;
6228 curr_bb = BLOCK_FOR_INSN (curr_insn);
6229 if (curr_bb != prev_bb)
6231 if (prev_bb != NULL)
6233 /* Update df_get_live_in (prev_bb): */
6234 EXECUTE_IF_SET_IN_BITMAP (&check_only_regs, 0, j, bi)
6235 if (bitmap_bit_p (&live_regs, j))
6236 bitmap_set_bit (df_get_live_in (prev_bb), j);
6237 else
6238 bitmap_clear_bit (df_get_live_in (prev_bb), j);
6240 if (curr_bb != last_bb)
6242 /* Update df_get_live_out (curr_bb): */
6243 EXECUTE_IF_SET_IN_BITMAP (&check_only_regs, 0, j, bi)
6245 live_p = bitmap_bit_p (&live_regs, j);
6246 if (! live_p)
6247 FOR_EACH_EDGE (e, ei, curr_bb->succs)
6248 if (bitmap_bit_p (df_get_live_in (e->dest), j))
6250 live_p = true;
6251 break;
6253 if (live_p)
6254 bitmap_set_bit (df_get_live_out (curr_bb), j);
6255 else
6256 bitmap_clear_bit (df_get_live_out (curr_bb), j);
6259 prev_bb = curr_bb;
6260 bitmap_and (&live_regs, &check_only_regs, df_get_live_out (curr_bb));
6262 if (! NONDEBUG_INSN_P (curr_insn))
6263 continue;
6264 curr_id = lra_get_insn_recog_data (curr_insn);
6265 curr_static_id = curr_id->insn_static_data;
6266 remove_p = false;
6267 if ((set = single_set (curr_insn)) != NULL_RTX
6268 && REG_P (SET_DEST (set))
6269 && (regno = REGNO (SET_DEST (set))) >= FIRST_PSEUDO_REGISTER
6270 && SET_DEST (set) != pic_offset_table_rtx
6271 && bitmap_bit_p (&check_only_regs, regno)
6272 && ! bitmap_bit_p (&live_regs, regno))
6273 remove_p = true;
6274 /* See which defined values die here. */
6275 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
6276 if (reg->type == OP_OUT && ! reg->subreg_p)
6277 bitmap_clear_bit (&live_regs, reg->regno);
6278 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
6279 if (reg->type == OP_OUT && ! reg->subreg_p)
6280 bitmap_clear_bit (&live_regs, reg->regno);
6281 if (curr_id->arg_hard_regs != NULL)
6282 /* Make clobbered argument hard registers die. */
6283 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6284 if (regno >= FIRST_PSEUDO_REGISTER)
6285 bitmap_clear_bit (&live_regs, regno - FIRST_PSEUDO_REGISTER);
6286 /* Mark each used value as live. */
6287 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
6288 if (reg->type != OP_OUT
6289 && bitmap_bit_p (&check_only_regs, reg->regno))
6290 bitmap_set_bit (&live_regs, reg->regno);
6291 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
6292 if (reg->type != OP_OUT
6293 && bitmap_bit_p (&check_only_regs, reg->regno))
6294 bitmap_set_bit (&live_regs, reg->regno);
6295 if (curr_id->arg_hard_regs != NULL)
6296 /* Make used argument hard registers live. */
6297 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6298 if (regno < FIRST_PSEUDO_REGISTER
6299 && bitmap_bit_p (&check_only_regs, regno))
6300 bitmap_set_bit (&live_regs, regno);
6301 /* It is quite important to remove dead move insns because it
6302 means removing dead store. We don't need to process them for
6303 constraints. */
6304 if (remove_p)
6306 if (lra_dump_file != NULL)
6308 fprintf (lra_dump_file, " Removing dead insn:\n ");
6309 dump_insn_slim (lra_dump_file, curr_insn);
6311 lra_set_insn_deleted (curr_insn);
6316 /* The structure describes info to do an inheritance for the current
6317 insn. We need to collect such info first before doing the
6318 transformations because the transformations change the insn
6319 internal representation. */
6320 struct to_inherit
6322 /* Original regno. */
6323 int regno;
6324 /* Subsequent insns which can inherit original reg value. */
6325 rtx insns;
6328 /* Array containing all info for doing inheritance from the current
6329 insn. */
6330 static struct to_inherit to_inherit[LRA_MAX_INSN_RELOADS];
6332 /* Number elements in the previous array. */
6333 static int to_inherit_num;
6335 /* Add inheritance info REGNO and INSNS. Their meaning is described in
6336 structure to_inherit. */
6337 static void
6338 add_to_inherit (int regno, rtx insns)
6340 int i;
6342 for (i = 0; i < to_inherit_num; i++)
6343 if (to_inherit[i].regno == regno)
6344 return;
6345 lra_assert (to_inherit_num < LRA_MAX_INSN_RELOADS);
6346 to_inherit[to_inherit_num].regno = regno;
6347 to_inherit[to_inherit_num++].insns = insns;
6350 /* Return the last non-debug insn in basic block BB, or the block begin
6351 note if none. */
6352 static rtx_insn *
6353 get_last_insertion_point (basic_block bb)
6355 rtx_insn *insn;
6357 FOR_BB_INSNS_REVERSE (bb, insn)
6358 if (NONDEBUG_INSN_P (insn) || NOTE_INSN_BASIC_BLOCK_P (insn))
6359 return insn;
6360 gcc_unreachable ();
6363 /* Set up RES by registers living on edges FROM except the edge (FROM,
6364 TO) or by registers set up in a jump insn in BB FROM. */
6365 static void
6366 get_live_on_other_edges (basic_block from, basic_block to, bitmap res)
6368 rtx_insn *last;
6369 struct lra_insn_reg *reg;
6370 edge e;
6371 edge_iterator ei;
6373 lra_assert (to != NULL);
6374 bitmap_clear (res);
6375 FOR_EACH_EDGE (e, ei, from->succs)
6376 if (e->dest != to)
6377 bitmap_ior_into (res, df_get_live_in (e->dest));
6378 last = get_last_insertion_point (from);
6379 if (! JUMP_P (last))
6380 return;
6381 curr_id = lra_get_insn_recog_data (last);
6382 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
6383 if (reg->type != OP_IN)
6384 bitmap_set_bit (res, reg->regno);
6387 /* Used as a temporary results of some bitmap calculations. */
6388 static bitmap_head temp_bitmap;
6390 /* We split for reloads of small class of hard regs. The following
6391 defines how many hard regs the class should have to be qualified as
6392 small. The code is mostly oriented to x86/x86-64 architecture
6393 where some insns need to use only specific register or pair of
6394 registers and these register can live in RTL explicitly, e.g. for
6395 parameter passing. */
6396 static const int max_small_class_regs_num = 2;
6398 /* Do inheritance/split transformations in EBB starting with HEAD and
6399 finishing on TAIL. We process EBB insns in the reverse order.
6400 Return true if we did any inheritance/split transformation in the
6401 EBB.
6403 We should avoid excessive splitting which results in worse code
6404 because of inaccurate cost calculations for spilling new split
6405 pseudos in such case. To achieve this we do splitting only if
6406 register pressure is high in given basic block and there are reload
6407 pseudos requiring hard registers. We could do more register
6408 pressure calculations at any given program point to avoid necessary
6409 splitting even more but it is to expensive and the current approach
6410 works well enough. */
6411 static bool
6412 inherit_in_ebb (rtx_insn *head, rtx_insn *tail)
6414 int i, src_regno, dst_regno, nregs;
6415 bool change_p, succ_p, update_reloads_num_p;
6416 rtx_insn *prev_insn, *last_insn;
6417 rtx next_usage_insns, curr_set;
6418 enum reg_class cl;
6419 struct lra_insn_reg *reg;
6420 basic_block last_processed_bb, curr_bb = NULL;
6421 HARD_REG_SET potential_reload_hard_regs, live_hard_regs;
6422 bitmap to_process;
6423 unsigned int j;
6424 bitmap_iterator bi;
6425 bool head_p, after_p;
6427 change_p = false;
6428 curr_usage_insns_check++;
6429 clear_invariants ();
6430 reloads_num = calls_num = 0;
6431 for (unsigned int i = 0; i < NUM_ABI_IDS; ++i)
6432 last_call_for_abi[i] = 0;
6433 CLEAR_HARD_REG_SET (full_and_partial_call_clobbers);
6434 bitmap_clear (&check_only_regs);
6435 bitmap_clear (&invalid_invariant_regs);
6436 last_processed_bb = NULL;
6437 CLEAR_HARD_REG_SET (potential_reload_hard_regs);
6438 live_hard_regs = eliminable_regset | lra_no_alloc_regs;
6439 /* We don't process new insns generated in the loop. */
6440 for (curr_insn = tail; curr_insn != PREV_INSN (head); curr_insn = prev_insn)
6442 prev_insn = PREV_INSN (curr_insn);
6443 if (BLOCK_FOR_INSN (curr_insn) != NULL)
6444 curr_bb = BLOCK_FOR_INSN (curr_insn);
6445 if (last_processed_bb != curr_bb)
6447 /* We are at the end of BB. Add qualified living
6448 pseudos for potential splitting. */
6449 to_process = df_get_live_out (curr_bb);
6450 if (last_processed_bb != NULL)
6452 /* We are somewhere in the middle of EBB. */
6453 get_live_on_other_edges (curr_bb, last_processed_bb,
6454 &temp_bitmap);
6455 to_process = &temp_bitmap;
6457 last_processed_bb = curr_bb;
6458 last_insn = get_last_insertion_point (curr_bb);
6459 after_p = (! JUMP_P (last_insn)
6460 && (! CALL_P (last_insn)
6461 || (find_reg_note (last_insn,
6462 REG_NORETURN, NULL_RTX) == NULL_RTX
6463 && ! SIBLING_CALL_P (last_insn))));
6464 CLEAR_HARD_REG_SET (potential_reload_hard_regs);
6465 EXECUTE_IF_SET_IN_BITMAP (to_process, 0, j, bi)
6467 if ((int) j >= lra_constraint_new_regno_start)
6468 break;
6469 if (j < FIRST_PSEUDO_REGISTER || reg_renumber[j] >= 0)
6471 if (j < FIRST_PSEUDO_REGISTER)
6472 SET_HARD_REG_BIT (live_hard_regs, j);
6473 else
6474 add_to_hard_reg_set (&live_hard_regs,
6475 PSEUDO_REGNO_MODE (j),
6476 reg_renumber[j]);
6477 setup_next_usage_insn (j, last_insn, reloads_num, after_p);
6481 src_regno = dst_regno = -1;
6482 curr_set = single_set (curr_insn);
6483 if (curr_set != NULL_RTX && REG_P (SET_DEST (curr_set)))
6484 dst_regno = REGNO (SET_DEST (curr_set));
6485 if (curr_set != NULL_RTX && REG_P (SET_SRC (curr_set)))
6486 src_regno = REGNO (SET_SRC (curr_set));
6487 update_reloads_num_p = true;
6488 if (src_regno < lra_constraint_new_regno_start
6489 && src_regno >= FIRST_PSEUDO_REGISTER
6490 && reg_renumber[src_regno] < 0
6491 && dst_regno >= lra_constraint_new_regno_start
6492 && (cl = lra_get_allocno_class (dst_regno)) != NO_REGS)
6494 /* 'reload_pseudo <- original_pseudo'. */
6495 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6496 reloads_num++;
6497 update_reloads_num_p = false;
6498 succ_p = false;
6499 if (usage_insns[src_regno].check == curr_usage_insns_check
6500 && (next_usage_insns = usage_insns[src_regno].insns) != NULL_RTX)
6501 succ_p = inherit_reload_reg (false, src_regno, cl,
6502 curr_insn, next_usage_insns);
6503 if (succ_p)
6504 change_p = true;
6505 else
6506 setup_next_usage_insn (src_regno, curr_insn, reloads_num, false);
6507 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6508 potential_reload_hard_regs |= reg_class_contents[cl];
6510 else if (src_regno < 0
6511 && dst_regno >= lra_constraint_new_regno_start
6512 && invariant_p (SET_SRC (curr_set))
6513 && (cl = lra_get_allocno_class (dst_regno)) != NO_REGS
6514 && ! bitmap_bit_p (&invalid_invariant_regs, dst_regno)
6515 && ! bitmap_bit_p (&invalid_invariant_regs,
6516 ORIGINAL_REGNO(regno_reg_rtx[dst_regno])))
6518 /* 'reload_pseudo <- invariant'. */
6519 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6520 reloads_num++;
6521 update_reloads_num_p = false;
6522 if (process_invariant_for_inheritance (SET_DEST (curr_set), SET_SRC (curr_set)))
6523 change_p = true;
6524 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6525 potential_reload_hard_regs |= reg_class_contents[cl];
6527 else if (src_regno >= lra_constraint_new_regno_start
6528 && dst_regno < lra_constraint_new_regno_start
6529 && dst_regno >= FIRST_PSEUDO_REGISTER
6530 && reg_renumber[dst_regno] < 0
6531 && (cl = lra_get_allocno_class (src_regno)) != NO_REGS
6532 && usage_insns[dst_regno].check == curr_usage_insns_check
6533 && (next_usage_insns
6534 = usage_insns[dst_regno].insns) != NULL_RTX)
6536 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6537 reloads_num++;
6538 update_reloads_num_p = false;
6539 /* 'original_pseudo <- reload_pseudo'. */
6540 if (! JUMP_P (curr_insn)
6541 && inherit_reload_reg (true, dst_regno, cl,
6542 curr_insn, next_usage_insns))
6543 change_p = true;
6544 /* Invalidate. */
6545 usage_insns[dst_regno].check = 0;
6546 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6547 potential_reload_hard_regs |= reg_class_contents[cl];
6549 else if (INSN_P (curr_insn))
6551 int iter;
6552 int max_uid = get_max_uid ();
6554 curr_id = lra_get_insn_recog_data (curr_insn);
6555 curr_static_id = curr_id->insn_static_data;
6556 to_inherit_num = 0;
6557 /* Process insn definitions. */
6558 for (iter = 0; iter < 2; iter++)
6559 for (reg = iter == 0 ? curr_id->regs : curr_static_id->hard_regs;
6560 reg != NULL;
6561 reg = reg->next)
6562 if (reg->type != OP_IN
6563 && (dst_regno = reg->regno) < lra_constraint_new_regno_start)
6565 if (dst_regno >= FIRST_PSEUDO_REGISTER && reg->type == OP_OUT
6566 && reg_renumber[dst_regno] < 0 && ! reg->subreg_p
6567 && usage_insns[dst_regno].check == curr_usage_insns_check
6568 && (next_usage_insns
6569 = usage_insns[dst_regno].insns) != NULL_RTX)
6571 struct lra_insn_reg *r;
6573 for (r = curr_id->regs; r != NULL; r = r->next)
6574 if (r->type != OP_OUT && r->regno == dst_regno)
6575 break;
6576 /* Don't do inheritance if the pseudo is also
6577 used in the insn. */
6578 if (r == NULL)
6579 /* We cannot do inheritance right now
6580 because the current insn reg info (chain
6581 regs) can change after that. */
6582 add_to_inherit (dst_regno, next_usage_insns);
6584 /* We cannot process one reg twice here because of
6585 usage_insns invalidation. */
6586 if ((dst_regno < FIRST_PSEUDO_REGISTER
6587 || reg_renumber[dst_regno] >= 0)
6588 && ! reg->subreg_p && reg->type != OP_IN)
6590 HARD_REG_SET s;
6592 if (split_if_necessary (dst_regno, reg->biggest_mode,
6593 potential_reload_hard_regs,
6594 false, curr_insn, max_uid))
6595 change_p = true;
6596 CLEAR_HARD_REG_SET (s);
6597 if (dst_regno < FIRST_PSEUDO_REGISTER)
6598 add_to_hard_reg_set (&s, reg->biggest_mode, dst_regno);
6599 else
6600 add_to_hard_reg_set (&s, PSEUDO_REGNO_MODE (dst_regno),
6601 reg_renumber[dst_regno]);
6602 live_hard_regs &= ~s;
6603 potential_reload_hard_regs &= ~s;
6605 /* We should invalidate potential inheritance or
6606 splitting for the current insn usages to the next
6607 usage insns (see code below) as the output pseudo
6608 prevents this. */
6609 if ((dst_regno >= FIRST_PSEUDO_REGISTER
6610 && reg_renumber[dst_regno] < 0)
6611 || (reg->type == OP_OUT && ! reg->subreg_p
6612 && (dst_regno < FIRST_PSEUDO_REGISTER
6613 || reg_renumber[dst_regno] >= 0)))
6615 /* Invalidate and mark definitions. */
6616 if (dst_regno >= FIRST_PSEUDO_REGISTER)
6617 usage_insns[dst_regno].check = -(int) INSN_UID (curr_insn);
6618 else
6620 nregs = hard_regno_nregs (dst_regno,
6621 reg->biggest_mode);
6622 for (i = 0; i < nregs; i++)
6623 usage_insns[dst_regno + i].check
6624 = -(int) INSN_UID (curr_insn);
6628 /* Process clobbered call regs. */
6629 if (curr_id->arg_hard_regs != NULL)
6630 for (i = 0; (dst_regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6631 if (dst_regno >= FIRST_PSEUDO_REGISTER)
6632 usage_insns[dst_regno - FIRST_PSEUDO_REGISTER].check
6633 = -(int) INSN_UID (curr_insn);
6634 if (! JUMP_P (curr_insn))
6635 for (i = 0; i < to_inherit_num; i++)
6636 if (inherit_reload_reg (true, to_inherit[i].regno,
6637 ALL_REGS, curr_insn,
6638 to_inherit[i].insns))
6639 change_p = true;
6640 if (CALL_P (curr_insn))
6642 rtx cheap, pat, dest;
6643 rtx_insn *restore;
6644 int regno, hard_regno;
6646 calls_num++;
6647 function_abi callee_abi = insn_callee_abi (curr_insn);
6648 last_call_for_abi[callee_abi.id ()] = calls_num;
6649 full_and_partial_call_clobbers
6650 |= callee_abi.full_and_partial_reg_clobbers ();
6651 if ((cheap = find_reg_note (curr_insn,
6652 REG_RETURNED, NULL_RTX)) != NULL_RTX
6653 && ((cheap = XEXP (cheap, 0)), true)
6654 && (regno = REGNO (cheap)) >= FIRST_PSEUDO_REGISTER
6655 && (hard_regno = reg_renumber[regno]) >= 0
6656 && usage_insns[regno].check == curr_usage_insns_check
6657 /* If there are pending saves/restores, the
6658 optimization is not worth. */
6659 && usage_insns[regno].calls_num == calls_num - 1
6660 && callee_abi.clobbers_reg_p (GET_MODE (cheap), hard_regno))
6662 /* Restore the pseudo from the call result as
6663 REG_RETURNED note says that the pseudo value is
6664 in the call result and the pseudo is an argument
6665 of the call. */
6666 pat = PATTERN (curr_insn);
6667 if (GET_CODE (pat) == PARALLEL)
6668 pat = XVECEXP (pat, 0, 0);
6669 dest = SET_DEST (pat);
6670 /* For multiple return values dest is PARALLEL.
6671 Currently we handle only single return value case. */
6672 if (REG_P (dest))
6674 start_sequence ();
6675 emit_move_insn (cheap, copy_rtx (dest));
6676 restore = get_insns ();
6677 end_sequence ();
6678 lra_process_new_insns (curr_insn, NULL, restore,
6679 "Inserting call parameter restore");
6680 /* We don't need to save/restore of the pseudo from
6681 this call. */
6682 usage_insns[regno].calls_num = calls_num;
6683 remove_from_hard_reg_set
6684 (&full_and_partial_call_clobbers,
6685 GET_MODE (cheap), hard_regno);
6686 bitmap_set_bit (&check_only_regs, regno);
6690 to_inherit_num = 0;
6691 /* Process insn usages. */
6692 for (iter = 0; iter < 2; iter++)
6693 for (reg = iter == 0 ? curr_id->regs : curr_static_id->hard_regs;
6694 reg != NULL;
6695 reg = reg->next)
6696 if ((reg->type != OP_OUT
6697 || (reg->type == OP_OUT && reg->subreg_p))
6698 && (src_regno = reg->regno) < lra_constraint_new_regno_start)
6700 if (src_regno >= FIRST_PSEUDO_REGISTER
6701 && reg_renumber[src_regno] < 0 && reg->type == OP_IN)
6703 if (usage_insns[src_regno].check == curr_usage_insns_check
6704 && (next_usage_insns
6705 = usage_insns[src_regno].insns) != NULL_RTX
6706 && NONDEBUG_INSN_P (curr_insn))
6707 add_to_inherit (src_regno, next_usage_insns);
6708 else if (usage_insns[src_regno].check
6709 != -(int) INSN_UID (curr_insn))
6710 /* Add usages but only if the reg is not set up
6711 in the same insn. */
6712 add_next_usage_insn (src_regno, curr_insn, reloads_num);
6714 else if (src_regno < FIRST_PSEUDO_REGISTER
6715 || reg_renumber[src_regno] >= 0)
6717 bool before_p;
6718 rtx_insn *use_insn = curr_insn;
6720 before_p = (JUMP_P (curr_insn)
6721 || (CALL_P (curr_insn) && reg->type == OP_IN));
6722 if (NONDEBUG_INSN_P (curr_insn)
6723 && (! JUMP_P (curr_insn) || reg->type == OP_IN)
6724 && split_if_necessary (src_regno, reg->biggest_mode,
6725 potential_reload_hard_regs,
6726 before_p, curr_insn, max_uid))
6728 if (reg->subreg_p)
6729 check_and_force_assignment_correctness_p = true;
6730 change_p = true;
6731 /* Invalidate. */
6732 usage_insns[src_regno].check = 0;
6733 if (before_p)
6734 use_insn = PREV_INSN (curr_insn);
6736 if (NONDEBUG_INSN_P (curr_insn))
6738 if (src_regno < FIRST_PSEUDO_REGISTER)
6739 add_to_hard_reg_set (&live_hard_regs,
6740 reg->biggest_mode, src_regno);
6741 else
6742 add_to_hard_reg_set (&live_hard_regs,
6743 PSEUDO_REGNO_MODE (src_regno),
6744 reg_renumber[src_regno]);
6746 if (src_regno >= FIRST_PSEUDO_REGISTER)
6747 add_next_usage_insn (src_regno, use_insn, reloads_num);
6748 else
6750 for (i = 0; i < hard_regno_nregs (src_regno, reg->biggest_mode); i++)
6751 add_next_usage_insn (src_regno + i, use_insn, reloads_num);
6755 /* Process used call regs. */
6756 if (curr_id->arg_hard_regs != NULL)
6757 for (i = 0; (src_regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6758 if (src_regno < FIRST_PSEUDO_REGISTER)
6760 SET_HARD_REG_BIT (live_hard_regs, src_regno);
6761 add_next_usage_insn (src_regno, curr_insn, reloads_num);
6763 for (i = 0; i < to_inherit_num; i++)
6765 src_regno = to_inherit[i].regno;
6766 if (inherit_reload_reg (false, src_regno, ALL_REGS,
6767 curr_insn, to_inherit[i].insns))
6768 change_p = true;
6769 else
6770 setup_next_usage_insn (src_regno, curr_insn, reloads_num, false);
6773 if (update_reloads_num_p
6774 && NONDEBUG_INSN_P (curr_insn) && curr_set != NULL_RTX)
6776 int regno = -1;
6777 if ((REG_P (SET_DEST (curr_set))
6778 && (regno = REGNO (SET_DEST (curr_set))) >= lra_constraint_new_regno_start
6779 && reg_renumber[regno] < 0
6780 && (cl = lra_get_allocno_class (regno)) != NO_REGS)
6781 || (REG_P (SET_SRC (curr_set))
6782 && (regno = REGNO (SET_SRC (curr_set))) >= lra_constraint_new_regno_start
6783 && reg_renumber[regno] < 0
6784 && (cl = lra_get_allocno_class (regno)) != NO_REGS))
6786 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6787 reloads_num++;
6788 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6789 potential_reload_hard_regs |= reg_class_contents[cl];
6792 if (NONDEBUG_INSN_P (curr_insn))
6794 int regno;
6796 /* Invalidate invariants with changed regs. */
6797 curr_id = lra_get_insn_recog_data (curr_insn);
6798 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
6799 if (reg->type != OP_IN)
6801 bitmap_set_bit (&invalid_invariant_regs, reg->regno);
6802 bitmap_set_bit (&invalid_invariant_regs,
6803 ORIGINAL_REGNO (regno_reg_rtx[reg->regno]));
6805 curr_static_id = curr_id->insn_static_data;
6806 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
6807 if (reg->type != OP_IN)
6808 bitmap_set_bit (&invalid_invariant_regs, reg->regno);
6809 if (curr_id->arg_hard_regs != NULL)
6810 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6811 if (regno >= FIRST_PSEUDO_REGISTER)
6812 bitmap_set_bit (&invalid_invariant_regs,
6813 regno - FIRST_PSEUDO_REGISTER);
6815 /* We reached the start of the current basic block. */
6816 if (prev_insn == NULL_RTX || prev_insn == PREV_INSN (head)
6817 || BLOCK_FOR_INSN (prev_insn) != curr_bb)
6819 /* We reached the beginning of the current block -- do
6820 rest of spliting in the current BB. */
6821 to_process = df_get_live_in (curr_bb);
6822 if (BLOCK_FOR_INSN (head) != curr_bb)
6824 /* We are somewhere in the middle of EBB. */
6825 get_live_on_other_edges (EDGE_PRED (curr_bb, 0)->src,
6826 curr_bb, &temp_bitmap);
6827 to_process = &temp_bitmap;
6829 head_p = true;
6830 EXECUTE_IF_SET_IN_BITMAP (to_process, 0, j, bi)
6832 if ((int) j >= lra_constraint_new_regno_start)
6833 break;
6834 if (((int) j < FIRST_PSEUDO_REGISTER || reg_renumber[j] >= 0)
6835 && usage_insns[j].check == curr_usage_insns_check
6836 && (next_usage_insns = usage_insns[j].insns) != NULL_RTX)
6838 if (need_for_split_p (potential_reload_hard_regs, j))
6840 if (lra_dump_file != NULL && head_p)
6842 fprintf (lra_dump_file,
6843 " ----------------------------------\n");
6844 head_p = false;
6846 if (split_reg (false, j, bb_note (curr_bb),
6847 next_usage_insns, NULL))
6848 change_p = true;
6850 usage_insns[j].check = 0;
6855 return change_p;
6858 /* This value affects EBB forming. If probability of edge from EBB to
6859 a BB is not greater than the following value, we don't add the BB
6860 to EBB. */
6861 #define EBB_PROBABILITY_CUTOFF \
6862 ((REG_BR_PROB_BASE * param_lra_inheritance_ebb_probability_cutoff) / 100)
6864 /* Current number of inheritance/split iteration. */
6865 int lra_inheritance_iter;
6867 /* Entry function for inheritance/split pass. */
6868 void
6869 lra_inheritance (void)
6871 int i;
6872 basic_block bb, start_bb;
6873 edge e;
6875 lra_inheritance_iter++;
6876 if (lra_inheritance_iter > LRA_MAX_INHERITANCE_PASSES)
6877 return;
6878 timevar_push (TV_LRA_INHERITANCE);
6879 if (lra_dump_file != NULL)
6880 fprintf (lra_dump_file, "\n********** Inheritance #%d: **********\n\n",
6881 lra_inheritance_iter);
6882 curr_usage_insns_check = 0;
6883 usage_insns = XNEWVEC (struct usage_insns, lra_constraint_new_regno_start);
6884 for (i = 0; i < lra_constraint_new_regno_start; i++)
6885 usage_insns[i].check = 0;
6886 bitmap_initialize (&check_only_regs, &reg_obstack);
6887 bitmap_initialize (&invalid_invariant_regs, &reg_obstack);
6888 bitmap_initialize (&live_regs, &reg_obstack);
6889 bitmap_initialize (&temp_bitmap, &reg_obstack);
6890 bitmap_initialize (&ebb_global_regs, &reg_obstack);
6891 FOR_EACH_BB_FN (bb, cfun)
6893 start_bb = bb;
6894 if (lra_dump_file != NULL)
6895 fprintf (lra_dump_file, "EBB");
6896 /* Form a EBB starting with BB. */
6897 bitmap_clear (&ebb_global_regs);
6898 bitmap_ior_into (&ebb_global_regs, df_get_live_in (bb));
6899 for (;;)
6901 if (lra_dump_file != NULL)
6902 fprintf (lra_dump_file, " %d", bb->index);
6903 if (bb->next_bb == EXIT_BLOCK_PTR_FOR_FN (cfun)
6904 || LABEL_P (BB_HEAD (bb->next_bb)))
6905 break;
6906 e = find_fallthru_edge (bb->succs);
6907 if (! e)
6908 break;
6909 if (e->probability.initialized_p ()
6910 && e->probability.to_reg_br_prob_base () < EBB_PROBABILITY_CUTOFF)
6911 break;
6912 bb = bb->next_bb;
6914 bitmap_ior_into (&ebb_global_regs, df_get_live_out (bb));
6915 if (lra_dump_file != NULL)
6916 fprintf (lra_dump_file, "\n");
6917 if (inherit_in_ebb (BB_HEAD (start_bb), BB_END (bb)))
6918 /* Remember that the EBB head and tail can change in
6919 inherit_in_ebb. */
6920 update_ebb_live_info (BB_HEAD (start_bb), BB_END (bb));
6922 bitmap_release (&ebb_global_regs);
6923 bitmap_release (&temp_bitmap);
6924 bitmap_release (&live_regs);
6925 bitmap_release (&invalid_invariant_regs);
6926 bitmap_release (&check_only_regs);
6927 free (usage_insns);
6929 timevar_pop (TV_LRA_INHERITANCE);
6934 /* This page contains code to undo failed inheritance/split
6935 transformations. */
6937 /* Current number of iteration undoing inheritance/split. */
6938 int lra_undo_inheritance_iter;
6940 /* Fix BB live info LIVE after removing pseudos created on pass doing
6941 inheritance/split which are REMOVED_PSEUDOS. */
6942 static void
6943 fix_bb_live_info (bitmap live, bitmap removed_pseudos)
6945 unsigned int regno;
6946 bitmap_iterator bi;
6948 EXECUTE_IF_SET_IN_BITMAP (removed_pseudos, 0, regno, bi)
6949 if (bitmap_clear_bit (live, regno)
6950 && REG_P (lra_reg_info[regno].restore_rtx))
6951 bitmap_set_bit (live, REGNO (lra_reg_info[regno].restore_rtx));
6954 /* Return regno of the (subreg of) REG. Otherwise, return a negative
6955 number. */
6956 static int
6957 get_regno (rtx reg)
6959 if (GET_CODE (reg) == SUBREG)
6960 reg = SUBREG_REG (reg);
6961 if (REG_P (reg))
6962 return REGNO (reg);
6963 return -1;
6966 /* Delete a move INSN with destination reg DREGNO and a previous
6967 clobber insn with the same regno. The inheritance/split code can
6968 generate moves with preceding clobber and when we delete such moves
6969 we should delete the clobber insn too to keep the correct life
6970 info. */
6971 static void
6972 delete_move_and_clobber (rtx_insn *insn, int dregno)
6974 rtx_insn *prev_insn = PREV_INSN (insn);
6976 lra_set_insn_deleted (insn);
6977 lra_assert (dregno >= 0);
6978 if (prev_insn != NULL && NONDEBUG_INSN_P (prev_insn)
6979 && GET_CODE (PATTERN (prev_insn)) == CLOBBER
6980 && dregno == get_regno (XEXP (PATTERN (prev_insn), 0)))
6981 lra_set_insn_deleted (prev_insn);
6984 /* Remove inheritance/split pseudos which are in REMOVE_PSEUDOS and
6985 return true if we did any change. The undo transformations for
6986 inheritance looks like
6987 i <- i2
6988 p <- i => p <- i2
6989 or removing
6990 p <- i, i <- p, and i <- i3
6991 where p is original pseudo from which inheritance pseudo i was
6992 created, i and i3 are removed inheritance pseudos, i2 is another
6993 not removed inheritance pseudo. All split pseudos or other
6994 occurrences of removed inheritance pseudos are changed on the
6995 corresponding original pseudos.
6997 The function also schedules insns changed and created during
6998 inheritance/split pass for processing by the subsequent constraint
6999 pass. */
7000 static bool
7001 remove_inheritance_pseudos (bitmap remove_pseudos)
7003 basic_block bb;
7004 int regno, sregno, prev_sregno, dregno;
7005 rtx restore_rtx;
7006 rtx set, prev_set;
7007 rtx_insn *prev_insn;
7008 bool change_p, done_p;
7010 change_p = ! bitmap_empty_p (remove_pseudos);
7011 /* We cannot finish the function right away if CHANGE_P is true
7012 because we need to marks insns affected by previous
7013 inheritance/split pass for processing by the subsequent
7014 constraint pass. */
7015 FOR_EACH_BB_FN (bb, cfun)
7017 fix_bb_live_info (df_get_live_in (bb), remove_pseudos);
7018 fix_bb_live_info (df_get_live_out (bb), remove_pseudos);
7019 FOR_BB_INSNS_REVERSE (bb, curr_insn)
7021 if (! INSN_P (curr_insn))
7022 continue;
7023 done_p = false;
7024 sregno = dregno = -1;
7025 if (change_p && NONDEBUG_INSN_P (curr_insn)
7026 && (set = single_set (curr_insn)) != NULL_RTX)
7028 dregno = get_regno (SET_DEST (set));
7029 sregno = get_regno (SET_SRC (set));
7032 if (sregno >= 0 && dregno >= 0)
7034 if (bitmap_bit_p (remove_pseudos, dregno)
7035 && ! REG_P (lra_reg_info[dregno].restore_rtx))
7037 /* invariant inheritance pseudo <- original pseudo */
7038 if (lra_dump_file != NULL)
7040 fprintf (lra_dump_file, " Removing invariant inheritance:\n");
7041 dump_insn_slim (lra_dump_file, curr_insn);
7042 fprintf (lra_dump_file, "\n");
7044 delete_move_and_clobber (curr_insn, dregno);
7045 done_p = true;
7047 else if (bitmap_bit_p (remove_pseudos, sregno)
7048 && ! REG_P (lra_reg_info[sregno].restore_rtx))
7050 /* reload pseudo <- invariant inheritance pseudo */
7051 start_sequence ();
7052 /* We cannot just change the source. It might be
7053 an insn different from the move. */
7054 emit_insn (lra_reg_info[sregno].restore_rtx);
7055 rtx_insn *new_insns = get_insns ();
7056 end_sequence ();
7057 lra_assert (single_set (new_insns) != NULL
7058 && SET_DEST (set) == SET_DEST (single_set (new_insns)));
7059 lra_process_new_insns (curr_insn, NULL, new_insns,
7060 "Changing reload<-invariant inheritance");
7061 delete_move_and_clobber (curr_insn, dregno);
7062 done_p = true;
7064 else if ((bitmap_bit_p (remove_pseudos, sregno)
7065 && (get_regno (lra_reg_info[sregno].restore_rtx) == dregno
7066 || (bitmap_bit_p (remove_pseudos, dregno)
7067 && get_regno (lra_reg_info[sregno].restore_rtx) >= 0
7068 && (get_regno (lra_reg_info[sregno].restore_rtx)
7069 == get_regno (lra_reg_info[dregno].restore_rtx)))))
7070 || (bitmap_bit_p (remove_pseudos, dregno)
7071 && get_regno (lra_reg_info[dregno].restore_rtx) == sregno))
7072 /* One of the following cases:
7073 original <- removed inheritance pseudo
7074 removed inherit pseudo <- another removed inherit pseudo
7075 removed inherit pseudo <- original pseudo
7077 removed_split_pseudo <- original_reg
7078 original_reg <- removed_split_pseudo */
7080 if (lra_dump_file != NULL)
7082 fprintf (lra_dump_file, " Removing %s:\n",
7083 bitmap_bit_p (&lra_split_regs, sregno)
7084 || bitmap_bit_p (&lra_split_regs, dregno)
7085 ? "split" : "inheritance");
7086 dump_insn_slim (lra_dump_file, curr_insn);
7088 delete_move_and_clobber (curr_insn, dregno);
7089 done_p = true;
7091 else if (bitmap_bit_p (remove_pseudos, sregno)
7092 && bitmap_bit_p (&lra_inheritance_pseudos, sregno))
7094 /* Search the following pattern:
7095 inherit_or_split_pseudo1 <- inherit_or_split_pseudo2
7096 original_pseudo <- inherit_or_split_pseudo1
7097 where the 2nd insn is the current insn and
7098 inherit_or_split_pseudo2 is not removed. If it is found,
7099 change the current insn onto:
7100 original_pseudo <- inherit_or_split_pseudo2. */
7101 for (prev_insn = PREV_INSN (curr_insn);
7102 prev_insn != NULL_RTX && ! NONDEBUG_INSN_P (prev_insn);
7103 prev_insn = PREV_INSN (prev_insn))
7105 if (prev_insn != NULL_RTX && BLOCK_FOR_INSN (prev_insn) == bb
7106 && (prev_set = single_set (prev_insn)) != NULL_RTX
7107 /* There should be no subregs in insn we are
7108 searching because only the original reg might
7109 be in subreg when we changed the mode of
7110 load/store for splitting. */
7111 && REG_P (SET_DEST (prev_set))
7112 && REG_P (SET_SRC (prev_set))
7113 && (int) REGNO (SET_DEST (prev_set)) == sregno
7114 && ((prev_sregno = REGNO (SET_SRC (prev_set)))
7115 >= FIRST_PSEUDO_REGISTER)
7116 && (lra_reg_info[prev_sregno].restore_rtx == NULL_RTX
7118 /* As we consider chain of inheritance or
7119 splitting described in above comment we should
7120 check that sregno and prev_sregno were
7121 inheritance/split pseudos created from the
7122 same original regno. */
7123 (get_regno (lra_reg_info[sregno].restore_rtx) >= 0
7124 && (get_regno (lra_reg_info[sregno].restore_rtx)
7125 == get_regno (lra_reg_info[prev_sregno].restore_rtx))))
7126 && ! bitmap_bit_p (remove_pseudos, prev_sregno))
7128 lra_assert (GET_MODE (SET_SRC (prev_set))
7129 == GET_MODE (regno_reg_rtx[sregno]));
7130 /* Although we have a single set, the insn can
7131 contain more one sregno register occurrence
7132 as a source. Change all occurrences. */
7133 lra_substitute_pseudo_within_insn (curr_insn, sregno,
7134 SET_SRC (prev_set),
7135 false);
7136 /* As we are finishing with processing the insn
7137 here, check the destination too as it might
7138 inheritance pseudo for another pseudo. */
7139 if (bitmap_bit_p (remove_pseudos, dregno)
7140 && bitmap_bit_p (&lra_inheritance_pseudos, dregno)
7141 && (restore_rtx
7142 = lra_reg_info[dregno].restore_rtx) != NULL_RTX)
7144 if (GET_CODE (SET_DEST (set)) == SUBREG)
7145 SUBREG_REG (SET_DEST (set)) = restore_rtx;
7146 else
7147 SET_DEST (set) = restore_rtx;
7149 lra_push_insn_and_update_insn_regno_info (curr_insn);
7150 lra_set_used_insn_alternative_by_uid
7151 (INSN_UID (curr_insn), LRA_UNKNOWN_ALT);
7152 done_p = true;
7153 if (lra_dump_file != NULL)
7155 fprintf (lra_dump_file, " Change reload insn:\n");
7156 dump_insn_slim (lra_dump_file, curr_insn);
7161 if (! done_p)
7163 struct lra_insn_reg *reg;
7164 bool restored_regs_p = false;
7165 bool kept_regs_p = false;
7167 curr_id = lra_get_insn_recog_data (curr_insn);
7168 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
7170 regno = reg->regno;
7171 restore_rtx = lra_reg_info[regno].restore_rtx;
7172 if (restore_rtx != NULL_RTX)
7174 if (change_p && bitmap_bit_p (remove_pseudos, regno))
7176 lra_substitute_pseudo_within_insn
7177 (curr_insn, regno, restore_rtx, false);
7178 restored_regs_p = true;
7180 else
7181 kept_regs_p = true;
7184 if (NONDEBUG_INSN_P (curr_insn) && kept_regs_p)
7186 /* The instruction has changed since the previous
7187 constraints pass. */
7188 lra_push_insn_and_update_insn_regno_info (curr_insn);
7189 lra_set_used_insn_alternative_by_uid
7190 (INSN_UID (curr_insn), LRA_UNKNOWN_ALT);
7192 else if (restored_regs_p)
7193 /* The instruction has been restored to the form that
7194 it had during the previous constraints pass. */
7195 lra_update_insn_regno_info (curr_insn);
7196 if (restored_regs_p && lra_dump_file != NULL)
7198 fprintf (lra_dump_file, " Insn after restoring regs:\n");
7199 dump_insn_slim (lra_dump_file, curr_insn);
7204 return change_p;
7207 /* If optional reload pseudos failed to get a hard register or was not
7208 inherited, it is better to remove optional reloads. We do this
7209 transformation after undoing inheritance to figure out necessity to
7210 remove optional reloads easier. Return true if we do any
7211 change. */
7212 static bool
7213 undo_optional_reloads (void)
7215 bool change_p, keep_p;
7216 unsigned int regno, uid;
7217 bitmap_iterator bi, bi2;
7218 rtx_insn *insn;
7219 rtx set, src, dest;
7220 auto_bitmap removed_optional_reload_pseudos (&reg_obstack);
7222 bitmap_copy (removed_optional_reload_pseudos, &lra_optional_reload_pseudos);
7223 EXECUTE_IF_SET_IN_BITMAP (&lra_optional_reload_pseudos, 0, regno, bi)
7225 keep_p = false;
7226 /* Keep optional reloads from previous subpasses. */
7227 if (lra_reg_info[regno].restore_rtx == NULL_RTX
7228 /* If the original pseudo changed its allocation, just
7229 removing the optional pseudo is dangerous as the original
7230 pseudo will have longer live range. */
7231 || reg_renumber[REGNO (lra_reg_info[regno].restore_rtx)] >= 0)
7232 keep_p = true;
7233 else if (reg_renumber[regno] >= 0)
7234 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi2)
7236 insn = lra_insn_recog_data[uid]->insn;
7237 if ((set = single_set (insn)) == NULL_RTX)
7238 continue;
7239 src = SET_SRC (set);
7240 dest = SET_DEST (set);
7241 if (! REG_P (src) || ! REG_P (dest))
7242 continue;
7243 if (REGNO (dest) == regno
7244 /* Ignore insn for optional reloads itself. */
7245 && REGNO (lra_reg_info[regno].restore_rtx) != REGNO (src)
7246 /* Check only inheritance on last inheritance pass. */
7247 && (int) REGNO (src) >= new_regno_start
7248 /* Check that the optional reload was inherited. */
7249 && bitmap_bit_p (&lra_inheritance_pseudos, REGNO (src)))
7251 keep_p = true;
7252 break;
7255 if (keep_p)
7257 bitmap_clear_bit (removed_optional_reload_pseudos, regno);
7258 if (lra_dump_file != NULL)
7259 fprintf (lra_dump_file, "Keep optional reload reg %d\n", regno);
7262 change_p = ! bitmap_empty_p (removed_optional_reload_pseudos);
7263 auto_bitmap insn_bitmap (&reg_obstack);
7264 EXECUTE_IF_SET_IN_BITMAP (removed_optional_reload_pseudos, 0, regno, bi)
7266 if (lra_dump_file != NULL)
7267 fprintf (lra_dump_file, "Remove optional reload reg %d\n", regno);
7268 bitmap_copy (insn_bitmap, &lra_reg_info[regno].insn_bitmap);
7269 EXECUTE_IF_SET_IN_BITMAP (insn_bitmap, 0, uid, bi2)
7271 insn = lra_insn_recog_data[uid]->insn;
7272 if ((set = single_set (insn)) != NULL_RTX)
7274 src = SET_SRC (set);
7275 dest = SET_DEST (set);
7276 if (REG_P (src) && REG_P (dest)
7277 && ((REGNO (src) == regno
7278 && (REGNO (lra_reg_info[regno].restore_rtx)
7279 == REGNO (dest)))
7280 || (REGNO (dest) == regno
7281 && (REGNO (lra_reg_info[regno].restore_rtx)
7282 == REGNO (src)))))
7284 if (lra_dump_file != NULL)
7286 fprintf (lra_dump_file, " Deleting move %u\n",
7287 INSN_UID (insn));
7288 dump_insn_slim (lra_dump_file, insn);
7290 delete_move_and_clobber (insn, REGNO (dest));
7291 continue;
7293 /* We should not worry about generation memory-memory
7294 moves here as if the corresponding inheritance did
7295 not work (inheritance pseudo did not get a hard reg),
7296 we remove the inheritance pseudo and the optional
7297 reload. */
7299 lra_substitute_pseudo_within_insn
7300 (insn, regno, lra_reg_info[regno].restore_rtx, false);
7301 lra_update_insn_regno_info (insn);
7302 if (lra_dump_file != NULL)
7304 fprintf (lra_dump_file,
7305 " Restoring original insn:\n");
7306 dump_insn_slim (lra_dump_file, insn);
7310 /* Clear restore_regnos. */
7311 EXECUTE_IF_SET_IN_BITMAP (&lra_optional_reload_pseudos, 0, regno, bi)
7312 lra_reg_info[regno].restore_rtx = NULL_RTX;
7313 return change_p;
7316 /* Entry function for undoing inheritance/split transformation. Return true
7317 if we did any RTL change in this pass. */
7318 bool
7319 lra_undo_inheritance (void)
7321 unsigned int regno;
7322 int hard_regno;
7323 int n_all_inherit, n_inherit, n_all_split, n_split;
7324 rtx restore_rtx;
7325 bitmap_iterator bi;
7326 bool change_p;
7328 lra_undo_inheritance_iter++;
7329 if (lra_undo_inheritance_iter > LRA_MAX_INHERITANCE_PASSES)
7330 return false;
7331 if (lra_dump_file != NULL)
7332 fprintf (lra_dump_file,
7333 "\n********** Undoing inheritance #%d: **********\n\n",
7334 lra_undo_inheritance_iter);
7335 auto_bitmap remove_pseudos (&reg_obstack);
7336 n_inherit = n_all_inherit = 0;
7337 EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, regno, bi)
7338 if (lra_reg_info[regno].restore_rtx != NULL_RTX)
7340 n_all_inherit++;
7341 if (reg_renumber[regno] < 0
7342 /* If the original pseudo changed its allocation, just
7343 removing inheritance is dangerous as for changing
7344 allocation we used shorter live-ranges. */
7345 && (! REG_P (lra_reg_info[regno].restore_rtx)
7346 || reg_renumber[REGNO (lra_reg_info[regno].restore_rtx)] < 0))
7347 bitmap_set_bit (remove_pseudos, regno);
7348 else
7349 n_inherit++;
7351 if (lra_dump_file != NULL && n_all_inherit != 0)
7352 fprintf (lra_dump_file, "Inherit %d out of %d (%.2f%%)\n",
7353 n_inherit, n_all_inherit,
7354 (double) n_inherit / n_all_inherit * 100);
7355 n_split = n_all_split = 0;
7356 EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, regno, bi)
7357 if ((restore_rtx = lra_reg_info[regno].restore_rtx) != NULL_RTX)
7359 int restore_regno = REGNO (restore_rtx);
7361 n_all_split++;
7362 hard_regno = (restore_regno >= FIRST_PSEUDO_REGISTER
7363 ? reg_renumber[restore_regno] : restore_regno);
7364 if (hard_regno < 0 || reg_renumber[regno] == hard_regno)
7365 bitmap_set_bit (remove_pseudos, regno);
7366 else
7368 n_split++;
7369 if (lra_dump_file != NULL)
7370 fprintf (lra_dump_file, " Keep split r%d (orig=r%d)\n",
7371 regno, restore_regno);
7374 if (lra_dump_file != NULL && n_all_split != 0)
7375 fprintf (lra_dump_file, "Split %d out of %d (%.2f%%)\n",
7376 n_split, n_all_split,
7377 (double) n_split / n_all_split * 100);
7378 change_p = remove_inheritance_pseudos (remove_pseudos);
7379 /* Clear restore_regnos. */
7380 EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, regno, bi)
7381 lra_reg_info[regno].restore_rtx = NULL_RTX;
7382 EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, regno, bi)
7383 lra_reg_info[regno].restore_rtx = NULL_RTX;
7384 change_p = undo_optional_reloads () || change_p;
7385 return change_p;