1 /* Perform various loop optimizations, including strength reduction.
2 Copyright (C) 1987, 1988, 1989, 1991, 1992, 1993, 1994, 1995,
3 1996, 1997, 1998, 1999, 2000, 2001, 2002, 2003, 2004, 2005
4 Free Software Foundation, Inc.
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify it under
9 the terms of the GNU General Public License as published by the Free
10 Software Foundation; either version 2, or (at your option) any later
13 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
14 WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING. If not, write to the Free
20 Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA
23 /* This is the loop optimization pass of the compiler.
24 It finds invariant computations within loops and moves them
25 to the beginning of the loop. Then it identifies basic and
26 general induction variables.
28 Basic induction variables (BIVs) are a pseudo registers which are set within
29 a loop only by incrementing or decrementing its value. General induction
30 variables (GIVs) are pseudo registers with a value which is a linear function
31 of a basic induction variable. BIVs are recognized by `basic_induction_var';
32 GIVs by `general_induction_var'.
34 Once induction variables are identified, strength reduction is applied to the
35 general induction variables, and induction variable elimination is applied to
36 the basic induction variables.
38 It also finds cases where
39 a register is set within the loop by zero-extending a narrower value
40 and changes these to zero the entire register once before the loop
41 and merely copy the low part within the loop.
43 Most of the complexity is in heuristics to decide when it is worth
44 while to do these things. */
48 #include "coretypes.h"
54 #include "hard-reg-set.h"
55 #include "basic-block.h"
56 #include "insn-config.h"
65 #include "insn-flags.h"
70 #include "tree-pass.h"
72 /* Get the loop info pointer of a loop. */
73 #define LOOP_INFO(LOOP) ((struct loop_info *) (LOOP)->aux)
75 /* Get a pointer to the loop movables structure. */
76 #define LOOP_MOVABLES(LOOP) (&LOOP_INFO (LOOP)->movables)
78 /* Get a pointer to the loop registers structure. */
79 #define LOOP_REGS(LOOP) (&LOOP_INFO (LOOP)->regs)
81 /* Get a pointer to the loop induction variables structure. */
82 #define LOOP_IVS(LOOP) (&LOOP_INFO (LOOP)->ivs)
84 /* Get the luid of an insn. Catch the error of trying to reference the LUID
85 of an insn added during loop, since these don't have LUIDs. */
87 #define INSN_LUID(INSN) \
88 (gcc_assert (INSN_UID (INSN) < max_uid_for_loop), uid_luid[INSN_UID (INSN)])
90 #define REGNO_FIRST_LUID(REGNO) \
91 (REGNO_FIRST_UID (REGNO) < max_uid_for_loop \
92 ? uid_luid[REGNO_FIRST_UID (REGNO)] \
94 #define REGNO_LAST_LUID(REGNO) \
95 (REGNO_LAST_UID (REGNO) < max_uid_for_loop \
96 ? uid_luid[REGNO_LAST_UID (REGNO)] \
99 /* A "basic induction variable" or biv is a pseudo reg that is set
100 (within this loop) only by incrementing or decrementing it. */
101 /* A "general induction variable" or giv is a pseudo reg whose
102 value is a linear function of a biv. */
104 /* Bivs are recognized by `basic_induction_var';
105 Givs by `general_induction_var'. */
107 /* An enum for the two different types of givs, those that are used
108 as memory addresses and those that are calculated into registers. */
116 /* A `struct induction' is created for every instruction that sets
117 an induction variable (either a biv or a giv). */
121 rtx insn; /* The insn that sets a biv or giv */
122 rtx new_reg; /* New register, containing strength reduced
123 version of this giv. */
124 rtx src_reg; /* Biv from which this giv is computed.
125 (If this is a biv, then this is the biv.) */
126 enum g_types giv_type; /* Indicate whether DEST_ADDR or DEST_REG */
127 rtx dest_reg; /* Destination register for insn: this is the
128 register which was the biv or giv.
129 For a biv, this equals src_reg.
130 For a DEST_ADDR type giv, this is 0. */
131 rtx *location; /* Place in the insn where this giv occurs.
132 If GIV_TYPE is DEST_REG, this is 0. */
133 /* For a biv, this is the place where add_val
135 enum machine_mode mode; /* The mode of this biv or giv */
136 rtx mem; /* For DEST_ADDR, the memory object. */
137 rtx mult_val; /* Multiplicative factor for src_reg. */
138 rtx add_val; /* Additive constant for that product. */
139 int benefit; /* Gain from eliminating this insn. */
140 rtx final_value; /* If the giv is used outside the loop, and its
141 final value could be calculated, it is put
142 here, and the giv is made replaceable. Set
143 the giv to this value before the loop. */
144 unsigned combined_with; /* The number of givs this giv has been
145 combined with. If nonzero, this giv
146 cannot combine with any other giv. */
147 unsigned replaceable : 1; /* 1 if we can substitute the strength-reduced
148 variable for the original variable.
149 0 means they must be kept separate and the
150 new one must be copied into the old pseudo
151 reg each time the old one is set. */
152 unsigned not_replaceable : 1; /* Used to prevent duplicating work. This is
153 1 if we know that the giv definitely can
154 not be made replaceable, in which case we
155 don't bother checking the variable again
156 even if further info is available.
157 Both this and the above can be zero. */
158 unsigned ignore : 1; /* 1 prohibits further processing of giv */
159 unsigned always_computable : 1;/* 1 if this value is computable every
161 unsigned always_executed : 1; /* 1 if this set occurs each iteration. */
162 unsigned maybe_multiple : 1; /* Only used for a biv and 1 if this biv
163 update may be done multiple times per
165 unsigned cant_derive : 1; /* For giv's, 1 if this giv cannot derive
166 another giv. This occurs in many cases
167 where a giv's lifetime spans an update to
169 unsigned maybe_dead : 1; /* 1 if this giv might be dead. In that case,
170 we won't use it to eliminate a biv, it
171 would probably lose. */
172 unsigned auto_inc_opt : 1; /* 1 if this giv had its increment output next
173 to it to try to form an auto-inc address. */
175 unsigned no_const_addval : 1; /* 1 if add_val does not contain a const. */
176 int lifetime; /* Length of life of this giv */
177 rtx derive_adjustment; /* If nonzero, is an adjustment to be
178 subtracted from add_val when this giv
179 derives another. This occurs when the
180 giv spans a biv update by incrementation. */
181 rtx ext_dependent; /* If nonzero, is a sign or zero extension
182 if a biv on which this giv is dependent. */
183 struct induction *next_iv; /* For givs, links together all givs that are
184 based on the same biv. For bivs, links
185 together all biv entries that refer to the
186 same biv register. */
187 struct induction *same; /* For givs, if the giv has been combined with
188 another giv, this points to the base giv.
189 The base giv will have COMBINED_WITH nonzero.
190 For bivs, if the biv has the same LOCATION
191 than another biv, this points to the base
193 struct induction *same_insn; /* If there are multiple identical givs in
194 the same insn, then all but one have this
195 field set, and they all point to the giv
196 that doesn't have this field set. */
197 rtx last_use; /* For a giv made from a biv increment, this is
198 a substitute for the lifetime information. */
202 /* A `struct iv_class' is created for each biv. */
206 unsigned int regno; /* Pseudo reg which is the biv. */
207 int biv_count; /* Number of insns setting this reg. */
208 struct induction *biv; /* List of all insns that set this reg. */
209 int giv_count; /* Number of DEST_REG givs computed from this
210 biv. The resulting count is only used in
212 struct induction *giv; /* List of all insns that compute a giv
214 int total_benefit; /* Sum of BENEFITs of all those givs. */
215 rtx initial_value; /* Value of reg at loop start. */
216 rtx initial_test; /* Test performed on BIV before loop. */
217 rtx final_value; /* Value of reg at loop end, if known. */
218 struct iv_class *next; /* Links all class structures together. */
219 rtx init_insn; /* insn which initializes biv, 0 if none. */
220 rtx init_set; /* SET of INIT_INSN, if any. */
221 unsigned incremented : 1; /* 1 if somewhere incremented/decremented */
222 unsigned eliminable : 1; /* 1 if plausible candidate for
224 unsigned nonneg : 1; /* 1 if we added a REG_NONNEG note for
226 unsigned reversed : 1; /* 1 if we reversed the loop that this
228 unsigned all_reduced : 1; /* 1 if all givs using this biv have
233 /* Definitions used by the basic induction variable discovery code. */
243 /* A `struct iv' is created for every register. */
250 struct iv_class *class;
251 struct induction *info;
256 #define REG_IV_TYPE(ivs, n) ivs->regs[n].type
257 #define REG_IV_INFO(ivs, n) ivs->regs[n].iv.info
258 #define REG_IV_CLASS(ivs, n) ivs->regs[n].iv.class
263 /* Indexed by register number, contains pointer to `struct
264 iv' if register is an induction variable. */
267 /* Size of regs array. */
270 /* The head of a list which links together (via the next field)
271 every iv class for the current loop. */
272 struct iv_class *list;
276 typedef struct loop_mem_info
278 rtx mem; /* The MEM itself. */
279 rtx reg; /* Corresponding pseudo, if any. */
280 int optimize; /* Nonzero if we can optimize access to this MEM. */
287 /* Number of times the reg is set during the loop being scanned.
288 During code motion, a negative value indicates a reg that has
289 been made a candidate; in particular -2 means that it is an
290 candidate that we know is equal to a constant and -1 means that
291 it is a candidate not known equal to a constant. After code
292 motion, regs moved have 0 (which is accurate now) while the
293 failed candidates have the original number of times set.
295 Therefore, at all times, == 0 indicates an invariant register;
296 < 0 a conditionally invariant one. */
299 /* Original value of set_in_loop; same except that this value
300 is not set negative for a reg whose sets have been made candidates
301 and not set to 0 for a reg that is moved. */
304 /* Contains the insn in which a register was used if it was used
305 exactly once; contains const0_rtx if it was used more than once. */
308 /* Nonzero indicates that the register cannot be moved or strength
310 char may_not_optimize;
312 /* Nonzero means reg N has already been moved out of one loop.
313 This reduces the desire to move it out of another. */
320 int num; /* Number of regs used in table. */
321 int size; /* Size of table. */
322 struct loop_reg *array; /* Register usage info. array. */
323 int multiple_uses; /* Nonzero if a reg has multiple uses. */
330 /* Head of movable chain. */
331 struct movable *head;
332 /* Last movable in chain. */
333 struct movable *last;
337 /* Information pertaining to a loop. */
341 /* Nonzero if there is a subroutine call in the current loop. */
343 /* Nonzero if there is a libcall in the current loop. */
345 /* Nonzero if there is a non constant call in the current loop. */
346 int has_nonconst_call;
347 /* Nonzero if there is a prefetch instruction in the current loop. */
349 /* Nonzero if there is a volatile memory reference in the current
352 /* Nonzero if there is a tablejump in the current loop. */
354 /* Nonzero if there are ways to leave the loop other than falling
356 int has_multiple_exit_targets;
357 /* Nonzero if there is an indirect jump in the current function. */
358 int has_indirect_jump;
359 /* Register or constant initial loop value. */
361 /* Register or constant value used for comparison test. */
362 rtx comparison_value;
363 /* Register or constant approximate final value. */
365 /* Register or constant initial loop value with term common to
366 final_value removed. */
367 rtx initial_equiv_value;
368 /* Register or constant final loop value with term common to
369 initial_value removed. */
370 rtx final_equiv_value;
371 /* Register corresponding to iteration variable. */
373 /* Constant loop increment. */
375 enum rtx_code comparison_code;
376 /* Holds the number of loop iterations. It is zero if the number
377 could not be calculated. Must be unsigned since the number of
378 iterations can be as high as 2^wordsize - 1. For loops with a
379 wider iterator, this number will be zero if the number of loop
380 iterations is too large for an unsigned integer to hold. */
381 unsigned HOST_WIDE_INT n_iterations;
382 int used_count_register;
383 /* The loop iterator induction variable. */
385 /* List of MEMs that are stored in this loop. */
387 /* Array of MEMs that are used (read or written) in this loop, but
388 cannot be aliased by anything in this loop, except perhaps
389 themselves. In other words, if mems[i] is altered during
390 the loop, it is altered by an expression that is rtx_equal_p to
393 /* The index of the next available slot in MEMS. */
395 /* The number of elements allocated in MEMS. */
397 /* Nonzero if we don't know what MEMs were changed in the current
398 loop. This happens if the loop contains a call (in which case
399 `has_call' will also be set) or if we store into more than
401 int unknown_address_altered;
402 /* The above doesn't count any readonly memory locations that are
403 stored. This does. */
404 int unknown_constant_address_altered;
405 /* Count of memory write instructions discovered in the loop. */
407 /* The insn where the first of these was found. */
408 rtx first_loop_store_insn;
409 /* The chain of movable insns in loop. */
410 struct loop_movables movables;
411 /* The registers used the in loop. */
412 struct loop_regs regs;
413 /* The induction variable information in loop. */
415 /* Nonzero if call is in pre_header extended basic block. */
416 int pre_header_has_call;
419 /* Not really meaningful values, but at least something. */
420 #ifndef SIMULTANEOUS_PREFETCHES
421 #define SIMULTANEOUS_PREFETCHES 3
423 #ifndef PREFETCH_BLOCK
424 #define PREFETCH_BLOCK 32
426 #ifndef HAVE_prefetch
427 #define HAVE_prefetch 0
428 #define CODE_FOR_prefetch 0
429 #define gen_prefetch(a,b,c) (gcc_unreachable (), NULL_RTX)
432 /* Give up the prefetch optimizations once we exceed a given threshold.
433 It is unlikely that we would be able to optimize something in a loop
434 with so many detected prefetches. */
435 #define MAX_PREFETCHES 100
436 /* The number of prefetch blocks that are beneficial to fetch at once before
437 a loop with a known (and low) iteration count. */
438 #define PREFETCH_BLOCKS_BEFORE_LOOP_MAX 6
439 /* For very tiny loops it is not worthwhile to prefetch even before the loop,
440 since it is likely that the data are already in the cache. */
441 #define PREFETCH_BLOCKS_BEFORE_LOOP_MIN 2
443 /* Parameterize some prefetch heuristics so they can be turned on and off
444 easily for performance testing on new architectures. These can be
445 defined in target-dependent files. */
447 /* Prefetch is worthwhile only when loads/stores are dense. */
448 #ifndef PREFETCH_ONLY_DENSE_MEM
449 #define PREFETCH_ONLY_DENSE_MEM 1
452 /* Define what we mean by "dense" loads and stores; This value divided by 256
453 is the minimum percentage of memory references that worth prefetching. */
454 #ifndef PREFETCH_DENSE_MEM
455 #define PREFETCH_DENSE_MEM 220
458 /* Do not prefetch for a loop whose iteration count is known to be low. */
459 #ifndef PREFETCH_NO_LOW_LOOPCNT
460 #define PREFETCH_NO_LOW_LOOPCNT 1
463 /* Define what we mean by a "low" iteration count. */
464 #ifndef PREFETCH_LOW_LOOPCNT
465 #define PREFETCH_LOW_LOOPCNT 32
468 /* Do not prefetch for a loop that contains a function call; such a loop is
469 probably not an internal loop. */
470 #ifndef PREFETCH_NO_CALL
471 #define PREFETCH_NO_CALL 1
474 /* Do not prefetch accesses with an extreme stride. */
475 #ifndef PREFETCH_NO_EXTREME_STRIDE
476 #define PREFETCH_NO_EXTREME_STRIDE 1
479 /* Define what we mean by an "extreme" stride. */
480 #ifndef PREFETCH_EXTREME_STRIDE
481 #define PREFETCH_EXTREME_STRIDE 4096
484 /* Define a limit to how far apart indices can be and still be merged
485 into a single prefetch. */
486 #ifndef PREFETCH_EXTREME_DIFFERENCE
487 #define PREFETCH_EXTREME_DIFFERENCE 4096
490 /* Issue prefetch instructions before the loop to fetch data to be used
491 in the first few loop iterations. */
492 #ifndef PREFETCH_BEFORE_LOOP
493 #define PREFETCH_BEFORE_LOOP 1
496 /* Do not handle reversed order prefetches (negative stride). */
497 #ifndef PREFETCH_NO_REVERSE_ORDER
498 #define PREFETCH_NO_REVERSE_ORDER 1
501 /* Prefetch even if the GIV is in conditional code. */
502 #ifndef PREFETCH_CONDITIONAL
503 #define PREFETCH_CONDITIONAL 1
506 #define LOOP_REG_LIFETIME(LOOP, REGNO) \
507 ((REGNO_LAST_LUID (REGNO) - REGNO_FIRST_LUID (REGNO)))
509 #define LOOP_REG_GLOBAL_P(LOOP, REGNO) \
510 ((REGNO_LAST_LUID (REGNO) > INSN_LUID ((LOOP)->end) \
511 || REGNO_FIRST_LUID (REGNO) < INSN_LUID ((LOOP)->start)))
513 #define LOOP_REGNO_NREGS(REGNO, SET_DEST) \
514 ((REGNO) < FIRST_PSEUDO_REGISTER \
515 ? (int) hard_regno_nregs[(REGNO)][GET_MODE (SET_DEST)] : 1)
518 /* Vector mapping INSN_UIDs to luids.
519 The luids are like uids but increase monotonically always.
520 We use them to see whether a jump comes from outside a given loop. */
522 static int *uid_luid;
524 /* Indexed by INSN_UID, contains the ordinal giving the (innermost) loop
525 number the insn is contained in. */
527 static struct loop **uid_loop;
529 /* 1 + largest uid of any insn. */
531 static int max_uid_for_loop;
533 /* Number of loops detected in current function. Used as index to the
536 static int max_loop_num;
538 /* Bound on pseudo register number before loop optimization.
539 A pseudo has valid regscan info if its number is < max_reg_before_loop. */
540 static unsigned int max_reg_before_loop;
542 /* The value to pass to the next call of reg_scan_update. */
543 static int loop_max_reg;
545 /* During the analysis of a loop, a chain of `struct movable's
546 is made to record all the movable insns found.
547 Then the entire chain can be scanned to decide which to move. */
551 rtx insn; /* A movable insn */
552 rtx set_src; /* The expression this reg is set from. */
553 rtx set_dest; /* The destination of this SET. */
554 rtx dependencies; /* When INSN is libcall, this is an EXPR_LIST
555 of any registers used within the LIBCALL. */
556 int consec; /* Number of consecutive following insns
557 that must be moved with this one. */
558 unsigned int regno; /* The register it sets */
559 short lifetime; /* lifetime of that register;
560 may be adjusted when matching movables
561 that load the same value are found. */
562 short savings; /* Number of insns we can move for this reg,
563 including other movables that force this
564 or match this one. */
565 ENUM_BITFIELD(machine_mode) savemode : 8; /* Nonzero means it is a mode for
566 a low part that we should avoid changing when
567 clearing the rest of the reg. */
568 unsigned int cond : 1; /* 1 if only conditionally movable */
569 unsigned int force : 1; /* 1 means MUST move this insn */
570 unsigned int global : 1; /* 1 means reg is live outside this loop */
571 /* If PARTIAL is 1, GLOBAL means something different:
572 that the reg is live outside the range from where it is set
573 to the following label. */
574 unsigned int done : 1; /* 1 inhibits further processing of this */
576 unsigned int partial : 1; /* 1 means this reg is used for zero-extending.
577 In particular, moving it does not make it
579 unsigned int move_insn : 1; /* 1 means that we call emit_move_insn to
580 load SRC, rather than copying INSN. */
581 unsigned int move_insn_first:1;/* Same as above, if this is necessary for the
582 first insn of a consecutive sets group. */
583 unsigned int is_equiv : 1; /* 1 means a REG_EQUIV is present on INSN. */
584 unsigned int insert_temp : 1; /* 1 means we copy to a new pseudo and replace
585 the original insn with a copy from that
586 pseudo, rather than deleting it. */
587 struct movable *match; /* First entry for same value */
588 struct movable *forces; /* An insn that must be moved if this is */
589 struct movable *next;
593 static FILE *loop_dump_stream;
595 /* Forward declarations. */
597 static void invalidate_loops_containing_label (rtx);
598 static void find_and_verify_loops (rtx, struct loops *);
599 static void mark_loop_jump (rtx, struct loop *);
600 static void prescan_loop (struct loop *);
601 static int reg_in_basic_block_p (rtx, rtx);
602 static int consec_sets_invariant_p (const struct loop *, rtx, int, rtx);
603 static int labels_in_range_p (rtx, int);
604 static void count_one_set (struct loop_regs *, rtx, rtx, rtx *);
605 static void note_addr_stored (rtx, rtx, void *);
606 static void note_set_pseudo_multiple_uses (rtx, rtx, void *);
607 static int loop_reg_used_before_p (const struct loop *, rtx, rtx);
608 static rtx find_regs_nested (rtx, rtx);
609 static void scan_loop (struct loop*, int);
611 static void replace_call_address (rtx, rtx, rtx);
613 static rtx skip_consec_insns (rtx, int);
614 static int libcall_benefit (rtx);
615 static rtx libcall_other_reg (rtx, rtx);
616 static void record_excess_regs (rtx, rtx, rtx *);
617 static void ignore_some_movables (struct loop_movables *);
618 static void force_movables (struct loop_movables *);
619 static void combine_movables (struct loop_movables *, struct loop_regs *);
620 static int num_unmoved_movables (const struct loop *);
621 static int regs_match_p (rtx, rtx, struct loop_movables *);
622 static int rtx_equal_for_loop_p (rtx, rtx, struct loop_movables *,
624 static void add_label_notes (rtx, rtx);
625 static void move_movables (struct loop *loop, struct loop_movables *, int,
627 static void loop_movables_add (struct loop_movables *, struct movable *);
628 static void loop_movables_free (struct loop_movables *);
629 static int count_nonfixed_reads (const struct loop *, rtx);
630 static void loop_bivs_find (struct loop *);
631 static void loop_bivs_init_find (struct loop *);
632 static void loop_bivs_check (struct loop *);
633 static void loop_givs_find (struct loop *);
634 static void loop_givs_check (struct loop *);
635 static int loop_biv_eliminable_p (struct loop *, struct iv_class *, int, int);
636 static int loop_giv_reduce_benefit (struct loop *, struct iv_class *,
637 struct induction *, rtx);
638 static void loop_givs_dead_check (struct loop *, struct iv_class *);
639 static void loop_givs_reduce (struct loop *, struct iv_class *);
640 static void loop_givs_rescan (struct loop *, struct iv_class *, rtx *);
641 static void loop_ivs_free (struct loop *);
642 static void strength_reduce (struct loop *, int);
643 static void find_single_use_in_loop (struct loop_regs *, rtx, rtx);
644 static int valid_initial_value_p (rtx, rtx, int, rtx);
645 static void find_mem_givs (const struct loop *, rtx, rtx, int, int);
646 static void record_biv (struct loop *, struct induction *, rtx, rtx, rtx,
647 rtx, rtx *, int, int);
648 static void check_final_value (const struct loop *, struct induction *);
649 static void loop_ivs_dump (const struct loop *, FILE *, int);
650 static void loop_iv_class_dump (const struct iv_class *, FILE *, int);
651 static void loop_biv_dump (const struct induction *, FILE *, int);
652 static void loop_giv_dump (const struct induction *, FILE *, int);
653 static void record_giv (const struct loop *, struct induction *, rtx, rtx,
654 rtx, rtx, rtx, rtx, int, enum g_types, int, int,
656 static void update_giv_derive (const struct loop *, rtx);
657 static HOST_WIDE_INT get_monotonic_increment (struct iv_class *);
658 static bool biased_biv_fits_mode_p (const struct loop *, struct iv_class *,
659 HOST_WIDE_INT, enum machine_mode,
660 unsigned HOST_WIDE_INT);
661 static bool biv_fits_mode_p (const struct loop *, struct iv_class *,
662 HOST_WIDE_INT, enum machine_mode, bool);
663 static bool extension_within_bounds_p (const struct loop *, struct iv_class *,
665 static void check_ext_dependent_givs (const struct loop *, struct iv_class *);
666 static int basic_induction_var (const struct loop *, rtx, enum machine_mode,
667 rtx, rtx, rtx *, rtx *, rtx **);
668 static rtx simplify_giv_expr (const struct loop *, rtx, rtx *, int *);
669 static int general_induction_var (const struct loop *loop, rtx, rtx *, rtx *,
670 rtx *, rtx *, int, int *, enum machine_mode);
671 static int consec_sets_giv (const struct loop *, int, rtx, rtx, rtx, rtx *,
672 rtx *, rtx *, rtx *);
673 static int check_dbra_loop (struct loop *, int);
674 static rtx express_from_1 (rtx, rtx, rtx);
675 static rtx combine_givs_p (struct induction *, struct induction *);
676 static int cmp_combine_givs_stats (const void *, const void *);
677 static void combine_givs (struct loop_regs *, struct iv_class *);
678 static int product_cheap_p (rtx, rtx);
679 static int maybe_eliminate_biv (const struct loop *, struct iv_class *, int,
681 static int maybe_eliminate_biv_1 (const struct loop *, rtx, rtx,
682 struct iv_class *, int, basic_block, rtx);
683 static int last_use_this_basic_block (rtx, rtx);
684 static void record_initial (rtx, rtx, void *);
685 static void update_reg_last_use (rtx, rtx);
686 static rtx next_insn_in_loop (const struct loop *, rtx);
687 static void loop_regs_scan (const struct loop *, int);
688 static int count_insns_in_loop (const struct loop *);
689 static int find_mem_in_note_1 (rtx *, void *);
690 static rtx find_mem_in_note (rtx);
691 static void load_mems (const struct loop *);
692 static int insert_loop_mem (rtx *, void *);
693 static int replace_loop_mem (rtx *, void *);
694 static void replace_loop_mems (rtx, rtx, rtx, int);
695 static int replace_loop_reg (rtx *, void *);
696 static void replace_loop_regs (rtx insn, rtx, rtx);
697 static void note_reg_stored (rtx, rtx, void *);
698 static void try_copy_prop (const struct loop *, rtx, unsigned int);
699 static void try_swap_copy_prop (const struct loop *, rtx, unsigned int);
700 static rtx check_insn_for_givs (struct loop *, rtx, int, int);
701 static rtx check_insn_for_bivs (struct loop *, rtx, int, int);
702 static rtx gen_add_mult (rtx, rtx, rtx, rtx);
703 static void loop_regs_update (const struct loop *, rtx);
704 static int iv_add_mult_cost (rtx, rtx, rtx, rtx);
705 static int loop_invariant_p (const struct loop *, rtx);
706 static rtx loop_insn_hoist (const struct loop *, rtx);
707 static void loop_iv_add_mult_emit_before (const struct loop *, rtx, rtx, rtx,
708 rtx, basic_block, rtx);
709 static rtx loop_insn_emit_before (const struct loop *, basic_block,
711 static int loop_insn_first_p (rtx, rtx);
712 static rtx get_condition_for_loop (const struct loop *, rtx);
713 static void loop_iv_add_mult_sink (const struct loop *, rtx, rtx, rtx, rtx);
714 static void loop_iv_add_mult_hoist (const struct loop *, rtx, rtx, rtx, rtx);
715 static rtx extend_value_for_giv (struct induction *, rtx);
716 static rtx loop_insn_sink (const struct loop *, rtx);
718 static rtx loop_insn_emit_after (const struct loop *, basic_block, rtx, rtx);
719 static rtx loop_call_insn_emit_before (const struct loop *, basic_block,
721 static rtx loop_call_insn_hoist (const struct loop *, rtx);
722 static rtx loop_insn_sink_or_swim (const struct loop *, rtx);
724 static void loop_dump_aux (const struct loop *, FILE *, int);
725 static void loop_delete_insns (rtx, rtx);
726 static HOST_WIDE_INT remove_constant_addition (rtx *);
727 static rtx gen_load_of_final_value (rtx, rtx);
728 void debug_ivs (const struct loop *);
729 void debug_iv_class (const struct iv_class *);
730 void debug_biv (const struct induction *);
731 void debug_giv (const struct induction *);
732 void debug_loop (const struct loop *);
733 void debug_loops (const struct loops *);
735 typedef struct loop_replace_args
742 /* Nonzero iff INSN is between START and END, inclusive. */
743 #define INSN_IN_RANGE_P(INSN, START, END) \
744 (INSN_UID (INSN) < max_uid_for_loop \
745 && INSN_LUID (INSN) >= INSN_LUID (START) \
746 && INSN_LUID (INSN) <= INSN_LUID (END))
748 /* Indirect_jump_in_function is computed once per function. */
749 static int indirect_jump_in_function;
750 static int indirect_jump_in_function_p (rtx);
752 static int compute_luids (rtx, rtx, int);
754 static int biv_elimination_giv_has_0_offset (struct induction *,
755 struct induction *, rtx);
757 /* Benefit penalty, if a giv is not replaceable, i.e. must emit an insn to
758 copy the value of the strength reduced giv to its original register. */
759 static int copy_cost;
761 /* Cost of using a register, to normalize the benefits of a giv. */
762 static int reg_address_cost;
767 rtx reg = gen_rtx_REG (word_mode, LAST_VIRTUAL_REGISTER + 1);
769 reg_address_cost = address_cost (reg, SImode);
771 copy_cost = COSTS_N_INSNS (1);
774 /* Compute the mapping from uids to luids.
775 LUIDs are numbers assigned to insns, like uids,
776 except that luids increase monotonically through the code.
777 Start at insn START and stop just before END. Assign LUIDs
778 starting with PREV_LUID + 1. Return the last assigned LUID + 1. */
780 compute_luids (rtx start, rtx end, int prev_luid)
785 for (insn = start, i = prev_luid; insn != end; insn = NEXT_INSN (insn))
787 if (INSN_UID (insn) >= max_uid_for_loop)
789 /* Don't assign luids to line-number NOTEs, so that the distance in
790 luids between two insns is not affected by -g. */
792 || NOTE_LINE_NUMBER (insn) <= 0)
793 uid_luid[INSN_UID (insn)] = ++i;
795 /* Give a line number note the same luid as preceding insn. */
796 uid_luid[INSN_UID (insn)] = i;
801 /* Entry point of this file. Perform loop optimization
802 on the current function. F is the first insn of the function
803 and DUMPFILE is a stream for output of a trace of actions taken
804 (or 0 if none should be output). */
807 loop_optimize (rtx f, FILE *dumpfile, int flags)
811 struct loops loops_data;
812 struct loops *loops = &loops_data;
813 struct loop_info *loops_info;
815 loop_dump_stream = dumpfile;
817 init_recog_no_volatile ();
819 max_reg_before_loop = max_reg_num ();
820 loop_max_reg = max_reg_before_loop;
824 /* Count the number of loops. */
827 for (insn = f; insn; insn = NEXT_INSN (insn))
830 && NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG)
834 /* Don't waste time if no loops. */
835 if (max_loop_num == 0)
838 loops->num = max_loop_num;
840 /* Get size to use for tables indexed by uids.
841 Leave some space for labels allocated by find_and_verify_loops. */
842 max_uid_for_loop = get_max_uid () + 1 + max_loop_num * 32;
844 uid_luid = xcalloc (max_uid_for_loop, sizeof (int));
845 uid_loop = xcalloc (max_uid_for_loop, sizeof (struct loop *));
847 /* Allocate storage for array of loops. */
848 loops->array = xcalloc (loops->num, sizeof (struct loop));
850 /* Find and process each loop.
851 First, find them, and record them in order of their beginnings. */
852 find_and_verify_loops (f, loops);
854 /* Allocate and initialize auxiliary loop information. */
855 loops_info = xcalloc (loops->num, sizeof (struct loop_info));
856 for (i = 0; i < (int) loops->num; i++)
857 loops->array[i].aux = loops_info + i;
859 /* Now find all register lifetimes. This must be done after
860 find_and_verify_loops, because it might reorder the insns in the
862 reg_scan (f, max_reg_before_loop);
864 /* This must occur after reg_scan so that registers created by gcse
865 will have entries in the register tables.
867 We could have added a call to reg_scan after gcse_main in toplev.c,
868 but moving this call to init_alias_analysis is more efficient. */
869 init_alias_analysis ();
871 /* See if we went too far. Note that get_max_uid already returns
872 one more that the maximum uid of all insn. */
873 gcc_assert (get_max_uid () <= max_uid_for_loop);
874 /* Now reset it to the actual size we need. See above. */
875 max_uid_for_loop = get_max_uid ();
877 /* find_and_verify_loops has already called compute_luids, but it
878 might have rearranged code afterwards, so we need to recompute
880 compute_luids (f, NULL_RTX, 0);
882 /* Don't leave gaps in uid_luid for insns that have been
883 deleted. It is possible that the first or last insn
884 using some register has been deleted by cross-jumping.
885 Make sure that uid_luid for that former insn's uid
886 points to the general area where that insn used to be. */
887 for (i = 0; i < max_uid_for_loop; i++)
889 uid_luid[0] = uid_luid[i];
890 if (uid_luid[0] != 0)
893 for (i = 0; i < max_uid_for_loop; i++)
894 if (uid_luid[i] == 0)
895 uid_luid[i] = uid_luid[i - 1];
897 /* Determine if the function has indirect jump. On some systems
898 this prevents low overhead loop instructions from being used. */
899 indirect_jump_in_function = indirect_jump_in_function_p (f);
901 /* Now scan the loops, last ones first, since this means inner ones are done
902 before outer ones. */
903 for (i = max_loop_num - 1; i >= 0; i--)
905 struct loop *loop = &loops->array[i];
907 if (! loop->invalid && loop->end)
909 scan_loop (loop, flags);
914 end_alias_analysis ();
917 for (i = 0; i < (int) loops->num; i++)
918 free (loops_info[i].mems);
926 /* Returns the next insn, in execution order, after INSN. START and
927 END are the NOTE_INSN_LOOP_BEG and NOTE_INSN_LOOP_END for the loop,
928 respectively. LOOP->TOP, if non-NULL, is the top of the loop in the
929 insn-stream; it is used with loops that are entered near the
933 next_insn_in_loop (const struct loop *loop, rtx insn)
935 insn = NEXT_INSN (insn);
937 if (insn == loop->end)
940 /* Go to the top of the loop, and continue there. */
947 if (insn == loop->scan_start)
954 /* Find any register references hidden inside X and add them to
955 the dependency list DEPS. This is used to look inside CLOBBER (MEM
956 when checking whether a PARALLEL can be pulled out of a loop. */
959 find_regs_nested (rtx deps, rtx x)
961 enum rtx_code code = GET_CODE (x);
963 deps = gen_rtx_EXPR_LIST (VOIDmode, x, deps);
966 const char *fmt = GET_RTX_FORMAT (code);
968 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
971 deps = find_regs_nested (deps, XEXP (x, i));
972 else if (fmt[i] == 'E')
973 for (j = 0; j < XVECLEN (x, i); j++)
974 deps = find_regs_nested (deps, XVECEXP (x, i, j));
980 /* Optimize one loop described by LOOP. */
982 /* ??? Could also move memory writes out of loops if the destination address
983 is invariant, the source is invariant, the memory write is not volatile,
984 and if we can prove that no read inside the loop can read this address
985 before the write occurs. If there is a read of this address after the
986 write, then we can also mark the memory read as invariant. */
989 scan_loop (struct loop *loop, int flags)
991 struct loop_info *loop_info = LOOP_INFO (loop);
992 struct loop_regs *regs = LOOP_REGS (loop);
994 rtx loop_start = loop->start;
995 rtx loop_end = loop->end;
997 /* 1 if we are scanning insns that could be executed zero times. */
999 /* 1 if we are scanning insns that might never be executed
1000 due to a subroutine call which might exit before they are reached. */
1001 int call_passed = 0;
1002 /* Number of insns in the loop. */
1005 rtx temp, update_start, update_end;
1006 /* The SET from an insn, if it is the only SET in the insn. */
1008 /* Chain describing insns movable in current loop. */
1009 struct loop_movables *movables = LOOP_MOVABLES (loop);
1010 /* Ratio of extra register life span we can justify
1011 for saving an instruction. More if loop doesn't call subroutines
1012 since in that case saving an insn makes more difference
1013 and more registers are available. */
1022 /* Determine whether this loop starts with a jump down to a test at
1023 the end. This will occur for a small number of loops with a test
1024 that is too complex to duplicate in front of the loop.
1026 We search for the first insn or label in the loop, skipping NOTEs.
1027 However, we must be careful not to skip past a NOTE_INSN_LOOP_BEG
1028 (because we might have a loop executed only once that contains a
1029 loop which starts with a jump to its exit test) or a NOTE_INSN_LOOP_END
1030 (in case we have a degenerate loop).
1032 Note that if we mistakenly think that a loop is entered at the top
1033 when, in fact, it is entered at the exit test, the only effect will be
1034 slightly poorer optimization. Making the opposite error can generate
1035 incorrect code. Since very few loops now start with a jump to the
1036 exit test, the code here to detect that case is very conservative. */
1038 for (p = NEXT_INSN (loop_start);
1040 && !LABEL_P (p) && ! INSN_P (p)
1042 || (NOTE_LINE_NUMBER (p) != NOTE_INSN_LOOP_BEG
1043 && NOTE_LINE_NUMBER (p) != NOTE_INSN_LOOP_END));
1047 loop->scan_start = p;
1049 /* If loop end is the end of the current function, then emit a
1050 NOTE_INSN_DELETED after loop_end and set loop->sink to the dummy
1051 note insn. This is the position we use when sinking insns out of
1053 if (NEXT_INSN (loop->end) != 0)
1054 loop->sink = NEXT_INSN (loop->end);
1056 loop->sink = emit_note_after (NOTE_INSN_DELETED, loop->end);
1058 /* Set up variables describing this loop. */
1059 prescan_loop (loop);
1060 threshold = (loop_info->has_call ? 1 : 2) * (1 + n_non_fixed_regs);
1062 /* If loop has a jump before the first label,
1063 the true entry is the target of that jump.
1064 Start scan from there.
1065 But record in LOOP->TOP the place where the end-test jumps
1066 back to so we can scan that after the end of the loop. */
1068 /* Loop entry must be unconditional jump (and not a RETURN) */
1069 && any_uncondjump_p (p)
1070 && JUMP_LABEL (p) != 0
1071 /* Check to see whether the jump actually
1072 jumps out of the loop (meaning it's no loop).
1073 This case can happen for things like
1074 do {..} while (0). If this label was generated previously
1075 by loop, we can't tell anything about it and have to reject
1077 && INSN_IN_RANGE_P (JUMP_LABEL (p), loop_start, loop_end))
1079 loop->top = next_label (loop->scan_start);
1080 loop->scan_start = JUMP_LABEL (p);
1083 /* If LOOP->SCAN_START was an insn created by loop, we don't know its luid
1084 as required by loop_reg_used_before_p. So skip such loops. (This
1085 test may never be true, but it's best to play it safe.)
1087 Also, skip loops where we do not start scanning at a label. This
1088 test also rejects loops starting with a JUMP_INSN that failed the
1091 if (INSN_UID (loop->scan_start) >= max_uid_for_loop
1092 || !LABEL_P (loop->scan_start))
1094 if (loop_dump_stream)
1095 fprintf (loop_dump_stream, "\nLoop from %d to %d is phony.\n\n",
1096 INSN_UID (loop_start), INSN_UID (loop_end));
1100 /* Allocate extra space for REGs that might be created by load_mems.
1101 We allocate a little extra slop as well, in the hopes that we
1102 won't have to reallocate the regs array. */
1103 loop_regs_scan (loop, loop_info->mems_idx + 16);
1104 insn_count = count_insns_in_loop (loop);
1106 if (loop_dump_stream)
1107 fprintf (loop_dump_stream, "\nLoop from %d to %d: %d real insns.\n",
1108 INSN_UID (loop_start), INSN_UID (loop_end), insn_count);
1110 /* Scan through the loop finding insns that are safe to move.
1111 Set REGS->ARRAY[I].SET_IN_LOOP negative for the reg I being set, so that
1112 this reg will be considered invariant for subsequent insns.
1113 We consider whether subsequent insns use the reg
1114 in deciding whether it is worth actually moving.
1116 MAYBE_NEVER is nonzero if we have passed a conditional jump insn
1117 and therefore it is possible that the insns we are scanning
1118 would never be executed. At such times, we must make sure
1119 that it is safe to execute the insn once instead of zero times.
1120 When MAYBE_NEVER is 0, all insns will be executed at least once
1121 so that is not a problem. */
1123 for (in_libcall = 0, p = next_insn_in_loop (loop, loop->scan_start);
1125 p = next_insn_in_loop (loop, p))
1127 if (in_libcall && INSN_P (p) && find_reg_note (p, REG_RETVAL, NULL_RTX))
1129 if (NONJUMP_INSN_P (p))
1131 /* Do not scan past an optimization barrier. */
1132 if (GET_CODE (PATTERN (p)) == ASM_INPUT)
1134 temp = find_reg_note (p, REG_LIBCALL, NULL_RTX);
1138 && (set = single_set (p))
1139 && REG_P (SET_DEST (set))
1140 && SET_DEST (set) != frame_pointer_rtx
1141 #ifdef PIC_OFFSET_TABLE_REG_CALL_CLOBBERED
1142 && SET_DEST (set) != pic_offset_table_rtx
1144 && ! regs->array[REGNO (SET_DEST (set))].may_not_optimize)
1149 int insert_temp = 0;
1150 rtx src = SET_SRC (set);
1151 rtx dependencies = 0;
1153 /* Figure out what to use as a source of this insn. If a
1154 REG_EQUIV note is given or if a REG_EQUAL note with a
1155 constant operand is specified, use it as the source and
1156 mark that we should move this insn by calling
1157 emit_move_insn rather that duplicating the insn.
1159 Otherwise, only use the REG_EQUAL contents if a REG_RETVAL
1161 temp = find_reg_note (p, REG_EQUIV, NULL_RTX);
1163 src = XEXP (temp, 0), move_insn = 1;
1166 temp = find_reg_note (p, REG_EQUAL, NULL_RTX);
1167 if (temp && CONSTANT_P (XEXP (temp, 0)))
1168 src = XEXP (temp, 0), move_insn = 1;
1169 if (temp && find_reg_note (p, REG_RETVAL, NULL_RTX))
1171 src = XEXP (temp, 0);
1172 /* A libcall block can use regs that don't appear in
1173 the equivalent expression. To move the libcall,
1174 we must move those regs too. */
1175 dependencies = libcall_other_reg (p, src);
1179 /* For parallels, add any possible uses to the dependencies, as
1180 we can't move the insn without resolving them first.
1181 MEMs inside CLOBBERs may also reference registers; these
1182 count as implicit uses. */
1183 if (GET_CODE (PATTERN (p)) == PARALLEL)
1185 for (i = 0; i < XVECLEN (PATTERN (p), 0); i++)
1187 rtx x = XVECEXP (PATTERN (p), 0, i);
1188 if (GET_CODE (x) == USE)
1190 = gen_rtx_EXPR_LIST (VOIDmode, XEXP (x, 0),
1192 else if (GET_CODE (x) == CLOBBER
1193 && MEM_P (XEXP (x, 0)))
1194 dependencies = find_regs_nested (dependencies,
1195 XEXP (XEXP (x, 0), 0));
1199 if (/* The register is used in basic blocks other
1200 than the one where it is set (meaning that
1201 something after this point in the loop might
1202 depend on its value before the set). */
1203 ! reg_in_basic_block_p (p, SET_DEST (set))
1204 /* And the set is not guaranteed to be executed once
1205 the loop starts, or the value before the set is
1206 needed before the set occurs...
1208 ??? Note we have quadratic behavior here, mitigated
1209 by the fact that the previous test will often fail for
1210 large loops. Rather than re-scanning the entire loop
1211 each time for register usage, we should build tables
1212 of the register usage and use them here instead. */
1214 || loop_reg_used_before_p (loop, set, p)))
1215 /* It is unsafe to move the set. However, it may be OK to
1216 move the source into a new pseudo, and substitute a
1217 reg-to-reg copy for the original insn.
1219 This code used to consider it OK to move a set of a variable
1220 which was not created by the user and not used in an exit
1222 That behavior is incorrect and was removed. */
1225 /* Don't try to optimize a MODE_CC set with a constant
1226 source. It probably will be combined with a conditional
1228 if (GET_MODE_CLASS (GET_MODE (SET_DEST (set))) == MODE_CC
1229 && CONSTANT_P (src))
1231 /* Don't try to optimize a register that was made
1232 by loop-optimization for an inner loop.
1233 We don't know its life-span, so we can't compute
1235 else if (REGNO (SET_DEST (set)) >= max_reg_before_loop)
1237 /* Don't move the source and add a reg-to-reg copy:
1238 - with -Os (this certainly increases size),
1239 - if the mode doesn't support copy operations (obviously),
1240 - if the source is already a reg (the motion will gain nothing),
1241 - if the source is a legitimate constant (likewise). */
1242 else if (insert_temp
1244 || ! can_copy_p (GET_MODE (SET_SRC (set)))
1245 || REG_P (SET_SRC (set))
1246 || (CONSTANT_P (SET_SRC (set))
1247 && LEGITIMATE_CONSTANT_P (SET_SRC (set)))))
1249 else if ((tem = loop_invariant_p (loop, src))
1250 && (dependencies == 0
1252 = loop_invariant_p (loop, dependencies)) != 0)
1253 && (regs->array[REGNO (SET_DEST (set))].set_in_loop == 1
1255 = consec_sets_invariant_p
1256 (loop, SET_DEST (set),
1257 regs->array[REGNO (SET_DEST (set))].set_in_loop,
1259 /* If the insn can cause a trap (such as divide by zero),
1260 can't move it unless it's guaranteed to be executed
1261 once loop is entered. Even a function call might
1262 prevent the trap insn from being reached
1263 (since it might exit!) */
1264 && ! ((maybe_never || call_passed)
1265 && may_trap_p (src)))
1268 int regno = REGNO (SET_DEST (set));
1271 /* A potential lossage is where we have a case where two
1272 insns can be combined as long as they are both in the
1273 loop, but we move one of them outside the loop. For
1274 large loops, this can lose. The most common case of
1275 this is the address of a function being called.
1277 Therefore, if this register is marked as being used
1278 exactly once if we are in a loop with calls
1279 (a "large loop"), see if we can replace the usage of
1280 this register with the source of this SET. If we can,
1284 (1) P has a REG_RETVAL note or
1285 (2) if we have SMALL_REGISTER_CLASSES and
1286 (a) SET_SRC is a hard register or
1287 (b) the destination of the user is a hard register. */
1289 if (loop_info->has_call
1290 && regno >= FIRST_PSEUDO_REGISTER
1291 && (user = regs->array[regno].single_usage) != NULL
1292 && user != const0_rtx
1293 && REGNO_FIRST_UID (regno) == INSN_UID (p)
1294 && REGNO_LAST_UID (regno) == INSN_UID (user)
1295 && regs->array[regno].set_in_loop == 1
1296 && GET_CODE (SET_SRC (set)) != ASM_OPERANDS
1297 && ! side_effects_p (SET_SRC (set))
1298 && ! find_reg_note (p, REG_RETVAL, NULL_RTX)
1299 && (!SMALL_REGISTER_CLASSES
1300 || !REG_P (SET_SRC (set))
1301 || !HARD_REGISTER_P (SET_SRC (set)))
1302 && (!SMALL_REGISTER_CLASSES
1303 || !NONJUMP_INSN_P (user)
1304 || !(user_set = single_set (user))
1305 || !REG_P (SET_DEST (user_set))
1306 || !HARD_REGISTER_P (SET_DEST (user_set)))
1307 /* This test is not redundant; SET_SRC (set) might be
1308 a call-clobbered register and the life of REGNO
1309 might span a call. */
1310 && ! modified_between_p (SET_SRC (set), p, user)
1311 && no_labels_between_p (p, user)
1312 && validate_replace_rtx (SET_DEST (set),
1313 SET_SRC (set), user))
1315 /* Replace any usage in a REG_EQUAL note. Must copy
1316 the new source, so that we don't get rtx sharing
1317 between the SET_SOURCE and REG_NOTES of insn p. */
1319 = replace_rtx (REG_NOTES (user), SET_DEST (set),
1320 copy_rtx (SET_SRC (set)));
1323 for (i = 0; i < LOOP_REGNO_NREGS (regno, SET_DEST (set));
1325 regs->array[regno+i].set_in_loop = 0;
1329 m = xmalloc (sizeof (struct movable));
1333 m->dependencies = dependencies;
1334 m->set_dest = SET_DEST (set);
1337 = regs->array[REGNO (SET_DEST (set))].set_in_loop - 1;
1341 m->move_insn = move_insn;
1342 m->move_insn_first = 0;
1343 m->insert_temp = insert_temp;
1344 m->is_equiv = (find_reg_note (p, REG_EQUIV, NULL_RTX) != 0);
1345 m->savemode = VOIDmode;
1347 /* Set M->cond if either loop_invariant_p
1348 or consec_sets_invariant_p returned 2
1349 (only conditionally invariant). */
1350 m->cond = ((tem | tem1 | tem2) > 1);
1351 m->global = LOOP_REG_GLOBAL_P (loop, regno);
1353 m->lifetime = LOOP_REG_LIFETIME (loop, regno);
1354 m->savings = regs->array[regno].n_times_set;
1355 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
1356 m->savings += libcall_benefit (p);
1357 for (i = 0; i < LOOP_REGNO_NREGS (regno, SET_DEST (set)); i++)
1358 regs->array[regno+i].set_in_loop = move_insn ? -2 : -1;
1359 /* Add M to the end of the chain MOVABLES. */
1360 loop_movables_add (movables, m);
1364 /* It is possible for the first instruction to have a
1365 REG_EQUAL note but a non-invariant SET_SRC, so we must
1366 remember the status of the first instruction in case
1367 the last instruction doesn't have a REG_EQUAL note. */
1368 m->move_insn_first = m->move_insn;
1370 /* Skip this insn, not checking REG_LIBCALL notes. */
1371 p = next_nonnote_insn (p);
1372 /* Skip the consecutive insns, if there are any. */
1373 p = skip_consec_insns (p, m->consec);
1374 /* Back up to the last insn of the consecutive group. */
1375 p = prev_nonnote_insn (p);
1377 /* We must now reset m->move_insn, m->is_equiv, and
1378 possibly m->set_src to correspond to the effects of
1380 temp = find_reg_note (p, REG_EQUIV, NULL_RTX);
1382 m->set_src = XEXP (temp, 0), m->move_insn = 1;
1385 temp = find_reg_note (p, REG_EQUAL, NULL_RTX);
1386 if (temp && CONSTANT_P (XEXP (temp, 0)))
1387 m->set_src = XEXP (temp, 0), m->move_insn = 1;
1393 = (find_reg_note (p, REG_EQUIV, NULL_RTX) != 0);
1396 /* If this register is always set within a STRICT_LOW_PART
1397 or set to zero, then its high bytes are constant.
1398 So clear them outside the loop and within the loop
1399 just load the low bytes.
1400 We must check that the machine has an instruction to do so.
1401 Also, if the value loaded into the register
1402 depends on the same register, this cannot be done. */
1403 else if (SET_SRC (set) == const0_rtx
1404 && NONJUMP_INSN_P (NEXT_INSN (p))
1405 && (set1 = single_set (NEXT_INSN (p)))
1406 && GET_CODE (set1) == SET
1407 && (GET_CODE (SET_DEST (set1)) == STRICT_LOW_PART)
1408 && (GET_CODE (XEXP (SET_DEST (set1), 0)) == SUBREG)
1409 && (SUBREG_REG (XEXP (SET_DEST (set1), 0))
1411 && !reg_mentioned_p (SET_DEST (set), SET_SRC (set1)))
1413 int regno = REGNO (SET_DEST (set));
1414 if (regs->array[regno].set_in_loop == 2)
1417 m = xmalloc (sizeof (struct movable));
1420 m->set_dest = SET_DEST (set);
1421 m->dependencies = 0;
1427 m->move_insn_first = 0;
1428 m->insert_temp = insert_temp;
1430 /* If the insn may not be executed on some cycles,
1431 we can't clear the whole reg; clear just high part.
1432 Not even if the reg is used only within this loop.
1439 Clearing x before the inner loop could clobber a value
1440 being saved from the last time around the outer loop.
1441 However, if the reg is not used outside this loop
1442 and all uses of the register are in the same
1443 basic block as the store, there is no problem.
1445 If this insn was made by loop, we don't know its
1446 INSN_LUID and hence must make a conservative
1448 m->global = (INSN_UID (p) >= max_uid_for_loop
1449 || LOOP_REG_GLOBAL_P (loop, regno)
1450 || (labels_in_range_p
1451 (p, REGNO_FIRST_LUID (regno))));
1452 if (maybe_never && m->global)
1453 m->savemode = GET_MODE (SET_SRC (set1));
1455 m->savemode = VOIDmode;
1459 m->lifetime = LOOP_REG_LIFETIME (loop, regno);
1462 i < LOOP_REGNO_NREGS (regno, SET_DEST (set));
1464 regs->array[regno+i].set_in_loop = -1;
1465 /* Add M to the end of the chain MOVABLES. */
1466 loop_movables_add (movables, m);
1471 /* Past a call insn, we get to insns which might not be executed
1472 because the call might exit. This matters for insns that trap.
1473 Constant and pure call insns always return, so they don't count. */
1474 else if (CALL_P (p) && ! CONST_OR_PURE_CALL_P (p))
1476 /* Past a label or a jump, we get to insns for which we
1477 can't count on whether or how many times they will be
1478 executed during each iteration. Therefore, we can
1479 only move out sets of trivial variables
1480 (those not used after the loop). */
1481 /* Similar code appears twice in strength_reduce. */
1482 else if ((LABEL_P (p) || JUMP_P (p))
1483 /* If we enter the loop in the middle, and scan around to the
1484 beginning, don't set maybe_never for that. This must be an
1485 unconditional jump, otherwise the code at the top of the
1486 loop might never be executed. Unconditional jumps are
1487 followed by a barrier then the loop_end. */
1488 && ! (JUMP_P (p) && JUMP_LABEL (p) == loop->top
1489 && NEXT_INSN (NEXT_INSN (p)) == loop_end
1490 && any_uncondjump_p (p)))
1494 /* If one movable subsumes another, ignore that other. */
1496 ignore_some_movables (movables);
1498 /* For each movable insn, see if the reg that it loads
1499 leads when it dies right into another conditionally movable insn.
1500 If so, record that the second insn "forces" the first one,
1501 since the second can be moved only if the first is. */
1503 force_movables (movables);
1505 /* See if there are multiple movable insns that load the same value.
1506 If there are, make all but the first point at the first one
1507 through the `match' field, and add the priorities of them
1508 all together as the priority of the first. */
1510 combine_movables (movables, regs);
1512 /* Now consider each movable insn to decide whether it is worth moving.
1513 Store 0 in regs->array[I].set_in_loop for each reg I that is moved.
1515 For machines with few registers this increases code size, so do not
1516 move moveables when optimizing for code size on such machines.
1517 (The 18 below is the value for i386.) */
1520 || (reg_class_size[GENERAL_REGS] > 18 && !loop_info->has_call))
1522 move_movables (loop, movables, threshold, insn_count);
1524 /* Recalculate regs->array if move_movables has created new
1526 if (max_reg_num () > regs->num)
1528 loop_regs_scan (loop, 0);
1529 for (update_start = loop_start;
1530 PREV_INSN (update_start)
1531 && !LABEL_P (PREV_INSN (update_start));
1532 update_start = PREV_INSN (update_start))
1534 update_end = NEXT_INSN (loop_end);
1536 reg_scan_update (update_start, update_end, loop_max_reg);
1537 loop_max_reg = max_reg_num ();
1541 /* Now candidates that still are negative are those not moved.
1542 Change regs->array[I].set_in_loop to indicate that those are not actually
1544 for (i = 0; i < regs->num; i++)
1545 if (regs->array[i].set_in_loop < 0)
1546 regs->array[i].set_in_loop = regs->array[i].n_times_set;
1548 /* Now that we've moved some things out of the loop, we might be able to
1549 hoist even more memory references. */
1552 /* Recalculate regs->array if load_mems has created new registers. */
1553 if (max_reg_num () > regs->num)
1554 loop_regs_scan (loop, 0);
1556 for (update_start = loop_start;
1557 PREV_INSN (update_start)
1558 && !LABEL_P (PREV_INSN (update_start));
1559 update_start = PREV_INSN (update_start))
1561 update_end = NEXT_INSN (loop_end);
1563 reg_scan_update (update_start, update_end, loop_max_reg);
1564 loop_max_reg = max_reg_num ();
1566 if (flag_strength_reduce)
1568 if (update_end && LABEL_P (update_end))
1569 /* Ensure our label doesn't go away. */
1570 LABEL_NUSES (update_end)++;
1572 strength_reduce (loop, flags);
1574 reg_scan_update (update_start, update_end, loop_max_reg);
1575 loop_max_reg = max_reg_num ();
1577 if (update_end && LABEL_P (update_end)
1578 && --LABEL_NUSES (update_end) == 0)
1579 delete_related_insns (update_end);
1583 /* The movable information is required for strength reduction. */
1584 loop_movables_free (movables);
1591 /* Add elements to *OUTPUT to record all the pseudo-regs
1592 mentioned in IN_THIS but not mentioned in NOT_IN_THIS. */
1595 record_excess_regs (rtx in_this, rtx not_in_this, rtx *output)
1601 code = GET_CODE (in_this);
1615 if (REGNO (in_this) >= FIRST_PSEUDO_REGISTER
1616 && ! reg_mentioned_p (in_this, not_in_this))
1617 *output = gen_rtx_EXPR_LIST (VOIDmode, in_this, *output);
1624 fmt = GET_RTX_FORMAT (code);
1625 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1632 for (j = 0; j < XVECLEN (in_this, i); j++)
1633 record_excess_regs (XVECEXP (in_this, i, j), not_in_this, output);
1637 record_excess_regs (XEXP (in_this, i), not_in_this, output);
1643 /* Check what regs are referred to in the libcall block ending with INSN,
1644 aside from those mentioned in the equivalent value.
1645 If there are none, return 0.
1646 If there are one or more, return an EXPR_LIST containing all of them. */
1649 libcall_other_reg (rtx insn, rtx equiv)
1651 rtx note = find_reg_note (insn, REG_RETVAL, NULL_RTX);
1652 rtx p = XEXP (note, 0);
1655 /* First, find all the regs used in the libcall block
1656 that are not mentioned as inputs to the result. */
1661 record_excess_regs (PATTERN (p), equiv, &output);
1668 /* Return 1 if all uses of REG
1669 are between INSN and the end of the basic block. */
1672 reg_in_basic_block_p (rtx insn, rtx reg)
1674 int regno = REGNO (reg);
1677 if (REGNO_FIRST_UID (regno) != INSN_UID (insn))
1680 /* Search this basic block for the already recorded last use of the reg. */
1681 for (p = insn; p; p = NEXT_INSN (p))
1683 switch (GET_CODE (p))
1690 /* Ordinary insn: if this is the last use, we win. */
1691 if (REGNO_LAST_UID (regno) == INSN_UID (p))
1696 /* Jump insn: if this is the last use, we win. */
1697 if (REGNO_LAST_UID (regno) == INSN_UID (p))
1699 /* Otherwise, it's the end of the basic block, so we lose. */
1704 /* It's the end of the basic block, so we lose. */
1712 /* The "last use" that was recorded can't be found after the first
1713 use. This can happen when the last use was deleted while
1714 processing an inner loop, this inner loop was then completely
1715 unrolled, and the outer loop is always exited after the inner loop,
1716 so that everything after the first use becomes a single basic block. */
1720 /* Compute the benefit of eliminating the insns in the block whose
1721 last insn is LAST. This may be a group of insns used to compute a
1722 value directly or can contain a library call. */
1725 libcall_benefit (rtx last)
1730 for (insn = XEXP (find_reg_note (last, REG_RETVAL, NULL_RTX), 0);
1731 insn != last; insn = NEXT_INSN (insn))
1734 benefit += 10; /* Assume at least this many insns in a library
1736 else if (NONJUMP_INSN_P (insn)
1737 && GET_CODE (PATTERN (insn)) != USE
1738 && GET_CODE (PATTERN (insn)) != CLOBBER)
1745 /* Skip COUNT insns from INSN, counting library calls as 1 insn. */
1748 skip_consec_insns (rtx insn, int count)
1750 for (; count > 0; count--)
1754 /* If first insn of libcall sequence, skip to end. */
1755 /* Do this at start of loop, since INSN is guaranteed to
1758 && (temp = find_reg_note (insn, REG_LIBCALL, NULL_RTX)))
1759 insn = XEXP (temp, 0);
1762 insn = NEXT_INSN (insn);
1763 while (NOTE_P (insn));
1769 /* Ignore any movable whose insn falls within a libcall
1770 which is part of another movable.
1771 We make use of the fact that the movable for the libcall value
1772 was made later and so appears later on the chain. */
1775 ignore_some_movables (struct loop_movables *movables)
1777 struct movable *m, *m1;
1779 for (m = movables->head; m; m = m->next)
1781 /* Is this a movable for the value of a libcall? */
1782 rtx note = find_reg_note (m->insn, REG_RETVAL, NULL_RTX);
1786 /* Check for earlier movables inside that range,
1787 and mark them invalid. We cannot use LUIDs here because
1788 insns created by loop.c for prior loops don't have LUIDs.
1789 Rather than reject all such insns from movables, we just
1790 explicitly check each insn in the libcall (since invariant
1791 libcalls aren't that common). */
1792 for (insn = XEXP (note, 0); insn != m->insn; insn = NEXT_INSN (insn))
1793 for (m1 = movables->head; m1 != m; m1 = m1->next)
1794 if (m1->insn == insn)
1800 /* For each movable insn, see if the reg that it loads
1801 leads when it dies right into another conditionally movable insn.
1802 If so, record that the second insn "forces" the first one,
1803 since the second can be moved only if the first is. */
1806 force_movables (struct loop_movables *movables)
1808 struct movable *m, *m1;
1810 for (m1 = movables->head; m1; m1 = m1->next)
1811 /* Omit this if moving just the (SET (REG) 0) of a zero-extend. */
1812 if (!m1->partial && !m1->done)
1814 int regno = m1->regno;
1815 for (m = m1->next; m; m = m->next)
1816 /* ??? Could this be a bug? What if CSE caused the
1817 register of M1 to be used after this insn?
1818 Since CSE does not update regno_last_uid,
1819 this insn M->insn might not be where it dies.
1820 But very likely this doesn't matter; what matters is
1821 that M's reg is computed from M1's reg. */
1822 if (INSN_UID (m->insn) == REGNO_LAST_UID (regno)
1825 if (m != 0 && m->set_src == m1->set_dest
1826 /* If m->consec, m->set_src isn't valid. */
1830 /* Increase the priority of the moving the first insn
1831 since it permits the second to be moved as well.
1832 Likewise for insns already forced by the first insn. */
1838 for (m2 = m1; m2; m2 = m2->forces)
1840 m2->lifetime += m->lifetime;
1841 m2->savings += m->savings;
1847 /* Find invariant expressions that are equal and can be combined into
1851 combine_movables (struct loop_movables *movables, struct loop_regs *regs)
1854 char *matched_regs = xmalloc (regs->num);
1855 enum machine_mode mode;
1857 /* Regs that are set more than once are not allowed to match
1858 or be matched. I'm no longer sure why not. */
1859 /* Only pseudo registers are allowed to match or be matched,
1860 since move_movables does not validate the change. */
1861 /* Perhaps testing m->consec_sets would be more appropriate here? */
1863 for (m = movables->head; m; m = m->next)
1864 if (m->match == 0 && regs->array[m->regno].n_times_set == 1
1865 && m->regno >= FIRST_PSEUDO_REGISTER
1870 int regno = m->regno;
1872 memset (matched_regs, 0, regs->num);
1873 matched_regs[regno] = 1;
1875 /* We want later insns to match the first one. Don't make the first
1876 one match any later ones. So start this loop at m->next. */
1877 for (m1 = m->next; m1; m1 = m1->next)
1878 if (m != m1 && m1->match == 0
1880 && regs->array[m1->regno].n_times_set == 1
1881 && m1->regno >= FIRST_PSEUDO_REGISTER
1882 /* A reg used outside the loop mustn't be eliminated. */
1884 /* A reg used for zero-extending mustn't be eliminated. */
1886 && (matched_regs[m1->regno]
1888 (GET_MODE (m->set_dest) == GET_MODE (m1->set_dest)
1889 /* See if the source of M1 says it matches M. */
1890 && ((REG_P (m1->set_src)
1891 && matched_regs[REGNO (m1->set_src)])
1892 || rtx_equal_for_loop_p (m->set_src, m1->set_src,
1894 && ((m->dependencies == m1->dependencies)
1895 || rtx_equal_p (m->dependencies, m1->dependencies)))
1897 m->lifetime += m1->lifetime;
1898 m->savings += m1->savings;
1901 matched_regs[m1->regno] = 1;
1905 /* Now combine the regs used for zero-extension.
1906 This can be done for those not marked `global'
1907 provided their lives don't overlap. */
1909 for (mode = GET_CLASS_NARROWEST_MODE (MODE_INT); mode != VOIDmode;
1910 mode = GET_MODE_WIDER_MODE (mode))
1912 struct movable *m0 = 0;
1914 /* Combine all the registers for extension from mode MODE.
1915 Don't combine any that are used outside this loop. */
1916 for (m = movables->head; m; m = m->next)
1917 if (m->partial && ! m->global
1918 && mode == GET_MODE (SET_SRC (PATTERN (NEXT_INSN (m->insn)))))
1922 int first = REGNO_FIRST_LUID (m->regno);
1923 int last = REGNO_LAST_LUID (m->regno);
1927 /* First one: don't check for overlap, just record it. */
1932 /* Make sure they extend to the same mode.
1933 (Almost always true.) */
1934 if (GET_MODE (m->set_dest) != GET_MODE (m0->set_dest))
1937 /* We already have one: check for overlap with those
1938 already combined together. */
1939 for (m1 = movables->head; m1 != m; m1 = m1->next)
1940 if (m1 == m0 || (m1->partial && m1->match == m0))
1941 if (! (REGNO_FIRST_LUID (m1->regno) > last
1942 || REGNO_LAST_LUID (m1->regno) < first))
1945 /* No overlap: we can combine this with the others. */
1946 m0->lifetime += m->lifetime;
1947 m0->savings += m->savings;
1957 free (matched_regs);
1960 /* Returns the number of movable instructions in LOOP that were not
1961 moved outside the loop. */
1964 num_unmoved_movables (const struct loop *loop)
1969 for (m = LOOP_MOVABLES (loop)->head; m; m = m->next)
1977 /* Return 1 if regs X and Y will become the same if moved. */
1980 regs_match_p (rtx x, rtx y, struct loop_movables *movables)
1982 unsigned int xn = REGNO (x);
1983 unsigned int yn = REGNO (y);
1984 struct movable *mx, *my;
1986 for (mx = movables->head; mx; mx = mx->next)
1987 if (mx->regno == xn)
1990 for (my = movables->head; my; my = my->next)
1991 if (my->regno == yn)
1995 && ((mx->match == my->match && mx->match != 0)
1997 || mx == my->match));
2000 /* Return 1 if X and Y are identical-looking rtx's.
2001 This is the Lisp function EQUAL for rtx arguments.
2003 If two registers are matching movables or a movable register and an
2004 equivalent constant, consider them equal. */
2007 rtx_equal_for_loop_p (rtx x, rtx y, struct loop_movables *movables,
2008 struct loop_regs *regs)
2018 if (x == 0 || y == 0)
2021 code = GET_CODE (x);
2023 /* If we have a register and a constant, they may sometimes be
2025 if (REG_P (x) && regs->array[REGNO (x)].set_in_loop == -2
2028 for (m = movables->head; m; m = m->next)
2029 if (m->move_insn && m->regno == REGNO (x)
2030 && rtx_equal_p (m->set_src, y))
2033 else if (REG_P (y) && regs->array[REGNO (y)].set_in_loop == -2
2036 for (m = movables->head; m; m = m->next)
2037 if (m->move_insn && m->regno == REGNO (y)
2038 && rtx_equal_p (m->set_src, x))
2042 /* Otherwise, rtx's of different codes cannot be equal. */
2043 if (code != GET_CODE (y))
2046 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent.
2047 (REG:SI x) and (REG:HI x) are NOT equivalent. */
2049 if (GET_MODE (x) != GET_MODE (y))
2052 /* These types of rtx's can be compared nonrecursively. */
2062 return (REGNO (x) == REGNO (y) || regs_match_p (x, y, movables));
2065 return XEXP (x, 0) == XEXP (y, 0);
2067 return XSTR (x, 0) == XSTR (y, 0);
2073 /* Compare the elements. If any pair of corresponding elements
2074 fail to match, return 0 for the whole things. */
2076 fmt = GET_RTX_FORMAT (code);
2077 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2082 if (XWINT (x, i) != XWINT (y, i))
2087 if (XINT (x, i) != XINT (y, i))
2092 /* Two vectors must have the same length. */
2093 if (XVECLEN (x, i) != XVECLEN (y, i))
2096 /* And the corresponding elements must match. */
2097 for (j = 0; j < XVECLEN (x, i); j++)
2098 if (rtx_equal_for_loop_p (XVECEXP (x, i, j), XVECEXP (y, i, j),
2099 movables, regs) == 0)
2104 if (rtx_equal_for_loop_p (XEXP (x, i), XEXP (y, i), movables, regs)
2110 if (strcmp (XSTR (x, i), XSTR (y, i)))
2115 /* These are just backpointers, so they don't matter. */
2121 /* It is believed that rtx's at this level will never
2122 contain anything but integers and other rtx's,
2123 except for within LABEL_REFs and SYMBOL_REFs. */
2131 /* If X contains any LABEL_REF's, add REG_LABEL notes for them to all
2132 insns in INSNS which use the reference. LABEL_NUSES for CODE_LABEL
2133 references is incremented once for each added note. */
2136 add_label_notes (rtx x, rtx insns)
2138 enum rtx_code code = GET_CODE (x);
2143 if (code == LABEL_REF && !LABEL_REF_NONLOCAL_P (x))
2145 /* This code used to ignore labels that referred to dispatch tables to
2146 avoid flow generating (slightly) worse code.
2148 We no longer ignore such label references (see LABEL_REF handling in
2149 mark_jump_label for additional information). */
2150 for (insn = insns; insn; insn = NEXT_INSN (insn))
2151 if (reg_mentioned_p (XEXP (x, 0), insn))
2153 REG_NOTES (insn) = gen_rtx_INSN_LIST (REG_LABEL, XEXP (x, 0),
2155 if (LABEL_P (XEXP (x, 0)))
2156 LABEL_NUSES (XEXP (x, 0))++;
2160 fmt = GET_RTX_FORMAT (code);
2161 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2164 add_label_notes (XEXP (x, i), insns);
2165 else if (fmt[i] == 'E')
2166 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
2167 add_label_notes (XVECEXP (x, i, j), insns);
2171 /* Scan MOVABLES, and move the insns that deserve to be moved.
2172 If two matching movables are combined, replace one reg with the
2173 other throughout. */
2176 move_movables (struct loop *loop, struct loop_movables *movables,
2177 int threshold, int insn_count)
2179 struct loop_regs *regs = LOOP_REGS (loop);
2180 int nregs = regs->num;
2184 rtx loop_start = loop->start;
2185 rtx loop_end = loop->end;
2186 /* Map of pseudo-register replacements to handle combining
2187 when we move several insns that load the same value
2188 into different pseudo-registers. */
2189 rtx *reg_map = xcalloc (nregs, sizeof (rtx));
2190 char *already_moved = xcalloc (nregs, sizeof (char));
2192 for (m = movables->head; m; m = m->next)
2194 /* Describe this movable insn. */
2196 if (loop_dump_stream)
2198 fprintf (loop_dump_stream, "Insn %d: regno %d (life %d), ",
2199 INSN_UID (m->insn), m->regno, m->lifetime);
2201 fprintf (loop_dump_stream, "consec %d, ", m->consec);
2203 fprintf (loop_dump_stream, "cond ");
2205 fprintf (loop_dump_stream, "force ");
2207 fprintf (loop_dump_stream, "global ");
2209 fprintf (loop_dump_stream, "done ");
2211 fprintf (loop_dump_stream, "move-insn ");
2213 fprintf (loop_dump_stream, "matches %d ",
2214 INSN_UID (m->match->insn));
2216 fprintf (loop_dump_stream, "forces %d ",
2217 INSN_UID (m->forces->insn));
2220 /* Ignore the insn if it's already done (it matched something else).
2221 Otherwise, see if it is now safe to move. */
2225 || (1 == loop_invariant_p (loop, m->set_src)
2226 && (m->dependencies == 0
2227 || 1 == loop_invariant_p (loop, m->dependencies))
2229 || 1 == consec_sets_invariant_p (loop, m->set_dest,
2232 && (! m->forces || m->forces->done))
2236 int savings = m->savings;
2238 /* We have an insn that is safe to move.
2239 Compute its desirability. */
2244 if (loop_dump_stream)
2245 fprintf (loop_dump_stream, "savings %d ", savings);
2247 if (regs->array[regno].moved_once && loop_dump_stream)
2248 fprintf (loop_dump_stream, "halved since already moved ");
2250 /* An insn MUST be moved if we already moved something else
2251 which is safe only if this one is moved too: that is,
2252 if already_moved[REGNO] is nonzero. */
2254 /* An insn is desirable to move if the new lifetime of the
2255 register is no more than THRESHOLD times the old lifetime.
2256 If it's not desirable, it means the loop is so big
2257 that moving won't speed things up much,
2258 and it is liable to make register usage worse. */
2260 /* It is also desirable to move if it can be moved at no
2261 extra cost because something else was already moved. */
2263 if (already_moved[regno]
2264 || (threshold * savings * m->lifetime) >=
2265 (regs->array[regno].moved_once ? insn_count * 2 : insn_count)
2266 || (m->forces && m->forces->done
2267 && regs->array[m->forces->regno].n_times_set == 1))
2271 rtx first = NULL_RTX;
2272 rtx newreg = NULL_RTX;
2275 newreg = gen_reg_rtx (GET_MODE (m->set_dest));
2277 /* Now move the insns that set the reg. */
2279 if (m->partial && m->match)
2283 /* Find the end of this chain of matching regs.
2284 Thus, we load each reg in the chain from that one reg.
2285 And that reg is loaded with 0 directly,
2286 since it has ->match == 0. */
2287 for (m1 = m; m1->match; m1 = m1->match);
2288 newpat = gen_move_insn (SET_DEST (PATTERN (m->insn)),
2289 SET_DEST (PATTERN (m1->insn)));
2290 i1 = loop_insn_hoist (loop, newpat);
2292 /* Mark the moved, invariant reg as being allowed to
2293 share a hard reg with the other matching invariant. */
2294 REG_NOTES (i1) = REG_NOTES (m->insn);
2295 r1 = SET_DEST (PATTERN (m->insn));
2296 r2 = SET_DEST (PATTERN (m1->insn));
2298 = gen_rtx_EXPR_LIST (VOIDmode, r1,
2299 gen_rtx_EXPR_LIST (VOIDmode, r2,
2301 delete_insn (m->insn);
2306 if (loop_dump_stream)
2307 fprintf (loop_dump_stream, " moved to %d", INSN_UID (i1));
2309 /* If we are to re-generate the item being moved with a
2310 new move insn, first delete what we have and then emit
2311 the move insn before the loop. */
2312 else if (m->move_insn)
2316 for (count = m->consec; count >= 0; count--)
2320 /* If this is the first insn of a library
2321 call sequence, something is very
2323 gcc_assert (!find_reg_note
2324 (p, REG_LIBCALL, NULL_RTX));
2326 /* If this is the last insn of a libcall
2327 sequence, then delete every insn in the
2328 sequence except the last. The last insn
2329 is handled in the normal manner. */
2330 temp = find_reg_note (p, REG_RETVAL, NULL_RTX);
2334 temp = XEXP (temp, 0);
2336 temp = delete_insn (temp);
2341 p = delete_insn (p);
2343 /* simplify_giv_expr expects that it can walk the insns
2344 at m->insn forwards and see this old sequence we are
2345 tossing here. delete_insn does preserve the next
2346 pointers, but when we skip over a NOTE we must fix
2347 it up. Otherwise that code walks into the non-deleted
2349 while (p && NOTE_P (p))
2350 p = NEXT_INSN (temp) = NEXT_INSN (p);
2354 /* Replace the original insn with a move from
2355 our newly created temp. */
2357 emit_move_insn (m->set_dest, newreg);
2360 emit_insn_before (seq, p);
2365 emit_move_insn (m->insert_temp ? newreg : m->set_dest,
2370 add_label_notes (m->set_src, seq);
2372 i1 = loop_insn_hoist (loop, seq);
2373 if (! find_reg_note (i1, REG_EQUAL, NULL_RTX))
2374 set_unique_reg_note (i1,
2375 m->is_equiv ? REG_EQUIV : REG_EQUAL,
2378 if (loop_dump_stream)
2379 fprintf (loop_dump_stream, " moved to %d", INSN_UID (i1));
2381 /* The more regs we move, the less we like moving them. */
2386 for (count = m->consec; count >= 0; count--)
2390 /* If first insn of libcall sequence, skip to end. */
2391 /* Do this at start of loop, since p is guaranteed to
2394 && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
2397 /* If last insn of libcall sequence, move all
2398 insns except the last before the loop. The last
2399 insn is handled in the normal manner. */
2401 && (temp = find_reg_note (p, REG_RETVAL, NULL_RTX)))
2405 rtx fn_address_insn = 0;
2408 for (temp = XEXP (temp, 0); temp != p;
2409 temp = NEXT_INSN (temp))
2418 body = PATTERN (temp);
2420 /* Find the next insn after TEMP,
2421 not counting USE or NOTE insns. */
2422 for (next = NEXT_INSN (temp); next != p;
2423 next = NEXT_INSN (next))
2424 if (! (NONJUMP_INSN_P (next)
2425 && GET_CODE (PATTERN (next)) == USE)
2429 /* If that is the call, this may be the insn
2430 that loads the function address.
2432 Extract the function address from the insn
2433 that loads it into a register.
2434 If this insn was cse'd, we get incorrect code.
2436 So emit a new move insn that copies the
2437 function address into the register that the
2438 call insn will use. flow.c will delete any
2439 redundant stores that we have created. */
2441 && GET_CODE (body) == SET
2442 && REG_P (SET_DEST (body))
2443 && (n = find_reg_note (temp, REG_EQUAL,
2446 fn_reg = SET_SRC (body);
2447 if (!REG_P (fn_reg))
2448 fn_reg = SET_DEST (body);
2449 fn_address = XEXP (n, 0);
2450 fn_address_insn = temp;
2452 /* We have the call insn.
2453 If it uses the register we suspect it might,
2454 load it with the correct address directly. */
2457 && reg_referenced_p (fn_reg, body))
2458 loop_insn_emit_after (loop, 0, fn_address_insn,
2460 (fn_reg, fn_address));
2464 i1 = loop_call_insn_hoist (loop, body);
2465 /* Because the USAGE information potentially
2466 contains objects other than hard registers
2467 we need to copy it. */
2468 if (CALL_INSN_FUNCTION_USAGE (temp))
2469 CALL_INSN_FUNCTION_USAGE (i1)
2470 = copy_rtx (CALL_INSN_FUNCTION_USAGE (temp));
2473 i1 = loop_insn_hoist (loop, body);
2476 if (temp == fn_address_insn)
2477 fn_address_insn = i1;
2478 REG_NOTES (i1) = REG_NOTES (temp);
2479 REG_NOTES (temp) = NULL;
2485 if (m->savemode != VOIDmode)
2487 /* P sets REG to zero; but we should clear only
2488 the bits that are not covered by the mode
2490 rtx reg = m->set_dest;
2495 tem = expand_simple_binop
2496 (GET_MODE (reg), AND, reg,
2497 GEN_INT ((((HOST_WIDE_INT) 1
2498 << GET_MODE_BITSIZE (m->savemode)))
2500 reg, 1, OPTAB_LIB_WIDEN);
2503 emit_move_insn (reg, tem);
2504 sequence = get_insns ();
2506 i1 = loop_insn_hoist (loop, sequence);
2508 else if (CALL_P (p))
2510 i1 = loop_call_insn_hoist (loop, PATTERN (p));
2511 /* Because the USAGE information potentially
2512 contains objects other than hard registers
2513 we need to copy it. */
2514 if (CALL_INSN_FUNCTION_USAGE (p))
2515 CALL_INSN_FUNCTION_USAGE (i1)
2516 = copy_rtx (CALL_INSN_FUNCTION_USAGE (p));
2518 else if (count == m->consec && m->move_insn_first)
2521 /* The SET_SRC might not be invariant, so we must
2522 use the REG_EQUAL note. */
2524 emit_move_insn (m->insert_temp ? newreg : m->set_dest,
2529 add_label_notes (m->set_src, seq);
2531 i1 = loop_insn_hoist (loop, seq);
2532 if (! find_reg_note (i1, REG_EQUAL, NULL_RTX))
2533 set_unique_reg_note (i1, m->is_equiv ? REG_EQUIV
2534 : REG_EQUAL, m->set_src);
2536 else if (m->insert_temp)
2538 rtx *reg_map2 = xcalloc (REGNO (newreg),
2540 reg_map2 [m->regno] = newreg;
2542 i1 = loop_insn_hoist (loop, copy_rtx (PATTERN (p)));
2543 replace_regs (i1, reg_map2, REGNO (newreg), 1);
2547 i1 = loop_insn_hoist (loop, PATTERN (p));
2549 if (REG_NOTES (i1) == 0)
2551 REG_NOTES (i1) = REG_NOTES (p);
2552 REG_NOTES (p) = NULL;
2554 /* If there is a REG_EQUAL note present whose value
2555 is not loop invariant, then delete it, since it
2556 may cause problems with later optimization passes.
2557 It is possible for cse to create such notes
2558 like this as a result of record_jump_cond. */
2560 if ((temp = find_reg_note (i1, REG_EQUAL, NULL_RTX))
2561 && ! loop_invariant_p (loop, XEXP (temp, 0)))
2562 remove_note (i1, temp);
2568 if (loop_dump_stream)
2569 fprintf (loop_dump_stream, " moved to %d",
2572 /* If library call, now fix the REG_NOTES that contain
2573 insn pointers, namely REG_LIBCALL on FIRST
2574 and REG_RETVAL on I1. */
2575 if ((temp = find_reg_note (i1, REG_RETVAL, NULL_RTX)))
2577 XEXP (temp, 0) = first;
2578 temp = find_reg_note (first, REG_LIBCALL, NULL_RTX);
2579 XEXP (temp, 0) = i1;
2586 /* simplify_giv_expr expects that it can walk the insns
2587 at m->insn forwards and see this old sequence we are
2588 tossing here. delete_insn does preserve the next
2589 pointers, but when we skip over a NOTE we must fix
2590 it up. Otherwise that code walks into the non-deleted
2592 while (p && NOTE_P (p))
2593 p = NEXT_INSN (temp) = NEXT_INSN (p);
2598 /* Replace the original insn with a move from
2599 our newly created temp. */
2601 emit_move_insn (m->set_dest, newreg);
2604 emit_insn_before (seq, p);
2608 /* The more regs we move, the less we like moving them. */
2614 if (!m->insert_temp)
2616 /* Any other movable that loads the same register
2618 already_moved[regno] = 1;
2620 /* This reg has been moved out of one loop. */
2621 regs->array[regno].moved_once = 1;
2623 /* The reg set here is now invariant. */
2627 for (i = 0; i < LOOP_REGNO_NREGS (regno, m->set_dest); i++)
2628 regs->array[regno+i].set_in_loop = 0;
2631 /* Change the length-of-life info for the register
2632 to say it lives at least the full length of this loop.
2633 This will help guide optimizations in outer loops. */
2635 if (REGNO_FIRST_LUID (regno) > INSN_LUID (loop_start))
2636 /* This is the old insn before all the moved insns.
2637 We can't use the moved insn because it is out of range
2638 in uid_luid. Only the old insns have luids. */
2639 REGNO_FIRST_UID (regno) = INSN_UID (loop_start);
2640 if (REGNO_LAST_LUID (regno) < INSN_LUID (loop_end))
2641 REGNO_LAST_UID (regno) = INSN_UID (loop_end);
2644 /* Combine with this moved insn any other matching movables. */
2647 for (m1 = movables->head; m1; m1 = m1->next)
2652 reg_map[m1->regno] = m->set_dest;
2654 /* Get rid of the matching insn
2655 and prevent further processing of it. */
2658 /* If library call, delete all insns. */
2659 if ((temp = find_reg_note (m1->insn, REG_RETVAL,
2661 delete_insn_chain (XEXP (temp, 0), m1->insn);
2663 delete_insn (m1->insn);
2665 /* Any other movable that loads the same register
2667 already_moved[m1->regno] = 1;
2669 /* The reg merged here is now invariant,
2670 if the reg it matches is invariant. */
2675 i < LOOP_REGNO_NREGS (regno, m1->set_dest);
2677 regs->array[m1->regno+i].set_in_loop = 0;
2681 else if (loop_dump_stream)
2682 fprintf (loop_dump_stream, "not desirable");
2684 else if (loop_dump_stream && !m->match)
2685 fprintf (loop_dump_stream, "not safe");
2687 if (loop_dump_stream)
2688 fprintf (loop_dump_stream, "\n");
2692 new_start = loop_start;
2694 /* Go through all the instructions in the loop, making
2695 all the register substitutions scheduled in REG_MAP. */
2696 for (p = new_start; p != loop_end; p = NEXT_INSN (p))
2699 replace_regs (PATTERN (p), reg_map, nregs, 0);
2700 replace_regs (REG_NOTES (p), reg_map, nregs, 0);
2706 free (already_moved);
2711 loop_movables_add (struct loop_movables *movables, struct movable *m)
2713 if (movables->head == 0)
2716 movables->last->next = m;
2722 loop_movables_free (struct loop_movables *movables)
2725 struct movable *m_next;
2727 for (m = movables->head; m; m = m_next)
2735 /* Scan X and replace the address of any MEM in it with ADDR.
2736 REG is the address that MEM should have before the replacement. */
2739 replace_call_address (rtx x, rtx reg, rtx addr)
2747 code = GET_CODE (x);
2761 /* Short cut for very common case. */
2762 replace_call_address (XEXP (x, 1), reg, addr);
2766 /* Short cut for very common case. */
2767 replace_call_address (XEXP (x, 0), reg, addr);
2771 /* If this MEM uses a reg other than the one we expected,
2772 something is wrong. */
2773 gcc_assert (XEXP (x, 0) == reg);
2781 fmt = GET_RTX_FORMAT (code);
2782 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2785 replace_call_address (XEXP (x, i), reg, addr);
2786 else if (fmt[i] == 'E')
2789 for (j = 0; j < XVECLEN (x, i); j++)
2790 replace_call_address (XVECEXP (x, i, j), reg, addr);
2796 /* Return the number of memory refs to addresses that vary
2800 count_nonfixed_reads (const struct loop *loop, rtx x)
2810 code = GET_CODE (x);
2824 return ((loop_invariant_p (loop, XEXP (x, 0)) != 1)
2825 + count_nonfixed_reads (loop, XEXP (x, 0)));
2832 fmt = GET_RTX_FORMAT (code);
2833 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2836 value += count_nonfixed_reads (loop, XEXP (x, i));
2840 for (j = 0; j < XVECLEN (x, i); j++)
2841 value += count_nonfixed_reads (loop, XVECEXP (x, i, j));
2847 /* Scan a loop setting the elements `loops_enclosed',
2848 `has_call', `has_nonconst_call', `has_volatile', `has_tablejump',
2849 `unknown_address_altered', `unknown_constant_address_altered', and
2850 `num_mem_sets' in LOOP. Also, fill in the array `mems' and the
2851 list `store_mems' in LOOP. */
2854 prescan_loop (struct loop *loop)
2858 struct loop_info *loop_info = LOOP_INFO (loop);
2859 rtx start = loop->start;
2860 rtx end = loop->end;
2861 /* The label after END. Jumping here is just like falling off the
2862 end of the loop. We use next_nonnote_insn instead of next_label
2863 as a hedge against the (pathological) case where some actual insn
2864 might end up between the two. */
2865 rtx exit_target = next_nonnote_insn (end);
2867 loop_info->has_indirect_jump = indirect_jump_in_function;
2868 loop_info->pre_header_has_call = 0;
2869 loop_info->has_call = 0;
2870 loop_info->has_nonconst_call = 0;
2871 loop_info->has_prefetch = 0;
2872 loop_info->has_volatile = 0;
2873 loop_info->has_tablejump = 0;
2874 loop_info->has_multiple_exit_targets = 0;
2877 loop_info->unknown_address_altered = 0;
2878 loop_info->unknown_constant_address_altered = 0;
2879 loop_info->store_mems = NULL_RTX;
2880 loop_info->first_loop_store_insn = NULL_RTX;
2881 loop_info->mems_idx = 0;
2882 loop_info->num_mem_sets = 0;
2884 for (insn = start; insn && !LABEL_P (insn);
2885 insn = PREV_INSN (insn))
2889 loop_info->pre_header_has_call = 1;
2894 for (insn = NEXT_INSN (start); insn != NEXT_INSN (end);
2895 insn = NEXT_INSN (insn))
2897 switch (GET_CODE (insn))
2900 if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG)
2903 /* Count number of loops contained in this one. */
2906 else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_END)
2911 if (! CONST_OR_PURE_CALL_P (insn))
2913 loop_info->unknown_address_altered = 1;
2914 loop_info->has_nonconst_call = 1;
2916 else if (pure_call_p (insn))
2917 loop_info->has_nonconst_call = 1;
2918 loop_info->has_call = 1;
2919 if (can_throw_internal (insn))
2920 loop_info->has_multiple_exit_targets = 1;
2924 if (! loop_info->has_multiple_exit_targets)
2926 rtx set = pc_set (insn);
2930 rtx src = SET_SRC (set);
2933 if (GET_CODE (src) == IF_THEN_ELSE)
2935 label1 = XEXP (src, 1);
2936 label2 = XEXP (src, 2);
2946 if (label1 && label1 != pc_rtx)
2948 if (GET_CODE (label1) != LABEL_REF)
2950 /* Something tricky. */
2951 loop_info->has_multiple_exit_targets = 1;
2954 else if (XEXP (label1, 0) != exit_target
2955 && LABEL_OUTSIDE_LOOP_P (label1))
2957 /* A jump outside the current loop. */
2958 loop_info->has_multiple_exit_targets = 1;
2970 /* A return, or something tricky. */
2971 loop_info->has_multiple_exit_targets = 1;
2977 if (volatile_refs_p (PATTERN (insn)))
2978 loop_info->has_volatile = 1;
2981 && (GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC
2982 || GET_CODE (PATTERN (insn)) == ADDR_VEC))
2983 loop_info->has_tablejump = 1;
2985 note_stores (PATTERN (insn), note_addr_stored, loop_info);
2986 if (! loop_info->first_loop_store_insn && loop_info->store_mems)
2987 loop_info->first_loop_store_insn = insn;
2989 if (flag_non_call_exceptions && can_throw_internal (insn))
2990 loop_info->has_multiple_exit_targets = 1;
2998 /* Now, rescan the loop, setting up the LOOP_MEMS array. */
2999 if (/* An exception thrown by a called function might land us
3001 ! loop_info->has_nonconst_call
3002 /* We don't want loads for MEMs moved to a location before the
3003 one at which their stack memory becomes allocated. (Note
3004 that this is not a problem for malloc, etc., since those
3005 require actual function calls. */
3006 && ! current_function_calls_alloca
3007 /* There are ways to leave the loop other than falling off the
3009 && ! loop_info->has_multiple_exit_targets)
3010 for (insn = NEXT_INSN (start); insn != NEXT_INSN (end);
3011 insn = NEXT_INSN (insn))
3012 for_each_rtx (&insn, insert_loop_mem, loop_info);
3014 /* BLKmode MEMs are added to LOOP_STORE_MEM as necessary so
3015 that loop_invariant_p and load_mems can use true_dependence
3016 to determine what is really clobbered. */
3017 if (loop_info->unknown_address_altered)
3019 rtx mem = gen_rtx_MEM (BLKmode, const0_rtx);
3021 loop_info->store_mems
3022 = gen_rtx_EXPR_LIST (VOIDmode, mem, loop_info->store_mems);
3024 if (loop_info->unknown_constant_address_altered)
3026 rtx mem = gen_rtx_MEM (BLKmode, const0_rtx);
3027 MEM_READONLY_P (mem) = 1;
3028 loop_info->store_mems
3029 = gen_rtx_EXPR_LIST (VOIDmode, mem, loop_info->store_mems);
3033 /* Invalidate all loops containing LABEL. */
3036 invalidate_loops_containing_label (rtx label)
3039 for (loop = uid_loop[INSN_UID (label)]; loop; loop = loop->outer)
3043 /* Scan the function looking for loops. Record the start and end of each loop.
3044 Also mark as invalid loops any loops that contain a setjmp or are branched
3045 to from outside the loop. */
3048 find_and_verify_loops (rtx f, struct loops *loops)
3053 struct loop *current_loop;
3054 struct loop *next_loop;
3057 num_loops = loops->num;
3059 compute_luids (f, NULL_RTX, 0);
3061 /* If there are jumps to undefined labels,
3062 treat them as jumps out of any/all loops.
3063 This also avoids writing past end of tables when there are no loops. */
3066 /* Find boundaries of loops, mark which loops are contained within
3067 loops, and invalidate loops that have setjmp. */
3070 current_loop = NULL;
3071 for (insn = f; insn; insn = NEXT_INSN (insn))
3074 switch (NOTE_LINE_NUMBER (insn))
3076 case NOTE_INSN_LOOP_BEG:
3077 next_loop = loops->array + num_loops;
3078 next_loop->num = num_loops;
3080 next_loop->start = insn;
3081 next_loop->outer = current_loop;
3082 current_loop = next_loop;
3085 case NOTE_INSN_LOOP_END:
3086 gcc_assert (current_loop);
3088 current_loop->end = insn;
3089 current_loop = current_loop->outer;
3097 && find_reg_note (insn, REG_SETJMP, NULL))
3099 /* In this case, we must invalidate our current loop and any
3101 for (loop = current_loop; loop; loop = loop->outer)
3104 if (loop_dump_stream)
3105 fprintf (loop_dump_stream,
3106 "\nLoop at %d ignored due to setjmp.\n",
3107 INSN_UID (loop->start));
3111 /* Note that this will mark the NOTE_INSN_LOOP_END note as being in the
3112 enclosing loop, but this doesn't matter. */
3113 uid_loop[INSN_UID (insn)] = current_loop;
3116 /* Any loop containing a label used in an initializer must be invalidated,
3117 because it can be jumped into from anywhere. */
3118 for (label = forced_labels; label; label = XEXP (label, 1))
3119 invalidate_loops_containing_label (XEXP (label, 0));
3121 /* Any loop containing a label used for an exception handler must be
3122 invalidated, because it can be jumped into from anywhere. */
3123 for_each_eh_label (invalidate_loops_containing_label);
3125 /* Now scan all insn's in the function. If any JUMP_INSN branches into a
3126 loop that it is not contained within, that loop is marked invalid.
3127 If any INSN or CALL_INSN uses a label's address, then the loop containing
3128 that label is marked invalid, because it could be jumped into from
3131 Also look for blocks of code ending in an unconditional branch that
3132 exits the loop. If such a block is surrounded by a conditional
3133 branch around the block, move the block elsewhere (see below) and
3134 invert the jump to point to the code block. This may eliminate a
3135 label in our loop and will simplify processing by both us and a
3136 possible second cse pass. */
3138 for (insn = f; insn; insn = NEXT_INSN (insn))
3141 struct loop *this_loop = uid_loop[INSN_UID (insn)];
3143 if (NONJUMP_INSN_P (insn) || CALL_P (insn))
3145 rtx note = find_reg_note (insn, REG_LABEL, NULL_RTX);
3147 invalidate_loops_containing_label (XEXP (note, 0));
3153 mark_loop_jump (PATTERN (insn), this_loop);
3155 /* See if this is an unconditional branch outside the loop. */
3157 && (GET_CODE (PATTERN (insn)) == RETURN
3158 || (any_uncondjump_p (insn)
3159 && onlyjump_p (insn)
3160 && (uid_loop[INSN_UID (JUMP_LABEL (insn))]
3162 && get_max_uid () < max_uid_for_loop)
3165 rtx our_next = next_real_insn (insn);
3166 rtx last_insn_to_move = NEXT_INSN (insn);
3167 struct loop *dest_loop;
3168 struct loop *outer_loop = NULL;
3170 /* Go backwards until we reach the start of the loop, a label,
3172 for (p = PREV_INSN (insn);
3175 && NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
3180 /* Check for the case where we have a jump to an inner nested
3181 loop, and do not perform the optimization in that case. */
3183 if (JUMP_LABEL (insn))
3185 dest_loop = uid_loop[INSN_UID (JUMP_LABEL (insn))];
3188 for (outer_loop = dest_loop; outer_loop;
3189 outer_loop = outer_loop->outer)
3190 if (outer_loop == this_loop)
3195 /* Make sure that the target of P is within the current loop. */
3197 if (JUMP_P (p) && JUMP_LABEL (p)
3198 && uid_loop[INSN_UID (JUMP_LABEL (p))] != this_loop)
3199 outer_loop = this_loop;
3201 /* If we stopped on a JUMP_INSN to the next insn after INSN,
3202 we have a block of code to try to move.
3204 We look backward and then forward from the target of INSN
3205 to find a BARRIER at the same loop depth as the target.
3206 If we find such a BARRIER, we make a new label for the start
3207 of the block, invert the jump in P and point it to that label,
3208 and move the block of code to the spot we found. */
3212 && JUMP_LABEL (p) != 0
3213 /* Just ignore jumps to labels that were never emitted.
3214 These always indicate compilation errors. */
3215 && INSN_UID (JUMP_LABEL (p)) != 0
3216 && any_condjump_p (p) && onlyjump_p (p)
3217 && next_real_insn (JUMP_LABEL (p)) == our_next
3218 /* If it's not safe to move the sequence, then we
3220 && insns_safe_to_move_p (p, NEXT_INSN (insn),
3221 &last_insn_to_move))
3224 = JUMP_LABEL (insn) ? JUMP_LABEL (insn) : get_last_insn ();
3225 struct loop *target_loop = uid_loop[INSN_UID (target)];
3229 /* Search for possible garbage past the conditional jumps
3230 and look for the last barrier. */
3231 for (tmp = last_insn_to_move;
3232 tmp && !LABEL_P (tmp); tmp = NEXT_INSN (tmp))
3233 if (BARRIER_P (tmp))
3234 last_insn_to_move = tmp;
3236 for (loc = target; loc; loc = PREV_INSN (loc))
3238 /* Don't move things inside a tablejump. */
3239 && ((loc2 = next_nonnote_insn (loc)) == 0
3241 || (loc2 = next_nonnote_insn (loc2)) == 0
3243 || (GET_CODE (PATTERN (loc2)) != ADDR_VEC
3244 && GET_CODE (PATTERN (loc2)) != ADDR_DIFF_VEC))
3245 && uid_loop[INSN_UID (loc)] == target_loop)
3249 for (loc = target; loc; loc = NEXT_INSN (loc))
3251 /* Don't move things inside a tablejump. */
3252 && ((loc2 = next_nonnote_insn (loc)) == 0
3254 || (loc2 = next_nonnote_insn (loc2)) == 0
3256 || (GET_CODE (PATTERN (loc2)) != ADDR_VEC
3257 && GET_CODE (PATTERN (loc2)) != ADDR_DIFF_VEC))
3258 && uid_loop[INSN_UID (loc)] == target_loop)
3263 rtx cond_label = JUMP_LABEL (p);
3264 rtx new_label = get_label_after (p);
3266 /* Ensure our label doesn't go away. */
3267 LABEL_NUSES (cond_label)++;
3269 /* Verify that uid_loop is large enough and that
3271 if (invert_jump (p, new_label, 1))
3276 /* If no suitable BARRIER was found, create a suitable
3277 one before TARGET. Since TARGET is a fall through
3278 path, we'll need to insert a jump around our block
3279 and add a BARRIER before TARGET.
3281 This creates an extra unconditional jump outside
3282 the loop. However, the benefits of removing rarely
3283 executed instructions from inside the loop usually
3284 outweighs the cost of the extra unconditional jump
3285 outside the loop. */
3290 temp = gen_jump (JUMP_LABEL (insn));
3291 temp = emit_jump_insn_before (temp, target);
3292 JUMP_LABEL (temp) = JUMP_LABEL (insn);
3293 LABEL_NUSES (JUMP_LABEL (insn))++;
3294 loc = emit_barrier_before (target);
3297 /* Include the BARRIER after INSN and copy the
3299 only_notes = squeeze_notes (&new_label,
3300 &last_insn_to_move);
3301 gcc_assert (!only_notes);
3303 reorder_insns (new_label, last_insn_to_move, loc);
3305 /* All those insns are now in TARGET_LOOP. */
3307 q != NEXT_INSN (last_insn_to_move);
3309 uid_loop[INSN_UID (q)] = target_loop;
3311 /* The label jumped to by INSN is no longer a loop
3312 exit. Unless INSN does not have a label (e.g.,
3313 it is a RETURN insn), search loop->exit_labels
3314 to find its label_ref, and remove it. Also turn
3315 off LABEL_OUTSIDE_LOOP_P bit. */
3316 if (JUMP_LABEL (insn))
3318 for (q = 0, r = this_loop->exit_labels;
3320 q = r, r = LABEL_NEXTREF (r))
3321 if (XEXP (r, 0) == JUMP_LABEL (insn))
3323 LABEL_OUTSIDE_LOOP_P (r) = 0;
3325 LABEL_NEXTREF (q) = LABEL_NEXTREF (r);
3327 this_loop->exit_labels = LABEL_NEXTREF (r);
3331 for (loop = this_loop; loop && loop != target_loop;
3335 /* If we didn't find it, then something is
3340 /* P is now a jump outside the loop, so it must be put
3341 in loop->exit_labels, and marked as such.
3342 The easiest way to do this is to just call
3343 mark_loop_jump again for P. */
3344 mark_loop_jump (PATTERN (p), this_loop);
3346 /* If INSN now jumps to the insn after it,
3348 if (JUMP_LABEL (insn) != 0
3349 && (next_real_insn (JUMP_LABEL (insn))
3350 == next_real_insn (insn)))
3351 delete_related_insns (insn);
3354 /* Continue the loop after where the conditional
3355 branch used to jump, since the only branch insn
3356 in the block (if it still remains) is an inter-loop
3357 branch and hence needs no processing. */
3358 insn = NEXT_INSN (cond_label);
3360 if (--LABEL_NUSES (cond_label) == 0)
3361 delete_related_insns (cond_label);
3363 /* This loop will be continued with NEXT_INSN (insn). */
3364 insn = PREV_INSN (insn);
3371 /* If any label in X jumps to a loop different from LOOP_NUM and any of the
3372 loops it is contained in, mark the target loop invalid.
3374 For speed, we assume that X is part of a pattern of a JUMP_INSN. */
3377 mark_loop_jump (rtx x, struct loop *loop)
3379 struct loop *dest_loop;
3380 struct loop *outer_loop;
3383 switch (GET_CODE (x))
3396 /* There could be a label reference in here. */
3397 mark_loop_jump (XEXP (x, 0), loop);
3403 mark_loop_jump (XEXP (x, 0), loop);
3404 mark_loop_jump (XEXP (x, 1), loop);
3408 /* This may refer to a LABEL_REF or SYMBOL_REF. */
3409 mark_loop_jump (XEXP (x, 1), loop);
3414 mark_loop_jump (XEXP (x, 0), loop);
3418 dest_loop = uid_loop[INSN_UID (XEXP (x, 0))];
3420 /* Link together all labels that branch outside the loop. This
3421 is used by final_[bg]iv_value and the loop unrolling code. Also
3422 mark this LABEL_REF so we know that this branch should predict
3425 /* A check to make sure the label is not in an inner nested loop,
3426 since this does not count as a loop exit. */
3429 for (outer_loop = dest_loop; outer_loop;
3430 outer_loop = outer_loop->outer)
3431 if (outer_loop == loop)
3437 if (loop && ! outer_loop)
3439 LABEL_OUTSIDE_LOOP_P (x) = 1;
3440 LABEL_NEXTREF (x) = loop->exit_labels;
3441 loop->exit_labels = x;
3443 for (outer_loop = loop;
3444 outer_loop && outer_loop != dest_loop;
3445 outer_loop = outer_loop->outer)
3446 outer_loop->exit_count++;
3449 /* If this is inside a loop, but not in the current loop or one enclosed
3450 by it, it invalidates at least one loop. */
3455 /* We must invalidate every nested loop containing the target of this
3456 label, except those that also contain the jump insn. */
3458 for (; dest_loop; dest_loop = dest_loop->outer)
3460 /* Stop when we reach a loop that also contains the jump insn. */
3461 for (outer_loop = loop; outer_loop; outer_loop = outer_loop->outer)
3462 if (dest_loop == outer_loop)
3465 /* If we get here, we know we need to invalidate a loop. */
3466 if (loop_dump_stream && ! dest_loop->invalid)
3467 fprintf (loop_dump_stream,
3468 "\nLoop at %d ignored due to multiple entry points.\n",
3469 INSN_UID (dest_loop->start));
3471 dest_loop->invalid = 1;
3476 /* If this is not setting pc, ignore. */
3477 if (SET_DEST (x) == pc_rtx)
3478 mark_loop_jump (SET_SRC (x), loop);
3482 mark_loop_jump (XEXP (x, 1), loop);
3483 mark_loop_jump (XEXP (x, 2), loop);
3488 for (i = 0; i < XVECLEN (x, 0); i++)
3489 mark_loop_jump (XVECEXP (x, 0, i), loop);
3493 for (i = 0; i < XVECLEN (x, 1); i++)
3494 mark_loop_jump (XVECEXP (x, 1, i), loop);
3498 /* Strictly speaking this is not a jump into the loop, only a possible
3499 jump out of the loop. However, we have no way to link the destination
3500 of this jump onto the list of exit labels. To be safe we mark this
3501 loop and any containing loops as invalid. */
3504 for (outer_loop = loop; outer_loop; outer_loop = outer_loop->outer)
3506 if (loop_dump_stream && ! outer_loop->invalid)
3507 fprintf (loop_dump_stream,
3508 "\nLoop at %d ignored due to unknown exit jump.\n",
3509 INSN_UID (outer_loop->start));
3510 outer_loop->invalid = 1;
3517 /* Return nonzero if there is a label in the range from
3518 insn INSN to and including the insn whose luid is END
3519 INSN must have an assigned luid (i.e., it must not have
3520 been previously created by loop.c). */
3523 labels_in_range_p (rtx insn, int end)
3525 while (insn && INSN_LUID (insn) <= end)
3529 insn = NEXT_INSN (insn);
3535 /* Record that a memory reference X is being set. */
3538 note_addr_stored (rtx x, rtx y ATTRIBUTE_UNUSED,
3539 void *data ATTRIBUTE_UNUSED)
3541 struct loop_info *loop_info = data;
3543 if (x == 0 || !MEM_P (x))
3546 /* Count number of memory writes.
3547 This affects heuristics in strength_reduce. */
3548 loop_info->num_mem_sets++;
3550 /* BLKmode MEM means all memory is clobbered. */
3551 if (GET_MODE (x) == BLKmode)
3553 if (MEM_READONLY_P (x))
3554 loop_info->unknown_constant_address_altered = 1;
3556 loop_info->unknown_address_altered = 1;
3561 loop_info->store_mems = gen_rtx_EXPR_LIST (VOIDmode, x,
3562 loop_info->store_mems);
3565 /* X is a value modified by an INSN that references a biv inside a loop
3566 exit test (i.e., X is somehow related to the value of the biv). If X
3567 is a pseudo that is used more than once, then the biv is (effectively)
3568 used more than once. DATA is a pointer to a loop_regs structure. */
3571 note_set_pseudo_multiple_uses (rtx x, rtx y ATTRIBUTE_UNUSED, void *data)
3573 struct loop_regs *regs = (struct loop_regs *) data;
3578 while (GET_CODE (x) == STRICT_LOW_PART
3579 || GET_CODE (x) == SIGN_EXTRACT
3580 || GET_CODE (x) == ZERO_EXTRACT
3581 || GET_CODE (x) == SUBREG)
3584 if (!REG_P (x) || REGNO (x) < FIRST_PSEUDO_REGISTER)
3587 /* If we do not have usage information, or if we know the register
3588 is used more than once, note that fact for check_dbra_loop. */
3589 if (REGNO (x) >= max_reg_before_loop
3590 || ! regs->array[REGNO (x)].single_usage
3591 || regs->array[REGNO (x)].single_usage == const0_rtx)
3592 regs->multiple_uses = 1;
3595 /* Return nonzero if the rtx X is invariant over the current loop.
3597 The value is 2 if we refer to something only conditionally invariant.
3599 A memory ref is invariant if it is not volatile and does not conflict
3600 with anything stored in `loop_info->store_mems'. */
3603 loop_invariant_p (const struct loop *loop, rtx x)
3605 struct loop_info *loop_info = LOOP_INFO (loop);
3606 struct loop_regs *regs = LOOP_REGS (loop);
3610 int conditional = 0;
3615 code = GET_CODE (x);
3629 case UNSPEC_VOLATILE:
3633 if ((x == frame_pointer_rtx || x == hard_frame_pointer_rtx
3634 || x == arg_pointer_rtx || x == pic_offset_table_rtx)
3635 && ! current_function_has_nonlocal_goto)
3638 if (LOOP_INFO (loop)->has_call
3639 && REGNO (x) < FIRST_PSEUDO_REGISTER && call_used_regs[REGNO (x)])
3642 /* Out-of-range regs can occur when we are called from unrolling.
3643 These registers created by the unroller are set in the loop,
3644 hence are never invariant.
3645 Other out-of-range regs can be generated by load_mems; those that
3646 are written to in the loop are not invariant, while those that are
3647 not written to are invariant. It would be easy for load_mems
3648 to set n_times_set correctly for these registers, however, there
3649 is no easy way to distinguish them from registers created by the
3652 if (REGNO (x) >= (unsigned) regs->num)
3655 if (regs->array[REGNO (x)].set_in_loop < 0)
3658 return regs->array[REGNO (x)].set_in_loop == 0;
3661 /* Volatile memory references must be rejected. Do this before
3662 checking for read-only items, so that volatile read-only items
3663 will be rejected also. */
3664 if (MEM_VOLATILE_P (x))
3667 /* See if there is any dependence between a store and this load. */
3668 mem_list_entry = loop_info->store_mems;
3669 while (mem_list_entry)
3671 if (true_dependence (XEXP (mem_list_entry, 0), VOIDmode,
3675 mem_list_entry = XEXP (mem_list_entry, 1);
3678 /* It's not invalidated by a store in memory
3679 but we must still verify the address is invariant. */
3683 /* Don't mess with insns declared volatile. */
3684 if (MEM_VOLATILE_P (x))
3692 fmt = GET_RTX_FORMAT (code);
3693 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3697 int tem = loop_invariant_p (loop, XEXP (x, i));
3703 else if (fmt[i] == 'E')
3706 for (j = 0; j < XVECLEN (x, i); j++)
3708 int tem = loop_invariant_p (loop, XVECEXP (x, i, j));
3718 return 1 + conditional;
3721 /* Return nonzero if all the insns in the loop that set REG
3722 are INSN and the immediately following insns,
3723 and if each of those insns sets REG in an invariant way
3724 (not counting uses of REG in them).
3726 The value is 2 if some of these insns are only conditionally invariant.
3728 We assume that INSN itself is the first set of REG
3729 and that its source is invariant. */
3732 consec_sets_invariant_p (const struct loop *loop, rtx reg, int n_sets,
3735 struct loop_regs *regs = LOOP_REGS (loop);
3737 unsigned int regno = REGNO (reg);
3739 /* Number of sets we have to insist on finding after INSN. */
3740 int count = n_sets - 1;
3741 int old = regs->array[regno].set_in_loop;
3745 /* If N_SETS hit the limit, we can't rely on its value. */
3749 regs->array[regno].set_in_loop = 0;
3757 code = GET_CODE (p);
3759 /* If library call, skip to end of it. */
3760 if (code == INSN && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
3765 && (set = single_set (p))
3766 && REG_P (SET_DEST (set))
3767 && REGNO (SET_DEST (set)) == regno)
3769 this = loop_invariant_p (loop, SET_SRC (set));
3772 else if ((temp = find_reg_note (p, REG_EQUAL, NULL_RTX)))
3774 /* If this is a libcall, then any invariant REG_EQUAL note is OK.
3775 If this is an ordinary insn, then only CONSTANT_P REG_EQUAL
3777 this = (CONSTANT_P (XEXP (temp, 0))
3778 || (find_reg_note (p, REG_RETVAL, NULL_RTX)
3779 && loop_invariant_p (loop, XEXP (temp, 0))));
3786 else if (code != NOTE)
3788 regs->array[regno].set_in_loop = old;
3793 regs->array[regno].set_in_loop = old;
3794 /* If loop_invariant_p ever returned 2, we return 2. */
3795 return 1 + (value & 2);
3798 /* Look at all uses (not sets) of registers in X. For each, if it is
3799 the single use, set USAGE[REGNO] to INSN; if there was a previous use in
3800 a different insn, set USAGE[REGNO] to const0_rtx. */
3803 find_single_use_in_loop (struct loop_regs *regs, rtx insn, rtx x)
3805 enum rtx_code code = GET_CODE (x);
3806 const char *fmt = GET_RTX_FORMAT (code);
3810 regs->array[REGNO (x)].single_usage
3811 = (regs->array[REGNO (x)].single_usage != 0
3812 && regs->array[REGNO (x)].single_usage != insn)
3813 ? const0_rtx : insn;
3815 else if (code == SET)
3817 /* Don't count SET_DEST if it is a REG; otherwise count things
3818 in SET_DEST because if a register is partially modified, it won't
3819 show up as a potential movable so we don't care how USAGE is set
3821 if (!REG_P (SET_DEST (x)))
3822 find_single_use_in_loop (regs, insn, SET_DEST (x));
3823 find_single_use_in_loop (regs, insn, SET_SRC (x));
3826 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3828 if (fmt[i] == 'e' && XEXP (x, i) != 0)
3829 find_single_use_in_loop (regs, insn, XEXP (x, i));
3830 else if (fmt[i] == 'E')
3831 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
3832 find_single_use_in_loop (regs, insn, XVECEXP (x, i, j));
3836 /* Count and record any set in X which is contained in INSN. Update
3837 REGS->array[I].MAY_NOT_OPTIMIZE and LAST_SET for any register I set
3841 count_one_set (struct loop_regs *regs, rtx insn, rtx x, rtx *last_set)
3843 if (GET_CODE (x) == CLOBBER && REG_P (XEXP (x, 0)))
3844 /* Don't move a reg that has an explicit clobber.
3845 It's not worth the pain to try to do it correctly. */
3846 regs->array[REGNO (XEXP (x, 0))].may_not_optimize = 1;
3848 if (GET_CODE (x) == SET || GET_CODE (x) == CLOBBER)
3850 rtx dest = SET_DEST (x);
3851 while (GET_CODE (dest) == SUBREG
3852 || GET_CODE (dest) == ZERO_EXTRACT
3853 || GET_CODE (dest) == STRICT_LOW_PART)
3854 dest = XEXP (dest, 0);
3858 int regno = REGNO (dest);
3859 for (i = 0; i < LOOP_REGNO_NREGS (regno, dest); i++)
3861 /* If this is the first setting of this reg
3862 in current basic block, and it was set before,
3863 it must be set in two basic blocks, so it cannot
3864 be moved out of the loop. */
3865 if (regs->array[regno].set_in_loop > 0
3866 && last_set[regno] == 0)
3867 regs->array[regno+i].may_not_optimize = 1;
3868 /* If this is not first setting in current basic block,
3869 see if reg was used in between previous one and this.
3870 If so, neither one can be moved. */
3871 if (last_set[regno] != 0
3872 && reg_used_between_p (dest, last_set[regno], insn))
3873 regs->array[regno+i].may_not_optimize = 1;
3874 if (regs->array[regno+i].set_in_loop < 127)
3875 ++regs->array[regno+i].set_in_loop;
3876 last_set[regno+i] = insn;
3882 /* Given a loop that is bounded by LOOP->START and LOOP->END and that
3883 is entered at LOOP->SCAN_START, return 1 if the register set in SET
3884 contained in insn INSN is used by any insn that precedes INSN in
3885 cyclic order starting from the loop entry point.
3887 We don't want to use INSN_LUID here because if we restrict INSN to those
3888 that have a valid INSN_LUID, it means we cannot move an invariant out
3889 from an inner loop past two loops. */
3892 loop_reg_used_before_p (const struct loop *loop, rtx set, rtx insn)
3894 rtx reg = SET_DEST (set);
3897 /* Scan forward checking for register usage. If we hit INSN, we
3898 are done. Otherwise, if we hit LOOP->END, wrap around to LOOP->START. */
3899 for (p = loop->scan_start; p != insn; p = NEXT_INSN (p))
3901 if (INSN_P (p) && reg_overlap_mentioned_p (reg, PATTERN (p)))
3912 /* Information we collect about arrays that we might want to prefetch. */
3913 struct prefetch_info
3915 struct iv_class *class; /* Class this prefetch is based on. */
3916 struct induction *giv; /* GIV this prefetch is based on. */
3917 rtx base_address; /* Start prefetching from this address plus
3919 HOST_WIDE_INT index;
3920 HOST_WIDE_INT stride; /* Prefetch stride in bytes in each
3922 unsigned int bytes_accessed; /* Sum of sizes of all accesses to this
3923 prefetch area in one iteration. */
3924 unsigned int total_bytes; /* Total bytes loop will access in this block.
3925 This is set only for loops with known
3926 iteration counts and is 0xffffffff
3928 int prefetch_in_loop; /* Number of prefetch insns in loop. */
3929 int prefetch_before_loop; /* Number of prefetch insns before loop. */
3930 unsigned int write : 1; /* 1 for read/write prefetches. */
3933 /* Data used by check_store function. */
3934 struct check_store_data
3940 static void check_store (rtx, rtx, void *);
3941 static void emit_prefetch_instructions (struct loop *);
3942 static int rtx_equal_for_prefetch_p (rtx, rtx);
3944 /* Set mem_write when mem_address is found. Used as callback to
3947 check_store (rtx x, rtx pat ATTRIBUTE_UNUSED, void *data)
3949 struct check_store_data *d = (struct check_store_data *) data;
3951 if ((MEM_P (x)) && rtx_equal_p (d->mem_address, XEXP (x, 0)))
3955 /* Like rtx_equal_p, but attempts to swap commutative operands. This is
3956 important to get some addresses combined. Later more sophisticated
3957 transformations can be added when necessary.
3959 ??? Same trick with swapping operand is done at several other places.
3960 It can be nice to develop some common way to handle this. */
3963 rtx_equal_for_prefetch_p (rtx x, rtx y)
3967 enum rtx_code code = GET_CODE (x);
3972 if (code != GET_CODE (y))
3975 if (GET_MODE (x) != GET_MODE (y))
3987 return XEXP (x, 0) == XEXP (y, 0);
3993 if (COMMUTATIVE_ARITH_P (x))
3995 return ((rtx_equal_for_prefetch_p (XEXP (x, 0), XEXP (y, 0))
3996 && rtx_equal_for_prefetch_p (XEXP (x, 1), XEXP (y, 1)))
3997 || (rtx_equal_for_prefetch_p (XEXP (x, 0), XEXP (y, 1))
3998 && rtx_equal_for_prefetch_p (XEXP (x, 1), XEXP (y, 0))));
4001 /* Compare the elements. If any pair of corresponding elements fails to
4002 match, return 0 for the whole thing. */
4004 fmt = GET_RTX_FORMAT (code);
4005 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4010 if (XWINT (x, i) != XWINT (y, i))
4015 if (XINT (x, i) != XINT (y, i))
4020 /* Two vectors must have the same length. */
4021 if (XVECLEN (x, i) != XVECLEN (y, i))
4024 /* And the corresponding elements must match. */
4025 for (j = 0; j < XVECLEN (x, i); j++)
4026 if (rtx_equal_for_prefetch_p (XVECEXP (x, i, j),
4027 XVECEXP (y, i, j)) == 0)
4032 if (rtx_equal_for_prefetch_p (XEXP (x, i), XEXP (y, i)) == 0)
4037 if (strcmp (XSTR (x, i), XSTR (y, i)))
4042 /* These are just backpointers, so they don't matter. */
4048 /* It is believed that rtx's at this level will never
4049 contain anything but integers and other rtx's,
4050 except for within LABEL_REFs and SYMBOL_REFs. */
4058 /* Remove constant addition value from the expression X (when present)
4061 static HOST_WIDE_INT
4062 remove_constant_addition (rtx *x)
4064 HOST_WIDE_INT addval = 0;
4067 /* Avoid clobbering a shared CONST expression. */
4068 if (GET_CODE (exp) == CONST)
4070 if (GET_CODE (XEXP (exp, 0)) == PLUS
4071 && GET_CODE (XEXP (XEXP (exp, 0), 0)) == SYMBOL_REF
4072 && GET_CODE (XEXP (XEXP (exp, 0), 1)) == CONST_INT)
4074 *x = XEXP (XEXP (exp, 0), 0);
4075 return INTVAL (XEXP (XEXP (exp, 0), 1));
4080 if (GET_CODE (exp) == CONST_INT)
4082 addval = INTVAL (exp);
4086 /* For plus expression recurse on ourself. */
4087 else if (GET_CODE (exp) == PLUS)
4089 addval += remove_constant_addition (&XEXP (exp, 0));
4090 addval += remove_constant_addition (&XEXP (exp, 1));
4092 /* In case our parameter was constant, remove extra zero from the
4094 if (XEXP (exp, 0) == const0_rtx)
4096 else if (XEXP (exp, 1) == const0_rtx)
4103 /* Attempt to identify accesses to arrays that are most likely to cause cache
4104 misses, and emit prefetch instructions a few prefetch blocks forward.
4106 To detect the arrays we use the GIV information that was collected by the
4107 strength reduction pass.
4109 The prefetch instructions are generated after the GIV information is done
4110 and before the strength reduction process. The new GIVs are injected into
4111 the strength reduction tables, so the prefetch addresses are optimized as
4114 GIVs are split into base address, stride, and constant addition values.
4115 GIVs with the same address, stride and close addition values are combined
4116 into a single prefetch. Also writes to GIVs are detected, so that prefetch
4117 for write instructions can be used for the block we write to, on machines
4118 that support write prefetches.
4120 Several heuristics are used to determine when to prefetch. They are
4121 controlled by defined symbols that can be overridden for each target. */
4124 emit_prefetch_instructions (struct loop *loop)
4126 int num_prefetches = 0;
4127 int num_real_prefetches = 0;
4128 int num_real_write_prefetches = 0;
4129 int num_prefetches_before = 0;
4130 int num_write_prefetches_before = 0;
4133 struct iv_class *bl;
4134 struct induction *iv;
4135 struct prefetch_info info[MAX_PREFETCHES];
4136 struct loop_ivs *ivs = LOOP_IVS (loop);
4138 if (!HAVE_prefetch || PREFETCH_BLOCK == 0)
4141 /* Consider only loops w/o calls. When a call is done, the loop is probably
4142 slow enough to read the memory. */
4143 if (PREFETCH_NO_CALL && LOOP_INFO (loop)->has_call)
4145 if (loop_dump_stream)
4146 fprintf (loop_dump_stream, "Prefetch: ignoring loop: has call.\n");
4151 /* Don't prefetch in loops known to have few iterations. */
4152 if (PREFETCH_NO_LOW_LOOPCNT
4153 && LOOP_INFO (loop)->n_iterations
4154 && LOOP_INFO (loop)->n_iterations <= PREFETCH_LOW_LOOPCNT)
4156 if (loop_dump_stream)
4157 fprintf (loop_dump_stream,
4158 "Prefetch: ignoring loop: not enough iterations.\n");
4162 /* Search all induction variables and pick those interesting for the prefetch
4164 for (bl = ivs->list; bl; bl = bl->next)
4166 struct induction *biv = bl->biv, *biv1;
4171 /* Expect all BIVs to be executed in each iteration. This makes our
4172 analysis more conservative. */
4175 /* Discard non-constant additions that we can't handle well yet, and
4176 BIVs that are executed multiple times; such BIVs ought to be
4177 handled in the nested loop. We accept not_every_iteration BIVs,
4178 since these only result in larger strides and make our
4179 heuristics more conservative. */
4180 if (GET_CODE (biv->add_val) != CONST_INT)
4182 if (loop_dump_stream)
4184 fprintf (loop_dump_stream,
4185 "Prefetch: ignoring biv %d: non-constant addition at insn %d:",
4186 REGNO (biv->src_reg), INSN_UID (biv->insn));
4187 print_rtl (loop_dump_stream, biv->add_val);
4188 fprintf (loop_dump_stream, "\n");
4193 if (biv->maybe_multiple)
4195 if (loop_dump_stream)
4197 fprintf (loop_dump_stream,
4198 "Prefetch: ignoring biv %d: maybe_multiple at insn %i:",
4199 REGNO (biv->src_reg), INSN_UID (biv->insn));
4200 print_rtl (loop_dump_stream, biv->add_val);
4201 fprintf (loop_dump_stream, "\n");
4206 basestride += INTVAL (biv1->add_val);
4207 biv1 = biv1->next_iv;
4210 if (biv1 || !basestride)
4213 for (iv = bl->giv; iv; iv = iv->next_iv)
4217 HOST_WIDE_INT index = 0;
4219 HOST_WIDE_INT stride = 0;
4220 int stride_sign = 1;
4221 struct check_store_data d;
4222 const char *ignore_reason = NULL;
4223 int size = GET_MODE_SIZE (GET_MODE (iv));
4225 /* See whether an induction variable is interesting to us and if
4226 not, report the reason. */
4227 if (iv->giv_type != DEST_ADDR)
4228 ignore_reason = "giv is not a destination address";
4230 /* We are interested only in constant stride memory references
4231 in order to be able to compute density easily. */
4232 else if (GET_CODE (iv->mult_val) != CONST_INT)
4233 ignore_reason = "stride is not constant";
4237 stride = INTVAL (iv->mult_val) * basestride;
4244 /* On some targets, reversed order prefetches are not
4246 if (PREFETCH_NO_REVERSE_ORDER && stride_sign < 0)
4247 ignore_reason = "reversed order stride";
4249 /* Prefetch of accesses with an extreme stride might not be
4250 worthwhile, either. */
4251 else if (PREFETCH_NO_EXTREME_STRIDE
4252 && stride > PREFETCH_EXTREME_STRIDE)
4253 ignore_reason = "extreme stride";
4255 /* Ignore GIVs with varying add values; we can't predict the
4256 value for the next iteration. */
4257 else if (!loop_invariant_p (loop, iv->add_val))
4258 ignore_reason = "giv has varying add value";
4260 /* Ignore GIVs in the nested loops; they ought to have been
4262 else if (iv->maybe_multiple)
4263 ignore_reason = "giv is in nested loop";
4266 if (ignore_reason != NULL)
4268 if (loop_dump_stream)
4269 fprintf (loop_dump_stream,
4270 "Prefetch: ignoring giv at %d: %s.\n",
4271 INSN_UID (iv->insn), ignore_reason);
4275 /* Determine the pointer to the basic array we are examining. It is
4276 the sum of the BIV's initial value and the GIV's add_val. */
4277 address = copy_rtx (iv->add_val);
4278 temp = copy_rtx (bl->initial_value);
4280 address = simplify_gen_binary (PLUS, Pmode, temp, address);
4281 index = remove_constant_addition (&address);
4284 d.mem_address = *iv->location;
4286 /* When the GIV is not always executed, we might be better off by
4287 not dirtying the cache pages. */
4288 if (PREFETCH_CONDITIONAL || iv->always_executed)
4289 note_stores (PATTERN (iv->insn), check_store, &d);
4292 if (loop_dump_stream)
4293 fprintf (loop_dump_stream, "Prefetch: Ignoring giv at %d: %s\n",
4294 INSN_UID (iv->insn), "in conditional code.");
4298 /* Attempt to find another prefetch to the same array and see if we
4299 can merge this one. */
4300 for (i = 0; i < num_prefetches; i++)
4301 if (rtx_equal_for_prefetch_p (address, info[i].base_address)
4302 && stride == info[i].stride)
4304 /* In case both access same array (same location
4305 just with small difference in constant indexes), merge
4306 the prefetches. Just do the later and the earlier will
4307 get prefetched from previous iteration.
4308 The artificial threshold should not be too small,
4309 but also not bigger than small portion of memory usually
4310 traversed by single loop. */
4311 if (index >= info[i].index
4312 && index - info[i].index < PREFETCH_EXTREME_DIFFERENCE)
4314 info[i].write |= d.mem_write;
4315 info[i].bytes_accessed += size;
4316 info[i].index = index;
4319 info[num_prefetches].base_address = address;
4324 if (index < info[i].index
4325 && info[i].index - index < PREFETCH_EXTREME_DIFFERENCE)
4327 info[i].write |= d.mem_write;
4328 info[i].bytes_accessed += size;
4334 /* Merging failed. */
4337 info[num_prefetches].giv = iv;
4338 info[num_prefetches].class = bl;
4339 info[num_prefetches].index = index;
4340 info[num_prefetches].stride = stride;
4341 info[num_prefetches].base_address = address;
4342 info[num_prefetches].write = d.mem_write;
4343 info[num_prefetches].bytes_accessed = size;
4345 if (num_prefetches >= MAX_PREFETCHES)
4347 if (loop_dump_stream)
4348 fprintf (loop_dump_stream,
4349 "Maximal number of prefetches exceeded.\n");
4356 for (i = 0; i < num_prefetches; i++)
4360 /* Attempt to calculate the total number of bytes fetched by all
4361 iterations of the loop. Avoid overflow. */
4362 if (LOOP_INFO (loop)->n_iterations
4363 && ((unsigned HOST_WIDE_INT) (0xffffffff / info[i].stride)
4364 >= LOOP_INFO (loop)->n_iterations))
4365 info[i].total_bytes = info[i].stride * LOOP_INFO (loop)->n_iterations;
4367 info[i].total_bytes = 0xffffffff;
4369 density = info[i].bytes_accessed * 100 / info[i].stride;
4371 /* Prefetch might be worthwhile only when the loads/stores are dense. */
4372 if (PREFETCH_ONLY_DENSE_MEM)
4373 if (density * 256 > PREFETCH_DENSE_MEM * 100
4374 && (info[i].total_bytes / PREFETCH_BLOCK
4375 >= PREFETCH_BLOCKS_BEFORE_LOOP_MIN))
4377 info[i].prefetch_before_loop = 1;
4378 info[i].prefetch_in_loop
4379 = (info[i].total_bytes / PREFETCH_BLOCK
4380 > PREFETCH_BLOCKS_BEFORE_LOOP_MAX);
4384 info[i].prefetch_in_loop = 0, info[i].prefetch_before_loop = 0;
4385 if (loop_dump_stream)
4386 fprintf (loop_dump_stream,
4387 "Prefetch: ignoring giv at %d: %d%% density is too low.\n",
4388 INSN_UID (info[i].giv->insn), density);
4391 info[i].prefetch_in_loop = 1, info[i].prefetch_before_loop = 1;
4393 /* Find how many prefetch instructions we'll use within the loop. */
4394 if (info[i].prefetch_in_loop != 0)
4396 info[i].prefetch_in_loop = ((info[i].stride + PREFETCH_BLOCK - 1)
4398 num_real_prefetches += info[i].prefetch_in_loop;
4400 num_real_write_prefetches += info[i].prefetch_in_loop;
4404 /* Determine how many iterations ahead to prefetch within the loop, based
4405 on how many prefetches we currently expect to do within the loop. */
4406 if (num_real_prefetches != 0)
4408 if ((ahead = SIMULTANEOUS_PREFETCHES / num_real_prefetches) == 0)
4410 if (loop_dump_stream)
4411 fprintf (loop_dump_stream,
4412 "Prefetch: ignoring prefetches within loop: ahead is zero; %d < %d\n",
4413 SIMULTANEOUS_PREFETCHES, num_real_prefetches);
4414 num_real_prefetches = 0, num_real_write_prefetches = 0;
4417 /* We'll also use AHEAD to determine how many prefetch instructions to
4418 emit before a loop, so don't leave it zero. */
4420 ahead = PREFETCH_BLOCKS_BEFORE_LOOP_MAX;
4422 for (i = 0; i < num_prefetches; i++)
4424 /* Update if we've decided not to prefetch anything within the loop. */
4425 if (num_real_prefetches == 0)
4426 info[i].prefetch_in_loop = 0;
4428 /* Find how many prefetch instructions we'll use before the loop. */
4429 if (info[i].prefetch_before_loop != 0)
4431 int n = info[i].total_bytes / PREFETCH_BLOCK;
4434 info[i].prefetch_before_loop = n;
4435 num_prefetches_before += n;
4437 num_write_prefetches_before += n;
4440 if (loop_dump_stream)
4442 if (info[i].prefetch_in_loop == 0
4443 && info[i].prefetch_before_loop == 0)
4445 fprintf (loop_dump_stream, "Prefetch insn: %d",
4446 INSN_UID (info[i].giv->insn));
4447 fprintf (loop_dump_stream,
4448 "; in loop: %d; before: %d; %s\n",
4449 info[i].prefetch_in_loop,
4450 info[i].prefetch_before_loop,
4451 info[i].write ? "read/write" : "read only");
4452 fprintf (loop_dump_stream,
4453 " density: %d%%; bytes_accessed: %u; total_bytes: %u\n",
4454 (int) (info[i].bytes_accessed * 100 / info[i].stride),
4455 info[i].bytes_accessed, info[i].total_bytes);
4456 fprintf (loop_dump_stream, " index: " HOST_WIDE_INT_PRINT_DEC
4457 "; stride: " HOST_WIDE_INT_PRINT_DEC "; address: ",
4458 info[i].index, info[i].stride);
4459 print_rtl (loop_dump_stream, info[i].base_address);
4460 fprintf (loop_dump_stream, "\n");
4464 if (num_real_prefetches + num_prefetches_before > 0)
4466 /* Record that this loop uses prefetch instructions. */
4467 LOOP_INFO (loop)->has_prefetch = 1;
4469 if (loop_dump_stream)
4471 fprintf (loop_dump_stream, "Real prefetches needed within loop: %d (write: %d)\n",
4472 num_real_prefetches, num_real_write_prefetches);
4473 fprintf (loop_dump_stream, "Real prefetches needed before loop: %d (write: %d)\n",
4474 num_prefetches_before, num_write_prefetches_before);
4478 for (i = 0; i < num_prefetches; i++)
4482 for (y = 0; y < info[i].prefetch_in_loop; y++)
4484 rtx loc = copy_rtx (*info[i].giv->location);
4486 int bytes_ahead = PREFETCH_BLOCK * (ahead + y);
4487 rtx before_insn = info[i].giv->insn;
4488 rtx prev_insn = PREV_INSN (info[i].giv->insn);
4491 /* We can save some effort by offsetting the address on
4492 architectures with offsettable memory references. */
4493 if (offsettable_address_p (0, VOIDmode, loc))
4494 loc = plus_constant (loc, bytes_ahead);
4497 rtx reg = gen_reg_rtx (Pmode);
4498 loop_iv_add_mult_emit_before (loop, loc, const1_rtx,
4499 GEN_INT (bytes_ahead), reg,
4505 /* Make sure the address operand is valid for prefetch. */
4506 if (! (*insn_data[(int)CODE_FOR_prefetch].operand[0].predicate)
4507 (loc, insn_data[(int)CODE_FOR_prefetch].operand[0].mode))
4508 loc = force_reg (Pmode, loc);
4509 emit_insn (gen_prefetch (loc, GEN_INT (info[i].write),
4513 emit_insn_before (seq, before_insn);
4515 /* Check all insns emitted and record the new GIV
4517 insn = NEXT_INSN (prev_insn);
4518 while (insn != before_insn)
4520 insn = check_insn_for_givs (loop, insn,
4521 info[i].giv->always_executed,
4522 info[i].giv->maybe_multiple);
4523 insn = NEXT_INSN (insn);
4527 if (PREFETCH_BEFORE_LOOP)
4529 /* Emit insns before the loop to fetch the first cache lines or,
4530 if we're not prefetching within the loop, everything we expect
4532 for (y = 0; y < info[i].prefetch_before_loop; y++)
4534 rtx reg = gen_reg_rtx (Pmode);
4535 rtx loop_start = loop->start;
4536 rtx init_val = info[i].class->initial_value;
4537 rtx add_val = simplify_gen_binary (PLUS, Pmode,
4538 info[i].giv->add_val,
4539 GEN_INT (y * PREFETCH_BLOCK));
4541 /* Functions called by LOOP_IV_ADD_EMIT_BEFORE expect a
4542 non-constant INIT_VAL to have the same mode as REG, which
4543 in this case we know to be Pmode. */
4544 if (GET_MODE (init_val) != Pmode && !CONSTANT_P (init_val))
4549 init_val = convert_to_mode (Pmode, init_val, 0);
4552 loop_insn_emit_before (loop, 0, loop_start, seq);
4554 loop_iv_add_mult_emit_before (loop, init_val,
4555 info[i].giv->mult_val,
4556 add_val, reg, 0, loop_start);
4557 emit_insn_before (gen_prefetch (reg, GEN_INT (info[i].write),
4567 /* Communication with routines called via `note_stores'. */
4569 static rtx note_insn;
4571 /* Dummy register to have nonzero DEST_REG for DEST_ADDR type givs. */
4573 static rtx addr_placeholder;
4575 /* ??? Unfinished optimizations, and possible future optimizations,
4576 for the strength reduction code. */
4578 /* ??? The interaction of biv elimination, and recognition of 'constant'
4579 bivs, may cause problems. */
4581 /* ??? Add heuristics so that DEST_ADDR strength reduction does not cause
4582 performance problems.
4584 Perhaps don't eliminate things that can be combined with an addressing
4585 mode. Find all givs that have the same biv, mult_val, and add_val;
4586 then for each giv, check to see if its only use dies in a following
4587 memory address. If so, generate a new memory address and check to see
4588 if it is valid. If it is valid, then store the modified memory address,
4589 otherwise, mark the giv as not done so that it will get its own iv. */
4591 /* ??? Could try to optimize branches when it is known that a biv is always
4594 /* ??? When replace a biv in a compare insn, we should replace with closest
4595 giv so that an optimized branch can still be recognized by the combiner,
4596 e.g. the VAX acb insn. */
4598 /* ??? Many of the checks involving uid_luid could be simplified if regscan
4599 was rerun in loop_optimize whenever a register was added or moved.
4600 Also, some of the optimizations could be a little less conservative. */
4602 /* Searches the insns between INSN and LOOP->END. Returns 1 if there
4603 is a backward branch in that range that branches to somewhere between
4604 LOOP->START and INSN. Returns 0 otherwise. */
4606 /* ??? This is quadratic algorithm. Could be rewritten to be linear.
4607 In practice, this is not a problem, because this function is seldom called,
4608 and uses a negligible amount of CPU time on average. */
4611 back_branch_in_range_p (const struct loop *loop, rtx insn)
4613 rtx p, q, target_insn;
4614 rtx loop_start = loop->start;
4615 rtx loop_end = loop->end;
4616 rtx orig_loop_end = loop->end;
4618 /* Stop before we get to the backward branch at the end of the loop. */
4619 loop_end = prev_nonnote_insn (loop_end);
4620 if (BARRIER_P (loop_end))
4621 loop_end = PREV_INSN (loop_end);
4623 /* Check in case insn has been deleted, search forward for first non
4624 deleted insn following it. */
4625 while (INSN_DELETED_P (insn))
4626 insn = NEXT_INSN (insn);
4628 /* Check for the case where insn is the last insn in the loop. Deal
4629 with the case where INSN was a deleted loop test insn, in which case
4630 it will now be the NOTE_LOOP_END. */
4631 if (insn == loop_end || insn == orig_loop_end)
4634 for (p = NEXT_INSN (insn); p != loop_end; p = NEXT_INSN (p))
4638 target_insn = JUMP_LABEL (p);
4640 /* Search from loop_start to insn, to see if one of them is
4641 the target_insn. We can't use INSN_LUID comparisons here,
4642 since insn may not have an LUID entry. */
4643 for (q = loop_start; q != insn; q = NEXT_INSN (q))
4644 if (q == target_insn)
4652 /* Scan the loop body and call FNCALL for each insn. In the addition to the
4653 LOOP and INSN parameters pass MAYBE_MULTIPLE and NOT_EVERY_ITERATION to the
4656 NOT_EVERY_ITERATION is 1 if current insn is not known to be executed at
4657 least once for every loop iteration except for the last one.
4659 MAYBE_MULTIPLE is 1 if current insn may be executed more than once for every
4662 typedef rtx (*loop_insn_callback) (struct loop *, rtx, int, int);
4664 for_each_insn_in_loop (struct loop *loop, loop_insn_callback fncall)
4666 int not_every_iteration = 0;
4667 int maybe_multiple = 0;
4668 int past_loop_latch = 0;
4669 bool exit_test_is_entry = false;
4672 /* If loop_scan_start points to the loop exit test, the loop body
4673 cannot be counted on running on every iteration, and we have to
4674 be wary of subversive use of gotos inside expression
4676 if (prev_nonnote_insn (loop->scan_start) != prev_nonnote_insn (loop->start))
4678 exit_test_is_entry = true;
4679 maybe_multiple = back_branch_in_range_p (loop, loop->scan_start);
4682 /* Scan through loop and update NOT_EVERY_ITERATION and MAYBE_MULTIPLE. */
4683 for (p = next_insn_in_loop (loop, loop->scan_start);
4685 p = next_insn_in_loop (loop, p))
4687 p = fncall (loop, p, not_every_iteration, maybe_multiple);
4689 /* Past CODE_LABEL, we get to insns that may be executed multiple
4690 times. The only way we can be sure that they can't is if every
4691 jump insn between here and the end of the loop either
4692 returns, exits the loop, is a jump to a location that is still
4693 behind the label, or is a jump to the loop start. */
4703 insn = NEXT_INSN (insn);
4704 if (insn == loop->scan_start)
4706 if (insn == loop->end)
4712 if (insn == loop->scan_start)
4717 && GET_CODE (PATTERN (insn)) != RETURN
4718 && (!any_condjump_p (insn)
4719 || (JUMP_LABEL (insn) != 0
4720 && JUMP_LABEL (insn) != loop->scan_start
4721 && !loop_insn_first_p (p, JUMP_LABEL (insn)))))
4729 /* Past a jump, we get to insns for which we can't count
4730 on whether they will be executed during each iteration. */
4731 /* This code appears twice in strength_reduce. There is also similar
4732 code in scan_loop. */
4734 /* If we enter the loop in the middle, and scan around to the
4735 beginning, don't set not_every_iteration for that.
4736 This can be any kind of jump, since we want to know if insns
4737 will be executed if the loop is executed. */
4738 && (exit_test_is_entry
4739 || !(JUMP_LABEL (p) == loop->top
4740 && ((NEXT_INSN (NEXT_INSN (p)) == loop->end
4741 && any_uncondjump_p (p))
4742 || (NEXT_INSN (p) == loop->end
4743 && any_condjump_p (p))))))
4747 /* If this is a jump outside the loop, then it also doesn't
4748 matter. Check to see if the target of this branch is on the
4749 loop->exits_labels list. */
4751 for (label = loop->exit_labels; label; label = LABEL_NEXTREF (label))
4752 if (XEXP (label, 0) == JUMP_LABEL (p))
4756 not_every_iteration = 1;
4759 /* Note if we pass a loop latch. If we do, then we can not clear
4760 NOT_EVERY_ITERATION below when we pass the last CODE_LABEL in
4761 a loop since a jump before the last CODE_LABEL may have started
4762 a new loop iteration.
4764 Note that LOOP_TOP is only set for rotated loops and we need
4765 this check for all loops, so compare against the CODE_LABEL
4766 which immediately follows LOOP_START. */
4768 && JUMP_LABEL (p) == NEXT_INSN (loop->start))
4769 past_loop_latch = 1;
4771 /* Unlike in the code motion pass where MAYBE_NEVER indicates that
4772 an insn may never be executed, NOT_EVERY_ITERATION indicates whether
4773 or not an insn is known to be executed each iteration of the
4774 loop, whether or not any iterations are known to occur.
4776 Therefore, if we have just passed a label and have no more labels
4777 between here and the test insn of the loop, and we have not passed
4778 a jump to the top of the loop, then we know these insns will be
4779 executed each iteration. */
4781 if (not_every_iteration
4784 && no_labels_between_p (p, loop->end))
4785 not_every_iteration = 0;
4790 loop_bivs_find (struct loop *loop)
4792 struct loop_regs *regs = LOOP_REGS (loop);
4793 struct loop_ivs *ivs = LOOP_IVS (loop);
4794 /* Temporary list pointers for traversing ivs->list. */
4795 struct iv_class *bl, **backbl;
4799 for_each_insn_in_loop (loop, check_insn_for_bivs);
4801 /* Scan ivs->list to remove all regs that proved not to be bivs.
4802 Make a sanity check against regs->n_times_set. */
4803 for (backbl = &ivs->list, bl = *backbl; bl; bl = bl->next)
4805 if (REG_IV_TYPE (ivs, bl->regno) != BASIC_INDUCT
4806 /* Above happens if register modified by subreg, etc. */
4807 /* Make sure it is not recognized as a basic induction var: */
4808 || regs->array[bl->regno].n_times_set != bl->biv_count
4809 /* If never incremented, it is invariant that we decided not to
4810 move. So leave it alone. */
4811 || ! bl->incremented)
4813 if (loop_dump_stream)
4814 fprintf (loop_dump_stream, "Biv %d: discarded, %s\n",
4816 (REG_IV_TYPE (ivs, bl->regno) != BASIC_INDUCT
4817 ? "not induction variable"
4818 : (! bl->incremented ? "never incremented"
4821 REG_IV_TYPE (ivs, bl->regno) = NOT_BASIC_INDUCT;
4828 if (loop_dump_stream)
4829 fprintf (loop_dump_stream, "Biv %d: verified\n", bl->regno);
4835 /* Determine how BIVS are initialized by looking through pre-header
4836 extended basic block. */
4838 loop_bivs_init_find (struct loop *loop)
4840 struct loop_ivs *ivs = LOOP_IVS (loop);
4841 /* Temporary list pointers for traversing ivs->list. */
4842 struct iv_class *bl;
4846 /* Find initial value for each biv by searching backwards from loop_start,
4847 halting at first label. Also record any test condition. */
4850 for (p = loop->start; p && !LABEL_P (p); p = PREV_INSN (p))
4860 note_stores (PATTERN (p), record_initial, ivs);
4862 /* Record any test of a biv that branches around the loop if no store
4863 between it and the start of loop. We only care about tests with
4864 constants and registers and only certain of those. */
4866 && JUMP_LABEL (p) != 0
4867 && next_real_insn (JUMP_LABEL (p)) == next_real_insn (loop->end)
4868 && (test = get_condition_for_loop (loop, p)) != 0
4869 && REG_P (XEXP (test, 0))
4870 && REGNO (XEXP (test, 0)) < max_reg_before_loop
4871 && (bl = REG_IV_CLASS (ivs, REGNO (XEXP (test, 0)))) != 0
4872 && valid_initial_value_p (XEXP (test, 1), p, call_seen, loop->start)
4873 && bl->init_insn == 0)
4875 /* If an NE test, we have an initial value! */
4876 if (GET_CODE (test) == NE)
4879 bl->init_set = gen_rtx_SET (VOIDmode,
4880 XEXP (test, 0), XEXP (test, 1));
4883 bl->initial_test = test;
4889 /* Look at the each biv and see if we can say anything better about its
4890 initial value from any initializing insns set up above. (This is done
4891 in two passes to avoid missing SETs in a PARALLEL.) */
4893 loop_bivs_check (struct loop *loop)
4895 struct loop_ivs *ivs = LOOP_IVS (loop);
4896 /* Temporary list pointers for traversing ivs->list. */
4897 struct iv_class *bl;
4898 struct iv_class **backbl;
4900 for (backbl = &ivs->list; (bl = *backbl); backbl = &bl->next)
4905 if (! bl->init_insn)
4908 /* IF INIT_INSN has a REG_EQUAL or REG_EQUIV note and the value
4909 is a constant, use the value of that. */
4910 if (((note = find_reg_note (bl->init_insn, REG_EQUAL, 0)) != NULL
4911 && CONSTANT_P (XEXP (note, 0)))
4912 || ((note = find_reg_note (bl->init_insn, REG_EQUIV, 0)) != NULL
4913 && CONSTANT_P (XEXP (note, 0))))
4914 src = XEXP (note, 0);
4916 src = SET_SRC (bl->init_set);
4918 if (loop_dump_stream)
4919 fprintf (loop_dump_stream,
4920 "Biv %d: initialized at insn %d: initial value ",
4921 bl->regno, INSN_UID (bl->init_insn));
4923 if ((GET_MODE (src) == GET_MODE (regno_reg_rtx[bl->regno])
4924 || GET_MODE (src) == VOIDmode)
4925 && valid_initial_value_p (src, bl->init_insn,
4926 LOOP_INFO (loop)->pre_header_has_call,
4929 bl->initial_value = src;
4931 if (loop_dump_stream)
4933 print_simple_rtl (loop_dump_stream, src);
4934 fputc ('\n', loop_dump_stream);
4937 /* If we can't make it a giv,
4938 let biv keep initial value of "itself". */
4939 else if (loop_dump_stream)
4940 fprintf (loop_dump_stream, "is complex\n");
4945 /* Search the loop for general induction variables. */
4948 loop_givs_find (struct loop* loop)
4950 for_each_insn_in_loop (loop, check_insn_for_givs);
4954 /* For each giv for which we still don't know whether or not it is
4955 replaceable, check to see if it is replaceable because its final value
4956 can be calculated. */
4959 loop_givs_check (struct loop *loop)
4961 struct loop_ivs *ivs = LOOP_IVS (loop);
4962 struct iv_class *bl;
4964 for (bl = ivs->list; bl; bl = bl->next)
4966 struct induction *v;
4968 for (v = bl->giv; v; v = v->next_iv)
4969 if (! v->replaceable && ! v->not_replaceable)
4970 check_final_value (loop, v);
4974 /* Try to generate the simplest rtx for the expression
4975 (PLUS (MULT mult1 mult2) add1). This is used to calculate the initial
4979 fold_rtx_mult_add (rtx mult1, rtx mult2, rtx add1, enum machine_mode mode)
4984 /* The modes must all be the same. This should always be true. For now,
4985 check to make sure. */
4986 gcc_assert (GET_MODE (mult1) == mode || GET_MODE (mult1) == VOIDmode);
4987 gcc_assert (GET_MODE (mult2) == mode || GET_MODE (mult2) == VOIDmode);
4988 gcc_assert (GET_MODE (add1) == mode || GET_MODE (add1) == VOIDmode);
4990 /* Ensure that if at least one of mult1/mult2 are constant, then mult2
4991 will be a constant. */
4992 if (GET_CODE (mult1) == CONST_INT)
4999 mult_res = simplify_binary_operation (MULT, mode, mult1, mult2);
5001 mult_res = gen_rtx_MULT (mode, mult1, mult2);
5003 /* Again, put the constant second. */
5004 if (GET_CODE (add1) == CONST_INT)
5011 result = simplify_binary_operation (PLUS, mode, add1, mult_res);
5013 result = gen_rtx_PLUS (mode, add1, mult_res);
5018 /* Searches the list of induction struct's for the biv BL, to try to calculate
5019 the total increment value for one iteration of the loop as a constant.
5021 Returns the increment value as an rtx, simplified as much as possible,
5022 if it can be calculated. Otherwise, returns 0. */
5025 biv_total_increment (const struct iv_class *bl)
5027 struct induction *v;
5030 /* For increment, must check every instruction that sets it. Each
5031 instruction must be executed only once each time through the loop.
5032 To verify this, we check that the insn is always executed, and that
5033 there are no backward branches after the insn that branch to before it.
5034 Also, the insn must have a mult_val of one (to make sure it really is
5037 result = const0_rtx;
5038 for (v = bl->biv; v; v = v->next_iv)
5040 if (v->always_computable && v->mult_val == const1_rtx
5041 && ! v->maybe_multiple
5042 && SCALAR_INT_MODE_P (v->mode))
5044 /* If we have already counted it, skip it. */
5048 result = fold_rtx_mult_add (result, const1_rtx, v->add_val, v->mode);
5057 /* Try to prove that the register is dead after the loop exits. Trace every
5058 loop exit looking for an insn that will always be executed, which sets
5059 the register to some value, and appears before the first use of the register
5060 is found. If successful, then return 1, otherwise return 0. */
5062 /* ?? Could be made more intelligent in the handling of jumps, so that
5063 it can search past if statements and other similar structures. */
5066 reg_dead_after_loop (const struct loop *loop, rtx reg)
5070 int label_count = 0;
5072 /* In addition to checking all exits of this loop, we must also check
5073 all exits of inner nested loops that would exit this loop. We don't
5074 have any way to identify those, so we just give up if there are any
5075 such inner loop exits. */
5077 for (label = loop->exit_labels; label; label = LABEL_NEXTREF (label))
5080 if (label_count != loop->exit_count)
5083 /* HACK: Must also search the loop fall through exit, create a label_ref
5084 here which points to the loop->end, and append the loop_number_exit_labels
5086 label = gen_rtx_LABEL_REF (Pmode, loop->end);
5087 LABEL_NEXTREF (label) = loop->exit_labels;
5089 for (; label; label = LABEL_NEXTREF (label))
5091 /* Succeed if find an insn which sets the biv or if reach end of
5092 function. Fail if find an insn that uses the biv, or if come to
5093 a conditional jump. */
5095 insn = NEXT_INSN (XEXP (label, 0));
5102 if (reg_referenced_p (reg, PATTERN (insn)))
5105 note = find_reg_equal_equiv_note (insn);
5106 if (note && reg_overlap_mentioned_p (reg, XEXP (note, 0)))
5109 set = single_set (insn);
5110 if (set && rtx_equal_p (SET_DEST (set), reg))
5115 if (GET_CODE (PATTERN (insn)) == RETURN)
5117 else if (!any_uncondjump_p (insn)
5118 /* Prevent infinite loop following infinite loops. */
5119 || jump_count++ > 20)
5122 insn = JUMP_LABEL (insn);
5126 insn = NEXT_INSN (insn);
5130 /* Success, the register is dead on all loop exits. */
5134 /* Try to calculate the final value of the biv, the value it will have at
5135 the end of the loop. If we can do it, return that value. */
5138 final_biv_value (const struct loop *loop, struct iv_class *bl)
5140 unsigned HOST_WIDE_INT n_iterations = LOOP_INFO (loop)->n_iterations;
5143 /* ??? This only works for MODE_INT biv's. Reject all others for now. */
5145 if (GET_MODE_CLASS (bl->biv->mode) != MODE_INT)
5148 /* The final value for reversed bivs must be calculated differently than
5149 for ordinary bivs. In this case, there is already an insn after the
5150 loop which sets this biv's final value (if necessary), and there are
5151 no other loop exits, so we can return any value. */
5154 if (loop_dump_stream)
5155 fprintf (loop_dump_stream,
5156 "Final biv value for %d, reversed biv.\n", bl->regno);
5161 /* Try to calculate the final value as initial value + (number of iterations
5162 * increment). For this to work, increment must be invariant, the only
5163 exit from the loop must be the fall through at the bottom (otherwise
5164 it may not have its final value when the loop exits), and the initial
5165 value of the biv must be invariant. */
5167 if (n_iterations != 0
5168 && ! loop->exit_count
5169 && loop_invariant_p (loop, bl->initial_value))
5171 increment = biv_total_increment (bl);
5173 if (increment && loop_invariant_p (loop, increment))
5175 /* Can calculate the loop exit value, emit insns after loop
5176 end to calculate this value into a temporary register in
5177 case it is needed later. */
5179 tem = gen_reg_rtx (bl->biv->mode);
5180 record_base_value (REGNO (tem), bl->biv->add_val, 0);
5181 loop_iv_add_mult_sink (loop, increment, GEN_INT (n_iterations),
5182 bl->initial_value, tem);
5184 if (loop_dump_stream)
5185 fprintf (loop_dump_stream,
5186 "Final biv value for %d, calculated.\n", bl->regno);
5192 /* Check to see if the biv is dead at all loop exits. */
5193 if (reg_dead_after_loop (loop, bl->biv->src_reg))
5195 if (loop_dump_stream)
5196 fprintf (loop_dump_stream,
5197 "Final biv value for %d, biv dead after loop exit.\n",
5206 /* Return nonzero if it is possible to eliminate the biv BL provided
5207 all givs are reduced. This is possible if either the reg is not
5208 used outside the loop, or we can compute what its final value will
5212 loop_biv_eliminable_p (struct loop *loop, struct iv_class *bl,
5213 int threshold, int insn_count)
5215 /* For architectures with a decrement_and_branch_until_zero insn,
5216 don't do this if we put a REG_NONNEG note on the endtest for this
5219 #ifdef HAVE_decrement_and_branch_until_zero
5222 if (loop_dump_stream)
5223 fprintf (loop_dump_stream,
5224 "Cannot eliminate nonneg biv %d.\n", bl->regno);
5229 /* Check that biv is used outside loop or if it has a final value.
5230 Compare against bl->init_insn rather than loop->start. We aren't
5231 concerned with any uses of the biv between init_insn and
5232 loop->start since these won't be affected by the value of the biv
5233 elsewhere in the function, so long as init_insn doesn't use the
5236 if ((REGNO_LAST_LUID (bl->regno) < INSN_LUID (loop->end)
5238 && INSN_UID (bl->init_insn) < max_uid_for_loop
5239 && REGNO_FIRST_LUID (bl->regno) >= INSN_LUID (bl->init_insn)
5240 && ! reg_mentioned_p (bl->biv->dest_reg, SET_SRC (bl->init_set)))
5241 || (bl->final_value = final_biv_value (loop, bl)))
5242 return maybe_eliminate_biv (loop, bl, 0, threshold, insn_count);
5244 if (loop_dump_stream)
5246 fprintf (loop_dump_stream,
5247 "Cannot eliminate biv %d.\n",
5249 fprintf (loop_dump_stream,
5250 "First use: insn %d, last use: insn %d.\n",
5251 REGNO_FIRST_UID (bl->regno),
5252 REGNO_LAST_UID (bl->regno));
5258 /* Reduce each giv of BL that we have decided to reduce. */
5261 loop_givs_reduce (struct loop *loop, struct iv_class *bl)
5263 struct induction *v;
5265 for (v = bl->giv; v; v = v->next_iv)
5267 struct induction *tv;
5268 if (! v->ignore && v->same == 0)
5270 int auto_inc_opt = 0;
5272 /* If the code for derived givs immediately below has already
5273 allocated a new_reg, we must keep it. */
5275 v->new_reg = gen_reg_rtx (v->mode);
5278 /* If the target has auto-increment addressing modes, and
5279 this is an address giv, then try to put the increment
5280 immediately after its use, so that flow can create an
5281 auto-increment addressing mode. */
5282 /* Don't do this for loops entered at the bottom, to avoid
5283 this invalid transformation:
5292 if (v->giv_type == DEST_ADDR && bl->biv_count == 1
5293 && bl->biv->always_executed && ! bl->biv->maybe_multiple
5294 /* We don't handle reversed biv's because bl->biv->insn
5295 does not have a valid INSN_LUID. */
5297 && v->always_executed && ! v->maybe_multiple
5298 && INSN_UID (v->insn) < max_uid_for_loop
5301 /* If other giv's have been combined with this one, then
5302 this will work only if all uses of the other giv's occur
5303 before this giv's insn. This is difficult to check.
5305 We simplify this by looking for the common case where
5306 there is one DEST_REG giv, and this giv's insn is the
5307 last use of the dest_reg of that DEST_REG giv. If the
5308 increment occurs after the address giv, then we can
5309 perform the optimization. (Otherwise, the increment
5310 would have to go before other_giv, and we would not be
5311 able to combine it with the address giv to get an
5312 auto-inc address.) */
5313 if (v->combined_with)
5315 struct induction *other_giv = 0;
5317 for (tv = bl->giv; tv; tv = tv->next_iv)
5325 if (! tv && other_giv
5326 && REGNO (other_giv->dest_reg) < max_reg_before_loop
5327 && (REGNO_LAST_UID (REGNO (other_giv->dest_reg))
5328 == INSN_UID (v->insn))
5329 && INSN_LUID (v->insn) < INSN_LUID (bl->biv->insn))
5332 /* Check for case where increment is before the address
5333 giv. Do this test in "loop order". */
5334 else if ((INSN_LUID (v->insn) > INSN_LUID (bl->biv->insn)
5335 && (INSN_LUID (v->insn) < INSN_LUID (loop->scan_start)
5336 || (INSN_LUID (bl->biv->insn)
5337 > INSN_LUID (loop->scan_start))))
5338 || (INSN_LUID (v->insn) < INSN_LUID (loop->scan_start)
5339 && (INSN_LUID (loop->scan_start)
5340 < INSN_LUID (bl->biv->insn))))
5349 /* We can't put an insn immediately after one setting
5350 cc0, or immediately before one using cc0. */
5351 if ((auto_inc_opt == 1 && sets_cc0_p (PATTERN (v->insn)))
5352 || (auto_inc_opt == -1
5353 && (prev = prev_nonnote_insn (v->insn)) != 0
5355 && sets_cc0_p (PATTERN (prev))))
5361 v->auto_inc_opt = 1;
5365 /* For each place where the biv is incremented, add an insn
5366 to increment the new, reduced reg for the giv. */
5367 for (tv = bl->biv; tv; tv = tv->next_iv)
5371 /* Skip if location is the same as a previous one. */
5375 insert_before = NEXT_INSN (tv->insn);
5376 else if (auto_inc_opt == 1)
5377 insert_before = NEXT_INSN (v->insn);
5379 insert_before = v->insn;
5381 if (tv->mult_val == const1_rtx)
5382 loop_iv_add_mult_emit_before (loop, tv->add_val, v->mult_val,
5383 v->new_reg, v->new_reg,
5385 else /* tv->mult_val == const0_rtx */
5386 /* A multiply is acceptable here
5387 since this is presumed to be seldom executed. */
5388 loop_iv_add_mult_emit_before (loop, tv->add_val, v->mult_val,
5389 v->add_val, v->new_reg,
5393 /* Add code at loop start to initialize giv's reduced reg. */
5395 loop_iv_add_mult_hoist (loop,
5396 extend_value_for_giv (v, bl->initial_value),
5397 v->mult_val, v->add_val, v->new_reg);
5403 /* Check for givs whose first use is their definition and whose
5404 last use is the definition of another giv. If so, it is likely
5405 dead and should not be used to derive another giv nor to
5409 loop_givs_dead_check (struct loop *loop ATTRIBUTE_UNUSED, struct iv_class *bl)
5411 struct induction *v;
5413 for (v = bl->giv; v; v = v->next_iv)
5416 || (v->same && v->same->ignore))
5419 if (v->giv_type == DEST_REG
5420 && REGNO_FIRST_UID (REGNO (v->dest_reg)) == INSN_UID (v->insn))
5422 struct induction *v1;
5424 for (v1 = bl->giv; v1; v1 = v1->next_iv)
5425 if (REGNO_LAST_UID (REGNO (v->dest_reg)) == INSN_UID (v1->insn))
5433 loop_givs_rescan (struct loop *loop, struct iv_class *bl, rtx *reg_map)
5435 struct induction *v;
5437 for (v = bl->giv; v; v = v->next_iv)
5439 if (v->same && v->same->ignore)
5445 /* Update expression if this was combined, in case other giv was
5448 v->new_reg = replace_rtx (v->new_reg,
5449 v->same->dest_reg, v->same->new_reg);
5451 /* See if this register is known to be a pointer to something. If
5452 so, see if we can find the alignment. First see if there is a
5453 destination register that is a pointer. If so, this shares the
5454 alignment too. Next see if we can deduce anything from the
5455 computational information. If not, and this is a DEST_ADDR
5456 giv, at least we know that it's a pointer, though we don't know
5458 if (REG_P (v->new_reg)
5459 && v->giv_type == DEST_REG
5460 && REG_POINTER (v->dest_reg))
5461 mark_reg_pointer (v->new_reg,
5462 REGNO_POINTER_ALIGN (REGNO (v->dest_reg)));
5463 else if (REG_P (v->new_reg)
5464 && REG_POINTER (v->src_reg))
5466 unsigned int align = REGNO_POINTER_ALIGN (REGNO (v->src_reg));
5469 || GET_CODE (v->add_val) != CONST_INT
5470 || INTVAL (v->add_val) % (align / BITS_PER_UNIT) != 0)
5473 mark_reg_pointer (v->new_reg, align);
5475 else if (REG_P (v->new_reg)
5476 && REG_P (v->add_val)
5477 && REG_POINTER (v->add_val))
5479 unsigned int align = REGNO_POINTER_ALIGN (REGNO (v->add_val));
5481 if (align == 0 || GET_CODE (v->mult_val) != CONST_INT
5482 || INTVAL (v->mult_val) % (align / BITS_PER_UNIT) != 0)
5485 mark_reg_pointer (v->new_reg, align);
5487 else if (REG_P (v->new_reg) && v->giv_type == DEST_ADDR)
5488 mark_reg_pointer (v->new_reg, 0);
5490 if (v->giv_type == DEST_ADDR)
5492 /* Store reduced reg as the address in the memref where we found
5494 if (validate_change_maybe_volatile (v->insn, v->location,
5496 /* Yay, it worked! */;
5497 /* Not replaceable; emit an insn to set the original
5498 giv reg from the reduced giv. */
5499 else if (REG_P (*v->location))
5503 tem = force_operand (v->new_reg, *v->location);
5504 if (tem != *v->location)
5505 emit_move_insn (*v->location, tem);
5508 loop_insn_emit_before (loop, 0, v->insn, tem);
5510 else if (GET_CODE (*v->location) == PLUS
5511 && REG_P (XEXP (*v->location, 0))
5512 && CONSTANT_P (XEXP (*v->location, 1)))
5516 tem = expand_simple_binop (GET_MODE (*v->location), MINUS,
5517 v->new_reg, XEXP (*v->location, 1),
5518 NULL_RTX, 0, OPTAB_LIB_WIDEN);
5519 emit_move_insn (XEXP (*v->location, 0), tem);
5522 loop_insn_emit_before (loop, 0, v->insn, tem);
5526 /* If it wasn't a reg, create a pseudo and use that. */
5529 reg = force_reg (v->mode, *v->location);
5530 if (validate_change_maybe_volatile (v->insn, v->location, reg))
5534 loop_insn_emit_before (loop, 0, v->insn, seq);
5539 if (loop_dump_stream)
5540 fprintf (loop_dump_stream,
5541 "unable to reduce iv in insn %d\n",
5542 INSN_UID (v->insn));
5543 bl->all_reduced = 0;
5549 else if (v->replaceable)
5551 reg_map[REGNO (v->dest_reg)] = v->new_reg;
5555 rtx original_insn = v->insn;
5558 /* Not replaceable; emit an insn to set the original giv reg from
5559 the reduced giv, same as above. */
5560 v->insn = loop_insn_emit_after (loop, 0, original_insn,
5561 gen_move_insn (v->dest_reg,
5564 /* The original insn may have a REG_EQUAL note. This note is
5565 now incorrect and may result in invalid substitutions later.
5566 The original insn is dead, but may be part of a libcall
5567 sequence, which doesn't seem worth the bother of handling. */
5568 note = find_reg_note (original_insn, REG_EQUAL, NULL_RTX);
5570 remove_note (original_insn, note);
5573 /* When a loop is reversed, givs which depend on the reversed
5574 biv, and which are live outside the loop, must be set to their
5575 correct final value. This insn is only needed if the giv is
5576 not replaceable. The correct final value is the same as the
5577 value that the giv starts the reversed loop with. */
5578 if (bl->reversed && ! v->replaceable)
5579 loop_iv_add_mult_sink (loop,
5580 extend_value_for_giv (v, bl->initial_value),
5581 v->mult_val, v->add_val, v->dest_reg);
5582 else if (v->final_value)
5583 loop_insn_sink_or_swim (loop,
5584 gen_load_of_final_value (v->dest_reg,
5587 if (loop_dump_stream)
5589 fprintf (loop_dump_stream, "giv at %d reduced to ",
5590 INSN_UID (v->insn));
5591 print_simple_rtl (loop_dump_stream, v->new_reg);
5592 fprintf (loop_dump_stream, "\n");
5599 loop_giv_reduce_benefit (struct loop *loop ATTRIBUTE_UNUSED,
5600 struct iv_class *bl, struct induction *v,
5606 benefit = v->benefit;
5607 PUT_MODE (test_reg, v->mode);
5608 add_cost = iv_add_mult_cost (bl->biv->add_val, v->mult_val,
5609 test_reg, test_reg);
5611 /* Reduce benefit if not replaceable, since we will insert a
5612 move-insn to replace the insn that calculates this giv. Don't do
5613 this unless the giv is a user variable, since it will often be
5614 marked non-replaceable because of the duplication of the exit
5615 code outside the loop. In such a case, the copies we insert are
5616 dead and will be deleted. So they don't have a cost. Similar
5617 situations exist. */
5618 /* ??? The new final_[bg]iv_value code does a much better job of
5619 finding replaceable giv's, and hence this code may no longer be
5621 if (! v->replaceable && ! bl->eliminable
5622 && REG_USERVAR_P (v->dest_reg))
5623 benefit -= copy_cost;
5625 /* Decrease the benefit to count the add-insns that we will insert
5626 to increment the reduced reg for the giv. ??? This can
5627 overestimate the run-time cost of the additional insns, e.g. if
5628 there are multiple basic blocks that increment the biv, but only
5629 one of these blocks is executed during each iteration. There is
5630 no good way to detect cases like this with the current structure
5631 of the loop optimizer. This code is more accurate for
5632 determining code size than run-time benefits. */
5633 benefit -= add_cost * bl->biv_count;
5635 /* Decide whether to strength-reduce this giv or to leave the code
5636 unchanged (recompute it from the biv each time it is used). This
5637 decision can be made independently for each giv. */
5640 /* Attempt to guess whether autoincrement will handle some of the
5641 new add insns; if so, increase BENEFIT (undo the subtraction of
5642 add_cost that was done above). */
5643 if (v->giv_type == DEST_ADDR
5644 /* Increasing the benefit is risky, since this is only a guess.
5645 Avoid increasing register pressure in cases where there would
5646 be no other benefit from reducing this giv. */
5648 && GET_CODE (v->mult_val) == CONST_INT)
5650 int size = GET_MODE_SIZE (GET_MODE (v->mem));
5652 if (HAVE_POST_INCREMENT
5653 && INTVAL (v->mult_val) == size)
5654 benefit += add_cost * bl->biv_count;
5655 else if (HAVE_PRE_INCREMENT
5656 && INTVAL (v->mult_val) == size)
5657 benefit += add_cost * bl->biv_count;
5658 else if (HAVE_POST_DECREMENT
5659 && -INTVAL (v->mult_val) == size)
5660 benefit += add_cost * bl->biv_count;
5661 else if (HAVE_PRE_DECREMENT
5662 && -INTVAL (v->mult_val) == size)
5663 benefit += add_cost * bl->biv_count;
5671 /* Free IV structures for LOOP. */
5674 loop_ivs_free (struct loop *loop)
5676 struct loop_ivs *ivs = LOOP_IVS (loop);
5677 struct iv_class *iv = ivs->list;
5683 struct iv_class *next = iv->next;
5684 struct induction *induction;
5685 struct induction *next_induction;
5687 for (induction = iv->biv; induction; induction = next_induction)
5689 next_induction = induction->next_iv;
5692 for (induction = iv->giv; induction; induction = next_induction)
5694 next_induction = induction->next_iv;
5703 /* Look back before LOOP->START for the insn that sets REG and return
5704 the equivalent constant if there is a REG_EQUAL note otherwise just
5705 the SET_SRC of REG. */
5708 loop_find_equiv_value (const struct loop *loop, rtx reg)
5710 rtx loop_start = loop->start;
5715 for (insn = PREV_INSN (loop_start); insn; insn = PREV_INSN (insn))
5720 else if (INSN_P (insn) && reg_set_p (reg, insn))
5722 /* We found the last insn before the loop that sets the register.
5723 If it sets the entire register, and has a REG_EQUAL note,
5724 then use the value of the REG_EQUAL note. */
5725 if ((set = single_set (insn))
5726 && (SET_DEST (set) == reg))
5728 rtx note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
5730 /* Only use the REG_EQUAL note if it is a constant.
5731 Other things, divide in particular, will cause
5732 problems later if we use them. */
5733 if (note && GET_CODE (XEXP (note, 0)) != EXPR_LIST
5734 && CONSTANT_P (XEXP (note, 0)))
5735 ret = XEXP (note, 0);
5737 ret = SET_SRC (set);
5739 /* We cannot do this if it changes between the
5740 assignment and loop start though. */
5741 if (modified_between_p (ret, insn, loop_start))
5750 /* Find and return register term common to both expressions OP0 and
5751 OP1 or NULL_RTX if no such term exists. Each expression must be a
5752 REG or a PLUS of a REG. */
5755 find_common_reg_term (rtx op0, rtx op1)
5757 if ((REG_P (op0) || GET_CODE (op0) == PLUS)
5758 && (REG_P (op1) || GET_CODE (op1) == PLUS))
5765 if (GET_CODE (op0) == PLUS)
5766 op01 = XEXP (op0, 1), op00 = XEXP (op0, 0);
5768 op01 = const0_rtx, op00 = op0;
5770 if (GET_CODE (op1) == PLUS)
5771 op11 = XEXP (op1, 1), op10 = XEXP (op1, 0);
5773 op11 = const0_rtx, op10 = op1;
5775 /* Find and return common register term if present. */
5776 if (REG_P (op00) && (op00 == op10 || op00 == op11))
5778 else if (REG_P (op01) && (op01 == op10 || op01 == op11))
5782 /* No common register term found. */
5786 /* Determine the loop iterator and calculate the number of loop
5787 iterations. Returns the exact number of loop iterations if it can
5788 be calculated, otherwise returns zero. */
5790 static unsigned HOST_WIDE_INT
5791 loop_iterations (struct loop *loop)
5793 struct loop_info *loop_info = LOOP_INFO (loop);
5794 struct loop_ivs *ivs = LOOP_IVS (loop);
5795 rtx comparison, comparison_value;
5796 rtx iteration_var, initial_value, increment, final_value;
5797 enum rtx_code comparison_code;
5799 unsigned HOST_WIDE_INT abs_inc;
5800 unsigned HOST_WIDE_INT abs_diff;
5803 int unsigned_p, compare_dir, final_larger;
5805 struct iv_class *bl;
5807 loop_info->n_iterations = 0;
5808 loop_info->initial_value = 0;
5809 loop_info->initial_equiv_value = 0;
5810 loop_info->comparison_value = 0;
5811 loop_info->final_value = 0;
5812 loop_info->final_equiv_value = 0;
5813 loop_info->increment = 0;
5814 loop_info->iteration_var = 0;
5817 /* We used to use prev_nonnote_insn here, but that fails because it might
5818 accidentally get the branch for a contained loop if the branch for this
5819 loop was deleted. We can only trust branches immediately before the
5821 last_loop_insn = PREV_INSN (loop->end);
5823 /* ??? We should probably try harder to find the jump insn
5824 at the end of the loop. The following code assumes that
5825 the last loop insn is a jump to the top of the loop. */
5826 if (!JUMP_P (last_loop_insn))
5828 if (loop_dump_stream)
5829 fprintf (loop_dump_stream,
5830 "Loop iterations: No final conditional branch found.\n");
5834 /* If there is a more than a single jump to the top of the loop
5835 we cannot (easily) determine the iteration count. */
5836 if (LABEL_NUSES (JUMP_LABEL (last_loop_insn)) > 1)
5838 if (loop_dump_stream)
5839 fprintf (loop_dump_stream,
5840 "Loop iterations: Loop has multiple back edges.\n");
5844 /* Find the iteration variable. If the last insn is a conditional
5845 branch, and the insn before tests a register value, make that the
5846 iteration variable. */
5848 comparison = get_condition_for_loop (loop, last_loop_insn);
5849 if (comparison == 0)
5851 if (loop_dump_stream)
5852 fprintf (loop_dump_stream,
5853 "Loop iterations: No final comparison found.\n");
5857 /* ??? Get_condition may switch position of induction variable and
5858 invariant register when it canonicalizes the comparison. */
5860 comparison_code = GET_CODE (comparison);
5861 iteration_var = XEXP (comparison, 0);
5862 comparison_value = XEXP (comparison, 1);
5864 if (!REG_P (iteration_var))
5866 if (loop_dump_stream)
5867 fprintf (loop_dump_stream,
5868 "Loop iterations: Comparison not against register.\n");
5872 /* The only new registers that are created before loop iterations
5873 are givs made from biv increments or registers created by
5874 load_mems. In the latter case, it is possible that try_copy_prop
5875 will propagate a new pseudo into the old iteration register but
5876 this will be marked by having the REG_USERVAR_P bit set. */
5878 gcc_assert ((unsigned) REGNO (iteration_var) < ivs->n_regs
5879 || REG_USERVAR_P (iteration_var));
5881 /* Determine the initial value of the iteration variable, and the amount
5882 that it is incremented each loop. Use the tables constructed by
5883 the strength reduction pass to calculate these values. */
5885 /* Clear the result values, in case no answer can be found. */
5889 /* The iteration variable can be either a giv or a biv. Check to see
5890 which it is, and compute the variable's initial value, and increment
5891 value if possible. */
5893 /* If this is a new register, can't handle it since we don't have any
5894 reg_iv_type entry for it. */
5895 if ((unsigned) REGNO (iteration_var) >= ivs->n_regs)
5897 if (loop_dump_stream)
5898 fprintf (loop_dump_stream,
5899 "Loop iterations: No reg_iv_type entry for iteration var.\n");
5903 /* Reject iteration variables larger than the host wide int size, since they
5904 could result in a number of iterations greater than the range of our
5905 `unsigned HOST_WIDE_INT' variable loop_info->n_iterations. */
5906 else if ((GET_MODE_BITSIZE (GET_MODE (iteration_var))
5907 > HOST_BITS_PER_WIDE_INT))
5909 if (loop_dump_stream)
5910 fprintf (loop_dump_stream,
5911 "Loop iterations: Iteration var rejected because mode too large.\n");
5914 else if (GET_MODE_CLASS (GET_MODE (iteration_var)) != MODE_INT)
5916 if (loop_dump_stream)
5917 fprintf (loop_dump_stream,
5918 "Loop iterations: Iteration var not an integer.\n");
5922 /* Try swapping the comparison to identify a suitable iv. */
5923 if (REG_IV_TYPE (ivs, REGNO (iteration_var)) != BASIC_INDUCT
5924 && REG_IV_TYPE (ivs, REGNO (iteration_var)) != GENERAL_INDUCT
5925 && REG_P (comparison_value)
5926 && REGNO (comparison_value) < ivs->n_regs)
5928 rtx temp = comparison_value;
5929 comparison_code = swap_condition (comparison_code);
5930 comparison_value = iteration_var;
5931 iteration_var = temp;
5934 if (REG_IV_TYPE (ivs, REGNO (iteration_var)) == BASIC_INDUCT)
5936 gcc_assert (REGNO (iteration_var) < ivs->n_regs);
5938 /* Grab initial value, only useful if it is a constant. */
5939 bl = REG_IV_CLASS (ivs, REGNO (iteration_var));
5940 initial_value = bl->initial_value;
5941 if (!bl->biv->always_executed || bl->biv->maybe_multiple)
5943 if (loop_dump_stream)
5944 fprintf (loop_dump_stream,
5945 "Loop iterations: Basic induction var not set once in each iteration.\n");
5949 increment = biv_total_increment (bl);
5951 else if (REG_IV_TYPE (ivs, REGNO (iteration_var)) == GENERAL_INDUCT)
5953 HOST_WIDE_INT offset = 0;
5954 struct induction *v = REG_IV_INFO (ivs, REGNO (iteration_var));
5955 rtx biv_initial_value;
5957 gcc_assert (REGNO (v->src_reg) < ivs->n_regs);
5959 if (!v->always_executed || v->maybe_multiple)
5961 if (loop_dump_stream)
5962 fprintf (loop_dump_stream,
5963 "Loop iterations: General induction var not set once in each iteration.\n");
5967 bl = REG_IV_CLASS (ivs, REGNO (v->src_reg));
5969 /* Increment value is mult_val times the increment value of the biv. */
5971 increment = biv_total_increment (bl);
5974 struct induction *biv_inc;
5976 increment = fold_rtx_mult_add (v->mult_val,
5977 extend_value_for_giv (v, increment),
5978 const0_rtx, v->mode);
5979 /* The caller assumes that one full increment has occurred at the
5980 first loop test. But that's not true when the biv is incremented
5981 after the giv is set (which is the usual case), e.g.:
5982 i = 6; do {;} while (i++ < 9) .
5983 Therefore, we bias the initial value by subtracting the amount of
5984 the increment that occurs between the giv set and the giv test. */
5985 for (biv_inc = bl->biv; biv_inc; biv_inc = biv_inc->next_iv)
5987 if (loop_insn_first_p (v->insn, biv_inc->insn))
5989 if (REG_P (biv_inc->add_val))
5991 if (loop_dump_stream)
5992 fprintf (loop_dump_stream,
5993 "Loop iterations: Basic induction var add_val is REG %d.\n",
5994 REGNO (biv_inc->add_val));
5998 /* If we have already counted it, skip it. */
6002 offset -= INTVAL (biv_inc->add_val);
6006 if (loop_dump_stream)
6007 fprintf (loop_dump_stream,
6008 "Loop iterations: Giv iterator, initial value bias %ld.\n",
6011 /* Initial value is mult_val times the biv's initial value plus
6012 add_val. Only useful if it is a constant. */
6013 biv_initial_value = extend_value_for_giv (v, bl->initial_value);
6015 = fold_rtx_mult_add (v->mult_val,
6016 plus_constant (biv_initial_value, offset),
6017 v->add_val, v->mode);
6021 if (loop_dump_stream)
6022 fprintf (loop_dump_stream,
6023 "Loop iterations: Not basic or general induction var.\n");
6027 if (initial_value == 0)
6032 switch (comparison_code)
6047 /* Cannot determine loop iterations with this case. */
6067 /* If the comparison value is an invariant register, then try to find
6068 its value from the insns before the start of the loop. */
6070 final_value = comparison_value;
6071 if (REG_P (comparison_value)
6072 && loop_invariant_p (loop, comparison_value))
6074 final_value = loop_find_equiv_value (loop, comparison_value);
6076 /* If we don't get an invariant final value, we are better
6077 off with the original register. */
6078 if (! loop_invariant_p (loop, final_value))
6079 final_value = comparison_value;
6082 /* Calculate the approximate final value of the induction variable
6083 (on the last successful iteration). The exact final value
6084 depends on the branch operator, and increment sign. It will be
6085 wrong if the iteration variable is not incremented by one each
6086 time through the loop and (comparison_value + off_by_one -
6087 initial_value) % increment != 0.
6088 ??? Note that the final_value may overflow and thus final_larger
6089 will be bogus. A potentially infinite loop will be classified
6090 as immediate, e.g. for (i = 0x7ffffff0; i <= 0x7fffffff; i++) */
6092 final_value = plus_constant (final_value, off_by_one);
6094 /* Save the calculated values describing this loop's bounds, in case
6095 precondition_loop_p will need them later. These values can not be
6096 recalculated inside precondition_loop_p because strength reduction
6097 optimizations may obscure the loop's structure.
6099 These values are only required by precondition_loop_p and insert_bct
6100 whenever the number of iterations cannot be computed at compile time.
6101 Only the difference between final_value and initial_value is
6102 important. Note that final_value is only approximate. */
6103 loop_info->initial_value = initial_value;
6104 loop_info->comparison_value = comparison_value;
6105 loop_info->final_value = plus_constant (comparison_value, off_by_one);
6106 loop_info->increment = increment;
6107 loop_info->iteration_var = iteration_var;
6108 loop_info->comparison_code = comparison_code;
6111 /* Try to determine the iteration count for loops such
6112 as (for i = init; i < init + const; i++). When running the
6113 loop optimization twice, the first pass often converts simple
6114 loops into this form. */
6116 if (REG_P (initial_value))
6122 reg1 = initial_value;
6123 if (GET_CODE (final_value) == PLUS)
6124 reg2 = XEXP (final_value, 0), const2 = XEXP (final_value, 1);
6126 reg2 = final_value, const2 = const0_rtx;
6128 /* Check for initial_value = reg1, final_value = reg2 + const2,
6129 where reg1 != reg2. */
6130 if (REG_P (reg2) && reg2 != reg1)
6134 /* Find what reg1 is equivalent to. Hopefully it will
6135 either be reg2 or reg2 plus a constant. */
6136 temp = loop_find_equiv_value (loop, reg1);
6138 if (find_common_reg_term (temp, reg2))
6139 initial_value = temp;
6140 else if (loop_invariant_p (loop, reg2))
6142 /* Find what reg2 is equivalent to. Hopefully it will
6143 either be reg1 or reg1 plus a constant. Let's ignore
6144 the latter case for now since it is not so common. */
6145 temp = loop_find_equiv_value (loop, reg2);
6147 if (temp == loop_info->iteration_var)
6148 temp = initial_value;
6150 final_value = (const2 == const0_rtx)
6151 ? reg1 : gen_rtx_PLUS (GET_MODE (reg1), reg1, const2);
6156 loop_info->initial_equiv_value = initial_value;
6157 loop_info->final_equiv_value = final_value;
6159 /* For EQ comparison loops, we don't have a valid final value.
6160 Check this now so that we won't leave an invalid value if we
6161 return early for any other reason. */
6162 if (comparison_code == EQ)
6163 loop_info->final_equiv_value = loop_info->final_value = 0;
6167 if (loop_dump_stream)
6168 fprintf (loop_dump_stream,
6169 "Loop iterations: Increment value can't be calculated.\n");
6173 if (GET_CODE (increment) != CONST_INT)
6175 /* If we have a REG, check to see if REG holds a constant value. */
6176 /* ??? Other RTL, such as (neg (reg)) is possible here, but it isn't
6177 clear if it is worthwhile to try to handle such RTL. */
6178 if (REG_P (increment) || GET_CODE (increment) == SUBREG)
6179 increment = loop_find_equiv_value (loop, increment);
6181 if (GET_CODE (increment) != CONST_INT)
6183 if (loop_dump_stream)
6185 fprintf (loop_dump_stream,
6186 "Loop iterations: Increment value not constant ");
6187 print_simple_rtl (loop_dump_stream, increment);
6188 fprintf (loop_dump_stream, ".\n");
6192 loop_info->increment = increment;
6195 if (GET_CODE (initial_value) != CONST_INT)
6197 if (loop_dump_stream)
6199 fprintf (loop_dump_stream,
6200 "Loop iterations: Initial value not constant ");
6201 print_simple_rtl (loop_dump_stream, initial_value);
6202 fprintf (loop_dump_stream, ".\n");
6206 else if (GET_CODE (final_value) != CONST_INT)
6208 if (loop_dump_stream)
6210 fprintf (loop_dump_stream,
6211 "Loop iterations: Final value not constant ");
6212 print_simple_rtl (loop_dump_stream, final_value);
6213 fprintf (loop_dump_stream, ".\n");
6217 else if (comparison_code == EQ)
6221 if (loop_dump_stream)
6222 fprintf (loop_dump_stream, "Loop iterations: EQ comparison loop.\n");
6224 inc_once = gen_int_mode (INTVAL (initial_value) + INTVAL (increment),
6225 GET_MODE (iteration_var));
6227 if (inc_once == final_value)
6229 /* The iterator value once through the loop is equal to the
6230 comparison value. Either we have an infinite loop, or
6231 we'll loop twice. */
6232 if (increment == const0_rtx)
6234 loop_info->n_iterations = 2;
6237 loop_info->n_iterations = 1;
6239 if (GET_CODE (loop_info->initial_value) == CONST_INT)
6240 loop_info->final_value
6241 = gen_int_mode ((INTVAL (loop_info->initial_value)
6242 + loop_info->n_iterations * INTVAL (increment)),
6243 GET_MODE (iteration_var));
6245 loop_info->final_value
6246 = plus_constant (loop_info->initial_value,
6247 loop_info->n_iterations * INTVAL (increment));
6248 loop_info->final_equiv_value
6249 = gen_int_mode ((INTVAL (initial_value)
6250 + loop_info->n_iterations * INTVAL (increment)),
6251 GET_MODE (iteration_var));
6252 return loop_info->n_iterations;
6255 /* Final_larger is 1 if final larger, 0 if they are equal, otherwise -1. */
6258 = ((unsigned HOST_WIDE_INT) INTVAL (final_value)
6259 > (unsigned HOST_WIDE_INT) INTVAL (initial_value))
6260 - ((unsigned HOST_WIDE_INT) INTVAL (final_value)
6261 < (unsigned HOST_WIDE_INT) INTVAL (initial_value));
6263 final_larger = (INTVAL (final_value) > INTVAL (initial_value))
6264 - (INTVAL (final_value) < INTVAL (initial_value));
6266 if (INTVAL (increment) > 0)
6268 else if (INTVAL (increment) == 0)
6273 /* There are 27 different cases: compare_dir = -1, 0, 1;
6274 final_larger = -1, 0, 1; increment_dir = -1, 0, 1.
6275 There are 4 normal cases, 4 reverse cases (where the iteration variable
6276 will overflow before the loop exits), 4 infinite loop cases, and 15
6277 immediate exit (0 or 1 iteration depending on loop type) cases.
6278 Only try to optimize the normal cases. */
6280 /* (compare_dir/final_larger/increment_dir)
6281 Normal cases: (0/-1/-1), (0/1/1), (-1/-1/-1), (1/1/1)
6282 Reverse cases: (0/-1/1), (0/1/-1), (-1/-1/1), (1/1/-1)
6283 Infinite loops: (0/-1/0), (0/1/0), (-1/-1/0), (1/1/0)
6284 Immediate exit: (0/0/X), (-1/0/X), (-1/1/X), (1/0/X), (1/-1/X) */
6286 /* ?? If the meaning of reverse loops (where the iteration variable
6287 will overflow before the loop exits) is undefined, then could
6288 eliminate all of these special checks, and just always assume
6289 the loops are normal/immediate/infinite. Note that this means
6290 the sign of increment_dir does not have to be known. Also,
6291 since it does not really hurt if immediate exit loops or infinite loops
6292 are optimized, then that case could be ignored also, and hence all
6293 loops can be optimized.
6295 According to ANSI Spec, the reverse loop case result is undefined,
6296 because the action on overflow is undefined.
6298 See also the special test for NE loops below. */
6300 if (final_larger == increment_dir && final_larger != 0
6301 && (final_larger == compare_dir || compare_dir == 0))
6306 if (loop_dump_stream)
6307 fprintf (loop_dump_stream, "Loop iterations: Not normal loop.\n");
6311 /* Calculate the number of iterations, final_value is only an approximation,
6312 so correct for that. Note that abs_diff and n_iterations are
6313 unsigned, because they can be as large as 2^n - 1. */
6315 inc = INTVAL (increment);
6319 abs_diff = INTVAL (final_value) - INTVAL (initial_value);
6324 abs_diff = INTVAL (initial_value) - INTVAL (final_value);
6328 /* Given that iteration_var is going to iterate over its own mode,
6329 not HOST_WIDE_INT, disregard higher bits that might have come
6330 into the picture due to sign extension of initial and final
6332 abs_diff &= ((unsigned HOST_WIDE_INT) 1
6333 << (GET_MODE_BITSIZE (GET_MODE (iteration_var)) - 1)
6336 /* For NE tests, make sure that the iteration variable won't miss
6337 the final value. If abs_diff mod abs_incr is not zero, then the
6338 iteration variable will overflow before the loop exits, and we
6339 can not calculate the number of iterations. */
6340 if (compare_dir == 0 && (abs_diff % abs_inc) != 0)
6343 /* Note that the number of iterations could be calculated using
6344 (abs_diff + abs_inc - 1) / abs_inc, provided care was taken to
6345 handle potential overflow of the summation. */
6346 loop_info->n_iterations = abs_diff / abs_inc + ((abs_diff % abs_inc) != 0);
6347 return loop_info->n_iterations;
6350 /* Perform strength reduction and induction variable elimination.
6352 Pseudo registers created during this function will be beyond the
6353 last valid index in several tables including
6354 REGS->ARRAY[I].N_TIMES_SET and REGNO_LAST_UID. This does not cause a
6355 problem here, because the added registers cannot be givs outside of
6356 their loop, and hence will never be reconsidered. But scan_loop
6357 must check regnos to make sure they are in bounds. */
6360 strength_reduce (struct loop *loop, int flags)
6362 struct loop_info *loop_info = LOOP_INFO (loop);
6363 struct loop_regs *regs = LOOP_REGS (loop);
6364 struct loop_ivs *ivs = LOOP_IVS (loop);
6366 /* Temporary list pointer for traversing ivs->list. */
6367 struct iv_class *bl;
6368 /* Ratio of extra register life span we can justify
6369 for saving an instruction. More if loop doesn't call subroutines
6370 since in that case saving an insn makes more difference
6371 and more registers are available. */
6372 /* ??? could set this to last value of threshold in move_movables */
6373 int threshold = (loop_info->has_call ? 1 : 2) * (3 + n_non_fixed_regs);
6374 /* Map of pseudo-register replacements. */
6375 rtx *reg_map = NULL;
6377 rtx test_reg = gen_rtx_REG (word_mode, LAST_VIRTUAL_REGISTER + 1);
6378 int insn_count = count_insns_in_loop (loop);
6380 addr_placeholder = gen_reg_rtx (Pmode);
6382 ivs->n_regs = max_reg_before_loop;
6383 ivs->regs = xcalloc (ivs->n_regs, sizeof (struct iv));
6385 /* Find all BIVs in loop. */
6386 loop_bivs_find (loop);
6388 /* Exit if there are no bivs. */
6391 loop_ivs_free (loop);
6395 /* Determine how BIVS are initialized by looking through pre-header
6396 extended basic block. */
6397 loop_bivs_init_find (loop);
6399 /* Look at the each biv and see if we can say anything better about its
6400 initial value from any initializing insns set up above. */
6401 loop_bivs_check (loop);
6403 /* Search the loop for general induction variables. */
6404 loop_givs_find (loop);
6406 /* Try to calculate and save the number of loop iterations. This is
6407 set to zero if the actual number can not be calculated. This must
6408 be called after all giv's have been identified, since otherwise it may
6409 fail if the iteration variable is a giv. */
6410 loop_iterations (loop);
6412 #ifdef HAVE_prefetch
6413 if (flags & LOOP_PREFETCH)
6414 emit_prefetch_instructions (loop);
6417 /* Now for each giv for which we still don't know whether or not it is
6418 replaceable, check to see if it is replaceable because its final value
6419 can be calculated. This must be done after loop_iterations is called,
6420 so that final_giv_value will work correctly. */
6421 loop_givs_check (loop);
6423 /* Try to prove that the loop counter variable (if any) is always
6424 nonnegative; if so, record that fact with a REG_NONNEG note
6425 so that "decrement and branch until zero" insn can be used. */
6426 check_dbra_loop (loop, insn_count);
6428 /* Create reg_map to hold substitutions for replaceable giv regs.
6429 Some givs might have been made from biv increments, so look at
6430 ivs->reg_iv_type for a suitable size. */
6431 reg_map_size = ivs->n_regs;
6432 reg_map = xcalloc (reg_map_size, sizeof (rtx));
6434 /* Examine each iv class for feasibility of strength reduction/induction
6435 variable elimination. */
6437 for (bl = ivs->list; bl; bl = bl->next)
6439 struct induction *v;
6442 /* Test whether it will be possible to eliminate this biv
6443 provided all givs are reduced. */
6444 bl->eliminable = loop_biv_eliminable_p (loop, bl, threshold, insn_count);
6446 /* This will be true at the end, if all givs which depend on this
6447 biv have been strength reduced.
6448 We can't (currently) eliminate the biv unless this is so. */
6449 bl->all_reduced = 1;
6451 /* Check each extension dependent giv in this class to see if its
6452 root biv is safe from wrapping in the interior mode. */
6453 check_ext_dependent_givs (loop, bl);
6455 /* Combine all giv's for this iv_class. */
6456 combine_givs (regs, bl);
6458 for (v = bl->giv; v; v = v->next_iv)
6460 struct induction *tv;
6462 if (v->ignore || v->same)
6465 benefit = loop_giv_reduce_benefit (loop, bl, v, test_reg);
6467 /* If an insn is not to be strength reduced, then set its ignore
6468 flag, and clear bl->all_reduced. */
6470 /* A giv that depends on a reversed biv must be reduced if it is
6471 used after the loop exit, otherwise, it would have the wrong
6472 value after the loop exit. To make it simple, just reduce all
6473 of such giv's whether or not we know they are used after the loop
6476 if (v->lifetime * threshold * benefit < insn_count
6479 if (loop_dump_stream)
6480 fprintf (loop_dump_stream,
6481 "giv of insn %d not worth while, %d vs %d.\n",
6483 v->lifetime * threshold * benefit, insn_count);
6485 bl->all_reduced = 0;
6489 /* Check that we can increment the reduced giv without a
6490 multiply insn. If not, reject it. */
6492 for (tv = bl->biv; tv; tv = tv->next_iv)
6493 if (tv->mult_val == const1_rtx
6494 && ! product_cheap_p (tv->add_val, v->mult_val))
6496 if (loop_dump_stream)
6497 fprintf (loop_dump_stream,
6498 "giv of insn %d: would need a multiply.\n",
6499 INSN_UID (v->insn));
6501 bl->all_reduced = 0;
6507 /* Check for givs whose first use is their definition and whose
6508 last use is the definition of another giv. If so, it is likely
6509 dead and should not be used to derive another giv nor to
6511 loop_givs_dead_check (loop, bl);
6513 /* Reduce each giv that we decided to reduce. */
6514 loop_givs_reduce (loop, bl);
6516 /* Rescan all givs. If a giv is the same as a giv not reduced, mark it
6519 For each giv register that can be reduced now: if replaceable,
6520 substitute reduced reg wherever the old giv occurs;
6521 else add new move insn "giv_reg = reduced_reg". */
6522 loop_givs_rescan (loop, bl, reg_map);
6524 /* All the givs based on the biv bl have been reduced if they
6527 /* For each giv not marked as maybe dead that has been combined with a
6528 second giv, clear any "maybe dead" mark on that second giv.
6529 v->new_reg will either be or refer to the register of the giv it
6532 Doing this clearing avoids problems in biv elimination where
6533 a giv's new_reg is a complex value that can't be put in the
6534 insn but the giv combined with (with a reg as new_reg) is
6535 marked maybe_dead. Since the register will be used in either
6536 case, we'd prefer it be used from the simpler giv. */
6538 for (v = bl->giv; v; v = v->next_iv)
6539 if (! v->maybe_dead && v->same)
6540 v->same->maybe_dead = 0;
6542 /* Try to eliminate the biv, if it is a candidate.
6543 This won't work if ! bl->all_reduced,
6544 since the givs we planned to use might not have been reduced.
6546 We have to be careful that we didn't initially think we could
6547 eliminate this biv because of a giv that we now think may be
6548 dead and shouldn't be used as a biv replacement.
6550 Also, there is the possibility that we may have a giv that looks
6551 like it can be used to eliminate a biv, but the resulting insn
6552 isn't valid. This can happen, for example, on the 88k, where a
6553 JUMP_INSN can compare a register only with zero. Attempts to
6554 replace it with a compare with a constant will fail.
6556 Note that in cases where this call fails, we may have replaced some
6557 of the occurrences of the biv with a giv, but no harm was done in
6558 doing so in the rare cases where it can occur. */
6560 if (bl->all_reduced == 1 && bl->eliminable
6561 && maybe_eliminate_biv (loop, bl, 1, threshold, insn_count))
6563 /* ?? If we created a new test to bypass the loop entirely,
6564 or otherwise drop straight in, based on this test, then
6565 we might want to rewrite it also. This way some later
6566 pass has more hope of removing the initialization of this
6569 /* If final_value != 0, then the biv may be used after loop end
6570 and we must emit an insn to set it just in case.
6572 Reversed bivs already have an insn after the loop setting their
6573 value, so we don't need another one. We can't calculate the
6574 proper final value for such a biv here anyways. */
6575 if (bl->final_value && ! bl->reversed)
6576 loop_insn_sink_or_swim (loop,
6577 gen_load_of_final_value (bl->biv->dest_reg,
6580 if (loop_dump_stream)
6581 fprintf (loop_dump_stream, "Reg %d: biv eliminated\n",
6584 /* See above note wrt final_value. But since we couldn't eliminate
6585 the biv, we must set the value after the loop instead of before. */
6586 else if (bl->final_value && ! bl->reversed)
6587 loop_insn_sink (loop, gen_load_of_final_value (bl->biv->dest_reg,
6591 /* Go through all the instructions in the loop, making all the
6592 register substitutions scheduled in REG_MAP. */
6594 for (p = loop->start; p != loop->end; p = NEXT_INSN (p))
6597 replace_regs (PATTERN (p), reg_map, reg_map_size, 0);
6598 replace_regs (REG_NOTES (p), reg_map, reg_map_size, 0);
6602 if (loop_dump_stream)
6603 fprintf (loop_dump_stream, "\n");
6605 loop_ivs_free (loop);
6610 /*Record all basic induction variables calculated in the insn. */
6612 check_insn_for_bivs (struct loop *loop, rtx p, int not_every_iteration,
6615 struct loop_ivs *ivs = LOOP_IVS (loop);
6622 if (NONJUMP_INSN_P (p)
6623 && (set = single_set (p))
6624 && REG_P (SET_DEST (set)))
6626 dest_reg = SET_DEST (set);
6627 if (REGNO (dest_reg) < max_reg_before_loop
6628 && REGNO (dest_reg) >= FIRST_PSEUDO_REGISTER
6629 && REG_IV_TYPE (ivs, REGNO (dest_reg)) != NOT_BASIC_INDUCT)
6631 if (basic_induction_var (loop, SET_SRC (set),
6632 GET_MODE (SET_SRC (set)),
6633 dest_reg, p, &inc_val, &mult_val,
6636 /* It is a possible basic induction variable.
6637 Create and initialize an induction structure for it. */
6639 struct induction *v = xmalloc (sizeof (struct induction));
6641 record_biv (loop, v, p, dest_reg, inc_val, mult_val, location,
6642 not_every_iteration, maybe_multiple);
6643 REG_IV_TYPE (ivs, REGNO (dest_reg)) = BASIC_INDUCT;
6645 else if (REGNO (dest_reg) < ivs->n_regs)
6646 REG_IV_TYPE (ivs, REGNO (dest_reg)) = NOT_BASIC_INDUCT;
6652 /* Record all givs calculated in the insn.
6653 A register is a giv if: it is only set once, it is a function of a
6654 biv and a constant (or invariant), and it is not a biv. */
6656 check_insn_for_givs (struct loop *loop, rtx p, int not_every_iteration,
6659 struct loop_regs *regs = LOOP_REGS (loop);
6662 /* Look for a general induction variable in a register. */
6663 if (NONJUMP_INSN_P (p)
6664 && (set = single_set (p))
6665 && REG_P (SET_DEST (set))
6666 && ! regs->array[REGNO (SET_DEST (set))].may_not_optimize)
6675 rtx last_consec_insn;
6677 dest_reg = SET_DEST (set);
6678 if (REGNO (dest_reg) < FIRST_PSEUDO_REGISTER)
6681 if (/* SET_SRC is a giv. */
6682 (general_induction_var (loop, SET_SRC (set), &src_reg, &add_val,
6683 &mult_val, &ext_val, 0, &benefit, VOIDmode)
6684 /* Equivalent expression is a giv. */
6685 || ((regnote = find_reg_note (p, REG_EQUAL, NULL_RTX))
6686 && general_induction_var (loop, XEXP (regnote, 0), &src_reg,
6687 &add_val, &mult_val, &ext_val, 0,
6688 &benefit, VOIDmode)))
6689 /* Don't try to handle any regs made by loop optimization.
6690 We have nothing on them in regno_first_uid, etc. */
6691 && REGNO (dest_reg) < max_reg_before_loop
6692 /* Don't recognize a BASIC_INDUCT_VAR here. */
6693 && dest_reg != src_reg
6694 /* This must be the only place where the register is set. */
6695 && (regs->array[REGNO (dest_reg)].n_times_set == 1
6696 /* or all sets must be consecutive and make a giv. */
6697 || (benefit = consec_sets_giv (loop, benefit, p,
6699 &add_val, &mult_val, &ext_val,
6700 &last_consec_insn))))
6702 struct induction *v = xmalloc (sizeof (struct induction));
6704 /* If this is a library call, increase benefit. */
6705 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
6706 benefit += libcall_benefit (p);
6708 /* Skip the consecutive insns, if there are any. */
6709 if (regs->array[REGNO (dest_reg)].n_times_set != 1)
6710 p = last_consec_insn;
6712 record_giv (loop, v, p, src_reg, dest_reg, mult_val, add_val,
6713 ext_val, benefit, DEST_REG, not_every_iteration,
6714 maybe_multiple, (rtx*) 0);
6719 /* Look for givs which are memory addresses. */
6720 if (NONJUMP_INSN_P (p))
6721 find_mem_givs (loop, PATTERN (p), p, not_every_iteration,
6724 /* Update the status of whether giv can derive other givs. This can
6725 change when we pass a label or an insn that updates a biv. */
6726 if (INSN_P (p) || LABEL_P (p))
6727 update_giv_derive (loop, p);
6731 /* Return 1 if X is a valid source for an initial value (or as value being
6732 compared against in an initial test).
6734 X must be either a register or constant and must not be clobbered between
6735 the current insn and the start of the loop.
6737 INSN is the insn containing X. */
6740 valid_initial_value_p (rtx x, rtx insn, int call_seen, rtx loop_start)
6745 /* Only consider pseudos we know about initialized in insns whose luids
6748 || REGNO (x) >= max_reg_before_loop)
6751 /* Don't use call-clobbered registers across a call which clobbers it. On
6752 some machines, don't use any hard registers at all. */
6753 if (REGNO (x) < FIRST_PSEUDO_REGISTER
6754 && (SMALL_REGISTER_CLASSES
6755 || (call_seen && call_used_regs[REGNO (x)])))
6758 /* Don't use registers that have been clobbered before the start of the
6760 if (reg_set_between_p (x, insn, loop_start))
6766 /* Scan X for memory refs and check each memory address
6767 as a possible giv. INSN is the insn whose pattern X comes from.
6768 NOT_EVERY_ITERATION is 1 if the insn might not be executed during
6769 every loop iteration. MAYBE_MULTIPLE is 1 if the insn might be executed
6770 more than once in each loop iteration. */
6773 find_mem_givs (const struct loop *loop, rtx x, rtx insn,
6774 int not_every_iteration, int maybe_multiple)
6783 code = GET_CODE (x);
6808 /* This code used to disable creating GIVs with mult_val == 1 and
6809 add_val == 0. However, this leads to lost optimizations when
6810 it comes time to combine a set of related DEST_ADDR GIVs, since
6811 this one would not be seen. */
6813 if (general_induction_var (loop, XEXP (x, 0), &src_reg, &add_val,
6814 &mult_val, &ext_val, 1, &benefit,
6817 /* Found one; record it. */
6818 struct induction *v = xmalloc (sizeof (struct induction));
6820 record_giv (loop, v, insn, src_reg, addr_placeholder, mult_val,
6821 add_val, ext_val, benefit, DEST_ADDR,
6822 not_every_iteration, maybe_multiple, &XEXP (x, 0));
6833 /* Recursively scan the subexpressions for other mem refs. */
6835 fmt = GET_RTX_FORMAT (code);
6836 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
6838 find_mem_givs (loop, XEXP (x, i), insn, not_every_iteration,
6840 else if (fmt[i] == 'E')
6841 for (j = 0; j < XVECLEN (x, i); j++)
6842 find_mem_givs (loop, XVECEXP (x, i, j), insn, not_every_iteration,
6846 /* Fill in the data about one biv update.
6847 V is the `struct induction' in which we record the biv. (It is
6848 allocated by the caller, with alloca.)
6849 INSN is the insn that sets it.
6850 DEST_REG is the biv's reg.
6852 MULT_VAL is const1_rtx if the biv is being incremented here, in which case
6853 INC_VAL is the increment. Otherwise, MULT_VAL is const0_rtx and the biv is
6854 being set to INC_VAL.
6856 NOT_EVERY_ITERATION is nonzero if this biv update is not know to be
6857 executed every iteration; MAYBE_MULTIPLE is nonzero if this biv update
6858 can be executed more than once per iteration. If MAYBE_MULTIPLE
6859 and NOT_EVERY_ITERATION are both zero, we know that the biv update is
6860 executed exactly once per iteration. */
6863 record_biv (struct loop *loop, struct induction *v, rtx insn, rtx dest_reg,
6864 rtx inc_val, rtx mult_val, rtx *location,
6865 int not_every_iteration, int maybe_multiple)
6867 struct loop_ivs *ivs = LOOP_IVS (loop);
6868 struct iv_class *bl;
6871 v->src_reg = dest_reg;
6872 v->dest_reg = dest_reg;
6873 v->mult_val = mult_val;
6874 v->add_val = inc_val;
6875 v->ext_dependent = NULL_RTX;
6876 v->location = location;
6877 v->mode = GET_MODE (dest_reg);
6878 v->always_computable = ! not_every_iteration;
6879 v->always_executed = ! not_every_iteration;
6880 v->maybe_multiple = maybe_multiple;
6883 /* Add this to the reg's iv_class, creating a class
6884 if this is the first incrementation of the reg. */
6886 bl = REG_IV_CLASS (ivs, REGNO (dest_reg));
6889 /* Create and initialize new iv_class. */
6891 bl = xmalloc (sizeof (struct iv_class));
6893 bl->regno = REGNO (dest_reg);
6899 /* Set initial value to the reg itself. */
6900 bl->initial_value = dest_reg;
6901 bl->final_value = 0;
6902 /* We haven't seen the initializing insn yet. */
6905 bl->initial_test = 0;
6906 bl->incremented = 0;
6910 bl->total_benefit = 0;
6912 /* Add this class to ivs->list. */
6913 bl->next = ivs->list;
6916 /* Put it in the array of biv register classes. */
6917 REG_IV_CLASS (ivs, REGNO (dest_reg)) = bl;
6921 /* Check if location is the same as a previous one. */
6922 struct induction *induction;
6923 for (induction = bl->biv; induction; induction = induction->next_iv)
6924 if (location == induction->location)
6926 v->same = induction;
6931 /* Update IV_CLASS entry for this biv. */
6932 v->next_iv = bl->biv;
6935 if (mult_val == const1_rtx)
6936 bl->incremented = 1;
6938 if (loop_dump_stream)
6939 loop_biv_dump (v, loop_dump_stream, 0);
6942 /* Fill in the data about one giv.
6943 V is the `struct induction' in which we record the giv. (It is
6944 allocated by the caller, with alloca.)
6945 INSN is the insn that sets it.
6946 BENEFIT estimates the savings from deleting this insn.
6947 TYPE is DEST_REG or DEST_ADDR; it says whether the giv is computed
6948 into a register or is used as a memory address.
6950 SRC_REG is the biv reg which the giv is computed from.
6951 DEST_REG is the giv's reg (if the giv is stored in a reg).
6952 MULT_VAL and ADD_VAL are the coefficients used to compute the giv.
6953 LOCATION points to the place where this giv's value appears in INSN. */
6956 record_giv (const struct loop *loop, struct induction *v, rtx insn,
6957 rtx src_reg, rtx dest_reg, rtx mult_val, rtx add_val,
6958 rtx ext_val, int benefit, enum g_types type,
6959 int not_every_iteration, int maybe_multiple, rtx *location)
6961 struct loop_ivs *ivs = LOOP_IVS (loop);
6962 struct induction *b;
6963 struct iv_class *bl;
6964 rtx set = single_set (insn);
6967 /* Attempt to prove constantness of the values. Don't let simplify_rtx
6968 undo the MULT canonicalization that we performed earlier. */
6969 temp = simplify_rtx (add_val);
6971 && ! (GET_CODE (add_val) == MULT
6972 && GET_CODE (temp) == ASHIFT))
6976 v->src_reg = src_reg;
6978 v->dest_reg = dest_reg;
6979 v->mult_val = mult_val;
6980 v->add_val = add_val;
6981 v->ext_dependent = ext_val;
6982 v->benefit = benefit;
6983 v->location = location;
6985 v->combined_with = 0;
6986 v->maybe_multiple = maybe_multiple;
6988 v->derive_adjustment = 0;
6994 v->auto_inc_opt = 0;
6997 /* The v->always_computable field is used in update_giv_derive, to
6998 determine whether a giv can be used to derive another giv. For a
6999 DEST_REG giv, INSN computes a new value for the giv, so its value
7000 isn't computable if INSN insn't executed every iteration.
7001 However, for a DEST_ADDR giv, INSN merely uses the value of the giv;
7002 it does not compute a new value. Hence the value is always computable
7003 regardless of whether INSN is executed each iteration. */
7005 if (type == DEST_ADDR)
7006 v->always_computable = 1;
7008 v->always_computable = ! not_every_iteration;
7010 v->always_executed = ! not_every_iteration;
7012 if (type == DEST_ADDR)
7014 v->mode = GET_MODE (*location);
7017 else /* type == DEST_REG */
7019 v->mode = GET_MODE (SET_DEST (set));
7021 v->lifetime = LOOP_REG_LIFETIME (loop, REGNO (dest_reg));
7023 /* If the lifetime is zero, it means that this register is
7024 really a dead store. So mark this as a giv that can be
7025 ignored. This will not prevent the biv from being eliminated. */
7026 if (v->lifetime == 0)
7029 REG_IV_TYPE (ivs, REGNO (dest_reg)) = GENERAL_INDUCT;
7030 REG_IV_INFO (ivs, REGNO (dest_reg)) = v;
7033 /* Add the giv to the class of givs computed from one biv. */
7035 bl = REG_IV_CLASS (ivs, REGNO (src_reg));
7037 v->next_iv = bl->giv;
7040 /* Don't count DEST_ADDR. This is supposed to count the number of
7041 insns that calculate givs. */
7042 if (type == DEST_REG)
7044 bl->total_benefit += benefit;
7046 if (type == DEST_ADDR)
7049 v->not_replaceable = 0;
7053 /* The giv can be replaced outright by the reduced register only if all
7054 of the following conditions are true:
7055 - the insn that sets the giv is always executed on any iteration
7056 on which the giv is used at all
7057 (there are two ways to deduce this:
7058 either the insn is executed on every iteration,
7059 or all uses follow that insn in the same basic block),
7060 - the giv is not used outside the loop
7061 - no assignments to the biv occur during the giv's lifetime. */
7063 if (REGNO_FIRST_UID (REGNO (dest_reg)) == INSN_UID (insn)
7064 /* Previous line always fails if INSN was moved by loop opt. */
7065 && REGNO_LAST_LUID (REGNO (dest_reg))
7066 < INSN_LUID (loop->end)
7067 && (! not_every_iteration
7068 || last_use_this_basic_block (dest_reg, insn)))
7070 /* Now check that there are no assignments to the biv within the
7071 giv's lifetime. This requires two separate checks. */
7073 /* Check each biv update, and fail if any are between the first
7074 and last use of the giv.
7076 If this loop contains an inner loop that was unrolled, then
7077 the insn modifying the biv may have been emitted by the loop
7078 unrolling code, and hence does not have a valid luid. Just
7079 mark the biv as not replaceable in this case. It is not very
7080 useful as a biv, because it is used in two different loops.
7081 It is very unlikely that we would be able to optimize the giv
7082 using this biv anyways. */
7085 v->not_replaceable = 0;
7086 for (b = bl->biv; b; b = b->next_iv)
7088 if (INSN_UID (b->insn) >= max_uid_for_loop
7089 || ((INSN_LUID (b->insn)
7090 >= REGNO_FIRST_LUID (REGNO (dest_reg)))
7091 && (INSN_LUID (b->insn)
7092 <= REGNO_LAST_LUID (REGNO (dest_reg)))))
7095 v->not_replaceable = 1;
7100 /* If there are any backwards branches that go from after the
7101 biv update to before it, then this giv is not replaceable. */
7103 for (b = bl->biv; b; b = b->next_iv)
7104 if (back_branch_in_range_p (loop, b->insn))
7107 v->not_replaceable = 1;
7113 /* May still be replaceable, we don't have enough info here to
7116 v->not_replaceable = 0;
7120 /* Record whether the add_val contains a const_int, for later use by
7125 v->no_const_addval = 1;
7126 if (tem == const0_rtx)
7128 else if (CONSTANT_P (add_val))
7129 v->no_const_addval = 0;
7130 if (GET_CODE (tem) == PLUS)
7134 if (GET_CODE (XEXP (tem, 0)) == PLUS)
7135 tem = XEXP (tem, 0);
7136 else if (GET_CODE (XEXP (tem, 1)) == PLUS)
7137 tem = XEXP (tem, 1);
7141 if (CONSTANT_P (XEXP (tem, 1)))
7142 v->no_const_addval = 0;
7146 if (loop_dump_stream)
7147 loop_giv_dump (v, loop_dump_stream, 0);
7150 /* Try to calculate the final value of the giv, the value it will have at
7151 the end of the loop. If we can do it, return that value. */
7154 final_giv_value (const struct loop *loop, struct induction *v)
7156 struct loop_ivs *ivs = LOOP_IVS (loop);
7157 struct iv_class *bl;
7161 rtx loop_end = loop->end;
7162 unsigned HOST_WIDE_INT n_iterations = LOOP_INFO (loop)->n_iterations;
7164 bl = REG_IV_CLASS (ivs, REGNO (v->src_reg));
7166 /* The final value for givs which depend on reversed bivs must be calculated
7167 differently than for ordinary givs. In this case, there is already an
7168 insn after the loop which sets this giv's final value (if necessary),
7169 and there are no other loop exits, so we can return any value. */
7172 if (loop_dump_stream)
7173 fprintf (loop_dump_stream,
7174 "Final giv value for %d, depends on reversed biv\n",
7175 REGNO (v->dest_reg));
7179 /* Try to calculate the final value as a function of the biv it depends
7180 upon. The only exit from the loop must be the fall through at the bottom
7181 and the insn that sets the giv must be executed on every iteration
7182 (otherwise the giv may not have its final value when the loop exits). */
7184 /* ??? Can calculate the final giv value by subtracting off the
7185 extra biv increments times the giv's mult_val. The loop must have
7186 only one exit for this to work, but the loop iterations does not need
7189 if (n_iterations != 0
7190 && ! loop->exit_count
7191 && v->always_executed)
7193 /* ?? It is tempting to use the biv's value here since these insns will
7194 be put after the loop, and hence the biv will have its final value
7195 then. However, this fails if the biv is subsequently eliminated.
7196 Perhaps determine whether biv's are eliminable before trying to
7197 determine whether giv's are replaceable so that we can use the
7198 biv value here if it is not eliminable. */
7200 /* We are emitting code after the end of the loop, so we must make
7201 sure that bl->initial_value is still valid then. It will still
7202 be valid if it is invariant. */
7204 increment = biv_total_increment (bl);
7206 if (increment && loop_invariant_p (loop, increment)
7207 && loop_invariant_p (loop, bl->initial_value))
7209 /* Can calculate the loop exit value of its biv as
7210 (n_iterations * increment) + initial_value */
7212 /* The loop exit value of the giv is then
7213 (final_biv_value - extra increments) * mult_val + add_val.
7214 The extra increments are any increments to the biv which
7215 occur in the loop after the giv's value is calculated.
7216 We must search from the insn that sets the giv to the end
7217 of the loop to calculate this value. */
7219 /* Put the final biv value in tem. */
7220 tem = gen_reg_rtx (v->mode);
7221 record_base_value (REGNO (tem), bl->biv->add_val, 0);
7222 loop_iv_add_mult_sink (loop, extend_value_for_giv (v, increment),
7223 GEN_INT (n_iterations),
7224 extend_value_for_giv (v, bl->initial_value),
7227 /* Subtract off extra increments as we find them. */
7228 for (insn = NEXT_INSN (v->insn); insn != loop_end;
7229 insn = NEXT_INSN (insn))
7231 struct induction *biv;
7233 for (biv = bl->biv; biv; biv = biv->next_iv)
7234 if (biv->insn == insn)
7237 tem = expand_simple_binop (GET_MODE (tem), MINUS, tem,
7238 biv->add_val, NULL_RTX, 0,
7242 loop_insn_sink (loop, seq);
7246 /* Now calculate the giv's final value. */
7247 loop_iv_add_mult_sink (loop, tem, v->mult_val, v->add_val, tem);
7249 if (loop_dump_stream)
7250 fprintf (loop_dump_stream,
7251 "Final giv value for %d, calc from biv's value.\n",
7252 REGNO (v->dest_reg));
7258 /* Replaceable giv's should never reach here. */
7259 gcc_assert (!v->replaceable);
7261 /* Check to see if the biv is dead at all loop exits. */
7262 if (reg_dead_after_loop (loop, v->dest_reg))
7264 if (loop_dump_stream)
7265 fprintf (loop_dump_stream,
7266 "Final giv value for %d, giv dead after loop exit.\n",
7267 REGNO (v->dest_reg));
7275 /* All this does is determine whether a giv can be made replaceable because
7276 its final value can be calculated. This code can not be part of record_giv
7277 above, because final_giv_value requires that the number of loop iterations
7278 be known, and that can not be accurately calculated until after all givs
7279 have been identified. */
7282 check_final_value (const struct loop *loop, struct induction *v)
7284 rtx final_value = 0;
7286 /* DEST_ADDR givs will never reach here, because they are always marked
7287 replaceable above in record_giv. */
7289 /* The giv can be replaced outright by the reduced register only if all
7290 of the following conditions are true:
7291 - the insn that sets the giv is always executed on any iteration
7292 on which the giv is used at all
7293 (there are two ways to deduce this:
7294 either the insn is executed on every iteration,
7295 or all uses follow that insn in the same basic block),
7296 - its final value can be calculated (this condition is different
7297 than the one above in record_giv)
7298 - it's not used before the it's set
7299 - no assignments to the biv occur during the giv's lifetime. */
7302 /* This is only called now when replaceable is known to be false. */
7303 /* Clear replaceable, so that it won't confuse final_giv_value. */
7307 if ((final_value = final_giv_value (loop, v))
7308 && (v->always_executed
7309 || last_use_this_basic_block (v->dest_reg, v->insn)))
7311 int biv_increment_seen = 0, before_giv_insn = 0;
7316 v->not_replaceable = 0;
7318 /* When trying to determine whether or not a biv increment occurs
7319 during the lifetime of the giv, we can ignore uses of the variable
7320 outside the loop because final_value is true. Hence we can not
7321 use regno_last_uid and regno_first_uid as above in record_giv. */
7323 /* Search the loop to determine whether any assignments to the
7324 biv occur during the giv's lifetime. Start with the insn
7325 that sets the giv, and search around the loop until we come
7326 back to that insn again.
7328 Also fail if there is a jump within the giv's lifetime that jumps
7329 to somewhere outside the lifetime but still within the loop. This
7330 catches spaghetti code where the execution order is not linear, and
7331 hence the above test fails. Here we assume that the giv lifetime
7332 does not extend from one iteration of the loop to the next, so as
7333 to make the test easier. Since the lifetime isn't known yet,
7334 this requires two loops. See also record_giv above. */
7336 last_giv_use = v->insn;
7343 before_giv_insn = 1;
7344 p = NEXT_INSN (loop->start);
7351 /* It is possible for the BIV increment to use the GIV if we
7352 have a cycle. Thus we must be sure to check each insn for
7353 both BIV and GIV uses, and we must check for BIV uses
7356 if (! biv_increment_seen
7357 && reg_set_p (v->src_reg, PATTERN (p)))
7358 biv_increment_seen = 1;
7360 if (reg_mentioned_p (v->dest_reg, PATTERN (p)))
7362 if (biv_increment_seen || before_giv_insn)
7365 v->not_replaceable = 1;
7373 /* Now that the lifetime of the giv is known, check for branches
7374 from within the lifetime to outside the lifetime if it is still
7384 p = NEXT_INSN (loop->start);
7385 if (p == last_giv_use)
7388 if (JUMP_P (p) && JUMP_LABEL (p)
7389 && LABEL_NAME (JUMP_LABEL (p))
7390 && ((loop_insn_first_p (JUMP_LABEL (p), v->insn)
7391 && loop_insn_first_p (loop->start, JUMP_LABEL (p)))
7392 || (loop_insn_first_p (last_giv_use, JUMP_LABEL (p))
7393 && loop_insn_first_p (JUMP_LABEL (p), loop->end))))
7396 v->not_replaceable = 1;
7398 if (loop_dump_stream)
7399 fprintf (loop_dump_stream,
7400 "Found branch outside giv lifetime.\n");
7407 /* If it is replaceable, then save the final value. */
7409 v->final_value = final_value;
7412 if (loop_dump_stream && v->replaceable)
7413 fprintf (loop_dump_stream, "Insn %d: giv reg %d final_value replaceable\n",
7414 INSN_UID (v->insn), REGNO (v->dest_reg));
7417 /* Update the status of whether a giv can derive other givs.
7419 We need to do something special if there is or may be an update to the biv
7420 between the time the giv is defined and the time it is used to derive
7423 In addition, a giv that is only conditionally set is not allowed to
7424 derive another giv once a label has been passed.
7426 The cases we look at are when a label or an update to a biv is passed. */
7429 update_giv_derive (const struct loop *loop, rtx p)
7431 struct loop_ivs *ivs = LOOP_IVS (loop);
7432 struct iv_class *bl;
7433 struct induction *biv, *giv;
7437 /* Search all IV classes, then all bivs, and finally all givs.
7439 There are three cases we are concerned with. First we have the situation
7440 of a giv that is only updated conditionally. In that case, it may not
7441 derive any givs after a label is passed.
7443 The second case is when a biv update occurs, or may occur, after the
7444 definition of a giv. For certain biv updates (see below) that are
7445 known to occur between the giv definition and use, we can adjust the
7446 giv definition. For others, or when the biv update is conditional,
7447 we must prevent the giv from deriving any other givs. There are two
7448 sub-cases within this case.
7450 If this is a label, we are concerned with any biv update that is done
7451 conditionally, since it may be done after the giv is defined followed by
7452 a branch here (actually, we need to pass both a jump and a label, but
7453 this extra tracking doesn't seem worth it).
7455 If this is a jump, we are concerned about any biv update that may be
7456 executed multiple times. We are actually only concerned about
7457 backward jumps, but it is probably not worth performing the test
7458 on the jump again here.
7460 If this is a biv update, we must adjust the giv status to show that a
7461 subsequent biv update was performed. If this adjustment cannot be done,
7462 the giv cannot derive further givs. */
7464 for (bl = ivs->list; bl; bl = bl->next)
7465 for (biv = bl->biv; biv; biv = biv->next_iv)
7466 if (LABEL_P (p) || JUMP_P (p)
7469 /* Skip if location is the same as a previous one. */
7473 for (giv = bl->giv; giv; giv = giv->next_iv)
7475 /* If cant_derive is already true, there is no point in
7476 checking all of these conditions again. */
7477 if (giv->cant_derive)
7480 /* If this giv is conditionally set and we have passed a label,
7481 it cannot derive anything. */
7482 if (LABEL_P (p) && ! giv->always_computable)
7483 giv->cant_derive = 1;
7485 /* Skip givs that have mult_val == 0, since
7486 they are really invariants. Also skip those that are
7487 replaceable, since we know their lifetime doesn't contain
7489 else if (giv->mult_val == const0_rtx || giv->replaceable)
7492 /* The only way we can allow this giv to derive another
7493 is if this is a biv increment and we can form the product
7494 of biv->add_val and giv->mult_val. In this case, we will
7495 be able to compute a compensation. */
7496 else if (biv->insn == p)
7501 if (biv->mult_val == const1_rtx)
7502 tem = simplify_giv_expr (loop,
7503 gen_rtx_MULT (giv->mode,
7506 &ext_val_dummy, &dummy);
7508 if (tem && giv->derive_adjustment)
7509 tem = simplify_giv_expr
7511 gen_rtx_PLUS (giv->mode, tem, giv->derive_adjustment),
7512 &ext_val_dummy, &dummy);
7515 giv->derive_adjustment = tem;
7517 giv->cant_derive = 1;
7519 else if ((LABEL_P (p) && ! biv->always_computable)
7520 || (JUMP_P (p) && biv->maybe_multiple))
7521 giv->cant_derive = 1;
7526 /* Check whether an insn is an increment legitimate for a basic induction var.
7527 X is the source of insn P, or a part of it.
7528 MODE is the mode in which X should be interpreted.
7530 DEST_REG is the putative biv, also the destination of the insn.
7531 We accept patterns of these forms:
7532 REG = REG + INVARIANT (includes REG = REG - CONSTANT)
7533 REG = INVARIANT + REG
7535 If X is suitable, we return 1, set *MULT_VAL to CONST1_RTX,
7536 store the additive term into *INC_VAL, and store the place where
7537 we found the additive term into *LOCATION.
7539 If X is an assignment of an invariant into DEST_REG, we set
7540 *MULT_VAL to CONST0_RTX, and store the invariant into *INC_VAL.
7542 We also want to detect a BIV when it corresponds to a variable
7543 whose mode was promoted. In that case, an increment
7544 of the variable may be a PLUS that adds a SUBREG of that variable to
7545 an invariant and then sign- or zero-extends the result of the PLUS
7548 Most GIVs in such cases will be in the promoted mode, since that is the
7549 probably the natural computation mode (and almost certainly the mode
7550 used for addresses) on the machine. So we view the pseudo-reg containing
7551 the variable as the BIV, as if it were simply incremented.
7553 Note that treating the entire pseudo as a BIV will result in making
7554 simple increments to any GIVs based on it. However, if the variable
7555 overflows in its declared mode but not its promoted mode, the result will
7556 be incorrect. This is acceptable if the variable is signed, since
7557 overflows in such cases are undefined, but not if it is unsigned, since
7558 those overflows are defined. So we only check for SIGN_EXTEND and
7561 If we cannot find a biv, we return 0. */
7564 basic_induction_var (const struct loop *loop, rtx x, enum machine_mode mode,
7565 rtx dest_reg, rtx p, rtx *inc_val, rtx *mult_val,
7570 rtx insn, set = 0, last, inc;
7572 code = GET_CODE (x);
7577 if (rtx_equal_p (XEXP (x, 0), dest_reg)
7578 || (GET_CODE (XEXP (x, 0)) == SUBREG
7579 && SUBREG_PROMOTED_VAR_P (XEXP (x, 0))
7580 && SUBREG_REG (XEXP (x, 0)) == dest_reg))
7582 argp = &XEXP (x, 1);
7584 else if (rtx_equal_p (XEXP (x, 1), dest_reg)
7585 || (GET_CODE (XEXP (x, 1)) == SUBREG
7586 && SUBREG_PROMOTED_VAR_P (XEXP (x, 1))
7587 && SUBREG_REG (XEXP (x, 1)) == dest_reg))
7589 argp = &XEXP (x, 0);
7595 if (loop_invariant_p (loop, arg) != 1)
7598 /* convert_modes can emit new instructions, e.g. when arg is a loop
7599 invariant MEM and dest_reg has a different mode.
7600 These instructions would be emitted after the end of the function
7601 and then *inc_val would be an uninitialized pseudo.
7602 Detect this and bail in this case.
7603 Other alternatives to solve this can be introducing a convert_modes
7604 variant which is allowed to fail but not allowed to emit new
7605 instructions, emit these instructions before loop start and let
7606 it be garbage collected if *inc_val is never used or saving the
7607 *inc_val initialization sequence generated here and when *inc_val
7608 is going to be actually used, emit it at some suitable place. */
7609 last = get_last_insn ();
7610 inc = convert_modes (GET_MODE (dest_reg), GET_MODE (x), arg, 0);
7611 if (get_last_insn () != last)
7613 delete_insns_since (last);
7618 *mult_val = const1_rtx;
7623 /* If what's inside the SUBREG is a BIV, then the SUBREG. This will
7624 handle addition of promoted variables.
7625 ??? The comment at the start of this function is wrong: promoted
7626 variable increments don't look like it says they do. */
7627 return basic_induction_var (loop, SUBREG_REG (x),
7628 GET_MODE (SUBREG_REG (x)),
7629 dest_reg, p, inc_val, mult_val, location);
7632 /* If this register is assigned in a previous insn, look at its
7633 source, but don't go outside the loop or past a label. */
7635 /* If this sets a register to itself, we would repeat any previous
7636 biv increment if we applied this strategy blindly. */
7637 if (rtx_equal_p (dest_reg, x))
7646 insn = PREV_INSN (insn);
7648 while (insn && NOTE_P (insn)
7649 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_BEG);
7653 set = single_set (insn);
7656 dest = SET_DEST (set);
7658 || (GET_CODE (dest) == SUBREG
7659 && (GET_MODE_SIZE (GET_MODE (dest)) <= UNITS_PER_WORD)
7660 && (GET_MODE_CLASS (GET_MODE (dest)) == MODE_INT)
7661 && SUBREG_REG (dest) == x))
7662 return basic_induction_var (loop, SET_SRC (set),
7663 (GET_MODE (SET_SRC (set)) == VOIDmode
7665 : GET_MODE (SET_SRC (set))),
7667 inc_val, mult_val, location);
7669 while (GET_CODE (dest) == SUBREG
7670 || GET_CODE (dest) == ZERO_EXTRACT
7671 || GET_CODE (dest) == STRICT_LOW_PART)
7672 dest = XEXP (dest, 0);
7678 /* Can accept constant setting of biv only when inside inner most loop.
7679 Otherwise, a biv of an inner loop may be incorrectly recognized
7680 as a biv of the outer loop,
7681 causing code to be moved INTO the inner loop. */
7683 if (loop_invariant_p (loop, x) != 1)
7688 /* convert_modes dies if we try to convert to or from CCmode, so just
7689 exclude that case. It is very unlikely that a condition code value
7690 would be a useful iterator anyways. convert_modes dies if we try to
7691 convert a float mode to non-float or vice versa too. */
7692 if (loop->level == 1
7693 && GET_MODE_CLASS (mode) == GET_MODE_CLASS (GET_MODE (dest_reg))
7694 && GET_MODE_CLASS (mode) != MODE_CC)
7696 /* Possible bug here? Perhaps we don't know the mode of X. */
7697 last = get_last_insn ();
7698 inc = convert_modes (GET_MODE (dest_reg), mode, x, 0);
7699 if (get_last_insn () != last)
7701 delete_insns_since (last);
7706 *mult_val = const0_rtx;
7713 /* Ignore this BIV if signed arithmetic overflow is defined. */
7716 return basic_induction_var (loop, XEXP (x, 0), GET_MODE (XEXP (x, 0)),
7717 dest_reg, p, inc_val, mult_val, location);
7720 /* Similar, since this can be a sign extension. */
7721 for (insn = PREV_INSN (p);
7722 (insn && NOTE_P (insn)
7723 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_BEG);
7724 insn = PREV_INSN (insn))
7728 set = single_set (insn);
7730 if (! rtx_equal_p (dest_reg, XEXP (x, 0))
7731 && set && SET_DEST (set) == XEXP (x, 0)
7732 && GET_CODE (XEXP (x, 1)) == CONST_INT
7733 && INTVAL (XEXP (x, 1)) >= 0
7734 && GET_CODE (SET_SRC (set)) == ASHIFT
7735 && XEXP (x, 1) == XEXP (SET_SRC (set), 1))
7736 return basic_induction_var (loop, XEXP (SET_SRC (set), 0),
7737 GET_MODE (XEXP (x, 0)),
7738 dest_reg, insn, inc_val, mult_val,
7747 /* A general induction variable (giv) is any quantity that is a linear
7748 function of a basic induction variable,
7749 i.e. giv = biv * mult_val + add_val.
7750 The coefficients can be any loop invariant quantity.
7751 A giv need not be computed directly from the biv;
7752 it can be computed by way of other givs. */
7754 /* Determine whether X computes a giv.
7755 If it does, return a nonzero value
7756 which is the benefit from eliminating the computation of X;
7757 set *SRC_REG to the register of the biv that it is computed from;
7758 set *ADD_VAL and *MULT_VAL to the coefficients,
7759 such that the value of X is biv * mult + add; */
7762 general_induction_var (const struct loop *loop, rtx x, rtx *src_reg,
7763 rtx *add_val, rtx *mult_val, rtx *ext_val,
7764 int is_addr, int *pbenefit,
7765 enum machine_mode addr_mode)
7767 struct loop_ivs *ivs = LOOP_IVS (loop);
7770 /* If this is an invariant, forget it, it isn't a giv. */
7771 if (loop_invariant_p (loop, x) == 1)
7775 *ext_val = NULL_RTX;
7776 x = simplify_giv_expr (loop, x, ext_val, pbenefit);
7780 switch (GET_CODE (x))
7784 /* Since this is now an invariant and wasn't before, it must be a giv
7785 with MULT_VAL == 0. It doesn't matter which BIV we associate this
7787 *src_reg = ivs->list->biv->dest_reg;
7788 *mult_val = const0_rtx;
7793 /* This is equivalent to a BIV. */
7795 *mult_val = const1_rtx;
7796 *add_val = const0_rtx;
7800 /* Either (plus (biv) (invar)) or
7801 (plus (mult (biv) (invar_1)) (invar_2)). */
7802 if (GET_CODE (XEXP (x, 0)) == MULT)
7804 *src_reg = XEXP (XEXP (x, 0), 0);
7805 *mult_val = XEXP (XEXP (x, 0), 1);
7809 *src_reg = XEXP (x, 0);
7810 *mult_val = const1_rtx;
7812 *add_val = XEXP (x, 1);
7816 /* ADD_VAL is zero. */
7817 *src_reg = XEXP (x, 0);
7818 *mult_val = XEXP (x, 1);
7819 *add_val = const0_rtx;
7826 /* Remove any enclosing USE from ADD_VAL and MULT_VAL (there will be
7827 unless they are CONST_INT). */
7828 if (GET_CODE (*add_val) == USE)
7829 *add_val = XEXP (*add_val, 0);
7830 if (GET_CODE (*mult_val) == USE)
7831 *mult_val = XEXP (*mult_val, 0);
7834 *pbenefit += address_cost (orig_x, addr_mode) - reg_address_cost;
7836 *pbenefit += rtx_cost (orig_x, SET);
7838 /* Always return true if this is a giv so it will be detected as such,
7839 even if the benefit is zero or negative. This allows elimination
7840 of bivs that might otherwise not be eliminated. */
7844 /* Given an expression, X, try to form it as a linear function of a biv.
7845 We will canonicalize it to be of the form
7846 (plus (mult (BIV) (invar_1))
7848 with possible degeneracies.
7850 The invariant expressions must each be of a form that can be used as a
7851 machine operand. We surround then with a USE rtx (a hack, but localized
7852 and certainly unambiguous!) if not a CONST_INT for simplicity in this
7853 routine; it is the caller's responsibility to strip them.
7855 If no such canonicalization is possible (i.e., two biv's are used or an
7856 expression that is neither invariant nor a biv or giv), this routine
7859 For a nonzero return, the result will have a code of CONST_INT, USE,
7860 REG (for a BIV), PLUS, or MULT. No other codes will occur.
7862 *BENEFIT will be incremented by the benefit of any sub-giv encountered. */
7864 static rtx sge_plus (enum machine_mode, rtx, rtx);
7865 static rtx sge_plus_constant (rtx, rtx);
7868 simplify_giv_expr (const struct loop *loop, rtx x, rtx *ext_val, int *benefit)
7870 struct loop_ivs *ivs = LOOP_IVS (loop);
7871 struct loop_regs *regs = LOOP_REGS (loop);
7872 enum machine_mode mode = GET_MODE (x);
7876 /* If this is not an integer mode, or if we cannot do arithmetic in this
7877 mode, this can't be a giv. */
7878 if (mode != VOIDmode
7879 && (GET_MODE_CLASS (mode) != MODE_INT
7880 || GET_MODE_BITSIZE (mode) > HOST_BITS_PER_WIDE_INT))
7883 switch (GET_CODE (x))
7886 arg0 = simplify_giv_expr (loop, XEXP (x, 0), ext_val, benefit);
7887 arg1 = simplify_giv_expr (loop, XEXP (x, 1), ext_val, benefit);
7888 if (arg0 == 0 || arg1 == 0)
7891 /* Put constant last, CONST_INT last if both constant. */
7892 if ((GET_CODE (arg0) == USE
7893 || GET_CODE (arg0) == CONST_INT)
7894 && ! ((GET_CODE (arg0) == USE
7895 && GET_CODE (arg1) == USE)
7896 || GET_CODE (arg1) == CONST_INT))
7897 tem = arg0, arg0 = arg1, arg1 = tem;
7899 /* Handle addition of zero, then addition of an invariant. */
7900 if (arg1 == const0_rtx)
7902 else if (GET_CODE (arg1) == CONST_INT || GET_CODE (arg1) == USE)
7903 switch (GET_CODE (arg0))
7907 /* Adding two invariants must result in an invariant, so enclose
7908 addition operation inside a USE and return it. */
7909 if (GET_CODE (arg0) == USE)
7910 arg0 = XEXP (arg0, 0);
7911 if (GET_CODE (arg1) == USE)
7912 arg1 = XEXP (arg1, 0);
7914 if (GET_CODE (arg0) == CONST_INT)
7915 tem = arg0, arg0 = arg1, arg1 = tem;
7916 if (GET_CODE (arg1) == CONST_INT)
7917 tem = sge_plus_constant (arg0, arg1);
7919 tem = sge_plus (mode, arg0, arg1);
7921 if (GET_CODE (tem) != CONST_INT)
7922 tem = gen_rtx_USE (mode, tem);
7927 /* biv + invar or mult + invar. Return sum. */
7928 return gen_rtx_PLUS (mode, arg0, arg1);
7931 /* (a + invar_1) + invar_2. Associate. */
7933 simplify_giv_expr (loop,
7945 /* Each argument must be either REG, PLUS, or MULT. Convert REG to
7946 MULT to reduce cases. */
7948 arg0 = gen_rtx_MULT (mode, arg0, const1_rtx);
7950 arg1 = gen_rtx_MULT (mode, arg1, const1_rtx);
7952 /* Now have PLUS + PLUS, PLUS + MULT, MULT + PLUS, or MULT + MULT.
7953 Put a MULT first, leaving PLUS + PLUS, MULT + PLUS, or MULT + MULT.
7954 Recurse to associate the second PLUS. */
7955 if (GET_CODE (arg1) == MULT)
7956 tem = arg0, arg0 = arg1, arg1 = tem;
7958 if (GET_CODE (arg1) == PLUS)
7960 simplify_giv_expr (loop,
7962 gen_rtx_PLUS (mode, arg0,
7967 /* Now must have MULT + MULT. Distribute if same biv, else not giv. */
7968 if (GET_CODE (arg0) != MULT || GET_CODE (arg1) != MULT)
7971 if (!rtx_equal_p (arg0, arg1))
7974 return simplify_giv_expr (loop,
7983 /* Handle "a - b" as "a + b * (-1)". */
7984 return simplify_giv_expr (loop,
7993 arg0 = simplify_giv_expr (loop, XEXP (x, 0), ext_val, benefit);
7994 arg1 = simplify_giv_expr (loop, XEXP (x, 1), ext_val, benefit);
7995 if (arg0 == 0 || arg1 == 0)
7998 /* Put constant last, CONST_INT last if both constant. */
7999 if ((GET_CODE (arg0) == USE || GET_CODE (arg0) == CONST_INT)
8000 && GET_CODE (arg1) != CONST_INT)
8001 tem = arg0, arg0 = arg1, arg1 = tem;
8003 /* If second argument is not now constant, not giv. */
8004 if (GET_CODE (arg1) != USE && GET_CODE (arg1) != CONST_INT)
8007 /* Handle multiply by 0 or 1. */
8008 if (arg1 == const0_rtx)
8011 else if (arg1 == const1_rtx)
8014 switch (GET_CODE (arg0))
8017 /* biv * invar. Done. */
8018 return gen_rtx_MULT (mode, arg0, arg1);
8021 /* Product of two constants. */
8022 return GEN_INT (INTVAL (arg0) * INTVAL (arg1));
8025 /* invar * invar is a giv, but attempt to simplify it somehow. */
8026 if (GET_CODE (arg1) != CONST_INT)
8029 arg0 = XEXP (arg0, 0);
8030 if (GET_CODE (arg0) == MULT)
8032 /* (invar_0 * invar_1) * invar_2. Associate. */
8033 return simplify_giv_expr (loop,
8042 /* Propagate the MULT expressions to the innermost nodes. */
8043 else if (GET_CODE (arg0) == PLUS)
8045 /* (invar_0 + invar_1) * invar_2. Distribute. */
8046 return simplify_giv_expr (loop,
8058 return gen_rtx_USE (mode, gen_rtx_MULT (mode, arg0, arg1));
8061 /* (a * invar_1) * invar_2. Associate. */
8062 return simplify_giv_expr (loop,
8071 /* (a + invar_1) * invar_2. Distribute. */
8072 return simplify_giv_expr (loop,
8087 /* Shift by constant is multiply by power of two. */
8088 if (GET_CODE (XEXP (x, 1)) != CONST_INT)
8092 simplify_giv_expr (loop,
8095 GEN_INT ((HOST_WIDE_INT) 1
8096 << INTVAL (XEXP (x, 1)))),
8100 /* "-a" is "a * (-1)" */
8101 return simplify_giv_expr (loop,
8102 gen_rtx_MULT (mode, XEXP (x, 0), constm1_rtx),
8106 /* "~a" is "-a - 1". Silly, but easy. */
8107 return simplify_giv_expr (loop,
8108 gen_rtx_MINUS (mode,
8109 gen_rtx_NEG (mode, XEXP (x, 0)),
8114 /* Already in proper form for invariant. */
8120 /* Conditionally recognize extensions of simple IVs. After we've
8121 computed loop traversal counts and verified the range of the
8122 source IV, we'll reevaluate this as a GIV. */
8123 if (*ext_val == NULL_RTX)
8125 arg0 = simplify_giv_expr (loop, XEXP (x, 0), ext_val, benefit);
8126 if (arg0 && *ext_val == NULL_RTX && REG_P (arg0))
8128 *ext_val = gen_rtx_fmt_e (GET_CODE (x), mode, arg0);
8135 /* If this is a new register, we can't deal with it. */
8136 if (REGNO (x) >= max_reg_before_loop)
8139 /* Check for biv or giv. */
8140 switch (REG_IV_TYPE (ivs, REGNO (x)))
8144 case GENERAL_INDUCT:
8146 struct induction *v = REG_IV_INFO (ivs, REGNO (x));
8148 /* Form expression from giv and add benefit. Ensure this giv
8149 can derive another and subtract any needed adjustment if so. */
8151 /* Increasing the benefit here is risky. The only case in which it
8152 is arguably correct is if this is the only use of V. In other
8153 cases, this will artificially inflate the benefit of the current
8154 giv, and lead to suboptimal code. Thus, it is disabled, since
8155 potentially not reducing an only marginally beneficial giv is
8156 less harmful than reducing many givs that are not really
8159 rtx single_use = regs->array[REGNO (x)].single_usage;
8160 if (single_use && single_use != const0_rtx)
8161 *benefit += v->benefit;
8167 tem = gen_rtx_PLUS (mode, gen_rtx_MULT (mode,
8168 v->src_reg, v->mult_val),
8171 if (v->derive_adjustment)
8172 tem = gen_rtx_MINUS (mode, tem, v->derive_adjustment);
8173 arg0 = simplify_giv_expr (loop, tem, ext_val, benefit);
8176 if (!v->ext_dependent)
8181 *ext_val = v->ext_dependent;
8189 /* If it isn't an induction variable, and it is invariant, we
8190 may be able to simplify things further by looking through
8191 the bits we just moved outside the loop. */
8192 if (loop_invariant_p (loop, x) == 1)
8195 struct loop_movables *movables = LOOP_MOVABLES (loop);
8197 for (m = movables->head; m; m = m->next)
8198 if (rtx_equal_p (x, m->set_dest))
8200 /* Ok, we found a match. Substitute and simplify. */
8202 /* If we match another movable, we must use that, as
8203 this one is going away. */
8205 return simplify_giv_expr (loop, m->match->set_dest,
8208 /* If consec is nonzero, this is a member of a group of
8209 instructions that were moved together. We handle this
8210 case only to the point of seeking to the last insn and
8211 looking for a REG_EQUAL. Fail if we don't find one. */
8218 tem = NEXT_INSN (tem);
8222 tem = find_reg_note (tem, REG_EQUAL, NULL_RTX);
8224 tem = XEXP (tem, 0);
8228 tem = single_set (m->insn);
8230 tem = SET_SRC (tem);
8235 /* What we are most interested in is pointer
8236 arithmetic on invariants -- only take
8237 patterns we may be able to do something with. */
8238 if (GET_CODE (tem) == PLUS
8239 || GET_CODE (tem) == MULT
8240 || GET_CODE (tem) == ASHIFT
8241 || GET_CODE (tem) == CONST_INT
8242 || GET_CODE (tem) == SYMBOL_REF)
8244 tem = simplify_giv_expr (loop, tem, ext_val,
8249 else if (GET_CODE (tem) == CONST
8250 && GET_CODE (XEXP (tem, 0)) == PLUS
8251 && GET_CODE (XEXP (XEXP (tem, 0), 0)) == SYMBOL_REF
8252 && GET_CODE (XEXP (XEXP (tem, 0), 1)) == CONST_INT)
8254 tem = simplify_giv_expr (loop, XEXP (tem, 0),
8266 /* Fall through to general case. */
8268 /* If invariant, return as USE (unless CONST_INT).
8269 Otherwise, not giv. */
8270 if (GET_CODE (x) == USE)
8273 if (loop_invariant_p (loop, x) == 1)
8275 if (GET_CODE (x) == CONST_INT)
8277 if (GET_CODE (x) == CONST
8278 && GET_CODE (XEXP (x, 0)) == PLUS
8279 && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF
8280 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT)
8282 return gen_rtx_USE (mode, x);
8289 /* This routine folds invariants such that there is only ever one
8290 CONST_INT in the summation. It is only used by simplify_giv_expr. */
8293 sge_plus_constant (rtx x, rtx c)
8295 if (GET_CODE (x) == CONST_INT)
8296 return GEN_INT (INTVAL (x) + INTVAL (c));
8297 else if (GET_CODE (x) != PLUS)
8298 return gen_rtx_PLUS (GET_MODE (x), x, c);
8299 else if (GET_CODE (XEXP (x, 1)) == CONST_INT)
8301 return gen_rtx_PLUS (GET_MODE (x), XEXP (x, 0),
8302 GEN_INT (INTVAL (XEXP (x, 1)) + INTVAL (c)));
8304 else if (GET_CODE (XEXP (x, 0)) == PLUS
8305 || GET_CODE (XEXP (x, 1)) != PLUS)
8307 return gen_rtx_PLUS (GET_MODE (x),
8308 sge_plus_constant (XEXP (x, 0), c), XEXP (x, 1));
8312 return gen_rtx_PLUS (GET_MODE (x),
8313 sge_plus_constant (XEXP (x, 1), c), XEXP (x, 0));
8318 sge_plus (enum machine_mode mode, rtx x, rtx y)
8320 while (GET_CODE (y) == PLUS)
8322 rtx a = XEXP (y, 0);
8323 if (GET_CODE (a) == CONST_INT)
8324 x = sge_plus_constant (x, a);
8326 x = gen_rtx_PLUS (mode, x, a);
8329 if (GET_CODE (y) == CONST_INT)
8330 x = sge_plus_constant (x, y);
8332 x = gen_rtx_PLUS (mode, x, y);
8336 /* Help detect a giv that is calculated by several consecutive insns;
8340 The caller has already identified the first insn P as having a giv as dest;
8341 we check that all other insns that set the same register follow
8342 immediately after P, that they alter nothing else,
8343 and that the result of the last is still a giv.
8345 The value is 0 if the reg set in P is not really a giv.
8346 Otherwise, the value is the amount gained by eliminating
8347 all the consecutive insns that compute the value.
8349 FIRST_BENEFIT is the amount gained by eliminating the first insn, P.
8350 SRC_REG is the reg of the biv; DEST_REG is the reg of the giv.
8352 The coefficients of the ultimate giv value are stored in
8353 *MULT_VAL and *ADD_VAL. */
8356 consec_sets_giv (const struct loop *loop, int first_benefit, rtx p,
8357 rtx src_reg, rtx dest_reg, rtx *add_val, rtx *mult_val,
8358 rtx *ext_val, rtx *last_consec_insn)
8360 struct loop_ivs *ivs = LOOP_IVS (loop);
8361 struct loop_regs *regs = LOOP_REGS (loop);
8368 /* Indicate that this is a giv so that we can update the value produced in
8369 each insn of the multi-insn sequence.
8371 This induction structure will be used only by the call to
8372 general_induction_var below, so we can allocate it on our stack.
8373 If this is a giv, our caller will replace the induct var entry with
8374 a new induction structure. */
8375 struct induction *v;
8377 if (REG_IV_TYPE (ivs, REGNO (dest_reg)) != UNKNOWN_INDUCT)
8380 v = alloca (sizeof (struct induction));
8381 v->src_reg = src_reg;
8382 v->mult_val = *mult_val;
8383 v->add_val = *add_val;
8384 v->benefit = first_benefit;
8386 v->derive_adjustment = 0;
8387 v->ext_dependent = NULL_RTX;
8389 REG_IV_TYPE (ivs, REGNO (dest_reg)) = GENERAL_INDUCT;
8390 REG_IV_INFO (ivs, REGNO (dest_reg)) = v;
8392 count = regs->array[REGNO (dest_reg)].n_times_set - 1;
8397 code = GET_CODE (p);
8399 /* If libcall, skip to end of call sequence. */
8400 if (code == INSN && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
8404 && (set = single_set (p))
8405 && REG_P (SET_DEST (set))
8406 && SET_DEST (set) == dest_reg
8407 && (general_induction_var (loop, SET_SRC (set), &src_reg,
8408 add_val, mult_val, ext_val, 0,
8410 /* Giv created by equivalent expression. */
8411 || ((temp = find_reg_note (p, REG_EQUAL, NULL_RTX))
8412 && general_induction_var (loop, XEXP (temp, 0), &src_reg,
8413 add_val, mult_val, ext_val, 0,
8414 &benefit, VOIDmode)))
8415 && src_reg == v->src_reg)
8417 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
8418 benefit += libcall_benefit (p);
8421 v->mult_val = *mult_val;
8422 v->add_val = *add_val;
8423 v->benefit += benefit;
8425 else if (code != NOTE)
8427 /* Allow insns that set something other than this giv to a
8428 constant. Such insns are needed on machines which cannot
8429 include long constants and should not disqualify a giv. */
8431 && (set = single_set (p))
8432 && SET_DEST (set) != dest_reg
8433 && CONSTANT_P (SET_SRC (set)))
8436 REG_IV_TYPE (ivs, REGNO (dest_reg)) = UNKNOWN_INDUCT;
8441 REG_IV_TYPE (ivs, REGNO (dest_reg)) = UNKNOWN_INDUCT;
8442 *last_consec_insn = p;
8446 /* Return an rtx, if any, that expresses giv G2 as a function of the register
8447 represented by G1. If no such expression can be found, or it is clear that
8448 it cannot possibly be a valid address, 0 is returned.
8450 To perform the computation, we note that
8453 where `v' is the biv.
8455 So G2 = (y/b) * G1 + (b - a*y/x).
8457 Note that MULT = y/x.
8459 Update: A and B are now allowed to be additive expressions such that
8460 B contains all variables in A. That is, computing B-A will not require
8461 subtracting variables. */
8464 express_from_1 (rtx a, rtx b, rtx mult)
8466 /* If MULT is zero, then A*MULT is zero, and our expression is B. */
8468 if (mult == const0_rtx)
8471 /* If MULT is not 1, we cannot handle A with non-constants, since we
8472 would then be required to subtract multiples of the registers in A.
8473 This is theoretically possible, and may even apply to some Fortran
8474 constructs, but it is a lot of work and we do not attempt it here. */
8476 if (mult != const1_rtx && GET_CODE (a) != CONST_INT)
8479 /* In general these structures are sorted top to bottom (down the PLUS
8480 chain), but not left to right across the PLUS. If B is a higher
8481 order giv than A, we can strip one level and recurse. If A is higher
8482 order, we'll eventually bail out, but won't know that until the end.
8483 If they are the same, we'll strip one level around this loop. */
8485 while (GET_CODE (a) == PLUS && GET_CODE (b) == PLUS)
8487 rtx ra, rb, oa, ob, tmp;
8489 ra = XEXP (a, 0), oa = XEXP (a, 1);
8490 if (GET_CODE (ra) == PLUS)
8491 tmp = ra, ra = oa, oa = tmp;
8493 rb = XEXP (b, 0), ob = XEXP (b, 1);
8494 if (GET_CODE (rb) == PLUS)
8495 tmp = rb, rb = ob, ob = tmp;
8497 if (rtx_equal_p (ra, rb))
8498 /* We matched: remove one reg completely. */
8500 else if (GET_CODE (ob) != PLUS && rtx_equal_p (ra, ob))
8501 /* An alternate match. */
8503 else if (GET_CODE (oa) != PLUS && rtx_equal_p (oa, rb))
8504 /* An alternate match. */
8508 /* Indicates an extra register in B. Strip one level from B and
8509 recurse, hoping B was the higher order expression. */
8510 ob = express_from_1 (a, ob, mult);
8513 return gen_rtx_PLUS (GET_MODE (b), rb, ob);
8517 /* Here we are at the last level of A, go through the cases hoping to
8518 get rid of everything but a constant. */
8520 if (GET_CODE (a) == PLUS)
8524 ra = XEXP (a, 0), oa = XEXP (a, 1);
8525 if (rtx_equal_p (oa, b))
8527 else if (!rtx_equal_p (ra, b))
8530 if (GET_CODE (oa) != CONST_INT)
8533 return GEN_INT (-INTVAL (oa) * INTVAL (mult));
8535 else if (GET_CODE (a) == CONST_INT)
8537 return plus_constant (b, -INTVAL (a) * INTVAL (mult));
8539 else if (CONSTANT_P (a))
8541 enum machine_mode mode_a = GET_MODE (a);
8542 enum machine_mode mode_b = GET_MODE (b);
8543 enum machine_mode mode = mode_b == VOIDmode ? mode_a : mode_b;
8544 return simplify_gen_binary (MINUS, mode, b, a);
8546 else if (GET_CODE (b) == PLUS)
8548 if (rtx_equal_p (a, XEXP (b, 0)))
8550 else if (rtx_equal_p (a, XEXP (b, 1)))
8555 else if (rtx_equal_p (a, b))
8562 express_from (struct induction *g1, struct induction *g2)
8566 /* The value that G1 will be multiplied by must be a constant integer. Also,
8567 the only chance we have of getting a valid address is if b*c/a (see above
8568 for notation) is also an integer. */
8569 if (GET_CODE (g1->mult_val) == CONST_INT
8570 && GET_CODE (g2->mult_val) == CONST_INT)
8572 if (g1->mult_val == const0_rtx
8573 || (g1->mult_val == constm1_rtx
8574 && INTVAL (g2->mult_val)
8575 == (HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT - 1))
8576 || INTVAL (g2->mult_val) % INTVAL (g1->mult_val) != 0)
8578 mult = GEN_INT (INTVAL (g2->mult_val) / INTVAL (g1->mult_val));
8580 else if (rtx_equal_p (g1->mult_val, g2->mult_val))
8584 /* ??? Find out if the one is a multiple of the other? */
8588 add = express_from_1 (g1->add_val, g2->add_val, mult);
8589 if (add == NULL_RTX)
8591 /* Failed. If we've got a multiplication factor between G1 and G2,
8592 scale G1's addend and try again. */
8593 if (INTVAL (mult) > 1)
8595 rtx g1_add_val = g1->add_val;
8596 if (GET_CODE (g1_add_val) == MULT
8597 && GET_CODE (XEXP (g1_add_val, 1)) == CONST_INT)
8600 m = INTVAL (mult) * INTVAL (XEXP (g1_add_val, 1));
8601 g1_add_val = gen_rtx_MULT (GET_MODE (g1_add_val),
8602 XEXP (g1_add_val, 0), GEN_INT (m));
8606 g1_add_val = gen_rtx_MULT (GET_MODE (g1_add_val), g1_add_val,
8610 add = express_from_1 (g1_add_val, g2->add_val, const1_rtx);
8613 if (add == NULL_RTX)
8616 /* Form simplified final result. */
8617 if (mult == const0_rtx)
8619 else if (mult == const1_rtx)
8620 mult = g1->dest_reg;
8622 mult = gen_rtx_MULT (g2->mode, g1->dest_reg, mult);
8624 if (add == const0_rtx)
8628 if (GET_CODE (add) == PLUS
8629 && CONSTANT_P (XEXP (add, 1)))
8631 rtx tem = XEXP (add, 1);
8632 mult = gen_rtx_PLUS (g2->mode, mult, XEXP (add, 0));
8636 return gen_rtx_PLUS (g2->mode, mult, add);
8640 /* Return an rtx, if any, that expresses giv G2 as a function of the register
8641 represented by G1. This indicates that G2 should be combined with G1 and
8642 that G2 can use (either directly or via an address expression) a register
8643 used to represent G1. */
8646 combine_givs_p (struct induction *g1, struct induction *g2)
8650 /* With the introduction of ext dependent givs, we must care for modes.
8651 G2 must not use a wider mode than G1. */
8652 if (GET_MODE_SIZE (g1->mode) < GET_MODE_SIZE (g2->mode))
8655 ret = comb = express_from (g1, g2);
8656 if (comb == NULL_RTX)
8658 if (g1->mode != g2->mode)
8659 ret = gen_lowpart (g2->mode, comb);
8661 /* If these givs are identical, they can be combined. We use the results
8662 of express_from because the addends are not in a canonical form, so
8663 rtx_equal_p is a weaker test. */
8664 /* But don't combine a DEST_REG giv with a DEST_ADDR giv; we want the
8665 combination to be the other way round. */
8666 if (comb == g1->dest_reg
8667 && (g1->giv_type == DEST_REG || g2->giv_type == DEST_ADDR))
8672 /* If G2 can be expressed as a function of G1 and that function is valid
8673 as an address and no more expensive than using a register for G2,
8674 the expression of G2 in terms of G1 can be used. */
8676 && g2->giv_type == DEST_ADDR
8677 && memory_address_p (GET_MODE (g2->mem), ret))
8683 /* See if BL is monotonic and has a constant per-iteration increment.
8684 Return the increment if so, otherwise return 0. */
8686 static HOST_WIDE_INT
8687 get_monotonic_increment (struct iv_class *bl)
8689 struct induction *v;
8692 /* Get the total increment and check that it is constant. */
8693 incr = biv_total_increment (bl);
8694 if (incr == 0 || GET_CODE (incr) != CONST_INT)
8697 for (v = bl->biv; v != 0; v = v->next_iv)
8699 if (GET_CODE (v->add_val) != CONST_INT)
8702 if (INTVAL (v->add_val) < 0 && INTVAL (incr) >= 0)
8705 if (INTVAL (v->add_val) > 0 && INTVAL (incr) <= 0)
8708 return INTVAL (incr);
8712 /* Subroutine of biv_fits_mode_p. Return true if biv BL, when biased by
8713 BIAS, will never exceed the unsigned range of MODE. LOOP is the loop
8714 to which the biv belongs and INCR is its per-iteration increment. */
8717 biased_biv_fits_mode_p (const struct loop *loop, struct iv_class *bl,
8718 HOST_WIDE_INT incr, enum machine_mode mode,
8719 unsigned HOST_WIDE_INT bias)
8721 unsigned HOST_WIDE_INT initial, maximum, span, delta;
8723 /* We need to be able to manipulate MODE-size constants. */
8724 if (HOST_BITS_PER_WIDE_INT < GET_MODE_BITSIZE (mode))
8727 /* The number of loop iterations must be constant. */
8728 if (LOOP_INFO (loop)->n_iterations == 0)
8731 /* So must the biv's initial value. */
8732 if (bl->initial_value == 0 || GET_CODE (bl->initial_value) != CONST_INT)
8735 initial = bias + INTVAL (bl->initial_value);
8736 maximum = GET_MODE_MASK (mode);
8738 /* Make sure that the initial value is within range. */
8739 if (initial > maximum)
8742 /* Set up DELTA and SPAN such that the number of iterations * DELTA
8743 (calculated to arbitrary precision) must be <= SPAN. */
8752 /* Handle the special case in which MAXIMUM is the largest
8753 unsigned HOST_WIDE_INT and INITIAL is 0. */
8754 if (maximum + 1 == initial)
8755 span = LOOP_INFO (loop)->n_iterations * delta;
8757 span = maximum + 1 - initial;
8759 return (span / LOOP_INFO (loop)->n_iterations >= delta);
8763 /* Return true if biv BL will never exceed the bounds of MODE. LOOP is
8764 the loop to which BL belongs and INCR is its per-iteration increment.
8765 UNSIGNEDP is true if the biv should be treated as unsigned. */
8768 biv_fits_mode_p (const struct loop *loop, struct iv_class *bl,
8769 HOST_WIDE_INT incr, enum machine_mode mode, bool unsignedp)
8771 struct loop_info *loop_info;
8772 unsigned HOST_WIDE_INT bias;
8774 /* A biv's value will always be limited to its natural mode.
8775 Larger modes will observe the same wrap-around. */
8776 if (GET_MODE_SIZE (mode) > GET_MODE_SIZE (GET_MODE (bl->biv->src_reg)))
8777 mode = GET_MODE (bl->biv->src_reg);
8779 loop_info = LOOP_INFO (loop);
8781 bias = (unsignedp ? 0 : (GET_MODE_MASK (mode) >> 1) + 1);
8782 if (biased_biv_fits_mode_p (loop, bl, incr, mode, bias))
8785 if (mode == GET_MODE (bl->biv->src_reg)
8786 && bl->biv->src_reg == loop_info->iteration_var
8787 && loop_info->comparison_value
8788 && loop_invariant_p (loop, loop_info->comparison_value))
8790 /* If the increment is +1, and the exit test is a <, the BIV
8791 cannot overflow. (For <=, we have the problematic case that
8792 the comparison value might be the maximum value of the range.) */
8795 if (loop_info->comparison_code == LT)
8797 if (loop_info->comparison_code == LTU && unsignedp)
8801 /* Likewise for increment -1 and exit test >. */
8804 if (loop_info->comparison_code == GT)
8806 if (loop_info->comparison_code == GTU && unsignedp)
8814 /* Return false iff it is provable that biv BL plus BIAS will not wrap
8815 at any point in its update sequence. Note that at the rtl level we
8816 may not have information about the signedness of BL; in that case,
8817 check for both signed and unsigned overflow. */
8820 biased_biv_may_wrap_p (const struct loop *loop, struct iv_class *bl,
8821 unsigned HOST_WIDE_INT bias)
8824 bool check_signed, check_unsigned;
8825 enum machine_mode mode;
8827 /* If the increment is not monotonic, we'd have to check separately
8828 at each increment step. Not Worth It. */
8829 incr = get_monotonic_increment (bl);
8833 /* If this biv is the loop iteration variable, then we may be able to
8834 deduce a sign based on the loop condition. */
8835 /* ??? This is not 100% reliable; consider an unsigned biv that is cast
8836 to signed for the comparison. However, this same bug appears all
8838 check_signed = check_unsigned = true;
8839 if (bl->biv->src_reg == LOOP_INFO (loop)->iteration_var)
8841 switch (LOOP_INFO (loop)->comparison_code)
8843 case GTU: case GEU: case LTU: case LEU:
8844 check_signed = false;
8846 case GT: case GE: case LT: case LE:
8847 check_unsigned = false;
8854 mode = GET_MODE (bl->biv->src_reg);
8857 && !biased_biv_fits_mode_p (loop, bl, incr, mode, bias))
8862 bias += (GET_MODE_MASK (mode) >> 1) + 1;
8863 if (!biased_biv_fits_mode_p (loop, bl, incr, mode, bias))
8871 /* Given that X is an extension or truncation of BL, return true
8872 if it is unaffected by overflow. LOOP is the loop to which
8873 BL belongs and INCR is its per-iteration increment. */
8876 extension_within_bounds_p (const struct loop *loop, struct iv_class *bl,
8877 HOST_WIDE_INT incr, rtx x)
8879 enum machine_mode mode;
8880 bool signedp, unsignedp;
8882 switch (GET_CODE (x))
8886 mode = GET_MODE (XEXP (x, 0));
8887 signedp = (GET_CODE (x) == SIGN_EXTEND);
8888 unsignedp = (GET_CODE (x) == ZERO_EXTEND);
8892 /* We don't know whether this value is being used as signed
8893 or unsigned, so check the conditions for both. */
8894 mode = GET_MODE (x);
8895 signedp = unsignedp = true;
8902 return ((!signedp || biv_fits_mode_p (loop, bl, incr, mode, false))
8903 && (!unsignedp || biv_fits_mode_p (loop, bl, incr, mode, true)));
8907 /* Check each extension dependent giv in this class to see if its
8908 root biv is safe from wrapping in the interior mode, which would
8909 make the giv illegal. */
8912 check_ext_dependent_givs (const struct loop *loop, struct iv_class *bl)
8914 struct induction *v;
8917 incr = get_monotonic_increment (bl);
8919 /* Invalidate givs that fail the tests. */
8920 for (v = bl->giv; v; v = v->next_iv)
8921 if (v->ext_dependent)
8924 && extension_within_bounds_p (loop, bl, incr, v->ext_dependent))
8926 if (loop_dump_stream)
8927 fprintf (loop_dump_stream,
8928 "Verified ext dependent giv at %d of reg %d\n",
8929 INSN_UID (v->insn), bl->regno);
8933 if (loop_dump_stream)
8934 fprintf (loop_dump_stream,
8935 "Failed ext dependent giv at %d\n",
8936 INSN_UID (v->insn));
8939 bl->all_reduced = 0;
8944 /* Generate a version of VALUE in a mode appropriate for initializing V. */
8947 extend_value_for_giv (struct induction *v, rtx value)
8949 rtx ext_dep = v->ext_dependent;
8954 /* Recall that check_ext_dependent_givs verified that the known bounds
8955 of a biv did not overflow or wrap with respect to the extension for
8956 the giv. Therefore, constants need no additional adjustment. */
8957 if (CONSTANT_P (value) && GET_MODE (value) == VOIDmode)
8960 /* Otherwise, we must adjust the value to compensate for the
8961 differing modes of the biv and the giv. */
8962 return gen_rtx_fmt_e (GET_CODE (ext_dep), GET_MODE (ext_dep), value);
8965 struct combine_givs_stats
8972 cmp_combine_givs_stats (const void *xp, const void *yp)
8974 const struct combine_givs_stats * const x =
8975 (const struct combine_givs_stats *) xp;
8976 const struct combine_givs_stats * const y =
8977 (const struct combine_givs_stats *) yp;
8979 d = y->total_benefit - x->total_benefit;
8980 /* Stabilize the sort. */
8982 d = x->giv_number - y->giv_number;
8986 /* Check all pairs of givs for iv_class BL and see if any can be combined with
8987 any other. If so, point SAME to the giv combined with and set NEW_REG to
8988 be an expression (in terms of the other giv's DEST_REG) equivalent to the
8989 giv. Also, update BENEFIT and related fields for cost/benefit analysis. */
8992 combine_givs (struct loop_regs *regs, struct iv_class *bl)
8994 /* Additional benefit to add for being combined multiple times. */
8995 const int extra_benefit = 3;
8997 struct induction *g1, *g2, **giv_array;
8998 int i, j, k, giv_count;
8999 struct combine_givs_stats *stats;
9002 /* Count givs, because bl->giv_count is incorrect here. */
9004 for (g1 = bl->giv; g1; g1 = g1->next_iv)
9008 giv_array = alloca (giv_count * sizeof (struct induction *));
9010 for (g1 = bl->giv; g1; g1 = g1->next_iv)
9012 giv_array[i++] = g1;
9014 stats = xcalloc (giv_count, sizeof (*stats));
9015 can_combine = xcalloc (giv_count, giv_count * sizeof (rtx));
9017 for (i = 0; i < giv_count; i++)
9023 stats[i].giv_number = i;
9025 /* If a DEST_REG GIV is used only once, do not allow it to combine
9026 with anything, for in doing so we will gain nothing that cannot
9027 be had by simply letting the GIV with which we would have combined
9028 to be reduced on its own. The lossage shows up in particular with
9029 DEST_ADDR targets on hosts with reg+reg addressing, though it can
9030 be seen elsewhere as well. */
9031 if (g1->giv_type == DEST_REG
9032 && (single_use = regs->array[REGNO (g1->dest_reg)].single_usage)
9033 && single_use != const0_rtx)
9036 this_benefit = g1->benefit;
9037 /* Add an additional weight for zero addends. */
9038 if (g1->no_const_addval)
9041 for (j = 0; j < giv_count; j++)
9047 && (this_combine = combine_givs_p (g1, g2)) != NULL_RTX)
9049 can_combine[i * giv_count + j] = this_combine;
9050 this_benefit += g2->benefit + extra_benefit;
9053 stats[i].total_benefit = this_benefit;
9056 /* Iterate, combining until we can't. */
9058 qsort (stats, giv_count, sizeof (*stats), cmp_combine_givs_stats);
9060 if (loop_dump_stream)
9062 fprintf (loop_dump_stream, "Sorted combine statistics:\n");
9063 for (k = 0; k < giv_count; k++)
9065 g1 = giv_array[stats[k].giv_number];
9066 if (!g1->combined_with && !g1->same)
9067 fprintf (loop_dump_stream, " {%d, %d}",
9068 INSN_UID (giv_array[stats[k].giv_number]->insn),
9069 stats[k].total_benefit);
9071 putc ('\n', loop_dump_stream);
9074 for (k = 0; k < giv_count; k++)
9076 int g1_add_benefit = 0;
9078 i = stats[k].giv_number;
9081 /* If it has already been combined, skip. */
9082 if (g1->combined_with || g1->same)
9085 for (j = 0; j < giv_count; j++)
9088 if (g1 != g2 && can_combine[i * giv_count + j]
9089 /* If it has already been combined, skip. */
9090 && ! g2->same && ! g2->combined_with)
9094 g2->new_reg = can_combine[i * giv_count + j];
9096 /* For destination, we now may replace by mem expression instead
9097 of register. This changes the costs considerably, so add the
9099 if (g2->giv_type == DEST_ADDR)
9100 g2->benefit = (g2->benefit + reg_address_cost
9101 - address_cost (g2->new_reg,
9102 GET_MODE (g2->mem)));
9103 g1->combined_with++;
9104 g1->lifetime += g2->lifetime;
9106 g1_add_benefit += g2->benefit;
9108 /* ??? The new final_[bg]iv_value code does a much better job
9109 of finding replaceable giv's, and hence this code may no
9110 longer be necessary. */
9111 if (! g2->replaceable && REG_USERVAR_P (g2->dest_reg))
9112 g1_add_benefit -= copy_cost;
9114 /* To help optimize the next set of combinations, remove
9115 this giv from the benefits of other potential mates. */
9116 for (l = 0; l < giv_count; ++l)
9118 int m = stats[l].giv_number;
9119 if (can_combine[m * giv_count + j])
9120 stats[l].total_benefit -= g2->benefit + extra_benefit;
9123 if (loop_dump_stream)
9124 fprintf (loop_dump_stream,
9125 "giv at %d combined with giv at %d; new benefit %d + %d, lifetime %d\n",
9126 INSN_UID (g2->insn), INSN_UID (g1->insn),
9127 g1->benefit, g1_add_benefit, g1->lifetime);
9131 /* To help optimize the next set of combinations, remove
9132 this giv from the benefits of other potential mates. */
9133 if (g1->combined_with)
9135 for (j = 0; j < giv_count; ++j)
9137 int m = stats[j].giv_number;
9138 if (can_combine[m * giv_count + i])
9139 stats[j].total_benefit -= g1->benefit + extra_benefit;
9142 g1->benefit += g1_add_benefit;
9144 /* We've finished with this giv, and everything it touched.
9145 Restart the combination so that proper weights for the
9146 rest of the givs are properly taken into account. */
9147 /* ??? Ideally we would compact the arrays at this point, so
9148 as to not cover old ground. But sanely compacting
9149 can_combine is tricky. */
9159 /* Generate sequence for REG = B * M + A. B is the initial value of
9160 the basic induction variable, M a multiplicative constant, A an
9161 additive constant and REG the destination register. */
9164 gen_add_mult (rtx b, rtx m, rtx a, rtx reg)
9170 /* Use unsigned arithmetic. */
9171 result = expand_mult_add (b, reg, m, a, GET_MODE (reg), 1);
9173 emit_move_insn (reg, result);
9181 /* Update registers created in insn sequence SEQ. */
9184 loop_regs_update (const struct loop *loop ATTRIBUTE_UNUSED, rtx seq)
9188 /* Update register info for alias analysis. */
9191 while (insn != NULL_RTX)
9193 rtx set = single_set (insn);
9195 if (set && REG_P (SET_DEST (set)))
9196 record_base_value (REGNO (SET_DEST (set)), SET_SRC (set), 0);
9198 insn = NEXT_INSN (insn);
9203 /* EMIT code before BEFORE_BB/BEFORE_INSN to set REG = B * M + A. B
9204 is the initial value of the basic induction variable, M a
9205 multiplicative constant, A an additive constant and REG the
9206 destination register. */
9209 loop_iv_add_mult_emit_before (const struct loop *loop, rtx b, rtx m, rtx a,
9210 rtx reg, basic_block before_bb, rtx before_insn)
9216 loop_iv_add_mult_hoist (loop, b, m, a, reg);
9220 /* Use copy_rtx to prevent unexpected sharing of these rtx. */
9221 seq = gen_add_mult (copy_rtx (b), copy_rtx (m), copy_rtx (a), reg);
9223 /* Increase the lifetime of any invariants moved further in code. */
9224 update_reg_last_use (a, before_insn);
9225 update_reg_last_use (b, before_insn);
9226 update_reg_last_use (m, before_insn);
9228 /* It is possible that the expansion created lots of new registers.
9229 Iterate over the sequence we just created and record them all. We
9230 must do this before inserting the sequence. */
9231 loop_regs_update (loop, seq);
9233 loop_insn_emit_before (loop, before_bb, before_insn, seq);
9237 /* Emit insns in loop pre-header to set REG = B * M + A. B is the
9238 initial value of the basic induction variable, M a multiplicative
9239 constant, A an additive constant and REG the destination
9243 loop_iv_add_mult_sink (const struct loop *loop, rtx b, rtx m, rtx a, rtx reg)
9247 /* Use copy_rtx to prevent unexpected sharing of these rtx. */
9248 seq = gen_add_mult (copy_rtx (b), copy_rtx (m), copy_rtx (a), reg);
9250 /* Increase the lifetime of any invariants moved further in code.
9251 ???? Is this really necessary? */
9252 update_reg_last_use (a, loop->sink);
9253 update_reg_last_use (b, loop->sink);
9254 update_reg_last_use (m, loop->sink);
9256 /* It is possible that the expansion created lots of new registers.
9257 Iterate over the sequence we just created and record them all. We
9258 must do this before inserting the sequence. */
9259 loop_regs_update (loop, seq);
9261 loop_insn_sink (loop, seq);
9265 /* Emit insns after loop to set REG = B * M + A. B is the initial
9266 value of the basic induction variable, M a multiplicative constant,
9267 A an additive constant and REG the destination register. */
9270 loop_iv_add_mult_hoist (const struct loop *loop, rtx b, rtx m, rtx a, rtx reg)
9274 /* Use copy_rtx to prevent unexpected sharing of these rtx. */
9275 seq = gen_add_mult (copy_rtx (b), copy_rtx (m), copy_rtx (a), reg);
9277 /* It is possible that the expansion created lots of new registers.
9278 Iterate over the sequence we just created and record them all. We
9279 must do this before inserting the sequence. */
9280 loop_regs_update (loop, seq);
9282 loop_insn_hoist (loop, seq);
9287 /* Similar to gen_add_mult, but compute cost rather than generating
9291 iv_add_mult_cost (rtx b, rtx m, rtx a, rtx reg)
9297 result = expand_mult_add (b, reg, m, a, GET_MODE (reg), 1);
9299 emit_move_insn (reg, result);
9300 last = get_last_insn ();
9303 rtx t = single_set (last);
9305 cost += rtx_cost (SET_SRC (t), SET);
9306 last = PREV_INSN (last);
9312 /* Test whether A * B can be computed without
9313 an actual multiply insn. Value is 1 if so.
9315 ??? This function stinks because it generates a ton of wasted RTL
9316 ??? and as a result fragments GC memory to no end. There are other
9317 ??? places in the compiler which are invoked a lot and do the same
9318 ??? thing, generate wasted RTL just to see if something is possible. */
9321 product_cheap_p (rtx a, rtx b)
9326 /* If only one is constant, make it B. */
9327 if (GET_CODE (a) == CONST_INT)
9328 tmp = a, a = b, b = tmp;
9330 /* If first constant, both constant, so don't need multiply. */
9331 if (GET_CODE (a) == CONST_INT)
9334 /* If second not constant, neither is constant, so would need multiply. */
9335 if (GET_CODE (b) != CONST_INT)
9338 /* One operand is constant, so might not need multiply insn. Generate the
9339 code for the multiply and see if a call or multiply, or long sequence
9340 of insns is generated. */
9343 expand_mult (GET_MODE (a), a, b, NULL_RTX, 1);
9348 if (tmp == NULL_RTX)
9350 else if (INSN_P (tmp))
9353 while (tmp != NULL_RTX)
9355 rtx next = NEXT_INSN (tmp);
9358 || !NONJUMP_INSN_P (tmp)
9359 || (GET_CODE (PATTERN (tmp)) == SET
9360 && GET_CODE (SET_SRC (PATTERN (tmp))) == MULT)
9361 || (GET_CODE (PATTERN (tmp)) == PARALLEL
9362 && GET_CODE (XVECEXP (PATTERN (tmp), 0, 0)) == SET
9363 && GET_CODE (SET_SRC (XVECEXP (PATTERN (tmp), 0, 0))) == MULT))
9372 else if (GET_CODE (tmp) == SET
9373 && GET_CODE (SET_SRC (tmp)) == MULT)
9375 else if (GET_CODE (tmp) == PARALLEL
9376 && GET_CODE (XVECEXP (tmp, 0, 0)) == SET
9377 && GET_CODE (SET_SRC (XVECEXP (tmp, 0, 0))) == MULT)
9383 /* Check to see if loop can be terminated by a "decrement and branch until
9384 zero" instruction. If so, add a REG_NONNEG note to the branch insn if so.
9385 Also try reversing an increment loop to a decrement loop
9386 to see if the optimization can be performed.
9387 Value is nonzero if optimization was performed. */
9389 /* This is useful even if the architecture doesn't have such an insn,
9390 because it might change a loops which increments from 0 to n to a loop
9391 which decrements from n to 0. A loop that decrements to zero is usually
9392 faster than one that increments from zero. */
9394 /* ??? This could be rewritten to use some of the loop unrolling procedures,
9395 such as approx_final_value, biv_total_increment, loop_iterations, and
9396 final_[bg]iv_value. */
9399 check_dbra_loop (struct loop *loop, int insn_count)
9401 struct loop_info *loop_info = LOOP_INFO (loop);
9402 struct loop_regs *regs = LOOP_REGS (loop);
9403 struct loop_ivs *ivs = LOOP_IVS (loop);
9404 struct iv_class *bl;
9406 enum machine_mode mode;
9412 rtx before_comparison;
9416 int compare_and_branch;
9417 rtx loop_start = loop->start;
9418 rtx loop_end = loop->end;
9420 /* If last insn is a conditional branch, and the insn before tests a
9421 register value, try to optimize it. Otherwise, we can't do anything. */
9423 jump = PREV_INSN (loop_end);
9424 comparison = get_condition_for_loop (loop, jump);
9425 if (comparison == 0)
9427 if (!onlyjump_p (jump))
9430 /* Try to compute whether the compare/branch at the loop end is one or
9431 two instructions. */
9432 get_condition (jump, &first_compare, false, true);
9433 if (first_compare == jump)
9434 compare_and_branch = 1;
9435 else if (first_compare == prev_nonnote_insn (jump))
9436 compare_and_branch = 2;
9441 /* If more than one condition is present to control the loop, then
9442 do not proceed, as this function does not know how to rewrite
9443 loop tests with more than one condition.
9445 Look backwards from the first insn in the last comparison
9446 sequence and see if we've got another comparison sequence. */
9449 if ((jump1 = prev_nonnote_insn (first_compare))
9454 /* Check all of the bivs to see if the compare uses one of them.
9455 Skip biv's set more than once because we can't guarantee that
9456 it will be zero on the last iteration. Also skip if the biv is
9457 used between its update and the test insn. */
9459 for (bl = ivs->list; bl; bl = bl->next)
9461 if (bl->biv_count == 1
9462 && ! bl->biv->maybe_multiple
9463 && bl->biv->dest_reg == XEXP (comparison, 0)
9464 && ! reg_used_between_p (regno_reg_rtx[bl->regno], bl->biv->insn,
9469 /* Try swapping the comparison to identify a suitable biv. */
9471 for (bl = ivs->list; bl; bl = bl->next)
9472 if (bl->biv_count == 1
9473 && ! bl->biv->maybe_multiple
9474 && bl->biv->dest_reg == XEXP (comparison, 1)
9475 && ! reg_used_between_p (regno_reg_rtx[bl->regno], bl->biv->insn,
9478 comparison = gen_rtx_fmt_ee (swap_condition (GET_CODE (comparison)),
9480 XEXP (comparison, 1),
9481 XEXP (comparison, 0));
9488 /* Look for the case where the basic induction variable is always
9489 nonnegative, and equals zero on the last iteration.
9490 In this case, add a reg_note REG_NONNEG, which allows the
9491 m68k DBRA instruction to be used. */
9493 if (((GET_CODE (comparison) == GT && XEXP (comparison, 1) == constm1_rtx)
9494 || (GET_CODE (comparison) == NE && XEXP (comparison, 1) == const0_rtx))
9495 && GET_CODE (bl->biv->add_val) == CONST_INT
9496 && INTVAL (bl->biv->add_val) < 0)
9498 /* Initial value must be greater than 0,
9499 init_val % -dec_value == 0 to ensure that it equals zero on
9500 the last iteration */
9502 if (GET_CODE (bl->initial_value) == CONST_INT
9503 && INTVAL (bl->initial_value) > 0
9504 && (INTVAL (bl->initial_value)
9505 % (-INTVAL (bl->biv->add_val))) == 0)
9507 /* Register always nonnegative, add REG_NOTE to branch. */
9508 if (! find_reg_note (jump, REG_NONNEG, NULL_RTX))
9510 = gen_rtx_EXPR_LIST (REG_NONNEG, bl->biv->dest_reg,
9517 /* If the decrement is 1 and the value was tested as >= 0 before
9518 the loop, then we can safely optimize. */
9519 for (p = loop_start; p; p = PREV_INSN (p))
9526 before_comparison = get_condition_for_loop (loop, p);
9527 if (before_comparison
9528 && XEXP (before_comparison, 0) == bl->biv->dest_reg
9529 && (GET_CODE (before_comparison) == LT
9530 || GET_CODE (before_comparison) == LTU)
9531 && XEXP (before_comparison, 1) == const0_rtx
9532 && ! reg_set_between_p (bl->biv->dest_reg, p, loop_start)
9533 && INTVAL (bl->biv->add_val) == -1)
9535 if (! find_reg_note (jump, REG_NONNEG, NULL_RTX))
9537 = gen_rtx_EXPR_LIST (REG_NONNEG, bl->biv->dest_reg,
9545 else if (GET_CODE (bl->biv->add_val) == CONST_INT
9546 && INTVAL (bl->biv->add_val) > 0)
9548 /* Try to change inc to dec, so can apply above optimization. */
9550 all registers modified are induction variables or invariant,
9551 all memory references have non-overlapping addresses
9552 (obviously true if only one write)
9553 allow 2 insns for the compare/jump at the end of the loop. */
9554 /* Also, we must avoid any instructions which use both the reversed
9555 biv and another biv. Such instructions will fail if the loop is
9556 reversed. We meet this condition by requiring that either
9557 no_use_except_counting is true, or else that there is only
9559 int num_nonfixed_reads = 0;
9560 /* 1 if the iteration var is used only to count iterations. */
9561 int no_use_except_counting = 0;
9562 /* 1 if the loop has no memory store, or it has a single memory store
9563 which is reversible. */
9564 int reversible_mem_store = 1;
9566 if (bl->giv_count == 0
9567 && !loop->exit_count
9568 && !loop_info->has_multiple_exit_targets)
9570 rtx bivreg = regno_reg_rtx[bl->regno];
9571 struct iv_class *blt;
9573 /* If there are no givs for this biv, and the only exit is the
9574 fall through at the end of the loop, then
9575 see if perhaps there are no uses except to count. */
9576 no_use_except_counting = 1;
9577 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
9580 rtx set = single_set (p);
9582 if (set && REG_P (SET_DEST (set))
9583 && REGNO (SET_DEST (set)) == bl->regno)
9584 /* An insn that sets the biv is okay. */
9586 else if (!reg_mentioned_p (bivreg, PATTERN (p)))
9587 /* An insn that doesn't mention the biv is okay. */
9589 else if (p == prev_nonnote_insn (prev_nonnote_insn (loop_end))
9590 || p == prev_nonnote_insn (loop_end))
9592 /* If either of these insns uses the biv and sets a pseudo
9593 that has more than one usage, then the biv has uses
9594 other than counting since it's used to derive a value
9595 that is used more than one time. */
9596 note_stores (PATTERN (p), note_set_pseudo_multiple_uses,
9598 if (regs->multiple_uses)
9600 no_use_except_counting = 0;
9606 no_use_except_counting = 0;
9611 /* A biv has uses besides counting if it is used to set
9613 for (blt = ivs->list; blt; blt = blt->next)
9615 && reg_mentioned_p (bivreg, SET_SRC (blt->init_set)))
9617 no_use_except_counting = 0;
9622 if (no_use_except_counting)
9623 /* No need to worry about MEMs. */
9625 else if (loop_info->num_mem_sets <= 1)
9627 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
9629 num_nonfixed_reads += count_nonfixed_reads (loop, PATTERN (p));
9631 /* If the loop has a single store, and the destination address is
9632 invariant, then we can't reverse the loop, because this address
9633 might then have the wrong value at loop exit.
9634 This would work if the source was invariant also, however, in that
9635 case, the insn should have been moved out of the loop. */
9637 if (loop_info->num_mem_sets == 1)
9639 struct induction *v;
9641 /* If we could prove that each of the memory locations
9642 written to was different, then we could reverse the
9643 store -- but we don't presently have any way of
9645 reversible_mem_store = 0;
9647 /* If the store depends on a register that is set after the
9648 store, it depends on the initial value, and is thus not
9650 for (v = bl->giv; reversible_mem_store && v; v = v->next_iv)
9652 if (v->giv_type == DEST_REG
9653 && reg_mentioned_p (v->dest_reg,
9654 PATTERN (loop_info->first_loop_store_insn))
9655 && loop_insn_first_p (loop_info->first_loop_store_insn,
9657 reversible_mem_store = 0;
9664 /* This code only acts for innermost loops. Also it simplifies
9665 the memory address check by only reversing loops with
9666 zero or one memory access.
9667 Two memory accesses could involve parts of the same array,
9668 and that can't be reversed.
9669 If the biv is used only for counting, than we don't need to worry
9670 about all these things. */
9672 if ((num_nonfixed_reads <= 1
9673 && ! loop_info->has_nonconst_call
9674 && ! loop_info->has_prefetch
9675 && ! loop_info->has_volatile
9676 && reversible_mem_store
9677 && (bl->giv_count + bl->biv_count + loop_info->num_mem_sets
9678 + num_unmoved_movables (loop) + compare_and_branch == insn_count)
9679 && (bl == ivs->list && bl->next == 0))
9680 || (no_use_except_counting && ! loop_info->has_prefetch))
9684 /* Loop can be reversed. */
9685 if (loop_dump_stream)
9686 fprintf (loop_dump_stream, "Can reverse loop\n");
9688 /* Now check other conditions:
9690 The increment must be a constant, as must the initial value,
9691 and the comparison code must be LT.
9693 This test can probably be improved since +/- 1 in the constant
9694 can be obtained by changing LT to LE and vice versa; this is
9698 /* for constants, LE gets turned into LT */
9699 && (GET_CODE (comparison) == LT
9700 || (GET_CODE (comparison) == LE
9701 && no_use_except_counting)
9702 || GET_CODE (comparison) == LTU))
9704 HOST_WIDE_INT add_val, add_adjust, comparison_val = 0;
9705 rtx initial_value, comparison_value;
9707 enum rtx_code cmp_code;
9708 int comparison_const_width;
9709 unsigned HOST_WIDE_INT comparison_sign_mask;
9710 bool keep_first_compare;
9712 add_val = INTVAL (bl->biv->add_val);
9713 comparison_value = XEXP (comparison, 1);
9714 if (GET_MODE (comparison_value) == VOIDmode)
9715 comparison_const_width
9716 = GET_MODE_BITSIZE (GET_MODE (XEXP (comparison, 0)));
9718 comparison_const_width
9719 = GET_MODE_BITSIZE (GET_MODE (comparison_value));
9720 if (comparison_const_width > HOST_BITS_PER_WIDE_INT)
9721 comparison_const_width = HOST_BITS_PER_WIDE_INT;
9722 comparison_sign_mask
9723 = (unsigned HOST_WIDE_INT) 1 << (comparison_const_width - 1);
9725 /* If the comparison value is not a loop invariant, then we
9726 can not reverse this loop.
9728 ??? If the insns which initialize the comparison value as
9729 a whole compute an invariant result, then we could move
9730 them out of the loop and proceed with loop reversal. */
9731 if (! loop_invariant_p (loop, comparison_value))
9734 if (GET_CODE (comparison_value) == CONST_INT)
9735 comparison_val = INTVAL (comparison_value);
9736 initial_value = bl->initial_value;
9738 /* Normalize the initial value if it is an integer and
9739 has no other use except as a counter. This will allow
9740 a few more loops to be reversed. */
9741 if (no_use_except_counting
9742 && GET_CODE (comparison_value) == CONST_INT
9743 && GET_CODE (initial_value) == CONST_INT)
9745 comparison_val = comparison_val - INTVAL (bl->initial_value);
9746 /* The code below requires comparison_val to be a multiple
9747 of add_val in order to do the loop reversal, so
9748 round up comparison_val to a multiple of add_val.
9749 Since comparison_value is constant, we know that the
9750 current comparison code is LT. */
9751 comparison_val = comparison_val + add_val - 1;
9753 -= (unsigned HOST_WIDE_INT) comparison_val % add_val;
9754 /* We postpone overflow checks for COMPARISON_VAL here;
9755 even if there is an overflow, we might still be able to
9756 reverse the loop, if converting the loop exit test to
9758 initial_value = const0_rtx;
9761 /* First check if we can do a vanilla loop reversal. */
9762 if (initial_value == const0_rtx
9763 && GET_CODE (comparison_value) == CONST_INT
9764 /* Now do postponed overflow checks on COMPARISON_VAL. */
9765 && ! (((comparison_val - add_val) ^ INTVAL (comparison_value))
9766 & comparison_sign_mask))
9768 /* Register will always be nonnegative, with value
9769 0 on last iteration */
9770 add_adjust = add_val;
9777 if (GET_CODE (comparison) == LE)
9778 add_adjust -= add_val;
9780 /* If the initial value is not zero, or if the comparison
9781 value is not an exact multiple of the increment, then we
9782 can not reverse this loop. */
9783 if (initial_value == const0_rtx
9784 && GET_CODE (comparison_value) == CONST_INT)
9786 if (((unsigned HOST_WIDE_INT) comparison_val % add_val) != 0)
9791 if (! no_use_except_counting || add_val != 1)
9795 final_value = comparison_value;
9797 /* Reset these in case we normalized the initial value
9798 and comparison value above. */
9799 if (GET_CODE (comparison_value) == CONST_INT
9800 && GET_CODE (initial_value) == CONST_INT)
9802 comparison_value = GEN_INT (comparison_val);
9804 = GEN_INT (comparison_val + INTVAL (bl->initial_value));
9806 bl->initial_value = initial_value;
9808 /* Save some info needed to produce the new insns. */
9809 reg = bl->biv->dest_reg;
9810 mode = GET_MODE (reg);
9811 jump_label = condjump_label (PREV_INSN (loop_end));
9812 new_add_val = GEN_INT (-INTVAL (bl->biv->add_val));
9814 /* Set start_value; if this is not a CONST_INT, we need
9816 Initialize biv to start_value before loop start.
9817 The old initializing insn will be deleted as a
9818 dead store by flow.c. */
9819 if (initial_value == const0_rtx
9820 && GET_CODE (comparison_value) == CONST_INT)
9823 = gen_int_mode (comparison_val - add_adjust, mode);
9824 loop_insn_hoist (loop, gen_move_insn (reg, start_value));
9826 else if (GET_CODE (initial_value) == CONST_INT)
9828 rtx offset = GEN_INT (-INTVAL (initial_value) - add_adjust);
9829 rtx add_insn = gen_add3_insn (reg, comparison_value, offset);
9835 = gen_rtx_PLUS (mode, comparison_value, offset);
9836 loop_insn_hoist (loop, add_insn);
9837 if (GET_CODE (comparison) == LE)
9838 final_value = gen_rtx_PLUS (mode, comparison_value,
9841 else if (! add_adjust)
9843 rtx sub_insn = gen_sub3_insn (reg, comparison_value,
9849 = gen_rtx_MINUS (mode, comparison_value, initial_value);
9850 loop_insn_hoist (loop, sub_insn);
9853 /* We could handle the other cases too, but it'll be
9854 better to have a testcase first. */
9857 /* We may not have a single insn which can increment a reg, so
9858 create a sequence to hold all the insns from expand_inc. */
9860 expand_inc (reg, new_add_val);
9864 p = loop_insn_emit_before (loop, 0, bl->biv->insn, tem);
9865 delete_insn (bl->biv->insn);
9867 /* Update biv info to reflect its new status. */
9869 bl->initial_value = start_value;
9870 bl->biv->add_val = new_add_val;
9872 /* Update loop info. */
9873 loop_info->initial_value = reg;
9874 loop_info->initial_equiv_value = reg;
9875 loop_info->final_value = const0_rtx;
9876 loop_info->final_equiv_value = const0_rtx;
9877 loop_info->comparison_value = const0_rtx;
9878 loop_info->comparison_code = cmp_code;
9879 loop_info->increment = new_add_val;
9881 /* Inc LABEL_NUSES so that delete_insn will
9882 not delete the label. */
9883 LABEL_NUSES (XEXP (jump_label, 0))++;
9885 /* If we have a separate comparison insn that does more
9886 than just set cc0, the result of the comparison might
9887 be used outside the loop. */
9888 keep_first_compare = (compare_and_branch == 2
9890 && sets_cc0_p (first_compare) <= 0
9894 /* Emit an insn after the end of the loop to set the biv's
9895 proper exit value if it is used anywhere outside the loop. */
9896 if (keep_first_compare
9897 || (REGNO_LAST_UID (bl->regno) != INSN_UID (first_compare))
9899 || REGNO_FIRST_UID (bl->regno) != INSN_UID (bl->init_insn))
9900 loop_insn_sink (loop, gen_load_of_final_value (reg, final_value));
9902 if (keep_first_compare)
9903 loop_insn_sink (loop, PATTERN (first_compare));
9905 /* Delete compare/branch at end of loop. */
9906 delete_related_insns (PREV_INSN (loop_end));
9907 if (compare_and_branch == 2)
9908 delete_related_insns (first_compare);
9910 /* Add new compare/branch insn at end of loop. */
9912 emit_cmp_and_jump_insns (reg, const0_rtx, cmp_code, NULL_RTX,
9914 XEXP (jump_label, 0));
9917 emit_jump_insn_before (tem, loop_end);
9919 for (tem = PREV_INSN (loop_end);
9920 tem && !JUMP_P (tem);
9921 tem = PREV_INSN (tem))
9925 JUMP_LABEL (tem) = XEXP (jump_label, 0);
9931 /* Increment of LABEL_NUSES done above. */
9932 /* Register is now always nonnegative,
9933 so add REG_NONNEG note to the branch. */
9934 REG_NOTES (tem) = gen_rtx_EXPR_LIST (REG_NONNEG, reg,
9940 /* No insn may reference both the reversed and another biv or it
9941 will fail (see comment near the top of the loop reversal
9943 Earlier on, we have verified that the biv has no use except
9944 counting, or it is the only biv in this function.
9945 However, the code that computes no_use_except_counting does
9946 not verify reg notes. It's possible to have an insn that
9947 references another biv, and has a REG_EQUAL note with an
9948 expression based on the reversed biv. To avoid this case,
9949 remove all REG_EQUAL notes based on the reversed biv
9951 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
9955 rtx set = single_set (p);
9956 /* If this is a set of a GIV based on the reversed biv, any
9957 REG_EQUAL notes should still be correct. */
9959 || !REG_P (SET_DEST (set))
9960 || (size_t) REGNO (SET_DEST (set)) >= ivs->n_regs
9961 || REG_IV_TYPE (ivs, REGNO (SET_DEST (set))) != GENERAL_INDUCT
9962 || REG_IV_INFO (ivs, REGNO (SET_DEST (set)))->src_reg != bl->biv->src_reg)
9963 for (pnote = ®_NOTES (p); *pnote;)
9965 if (REG_NOTE_KIND (*pnote) == REG_EQUAL
9966 && reg_mentioned_p (regno_reg_rtx[bl->regno],
9968 *pnote = XEXP (*pnote, 1);
9970 pnote = &XEXP (*pnote, 1);
9974 /* Mark that this biv has been reversed. Each giv which depends
9975 on this biv, and which is also live past the end of the loop
9976 will have to be fixed up. */
9980 if (loop_dump_stream)
9982 fprintf (loop_dump_stream, "Reversed loop");
9984 fprintf (loop_dump_stream, " and added reg_nonneg\n");
9986 fprintf (loop_dump_stream, "\n");
9997 /* Verify whether the biv BL appears to be eliminable,
9998 based on the insns in the loop that refer to it.
10000 If ELIMINATE_P is nonzero, actually do the elimination.
10002 THRESHOLD and INSN_COUNT are from loop_optimize and are used to
10003 determine whether invariant insns should be placed inside or at the
10004 start of the loop. */
10007 maybe_eliminate_biv (const struct loop *loop, struct iv_class *bl,
10008 int eliminate_p, int threshold, int insn_count)
10010 struct loop_ivs *ivs = LOOP_IVS (loop);
10011 rtx reg = bl->biv->dest_reg;
10014 /* Scan all insns in the loop, stopping if we find one that uses the
10015 biv in a way that we cannot eliminate. */
10017 for (p = loop->start; p != loop->end; p = NEXT_INSN (p))
10019 enum rtx_code code = GET_CODE (p);
10020 basic_block where_bb = 0;
10021 rtx where_insn = threshold >= insn_count ? 0 : p;
10024 /* If this is a libcall that sets a giv, skip ahead to its end. */
10027 note = find_reg_note (p, REG_LIBCALL, NULL_RTX);
10031 rtx last = XEXP (note, 0);
10032 rtx set = single_set (last);
10034 if (set && REG_P (SET_DEST (set)))
10036 unsigned int regno = REGNO (SET_DEST (set));
10038 if (regno < ivs->n_regs
10039 && REG_IV_TYPE (ivs, regno) == GENERAL_INDUCT
10040 && REG_IV_INFO (ivs, regno)->src_reg == bl->biv->src_reg)
10046 /* Closely examine the insn if the biv is mentioned. */
10047 if ((code == INSN || code == JUMP_INSN || code == CALL_INSN)
10048 && reg_mentioned_p (reg, PATTERN (p))
10049 && ! maybe_eliminate_biv_1 (loop, PATTERN (p), p, bl,
10050 eliminate_p, where_bb, where_insn))
10052 if (loop_dump_stream)
10053 fprintf (loop_dump_stream,
10054 "Cannot eliminate biv %d: biv used in insn %d.\n",
10055 bl->regno, INSN_UID (p));
10059 /* If we are eliminating, kill REG_EQUAL notes mentioning the biv. */
10061 && (note = find_reg_note (p, REG_EQUAL, NULL_RTX)) != NULL_RTX
10062 && reg_mentioned_p (reg, XEXP (note, 0)))
10063 remove_note (p, note);
10066 if (p == loop->end)
10068 if (loop_dump_stream)
10069 fprintf (loop_dump_stream, "biv %d %s eliminated.\n",
10070 bl->regno, eliminate_p ? "was" : "can be");
10077 /* INSN and REFERENCE are instructions in the same insn chain.
10078 Return nonzero if INSN is first. */
10081 loop_insn_first_p (rtx insn, rtx reference)
10085 for (p = insn, q = reference;;)
10087 /* Start with test for not first so that INSN == REFERENCE yields not
10089 if (q == insn || ! p)
10091 if (p == reference || ! q)
10094 /* Either of P or Q might be a NOTE. Notes have the same LUID as the
10095 previous insn, hence the <= comparison below does not work if
10097 if (INSN_UID (p) < max_uid_for_loop
10098 && INSN_UID (q) < max_uid_for_loop
10100 return INSN_LUID (p) <= INSN_LUID (q);
10102 if (INSN_UID (p) >= max_uid_for_loop
10105 if (INSN_UID (q) >= max_uid_for_loop)
10110 /* We are trying to eliminate BIV in INSN using GIV. Return nonzero if
10111 the offset that we have to take into account due to auto-increment /
10112 div derivation is zero. */
10114 biv_elimination_giv_has_0_offset (struct induction *biv,
10115 struct induction *giv, rtx insn)
10117 /* If the giv V had the auto-inc address optimization applied
10118 to it, and INSN occurs between the giv insn and the biv
10119 insn, then we'd have to adjust the value used here.
10120 This is rare, so we don't bother to make this possible. */
10121 if (giv->auto_inc_opt
10122 && ((loop_insn_first_p (giv->insn, insn)
10123 && loop_insn_first_p (insn, biv->insn))
10124 || (loop_insn_first_p (biv->insn, insn)
10125 && loop_insn_first_p (insn, giv->insn))))
10131 /* If BL appears in X (part of the pattern of INSN), see if we can
10132 eliminate its use. If so, return 1. If not, return 0.
10134 If BIV does not appear in X, return 1.
10136 If ELIMINATE_P is nonzero, actually do the elimination.
10137 WHERE_INSN/WHERE_BB indicate where extra insns should be added.
10138 Depending on how many items have been moved out of the loop, it
10139 will either be before INSN (when WHERE_INSN is nonzero) or at the
10140 start of the loop (when WHERE_INSN is zero). */
10143 maybe_eliminate_biv_1 (const struct loop *loop, rtx x, rtx insn,
10144 struct iv_class *bl, int eliminate_p,
10145 basic_block where_bb, rtx where_insn)
10147 enum rtx_code code = GET_CODE (x);
10148 rtx reg = bl->biv->dest_reg;
10149 enum machine_mode mode = GET_MODE (reg);
10150 struct induction *v;
10162 /* If we haven't already been able to do something with this BIV,
10163 we can't eliminate it. */
10169 /* If this sets the BIV, it is not a problem. */
10170 if (SET_DEST (x) == reg)
10173 /* If this is an insn that defines a giv, it is also ok because
10174 it will go away when the giv is reduced. */
10175 for (v = bl->giv; v; v = v->next_iv)
10176 if (v->giv_type == DEST_REG && SET_DEST (x) == v->dest_reg)
10180 if (SET_DEST (x) == cc0_rtx && SET_SRC (x) == reg)
10182 /* Can replace with any giv that was reduced and
10183 that has (MULT_VAL != 0) and (ADD_VAL == 0).
10184 Require a constant for MULT_VAL, so we know it's nonzero.
10185 ??? We disable this optimization to avoid potential
10188 for (v = bl->giv; v; v = v->next_iv)
10189 if (GET_CODE (v->mult_val) == CONST_INT && v->mult_val != const0_rtx
10190 && v->add_val == const0_rtx
10191 && ! v->ignore && ! v->maybe_dead && v->always_computable
10195 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
10201 /* If the giv has the opposite direction of change,
10202 then reverse the comparison. */
10203 if (INTVAL (v->mult_val) < 0)
10204 new = gen_rtx_COMPARE (GET_MODE (v->new_reg),
10205 const0_rtx, v->new_reg);
10209 /* We can probably test that giv's reduced reg. */
10210 if (validate_change (insn, &SET_SRC (x), new, 0))
10214 /* Look for a giv with (MULT_VAL != 0) and (ADD_VAL != 0);
10215 replace test insn with a compare insn (cmp REDUCED_GIV ADD_VAL).
10216 Require a constant for MULT_VAL, so we know it's nonzero.
10217 ??? Do this only if ADD_VAL is a pointer to avoid a potential
10218 overflow problem. */
10220 for (v = bl->giv; v; v = v->next_iv)
10221 if (GET_CODE (v->mult_val) == CONST_INT
10222 && v->mult_val != const0_rtx
10223 && ! v->ignore && ! v->maybe_dead && v->always_computable
10225 && (GET_CODE (v->add_val) == SYMBOL_REF
10226 || GET_CODE (v->add_val) == LABEL_REF
10227 || GET_CODE (v->add_val) == CONST
10228 || (REG_P (v->add_val)
10229 && REG_POINTER (v->add_val))))
10231 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
10237 /* If the giv has the opposite direction of change,
10238 then reverse the comparison. */
10239 if (INTVAL (v->mult_val) < 0)
10240 new = gen_rtx_COMPARE (VOIDmode, copy_rtx (v->add_val),
10243 new = gen_rtx_COMPARE (VOIDmode, v->new_reg,
10244 copy_rtx (v->add_val));
10246 /* Replace biv with the giv's reduced register. */
10247 update_reg_last_use (v->add_val, insn);
10248 if (validate_change (insn, &SET_SRC (PATTERN (insn)), new, 0))
10251 /* Insn doesn't support that constant or invariant. Copy it
10252 into a register (it will be a loop invariant.) */
10253 tem = gen_reg_rtx (GET_MODE (v->new_reg));
10255 loop_insn_emit_before (loop, 0, where_insn,
10256 gen_move_insn (tem,
10257 copy_rtx (v->add_val)));
10259 /* Substitute the new register for its invariant value in
10260 the compare expression. */
10261 XEXP (new, (INTVAL (v->mult_val) < 0) ? 0 : 1) = tem;
10262 if (validate_change (insn, &SET_SRC (PATTERN (insn)), new, 0))
10271 case GT: case GE: case GTU: case GEU:
10272 case LT: case LE: case LTU: case LEU:
10273 /* See if either argument is the biv. */
10274 if (XEXP (x, 0) == reg)
10275 arg = XEXP (x, 1), arg_operand = 1;
10276 else if (XEXP (x, 1) == reg)
10277 arg = XEXP (x, 0), arg_operand = 0;
10281 if (GET_CODE (arg) != CONST_INT)
10284 /* Unless we're dealing with an equality comparison, if we can't
10285 determine that the original biv doesn't wrap, then we must not
10286 apply the transformation. */
10287 /* ??? Actually, what we must do is verify that the transformed
10288 giv doesn't wrap. But the general case of this transformation
10289 was disabled long ago due to wrapping problems, and there's no
10290 point reviving it this close to end-of-life for loop.c. The
10291 only case still enabled is known (via the check on add_val) to
10292 be pointer arithmetic, which in theory never overflows for
10294 /* Without lifetime analysis, we don't know how COMPARE will be
10295 used, so we must assume the worst. */
10296 if (code != EQ && code != NE
10297 && biased_biv_may_wrap_p (loop, bl, INTVAL (arg)))
10300 /* Try to replace with any giv that has constant positive mult_val
10301 and a pointer add_val. */
10302 for (v = bl->giv; v; v = v->next_iv)
10303 if (GET_CODE (v->mult_val) == CONST_INT
10304 && INTVAL (v->mult_val) > 0
10305 && (GET_CODE (v->add_val) == SYMBOL_REF
10306 || GET_CODE (v->add_val) == LABEL_REF
10307 || GET_CODE (v->add_val) == CONST
10308 || (REG_P (v->add_val) && REG_POINTER (v->add_val)))
10309 && ! v->ignore && ! v->maybe_dead && v->always_computable
10310 && v->mode == mode)
10312 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
10318 /* Replace biv with the giv's reduced reg. */
10319 validate_change (insn, &XEXP (x, 1 - arg_operand), v->new_reg, 1);
10321 /* Load the value into a register. */
10322 tem = gen_reg_rtx (mode);
10323 loop_iv_add_mult_emit_before (loop, arg, v->mult_val, v->add_val,
10324 tem, where_bb, where_insn);
10326 validate_change (insn, &XEXP (x, arg_operand), tem, 1);
10328 if (apply_change_group ())
10332 /* If we get here, the biv can't be eliminated. */
10336 /* If this address is a DEST_ADDR giv, it doesn't matter if the
10337 biv is used in it, since it will be replaced. */
10338 for (v = bl->giv; v; v = v->next_iv)
10339 if (v->giv_type == DEST_ADDR && v->location == &XEXP (x, 0))
10347 /* See if any subexpression fails elimination. */
10348 fmt = GET_RTX_FORMAT (code);
10349 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
10354 if (! maybe_eliminate_biv_1 (loop, XEXP (x, i), insn, bl,
10355 eliminate_p, where_bb, where_insn))
10360 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
10361 if (! maybe_eliminate_biv_1 (loop, XVECEXP (x, i, j), insn, bl,
10362 eliminate_p, where_bb, where_insn))
10371 /* Return nonzero if the last use of REG
10372 is in an insn following INSN in the same basic block. */
10375 last_use_this_basic_block (rtx reg, rtx insn)
10379 n && !LABEL_P (n) && !JUMP_P (n);
10382 if (REGNO_LAST_UID (REGNO (reg)) == INSN_UID (n))
10388 /* Called via `note_stores' to record the initial value of a biv. Here we
10389 just record the location of the set and process it later. */
10392 record_initial (rtx dest, rtx set, void *data ATTRIBUTE_UNUSED)
10394 struct loop_ivs *ivs = (struct loop_ivs *) data;
10395 struct iv_class *bl;
10398 || REGNO (dest) >= ivs->n_regs
10399 || REG_IV_TYPE (ivs, REGNO (dest)) != BASIC_INDUCT)
10402 bl = REG_IV_CLASS (ivs, REGNO (dest));
10404 /* If this is the first set found, record it. */
10405 if (bl->init_insn == 0)
10407 bl->init_insn = note_insn;
10408 bl->init_set = set;
10412 /* If any of the registers in X are "old" and currently have a last use earlier
10413 than INSN, update them to have a last use of INSN. Their actual last use
10414 will be the previous insn but it will not have a valid uid_luid so we can't
10415 use it. X must be a source expression only. */
10418 update_reg_last_use (rtx x, rtx insn)
10420 /* Check for the case where INSN does not have a valid luid. In this case,
10421 there is no need to modify the regno_last_uid, as this can only happen
10422 when code is inserted after the loop_end to set a pseudo's final value,
10423 and hence this insn will never be the last use of x.
10424 ???? This comment is not correct. See for example loop_givs_reduce.
10425 This may insert an insn before another new insn. */
10426 if (REG_P (x) && REGNO (x) < max_reg_before_loop
10427 && INSN_UID (insn) < max_uid_for_loop
10428 && REGNO_LAST_LUID (REGNO (x)) < INSN_LUID (insn))
10430 REGNO_LAST_UID (REGNO (x)) = INSN_UID (insn);
10435 const char *fmt = GET_RTX_FORMAT (GET_CODE (x));
10436 for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
10439 update_reg_last_use (XEXP (x, i), insn);
10440 else if (fmt[i] == 'E')
10441 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
10442 update_reg_last_use (XVECEXP (x, i, j), insn);
10447 /* Similar to rtlanal.c:get_condition, except that we also put an
10448 invariant last unless both operands are invariants. */
10451 get_condition_for_loop (const struct loop *loop, rtx x)
10453 rtx comparison = get_condition (x, (rtx*) 0, false, true);
10455 if (comparison == 0
10456 || ! loop_invariant_p (loop, XEXP (comparison, 0))
10457 || loop_invariant_p (loop, XEXP (comparison, 1)))
10460 return gen_rtx_fmt_ee (swap_condition (GET_CODE (comparison)), VOIDmode,
10461 XEXP (comparison, 1), XEXP (comparison, 0));
10464 /* Scan the function and determine whether it has indirect (computed) jumps.
10466 This is taken mostly from flow.c; similar code exists elsewhere
10467 in the compiler. It may be useful to put this into rtlanal.c. */
10469 indirect_jump_in_function_p (rtx start)
10473 for (insn = start; insn; insn = NEXT_INSN (insn))
10474 if (computed_jump_p (insn))
10480 /* Add MEM to the LOOP_MEMS array, if appropriate. See the
10481 documentation for LOOP_MEMS for the definition of `appropriate'.
10482 This function is called from prescan_loop via for_each_rtx. */
10485 insert_loop_mem (rtx *mem, void *data ATTRIBUTE_UNUSED)
10487 struct loop_info *loop_info = data;
10494 switch (GET_CODE (m))
10500 /* We're not interested in MEMs that are only clobbered. */
10504 /* We're not interested in the MEM associated with a
10505 CONST_DOUBLE, so there's no need to traverse into this. */
10509 /* We're not interested in any MEMs that only appear in notes. */
10513 /* This is not a MEM. */
10517 /* See if we've already seen this MEM. */
10518 for (i = 0; i < loop_info->mems_idx; ++i)
10519 if (rtx_equal_p (m, loop_info->mems[i].mem))
10521 if (MEM_VOLATILE_P (m) && !MEM_VOLATILE_P (loop_info->mems[i].mem))
10522 loop_info->mems[i].mem = m;
10523 if (GET_MODE (m) != GET_MODE (loop_info->mems[i].mem))
10524 /* The modes of the two memory accesses are different. If
10525 this happens, something tricky is going on, and we just
10526 don't optimize accesses to this MEM. */
10527 loop_info->mems[i].optimize = 0;
10532 /* Resize the array, if necessary. */
10533 if (loop_info->mems_idx == loop_info->mems_allocated)
10535 if (loop_info->mems_allocated != 0)
10536 loop_info->mems_allocated *= 2;
10538 loop_info->mems_allocated = 32;
10540 loop_info->mems = xrealloc (loop_info->mems,
10541 loop_info->mems_allocated * sizeof (loop_mem_info));
10544 /* Actually insert the MEM. */
10545 loop_info->mems[loop_info->mems_idx].mem = m;
10546 /* We can't hoist this MEM out of the loop if it's a BLKmode MEM
10547 because we can't put it in a register. We still store it in the
10548 table, though, so that if we see the same address later, but in a
10549 non-BLK mode, we'll not think we can optimize it at that point. */
10550 loop_info->mems[loop_info->mems_idx].optimize = (GET_MODE (m) != BLKmode);
10551 loop_info->mems[loop_info->mems_idx].reg = NULL_RTX;
10552 ++loop_info->mems_idx;
10558 /* Allocate REGS->ARRAY or reallocate it if it is too small.
10560 Increment REGS->ARRAY[I].SET_IN_LOOP at the index I of each
10561 register that is modified by an insn between FROM and TO. If the
10562 value of an element of REGS->array[I].SET_IN_LOOP becomes 127 or
10563 more, stop incrementing it, to avoid overflow.
10565 Store in REGS->ARRAY[I].SINGLE_USAGE the single insn in which
10566 register I is used, if it is only used once. Otherwise, it is set
10567 to 0 (for no uses) or const0_rtx for more than one use. This
10568 parameter may be zero, in which case this processing is not done.
10570 Set REGS->ARRAY[I].MAY_NOT_OPTIMIZE nonzero if we should not
10571 optimize register I. */
10574 loop_regs_scan (const struct loop *loop, int extra_size)
10576 struct loop_regs *regs = LOOP_REGS (loop);
10578 /* last_set[n] is nonzero iff reg n has been set in the current
10579 basic block. In that case, it is the insn that last set reg n. */
10584 old_nregs = regs->num;
10585 regs->num = max_reg_num ();
10587 /* Grow the regs array if not allocated or too small. */
10588 if (regs->num >= regs->size)
10590 regs->size = regs->num + extra_size;
10592 regs->array = xrealloc (regs->array, regs->size * sizeof (*regs->array));
10594 /* Zero the new elements. */
10595 memset (regs->array + old_nregs, 0,
10596 (regs->size - old_nregs) * sizeof (*regs->array));
10599 /* Clear previously scanned fields but do not clear n_times_set. */
10600 for (i = 0; i < old_nregs; i++)
10602 regs->array[i].set_in_loop = 0;
10603 regs->array[i].may_not_optimize = 0;
10604 regs->array[i].single_usage = NULL_RTX;
10607 last_set = xcalloc (regs->num, sizeof (rtx));
10609 /* Scan the loop, recording register usage. */
10610 for (insn = loop->top ? loop->top : loop->start; insn != loop->end;
10611 insn = NEXT_INSN (insn))
10615 /* Record registers that have exactly one use. */
10616 find_single_use_in_loop (regs, insn, PATTERN (insn));
10618 /* Include uses in REG_EQUAL notes. */
10619 if (REG_NOTES (insn))
10620 find_single_use_in_loop (regs, insn, REG_NOTES (insn));
10622 if (GET_CODE (PATTERN (insn)) == SET
10623 || GET_CODE (PATTERN (insn)) == CLOBBER)
10624 count_one_set (regs, insn, PATTERN (insn), last_set);
10625 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
10628 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
10629 count_one_set (regs, insn, XVECEXP (PATTERN (insn), 0, i),
10634 if (LABEL_P (insn) || JUMP_P (insn))
10635 memset (last_set, 0, regs->num * sizeof (rtx));
10637 /* Invalidate all registers used for function argument passing.
10638 We check rtx_varies_p for the same reason as below, to allow
10639 optimizing PIC calculations. */
10643 for (link = CALL_INSN_FUNCTION_USAGE (insn);
10645 link = XEXP (link, 1))
10649 if (GET_CODE (op = XEXP (link, 0)) == USE
10650 && REG_P (reg = XEXP (op, 0))
10651 && rtx_varies_p (reg, 1))
10652 regs->array[REGNO (reg)].may_not_optimize = 1;
10657 /* Invalidate all hard registers clobbered by calls. With one exception:
10658 a call-clobbered PIC register is still function-invariant for our
10659 purposes, since we can hoist any PIC calculations out of the loop.
10660 Thus the call to rtx_varies_p. */
10661 if (LOOP_INFO (loop)->has_call)
10662 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
10663 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, i)
10664 && rtx_varies_p (regno_reg_rtx[i], 1))
10666 regs->array[i].may_not_optimize = 1;
10667 regs->array[i].set_in_loop = 1;
10670 #ifdef AVOID_CCMODE_COPIES
10671 /* Don't try to move insns which set CC registers if we should not
10672 create CCmode register copies. */
10673 for (i = regs->num - 1; i >= FIRST_PSEUDO_REGISTER; i--)
10674 if (GET_MODE_CLASS (GET_MODE (regno_reg_rtx[i])) == MODE_CC)
10675 regs->array[i].may_not_optimize = 1;
10678 /* Set regs->array[I].n_times_set for the new registers. */
10679 for (i = old_nregs; i < regs->num; i++)
10680 regs->array[i].n_times_set = regs->array[i].set_in_loop;
10685 /* Returns the number of real INSNs in the LOOP. */
10688 count_insns_in_loop (const struct loop *loop)
10693 for (insn = loop->top ? loop->top : loop->start; insn != loop->end;
10694 insn = NEXT_INSN (insn))
10701 /* Move MEMs into registers for the duration of the loop. */
10704 load_mems (const struct loop *loop)
10706 struct loop_info *loop_info = LOOP_INFO (loop);
10707 struct loop_regs *regs = LOOP_REGS (loop);
10708 int maybe_never = 0;
10710 rtx p, prev_ebb_head;
10711 rtx label = NULL_RTX;
10713 /* Nonzero if the next instruction may never be executed. */
10714 int next_maybe_never = 0;
10715 unsigned int last_max_reg = max_reg_num ();
10717 if (loop_info->mems_idx == 0)
10720 /* We cannot use next_label here because it skips over normal insns. */
10721 end_label = next_nonnote_insn (loop->end);
10722 if (end_label && !LABEL_P (end_label))
10723 end_label = NULL_RTX;
10725 /* Check to see if it's possible that some instructions in the loop are
10726 never executed. Also check if there is a goto out of the loop other
10727 than right after the end of the loop. */
10728 for (p = next_insn_in_loop (loop, loop->scan_start);
10730 p = next_insn_in_loop (loop, p))
10734 else if (JUMP_P (p)
10735 /* If we enter the loop in the middle, and scan
10736 around to the beginning, don't set maybe_never
10737 for that. This must be an unconditional jump,
10738 otherwise the code at the top of the loop might
10739 never be executed. Unconditional jumps are
10740 followed a by barrier then loop end. */
10742 && JUMP_LABEL (p) == loop->top
10743 && NEXT_INSN (NEXT_INSN (p)) == loop->end
10744 && any_uncondjump_p (p)))
10746 /* If this is a jump outside of the loop but not right
10747 after the end of the loop, we would have to emit new fixup
10748 sequences for each such label. */
10749 if (/* If we can't tell where control might go when this
10750 JUMP_INSN is executed, we must be conservative. */
10752 || (JUMP_LABEL (p) != end_label
10753 && (INSN_UID (JUMP_LABEL (p)) >= max_uid_for_loop
10754 || INSN_LUID (JUMP_LABEL (p)) < INSN_LUID (loop->start)
10755 || INSN_LUID (JUMP_LABEL (p)) > INSN_LUID (loop->end))))
10758 if (!any_condjump_p (p))
10759 /* Something complicated. */
10762 /* If there are any more instructions in the loop, they
10763 might not be reached. */
10764 next_maybe_never = 1;
10766 else if (next_maybe_never)
10770 /* Find start of the extended basic block that enters the loop. */
10771 for (p = loop->start;
10772 PREV_INSN (p) && !LABEL_P (p);
10777 cselib_init (true);
10779 /* Build table of mems that get set to constant values before the
10781 for (; p != loop->start; p = NEXT_INSN (p))
10782 cselib_process_insn (p);
10784 /* Actually move the MEMs. */
10785 for (i = 0; i < loop_info->mems_idx; ++i)
10787 regset_head load_copies;
10788 regset_head store_copies;
10791 rtx mem = loop_info->mems[i].mem;
10792 rtx mem_list_entry;
10794 if (MEM_VOLATILE_P (mem)
10795 || loop_invariant_p (loop, XEXP (mem, 0)) != 1)
10796 /* There's no telling whether or not MEM is modified. */
10797 loop_info->mems[i].optimize = 0;
10799 /* Go through the MEMs written to in the loop to see if this
10800 one is aliased by one of them. */
10801 mem_list_entry = loop_info->store_mems;
10802 while (mem_list_entry)
10804 if (rtx_equal_p (mem, XEXP (mem_list_entry, 0)))
10806 else if (true_dependence (XEXP (mem_list_entry, 0), VOIDmode,
10807 mem, rtx_varies_p))
10809 /* MEM is indeed aliased by this store. */
10810 loop_info->mems[i].optimize = 0;
10813 mem_list_entry = XEXP (mem_list_entry, 1);
10816 if (flag_float_store && written
10817 && GET_MODE_CLASS (GET_MODE (mem)) == MODE_FLOAT)
10818 loop_info->mems[i].optimize = 0;
10820 /* If this MEM is written to, we must be sure that there
10821 are no reads from another MEM that aliases this one. */
10822 if (loop_info->mems[i].optimize && written)
10826 for (j = 0; j < loop_info->mems_idx; ++j)
10830 else if (true_dependence (mem,
10832 loop_info->mems[j].mem,
10835 /* It's not safe to hoist loop_info->mems[i] out of
10836 the loop because writes to it might not be
10837 seen by reads from loop_info->mems[j]. */
10838 loop_info->mems[i].optimize = 0;
10844 if (maybe_never && may_trap_p (mem))
10845 /* We can't access the MEM outside the loop; it might
10846 cause a trap that wouldn't have happened otherwise. */
10847 loop_info->mems[i].optimize = 0;
10849 if (!loop_info->mems[i].optimize)
10850 /* We thought we were going to lift this MEM out of the
10851 loop, but later discovered that we could not. */
10854 INIT_REG_SET (&load_copies);
10855 INIT_REG_SET (&store_copies);
10857 /* Allocate a pseudo for this MEM. We set REG_USERVAR_P in
10858 order to keep scan_loop from moving stores to this MEM
10859 out of the loop just because this REG is neither a
10860 user-variable nor used in the loop test. */
10861 reg = gen_reg_rtx (GET_MODE (mem));
10862 REG_USERVAR_P (reg) = 1;
10863 loop_info->mems[i].reg = reg;
10865 /* Now, replace all references to the MEM with the
10866 corresponding pseudos. */
10868 for (p = next_insn_in_loop (loop, loop->scan_start);
10870 p = next_insn_in_loop (loop, p))
10876 set = single_set (p);
10878 /* See if this copies the mem into a register that isn't
10879 modified afterwards. We'll try to do copy propagation
10880 a little further on. */
10882 /* @@@ This test is _way_ too conservative. */
10884 && REG_P (SET_DEST (set))
10885 && REGNO (SET_DEST (set)) >= FIRST_PSEUDO_REGISTER
10886 && REGNO (SET_DEST (set)) < last_max_reg
10887 && regs->array[REGNO (SET_DEST (set))].n_times_set == 1
10888 && rtx_equal_p (SET_SRC (set), mem))
10889 SET_REGNO_REG_SET (&load_copies, REGNO (SET_DEST (set)));
10891 /* See if this copies the mem from a register that isn't
10892 modified afterwards. We'll try to remove the
10893 redundant copy later on by doing a little register
10894 renaming and copy propagation. This will help
10895 to untangle things for the BIV detection code. */
10898 && REG_P (SET_SRC (set))
10899 && REGNO (SET_SRC (set)) >= FIRST_PSEUDO_REGISTER
10900 && REGNO (SET_SRC (set)) < last_max_reg
10901 && regs->array[REGNO (SET_SRC (set))].n_times_set == 1
10902 && rtx_equal_p (SET_DEST (set), mem))
10903 SET_REGNO_REG_SET (&store_copies, REGNO (SET_SRC (set)));
10905 /* If this is a call which uses / clobbers this memory
10906 location, we must not change the interface here. */
10908 && reg_mentioned_p (loop_info->mems[i].mem,
10909 CALL_INSN_FUNCTION_USAGE (p)))
10911 cancel_changes (0);
10912 loop_info->mems[i].optimize = 0;
10916 /* Replace the memory reference with the shadow register. */
10917 replace_loop_mems (p, loop_info->mems[i].mem,
10918 loop_info->mems[i].reg, written);
10926 if (! loop_info->mems[i].optimize)
10927 ; /* We found we couldn't do the replacement, so do nothing. */
10928 else if (! apply_change_group ())
10929 /* We couldn't replace all occurrences of the MEM. */
10930 loop_info->mems[i].optimize = 0;
10933 /* Load the memory immediately before LOOP->START, which is
10934 the NOTE_LOOP_BEG. */
10935 cselib_val *e = cselib_lookup (mem, VOIDmode, 0);
10939 struct elt_loc_list *const_equiv = 0;
10940 reg_set_iterator rsi;
10944 struct elt_loc_list *equiv;
10945 struct elt_loc_list *best_equiv = 0;
10946 for (equiv = e->locs; equiv; equiv = equiv->next)
10948 if (CONSTANT_P (equiv->loc))
10949 const_equiv = equiv;
10950 else if (REG_P (equiv->loc)
10951 /* Extending hard register lifetimes causes crash
10952 on SRC targets. Doing so on non-SRC is
10953 probably also not good idea, since we most
10954 probably have pseudoregister equivalence as
10956 && REGNO (equiv->loc) >= FIRST_PSEUDO_REGISTER)
10957 best_equiv = equiv;
10959 /* Use the constant equivalence if that is cheap enough. */
10961 best_equiv = const_equiv;
10962 else if (const_equiv
10963 && (rtx_cost (const_equiv->loc, SET)
10964 <= rtx_cost (best_equiv->loc, SET)))
10966 best_equiv = const_equiv;
10970 /* If best_equiv is nonzero, we know that MEM is set to a
10971 constant or register before the loop. We will use this
10972 knowledge to initialize the shadow register with that
10973 constant or reg rather than by loading from MEM. */
10975 best = copy_rtx (best_equiv->loc);
10978 set = gen_move_insn (reg, best);
10979 set = loop_insn_hoist (loop, set);
10982 for (p = prev_ebb_head; p != loop->start; p = NEXT_INSN (p))
10983 if (REGNO_LAST_UID (REGNO (best)) == INSN_UID (p))
10985 REGNO_LAST_UID (REGNO (best)) = INSN_UID (set);
10991 set_unique_reg_note (set, REG_EQUAL, copy_rtx (const_equiv->loc));
10995 if (label == NULL_RTX)
10997 label = gen_label_rtx ();
10998 emit_label_after (label, loop->end);
11001 /* Store the memory immediately after END, which is
11002 the NOTE_LOOP_END. */
11003 set = gen_move_insn (copy_rtx (mem), reg);
11004 loop_insn_emit_after (loop, 0, label, set);
11007 if (loop_dump_stream)
11009 fprintf (loop_dump_stream, "Hoisted regno %d %s from ",
11010 REGNO (reg), (written ? "r/w" : "r/o"));
11011 print_rtl (loop_dump_stream, mem);
11012 fputc ('\n', loop_dump_stream);
11015 /* Attempt a bit of copy propagation. This helps untangle the
11016 data flow, and enables {basic,general}_induction_var to find
11018 EXECUTE_IF_SET_IN_REG_SET
11019 (&load_copies, FIRST_PSEUDO_REGISTER, j, rsi)
11021 try_copy_prop (loop, reg, j);
11023 CLEAR_REG_SET (&load_copies);
11025 EXECUTE_IF_SET_IN_REG_SET
11026 (&store_copies, FIRST_PSEUDO_REGISTER, j, rsi)
11028 try_swap_copy_prop (loop, reg, j);
11030 CLEAR_REG_SET (&store_copies);
11034 /* Now, we need to replace all references to the previous exit
11035 label with the new one. */
11036 if (label != NULL_RTX && end_label != NULL_RTX)
11037 for (p = loop->start; p != loop->end; p = NEXT_INSN (p))
11038 if (JUMP_P (p) && JUMP_LABEL (p) == end_label)
11039 redirect_jump (p, label, false);
11044 /* For communication between note_reg_stored and its caller. */
11045 struct note_reg_stored_arg
11051 /* Called via note_stores, record in SET_SEEN whether X, which is written,
11052 is equal to ARG. */
11054 note_reg_stored (rtx x, rtx setter ATTRIBUTE_UNUSED, void *arg)
11056 struct note_reg_stored_arg *t = (struct note_reg_stored_arg *) arg;
11061 /* Try to replace every occurrence of pseudo REGNO with REPLACEMENT.
11062 There must be exactly one insn that sets this pseudo; it will be
11063 deleted if all replacements succeed and we can prove that the register
11064 is not used after the loop. */
11067 try_copy_prop (const struct loop *loop, rtx replacement, unsigned int regno)
11069 /* This is the reg that we are copying from. */
11070 rtx reg_rtx = regno_reg_rtx[regno];
11073 /* These help keep track of whether we replaced all uses of the reg. */
11074 int replaced_last = 0;
11075 int store_is_first = 0;
11077 for (insn = next_insn_in_loop (loop, loop->scan_start);
11079 insn = next_insn_in_loop (loop, insn))
11083 /* Only substitute within one extended basic block from the initializing
11085 if (LABEL_P (insn) && init_insn)
11088 if (! INSN_P (insn))
11091 /* Is this the initializing insn? */
11092 set = single_set (insn);
11094 && REG_P (SET_DEST (set))
11095 && REGNO (SET_DEST (set)) == regno)
11097 gcc_assert (!init_insn);
11100 if (REGNO_FIRST_UID (regno) == INSN_UID (insn))
11101 store_is_first = 1;
11104 /* Only substitute after seeing the initializing insn. */
11105 if (init_insn && insn != init_insn)
11107 struct note_reg_stored_arg arg;
11109 replace_loop_regs (insn, reg_rtx, replacement);
11110 if (REGNO_LAST_UID (regno) == INSN_UID (insn))
11113 /* Stop replacing when REPLACEMENT is modified. */
11114 arg.reg = replacement;
11116 note_stores (PATTERN (insn), note_reg_stored, &arg);
11119 rtx note = find_reg_note (insn, REG_EQUAL, NULL);
11121 /* It is possible that we've turned previously valid REG_EQUAL to
11122 invalid, as we change the REGNO to REPLACEMENT and unlike REGNO,
11123 REPLACEMENT is modified, we get different meaning. */
11124 if (note && reg_mentioned_p (replacement, XEXP (note, 0)))
11125 remove_note (insn, note);
11130 gcc_assert (init_insn);
11131 if (apply_change_group ())
11133 if (loop_dump_stream)
11134 fprintf (loop_dump_stream, " Replaced reg %d", regno);
11135 if (store_is_first && replaced_last)
11140 /* Assume we're just deleting INIT_INSN. */
11142 /* Look for REG_RETVAL note. If we're deleting the end of
11143 the libcall sequence, the whole sequence can go. */
11144 retval_note = find_reg_note (init_insn, REG_RETVAL, NULL_RTX);
11145 /* If we found a REG_RETVAL note, find the first instruction
11146 in the sequence. */
11148 first = XEXP (retval_note, 0);
11150 /* Delete the instructions. */
11151 loop_delete_insns (first, init_insn);
11153 if (loop_dump_stream)
11154 fprintf (loop_dump_stream, ".\n");
11158 /* Replace all the instructions from FIRST up to and including LAST
11159 with NOTE_INSN_DELETED notes. */
11162 loop_delete_insns (rtx first, rtx last)
11166 if (loop_dump_stream)
11167 fprintf (loop_dump_stream, ", deleting init_insn (%d)",
11169 delete_insn (first);
11171 /* If this was the LAST instructions we're supposed to delete,
11176 first = NEXT_INSN (first);
11180 /* Try to replace occurrences of pseudo REGNO with REPLACEMENT within
11181 loop LOOP if the order of the sets of these registers can be
11182 swapped. There must be exactly one insn within the loop that sets
11183 this pseudo followed immediately by a move insn that sets
11184 REPLACEMENT with REGNO. */
11186 try_swap_copy_prop (const struct loop *loop, rtx replacement,
11187 unsigned int regno)
11190 rtx set = NULL_RTX;
11191 unsigned int new_regno;
11193 new_regno = REGNO (replacement);
11195 for (insn = next_insn_in_loop (loop, loop->scan_start);
11197 insn = next_insn_in_loop (loop, insn))
11199 /* Search for the insn that copies REGNO to NEW_REGNO? */
11201 && (set = single_set (insn))
11202 && REG_P (SET_DEST (set))
11203 && REGNO (SET_DEST (set)) == new_regno
11204 && REG_P (SET_SRC (set))
11205 && REGNO (SET_SRC (set)) == regno)
11209 if (insn != NULL_RTX)
11214 /* Some DEF-USE info would come in handy here to make this
11215 function more general. For now, just check the previous insn
11216 which is the most likely candidate for setting REGNO. */
11218 prev_insn = PREV_INSN (insn);
11221 && (prev_set = single_set (prev_insn))
11222 && REG_P (SET_DEST (prev_set))
11223 && REGNO (SET_DEST (prev_set)) == regno)
11226 (set (reg regno) (expr))
11227 (set (reg new_regno) (reg regno))
11229 so try converting this to:
11230 (set (reg new_regno) (expr))
11231 (set (reg regno) (reg new_regno))
11233 The former construct is often generated when a global
11234 variable used for an induction variable is shadowed by a
11235 register (NEW_REGNO). The latter construct improves the
11236 chances of GIV replacement and BIV elimination. */
11238 validate_change (prev_insn, &SET_DEST (prev_set),
11240 validate_change (insn, &SET_DEST (set),
11242 validate_change (insn, &SET_SRC (set),
11245 if (apply_change_group ())
11247 if (loop_dump_stream)
11248 fprintf (loop_dump_stream,
11249 " Swapped set of reg %d at %d with reg %d at %d.\n",
11250 regno, INSN_UID (insn),
11251 new_regno, INSN_UID (prev_insn));
11253 /* Update first use of REGNO. */
11254 if (REGNO_FIRST_UID (regno) == INSN_UID (prev_insn))
11255 REGNO_FIRST_UID (regno) = INSN_UID (insn);
11257 /* Now perform copy propagation to hopefully
11258 remove all uses of REGNO within the loop. */
11259 try_copy_prop (loop, replacement, regno);
11265 /* Worker function for find_mem_in_note, called via for_each_rtx. */
11268 find_mem_in_note_1 (rtx *x, void *data)
11270 if (*x != NULL_RTX && MEM_P (*x))
11272 rtx *res = (rtx *) data;
11279 /* Returns the first MEM found in NOTE by depth-first search. */
11282 find_mem_in_note (rtx note)
11284 if (note && for_each_rtx (¬e, find_mem_in_note_1, ¬e))
11289 /* Replace MEM with its associated pseudo register. This function is
11290 called from load_mems via for_each_rtx. DATA is actually a pointer
11291 to a structure describing the instruction currently being scanned
11292 and the MEM we are currently replacing. */
11295 replace_loop_mem (rtx *mem, void *data)
11297 loop_replace_args *args = (loop_replace_args *) data;
11303 switch (GET_CODE (m))
11309 /* We're not interested in the MEM associated with a
11310 CONST_DOUBLE, so there's no need to traverse into one. */
11314 /* This is not a MEM. */
11318 if (!rtx_equal_p (args->match, m))
11319 /* This is not the MEM we are currently replacing. */
11322 /* Actually replace the MEM. */
11323 validate_change (args->insn, mem, args->replacement, 1);
11329 replace_loop_mems (rtx insn, rtx mem, rtx reg, int written)
11331 loop_replace_args args;
11335 args.replacement = reg;
11337 for_each_rtx (&insn, replace_loop_mem, &args);
11339 /* If we hoist a mem write out of the loop, then REG_EQUAL
11340 notes referring to the mem are no longer valid. */
11346 for (link = ®_NOTES (insn); (note = *link); link = &XEXP (note, 1))
11348 if (REG_NOTE_KIND (note) == REG_EQUAL
11349 && (sub = find_mem_in_note (note))
11350 && true_dependence (mem, VOIDmode, sub, rtx_varies_p))
11352 /* Remove the note. */
11353 validate_change (NULL_RTX, link, XEXP (note, 1), 1);
11360 /* Replace one register with another. Called through for_each_rtx; PX points
11361 to the rtx being scanned. DATA is actually a pointer to
11362 a structure of arguments. */
11365 replace_loop_reg (rtx *px, void *data)
11368 loop_replace_args *args = (loop_replace_args *) data;
11373 if (x == args->match)
11374 validate_change (args->insn, px, args->replacement, 1);
11380 replace_loop_regs (rtx insn, rtx reg, rtx replacement)
11382 loop_replace_args args;
11386 args.replacement = replacement;
11388 for_each_rtx (&insn, replace_loop_reg, &args);
11391 /* Emit insn for PATTERN after WHERE_INSN in basic block WHERE_BB
11392 (ignored in the interim). */
11395 loop_insn_emit_after (const struct loop *loop ATTRIBUTE_UNUSED,
11396 basic_block where_bb ATTRIBUTE_UNUSED, rtx where_insn,
11399 return emit_insn_after (pattern, where_insn);
11403 /* If WHERE_INSN is nonzero emit insn for PATTERN before WHERE_INSN
11404 in basic block WHERE_BB (ignored in the interim) within the loop
11405 otherwise hoist PATTERN into the loop pre-header. */
11408 loop_insn_emit_before (const struct loop *loop,
11409 basic_block where_bb ATTRIBUTE_UNUSED,
11410 rtx where_insn, rtx pattern)
11413 return loop_insn_hoist (loop, pattern);
11414 return emit_insn_before (pattern, where_insn);
11418 /* Emit call insn for PATTERN before WHERE_INSN in basic block
11419 WHERE_BB (ignored in the interim) within the loop. */
11422 loop_call_insn_emit_before (const struct loop *loop ATTRIBUTE_UNUSED,
11423 basic_block where_bb ATTRIBUTE_UNUSED,
11424 rtx where_insn, rtx pattern)
11426 return emit_call_insn_before (pattern, where_insn);
11430 /* Hoist insn for PATTERN into the loop pre-header. */
11433 loop_insn_hoist (const struct loop *loop, rtx pattern)
11435 return loop_insn_emit_before (loop, 0, loop->start, pattern);
11439 /* Hoist call insn for PATTERN into the loop pre-header. */
11442 loop_call_insn_hoist (const struct loop *loop, rtx pattern)
11444 return loop_call_insn_emit_before (loop, 0, loop->start, pattern);
11448 /* Sink insn for PATTERN after the loop end. */
11451 loop_insn_sink (const struct loop *loop, rtx pattern)
11453 return loop_insn_emit_before (loop, 0, loop->sink, pattern);
11456 /* bl->final_value can be either general_operand or PLUS of general_operand
11457 and constant. Emit sequence of instructions to load it into REG. */
11459 gen_load_of_final_value (rtx reg, rtx final_value)
11463 final_value = force_operand (final_value, reg);
11464 if (final_value != reg)
11465 emit_move_insn (reg, final_value);
11466 seq = get_insns ();
11471 /* If the loop has multiple exits, emit insn for PATTERN before the
11472 loop to ensure that it will always be executed no matter how the
11473 loop exits. Otherwise, emit the insn for PATTERN after the loop,
11474 since this is slightly more efficient. */
11477 loop_insn_sink_or_swim (const struct loop *loop, rtx pattern)
11479 if (loop->exit_count)
11480 return loop_insn_hoist (loop, pattern);
11482 return loop_insn_sink (loop, pattern);
11486 loop_ivs_dump (const struct loop *loop, FILE *file, int verbose)
11488 struct iv_class *bl;
11491 if (! loop || ! file)
11494 for (bl = LOOP_IVS (loop)->list; bl; bl = bl->next)
11497 fprintf (file, "Loop %d: %d IV classes\n", loop->num, iv_num);
11499 for (bl = LOOP_IVS (loop)->list; bl; bl = bl->next)
11501 loop_iv_class_dump (bl, file, verbose);
11502 fputc ('\n', file);
11508 loop_iv_class_dump (const struct iv_class *bl, FILE *file,
11509 int verbose ATTRIBUTE_UNUSED)
11511 struct induction *v;
11515 if (! bl || ! file)
11518 fprintf (file, "IV class for reg %d, benefit %d\n",
11519 bl->regno, bl->total_benefit);
11521 fprintf (file, " Init insn %d", INSN_UID (bl->init_insn));
11522 if (bl->initial_value)
11524 fprintf (file, ", init val: ");
11525 print_simple_rtl (file, bl->initial_value);
11527 if (bl->initial_test)
11529 fprintf (file, ", init test: ");
11530 print_simple_rtl (file, bl->initial_test);
11532 fputc ('\n', file);
11534 if (bl->final_value)
11536 fprintf (file, " Final val: ");
11537 print_simple_rtl (file, bl->final_value);
11538 fputc ('\n', file);
11541 if ((incr = biv_total_increment (bl)))
11543 fprintf (file, " Total increment: ");
11544 print_simple_rtl (file, incr);
11545 fputc ('\n', file);
11548 /* List the increments. */
11549 for (i = 0, v = bl->biv; v; v = v->next_iv, i++)
11551 fprintf (file, " Inc%d: insn %d, incr: ", i, INSN_UID (v->insn));
11552 print_simple_rtl (file, v->add_val);
11553 fputc ('\n', file);
11556 /* List the givs. */
11557 for (i = 0, v = bl->giv; v; v = v->next_iv, i++)
11559 fprintf (file, " Giv%d: insn %d, benefit %d, ",
11560 i, INSN_UID (v->insn), v->benefit);
11561 if (v->giv_type == DEST_ADDR)
11562 print_simple_rtl (file, v->mem);
11564 print_simple_rtl (file, single_set (v->insn));
11565 fputc ('\n', file);
11571 loop_biv_dump (const struct induction *v, FILE *file, int verbose)
11578 REGNO (v->dest_reg), INSN_UID (v->insn));
11579 fprintf (file, " const ");
11580 print_simple_rtl (file, v->add_val);
11582 if (verbose && v->final_value)
11584 fputc ('\n', file);
11585 fprintf (file, " final ");
11586 print_simple_rtl (file, v->final_value);
11589 fputc ('\n', file);
11594 loop_giv_dump (const struct induction *v, FILE *file, int verbose)
11599 if (v->giv_type == DEST_REG)
11600 fprintf (file, "Giv %d: insn %d",
11601 REGNO (v->dest_reg), INSN_UID (v->insn));
11603 fprintf (file, "Dest address: insn %d",
11604 INSN_UID (v->insn));
11606 fprintf (file, " src reg %d benefit %d",
11607 REGNO (v->src_reg), v->benefit);
11608 fprintf (file, " lifetime %d",
11611 if (v->replaceable)
11612 fprintf (file, " replaceable");
11614 if (v->no_const_addval)
11615 fprintf (file, " ncav");
11617 if (v->ext_dependent)
11619 switch (GET_CODE (v->ext_dependent))
11622 fprintf (file, " ext se");
11625 fprintf (file, " ext ze");
11628 fprintf (file, " ext tr");
11631 gcc_unreachable ();
11635 fputc ('\n', file);
11636 fprintf (file, " mult ");
11637 print_simple_rtl (file, v->mult_val);
11639 fputc ('\n', file);
11640 fprintf (file, " add ");
11641 print_simple_rtl (file, v->add_val);
11643 if (verbose && v->final_value)
11645 fputc ('\n', file);
11646 fprintf (file, " final ");
11647 print_simple_rtl (file, v->final_value);
11650 fputc ('\n', file);
11655 debug_ivs (const struct loop *loop)
11657 loop_ivs_dump (loop, stderr, 1);
11662 debug_iv_class (const struct iv_class *bl)
11664 loop_iv_class_dump (bl, stderr, 1);
11669 debug_biv (const struct induction *v)
11671 loop_biv_dump (v, stderr, 1);
11676 debug_giv (const struct induction *v)
11678 loop_giv_dump (v, stderr, 1);
11682 #define LOOP_BLOCK_NUM_1(INSN) \
11683 ((INSN) ? (BLOCK_FOR_INSN (INSN) ? BLOCK_NUM (INSN) : - 1) : -1)
11685 /* The notes do not have an assigned block, so look at the next insn. */
11686 #define LOOP_BLOCK_NUM(INSN) \
11687 ((INSN) ? (NOTE_P (INSN) \
11688 ? LOOP_BLOCK_NUM_1 (next_nonnote_insn (INSN)) \
11689 : LOOP_BLOCK_NUM_1 (INSN)) \
11692 #define LOOP_INSN_UID(INSN) ((INSN) ? INSN_UID (INSN) : -1)
11695 loop_dump_aux (const struct loop *loop, FILE *file,
11696 int verbose ATTRIBUTE_UNUSED)
11700 if (! loop || ! file || !BB_HEAD (loop->first))
11703 /* Print diagnostics to compare our concept of a loop with
11704 what the loop notes say. */
11705 if (! PREV_INSN (BB_HEAD (loop->first))
11706 || !NOTE_P (PREV_INSN (BB_HEAD (loop->first)))
11707 || NOTE_LINE_NUMBER (PREV_INSN (BB_HEAD (loop->first)))
11708 != NOTE_INSN_LOOP_BEG)
11709 fprintf (file, ";; No NOTE_INSN_LOOP_BEG at %d\n",
11710 INSN_UID (PREV_INSN (BB_HEAD (loop->first))));
11711 if (! NEXT_INSN (BB_END (loop->last))
11712 || !NOTE_P (NEXT_INSN (BB_END (loop->last)))
11713 || NOTE_LINE_NUMBER (NEXT_INSN (BB_END (loop->last)))
11714 != NOTE_INSN_LOOP_END)
11715 fprintf (file, ";; No NOTE_INSN_LOOP_END at %d\n",
11716 INSN_UID (NEXT_INSN (BB_END (loop->last))));
11721 ";; start %d (%d), end %d (%d)\n",
11722 LOOP_BLOCK_NUM (loop->start),
11723 LOOP_INSN_UID (loop->start),
11724 LOOP_BLOCK_NUM (loop->end),
11725 LOOP_INSN_UID (loop->end));
11726 fprintf (file, ";; top %d (%d), scan start %d (%d)\n",
11727 LOOP_BLOCK_NUM (loop->top),
11728 LOOP_INSN_UID (loop->top),
11729 LOOP_BLOCK_NUM (loop->scan_start),
11730 LOOP_INSN_UID (loop->scan_start));
11731 fprintf (file, ";; exit_count %d", loop->exit_count);
11732 if (loop->exit_count)
11734 fputs (", labels:", file);
11735 for (label = loop->exit_labels; label; label = LABEL_NEXTREF (label))
11737 fprintf (file, " %d ",
11738 LOOP_INSN_UID (XEXP (label, 0)));
11741 fputs ("\n", file);
11745 /* Call this function from the debugger to dump LOOP. */
11748 debug_loop (const struct loop *loop)
11750 flow_loop_dump (loop, stderr, loop_dump_aux, 1);
11753 /* Call this function from the debugger to dump LOOPS. */
11756 debug_loops (const struct loops *loops)
11758 flow_loops_dump (loops, stderr, loop_dump_aux, 1);
11762 gate_handle_loop_optimize (void)
11764 return (optimize > 0 && flag_loop_optimize);
11767 /* Move constant computations out of loops. */
11769 rest_of_handle_loop_optimize (void)
11773 /* CFG is no longer maintained up-to-date. */
11774 free_bb_for_insn ();
11775 profile_status = PROFILE_ABSENT;
11777 do_prefetch = flag_prefetch_loop_arrays ? LOOP_PREFETCH : 0;
11779 if (flag_rerun_loop_opt)
11781 cleanup_barriers ();
11783 /* We only want to perform unrolling once. */
11784 loop_optimize (get_insns (), dump_file, 0);
11786 /* The first call to loop_optimize makes some instructions
11787 trivially dead. We delete those instructions now in the
11788 hope that doing so will make the heuristics in loop work
11789 better and possibly speed up compilation. */
11790 delete_trivially_dead_insns (get_insns (), max_reg_num ());
11792 /* The regscan pass is currently necessary as the alias
11793 analysis code depends on this information. */
11794 reg_scan (get_insns (), max_reg_num ());
11796 cleanup_barriers ();
11797 loop_optimize (get_insns (), dump_file, do_prefetch);
11799 /* Loop can create trivially dead instructions. */
11800 delete_trivially_dead_insns (get_insns (), max_reg_num ());
11801 find_basic_blocks (get_insns ());
11804 struct tree_opt_pass pass_loop_optimize =
11806 "old-loop", /* name */
11807 gate_handle_loop_optimize, /* gate */
11808 rest_of_handle_loop_optimize, /* execute */
11811 0, /* static_pass_number */
11812 TV_LOOP, /* tv_id */
11813 0, /* properties_required */
11814 0, /* properties_provided */
11815 0, /* properties_destroyed */
11816 0, /* todo_flags_start */
11818 TODO_ggc_collect, /* todo_flags_finish */