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. */
22 /* This is the loop optimization pass of the compiler.
23 It finds invariant computations within loops and moves them
24 to the beginning of the loop. Then it identifies basic and
25 general induction variables. Strength reduction is applied to the general
26 induction variables, and induction variable elimination is applied to
27 the basic induction variables.
29 It also finds cases where
30 a register is set within the loop by zero-extending a narrower value
31 and changes these to zero the entire register once before the loop
32 and merely copy the low part within the loop.
34 Most of the complexity is in heuristics to decide when it is worth
35 while to do these things. */
44 #include "hard-reg-set.h"
45 #include "basic-block.h"
46 #include "insn-config.h"
47 #include "insn-flags.h"
57 /* Vector mapping INSN_UIDs to luids.
58 The luids are like uids but increase monotonically always.
59 We use them to see whether a jump comes from outside a given loop. */
63 /* Indexed by INSN_UID, contains the ordinal giving the (innermost) loop
64 number the insn is contained in. */
66 struct loop **uid_loop;
68 /* 1 + largest uid of any insn. */
72 /* 1 + luid of last insn. */
76 /* Number of loops detected in current function. Used as index to the
79 static int max_loop_num;
81 /* Indexed by register number, contains the number of times the reg
82 is set during the loop being scanned.
83 During code motion, a negative value indicates a reg that has been
84 made a candidate; in particular -2 means that it is an candidate that
85 we know is equal to a constant and -1 means that it is an candidate
86 not known equal to a constant.
87 After code motion, regs moved have 0 (which is accurate now)
88 while the failed candidates have the original number of times set.
90 Therefore, at all times, == 0 indicates an invariant register;
91 < 0 a conditionally invariant one. */
93 static varray_type set_in_loop;
95 /* Original value of set_in_loop; same except that this value
96 is not set negative for a reg whose sets have been made candidates
97 and not set to 0 for a reg that is moved. */
99 static varray_type n_times_set;
101 /* Index by register number, 1 indicates that the register
102 cannot be moved or strength reduced. */
104 static varray_type may_not_optimize;
106 /* Contains the insn in which a register was used if it was used
107 exactly once; contains const0_rtx if it was used more than once. */
109 static varray_type reg_single_usage;
111 /* Nonzero means reg N has already been moved out of one loop.
112 This reduces the desire to move it out of another. */
114 static char *moved_once;
116 /* List of MEMs that are stored in this loop. */
118 static rtx loop_store_mems;
120 /* The insn where the first of these was found. */
121 static rtx first_loop_store_insn;
123 typedef struct loop_mem_info {
124 rtx mem; /* The MEM itself. */
125 rtx reg; /* Corresponding pseudo, if any. */
126 int optimize; /* Nonzero if we can optimize access to this MEM. */
129 /* Array of MEMs that are used (read or written) in this loop, but
130 cannot be aliased by anything in this loop, except perhaps
131 themselves. In other words, if loop_mems[i] is altered during the
132 loop, it is altered by an expression that is rtx_equal_p to it. */
134 static loop_mem_info *loop_mems;
136 /* The index of the next available slot in LOOP_MEMS. */
138 static int loop_mems_idx;
140 /* The number of elements allocated in LOOP_MEMs. */
142 static int loop_mems_allocated;
144 /* Nonzero if we don't know what MEMs were changed in the current
145 loop. This happens if the loop contains a call (in which case
146 `loop_info->has_call' will also be set) or if we store into more
147 than NUM_STORES MEMs. */
149 static int unknown_address_altered;
151 /* The above doesn't count any readonly memory locations that are stored.
154 static int unknown_constant_address_altered;
156 /* Count of movable (i.e. invariant) instructions discovered in the loop. */
157 static int num_movables;
159 /* Count of memory write instructions discovered in the loop. */
160 static int num_mem_sets;
162 /* Bound on pseudo register number before loop optimization.
163 A pseudo has valid regscan info if its number is < max_reg_before_loop. */
164 unsigned int max_reg_before_loop;
166 /* The value to pass to the next call of reg_scan_update. */
167 static int loop_max_reg;
169 /* This obstack is used in product_cheap_p to allocate its rtl. It
170 may call gen_reg_rtx which, in turn, may reallocate regno_reg_rtx.
171 If we used the same obstack that it did, we would be deallocating
174 static struct obstack temp_obstack;
176 /* This is where the pointer to the obstack being used for RTL is stored. */
178 extern struct obstack *rtl_obstack;
180 #define obstack_chunk_alloc xmalloc
181 #define obstack_chunk_free free
183 /* During the analysis of a loop, a chain of `struct movable's
184 is made to record all the movable insns found.
185 Then the entire chain can be scanned to decide which to move. */
189 rtx insn; /* A movable insn */
190 rtx set_src; /* The expression this reg is set from. */
191 rtx set_dest; /* The destination of this SET. */
192 rtx dependencies; /* When INSN is libcall, this is an EXPR_LIST
193 of any registers used within the LIBCALL. */
194 int consec; /* Number of consecutive following insns
195 that must be moved with this one. */
196 unsigned int regno; /* The register it sets */
197 short lifetime; /* lifetime of that register;
198 may be adjusted when matching movables
199 that load the same value are found. */
200 short savings; /* Number of insns we can move for this reg,
201 including other movables that force this
202 or match this one. */
203 unsigned int cond : 1; /* 1 if only conditionally movable */
204 unsigned int force : 1; /* 1 means MUST move this insn */
205 unsigned int global : 1; /* 1 means reg is live outside this loop */
206 /* If PARTIAL is 1, GLOBAL means something different:
207 that the reg is live outside the range from where it is set
208 to the following label. */
209 unsigned int done : 1; /* 1 inhibits further processing of this */
211 unsigned int partial : 1; /* 1 means this reg is used for zero-extending.
212 In particular, moving it does not make it
214 unsigned int move_insn : 1; /* 1 means that we call emit_move_insn to
215 load SRC, rather than copying INSN. */
216 unsigned int move_insn_first:1;/* Same as above, if this is necessary for the
217 first insn of a consecutive sets group. */
218 unsigned int is_equiv : 1; /* 1 means a REG_EQUIV is present on INSN. */
219 enum machine_mode savemode; /* Nonzero means it is a mode for a low part
220 that we should avoid changing when clearing
221 the rest of the reg. */
222 struct movable *match; /* First entry for same value */
223 struct movable *forces; /* An insn that must be moved if this is */
224 struct movable *next;
227 static struct movable *the_movables;
229 FILE *loop_dump_stream;
231 /* Forward declarations. */
233 static void verify_dominator PARAMS ((struct loop *));
234 static void find_and_verify_loops PARAMS ((rtx, struct loops *));
235 static void mark_loop_jump PARAMS ((rtx, struct loop *));
236 static void prescan_loop PARAMS ((struct loop *));
237 static int reg_in_basic_block_p PARAMS ((rtx, rtx));
238 static int consec_sets_invariant_p PARAMS ((const struct loop *,
240 static int labels_in_range_p PARAMS ((rtx, int));
241 static void count_one_set PARAMS ((rtx, rtx, varray_type, rtx *));
243 static void count_loop_regs_set PARAMS ((const struct loop*,
244 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));
315 static int iv_add_mult_cost PARAMS ((rtx, rtx, rtx, rtx));
317 static void loop_dump_aux PARAMS ((const struct loop *, FILE *, int));
318 void debug_loop PARAMS ((const struct loop *));
320 typedef struct rtx_and_int {
325 typedef struct rtx_pair {
330 /* Nonzero iff INSN is between START and END, inclusive. */
331 #define INSN_IN_RANGE_P(INSN, START, END) \
332 (INSN_UID (INSN) < max_uid_for_loop \
333 && INSN_LUID (INSN) >= INSN_LUID (START) \
334 && INSN_LUID (INSN) <= INSN_LUID (END))
336 /* Indirect_jump_in_function is computed once per function. */
337 static int indirect_jump_in_function;
338 static int indirect_jump_in_function_p PARAMS ((rtx));
340 static int compute_luids PARAMS ((rtx, rtx, int));
342 static int biv_elimination_giv_has_0_offset PARAMS ((struct induction *,
343 struct induction *, rtx));
345 /* Benefit penalty, if a giv is not replaceable, i.e. must emit an insn to
346 copy the value of the strength reduced giv to its original register. */
347 static int copy_cost;
349 /* Cost of using a register, to normalize the benefits of a giv. */
350 static int reg_address_cost;
355 char *free_point = (char *) oballoc (1);
356 rtx reg = gen_rtx_REG (word_mode, LAST_VIRTUAL_REGISTER + 1);
358 reg_address_cost = address_cost (reg, SImode);
362 /* Free the objects we just allocated. */
365 /* Initialize the obstack used for rtl in product_cheap_p. */
366 gcc_obstack_init (&temp_obstack);
369 /* Compute the mapping from uids to luids.
370 LUIDs are numbers assigned to insns, like uids,
371 except that luids increase monotonically through the code.
372 Start at insn START and stop just before END. Assign LUIDs
373 starting with PREV_LUID + 1. Return the last assigned LUID + 1. */
375 compute_luids (start, end, prev_luid)
382 for (insn = start, i = prev_luid; insn != end; insn = NEXT_INSN (insn))
384 if (INSN_UID (insn) >= max_uid_for_loop)
386 /* Don't assign luids to line-number NOTEs, so that the distance in
387 luids between two insns is not affected by -g. */
388 if (GET_CODE (insn) != NOTE
389 || NOTE_LINE_NUMBER (insn) <= 0)
390 uid_luid[INSN_UID (insn)] = ++i;
392 /* Give a line number note the same luid as preceding insn. */
393 uid_luid[INSN_UID (insn)] = i;
398 /* Entry point of this file. Perform loop optimization
399 on the current function. F is the first insn of the function
400 and DUMPFILE is a stream for output of a trace of actions taken
401 (or 0 if none should be output). */
404 loop_optimize (f, dumpfile, flags)
405 /* f is the first instruction of a chain of insns for one function */
412 struct loops loops_data;
413 struct loops *loops = &loops_data;
414 struct loop_info *loops_info;
416 loop_dump_stream = dumpfile;
418 init_recog_no_volatile ();
420 max_reg_before_loop = max_reg_num ();
421 loop_max_reg = max_reg_before_loop;
425 /* Count the number of loops. */
428 for (insn = f; insn; insn = NEXT_INSN (insn))
430 if (GET_CODE (insn) == NOTE
431 && NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG)
435 /* Don't waste time if no loops. */
436 if (max_loop_num == 0)
439 loops->num = max_loop_num;
441 moved_once = (char *) xcalloc (max_reg_before_loop, sizeof (char));
443 /* Get size to use for tables indexed by uids.
444 Leave some space for labels allocated by find_and_verify_loops. */
445 max_uid_for_loop = get_max_uid () + 1 + max_loop_num * 32;
447 uid_luid = (int *) xcalloc (max_uid_for_loop, sizeof (int));
448 uid_loop = (struct loop **) xcalloc (max_uid_for_loop,
449 sizeof (struct loop *));
451 /* Allocate storage for array of loops. */
452 loops->array = (struct loop *)
453 xcalloc (loops->num, sizeof (struct loop));
455 /* Find and process each loop.
456 First, find them, and record them in order of their beginnings. */
457 find_and_verify_loops (f, loops);
459 /* Allocate and initialize auxiliary loop information. */
460 loops_info = xcalloc (loops->num, sizeof (struct loop_info));
461 for (i = 0; i < loops->num; i++)
462 loops->array[i].aux = loops_info + i;
464 /* Now find all register lifetimes. This must be done after
465 find_and_verify_loops, because it might reorder the insns in the
467 reg_scan (f, max_reg_before_loop, 1);
469 /* This must occur after reg_scan so that registers created by gcse
470 will have entries in the register tables.
472 We could have added a call to reg_scan after gcse_main in toplev.c,
473 but moving this call to init_alias_analysis is more efficient. */
474 init_alias_analysis ();
476 /* See if we went too far. Note that get_max_uid already returns
477 one more that the maximum uid of all insn. */
478 if (get_max_uid () > max_uid_for_loop)
480 /* Now reset it to the actual size we need. See above. */
481 max_uid_for_loop = get_max_uid ();
483 /* find_and_verify_loops has already called compute_luids, but it
484 might have rearranged code afterwards, so we need to recompute
486 max_luid = compute_luids (f, NULL_RTX, 0);
488 /* Don't leave gaps in uid_luid for insns that have been
489 deleted. It is possible that the first or last insn
490 using some register has been deleted by cross-jumping.
491 Make sure that uid_luid for that former insn's uid
492 points to the general area where that insn used to be. */
493 for (i = 0; i < max_uid_for_loop; i++)
495 uid_luid[0] = uid_luid[i];
496 if (uid_luid[0] != 0)
499 for (i = 0; i < max_uid_for_loop; i++)
500 if (uid_luid[i] == 0)
501 uid_luid[i] = uid_luid[i - 1];
503 /* Determine if the function has indirect jump. On some systems
504 this prevents low overhead loop instructions from being used. */
505 indirect_jump_in_function = indirect_jump_in_function_p (f);
507 /* Now scan the loops, last ones first, since this means inner ones are done
508 before outer ones. */
509 for (i = max_loop_num - 1; i >= 0; i--)
511 struct loop *loop = &loops->array[i];
513 if (! loop->invalid && loop->end)
514 scan_loop (loop, flags);
517 /* If there were lexical blocks inside the loop, they have been
518 replicated. We will now have more than one NOTE_INSN_BLOCK_BEG
519 and NOTE_INSN_BLOCK_END for each such block. We must duplicate
520 the BLOCKs as well. */
521 if (write_symbols != NO_DEBUG)
524 end_alias_analysis ();
534 /* Returns the next insn, in execution order, after INSN. START and
535 END are the NOTE_INSN_LOOP_BEG and NOTE_INSN_LOOP_END for the loop,
536 respectively. LOOP->TOP, if non-NULL, is the top of the loop in the
537 insn-stream; it is used with loops that are entered near the
541 next_insn_in_loop (loop, insn)
542 const struct loop *loop;
545 insn = NEXT_INSN (insn);
547 if (insn == loop->end)
550 /* Go to the top of the loop, and continue there. */
557 if (insn == loop->scan_start)
564 /* Optimize one loop described by LOOP. */
566 /* ??? Could also move memory writes out of loops if the destination address
567 is invariant, the source is invariant, the memory write is not volatile,
568 and if we can prove that no read inside the loop can read this address
569 before the write occurs. If there is a read of this address after the
570 write, then we can also mark the memory read as invariant. */
573 scan_loop (loop, flags)
578 rtx loop_start = loop->start;
579 rtx loop_end = loop->end;
580 /* Additional information about the current loop being processed
581 that is used to compute the number of loop iterations for loop
582 unrolling and doloop optimization. */
583 struct loop_info *loop_info = LOOP_INFO (loop);
585 /* 1 if we are scanning insns that could be executed zero times. */
587 /* 1 if we are scanning insns that might never be executed
588 due to a subroutine call which might exit before they are reached. */
590 /* Jump insn that enters the loop, or 0 if control drops in. */
591 rtx loop_entry_jump = 0;
592 /* Number of insns in the loop. */
595 rtx temp, update_start, update_end;
596 /* The SET from an insn, if it is the only SET in the insn. */
598 /* Chain describing insns movable in current loop. */
599 struct movable *movables = 0;
600 /* Last element in `movables' -- so we can add elements at the end. */
601 struct movable *last_movable = 0;
602 /* Ratio of extra register life span we can justify
603 for saving an instruction. More if loop doesn't call subroutines
604 since in that case saving an insn makes more difference
605 and more registers are available. */
607 /* Nonzero if we are scanning instructions in a sub-loop. */
613 /* Determine whether this loop starts with a jump down to a test at
614 the end. This will occur for a small number of loops with a test
615 that is too complex to duplicate in front of the loop.
617 We search for the first insn or label in the loop, skipping NOTEs.
618 However, we must be careful not to skip past a NOTE_INSN_LOOP_BEG
619 (because we might have a loop executed only once that contains a
620 loop which starts with a jump to its exit test) or a NOTE_INSN_LOOP_END
621 (in case we have a degenerate loop).
623 Note that if we mistakenly think that a loop is entered at the top
624 when, in fact, it is entered at the exit test, the only effect will be
625 slightly poorer optimization. Making the opposite error can generate
626 incorrect code. Since very few loops now start with a jump to the
627 exit test, the code here to detect that case is very conservative. */
629 for (p = NEXT_INSN (loop_start);
631 && GET_CODE (p) != CODE_LABEL && ! INSN_P (p)
632 && (GET_CODE (p) != NOTE
633 || (NOTE_LINE_NUMBER (p) != NOTE_INSN_LOOP_BEG
634 && NOTE_LINE_NUMBER (p) != NOTE_INSN_LOOP_END));
638 loop->scan_start = p;
640 /* Set up variables describing this loop. */
642 threshold = (loop_info->has_call ? 1 : 2) * (1 + n_non_fixed_regs);
644 /* If loop has a jump before the first label,
645 the true entry is the target of that jump.
646 Start scan from there.
647 But record in LOOP->TOP the place where the end-test jumps
648 back to so we can scan that after the end of the loop. */
649 if (GET_CODE (p) == JUMP_INSN)
653 /* Loop entry must be unconditional jump (and not a RETURN) */
654 if (any_uncondjump_p (p)
655 && JUMP_LABEL (p) != 0
656 /* Check to see whether the jump actually
657 jumps out of the loop (meaning it's no loop).
658 This case can happen for things like
659 do {..} while (0). If this label was generated previously
660 by loop, we can't tell anything about it and have to reject
662 && INSN_IN_RANGE_P (JUMP_LABEL (p), loop_start, loop_end))
664 loop->top = next_label (loop->scan_start);
665 loop->scan_start = JUMP_LABEL (p);
669 /* If LOOP->SCAN_START was an insn created by loop, we don't know its luid
670 as required by loop_reg_used_before_p. So skip such loops. (This
671 test may never be true, but it's best to play it safe.)
673 Also, skip loops where we do not start scanning at a label. This
674 test also rejects loops starting with a JUMP_INSN that failed the
677 if (INSN_UID (loop->scan_start) >= max_uid_for_loop
678 || GET_CODE (loop->scan_start) != CODE_LABEL)
680 if (loop_dump_stream)
681 fprintf (loop_dump_stream, "\nLoop from %d to %d is phony.\n\n",
682 INSN_UID (loop_start), INSN_UID (loop_end));
686 /* Count number of times each reg is set during this loop.
687 Set VARRAY_CHAR (may_not_optimize, I) if it is not safe to move out
688 the setting of register I. Set VARRAY_RTX (reg_single_usage, I). */
690 /* Allocate extra space for REGS that might be created by
691 load_mems. We allocate a little extra slop as well, in the hopes
692 that even after the moving of movables creates some new registers
693 we won't have to reallocate these arrays. However, we do grow
694 the arrays, if necessary, in load_mems_recount_loop_regs_set. */
695 nregs = max_reg_num () + loop_mems_idx + 16;
696 VARRAY_INT_INIT (set_in_loop, nregs, "set_in_loop");
697 VARRAY_INT_INIT (n_times_set, nregs, "n_times_set");
698 VARRAY_CHAR_INIT (may_not_optimize, nregs, "may_not_optimize");
699 VARRAY_RTX_INIT (reg_single_usage, nregs, "reg_single_usage");
701 count_loop_regs_set (loop, may_not_optimize, reg_single_usage,
704 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
706 VARRAY_CHAR (may_not_optimize, i) = 1;
707 VARRAY_INT (set_in_loop, i) = 1;
710 #ifdef AVOID_CCMODE_COPIES
711 /* Don't try to move insns which set CC registers if we should not
712 create CCmode register copies. */
713 for (i = max_reg_num () - 1; i >= FIRST_PSEUDO_REGISTER; i--)
714 if (GET_MODE_CLASS (GET_MODE (regno_reg_rtx[i])) == MODE_CC)
715 VARRAY_CHAR (may_not_optimize, i) = 1;
718 bcopy ((char *) &set_in_loop->data,
719 (char *) &n_times_set->data, nregs * sizeof (int));
721 if (loop_dump_stream)
723 fprintf (loop_dump_stream, "\nLoop from %d to %d: %d real insns.\n",
724 INSN_UID (loop_start), INSN_UID (loop_end), insn_count);
726 fprintf (loop_dump_stream, "Continue at insn %d.\n",
727 INSN_UID (loop->cont));
730 /* Scan through the loop finding insns that are safe to move.
731 Set set_in_loop negative for the reg being set, so that
732 this reg will be considered invariant for subsequent insns.
733 We consider whether subsequent insns use the reg
734 in deciding whether it is worth actually moving.
736 MAYBE_NEVER is nonzero if we have passed a conditional jump insn
737 and therefore it is possible that the insns we are scanning
738 would never be executed. At such times, we must make sure
739 that it is safe to execute the insn once instead of zero times.
740 When MAYBE_NEVER is 0, all insns will be executed at least once
741 so that is not a problem. */
743 for (p = next_insn_in_loop (loop, loop->scan_start);
745 p = next_insn_in_loop (loop, p))
747 if (GET_CODE (p) == INSN
748 && (set = single_set (p))
749 && GET_CODE (SET_DEST (set)) == REG
750 && ! VARRAY_CHAR (may_not_optimize, REGNO (SET_DEST (set))))
755 rtx src = SET_SRC (set);
756 rtx dependencies = 0;
758 /* Figure out what to use as a source of this insn. If a REG_EQUIV
759 note is given or if a REG_EQUAL note with a constant operand is
760 specified, use it as the source and mark that we should move
761 this insn by calling emit_move_insn rather that duplicating the
764 Otherwise, only use the REG_EQUAL contents if a REG_RETVAL note
766 temp = find_reg_note (p, REG_EQUIV, NULL_RTX);
768 src = XEXP (temp, 0), move_insn = 1;
771 temp = find_reg_note (p, REG_EQUAL, NULL_RTX);
772 if (temp && CONSTANT_P (XEXP (temp, 0)))
773 src = XEXP (temp, 0), move_insn = 1;
774 if (temp && find_reg_note (p, REG_RETVAL, NULL_RTX))
776 src = XEXP (temp, 0);
777 /* A libcall block can use regs that don't appear in
778 the equivalent expression. To move the libcall,
779 we must move those regs too. */
780 dependencies = libcall_other_reg (p, src);
784 /* Don't try to optimize a register that was made
785 by loop-optimization for an inner loop.
786 We don't know its life-span, so we can't compute the benefit. */
787 if (REGNO (SET_DEST (set)) >= max_reg_before_loop)
789 else if (/* The register is used in basic blocks other
790 than the one where it is set (meaning that
791 something after this point in the loop might
792 depend on its value before the set). */
793 ! reg_in_basic_block_p (p, SET_DEST (set))
794 /* And the set is not guaranteed to be executed one
795 the loop starts, or the value before the set is
796 needed before the set occurs...
798 ??? Note we have quadratic behaviour here, mitigated
799 by the fact that the previous test will often fail for
800 large loops. Rather than re-scanning the entire loop
801 each time for register usage, we should build tables
802 of the register usage and use them here instead. */
804 || loop_reg_used_before_p (loop, set, p)))
805 /* It is unsafe to move the set.
807 This code used to consider it OK to move a set of a variable
808 which was not created by the user and not used in an exit test.
809 That behavior is incorrect and was removed. */
811 else if ((tem = loop_invariant_p (loop, src))
812 && (dependencies == 0
813 || (tem2 = loop_invariant_p (loop, dependencies)) != 0)
814 && (VARRAY_INT (set_in_loop,
815 REGNO (SET_DEST (set))) == 1
817 = consec_sets_invariant_p
818 (loop, SET_DEST (set),
819 VARRAY_INT (set_in_loop, REGNO (SET_DEST (set))),
821 /* If the insn can cause a trap (such as divide by zero),
822 can't move it unless it's guaranteed to be executed
823 once loop is entered. Even a function call might
824 prevent the trap insn from being reached
825 (since it might exit!) */
826 && ! ((maybe_never || call_passed)
827 && may_trap_p (src)))
829 register struct movable *m;
830 register int regno = REGNO (SET_DEST (set));
832 /* A potential lossage is where we have a case where two insns
833 can be combined as long as they are both in the loop, but
834 we move one of them outside the loop. For large loops,
835 this can lose. The most common case of this is the address
836 of a function being called.
838 Therefore, if this register is marked as being used exactly
839 once if we are in a loop with calls (a "large loop"), see if
840 we can replace the usage of this register with the source
841 of this SET. If we can, delete this insn.
843 Don't do this if P has a REG_RETVAL note or if we have
844 SMALL_REGISTER_CLASSES and SET_SRC is a hard register. */
846 if (loop_info->has_call
847 && VARRAY_RTX (reg_single_usage, regno) != 0
848 && VARRAY_RTX (reg_single_usage, regno) != const0_rtx
849 && REGNO_FIRST_UID (regno) == INSN_UID (p)
850 && (REGNO_LAST_UID (regno)
851 == INSN_UID (VARRAY_RTX (reg_single_usage, regno)))
852 && VARRAY_INT (set_in_loop, regno) == 1
853 && ! side_effects_p (SET_SRC (set))
854 && ! find_reg_note (p, REG_RETVAL, NULL_RTX)
855 && (! SMALL_REGISTER_CLASSES
856 || (! (GET_CODE (SET_SRC (set)) == REG
857 && REGNO (SET_SRC (set)) < FIRST_PSEUDO_REGISTER)))
858 /* This test is not redundant; SET_SRC (set) might be
859 a call-clobbered register and the life of REGNO
860 might span a call. */
861 && ! modified_between_p (SET_SRC (set), p,
863 (reg_single_usage, regno))
864 && no_labels_between_p (p, VARRAY_RTX (reg_single_usage, regno))
865 && validate_replace_rtx (SET_DEST (set), SET_SRC (set),
867 (reg_single_usage, regno)))
869 /* Replace any usage in a REG_EQUAL note. Must copy the
870 new source, so that we don't get rtx sharing between the
871 SET_SOURCE and REG_NOTES of insn p. */
872 REG_NOTES (VARRAY_RTX (reg_single_usage, regno))
873 = replace_rtx (REG_NOTES (VARRAY_RTX
874 (reg_single_usage, regno)),
875 SET_DEST (set), copy_rtx (SET_SRC (set)));
878 NOTE_LINE_NUMBER (p) = NOTE_INSN_DELETED;
879 NOTE_SOURCE_FILE (p) = 0;
880 VARRAY_INT (set_in_loop, regno) = 0;
884 m = (struct movable *) alloca (sizeof (struct movable));
888 m->dependencies = dependencies;
889 m->set_dest = SET_DEST (set);
891 m->consec = VARRAY_INT (set_in_loop,
892 REGNO (SET_DEST (set))) - 1;
896 m->move_insn = move_insn;
897 m->move_insn_first = 0;
898 m->is_equiv = (find_reg_note (p, REG_EQUIV, NULL_RTX) != 0);
899 m->savemode = VOIDmode;
901 /* Set M->cond if either loop_invariant_p
902 or consec_sets_invariant_p returned 2
903 (only conditionally invariant). */
904 m->cond = ((tem | tem1 | tem2) > 1);
905 m->global = (uid_luid[REGNO_LAST_UID (regno)]
906 > INSN_LUID (loop_end)
907 || uid_luid[REGNO_FIRST_UID (regno)] < INSN_LUID (loop_start));
909 m->lifetime = (uid_luid[REGNO_LAST_UID (regno)]
910 - uid_luid[REGNO_FIRST_UID (regno)]);
911 m->savings = VARRAY_INT (n_times_set, regno);
912 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
913 m->savings += libcall_benefit (p);
914 VARRAY_INT (set_in_loop, regno) = move_insn ? -2 : -1;
915 /* Add M to the end of the chain MOVABLES. */
919 last_movable->next = m;
924 /* It is possible for the first instruction to have a
925 REG_EQUAL note but a non-invariant SET_SRC, so we must
926 remember the status of the first instruction in case
927 the last instruction doesn't have a REG_EQUAL note. */
928 m->move_insn_first = m->move_insn;
930 /* Skip this insn, not checking REG_LIBCALL notes. */
931 p = next_nonnote_insn (p);
932 /* Skip the consecutive insns, if there are any. */
933 p = skip_consec_insns (p, m->consec);
934 /* Back up to the last insn of the consecutive group. */
935 p = prev_nonnote_insn (p);
937 /* We must now reset m->move_insn, m->is_equiv, and possibly
938 m->set_src to correspond to the effects of all the
940 temp = find_reg_note (p, REG_EQUIV, NULL_RTX);
942 m->set_src = XEXP (temp, 0), m->move_insn = 1;
945 temp = find_reg_note (p, REG_EQUAL, NULL_RTX);
946 if (temp && CONSTANT_P (XEXP (temp, 0)))
947 m->set_src = XEXP (temp, 0), m->move_insn = 1;
952 m->is_equiv = (find_reg_note (p, REG_EQUIV, NULL_RTX) != 0);
955 /* If this register is always set within a STRICT_LOW_PART
956 or set to zero, then its high bytes are constant.
957 So clear them outside the loop and within the loop
958 just load the low bytes.
959 We must check that the machine has an instruction to do so.
960 Also, if the value loaded into the register
961 depends on the same register, this cannot be done. */
962 else if (SET_SRC (set) == const0_rtx
963 && GET_CODE (NEXT_INSN (p)) == INSN
964 && (set1 = single_set (NEXT_INSN (p)))
965 && GET_CODE (set1) == SET
966 && (GET_CODE (SET_DEST (set1)) == STRICT_LOW_PART)
967 && (GET_CODE (XEXP (SET_DEST (set1), 0)) == SUBREG)
968 && (SUBREG_REG (XEXP (SET_DEST (set1), 0))
970 && !reg_mentioned_p (SET_DEST (set), SET_SRC (set1)))
972 register int regno = REGNO (SET_DEST (set));
973 if (VARRAY_INT (set_in_loop, regno) == 2)
975 register struct movable *m;
976 m = (struct movable *) alloca (sizeof (struct movable));
979 m->set_dest = SET_DEST (set);
986 m->move_insn_first = 0;
988 /* If the insn may not be executed on some cycles,
989 we can't clear the whole reg; clear just high part.
990 Not even if the reg is used only within this loop.
997 Clearing x before the inner loop could clobber a value
998 being saved from the last time around the outer loop.
999 However, if the reg is not used outside this loop
1000 and all uses of the register are in the same
1001 basic block as the store, there is no problem.
1003 If this insn was made by loop, we don't know its
1004 INSN_LUID and hence must make a conservative
1006 m->global = (INSN_UID (p) >= max_uid_for_loop
1007 || (uid_luid[REGNO_LAST_UID (regno)]
1008 > INSN_LUID (loop_end))
1009 || (uid_luid[REGNO_FIRST_UID (regno)]
1011 || (labels_in_range_p
1012 (p, uid_luid[REGNO_FIRST_UID (regno)])));
1013 if (maybe_never && m->global)
1014 m->savemode = GET_MODE (SET_SRC (set1));
1016 m->savemode = VOIDmode;
1020 m->lifetime = (uid_luid[REGNO_LAST_UID (regno)]
1021 - uid_luid[REGNO_FIRST_UID (regno)]);
1023 VARRAY_INT (set_in_loop, regno) = -1;
1024 /* Add M to the end of the chain MOVABLES. */
1028 last_movable->next = m;
1033 /* Past a call insn, we get to insns which might not be executed
1034 because the call might exit. This matters for insns that trap.
1035 Constant and pure call insns always return, so they don't count. */
1036 else if (GET_CODE (p) == CALL_INSN && ! CONST_CALL_P (p))
1038 /* Past a label or a jump, we get to insns for which we
1039 can't count on whether or how many times they will be
1040 executed during each iteration. Therefore, we can
1041 only move out sets of trivial variables
1042 (those not used after the loop). */
1043 /* Similar code appears twice in strength_reduce. */
1044 else if ((GET_CODE (p) == CODE_LABEL || GET_CODE (p) == JUMP_INSN)
1045 /* If we enter the loop in the middle, and scan around to the
1046 beginning, don't set maybe_never for that. This must be an
1047 unconditional jump, otherwise the code at the top of the
1048 loop might never be executed. Unconditional jumps are
1049 followed a by barrier then loop end. */
1050 && ! (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) == loop->top
1051 && NEXT_INSN (NEXT_INSN (p)) == loop_end
1052 && any_uncondjump_p (p)))
1054 else if (GET_CODE (p) == NOTE)
1056 /* At the virtual top of a converted loop, insns are again known to
1057 be executed: logically, the loop begins here even though the exit
1058 code has been duplicated. */
1059 if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP && loop_depth == 0)
1060 maybe_never = call_passed = 0;
1061 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
1063 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END)
1068 /* If one movable subsumes another, ignore that other. */
1070 ignore_some_movables (movables);
1072 /* For each movable insn, see if the reg that it loads
1073 leads when it dies right into another conditionally movable insn.
1074 If so, record that the second insn "forces" the first one,
1075 since the second can be moved only if the first is. */
1077 force_movables (movables);
1079 /* See if there are multiple movable insns that load the same value.
1080 If there are, make all but the first point at the first one
1081 through the `match' field, and add the priorities of them
1082 all together as the priority of the first. */
1084 combine_movables (movables, nregs);
1086 /* Now consider each movable insn to decide whether it is worth moving.
1087 Store 0 in set_in_loop for each reg that is moved.
1089 Generally this increases code size, so do not move moveables when
1090 optimizing for code size. */
1092 if (! optimize_size)
1093 move_movables (loop, movables, threshold, insn_count, nregs);
1095 /* Now candidates that still are negative are those not moved.
1096 Change set_in_loop to indicate that those are not actually invariant. */
1097 for (i = 0; i < nregs; i++)
1098 if (VARRAY_INT (set_in_loop, i) < 0)
1099 VARRAY_INT (set_in_loop, i) = VARRAY_INT (n_times_set, i);
1101 /* Now that we've moved some things out of the loop, we might be able to
1102 hoist even more memory references. */
1103 load_mems_and_recount_loop_regs_set (loop, &insn_count);
1105 for (update_start = loop_start;
1106 PREV_INSN (update_start)
1107 && GET_CODE (PREV_INSN (update_start)) != CODE_LABEL;
1108 update_start = PREV_INSN (update_start))
1110 update_end = NEXT_INSN (loop_end);
1112 reg_scan_update (update_start, update_end, loop_max_reg);
1113 loop_max_reg = max_reg_num ();
1115 if (flag_strength_reduce)
1117 if (update_end && GET_CODE (update_end) == CODE_LABEL)
1118 /* Ensure our label doesn't go away. */
1119 LABEL_NUSES (update_end)++;
1121 the_movables = movables;
1122 strength_reduce (loop, insn_count, flags);
1124 reg_scan_update (update_start, update_end, loop_max_reg);
1125 loop_max_reg = max_reg_num ();
1127 if (update_end && GET_CODE (update_end) == CODE_LABEL
1128 && --LABEL_NUSES (update_end) == 0)
1129 delete_insn (update_end);
1132 VARRAY_FREE (reg_single_usage);
1133 VARRAY_FREE (set_in_loop);
1134 VARRAY_FREE (n_times_set);
1135 VARRAY_FREE (may_not_optimize);
1138 /* Add elements to *OUTPUT to record all the pseudo-regs
1139 mentioned in IN_THIS but not mentioned in NOT_IN_THIS. */
1142 record_excess_regs (in_this, not_in_this, output)
1143 rtx in_this, not_in_this;
1150 code = GET_CODE (in_this);
1164 if (REGNO (in_this) >= FIRST_PSEUDO_REGISTER
1165 && ! reg_mentioned_p (in_this, not_in_this))
1166 *output = gen_rtx_EXPR_LIST (VOIDmode, in_this, *output);
1173 fmt = GET_RTX_FORMAT (code);
1174 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1181 for (j = 0; j < XVECLEN (in_this, i); j++)
1182 record_excess_regs (XVECEXP (in_this, i, j), not_in_this, output);
1186 record_excess_regs (XEXP (in_this, i), not_in_this, output);
1192 /* Check what regs are referred to in the libcall block ending with INSN,
1193 aside from those mentioned in the equivalent value.
1194 If there are none, return 0.
1195 If there are one or more, return an EXPR_LIST containing all of them. */
1198 libcall_other_reg (insn, equiv)
1201 rtx note = find_reg_note (insn, REG_RETVAL, NULL_RTX);
1202 rtx p = XEXP (note, 0);
1205 /* First, find all the regs used in the libcall block
1206 that are not mentioned as inputs to the result. */
1210 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
1211 || GET_CODE (p) == CALL_INSN)
1212 record_excess_regs (PATTERN (p), equiv, &output);
1219 /* Return 1 if all uses of REG
1220 are between INSN and the end of the basic block. */
1223 reg_in_basic_block_p (insn, reg)
1226 int regno = REGNO (reg);
1229 if (REGNO_FIRST_UID (regno) != INSN_UID (insn))
1232 /* Search this basic block for the already recorded last use of the reg. */
1233 for (p = insn; p; p = NEXT_INSN (p))
1235 switch (GET_CODE (p))
1242 /* Ordinary insn: if this is the last use, we win. */
1243 if (REGNO_LAST_UID (regno) == INSN_UID (p))
1248 /* Jump insn: if this is the last use, we win. */
1249 if (REGNO_LAST_UID (regno) == INSN_UID (p))
1251 /* Otherwise, it's the end of the basic block, so we lose. */
1256 /* It's the end of the basic block, so we lose. */
1264 /* The "last use" that was recorded can't be found after the first
1265 use. This can happen when the last use was deleted while
1266 processing an inner loop, this inner loop was then completely
1267 unrolled, and the outer loop is always exited after the inner loop,
1268 so that everything after the first use becomes a single basic block. */
1272 /* Compute the benefit of eliminating the insns in the block whose
1273 last insn is LAST. This may be a group of insns used to compute a
1274 value directly or can contain a library call. */
1277 libcall_benefit (last)
1283 for (insn = XEXP (find_reg_note (last, REG_RETVAL, NULL_RTX), 0);
1284 insn != last; insn = NEXT_INSN (insn))
1286 if (GET_CODE (insn) == CALL_INSN)
1287 benefit += 10; /* Assume at least this many insns in a library
1289 else if (GET_CODE (insn) == INSN
1290 && GET_CODE (PATTERN (insn)) != USE
1291 && GET_CODE (PATTERN (insn)) != CLOBBER)
1298 /* Skip COUNT insns from INSN, counting library calls as 1 insn. */
1301 skip_consec_insns (insn, count)
1305 for (; count > 0; count--)
1309 /* If first insn of libcall sequence, skip to end. */
1310 /* Do this at start of loop, since INSN is guaranteed to
1312 if (GET_CODE (insn) != NOTE
1313 && (temp = find_reg_note (insn, REG_LIBCALL, NULL_RTX)))
1314 insn = XEXP (temp, 0);
1317 insn = NEXT_INSN (insn);
1318 while (GET_CODE (insn) == NOTE);
1324 /* Ignore any movable whose insn falls within a libcall
1325 which is part of another movable.
1326 We make use of the fact that the movable for the libcall value
1327 was made later and so appears later on the chain. */
1330 ignore_some_movables (movables)
1331 struct movable *movables;
1333 register struct movable *m, *m1;
1335 for (m = movables; m; m = m->next)
1337 /* Is this a movable for the value of a libcall? */
1338 rtx note = find_reg_note (m->insn, REG_RETVAL, NULL_RTX);
1342 /* Check for earlier movables inside that range,
1343 and mark them invalid. We cannot use LUIDs here because
1344 insns created by loop.c for prior loops don't have LUIDs.
1345 Rather than reject all such insns from movables, we just
1346 explicitly check each insn in the libcall (since invariant
1347 libcalls aren't that common). */
1348 for (insn = XEXP (note, 0); insn != m->insn; insn = NEXT_INSN (insn))
1349 for (m1 = movables; m1 != m; m1 = m1->next)
1350 if (m1->insn == insn)
1356 /* For each movable insn, see if the reg that it loads
1357 leads when it dies right into another conditionally movable insn.
1358 If so, record that the second insn "forces" the first one,
1359 since the second can be moved only if the first is. */
1362 force_movables (movables)
1363 struct movable *movables;
1365 register struct movable *m, *m1;
1366 for (m1 = movables; m1; m1 = m1->next)
1367 /* Omit this if moving just the (SET (REG) 0) of a zero-extend. */
1368 if (!m1->partial && !m1->done)
1370 int regno = m1->regno;
1371 for (m = m1->next; m; m = m->next)
1372 /* ??? Could this be a bug? What if CSE caused the
1373 register of M1 to be used after this insn?
1374 Since CSE does not update regno_last_uid,
1375 this insn M->insn might not be where it dies.
1376 But very likely this doesn't matter; what matters is
1377 that M's reg is computed from M1's reg. */
1378 if (INSN_UID (m->insn) == REGNO_LAST_UID (regno)
1381 if (m != 0 && m->set_src == m1->set_dest
1382 /* If m->consec, m->set_src isn't valid. */
1386 /* Increase the priority of the moving the first insn
1387 since it permits the second to be moved as well. */
1391 m1->lifetime += m->lifetime;
1392 m1->savings += m->savings;
1397 /* Find invariant expressions that are equal and can be combined into
1401 combine_movables (movables, nregs)
1402 struct movable *movables;
1405 register struct movable *m;
1406 char *matched_regs = (char *) xmalloc (nregs);
1407 enum machine_mode mode;
1409 /* Regs that are set more than once are not allowed to match
1410 or be matched. I'm no longer sure why not. */
1411 /* Perhaps testing m->consec_sets would be more appropriate here? */
1413 for (m = movables; m; m = m->next)
1414 if (m->match == 0 && VARRAY_INT (n_times_set, m->regno) == 1
1417 register struct movable *m1;
1418 int regno = m->regno;
1420 bzero (matched_regs, nregs);
1421 matched_regs[regno] = 1;
1423 /* We want later insns to match the first one. Don't make the first
1424 one match any later ones. So start this loop at m->next. */
1425 for (m1 = m->next; m1; m1 = m1->next)
1426 if (m != m1 && m1->match == 0 && VARRAY_INT (n_times_set, m1->regno) == 1
1427 /* A reg used outside the loop mustn't be eliminated. */
1429 /* A reg used for zero-extending mustn't be eliminated. */
1431 && (matched_regs[m1->regno]
1434 /* Can combine regs with different modes loaded from the
1435 same constant only if the modes are the same or
1436 if both are integer modes with M wider or the same
1437 width as M1. The check for integer is redundant, but
1438 safe, since the only case of differing destination
1439 modes with equal sources is when both sources are
1440 VOIDmode, i.e., CONST_INT. */
1441 (GET_MODE (m->set_dest) == GET_MODE (m1->set_dest)
1442 || (GET_MODE_CLASS (GET_MODE (m->set_dest)) == MODE_INT
1443 && GET_MODE_CLASS (GET_MODE (m1->set_dest)) == MODE_INT
1444 && (GET_MODE_BITSIZE (GET_MODE (m->set_dest))
1445 >= GET_MODE_BITSIZE (GET_MODE (m1->set_dest)))))
1446 /* See if the source of M1 says it matches M. */
1447 && ((GET_CODE (m1->set_src) == REG
1448 && matched_regs[REGNO (m1->set_src)])
1449 || rtx_equal_for_loop_p (m->set_src, m1->set_src,
1451 && ((m->dependencies == m1->dependencies)
1452 || rtx_equal_p (m->dependencies, m1->dependencies)))
1454 m->lifetime += m1->lifetime;
1455 m->savings += m1->savings;
1458 matched_regs[m1->regno] = 1;
1462 /* Now combine the regs used for zero-extension.
1463 This can be done for those not marked `global'
1464 provided their lives don't overlap. */
1466 for (mode = GET_CLASS_NARROWEST_MODE (MODE_INT); mode != VOIDmode;
1467 mode = GET_MODE_WIDER_MODE (mode))
1469 register struct movable *m0 = 0;
1471 /* Combine all the registers for extension from mode MODE.
1472 Don't combine any that are used outside this loop. */
1473 for (m = movables; m; m = m->next)
1474 if (m->partial && ! m->global
1475 && mode == GET_MODE (SET_SRC (PATTERN (NEXT_INSN (m->insn)))))
1477 register struct movable *m1;
1478 int first = uid_luid[REGNO_FIRST_UID (m->regno)];
1479 int last = uid_luid[REGNO_LAST_UID (m->regno)];
1483 /* First one: don't check for overlap, just record it. */
1488 /* Make sure they extend to the same mode.
1489 (Almost always true.) */
1490 if (GET_MODE (m->set_dest) != GET_MODE (m0->set_dest))
1493 /* We already have one: check for overlap with those
1494 already combined together. */
1495 for (m1 = movables; m1 != m; m1 = m1->next)
1496 if (m1 == m0 || (m1->partial && m1->match == m0))
1497 if (! (uid_luid[REGNO_FIRST_UID (m1->regno)] > last
1498 || uid_luid[REGNO_LAST_UID (m1->regno)] < first))
1501 /* No overlap: we can combine this with the others. */
1502 m0->lifetime += m->lifetime;
1503 m0->savings += m->savings;
1513 free (matched_regs);
1516 /* Return 1 if regs X and Y will become the same if moved. */
1519 regs_match_p (x, y, movables)
1521 struct movable *movables;
1523 unsigned int xn = REGNO (x);
1524 unsigned int yn = REGNO (y);
1525 struct movable *mx, *my;
1527 for (mx = movables; mx; mx = mx->next)
1528 if (mx->regno == xn)
1531 for (my = movables; my; my = my->next)
1532 if (my->regno == yn)
1536 && ((mx->match == my->match && mx->match != 0)
1538 || mx == my->match));
1541 /* Return 1 if X and Y are identical-looking rtx's.
1542 This is the Lisp function EQUAL for rtx arguments.
1544 If two registers are matching movables or a movable register and an
1545 equivalent constant, consider them equal. */
1548 rtx_equal_for_loop_p (x, y, movables)
1550 struct movable *movables;
1554 register struct movable *m;
1555 register enum rtx_code code;
1556 register const char *fmt;
1560 if (x == 0 || y == 0)
1563 code = GET_CODE (x);
1565 /* If we have a register and a constant, they may sometimes be
1567 if (GET_CODE (x) == REG && VARRAY_INT (set_in_loop, REGNO (x)) == -2
1570 for (m = movables; m; m = m->next)
1571 if (m->move_insn && m->regno == REGNO (x)
1572 && rtx_equal_p (m->set_src, y))
1575 else if (GET_CODE (y) == REG && VARRAY_INT (set_in_loop, REGNO (y)) == -2
1578 for (m = movables; m; m = m->next)
1579 if (m->move_insn && m->regno == REGNO (y)
1580 && rtx_equal_p (m->set_src, x))
1584 /* Otherwise, rtx's of different codes cannot be equal. */
1585 if (code != GET_CODE (y))
1588 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent.
1589 (REG:SI x) and (REG:HI x) are NOT equivalent. */
1591 if (GET_MODE (x) != GET_MODE (y))
1594 /* These three types of rtx's can be compared nonrecursively. */
1596 return (REGNO (x) == REGNO (y) || regs_match_p (x, y, movables));
1598 if (code == LABEL_REF)
1599 return XEXP (x, 0) == XEXP (y, 0);
1600 if (code == SYMBOL_REF)
1601 return XSTR (x, 0) == XSTR (y, 0);
1603 /* Compare the elements. If any pair of corresponding elements
1604 fail to match, return 0 for the whole things. */
1606 fmt = GET_RTX_FORMAT (code);
1607 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1612 if (XWINT (x, i) != XWINT (y, i))
1617 if (XINT (x, i) != XINT (y, i))
1622 /* Two vectors must have the same length. */
1623 if (XVECLEN (x, i) != XVECLEN (y, i))
1626 /* And the corresponding elements must match. */
1627 for (j = 0; j < XVECLEN (x, i); j++)
1628 if (rtx_equal_for_loop_p (XVECEXP (x, i, j), XVECEXP (y, i, j), movables) == 0)
1633 if (rtx_equal_for_loop_p (XEXP (x, i), XEXP (y, i), movables) == 0)
1638 if (strcmp (XSTR (x, i), XSTR (y, i)))
1643 /* These are just backpointers, so they don't matter. */
1649 /* It is believed that rtx's at this level will never
1650 contain anything but integers and other rtx's,
1651 except for within LABEL_REFs and SYMBOL_REFs. */
1659 /* If X contains any LABEL_REF's, add REG_LABEL notes for them to all
1660 insns in INSNS which use the reference. */
1663 add_label_notes (x, insns)
1667 enum rtx_code code = GET_CODE (x);
1672 if (code == LABEL_REF && !LABEL_REF_NONLOCAL_P (x))
1674 /* This code used to ignore labels that referred to dispatch tables to
1675 avoid flow generating (slighly) worse code.
1677 We no longer ignore such label references (see LABEL_REF handling in
1678 mark_jump_label for additional information). */
1679 for (insn = insns; insn; insn = NEXT_INSN (insn))
1680 if (reg_mentioned_p (XEXP (x, 0), insn))
1681 REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_LABEL, XEXP (x, 0),
1685 fmt = GET_RTX_FORMAT (code);
1686 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1689 add_label_notes (XEXP (x, i), insns);
1690 else if (fmt[i] == 'E')
1691 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
1692 add_label_notes (XVECEXP (x, i, j), insns);
1696 /* Scan MOVABLES, and move the insns that deserve to be moved.
1697 If two matching movables are combined, replace one reg with the
1698 other throughout. */
1701 move_movables (loop, movables, threshold, insn_count, nregs)
1703 struct movable *movables;
1709 register struct movable *m;
1711 rtx loop_start = loop->start;
1712 rtx loop_end = loop->end;
1713 /* Map of pseudo-register replacements to handle combining
1714 when we move several insns that load the same value
1715 into different pseudo-registers. */
1716 rtx *reg_map = (rtx *) xcalloc (nregs, sizeof (rtx));
1717 char *already_moved = (char *) xcalloc (nregs, sizeof (char));
1721 for (m = movables; m; m = m->next)
1723 /* Describe this movable insn. */
1725 if (loop_dump_stream)
1727 fprintf (loop_dump_stream, "Insn %d: regno %d (life %d), ",
1728 INSN_UID (m->insn), m->regno, m->lifetime);
1730 fprintf (loop_dump_stream, "consec %d, ", m->consec);
1732 fprintf (loop_dump_stream, "cond ");
1734 fprintf (loop_dump_stream, "force ");
1736 fprintf (loop_dump_stream, "global ");
1738 fprintf (loop_dump_stream, "done ");
1740 fprintf (loop_dump_stream, "move-insn ");
1742 fprintf (loop_dump_stream, "matches %d ",
1743 INSN_UID (m->match->insn));
1745 fprintf (loop_dump_stream, "forces %d ",
1746 INSN_UID (m->forces->insn));
1749 /* Count movables. Value used in heuristics in strength_reduce. */
1752 /* Ignore the insn if it's already done (it matched something else).
1753 Otherwise, see if it is now safe to move. */
1757 || (1 == loop_invariant_p (loop, m->set_src)
1758 && (m->dependencies == 0
1759 || 1 == loop_invariant_p (loop, m->dependencies))
1761 || 1 == consec_sets_invariant_p (loop, m->set_dest,
1764 && (! m->forces || m->forces->done))
1768 int savings = m->savings;
1770 /* We have an insn that is safe to move.
1771 Compute its desirability. */
1776 if (loop_dump_stream)
1777 fprintf (loop_dump_stream, "savings %d ", savings);
1779 if (moved_once[regno] && loop_dump_stream)
1780 fprintf (loop_dump_stream, "halved since already moved ");
1782 /* An insn MUST be moved if we already moved something else
1783 which is safe only if this one is moved too: that is,
1784 if already_moved[REGNO] is nonzero. */
1786 /* An insn is desirable to move if the new lifetime of the
1787 register is no more than THRESHOLD times the old lifetime.
1788 If it's not desirable, it means the loop is so big
1789 that moving won't speed things up much,
1790 and it is liable to make register usage worse. */
1792 /* It is also desirable to move if it can be moved at no
1793 extra cost because something else was already moved. */
1795 if (already_moved[regno]
1796 || flag_move_all_movables
1797 || (threshold * savings * m->lifetime) >=
1798 (moved_once[regno] ? insn_count * 2 : insn_count)
1799 || (m->forces && m->forces->done
1800 && VARRAY_INT (n_times_set, m->forces->regno) == 1))
1803 register struct movable *m1;
1804 rtx first = NULL_RTX;
1806 /* Now move the insns that set the reg. */
1808 if (m->partial && m->match)
1812 /* Find the end of this chain of matching regs.
1813 Thus, we load each reg in the chain from that one reg.
1814 And that reg is loaded with 0 directly,
1815 since it has ->match == 0. */
1816 for (m1 = m; m1->match; m1 = m1->match);
1817 newpat = gen_move_insn (SET_DEST (PATTERN (m->insn)),
1818 SET_DEST (PATTERN (m1->insn)));
1819 i1 = emit_insn_before (newpat, loop_start);
1821 /* Mark the moved, invariant reg as being allowed to
1822 share a hard reg with the other matching invariant. */
1823 REG_NOTES (i1) = REG_NOTES (m->insn);
1824 r1 = SET_DEST (PATTERN (m->insn));
1825 r2 = SET_DEST (PATTERN (m1->insn));
1827 = gen_rtx_EXPR_LIST (VOIDmode, r1,
1828 gen_rtx_EXPR_LIST (VOIDmode, r2,
1830 delete_insn (m->insn);
1835 if (loop_dump_stream)
1836 fprintf (loop_dump_stream, " moved to %d", INSN_UID (i1));
1838 /* If we are to re-generate the item being moved with a
1839 new move insn, first delete what we have and then emit
1840 the move insn before the loop. */
1841 else if (m->move_insn)
1845 for (count = m->consec; count >= 0; count--)
1847 /* If this is the first insn of a library call sequence,
1849 if (GET_CODE (p) != NOTE
1850 && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
1853 /* If this is the last insn of a libcall sequence, then
1854 delete every insn in the sequence except the last.
1855 The last insn is handled in the normal manner. */
1856 if (GET_CODE (p) != NOTE
1857 && (temp = find_reg_note (p, REG_RETVAL, NULL_RTX)))
1859 temp = XEXP (temp, 0);
1861 temp = delete_insn (temp);
1865 p = delete_insn (p);
1867 /* simplify_giv_expr expects that it can walk the insns
1868 at m->insn forwards and see this old sequence we are
1869 tossing here. delete_insn does preserve the next
1870 pointers, but when we skip over a NOTE we must fix
1871 it up. Otherwise that code walks into the non-deleted
1873 while (p && GET_CODE (p) == NOTE)
1874 p = NEXT_INSN (temp) = NEXT_INSN (p);
1878 emit_move_insn (m->set_dest, m->set_src);
1879 temp = get_insns ();
1882 add_label_notes (m->set_src, temp);
1884 i1 = emit_insns_before (temp, loop_start);
1885 if (! find_reg_note (i1, REG_EQUAL, NULL_RTX))
1887 = gen_rtx_EXPR_LIST (m->is_equiv ? REG_EQUIV : REG_EQUAL,
1888 m->set_src, REG_NOTES (i1));
1890 if (loop_dump_stream)
1891 fprintf (loop_dump_stream, " moved to %d", INSN_UID (i1));
1893 /* The more regs we move, the less we like moving them. */
1898 for (count = m->consec; count >= 0; count--)
1902 /* If first insn of libcall sequence, skip to end. */
1903 /* Do this at start of loop, since p is guaranteed to
1905 if (GET_CODE (p) != NOTE
1906 && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
1909 /* If last insn of libcall sequence, move all
1910 insns except the last before the loop. The last
1911 insn is handled in the normal manner. */
1912 if (GET_CODE (p) != NOTE
1913 && (temp = find_reg_note (p, REG_RETVAL, NULL_RTX)))
1917 rtx fn_address_insn = 0;
1920 for (temp = XEXP (temp, 0); temp != p;
1921 temp = NEXT_INSN (temp))
1927 if (GET_CODE (temp) == NOTE)
1930 body = PATTERN (temp);
1932 /* Find the next insn after TEMP,
1933 not counting USE or NOTE insns. */
1934 for (next = NEXT_INSN (temp); next != p;
1935 next = NEXT_INSN (next))
1936 if (! (GET_CODE (next) == INSN
1937 && GET_CODE (PATTERN (next)) == USE)
1938 && GET_CODE (next) != NOTE)
1941 /* If that is the call, this may be the insn
1942 that loads the function address.
1944 Extract the function address from the insn
1945 that loads it into a register.
1946 If this insn was cse'd, we get incorrect code.
1948 So emit a new move insn that copies the
1949 function address into the register that the
1950 call insn will use. flow.c will delete any
1951 redundant stores that we have created. */
1952 if (GET_CODE (next) == CALL_INSN
1953 && GET_CODE (body) == SET
1954 && GET_CODE (SET_DEST (body)) == REG
1955 && (n = find_reg_note (temp, REG_EQUAL,
1958 fn_reg = SET_SRC (body);
1959 if (GET_CODE (fn_reg) != REG)
1960 fn_reg = SET_DEST (body);
1961 fn_address = XEXP (n, 0);
1962 fn_address_insn = temp;
1964 /* We have the call insn.
1965 If it uses the register we suspect it might,
1966 load it with the correct address directly. */
1967 if (GET_CODE (temp) == CALL_INSN
1969 && reg_referenced_p (fn_reg, body))
1970 emit_insn_after (gen_move_insn (fn_reg,
1974 if (GET_CODE (temp) == CALL_INSN)
1976 i1 = emit_call_insn_before (body, loop_start);
1977 /* Because the USAGE information potentially
1978 contains objects other than hard registers
1979 we need to copy it. */
1980 if (CALL_INSN_FUNCTION_USAGE (temp))
1981 CALL_INSN_FUNCTION_USAGE (i1)
1982 = copy_rtx (CALL_INSN_FUNCTION_USAGE (temp));
1985 i1 = emit_insn_before (body, loop_start);
1988 if (temp == fn_address_insn)
1989 fn_address_insn = i1;
1990 REG_NOTES (i1) = REG_NOTES (temp);
1996 if (m->savemode != VOIDmode)
1998 /* P sets REG to zero; but we should clear only
1999 the bits that are not covered by the mode
2001 rtx reg = m->set_dest;
2007 (GET_MODE (reg), and_optab, reg,
2008 GEN_INT ((((HOST_WIDE_INT) 1
2009 << GET_MODE_BITSIZE (m->savemode)))
2011 reg, 1, OPTAB_LIB_WIDEN);
2015 emit_move_insn (reg, tem);
2016 sequence = gen_sequence ();
2018 i1 = emit_insn_before (sequence, loop_start);
2020 else if (GET_CODE (p) == CALL_INSN)
2022 i1 = emit_call_insn_before (PATTERN (p), loop_start);
2023 /* Because the USAGE information potentially
2024 contains objects other than hard registers
2025 we need to copy it. */
2026 if (CALL_INSN_FUNCTION_USAGE (p))
2027 CALL_INSN_FUNCTION_USAGE (i1)
2028 = copy_rtx (CALL_INSN_FUNCTION_USAGE (p));
2030 else if (count == m->consec && m->move_insn_first)
2032 /* The SET_SRC might not be invariant, so we must
2033 use the REG_EQUAL note. */
2035 emit_move_insn (m->set_dest, m->set_src);
2036 temp = get_insns ();
2039 add_label_notes (m->set_src, temp);
2041 i1 = emit_insns_before (temp, loop_start);
2042 if (! find_reg_note (i1, REG_EQUAL, NULL_RTX))
2044 = gen_rtx_EXPR_LIST ((m->is_equiv ? REG_EQUIV
2046 m->set_src, REG_NOTES (i1));
2049 i1 = emit_insn_before (PATTERN (p), loop_start);
2051 if (REG_NOTES (i1) == 0)
2053 REG_NOTES (i1) = REG_NOTES (p);
2055 /* If there is a REG_EQUAL note present whose value
2056 is not loop invariant, then delete it, since it
2057 may cause problems with later optimization passes.
2058 It is possible for cse to create such notes
2059 like this as a result of record_jump_cond. */
2061 if ((temp = find_reg_note (i1, REG_EQUAL, NULL_RTX))
2062 && ! loop_invariant_p (loop, XEXP (temp, 0)))
2063 remove_note (i1, temp);
2069 if (loop_dump_stream)
2070 fprintf (loop_dump_stream, " moved to %d",
2073 /* If library call, now fix the REG_NOTES that contain
2074 insn pointers, namely REG_LIBCALL on FIRST
2075 and REG_RETVAL on I1. */
2076 if ((temp = find_reg_note (i1, REG_RETVAL, NULL_RTX)))
2078 XEXP (temp, 0) = first;
2079 temp = find_reg_note (first, REG_LIBCALL, NULL_RTX);
2080 XEXP (temp, 0) = i1;
2087 /* simplify_giv_expr expects that it can walk the insns
2088 at m->insn forwards and see this old sequence we are
2089 tossing here. delete_insn does preserve the next
2090 pointers, but when we skip over a NOTE we must fix
2091 it up. Otherwise that code walks into the non-deleted
2093 while (p && GET_CODE (p) == NOTE)
2094 p = NEXT_INSN (temp) = NEXT_INSN (p);
2097 /* The more regs we move, the less we like moving them. */
2101 /* Any other movable that loads the same register
2103 already_moved[regno] = 1;
2105 /* This reg has been moved out of one loop. */
2106 moved_once[regno] = 1;
2108 /* The reg set here is now invariant. */
2110 VARRAY_INT (set_in_loop, regno) = 0;
2114 /* Change the length-of-life info for the register
2115 to say it lives at least the full length of this loop.
2116 This will help guide optimizations in outer loops. */
2118 if (uid_luid[REGNO_FIRST_UID (regno)] > INSN_LUID (loop_start))
2119 /* This is the old insn before all the moved insns.
2120 We can't use the moved insn because it is out of range
2121 in uid_luid. Only the old insns have luids. */
2122 REGNO_FIRST_UID (regno) = INSN_UID (loop_start);
2123 if (uid_luid[REGNO_LAST_UID (regno)] < INSN_LUID (loop_end))
2124 REGNO_LAST_UID (regno) = INSN_UID (loop_end);
2126 /* Combine with this moved insn any other matching movables. */
2129 for (m1 = movables; m1; m1 = m1->next)
2134 /* Schedule the reg loaded by M1
2135 for replacement so that shares the reg of M.
2136 If the modes differ (only possible in restricted
2137 circumstances, make a SUBREG.
2139 Note this assumes that the target dependent files
2140 treat REG and SUBREG equally, including within
2141 GO_IF_LEGITIMATE_ADDRESS and in all the
2142 predicates since we never verify that replacing the
2143 original register with a SUBREG results in a
2144 recognizable insn. */
2145 if (GET_MODE (m->set_dest) == GET_MODE (m1->set_dest))
2146 reg_map[m1->regno] = m->set_dest;
2149 = gen_lowpart_common (GET_MODE (m1->set_dest),
2152 /* Get rid of the matching insn
2153 and prevent further processing of it. */
2156 /* if library call, delete all insn except last, which
2158 if ((temp = find_reg_note (m1->insn, REG_RETVAL,
2161 for (temp = XEXP (temp, 0); temp != m1->insn;
2162 temp = NEXT_INSN (temp))
2165 delete_insn (m1->insn);
2167 /* Any other movable that loads the same register
2169 already_moved[m1->regno] = 1;
2171 /* The reg merged here is now invariant,
2172 if the reg it matches is invariant. */
2174 VARRAY_INT (set_in_loop, m1->regno) = 0;
2177 else if (loop_dump_stream)
2178 fprintf (loop_dump_stream, "not desirable");
2180 else if (loop_dump_stream && !m->match)
2181 fprintf (loop_dump_stream, "not safe");
2183 if (loop_dump_stream)
2184 fprintf (loop_dump_stream, "\n");
2188 new_start = loop_start;
2190 /* Go through all the instructions in the loop, making
2191 all the register substitutions scheduled in REG_MAP. */
2192 for (p = new_start; p != loop_end; p = NEXT_INSN (p))
2193 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
2194 || GET_CODE (p) == CALL_INSN)
2196 replace_regs (PATTERN (p), reg_map, nregs, 0);
2197 replace_regs (REG_NOTES (p), reg_map, nregs, 0);
2203 free (already_moved);
2207 /* Scan X and replace the address of any MEM in it with ADDR.
2208 REG is the address that MEM should have before the replacement. */
2211 replace_call_address (x, reg, addr)
2214 register enum rtx_code code;
2216 register const char *fmt;
2220 code = GET_CODE (x);
2234 /* Short cut for very common case. */
2235 replace_call_address (XEXP (x, 1), reg, addr);
2239 /* Short cut for very common case. */
2240 replace_call_address (XEXP (x, 0), reg, addr);
2244 /* If this MEM uses a reg other than the one we expected,
2245 something is wrong. */
2246 if (XEXP (x, 0) != reg)
2255 fmt = GET_RTX_FORMAT (code);
2256 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2259 replace_call_address (XEXP (x, i), reg, addr);
2260 else if (fmt[i] == 'E')
2263 for (j = 0; j < XVECLEN (x, i); j++)
2264 replace_call_address (XVECEXP (x, i, j), reg, addr);
2270 /* Return the number of memory refs to addresses that vary
2274 count_nonfixed_reads (loop, x)
2275 const struct loop *loop;
2278 register enum rtx_code code;
2280 register const char *fmt;
2286 code = GET_CODE (x);
2300 return ((loop_invariant_p (loop, XEXP (x, 0)) != 1)
2301 + count_nonfixed_reads (loop, XEXP (x, 0)));
2308 fmt = GET_RTX_FORMAT (code);
2309 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2312 value += count_nonfixed_reads (loop, XEXP (x, i));
2316 for (j = 0; j < XVECLEN (x, i); j++)
2317 value += count_nonfixed_reads (loop, XVECEXP (x, i, j));
2323 /* Scan a loop setting the elements `cont', `vtop', `loops_enclosed',
2324 `has_call', `has_volatile', and `has_tablejump' within LOOP.
2325 Set the global variables `unknown_address_altered',
2326 `unknown_constant_address_altered', and `num_mem_sets'. Also, fill
2327 in the array `loop_mems' and the list `loop_store_mems'. */
2333 register int level = 1;
2335 struct loop_info *loop_info = LOOP_INFO (loop);
2336 rtx start = loop->start;
2337 rtx end = loop->end;
2338 /* The label after END. Jumping here is just like falling off the
2339 end of the loop. We use next_nonnote_insn instead of next_label
2340 as a hedge against the (pathological) case where some actual insn
2341 might end up between the two. */
2342 rtx exit_target = next_nonnote_insn (end);
2344 loop_info->has_indirect_jump = indirect_jump_in_function;
2345 loop_info->has_call = 0;
2346 loop_info->has_volatile = 0;
2347 loop_info->has_tablejump = 0;
2348 loop_info->has_multiple_exit_targets = 0;
2351 unknown_address_altered = 0;
2352 unknown_constant_address_altered = 0;
2353 loop_store_mems = NULL_RTX;
2354 first_loop_store_insn = NULL_RTX;
2358 for (insn = NEXT_INSN (start); insn != NEXT_INSN (end);
2359 insn = NEXT_INSN (insn))
2361 if (GET_CODE (insn) == NOTE)
2363 if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG)
2366 /* Count number of loops contained in this one. */
2369 else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_END)
2374 else if (GET_CODE (insn) == CALL_INSN)
2376 if (! CONST_CALL_P (insn))
2377 unknown_address_altered = 1;
2378 loop_info->has_call = 1;
2380 else if (GET_CODE (insn) == INSN || GET_CODE (insn) == JUMP_INSN)
2382 rtx label1 = NULL_RTX;
2383 rtx label2 = NULL_RTX;
2385 if (volatile_refs_p (PATTERN (insn)))
2386 loop_info->has_volatile = 1;
2388 if (GET_CODE (insn) == JUMP_INSN
2389 && (GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC
2390 || GET_CODE (PATTERN (insn)) == ADDR_VEC))
2391 loop_info->has_tablejump = 1;
2393 note_stores (PATTERN (insn), note_addr_stored, NULL);
2394 if (! first_loop_store_insn && loop_store_mems)
2395 first_loop_store_insn = insn;
2397 if (! loop_info->has_multiple_exit_targets
2398 && GET_CODE (insn) == JUMP_INSN
2399 && GET_CODE (PATTERN (insn)) == SET
2400 && SET_DEST (PATTERN (insn)) == pc_rtx)
2402 if (GET_CODE (SET_SRC (PATTERN (insn))) == IF_THEN_ELSE)
2404 label1 = XEXP (SET_SRC (PATTERN (insn)), 1);
2405 label2 = XEXP (SET_SRC (PATTERN (insn)), 2);
2409 label1 = SET_SRC (PATTERN (insn));
2414 if (label1 && label1 != pc_rtx)
2416 if (GET_CODE (label1) != LABEL_REF)
2418 /* Something tricky. */
2419 loop_info->has_multiple_exit_targets = 1;
2422 else if (XEXP (label1, 0) != exit_target
2423 && LABEL_OUTSIDE_LOOP_P (label1))
2425 /* A jump outside the current loop. */
2426 loop_info->has_multiple_exit_targets = 1;
2437 else if (GET_CODE (insn) == RETURN)
2438 loop_info->has_multiple_exit_targets = 1;
2441 /* Now, rescan the loop, setting up the LOOP_MEMS array. */
2442 if (/* An exception thrown by a called function might land us
2444 ! loop_info->has_call
2445 /* We don't want loads for MEMs moved to a location before the
2446 one at which their stack memory becomes allocated. (Note
2447 that this is not a problem for malloc, etc., since those
2448 require actual function calls. */
2449 && ! current_function_calls_alloca
2450 /* There are ways to leave the loop other than falling off the
2452 && ! loop_info->has_multiple_exit_targets)
2453 for (insn = NEXT_INSN (start); insn != NEXT_INSN (end);
2454 insn = NEXT_INSN (insn))
2455 for_each_rtx (&insn, insert_loop_mem, 0);
2457 /* BLKmode MEMs are added to LOOP_STORE_MEM as necessary so
2458 that loop_invariant_p and load_mems can use true_dependence
2459 to determine what is really clobbered. */
2460 if (unknown_address_altered)
2462 rtx mem = gen_rtx_MEM (BLKmode, const0_rtx);
2464 loop_store_mems = gen_rtx_EXPR_LIST (VOIDmode, mem, loop_store_mems);
2466 if (unknown_constant_address_altered)
2468 rtx mem = gen_rtx_MEM (BLKmode, const0_rtx);
2470 RTX_UNCHANGING_P (mem) = 1;
2471 loop_store_mems = gen_rtx_EXPR_LIST (VOIDmode, mem, loop_store_mems);
2475 /* LOOP->CONT_DOMINATOR is now the last label between the loop start
2476 and the continue note that is a the destination of a (cond)jump after
2477 the continue note. If there is any (cond)jump between the loop start
2478 and what we have so far as LOOP->CONT_DOMINATOR that has a
2479 target between LOOP->CONT_DOMINATOR and the continue note, move
2480 LOOP->CONT_DOMINATOR forward to that label; if a jump's
2481 destination cannot be determined, clear LOOP->CONT_DOMINATOR. */
2484 verify_dominator (loop)
2489 if (! loop->cont_dominator)
2490 /* This can happen for an empty loop, e.g. in
2491 gcc.c-torture/compile/920410-2.c */
2493 if (loop->cont_dominator == const0_rtx)
2495 loop->cont_dominator = 0;
2498 for (insn = loop->start; insn != loop->cont_dominator;
2499 insn = NEXT_INSN (insn))
2501 if (GET_CODE (insn) == JUMP_INSN
2502 && GET_CODE (PATTERN (insn)) != RETURN)
2504 rtx label = JUMP_LABEL (insn);
2507 /* If it is not a jump we can easily understand or for
2508 which we do not have jump target information in the JUMP_LABEL
2509 field (consider ADDR_VEC and ADDR_DIFF_VEC insns), then clear
2510 LOOP->CONT_DOMINATOR. */
2511 if (! any_condjump_p (insn)
2512 || label == NULL_RTX)
2514 loop->cont_dominator = NULL_RTX;
2518 label_luid = INSN_LUID (label);
2519 if (label_luid < INSN_LUID (loop->cont)
2521 > INSN_LUID (loop->cont)))
2522 loop->cont_dominator = label;
2527 /* Scan the function looking for loops. Record the start and end of each loop.
2528 Also mark as invalid loops any loops that contain a setjmp or are branched
2529 to from outside the loop. */
2532 find_and_verify_loops (f, loops)
2534 struct loops *loops;
2539 struct loop *current_loop;
2540 struct loop *next_loop;
2543 num_loops = loops->num;
2545 compute_luids (f, NULL_RTX, 0);
2547 /* If there are jumps to undefined labels,
2548 treat them as jumps out of any/all loops.
2549 This also avoids writing past end of tables when there are no loops. */
2552 /* Find boundaries of loops, mark which loops are contained within
2553 loops, and invalidate loops that have setjmp. */
2556 current_loop = NULL;
2557 for (insn = f; insn; insn = NEXT_INSN (insn))
2559 if (GET_CODE (insn) == NOTE)
2560 switch (NOTE_LINE_NUMBER (insn))
2562 case NOTE_INSN_LOOP_BEG:
2563 next_loop = loops->array + num_loops;
2564 next_loop->num = num_loops;
2566 next_loop->start = insn;
2567 next_loop->outer = current_loop;
2568 current_loop = next_loop;
2571 case NOTE_INSN_SETJMP:
2572 /* In this case, we must invalidate our current loop and any
2574 for (loop = current_loop; loop; loop = loop->outer)
2577 if (loop_dump_stream)
2578 fprintf (loop_dump_stream,
2579 "\nLoop at %d ignored due to setjmp.\n",
2580 INSN_UID (loop->start));
2584 case NOTE_INSN_LOOP_CONT:
2585 current_loop->cont = insn;
2588 case NOTE_INSN_LOOP_VTOP:
2589 current_loop->vtop = insn;
2592 case NOTE_INSN_LOOP_END:
2596 current_loop->end = insn;
2597 verify_dominator (current_loop);
2598 current_loop = current_loop->outer;
2604 /* If for any loop, this is a jump insn between the NOTE_INSN_LOOP_CONT
2605 and NOTE_INSN_LOOP_END notes, update loop->cont_dominator. */
2606 else if (GET_CODE (insn) == JUMP_INSN
2607 && GET_CODE (PATTERN (insn)) != RETURN
2610 rtx label = JUMP_LABEL (insn);
2612 if (! any_condjump_p (insn))
2615 loop = current_loop;
2618 /* First see if we care about this loop. */
2619 if (loop->cont && loop->cont_dominator != const0_rtx)
2621 /* If the jump destination is not known, invalidate
2622 loop->cont_dominator. */
2624 loop->cont_dominator = const0_rtx;
2626 /* Check if the destination is between loop start and
2628 if ((INSN_LUID (label)
2629 < INSN_LUID (loop->cont))
2630 && (INSN_LUID (label)
2631 > INSN_LUID (loop->start))
2632 /* And if there is no later destination already
2634 && (! loop->cont_dominator
2635 || (INSN_LUID (label)
2636 > INSN_LUID (loop->cont_dominator))))
2637 loop->cont_dominator = label;
2644 /* Note that this will mark the NOTE_INSN_LOOP_END note as being in the
2645 enclosing loop, but this doesn't matter. */
2646 uid_loop[INSN_UID (insn)] = current_loop;
2649 /* Any loop containing a label used in an initializer must be invalidated,
2650 because it can be jumped into from anywhere. */
2652 for (label = forced_labels; label; label = XEXP (label, 1))
2654 for (loop = uid_loop[INSN_UID (XEXP (label, 0))];
2655 loop; loop = loop->outer)
2659 /* Any loop containing a label used for an exception handler must be
2660 invalidated, because it can be jumped into from anywhere. */
2662 for (label = exception_handler_labels; label; label = XEXP (label, 1))
2664 for (loop = uid_loop[INSN_UID (XEXP (label, 0))];
2665 loop; loop = loop->outer)
2669 /* Now scan all insn's in the function. If any JUMP_INSN branches into a
2670 loop that it is not contained within, that loop is marked invalid.
2671 If any INSN or CALL_INSN uses a label's address, then the loop containing
2672 that label is marked invalid, because it could be jumped into from
2675 Also look for blocks of code ending in an unconditional branch that
2676 exits the loop. If such a block is surrounded by a conditional
2677 branch around the block, move the block elsewhere (see below) and
2678 invert the jump to point to the code block. This may eliminate a
2679 label in our loop and will simplify processing by both us and a
2680 possible second cse pass. */
2682 for (insn = f; insn; insn = NEXT_INSN (insn))
2685 struct loop *this_loop = uid_loop[INSN_UID (insn)];
2687 if (GET_CODE (insn) == INSN || GET_CODE (insn) == CALL_INSN)
2689 rtx note = find_reg_note (insn, REG_LABEL, NULL_RTX);
2692 for (loop = uid_loop[INSN_UID (XEXP (note, 0))];
2693 loop; loop = loop->outer)
2698 if (GET_CODE (insn) != JUMP_INSN)
2701 mark_loop_jump (PATTERN (insn), this_loop);
2703 /* See if this is an unconditional branch outside the loop. */
2705 && (GET_CODE (PATTERN (insn)) == RETURN
2706 || (any_uncondjump_p (insn)
2707 && onlyjump_p (insn)
2708 && (uid_loop[INSN_UID (JUMP_LABEL (insn))]
2710 && get_max_uid () < max_uid_for_loop)
2713 rtx our_next = next_real_insn (insn);
2714 rtx last_insn_to_move = NEXT_INSN (insn);
2715 struct loop *dest_loop;
2716 struct loop *outer_loop = NULL;
2718 /* Go backwards until we reach the start of the loop, a label,
2720 for (p = PREV_INSN (insn);
2721 GET_CODE (p) != CODE_LABEL
2722 && ! (GET_CODE (p) == NOTE
2723 && NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
2724 && GET_CODE (p) != JUMP_INSN;
2728 /* Check for the case where we have a jump to an inner nested
2729 loop, and do not perform the optimization in that case. */
2731 if (JUMP_LABEL (insn))
2733 dest_loop = uid_loop[INSN_UID (JUMP_LABEL (insn))];
2736 for (outer_loop = dest_loop; outer_loop;
2737 outer_loop = outer_loop->outer)
2738 if (outer_loop == this_loop)
2743 /* Make sure that the target of P is within the current loop. */
2745 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p)
2746 && uid_loop[INSN_UID (JUMP_LABEL (p))] != this_loop)
2747 outer_loop = this_loop;
2749 /* If we stopped on a JUMP_INSN to the next insn after INSN,
2750 we have a block of code to try to move.
2752 We look backward and then forward from the target of INSN
2753 to find a BARRIER at the same loop depth as the target.
2754 If we find such a BARRIER, we make a new label for the start
2755 of the block, invert the jump in P and point it to that label,
2756 and move the block of code to the spot we found. */
2759 && GET_CODE (p) == JUMP_INSN
2760 && JUMP_LABEL (p) != 0
2761 /* Just ignore jumps to labels that were never emitted.
2762 These always indicate compilation errors. */
2763 && INSN_UID (JUMP_LABEL (p)) != 0
2764 && any_condjump_p (p) && onlyjump_p (p)
2765 && next_real_insn (JUMP_LABEL (p)) == our_next
2766 /* If it's not safe to move the sequence, then we
2768 && insns_safe_to_move_p (p, NEXT_INSN (insn),
2769 &last_insn_to_move))
2772 = JUMP_LABEL (insn) ? JUMP_LABEL (insn) : get_last_insn ();
2773 struct loop *target_loop = uid_loop[INSN_UID (target)];
2776 for (loc = target; loc; loc = PREV_INSN (loc))
2777 if (GET_CODE (loc) == BARRIER
2778 /* Don't move things inside a tablejump. */
2779 && ((loc2 = next_nonnote_insn (loc)) == 0
2780 || GET_CODE (loc2) != CODE_LABEL
2781 || (loc2 = next_nonnote_insn (loc2)) == 0
2782 || GET_CODE (loc2) != JUMP_INSN
2783 || (GET_CODE (PATTERN (loc2)) != ADDR_VEC
2784 && GET_CODE (PATTERN (loc2)) != ADDR_DIFF_VEC))
2785 && uid_loop[INSN_UID (loc)] == target_loop)
2789 for (loc = target; loc; loc = NEXT_INSN (loc))
2790 if (GET_CODE (loc) == BARRIER
2791 /* Don't move things inside a tablejump. */
2792 && ((loc2 = next_nonnote_insn (loc)) == 0
2793 || GET_CODE (loc2) != CODE_LABEL
2794 || (loc2 = next_nonnote_insn (loc2)) == 0
2795 || GET_CODE (loc2) != JUMP_INSN
2796 || (GET_CODE (PATTERN (loc2)) != ADDR_VEC
2797 && GET_CODE (PATTERN (loc2)) != ADDR_DIFF_VEC))
2798 && uid_loop[INSN_UID (loc)] == target_loop)
2803 rtx cond_label = JUMP_LABEL (p);
2804 rtx new_label = get_label_after (p);
2806 /* Ensure our label doesn't go away. */
2807 LABEL_NUSES (cond_label)++;
2809 /* Verify that uid_loop is large enough and that
2811 if (invert_jump (p, new_label, 1))
2815 /* If no suitable BARRIER was found, create a suitable
2816 one before TARGET. Since TARGET is a fall through
2817 path, we'll need to insert an jump around our block
2818 and a add a BARRIER before TARGET.
2820 This creates an extra unconditional jump outside
2821 the loop. However, the benefits of removing rarely
2822 executed instructions from inside the loop usually
2823 outweighs the cost of the extra unconditional jump
2824 outside the loop. */
2829 temp = gen_jump (JUMP_LABEL (insn));
2830 temp = emit_jump_insn_before (temp, target);
2831 JUMP_LABEL (temp) = JUMP_LABEL (insn);
2832 LABEL_NUSES (JUMP_LABEL (insn))++;
2833 loc = emit_barrier_before (target);
2836 /* Include the BARRIER after INSN and copy the
2838 new_label = squeeze_notes (new_label,
2840 reorder_insns (new_label, last_insn_to_move, loc);
2842 /* All those insns are now in TARGET_LOOP. */
2844 q != NEXT_INSN (last_insn_to_move);
2846 uid_loop[INSN_UID (q)] = target_loop;
2848 /* The label jumped to by INSN is no longer a loop
2849 exit. Unless INSN does not have a label (e.g.,
2850 it is a RETURN insn), search loop->exit_labels
2851 to find its label_ref, and remove it. Also turn
2852 off LABEL_OUTSIDE_LOOP_P bit. */
2853 if (JUMP_LABEL (insn))
2856 r = this_loop->exit_labels;
2857 r; q = r, r = LABEL_NEXTREF (r))
2858 if (XEXP (r, 0) == JUMP_LABEL (insn))
2860 LABEL_OUTSIDE_LOOP_P (r) = 0;
2862 LABEL_NEXTREF (q) = LABEL_NEXTREF (r);
2864 this_loop->exit_labels = LABEL_NEXTREF (r);
2868 for (loop = this_loop; loop && loop != target_loop;
2872 /* If we didn't find it, then something is
2878 /* P is now a jump outside the loop, so it must be put
2879 in loop->exit_labels, and marked as such.
2880 The easiest way to do this is to just call
2881 mark_loop_jump again for P. */
2882 mark_loop_jump (PATTERN (p), this_loop);
2884 /* If INSN now jumps to the insn after it,
2886 if (JUMP_LABEL (insn) != 0
2887 && (next_real_insn (JUMP_LABEL (insn))
2888 == next_real_insn (insn)))
2892 /* Continue the loop after where the conditional
2893 branch used to jump, since the only branch insn
2894 in the block (if it still remains) is an inter-loop
2895 branch and hence needs no processing. */
2896 insn = NEXT_INSN (cond_label);
2898 if (--LABEL_NUSES (cond_label) == 0)
2899 delete_insn (cond_label);
2901 /* This loop will be continued with NEXT_INSN (insn). */
2902 insn = PREV_INSN (insn);
2909 /* If any label in X jumps to a loop different from LOOP_NUM and any of the
2910 loops it is contained in, mark the target loop invalid.
2912 For speed, we assume that X is part of a pattern of a JUMP_INSN. */
2915 mark_loop_jump (x, loop)
2919 struct loop *dest_loop;
2920 struct loop *outer_loop;
2923 switch (GET_CODE (x))
2936 /* There could be a label reference in here. */
2937 mark_loop_jump (XEXP (x, 0), loop);
2943 mark_loop_jump (XEXP (x, 0), loop);
2944 mark_loop_jump (XEXP (x, 1), loop);
2948 /* This may refer to a LABEL_REF or SYMBOL_REF. */
2949 mark_loop_jump (XEXP (x, 1), loop);
2954 mark_loop_jump (XEXP (x, 0), loop);
2958 dest_loop = uid_loop[INSN_UID (XEXP (x, 0))];
2960 /* Link together all labels that branch outside the loop. This
2961 is used by final_[bg]iv_value and the loop unrolling code. Also
2962 mark this LABEL_REF so we know that this branch should predict
2965 /* A check to make sure the label is not in an inner nested loop,
2966 since this does not count as a loop exit. */
2969 for (outer_loop = dest_loop; outer_loop;
2970 outer_loop = outer_loop->outer)
2971 if (outer_loop == loop)
2977 if (loop && ! outer_loop)
2979 LABEL_OUTSIDE_LOOP_P (x) = 1;
2980 LABEL_NEXTREF (x) = loop->exit_labels;
2981 loop->exit_labels = x;
2983 for (outer_loop = loop;
2984 outer_loop && outer_loop != dest_loop;
2985 outer_loop = outer_loop->outer)
2986 outer_loop->exit_count++;
2989 /* If this is inside a loop, but not in the current loop or one enclosed
2990 by it, it invalidates at least one loop. */
2995 /* We must invalidate every nested loop containing the target of this
2996 label, except those that also contain the jump insn. */
2998 for (; dest_loop; dest_loop = dest_loop->outer)
3000 /* Stop when we reach a loop that also contains the jump insn. */
3001 for (outer_loop = loop; outer_loop; outer_loop = outer_loop->outer)
3002 if (dest_loop == outer_loop)
3005 /* If we get here, we know we need to invalidate a loop. */
3006 if (loop_dump_stream && ! dest_loop->invalid)
3007 fprintf (loop_dump_stream,
3008 "\nLoop at %d ignored due to multiple entry points.\n",
3009 INSN_UID (dest_loop->start));
3011 dest_loop->invalid = 1;
3016 /* If this is not setting pc, ignore. */
3017 if (SET_DEST (x) == pc_rtx)
3018 mark_loop_jump (SET_SRC (x), loop);
3022 mark_loop_jump (XEXP (x, 1), loop);
3023 mark_loop_jump (XEXP (x, 2), loop);
3028 for (i = 0; i < XVECLEN (x, 0); i++)
3029 mark_loop_jump (XVECEXP (x, 0, i), loop);
3033 for (i = 0; i < XVECLEN (x, 1); i++)
3034 mark_loop_jump (XVECEXP (x, 1, i), loop);
3038 /* Strictly speaking this is not a jump into the loop, only a possible
3039 jump out of the loop. However, we have no way to link the destination
3040 of this jump onto the list of exit labels. To be safe we mark this
3041 loop and any containing loops as invalid. */
3044 for (outer_loop = loop; outer_loop; outer_loop = outer_loop->outer)
3046 if (loop_dump_stream && ! outer_loop->invalid)
3047 fprintf (loop_dump_stream,
3048 "\nLoop at %d ignored due to unknown exit jump.\n",
3049 INSN_UID (outer_loop->start));
3050 outer_loop->invalid = 1;
3057 /* Return nonzero if there is a label in the range from
3058 insn INSN to and including the insn whose luid is END
3059 INSN must have an assigned luid (i.e., it must not have
3060 been previously created by loop.c). */
3063 labels_in_range_p (insn, end)
3067 while (insn && INSN_LUID (insn) <= end)
3069 if (GET_CODE (insn) == CODE_LABEL)
3071 insn = NEXT_INSN (insn);
3077 /* Record that a memory reference X is being set. */
3080 note_addr_stored (x, y, data)
3082 rtx y ATTRIBUTE_UNUSED;
3083 void *data ATTRIBUTE_UNUSED;
3085 if (x == 0 || GET_CODE (x) != MEM)
3088 /* Count number of memory writes.
3089 This affects heuristics in strength_reduce. */
3092 /* BLKmode MEM means all memory is clobbered. */
3093 if (GET_MODE (x) == BLKmode)
3095 if (RTX_UNCHANGING_P (x))
3096 unknown_constant_address_altered = 1;
3098 unknown_address_altered = 1;
3103 loop_store_mems = gen_rtx_EXPR_LIST (VOIDmode, x, loop_store_mems);
3106 /* X is a value modified by an INSN that references a biv inside a loop
3107 exit test (ie, X is somehow related to the value of the biv). If X
3108 is a pseudo that is used more than once, then the biv is (effectively)
3109 used more than once. DATA is really an `int *', and is set if the
3110 biv is used more than once. */
3113 note_set_pseudo_multiple_uses (x, y, data)
3115 rtx y ATTRIBUTE_UNUSED;
3121 while (GET_CODE (x) == STRICT_LOW_PART
3122 || GET_CODE (x) == SIGN_EXTRACT
3123 || GET_CODE (x) == ZERO_EXTRACT
3124 || GET_CODE (x) == SUBREG)
3127 if (GET_CODE (x) != REG || REGNO (x) < FIRST_PSEUDO_REGISTER)
3130 /* If we do not have usage information, or if we know the register
3131 is used more than once, note that fact for check_dbra_loop. */
3132 if (REGNO (x) >= max_reg_before_loop
3133 || ! VARRAY_RTX (reg_single_usage, REGNO (x))
3134 || VARRAY_RTX (reg_single_usage, REGNO (x)) == const0_rtx)
3135 *((int *) data) = 1;
3138 /* Return nonzero if the rtx X is invariant over the current loop.
3140 The value is 2 if we refer to something only conditionally invariant.
3142 A memory ref is invariant if it is not volatile and does not conflict
3143 with anything stored in `loop_store_mems'. */
3146 loop_invariant_p (loop, x)
3147 const struct loop *loop;
3151 register enum rtx_code code;
3152 register const char *fmt;
3153 int conditional = 0;
3158 code = GET_CODE (x);
3168 /* A LABEL_REF is normally invariant, however, if we are unrolling
3169 loops, and this label is inside the loop, then it isn't invariant.
3170 This is because each unrolled copy of the loop body will have
3171 a copy of this label. If this was invariant, then an insn loading
3172 the address of this label into a register might get moved outside
3173 the loop, and then each loop body would end up using the same label.
3175 We don't know the loop bounds here though, so just fail for all
3177 if (flag_unroll_loops)
3184 case UNSPEC_VOLATILE:
3188 /* We used to check RTX_UNCHANGING_P (x) here, but that is invalid
3189 since the reg might be set by initialization within the loop. */
3191 if ((x == frame_pointer_rtx || x == hard_frame_pointer_rtx
3192 || x == arg_pointer_rtx)
3193 && ! current_function_has_nonlocal_goto)
3196 if (LOOP_INFO (loop)->has_call
3197 && REGNO (x) < FIRST_PSEUDO_REGISTER && call_used_regs[REGNO (x)])
3200 if (VARRAY_INT (set_in_loop, REGNO (x)) < 0)
3203 return VARRAY_INT (set_in_loop, REGNO (x)) == 0;
3206 /* Volatile memory references must be rejected. Do this before
3207 checking for read-only items, so that volatile read-only items
3208 will be rejected also. */
3209 if (MEM_VOLATILE_P (x))
3212 /* See if there is any dependence between a store and this load. */
3213 mem_list_entry = loop_store_mems;
3214 while (mem_list_entry)
3216 if (true_dependence (XEXP (mem_list_entry, 0), VOIDmode,
3220 mem_list_entry = XEXP (mem_list_entry, 1);
3223 /* It's not invalidated by a store in memory
3224 but we must still verify the address is invariant. */
3228 /* Don't mess with insns declared volatile. */
3229 if (MEM_VOLATILE_P (x))
3237 fmt = GET_RTX_FORMAT (code);
3238 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3242 int tem = loop_invariant_p (loop, XEXP (x, i));
3248 else if (fmt[i] == 'E')
3251 for (j = 0; j < XVECLEN (x, i); j++)
3253 int tem = loop_invariant_p (loop, XVECEXP (x, i, j));
3263 return 1 + conditional;
3266 /* Return nonzero if all the insns in the loop that set REG
3267 are INSN and the immediately following insns,
3268 and if each of those insns sets REG in an invariant way
3269 (not counting uses of REG in them).
3271 The value is 2 if some of these insns are only conditionally invariant.
3273 We assume that INSN itself is the first set of REG
3274 and that its source is invariant. */
3277 consec_sets_invariant_p (loop, reg, n_sets, insn)
3278 const struct loop *loop;
3283 unsigned int regno = REGNO (reg);
3285 /* Number of sets we have to insist on finding after INSN. */
3286 int count = n_sets - 1;
3287 int old = VARRAY_INT (set_in_loop, regno);
3291 /* If N_SETS hit the limit, we can't rely on its value. */
3295 VARRAY_INT (set_in_loop, regno) = 0;
3299 register enum rtx_code code;
3303 code = GET_CODE (p);
3305 /* If library call, skip to end of it. */
3306 if (code == INSN && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
3311 && (set = single_set (p))
3312 && GET_CODE (SET_DEST (set)) == REG
3313 && REGNO (SET_DEST (set)) == regno)
3315 this = loop_invariant_p (loop, SET_SRC (set));
3318 else if ((temp = find_reg_note (p, REG_EQUAL, NULL_RTX)))
3320 /* If this is a libcall, then any invariant REG_EQUAL note is OK.
3321 If this is an ordinary insn, then only CONSTANT_P REG_EQUAL
3323 this = (CONSTANT_P (XEXP (temp, 0))
3324 || (find_reg_note (p, REG_RETVAL, NULL_RTX)
3325 && loop_invariant_p (loop, XEXP (temp, 0))));
3332 else if (code != NOTE)
3334 VARRAY_INT (set_in_loop, regno) = old;
3339 VARRAY_INT (set_in_loop, regno) = old;
3340 /* If loop_invariant_p ever returned 2, we return 2. */
3341 return 1 + (value & 2);
3345 /* I don't think this condition is sufficient to allow INSN
3346 to be moved, so we no longer test it. */
3348 /* Return 1 if all insns in the basic block of INSN and following INSN
3349 that set REG are invariant according to TABLE. */
3352 all_sets_invariant_p (reg, insn, table)
3356 register rtx p = insn;
3357 register int regno = REGNO (reg);
3361 register enum rtx_code code;
3363 code = GET_CODE (p);
3364 if (code == CODE_LABEL || code == JUMP_INSN)
3366 if (code == INSN && GET_CODE (PATTERN (p)) == SET
3367 && GET_CODE (SET_DEST (PATTERN (p))) == REG
3368 && REGNO (SET_DEST (PATTERN (p))) == regno)
3370 if (! loop_invariant_p (loop, SET_SRC (PATTERN (p)), table))
3377 /* Look at all uses (not sets) of registers in X. For each, if it is
3378 the single use, set USAGE[REGNO] to INSN; if there was a previous use in
3379 a different insn, set USAGE[REGNO] to const0_rtx. */
3382 find_single_use_in_loop (insn, x, usage)
3387 enum rtx_code code = GET_CODE (x);
3388 const char *fmt = GET_RTX_FORMAT (code);
3392 VARRAY_RTX (usage, REGNO (x))
3393 = (VARRAY_RTX (usage, REGNO (x)) != 0
3394 && VARRAY_RTX (usage, REGNO (x)) != insn)
3395 ? const0_rtx : insn;
3397 else if (code == SET)
3399 /* Don't count SET_DEST if it is a REG; otherwise count things
3400 in SET_DEST because if a register is partially modified, it won't
3401 show up as a potential movable so we don't care how USAGE is set
3403 if (GET_CODE (SET_DEST (x)) != REG)
3404 find_single_use_in_loop (insn, SET_DEST (x), usage);
3405 find_single_use_in_loop (insn, SET_SRC (x), usage);
3408 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3410 if (fmt[i] == 'e' && XEXP (x, i) != 0)
3411 find_single_use_in_loop (insn, XEXP (x, i), usage);
3412 else if (fmt[i] == 'E')
3413 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
3414 find_single_use_in_loop (insn, XVECEXP (x, i, j), usage);
3418 /* Count and record any set in X which is contained in INSN. Update
3419 MAY_NOT_MOVE and LAST_SET for any register set in X. */
3422 count_one_set (insn, x, may_not_move, last_set)
3424 varray_type may_not_move;
3427 if (GET_CODE (x) == CLOBBER && GET_CODE (XEXP (x, 0)) == REG)
3428 /* Don't move a reg that has an explicit clobber.
3429 It's not worth the pain to try to do it correctly. */
3430 VARRAY_CHAR (may_not_move, REGNO (XEXP (x, 0))) = 1;
3432 if (GET_CODE (x) == SET || GET_CODE (x) == CLOBBER)
3434 rtx dest = SET_DEST (x);
3435 while (GET_CODE (dest) == SUBREG
3436 || GET_CODE (dest) == ZERO_EXTRACT
3437 || GET_CODE (dest) == SIGN_EXTRACT
3438 || GET_CODE (dest) == STRICT_LOW_PART)
3439 dest = XEXP (dest, 0);
3440 if (GET_CODE (dest) == REG)
3442 register int regno = REGNO (dest);
3443 /* If this is the first setting of this reg
3444 in current basic block, and it was set before,
3445 it must be set in two basic blocks, so it cannot
3446 be moved out of the loop. */
3447 if (VARRAY_INT (set_in_loop, regno) > 0
3448 && last_set[regno] == 0)
3449 VARRAY_CHAR (may_not_move, regno) = 1;
3450 /* If this is not first setting in current basic block,
3451 see if reg was used in between previous one and this.
3452 If so, neither one can be moved. */
3453 if (last_set[regno] != 0
3454 && reg_used_between_p (dest, last_set[regno], insn))
3455 VARRAY_CHAR (may_not_move, regno) = 1;
3456 if (VARRAY_INT (set_in_loop, regno) < 127)
3457 ++VARRAY_INT (set_in_loop, regno);
3458 last_set[regno] = insn;
3463 /* Increment SET_IN_LOOP at the index of each register
3464 that is modified by an insn between FROM and TO.
3465 If the value of an element of SET_IN_LOOP becomes 127 or more,
3466 stop incrementing it, to avoid overflow.
3468 Store in SINGLE_USAGE[I] the single insn in which register I is
3469 used, if it is only used once. Otherwise, it is set to 0 (for no
3470 uses) or const0_rtx for more than one use. This parameter may be zero,
3471 in which case this processing is not done.
3473 Store in *COUNT_PTR the number of actual instruction
3474 in the loop. We use this to decide what is worth moving out. */
3476 /* last_set[n] is nonzero iff reg n has been set in the current basic block.
3477 In that case, it is the insn that last set reg n. */
3480 count_loop_regs_set (loop, may_not_move, single_usage, count_ptr, nregs)
3481 const struct loop *loop;
3482 varray_type may_not_move;
3483 varray_type single_usage;
3487 register rtx *last_set = (rtx *) xcalloc (nregs, sizeof (rtx));
3489 register int count = 0;
3491 for (insn = loop->top ? loop->top : loop->start; insn != loop->end;
3492 insn = NEXT_INSN (insn))
3498 /* Record registers that have exactly one use. */
3499 find_single_use_in_loop (insn, PATTERN (insn), single_usage);
3501 /* Include uses in REG_EQUAL notes. */
3502 if (REG_NOTES (insn))
3503 find_single_use_in_loop (insn, REG_NOTES (insn), single_usage);
3505 if (GET_CODE (PATTERN (insn)) == SET
3506 || GET_CODE (PATTERN (insn)) == CLOBBER)
3507 count_one_set (insn, PATTERN (insn), may_not_move, last_set);
3508 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
3511 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
3512 count_one_set (insn, XVECEXP (PATTERN (insn), 0, i),
3513 may_not_move, last_set);
3517 if (GET_CODE (insn) == CODE_LABEL || GET_CODE (insn) == JUMP_INSN)
3518 bzero ((char *) last_set, nregs * sizeof (rtx));
3526 /* Given a loop that is bounded by LOOP->START and LOOP->END and that
3527 is entered at LOOP->SCAN_START, return 1 if the register set in SET
3528 contained in insn INSN is used by any insn that precedes INSN in
3529 cyclic order starting from the loop entry point.
3531 We don't want to use INSN_LUID here because if we restrict INSN to those
3532 that have a valid INSN_LUID, it means we cannot move an invariant out
3533 from an inner loop past two loops. */
3536 loop_reg_used_before_p (loop, set, insn)
3537 const struct loop *loop;
3540 rtx reg = SET_DEST (set);
3543 /* Scan forward checking for register usage. If we hit INSN, we
3544 are done. Otherwise, if we hit LOOP->END, wrap around to LOOP->START. */
3545 for (p = loop->scan_start; p != insn; p = NEXT_INSN (p))
3547 if (INSN_P (p) && reg_overlap_mentioned_p (reg, PATTERN (p)))
3557 /* A "basic induction variable" or biv is a pseudo reg that is set
3558 (within this loop) only by incrementing or decrementing it. */
3559 /* A "general induction variable" or giv is a pseudo reg whose
3560 value is a linear function of a biv. */
3562 /* Bivs are recognized by `basic_induction_var';
3563 Givs by `general_induction_var'. */
3565 /* Indexed by register number, indicates whether or not register is an
3566 induction variable, and if so what type. */
3568 varray_type reg_iv_type;
3570 /* Indexed by register number, contains pointer to `struct induction'
3571 if register is an induction variable. This holds general info for
3572 all induction variables. */
3574 varray_type reg_iv_info;
3576 /* Indexed by register number, contains pointer to `struct iv_class'
3577 if register is a basic induction variable. This holds info describing
3578 the class (a related group) of induction variables that the biv belongs
3581 struct iv_class **reg_biv_class;
3583 /* The head of a list which links together (via the next field)
3584 every iv class for the current loop. */
3586 struct iv_class *loop_iv_list;
3588 /* Givs made from biv increments are always splittable for loop unrolling.
3589 Since there is no regscan info for them, we have to keep track of them
3591 unsigned int first_increment_giv, last_increment_giv;
3593 /* Communication with routines called via `note_stores'. */
3595 static rtx note_insn;
3597 /* Dummy register to have non-zero DEST_REG for DEST_ADDR type givs. */
3599 static rtx addr_placeholder;
3601 /* ??? Unfinished optimizations, and possible future optimizations,
3602 for the strength reduction code. */
3604 /* ??? The interaction of biv elimination, and recognition of 'constant'
3605 bivs, may cause problems. */
3607 /* ??? Add heuristics so that DEST_ADDR strength reduction does not cause
3608 performance problems.
3610 Perhaps don't eliminate things that can be combined with an addressing
3611 mode. Find all givs that have the same biv, mult_val, and add_val;
3612 then for each giv, check to see if its only use dies in a following
3613 memory address. If so, generate a new memory address and check to see
3614 if it is valid. If it is valid, then store the modified memory address,
3615 otherwise, mark the giv as not done so that it will get its own iv. */
3617 /* ??? Could try to optimize branches when it is known that a biv is always
3620 /* ??? When replace a biv in a compare insn, we should replace with closest
3621 giv so that an optimized branch can still be recognized by the combiner,
3622 e.g. the VAX acb insn. */
3624 /* ??? Many of the checks involving uid_luid could be simplified if regscan
3625 was rerun in loop_optimize whenever a register was added or moved.
3626 Also, some of the optimizations could be a little less conservative. */
3628 /* Scan the loop body and call FNCALL for each insn. In the addition to the
3629 LOOP and INSN parameters pass MAYBE_MULTIPLE and NOT_EVERY_ITERATION to the
3632 NOT_EVERY_ITERATION if current insn is not executed at least once for every
3633 loop iteration except for the last one.
3635 MAYBE_MULTIPLE is 1 if current insn may be executed more than once for every
3639 for_each_insn_in_loop (loop, fncall)
3641 loop_insn_callback fncall;
3643 /* This is 1 if current insn is not executed at least once for every loop
3645 int not_every_iteration = 0;
3646 int maybe_multiple = 0;
3647 int past_loop_latch = 0;
3651 /* If loop_scan_start points to the loop exit test, we have to be wary of
3652 subversive use of gotos inside expression statements. */
3653 if (prev_nonnote_insn (loop->scan_start) != prev_nonnote_insn (loop->start))
3654 maybe_multiple = back_branch_in_range_p (loop, loop->scan_start);
3656 /* Scan through loop to find all possible bivs. */
3658 for (p = next_insn_in_loop (loop, loop->scan_start);
3660 p = next_insn_in_loop (loop, p))
3662 p = fncall (loop, p, not_every_iteration, maybe_multiple);
3664 /* Past CODE_LABEL, we get to insns that may be executed multiple
3665 times. The only way we can be sure that they can't is if every
3666 jump insn between here and the end of the loop either
3667 returns, exits the loop, is a jump to a location that is still
3668 behind the label, or is a jump to the loop start. */
3670 if (GET_CODE (p) == CODE_LABEL)
3678 insn = NEXT_INSN (insn);
3679 if (insn == loop->scan_start)
3681 if (insn == loop->end)
3687 if (insn == loop->scan_start)
3691 if (GET_CODE (insn) == JUMP_INSN
3692 && GET_CODE (PATTERN (insn)) != RETURN
3693 && (!any_condjump_p (insn)
3694 || (JUMP_LABEL (insn) != 0
3695 && JUMP_LABEL (insn) != loop->scan_start
3696 && !loop_insn_first_p (p, JUMP_LABEL (insn)))))
3704 /* Past a jump, we get to insns for which we can't count
3705 on whether they will be executed during each iteration. */
3706 /* This code appears twice in strength_reduce. There is also similar
3707 code in scan_loop. */
3708 if (GET_CODE (p) == JUMP_INSN
3709 /* If we enter the loop in the middle, and scan around to the
3710 beginning, don't set not_every_iteration for that.
3711 This can be any kind of jump, since we want to know if insns
3712 will be executed if the loop is executed. */
3713 && !(JUMP_LABEL (p) == loop->top
3714 && ((NEXT_INSN (NEXT_INSN (p)) == loop->end
3715 && any_uncondjump_p (p))
3716 || (NEXT_INSN (p) == loop->end && any_condjump_p (p)))))
3720 /* If this is a jump outside the loop, then it also doesn't
3721 matter. Check to see if the target of this branch is on the
3722 loop->exits_labels list. */
3724 for (label = loop->exit_labels; label; label = LABEL_NEXTREF (label))
3725 if (XEXP (label, 0) == JUMP_LABEL (p))
3729 not_every_iteration = 1;
3732 else if (GET_CODE (p) == NOTE)
3734 /* At the virtual top of a converted loop, insns are again known to
3735 be executed each iteration: logically, the loop begins here
3736 even though the exit code has been duplicated.
3738 Insns are also again known to be executed each iteration at
3739 the LOOP_CONT note. */
3740 if ((NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP
3741 || NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_CONT)
3743 not_every_iteration = 0;
3744 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
3746 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END)
3750 /* Note if we pass a loop latch. If we do, then we can not clear
3751 NOT_EVERY_ITERATION below when we pass the last CODE_LABEL in
3752 a loop since a jump before the last CODE_LABEL may have started
3753 a new loop iteration.
3755 Note that LOOP_TOP is only set for rotated loops and we need
3756 this check for all loops, so compare against the CODE_LABEL
3757 which immediately follows LOOP_START. */
3758 if (GET_CODE (p) == JUMP_INSN
3759 && JUMP_LABEL (p) == NEXT_INSN (loop->start))
3760 past_loop_latch = 1;
3762 /* Unlike in the code motion pass where MAYBE_NEVER indicates that
3763 an insn may never be executed, NOT_EVERY_ITERATION indicates whether
3764 or not an insn is known to be executed each iteration of the
3765 loop, whether or not any iterations are known to occur.
3767 Therefore, if we have just passed a label and have no more labels
3768 between here and the test insn of the loop, and we have not passed
3769 a jump to the top of the loop, then we know these insns will be
3770 executed each iteration. */
3772 if (not_every_iteration
3774 && GET_CODE (p) == CODE_LABEL
3775 && no_labels_between_p (p, loop->end)
3776 && loop_insn_first_p (p, loop->cont))
3777 not_every_iteration = 0;
3781 /* Perform strength reduction and induction variable elimination.
3783 Pseudo registers created during this function will be beyond the last
3784 valid index in several tables including n_times_set and regno_last_uid.
3785 This does not cause a problem here, because the added registers cannot be
3786 givs outside of their loop, and hence will never be reconsidered.
3787 But scan_loop must check regnos to make sure they are in bounds. */
3790 strength_reduce (loop, insn_count, flags)
3796 /* Temporary list pointers for traversing loop_iv_list. */
3797 struct iv_class *bl, **backbl;
3798 struct loop_info *loop_info = LOOP_INFO (loop);
3799 /* Ratio of extra register life span we can justify
3800 for saving an instruction. More if loop doesn't call subroutines
3801 since in that case saving an insn makes more difference
3802 and more registers are available. */
3803 /* ??? could set this to last value of threshold in move_movables */
3804 int threshold = (loop_info->has_call ? 1 : 2) * (3 + n_non_fixed_regs);
3805 /* Map of pseudo-register replacements. */
3806 rtx *reg_map = NULL;
3810 rtx end_insert_before;
3811 int n_extra_increment;
3812 int unrolled_insn_copies = 0;
3813 rtx loop_start = loop->start;
3814 rtx loop_end = loop->end;
3815 rtx loop_scan_start = loop->scan_start;
3816 rtx test_reg = gen_rtx_REG (word_mode, LAST_VIRTUAL_REGISTER + 1);
3818 VARRAY_INT_INIT (reg_iv_type, max_reg_before_loop, "reg_iv_type");
3819 VARRAY_GENERIC_PTR_INIT (reg_iv_info, max_reg_before_loop, "reg_iv_info");
3820 reg_biv_class = (struct iv_class **)
3821 xcalloc (max_reg_before_loop, sizeof (struct iv_class *));
3824 addr_placeholder = gen_reg_rtx (Pmode);
3826 /* Save insn immediately after the loop_end. Insns inserted after loop_end
3827 must be put before this insn, so that they will appear in the right
3828 order (i.e. loop order).
3830 If loop_end is the end of the current function, then emit a
3831 NOTE_INSN_DELETED after loop_end and set end_insert_before to the
3833 if (NEXT_INSN (loop_end) != 0)
3834 end_insert_before = NEXT_INSN (loop_end);
3836 end_insert_before = emit_note_after (NOTE_INSN_DELETED, loop_end);
3838 for_each_insn_in_loop (loop, check_insn_for_bivs);
3840 /* Scan loop_iv_list to remove all regs that proved not to be bivs.
3841 Make a sanity check against n_times_set. */
3842 for (backbl = &loop_iv_list, bl = *backbl; bl; bl = bl->next)
3844 if (REG_IV_TYPE (bl->regno) != BASIC_INDUCT
3845 /* Above happens if register modified by subreg, etc. */
3846 /* Make sure it is not recognized as a basic induction var: */
3847 || VARRAY_INT (n_times_set, bl->regno) != bl->biv_count
3848 /* If never incremented, it is invariant that we decided not to
3849 move. So leave it alone. */
3850 || ! bl->incremented)
3852 if (loop_dump_stream)
3853 fprintf (loop_dump_stream, "Reg %d: biv discarded, %s\n",
3855 (REG_IV_TYPE (bl->regno) != BASIC_INDUCT
3856 ? "not induction variable"
3857 : (! bl->incremented ? "never incremented"
3860 REG_IV_TYPE (bl->regno) = NOT_BASIC_INDUCT;
3867 if (loop_dump_stream)
3868 fprintf (loop_dump_stream, "Reg %d: biv verified\n", bl->regno);
3872 /* Exit if there are no bivs. */
3875 /* Can still unroll the loop anyways, but indicate that there is no
3876 strength reduction info available. */
3877 if (flags & LOOP_UNROLL)
3878 unroll_loop (loop, insn_count, end_insert_before, 0);
3883 /* Find initial value for each biv by searching backwards from loop_start,
3884 halting at first label. Also record any test condition. */
3887 for (p = loop_start; p && GET_CODE (p) != CODE_LABEL; p = PREV_INSN (p))
3891 if (GET_CODE (p) == CALL_INSN)
3895 note_stores (PATTERN (p), record_initial, NULL);
3897 /* Record any test of a biv that branches around the loop if no store
3898 between it and the start of loop. We only care about tests with
3899 constants and registers and only certain of those. */
3900 if (GET_CODE (p) == JUMP_INSN
3901 && JUMP_LABEL (p) != 0
3902 && next_real_insn (JUMP_LABEL (p)) == next_real_insn (loop_end)
3903 && (test = get_condition_for_loop (loop, p)) != 0
3904 && GET_CODE (XEXP (test, 0)) == REG
3905 && REGNO (XEXP (test, 0)) < max_reg_before_loop
3906 && (bl = reg_biv_class[REGNO (XEXP (test, 0))]) != 0
3907 && valid_initial_value_p (XEXP (test, 1), p, call_seen, loop_start)
3908 && bl->init_insn == 0)
3910 /* If an NE test, we have an initial value! */
3911 if (GET_CODE (test) == NE)
3914 bl->init_set = gen_rtx_SET (VOIDmode,
3915 XEXP (test, 0), XEXP (test, 1));
3918 bl->initial_test = test;
3922 /* Look at the each biv and see if we can say anything better about its
3923 initial value from any initializing insns set up above. (This is done
3924 in two passes to avoid missing SETs in a PARALLEL.) */
3925 for (backbl = &loop_iv_list; (bl = *backbl); backbl = &bl->next)
3930 if (! bl->init_insn)
3933 /* IF INIT_INSN has a REG_EQUAL or REG_EQUIV note and the value
3934 is a constant, use the value of that. */
3935 if (((note = find_reg_note (bl->init_insn, REG_EQUAL, 0)) != NULL
3936 && CONSTANT_P (XEXP (note, 0)))
3937 || ((note = find_reg_note (bl->init_insn, REG_EQUIV, 0)) != NULL
3938 && CONSTANT_P (XEXP (note, 0))))
3939 src = XEXP (note, 0);
3941 src = SET_SRC (bl->init_set);
3943 if (loop_dump_stream)
3944 fprintf (loop_dump_stream,
3945 "Biv %d initialized at insn %d: initial value ",
3946 bl->regno, INSN_UID (bl->init_insn));
3948 if ((GET_MODE (src) == GET_MODE (regno_reg_rtx[bl->regno])
3949 || GET_MODE (src) == VOIDmode)
3950 && valid_initial_value_p (src, bl->init_insn, call_seen, loop_start))
3952 bl->initial_value = src;
3954 if (loop_dump_stream)
3956 if (GET_CODE (src) == CONST_INT)
3958 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (src));
3959 fputc ('\n', loop_dump_stream);
3963 print_rtl (loop_dump_stream, src);
3964 fprintf (loop_dump_stream, "\n");
3970 struct iv_class *bl2 = 0;
3971 rtx increment = NULL_RTX;
3973 /* Biv initial value is not a simple move. If it is the sum of
3974 another biv and a constant, check if both bivs are incremented
3975 in lockstep. Then we are actually looking at a giv.
3976 For simplicity, we only handle the case where there is but a
3977 single increment, and the register is not used elsewhere. */
3978 if (bl->biv_count == 1
3979 && bl->regno < max_reg_before_loop
3980 && uid_luid[REGNO_LAST_UID (bl->regno)] < INSN_LUID (loop_end)
3981 && GET_CODE (src) == PLUS
3982 && GET_CODE (XEXP (src, 0)) == REG
3983 && CONSTANT_P (XEXP (src, 1))
3984 && ((increment = biv_total_increment (bl)) != NULL_RTX))
3986 unsigned int regno = REGNO (XEXP (src, 0));
3988 for (bl2 = loop_iv_list; bl2; bl2 = bl2->next)
3989 if (bl2->regno == regno)
3993 /* Now, can we transform this biv into a giv? */
3995 && bl2->biv_count == 1
3996 && rtx_equal_p (increment, biv_total_increment (bl2))
3997 /* init_insn is only set to insns that are before loop_start
3998 without any intervening labels. */
3999 && ! reg_set_between_p (bl2->biv->src_reg,
4000 PREV_INSN (bl->init_insn), loop_start)
4001 /* The register from BL2 must be set before the register from
4002 BL is set, or we must be able to move the latter set after
4003 the former set. Currently there can't be any labels
4004 in-between when biv_total_increment returns nonzero both times
4005 but we test it here in case some day some real cfg analysis
4006 gets used to set always_computable. */
4007 && (loop_insn_first_p (bl2->biv->insn, bl->biv->insn)
4008 ? no_labels_between_p (bl2->biv->insn, bl->biv->insn)
4009 : (! reg_used_between_p (bl->biv->src_reg, bl->biv->insn,
4011 && no_jumps_between_p (bl->biv->insn, bl2->biv->insn)))
4012 && validate_change (bl->biv->insn,
4013 &SET_SRC (single_set (bl->biv->insn)),
4016 rtx dominator = loop->cont_dominator;
4017 rtx giv = bl->biv->src_reg;
4018 rtx giv_insn = bl->biv->insn;
4019 rtx after_giv = NEXT_INSN (giv_insn);
4021 if (loop_dump_stream)
4022 fprintf (loop_dump_stream, "is giv of biv %d\n", bl2->regno);
4023 /* Let this giv be discovered by the generic code. */
4024 REG_IV_TYPE (bl->regno) = UNKNOWN_INDUCT;
4025 reg_biv_class[bl->regno] = (struct iv_class *) NULL_PTR;
4026 /* We can get better optimization if we can move the giv setting
4027 before the first giv use. */
4029 && ! loop_insn_first_p (dominator, loop_scan_start)
4030 && ! reg_set_between_p (bl2->biv->src_reg, loop_start,
4032 && ! reg_used_between_p (giv, loop_start, dominator)
4033 && ! reg_used_between_p (giv, giv_insn, loop_end))
4038 for (next = NEXT_INSN (dominator);; next = NEXT_INSN (next))
4040 if (GET_CODE (next) == JUMP_INSN
4042 && insn_dependent_p (giv_insn, next)))
4045 if (! INSN_P (next) || ! sets_cc0_p (PATTERN (next)))
4049 if (loop_dump_stream)
4050 fprintf (loop_dump_stream, "move after insn %d\n",
4051 INSN_UID (dominator));
4052 /* Avoid problems with luids by actually moving the insn
4053 and adjusting all luids in the range. */
4054 reorder_insns (giv_insn, giv_insn, dominator);
4055 for (p = dominator; INSN_UID (p) >= max_uid_for_loop;)
4057 compute_luids (giv_insn, after_giv, INSN_LUID (p));
4058 /* If the only purpose of the init insn is to initialize
4059 this giv, delete it. */
4060 if (single_set (bl->init_insn)
4061 && ! reg_used_between_p (giv, bl->init_insn, loop_start))
4062 delete_insn (bl->init_insn);
4064 else if (! loop_insn_first_p (bl2->biv->insn, bl->biv->insn))
4066 rtx p = PREV_INSN (giv_insn);
4067 while (INSN_UID (p) >= max_uid_for_loop)
4069 reorder_insns (giv_insn, giv_insn, bl2->biv->insn);
4070 compute_luids (after_giv, NEXT_INSN (giv_insn),
4073 /* Remove this biv from the chain. */
4077 /* If we can't make it a giv,
4078 let biv keep initial value of "itself". */
4079 else if (loop_dump_stream)
4080 fprintf (loop_dump_stream, "is complex\n");
4084 /* If a biv is unconditionally incremented several times in a row, convert
4085 all but the last increment into a giv. */
4087 /* Get an upper bound for the number of registers
4088 we might have after all bivs have been processed. */
4089 first_increment_giv = max_reg_num ();
4090 for (n_extra_increment = 0, bl = loop_iv_list; bl; bl = bl->next)
4091 n_extra_increment += bl->biv_count - 1;
4093 /* If the loop contains volatile memory references do not allow any
4094 replacements to take place, since this could loose the volatile
4096 if (n_extra_increment && ! loop_info->has_volatile)
4098 unsigned int nregs = first_increment_giv + n_extra_increment;
4100 /* Reallocate reg_iv_type and reg_iv_info. */
4101 VARRAY_GROW (reg_iv_type, nregs);
4102 VARRAY_GROW (reg_iv_info, nregs);
4104 for (bl = loop_iv_list; bl; bl = bl->next)
4106 struct induction **vp, *v, *next;
4107 int biv_dead_after_loop = 0;
4109 /* The biv increments lists are in reverse order. Fix this
4111 for (v = bl->biv, bl->biv = 0; v; v = next)
4114 v->next_iv = bl->biv;
4118 /* We must guard against the case that an early exit between v->insn
4119 and next->insn leaves the biv live after the loop, since that
4120 would mean that we'd be missing an increment for the final
4121 value. The following test to set biv_dead_after_loop is like
4122 the first part of the test to set bl->eliminable.
4123 We don't check here if we can calculate the final value, since
4124 this can't succeed if we already know that there is a jump
4125 between v->insn and next->insn, yet next->always_executed is
4126 set and next->maybe_multiple is cleared. Such a combination
4127 implies that the jump destination is outside the loop.
4128 If we want to make this check more sophisticated, we should
4129 check each branch between v->insn and next->insn individually
4130 to see if the biv is dead at its destination. */
4132 if (uid_luid[REGNO_LAST_UID (bl->regno)] < INSN_LUID (loop_end)
4134 && INSN_UID (bl->init_insn) < max_uid_for_loop
4135 && (uid_luid[REGNO_FIRST_UID (bl->regno)]
4136 >= INSN_LUID (bl->init_insn))
4137 #ifdef HAVE_decrement_and_branch_until_zero
4140 && ! reg_mentioned_p (bl->biv->dest_reg, SET_SRC (bl->init_set)))
4141 biv_dead_after_loop = 1;
4143 for (vp = &bl->biv, next = *vp; v = next, next = v->next_iv;)
4145 HOST_WIDE_INT offset;
4146 rtx set, add_val, old_reg, dest_reg, last_use_insn, note;
4147 int old_regno, new_regno;
4150 if (! v->always_executed
4151 || v->maybe_multiple
4152 || GET_CODE (v->add_val) != CONST_INT
4153 || ! next->always_executed
4154 || next->maybe_multiple
4155 || ! CONSTANT_P (next->add_val)
4156 || v->mult_val != const1_rtx
4157 || next->mult_val != const1_rtx
4158 || ! (biv_dead_after_loop
4159 || no_jumps_between_p (v->insn, next->insn)))
4164 offset = INTVAL (v->add_val);
4165 set = single_set (v->insn);
4166 add_val = plus_constant (next->add_val, offset);
4167 old_reg = v->dest_reg;
4168 dest_reg = gen_reg_rtx (v->mode);
4170 /* Unlike reg_iv_type / reg_iv_info, the other three arrays
4171 have been allocated with some slop space, so we may not
4172 actually need to reallocate them. If we do, the following
4173 if statement will be executed just once in this loop. */
4174 if ((unsigned) max_reg_num () > n_times_set->num_elements)
4176 /* Grow all the remaining arrays. */
4177 VARRAY_GROW (set_in_loop, nregs);
4178 VARRAY_GROW (n_times_set, nregs);
4179 VARRAY_GROW (may_not_optimize, nregs);
4180 VARRAY_GROW (reg_single_usage, nregs);
4183 /* Some bivs are incremented with a multi-insn sequence.
4184 The first insn contains the add. */
4185 next_loc_insn = next->insn;
4186 while (NOTE_P (next_loc_insn)
4187 || ! loc_mentioned_in_p (next->location,
4188 PATTERN (next_loc_insn)))
4189 next_loc_insn = PREV_INSN (next_loc_insn);
4191 if (next_loc_insn == v->insn)
4194 if (! validate_change (next_loc_insn, next->location, add_val, 0))
4200 /* Here we can try to eliminate the increment by combining
4201 it into the uses. */
4203 /* Set last_use_insn so that we can check against it. */
4205 for (last_use_insn = v->insn, p = NEXT_INSN (v->insn);
4207 p = next_insn_in_loop (loop, p))
4211 if (reg_mentioned_p (old_reg, PATTERN (p)))
4217 /* If we can't get the LUIDs for the insns, we can't
4218 calculate the lifetime. This is likely from unrolling
4219 of an inner loop, so there is little point in making this
4220 a DEST_REG giv anyways. */
4221 if (INSN_UID (v->insn) >= max_uid_for_loop
4222 || INSN_UID (last_use_insn) >= max_uid_for_loop
4223 || ! validate_change (v->insn, &SET_DEST (set), dest_reg, 0))
4225 /* Change the increment at NEXT back to what it was. */
4226 if (! validate_change (next_loc_insn, next->location,
4232 next->add_val = add_val;
4233 v->dest_reg = dest_reg;
4234 v->giv_type = DEST_REG;
4235 v->location = &SET_SRC (set);
4237 v->combined_with = 0;
4239 v->derive_adjustment = 0;
4245 v->auto_inc_opt = 0;
4248 v->derived_from = 0;
4249 v->always_computable = 1;
4250 v->always_executed = 1;
4252 v->no_const_addval = 0;
4254 old_regno = REGNO (old_reg);
4255 new_regno = REGNO (dest_reg);
4256 VARRAY_INT (set_in_loop, old_regno)--;
4257 VARRAY_INT (set_in_loop, new_regno) = 1;
4258 VARRAY_INT (n_times_set, old_regno)--;
4259 VARRAY_INT (n_times_set, new_regno) = 1;
4260 VARRAY_CHAR (may_not_optimize, new_regno) = 0;
4262 REG_IV_TYPE (new_regno) = GENERAL_INDUCT;
4263 REG_IV_INFO (new_regno) = v;
4265 /* If next_insn has a REG_EQUAL note that mentiones OLD_REG,
4266 it must be replaced. */
4267 note = find_reg_note (next->insn, REG_EQUAL, NULL_RTX);
4268 if (note && reg_mentioned_p (old_reg, XEXP (note, 0)))
4269 XEXP (note, 0) = copy_rtx (SET_SRC (single_set (next->insn)));
4271 /* Remove the increment from the list of biv increments,
4272 and record it as a giv. */
4275 v->next_iv = bl->giv;
4278 v->benefit = rtx_cost (SET_SRC (set), SET);
4279 bl->total_benefit += v->benefit;
4281 /* Now replace the biv with DEST_REG in all insns between
4282 the replaced increment and the next increment, and
4283 remember the last insn that needed a replacement. */
4284 for (last_use_insn = v->insn, p = NEXT_INSN (v->insn);
4286 p = next_insn_in_loop (loop, p))
4292 if (reg_mentioned_p (old_reg, PATTERN (p)))
4295 if (! validate_replace_rtx (old_reg, dest_reg, p))
4298 for (note = REG_NOTES (p); note; note = XEXP (note, 1))
4300 if (GET_CODE (note) == EXPR_LIST)
4302 = replace_rtx (XEXP (note, 0), old_reg, dest_reg);
4306 v->last_use = last_use_insn;
4307 v->lifetime = INSN_LUID (last_use_insn) - INSN_LUID (v->insn);
4308 /* If the lifetime is zero, it means that this register is really
4309 a dead store. So mark this as a giv that can be ignored.
4310 This will not prevent the biv from being eliminated. */
4311 if (v->lifetime == 0)
4314 if (loop_dump_stream)
4315 fprintf (loop_dump_stream,
4316 "Increment %d of biv %d converted to giv %d.\n\n",
4317 INSN_UID (v->insn), old_regno, new_regno);
4321 last_increment_giv = max_reg_num () - 1;
4323 /* Search the loop for general induction variables. */
4325 for_each_insn_in_loop (loop, check_insn_for_givs);
4327 /* Try to calculate and save the number of loop iterations. This is
4328 set to zero if the actual number can not be calculated. This must
4329 be called after all giv's have been identified, since otherwise it may
4330 fail if the iteration variable is a giv. */
4332 loop_iterations (loop);
4334 /* Now for each giv for which we still don't know whether or not it is
4335 replaceable, check to see if it is replaceable because its final value
4336 can be calculated. This must be done after loop_iterations is called,
4337 so that final_giv_value will work correctly. */
4339 for (bl = loop_iv_list; bl; bl = bl->next)
4341 struct induction *v;
4343 for (v = bl->giv; v; v = v->next_iv)
4344 if (! v->replaceable && ! v->not_replaceable)
4345 check_final_value (loop, v);
4348 /* Try to prove that the loop counter variable (if any) is always
4349 nonnegative; if so, record that fact with a REG_NONNEG note
4350 so that "decrement and branch until zero" insn can be used. */
4351 check_dbra_loop (loop, insn_count);
4353 /* Create reg_map to hold substitutions for replaceable giv regs.
4354 Some givs might have been made from biv increments, so look at
4355 reg_iv_type for a suitable size. */
4356 reg_map_size = reg_iv_type->num_elements;
4357 reg_map = (rtx *) xcalloc (reg_map_size, sizeof (rtx));
4359 /* Examine each iv class for feasibility of strength reduction/induction
4360 variable elimination. */
4362 for (bl = loop_iv_list; bl; bl = bl->next)
4364 struct induction *v;
4367 rtx final_value = 0;
4370 /* Test whether it will be possible to eliminate this biv
4371 provided all givs are reduced. This is possible if either
4372 the reg is not used outside the loop, or we can compute
4373 what its final value will be.
4375 For architectures with a decrement_and_branch_until_zero insn,
4376 don't do this if we put a REG_NONNEG note on the endtest for
4379 /* Compare against bl->init_insn rather than loop_start.
4380 We aren't concerned with any uses of the biv between
4381 init_insn and loop_start since these won't be affected
4382 by the value of the biv elsewhere in the function, so
4383 long as init_insn doesn't use the biv itself.
4384 March 14, 1989 -- self@bayes.arc.nasa.gov */
4386 if ((uid_luid[REGNO_LAST_UID (bl->regno)] < INSN_LUID (loop_end)
4388 && INSN_UID (bl->init_insn) < max_uid_for_loop
4389 && uid_luid[REGNO_FIRST_UID (bl->regno)] >= INSN_LUID (bl->init_insn)
4390 #ifdef HAVE_decrement_and_branch_until_zero
4393 && ! reg_mentioned_p (bl->biv->dest_reg, SET_SRC (bl->init_set)))
4394 || ((final_value = final_biv_value (loop, bl))
4395 #ifdef HAVE_decrement_and_branch_until_zero
4399 bl->eliminable = maybe_eliminate_biv (loop, bl, 0, threshold,
4403 if (loop_dump_stream)
4405 fprintf (loop_dump_stream,
4406 "Cannot eliminate biv %d.\n",
4408 fprintf (loop_dump_stream,
4409 "First use: insn %d, last use: insn %d.\n",
4410 REGNO_FIRST_UID (bl->regno),
4411 REGNO_LAST_UID (bl->regno));
4415 /* Combine all giv's for this iv_class. */
4418 /* This will be true at the end, if all givs which depend on this
4419 biv have been strength reduced.
4420 We can't (currently) eliminate the biv unless this is so. */
4423 /* Check each giv in this class to see if we will benefit by reducing
4424 it. Skip giv's combined with others. */
4425 for (v = bl->giv; v; v = v->next_iv)
4427 struct induction *tv;
4430 if (v->ignore || v->same)
4433 benefit = v->benefit;
4434 PUT_MODE (test_reg, v->mode);
4435 add_cost = iv_add_mult_cost (bl->biv->add_val, v->mult_val,
4436 test_reg, test_reg);
4438 /* Reduce benefit if not replaceable, since we will insert
4439 a move-insn to replace the insn that calculates this giv.
4440 Don't do this unless the giv is a user variable, since it
4441 will often be marked non-replaceable because of the duplication
4442 of the exit code outside the loop. In such a case, the copies
4443 we insert are dead and will be deleted. So they don't have
4444 a cost. Similar situations exist. */
4445 /* ??? The new final_[bg]iv_value code does a much better job
4446 of finding replaceable giv's, and hence this code may no longer
4448 if (! v->replaceable && ! bl->eliminable
4449 && REG_USERVAR_P (v->dest_reg))
4450 benefit -= copy_cost;
4452 /* Decrease the benefit to count the add-insns that we will
4453 insert to increment the reduced reg for the giv.
4454 ??? This can overestimate the run-time cost of the additional
4455 insns, e.g. if there are multiple basic blocks that increment
4456 the biv, but only one of these blocks is executed during each
4457 iteration. There is no good way to detect cases like this with
4458 the current structure of the loop optimizer.
4459 This code is more accurate for determining code size than
4460 run-time benefits. */
4461 benefit -= add_cost * bl->biv_count;
4463 /* Decide whether to strength-reduce this giv or to leave the code
4464 unchanged (recompute it from the biv each time it is used).
4465 This decision can be made independently for each giv. */
4468 /* Attempt to guess whether autoincrement will handle some of the
4469 new add insns; if so, increase BENEFIT (undo the subtraction of
4470 add_cost that was done above). */
4471 if (v->giv_type == DEST_ADDR
4472 /* Increasing the benefit is risky, since this is only a guess.
4473 Avoid increasing register pressure in cases where there would
4474 be no other benefit from reducing this giv. */
4476 && GET_CODE (v->mult_val) == CONST_INT)
4478 if (HAVE_POST_INCREMENT
4479 && INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
4480 benefit += add_cost * bl->biv_count;
4481 else if (HAVE_PRE_INCREMENT
4482 && INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
4483 benefit += add_cost * bl->biv_count;
4484 else if (HAVE_POST_DECREMENT
4485 && -INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
4486 benefit += add_cost * bl->biv_count;
4487 else if (HAVE_PRE_DECREMENT
4488 && -INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
4489 benefit += add_cost * bl->biv_count;
4493 /* If an insn is not to be strength reduced, then set its ignore
4494 flag, and clear all_reduced. */
4496 /* A giv that depends on a reversed biv must be reduced if it is
4497 used after the loop exit, otherwise, it would have the wrong
4498 value after the loop exit. To make it simple, just reduce all
4499 of such giv's whether or not we know they are used after the loop
4502 if ( ! flag_reduce_all_givs && v->lifetime * threshold * benefit < insn_count
4505 if (loop_dump_stream)
4506 fprintf (loop_dump_stream,
4507 "giv of insn %d not worth while, %d vs %d.\n",
4509 v->lifetime * threshold * benefit, insn_count);
4515 /* Check that we can increment the reduced giv without a
4516 multiply insn. If not, reject it. */
4518 for (tv = bl->biv; tv; tv = tv->next_iv)
4519 if (tv->mult_val == const1_rtx
4520 && ! product_cheap_p (tv->add_val, v->mult_val))
4522 if (loop_dump_stream)
4523 fprintf (loop_dump_stream,
4524 "giv of insn %d: would need a multiply.\n",
4525 INSN_UID (v->insn));
4533 /* Check for givs whose first use is their definition and whose
4534 last use is the definition of another giv. If so, it is likely
4535 dead and should not be used to derive another giv nor to
4537 for (v = bl->giv; v; v = v->next_iv)
4540 || (v->same && v->same->ignore))
4545 struct induction *v1;
4547 for (v1 = bl->giv; v1; v1 = v1->next_iv)
4548 if (v->last_use == v1->insn)
4551 else if (v->giv_type == DEST_REG
4552 && REGNO_FIRST_UID (REGNO (v->dest_reg)) == INSN_UID (v->insn))
4554 struct induction *v1;
4556 for (v1 = bl->giv; v1; v1 = v1->next_iv)
4557 if (REGNO_LAST_UID (REGNO (v->dest_reg)) == INSN_UID (v1->insn))
4562 /* Now that we know which givs will be reduced, try to rearrange the
4563 combinations to reduce register pressure.
4564 recombine_givs calls find_life_end, which needs reg_iv_type and
4565 reg_iv_info to be valid for all pseudos. We do the necessary
4566 reallocation here since it allows to check if there are still
4567 more bivs to process. */
4568 nregs = max_reg_num ();
4569 if (nregs > reg_iv_type->num_elements)
4571 /* If there are still more bivs to process, allocate some slack
4572 space so that we're not constantly reallocating these arrays. */
4575 /* Reallocate reg_iv_type and reg_iv_info. */
4576 VARRAY_GROW (reg_iv_type, nregs);
4577 VARRAY_GROW (reg_iv_info, nregs);
4579 recombine_givs (loop, bl, flags & LOOP_UNROLL);
4581 /* Reduce each giv that we decided to reduce. */
4583 for (v = bl->giv; v; v = v->next_iv)
4585 struct induction *tv;
4586 if (! v->ignore && v->same == 0)
4588 int auto_inc_opt = 0;
4590 /* If the code for derived givs immediately below has already
4591 allocated a new_reg, we must keep it. */
4593 v->new_reg = gen_reg_rtx (v->mode);
4595 if (v->derived_from)
4597 struct induction *d = v->derived_from;
4599 /* In case d->dest_reg is not replaceable, we have
4600 to replace it in v->insn now. */
4602 d->new_reg = gen_reg_rtx (d->mode);
4604 = replace_rtx (PATTERN (v->insn), d->dest_reg, d->new_reg);
4606 = replace_rtx (PATTERN (v->insn), v->dest_reg, v->new_reg);
4607 /* For each place where the biv is incremented, add an
4608 insn to set the new, reduced reg for the giv.
4609 We used to do this only for biv_count != 1, but
4610 this fails when there is a giv after a single biv
4611 increment, e.g. when the last giv was expressed as
4613 for (tv = bl->biv; tv; tv = tv->next_iv)
4615 /* We always emit reduced giv increments before the
4616 biv increment when bl->biv_count != 1. So by
4617 emitting the add insns for derived givs after the
4618 biv increment, they pick up the updated value of
4620 If the reduced giv is processed with
4621 auto_inc_opt == 1, then it is incremented earlier
4622 than the biv, hence we'll still pick up the right
4624 If it's processed with auto_inc_opt == -1,
4625 that implies that the biv increment is before the
4626 first reduced giv's use. The derived giv's lifetime
4627 is after the reduced giv's lifetime, hence in this
4628 case, the biv increment doesn't matter. */
4629 emit_insn_after (copy_rtx (PATTERN (v->insn)), tv->insn);
4635 /* If the target has auto-increment addressing modes, and
4636 this is an address giv, then try to put the increment
4637 immediately after its use, so that flow can create an
4638 auto-increment addressing mode. */
4639 if (v->giv_type == DEST_ADDR && bl->biv_count == 1
4640 && bl->biv->always_executed && ! bl->biv->maybe_multiple
4641 /* We don't handle reversed biv's because bl->biv->insn
4642 does not have a valid INSN_LUID. */
4644 && v->always_executed && ! v->maybe_multiple
4645 && INSN_UID (v->insn) < max_uid_for_loop)
4647 /* If other giv's have been combined with this one, then
4648 this will work only if all uses of the other giv's occur
4649 before this giv's insn. This is difficult to check.
4651 We simplify this by looking for the common case where
4652 there is one DEST_REG giv, and this giv's insn is the
4653 last use of the dest_reg of that DEST_REG giv. If the
4654 increment occurs after the address giv, then we can
4655 perform the optimization. (Otherwise, the increment
4656 would have to go before other_giv, and we would not be
4657 able to combine it with the address giv to get an
4658 auto-inc address.) */
4659 if (v->combined_with)
4661 struct induction *other_giv = 0;
4663 for (tv = bl->giv; tv; tv = tv->next_iv)
4671 if (! tv && other_giv
4672 && REGNO (other_giv->dest_reg) < max_reg_before_loop
4673 && (REGNO_LAST_UID (REGNO (other_giv->dest_reg))
4674 == INSN_UID (v->insn))
4675 && INSN_LUID (v->insn) < INSN_LUID (bl->biv->insn))
4678 /* Check for case where increment is before the address
4679 giv. Do this test in "loop order". */
4680 else if ((INSN_LUID (v->insn) > INSN_LUID (bl->biv->insn)
4681 && (INSN_LUID (v->insn) < INSN_LUID (loop_scan_start)
4682 || (INSN_LUID (bl->biv->insn)
4683 > INSN_LUID (loop_scan_start))))
4684 || (INSN_LUID (v->insn) < INSN_LUID (loop_scan_start)
4685 && (INSN_LUID (loop_scan_start)
4686 < INSN_LUID (bl->biv->insn))))
4695 /* We can't put an insn immediately after one setting
4696 cc0, or immediately before one using cc0. */
4697 if ((auto_inc_opt == 1 && sets_cc0_p (PATTERN (v->insn)))
4698 || (auto_inc_opt == -1
4699 && (prev = prev_nonnote_insn (v->insn)) != 0
4701 && sets_cc0_p (PATTERN (prev))))
4707 v->auto_inc_opt = 1;
4711 /* For each place where the biv is incremented, add an insn
4712 to increment the new, reduced reg for the giv. */
4713 for (tv = bl->biv; tv; tv = tv->next_iv)
4718 insert_before = tv->insn;
4719 else if (auto_inc_opt == 1)
4720 insert_before = NEXT_INSN (v->insn);
4722 insert_before = v->insn;
4724 if (tv->mult_val == const1_rtx)
4725 emit_iv_add_mult (tv->add_val, v->mult_val,
4726 v->new_reg, v->new_reg, insert_before);
4727 else /* tv->mult_val == const0_rtx */
4728 /* A multiply is acceptable here
4729 since this is presumed to be seldom executed. */
4730 emit_iv_add_mult (tv->add_val, v->mult_val,
4731 v->add_val, v->new_reg, insert_before);
4734 /* Add code at loop start to initialize giv's reduced reg. */
4736 emit_iv_add_mult (bl->initial_value, v->mult_val,
4737 v->add_val, v->new_reg, loop_start);
4741 /* Rescan all givs. If a giv is the same as a giv not reduced, mark it
4744 For each giv register that can be reduced now: if replaceable,
4745 substitute reduced reg wherever the old giv occurs;
4746 else add new move insn "giv_reg = reduced_reg". */
4748 for (v = bl->giv; v; v = v->next_iv)
4750 if (v->same && v->same->ignore)
4756 /* Update expression if this was combined, in case other giv was
4759 v->new_reg = replace_rtx (v->new_reg,
4760 v->same->dest_reg, v->same->new_reg);
4762 if (v->giv_type == DEST_ADDR)
4763 /* Store reduced reg as the address in the memref where we found
4765 validate_change (v->insn, v->location, v->new_reg, 0);
4766 else if (v->replaceable)
4768 reg_map[REGNO (v->dest_reg)] = v->new_reg;
4771 /* I can no longer duplicate the original problem. Perhaps
4772 this is unnecessary now? */
4774 /* Replaceable; it isn't strictly necessary to delete the old
4775 insn and emit a new one, because v->dest_reg is now dead.
4777 However, especially when unrolling loops, the special
4778 handling for (set REG0 REG1) in the second cse pass may
4779 make v->dest_reg live again. To avoid this problem, emit
4780 an insn to set the original giv reg from the reduced giv.
4781 We can not delete the original insn, since it may be part
4782 of a LIBCALL, and the code in flow that eliminates dead
4783 libcalls will fail if it is deleted. */
4784 emit_insn_after (gen_move_insn (v->dest_reg, v->new_reg),
4790 /* Not replaceable; emit an insn to set the original giv reg from
4791 the reduced giv, same as above. */
4792 emit_insn_after (gen_move_insn (v->dest_reg, v->new_reg),
4796 /* When a loop is reversed, givs which depend on the reversed
4797 biv, and which are live outside the loop, must be set to their
4798 correct final value. This insn is only needed if the giv is
4799 not replaceable. The correct final value is the same as the
4800 value that the giv starts the reversed loop with. */
4801 if (bl->reversed && ! v->replaceable)
4802 emit_iv_add_mult (bl->initial_value, v->mult_val,
4803 v->add_val, v->dest_reg, end_insert_before);
4804 else if (v->final_value)
4808 /* If the loop has multiple exits, emit the insn before the
4809 loop to ensure that it will always be executed no matter
4810 how the loop exits. Otherwise, emit the insn after the loop,
4811 since this is slightly more efficient. */
4812 if (loop->exit_count)
4813 insert_before = loop_start;
4815 insert_before = end_insert_before;
4816 emit_insn_before (gen_move_insn (v->dest_reg, v->final_value),
4820 /* If the insn to set the final value of the giv was emitted
4821 before the loop, then we must delete the insn inside the loop
4822 that sets it. If this is a LIBCALL, then we must delete
4823 every insn in the libcall. Note, however, that
4824 final_giv_value will only succeed when there are multiple
4825 exits if the giv is dead at each exit, hence it does not
4826 matter that the original insn remains because it is dead
4828 /* Delete the insn inside the loop that sets the giv since
4829 the giv is now set before (or after) the loop. */
4830 delete_insn (v->insn);
4834 if (loop_dump_stream)
4836 fprintf (loop_dump_stream, "giv at %d reduced to ",
4837 INSN_UID (v->insn));
4838 print_rtl (loop_dump_stream, v->new_reg);
4839 fprintf (loop_dump_stream, "\n");
4843 /* All the givs based on the biv bl have been reduced if they
4846 /* For each giv not marked as maybe dead that has been combined with a
4847 second giv, clear any "maybe dead" mark on that second giv.
4848 v->new_reg will either be or refer to the register of the giv it
4851 Doing this clearing avoids problems in biv elimination where a
4852 giv's new_reg is a complex value that can't be put in the insn but
4853 the giv combined with (with a reg as new_reg) is marked maybe_dead.
4854 Since the register will be used in either case, we'd prefer it be
4855 used from the simpler giv. */
4857 for (v = bl->giv; v; v = v->next_iv)
4858 if (! v->maybe_dead && v->same)
4859 v->same->maybe_dead = 0;
4861 /* Try to eliminate the biv, if it is a candidate.
4862 This won't work if ! all_reduced,
4863 since the givs we planned to use might not have been reduced.
4865 We have to be careful that we didn't initially think we could eliminate
4866 this biv because of a giv that we now think may be dead and shouldn't
4867 be used as a biv replacement.
4869 Also, there is the possibility that we may have a giv that looks
4870 like it can be used to eliminate a biv, but the resulting insn
4871 isn't valid. This can happen, for example, on the 88k, where a
4872 JUMP_INSN can compare a register only with zero. Attempts to
4873 replace it with a compare with a constant will fail.
4875 Note that in cases where this call fails, we may have replaced some
4876 of the occurrences of the biv with a giv, but no harm was done in
4877 doing so in the rare cases where it can occur. */
4879 if (all_reduced == 1 && bl->eliminable
4880 && maybe_eliminate_biv (loop, bl, 1, threshold, insn_count))
4882 /* ?? If we created a new test to bypass the loop entirely,
4883 or otherwise drop straight in, based on this test, then
4884 we might want to rewrite it also. This way some later
4885 pass has more hope of removing the initialization of this
4888 /* If final_value != 0, then the biv may be used after loop end
4889 and we must emit an insn to set it just in case.
4891 Reversed bivs already have an insn after the loop setting their
4892 value, so we don't need another one. We can't calculate the
4893 proper final value for such a biv here anyways. */
4894 if (final_value != 0 && ! bl->reversed)
4898 /* If the loop has multiple exits, emit the insn before the
4899 loop to ensure that it will always be executed no matter
4900 how the loop exits. Otherwise, emit the insn after the
4901 loop, since this is slightly more efficient. */
4902 if (loop->exit_count)
4903 insert_before = loop_start;
4905 insert_before = end_insert_before;
4907 emit_insn_before (gen_move_insn (bl->biv->dest_reg, final_value),
4912 /* Delete all of the instructions inside the loop which set
4913 the biv, as they are all dead. If is safe to delete them,
4914 because an insn setting a biv will never be part of a libcall. */
4915 /* However, deleting them will invalidate the regno_last_uid info,
4916 so keeping them around is more convenient. Final_biv_value
4917 will only succeed when there are multiple exits if the biv
4918 is dead at each exit, hence it does not matter that the original
4919 insn remains, because it is dead anyways. */
4920 for (v = bl->biv; v; v = v->next_iv)
4921 delete_insn (v->insn);
4924 if (loop_dump_stream)
4925 fprintf (loop_dump_stream, "Reg %d: biv eliminated\n",
4930 /* Go through all the instructions in the loop, making all the
4931 register substitutions scheduled in REG_MAP. */
4933 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
4934 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
4935 || GET_CODE (p) == CALL_INSN)
4937 replace_regs (PATTERN (p), reg_map, reg_map_size, 0);
4938 replace_regs (REG_NOTES (p), reg_map, reg_map_size, 0);
4942 if (loop_info->n_iterations > 0)
4944 /* When we completely unroll a loop we will likely not need the increment
4945 of the loop BIV and we will not need the conditional branch at the
4947 unrolled_insn_copies = insn_count - 2;
4950 /* When we completely unroll a loop on a HAVE_cc0 machine we will not
4951 need the comparison before the conditional branch at the end of the
4953 unrolled_insn_copies -= 1;
4956 /* We'll need one copy for each loop iteration. */
4957 unrolled_insn_copies *= loop_info->n_iterations;
4959 /* A little slop to account for the ability to remove initialization
4960 code, better CSE, and other secondary benefits of completely
4961 unrolling some loops. */
4962 unrolled_insn_copies -= 1;
4964 /* Clamp the value. */
4965 if (unrolled_insn_copies < 0)
4966 unrolled_insn_copies = 0;
4969 /* Unroll loops from within strength reduction so that we can use the
4970 induction variable information that strength_reduce has already
4971 collected. Always unroll loops that would be as small or smaller
4972 unrolled than when rolled. */
4973 if ((flags & LOOP_UNROLL)
4974 || (loop_info->n_iterations > 0
4975 && unrolled_insn_copies <= insn_count))
4976 unroll_loop (loop, insn_count, end_insert_before, 1);
4978 #ifdef HAVE_doloop_end
4979 if (HAVE_doloop_end && (flags & LOOP_BCT) && flag_branch_on_count_reg)
4980 doloop_optimize (loop);
4981 #endif /* HAVE_doloop_end */
4983 if (loop_dump_stream)
4984 fprintf (loop_dump_stream, "\n");
4987 VARRAY_FREE (reg_iv_type);
4988 VARRAY_FREE (reg_iv_info);
4989 free (reg_biv_class);
4994 /*Record all basic induction variables calculated in the insn. */
4996 check_insn_for_bivs (loop, p, not_every_iteration, maybe_multiple)
4999 int not_every_iteration;
5008 if (GET_CODE (p) == INSN
5009 && (set = single_set (p))
5010 && GET_CODE (SET_DEST (set)) == REG)
5012 dest_reg = SET_DEST (set);
5013 if (REGNO (dest_reg) < max_reg_before_loop
5014 && REGNO (dest_reg) >= FIRST_PSEUDO_REGISTER
5015 && REG_IV_TYPE (REGNO (dest_reg)) != NOT_BASIC_INDUCT)
5017 if (basic_induction_var (loop, SET_SRC (set),
5018 GET_MODE (SET_SRC (set)),
5019 dest_reg, p, &inc_val, &mult_val,
5022 /* It is a possible basic induction variable.
5023 Create and initialize an induction structure for it. */
5026 = (struct induction *) oballoc (sizeof (struct induction));
5028 record_biv (v, p, dest_reg, inc_val, mult_val, location,
5029 not_every_iteration, maybe_multiple);
5030 REG_IV_TYPE (REGNO (dest_reg)) = BASIC_INDUCT;
5032 else if (REGNO (dest_reg) < max_reg_before_loop)
5033 REG_IV_TYPE (REGNO (dest_reg)) = NOT_BASIC_INDUCT;
5039 /* Record all givs calculated in the insn.
5040 A register is a giv if: it is only set once, it is a function of a
5041 biv and a constant (or invariant), and it is not a biv. */
5043 check_insn_for_givs (loop, p, not_every_iteration, maybe_multiple)
5046 int not_every_iteration;
5050 /* Look for a general induction variable in a register. */
5051 if (GET_CODE (p) == INSN
5052 && (set = single_set (p))
5053 && GET_CODE (SET_DEST (set)) == REG
5054 && ! VARRAY_CHAR (may_not_optimize, REGNO (SET_DEST (set))))
5062 rtx last_consec_insn;
5064 dest_reg = SET_DEST (set);
5065 if (REGNO (dest_reg) < FIRST_PSEUDO_REGISTER)
5068 if (/* SET_SRC is a giv. */
5069 (general_induction_var (loop, SET_SRC (set), &src_reg, &add_val,
5070 &mult_val, 0, &benefit, VOIDmode)
5071 /* Equivalent expression is a giv. */
5072 || ((regnote = find_reg_note (p, REG_EQUAL, NULL_RTX))
5073 && general_induction_var (loop, XEXP (regnote, 0), &src_reg,
5074 &add_val, &mult_val, 0,
5075 &benefit, VOIDmode)))
5076 /* Don't try to handle any regs made by loop optimization.
5077 We have nothing on them in regno_first_uid, etc. */
5078 && REGNO (dest_reg) < max_reg_before_loop
5079 /* Don't recognize a BASIC_INDUCT_VAR here. */
5080 && dest_reg != src_reg
5081 /* This must be the only place where the register is set. */
5082 && (VARRAY_INT (n_times_set, REGNO (dest_reg)) == 1
5083 /* or all sets must be consecutive and make a giv. */
5084 || (benefit = consec_sets_giv (loop, benefit, p,
5086 &add_val, &mult_val,
5087 &last_consec_insn))))
5090 = (struct induction *) oballoc (sizeof (struct induction));
5092 /* If this is a library call, increase benefit. */
5093 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
5094 benefit += libcall_benefit (p);
5096 /* Skip the consecutive insns, if there are any. */
5097 if (VARRAY_INT (n_times_set, REGNO (dest_reg)) != 1)
5098 p = last_consec_insn;
5100 record_giv (loop, v, p, src_reg, dest_reg, mult_val, add_val,
5101 benefit, DEST_REG, not_every_iteration,
5102 maybe_multiple, NULL_PTR);
5107 #ifndef DONT_REDUCE_ADDR
5108 /* Look for givs which are memory addresses. */
5109 /* This resulted in worse code on a VAX 8600. I wonder if it
5111 if (GET_CODE (p) == INSN)
5112 find_mem_givs (loop, PATTERN (p), p, not_every_iteration,
5116 /* Update the status of whether giv can derive other givs. This can
5117 change when we pass a label or an insn that updates a biv. */
5118 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
5119 || GET_CODE (p) == CODE_LABEL)
5120 update_giv_derive (loop, p);
5124 /* Return 1 if X is a valid source for an initial value (or as value being
5125 compared against in an initial test).
5127 X must be either a register or constant and must not be clobbered between
5128 the current insn and the start of the loop.
5130 INSN is the insn containing X. */
5133 valid_initial_value_p (x, insn, call_seen, loop_start)
5142 /* Only consider pseudos we know about initialized in insns whose luids
5144 if (GET_CODE (x) != REG
5145 || REGNO (x) >= max_reg_before_loop)
5148 /* Don't use call-clobbered registers across a call which clobbers it. On
5149 some machines, don't use any hard registers at all. */
5150 if (REGNO (x) < FIRST_PSEUDO_REGISTER
5151 && (SMALL_REGISTER_CLASSES
5152 || (call_used_regs[REGNO (x)] && call_seen)))
5155 /* Don't use registers that have been clobbered before the start of the
5157 if (reg_set_between_p (x, insn, loop_start))
5163 /* Scan X for memory refs and check each memory address
5164 as a possible giv. INSN is the insn whose pattern X comes from.
5165 NOT_EVERY_ITERATION is 1 if the insn might not be executed during
5166 every loop iteration. MAYBE_MULTIPLE is 1 if the insn might be executed
5167 more thanonce in each loop iteration. */
5170 find_mem_givs (loop, x, insn, not_every_iteration, maybe_multiple)
5171 const struct loop *loop;
5174 int not_every_iteration, maybe_multiple;
5177 register enum rtx_code code;
5178 register const char *fmt;
5183 code = GET_CODE (x);
5207 /* This code used to disable creating GIVs with mult_val == 1 and
5208 add_val == 0. However, this leads to lost optimizations when
5209 it comes time to combine a set of related DEST_ADDR GIVs, since
5210 this one would not be seen. */
5212 if (general_induction_var (loop, XEXP (x, 0), &src_reg, &add_val,
5213 &mult_val, 1, &benefit, GET_MODE (x)))
5215 /* Found one; record it. */
5217 = (struct induction *) oballoc (sizeof (struct induction));
5219 record_giv (loop, v, insn, src_reg, addr_placeholder, mult_val,
5220 add_val, benefit, DEST_ADDR, not_every_iteration,
5221 maybe_multiple, &XEXP (x, 0));
5223 v->mem_mode = GET_MODE (x);
5232 /* Recursively scan the subexpressions for other mem refs. */
5234 fmt = GET_RTX_FORMAT (code);
5235 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
5237 find_mem_givs (loop, XEXP (x, i), insn, not_every_iteration,
5239 else if (fmt[i] == 'E')
5240 for (j = 0; j < XVECLEN (x, i); j++)
5241 find_mem_givs (loop, XVECEXP (x, i, j), insn, not_every_iteration,
5245 /* Fill in the data about one biv update.
5246 V is the `struct induction' in which we record the biv. (It is
5247 allocated by the caller, with alloca.)
5248 INSN is the insn that sets it.
5249 DEST_REG is the biv's reg.
5251 MULT_VAL is const1_rtx if the biv is being incremented here, in which case
5252 INC_VAL is the increment. Otherwise, MULT_VAL is const0_rtx and the biv is
5253 being set to INC_VAL.
5255 NOT_EVERY_ITERATION is nonzero if this biv update is not know to be
5256 executed every iteration; MAYBE_MULTIPLE is nonzero if this biv update
5257 can be executed more than once per iteration. If MAYBE_MULTIPLE
5258 and NOT_EVERY_ITERATION are both zero, we know that the biv update is
5259 executed exactly once per iteration. */
5262 record_biv (v, insn, dest_reg, inc_val, mult_val, location,
5263 not_every_iteration, maybe_multiple)
5264 struct induction *v;
5270 int not_every_iteration;
5273 struct iv_class *bl;
5276 v->src_reg = dest_reg;
5277 v->dest_reg = dest_reg;
5278 v->mult_val = mult_val;
5279 v->add_val = inc_val;
5280 v->location = location;
5281 v->mode = GET_MODE (dest_reg);
5282 v->always_computable = ! not_every_iteration;
5283 v->always_executed = ! not_every_iteration;
5284 v->maybe_multiple = maybe_multiple;
5286 /* Add this to the reg's iv_class, creating a class
5287 if this is the first incrementation of the reg. */
5289 bl = reg_biv_class[REGNO (dest_reg)];
5292 /* Create and initialize new iv_class. */
5294 bl = (struct iv_class *) oballoc (sizeof (struct iv_class));
5296 bl->regno = REGNO (dest_reg);
5302 /* Set initial value to the reg itself. */
5303 bl->initial_value = dest_reg;
5304 /* We haven't seen the initializing insn yet */
5307 bl->initial_test = 0;
5308 bl->incremented = 0;
5312 bl->total_benefit = 0;
5314 /* Add this class to loop_iv_list. */
5315 bl->next = loop_iv_list;
5318 /* Put it in the array of biv register classes. */
5319 reg_biv_class[REGNO (dest_reg)] = bl;
5322 /* Update IV_CLASS entry for this biv. */
5323 v->next_iv = bl->biv;
5326 if (mult_val == const1_rtx)
5327 bl->incremented = 1;
5329 if (loop_dump_stream)
5331 fprintf (loop_dump_stream,
5332 "Insn %d: possible biv, reg %d,",
5333 INSN_UID (insn), REGNO (dest_reg));
5334 if (GET_CODE (inc_val) == CONST_INT)
5336 fprintf (loop_dump_stream, " const =");
5337 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (inc_val));
5338 fputc ('\n', loop_dump_stream);
5342 fprintf (loop_dump_stream, " const = ");
5343 print_rtl (loop_dump_stream, inc_val);
5344 fprintf (loop_dump_stream, "\n");
5349 /* Fill in the data about one giv.
5350 V is the `struct induction' in which we record the giv. (It is
5351 allocated by the caller, with alloca.)
5352 INSN is the insn that sets it.
5353 BENEFIT estimates the savings from deleting this insn.
5354 TYPE is DEST_REG or DEST_ADDR; it says whether the giv is computed
5355 into a register or is used as a memory address.
5357 SRC_REG is the biv reg which the giv is computed from.
5358 DEST_REG is the giv's reg (if the giv is stored in a reg).
5359 MULT_VAL and ADD_VAL are the coefficients used to compute the giv.
5360 LOCATION points to the place where this giv's value appears in INSN. */
5363 record_giv (loop, v, insn, src_reg, dest_reg, mult_val, add_val, benefit,
5364 type, not_every_iteration, maybe_multiple, location)
5365 const struct loop *loop;
5366 struct induction *v;
5370 rtx mult_val, add_val;
5373 int not_every_iteration, maybe_multiple;
5376 struct induction *b;
5377 struct iv_class *bl;
5378 rtx set = single_set (insn);
5381 /* Attempt to prove constantness of the values. */
5382 temp = simplify_rtx (add_val);
5387 v->src_reg = src_reg;
5389 v->dest_reg = dest_reg;
5390 v->mult_val = mult_val;
5391 v->add_val = add_val;
5392 v->benefit = benefit;
5393 v->location = location;
5395 v->combined_with = 0;
5396 v->maybe_multiple = maybe_multiple;
5398 v->derive_adjustment = 0;
5404 v->auto_inc_opt = 0;
5407 v->derived_from = 0;
5410 /* The v->always_computable field is used in update_giv_derive, to
5411 determine whether a giv can be used to derive another giv. For a
5412 DEST_REG giv, INSN computes a new value for the giv, so its value
5413 isn't computable if INSN insn't executed every iteration.
5414 However, for a DEST_ADDR giv, INSN merely uses the value of the giv;
5415 it does not compute a new value. Hence the value is always computable
5416 regardless of whether INSN is executed each iteration. */
5418 if (type == DEST_ADDR)
5419 v->always_computable = 1;
5421 v->always_computable = ! not_every_iteration;
5423 v->always_executed = ! not_every_iteration;
5425 if (type == DEST_ADDR)
5427 v->mode = GET_MODE (*location);
5430 else /* type == DEST_REG */
5432 v->mode = GET_MODE (SET_DEST (set));
5434 v->lifetime = (uid_luid[REGNO_LAST_UID (REGNO (dest_reg))]
5435 - uid_luid[REGNO_FIRST_UID (REGNO (dest_reg))]);
5437 /* If the lifetime is zero, it means that this register is
5438 really a dead store. So mark this as a giv that can be
5439 ignored. This will not prevent the biv from being eliminated. */
5440 if (v->lifetime == 0)
5443 REG_IV_TYPE (REGNO (dest_reg)) = GENERAL_INDUCT;
5444 REG_IV_INFO (REGNO (dest_reg)) = v;
5447 /* Add the giv to the class of givs computed from one biv. */
5449 bl = reg_biv_class[REGNO (src_reg)];
5452 v->next_iv = bl->giv;
5454 /* Don't count DEST_ADDR. This is supposed to count the number of
5455 insns that calculate givs. */
5456 if (type == DEST_REG)
5458 bl->total_benefit += benefit;
5461 /* Fatal error, biv missing for this giv? */
5464 if (type == DEST_ADDR)
5468 /* The giv can be replaced outright by the reduced register only if all
5469 of the following conditions are true:
5470 - the insn that sets the giv is always executed on any iteration
5471 on which the giv is used at all
5472 (there are two ways to deduce this:
5473 either the insn is executed on every iteration,
5474 or all uses follow that insn in the same basic block),
5475 - the giv is not used outside the loop
5476 - no assignments to the biv occur during the giv's lifetime. */
5478 if (REGNO_FIRST_UID (REGNO (dest_reg)) == INSN_UID (insn)
5479 /* Previous line always fails if INSN was moved by loop opt. */
5480 && uid_luid[REGNO_LAST_UID (REGNO (dest_reg))]
5481 < INSN_LUID (loop->end)
5482 && (! not_every_iteration
5483 || last_use_this_basic_block (dest_reg, insn)))
5485 /* Now check that there are no assignments to the biv within the
5486 giv's lifetime. This requires two separate checks. */
5488 /* Check each biv update, and fail if any are between the first
5489 and last use of the giv.
5491 If this loop contains an inner loop that was unrolled, then
5492 the insn modifying the biv may have been emitted by the loop
5493 unrolling code, and hence does not have a valid luid. Just
5494 mark the biv as not replaceable in this case. It is not very
5495 useful as a biv, because it is used in two different loops.
5496 It is very unlikely that we would be able to optimize the giv
5497 using this biv anyways. */
5500 for (b = bl->biv; b; b = b->next_iv)
5502 if (INSN_UID (b->insn) >= max_uid_for_loop
5503 || ((uid_luid[INSN_UID (b->insn)]
5504 >= uid_luid[REGNO_FIRST_UID (REGNO (dest_reg))])
5505 && (uid_luid[INSN_UID (b->insn)]
5506 <= uid_luid[REGNO_LAST_UID (REGNO (dest_reg))])))
5509 v->not_replaceable = 1;
5514 /* If there are any backwards branches that go from after the
5515 biv update to before it, then this giv is not replaceable. */
5517 for (b = bl->biv; b; b = b->next_iv)
5518 if (back_branch_in_range_p (loop, b->insn))
5521 v->not_replaceable = 1;
5527 /* May still be replaceable, we don't have enough info here to
5530 v->not_replaceable = 0;
5534 /* Record whether the add_val contains a const_int, for later use by
5539 v->no_const_addval = 1;
5540 if (tem == const0_rtx)
5542 else if (CONSTANT_P (add_val))
5543 v->no_const_addval = 0;
5544 if (GET_CODE (tem) == PLUS)
5548 if (GET_CODE (XEXP (tem, 0)) == PLUS)
5549 tem = XEXP (tem, 0);
5550 else if (GET_CODE (XEXP (tem, 1)) == PLUS)
5551 tem = XEXP (tem, 1);
5555 if (CONSTANT_P (XEXP (tem, 1)))
5556 v->no_const_addval = 0;
5560 if (loop_dump_stream)
5562 if (type == DEST_REG)
5563 fprintf (loop_dump_stream, "Insn %d: giv reg %d",
5564 INSN_UID (insn), REGNO (dest_reg));
5566 fprintf (loop_dump_stream, "Insn %d: dest address",
5569 fprintf (loop_dump_stream, " src reg %d benefit %d",
5570 REGNO (src_reg), v->benefit);
5571 fprintf (loop_dump_stream, " lifetime %d",
5575 fprintf (loop_dump_stream, " replaceable");
5577 if (v->no_const_addval)
5578 fprintf (loop_dump_stream, " ncav");
5580 if (GET_CODE (mult_val) == CONST_INT)
5582 fprintf (loop_dump_stream, " mult ");
5583 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (mult_val));
5587 fprintf (loop_dump_stream, " mult ");
5588 print_rtl (loop_dump_stream, mult_val);
5591 if (GET_CODE (add_val) == CONST_INT)
5593 fprintf (loop_dump_stream, " add ");
5594 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (add_val));
5598 fprintf (loop_dump_stream, " add ");
5599 print_rtl (loop_dump_stream, add_val);
5603 if (loop_dump_stream)
5604 fprintf (loop_dump_stream, "\n");
5608 /* All this does is determine whether a giv can be made replaceable because
5609 its final value can be calculated. This code can not be part of record_giv
5610 above, because final_giv_value requires that the number of loop iterations
5611 be known, and that can not be accurately calculated until after all givs
5612 have been identified. */
5615 check_final_value (loop, v)
5616 const struct loop *loop;
5617 struct induction *v;
5619 struct iv_class *bl;
5620 rtx final_value = 0;
5622 bl = reg_biv_class[REGNO (v->src_reg)];
5624 /* DEST_ADDR givs will never reach here, because they are always marked
5625 replaceable above in record_giv. */
5627 /* The giv can be replaced outright by the reduced register only if all
5628 of the following conditions are true:
5629 - the insn that sets the giv is always executed on any iteration
5630 on which the giv is used at all
5631 (there are two ways to deduce this:
5632 either the insn is executed on every iteration,
5633 or all uses follow that insn in the same basic block),
5634 - its final value can be calculated (this condition is different
5635 than the one above in record_giv)
5636 - no assignments to the biv occur during the giv's lifetime. */
5639 /* This is only called now when replaceable is known to be false. */
5640 /* Clear replaceable, so that it won't confuse final_giv_value. */
5644 if ((final_value = final_giv_value (loop, v))
5645 && (v->always_computable || last_use_this_basic_block (v->dest_reg, v->insn)))
5647 int biv_increment_seen = 0;
5653 /* When trying to determine whether or not a biv increment occurs
5654 during the lifetime of the giv, we can ignore uses of the variable
5655 outside the loop because final_value is true. Hence we can not
5656 use regno_last_uid and regno_first_uid as above in record_giv. */
5658 /* Search the loop to determine whether any assignments to the
5659 biv occur during the giv's lifetime. Start with the insn
5660 that sets the giv, and search around the loop until we come
5661 back to that insn again.
5663 Also fail if there is a jump within the giv's lifetime that jumps
5664 to somewhere outside the lifetime but still within the loop. This
5665 catches spaghetti code where the execution order is not linear, and
5666 hence the above test fails. Here we assume that the giv lifetime
5667 does not extend from one iteration of the loop to the next, so as
5668 to make the test easier. Since the lifetime isn't known yet,
5669 this requires two loops. See also record_giv above. */
5671 last_giv_use = v->insn;
5677 p = NEXT_INSN (loop->start);
5681 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
5682 || GET_CODE (p) == CALL_INSN)
5684 /* It is possible for the BIV increment to use the GIV if we
5685 have a cycle. Thus we must be sure to check each insn for
5686 both BIV and GIV uses, and we must check for BIV uses
5689 if (! biv_increment_seen
5690 && reg_set_p (v->src_reg, PATTERN (p)))
5691 biv_increment_seen = 1;
5693 if (reg_mentioned_p (v->dest_reg, PATTERN (p)))
5695 if (biv_increment_seen)
5698 v->not_replaceable = 1;
5706 /* Now that the lifetime of the giv is known, check for branches
5707 from within the lifetime to outside the lifetime if it is still
5717 p = NEXT_INSN (loop->start);
5718 if (p == last_giv_use)
5721 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p)
5722 && LABEL_NAME (JUMP_LABEL (p))
5723 && ((loop_insn_first_p (JUMP_LABEL (p), v->insn)
5724 && loop_insn_first_p (loop->start, JUMP_LABEL (p)))
5725 || (loop_insn_first_p (last_giv_use, JUMP_LABEL (p))
5726 && loop_insn_first_p (JUMP_LABEL (p), loop->end))))
5729 v->not_replaceable = 1;
5731 if (loop_dump_stream)
5732 fprintf (loop_dump_stream,
5733 "Found branch outside giv lifetime.\n");
5740 /* If it is replaceable, then save the final value. */
5742 v->final_value = final_value;
5745 if (loop_dump_stream && v->replaceable)
5746 fprintf (loop_dump_stream, "Insn %d: giv reg %d final_value replaceable\n",
5747 INSN_UID (v->insn), REGNO (v->dest_reg));
5750 /* Update the status of whether a giv can derive other givs.
5752 We need to do something special if there is or may be an update to the biv
5753 between the time the giv is defined and the time it is used to derive
5756 In addition, a giv that is only conditionally set is not allowed to
5757 derive another giv once a label has been passed.
5759 The cases we look at are when a label or an update to a biv is passed. */
5762 update_giv_derive (loop, p)
5763 const struct loop *loop;
5766 struct iv_class *bl;
5767 struct induction *biv, *giv;
5771 /* Search all IV classes, then all bivs, and finally all givs.
5773 There are three cases we are concerned with. First we have the situation
5774 of a giv that is only updated conditionally. In that case, it may not
5775 derive any givs after a label is passed.
5777 The second case is when a biv update occurs, or may occur, after the
5778 definition of a giv. For certain biv updates (see below) that are
5779 known to occur between the giv definition and use, we can adjust the
5780 giv definition. For others, or when the biv update is conditional,
5781 we must prevent the giv from deriving any other givs. There are two
5782 sub-cases within this case.
5784 If this is a label, we are concerned with any biv update that is done
5785 conditionally, since it may be done after the giv is defined followed by
5786 a branch here (actually, we need to pass both a jump and a label, but
5787 this extra tracking doesn't seem worth it).
5789 If this is a jump, we are concerned about any biv update that may be
5790 executed multiple times. We are actually only concerned about
5791 backward jumps, but it is probably not worth performing the test
5792 on the jump again here.
5794 If this is a biv update, we must adjust the giv status to show that a
5795 subsequent biv update was performed. If this adjustment cannot be done,
5796 the giv cannot derive further givs. */
5798 for (bl = loop_iv_list; bl; bl = bl->next)
5799 for (biv = bl->biv; biv; biv = biv->next_iv)
5800 if (GET_CODE (p) == CODE_LABEL || GET_CODE (p) == JUMP_INSN
5803 for (giv = bl->giv; giv; giv = giv->next_iv)
5805 /* If cant_derive is already true, there is no point in
5806 checking all of these conditions again. */
5807 if (giv->cant_derive)
5810 /* If this giv is conditionally set and we have passed a label,
5811 it cannot derive anything. */
5812 if (GET_CODE (p) == CODE_LABEL && ! giv->always_computable)
5813 giv->cant_derive = 1;
5815 /* Skip givs that have mult_val == 0, since
5816 they are really invariants. Also skip those that are
5817 replaceable, since we know their lifetime doesn't contain
5819 else if (giv->mult_val == const0_rtx || giv->replaceable)
5822 /* The only way we can allow this giv to derive another
5823 is if this is a biv increment and we can form the product
5824 of biv->add_val and giv->mult_val. In this case, we will
5825 be able to compute a compensation. */
5826 else if (biv->insn == p)
5830 if (biv->mult_val == const1_rtx)
5831 tem = simplify_giv_expr (loop,
5832 gen_rtx_MULT (giv->mode,
5837 if (tem && giv->derive_adjustment)
5838 tem = simplify_giv_expr
5840 gen_rtx_PLUS (giv->mode, tem, giv->derive_adjustment),
5844 giv->derive_adjustment = tem;
5846 giv->cant_derive = 1;
5848 else if ((GET_CODE (p) == CODE_LABEL && ! biv->always_computable)
5849 || (GET_CODE (p) == JUMP_INSN && biv->maybe_multiple))
5850 giv->cant_derive = 1;
5855 /* Check whether an insn is an increment legitimate for a basic induction var.
5856 X is the source of insn P, or a part of it.
5857 MODE is the mode in which X should be interpreted.
5859 DEST_REG is the putative biv, also the destination of the insn.
5860 We accept patterns of these forms:
5861 REG = REG + INVARIANT (includes REG = REG - CONSTANT)
5862 REG = INVARIANT + REG
5864 If X is suitable, we return 1, set *MULT_VAL to CONST1_RTX,
5865 store the additive term into *INC_VAL, and store the place where
5866 we found the additive term into *LOCATION.
5868 If X is an assignment of an invariant into DEST_REG, we set
5869 *MULT_VAL to CONST0_RTX, and store the invariant into *INC_VAL.
5871 We also want to detect a BIV when it corresponds to a variable
5872 whose mode was promoted via PROMOTED_MODE. In that case, an increment
5873 of the variable may be a PLUS that adds a SUBREG of that variable to
5874 an invariant and then sign- or zero-extends the result of the PLUS
5877 Most GIVs in such cases will be in the promoted mode, since that is the
5878 probably the natural computation mode (and almost certainly the mode
5879 used for addresses) on the machine. So we view the pseudo-reg containing
5880 the variable as the BIV, as if it were simply incremented.
5882 Note that treating the entire pseudo as a BIV will result in making
5883 simple increments to any GIVs based on it. However, if the variable
5884 overflows in its declared mode but not its promoted mode, the result will
5885 be incorrect. This is acceptable if the variable is signed, since
5886 overflows in such cases are undefined, but not if it is unsigned, since
5887 those overflows are defined. So we only check for SIGN_EXTEND and
5890 If we cannot find a biv, we return 0. */
5893 basic_induction_var (loop, x, mode, dest_reg, p, inc_val, mult_val, location)
5894 const struct loop *loop;
5896 enum machine_mode mode;
5903 register enum rtx_code code;
5907 code = GET_CODE (x);
5912 if (rtx_equal_p (XEXP (x, 0), dest_reg)
5913 || (GET_CODE (XEXP (x, 0)) == SUBREG
5914 && SUBREG_PROMOTED_VAR_P (XEXP (x, 0))
5915 && SUBREG_REG (XEXP (x, 0)) == dest_reg))
5917 argp = &XEXP (x, 1);
5919 else if (rtx_equal_p (XEXP (x, 1), dest_reg)
5920 || (GET_CODE (XEXP (x, 1)) == SUBREG
5921 && SUBREG_PROMOTED_VAR_P (XEXP (x, 1))
5922 && SUBREG_REG (XEXP (x, 1)) == dest_reg))
5924 argp = &XEXP (x, 0);
5930 if (loop_invariant_p (loop, arg) != 1)
5933 *inc_val = convert_modes (GET_MODE (dest_reg), GET_MODE (x), arg, 0);
5934 *mult_val = const1_rtx;
5939 /* If this is a SUBREG for a promoted variable, check the inner
5941 if (SUBREG_PROMOTED_VAR_P (x))
5942 return basic_induction_var (loop, SUBREG_REG (x),
5943 GET_MODE (SUBREG_REG (x)),
5944 dest_reg, p, inc_val, mult_val, location);
5948 /* If this register is assigned in a previous insn, look at its
5949 source, but don't go outside the loop or past a label. */
5951 /* If this sets a register to itself, we would repeat any previous
5952 biv increment if we applied this strategy blindly. */
5953 if (rtx_equal_p (dest_reg, x))
5961 insn = PREV_INSN (insn);
5963 while (insn && GET_CODE (insn) == NOTE
5964 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_BEG);
5968 set = single_set (insn);
5972 if ((SET_DEST (set) == x
5973 || (GET_CODE (SET_DEST (set)) == SUBREG
5974 && (GET_MODE_SIZE (GET_MODE (SET_DEST (set)))
5976 && (GET_MODE_CLASS (GET_MODE (SET_DEST (set)))
5978 && SUBREG_REG (SET_DEST (set)) == x))
5979 && basic_induction_var (loop, SET_SRC (set),
5980 (GET_MODE (SET_SRC (set)) == VOIDmode
5982 : GET_MODE (SET_SRC (set))),
5984 inc_val, mult_val, location))
5987 /* ... fall through ... */
5989 /* Can accept constant setting of biv only when inside inner most loop.
5990 Otherwise, a biv of an inner loop may be incorrectly recognized
5991 as a biv of the outer loop,
5992 causing code to be moved INTO the inner loop. */
5994 if (loop_invariant_p (loop, x) != 1)
5999 /* convert_modes aborts if we try to convert to or from CCmode, so just
6000 exclude that case. It is very unlikely that a condition code value
6001 would be a useful iterator anyways. */
6002 if (loop->level == 1
6003 && GET_MODE_CLASS (mode) != MODE_CC
6004 && GET_MODE_CLASS (GET_MODE (dest_reg)) != MODE_CC)
6006 /* Possible bug here? Perhaps we don't know the mode of X. */
6007 *inc_val = convert_modes (GET_MODE (dest_reg), mode, x, 0);
6008 *mult_val = const0_rtx;
6015 return basic_induction_var (loop, XEXP (x, 0), GET_MODE (XEXP (x, 0)),
6016 dest_reg, p, inc_val, mult_val, location);
6019 /* Similar, since this can be a sign extension. */
6020 for (insn = PREV_INSN (p);
6021 (insn && GET_CODE (insn) == NOTE
6022 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_BEG);
6023 insn = PREV_INSN (insn))
6027 set = single_set (insn);
6029 if (! rtx_equal_p (dest_reg, XEXP (x, 0))
6030 && set && SET_DEST (set) == XEXP (x, 0)
6031 && GET_CODE (XEXP (x, 1)) == CONST_INT
6032 && INTVAL (XEXP (x, 1)) >= 0
6033 && GET_CODE (SET_SRC (set)) == ASHIFT
6034 && XEXP (x, 1) == XEXP (SET_SRC (set), 1))
6035 return basic_induction_var (loop, XEXP (SET_SRC (set), 0),
6036 GET_MODE (XEXP (x, 0)),
6037 dest_reg, insn, inc_val, mult_val,
6046 /* A general induction variable (giv) is any quantity that is a linear
6047 function of a basic induction variable,
6048 i.e. giv = biv * mult_val + add_val.
6049 The coefficients can be any loop invariant quantity.
6050 A giv need not be computed directly from the biv;
6051 it can be computed by way of other givs. */
6053 /* Determine whether X computes a giv.
6054 If it does, return a nonzero value
6055 which is the benefit from eliminating the computation of X;
6056 set *SRC_REG to the register of the biv that it is computed from;
6057 set *ADD_VAL and *MULT_VAL to the coefficients,
6058 such that the value of X is biv * mult + add; */
6061 general_induction_var (loop, x, src_reg, add_val, mult_val, is_addr,
6062 pbenefit, addr_mode)
6063 const struct loop *loop;
6070 enum machine_mode addr_mode;
6075 /* If this is an invariant, forget it, it isn't a giv. */
6076 if (loop_invariant_p (loop, x) == 1)
6079 /* See if the expression could be a giv and get its form.
6080 Mark our place on the obstack in case we don't find a giv. */
6081 storage = (char *) oballoc (0);
6083 x = simplify_giv_expr (loop, x, pbenefit);
6090 switch (GET_CODE (x))
6094 /* Since this is now an invariant and wasn't before, it must be a giv
6095 with MULT_VAL == 0. It doesn't matter which BIV we associate this
6097 *src_reg = loop_iv_list->biv->dest_reg;
6098 *mult_val = const0_rtx;
6103 /* This is equivalent to a BIV. */
6105 *mult_val = const1_rtx;
6106 *add_val = const0_rtx;
6110 /* Either (plus (biv) (invar)) or
6111 (plus (mult (biv) (invar_1)) (invar_2)). */
6112 if (GET_CODE (XEXP (x, 0)) == MULT)
6114 *src_reg = XEXP (XEXP (x, 0), 0);
6115 *mult_val = XEXP (XEXP (x, 0), 1);
6119 *src_reg = XEXP (x, 0);
6120 *mult_val = const1_rtx;
6122 *add_val = XEXP (x, 1);
6126 /* ADD_VAL is zero. */
6127 *src_reg = XEXP (x, 0);
6128 *mult_val = XEXP (x, 1);
6129 *add_val = const0_rtx;
6136 /* Remove any enclosing USE from ADD_VAL and MULT_VAL (there will be
6137 unless they are CONST_INT). */
6138 if (GET_CODE (*add_val) == USE)
6139 *add_val = XEXP (*add_val, 0);
6140 if (GET_CODE (*mult_val) == USE)
6141 *mult_val = XEXP (*mult_val, 0);
6144 *pbenefit += address_cost (orig_x, addr_mode) - reg_address_cost;
6146 *pbenefit += rtx_cost (orig_x, SET);
6148 /* Always return true if this is a giv so it will be detected as such,
6149 even if the benefit is zero or negative. This allows elimination
6150 of bivs that might otherwise not be eliminated. */
6154 /* Given an expression, X, try to form it as a linear function of a biv.
6155 We will canonicalize it to be of the form
6156 (plus (mult (BIV) (invar_1))
6158 with possible degeneracies.
6160 The invariant expressions must each be of a form that can be used as a
6161 machine operand. We surround then with a USE rtx (a hack, but localized
6162 and certainly unambiguous!) if not a CONST_INT for simplicity in this
6163 routine; it is the caller's responsibility to strip them.
6165 If no such canonicalization is possible (i.e., two biv's are used or an
6166 expression that is neither invariant nor a biv or giv), this routine
6169 For a non-zero return, the result will have a code of CONST_INT, USE,
6170 REG (for a BIV), PLUS, or MULT. No other codes will occur.
6172 *BENEFIT will be incremented by the benefit of any sub-giv encountered. */
6174 static rtx sge_plus PARAMS ((enum machine_mode, rtx, rtx));
6175 static rtx sge_plus_constant PARAMS ((rtx, rtx));
6176 static int cmp_combine_givs_stats PARAMS ((const PTR, const PTR));
6177 static int cmp_recombine_givs_stats PARAMS ((const PTR, const PTR));
6180 simplify_giv_expr (loop, x, benefit)
6181 const struct loop *loop;
6185 enum machine_mode mode = GET_MODE (x);
6189 /* If this is not an integer mode, or if we cannot do arithmetic in this
6190 mode, this can't be a giv. */
6191 if (mode != VOIDmode
6192 && (GET_MODE_CLASS (mode) != MODE_INT
6193 || GET_MODE_BITSIZE (mode) > HOST_BITS_PER_WIDE_INT))
6196 switch (GET_CODE (x))
6199 arg0 = simplify_giv_expr (loop, XEXP (x, 0), benefit);
6200 arg1 = simplify_giv_expr (loop, XEXP (x, 1), benefit);
6201 if (arg0 == 0 || arg1 == 0)
6204 /* Put constant last, CONST_INT last if both constant. */
6205 if ((GET_CODE (arg0) == USE
6206 || GET_CODE (arg0) == CONST_INT)
6207 && ! ((GET_CODE (arg0) == USE
6208 && GET_CODE (arg1) == USE)
6209 || GET_CODE (arg1) == CONST_INT))
6210 tem = arg0, arg0 = arg1, arg1 = tem;
6212 /* Handle addition of zero, then addition of an invariant. */
6213 if (arg1 == const0_rtx)
6215 else if (GET_CODE (arg1) == CONST_INT || GET_CODE (arg1) == USE)
6216 switch (GET_CODE (arg0))
6220 /* Adding two invariants must result in an invariant, so enclose
6221 addition operation inside a USE and return it. */
6222 if (GET_CODE (arg0) == USE)
6223 arg0 = XEXP (arg0, 0);
6224 if (GET_CODE (arg1) == USE)
6225 arg1 = XEXP (arg1, 0);
6227 if (GET_CODE (arg0) == CONST_INT)
6228 tem = arg0, arg0 = arg1, arg1 = tem;
6229 if (GET_CODE (arg1) == CONST_INT)
6230 tem = sge_plus_constant (arg0, arg1);
6232 tem = sge_plus (mode, arg0, arg1);
6234 if (GET_CODE (tem) != CONST_INT)
6235 tem = gen_rtx_USE (mode, tem);
6240 /* biv + invar or mult + invar. Return sum. */
6241 return gen_rtx_PLUS (mode, arg0, arg1);
6244 /* (a + invar_1) + invar_2. Associate. */
6246 simplify_giv_expr (loop,
6258 /* Each argument must be either REG, PLUS, or MULT. Convert REG to
6259 MULT to reduce cases. */
6260 if (GET_CODE (arg0) == REG)
6261 arg0 = gen_rtx_MULT (mode, arg0, const1_rtx);
6262 if (GET_CODE (arg1) == REG)
6263 arg1 = gen_rtx_MULT (mode, arg1, const1_rtx);
6265 /* Now have PLUS + PLUS, PLUS + MULT, MULT + PLUS, or MULT + MULT.
6266 Put a MULT first, leaving PLUS + PLUS, MULT + PLUS, or MULT + MULT.
6267 Recurse to associate the second PLUS. */
6268 if (GET_CODE (arg1) == MULT)
6269 tem = arg0, arg0 = arg1, arg1 = tem;
6271 if (GET_CODE (arg1) == PLUS)
6273 simplify_giv_expr (loop,
6275 gen_rtx_PLUS (mode, arg0,
6280 /* Now must have MULT + MULT. Distribute if same biv, else not giv. */
6281 if (GET_CODE (arg0) != MULT || GET_CODE (arg1) != MULT)
6284 if (!rtx_equal_p (arg0, arg1))
6287 return simplify_giv_expr (loop,
6296 /* Handle "a - b" as "a + b * (-1)". */
6297 return simplify_giv_expr (loop,
6306 arg0 = simplify_giv_expr (loop, XEXP (x, 0), benefit);
6307 arg1 = simplify_giv_expr (loop, XEXP (x, 1), benefit);
6308 if (arg0 == 0 || arg1 == 0)
6311 /* Put constant last, CONST_INT last if both constant. */
6312 if ((GET_CODE (arg0) == USE || GET_CODE (arg0) == CONST_INT)
6313 && GET_CODE (arg1) != CONST_INT)
6314 tem = arg0, arg0 = arg1, arg1 = tem;
6316 /* If second argument is not now constant, not giv. */
6317 if (GET_CODE (arg1) != USE && GET_CODE (arg1) != CONST_INT)
6320 /* Handle multiply by 0 or 1. */
6321 if (arg1 == const0_rtx)
6324 else if (arg1 == const1_rtx)
6327 switch (GET_CODE (arg0))
6330 /* biv * invar. Done. */
6331 return gen_rtx_MULT (mode, arg0, arg1);
6334 /* Product of two constants. */
6335 return GEN_INT (INTVAL (arg0) * INTVAL (arg1));
6338 /* invar * invar is a giv, but attempt to simplify it somehow. */
6339 if (GET_CODE (arg1) != CONST_INT)
6342 arg0 = XEXP (arg0, 0);
6343 if (GET_CODE (arg0) == MULT)
6345 /* (invar_0 * invar_1) * invar_2. Associate. */
6346 return simplify_giv_expr (loop,
6355 /* Porpagate the MULT expressions to the intermost nodes. */
6356 else if (GET_CODE (arg0) == PLUS)
6358 /* (invar_0 + invar_1) * invar_2. Distribute. */
6359 return simplify_giv_expr (loop,
6371 return gen_rtx_USE (mode, gen_rtx_MULT (mode, arg0, arg1));
6374 /* (a * invar_1) * invar_2. Associate. */
6375 return simplify_giv_expr (loop,
6384 /* (a + invar_1) * invar_2. Distribute. */
6385 return simplify_giv_expr (loop,
6400 /* Shift by constant is multiply by power of two. */
6401 if (GET_CODE (XEXP (x, 1)) != CONST_INT)
6405 simplify_giv_expr (loop,
6408 GEN_INT ((HOST_WIDE_INT) 1
6409 << INTVAL (XEXP (x, 1)))),
6413 /* "-a" is "a * (-1)" */
6414 return simplify_giv_expr (loop,
6415 gen_rtx_MULT (mode, XEXP (x, 0), constm1_rtx),
6419 /* "~a" is "-a - 1". Silly, but easy. */
6420 return simplify_giv_expr (loop,
6421 gen_rtx_MINUS (mode,
6422 gen_rtx_NEG (mode, XEXP (x, 0)),
6427 /* Already in proper form for invariant. */
6431 /* If this is a new register, we can't deal with it. */
6432 if (REGNO (x) >= max_reg_before_loop)
6435 /* Check for biv or giv. */
6436 switch (REG_IV_TYPE (REGNO (x)))
6440 case GENERAL_INDUCT:
6442 struct induction *v = REG_IV_INFO (REGNO (x));
6444 /* Form expression from giv and add benefit. Ensure this giv
6445 can derive another and subtract any needed adjustment if so. */
6447 /* Increasing the benefit here is risky. The only case in which it
6448 is arguably correct is if this is the only use of V. In other
6449 cases, this will artificially inflate the benefit of the current
6450 giv, and lead to suboptimal code. Thus, it is disabled, since
6451 potentially not reducing an only marginally beneficial giv is
6452 less harmful than reducing many givs that are not really
6455 rtx single_use = VARRAY_RTX (reg_single_usage, REGNO (x));
6456 if (single_use && single_use != const0_rtx)
6457 *benefit += v->benefit;
6463 tem = gen_rtx_PLUS (mode, gen_rtx_MULT (mode,
6464 v->src_reg, v->mult_val),
6467 if (v->derive_adjustment)
6468 tem = gen_rtx_MINUS (mode, tem, v->derive_adjustment);
6469 return simplify_giv_expr (loop, tem, benefit);
6473 /* If it isn't an induction variable, and it is invariant, we
6474 may be able to simplify things further by looking through
6475 the bits we just moved outside the loop. */
6476 if (loop_invariant_p (loop, x) == 1)
6480 for (m = the_movables; m; m = m->next)
6481 if (rtx_equal_p (x, m->set_dest))
6483 /* Ok, we found a match. Substitute and simplify. */
6485 /* If we match another movable, we must use that, as
6486 this one is going away. */
6488 return simplify_giv_expr (loop, m->match->set_dest,
6491 /* If consec is non-zero, this is a member of a group of
6492 instructions that were moved together. We handle this
6493 case only to the point of seeking to the last insn and
6494 looking for a REG_EQUAL. Fail if we don't find one. */
6499 do { tem = NEXT_INSN (tem); } while (--i > 0);
6501 tem = find_reg_note (tem, REG_EQUAL, NULL_RTX);
6503 tem = XEXP (tem, 0);
6507 tem = single_set (m->insn);
6509 tem = SET_SRC (tem);
6514 /* What we are most interested in is pointer
6515 arithmetic on invariants -- only take
6516 patterns we may be able to do something with. */
6517 if (GET_CODE (tem) == PLUS
6518 || GET_CODE (tem) == MULT
6519 || GET_CODE (tem) == ASHIFT
6520 || GET_CODE (tem) == CONST_INT
6521 || GET_CODE (tem) == SYMBOL_REF)
6523 tem = simplify_giv_expr (loop, tem, benefit);
6527 else if (GET_CODE (tem) == CONST
6528 && GET_CODE (XEXP (tem, 0)) == PLUS
6529 && GET_CODE (XEXP (XEXP (tem, 0), 0)) == SYMBOL_REF
6530 && GET_CODE (XEXP (XEXP (tem, 0), 1)) == CONST_INT)
6532 tem = simplify_giv_expr (loop, XEXP (tem, 0),
6544 /* Fall through to general case. */
6546 /* If invariant, return as USE (unless CONST_INT).
6547 Otherwise, not giv. */
6548 if (GET_CODE (x) == USE)
6551 if (loop_invariant_p (loop, x) == 1)
6553 if (GET_CODE (x) == CONST_INT)
6555 if (GET_CODE (x) == CONST
6556 && GET_CODE (XEXP (x, 0)) == PLUS
6557 && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF
6558 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT)
6560 return gen_rtx_USE (mode, x);
6567 /* This routine folds invariants such that there is only ever one
6568 CONST_INT in the summation. It is only used by simplify_giv_expr. */
6571 sge_plus_constant (x, c)
6574 if (GET_CODE (x) == CONST_INT)
6575 return GEN_INT (INTVAL (x) + INTVAL (c));
6576 else if (GET_CODE (x) != PLUS)
6577 return gen_rtx_PLUS (GET_MODE (x), x, c);
6578 else if (GET_CODE (XEXP (x, 1)) == CONST_INT)
6580 return gen_rtx_PLUS (GET_MODE (x), XEXP (x, 0),
6581 GEN_INT (INTVAL (XEXP (x, 1)) + INTVAL (c)));
6583 else if (GET_CODE (XEXP (x, 0)) == PLUS
6584 || GET_CODE (XEXP (x, 1)) != PLUS)
6586 return gen_rtx_PLUS (GET_MODE (x),
6587 sge_plus_constant (XEXP (x, 0), c), XEXP (x, 1));
6591 return gen_rtx_PLUS (GET_MODE (x),
6592 sge_plus_constant (XEXP (x, 1), c), XEXP (x, 0));
6597 sge_plus (mode, x, y)
6598 enum machine_mode mode;
6601 while (GET_CODE (y) == PLUS)
6603 rtx a = XEXP (y, 0);
6604 if (GET_CODE (a) == CONST_INT)
6605 x = sge_plus_constant (x, a);
6607 x = gen_rtx_PLUS (mode, x, a);
6610 if (GET_CODE (y) == CONST_INT)
6611 x = sge_plus_constant (x, y);
6613 x = gen_rtx_PLUS (mode, x, y);
6617 /* Help detect a giv that is calculated by several consecutive insns;
6621 The caller has already identified the first insn P as having a giv as dest;
6622 we check that all other insns that set the same register follow
6623 immediately after P, that they alter nothing else,
6624 and that the result of the last is still a giv.
6626 The value is 0 if the reg set in P is not really a giv.
6627 Otherwise, the value is the amount gained by eliminating
6628 all the consecutive insns that compute the value.
6630 FIRST_BENEFIT is the amount gained by eliminating the first insn, P.
6631 SRC_REG is the reg of the biv; DEST_REG is the reg of the giv.
6633 The coefficients of the ultimate giv value are stored in
6634 *MULT_VAL and *ADD_VAL. */
6637 consec_sets_giv (loop, first_benefit, p, src_reg, dest_reg,
6638 add_val, mult_val, last_consec_insn)
6639 const struct loop *loop;
6646 rtx *last_consec_insn;
6654 /* Indicate that this is a giv so that we can update the value produced in
6655 each insn of the multi-insn sequence.
6657 This induction structure will be used only by the call to
6658 general_induction_var below, so we can allocate it on our stack.
6659 If this is a giv, our caller will replace the induct var entry with
6660 a new induction structure. */
6662 = (struct induction *) alloca (sizeof (struct induction));
6663 v->src_reg = src_reg;
6664 v->mult_val = *mult_val;
6665 v->add_val = *add_val;
6666 v->benefit = first_benefit;
6668 v->derive_adjustment = 0;
6670 REG_IV_TYPE (REGNO (dest_reg)) = GENERAL_INDUCT;
6671 REG_IV_INFO (REGNO (dest_reg)) = v;
6673 count = VARRAY_INT (n_times_set, REGNO (dest_reg)) - 1;
6678 code = GET_CODE (p);
6680 /* If libcall, skip to end of call sequence. */
6681 if (code == INSN && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
6685 && (set = single_set (p))
6686 && GET_CODE (SET_DEST (set)) == REG
6687 && SET_DEST (set) == dest_reg
6688 && (general_induction_var (loop, SET_SRC (set), &src_reg,
6689 add_val, mult_val, 0, &benefit, VOIDmode)
6690 /* Giv created by equivalent expression. */
6691 || ((temp = find_reg_note (p, REG_EQUAL, NULL_RTX))
6692 && general_induction_var (loop, XEXP (temp, 0), &src_reg,
6693 add_val, mult_val, 0, &benefit,
6695 && src_reg == v->src_reg)
6697 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
6698 benefit += libcall_benefit (p);
6701 v->mult_val = *mult_val;
6702 v->add_val = *add_val;
6703 v->benefit += benefit;
6705 else if (code != NOTE)
6707 /* Allow insns that set something other than this giv to a
6708 constant. Such insns are needed on machines which cannot
6709 include long constants and should not disqualify a giv. */
6711 && (set = single_set (p))
6712 && SET_DEST (set) != dest_reg
6713 && CONSTANT_P (SET_SRC (set)))
6716 REG_IV_TYPE (REGNO (dest_reg)) = UNKNOWN_INDUCT;
6721 *last_consec_insn = p;
6725 /* Return an rtx, if any, that expresses giv G2 as a function of the register
6726 represented by G1. If no such expression can be found, or it is clear that
6727 it cannot possibly be a valid address, 0 is returned.
6729 To perform the computation, we note that
6732 where `v' is the biv.
6734 So G2 = (y/b) * G1 + (b - a*y/x).
6736 Note that MULT = y/x.
6738 Update: A and B are now allowed to be additive expressions such that
6739 B contains all variables in A. That is, computing B-A will not require
6740 subtracting variables. */
6743 express_from_1 (a, b, mult)
6746 /* If MULT is zero, then A*MULT is zero, and our expression is B. */
6748 if (mult == const0_rtx)
6751 /* If MULT is not 1, we cannot handle A with non-constants, since we
6752 would then be required to subtract multiples of the registers in A.
6753 This is theoretically possible, and may even apply to some Fortran
6754 constructs, but it is a lot of work and we do not attempt it here. */
6756 if (mult != const1_rtx && GET_CODE (a) != CONST_INT)
6759 /* In general these structures are sorted top to bottom (down the PLUS
6760 chain), but not left to right across the PLUS. If B is a higher
6761 order giv than A, we can strip one level and recurse. If A is higher
6762 order, we'll eventually bail out, but won't know that until the end.
6763 If they are the same, we'll strip one level around this loop. */
6765 while (GET_CODE (a) == PLUS && GET_CODE (b) == PLUS)
6767 rtx ra, rb, oa, ob, tmp;
6769 ra = XEXP (a, 0), oa = XEXP (a, 1);
6770 if (GET_CODE (ra) == PLUS)
6771 tmp = ra, ra = oa, oa = tmp;
6773 rb = XEXP (b, 0), ob = XEXP (b, 1);
6774 if (GET_CODE (rb) == PLUS)
6775 tmp = rb, rb = ob, ob = tmp;
6777 if (rtx_equal_p (ra, rb))
6778 /* We matched: remove one reg completely. */
6780 else if (GET_CODE (ob) != PLUS && rtx_equal_p (ra, ob))
6781 /* An alternate match. */
6783 else if (GET_CODE (oa) != PLUS && rtx_equal_p (oa, rb))
6784 /* An alternate match. */
6788 /* Indicates an extra register in B. Strip one level from B and
6789 recurse, hoping B was the higher order expression. */
6790 ob = express_from_1 (a, ob, mult);
6793 return gen_rtx_PLUS (GET_MODE (b), rb, ob);
6797 /* Here we are at the last level of A, go through the cases hoping to
6798 get rid of everything but a constant. */
6800 if (GET_CODE (a) == PLUS)
6804 ra = XEXP (a, 0), oa = XEXP (a, 1);
6805 if (rtx_equal_p (oa, b))
6807 else if (!rtx_equal_p (ra, b))
6810 if (GET_CODE (oa) != CONST_INT)
6813 return GEN_INT (-INTVAL (oa) * INTVAL (mult));
6815 else if (GET_CODE (a) == CONST_INT)
6817 return plus_constant (b, -INTVAL (a) * INTVAL (mult));
6819 else if (CONSTANT_P (a))
6821 return simplify_gen_binary (MINUS, GET_MODE (b) != VOIDmode ? GET_MODE (b) : GET_MODE (a), const0_rtx, a);
6823 else if (GET_CODE (b) == PLUS)
6825 if (rtx_equal_p (a, XEXP (b, 0)))
6827 else if (rtx_equal_p (a, XEXP (b, 1)))
6832 else if (rtx_equal_p (a, b))
6839 express_from (g1, g2)
6840 struct induction *g1, *g2;
6844 /* The value that G1 will be multiplied by must be a constant integer. Also,
6845 the only chance we have of getting a valid address is if b*c/a (see above
6846 for notation) is also an integer. */
6847 if (GET_CODE (g1->mult_val) == CONST_INT
6848 && GET_CODE (g2->mult_val) == CONST_INT)
6850 if (g1->mult_val == const0_rtx
6851 || INTVAL (g2->mult_val) % INTVAL (g1->mult_val) != 0)
6853 mult = GEN_INT (INTVAL (g2->mult_val) / INTVAL (g1->mult_val));
6855 else if (rtx_equal_p (g1->mult_val, g2->mult_val))
6859 /* ??? Find out if the one is a multiple of the other? */
6863 add = express_from_1 (g1->add_val, g2->add_val, mult);
6864 if (add == NULL_RTX)
6866 /* Failed. If we've got a multiplication factor between G1 and G2,
6867 scale G1's addend and try again. */
6868 if (INTVAL (mult) > 1)
6870 rtx g1_add_val = g1->add_val;
6871 if (GET_CODE (g1_add_val) == MULT
6872 && GET_CODE (XEXP (g1_add_val, 1)) == CONST_INT)
6875 m = INTVAL (mult) * INTVAL (XEXP (g1_add_val, 1));
6876 g1_add_val = gen_rtx_MULT (GET_MODE (g1_add_val),
6877 XEXP (g1_add_val, 0), GEN_INT (m));
6881 g1_add_val = gen_rtx_MULT (GET_MODE (g1_add_val), g1_add_val,
6885 add = express_from_1 (g1_add_val, g2->add_val, const1_rtx);
6888 if (add == NULL_RTX)
6891 /* Form simplified final result. */
6892 if (mult == const0_rtx)
6894 else if (mult == const1_rtx)
6895 mult = g1->dest_reg;
6897 mult = gen_rtx_MULT (g2->mode, g1->dest_reg, mult);
6899 if (add == const0_rtx)
6903 if (GET_CODE (add) == PLUS
6904 && CONSTANT_P (XEXP (add, 1)))
6906 rtx tem = XEXP (add, 1);
6907 mult = gen_rtx_PLUS (g2->mode, mult, XEXP (add, 0));
6911 return gen_rtx_PLUS (g2->mode, mult, add);
6915 /* Return an rtx, if any, that expresses giv G2 as a function of the register
6916 represented by G1. This indicates that G2 should be combined with G1 and
6917 that G2 can use (either directly or via an address expression) a register
6918 used to represent G1. */
6921 combine_givs_p (g1, g2)
6922 struct induction *g1, *g2;
6924 rtx tem = express_from (g1, g2);
6926 /* If these givs are identical, they can be combined. We use the results
6927 of express_from because the addends are not in a canonical form, so
6928 rtx_equal_p is a weaker test. */
6929 /* But don't combine a DEST_REG giv with a DEST_ADDR giv; we want the
6930 combination to be the other way round. */
6931 if (tem == g1->dest_reg
6932 && (g1->giv_type == DEST_REG || g2->giv_type == DEST_ADDR))
6934 return g1->dest_reg;
6937 /* If G2 can be expressed as a function of G1 and that function is valid
6938 as an address and no more expensive than using a register for G2,
6939 the expression of G2 in terms of G1 can be used. */
6941 && g2->giv_type == DEST_ADDR
6942 && memory_address_p (g2->mem_mode, tem)
6943 /* ??? Looses, especially with -fforce-addr, where *g2->location
6944 will always be a register, and so anything more complicated
6948 && ADDRESS_COST (tem) <= ADDRESS_COST (*g2->location)
6950 && rtx_cost (tem, MEM) <= rtx_cost (*g2->location, MEM)
6961 struct combine_givs_stats
6968 cmp_combine_givs_stats (xp, yp)
6972 const struct combine_givs_stats * const x =
6973 (const struct combine_givs_stats *) xp;
6974 const struct combine_givs_stats * const y =
6975 (const struct combine_givs_stats *) yp;
6977 d = y->total_benefit - x->total_benefit;
6978 /* Stabilize the sort. */
6980 d = x->giv_number - y->giv_number;
6984 /* Check all pairs of givs for iv_class BL and see if any can be combined with
6985 any other. If so, point SAME to the giv combined with and set NEW_REG to
6986 be an expression (in terms of the other giv's DEST_REG) equivalent to the
6987 giv. Also, update BENEFIT and related fields for cost/benefit analysis. */
6991 struct iv_class *bl;
6993 /* Additional benefit to add for being combined multiple times. */
6994 const int extra_benefit = 3;
6996 struct induction *g1, *g2, **giv_array;
6997 int i, j, k, giv_count;
6998 struct combine_givs_stats *stats;
7001 /* Count givs, because bl->giv_count is incorrect here. */
7003 for (g1 = bl->giv; g1; g1 = g1->next_iv)
7008 = (struct induction **) alloca (giv_count * sizeof (struct induction *));
7010 for (g1 = bl->giv; g1; g1 = g1->next_iv)
7012 giv_array[i++] = g1;
7014 stats = (struct combine_givs_stats *) xcalloc (giv_count, sizeof (*stats));
7015 can_combine = (rtx *) xcalloc (giv_count, giv_count * sizeof (rtx));
7017 for (i = 0; i < giv_count; i++)
7023 stats[i].giv_number = i;
7025 /* If a DEST_REG GIV is used only once, do not allow it to combine
7026 with anything, for in doing so we will gain nothing that cannot
7027 be had by simply letting the GIV with which we would have combined
7028 to be reduced on its own. The losage shows up in particular with
7029 DEST_ADDR targets on hosts with reg+reg addressing, though it can
7030 be seen elsewhere as well. */
7031 if (g1->giv_type == DEST_REG
7032 && (single_use = VARRAY_RTX (reg_single_usage, REGNO (g1->dest_reg)))
7033 && single_use != const0_rtx)
7036 this_benefit = g1->benefit;
7037 /* Add an additional weight for zero addends. */
7038 if (g1->no_const_addval)
7041 for (j = 0; j < giv_count; j++)
7047 && (this_combine = combine_givs_p (g1, g2)) != NULL_RTX)
7049 can_combine[i * giv_count + j] = this_combine;
7050 this_benefit += g2->benefit + extra_benefit;
7053 stats[i].total_benefit = this_benefit;
7056 /* Iterate, combining until we can't. */
7058 qsort (stats, giv_count, sizeof (*stats), cmp_combine_givs_stats);
7060 if (loop_dump_stream)
7062 fprintf (loop_dump_stream, "Sorted combine statistics:\n");
7063 for (k = 0; k < giv_count; k++)
7065 g1 = giv_array[stats[k].giv_number];
7066 if (!g1->combined_with && !g1->same)
7067 fprintf (loop_dump_stream, " {%d, %d}",
7068 INSN_UID (giv_array[stats[k].giv_number]->insn),
7069 stats[k].total_benefit);
7071 putc ('\n', loop_dump_stream);
7074 for (k = 0; k < giv_count; k++)
7076 int g1_add_benefit = 0;
7078 i = stats[k].giv_number;
7081 /* If it has already been combined, skip. */
7082 if (g1->combined_with || g1->same)
7085 for (j = 0; j < giv_count; j++)
7088 if (g1 != g2 && can_combine[i * giv_count + j]
7089 /* If it has already been combined, skip. */
7090 && ! g2->same && ! g2->combined_with)
7094 g2->new_reg = can_combine[i * giv_count + j];
7096 g1->combined_with++;
7097 g1->lifetime += g2->lifetime;
7099 g1_add_benefit += g2->benefit;
7101 /* ??? The new final_[bg]iv_value code does a much better job
7102 of finding replaceable giv's, and hence this code may no
7103 longer be necessary. */
7104 if (! g2->replaceable && REG_USERVAR_P (g2->dest_reg))
7105 g1_add_benefit -= copy_cost;
7107 /* To help optimize the next set of combinations, remove
7108 this giv from the benefits of other potential mates. */
7109 for (l = 0; l < giv_count; ++l)
7111 int m = stats[l].giv_number;
7112 if (can_combine[m * giv_count + j])
7113 stats[l].total_benefit -= g2->benefit + extra_benefit;
7116 if (loop_dump_stream)
7117 fprintf (loop_dump_stream,
7118 "giv at %d combined with giv at %d; new benefit %d + %d, lifetime %d\n",
7119 INSN_UID (g2->insn), INSN_UID (g1->insn),
7120 g1->benefit, g1_add_benefit, g1->lifetime);
7124 /* To help optimize the next set of combinations, remove
7125 this giv from the benefits of other potential mates. */
7126 if (g1->combined_with)
7128 for (j = 0; j < giv_count; ++j)
7130 int m = stats[j].giv_number;
7131 if (can_combine[m * giv_count + i])
7132 stats[j].total_benefit -= g1->benefit + extra_benefit;
7135 g1->benefit += g1_add_benefit;
7137 /* We've finished with this giv, and everything it touched.
7138 Restart the combination so that proper weights for the
7139 rest of the givs are properly taken into account. */
7140 /* ??? Ideally we would compact the arrays at this point, so
7141 as to not cover old ground. But sanely compacting
7142 can_combine is tricky. */
7152 struct recombine_givs_stats
7155 int start_luid, end_luid;
7158 /* Used below as comparison function for qsort. We want a ascending luid
7159 when scanning the array starting at the end, thus the arguments are
7162 cmp_recombine_givs_stats (xp, yp)
7166 const struct recombine_givs_stats * const x =
7167 (const struct recombine_givs_stats *) xp;
7168 const struct recombine_givs_stats * const y =
7169 (const struct recombine_givs_stats *) yp;
7171 d = y->start_luid - x->start_luid;
7172 /* Stabilize the sort. */
7174 d = y->giv_number - x->giv_number;
7178 /* Scan X, which is a part of INSN, for the end of life of a giv. Also
7179 look for the start of life of a giv where the start has not been seen
7180 yet to unlock the search for the end of its life.
7181 Only consider givs that belong to BIV.
7182 Return the total number of lifetime ends that have been found. */
7184 find_life_end (x, stats, insn, biv)
7186 struct recombine_givs_stats *stats;
7193 code = GET_CODE (x);
7198 rtx reg = SET_DEST (x);
7199 if (GET_CODE (reg) == REG)
7201 int regno = REGNO (reg);
7202 struct induction *v = REG_IV_INFO (regno);
7204 if (REG_IV_TYPE (regno) == GENERAL_INDUCT
7206 && v->src_reg == biv
7207 && stats[v->ix].end_luid <= 0)
7209 /* If we see a 0 here for end_luid, it means that we have
7210 scanned the entire loop without finding any use at all.
7211 We must not predicate this code on a start_luid match
7212 since that would make the test fail for givs that have
7213 been hoisted out of inner loops. */
7214 if (stats[v->ix].end_luid == 0)
7216 stats[v->ix].end_luid = stats[v->ix].start_luid;
7217 return 1 + find_life_end (SET_SRC (x), stats, insn, biv);
7219 else if (stats[v->ix].start_luid == INSN_LUID (insn))
7220 stats[v->ix].end_luid = 0;
7222 return find_life_end (SET_SRC (x), stats, insn, biv);
7228 int regno = REGNO (x);
7229 struct induction *v = REG_IV_INFO (regno);
7231 if (REG_IV_TYPE (regno) == GENERAL_INDUCT
7233 && v->src_reg == biv
7234 && stats[v->ix].end_luid == 0)
7236 while (INSN_UID (insn) >= max_uid_for_loop)
7237 insn = NEXT_INSN (insn);
7238 stats[v->ix].end_luid = INSN_LUID (insn);
7251 fmt = GET_RTX_FORMAT (code);
7253 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
7256 retval += find_life_end (XEXP (x, i), stats, insn, biv);
7258 else if (fmt[i] == 'E')
7259 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
7260 retval += find_life_end (XVECEXP (x, i, j), stats, insn, biv);
7265 /* For each giv that has been combined with another, look if
7266 we can combine it with the most recently used one instead.
7267 This tends to shorten giv lifetimes, and helps the next step:
7268 try to derive givs from other givs. */
7270 recombine_givs (loop, bl, unroll_p)
7271 const struct loop *loop;
7272 struct iv_class *bl;
7275 struct induction *v, **giv_array, *last_giv;
7276 struct recombine_givs_stats *stats;
7279 int ends_need_computing;
7281 for (giv_count = 0, v = bl->giv; v; v = v->next_iv)
7287 = (struct induction **) xmalloc (giv_count * sizeof (struct induction *));
7288 stats = (struct recombine_givs_stats *) xmalloc (giv_count * sizeof *stats);
7290 /* Initialize stats and set up the ix field for each giv in stats to name
7291 the corresponding index into stats. */
7292 for (i = 0, v = bl->giv; v; v = v->next_iv)
7299 stats[i].giv_number = i;
7300 /* If this giv has been hoisted out of an inner loop, use the luid of
7301 the previous insn. */
7302 for (p = v->insn; INSN_UID (p) >= max_uid_for_loop; )
7304 stats[i].start_luid = INSN_LUID (p);
7308 qsort (stats, giv_count, sizeof (*stats), cmp_recombine_givs_stats);
7310 /* Set up the ix field for each giv in stats to name
7311 the corresponding index into stats, and
7312 do the actual most-recently-used recombination. */
7313 for (last_giv = 0, i = giv_count - 1; i >= 0; i--)
7315 v = giv_array[stats[i].giv_number];
7319 struct induction *old_same = v->same;
7322 /* combine_givs_p actually says if we can make this transformation.
7323 The other tests are here only to avoid keeping a giv alive
7324 that could otherwise be eliminated. */
7326 && ((old_same->maybe_dead && ! old_same->combined_with)
7327 || ! last_giv->maybe_dead
7328 || last_giv->combined_with)
7329 && (new_combine = combine_givs_p (last_giv, v)))
7331 old_same->combined_with--;
7332 v->new_reg = new_combine;
7334 last_giv->combined_with++;
7335 /* No need to update lifetimes / benefits here since we have
7336 already decided what to reduce. */
7338 if (loop_dump_stream)
7340 fprintf (loop_dump_stream,
7341 "giv at %d recombined with giv at %d as ",
7342 INSN_UID (v->insn), INSN_UID (last_giv->insn));
7343 print_rtl (loop_dump_stream, v->new_reg);
7344 putc ('\n', loop_dump_stream);
7350 else if (v->giv_type != DEST_REG)
7353 || (last_giv->maybe_dead && ! last_giv->combined_with)
7355 || v->combined_with)
7359 ends_need_computing = 0;
7360 /* For each DEST_REG giv, compute lifetime starts, and try to compute
7361 lifetime ends from regscan info. */
7362 for (i = giv_count - 1; i >= 0; i--)
7364 v = giv_array[stats[i].giv_number];
7367 if (v->giv_type == DEST_ADDR)
7369 /* Loop unrolling of an inner loop can even create new DEST_REG
7372 for (p = v->insn; INSN_UID (p) >= max_uid_for_loop;)
7374 stats[i].start_luid = stats[i].end_luid = INSN_LUID (p);
7376 stats[i].end_luid++;
7378 else /* v->giv_type == DEST_REG */
7382 stats[i].start_luid = INSN_LUID (v->insn);
7383 stats[i].end_luid = INSN_LUID (v->last_use);
7385 else if (INSN_UID (v->insn) >= max_uid_for_loop)
7388 /* This insn has been created by loop optimization on an inner
7389 loop. We don't have a proper start_luid that will match
7390 when we see the first set. But we do know that there will
7391 be no use before the set, so we can set end_luid to 0 so that
7392 we'll start looking for the last use right away. */
7393 for (p = PREV_INSN (v->insn); INSN_UID (p) >= max_uid_for_loop; )
7395 stats[i].start_luid = INSN_LUID (p);
7396 stats[i].end_luid = 0;
7397 ends_need_computing++;
7401 int regno = REGNO (v->dest_reg);
7402 int count = VARRAY_INT (n_times_set, regno) - 1;
7405 /* Find the first insn that sets the giv, so that we can verify
7406 if this giv's lifetime wraps around the loop. We also need
7407 the luid of the first setting insn in order to detect the
7408 last use properly. */
7411 p = prev_nonnote_insn (p);
7412 if (reg_set_p (v->dest_reg, p))
7416 stats[i].start_luid = INSN_LUID (p);
7417 if (stats[i].start_luid > uid_luid[REGNO_FIRST_UID (regno)])
7419 stats[i].end_luid = -1;
7420 ends_need_computing++;
7424 stats[i].end_luid = uid_luid[REGNO_LAST_UID (regno)];
7425 if (stats[i].end_luid > INSN_LUID (loop->end))
7427 stats[i].end_luid = -1;
7428 ends_need_computing++;
7435 /* If the regscan information was unconclusive for one or more DEST_REG
7436 givs, scan the all insn in the loop to find out lifetime ends. */
7437 if (ends_need_computing)
7439 rtx biv = bl->biv->src_reg;
7444 if (p == loop->start)
7449 ends_need_computing -= find_life_end (PATTERN (p), stats, p, biv);
7451 while (ends_need_computing);
7454 /* Set start_luid back to the last insn that sets the giv. This allows
7455 more combinations. */
7456 for (i = giv_count - 1; i >= 0; i--)
7458 v = giv_array[stats[i].giv_number];
7461 if (INSN_UID (v->insn) < max_uid_for_loop)
7462 stats[i].start_luid = INSN_LUID (v->insn);
7465 /* Now adjust lifetime ends by taking combined givs into account. */
7466 for (i = giv_count - 1; i >= 0; i--)
7471 v = giv_array[stats[i].giv_number];
7474 if (v->same && ! v->same->ignore)
7477 luid = stats[i].start_luid;
7478 /* Use unsigned arithmetic to model loop wrap-around. */
7479 if (luid - stats[j].start_luid
7480 > (unsigned) stats[j].end_luid - stats[j].start_luid)
7481 stats[j].end_luid = luid;
7485 qsort (stats, giv_count, sizeof (*stats), cmp_recombine_givs_stats);
7487 /* Try to derive DEST_REG givs from previous DEST_REG givs with the
7488 same mult_val and non-overlapping lifetime. This reduces register
7490 Once we find a DEST_REG giv that is suitable to derive others from,
7491 we set last_giv to this giv, and try to derive as many other DEST_REG
7492 givs from it without joining overlapping lifetimes. If we then
7493 encounter a DEST_REG giv that we can't derive, we set rescan to the
7494 index for this giv (unless rescan is already set).
7495 When we are finished with the current LAST_GIV (i.e. the inner loop
7496 terminates), we start again with rescan, which then becomes the new
7498 for (i = giv_count - 1; i >= 0; i = rescan)
7500 int life_start = 0, life_end = 0;
7502 for (last_giv = 0, rescan = -1; i >= 0; i--)
7506 v = giv_array[stats[i].giv_number];
7507 if (v->giv_type != DEST_REG || v->derived_from || v->same)
7511 /* Don't use a giv that's likely to be dead to derive
7512 others - that would be likely to keep that giv alive. */
7513 if (! v->maybe_dead || v->combined_with)
7516 life_start = stats[i].start_luid;
7517 life_end = stats[i].end_luid;
7521 /* Use unsigned arithmetic to model loop wrap around. */
7522 if (((unsigned) stats[i].start_luid - life_start
7523 >= (unsigned) life_end - life_start)
7524 && ((unsigned) stats[i].end_luid - life_start
7525 > (unsigned) life_end - life_start)
7526 /* Check that the giv insn we're about to use for deriving
7527 precedes all uses of that giv. Note that initializing the
7528 derived giv would defeat the purpose of reducing register
7530 ??? We could arrange to move the insn. */
7531 && ((unsigned) stats[i].end_luid - INSN_LUID (loop->start)
7532 > (unsigned) stats[i].start_luid - INSN_LUID (loop->start))
7533 && rtx_equal_p (last_giv->mult_val, v->mult_val)
7534 /* ??? Could handle libcalls, but would need more logic. */
7535 && ! find_reg_note (v->insn, REG_RETVAL, NULL_RTX)
7536 /* We would really like to know if for any giv that v
7537 is combined with, v->insn or any intervening biv increment
7538 dominates that combined giv. However, we
7539 don't have this detailed control flow information.
7540 N.B. since last_giv will be reduced, it is valid
7541 anywhere in the loop, so we don't need to check the
7542 validity of last_giv.
7543 We rely here on the fact that v->always_executed implies that
7544 there is no jump to someplace else in the loop before the
7545 giv insn, and hence any insn that is executed before the
7546 giv insn in the loop will have a lower luid. */
7547 && (v->always_executed || ! v->combined_with)
7548 && (sum = express_from (last_giv, v))
7549 /* Make sure we don't make the add more expensive. ADD_COST
7550 doesn't take different costs of registers and constants into
7551 account, so compare the cost of the actual SET_SRCs. */
7552 && (rtx_cost (sum, SET)
7553 <= rtx_cost (SET_SRC (single_set (v->insn)), SET))
7554 /* ??? unroll can't understand anything but reg + const_int
7555 sums. It would be cleaner to fix unroll. */
7556 && ((GET_CODE (sum) == PLUS
7557 && GET_CODE (XEXP (sum, 0)) == REG
7558 && GET_CODE (XEXP (sum, 1)) == CONST_INT)
7560 && validate_change (v->insn, &PATTERN (v->insn),
7561 gen_rtx_SET (VOIDmode, v->dest_reg, sum), 0))
7563 v->derived_from = last_giv;
7564 life_end = stats[i].end_luid;
7566 if (loop_dump_stream)
7568 fprintf (loop_dump_stream,
7569 "giv at %d derived from %d as ",
7570 INSN_UID (v->insn), INSN_UID (last_giv->insn));
7571 print_rtl (loop_dump_stream, sum);
7572 putc ('\n', loop_dump_stream);
7575 else if (rescan < 0)
7585 /* EMIT code before INSERT_BEFORE to set REG = B * M + A. */
7588 emit_iv_add_mult (b, m, a, reg, insert_before)
7589 rtx b; /* initial value of basic induction variable */
7590 rtx m; /* multiplicative constant */
7591 rtx a; /* additive constant */
7592 rtx reg; /* destination register */
7598 /* Prevent unexpected sharing of these rtx. */
7602 /* Increase the lifetime of any invariants moved further in code. */
7603 update_reg_last_use (a, insert_before);
7604 update_reg_last_use (b, insert_before);
7605 update_reg_last_use (m, insert_before);
7608 result = expand_mult_add (b, reg, m, a, GET_MODE (reg), 0);
7610 emit_move_insn (reg, result);
7611 seq = gen_sequence ();
7614 emit_insn_before (seq, insert_before);
7616 /* It is entirely possible that the expansion created lots of new
7617 registers. Iterate over the sequence we just created and
7620 if (GET_CODE (seq) == SEQUENCE)
7623 for (i = 0; i < XVECLEN (seq, 0); ++i)
7625 rtx set = single_set (XVECEXP (seq, 0, i));
7626 if (set && GET_CODE (SET_DEST (set)) == REG)
7627 record_base_value (REGNO (SET_DEST (set)), SET_SRC (set), 0);
7630 else if (GET_CODE (seq) == SET
7631 && GET_CODE (SET_DEST (seq)) == REG)
7632 record_base_value (REGNO (SET_DEST (seq)), SET_SRC (seq), 0);
7635 /* Similar to emit_iv_add_mult, but compute cost rather than emitting
7638 iv_add_mult_cost (b, m, a, reg)
7639 rtx b; /* initial value of basic induction variable */
7640 rtx m; /* multiplicative constant */
7641 rtx a; /* additive constant */
7642 rtx reg; /* destination register */
7648 result = expand_mult_add (b, reg, m, a, GET_MODE (reg), 0);
7650 emit_move_insn (reg, result);
7651 last = get_last_insn ();
7654 rtx t = single_set (last);
7656 cost += rtx_cost (SET_SRC (t), SET);
7657 last = PREV_INSN (last);
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. */
7950 else if (num_mem_sets <= 1)
7952 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
7954 num_nonfixed_reads += count_nonfixed_reads (loop, PATTERN (p));
7956 /* If the loop has a single store, and the destination address is
7957 invariant, then we can't reverse the loop, because this address
7958 might then have the wrong value at loop exit.
7959 This would work if the source was invariant also, however, in that
7960 case, the insn should have been moved out of the loop. */
7962 if (num_mem_sets == 1)
7964 struct induction *v;
7966 reversible_mem_store
7967 = (! unknown_address_altered
7968 && ! unknown_constant_address_altered
7969 && ! loop_invariant_p (loop,
7970 XEXP (XEXP (loop_store_mems, 0),
7973 /* If the store depends on a register that is set after the
7974 store, it depends on the initial value, and is thus not
7976 for (v = bl->giv; reversible_mem_store && v; v = v->next_iv)
7978 if (v->giv_type == DEST_REG
7979 && reg_mentioned_p (v->dest_reg,
7980 PATTERN (first_loop_store_insn))
7981 && loop_insn_first_p (first_loop_store_insn, v->insn))
7982 reversible_mem_store = 0;
7989 /* This code only acts for innermost loops. Also it simplifies
7990 the memory address check by only reversing loops with
7991 zero or one memory access.
7992 Two memory accesses could involve parts of the same array,
7993 and that can't be reversed.
7994 If the biv is used only for counting, than we don't need to worry
7995 about all these things. */
7997 if ((num_nonfixed_reads <= 1
7998 && ! loop_info->has_call
7999 && ! loop_info->has_volatile
8000 && reversible_mem_store
8001 && (bl->giv_count + bl->biv_count + num_mem_sets
8002 + num_movables + compare_and_branch == insn_count)
8003 && (bl == loop_iv_list && bl->next == 0))
8004 || no_use_except_counting)
8008 /* Loop can be reversed. */
8009 if (loop_dump_stream)
8010 fprintf (loop_dump_stream, "Can reverse loop\n");
8012 /* Now check other conditions:
8014 The increment must be a constant, as must the initial value,
8015 and the comparison code must be LT.
8017 This test can probably be improved since +/- 1 in the constant
8018 can be obtained by changing LT to LE and vice versa; this is
8022 /* for constants, LE gets turned into LT */
8023 && (GET_CODE (comparison) == LT
8024 || (GET_CODE (comparison) == LE
8025 && no_use_except_counting)))
8027 HOST_WIDE_INT add_val, add_adjust, comparison_val = 0;
8028 rtx initial_value, comparison_value;
8030 enum rtx_code cmp_code;
8031 int comparison_const_width;
8032 unsigned HOST_WIDE_INT comparison_sign_mask;
8034 add_val = INTVAL (bl->biv->add_val);
8035 comparison_value = XEXP (comparison, 1);
8036 if (GET_MODE (comparison_value) == VOIDmode)
8037 comparison_const_width
8038 = GET_MODE_BITSIZE (GET_MODE (XEXP (comparison, 0)));
8040 comparison_const_width
8041 = GET_MODE_BITSIZE (GET_MODE (comparison_value));
8042 if (comparison_const_width > HOST_BITS_PER_WIDE_INT)
8043 comparison_const_width = HOST_BITS_PER_WIDE_INT;
8044 comparison_sign_mask
8045 = (unsigned HOST_WIDE_INT) 1 << (comparison_const_width - 1);
8047 /* If the comparison value is not a loop invariant, then we
8048 can not reverse this loop.
8050 ??? If the insns which initialize the comparison value as
8051 a whole compute an invariant result, then we could move
8052 them out of the loop and proceed with loop reversal. */
8053 if (! loop_invariant_p (loop, comparison_value))
8056 if (GET_CODE (comparison_value) == CONST_INT)
8057 comparison_val = INTVAL (comparison_value);
8058 initial_value = bl->initial_value;
8060 /* Normalize the initial value if it is an integer and
8061 has no other use except as a counter. This will allow
8062 a few more loops to be reversed. */
8063 if (no_use_except_counting
8064 && GET_CODE (comparison_value) == CONST_INT
8065 && GET_CODE (initial_value) == CONST_INT)
8067 comparison_val = comparison_val - INTVAL (bl->initial_value);
8068 /* The code below requires comparison_val to be a multiple
8069 of add_val in order to do the loop reversal, so
8070 round up comparison_val to a multiple of add_val.
8071 Since comparison_value is constant, we know that the
8072 current comparison code is LT. */
8073 comparison_val = comparison_val + add_val - 1;
8075 -= (unsigned HOST_WIDE_INT) comparison_val % add_val;
8076 /* We postpone overflow checks for COMPARISON_VAL here;
8077 even if there is an overflow, we might still be able to
8078 reverse the loop, if converting the loop exit test to
8080 initial_value = const0_rtx;
8083 /* First check if we can do a vanilla loop reversal. */
8084 if (initial_value == const0_rtx
8085 /* If we have a decrement_and_branch_on_count,
8086 prefer the NE test, since this will allow that
8087 instruction to be generated. Note that we must
8088 use a vanilla loop reversal if the biv is used to
8089 calculate a giv or has a non-counting use. */
8090 #if ! defined (HAVE_decrement_and_branch_until_zero) \
8091 && defined (HAVE_decrement_and_branch_on_count)
8092 && (! (add_val == 1 && loop->vtop
8093 && (bl->biv_count == 0
8094 || no_use_except_counting)))
8096 && GET_CODE (comparison_value) == CONST_INT
8097 /* Now do postponed overflow checks on COMPARISON_VAL. */
8098 && ! (((comparison_val - add_val) ^ INTVAL (comparison_value))
8099 & comparison_sign_mask))
8101 /* Register will always be nonnegative, with value
8102 0 on last iteration */
8103 add_adjust = add_val;
8107 else if (add_val == 1 && loop->vtop
8108 && (bl->biv_count == 0
8109 || no_use_except_counting))
8117 if (GET_CODE (comparison) == LE)
8118 add_adjust -= add_val;
8120 /* If the initial value is not zero, or if the comparison
8121 value is not an exact multiple of the increment, then we
8122 can not reverse this loop. */
8123 if (initial_value == const0_rtx
8124 && GET_CODE (comparison_value) == CONST_INT)
8126 if (((unsigned HOST_WIDE_INT) comparison_val % add_val) != 0)
8131 if (! no_use_except_counting || add_val != 1)
8135 final_value = comparison_value;
8137 /* Reset these in case we normalized the initial value
8138 and comparison value above. */
8139 if (GET_CODE (comparison_value) == CONST_INT
8140 && GET_CODE (initial_value) == CONST_INT)
8142 comparison_value = GEN_INT (comparison_val);
8144 = GEN_INT (comparison_val + INTVAL (bl->initial_value));
8146 bl->initial_value = initial_value;
8148 /* Save some info needed to produce the new insns. */
8149 reg = bl->biv->dest_reg;
8150 jump_label = XEXP (SET_SRC (PATTERN (PREV_INSN (loop_end))), 1);
8151 if (jump_label == pc_rtx)
8152 jump_label = XEXP (SET_SRC (PATTERN (PREV_INSN (loop_end))), 2);
8153 new_add_val = GEN_INT (- INTVAL (bl->biv->add_val));
8155 /* Set start_value; if this is not a CONST_INT, we need
8157 Initialize biv to start_value before loop start.
8158 The old initializing insn will be deleted as a
8159 dead store by flow.c. */
8160 if (initial_value == const0_rtx
8161 && GET_CODE (comparison_value) == CONST_INT)
8163 start_value = GEN_INT (comparison_val - add_adjust);
8164 emit_insn_before (gen_move_insn (reg, start_value),
8167 else if (GET_CODE (initial_value) == CONST_INT)
8169 rtx offset = GEN_INT (-INTVAL (initial_value) - add_adjust);
8170 enum machine_mode mode = GET_MODE (reg);
8171 enum insn_code icode
8172 = add_optab->handlers[(int) mode].insn_code;
8174 if (! (*insn_data[icode].operand[0].predicate) (reg, mode)
8175 || ! ((*insn_data[icode].operand[1].predicate)
8176 (comparison_value, mode))
8177 || ! ((*insn_data[icode].operand[2].predicate)
8181 = gen_rtx_PLUS (mode, comparison_value, offset);
8182 emit_insn_before ((GEN_FCN (icode)
8183 (reg, comparison_value, offset)),
8185 if (GET_CODE (comparison) == LE)
8186 final_value = gen_rtx_PLUS (mode, comparison_value,
8189 else if (! add_adjust)
8191 enum machine_mode mode = GET_MODE (reg);
8192 enum insn_code icode
8193 = sub_optab->handlers[(int) mode].insn_code;
8194 if (! (*insn_data[icode].operand[0].predicate) (reg, mode)
8195 || ! ((*insn_data[icode].operand[1].predicate)
8196 (comparison_value, mode))
8197 || ! ((*insn_data[icode].operand[2].predicate)
8198 (initial_value, mode)))
8201 = gen_rtx_MINUS (mode, comparison_value, initial_value);
8202 emit_insn_before ((GEN_FCN (icode)
8203 (reg, comparison_value, initial_value)),
8207 /* We could handle the other cases too, but it'll be
8208 better to have a testcase first. */
8211 /* We may not have a single insn which can increment a reg, so
8212 create a sequence to hold all the insns from expand_inc. */
8214 expand_inc (reg, new_add_val);
8215 tem = gen_sequence ();
8218 p = emit_insn_before (tem, bl->biv->insn);
8219 delete_insn (bl->biv->insn);
8221 /* Update biv info to reflect its new status. */
8223 bl->initial_value = start_value;
8224 bl->biv->add_val = new_add_val;
8226 /* Update loop info. */
8227 loop_info->initial_value = reg;
8228 loop_info->initial_equiv_value = reg;
8229 loop_info->final_value = const0_rtx;
8230 loop_info->final_equiv_value = const0_rtx;
8231 loop_info->comparison_value = const0_rtx;
8232 loop_info->comparison_code = cmp_code;
8233 loop_info->increment = new_add_val;
8235 /* Inc LABEL_NUSES so that delete_insn will
8236 not delete the label. */
8237 LABEL_NUSES (XEXP (jump_label, 0)) ++;
8239 /* Emit an insn after the end of the loop to set the biv's
8240 proper exit value if it is used anywhere outside the loop. */
8241 if ((REGNO_LAST_UID (bl->regno) != INSN_UID (first_compare))
8243 || REGNO_FIRST_UID (bl->regno) != INSN_UID (bl->init_insn))
8244 emit_insn_after (gen_move_insn (reg, final_value),
8247 /* Delete compare/branch at end of loop. */
8248 delete_insn (PREV_INSN (loop_end));
8249 if (compare_and_branch == 2)
8250 delete_insn (first_compare);
8252 /* Add new compare/branch insn at end of loop. */
8254 emit_cmp_and_jump_insns (reg, const0_rtx, cmp_code, NULL_RTX,
8255 GET_MODE (reg), 0, 0,
8256 XEXP (jump_label, 0));
8257 tem = gen_sequence ();
8259 emit_jump_insn_before (tem, loop_end);
8261 for (tem = PREV_INSN (loop_end);
8262 tem && GET_CODE (tem) != JUMP_INSN;
8263 tem = PREV_INSN (tem))
8267 JUMP_LABEL (tem) = XEXP (jump_label, 0);
8273 /* Increment of LABEL_NUSES done above. */
8274 /* Register is now always nonnegative,
8275 so add REG_NONNEG note to the branch. */
8276 REG_NOTES (tem) = gen_rtx_EXPR_LIST (REG_NONNEG, reg,
8282 /* No insn may reference both the reversed and another biv or it
8283 will fail (see comment near the top of the loop reversal
8285 Earlier on, we have verified that the biv has no use except
8286 counting, or it is the only biv in this function.
8287 However, the code that computes no_use_except_counting does
8288 not verify reg notes. It's possible to have an insn that
8289 references another biv, and has a REG_EQUAL note with an
8290 expression based on the reversed biv. To avoid this case,
8291 remove all REG_EQUAL notes based on the reversed biv
8293 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
8297 rtx set = single_set (p);
8298 /* If this is a set of a GIV based on the reversed biv, any
8299 REG_EQUAL notes should still be correct. */
8301 || GET_CODE (SET_DEST (set)) != REG
8302 || (size_t) REGNO (SET_DEST (set)) >= reg_iv_type->num_elements
8303 || REG_IV_TYPE (REGNO (SET_DEST (set))) != GENERAL_INDUCT
8304 || REG_IV_INFO (REGNO (SET_DEST (set)))->src_reg != bl->biv->src_reg)
8305 for (pnote = ®_NOTES (p); *pnote;)
8307 if (REG_NOTE_KIND (*pnote) == REG_EQUAL
8308 && reg_mentioned_p (regno_reg_rtx[bl->regno],
8310 *pnote = XEXP (*pnote, 1);
8312 pnote = &XEXP (*pnote, 1);
8316 /* Mark that this biv has been reversed. Each giv which depends
8317 on this biv, and which is also live past the end of the loop
8318 will have to be fixed up. */
8322 if (loop_dump_stream)
8324 fprintf (loop_dump_stream, "Reversed loop");
8326 fprintf (loop_dump_stream, " and added reg_nonneg\n");
8328 fprintf (loop_dump_stream, "\n");
8339 /* Verify whether the biv BL appears to be eliminable,
8340 based on the insns in the loop that refer to it.
8342 If ELIMINATE_P is non-zero, actually do the elimination.
8344 THRESHOLD and INSN_COUNT are from loop_optimize and are used to
8345 determine whether invariant insns should be placed inside or at the
8346 start of the loop. */
8349 maybe_eliminate_biv (loop, bl, eliminate_p, threshold, insn_count)
8350 const struct loop *loop;
8351 struct iv_class *bl;
8353 int threshold, insn_count;
8355 rtx reg = bl->biv->dest_reg;
8356 rtx loop_start = loop->start;
8357 rtx loop_end = loop->end;
8360 /* Scan all insns in the loop, stopping if we find one that uses the
8361 biv in a way that we cannot eliminate. */
8363 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
8365 enum rtx_code code = GET_CODE (p);
8366 rtx where = threshold >= insn_count ? loop_start : p;
8368 /* If this is a libcall that sets a giv, skip ahead to its end. */
8369 if (GET_RTX_CLASS (code) == 'i')
8371 rtx note = find_reg_note (p, REG_LIBCALL, NULL_RTX);
8375 rtx last = XEXP (note, 0);
8376 rtx set = single_set (last);
8378 if (set && GET_CODE (SET_DEST (set)) == REG)
8380 unsigned int regno = REGNO (SET_DEST (set));
8382 if (regno < max_reg_before_loop
8383 && REG_IV_TYPE (regno) == GENERAL_INDUCT
8384 && REG_IV_INFO (regno)->src_reg == bl->biv->src_reg)
8389 if ((code == INSN || code == JUMP_INSN || code == CALL_INSN)
8390 && reg_mentioned_p (reg, PATTERN (p))
8391 && ! maybe_eliminate_biv_1 (loop, PATTERN (p), p, bl,
8392 eliminate_p, where))
8394 if (loop_dump_stream)
8395 fprintf (loop_dump_stream,
8396 "Cannot eliminate biv %d: biv used in insn %d.\n",
8397 bl->regno, INSN_UID (p));
8404 if (loop_dump_stream)
8405 fprintf (loop_dump_stream, "biv %d %s eliminated.\n",
8406 bl->regno, eliminate_p ? "was" : "can be");
8413 /* INSN and REFERENCE are instructions in the same insn chain.
8414 Return non-zero if INSN is first. */
8417 loop_insn_first_p (insn, reference)
8418 rtx insn, reference;
8422 for (p = insn, q = reference;;)
8424 /* Start with test for not first so that INSN == REFERENCE yields not
8426 if (q == insn || ! p)
8428 if (p == reference || ! q)
8431 /* Either of P or Q might be a NOTE. Notes have the same LUID as the
8432 previous insn, hence the <= comparison below does not work if
8434 if (INSN_UID (p) < max_uid_for_loop
8435 && INSN_UID (q) < max_uid_for_loop
8436 && GET_CODE (p) != NOTE)
8437 return INSN_LUID (p) <= INSN_LUID (q);
8439 if (INSN_UID (p) >= max_uid_for_loop
8440 || GET_CODE (p) == NOTE)
8442 if (INSN_UID (q) >= max_uid_for_loop)
8447 /* We are trying to eliminate BIV in INSN using GIV. Return non-zero if
8448 the offset that we have to take into account due to auto-increment /
8449 div derivation is zero. */
8451 biv_elimination_giv_has_0_offset (biv, giv, insn)
8452 struct induction *biv, *giv;
8455 /* If the giv V had the auto-inc address optimization applied
8456 to it, and INSN occurs between the giv insn and the biv
8457 insn, then we'd have to adjust the value used here.
8458 This is rare, so we don't bother to make this possible. */
8459 if (giv->auto_inc_opt
8460 && ((loop_insn_first_p (giv->insn, insn)
8461 && loop_insn_first_p (insn, biv->insn))
8462 || (loop_insn_first_p (biv->insn, insn)
8463 && loop_insn_first_p (insn, giv->insn))))
8466 /* If the giv V was derived from another giv, and INSN does
8467 not occur between the giv insn and the biv insn, then we'd
8468 have to adjust the value used here. This is rare, so we don't
8469 bother to make this possible. */
8470 if (giv->derived_from
8471 && ! (giv->always_executed
8472 && loop_insn_first_p (giv->insn, insn)
8473 && loop_insn_first_p (insn, biv->insn)))
8476 && giv->same->derived_from
8477 && ! (giv->same->always_executed
8478 && loop_insn_first_p (giv->same->insn, insn)
8479 && loop_insn_first_p (insn, biv->insn)))
8485 /* If BL appears in X (part of the pattern of INSN), see if we can
8486 eliminate its use. If so, return 1. If not, return 0.
8488 If BIV does not appear in X, return 1.
8490 If ELIMINATE_P is non-zero, actually do the elimination. WHERE indicates
8491 where extra insns should be added. Depending on how many items have been
8492 moved out of the loop, it will either be before INSN or at the start of
8496 maybe_eliminate_biv_1 (loop, x, insn, bl, eliminate_p, where)
8497 const struct loop *loop;
8499 struct iv_class *bl;
8503 enum rtx_code code = GET_CODE (x);
8504 rtx reg = bl->biv->dest_reg;
8505 enum machine_mode mode = GET_MODE (reg);
8506 struct induction *v;
8518 /* If we haven't already been able to do something with this BIV,
8519 we can't eliminate it. */
8525 /* If this sets the BIV, it is not a problem. */
8526 if (SET_DEST (x) == reg)
8529 /* If this is an insn that defines a giv, it is also ok because
8530 it will go away when the giv is reduced. */
8531 for (v = bl->giv; v; v = v->next_iv)
8532 if (v->giv_type == DEST_REG && SET_DEST (x) == v->dest_reg)
8536 if (SET_DEST (x) == cc0_rtx && SET_SRC (x) == reg)
8538 /* Can replace with any giv that was reduced and
8539 that has (MULT_VAL != 0) and (ADD_VAL == 0).
8540 Require a constant for MULT_VAL, so we know it's nonzero.
8541 ??? We disable this optimization to avoid potential
8544 for (v = bl->giv; v; v = v->next_iv)
8545 if (GET_CODE (v->mult_val) == CONST_INT && v->mult_val != const0_rtx
8546 && v->add_val == const0_rtx
8547 && ! v->ignore && ! v->maybe_dead && v->always_computable
8551 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8557 /* If the giv has the opposite direction of change,
8558 then reverse the comparison. */
8559 if (INTVAL (v->mult_val) < 0)
8560 new = gen_rtx_COMPARE (GET_MODE (v->new_reg),
8561 const0_rtx, v->new_reg);
8565 /* We can probably test that giv's reduced reg. */
8566 if (validate_change (insn, &SET_SRC (x), new, 0))
8570 /* Look for a giv with (MULT_VAL != 0) and (ADD_VAL != 0);
8571 replace test insn with a compare insn (cmp REDUCED_GIV ADD_VAL).
8572 Require a constant for MULT_VAL, so we know it's nonzero.
8573 ??? Do this only if ADD_VAL is a pointer to avoid a potential
8574 overflow problem. */
8576 for (v = bl->giv; v; v = v->next_iv)
8577 if (GET_CODE (v->mult_val) == CONST_INT
8578 && v->mult_val != const0_rtx
8579 && ! v->ignore && ! v->maybe_dead && v->always_computable
8581 && (GET_CODE (v->add_val) == SYMBOL_REF
8582 || GET_CODE (v->add_val) == LABEL_REF
8583 || GET_CODE (v->add_val) == CONST
8584 || (GET_CODE (v->add_val) == REG
8585 && REGNO_POINTER_FLAG (REGNO (v->add_val)))))
8587 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8593 /* If the giv has the opposite direction of change,
8594 then reverse the comparison. */
8595 if (INTVAL (v->mult_val) < 0)
8596 new = gen_rtx_COMPARE (VOIDmode, copy_rtx (v->add_val),
8599 new = gen_rtx_COMPARE (VOIDmode, v->new_reg,
8600 copy_rtx (v->add_val));
8602 /* Replace biv with the giv's reduced register. */
8603 update_reg_last_use (v->add_val, insn);
8604 if (validate_change (insn, &SET_SRC (PATTERN (insn)), new, 0))
8607 /* Insn doesn't support that constant or invariant. Copy it
8608 into a register (it will be a loop invariant.) */
8609 tem = gen_reg_rtx (GET_MODE (v->new_reg));
8611 emit_insn_before (gen_move_insn (tem, copy_rtx (v->add_val)),
8614 /* Substitute the new register for its invariant value in
8615 the compare expression. */
8616 XEXP (new, (INTVAL (v->mult_val) < 0) ? 0 : 1) = tem;
8617 if (validate_change (insn, &SET_SRC (PATTERN (insn)), new, 0))
8626 case GT: case GE: case GTU: case GEU:
8627 case LT: case LE: case LTU: case LEU:
8628 /* See if either argument is the biv. */
8629 if (XEXP (x, 0) == reg)
8630 arg = XEXP (x, 1), arg_operand = 1;
8631 else if (XEXP (x, 1) == reg)
8632 arg = XEXP (x, 0), arg_operand = 0;
8636 if (CONSTANT_P (arg))
8638 /* First try to replace with any giv that has constant positive
8639 mult_val and constant add_val. We might be able to support
8640 negative mult_val, but it seems complex to do it in general. */
8642 for (v = bl->giv; v; v = v->next_iv)
8643 if (GET_CODE (v->mult_val) == CONST_INT
8644 && INTVAL (v->mult_val) > 0
8645 && (GET_CODE (v->add_val) == SYMBOL_REF
8646 || GET_CODE (v->add_val) == LABEL_REF
8647 || GET_CODE (v->add_val) == CONST
8648 || (GET_CODE (v->add_val) == REG
8649 && REGNO_POINTER_FLAG (REGNO (v->add_val))))
8650 && ! v->ignore && ! v->maybe_dead && v->always_computable
8653 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8659 /* Replace biv with the giv's reduced reg. */
8660 validate_change (insn, &XEXP (x, 1 - arg_operand), v->new_reg, 1);
8662 /* If all constants are actually constant integers and
8663 the derived constant can be directly placed in the COMPARE,
8665 if (GET_CODE (arg) == CONST_INT
8666 && GET_CODE (v->mult_val) == CONST_INT
8667 && GET_CODE (v->add_val) == CONST_INT)
8669 validate_change (insn, &XEXP (x, arg_operand),
8670 GEN_INT (INTVAL (arg)
8671 * INTVAL (v->mult_val)
8672 + INTVAL (v->add_val)), 1);
8676 /* Otherwise, load it into a register. */
8677 tem = gen_reg_rtx (mode);
8678 emit_iv_add_mult (arg, v->mult_val, v->add_val, tem, where);
8679 validate_change (insn, &XEXP (x, arg_operand), tem, 1);
8681 if (apply_change_group ())
8685 /* Look for giv with positive constant mult_val and nonconst add_val.
8686 Insert insns to calculate new compare value.
8687 ??? Turn this off due to possible overflow. */
8689 for (v = bl->giv; v; v = v->next_iv)
8690 if (GET_CODE (v->mult_val) == CONST_INT
8691 && INTVAL (v->mult_val) > 0
8692 && ! v->ignore && ! v->maybe_dead && v->always_computable
8698 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8704 tem = gen_reg_rtx (mode);
8706 /* Replace biv with giv's reduced register. */
8707 validate_change (insn, &XEXP (x, 1 - arg_operand),
8710 /* Compute value to compare against. */
8711 emit_iv_add_mult (arg, v->mult_val, v->add_val, tem, where);
8712 /* Use it in this insn. */
8713 validate_change (insn, &XEXP (x, arg_operand), tem, 1);
8714 if (apply_change_group ())
8718 else if (GET_CODE (arg) == REG || GET_CODE (arg) == MEM)
8720 if (loop_invariant_p (loop, arg) == 1)
8722 /* Look for giv with constant positive mult_val and nonconst
8723 add_val. Insert insns to compute new compare value.
8724 ??? Turn this off due to possible overflow. */
8726 for (v = bl->giv; v; v = v->next_iv)
8727 if (GET_CODE (v->mult_val) == CONST_INT && INTVAL (v->mult_val) > 0
8728 && ! v->ignore && ! v->maybe_dead && v->always_computable
8734 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8740 tem = gen_reg_rtx (mode);
8742 /* Replace biv with giv's reduced register. */
8743 validate_change (insn, &XEXP (x, 1 - arg_operand),
8746 /* Compute value to compare against. */
8747 emit_iv_add_mult (arg, v->mult_val, v->add_val,
8749 validate_change (insn, &XEXP (x, arg_operand), tem, 1);
8750 if (apply_change_group ())
8755 /* This code has problems. Basically, you can't know when
8756 seeing if we will eliminate BL, whether a particular giv
8757 of ARG will be reduced. If it isn't going to be reduced,
8758 we can't eliminate BL. We can try forcing it to be reduced,
8759 but that can generate poor code.
8761 The problem is that the benefit of reducing TV, below should
8762 be increased if BL can actually be eliminated, but this means
8763 we might have to do a topological sort of the order in which
8764 we try to process biv. It doesn't seem worthwhile to do
8765 this sort of thing now. */
8768 /* Otherwise the reg compared with had better be a biv. */
8769 if (GET_CODE (arg) != REG
8770 || REG_IV_TYPE (REGNO (arg)) != BASIC_INDUCT)
8773 /* Look for a pair of givs, one for each biv,
8774 with identical coefficients. */
8775 for (v = bl->giv; v; v = v->next_iv)
8777 struct induction *tv;
8779 if (v->ignore || v->maybe_dead || v->mode != mode)
8782 for (tv = reg_biv_class[REGNO (arg)]->giv; tv; tv = tv->next_iv)
8783 if (! tv->ignore && ! tv->maybe_dead
8784 && rtx_equal_p (tv->mult_val, v->mult_val)
8785 && rtx_equal_p (tv->add_val, v->add_val)
8786 && tv->mode == mode)
8788 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8794 /* Replace biv with its giv's reduced reg. */
8795 XEXP (x, 1 - arg_operand) = v->new_reg;
8796 /* Replace other operand with the other giv's
8798 XEXP (x, arg_operand) = tv->new_reg;
8805 /* If we get here, the biv can't be eliminated. */
8809 /* If this address is a DEST_ADDR giv, it doesn't matter if the
8810 biv is used in it, since it will be replaced. */
8811 for (v = bl->giv; v; v = v->next_iv)
8812 if (v->giv_type == DEST_ADDR && v->location == &XEXP (x, 0))
8820 /* See if any subexpression fails elimination. */
8821 fmt = GET_RTX_FORMAT (code);
8822 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
8827 if (! maybe_eliminate_biv_1 (loop, XEXP (x, i), insn, bl,
8828 eliminate_p, where))
8833 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
8834 if (! maybe_eliminate_biv_1 (loop, XVECEXP (x, i, j), insn, bl,
8835 eliminate_p, where))
8844 /* Return nonzero if the last use of REG
8845 is in an insn following INSN in the same basic block. */
8848 last_use_this_basic_block (reg, insn)
8854 n && GET_CODE (n) != CODE_LABEL && GET_CODE (n) != JUMP_INSN;
8857 if (REGNO_LAST_UID (REGNO (reg)) == INSN_UID (n))
8863 /* Called via `note_stores' to record the initial value of a biv. Here we
8864 just record the location of the set and process it later. */
8867 record_initial (dest, set, data)
8870 void *data ATTRIBUTE_UNUSED;
8872 struct iv_class *bl;
8874 if (GET_CODE (dest) != REG
8875 || REGNO (dest) >= max_reg_before_loop
8876 || REG_IV_TYPE (REGNO (dest)) != BASIC_INDUCT)
8879 bl = reg_biv_class[REGNO (dest)];
8881 /* If this is the first set found, record it. */
8882 if (bl->init_insn == 0)
8884 bl->init_insn = note_insn;
8889 /* If any of the registers in X are "old" and currently have a last use earlier
8890 than INSN, update them to have a last use of INSN. Their actual last use
8891 will be the previous insn but it will not have a valid uid_luid so we can't
8895 update_reg_last_use (x, insn)
8899 /* Check for the case where INSN does not have a valid luid. In this case,
8900 there is no need to modify the regno_last_uid, as this can only happen
8901 when code is inserted after the loop_end to set a pseudo's final value,
8902 and hence this insn will never be the last use of x. */
8903 if (GET_CODE (x) == REG && REGNO (x) < max_reg_before_loop
8904 && INSN_UID (insn) < max_uid_for_loop
8905 && uid_luid[REGNO_LAST_UID (REGNO (x))] < uid_luid[INSN_UID (insn)])
8906 REGNO_LAST_UID (REGNO (x)) = INSN_UID (insn);
8910 register const char *fmt = GET_RTX_FORMAT (GET_CODE (x));
8911 for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
8914 update_reg_last_use (XEXP (x, i), insn);
8915 else if (fmt[i] == 'E')
8916 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
8917 update_reg_last_use (XVECEXP (x, i, j), insn);
8922 /* Given an insn INSN and condition COND, return the condition in a
8923 canonical form to simplify testing by callers. Specifically:
8925 (1) The code will always be a comparison operation (EQ, NE, GT, etc.).
8926 (2) Both operands will be machine operands; (cc0) will have been replaced.
8927 (3) If an operand is a constant, it will be the second operand.
8928 (4) (LE x const) will be replaced with (LT x <const+1>) and similarly
8929 for GE, GEU, and LEU.
8931 If the condition cannot be understood, or is an inequality floating-point
8932 comparison which needs to be reversed, 0 will be returned.
8934 If REVERSE is non-zero, then reverse the condition prior to canonizing it.
8936 If EARLIEST is non-zero, it is a pointer to a place where the earliest
8937 insn used in locating the condition was found. If a replacement test
8938 of the condition is desired, it should be placed in front of that
8939 insn and we will be sure that the inputs are still valid.
8941 If WANT_REG is non-zero, we wish the condition to be relative to that
8942 register, if possible. Therefore, do not canonicalize the condition
8946 canonicalize_condition (insn, cond, reverse, earliest, want_reg)
8958 int reverse_code = 0;
8959 int did_reverse_condition = 0;
8960 enum machine_mode mode;
8962 code = GET_CODE (cond);
8963 mode = GET_MODE (cond);
8964 op0 = XEXP (cond, 0);
8965 op1 = XEXP (cond, 1);
8969 code = reverse_condition (code);
8970 did_reverse_condition ^= 1;
8976 /* If we are comparing a register with zero, see if the register is set
8977 in the previous insn to a COMPARE or a comparison operation. Perform
8978 the same tests as a function of STORE_FLAG_VALUE as find_comparison_args
8981 while (GET_RTX_CLASS (code) == '<'
8982 && op1 == CONST0_RTX (GET_MODE (op0))
8985 /* Set non-zero when we find something of interest. */
8989 /* If comparison with cc0, import actual comparison from compare
8993 if ((prev = prev_nonnote_insn (prev)) == 0
8994 || GET_CODE (prev) != INSN
8995 || (set = single_set (prev)) == 0
8996 || SET_DEST (set) != cc0_rtx)
8999 op0 = SET_SRC (set);
9000 op1 = CONST0_RTX (GET_MODE (op0));
9006 /* If this is a COMPARE, pick up the two things being compared. */
9007 if (GET_CODE (op0) == COMPARE)
9009 op1 = XEXP (op0, 1);
9010 op0 = XEXP (op0, 0);
9013 else if (GET_CODE (op0) != REG)
9016 /* Go back to the previous insn. Stop if it is not an INSN. We also
9017 stop if it isn't a single set or if it has a REG_INC note because
9018 we don't want to bother dealing with it. */
9020 if ((prev = prev_nonnote_insn (prev)) == 0
9021 || GET_CODE (prev) != INSN
9022 || FIND_REG_INC_NOTE (prev, 0)
9023 || (set = single_set (prev)) == 0)
9026 /* If this is setting OP0, get what it sets it to if it looks
9028 if (rtx_equal_p (SET_DEST (set), op0))
9030 enum machine_mode inner_mode = GET_MODE (SET_DEST (set));
9032 /* ??? We may not combine comparisons done in a CCmode with
9033 comparisons not done in a CCmode. This is to aid targets
9034 like Alpha that have an IEEE compliant EQ instruction, and
9035 a non-IEEE compliant BEQ instruction. The use of CCmode is
9036 actually artificial, simply to prevent the combination, but
9037 should not affect other platforms.
9039 However, we must allow VOIDmode comparisons to match either
9040 CCmode or non-CCmode comparison, because some ports have
9041 modeless comparisons inside branch patterns.
9043 ??? This mode check should perhaps look more like the mode check
9044 in simplify_comparison in combine. */
9046 if ((GET_CODE (SET_SRC (set)) == COMPARE
9049 && GET_MODE_CLASS (inner_mode) == MODE_INT
9050 && (GET_MODE_BITSIZE (inner_mode)
9051 <= HOST_BITS_PER_WIDE_INT)
9052 && (STORE_FLAG_VALUE
9053 & ((HOST_WIDE_INT) 1
9054 << (GET_MODE_BITSIZE (inner_mode) - 1))))
9055 #ifdef FLOAT_STORE_FLAG_VALUE
9057 && GET_MODE_CLASS (inner_mode) == MODE_FLOAT
9058 && (REAL_VALUE_NEGATIVE
9059 (FLOAT_STORE_FLAG_VALUE (inner_mode))))
9062 && GET_RTX_CLASS (GET_CODE (SET_SRC (set))) == '<'))
9063 && (((GET_MODE_CLASS (mode) == MODE_CC)
9064 == (GET_MODE_CLASS (inner_mode) == MODE_CC))
9065 || mode == VOIDmode || inner_mode == VOIDmode))
9067 else if (((code == EQ
9069 && (GET_MODE_BITSIZE (inner_mode)
9070 <= HOST_BITS_PER_WIDE_INT)
9071 && GET_MODE_CLASS (inner_mode) == MODE_INT
9072 && (STORE_FLAG_VALUE
9073 & ((HOST_WIDE_INT) 1
9074 << (GET_MODE_BITSIZE (inner_mode) - 1))))
9075 #ifdef FLOAT_STORE_FLAG_VALUE
9077 && GET_MODE_CLASS (inner_mode) == MODE_FLOAT
9078 && (REAL_VALUE_NEGATIVE
9079 (FLOAT_STORE_FLAG_VALUE (inner_mode))))
9082 && GET_RTX_CLASS (GET_CODE (SET_SRC (set))) == '<'
9083 && (((GET_MODE_CLASS (mode) == MODE_CC)
9084 == (GET_MODE_CLASS (inner_mode) == MODE_CC))
9085 || mode == VOIDmode || inner_mode == VOIDmode))
9088 /* We might have reversed a LT to get a GE here. But this wasn't
9089 actually the comparison of data, so we don't flag that we
9090 have had to reverse the condition. */
9091 did_reverse_condition ^= 1;
9099 else if (reg_set_p (op0, prev))
9100 /* If this sets OP0, but not directly, we have to give up. */
9105 if (GET_RTX_CLASS (GET_CODE (x)) == '<')
9106 code = GET_CODE (x);
9109 code = reverse_condition (code);
9110 if (code == UNKNOWN)
9112 did_reverse_condition ^= 1;
9116 op0 = XEXP (x, 0), op1 = XEXP (x, 1);
9122 /* If constant is first, put it last. */
9123 if (CONSTANT_P (op0))
9124 code = swap_condition (code), tem = op0, op0 = op1, op1 = tem;
9126 /* If OP0 is the result of a comparison, we weren't able to find what
9127 was really being compared, so fail. */
9128 if (GET_MODE_CLASS (GET_MODE (op0)) == MODE_CC)
9131 /* Canonicalize any ordered comparison with integers involving equality
9132 if we can do computations in the relevant mode and we do not
9135 if (GET_CODE (op1) == CONST_INT
9136 && GET_MODE (op0) != VOIDmode
9137 && GET_MODE_BITSIZE (GET_MODE (op0)) <= HOST_BITS_PER_WIDE_INT)
9139 HOST_WIDE_INT const_val = INTVAL (op1);
9140 unsigned HOST_WIDE_INT uconst_val = const_val;
9141 unsigned HOST_WIDE_INT max_val
9142 = (unsigned HOST_WIDE_INT) GET_MODE_MASK (GET_MODE (op0));
9147 if ((unsigned HOST_WIDE_INT) const_val != max_val >> 1)
9148 code = LT, op1 = GEN_INT (const_val + 1);
9151 /* When cross-compiling, const_val might be sign-extended from
9152 BITS_PER_WORD to HOST_BITS_PER_WIDE_INT */
9154 if ((HOST_WIDE_INT) (const_val & max_val)
9155 != (((HOST_WIDE_INT) 1
9156 << (GET_MODE_BITSIZE (GET_MODE (op0)) - 1))))
9157 code = GT, op1 = GEN_INT (const_val - 1);
9161 if (uconst_val < max_val)
9162 code = LTU, op1 = GEN_INT (uconst_val + 1);
9166 if (uconst_val != 0)
9167 code = GTU, op1 = GEN_INT (uconst_val - 1);
9175 /* If this was floating-point and we reversed anything other than an
9176 EQ or NE or (UN)ORDERED, return zero. */
9177 if (TARGET_FLOAT_FORMAT == IEEE_FLOAT_FORMAT
9178 && did_reverse_condition
9179 && code != NE && code != EQ && code != UNORDERED && code != ORDERED
9181 && GET_MODE_CLASS (GET_MODE (op0)) == MODE_FLOAT)
9185 /* Never return CC0; return zero instead. */
9190 return gen_rtx_fmt_ee (code, VOIDmode, op0, op1);
9193 /* Given a jump insn JUMP, return the condition that will cause it to branch
9194 to its JUMP_LABEL. If the condition cannot be understood, or is an
9195 inequality floating-point comparison which needs to be reversed, 0 will
9198 If EARLIEST is non-zero, it is a pointer to a place where the earliest
9199 insn used in locating the condition was found. If a replacement test
9200 of the condition is desired, it should be placed in front of that
9201 insn and we will be sure that the inputs are still valid. */
9204 get_condition (jump, earliest)
9212 /* If this is not a standard conditional jump, we can't parse it. */
9213 if (GET_CODE (jump) != JUMP_INSN
9214 || ! any_condjump_p (jump))
9216 set = pc_set (jump);
9218 cond = XEXP (SET_SRC (set), 0);
9220 /* If this branches to JUMP_LABEL when the condition is false, reverse
9223 = GET_CODE (XEXP (SET_SRC (set), 2)) == LABEL_REF
9224 && XEXP (XEXP (SET_SRC (set), 2), 0) == JUMP_LABEL (jump);
9226 return canonicalize_condition (jump, cond, reverse, earliest, NULL_RTX);
9229 /* Similar to above routine, except that we also put an invariant last
9230 unless both operands are invariants. */
9233 get_condition_for_loop (loop, x)
9234 const struct loop *loop;
9237 rtx comparison = get_condition (x, NULL_PTR);
9240 || ! loop_invariant_p (loop, XEXP (comparison, 0))
9241 || loop_invariant_p (loop, XEXP (comparison, 1)))
9244 return gen_rtx_fmt_ee (swap_condition (GET_CODE (comparison)), VOIDmode,
9245 XEXP (comparison, 1), XEXP (comparison, 0));
9248 /* Scan the function and determine whether it has indirect (computed) jumps.
9250 This is taken mostly from flow.c; similar code exists elsewhere
9251 in the compiler. It may be useful to put this into rtlanal.c. */
9253 indirect_jump_in_function_p (start)
9258 for (insn = start; insn; insn = NEXT_INSN (insn))
9259 if (computed_jump_p (insn))
9265 /* Add MEM to the LOOP_MEMS array, if appropriate. See the
9266 documentation for LOOP_MEMS for the definition of `appropriate'.
9267 This function is called from prescan_loop via for_each_rtx. */
9270 insert_loop_mem (mem, data)
9272 void *data ATTRIBUTE_UNUSED;
9280 switch (GET_CODE (m))
9286 /* We're not interested in MEMs that are only clobbered. */
9290 /* We're not interested in the MEM associated with a
9291 CONST_DOUBLE, so there's no need to traverse into this. */
9295 /* We're not interested in any MEMs that only appear in notes. */
9299 /* This is not a MEM. */
9303 /* See if we've already seen this MEM. */
9304 for (i = 0; i < loop_mems_idx; ++i)
9305 if (rtx_equal_p (m, loop_mems[i].mem))
9307 if (GET_MODE (m) != GET_MODE (loop_mems[i].mem))
9308 /* The modes of the two memory accesses are different. If
9309 this happens, something tricky is going on, and we just
9310 don't optimize accesses to this MEM. */
9311 loop_mems[i].optimize = 0;
9316 /* Resize the array, if necessary. */
9317 if (loop_mems_idx == loop_mems_allocated)
9319 if (loop_mems_allocated != 0)
9320 loop_mems_allocated *= 2;
9322 loop_mems_allocated = 32;
9324 loop_mems = (loop_mem_info*)
9325 xrealloc (loop_mems,
9326 loop_mems_allocated * sizeof (loop_mem_info));
9329 /* Actually insert the MEM. */
9330 loop_mems[loop_mems_idx].mem = m;
9331 /* We can't hoist this MEM out of the loop if it's a BLKmode MEM
9332 because we can't put it in a register. We still store it in the
9333 table, though, so that if we see the same address later, but in a
9334 non-BLK mode, we'll not think we can optimize it at that point. */
9335 loop_mems[loop_mems_idx].optimize = (GET_MODE (m) != BLKmode);
9336 loop_mems[loop_mems_idx].reg = NULL_RTX;
9342 /* Like load_mems, but also ensures that SET_IN_LOOP,
9343 MAY_NOT_OPTIMIZE, REG_SINGLE_USAGE, and INSN_COUNT have the correct
9344 values after load_mems. */
9347 load_mems_and_recount_loop_regs_set (loop, insn_count)
9348 const struct loop *loop;
9351 int nregs = max_reg_num ();
9355 /* Recalculate set_in_loop and friends since load_mems may have
9356 created new registers. */
9357 if (max_reg_num () > nregs)
9363 nregs = max_reg_num ();
9365 if ((unsigned) nregs > set_in_loop->num_elements)
9367 /* Grow all the arrays. */
9368 VARRAY_GROW (set_in_loop, nregs);
9369 VARRAY_GROW (n_times_set, nregs);
9370 VARRAY_GROW (may_not_optimize, nregs);
9371 VARRAY_GROW (reg_single_usage, nregs);
9373 /* Clear the arrays */
9374 bzero ((char *) &set_in_loop->data, nregs * sizeof (int));
9375 bzero ((char *) &may_not_optimize->data, nregs * sizeof (char));
9376 bzero ((char *) ®_single_usage->data, nregs * sizeof (rtx));
9378 count_loop_regs_set (loop, may_not_optimize, reg_single_usage,
9381 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
9383 VARRAY_CHAR (may_not_optimize, i) = 1;
9384 VARRAY_INT (set_in_loop, i) = 1;
9387 #ifdef AVOID_CCMODE_COPIES
9388 /* Don't try to move insns which set CC registers if we should not
9389 create CCmode register copies. */
9390 for (i = max_reg_num () - 1; i >= FIRST_PSEUDO_REGISTER; i--)
9391 if (GET_MODE_CLASS (GET_MODE (regno_reg_rtx[i])) == MODE_CC)
9392 VARRAY_CHAR (may_not_optimize, i) = 1;
9395 /* Set n_times_set for the new registers. */
9396 bcopy ((char *) (&set_in_loop->data.i[0] + old_nregs),
9397 (char *) (&n_times_set->data.i[0] + old_nregs),
9398 (nregs - old_nregs) * sizeof (int));
9402 /* Move MEMs into registers for the duration of the loop. */
9406 const struct loop *loop;
9408 int maybe_never = 0;
9411 rtx label = NULL_RTX;
9412 rtx end_label = NULL_RTX;
9413 /* Nonzero if the next instruction may never be executed. */
9414 int next_maybe_never = 0;
9415 int last_max_reg = max_reg_num ();
9417 if (loop_mems_idx == 0)
9420 /* Find start of the extended basic block that enters the loop. */
9421 for (p = loop->start;
9422 PREV_INSN (p) && GET_CODE (p) != CODE_LABEL;
9428 /* Build table of mems that get set to constant values before the
9430 for (; p != loop->start; p = NEXT_INSN (p))
9431 cselib_process_insn (p);
9433 /* Check to see if it's possible that some instructions in the
9434 loop are never executed. */
9435 for (p = next_insn_in_loop (loop, loop->scan_start);
9436 p != NULL_RTX && ! maybe_never;
9437 p = next_insn_in_loop (loop, p))
9439 if (GET_CODE (p) == CODE_LABEL)
9441 else if (GET_CODE (p) == JUMP_INSN
9442 /* If we enter the loop in the middle, and scan
9443 around to the beginning, don't set maybe_never
9444 for that. This must be an unconditional jump,
9445 otherwise the code at the top of the loop might
9446 never be executed. Unconditional jumps are
9447 followed a by barrier then loop end. */
9448 && ! (GET_CODE (p) == JUMP_INSN
9449 && JUMP_LABEL (p) == loop->top
9450 && NEXT_INSN (NEXT_INSN (p)) == loop->end
9451 && any_uncondjump_p (p)))
9453 if (!any_condjump_p (p))
9454 /* Something complicated. */
9457 /* If there are any more instructions in the loop, they
9458 might not be reached. */
9459 next_maybe_never = 1;
9461 else if (next_maybe_never)
9465 /* Actually move the MEMs. */
9466 for (i = 0; i < loop_mems_idx; ++i)
9468 regset_head load_copies;
9469 regset_head store_copies;
9472 rtx mem = loop_mems[i].mem;
9475 if (MEM_VOLATILE_P (mem)
9476 || loop_invariant_p (loop, XEXP (mem, 0)) != 1)
9477 /* There's no telling whether or not MEM is modified. */
9478 loop_mems[i].optimize = 0;
9480 /* Go through the MEMs written to in the loop to see if this
9481 one is aliased by one of them. */
9482 mem_list_entry = loop_store_mems;
9483 while (mem_list_entry)
9485 if (rtx_equal_p (mem, XEXP (mem_list_entry, 0)))
9487 else if (true_dependence (XEXP (mem_list_entry, 0), VOIDmode,
9490 /* MEM is indeed aliased by this store. */
9491 loop_mems[i].optimize = 0;
9494 mem_list_entry = XEXP (mem_list_entry, 1);
9497 if (flag_float_store && written
9498 && GET_MODE_CLASS (GET_MODE (mem)) == MODE_FLOAT)
9499 loop_mems[i].optimize = 0;
9501 /* If this MEM is written to, we must be sure that there
9502 are no reads from another MEM that aliases this one. */
9503 if (loop_mems[i].optimize && written)
9507 for (j = 0; j < loop_mems_idx; ++j)
9511 else if (true_dependence (mem,
9516 /* It's not safe to hoist loop_mems[i] out of
9517 the loop because writes to it might not be
9518 seen by reads from loop_mems[j]. */
9519 loop_mems[i].optimize = 0;
9525 if (maybe_never && may_trap_p (mem))
9526 /* We can't access the MEM outside the loop; it might
9527 cause a trap that wouldn't have happened otherwise. */
9528 loop_mems[i].optimize = 0;
9530 if (!loop_mems[i].optimize)
9531 /* We thought we were going to lift this MEM out of the
9532 loop, but later discovered that we could not. */
9535 INIT_REG_SET (&load_copies);
9536 INIT_REG_SET (&store_copies);
9538 /* Allocate a pseudo for this MEM. We set REG_USERVAR_P in
9539 order to keep scan_loop from moving stores to this MEM
9540 out of the loop just because this REG is neither a
9541 user-variable nor used in the loop test. */
9542 reg = gen_reg_rtx (GET_MODE (mem));
9543 REG_USERVAR_P (reg) = 1;
9544 loop_mems[i].reg = reg;
9546 /* Now, replace all references to the MEM with the
9547 corresponding pesudos. */
9549 for (p = next_insn_in_loop (loop, loop->scan_start);
9551 p = next_insn_in_loop (loop, p))
9559 set = single_set (p);
9561 /* See if this copies the mem into a register that isn't
9562 modified afterwards. We'll try to do copy propagation
9563 a little further on. */
9565 /* @@@ This test is _way_ too conservative. */
9567 && GET_CODE (SET_DEST (set)) == REG
9568 && REGNO (SET_DEST (set)) >= FIRST_PSEUDO_REGISTER
9569 && REGNO (SET_DEST (set)) < last_max_reg
9570 && VARRAY_INT (n_times_set, REGNO (SET_DEST (set))) == 1
9571 && rtx_equal_p (SET_SRC (set), mem))
9572 SET_REGNO_REG_SET (&load_copies, REGNO (SET_DEST (set)));
9574 /* See if this copies the mem from a register that isn't
9575 modified afterwards. We'll try to remove the
9576 redundant copy later on by doing a little register
9577 renaming and copy propagation. This will help
9578 to untangle things for the BIV detection code. */
9581 && GET_CODE (SET_SRC (set)) == REG
9582 && REGNO (SET_SRC (set)) >= FIRST_PSEUDO_REGISTER
9583 && REGNO (SET_SRC (set)) < last_max_reg
9584 && VARRAY_INT (n_times_set, REGNO (SET_SRC (set))) == 1
9585 && rtx_equal_p (SET_DEST (set), mem))
9586 SET_REGNO_REG_SET (&store_copies, REGNO (SET_SRC (set)));
9588 /* Replace the memory reference with the shadow register. */
9591 for_each_rtx (&p, replace_loop_mem, &ri);
9594 if (GET_CODE (p) == CODE_LABEL
9595 || GET_CODE (p) == JUMP_INSN)
9599 if (! apply_change_group ())
9600 /* We couldn't replace all occurrences of the MEM. */
9601 loop_mems[i].optimize = 0;
9604 /* Load the memory immediately before LOOP->START, which is
9605 the NOTE_LOOP_BEG. */
9606 cselib_val *e = cselib_lookup (mem, VOIDmode, 0);
9610 struct elt_loc_list *const_equiv = 0;
9614 struct elt_loc_list *equiv;
9615 struct elt_loc_list *best_equiv = 0;
9616 for (equiv = e->locs; equiv; equiv = equiv->next)
9618 if (CONSTANT_P (equiv->loc))
9619 const_equiv = equiv;
9620 else if (GET_CODE (equiv->loc) == REG
9621 /* Extending hard register lifetimes cuases crash
9622 on SRC targets. Doing so on non-SRC is
9623 probably also not good idea, since we most
9624 probably have pseudoregister equivalence as
9626 && REGNO (equiv->loc) >= FIRST_PSEUDO_REGISTER)
9629 /* Use the constant equivalence if that is cheap enough. */
9631 best_equiv = const_equiv;
9632 else if (const_equiv
9633 && (rtx_cost (const_equiv->loc, SET)
9634 <= rtx_cost (best_equiv->loc, SET)))
9636 best_equiv = const_equiv;
9640 /* If best_equiv is nonzero, we know that MEM is set to a
9641 constant or register before the loop. We will use this
9642 knowledge to initialize the shadow register with that
9643 constant or reg rather than by loading from MEM. */
9645 best = copy_rtx (best_equiv->loc);
9647 set = gen_move_insn (reg, best);
9648 set = emit_insn_before (set, loop->start);
9650 REG_NOTES (set) = gen_rtx_EXPR_LIST (REG_EQUAL,
9651 copy_rtx (const_equiv->loc),
9656 if (label == NULL_RTX)
9658 /* We must compute the former
9659 right-after-the-end label before we insert
9661 end_label = next_label (loop->end);
9662 label = gen_label_rtx ();
9663 emit_label_after (label, loop->end);
9666 /* Store the memory immediately after END, which is
9667 the NOTE_LOOP_END. */
9668 set = gen_move_insn (copy_rtx (mem), reg);
9669 emit_insn_after (set, label);
9672 if (loop_dump_stream)
9674 fprintf (loop_dump_stream, "Hoisted regno %d %s from ",
9675 REGNO (reg), (written ? "r/w" : "r/o"));
9676 print_rtl (loop_dump_stream, mem);
9677 fputc ('\n', loop_dump_stream);
9680 /* Attempt a bit of copy propagation. This helps untangle the
9681 data flow, and enables {basic,general}_induction_var to find
9683 EXECUTE_IF_SET_IN_REG_SET
9684 (&load_copies, FIRST_PSEUDO_REGISTER, j,
9686 try_copy_prop (loop, reg, j);
9688 CLEAR_REG_SET (&load_copies);
9690 EXECUTE_IF_SET_IN_REG_SET
9691 (&store_copies, FIRST_PSEUDO_REGISTER, j,
9693 try_swap_copy_prop (loop, reg, j);
9695 CLEAR_REG_SET (&store_copies);
9699 if (label != NULL_RTX)
9701 /* Now, we need to replace all references to the previous exit
9702 label with the new one. */
9707 for (p = loop->start; p != loop->end; p = NEXT_INSN (p))
9709 for_each_rtx (&p, replace_label, &rr);
9711 /* If this is a JUMP_INSN, then we also need to fix the JUMP_LABEL
9712 field. This is not handled by for_each_rtx because it doesn't
9713 handle unprinted ('0') fields. We need to update JUMP_LABEL
9714 because the immediately following unroll pass will use it.
9715 replace_label would not work anyways, because that only handles
9717 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) == end_label)
9718 JUMP_LABEL (p) = label;
9725 /* For communication between note_reg_stored and its caller. */
9726 struct note_reg_stored_arg
9732 /* Called via note_stores, record in SET_SEEN whether X, which is written,
9735 note_reg_stored (x, setter, arg)
9736 rtx x, setter ATTRIBUTE_UNUSED;
9739 struct note_reg_stored_arg *t = (struct note_reg_stored_arg *) arg;
9744 /* Try to replace every occurrence of pseudo REGNO with REPLACEMENT.
9745 There must be exactly one insn that sets this pseudo; it will be
9746 deleted if all replacements succeed and we can prove that the register
9747 is not used after the loop. */
9750 try_copy_prop (loop, replacement, regno)
9751 const struct loop *loop;
9755 /* This is the reg that we are copying from. */
9756 rtx reg_rtx = regno_reg_rtx[regno];
9759 /* These help keep track of whether we replaced all uses of the reg. */
9760 int replaced_last = 0;
9761 int store_is_first = 0;
9763 for (insn = next_insn_in_loop (loop, loop->scan_start);
9765 insn = next_insn_in_loop (loop, insn))
9769 /* Only substitute within one extended basic block from the initializing
9771 if (GET_CODE (insn) == CODE_LABEL && init_insn)
9774 if (! INSN_P (insn))
9777 /* Is this the initializing insn? */
9778 set = single_set (insn);
9780 && GET_CODE (SET_DEST (set)) == REG
9781 && REGNO (SET_DEST (set)) == regno)
9787 if (REGNO_FIRST_UID (regno) == INSN_UID (insn))
9791 /* Only substitute after seeing the initializing insn. */
9792 if (init_insn && insn != init_insn)
9794 struct note_reg_stored_arg arg;
9797 array[1] = replacement;
9800 for_each_rtx (&insn, replace_loop_reg, array);
9801 if (REGNO_LAST_UID (regno) == INSN_UID (insn))
9804 /* Stop replacing when REPLACEMENT is modified. */
9805 arg.reg = replacement;
9807 note_stores (PATTERN (insn), note_reg_stored, &arg);
9814 if (apply_change_group ())
9816 if (loop_dump_stream)
9817 fprintf (loop_dump_stream, " Replaced reg %d", regno);
9818 if (store_is_first && replaced_last)
9820 PUT_CODE (init_insn, NOTE);
9821 NOTE_LINE_NUMBER (init_insn) = NOTE_INSN_DELETED;
9822 if (loop_dump_stream)
9823 fprintf (loop_dump_stream, ", deleting init_insn (%d)",
9824 INSN_UID (init_insn));
9826 if (loop_dump_stream)
9827 fprintf (loop_dump_stream, ".\n");
9832 /* Try to replace occurrences of pseudo REGNO with REPLACEMENT within
9833 loop LOOP if the order of the sets of these registers can be
9834 swapped. There must be exactly one insn within the loop that sets
9835 this pseudo followed immediately by a move insn that sets
9836 REPLACEMENT with REGNO. */
9838 try_swap_copy_prop (loop, replacement, regno)
9839 const struct loop *loop;
9845 unsigned int new_regno;
9847 new_regno = REGNO (replacement);
9849 for (insn = next_insn_in_loop (loop, loop->scan_start);
9851 insn = next_insn_in_loop (loop, insn))
9853 /* Search for the insn that copies REGNO to NEW_REGNO? */
9854 if (GET_RTX_CLASS (GET_CODE (insn)) == 'i'
9855 && (set = single_set (insn))
9856 && GET_CODE (SET_DEST (set)) == REG
9857 && REGNO (SET_DEST (set)) == new_regno
9858 && GET_CODE (SET_SRC (set)) == REG
9859 && REGNO (SET_SRC (set)) == regno)
9863 if (insn != NULL_RTX)
9868 /* Some DEF-USE info would come in handy here to make this
9869 function more general. For now, just check the previous insn
9870 which is the most likely candidate for setting REGNO. */
9872 prev_insn = PREV_INSN (insn);
9874 if (GET_RTX_CLASS (GET_CODE (insn)) == 'i'
9875 && (prev_set = single_set (prev_insn))
9876 && GET_CODE (SET_DEST (prev_set)) == REG
9877 && REGNO (SET_DEST (prev_set)) == regno)
9880 (set (reg regno) (expr))
9881 (set (reg new_regno) (reg regno))
9883 so try converting this to:
9884 (set (reg new_regno) (expr))
9885 (set (reg regno) (reg new_regno))
9887 The former construct is often generated when a global
9888 variable used for an induction variable is shadowed by a
9889 register (NEW_REGNO). The latter construct improves the
9890 chances of GIV replacement and BIV elimination. */
9892 validate_change (prev_insn, &SET_DEST (prev_set),
9894 validate_change (insn, &SET_DEST (set),
9896 validate_change (insn, &SET_SRC (set),
9899 if (apply_change_group ())
9901 if (loop_dump_stream)
9902 fprintf (loop_dump_stream,
9903 " Swapped set of reg %d at %d with reg %d at %d.\n",
9904 regno, INSN_UID (insn),
9905 new_regno, INSN_UID (prev_insn));
9907 /* Update first use of REGNO. */
9908 if (REGNO_FIRST_UID (regno) == INSN_UID (prev_insn))
9909 REGNO_FIRST_UID (regno) = INSN_UID (insn);
9911 /* Now perform copy propagation to hopefully
9912 remove all uses of REGNO within the loop. */
9913 try_copy_prop (loop, replacement, regno);
9920 /* Replace MEM with its associated pseudo register. This function is
9921 called from load_mems via for_each_rtx. DATA is actually an
9922 rtx_and_int * describing the instruction currently being scanned
9923 and the MEM we are currently replacing. */
9926 replace_loop_mem (mem, data)
9938 switch (GET_CODE (m))
9944 /* We're not interested in the MEM associated with a
9945 CONST_DOUBLE, so there's no need to traverse into one. */
9949 /* This is not a MEM. */
9953 ri = (rtx_and_int *) data;
9956 if (!rtx_equal_p (loop_mems[i].mem, m))
9957 /* This is not the MEM we are currently replacing. */
9962 /* Actually replace the MEM. */
9963 validate_change (insn, mem, loop_mems[i].reg, 1);
9968 /* Replace one register with another. Called through for_each_rtx; PX points
9969 to the rtx being scanned. DATA is actually an array of three rtx's; the
9970 first one is the one to be replaced, and the second one the replacement.
9971 The third one is the current insn. */
9974 replace_loop_reg (px, data)
9979 rtx *array = (rtx *) data;
9985 validate_change (array[2], px, array[1], 1);
9990 /* Replace occurrences of the old exit label for the loop with the new
9991 one. DATA is an rtx_pair containing the old and new labels,
9995 replace_label (x, data)
10000 rtx old_label = ((rtx_pair *) data)->r1;
10001 rtx new_label = ((rtx_pair *) data)->r2;
10006 if (GET_CODE (l) != LABEL_REF)
10009 if (XEXP (l, 0) != old_label)
10012 XEXP (l, 0) = new_label;
10013 ++LABEL_NUSES (new_label);
10014 --LABEL_NUSES (old_label);
10019 #define LOOP_BLOCK_NUM_1(INSN) \
10020 ((INSN) ? (BLOCK_FOR_INSN (INSN) ? BLOCK_NUM (INSN) : - 1) : -1)
10022 /* The notes do not have an assigned block, so look at the next insn. */
10023 #define LOOP_BLOCK_NUM(INSN) \
10024 ((INSN) ? (GET_CODE (INSN) == NOTE \
10025 ? LOOP_BLOCK_NUM_1 (next_nonnote_insn (INSN)) \
10026 : LOOP_BLOCK_NUM_1 (INSN)) \
10029 #define LOOP_INSN_UID(INSN) ((INSN) ? INSN_UID (INSN) : -1)
10031 static void loop_dump_aux (loop, file, verbose)
10032 const struct loop *loop;
10038 if (! loop || ! file)
10041 /* Print diagnostics to compare our concept of a loop with
10042 what the loop notes say. */
10043 if (! PREV_INSN (loop->first->head)
10044 || GET_CODE (PREV_INSN (loop->first->head)) != NOTE
10045 || NOTE_LINE_NUMBER (PREV_INSN (loop->first->head))
10046 != NOTE_INSN_LOOP_BEG)
10047 fprintf (file, ";; No NOTE_INSN_LOOP_BEG at %d\n",
10048 INSN_UID (PREV_INSN (loop->first->head)));
10049 if (! NEXT_INSN (loop->last->end)
10050 || GET_CODE (NEXT_INSN (loop->last->end)) != NOTE
10051 || NOTE_LINE_NUMBER (NEXT_INSN (loop->last->end))
10052 != NOTE_INSN_LOOP_END)
10053 fprintf (file, ";; No NOTE_INSN_LOOP_END at %d\n",
10054 INSN_UID (NEXT_INSN (loop->last->end)));
10059 ";; start %d (%d), cont dom %d (%d), cont %d (%d), vtop %d (%d), end %d (%d)\n",
10060 LOOP_BLOCK_NUM (loop->start),
10061 LOOP_INSN_UID (loop->start),
10062 LOOP_BLOCK_NUM (loop->cont),
10063 LOOP_INSN_UID (loop->cont),
10064 LOOP_BLOCK_NUM (loop->cont),
10065 LOOP_INSN_UID (loop->cont),
10066 LOOP_BLOCK_NUM (loop->vtop),
10067 LOOP_INSN_UID (loop->vtop),
10068 LOOP_BLOCK_NUM (loop->end),
10069 LOOP_INSN_UID (loop->end));
10070 fprintf (file, ";; top %d (%d), scan start %d (%d)\n",
10071 LOOP_BLOCK_NUM (loop->top),
10072 LOOP_INSN_UID (loop->top) ,
10073 LOOP_BLOCK_NUM (loop->scan_start),
10074 LOOP_INSN_UID (loop->scan_start));
10075 fprintf (file, ";; exit_count %d", loop->exit_count);
10076 if (loop->exit_count)
10078 fputs (", labels:", file);
10079 for (label = loop->exit_labels; label; label = LABEL_NEXTREF (label))
10081 fprintf (file, " %d ",
10082 LOOP_INSN_UID (XEXP (label, 0)));
10085 fputs ("\n", file);
10087 /* This can happen when a marked loop appears as two nested loops,
10088 say from while (a || b) {}. The inner loop won't match
10089 the loop markers but the outer one will. */
10090 if (LOOP_BLOCK_NUM (loop->cont) != loop->latch->index)
10091 fprintf (file, ";; NOTE_INSN_LOOP_CONT not in loop latch\n");
10096 /* Call this function from the debugger to dump LOOP. */
10100 const struct loop *loop;
10102 flow_loop_dump (loop, stderr, loop_dump_aux, 1);