1 /* Perform various loop optimizations, including strength reduction.
2 Copyright (C) 1987, 1988, 1989, 1991, 1992, 1993, 1994, 1995, 1996, 1997,
3 1998, 1999, 2000 Free Software Foundation, Inc.
5 This file is part of GNU CC.
7 GNU CC is free software; you can redistribute it and/or modify
8 it under the terms of the GNU General Public License as published by
9 the Free Software Foundation; either version 2, or (at your option)
12 GNU CC is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License for more details.
17 You should have received a copy of the GNU General Public License
18 along with GNU CC; see the file COPYING. If not, write to
19 the Free Software Foundation, 59 Temple Place - Suite 330,
20 Boston, MA 02111-1307, USA. */
23 /* This is the loop optimization pass of the compiler.
24 It finds invariant computations within loops and moves them
25 to the beginning of the loop. Then it identifies basic and
26 general induction variables. Strength reduction is applied to the general
27 induction variables, and induction variable elimination is applied to
28 the basic induction variables.
30 It also finds cases where
31 a register is set within the loop by zero-extending a narrower value
32 and changes these to zero the entire register once before the loop
33 and merely copy the low part within the loop.
35 Most of the complexity is in heuristics to decide when it is worth
36 while to do these things. */
45 #include "basic-block.h"
46 #include "insn-config.h"
47 #include "insn-flags.h"
49 #include "hard-reg-set.h"
58 /* Vector mapping INSN_UIDs to luids.
59 The luids are like uids but increase monotonically always.
60 We use them to see whether a jump comes from outside a given loop. */
64 /* Indexed by INSN_UID, contains the ordinal giving the (innermost) loop
65 number the insn is contained in. */
67 struct loop **uid_loop;
69 /* 1 + largest uid of any insn. */
73 /* 1 + luid of last insn. */
77 /* Number of loops detected in current function. Used as index to the
80 static int max_loop_num;
82 /* Indexed by register number, contains the number of times the reg
83 is set during the loop being scanned.
84 During code motion, a negative value indicates a reg that has been
85 made a candidate; in particular -2 means that it is an candidate that
86 we know is equal to a constant and -1 means that it is an candidate
87 not known equal to a constant.
88 After code motion, regs moved have 0 (which is accurate now)
89 while the failed candidates have the original number of times set.
91 Therefore, at all times, == 0 indicates an invariant register;
92 < 0 a conditionally invariant one. */
94 static varray_type set_in_loop;
96 /* Original value of set_in_loop; same except that this value
97 is not set negative for a reg whose sets have been made candidates
98 and not set to 0 for a reg that is moved. */
100 static varray_type n_times_set;
102 /* Index by register number, 1 indicates that the register
103 cannot be moved or strength reduced. */
105 static varray_type may_not_optimize;
107 /* Contains the insn in which a register was used if it was used
108 exactly once; contains const0_rtx if it was used more than once. */
110 static varray_type reg_single_usage;
112 /* Nonzero means reg N has already been moved out of one loop.
113 This reduces the desire to move it out of another. */
115 static char *moved_once;
117 /* List of MEMs that are stored in this loop. */
119 static rtx loop_store_mems;
121 /* The insn where the first of these was found. */
122 static rtx first_loop_store_insn;
124 typedef struct loop_mem_info {
125 rtx mem; /* The MEM itself. */
126 rtx reg; /* Corresponding pseudo, if any. */
127 int optimize; /* Nonzero if we can optimize access to this MEM. */
130 /* Array of MEMs that are used (read or written) in this loop, but
131 cannot be aliased by anything in this loop, except perhaps
132 themselves. In other words, if loop_mems[i] is altered during the
133 loop, it is altered by an expression that is rtx_equal_p to it. */
135 static loop_mem_info *loop_mems;
137 /* The index of the next available slot in LOOP_MEMS. */
139 static int loop_mems_idx;
141 /* The number of elements allocated in LOOP_MEMs. */
143 static int loop_mems_allocated;
145 /* Nonzero if we don't know what MEMs were changed in the current
146 loop. This happens if the loop contains a call (in which case
147 `loop_info->has_call' will also be set) or if we store into more
148 than NUM_STORES MEMs. */
150 static int unknown_address_altered;
152 /* The above doesn't count any readonly memory locations that are stored.
155 static int unknown_constant_address_altered;
157 /* Count of movable (i.e. invariant) instructions discovered in the loop. */
158 static int num_movables;
160 /* Count of memory write instructions discovered in the loop. */
161 static int num_mem_sets;
163 /* Bound on pseudo register number before loop optimization.
164 A pseudo has valid regscan info if its number is < max_reg_before_loop. */
165 unsigned int max_reg_before_loop;
167 /* The value to pass to the next call of reg_scan_update. */
168 static int loop_max_reg;
170 /* This obstack is used in product_cheap_p to allocate its rtl. It
171 may call gen_reg_rtx which, in turn, may reallocate regno_reg_rtx.
172 If we used the same obstack that it did, we would be deallocating
175 static struct obstack temp_obstack;
177 /* This is where the pointer to the obstack being used for RTL is stored. */
179 extern struct obstack *rtl_obstack;
181 #define obstack_chunk_alloc xmalloc
182 #define obstack_chunk_free free
184 /* During the analysis of a loop, a chain of `struct movable's
185 is made to record all the movable insns found.
186 Then the entire chain can be scanned to decide which to move. */
190 rtx insn; /* A movable insn */
191 rtx set_src; /* The expression this reg is set from. */
192 rtx set_dest; /* The destination of this SET. */
193 rtx dependencies; /* When INSN is libcall, this is an EXPR_LIST
194 of any registers used within the LIBCALL. */
195 int consec; /* Number of consecutive following insns
196 that must be moved with this one. */
197 unsigned int regno; /* The register it sets */
198 short lifetime; /* lifetime of that register;
199 may be adjusted when matching movables
200 that load the same value are found. */
201 short savings; /* Number of insns we can move for this reg,
202 including other movables that force this
203 or match this one. */
204 unsigned int cond : 1; /* 1 if only conditionally movable */
205 unsigned int force : 1; /* 1 means MUST move this insn */
206 unsigned int global : 1; /* 1 means reg is live outside this loop */
207 /* If PARTIAL is 1, GLOBAL means something different:
208 that the reg is live outside the range from where it is set
209 to the following label. */
210 unsigned int done : 1; /* 1 inhibits further processing of this */
212 unsigned int partial : 1; /* 1 means this reg is used for zero-extending.
213 In particular, moving it does not make it
215 unsigned int move_insn : 1; /* 1 means that we call emit_move_insn to
216 load SRC, rather than copying INSN. */
217 unsigned int move_insn_first:1;/* Same as above, if this is necessary for the
218 first insn of a consecutive sets group. */
219 unsigned int is_equiv : 1; /* 1 means a REG_EQUIV is present on INSN. */
220 enum machine_mode savemode; /* Nonzero means it is a mode for a low part
221 that we should avoid changing when clearing
222 the rest of the reg. */
223 struct movable *match; /* First entry for same value */
224 struct movable *forces; /* An insn that must be moved if this is */
225 struct movable *next;
228 static struct movable *the_movables;
230 FILE *loop_dump_stream;
232 /* Forward declarations. */
234 static void verify_dominator PARAMS ((struct loop *));
235 static void find_and_verify_loops PARAMS ((rtx, struct loops *));
236 static void mark_loop_jump PARAMS ((rtx, struct loop *));
237 static void prescan_loop PARAMS ((struct loop *));
238 static int reg_in_basic_block_p PARAMS ((rtx, rtx));
239 static int consec_sets_invariant_p PARAMS ((const struct loop *,
241 static int labels_in_range_p PARAMS ((rtx, int));
242 static void count_one_set PARAMS ((rtx, rtx, varray_type, rtx *));
244 static void count_loop_regs_set PARAMS ((rtx, rtx, varray_type, varray_type,
246 static void note_addr_stored PARAMS ((rtx, rtx, void *));
247 static void note_set_pseudo_multiple_uses PARAMS ((rtx, rtx, void *));
248 static int loop_reg_used_before_p PARAMS ((const struct loop *, rtx, rtx));
249 static void scan_loop PARAMS ((struct loop*, int, 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, 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 **, int *));
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 *));
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 int replace_label PARAMS ((rtx *, void *));
311 static rtx check_insn_for_givs PARAMS((struct loop *, rtx, int, int));
312 static rtx check_insn_for_bivs PARAMS((struct loop *, rtx, int, int));
314 typedef struct rtx_and_int {
319 typedef struct rtx_pair {
324 /* Nonzero iff INSN is between START and END, inclusive. */
325 #define INSN_IN_RANGE_P(INSN, START, END) \
326 (INSN_UID (INSN) < max_uid_for_loop \
327 && INSN_LUID (INSN) >= INSN_LUID (START) \
328 && INSN_LUID (INSN) <= INSN_LUID (END))
330 #ifdef HAVE_decrement_and_branch_on_count
331 /* Test whether BCT applicable and safe. */
332 static void insert_bct PARAMS ((struct loop *));
334 /* Auxiliary function that inserts the BCT pattern into the loop. */
335 static void instrument_loop_bct PARAMS ((rtx, rtx, rtx));
336 #endif /* HAVE_decrement_and_branch_on_count */
338 /* Indirect_jump_in_function is computed once per function. */
339 int indirect_jump_in_function = 0;
340 static int indirect_jump_in_function_p PARAMS ((rtx));
342 static int compute_luids PARAMS ((rtx, rtx, int));
344 static int biv_elimination_giv_has_0_offset PARAMS ((struct induction *,
345 struct induction *, rtx));
347 /* Relative gain of eliminating various kinds of operations. */
350 static int shift_cost;
351 static int mult_cost;
354 /* Benefit penalty, if a giv is not replaceable, i.e. must emit an insn to
355 copy the value of the strength reduced giv to its original register. */
356 static int copy_cost;
358 /* Cost of using a register, to normalize the benefits of a giv. */
359 static int reg_address_cost;
365 char *free_point = (char *) oballoc (1);
366 rtx reg = gen_rtx_REG (word_mode, LAST_VIRTUAL_REGISTER + 1);
368 add_cost = rtx_cost (gen_rtx_PLUS (word_mode, reg, reg), SET);
371 reg_address_cost = ADDRESS_COST (reg);
373 reg_address_cost = rtx_cost (reg, MEM);
376 /* We multiply by 2 to reconcile the difference in scale between
377 these two ways of computing costs. Otherwise the cost of a copy
378 will be far less than the cost of an add. */
382 /* Free the objects we just allocated. */
385 /* Initialize the obstack used for rtl in product_cheap_p. */
386 gcc_obstack_init (&temp_obstack);
389 /* Compute the mapping from uids to luids.
390 LUIDs are numbers assigned to insns, like uids,
391 except that luids increase monotonically through the code.
392 Start at insn START and stop just before END. Assign LUIDs
393 starting with PREV_LUID + 1. Return the last assigned LUID + 1. */
395 compute_luids (start, end, prev_luid)
402 for (insn = start, i = prev_luid; insn != end; insn = NEXT_INSN (insn))
404 if (INSN_UID (insn) >= max_uid_for_loop)
406 /* Don't assign luids to line-number NOTEs, so that the distance in
407 luids between two insns is not affected by -g. */
408 if (GET_CODE (insn) != NOTE
409 || NOTE_LINE_NUMBER (insn) <= 0)
410 uid_luid[INSN_UID (insn)] = ++i;
412 /* Give a line number note the same luid as preceding insn. */
413 uid_luid[INSN_UID (insn)] = i;
418 /* Entry point of this file. Perform loop optimization
419 on the current function. F is the first insn of the function
420 and DUMPFILE is a stream for output of a trace of actions taken
421 (or 0 if none should be output). */
424 loop_optimize (f, dumpfile, unroll_p, bct_p)
425 /* f is the first instruction of a chain of insns for one function */
432 struct loops loops_data;
433 struct loops *loops = &loops_data;
434 struct loop_info *loops_info;
436 loop_dump_stream = dumpfile;
438 init_recog_no_volatile ();
440 max_reg_before_loop = max_reg_num ();
441 loop_max_reg = max_reg_before_loop;
445 /* Count the number of loops. */
448 for (insn = f; insn; insn = NEXT_INSN (insn))
450 if (GET_CODE (insn) == NOTE
451 && NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG)
455 /* Don't waste time if no loops. */
456 if (max_loop_num == 0)
459 loops->num = max_loop_num;
461 moved_once = (char *) xcalloc (max_reg_before_loop, sizeof (char));
463 /* Get size to use for tables indexed by uids.
464 Leave some space for labels allocated by find_and_verify_loops. */
465 max_uid_for_loop = get_max_uid () + 1 + max_loop_num * 32;
467 uid_luid = (int *) xcalloc (max_uid_for_loop, sizeof (int));
468 uid_loop = (struct loop **) xcalloc (max_uid_for_loop,
469 sizeof (struct loop *));
471 /* Allocate storage for array of loops. */
472 loops->array = (struct loop *)
473 xcalloc (loops->num, sizeof (struct loop));
475 /* Find and process each loop.
476 First, find them, and record them in order of their beginnings. */
477 find_and_verify_loops (f, loops);
479 /* Allocate and initialize auxiliary loop information. */
480 loops_info = xcalloc (loops->num, sizeof (struct loop_info));
481 for (i = 0; i < loops->num; i++)
482 loops->array[i].aux = loops_info + i;
484 /* Now find all register lifetimes. This must be done after
485 find_and_verify_loops, because it might reorder the insns in the
487 reg_scan (f, max_reg_before_loop, 1);
489 /* This must occur after reg_scan so that registers created by gcse
490 will have entries in the register tables.
492 We could have added a call to reg_scan after gcse_main in toplev.c,
493 but moving this call to init_alias_analysis is more efficient. */
494 init_alias_analysis ();
496 /* See if we went too far. Note that get_max_uid already returns
497 one more that the maximum uid of all insn. */
498 if (get_max_uid () > max_uid_for_loop)
500 /* Now reset it to the actual size we need. See above. */
501 max_uid_for_loop = get_max_uid ();
503 /* find_and_verify_loops has already called compute_luids, but it
504 might have rearranged code afterwards, so we need to recompute
506 max_luid = compute_luids (f, NULL_RTX, 0);
508 /* Don't leave gaps in uid_luid for insns that have been
509 deleted. It is possible that the first or last insn
510 using some register has been deleted by cross-jumping.
511 Make sure that uid_luid for that former insn's uid
512 points to the general area where that insn used to be. */
513 for (i = 0; i < max_uid_for_loop; i++)
515 uid_luid[0] = uid_luid[i];
516 if (uid_luid[0] != 0)
519 for (i = 0; i < max_uid_for_loop; i++)
520 if (uid_luid[i] == 0)
521 uid_luid[i] = uid_luid[i - 1];
523 /* Determine if the function has indirect jump. On some systems
524 this prevents low overhead loop instructions from being used. */
525 indirect_jump_in_function = indirect_jump_in_function_p (f);
527 /* Now scan the loops, last ones first, since this means inner ones are done
528 before outer ones. */
529 for (i = max_loop_num - 1; i >= 0; i--)
531 struct loop *loop = &loops->array[i];
533 if (! loop->invalid && loop->end)
534 scan_loop (loop, unroll_p, bct_p);
537 /* If there were lexical blocks inside the loop, they have been
538 replicated. We will now have more than one NOTE_INSN_BLOCK_BEG
539 and NOTE_INSN_BLOCK_END for each such block. We must duplicate
540 the BLOCKs as well. */
541 if (write_symbols != NO_DEBUG)
544 end_alias_analysis ();
554 /* Returns the next insn, in execution order, after INSN. START and
555 END are the NOTE_INSN_LOOP_BEG and NOTE_INSN_LOOP_END for the loop,
556 respectively. LOOP->TOP, if non-NULL, is the top of the loop in the
557 insn-stream; it is used with loops that are entered near the
561 next_insn_in_loop (loop, insn)
562 const struct loop *loop;
565 insn = NEXT_INSN (insn);
567 if (insn == loop->end)
570 /* Go to the top of the loop, and continue there. */
577 if (insn == loop->scan_start)
584 /* Optimize one loop described by LOOP. */
586 /* ??? Could also move memory writes out of loops if the destination address
587 is invariant, the source is invariant, the memory write is not volatile,
588 and if we can prove that no read inside the loop can read this address
589 before the write occurs. If there is a read of this address after the
590 write, then we can also mark the memory read as invariant. */
593 scan_loop (loop, unroll_p, bct_p)
598 rtx loop_start = loop->start;
599 rtx loop_end = loop->end;
600 /* Additional information about the current loop being processed
601 that is used to compute the number of loop iterations for loop
602 unrolling and doloop optimization. */
603 struct loop_info *loop_info = LOOP_INFO (loop);
605 /* 1 if we are scanning insns that could be executed zero times. */
607 /* 1 if we are scanning insns that might never be executed
608 due to a subroutine call which might exit before they are reached. */
610 /* Jump insn that enters the loop, or 0 if control drops in. */
611 rtx loop_entry_jump = 0;
612 /* Number of insns in the loop. */
616 rtx temp, update_start, update_end;
617 /* The SET from an insn, if it is the only SET in the insn. */
619 /* Chain describing insns movable in current loop. */
620 struct movable *movables = 0;
621 /* Last element in `movables' -- so we can add elements at the end. */
622 struct movable *last_movable = 0;
623 /* Ratio of extra register life span we can justify
624 for saving an instruction. More if loop doesn't call subroutines
625 since in that case saving an insn makes more difference
626 and more registers are available. */
628 /* Nonzero if we are scanning instructions in a sub-loop. */
634 /* Determine whether this loop starts with a jump down to a test at
635 the end. This will occur for a small number of loops with a test
636 that is too complex to duplicate in front of the loop.
638 We search for the first insn or label in the loop, skipping NOTEs.
639 However, we must be careful not to skip past a NOTE_INSN_LOOP_BEG
640 (because we might have a loop executed only once that contains a
641 loop which starts with a jump to its exit test) or a NOTE_INSN_LOOP_END
642 (in case we have a degenerate loop).
644 Note that if we mistakenly think that a loop is entered at the top
645 when, in fact, it is entered at the exit test, the only effect will be
646 slightly poorer optimization. Making the opposite error can generate
647 incorrect code. Since very few loops now start with a jump to the
648 exit test, the code here to detect that case is very conservative. */
650 for (p = NEXT_INSN (loop_start);
652 && GET_CODE (p) != CODE_LABEL && GET_RTX_CLASS (GET_CODE (p)) != 'i'
653 && (GET_CODE (p) != NOTE
654 || (NOTE_LINE_NUMBER (p) != NOTE_INSN_LOOP_BEG
655 && NOTE_LINE_NUMBER (p) != NOTE_INSN_LOOP_END));
659 loop->scan_start = p;
661 /* Set up variables describing this loop. */
663 threshold = (loop_info->has_call ? 1 : 2) * (1 + n_non_fixed_regs);
665 /* If loop has a jump before the first label,
666 the true entry is the target of that jump.
667 Start scan from there.
668 But record in LOOP->TOP the place where the end-test jumps
669 back to so we can scan that after the end of the loop. */
670 if (GET_CODE (p) == JUMP_INSN)
674 /* Loop entry must be unconditional jump (and not a RETURN) */
676 && JUMP_LABEL (p) != 0
677 /* Check to see whether the jump actually
678 jumps out of the loop (meaning it's no loop).
679 This case can happen for things like
680 do {..} while (0). If this label was generated previously
681 by loop, we can't tell anything about it and have to reject
683 && INSN_IN_RANGE_P (JUMP_LABEL (p), loop_start, loop_end))
685 loop->top = next_label (loop->scan_start);
686 loop->scan_start = JUMP_LABEL (p);
690 /* If LOOP->SCAN_START was an insn created by loop, we don't know its luid
691 as required by loop_reg_used_before_p. So skip such loops. (This
692 test may never be true, but it's best to play it safe.)
694 Also, skip loops where we do not start scanning at a label. This
695 test also rejects loops starting with a JUMP_INSN that failed the
698 if (INSN_UID (loop->scan_start) >= max_uid_for_loop
699 || GET_CODE (loop->scan_start) != CODE_LABEL)
701 if (loop_dump_stream)
702 fprintf (loop_dump_stream, "\nLoop from %d to %d is phony.\n\n",
703 INSN_UID (loop_start), INSN_UID (loop_end));
707 /* Count number of times each reg is set during this loop.
708 Set VARRAY_CHAR (may_not_optimize, I) if it is not safe to move out
709 the setting of register I. Set VARRAY_RTX (reg_single_usage, I). */
711 /* Allocate extra space for REGS that might be created by
712 load_mems. We allocate a little extra slop as well, in the hopes
713 that even after the moving of movables creates some new registers
714 we won't have to reallocate these arrays. However, we do grow
715 the arrays, if necessary, in load_mems_recount_loop_regs_set. */
716 nregs = max_reg_num () + loop_mems_idx + 16;
717 VARRAY_INT_INIT (set_in_loop, nregs, "set_in_loop");
718 VARRAY_INT_INIT (n_times_set, nregs, "n_times_set");
719 VARRAY_CHAR_INIT (may_not_optimize, nregs, "may_not_optimize");
720 VARRAY_RTX_INIT (reg_single_usage, nregs, "reg_single_usage");
722 count_loop_regs_set (loop->top ? loop->top : loop->start, loop->end,
723 may_not_optimize, reg_single_usage, &insn_count, nregs);
725 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
727 VARRAY_CHAR (may_not_optimize, i) = 1;
728 VARRAY_INT (set_in_loop, i) = 1;
731 #ifdef AVOID_CCMODE_COPIES
732 /* Don't try to move insns which set CC registers if we should not
733 create CCmode register copies. */
734 for (i = max_reg_num () - 1; i >= FIRST_PSEUDO_REGISTER; i--)
735 if (GET_MODE_CLASS (GET_MODE (regno_reg_rtx[i])) == MODE_CC)
736 VARRAY_CHAR (may_not_optimize, i) = 1;
739 bcopy ((char *) &set_in_loop->data,
740 (char *) &n_times_set->data, nregs * sizeof (int));
742 if (loop_dump_stream)
744 fprintf (loop_dump_stream, "\nLoop from %d to %d: %d real insns.\n",
745 INSN_UID (loop_start), INSN_UID (loop_end), insn_count);
747 fprintf (loop_dump_stream, "Continue at insn %d.\n",
748 INSN_UID (loop->cont));
751 /* Scan through the loop finding insns that are safe to move.
752 Set set_in_loop negative for the reg being set, so that
753 this reg will be considered invariant for subsequent insns.
754 We consider whether subsequent insns use the reg
755 in deciding whether it is worth actually moving.
757 MAYBE_NEVER is nonzero if we have passed a conditional jump insn
758 and therefore it is possible that the insns we are scanning
759 would never be executed. At such times, we must make sure
760 that it is safe to execute the insn once instead of zero times.
761 When MAYBE_NEVER is 0, all insns will be executed at least once
762 so that is not a problem. */
764 for (p = next_insn_in_loop (loop, loop->scan_start);
766 p = next_insn_in_loop (loop, p))
768 if (GET_RTX_CLASS (GET_CODE (p)) == 'i'
769 && find_reg_note (p, REG_LIBCALL, NULL_RTX))
771 else if (GET_RTX_CLASS (GET_CODE (p)) == 'i'
772 && find_reg_note (p, REG_RETVAL, NULL_RTX))
775 if (GET_CODE (p) == INSN
776 && (set = single_set (p))
777 && GET_CODE (SET_DEST (set)) == REG
778 && ! VARRAY_CHAR (may_not_optimize, REGNO (SET_DEST (set))))
783 rtx src = SET_SRC (set);
784 rtx dependencies = 0;
786 /* Figure out what to use as a source of this insn. If a REG_EQUIV
787 note is given or if a REG_EQUAL note with a constant operand is
788 specified, use it as the source and mark that we should move
789 this insn by calling emit_move_insn rather that duplicating the
792 Otherwise, only use the REG_EQUAL contents if a REG_RETVAL note
794 temp = find_reg_note (p, REG_EQUIV, NULL_RTX);
796 src = XEXP (temp, 0), move_insn = 1;
799 temp = find_reg_note (p, REG_EQUAL, NULL_RTX);
800 if (temp && CONSTANT_P (XEXP (temp, 0)))
801 src = XEXP (temp, 0), move_insn = 1;
802 if (temp && find_reg_note (p, REG_RETVAL, NULL_RTX))
804 src = XEXP (temp, 0);
805 /* A libcall block can use regs that don't appear in
806 the equivalent expression. To move the libcall,
807 we must move those regs too. */
808 dependencies = libcall_other_reg (p, src);
812 /* Don't try to optimize a register that was made
813 by loop-optimization for an inner loop.
814 We don't know its life-span, so we can't compute the benefit. */
815 if (REGNO (SET_DEST (set)) >= max_reg_before_loop)
817 else if (/* The register is used in basic blocks other
818 than the one where it is set (meaning that
819 something after this point in the loop might
820 depend on its value before the set). */
821 ! reg_in_basic_block_p (p, SET_DEST (set))
822 /* And the set is not guaranteed to be executed one
823 the loop starts, or the value before the set is
824 needed before the set occurs...
826 ??? Note we have quadratic behaviour here, mitigated
827 by the fact that the previous test will often fail for
828 large loops. Rather than re-scanning the entire loop
829 each time for register usage, we should build tables
830 of the register usage and use them here instead. */
832 || loop_reg_used_before_p (loop, set, p)))
833 /* It is unsafe to move the set.
835 This code used to consider it OK to move a set of a variable
836 which was not created by the user and not used in an exit test.
837 That behavior is incorrect and was removed. */
839 else if ((tem = loop_invariant_p (loop, src))
840 && (dependencies == 0
841 || (tem2 = loop_invariant_p (loop, dependencies)) != 0)
842 && (VARRAY_INT (set_in_loop,
843 REGNO (SET_DEST (set))) == 1
845 = consec_sets_invariant_p
846 (loop, SET_DEST (set),
847 VARRAY_INT (set_in_loop, REGNO (SET_DEST (set))),
849 /* If the insn can cause a trap (such as divide by zero),
850 can't move it unless it's guaranteed to be executed
851 once loop is entered. Even a function call might
852 prevent the trap insn from being reached
853 (since it might exit!) */
854 && ! ((maybe_never || call_passed)
855 && may_trap_p (src)))
857 register struct movable *m;
858 register int regno = REGNO (SET_DEST (set));
860 /* A potential lossage is where we have a case where two insns
861 can be combined as long as they are both in the loop, but
862 we move one of them outside the loop. For large loops,
863 this can lose. The most common case of this is the address
864 of a function being called.
866 Therefore, if this register is marked as being used exactly
867 once if we are in a loop with calls (a "large loop"), see if
868 we can replace the usage of this register with the source
869 of this SET. If we can, delete this insn.
871 Don't do this if P has a REG_RETVAL note or if we have
872 SMALL_REGISTER_CLASSES and SET_SRC is a hard register. */
874 if (loop_info->has_call
875 && VARRAY_RTX (reg_single_usage, regno) != 0
876 && VARRAY_RTX (reg_single_usage, regno) != const0_rtx
877 && REGNO_FIRST_UID (regno) == INSN_UID (p)
878 && (REGNO_LAST_UID (regno)
879 == INSN_UID (VARRAY_RTX (reg_single_usage, regno)))
880 && VARRAY_INT (set_in_loop, regno) == 1
881 && ! side_effects_p (SET_SRC (set))
882 && ! find_reg_note (p, REG_RETVAL, NULL_RTX)
883 && (! SMALL_REGISTER_CLASSES
884 || (! (GET_CODE (SET_SRC (set)) == REG
885 && REGNO (SET_SRC (set)) < FIRST_PSEUDO_REGISTER)))
886 /* This test is not redundant; SET_SRC (set) might be
887 a call-clobbered register and the life of REGNO
888 might span a call. */
889 && ! modified_between_p (SET_SRC (set), p,
891 (reg_single_usage, regno))
892 && no_labels_between_p (p, VARRAY_RTX (reg_single_usage, regno))
893 && validate_replace_rtx (SET_DEST (set), SET_SRC (set),
895 (reg_single_usage, regno)))
897 /* Replace any usage in a REG_EQUAL note. Must copy the
898 new source, so that we don't get rtx sharing between the
899 SET_SOURCE and REG_NOTES of insn p. */
900 REG_NOTES (VARRAY_RTX (reg_single_usage, regno))
901 = replace_rtx (REG_NOTES (VARRAY_RTX
902 (reg_single_usage, regno)),
903 SET_DEST (set), copy_rtx (SET_SRC (set)));
906 NOTE_LINE_NUMBER (p) = NOTE_INSN_DELETED;
907 NOTE_SOURCE_FILE (p) = 0;
908 VARRAY_INT (set_in_loop, regno) = 0;
912 m = (struct movable *) alloca (sizeof (struct movable));
916 m->dependencies = dependencies;
917 m->set_dest = SET_DEST (set);
919 m->consec = VARRAY_INT (set_in_loop,
920 REGNO (SET_DEST (set))) - 1;
924 m->move_insn = move_insn;
925 m->move_insn_first = 0;
926 m->is_equiv = (find_reg_note (p, REG_EQUIV, NULL_RTX) != 0);
927 m->savemode = VOIDmode;
929 /* Set M->cond if either loop_invariant_p
930 or consec_sets_invariant_p returned 2
931 (only conditionally invariant). */
932 m->cond = ((tem | tem1 | tem2) > 1);
933 m->global = (uid_luid[REGNO_LAST_UID (regno)]
934 > INSN_LUID (loop_end)
935 || uid_luid[REGNO_FIRST_UID (regno)] < INSN_LUID (loop_start));
937 m->lifetime = (uid_luid[REGNO_LAST_UID (regno)]
938 - uid_luid[REGNO_FIRST_UID (regno)]);
939 m->savings = VARRAY_INT (n_times_set, regno);
940 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
941 m->savings += libcall_benefit (p);
942 VARRAY_INT (set_in_loop, regno) = move_insn ? -2 : -1;
943 /* Add M to the end of the chain MOVABLES. */
947 last_movable->next = m;
952 /* It is possible for the first instruction to have a
953 REG_EQUAL note but a non-invariant SET_SRC, so we must
954 remember the status of the first instruction in case
955 the last instruction doesn't have a REG_EQUAL note. */
956 m->move_insn_first = m->move_insn;
958 /* Skip this insn, not checking REG_LIBCALL notes. */
959 p = next_nonnote_insn (p);
960 /* Skip the consecutive insns, if there are any. */
961 p = skip_consec_insns (p, m->consec);
962 /* Back up to the last insn of the consecutive group. */
963 p = prev_nonnote_insn (p);
965 /* We must now reset m->move_insn, m->is_equiv, and possibly
966 m->set_src to correspond to the effects of all the
968 temp = find_reg_note (p, REG_EQUIV, NULL_RTX);
970 m->set_src = XEXP (temp, 0), m->move_insn = 1;
973 temp = find_reg_note (p, REG_EQUAL, NULL_RTX);
974 if (temp && CONSTANT_P (XEXP (temp, 0)))
975 m->set_src = XEXP (temp, 0), m->move_insn = 1;
980 m->is_equiv = (find_reg_note (p, REG_EQUIV, NULL_RTX) != 0);
983 /* If this register is always set within a STRICT_LOW_PART
984 or set to zero, then its high bytes are constant.
985 So clear them outside the loop and within the loop
986 just load the low bytes.
987 We must check that the machine has an instruction to do so.
988 Also, if the value loaded into the register
989 depends on the same register, this cannot be done. */
990 else if (SET_SRC (set) == const0_rtx
991 && GET_CODE (NEXT_INSN (p)) == INSN
992 && (set1 = single_set (NEXT_INSN (p)))
993 && GET_CODE (set1) == SET
994 && (GET_CODE (SET_DEST (set1)) == STRICT_LOW_PART)
995 && (GET_CODE (XEXP (SET_DEST (set1), 0)) == SUBREG)
996 && (SUBREG_REG (XEXP (SET_DEST (set1), 0))
998 && !reg_mentioned_p (SET_DEST (set), SET_SRC (set1)))
1000 register int regno = REGNO (SET_DEST (set));
1001 if (VARRAY_INT (set_in_loop, regno) == 2)
1003 register struct movable *m;
1004 m = (struct movable *) alloca (sizeof (struct movable));
1007 m->set_dest = SET_DEST (set);
1008 m->dependencies = 0;
1014 m->move_insn_first = 0;
1016 /* If the insn may not be executed on some cycles,
1017 we can't clear the whole reg; clear just high part.
1018 Not even if the reg is used only within this loop.
1025 Clearing x before the inner loop could clobber a value
1026 being saved from the last time around the outer loop.
1027 However, if the reg is not used outside this loop
1028 and all uses of the register are in the same
1029 basic block as the store, there is no problem.
1031 If this insn was made by loop, we don't know its
1032 INSN_LUID and hence must make a conservative
1034 m->global = (INSN_UID (p) >= max_uid_for_loop
1035 || (uid_luid[REGNO_LAST_UID (regno)]
1036 > INSN_LUID (loop_end))
1037 || (uid_luid[REGNO_FIRST_UID (regno)]
1039 || (labels_in_range_p
1040 (p, uid_luid[REGNO_FIRST_UID (regno)])));
1041 if (maybe_never && m->global)
1042 m->savemode = GET_MODE (SET_SRC (set1));
1044 m->savemode = VOIDmode;
1048 m->lifetime = (uid_luid[REGNO_LAST_UID (regno)]
1049 - uid_luid[REGNO_FIRST_UID (regno)]);
1051 VARRAY_INT (set_in_loop, regno) = -1;
1052 /* Add M to the end of the chain MOVABLES. */
1056 last_movable->next = m;
1061 /* Past a call insn, we get to insns which might not be executed
1062 because the call might exit. This matters for insns that trap.
1063 Call insns inside a REG_LIBCALL/REG_RETVAL block always return,
1064 so they don't count. */
1065 else if (GET_CODE (p) == CALL_INSN && ! in_libcall)
1067 /* Past a label or a jump, we get to insns for which we
1068 can't count on whether or how many times they will be
1069 executed during each iteration. Therefore, we can
1070 only move out sets of trivial variables
1071 (those not used after the loop). */
1072 /* Similar code appears twice in strength_reduce. */
1073 else if ((GET_CODE (p) == CODE_LABEL || GET_CODE (p) == JUMP_INSN)
1074 /* If we enter the loop in the middle, and scan around to the
1075 beginning, don't set maybe_never for that. This must be an
1076 unconditional jump, otherwise the code at the top of the
1077 loop might never be executed. Unconditional jumps are
1078 followed a by barrier then loop end. */
1079 && ! (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) == loop->top
1080 && NEXT_INSN (NEXT_INSN (p)) == loop_end
1081 && simplejump_p (p)))
1083 else if (GET_CODE (p) == NOTE)
1085 /* At the virtual top of a converted loop, insns are again known to
1086 be executed: logically, the loop begins here even though the exit
1087 code has been duplicated. */
1088 if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP && loop_depth == 0)
1089 maybe_never = call_passed = 0;
1090 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
1092 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END)
1097 /* If one movable subsumes another, ignore that other. */
1099 ignore_some_movables (movables);
1101 /* For each movable insn, see if the reg that it loads
1102 leads when it dies right into another conditionally movable insn.
1103 If so, record that the second insn "forces" the first one,
1104 since the second can be moved only if the first is. */
1106 force_movables (movables);
1108 /* See if there are multiple movable insns that load the same value.
1109 If there are, make all but the first point at the first one
1110 through the `match' field, and add the priorities of them
1111 all together as the priority of the first. */
1113 combine_movables (movables, nregs);
1115 /* Now consider each movable insn to decide whether it is worth moving.
1116 Store 0 in set_in_loop for each reg that is moved.
1118 Generally this increases code size, so do not move moveables when
1119 optimizing for code size. */
1121 if (! optimize_size)
1122 move_movables (loop, movables, threshold, insn_count, nregs);
1124 /* Now candidates that still are negative are those not moved.
1125 Change set_in_loop to indicate that those are not actually invariant. */
1126 for (i = 0; i < nregs; i++)
1127 if (VARRAY_INT (set_in_loop, i) < 0)
1128 VARRAY_INT (set_in_loop, i) = VARRAY_INT (n_times_set, i);
1130 /* Now that we've moved some things out of the loop, we might be able to
1131 hoist even more memory references. */
1132 load_mems_and_recount_loop_regs_set (loop, &insn_count);
1134 for (update_start = loop_start;
1135 PREV_INSN (update_start)
1136 && GET_CODE (PREV_INSN (update_start)) != CODE_LABEL;
1137 update_start = PREV_INSN (update_start))
1139 update_end = NEXT_INSN (loop_end);
1141 reg_scan_update (update_start, update_end, loop_max_reg);
1142 loop_max_reg = max_reg_num ();
1144 if (flag_strength_reduce)
1146 the_movables = movables;
1147 strength_reduce (loop, insn_count, unroll_p, bct_p);
1149 reg_scan_update (update_start, update_end, loop_max_reg);
1150 loop_max_reg = max_reg_num ();
1153 VARRAY_FREE (reg_single_usage);
1154 VARRAY_FREE (set_in_loop);
1155 VARRAY_FREE (n_times_set);
1156 VARRAY_FREE (may_not_optimize);
1159 /* Add elements to *OUTPUT to record all the pseudo-regs
1160 mentioned in IN_THIS but not mentioned in NOT_IN_THIS. */
1163 record_excess_regs (in_this, not_in_this, output)
1164 rtx in_this, not_in_this;
1171 code = GET_CODE (in_this);
1185 if (REGNO (in_this) >= FIRST_PSEUDO_REGISTER
1186 && ! reg_mentioned_p (in_this, not_in_this))
1187 *output = gen_rtx_EXPR_LIST (VOIDmode, in_this, *output);
1194 fmt = GET_RTX_FORMAT (code);
1195 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1202 for (j = 0; j < XVECLEN (in_this, i); j++)
1203 record_excess_regs (XVECEXP (in_this, i, j), not_in_this, output);
1207 record_excess_regs (XEXP (in_this, i), not_in_this, output);
1213 /* Check what regs are referred to in the libcall block ending with INSN,
1214 aside from those mentioned in the equivalent value.
1215 If there are none, return 0.
1216 If there are one or more, return an EXPR_LIST containing all of them. */
1219 libcall_other_reg (insn, equiv)
1222 rtx note = find_reg_note (insn, REG_RETVAL, NULL_RTX);
1223 rtx p = XEXP (note, 0);
1226 /* First, find all the regs used in the libcall block
1227 that are not mentioned as inputs to the result. */
1231 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
1232 || GET_CODE (p) == CALL_INSN)
1233 record_excess_regs (PATTERN (p), equiv, &output);
1240 /* Return 1 if all uses of REG
1241 are between INSN and the end of the basic block. */
1244 reg_in_basic_block_p (insn, reg)
1247 int regno = REGNO (reg);
1250 if (REGNO_FIRST_UID (regno) != INSN_UID (insn))
1253 /* Search this basic block for the already recorded last use of the reg. */
1254 for (p = insn; p; p = NEXT_INSN (p))
1256 switch (GET_CODE (p))
1263 /* Ordinary insn: if this is the last use, we win. */
1264 if (REGNO_LAST_UID (regno) == INSN_UID (p))
1269 /* Jump insn: if this is the last use, we win. */
1270 if (REGNO_LAST_UID (regno) == INSN_UID (p))
1272 /* Otherwise, it's the end of the basic block, so we lose. */
1277 /* It's the end of the basic block, so we lose. */
1285 /* The "last use" that was recorded can't be found after the first
1286 use. This can happen when the last use was deleted while
1287 processing an inner loop, this inner loop was then completely
1288 unrolled, and the outer loop is always exited after the inner loop,
1289 so that everything after the first use becomes a single basic block. */
1293 /* Compute the benefit of eliminating the insns in the block whose
1294 last insn is LAST. This may be a group of insns used to compute a
1295 value directly or can contain a library call. */
1298 libcall_benefit (last)
1304 for (insn = XEXP (find_reg_note (last, REG_RETVAL, NULL_RTX), 0);
1305 insn != last; insn = NEXT_INSN (insn))
1307 if (GET_CODE (insn) == CALL_INSN)
1308 benefit += 10; /* Assume at least this many insns in a library
1310 else if (GET_CODE (insn) == INSN
1311 && GET_CODE (PATTERN (insn)) != USE
1312 && GET_CODE (PATTERN (insn)) != CLOBBER)
1319 /* Skip COUNT insns from INSN, counting library calls as 1 insn. */
1322 skip_consec_insns (insn, count)
1326 for (; count > 0; count--)
1330 /* If first insn of libcall sequence, skip to end. */
1331 /* Do this at start of loop, since INSN is guaranteed to
1333 if (GET_CODE (insn) != NOTE
1334 && (temp = find_reg_note (insn, REG_LIBCALL, NULL_RTX)))
1335 insn = XEXP (temp, 0);
1337 do insn = NEXT_INSN (insn);
1338 while (GET_CODE (insn) == NOTE);
1344 /* Ignore any movable whose insn falls within a libcall
1345 which is part of another movable.
1346 We make use of the fact that the movable for the libcall value
1347 was made later and so appears later on the chain. */
1350 ignore_some_movables (movables)
1351 struct movable *movables;
1353 register struct movable *m, *m1;
1355 for (m = movables; m; m = m->next)
1357 /* Is this a movable for the value of a libcall? */
1358 rtx note = find_reg_note (m->insn, REG_RETVAL, NULL_RTX);
1362 /* Check for earlier movables inside that range,
1363 and mark them invalid. We cannot use LUIDs here because
1364 insns created by loop.c for prior loops don't have LUIDs.
1365 Rather than reject all such insns from movables, we just
1366 explicitly check each insn in the libcall (since invariant
1367 libcalls aren't that common). */
1368 for (insn = XEXP (note, 0); insn != m->insn; insn = NEXT_INSN (insn))
1369 for (m1 = movables; m1 != m; m1 = m1->next)
1370 if (m1->insn == insn)
1376 /* For each movable insn, see if the reg that it loads
1377 leads when it dies right into another conditionally movable insn.
1378 If so, record that the second insn "forces" the first one,
1379 since the second can be moved only if the first is. */
1382 force_movables (movables)
1383 struct movable *movables;
1385 register struct movable *m, *m1;
1386 for (m1 = movables; m1; m1 = m1->next)
1387 /* Omit this if moving just the (SET (REG) 0) of a zero-extend. */
1388 if (!m1->partial && !m1->done)
1390 int regno = m1->regno;
1391 for (m = m1->next; m; m = m->next)
1392 /* ??? Could this be a bug? What if CSE caused the
1393 register of M1 to be used after this insn?
1394 Since CSE does not update regno_last_uid,
1395 this insn M->insn might not be where it dies.
1396 But very likely this doesn't matter; what matters is
1397 that M's reg is computed from M1's reg. */
1398 if (INSN_UID (m->insn) == REGNO_LAST_UID (regno)
1401 if (m != 0 && m->set_src == m1->set_dest
1402 /* If m->consec, m->set_src isn't valid. */
1406 /* Increase the priority of the moving the first insn
1407 since it permits the second to be moved as well. */
1411 m1->lifetime += m->lifetime;
1412 m1->savings += m->savings;
1417 /* Find invariant expressions that are equal and can be combined into
1421 combine_movables (movables, nregs)
1422 struct movable *movables;
1425 register struct movable *m;
1426 char *matched_regs = (char *) xmalloc (nregs);
1427 enum machine_mode mode;
1429 /* Regs that are set more than once are not allowed to match
1430 or be matched. I'm no longer sure why not. */
1431 /* Perhaps testing m->consec_sets would be more appropriate here? */
1433 for (m = movables; m; m = m->next)
1434 if (m->match == 0 && VARRAY_INT (n_times_set, m->regno) == 1 && !m->partial)
1436 register struct movable *m1;
1437 int regno = m->regno;
1439 bzero (matched_regs, nregs);
1440 matched_regs[regno] = 1;
1442 /* We want later insns to match the first one. Don't make the first
1443 one match any later ones. So start this loop at m->next. */
1444 for (m1 = m->next; m1; m1 = m1->next)
1445 if (m != m1 && m1->match == 0 && VARRAY_INT (n_times_set, m1->regno) == 1
1446 /* A reg used outside the loop mustn't be eliminated. */
1448 /* A reg used for zero-extending mustn't be eliminated. */
1450 && (matched_regs[m1->regno]
1453 /* Can combine regs with different modes loaded from the
1454 same constant only if the modes are the same or
1455 if both are integer modes with M wider or the same
1456 width as M1. The check for integer is redundant, but
1457 safe, since the only case of differing destination
1458 modes with equal sources is when both sources are
1459 VOIDmode, i.e., CONST_INT. */
1460 (GET_MODE (m->set_dest) == GET_MODE (m1->set_dest)
1461 || (GET_MODE_CLASS (GET_MODE (m->set_dest)) == MODE_INT
1462 && GET_MODE_CLASS (GET_MODE (m1->set_dest)) == MODE_INT
1463 && (GET_MODE_BITSIZE (GET_MODE (m->set_dest))
1464 >= GET_MODE_BITSIZE (GET_MODE (m1->set_dest)))))
1465 /* See if the source of M1 says it matches M. */
1466 && ((GET_CODE (m1->set_src) == REG
1467 && matched_regs[REGNO (m1->set_src)])
1468 || rtx_equal_for_loop_p (m->set_src, m1->set_src,
1470 && ((m->dependencies == m1->dependencies)
1471 || rtx_equal_p (m->dependencies, m1->dependencies)))
1473 m->lifetime += m1->lifetime;
1474 m->savings += m1->savings;
1477 matched_regs[m1->regno] = 1;
1481 /* Now combine the regs used for zero-extension.
1482 This can be done for those not marked `global'
1483 provided their lives don't overlap. */
1485 for (mode = GET_CLASS_NARROWEST_MODE (MODE_INT); mode != VOIDmode;
1486 mode = GET_MODE_WIDER_MODE (mode))
1488 register struct movable *m0 = 0;
1490 /* Combine all the registers for extension from mode MODE.
1491 Don't combine any that are used outside this loop. */
1492 for (m = movables; m; m = m->next)
1493 if (m->partial && ! m->global
1494 && mode == GET_MODE (SET_SRC (PATTERN (NEXT_INSN (m->insn)))))
1496 register struct movable *m1;
1497 int first = uid_luid[REGNO_FIRST_UID (m->regno)];
1498 int last = uid_luid[REGNO_LAST_UID (m->regno)];
1502 /* First one: don't check for overlap, just record it. */
1507 /* Make sure they extend to the same mode.
1508 (Almost always true.) */
1509 if (GET_MODE (m->set_dest) != GET_MODE (m0->set_dest))
1512 /* We already have one: check for overlap with those
1513 already combined together. */
1514 for (m1 = movables; m1 != m; m1 = m1->next)
1515 if (m1 == m0 || (m1->partial && m1->match == m0))
1516 if (! (uid_luid[REGNO_FIRST_UID (m1->regno)] > last
1517 || uid_luid[REGNO_LAST_UID (m1->regno)] < first))
1520 /* No overlap: we can combine this with the others. */
1521 m0->lifetime += m->lifetime;
1522 m0->savings += m->savings;
1531 free (matched_regs);
1534 /* Return 1 if regs X and Y will become the same if moved. */
1537 regs_match_p (x, y, movables)
1539 struct movable *movables;
1541 unsigned int xn = REGNO (x);
1542 unsigned int yn = REGNO (y);
1543 struct movable *mx, *my;
1545 for (mx = movables; mx; mx = mx->next)
1546 if (mx->regno == xn)
1549 for (my = movables; my; my = my->next)
1550 if (my->regno == yn)
1554 && ((mx->match == my->match && mx->match != 0)
1556 || mx == my->match));
1559 /* Return 1 if X and Y are identical-looking rtx's.
1560 This is the Lisp function EQUAL for rtx arguments.
1562 If two registers are matching movables or a movable register and an
1563 equivalent constant, consider them equal. */
1566 rtx_equal_for_loop_p (x, y, movables)
1568 struct movable *movables;
1572 register struct movable *m;
1573 register enum rtx_code code;
1574 register const char *fmt;
1578 if (x == 0 || y == 0)
1581 code = GET_CODE (x);
1583 /* If we have a register and a constant, they may sometimes be
1585 if (GET_CODE (x) == REG && VARRAY_INT (set_in_loop, REGNO (x)) == -2
1588 for (m = movables; m; m = m->next)
1589 if (m->move_insn && m->regno == REGNO (x)
1590 && rtx_equal_p (m->set_src, y))
1593 else if (GET_CODE (y) == REG && VARRAY_INT (set_in_loop, REGNO (y)) == -2
1596 for (m = movables; m; m = m->next)
1597 if (m->move_insn && m->regno == REGNO (y)
1598 && rtx_equal_p (m->set_src, x))
1602 /* Otherwise, rtx's of different codes cannot be equal. */
1603 if (code != GET_CODE (y))
1606 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent.
1607 (REG:SI x) and (REG:HI x) are NOT equivalent. */
1609 if (GET_MODE (x) != GET_MODE (y))
1612 /* These three types of rtx's can be compared nonrecursively. */
1614 return (REGNO (x) == REGNO (y) || regs_match_p (x, y, movables));
1616 if (code == LABEL_REF)
1617 return XEXP (x, 0) == XEXP (y, 0);
1618 if (code == SYMBOL_REF)
1619 return XSTR (x, 0) == XSTR (y, 0);
1621 /* Compare the elements. If any pair of corresponding elements
1622 fail to match, return 0 for the whole things. */
1624 fmt = GET_RTX_FORMAT (code);
1625 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1630 if (XWINT (x, i) != XWINT (y, i))
1635 if (XINT (x, i) != XINT (y, i))
1640 /* Two vectors must have the same length. */
1641 if (XVECLEN (x, i) != XVECLEN (y, i))
1644 /* And the corresponding elements must match. */
1645 for (j = 0; j < XVECLEN (x, i); j++)
1646 if (rtx_equal_for_loop_p (XVECEXP (x, i, j), XVECEXP (y, i, j), movables) == 0)
1651 if (rtx_equal_for_loop_p (XEXP (x, i), XEXP (y, i), movables) == 0)
1656 if (strcmp (XSTR (x, i), XSTR (y, i)))
1661 /* These are just backpointers, so they don't matter. */
1667 /* It is believed that rtx's at this level will never
1668 contain anything but integers and other rtx's,
1669 except for within LABEL_REFs and SYMBOL_REFs. */
1677 /* If X contains any LABEL_REF's, add REG_LABEL notes for them to all
1678 insns in INSNS which use the reference. */
1681 add_label_notes (x, insns)
1685 enum rtx_code code = GET_CODE (x);
1690 if (code == LABEL_REF && !LABEL_REF_NONLOCAL_P (x))
1692 /* This code used to ignore labels that referred to dispatch tables to
1693 avoid flow generating (slighly) worse code.
1695 We no longer ignore such label references (see LABEL_REF handling in
1696 mark_jump_label for additional information). */
1697 for (insn = insns; insn; insn = NEXT_INSN (insn))
1698 if (reg_mentioned_p (XEXP (x, 0), insn))
1699 REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_LABEL, XEXP (x, 0),
1703 fmt = GET_RTX_FORMAT (code);
1704 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1707 add_label_notes (XEXP (x, i), insns);
1708 else if (fmt[i] == 'E')
1709 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
1710 add_label_notes (XVECEXP (x, i, j), insns);
1714 /* Scan MOVABLES, and move the insns that deserve to be moved.
1715 If two matching movables are combined, replace one reg with the
1716 other throughout. */
1719 move_movables (loop, movables, threshold, insn_count, nregs)
1721 struct movable *movables;
1727 register struct movable *m;
1729 rtx loop_start = loop->start;
1730 rtx loop_end = loop->end;
1731 /* Map of pseudo-register replacements to handle combining
1732 when we move several insns that load the same value
1733 into different pseudo-registers. */
1734 rtx *reg_map = (rtx *) xcalloc (nregs, sizeof (rtx));
1735 char *already_moved = (char *) xcalloc (nregs, sizeof (char));
1739 for (m = movables; m; m = m->next)
1741 /* Describe this movable insn. */
1743 if (loop_dump_stream)
1745 fprintf (loop_dump_stream, "Insn %d: regno %d (life %d), ",
1746 INSN_UID (m->insn), m->regno, m->lifetime);
1748 fprintf (loop_dump_stream, "consec %d, ", m->consec);
1750 fprintf (loop_dump_stream, "cond ");
1752 fprintf (loop_dump_stream, "force ");
1754 fprintf (loop_dump_stream, "global ");
1756 fprintf (loop_dump_stream, "done ");
1758 fprintf (loop_dump_stream, "move-insn ");
1760 fprintf (loop_dump_stream, "matches %d ",
1761 INSN_UID (m->match->insn));
1763 fprintf (loop_dump_stream, "forces %d ",
1764 INSN_UID (m->forces->insn));
1767 /* Count movables. Value used in heuristics in strength_reduce. */
1770 /* Ignore the insn if it's already done (it matched something else).
1771 Otherwise, see if it is now safe to move. */
1775 || (1 == loop_invariant_p (loop, m->set_src)
1776 && (m->dependencies == 0
1777 || 1 == loop_invariant_p (loop, m->dependencies))
1779 || 1 == consec_sets_invariant_p (loop, m->set_dest,
1782 && (! m->forces || m->forces->done))
1786 int savings = m->savings;
1788 /* We have an insn that is safe to move.
1789 Compute its desirability. */
1794 if (loop_dump_stream)
1795 fprintf (loop_dump_stream, "savings %d ", savings);
1797 if (moved_once[regno] && loop_dump_stream)
1798 fprintf (loop_dump_stream, "halved since already moved ");
1800 /* An insn MUST be moved if we already moved something else
1801 which is safe only if this one is moved too: that is,
1802 if already_moved[REGNO] is nonzero. */
1804 /* An insn is desirable to move if the new lifetime of the
1805 register is no more than THRESHOLD times the old lifetime.
1806 If it's not desirable, it means the loop is so big
1807 that moving won't speed things up much,
1808 and it is liable to make register usage worse. */
1810 /* It is also desirable to move if it can be moved at no
1811 extra cost because something else was already moved. */
1813 if (already_moved[regno]
1814 || flag_move_all_movables
1815 || (threshold * savings * m->lifetime) >=
1816 (moved_once[regno] ? insn_count * 2 : insn_count)
1817 || (m->forces && m->forces->done
1818 && VARRAY_INT (n_times_set, m->forces->regno) == 1))
1821 register struct movable *m1;
1822 rtx first = NULL_RTX;
1824 /* Now move the insns that set the reg. */
1826 if (m->partial && m->match)
1830 /* Find the end of this chain of matching regs.
1831 Thus, we load each reg in the chain from that one reg.
1832 And that reg is loaded with 0 directly,
1833 since it has ->match == 0. */
1834 for (m1 = m; m1->match; m1 = m1->match);
1835 newpat = gen_move_insn (SET_DEST (PATTERN (m->insn)),
1836 SET_DEST (PATTERN (m1->insn)));
1837 i1 = emit_insn_before (newpat, loop_start);
1839 /* Mark the moved, invariant reg as being allowed to
1840 share a hard reg with the other matching invariant. */
1841 REG_NOTES (i1) = REG_NOTES (m->insn);
1842 r1 = SET_DEST (PATTERN (m->insn));
1843 r2 = SET_DEST (PATTERN (m1->insn));
1845 = gen_rtx_EXPR_LIST (VOIDmode, r1,
1846 gen_rtx_EXPR_LIST (VOIDmode, r2,
1848 delete_insn (m->insn);
1853 if (loop_dump_stream)
1854 fprintf (loop_dump_stream, " moved to %d", INSN_UID (i1));
1856 /* If we are to re-generate the item being moved with a
1857 new move insn, first delete what we have and then emit
1858 the move insn before the loop. */
1859 else if (m->move_insn)
1863 for (count = m->consec; count >= 0; count--)
1865 /* If this is the first insn of a library call sequence,
1867 if (GET_CODE (p) != NOTE
1868 && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
1871 /* If this is the last insn of a libcall sequence, then
1872 delete every insn in the sequence except the last.
1873 The last insn is handled in the normal manner. */
1874 if (GET_CODE (p) != NOTE
1875 && (temp = find_reg_note (p, REG_RETVAL, NULL_RTX)))
1877 temp = XEXP (temp, 0);
1879 temp = delete_insn (temp);
1883 p = delete_insn (p);
1885 /* simplify_giv_expr expects that it can walk the insns
1886 at m->insn forwards and see this old sequence we are
1887 tossing here. delete_insn does preserve the next
1888 pointers, but when we skip over a NOTE we must fix
1889 it up. Otherwise that code walks into the non-deleted
1891 while (p && GET_CODE (p) == NOTE)
1892 p = NEXT_INSN (temp) = NEXT_INSN (p);
1896 emit_move_insn (m->set_dest, m->set_src);
1897 temp = get_insns ();
1900 add_label_notes (m->set_src, temp);
1902 i1 = emit_insns_before (temp, loop_start);
1903 if (! find_reg_note (i1, REG_EQUAL, NULL_RTX))
1905 = gen_rtx_EXPR_LIST (m->is_equiv ? REG_EQUIV : REG_EQUAL,
1906 m->set_src, REG_NOTES (i1));
1908 if (loop_dump_stream)
1909 fprintf (loop_dump_stream, " moved to %d", INSN_UID (i1));
1911 /* The more regs we move, the less we like moving them. */
1916 for (count = m->consec; count >= 0; count--)
1920 /* If first insn of libcall sequence, skip to end. */
1921 /* Do this at start of loop, since p is guaranteed to
1923 if (GET_CODE (p) != NOTE
1924 && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
1927 /* If last insn of libcall sequence, move all
1928 insns except the last before the loop. The last
1929 insn is handled in the normal manner. */
1930 if (GET_CODE (p) != NOTE
1931 && (temp = find_reg_note (p, REG_RETVAL, NULL_RTX)))
1935 rtx fn_address_insn = 0;
1938 for (temp = XEXP (temp, 0); temp != p;
1939 temp = NEXT_INSN (temp))
1945 if (GET_CODE (temp) == NOTE)
1948 body = PATTERN (temp);
1950 /* Find the next insn after TEMP,
1951 not counting USE or NOTE insns. */
1952 for (next = NEXT_INSN (temp); next != p;
1953 next = NEXT_INSN (next))
1954 if (! (GET_CODE (next) == INSN
1955 && GET_CODE (PATTERN (next)) == USE)
1956 && GET_CODE (next) != NOTE)
1959 /* If that is the call, this may be the insn
1960 that loads the function address.
1962 Extract the function address from the insn
1963 that loads it into a register.
1964 If this insn was cse'd, we get incorrect code.
1966 So emit a new move insn that copies the
1967 function address into the register that the
1968 call insn will use. flow.c will delete any
1969 redundant stores that we have created. */
1970 if (GET_CODE (next) == CALL_INSN
1971 && GET_CODE (body) == SET
1972 && GET_CODE (SET_DEST (body)) == REG
1973 && (n = find_reg_note (temp, REG_EQUAL,
1976 fn_reg = SET_SRC (body);
1977 if (GET_CODE (fn_reg) != REG)
1978 fn_reg = SET_DEST (body);
1979 fn_address = XEXP (n, 0);
1980 fn_address_insn = temp;
1982 /* We have the call insn.
1983 If it uses the register we suspect it might,
1984 load it with the correct address directly. */
1985 if (GET_CODE (temp) == CALL_INSN
1987 && reg_referenced_p (fn_reg, body))
1988 emit_insn_after (gen_move_insn (fn_reg,
1992 if (GET_CODE (temp) == CALL_INSN)
1994 i1 = emit_call_insn_before (body, loop_start);
1995 /* Because the USAGE information potentially
1996 contains objects other than hard registers
1997 we need to copy it. */
1998 if (CALL_INSN_FUNCTION_USAGE (temp))
1999 CALL_INSN_FUNCTION_USAGE (i1)
2000 = copy_rtx (CALL_INSN_FUNCTION_USAGE (temp));
2003 i1 = emit_insn_before (body, loop_start);
2006 if (temp == fn_address_insn)
2007 fn_address_insn = i1;
2008 REG_NOTES (i1) = REG_NOTES (temp);
2014 if (m->savemode != VOIDmode)
2016 /* P sets REG to zero; but we should clear only
2017 the bits that are not covered by the mode
2019 rtx reg = m->set_dest;
2025 (GET_MODE (reg), and_optab, reg,
2026 GEN_INT ((((HOST_WIDE_INT) 1
2027 << GET_MODE_BITSIZE (m->savemode)))
2029 reg, 1, OPTAB_LIB_WIDEN);
2033 emit_move_insn (reg, tem);
2034 sequence = gen_sequence ();
2036 i1 = emit_insn_before (sequence, loop_start);
2038 else if (GET_CODE (p) == CALL_INSN)
2040 i1 = emit_call_insn_before (PATTERN (p), loop_start);
2041 /* Because the USAGE information potentially
2042 contains objects other than hard registers
2043 we need to copy it. */
2044 if (CALL_INSN_FUNCTION_USAGE (p))
2045 CALL_INSN_FUNCTION_USAGE (i1)
2046 = copy_rtx (CALL_INSN_FUNCTION_USAGE (p));
2048 else if (count == m->consec && m->move_insn_first)
2050 /* The SET_SRC might not be invariant, so we must
2051 use the REG_EQUAL note. */
2053 emit_move_insn (m->set_dest, m->set_src);
2054 temp = get_insns ();
2057 add_label_notes (m->set_src, temp);
2059 i1 = emit_insns_before (temp, loop_start);
2060 if (! find_reg_note (i1, REG_EQUAL, NULL_RTX))
2062 = gen_rtx_EXPR_LIST ((m->is_equiv ? REG_EQUIV
2064 m->set_src, REG_NOTES (i1));
2067 i1 = emit_insn_before (PATTERN (p), loop_start);
2069 if (REG_NOTES (i1) == 0)
2071 REG_NOTES (i1) = REG_NOTES (p);
2073 /* If there is a REG_EQUAL note present whose value
2074 is not loop invariant, then delete it, since it
2075 may cause problems with later optimization passes.
2076 It is possible for cse to create such notes
2077 like this as a result of record_jump_cond. */
2079 if ((temp = find_reg_note (i1, REG_EQUAL, NULL_RTX))
2080 && ! loop_invariant_p (loop, XEXP (temp, 0)))
2081 remove_note (i1, temp);
2087 if (loop_dump_stream)
2088 fprintf (loop_dump_stream, " moved to %d",
2091 /* If library call, now fix the REG_NOTES that contain
2092 insn pointers, namely REG_LIBCALL on FIRST
2093 and REG_RETVAL on I1. */
2094 if ((temp = find_reg_note (i1, REG_RETVAL, NULL_RTX)))
2096 XEXP (temp, 0) = first;
2097 temp = find_reg_note (first, REG_LIBCALL, NULL_RTX);
2098 XEXP (temp, 0) = i1;
2105 /* simplify_giv_expr expects that it can walk the insns
2106 at m->insn forwards and see this old sequence we are
2107 tossing here. delete_insn does preserve the next
2108 pointers, but when we skip over a NOTE we must fix
2109 it up. Otherwise that code walks into the non-deleted
2111 while (p && GET_CODE (p) == NOTE)
2112 p = NEXT_INSN (temp) = NEXT_INSN (p);
2115 /* The more regs we move, the less we like moving them. */
2119 /* Any other movable that loads the same register
2121 already_moved[regno] = 1;
2123 /* This reg has been moved out of one loop. */
2124 moved_once[regno] = 1;
2126 /* The reg set here is now invariant. */
2128 VARRAY_INT (set_in_loop, regno) = 0;
2132 /* Change the length-of-life info for the register
2133 to say it lives at least the full length of this loop.
2134 This will help guide optimizations in outer loops. */
2136 if (uid_luid[REGNO_FIRST_UID (regno)] > INSN_LUID (loop_start))
2137 /* This is the old insn before all the moved insns.
2138 We can't use the moved insn because it is out of range
2139 in uid_luid. Only the old insns have luids. */
2140 REGNO_FIRST_UID (regno) = INSN_UID (loop_start);
2141 if (uid_luid[REGNO_LAST_UID (regno)] < INSN_LUID (loop_end))
2142 REGNO_LAST_UID (regno) = INSN_UID (loop_end);
2144 /* Combine with this moved insn any other matching movables. */
2147 for (m1 = movables; m1; m1 = m1->next)
2152 /* Schedule the reg loaded by M1
2153 for replacement so that shares the reg of M.
2154 If the modes differ (only possible in restricted
2155 circumstances, make a SUBREG.
2157 Note this assumes that the target dependent files
2158 treat REG and SUBREG equally, including within
2159 GO_IF_LEGITIMATE_ADDRESS and in all the
2160 predicates since we never verify that replacing the
2161 original register with a SUBREG results in a
2162 recognizable insn. */
2163 if (GET_MODE (m->set_dest) == GET_MODE (m1->set_dest))
2164 reg_map[m1->regno] = m->set_dest;
2167 = gen_lowpart_common (GET_MODE (m1->set_dest),
2170 /* Get rid of the matching insn
2171 and prevent further processing of it. */
2174 /* if library call, delete all insn except last, which
2176 if ((temp = find_reg_note (m1->insn, REG_RETVAL,
2179 for (temp = XEXP (temp, 0); temp != m1->insn;
2180 temp = NEXT_INSN (temp))
2183 delete_insn (m1->insn);
2185 /* Any other movable that loads the same register
2187 already_moved[m1->regno] = 1;
2189 /* The reg merged here is now invariant,
2190 if the reg it matches is invariant. */
2192 VARRAY_INT (set_in_loop, m1->regno) = 0;
2195 else if (loop_dump_stream)
2196 fprintf (loop_dump_stream, "not desirable");
2198 else if (loop_dump_stream && !m->match)
2199 fprintf (loop_dump_stream, "not safe");
2201 if (loop_dump_stream)
2202 fprintf (loop_dump_stream, "\n");
2206 new_start = loop_start;
2208 /* Go through all the instructions in the loop, making
2209 all the register substitutions scheduled in REG_MAP. */
2210 for (p = new_start; p != loop_end; p = NEXT_INSN (p))
2211 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
2212 || GET_CODE (p) == CALL_INSN)
2214 replace_regs (PATTERN (p), reg_map, nregs, 0);
2215 replace_regs (REG_NOTES (p), reg_map, nregs, 0);
2221 free (already_moved);
2225 /* Scan X and replace the address of any MEM in it with ADDR.
2226 REG is the address that MEM should have before the replacement. */
2229 replace_call_address (x, reg, addr)
2232 register enum rtx_code code;
2234 register const char *fmt;
2238 code = GET_CODE (x);
2252 /* Short cut for very common case. */
2253 replace_call_address (XEXP (x, 1), reg, addr);
2257 /* Short cut for very common case. */
2258 replace_call_address (XEXP (x, 0), reg, addr);
2262 /* If this MEM uses a reg other than the one we expected,
2263 something is wrong. */
2264 if (XEXP (x, 0) != reg)
2273 fmt = GET_RTX_FORMAT (code);
2274 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2277 replace_call_address (XEXP (x, i), reg, addr);
2278 else if (fmt[i] == 'E')
2281 for (j = 0; j < XVECLEN (x, i); j++)
2282 replace_call_address (XVECEXP (x, i, j), reg, addr);
2288 /* Return the number of memory refs to addresses that vary
2292 count_nonfixed_reads (loop, x)
2293 const struct loop *loop;
2296 register enum rtx_code code;
2298 register const char *fmt;
2304 code = GET_CODE (x);
2318 return ((loop_invariant_p (loop, XEXP (x, 0)) != 1)
2319 + count_nonfixed_reads (loop, XEXP (x, 0)));
2326 fmt = GET_RTX_FORMAT (code);
2327 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2330 value += count_nonfixed_reads (loop, XEXP (x, i));
2334 for (j = 0; j < XVECLEN (x, i); j++)
2335 value += count_nonfixed_reads (loop, XVECEXP (x, i, j));
2343 /* P is an instruction that sets a register to the result of a ZERO_EXTEND.
2344 Replace it with an instruction to load just the low bytes
2345 if the machine supports such an instruction,
2346 and insert above LOOP_START an instruction to clear the register. */
2349 constant_high_bytes (p, loop_start)
2353 register int insn_code_number;
2355 /* Try to change (SET (REG ...) (ZERO_EXTEND (..:B ...)))
2356 to (SET (STRICT_LOW_PART (SUBREG:B (REG...))) ...). */
2361 gen_rtx_STRICT_LOW_PART
2363 gen_rtx_SUBREG (GET_MODE (XEXP (SET_SRC (PATTERN (p)), 0)),
2364 SET_DEST (PATTERN (p)), 0)),
2365 XEXP (SET_SRC (PATTERN (p)), 0));
2367 insn_code_number = recog (new, p);
2369 if (insn_code_number)
2373 /* Clear destination register before the loop. */
2374 emit_insn_before (gen_rtx_SET (VOIDmode,
2375 SET_DEST (PATTERN (p)), const0_rtx),
2378 /* Inside the loop, just load the low part. */
2384 /* Scan a loop setting the elements `cont', `vtop', `loops_enclosed',
2385 `has_call', `has_volatile', and `has_tablejump' within LOOP.
2386 Set the global variables `unknown_address_altered',
2387 `unknown_constant_address_altered', and `num_mem_sets'. Also, fill
2388 in the array `loop_mems' and the list `loop_store_mems'. */
2394 register int level = 1;
2396 struct loop_info *loop_info = LOOP_INFO (loop);
2397 rtx start = loop->start;
2398 rtx end = loop->end;
2399 /* The label after END. Jumping here is just like falling off the
2400 end of the loop. We use next_nonnote_insn instead of next_label
2401 as a hedge against the (pathological) case where some actual insn
2402 might end up between the two. */
2403 rtx exit_target = next_nonnote_insn (end);
2405 loop_info->has_indirect_jump = indirect_jump_in_function;
2406 loop_info->has_call = 0;
2407 loop_info->has_volatile = 0;
2408 loop_info->has_tablejump = 0;
2409 loop_info->has_multiple_exit_targets = 0;
2414 unknown_address_altered = 0;
2415 unknown_constant_address_altered = 0;
2416 loop_store_mems = NULL_RTX;
2417 first_loop_store_insn = NULL_RTX;
2421 for (insn = NEXT_INSN (start); insn != NEXT_INSN (end);
2422 insn = NEXT_INSN (insn))
2424 if (GET_CODE (insn) == NOTE)
2426 if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG)
2429 /* Count number of loops contained in this one. */
2432 else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_END)
2441 else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_CONT)
2446 else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_VTOP)
2448 /* If there is a NOTE_INSN_LOOP_VTOP, then this is a for
2449 or while style loop, with a loop exit test at the
2450 start. Thus, we can assume that the loop condition
2451 was true when the loop was entered. */
2456 else if (GET_CODE (insn) == CALL_INSN)
2458 if (! CONST_CALL_P (insn))
2459 unknown_address_altered = 1;
2460 loop_info->has_call = 1;
2462 else if (GET_CODE (insn) == INSN || GET_CODE (insn) == JUMP_INSN)
2464 rtx label1 = NULL_RTX;
2465 rtx label2 = NULL_RTX;
2467 if (volatile_refs_p (PATTERN (insn)))
2468 loop_info->has_volatile = 1;
2470 if (GET_CODE (insn) == JUMP_INSN
2471 && (GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC
2472 || GET_CODE (PATTERN (insn)) == ADDR_VEC))
2473 loop_info->has_tablejump = 1;
2475 note_stores (PATTERN (insn), note_addr_stored, NULL);
2476 if (! first_loop_store_insn && loop_store_mems)
2477 first_loop_store_insn = insn;
2479 if (! loop_info->has_multiple_exit_targets
2480 && GET_CODE (insn) == JUMP_INSN
2481 && GET_CODE (PATTERN (insn)) == SET
2482 && SET_DEST (PATTERN (insn)) == pc_rtx)
2484 if (GET_CODE (SET_SRC (PATTERN (insn))) == IF_THEN_ELSE)
2486 label1 = XEXP (SET_SRC (PATTERN (insn)), 1);
2487 label2 = XEXP (SET_SRC (PATTERN (insn)), 2);
2491 label1 = SET_SRC (PATTERN (insn));
2495 if (label1 && label1 != pc_rtx)
2497 if (GET_CODE (label1) != LABEL_REF)
2499 /* Something tricky. */
2500 loop_info->has_multiple_exit_targets = 1;
2503 else if (XEXP (label1, 0) != exit_target
2504 && LABEL_OUTSIDE_LOOP_P (label1))
2506 /* A jump outside the current loop. */
2507 loop_info->has_multiple_exit_targets = 1;
2517 else if (GET_CODE (insn) == RETURN)
2518 loop_info->has_multiple_exit_targets = 1;
2521 /* Now, rescan the loop, setting up the LOOP_MEMS array. */
2522 if (/* We can't tell what MEMs are aliased by what. */
2523 ! unknown_address_altered
2524 /* An exception thrown by a called function might land us
2526 && ! loop_info->has_call
2527 /* We don't want loads for MEMs moved to a location before the
2528 one at which their stack memory becomes allocated. (Note
2529 that this is not a problem for malloc, etc., since those
2530 require actual function calls. */
2531 && ! current_function_calls_alloca
2532 /* There are ways to leave the loop other than falling off the
2534 && ! loop_info->has_multiple_exit_targets)
2535 for (insn = NEXT_INSN (start); insn != NEXT_INSN (end);
2536 insn = NEXT_INSN (insn))
2537 for_each_rtx (&insn, insert_loop_mem, 0);
2540 /* LOOP->CONT_DOMINATOR is now the last label between the loop start
2541 and the continue note that is a the destination of a (cond)jump after
2542 the continue note. If there is any (cond)jump between the loop start
2543 and what we have so far as LOOP->CONT_DOMINATOR that has a
2544 target between LOOP->DOMINATOR and the continue note, move
2545 LOOP->CONT_DOMINATOR forward to that label; if a jump's
2546 destination cannot be determined, clear LOOP->CONT_DOMINATOR. */
2549 verify_dominator (loop)
2554 if (! loop->cont_dominator)
2555 /* This can happen for an empty loop, e.g. in
2556 gcc.c-torture/compile/920410-2.c */
2558 if (loop->cont_dominator == const0_rtx)
2560 loop->cont_dominator = 0;
2563 for (insn = loop->start; insn != loop->cont_dominator;
2564 insn = NEXT_INSN (insn))
2566 if (GET_CODE (insn) == JUMP_INSN
2567 && GET_CODE (PATTERN (insn)) != RETURN)
2569 rtx label = JUMP_LABEL (insn);
2572 /* If it is not a jump we can easily understand or for
2573 which we do not have jump target information in the JUMP_LABEL
2574 field (consider ADDR_VEC and ADDR_DIFF_VEC insns), then clear
2575 LOOP->CONT_DOMINATOR. */
2576 if ((! condjump_p (insn)
2577 && ! condjump_in_parallel_p (insn))
2578 || label == NULL_RTX)
2580 loop->cont_dominator = NULL_RTX;
2584 label_luid = INSN_LUID (label);
2585 if (label_luid < INSN_LUID (loop->cont)
2587 > INSN_LUID (loop->cont)))
2588 loop->cont_dominator = label;
2593 /* Scan the function looking for loops. Record the start and end of each loop.
2594 Also mark as invalid loops any loops that contain a setjmp or are branched
2595 to from outside the loop. */
2598 find_and_verify_loops (f, loops)
2600 struct loops *loops;
2605 struct loop *current_loop;
2606 struct loop *next_loop;
2609 num_loops = loops->num;
2611 compute_luids (f, NULL_RTX, 0);
2613 /* If there are jumps to undefined labels,
2614 treat them as jumps out of any/all loops.
2615 This also avoids writing past end of tables when there are no loops. */
2618 /* Find boundaries of loops, mark which loops are contained within
2619 loops, and invalidate loops that have setjmp. */
2622 current_loop = NULL;
2623 for (insn = f; insn; insn = NEXT_INSN (insn))
2625 if (GET_CODE (insn) == NOTE)
2626 switch (NOTE_LINE_NUMBER (insn))
2628 case NOTE_INSN_LOOP_BEG:
2629 next_loop = loops->array + num_loops;
2630 next_loop->num = num_loops;
2632 next_loop->start = insn;
2633 next_loop->outer = current_loop;
2634 current_loop = next_loop;
2637 case NOTE_INSN_SETJMP:
2638 /* In this case, we must invalidate our current loop and any
2640 for (loop = current_loop; loop; loop = loop->outer)
2643 if (loop_dump_stream)
2644 fprintf (loop_dump_stream,
2645 "\nLoop at %d ignored due to setjmp.\n",
2646 INSN_UID (loop->start));
2650 case NOTE_INSN_LOOP_CONT:
2651 current_loop->cont = insn;
2653 case NOTE_INSN_LOOP_END:
2657 current_loop->end = insn;
2658 verify_dominator (current_loop);
2659 current_loop = current_loop->outer;
2665 /* If for any loop, this is a jump insn between the NOTE_INSN_LOOP_CONT
2666 and NOTE_INSN_LOOP_END notes, update loop->dominator. */
2667 else if (GET_CODE (insn) == JUMP_INSN
2668 && GET_CODE (PATTERN (insn)) != RETURN
2671 rtx label = JUMP_LABEL (insn);
2673 if (! condjump_p (insn) && ! condjump_in_parallel_p (insn))
2676 loop = current_loop;
2679 /* First see if we care about this loop. */
2680 if (loop->cont && loop->cont_dominator != const0_rtx)
2682 /* If the jump destination is not known, invalidate
2683 loop->const_dominator. */
2685 loop->cont_dominator = const0_rtx;
2687 /* Check if the destination is between loop start and
2689 if ((INSN_LUID (label)
2690 < INSN_LUID (loop->cont))
2691 && (INSN_LUID (label)
2692 > INSN_LUID (loop->start))
2693 /* And if there is no later destination already
2695 && (! loop->cont_dominator
2696 || (INSN_LUID (label)
2697 > INSN_LUID (loop->cont_dominator))))
2698 loop->cont_dominator = label;
2705 /* Note that this will mark the NOTE_INSN_LOOP_END note as being in the
2706 enclosing loop, but this doesn't matter. */
2707 uid_loop[INSN_UID (insn)] = current_loop;
2710 /* Any loop containing a label used in an initializer must be invalidated,
2711 because it can be jumped into from anywhere. */
2713 for (label = forced_labels; label; label = XEXP (label, 1))
2715 for (loop = uid_loop[INSN_UID (XEXP (label, 0))];
2716 loop; loop = loop->outer)
2720 /* Any loop containing a label used for an exception handler must be
2721 invalidated, because it can be jumped into from anywhere. */
2723 for (label = exception_handler_labels; label; label = XEXP (label, 1))
2725 for (loop = uid_loop[INSN_UID (XEXP (label, 0))];
2726 loop; loop = loop->outer)
2730 /* Now scan all insn's in the function. If any JUMP_INSN branches into a
2731 loop that it is not contained within, that loop is marked invalid.
2732 If any INSN or CALL_INSN uses a label's address, then the loop containing
2733 that label is marked invalid, because it could be jumped into from
2736 Also look for blocks of code ending in an unconditional branch that
2737 exits the loop. If such a block is surrounded by a conditional
2738 branch around the block, move the block elsewhere (see below) and
2739 invert the jump to point to the code block. This may eliminate a
2740 label in our loop and will simplify processing by both us and a
2741 possible second cse pass. */
2743 for (insn = f; insn; insn = NEXT_INSN (insn))
2744 if (GET_RTX_CLASS (GET_CODE (insn)) == 'i')
2746 struct loop *this_loop = uid_loop[INSN_UID (insn)];
2748 if (GET_CODE (insn) == INSN || GET_CODE (insn) == CALL_INSN)
2750 rtx note = find_reg_note (insn, REG_LABEL, NULL_RTX);
2753 for (loop = uid_loop[INSN_UID (XEXP (note, 0))];
2754 loop; loop = loop->outer)
2759 if (GET_CODE (insn) != JUMP_INSN)
2762 mark_loop_jump (PATTERN (insn), this_loop);
2764 /* See if this is an unconditional branch outside the loop. */
2766 && (GET_CODE (PATTERN (insn)) == RETURN
2767 || (simplejump_p (insn)
2768 && (uid_loop[INSN_UID (JUMP_LABEL (insn))]
2770 && get_max_uid () < max_uid_for_loop)
2773 rtx our_next = next_real_insn (insn);
2774 rtx last_insn_to_move = NEXT_INSN (insn);
2775 struct loop *dest_loop;
2776 struct loop *outer_loop = NULL;
2778 /* Go backwards until we reach the start of the loop, a label,
2780 for (p = PREV_INSN (insn);
2781 GET_CODE (p) != CODE_LABEL
2782 && ! (GET_CODE (p) == NOTE
2783 && NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
2784 && GET_CODE (p) != JUMP_INSN;
2788 /* Check for the case where we have a jump to an inner nested
2789 loop, and do not perform the optimization in that case. */
2791 if (JUMP_LABEL (insn))
2793 dest_loop = uid_loop[INSN_UID (JUMP_LABEL (insn))];
2796 for (outer_loop = dest_loop; outer_loop;
2797 outer_loop = outer_loop->outer)
2798 if (outer_loop == this_loop)
2803 /* Make sure that the target of P is within the current loop. */
2805 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p)
2806 && uid_loop[INSN_UID (JUMP_LABEL (p))] != this_loop)
2807 outer_loop = this_loop;
2809 /* If we stopped on a JUMP_INSN to the next insn after INSN,
2810 we have a block of code to try to move.
2812 We look backward and then forward from the target of INSN
2813 to find a BARRIER at the same loop depth as the target.
2814 If we find such a BARRIER, we make a new label for the start
2815 of the block, invert the jump in P and point it to that label,
2816 and move the block of code to the spot we found. */
2819 && GET_CODE (p) == JUMP_INSN
2820 && JUMP_LABEL (p) != 0
2821 /* Just ignore jumps to labels that were never emitted.
2822 These always indicate compilation errors. */
2823 && INSN_UID (JUMP_LABEL (p)) != 0
2825 && ! simplejump_p (p)
2826 && next_real_insn (JUMP_LABEL (p)) == our_next
2827 /* If it's not safe to move the sequence, then we
2829 && insns_safe_to_move_p (p, NEXT_INSN (insn),
2830 &last_insn_to_move))
2833 = JUMP_LABEL (insn) ? JUMP_LABEL (insn) : get_last_insn ();
2834 struct loop *target_loop = uid_loop[INSN_UID (target)];
2837 for (loc = target; loc; loc = PREV_INSN (loc))
2838 if (GET_CODE (loc) == BARRIER
2839 /* Don't move things inside a tablejump. */
2840 && ((loc2 = next_nonnote_insn (loc)) == 0
2841 || GET_CODE (loc2) != CODE_LABEL
2842 || (loc2 = next_nonnote_insn (loc2)) == 0
2843 || GET_CODE (loc2) != JUMP_INSN
2844 || (GET_CODE (PATTERN (loc2)) != ADDR_VEC
2845 && GET_CODE (PATTERN (loc2)) != ADDR_DIFF_VEC))
2846 && uid_loop[INSN_UID (loc)] == target_loop)
2850 for (loc = target; loc; loc = NEXT_INSN (loc))
2851 if (GET_CODE (loc) == BARRIER
2852 /* Don't move things inside a tablejump. */
2853 && ((loc2 = next_nonnote_insn (loc)) == 0
2854 || GET_CODE (loc2) != CODE_LABEL
2855 || (loc2 = next_nonnote_insn (loc2)) == 0
2856 || GET_CODE (loc2) != JUMP_INSN
2857 || (GET_CODE (PATTERN (loc2)) != ADDR_VEC
2858 && GET_CODE (PATTERN (loc2)) != ADDR_DIFF_VEC))
2859 && uid_loop[INSN_UID (loc)] == target_loop)
2864 rtx cond_label = JUMP_LABEL (p);
2865 rtx new_label = get_label_after (p);
2867 /* Ensure our label doesn't go away. */
2868 LABEL_NUSES (cond_label)++;
2870 /* Verify that uid_loop is large enough and that
2872 if (invert_jump (p, new_label))
2876 /* If no suitable BARRIER was found, create a suitable
2877 one before TARGET. Since TARGET is a fall through
2878 path, we'll need to insert an jump around our block
2879 and a add a BARRIER before TARGET.
2881 This creates an extra unconditional jump outside
2882 the loop. However, the benefits of removing rarely
2883 executed instructions from inside the loop usually
2884 outweighs the cost of the extra unconditional jump
2885 outside the loop. */
2890 temp = gen_jump (JUMP_LABEL (insn));
2891 temp = emit_jump_insn_before (temp, target);
2892 JUMP_LABEL (temp) = JUMP_LABEL (insn);
2893 LABEL_NUSES (JUMP_LABEL (insn))++;
2894 loc = emit_barrier_before (target);
2897 /* Include the BARRIER after INSN and copy the
2899 new_label = squeeze_notes (new_label,
2901 reorder_insns (new_label, last_insn_to_move, loc);
2903 /* All those insns are now in TARGET_LOOP. */
2905 q != NEXT_INSN (last_insn_to_move);
2907 uid_loop[INSN_UID (q)] = target_loop;
2909 /* The label jumped to by INSN is no longer a loop exit.
2910 Unless INSN does not have a label (e.g., it is a
2911 RETURN insn), search loop->exit_labels to find
2912 its label_ref, and remove it. Also turn off
2913 LABEL_OUTSIDE_LOOP_P bit. */
2914 if (JUMP_LABEL (insn))
2917 r = this_loop->exit_labels;
2918 r; q = r, r = LABEL_NEXTREF (r))
2919 if (XEXP (r, 0) == JUMP_LABEL (insn))
2921 LABEL_OUTSIDE_LOOP_P (r) = 0;
2923 LABEL_NEXTREF (q) = LABEL_NEXTREF (r);
2925 this_loop->exit_labels = LABEL_NEXTREF (r);
2929 for (loop = this_loop; loop && loop != target_loop;
2933 /* If we didn't find it, then something is
2939 /* P is now a jump outside the loop, so it must be put
2940 in loop->exit_labels, and marked as such.
2941 The easiest way to do this is to just call
2942 mark_loop_jump again for P. */
2943 mark_loop_jump (PATTERN (p), this_loop);
2945 /* If INSN now jumps to the insn after it,
2947 if (JUMP_LABEL (insn) != 0
2948 && (next_real_insn (JUMP_LABEL (insn))
2949 == next_real_insn (insn)))
2953 /* Continue the loop after where the conditional
2954 branch used to jump, since the only branch insn
2955 in the block (if it still remains) is an inter-loop
2956 branch and hence needs no processing. */
2957 insn = NEXT_INSN (cond_label);
2959 if (--LABEL_NUSES (cond_label) == 0)
2960 delete_insn (cond_label);
2962 /* This loop will be continued with NEXT_INSN (insn). */
2963 insn = PREV_INSN (insn);
2970 /* If any label in X jumps to a loop different from LOOP_NUM and any of the
2971 loops it is contained in, mark the target loop invalid.
2973 For speed, we assume that X is part of a pattern of a JUMP_INSN. */
2976 mark_loop_jump (x, loop)
2980 struct loop *dest_loop;
2981 struct loop *outer_loop;
2984 switch (GET_CODE (x))
2997 /* There could be a label reference in here. */
2998 mark_loop_jump (XEXP (x, 0), loop);
3004 mark_loop_jump (XEXP (x, 0), loop);
3005 mark_loop_jump (XEXP (x, 1), loop);
3009 /* This may refer to a LABEL_REF or SYMBOL_REF. */
3010 mark_loop_jump (XEXP (x, 1), loop);
3015 mark_loop_jump (XEXP (x, 0), loop);
3019 dest_loop = uid_loop[INSN_UID (XEXP (x, 0))];
3021 /* Link together all labels that branch outside the loop. This
3022 is used by final_[bg]iv_value and the loop unrolling code. Also
3023 mark this LABEL_REF so we know that this branch should predict
3026 /* A check to make sure the label is not in an inner nested loop,
3027 since this does not count as a loop exit. */
3030 for (outer_loop = dest_loop; outer_loop;
3031 outer_loop = outer_loop->outer)
3032 if (outer_loop == loop)
3038 if (loop && ! outer_loop)
3040 LABEL_OUTSIDE_LOOP_P (x) = 1;
3041 LABEL_NEXTREF (x) = loop->exit_labels;
3042 loop->exit_labels = x;
3044 for (outer_loop = loop;
3045 outer_loop && outer_loop != dest_loop;
3046 outer_loop = outer_loop->outer)
3047 outer_loop->exit_count++;
3050 /* If this is inside a loop, but not in the current loop or one enclosed
3051 by it, it invalidates at least one loop. */
3056 /* We must invalidate every nested loop containing the target of this
3057 label, except those that also contain the jump insn. */
3059 for (; dest_loop; dest_loop = dest_loop->outer)
3061 /* Stop when we reach a loop that also contains the jump insn. */
3062 for (outer_loop = loop; outer_loop; outer_loop = outer_loop->outer)
3063 if (dest_loop == outer_loop)
3066 /* If we get here, we know we need to invalidate a loop. */
3067 if (loop_dump_stream && ! dest_loop->invalid)
3068 fprintf (loop_dump_stream,
3069 "\nLoop at %d ignored due to multiple entry points.\n",
3070 INSN_UID (dest_loop->start));
3072 dest_loop->invalid = 1;
3077 /* If this is not setting pc, ignore. */
3078 if (SET_DEST (x) == pc_rtx)
3079 mark_loop_jump (SET_SRC (x), loop);
3083 mark_loop_jump (XEXP (x, 1), loop);
3084 mark_loop_jump (XEXP (x, 2), loop);
3089 for (i = 0; i < XVECLEN (x, 0); i++)
3090 mark_loop_jump (XVECEXP (x, 0, i), loop);
3094 for (i = 0; i < XVECLEN (x, 1); i++)
3095 mark_loop_jump (XVECEXP (x, 1, i), loop);
3099 /* Strictly speaking this is not a jump into the loop, only a possible
3100 jump out of the loop. However, we have no way to link the destination
3101 of this jump onto the list of exit labels. To be safe we mark this
3102 loop and any containing loops as invalid. */
3105 for (outer_loop = loop; outer_loop; outer_loop = outer_loop->outer)
3107 if (loop_dump_stream && ! outer_loop->invalid)
3108 fprintf (loop_dump_stream,
3109 "\nLoop at %d ignored due to unknown exit jump.\n",
3110 INSN_UID (outer_loop->start));
3111 outer_loop->invalid = 1;
3118 /* Return nonzero if there is a label in the range from
3119 insn INSN to and including the insn whose luid is END
3120 INSN must have an assigned luid (i.e., it must not have
3121 been previously created by loop.c). */
3124 labels_in_range_p (insn, end)
3128 while (insn && INSN_LUID (insn) <= end)
3130 if (GET_CODE (insn) == CODE_LABEL)
3132 insn = NEXT_INSN (insn);
3138 /* Record that a memory reference X is being set. */
3141 note_addr_stored (x, y, data)
3143 rtx y ATTRIBUTE_UNUSED;
3144 void *data ATTRIBUTE_UNUSED;
3146 if (x == 0 || GET_CODE (x) != MEM)
3149 /* Count number of memory writes.
3150 This affects heuristics in strength_reduce. */
3153 /* BLKmode MEM means all memory is clobbered. */
3154 if (GET_MODE (x) == BLKmode)
3156 if (RTX_UNCHANGING_P (x))
3157 unknown_constant_address_altered = 1;
3159 unknown_address_altered = 1;
3164 loop_store_mems = gen_rtx_EXPR_LIST (VOIDmode, x, loop_store_mems);
3167 /* X is a value modified by an INSN that references a biv inside a loop
3168 exit test (ie, X is somehow related to the value of the biv). If X
3169 is a pseudo that is used more than once, then the biv is (effectively)
3170 used more than once. DATA is really an `int *', and is set if the
3171 biv is used more than once. */
3174 note_set_pseudo_multiple_uses (x, y, data)
3176 rtx y ATTRIBUTE_UNUSED;
3182 while (GET_CODE (x) == STRICT_LOW_PART
3183 || GET_CODE (x) == SIGN_EXTRACT
3184 || GET_CODE (x) == ZERO_EXTRACT
3185 || GET_CODE (x) == SUBREG)
3188 if (GET_CODE (x) != REG || REGNO (x) < FIRST_PSEUDO_REGISTER)
3191 /* If we do not have usage information, or if we know the register
3192 is used more than once, note that fact for check_dbra_loop. */
3193 if (REGNO (x) >= max_reg_before_loop
3194 || ! VARRAY_RTX (reg_single_usage, REGNO (x))
3195 || VARRAY_RTX (reg_single_usage, REGNO (x)) == const0_rtx)
3196 *((int *) data) = 1;
3199 /* Return nonzero if the rtx X is invariant over the current loop.
3201 The value is 2 if we refer to something only conditionally invariant.
3203 If `unknown_address_altered' is nonzero, no memory ref is invariant.
3204 Otherwise, a memory ref is invariant if it does not conflict with
3205 anything stored in `loop_store_mems'. */
3208 loop_invariant_p (loop, x)
3209 const struct loop *loop;
3213 register enum rtx_code code;
3214 register const char *fmt;
3215 int conditional = 0;
3220 code = GET_CODE (x);
3230 /* A LABEL_REF is normally invariant, however, if we are unrolling
3231 loops, and this label is inside the loop, then it isn't invariant.
3232 This is because each unrolled copy of the loop body will have
3233 a copy of this label. If this was invariant, then an insn loading
3234 the address of this label into a register might get moved outside
3235 the loop, and then each loop body would end up using the same label.
3237 We don't know the loop bounds here though, so just fail for all
3239 if (flag_unroll_loops)
3246 case UNSPEC_VOLATILE:
3250 /* We used to check RTX_UNCHANGING_P (x) here, but that is invalid
3251 since the reg might be set by initialization within the loop. */
3253 if ((x == frame_pointer_rtx || x == hard_frame_pointer_rtx
3254 || x == arg_pointer_rtx)
3255 && ! current_function_has_nonlocal_goto)
3258 if (LOOP_INFO (loop)->has_call
3259 && REGNO (x) < FIRST_PSEUDO_REGISTER && call_used_regs[REGNO (x)])
3262 if (VARRAY_INT (set_in_loop, REGNO (x)) < 0)
3265 return VARRAY_INT (set_in_loop, REGNO (x)) == 0;
3268 /* Volatile memory references must be rejected. Do this before
3269 checking for read-only items, so that volatile read-only items
3270 will be rejected also. */
3271 if (MEM_VOLATILE_P (x))
3274 /* If we had a subroutine call, any location in memory could
3275 have been clobbered. We used to test here for volatile and
3276 readonly, but true_dependence knows how to do that better
3278 if (RTX_UNCHANGING_P (x)
3279 ? unknown_constant_address_altered : unknown_address_altered)
3282 /* See if there is any dependence between a store and this load. */
3283 mem_list_entry = loop_store_mems;
3284 while (mem_list_entry)
3286 if (true_dependence (XEXP (mem_list_entry, 0), VOIDmode,
3290 mem_list_entry = XEXP (mem_list_entry, 1);
3293 /* It's not invalidated by a store in memory
3294 but we must still verify the address is invariant. */
3298 /* Don't mess with insns declared volatile. */
3299 if (MEM_VOLATILE_P (x))
3307 fmt = GET_RTX_FORMAT (code);
3308 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3312 int tem = loop_invariant_p (loop, XEXP (x, i));
3318 else if (fmt[i] == 'E')
3321 for (j = 0; j < XVECLEN (x, i); j++)
3323 int tem = loop_invariant_p (loop, XVECEXP (x, i, j));
3333 return 1 + conditional;
3337 /* Return nonzero if all the insns in the loop that set REG
3338 are INSN and the immediately following insns,
3339 and if each of those insns sets REG in an invariant way
3340 (not counting uses of REG in them).
3342 The value is 2 if some of these insns are only conditionally invariant.
3344 We assume that INSN itself is the first set of REG
3345 and that its source is invariant. */
3348 consec_sets_invariant_p (loop, reg, n_sets, insn)
3349 const struct loop *loop;
3354 unsigned int regno = REGNO (reg);
3356 /* Number of sets we have to insist on finding after INSN. */
3357 int count = n_sets - 1;
3358 int old = VARRAY_INT (set_in_loop, regno);
3362 /* If N_SETS hit the limit, we can't rely on its value. */
3366 VARRAY_INT (set_in_loop, regno) = 0;
3370 register enum rtx_code code;
3374 code = GET_CODE (p);
3376 /* If library call, skip to end of it. */
3377 if (code == INSN && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
3382 && (set = single_set (p))
3383 && GET_CODE (SET_DEST (set)) == REG
3384 && REGNO (SET_DEST (set)) == regno)
3386 this = loop_invariant_p (loop, SET_SRC (set));
3389 else if ((temp = find_reg_note (p, REG_EQUAL, NULL_RTX)))
3391 /* If this is a libcall, then any invariant REG_EQUAL note is OK.
3392 If this is an ordinary insn, then only CONSTANT_P REG_EQUAL
3394 this = (CONSTANT_P (XEXP (temp, 0))
3395 || (find_reg_note (p, REG_RETVAL, NULL_RTX)
3396 && loop_invariant_p (loop, XEXP (temp, 0))));
3403 else if (code != NOTE)
3405 VARRAY_INT (set_in_loop, regno) = old;
3410 VARRAY_INT (set_in_loop, regno) = old;
3411 /* If loop_invariant_p ever returned 2, we return 2. */
3412 return 1 + (value & 2);
3416 /* I don't think this condition is sufficient to allow INSN
3417 to be moved, so we no longer test it. */
3419 /* Return 1 if all insns in the basic block of INSN and following INSN
3420 that set REG are invariant according to TABLE. */
3423 all_sets_invariant_p (reg, insn, table)
3427 register rtx p = insn;
3428 register int regno = REGNO (reg);
3432 register enum rtx_code code;
3434 code = GET_CODE (p);
3435 if (code == CODE_LABEL || code == JUMP_INSN)
3437 if (code == INSN && GET_CODE (PATTERN (p)) == SET
3438 && GET_CODE (SET_DEST (PATTERN (p))) == REG
3439 && REGNO (SET_DEST (PATTERN (p))) == regno)
3441 if (! loop_invariant_p (loop, SET_SRC (PATTERN (p)), table))
3448 /* Look at all uses (not sets) of registers in X. For each, if it is
3449 the single use, set USAGE[REGNO] to INSN; if there was a previous use in
3450 a different insn, set USAGE[REGNO] to const0_rtx. */
3453 find_single_use_in_loop (insn, x, usage)
3458 enum rtx_code code = GET_CODE (x);
3459 const char *fmt = GET_RTX_FORMAT (code);
3463 VARRAY_RTX (usage, REGNO (x))
3464 = (VARRAY_RTX (usage, REGNO (x)) != 0
3465 && VARRAY_RTX (usage, REGNO (x)) != insn)
3466 ? const0_rtx : insn;
3468 else if (code == SET)
3470 /* Don't count SET_DEST if it is a REG; otherwise count things
3471 in SET_DEST because if a register is partially modified, it won't
3472 show up as a potential movable so we don't care how USAGE is set
3474 if (GET_CODE (SET_DEST (x)) != REG)
3475 find_single_use_in_loop (insn, SET_DEST (x), usage);
3476 find_single_use_in_loop (insn, SET_SRC (x), usage);
3479 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3481 if (fmt[i] == 'e' && XEXP (x, i) != 0)
3482 find_single_use_in_loop (insn, XEXP (x, i), usage);
3483 else if (fmt[i] == 'E')
3484 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
3485 find_single_use_in_loop (insn, XVECEXP (x, i, j), usage);
3489 /* Count and record any set in X which is contained in INSN. Update
3490 MAY_NOT_MOVE and LAST_SET for any register set in X. */
3493 count_one_set (insn, x, may_not_move, last_set)
3495 varray_type may_not_move;
3498 if (GET_CODE (x) == CLOBBER && GET_CODE (XEXP (x, 0)) == REG)
3499 /* Don't move a reg that has an explicit clobber.
3500 It's not worth the pain to try to do it correctly. */
3501 VARRAY_CHAR (may_not_move, REGNO (XEXP (x, 0))) = 1;
3503 if (GET_CODE (x) == SET || GET_CODE (x) == CLOBBER)
3505 rtx dest = SET_DEST (x);
3506 while (GET_CODE (dest) == SUBREG
3507 || GET_CODE (dest) == ZERO_EXTRACT
3508 || GET_CODE (dest) == SIGN_EXTRACT
3509 || GET_CODE (dest) == STRICT_LOW_PART)
3510 dest = XEXP (dest, 0);
3511 if (GET_CODE (dest) == REG)
3513 register int regno = REGNO (dest);
3514 /* If this is the first setting of this reg
3515 in current basic block, and it was set before,
3516 it must be set in two basic blocks, so it cannot
3517 be moved out of the loop. */
3518 if (VARRAY_INT (set_in_loop, regno) > 0
3519 && last_set[regno] == 0)
3520 VARRAY_CHAR (may_not_move, regno) = 1;
3521 /* If this is not first setting in current basic block,
3522 see if reg was used in between previous one and this.
3523 If so, neither one can be moved. */
3524 if (last_set[regno] != 0
3525 && reg_used_between_p (dest, last_set[regno], insn))
3526 VARRAY_CHAR (may_not_move, regno) = 1;
3527 if (VARRAY_INT (set_in_loop, regno) < 127)
3528 ++VARRAY_INT (set_in_loop, regno);
3529 last_set[regno] = insn;
3534 /* Increment SET_IN_LOOP at the index of each register
3535 that is modified by an insn between FROM and TO.
3536 If the value of an element of SET_IN_LOOP becomes 127 or more,
3537 stop incrementing it, to avoid overflow.
3539 Store in SINGLE_USAGE[I] the single insn in which register I is
3540 used, if it is only used once. Otherwise, it is set to 0 (for no
3541 uses) or const0_rtx for more than one use. This parameter may be zero,
3542 in which case this processing is not done.
3544 Store in *COUNT_PTR the number of actual instruction
3545 in the loop. We use this to decide what is worth moving out. */
3547 /* last_set[n] is nonzero iff reg n has been set in the current basic block.
3548 In that case, it is the insn that last set reg n. */
3551 count_loop_regs_set (from, to, may_not_move, single_usage, count_ptr, nregs)
3552 register rtx from, to;
3553 varray_type may_not_move;
3554 varray_type single_usage;
3558 register rtx *last_set = (rtx *) xcalloc (nregs, sizeof (rtx));
3560 register int count = 0;
3562 for (insn = from; insn != to; insn = NEXT_INSN (insn))
3564 if (GET_RTX_CLASS (GET_CODE (insn)) == 'i')
3568 /* Record registers that have exactly one use. */
3569 find_single_use_in_loop (insn, PATTERN (insn), single_usage);
3571 /* Include uses in REG_EQUAL notes. */
3572 if (REG_NOTES (insn))
3573 find_single_use_in_loop (insn, REG_NOTES (insn), single_usage);
3575 if (GET_CODE (PATTERN (insn)) == SET
3576 || GET_CODE (PATTERN (insn)) == CLOBBER)
3577 count_one_set (insn, PATTERN (insn), may_not_move, last_set);
3578 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
3581 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
3582 count_one_set (insn, XVECEXP (PATTERN (insn), 0, i),
3583 may_not_move, last_set);
3587 if (GET_CODE (insn) == CODE_LABEL || GET_CODE (insn) == JUMP_INSN)
3588 bzero ((char *) last_set, nregs * sizeof (rtx));
3596 /* Given a loop that is bounded by LOOP->START and LOOP->END and that
3597 is entered at LOOP->SCAN_START, return 1 if the register set in SET
3598 contained in insn INSN is used by any insn that precedes INSN in
3599 cyclic order starting from the loop entry point.
3601 We don't want to use INSN_LUID here because if we restrict INSN to those
3602 that have a valid INSN_LUID, it means we cannot move an invariant out
3603 from an inner loop past two loops. */
3606 loop_reg_used_before_p (loop, set, insn)
3607 const struct loop *loop;
3610 rtx reg = SET_DEST (set);
3613 /* Scan forward checking for register usage. If we hit INSN, we
3614 are done. Otherwise, if we hit LOOP->END, wrap around to LOOP->START. */
3615 for (p = loop->scan_start; p != insn; p = NEXT_INSN (p))
3617 if (GET_RTX_CLASS (GET_CODE (p)) == 'i'
3618 && reg_overlap_mentioned_p (reg, PATTERN (p)))
3628 /* A "basic induction variable" or biv is a pseudo reg that is set
3629 (within this loop) only by incrementing or decrementing it. */
3630 /* A "general induction variable" or giv is a pseudo reg whose
3631 value is a linear function of a biv. */
3633 /* Bivs are recognized by `basic_induction_var';
3634 Givs by `general_induction_var'. */
3636 /* Indexed by register number, indicates whether or not register is an
3637 induction variable, and if so what type. */
3639 varray_type reg_iv_type;
3641 /* Indexed by register number, contains pointer to `struct induction'
3642 if register is an induction variable. This holds general info for
3643 all induction variables. */
3645 varray_type reg_iv_info;
3647 /* Indexed by register number, contains pointer to `struct iv_class'
3648 if register is a basic induction variable. This holds info describing
3649 the class (a related group) of induction variables that the biv belongs
3652 struct iv_class **reg_biv_class;
3654 /* The head of a list which links together (via the next field)
3655 every iv class for the current loop. */
3657 struct iv_class *loop_iv_list;
3659 /* Givs made from biv increments are always splittable for loop unrolling.
3660 Since there is no regscan info for them, we have to keep track of them
3662 unsigned int first_increment_giv, last_increment_giv;
3664 /* Communication with routines called via `note_stores'. */
3666 static rtx note_insn;
3668 /* Dummy register to have non-zero DEST_REG for DEST_ADDR type givs. */
3670 static rtx addr_placeholder;
3672 /* ??? Unfinished optimizations, and possible future optimizations,
3673 for the strength reduction code. */
3675 /* ??? The interaction of biv elimination, and recognition of 'constant'
3676 bivs, may cause problems. */
3678 /* ??? Add heuristics so that DEST_ADDR strength reduction does not cause
3679 performance problems.
3681 Perhaps don't eliminate things that can be combined with an addressing
3682 mode. Find all givs that have the same biv, mult_val, and add_val;
3683 then for each giv, check to see if its only use dies in a following
3684 memory address. If so, generate a new memory address and check to see
3685 if it is valid. If it is valid, then store the modified memory address,
3686 otherwise, mark the giv as not done so that it will get its own iv. */
3688 /* ??? Could try to optimize branches when it is known that a biv is always
3691 /* ??? When replace a biv in a compare insn, we should replace with closest
3692 giv so that an optimized branch can still be recognized by the combiner,
3693 e.g. the VAX acb insn. */
3695 /* ??? Many of the checks involving uid_luid could be simplified if regscan
3696 was rerun in loop_optimize whenever a register was added or moved.
3697 Also, some of the optimizations could be a little less conservative. */
3699 /* Scan the loop body and call FNCALL for each insn. In the addition to the
3700 LOOP and INSN parameters pass MAYBE_MULTIPLE and NOT_EVERY_ITERATION to the
3703 NOT_EVERY_ITERATION if current insn is not executed at least once for every
3704 loop iteration except for the last one.
3706 MAYBE_MULTIPLE is 1 if current insn may be executed more than once for every
3710 for_each_insn_in_loop (loop, fncall)
3712 loop_insn_callback fncall;
3714 /* This is 1 if current insn is not executed at least once for every loop
3716 int not_every_iteration = 0;
3717 int maybe_multiple = 0;
3718 int past_loop_latch = 0;
3722 /* If loop_scan_start points to the loop exit test, we have to be wary of
3723 subversive use of gotos inside expression statements. */
3724 if (prev_nonnote_insn (loop->scan_start) != prev_nonnote_insn (loop->start))
3725 maybe_multiple = back_branch_in_range_p (loop, loop->scan_start);
3727 /* Scan through loop to find all possible bivs. */
3729 for (p = next_insn_in_loop (loop, loop->scan_start);
3731 p = next_insn_in_loop (loop, p))
3733 p = fncall (loop, p, not_every_iteration, maybe_multiple);
3735 /* Past CODE_LABEL, we get to insns that may be executed multiple
3736 times. The only way we can be sure that they can't is if every
3737 jump insn between here and the end of the loop either
3738 returns, exits the loop, is a jump to a location that is still
3739 behind the label, or is a jump to the loop start. */
3741 if (GET_CODE (p) == CODE_LABEL)
3749 insn = NEXT_INSN (insn);
3750 if (insn == loop->scan_start)
3752 if (insn == loop->end)
3758 if (insn == loop->scan_start)
3762 if (GET_CODE (insn) == JUMP_INSN
3763 && GET_CODE (PATTERN (insn)) != RETURN
3764 && (!condjump_p (insn)
3765 || (JUMP_LABEL (insn) != 0
3766 && JUMP_LABEL (insn) != loop->scan_start
3767 && !loop_insn_first_p (p, JUMP_LABEL (insn)))))
3775 /* Past a jump, we get to insns for which we can't count
3776 on whether they will be executed during each iteration. */
3777 /* This code appears twice in strength_reduce. There is also similar
3778 code in scan_loop. */
3779 if (GET_CODE (p) == JUMP_INSN
3780 /* If we enter the loop in the middle, and scan around to the
3781 beginning, don't set not_every_iteration for that.
3782 This can be any kind of jump, since we want to know if insns
3783 will be executed if the loop is executed. */
3784 && !(JUMP_LABEL (p) == loop->top
3785 && ((NEXT_INSN (NEXT_INSN (p)) == loop->end && simplejump_p (p))
3786 || (NEXT_INSN (p) == loop->end && condjump_p (p)))))
3790 /* If this is a jump outside the loop, then it also doesn't
3791 matter. Check to see if the target of this branch is on the
3792 loop->exits_labels list. */
3794 for (label = loop->exit_labels; label; label = LABEL_NEXTREF (label))
3795 if (XEXP (label, 0) == JUMP_LABEL (p))
3799 not_every_iteration = 1;
3802 else if (GET_CODE (p) == NOTE)
3804 /* At the virtual top of a converted loop, insns are again known to
3805 be executed each iteration: logically, the loop begins here
3806 even though the exit code has been duplicated.
3808 Insns are also again known to be executed each iteration at
3809 the LOOP_CONT note. */
3810 if ((NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP
3811 || NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_CONT)
3813 not_every_iteration = 0;
3814 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
3816 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END)
3820 /* Note if we pass a loop latch. If we do, then we can not clear
3821 NOT_EVERY_ITERATION below when we pass the last CODE_LABEL in
3822 a loop since a jump before the last CODE_LABEL may have started
3823 a new loop iteration.
3825 Note that LOOP_TOP is only set for rotated loops and we need
3826 this check for all loops, so compare against the CODE_LABEL
3827 which immediately follows LOOP_START. */
3828 if (GET_CODE (p) == JUMP_INSN
3829 && JUMP_LABEL (p) == NEXT_INSN (loop->start))
3830 past_loop_latch = 1;
3832 /* Unlike in the code motion pass where MAYBE_NEVER indicates that
3833 an insn may never be executed, NOT_EVERY_ITERATION indicates whether
3834 or not an insn is known to be executed each iteration of the
3835 loop, whether or not any iterations are known to occur.
3837 Therefore, if we have just passed a label and have no more labels
3838 between here and the test insn of the loop, and we have not passed
3839 a jump to the top of the loop, then we know these insns will be
3840 executed each iteration. */
3842 if (not_every_iteration
3844 && GET_CODE (p) == CODE_LABEL
3845 && no_labels_between_p (p, loop->end)
3846 && loop_insn_first_p (p, loop->cont))
3847 not_every_iteration = 0;
3851 /* Perform strength reduction and induction variable elimination.
3853 Pseudo registers created during this function will be beyond the last
3854 valid index in several tables including n_times_set and regno_last_uid.
3855 This does not cause a problem here, because the added registers cannot be
3856 givs outside of their loop, and hence will never be reconsidered.
3857 But scan_loop must check regnos to make sure they are in bounds. */
3860 strength_reduce (loop, insn_count, unroll_p, bct_p)
3863 int unroll_p, bct_p ATTRIBUTE_UNUSED;
3866 /* Temporary list pointers for traversing loop_iv_list. */
3867 struct iv_class *bl, **backbl;
3868 struct loop_info *loop_info = LOOP_INFO (loop);
3869 /* Ratio of extra register life span we can justify
3870 for saving an instruction. More if loop doesn't call subroutines
3871 since in that case saving an insn makes more difference
3872 and more registers are available. */
3873 /* ??? could set this to last value of threshold in move_movables */
3874 int threshold = (loop_info->has_call ? 1 : 2) * (3 + n_non_fixed_regs);
3875 /* Map of pseudo-register replacements. */
3876 rtx *reg_map = NULL;
3880 rtx end_insert_before;
3881 int n_extra_increment;
3882 int unrolled_insn_copies = 0;
3883 rtx loop_start = loop->start;
3884 rtx loop_end = loop->end;
3885 rtx loop_scan_start = loop->scan_start;
3887 VARRAY_INT_INIT (reg_iv_type, max_reg_before_loop, "reg_iv_type");
3888 VARRAY_GENERIC_PTR_INIT (reg_iv_info, max_reg_before_loop, "reg_iv_info");
3889 reg_biv_class = (struct iv_class **)
3890 xcalloc (max_reg_before_loop, sizeof (struct iv_class *));
3893 addr_placeholder = gen_reg_rtx (Pmode);
3895 /* Save insn immediately after the loop_end. Insns inserted after loop_end
3896 must be put before this insn, so that they will appear in the right
3897 order (i.e. loop order).
3899 If loop_end is the end of the current function, then emit a
3900 NOTE_INSN_DELETED after loop_end and set end_insert_before to the
3902 if (NEXT_INSN (loop_end) != 0)
3903 end_insert_before = NEXT_INSN (loop_end);
3905 end_insert_before = emit_note_after (NOTE_INSN_DELETED, loop_end);
3907 for_each_insn_in_loop (loop, check_insn_for_bivs);
3909 /* Scan loop_iv_list to remove all regs that proved not to be bivs.
3910 Make a sanity check against n_times_set. */
3911 for (backbl = &loop_iv_list, bl = *backbl; bl; bl = bl->next)
3915 if (REG_IV_TYPE (bl->regno) != BASIC_INDUCT
3916 /* Above happens if register modified by subreg, etc. */
3917 /* Make sure it is not recognized as a basic induction var: */
3918 || VARRAY_INT (n_times_set, bl->regno) != bl->biv_count
3919 /* If never incremented, it is invariant that we decided not to
3920 move. So leave it alone. */
3921 || ! bl->incremented)
3923 else if (bl->biv_count > 1)
3925 /* ??? If we have multiple increments for this BIV, and any of
3926 them take multiple insns to perform the increment, drop the
3927 BIV, since the bit below that converts the extra increments
3928 into GIVs can't handle the multiple insn increment. */
3930 struct induction *v;
3931 for (v = bl->biv; v ; v = v->next_iv)
3932 if (v->multi_insn_incr)
3938 if (loop_dump_stream)
3939 fprintf (loop_dump_stream, "Reg %d: biv discarded, %s\n",
3941 (REG_IV_TYPE (bl->regno) != BASIC_INDUCT
3942 ? "not induction variable"
3943 : (! bl->incremented ? "never incremented"
3946 REG_IV_TYPE (bl->regno) = NOT_BASIC_INDUCT;
3953 if (loop_dump_stream)
3954 fprintf (loop_dump_stream, "Reg %d: biv verified\n", bl->regno);
3958 /* Exit if there are no bivs. */
3961 /* Can still unroll the loop anyways, but indicate that there is no
3962 strength reduction info available. */
3964 unroll_loop (loop, insn_count, end_insert_before, 0);
3969 /* Find initial value for each biv by searching backwards from loop_start,
3970 halting at first label. Also record any test condition. */
3973 for (p = loop_start; p && GET_CODE (p) != CODE_LABEL; p = PREV_INSN (p))
3977 if (GET_CODE (p) == CALL_INSN)
3981 note_stores (PATTERN (p), record_initial, NULL);
3983 /* Record any test of a biv that branches around the loop if no store
3984 between it and the start of loop. We only care about tests with
3985 constants and registers and only certain of those. */
3986 if (GET_CODE (p) == JUMP_INSN
3987 && JUMP_LABEL (p) != 0
3988 && next_real_insn (JUMP_LABEL (p)) == next_real_insn (loop_end)
3989 && (test = get_condition_for_loop (loop, p)) != 0
3990 && GET_CODE (XEXP (test, 0)) == REG
3991 && REGNO (XEXP (test, 0)) < max_reg_before_loop
3992 && (bl = reg_biv_class[REGNO (XEXP (test, 0))]) != 0
3993 && valid_initial_value_p (XEXP (test, 1), p, call_seen, loop_start)
3994 && bl->init_insn == 0)
3996 /* If an NE test, we have an initial value! */
3997 if (GET_CODE (test) == NE)
4000 bl->init_set = gen_rtx_SET (VOIDmode,
4001 XEXP (test, 0), XEXP (test, 1));
4004 bl->initial_test = test;
4008 /* Look at the each biv and see if we can say anything better about its
4009 initial value from any initializing insns set up above. (This is done
4010 in two passes to avoid missing SETs in a PARALLEL.) */
4011 for (backbl = &loop_iv_list; (bl = *backbl); backbl = &bl->next)
4016 if (! bl->init_insn)
4019 /* IF INIT_INSN has a REG_EQUAL or REG_EQUIV note and the value
4020 is a constant, use the value of that. */
4021 if (((note = find_reg_note (bl->init_insn, REG_EQUAL, 0)) != NULL
4022 && CONSTANT_P (XEXP (note, 0)))
4023 || ((note = find_reg_note (bl->init_insn, REG_EQUIV, 0)) != NULL
4024 && CONSTANT_P (XEXP (note, 0))))
4025 src = XEXP (note, 0);
4027 src = SET_SRC (bl->init_set);
4029 if (loop_dump_stream)
4030 fprintf (loop_dump_stream,
4031 "Biv %d initialized at insn %d: initial value ",
4032 bl->regno, INSN_UID (bl->init_insn));
4034 if ((GET_MODE (src) == GET_MODE (regno_reg_rtx[bl->regno])
4035 || GET_MODE (src) == VOIDmode)
4036 && valid_initial_value_p (src, bl->init_insn, call_seen, loop_start))
4038 bl->initial_value = src;
4040 if (loop_dump_stream)
4042 if (GET_CODE (src) == CONST_INT)
4044 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (src));
4045 fputc ('\n', loop_dump_stream);
4049 print_rtl (loop_dump_stream, src);
4050 fprintf (loop_dump_stream, "\n");
4056 struct iv_class *bl2 = 0;
4057 rtx increment = NULL_RTX;
4059 /* Biv initial value is not a simple move. If it is the sum of
4060 another biv and a constant, check if both bivs are incremented
4061 in lockstep. Then we are actually looking at a giv.
4062 For simplicity, we only handle the case where there is but a
4063 single increment, and the register is not used elsewhere. */
4064 if (bl->biv_count == 1
4065 && bl->regno < max_reg_before_loop
4066 && uid_luid[REGNO_LAST_UID (bl->regno)] < INSN_LUID (loop_end)
4067 && GET_CODE (src) == PLUS
4068 && GET_CODE (XEXP (src, 0)) == REG
4069 && CONSTANT_P (XEXP (src, 1))
4070 && ((increment = biv_total_increment (bl)) != NULL_RTX))
4072 unsigned int regno = REGNO (XEXP (src, 0));
4074 for (bl2 = loop_iv_list; bl2; bl2 = bl2->next)
4075 if (bl2->regno == regno)
4079 /* Now, can we transform this biv into a giv? */
4081 && bl2->biv_count == 1
4082 && rtx_equal_p (increment, biv_total_increment (bl2))
4083 /* init_insn is only set to insns that are before loop_start
4084 without any intervening labels. */
4085 && ! reg_set_between_p (bl2->biv->src_reg,
4086 PREV_INSN (bl->init_insn), loop_start)
4087 /* The register from BL2 must be set before the register from
4088 BL is set, or we must be able to move the latter set after
4089 the former set. Currently there can't be any labels
4090 in-between when biv_total_increment returns nonzero both times
4091 but we test it here in case some day some real cfg analysis
4092 gets used to set always_computable. */
4093 && (loop_insn_first_p (bl2->biv->insn, bl->biv->insn)
4094 ? no_labels_between_p (bl2->biv->insn, bl->biv->insn)
4095 : (! reg_used_between_p (bl->biv->src_reg, bl->biv->insn,
4097 && no_jumps_between_p (bl->biv->insn, bl2->biv->insn)))
4098 && validate_change (bl->biv->insn,
4099 &SET_SRC (single_set (bl->biv->insn)),
4102 rtx dominator = loop->cont_dominator;
4103 rtx giv = bl->biv->src_reg;
4104 rtx giv_insn = bl->biv->insn;
4105 rtx after_giv = NEXT_INSN (giv_insn);
4107 if (loop_dump_stream)
4108 fprintf (loop_dump_stream, "is giv of biv %d\n", bl2->regno);
4109 /* Let this giv be discovered by the generic code. */
4110 REG_IV_TYPE (bl->regno) = UNKNOWN_INDUCT;
4111 reg_biv_class[bl->regno] = (struct iv_class *) NULL_PTR;
4112 /* We can get better optimization if we can move the giv setting
4113 before the first giv use. */
4115 && ! loop_insn_first_p (dominator, loop_scan_start)
4116 && ! reg_set_between_p (bl2->biv->src_reg, loop_start,
4118 && ! reg_used_between_p (giv, loop_start, dominator)
4119 && ! reg_used_between_p (giv, giv_insn, loop_end))
4124 for (next = NEXT_INSN (dominator); ; next = NEXT_INSN (next))
4127 && (reg_mentioned_p (giv, PATTERN (next))
4128 || reg_set_p (bl2->biv->src_reg, next)))
4129 || GET_CODE (next) == JUMP_INSN)
4133 || ! sets_cc0_p (PATTERN (next)))
4137 if (loop_dump_stream)
4138 fprintf (loop_dump_stream, "move after insn %d\n",
4139 INSN_UID (dominator));
4140 /* Avoid problems with luids by actually moving the insn
4141 and adjusting all luids in the range. */
4142 reorder_insns (giv_insn, giv_insn, dominator);
4143 for (p = dominator; INSN_UID (p) >= max_uid_for_loop; )
4145 compute_luids (giv_insn, after_giv, INSN_LUID (p));
4146 /* If the only purpose of the init insn is to initialize
4147 this giv, delete it. */
4148 if (single_set (bl->init_insn)
4149 && ! reg_used_between_p (giv, bl->init_insn, loop_start))
4150 delete_insn (bl->init_insn);
4152 else if (! loop_insn_first_p (bl2->biv->insn, bl->biv->insn))
4154 rtx p = PREV_INSN (giv_insn);
4155 while (INSN_UID (p) >= max_uid_for_loop)
4157 reorder_insns (giv_insn, giv_insn, bl2->biv->insn);
4158 compute_luids (after_giv, NEXT_INSN (giv_insn),
4161 /* Remove this biv from the chain. */
4165 /* If we can't make it a giv,
4166 let biv keep initial value of "itself". */
4167 else if (loop_dump_stream)
4168 fprintf (loop_dump_stream, "is complex\n");
4172 /* If a biv is unconditionally incremented several times in a row, convert
4173 all but the last increment into a giv. */
4175 /* Get an upper bound for the number of registers
4176 we might have after all bivs have been processed. */
4177 first_increment_giv = max_reg_num ();
4178 for (n_extra_increment = 0, bl = loop_iv_list; bl; bl = bl->next)
4179 n_extra_increment += bl->biv_count - 1;
4181 /* If the loop contains volatile memory references do not allow any
4182 replacements to take place, since this could loose the volatile
4184 if (n_extra_increment && ! loop_info->has_volatile)
4186 unsigned int nregs = first_increment_giv + n_extra_increment;
4188 /* Reallocate reg_iv_type and reg_iv_info. */
4189 VARRAY_GROW (reg_iv_type, nregs);
4190 VARRAY_GROW (reg_iv_info, nregs);
4192 for (bl = loop_iv_list; bl; bl = bl->next)
4194 struct induction **vp, *v, *next;
4195 int biv_dead_after_loop = 0;
4197 /* The biv increments lists are in reverse order. Fix this
4199 for (v = bl->biv, bl->biv = 0; v; v = next)
4202 v->next_iv = bl->biv;
4206 /* We must guard against the case that an early exit between v->insn
4207 and next->insn leaves the biv live after the loop, since that
4208 would mean that we'd be missing an increment for the final
4209 value. The following test to set biv_dead_after_loop is like
4210 the first part of the test to set bl->eliminable.
4211 We don't check here if we can calculate the final value, since
4212 this can't succeed if we already know that there is a jump
4213 between v->insn and next->insn, yet next->always_executed is
4214 set and next->maybe_multiple is cleared. Such a combination
4215 implies that the jump destination is outside the loop.
4216 If we want to make this check more sophisticated, we should
4217 check each branch between v->insn and next->insn individually
4218 to see if the biv is dead at its destination. */
4220 if (uid_luid[REGNO_LAST_UID (bl->regno)] < INSN_LUID (loop_end)
4222 && INSN_UID (bl->init_insn) < max_uid_for_loop
4223 && (uid_luid[REGNO_FIRST_UID (bl->regno)]
4224 >= INSN_LUID (bl->init_insn))
4225 #ifdef HAVE_decrement_and_branch_until_zero
4228 && ! reg_mentioned_p (bl->biv->dest_reg, SET_SRC (bl->init_set)))
4229 biv_dead_after_loop = 1;
4231 for (vp = &bl->biv, next = *vp; v = next, next = v->next_iv;)
4233 HOST_WIDE_INT offset;
4234 rtx set, add_val, old_reg, dest_reg, last_use_insn, note;
4235 int old_regno, new_regno;
4237 if (! v->always_executed
4238 || v->maybe_multiple
4239 || GET_CODE (v->add_val) != CONST_INT
4240 || ! next->always_executed
4241 || next->maybe_multiple
4242 || ! CONSTANT_P (next->add_val)
4243 || v->mult_val != const1_rtx
4244 || next->mult_val != const1_rtx
4245 || ! (biv_dead_after_loop
4246 || no_jumps_between_p (v->insn, next->insn)))
4251 offset = INTVAL (v->add_val);
4252 set = single_set (v->insn);
4253 add_val = plus_constant (next->add_val, offset);
4254 old_reg = v->dest_reg;
4255 dest_reg = gen_reg_rtx (v->mode);
4257 /* Unlike reg_iv_type / reg_iv_info, the other three arrays
4258 have been allocated with some slop space, so we may not
4259 actually need to reallocate them. If we do, the following
4260 if statement will be executed just once in this loop. */
4261 if ((unsigned) max_reg_num () > n_times_set->num_elements)
4263 /* Grow all the remaining arrays. */
4264 VARRAY_GROW (set_in_loop, nregs);
4265 VARRAY_GROW (n_times_set, nregs);
4266 VARRAY_GROW (may_not_optimize, nregs);
4267 VARRAY_GROW (reg_single_usage, nregs);
4270 if (! validate_change (next->insn, next->location, add_val, 0))
4276 /* Here we can try to eliminate the increment by combining
4277 it into the uses. */
4279 /* Set last_use_insn so that we can check against it. */
4281 for (last_use_insn = v->insn, p = NEXT_INSN (v->insn);
4283 p = next_insn_in_loop (loop, p))
4287 if (reg_mentioned_p (old_reg, PATTERN (p)))
4293 /* If we can't get the LUIDs for the insns, we can't
4294 calculate the lifetime. This is likely from unrolling
4295 of an inner loop, so there is little point in making this
4296 a DEST_REG giv anyways. */
4297 if (INSN_UID (v->insn) >= max_uid_for_loop
4298 || INSN_UID (last_use_insn) >= max_uid_for_loop
4299 || ! validate_change (v->insn, &SET_DEST (set), dest_reg, 0))
4301 /* Change the increment at NEXT back to what it was. */
4302 if (! validate_change (next->insn, next->location,
4308 next->add_val = add_val;
4309 v->dest_reg = dest_reg;
4310 v->giv_type = DEST_REG;
4311 v->location = &SET_SRC (set);
4313 v->combined_with = 0;
4315 v->derive_adjustment = 0;
4321 v->auto_inc_opt = 0;
4324 v->derived_from = 0;
4325 v->always_computable = 1;
4326 v->always_executed = 1;
4328 v->no_const_addval = 0;
4330 old_regno = REGNO (old_reg);
4331 new_regno = REGNO (dest_reg);
4332 VARRAY_INT (set_in_loop, old_regno)--;
4333 VARRAY_INT (set_in_loop, new_regno) = 1;
4334 VARRAY_INT (n_times_set, old_regno)--;
4335 VARRAY_INT (n_times_set, new_regno) = 1;
4336 VARRAY_CHAR (may_not_optimize, new_regno) = 0;
4338 REG_IV_TYPE (new_regno) = GENERAL_INDUCT;
4339 REG_IV_INFO (new_regno) = v;
4341 /* If next_insn has a REG_EQUAL note that mentiones OLD_REG,
4342 it must be replaced. */
4343 note = find_reg_note (next->insn, REG_EQUAL, NULL_RTX);
4344 if (note && reg_mentioned_p (old_reg, XEXP (note, 0)))
4345 XEXP (note, 0) = copy_rtx (SET_SRC (single_set (next->insn)));
4347 /* Remove the increment from the list of biv increments,
4348 and record it as a giv. */
4351 v->next_iv = bl->giv;
4354 v->benefit = rtx_cost (SET_SRC (set), SET);
4355 bl->total_benefit += v->benefit;
4357 /* Now replace the biv with DEST_REG in all insns between
4358 the replaced increment and the next increment, and
4359 remember the last insn that needed a replacement. */
4360 for (last_use_insn = v->insn, p = NEXT_INSN (v->insn);
4362 p = next_insn_in_loop (loop, p))
4366 if (GET_RTX_CLASS (GET_CODE (p)) != 'i')
4368 if (reg_mentioned_p (old_reg, PATTERN (p)))
4371 if (! validate_replace_rtx (old_reg, dest_reg, p))
4374 for (note = REG_NOTES (p); note; note = XEXP (note, 1))
4376 if (GET_CODE (note) == EXPR_LIST)
4378 = replace_rtx (XEXP (note, 0), old_reg, dest_reg);
4382 v->last_use = last_use_insn;
4383 v->lifetime = INSN_LUID (last_use_insn) - INSN_LUID (v->insn);
4384 /* If the lifetime is zero, it means that this register is really
4385 a dead store. So mark this as a giv that can be ignored.
4386 This will not prevent the biv from being eliminated. */
4387 if (v->lifetime == 0)
4390 if (loop_dump_stream)
4391 fprintf (loop_dump_stream,
4392 "Increment %d of biv %d converted to giv %d.\n",
4393 INSN_UID (v->insn), old_regno, new_regno);
4397 last_increment_giv = max_reg_num () - 1;
4399 /* Search the loop for general induction variables. */
4401 for_each_insn_in_loop (loop, check_insn_for_givs);
4403 /* Try to calculate and save the number of loop iterations. This is
4404 set to zero if the actual number can not be calculated. This must
4405 be called after all giv's have been identified, since otherwise it may
4406 fail if the iteration variable is a giv. */
4408 loop_iterations (loop);
4410 /* Now for each giv for which we still don't know whether or not it is
4411 replaceable, check to see if it is replaceable because its final value
4412 can be calculated. This must be done after loop_iterations is called,
4413 so that final_giv_value will work correctly. */
4415 for (bl = loop_iv_list; bl; bl = bl->next)
4417 struct induction *v;
4419 for (v = bl->giv; v; v = v->next_iv)
4420 if (! v->replaceable && ! v->not_replaceable)
4421 check_final_value (loop, v);
4424 /* Try to prove that the loop counter variable (if any) is always
4425 nonnegative; if so, record that fact with a REG_NONNEG note
4426 so that "decrement and branch until zero" insn can be used. */
4427 check_dbra_loop (loop, insn_count);
4429 /* Create reg_map to hold substitutions for replaceable giv regs.
4430 Some givs might have been made from biv increments, so look at
4431 reg_iv_type for a suitable size. */
4432 reg_map_size = reg_iv_type->num_elements;
4433 reg_map = (rtx *) xcalloc (reg_map_size, sizeof (rtx));
4435 /* Examine each iv class for feasibility of strength reduction/induction
4436 variable elimination. */
4438 for (bl = loop_iv_list; bl; bl = bl->next)
4440 struct induction *v;
4443 rtx final_value = 0;
4446 /* Test whether it will be possible to eliminate this biv
4447 provided all givs are reduced. This is possible if either
4448 the reg is not used outside the loop, or we can compute
4449 what its final value will be.
4451 For architectures with a decrement_and_branch_until_zero insn,
4452 don't do this if we put a REG_NONNEG note on the endtest for
4455 /* Compare against bl->init_insn rather than loop_start.
4456 We aren't concerned with any uses of the biv between
4457 init_insn and loop_start since these won't be affected
4458 by the value of the biv elsewhere in the function, so
4459 long as init_insn doesn't use the biv itself.
4460 March 14, 1989 -- self@bayes.arc.nasa.gov */
4462 if ((uid_luid[REGNO_LAST_UID (bl->regno)] < INSN_LUID (loop_end)
4464 && INSN_UID (bl->init_insn) < max_uid_for_loop
4465 && uid_luid[REGNO_FIRST_UID (bl->regno)] >= INSN_LUID (bl->init_insn)
4466 #ifdef HAVE_decrement_and_branch_until_zero
4469 && ! reg_mentioned_p (bl->biv->dest_reg, SET_SRC (bl->init_set)))
4470 || ((final_value = final_biv_value (loop, bl))
4471 #ifdef HAVE_decrement_and_branch_until_zero
4475 bl->eliminable = maybe_eliminate_biv (loop, bl, 0, threshold,
4479 if (loop_dump_stream)
4481 fprintf (loop_dump_stream,
4482 "Cannot eliminate biv %d.\n",
4484 fprintf (loop_dump_stream,
4485 "First use: insn %d, last use: insn %d.\n",
4486 REGNO_FIRST_UID (bl->regno),
4487 REGNO_LAST_UID (bl->regno));
4491 /* Combine all giv's for this iv_class. */
4494 /* This will be true at the end, if all givs which depend on this
4495 biv have been strength reduced.
4496 We can't (currently) eliminate the biv unless this is so. */
4499 /* Check each giv in this class to see if we will benefit by reducing
4500 it. Skip giv's combined with others. */
4501 for (v = bl->giv; v; v = v->next_iv)
4503 struct induction *tv;
4505 if (v->ignore || v->same)
4508 benefit = v->benefit;
4510 /* Reduce benefit if not replaceable, since we will insert
4511 a move-insn to replace the insn that calculates this giv.
4512 Don't do this unless the giv is a user variable, since it
4513 will often be marked non-replaceable because of the duplication
4514 of the exit code outside the loop. In such a case, the copies
4515 we insert are dead and will be deleted. So they don't have
4516 a cost. Similar situations exist. */
4517 /* ??? The new final_[bg]iv_value code does a much better job
4518 of finding replaceable giv's, and hence this code may no longer
4520 if (! v->replaceable && ! bl->eliminable
4521 && REG_USERVAR_P (v->dest_reg))
4522 benefit -= copy_cost;
4524 /* Decrease the benefit to count the add-insns that we will
4525 insert to increment the reduced reg for the giv. */
4526 benefit -= add_cost * bl->biv_count;
4528 /* Decide whether to strength-reduce this giv or to leave the code
4529 unchanged (recompute it from the biv each time it is used).
4530 This decision can be made independently for each giv. */
4533 /* Attempt to guess whether autoincrement will handle some of the
4534 new add insns; if so, increase BENEFIT (undo the subtraction of
4535 add_cost that was done above). */
4536 if (v->giv_type == DEST_ADDR
4537 && GET_CODE (v->mult_val) == CONST_INT)
4539 if (HAVE_POST_INCREMENT
4540 && INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
4541 benefit += add_cost * bl->biv_count;
4542 else if (HAVE_PRE_INCREMENT
4543 && INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
4544 benefit += add_cost * bl->biv_count;
4545 else if (HAVE_POST_DECREMENT
4546 && -INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
4547 benefit += add_cost * bl->biv_count;
4548 else if (HAVE_PRE_DECREMENT
4549 && -INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
4550 benefit += add_cost * bl->biv_count;
4554 /* If an insn is not to be strength reduced, then set its ignore
4555 flag, and clear all_reduced. */
4557 /* A giv that depends on a reversed biv must be reduced if it is
4558 used after the loop exit, otherwise, it would have the wrong
4559 value after the loop exit. To make it simple, just reduce all
4560 of such giv's whether or not we know they are used after the loop
4563 if ( ! flag_reduce_all_givs && v->lifetime * threshold * benefit < insn_count
4566 if (loop_dump_stream)
4567 fprintf (loop_dump_stream,
4568 "giv of insn %d not worth while, %d vs %d.\n",
4570 v->lifetime * threshold * benefit, insn_count);
4576 /* Check that we can increment the reduced giv without a
4577 multiply insn. If not, reject it. */
4579 for (tv = bl->biv; tv; tv = tv->next_iv)
4580 if (tv->mult_val == const1_rtx
4581 && ! product_cheap_p (tv->add_val, v->mult_val))
4583 if (loop_dump_stream)
4584 fprintf (loop_dump_stream,
4585 "giv of insn %d: would need a multiply.\n",
4586 INSN_UID (v->insn));
4594 /* Check for givs whose first use is their definition and whose
4595 last use is the definition of another giv. If so, it is likely
4596 dead and should not be used to derive another giv nor to
4598 for (v = bl->giv; v; v = v->next_iv)
4601 || (v->same && v->same->ignore))
4606 struct induction *v1;
4608 for (v1 = bl->giv; v1; v1 = v1->next_iv)
4609 if (v->last_use == v1->insn)
4612 else if (v->giv_type == DEST_REG
4613 && REGNO_FIRST_UID (REGNO (v->dest_reg)) == INSN_UID (v->insn))
4615 struct induction *v1;
4617 for (v1 = bl->giv; v1; v1 = v1->next_iv)
4618 if (REGNO_LAST_UID (REGNO (v->dest_reg)) == INSN_UID (v1->insn))
4623 /* Now that we know which givs will be reduced, try to rearrange the
4624 combinations to reduce register pressure.
4625 recombine_givs calls find_life_end, which needs reg_iv_type and
4626 reg_iv_info to be valid for all pseudos. We do the necessary
4627 reallocation here since it allows to check if there are still
4628 more bivs to process. */
4629 nregs = max_reg_num ();
4630 if (nregs > reg_iv_type->num_elements)
4632 /* If there are still more bivs to process, allocate some slack
4633 space so that we're not constantly reallocating these arrays. */
4636 /* Reallocate reg_iv_type and reg_iv_info. */
4637 VARRAY_GROW (reg_iv_type, nregs);
4638 VARRAY_GROW (reg_iv_info, nregs);
4640 recombine_givs (loop, bl, unroll_p);
4642 /* Reduce each giv that we decided to reduce. */
4644 for (v = bl->giv; v; v = v->next_iv)
4646 struct induction *tv;
4647 if (! v->ignore && v->same == 0)
4649 int auto_inc_opt = 0;
4651 /* If the code for derived givs immediately below has already
4652 allocated a new_reg, we must keep it. */
4654 v->new_reg = gen_reg_rtx (v->mode);
4656 if (v->derived_from)
4658 struct induction *d = v->derived_from;
4660 /* In case d->dest_reg is not replaceable, we have
4661 to replace it in v->insn now. */
4663 d->new_reg = gen_reg_rtx (d->mode);
4665 = replace_rtx (PATTERN (v->insn), d->dest_reg, d->new_reg);
4667 = replace_rtx (PATTERN (v->insn), v->dest_reg, v->new_reg);
4668 /* For each place where the biv is incremented, add an
4669 insn to set the new, reduced reg for the giv.
4670 We used to do this only for biv_count != 1, but
4671 this fails when there is a giv after a single biv
4672 increment, e.g. when the last giv was expressed as
4674 for (tv = bl->biv; tv; tv = tv->next_iv)
4676 /* We always emit reduced giv increments before the
4677 biv increment when bl->biv_count != 1. So by
4678 emitting the add insns for derived givs after the
4679 biv increment, they pick up the updated value of
4681 If the reduced giv is processed with
4682 auto_inc_opt == 1, then it is incremented earlier
4683 than the biv, hence we'll still pick up the right
4685 If it's processed with auto_inc_opt == -1,
4686 that implies that the biv increment is before the
4687 first reduced giv's use. The derived giv's lifetime
4688 is after the reduced giv's lifetime, hence in this
4689 case, the biv increment doesn't matter. */
4690 emit_insn_after (copy_rtx (PATTERN (v->insn)), tv->insn);
4696 /* If the target has auto-increment addressing modes, and
4697 this is an address giv, then try to put the increment
4698 immediately after its use, so that flow can create an
4699 auto-increment addressing mode. */
4700 if (v->giv_type == DEST_ADDR && bl->biv_count == 1
4701 && bl->biv->always_executed && ! bl->biv->maybe_multiple
4702 /* We don't handle reversed biv's because bl->biv->insn
4703 does not have a valid INSN_LUID. */
4705 && v->always_executed && ! v->maybe_multiple
4706 && INSN_UID (v->insn) < max_uid_for_loop)
4708 /* If other giv's have been combined with this one, then
4709 this will work only if all uses of the other giv's occur
4710 before this giv's insn. This is difficult to check.
4712 We simplify this by looking for the common case where
4713 there is one DEST_REG giv, and this giv's insn is the
4714 last use of the dest_reg of that DEST_REG giv. If the
4715 increment occurs after the address giv, then we can
4716 perform the optimization. (Otherwise, the increment
4717 would have to go before other_giv, and we would not be
4718 able to combine it with the address giv to get an
4719 auto-inc address.) */
4720 if (v->combined_with)
4722 struct induction *other_giv = 0;
4724 for (tv = bl->giv; tv; tv = tv->next_iv)
4732 if (! tv && other_giv
4733 && REGNO (other_giv->dest_reg) < max_reg_before_loop
4734 && (REGNO_LAST_UID (REGNO (other_giv->dest_reg))
4735 == INSN_UID (v->insn))
4736 && INSN_LUID (v->insn) < INSN_LUID (bl->biv->insn))
4739 /* Check for case where increment is before the address
4740 giv. Do this test in "loop order". */
4741 else if ((INSN_LUID (v->insn) > INSN_LUID (bl->biv->insn)
4742 && (INSN_LUID (v->insn) < INSN_LUID (loop_scan_start)
4743 || (INSN_LUID (bl->biv->insn)
4744 > INSN_LUID (loop_scan_start))))
4745 || (INSN_LUID (v->insn) < INSN_LUID (loop_scan_start)
4746 && (INSN_LUID (loop_scan_start)
4747 < INSN_LUID (bl->biv->insn))))
4756 /* We can't put an insn immediately after one setting
4757 cc0, or immediately before one using cc0. */
4758 if ((auto_inc_opt == 1 && sets_cc0_p (PATTERN (v->insn)))
4759 || (auto_inc_opt == -1
4760 && (prev = prev_nonnote_insn (v->insn)) != 0
4761 && GET_RTX_CLASS (GET_CODE (prev)) == 'i'
4762 && sets_cc0_p (PATTERN (prev))))
4768 v->auto_inc_opt = 1;
4772 /* For each place where the biv is incremented, add an insn
4773 to increment the new, reduced reg for the giv. */
4774 for (tv = bl->biv; tv; tv = tv->next_iv)
4779 insert_before = tv->insn;
4780 else if (auto_inc_opt == 1)
4781 insert_before = NEXT_INSN (v->insn);
4783 insert_before = v->insn;
4785 if (tv->mult_val == const1_rtx)
4786 emit_iv_add_mult (tv->add_val, v->mult_val,
4787 v->new_reg, v->new_reg, insert_before);
4788 else /* tv->mult_val == const0_rtx */
4789 /* A multiply is acceptable here
4790 since this is presumed to be seldom executed. */
4791 emit_iv_add_mult (tv->add_val, v->mult_val,
4792 v->add_val, v->new_reg, insert_before);
4795 /* Add code at loop start to initialize giv's reduced reg. */
4797 emit_iv_add_mult (bl->initial_value, v->mult_val,
4798 v->add_val, v->new_reg, loop_start);
4802 /* Rescan all givs. If a giv is the same as a giv not reduced, mark it
4805 For each giv register that can be reduced now: if replaceable,
4806 substitute reduced reg wherever the old giv occurs;
4807 else add new move insn "giv_reg = reduced_reg". */
4809 for (v = bl->giv; v; v = v->next_iv)
4811 if (v->same && v->same->ignore)
4817 /* Update expression if this was combined, in case other giv was
4820 v->new_reg = replace_rtx (v->new_reg,
4821 v->same->dest_reg, v->same->new_reg);
4823 if (v->giv_type == DEST_ADDR)
4824 /* Store reduced reg as the address in the memref where we found
4826 validate_change (v->insn, v->location, v->new_reg, 0);
4827 else if (v->replaceable)
4829 reg_map[REGNO (v->dest_reg)] = v->new_reg;
4832 /* I can no longer duplicate the original problem. Perhaps
4833 this is unnecessary now? */
4835 /* Replaceable; it isn't strictly necessary to delete the old
4836 insn and emit a new one, because v->dest_reg is now dead.
4838 However, especially when unrolling loops, the special
4839 handling for (set REG0 REG1) in the second cse pass may
4840 make v->dest_reg live again. To avoid this problem, emit
4841 an insn to set the original giv reg from the reduced giv.
4842 We can not delete the original insn, since it may be part
4843 of a LIBCALL, and the code in flow that eliminates dead
4844 libcalls will fail if it is deleted. */
4845 emit_insn_after (gen_move_insn (v->dest_reg, v->new_reg),
4851 /* Not replaceable; emit an insn to set the original giv reg from
4852 the reduced giv, same as above. */
4853 emit_insn_after (gen_move_insn (v->dest_reg, v->new_reg),
4857 /* When a loop is reversed, givs which depend on the reversed
4858 biv, and which are live outside the loop, must be set to their
4859 correct final value. This insn is only needed if the giv is
4860 not replaceable. The correct final value is the same as the
4861 value that the giv starts the reversed loop with. */
4862 if (bl->reversed && ! v->replaceable)
4863 emit_iv_add_mult (bl->initial_value, v->mult_val,
4864 v->add_val, v->dest_reg, end_insert_before);
4865 else if (v->final_value)
4869 /* If the loop has multiple exits, emit the insn before the
4870 loop to ensure that it will always be executed no matter
4871 how the loop exits. Otherwise, emit the insn after the loop,
4872 since this is slightly more efficient. */
4873 if (loop->exit_count)
4874 insert_before = loop_start;
4876 insert_before = end_insert_before;
4877 emit_insn_before (gen_move_insn (v->dest_reg, v->final_value),
4881 /* If the insn to set the final value of the giv was emitted
4882 before the loop, then we must delete the insn inside the loop
4883 that sets it. If this is a LIBCALL, then we must delete
4884 every insn in the libcall. Note, however, that
4885 final_giv_value will only succeed when there are multiple
4886 exits if the giv is dead at each exit, hence it does not
4887 matter that the original insn remains because it is dead
4889 /* Delete the insn inside the loop that sets the giv since
4890 the giv is now set before (or after) the loop. */
4891 delete_insn (v->insn);
4895 if (loop_dump_stream)
4897 fprintf (loop_dump_stream, "giv at %d reduced to ",
4898 INSN_UID (v->insn));
4899 print_rtl (loop_dump_stream, v->new_reg);
4900 fprintf (loop_dump_stream, "\n");
4904 /* All the givs based on the biv bl have been reduced if they
4907 /* For each giv not marked as maybe dead that has been combined with a
4908 second giv, clear any "maybe dead" mark on that second giv.
4909 v->new_reg will either be or refer to the register of the giv it
4912 Doing this clearing avoids problems in biv elimination where a
4913 giv's new_reg is a complex value that can't be put in the insn but
4914 the giv combined with (with a reg as new_reg) is marked maybe_dead.
4915 Since the register will be used in either case, we'd prefer it be
4916 used from the simpler giv. */
4918 for (v = bl->giv; v; v = v->next_iv)
4919 if (! v->maybe_dead && v->same)
4920 v->same->maybe_dead = 0;
4922 /* Try to eliminate the biv, if it is a candidate.
4923 This won't work if ! all_reduced,
4924 since the givs we planned to use might not have been reduced.
4926 We have to be careful that we didn't initially think we could eliminate
4927 this biv because of a giv that we now think may be dead and shouldn't
4928 be used as a biv replacement.
4930 Also, there is the possibility that we may have a giv that looks
4931 like it can be used to eliminate a biv, but the resulting insn
4932 isn't valid. This can happen, for example, on the 88k, where a
4933 JUMP_INSN can compare a register only with zero. Attempts to
4934 replace it with a compare with a constant will fail.
4936 Note that in cases where this call fails, we may have replaced some
4937 of the occurrences of the biv with a giv, but no harm was done in
4938 doing so in the rare cases where it can occur. */
4940 if (all_reduced == 1 && bl->eliminable
4941 && maybe_eliminate_biv (loop, bl, 1, threshold, insn_count))
4943 /* ?? If we created a new test to bypass the loop entirely,
4944 or otherwise drop straight in, based on this test, then
4945 we might want to rewrite it also. This way some later
4946 pass has more hope of removing the initialization of this
4949 /* If final_value != 0, then the biv may be used after loop end
4950 and we must emit an insn to set it just in case.
4952 Reversed bivs already have an insn after the loop setting their
4953 value, so we don't need another one. We can't calculate the
4954 proper final value for such a biv here anyways. */
4955 if (final_value != 0 && ! bl->reversed)
4959 /* If the loop has multiple exits, emit the insn before the
4960 loop to ensure that it will always be executed no matter
4961 how the loop exits. Otherwise, emit the insn after the
4962 loop, since this is slightly more efficient. */
4963 if (loop->exit_count)
4964 insert_before = loop_start;
4966 insert_before = end_insert_before;
4968 emit_insn_before (gen_move_insn (bl->biv->dest_reg, final_value),
4973 /* Delete all of the instructions inside the loop which set
4974 the biv, as they are all dead. If is safe to delete them,
4975 because an insn setting a biv will never be part of a libcall. */
4976 /* However, deleting them will invalidate the regno_last_uid info,
4977 so keeping them around is more convenient. Final_biv_value
4978 will only succeed when there are multiple exits if the biv
4979 is dead at each exit, hence it does not matter that the original
4980 insn remains, because it is dead anyways. */
4981 for (v = bl->biv; v; v = v->next_iv)
4982 delete_insn (v->insn);
4985 if (loop_dump_stream)
4986 fprintf (loop_dump_stream, "Reg %d: biv eliminated\n",
4991 /* Go through all the instructions in the loop, making all the
4992 register substitutions scheduled in REG_MAP. */
4994 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
4995 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
4996 || GET_CODE (p) == CALL_INSN)
4998 replace_regs (PATTERN (p), reg_map, reg_map_size, 0);
4999 replace_regs (REG_NOTES (p), reg_map, reg_map_size, 0);
5003 if (loop_info->n_iterations > 0)
5005 /* When we completely unroll a loop we will likely not need the increment
5006 of the loop BIV and we will not need the conditional branch at the
5008 unrolled_insn_copies = insn_count - 2;
5011 /* When we completely unroll a loop on a HAVE_cc0 machine we will not
5012 need the comparison before the conditional branch at the end of the
5014 unrolled_insn_copies -= 1;
5017 /* We'll need one copy for each loop iteration. */
5018 unrolled_insn_copies *= loop_info->n_iterations;
5020 /* A little slop to account for the ability to remove initialization
5021 code, better CSE, and other secondary benefits of completely
5022 unrolling some loops. */
5023 unrolled_insn_copies -= 1;
5025 /* Clamp the value. */
5026 if (unrolled_insn_copies < 0)
5027 unrolled_insn_copies = 0;
5030 /* Unroll loops from within strength reduction so that we can use the
5031 induction variable information that strength_reduce has already
5032 collected. Always unroll loops that would be as small or smaller
5033 unrolled than when rolled. */
5035 || (loop_info->n_iterations > 0
5036 && unrolled_insn_copies <= insn_count))
5037 unroll_loop (loop, insn_count, end_insert_before, 1);
5039 #ifdef HAVE_decrement_and_branch_on_count
5040 /* Instrument the loop with BCT insn. */
5041 if (HAVE_decrement_and_branch_on_count && bct_p
5042 && flag_branch_on_count_reg)
5044 #endif /* HAVE_decrement_and_branch_on_count */
5046 if (loop_dump_stream)
5047 fprintf (loop_dump_stream, "\n");
5050 VARRAY_FREE (reg_iv_type);
5051 VARRAY_FREE (reg_iv_info);
5052 free (reg_biv_class);
5057 /*Record all basic induction variables calculated in the insn. */
5059 check_insn_for_bivs (loop, p, not_every_iteration, maybe_multiple)
5062 int not_every_iteration;
5071 if (GET_CODE (p) == INSN
5072 && (set = single_set (p))
5073 && GET_CODE (SET_DEST (set)) == REG)
5075 dest_reg = SET_DEST (set);
5076 if (REGNO (dest_reg) < max_reg_before_loop
5077 && REGNO (dest_reg) >= FIRST_PSEUDO_REGISTER
5078 && REG_IV_TYPE (REGNO (dest_reg)) != NOT_BASIC_INDUCT)
5080 int multi_insn_incr = 0;
5082 if (basic_induction_var (loop, SET_SRC (set),
5083 GET_MODE (SET_SRC (set)),
5084 dest_reg, p, &inc_val, &mult_val,
5085 &location, &multi_insn_incr))
5087 /* It is a possible basic induction variable.
5088 Create and initialize an induction structure for it. */
5091 = (struct induction *) oballoc (sizeof (struct induction));
5093 record_biv (v, p, dest_reg, inc_val, mult_val, location,
5094 not_every_iteration, maybe_multiple,
5096 REG_IV_TYPE (REGNO (dest_reg)) = BASIC_INDUCT;
5098 else if (REGNO (dest_reg) < max_reg_before_loop)
5099 REG_IV_TYPE (REGNO (dest_reg)) = NOT_BASIC_INDUCT;
5105 /* Record all givs calculated in the insn.
5106 A register is a giv if: it is only set once, it is a function of a
5107 biv and a constant (or invariant), and it is not a biv. */
5109 check_insn_for_givs (loop, p, not_every_iteration, maybe_multiple)
5112 int not_every_iteration;
5116 /* Look for a general induction variable in a register. */
5117 if (GET_CODE (p) == INSN
5118 && (set = single_set (p))
5119 && GET_CODE (SET_DEST (set)) == REG
5120 && ! VARRAY_CHAR (may_not_optimize, REGNO (SET_DEST (set))))
5128 rtx last_consec_insn;
5130 dest_reg = SET_DEST (set);
5131 if (REGNO (dest_reg) < FIRST_PSEUDO_REGISTER)
5134 if (/* SET_SRC is a giv. */
5135 (general_induction_var (loop, SET_SRC (set), &src_reg, &add_val,
5136 &mult_val, 0, &benefit)
5137 /* Equivalent expression is a giv. */
5138 || ((regnote = find_reg_note (p, REG_EQUAL, NULL_RTX))
5139 && general_induction_var (loop, XEXP (regnote, 0), &src_reg,
5140 &add_val, &mult_val, 0,
5142 /* Don't try to handle any regs made by loop optimization.
5143 We have nothing on them in regno_first_uid, etc. */
5144 && REGNO (dest_reg) < max_reg_before_loop
5145 /* Don't recognize a BASIC_INDUCT_VAR here. */
5146 && dest_reg != src_reg
5147 /* This must be the only place where the register is set. */
5148 && (VARRAY_INT (n_times_set, REGNO (dest_reg)) == 1
5149 /* or all sets must be consecutive and make a giv. */
5150 || (benefit = consec_sets_giv (loop, benefit, p,
5152 &add_val, &mult_val,
5153 &last_consec_insn))))
5156 = (struct induction *) oballoc (sizeof (struct induction));
5158 /* If this is a library call, increase benefit. */
5159 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
5160 benefit += libcall_benefit (p);
5162 /* Skip the consecutive insns, if there are any. */
5163 if (VARRAY_INT (n_times_set, REGNO (dest_reg)) != 1)
5164 p = last_consec_insn;
5166 record_giv (loop, v, p, src_reg, dest_reg, mult_val, add_val,
5167 benefit, DEST_REG, not_every_iteration,
5168 maybe_multiple, NULL_PTR);
5173 #ifndef DONT_REDUCE_ADDR
5174 /* Look for givs which are memory addresses. */
5175 /* This resulted in worse code on a VAX 8600. I wonder if it
5177 if (GET_CODE (p) == INSN)
5178 find_mem_givs (loop, PATTERN (p), p, not_every_iteration,
5182 /* Update the status of whether giv can derive other givs. This can
5183 change when we pass a label or an insn that updates a biv. */
5184 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
5185 || GET_CODE (p) == CODE_LABEL)
5186 update_giv_derive (loop, p);
5190 /* Return 1 if X is a valid source for an initial value (or as value being
5191 compared against in an initial test).
5193 X must be either a register or constant and must not be clobbered between
5194 the current insn and the start of the loop.
5196 INSN is the insn containing X. */
5199 valid_initial_value_p (x, insn, call_seen, loop_start)
5208 /* Only consider pseudos we know about initialized in insns whose luids
5210 if (GET_CODE (x) != REG
5211 || REGNO (x) >= max_reg_before_loop)
5214 /* Don't use call-clobbered registers across a call which clobbers it. On
5215 some machines, don't use any hard registers at all. */
5216 if (REGNO (x) < FIRST_PSEUDO_REGISTER
5217 && (SMALL_REGISTER_CLASSES
5218 || (call_used_regs[REGNO (x)] && call_seen)))
5221 /* Don't use registers that have been clobbered before the start of the
5223 if (reg_set_between_p (x, insn, loop_start))
5229 /* Scan X for memory refs and check each memory address
5230 as a possible giv. INSN is the insn whose pattern X comes from.
5231 NOT_EVERY_ITERATION is 1 if the insn might not be executed during
5232 every loop iteration. MAYBE_MULTIPLE is 1 if the insn might be executed
5233 more thanonce in each loop iteration. */
5236 find_mem_givs (loop, x, insn, not_every_iteration, maybe_multiple)
5237 const struct loop *loop;
5240 int not_every_iteration, maybe_multiple;
5243 register enum rtx_code code;
5244 register const char *fmt;
5249 code = GET_CODE (x);
5273 /* This code used to disable creating GIVs with mult_val == 1 and
5274 add_val == 0. However, this leads to lost optimizations when
5275 it comes time to combine a set of related DEST_ADDR GIVs, since
5276 this one would not be seen. */
5278 if (general_induction_var (loop, XEXP (x, 0), &src_reg, &add_val,
5279 &mult_val, 1, &benefit))
5281 /* Found one; record it. */
5283 = (struct induction *) oballoc (sizeof (struct induction));
5285 record_giv (loop, v, insn, src_reg, addr_placeholder, mult_val,
5286 add_val, benefit, DEST_ADDR, not_every_iteration,
5287 maybe_multiple, &XEXP (x, 0));
5289 v->mem_mode = GET_MODE (x);
5298 /* Recursively scan the subexpressions for other mem refs. */
5300 fmt = GET_RTX_FORMAT (code);
5301 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
5303 find_mem_givs (loop, XEXP (x, i), insn, not_every_iteration,
5305 else if (fmt[i] == 'E')
5306 for (j = 0; j < XVECLEN (x, i); j++)
5307 find_mem_givs (loop, XVECEXP (x, i, j), insn, not_every_iteration,
5311 /* Fill in the data about one biv update.
5312 V is the `struct induction' in which we record the biv. (It is
5313 allocated by the caller, with alloca.)
5314 INSN is the insn that sets it.
5315 DEST_REG is the biv's reg.
5317 MULT_VAL is const1_rtx if the biv is being incremented here, in which case
5318 INC_VAL is the increment. Otherwise, MULT_VAL is const0_rtx and the biv is
5319 being set to INC_VAL.
5321 NOT_EVERY_ITERATION is nonzero if this biv update is not know to be
5322 executed every iteration; MAYBE_MULTIPLE is nonzero if this biv update
5323 can be executed more than once per iteration. If MAYBE_MULTIPLE
5324 and NOT_EVERY_ITERATION are both zero, we know that the biv update is
5325 executed exactly once per iteration. */
5328 record_biv (v, insn, dest_reg, inc_val, mult_val, location,
5329 not_every_iteration, maybe_multiple, multi_insn_incr)
5330 struct induction *v;
5336 int not_every_iteration;
5338 int multi_insn_incr;
5340 struct iv_class *bl;
5343 v->src_reg = dest_reg;
5344 v->dest_reg = dest_reg;
5345 v->mult_val = mult_val;
5346 v->add_val = inc_val;
5347 v->location = location;
5348 v->mode = GET_MODE (dest_reg);
5349 v->always_computable = ! not_every_iteration;
5350 v->always_executed = ! not_every_iteration;
5351 v->maybe_multiple = maybe_multiple;
5352 v->multi_insn_incr = multi_insn_incr;
5354 /* Add this to the reg's iv_class, creating a class
5355 if this is the first incrementation of the reg. */
5357 bl = reg_biv_class[REGNO (dest_reg)];
5360 /* Create and initialize new iv_class. */
5362 bl = (struct iv_class *) oballoc (sizeof (struct iv_class));
5364 bl->regno = REGNO (dest_reg);
5370 /* Set initial value to the reg itself. */
5371 bl->initial_value = dest_reg;
5372 /* We haven't seen the initializing insn yet */
5375 bl->initial_test = 0;
5376 bl->incremented = 0;
5380 bl->total_benefit = 0;
5382 /* Add this class to loop_iv_list. */
5383 bl->next = loop_iv_list;
5386 /* Put it in the array of biv register classes. */
5387 reg_biv_class[REGNO (dest_reg)] = bl;
5390 /* Update IV_CLASS entry for this biv. */
5391 v->next_iv = bl->biv;
5394 if (mult_val == const1_rtx)
5395 bl->incremented = 1;
5397 if (loop_dump_stream)
5399 fprintf (loop_dump_stream,
5400 "Insn %d: possible biv, reg %d,",
5401 INSN_UID (insn), REGNO (dest_reg));
5402 if (GET_CODE (inc_val) == CONST_INT)
5404 fprintf (loop_dump_stream, " const =");
5405 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (inc_val));
5406 fputc ('\n', loop_dump_stream);
5410 fprintf (loop_dump_stream, " const = ");
5411 print_rtl (loop_dump_stream, inc_val);
5412 fprintf (loop_dump_stream, "\n");
5417 /* Fill in the data about one giv.
5418 V is the `struct induction' in which we record the giv. (It is
5419 allocated by the caller, with alloca.)
5420 INSN is the insn that sets it.
5421 BENEFIT estimates the savings from deleting this insn.
5422 TYPE is DEST_REG or DEST_ADDR; it says whether the giv is computed
5423 into a register or is used as a memory address.
5425 SRC_REG is the biv reg which the giv is computed from.
5426 DEST_REG is the giv's reg (if the giv is stored in a reg).
5427 MULT_VAL and ADD_VAL are the coefficients used to compute the giv.
5428 LOCATION points to the place where this giv's value appears in INSN. */
5431 record_giv (loop, v, insn, src_reg, dest_reg, mult_val, add_val, benefit,
5432 type, not_every_iteration, maybe_multiple, location)
5433 const struct loop *loop;
5434 struct induction *v;
5438 rtx mult_val, add_val;
5441 int not_every_iteration, maybe_multiple;
5444 struct induction *b;
5445 struct iv_class *bl;
5446 rtx set = single_set (insn);
5449 /* Attempt to prove constantness of the values. */
5450 temp = simplify_rtx (add_val);
5455 v->src_reg = src_reg;
5457 v->dest_reg = dest_reg;
5458 v->mult_val = mult_val;
5459 v->add_val = add_val;
5460 v->benefit = benefit;
5461 v->location = location;
5463 v->combined_with = 0;
5464 v->maybe_multiple = maybe_multiple;
5465 v->multi_insn_incr = 0;
5467 v->derive_adjustment = 0;
5473 v->auto_inc_opt = 0;
5476 v->derived_from = 0;
5479 /* The v->always_computable field is used in update_giv_derive, to
5480 determine whether a giv can be used to derive another giv. For a
5481 DEST_REG giv, INSN computes a new value for the giv, so its value
5482 isn't computable if INSN insn't executed every iteration.
5483 However, for a DEST_ADDR giv, INSN merely uses the value of the giv;
5484 it does not compute a new value. Hence the value is always computable
5485 regardless of whether INSN is executed each iteration. */
5487 if (type == DEST_ADDR)
5488 v->always_computable = 1;
5490 v->always_computable = ! not_every_iteration;
5492 v->always_executed = ! not_every_iteration;
5494 if (type == DEST_ADDR)
5496 v->mode = GET_MODE (*location);
5499 else /* type == DEST_REG */
5501 v->mode = GET_MODE (SET_DEST (set));
5503 v->lifetime = (uid_luid[REGNO_LAST_UID (REGNO (dest_reg))]
5504 - uid_luid[REGNO_FIRST_UID (REGNO (dest_reg))]);
5506 /* If the lifetime is zero, it means that this register is
5507 really a dead store. So mark this as a giv that can be
5508 ignored. This will not prevent the biv from being eliminated. */
5509 if (v->lifetime == 0)
5512 REG_IV_TYPE (REGNO (dest_reg)) = GENERAL_INDUCT;
5513 REG_IV_INFO (REGNO (dest_reg)) = v;
5516 /* Add the giv to the class of givs computed from one biv. */
5518 bl = reg_biv_class[REGNO (src_reg)];
5521 v->next_iv = bl->giv;
5523 /* Don't count DEST_ADDR. This is supposed to count the number of
5524 insns that calculate givs. */
5525 if (type == DEST_REG)
5527 bl->total_benefit += benefit;
5530 /* Fatal error, biv missing for this giv? */
5533 if (type == DEST_ADDR)
5537 /* The giv can be replaced outright by the reduced register only if all
5538 of the following conditions are true:
5539 - the insn that sets the giv is always executed on any iteration
5540 on which the giv is used at all
5541 (there are two ways to deduce this:
5542 either the insn is executed on every iteration,
5543 or all uses follow that insn in the same basic block),
5544 - the giv is not used outside the loop
5545 - no assignments to the biv occur during the giv's lifetime. */
5547 if (REGNO_FIRST_UID (REGNO (dest_reg)) == INSN_UID (insn)
5548 /* Previous line always fails if INSN was moved by loop opt. */
5549 && uid_luid[REGNO_LAST_UID (REGNO (dest_reg))]
5550 < INSN_LUID (loop->end)
5551 && (! not_every_iteration
5552 || last_use_this_basic_block (dest_reg, insn)))
5554 /* Now check that there are no assignments to the biv within the
5555 giv's lifetime. This requires two separate checks. */
5557 /* Check each biv update, and fail if any are between the first
5558 and last use of the giv.
5560 If this loop contains an inner loop that was unrolled, then
5561 the insn modifying the biv may have been emitted by the loop
5562 unrolling code, and hence does not have a valid luid. Just
5563 mark the biv as not replaceable in this case. It is not very
5564 useful as a biv, because it is used in two different loops.
5565 It is very unlikely that we would be able to optimize the giv
5566 using this biv anyways. */
5569 for (b = bl->biv; b; b = b->next_iv)
5571 if (INSN_UID (b->insn) >= max_uid_for_loop
5572 || ((uid_luid[INSN_UID (b->insn)]
5573 >= uid_luid[REGNO_FIRST_UID (REGNO (dest_reg))])
5574 && (uid_luid[INSN_UID (b->insn)]
5575 <= uid_luid[REGNO_LAST_UID (REGNO (dest_reg))])))
5578 v->not_replaceable = 1;
5583 /* If there are any backwards branches that go from after the
5584 biv update to before it, then this giv is not replaceable. */
5586 for (b = bl->biv; b; b = b->next_iv)
5587 if (back_branch_in_range_p (loop, b->insn))
5590 v->not_replaceable = 1;
5596 /* May still be replaceable, we don't have enough info here to
5599 v->not_replaceable = 0;
5603 /* Record whether the add_val contains a const_int, for later use by
5608 v->no_const_addval = 1;
5609 if (tem == const0_rtx)
5611 else if (CONSTANT_P (add_val))
5612 v->no_const_addval = 0;
5613 if (GET_CODE (tem) == PLUS)
5617 if (GET_CODE (XEXP (tem, 0)) == PLUS)
5618 tem = XEXP (tem, 0);
5619 else if (GET_CODE (XEXP (tem, 1)) == PLUS)
5620 tem = XEXP (tem, 1);
5624 if (CONSTANT_P (XEXP (tem, 1)))
5625 v->no_const_addval = 0;
5629 if (loop_dump_stream)
5631 if (type == DEST_REG)
5632 fprintf (loop_dump_stream, "Insn %d: giv reg %d",
5633 INSN_UID (insn), REGNO (dest_reg));
5635 fprintf (loop_dump_stream, "Insn %d: dest address",
5638 fprintf (loop_dump_stream, " src reg %d benefit %d",
5639 REGNO (src_reg), v->benefit);
5640 fprintf (loop_dump_stream, " lifetime %d",
5644 fprintf (loop_dump_stream, " replaceable");
5646 if (v->no_const_addval)
5647 fprintf (loop_dump_stream, " ncav");
5649 if (GET_CODE (mult_val) == CONST_INT)
5651 fprintf (loop_dump_stream, " mult ");
5652 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (mult_val));
5656 fprintf (loop_dump_stream, " mult ");
5657 print_rtl (loop_dump_stream, mult_val);
5660 if (GET_CODE (add_val) == CONST_INT)
5662 fprintf (loop_dump_stream, " add ");
5663 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (add_val));
5667 fprintf (loop_dump_stream, " add ");
5668 print_rtl (loop_dump_stream, add_val);
5672 if (loop_dump_stream)
5673 fprintf (loop_dump_stream, "\n");
5678 /* All this does is determine whether a giv can be made replaceable because
5679 its final value can be calculated. This code can not be part of record_giv
5680 above, because final_giv_value requires that the number of loop iterations
5681 be known, and that can not be accurately calculated until after all givs
5682 have been identified. */
5685 check_final_value (loop, v)
5686 const struct loop *loop;
5687 struct induction *v;
5689 struct iv_class *bl;
5690 rtx final_value = 0;
5692 bl = reg_biv_class[REGNO (v->src_reg)];
5694 /* DEST_ADDR givs will never reach here, because they are always marked
5695 replaceable above in record_giv. */
5697 /* The giv can be replaced outright by the reduced register only if all
5698 of the following conditions are true:
5699 - the insn that sets the giv is always executed on any iteration
5700 on which the giv is used at all
5701 (there are two ways to deduce this:
5702 either the insn is executed on every iteration,
5703 or all uses follow that insn in the same basic block),
5704 - its final value can be calculated (this condition is different
5705 than the one above in record_giv)
5706 - no assignments to the biv occur during the giv's lifetime. */
5709 /* This is only called now when replaceable is known to be false. */
5710 /* Clear replaceable, so that it won't confuse final_giv_value. */
5714 if ((final_value = final_giv_value (loop, v))
5715 && (v->always_computable || last_use_this_basic_block (v->dest_reg, v->insn)))
5717 int biv_increment_seen = 0;
5723 /* When trying to determine whether or not a biv increment occurs
5724 during the lifetime of the giv, we can ignore uses of the variable
5725 outside the loop because final_value is true. Hence we can not
5726 use regno_last_uid and regno_first_uid as above in record_giv. */
5728 /* Search the loop to determine whether any assignments to the
5729 biv occur during the giv's lifetime. Start with the insn
5730 that sets the giv, and search around the loop until we come
5731 back to that insn again.
5733 Also fail if there is a jump within the giv's lifetime that jumps
5734 to somewhere outside the lifetime but still within the loop. This
5735 catches spaghetti code where the execution order is not linear, and
5736 hence the above test fails. Here we assume that the giv lifetime
5737 does not extend from one iteration of the loop to the next, so as
5738 to make the test easier. Since the lifetime isn't known yet,
5739 this requires two loops. See also record_giv above. */
5741 last_giv_use = v->insn;
5747 p = NEXT_INSN (loop->start);
5751 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
5752 || GET_CODE (p) == CALL_INSN)
5754 if (biv_increment_seen)
5756 if (reg_mentioned_p (v->dest_reg, PATTERN (p)))
5759 v->not_replaceable = 1;
5763 else if (reg_set_p (v->src_reg, PATTERN (p)))
5764 biv_increment_seen = 1;
5765 else if (reg_mentioned_p (v->dest_reg, PATTERN (p)))
5770 /* Now that the lifetime of the giv is known, check for branches
5771 from within the lifetime to outside the lifetime if it is still
5781 p = NEXT_INSN (loop->start);
5782 if (p == last_giv_use)
5785 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p)
5786 && LABEL_NAME (JUMP_LABEL (p))
5787 && ((loop_insn_first_p (JUMP_LABEL (p), v->insn)
5788 && loop_insn_first_p (loop->start, JUMP_LABEL (p)))
5789 || (loop_insn_first_p (last_giv_use, JUMP_LABEL (p))
5790 && loop_insn_first_p (JUMP_LABEL (p), loop->end))))
5793 v->not_replaceable = 1;
5795 if (loop_dump_stream)
5796 fprintf (loop_dump_stream,
5797 "Found branch outside giv lifetime.\n");
5804 /* If it is replaceable, then save the final value. */
5806 v->final_value = final_value;
5809 if (loop_dump_stream && v->replaceable)
5810 fprintf (loop_dump_stream, "Insn %d: giv reg %d final_value replaceable\n",
5811 INSN_UID (v->insn), REGNO (v->dest_reg));
5814 /* Update the status of whether a giv can derive other givs.
5816 We need to do something special if there is or may be an update to the biv
5817 between the time the giv is defined and the time it is used to derive
5820 In addition, a giv that is only conditionally set is not allowed to
5821 derive another giv once a label has been passed.
5823 The cases we look at are when a label or an update to a biv is passed. */
5826 update_giv_derive (loop, p)
5827 const struct loop *loop;
5830 struct iv_class *bl;
5831 struct induction *biv, *giv;
5835 /* Search all IV classes, then all bivs, and finally all givs.
5837 There are three cases we are concerned with. First we have the situation
5838 of a giv that is only updated conditionally. In that case, it may not
5839 derive any givs after a label is passed.
5841 The second case is when a biv update occurs, or may occur, after the
5842 definition of a giv. For certain biv updates (see below) that are
5843 known to occur between the giv definition and use, we can adjust the
5844 giv definition. For others, or when the biv update is conditional,
5845 we must prevent the giv from deriving any other givs. There are two
5846 sub-cases within this case.
5848 If this is a label, we are concerned with any biv update that is done
5849 conditionally, since it may be done after the giv is defined followed by
5850 a branch here (actually, we need to pass both a jump and a label, but
5851 this extra tracking doesn't seem worth it).
5853 If this is a jump, we are concerned about any biv update that may be
5854 executed multiple times. We are actually only concerned about
5855 backward jumps, but it is probably not worth performing the test
5856 on the jump again here.
5858 If this is a biv update, we must adjust the giv status to show that a
5859 subsequent biv update was performed. If this adjustment cannot be done,
5860 the giv cannot derive further givs. */
5862 for (bl = loop_iv_list; bl; bl = bl->next)
5863 for (biv = bl->biv; biv; biv = biv->next_iv)
5864 if (GET_CODE (p) == CODE_LABEL || GET_CODE (p) == JUMP_INSN
5867 for (giv = bl->giv; giv; giv = giv->next_iv)
5869 /* If cant_derive is already true, there is no point in
5870 checking all of these conditions again. */
5871 if (giv->cant_derive)
5874 /* If this giv is conditionally set and we have passed a label,
5875 it cannot derive anything. */
5876 if (GET_CODE (p) == CODE_LABEL && ! giv->always_computable)
5877 giv->cant_derive = 1;
5879 /* Skip givs that have mult_val == 0, since
5880 they are really invariants. Also skip those that are
5881 replaceable, since we know their lifetime doesn't contain
5883 else if (giv->mult_val == const0_rtx || giv->replaceable)
5886 /* The only way we can allow this giv to derive another
5887 is if this is a biv increment and we can form the product
5888 of biv->add_val and giv->mult_val. In this case, we will
5889 be able to compute a compensation. */
5890 else if (biv->insn == p)
5894 if (biv->mult_val == const1_rtx)
5895 tem = simplify_giv_expr (loop,
5896 gen_rtx_MULT (giv->mode,
5901 if (tem && giv->derive_adjustment)
5902 tem = simplify_giv_expr
5904 gen_rtx_PLUS (giv->mode, tem, giv->derive_adjustment),
5908 giv->derive_adjustment = tem;
5910 giv->cant_derive = 1;
5912 else if ((GET_CODE (p) == CODE_LABEL && ! biv->always_computable)
5913 || (GET_CODE (p) == JUMP_INSN && biv->maybe_multiple))
5914 giv->cant_derive = 1;
5919 /* Check whether an insn is an increment legitimate for a basic induction var.
5920 X is the source of insn P, or a part of it.
5921 MODE is the mode in which X should be interpreted.
5923 DEST_REG is the putative biv, also the destination of the insn.
5924 We accept patterns of these forms:
5925 REG = REG + INVARIANT (includes REG = REG - CONSTANT)
5926 REG = INVARIANT + REG
5928 If X is suitable, we return 1, set *MULT_VAL to CONST1_RTX,
5929 store the additive term into *INC_VAL, and store the place where
5930 we found the additive term into *LOCATION.
5932 If X is an assignment of an invariant into DEST_REG, we set
5933 *MULT_VAL to CONST0_RTX, and store the invariant into *INC_VAL.
5935 We also want to detect a BIV when it corresponds to a variable
5936 whose mode was promoted via PROMOTED_MODE. In that case, an increment
5937 of the variable may be a PLUS that adds a SUBREG of that variable to
5938 an invariant and then sign- or zero-extends the result of the PLUS
5941 Most GIVs in such cases will be in the promoted mode, since that is the
5942 probably the natural computation mode (and almost certainly the mode
5943 used for addresses) on the machine. So we view the pseudo-reg containing
5944 the variable as the BIV, as if it were simply incremented.
5946 Note that treating the entire pseudo as a BIV will result in making
5947 simple increments to any GIVs based on it. However, if the variable
5948 overflows in its declared mode but not its promoted mode, the result will
5949 be incorrect. This is acceptable if the variable is signed, since
5950 overflows in such cases are undefined, but not if it is unsigned, since
5951 those overflows are defined. So we only check for SIGN_EXTEND and
5954 If we cannot find a biv, we return 0. */
5957 basic_induction_var (loop, x, mode, dest_reg, p, inc_val, mult_val,
5958 location, multi_insn_incr)
5959 const struct loop *loop;
5961 enum machine_mode mode;
5967 int *multi_insn_incr;
5969 register enum rtx_code code;
5973 code = GET_CODE (x);
5978 if (rtx_equal_p (XEXP (x, 0), dest_reg)
5979 || (GET_CODE (XEXP (x, 0)) == SUBREG
5980 && SUBREG_PROMOTED_VAR_P (XEXP (x, 0))
5981 && SUBREG_REG (XEXP (x, 0)) == dest_reg))
5983 argp = &XEXP (x, 1);
5985 else if (rtx_equal_p (XEXP (x, 1), dest_reg)
5986 || (GET_CODE (XEXP (x, 1)) == SUBREG
5987 && SUBREG_PROMOTED_VAR_P (XEXP (x, 1))
5988 && SUBREG_REG (XEXP (x, 1)) == dest_reg))
5990 argp = &XEXP (x, 0);
5996 if (loop_invariant_p (loop, arg) != 1)
5999 *inc_val = convert_modes (GET_MODE (dest_reg), GET_MODE (x), arg, 0);
6000 *mult_val = const1_rtx;
6005 /* If this is a SUBREG for a promoted variable, check the inner
6007 if (SUBREG_PROMOTED_VAR_P (x))
6008 return basic_induction_var (loop, SUBREG_REG (x),
6009 GET_MODE (SUBREG_REG (x)),
6010 dest_reg, p, inc_val, mult_val, location,
6015 /* If this register is assigned in a previous insn, look at its
6016 source, but don't go outside the loop or past a label. */
6022 insn = PREV_INSN (insn);
6023 } while (insn && GET_CODE (insn) == NOTE
6024 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_BEG);
6028 set = single_set (insn);
6032 if ((SET_DEST (set) == x
6033 || (GET_CODE (SET_DEST (set)) == SUBREG
6034 && (GET_MODE_SIZE (GET_MODE (SET_DEST (set)))
6036 && (GET_MODE_CLASS (GET_MODE (SET_DEST (set)))
6038 && SUBREG_REG (SET_DEST (set)) == x))
6039 && basic_induction_var (loop, SET_SRC (set),
6040 (GET_MODE (SET_SRC (set)) == VOIDmode
6042 : GET_MODE (SET_SRC (set))),
6044 inc_val, mult_val, location,
6047 *multi_insn_incr = 1;
6051 /* ... fall through ... */
6053 /* Can accept constant setting of biv only when inside inner most loop.
6054 Otherwise, a biv of an inner loop may be incorrectly recognized
6055 as a biv of the outer loop,
6056 causing code to be moved INTO the inner loop. */
6058 if (loop_invariant_p (loop, x) != 1)
6063 /* convert_modes aborts if we try to convert to or from CCmode, so just
6064 exclude that case. It is very unlikely that a condition code value
6065 would be a useful iterator anyways. */
6066 if (loop->level == 1
6067 && GET_MODE_CLASS (mode) != MODE_CC
6068 && GET_MODE_CLASS (GET_MODE (dest_reg)) != MODE_CC)
6070 /* Possible bug here? Perhaps we don't know the mode of X. */
6071 *inc_val = convert_modes (GET_MODE (dest_reg), mode, x, 0);
6072 *mult_val = const0_rtx;
6079 return basic_induction_var (loop, XEXP (x, 0), GET_MODE (XEXP (x, 0)),
6080 dest_reg, p, inc_val, mult_val, location,
6084 /* Similar, since this can be a sign extension. */
6085 for (insn = PREV_INSN (p);
6086 (insn && GET_CODE (insn) == NOTE
6087 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_BEG);
6088 insn = PREV_INSN (insn))
6092 set = single_set (insn);
6094 if (set && SET_DEST (set) == XEXP (x, 0)
6095 && GET_CODE (XEXP (x, 1)) == CONST_INT
6096 && INTVAL (XEXP (x, 1)) >= 0
6097 && GET_CODE (SET_SRC (set)) == ASHIFT
6098 && XEXP (x, 1) == XEXP (SET_SRC (set), 1)
6099 && basic_induction_var (loop, XEXP (SET_SRC (set), 0),
6100 GET_MODE (XEXP (x, 0)),
6101 dest_reg, insn, inc_val, mult_val,
6102 location, multi_insn_incr))
6104 *multi_insn_incr = 1;
6114 /* A general induction variable (giv) is any quantity that is a linear
6115 function of a basic induction variable,
6116 i.e. giv = biv * mult_val + add_val.
6117 The coefficients can be any loop invariant quantity.
6118 A giv need not be computed directly from the biv;
6119 it can be computed by way of other givs. */
6121 /* Determine whether X computes a giv.
6122 If it does, return a nonzero value
6123 which is the benefit from eliminating the computation of X;
6124 set *SRC_REG to the register of the biv that it is computed from;
6125 set *ADD_VAL and *MULT_VAL to the coefficients,
6126 such that the value of X is biv * mult + add; */
6129 general_induction_var (loop, x, src_reg, add_val, mult_val, is_addr, pbenefit)
6130 const struct loop *loop;
6141 /* If this is an invariant, forget it, it isn't a giv. */
6142 if (loop_invariant_p (loop, x) == 1)
6145 /* See if the expression could be a giv and get its form.
6146 Mark our place on the obstack in case we don't find a giv. */
6147 storage = (char *) oballoc (0);
6149 x = simplify_giv_expr (loop, x, pbenefit);
6156 switch (GET_CODE (x))
6160 /* Since this is now an invariant and wasn't before, it must be a giv
6161 with MULT_VAL == 0. It doesn't matter which BIV we associate this
6163 *src_reg = loop_iv_list->biv->dest_reg;
6164 *mult_val = const0_rtx;
6169 /* This is equivalent to a BIV. */
6171 *mult_val = const1_rtx;
6172 *add_val = const0_rtx;
6176 /* Either (plus (biv) (invar)) or
6177 (plus (mult (biv) (invar_1)) (invar_2)). */
6178 if (GET_CODE (XEXP (x, 0)) == MULT)
6180 *src_reg = XEXP (XEXP (x, 0), 0);
6181 *mult_val = XEXP (XEXP (x, 0), 1);
6185 *src_reg = XEXP (x, 0);
6186 *mult_val = const1_rtx;
6188 *add_val = XEXP (x, 1);
6192 /* ADD_VAL is zero. */
6193 *src_reg = XEXP (x, 0);
6194 *mult_val = XEXP (x, 1);
6195 *add_val = const0_rtx;
6202 /* Remove any enclosing USE from ADD_VAL and MULT_VAL (there will be
6203 unless they are CONST_INT). */
6204 if (GET_CODE (*add_val) == USE)
6205 *add_val = XEXP (*add_val, 0);
6206 if (GET_CODE (*mult_val) == USE)
6207 *mult_val = XEXP (*mult_val, 0);
6212 *pbenefit += ADDRESS_COST (orig_x) - reg_address_cost;
6214 *pbenefit += rtx_cost (orig_x, MEM) - reg_address_cost;
6218 *pbenefit += rtx_cost (orig_x, SET);
6220 /* Always return true if this is a giv so it will be detected as such,
6221 even if the benefit is zero or negative. This allows elimination
6222 of bivs that might otherwise not be eliminated. */
6226 /* Given an expression, X, try to form it as a linear function of a biv.
6227 We will canonicalize it to be of the form
6228 (plus (mult (BIV) (invar_1))
6230 with possible degeneracies.
6232 The invariant expressions must each be of a form that can be used as a
6233 machine operand. We surround then with a USE rtx (a hack, but localized
6234 and certainly unambiguous!) if not a CONST_INT for simplicity in this
6235 routine; it is the caller's responsibility to strip them.
6237 If no such canonicalization is possible (i.e., two biv's are used or an
6238 expression that is neither invariant nor a biv or giv), this routine
6241 For a non-zero return, the result will have a code of CONST_INT, USE,
6242 REG (for a BIV), PLUS, or MULT. No other codes will occur.
6244 *BENEFIT will be incremented by the benefit of any sub-giv encountered. */
6246 static rtx sge_plus PARAMS ((enum machine_mode, rtx, rtx));
6247 static rtx sge_plus_constant PARAMS ((rtx, rtx));
6248 static int cmp_combine_givs_stats PARAMS ((const PTR, const PTR));
6249 static int cmp_recombine_givs_stats PARAMS ((const PTR, const PTR));
6252 simplify_giv_expr (loop, x, benefit)
6253 const struct loop *loop;
6257 enum machine_mode mode = GET_MODE (x);
6261 /* If this is not an integer mode, or if we cannot do arithmetic in this
6262 mode, this can't be a giv. */
6263 if (mode != VOIDmode
6264 && (GET_MODE_CLASS (mode) != MODE_INT
6265 || GET_MODE_BITSIZE (mode) > HOST_BITS_PER_WIDE_INT))
6268 switch (GET_CODE (x))
6271 arg0 = simplify_giv_expr (loop, XEXP (x, 0), benefit);
6272 arg1 = simplify_giv_expr (loop, XEXP (x, 1), benefit);
6273 if (arg0 == 0 || arg1 == 0)
6276 /* Put constant last, CONST_INT last if both constant. */
6277 if ((GET_CODE (arg0) == USE
6278 || GET_CODE (arg0) == CONST_INT)
6279 && ! ((GET_CODE (arg0) == USE
6280 && GET_CODE (arg1) == USE)
6281 || GET_CODE (arg1) == CONST_INT))
6282 tem = arg0, arg0 = arg1, arg1 = tem;
6284 /* Handle addition of zero, then addition of an invariant. */
6285 if (arg1 == const0_rtx)
6287 else if (GET_CODE (arg1) == CONST_INT || GET_CODE (arg1) == USE)
6288 switch (GET_CODE (arg0))
6292 /* Adding two invariants must result in an invariant, so enclose
6293 addition operation inside a USE and return it. */
6294 if (GET_CODE (arg0) == USE)
6295 arg0 = XEXP (arg0, 0);
6296 if (GET_CODE (arg1) == USE)
6297 arg1 = XEXP (arg1, 0);
6299 if (GET_CODE (arg0) == CONST_INT)
6300 tem = arg0, arg0 = arg1, arg1 = tem;
6301 if (GET_CODE (arg1) == CONST_INT)
6302 tem = sge_plus_constant (arg0, arg1);
6304 tem = sge_plus (mode, arg0, arg1);
6306 if (GET_CODE (tem) != CONST_INT)
6307 tem = gen_rtx_USE (mode, tem);
6312 /* biv + invar or mult + invar. Return sum. */
6313 return gen_rtx_PLUS (mode, arg0, arg1);
6316 /* (a + invar_1) + invar_2. Associate. */
6318 simplify_giv_expr (loop,
6330 /* Each argument must be either REG, PLUS, or MULT. Convert REG to
6331 MULT to reduce cases. */
6332 if (GET_CODE (arg0) == REG)
6333 arg0 = gen_rtx_MULT (mode, arg0, const1_rtx);
6334 if (GET_CODE (arg1) == REG)
6335 arg1 = gen_rtx_MULT (mode, arg1, const1_rtx);
6337 /* Now have PLUS + PLUS, PLUS + MULT, MULT + PLUS, or MULT + MULT.
6338 Put a MULT first, leaving PLUS + PLUS, MULT + PLUS, or MULT + MULT.
6339 Recurse to associate the second PLUS. */
6340 if (GET_CODE (arg1) == MULT)
6341 tem = arg0, arg0 = arg1, arg1 = tem;
6343 if (GET_CODE (arg1) == PLUS)
6345 simplify_giv_expr (loop,
6347 gen_rtx_PLUS (mode, arg0,
6352 /* Now must have MULT + MULT. Distribute if same biv, else not giv. */
6353 if (GET_CODE (arg0) != MULT || GET_CODE (arg1) != MULT)
6356 if (!rtx_equal_p (arg0, arg1))
6359 return simplify_giv_expr (loop,
6368 /* Handle "a - b" as "a + b * (-1)". */
6369 return simplify_giv_expr (loop,
6378 arg0 = simplify_giv_expr (loop, XEXP (x, 0), benefit);
6379 arg1 = simplify_giv_expr (loop, XEXP (x, 1), benefit);
6380 if (arg0 == 0 || arg1 == 0)
6383 /* Put constant last, CONST_INT last if both constant. */
6384 if ((GET_CODE (arg0) == USE || GET_CODE (arg0) == CONST_INT)
6385 && GET_CODE (arg1) != CONST_INT)
6386 tem = arg0, arg0 = arg1, arg1 = tem;
6388 /* If second argument is not now constant, not giv. */
6389 if (GET_CODE (arg1) != USE && GET_CODE (arg1) != CONST_INT)
6392 /* Handle multiply by 0 or 1. */
6393 if (arg1 == const0_rtx)
6396 else if (arg1 == const1_rtx)
6399 switch (GET_CODE (arg0))
6402 /* biv * invar. Done. */
6403 return gen_rtx_MULT (mode, arg0, arg1);
6406 /* Product of two constants. */
6407 return GEN_INT (INTVAL (arg0) * INTVAL (arg1));
6410 /* invar * invar is a giv, but attempt to simplify it somehow. */
6411 if (GET_CODE (arg1) != CONST_INT)
6414 arg0 = XEXP (arg0, 0);
6415 if (GET_CODE (arg0) == MULT)
6417 /* (invar_0 * invar_1) * invar_2. Associate. */
6418 return simplify_giv_expr (loop,
6427 /* Porpagate the MULT expressions to the intermost nodes. */
6428 else if (GET_CODE (arg0) == PLUS)
6430 /* (invar_0 + invar_1) * invar_2. Distribute. */
6431 return simplify_giv_expr (loop,
6443 return gen_rtx_USE (mode, gen_rtx_MULT (mode, arg0, arg1));
6446 /* (a * invar_1) * invar_2. Associate. */
6447 return simplify_giv_expr (loop,
6456 /* (a + invar_1) * invar_2. Distribute. */
6457 return simplify_giv_expr (loop,
6472 /* Shift by constant is multiply by power of two. */
6473 if (GET_CODE (XEXP (x, 1)) != CONST_INT)
6477 simplify_giv_expr (loop,
6480 GEN_INT ((HOST_WIDE_INT) 1
6481 << INTVAL (XEXP (x, 1)))),
6485 /* "-a" is "a * (-1)" */
6486 return simplify_giv_expr (loop,
6487 gen_rtx_MULT (mode, XEXP (x, 0), constm1_rtx),
6491 /* "~a" is "-a - 1". Silly, but easy. */
6492 return simplify_giv_expr (loop,
6493 gen_rtx_MINUS (mode,
6494 gen_rtx_NEG (mode, XEXP (x, 0)),
6499 /* Already in proper form for invariant. */
6503 /* If this is a new register, we can't deal with it. */
6504 if (REGNO (x) >= max_reg_before_loop)
6507 /* Check for biv or giv. */
6508 switch (REG_IV_TYPE (REGNO (x)))
6512 case GENERAL_INDUCT:
6514 struct induction *v = REG_IV_INFO (REGNO (x));
6516 /* Form expression from giv and add benefit. Ensure this giv
6517 can derive another and subtract any needed adjustment if so. */
6518 *benefit += v->benefit;
6522 tem = gen_rtx_PLUS (mode, gen_rtx_MULT (mode,
6523 v->src_reg, v->mult_val),
6526 if (v->derive_adjustment)
6527 tem = gen_rtx_MINUS (mode, tem, v->derive_adjustment);
6528 return simplify_giv_expr (loop, tem, benefit);
6532 /* If it isn't an induction variable, and it is invariant, we
6533 may be able to simplify things further by looking through
6534 the bits we just moved outside the loop. */
6535 if (loop_invariant_p (loop, x) == 1)
6539 for (m = the_movables; m ; m = m->next)
6540 if (rtx_equal_p (x, m->set_dest))
6542 /* Ok, we found a match. Substitute and simplify. */
6544 /* If we match another movable, we must use that, as
6545 this one is going away. */
6547 return simplify_giv_expr (loop, m->match->set_dest,
6550 /* If consec is non-zero, this is a member of a group of
6551 instructions that were moved together. We handle this
6552 case only to the point of seeking to the last insn and
6553 looking for a REG_EQUAL. Fail if we don't find one. */
6558 do { tem = NEXT_INSN (tem); } while (--i > 0);
6560 tem = find_reg_note (tem, REG_EQUAL, NULL_RTX);
6562 tem = XEXP (tem, 0);
6566 tem = single_set (m->insn);
6568 tem = SET_SRC (tem);
6573 /* What we are most interested in is pointer
6574 arithmetic on invariants -- only take
6575 patterns we may be able to do something with. */
6576 if (GET_CODE (tem) == PLUS
6577 || GET_CODE (tem) == MULT
6578 || GET_CODE (tem) == ASHIFT
6579 || GET_CODE (tem) == CONST_INT
6580 || GET_CODE (tem) == SYMBOL_REF)
6582 tem = simplify_giv_expr (loop, tem, benefit);
6586 else if (GET_CODE (tem) == CONST
6587 && GET_CODE (XEXP (tem, 0)) == PLUS
6588 && GET_CODE (XEXP (XEXP (tem, 0), 0)) == SYMBOL_REF
6589 && GET_CODE (XEXP (XEXP (tem, 0), 1)) == CONST_INT)
6591 tem = simplify_giv_expr (loop, XEXP (tem, 0),
6603 /* Fall through to general case. */
6605 /* If invariant, return as USE (unless CONST_INT).
6606 Otherwise, not giv. */
6607 if (GET_CODE (x) == USE)
6610 if (loop_invariant_p (loop, x) == 1)
6612 if (GET_CODE (x) == CONST_INT)
6614 if (GET_CODE (x) == CONST
6615 && GET_CODE (XEXP (x, 0)) == PLUS
6616 && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF
6617 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT)
6619 return gen_rtx_USE (mode, x);
6626 /* This routine folds invariants such that there is only ever one
6627 CONST_INT in the summation. It is only used by simplify_giv_expr. */
6630 sge_plus_constant (x, c)
6633 if (GET_CODE (x) == CONST_INT)
6634 return GEN_INT (INTVAL (x) + INTVAL (c));
6635 else if (GET_CODE (x) != PLUS)
6636 return gen_rtx_PLUS (GET_MODE (x), x, c);
6637 else if (GET_CODE (XEXP (x, 1)) == CONST_INT)
6639 return gen_rtx_PLUS (GET_MODE (x), XEXP (x, 0),
6640 GEN_INT (INTVAL (XEXP (x, 1)) + INTVAL (c)));
6642 else if (GET_CODE (XEXP (x, 0)) == PLUS
6643 || GET_CODE (XEXP (x, 1)) != PLUS)
6645 return gen_rtx_PLUS (GET_MODE (x),
6646 sge_plus_constant (XEXP (x, 0), c), XEXP (x, 1));
6650 return gen_rtx_PLUS (GET_MODE (x),
6651 sge_plus_constant (XEXP (x, 1), c), XEXP (x, 0));
6656 sge_plus (mode, x, y)
6657 enum machine_mode mode;
6660 while (GET_CODE (y) == PLUS)
6662 rtx a = XEXP (y, 0);
6663 if (GET_CODE (a) == CONST_INT)
6664 x = sge_plus_constant (x, a);
6666 x = gen_rtx_PLUS (mode, x, a);
6669 if (GET_CODE (y) == CONST_INT)
6670 x = sge_plus_constant (x, y);
6672 x = gen_rtx_PLUS (mode, x, y);
6676 /* Help detect a giv that is calculated by several consecutive insns;
6680 The caller has already identified the first insn P as having a giv as dest;
6681 we check that all other insns that set the same register follow
6682 immediately after P, that they alter nothing else,
6683 and that the result of the last is still a giv.
6685 The value is 0 if the reg set in P is not really a giv.
6686 Otherwise, the value is the amount gained by eliminating
6687 all the consecutive insns that compute the value.
6689 FIRST_BENEFIT is the amount gained by eliminating the first insn, P.
6690 SRC_REG is the reg of the biv; DEST_REG is the reg of the giv.
6692 The coefficients of the ultimate giv value are stored in
6693 *MULT_VAL and *ADD_VAL. */
6696 consec_sets_giv (loop, first_benefit, p, src_reg, dest_reg,
6697 add_val, mult_val, last_consec_insn)
6698 const struct loop *loop;
6705 rtx *last_consec_insn;
6713 /* Indicate that this is a giv so that we can update the value produced in
6714 each insn of the multi-insn sequence.
6716 This induction structure will be used only by the call to
6717 general_induction_var below, so we can allocate it on our stack.
6718 If this is a giv, our caller will replace the induct var entry with
6719 a new induction structure. */
6721 = (struct induction *) alloca (sizeof (struct induction));
6722 v->src_reg = src_reg;
6723 v->mult_val = *mult_val;
6724 v->add_val = *add_val;
6725 v->benefit = first_benefit;
6727 v->derive_adjustment = 0;
6729 REG_IV_TYPE (REGNO (dest_reg)) = GENERAL_INDUCT;
6730 REG_IV_INFO (REGNO (dest_reg)) = v;
6732 count = VARRAY_INT (n_times_set, REGNO (dest_reg)) - 1;
6737 code = GET_CODE (p);
6739 /* If libcall, skip to end of call sequence. */
6740 if (code == INSN && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
6744 && (set = single_set (p))
6745 && GET_CODE (SET_DEST (set)) == REG
6746 && SET_DEST (set) == dest_reg
6747 && (general_induction_var (loop, SET_SRC (set), &src_reg,
6748 add_val, mult_val, 0, &benefit)
6749 /* Giv created by equivalent expression. */
6750 || ((temp = find_reg_note (p, REG_EQUAL, NULL_RTX))
6751 && general_induction_var (loop, XEXP (temp, 0), &src_reg,
6752 add_val, mult_val, 0, &benefit)))
6753 && src_reg == v->src_reg)
6755 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
6756 benefit += libcall_benefit (p);
6759 v->mult_val = *mult_val;
6760 v->add_val = *add_val;
6761 v->benefit = benefit;
6763 else if (code != NOTE)
6765 /* Allow insns that set something other than this giv to a
6766 constant. Such insns are needed on machines which cannot
6767 include long constants and should not disqualify a giv. */
6769 && (set = single_set (p))
6770 && SET_DEST (set) != dest_reg
6771 && CONSTANT_P (SET_SRC (set)))
6774 REG_IV_TYPE (REGNO (dest_reg)) = UNKNOWN_INDUCT;
6779 *last_consec_insn = p;
6783 /* Return an rtx, if any, that expresses giv G2 as a function of the register
6784 represented by G1. If no such expression can be found, or it is clear that
6785 it cannot possibly be a valid address, 0 is returned.
6787 To perform the computation, we note that
6790 where `v' is the biv.
6792 So G2 = (y/b) * G1 + (b - a*y/x).
6794 Note that MULT = y/x.
6796 Update: A and B are now allowed to be additive expressions such that
6797 B contains all variables in A. That is, computing B-A will not require
6798 subtracting variables. */
6801 express_from_1 (a, b, mult)
6804 /* If MULT is zero, then A*MULT is zero, and our expression is B. */
6806 if (mult == const0_rtx)
6809 /* If MULT is not 1, we cannot handle A with non-constants, since we
6810 would then be required to subtract multiples of the registers in A.
6811 This is theoretically possible, and may even apply to some Fortran
6812 constructs, but it is a lot of work and we do not attempt it here. */
6814 if (mult != const1_rtx && GET_CODE (a) != CONST_INT)
6817 /* In general these structures are sorted top to bottom (down the PLUS
6818 chain), but not left to right across the PLUS. If B is a higher
6819 order giv than A, we can strip one level and recurse. If A is higher
6820 order, we'll eventually bail out, but won't know that until the end.
6821 If they are the same, we'll strip one level around this loop. */
6823 while (GET_CODE (a) == PLUS && GET_CODE (b) == PLUS)
6825 rtx ra, rb, oa, ob, tmp;
6827 ra = XEXP (a, 0), oa = XEXP (a, 1);
6828 if (GET_CODE (ra) == PLUS)
6829 tmp = ra, ra = oa, oa = tmp;
6831 rb = XEXP (b, 0), ob = XEXP (b, 1);
6832 if (GET_CODE (rb) == PLUS)
6833 tmp = rb, rb = ob, ob = tmp;
6835 if (rtx_equal_p (ra, rb))
6836 /* We matched: remove one reg completely. */
6838 else if (GET_CODE (ob) != PLUS && rtx_equal_p (ra, ob))
6839 /* An alternate match. */
6841 else if (GET_CODE (oa) != PLUS && rtx_equal_p (oa, rb))
6842 /* An alternate match. */
6846 /* Indicates an extra register in B. Strip one level from B and
6847 recurse, hoping B was the higher order expression. */
6848 ob = express_from_1 (a, ob, mult);
6851 return gen_rtx_PLUS (GET_MODE (b), rb, ob);
6855 /* Here we are at the last level of A, go through the cases hoping to
6856 get rid of everything but a constant. */
6858 if (GET_CODE (a) == PLUS)
6862 ra = XEXP (a, 0), oa = XEXP (a, 1);
6863 if (rtx_equal_p (oa, b))
6865 else if (!rtx_equal_p (ra, b))
6868 if (GET_CODE (oa) != CONST_INT)
6871 return GEN_INT (-INTVAL (oa) * INTVAL (mult));
6873 else if (GET_CODE (a) == CONST_INT)
6875 return plus_constant (b, -INTVAL (a) * INTVAL (mult));
6877 else if (CONSTANT_P (a))
6879 return simplify_gen_binary (MINUS, GET_MODE (b), const0_rtx, a);
6881 else if (GET_CODE (b) == PLUS)
6883 if (rtx_equal_p (a, XEXP (b, 0)))
6885 else if (rtx_equal_p (a, XEXP (b, 1)))
6890 else if (rtx_equal_p (a, b))
6897 express_from (g1, g2)
6898 struct induction *g1, *g2;
6902 /* The value that G1 will be multiplied by must be a constant integer. Also,
6903 the only chance we have of getting a valid address is if b*c/a (see above
6904 for notation) is also an integer. */
6905 if (GET_CODE (g1->mult_val) == CONST_INT
6906 && GET_CODE (g2->mult_val) == CONST_INT)
6908 if (g1->mult_val == const0_rtx
6909 || INTVAL (g2->mult_val) % INTVAL (g1->mult_val) != 0)
6911 mult = GEN_INT (INTVAL (g2->mult_val) / INTVAL (g1->mult_val));
6913 else if (rtx_equal_p (g1->mult_val, g2->mult_val))
6917 /* ??? Find out if the one is a multiple of the other? */
6921 add = express_from_1 (g1->add_val, g2->add_val, mult);
6922 if (add == NULL_RTX)
6924 /* Failed. If we've got a multiplication factor between G1 and G2,
6925 scale G1's addend and try again. */
6926 if (INTVAL (mult) > 1)
6928 rtx g1_add_val = g1->add_val;
6929 if (GET_CODE (g1_add_val) == MULT
6930 && GET_CODE (XEXP (g1_add_val, 1)) == CONST_INT)
6933 m = INTVAL (mult) * INTVAL (XEXP (g1_add_val, 1));
6934 g1_add_val = gen_rtx_MULT (GET_MODE (g1_add_val),
6935 XEXP (g1_add_val, 0), GEN_INT (m));
6939 g1_add_val = gen_rtx_MULT (GET_MODE (g1_add_val), g1_add_val,
6943 add = express_from_1 (g1_add_val, g2->add_val, const1_rtx);
6946 if (add == NULL_RTX)
6949 /* Form simplified final result. */
6950 if (mult == const0_rtx)
6952 else if (mult == const1_rtx)
6953 mult = g1->dest_reg;
6955 mult = gen_rtx_MULT (g2->mode, g1->dest_reg, mult);
6957 if (add == const0_rtx)
6961 if (GET_CODE (add) == PLUS
6962 && CONSTANT_P (XEXP (add, 1)))
6964 rtx tem = XEXP (add, 1);
6965 mult = gen_rtx_PLUS (g2->mode, mult, XEXP (add, 0));
6969 return gen_rtx_PLUS (g2->mode, mult, add);
6974 /* Return an rtx, if any, that expresses giv G2 as a function of the register
6975 represented by G1. This indicates that G2 should be combined with G1 and
6976 that G2 can use (either directly or via an address expression) a register
6977 used to represent G1. */
6980 combine_givs_p (g1, g2)
6981 struct induction *g1, *g2;
6983 rtx tem = express_from (g1, g2);
6985 /* If these givs are identical, they can be combined. We use the results
6986 of express_from because the addends are not in a canonical form, so
6987 rtx_equal_p is a weaker test. */
6988 /* But don't combine a DEST_REG giv with a DEST_ADDR giv; we want the
6989 combination to be the other way round. */
6990 if (tem == g1->dest_reg
6991 && (g1->giv_type == DEST_REG || g2->giv_type == DEST_ADDR))
6993 return g1->dest_reg;
6996 /* If G2 can be expressed as a function of G1 and that function is valid
6997 as an address and no more expensive than using a register for G2,
6998 the expression of G2 in terms of G1 can be used. */
7000 && g2->giv_type == DEST_ADDR
7001 && memory_address_p (g2->mem_mode, tem)
7002 /* ??? Looses, especially with -fforce-addr, where *g2->location
7003 will always be a register, and so anything more complicated
7007 && ADDRESS_COST (tem) <= ADDRESS_COST (*g2->location)
7009 && rtx_cost (tem, MEM) <= rtx_cost (*g2->location, MEM)
7020 struct combine_givs_stats
7027 cmp_combine_givs_stats (xp, yp)
7031 const struct combine_givs_stats * const x =
7032 (const struct combine_givs_stats *) xp;
7033 const struct combine_givs_stats * const y =
7034 (const struct combine_givs_stats *) yp;
7036 d = y->total_benefit - x->total_benefit;
7037 /* Stabilize the sort. */
7039 d = x->giv_number - y->giv_number;
7043 /* Check all pairs of givs for iv_class BL and see if any can be combined with
7044 any other. If so, point SAME to the giv combined with and set NEW_REG to
7045 be an expression (in terms of the other giv's DEST_REG) equivalent to the
7046 giv. Also, update BENEFIT and related fields for cost/benefit analysis. */
7050 struct iv_class *bl;
7052 /* Additional benefit to add for being combined multiple times. */
7053 const int extra_benefit = 3;
7055 struct induction *g1, *g2, **giv_array;
7056 int i, j, k, giv_count;
7057 struct combine_givs_stats *stats;
7060 /* Count givs, because bl->giv_count is incorrect here. */
7062 for (g1 = bl->giv; g1; g1 = g1->next_iv)
7067 = (struct induction **) alloca (giv_count * sizeof (struct induction *));
7069 for (g1 = bl->giv; g1; g1 = g1->next_iv)
7071 giv_array[i++] = g1;
7073 stats = (struct combine_givs_stats *) xcalloc (giv_count, sizeof (*stats));
7074 can_combine = (rtx *) xcalloc (giv_count, giv_count * sizeof(rtx));
7076 for (i = 0; i < giv_count; i++)
7082 stats[i].giv_number = i;
7084 /* If a DEST_REG GIV is used only once, do not allow it to combine
7085 with anything, for in doing so we will gain nothing that cannot
7086 be had by simply letting the GIV with which we would have combined
7087 to be reduced on its own. The losage shows up in particular with
7088 DEST_ADDR targets on hosts with reg+reg addressing, though it can
7089 be seen elsewhere as well. */
7090 if (g1->giv_type == DEST_REG
7091 && (single_use = VARRAY_RTX (reg_single_usage, REGNO (g1->dest_reg)))
7092 && single_use != const0_rtx)
7095 this_benefit = g1->benefit;
7096 /* Add an additional weight for zero addends. */
7097 if (g1->no_const_addval)
7100 for (j = 0; j < giv_count; j++)
7106 && (this_combine = combine_givs_p (g1, g2)) != NULL_RTX)
7108 can_combine[i*giv_count + j] = this_combine;
7109 this_benefit += g2->benefit + extra_benefit;
7112 stats[i].total_benefit = this_benefit;
7115 /* Iterate, combining until we can't. */
7117 qsort (stats, giv_count, sizeof(*stats), cmp_combine_givs_stats);
7119 if (loop_dump_stream)
7121 fprintf (loop_dump_stream, "Sorted combine statistics:\n");
7122 for (k = 0; k < giv_count; k++)
7124 g1 = giv_array[stats[k].giv_number];
7125 if (!g1->combined_with && !g1->same)
7126 fprintf (loop_dump_stream, " {%d, %d}",
7127 INSN_UID (giv_array[stats[k].giv_number]->insn),
7128 stats[k].total_benefit);
7130 putc ('\n', loop_dump_stream);
7133 for (k = 0; k < giv_count; k++)
7135 int g1_add_benefit = 0;
7137 i = stats[k].giv_number;
7140 /* If it has already been combined, skip. */
7141 if (g1->combined_with || g1->same)
7144 for (j = 0; j < giv_count; j++)
7147 if (g1 != g2 && can_combine[i*giv_count + j]
7148 /* If it has already been combined, skip. */
7149 && ! g2->same && ! g2->combined_with)
7153 g2->new_reg = can_combine[i*giv_count + j];
7155 g1->combined_with++;
7156 g1->lifetime += g2->lifetime;
7158 g1_add_benefit += g2->benefit;
7160 /* ??? The new final_[bg]iv_value code does a much better job
7161 of finding replaceable giv's, and hence this code may no
7162 longer be necessary. */
7163 if (! g2->replaceable && REG_USERVAR_P (g2->dest_reg))
7164 g1_add_benefit -= copy_cost;
7166 /* To help optimize the next set of combinations, remove
7167 this giv from the benefits of other potential mates. */
7168 for (l = 0; l < giv_count; ++l)
7170 int m = stats[l].giv_number;
7171 if (can_combine[m*giv_count + j])
7172 stats[l].total_benefit -= g2->benefit + extra_benefit;
7175 if (loop_dump_stream)
7176 fprintf (loop_dump_stream,
7177 "giv at %d combined with giv at %d\n",
7178 INSN_UID (g2->insn), INSN_UID (g1->insn));
7182 /* To help optimize the next set of combinations, remove
7183 this giv from the benefits of other potential mates. */
7184 if (g1->combined_with)
7186 for (j = 0; j < giv_count; ++j)
7188 int m = stats[j].giv_number;
7189 if (can_combine[m*giv_count + i])
7190 stats[j].total_benefit -= g1->benefit + extra_benefit;
7193 g1->benefit += g1_add_benefit;
7195 /* We've finished with this giv, and everything it touched.
7196 Restart the combination so that proper weights for the
7197 rest of the givs are properly taken into account. */
7198 /* ??? Ideally we would compact the arrays at this point, so
7199 as to not cover old ground. But sanely compacting
7200 can_combine is tricky. */
7210 struct recombine_givs_stats
7213 int start_luid, end_luid;
7216 /* Used below as comparison function for qsort. We want a ascending luid
7217 when scanning the array starting at the end, thus the arguments are
7220 cmp_recombine_givs_stats (xp, yp)
7224 const struct recombine_givs_stats * const x =
7225 (const struct recombine_givs_stats *) xp;
7226 const struct recombine_givs_stats * const y =
7227 (const struct recombine_givs_stats *) yp;
7229 d = y->start_luid - x->start_luid;
7230 /* Stabilize the sort. */
7232 d = y->giv_number - x->giv_number;
7236 /* Scan X, which is a part of INSN, for the end of life of a giv. Also
7237 look for the start of life of a giv where the start has not been seen
7238 yet to unlock the search for the end of its life.
7239 Only consider givs that belong to BIV.
7240 Return the total number of lifetime ends that have been found. */
7242 find_life_end (x, stats, insn, biv)
7244 struct recombine_givs_stats *stats;
7251 code = GET_CODE (x);
7256 rtx reg = SET_DEST (x);
7257 if (GET_CODE (reg) == REG)
7259 int regno = REGNO (reg);
7260 struct induction *v = REG_IV_INFO (regno);
7262 if (REG_IV_TYPE (regno) == GENERAL_INDUCT
7264 && v->src_reg == biv
7265 && stats[v->ix].end_luid <= 0)
7267 /* If we see a 0 here for end_luid, it means that we have
7268 scanned the entire loop without finding any use at all.
7269 We must not predicate this code on a start_luid match
7270 since that would make the test fail for givs that have
7271 been hoisted out of inner loops. */
7272 if (stats[v->ix].end_luid == 0)
7274 stats[v->ix].end_luid = stats[v->ix].start_luid;
7275 return 1 + find_life_end (SET_SRC (x), stats, insn, biv);
7277 else if (stats[v->ix].start_luid == INSN_LUID (insn))
7278 stats[v->ix].end_luid = 0;
7280 return find_life_end (SET_SRC (x), stats, insn, biv);
7286 int regno = REGNO (x);
7287 struct induction *v = REG_IV_INFO (regno);
7289 if (REG_IV_TYPE (regno) == GENERAL_INDUCT
7291 && v->src_reg == biv
7292 && stats[v->ix].end_luid == 0)
7294 while (INSN_UID (insn) >= max_uid_for_loop)
7295 insn = NEXT_INSN (insn);
7296 stats[v->ix].end_luid = INSN_LUID (insn);
7309 fmt = GET_RTX_FORMAT (code);
7311 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
7314 retval += find_life_end (XEXP (x, i), stats, insn, biv);
7316 else if (fmt[i] == 'E')
7317 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
7318 retval += find_life_end (XVECEXP (x, i, j), stats, insn, biv);
7323 /* For each giv that has been combined with another, look if
7324 we can combine it with the most recently used one instead.
7325 This tends to shorten giv lifetimes, and helps the next step:
7326 try to derive givs from other givs. */
7328 recombine_givs (loop, bl, unroll_p)
7329 const struct loop *loop;
7330 struct iv_class *bl;
7333 struct induction *v, **giv_array, *last_giv;
7334 struct recombine_givs_stats *stats;
7337 int ends_need_computing;
7339 for (giv_count = 0, v = bl->giv; v; v = v->next_iv)
7345 = (struct induction **) xmalloc (giv_count * sizeof (struct induction *));
7346 stats = (struct recombine_givs_stats *) xmalloc (giv_count * sizeof *stats);
7348 /* Initialize stats and set up the ix field for each giv in stats to name
7349 the corresponding index into stats. */
7350 for (i = 0, v = bl->giv; v; v = v->next_iv)
7357 stats[i].giv_number = i;
7358 /* If this giv has been hoisted out of an inner loop, use the luid of
7359 the previous insn. */
7360 for (p = v->insn; INSN_UID (p) >= max_uid_for_loop; )
7362 stats[i].start_luid = INSN_LUID (p);
7366 qsort (stats, giv_count, sizeof(*stats), cmp_recombine_givs_stats);
7368 /* Set up the ix field for each giv in stats to name
7369 the corresponding index into stats, and
7370 do the actual most-recently-used recombination. */
7371 for (last_giv = 0, i = giv_count - 1; i >= 0; i--)
7373 v = giv_array[stats[i].giv_number];
7377 struct induction *old_same = v->same;
7380 /* combine_givs_p actually says if we can make this transformation.
7381 The other tests are here only to avoid keeping a giv alive
7382 that could otherwise be eliminated. */
7384 && ((old_same->maybe_dead && ! old_same->combined_with)
7385 || ! last_giv->maybe_dead
7386 || last_giv->combined_with)
7387 && (new_combine = combine_givs_p (last_giv, v)))
7389 old_same->combined_with--;
7390 v->new_reg = new_combine;
7392 last_giv->combined_with++;
7393 /* No need to update lifetimes / benefits here since we have
7394 already decided what to reduce. */
7396 if (loop_dump_stream)
7398 fprintf (loop_dump_stream,
7399 "giv at %d recombined with giv at %d as ",
7400 INSN_UID (v->insn), INSN_UID (last_giv->insn));
7401 print_rtl (loop_dump_stream, v->new_reg);
7402 putc ('\n', loop_dump_stream);
7408 else if (v->giv_type != DEST_REG)
7411 || (last_giv->maybe_dead && ! last_giv->combined_with)
7413 || v->combined_with)
7417 ends_need_computing = 0;
7418 /* For each DEST_REG giv, compute lifetime starts, and try to compute
7419 lifetime ends from regscan info. */
7420 for (i = giv_count - 1; i >= 0; i--)
7422 v = giv_array[stats[i].giv_number];
7425 if (v->giv_type == DEST_ADDR)
7427 /* Loop unrolling of an inner loop can even create new DEST_REG
7430 for (p = v->insn; INSN_UID (p) >= max_uid_for_loop; )
7432 stats[i].start_luid = stats[i].end_luid = INSN_LUID (p);
7434 stats[i].end_luid++;
7436 else /* v->giv_type == DEST_REG */
7440 stats[i].start_luid = INSN_LUID (v->insn);
7441 stats[i].end_luid = INSN_LUID (v->last_use);
7443 else if (INSN_UID (v->insn) >= max_uid_for_loop)
7446 /* This insn has been created by loop optimization on an inner
7447 loop. We don't have a proper start_luid that will match
7448 when we see the first set. But we do know that there will
7449 be no use before the set, so we can set end_luid to 0 so that
7450 we'll start looking for the last use right away. */
7451 for (p = PREV_INSN (v->insn); INSN_UID (p) >= max_uid_for_loop; )
7453 stats[i].start_luid = INSN_LUID (p);
7454 stats[i].end_luid = 0;
7455 ends_need_computing++;
7459 int regno = REGNO (v->dest_reg);
7460 int count = VARRAY_INT (n_times_set, regno) - 1;
7463 /* Find the first insn that sets the giv, so that we can verify
7464 if this giv's lifetime wraps around the loop. We also need
7465 the luid of the first setting insn in order to detect the
7466 last use properly. */
7469 p = prev_nonnote_insn (p);
7470 if (reg_set_p (v->dest_reg, p))
7474 stats[i].start_luid = INSN_LUID (p);
7475 if (stats[i].start_luid > uid_luid[REGNO_FIRST_UID (regno)])
7477 stats[i].end_luid = -1;
7478 ends_need_computing++;
7482 stats[i].end_luid = uid_luid[REGNO_LAST_UID (regno)];
7483 if (stats[i].end_luid > INSN_LUID (loop->end))
7485 stats[i].end_luid = -1;
7486 ends_need_computing++;
7493 /* If the regscan information was unconclusive for one or more DEST_REG
7494 givs, scan the all insn in the loop to find out lifetime ends. */
7495 if (ends_need_computing)
7497 rtx biv = bl->biv->src_reg;
7502 if (p == loop->start)
7505 if (GET_RTX_CLASS (GET_CODE (p)) != 'i')
7507 ends_need_computing -= find_life_end (PATTERN (p), stats, p, biv);
7509 while (ends_need_computing);
7512 /* Set start_luid back to the last insn that sets the giv. This allows
7513 more combinations. */
7514 for (i = giv_count - 1; i >= 0; i--)
7516 v = giv_array[stats[i].giv_number];
7519 if (INSN_UID (v->insn) < max_uid_for_loop)
7520 stats[i].start_luid = INSN_LUID (v->insn);
7523 /* Now adjust lifetime ends by taking combined givs into account. */
7524 for (i = giv_count - 1; i >= 0; i--)
7529 v = giv_array[stats[i].giv_number];
7532 if (v->same && ! v->same->ignore)
7535 luid = stats[i].start_luid;
7536 /* Use unsigned arithmetic to model loop wrap-around. */
7537 if (luid - stats[j].start_luid
7538 > (unsigned) stats[j].end_luid - stats[j].start_luid)
7539 stats[j].end_luid = luid;
7543 qsort (stats, giv_count, sizeof(*stats), cmp_recombine_givs_stats);
7545 /* Try to derive DEST_REG givs from previous DEST_REG givs with the
7546 same mult_val and non-overlapping lifetime. This reduces register
7548 Once we find a DEST_REG giv that is suitable to derive others from,
7549 we set last_giv to this giv, and try to derive as many other DEST_REG
7550 givs from it without joining overlapping lifetimes. If we then
7551 encounter a DEST_REG giv that we can't derive, we set rescan to the
7552 index for this giv (unless rescan is already set).
7553 When we are finished with the current LAST_GIV (i.e. the inner loop
7554 terminates), we start again with rescan, which then becomes the new
7556 for (i = giv_count - 1; i >= 0; i = rescan)
7558 int life_start = 0, life_end = 0;
7560 for (last_giv = 0, rescan = -1; i >= 0; i--)
7564 v = giv_array[stats[i].giv_number];
7565 if (v->giv_type != DEST_REG || v->derived_from || v->same)
7569 /* Don't use a giv that's likely to be dead to derive
7570 others - that would be likely to keep that giv alive. */
7571 if (! v->maybe_dead || v->combined_with)
7574 life_start = stats[i].start_luid;
7575 life_end = stats[i].end_luid;
7579 /* Use unsigned arithmetic to model loop wrap around. */
7580 if (((unsigned) stats[i].start_luid - life_start
7581 >= (unsigned) life_end - life_start)
7582 && ((unsigned) stats[i].end_luid - life_start
7583 > (unsigned) life_end - life_start)
7584 /* Check that the giv insn we're about to use for deriving
7585 precedes all uses of that giv. Note that initializing the
7586 derived giv would defeat the purpose of reducing register
7588 ??? We could arrange to move the insn. */
7589 && ((unsigned) stats[i].end_luid - INSN_LUID (loop->start)
7590 > (unsigned) stats[i].start_luid - INSN_LUID (loop->start))
7591 && rtx_equal_p (last_giv->mult_val, v->mult_val)
7592 /* ??? Could handle libcalls, but would need more logic. */
7593 && ! find_reg_note (v->insn, REG_RETVAL, NULL_RTX)
7594 /* We would really like to know if for any giv that v
7595 is combined with, v->insn or any intervening biv increment
7596 dominates that combined giv. However, we
7597 don't have this detailed control flow information.
7598 N.B. since last_giv will be reduced, it is valid
7599 anywhere in the loop, so we don't need to check the
7600 validity of last_giv.
7601 We rely here on the fact that v->always_executed implies that
7602 there is no jump to someplace else in the loop before the
7603 giv insn, and hence any insn that is executed before the
7604 giv insn in the loop will have a lower luid. */
7605 && (v->always_executed || ! v->combined_with)
7606 && (sum = express_from (last_giv, v))
7607 /* Make sure we don't make the add more expensive. ADD_COST
7608 doesn't take different costs of registers and constants into
7609 account, so compare the cost of the actual SET_SRCs. */
7610 && (rtx_cost (sum, SET)
7611 <= rtx_cost (SET_SRC (single_set (v->insn)), SET))
7612 /* ??? unroll can't understand anything but reg + const_int
7613 sums. It would be cleaner to fix unroll. */
7614 && ((GET_CODE (sum) == PLUS
7615 && GET_CODE (XEXP (sum, 0)) == REG
7616 && GET_CODE (XEXP (sum, 1)) == CONST_INT)
7618 && validate_change (v->insn, &PATTERN (v->insn),
7619 gen_rtx_SET (VOIDmode, v->dest_reg, sum), 0))
7621 v->derived_from = last_giv;
7622 life_end = stats[i].end_luid;
7624 if (loop_dump_stream)
7626 fprintf (loop_dump_stream,
7627 "giv at %d derived from %d as ",
7628 INSN_UID (v->insn), INSN_UID (last_giv->insn));
7629 print_rtl (loop_dump_stream, sum);
7630 putc ('\n', loop_dump_stream);
7633 else if (rescan < 0)
7643 /* EMIT code before INSERT_BEFORE to set REG = B * M + A. */
7646 emit_iv_add_mult (b, m, a, reg, insert_before)
7647 rtx b; /* initial value of basic induction variable */
7648 rtx m; /* multiplicative constant */
7649 rtx a; /* additive constant */
7650 rtx reg; /* destination register */
7656 /* Prevent unexpected sharing of these rtx. */
7660 /* Increase the lifetime of any invariants moved further in code. */
7661 update_reg_last_use (a, insert_before);
7662 update_reg_last_use (b, insert_before);
7663 update_reg_last_use (m, insert_before);
7666 result = expand_mult_add (b, reg, m, a, GET_MODE (reg), 0);
7668 emit_move_insn (reg, result);
7669 seq = gen_sequence ();
7672 emit_insn_before (seq, insert_before);
7674 /* It is entirely possible that the expansion created lots of new
7675 registers. Iterate over the sequence we just created and
7678 if (GET_CODE (seq) == SEQUENCE)
7681 for (i = 0; i < XVECLEN (seq, 0); ++i)
7683 rtx set = single_set (XVECEXP (seq, 0, i));
7684 if (set && GET_CODE (SET_DEST (set)) == REG)
7685 record_base_value (REGNO (SET_DEST (set)), SET_SRC (set), 0);
7688 else if (GET_CODE (seq) == SET
7689 && GET_CODE (SET_DEST (seq)) == REG)
7690 record_base_value (REGNO (SET_DEST (seq)), SET_SRC (seq), 0);
7693 /* Test whether A * B can be computed without
7694 an actual multiply insn. Value is 1 if so. */
7697 product_cheap_p (a, b)
7703 struct obstack *old_rtl_obstack = rtl_obstack;
7704 char *storage = (char *) obstack_alloc (&temp_obstack, 0);
7707 /* If only one is constant, make it B. */
7708 if (GET_CODE (a) == CONST_INT)
7709 tmp = a, a = b, b = tmp;
7711 /* If first constant, both constant, so don't need multiply. */
7712 if (GET_CODE (a) == CONST_INT)
7715 /* If second not constant, neither is constant, so would need multiply. */
7716 if (GET_CODE (b) != CONST_INT)
7719 /* One operand is constant, so might not need multiply insn. Generate the
7720 code for the multiply and see if a call or multiply, or long sequence
7721 of insns is generated. */
7723 rtl_obstack = &temp_obstack;
7725 expand_mult (GET_MODE (a), a, b, NULL_RTX, 0);
7726 tmp = gen_sequence ();
7729 if (GET_CODE (tmp) == SEQUENCE)
7731 if (XVEC (tmp, 0) == 0)
7733 else if (XVECLEN (tmp, 0) > 3)
7736 for (i = 0; i < XVECLEN (tmp, 0); i++)
7738 rtx insn = XVECEXP (tmp, 0, i);
7740 if (GET_CODE (insn) != INSN
7741 || (GET_CODE (PATTERN (insn)) == SET
7742 && GET_CODE (SET_SRC (PATTERN (insn))) == MULT)
7743 || (GET_CODE (PATTERN (insn)) == PARALLEL
7744 && GET_CODE (XVECEXP (PATTERN (insn), 0, 0)) == SET
7745 && GET_CODE (SET_SRC (XVECEXP (PATTERN (insn), 0, 0))) == MULT))
7752 else if (GET_CODE (tmp) == SET
7753 && GET_CODE (SET_SRC (tmp)) == MULT)
7755 else if (GET_CODE (tmp) == PARALLEL
7756 && GET_CODE (XVECEXP (tmp, 0, 0)) == SET
7757 && GET_CODE (SET_SRC (XVECEXP (tmp, 0, 0))) == MULT)
7760 /* Free any storage we obtained in generating this multiply and restore rtl
7761 allocation to its normal obstack. */
7762 obstack_free (&temp_obstack, storage);
7763 rtl_obstack = old_rtl_obstack;
7768 /* Check to see if loop can be terminated by a "decrement and branch until
7769 zero" instruction. If so, add a REG_NONNEG note to the branch insn if so.
7770 Also try reversing an increment loop to a decrement loop
7771 to see if the optimization can be performed.
7772 Value is nonzero if optimization was performed. */
7774 /* This is useful even if the architecture doesn't have such an insn,
7775 because it might change a loops which increments from 0 to n to a loop
7776 which decrements from n to 0. A loop that decrements to zero is usually
7777 faster than one that increments from zero. */
7779 /* ??? This could be rewritten to use some of the loop unrolling procedures,
7780 such as approx_final_value, biv_total_increment, loop_iterations, and
7781 final_[bg]iv_value. */
7784 check_dbra_loop (loop, insn_count)
7788 struct iv_class *bl;
7795 rtx before_comparison;
7799 int compare_and_branch;
7800 rtx loop_start = loop->start;
7801 rtx loop_end = loop->end;
7802 struct loop_info *loop_info = LOOP_INFO (loop);
7804 /* If last insn is a conditional branch, and the insn before tests a
7805 register value, try to optimize it. Otherwise, we can't do anything. */
7807 jump = PREV_INSN (loop_end);
7808 comparison = get_condition_for_loop (loop, jump);
7809 if (comparison == 0)
7812 /* Try to compute whether the compare/branch at the loop end is one or
7813 two instructions. */
7814 get_condition (jump, &first_compare);
7815 if (first_compare == jump)
7816 compare_and_branch = 1;
7817 else if (first_compare == prev_nonnote_insn (jump))
7818 compare_and_branch = 2;
7822 /* Check all of the bivs to see if the compare uses one of them.
7823 Skip biv's set more than once because we can't guarantee that
7824 it will be zero on the last iteration. Also skip if the biv is
7825 used between its update and the test insn. */
7827 for (bl = loop_iv_list; bl; bl = bl->next)
7829 if (bl->biv_count == 1
7830 && ! bl->biv->maybe_multiple
7831 && bl->biv->dest_reg == XEXP (comparison, 0)
7832 && ! reg_used_between_p (regno_reg_rtx[bl->regno], bl->biv->insn,
7840 /* Look for the case where the basic induction variable is always
7841 nonnegative, and equals zero on the last iteration.
7842 In this case, add a reg_note REG_NONNEG, which allows the
7843 m68k DBRA instruction to be used. */
7845 if (((GET_CODE (comparison) == GT
7846 && GET_CODE (XEXP (comparison, 1)) == CONST_INT
7847 && INTVAL (XEXP (comparison, 1)) == -1)
7848 || (GET_CODE (comparison) == NE && XEXP (comparison, 1) == const0_rtx))
7849 && GET_CODE (bl->biv->add_val) == CONST_INT
7850 && INTVAL (bl->biv->add_val) < 0)
7852 /* Initial value must be greater than 0,
7853 init_val % -dec_value == 0 to ensure that it equals zero on
7854 the last iteration */
7856 if (GET_CODE (bl->initial_value) == CONST_INT
7857 && INTVAL (bl->initial_value) > 0
7858 && (INTVAL (bl->initial_value)
7859 % (-INTVAL (bl->biv->add_val))) == 0)
7861 /* register always nonnegative, add REG_NOTE to branch */
7862 REG_NOTES (PREV_INSN (loop_end))
7863 = gen_rtx_EXPR_LIST (REG_NONNEG, NULL_RTX,
7864 REG_NOTES (PREV_INSN (loop_end)));
7870 /* If the decrement is 1 and the value was tested as >= 0 before
7871 the loop, then we can safely optimize. */
7872 for (p = loop_start; p; p = PREV_INSN (p))
7874 if (GET_CODE (p) == CODE_LABEL)
7876 if (GET_CODE (p) != JUMP_INSN)
7879 before_comparison = get_condition_for_loop (loop, p);
7880 if (before_comparison
7881 && XEXP (before_comparison, 0) == bl->biv->dest_reg
7882 && GET_CODE (before_comparison) == LT
7883 && XEXP (before_comparison, 1) == const0_rtx
7884 && ! reg_set_between_p (bl->biv->dest_reg, p, loop_start)
7885 && INTVAL (bl->biv->add_val) == -1)
7887 REG_NOTES (PREV_INSN (loop_end))
7888 = gen_rtx_EXPR_LIST (REG_NONNEG, NULL_RTX,
7889 REG_NOTES (PREV_INSN (loop_end)));
7896 else if (GET_CODE (bl->biv->add_val) == CONST_INT
7897 && INTVAL (bl->biv->add_val) > 0)
7899 /* Try to change inc to dec, so can apply above optimization. */
7901 all registers modified are induction variables or invariant,
7902 all memory references have non-overlapping addresses
7903 (obviously true if only one write)
7904 allow 2 insns for the compare/jump at the end of the loop. */
7905 /* Also, we must avoid any instructions which use both the reversed
7906 biv and another biv. Such instructions will fail if the loop is
7907 reversed. We meet this condition by requiring that either
7908 no_use_except_counting is true, or else that there is only
7910 int num_nonfixed_reads = 0;
7911 /* 1 if the iteration var is used only to count iterations. */
7912 int no_use_except_counting = 0;
7913 /* 1 if the loop has no memory store, or it has a single memory store
7914 which is reversible. */
7915 int reversible_mem_store = 1;
7917 if (bl->giv_count == 0 && ! loop->exit_count)
7919 rtx bivreg = regno_reg_rtx[bl->regno];
7921 /* If there are no givs for this biv, and the only exit is the
7922 fall through at the end of the loop, then
7923 see if perhaps there are no uses except to count. */
7924 no_use_except_counting = 1;
7925 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
7926 if (GET_RTX_CLASS (GET_CODE (p)) == 'i')
7928 rtx set = single_set (p);
7930 if (set && GET_CODE (SET_DEST (set)) == REG
7931 && REGNO (SET_DEST (set)) == bl->regno)
7932 /* An insn that sets the biv is okay. */
7934 else if ((p == prev_nonnote_insn (prev_nonnote_insn (loop_end))
7935 || p == prev_nonnote_insn (loop_end))
7936 && reg_mentioned_p (bivreg, PATTERN (p)))
7938 /* If either of these insns uses the biv and sets a pseudo
7939 that has more than one usage, then the biv has uses
7940 other than counting since it's used to derive a value
7941 that is used more than one time. */
7942 int note_set_pseudo_multiple_uses_retval = 0;
7943 note_stores (PATTERN (p), note_set_pseudo_multiple_uses,
7944 ¬e_set_pseudo_multiple_uses_retval);
7945 if (note_set_pseudo_multiple_uses_retval)
7947 no_use_except_counting = 0;
7951 else if (reg_mentioned_p (bivreg, PATTERN (p)))
7953 no_use_except_counting = 0;
7959 if (no_use_except_counting)
7960 ; /* no need to worry about MEMs. */
7961 else if (num_mem_sets <= 1)
7963 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
7964 if (GET_RTX_CLASS (GET_CODE (p)) == 'i')
7965 num_nonfixed_reads += count_nonfixed_reads (loop, PATTERN (p));
7967 /* If the loop has a single store, and the destination address is
7968 invariant, then we can't reverse the loop, because this address
7969 might then have the wrong value at loop exit.
7970 This would work if the source was invariant also, however, in that
7971 case, the insn should have been moved out of the loop. */
7973 if (num_mem_sets == 1)
7975 struct induction *v;
7977 reversible_mem_store
7978 = (! unknown_address_altered
7979 && ! unknown_constant_address_altered
7980 && ! loop_invariant_p (loop,
7981 XEXP (XEXP (loop_store_mems, 0),
7984 /* If the store depends on a register that is set after the
7985 store, it depends on the initial value, and is thus not
7987 for (v = bl->giv; reversible_mem_store && v; v = v->next_iv)
7989 if (v->giv_type == DEST_REG
7990 && reg_mentioned_p (v->dest_reg,
7991 PATTERN (first_loop_store_insn))
7992 && loop_insn_first_p (first_loop_store_insn, v->insn))
7993 reversible_mem_store = 0;
8000 /* This code only acts for innermost loops. Also it simplifies
8001 the memory address check by only reversing loops with
8002 zero or one memory access.
8003 Two memory accesses could involve parts of the same array,
8004 and that can't be reversed.
8005 If the biv is used only for counting, than we don't need to worry
8006 about all these things. */
8008 if ((num_nonfixed_reads <= 1
8009 && ! loop_info->has_call
8010 && ! loop_info->has_volatile
8011 && reversible_mem_store
8012 && (bl->giv_count + bl->biv_count + num_mem_sets
8013 + num_movables + compare_and_branch == insn_count)
8014 && (bl == loop_iv_list && bl->next == 0))
8015 || no_use_except_counting)
8019 /* Loop can be reversed. */
8020 if (loop_dump_stream)
8021 fprintf (loop_dump_stream, "Can reverse loop\n");
8023 /* Now check other conditions:
8025 The increment must be a constant, as must the initial value,
8026 and the comparison code must be LT.
8028 This test can probably be improved since +/- 1 in the constant
8029 can be obtained by changing LT to LE and vice versa; this is
8033 /* for constants, LE gets turned into LT */
8034 && (GET_CODE (comparison) == LT
8035 || (GET_CODE (comparison) == LE
8036 && no_use_except_counting)))
8038 HOST_WIDE_INT add_val, add_adjust, comparison_val = 0;
8039 rtx initial_value, comparison_value;
8041 enum rtx_code cmp_code;
8042 int comparison_const_width;
8043 unsigned HOST_WIDE_INT comparison_sign_mask;
8045 add_val = INTVAL (bl->biv->add_val);
8046 comparison_value = XEXP (comparison, 1);
8047 if (GET_MODE (comparison_value) == VOIDmode)
8048 comparison_const_width
8049 = GET_MODE_BITSIZE (GET_MODE (XEXP (comparison, 0)));
8051 comparison_const_width
8052 = GET_MODE_BITSIZE (GET_MODE (comparison_value));
8053 if (comparison_const_width > HOST_BITS_PER_WIDE_INT)
8054 comparison_const_width = HOST_BITS_PER_WIDE_INT;
8055 comparison_sign_mask
8056 = (unsigned HOST_WIDE_INT)1 << (comparison_const_width - 1);
8058 /* If the comparison value is not a loop invariant, then we
8059 can not reverse this loop.
8061 ??? If the insns which initialize the comparison value as
8062 a whole compute an invariant result, then we could move
8063 them out of the loop and proceed with loop reversal. */
8064 if (! loop_invariant_p (loop, comparison_value))
8067 if (GET_CODE (comparison_value) == CONST_INT)
8068 comparison_val = INTVAL (comparison_value);
8069 initial_value = bl->initial_value;
8071 /* Normalize the initial value if it is an integer and
8072 has no other use except as a counter. This will allow
8073 a few more loops to be reversed. */
8074 if (no_use_except_counting
8075 && GET_CODE (comparison_value) == CONST_INT
8076 && GET_CODE (initial_value) == CONST_INT)
8078 comparison_val = comparison_val - INTVAL (bl->initial_value);
8079 /* The code below requires comparison_val to be a multiple
8080 of add_val in order to do the loop reversal, so
8081 round up comparison_val to a multiple of add_val.
8082 Since comparison_value is constant, we know that the
8083 current comparison code is LT. */
8084 comparison_val = comparison_val + add_val - 1;
8086 -= (unsigned HOST_WIDE_INT) comparison_val % add_val;
8087 /* We postpone overflow checks for COMPARISON_VAL here;
8088 even if there is an overflow, we might still be able to
8089 reverse the loop, if converting the loop exit test to
8091 initial_value = const0_rtx;
8094 /* First check if we can do a vanilla loop reversal. */
8095 if (initial_value == const0_rtx
8096 /* If we have a decrement_and_branch_on_count,
8097 prefer the NE test, since this will allow that
8098 instruction to be generated. Note that we must
8099 use a vanilla loop reversal if the biv is used to
8100 calculate a giv or has a non-counting use. */
8101 #if ! defined (HAVE_decrement_and_branch_until_zero) \
8102 && defined (HAVE_decrement_and_branch_on_count)
8103 && (! (add_val == 1 && loop->vtop
8104 && (bl->biv_count == 0
8105 || no_use_except_counting)))
8107 && GET_CODE (comparison_value) == CONST_INT
8108 /* Now do postponed overflow checks on COMPARISON_VAL. */
8109 && ! (((comparison_val - add_val) ^ INTVAL (comparison_value))
8110 & comparison_sign_mask))
8112 /* Register will always be nonnegative, with value
8113 0 on last iteration */
8114 add_adjust = add_val;
8118 else if (add_val == 1 && loop->vtop
8119 && (bl->biv_count == 0
8120 || no_use_except_counting))
8128 if (GET_CODE (comparison) == LE)
8129 add_adjust -= add_val;
8131 /* If the initial value is not zero, or if the comparison
8132 value is not an exact multiple of the increment, then we
8133 can not reverse this loop. */
8134 if (initial_value == const0_rtx
8135 && GET_CODE (comparison_value) == CONST_INT)
8137 if (((unsigned HOST_WIDE_INT) comparison_val % add_val) != 0)
8142 if (! no_use_except_counting || add_val != 1)
8146 final_value = comparison_value;
8148 /* Reset these in case we normalized the initial value
8149 and comparison value above. */
8150 if (GET_CODE (comparison_value) == CONST_INT
8151 && GET_CODE (initial_value) == CONST_INT)
8153 comparison_value = GEN_INT (comparison_val);
8155 = GEN_INT (comparison_val + INTVAL (bl->initial_value));
8157 bl->initial_value = initial_value;
8159 /* Save some info needed to produce the new insns. */
8160 reg = bl->biv->dest_reg;
8161 jump_label = XEXP (SET_SRC (PATTERN (PREV_INSN (loop_end))), 1);
8162 if (jump_label == pc_rtx)
8163 jump_label = XEXP (SET_SRC (PATTERN (PREV_INSN (loop_end))), 2);
8164 new_add_val = GEN_INT (- INTVAL (bl->biv->add_val));
8166 /* Set start_value; if this is not a CONST_INT, we need
8168 Initialize biv to start_value before loop start.
8169 The old initializing insn will be deleted as a
8170 dead store by flow.c. */
8171 if (initial_value == const0_rtx
8172 && GET_CODE (comparison_value) == CONST_INT)
8174 start_value = GEN_INT (comparison_val - add_adjust);
8175 emit_insn_before (gen_move_insn (reg, start_value),
8178 else if (GET_CODE (initial_value) == CONST_INT)
8180 rtx offset = GEN_INT (-INTVAL (initial_value) - add_adjust);
8181 enum machine_mode mode = GET_MODE (reg);
8182 enum insn_code icode
8183 = add_optab->handlers[(int) mode].insn_code;
8185 if (! (*insn_data[icode].operand[0].predicate) (reg, mode)
8186 || ! ((*insn_data[icode].operand[1].predicate)
8187 (comparison_value, mode))
8188 || ! ((*insn_data[icode].operand[2].predicate)
8192 = gen_rtx_PLUS (mode, comparison_value, offset);
8193 emit_insn_before ((GEN_FCN (icode)
8194 (reg, comparison_value, offset)),
8196 if (GET_CODE (comparison) == LE)
8197 final_value = gen_rtx_PLUS (mode, comparison_value,
8200 else if (! add_adjust)
8202 enum machine_mode mode = GET_MODE (reg);
8203 enum insn_code icode
8204 = sub_optab->handlers[(int) mode].insn_code;
8205 if (! (*insn_data[icode].operand[0].predicate) (reg, mode)
8206 || ! ((*insn_data[icode].operand[1].predicate)
8207 (comparison_value, mode))
8208 || ! ((*insn_data[icode].operand[2].predicate)
8209 (initial_value, mode)))
8212 = gen_rtx_MINUS (mode, comparison_value, initial_value);
8213 emit_insn_before ((GEN_FCN (icode)
8214 (reg, comparison_value, initial_value)),
8218 /* We could handle the other cases too, but it'll be
8219 better to have a testcase first. */
8222 /* We may not have a single insn which can increment a reg, so
8223 create a sequence to hold all the insns from expand_inc. */
8225 expand_inc (reg, new_add_val);
8226 tem = gen_sequence ();
8229 p = emit_insn_before (tem, bl->biv->insn);
8230 delete_insn (bl->biv->insn);
8232 /* Update biv info to reflect its new status. */
8234 bl->initial_value = start_value;
8235 bl->biv->add_val = new_add_val;
8237 /* Update loop info. */
8238 loop_info->initial_value = reg;
8239 loop_info->initial_equiv_value = reg;
8240 loop_info->final_value = const0_rtx;
8241 loop_info->final_equiv_value = const0_rtx;
8242 loop_info->comparison_value = const0_rtx;
8243 loop_info->comparison_code = cmp_code;
8244 loop_info->increment = new_add_val;
8246 /* Inc LABEL_NUSES so that delete_insn will
8247 not delete the label. */
8248 LABEL_NUSES (XEXP (jump_label, 0)) ++;
8250 /* Emit an insn after the end of the loop to set the biv's
8251 proper exit value if it is used anywhere outside the loop. */
8252 if ((REGNO_LAST_UID (bl->regno) != INSN_UID (first_compare))
8254 || REGNO_FIRST_UID (bl->regno) != INSN_UID (bl->init_insn))
8255 emit_insn_after (gen_move_insn (reg, final_value),
8258 /* Delete compare/branch at end of loop. */
8259 delete_insn (PREV_INSN (loop_end));
8260 if (compare_and_branch == 2)
8261 delete_insn (first_compare);
8263 /* Add new compare/branch insn at end of loop. */
8265 emit_cmp_and_jump_insns (reg, const0_rtx, cmp_code, NULL_RTX,
8266 GET_MODE (reg), 0, 0,
8267 XEXP (jump_label, 0));
8268 tem = gen_sequence ();
8270 emit_jump_insn_before (tem, loop_end);
8272 for (tem = PREV_INSN (loop_end);
8273 tem && GET_CODE (tem) != JUMP_INSN;
8274 tem = PREV_INSN (tem))
8278 JUMP_LABEL (tem) = XEXP (jump_label, 0);
8284 /* Increment of LABEL_NUSES done above. */
8285 /* Register is now always nonnegative,
8286 so add REG_NONNEG note to the branch. */
8287 REG_NOTES (tem) = gen_rtx_EXPR_LIST (REG_NONNEG, NULL_RTX,
8293 /* No insn may reference both the reversed and another biv or it
8294 will fail (see comment near the top of the loop reversal
8296 Earlier on, we have verified that the biv has no use except
8297 counting, or it is the only biv in this function.
8298 However, the code that computes no_use_except_counting does
8299 not verify reg notes. It's possible to have an insn that
8300 references another biv, and has a REG_EQUAL note with an
8301 expression based on the reversed biv. To avoid this case,
8302 remove all REG_EQUAL notes based on the reversed biv
8304 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
8305 if (GET_RTX_CLASS (GET_CODE (p)) == 'i')
8308 rtx set = single_set (p);
8309 /* If this is a set of a GIV based on the reversed biv, any
8310 REG_EQUAL notes should still be correct. */
8312 || GET_CODE (SET_DEST (set)) != REG
8313 || (size_t) REGNO (SET_DEST (set)) >= reg_iv_type->num_elements
8314 || REG_IV_TYPE (REGNO (SET_DEST (set))) != GENERAL_INDUCT
8315 || REG_IV_INFO (REGNO (SET_DEST (set)))->src_reg != bl->biv->src_reg)
8316 for (pnote = ®_NOTES (p); *pnote;)
8318 if (REG_NOTE_KIND (*pnote) == REG_EQUAL
8319 && reg_mentioned_p (regno_reg_rtx[bl->regno],
8321 *pnote = XEXP (*pnote, 1);
8323 pnote = &XEXP (*pnote, 1);
8327 /* Mark that this biv has been reversed. Each giv which depends
8328 on this biv, and which is also live past the end of the loop
8329 will have to be fixed up. */
8333 if (loop_dump_stream)
8335 fprintf (loop_dump_stream, "Reversed loop");
8337 fprintf (loop_dump_stream, " and added reg_nonneg\n");
8339 fprintf (loop_dump_stream, "\n");
8350 /* Verify whether the biv BL appears to be eliminable,
8351 based on the insns in the loop that refer to it.
8353 If ELIMINATE_P is non-zero, actually do the elimination.
8355 THRESHOLD and INSN_COUNT are from loop_optimize and are used to
8356 determine whether invariant insns should be placed inside or at the
8357 start of the loop. */
8360 maybe_eliminate_biv (loop, bl, eliminate_p, threshold, insn_count)
8361 const struct loop *loop;
8362 struct iv_class *bl;
8364 int threshold, insn_count;
8366 rtx reg = bl->biv->dest_reg;
8367 rtx loop_start = loop->start;
8368 rtx loop_end = loop->end;
8371 /* Scan all insns in the loop, stopping if we find one that uses the
8372 biv in a way that we cannot eliminate. */
8374 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
8376 enum rtx_code code = GET_CODE (p);
8377 rtx where = threshold >= insn_count ? loop_start : p;
8379 /* If this is a libcall that sets a giv, skip ahead to its end. */
8380 if (GET_RTX_CLASS (code) == 'i')
8382 rtx note = find_reg_note (p, REG_LIBCALL, NULL_RTX);
8386 rtx last = XEXP (note, 0);
8387 rtx set = single_set (last);
8389 if (set && GET_CODE (SET_DEST (set)) == REG)
8391 unsigned int regno = REGNO (SET_DEST (set));
8393 if (regno < max_reg_before_loop
8394 && REG_IV_TYPE (regno) == GENERAL_INDUCT
8395 && REG_IV_INFO (regno)->src_reg == bl->biv->src_reg)
8400 if ((code == INSN || code == JUMP_INSN || code == CALL_INSN)
8401 && reg_mentioned_p (reg, PATTERN (p))
8402 && ! maybe_eliminate_biv_1 (loop, PATTERN (p), p, bl,
8403 eliminate_p, where))
8405 if (loop_dump_stream)
8406 fprintf (loop_dump_stream,
8407 "Cannot eliminate biv %d: biv used in insn %d.\n",
8408 bl->regno, INSN_UID (p));
8415 if (loop_dump_stream)
8416 fprintf (loop_dump_stream, "biv %d %s eliminated.\n",
8417 bl->regno, eliminate_p ? "was" : "can be");
8424 /* INSN and REFERENCE are instructions in the same insn chain.
8425 Return non-zero if INSN is first. */
8428 loop_insn_first_p (insn, reference)
8429 rtx insn, reference;
8433 for (p = insn, q = reference; ;)
8435 /* Start with test for not first so that INSN == REFERENCE yields not
8437 if (q == insn || ! p)
8439 if (p == reference || ! q)
8442 /* Either of P or Q might be a NOTE. Notes have the same LUID as the
8443 previous insn, hence the <= comparison below does not work if
8445 if (INSN_UID (p) < max_uid_for_loop
8446 && INSN_UID (q) < max_uid_for_loop
8447 && GET_CODE (p) != NOTE)
8448 return INSN_LUID (p) <= INSN_LUID (q);
8450 if (INSN_UID (p) >= max_uid_for_loop
8451 || GET_CODE (p) == NOTE)
8453 if (INSN_UID (q) >= max_uid_for_loop)
8458 /* We are trying to eliminate BIV in INSN using GIV. Return non-zero if
8459 the offset that we have to take into account due to auto-increment /
8460 div derivation is zero. */
8462 biv_elimination_giv_has_0_offset (biv, giv, insn)
8463 struct induction *biv, *giv;
8466 /* If the giv V had the auto-inc address optimization applied
8467 to it, and INSN occurs between the giv insn and the biv
8468 insn, then we'd have to adjust the value used here.
8469 This is rare, so we don't bother to make this possible. */
8470 if (giv->auto_inc_opt
8471 && ((loop_insn_first_p (giv->insn, insn)
8472 && loop_insn_first_p (insn, biv->insn))
8473 || (loop_insn_first_p (biv->insn, insn)
8474 && loop_insn_first_p (insn, giv->insn))))
8477 /* If the giv V was derived from another giv, and INSN does
8478 not occur between the giv insn and the biv insn, then we'd
8479 have to adjust the value used here. This is rare, so we don't
8480 bother to make this possible. */
8481 if (giv->derived_from
8482 && ! (giv->always_executed
8483 && loop_insn_first_p (giv->insn, insn)
8484 && loop_insn_first_p (insn, biv->insn)))
8487 && giv->same->derived_from
8488 && ! (giv->same->always_executed
8489 && loop_insn_first_p (giv->same->insn, insn)
8490 && loop_insn_first_p (insn, biv->insn)))
8496 /* If BL appears in X (part of the pattern of INSN), see if we can
8497 eliminate its use. If so, return 1. If not, return 0.
8499 If BIV does not appear in X, return 1.
8501 If ELIMINATE_P is non-zero, actually do the elimination. WHERE indicates
8502 where extra insns should be added. Depending on how many items have been
8503 moved out of the loop, it will either be before INSN or at the start of
8507 maybe_eliminate_biv_1 (loop, x, insn, bl, eliminate_p, where)
8508 const struct loop *loop;
8510 struct iv_class *bl;
8514 enum rtx_code code = GET_CODE (x);
8515 rtx reg = bl->biv->dest_reg;
8516 enum machine_mode mode = GET_MODE (reg);
8517 struct induction *v;
8529 /* If we haven't already been able to do something with this BIV,
8530 we can't eliminate it. */
8536 /* If this sets the BIV, it is not a problem. */
8537 if (SET_DEST (x) == reg)
8540 /* If this is an insn that defines a giv, it is also ok because
8541 it will go away when the giv is reduced. */
8542 for (v = bl->giv; v; v = v->next_iv)
8543 if (v->giv_type == DEST_REG && SET_DEST (x) == v->dest_reg)
8547 if (SET_DEST (x) == cc0_rtx && SET_SRC (x) == reg)
8549 /* Can replace with any giv that was reduced and
8550 that has (MULT_VAL != 0) and (ADD_VAL == 0).
8551 Require a constant for MULT_VAL, so we know it's nonzero.
8552 ??? We disable this optimization to avoid potential
8555 for (v = bl->giv; v; v = v->next_iv)
8556 if (GET_CODE (v->mult_val) == CONST_INT && v->mult_val != const0_rtx
8557 && v->add_val == const0_rtx
8558 && ! v->ignore && ! v->maybe_dead && v->always_computable
8562 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8568 /* If the giv has the opposite direction of change,
8569 then reverse the comparison. */
8570 if (INTVAL (v->mult_val) < 0)
8571 new = gen_rtx_COMPARE (GET_MODE (v->new_reg),
8572 const0_rtx, v->new_reg);
8576 /* We can probably test that giv's reduced reg. */
8577 if (validate_change (insn, &SET_SRC (x), new, 0))
8581 /* Look for a giv with (MULT_VAL != 0) and (ADD_VAL != 0);
8582 replace test insn with a compare insn (cmp REDUCED_GIV ADD_VAL).
8583 Require a constant for MULT_VAL, so we know it's nonzero.
8584 ??? Do this only if ADD_VAL is a pointer to avoid a potential
8585 overflow problem. */
8587 for (v = bl->giv; v; v = v->next_iv)
8588 if (GET_CODE (v->mult_val) == CONST_INT && v->mult_val != const0_rtx
8589 && ! v->ignore && ! v->maybe_dead && v->always_computable
8591 && (GET_CODE (v->add_val) == SYMBOL_REF
8592 || GET_CODE (v->add_val) == LABEL_REF
8593 || GET_CODE (v->add_val) == CONST
8594 || (GET_CODE (v->add_val) == REG
8595 && REGNO_POINTER_FLAG (REGNO (v->add_val)))))
8597 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8603 /* If the giv has the opposite direction of change,
8604 then reverse the comparison. */
8605 if (INTVAL (v->mult_val) < 0)
8606 new = gen_rtx_COMPARE (VOIDmode, copy_rtx (v->add_val),
8609 new = gen_rtx_COMPARE (VOIDmode, v->new_reg,
8610 copy_rtx (v->add_val));
8612 /* Replace biv with the giv's reduced register. */
8613 update_reg_last_use (v->add_val, insn);
8614 if (validate_change (insn, &SET_SRC (PATTERN (insn)), new, 0))
8617 /* Insn doesn't support that constant or invariant. Copy it
8618 into a register (it will be a loop invariant.) */
8619 tem = gen_reg_rtx (GET_MODE (v->new_reg));
8621 emit_insn_before (gen_move_insn (tem, copy_rtx (v->add_val)),
8624 /* Substitute the new register for its invariant value in
8625 the compare expression. */
8626 XEXP (new, (INTVAL (v->mult_val) < 0) ? 0 : 1) = tem;
8627 if (validate_change (insn, &SET_SRC (PATTERN (insn)), new, 0))
8636 case GT: case GE: case GTU: case GEU:
8637 case LT: case LE: case LTU: case LEU:
8638 /* See if either argument is the biv. */
8639 if (XEXP (x, 0) == reg)
8640 arg = XEXP (x, 1), arg_operand = 1;
8641 else if (XEXP (x, 1) == reg)
8642 arg = XEXP (x, 0), arg_operand = 0;
8646 if (CONSTANT_P (arg))
8648 /* First try to replace with any giv that has constant positive
8649 mult_val and constant add_val. We might be able to support
8650 negative mult_val, but it seems complex to do it in general. */
8652 for (v = bl->giv; v; v = v->next_iv)
8653 if (GET_CODE (v->mult_val) == CONST_INT && INTVAL (v->mult_val) > 0
8654 && (GET_CODE (v->add_val) == SYMBOL_REF
8655 || GET_CODE (v->add_val) == LABEL_REF
8656 || GET_CODE (v->add_val) == CONST
8657 || (GET_CODE (v->add_val) == REG
8658 && REGNO_POINTER_FLAG (REGNO (v->add_val))))
8659 && ! v->ignore && ! v->maybe_dead && v->always_computable
8662 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8668 /* Replace biv with the giv's reduced reg. */
8669 validate_change (insn, &XEXP (x, 1-arg_operand), v->new_reg, 1);
8671 /* If all constants are actually constant integers and
8672 the derived constant can be directly placed in the COMPARE,
8674 if (GET_CODE (arg) == CONST_INT
8675 && GET_CODE (v->mult_val) == CONST_INT
8676 && GET_CODE (v->add_val) == CONST_INT)
8678 validate_change (insn, &XEXP (x, arg_operand),
8679 GEN_INT (INTVAL (arg)
8680 * INTVAL (v->mult_val)
8681 + INTVAL (v->add_val)), 1);
8685 /* Otherwise, load it into a register. */
8686 tem = gen_reg_rtx (mode);
8687 emit_iv_add_mult (arg, v->mult_val, v->add_val, tem, where);
8688 validate_change (insn, &XEXP (x, arg_operand), tem, 1);
8690 if (apply_change_group ())
8694 /* Look for giv with positive constant mult_val and nonconst add_val.
8695 Insert insns to calculate new compare value.
8696 ??? Turn this off due to possible overflow. */
8698 for (v = bl->giv; v; v = v->next_iv)
8699 if (GET_CODE (v->mult_val) == CONST_INT && INTVAL (v->mult_val) > 0
8700 && ! v->ignore && ! v->maybe_dead && v->always_computable
8706 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8712 tem = gen_reg_rtx (mode);
8714 /* Replace biv with giv's reduced register. */
8715 validate_change (insn, &XEXP (x, 1 - arg_operand),
8718 /* Compute value to compare against. */
8719 emit_iv_add_mult (arg, v->mult_val, v->add_val, tem, where);
8720 /* Use it in this insn. */
8721 validate_change (insn, &XEXP (x, arg_operand), tem, 1);
8722 if (apply_change_group ())
8726 else if (GET_CODE (arg) == REG || GET_CODE (arg) == MEM)
8728 if (loop_invariant_p (loop, arg) == 1)
8730 /* Look for giv with constant positive mult_val and nonconst
8731 add_val. Insert insns to compute new compare value.
8732 ??? Turn this off due to possible overflow. */
8734 for (v = bl->giv; v; v = v->next_iv)
8735 if (GET_CODE (v->mult_val) == CONST_INT && INTVAL (v->mult_val) > 0
8736 && ! v->ignore && ! v->maybe_dead && v->always_computable
8742 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8748 tem = gen_reg_rtx (mode);
8750 /* Replace biv with giv's reduced register. */
8751 validate_change (insn, &XEXP (x, 1 - arg_operand),
8754 /* Compute value to compare against. */
8755 emit_iv_add_mult (arg, v->mult_val, v->add_val,
8757 validate_change (insn, &XEXP (x, arg_operand), tem, 1);
8758 if (apply_change_group ())
8763 /* This code has problems. Basically, you can't know when
8764 seeing if we will eliminate BL, whether a particular giv
8765 of ARG will be reduced. If it isn't going to be reduced,
8766 we can't eliminate BL. We can try forcing it to be reduced,
8767 but that can generate poor code.
8769 The problem is that the benefit of reducing TV, below should
8770 be increased if BL can actually be eliminated, but this means
8771 we might have to do a topological sort of the order in which
8772 we try to process biv. It doesn't seem worthwhile to do
8773 this sort of thing now. */
8776 /* Otherwise the reg compared with had better be a biv. */
8777 if (GET_CODE (arg) != REG
8778 || REG_IV_TYPE (REGNO (arg)) != BASIC_INDUCT)
8781 /* Look for a pair of givs, one for each biv,
8782 with identical coefficients. */
8783 for (v = bl->giv; v; v = v->next_iv)
8785 struct induction *tv;
8787 if (v->ignore || v->maybe_dead || v->mode != mode)
8790 for (tv = reg_biv_class[REGNO (arg)]->giv; tv; tv = tv->next_iv)
8791 if (! tv->ignore && ! tv->maybe_dead
8792 && rtx_equal_p (tv->mult_val, v->mult_val)
8793 && rtx_equal_p (tv->add_val, v->add_val)
8794 && tv->mode == mode)
8796 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8802 /* Replace biv with its giv's reduced reg. */
8803 XEXP (x, 1-arg_operand) = v->new_reg;
8804 /* Replace other operand with the other giv's
8806 XEXP (x, arg_operand) = tv->new_reg;
8813 /* If we get here, the biv can't be eliminated. */
8817 /* If this address is a DEST_ADDR giv, it doesn't matter if the
8818 biv is used in it, since it will be replaced. */
8819 for (v = bl->giv; v; v = v->next_iv)
8820 if (v->giv_type == DEST_ADDR && v->location == &XEXP (x, 0))
8828 /* See if any subexpression fails elimination. */
8829 fmt = GET_RTX_FORMAT (code);
8830 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
8835 if (! maybe_eliminate_biv_1 (loop, XEXP (x, i), insn, bl,
8836 eliminate_p, where))
8841 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
8842 if (! maybe_eliminate_biv_1 (loop, XVECEXP (x, i, j), insn, bl,
8843 eliminate_p, where))
8852 /* Return nonzero if the last use of REG
8853 is in an insn following INSN in the same basic block. */
8856 last_use_this_basic_block (reg, insn)
8862 n && GET_CODE (n) != CODE_LABEL && GET_CODE (n) != JUMP_INSN;
8865 if (REGNO_LAST_UID (REGNO (reg)) == INSN_UID (n))
8871 /* Called via `note_stores' to record the initial value of a biv. Here we
8872 just record the location of the set and process it later. */
8875 record_initial (dest, set, data)
8878 void *data ATTRIBUTE_UNUSED;
8880 struct iv_class *bl;
8882 if (GET_CODE (dest) != REG
8883 || REGNO (dest) >= max_reg_before_loop
8884 || REG_IV_TYPE (REGNO (dest)) != BASIC_INDUCT)
8887 bl = reg_biv_class[REGNO (dest)];
8889 /* If this is the first set found, record it. */
8890 if (bl->init_insn == 0)
8892 bl->init_insn = note_insn;
8897 /* If any of the registers in X are "old" and currently have a last use earlier
8898 than INSN, update them to have a last use of INSN. Their actual last use
8899 will be the previous insn but it will not have a valid uid_luid so we can't
8903 update_reg_last_use (x, insn)
8907 /* Check for the case where INSN does not have a valid luid. In this case,
8908 there is no need to modify the regno_last_uid, as this can only happen
8909 when code is inserted after the loop_end to set a pseudo's final value,
8910 and hence this insn will never be the last use of x. */
8911 if (GET_CODE (x) == REG && REGNO (x) < max_reg_before_loop
8912 && INSN_UID (insn) < max_uid_for_loop
8913 && uid_luid[REGNO_LAST_UID (REGNO (x))] < uid_luid[INSN_UID (insn)])
8914 REGNO_LAST_UID (REGNO (x)) = INSN_UID (insn);
8918 register const char *fmt = GET_RTX_FORMAT (GET_CODE (x));
8919 for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
8922 update_reg_last_use (XEXP (x, i), insn);
8923 else if (fmt[i] == 'E')
8924 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
8925 update_reg_last_use (XVECEXP (x, i, j), insn);
8930 /* Given an insn INSN and condition COND, return the condition in a
8931 canonical form to simplify testing by callers. Specifically:
8933 (1) The code will always be a comparison operation (EQ, NE, GT, etc.).
8934 (2) Both operands will be machine operands; (cc0) will have been replaced.
8935 (3) If an operand is a constant, it will be the second operand.
8936 (4) (LE x const) will be replaced with (LT x <const+1>) and similarly
8937 for GE, GEU, and LEU.
8939 If the condition cannot be understood, or is an inequality floating-point
8940 comparison which needs to be reversed, 0 will be returned.
8942 If REVERSE is non-zero, then reverse the condition prior to canonizing it.
8944 If EARLIEST is non-zero, it is a pointer to a place where the earliest
8945 insn used in locating the condition was found. If a replacement test
8946 of the condition is desired, it should be placed in front of that
8947 insn and we will be sure that the inputs are still valid.
8949 If WANT_REG is non-zero, we wish the condition to be relative to that
8950 register, if possible. Therefore, do not canonicalize the condition
8954 canonicalize_condition (insn, cond, reverse, earliest, want_reg)
8966 int reverse_code = 0;
8967 int did_reverse_condition = 0;
8968 enum machine_mode mode;
8970 code = GET_CODE (cond);
8971 mode = GET_MODE (cond);
8972 op0 = XEXP (cond, 0);
8973 op1 = XEXP (cond, 1);
8977 code = reverse_condition (code);
8978 did_reverse_condition ^= 1;
8984 /* If we are comparing a register with zero, see if the register is set
8985 in the previous insn to a COMPARE or a comparison operation. Perform
8986 the same tests as a function of STORE_FLAG_VALUE as find_comparison_args
8989 while (GET_RTX_CLASS (code) == '<'
8990 && op1 == CONST0_RTX (GET_MODE (op0))
8993 /* Set non-zero when we find something of interest. */
8997 /* If comparison with cc0, import actual comparison from compare
9001 if ((prev = prev_nonnote_insn (prev)) == 0
9002 || GET_CODE (prev) != INSN
9003 || (set = single_set (prev)) == 0
9004 || SET_DEST (set) != cc0_rtx)
9007 op0 = SET_SRC (set);
9008 op1 = CONST0_RTX (GET_MODE (op0));
9014 /* If this is a COMPARE, pick up the two things being compared. */
9015 if (GET_CODE (op0) == COMPARE)
9017 op1 = XEXP (op0, 1);
9018 op0 = XEXP (op0, 0);
9021 else if (GET_CODE (op0) != REG)
9024 /* Go back to the previous insn. Stop if it is not an INSN. We also
9025 stop if it isn't a single set or if it has a REG_INC note because
9026 we don't want to bother dealing with it. */
9028 if ((prev = prev_nonnote_insn (prev)) == 0
9029 || GET_CODE (prev) != INSN
9030 || FIND_REG_INC_NOTE (prev, 0)
9031 || (set = single_set (prev)) == 0)
9034 /* If this is setting OP0, get what it sets it to if it looks
9036 if (rtx_equal_p (SET_DEST (set), op0))
9038 enum machine_mode inner_mode = GET_MODE (SET_SRC (set));
9040 /* ??? We may not combine comparisons done in a CCmode with
9041 comparisons not done in a CCmode. This is to aid targets
9042 like Alpha that have an IEEE compliant EQ instruction, and
9043 a non-IEEE compliant BEQ instruction. The use of CCmode is
9044 actually artificial, simply to prevent the combination, but
9045 should not affect other platforms.
9047 However, we must allow VOIDmode comparisons to match either
9048 CCmode or non-CCmode comparison, because some ports have
9049 modeless comparisons inside branch patterns.
9051 ??? This mode check should perhaps look more like the mode check
9052 in simplify_comparison in combine. */
9054 if ((GET_CODE (SET_SRC (set)) == COMPARE
9057 && GET_MODE_CLASS (inner_mode) == MODE_INT
9058 && (GET_MODE_BITSIZE (inner_mode)
9059 <= HOST_BITS_PER_WIDE_INT)
9060 && (STORE_FLAG_VALUE
9061 & ((HOST_WIDE_INT) 1
9062 << (GET_MODE_BITSIZE (inner_mode) - 1))))
9063 #ifdef FLOAT_STORE_FLAG_VALUE
9065 && GET_MODE_CLASS (inner_mode) == MODE_FLOAT
9066 && (REAL_VALUE_NEGATIVE
9067 (FLOAT_STORE_FLAG_VALUE (inner_mode))))
9070 && GET_RTX_CLASS (GET_CODE (SET_SRC (set))) == '<'))
9071 && (((GET_MODE_CLASS (mode) == MODE_CC)
9072 == (GET_MODE_CLASS (inner_mode) == MODE_CC))
9073 || mode == VOIDmode || inner_mode == VOIDmode))
9075 else if (((code == EQ
9077 && (GET_MODE_BITSIZE (inner_mode)
9078 <= HOST_BITS_PER_WIDE_INT)
9079 && GET_MODE_CLASS (inner_mode) == MODE_INT
9080 && (STORE_FLAG_VALUE
9081 & ((HOST_WIDE_INT) 1
9082 << (GET_MODE_BITSIZE (inner_mode) - 1))))
9083 #ifdef FLOAT_STORE_FLAG_VALUE
9085 && GET_MODE_CLASS (inner_mode) == MODE_FLOAT
9086 && (REAL_VALUE_NEGATIVE
9087 (FLOAT_STORE_FLAG_VALUE (inner_mode))))
9090 && GET_RTX_CLASS (GET_CODE (SET_SRC (set))) == '<'
9091 && (((GET_MODE_CLASS (mode) == MODE_CC)
9092 == (GET_MODE_CLASS (inner_mode) == MODE_CC))
9093 || mode == VOIDmode || inner_mode == VOIDmode))
9096 /* We might have reversed a LT to get a GE here. But this wasn't
9097 actually the comparison of data, so we don't flag that we
9098 have had to reverse the condition. */
9099 did_reverse_condition ^= 1;
9107 else if (reg_set_p (op0, prev))
9108 /* If this sets OP0, but not directly, we have to give up. */
9113 if (GET_RTX_CLASS (GET_CODE (x)) == '<')
9114 code = GET_CODE (x);
9117 code = reverse_condition (code);
9118 if (code == UNKNOWN)
9120 did_reverse_condition ^= 1;
9124 op0 = XEXP (x, 0), op1 = XEXP (x, 1);
9130 /* If constant is first, put it last. */
9131 if (CONSTANT_P (op0))
9132 code = swap_condition (code), tem = op0, op0 = op1, op1 = tem;
9134 /* If OP0 is the result of a comparison, we weren't able to find what
9135 was really being compared, so fail. */
9136 if (GET_MODE_CLASS (GET_MODE (op0)) == MODE_CC)
9139 /* Canonicalize any ordered comparison with integers involving equality
9140 if we can do computations in the relevant mode and we do not
9143 if (GET_CODE (op1) == CONST_INT
9144 && GET_MODE (op0) != VOIDmode
9145 && GET_MODE_BITSIZE (GET_MODE (op0)) <= HOST_BITS_PER_WIDE_INT)
9147 HOST_WIDE_INT const_val = INTVAL (op1);
9148 unsigned HOST_WIDE_INT uconst_val = const_val;
9149 unsigned HOST_WIDE_INT max_val
9150 = (unsigned HOST_WIDE_INT) GET_MODE_MASK (GET_MODE (op0));
9155 if ((unsigned HOST_WIDE_INT) const_val != max_val >> 1)
9156 code = LT, op1 = GEN_INT (const_val + 1);
9159 /* When cross-compiling, const_val might be sign-extended from
9160 BITS_PER_WORD to HOST_BITS_PER_WIDE_INT */
9162 if ((HOST_WIDE_INT) (const_val & max_val)
9163 != (((HOST_WIDE_INT) 1
9164 << (GET_MODE_BITSIZE (GET_MODE (op0)) - 1))))
9165 code = GT, op1 = GEN_INT (const_val - 1);
9169 if (uconst_val < max_val)
9170 code = LTU, op1 = GEN_INT (uconst_val + 1);
9174 if (uconst_val != 0)
9175 code = GTU, op1 = GEN_INT (uconst_val - 1);
9183 /* If this was floating-point and we reversed anything other than an
9184 EQ or NE or (UN)ORDERED, return zero. */
9185 if (TARGET_FLOAT_FORMAT == IEEE_FLOAT_FORMAT
9186 && did_reverse_condition
9187 && code != NE && code != EQ && code != UNORDERED && code != ORDERED
9189 && GET_MODE_CLASS (GET_MODE (op0)) == MODE_FLOAT)
9193 /* Never return CC0; return zero instead. */
9198 return gen_rtx_fmt_ee (code, VOIDmode, op0, op1);
9201 /* Given a jump insn JUMP, return the condition that will cause it to branch
9202 to its JUMP_LABEL. If the condition cannot be understood, or is an
9203 inequality floating-point comparison which needs to be reversed, 0 will
9206 If EARLIEST is non-zero, it is a pointer to a place where the earliest
9207 insn used in locating the condition was found. If a replacement test
9208 of the condition is desired, it should be placed in front of that
9209 insn and we will be sure that the inputs are still valid. */
9212 get_condition (jump, earliest)
9219 /* If this is not a standard conditional jump, we can't parse it. */
9220 if (GET_CODE (jump) != JUMP_INSN
9221 || ! condjump_p (jump) || simplejump_p (jump))
9224 cond = XEXP (SET_SRC (PATTERN (jump)), 0);
9226 /* If this branches to JUMP_LABEL when the condition is false, reverse
9229 = GET_CODE (XEXP (SET_SRC (PATTERN (jump)), 2)) == LABEL_REF
9230 && XEXP (XEXP (SET_SRC (PATTERN (jump)), 2), 0) == JUMP_LABEL (jump);
9232 return canonicalize_condition (jump, cond, reverse, earliest, NULL_RTX);
9235 /* Similar to above routine, except that we also put an invariant last
9236 unless both operands are invariants. */
9239 get_condition_for_loop (loop, x)
9240 const struct loop *loop;
9243 rtx comparison = get_condition (x, NULL_PTR);
9246 || ! loop_invariant_p (loop, XEXP (comparison, 0))
9247 || loop_invariant_p (loop, XEXP (comparison, 1)))
9250 return gen_rtx_fmt_ee (swap_condition (GET_CODE (comparison)), VOIDmode,
9251 XEXP (comparison, 1), XEXP (comparison, 0));
9254 #ifdef HAVE_decrement_and_branch_on_count
9255 /* Instrument loop for insertion of bct instruction. We distinguish between
9256 loops with compile-time bounds and those with run-time bounds.
9257 Information from loop_iterations() is used to compute compile-time bounds.
9258 Run-time bounds should use loop preconditioning, but currently ignored.
9265 unsigned HOST_WIDE_INT n_iterations;
9266 rtx loop_start = loop->start;
9267 rtx loop_end = loop->end;
9268 struct loop_info *loop_info = LOOP_INFO (loop);
9269 int loop_num = loop->num;
9272 int increment_direction, compare_direction;
9273 /* If the loop condition is <= or >=, the number of iteration
9274 is 1 more than the range of the bounds of the loop. */
9275 int add_iteration = 0;
9276 enum machine_mode loop_var_mode = word_mode;
9279 /* It's impossible to instrument a competely unrolled loop. */
9280 if (loop_info->unroll_number == loop_info->n_iterations)
9283 /* Make sure that the count register is not in use. */
9284 if (loop_info->used_count_register)
9286 if (loop_dump_stream)
9287 fprintf (loop_dump_stream,
9288 "insert_bct %d: BCT instrumentation failed: count register already in use\n",
9293 /* Make sure that the function has no indirect jumps. */
9294 if (indirect_jump_in_function)
9296 if (loop_dump_stream)
9297 fprintf (loop_dump_stream,
9298 "insert_bct %d: BCT instrumentation failed: indirect jump in function\n",
9303 /* Make sure that the last loop insn is a conditional jump. */
9304 if (GET_CODE (PREV_INSN (loop_end)) != JUMP_INSN
9305 || ! condjump_p (PREV_INSN (loop_end))
9306 || simplejump_p (PREV_INSN (loop_end)))
9308 if (loop_dump_stream)
9309 fprintf (loop_dump_stream,
9310 "insert_bct %d: BCT instrumentation failed: invalid jump at loop end\n",
9315 /* Make sure that the loop does not contain a function call
9316 (the count register might be altered by the called function). */
9317 if (loop_info->has_call)
9319 if (loop_dump_stream)
9320 fprintf (loop_dump_stream,
9321 "insert_bct %d: BCT instrumentation failed: function call in loop\n",
9326 /* Make sure that the loop does not jump via a table.
9327 (the count register might be used to perform the branch on table). */
9328 if (loop_info->has_tablejump)
9330 if (loop_dump_stream)
9331 fprintf (loop_dump_stream,
9332 "insert_bct %d: BCT instrumentation failed: computed branch in the loop\n",
9337 /* Account for loop unrolling in instrumented iteration count. */
9338 if (loop_info->unroll_number > 1)
9339 n_iterations = loop_info->n_iterations / loop_info->unroll_number;
9341 n_iterations = loop_info->n_iterations;
9343 if (n_iterations != 0 && n_iterations < 3)
9345 /* Allow an enclosing outer loop to benefit if possible. */
9346 if (loop_dump_stream)
9347 fprintf (loop_dump_stream,
9348 "insert_bct %d: Too few iterations to benefit from BCT optimization\n",
9353 /* Try to instrument the loop. */
9355 /* Handle the simpler case, where the bounds are known at compile time. */
9356 if (n_iterations > 0)
9358 struct loop *outer_loop;
9359 struct loop_info *outer_loop_info;
9361 /* Mark all enclosing loops that they cannot use count register. */
9362 for (outer_loop = loop; outer_loop; outer_loop = outer_loop->outer)
9364 outer_loop_info = LOOP_INFO (outer_loop);
9365 outer_loop_info->used_count_register = 1;
9367 instrument_loop_bct (loop_start, loop_end, GEN_INT (n_iterations));
9371 /* Handle the more complex case, that the bounds are NOT known
9372 at compile time. In this case we generate run_time calculation
9373 of the number of iterations. */
9375 if (loop_info->iteration_var == 0)
9377 if (loop_dump_stream)
9378 fprintf (loop_dump_stream,
9379 "insert_bct %d: BCT Runtime Instrumentation failed: no loop iteration variable found\n",
9384 if (GET_MODE_CLASS (GET_MODE (loop_info->iteration_var)) != MODE_INT
9385 || GET_MODE_SIZE (GET_MODE (loop_info->iteration_var)) != UNITS_PER_WORD)
9387 if (loop_dump_stream)
9388 fprintf (loop_dump_stream,
9389 "insert_bct %d: BCT Runtime Instrumentation failed: loop variable not integer\n",
9394 /* With runtime bounds, if the compare is of the form '!=' we give up */
9395 if (loop_info->comparison_code == NE)
9397 if (loop_dump_stream)
9398 fprintf (loop_dump_stream,
9399 "insert_bct %d: BCT Runtime Instrumentation failed: runtime bounds with != comparison\n",
9403 /* Use common loop preconditioning code instead. */
9407 /* We rely on the existence of run-time guard to ensure that the
9408 loop executes at least once. */
9410 rtx iterations_num_reg;
9412 unsigned HOST_WIDE_INT increment_value_abs
9413 = INTVAL (increment) * increment_direction;
9415 /* make sure that the increment is a power of two, otherwise (an
9416 expensive) divide is needed. */
9417 if (exact_log2 (increment_value_abs) == -1)
9419 if (loop_dump_stream)
9420 fprintf (loop_dump_stream,
9421 "insert_bct: not instrumenting BCT because the increment is not power of 2\n");
9425 /* compute the number of iterations */
9430 /* Again, the number of iterations is calculated by:
9432 ; compare-val - initial-val + (increment -1) + additional-iteration
9433 ; num_iterations = -----------------------------------------------------------------
9436 /* ??? Do we have to call copy_rtx here before passing rtx to
9438 if (compare_direction > 0)
9440 /* <, <= :the loop variable is increasing */
9441 temp_reg = expand_binop (loop_var_mode, sub_optab,
9442 comparison_value, initial_value,
9443 NULL_RTX, 0, OPTAB_LIB_WIDEN);
9447 temp_reg = expand_binop (loop_var_mode, sub_optab,
9448 initial_value, comparison_value,
9449 NULL_RTX, 0, OPTAB_LIB_WIDEN);
9452 if (increment_value_abs - 1 + add_iteration != 0)
9453 temp_reg = expand_binop (loop_var_mode, add_optab, temp_reg,
9454 GEN_INT (increment_value_abs - 1
9456 NULL_RTX, 0, OPTAB_LIB_WIDEN);
9458 if (increment_value_abs != 1)
9459 iterations_num_reg = expand_binop (loop_var_mode, asr_optab,
9461 GEN_INT (exact_log2 (increment_value_abs)),
9462 NULL_RTX, 0, OPTAB_LIB_WIDEN);
9464 iterations_num_reg = temp_reg;
9466 sequence = gen_sequence ();
9468 emit_insn_before (sequence, loop_start);
9469 instrument_loop_bct (loop_start, loop_end, iterations_num_reg);
9473 #endif /* Complex case */
9476 /* Instrument loop by inserting a bct in it as follows:
9477 1. A new counter register is created.
9478 2. In the head of the loop the new variable is initialized to the value
9479 passed in the loop_num_iterations parameter.
9480 3. At the end of the loop, comparison of the register with 0 is generated.
9481 The created comparison follows the pattern defined for the
9482 decrement_and_branch_on_count insn, so this insn will be generated.
9483 4. The branch on the old variable are deleted. The compare must remain
9484 because it might be used elsewhere. If the loop-variable or condition
9485 register are used elsewhere, they will be eliminated by flow. */
9488 instrument_loop_bct (loop_start, loop_end, loop_num_iterations)
9489 rtx loop_start, loop_end;
9490 rtx loop_num_iterations;
9496 if (HAVE_decrement_and_branch_on_count)
9498 if (loop_dump_stream)
9500 fputs ("instrument_bct: Inserting BCT (", loop_dump_stream);
9501 if (GET_CODE (loop_num_iterations) == CONST_INT)
9502 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC,
9503 INTVAL (loop_num_iterations));
9505 fputs ("runtime", loop_dump_stream);
9506 fputs (" iterations)", loop_dump_stream);
9509 /* Discard original jump to continue loop. Original compare result
9510 may still be live, so it cannot be discarded explicitly. */
9511 delete_insn (PREV_INSN (loop_end));
9513 /* Insert the label which will delimit the start of the loop. */
9514 start_label = gen_label_rtx ();
9515 emit_label_after (start_label, loop_start);
9517 /* Insert initialization of the count register into the loop header. */
9519 counter_reg = gen_reg_rtx (word_mode);
9520 emit_insn (gen_move_insn (counter_reg, loop_num_iterations));
9521 sequence = gen_sequence ();
9523 emit_insn_before (sequence, loop_start);
9525 /* Insert new comparison on the count register instead of the
9526 old one, generating the needed BCT pattern (that will be
9527 later recognized by assembly generation phase). */
9528 emit_jump_insn_before (gen_decrement_and_branch_on_count (counter_reg,
9531 LABEL_NUSES (start_label)++;
9535 #endif /* HAVE_decrement_and_branch_on_count */
9537 /* Scan the function and determine whether it has indirect (computed) jumps.
9539 This is taken mostly from flow.c; similar code exists elsewhere
9540 in the compiler. It may be useful to put this into rtlanal.c. */
9542 indirect_jump_in_function_p (start)
9547 for (insn = start; insn; insn = NEXT_INSN (insn))
9548 if (computed_jump_p (insn))
9554 /* Add MEM to the LOOP_MEMS array, if appropriate. See the
9555 documentation for LOOP_MEMS for the definition of `appropriate'.
9556 This function is called from prescan_loop via for_each_rtx. */
9559 insert_loop_mem (mem, data)
9561 void *data ATTRIBUTE_UNUSED;
9569 switch (GET_CODE (m))
9575 /* We're not interested in MEMs that are only clobbered. */
9579 /* We're not interested in the MEM associated with a
9580 CONST_DOUBLE, so there's no need to traverse into this. */
9584 /* We're not interested in any MEMs that only appear in notes. */
9588 /* This is not a MEM. */
9592 /* See if we've already seen this MEM. */
9593 for (i = 0; i < loop_mems_idx; ++i)
9594 if (rtx_equal_p (m, loop_mems[i].mem))
9596 if (GET_MODE (m) != GET_MODE (loop_mems[i].mem))
9597 /* The modes of the two memory accesses are different. If
9598 this happens, something tricky is going on, and we just
9599 don't optimize accesses to this MEM. */
9600 loop_mems[i].optimize = 0;
9605 /* Resize the array, if necessary. */
9606 if (loop_mems_idx == loop_mems_allocated)
9608 if (loop_mems_allocated != 0)
9609 loop_mems_allocated *= 2;
9611 loop_mems_allocated = 32;
9613 loop_mems = (loop_mem_info*)
9614 xrealloc (loop_mems,
9615 loop_mems_allocated * sizeof (loop_mem_info));
9618 /* Actually insert the MEM. */
9619 loop_mems[loop_mems_idx].mem = m;
9620 /* We can't hoist this MEM out of the loop if it's a BLKmode MEM
9621 because we can't put it in a register. We still store it in the
9622 table, though, so that if we see the same address later, but in a
9623 non-BLK mode, we'll not think we can optimize it at that point. */
9624 loop_mems[loop_mems_idx].optimize = (GET_MODE (m) != BLKmode);
9625 loop_mems[loop_mems_idx].reg = NULL_RTX;
9631 /* Like load_mems, but also ensures that SET_IN_LOOP,
9632 MAY_NOT_OPTIMIZE, REG_SINGLE_USAGE, and INSN_COUNT have the correct
9633 values after load_mems. */
9636 load_mems_and_recount_loop_regs_set (loop, insn_count)
9637 const struct loop *loop;
9640 int nregs = max_reg_num ();
9644 /* Recalculate set_in_loop and friends since load_mems may have
9645 created new registers. */
9646 if (max_reg_num () > nregs)
9652 nregs = max_reg_num ();
9654 if ((unsigned) nregs > set_in_loop->num_elements)
9656 /* Grow all the arrays. */
9657 VARRAY_GROW (set_in_loop, nregs);
9658 VARRAY_GROW (n_times_set, nregs);
9659 VARRAY_GROW (may_not_optimize, nregs);
9660 VARRAY_GROW (reg_single_usage, nregs);
9662 /* Clear the arrays */
9663 bzero ((char *) &set_in_loop->data, nregs * sizeof (int));
9664 bzero ((char *) &may_not_optimize->data, nregs * sizeof (char));
9665 bzero ((char *) ®_single_usage->data, nregs * sizeof (rtx));
9667 count_loop_regs_set (loop->top ? loop->top : loop->start, loop->end,
9668 may_not_optimize, reg_single_usage,
9671 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
9673 VARRAY_CHAR (may_not_optimize, i) = 1;
9674 VARRAY_INT (set_in_loop, i) = 1;
9677 #ifdef AVOID_CCMODE_COPIES
9678 /* Don't try to move insns which set CC registers if we should not
9679 create CCmode register copies. */
9680 for (i = max_reg_num () - 1; i >= FIRST_PSEUDO_REGISTER; i--)
9681 if (GET_MODE_CLASS (GET_MODE (regno_reg_rtx[i])) == MODE_CC)
9682 VARRAY_CHAR (may_not_optimize, i) = 1;
9685 /* Set n_times_set for the new registers. */
9686 bcopy ((char *) (&set_in_loop->data.i[0] + old_nregs),
9687 (char *) (&n_times_set->data.i[0] + old_nregs),
9688 (nregs - old_nregs) * sizeof (int));
9692 /* Move MEMs into registers for the duration of the loop. */
9696 const struct loop *loop;
9698 int maybe_never = 0;
9701 rtx label = NULL_RTX;
9702 rtx end_label = NULL_RTX;
9703 /* Nonzero if the next instruction may never be executed. */
9704 int next_maybe_never = 0;
9705 int last_max_reg = max_reg_num ();
9707 if (loop_mems_idx == 0)
9710 /* Find start of the extended basic block that enters the loop. */
9711 for (p = loop->start;
9712 PREV_INSN (p) && GET_CODE (p) != CODE_LABEL;
9718 /* Build table of mems that get set to constant values before the
9720 for (; p != loop->start; p = NEXT_INSN (p))
9721 cselib_process_insn (p);
9723 /* Check to see if it's possible that some instructions in the
9724 loop are never executed. */
9725 for (p = next_insn_in_loop (loop, loop->scan_start);
9726 p != NULL_RTX && ! maybe_never;
9727 p = next_insn_in_loop (loop, p))
9729 if (GET_CODE (p) == CODE_LABEL)
9731 else if (GET_CODE (p) == JUMP_INSN
9732 /* If we enter the loop in the middle, and scan
9733 around to the beginning, don't set maybe_never
9734 for that. This must be an unconditional jump,
9735 otherwise the code at the top of the loop might
9736 never be executed. Unconditional jumps are
9737 followed a by barrier then loop end. */
9738 && ! (GET_CODE (p) == JUMP_INSN
9739 && JUMP_LABEL (p) == loop->top
9740 && NEXT_INSN (NEXT_INSN (p)) == loop->end
9741 && simplejump_p (p)))
9743 if (!condjump_p (p))
9744 /* Something complicated. */
9747 /* If there are any more instructions in the loop, they
9748 might not be reached. */
9749 next_maybe_never = 1;
9751 else if (next_maybe_never)
9755 /* Actually move the MEMs. */
9756 for (i = 0; i < loop_mems_idx; ++i)
9761 rtx mem = loop_mems[i].mem;
9764 if (MEM_VOLATILE_P (mem)
9765 || loop_invariant_p (loop, XEXP (mem, 0)) != 1)
9766 /* There's no telling whether or not MEM is modified. */
9767 loop_mems[i].optimize = 0;
9769 /* Go through the MEMs written to in the loop to see if this
9770 one is aliased by one of them. */
9771 mem_list_entry = loop_store_mems;
9772 while (mem_list_entry)
9774 if (rtx_equal_p (mem, XEXP (mem_list_entry, 0)))
9776 else if (true_dependence (XEXP (mem_list_entry, 0), VOIDmode,
9779 /* MEM is indeed aliased by this store. */
9780 loop_mems[i].optimize = 0;
9783 mem_list_entry = XEXP (mem_list_entry, 1);
9786 if (flag_float_store && written
9787 && GET_MODE_CLASS (GET_MODE (mem)) == MODE_FLOAT)
9788 loop_mems[i].optimize = 0;
9790 /* If this MEM is written to, we must be sure that there
9791 are no reads from another MEM that aliases this one. */
9792 if (loop_mems[i].optimize && written)
9796 for (j = 0; j < loop_mems_idx; ++j)
9800 else if (true_dependence (mem,
9805 /* It's not safe to hoist loop_mems[i] out of
9806 the loop because writes to it might not be
9807 seen by reads from loop_mems[j]. */
9808 loop_mems[i].optimize = 0;
9814 if (maybe_never && may_trap_p (mem))
9815 /* We can't access the MEM outside the loop; it might
9816 cause a trap that wouldn't have happened otherwise. */
9817 loop_mems[i].optimize = 0;
9819 if (!loop_mems[i].optimize)
9820 /* We thought we were going to lift this MEM out of the
9821 loop, but later discovered that we could not. */
9824 INIT_REG_SET (&copies);
9826 /* Allocate a pseudo for this MEM. We set REG_USERVAR_P in
9827 order to keep scan_loop from moving stores to this MEM
9828 out of the loop just because this REG is neither a
9829 user-variable nor used in the loop test. */
9830 reg = gen_reg_rtx (GET_MODE (mem));
9831 REG_USERVAR_P (reg) = 1;
9832 loop_mems[i].reg = reg;
9834 /* Now, replace all references to the MEM with the
9835 corresponding pesudos. */
9837 for (p = next_insn_in_loop (loop, loop->scan_start);
9839 p = next_insn_in_loop (loop, p))
9844 if (GET_RTX_CLASS (GET_CODE (p)) == 'i')
9846 /* See if this copies the mem into a register that isn't
9847 modified afterwards. We'll try to do copy propagation
9848 a little further on. */
9849 set = single_set (p);
9851 /* @@@ This test is _way_ too conservative. */
9853 && GET_CODE (SET_DEST (set)) == REG
9854 && REGNO (SET_DEST (set)) >= FIRST_PSEUDO_REGISTER
9855 && REGNO (SET_DEST (set)) < last_max_reg
9856 && VARRAY_INT (n_times_set, REGNO (SET_DEST (set))) == 1
9857 && rtx_equal_p (SET_SRC (set), loop_mems[i].mem))
9858 SET_REGNO_REG_SET (&copies, REGNO (SET_DEST (set)));
9861 for_each_rtx (&p, replace_loop_mem, &ri);
9864 if (GET_CODE (p) == CODE_LABEL
9865 || GET_CODE (p) == JUMP_INSN)
9869 if (! apply_change_group ())
9870 /* We couldn't replace all occurrences of the MEM. */
9871 loop_mems[i].optimize = 0;
9874 /* Load the memory immediately before LOOP->START, which is
9875 the NOTE_LOOP_BEG. */
9876 cselib_val *e = cselib_lookup (mem, VOIDmode, 0);
9880 struct elt_loc_list *const_equiv = 0;
9884 struct elt_loc_list *equiv;
9885 struct elt_loc_list *best_equiv = 0;
9886 for (equiv = e->locs; equiv; equiv = equiv->next)
9888 if (CONSTANT_P (equiv->loc))
9889 const_equiv = equiv;
9890 else if (GET_CODE (equiv->loc) == REG
9891 /* Extending hard register lifetimes cuases crash
9892 on SRC targets. Doing so on non-SRC is
9893 probably also not good idea, since we most
9894 probably have pseudoregister equivalence as
9896 && REGNO (equiv->loc) >= FIRST_PSEUDO_REGISTER)
9899 /* Use the constant equivalence if that is cheap enough. */
9901 best_equiv = const_equiv;
9902 else if (const_equiv
9903 && (rtx_cost (const_equiv->loc, SET)
9904 <= rtx_cost (best_equiv->loc, SET)))
9906 best_equiv = const_equiv;
9910 /* If best_equiv is nonzero, we know that MEM is set to a
9911 constant or register before the loop. We will use this
9912 knowledge to initialize the shadow register with that
9913 constant or reg rather than by loading from MEM. */
9915 best = copy_rtx (best_equiv->loc);
9917 set = gen_move_insn (reg, best);
9918 set = emit_insn_before (set, loop->start);
9920 REG_NOTES (set) = gen_rtx_EXPR_LIST (REG_EQUAL,
9921 copy_rtx (const_equiv->loc),
9926 if (label == NULL_RTX)
9928 /* We must compute the former
9929 right-after-the-end label before we insert
9931 end_label = next_label (loop->end);
9932 label = gen_label_rtx ();
9933 emit_label_after (label, loop->end);
9936 /* Store the memory immediately after END, which is
9937 the NOTE_LOOP_END. */
9938 set = gen_move_insn (copy_rtx (mem), reg);
9939 emit_insn_after (set, label);
9942 if (loop_dump_stream)
9944 fprintf (loop_dump_stream, "Hoisted regno %d %s from ",
9945 REGNO (reg), (written ? "r/w" : "r/o"));
9946 print_rtl (loop_dump_stream, mem);
9947 fputc ('\n', loop_dump_stream);
9950 /* Attempt a bit of copy propagation. This helps untangle the
9951 data flow, and enables {basic,general}_induction_var to find
9953 EXECUTE_IF_SET_IN_REG_SET
9954 (&copies, FIRST_PSEUDO_REGISTER, j,
9956 try_copy_prop (loop, loop_mems[i].reg, j);
9958 CLEAR_REG_SET (&copies);
9962 if (label != NULL_RTX)
9964 /* Now, we need to replace all references to the previous exit
9965 label with the new one. */
9970 for (p = loop->start; p != loop->end; p = NEXT_INSN (p))
9972 for_each_rtx (&p, replace_label, &rr);
9974 /* If this is a JUMP_INSN, then we also need to fix the JUMP_LABEL
9975 field. This is not handled by for_each_rtx because it doesn't
9976 handle unprinted ('0') fields. We need to update JUMP_LABEL
9977 because the immediately following unroll pass will use it.
9978 replace_label would not work anyways, because that only handles
9980 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) == end_label)
9981 JUMP_LABEL (p) = label;
9988 /* For communication between note_reg_stored and its caller. */
9989 struct note_reg_stored_arg
9995 /* Called via note_stores, record in SET_SEEN whether X, which is written,
9998 note_reg_stored (x, setter, arg)
9999 rtx x, setter ATTRIBUTE_UNUSED;
10002 struct note_reg_stored_arg *t = (struct note_reg_stored_arg *)arg;
10007 /* Try to replace every occurrence of pseudo REGNO with REPLACEMENT.
10008 There must be exactly one insn that sets this pseudo; it will be
10009 deleted if all replacements succeed and we can prove that the register
10010 is not used after the loop. */
10013 try_copy_prop (loop, replacement, regno)
10014 const struct loop *loop;
10016 unsigned int regno;
10018 /* This is the reg that we are copying from. */
10019 rtx reg_rtx = regno_reg_rtx[regno];
10022 /* These help keep track of whether we replaced all uses of the reg. */
10023 int replaced_last = 0;
10024 int store_is_first = 0;
10026 for (insn = next_insn_in_loop (loop, loop->scan_start);
10028 insn = next_insn_in_loop (loop, insn))
10032 /* Only substitute within one extended basic block from the initializing
10034 if (GET_CODE (insn) == CODE_LABEL && init_insn)
10037 if (GET_RTX_CLASS (GET_CODE (insn)) != 'i')
10040 /* Is this the initializing insn? */
10041 set = single_set (insn);
10043 && GET_CODE (SET_DEST (set)) == REG
10044 && REGNO (SET_DEST (set)) == regno)
10050 if (REGNO_FIRST_UID (regno) == INSN_UID (insn))
10051 store_is_first = 1;
10054 /* Only substitute after seeing the initializing insn. */
10055 if (init_insn && insn != init_insn)
10057 struct note_reg_stored_arg arg;
10059 array[0] = reg_rtx;
10060 array[1] = replacement;
10063 for_each_rtx (&insn, replace_loop_reg, array);
10064 if (REGNO_LAST_UID (regno) == INSN_UID (insn))
10067 /* Stop replacing when REPLACEMENT is modified. */
10068 arg.reg = replacement;
10070 note_stores (PATTERN (insn), note_reg_stored, &arg);
10077 if (apply_change_group ())
10079 if (loop_dump_stream)
10080 fprintf (loop_dump_stream, " Replaced reg %d", regno);
10081 if (store_is_first && replaced_last)
10083 PUT_CODE (init_insn, NOTE);
10084 NOTE_LINE_NUMBER (init_insn) = NOTE_INSN_DELETED;
10085 if (loop_dump_stream)
10086 fprintf (loop_dump_stream, ", deleting init_insn (%d)",
10087 INSN_UID (init_insn));
10089 if (loop_dump_stream)
10090 fprintf (loop_dump_stream, ".\n");
10094 /* Replace MEM with its associated pseudo register. This function is
10095 called from load_mems via for_each_rtx. DATA is actually an
10096 rtx_and_int * describing the instruction currently being scanned
10097 and the MEM we are currently replacing. */
10100 replace_loop_mem (mem, data)
10112 switch (GET_CODE (m))
10118 /* We're not interested in the MEM associated with a
10119 CONST_DOUBLE, so there's no need to traverse into one. */
10123 /* This is not a MEM. */
10127 ri = (rtx_and_int*) data;
10130 if (!rtx_equal_p (loop_mems[i].mem, m))
10131 /* This is not the MEM we are currently replacing. */
10136 /* Actually replace the MEM. */
10137 validate_change (insn, mem, loop_mems[i].reg, 1);
10142 /* Replace one register with another. Called through for_each_rtx; PX points
10143 to the rtx being scanned. DATA is actually an array of three rtx's; the
10144 first one is the one to be replaced, and the second one the replacement.
10145 The third one is the current insn. */
10148 replace_loop_reg (px, data)
10153 rtx *array = (rtx *)data;
10159 validate_change (array[2], px, array[1], 1);
10164 /* Replace occurrences of the old exit label for the loop with the new
10165 one. DATA is an rtx_pair containing the old and new labels,
10169 replace_label (x, data)
10174 rtx old_label = ((rtx_pair*) data)->r1;
10175 rtx new_label = ((rtx_pair*) data)->r2;
10180 if (GET_CODE (l) != LABEL_REF)
10183 if (XEXP (l, 0) != old_label)
10186 XEXP (l, 0) = new_label;
10187 ++LABEL_NUSES (new_label);
10188 --LABEL_NUSES (old_label);