1 /* Perform various loop optimizations, including strength reduction.
2 Copyright (C) 1987, 1988, 1989, 1991, 1992, 1993, 1994, 1995, 1996, 1997,
3 1998, 1999, 2000 Free Software Foundation, Inc.
5 This file is part of GNU CC.
7 GNU CC is free software; you can redistribute it and/or modify
8 it under the terms of the GNU General Public License as published by
9 the Free Software Foundation; either version 2, or (at your option)
12 GNU CC is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License for more details.
17 You should have received a copy of the GNU General Public License
18 along with GNU CC; see the file COPYING. If not, write to
19 the Free Software Foundation, 59 Temple Place - Suite 330,
20 Boston, MA 02111-1307, USA. */
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. Strength reduction is applied to the general
27 induction variables, and induction variable elimination is applied to
28 the basic induction variables.
30 It also finds cases where
31 a register is set within the loop by zero-extending a narrower value
32 and changes these to zero the entire register once before the loop
33 and merely copy the low part within the loop.
35 Most of the complexity is in heuristics to decide when it is worth
36 while to do these things. */
45 #include "hard-reg-set.h"
46 #include "basic-block.h"
47 #include "insn-config.h"
48 #include "insn-flags.h"
58 /* Vector mapping INSN_UIDs to luids.
59 The luids are like uids but increase monotonically always.
60 We use them to see whether a jump comes from outside a given loop. */
64 /* Indexed by INSN_UID, contains the ordinal giving the (innermost) loop
65 number the insn is contained in. */
67 struct loop **uid_loop;
69 /* 1 + largest uid of any insn. */
73 /* 1 + luid of last insn. */
77 /* Number of loops detected in current function. Used as index to the
80 static int max_loop_num;
82 /* Indexed by register number, contains the number of times the reg
83 is set during the loop being scanned.
84 During code motion, a negative value indicates a reg that has been
85 made a candidate; in particular -2 means that it is an candidate that
86 we know is equal to a constant and -1 means that it is an candidate
87 not known equal to a constant.
88 After code motion, regs moved have 0 (which is accurate now)
89 while the failed candidates have the original number of times set.
91 Therefore, at all times, == 0 indicates an invariant register;
92 < 0 a conditionally invariant one. */
94 static varray_type set_in_loop;
96 /* Original value of set_in_loop; same except that this value
97 is not set negative for a reg whose sets have been made candidates
98 and not set to 0 for a reg that is moved. */
100 static varray_type n_times_set;
102 /* Index by register number, 1 indicates that the register
103 cannot be moved or strength reduced. */
105 static varray_type may_not_optimize;
107 /* Contains the insn in which a register was used if it was used
108 exactly once; contains const0_rtx if it was used more than once. */
110 static varray_type reg_single_usage;
112 /* Nonzero means reg N has already been moved out of one loop.
113 This reduces the desire to move it out of another. */
115 static char *moved_once;
117 /* List of MEMs that are stored in this loop. */
119 static rtx loop_store_mems;
121 /* The insn where the first of these was found. */
122 static rtx first_loop_store_insn;
124 typedef struct loop_mem_info {
125 rtx mem; /* The MEM itself. */
126 rtx reg; /* Corresponding pseudo, if any. */
127 int optimize; /* Nonzero if we can optimize access to this MEM. */
130 /* Array of MEMs that are used (read or written) in this loop, but
131 cannot be aliased by anything in this loop, except perhaps
132 themselves. In other words, if loop_mems[i] is altered during the
133 loop, it is altered by an expression that is rtx_equal_p to it. */
135 static loop_mem_info *loop_mems;
137 /* The index of the next available slot in LOOP_MEMS. */
139 static int loop_mems_idx;
141 /* The number of elements allocated in LOOP_MEMs. */
143 static int loop_mems_allocated;
145 /* Nonzero if we don't know what MEMs were changed in the current
146 loop. This happens if the loop contains a call (in which case
147 `loop_info->has_call' will also be set) or if we store into more
148 than NUM_STORES MEMs. */
150 static int unknown_address_altered;
152 /* The above doesn't count any readonly memory locations that are stored.
155 static int unknown_constant_address_altered;
157 /* Count of movable (i.e. invariant) instructions discovered in the loop. */
158 static int num_movables;
160 /* Count of memory write instructions discovered in the loop. */
161 static int num_mem_sets;
163 /* Bound on pseudo register number before loop optimization.
164 A pseudo has valid regscan info if its number is < max_reg_before_loop. */
165 unsigned int max_reg_before_loop;
167 /* The value to pass to the next call of reg_scan_update. */
168 static int loop_max_reg;
170 /* This obstack is used in product_cheap_p to allocate its rtl. It
171 may call gen_reg_rtx which, in turn, may reallocate regno_reg_rtx.
172 If we used the same obstack that it did, we would be deallocating
175 static struct obstack temp_obstack;
177 /* This is where the pointer to the obstack being used for RTL is stored. */
179 extern struct obstack *rtl_obstack;
181 #define obstack_chunk_alloc xmalloc
182 #define obstack_chunk_free free
184 /* During the analysis of a loop, a chain of `struct movable's
185 is made to record all the movable insns found.
186 Then the entire chain can be scanned to decide which to move. */
190 rtx insn; /* A movable insn */
191 rtx set_src; /* The expression this reg is set from. */
192 rtx set_dest; /* The destination of this SET. */
193 rtx dependencies; /* When INSN is libcall, this is an EXPR_LIST
194 of any registers used within the LIBCALL. */
195 int consec; /* Number of consecutive following insns
196 that must be moved with this one. */
197 unsigned int regno; /* The register it sets */
198 short lifetime; /* lifetime of that register;
199 may be adjusted when matching movables
200 that load the same value are found. */
201 short savings; /* Number of insns we can move for this reg,
202 including other movables that force this
203 or match this one. */
204 unsigned int cond : 1; /* 1 if only conditionally movable */
205 unsigned int force : 1; /* 1 means MUST move this insn */
206 unsigned int global : 1; /* 1 means reg is live outside this loop */
207 /* If PARTIAL is 1, GLOBAL means something different:
208 that the reg is live outside the range from where it is set
209 to the following label. */
210 unsigned int done : 1; /* 1 inhibits further processing of this */
212 unsigned int partial : 1; /* 1 means this reg is used for zero-extending.
213 In particular, moving it does not make it
215 unsigned int move_insn : 1; /* 1 means that we call emit_move_insn to
216 load SRC, rather than copying INSN. */
217 unsigned int move_insn_first:1;/* Same as above, if this is necessary for the
218 first insn of a consecutive sets group. */
219 unsigned int is_equiv : 1; /* 1 means a REG_EQUIV is present on INSN. */
220 enum machine_mode savemode; /* Nonzero means it is a mode for a low part
221 that we should avoid changing when clearing
222 the rest of the reg. */
223 struct movable *match; /* First entry for same value */
224 struct movable *forces; /* An insn that must be moved if this is */
225 struct movable *next;
228 static struct movable *the_movables;
230 FILE *loop_dump_stream;
232 /* Forward declarations. */
234 static void verify_dominator PARAMS ((struct loop *));
235 static void find_and_verify_loops PARAMS ((rtx, struct loops *));
236 static void mark_loop_jump PARAMS ((rtx, struct loop *));
237 static void prescan_loop PARAMS ((struct loop *));
238 static int reg_in_basic_block_p PARAMS ((rtx, rtx));
239 static int consec_sets_invariant_p PARAMS ((const struct loop *,
241 static int labels_in_range_p PARAMS ((rtx, int));
242 static void count_one_set PARAMS ((rtx, rtx, varray_type, rtx *));
244 static void count_loop_regs_set PARAMS ((rtx, rtx, varray_type, varray_type,
246 static void note_addr_stored PARAMS ((rtx, rtx, void *));
247 static void note_set_pseudo_multiple_uses PARAMS ((rtx, rtx, void *));
248 static int loop_reg_used_before_p PARAMS ((const struct loop *, rtx, rtx));
249 static void scan_loop PARAMS ((struct loop*, int));
251 static void replace_call_address PARAMS ((rtx, rtx, rtx));
253 static rtx skip_consec_insns PARAMS ((rtx, int));
254 static int libcall_benefit PARAMS ((rtx));
255 static void ignore_some_movables PARAMS ((struct movable *));
256 static void force_movables PARAMS ((struct movable *));
257 static void combine_movables PARAMS ((struct movable *, int));
258 static int regs_match_p PARAMS ((rtx, rtx, struct movable *));
259 static int rtx_equal_for_loop_p PARAMS ((rtx, rtx, struct movable *));
260 static void add_label_notes PARAMS ((rtx, rtx));
261 static void move_movables PARAMS ((struct loop *loop, struct movable *,
263 static int count_nonfixed_reads PARAMS ((const struct loop *, rtx));
264 static void strength_reduce PARAMS ((struct loop *, int, int));
265 static void find_single_use_in_loop PARAMS ((rtx, rtx, varray_type));
266 static int valid_initial_value_p PARAMS ((rtx, rtx, int, rtx));
267 static void find_mem_givs PARAMS ((const struct loop *, rtx, rtx, int, int));
268 static void record_biv PARAMS ((struct induction *, rtx, rtx, rtx, rtx, rtx *,
270 static void check_final_value PARAMS ((const struct loop *,
271 struct induction *));
272 static void record_giv PARAMS ((const struct loop *, struct induction *,
273 rtx, rtx, rtx, rtx, rtx, int, enum g_types,
275 static void update_giv_derive PARAMS ((const struct loop *, rtx));
276 static int basic_induction_var PARAMS ((const struct loop *, rtx,
277 enum machine_mode, rtx, rtx,
278 rtx *, rtx *, rtx **));
279 static rtx simplify_giv_expr PARAMS ((const struct loop *, rtx, int *));
280 static int general_induction_var PARAMS ((const struct loop *loop, rtx, rtx *,
281 rtx *, rtx *, int, int *, enum machine_mode));
282 static int consec_sets_giv PARAMS ((const struct loop *, int, rtx,
283 rtx, rtx, rtx *, rtx *, rtx *));
284 static int check_dbra_loop PARAMS ((struct loop *, int));
285 static rtx express_from_1 PARAMS ((rtx, rtx, rtx));
286 static rtx combine_givs_p PARAMS ((struct induction *, struct induction *));
287 static void combine_givs PARAMS ((struct iv_class *));
288 struct recombine_givs_stats;
289 static int find_life_end PARAMS ((rtx, struct recombine_givs_stats *,
291 static void recombine_givs PARAMS ((const struct loop *, struct iv_class *,
293 static int product_cheap_p PARAMS ((rtx, rtx));
294 static int maybe_eliminate_biv PARAMS ((const struct loop *, struct iv_class *,
296 static int maybe_eliminate_biv_1 PARAMS ((const struct loop *, rtx, rtx,
297 struct iv_class *, int, rtx));
298 static int last_use_this_basic_block PARAMS ((rtx, rtx));
299 static void record_initial PARAMS ((rtx, rtx, void *));
300 static void update_reg_last_use PARAMS ((rtx, rtx));
301 static rtx next_insn_in_loop PARAMS ((const struct loop *, rtx));
302 static void load_mems_and_recount_loop_regs_set PARAMS ((const struct loop*,
304 static void load_mems PARAMS ((const struct loop *));
305 static int insert_loop_mem PARAMS ((rtx *, void *));
306 static int replace_loop_mem PARAMS ((rtx *, void *));
307 static int replace_loop_reg PARAMS ((rtx *, void *));
308 static void note_reg_stored PARAMS ((rtx, rtx, void *));
309 static void try_copy_prop PARAMS ((const struct loop *, rtx, unsigned int));
310 static void try_swap_copy_prop PARAMS ((const struct loop *, rtx,
312 static int replace_label PARAMS ((rtx *, void *));
313 static rtx check_insn_for_givs PARAMS((struct loop *, rtx, int, int));
314 static rtx check_insn_for_bivs PARAMS((struct loop *, rtx, int, int));
316 typedef struct rtx_and_int {
321 typedef struct rtx_pair {
326 /* Nonzero iff INSN is between START and END, inclusive. */
327 #define INSN_IN_RANGE_P(INSN, START, END) \
328 (INSN_UID (INSN) < max_uid_for_loop \
329 && INSN_LUID (INSN) >= INSN_LUID (START) \
330 && INSN_LUID (INSN) <= INSN_LUID (END))
332 /* Indirect_jump_in_function is computed once per function. */
333 static int indirect_jump_in_function;
334 static int indirect_jump_in_function_p PARAMS ((rtx));
336 static int compute_luids PARAMS ((rtx, rtx, int));
338 static int biv_elimination_giv_has_0_offset PARAMS ((struct induction *,
339 struct induction *, rtx));
341 /* Relative gain of eliminating various kinds of operations. */
344 static int shift_cost;
345 static int mult_cost;
348 /* Benefit penalty, if a giv is not replaceable, i.e. must emit an insn to
349 copy the value of the strength reduced giv to its original register. */
350 static int copy_cost;
352 /* Cost of using a register, to normalize the benefits of a giv. */
353 static int reg_address_cost;
359 char *free_point = (char *) oballoc (1);
360 rtx reg = gen_rtx_REG (word_mode, LAST_VIRTUAL_REGISTER + 1);
362 add_cost = rtx_cost (gen_rtx_PLUS (word_mode, reg, reg), SET);
364 reg_address_cost = address_cost (reg, SImode);
366 /* We multiply by 2 to reconcile the difference in scale between
367 these two ways of computing costs. Otherwise the cost of a copy
368 will be far less than the cost of an add. */
372 /* Free the objects we just allocated. */
375 /* Initialize the obstack used for rtl in product_cheap_p. */
376 gcc_obstack_init (&temp_obstack);
379 /* Compute the mapping from uids to luids.
380 LUIDs are numbers assigned to insns, like uids,
381 except that luids increase monotonically through the code.
382 Start at insn START and stop just before END. Assign LUIDs
383 starting with PREV_LUID + 1. Return the last assigned LUID + 1. */
385 compute_luids (start, end, prev_luid)
392 for (insn = start, i = prev_luid; insn != end; insn = NEXT_INSN (insn))
394 if (INSN_UID (insn) >= max_uid_for_loop)
396 /* Don't assign luids to line-number NOTEs, so that the distance in
397 luids between two insns is not affected by -g. */
398 if (GET_CODE (insn) != NOTE
399 || NOTE_LINE_NUMBER (insn) <= 0)
400 uid_luid[INSN_UID (insn)] = ++i;
402 /* Give a line number note the same luid as preceding insn. */
403 uid_luid[INSN_UID (insn)] = i;
408 /* Entry point of this file. Perform loop optimization
409 on the current function. F is the first insn of the function
410 and DUMPFILE is a stream for output of a trace of actions taken
411 (or 0 if none should be output). */
414 loop_optimize (f, dumpfile, flags)
415 /* f is the first instruction of a chain of insns for one function */
422 struct loops loops_data;
423 struct loops *loops = &loops_data;
424 struct loop_info *loops_info;
426 loop_dump_stream = dumpfile;
428 init_recog_no_volatile ();
430 max_reg_before_loop = max_reg_num ();
431 loop_max_reg = max_reg_before_loop;
435 /* Count the number of loops. */
438 for (insn = f; insn; insn = NEXT_INSN (insn))
440 if (GET_CODE (insn) == NOTE
441 && NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG)
445 /* Don't waste time if no loops. */
446 if (max_loop_num == 0)
449 loops->num = max_loop_num;
451 moved_once = (char *) xcalloc (max_reg_before_loop, sizeof (char));
453 /* Get size to use for tables indexed by uids.
454 Leave some space for labels allocated by find_and_verify_loops. */
455 max_uid_for_loop = get_max_uid () + 1 + max_loop_num * 32;
457 uid_luid = (int *) xcalloc (max_uid_for_loop, sizeof (int));
458 uid_loop = (struct loop **) xcalloc (max_uid_for_loop,
459 sizeof (struct loop *));
461 /* Allocate storage for array of loops. */
462 loops->array = (struct loop *)
463 xcalloc (loops->num, sizeof (struct loop));
465 /* Find and process each loop.
466 First, find them, and record them in order of their beginnings. */
467 find_and_verify_loops (f, loops);
469 /* Allocate and initialize auxiliary loop information. */
470 loops_info = xcalloc (loops->num, sizeof (struct loop_info));
471 for (i = 0; i < loops->num; i++)
472 loops->array[i].aux = loops_info + i;
474 /* Now find all register lifetimes. This must be done after
475 find_and_verify_loops, because it might reorder the insns in the
477 reg_scan (f, max_reg_before_loop, 1);
479 /* This must occur after reg_scan so that registers created by gcse
480 will have entries in the register tables.
482 We could have added a call to reg_scan after gcse_main in toplev.c,
483 but moving this call to init_alias_analysis is more efficient. */
484 init_alias_analysis ();
486 /* See if we went too far. Note that get_max_uid already returns
487 one more that the maximum uid of all insn. */
488 if (get_max_uid () > max_uid_for_loop)
490 /* Now reset it to the actual size we need. See above. */
491 max_uid_for_loop = get_max_uid ();
493 /* find_and_verify_loops has already called compute_luids, but it
494 might have rearranged code afterwards, so we need to recompute
496 max_luid = compute_luids (f, NULL_RTX, 0);
498 /* Don't leave gaps in uid_luid for insns that have been
499 deleted. It is possible that the first or last insn
500 using some register has been deleted by cross-jumping.
501 Make sure that uid_luid for that former insn's uid
502 points to the general area where that insn used to be. */
503 for (i = 0; i < max_uid_for_loop; i++)
505 uid_luid[0] = uid_luid[i];
506 if (uid_luid[0] != 0)
509 for (i = 0; i < max_uid_for_loop; i++)
510 if (uid_luid[i] == 0)
511 uid_luid[i] = uid_luid[i - 1];
513 /* Determine if the function has indirect jump. On some systems
514 this prevents low overhead loop instructions from being used. */
515 indirect_jump_in_function = indirect_jump_in_function_p (f);
517 /* Now scan the loops, last ones first, since this means inner ones are done
518 before outer ones. */
519 for (i = max_loop_num - 1; i >= 0; i--)
521 struct loop *loop = &loops->array[i];
523 if (! loop->invalid && loop->end)
524 scan_loop (loop, flags);
527 /* If there were lexical blocks inside the loop, they have been
528 replicated. We will now have more than one NOTE_INSN_BLOCK_BEG
529 and NOTE_INSN_BLOCK_END for each such block. We must duplicate
530 the BLOCKs as well. */
531 if (write_symbols != NO_DEBUG)
534 end_alias_analysis ();
544 /* Returns the next insn, in execution order, after INSN. START and
545 END are the NOTE_INSN_LOOP_BEG and NOTE_INSN_LOOP_END for the loop,
546 respectively. LOOP->TOP, if non-NULL, is the top of the loop in the
547 insn-stream; it is used with loops that are entered near the
551 next_insn_in_loop (loop, insn)
552 const struct loop *loop;
555 insn = NEXT_INSN (insn);
557 if (insn == loop->end)
560 /* Go to the top of the loop, and continue there. */
567 if (insn == loop->scan_start)
574 /* Optimize one loop described by LOOP. */
576 /* ??? Could also move memory writes out of loops if the destination address
577 is invariant, the source is invariant, the memory write is not volatile,
578 and if we can prove that no read inside the loop can read this address
579 before the write occurs. If there is a read of this address after the
580 write, then we can also mark the memory read as invariant. */
583 scan_loop (loop, flags)
588 rtx loop_start = loop->start;
589 rtx loop_end = loop->end;
590 /* Additional information about the current loop being processed
591 that is used to compute the number of loop iterations for loop
592 unrolling and doloop optimization. */
593 struct loop_info *loop_info = LOOP_INFO (loop);
595 /* 1 if we are scanning insns that could be executed zero times. */
597 /* 1 if we are scanning insns that might never be executed
598 due to a subroutine call which might exit before they are reached. */
600 /* Jump insn that enters the loop, or 0 if control drops in. */
601 rtx loop_entry_jump = 0;
602 /* Number of insns in the loop. */
606 rtx temp, update_start, update_end;
607 /* The SET from an insn, if it is the only SET in the insn. */
609 /* Chain describing insns movable in current loop. */
610 struct movable *movables = 0;
611 /* Last element in `movables' -- so we can add elements at the end. */
612 struct movable *last_movable = 0;
613 /* Ratio of extra register life span we can justify
614 for saving an instruction. More if loop doesn't call subroutines
615 since in that case saving an insn makes more difference
616 and more registers are available. */
618 /* Nonzero if we are scanning instructions in a sub-loop. */
624 /* Determine whether this loop starts with a jump down to a test at
625 the end. This will occur for a small number of loops with a test
626 that is too complex to duplicate in front of the loop.
628 We search for the first insn or label in the loop, skipping NOTEs.
629 However, we must be careful not to skip past a NOTE_INSN_LOOP_BEG
630 (because we might have a loop executed only once that contains a
631 loop which starts with a jump to its exit test) or a NOTE_INSN_LOOP_END
632 (in case we have a degenerate loop).
634 Note that if we mistakenly think that a loop is entered at the top
635 when, in fact, it is entered at the exit test, the only effect will be
636 slightly poorer optimization. Making the opposite error can generate
637 incorrect code. Since very few loops now start with a jump to the
638 exit test, the code here to detect that case is very conservative. */
640 for (p = NEXT_INSN (loop_start);
642 && GET_CODE (p) != CODE_LABEL && ! INSN_P (p)
643 && (GET_CODE (p) != NOTE
644 || (NOTE_LINE_NUMBER (p) != NOTE_INSN_LOOP_BEG
645 && NOTE_LINE_NUMBER (p) != NOTE_INSN_LOOP_END));
649 loop->scan_start = p;
651 /* Set up variables describing this loop. */
653 threshold = (loop_info->has_call ? 1 : 2) * (1 + n_non_fixed_regs);
655 /* If loop has a jump before the first label,
656 the true entry is the target of that jump.
657 Start scan from there.
658 But record in LOOP->TOP the place where the end-test jumps
659 back to so we can scan that after the end of the loop. */
660 if (GET_CODE (p) == JUMP_INSN)
664 /* Loop entry must be unconditional jump (and not a RETURN) */
665 if (any_uncondjump_p (p)
666 && JUMP_LABEL (p) != 0
667 /* Check to see whether the jump actually
668 jumps out of the loop (meaning it's no loop).
669 This case can happen for things like
670 do {..} while (0). If this label was generated previously
671 by loop, we can't tell anything about it and have to reject
673 && INSN_IN_RANGE_P (JUMP_LABEL (p), loop_start, loop_end))
675 loop->top = next_label (loop->scan_start);
676 loop->scan_start = JUMP_LABEL (p);
680 /* If LOOP->SCAN_START was an insn created by loop, we don't know its luid
681 as required by loop_reg_used_before_p. So skip such loops. (This
682 test may never be true, but it's best to play it safe.)
684 Also, skip loops where we do not start scanning at a label. This
685 test also rejects loops starting with a JUMP_INSN that failed the
688 if (INSN_UID (loop->scan_start) >= max_uid_for_loop
689 || GET_CODE (loop->scan_start) != CODE_LABEL)
691 if (loop_dump_stream)
692 fprintf (loop_dump_stream, "\nLoop from %d to %d is phony.\n\n",
693 INSN_UID (loop_start), INSN_UID (loop_end));
697 /* Count number of times each reg is set during this loop.
698 Set VARRAY_CHAR (may_not_optimize, I) if it is not safe to move out
699 the setting of register I. Set VARRAY_RTX (reg_single_usage, I). */
701 /* Allocate extra space for REGS that might be created by
702 load_mems. We allocate a little extra slop as well, in the hopes
703 that even after the moving of movables creates some new registers
704 we won't have to reallocate these arrays. However, we do grow
705 the arrays, if necessary, in load_mems_recount_loop_regs_set. */
706 nregs = max_reg_num () + loop_mems_idx + 16;
707 VARRAY_INT_INIT (set_in_loop, nregs, "set_in_loop");
708 VARRAY_INT_INIT (n_times_set, nregs, "n_times_set");
709 VARRAY_CHAR_INIT (may_not_optimize, nregs, "may_not_optimize");
710 VARRAY_RTX_INIT (reg_single_usage, nregs, "reg_single_usage");
712 count_loop_regs_set (loop->top ? loop->top : loop->start, loop->end,
713 may_not_optimize, reg_single_usage, &insn_count, nregs);
715 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
717 VARRAY_CHAR (may_not_optimize, i) = 1;
718 VARRAY_INT (set_in_loop, i) = 1;
721 #ifdef AVOID_CCMODE_COPIES
722 /* Don't try to move insns which set CC registers if we should not
723 create CCmode register copies. */
724 for (i = max_reg_num () - 1; i >= FIRST_PSEUDO_REGISTER; i--)
725 if (GET_MODE_CLASS (GET_MODE (regno_reg_rtx[i])) == MODE_CC)
726 VARRAY_CHAR (may_not_optimize, i) = 1;
729 bcopy ((char *) &set_in_loop->data,
730 (char *) &n_times_set->data, nregs * sizeof (int));
732 if (loop_dump_stream)
734 fprintf (loop_dump_stream, "\nLoop from %d to %d: %d real insns.\n",
735 INSN_UID (loop_start), INSN_UID (loop_end), insn_count);
737 fprintf (loop_dump_stream, "Continue at insn %d.\n",
738 INSN_UID (loop->cont));
741 /* Scan through the loop finding insns that are safe to move.
742 Set set_in_loop negative for the reg being set, so that
743 this reg will be considered invariant for subsequent insns.
744 We consider whether subsequent insns use the reg
745 in deciding whether it is worth actually moving.
747 MAYBE_NEVER is nonzero if we have passed a conditional jump insn
748 and therefore it is possible that the insns we are scanning
749 would never be executed. At such times, we must make sure
750 that it is safe to execute the insn once instead of zero times.
751 When MAYBE_NEVER is 0, all insns will be executed at least once
752 so that is not a problem. */
754 for (p = next_insn_in_loop (loop, loop->scan_start);
756 p = next_insn_in_loop (loop, p))
758 if (INSN_P (p) && find_reg_note (p, REG_LIBCALL, NULL_RTX))
760 else if (INSN_P (p) && find_reg_note (p, REG_RETVAL, NULL_RTX))
763 if (GET_CODE (p) == INSN
764 && (set = single_set (p))
765 && GET_CODE (SET_DEST (set)) == REG
766 && ! VARRAY_CHAR (may_not_optimize, REGNO (SET_DEST (set))))
771 rtx src = SET_SRC (set);
772 rtx dependencies = 0;
774 /* Figure out what to use as a source of this insn. If a REG_EQUIV
775 note is given or if a REG_EQUAL note with a constant operand is
776 specified, use it as the source and mark that we should move
777 this insn by calling emit_move_insn rather that duplicating the
780 Otherwise, only use the REG_EQUAL contents if a REG_RETVAL note
782 temp = find_reg_note (p, REG_EQUIV, NULL_RTX);
784 src = XEXP (temp, 0), move_insn = 1;
787 temp = find_reg_note (p, REG_EQUAL, NULL_RTX);
788 if (temp && CONSTANT_P (XEXP (temp, 0)))
789 src = XEXP (temp, 0), move_insn = 1;
790 if (temp && find_reg_note (p, REG_RETVAL, NULL_RTX))
792 src = XEXP (temp, 0);
793 /* A libcall block can use regs that don't appear in
794 the equivalent expression. To move the libcall,
795 we must move those regs too. */
796 dependencies = libcall_other_reg (p, src);
800 /* Don't try to optimize a register that was made
801 by loop-optimization for an inner loop.
802 We don't know its life-span, so we can't compute the benefit. */
803 if (REGNO (SET_DEST (set)) >= max_reg_before_loop)
805 else if (/* The register is used in basic blocks other
806 than the one where it is set (meaning that
807 something after this point in the loop might
808 depend on its value before the set). */
809 ! reg_in_basic_block_p (p, SET_DEST (set))
810 /* And the set is not guaranteed to be executed one
811 the loop starts, or the value before the set is
812 needed before the set occurs...
814 ??? Note we have quadratic behaviour here, mitigated
815 by the fact that the previous test will often fail for
816 large loops. Rather than re-scanning the entire loop
817 each time for register usage, we should build tables
818 of the register usage and use them here instead. */
820 || loop_reg_used_before_p (loop, set, p)))
821 /* It is unsafe to move the set.
823 This code used to consider it OK to move a set of a variable
824 which was not created by the user and not used in an exit test.
825 That behavior is incorrect and was removed. */
827 else if ((tem = loop_invariant_p (loop, src))
828 && (dependencies == 0
829 || (tem2 = loop_invariant_p (loop, dependencies)) != 0)
830 && (VARRAY_INT (set_in_loop,
831 REGNO (SET_DEST (set))) == 1
833 = consec_sets_invariant_p
834 (loop, SET_DEST (set),
835 VARRAY_INT (set_in_loop, REGNO (SET_DEST (set))),
837 /* If the insn can cause a trap (such as divide by zero),
838 can't move it unless it's guaranteed to be executed
839 once loop is entered. Even a function call might
840 prevent the trap insn from being reached
841 (since it might exit!) */
842 && ! ((maybe_never || call_passed)
843 && may_trap_p (src)))
845 register struct movable *m;
846 register int regno = REGNO (SET_DEST (set));
848 /* A potential lossage is where we have a case where two insns
849 can be combined as long as they are both in the loop, but
850 we move one of them outside the loop. For large loops,
851 this can lose. The most common case of this is the address
852 of a function being called.
854 Therefore, if this register is marked as being used exactly
855 once if we are in a loop with calls (a "large loop"), see if
856 we can replace the usage of this register with the source
857 of this SET. If we can, delete this insn.
859 Don't do this if P has a REG_RETVAL note or if we have
860 SMALL_REGISTER_CLASSES and SET_SRC is a hard register. */
862 if (loop_info->has_call
863 && VARRAY_RTX (reg_single_usage, regno) != 0
864 && VARRAY_RTX (reg_single_usage, regno) != const0_rtx
865 && REGNO_FIRST_UID (regno) == INSN_UID (p)
866 && (REGNO_LAST_UID (regno)
867 == INSN_UID (VARRAY_RTX (reg_single_usage, regno)))
868 && VARRAY_INT (set_in_loop, regno) == 1
869 && ! side_effects_p (SET_SRC (set))
870 && ! find_reg_note (p, REG_RETVAL, NULL_RTX)
871 && (! SMALL_REGISTER_CLASSES
872 || (! (GET_CODE (SET_SRC (set)) == REG
873 && REGNO (SET_SRC (set)) < FIRST_PSEUDO_REGISTER)))
874 /* This test is not redundant; SET_SRC (set) might be
875 a call-clobbered register and the life of REGNO
876 might span a call. */
877 && ! modified_between_p (SET_SRC (set), p,
879 (reg_single_usage, regno))
880 && no_labels_between_p (p, VARRAY_RTX (reg_single_usage, regno))
881 && validate_replace_rtx (SET_DEST (set), SET_SRC (set),
883 (reg_single_usage, regno)))
885 /* Replace any usage in a REG_EQUAL note. Must copy the
886 new source, so that we don't get rtx sharing between the
887 SET_SOURCE and REG_NOTES of insn p. */
888 REG_NOTES (VARRAY_RTX (reg_single_usage, regno))
889 = replace_rtx (REG_NOTES (VARRAY_RTX
890 (reg_single_usage, regno)),
891 SET_DEST (set), copy_rtx (SET_SRC (set)));
894 NOTE_LINE_NUMBER (p) = NOTE_INSN_DELETED;
895 NOTE_SOURCE_FILE (p) = 0;
896 VARRAY_INT (set_in_loop, regno) = 0;
900 m = (struct movable *) alloca (sizeof (struct movable));
904 m->dependencies = dependencies;
905 m->set_dest = SET_DEST (set);
907 m->consec = VARRAY_INT (set_in_loop,
908 REGNO (SET_DEST (set))) - 1;
912 m->move_insn = move_insn;
913 m->move_insn_first = 0;
914 m->is_equiv = (find_reg_note (p, REG_EQUIV, NULL_RTX) != 0);
915 m->savemode = VOIDmode;
917 /* Set M->cond if either loop_invariant_p
918 or consec_sets_invariant_p returned 2
919 (only conditionally invariant). */
920 m->cond = ((tem | tem1 | tem2) > 1);
921 m->global = (uid_luid[REGNO_LAST_UID (regno)]
922 > INSN_LUID (loop_end)
923 || uid_luid[REGNO_FIRST_UID (regno)] < INSN_LUID (loop_start));
925 m->lifetime = (uid_luid[REGNO_LAST_UID (regno)]
926 - uid_luid[REGNO_FIRST_UID (regno)]);
927 m->savings = VARRAY_INT (n_times_set, regno);
928 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
929 m->savings += libcall_benefit (p);
930 VARRAY_INT (set_in_loop, regno) = move_insn ? -2 : -1;
931 /* Add M to the end of the chain MOVABLES. */
935 last_movable->next = m;
940 /* It is possible for the first instruction to have a
941 REG_EQUAL note but a non-invariant SET_SRC, so we must
942 remember the status of the first instruction in case
943 the last instruction doesn't have a REG_EQUAL note. */
944 m->move_insn_first = m->move_insn;
946 /* Skip this insn, not checking REG_LIBCALL notes. */
947 p = next_nonnote_insn (p);
948 /* Skip the consecutive insns, if there are any. */
949 p = skip_consec_insns (p, m->consec);
950 /* Back up to the last insn of the consecutive group. */
951 p = prev_nonnote_insn (p);
953 /* We must now reset m->move_insn, m->is_equiv, and possibly
954 m->set_src to correspond to the effects of all the
956 temp = find_reg_note (p, REG_EQUIV, NULL_RTX);
958 m->set_src = XEXP (temp, 0), m->move_insn = 1;
961 temp = find_reg_note (p, REG_EQUAL, NULL_RTX);
962 if (temp && CONSTANT_P (XEXP (temp, 0)))
963 m->set_src = XEXP (temp, 0), m->move_insn = 1;
968 m->is_equiv = (find_reg_note (p, REG_EQUIV, NULL_RTX) != 0);
971 /* If this register is always set within a STRICT_LOW_PART
972 or set to zero, then its high bytes are constant.
973 So clear them outside the loop and within the loop
974 just load the low bytes.
975 We must check that the machine has an instruction to do so.
976 Also, if the value loaded into the register
977 depends on the same register, this cannot be done. */
978 else if (SET_SRC (set) == const0_rtx
979 && GET_CODE (NEXT_INSN (p)) == INSN
980 && (set1 = single_set (NEXT_INSN (p)))
981 && GET_CODE (set1) == SET
982 && (GET_CODE (SET_DEST (set1)) == STRICT_LOW_PART)
983 && (GET_CODE (XEXP (SET_DEST (set1), 0)) == SUBREG)
984 && (SUBREG_REG (XEXP (SET_DEST (set1), 0))
986 && !reg_mentioned_p (SET_DEST (set), SET_SRC (set1)))
988 register int regno = REGNO (SET_DEST (set));
989 if (VARRAY_INT (set_in_loop, regno) == 2)
991 register struct movable *m;
992 m = (struct movable *) alloca (sizeof (struct movable));
995 m->set_dest = SET_DEST (set);
1002 m->move_insn_first = 0;
1004 /* If the insn may not be executed on some cycles,
1005 we can't clear the whole reg; clear just high part.
1006 Not even if the reg is used only within this loop.
1013 Clearing x before the inner loop could clobber a value
1014 being saved from the last time around the outer loop.
1015 However, if the reg is not used outside this loop
1016 and all uses of the register are in the same
1017 basic block as the store, there is no problem.
1019 If this insn was made by loop, we don't know its
1020 INSN_LUID and hence must make a conservative
1022 m->global = (INSN_UID (p) >= max_uid_for_loop
1023 || (uid_luid[REGNO_LAST_UID (regno)]
1024 > INSN_LUID (loop_end))
1025 || (uid_luid[REGNO_FIRST_UID (regno)]
1027 || (labels_in_range_p
1028 (p, uid_luid[REGNO_FIRST_UID (regno)])));
1029 if (maybe_never && m->global)
1030 m->savemode = GET_MODE (SET_SRC (set1));
1032 m->savemode = VOIDmode;
1036 m->lifetime = (uid_luid[REGNO_LAST_UID (regno)]
1037 - uid_luid[REGNO_FIRST_UID (regno)]);
1039 VARRAY_INT (set_in_loop, regno) = -1;
1040 /* Add M to the end of the chain MOVABLES. */
1044 last_movable->next = m;
1049 /* Past a call insn, we get to insns which might not be executed
1050 because the call might exit. This matters for insns that trap.
1051 Call insns inside a REG_LIBCALL/REG_RETVAL block always return,
1052 so they don't count. */
1053 else if (GET_CODE (p) == CALL_INSN && ! in_libcall)
1055 /* Past a label or a jump, we get to insns for which we
1056 can't count on whether or how many times they will be
1057 executed during each iteration. Therefore, we can
1058 only move out sets of trivial variables
1059 (those not used after the loop). */
1060 /* Similar code appears twice in strength_reduce. */
1061 else if ((GET_CODE (p) == CODE_LABEL || GET_CODE (p) == JUMP_INSN)
1062 /* If we enter the loop in the middle, and scan around to the
1063 beginning, don't set maybe_never for that. This must be an
1064 unconditional jump, otherwise the code at the top of the
1065 loop might never be executed. Unconditional jumps are
1066 followed a by barrier then loop end. */
1067 && ! (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) == loop->top
1068 && NEXT_INSN (NEXT_INSN (p)) == loop_end
1069 && any_uncondjump_p (p)))
1071 else if (GET_CODE (p) == NOTE)
1073 /* At the virtual top of a converted loop, insns are again known to
1074 be executed: logically, the loop begins here even though the exit
1075 code has been duplicated. */
1076 if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP && loop_depth == 0)
1077 maybe_never = call_passed = 0;
1078 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
1080 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END)
1085 /* If one movable subsumes another, ignore that other. */
1087 ignore_some_movables (movables);
1089 /* For each movable insn, see if the reg that it loads
1090 leads when it dies right into another conditionally movable insn.
1091 If so, record that the second insn "forces" the first one,
1092 since the second can be moved only if the first is. */
1094 force_movables (movables);
1096 /* See if there are multiple movable insns that load the same value.
1097 If there are, make all but the first point at the first one
1098 through the `match' field, and add the priorities of them
1099 all together as the priority of the first. */
1101 combine_movables (movables, nregs);
1103 /* Now consider each movable insn to decide whether it is worth moving.
1104 Store 0 in set_in_loop for each reg that is moved.
1106 Generally this increases code size, so do not move moveables when
1107 optimizing for code size. */
1109 if (! optimize_size)
1110 move_movables (loop, movables, threshold, insn_count, nregs);
1112 /* Now candidates that still are negative are those not moved.
1113 Change set_in_loop to indicate that those are not actually invariant. */
1114 for (i = 0; i < nregs; i++)
1115 if (VARRAY_INT (set_in_loop, i) < 0)
1116 VARRAY_INT (set_in_loop, i) = VARRAY_INT (n_times_set, i);
1118 /* Now that we've moved some things out of the loop, we might be able to
1119 hoist even more memory references. */
1120 load_mems_and_recount_loop_regs_set (loop, &insn_count);
1122 for (update_start = loop_start;
1123 PREV_INSN (update_start)
1124 && GET_CODE (PREV_INSN (update_start)) != CODE_LABEL;
1125 update_start = PREV_INSN (update_start))
1127 update_end = NEXT_INSN (loop_end);
1129 reg_scan_update (update_start, update_end, loop_max_reg);
1130 loop_max_reg = max_reg_num ();
1132 if (flag_strength_reduce)
1134 if (update_end && GET_CODE (update_end) == CODE_LABEL)
1135 /* Ensure our label doesn't go away. */
1136 LABEL_NUSES (update_end)++;
1138 the_movables = movables;
1139 strength_reduce (loop, insn_count, flags);
1141 reg_scan_update (update_start, update_end, loop_max_reg);
1142 loop_max_reg = max_reg_num ();
1144 if (update_end && GET_CODE (update_end) == CODE_LABEL
1145 && --LABEL_NUSES (update_end) == 0)
1146 delete_insn (update_end);
1149 VARRAY_FREE (reg_single_usage);
1150 VARRAY_FREE (set_in_loop);
1151 VARRAY_FREE (n_times_set);
1152 VARRAY_FREE (may_not_optimize);
1155 /* Add elements to *OUTPUT to record all the pseudo-regs
1156 mentioned in IN_THIS but not mentioned in NOT_IN_THIS. */
1159 record_excess_regs (in_this, not_in_this, output)
1160 rtx in_this, not_in_this;
1167 code = GET_CODE (in_this);
1181 if (REGNO (in_this) >= FIRST_PSEUDO_REGISTER
1182 && ! reg_mentioned_p (in_this, not_in_this))
1183 *output = gen_rtx_EXPR_LIST (VOIDmode, in_this, *output);
1190 fmt = GET_RTX_FORMAT (code);
1191 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1198 for (j = 0; j < XVECLEN (in_this, i); j++)
1199 record_excess_regs (XVECEXP (in_this, i, j), not_in_this, output);
1203 record_excess_regs (XEXP (in_this, i), not_in_this, output);
1209 /* Check what regs are referred to in the libcall block ending with INSN,
1210 aside from those mentioned in the equivalent value.
1211 If there are none, return 0.
1212 If there are one or more, return an EXPR_LIST containing all of them. */
1215 libcall_other_reg (insn, equiv)
1218 rtx note = find_reg_note (insn, REG_RETVAL, NULL_RTX);
1219 rtx p = XEXP (note, 0);
1222 /* First, find all the regs used in the libcall block
1223 that are not mentioned as inputs to the result. */
1227 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
1228 || GET_CODE (p) == CALL_INSN)
1229 record_excess_regs (PATTERN (p), equiv, &output);
1236 /* Return 1 if all uses of REG
1237 are between INSN and the end of the basic block. */
1240 reg_in_basic_block_p (insn, reg)
1243 int regno = REGNO (reg);
1246 if (REGNO_FIRST_UID (regno) != INSN_UID (insn))
1249 /* Search this basic block for the already recorded last use of the reg. */
1250 for (p = insn; p; p = NEXT_INSN (p))
1252 switch (GET_CODE (p))
1259 /* Ordinary insn: if this is the last use, we win. */
1260 if (REGNO_LAST_UID (regno) == INSN_UID (p))
1265 /* Jump insn: if this is the last use, we win. */
1266 if (REGNO_LAST_UID (regno) == INSN_UID (p))
1268 /* Otherwise, it's the end of the basic block, so we lose. */
1273 /* It's the end of the basic block, so we lose. */
1281 /* The "last use" that was recorded can't be found after the first
1282 use. This can happen when the last use was deleted while
1283 processing an inner loop, this inner loop was then completely
1284 unrolled, and the outer loop is always exited after the inner loop,
1285 so that everything after the first use becomes a single basic block. */
1289 /* Compute the benefit of eliminating the insns in the block whose
1290 last insn is LAST. This may be a group of insns used to compute a
1291 value directly or can contain a library call. */
1294 libcall_benefit (last)
1300 for (insn = XEXP (find_reg_note (last, REG_RETVAL, NULL_RTX), 0);
1301 insn != last; insn = NEXT_INSN (insn))
1303 if (GET_CODE (insn) == CALL_INSN)
1304 benefit += 10; /* Assume at least this many insns in a library
1306 else if (GET_CODE (insn) == INSN
1307 && GET_CODE (PATTERN (insn)) != USE
1308 && GET_CODE (PATTERN (insn)) != CLOBBER)
1315 /* Skip COUNT insns from INSN, counting library calls as 1 insn. */
1318 skip_consec_insns (insn, count)
1322 for (; count > 0; count--)
1326 /* If first insn of libcall sequence, skip to end. */
1327 /* Do this at start of loop, since INSN is guaranteed to
1329 if (GET_CODE (insn) != NOTE
1330 && (temp = find_reg_note (insn, REG_LIBCALL, NULL_RTX)))
1331 insn = XEXP (temp, 0);
1333 do insn = NEXT_INSN (insn);
1334 while (GET_CODE (insn) == NOTE);
1340 /* Ignore any movable whose insn falls within a libcall
1341 which is part of another movable.
1342 We make use of the fact that the movable for the libcall value
1343 was made later and so appears later on the chain. */
1346 ignore_some_movables (movables)
1347 struct movable *movables;
1349 register struct movable *m, *m1;
1351 for (m = movables; m; m = m->next)
1353 /* Is this a movable for the value of a libcall? */
1354 rtx note = find_reg_note (m->insn, REG_RETVAL, NULL_RTX);
1358 /* Check for earlier movables inside that range,
1359 and mark them invalid. We cannot use LUIDs here because
1360 insns created by loop.c for prior loops don't have LUIDs.
1361 Rather than reject all such insns from movables, we just
1362 explicitly check each insn in the libcall (since invariant
1363 libcalls aren't that common). */
1364 for (insn = XEXP (note, 0); insn != m->insn; insn = NEXT_INSN (insn))
1365 for (m1 = movables; m1 != m; m1 = m1->next)
1366 if (m1->insn == insn)
1372 /* For each movable insn, see if the reg that it loads
1373 leads when it dies right into another conditionally movable insn.
1374 If so, record that the second insn "forces" the first one,
1375 since the second can be moved only if the first is. */
1378 force_movables (movables)
1379 struct movable *movables;
1381 register struct movable *m, *m1;
1382 for (m1 = movables; m1; m1 = m1->next)
1383 /* Omit this if moving just the (SET (REG) 0) of a zero-extend. */
1384 if (!m1->partial && !m1->done)
1386 int regno = m1->regno;
1387 for (m = m1->next; m; m = m->next)
1388 /* ??? Could this be a bug? What if CSE caused the
1389 register of M1 to be used after this insn?
1390 Since CSE does not update regno_last_uid,
1391 this insn M->insn might not be where it dies.
1392 But very likely this doesn't matter; what matters is
1393 that M's reg is computed from M1's reg. */
1394 if (INSN_UID (m->insn) == REGNO_LAST_UID (regno)
1397 if (m != 0 && m->set_src == m1->set_dest
1398 /* If m->consec, m->set_src isn't valid. */
1402 /* Increase the priority of the moving the first insn
1403 since it permits the second to be moved as well. */
1407 m1->lifetime += m->lifetime;
1408 m1->savings += m->savings;
1413 /* Find invariant expressions that are equal and can be combined into
1417 combine_movables (movables, nregs)
1418 struct movable *movables;
1421 register struct movable *m;
1422 char *matched_regs = (char *) xmalloc (nregs);
1423 enum machine_mode mode;
1425 /* Regs that are set more than once are not allowed to match
1426 or be matched. I'm no longer sure why not. */
1427 /* Perhaps testing m->consec_sets would be more appropriate here? */
1429 for (m = movables; m; m = m->next)
1430 if (m->match == 0 && VARRAY_INT (n_times_set, m->regno) == 1 && !m->partial)
1432 register struct movable *m1;
1433 int regno = m->regno;
1435 bzero (matched_regs, nregs);
1436 matched_regs[regno] = 1;
1438 /* We want later insns to match the first one. Don't make the first
1439 one match any later ones. So start this loop at m->next. */
1440 for (m1 = m->next; m1; m1 = m1->next)
1441 if (m != m1 && m1->match == 0 && VARRAY_INT (n_times_set, m1->regno) == 1
1442 /* A reg used outside the loop mustn't be eliminated. */
1444 /* A reg used for zero-extending mustn't be eliminated. */
1446 && (matched_regs[m1->regno]
1449 /* Can combine regs with different modes loaded from the
1450 same constant only if the modes are the same or
1451 if both are integer modes with M wider or the same
1452 width as M1. The check for integer is redundant, but
1453 safe, since the only case of differing destination
1454 modes with equal sources is when both sources are
1455 VOIDmode, i.e., CONST_INT. */
1456 (GET_MODE (m->set_dest) == GET_MODE (m1->set_dest)
1457 || (GET_MODE_CLASS (GET_MODE (m->set_dest)) == MODE_INT
1458 && GET_MODE_CLASS (GET_MODE (m1->set_dest)) == MODE_INT
1459 && (GET_MODE_BITSIZE (GET_MODE (m->set_dest))
1460 >= GET_MODE_BITSIZE (GET_MODE (m1->set_dest)))))
1461 /* See if the source of M1 says it matches M. */
1462 && ((GET_CODE (m1->set_src) == REG
1463 && matched_regs[REGNO (m1->set_src)])
1464 || rtx_equal_for_loop_p (m->set_src, m1->set_src,
1466 && ((m->dependencies == m1->dependencies)
1467 || rtx_equal_p (m->dependencies, m1->dependencies)))
1469 m->lifetime += m1->lifetime;
1470 m->savings += m1->savings;
1473 matched_regs[m1->regno] = 1;
1477 /* Now combine the regs used for zero-extension.
1478 This can be done for those not marked `global'
1479 provided their lives don't overlap. */
1481 for (mode = GET_CLASS_NARROWEST_MODE (MODE_INT); mode != VOIDmode;
1482 mode = GET_MODE_WIDER_MODE (mode))
1484 register struct movable *m0 = 0;
1486 /* Combine all the registers for extension from mode MODE.
1487 Don't combine any that are used outside this loop. */
1488 for (m = movables; m; m = m->next)
1489 if (m->partial && ! m->global
1490 && mode == GET_MODE (SET_SRC (PATTERN (NEXT_INSN (m->insn)))))
1492 register struct movable *m1;
1493 int first = uid_luid[REGNO_FIRST_UID (m->regno)];
1494 int last = uid_luid[REGNO_LAST_UID (m->regno)];
1498 /* First one: don't check for overlap, just record it. */
1503 /* Make sure they extend to the same mode.
1504 (Almost always true.) */
1505 if (GET_MODE (m->set_dest) != GET_MODE (m0->set_dest))
1508 /* We already have one: check for overlap with those
1509 already combined together. */
1510 for (m1 = movables; m1 != m; m1 = m1->next)
1511 if (m1 == m0 || (m1->partial && m1->match == m0))
1512 if (! (uid_luid[REGNO_FIRST_UID (m1->regno)] > last
1513 || uid_luid[REGNO_LAST_UID (m1->regno)] < first))
1516 /* No overlap: we can combine this with the others. */
1517 m0->lifetime += m->lifetime;
1518 m0->savings += m->savings;
1527 free (matched_regs);
1530 /* Return 1 if regs X and Y will become the same if moved. */
1533 regs_match_p (x, y, movables)
1535 struct movable *movables;
1537 unsigned int xn = REGNO (x);
1538 unsigned int yn = REGNO (y);
1539 struct movable *mx, *my;
1541 for (mx = movables; mx; mx = mx->next)
1542 if (mx->regno == xn)
1545 for (my = movables; my; my = my->next)
1546 if (my->regno == yn)
1550 && ((mx->match == my->match && mx->match != 0)
1552 || mx == my->match));
1555 /* Return 1 if X and Y are identical-looking rtx's.
1556 This is the Lisp function EQUAL for rtx arguments.
1558 If two registers are matching movables or a movable register and an
1559 equivalent constant, consider them equal. */
1562 rtx_equal_for_loop_p (x, y, movables)
1564 struct movable *movables;
1568 register struct movable *m;
1569 register enum rtx_code code;
1570 register const char *fmt;
1574 if (x == 0 || y == 0)
1577 code = GET_CODE (x);
1579 /* If we have a register and a constant, they may sometimes be
1581 if (GET_CODE (x) == REG && VARRAY_INT (set_in_loop, REGNO (x)) == -2
1584 for (m = movables; m; m = m->next)
1585 if (m->move_insn && m->regno == REGNO (x)
1586 && rtx_equal_p (m->set_src, y))
1589 else if (GET_CODE (y) == REG && VARRAY_INT (set_in_loop, REGNO (y)) == -2
1592 for (m = movables; m; m = m->next)
1593 if (m->move_insn && m->regno == REGNO (y)
1594 && rtx_equal_p (m->set_src, x))
1598 /* Otherwise, rtx's of different codes cannot be equal. */
1599 if (code != GET_CODE (y))
1602 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent.
1603 (REG:SI x) and (REG:HI x) are NOT equivalent. */
1605 if (GET_MODE (x) != GET_MODE (y))
1608 /* These three types of rtx's can be compared nonrecursively. */
1610 return (REGNO (x) == REGNO (y) || regs_match_p (x, y, movables));
1612 if (code == LABEL_REF)
1613 return XEXP (x, 0) == XEXP (y, 0);
1614 if (code == SYMBOL_REF)
1615 return XSTR (x, 0) == XSTR (y, 0);
1617 /* Compare the elements. If any pair of corresponding elements
1618 fail to match, return 0 for the whole things. */
1620 fmt = GET_RTX_FORMAT (code);
1621 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1626 if (XWINT (x, i) != XWINT (y, i))
1631 if (XINT (x, i) != XINT (y, i))
1636 /* Two vectors must have the same length. */
1637 if (XVECLEN (x, i) != XVECLEN (y, i))
1640 /* And the corresponding elements must match. */
1641 for (j = 0; j < XVECLEN (x, i); j++)
1642 if (rtx_equal_for_loop_p (XVECEXP (x, i, j), XVECEXP (y, i, j), movables) == 0)
1647 if (rtx_equal_for_loop_p (XEXP (x, i), XEXP (y, i), movables) == 0)
1652 if (strcmp (XSTR (x, i), XSTR (y, i)))
1657 /* These are just backpointers, so they don't matter. */
1663 /* It is believed that rtx's at this level will never
1664 contain anything but integers and other rtx's,
1665 except for within LABEL_REFs and SYMBOL_REFs. */
1673 /* If X contains any LABEL_REF's, add REG_LABEL notes for them to all
1674 insns in INSNS which use the reference. */
1677 add_label_notes (x, insns)
1681 enum rtx_code code = GET_CODE (x);
1686 if (code == LABEL_REF && !LABEL_REF_NONLOCAL_P (x))
1688 /* This code used to ignore labels that referred to dispatch tables to
1689 avoid flow generating (slighly) worse code.
1691 We no longer ignore such label references (see LABEL_REF handling in
1692 mark_jump_label for additional information). */
1693 for (insn = insns; insn; insn = NEXT_INSN (insn))
1694 if (reg_mentioned_p (XEXP (x, 0), insn))
1695 REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_LABEL, XEXP (x, 0),
1699 fmt = GET_RTX_FORMAT (code);
1700 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1703 add_label_notes (XEXP (x, i), insns);
1704 else if (fmt[i] == 'E')
1705 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
1706 add_label_notes (XVECEXP (x, i, j), insns);
1710 /* Scan MOVABLES, and move the insns that deserve to be moved.
1711 If two matching movables are combined, replace one reg with the
1712 other throughout. */
1715 move_movables (loop, movables, threshold, insn_count, nregs)
1717 struct movable *movables;
1723 register struct movable *m;
1725 rtx loop_start = loop->start;
1726 rtx loop_end = loop->end;
1727 /* Map of pseudo-register replacements to handle combining
1728 when we move several insns that load the same value
1729 into different pseudo-registers. */
1730 rtx *reg_map = (rtx *) xcalloc (nregs, sizeof (rtx));
1731 char *already_moved = (char *) xcalloc (nregs, sizeof (char));
1735 for (m = movables; m; m = m->next)
1737 /* Describe this movable insn. */
1739 if (loop_dump_stream)
1741 fprintf (loop_dump_stream, "Insn %d: regno %d (life %d), ",
1742 INSN_UID (m->insn), m->regno, m->lifetime);
1744 fprintf (loop_dump_stream, "consec %d, ", m->consec);
1746 fprintf (loop_dump_stream, "cond ");
1748 fprintf (loop_dump_stream, "force ");
1750 fprintf (loop_dump_stream, "global ");
1752 fprintf (loop_dump_stream, "done ");
1754 fprintf (loop_dump_stream, "move-insn ");
1756 fprintf (loop_dump_stream, "matches %d ",
1757 INSN_UID (m->match->insn));
1759 fprintf (loop_dump_stream, "forces %d ",
1760 INSN_UID (m->forces->insn));
1763 /* Count movables. Value used in heuristics in strength_reduce. */
1766 /* Ignore the insn if it's already done (it matched something else).
1767 Otherwise, see if it is now safe to move. */
1771 || (1 == loop_invariant_p (loop, m->set_src)
1772 && (m->dependencies == 0
1773 || 1 == loop_invariant_p (loop, m->dependencies))
1775 || 1 == consec_sets_invariant_p (loop, m->set_dest,
1778 && (! m->forces || m->forces->done))
1782 int savings = m->savings;
1784 /* We have an insn that is safe to move.
1785 Compute its desirability. */
1790 if (loop_dump_stream)
1791 fprintf (loop_dump_stream, "savings %d ", savings);
1793 if (moved_once[regno] && loop_dump_stream)
1794 fprintf (loop_dump_stream, "halved since already moved ");
1796 /* An insn MUST be moved if we already moved something else
1797 which is safe only if this one is moved too: that is,
1798 if already_moved[REGNO] is nonzero. */
1800 /* An insn is desirable to move if the new lifetime of the
1801 register is no more than THRESHOLD times the old lifetime.
1802 If it's not desirable, it means the loop is so big
1803 that moving won't speed things up much,
1804 and it is liable to make register usage worse. */
1806 /* It is also desirable to move if it can be moved at no
1807 extra cost because something else was already moved. */
1809 if (already_moved[regno]
1810 || flag_move_all_movables
1811 || (threshold * savings * m->lifetime) >=
1812 (moved_once[regno] ? insn_count * 2 : insn_count)
1813 || (m->forces && m->forces->done
1814 && VARRAY_INT (n_times_set, m->forces->regno) == 1))
1817 register struct movable *m1;
1818 rtx first = NULL_RTX;
1820 /* Now move the insns that set the reg. */
1822 if (m->partial && m->match)
1826 /* Find the end of this chain of matching regs.
1827 Thus, we load each reg in the chain from that one reg.
1828 And that reg is loaded with 0 directly,
1829 since it has ->match == 0. */
1830 for (m1 = m; m1->match; m1 = m1->match);
1831 newpat = gen_move_insn (SET_DEST (PATTERN (m->insn)),
1832 SET_DEST (PATTERN (m1->insn)));
1833 i1 = emit_insn_before (newpat, loop_start);
1835 /* Mark the moved, invariant reg as being allowed to
1836 share a hard reg with the other matching invariant. */
1837 REG_NOTES (i1) = REG_NOTES (m->insn);
1838 r1 = SET_DEST (PATTERN (m->insn));
1839 r2 = SET_DEST (PATTERN (m1->insn));
1841 = gen_rtx_EXPR_LIST (VOIDmode, r1,
1842 gen_rtx_EXPR_LIST (VOIDmode, r2,
1844 delete_insn (m->insn);
1849 if (loop_dump_stream)
1850 fprintf (loop_dump_stream, " moved to %d", INSN_UID (i1));
1852 /* If we are to re-generate the item being moved with a
1853 new move insn, first delete what we have and then emit
1854 the move insn before the loop. */
1855 else if (m->move_insn)
1859 for (count = m->consec; count >= 0; count--)
1861 /* If this is the first insn of a library call sequence,
1863 if (GET_CODE (p) != NOTE
1864 && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
1867 /* If this is the last insn of a libcall sequence, then
1868 delete every insn in the sequence except the last.
1869 The last insn is handled in the normal manner. */
1870 if (GET_CODE (p) != NOTE
1871 && (temp = find_reg_note (p, REG_RETVAL, NULL_RTX)))
1873 temp = XEXP (temp, 0);
1875 temp = delete_insn (temp);
1879 p = delete_insn (p);
1881 /* simplify_giv_expr expects that it can walk the insns
1882 at m->insn forwards and see this old sequence we are
1883 tossing here. delete_insn does preserve the next
1884 pointers, but when we skip over a NOTE we must fix
1885 it up. Otherwise that code walks into the non-deleted
1887 while (p && GET_CODE (p) == NOTE)
1888 p = NEXT_INSN (temp) = NEXT_INSN (p);
1892 emit_move_insn (m->set_dest, m->set_src);
1893 temp = get_insns ();
1896 add_label_notes (m->set_src, temp);
1898 i1 = emit_insns_before (temp, loop_start);
1899 if (! find_reg_note (i1, REG_EQUAL, NULL_RTX))
1901 = gen_rtx_EXPR_LIST (m->is_equiv ? REG_EQUIV : REG_EQUAL,
1902 m->set_src, REG_NOTES (i1));
1904 if (loop_dump_stream)
1905 fprintf (loop_dump_stream, " moved to %d", INSN_UID (i1));
1907 /* The more regs we move, the less we like moving them. */
1912 for (count = m->consec; count >= 0; count--)
1916 /* If first insn of libcall sequence, skip to end. */
1917 /* Do this at start of loop, since p is guaranteed to
1919 if (GET_CODE (p) != NOTE
1920 && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
1923 /* If last insn of libcall sequence, move all
1924 insns except the last before the loop. The last
1925 insn is handled in the normal manner. */
1926 if (GET_CODE (p) != NOTE
1927 && (temp = find_reg_note (p, REG_RETVAL, NULL_RTX)))
1931 rtx fn_address_insn = 0;
1934 for (temp = XEXP (temp, 0); temp != p;
1935 temp = NEXT_INSN (temp))
1941 if (GET_CODE (temp) == NOTE)
1944 body = PATTERN (temp);
1946 /* Find the next insn after TEMP,
1947 not counting USE or NOTE insns. */
1948 for (next = NEXT_INSN (temp); next != p;
1949 next = NEXT_INSN (next))
1950 if (! (GET_CODE (next) == INSN
1951 && GET_CODE (PATTERN (next)) == USE)
1952 && GET_CODE (next) != NOTE)
1955 /* If that is the call, this may be the insn
1956 that loads the function address.
1958 Extract the function address from the insn
1959 that loads it into a register.
1960 If this insn was cse'd, we get incorrect code.
1962 So emit a new move insn that copies the
1963 function address into the register that the
1964 call insn will use. flow.c will delete any
1965 redundant stores that we have created. */
1966 if (GET_CODE (next) == CALL_INSN
1967 && GET_CODE (body) == SET
1968 && GET_CODE (SET_DEST (body)) == REG
1969 && (n = find_reg_note (temp, REG_EQUAL,
1972 fn_reg = SET_SRC (body);
1973 if (GET_CODE (fn_reg) != REG)
1974 fn_reg = SET_DEST (body);
1975 fn_address = XEXP (n, 0);
1976 fn_address_insn = temp;
1978 /* We have the call insn.
1979 If it uses the register we suspect it might,
1980 load it with the correct address directly. */
1981 if (GET_CODE (temp) == CALL_INSN
1983 && reg_referenced_p (fn_reg, body))
1984 emit_insn_after (gen_move_insn (fn_reg,
1988 if (GET_CODE (temp) == CALL_INSN)
1990 i1 = emit_call_insn_before (body, loop_start);
1991 /* Because the USAGE information potentially
1992 contains objects other than hard registers
1993 we need to copy it. */
1994 if (CALL_INSN_FUNCTION_USAGE (temp))
1995 CALL_INSN_FUNCTION_USAGE (i1)
1996 = copy_rtx (CALL_INSN_FUNCTION_USAGE (temp));
1999 i1 = emit_insn_before (body, loop_start);
2002 if (temp == fn_address_insn)
2003 fn_address_insn = i1;
2004 REG_NOTES (i1) = REG_NOTES (temp);
2010 if (m->savemode != VOIDmode)
2012 /* P sets REG to zero; but we should clear only
2013 the bits that are not covered by the mode
2015 rtx reg = m->set_dest;
2021 (GET_MODE (reg), and_optab, reg,
2022 GEN_INT ((((HOST_WIDE_INT) 1
2023 << GET_MODE_BITSIZE (m->savemode)))
2025 reg, 1, OPTAB_LIB_WIDEN);
2029 emit_move_insn (reg, tem);
2030 sequence = gen_sequence ();
2032 i1 = emit_insn_before (sequence, loop_start);
2034 else if (GET_CODE (p) == CALL_INSN)
2036 i1 = emit_call_insn_before (PATTERN (p), loop_start);
2037 /* Because the USAGE information potentially
2038 contains objects other than hard registers
2039 we need to copy it. */
2040 if (CALL_INSN_FUNCTION_USAGE (p))
2041 CALL_INSN_FUNCTION_USAGE (i1)
2042 = copy_rtx (CALL_INSN_FUNCTION_USAGE (p));
2044 else if (count == m->consec && m->move_insn_first)
2046 /* The SET_SRC might not be invariant, so we must
2047 use the REG_EQUAL note. */
2049 emit_move_insn (m->set_dest, m->set_src);
2050 temp = get_insns ();
2053 add_label_notes (m->set_src, temp);
2055 i1 = emit_insns_before (temp, loop_start);
2056 if (! find_reg_note (i1, REG_EQUAL, NULL_RTX))
2058 = gen_rtx_EXPR_LIST ((m->is_equiv ? REG_EQUIV
2060 m->set_src, REG_NOTES (i1));
2063 i1 = emit_insn_before (PATTERN (p), loop_start);
2065 if (REG_NOTES (i1) == 0)
2067 REG_NOTES (i1) = REG_NOTES (p);
2069 /* If there is a REG_EQUAL note present whose value
2070 is not loop invariant, then delete it, since it
2071 may cause problems with later optimization passes.
2072 It is possible for cse to create such notes
2073 like this as a result of record_jump_cond. */
2075 if ((temp = find_reg_note (i1, REG_EQUAL, NULL_RTX))
2076 && ! loop_invariant_p (loop, XEXP (temp, 0)))
2077 remove_note (i1, temp);
2083 if (loop_dump_stream)
2084 fprintf (loop_dump_stream, " moved to %d",
2087 /* If library call, now fix the REG_NOTES that contain
2088 insn pointers, namely REG_LIBCALL on FIRST
2089 and REG_RETVAL on I1. */
2090 if ((temp = find_reg_note (i1, REG_RETVAL, NULL_RTX)))
2092 XEXP (temp, 0) = first;
2093 temp = find_reg_note (first, REG_LIBCALL, NULL_RTX);
2094 XEXP (temp, 0) = i1;
2101 /* simplify_giv_expr expects that it can walk the insns
2102 at m->insn forwards and see this old sequence we are
2103 tossing here. delete_insn does preserve the next
2104 pointers, but when we skip over a NOTE we must fix
2105 it up. Otherwise that code walks into the non-deleted
2107 while (p && GET_CODE (p) == NOTE)
2108 p = NEXT_INSN (temp) = NEXT_INSN (p);
2111 /* The more regs we move, the less we like moving them. */
2115 /* Any other movable that loads the same register
2117 already_moved[regno] = 1;
2119 /* This reg has been moved out of one loop. */
2120 moved_once[regno] = 1;
2122 /* The reg set here is now invariant. */
2124 VARRAY_INT (set_in_loop, regno) = 0;
2128 /* Change the length-of-life info for the register
2129 to say it lives at least the full length of this loop.
2130 This will help guide optimizations in outer loops. */
2132 if (uid_luid[REGNO_FIRST_UID (regno)] > INSN_LUID (loop_start))
2133 /* This is the old insn before all the moved insns.
2134 We can't use the moved insn because it is out of range
2135 in uid_luid. Only the old insns have luids. */
2136 REGNO_FIRST_UID (regno) = INSN_UID (loop_start);
2137 if (uid_luid[REGNO_LAST_UID (regno)] < INSN_LUID (loop_end))
2138 REGNO_LAST_UID (regno) = INSN_UID (loop_end);
2140 /* Combine with this moved insn any other matching movables. */
2143 for (m1 = movables; m1; m1 = m1->next)
2148 /* Schedule the reg loaded by M1
2149 for replacement so that shares the reg of M.
2150 If the modes differ (only possible in restricted
2151 circumstances, make a SUBREG.
2153 Note this assumes that the target dependent files
2154 treat REG and SUBREG equally, including within
2155 GO_IF_LEGITIMATE_ADDRESS and in all the
2156 predicates since we never verify that replacing the
2157 original register with a SUBREG results in a
2158 recognizable insn. */
2159 if (GET_MODE (m->set_dest) == GET_MODE (m1->set_dest))
2160 reg_map[m1->regno] = m->set_dest;
2163 = gen_lowpart_common (GET_MODE (m1->set_dest),
2166 /* Get rid of the matching insn
2167 and prevent further processing of it. */
2170 /* if library call, delete all insn except last, which
2172 if ((temp = find_reg_note (m1->insn, REG_RETVAL,
2175 for (temp = XEXP (temp, 0); temp != m1->insn;
2176 temp = NEXT_INSN (temp))
2179 delete_insn (m1->insn);
2181 /* Any other movable that loads the same register
2183 already_moved[m1->regno] = 1;
2185 /* The reg merged here is now invariant,
2186 if the reg it matches is invariant. */
2188 VARRAY_INT (set_in_loop, m1->regno) = 0;
2191 else if (loop_dump_stream)
2192 fprintf (loop_dump_stream, "not desirable");
2194 else if (loop_dump_stream && !m->match)
2195 fprintf (loop_dump_stream, "not safe");
2197 if (loop_dump_stream)
2198 fprintf (loop_dump_stream, "\n");
2202 new_start = loop_start;
2204 /* Go through all the instructions in the loop, making
2205 all the register substitutions scheduled in REG_MAP. */
2206 for (p = new_start; p != loop_end; p = NEXT_INSN (p))
2207 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
2208 || GET_CODE (p) == CALL_INSN)
2210 replace_regs (PATTERN (p), reg_map, nregs, 0);
2211 replace_regs (REG_NOTES (p), reg_map, nregs, 0);
2217 free (already_moved);
2221 /* Scan X and replace the address of any MEM in it with ADDR.
2222 REG is the address that MEM should have before the replacement. */
2225 replace_call_address (x, reg, addr)
2228 register enum rtx_code code;
2230 register const char *fmt;
2234 code = GET_CODE (x);
2248 /* Short cut for very common case. */
2249 replace_call_address (XEXP (x, 1), reg, addr);
2253 /* Short cut for very common case. */
2254 replace_call_address (XEXP (x, 0), reg, addr);
2258 /* If this MEM uses a reg other than the one we expected,
2259 something is wrong. */
2260 if (XEXP (x, 0) != reg)
2269 fmt = GET_RTX_FORMAT (code);
2270 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2273 replace_call_address (XEXP (x, i), reg, addr);
2274 else if (fmt[i] == 'E')
2277 for (j = 0; j < XVECLEN (x, i); j++)
2278 replace_call_address (XVECEXP (x, i, j), reg, addr);
2284 /* Return the number of memory refs to addresses that vary
2288 count_nonfixed_reads (loop, x)
2289 const struct loop *loop;
2292 register enum rtx_code code;
2294 register const char *fmt;
2300 code = GET_CODE (x);
2314 return ((loop_invariant_p (loop, XEXP (x, 0)) != 1)
2315 + count_nonfixed_reads (loop, XEXP (x, 0)));
2322 fmt = GET_RTX_FORMAT (code);
2323 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2326 value += count_nonfixed_reads (loop, XEXP (x, i));
2330 for (j = 0; j < XVECLEN (x, i); j++)
2331 value += count_nonfixed_reads (loop, XVECEXP (x, i, j));
2339 /* P is an instruction that sets a register to the result of a ZERO_EXTEND.
2340 Replace it with an instruction to load just the low bytes
2341 if the machine supports such an instruction,
2342 and insert above LOOP_START an instruction to clear the register. */
2345 constant_high_bytes (p, loop_start)
2349 register int insn_code_number;
2351 /* Try to change (SET (REG ...) (ZERO_EXTEND (..:B ...)))
2352 to (SET (STRICT_LOW_PART (SUBREG:B (REG...))) ...). */
2357 gen_rtx_STRICT_LOW_PART
2359 gen_rtx_SUBREG (GET_MODE (XEXP (SET_SRC (PATTERN (p)), 0)),
2360 SET_DEST (PATTERN (p)), 0)),
2361 XEXP (SET_SRC (PATTERN (p)), 0));
2363 insn_code_number = recog (new, p);
2365 if (insn_code_number)
2369 /* Clear destination register before the loop. */
2370 emit_insn_before (gen_rtx_SET (VOIDmode,
2371 SET_DEST (PATTERN (p)), const0_rtx),
2374 /* Inside the loop, just load the low part. */
2380 /* Scan a loop setting the elements `cont', `vtop', `loops_enclosed',
2381 `has_call', `has_volatile', and `has_tablejump' within LOOP.
2382 Set the global variables `unknown_address_altered',
2383 `unknown_constant_address_altered', and `num_mem_sets'. Also, fill
2384 in the array `loop_mems' and the list `loop_store_mems'. */
2390 register int level = 1;
2392 struct loop_info *loop_info = LOOP_INFO (loop);
2393 rtx start = loop->start;
2394 rtx end = loop->end;
2395 /* The label after END. Jumping here is just like falling off the
2396 end of the loop. We use next_nonnote_insn instead of next_label
2397 as a hedge against the (pathological) case where some actual insn
2398 might end up between the two. */
2399 rtx exit_target = next_nonnote_insn (end);
2401 loop_info->has_indirect_jump = indirect_jump_in_function;
2402 loop_info->has_call = 0;
2403 loop_info->has_volatile = 0;
2404 loop_info->has_tablejump = 0;
2405 loop_info->has_multiple_exit_targets = 0;
2410 unknown_address_altered = 0;
2411 unknown_constant_address_altered = 0;
2412 loop_store_mems = NULL_RTX;
2413 first_loop_store_insn = NULL_RTX;
2417 for (insn = NEXT_INSN (start); insn != NEXT_INSN (end);
2418 insn = NEXT_INSN (insn))
2420 if (GET_CODE (insn) == NOTE)
2422 if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG)
2425 /* Count number of loops contained in this one. */
2428 else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_END)
2437 else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_CONT)
2442 else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_VTOP)
2444 /* If there is a NOTE_INSN_LOOP_VTOP, then this is a for
2445 or while style loop, with a loop exit test at the
2446 start. Thus, we can assume that the loop condition
2447 was true when the loop was entered. */
2452 else if (GET_CODE (insn) == CALL_INSN)
2454 if (! CONST_CALL_P (insn))
2455 unknown_address_altered = 1;
2456 loop_info->has_call = 1;
2458 else if (GET_CODE (insn) == INSN || GET_CODE (insn) == JUMP_INSN)
2460 rtx label1 = NULL_RTX;
2461 rtx label2 = NULL_RTX;
2463 if (volatile_refs_p (PATTERN (insn)))
2464 loop_info->has_volatile = 1;
2466 if (GET_CODE (insn) == JUMP_INSN
2467 && (GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC
2468 || GET_CODE (PATTERN (insn)) == ADDR_VEC))
2469 loop_info->has_tablejump = 1;
2471 note_stores (PATTERN (insn), note_addr_stored, NULL);
2472 if (! first_loop_store_insn && loop_store_mems)
2473 first_loop_store_insn = insn;
2475 if (! loop_info->has_multiple_exit_targets
2476 && GET_CODE (insn) == JUMP_INSN
2477 && GET_CODE (PATTERN (insn)) == SET
2478 && SET_DEST (PATTERN (insn)) == pc_rtx)
2480 if (GET_CODE (SET_SRC (PATTERN (insn))) == IF_THEN_ELSE)
2482 label1 = XEXP (SET_SRC (PATTERN (insn)), 1);
2483 label2 = XEXP (SET_SRC (PATTERN (insn)), 2);
2487 label1 = SET_SRC (PATTERN (insn));
2491 if (label1 && label1 != pc_rtx)
2493 if (GET_CODE (label1) != LABEL_REF)
2495 /* Something tricky. */
2496 loop_info->has_multiple_exit_targets = 1;
2499 else if (XEXP (label1, 0) != exit_target
2500 && LABEL_OUTSIDE_LOOP_P (label1))
2502 /* A jump outside the current loop. */
2503 loop_info->has_multiple_exit_targets = 1;
2513 else if (GET_CODE (insn) == RETURN)
2514 loop_info->has_multiple_exit_targets = 1;
2517 /* Now, rescan the loop, setting up the LOOP_MEMS array. */
2518 if (/* We can't tell what MEMs are aliased by what. */
2519 ! unknown_address_altered
2520 /* An exception thrown by a called function might land us
2522 && ! loop_info->has_call
2523 /* We don't want loads for MEMs moved to a location before the
2524 one at which their stack memory becomes allocated. (Note
2525 that this is not a problem for malloc, etc., since those
2526 require actual function calls. */
2527 && ! current_function_calls_alloca
2528 /* There are ways to leave the loop other than falling off the
2530 && ! loop_info->has_multiple_exit_targets)
2531 for (insn = NEXT_INSN (start); insn != NEXT_INSN (end);
2532 insn = NEXT_INSN (insn))
2533 for_each_rtx (&insn, insert_loop_mem, 0);
2536 /* LOOP->CONT_DOMINATOR is now the last label between the loop start
2537 and the continue note that is a the destination of a (cond)jump after
2538 the continue note. If there is any (cond)jump between the loop start
2539 and what we have so far as LOOP->CONT_DOMINATOR that has a
2540 target between LOOP->DOMINATOR and the continue note, move
2541 LOOP->CONT_DOMINATOR forward to that label; if a jump's
2542 destination cannot be determined, clear LOOP->CONT_DOMINATOR. */
2545 verify_dominator (loop)
2550 if (! loop->cont_dominator)
2551 /* This can happen for an empty loop, e.g. in
2552 gcc.c-torture/compile/920410-2.c */
2554 if (loop->cont_dominator == const0_rtx)
2556 loop->cont_dominator = 0;
2559 for (insn = loop->start; insn != loop->cont_dominator;
2560 insn = NEXT_INSN (insn))
2562 if (GET_CODE (insn) == JUMP_INSN
2563 && GET_CODE (PATTERN (insn)) != RETURN)
2565 rtx label = JUMP_LABEL (insn);
2568 /* If it is not a jump we can easily understand or for
2569 which we do not have jump target information in the JUMP_LABEL
2570 field (consider ADDR_VEC and ADDR_DIFF_VEC insns), then clear
2571 LOOP->CONT_DOMINATOR. */
2572 if (! any_condjump_p (insn)
2573 || label == NULL_RTX)
2575 loop->cont_dominator = NULL_RTX;
2579 label_luid = INSN_LUID (label);
2580 if (label_luid < INSN_LUID (loop->cont)
2582 > INSN_LUID (loop->cont)))
2583 loop->cont_dominator = label;
2588 /* Scan the function looking for loops. Record the start and end of each loop.
2589 Also mark as invalid loops any loops that contain a setjmp or are branched
2590 to from outside the loop. */
2593 find_and_verify_loops (f, loops)
2595 struct loops *loops;
2600 struct loop *current_loop;
2601 struct loop *next_loop;
2604 num_loops = loops->num;
2606 compute_luids (f, NULL_RTX, 0);
2608 /* If there are jumps to undefined labels,
2609 treat them as jumps out of any/all loops.
2610 This also avoids writing past end of tables when there are no loops. */
2613 /* Find boundaries of loops, mark which loops are contained within
2614 loops, and invalidate loops that have setjmp. */
2617 current_loop = NULL;
2618 for (insn = f; insn; insn = NEXT_INSN (insn))
2620 if (GET_CODE (insn) == NOTE)
2621 switch (NOTE_LINE_NUMBER (insn))
2623 case NOTE_INSN_LOOP_BEG:
2624 next_loop = loops->array + num_loops;
2625 next_loop->num = num_loops;
2627 next_loop->start = insn;
2628 next_loop->outer = current_loop;
2629 current_loop = next_loop;
2632 case NOTE_INSN_SETJMP:
2633 /* In this case, we must invalidate our current loop and any
2635 for (loop = current_loop; loop; loop = loop->outer)
2638 if (loop_dump_stream)
2639 fprintf (loop_dump_stream,
2640 "\nLoop at %d ignored due to setjmp.\n",
2641 INSN_UID (loop->start));
2645 case NOTE_INSN_LOOP_CONT:
2646 current_loop->cont = insn;
2648 case NOTE_INSN_LOOP_END:
2652 current_loop->end = insn;
2653 verify_dominator (current_loop);
2654 current_loop = current_loop->outer;
2660 /* If for any loop, this is a jump insn between the NOTE_INSN_LOOP_CONT
2661 and NOTE_INSN_LOOP_END notes, update loop->dominator. */
2662 else if (GET_CODE (insn) == JUMP_INSN
2663 && GET_CODE (PATTERN (insn)) != RETURN
2666 rtx label = JUMP_LABEL (insn);
2668 if (! any_condjump_p (insn))
2671 loop = current_loop;
2674 /* First see if we care about this loop. */
2675 if (loop->cont && loop->cont_dominator != const0_rtx)
2677 /* If the jump destination is not known, invalidate
2678 loop->const_dominator. */
2680 loop->cont_dominator = const0_rtx;
2682 /* Check if the destination is between loop start and
2684 if ((INSN_LUID (label)
2685 < INSN_LUID (loop->cont))
2686 && (INSN_LUID (label)
2687 > INSN_LUID (loop->start))
2688 /* And if there is no later destination already
2690 && (! loop->cont_dominator
2691 || (INSN_LUID (label)
2692 > INSN_LUID (loop->cont_dominator))))
2693 loop->cont_dominator = label;
2700 /* Note that this will mark the NOTE_INSN_LOOP_END note as being in the
2701 enclosing loop, but this doesn't matter. */
2702 uid_loop[INSN_UID (insn)] = current_loop;
2705 /* Any loop containing a label used in an initializer must be invalidated,
2706 because it can be jumped into from anywhere. */
2708 for (label = forced_labels; label; label = XEXP (label, 1))
2710 for (loop = uid_loop[INSN_UID (XEXP (label, 0))];
2711 loop; loop = loop->outer)
2715 /* Any loop containing a label used for an exception handler must be
2716 invalidated, because it can be jumped into from anywhere. */
2718 for (label = exception_handler_labels; label; label = XEXP (label, 1))
2720 for (loop = uid_loop[INSN_UID (XEXP (label, 0))];
2721 loop; loop = loop->outer)
2725 /* Now scan all insn's in the function. If any JUMP_INSN branches into a
2726 loop that it is not contained within, that loop is marked invalid.
2727 If any INSN or CALL_INSN uses a label's address, then the loop containing
2728 that label is marked invalid, because it could be jumped into from
2731 Also look for blocks of code ending in an unconditional branch that
2732 exits the loop. If such a block is surrounded by a conditional
2733 branch around the block, move the block elsewhere (see below) and
2734 invert the jump to point to the code block. This may eliminate a
2735 label in our loop and will simplify processing by both us and a
2736 possible second cse pass. */
2738 for (insn = f; insn; insn = NEXT_INSN (insn))
2741 struct loop *this_loop = uid_loop[INSN_UID (insn)];
2743 if (GET_CODE (insn) == INSN || GET_CODE (insn) == CALL_INSN)
2745 rtx note = find_reg_note (insn, REG_LABEL, NULL_RTX);
2748 for (loop = uid_loop[INSN_UID (XEXP (note, 0))];
2749 loop; loop = loop->outer)
2754 if (GET_CODE (insn) != JUMP_INSN)
2757 mark_loop_jump (PATTERN (insn), this_loop);
2759 /* See if this is an unconditional branch outside the loop. */
2761 && (GET_CODE (PATTERN (insn)) == RETURN
2762 || (any_uncondjump_p (insn)
2763 && onlyjump_p (insn)
2764 && (uid_loop[INSN_UID (JUMP_LABEL (insn))]
2766 && get_max_uid () < max_uid_for_loop)
2769 rtx our_next = next_real_insn (insn);
2770 rtx last_insn_to_move = NEXT_INSN (insn);
2771 struct loop *dest_loop;
2772 struct loop *outer_loop = NULL;
2774 /* Go backwards until we reach the start of the loop, a label,
2776 for (p = PREV_INSN (insn);
2777 GET_CODE (p) != CODE_LABEL
2778 && ! (GET_CODE (p) == NOTE
2779 && NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
2780 && GET_CODE (p) != JUMP_INSN;
2784 /* Check for the case where we have a jump to an inner nested
2785 loop, and do not perform the optimization in that case. */
2787 if (JUMP_LABEL (insn))
2789 dest_loop = uid_loop[INSN_UID (JUMP_LABEL (insn))];
2792 for (outer_loop = dest_loop; outer_loop;
2793 outer_loop = outer_loop->outer)
2794 if (outer_loop == this_loop)
2799 /* Make sure that the target of P is within the current loop. */
2801 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p)
2802 && uid_loop[INSN_UID (JUMP_LABEL (p))] != this_loop)
2803 outer_loop = this_loop;
2805 /* If we stopped on a JUMP_INSN to the next insn after INSN,
2806 we have a block of code to try to move.
2808 We look backward and then forward from the target of INSN
2809 to find a BARRIER at the same loop depth as the target.
2810 If we find such a BARRIER, we make a new label for the start
2811 of the block, invert the jump in P and point it to that label,
2812 and move the block of code to the spot we found. */
2815 && GET_CODE (p) == JUMP_INSN
2816 && JUMP_LABEL (p) != 0
2817 /* Just ignore jumps to labels that were never emitted.
2818 These always indicate compilation errors. */
2819 && INSN_UID (JUMP_LABEL (p)) != 0
2820 && any_condjump_p (p) && onlyjump_p (p)
2821 && next_real_insn (JUMP_LABEL (p)) == our_next
2822 /* If it's not safe to move the sequence, then we
2824 && insns_safe_to_move_p (p, NEXT_INSN (insn),
2825 &last_insn_to_move))
2828 = JUMP_LABEL (insn) ? JUMP_LABEL (insn) : get_last_insn ();
2829 struct loop *target_loop = uid_loop[INSN_UID (target)];
2832 for (loc = target; loc; loc = PREV_INSN (loc))
2833 if (GET_CODE (loc) == BARRIER
2834 /* Don't move things inside a tablejump. */
2835 && ((loc2 = next_nonnote_insn (loc)) == 0
2836 || GET_CODE (loc2) != CODE_LABEL
2837 || (loc2 = next_nonnote_insn (loc2)) == 0
2838 || GET_CODE (loc2) != JUMP_INSN
2839 || (GET_CODE (PATTERN (loc2)) != ADDR_VEC
2840 && GET_CODE (PATTERN (loc2)) != ADDR_DIFF_VEC))
2841 && uid_loop[INSN_UID (loc)] == target_loop)
2845 for (loc = target; loc; loc = NEXT_INSN (loc))
2846 if (GET_CODE (loc) == BARRIER
2847 /* Don't move things inside a tablejump. */
2848 && ((loc2 = next_nonnote_insn (loc)) == 0
2849 || GET_CODE (loc2) != CODE_LABEL
2850 || (loc2 = next_nonnote_insn (loc2)) == 0
2851 || GET_CODE (loc2) != JUMP_INSN
2852 || (GET_CODE (PATTERN (loc2)) != ADDR_VEC
2853 && GET_CODE (PATTERN (loc2)) != ADDR_DIFF_VEC))
2854 && uid_loop[INSN_UID (loc)] == target_loop)
2859 rtx cond_label = JUMP_LABEL (p);
2860 rtx new_label = get_label_after (p);
2862 /* Ensure our label doesn't go away. */
2863 LABEL_NUSES (cond_label)++;
2865 /* Verify that uid_loop is large enough and that
2867 if (invert_jump (p, new_label, 1))
2871 /* If no suitable BARRIER was found, create a suitable
2872 one before TARGET. Since TARGET is a fall through
2873 path, we'll need to insert an jump around our block
2874 and a add a BARRIER before TARGET.
2876 This creates an extra unconditional jump outside
2877 the loop. However, the benefits of removing rarely
2878 executed instructions from inside the loop usually
2879 outweighs the cost of the extra unconditional jump
2880 outside the loop. */
2885 temp = gen_jump (JUMP_LABEL (insn));
2886 temp = emit_jump_insn_before (temp, target);
2887 JUMP_LABEL (temp) = JUMP_LABEL (insn);
2888 LABEL_NUSES (JUMP_LABEL (insn))++;
2889 loc = emit_barrier_before (target);
2892 /* Include the BARRIER after INSN and copy the
2894 new_label = squeeze_notes (new_label,
2896 reorder_insns (new_label, last_insn_to_move, loc);
2898 /* All those insns are now in TARGET_LOOP. */
2900 q != NEXT_INSN (last_insn_to_move);
2902 uid_loop[INSN_UID (q)] = target_loop;
2904 /* The label jumped to by INSN is no longer a loop
2905 exit. Unless INSN does not have a label (e.g.,
2906 it is a RETURN insn), search loop->exit_labels
2907 to find its label_ref, and remove it. Also turn
2908 off LABEL_OUTSIDE_LOOP_P bit. */
2909 if (JUMP_LABEL (insn))
2912 r = this_loop->exit_labels;
2913 r; q = r, r = LABEL_NEXTREF (r))
2914 if (XEXP (r, 0) == JUMP_LABEL (insn))
2916 LABEL_OUTSIDE_LOOP_P (r) = 0;
2918 LABEL_NEXTREF (q) = LABEL_NEXTREF (r);
2920 this_loop->exit_labels = LABEL_NEXTREF (r);
2924 for (loop = this_loop; loop && loop != target_loop;
2928 /* If we didn't find it, then something is
2934 /* P is now a jump outside the loop, so it must be put
2935 in loop->exit_labels, and marked as such.
2936 The easiest way to do this is to just call
2937 mark_loop_jump again for P. */
2938 mark_loop_jump (PATTERN (p), this_loop);
2940 /* If INSN now jumps to the insn after it,
2942 if (JUMP_LABEL (insn) != 0
2943 && (next_real_insn (JUMP_LABEL (insn))
2944 == next_real_insn (insn)))
2948 /* Continue the loop after where the conditional
2949 branch used to jump, since the only branch insn
2950 in the block (if it still remains) is an inter-loop
2951 branch and hence needs no processing. */
2952 insn = NEXT_INSN (cond_label);
2954 if (--LABEL_NUSES (cond_label) == 0)
2955 delete_insn (cond_label);
2957 /* This loop will be continued with NEXT_INSN (insn). */
2958 insn = PREV_INSN (insn);
2965 /* If any label in X jumps to a loop different from LOOP_NUM and any of the
2966 loops it is contained in, mark the target loop invalid.
2968 For speed, we assume that X is part of a pattern of a JUMP_INSN. */
2971 mark_loop_jump (x, loop)
2975 struct loop *dest_loop;
2976 struct loop *outer_loop;
2979 switch (GET_CODE (x))
2992 /* There could be a label reference in here. */
2993 mark_loop_jump (XEXP (x, 0), loop);
2999 mark_loop_jump (XEXP (x, 0), loop);
3000 mark_loop_jump (XEXP (x, 1), loop);
3004 /* This may refer to a LABEL_REF or SYMBOL_REF. */
3005 mark_loop_jump (XEXP (x, 1), loop);
3010 mark_loop_jump (XEXP (x, 0), loop);
3014 dest_loop = uid_loop[INSN_UID (XEXP (x, 0))];
3016 /* Link together all labels that branch outside the loop. This
3017 is used by final_[bg]iv_value and the loop unrolling code. Also
3018 mark this LABEL_REF so we know that this branch should predict
3021 /* A check to make sure the label is not in an inner nested loop,
3022 since this does not count as a loop exit. */
3025 for (outer_loop = dest_loop; outer_loop;
3026 outer_loop = outer_loop->outer)
3027 if (outer_loop == loop)
3033 if (loop && ! outer_loop)
3035 LABEL_OUTSIDE_LOOP_P (x) = 1;
3036 LABEL_NEXTREF (x) = loop->exit_labels;
3037 loop->exit_labels = x;
3039 for (outer_loop = loop;
3040 outer_loop && outer_loop != dest_loop;
3041 outer_loop = outer_loop->outer)
3042 outer_loop->exit_count++;
3045 /* If this is inside a loop, but not in the current loop or one enclosed
3046 by it, it invalidates at least one loop. */
3051 /* We must invalidate every nested loop containing the target of this
3052 label, except those that also contain the jump insn. */
3054 for (; dest_loop; dest_loop = dest_loop->outer)
3056 /* Stop when we reach a loop that also contains the jump insn. */
3057 for (outer_loop = loop; outer_loop; outer_loop = outer_loop->outer)
3058 if (dest_loop == outer_loop)
3061 /* If we get here, we know we need to invalidate a loop. */
3062 if (loop_dump_stream && ! dest_loop->invalid)
3063 fprintf (loop_dump_stream,
3064 "\nLoop at %d ignored due to multiple entry points.\n",
3065 INSN_UID (dest_loop->start));
3067 dest_loop->invalid = 1;
3072 /* If this is not setting pc, ignore. */
3073 if (SET_DEST (x) == pc_rtx)
3074 mark_loop_jump (SET_SRC (x), loop);
3078 mark_loop_jump (XEXP (x, 1), loop);
3079 mark_loop_jump (XEXP (x, 2), loop);
3084 for (i = 0; i < XVECLEN (x, 0); i++)
3085 mark_loop_jump (XVECEXP (x, 0, i), loop);
3089 for (i = 0; i < XVECLEN (x, 1); i++)
3090 mark_loop_jump (XVECEXP (x, 1, i), loop);
3094 /* Strictly speaking this is not a jump into the loop, only a possible
3095 jump out of the loop. However, we have no way to link the destination
3096 of this jump onto the list of exit labels. To be safe we mark this
3097 loop and any containing loops as invalid. */
3100 for (outer_loop = loop; outer_loop; outer_loop = outer_loop->outer)
3102 if (loop_dump_stream && ! outer_loop->invalid)
3103 fprintf (loop_dump_stream,
3104 "\nLoop at %d ignored due to unknown exit jump.\n",
3105 INSN_UID (outer_loop->start));
3106 outer_loop->invalid = 1;
3113 /* Return nonzero if there is a label in the range from
3114 insn INSN to and including the insn whose luid is END
3115 INSN must have an assigned luid (i.e., it must not have
3116 been previously created by loop.c). */
3119 labels_in_range_p (insn, end)
3123 while (insn && INSN_LUID (insn) <= end)
3125 if (GET_CODE (insn) == CODE_LABEL)
3127 insn = NEXT_INSN (insn);
3133 /* Record that a memory reference X is being set. */
3136 note_addr_stored (x, y, data)
3138 rtx y ATTRIBUTE_UNUSED;
3139 void *data ATTRIBUTE_UNUSED;
3141 if (x == 0 || GET_CODE (x) != MEM)
3144 /* Count number of memory writes.
3145 This affects heuristics in strength_reduce. */
3148 /* BLKmode MEM means all memory is clobbered. */
3149 if (GET_MODE (x) == BLKmode)
3151 if (RTX_UNCHANGING_P (x))
3152 unknown_constant_address_altered = 1;
3154 unknown_address_altered = 1;
3159 loop_store_mems = gen_rtx_EXPR_LIST (VOIDmode, x, loop_store_mems);
3162 /* X is a value modified by an INSN that references a biv inside a loop
3163 exit test (ie, X is somehow related to the value of the biv). If X
3164 is a pseudo that is used more than once, then the biv is (effectively)
3165 used more than once. DATA is really an `int *', and is set if the
3166 biv is used more than once. */
3169 note_set_pseudo_multiple_uses (x, y, data)
3171 rtx y ATTRIBUTE_UNUSED;
3177 while (GET_CODE (x) == STRICT_LOW_PART
3178 || GET_CODE (x) == SIGN_EXTRACT
3179 || GET_CODE (x) == ZERO_EXTRACT
3180 || GET_CODE (x) == SUBREG)
3183 if (GET_CODE (x) != REG || REGNO (x) < FIRST_PSEUDO_REGISTER)
3186 /* If we do not have usage information, or if we know the register
3187 is used more than once, note that fact for check_dbra_loop. */
3188 if (REGNO (x) >= max_reg_before_loop
3189 || ! VARRAY_RTX (reg_single_usage, REGNO (x))
3190 || VARRAY_RTX (reg_single_usage, REGNO (x)) == const0_rtx)
3191 *((int *) data) = 1;
3194 /* Return nonzero if the rtx X is invariant over the current loop.
3196 The value is 2 if we refer to something only conditionally invariant.
3198 If `unknown_address_altered' is nonzero, no memory ref is invariant.
3199 Otherwise, a memory ref is invariant if it does not conflict with
3200 anything stored in `loop_store_mems'. */
3203 loop_invariant_p (loop, x)
3204 const struct loop *loop;
3208 register enum rtx_code code;
3209 register const char *fmt;
3210 int conditional = 0;
3215 code = GET_CODE (x);
3225 /* A LABEL_REF is normally invariant, however, if we are unrolling
3226 loops, and this label is inside the loop, then it isn't invariant.
3227 This is because each unrolled copy of the loop body will have
3228 a copy of this label. If this was invariant, then an insn loading
3229 the address of this label into a register might get moved outside
3230 the loop, and then each loop body would end up using the same label.
3232 We don't know the loop bounds here though, so just fail for all
3234 if (flag_unroll_loops)
3241 case UNSPEC_VOLATILE:
3245 /* We used to check RTX_UNCHANGING_P (x) here, but that is invalid
3246 since the reg might be set by initialization within the loop. */
3248 if ((x == frame_pointer_rtx || x == hard_frame_pointer_rtx
3249 || x == arg_pointer_rtx)
3250 && ! current_function_has_nonlocal_goto)
3253 if (LOOP_INFO (loop)->has_call
3254 && REGNO (x) < FIRST_PSEUDO_REGISTER && call_used_regs[REGNO (x)])
3257 if (VARRAY_INT (set_in_loop, REGNO (x)) < 0)
3260 return VARRAY_INT (set_in_loop, REGNO (x)) == 0;
3263 /* Volatile memory references must be rejected. Do this before
3264 checking for read-only items, so that volatile read-only items
3265 will be rejected also. */
3266 if (MEM_VOLATILE_P (x))
3269 /* If we had a subroutine call, any location in memory could
3270 have been clobbered. We used to test here for volatile and
3271 readonly, but true_dependence knows how to do that better
3273 if (RTX_UNCHANGING_P (x)
3274 ? unknown_constant_address_altered : unknown_address_altered)
3277 /* See if there is any dependence between a store and this load. */
3278 mem_list_entry = loop_store_mems;
3279 while (mem_list_entry)
3281 if (true_dependence (XEXP (mem_list_entry, 0), VOIDmode,
3285 mem_list_entry = XEXP (mem_list_entry, 1);
3288 /* It's not invalidated by a store in memory
3289 but we must still verify the address is invariant. */
3293 /* Don't mess with insns declared volatile. */
3294 if (MEM_VOLATILE_P (x))
3302 fmt = GET_RTX_FORMAT (code);
3303 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3307 int tem = loop_invariant_p (loop, XEXP (x, i));
3313 else if (fmt[i] == 'E')
3316 for (j = 0; j < XVECLEN (x, i); j++)
3318 int tem = loop_invariant_p (loop, XVECEXP (x, i, j));
3328 return 1 + conditional;
3332 /* Return nonzero if all the insns in the loop that set REG
3333 are INSN and the immediately following insns,
3334 and if each of those insns sets REG in an invariant way
3335 (not counting uses of REG in them).
3337 The value is 2 if some of these insns are only conditionally invariant.
3339 We assume that INSN itself is the first set of REG
3340 and that its source is invariant. */
3343 consec_sets_invariant_p (loop, reg, n_sets, insn)
3344 const struct loop *loop;
3349 unsigned int regno = REGNO (reg);
3351 /* Number of sets we have to insist on finding after INSN. */
3352 int count = n_sets - 1;
3353 int old = VARRAY_INT (set_in_loop, regno);
3357 /* If N_SETS hit the limit, we can't rely on its value. */
3361 VARRAY_INT (set_in_loop, regno) = 0;
3365 register enum rtx_code code;
3369 code = GET_CODE (p);
3371 /* If library call, skip to end of it. */
3372 if (code == INSN && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
3377 && (set = single_set (p))
3378 && GET_CODE (SET_DEST (set)) == REG
3379 && REGNO (SET_DEST (set)) == regno)
3381 this = loop_invariant_p (loop, SET_SRC (set));
3384 else if ((temp = find_reg_note (p, REG_EQUAL, NULL_RTX)))
3386 /* If this is a libcall, then any invariant REG_EQUAL note is OK.
3387 If this is an ordinary insn, then only CONSTANT_P REG_EQUAL
3389 this = (CONSTANT_P (XEXP (temp, 0))
3390 || (find_reg_note (p, REG_RETVAL, NULL_RTX)
3391 && loop_invariant_p (loop, XEXP (temp, 0))));
3398 else if (code != NOTE)
3400 VARRAY_INT (set_in_loop, regno) = old;
3405 VARRAY_INT (set_in_loop, regno) = old;
3406 /* If loop_invariant_p ever returned 2, we return 2. */
3407 return 1 + (value & 2);
3411 /* I don't think this condition is sufficient to allow INSN
3412 to be moved, so we no longer test it. */
3414 /* Return 1 if all insns in the basic block of INSN and following INSN
3415 that set REG are invariant according to TABLE. */
3418 all_sets_invariant_p (reg, insn, table)
3422 register rtx p = insn;
3423 register int regno = REGNO (reg);
3427 register enum rtx_code code;
3429 code = GET_CODE (p);
3430 if (code == CODE_LABEL || code == JUMP_INSN)
3432 if (code == INSN && GET_CODE (PATTERN (p)) == SET
3433 && GET_CODE (SET_DEST (PATTERN (p))) == REG
3434 && REGNO (SET_DEST (PATTERN (p))) == regno)
3436 if (! loop_invariant_p (loop, SET_SRC (PATTERN (p)), table))
3443 /* Look at all uses (not sets) of registers in X. For each, if it is
3444 the single use, set USAGE[REGNO] to INSN; if there was a previous use in
3445 a different insn, set USAGE[REGNO] to const0_rtx. */
3448 find_single_use_in_loop (insn, x, usage)
3453 enum rtx_code code = GET_CODE (x);
3454 const char *fmt = GET_RTX_FORMAT (code);
3458 VARRAY_RTX (usage, REGNO (x))
3459 = (VARRAY_RTX (usage, REGNO (x)) != 0
3460 && VARRAY_RTX (usage, REGNO (x)) != insn)
3461 ? const0_rtx : insn;
3463 else if (code == SET)
3465 /* Don't count SET_DEST if it is a REG; otherwise count things
3466 in SET_DEST because if a register is partially modified, it won't
3467 show up as a potential movable so we don't care how USAGE is set
3469 if (GET_CODE (SET_DEST (x)) != REG)
3470 find_single_use_in_loop (insn, SET_DEST (x), usage);
3471 find_single_use_in_loop (insn, SET_SRC (x), usage);
3474 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3476 if (fmt[i] == 'e' && XEXP (x, i) != 0)
3477 find_single_use_in_loop (insn, XEXP (x, i), usage);
3478 else if (fmt[i] == 'E')
3479 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
3480 find_single_use_in_loop (insn, XVECEXP (x, i, j), usage);
3484 /* Count and record any set in X which is contained in INSN. Update
3485 MAY_NOT_MOVE and LAST_SET for any register set in X. */
3488 count_one_set (insn, x, may_not_move, last_set)
3490 varray_type may_not_move;
3493 if (GET_CODE (x) == CLOBBER && GET_CODE (XEXP (x, 0)) == REG)
3494 /* Don't move a reg that has an explicit clobber.
3495 It's not worth the pain to try to do it correctly. */
3496 VARRAY_CHAR (may_not_move, REGNO (XEXP (x, 0))) = 1;
3498 if (GET_CODE (x) == SET || GET_CODE (x) == CLOBBER)
3500 rtx dest = SET_DEST (x);
3501 while (GET_CODE (dest) == SUBREG
3502 || GET_CODE (dest) == ZERO_EXTRACT
3503 || GET_CODE (dest) == SIGN_EXTRACT
3504 || GET_CODE (dest) == STRICT_LOW_PART)
3505 dest = XEXP (dest, 0);
3506 if (GET_CODE (dest) == REG)
3508 register int regno = REGNO (dest);
3509 /* If this is the first setting of this reg
3510 in current basic block, and it was set before,
3511 it must be set in two basic blocks, so it cannot
3512 be moved out of the loop. */
3513 if (VARRAY_INT (set_in_loop, regno) > 0
3514 && last_set[regno] == 0)
3515 VARRAY_CHAR (may_not_move, regno) = 1;
3516 /* If this is not first setting in current basic block,
3517 see if reg was used in between previous one and this.
3518 If so, neither one can be moved. */
3519 if (last_set[regno] != 0
3520 && reg_used_between_p (dest, last_set[regno], insn))
3521 VARRAY_CHAR (may_not_move, regno) = 1;
3522 if (VARRAY_INT (set_in_loop, regno) < 127)
3523 ++VARRAY_INT (set_in_loop, regno);
3524 last_set[regno] = insn;
3529 /* Increment SET_IN_LOOP at the index of each register
3530 that is modified by an insn between FROM and TO.
3531 If the value of an element of SET_IN_LOOP becomes 127 or more,
3532 stop incrementing it, to avoid overflow.
3534 Store in SINGLE_USAGE[I] the single insn in which register I is
3535 used, if it is only used once. Otherwise, it is set to 0 (for no
3536 uses) or const0_rtx for more than one use. This parameter may be zero,
3537 in which case this processing is not done.
3539 Store in *COUNT_PTR the number of actual instruction
3540 in the loop. We use this to decide what is worth moving out. */
3542 /* last_set[n] is nonzero iff reg n has been set in the current basic block.
3543 In that case, it is the insn that last set reg n. */
3546 count_loop_regs_set (from, to, may_not_move, single_usage, count_ptr, nregs)
3547 register rtx from, to;
3548 varray_type may_not_move;
3549 varray_type single_usage;
3553 register rtx *last_set = (rtx *) xcalloc (nregs, sizeof (rtx));
3555 register int count = 0;
3557 for (insn = from; insn != to; insn = NEXT_INSN (insn))
3563 /* Record registers that have exactly one use. */
3564 find_single_use_in_loop (insn, PATTERN (insn), single_usage);
3566 /* Include uses in REG_EQUAL notes. */
3567 if (REG_NOTES (insn))
3568 find_single_use_in_loop (insn, REG_NOTES (insn), single_usage);
3570 if (GET_CODE (PATTERN (insn)) == SET
3571 || GET_CODE (PATTERN (insn)) == CLOBBER)
3572 count_one_set (insn, PATTERN (insn), may_not_move, last_set);
3573 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
3576 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
3577 count_one_set (insn, XVECEXP (PATTERN (insn), 0, i),
3578 may_not_move, last_set);
3582 if (GET_CODE (insn) == CODE_LABEL || GET_CODE (insn) == JUMP_INSN)
3583 bzero ((char *) last_set, nregs * sizeof (rtx));
3591 /* Given a loop that is bounded by LOOP->START and LOOP->END and that
3592 is entered at LOOP->SCAN_START, return 1 if the register set in SET
3593 contained in insn INSN is used by any insn that precedes INSN in
3594 cyclic order starting from the loop entry point.
3596 We don't want to use INSN_LUID here because if we restrict INSN to those
3597 that have a valid INSN_LUID, it means we cannot move an invariant out
3598 from an inner loop past two loops. */
3601 loop_reg_used_before_p (loop, set, insn)
3602 const struct loop *loop;
3605 rtx reg = SET_DEST (set);
3608 /* Scan forward checking for register usage. If we hit INSN, we
3609 are done. Otherwise, if we hit LOOP->END, wrap around to LOOP->START. */
3610 for (p = loop->scan_start; p != insn; p = NEXT_INSN (p))
3612 if (INSN_P (p) && reg_overlap_mentioned_p (reg, PATTERN (p)))
3622 /* A "basic induction variable" or biv is a pseudo reg that is set
3623 (within this loop) only by incrementing or decrementing it. */
3624 /* A "general induction variable" or giv is a pseudo reg whose
3625 value is a linear function of a biv. */
3627 /* Bivs are recognized by `basic_induction_var';
3628 Givs by `general_induction_var'. */
3630 /* Indexed by register number, indicates whether or not register is an
3631 induction variable, and if so what type. */
3633 varray_type reg_iv_type;
3635 /* Indexed by register number, contains pointer to `struct induction'
3636 if register is an induction variable. This holds general info for
3637 all induction variables. */
3639 varray_type reg_iv_info;
3641 /* Indexed by register number, contains pointer to `struct iv_class'
3642 if register is a basic induction variable. This holds info describing
3643 the class (a related group) of induction variables that the biv belongs
3646 struct iv_class **reg_biv_class;
3648 /* The head of a list which links together (via the next field)
3649 every iv class for the current loop. */
3651 struct iv_class *loop_iv_list;
3653 /* Givs made from biv increments are always splittable for loop unrolling.
3654 Since there is no regscan info for them, we have to keep track of them
3656 unsigned int first_increment_giv, last_increment_giv;
3658 /* Communication with routines called via `note_stores'. */
3660 static rtx note_insn;
3662 /* Dummy register to have non-zero DEST_REG for DEST_ADDR type givs. */
3664 static rtx addr_placeholder;
3666 /* ??? Unfinished optimizations, and possible future optimizations,
3667 for the strength reduction code. */
3669 /* ??? The interaction of biv elimination, and recognition of 'constant'
3670 bivs, may cause problems. */
3672 /* ??? Add heuristics so that DEST_ADDR strength reduction does not cause
3673 performance problems.
3675 Perhaps don't eliminate things that can be combined with an addressing
3676 mode. Find all givs that have the same biv, mult_val, and add_val;
3677 then for each giv, check to see if its only use dies in a following
3678 memory address. If so, generate a new memory address and check to see
3679 if it is valid. If it is valid, then store the modified memory address,
3680 otherwise, mark the giv as not done so that it will get its own iv. */
3682 /* ??? Could try to optimize branches when it is known that a biv is always
3685 /* ??? When replace a biv in a compare insn, we should replace with closest
3686 giv so that an optimized branch can still be recognized by the combiner,
3687 e.g. the VAX acb insn. */
3689 /* ??? Many of the checks involving uid_luid could be simplified if regscan
3690 was rerun in loop_optimize whenever a register was added or moved.
3691 Also, some of the optimizations could be a little less conservative. */
3693 /* Scan the loop body and call FNCALL for each insn. In the addition to the
3694 LOOP and INSN parameters pass MAYBE_MULTIPLE and NOT_EVERY_ITERATION to the
3697 NOT_EVERY_ITERATION if current insn is not executed at least once for every
3698 loop iteration except for the last one.
3700 MAYBE_MULTIPLE is 1 if current insn may be executed more than once for every
3704 for_each_insn_in_loop (loop, fncall)
3706 loop_insn_callback fncall;
3708 /* This is 1 if current insn is not executed at least once for every loop
3710 int not_every_iteration = 0;
3711 int maybe_multiple = 0;
3712 int past_loop_latch = 0;
3716 /* If loop_scan_start points to the loop exit test, we have to be wary of
3717 subversive use of gotos inside expression statements. */
3718 if (prev_nonnote_insn (loop->scan_start) != prev_nonnote_insn (loop->start))
3719 maybe_multiple = back_branch_in_range_p (loop, loop->scan_start);
3721 /* Scan through loop to find all possible bivs. */
3723 for (p = next_insn_in_loop (loop, loop->scan_start);
3725 p = next_insn_in_loop (loop, p))
3727 p = fncall (loop, p, not_every_iteration, maybe_multiple);
3729 /* Past CODE_LABEL, we get to insns that may be executed multiple
3730 times. The only way we can be sure that they can't is if every
3731 jump insn between here and the end of the loop either
3732 returns, exits the loop, is a jump to a location that is still
3733 behind the label, or is a jump to the loop start. */
3735 if (GET_CODE (p) == CODE_LABEL)
3743 insn = NEXT_INSN (insn);
3744 if (insn == loop->scan_start)
3746 if (insn == loop->end)
3752 if (insn == loop->scan_start)
3756 if (GET_CODE (insn) == JUMP_INSN
3757 && GET_CODE (PATTERN (insn)) != RETURN
3758 && (!any_condjump_p (insn)
3759 || (JUMP_LABEL (insn) != 0
3760 && JUMP_LABEL (insn) != loop->scan_start
3761 && !loop_insn_first_p (p, JUMP_LABEL (insn)))))
3769 /* Past a jump, we get to insns for which we can't count
3770 on whether they will be executed during each iteration. */
3771 /* This code appears twice in strength_reduce. There is also similar
3772 code in scan_loop. */
3773 if (GET_CODE (p) == JUMP_INSN
3774 /* If we enter the loop in the middle, and scan around to the
3775 beginning, don't set not_every_iteration for that.
3776 This can be any kind of jump, since we want to know if insns
3777 will be executed if the loop is executed. */
3778 && !(JUMP_LABEL (p) == loop->top
3779 && ((NEXT_INSN (NEXT_INSN (p)) == loop->end
3780 && any_uncondjump_p (p))
3781 || (NEXT_INSN (p) == loop->end && any_condjump_p (p)))))
3785 /* If this is a jump outside the loop, then it also doesn't
3786 matter. Check to see if the target of this branch is on the
3787 loop->exits_labels list. */
3789 for (label = loop->exit_labels; label; label = LABEL_NEXTREF (label))
3790 if (XEXP (label, 0) == JUMP_LABEL (p))
3794 not_every_iteration = 1;
3797 else if (GET_CODE (p) == NOTE)
3799 /* At the virtual top of a converted loop, insns are again known to
3800 be executed each iteration: logically, the loop begins here
3801 even though the exit code has been duplicated.
3803 Insns are also again known to be executed each iteration at
3804 the LOOP_CONT note. */
3805 if ((NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP
3806 || NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_CONT)
3808 not_every_iteration = 0;
3809 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
3811 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END)
3815 /* Note if we pass a loop latch. If we do, then we can not clear
3816 NOT_EVERY_ITERATION below when we pass the last CODE_LABEL in
3817 a loop since a jump before the last CODE_LABEL may have started
3818 a new loop iteration.
3820 Note that LOOP_TOP is only set for rotated loops and we need
3821 this check for all loops, so compare against the CODE_LABEL
3822 which immediately follows LOOP_START. */
3823 if (GET_CODE (p) == JUMP_INSN
3824 && JUMP_LABEL (p) == NEXT_INSN (loop->start))
3825 past_loop_latch = 1;
3827 /* Unlike in the code motion pass where MAYBE_NEVER indicates that
3828 an insn may never be executed, NOT_EVERY_ITERATION indicates whether
3829 or not an insn is known to be executed each iteration of the
3830 loop, whether or not any iterations are known to occur.
3832 Therefore, if we have just passed a label and have no more labels
3833 between here and the test insn of the loop, and we have not passed
3834 a jump to the top of the loop, then we know these insns will be
3835 executed each iteration. */
3837 if (not_every_iteration
3839 && GET_CODE (p) == CODE_LABEL
3840 && no_labels_between_p (p, loop->end)
3841 && loop_insn_first_p (p, loop->cont))
3842 not_every_iteration = 0;
3846 /* Perform strength reduction and induction variable elimination.
3848 Pseudo registers created during this function will be beyond the last
3849 valid index in several tables including n_times_set and regno_last_uid.
3850 This does not cause a problem here, because the added registers cannot be
3851 givs outside of their loop, and hence will never be reconsidered.
3852 But scan_loop must check regnos to make sure they are in bounds. */
3855 strength_reduce (loop, insn_count, flags)
3861 /* Temporary list pointers for traversing loop_iv_list. */
3862 struct iv_class *bl, **backbl;
3863 struct loop_info *loop_info = LOOP_INFO (loop);
3864 /* Ratio of extra register life span we can justify
3865 for saving an instruction. More if loop doesn't call subroutines
3866 since in that case saving an insn makes more difference
3867 and more registers are available. */
3868 /* ??? could set this to last value of threshold in move_movables */
3869 int threshold = (loop_info->has_call ? 1 : 2) * (3 + n_non_fixed_regs);
3870 /* Map of pseudo-register replacements. */
3871 rtx *reg_map = NULL;
3875 rtx end_insert_before;
3876 int n_extra_increment;
3877 int unrolled_insn_copies = 0;
3878 rtx loop_start = loop->start;
3879 rtx loop_end = loop->end;
3880 rtx loop_scan_start = loop->scan_start;
3882 VARRAY_INT_INIT (reg_iv_type, max_reg_before_loop, "reg_iv_type");
3883 VARRAY_GENERIC_PTR_INIT (reg_iv_info, max_reg_before_loop, "reg_iv_info");
3884 reg_biv_class = (struct iv_class **)
3885 xcalloc (max_reg_before_loop, sizeof (struct iv_class *));
3888 addr_placeholder = gen_reg_rtx (Pmode);
3890 /* Save insn immediately after the loop_end. Insns inserted after loop_end
3891 must be put before this insn, so that they will appear in the right
3892 order (i.e. loop order).
3894 If loop_end is the end of the current function, then emit a
3895 NOTE_INSN_DELETED after loop_end and set end_insert_before to the
3897 if (NEXT_INSN (loop_end) != 0)
3898 end_insert_before = NEXT_INSN (loop_end);
3900 end_insert_before = emit_note_after (NOTE_INSN_DELETED, loop_end);
3902 for_each_insn_in_loop (loop, check_insn_for_bivs);
3904 /* Scan loop_iv_list to remove all regs that proved not to be bivs.
3905 Make a sanity check against n_times_set. */
3906 for (backbl = &loop_iv_list, bl = *backbl; bl; bl = bl->next)
3908 if (REG_IV_TYPE (bl->regno) != BASIC_INDUCT
3909 /* Above happens if register modified by subreg, etc. */
3910 /* Make sure it is not recognized as a basic induction var: */
3911 || VARRAY_INT (n_times_set, bl->regno) != bl->biv_count
3912 /* If never incremented, it is invariant that we decided not to
3913 move. So leave it alone. */
3914 || ! bl->incremented)
3916 if (loop_dump_stream)
3917 fprintf (loop_dump_stream, "Reg %d: biv discarded, %s\n",
3919 (REG_IV_TYPE (bl->regno) != BASIC_INDUCT
3920 ? "not induction variable"
3921 : (! bl->incremented ? "never incremented"
3924 REG_IV_TYPE (bl->regno) = NOT_BASIC_INDUCT;
3931 if (loop_dump_stream)
3932 fprintf (loop_dump_stream, "Reg %d: biv verified\n", bl->regno);
3936 /* Exit if there are no bivs. */
3939 /* Can still unroll the loop anyways, but indicate that there is no
3940 strength reduction info available. */
3941 if (flags & LOOP_UNROLL)
3942 unroll_loop (loop, insn_count, end_insert_before, 0);
3947 /* Find initial value for each biv by searching backwards from loop_start,
3948 halting at first label. Also record any test condition. */
3951 for (p = loop_start; p && GET_CODE (p) != CODE_LABEL; p = PREV_INSN (p))
3955 if (GET_CODE (p) == CALL_INSN)
3959 note_stores (PATTERN (p), record_initial, NULL);
3961 /* Record any test of a biv that branches around the loop if no store
3962 between it and the start of loop. We only care about tests with
3963 constants and registers and only certain of those. */
3964 if (GET_CODE (p) == JUMP_INSN
3965 && JUMP_LABEL (p) != 0
3966 && next_real_insn (JUMP_LABEL (p)) == next_real_insn (loop_end)
3967 && (test = get_condition_for_loop (loop, p)) != 0
3968 && GET_CODE (XEXP (test, 0)) == REG
3969 && REGNO (XEXP (test, 0)) < max_reg_before_loop
3970 && (bl = reg_biv_class[REGNO (XEXP (test, 0))]) != 0
3971 && valid_initial_value_p (XEXP (test, 1), p, call_seen, loop_start)
3972 && bl->init_insn == 0)
3974 /* If an NE test, we have an initial value! */
3975 if (GET_CODE (test) == NE)
3978 bl->init_set = gen_rtx_SET (VOIDmode,
3979 XEXP (test, 0), XEXP (test, 1));
3982 bl->initial_test = test;
3986 /* Look at the each biv and see if we can say anything better about its
3987 initial value from any initializing insns set up above. (This is done
3988 in two passes to avoid missing SETs in a PARALLEL.) */
3989 for (backbl = &loop_iv_list; (bl = *backbl); backbl = &bl->next)
3994 if (! bl->init_insn)
3997 /* IF INIT_INSN has a REG_EQUAL or REG_EQUIV note and the value
3998 is a constant, use the value of that. */
3999 if (((note = find_reg_note (bl->init_insn, REG_EQUAL, 0)) != NULL
4000 && CONSTANT_P (XEXP (note, 0)))
4001 || ((note = find_reg_note (bl->init_insn, REG_EQUIV, 0)) != NULL
4002 && CONSTANT_P (XEXP (note, 0))))
4003 src = XEXP (note, 0);
4005 src = SET_SRC (bl->init_set);
4007 if (loop_dump_stream)
4008 fprintf (loop_dump_stream,
4009 "Biv %d initialized at insn %d: initial value ",
4010 bl->regno, INSN_UID (bl->init_insn));
4012 if ((GET_MODE (src) == GET_MODE (regno_reg_rtx[bl->regno])
4013 || GET_MODE (src) == VOIDmode)
4014 && valid_initial_value_p (src, bl->init_insn, call_seen, loop_start))
4016 bl->initial_value = src;
4018 if (loop_dump_stream)
4020 if (GET_CODE (src) == CONST_INT)
4022 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (src));
4023 fputc ('\n', loop_dump_stream);
4027 print_rtl (loop_dump_stream, src);
4028 fprintf (loop_dump_stream, "\n");
4034 struct iv_class *bl2 = 0;
4035 rtx increment = NULL_RTX;
4037 /* Biv initial value is not a simple move. If it is the sum of
4038 another biv and a constant, check if both bivs are incremented
4039 in lockstep. Then we are actually looking at a giv.
4040 For simplicity, we only handle the case where there is but a
4041 single increment, and the register is not used elsewhere. */
4042 if (bl->biv_count == 1
4043 && bl->regno < max_reg_before_loop
4044 && uid_luid[REGNO_LAST_UID (bl->regno)] < INSN_LUID (loop_end)
4045 && GET_CODE (src) == PLUS
4046 && GET_CODE (XEXP (src, 0)) == REG
4047 && CONSTANT_P (XEXP (src, 1))
4048 && ((increment = biv_total_increment (bl)) != NULL_RTX))
4050 unsigned int regno = REGNO (XEXP (src, 0));
4052 for (bl2 = loop_iv_list; bl2; bl2 = bl2->next)
4053 if (bl2->regno == regno)
4057 /* Now, can we transform this biv into a giv? */
4059 && bl2->biv_count == 1
4060 && rtx_equal_p (increment, biv_total_increment (bl2))
4061 /* init_insn is only set to insns that are before loop_start
4062 without any intervening labels. */
4063 && ! reg_set_between_p (bl2->biv->src_reg,
4064 PREV_INSN (bl->init_insn), loop_start)
4065 /* The register from BL2 must be set before the register from
4066 BL is set, or we must be able to move the latter set after
4067 the former set. Currently there can't be any labels
4068 in-between when biv_total_increment returns nonzero both times
4069 but we test it here in case some day some real cfg analysis
4070 gets used to set always_computable. */
4071 && (loop_insn_first_p (bl2->biv->insn, bl->biv->insn)
4072 ? no_labels_between_p (bl2->biv->insn, bl->biv->insn)
4073 : (! reg_used_between_p (bl->biv->src_reg, bl->biv->insn,
4075 && no_jumps_between_p (bl->biv->insn, bl2->biv->insn)))
4076 && validate_change (bl->biv->insn,
4077 &SET_SRC (single_set (bl->biv->insn)),
4080 rtx dominator = loop->cont_dominator;
4081 rtx giv = bl->biv->src_reg;
4082 rtx giv_insn = bl->biv->insn;
4083 rtx after_giv = NEXT_INSN (giv_insn);
4085 if (loop_dump_stream)
4086 fprintf (loop_dump_stream, "is giv of biv %d\n", bl2->regno);
4087 /* Let this giv be discovered by the generic code. */
4088 REG_IV_TYPE (bl->regno) = UNKNOWN_INDUCT;
4089 reg_biv_class[bl->regno] = (struct iv_class *) NULL_PTR;
4090 /* We can get better optimization if we can move the giv setting
4091 before the first giv use. */
4093 && ! loop_insn_first_p (dominator, loop_scan_start)
4094 && ! reg_set_between_p (bl2->biv->src_reg, loop_start,
4096 && ! reg_used_between_p (giv, loop_start, dominator)
4097 && ! reg_used_between_p (giv, giv_insn, loop_end))
4102 for (next = NEXT_INSN (dominator); ; next = NEXT_INSN (next))
4104 if (GET_CODE (next) == JUMP_INSN
4106 && insn_dependent_p (giv_insn, next)))
4109 if (! INSN_P (next) || ! sets_cc0_p (PATTERN (next)))
4113 if (loop_dump_stream)
4114 fprintf (loop_dump_stream, "move after insn %d\n",
4115 INSN_UID (dominator));
4116 /* Avoid problems with luids by actually moving the insn
4117 and adjusting all luids in the range. */
4118 reorder_insns (giv_insn, giv_insn, dominator);
4119 for (p = dominator; INSN_UID (p) >= max_uid_for_loop; )
4121 compute_luids (giv_insn, after_giv, INSN_LUID (p));
4122 /* If the only purpose of the init insn is to initialize
4123 this giv, delete it. */
4124 if (single_set (bl->init_insn)
4125 && ! reg_used_between_p (giv, bl->init_insn, loop_start))
4126 delete_insn (bl->init_insn);
4128 else if (! loop_insn_first_p (bl2->biv->insn, bl->biv->insn))
4130 rtx p = PREV_INSN (giv_insn);
4131 while (INSN_UID (p) >= max_uid_for_loop)
4133 reorder_insns (giv_insn, giv_insn, bl2->biv->insn);
4134 compute_luids (after_giv, NEXT_INSN (giv_insn),
4137 /* Remove this biv from the chain. */
4141 /* If we can't make it a giv,
4142 let biv keep initial value of "itself". */
4143 else if (loop_dump_stream)
4144 fprintf (loop_dump_stream, "is complex\n");
4148 /* If a biv is unconditionally incremented several times in a row, convert
4149 all but the last increment into a giv. */
4151 /* Get an upper bound for the number of registers
4152 we might have after all bivs have been processed. */
4153 first_increment_giv = max_reg_num ();
4154 for (n_extra_increment = 0, bl = loop_iv_list; bl; bl = bl->next)
4155 n_extra_increment += bl->biv_count - 1;
4157 /* If the loop contains volatile memory references do not allow any
4158 replacements to take place, since this could loose the volatile
4160 if (n_extra_increment && ! loop_info->has_volatile)
4162 unsigned int nregs = first_increment_giv + n_extra_increment;
4164 /* Reallocate reg_iv_type and reg_iv_info. */
4165 VARRAY_GROW (reg_iv_type, nregs);
4166 VARRAY_GROW (reg_iv_info, nregs);
4168 for (bl = loop_iv_list; bl; bl = bl->next)
4170 struct induction **vp, *v, *next;
4171 int biv_dead_after_loop = 0;
4173 /* The biv increments lists are in reverse order. Fix this
4175 for (v = bl->biv, bl->biv = 0; v; v = next)
4178 v->next_iv = bl->biv;
4182 /* We must guard against the case that an early exit between v->insn
4183 and next->insn leaves the biv live after the loop, since that
4184 would mean that we'd be missing an increment for the final
4185 value. The following test to set biv_dead_after_loop is like
4186 the first part of the test to set bl->eliminable.
4187 We don't check here if we can calculate the final value, since
4188 this can't succeed if we already know that there is a jump
4189 between v->insn and next->insn, yet next->always_executed is
4190 set and next->maybe_multiple is cleared. Such a combination
4191 implies that the jump destination is outside the loop.
4192 If we want to make this check more sophisticated, we should
4193 check each branch between v->insn and next->insn individually
4194 to see if the biv is dead at its destination. */
4196 if (uid_luid[REGNO_LAST_UID (bl->regno)] < INSN_LUID (loop_end)
4198 && INSN_UID (bl->init_insn) < max_uid_for_loop
4199 && (uid_luid[REGNO_FIRST_UID (bl->regno)]
4200 >= INSN_LUID (bl->init_insn))
4201 #ifdef HAVE_decrement_and_branch_until_zero
4204 && ! reg_mentioned_p (bl->biv->dest_reg, SET_SRC (bl->init_set)))
4205 biv_dead_after_loop = 1;
4207 for (vp = &bl->biv, next = *vp; v = next, next = v->next_iv;)
4209 HOST_WIDE_INT offset;
4210 rtx set, add_val, old_reg, dest_reg, last_use_insn, note;
4211 int old_regno, new_regno;
4214 if (! v->always_executed
4215 || v->maybe_multiple
4216 || GET_CODE (v->add_val) != CONST_INT
4217 || ! next->always_executed
4218 || next->maybe_multiple
4219 || ! CONSTANT_P (next->add_val)
4220 || v->mult_val != const1_rtx
4221 || next->mult_val != const1_rtx
4222 || ! (biv_dead_after_loop
4223 || no_jumps_between_p (v->insn, next->insn)))
4228 offset = INTVAL (v->add_val);
4229 set = single_set (v->insn);
4230 add_val = plus_constant (next->add_val, offset);
4231 old_reg = v->dest_reg;
4232 dest_reg = gen_reg_rtx (v->mode);
4234 /* Unlike reg_iv_type / reg_iv_info, the other three arrays
4235 have been allocated with some slop space, so we may not
4236 actually need to reallocate them. If we do, the following
4237 if statement will be executed just once in this loop. */
4238 if ((unsigned) max_reg_num () > n_times_set->num_elements)
4240 /* Grow all the remaining arrays. */
4241 VARRAY_GROW (set_in_loop, nregs);
4242 VARRAY_GROW (n_times_set, nregs);
4243 VARRAY_GROW (may_not_optimize, nregs);
4244 VARRAY_GROW (reg_single_usage, nregs);
4247 /* Some bivs are incremented with a multi-insn sequence.
4248 The first insn contains the add. */
4249 next_loc_insn = next->insn;
4250 while (! loc_mentioned_in_p (next->location,
4251 PATTERN (next_loc_insn)))
4252 next_loc_insn = PREV_INSN (next_loc_insn);
4254 if (next_loc_insn == v->insn)
4257 if (! validate_change (next_loc_insn, next->location, add_val, 0))
4263 /* Here we can try to eliminate the increment by combining
4264 it into the uses. */
4266 /* Set last_use_insn so that we can check against it. */
4268 for (last_use_insn = v->insn, p = NEXT_INSN (v->insn);
4270 p = next_insn_in_loop (loop, p))
4274 if (reg_mentioned_p (old_reg, PATTERN (p)))
4280 /* If we can't get the LUIDs for the insns, we can't
4281 calculate the lifetime. This is likely from unrolling
4282 of an inner loop, so there is little point in making this
4283 a DEST_REG giv anyways. */
4284 if (INSN_UID (v->insn) >= max_uid_for_loop
4285 || INSN_UID (last_use_insn) >= max_uid_for_loop
4286 || ! validate_change (v->insn, &SET_DEST (set), dest_reg, 0))
4288 /* Change the increment at NEXT back to what it was. */
4289 if (! validate_change (next_loc_insn, next->location,
4295 next->add_val = add_val;
4296 v->dest_reg = dest_reg;
4297 v->giv_type = DEST_REG;
4298 v->location = &SET_SRC (set);
4300 v->combined_with = 0;
4302 v->derive_adjustment = 0;
4308 v->auto_inc_opt = 0;
4311 v->derived_from = 0;
4312 v->always_computable = 1;
4313 v->always_executed = 1;
4315 v->no_const_addval = 0;
4317 old_regno = REGNO (old_reg);
4318 new_regno = REGNO (dest_reg);
4319 VARRAY_INT (set_in_loop, old_regno)--;
4320 VARRAY_INT (set_in_loop, new_regno) = 1;
4321 VARRAY_INT (n_times_set, old_regno)--;
4322 VARRAY_INT (n_times_set, new_regno) = 1;
4323 VARRAY_CHAR (may_not_optimize, new_regno) = 0;
4325 REG_IV_TYPE (new_regno) = GENERAL_INDUCT;
4326 REG_IV_INFO (new_regno) = v;
4328 /* If next_insn has a REG_EQUAL note that mentiones OLD_REG,
4329 it must be replaced. */
4330 note = find_reg_note (next->insn, REG_EQUAL, NULL_RTX);
4331 if (note && reg_mentioned_p (old_reg, XEXP (note, 0)))
4332 XEXP (note, 0) = copy_rtx (SET_SRC (single_set (next->insn)));
4334 /* Remove the increment from the list of biv increments,
4335 and record it as a giv. */
4338 v->next_iv = bl->giv;
4341 v->benefit = rtx_cost (SET_SRC (set), SET);
4342 bl->total_benefit += v->benefit;
4344 /* Now replace the biv with DEST_REG in all insns between
4345 the replaced increment and the next increment, and
4346 remember the last insn that needed a replacement. */
4347 for (last_use_insn = v->insn, p = NEXT_INSN (v->insn);
4349 p = next_insn_in_loop (loop, p))
4355 if (reg_mentioned_p (old_reg, PATTERN (p)))
4358 if (! validate_replace_rtx (old_reg, dest_reg, p))
4361 for (note = REG_NOTES (p); note; note = XEXP (note, 1))
4363 if (GET_CODE (note) == EXPR_LIST)
4365 = replace_rtx (XEXP (note, 0), old_reg, dest_reg);
4369 v->last_use = last_use_insn;
4370 v->lifetime = INSN_LUID (last_use_insn) - INSN_LUID (v->insn);
4371 /* If the lifetime is zero, it means that this register is really
4372 a dead store. So mark this as a giv that can be ignored.
4373 This will not prevent the biv from being eliminated. */
4374 if (v->lifetime == 0)
4377 if (loop_dump_stream)
4378 fprintf (loop_dump_stream,
4379 "Increment %d of biv %d converted to giv %d.\n\n",
4380 INSN_UID (v->insn), old_regno, new_regno);
4384 last_increment_giv = max_reg_num () - 1;
4386 /* Search the loop for general induction variables. */
4388 for_each_insn_in_loop (loop, check_insn_for_givs);
4390 /* Try to calculate and save the number of loop iterations. This is
4391 set to zero if the actual number can not be calculated. This must
4392 be called after all giv's have been identified, since otherwise it may
4393 fail if the iteration variable is a giv. */
4395 loop_iterations (loop);
4397 /* Now for each giv for which we still don't know whether or not it is
4398 replaceable, check to see if it is replaceable because its final value
4399 can be calculated. This must be done after loop_iterations is called,
4400 so that final_giv_value will work correctly. */
4402 for (bl = loop_iv_list; bl; bl = bl->next)
4404 struct induction *v;
4406 for (v = bl->giv; v; v = v->next_iv)
4407 if (! v->replaceable && ! v->not_replaceable)
4408 check_final_value (loop, v);
4411 /* Try to prove that the loop counter variable (if any) is always
4412 nonnegative; if so, record that fact with a REG_NONNEG note
4413 so that "decrement and branch until zero" insn can be used. */
4414 check_dbra_loop (loop, insn_count);
4416 /* Create reg_map to hold substitutions for replaceable giv regs.
4417 Some givs might have been made from biv increments, so look at
4418 reg_iv_type for a suitable size. */
4419 reg_map_size = reg_iv_type->num_elements;
4420 reg_map = (rtx *) xcalloc (reg_map_size, sizeof (rtx));
4422 /* Examine each iv class for feasibility of strength reduction/induction
4423 variable elimination. */
4425 for (bl = loop_iv_list; bl; bl = bl->next)
4427 struct induction *v;
4430 rtx final_value = 0;
4433 /* Test whether it will be possible to eliminate this biv
4434 provided all givs are reduced. This is possible if either
4435 the reg is not used outside the loop, or we can compute
4436 what its final value will be.
4438 For architectures with a decrement_and_branch_until_zero insn,
4439 don't do this if we put a REG_NONNEG note on the endtest for
4442 /* Compare against bl->init_insn rather than loop_start.
4443 We aren't concerned with any uses of the biv between
4444 init_insn and loop_start since these won't be affected
4445 by the value of the biv elsewhere in the function, so
4446 long as init_insn doesn't use the biv itself.
4447 March 14, 1989 -- self@bayes.arc.nasa.gov */
4449 if ((uid_luid[REGNO_LAST_UID (bl->regno)] < INSN_LUID (loop_end)
4451 && INSN_UID (bl->init_insn) < max_uid_for_loop
4452 && uid_luid[REGNO_FIRST_UID (bl->regno)] >= INSN_LUID (bl->init_insn)
4453 #ifdef HAVE_decrement_and_branch_until_zero
4456 && ! reg_mentioned_p (bl->biv->dest_reg, SET_SRC (bl->init_set)))
4457 || ((final_value = final_biv_value (loop, bl))
4458 #ifdef HAVE_decrement_and_branch_until_zero
4462 bl->eliminable = maybe_eliminate_biv (loop, bl, 0, threshold,
4466 if (loop_dump_stream)
4468 fprintf (loop_dump_stream,
4469 "Cannot eliminate biv %d.\n",
4471 fprintf (loop_dump_stream,
4472 "First use: insn %d, last use: insn %d.\n",
4473 REGNO_FIRST_UID (bl->regno),
4474 REGNO_LAST_UID (bl->regno));
4478 /* Combine all giv's for this iv_class. */
4481 /* This will be true at the end, if all givs which depend on this
4482 biv have been strength reduced.
4483 We can't (currently) eliminate the biv unless this is so. */
4486 /* Check each giv in this class to see if we will benefit by reducing
4487 it. Skip giv's combined with others. */
4488 for (v = bl->giv; v; v = v->next_iv)
4490 struct induction *tv;
4492 if (v->ignore || v->same)
4495 benefit = v->benefit;
4497 /* Reduce benefit if not replaceable, since we will insert
4498 a move-insn to replace the insn that calculates this giv.
4499 Don't do this unless the giv is a user variable, since it
4500 will often be marked non-replaceable because of the duplication
4501 of the exit code outside the loop. In such a case, the copies
4502 we insert are dead and will be deleted. So they don't have
4503 a cost. Similar situations exist. */
4504 /* ??? The new final_[bg]iv_value code does a much better job
4505 of finding replaceable giv's, and hence this code may no longer
4507 if (! v->replaceable && ! bl->eliminable
4508 && REG_USERVAR_P (v->dest_reg))
4509 benefit -= copy_cost;
4511 /* Decrease the benefit to count the add-insns that we will
4512 insert to increment the reduced reg for the giv. */
4513 benefit -= add_cost * bl->biv_count;
4515 /* Decide whether to strength-reduce this giv or to leave the code
4516 unchanged (recompute it from the biv each time it is used).
4517 This decision can be made independently for each giv. */
4520 /* Attempt to guess whether autoincrement will handle some of the
4521 new add insns; if so, increase BENEFIT (undo the subtraction of
4522 add_cost that was done above). */
4523 if (v->giv_type == DEST_ADDR
4524 && GET_CODE (v->mult_val) == CONST_INT)
4526 if (HAVE_POST_INCREMENT
4527 && INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
4528 benefit += add_cost * bl->biv_count;
4529 else if (HAVE_PRE_INCREMENT
4530 && INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
4531 benefit += add_cost * bl->biv_count;
4532 else if (HAVE_POST_DECREMENT
4533 && -INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
4534 benefit += add_cost * bl->biv_count;
4535 else if (HAVE_PRE_DECREMENT
4536 && -INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
4537 benefit += add_cost * bl->biv_count;
4541 /* If an insn is not to be strength reduced, then set its ignore
4542 flag, and clear all_reduced. */
4544 /* A giv that depends on a reversed biv must be reduced if it is
4545 used after the loop exit, otherwise, it would have the wrong
4546 value after the loop exit. To make it simple, just reduce all
4547 of such giv's whether or not we know they are used after the loop
4550 if ( ! flag_reduce_all_givs && v->lifetime * threshold * benefit < insn_count
4553 if (loop_dump_stream)
4554 fprintf (loop_dump_stream,
4555 "giv of insn %d not worth while, %d vs %d.\n",
4557 v->lifetime * threshold * benefit, insn_count);
4563 /* Check that we can increment the reduced giv without a
4564 multiply insn. If not, reject it. */
4566 for (tv = bl->biv; tv; tv = tv->next_iv)
4567 if (tv->mult_val == const1_rtx
4568 && ! product_cheap_p (tv->add_val, v->mult_val))
4570 if (loop_dump_stream)
4571 fprintf (loop_dump_stream,
4572 "giv of insn %d: would need a multiply.\n",
4573 INSN_UID (v->insn));
4581 /* Check for givs whose first use is their definition and whose
4582 last use is the definition of another giv. If so, it is likely
4583 dead and should not be used to derive another giv nor to
4585 for (v = bl->giv; v; v = v->next_iv)
4588 || (v->same && v->same->ignore))
4593 struct induction *v1;
4595 for (v1 = bl->giv; v1; v1 = v1->next_iv)
4596 if (v->last_use == v1->insn)
4599 else if (v->giv_type == DEST_REG
4600 && REGNO_FIRST_UID (REGNO (v->dest_reg)) == INSN_UID (v->insn))
4602 struct induction *v1;
4604 for (v1 = bl->giv; v1; v1 = v1->next_iv)
4605 if (REGNO_LAST_UID (REGNO (v->dest_reg)) == INSN_UID (v1->insn))
4610 /* Now that we know which givs will be reduced, try to rearrange the
4611 combinations to reduce register pressure.
4612 recombine_givs calls find_life_end, which needs reg_iv_type and
4613 reg_iv_info to be valid for all pseudos. We do the necessary
4614 reallocation here since it allows to check if there are still
4615 more bivs to process. */
4616 nregs = max_reg_num ();
4617 if (nregs > reg_iv_type->num_elements)
4619 /* If there are still more bivs to process, allocate some slack
4620 space so that we're not constantly reallocating these arrays. */
4623 /* Reallocate reg_iv_type and reg_iv_info. */
4624 VARRAY_GROW (reg_iv_type, nregs);
4625 VARRAY_GROW (reg_iv_info, nregs);
4627 recombine_givs (loop, bl, flags & LOOP_UNROLL);
4629 /* Reduce each giv that we decided to reduce. */
4631 for (v = bl->giv; v; v = v->next_iv)
4633 struct induction *tv;
4634 if (! v->ignore && v->same == 0)
4636 int auto_inc_opt = 0;
4638 /* If the code for derived givs immediately below has already
4639 allocated a new_reg, we must keep it. */
4641 v->new_reg = gen_reg_rtx (v->mode);
4643 if (v->derived_from)
4645 struct induction *d = v->derived_from;
4647 /* In case d->dest_reg is not replaceable, we have
4648 to replace it in v->insn now. */
4650 d->new_reg = gen_reg_rtx (d->mode);
4652 = replace_rtx (PATTERN (v->insn), d->dest_reg, d->new_reg);
4654 = replace_rtx (PATTERN (v->insn), v->dest_reg, v->new_reg);
4655 /* For each place where the biv is incremented, add an
4656 insn to set the new, reduced reg for the giv.
4657 We used to do this only for biv_count != 1, but
4658 this fails when there is a giv after a single biv
4659 increment, e.g. when the last giv was expressed as
4661 for (tv = bl->biv; tv; tv = tv->next_iv)
4663 /* We always emit reduced giv increments before the
4664 biv increment when bl->biv_count != 1. So by
4665 emitting the add insns for derived givs after the
4666 biv increment, they pick up the updated value of
4668 If the reduced giv is processed with
4669 auto_inc_opt == 1, then it is incremented earlier
4670 than the biv, hence we'll still pick up the right
4672 If it's processed with auto_inc_opt == -1,
4673 that implies that the biv increment is before the
4674 first reduced giv's use. The derived giv's lifetime
4675 is after the reduced giv's lifetime, hence in this
4676 case, the biv increment doesn't matter. */
4677 emit_insn_after (copy_rtx (PATTERN (v->insn)), tv->insn);
4683 /* If the target has auto-increment addressing modes, and
4684 this is an address giv, then try to put the increment
4685 immediately after its use, so that flow can create an
4686 auto-increment addressing mode. */
4687 if (v->giv_type == DEST_ADDR && bl->biv_count == 1
4688 && bl->biv->always_executed && ! bl->biv->maybe_multiple
4689 /* We don't handle reversed biv's because bl->biv->insn
4690 does not have a valid INSN_LUID. */
4692 && v->always_executed && ! v->maybe_multiple
4693 && INSN_UID (v->insn) < max_uid_for_loop)
4695 /* If other giv's have been combined with this one, then
4696 this will work only if all uses of the other giv's occur
4697 before this giv's insn. This is difficult to check.
4699 We simplify this by looking for the common case where
4700 there is one DEST_REG giv, and this giv's insn is the
4701 last use of the dest_reg of that DEST_REG giv. If the
4702 increment occurs after the address giv, then we can
4703 perform the optimization. (Otherwise, the increment
4704 would have to go before other_giv, and we would not be
4705 able to combine it with the address giv to get an
4706 auto-inc address.) */
4707 if (v->combined_with)
4709 struct induction *other_giv = 0;
4711 for (tv = bl->giv; tv; tv = tv->next_iv)
4719 if (! tv && other_giv
4720 && REGNO (other_giv->dest_reg) < max_reg_before_loop
4721 && (REGNO_LAST_UID (REGNO (other_giv->dest_reg))
4722 == INSN_UID (v->insn))
4723 && INSN_LUID (v->insn) < INSN_LUID (bl->biv->insn))
4726 /* Check for case where increment is before the address
4727 giv. Do this test in "loop order". */
4728 else if ((INSN_LUID (v->insn) > INSN_LUID (bl->biv->insn)
4729 && (INSN_LUID (v->insn) < INSN_LUID (loop_scan_start)
4730 || (INSN_LUID (bl->biv->insn)
4731 > INSN_LUID (loop_scan_start))))
4732 || (INSN_LUID (v->insn) < INSN_LUID (loop_scan_start)
4733 && (INSN_LUID (loop_scan_start)
4734 < INSN_LUID (bl->biv->insn))))
4743 /* We can't put an insn immediately after one setting
4744 cc0, or immediately before one using cc0. */
4745 if ((auto_inc_opt == 1 && sets_cc0_p (PATTERN (v->insn)))
4746 || (auto_inc_opt == -1
4747 && (prev = prev_nonnote_insn (v->insn)) != 0
4749 && sets_cc0_p (PATTERN (prev))))
4755 v->auto_inc_opt = 1;
4759 /* For each place where the biv is incremented, add an insn
4760 to increment the new, reduced reg for the giv. */
4761 for (tv = bl->biv; tv; tv = tv->next_iv)
4766 insert_before = tv->insn;
4767 else if (auto_inc_opt == 1)
4768 insert_before = NEXT_INSN (v->insn);
4770 insert_before = v->insn;
4772 if (tv->mult_val == const1_rtx)
4773 emit_iv_add_mult (tv->add_val, v->mult_val,
4774 v->new_reg, v->new_reg, insert_before);
4775 else /* tv->mult_val == const0_rtx */
4776 /* A multiply is acceptable here
4777 since this is presumed to be seldom executed. */
4778 emit_iv_add_mult (tv->add_val, v->mult_val,
4779 v->add_val, v->new_reg, insert_before);
4782 /* Add code at loop start to initialize giv's reduced reg. */
4784 emit_iv_add_mult (bl->initial_value, v->mult_val,
4785 v->add_val, v->new_reg, loop_start);
4789 /* Rescan all givs. If a giv is the same as a giv not reduced, mark it
4792 For each giv register that can be reduced now: if replaceable,
4793 substitute reduced reg wherever the old giv occurs;
4794 else add new move insn "giv_reg = reduced_reg". */
4796 for (v = bl->giv; v; v = v->next_iv)
4798 if (v->same && v->same->ignore)
4804 /* Update expression if this was combined, in case other giv was
4807 v->new_reg = replace_rtx (v->new_reg,
4808 v->same->dest_reg, v->same->new_reg);
4810 if (v->giv_type == DEST_ADDR)
4811 /* Store reduced reg as the address in the memref where we found
4813 validate_change (v->insn, v->location, v->new_reg, 0);
4814 else if (v->replaceable)
4816 reg_map[REGNO (v->dest_reg)] = v->new_reg;
4819 /* I can no longer duplicate the original problem. Perhaps
4820 this is unnecessary now? */
4822 /* Replaceable; it isn't strictly necessary to delete the old
4823 insn and emit a new one, because v->dest_reg is now dead.
4825 However, especially when unrolling loops, the special
4826 handling for (set REG0 REG1) in the second cse pass may
4827 make v->dest_reg live again. To avoid this problem, emit
4828 an insn to set the original giv reg from the reduced giv.
4829 We can not delete the original insn, since it may be part
4830 of a LIBCALL, and the code in flow that eliminates dead
4831 libcalls will fail if it is deleted. */
4832 emit_insn_after (gen_move_insn (v->dest_reg, v->new_reg),
4838 /* Not replaceable; emit an insn to set the original giv reg from
4839 the reduced giv, same as above. */
4840 emit_insn_after (gen_move_insn (v->dest_reg, v->new_reg),
4844 /* When a loop is reversed, givs which depend on the reversed
4845 biv, and which are live outside the loop, must be set to their
4846 correct final value. This insn is only needed if the giv is
4847 not replaceable. The correct final value is the same as the
4848 value that the giv starts the reversed loop with. */
4849 if (bl->reversed && ! v->replaceable)
4850 emit_iv_add_mult (bl->initial_value, v->mult_val,
4851 v->add_val, v->dest_reg, end_insert_before);
4852 else if (v->final_value)
4856 /* If the loop has multiple exits, emit the insn before the
4857 loop to ensure that it will always be executed no matter
4858 how the loop exits. Otherwise, emit the insn after the loop,
4859 since this is slightly more efficient. */
4860 if (loop->exit_count)
4861 insert_before = loop_start;
4863 insert_before = end_insert_before;
4864 emit_insn_before (gen_move_insn (v->dest_reg, v->final_value),
4868 /* If the insn to set the final value of the giv was emitted
4869 before the loop, then we must delete the insn inside the loop
4870 that sets it. If this is a LIBCALL, then we must delete
4871 every insn in the libcall. Note, however, that
4872 final_giv_value will only succeed when there are multiple
4873 exits if the giv is dead at each exit, hence it does not
4874 matter that the original insn remains because it is dead
4876 /* Delete the insn inside the loop that sets the giv since
4877 the giv is now set before (or after) the loop. */
4878 delete_insn (v->insn);
4882 if (loop_dump_stream)
4884 fprintf (loop_dump_stream, "giv at %d reduced to ",
4885 INSN_UID (v->insn));
4886 print_rtl (loop_dump_stream, v->new_reg);
4887 fprintf (loop_dump_stream, "\n");
4891 /* All the givs based on the biv bl have been reduced if they
4894 /* For each giv not marked as maybe dead that has been combined with a
4895 second giv, clear any "maybe dead" mark on that second giv.
4896 v->new_reg will either be or refer to the register of the giv it
4899 Doing this clearing avoids problems in biv elimination where a
4900 giv's new_reg is a complex value that can't be put in the insn but
4901 the giv combined with (with a reg as new_reg) is marked maybe_dead.
4902 Since the register will be used in either case, we'd prefer it be
4903 used from the simpler giv. */
4905 for (v = bl->giv; v; v = v->next_iv)
4906 if (! v->maybe_dead && v->same)
4907 v->same->maybe_dead = 0;
4909 /* Try to eliminate the biv, if it is a candidate.
4910 This won't work if ! all_reduced,
4911 since the givs we planned to use might not have been reduced.
4913 We have to be careful that we didn't initially think we could eliminate
4914 this biv because of a giv that we now think may be dead and shouldn't
4915 be used as a biv replacement.
4917 Also, there is the possibility that we may have a giv that looks
4918 like it can be used to eliminate a biv, but the resulting insn
4919 isn't valid. This can happen, for example, on the 88k, where a
4920 JUMP_INSN can compare a register only with zero. Attempts to
4921 replace it with a compare with a constant will fail.
4923 Note that in cases where this call fails, we may have replaced some
4924 of the occurrences of the biv with a giv, but no harm was done in
4925 doing so in the rare cases where it can occur. */
4927 if (all_reduced == 1 && bl->eliminable
4928 && maybe_eliminate_biv (loop, bl, 1, threshold, insn_count))
4930 /* ?? If we created a new test to bypass the loop entirely,
4931 or otherwise drop straight in, based on this test, then
4932 we might want to rewrite it also. This way some later
4933 pass has more hope of removing the initialization of this
4936 /* If final_value != 0, then the biv may be used after loop end
4937 and we must emit an insn to set it just in case.
4939 Reversed bivs already have an insn after the loop setting their
4940 value, so we don't need another one. We can't calculate the
4941 proper final value for such a biv here anyways. */
4942 if (final_value != 0 && ! bl->reversed)
4946 /* If the loop has multiple exits, emit the insn before the
4947 loop to ensure that it will always be executed no matter
4948 how the loop exits. Otherwise, emit the insn after the
4949 loop, since this is slightly more efficient. */
4950 if (loop->exit_count)
4951 insert_before = loop_start;
4953 insert_before = end_insert_before;
4955 emit_insn_before (gen_move_insn (bl->biv->dest_reg, final_value),
4960 /* Delete all of the instructions inside the loop which set
4961 the biv, as they are all dead. If is safe to delete them,
4962 because an insn setting a biv will never be part of a libcall. */
4963 /* However, deleting them will invalidate the regno_last_uid info,
4964 so keeping them around is more convenient. Final_biv_value
4965 will only succeed when there are multiple exits if the biv
4966 is dead at each exit, hence it does not matter that the original
4967 insn remains, because it is dead anyways. */
4968 for (v = bl->biv; v; v = v->next_iv)
4969 delete_insn (v->insn);
4972 if (loop_dump_stream)
4973 fprintf (loop_dump_stream, "Reg %d: biv eliminated\n",
4978 /* Go through all the instructions in the loop, making all the
4979 register substitutions scheduled in REG_MAP. */
4981 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
4982 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
4983 || GET_CODE (p) == CALL_INSN)
4985 replace_regs (PATTERN (p), reg_map, reg_map_size, 0);
4986 replace_regs (REG_NOTES (p), reg_map, reg_map_size, 0);
4990 if (loop_info->n_iterations > 0)
4992 /* When we completely unroll a loop we will likely not need the increment
4993 of the loop BIV and we will not need the conditional branch at the
4995 unrolled_insn_copies = insn_count - 2;
4998 /* When we completely unroll a loop on a HAVE_cc0 machine we will not
4999 need the comparison before the conditional branch at the end of the
5001 unrolled_insn_copies -= 1;
5004 /* We'll need one copy for each loop iteration. */
5005 unrolled_insn_copies *= loop_info->n_iterations;
5007 /* A little slop to account for the ability to remove initialization
5008 code, better CSE, and other secondary benefits of completely
5009 unrolling some loops. */
5010 unrolled_insn_copies -= 1;
5012 /* Clamp the value. */
5013 if (unrolled_insn_copies < 0)
5014 unrolled_insn_copies = 0;
5017 /* Unroll loops from within strength reduction so that we can use the
5018 induction variable information that strength_reduce has already
5019 collected. Always unroll loops that would be as small or smaller
5020 unrolled than when rolled. */
5021 if ((flags & LOOP_UNROLL)
5022 || (loop_info->n_iterations > 0
5023 && unrolled_insn_copies <= insn_count))
5024 unroll_loop (loop, insn_count, end_insert_before, 1);
5026 #ifdef HAVE_doloop_end
5027 if (HAVE_doloop_end && (flags & LOOP_BCT) && flag_branch_on_count_reg)
5028 doloop_optimize (loop);
5029 #endif /* HAVE_doloop_end */
5031 if (loop_dump_stream)
5032 fprintf (loop_dump_stream, "\n");
5035 VARRAY_FREE (reg_iv_type);
5036 VARRAY_FREE (reg_iv_info);
5037 free (reg_biv_class);
5042 /*Record all basic induction variables calculated in the insn. */
5044 check_insn_for_bivs (loop, p, not_every_iteration, maybe_multiple)
5047 int not_every_iteration;
5056 if (GET_CODE (p) == INSN
5057 && (set = single_set (p))
5058 && GET_CODE (SET_DEST (set)) == REG)
5060 dest_reg = SET_DEST (set);
5061 if (REGNO (dest_reg) < max_reg_before_loop
5062 && REGNO (dest_reg) >= FIRST_PSEUDO_REGISTER
5063 && REG_IV_TYPE (REGNO (dest_reg)) != NOT_BASIC_INDUCT)
5065 if (basic_induction_var (loop, SET_SRC (set),
5066 GET_MODE (SET_SRC (set)),
5067 dest_reg, p, &inc_val, &mult_val,
5070 /* It is a possible basic induction variable.
5071 Create and initialize an induction structure for it. */
5074 = (struct induction *) oballoc (sizeof (struct induction));
5076 record_biv (v, p, dest_reg, inc_val, mult_val, location,
5077 not_every_iteration, maybe_multiple);
5078 REG_IV_TYPE (REGNO (dest_reg)) = BASIC_INDUCT;
5080 else if (REGNO (dest_reg) < max_reg_before_loop)
5081 REG_IV_TYPE (REGNO (dest_reg)) = NOT_BASIC_INDUCT;
5087 /* Record all givs calculated in the insn.
5088 A register is a giv if: it is only set once, it is a function of a
5089 biv and a constant (or invariant), and it is not a biv. */
5091 check_insn_for_givs (loop, p, not_every_iteration, maybe_multiple)
5094 int not_every_iteration;
5098 /* Look for a general induction variable in a register. */
5099 if (GET_CODE (p) == INSN
5100 && (set = single_set (p))
5101 && GET_CODE (SET_DEST (set)) == REG
5102 && ! VARRAY_CHAR (may_not_optimize, REGNO (SET_DEST (set))))
5110 rtx last_consec_insn;
5112 dest_reg = SET_DEST (set);
5113 if (REGNO (dest_reg) < FIRST_PSEUDO_REGISTER)
5116 if (/* SET_SRC is a giv. */
5117 (general_induction_var (loop, SET_SRC (set), &src_reg, &add_val,
5118 &mult_val, 0, &benefit, VOIDmode)
5119 /* Equivalent expression is a giv. */
5120 || ((regnote = find_reg_note (p, REG_EQUAL, NULL_RTX))
5121 && general_induction_var (loop, XEXP (regnote, 0), &src_reg,
5122 &add_val, &mult_val, 0,
5123 &benefit, VOIDmode)))
5124 /* Don't try to handle any regs made by loop optimization.
5125 We have nothing on them in regno_first_uid, etc. */
5126 && REGNO (dest_reg) < max_reg_before_loop
5127 /* Don't recognize a BASIC_INDUCT_VAR here. */
5128 && dest_reg != src_reg
5129 /* This must be the only place where the register is set. */
5130 && (VARRAY_INT (n_times_set, REGNO (dest_reg)) == 1
5131 /* or all sets must be consecutive and make a giv. */
5132 || (benefit = consec_sets_giv (loop, benefit, p,
5134 &add_val, &mult_val,
5135 &last_consec_insn))))
5138 = (struct induction *) oballoc (sizeof (struct induction));
5140 /* If this is a library call, increase benefit. */
5141 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
5142 benefit += libcall_benefit (p);
5144 /* Skip the consecutive insns, if there are any. */
5145 if (VARRAY_INT (n_times_set, REGNO (dest_reg)) != 1)
5146 p = last_consec_insn;
5148 record_giv (loop, v, p, src_reg, dest_reg, mult_val, add_val,
5149 benefit, DEST_REG, not_every_iteration,
5150 maybe_multiple, NULL_PTR);
5155 #ifndef DONT_REDUCE_ADDR
5156 /* Look for givs which are memory addresses. */
5157 /* This resulted in worse code on a VAX 8600. I wonder if it
5159 if (GET_CODE (p) == INSN)
5160 find_mem_givs (loop, PATTERN (p), p, not_every_iteration,
5164 /* Update the status of whether giv can derive other givs. This can
5165 change when we pass a label or an insn that updates a biv. */
5166 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
5167 || GET_CODE (p) == CODE_LABEL)
5168 update_giv_derive (loop, p);
5172 /* Return 1 if X is a valid source for an initial value (or as value being
5173 compared against in an initial test).
5175 X must be either a register or constant and must not be clobbered between
5176 the current insn and the start of the loop.
5178 INSN is the insn containing X. */
5181 valid_initial_value_p (x, insn, call_seen, loop_start)
5190 /* Only consider pseudos we know about initialized in insns whose luids
5192 if (GET_CODE (x) != REG
5193 || REGNO (x) >= max_reg_before_loop)
5196 /* Don't use call-clobbered registers across a call which clobbers it. On
5197 some machines, don't use any hard registers at all. */
5198 if (REGNO (x) < FIRST_PSEUDO_REGISTER
5199 && (SMALL_REGISTER_CLASSES
5200 || (call_used_regs[REGNO (x)] && call_seen)))
5203 /* Don't use registers that have been clobbered before the start of the
5205 if (reg_set_between_p (x, insn, loop_start))
5211 /* Scan X for memory refs and check each memory address
5212 as a possible giv. INSN is the insn whose pattern X comes from.
5213 NOT_EVERY_ITERATION is 1 if the insn might not be executed during
5214 every loop iteration. MAYBE_MULTIPLE is 1 if the insn might be executed
5215 more thanonce in each loop iteration. */
5218 find_mem_givs (loop, x, insn, not_every_iteration, maybe_multiple)
5219 const struct loop *loop;
5222 int not_every_iteration, maybe_multiple;
5225 register enum rtx_code code;
5226 register const char *fmt;
5231 code = GET_CODE (x);
5255 /* This code used to disable creating GIVs with mult_val == 1 and
5256 add_val == 0. However, this leads to lost optimizations when
5257 it comes time to combine a set of related DEST_ADDR GIVs, since
5258 this one would not be seen. */
5260 if (general_induction_var (loop, XEXP (x, 0), &src_reg, &add_val,
5261 &mult_val, 1, &benefit, GET_MODE (x)))
5263 /* Found one; record it. */
5265 = (struct induction *) oballoc (sizeof (struct induction));
5267 record_giv (loop, v, insn, src_reg, addr_placeholder, mult_val,
5268 add_val, benefit, DEST_ADDR, not_every_iteration,
5269 maybe_multiple, &XEXP (x, 0));
5271 v->mem_mode = GET_MODE (x);
5280 /* Recursively scan the subexpressions for other mem refs. */
5282 fmt = GET_RTX_FORMAT (code);
5283 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
5285 find_mem_givs (loop, XEXP (x, i), insn, not_every_iteration,
5287 else if (fmt[i] == 'E')
5288 for (j = 0; j < XVECLEN (x, i); j++)
5289 find_mem_givs (loop, XVECEXP (x, i, j), insn, not_every_iteration,
5293 /* Fill in the data about one biv update.
5294 V is the `struct induction' in which we record the biv. (It is
5295 allocated by the caller, with alloca.)
5296 INSN is the insn that sets it.
5297 DEST_REG is the biv's reg.
5299 MULT_VAL is const1_rtx if the biv is being incremented here, in which case
5300 INC_VAL is the increment. Otherwise, MULT_VAL is const0_rtx and the biv is
5301 being set to INC_VAL.
5303 NOT_EVERY_ITERATION is nonzero if this biv update is not know to be
5304 executed every iteration; MAYBE_MULTIPLE is nonzero if this biv update
5305 can be executed more than once per iteration. If MAYBE_MULTIPLE
5306 and NOT_EVERY_ITERATION are both zero, we know that the biv update is
5307 executed exactly once per iteration. */
5310 record_biv (v, insn, dest_reg, inc_val, mult_val, location,
5311 not_every_iteration, maybe_multiple)
5312 struct induction *v;
5318 int not_every_iteration;
5321 struct iv_class *bl;
5324 v->src_reg = dest_reg;
5325 v->dest_reg = dest_reg;
5326 v->mult_val = mult_val;
5327 v->add_val = inc_val;
5328 v->location = location;
5329 v->mode = GET_MODE (dest_reg);
5330 v->always_computable = ! not_every_iteration;
5331 v->always_executed = ! not_every_iteration;
5332 v->maybe_multiple = maybe_multiple;
5334 /* Add this to the reg's iv_class, creating a class
5335 if this is the first incrementation of the reg. */
5337 bl = reg_biv_class[REGNO (dest_reg)];
5340 /* Create and initialize new iv_class. */
5342 bl = (struct iv_class *) oballoc (sizeof (struct iv_class));
5344 bl->regno = REGNO (dest_reg);
5350 /* Set initial value to the reg itself. */
5351 bl->initial_value = dest_reg;
5352 /* We haven't seen the initializing insn yet */
5355 bl->initial_test = 0;
5356 bl->incremented = 0;
5360 bl->total_benefit = 0;
5362 /* Add this class to loop_iv_list. */
5363 bl->next = loop_iv_list;
5366 /* Put it in the array of biv register classes. */
5367 reg_biv_class[REGNO (dest_reg)] = bl;
5370 /* Update IV_CLASS entry for this biv. */
5371 v->next_iv = bl->biv;
5374 if (mult_val == const1_rtx)
5375 bl->incremented = 1;
5377 if (loop_dump_stream)
5379 fprintf (loop_dump_stream,
5380 "Insn %d: possible biv, reg %d,",
5381 INSN_UID (insn), REGNO (dest_reg));
5382 if (GET_CODE (inc_val) == CONST_INT)
5384 fprintf (loop_dump_stream, " const =");
5385 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (inc_val));
5386 fputc ('\n', loop_dump_stream);
5390 fprintf (loop_dump_stream, " const = ");
5391 print_rtl (loop_dump_stream, inc_val);
5392 fprintf (loop_dump_stream, "\n");
5397 /* Fill in the data about one giv.
5398 V is the `struct induction' in which we record the giv. (It is
5399 allocated by the caller, with alloca.)
5400 INSN is the insn that sets it.
5401 BENEFIT estimates the savings from deleting this insn.
5402 TYPE is DEST_REG or DEST_ADDR; it says whether the giv is computed
5403 into a register or is used as a memory address.
5405 SRC_REG is the biv reg which the giv is computed from.
5406 DEST_REG is the giv's reg (if the giv is stored in a reg).
5407 MULT_VAL and ADD_VAL are the coefficients used to compute the giv.
5408 LOCATION points to the place where this giv's value appears in INSN. */
5411 record_giv (loop, v, insn, src_reg, dest_reg, mult_val, add_val, benefit,
5412 type, not_every_iteration, maybe_multiple, location)
5413 const struct loop *loop;
5414 struct induction *v;
5418 rtx mult_val, add_val;
5421 int not_every_iteration, maybe_multiple;
5424 struct induction *b;
5425 struct iv_class *bl;
5426 rtx set = single_set (insn);
5429 /* Attempt to prove constantness of the values. */
5430 temp = simplify_rtx (add_val);
5435 v->src_reg = src_reg;
5437 v->dest_reg = dest_reg;
5438 v->mult_val = mult_val;
5439 v->add_val = add_val;
5440 v->benefit = benefit;
5441 v->location = location;
5443 v->combined_with = 0;
5444 v->maybe_multiple = maybe_multiple;
5446 v->derive_adjustment = 0;
5452 v->auto_inc_opt = 0;
5455 v->derived_from = 0;
5458 /* The v->always_computable field is used in update_giv_derive, to
5459 determine whether a giv can be used to derive another giv. For a
5460 DEST_REG giv, INSN computes a new value for the giv, so its value
5461 isn't computable if INSN insn't executed every iteration.
5462 However, for a DEST_ADDR giv, INSN merely uses the value of the giv;
5463 it does not compute a new value. Hence the value is always computable
5464 regardless of whether INSN is executed each iteration. */
5466 if (type == DEST_ADDR)
5467 v->always_computable = 1;
5469 v->always_computable = ! not_every_iteration;
5471 v->always_executed = ! not_every_iteration;
5473 if (type == DEST_ADDR)
5475 v->mode = GET_MODE (*location);
5478 else /* type == DEST_REG */
5480 v->mode = GET_MODE (SET_DEST (set));
5482 v->lifetime = (uid_luid[REGNO_LAST_UID (REGNO (dest_reg))]
5483 - uid_luid[REGNO_FIRST_UID (REGNO (dest_reg))]);
5485 /* If the lifetime is zero, it means that this register is
5486 really a dead store. So mark this as a giv that can be
5487 ignored. This will not prevent the biv from being eliminated. */
5488 if (v->lifetime == 0)
5491 REG_IV_TYPE (REGNO (dest_reg)) = GENERAL_INDUCT;
5492 REG_IV_INFO (REGNO (dest_reg)) = v;
5495 /* Add the giv to the class of givs computed from one biv. */
5497 bl = reg_biv_class[REGNO (src_reg)];
5500 v->next_iv = bl->giv;
5502 /* Don't count DEST_ADDR. This is supposed to count the number of
5503 insns that calculate givs. */
5504 if (type == DEST_REG)
5506 bl->total_benefit += benefit;
5509 /* Fatal error, biv missing for this giv? */
5512 if (type == DEST_ADDR)
5516 /* The giv can be replaced outright by the reduced register only if all
5517 of the following conditions are true:
5518 - the insn that sets the giv is always executed on any iteration
5519 on which the giv is used at all
5520 (there are two ways to deduce this:
5521 either the insn is executed on every iteration,
5522 or all uses follow that insn in the same basic block),
5523 - the giv is not used outside the loop
5524 - no assignments to the biv occur during the giv's lifetime. */
5526 if (REGNO_FIRST_UID (REGNO (dest_reg)) == INSN_UID (insn)
5527 /* Previous line always fails if INSN was moved by loop opt. */
5528 && uid_luid[REGNO_LAST_UID (REGNO (dest_reg))]
5529 < INSN_LUID (loop->end)
5530 && (! not_every_iteration
5531 || last_use_this_basic_block (dest_reg, insn)))
5533 /* Now check that there are no assignments to the biv within the
5534 giv's lifetime. This requires two separate checks. */
5536 /* Check each biv update, and fail if any are between the first
5537 and last use of the giv.
5539 If this loop contains an inner loop that was unrolled, then
5540 the insn modifying the biv may have been emitted by the loop
5541 unrolling code, and hence does not have a valid luid. Just
5542 mark the biv as not replaceable in this case. It is not very
5543 useful as a biv, because it is used in two different loops.
5544 It is very unlikely that we would be able to optimize the giv
5545 using this biv anyways. */
5548 for (b = bl->biv; b; b = b->next_iv)
5550 if (INSN_UID (b->insn) >= max_uid_for_loop
5551 || ((uid_luid[INSN_UID (b->insn)]
5552 >= uid_luid[REGNO_FIRST_UID (REGNO (dest_reg))])
5553 && (uid_luid[INSN_UID (b->insn)]
5554 <= uid_luid[REGNO_LAST_UID (REGNO (dest_reg))])))
5557 v->not_replaceable = 1;
5562 /* If there are any backwards branches that go from after the
5563 biv update to before it, then this giv is not replaceable. */
5565 for (b = bl->biv; b; b = b->next_iv)
5566 if (back_branch_in_range_p (loop, b->insn))
5569 v->not_replaceable = 1;
5575 /* May still be replaceable, we don't have enough info here to
5578 v->not_replaceable = 0;
5582 /* Record whether the add_val contains a const_int, for later use by
5587 v->no_const_addval = 1;
5588 if (tem == const0_rtx)
5590 else if (CONSTANT_P (add_val))
5591 v->no_const_addval = 0;
5592 if (GET_CODE (tem) == PLUS)
5596 if (GET_CODE (XEXP (tem, 0)) == PLUS)
5597 tem = XEXP (tem, 0);
5598 else if (GET_CODE (XEXP (tem, 1)) == PLUS)
5599 tem = XEXP (tem, 1);
5603 if (CONSTANT_P (XEXP (tem, 1)))
5604 v->no_const_addval = 0;
5608 if (loop_dump_stream)
5610 if (type == DEST_REG)
5611 fprintf (loop_dump_stream, "Insn %d: giv reg %d",
5612 INSN_UID (insn), REGNO (dest_reg));
5614 fprintf (loop_dump_stream, "Insn %d: dest address",
5617 fprintf (loop_dump_stream, " src reg %d benefit %d",
5618 REGNO (src_reg), v->benefit);
5619 fprintf (loop_dump_stream, " lifetime %d",
5623 fprintf (loop_dump_stream, " replaceable");
5625 if (v->no_const_addval)
5626 fprintf (loop_dump_stream, " ncav");
5628 if (GET_CODE (mult_val) == CONST_INT)
5630 fprintf (loop_dump_stream, " mult ");
5631 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (mult_val));
5635 fprintf (loop_dump_stream, " mult ");
5636 print_rtl (loop_dump_stream, mult_val);
5639 if (GET_CODE (add_val) == CONST_INT)
5641 fprintf (loop_dump_stream, " add ");
5642 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (add_val));
5646 fprintf (loop_dump_stream, " add ");
5647 print_rtl (loop_dump_stream, add_val);
5651 if (loop_dump_stream)
5652 fprintf (loop_dump_stream, "\n");
5657 /* All this does is determine whether a giv can be made replaceable because
5658 its final value can be calculated. This code can not be part of record_giv
5659 above, because final_giv_value requires that the number of loop iterations
5660 be known, and that can not be accurately calculated until after all givs
5661 have been identified. */
5664 check_final_value (loop, v)
5665 const struct loop *loop;
5666 struct induction *v;
5668 struct iv_class *bl;
5669 rtx final_value = 0;
5671 bl = reg_biv_class[REGNO (v->src_reg)];
5673 /* DEST_ADDR givs will never reach here, because they are always marked
5674 replaceable above in record_giv. */
5676 /* The giv can be replaced outright by the reduced register only if all
5677 of the following conditions are true:
5678 - the insn that sets the giv is always executed on any iteration
5679 on which the giv is used at all
5680 (there are two ways to deduce this:
5681 either the insn is executed on every iteration,
5682 or all uses follow that insn in the same basic block),
5683 - its final value can be calculated (this condition is different
5684 than the one above in record_giv)
5685 - no assignments to the biv occur during the giv's lifetime. */
5688 /* This is only called now when replaceable is known to be false. */
5689 /* Clear replaceable, so that it won't confuse final_giv_value. */
5693 if ((final_value = final_giv_value (loop, v))
5694 && (v->always_computable || last_use_this_basic_block (v->dest_reg, v->insn)))
5696 int biv_increment_seen = 0;
5702 /* When trying to determine whether or not a biv increment occurs
5703 during the lifetime of the giv, we can ignore uses of the variable
5704 outside the loop because final_value is true. Hence we can not
5705 use regno_last_uid and regno_first_uid as above in record_giv. */
5707 /* Search the loop to determine whether any assignments to the
5708 biv occur during the giv's lifetime. Start with the insn
5709 that sets the giv, and search around the loop until we come
5710 back to that insn again.
5712 Also fail if there is a jump within the giv's lifetime that jumps
5713 to somewhere outside the lifetime but still within the loop. This
5714 catches spaghetti code where the execution order is not linear, and
5715 hence the above test fails. Here we assume that the giv lifetime
5716 does not extend from one iteration of the loop to the next, so as
5717 to make the test easier. Since the lifetime isn't known yet,
5718 this requires two loops. See also record_giv above. */
5720 last_giv_use = v->insn;
5726 p = NEXT_INSN (loop->start);
5730 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
5731 || GET_CODE (p) == CALL_INSN)
5733 if (biv_increment_seen)
5735 if (reg_mentioned_p (v->dest_reg, PATTERN (p)))
5738 v->not_replaceable = 1;
5742 else if (reg_set_p (v->src_reg, PATTERN (p)))
5743 biv_increment_seen = 1;
5744 else if (reg_mentioned_p (v->dest_reg, PATTERN (p)))
5749 /* Now that the lifetime of the giv is known, check for branches
5750 from within the lifetime to outside the lifetime if it is still
5760 p = NEXT_INSN (loop->start);
5761 if (p == last_giv_use)
5764 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p)
5765 && LABEL_NAME (JUMP_LABEL (p))
5766 && ((loop_insn_first_p (JUMP_LABEL (p), v->insn)
5767 && loop_insn_first_p (loop->start, JUMP_LABEL (p)))
5768 || (loop_insn_first_p (last_giv_use, JUMP_LABEL (p))
5769 && loop_insn_first_p (JUMP_LABEL (p), loop->end))))
5772 v->not_replaceable = 1;
5774 if (loop_dump_stream)
5775 fprintf (loop_dump_stream,
5776 "Found branch outside giv lifetime.\n");
5783 /* If it is replaceable, then save the final value. */
5785 v->final_value = final_value;
5788 if (loop_dump_stream && v->replaceable)
5789 fprintf (loop_dump_stream, "Insn %d: giv reg %d final_value replaceable\n",
5790 INSN_UID (v->insn), REGNO (v->dest_reg));
5793 /* Update the status of whether a giv can derive other givs.
5795 We need to do something special if there is or may be an update to the biv
5796 between the time the giv is defined and the time it is used to derive
5799 In addition, a giv that is only conditionally set is not allowed to
5800 derive another giv once a label has been passed.
5802 The cases we look at are when a label or an update to a biv is passed. */
5805 update_giv_derive (loop, p)
5806 const struct loop *loop;
5809 struct iv_class *bl;
5810 struct induction *biv, *giv;
5814 /* Search all IV classes, then all bivs, and finally all givs.
5816 There are three cases we are concerned with. First we have the situation
5817 of a giv that is only updated conditionally. In that case, it may not
5818 derive any givs after a label is passed.
5820 The second case is when a biv update occurs, or may occur, after the
5821 definition of a giv. For certain biv updates (see below) that are
5822 known to occur between the giv definition and use, we can adjust the
5823 giv definition. For others, or when the biv update is conditional,
5824 we must prevent the giv from deriving any other givs. There are two
5825 sub-cases within this case.
5827 If this is a label, we are concerned with any biv update that is done
5828 conditionally, since it may be done after the giv is defined followed by
5829 a branch here (actually, we need to pass both a jump and a label, but
5830 this extra tracking doesn't seem worth it).
5832 If this is a jump, we are concerned about any biv update that may be
5833 executed multiple times. We are actually only concerned about
5834 backward jumps, but it is probably not worth performing the test
5835 on the jump again here.
5837 If this is a biv update, we must adjust the giv status to show that a
5838 subsequent biv update was performed. If this adjustment cannot be done,
5839 the giv cannot derive further givs. */
5841 for (bl = loop_iv_list; bl; bl = bl->next)
5842 for (biv = bl->biv; biv; biv = biv->next_iv)
5843 if (GET_CODE (p) == CODE_LABEL || GET_CODE (p) == JUMP_INSN
5846 for (giv = bl->giv; giv; giv = giv->next_iv)
5848 /* If cant_derive is already true, there is no point in
5849 checking all of these conditions again. */
5850 if (giv->cant_derive)
5853 /* If this giv is conditionally set and we have passed a label,
5854 it cannot derive anything. */
5855 if (GET_CODE (p) == CODE_LABEL && ! giv->always_computable)
5856 giv->cant_derive = 1;
5858 /* Skip givs that have mult_val == 0, since
5859 they are really invariants. Also skip those that are
5860 replaceable, since we know their lifetime doesn't contain
5862 else if (giv->mult_val == const0_rtx || giv->replaceable)
5865 /* The only way we can allow this giv to derive another
5866 is if this is a biv increment and we can form the product
5867 of biv->add_val and giv->mult_val. In this case, we will
5868 be able to compute a compensation. */
5869 else if (biv->insn == p)
5873 if (biv->mult_val == const1_rtx)
5874 tem = simplify_giv_expr (loop,
5875 gen_rtx_MULT (giv->mode,
5880 if (tem && giv->derive_adjustment)
5881 tem = simplify_giv_expr
5883 gen_rtx_PLUS (giv->mode, tem, giv->derive_adjustment),
5887 giv->derive_adjustment = tem;
5889 giv->cant_derive = 1;
5891 else if ((GET_CODE (p) == CODE_LABEL && ! biv->always_computable)
5892 || (GET_CODE (p) == JUMP_INSN && biv->maybe_multiple))
5893 giv->cant_derive = 1;
5898 /* Check whether an insn is an increment legitimate for a basic induction var.
5899 X is the source of insn P, or a part of it.
5900 MODE is the mode in which X should be interpreted.
5902 DEST_REG is the putative biv, also the destination of the insn.
5903 We accept patterns of these forms:
5904 REG = REG + INVARIANT (includes REG = REG - CONSTANT)
5905 REG = INVARIANT + REG
5907 If X is suitable, we return 1, set *MULT_VAL to CONST1_RTX,
5908 store the additive term into *INC_VAL, and store the place where
5909 we found the additive term into *LOCATION.
5911 If X is an assignment of an invariant into DEST_REG, we set
5912 *MULT_VAL to CONST0_RTX, and store the invariant into *INC_VAL.
5914 We also want to detect a BIV when it corresponds to a variable
5915 whose mode was promoted via PROMOTED_MODE. In that case, an increment
5916 of the variable may be a PLUS that adds a SUBREG of that variable to
5917 an invariant and then sign- or zero-extends the result of the PLUS
5920 Most GIVs in such cases will be in the promoted mode, since that is the
5921 probably the natural computation mode (and almost certainly the mode
5922 used for addresses) on the machine. So we view the pseudo-reg containing
5923 the variable as the BIV, as if it were simply incremented.
5925 Note that treating the entire pseudo as a BIV will result in making
5926 simple increments to any GIVs based on it. However, if the variable
5927 overflows in its declared mode but not its promoted mode, the result will
5928 be incorrect. This is acceptable if the variable is signed, since
5929 overflows in such cases are undefined, but not if it is unsigned, since
5930 those overflows are defined. So we only check for SIGN_EXTEND and
5933 If we cannot find a biv, we return 0. */
5936 basic_induction_var (loop, x, mode, dest_reg, p, inc_val, mult_val, location)
5937 const struct loop *loop;
5939 enum machine_mode mode;
5946 register enum rtx_code code;
5950 code = GET_CODE (x);
5955 if (rtx_equal_p (XEXP (x, 0), dest_reg)
5956 || (GET_CODE (XEXP (x, 0)) == SUBREG
5957 && SUBREG_PROMOTED_VAR_P (XEXP (x, 0))
5958 && SUBREG_REG (XEXP (x, 0)) == dest_reg))
5960 argp = &XEXP (x, 1);
5962 else if (rtx_equal_p (XEXP (x, 1), dest_reg)
5963 || (GET_CODE (XEXP (x, 1)) == SUBREG
5964 && SUBREG_PROMOTED_VAR_P (XEXP (x, 1))
5965 && SUBREG_REG (XEXP (x, 1)) == dest_reg))
5967 argp = &XEXP (x, 0);
5973 if (loop_invariant_p (loop, arg) != 1)
5976 *inc_val = convert_modes (GET_MODE (dest_reg), GET_MODE (x), arg, 0);
5977 *mult_val = const1_rtx;
5982 /* If this is a SUBREG for a promoted variable, check the inner
5984 if (SUBREG_PROMOTED_VAR_P (x))
5985 return basic_induction_var (loop, SUBREG_REG (x),
5986 GET_MODE (SUBREG_REG (x)),
5987 dest_reg, p, inc_val, mult_val, location);
5991 /* If this register is assigned in a previous insn, look at its
5992 source, but don't go outside the loop or past a label. */
5994 /* If this sets a register to itself, we would repeat any previous
5995 biv increment if we applied this strategy blindly. */
5996 if (rtx_equal_p (dest_reg, x))
6003 insn = PREV_INSN (insn);
6004 } while (insn && GET_CODE (insn) == NOTE
6005 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_BEG);
6009 set = single_set (insn);
6013 if ((SET_DEST (set) == x
6014 || (GET_CODE (SET_DEST (set)) == SUBREG
6015 && (GET_MODE_SIZE (GET_MODE (SET_DEST (set)))
6017 && (GET_MODE_CLASS (GET_MODE (SET_DEST (set)))
6019 && SUBREG_REG (SET_DEST (set)) == x))
6020 && basic_induction_var (loop, SET_SRC (set),
6021 (GET_MODE (SET_SRC (set)) == VOIDmode
6023 : GET_MODE (SET_SRC (set))),
6025 inc_val, mult_val, location))
6028 /* ... fall through ... */
6030 /* Can accept constant setting of biv only when inside inner most loop.
6031 Otherwise, a biv of an inner loop may be incorrectly recognized
6032 as a biv of the outer loop,
6033 causing code to be moved INTO the inner loop. */
6035 if (loop_invariant_p (loop, x) != 1)
6040 /* convert_modes aborts if we try to convert to or from CCmode, so just
6041 exclude that case. It is very unlikely that a condition code value
6042 would be a useful iterator anyways. */
6043 if (loop->level == 1
6044 && GET_MODE_CLASS (mode) != MODE_CC
6045 && GET_MODE_CLASS (GET_MODE (dest_reg)) != MODE_CC)
6047 /* Possible bug here? Perhaps we don't know the mode of X. */
6048 *inc_val = convert_modes (GET_MODE (dest_reg), mode, x, 0);
6049 *mult_val = const0_rtx;
6056 return basic_induction_var (loop, XEXP (x, 0), GET_MODE (XEXP (x, 0)),
6057 dest_reg, p, inc_val, mult_val, location);
6060 /* Similar, since this can be a sign extension. */
6061 for (insn = PREV_INSN (p);
6062 (insn && GET_CODE (insn) == NOTE
6063 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_BEG);
6064 insn = PREV_INSN (insn))
6068 set = single_set (insn);
6070 if (! rtx_equal_p (dest_reg, XEXP (x, 0))
6071 && set && SET_DEST (set) == XEXP (x, 0)
6072 && GET_CODE (XEXP (x, 1)) == CONST_INT
6073 && INTVAL (XEXP (x, 1)) >= 0
6074 && GET_CODE (SET_SRC (set)) == ASHIFT
6075 && XEXP (x, 1) == XEXP (SET_SRC (set), 1))
6076 return basic_induction_var (loop, XEXP (SET_SRC (set), 0),
6077 GET_MODE (XEXP (x, 0)),
6078 dest_reg, insn, inc_val, mult_val,
6087 /* A general induction variable (giv) is any quantity that is a linear
6088 function of a basic induction variable,
6089 i.e. giv = biv * mult_val + add_val.
6090 The coefficients can be any loop invariant quantity.
6091 A giv need not be computed directly from the biv;
6092 it can be computed by way of other givs. */
6094 /* Determine whether X computes a giv.
6095 If it does, return a nonzero value
6096 which is the benefit from eliminating the computation of X;
6097 set *SRC_REG to the register of the biv that it is computed from;
6098 set *ADD_VAL and *MULT_VAL to the coefficients,
6099 such that the value of X is biv * mult + add; */
6102 general_induction_var (loop, x, src_reg, add_val, mult_val, is_addr,
6103 pbenefit, addr_mode)
6104 const struct loop *loop;
6111 enum machine_mode addr_mode;
6116 /* If this is an invariant, forget it, it isn't a giv. */
6117 if (loop_invariant_p (loop, x) == 1)
6120 /* See if the expression could be a giv and get its form.
6121 Mark our place on the obstack in case we don't find a giv. */
6122 storage = (char *) oballoc (0);
6124 x = simplify_giv_expr (loop, x, pbenefit);
6131 switch (GET_CODE (x))
6135 /* Since this is now an invariant and wasn't before, it must be a giv
6136 with MULT_VAL == 0. It doesn't matter which BIV we associate this
6138 *src_reg = loop_iv_list->biv->dest_reg;
6139 *mult_val = const0_rtx;
6144 /* This is equivalent to a BIV. */
6146 *mult_val = const1_rtx;
6147 *add_val = const0_rtx;
6151 /* Either (plus (biv) (invar)) or
6152 (plus (mult (biv) (invar_1)) (invar_2)). */
6153 if (GET_CODE (XEXP (x, 0)) == MULT)
6155 *src_reg = XEXP (XEXP (x, 0), 0);
6156 *mult_val = XEXP (XEXP (x, 0), 1);
6160 *src_reg = XEXP (x, 0);
6161 *mult_val = const1_rtx;
6163 *add_val = XEXP (x, 1);
6167 /* ADD_VAL is zero. */
6168 *src_reg = XEXP (x, 0);
6169 *mult_val = XEXP (x, 1);
6170 *add_val = const0_rtx;
6177 /* Remove any enclosing USE from ADD_VAL and MULT_VAL (there will be
6178 unless they are CONST_INT). */
6179 if (GET_CODE (*add_val) == USE)
6180 *add_val = XEXP (*add_val, 0);
6181 if (GET_CODE (*mult_val) == USE)
6182 *mult_val = XEXP (*mult_val, 0);
6185 *pbenefit += address_cost (orig_x, addr_mode) - reg_address_cost;
6187 *pbenefit += rtx_cost (orig_x, SET);
6189 /* Always return true if this is a giv so it will be detected as such,
6190 even if the benefit is zero or negative. This allows elimination
6191 of bivs that might otherwise not be eliminated. */
6195 /* Given an expression, X, try to form it as a linear function of a biv.
6196 We will canonicalize it to be of the form
6197 (plus (mult (BIV) (invar_1))
6199 with possible degeneracies.
6201 The invariant expressions must each be of a form that can be used as a
6202 machine operand. We surround then with a USE rtx (a hack, but localized
6203 and certainly unambiguous!) if not a CONST_INT for simplicity in this
6204 routine; it is the caller's responsibility to strip them.
6206 If no such canonicalization is possible (i.e., two biv's are used or an
6207 expression that is neither invariant nor a biv or giv), this routine
6210 For a non-zero return, the result will have a code of CONST_INT, USE,
6211 REG (for a BIV), PLUS, or MULT. No other codes will occur.
6213 *BENEFIT will be incremented by the benefit of any sub-giv encountered. */
6215 static rtx sge_plus PARAMS ((enum machine_mode, rtx, rtx));
6216 static rtx sge_plus_constant PARAMS ((rtx, rtx));
6217 static int cmp_combine_givs_stats PARAMS ((const PTR, const PTR));
6218 static int cmp_recombine_givs_stats PARAMS ((const PTR, const PTR));
6221 simplify_giv_expr (loop, x, benefit)
6222 const struct loop *loop;
6226 enum machine_mode mode = GET_MODE (x);
6230 /* If this is not an integer mode, or if we cannot do arithmetic in this
6231 mode, this can't be a giv. */
6232 if (mode != VOIDmode
6233 && (GET_MODE_CLASS (mode) != MODE_INT
6234 || GET_MODE_BITSIZE (mode) > HOST_BITS_PER_WIDE_INT))
6237 switch (GET_CODE (x))
6240 arg0 = simplify_giv_expr (loop, XEXP (x, 0), benefit);
6241 arg1 = simplify_giv_expr (loop, XEXP (x, 1), benefit);
6242 if (arg0 == 0 || arg1 == 0)
6245 /* Put constant last, CONST_INT last if both constant. */
6246 if ((GET_CODE (arg0) == USE
6247 || GET_CODE (arg0) == CONST_INT)
6248 && ! ((GET_CODE (arg0) == USE
6249 && GET_CODE (arg1) == USE)
6250 || GET_CODE (arg1) == CONST_INT))
6251 tem = arg0, arg0 = arg1, arg1 = tem;
6253 /* Handle addition of zero, then addition of an invariant. */
6254 if (arg1 == const0_rtx)
6256 else if (GET_CODE (arg1) == CONST_INT || GET_CODE (arg1) == USE)
6257 switch (GET_CODE (arg0))
6261 /* Adding two invariants must result in an invariant, so enclose
6262 addition operation inside a USE and return it. */
6263 if (GET_CODE (arg0) == USE)
6264 arg0 = XEXP (arg0, 0);
6265 if (GET_CODE (arg1) == USE)
6266 arg1 = XEXP (arg1, 0);
6268 if (GET_CODE (arg0) == CONST_INT)
6269 tem = arg0, arg0 = arg1, arg1 = tem;
6270 if (GET_CODE (arg1) == CONST_INT)
6271 tem = sge_plus_constant (arg0, arg1);
6273 tem = sge_plus (mode, arg0, arg1);
6275 if (GET_CODE (tem) != CONST_INT)
6276 tem = gen_rtx_USE (mode, tem);
6281 /* biv + invar or mult + invar. Return sum. */
6282 return gen_rtx_PLUS (mode, arg0, arg1);
6285 /* (a + invar_1) + invar_2. Associate. */
6287 simplify_giv_expr (loop,
6299 /* Each argument must be either REG, PLUS, or MULT. Convert REG to
6300 MULT to reduce cases. */
6301 if (GET_CODE (arg0) == REG)
6302 arg0 = gen_rtx_MULT (mode, arg0, const1_rtx);
6303 if (GET_CODE (arg1) == REG)
6304 arg1 = gen_rtx_MULT (mode, arg1, const1_rtx);
6306 /* Now have PLUS + PLUS, PLUS + MULT, MULT + PLUS, or MULT + MULT.
6307 Put a MULT first, leaving PLUS + PLUS, MULT + PLUS, or MULT + MULT.
6308 Recurse to associate the second PLUS. */
6309 if (GET_CODE (arg1) == MULT)
6310 tem = arg0, arg0 = arg1, arg1 = tem;
6312 if (GET_CODE (arg1) == PLUS)
6314 simplify_giv_expr (loop,
6316 gen_rtx_PLUS (mode, arg0,
6321 /* Now must have MULT + MULT. Distribute if same biv, else not giv. */
6322 if (GET_CODE (arg0) != MULT || GET_CODE (arg1) != MULT)
6325 if (!rtx_equal_p (arg0, arg1))
6328 return simplify_giv_expr (loop,
6337 /* Handle "a - b" as "a + b * (-1)". */
6338 return simplify_giv_expr (loop,
6347 arg0 = simplify_giv_expr (loop, XEXP (x, 0), benefit);
6348 arg1 = simplify_giv_expr (loop, XEXP (x, 1), benefit);
6349 if (arg0 == 0 || arg1 == 0)
6352 /* Put constant last, CONST_INT last if both constant. */
6353 if ((GET_CODE (arg0) == USE || GET_CODE (arg0) == CONST_INT)
6354 && GET_CODE (arg1) != CONST_INT)
6355 tem = arg0, arg0 = arg1, arg1 = tem;
6357 /* If second argument is not now constant, not giv. */
6358 if (GET_CODE (arg1) != USE && GET_CODE (arg1) != CONST_INT)
6361 /* Handle multiply by 0 or 1. */
6362 if (arg1 == const0_rtx)
6365 else if (arg1 == const1_rtx)
6368 switch (GET_CODE (arg0))
6371 /* biv * invar. Done. */
6372 return gen_rtx_MULT (mode, arg0, arg1);
6375 /* Product of two constants. */
6376 return GEN_INT (INTVAL (arg0) * INTVAL (arg1));
6379 /* invar * invar is a giv, but attempt to simplify it somehow. */
6380 if (GET_CODE (arg1) != CONST_INT)
6383 arg0 = XEXP (arg0, 0);
6384 if (GET_CODE (arg0) == MULT)
6386 /* (invar_0 * invar_1) * invar_2. Associate. */
6387 return simplify_giv_expr (loop,
6396 /* Porpagate the MULT expressions to the intermost nodes. */
6397 else if (GET_CODE (arg0) == PLUS)
6399 /* (invar_0 + invar_1) * invar_2. Distribute. */
6400 return simplify_giv_expr (loop,
6412 return gen_rtx_USE (mode, gen_rtx_MULT (mode, arg0, arg1));
6415 /* (a * invar_1) * invar_2. Associate. */
6416 return simplify_giv_expr (loop,
6425 /* (a + invar_1) * invar_2. Distribute. */
6426 return simplify_giv_expr (loop,
6441 /* Shift by constant is multiply by power of two. */
6442 if (GET_CODE (XEXP (x, 1)) != CONST_INT)
6446 simplify_giv_expr (loop,
6449 GEN_INT ((HOST_WIDE_INT) 1
6450 << INTVAL (XEXP (x, 1)))),
6454 /* "-a" is "a * (-1)" */
6455 return simplify_giv_expr (loop,
6456 gen_rtx_MULT (mode, XEXP (x, 0), constm1_rtx),
6460 /* "~a" is "-a - 1". Silly, but easy. */
6461 return simplify_giv_expr (loop,
6462 gen_rtx_MINUS (mode,
6463 gen_rtx_NEG (mode, XEXP (x, 0)),
6468 /* Already in proper form for invariant. */
6472 /* If this is a new register, we can't deal with it. */
6473 if (REGNO (x) >= max_reg_before_loop)
6476 /* Check for biv or giv. */
6477 switch (REG_IV_TYPE (REGNO (x)))
6481 case GENERAL_INDUCT:
6483 struct induction *v = REG_IV_INFO (REGNO (x));
6485 /* Form expression from giv and add benefit. Ensure this giv
6486 can derive another and subtract any needed adjustment if so. */
6487 *benefit += v->benefit;
6491 tem = gen_rtx_PLUS (mode, gen_rtx_MULT (mode,
6492 v->src_reg, v->mult_val),
6495 if (v->derive_adjustment)
6496 tem = gen_rtx_MINUS (mode, tem, v->derive_adjustment);
6497 return simplify_giv_expr (loop, tem, benefit);
6501 /* If it isn't an induction variable, and it is invariant, we
6502 may be able to simplify things further by looking through
6503 the bits we just moved outside the loop. */
6504 if (loop_invariant_p (loop, x) == 1)
6508 for (m = the_movables; m ; m = m->next)
6509 if (rtx_equal_p (x, m->set_dest))
6511 /* Ok, we found a match. Substitute and simplify. */
6513 /* If we match another movable, we must use that, as
6514 this one is going away. */
6516 return simplify_giv_expr (loop, m->match->set_dest,
6519 /* If consec is non-zero, this is a member of a group of
6520 instructions that were moved together. We handle this
6521 case only to the point of seeking to the last insn and
6522 looking for a REG_EQUAL. Fail if we don't find one. */
6527 do { tem = NEXT_INSN (tem); } while (--i > 0);
6529 tem = find_reg_note (tem, REG_EQUAL, NULL_RTX);
6531 tem = XEXP (tem, 0);
6535 tem = single_set (m->insn);
6537 tem = SET_SRC (tem);
6542 /* What we are most interested in is pointer
6543 arithmetic on invariants -- only take
6544 patterns we may be able to do something with. */
6545 if (GET_CODE (tem) == PLUS
6546 || GET_CODE (tem) == MULT
6547 || GET_CODE (tem) == ASHIFT
6548 || GET_CODE (tem) == CONST_INT
6549 || GET_CODE (tem) == SYMBOL_REF)
6551 tem = simplify_giv_expr (loop, tem, benefit);
6555 else if (GET_CODE (tem) == CONST
6556 && GET_CODE (XEXP (tem, 0)) == PLUS
6557 && GET_CODE (XEXP (XEXP (tem, 0), 0)) == SYMBOL_REF
6558 && GET_CODE (XEXP (XEXP (tem, 0), 1)) == CONST_INT)
6560 tem = simplify_giv_expr (loop, XEXP (tem, 0),
6572 /* Fall through to general case. */
6574 /* If invariant, return as USE (unless CONST_INT).
6575 Otherwise, not giv. */
6576 if (GET_CODE (x) == USE)
6579 if (loop_invariant_p (loop, x) == 1)
6581 if (GET_CODE (x) == CONST_INT)
6583 if (GET_CODE (x) == CONST
6584 && GET_CODE (XEXP (x, 0)) == PLUS
6585 && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF
6586 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT)
6588 return gen_rtx_USE (mode, x);
6595 /* This routine folds invariants such that there is only ever one
6596 CONST_INT in the summation. It is only used by simplify_giv_expr. */
6599 sge_plus_constant (x, c)
6602 if (GET_CODE (x) == CONST_INT)
6603 return GEN_INT (INTVAL (x) + INTVAL (c));
6604 else if (GET_CODE (x) != PLUS)
6605 return gen_rtx_PLUS (GET_MODE (x), x, c);
6606 else if (GET_CODE (XEXP (x, 1)) == CONST_INT)
6608 return gen_rtx_PLUS (GET_MODE (x), XEXP (x, 0),
6609 GEN_INT (INTVAL (XEXP (x, 1)) + INTVAL (c)));
6611 else if (GET_CODE (XEXP (x, 0)) == PLUS
6612 || GET_CODE (XEXP (x, 1)) != PLUS)
6614 return gen_rtx_PLUS (GET_MODE (x),
6615 sge_plus_constant (XEXP (x, 0), c), XEXP (x, 1));
6619 return gen_rtx_PLUS (GET_MODE (x),
6620 sge_plus_constant (XEXP (x, 1), c), XEXP (x, 0));
6625 sge_plus (mode, x, y)
6626 enum machine_mode mode;
6629 while (GET_CODE (y) == PLUS)
6631 rtx a = XEXP (y, 0);
6632 if (GET_CODE (a) == CONST_INT)
6633 x = sge_plus_constant (x, a);
6635 x = gen_rtx_PLUS (mode, x, a);
6638 if (GET_CODE (y) == CONST_INT)
6639 x = sge_plus_constant (x, y);
6641 x = gen_rtx_PLUS (mode, x, y);
6645 /* Help detect a giv that is calculated by several consecutive insns;
6649 The caller has already identified the first insn P as having a giv as dest;
6650 we check that all other insns that set the same register follow
6651 immediately after P, that they alter nothing else,
6652 and that the result of the last is still a giv.
6654 The value is 0 if the reg set in P is not really a giv.
6655 Otherwise, the value is the amount gained by eliminating
6656 all the consecutive insns that compute the value.
6658 FIRST_BENEFIT is the amount gained by eliminating the first insn, P.
6659 SRC_REG is the reg of the biv; DEST_REG is the reg of the giv.
6661 The coefficients of the ultimate giv value are stored in
6662 *MULT_VAL and *ADD_VAL. */
6665 consec_sets_giv (loop, first_benefit, p, src_reg, dest_reg,
6666 add_val, mult_val, last_consec_insn)
6667 const struct loop *loop;
6674 rtx *last_consec_insn;
6682 /* Indicate that this is a giv so that we can update the value produced in
6683 each insn of the multi-insn sequence.
6685 This induction structure will be used only by the call to
6686 general_induction_var below, so we can allocate it on our stack.
6687 If this is a giv, our caller will replace the induct var entry with
6688 a new induction structure. */
6690 = (struct induction *) alloca (sizeof (struct induction));
6691 v->src_reg = src_reg;
6692 v->mult_val = *mult_val;
6693 v->add_val = *add_val;
6694 v->benefit = first_benefit;
6696 v->derive_adjustment = 0;
6698 REG_IV_TYPE (REGNO (dest_reg)) = GENERAL_INDUCT;
6699 REG_IV_INFO (REGNO (dest_reg)) = v;
6701 count = VARRAY_INT (n_times_set, REGNO (dest_reg)) - 1;
6706 code = GET_CODE (p);
6708 /* If libcall, skip to end of call sequence. */
6709 if (code == INSN && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
6713 && (set = single_set (p))
6714 && GET_CODE (SET_DEST (set)) == REG
6715 && SET_DEST (set) == dest_reg
6716 && (general_induction_var (loop, SET_SRC (set), &src_reg,
6717 add_val, mult_val, 0, &benefit, VOIDmode)
6718 /* Giv created by equivalent expression. */
6719 || ((temp = find_reg_note (p, REG_EQUAL, NULL_RTX))
6720 && general_induction_var (loop, XEXP (temp, 0), &src_reg,
6721 add_val, mult_val, 0, &benefit,
6723 && src_reg == v->src_reg)
6725 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
6726 benefit += libcall_benefit (p);
6729 v->mult_val = *mult_val;
6730 v->add_val = *add_val;
6731 v->benefit = benefit;
6733 else if (code != NOTE)
6735 /* Allow insns that set something other than this giv to a
6736 constant. Such insns are needed on machines which cannot
6737 include long constants and should not disqualify a giv. */
6739 && (set = single_set (p))
6740 && SET_DEST (set) != dest_reg
6741 && CONSTANT_P (SET_SRC (set)))
6744 REG_IV_TYPE (REGNO (dest_reg)) = UNKNOWN_INDUCT;
6749 *last_consec_insn = p;
6753 /* Return an rtx, if any, that expresses giv G2 as a function of the register
6754 represented by G1. If no such expression can be found, or it is clear that
6755 it cannot possibly be a valid address, 0 is returned.
6757 To perform the computation, we note that
6760 where `v' is the biv.
6762 So G2 = (y/b) * G1 + (b - a*y/x).
6764 Note that MULT = y/x.
6766 Update: A and B are now allowed to be additive expressions such that
6767 B contains all variables in A. That is, computing B-A will not require
6768 subtracting variables. */
6771 express_from_1 (a, b, mult)
6774 /* If MULT is zero, then A*MULT is zero, and our expression is B. */
6776 if (mult == const0_rtx)
6779 /* If MULT is not 1, we cannot handle A with non-constants, since we
6780 would then be required to subtract multiples of the registers in A.
6781 This is theoretically possible, and may even apply to some Fortran
6782 constructs, but it is a lot of work and we do not attempt it here. */
6784 if (mult != const1_rtx && GET_CODE (a) != CONST_INT)
6787 /* In general these structures are sorted top to bottom (down the PLUS
6788 chain), but not left to right across the PLUS. If B is a higher
6789 order giv than A, we can strip one level and recurse. If A is higher
6790 order, we'll eventually bail out, but won't know that until the end.
6791 If they are the same, we'll strip one level around this loop. */
6793 while (GET_CODE (a) == PLUS && GET_CODE (b) == PLUS)
6795 rtx ra, rb, oa, ob, tmp;
6797 ra = XEXP (a, 0), oa = XEXP (a, 1);
6798 if (GET_CODE (ra) == PLUS)
6799 tmp = ra, ra = oa, oa = tmp;
6801 rb = XEXP (b, 0), ob = XEXP (b, 1);
6802 if (GET_CODE (rb) == PLUS)
6803 tmp = rb, rb = ob, ob = tmp;
6805 if (rtx_equal_p (ra, rb))
6806 /* We matched: remove one reg completely. */
6808 else if (GET_CODE (ob) != PLUS && rtx_equal_p (ra, ob))
6809 /* An alternate match. */
6811 else if (GET_CODE (oa) != PLUS && rtx_equal_p (oa, rb))
6812 /* An alternate match. */
6816 /* Indicates an extra register in B. Strip one level from B and
6817 recurse, hoping B was the higher order expression. */
6818 ob = express_from_1 (a, ob, mult);
6821 return gen_rtx_PLUS (GET_MODE (b), rb, ob);
6825 /* Here we are at the last level of A, go through the cases hoping to
6826 get rid of everything but a constant. */
6828 if (GET_CODE (a) == PLUS)
6832 ra = XEXP (a, 0), oa = XEXP (a, 1);
6833 if (rtx_equal_p (oa, b))
6835 else if (!rtx_equal_p (ra, b))
6838 if (GET_CODE (oa) != CONST_INT)
6841 return GEN_INT (-INTVAL (oa) * INTVAL (mult));
6843 else if (GET_CODE (a) == CONST_INT)
6845 return plus_constant (b, -INTVAL (a) * INTVAL (mult));
6847 else if (CONSTANT_P (a))
6849 return simplify_gen_binary (MINUS, GET_MODE (b) != VOIDmode ? GET_MODE (b) : GET_MODE (a), const0_rtx, a);
6851 else if (GET_CODE (b) == PLUS)
6853 if (rtx_equal_p (a, XEXP (b, 0)))
6855 else if (rtx_equal_p (a, XEXP (b, 1)))
6860 else if (rtx_equal_p (a, b))
6867 express_from (g1, g2)
6868 struct induction *g1, *g2;
6872 /* The value that G1 will be multiplied by must be a constant integer. Also,
6873 the only chance we have of getting a valid address is if b*c/a (see above
6874 for notation) is also an integer. */
6875 if (GET_CODE (g1->mult_val) == CONST_INT
6876 && GET_CODE (g2->mult_val) == CONST_INT)
6878 if (g1->mult_val == const0_rtx
6879 || INTVAL (g2->mult_val) % INTVAL (g1->mult_val) != 0)
6881 mult = GEN_INT (INTVAL (g2->mult_val) / INTVAL (g1->mult_val));
6883 else if (rtx_equal_p (g1->mult_val, g2->mult_val))
6887 /* ??? Find out if the one is a multiple of the other? */
6891 add = express_from_1 (g1->add_val, g2->add_val, mult);
6892 if (add == NULL_RTX)
6894 /* Failed. If we've got a multiplication factor between G1 and G2,
6895 scale G1's addend and try again. */
6896 if (INTVAL (mult) > 1)
6898 rtx g1_add_val = g1->add_val;
6899 if (GET_CODE (g1_add_val) == MULT
6900 && GET_CODE (XEXP (g1_add_val, 1)) == CONST_INT)
6903 m = INTVAL (mult) * INTVAL (XEXP (g1_add_val, 1));
6904 g1_add_val = gen_rtx_MULT (GET_MODE (g1_add_val),
6905 XEXP (g1_add_val, 0), GEN_INT (m));
6909 g1_add_val = gen_rtx_MULT (GET_MODE (g1_add_val), g1_add_val,
6913 add = express_from_1 (g1_add_val, g2->add_val, const1_rtx);
6916 if (add == NULL_RTX)
6919 /* Form simplified final result. */
6920 if (mult == const0_rtx)
6922 else if (mult == const1_rtx)
6923 mult = g1->dest_reg;
6925 mult = gen_rtx_MULT (g2->mode, g1->dest_reg, mult);
6927 if (add == const0_rtx)
6931 if (GET_CODE (add) == PLUS
6932 && CONSTANT_P (XEXP (add, 1)))
6934 rtx tem = XEXP (add, 1);
6935 mult = gen_rtx_PLUS (g2->mode, mult, XEXP (add, 0));
6939 return gen_rtx_PLUS (g2->mode, mult, add);
6944 /* Return an rtx, if any, that expresses giv G2 as a function of the register
6945 represented by G1. This indicates that G2 should be combined with G1 and
6946 that G2 can use (either directly or via an address expression) a register
6947 used to represent G1. */
6950 combine_givs_p (g1, g2)
6951 struct induction *g1, *g2;
6953 rtx tem = express_from (g1, g2);
6955 /* If these givs are identical, they can be combined. We use the results
6956 of express_from because the addends are not in a canonical form, so
6957 rtx_equal_p is a weaker test. */
6958 /* But don't combine a DEST_REG giv with a DEST_ADDR giv; we want the
6959 combination to be the other way round. */
6960 if (tem == g1->dest_reg
6961 && (g1->giv_type == DEST_REG || g2->giv_type == DEST_ADDR))
6963 return g1->dest_reg;
6966 /* If G2 can be expressed as a function of G1 and that function is valid
6967 as an address and no more expensive than using a register for G2,
6968 the expression of G2 in terms of G1 can be used. */
6970 && g2->giv_type == DEST_ADDR
6971 && memory_address_p (g2->mem_mode, tem)
6972 /* ??? Looses, especially with -fforce-addr, where *g2->location
6973 will always be a register, and so anything more complicated
6977 && ADDRESS_COST (tem) <= ADDRESS_COST (*g2->location)
6979 && rtx_cost (tem, MEM) <= rtx_cost (*g2->location, MEM)
6990 struct combine_givs_stats
6997 cmp_combine_givs_stats (xp, yp)
7001 const struct combine_givs_stats * const x =
7002 (const struct combine_givs_stats *) xp;
7003 const struct combine_givs_stats * const y =
7004 (const struct combine_givs_stats *) yp;
7006 d = y->total_benefit - x->total_benefit;
7007 /* Stabilize the sort. */
7009 d = x->giv_number - y->giv_number;
7013 /* Check all pairs of givs for iv_class BL and see if any can be combined with
7014 any other. If so, point SAME to the giv combined with and set NEW_REG to
7015 be an expression (in terms of the other giv's DEST_REG) equivalent to the
7016 giv. Also, update BENEFIT and related fields for cost/benefit analysis. */
7020 struct iv_class *bl;
7022 /* Additional benefit to add for being combined multiple times. */
7023 const int extra_benefit = 3;
7025 struct induction *g1, *g2, **giv_array;
7026 int i, j, k, giv_count;
7027 struct combine_givs_stats *stats;
7030 /* Count givs, because bl->giv_count is incorrect here. */
7032 for (g1 = bl->giv; g1; g1 = g1->next_iv)
7037 = (struct induction **) alloca (giv_count * sizeof (struct induction *));
7039 for (g1 = bl->giv; g1; g1 = g1->next_iv)
7041 giv_array[i++] = g1;
7043 stats = (struct combine_givs_stats *) xcalloc (giv_count, sizeof (*stats));
7044 can_combine = (rtx *) xcalloc (giv_count, giv_count * sizeof(rtx));
7046 for (i = 0; i < giv_count; i++)
7052 stats[i].giv_number = i;
7054 /* If a DEST_REG GIV is used only once, do not allow it to combine
7055 with anything, for in doing so we will gain nothing that cannot
7056 be had by simply letting the GIV with which we would have combined
7057 to be reduced on its own. The losage shows up in particular with
7058 DEST_ADDR targets on hosts with reg+reg addressing, though it can
7059 be seen elsewhere as well. */
7060 if (g1->giv_type == DEST_REG
7061 && (single_use = VARRAY_RTX (reg_single_usage, REGNO (g1->dest_reg)))
7062 && single_use != const0_rtx)
7065 this_benefit = g1->benefit;
7066 /* Add an additional weight for zero addends. */
7067 if (g1->no_const_addval)
7070 for (j = 0; j < giv_count; j++)
7076 && (this_combine = combine_givs_p (g1, g2)) != NULL_RTX)
7078 can_combine[i*giv_count + j] = this_combine;
7079 this_benefit += g2->benefit + extra_benefit;
7082 stats[i].total_benefit = this_benefit;
7085 /* Iterate, combining until we can't. */
7087 qsort (stats, giv_count, sizeof(*stats), cmp_combine_givs_stats);
7089 if (loop_dump_stream)
7091 fprintf (loop_dump_stream, "Sorted combine statistics:\n");
7092 for (k = 0; k < giv_count; k++)
7094 g1 = giv_array[stats[k].giv_number];
7095 if (!g1->combined_with && !g1->same)
7096 fprintf (loop_dump_stream, " {%d, %d}",
7097 INSN_UID (giv_array[stats[k].giv_number]->insn),
7098 stats[k].total_benefit);
7100 putc ('\n', loop_dump_stream);
7103 for (k = 0; k < giv_count; k++)
7105 int g1_add_benefit = 0;
7107 i = stats[k].giv_number;
7110 /* If it has already been combined, skip. */
7111 if (g1->combined_with || g1->same)
7114 for (j = 0; j < giv_count; j++)
7117 if (g1 != g2 && can_combine[i*giv_count + j]
7118 /* If it has already been combined, skip. */
7119 && ! g2->same && ! g2->combined_with)
7123 g2->new_reg = can_combine[i*giv_count + j];
7125 g1->combined_with++;
7126 g1->lifetime += g2->lifetime;
7128 g1_add_benefit += g2->benefit;
7130 /* ??? The new final_[bg]iv_value code does a much better job
7131 of finding replaceable giv's, and hence this code may no
7132 longer be necessary. */
7133 if (! g2->replaceable && REG_USERVAR_P (g2->dest_reg))
7134 g1_add_benefit -= copy_cost;
7136 /* To help optimize the next set of combinations, remove
7137 this giv from the benefits of other potential mates. */
7138 for (l = 0; l < giv_count; ++l)
7140 int m = stats[l].giv_number;
7141 if (can_combine[m*giv_count + j])
7142 stats[l].total_benefit -= g2->benefit + extra_benefit;
7145 if (loop_dump_stream)
7146 fprintf (loop_dump_stream,
7147 "giv at %d combined with giv at %d\n",
7148 INSN_UID (g2->insn), INSN_UID (g1->insn));
7152 /* To help optimize the next set of combinations, remove
7153 this giv from the benefits of other potential mates. */
7154 if (g1->combined_with)
7156 for (j = 0; j < giv_count; ++j)
7158 int m = stats[j].giv_number;
7159 if (can_combine[m*giv_count + i])
7160 stats[j].total_benefit -= g1->benefit + extra_benefit;
7163 g1->benefit += g1_add_benefit;
7165 /* We've finished with this giv, and everything it touched.
7166 Restart the combination so that proper weights for the
7167 rest of the givs are properly taken into account. */
7168 /* ??? Ideally we would compact the arrays at this point, so
7169 as to not cover old ground. But sanely compacting
7170 can_combine is tricky. */
7180 struct recombine_givs_stats
7183 int start_luid, end_luid;
7186 /* Used below as comparison function for qsort. We want a ascending luid
7187 when scanning the array starting at the end, thus the arguments are
7190 cmp_recombine_givs_stats (xp, yp)
7194 const struct recombine_givs_stats * const x =
7195 (const struct recombine_givs_stats *) xp;
7196 const struct recombine_givs_stats * const y =
7197 (const struct recombine_givs_stats *) yp;
7199 d = y->start_luid - x->start_luid;
7200 /* Stabilize the sort. */
7202 d = y->giv_number - x->giv_number;
7206 /* Scan X, which is a part of INSN, for the end of life of a giv. Also
7207 look for the start of life of a giv where the start has not been seen
7208 yet to unlock the search for the end of its life.
7209 Only consider givs that belong to BIV.
7210 Return the total number of lifetime ends that have been found. */
7212 find_life_end (x, stats, insn, biv)
7214 struct recombine_givs_stats *stats;
7221 code = GET_CODE (x);
7226 rtx reg = SET_DEST (x);
7227 if (GET_CODE (reg) == REG)
7229 int regno = REGNO (reg);
7230 struct induction *v = REG_IV_INFO (regno);
7232 if (REG_IV_TYPE (regno) == GENERAL_INDUCT
7234 && v->src_reg == biv
7235 && stats[v->ix].end_luid <= 0)
7237 /* If we see a 0 here for end_luid, it means that we have
7238 scanned the entire loop without finding any use at all.
7239 We must not predicate this code on a start_luid match
7240 since that would make the test fail for givs that have
7241 been hoisted out of inner loops. */
7242 if (stats[v->ix].end_luid == 0)
7244 stats[v->ix].end_luid = stats[v->ix].start_luid;
7245 return 1 + find_life_end (SET_SRC (x), stats, insn, biv);
7247 else if (stats[v->ix].start_luid == INSN_LUID (insn))
7248 stats[v->ix].end_luid = 0;
7250 return find_life_end (SET_SRC (x), stats, insn, biv);
7256 int regno = REGNO (x);
7257 struct induction *v = REG_IV_INFO (regno);
7259 if (REG_IV_TYPE (regno) == GENERAL_INDUCT
7261 && v->src_reg == biv
7262 && stats[v->ix].end_luid == 0)
7264 while (INSN_UID (insn) >= max_uid_for_loop)
7265 insn = NEXT_INSN (insn);
7266 stats[v->ix].end_luid = INSN_LUID (insn);
7279 fmt = GET_RTX_FORMAT (code);
7281 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
7284 retval += find_life_end (XEXP (x, i), stats, insn, biv);
7286 else if (fmt[i] == 'E')
7287 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
7288 retval += find_life_end (XVECEXP (x, i, j), stats, insn, biv);
7293 /* For each giv that has been combined with another, look if
7294 we can combine it with the most recently used one instead.
7295 This tends to shorten giv lifetimes, and helps the next step:
7296 try to derive givs from other givs. */
7298 recombine_givs (loop, bl, unroll_p)
7299 const struct loop *loop;
7300 struct iv_class *bl;
7303 struct induction *v, **giv_array, *last_giv;
7304 struct recombine_givs_stats *stats;
7307 int ends_need_computing;
7309 for (giv_count = 0, v = bl->giv; v; v = v->next_iv)
7315 = (struct induction **) xmalloc (giv_count * sizeof (struct induction *));
7316 stats = (struct recombine_givs_stats *) xmalloc (giv_count * sizeof *stats);
7318 /* Initialize stats and set up the ix field for each giv in stats to name
7319 the corresponding index into stats. */
7320 for (i = 0, v = bl->giv; v; v = v->next_iv)
7327 stats[i].giv_number = i;
7328 /* If this giv has been hoisted out of an inner loop, use the luid of
7329 the previous insn. */
7330 for (p = v->insn; INSN_UID (p) >= max_uid_for_loop; )
7332 stats[i].start_luid = INSN_LUID (p);
7336 qsort (stats, giv_count, sizeof(*stats), cmp_recombine_givs_stats);
7338 /* Set up the ix field for each giv in stats to name
7339 the corresponding index into stats, and
7340 do the actual most-recently-used recombination. */
7341 for (last_giv = 0, i = giv_count - 1; i >= 0; i--)
7343 v = giv_array[stats[i].giv_number];
7347 struct induction *old_same = v->same;
7350 /* combine_givs_p actually says if we can make this transformation.
7351 The other tests are here only to avoid keeping a giv alive
7352 that could otherwise be eliminated. */
7354 && ((old_same->maybe_dead && ! old_same->combined_with)
7355 || ! last_giv->maybe_dead
7356 || last_giv->combined_with)
7357 && (new_combine = combine_givs_p (last_giv, v)))
7359 old_same->combined_with--;
7360 v->new_reg = new_combine;
7362 last_giv->combined_with++;
7363 /* No need to update lifetimes / benefits here since we have
7364 already decided what to reduce. */
7366 if (loop_dump_stream)
7368 fprintf (loop_dump_stream,
7369 "giv at %d recombined with giv at %d as ",
7370 INSN_UID (v->insn), INSN_UID (last_giv->insn));
7371 print_rtl (loop_dump_stream, v->new_reg);
7372 putc ('\n', loop_dump_stream);
7378 else if (v->giv_type != DEST_REG)
7381 || (last_giv->maybe_dead && ! last_giv->combined_with)
7383 || v->combined_with)
7387 ends_need_computing = 0;
7388 /* For each DEST_REG giv, compute lifetime starts, and try to compute
7389 lifetime ends from regscan info. */
7390 for (i = giv_count - 1; i >= 0; i--)
7392 v = giv_array[stats[i].giv_number];
7395 if (v->giv_type == DEST_ADDR)
7397 /* Loop unrolling of an inner loop can even create new DEST_REG
7400 for (p = v->insn; INSN_UID (p) >= max_uid_for_loop; )
7402 stats[i].start_luid = stats[i].end_luid = INSN_LUID (p);
7404 stats[i].end_luid++;
7406 else /* v->giv_type == DEST_REG */
7410 stats[i].start_luid = INSN_LUID (v->insn);
7411 stats[i].end_luid = INSN_LUID (v->last_use);
7413 else if (INSN_UID (v->insn) >= max_uid_for_loop)
7416 /* This insn has been created by loop optimization on an inner
7417 loop. We don't have a proper start_luid that will match
7418 when we see the first set. But we do know that there will
7419 be no use before the set, so we can set end_luid to 0 so that
7420 we'll start looking for the last use right away. */
7421 for (p = PREV_INSN (v->insn); INSN_UID (p) >= max_uid_for_loop; )
7423 stats[i].start_luid = INSN_LUID (p);
7424 stats[i].end_luid = 0;
7425 ends_need_computing++;
7429 int regno = REGNO (v->dest_reg);
7430 int count = VARRAY_INT (n_times_set, regno) - 1;
7433 /* Find the first insn that sets the giv, so that we can verify
7434 if this giv's lifetime wraps around the loop. We also need
7435 the luid of the first setting insn in order to detect the
7436 last use properly. */
7439 p = prev_nonnote_insn (p);
7440 if (reg_set_p (v->dest_reg, p))
7444 stats[i].start_luid = INSN_LUID (p);
7445 if (stats[i].start_luid > uid_luid[REGNO_FIRST_UID (regno)])
7447 stats[i].end_luid = -1;
7448 ends_need_computing++;
7452 stats[i].end_luid = uid_luid[REGNO_LAST_UID (regno)];
7453 if (stats[i].end_luid > INSN_LUID (loop->end))
7455 stats[i].end_luid = -1;
7456 ends_need_computing++;
7463 /* If the regscan information was unconclusive for one or more DEST_REG
7464 givs, scan the all insn in the loop to find out lifetime ends. */
7465 if (ends_need_computing)
7467 rtx biv = bl->biv->src_reg;
7472 if (p == loop->start)
7477 ends_need_computing -= find_life_end (PATTERN (p), stats, p, biv);
7479 while (ends_need_computing);
7482 /* Set start_luid back to the last insn that sets the giv. This allows
7483 more combinations. */
7484 for (i = giv_count - 1; i >= 0; i--)
7486 v = giv_array[stats[i].giv_number];
7489 if (INSN_UID (v->insn) < max_uid_for_loop)
7490 stats[i].start_luid = INSN_LUID (v->insn);
7493 /* Now adjust lifetime ends by taking combined givs into account. */
7494 for (i = giv_count - 1; i >= 0; i--)
7499 v = giv_array[stats[i].giv_number];
7502 if (v->same && ! v->same->ignore)
7505 luid = stats[i].start_luid;
7506 /* Use unsigned arithmetic to model loop wrap-around. */
7507 if (luid - stats[j].start_luid
7508 > (unsigned) stats[j].end_luid - stats[j].start_luid)
7509 stats[j].end_luid = luid;
7513 qsort (stats, giv_count, sizeof(*stats), cmp_recombine_givs_stats);
7515 /* Try to derive DEST_REG givs from previous DEST_REG givs with the
7516 same mult_val and non-overlapping lifetime. This reduces register
7518 Once we find a DEST_REG giv that is suitable to derive others from,
7519 we set last_giv to this giv, and try to derive as many other DEST_REG
7520 givs from it without joining overlapping lifetimes. If we then
7521 encounter a DEST_REG giv that we can't derive, we set rescan to the
7522 index for this giv (unless rescan is already set).
7523 When we are finished with the current LAST_GIV (i.e. the inner loop
7524 terminates), we start again with rescan, which then becomes the new
7526 for (i = giv_count - 1; i >= 0; i = rescan)
7528 int life_start = 0, life_end = 0;
7530 for (last_giv = 0, rescan = -1; i >= 0; i--)
7534 v = giv_array[stats[i].giv_number];
7535 if (v->giv_type != DEST_REG || v->derived_from || v->same)
7539 /* Don't use a giv that's likely to be dead to derive
7540 others - that would be likely to keep that giv alive. */
7541 if (! v->maybe_dead || v->combined_with)
7544 life_start = stats[i].start_luid;
7545 life_end = stats[i].end_luid;
7549 /* Use unsigned arithmetic to model loop wrap around. */
7550 if (((unsigned) stats[i].start_luid - life_start
7551 >= (unsigned) life_end - life_start)
7552 && ((unsigned) stats[i].end_luid - life_start
7553 > (unsigned) life_end - life_start)
7554 /* Check that the giv insn we're about to use for deriving
7555 precedes all uses of that giv. Note that initializing the
7556 derived giv would defeat the purpose of reducing register
7558 ??? We could arrange to move the insn. */
7559 && ((unsigned) stats[i].end_luid - INSN_LUID (loop->start)
7560 > (unsigned) stats[i].start_luid - INSN_LUID (loop->start))
7561 && rtx_equal_p (last_giv->mult_val, v->mult_val)
7562 /* ??? Could handle libcalls, but would need more logic. */
7563 && ! find_reg_note (v->insn, REG_RETVAL, NULL_RTX)
7564 /* We would really like to know if for any giv that v
7565 is combined with, v->insn or any intervening biv increment
7566 dominates that combined giv. However, we
7567 don't have this detailed control flow information.
7568 N.B. since last_giv will be reduced, it is valid
7569 anywhere in the loop, so we don't need to check the
7570 validity of last_giv.
7571 We rely here on the fact that v->always_executed implies that
7572 there is no jump to someplace else in the loop before the
7573 giv insn, and hence any insn that is executed before the
7574 giv insn in the loop will have a lower luid. */
7575 && (v->always_executed || ! v->combined_with)
7576 && (sum = express_from (last_giv, v))
7577 /* Make sure we don't make the add more expensive. ADD_COST
7578 doesn't take different costs of registers and constants into
7579 account, so compare the cost of the actual SET_SRCs. */
7580 && (rtx_cost (sum, SET)
7581 <= rtx_cost (SET_SRC (single_set (v->insn)), SET))
7582 /* ??? unroll can't understand anything but reg + const_int
7583 sums. It would be cleaner to fix unroll. */
7584 && ((GET_CODE (sum) == PLUS
7585 && GET_CODE (XEXP (sum, 0)) == REG
7586 && GET_CODE (XEXP (sum, 1)) == CONST_INT)
7588 && validate_change (v->insn, &PATTERN (v->insn),
7589 gen_rtx_SET (VOIDmode, v->dest_reg, sum), 0))
7591 v->derived_from = last_giv;
7592 life_end = stats[i].end_luid;
7594 if (loop_dump_stream)
7596 fprintf (loop_dump_stream,
7597 "giv at %d derived from %d as ",
7598 INSN_UID (v->insn), INSN_UID (last_giv->insn));
7599 print_rtl (loop_dump_stream, sum);
7600 putc ('\n', loop_dump_stream);
7603 else if (rescan < 0)
7613 /* EMIT code before INSERT_BEFORE to set REG = B * M + A. */
7616 emit_iv_add_mult (b, m, a, reg, insert_before)
7617 rtx b; /* initial value of basic induction variable */
7618 rtx m; /* multiplicative constant */
7619 rtx a; /* additive constant */
7620 rtx reg; /* destination register */
7626 /* Prevent unexpected sharing of these rtx. */
7630 /* Increase the lifetime of any invariants moved further in code. */
7631 update_reg_last_use (a, insert_before);
7632 update_reg_last_use (b, insert_before);
7633 update_reg_last_use (m, insert_before);
7636 result = expand_mult_add (b, reg, m, a, GET_MODE (reg), 0);
7638 emit_move_insn (reg, result);
7639 seq = gen_sequence ();
7642 emit_insn_before (seq, insert_before);
7644 /* It is entirely possible that the expansion created lots of new
7645 registers. Iterate over the sequence we just created and
7648 if (GET_CODE (seq) == SEQUENCE)
7651 for (i = 0; i < XVECLEN (seq, 0); ++i)
7653 rtx set = single_set (XVECEXP (seq, 0, i));
7654 if (set && GET_CODE (SET_DEST (set)) == REG)
7655 record_base_value (REGNO (SET_DEST (set)), SET_SRC (set), 0);
7658 else if (GET_CODE (seq) == SET
7659 && GET_CODE (SET_DEST (seq)) == REG)
7660 record_base_value (REGNO (SET_DEST (seq)), SET_SRC (seq), 0);
7663 /* Test whether A * B can be computed without
7664 an actual multiply insn. Value is 1 if so. */
7667 product_cheap_p (a, b)
7673 struct obstack *old_rtl_obstack = rtl_obstack;
7674 char *storage = (char *) obstack_alloc (&temp_obstack, 0);
7677 /* If only one is constant, make it B. */
7678 if (GET_CODE (a) == CONST_INT)
7679 tmp = a, a = b, b = tmp;
7681 /* If first constant, both constant, so don't need multiply. */
7682 if (GET_CODE (a) == CONST_INT)
7685 /* If second not constant, neither is constant, so would need multiply. */
7686 if (GET_CODE (b) != CONST_INT)
7689 /* One operand is constant, so might not need multiply insn. Generate the
7690 code for the multiply and see if a call or multiply, or long sequence
7691 of insns is generated. */
7693 rtl_obstack = &temp_obstack;
7695 expand_mult (GET_MODE (a), a, b, NULL_RTX, 0);
7696 tmp = gen_sequence ();
7699 if (GET_CODE (tmp) == SEQUENCE)
7701 if (XVEC (tmp, 0) == 0)
7703 else if (XVECLEN (tmp, 0) > 3)
7706 for (i = 0; i < XVECLEN (tmp, 0); i++)
7708 rtx insn = XVECEXP (tmp, 0, i);
7710 if (GET_CODE (insn) != INSN
7711 || (GET_CODE (PATTERN (insn)) == SET
7712 && GET_CODE (SET_SRC (PATTERN (insn))) == MULT)
7713 || (GET_CODE (PATTERN (insn)) == PARALLEL
7714 && GET_CODE (XVECEXP (PATTERN (insn), 0, 0)) == SET
7715 && GET_CODE (SET_SRC (XVECEXP (PATTERN (insn), 0, 0))) == MULT))
7722 else if (GET_CODE (tmp) == SET
7723 && GET_CODE (SET_SRC (tmp)) == MULT)
7725 else if (GET_CODE (tmp) == PARALLEL
7726 && GET_CODE (XVECEXP (tmp, 0, 0)) == SET
7727 && GET_CODE (SET_SRC (XVECEXP (tmp, 0, 0))) == MULT)
7730 /* Free any storage we obtained in generating this multiply and restore rtl
7731 allocation to its normal obstack. */
7732 obstack_free (&temp_obstack, storage);
7733 rtl_obstack = old_rtl_obstack;
7738 /* Check to see if loop can be terminated by a "decrement and branch until
7739 zero" instruction. If so, add a REG_NONNEG note to the branch insn if so.
7740 Also try reversing an increment loop to a decrement loop
7741 to see if the optimization can be performed.
7742 Value is nonzero if optimization was performed. */
7744 /* This is useful even if the architecture doesn't have such an insn,
7745 because it might change a loops which increments from 0 to n to a loop
7746 which decrements from n to 0. A loop that decrements to zero is usually
7747 faster than one that increments from zero. */
7749 /* ??? This could be rewritten to use some of the loop unrolling procedures,
7750 such as approx_final_value, biv_total_increment, loop_iterations, and
7751 final_[bg]iv_value. */
7754 check_dbra_loop (loop, insn_count)
7758 struct iv_class *bl;
7765 rtx before_comparison;
7769 int compare_and_branch;
7770 rtx loop_start = loop->start;
7771 rtx loop_end = loop->end;
7772 struct loop_info *loop_info = LOOP_INFO (loop);
7774 /* If last insn is a conditional branch, and the insn before tests a
7775 register value, try to optimize it. Otherwise, we can't do anything. */
7777 jump = PREV_INSN (loop_end);
7778 comparison = get_condition_for_loop (loop, jump);
7779 if (comparison == 0)
7781 if (!onlyjump_p (jump))
7784 /* Try to compute whether the compare/branch at the loop end is one or
7785 two instructions. */
7786 get_condition (jump, &first_compare);
7787 if (first_compare == jump)
7788 compare_and_branch = 1;
7789 else if (first_compare == prev_nonnote_insn (jump))
7790 compare_and_branch = 2;
7795 /* If more than one condition is present to control the loop, then
7796 do not proceed, as this function does not know how to rewrite
7797 loop tests with more than one condition.
7799 Look backwards from the first insn in the last comparison
7800 sequence and see if we've got another comparison sequence. */
7803 if ((jump1 = prev_nonnote_insn (first_compare)) != loop->cont)
7804 if (GET_CODE (jump1) == JUMP_INSN)
7808 /* Check all of the bivs to see if the compare uses one of them.
7809 Skip biv's set more than once because we can't guarantee that
7810 it will be zero on the last iteration. Also skip if the biv is
7811 used between its update and the test insn. */
7813 for (bl = loop_iv_list; bl; bl = bl->next)
7815 if (bl->biv_count == 1
7816 && ! bl->biv->maybe_multiple
7817 && bl->biv->dest_reg == XEXP (comparison, 0)
7818 && ! reg_used_between_p (regno_reg_rtx[bl->regno], bl->biv->insn,
7826 /* Look for the case where the basic induction variable is always
7827 nonnegative, and equals zero on the last iteration.
7828 In this case, add a reg_note REG_NONNEG, which allows the
7829 m68k DBRA instruction to be used. */
7831 if (((GET_CODE (comparison) == GT
7832 && GET_CODE (XEXP (comparison, 1)) == CONST_INT
7833 && INTVAL (XEXP (comparison, 1)) == -1)
7834 || (GET_CODE (comparison) == NE && XEXP (comparison, 1) == const0_rtx))
7835 && GET_CODE (bl->biv->add_val) == CONST_INT
7836 && INTVAL (bl->biv->add_val) < 0)
7838 /* Initial value must be greater than 0,
7839 init_val % -dec_value == 0 to ensure that it equals zero on
7840 the last iteration */
7842 if (GET_CODE (bl->initial_value) == CONST_INT
7843 && INTVAL (bl->initial_value) > 0
7844 && (INTVAL (bl->initial_value)
7845 % (-INTVAL (bl->biv->add_val))) == 0)
7847 /* register always nonnegative, add REG_NOTE to branch */
7848 if (! find_reg_note (jump, REG_NONNEG, NULL_RTX))
7850 = gen_rtx_EXPR_LIST (REG_NONNEG, bl->biv->dest_reg,
7857 /* If the decrement is 1 and the value was tested as >= 0 before
7858 the loop, then we can safely optimize. */
7859 for (p = loop_start; p; p = PREV_INSN (p))
7861 if (GET_CODE (p) == CODE_LABEL)
7863 if (GET_CODE (p) != JUMP_INSN)
7866 before_comparison = get_condition_for_loop (loop, p);
7867 if (before_comparison
7868 && XEXP (before_comparison, 0) == bl->biv->dest_reg
7869 && GET_CODE (before_comparison) == LT
7870 && XEXP (before_comparison, 1) == const0_rtx
7871 && ! reg_set_between_p (bl->biv->dest_reg, p, loop_start)
7872 && INTVAL (bl->biv->add_val) == -1)
7874 if (! find_reg_note (jump, REG_NONNEG, NULL_RTX))
7876 = gen_rtx_EXPR_LIST (REG_NONNEG, bl->biv->dest_reg,
7884 else if (GET_CODE (bl->biv->add_val) == CONST_INT
7885 && INTVAL (bl->biv->add_val) > 0)
7887 /* Try to change inc to dec, so can apply above optimization. */
7889 all registers modified are induction variables or invariant,
7890 all memory references have non-overlapping addresses
7891 (obviously true if only one write)
7892 allow 2 insns for the compare/jump at the end of the loop. */
7893 /* Also, we must avoid any instructions which use both the reversed
7894 biv and another biv. Such instructions will fail if the loop is
7895 reversed. We meet this condition by requiring that either
7896 no_use_except_counting is true, or else that there is only
7898 int num_nonfixed_reads = 0;
7899 /* 1 if the iteration var is used only to count iterations. */
7900 int no_use_except_counting = 0;
7901 /* 1 if the loop has no memory store, or it has a single memory store
7902 which is reversible. */
7903 int reversible_mem_store = 1;
7905 if (bl->giv_count == 0 && ! loop->exit_count)
7907 rtx bivreg = regno_reg_rtx[bl->regno];
7909 /* If there are no givs for this biv, and the only exit is the
7910 fall through at the end of the loop, then
7911 see if perhaps there are no uses except to count. */
7912 no_use_except_counting = 1;
7913 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
7916 rtx set = single_set (p);
7918 if (set && GET_CODE (SET_DEST (set)) == REG
7919 && REGNO (SET_DEST (set)) == bl->regno)
7920 /* An insn that sets the biv is okay. */
7922 else if ((p == prev_nonnote_insn (prev_nonnote_insn (loop_end))
7923 || p == prev_nonnote_insn (loop_end))
7924 && reg_mentioned_p (bivreg, PATTERN (p)))
7926 /* If either of these insns uses the biv and sets a pseudo
7927 that has more than one usage, then the biv has uses
7928 other than counting since it's used to derive a value
7929 that is used more than one time. */
7930 int note_set_pseudo_multiple_uses_retval = 0;
7931 note_stores (PATTERN (p), note_set_pseudo_multiple_uses,
7932 ¬e_set_pseudo_multiple_uses_retval);
7933 if (note_set_pseudo_multiple_uses_retval)
7935 no_use_except_counting = 0;
7939 else if (reg_mentioned_p (bivreg, PATTERN (p)))
7941 no_use_except_counting = 0;
7947 if (no_use_except_counting)
7948 ; /* no need to worry about MEMs. */
7949 else if (num_mem_sets <= 1)
7951 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
7953 num_nonfixed_reads += count_nonfixed_reads (loop, PATTERN (p));
7955 /* If the loop has a single store, and the destination address is
7956 invariant, then we can't reverse the loop, because this address
7957 might then have the wrong value at loop exit.
7958 This would work if the source was invariant also, however, in that
7959 case, the insn should have been moved out of the loop. */
7961 if (num_mem_sets == 1)
7963 struct induction *v;
7965 reversible_mem_store
7966 = (! unknown_address_altered
7967 && ! unknown_constant_address_altered
7968 && ! loop_invariant_p (loop,
7969 XEXP (XEXP (loop_store_mems, 0),
7972 /* If the store depends on a register that is set after the
7973 store, it depends on the initial value, and is thus not
7975 for (v = bl->giv; reversible_mem_store && v; v = v->next_iv)
7977 if (v->giv_type == DEST_REG
7978 && reg_mentioned_p (v->dest_reg,
7979 PATTERN (first_loop_store_insn))
7980 && loop_insn_first_p (first_loop_store_insn, v->insn))
7981 reversible_mem_store = 0;
7988 /* This code only acts for innermost loops. Also it simplifies
7989 the memory address check by only reversing loops with
7990 zero or one memory access.
7991 Two memory accesses could involve parts of the same array,
7992 and that can't be reversed.
7993 If the biv is used only for counting, than we don't need to worry
7994 about all these things. */
7996 if ((num_nonfixed_reads <= 1
7997 && ! loop_info->has_call
7998 && ! loop_info->has_volatile
7999 && reversible_mem_store
8000 && (bl->giv_count + bl->biv_count + num_mem_sets
8001 + num_movables + compare_and_branch == insn_count)
8002 && (bl == loop_iv_list && bl->next == 0))
8003 || no_use_except_counting)
8007 /* Loop can be reversed. */
8008 if (loop_dump_stream)
8009 fprintf (loop_dump_stream, "Can reverse loop\n");
8011 /* Now check other conditions:
8013 The increment must be a constant, as must the initial value,
8014 and the comparison code must be LT.
8016 This test can probably be improved since +/- 1 in the constant
8017 can be obtained by changing LT to LE and vice versa; this is
8021 /* for constants, LE gets turned into LT */
8022 && (GET_CODE (comparison) == LT
8023 || (GET_CODE (comparison) == LE
8024 && no_use_except_counting)))
8026 HOST_WIDE_INT add_val, add_adjust, comparison_val = 0;
8027 rtx initial_value, comparison_value;
8029 enum rtx_code cmp_code;
8030 int comparison_const_width;
8031 unsigned HOST_WIDE_INT comparison_sign_mask;
8033 add_val = INTVAL (bl->biv->add_val);
8034 comparison_value = XEXP (comparison, 1);
8035 if (GET_MODE (comparison_value) == VOIDmode)
8036 comparison_const_width
8037 = GET_MODE_BITSIZE (GET_MODE (XEXP (comparison, 0)));
8039 comparison_const_width
8040 = GET_MODE_BITSIZE (GET_MODE (comparison_value));
8041 if (comparison_const_width > HOST_BITS_PER_WIDE_INT)
8042 comparison_const_width = HOST_BITS_PER_WIDE_INT;
8043 comparison_sign_mask
8044 = (unsigned HOST_WIDE_INT)1 << (comparison_const_width - 1);
8046 /* If the comparison value is not a loop invariant, then we
8047 can not reverse this loop.
8049 ??? If the insns which initialize the comparison value as
8050 a whole compute an invariant result, then we could move
8051 them out of the loop and proceed with loop reversal. */
8052 if (! loop_invariant_p (loop, comparison_value))
8055 if (GET_CODE (comparison_value) == CONST_INT)
8056 comparison_val = INTVAL (comparison_value);
8057 initial_value = bl->initial_value;
8059 /* Normalize the initial value if it is an integer and
8060 has no other use except as a counter. This will allow
8061 a few more loops to be reversed. */
8062 if (no_use_except_counting
8063 && GET_CODE (comparison_value) == CONST_INT
8064 && GET_CODE (initial_value) == CONST_INT)
8066 comparison_val = comparison_val - INTVAL (bl->initial_value);
8067 /* The code below requires comparison_val to be a multiple
8068 of add_val in order to do the loop reversal, so
8069 round up comparison_val to a multiple of add_val.
8070 Since comparison_value is constant, we know that the
8071 current comparison code is LT. */
8072 comparison_val = comparison_val + add_val - 1;
8074 -= (unsigned HOST_WIDE_INT) comparison_val % add_val;
8075 /* We postpone overflow checks for COMPARISON_VAL here;
8076 even if there is an overflow, we might still be able to
8077 reverse the loop, if converting the loop exit test to
8079 initial_value = const0_rtx;
8082 /* First check if we can do a vanilla loop reversal. */
8083 if (initial_value == const0_rtx
8084 /* If we have a decrement_and_branch_on_count,
8085 prefer the NE test, since this will allow that
8086 instruction to be generated. Note that we must
8087 use a vanilla loop reversal if the biv is used to
8088 calculate a giv or has a non-counting use. */
8089 #if ! defined (HAVE_decrement_and_branch_until_zero) \
8090 && defined (HAVE_decrement_and_branch_on_count)
8091 && (! (add_val == 1 && loop->vtop
8092 && (bl->biv_count == 0
8093 || no_use_except_counting)))
8095 && GET_CODE (comparison_value) == CONST_INT
8096 /* Now do postponed overflow checks on COMPARISON_VAL. */
8097 && ! (((comparison_val - add_val) ^ INTVAL (comparison_value))
8098 & comparison_sign_mask))
8100 /* Register will always be nonnegative, with value
8101 0 on last iteration */
8102 add_adjust = add_val;
8106 else if (add_val == 1 && loop->vtop
8107 && (bl->biv_count == 0
8108 || no_use_except_counting))
8116 if (GET_CODE (comparison) == LE)
8117 add_adjust -= add_val;
8119 /* If the initial value is not zero, or if the comparison
8120 value is not an exact multiple of the increment, then we
8121 can not reverse this loop. */
8122 if (initial_value == const0_rtx
8123 && GET_CODE (comparison_value) == CONST_INT)
8125 if (((unsigned HOST_WIDE_INT) comparison_val % add_val) != 0)
8130 if (! no_use_except_counting || add_val != 1)
8134 final_value = comparison_value;
8136 /* Reset these in case we normalized the initial value
8137 and comparison value above. */
8138 if (GET_CODE (comparison_value) == CONST_INT
8139 && GET_CODE (initial_value) == CONST_INT)
8141 comparison_value = GEN_INT (comparison_val);
8143 = GEN_INT (comparison_val + INTVAL (bl->initial_value));
8145 bl->initial_value = initial_value;
8147 /* Save some info needed to produce the new insns. */
8148 reg = bl->biv->dest_reg;
8149 jump_label = XEXP (SET_SRC (PATTERN (PREV_INSN (loop_end))), 1);
8150 if (jump_label == pc_rtx)
8151 jump_label = XEXP (SET_SRC (PATTERN (PREV_INSN (loop_end))), 2);
8152 new_add_val = GEN_INT (- INTVAL (bl->biv->add_val));
8154 /* Set start_value; if this is not a CONST_INT, we need
8156 Initialize biv to start_value before loop start.
8157 The old initializing insn will be deleted as a
8158 dead store by flow.c. */
8159 if (initial_value == const0_rtx
8160 && GET_CODE (comparison_value) == CONST_INT)
8162 start_value = GEN_INT (comparison_val - add_adjust);
8163 emit_insn_before (gen_move_insn (reg, start_value),
8166 else if (GET_CODE (initial_value) == CONST_INT)
8168 rtx offset = GEN_INT (-INTVAL (initial_value) - add_adjust);
8169 enum machine_mode mode = GET_MODE (reg);
8170 enum insn_code icode
8171 = add_optab->handlers[(int) mode].insn_code;
8173 if (! (*insn_data[icode].operand[0].predicate) (reg, mode)
8174 || ! ((*insn_data[icode].operand[1].predicate)
8175 (comparison_value, mode))
8176 || ! ((*insn_data[icode].operand[2].predicate)
8180 = gen_rtx_PLUS (mode, comparison_value, offset);
8181 emit_insn_before ((GEN_FCN (icode)
8182 (reg, comparison_value, offset)),
8184 if (GET_CODE (comparison) == LE)
8185 final_value = gen_rtx_PLUS (mode, comparison_value,
8188 else if (! add_adjust)
8190 enum machine_mode mode = GET_MODE (reg);
8191 enum insn_code icode
8192 = sub_optab->handlers[(int) mode].insn_code;
8193 if (! (*insn_data[icode].operand[0].predicate) (reg, mode)
8194 || ! ((*insn_data[icode].operand[1].predicate)
8195 (comparison_value, mode))
8196 || ! ((*insn_data[icode].operand[2].predicate)
8197 (initial_value, mode)))
8200 = gen_rtx_MINUS (mode, comparison_value, initial_value);
8201 emit_insn_before ((GEN_FCN (icode)
8202 (reg, comparison_value, initial_value)),
8206 /* We could handle the other cases too, but it'll be
8207 better to have a testcase first. */
8210 /* We may not have a single insn which can increment a reg, so
8211 create a sequence to hold all the insns from expand_inc. */
8213 expand_inc (reg, new_add_val);
8214 tem = gen_sequence ();
8217 p = emit_insn_before (tem, bl->biv->insn);
8218 delete_insn (bl->biv->insn);
8220 /* Update biv info to reflect its new status. */
8222 bl->initial_value = start_value;
8223 bl->biv->add_val = new_add_val;
8225 /* Update loop info. */
8226 loop_info->initial_value = reg;
8227 loop_info->initial_equiv_value = reg;
8228 loop_info->final_value = const0_rtx;
8229 loop_info->final_equiv_value = const0_rtx;
8230 loop_info->comparison_value = const0_rtx;
8231 loop_info->comparison_code = cmp_code;
8232 loop_info->increment = new_add_val;
8234 /* Inc LABEL_NUSES so that delete_insn will
8235 not delete the label. */
8236 LABEL_NUSES (XEXP (jump_label, 0)) ++;
8238 /* Emit an insn after the end of the loop to set the biv's
8239 proper exit value if it is used anywhere outside the loop. */
8240 if ((REGNO_LAST_UID (bl->regno) != INSN_UID (first_compare))
8242 || REGNO_FIRST_UID (bl->regno) != INSN_UID (bl->init_insn))
8243 emit_insn_after (gen_move_insn (reg, final_value),
8246 /* Delete compare/branch at end of loop. */
8247 delete_insn (PREV_INSN (loop_end));
8248 if (compare_and_branch == 2)
8249 delete_insn (first_compare);
8251 /* Add new compare/branch insn at end of loop. */
8253 emit_cmp_and_jump_insns (reg, const0_rtx, cmp_code, NULL_RTX,
8254 GET_MODE (reg), 0, 0,
8255 XEXP (jump_label, 0));
8256 tem = gen_sequence ();
8258 emit_jump_insn_before (tem, loop_end);
8260 for (tem = PREV_INSN (loop_end);
8261 tem && GET_CODE (tem) != JUMP_INSN;
8262 tem = PREV_INSN (tem))
8266 JUMP_LABEL (tem) = XEXP (jump_label, 0);
8272 /* Increment of LABEL_NUSES done above. */
8273 /* Register is now always nonnegative,
8274 so add REG_NONNEG note to the branch. */
8275 REG_NOTES (tem) = gen_rtx_EXPR_LIST (REG_NONNEG, reg,
8281 /* No insn may reference both the reversed and another biv or it
8282 will fail (see comment near the top of the loop reversal
8284 Earlier on, we have verified that the biv has no use except
8285 counting, or it is the only biv in this function.
8286 However, the code that computes no_use_except_counting does
8287 not verify reg notes. It's possible to have an insn that
8288 references another biv, and has a REG_EQUAL note with an
8289 expression based on the reversed biv. To avoid this case,
8290 remove all REG_EQUAL notes based on the reversed biv
8292 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
8296 rtx set = single_set (p);
8297 /* If this is a set of a GIV based on the reversed biv, any
8298 REG_EQUAL notes should still be correct. */
8300 || GET_CODE (SET_DEST (set)) != REG
8301 || (size_t) REGNO (SET_DEST (set)) >= reg_iv_type->num_elements
8302 || REG_IV_TYPE (REGNO (SET_DEST (set))) != GENERAL_INDUCT
8303 || REG_IV_INFO (REGNO (SET_DEST (set)))->src_reg != bl->biv->src_reg)
8304 for (pnote = ®_NOTES (p); *pnote;)
8306 if (REG_NOTE_KIND (*pnote) == REG_EQUAL
8307 && reg_mentioned_p (regno_reg_rtx[bl->regno],
8309 *pnote = XEXP (*pnote, 1);
8311 pnote = &XEXP (*pnote, 1);
8315 /* Mark that this biv has been reversed. Each giv which depends
8316 on this biv, and which is also live past the end of the loop
8317 will have to be fixed up. */
8321 if (loop_dump_stream)
8323 fprintf (loop_dump_stream, "Reversed loop");
8325 fprintf (loop_dump_stream, " and added reg_nonneg\n");
8327 fprintf (loop_dump_stream, "\n");
8338 /* Verify whether the biv BL appears to be eliminable,
8339 based on the insns in the loop that refer to it.
8341 If ELIMINATE_P is non-zero, actually do the elimination.
8343 THRESHOLD and INSN_COUNT are from loop_optimize and are used to
8344 determine whether invariant insns should be placed inside or at the
8345 start of the loop. */
8348 maybe_eliminate_biv (loop, bl, eliminate_p, threshold, insn_count)
8349 const struct loop *loop;
8350 struct iv_class *bl;
8352 int threshold, insn_count;
8354 rtx reg = bl->biv->dest_reg;
8355 rtx loop_start = loop->start;
8356 rtx loop_end = loop->end;
8359 /* Scan all insns in the loop, stopping if we find one that uses the
8360 biv in a way that we cannot eliminate. */
8362 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
8364 enum rtx_code code = GET_CODE (p);
8365 rtx where = threshold >= insn_count ? loop_start : p;
8367 /* If this is a libcall that sets a giv, skip ahead to its end. */
8368 if (GET_RTX_CLASS (code) == 'i')
8370 rtx note = find_reg_note (p, REG_LIBCALL, NULL_RTX);
8374 rtx last = XEXP (note, 0);
8375 rtx set = single_set (last);
8377 if (set && GET_CODE (SET_DEST (set)) == REG)
8379 unsigned int regno = REGNO (SET_DEST (set));
8381 if (regno < max_reg_before_loop
8382 && REG_IV_TYPE (regno) == GENERAL_INDUCT
8383 && REG_IV_INFO (regno)->src_reg == bl->biv->src_reg)
8388 if ((code == INSN || code == JUMP_INSN || code == CALL_INSN)
8389 && reg_mentioned_p (reg, PATTERN (p))
8390 && ! maybe_eliminate_biv_1 (loop, PATTERN (p), p, bl,
8391 eliminate_p, where))
8393 if (loop_dump_stream)
8394 fprintf (loop_dump_stream,
8395 "Cannot eliminate biv %d: biv used in insn %d.\n",
8396 bl->regno, INSN_UID (p));
8403 if (loop_dump_stream)
8404 fprintf (loop_dump_stream, "biv %d %s eliminated.\n",
8405 bl->regno, eliminate_p ? "was" : "can be");
8412 /* INSN and REFERENCE are instructions in the same insn chain.
8413 Return non-zero if INSN is first. */
8416 loop_insn_first_p (insn, reference)
8417 rtx insn, reference;
8421 for (p = insn, q = reference; ;)
8423 /* Start with test for not first so that INSN == REFERENCE yields not
8425 if (q == insn || ! p)
8427 if (p == reference || ! q)
8430 /* Either of P or Q might be a NOTE. Notes have the same LUID as the
8431 previous insn, hence the <= comparison below does not work if
8433 if (INSN_UID (p) < max_uid_for_loop
8434 && INSN_UID (q) < max_uid_for_loop
8435 && GET_CODE (p) != NOTE)
8436 return INSN_LUID (p) <= INSN_LUID (q);
8438 if (INSN_UID (p) >= max_uid_for_loop
8439 || GET_CODE (p) == NOTE)
8441 if (INSN_UID (q) >= max_uid_for_loop)
8446 /* We are trying to eliminate BIV in INSN using GIV. Return non-zero if
8447 the offset that we have to take into account due to auto-increment /
8448 div derivation is zero. */
8450 biv_elimination_giv_has_0_offset (biv, giv, insn)
8451 struct induction *biv, *giv;
8454 /* If the giv V had the auto-inc address optimization applied
8455 to it, and INSN occurs between the giv insn and the biv
8456 insn, then we'd have to adjust the value used here.
8457 This is rare, so we don't bother to make this possible. */
8458 if (giv->auto_inc_opt
8459 && ((loop_insn_first_p (giv->insn, insn)
8460 && loop_insn_first_p (insn, biv->insn))
8461 || (loop_insn_first_p (biv->insn, insn)
8462 && loop_insn_first_p (insn, giv->insn))))
8465 /* If the giv V was derived from another giv, and INSN does
8466 not occur between the giv insn and the biv insn, then we'd
8467 have to adjust the value used here. This is rare, so we don't
8468 bother to make this possible. */
8469 if (giv->derived_from
8470 && ! (giv->always_executed
8471 && loop_insn_first_p (giv->insn, insn)
8472 && loop_insn_first_p (insn, biv->insn)))
8475 && giv->same->derived_from
8476 && ! (giv->same->always_executed
8477 && loop_insn_first_p (giv->same->insn, insn)
8478 && loop_insn_first_p (insn, biv->insn)))
8484 /* If BL appears in X (part of the pattern of INSN), see if we can
8485 eliminate its use. If so, return 1. If not, return 0.
8487 If BIV does not appear in X, return 1.
8489 If ELIMINATE_P is non-zero, actually do the elimination. WHERE indicates
8490 where extra insns should be added. Depending on how many items have been
8491 moved out of the loop, it will either be before INSN or at the start of
8495 maybe_eliminate_biv_1 (loop, x, insn, bl, eliminate_p, where)
8496 const struct loop *loop;
8498 struct iv_class *bl;
8502 enum rtx_code code = GET_CODE (x);
8503 rtx reg = bl->biv->dest_reg;
8504 enum machine_mode mode = GET_MODE (reg);
8505 struct induction *v;
8517 /* If we haven't already been able to do something with this BIV,
8518 we can't eliminate it. */
8524 /* If this sets the BIV, it is not a problem. */
8525 if (SET_DEST (x) == reg)
8528 /* If this is an insn that defines a giv, it is also ok because
8529 it will go away when the giv is reduced. */
8530 for (v = bl->giv; v; v = v->next_iv)
8531 if (v->giv_type == DEST_REG && SET_DEST (x) == v->dest_reg)
8535 if (SET_DEST (x) == cc0_rtx && SET_SRC (x) == reg)
8537 /* Can replace with any giv that was reduced and
8538 that has (MULT_VAL != 0) and (ADD_VAL == 0).
8539 Require a constant for MULT_VAL, so we know it's nonzero.
8540 ??? We disable this optimization to avoid potential
8543 for (v = bl->giv; v; v = v->next_iv)
8544 if (GET_CODE (v->mult_val) == CONST_INT && v->mult_val != const0_rtx
8545 && v->add_val == const0_rtx
8546 && ! v->ignore && ! v->maybe_dead && v->always_computable
8550 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8556 /* If the giv has the opposite direction of change,
8557 then reverse the comparison. */
8558 if (INTVAL (v->mult_val) < 0)
8559 new = gen_rtx_COMPARE (GET_MODE (v->new_reg),
8560 const0_rtx, v->new_reg);
8564 /* We can probably test that giv's reduced reg. */
8565 if (validate_change (insn, &SET_SRC (x), new, 0))
8569 /* Look for a giv with (MULT_VAL != 0) and (ADD_VAL != 0);
8570 replace test insn with a compare insn (cmp REDUCED_GIV ADD_VAL).
8571 Require a constant for MULT_VAL, so we know it's nonzero.
8572 ??? Do this only if ADD_VAL is a pointer to avoid a potential
8573 overflow problem. */
8575 for (v = bl->giv; v; v = v->next_iv)
8576 if (GET_CODE (v->mult_val) == CONST_INT && v->mult_val != const0_rtx
8577 && ! v->ignore && ! v->maybe_dead && v->always_computable
8579 && (GET_CODE (v->add_val) == SYMBOL_REF
8580 || GET_CODE (v->add_val) == LABEL_REF
8581 || GET_CODE (v->add_val) == CONST
8582 || (GET_CODE (v->add_val) == REG
8583 && REGNO_POINTER_FLAG (REGNO (v->add_val)))))
8585 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8591 /* If the giv has the opposite direction of change,
8592 then reverse the comparison. */
8593 if (INTVAL (v->mult_val) < 0)
8594 new = gen_rtx_COMPARE (VOIDmode, copy_rtx (v->add_val),
8597 new = gen_rtx_COMPARE (VOIDmode, v->new_reg,
8598 copy_rtx (v->add_val));
8600 /* Replace biv with the giv's reduced register. */
8601 update_reg_last_use (v->add_val, insn);
8602 if (validate_change (insn, &SET_SRC (PATTERN (insn)), new, 0))
8605 /* Insn doesn't support that constant or invariant. Copy it
8606 into a register (it will be a loop invariant.) */
8607 tem = gen_reg_rtx (GET_MODE (v->new_reg));
8609 emit_insn_before (gen_move_insn (tem, copy_rtx (v->add_val)),
8612 /* Substitute the new register for its invariant value in
8613 the compare expression. */
8614 XEXP (new, (INTVAL (v->mult_val) < 0) ? 0 : 1) = tem;
8615 if (validate_change (insn, &SET_SRC (PATTERN (insn)), new, 0))
8624 case GT: case GE: case GTU: case GEU:
8625 case LT: case LE: case LTU: case LEU:
8626 /* See if either argument is the biv. */
8627 if (XEXP (x, 0) == reg)
8628 arg = XEXP (x, 1), arg_operand = 1;
8629 else if (XEXP (x, 1) == reg)
8630 arg = XEXP (x, 0), arg_operand = 0;
8634 if (CONSTANT_P (arg))
8636 /* First try to replace with any giv that has constant positive
8637 mult_val and constant add_val. We might be able to support
8638 negative mult_val, but it seems complex to do it in general. */
8640 for (v = bl->giv; v; v = v->next_iv)
8641 if (GET_CODE (v->mult_val) == CONST_INT && INTVAL (v->mult_val) > 0
8642 && (GET_CODE (v->add_val) == SYMBOL_REF
8643 || GET_CODE (v->add_val) == LABEL_REF
8644 || GET_CODE (v->add_val) == CONST
8645 || (GET_CODE (v->add_val) == REG
8646 && REGNO_POINTER_FLAG (REGNO (v->add_val))))
8647 && ! v->ignore && ! v->maybe_dead && v->always_computable
8650 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8656 /* Replace biv with the giv's reduced reg. */
8657 validate_change (insn, &XEXP (x, 1-arg_operand), v->new_reg, 1);
8659 /* If all constants are actually constant integers and
8660 the derived constant can be directly placed in the COMPARE,
8662 if (GET_CODE (arg) == CONST_INT
8663 && GET_CODE (v->mult_val) == CONST_INT
8664 && GET_CODE (v->add_val) == CONST_INT)
8666 validate_change (insn, &XEXP (x, arg_operand),
8667 GEN_INT (INTVAL (arg)
8668 * INTVAL (v->mult_val)
8669 + INTVAL (v->add_val)), 1);
8673 /* Otherwise, load it into a register. */
8674 tem = gen_reg_rtx (mode);
8675 emit_iv_add_mult (arg, v->mult_val, v->add_val, tem, where);
8676 validate_change (insn, &XEXP (x, arg_operand), tem, 1);
8678 if (apply_change_group ())
8682 /* Look for giv with positive constant mult_val and nonconst add_val.
8683 Insert insns to calculate new compare value.
8684 ??? Turn this off due to possible overflow. */
8686 for (v = bl->giv; v; v = v->next_iv)
8687 if (GET_CODE (v->mult_val) == CONST_INT && INTVAL (v->mult_val) > 0
8688 && ! v->ignore && ! v->maybe_dead && v->always_computable
8694 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8700 tem = gen_reg_rtx (mode);
8702 /* Replace biv with giv's reduced register. */
8703 validate_change (insn, &XEXP (x, 1 - arg_operand),
8706 /* Compute value to compare against. */
8707 emit_iv_add_mult (arg, v->mult_val, v->add_val, tem, where);
8708 /* Use it in this insn. */
8709 validate_change (insn, &XEXP (x, arg_operand), tem, 1);
8710 if (apply_change_group ())
8714 else if (GET_CODE (arg) == REG || GET_CODE (arg) == MEM)
8716 if (loop_invariant_p (loop, arg) == 1)
8718 /* Look for giv with constant positive mult_val and nonconst
8719 add_val. Insert insns to compute new compare value.
8720 ??? Turn this off due to possible overflow. */
8722 for (v = bl->giv; v; v = v->next_iv)
8723 if (GET_CODE (v->mult_val) == CONST_INT && INTVAL (v->mult_val) > 0
8724 && ! v->ignore && ! v->maybe_dead && v->always_computable
8730 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8736 tem = gen_reg_rtx (mode);
8738 /* Replace biv with giv's reduced register. */
8739 validate_change (insn, &XEXP (x, 1 - arg_operand),
8742 /* Compute value to compare against. */
8743 emit_iv_add_mult (arg, v->mult_val, v->add_val,
8745 validate_change (insn, &XEXP (x, arg_operand), tem, 1);
8746 if (apply_change_group ())
8751 /* This code has problems. Basically, you can't know when
8752 seeing if we will eliminate BL, whether a particular giv
8753 of ARG will be reduced. If it isn't going to be reduced,
8754 we can't eliminate BL. We can try forcing it to be reduced,
8755 but that can generate poor code.
8757 The problem is that the benefit of reducing TV, below should
8758 be increased if BL can actually be eliminated, but this means
8759 we might have to do a topological sort of the order in which
8760 we try to process biv. It doesn't seem worthwhile to do
8761 this sort of thing now. */
8764 /* Otherwise the reg compared with had better be a biv. */
8765 if (GET_CODE (arg) != REG
8766 || REG_IV_TYPE (REGNO (arg)) != BASIC_INDUCT)
8769 /* Look for a pair of givs, one for each biv,
8770 with identical coefficients. */
8771 for (v = bl->giv; v; v = v->next_iv)
8773 struct induction *tv;
8775 if (v->ignore || v->maybe_dead || v->mode != mode)
8778 for (tv = reg_biv_class[REGNO (arg)]->giv; tv; tv = tv->next_iv)
8779 if (! tv->ignore && ! tv->maybe_dead
8780 && rtx_equal_p (tv->mult_val, v->mult_val)
8781 && rtx_equal_p (tv->add_val, v->add_val)
8782 && tv->mode == mode)
8784 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8790 /* Replace biv with its giv's reduced reg. */
8791 XEXP (x, 1-arg_operand) = v->new_reg;
8792 /* Replace other operand with the other giv's
8794 XEXP (x, arg_operand) = tv->new_reg;
8801 /* If we get here, the biv can't be eliminated. */
8805 /* If this address is a DEST_ADDR giv, it doesn't matter if the
8806 biv is used in it, since it will be replaced. */
8807 for (v = bl->giv; v; v = v->next_iv)
8808 if (v->giv_type == DEST_ADDR && v->location == &XEXP (x, 0))
8816 /* See if any subexpression fails elimination. */
8817 fmt = GET_RTX_FORMAT (code);
8818 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
8823 if (! maybe_eliminate_biv_1 (loop, XEXP (x, i), insn, bl,
8824 eliminate_p, where))
8829 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
8830 if (! maybe_eliminate_biv_1 (loop, XVECEXP (x, i, j), insn, bl,
8831 eliminate_p, where))
8840 /* Return nonzero if the last use of REG
8841 is in an insn following INSN in the same basic block. */
8844 last_use_this_basic_block (reg, insn)
8850 n && GET_CODE (n) != CODE_LABEL && GET_CODE (n) != JUMP_INSN;
8853 if (REGNO_LAST_UID (REGNO (reg)) == INSN_UID (n))
8859 /* Called via `note_stores' to record the initial value of a biv. Here we
8860 just record the location of the set and process it later. */
8863 record_initial (dest, set, data)
8866 void *data ATTRIBUTE_UNUSED;
8868 struct iv_class *bl;
8870 if (GET_CODE (dest) != REG
8871 || REGNO (dest) >= max_reg_before_loop
8872 || REG_IV_TYPE (REGNO (dest)) != BASIC_INDUCT)
8875 bl = reg_biv_class[REGNO (dest)];
8877 /* If this is the first set found, record it. */
8878 if (bl->init_insn == 0)
8880 bl->init_insn = note_insn;
8885 /* If any of the registers in X are "old" and currently have a last use earlier
8886 than INSN, update them to have a last use of INSN. Their actual last use
8887 will be the previous insn but it will not have a valid uid_luid so we can't
8891 update_reg_last_use (x, insn)
8895 /* Check for the case where INSN does not have a valid luid. In this case,
8896 there is no need to modify the regno_last_uid, as this can only happen
8897 when code is inserted after the loop_end to set a pseudo's final value,
8898 and hence this insn will never be the last use of x. */
8899 if (GET_CODE (x) == REG && REGNO (x) < max_reg_before_loop
8900 && INSN_UID (insn) < max_uid_for_loop
8901 && uid_luid[REGNO_LAST_UID (REGNO (x))] < uid_luid[INSN_UID (insn)])
8902 REGNO_LAST_UID (REGNO (x)) = INSN_UID (insn);
8906 register const char *fmt = GET_RTX_FORMAT (GET_CODE (x));
8907 for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
8910 update_reg_last_use (XEXP (x, i), insn);
8911 else if (fmt[i] == 'E')
8912 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
8913 update_reg_last_use (XVECEXP (x, i, j), insn);
8918 /* Given an insn INSN and condition COND, return the condition in a
8919 canonical form to simplify testing by callers. Specifically:
8921 (1) The code will always be a comparison operation (EQ, NE, GT, etc.).
8922 (2) Both operands will be machine operands; (cc0) will have been replaced.
8923 (3) If an operand is a constant, it will be the second operand.
8924 (4) (LE x const) will be replaced with (LT x <const+1>) and similarly
8925 for GE, GEU, and LEU.
8927 If the condition cannot be understood, or is an inequality floating-point
8928 comparison which needs to be reversed, 0 will be returned.
8930 If REVERSE is non-zero, then reverse the condition prior to canonizing it.
8932 If EARLIEST is non-zero, it is a pointer to a place where the earliest
8933 insn used in locating the condition was found. If a replacement test
8934 of the condition is desired, it should be placed in front of that
8935 insn and we will be sure that the inputs are still valid.
8937 If WANT_REG is non-zero, we wish the condition to be relative to that
8938 register, if possible. Therefore, do not canonicalize the condition
8942 canonicalize_condition (insn, cond, reverse, earliest, want_reg)
8954 int reverse_code = 0;
8955 int did_reverse_condition = 0;
8956 enum machine_mode mode;
8958 code = GET_CODE (cond);
8959 mode = GET_MODE (cond);
8960 op0 = XEXP (cond, 0);
8961 op1 = XEXP (cond, 1);
8965 code = reverse_condition (code);
8966 did_reverse_condition ^= 1;
8972 /* If we are comparing a register with zero, see if the register is set
8973 in the previous insn to a COMPARE or a comparison operation. Perform
8974 the same tests as a function of STORE_FLAG_VALUE as find_comparison_args
8977 while (GET_RTX_CLASS (code) == '<'
8978 && op1 == CONST0_RTX (GET_MODE (op0))
8981 /* Set non-zero when we find something of interest. */
8985 /* If comparison with cc0, import actual comparison from compare
8989 if ((prev = prev_nonnote_insn (prev)) == 0
8990 || GET_CODE (prev) != INSN
8991 || (set = single_set (prev)) == 0
8992 || SET_DEST (set) != cc0_rtx)
8995 op0 = SET_SRC (set);
8996 op1 = CONST0_RTX (GET_MODE (op0));
9002 /* If this is a COMPARE, pick up the two things being compared. */
9003 if (GET_CODE (op0) == COMPARE)
9005 op1 = XEXP (op0, 1);
9006 op0 = XEXP (op0, 0);
9009 else if (GET_CODE (op0) != REG)
9012 /* Go back to the previous insn. Stop if it is not an INSN. We also
9013 stop if it isn't a single set or if it has a REG_INC note because
9014 we don't want to bother dealing with it. */
9016 if ((prev = prev_nonnote_insn (prev)) == 0
9017 || GET_CODE (prev) != INSN
9018 || FIND_REG_INC_NOTE (prev, 0)
9019 || (set = single_set (prev)) == 0)
9022 /* If this is setting OP0, get what it sets it to if it looks
9024 if (rtx_equal_p (SET_DEST (set), op0))
9026 enum machine_mode inner_mode = GET_MODE (SET_DEST (set));
9028 /* ??? We may not combine comparisons done in a CCmode with
9029 comparisons not done in a CCmode. This is to aid targets
9030 like Alpha that have an IEEE compliant EQ instruction, and
9031 a non-IEEE compliant BEQ instruction. The use of CCmode is
9032 actually artificial, simply to prevent the combination, but
9033 should not affect other platforms.
9035 However, we must allow VOIDmode comparisons to match either
9036 CCmode or non-CCmode comparison, because some ports have
9037 modeless comparisons inside branch patterns.
9039 ??? This mode check should perhaps look more like the mode check
9040 in simplify_comparison in combine. */
9042 if ((GET_CODE (SET_SRC (set)) == COMPARE
9045 && GET_MODE_CLASS (inner_mode) == MODE_INT
9046 && (GET_MODE_BITSIZE (inner_mode)
9047 <= HOST_BITS_PER_WIDE_INT)
9048 && (STORE_FLAG_VALUE
9049 & ((HOST_WIDE_INT) 1
9050 << (GET_MODE_BITSIZE (inner_mode) - 1))))
9051 #ifdef FLOAT_STORE_FLAG_VALUE
9053 && GET_MODE_CLASS (inner_mode) == MODE_FLOAT
9054 && (REAL_VALUE_NEGATIVE
9055 (FLOAT_STORE_FLAG_VALUE (inner_mode))))
9058 && GET_RTX_CLASS (GET_CODE (SET_SRC (set))) == '<'))
9059 && (((GET_MODE_CLASS (mode) == MODE_CC)
9060 == (GET_MODE_CLASS (inner_mode) == MODE_CC))
9061 || mode == VOIDmode || inner_mode == VOIDmode))
9063 else if (((code == EQ
9065 && (GET_MODE_BITSIZE (inner_mode)
9066 <= HOST_BITS_PER_WIDE_INT)
9067 && GET_MODE_CLASS (inner_mode) == MODE_INT
9068 && (STORE_FLAG_VALUE
9069 & ((HOST_WIDE_INT) 1
9070 << (GET_MODE_BITSIZE (inner_mode) - 1))))
9071 #ifdef FLOAT_STORE_FLAG_VALUE
9073 && GET_MODE_CLASS (inner_mode) == MODE_FLOAT
9074 && (REAL_VALUE_NEGATIVE
9075 (FLOAT_STORE_FLAG_VALUE (inner_mode))))
9078 && GET_RTX_CLASS (GET_CODE (SET_SRC (set))) == '<'
9079 && (((GET_MODE_CLASS (mode) == MODE_CC)
9080 == (GET_MODE_CLASS (inner_mode) == MODE_CC))
9081 || mode == VOIDmode || inner_mode == VOIDmode))
9084 /* We might have reversed a LT to get a GE here. But this wasn't
9085 actually the comparison of data, so we don't flag that we
9086 have had to reverse the condition. */
9087 did_reverse_condition ^= 1;
9095 else if (reg_set_p (op0, prev))
9096 /* If this sets OP0, but not directly, we have to give up. */
9101 if (GET_RTX_CLASS (GET_CODE (x)) == '<')
9102 code = GET_CODE (x);
9105 code = reverse_condition (code);
9106 if (code == UNKNOWN)
9108 did_reverse_condition ^= 1;
9112 op0 = XEXP (x, 0), op1 = XEXP (x, 1);
9118 /* If constant is first, put it last. */
9119 if (CONSTANT_P (op0))
9120 code = swap_condition (code), tem = op0, op0 = op1, op1 = tem;
9122 /* If OP0 is the result of a comparison, we weren't able to find what
9123 was really being compared, so fail. */
9124 if (GET_MODE_CLASS (GET_MODE (op0)) == MODE_CC)
9127 /* Canonicalize any ordered comparison with integers involving equality
9128 if we can do computations in the relevant mode and we do not
9131 if (GET_CODE (op1) == CONST_INT
9132 && GET_MODE (op0) != VOIDmode
9133 && GET_MODE_BITSIZE (GET_MODE (op0)) <= HOST_BITS_PER_WIDE_INT)
9135 HOST_WIDE_INT const_val = INTVAL (op1);
9136 unsigned HOST_WIDE_INT uconst_val = const_val;
9137 unsigned HOST_WIDE_INT max_val
9138 = (unsigned HOST_WIDE_INT) GET_MODE_MASK (GET_MODE (op0));
9143 if ((unsigned HOST_WIDE_INT) const_val != max_val >> 1)
9144 code = LT, op1 = GEN_INT (const_val + 1);
9147 /* When cross-compiling, const_val might be sign-extended from
9148 BITS_PER_WORD to HOST_BITS_PER_WIDE_INT */
9150 if ((HOST_WIDE_INT) (const_val & max_val)
9151 != (((HOST_WIDE_INT) 1
9152 << (GET_MODE_BITSIZE (GET_MODE (op0)) - 1))))
9153 code = GT, op1 = GEN_INT (const_val - 1);
9157 if (uconst_val < max_val)
9158 code = LTU, op1 = GEN_INT (uconst_val + 1);
9162 if (uconst_val != 0)
9163 code = GTU, op1 = GEN_INT (uconst_val - 1);
9171 /* If this was floating-point and we reversed anything other than an
9172 EQ or NE or (UN)ORDERED, return zero. */
9173 if (TARGET_FLOAT_FORMAT == IEEE_FLOAT_FORMAT
9174 && did_reverse_condition
9175 && code != NE && code != EQ && code != UNORDERED && code != ORDERED
9177 && GET_MODE_CLASS (GET_MODE (op0)) == MODE_FLOAT)
9181 /* Never return CC0; return zero instead. */
9186 return gen_rtx_fmt_ee (code, VOIDmode, op0, op1);
9190 /* Given a jump insn JUMP, return the condition that will cause it to branch
9191 to its JUMP_LABEL. If the condition cannot be understood, or is an
9192 inequality floating-point comparison which needs to be reversed, 0 will
9195 If EARLIEST is non-zero, it is a pointer to a place where the earliest
9196 insn used in locating the condition was found. If a replacement test
9197 of the condition is desired, it should be placed in front of that
9198 insn and we will be sure that the inputs are still valid. */
9201 get_condition (jump, earliest)
9209 /* If this is not a standard conditional jump, we can't parse it. */
9210 if (GET_CODE (jump) != JUMP_INSN
9211 || ! any_condjump_p (jump))
9213 set = pc_set (jump);
9215 cond = XEXP (SET_SRC (set), 0);
9217 /* If this branches to JUMP_LABEL when the condition is false, reverse
9220 = GET_CODE (XEXP (SET_SRC (set), 2)) == LABEL_REF
9221 && XEXP (XEXP (SET_SRC (set), 2), 0) == JUMP_LABEL (jump);
9223 return canonicalize_condition (jump, cond, reverse, earliest, NULL_RTX);
9226 /* Similar to above routine, except that we also put an invariant last
9227 unless both operands are invariants. */
9230 get_condition_for_loop (loop, x)
9231 const struct loop *loop;
9234 rtx comparison = get_condition (x, NULL_PTR);
9237 || ! loop_invariant_p (loop, XEXP (comparison, 0))
9238 || loop_invariant_p (loop, XEXP (comparison, 1)))
9241 return gen_rtx_fmt_ee (swap_condition (GET_CODE (comparison)), VOIDmode,
9242 XEXP (comparison, 1), XEXP (comparison, 0));
9246 /* Scan the function and determine whether it has indirect (computed) jumps.
9248 This is taken mostly from flow.c; similar code exists elsewhere
9249 in the compiler. It may be useful to put this into rtlanal.c. */
9251 indirect_jump_in_function_p (start)
9256 for (insn = start; insn; insn = NEXT_INSN (insn))
9257 if (computed_jump_p (insn))
9263 /* Add MEM to the LOOP_MEMS array, if appropriate. See the
9264 documentation for LOOP_MEMS for the definition of `appropriate'.
9265 This function is called from prescan_loop via for_each_rtx. */
9268 insert_loop_mem (mem, data)
9270 void *data ATTRIBUTE_UNUSED;
9278 switch (GET_CODE (m))
9284 /* We're not interested in MEMs that are only clobbered. */
9288 /* We're not interested in the MEM associated with a
9289 CONST_DOUBLE, so there's no need to traverse into this. */
9293 /* We're not interested in any MEMs that only appear in notes. */
9297 /* This is not a MEM. */
9301 /* See if we've already seen this MEM. */
9302 for (i = 0; i < loop_mems_idx; ++i)
9303 if (rtx_equal_p (m, loop_mems[i].mem))
9305 if (GET_MODE (m) != GET_MODE (loop_mems[i].mem))
9306 /* The modes of the two memory accesses are different. If
9307 this happens, something tricky is going on, and we just
9308 don't optimize accesses to this MEM. */
9309 loop_mems[i].optimize = 0;
9314 /* Resize the array, if necessary. */
9315 if (loop_mems_idx == loop_mems_allocated)
9317 if (loop_mems_allocated != 0)
9318 loop_mems_allocated *= 2;
9320 loop_mems_allocated = 32;
9322 loop_mems = (loop_mem_info*)
9323 xrealloc (loop_mems,
9324 loop_mems_allocated * sizeof (loop_mem_info));
9327 /* Actually insert the MEM. */
9328 loop_mems[loop_mems_idx].mem = m;
9329 /* We can't hoist this MEM out of the loop if it's a BLKmode MEM
9330 because we can't put it in a register. We still store it in the
9331 table, though, so that if we see the same address later, but in a
9332 non-BLK mode, we'll not think we can optimize it at that point. */
9333 loop_mems[loop_mems_idx].optimize = (GET_MODE (m) != BLKmode);
9334 loop_mems[loop_mems_idx].reg = NULL_RTX;
9340 /* Like load_mems, but also ensures that SET_IN_LOOP,
9341 MAY_NOT_OPTIMIZE, REG_SINGLE_USAGE, and INSN_COUNT have the correct
9342 values after load_mems. */
9345 load_mems_and_recount_loop_regs_set (loop, insn_count)
9346 const struct loop *loop;
9349 int nregs = max_reg_num ();
9353 /* Recalculate set_in_loop and friends since load_mems may have
9354 created new registers. */
9355 if (max_reg_num () > nregs)
9361 nregs = max_reg_num ();
9363 if ((unsigned) nregs > set_in_loop->num_elements)
9365 /* Grow all the arrays. */
9366 VARRAY_GROW (set_in_loop, nregs);
9367 VARRAY_GROW (n_times_set, nregs);
9368 VARRAY_GROW (may_not_optimize, nregs);
9369 VARRAY_GROW (reg_single_usage, nregs);
9371 /* Clear the arrays */
9372 bzero ((char *) &set_in_loop->data, nregs * sizeof (int));
9373 bzero ((char *) &may_not_optimize->data, nregs * sizeof (char));
9374 bzero ((char *) ®_single_usage->data, nregs * sizeof (rtx));
9376 count_loop_regs_set (loop->top ? loop->top : loop->start, loop->end,
9377 may_not_optimize, reg_single_usage,
9380 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
9382 VARRAY_CHAR (may_not_optimize, i) = 1;
9383 VARRAY_INT (set_in_loop, i) = 1;
9386 #ifdef AVOID_CCMODE_COPIES
9387 /* Don't try to move insns which set CC registers if we should not
9388 create CCmode register copies. */
9389 for (i = max_reg_num () - 1; i >= FIRST_PSEUDO_REGISTER; i--)
9390 if (GET_MODE_CLASS (GET_MODE (regno_reg_rtx[i])) == MODE_CC)
9391 VARRAY_CHAR (may_not_optimize, i) = 1;
9394 /* Set n_times_set for the new registers. */
9395 bcopy ((char *) (&set_in_loop->data.i[0] + old_nregs),
9396 (char *) (&n_times_set->data.i[0] + old_nregs),
9397 (nregs - old_nregs) * sizeof (int));
9401 /* Move MEMs into registers for the duration of the loop. */
9405 const struct loop *loop;
9407 int maybe_never = 0;
9410 rtx label = NULL_RTX;
9411 rtx end_label = NULL_RTX;
9412 /* Nonzero if the next instruction may never be executed. */
9413 int next_maybe_never = 0;
9414 int last_max_reg = max_reg_num ();
9416 if (loop_mems_idx == 0)
9419 /* Find start of the extended basic block that enters the loop. */
9420 for (p = loop->start;
9421 PREV_INSN (p) && GET_CODE (p) != CODE_LABEL;
9427 /* Build table of mems that get set to constant values before the
9429 for (; p != loop->start; p = NEXT_INSN (p))
9430 cselib_process_insn (p);
9432 /* Check to see if it's possible that some instructions in the
9433 loop are never executed. */
9434 for (p = next_insn_in_loop (loop, loop->scan_start);
9435 p != NULL_RTX && ! maybe_never;
9436 p = next_insn_in_loop (loop, p))
9438 if (GET_CODE (p) == CODE_LABEL)
9440 else if (GET_CODE (p) == JUMP_INSN
9441 /* If we enter the loop in the middle, and scan
9442 around to the beginning, don't set maybe_never
9443 for that. This must be an unconditional jump,
9444 otherwise the code at the top of the loop might
9445 never be executed. Unconditional jumps are
9446 followed a by barrier then loop end. */
9447 && ! (GET_CODE (p) == JUMP_INSN
9448 && JUMP_LABEL (p) == loop->top
9449 && NEXT_INSN (NEXT_INSN (p)) == loop->end
9450 && any_uncondjump_p (p)))
9452 if (!any_condjump_p (p))
9453 /* Something complicated. */
9456 /* If there are any more instructions in the loop, they
9457 might not be reached. */
9458 next_maybe_never = 1;
9460 else if (next_maybe_never)
9464 /* Actually move the MEMs. */
9465 for (i = 0; i < loop_mems_idx; ++i)
9467 regset_head load_copies;
9468 regset_head store_copies;
9471 rtx mem = loop_mems[i].mem;
9474 if (MEM_VOLATILE_P (mem)
9475 || loop_invariant_p (loop, XEXP (mem, 0)) != 1)
9476 /* There's no telling whether or not MEM is modified. */
9477 loop_mems[i].optimize = 0;
9479 /* Go through the MEMs written to in the loop to see if this
9480 one is aliased by one of them. */
9481 mem_list_entry = loop_store_mems;
9482 while (mem_list_entry)
9484 if (rtx_equal_p (mem, XEXP (mem_list_entry, 0)))
9486 else if (true_dependence (XEXP (mem_list_entry, 0), VOIDmode,
9489 /* MEM is indeed aliased by this store. */
9490 loop_mems[i].optimize = 0;
9493 mem_list_entry = XEXP (mem_list_entry, 1);
9496 if (flag_float_store && written
9497 && GET_MODE_CLASS (GET_MODE (mem)) == MODE_FLOAT)
9498 loop_mems[i].optimize = 0;
9500 /* If this MEM is written to, we must be sure that there
9501 are no reads from another MEM that aliases this one. */
9502 if (loop_mems[i].optimize && written)
9506 for (j = 0; j < loop_mems_idx; ++j)
9510 else if (true_dependence (mem,
9515 /* It's not safe to hoist loop_mems[i] out of
9516 the loop because writes to it might not be
9517 seen by reads from loop_mems[j]. */
9518 loop_mems[i].optimize = 0;
9524 if (maybe_never && may_trap_p (mem))
9525 /* We can't access the MEM outside the loop; it might
9526 cause a trap that wouldn't have happened otherwise. */
9527 loop_mems[i].optimize = 0;
9529 if (!loop_mems[i].optimize)
9530 /* We thought we were going to lift this MEM out of the
9531 loop, but later discovered that we could not. */
9534 INIT_REG_SET (&load_copies);
9535 INIT_REG_SET (&store_copies);
9537 /* Allocate a pseudo for this MEM. We set REG_USERVAR_P in
9538 order to keep scan_loop from moving stores to this MEM
9539 out of the loop just because this REG is neither a
9540 user-variable nor used in the loop test. */
9541 reg = gen_reg_rtx (GET_MODE (mem));
9542 REG_USERVAR_P (reg) = 1;
9543 loop_mems[i].reg = reg;
9545 /* Now, replace all references to the MEM with the
9546 corresponding pesudos. */
9548 for (p = next_insn_in_loop (loop, loop->scan_start);
9550 p = next_insn_in_loop (loop, p))
9558 set = single_set (p);
9560 /* See if this copies the mem into a register that isn't
9561 modified afterwards. We'll try to do copy propagation
9562 a little further on. */
9564 /* @@@ This test is _way_ too conservative. */
9566 && GET_CODE (SET_DEST (set)) == REG
9567 && REGNO (SET_DEST (set)) >= FIRST_PSEUDO_REGISTER
9568 && REGNO (SET_DEST (set)) < last_max_reg
9569 && VARRAY_INT (n_times_set, REGNO (SET_DEST (set))) == 1U
9570 && rtx_equal_p (SET_SRC (set), mem))
9571 SET_REGNO_REG_SET (&load_copies, REGNO (SET_DEST (set)));
9573 /* See if this copies the mem from a register that isn't
9574 modified afterwards. We'll try to remove the
9575 redundant copy later on by doing a little register
9576 renaming and copy propagation. This will help
9577 to untangle things for the BIV detection code. */
9580 && GET_CODE (SET_SRC (set)) == REG
9581 && REGNO (SET_SRC (set)) >= FIRST_PSEUDO_REGISTER
9582 && REGNO (SET_SRC (set)) < last_max_reg
9583 && VARRAY_INT (n_times_set, REGNO (SET_SRC (set))) == 1U
9584 && rtx_equal_p (SET_DEST (set), mem))
9585 SET_REGNO_REG_SET (&store_copies, REGNO (SET_SRC (set)));
9587 /* Replace the memory reference with the shadow register. */
9590 for_each_rtx (&p, replace_loop_mem, &ri);
9593 if (GET_CODE (p) == CODE_LABEL
9594 || GET_CODE (p) == JUMP_INSN)
9598 if (! apply_change_group ())
9599 /* We couldn't replace all occurrences of the MEM. */
9600 loop_mems[i].optimize = 0;
9603 /* Load the memory immediately before LOOP->START, which is
9604 the NOTE_LOOP_BEG. */
9605 cselib_val *e = cselib_lookup (mem, VOIDmode, 0);
9609 struct elt_loc_list *const_equiv = 0;
9613 struct elt_loc_list *equiv;
9614 struct elt_loc_list *best_equiv = 0;
9615 for (equiv = e->locs; equiv; equiv = equiv->next)
9617 if (CONSTANT_P (equiv->loc))
9618 const_equiv = equiv;
9619 else if (GET_CODE (equiv->loc) == REG
9620 /* Extending hard register lifetimes cuases crash
9621 on SRC targets. Doing so on non-SRC is
9622 probably also not good idea, since we most
9623 probably have pseudoregister equivalence as
9625 && REGNO (equiv->loc) >= FIRST_PSEUDO_REGISTER)
9628 /* Use the constant equivalence if that is cheap enough. */
9630 best_equiv = const_equiv;
9631 else if (const_equiv
9632 && (rtx_cost (const_equiv->loc, SET)
9633 <= rtx_cost (best_equiv->loc, SET)))
9635 best_equiv = const_equiv;
9639 /* If best_equiv is nonzero, we know that MEM is set to a
9640 constant or register before the loop. We will use this
9641 knowledge to initialize the shadow register with that
9642 constant or reg rather than by loading from MEM. */
9644 best = copy_rtx (best_equiv->loc);
9646 set = gen_move_insn (reg, best);
9647 set = emit_insn_before (set, loop->start);
9649 REG_NOTES (set) = gen_rtx_EXPR_LIST (REG_EQUAL,
9650 copy_rtx (const_equiv->loc),
9655 if (label == NULL_RTX)
9657 /* We must compute the former
9658 right-after-the-end label before we insert
9660 end_label = next_label (loop->end);
9661 label = gen_label_rtx ();
9662 emit_label_after (label, loop->end);
9665 /* Store the memory immediately after END, which is
9666 the NOTE_LOOP_END. */
9667 set = gen_move_insn (copy_rtx (mem), reg);
9668 emit_insn_after (set, label);
9671 if (loop_dump_stream)
9673 fprintf (loop_dump_stream, "Hoisted regno %d %s from ",
9674 REGNO (reg), (written ? "r/w" : "r/o"));
9675 print_rtl (loop_dump_stream, mem);
9676 fputc ('\n', loop_dump_stream);
9679 /* Attempt a bit of copy propagation. This helps untangle the
9680 data flow, and enables {basic,general}_induction_var to find
9682 EXECUTE_IF_SET_IN_REG_SET
9683 (&load_copies, FIRST_PSEUDO_REGISTER, j,
9685 try_copy_prop (loop, reg, j);
9687 CLEAR_REG_SET (&load_copies);
9689 EXECUTE_IF_SET_IN_REG_SET
9690 (&store_copies, FIRST_PSEUDO_REGISTER, j,
9692 try_swap_copy_prop (loop, reg, j);
9694 CLEAR_REG_SET (&store_copies);
9698 if (label != NULL_RTX)
9700 /* Now, we need to replace all references to the previous exit
9701 label with the new one. */
9706 for (p = loop->start; p != loop->end; p = NEXT_INSN (p))
9708 for_each_rtx (&p, replace_label, &rr);
9710 /* If this is a JUMP_INSN, then we also need to fix the JUMP_LABEL
9711 field. This is not handled by for_each_rtx because it doesn't
9712 handle unprinted ('0') fields. We need to update JUMP_LABEL
9713 because the immediately following unroll pass will use it.
9714 replace_label would not work anyways, because that only handles
9716 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) == end_label)
9717 JUMP_LABEL (p) = label;
9724 /* For communication between note_reg_stored and its caller. */
9725 struct note_reg_stored_arg
9731 /* Called via note_stores, record in SET_SEEN whether X, which is written,
9734 note_reg_stored (x, setter, arg)
9735 rtx x, setter ATTRIBUTE_UNUSED;
9738 struct note_reg_stored_arg *t = (struct note_reg_stored_arg *)arg;
9743 /* Try to replace every occurrence of pseudo REGNO with REPLACEMENT.
9744 There must be exactly one insn that sets this pseudo; it will be
9745 deleted if all replacements succeed and we can prove that the register
9746 is not used after the loop. */
9749 try_copy_prop (loop, replacement, regno)
9750 const struct loop *loop;
9754 /* This is the reg that we are copying from. */
9755 rtx reg_rtx = regno_reg_rtx[regno];
9758 /* These help keep track of whether we replaced all uses of the reg. */
9759 int replaced_last = 0;
9760 int store_is_first = 0;
9762 for (insn = next_insn_in_loop (loop, loop->scan_start);
9764 insn = next_insn_in_loop (loop, insn))
9768 /* Only substitute within one extended basic block from the initializing
9770 if (GET_CODE (insn) == CODE_LABEL && init_insn)
9773 if (! INSN_P (insn))
9776 /* Is this the initializing insn? */
9777 set = single_set (insn);
9779 && GET_CODE (SET_DEST (set)) == REG
9780 && REGNO (SET_DEST (set)) == regno)
9786 if (REGNO_FIRST_UID (regno) == INSN_UID (insn))
9790 /* Only substitute after seeing the initializing insn. */
9791 if (init_insn && insn != init_insn)
9793 struct note_reg_stored_arg arg;
9796 array[1] = replacement;
9799 for_each_rtx (&insn, replace_loop_reg, array);
9800 if (REGNO_LAST_UID (regno) == INSN_UID (insn))
9803 /* Stop replacing when REPLACEMENT is modified. */
9804 arg.reg = replacement;
9806 note_stores (PATTERN (insn), note_reg_stored, &arg);
9813 if (apply_change_group ())
9815 if (loop_dump_stream)
9816 fprintf (loop_dump_stream, " Replaced reg %d", regno);
9817 if (store_is_first && replaced_last)
9819 PUT_CODE (init_insn, NOTE);
9820 NOTE_LINE_NUMBER (init_insn) = NOTE_INSN_DELETED;
9821 if (loop_dump_stream)
9822 fprintf (loop_dump_stream, ", deleting init_insn (%d)",
9823 INSN_UID (init_insn));
9825 if (loop_dump_stream)
9826 fprintf (loop_dump_stream, ".\n");
9831 /* Try to replace occurrences of pseudo REGNO with REPLACEMENT within
9832 loop LOOP if the order of the sets of these registers can be
9833 swapped. There must be exactly one insn within the loop that sets
9834 this pseudo followed immediately by a move insn that sets
9835 REPLACEMENT with REGNO. */
9837 try_swap_copy_prop (loop, replacement, regno)
9838 const struct loop *loop;
9844 unsigned int new_regno;
9846 new_regno = REGNO (replacement);
9848 for (insn = next_insn_in_loop (loop, loop->scan_start);
9850 insn = next_insn_in_loop (loop, insn))
9852 /* Search for the insn that copies REGNO to NEW_REGNO? */
9853 if (GET_RTX_CLASS (GET_CODE (insn)) == 'i'
9854 && (set = single_set (insn))
9855 && GET_CODE (SET_DEST (set)) == REG
9856 && REGNO (SET_DEST (set)) == new_regno
9857 && GET_CODE (SET_SRC (set)) == REG
9858 && REGNO (SET_SRC (set)) == regno)
9862 if (insn != NULL_RTX)
9867 /* Some DEF-USE info would come in handy here to make this
9868 function more general. For now, just check the previous insn
9869 which is the most likely candidate for setting REGNO. */
9871 prev_insn = PREV_INSN (insn);
9873 if (GET_RTX_CLASS (GET_CODE (insn)) == 'i'
9874 && (prev_set = single_set (prev_insn))
9875 && GET_CODE (SET_DEST (prev_set)) == REG
9876 && REGNO (SET_DEST (prev_set)) == regno)
9879 (set (reg regno) (expr))
9880 (set (reg new_regno) (reg regno))
9882 so try converting this to:
9883 (set (reg new_regno) (expr))
9884 (set (reg regno) (reg new_regno))
9886 The former construct is often generated when a global
9887 variable used for an induction variable is shadowed by a
9888 register (NEW_REGNO). The latter construct improves the
9889 chances of GIV replacement and BIV elimination. */
9891 validate_change (prev_insn, &SET_DEST (prev_set),
9893 validate_change (insn, &SET_DEST (set),
9895 validate_change (insn, &SET_SRC (set),
9898 if (apply_change_group ())
9900 if (loop_dump_stream)
9901 fprintf (loop_dump_stream,
9902 " Swapped set of reg %d at %d with reg %d at %d.\n",
9903 regno, INSN_UID (insn),
9904 new_regno, INSN_UID (prev_insn));
9906 /* Update first use of REGNO. */
9907 if (REGNO_FIRST_UID (regno) == INSN_UID (prev_insn))
9908 REGNO_FIRST_UID (regno) = INSN_UID (insn);
9910 /* Now perform copy propagation to hopefully
9911 remove all uses of REGNO within the loop. */
9912 try_copy_prop (loop, replacement, regno);
9919 /* Replace MEM with its associated pseudo register. This function is
9920 called from load_mems via for_each_rtx. DATA is actually an
9921 rtx_and_int * describing the instruction currently being scanned
9922 and the MEM we are currently replacing. */
9925 replace_loop_mem (mem, data)
9937 switch (GET_CODE (m))
9943 /* We're not interested in the MEM associated with a
9944 CONST_DOUBLE, so there's no need to traverse into one. */
9948 /* This is not a MEM. */
9952 ri = (rtx_and_int*) data;
9955 if (!rtx_equal_p (loop_mems[i].mem, m))
9956 /* This is not the MEM we are currently replacing. */
9961 /* Actually replace the MEM. */
9962 validate_change (insn, mem, loop_mems[i].reg, 1);
9967 /* Replace one register with another. Called through for_each_rtx; PX points
9968 to the rtx being scanned. DATA is actually an array of three rtx's; the
9969 first one is the one to be replaced, and the second one the replacement.
9970 The third one is the current insn. */
9973 replace_loop_reg (px, data)
9978 rtx *array = (rtx *)data;
9984 validate_change (array[2], px, array[1], 1);
9989 /* Replace occurrences of the old exit label for the loop with the new
9990 one. DATA is an rtx_pair containing the old and new labels,
9994 replace_label (x, data)
9999 rtx old_label = ((rtx_pair*) data)->r1;
10000 rtx new_label = ((rtx_pair*) data)->r2;
10005 if (GET_CODE (l) != LABEL_REF)
10008 if (XEXP (l, 0) != old_label)
10011 XEXP (l, 0) = new_label;
10012 ++LABEL_NUSES (new_label);
10013 --LABEL_NUSES (old_label);