]> gcc.gnu.org Git - gcc.git/blob - gcc/loop.c
*** empty log message ***
[gcc.git] / gcc / loop.c
1 /* Move constant computations out of loops.
2 Copyright (C) 1987, 1988, 1989, 1991 Free Software Foundation, Inc.
3
4 This file is part of GNU CC.
5
6 GNU CC is free software; you can redistribute it and/or modify
7 it under the terms of the GNU General Public License as published by
8 the Free Software Foundation; either version 2, or (at your option)
9 any later version.
10
11 GNU CC is distributed in the hope that it will be useful,
12 but WITHOUT ANY WARRANTY; without even the implied warranty of
13 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 GNU General Public License for more details.
15
16 You should have received a copy of the GNU General Public License
17 along with GNU CC; see the file COPYING. If not, write to
18 the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA. */
19
20
21 /* This is the loop optimization pass of the compiler.
22 It finds invariant computations within loops and moves them
23 to the beginning of the loop. Then it identifies basic and
24 general induction variables. Strength reduction is applied to the general
25 induction variables, and induction variable elimination is applied to
26 the basic induction variables.
27
28 It also finds cases where
29 a register is set within the loop by zero-extending a narrower value
30 and changes these to zero the entire register once before the loop
31 and merely copy the low part within the loop.
32
33 Most of the complexity is in heuristics to decide when it is worth
34 while to do these things. */
35
36 #include "config.h"
37 #include "rtl.h"
38 #include "obstack.h"
39 #include "expr.h"
40 #include "insn-config.h"
41 #include "insn-flags.h"
42 #include "regs.h"
43 #include "hard-reg-set.h"
44 #include "recog.h"
45 #include "flags.h"
46 #include "real.h"
47 #include <stdio.h>
48 #include "loop.h"
49
50 /* Vector mapping INSN_UIDs to luids.
51 The luids are like uids but increase monononically always.
52 We use them to see whether a jump comes from outside a given loop. */
53
54 int *uid_luid;
55
56 /* Indexed by INSN_UID, contains the ordinal giving the (innermost) loop
57 number the insn is contained in. */
58
59 int *uid_loop_num;
60
61 /* 1 + largest uid of any insn. */
62
63 int max_uid_for_loop;
64
65 /* 1 + luid of last insn. */
66
67 static int max_luid;
68
69 /* Number of loops detected in current function. Used as index to the
70 next few tables. */
71
72 static int max_loop_num;
73
74 /* Indexed by loop number, contains the first and last insn of each loop. */
75
76 static rtx *loop_number_loop_starts, *loop_number_loop_ends;
77
78 /* For each loop, gives the containing loop number, -1 if none. */
79
80 int *loop_outer_loop;
81
82 /* Indexed by loop number, contains a nonzero value if the "loop" isn't
83 really a loop (an insn outside the loop branches into it). */
84
85 static char *loop_invalid;
86
87 /* Indexed by loop number, links together all LABEL_REFs which refer to
88 code labels outside the loop. Used by routines that need to know all
89 loop exits, such as final_biv_value and final_giv_value.
90
91 This does not include loop exits due to return instructions. This is
92 because all bivs and givs are pseudos, and hence must be dead after a
93 return, so the presense of a return does not affect any of the
94 optimizations that use this info. It is simpler to just not include return
95 instructions on this list. */
96
97 rtx *loop_number_exit_labels;
98
99 /* Holds the number of loop iterations. It is zero if the number could not be
100 calculated. Must be unsigned long since the number of iterations can
101 be as high as 2^31-1. For loops with a DImode iterator, this number will
102 will be zero if the number of loop iterations is too large for an
103 unsigned long to hold. */
104
105 unsigned long loop_n_iterations;
106
107 /* Nonzero if there is a subroutine call in the current loop.
108 (unknown_address_altered is also nonzero in this case.) */
109
110 static int loop_has_call;
111
112 /* Added loop_continue which is the NOTE_INSN_LOOP_CONT of the
113 current loop. A continue statement will generate a branch to
114 NEXT_INSN (loop_continue). */
115
116 static rtx loop_continue;
117
118 /* Indexed by register number, contains the number of times the reg
119 is set during the loop being scanned.
120 During code motion, a negative value indicates a reg that has been
121 made a candidate; in particular -2 means that it is an candidate that
122 we know is equal to a constant and -1 means that it is an condidate
123 not known equal to a constant.
124 After code motion, regs moved have 0 (which is accurate now)
125 while the failed candidates have the original number of times set.
126
127 Therefore, at all times, == 0 indicates an invariant register;
128 < 0 a conditionally invariant one. */
129
130 static short *n_times_set;
131
132 /* Original value of n_times_set; same except that this value
133 is not set negative for a reg whose sets have been made candidates
134 and not set to 0 for a reg that is moved. */
135
136 static short *n_times_used;
137
138 /* Index by register number, 1 indicates that the register
139 cannot be moved or strength reduced. */
140
141 static char *may_not_optimize;
142
143 /* Nonzero means reg N has already been moved out of one loop.
144 This reduces the desire to move it out of another. */
145
146 static char *moved_once;
147
148 /* Array of MEMs that are stored in this loop. If there are too many to fit
149 here, we just turn on unknown_address_altered. */
150
151 #define NUM_STORES 20
152 static rtx loop_store_mems[NUM_STORES];
153
154 /* Index of first available slot in above array. */
155 static int loop_store_mems_idx;
156
157 /* Nonzero if we don't know what MEMs were changed in the current loop.
158 This happens if the loop contains a call (in which call `loop_has_call'
159 will also be set) or if we store into more than NUM_STORES MEMs. */
160
161 static int unknown_address_altered;
162
163 /* Count of movable (i.e. invariant) instructions discovered in the loop. */
164 static int num_movables;
165
166 /* Count of memory write instructions discovered in the loop. */
167 static int num_mem_sets;
168
169 /* Number of loops contained within the current one, including itself. */
170 static int loops_enclosed;
171
172 /* Bound on pseudo register number before loop optimization.
173 A pseudo has valid regscan info if its number is < max_reg_before_loop. */
174 int max_reg_before_loop;
175
176 /* This obstack is used in product_cheap_p to allocate its rtl. It
177 may call gen_reg_rtx which, in turn, may reallocate regno_reg_rtx.
178 If we used the same obstack that it did, we would be deallocating
179 that array. */
180
181 static struct obstack temp_obstack;
182
183 /* This is where the pointer to the obstack being used for RTL is stored. */
184
185 extern struct obstack *rtl_obstack;
186
187 #define obstack_chunk_alloc xmalloc
188 #define obstack_chunk_free free
189
190 extern char *oballoc ();
191 extern int xmalloc ();
192 extern void free ();
193 \f
194 /* During the analysis of a loop, a chain of `struct movable's
195 is made to record all the movable insns found.
196 Then the entire chain can be scanned to decide which to move. */
197
198 struct movable
199 {
200 rtx insn; /* A movable insn */
201 rtx set_src; /* The expression this reg is set from. */
202 rtx set_dest; /* The destination of this SET. */
203 rtx dependencies; /* When INSN is libcall, this is an EXPR_LIST
204 of any registers used within the LIBCALL. */
205 int consec; /* Number of consecutive following insns
206 that must be moved with this one. */
207 int regno; /* The register it sets */
208 short lifetime; /* lifetime of that register;
209 may be adjusted when matching movables
210 that load the same value are found. */
211 short savings; /* Number of insns we can move for this reg,
212 including other movables that force this
213 or match this one. */
214 unsigned int cond : 1; /* 1 if only conditionally movable */
215 unsigned int force : 1; /* 1 means MUST move this insn */
216 unsigned int global : 1; /* 1 means reg is live outside this loop */
217 /* If PARTIAL is 1, GLOBAL means something different:
218 that the reg is live outside the range from where it is set
219 to the following label. */
220 unsigned int done : 1; /* 1 inhibits further processing of this */
221
222 unsigned int partial : 1; /* 1 means this reg is used for zero-extending.
223 In particular, moving it does not make it
224 invariant. */
225 unsigned int move_insn : 1; /* 1 means that we call emit_move_insn to
226 load SRC, rather than copying INSN. */
227 unsigned int is_equiv : 1; /* 1 means a REG_EQUIV is present on INSN. */
228 enum machine_mode savemode; /* Nonzero means it is a mode for a low part
229 that we should avoid changing when clearing
230 the rest of the reg. */
231 struct movable *match; /* First entry for same value */
232 struct movable *forces; /* An insn that must be moved if this is */
233 struct movable *next;
234 };
235
236 FILE *loop_dump_stream;
237
238 /* Forward declarations. */
239
240 static void find_and_verify_loops ();
241 static void mark_loop_jump ();
242 static void prescan_loop ();
243 static int reg_in_basic_block_p ();
244 static int consec_sets_invariant_p ();
245 static rtx libcall_other_reg ();
246 static int labels_in_range_p ();
247 static void count_loop_regs_set ();
248 static void note_addr_stored ();
249 static int loop_reg_used_before_p ();
250 static void scan_loop ();
251 static void replace_call_address ();
252 static rtx skip_consec_insns ();
253 static int libcall_benefit ();
254 static void ignore_some_movables ();
255 static void force_movables ();
256 static void combine_movables ();
257 static int rtx_equal_for_loop_p ();
258 static void move_movables ();
259 static void strength_reduce ();
260 static int valid_initial_value_p ();
261 static void find_mem_givs ();
262 static void record_biv ();
263 static void check_final_value ();
264 static void record_giv ();
265 static void update_giv_derive ();
266 static void delete_insn_forces ();
267 static int basic_induction_var ();
268 static rtx simplify_giv_expr ();
269 static int general_induction_var ();
270 static int consec_sets_giv ();
271 static int check_dbra_loop ();
272 static rtx express_from ();
273 static int combine_givs_p ();
274 static void combine_givs ();
275 static int product_cheap_p ();
276 static int maybe_eliminate_biv ();
277 static int maybe_eliminate_biv_1 ();
278 static int last_use_this_basic_block ();
279 static void record_initial ();
280 static void update_reg_last_use ();
281 \f
282 /* Relative gain of eliminating various kinds of operations. */
283 int add_cost;
284 #if 0
285 int shift_cost;
286 int mult_cost;
287 #endif
288
289 /* Benefit penalty, if a giv is not replaceable, i.e. must emit an insn to
290 copy the value of the strength reduced giv to its original register. */
291 int copy_cost;
292
293 void
294 init_loop ()
295 {
296 char *free_point = (char *) oballoc (1);
297 rtx reg = gen_rtx (REG, SImode, 0);
298 rtx pow2 = gen_rtx (CONST_INT, VOIDmode, 32);
299 rtx lea;
300 int i;
301
302 add_cost = rtx_cost (gen_rtx (PLUS, SImode, reg, reg));
303
304 /* We multiply by 2 to reconcile the difference in scale between
305 these two ways of computing costs. Otherwise the cost of a copy
306 will be far less than the cost of an add. */
307 #ifdef REGISTER_MOVE_COST
308 copy_cost = REGISTER_MOVE_COST (GENERAL_REGS, GENERAL_REGS) * 2;
309 #else
310 copy_cost = 2 * 2;
311 #endif
312
313 /* Free the objects we just allocated. */
314 obfree (free_point);
315
316 /* Initialize the obstack used for rtl in product_cheap_p. */
317 gcc_obstack_init (&temp_obstack);
318 }
319 \f
320 /* Entry point of this file. Perform loop optimization
321 on the current function. F is the first insn of the function
322 and DUMPFILE is a stream for output of a trace of actions taken
323 (or 0 if none should be output). */
324
325 void
326 loop_optimize (f, dumpfile)
327 /* f is the first instruction of a chain of insns for one function */
328 rtx f;
329 FILE *dumpfile;
330 {
331 register rtx insn;
332 register int i;
333 rtx end;
334 rtx last_insn;
335
336 loop_dump_stream = dumpfile;
337
338 init_recog_no_volatile ();
339 init_alias_analysis ();
340
341 max_reg_before_loop = max_reg_num ();
342
343 moved_once = (char *) alloca (max_reg_before_loop);
344 bzero (moved_once, max_reg_before_loop);
345
346 regs_may_share = 0;
347
348 /* Count the number of loops. */
349
350 max_loop_num = 0;
351 for (insn = f; insn; insn = NEXT_INSN (insn))
352 {
353 if (GET_CODE (insn) == NOTE
354 && NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG)
355 max_loop_num++;
356 }
357
358 /* Don't waste time if no loops. */
359 if (max_loop_num == 0)
360 return;
361
362 /* Get size to use for tables indexed by uids.
363 Leave some space for labels allocated by find_and_verify_loops. */
364 max_uid_for_loop = get_max_uid () + 1 + max_loop_num * 4;
365
366 uid_luid = (int *) alloca (max_uid_for_loop * sizeof (int));
367 uid_loop_num = (int *) alloca (max_uid_for_loop * sizeof (int));
368
369 bzero (uid_luid, max_uid_for_loop * sizeof (int));
370 bzero (uid_loop_num, max_uid_for_loop * sizeof (int));
371
372 /* Allocate tables for recording each loop. We set each entry, so they need
373 not be zeroed. */
374 loop_number_loop_starts = (rtx *) alloca (max_loop_num * sizeof (rtx));
375 loop_number_loop_ends = (rtx *) alloca (max_loop_num * sizeof (rtx));
376 loop_outer_loop = (int *) alloca (max_loop_num * sizeof (int));
377 loop_invalid = (char *) alloca (max_loop_num * sizeof (char));
378 loop_number_exit_labels = (rtx *) alloca (max_loop_num * sizeof (rtx));
379
380 if (flag_unroll_loops && write_symbols != NO_DEBUG)
381 {
382 loop_number_first_block
383 = (union tree_node **) alloca (max_loop_num
384 * sizeof (union tree_node *));
385 loop_number_last_block
386 = (union tree_node **) alloca (max_loop_num
387 * sizeof (union tree_node *));
388 loop_number_block_level = (int *) alloca (max_loop_num * sizeof (int));
389 }
390
391 /* Find and process each loop.
392 First, find them, and record them in order of their beginnings. */
393 find_and_verify_loops (f);
394
395 /* Now find all register lifetimes. This must be done after
396 find_and_verify_loops, because it might reorder the insns in the
397 function. */
398 reg_scan (f, max_reg_num (), 1);
399
400 /* Compute the mapping from uids to luids.
401 LUIDs are numbers assigned to insns, like uids,
402 except that luids increase monotonically through the code.
403 Don't assign luids to line-number NOTEs, so that the distance in luids
404 between two insns is not affected by -g. */
405
406 for (insn = f, i = 0; insn; insn = NEXT_INSN (insn))
407 {
408 last_insn = insn;
409 if (GET_CODE (insn) != NOTE
410 || NOTE_LINE_NUMBER (insn) <= 0)
411 uid_luid[INSN_UID (insn)] = ++i;
412 else
413 /* Give a line number note the same luid as preceding insn. */
414 uid_luid[INSN_UID (insn)] = i;
415 }
416
417 max_luid = i + 1;
418
419 /* Don't leave gaps in uid_luid for insns that have been
420 deleted. It is possible that the first or last insn
421 using some register has been deleted by cross-jumping.
422 Make sure that uid_luid for that former insn's uid
423 points to the general area where that insn used to be. */
424 for (i = 0; i < max_uid_for_loop; i++)
425 {
426 uid_luid[0] = uid_luid[i];
427 if (uid_luid[0] != 0)
428 break;
429 }
430 for (i = 0; i < max_uid_for_loop; i++)
431 if (uid_luid[i] == 0)
432 uid_luid[i] = uid_luid[i - 1];
433
434 /* Create a mapping from loops to BLOCK tree nodes. */
435 if (flag_unroll_loops && write_symbols != NO_DEBUG)
436 find_loop_tree_blocks (f);
437
438 /* Now scan the loops, last ones first, since this means inner ones are done
439 before outer ones. */
440 for (i = max_loop_num-1; i >= 0; i--)
441 if (! loop_invalid[i] && loop_number_loop_ends[i])
442 scan_loop (loop_number_loop_starts[i], loop_number_loop_ends[i],
443 max_reg_num ());
444 }
445 \f
446 /* Optimize one loop whose start is LOOP_START and end is END.
447 LOOP_START is the NOTE_INSN_LOOP_BEG and END is the matching
448 NOTE_INSN_LOOP_END. */
449
450 /* ??? Could also move memory writes out of loops if the destination address
451 is invariant, the source is invariant, the memory write is not volatile,
452 and if we can prove that no read inside the loop can read this address
453 before the write occurs. If there is a read of this address after the
454 write, then we can also mark the memory read as invariant. */
455
456 static void
457 scan_loop (loop_start, end, nregs)
458 rtx loop_start, end;
459 int nregs;
460 {
461 register int i;
462 register rtx p;
463 /* 1 if we are scanning insns that could be executed zero times. */
464 int maybe_never = 0;
465 /* 1 if we are scanning insns that might never be executed
466 due to a subroutine call which might exit before they are reached. */
467 int call_passed = 0;
468 /* For a rotated loop that is entered near the bottom,
469 this is the label at the top. Otherwise it is zero. */
470 rtx loop_top = 0;
471 /* Jump insn that enters the loop, or 0 if control drops in. */
472 rtx loop_entry_jump = 0;
473 /* Place in the loop where control enters. */
474 rtx scan_start;
475 /* Number of insns in the loop. */
476 int insn_count;
477 int in_libcall = 0;
478 int tem;
479 rtx temp;
480 /* The SET from an insn, if it is the only SET in the insn. */
481 rtx set, set1;
482 /* Chain describing insns movable in current loop. */
483 struct movable *movables = 0;
484 /* Last element in `movables' -- so we can add elements at the end. */
485 struct movable *last_movable = 0;
486 /* Ratio of extra register life span we can justify
487 for saving an instruction. More if loop doesn't call subroutines
488 since in that case saving an insn makes more difference
489 and more registers are available. */
490 int threshold;
491 /* If we have calls, contains the insn in which a register was used
492 if it was used exactly once; contains const0_rtx if it was used more
493 than once. */
494 rtx *reg_single_usage = 0;
495
496 n_times_set = (short *) alloca (nregs * sizeof (short));
497 n_times_used = (short *) alloca (nregs * sizeof (short));
498 may_not_optimize = (char *) alloca (nregs);
499
500 /* Determine whether this loop starts with a jump down to a test at
501 the end. This will occur for a small number of loops with a test
502 that is too complex to duplicate in front of the loop.
503
504 We search for the first insn or label in the loop, skipping NOTEs.
505 However, we must be careful not to skip past a NOTE_INSN_LOOP_BEG
506 (because we might have a loop executed only once that contains a
507 loop which starts with a jump to its exit test) or a NOTE_INSN_LOOP_END
508 (in case we have a degenerate loop).
509
510 Note that if we mistakenly think that a loop is entered at the top
511 when, in fact, it is entered at the exit test, the only effect will be
512 slightly poorer optimization. Making the opposite error can generate
513 incorrect code. Since very few loops now start with a jump to the
514 exit test, the code here to detect that case is very conservative. */
515
516 for (p = NEXT_INSN (loop_start);
517 p != end
518 && GET_CODE (p) != CODE_LABEL && GET_RTX_CLASS (GET_CODE (p)) != 'i'
519 && (GET_CODE (p) != NOTE
520 || (NOTE_LINE_NUMBER (p) != NOTE_INSN_LOOP_BEG
521 && NOTE_LINE_NUMBER (p) != NOTE_INSN_LOOP_END));
522 p = NEXT_INSN (p))
523 ;
524
525 scan_start = p;
526
527 /* Set up variables describing this loop. */
528 prescan_loop (loop_start, end);
529 threshold = (loop_has_call ? 1 : 2) * (1 + n_non_fixed_regs);
530
531 /* If loop has a jump before the first label,
532 the true entry is the target of that jump.
533 Start scan from there.
534 But record in LOOP_TOP the place where the end-test jumps
535 back to so we can scan that after the end of the loop. */
536 if (GET_CODE (p) == JUMP_INSN)
537 {
538 loop_entry_jump = p;
539
540 /* Loop entry must be unconditional jump (and not a RETURN) */
541 if (simplejump_p (p)
542 && JUMP_LABEL (p) != 0
543 /* Check to see whether the jump actually
544 jumps out of the loop (meaning it's no loop).
545 This case can happen for things like
546 do {..} while (0). If this label was generated previously
547 by loop, we can't tell anything about it and have to reject
548 the loop. */
549 && INSN_UID (JUMP_LABEL (p)) < max_uid_for_loop
550 && INSN_LUID (JUMP_LABEL (p)) >= INSN_LUID (loop_start)
551 && INSN_LUID (JUMP_LABEL (p)) < INSN_LUID (end))
552 {
553 loop_top = next_label (scan_start);
554 scan_start = JUMP_LABEL (p);
555 }
556 }
557
558 /* If SCAN_START was an insn created by loop, we don't know its luid
559 as required by loop_reg_used_before_p. So skip such loops. (This
560 test may never be true, but it's best to play it safe.)
561
562 Also, skip loops where we do not start scanning at a label. This
563 test also rejects loops starting with a JUMP_INSN that failed the
564 test above. */
565
566 if (INSN_UID (scan_start) >= max_uid_for_loop
567 || GET_CODE (scan_start) != CODE_LABEL)
568 {
569 if (loop_dump_stream)
570 fprintf (loop_dump_stream, "\nLoop from %d to %d is phony.\n\n",
571 INSN_UID (loop_start), INSN_UID (end));
572 return;
573 }
574
575 /* Count number of times each reg is set during this loop.
576 Set may_not_optimize[I] if it is not safe to move out
577 the setting of register I. If this loop has calls, set
578 reg_single_usage[I]. */
579
580 bzero (n_times_set, nregs * sizeof (short));
581 bzero (may_not_optimize, nregs);
582
583 if (loop_has_call)
584 {
585 reg_single_usage = (rtx *) alloca (nregs * sizeof (rtx));
586 bzero (reg_single_usage, nregs * sizeof (rtx));
587 }
588
589 count_loop_regs_set (loop_top ? loop_top : loop_start, end,
590 may_not_optimize, reg_single_usage, &insn_count, nregs);
591
592 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
593 may_not_optimize[i] = 1, n_times_set[i] = 1;
594 bcopy (n_times_set, n_times_used, nregs * sizeof (short));
595
596 if (loop_dump_stream)
597 {
598 fprintf (loop_dump_stream, "\nLoop from %d to %d: %d real insns.\n",
599 INSN_UID (loop_start), INSN_UID (end), insn_count);
600 if (loop_continue)
601 fprintf (loop_dump_stream, "Continue at insn %d.\n",
602 INSN_UID (loop_continue));
603 }
604
605 /* Scan through the loop finding insns that are safe to move.
606 In each such insn, store QImode as the mode, to mark it.
607 Then set n_times_set negative for the reg being set, so that
608 this reg will be considered invariant for subsequent insns.
609 We consider whether subsequent insns use the reg
610 in deciding whether it is worth actually moving.
611
612 MAYBE_NEVER is nonzero if we have passed a conditional jump insn
613 and therefore it is possible that the insns we are scanning
614 would never be executed. At such times, we must make sure
615 that it is safe to execute the insn once instead of zero times.
616 When MAYBE_NEVER is 0, all insns will be executed at least once
617 so that is not a problem. */
618
619 p = scan_start;
620 while (1)
621 {
622 p = NEXT_INSN (p);
623 /* At end of a straight-in loop, we are done.
624 At end of a loop entered at the bottom, scan the top. */
625 if (p == scan_start)
626 break;
627 if (p == end)
628 {
629 if (loop_top != 0)
630 p = NEXT_INSN (loop_top);
631 else
632 break;
633 if (p == scan_start)
634 break;
635 }
636
637 if (GET_RTX_CLASS (GET_CODE (p)) == 'i'
638 && find_reg_note (p, REG_LIBCALL, 0))
639 in_libcall = 1;
640 else if (GET_RTX_CLASS (GET_CODE (p)) == 'i'
641 && find_reg_note (p, REG_RETVAL, 0))
642 in_libcall = 0;
643
644 if (GET_CODE (p) == INSN
645 && (set = single_set (p))
646 && GET_CODE (SET_DEST (set)) == REG
647 && ! may_not_optimize[REGNO (SET_DEST (set))])
648 {
649 int tem1 = 0;
650 int tem2 = 0;
651 int move_insn = 0;
652 rtx src = SET_SRC (set);
653 rtx dependencies = 0;
654
655 /* Figure out what to use as a source of this insn. If a REG_EQUIV
656 note is given or if a REG_EQUAL note with a constant operand is
657 specified, use it as the source and mark that we should move
658 this insn by calling emit_move_insn rather that duplicating the
659 insn.
660
661 Otherwise, only use the REG_EQUAL contents if a REG_RETVAL note
662 is present. */
663 temp = find_reg_note (p, REG_EQUIV, 0);
664 if (temp)
665 src = XEXP (temp, 0), move_insn = 1;
666 else
667 {
668 temp = find_reg_note (p, REG_EQUAL, 0);
669 if (temp && CONSTANT_P (XEXP (temp, 0)))
670 src = XEXP (temp, 0), move_insn = 1;
671 if (temp && find_reg_note (p, REG_RETVAL, 0))
672 {
673 src = XEXP (temp, 0);
674 /* A libcall block can use regs that don't appear in
675 the equivalent expression. To move the libcall,
676 we must move those regs too. */
677 dependencies = libcall_other_reg (p, src);
678 }
679 }
680
681 /* Don't try to optimize a register that was made
682 by loop-optimization for an inner loop.
683 We don't know its life-span, so we can't compute the benefit. */
684 if (REGNO (SET_DEST (set)) >= max_reg_before_loop)
685 ;
686 /* In order to move a register, we need to have one of three cases:
687 (1) it is used only in the same basic block as the set
688 (2) it is not a user variable.
689 (3) the set is guaranteed to be executed once the loop starts,
690 and the reg is not used until after that. */
691 else if (! ((! maybe_never
692 && ! loop_reg_used_before_p (set, p, loop_start,
693 scan_start, end))
694 || ! REG_USERVAR_P (SET_DEST (PATTERN (p)))
695 || reg_in_basic_block_p (p, SET_DEST (PATTERN (p)))))
696 ;
697 else if ((tem = invariant_p (src))
698 && (dependencies == 0
699 || (tem2 = invariant_p (dependencies)) != 0)
700 && (n_times_set[REGNO (SET_DEST (set))] == 1
701 || (tem1
702 = consec_sets_invariant_p (SET_DEST (set),
703 n_times_set[REGNO (SET_DEST (set))],
704 p)))
705 /* If the insn can cause a trap (such as divide by zero),
706 can't move it unless it's guaranteed to be executed
707 once loop is entered. Even a function call might
708 prevent the trap insn from being reached
709 (since it might exit!) */
710 && ! ((maybe_never || call_passed)
711 && may_trap_p (src)))
712 {
713 register struct movable *m;
714 register int regno = REGNO (SET_DEST (set));
715
716 /* A potential lossage is where we have a case where two insns
717 can be combined as long as they are both in the loop, but
718 we move one of them outside the loop. For large loops,
719 this can lose. The most common case of this is the address
720 of a function being called.
721
722 Therefore, if this register is marked as being used exactly
723 once if we are in a loop with calls (a "large loop"), see if
724 we can replace the usage of this register with the source
725 of this SET. If we can, delete this insn.
726
727 Don't do this if P has a REG_RETVAL note or if we have
728 SMALL_REGISTER_CLASSES and SET_SRC is a hard register. */
729
730 if (reg_single_usage && reg_single_usage[regno] != 0
731 && reg_single_usage[regno] != const0_rtx
732 && regno_first_uid[regno] == INSN_UID (p)
733 && (regno_last_uid[regno]
734 == INSN_UID (reg_single_usage[regno]))
735 && n_times_set[REGNO (SET_DEST (set))] == 1
736 && ! side_effects_p (SET_SRC (set))
737 && ! find_reg_note (p, REG_RETVAL, 0)
738 #ifdef SMALL_REGISTER_CLASSES
739 && ! (GET_CODE (SET_SRC (set)) == REG
740 && REGNO (SET_SRC (set)) < FIRST_PSEUDO_REGISTER)
741 #endif
742 /* This test is not redundant; SET_SRC (set) might be
743 a call-clobbered register and the life of REGNO
744 might span a call. */
745 && ! modified_between_p (SET_SRC (set), p,
746 reg_single_usage[regno])
747 && validate_replace_rtx (SET_DEST (set), SET_SRC (set),
748 reg_single_usage[regno]))
749 {
750 /* Replace any usage in a REG_EQUAL note. */
751 REG_NOTES (reg_single_usage[regno])
752 = replace_rtx (REG_NOTES (reg_single_usage[regno]),
753 SET_DEST (set), SET_SRC (set));
754
755 PUT_CODE (p, NOTE);
756 NOTE_LINE_NUMBER (p) = NOTE_INSN_DELETED;
757 NOTE_SOURCE_FILE (p) = 0;
758 n_times_set[regno] = 0;
759 continue;
760 }
761
762 m = (struct movable *) alloca (sizeof (struct movable));
763 m->next = 0;
764 m->insn = p;
765 m->set_src = src;
766 m->dependencies = dependencies;
767 m->set_dest = SET_DEST (set);
768 m->force = 0;
769 m->consec = n_times_set[REGNO (SET_DEST (set))] - 1;
770 m->done = 0;
771 m->forces = 0;
772 m->partial = 0;
773 m->move_insn = move_insn;
774 m->is_equiv = (find_reg_note (p, REG_EQUIV, 0) != 0);
775 m->savemode = VOIDmode;
776 m->regno = regno;
777 /* Set M->cond if either invariant_p or consec_sets_invariant_p
778 returned 2 (only conditionally invariant). */
779 m->cond = ((tem | tem1 | tem2) > 1);
780 m->global = (uid_luid[regno_last_uid[regno]] > INSN_LUID (end)
781 || uid_luid[regno_first_uid[regno]] < INSN_LUID (loop_start));
782 m->match = 0;
783 m->lifetime = (uid_luid[regno_last_uid[regno]]
784 - uid_luid[regno_first_uid[regno]]);
785 m->savings = n_times_used[regno];
786 if (find_reg_note (p, REG_RETVAL, 0))
787 m->savings += libcall_benefit (p);
788 n_times_set[regno] = move_insn ? -2 : -1;
789 /* Add M to the end of the chain MOVABLES. */
790 if (movables == 0)
791 movables = m;
792 else
793 last_movable->next = m;
794 last_movable = m;
795
796 if (m->consec > 0)
797 {
798 /* Skip this insn, not checking REG_LIBCALL notes. */
799 p = NEXT_INSN (p);
800 /* Skip the consecutive insns, if there are any. */
801 p = skip_consec_insns (p, m->consec);
802 /* Back up to the last insn of the consecutive group. */
803 p = prev_nonnote_insn (p);
804
805 /* We must now reset m->move_insn, m->is_equiv, and possibly
806 m->set_src to correspond to the effects of all the
807 insns. */
808 temp = find_reg_note (p, REG_EQUIV, 0);
809 if (temp)
810 m->set_src = XEXP (temp, 0), m->move_insn = 1;
811 else
812 {
813 temp = find_reg_note (p, REG_EQUAL, 0);
814 if (temp && CONSTANT_P (XEXP (temp, 0)))
815 m->set_src = XEXP (temp, 0), m->move_insn = 1;
816 else
817 m->move_insn = 0;
818
819 }
820 m->is_equiv = (find_reg_note (p, REG_EQUIV, 0) != 0);
821 }
822 }
823 /* If this register is always set within a STRICT_LOW_PART
824 or set to zero, then its high bytes are constant.
825 So clear them outside the loop and within the loop
826 just load the low bytes.
827 We must check that the machine has an instruction to do so.
828 Also, if the value loaded into the register
829 depends on the same register, this cannot be done. */
830 else if (SET_SRC (set) == const0_rtx
831 && GET_CODE (NEXT_INSN (p)) == INSN
832 && (set1 = single_set (NEXT_INSN (p)))
833 && GET_CODE (set1) == SET
834 && (GET_CODE (SET_DEST (set1)) == STRICT_LOW_PART)
835 && (GET_CODE (XEXP (SET_DEST (set1), 0)) == SUBREG)
836 && (SUBREG_REG (XEXP (SET_DEST (set1), 0))
837 == SET_DEST (set))
838 && !reg_mentioned_p (SET_DEST (set), SET_SRC (set1)))
839 {
840 register int regno = REGNO (SET_DEST (set));
841 if (n_times_set[regno] == 2)
842 {
843 register struct movable *m;
844 m = (struct movable *) alloca (sizeof (struct movable));
845 m->next = 0;
846 m->insn = p;
847 m->set_dest = SET_DEST (set);
848 m->dependencies = 0;
849 m->force = 0;
850 m->consec = 0;
851 m->done = 0;
852 m->forces = 0;
853 m->move_insn = 0;
854 m->partial = 1;
855 /* If the insn may not be executed on some cycles,
856 we can't clear the whole reg; clear just high part.
857 Not even if the reg is used only within this loop.
858 Consider this:
859 while (1)
860 while (s != t) {
861 if (foo ()) x = *s;
862 use (x);
863 }
864 Clearing x before the inner loop could clobber a value
865 being saved from the last time around the outer loop.
866 However, if the reg is not used outside this loop
867 and all uses of the register are in the same
868 basic block as the store, there is no problem.
869
870 If this insn was made by loop, we don't know its
871 INSN_LUID and hence must make a conservative
872 assumption. */
873 m->global = (INSN_UID (p) >= max_uid_for_loop
874 || (uid_luid[regno_last_uid[regno]]
875 > INSN_LUID (end))
876 || (uid_luid[regno_first_uid[regno]]
877 < INSN_LUID (p))
878 || (labels_in_range_p
879 (p, uid_luid[regno_first_uid[regno]])));
880 if (maybe_never && m->global)
881 m->savemode = GET_MODE (SET_SRC (set1));
882 else
883 m->savemode = VOIDmode;
884 m->regno = regno;
885 m->cond = 0;
886 m->match = 0;
887 m->lifetime = (uid_luid[regno_last_uid[regno]]
888 - uid_luid[regno_first_uid[regno]]);
889 m->savings = 1;
890 n_times_set[regno] = -1;
891 /* Add M to the end of the chain MOVABLES. */
892 if (movables == 0)
893 movables = m;
894 else
895 last_movable->next = m;
896 last_movable = m;
897 }
898 }
899 }
900 /* Past a call insn, we get to insns which might not be executed
901 because the call might exit. This matters for insns that trap.
902 Call insns inside a REG_LIBCALL/REG_RETVAL block always return,
903 so they don't count. */
904 else if (GET_CODE (p) == CALL_INSN && ! in_libcall)
905 call_passed = 1;
906 /* Past a label or a jump, we get to insns for which we
907 can't count on whether or how many times they will be
908 executed during each iteration. Therefore, we can
909 only move out sets of trivial variables
910 (those not used after the loop). */
911 /* This code appears in three places, once in scan_loop, and twice
912 in strength_reduce. */
913 else if ((GET_CODE (p) == CODE_LABEL || GET_CODE (p) == JUMP_INSN)
914 /* If we enter the loop in the middle, and scan around to the
915 beginning, don't set maybe_never for that. This must be an
916 unconditional jump, otherwise the code at the top of the
917 loop might never be executed. Unconditional jumps are
918 followed a by barrier then loop end. */
919 && ! (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) == loop_top
920 && NEXT_INSN (NEXT_INSN (p)) == end
921 && simplejump_p (p)))
922 maybe_never = 1;
923 /* At the virtual top of a converted loop, insns are again known to
924 be executed: logically, the loop begins here even though the exit
925 code has been duplicated. */
926 else if (GET_CODE (p) == NOTE
927 && NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP)
928 maybe_never = call_passed = 0;
929 }
930
931 /* If one movable subsumes another, ignore that other. */
932
933 ignore_some_movables (movables);
934
935 /* For each movable insn, see if the reg that it loads
936 leads when it dies right into another conditionally movable insn.
937 If so, record that the second insn "forces" the first one,
938 since the second can be moved only if the first is. */
939
940 force_movables (movables);
941
942 /* See if there are multiple movable insns that load the same value.
943 If there are, make all but the first point at the first one
944 through the `match' field, and add the priorities of them
945 all together as the priority of the first. */
946
947 combine_movables (movables, nregs);
948
949 /* Now consider each movable insn to decide whether it is worth moving.
950 Store 0 in n_times_set for each reg that is moved. */
951
952 move_movables (movables, threshold,
953 insn_count, loop_start, end, nregs);
954
955 /* Now candidates that still are negative are those not moved.
956 Change n_times_set to indicate that those are not actually invariant. */
957 for (i = 0; i < nregs; i++)
958 if (n_times_set[i] < 0)
959 n_times_set[i] = n_times_used[i];
960
961 if (flag_strength_reduce)
962 strength_reduce (scan_start, end, loop_top,
963 insn_count, loop_start, end);
964 }
965 \f
966 /* Add elements to *OUTPUT to record all the pseudo-regs
967 mentioned in IN_THIS but not mentioned in NOT_IN_THIS. */
968
969 void
970 record_excess_regs (in_this, not_in_this, output)
971 rtx in_this, not_in_this;
972 rtx *output;
973 {
974 enum rtx_code code;
975 char *fmt;
976 int i;
977
978 code = GET_CODE (in_this);
979
980 switch (code)
981 {
982 case PC:
983 case CC0:
984 case CONST_INT:
985 case CONST_DOUBLE:
986 case CONST:
987 case SYMBOL_REF:
988 case LABEL_REF:
989 return;
990
991 case REG:
992 if (REGNO (in_this) >= FIRST_PSEUDO_REGISTER
993 && ! reg_mentioned_p (in_this, not_in_this))
994 *output = gen_rtx (EXPR_LIST, VOIDmode, in_this, *output);
995 return;
996 }
997
998 fmt = GET_RTX_FORMAT (code);
999 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1000 {
1001 int j;
1002
1003 switch (fmt[i])
1004 {
1005 case 'E':
1006 for (j = 0; j < XVECLEN (in_this, i); j++)
1007 record_excess_regs (XVECEXP (in_this, i, j), not_in_this, output);
1008 break;
1009
1010 case 'e':
1011 record_excess_regs (XEXP (in_this, i), not_in_this, output);
1012 break;
1013 }
1014 }
1015 }
1016 \f
1017 /* Check what regs are referred to in the libcall block ending with INSN,
1018 aside from those mentioned in the equivalent value.
1019 If there are none, return 0.
1020 If there are one or more, return an EXPR_LIST containing all of them. */
1021
1022 static rtx
1023 libcall_other_reg (insn, equiv)
1024 rtx insn, equiv;
1025 {
1026 rtx note = find_reg_note (insn, REG_RETVAL, 0);
1027 rtx p = XEXP (note, 0);
1028 rtx output = 0;
1029
1030 /* First, find all the regs used in the libcall block
1031 that are not mentioned as inputs to the result. */
1032
1033 while (p != insn)
1034 {
1035 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
1036 || GET_CODE (p) == CALL_INSN)
1037 record_excess_regs (PATTERN (p), equiv, &output);
1038 p = NEXT_INSN (p);
1039 }
1040
1041 return output;
1042 }
1043 \f
1044 /* Return 1 if all uses of REG
1045 are between INSN and the end of the basic block. */
1046
1047 static int
1048 reg_in_basic_block_p (insn, reg)
1049 rtx insn, reg;
1050 {
1051 int regno = REGNO (reg);
1052 rtx p;
1053
1054 if (regno_first_uid[regno] != INSN_UID (insn))
1055 return 0;
1056
1057 /* Search this basic block for the already recorded last use of the reg. */
1058 for (p = insn; p; p = NEXT_INSN (p))
1059 {
1060 switch (GET_CODE (p))
1061 {
1062 case NOTE:
1063 break;
1064
1065 case INSN:
1066 case CALL_INSN:
1067 /* Ordinary insn: if this is the last use, we win. */
1068 if (regno_last_uid[regno] == INSN_UID (p))
1069 return 1;
1070 break;
1071
1072 case JUMP_INSN:
1073 /* Jump insn: if this is the last use, we win. */
1074 if (regno_last_uid[regno] == INSN_UID (p))
1075 return 1;
1076 /* Otherwise, it's the end of the basic block, so we lose. */
1077 return 0;
1078
1079 case CODE_LABEL:
1080 case BARRIER:
1081 /* It's the end of the basic block, so we lose. */
1082 return 0;
1083 }
1084 }
1085
1086 /* The "last use" doesn't follow the "first use"?? */
1087 abort ();
1088 }
1089 \f
1090 /* Compute the benefit of eliminating the insns in the block whose
1091 last insn is LAST. This may be a group of insns used to compute a
1092 value directly or can contain a library call. */
1093
1094 static int
1095 libcall_benefit (last)
1096 rtx last;
1097 {
1098 rtx insn;
1099 int benefit = 0;
1100
1101 for (insn = XEXP (find_reg_note (last, REG_RETVAL, 0), 0);
1102 insn != last; insn = NEXT_INSN (insn))
1103 {
1104 if (GET_CODE (insn) == CALL_INSN)
1105 benefit += 10; /* Assume at least this many insns in a library
1106 routine. */
1107 else if (GET_CODE (insn) == INSN
1108 && GET_CODE (PATTERN (insn)) != USE
1109 && GET_CODE (PATTERN (insn)) != CLOBBER)
1110 benefit++;
1111 }
1112
1113 return benefit;
1114 }
1115 \f
1116 /* Skip COUNT insns from INSN, counting library calls as 1 insn. */
1117
1118 static rtx
1119 skip_consec_insns (insn, count)
1120 rtx insn;
1121 int count;
1122 {
1123 for (; count > 0; count--)
1124 {
1125 rtx temp;
1126
1127 /* If first insn of libcall sequence, skip to end. */
1128 /* Do this at start of loop, since INSN is guaranteed to
1129 be an insn here. */
1130 if (GET_CODE (insn) != NOTE
1131 && (temp = find_reg_note (insn, REG_LIBCALL, 0)))
1132 insn = XEXP (temp, 0);
1133
1134 do insn = NEXT_INSN (insn);
1135 while (GET_CODE (insn) == NOTE);
1136 }
1137
1138 return insn;
1139 }
1140
1141 /* Ignore any movable whose insn falls within a libcall
1142 which is part of another movable.
1143 We make use of the fact that the movable for the libcall value
1144 was made later and so appears later on the chain. */
1145
1146 static void
1147 ignore_some_movables (movables)
1148 struct movable *movables;
1149 {
1150 register struct movable *m, *m1;
1151
1152 for (m = movables; m; m = m->next)
1153 {
1154 /* Is this a movable for the value of a libcall? */
1155 rtx note = find_reg_note (m->insn, REG_RETVAL, 0);
1156 if (note)
1157 {
1158 rtx insn;
1159 /* Check for earlier movables inside that range,
1160 and mark them invalid. We cannot use LUIDs here because
1161 insns created by loop.c for prior loops don't have LUIDs.
1162 Rather than reject all such insns from movables, we just
1163 explicitly check each insn in the libcall (since invariant
1164 libcalls aren't that common). */
1165 for (insn = XEXP (note, 0); insn != m->insn; insn = NEXT_INSN (insn))
1166 for (m1 = movables; m1 != m; m1 = m1->next)
1167 if (m1->insn == insn)
1168 m1->done = 1;
1169 }
1170 }
1171 }
1172
1173 /* For each movable insn, see if the reg that it loads
1174 leads when it dies right into another conditionally movable insn.
1175 If so, record that the second insn "forces" the first one,
1176 since the second can be moved only if the first is. */
1177
1178 static void
1179 force_movables (movables)
1180 struct movable *movables;
1181 {
1182 register struct movable *m, *m1;
1183 for (m1 = movables; m1; m1 = m1->next)
1184 /* Omit this if moving just the (SET (REG) 0) of a zero-extend. */
1185 if (!m1->partial && !m1->done)
1186 {
1187 int regno = m1->regno;
1188 for (m = m1->next; m; m = m->next)
1189 /* ??? Could this be a bug? What if CSE caused the
1190 register of M1 to be used after this insn?
1191 Since CSE does not update regno_last_uid,
1192 this insn M->insn might not be where it dies.
1193 But very likely this doesn't matter; what matters is
1194 that M's reg is computed from M1's reg. */
1195 if (INSN_UID (m->insn) == regno_last_uid[regno]
1196 && !m->done)
1197 break;
1198 if (m != 0 && m->set_src == m1->set_dest
1199 /* If m->consec, m->set_src isn't valid. */
1200 && m->consec == 0)
1201 m = 0;
1202
1203 /* Increase the priority of the moving the first insn
1204 since it permits the second to be moved as well. */
1205 if (m != 0)
1206 {
1207 m->forces = m1;
1208 m1->lifetime += m->lifetime;
1209 m1->savings += m1->savings;
1210 }
1211 }
1212 }
1213 \f
1214 /* Find invariant expressions that are equal and can be combined into
1215 one register. */
1216
1217 static void
1218 combine_movables (movables, nregs)
1219 struct movable *movables;
1220 int nregs;
1221 {
1222 register struct movable *m;
1223 char *matched_regs = (char *) alloca (nregs);
1224 enum machine_mode mode;
1225
1226 /* Regs that are set more than once are not allowed to match
1227 or be matched. I'm no longer sure why not. */
1228 /* Perhaps testing m->consec_sets would be more appropriate here? */
1229
1230 for (m = movables; m; m = m->next)
1231 if (m->match == 0 && n_times_used[m->regno] == 1 && !m->partial)
1232 {
1233 register struct movable *m1;
1234 int regno = m->regno;
1235 rtx reg_note, reg_note1;
1236
1237 bzero (matched_regs, nregs);
1238 matched_regs[regno] = 1;
1239
1240 for (m1 = movables; m1; m1 = m1->next)
1241 if (m != m1 && m1->match == 0 && n_times_used[m1->regno] == 1
1242 /* A reg used outside the loop mustn't be eliminated. */
1243 && !m1->global
1244 /* A reg used for zero-extending mustn't be eliminated. */
1245 && !m1->partial
1246 && (matched_regs[m1->regno]
1247 ||
1248 (
1249 /* Can combine regs with different modes loaded from the
1250 same constant only if the modes are the same or
1251 if both are integer modes with M wider or the same
1252 width as M1. The check for integer is redundant, but
1253 safe, since the only case of differing destination
1254 modes with equal sources is when both sources are
1255 VOIDmode, i.e., CONST_INT. */
1256 (GET_MODE (m->set_dest) == GET_MODE (m1->set_dest)
1257 || (GET_MODE_CLASS (GET_MODE (m->set_dest)) == MODE_INT
1258 && GET_MODE_CLASS (GET_MODE (m1->set_dest)) == MODE_INT
1259 && (GET_MODE_BITSIZE (GET_MODE (m->set_dest))
1260 >= GET_MODE_BITSIZE (GET_MODE (m1->set_dest)))))
1261 /* See if the source of M1 says it matches M. */
1262 && ((GET_CODE (m1->set_src) == REG
1263 && matched_regs[REGNO (m1->set_src)])
1264 || rtx_equal_for_loop_p (m->set_src, m1->set_src,
1265 movables))))
1266 && ((m->dependencies == m1->dependencies)
1267 || rtx_equal_p (m->dependencies, m1->dependencies)))
1268 {
1269 m->lifetime += m1->lifetime;
1270 m->savings += m1->savings;
1271 m1->done = 1;
1272 m1->match = m;
1273 matched_regs[m1->regno] = 1;
1274 }
1275 }
1276
1277 /* Now combine the regs used for zero-extension.
1278 This can be done for those not marked `global'
1279 provided their lives don't overlap. */
1280
1281 for (mode = GET_CLASS_NARROWEST_MODE (MODE_INT); mode != VOIDmode;
1282 mode = GET_MODE_WIDER_MODE (mode))
1283 {
1284 register struct movable *m0 = 0;
1285
1286 /* Combine all the registers for extension from mode MODE.
1287 Don't combine any that are used outside this loop. */
1288 for (m = movables; m; m = m->next)
1289 if (m->partial && ! m->global
1290 && mode == GET_MODE (SET_SRC (PATTERN (NEXT_INSN (m->insn)))))
1291 {
1292 register struct movable *m1;
1293 int first = uid_luid[regno_first_uid[m->regno]];
1294 int last = uid_luid[regno_last_uid[m->regno]];
1295
1296 if (m0 == 0)
1297 {
1298 /* First one: don't check for overlap, just record it. */
1299 m0 = m;
1300 continue;
1301 }
1302
1303 /* Make sure they extend to the same mode.
1304 (Almost always true.) */
1305 if (GET_MODE (m->set_dest) != GET_MODE (m0->set_dest))
1306 continue;
1307
1308 /* We already have one: check for overlap with those
1309 already combined together. */
1310 for (m1 = movables; m1 != m; m1 = m1->next)
1311 if (m1 == m0 || (m1->partial && m1->match == m0))
1312 if (! (uid_luid[regno_first_uid[m1->regno]] > last
1313 || uid_luid[regno_last_uid[m1->regno]] < first))
1314 goto overlap;
1315
1316 /* No overlap: we can combine this with the others. */
1317 m0->lifetime += m->lifetime;
1318 m0->savings += m->savings;
1319 m->done = 1;
1320 m->match = m0;
1321
1322 overlap: ;
1323 }
1324 }
1325 }
1326 \f
1327 /* Return 1 if regs X and Y will become the same if moved. */
1328
1329 static int
1330 regs_match_p (x, y, movables)
1331 rtx x, y;
1332 struct movable *movables;
1333 {
1334 int xn = REGNO (x);
1335 int yn = REGNO (y);
1336 struct movable *mx, *my;
1337
1338 for (mx = movables; mx; mx = mx->next)
1339 if (mx->regno == xn)
1340 break;
1341
1342 for (my = movables; my; my = my->next)
1343 if (my->regno == yn)
1344 break;
1345
1346 return (mx && my
1347 && ((mx->match == my->match && mx->match != 0)
1348 || mx->match == my
1349 || mx == my->match));
1350 }
1351
1352 /* Return 1 if X and Y are identical-looking rtx's.
1353 This is the Lisp function EQUAL for rtx arguments.
1354
1355 If two registers are matching movables or a movable register and an
1356 equivalent constant, consider them equal. */
1357
1358 static int
1359 rtx_equal_for_loop_p (x, y, movables)
1360 rtx x, y;
1361 struct movable *movables;
1362 {
1363 register int i;
1364 register int j;
1365 register struct movable *m;
1366 register enum rtx_code code;
1367 register char *fmt;
1368
1369 if (x == y)
1370 return 1;
1371 if (x == 0 || y == 0)
1372 return 0;
1373
1374 code = GET_CODE (x);
1375
1376 /* If we have a register and a constant, they may sometimes be
1377 equal. */
1378 if (GET_CODE (x) == REG && n_times_set[REGNO (x)] == -2
1379 && CONSTANT_P (y))
1380 for (m = movables; m; m = m->next)
1381 if (m->move_insn && m->regno == REGNO (x)
1382 && rtx_equal_p (m->set_src, y))
1383 return 1;
1384
1385 else if (GET_CODE (y) == REG && n_times_set[REGNO (y)] == -2
1386 && CONSTANT_P (x))
1387 for (m = movables; m; m = m->next)
1388 if (m->move_insn && m->regno == REGNO (y)
1389 && rtx_equal_p (m->set_src, x))
1390 return 1;
1391
1392 /* Otherwise, rtx's of different codes cannot be equal. */
1393 if (code != GET_CODE (y))
1394 return 0;
1395
1396 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent.
1397 (REG:SI x) and (REG:HI x) are NOT equivalent. */
1398
1399 if (GET_MODE (x) != GET_MODE (y))
1400 return 0;
1401
1402 /* These three types of rtx's can be compared nonrecursively. */
1403 if (code == REG)
1404 return (REGNO (x) == REGNO (y) || regs_match_p (x, y, movables));
1405
1406 if (code == LABEL_REF)
1407 return XEXP (x, 0) == XEXP (y, 0);
1408 if (code == SYMBOL_REF)
1409 return XSTR (x, 0) == XSTR (y, 0);
1410
1411 /* Compare the elements. If any pair of corresponding elements
1412 fail to match, return 0 for the whole things. */
1413
1414 fmt = GET_RTX_FORMAT (code);
1415 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1416 {
1417 switch (fmt[i])
1418 {
1419 case 'i':
1420 if (XINT (x, i) != XINT (y, i))
1421 return 0;
1422 break;
1423
1424 case 'E':
1425 /* Two vectors must have the same length. */
1426 if (XVECLEN (x, i) != XVECLEN (y, i))
1427 return 0;
1428
1429 /* And the corresponding elements must match. */
1430 for (j = 0; j < XVECLEN (x, i); j++)
1431 if (rtx_equal_for_loop_p (XVECEXP (x, i, j), XVECEXP (y, i, j), movables) == 0)
1432 return 0;
1433 break;
1434
1435 case 'e':
1436 if (rtx_equal_for_loop_p (XEXP (x, i), XEXP (y, i), movables) == 0)
1437 return 0;
1438 break;
1439
1440 case 's':
1441 if (strcmp (XSTR (x, i), XSTR (y, i)))
1442 return 0;
1443 break;
1444
1445 case 'u':
1446 /* These are just backpointers, so they don't matter. */
1447 break;
1448
1449 case '0':
1450 break;
1451
1452 /* It is believed that rtx's at this level will never
1453 contain anything but integers and other rtx's,
1454 except for within LABEL_REFs and SYMBOL_REFs. */
1455 default:
1456 abort ();
1457 }
1458 }
1459 return 1;
1460 }
1461 \f
1462 /* Scan MOVABLES, and move the insns that deserve to be moved.
1463 If two matching movables are combined, replace one reg with the
1464 other throughout. */
1465
1466 static void
1467 move_movables (movables, threshold, insn_count, loop_start, end, nregs)
1468 struct movable *movables;
1469 int threshold;
1470 int insn_count;
1471 rtx loop_start;
1472 rtx end;
1473 int nregs;
1474 {
1475 rtx new_start = 0;
1476 register struct movable *m;
1477 register rtx p;
1478 /* Map of pseudo-register replacements to handle combining
1479 when we move several insns that load the same value
1480 into different pseudo-registers. */
1481 rtx *reg_map = (rtx *) alloca (nregs * sizeof (rtx));
1482 char *already_moved = (char *) alloca (nregs);
1483
1484 bzero (already_moved, nregs);
1485 bzero (reg_map, nregs * sizeof (rtx));
1486
1487 num_movables = 0;
1488
1489 for (m = movables; m; m = m->next)
1490 {
1491 /* Describe this movable insn. */
1492
1493 if (loop_dump_stream)
1494 {
1495 fprintf (loop_dump_stream, "Insn %d: regno %d (life %d), ",
1496 INSN_UID (m->insn), m->regno, m->lifetime);
1497 if (m->consec > 0)
1498 fprintf (loop_dump_stream, "consec %d, ", m->consec);
1499 if (m->cond)
1500 fprintf (loop_dump_stream, "cond ");
1501 if (m->force)
1502 fprintf (loop_dump_stream, "force ");
1503 if (m->global)
1504 fprintf (loop_dump_stream, "global ");
1505 if (m->done)
1506 fprintf (loop_dump_stream, "done ");
1507 if (m->move_insn)
1508 fprintf (loop_dump_stream, "move-insn ");
1509 if (m->match)
1510 fprintf (loop_dump_stream, "matches %d ",
1511 INSN_UID (m->match->insn));
1512 if (m->forces)
1513 fprintf (loop_dump_stream, "forces %d ",
1514 INSN_UID (m->forces->insn));
1515 }
1516
1517 /* Count movables. Value used in heuristics in strength_reduce. */
1518 num_movables++;
1519
1520 /* Ignore the insn if it's already done (it matched something else).
1521 Otherwise, see if it is now safe to move. */
1522
1523 if (!m->done
1524 && (! m->cond
1525 || (1 == invariant_p (m->set_src)
1526 && (m->dependencies == 0
1527 || 1 == invariant_p (m->dependencies))
1528 && (m->consec == 0
1529 || 1 == consec_sets_invariant_p (m->set_dest,
1530 m->consec + 1,
1531 m->insn))))
1532 && (! m->forces || m->forces->done))
1533 {
1534 register int regno;
1535 register rtx p;
1536 int savings = m->savings;
1537
1538 /* We have an insn that is safe to move.
1539 Compute its desirability. */
1540
1541 p = m->insn;
1542 regno = m->regno;
1543
1544 if (loop_dump_stream)
1545 fprintf (loop_dump_stream, "savings %d ", savings);
1546
1547 if (moved_once[regno])
1548 {
1549 insn_count *= 2;
1550
1551 if (loop_dump_stream)
1552 fprintf (loop_dump_stream, "halved since already moved ");
1553 }
1554
1555 /* An insn MUST be moved if we already moved something else
1556 which is safe only if this one is moved too: that is,
1557 if already_moved[REGNO] is nonzero. */
1558
1559 /* An insn is desirable to move if the new lifetime of the
1560 register is no more than THRESHOLD times the old lifetime.
1561 If it's not desirable, it means the loop is so big
1562 that moving won't speed things up much,
1563 and it is liable to make register usage worse. */
1564
1565 /* It is also desirable to move if it can be moved at no
1566 extra cost because something else was already moved. */
1567
1568 if (already_moved[regno]
1569 || (threshold * savings * m->lifetime) >= insn_count
1570 || (m->forces && m->forces->done
1571 && n_times_used[m->forces->regno] == 1))
1572 {
1573 int count;
1574 register struct movable *m1;
1575 rtx first;
1576
1577 /* Now move the insns that set the reg. */
1578
1579 if (m->partial && m->match)
1580 {
1581 rtx newpat, i1;
1582 rtx r1, r2;
1583 /* Find the end of this chain of matching regs.
1584 Thus, we load each reg in the chain from that one reg.
1585 And that reg is loaded with 0 directly,
1586 since it has ->match == 0. */
1587 for (m1 = m; m1->match; m1 = m1->match);
1588 newpat = gen_move_insn (SET_DEST (PATTERN (m->insn)),
1589 SET_DEST (PATTERN (m1->insn)));
1590 i1 = emit_insn_before (newpat, loop_start);
1591
1592 /* Mark the moved, invariant reg as being allowed to
1593 share a hard reg with the other matching invariant. */
1594 REG_NOTES (i1) = REG_NOTES (m->insn);
1595 r1 = SET_DEST (PATTERN (m->insn));
1596 r2 = SET_DEST (PATTERN (m1->insn));
1597 regs_may_share = gen_rtx (EXPR_LIST, VOIDmode, r1,
1598 gen_rtx (EXPR_LIST, VOIDmode, r2,
1599 regs_may_share));
1600 delete_insn (m->insn);
1601
1602 if (new_start == 0)
1603 new_start = i1;
1604
1605 if (loop_dump_stream)
1606 fprintf (loop_dump_stream, " moved to %d", INSN_UID (i1));
1607 }
1608 /* If we are to re-generate the item being moved with a
1609 new move insn, first delete what we have and then emit
1610 the move insn before the loop. */
1611 else if (m->move_insn)
1612 {
1613 rtx i1, temp;
1614
1615 for (count = m->consec; count >= 0; count--)
1616 {
1617 /* If this is the first insn of a library call sequence,
1618 skip to the end. */
1619 if (GET_CODE (p) != NOTE
1620 && (temp = find_reg_note (p, REG_LIBCALL, 0)))
1621 p = XEXP (temp, 0);
1622
1623 /* If this is the last insn of a libcall sequence, then
1624 delete every insn in the sequence except the last.
1625 The last insn is handled in the normal manner. */
1626 if (GET_CODE (p) != NOTE
1627 && (temp = find_reg_note (p, REG_RETVAL, 0)))
1628 {
1629 temp = XEXP (temp, 0);
1630 while (temp != p)
1631 temp = delete_insn (temp);
1632 }
1633
1634 p = delete_insn (p);
1635 }
1636
1637 start_sequence ();
1638 emit_move_insn (m->set_dest, m->set_src);
1639 temp = gen_sequence ();
1640 end_sequence ();
1641
1642 i1 = emit_insn_before (temp, loop_start);
1643 if (! find_reg_note (i1, REG_EQUAL, 0))
1644 REG_NOTES (i1)
1645 = gen_rtx (EXPR_LIST,
1646 m->is_equiv ? REG_EQUIV : REG_EQUAL,
1647 m->set_src, REG_NOTES (i1));
1648
1649 if (loop_dump_stream)
1650 fprintf (loop_dump_stream, " moved to %d", INSN_UID (i1));
1651
1652 /* The more regs we move, the less we like moving them. */
1653 threshold -= 3;
1654 }
1655 else
1656 {
1657 for (count = m->consec; count >= 0; count--)
1658 {
1659 rtx i1, temp;
1660
1661 /* If first insn of libcall sequence, skip to end. */
1662 /* Do this at start of loop, since p is guaranteed to
1663 be an insn here. */
1664 if (GET_CODE (p) != NOTE
1665 && (temp = find_reg_note (p, REG_LIBCALL, 0)))
1666 p = XEXP (temp, 0);
1667
1668 /* If last insn of libcall sequence, move all
1669 insns except the last before the loop. The last
1670 insn is handled in the normal manner. */
1671 if (GET_CODE (p) != NOTE
1672 && (temp = find_reg_note (p, REG_RETVAL, 0)))
1673 {
1674 rtx fn_address = 0;
1675 rtx fn_reg = 0;
1676 rtx fn_address_insn = 0;
1677
1678 first = 0;
1679 for (temp = XEXP (temp, 0); temp != p;
1680 temp = NEXT_INSN (temp))
1681 {
1682 rtx body;
1683 rtx n;
1684 rtx next;
1685
1686 if (GET_CODE (temp) == NOTE)
1687 continue;
1688
1689 body = PATTERN (temp);
1690
1691 /* Find the next insn after TEMP,
1692 not counting USE or NOTE insns. */
1693 for (next = NEXT_INSN (temp); next != p;
1694 next = NEXT_INSN (next))
1695 if (! (GET_CODE (next) == INSN
1696 && GET_CODE (PATTERN (next)) == USE)
1697 && GET_CODE (next) != NOTE)
1698 break;
1699
1700 /* If that is the call, this may be the insn
1701 that loads the function address.
1702
1703 Extract the function address from the insn
1704 that loads it into a register.
1705 If this insn was cse'd, we get incorrect code.
1706
1707 So emit a new move insn that copies the
1708 function address into the register that the
1709 call insn will use. flow.c will delete any
1710 redundant stores that we have created. */
1711 if (GET_CODE (next) == CALL_INSN
1712 && GET_CODE (body) == SET
1713 && GET_CODE (SET_DEST (body)) == REG
1714 && (n = find_reg_note (temp, REG_EQUAL, 0)))
1715 {
1716 fn_reg = SET_SRC (body);
1717 if (GET_CODE (fn_reg) != REG)
1718 fn_reg = SET_DEST (body);
1719 fn_address = XEXP (n, 0);
1720 fn_address_insn = temp;
1721 }
1722 /* We have the call insn.
1723 If it uses the register we suspect it might,
1724 load it with the correct address directly. */
1725 if (GET_CODE (temp) == CALL_INSN
1726 && fn_address != 0
1727 && reg_mentioned_p (fn_reg, body))
1728 emit_insn_after (gen_move_insn (fn_reg,
1729 fn_address),
1730 fn_address_insn);
1731
1732 if (GET_CODE (temp) == CALL_INSN)
1733 i1 = emit_call_insn_before (body, loop_start);
1734 else
1735 i1 = emit_insn_before (body, loop_start);
1736 if (first == 0)
1737 first = i1;
1738 if (temp == fn_address_insn)
1739 fn_address_insn = i1;
1740 REG_NOTES (i1) = REG_NOTES (temp);
1741 delete_insn (temp);
1742 }
1743 }
1744 if (m->savemode != VOIDmode)
1745 {
1746 /* P sets REG to zero; but we should clear only
1747 the bits that are not covered by the mode
1748 m->savemode. */
1749 rtx reg = m->set_dest;
1750 rtx sequence;
1751 rtx tem;
1752
1753 start_sequence ();
1754 tem = expand_binop
1755 (GET_MODE (reg), and_optab, reg,
1756 gen_rtx (CONST_INT, VOIDmode,
1757 ((1 << GET_MODE_BITSIZE (m->savemode)))
1758 - 1),
1759 reg, 1, OPTAB_LIB_WIDEN);
1760 if (tem == 0)
1761 abort ();
1762 if (tem != reg)
1763 emit_move_insn (reg, tem);
1764 sequence = gen_sequence ();
1765 end_sequence ();
1766 i1 = emit_insn_before (sequence, loop_start);
1767 }
1768 else if (GET_CODE (p) == CALL_INSN)
1769 i1 = emit_call_insn_before (PATTERN (p), loop_start);
1770 else
1771 i1 = emit_insn_before (PATTERN (p), loop_start);
1772
1773 REG_NOTES (i1) = REG_NOTES (p);
1774
1775 if (new_start == 0)
1776 new_start = i1;
1777
1778 if (loop_dump_stream)
1779 fprintf (loop_dump_stream, " moved to %d",
1780 INSN_UID (i1));
1781
1782 #if 0
1783 /* This isn't needed because REG_NOTES is copied
1784 below and is wrong since P might be a PARALLEL. */
1785 if (REG_NOTES (i1) == 0
1786 && ! m->partial /* But not if it's a zero-extend clr. */
1787 && ! m->global /* and not if used outside the loop
1788 (since it might get set outside). */
1789 && CONSTANT_P (SET_SRC (PATTERN (p))))
1790 REG_NOTES (i1)
1791 = gen_rtx (EXPR_LIST, REG_EQUAL,
1792 SET_SRC (PATTERN (p)), REG_NOTES (i1));
1793 #endif
1794
1795 /* If library call, now fix the REG_NOTES that contain
1796 insn pointers, namely REG_LIBCALL on FIRST
1797 and REG_RETVAL on I1. */
1798 if (temp = find_reg_note (i1, REG_RETVAL, 0))
1799 {
1800 XEXP (temp, 0) = first;
1801 temp = find_reg_note (first, REG_LIBCALL, 0);
1802 XEXP (temp, 0) = i1;
1803 }
1804
1805 delete_insn (p);
1806 do p = NEXT_INSN (p);
1807 while (p && GET_CODE (p) == NOTE);
1808 }
1809
1810 /* The more regs we move, the less we like moving them. */
1811 threshold -= 3;
1812 }
1813
1814 /* Any other movable that loads the same register
1815 MUST be moved. */
1816 already_moved[regno] = 1;
1817
1818 /* This reg has been moved out of one loop. */
1819 moved_once[regno] = 1;
1820
1821 /* The reg set here is now invariant. */
1822 if (! m->partial)
1823 n_times_set[regno] = 0;
1824
1825 m->done = 1;
1826
1827 /* Change the length-of-life info for the register
1828 to say it lives at least the full length of this loop.
1829 This will help guide optimizations in outer loops. */
1830
1831 if (uid_luid[regno_first_uid[regno]] > INSN_LUID (loop_start))
1832 /* This is the old insn before all the moved insns.
1833 We can't use the moved insn because it is out of range
1834 in uid_luid. Only the old insns have luids. */
1835 regno_first_uid[regno] = INSN_UID (loop_start);
1836 if (uid_luid[regno_last_uid[regno]] < INSN_LUID (end))
1837 regno_last_uid[regno] = INSN_UID (end);
1838
1839 /* Combine with this moved insn any other matching movables. */
1840
1841 if (! m->partial)
1842 for (m1 = movables; m1; m1 = m1->next)
1843 if (m1->match == m)
1844 {
1845 rtx temp;
1846
1847 /* Schedule the reg loaded by M1
1848 for replacement so that shares the reg of M.
1849 If the modes differ (only possible in restricted
1850 circumstances, make a SUBREG. */
1851 if (GET_MODE (m->set_dest) == GET_MODE (m1->set_dest))
1852 reg_map[m1->regno] = m->set_dest;
1853 else
1854 reg_map[m1->regno]
1855 = gen_lowpart_common (GET_MODE (m1->set_dest),
1856 m->set_dest);
1857
1858 /* Get rid of the matching insn
1859 and prevent further processing of it. */
1860 m1->done = 1;
1861
1862 /* if library call, delete all insn except last, which
1863 is deleted below */
1864 if (temp = find_reg_note (m1->insn, REG_RETVAL, 0))
1865 {
1866 for (temp = XEXP (temp, 0); temp != m1->insn;
1867 temp = NEXT_INSN (temp))
1868 delete_insn (temp);
1869 }
1870 delete_insn (m1->insn);
1871
1872 /* Any other movable that loads the same register
1873 MUST be moved. */
1874 already_moved[m1->regno] = 1;
1875
1876 /* The reg merged here is now invariant,
1877 if the reg it matches is invariant. */
1878 if (! m->partial)
1879 n_times_set[m1->regno] = 0;
1880 }
1881 }
1882 else if (loop_dump_stream)
1883 fprintf (loop_dump_stream, "not desirable");
1884 }
1885 else if (loop_dump_stream && !m->match)
1886 fprintf (loop_dump_stream, "not safe");
1887
1888 if (loop_dump_stream)
1889 fprintf (loop_dump_stream, "\n");
1890 }
1891
1892 if (new_start == 0)
1893 new_start = loop_start;
1894
1895 /* Go through all the instructions in the loop, making
1896 all the register substitutions scheduled in REG_MAP. */
1897 for (p = new_start; p != end; p = NEXT_INSN (p))
1898 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
1899 || GET_CODE (p) == CALL_INSN)
1900 {
1901 replace_regs (PATTERN (p), reg_map, nregs, 0);
1902 replace_regs (REG_NOTES (p), reg_map, nregs, 0);
1903 }
1904 }
1905 \f
1906 #if 0
1907 /* Scan X and replace the address of any MEM in it with ADDR.
1908 REG is the address that MEM should have before the replacement. */
1909
1910 static void
1911 replace_call_address (x, reg, addr)
1912 rtx x, reg, addr;
1913 {
1914 register enum rtx_code code;
1915 register int i;
1916 register char *fmt;
1917
1918 if (x == 0)
1919 return;
1920 code = GET_CODE (x);
1921 switch (code)
1922 {
1923 case PC:
1924 case CC0:
1925 case CONST_INT:
1926 case CONST_DOUBLE:
1927 case CONST:
1928 case SYMBOL_REF:
1929 case LABEL_REF:
1930 case REG:
1931 return;
1932
1933 case SET:
1934 /* Short cut for very common case. */
1935 replace_call_address (XEXP (x, 1), reg, addr);
1936 return;
1937
1938 case CALL:
1939 /* Short cut for very common case. */
1940 replace_call_address (XEXP (x, 0), reg, addr);
1941 return;
1942
1943 case MEM:
1944 /* If this MEM uses a reg other than the one we expected,
1945 something is wrong. */
1946 if (XEXP (x, 0) != reg)
1947 abort ();
1948 XEXP (x, 0) = addr;
1949 return;
1950 }
1951
1952 fmt = GET_RTX_FORMAT (code);
1953 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1954 {
1955 if (fmt[i] == 'e')
1956 replace_call_address (XEXP (x, i), reg, addr);
1957 if (fmt[i] == 'E')
1958 {
1959 register int j;
1960 for (j = 0; j < XVECLEN (x, i); j++)
1961 replace_call_address (XVECEXP (x, i, j), reg, addr);
1962 }
1963 }
1964 }
1965 #endif
1966 \f
1967 /* Return the number of memory refs to addresses that vary
1968 in the rtx X. */
1969
1970 static int
1971 count_nonfixed_reads (x)
1972 rtx x;
1973 {
1974 register enum rtx_code code;
1975 register int i;
1976 register char *fmt;
1977 int value;
1978
1979 if (x == 0)
1980 return 0;
1981
1982 code = GET_CODE (x);
1983 switch (code)
1984 {
1985 case PC:
1986 case CC0:
1987 case CONST_INT:
1988 case CONST_DOUBLE:
1989 case CONST:
1990 case SYMBOL_REF:
1991 case LABEL_REF:
1992 case REG:
1993 return 0;
1994
1995 case MEM:
1996 return ((invariant_p (XEXP (x, 0)) != 1)
1997 + count_nonfixed_reads (XEXP (x, 0)));
1998 }
1999
2000 value = 0;
2001 fmt = GET_RTX_FORMAT (code);
2002 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2003 {
2004 if (fmt[i] == 'e')
2005 value += count_nonfixed_reads (XEXP (x, i));
2006 if (fmt[i] == 'E')
2007 {
2008 register int j;
2009 for (j = 0; j < XVECLEN (x, i); j++)
2010 value += count_nonfixed_reads (XVECEXP (x, i, j));
2011 }
2012 }
2013 return value;
2014 }
2015
2016 \f
2017 #if 0
2018 /* P is an instruction that sets a register to the result of a ZERO_EXTEND.
2019 Replace it with an instruction to load just the low bytes
2020 if the machine supports such an instruction,
2021 and insert above LOOP_START an instruction to clear the register. */
2022
2023 static void
2024 constant_high_bytes (p, loop_start)
2025 rtx p, loop_start;
2026 {
2027 register rtx new;
2028 register int insn_code_number;
2029
2030 /* Try to change (SET (REG ...) (ZERO_EXTEND (..:B ...)))
2031 to (SET (STRICT_LOW_PART (SUBREG:B (REG...))) ...). */
2032
2033 new = gen_rtx (SET, VOIDmode,
2034 gen_rtx (STRICT_LOW_PART, VOIDmode,
2035 gen_rtx (SUBREG, GET_MODE (XEXP (SET_SRC (PATTERN (p)), 0)),
2036 SET_DEST (PATTERN (p)),
2037 0)),
2038 XEXP (SET_SRC (PATTERN (p)), 0));
2039 insn_code_number = recog (new, p);
2040
2041 if (insn_code_number)
2042 {
2043 register int i;
2044
2045 /* Clear destination register before the loop. */
2046 emit_insn_before (gen_rtx (SET, VOIDmode,
2047 SET_DEST (PATTERN (p)),
2048 const0_rtx),
2049 loop_start);
2050
2051 /* Inside the loop, just load the low part. */
2052 PATTERN (p) = new;
2053 }
2054 }
2055 #endif
2056 \f
2057 /* Scan a loop setting the variables `unknown_address_altered',
2058 `num_mem_sets', `loop_continue', loops_enclosed' and `loop_has_call'.
2059 Also, fill in the array `loop_store_mems'. */
2060
2061 static void
2062 prescan_loop (start, end)
2063 rtx start, end;
2064 {
2065 register int level = 1;
2066 register rtx insn;
2067
2068 unknown_address_altered = 0;
2069 loop_has_call = 0;
2070 loop_store_mems_idx = 0;
2071
2072 num_mem_sets = 0;
2073 loops_enclosed = 1;
2074 loop_continue = 0;
2075
2076 for (insn = NEXT_INSN (start); insn != NEXT_INSN (end);
2077 insn = NEXT_INSN (insn))
2078 {
2079 if (GET_CODE (insn) == NOTE)
2080 {
2081 if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG)
2082 {
2083 ++level;
2084 /* Count number of loops contained in this one. */
2085 loops_enclosed++;
2086 }
2087 else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_END)
2088 {
2089 --level;
2090 if (level == 0)
2091 {
2092 end = insn;
2093 break;
2094 }
2095 }
2096 else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_CONT)
2097 {
2098 if (level == 1)
2099 loop_continue = insn;
2100 }
2101 }
2102 else if (GET_CODE (insn) == CALL_INSN)
2103 {
2104 unknown_address_altered = 1;
2105 loop_has_call = 1;
2106 }
2107 else
2108 {
2109 if (GET_CODE (insn) == INSN || GET_CODE (insn) == JUMP_INSN)
2110 note_stores (PATTERN (insn), note_addr_stored);
2111 }
2112 }
2113 }
2114 \f
2115 /* Scan the function looking for loops. Record the start and end of each loop.
2116 Also mark as invalid loops any loops that contain a setjmp or are branched
2117 to from outside the loop. */
2118
2119 static void
2120 find_and_verify_loops (f)
2121 rtx f;
2122 {
2123 rtx insn;
2124 int current_loop = -1;
2125 int next_loop = -1;
2126 int loop;
2127
2128 /* If there are jumps to undefined labels,
2129 treat them as jumps out of any/all loops.
2130 This also avoids writing past end of tables when there are no loops. */
2131 uid_loop_num[0] = -1;
2132
2133 /* Find boundaries of loops, mark which loops are contained within
2134 loops, and invalidate loops that have setjmp. */
2135
2136 for (insn = f; insn; insn = NEXT_INSN (insn))
2137 {
2138 if (GET_CODE (insn) == NOTE)
2139 switch (NOTE_LINE_NUMBER (insn))
2140 {
2141 case NOTE_INSN_LOOP_BEG:
2142 loop_number_loop_starts[++next_loop] = insn;
2143 loop_number_loop_ends[next_loop] = 0;
2144 loop_outer_loop[next_loop] = current_loop;
2145 loop_invalid[next_loop] = 0;
2146 loop_number_exit_labels[next_loop] = 0;
2147 current_loop = next_loop;
2148 break;
2149
2150 case NOTE_INSN_SETJMP:
2151 /* In this case, we must invalidate our current loop and any
2152 enclosing loop. */
2153 for (loop = current_loop; loop != -1; loop = loop_outer_loop[loop])
2154 {
2155 loop_invalid[loop] = 1;
2156 if (loop_dump_stream)
2157 fprintf (loop_dump_stream,
2158 "\nLoop at %d ignored due to setjmp.\n",
2159 INSN_UID (loop_number_loop_starts[loop]));
2160 }
2161 break;
2162
2163 case NOTE_INSN_LOOP_END:
2164 if (current_loop == -1)
2165 abort ();
2166
2167 loop_number_loop_ends[current_loop] = insn;
2168 current_loop = loop_outer_loop[current_loop];
2169 break;
2170
2171 }
2172
2173 /* Note that this will mark the NOTE_INSN_LOOP_END note as being in the
2174 enclosing loop, but this doesn't matter. */
2175 uid_loop_num[INSN_UID (insn)] = current_loop;
2176 }
2177
2178 /* Now scan all JUMP_INSN's in the function. If any branches into a loop
2179 that it is not contained within, that loop is marked invalid.
2180
2181 Also look for blocks of code ending in an unconditional branch that
2182 exits the loop. If such a block is surrounded by a conditional
2183 branch around the block, move the block elsewhere (see below) and
2184 invert the jump to point to the code block. This may eliminate a
2185 label in our loop and will simplify processing by both us and a
2186 possible second cse pass. */
2187
2188 for (insn = f; insn; insn = NEXT_INSN (insn))
2189 if (GET_CODE (insn) == JUMP_INSN)
2190 {
2191 int this_loop_num = uid_loop_num[INSN_UID (insn)];
2192
2193 mark_loop_jump (PATTERN (insn), this_loop_num);
2194
2195 /* See if this is an unconditional branch outside the loop. */
2196 if (this_loop_num != -1
2197 && (GET_CODE (PATTERN (insn)) == RETURN
2198 || (simplejump_p (insn)
2199 && (uid_loop_num[INSN_UID (JUMP_LABEL (insn))]
2200 != this_loop_num))))
2201 {
2202 rtx p;
2203 rtx our_next = next_real_insn (insn);
2204
2205 /* Go backwards until we reach the start of the loop, a label,
2206 or a JUMP_INSN. */
2207 for (p = PREV_INSN (insn);
2208 GET_CODE (p) != CODE_LABEL
2209 && ! (GET_CODE (p) == NOTE
2210 && NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
2211 && GET_CODE (p) != JUMP_INSN;
2212 p = PREV_INSN (p))
2213 ;
2214
2215 /* If we stopped on a JUMP_INSN to the next insn after INSN,
2216 we have a block of code to try to move.
2217
2218 We look backward and then forward from the target of INSN
2219 to find a BARRIER at the same loop depth as the target.
2220 If we find such a BARRIER, we make a new label for the start
2221 of the block, invert the jump in P and point it to that label,
2222 and move the block of code to the spot we found. */
2223
2224 if (GET_CODE (p) == JUMP_INSN
2225 && JUMP_LABEL (p) != 0
2226 && condjump_p (p)
2227 && ! simplejump_p (p)
2228 && next_real_insn (JUMP_LABEL (p)) == our_next)
2229 {
2230 rtx target
2231 = JUMP_LABEL (insn) ? JUMP_LABEL (insn) : get_last_insn ();
2232 int target_loop_num = uid_loop_num[INSN_UID (target)];
2233 rtx loc;
2234
2235 for (loc = target; loc; loc = PREV_INSN (loc))
2236 if (GET_CODE (loc) == BARRIER
2237 && uid_loop_num[INSN_UID (loc)] == target_loop_num)
2238 break;
2239
2240 if (loc == 0)
2241 for (loc = target; loc; loc = NEXT_INSN (loc))
2242 if (GET_CODE (loc) == BARRIER
2243 && uid_loop_num[INSN_UID (loc)] == target_loop_num)
2244 break;
2245
2246 if (loc)
2247 {
2248 rtx cond_label = JUMP_LABEL (p);
2249 rtx new_label = get_label_after (p);
2250
2251 /* Ensure our label doesn't go away. */
2252 LABEL_NUSES (cond_label)++;
2253
2254 /* Verify that uid_loop_num is large enough and that
2255 we can invert P. */
2256 if (INSN_UID (new_label) < max_uid_for_loop
2257 && invert_jump (p, new_label))
2258 {
2259 rtx q, r;
2260
2261 /* Include the BARRIER after INSN and copy the
2262 block after LOC. */
2263 squeeze_notes (new_label, NEXT_INSN (insn));
2264 reorder_insns (new_label, NEXT_INSN (insn), loc);
2265
2266 /* All those insns are now in TARGET_LOOP_NUM. */
2267 for (q = new_label; q != NEXT_INSN (NEXT_INSN (insn));
2268 q = NEXT_INSN (q))
2269 uid_loop_num[INSN_UID (q)] = target_loop_num;
2270
2271 /* The label jumped to by INSN is no longer a loop exit.
2272 Unless INSN does not have a label (e.g., it is a
2273 RETURN insn), search loop_number_exit_labels to find
2274 its label_ref, and remove it. Also turn off
2275 LABEL_OUTSIDE_LOOP_P bit. */
2276 if (JUMP_LABEL (insn))
2277 {
2278 for (q = 0,
2279 r = loop_number_exit_labels[this_loop_num];
2280 r; q = r, r = LABEL_NEXTREF (r))
2281 if (XEXP (r, 0) == JUMP_LABEL (insn))
2282 {
2283 LABEL_OUTSIDE_LOOP_P (r) = 0;
2284 if (q)
2285 LABEL_NEXTREF (q) = LABEL_NEXTREF (r);
2286 else
2287 loop_number_exit_labels[this_loop_num]
2288 = LABEL_NEXTREF (r);
2289 break;
2290 }
2291
2292 /* If we didn't find it, then something is wrong. */
2293 if (! r)
2294 abort ();
2295 }
2296
2297 /* P is now a jump outside the loop, so it must be put
2298 in loop_number_exit_labels, and marked as such.
2299 The easiest way to do this is to just call
2300 mark_loop_jump again for P. */
2301 mark_loop_jump (PATTERN (p), this_loop_num);
2302
2303 /* If INSN now jumps to the insn after it,
2304 delete INSN. */
2305 if (JUMP_LABEL (insn) != 0
2306 && (next_real_insn (JUMP_LABEL (insn))
2307 == next_real_insn (insn)))
2308 delete_insn (insn);
2309 }
2310
2311 /* Continue the loop after where the conditional
2312 branch used to jump, since the only branch insn
2313 in the block (if it still remains) is an inter-loop
2314 branch and hence needs no processing. */
2315 insn = NEXT_INSN (cond_label);
2316
2317 if (--LABEL_NUSES (cond_label) == 0)
2318 delete_insn (cond_label);
2319 }
2320 }
2321 }
2322 }
2323 }
2324
2325 /* If any label in X jumps to a loop different from LOOP_NUM and any of the
2326 loops it is contained in, mark the target loop invalid.
2327
2328 For speed, we assume that X is part of a pattern of a JUMP_INSN. */
2329
2330 static void
2331 mark_loop_jump (x, loop_num)
2332 rtx x;
2333 int loop_num;
2334 {
2335 int dest_loop;
2336 int outer_loop;
2337 int i;
2338
2339 switch (GET_CODE (x))
2340 {
2341 case PC:
2342 case USE:
2343 case CLOBBER:
2344 case REG:
2345 case MEM:
2346 case CONST_INT:
2347 case CONST_DOUBLE:
2348 case RETURN:
2349 return;
2350
2351 case CONST:
2352 /* There could be a label reference in here. */
2353 mark_loop_jump (XEXP (x, 0), loop_num);
2354 return;
2355
2356 case PLUS:
2357 case MINUS:
2358 case MULT:
2359 case LSHIFT:
2360 mark_loop_jump (XEXP (x, 0), loop_num);
2361 mark_loop_jump (XEXP (x, 1), loop_num);
2362 return;
2363
2364 case SIGN_EXTEND:
2365 case ZERO_EXTEND:
2366 mark_loop_jump (XEXP (x, 0), loop_num);
2367 return;
2368
2369 case LABEL_REF:
2370 dest_loop = uid_loop_num[INSN_UID (XEXP (x, 0))];
2371
2372 /* Link together all labels that branch outside the loop. This
2373 is used by final_[bg]iv_value and the loop unrolling code. Also
2374 mark this LABEL_REF so we know that this branch should predict
2375 false. */
2376
2377 if (dest_loop != loop_num && loop_num != -1)
2378 {
2379 LABEL_OUTSIDE_LOOP_P (x) = 1;
2380 LABEL_NEXTREF (x) = loop_number_exit_labels[loop_num];
2381 loop_number_exit_labels[loop_num] = x;
2382 }
2383
2384 /* If this is inside a loop, but not in the current loop or one enclosed
2385 by it, it invalidates at least one loop. */
2386
2387 if (dest_loop == -1)
2388 return;
2389
2390 /* We must invalidate every nested loop containing the target of this
2391 label, except those that also contain the jump insn. */
2392
2393 for (; dest_loop != -1; dest_loop = loop_outer_loop[dest_loop])
2394 {
2395 /* Stop when we reach a loop that also contains the jump insn. */
2396 for (outer_loop = loop_num; outer_loop != -1;
2397 outer_loop = loop_outer_loop[outer_loop])
2398 if (dest_loop == outer_loop)
2399 return;
2400
2401 /* If we get here, we know we need to invalidate a loop. */
2402 if (loop_dump_stream && ! loop_invalid[dest_loop])
2403 fprintf (loop_dump_stream,
2404 "\nLoop at %d ignored due to multiple entry points.\n",
2405 INSN_UID (loop_number_loop_starts[dest_loop]));
2406
2407 loop_invalid[dest_loop] = 1;
2408 }
2409 return;
2410
2411 case SET:
2412 /* If this is not setting pc, ignore. */
2413 if (SET_DEST (x) == pc_rtx)
2414 mark_loop_jump (SET_SRC (x), loop_num);
2415 return;
2416
2417 case IF_THEN_ELSE:
2418 mark_loop_jump (XEXP (x, 1), loop_num);
2419 mark_loop_jump (XEXP (x, 2), loop_num);
2420 return;
2421
2422 case PARALLEL:
2423 case ADDR_VEC:
2424 for (i = 0; i < XVECLEN (x, 0); i++)
2425 mark_loop_jump (XVECEXP (x, 0, i), loop_num);
2426 return;
2427
2428 case ADDR_DIFF_VEC:
2429 for (i = 0; i < XVECLEN (x, 1); i++)
2430 mark_loop_jump (XVECEXP (x, 1, i), loop_num);
2431 return;
2432
2433 default:
2434 /* Nothing else should occur in a JUMP_INSN. */
2435 abort ();
2436 }
2437 }
2438 \f
2439 /* Return nonzero if there is a label in the range from
2440 insn INSN to and including the insn whose luid is END
2441 INSN must have an assigned luid (i.e., it must not have
2442 been previously created by loop.c). */
2443
2444 static int
2445 labels_in_range_p (insn, end)
2446 rtx insn;
2447 int end;
2448 {
2449 while (insn && INSN_LUID (insn) <= end)
2450 {
2451 if (GET_CODE (insn) == CODE_LABEL)
2452 return 1;
2453 insn = NEXT_INSN (insn);
2454 }
2455
2456 return 0;
2457 }
2458
2459 /* Record that a memory reference X is being set. */
2460
2461 static void
2462 note_addr_stored (x)
2463 rtx x;
2464 {
2465 register int i;
2466
2467 if (x == 0 || GET_CODE (x) != MEM)
2468 return;
2469
2470 /* Count number of memory writes.
2471 This affects heuristics in strength_reduce. */
2472 num_mem_sets++;
2473
2474 if (unknown_address_altered)
2475 return;
2476
2477 for (i = 0; i < loop_store_mems_idx; i++)
2478 if (rtx_equal_p (XEXP (loop_store_mems[i], 0), XEXP (x, 0))
2479 && MEM_IN_STRUCT_P (x) == MEM_IN_STRUCT_P (loop_store_mems[i]))
2480 {
2481 /* We are storing at the same address as previously noted. Save the
2482 wider reference, treating BLKmode as wider. */
2483 if (GET_MODE (x) == BLKmode
2484 || (GET_MODE_SIZE (GET_MODE (x))
2485 > GET_MODE_SIZE (GET_MODE (loop_store_mems[i]))))
2486 loop_store_mems[i] = x;
2487 break;
2488 }
2489
2490 if (i == NUM_STORES)
2491 unknown_address_altered = 1;
2492
2493 else if (i == loop_store_mems_idx)
2494 loop_store_mems[loop_store_mems_idx++] = x;
2495 }
2496 \f
2497 /* Return nonzero if the rtx X is invariant over the current loop.
2498
2499 The value is 2 if we refer to something only conditionally invariant.
2500
2501 If `unknown_address_altered' is nonzero, no memory ref is invariant.
2502 Otherwise, a memory ref is invariant if it does not conflict with
2503 anything stored in `loop_store_mems'. */
2504
2505 int
2506 invariant_p (x)
2507 register rtx x;
2508 {
2509 register int i;
2510 register enum rtx_code code;
2511 register char *fmt;
2512 int conditional = 0;
2513
2514 if (x == 0)
2515 return 1;
2516 code = GET_CODE (x);
2517 switch (code)
2518 {
2519 case CONST_INT:
2520 case CONST_DOUBLE:
2521 case SYMBOL_REF:
2522 case CONST:
2523 return 1;
2524
2525 case LABEL_REF:
2526 /* A LABEL_REF is normally invariant, however, if we are unrolling
2527 loops, and this label is inside the loop, then it isn't invariant.
2528 This is because each unrolled copy of the loop body will have
2529 a copy of this label. If this was invariant, then an insn loading
2530 the address of this label into a register might get moved outside
2531 the loop, and then each loop body would end up using the same label.
2532
2533 We don't know the loop bounds here though, so just fail for all
2534 labels. */
2535 if (flag_unroll_loops)
2536 return 0;
2537 else
2538 return 1;
2539
2540 case PC:
2541 case CC0:
2542 case UNSPEC_VOLATILE:
2543 return 0;
2544
2545 case REG:
2546 /* We used to check RTX_UNCHANGING_P (x) here, but that is invalid
2547 since the reg might be set by initialization within the loop. */
2548 if (x == frame_pointer_rtx || x == arg_pointer_rtx)
2549 return 1;
2550 if (loop_has_call
2551 && REGNO (x) < FIRST_PSEUDO_REGISTER && call_used_regs[REGNO (x)])
2552 return 0;
2553 if (n_times_set[REGNO (x)] < 0)
2554 return 2;
2555 return n_times_set[REGNO (x)] == 0;
2556
2557 case MEM:
2558 /* Read-only items (such as constants in a constant pool) are
2559 invariant if their address is. */
2560 if (RTX_UNCHANGING_P (x))
2561 break;
2562
2563 /* If we filled the table (or had a subroutine call), any location
2564 in memory could have been clobbered. */
2565 if (unknown_address_altered
2566 /* Don't mess with volatile memory references. */
2567 || MEM_VOLATILE_P (x))
2568 return 0;
2569
2570 /* See if there is any dependence between a store and this load. */
2571 for (i = loop_store_mems_idx - 1; i >= 0; i--)
2572 if (true_dependence (loop_store_mems[i], x))
2573 return 0;
2574
2575 /* It's not invalidated by a store in memory
2576 but we must still verify the address is invariant. */
2577 break;
2578
2579 case ASM_OPERANDS:
2580 /* Don't mess with insns declared volatile. */
2581 if (MEM_VOLATILE_P (x))
2582 return 0;
2583 }
2584
2585 fmt = GET_RTX_FORMAT (code);
2586 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2587 {
2588 if (fmt[i] == 'e')
2589 {
2590 int tem = invariant_p (XEXP (x, i));
2591 if (tem == 0)
2592 return 0;
2593 if (tem == 2)
2594 conditional = 1;
2595 }
2596 else if (fmt[i] == 'E')
2597 {
2598 register int j;
2599 for (j = 0; j < XVECLEN (x, i); j++)
2600 {
2601 int tem = invariant_p (XVECEXP (x, i, j));
2602 if (tem == 0)
2603 return 0;
2604 if (tem == 2)
2605 conditional = 1;
2606 }
2607
2608 }
2609 }
2610
2611 return 1 + conditional;
2612 }
2613
2614 /* Return 1 if OTHER (a mem ref) overlaps the area of memory
2615 which is SIZE bytes starting at BASE. */
2616
2617 int
2618 addr_overlap_p (other, base, size)
2619 rtx other;
2620 rtx base;
2621 int size;
2622 {
2623 int start = 0, end;
2624
2625 if (GET_CODE (base) == CONST)
2626 base = XEXP (base, 0);
2627 if (GET_CODE (base) == PLUS
2628 && GET_CODE (XEXP (base, 1)) == CONST_INT)
2629 {
2630 start = INTVAL (XEXP (base, 1));
2631 base = XEXP (base, 0);
2632 }
2633
2634 end = start + size;
2635 return refers_to_mem_p (other, base, start, end);
2636 }
2637 \f
2638 /* Return nonzero if all the insns in the loop that set REG
2639 are INSN and the immediately following insns,
2640 and if each of those insns sets REG in an invariant way
2641 (not counting uses of REG in them).
2642
2643 The value is 2 if some of these insns are only conditionally invariant.
2644
2645 We assume that INSN itself is the first set of REG
2646 and that its source is invariant. */
2647
2648 static int
2649 consec_sets_invariant_p (reg, n_sets, insn)
2650 int n_sets;
2651 rtx reg, insn;
2652 {
2653 register rtx p = insn;
2654 register int regno = REGNO (reg);
2655 rtx temp;
2656 /* Number of sets we have to insist on finding after INSN. */
2657 int count = n_sets - 1;
2658 int old = n_times_set[regno];
2659 int value = 0;
2660 int this;
2661
2662 /* If N_SETS hit the limit, we can't rely on its value. */
2663 if (n_sets == 127)
2664 return 0;
2665
2666 n_times_set[regno] = 0;
2667
2668 while (count > 0)
2669 {
2670 register enum rtx_code code;
2671 rtx set;
2672
2673 p = NEXT_INSN (p);
2674 code = GET_CODE (p);
2675
2676 /* If library call, skip to end of of it. */
2677 if (code == INSN && (temp = find_reg_note (p, REG_LIBCALL, 0)))
2678 p = XEXP (temp, 0);
2679
2680 this = 0;
2681 if (code == INSN
2682 && (set = single_set (p))
2683 && GET_CODE (SET_DEST (set)) == REG
2684 && REGNO (SET_DEST (set)) == regno)
2685 {
2686 this = invariant_p (SET_SRC (set));
2687 if (this != 0)
2688 value |= this;
2689 else if (temp = find_reg_note (p, REG_EQUAL, 0))
2690 {
2691 this = invariant_p (XEXP (temp, 0));
2692 if (this != 0)
2693 value |= this;
2694 }
2695 }
2696 if (this != 0)
2697 count--;
2698 else if (code != NOTE)
2699 {
2700 n_times_set[regno] = old;
2701 return 0;
2702 }
2703 }
2704
2705 n_times_set[regno] = old;
2706 /* If invariant_p ever returned 2, we return 2. */
2707 return 1 + (value & 2);
2708 }
2709
2710 #if 0
2711 /* I don't think this condition is sufficient to allow INSN
2712 to be moved, so we no longer test it. */
2713
2714 /* Return 1 if all insns in the basic block of INSN and following INSN
2715 that set REG are invariant according to TABLE. */
2716
2717 static int
2718 all_sets_invariant_p (reg, insn, table)
2719 rtx reg, insn;
2720 short *table;
2721 {
2722 register rtx p = insn;
2723 register int regno = REGNO (reg);
2724
2725 while (1)
2726 {
2727 register enum rtx_code code;
2728 p = NEXT_INSN (p);
2729 code = GET_CODE (p);
2730 if (code == CODE_LABEL || code == JUMP_INSN)
2731 return 1;
2732 if (code == INSN && GET_CODE (PATTERN (p)) == SET
2733 && GET_CODE (SET_DEST (PATTERN (p))) == REG
2734 && REGNO (SET_DEST (PATTERN (p))) == regno)
2735 {
2736 if (!invariant_p (SET_SRC (PATTERN (p)), table))
2737 return 0;
2738 }
2739 }
2740 }
2741 #endif /* 0 */
2742 \f
2743 /* Look at all uses (not sets) of registers in X. For each, if it is
2744 the single use, set USAGE[REGNO] to INSN; if there was a previous use in
2745 a different insn, set USAGE[REGNO] to const0_rtx. */
2746
2747 static void
2748 find_single_use_in_loop (insn, x, usage)
2749 rtx insn;
2750 rtx x;
2751 rtx *usage;
2752 {
2753 enum rtx_code code = GET_CODE (x);
2754 char *fmt = GET_RTX_FORMAT (code);
2755 int i, j;
2756
2757 if (code == REG)
2758 usage[REGNO (x)]
2759 = (usage[REGNO (x)] != 0 && usage[REGNO (x)] != insn)
2760 ? const0_rtx : insn;
2761
2762 else if (code == SET)
2763 {
2764 /* Don't count SET_DEST if it is a REG; otherwise count things
2765 in SET_DEST because if a register is partially modified, it won't
2766 show up as a potential movable so we don't care how USAGE is set
2767 for it. */
2768 if (GET_CODE (SET_DEST (x)) != REG)
2769 find_single_use_in_loop (insn, SET_DEST (x), usage);
2770 find_single_use_in_loop (insn, SET_SRC (x), usage);
2771 }
2772 else
2773 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2774 {
2775 if (fmt[i] == 'e' && XEXP (x, i) != 0)
2776 find_single_use_in_loop (insn, XEXP (x, i), usage);
2777 else if (fmt[i] == 'E')
2778 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
2779 find_single_use_in_loop (insn, XVECEXP (x, i, j), usage);
2780 }
2781 }
2782 \f
2783 /* Increment N_TIMES_SET at the index of each register
2784 that is modified by an insn between FROM and TO.
2785 If the value of an element of N_TIMES_SET becomes 127 or more,
2786 stop incrementing it, to avoid overflow.
2787
2788 Store in SINGLE_USAGE[I] the single insn in which register I is
2789 used, if it is only used once. Otherwise, it is set to 0 (for no
2790 uses) or const0_rtx for more than one use. This parameter may be zero,
2791 in which case this processing is not done.
2792
2793 Store in *COUNT_PTR the number of actual instruction
2794 in the loop. We use this to decide what is worth moving out. */
2795
2796 /* last_set[n] is nonzero iff reg n has been set in the current basic block.
2797 In that case, it is the insn that last set reg n. */
2798
2799 static void
2800 count_loop_regs_set (from, to, may_not_move, single_usage, count_ptr, nregs)
2801 register rtx from, to;
2802 char *may_not_move;
2803 rtx *single_usage;
2804 int *count_ptr;
2805 int nregs;
2806 {
2807 register rtx *last_set = (rtx *) alloca (nregs * sizeof (rtx));
2808 register rtx insn;
2809 register int count = 0;
2810 register rtx dest;
2811
2812 bzero (last_set, nregs * sizeof (rtx));
2813 for (insn = from; insn != to; insn = NEXT_INSN (insn))
2814 {
2815 if (GET_RTX_CLASS (GET_CODE (insn)) == 'i')
2816 {
2817 ++count;
2818
2819 /* If requested, record registers that have exactly one use. */
2820 if (single_usage)
2821 {
2822 find_single_use_in_loop (insn, PATTERN (insn), single_usage);
2823
2824 /* Include uses in REG_EQUAL notes. */
2825 if (REG_NOTES (insn))
2826 find_single_use_in_loop (insn, REG_NOTES (insn), single_usage);
2827 }
2828
2829 if (GET_CODE (PATTERN (insn)) == CLOBBER
2830 && GET_CODE (XEXP (PATTERN (insn), 0)) == REG)
2831 /* Don't move a reg that has an explicit clobber.
2832 We might do so sometimes, but it's not worth the pain. */
2833 may_not_move[REGNO (XEXP (PATTERN (insn), 0))] = 1;
2834
2835 if (GET_CODE (PATTERN (insn)) == SET
2836 || GET_CODE (PATTERN (insn)) == CLOBBER)
2837 {
2838 dest = SET_DEST (PATTERN (insn));
2839 while (GET_CODE (dest) == SUBREG
2840 || GET_CODE (dest) == ZERO_EXTRACT
2841 || GET_CODE (dest) == SIGN_EXTRACT
2842 || GET_CODE (dest) == STRICT_LOW_PART)
2843 dest = XEXP (dest, 0);
2844 if (GET_CODE (dest) == REG)
2845 {
2846 register int regno = REGNO (dest);
2847 /* If this is the first setting of this reg
2848 in current basic block, and it was set before,
2849 it must be set in two basic blocks, so it cannot
2850 be moved out of the loop. */
2851 if (n_times_set[regno] > 0 && last_set[regno] == 0)
2852 may_not_move[regno] = 1;
2853 /* If this is not first setting in current basic block,
2854 see if reg was used in between previous one and this.
2855 If so, neither one can be moved. */
2856 if (last_set[regno] != 0
2857 && reg_used_between_p (dest, last_set[regno], insn))
2858 may_not_move[regno] = 1;
2859 if (n_times_set[regno] < 127)
2860 ++n_times_set[regno];
2861 last_set[regno] = insn;
2862 }
2863 }
2864 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
2865 {
2866 register int i;
2867 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
2868 {
2869 register rtx x = XVECEXP (PATTERN (insn), 0, i);
2870 if (GET_CODE (x) == CLOBBER && GET_CODE (XEXP (x, 0)) == REG)
2871 /* Don't move a reg that has an explicit clobber.
2872 It's not worth the pain to try to do it correctly. */
2873 may_not_move[REGNO (XEXP (x, 0))] = 1;
2874
2875 if (GET_CODE (x) == SET || GET_CODE (x) == CLOBBER)
2876 {
2877 dest = SET_DEST (x);
2878 while (GET_CODE (dest) == SUBREG
2879 || GET_CODE (dest) == ZERO_EXTRACT
2880 || GET_CODE (dest) == SIGN_EXTRACT
2881 || GET_CODE (dest) == STRICT_LOW_PART)
2882 dest = XEXP (dest, 0);
2883 if (GET_CODE (dest) == REG)
2884 {
2885 register int regno = REGNO (dest);
2886 if (n_times_set[regno] > 0 && last_set[regno] == 0)
2887 may_not_move[regno] = 1;
2888 if (last_set[regno] != 0
2889 && reg_used_between_p (dest, last_set[regno], insn))
2890 may_not_move[regno] = 1;
2891 if (n_times_set[regno] < 127)
2892 ++n_times_set[regno];
2893 last_set[regno] = insn;
2894 }
2895 }
2896 }
2897 }
2898 }
2899 if (GET_CODE (insn) == CODE_LABEL || GET_CODE (insn) == JUMP_INSN)
2900 bzero (last_set, nregs * sizeof (rtx));
2901 }
2902 *count_ptr = count;
2903 }
2904 \f
2905 /* Given a loop that is bounded by LOOP_START and LOOP_END
2906 and that is entered at SCAN_START,
2907 return 1 if the register set in SET contained in insn INSN is used by
2908 any insn that precedes INSN in cyclic order starting
2909 from the loop entry point.
2910
2911 We don't want to use INSN_LUID here because if we restrict INSN to those
2912 that have a valid INSN_LUID, it means we cannot move an invariant out
2913 from an inner loop past two loops. */
2914
2915 static int
2916 loop_reg_used_before_p (set, insn, loop_start, scan_start, loop_end)
2917 rtx set, insn, loop_start, scan_start, loop_end;
2918 {
2919 rtx reg = SET_DEST (set);
2920 rtx p;
2921
2922 /* Scan forward checking for register usage. If we hit INSN, we
2923 are done. Otherwise, if we hit LOOP_END, wrap around to LOOP_START. */
2924 for (p = scan_start; p != insn; p = NEXT_INSN (p))
2925 {
2926 if (GET_RTX_CLASS (GET_CODE (p)) == 'i'
2927 && reg_overlap_mentioned_p (reg, PATTERN (p)))
2928 return 1;
2929
2930 if (p == loop_end)
2931 p = loop_start;
2932 }
2933
2934 return 0;
2935 }
2936 \f
2937 /* A "basic induction variable" or biv is a pseudo reg that is set
2938 (within this loop) only by incrementing or decrementing it. */
2939 /* A "general induction variable" or giv is a pseudo reg whose
2940 value is a linear function of a biv. */
2941
2942 /* Bivs are recognized by `basic_induction_var';
2943 Givs by `general_induct_var'. */
2944
2945 /* Indexed by register number, indicates whether or not register is an
2946 induction variable, and if so what type. */
2947
2948 enum iv_mode *reg_iv_type;
2949
2950 /* Indexed by register number, contains pointer to `struct induction'
2951 if register is an induction variable. This holds general info for
2952 all induction variables. */
2953
2954 struct induction **reg_iv_info;
2955
2956 /* Indexed by register number, contains pointer to `struct iv_class'
2957 if register is a basic induction variable. This holds info describing
2958 the class (a related group) of induction variables that the biv belongs
2959 to. */
2960
2961 struct iv_class **reg_biv_class;
2962
2963 /* The head of a list which links together (via the next field)
2964 every iv class for the current loop. */
2965
2966 struct iv_class *loop_iv_list;
2967
2968 /* Communication with routines called via `note_stores'. */
2969
2970 static rtx note_insn;
2971
2972 /* Dummy register to have non-zero DEST_REG for DEST_ADDR type givs. */
2973
2974 static rtx addr_placeholder;
2975
2976 /* ??? Unfinished optimizations, and possible future optimizations,
2977 for the strength reduction code. */
2978
2979 /* ??? There is one more optimization you might be interested in doing: to
2980 allocate pseudo registers for frequently-accessed memory locations.
2981 If the same memory location is referenced each time around, it might
2982 be possible to copy it into a register before and out after.
2983 This is especially useful when the memory location is a variable which
2984 is in a stack slot because somewhere its address is taken. If the
2985 loop doesn't contain a function call and the variable isn't volatile,
2986 it is safe to keep the value in a register for the duration of the
2987 loop. One tricky thing is that the copying of the value back from the
2988 register has to be done on all exits from the loop. You need to check that
2989 all the exits from the loop go to the same place. */
2990
2991 /* ??? The interaction of biv elimination, and recognition of 'constant'
2992 bivs, may cause problems. */
2993
2994 /* ??? Add heuristics so that DEST_ADDR strength reduction does not cause
2995 performance problems.
2996
2997 Perhaps don't eliminate things that can be combined with an addressing
2998 mode. Find all givs that have the same biv, mult_val, and add_val;
2999 then for each giv, check to see if its only use dies in a following
3000 memory address. If so, generate a new memory address and check to see
3001 if it is valid. If it is valid, then store the modified memory address,
3002 otherwise, mark the giv as not done so that it will get its own iv. */
3003
3004 /* ??? Could try to optimize branches when it is known that a biv is always
3005 positive. */
3006
3007 /* ??? When replace a biv in a compare insn, we should replace with closest
3008 giv so that an optimized branch can still be recognized by the combiner,
3009 e.g. the VAX acb insn. */
3010
3011 /* ??? Many of the checks involving uid_luid could be simplified if regscan
3012 was rerun in loop_optimize whenever a register was added or moved.
3013 Also, some of the optimizations could be a little less conservative. */
3014 \f
3015 /* Perform strength reduction and induction variable elimination. */
3016
3017 /* Pseudo registers created during this function will be beyond the last
3018 valid index in several tables including n_times_set and regno_last_uid.
3019 This does not cause a problem here, because the added registers cannot be
3020 givs outside of their loop, and hence will never be reconsidered.
3021 But scan_loop must check regnos to make sure they are in bounds. */
3022
3023 static void
3024 strength_reduce (scan_start, end, loop_top, insn_count,
3025 loop_start, loop_end)
3026 rtx scan_start;
3027 rtx end;
3028 rtx loop_top;
3029 int insn_count;
3030 rtx loop_start;
3031 rtx loop_end;
3032 {
3033 rtx p;
3034 rtx set;
3035 rtx inc_val;
3036 rtx mult_val;
3037 rtx dest_reg;
3038 /* This is 1 if current insn is not executed at least once for every loop
3039 iteration. */
3040 int not_every_iteration = 0;
3041 /* Temporary list pointers for traversing loop_iv_list. */
3042 struct iv_class *bl, **backbl;
3043 /* Ratio of extra register life span we can justify
3044 for saving an instruction. More if loop doesn't call subroutines
3045 since in that case saving an insn makes more difference
3046 and more registers are available. */
3047 /* ??? could set this to last value of threshold in move_movables */
3048 int threshold = (loop_has_call ? 1 : 2) * (3 + n_non_fixed_regs);
3049 /* Map of pseudo-register replacements. */
3050 rtx *reg_map;
3051 int call_seen;
3052 rtx test;
3053 rtx end_insert_before;
3054
3055 reg_iv_type = (enum iv_mode *) alloca (max_reg_before_loop
3056 * sizeof (enum iv_mode *));
3057 bzero ((char *) reg_iv_type, max_reg_before_loop * sizeof (enum iv_mode *));
3058 reg_iv_info = (struct induction **)
3059 alloca (max_reg_before_loop * sizeof (struct induction *));
3060 bzero ((char *) reg_iv_info, (max_reg_before_loop
3061 * sizeof (struct induction *)));
3062 reg_biv_class = (struct iv_class **)
3063 alloca (max_reg_before_loop * sizeof (struct iv_class *));
3064 bzero ((char *) reg_biv_class, (max_reg_before_loop
3065 * sizeof (struct iv_class *)));
3066
3067 loop_iv_list = 0;
3068 addr_placeholder = gen_reg_rtx (Pmode);
3069
3070 /* Save insn immediately after the loop_end. Insns inserted after loop_end
3071 must be put before this insn, so that they will appear in the right
3072 order (i.e. loop order). */
3073
3074 end_insert_before = NEXT_INSN (loop_end);
3075
3076 /* Scan through loop to find all possible bivs. */
3077
3078 p = scan_start;
3079 while (1)
3080 {
3081 p = NEXT_INSN (p);
3082 /* At end of a straight-in loop, we are done.
3083 At end of a loop entered at the bottom, scan the top. */
3084 if (p == scan_start)
3085 break;
3086 if (p == end)
3087 {
3088 if (loop_top != 0)
3089 p = NEXT_INSN (loop_top);
3090 else
3091 break;
3092 if (p == scan_start)
3093 break;
3094 }
3095
3096 if (GET_CODE (p) == INSN
3097 && (set = single_set (p))
3098 && GET_CODE (SET_DEST (set)) == REG)
3099 {
3100 dest_reg = SET_DEST (set);
3101 if (REGNO (dest_reg) < max_reg_before_loop
3102 && REGNO (dest_reg) >= FIRST_PSEUDO_REGISTER
3103 && reg_iv_type[REGNO (dest_reg)] != NOT_BASIC_INDUCT)
3104 {
3105 if (basic_induction_var (SET_SRC (set), dest_reg,
3106 &inc_val, &mult_val))
3107 {
3108 /* It is a possible basic induction variable.
3109 Create and initialize an induction structure for it. */
3110
3111 struct induction *v
3112 = (struct induction *) alloca (sizeof (struct induction));
3113
3114 record_biv (v, p, dest_reg, inc_val, mult_val,
3115 not_every_iteration);
3116 reg_iv_type[REGNO (dest_reg)] = BASIC_INDUCT;
3117 }
3118 else if (REGNO (dest_reg) < max_reg_before_loop)
3119 reg_iv_type[REGNO (dest_reg)] = NOT_BASIC_INDUCT;
3120 }
3121 }
3122
3123 /* Past a label or a jump, we get to insns for which we can't count
3124 on whether or how many times they will be executed during each
3125 iteration. */
3126 /* This code appears in three places, once in scan_loop, and twice
3127 in strength_reduce. */
3128 if ((GET_CODE (p) == CODE_LABEL || GET_CODE (p) == JUMP_INSN)
3129 /* If we enter the loop in the middle, and scan around to the
3130 beginning, don't set not_every_iteration for that.
3131 This can be any kind of jump, since we want to know if insns
3132 will be executed if the loop is executed. */
3133 && ! (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) == loop_top
3134 && ((NEXT_INSN (NEXT_INSN (p)) == loop_end && simplejump_p (p))
3135 || (NEXT_INSN (p) == loop_end && condjump_p (p)))))
3136 not_every_iteration = 1;
3137
3138 /* At the virtual top of a converted loop, insns are again known to
3139 be executed each iteration: logically, the loop begins here
3140 even though the exit code has been duplicated. */
3141
3142 else if (GET_CODE (p) == NOTE
3143 && NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP)
3144 not_every_iteration = 0;
3145
3146 /* Unlike in the code motion pass where MAYBE_NEVER indicates that
3147 an insn may never be executed, NOT_EVERY_ITERATION indicates whether
3148 or not an insn is known to be executed each iteration of the
3149 loop, whether or not any iterations are known to occur.
3150
3151 Therefore, if we have just passed a label and have no more labels
3152 between here and the test insn of the loop, we know these insns
3153 will be executed each iteration. This can also happen if we
3154 have just passed a jump, for example, when there are nested loops. */
3155
3156 if (not_every_iteration && GET_CODE (p) == CODE_LABEL
3157 && no_labels_between_p (p, loop_end))
3158 not_every_iteration = 0;
3159 }
3160
3161 /* Scan loop_iv_list to remove all regs that proved not to be bivs.
3162 Make a sanity check against n_times_set. */
3163 for (backbl = &loop_iv_list, bl = *backbl; bl; bl = bl->next)
3164 {
3165 if (reg_iv_type[bl->regno] != BASIC_INDUCT
3166 /* Above happens if register modified by subreg, etc. */
3167 /* Make sure it is not recognized as a basic induction var: */
3168 || n_times_set[bl->regno] != bl->biv_count
3169 /* If never incremented, it is invariant that we decided not to
3170 move. So leave it alone. */
3171 || ! bl->incremented)
3172 {
3173 if (loop_dump_stream)
3174 fprintf (loop_dump_stream, "Reg %d: biv discarded, %s\n",
3175 bl->regno,
3176 (reg_iv_type[bl->regno] != BASIC_INDUCT
3177 ? "not induction variable"
3178 : (! bl->incremented ? "never incremented"
3179 : "count error")));
3180
3181 reg_iv_type[bl->regno] = NOT_BASIC_INDUCT;
3182 *backbl = bl->next;
3183 }
3184 else
3185 {
3186 backbl = &bl->next;
3187
3188 if (loop_dump_stream)
3189 fprintf (loop_dump_stream, "Reg %d: biv verified\n", bl->regno);
3190 }
3191 }
3192
3193 /* Exit if there are no bivs. */
3194 if (! loop_iv_list)
3195 {
3196 /* Can still unroll the loop anyways, but indicate that there is no
3197 strength reduction info available. */
3198 if (flag_unroll_loops)
3199 unroll_loop (loop_end, insn_count, loop_start, end_insert_before, 0);
3200
3201 return;
3202 }
3203
3204 /* Find initial value for each biv by searching backwards from loop_start,
3205 halting at first label. Also record any test condition. */
3206
3207 call_seen = 0;
3208 for (p = loop_start; p && GET_CODE (p) != CODE_LABEL; p = PREV_INSN (p))
3209 {
3210 note_insn = p;
3211
3212 if (GET_CODE (p) == CALL_INSN)
3213 call_seen = 1;
3214
3215 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
3216 || GET_CODE (p) == CALL_INSN)
3217 note_stores (PATTERN (p), record_initial);
3218
3219 /* Record any test of a biv that branches around the loop if no store
3220 between it and the start of loop. We only care about tests with
3221 constants and registers and only certain of those. */
3222 if (GET_CODE (p) == JUMP_INSN
3223 && JUMP_LABEL (p) != 0
3224 && next_real_insn (JUMP_LABEL (p)) == next_real_insn (loop_end)
3225 && (test = get_condition_for_loop (p)) != 0
3226 && GET_CODE (XEXP (test, 0)) == REG
3227 && REGNO (XEXP (test, 0)) < max_reg_before_loop
3228 && (bl = reg_biv_class[REGNO (XEXP (test, 0))]) != 0
3229 && valid_initial_value_p (XEXP (test, 1), p, call_seen, loop_start)
3230 && bl->init_insn == 0)
3231 {
3232 /* If an NE test, we have an initial value! */
3233 if (GET_CODE (test) == NE)
3234 {
3235 bl->init_insn = p;
3236 bl->init_set = gen_rtx (SET, VOIDmode,
3237 XEXP (test, 0), XEXP (test, 1));
3238 }
3239 else
3240 bl->initial_test = test;
3241 }
3242 }
3243
3244 /* Look at the each biv and see if we can say anything better about its
3245 initial value from any initializing insns set up above. (This is done
3246 in two passes to avoid missing SETs in a PARALLEL.) */
3247 for (bl = loop_iv_list; bl; bl = bl->next)
3248 {
3249 rtx src;
3250
3251 if (! bl->init_insn)
3252 continue;
3253
3254 src = SET_SRC (bl->init_set);
3255
3256 if (loop_dump_stream)
3257 fprintf (loop_dump_stream,
3258 "Biv %d initialized at insn %d: initial value ",
3259 bl->regno, INSN_UID (bl->init_insn));
3260
3261 if (valid_initial_value_p (src, bl->init_insn, call_seen, loop_start))
3262 {
3263 bl->initial_value = src;
3264
3265 if (loop_dump_stream)
3266 {
3267 if (GET_CODE (src) == CONST_INT)
3268 fprintf (loop_dump_stream, "%d\n", INTVAL (src));
3269 else
3270 {
3271 print_rtl (loop_dump_stream, src);
3272 fprintf (loop_dump_stream, "\n");
3273 }
3274 }
3275 }
3276 else
3277 {
3278 /* Biv initial value is not simple move,
3279 so let it keep intial value of "itself". */
3280
3281 if (loop_dump_stream)
3282 fprintf (loop_dump_stream, "is complex\n");
3283 }
3284 }
3285
3286 /* Search the loop for general induction variables. */
3287
3288 /* A register is a giv if: it is only set once, it is a function of a
3289 biv and a constant (or invariant), and it is not a biv. */
3290
3291 not_every_iteration = 0;
3292 p = scan_start;
3293 while (1)
3294 {
3295 p = NEXT_INSN (p);
3296 /* At end of a straight-in loop, we are done.
3297 At end of a loop entered at the bottom, scan the top. */
3298 if (p == scan_start)
3299 break;
3300 if (p == end)
3301 {
3302 if (loop_top != 0)
3303 p = NEXT_INSN (loop_top);
3304 else
3305 break;
3306 if (p == scan_start)
3307 break;
3308 }
3309
3310 /* Look for a general induction variable in a register. */
3311 if (GET_CODE (p) == INSN
3312 && (set = single_set (p))
3313 && GET_CODE (SET_DEST (set)) == REG
3314 && ! may_not_optimize[REGNO (SET_DEST (set))])
3315 {
3316 rtx src_reg;
3317 rtx add_val;
3318 rtx mult_val;
3319 int benefit;
3320 rtx regnote = 0;
3321
3322 dest_reg = SET_DEST (set);
3323 if (REGNO (dest_reg) < FIRST_PSEUDO_REGISTER)
3324 continue;
3325
3326 if (/* SET_SRC is a giv. */
3327 ((benefit = general_induction_var (SET_SRC (set),
3328 &src_reg, &add_val,
3329 &mult_val))
3330 /* Equivalent expression is a giv. */
3331 || ((regnote = find_reg_note (p, REG_EQUAL, 0))
3332 && (benefit = general_induction_var (XEXP (regnote, 0),
3333 &src_reg,
3334 &add_val, &mult_val))))
3335 /* Don't try to handle any regs made by loop optimization.
3336 We have nothing on them in regno_first_uid, etc. */
3337 && REGNO (dest_reg) < max_reg_before_loop
3338 /* Don't recognize a BASIC_INDUCT_VAR here. */
3339 && dest_reg != src_reg
3340 /* This must be the only place where the register is set. */
3341 && (n_times_set[REGNO (dest_reg)] == 1
3342 /* or all sets must be consecutive and make a giv. */
3343 || (benefit = consec_sets_giv (benefit, p,
3344 src_reg, dest_reg,
3345 &add_val, &mult_val))))
3346 {
3347 int count;
3348 struct induction *v
3349 = (struct induction *) alloca (sizeof (struct induction));
3350 rtx temp;
3351
3352 /* If this is a library call, increase benefit. */
3353 if (find_reg_note (p, REG_RETVAL, 0))
3354 benefit += libcall_benefit (p);
3355
3356 /* Skip the consecutive insns, if there are any. */
3357 for (count = n_times_set[REGNO (dest_reg)] - 1;
3358 count > 0; count--)
3359 {
3360 /* If first insn of libcall sequence, skip to end.
3361 Do this at start of loop, since INSN is guaranteed to
3362 be an insn here. */
3363 if (GET_CODE (p) != NOTE
3364 && (temp = find_reg_note (p, REG_LIBCALL, 0)))
3365 p = XEXP (temp, 0);
3366
3367 do p = NEXT_INSN (p);
3368 while (GET_CODE (p) == NOTE);
3369 }
3370
3371 record_giv (v, p, src_reg, dest_reg, mult_val, add_val, benefit,
3372 DEST_REG, not_every_iteration, 0, loop_start,
3373 loop_end);
3374
3375 }
3376 }
3377
3378 #ifndef DONT_REDUCE_ADDR
3379 /* Look for givs which are memory addresses. */
3380 /* This resulted in worse code on a VAX 8600. I wonder if it
3381 still does. */
3382 if (GET_CODE (p) == INSN)
3383 find_mem_givs (PATTERN (p), p, not_every_iteration, loop_start,
3384 loop_end);
3385 #endif
3386
3387 /* Update the status of whether giv can derive other givs. This can
3388 change when we pass a label or an insn that updates a biv. */
3389 if (GET_CODE (p) == INSN || GET_CODE (p) == CODE_LABEL)
3390 update_giv_derive (p);
3391
3392 /* Past a label or a jump, we get to insns for which we can't count
3393 on whether or how many times they will be executed during each
3394 iteration. */
3395 /* This code appears in three places, once in scan_loop, and twice
3396 in strength_reduce. */
3397 if ((GET_CODE (p) == CODE_LABEL || GET_CODE (p) == JUMP_INSN)
3398 /* If we enter the loop in the middle, and scan around
3399 to the beginning, don't set not_every_iteration for that.
3400 This can be any kind of jump, since we want to know if insns
3401 will be executed if the loop is executed. */
3402 && ! (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) == loop_top
3403 && ((NEXT_INSN (NEXT_INSN (p)) == loop_end && simplejump_p (p))
3404 || (NEXT_INSN (p) == loop_end && condjump_p (p)))))
3405 not_every_iteration = 1;
3406
3407 /* At the virtual top of a converted loop, insns are again known to
3408 be executed each iteration: logically, the loop begins here
3409 even though the exit code has been duplicated. */
3410
3411 else if (GET_CODE (p) == NOTE
3412 && NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP)
3413 not_every_iteration = 0;
3414
3415 /* Unlike in the code motion pass where MAYBE_NEVER indicates that
3416 an insn may never be executed, NOT_EVERY_ITERATION indicates whether
3417 or not an insn is known to be executed each iteration of the
3418 loop, whether or not any iterations are known to occur.
3419
3420 Therefore, if we have just passed a label and have no more labels
3421 between here and the test insn of the loop, we know these insns
3422 will be executed each iteration. */
3423
3424 if (not_every_iteration && GET_CODE (p) == CODE_LABEL
3425 && no_labels_between_p (p, loop_end))
3426 not_every_iteration = 0;
3427 }
3428
3429 /* Try to calculate and save the number of loop iterations. This is
3430 set to zero if the actual number can not be calculated. This must
3431 be called after all giv's have been identified, since otherwise it may
3432 fail if the iteration variable is a giv. */
3433
3434 loop_n_iterations = loop_iterations (loop_start, loop_end);
3435
3436 /* Now for each giv for which we still don't know whether or not it is
3437 replaceable, check to see if it is replaceable because its final value
3438 can be calculated. This must be done after loop_iterations is called,
3439 so that final_giv_value will work correctly. */
3440
3441 for (bl = loop_iv_list; bl; bl = bl->next)
3442 {
3443 struct induction *v;
3444
3445 for (v = bl->giv; v; v = v->next_iv)
3446 if (! v->replaceable && ! v->not_replaceable)
3447 check_final_value (v, loop_start, loop_end);
3448 }
3449
3450 /* Try to prove that the loop counter variable (if any) is always
3451 nonnegative; if so, record that fact with a REG_NONNEG note
3452 so that "decrement and branch until zero" insn can be used. */
3453 check_dbra_loop (loop_end, insn_count, loop_start);
3454
3455 /* Create reg_map to hold substitutions for replaceable giv regs. */
3456 reg_map = (rtx *) alloca (max_reg_before_loop * sizeof (rtx));
3457 bzero ((char *) reg_map, max_reg_before_loop * sizeof (rtx));
3458
3459 /* Examine each iv class for feasibility of strength reduction/induction
3460 variable elimination. */
3461
3462 for (bl = loop_iv_list; bl; bl = bl->next)
3463 {
3464 struct induction *v;
3465 int benefit;
3466 int all_reduced;
3467 rtx final_value = 0;
3468
3469 /* Test whether it will be possible to eliminate this biv
3470 provided all givs are reduced. This is possible if either
3471 the reg is not used outside the loop, or we can compute
3472 what its final value will be.
3473
3474 For architectures with a decrement_and_branch_until_zero insn,
3475 don't do this if we put a REG_NONNEG note on the endtest for
3476 this biv. */
3477
3478 /* Compare against bl->init_insn rather than loop_start.
3479 We aren't concerned with any uses of the biv between
3480 init_insn and loop_start since these won't be affected
3481 by the value of the biv elsewhere in the function, so
3482 long as init_insn doesn't use the biv itself.
3483 March 14, 1989 -- self@bayes.arc.nasa.gov */
3484
3485 if ((uid_luid[regno_last_uid[bl->regno]] < INSN_LUID (loop_end)
3486 && bl->init_insn
3487 && INSN_UID (bl->init_insn) < max_uid_for_loop
3488 && uid_luid[regno_first_uid[bl->regno]] >= INSN_LUID (bl->init_insn)
3489 #ifdef HAVE_decrement_and_branch_until_zero
3490 && ! bl->nonneg
3491 #endif
3492 && ! reg_mentioned_p (bl->biv->dest_reg, SET_SRC (bl->init_set)))
3493 || ((final_value = final_biv_value (bl, loop_start, loop_end))
3494 #ifdef HAVE_decrement_and_branch_until_zero
3495 && ! bl->nonneg
3496 #endif
3497 ))
3498 bl->eliminable = maybe_eliminate_biv (bl, loop_start, end, 0,
3499 threshold, insn_count);
3500 else
3501 {
3502 if (loop_dump_stream)
3503 {
3504 fprintf (loop_dump_stream,
3505 "Cannot eliminate biv %d.\n",
3506 bl->regno);
3507 fprintf (loop_dump_stream,
3508 "First use: insn %d, last use: insn %d.\n",
3509 regno_first_uid[bl->regno],
3510 regno_last_uid[bl->regno]);
3511 }
3512 }
3513
3514 /* Combine all giv's for this iv_class. */
3515 combine_givs (bl);
3516
3517 /* This will be true at the end, if all givs which depend on this
3518 biv have been strength reduced.
3519 We can't (currently) eliminate the biv unless this is so. */
3520 all_reduced = 1;
3521
3522 /* Check each giv in this class to see if we will benefit by reducing
3523 it. Skip giv's combined with others. */
3524 for (v = bl->giv; v; v = v->next_iv)
3525 {
3526 struct induction *tv;
3527
3528 if (v->ignore || v->same)
3529 continue;
3530
3531 benefit = v->benefit;
3532
3533 /* Reduce benefit if not replaceable, since we will insert
3534 a move-insn to replace the insn that calculates this giv.
3535 Don't do this unless the giv is a user variable, since it
3536 will often be marked non-replaceable because of the duplication
3537 of the exit code outside the loop. In such a case, the copies
3538 we insert are dead and will be deleted. So they don't have
3539 a cost. Similar situations exist. */
3540 /* ??? The new final_[bg]iv_value code does a much better job
3541 of finding replaceable giv's, and hence this code may no longer
3542 be necessary. */
3543 if (! v->replaceable && ! bl->eliminable
3544 && REG_USERVAR_P (v->dest_reg))
3545 benefit -= copy_cost;
3546
3547 /* Decrease the benefit to count the add-insns that we will
3548 insert to increment the reduced reg for the giv. */
3549 benefit -= add_cost * bl->biv_count;
3550
3551 /* Decide whether to strength-reduce this giv or to leave the code
3552 unchanged (recompute it from the biv each time it is used).
3553 This decision can be made independently for each giv. */
3554
3555 /* ??? Perhaps attempt to guess whether autoincrement will handle
3556 some of the new add insns; if so, can increase BENEFIT
3557 (undo the subtraction of add_cost that was done above). */
3558
3559 /* If an insn is not to be strength reduced, then set its ignore
3560 flag, and clear all_reduced. */
3561
3562 if (v->lifetime * threshold * benefit < insn_count)
3563 {
3564 if (loop_dump_stream)
3565 fprintf (loop_dump_stream,
3566 "giv of insn %d not worth while, %d vs %d.\n",
3567 INSN_UID (v->insn),
3568 v->lifetime * threshold * benefit, insn_count);
3569 v->ignore = 1;
3570 all_reduced = 0;
3571 }
3572 else
3573 {
3574 /* Check that we can increment the reduced giv without a
3575 multiply insn. If not, reject it. */
3576
3577 for (tv = bl->biv; tv; tv = tv->next_iv)
3578 if (tv->mult_val == const1_rtx
3579 && ! product_cheap_p (tv->add_val, v->mult_val))
3580 {
3581 if (loop_dump_stream)
3582 fprintf (loop_dump_stream,
3583 "giv of insn %d: would need a multiply.\n",
3584 INSN_UID (v->insn));
3585 v->ignore = 1;
3586 all_reduced = 0;
3587 break;
3588 }
3589 }
3590 }
3591
3592 /* Reduce each giv that we decided to reduce. */
3593
3594 for (v = bl->giv; v; v = v->next_iv)
3595 {
3596 struct induction *tv;
3597 if (! v->ignore && v->same == 0)
3598 {
3599 v->new_reg = gen_reg_rtx (v->mode);
3600
3601 /* For each place where the biv is incremented,
3602 add an insn to increment the new, reduced reg for the giv. */
3603 for (tv = bl->biv; tv; tv = tv->next_iv)
3604 {
3605 if (tv->mult_val == const1_rtx)
3606 emit_iv_add_mult (tv->add_val, v->mult_val,
3607 v->new_reg, v->new_reg, tv->insn);
3608 else /* tv->mult_val == const0_rtx */
3609 /* A multiply is acceptable here
3610 since this is presumed to be seldom executed. */
3611 emit_iv_add_mult (tv->add_val, v->mult_val,
3612 v->add_val, v->new_reg, tv->insn);
3613 }
3614
3615 /* Add code at loop start to initialize giv's reduced reg. */
3616
3617 emit_iv_add_mult (bl->initial_value, v->mult_val,
3618 v->add_val, v->new_reg, loop_start);
3619 }
3620 }
3621
3622 /* Rescan all givs. If a giv is the same as a giv not reduced, mark it
3623 as not reduced.
3624
3625 For each giv register that can be reduced now: if replaceable,
3626 substitute reduced reg wherever the old giv occurs;
3627 else add new move insn "giv_reg = reduced_reg".
3628
3629 Also check for givs whose first use is their definition and whose
3630 last use is the definition of another giv. If so, it is likely
3631 dead and should not be used to eliminate a biv. */
3632 for (v = bl->giv; v; v = v->next_iv)
3633 {
3634 if (v->same && v->same->ignore)
3635 v->ignore = 1;
3636
3637 if (v->ignore)
3638 continue;
3639
3640 if (v->giv_type == DEST_REG
3641 && regno_first_uid[REGNO (v->dest_reg)] == INSN_UID (v->insn))
3642 {
3643 struct induction *v1;
3644
3645 for (v1 = bl->giv; v1; v1 = v1->next_iv)
3646 if (regno_last_uid[REGNO (v->dest_reg)] == INSN_UID (v1->insn))
3647 v->maybe_dead = 1;
3648 }
3649
3650 /* Update expression if this was combined, in case other giv was
3651 replaced. */
3652 if (v->same)
3653 v->new_reg = replace_rtx (v->new_reg,
3654 v->same->dest_reg, v->same->new_reg);
3655
3656 if (v->giv_type == DEST_ADDR)
3657 /* Store reduced reg as the address in the memref where we found
3658 this giv. */
3659 *v->location = v->new_reg;
3660 else if (v->replaceable)
3661 {
3662 reg_map[REGNO (v->dest_reg)] = v->new_reg;
3663
3664 #if 0
3665 /* I can no longer duplicate the original problem. Perhaps
3666 this is unnecessary now? */
3667
3668 /* Replaceable; it isn't strictly necessary to delete the old
3669 insn and emit a new one, because v->dest_reg is now dead.
3670
3671 However, especially when unrolling loops, the special
3672 handling for (set REG0 REG1) in the second cse pass may
3673 make v->dest_reg live again. To avoid this problem, emit
3674 an insn to set the original giv reg from the reduced giv.
3675 We can not delete the original insn, since it may be part
3676 of a LIBCALL, and the code in flow that eliminates dead
3677 libcalls will fail if it is deleted. */
3678 emit_insn_after (gen_move_insn (v->dest_reg, v->new_reg),
3679 v->insn);
3680 #endif
3681 }
3682 else
3683 {
3684 /* Not replaceable; emit an insn to set the original giv reg from
3685 the reduced giv, same as above. */
3686 emit_insn_after (gen_move_insn (v->dest_reg, v->new_reg),
3687 v->insn);
3688 }
3689
3690 /* When a loop is reversed, givs which depend on the reversed
3691 biv, and which are live outside the loop, must be set to their
3692 correct final value. This insn is only needed if the giv is
3693 not replaceable. The correct final value is the same as the
3694 value that the giv starts the reversed loop with. */
3695 if (bl->reversed && ! v->replaceable)
3696 emit_iv_add_mult (bl->initial_value, v->mult_val,
3697 v->add_val, v->dest_reg, end_insert_before);
3698 else if (v->final_value)
3699 {
3700 rtx insert_before;
3701
3702 /* If the loop has multiple exits, emit the insn before the
3703 loop to ensure that it will always be executed no matter
3704 how the loop exits. Otherwise, emit the insn after the loop,
3705 since this is slightly more efficient. */
3706 if (loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]])
3707 insert_before = loop_start;
3708 else
3709 insert_before = end_insert_before;
3710 emit_insn_before (gen_move_insn (v->dest_reg, v->final_value),
3711 insert_before);
3712
3713 #if 0
3714 /* If the insn to set the final value of the giv was emitted
3715 before the loop, then we must delete the insn inside the loop
3716 that sets it. If this is a LIBCALL, then we must delete
3717 every insn in the libcall. Note, however, that
3718 final_giv_value will only succeed when there are multiple
3719 exits if the giv is dead at each exit, hence it does not
3720 matter that the original insn remains because it is dead
3721 anyways. */
3722 /* Delete the insn inside the loop that sets the giv since
3723 the giv is now set before (or after) the loop. */
3724 delete_insn (v->insn);
3725 #endif
3726 }
3727
3728 if (loop_dump_stream)
3729 {
3730 fprintf (loop_dump_stream, "giv at %d reduced to ",
3731 INSN_UID (v->insn));
3732 print_rtl (loop_dump_stream, v->new_reg);
3733 fprintf (loop_dump_stream, "\n");
3734 }
3735 }
3736
3737 /* All the givs based on the biv bl have been reduced if they
3738 merit it. */
3739
3740 /* For each giv not marked as maybe dead that has been combined with a
3741 second giv, clear any "maybe dead" mark on that second giv.
3742 v->new_reg will either be or refer to the register of the giv it
3743 combined with.
3744
3745 Doing this clearing avoids problems in biv elimination where a
3746 giv's new_reg is a complex value that can't be put in the insn but
3747 the giv combined with (with a reg as new_reg) is marked maybe_dead.
3748 Since the register will be used in either case, we'd prefer it be
3749 used from the simpler giv. */
3750
3751 for (v = bl->giv; v; v = v->next_iv)
3752 if (! v->maybe_dead && v->same)
3753 v->same->maybe_dead = 0;
3754
3755 /* Try to eliminate the biv, if it is a candidate.
3756 This won't work if ! all_reduced,
3757 since the givs we planned to use might not have been reduced.
3758
3759 We have to be careful that we didn't initially think we could elminate
3760 this biv because of a giv that we now think may be dead and shouldn't
3761 be used as a biv replacement.
3762
3763 Also, there is the possibility that we may have a giv that looks
3764 like it can be used to eliminate a biv, but the resulting insn
3765 isn't valid. This can happen, for example, on the 88k, where a
3766 JUMP_INSN can compare a register only with zero. Attempts to
3767 replace it with a comapare with a constant will fail.
3768
3769 Note that in cases where this call fails, we may have replaced some
3770 of the occurrences of the biv with a giv, but no harm was done in
3771 doing so in the rare cases where it can occur. */
3772
3773 if (all_reduced == 1 && bl->eliminable
3774 && maybe_eliminate_biv (bl, loop_start, end, 1,
3775 threshold, insn_count))
3776
3777 {
3778 /* ?? If we created a new test to bypass the loop entirely,
3779 or otherwise drop straight in, based on this test, then
3780 we might want to rewrite it also. This way some later
3781 pass has more hope of removing the initialization of this
3782 biv entirely. */
3783
3784 /* If final_value != 0, then the biv may be used after loop end
3785 and we must emit an insn to set it just in case.
3786
3787 Reversed bivs already have an insn after the loop setting their
3788 value, so we don't need another one. We can't calculate the
3789 proper final value for such a biv here anyways. */
3790 if (final_value != 0 && ! bl->reversed)
3791 {
3792 rtx insert_before;
3793
3794 /* If the loop has multiple exits, emit the insn before the
3795 loop to ensure that it will always be executed no matter
3796 how the loop exits. Otherwise, emit the insn after the
3797 loop, since this is slightly more efficient. */
3798 if (loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]])
3799 insert_before = loop_start;
3800 else
3801 insert_before = end_insert_before;
3802
3803 emit_insn_before (gen_move_insn (bl->biv->dest_reg, final_value),
3804 end_insert_before);
3805 }
3806
3807 #if 0
3808 /* Delete all of the instructions inside the loop which set
3809 the biv, as they are all dead. If is safe to delete them,
3810 because an insn setting a biv will never be part of a libcall. */
3811 /* However, deleting them will invalidate the regno_last_uid info,
3812 so keeping them around is more convenient. Final_biv_value
3813 will only succeed when there are multiple exits if the biv
3814 is dead at each exit, hence it does not matter that the original
3815 insn remains, because it is dead anyways. */
3816 for (v = bl->biv; v; v = v->next_iv)
3817 delete_insn (v->insn);
3818 #endif
3819
3820 if (loop_dump_stream)
3821 fprintf (loop_dump_stream, "Reg %d: biv eliminated\n",
3822 bl->regno);
3823 }
3824 }
3825
3826 /* Go through all the instructions in the loop, making all the
3827 register substitutions scheduled in REG_MAP. */
3828
3829 for (p = loop_start; p != end; p = NEXT_INSN (p))
3830 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
3831 || GET_CODE (p) == CALL_INSN)
3832 {
3833 replace_regs (PATTERN (p), reg_map, max_reg_before_loop, 0);
3834 replace_regs (REG_NOTES (p), reg_map, max_reg_before_loop, 0);
3835 }
3836
3837 /* Unroll loops from within strength reduction so that we can use the
3838 induction variable information that strength_reduce has already
3839 collected. */
3840
3841 if (flag_unroll_loops)
3842 unroll_loop (loop_end, insn_count, loop_start, end_insert_before, 1);
3843
3844 if (loop_dump_stream)
3845 fprintf (loop_dump_stream, "\n");
3846 }
3847 \f
3848 /* Return 1 if X is a valid source for an initial value (or as value being
3849 compared against in an initial test).
3850
3851 X must be either a register or constant and must not be clobbered between
3852 the current insn and the start of the loop.
3853
3854 INSN is the insn containing X. */
3855
3856 static int
3857 valid_initial_value_p (x, insn, call_seen, loop_start)
3858 rtx x;
3859 rtx insn;
3860 int call_seen;
3861 rtx loop_start;
3862 {
3863 if (CONSTANT_P (x))
3864 return 1;
3865
3866 /* Only consider psuedos we know about initialized in insns whose luids
3867 we know. */
3868 if (GET_CODE (x) != REG
3869 || REGNO (x) >= max_reg_before_loop)
3870 return 0;
3871
3872 /* Don't use call-clobbered registers across a call which clobbers it. On
3873 some machines, don't use any hard registers at all. */
3874 if (REGNO (x) < FIRST_PSEUDO_REGISTER
3875 #ifndef SMALL_REGISTER_CLASSES
3876 && call_used_regs[REGNO (x)] && call_seen
3877 #endif
3878 )
3879 return 0;
3880
3881 /* Don't use registers that have been clobbered before the start of the
3882 loop. */
3883 if (reg_set_between_p (x, insn, loop_start))
3884 return 0;
3885
3886 return 1;
3887 }
3888 \f
3889 /* Scan X for memory refs and check each memory address
3890 as a possible giv. INSN is the insn whose pattern X comes from.
3891 NOT_EVERY_ITERATION is 1 if the insn might not be executed during
3892 every loop iteration. */
3893
3894 static void
3895 find_mem_givs (x, insn, not_every_iteration, loop_start, loop_end)
3896 rtx x;
3897 rtx insn;
3898 int not_every_iteration;
3899 rtx loop_start, loop_end;
3900 {
3901 register int i, j;
3902 register enum rtx_code code;
3903 register char *fmt;
3904
3905 if (x == 0)
3906 return;
3907
3908 code = GET_CODE (x);
3909 switch (code)
3910 {
3911 case REG:
3912 case CONST_INT:
3913 case CONST:
3914 case CONST_DOUBLE:
3915 case SYMBOL_REF:
3916 case LABEL_REF:
3917 case PC:
3918 case CC0:
3919 case ADDR_VEC:
3920 case ADDR_DIFF_VEC:
3921 case USE:
3922 case CLOBBER:
3923 return;
3924
3925 case MEM:
3926 {
3927 rtx src_reg;
3928 rtx add_val;
3929 rtx mult_val;
3930 int benefit;
3931
3932 benefit = general_induction_var (XEXP (x, 0),
3933 &src_reg, &add_val, &mult_val);
3934
3935 /* Don't make a DEST_ADDR giv with mult_val == 1 && add_val == 0.
3936 Such a giv isn't useful. */
3937 if (benefit > 0 && (mult_val != const1_rtx || add_val != const0_rtx))
3938 {
3939 /* Found one; record it. */
3940 struct induction *v
3941 = (struct induction *) oballoc (sizeof (struct induction));
3942
3943 record_giv (v, insn, src_reg, addr_placeholder, mult_val,
3944 add_val, benefit, DEST_ADDR, not_every_iteration,
3945 &XEXP (x, 0), loop_start, loop_end);
3946
3947 v->mem_mode = GET_MODE (x);
3948 }
3949 return;
3950 }
3951 }
3952
3953 /* Recursively scan the subexpressions for other mem refs. */
3954
3955 fmt = GET_RTX_FORMAT (code);
3956 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3957 if (fmt[i] == 'e')
3958 find_mem_givs (XEXP (x, i), insn, not_every_iteration, loop_start,
3959 loop_end);
3960 else if (fmt[i] == 'E')
3961 for (j = 0; j < XVECLEN (x, i); j++)
3962 find_mem_givs (XVECEXP (x, i, j), insn, not_every_iteration,
3963 loop_start, loop_end);
3964 }
3965 \f
3966 /* Fill in the data about one biv update.
3967 V is the `struct induction' in which we record the biv. (It is
3968 allocated by the caller, with alloca.)
3969 INSN is the insn that sets it.
3970 DEST_REG is the biv's reg.
3971
3972 MULT_VAL is const1_rtx if the biv is being incremented here, in which case
3973 INC_VAL is the increment. Otherwise, MULT_VAL is const0_rtx and the biv is
3974 being set to INC_VAL. */
3975
3976 static void
3977 record_biv (v, insn, dest_reg, inc_val, mult_val, not_every_iteration)
3978 struct induction *v;
3979 rtx insn;
3980 rtx dest_reg;
3981 rtx inc_val;
3982 rtx mult_val;
3983 int not_every_iteration;
3984 {
3985 struct iv_class *bl;
3986
3987 v->insn = insn;
3988 v->src_reg = dest_reg;
3989 v->dest_reg = dest_reg;
3990 v->mult_val = mult_val;
3991 v->add_val = inc_val;
3992 v->mode = GET_MODE (dest_reg);
3993 v->always_computable = ! not_every_iteration;
3994
3995 /* Add this to the reg's iv_class, creating a class
3996 if this is the first incrementation of the reg. */
3997
3998 bl = reg_biv_class[REGNO (dest_reg)];
3999 if (bl == 0)
4000 {
4001 /* Create and initialize new iv_class. */
4002
4003 bl = (struct iv_class *) oballoc (sizeof (struct iv_class));
4004
4005 bl->regno = REGNO (dest_reg);
4006 bl->biv = 0;
4007 bl->giv = 0;
4008 bl->biv_count = 0;
4009 bl->giv_count = 0;
4010
4011 /* Set initial value to the reg itself. */
4012 bl->initial_value = dest_reg;
4013 /* We haven't seen the intializing insn yet */
4014 bl->init_insn = 0;
4015 bl->init_set = 0;
4016 bl->initial_test = 0;
4017 bl->incremented = 0;
4018 bl->eliminable = 0;
4019 bl->nonneg = 0;
4020 bl->reversed = 0;
4021
4022 /* Add this class to loop_iv_list. */
4023 bl->next = loop_iv_list;
4024 loop_iv_list = bl;
4025
4026 /* Put it in the array of biv register classes. */
4027 reg_biv_class[REGNO (dest_reg)] = bl;
4028 }
4029
4030 /* Update IV_CLASS entry for this biv. */
4031 v->next_iv = bl->biv;
4032 bl->biv = v;
4033 bl->biv_count++;
4034 if (mult_val == const1_rtx)
4035 bl->incremented = 1;
4036
4037 if (loop_dump_stream)
4038 {
4039 fprintf (loop_dump_stream,
4040 "Insn %d: possible biv, reg %d,",
4041 INSN_UID (insn), REGNO (dest_reg));
4042 if (GET_CODE (inc_val) == CONST_INT)
4043 fprintf (loop_dump_stream, " const = %d\n",
4044 INTVAL (inc_val));
4045 else
4046 {
4047 fprintf (loop_dump_stream, " const = ");
4048 print_rtl (loop_dump_stream, inc_val);
4049 fprintf (loop_dump_stream, "\n");
4050 }
4051 }
4052 }
4053 \f
4054 /* Fill in the data about one giv.
4055 V is the `struct induction' in which we record the giv. (It is
4056 allocated by the caller, with alloca.)
4057 INSN is the insn that sets it.
4058 BENEFIT estimates the savings from deleting this insn.
4059 TYPE is DEST_REG or DEST_ADDR; it says whether the giv is computed
4060 into a register or is used as a memory address.
4061
4062 SRC_REG is the biv reg which the giv is computed from.
4063 DEST_REG is the giv's reg (if the giv is stored in a reg).
4064 MULT_VAL and ADD_VAL are the coefficients used to compute the giv.
4065 LOCATION points to the place where this giv's value appears in INSN. */
4066
4067 static void
4068 record_giv (v, insn, src_reg, dest_reg, mult_val, add_val, benefit,
4069 type, not_every_iteration, location, loop_start, loop_end)
4070 struct induction *v;
4071 rtx insn;
4072 rtx src_reg;
4073 rtx dest_reg;
4074 rtx mult_val, add_val;
4075 int benefit;
4076 enum g_types type;
4077 int not_every_iteration;
4078 rtx *location;
4079 rtx loop_start, loop_end;
4080 {
4081 struct induction *b;
4082 struct iv_class *bl;
4083 rtx set = single_set (insn);
4084 rtx p;
4085
4086 v->insn = insn;
4087 v->src_reg = src_reg;
4088 v->giv_type = type;
4089 v->dest_reg = dest_reg;
4090 v->mult_val = mult_val;
4091 v->add_val = add_val;
4092 v->benefit = benefit;
4093 v->location = location;
4094 v->cant_derive = 0;
4095 v->combined_with = 0;
4096 v->maybe_dead = 0;
4097 v->derive_adjustment = 0;
4098 v->same = 0;
4099 v->ignore = 0;
4100 v->new_reg = 0;
4101 v->final_value = 0;
4102
4103 /* The v->always_computable field is used in update_giv_derive, to
4104 determine whether a giv can be used to derive another giv. For a
4105 DEST_REG giv, INSN computes a new value for the giv, so its value
4106 isn't computable if INSN insn't executed every iteration.
4107 However, for a DEST_ADDR giv, INSN merely uses the value of the giv;
4108 it does not compute a new value. Hence the value is always computable
4109 irregardless of whether INSN is executed each iteration. */
4110
4111 if (type == DEST_ADDR)
4112 v->always_computable = 1;
4113 else
4114 v->always_computable = ! not_every_iteration;
4115
4116 if (type == DEST_ADDR)
4117 {
4118 v->mode = GET_MODE (*location);
4119 v->lifetime = 1;
4120 v->times_used = 1;
4121 }
4122 else /* type == DEST_REG */
4123 {
4124 v->mode = GET_MODE (SET_DEST (set));
4125
4126 v->lifetime = (uid_luid[regno_last_uid[REGNO (dest_reg)]]
4127 - uid_luid[regno_first_uid[REGNO (dest_reg)]]);
4128
4129 v->times_used = n_times_used[REGNO (dest_reg)];
4130
4131 /* If the lifetime is zero, it means that this register is
4132 really a dead store. So mark this as a giv that can be
4133 ignored. This will not prevent the biv from being eliminated. */
4134 if (v->lifetime == 0)
4135 v->ignore = 1;
4136
4137 reg_iv_type[REGNO (dest_reg)] = GENERAL_INDUCT;
4138 reg_iv_info[REGNO (dest_reg)] = v;
4139 }
4140
4141 /* Add the giv to the class of givs computed from one biv. */
4142
4143 bl = reg_biv_class[REGNO (src_reg)];
4144 if (bl)
4145 {
4146 v->next_iv = bl->giv;
4147 bl->giv = v;
4148 /* Don't count DEST_ADDR. This is supposed to count the number of
4149 insns that calculate givs. */
4150 if (type == DEST_REG)
4151 bl->giv_count++;
4152 bl->total_benefit += benefit;
4153 }
4154 else
4155 /* Fatal error, biv missing for this giv? */
4156 abort ();
4157
4158 if (type == DEST_ADDR)
4159 v->replaceable = 1;
4160 else
4161 {
4162 /* The giv can be replaced outright by the reduced register only if all
4163 of the following conditions are true:
4164 - the insn that sets the giv is always executed on any iteration
4165 on which the giv is used at all
4166 (there are two ways to deduce this:
4167 either the insn is executed on every iteration,
4168 or all uses follow that insn in the same basic block),
4169 - the giv is not used outside the loop
4170 - no assignments to the biv occur during the giv's lifetime. */
4171
4172 if (regno_first_uid[REGNO (dest_reg)] == INSN_UID (insn)
4173 /* Previous line always fails if INSN was moved by loop opt. */
4174 && uid_luid[regno_last_uid[REGNO (dest_reg)]] < INSN_LUID (loop_end)
4175 && (! not_every_iteration
4176 || last_use_this_basic_block (dest_reg, insn)))
4177 {
4178 /* Now check that there are no assignments to the biv within the
4179 giv's lifetime. This requires two separate checks. */
4180
4181 /* Check each biv update, and fail if any are between the first
4182 and last use of the giv.
4183
4184 If this loop contains an inner loop that was unrolled, then
4185 the insn modifying the biv may have been emitted by the loop
4186 unrolling code, and hence does not have a valid luid. Just
4187 mark the biv as not replaceable in this case. It is not very
4188 useful as a biv, because it is used in two different loops.
4189 It is very unlikely that we would be able to optimize the giv
4190 using this biv anyways. */
4191
4192 v->replaceable = 1;
4193 for (b = bl->biv; b; b = b->next_iv)
4194 {
4195 if (INSN_UID (b->insn) >= max_uid_for_loop
4196 || ((uid_luid[INSN_UID (b->insn)]
4197 >= uid_luid[regno_first_uid[REGNO (dest_reg)]])
4198 && (uid_luid[INSN_UID (b->insn)]
4199 <= uid_luid[regno_last_uid[REGNO (dest_reg)]])))
4200 {
4201 v->replaceable = 0;
4202 v->not_replaceable = 1;
4203 break;
4204 }
4205 }
4206
4207 /* Check each insn between the first and last use of the giv,
4208 and fail if any of them are branches that jump to a named label
4209 outside this range, but still inside the loop. This catches
4210 cases of spaghetti code where the execution order of insns
4211 is not linear, and hence the above test fails. For example,
4212 in the following code, j is not replaceable:
4213 for (i = 0; i < 100; ) {
4214 L0: j = 4*i; goto L1;
4215 L2: k = j; goto L3;
4216 L1: i++; goto L2;
4217 L3: ; }
4218 printf ("k = %d\n", k); }
4219 This test is conservative, but this test succeeds rarely enough
4220 that it isn't a problem. See also check_final_value below. */
4221
4222 if (v->replaceable)
4223 for (p = insn;
4224 INSN_UID (p) >= max_uid_for_loop
4225 || INSN_LUID (p) < uid_luid[regno_last_uid[REGNO (dest_reg)]];
4226 p = NEXT_INSN (p))
4227 {
4228 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p)
4229 && LABEL_NAME (JUMP_LABEL (p))
4230 && ((INSN_LUID (JUMP_LABEL (p)) > INSN_LUID (loop_start)
4231 && (INSN_LUID (JUMP_LABEL (p))
4232 < uid_luid[regno_first_uid[REGNO (dest_reg)]]))
4233 || (INSN_LUID (JUMP_LABEL (p)) < INSN_LUID (loop_end)
4234 && (INSN_LUID (JUMP_LABEL (p))
4235 > uid_luid[regno_last_uid[REGNO (dest_reg)]]))))
4236 {
4237 v->replaceable = 0;
4238 v->not_replaceable = 1;
4239
4240 if (loop_dump_stream)
4241 fprintf (loop_dump_stream,
4242 "Found branch outside giv lifetime.\n");
4243
4244 break;
4245 }
4246 }
4247 }
4248 else
4249 {
4250 /* May still be replaceable, we don't have enough info here to
4251 decide. */
4252 v->replaceable = 0;
4253 v->not_replaceable = 0;
4254 }
4255 }
4256
4257 if (loop_dump_stream)
4258 {
4259 if (type == DEST_REG)
4260 fprintf (loop_dump_stream, "Insn %d: giv reg %d",
4261 INSN_UID (insn), REGNO (dest_reg));
4262 else
4263 fprintf (loop_dump_stream, "Insn %d: dest address",
4264 INSN_UID (insn));
4265
4266 fprintf (loop_dump_stream, " src reg %d benefit %d",
4267 REGNO (src_reg), v->benefit);
4268 fprintf (loop_dump_stream, " used %d lifetime %d",
4269 v->times_used, v->lifetime);
4270
4271 if (v->replaceable)
4272 fprintf (loop_dump_stream, " replaceable");
4273
4274 if (GET_CODE (mult_val) == CONST_INT)
4275 fprintf (loop_dump_stream, " mult %d",
4276 INTVAL (mult_val));
4277 else
4278 {
4279 fprintf (loop_dump_stream, " mult ");
4280 print_rtl (loop_dump_stream, mult_val);
4281 }
4282
4283 if (GET_CODE (add_val) == CONST_INT)
4284 fprintf (loop_dump_stream, " add %d",
4285 INTVAL (add_val));
4286 else
4287 {
4288 fprintf (loop_dump_stream, " add ");
4289 print_rtl (loop_dump_stream, add_val);
4290 }
4291 }
4292
4293 if (loop_dump_stream)
4294 fprintf (loop_dump_stream, "\n");
4295
4296 }
4297
4298
4299 /* All this does is determine whether a giv can be made replaceable because
4300 its final value can be calculated. This code can not be part of record_giv
4301 above, because final_giv_value requires that the number of loop iterations
4302 be known, and that can not be accurately calculated until after all givs
4303 have been identified. */
4304
4305 static void
4306 check_final_value (v, loop_start, loop_end)
4307 struct induction *v;
4308 rtx loop_start, loop_end;
4309 {
4310 struct iv_class *bl;
4311 rtx final_value = 0;
4312 rtx tem;
4313
4314 bl = reg_biv_class[REGNO (v->src_reg)];
4315
4316 /* DEST_ADDR givs will never reach here, because they are always marked
4317 replaceable above in record_giv. */
4318
4319 /* The giv can be replaced outright by the reduced register only if all
4320 of the following conditions are true:
4321 - the insn that sets the giv is always executed on any iteration
4322 on which the giv is used at all
4323 (there are two ways to deduce this:
4324 either the insn is executed on every iteration,
4325 or all uses follow that insn in the same basic block),
4326 - its final value can be calculated (this condition is different
4327 than the one above in record_giv)
4328 - no assignments to the biv occur during the giv's lifetime. */
4329
4330 #if 0
4331 /* This is only called now when replaceable is known to be false. */
4332 /* Clear replaceable, so that it won't confuse final_giv_value. */
4333 v->replaceable = 0;
4334 #endif
4335
4336 if ((final_value = final_giv_value (v, loop_start, loop_end))
4337 && (v->always_computable || last_use_this_basic_block (v->dest_reg, v->insn)))
4338 {
4339 int biv_increment_seen = 0;
4340 rtx p = v->insn;
4341 rtx last_giv_use;
4342
4343 v->replaceable = 1;
4344
4345 /* When trying to determine whether or not a biv increment occurs
4346 during the lifetime of the giv, we can ignore uses of the variable
4347 outside the loop because final_value is true. Hence we can not
4348 use regno_last_uid and regno_first_uid as above in record_giv. */
4349
4350 /* Search the loop to determine whether any assignments to the
4351 biv occur during the giv's lifetime. Start with the insn
4352 that sets the giv, and search around the loop until we come
4353 back to that insn again.
4354
4355 Also fail if there is a jump within the giv's lifetime that jumps
4356 to somewhere outside the lifetime but still within the loop. This
4357 catches spaghetti code where the execution order is not linear, and
4358 hence the above test fails. Here we assume that the giv lifetime
4359 does not extend from one iteration of the loop to the next, so as
4360 to make the test easier. Since the lifetime isn't known yet,
4361 this requires two loops. See also record_giv above. */
4362
4363 last_giv_use = v->insn;
4364
4365 while (1)
4366 {
4367 p = NEXT_INSN (p);
4368 if (p == loop_end)
4369 p = NEXT_INSN (loop_start);
4370 if (p == v->insn)
4371 break;
4372
4373 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
4374 || GET_CODE (p) == CALL_INSN)
4375 {
4376 if (biv_increment_seen)
4377 {
4378 if (reg_mentioned_p (v->dest_reg, PATTERN (p)))
4379 {
4380 v->replaceable = 0;
4381 v->not_replaceable = 1;
4382 break;
4383 }
4384 }
4385 else if (GET_CODE (PATTERN (p)) == SET
4386 && SET_DEST (PATTERN (p)) == v->src_reg)
4387 biv_increment_seen = 1;
4388 else if (reg_mentioned_p (v->dest_reg, PATTERN (p)))
4389 last_giv_use = p;
4390 }
4391 }
4392
4393 /* Now that the lifetime of the giv is known, check for branches
4394 from within the lifetime to outside the lifetime if it is still
4395 replaceable. */
4396
4397 if (v->replaceable)
4398 {
4399 p = v->insn;
4400 while (1)
4401 {
4402 p = NEXT_INSN (p);
4403 if (p == loop_end)
4404 p = NEXT_INSN (loop_start);
4405 if (p == last_giv_use)
4406 break;
4407
4408 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p)
4409 && LABEL_NAME (JUMP_LABEL (p))
4410 && ((INSN_LUID (JUMP_LABEL (p)) < INSN_LUID (v->insn)
4411 && INSN_LUID (JUMP_LABEL (p)) > INSN_LUID (loop_start))
4412 || (INSN_LUID (JUMP_LABEL (p)) > INSN_LUID (last_giv_use)
4413 && INSN_LUID (JUMP_LABEL (p)) < INSN_LUID (loop_end))))
4414 {
4415 v->replaceable = 0;
4416 v->not_replaceable = 1;
4417
4418 if (loop_dump_stream)
4419 fprintf (loop_dump_stream,
4420 "Found branch outside giv lifetime.\n");
4421
4422 break;
4423 }
4424 }
4425 }
4426
4427 /* If it is replaceable, then save the final value. */
4428 if (v->replaceable)
4429 v->final_value = final_value;
4430 }
4431
4432 if (loop_dump_stream && v->replaceable)
4433 fprintf (loop_dump_stream, "Insn %d: giv reg %d final_value replaceable\n",
4434 INSN_UID (v->insn), REGNO (v->dest_reg));
4435 }
4436 \f
4437 /* Update the status of whether a giv can derive other givs.
4438
4439 We need to do something special if there is or may be an update to the biv
4440 between the time the giv is defined and the time it is used to derive
4441 another giv.
4442
4443 In addition, a giv that is only conditionally set is not allowed to
4444 derive another giv once a label has been passed.
4445
4446 The cases we look at are when a label or an update to a biv is passed. */
4447
4448 static void
4449 update_giv_derive (p)
4450 rtx p;
4451 {
4452 struct iv_class *bl;
4453 struct induction *biv, *giv;
4454 rtx tem;
4455 int dummy;
4456
4457 /* Search all IV classes, then all bivs, and finally all givs.
4458
4459 There are two cases we are concerned with. First we have the situation
4460 of a giv that is only updated conditionally. In that case, it may not
4461 derive any givs after a label is passed.
4462
4463 The second case is when a biv update occurs, or may occur, after the
4464 definition of a giv. For certain biv updates (see below) that are
4465 known to occur between the giv definition and use, we can adjust the
4466 giv definition. For others, or when the biv update is conditional,
4467 we must prevent the giv from deriving any other givs. There are two
4468 sub-cases within this case.
4469
4470 If this is a label, we are concerned with any biv update that is done
4471 conditionally, since it may be done after the giv is defined followed by
4472 a branch here (actually, we need to pass both a jump and a label, but
4473 this extra tracking doesn't seem worth it).
4474
4475 If this is a giv update, we must adjust the giv status to show that a
4476 subsequent biv update was performed. If this adjustment cannot be done,
4477 the giv cannot derive further givs. */
4478
4479 for (bl = loop_iv_list; bl; bl = bl->next)
4480 for (biv = bl->biv; biv; biv = biv->next_iv)
4481 if (GET_CODE (p) == CODE_LABEL || biv->insn == p)
4482 {
4483 for (giv = bl->giv; giv; giv = giv->next_iv)
4484 {
4485 /* If cant_derive is already true, there is no point in
4486 checking all of these conditions again. */
4487 if (giv->cant_derive)
4488 continue;
4489
4490 /* If this giv is conditionally set and we have passed a label,
4491 it cannot derive anything. */
4492 if (GET_CODE (p) == CODE_LABEL && ! giv->always_computable)
4493 giv->cant_derive = 1;
4494
4495 /* Skip givs that have mult_val == 0, since
4496 they are really invariants. Also skip those that are
4497 replaceable, since we know their lifetime doesn't contain
4498 any biv update. */
4499 else if (giv->mult_val == const0_rtx || giv->replaceable)
4500 continue;
4501
4502 /* The only way we can allow this giv to derive another
4503 is if this is a biv increment and we can form the product
4504 of biv->add_val and giv->mult_val. In this case, we will
4505 be able to compute a compensation. */
4506 else if (biv->insn == p)
4507 {
4508 if (biv->mult_val == const1_rtx
4509 && (tem = simplify_giv_expr (gen_rtx (MULT, giv->mode,
4510 biv->add_val,
4511 giv->mult_val),
4512 &dummy)))
4513 giv->derive_adjustment = tem;
4514 else
4515 giv->cant_derive = 1;
4516 }
4517 else if (GET_CODE (p) == CODE_LABEL && ! biv->always_computable)
4518 giv->cant_derive = 1;
4519 }
4520 }
4521 }
4522 \f
4523 /* Check whether an insn is an increment legitimate for a basic induction var.
4524 X is the source of the insn.
4525 DEST_REG is the putative biv, also the destination of the insn.
4526 We accept patterns of these forms:
4527 REG = REG + INVARIANT
4528 REG = INVARIANT + REG
4529 REG = REG - CONSTANT
4530
4531 If X is suitable, we return 1, set *MULT_VAL to CONST1_RTX,
4532 and store the additive term into *INC_VAL.
4533
4534 If X is an assignment of an invariant into DEST_REG, we set
4535 *MULT_VAL to CONST0_RTX, and store the invariant into *INC_VAL.
4536
4537 Otherwise we return 0. */
4538
4539 static int
4540 basic_induction_var (x, dest_reg, inc_val, mult_val)
4541 register rtx x;
4542 rtx dest_reg;
4543 rtx *inc_val;
4544 rtx *mult_val;
4545 {
4546 register enum rtx_code code;
4547 rtx arg;
4548
4549 code = GET_CODE (x);
4550 switch (code)
4551 {
4552 case PLUS:
4553 if (XEXP (x, 0) == dest_reg)
4554 arg = XEXP (x, 1);
4555 else if (XEXP (x, 1) == dest_reg)
4556 arg = XEXP (x, 0);
4557 else
4558 return 0;
4559
4560 if (invariant_p (arg) != 1)
4561 return 0;
4562
4563 *inc_val = arg;
4564 *mult_val = const1_rtx;
4565 return 1;
4566
4567 case MINUS:
4568 if (XEXP (x, 0) == dest_reg
4569 && GET_CODE (XEXP (x, 1)) == CONST_INT)
4570 *inc_val = gen_rtx (CONST_INT, VOIDmode,
4571 - INTVAL (XEXP (x, 1)));
4572 else
4573 return 0;
4574
4575 *mult_val = const1_rtx;
4576 return 1;
4577
4578 /* Can accept constant setting of biv only when inside inner most loop.
4579 Otherwise, a biv of an inner loop may be incorrectly recognized
4580 as a biv of the outer loop,
4581 causing code to be moved INTO the inner loop. */
4582 case MEM:
4583 case REG:
4584 if (invariant_p (x) != 1)
4585 return 0;
4586 case CONST_INT:
4587 case SYMBOL_REF:
4588 case CONST:
4589 if (loops_enclosed == 1)
4590 {
4591 *inc_val = x;
4592 *mult_val = const0_rtx;
4593 return 1;
4594 }
4595 else
4596 return 0;
4597
4598 default:
4599 return 0;
4600 }
4601 }
4602 \f
4603 /* A general induction variable (giv) is any quantity that is a linear
4604 function of a basic induction variable,
4605 i.e. giv = biv * mult_val + add_val.
4606 The coefficients can be any loop invariant quantity.
4607 A giv need not be computed directly from the biv;
4608 it can be computed by way of other givs. */
4609
4610 /* Determine whether X computes a giv.
4611 If it does, return a nonzero value
4612 which is the benefit from eliminating the computation of X;
4613 set *SRC_REG to the register of the biv that it is computed from;
4614 set *ADD_VAL and *MULT_VAL to the coefficients,
4615 such that the value of X is biv * mult + add; */
4616
4617 static int
4618 general_induction_var (x, src_reg, add_val, mult_val)
4619 rtx x;
4620 rtx *src_reg;
4621 rtx *add_val;
4622 rtx *mult_val;
4623 {
4624 rtx orig_x = x;
4625 int benefit = 0;
4626 char *storage;
4627
4628 /* If this is an invariant, forget it, it isn't a giv. */
4629 if (invariant_p (x) == 1)
4630 return 0;
4631
4632 /* See if the expression could be a giv and get its form.
4633 Mark our place on the obstack in case we don't find a giv. */
4634 storage = (char *) oballoc (0);
4635 x = simplify_giv_expr (x, &benefit);
4636 if (x == 0)
4637 {
4638 obfree (storage);
4639 return 0;
4640 }
4641
4642 switch (GET_CODE (x))
4643 {
4644 case USE:
4645 case CONST_INT:
4646 /* Since this is now an invariant and wasn't before, it must be a giv
4647 with MULT_VAL == 0. It doesn't matter which BIV we associate this
4648 with. */
4649 *src_reg = loop_iv_list->biv->dest_reg;
4650 *mult_val = const0_rtx;
4651 *add_val = x;
4652 break;
4653
4654 case REG:
4655 /* This is equivalent to a BIV. */
4656 *src_reg = x;
4657 *mult_val = const1_rtx;
4658 *add_val = const0_rtx;
4659 break;
4660
4661 case PLUS:
4662 /* Either (plus (biv) (invar)) or
4663 (plus (mult (biv) (invar_1)) (invar_2)). */
4664 if (GET_CODE (XEXP (x, 0)) == MULT)
4665 {
4666 *src_reg = XEXP (XEXP (x, 0), 0);
4667 *mult_val = XEXP (XEXP (x, 0), 1);
4668 }
4669 else
4670 {
4671 *src_reg = XEXP (x, 0);
4672 *mult_val = const1_rtx;
4673 }
4674 *add_val = XEXP (x, 1);
4675 break;
4676
4677 case MULT:
4678 /* ADD_VAL is zero. */
4679 *src_reg = XEXP (x, 0);
4680 *mult_val = XEXP (x, 1);
4681 *add_val = const0_rtx;
4682 break;
4683
4684 default:
4685 abort ();
4686 }
4687
4688 /* Remove any enclosing USE from ADD_VAL and MULT_VAL (there will be
4689 unless they are CONST_INT). */
4690 if (GET_CODE (*add_val) == USE)
4691 *add_val = XEXP (*add_val, 0);
4692 if (GET_CODE (*mult_val) == USE)
4693 *mult_val = XEXP (*mult_val, 0);
4694
4695 benefit += rtx_cost (orig_x);
4696
4697 /* Always return some benefit if this is a giv so it will be detected
4698 as such. This allows elimination of bivs that might otherwise
4699 not be eliminated. */
4700 return benefit == 0 ? 1 : benefit;
4701 }
4702 \f
4703 /* Given an expression, X, try to form it as a linear function of a biv.
4704 We will canonicalize it to be of the form
4705 (plus (mult (BIV) (invar_1))
4706 (invar_2))
4707 with possibile degeneracies.
4708
4709 The invariant expressions must each be of a form that can be used as a
4710 machine operand. We surround then with a USE rtx (a hack, but localized
4711 and certainly unambiguous!) if not a CONST_INT for simplicity in this
4712 routine; it is the caller's responsibility to strip them.
4713
4714 If no such canonicalization is possible (i.e., two biv's are used or an
4715 expression that is neither invariant nor a biv or giv), this routine
4716 returns 0.
4717
4718 For a non-zero return, the result will have a code of CONST_INT, USE,
4719 REG (for a BIV), PLUS, or MULT. No other codes will occur.
4720
4721 *BENEFIT will be incremented by the benefit of any sub-giv encountered. */
4722
4723 static rtx
4724 simplify_giv_expr (x, benefit)
4725 rtx x;
4726 int *benefit;
4727 {
4728 enum machine_mode mode = GET_MODE (x);
4729 rtx arg0, arg1;
4730 rtx tem;
4731
4732 /* If this is not an integer mode, or if we cannot do arithmetic in this
4733 mode, this can't be a giv. */
4734 if (mode != VOIDmode
4735 && (GET_MODE_CLASS (mode) != MODE_INT
4736 || GET_MODE_BITSIZE (mode) > HOST_BITS_PER_INT))
4737 return 0;
4738
4739 switch (GET_CODE (x))
4740 {
4741 case PLUS:
4742 arg0 = simplify_giv_expr (XEXP (x, 0), benefit);
4743 arg1 = simplify_giv_expr (XEXP (x, 1), benefit);
4744 if (arg0 == 0 || arg1 == 0)
4745 return 0;
4746
4747 /* Put constant last, CONST_INT last if both constant. */
4748 if ((GET_CODE (arg0) == USE
4749 || GET_CODE (arg0) == CONST_INT)
4750 && GET_CODE (arg1) != CONST_INT)
4751 tem = arg0, arg0 = arg1, arg1 = tem;
4752
4753 /* Handle addition of zero, then addition of an invariant. */
4754 if (arg1 == const0_rtx)
4755 return arg0;
4756 else if (GET_CODE (arg1) == CONST_INT || GET_CODE (arg1) == USE)
4757 switch (GET_CODE (arg0))
4758 {
4759 case CONST_INT:
4760 case USE:
4761 /* Both invariant. Only valid if sum is machine operand.
4762 First strip off possible USE on first operand. */
4763 if (GET_CODE (arg0) == USE)
4764 arg0 = XEXP (arg0, 0);
4765
4766 tem = 0;
4767 if (CONSTANT_P (arg0) && GET_CODE (arg1) == CONST_INT)
4768 {
4769 tem = plus_constant (arg0, INTVAL (arg1));
4770 if (GET_CODE (tem) != CONST_INT)
4771 tem = gen_rtx (USE, mode, tem);
4772 }
4773
4774 return tem;
4775
4776 case REG:
4777 case MULT:
4778 /* biv + invar or mult + invar. Return sum. */
4779 return gen_rtx (PLUS, mode, arg0, arg1);
4780
4781 case PLUS:
4782 /* (a + invar_1) + invar_2. Associate. */
4783 return simplify_giv_expr (gen_rtx (PLUS, mode,
4784 XEXP (arg0, 0),
4785 gen_rtx (PLUS, mode,
4786 XEXP (arg0, 1), arg1)),
4787 benefit);
4788
4789 default:
4790 abort ();
4791 }
4792
4793 /* Each argument must be either REG, PLUS, or MULT. Convert REG to
4794 MULT to reduce cases. */
4795 if (GET_CODE (arg0) == REG)
4796 arg0 = gen_rtx (MULT, mode, arg0, const1_rtx);
4797 if (GET_CODE (arg1) == REG)
4798 arg1 = gen_rtx (MULT, mode, arg1, const1_rtx);
4799
4800 /* Now have PLUS + PLUS, PLUS + MULT, MULT + PLUS, or MULT + MULT.
4801 Put a MULT first, leaving PLUS + PLUS, MULT + PLUS, or MULT + MULT.
4802 Recurse to associate the second PLUS. */
4803 if (GET_CODE (arg1) == MULT)
4804 tem = arg0, arg0 = arg1, arg1 = tem;
4805
4806 if (GET_CODE (arg1) == PLUS)
4807 return simplify_giv_expr (gen_rtx (PLUS, mode,
4808 gen_rtx (PLUS, mode,
4809 arg0, XEXP (arg1, 0)),
4810 XEXP (arg1, 1)),
4811 benefit);
4812
4813 /* Now must have MULT + MULT. Distribute if same biv, else not giv. */
4814 if (GET_CODE (arg0) != MULT || GET_CODE (arg1) != MULT)
4815 abort ();
4816
4817 if (XEXP (arg0, 0) != XEXP (arg1, 0))
4818 return 0;
4819
4820 return simplify_giv_expr (gen_rtx (MULT, mode,
4821 XEXP (arg0, 0),
4822 gen_rtx (PLUS, mode,
4823 XEXP (arg0, 1),
4824 XEXP (arg1, 1))),
4825 benefit);
4826
4827 case MINUS:
4828 /* Handle "a - b" as "a + b * (-1)". */
4829 return simplify_giv_expr (gen_rtx (PLUS, mode,
4830 XEXP (x, 0),
4831 gen_rtx (MULT, mode,
4832 XEXP (x, 1),
4833 gen_rtx (CONST_INT,
4834 VOIDmode, -1))),
4835 benefit);
4836
4837 case MULT:
4838 arg0 = simplify_giv_expr (XEXP (x, 0), benefit);
4839 arg1 = simplify_giv_expr (XEXP (x, 1), benefit);
4840 if (arg0 == 0 || arg1 == 0)
4841 return 0;
4842
4843 /* Put constant last, CONST_INT last if both constant. */
4844 if ((GET_CODE (arg0) == USE || GET_CODE (arg0) == CONST_INT)
4845 && GET_CODE (arg1) != CONST_INT)
4846 tem = arg0, arg0 = arg1, arg1 = tem;
4847
4848 /* If second argument is not now constant, not giv. */
4849 if (GET_CODE (arg1) != USE && GET_CODE (arg1) != CONST_INT)
4850 return 0;
4851
4852 /* Handle multiply by 0 or 1. */
4853 if (arg1 == const0_rtx)
4854 return const0_rtx;
4855
4856 else if (arg1 == const1_rtx)
4857 return arg0;
4858
4859 switch (GET_CODE (arg0))
4860 {
4861 case REG:
4862 /* biv * invar. Done. */
4863 return gen_rtx (MULT, mode, arg0, arg1);
4864
4865 case CONST_INT:
4866 /* Product of two constants. */
4867 return gen_rtx (CONST_INT, mode, INTVAL (arg0) * INTVAL (arg1));
4868
4869 case USE:
4870 /* invar * invar. Not giv. */
4871 return 0;
4872
4873 case MULT:
4874 /* (a * invar_1) * invar_2. Associate. */
4875 return simplify_giv_expr (gen_rtx (MULT, mode,
4876 XEXP (arg0, 0),
4877 gen_rtx (MULT, mode,
4878 XEXP (arg0, 1), arg1)),
4879 benefit);
4880
4881 case PLUS:
4882 /* (a + invar_1) * invar_2. Distribute. */
4883 return simplify_giv_expr (gen_rtx (PLUS, mode,
4884 gen_rtx (MULT, mode,
4885 XEXP (arg0, 0), arg1),
4886 gen_rtx (MULT, mode,
4887 XEXP (arg0, 1), arg1)),
4888 benefit);
4889
4890 default:
4891 abort ();
4892 }
4893
4894 case ASHIFT:
4895 case LSHIFT:
4896 /* Shift by constant is multiply by power of two. */
4897 if (GET_CODE (XEXP (x, 1)) != CONST_INT)
4898 return 0;
4899
4900 return simplify_giv_expr (gen_rtx (MULT, mode,
4901 XEXP (x, 0),
4902 gen_rtx (CONST_INT, VOIDmode,
4903 1 << INTVAL (XEXP (x, 1)))),
4904 benefit);
4905
4906 case NEG:
4907 /* "-a" is "a * (-1)" */
4908 return simplify_giv_expr (gen_rtx (MULT, mode,
4909 XEXP (x, 0),
4910 gen_rtx (CONST_INT, VOIDmode, -1)),
4911 benefit);
4912
4913 case NOT:
4914 /* "~a" is "-a - 1". Silly, but easy. */
4915 return simplify_giv_expr (gen_rtx (MINUS, mode,
4916 gen_rtx (NEG, mode, XEXP (x, 0)),
4917 const1_rtx),
4918 benefit);
4919
4920 case USE:
4921 /* Already in proper form for invariant. */
4922 return x;
4923
4924 case REG:
4925 /* If this is a new register, we can't deal with it. */
4926 if (REGNO (x) >= max_reg_before_loop)
4927 return 0;
4928
4929 /* Check for biv or giv. */
4930 switch (reg_iv_type[REGNO (x)])
4931 {
4932 case BASIC_INDUCT:
4933 return x;
4934 case GENERAL_INDUCT:
4935 {
4936 struct induction *v = reg_iv_info[REGNO (x)];
4937
4938 /* Form expression from giv and add benefit. Ensure this giv
4939 can derive another and subtract any needed adjustment if so. */
4940 *benefit += v->benefit;
4941 if (v->cant_derive)
4942 return 0;
4943
4944 tem = gen_rtx (PLUS, mode, gen_rtx (MULT, mode,
4945 v->src_reg, v->mult_val),
4946 v->add_val);
4947 if (v->derive_adjustment)
4948 tem = gen_rtx (MINUS, mode, tem, v->derive_adjustment);
4949 return simplify_giv_expr (tem, benefit);
4950 }
4951 }
4952
4953 /* Fall through to general case. */
4954 default:
4955 /* If invariant, return as USE (unless CONST_INT).
4956 Otherwise, not giv. */
4957 if (GET_CODE (x) == USE)
4958 x = XEXP (x, 0);
4959
4960 if (invariant_p (x) == 1)
4961 {
4962 if (GET_CODE (x) == CONST_INT)
4963 return x;
4964 else
4965 return gen_rtx (USE, mode, x);
4966 }
4967 else
4968 return 0;
4969 }
4970 }
4971 \f
4972 /* Help detect a giv that is calculated by several consecutive insns;
4973 for example,
4974 giv = biv * M
4975 giv = giv + A
4976 The caller has already identified the first insn P as having a giv as dest;
4977 we check that all other insns that set the same register follow
4978 immediately after P, that they alter nothing else,
4979 and that the result of the last is still a giv.
4980
4981 The value is 0 if the reg set in P is not really a giv.
4982 Otherwise, the value is the amount gained by eliminating
4983 all the consecutive insns that compute the value.
4984
4985 FIRST_BENEFIT is the amount gained by eliminating the first insn, P.
4986 SRC_REG is the reg of the biv; DEST_REG is the reg of the giv.
4987
4988 The coefficients of the ultimate giv value are stored in
4989 *MULT_VAL and *ADD_VAL. */
4990
4991 static int
4992 consec_sets_giv (first_benefit, p, src_reg, dest_reg,
4993 add_val, mult_val)
4994 int first_benefit;
4995 rtx p;
4996 rtx src_reg;
4997 rtx dest_reg;
4998 rtx *add_val;
4999 rtx *mult_val;
5000 {
5001 int count;
5002 enum rtx_code code;
5003 int benefit;
5004 rtx temp;
5005 rtx set;
5006
5007 /* Indicate that this is a giv so that we can update the value produced in
5008 each insn of the multi-insn sequence.
5009
5010 This induction structure will be used only by the call to
5011 general_induction_var below, so we can allocate it on our stack.
5012 If this is a giv, our caller will replace the induct var entry with
5013 a new induction structure. */
5014 struct induction *v
5015 = (struct induction *) alloca (sizeof (struct induction));
5016 v->src_reg = src_reg;
5017 v->mult_val = *mult_val;
5018 v->add_val = *add_val;
5019 v->benefit = first_benefit;
5020 v->cant_derive = 0;
5021 v->derive_adjustment = 0;
5022
5023 reg_iv_type[REGNO (dest_reg)] = GENERAL_INDUCT;
5024 reg_iv_info[REGNO (dest_reg)] = v;
5025
5026 count = n_times_set[REGNO (dest_reg)] - 1;
5027
5028 while (count > 0)
5029 {
5030 p = NEXT_INSN (p);
5031 code = GET_CODE (p);
5032
5033 /* If libcall, skip to end of call sequence. */
5034 if (code == INSN && (temp = find_reg_note (p, REG_LIBCALL, 0)))
5035 p = XEXP (temp, 0);
5036
5037 if (code == INSN
5038 && (set = single_set (p))
5039 && GET_CODE (SET_DEST (set)) == REG
5040 && SET_DEST (set) == dest_reg
5041 && ((benefit = general_induction_var (SET_SRC (set), &src_reg,
5042 add_val, mult_val))
5043 /* Giv created by equivalent expression. */
5044 || ((temp = find_reg_note (p, REG_EQUAL, 0))
5045 && (benefit = general_induction_var (XEXP (temp, 0), &src_reg,
5046 add_val, mult_val))))
5047 && src_reg == v->src_reg)
5048 {
5049 if (find_reg_note (p, REG_RETVAL, 0))
5050 benefit += libcall_benefit (p);
5051
5052 count--;
5053 v->mult_val = *mult_val;
5054 v->add_val = *add_val;
5055 v->benefit = benefit;
5056 }
5057 else if (code != NOTE)
5058 {
5059 /* Allow insns that set something other than this giv to a
5060 constant. Such insns are needed on machines which cannot
5061 include long constants and should not disqualify a giv. */
5062 if (code == INSN
5063 && (set = single_set (p))
5064 && SET_DEST (set) != dest_reg
5065 && CONSTANT_P (SET_SRC (set)))
5066 continue;
5067
5068 reg_iv_type[REGNO (dest_reg)] = UNKNOWN_INDUCT;
5069 return 0;
5070 }
5071 }
5072
5073 return v->benefit;
5074 }
5075 \f
5076 /* Return an rtx, if any, that expresses giv G2 as a function of the register
5077 represented by G1. If no such expression can be found, or it is clear that
5078 it cannot possibly be a valid address, 0 is returned.
5079
5080 To perform the computation, we note that
5081 G1 = a * v + b and
5082 G2 = c * v + d
5083 where `v' is the biv.
5084
5085 So G2 = (c/a) * G1 + (d - b*c/a) */
5086
5087 #ifdef ADDRESS_COST
5088 static rtx
5089 express_from (g1, g2)
5090 struct induction *g1, *g2;
5091 {
5092 rtx mult, add;
5093
5094 /* The value that G1 will be multiplied by must be a constant integer. Also,
5095 the only chance we have of getting a valid address is if b*c/a (see above
5096 for notation) is also an integer. */
5097 if (GET_CODE (g1->mult_val) != CONST_INT
5098 || GET_CODE (g2->mult_val) != CONST_INT
5099 || GET_CODE (g1->add_val) != CONST_INT
5100 || g1->mult_val == const0_rtx
5101 || INTVAL (g2->mult_val) % INTVAL (g1->mult_val) != 0)
5102 return 0;
5103
5104 mult = gen_rtx (CONST_INT, VOIDmode,
5105 INTVAL (g2->mult_val) / INTVAL (g1->mult_val));
5106 add = plus_constant (g2->add_val, - INTVAL (g1->add_val) * INTVAL (mult));
5107
5108 /* Form simplified final result. */
5109 if (mult == const0_rtx)
5110 return add;
5111 else if (mult == const1_rtx)
5112 mult = g1->dest_reg;
5113 else
5114 mult = gen_rtx (MULT, g2->mode, g1->dest_reg, mult);
5115
5116 if (add == const0_rtx)
5117 return mult;
5118 else
5119 return gen_rtx (PLUS, g2->mode, mult, add);
5120 }
5121 #endif
5122 \f
5123 /* Return 1 if giv G2 can be combined with G1. This means that G2 can use
5124 (either directly or via an address expression) a register used to represent
5125 G1. Set g2->new_reg to a represtation of G1 (normally just
5126 g1->dest_reg). */
5127
5128 static int
5129 combine_givs_p (g1, g2)
5130 struct induction *g1, *g2;
5131 {
5132 rtx tem;
5133
5134 /* If these givs are identical, they can be combined. */
5135 if (rtx_equal_p (g1->mult_val, g2->mult_val)
5136 && rtx_equal_p (g1->add_val, g2->add_val))
5137 {
5138 g2->new_reg = g1->dest_reg;
5139 return 1;
5140 }
5141
5142 #ifdef ADDRESS_COST
5143 /* If G2 can be expressed as a function of G1 and that function is valid
5144 as an address and no more expensive than using a register for G2,
5145 the expression of G2 in terms of G1 can be used. */
5146 if (g2->giv_type == DEST_ADDR
5147 && (tem = express_from (g1, g2)) != 0
5148 && memory_address_p (g2->mem_mode, tem)
5149 && ADDRESS_COST (tem) <= ADDRESS_COST (*g2->location))
5150 {
5151 g2->new_reg = tem;
5152 return 1;
5153 }
5154 #endif
5155
5156 return 0;
5157 }
5158 \f
5159 /* Check all pairs of givs for iv_class BL and see if any can be combined with
5160 any other. If so, point SAME to the giv combined with and set NEW_REG to
5161 be an expression (in terms of the other giv's DEST_REG) equivalent to the
5162 giv. Also, update BENEFIT and related fields for cost/benefit analysis. */
5163
5164 static void
5165 combine_givs (bl)
5166 struct iv_class *bl;
5167 {
5168 struct induction *g1, *g2;
5169 int pass;
5170
5171 for (g1 = bl->giv; g1; g1 = g1->next_iv)
5172 for (pass = 0; pass <= 1; pass++)
5173 for (g2 = bl->giv; g2; g2 = g2->next_iv)
5174 if (g1 != g2
5175 /* First try to combine with replaceable givs, then all givs. */
5176 && (g1->replaceable || pass == 1)
5177 /* If either has already been combined or is to be ignored, can't
5178 combine. */
5179 && ! g1->ignore && ! g2->ignore && ! g1->same && ! g2->same
5180 /* If something has been based on G2, G2 cannot itself be based
5181 on something else. */
5182 && ! g2->combined_with
5183 && combine_givs_p (g1, g2))
5184 {
5185 /* g2->new_reg set by `combine_givs_p' */
5186 g2->same = g1;
5187 g1->combined_with = 1;
5188 g1->benefit += g2->benefit;
5189 /* ??? The new final_[bg]iv_value code does a much better job
5190 of finding replaceable giv's, and hence this code may no
5191 longer be necessary. */
5192 if (! g2->replaceable && REG_USERVAR_P (g2->dest_reg))
5193 g1->benefit -= copy_cost;
5194 g1->lifetime += g2->lifetime;
5195 g1->times_used += g2->times_used;
5196
5197 if (loop_dump_stream)
5198 fprintf (loop_dump_stream, "giv at %d combined with giv at %d\n",
5199 INSN_UID (g2->insn), INSN_UID (g1->insn));
5200 }
5201 }
5202 \f
5203 /* EMIT code before INSERT_BEFORE to set REG = B * M + A. */
5204
5205 void
5206 emit_iv_add_mult (b, m, a, reg, insert_before)
5207 rtx b; /* initial value of basic induction variable */
5208 rtx m; /* multiplicative constant */
5209 rtx a; /* additive constant */
5210 rtx reg; /* destination register */
5211 rtx insert_before;
5212 {
5213 rtx seq;
5214 rtx result;
5215
5216 /* Prevent unexpected sharing of these rtx. */
5217 a = copy_rtx (a);
5218 b = copy_rtx (b);
5219
5220 /* Increase the lifetime of any invariants moved further in code. */
5221 update_reg_last_use (a, insert_before);
5222 update_reg_last_use (b, insert_before);
5223 update_reg_last_use (m, insert_before);
5224
5225 start_sequence ();
5226 result = expand_mult_add (b, reg, m, a, GET_MODE (reg), 0);
5227 if (reg != result)
5228 emit_move_insn (reg, result);
5229 seq = gen_sequence ();
5230 end_sequence ();
5231
5232 emit_insn_before (seq, insert_before);
5233 }
5234 \f
5235 /* Test whether A * B can be computed without
5236 an actual multiply insn. Value is 1 if so. */
5237
5238 static int
5239 product_cheap_p (a, b)
5240 rtx a;
5241 rtx b;
5242 {
5243 int i;
5244 rtx tmp;
5245 struct obstack *old_rtl_obstack = rtl_obstack;
5246 char *storage = (char *) obstack_alloc (&temp_obstack, 0);
5247 int win = 1;
5248
5249 /* If only one is constant, make it B. */
5250 if (GET_CODE (a) == CONST_INT)
5251 tmp = a, a = b, b = tmp;
5252
5253 /* If first constant, both constant, so don't need multiply. */
5254 if (GET_CODE (a) == CONST_INT)
5255 return 1;
5256
5257 /* If second not constant, neither is constant, so would need multiply. */
5258 if (GET_CODE (b) != CONST_INT)
5259 return 0;
5260
5261 /* One operand is constant, so might not need multiply insn. Generate the
5262 code for the multiply and see if a call or multiply, or long sequence
5263 of insns is generated. */
5264
5265 rtl_obstack = &temp_obstack;
5266 start_sequence ();
5267 expand_mult (GET_MODE (a), a, b, 0, 0);
5268 tmp = gen_sequence ();
5269 end_sequence ();
5270
5271 if (GET_CODE (tmp) == SEQUENCE)
5272 {
5273 if (XVEC (tmp, 0) == 0)
5274 win = 1;
5275 else if (XVECLEN (tmp, 0) > 3)
5276 win = 0;
5277 else
5278 for (i = 0; i < XVECLEN (tmp, 0); i++)
5279 {
5280 rtx insn = XVECEXP (tmp, 0, i);
5281
5282 if (GET_CODE (insn) != INSN
5283 || (GET_CODE (PATTERN (insn)) == SET
5284 && GET_CODE (SET_SRC (PATTERN (insn))) == MULT)
5285 || (GET_CODE (PATTERN (insn)) == PARALLEL
5286 && GET_CODE (XVECEXP (PATTERN (insn), 0, 0)) == SET
5287 && GET_CODE (SET_SRC (XVECEXP (PATTERN (insn), 0, 0))) == MULT))
5288 {
5289 win = 0;
5290 break;
5291 }
5292 }
5293 }
5294 else if (GET_CODE (tmp) == SET
5295 && GET_CODE (SET_SRC (tmp)) == MULT)
5296 win = 0;
5297 else if (GET_CODE (tmp) == PARALLEL
5298 && GET_CODE (XVECEXP (tmp, 0, 0)) == SET
5299 && GET_CODE (SET_SRC (XVECEXP (tmp, 0, 0))) == MULT)
5300 win = 0;
5301
5302 /* Free any storage we obtained in generating this multiply and restore rtl
5303 allocation to its normal obstack. */
5304 obstack_free (&temp_obstack, storage);
5305 rtl_obstack = old_rtl_obstack;
5306
5307 return win;
5308 }
5309 \f
5310 /* Check to see if loop can be terminated by a "decrement and branch until
5311 zero" instruction. If so, add a REG_NONNEG note to the branch insn if so.
5312 Also try reversing an increment loop to a decrement loop
5313 to see if the optimization can be performed.
5314 Value is nonzero if optimization was performed. */
5315
5316 /* This is useful even if the architecture doesn't have such an insn,
5317 because it might change a loops which increments from 0 to n to a loop
5318 which decrements from n to 0. A loop that decrements to zero is usually
5319 faster than one that increments from zero. */
5320
5321 /* ??? This could be rewritten to use some of the loop unrolling procedures,
5322 such as approx_final_value, biv_total_increment, loop_iterations, and
5323 final_[bg]iv_value. */
5324
5325 static int
5326 check_dbra_loop (loop_end, insn_count, loop_start)
5327 rtx loop_end;
5328 int insn_count;
5329 rtx loop_start;
5330 {
5331 struct iv_class *bl;
5332 rtx reg;
5333 rtx jump_label;
5334 rtx final_value;
5335 rtx start_value;
5336 enum rtx_code branch_code;
5337 rtx new_add_val;
5338 rtx comparison;
5339 rtx before_comparison;
5340 rtx p;
5341
5342 /* If last insn is a conditional branch, and the insn before tests a
5343 register value, try to optimize it. Otherwise, we can't do anything. */
5344
5345 comparison = get_condition_for_loop (PREV_INSN (loop_end));
5346 if (comparison == 0)
5347 return 0;
5348
5349 /* Check all of the bivs to see if the compare uses one of them.
5350 Skip biv's set more than once because we can't guarantee that
5351 it will be zero on the last iteration. Also skip if the biv is
5352 used between its update and the test insn. */
5353
5354 for (bl = loop_iv_list; bl; bl = bl->next)
5355 {
5356 if (bl->biv_count == 1
5357 && bl->biv->dest_reg == XEXP (comparison, 0)
5358 && ! reg_used_between_p (regno_reg_rtx[bl->regno], bl->biv->insn,
5359 PREV_INSN (PREV_INSN (loop_end))))
5360 break;
5361 }
5362
5363 if (! bl)
5364 return 0;
5365
5366 /* Look for the case where the basic induction variable is always
5367 nonnegative, and equals zero on the last iteration.
5368 In this case, add a reg_note REG_NONNEG, which allows the
5369 m68k DBRA instruction to be used. */
5370
5371 if (((GET_CODE (comparison) == GT
5372 && GET_CODE (XEXP (comparison, 1)) == CONST_INT
5373 && INTVAL (XEXP (comparison, 1)) == -1)
5374 || (GET_CODE (comparison) == NE && XEXP (comparison, 1) == const0_rtx))
5375 && GET_CODE (bl->biv->add_val) == CONST_INT
5376 && INTVAL (bl->biv->add_val) < 0)
5377 {
5378 /* Initial value must be greater than 0,
5379 init_val % -dec_value == 0 to ensure that it equals zero on
5380 the last iteration */
5381
5382 if (GET_CODE (bl->initial_value) == CONST_INT
5383 && INTVAL (bl->initial_value) > 0
5384 && (INTVAL (bl->initial_value) %
5385 (-INTVAL (bl->biv->add_val))) == 0)
5386 {
5387 /* register always nonnegative, add REG_NOTE to branch */
5388 REG_NOTES (PREV_INSN (loop_end))
5389 = gen_rtx (EXPR_LIST, REG_NONNEG, 0,
5390 REG_NOTES (PREV_INSN (loop_end)));
5391 bl->nonneg = 1;
5392
5393 return 1;
5394 }
5395
5396 /* If the decrement is 1 and the value was tested as >= 0 before
5397 the loop, then we can safely optimize. */
5398 for (p = loop_start; p; p = PREV_INSN (p))
5399 {
5400 if (GET_CODE (p) == CODE_LABEL)
5401 break;
5402 if (GET_CODE (p) != JUMP_INSN)
5403 continue;
5404
5405 before_comparison = get_condition_for_loop (p);
5406 if (before_comparison
5407 && XEXP (before_comparison, 0) == bl->biv->dest_reg
5408 && GET_CODE (before_comparison) == LT
5409 && XEXP (before_comparison, 1) == const0_rtx
5410 && ! reg_set_between_p (bl->biv->dest_reg, p, loop_start)
5411 && INTVAL (bl->biv->add_val) == -1)
5412 {
5413 REG_NOTES (PREV_INSN (loop_end))
5414 = gen_rtx (EXPR_LIST, REG_NONNEG, 0,
5415 REG_NOTES (PREV_INSN (loop_end)));
5416 bl->nonneg = 1;
5417
5418 return 1;
5419 }
5420 }
5421 }
5422 else if (num_mem_sets <= 1)
5423 {
5424 /* Try to change inc to dec, so can apply above optimization. */
5425 /* Can do this if:
5426 all registers modified are induction variables or invariant,
5427 all memory references have non-overlapping addresses
5428 (obviously true if only one write)
5429 allow 2 insns for the compare/jump at the end of the loop. */
5430 int num_nonfixed_reads = 0;
5431 /* 1 if the iteration var is used only to count iterations. */
5432 int no_use_except_counting = 0;
5433
5434 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
5435 if (GET_RTX_CLASS (GET_CODE (p)) == 'i')
5436 num_nonfixed_reads += count_nonfixed_reads (PATTERN (p));
5437
5438 if (bl->giv_count == 0
5439 && ! loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]])
5440 {
5441 rtx bivreg = regno_reg_rtx[bl->regno];
5442
5443 /* If there are no givs for this biv, and the only exit is the
5444 fall through at the end of the the loop, then
5445 see if perhaps there are no uses except to count. */
5446 no_use_except_counting = 1;
5447 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
5448 if (GET_RTX_CLASS (GET_CODE (p)) == 'i')
5449 {
5450 rtx set = single_set (p);
5451
5452 if (set && GET_CODE (SET_DEST (set)) == REG
5453 && REGNO (SET_DEST (set)) == bl->regno)
5454 /* An insn that sets the biv is okay. */
5455 ;
5456 else if (p == prev_nonnote_insn (prev_nonnote_insn (loop_end))
5457 || p == prev_nonnote_insn (loop_end))
5458 /* Don't bother about the end test. */
5459 ;
5460 else if (reg_mentioned_p (bivreg, PATTERN (p)))
5461 /* Any other use of the biv is no good. */
5462 {
5463 no_use_except_counting = 0;
5464 break;
5465 }
5466 }
5467 }
5468
5469 /* This code only acts for innermost loops. Also it simplifies
5470 the memory address check by only reversing loops with
5471 zero or one memory access.
5472 Two memory accesses could involve parts of the same array,
5473 and that can't be reversed. */
5474
5475 if (num_nonfixed_reads <= 1
5476 && !loop_has_call
5477 && (no_use_except_counting
5478 || (bl->giv_count + bl->biv_count + num_mem_sets
5479 + num_movables + 2 == insn_count)))
5480 {
5481 rtx condition = get_condition_for_loop (PREV_INSN (loop_end));
5482 int win;
5483 rtx tem;
5484
5485 /* Loop can be reversed. */
5486 if (loop_dump_stream)
5487 fprintf (loop_dump_stream, "Can reverse loop\n");
5488
5489 /* Now check other conditions:
5490 initial_value must be zero,
5491 final_value % add_val == 0, so that when reversed, the
5492 biv will be zero on the last iteration.
5493
5494 This test can probably be improved since +/- 1 in the constant
5495 can be obtained by changing LT to LE and vice versa; this is
5496 confusing. */
5497
5498 if (comparison && bl->initial_value == const0_rtx
5499 && GET_CODE (XEXP (comparison, 1)) == CONST_INT
5500 /* LE gets turned into LT */
5501 && GET_CODE (comparison) == LT
5502 && (INTVAL (XEXP (comparison, 1))
5503 % INTVAL (bl->biv->add_val)) == 0)
5504 {
5505 /* Register will always be nonnegative, with value
5506 0 on last iteration if loop reversed */
5507
5508 /* Save some info needed to produce the new insns. */
5509 reg = bl->biv->dest_reg;
5510 jump_label = XEXP (SET_SRC (PATTERN (PREV_INSN (loop_end))), 1);
5511 new_add_val = gen_rtx (CONST_INT, VOIDmode,
5512 - INTVAL (bl->biv->add_val));
5513
5514 final_value = XEXP (comparison, 1);
5515 start_value = gen_rtx (CONST_INT, VOIDmode,
5516 (INTVAL (XEXP (comparison, 1))
5517 - INTVAL (bl->biv->add_val)));
5518
5519 /* Initialize biv to start_value before loop start.
5520 The old initializing insn will be deleted as a
5521 dead store by flow.c. */
5522 emit_insn_before (gen_move_insn (reg, start_value), loop_start);
5523
5524 /* Add insn to decrement register, and delete insn
5525 that incremented the register. */
5526 p = emit_insn_before (gen_add2_insn (reg, new_add_val),
5527 bl->biv->insn);
5528 delete_insn (bl->biv->insn);
5529
5530 /* Update biv info to reflect its new status. */
5531 bl->biv->insn = p;
5532 bl->initial_value = start_value;
5533 bl->biv->add_val = new_add_val;
5534
5535 /* Inc LABEL_NUSES so that delete_insn will
5536 not delete the label. */
5537 LABEL_NUSES (XEXP (jump_label, 0)) ++;
5538
5539 /* Emit an insn after the end of the loop to set the biv's
5540 proper exit value if it is used anywhere outside the loop. */
5541 if ((regno_last_uid[bl->regno]
5542 != INSN_UID (PREV_INSN (PREV_INSN (loop_end))))
5543 || ! bl->init_insn
5544 || regno_first_uid[bl->regno] != INSN_UID (bl->init_insn))
5545 emit_insn_after (gen_move_insn (reg, final_value),
5546 loop_end);
5547
5548 /* Delete compare/branch at end of loop. */
5549 delete_insn (PREV_INSN (loop_end));
5550 delete_insn (PREV_INSN (loop_end));
5551
5552 /* Add new compare/branch insn at end of loop. */
5553 start_sequence ();
5554 emit_cmp_insn (reg, const0_rtx, GE, 0, GET_MODE (reg), 0, 0);
5555 emit_jump_insn (gen_bge (XEXP (jump_label, 0)));
5556 tem = gen_sequence ();
5557 end_sequence ();
5558 emit_jump_insn_before (tem, loop_end);
5559
5560 for (tem = PREV_INSN (loop_end);
5561 tem && GET_CODE (tem) != JUMP_INSN; tem = PREV_INSN (tem))
5562 ;
5563 if (tem)
5564 {
5565 JUMP_LABEL (tem) = XEXP (jump_label, 0);
5566
5567 /* Increment of LABEL_NUSES done above. */
5568 /* Register is now always nonnegative,
5569 so add REG_NONNEG note to the branch. */
5570 REG_NOTES (tem) = gen_rtx (EXPR_LIST, REG_NONNEG, 0,
5571 REG_NOTES (tem));
5572 }
5573
5574 bl->nonneg = 1;
5575
5576 /* Mark that this biv has been reversed. Each giv which depends
5577 on this biv, and which is also live past the end of the loop
5578 will have to be fixed up. */
5579
5580 bl->reversed = 1;
5581
5582 if (loop_dump_stream)
5583 fprintf (loop_dump_stream,
5584 "Reversed loop and added reg_nonneg\n");
5585
5586 return 1;
5587 }
5588 }
5589 }
5590
5591 return 0;
5592 }
5593 \f
5594 /* Verify whether the biv BL appears to be eliminable,
5595 based on the insns in the loop that refer to it.
5596 LOOP_START is the first insn of the loop, and END is the end insn.
5597
5598 If ELIMINATE_P is non-zero, actually do the elimination.
5599
5600 THRESHOLD and INSN_COUNT are from loop_optimize and are used to
5601 determine whether invariant insns should be placed inside or at the
5602 start of the loop. */
5603
5604 static int
5605 maybe_eliminate_biv (bl, loop_start, end, eliminate_p, threshold, insn_count)
5606 struct iv_class *bl;
5607 rtx loop_start;
5608 rtx end;
5609 int eliminate_p;
5610 int threshold, insn_count;
5611 {
5612 rtx reg = bl->biv->dest_reg;
5613 rtx p, set;
5614 struct induction *v;
5615
5616 /* Scan all insns in the loop, stopping if we find one that uses the
5617 biv in a way that we cannot eliminate. */
5618
5619 for (p = loop_start; p != end; p = NEXT_INSN (p))
5620 {
5621 enum rtx_code code = GET_CODE (p);
5622 rtx where = threshold >= insn_count ? loop_start : p;
5623
5624 if ((code == INSN || code == JUMP_INSN || code == CALL_INSN)
5625 && reg_mentioned_p (reg, PATTERN (p))
5626 && ! maybe_eliminate_biv_1 (PATTERN (p), p, bl, eliminate_p, where))
5627 {
5628 if (loop_dump_stream)
5629 fprintf (loop_dump_stream,
5630 "Cannot eliminate biv %d: biv used in insn %d.\n",
5631 bl->regno, INSN_UID (p));
5632 break;
5633 }
5634 }
5635
5636 if (p == end)
5637 {
5638 if (loop_dump_stream)
5639 fprintf (loop_dump_stream, "biv %d %s eliminated.\n",
5640 bl->regno, eliminate_p ? "was" : "can be");
5641 return 1;
5642 }
5643
5644 return 0;
5645 }
5646 \f
5647 /* If BL appears in X (part of the pattern of INSN), see if we can
5648 eliminate its use. If so, return 1. If not, return 0.
5649
5650 If BIV does not appear in X, return 1.
5651
5652 If ELIMINATE_P is non-zero, actually do the elimination. WHERE indicates
5653 where extra insns should be added. Depending on how many items have been
5654 moved out of the loop, it will either be before INSN or at the start of
5655 the loop. */
5656
5657 static int
5658 maybe_eliminate_biv_1 (x, insn, bl, eliminate_p, where)
5659 rtx x, insn;
5660 struct iv_class *bl;
5661 int eliminate_p;
5662 rtx where;
5663 {
5664 enum rtx_code code = GET_CODE (x);
5665 rtx reg = bl->biv->dest_reg;
5666 enum machine_mode mode = GET_MODE (reg);
5667 struct induction *v;
5668 rtx arg, new, tem;
5669 int arg_operand;
5670 char *fmt;
5671 int i, j;
5672
5673 switch (code)
5674 {
5675 case REG:
5676 /* If we haven't already been able to do something with this BIV,
5677 we can't eliminate it. */
5678 if (x == reg)
5679 return 0;
5680 return 1;
5681
5682 case SET:
5683 /* If this sets the BIV, it is not a problem. */
5684 if (SET_DEST (x) == reg)
5685 return 1;
5686
5687 /* If this is an insn that defines a giv, it is also ok because
5688 it will go away when the giv is reduced. */
5689 for (v = bl->giv; v; v = v->next_iv)
5690 if (v->giv_type == DEST_REG && SET_DEST (x) == v->dest_reg)
5691 return 1;
5692
5693 #ifdef HAVE_cc0
5694 if (SET_DEST (x) == cc0_rtx && SET_SRC (x) == reg)
5695 {
5696 /* Can replace with any giv that was reduced and
5697 that has (MULT_VAL != 0) and (ADD_VAL == 0).
5698 Require a constant for MULT_VAL, so we know it's nonzero. */
5699
5700 for (v = bl->giv; v; v = v->next_iv)
5701 if (CONSTANT_P (v->mult_val) && v->mult_val != const0_rtx
5702 && v->add_val == const0_rtx
5703 && ! v->ignore && ! v->maybe_dead
5704 && v->mode == mode)
5705 {
5706 if (! eliminate_p)
5707 return 1;
5708
5709 /* If the giv has the opposite direction of change,
5710 then reverse the comparison. */
5711 if (INTVAL (v->mult_val) < 0)
5712 new = gen_rtx (COMPARE, GET_MODE (v->new_reg),
5713 const0_rtx, v->new_reg);
5714 else
5715 new = v->new_reg;
5716
5717 /* We can probably test that giv's reduced reg. */
5718 if (validate_change (insn, &SET_SRC (x), new, 0))
5719 return 1;
5720 }
5721
5722 /* Look for a giv with (MULT_VAL != 0) and (ADD_VAL != 0);
5723 replace test insn with a compare insn (cmp REDUCED_GIV ADD_VAL).
5724 Require a constant for MULT_VAL, so we know it's nonzero. */
5725
5726 for (v = bl->giv; v; v = v->next_iv)
5727 if (CONSTANT_P (v->mult_val) && v->mult_val != const0_rtx
5728 && ! v->ignore && ! v->maybe_dead
5729 && v->mode == mode)
5730 {
5731 if (! eliminate_p)
5732 return 1;
5733
5734 /* If the giv has the opposite direction of change,
5735 then reverse the comparison. */
5736 if (INTVAL (v->mult_val) < 0)
5737 new = gen_rtx (COMPARE, VOIDmode, copy_rtx (v->add_val),
5738 v->new_reg);
5739 else
5740 new = gen_rtx (COMPARE, VOIDmode, v->new_reg,
5741 copy_rtx (v->add_val));
5742
5743 /* Replace biv with the giv's reduced register. */
5744 update_reg_last_use (v->add_val, insn);
5745 if (validate_change (insn, &SET_SRC (PATTERN (insn)), new, 0))
5746 return 1;
5747
5748 /* Insn doesn't support that constant or invariant. Copy it
5749 into a register (it will be a loop invariant.) */
5750 tem = gen_reg_rtx (GET_MODE (v->new_reg));
5751
5752 emit_insn_before (gen_move_insn (tem, copy_rtx (v->add_val)),
5753 where);
5754
5755 if (validate_change (insn, &SET_SRC (PATTERN (insn)),
5756 gen_rtx (COMPARE, VOIDmode,
5757 v->new_reg, tem), 0))
5758 return 1;
5759 }
5760 }
5761 #endif
5762 break;
5763
5764 case COMPARE:
5765 case EQ: case NE:
5766 case GT: case GE: case GTU: case GEU:
5767 case LT: case LE: case LTU: case LEU:
5768 /* See if either argument is the biv. */
5769 if (XEXP (x, 0) == reg)
5770 arg = XEXP (x, 1), arg_operand = 1;
5771 else if (XEXP (x, 1) == reg)
5772 arg = XEXP (x, 0), arg_operand = 0;
5773 else
5774 break;
5775
5776 if (CONSTANT_P (arg))
5777 {
5778 /* First try to replace with any giv that has constant positive
5779 mult_val and constant add_val. We might be able to support
5780 negative mult_val, but it seems complex to do it in general. */
5781
5782 for (v = bl->giv; v; v = v->next_iv)
5783 if (CONSTANT_P (v->mult_val) && INTVAL (v->mult_val) > 0
5784 && CONSTANT_P (v->add_val)
5785 && ! v->ignore && ! v->maybe_dead
5786 && v->mode == mode)
5787 {
5788 if (! eliminate_p)
5789 return 1;
5790
5791 /* Replace biv with the giv's reduced reg. */
5792 XEXP (x, 1-arg_operand) = v->new_reg;
5793
5794 /* If all constants are actually constant integers and
5795 the derived constant can be directly placed in the COMPARE,
5796 do so. */
5797 if (GET_CODE (arg) == CONST_INT
5798 && GET_CODE (v->mult_val) == CONST_INT
5799 && GET_CODE (v->add_val) == CONST_INT
5800 && validate_change (insn, &XEXP (x, arg_operand),
5801 gen_rtx (CONST_INT, VOIDmode,
5802 (INTVAL (arg)
5803 * INTVAL (v->mult_val)
5804 + INTVAL (v->add_val))), 0))
5805 return 1;
5806
5807 /* Otherwise, load it into a register. */
5808 tem = gen_reg_rtx (mode);
5809 emit_iv_add_mult (arg, v->mult_val, v->add_val, tem, where);
5810 if (validate_change (insn, &XEXP (x, arg_operand), tem, 0))
5811 return 1;
5812
5813 /* If that failed, put back the change we made above. */
5814 XEXP (x, 1-arg_operand) = reg;
5815 }
5816
5817 /* Look for giv with positive constant mult_val and nonconst add_val.
5818 Insert insns to calculate new compare value. */
5819
5820 for (v = bl->giv; v; v = v->next_iv)
5821 if (CONSTANT_P (v->mult_val)
5822 && ! v->ignore && ! v->maybe_dead
5823 && v->mode == mode)
5824 {
5825 rtx tem;
5826
5827 if (! eliminate_p)
5828 return 1;
5829
5830 tem = gen_reg_rtx (mode);
5831
5832 /* Replace biv with giv's reduced register. */
5833 validate_change (insn, &XEXP (x, 1 - arg_operand),
5834 v->new_reg, 1);
5835
5836 /* Compute value to compare against. */
5837 emit_iv_add_mult (arg, v->mult_val, v->add_val, tem, where);
5838 /* Use it in this insn. */
5839 validate_change (insn, &XEXP (x, arg_operand), tem, 1);
5840 if (apply_change_group ())
5841 return 1;
5842 }
5843 }
5844 else if (GET_CODE (arg) == REG || GET_CODE (arg) == MEM)
5845 {
5846 if (invariant_p (arg) == 1)
5847 {
5848 /* Look for giv with constant positive mult_val and nonconst
5849 add_val. Insert insns to compute new compare value. */
5850
5851 for (v = bl->giv; v; v = v->next_iv)
5852 if (CONSTANT_P (v->mult_val) && INTVAL (v->mult_val) > 0
5853 && ! v->ignore && ! v->maybe_dead
5854 && v->mode == mode)
5855 {
5856 rtx tem;
5857
5858 if (! eliminate_p)
5859 return 1;
5860
5861 tem = gen_reg_rtx (mode);
5862
5863 /* Replace biv with giv's reduced register. */
5864 validate_change (insn, &XEXP (x, 1 - arg_operand),
5865 v->new_reg, 1);
5866
5867 /* Compute value to compare against. */
5868 emit_iv_add_mult (arg, v->mult_val, v->add_val,
5869 tem, where);
5870 validate_change (insn, &XEXP (x, arg_operand), tem, 1);
5871 if (apply_change_group ())
5872 return 1;
5873 }
5874 }
5875
5876 /* This code has problems. Basically, you can't know when
5877 seeing if we will eliminate BL, whether a particular giv
5878 of ARG will be reduced. If it isn't going to be reduced,
5879 we can't eliminate BL. We can try forcing it to be reduced,
5880 but that can generate poor code.
5881
5882 The problem is that the benefit of reducing TV, below should
5883 be increased if BL can actually be eliminated, but this means
5884 we might have to do a topological sort of the order in which
5885 we try to process biv. It doesn't seem worthwhile to do
5886 this sort of thing now. */
5887
5888 #if 0
5889 /* Otherwise the reg compared with had better be a biv. */
5890 if (GET_CODE (arg) != REG
5891 || reg_iv_type[REGNO (arg)] != BASIC_INDUCT)
5892 return 0;
5893
5894 /* Look for a pair of givs, one for each biv,
5895 with identical coefficients. */
5896 for (v = bl->giv; v; v = v->next_iv)
5897 {
5898 struct induction *tv;
5899
5900 if (v->ignore || v->maybe_dead || v->mode != mode)
5901 continue;
5902
5903 for (tv = reg_biv_class[REGNO (arg)]->giv; tv; tv = tv->next_iv)
5904 if (! tv->ignore && ! tv->maybe_dead
5905 && rtx_equal_p (tv->mult_val, v->mult_val)
5906 && rtx_equal_p (tv->add_val, v->add_val)
5907 && tv->mode == mode)
5908 {
5909 if (! eliminate_p)
5910 return 1;
5911
5912 /* Replace biv with its giv's reduced reg. */
5913 XEXP (x, 1-arg_operand) = v->new_reg;
5914 /* Replace other operand with the other giv's
5915 reduced reg. */
5916 XEXP (x, arg_operand) = tv->new_reg;
5917 return 1;
5918 }
5919 }
5920 #endif
5921 }
5922
5923 /* If we get here, the biv can't be eliminated. */
5924 return 0;
5925
5926 case MEM:
5927 /* If this address is a DEST_ADDR giv, it doesn't matter if the
5928 biv is used in it, since it will be replaced. */
5929 for (v = bl->giv; v; v = v->next_iv)
5930 if (v->giv_type == DEST_ADDR && v->location == &XEXP (x, 0))
5931 return 1;
5932 break;
5933 }
5934
5935 /* See if any subexpression fails elimination. */
5936 fmt = GET_RTX_FORMAT (code);
5937 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
5938 {
5939 switch (fmt[i])
5940 {
5941 case 'e':
5942 if (! maybe_eliminate_biv_1 (XEXP (x, i), insn, bl,
5943 eliminate_p, where))
5944 return 0;
5945 break;
5946
5947 case 'E':
5948 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
5949 if (! maybe_eliminate_biv_1 (XVECEXP (x, i, j), insn, bl,
5950 eliminate_p, where))
5951 return 0;
5952 break;
5953 }
5954 }
5955
5956 return 1;
5957 }
5958 \f
5959 /* Return nonzero if the last use of REG
5960 is in an insn following INSN in the same basic block. */
5961
5962 static int
5963 last_use_this_basic_block (reg, insn)
5964 rtx reg;
5965 rtx insn;
5966 {
5967 rtx n;
5968 for (n = insn;
5969 n && GET_CODE (n) != CODE_LABEL && GET_CODE (n) != JUMP_INSN;
5970 n = NEXT_INSN (n))
5971 {
5972 if (regno_last_uid[REGNO (reg)] == INSN_UID (n))
5973 return 1;
5974 }
5975 return 0;
5976 }
5977 \f
5978 /* Called via `note_stores' to record the initial value of a biv. Here we
5979 just record the location of the set and process it later. */
5980
5981 static void
5982 record_initial (dest, set)
5983 rtx dest;
5984 rtx set;
5985 {
5986 struct iv_class *bl;
5987
5988 if (GET_CODE (dest) != REG
5989 || REGNO (dest) >= max_reg_before_loop
5990 || reg_iv_type[REGNO (dest)] != BASIC_INDUCT)
5991 return;
5992
5993 bl = reg_biv_class[REGNO (dest)];
5994
5995 /* If this is the first set found, record it. */
5996 if (bl->init_insn == 0)
5997 {
5998 bl->init_insn = note_insn;
5999 bl->init_set = set;
6000 }
6001 }
6002 \f
6003 /* If any of the registers in X are "old" and currently have a last use earlier
6004 than INSN, update them to have a last use of INSN. Their actual last use
6005 will be the previous insn but it will not have a valid uid_luid so we can't
6006 use it. */
6007
6008 static void
6009 update_reg_last_use (x, insn)
6010 rtx x;
6011 rtx insn;
6012 {
6013 /* Check for the case where INSN does not have a valid luid. In this case,
6014 there is no need to modify the regno_last_uid, as this can only happen
6015 when code is inserted after the loop_end to set a pseudo's final value,
6016 and hence this insn will never be the last use of x. */
6017 if (GET_CODE (x) == REG && REGNO (x) < max_reg_before_loop
6018 && INSN_UID (insn) < max_uid_for_loop
6019 && uid_luid[regno_last_uid[REGNO (x)]] < uid_luid[INSN_UID (insn)])
6020 regno_last_uid[REGNO (x)] = INSN_UID (insn);
6021 else
6022 {
6023 register int i, j;
6024 register char *fmt = GET_RTX_FORMAT (GET_CODE (x));
6025 for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
6026 {
6027 if (fmt[i] == 'e')
6028 update_reg_last_use (XEXP (x, i), insn);
6029 else if (fmt[i] == 'E')
6030 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
6031 update_reg_last_use (XVECEXP (x, i, j), insn);
6032 }
6033 }
6034 }
6035 \f
6036 /* Given a jump insn JUMP, return the condition that will cause it to branch
6037 to its JUMP_LABEL. If the condition cannot be understood, or is an
6038 inequality floating-point comparison which needs to be reversed, 0 will
6039 be returned.
6040
6041 If EARLIEST is non-zero, it is a pointer to a place where the earliest
6042 insn used in locating the condition was found. If a replacement test
6043 of the condition is desired, it should be placed in front of that
6044 insn and we will be sure that the inputs are still valid.
6045
6046 The condition will be returned in a canonical form to simplify testing by
6047 callers. Specifically:
6048
6049 (1) The code will always be a comparison operation (EQ, NE, GT, etc.).
6050 (2) Both operands will be machine operands; (cc0) will have been replaced.
6051 (3) If an operand is a constant, it will be the second operand.
6052 (4) (LE x const) will be replaced with (LT x <const+1>) and similarly
6053 for GE, GEU, and LEU. */
6054
6055 rtx
6056 get_condition (jump, earliest)
6057 rtx jump;
6058 rtx *earliest;
6059 {
6060 enum rtx_code code;
6061 rtx prev = jump;
6062 rtx set;
6063 rtx tem;
6064 rtx op0, op1;
6065 int reverse_code = 0;
6066 int did_reverse_condition = 0;
6067
6068 /* If this is not a standard conditional jump, we can't parse it. */
6069 if (GET_CODE (jump) != JUMP_INSN
6070 || ! condjump_p (jump) || simplejump_p (jump))
6071 return 0;
6072
6073 code = GET_CODE (XEXP (SET_SRC (PATTERN (jump)), 0));
6074 op0 = XEXP (XEXP (SET_SRC (PATTERN (jump)), 0), 0);
6075 op1 = XEXP (XEXP (SET_SRC (PATTERN (jump)), 0), 1);
6076
6077 if (earliest)
6078 *earliest = jump;
6079
6080 /* If this branches to JUMP_LABEL when the condition is false, reverse
6081 the condition. */
6082 if (XEXP (XEXP (SET_SRC (PATTERN (jump)), 2), 0) == JUMP_LABEL (jump))
6083 code = reverse_condition (code), did_reverse_condition ^= 1;
6084
6085 /* If we are comparing a register with zero, see if the register is set
6086 in the previous insn to a COMPARE or a comparison operation. Perform
6087 the same tests as a function of STORE_FLAG_VALUE as find_comparison_args
6088 in cse.c */
6089
6090 while (GET_RTX_CLASS (code) == '<' && op1 == const0_rtx)
6091 {
6092 /* Set non-zero when we find something of interest. */
6093 rtx x = 0;
6094
6095 #ifdef HAVE_cc0
6096 /* If comparison with cc0, import actual comparison from compare
6097 insn. */
6098 if (op0 == cc0_rtx)
6099 {
6100 if ((prev = prev_nonnote_insn (prev)) == 0
6101 || GET_CODE (prev) != INSN
6102 || (set = single_set (prev)) == 0
6103 || SET_DEST (set) != cc0_rtx)
6104 return 0;
6105
6106 op0 = SET_SRC (set);
6107 op1 = CONST0_RTX (GET_MODE (op0));
6108 if (earliest)
6109 *earliest = prev;
6110 }
6111 #endif
6112
6113 /* If this is a COMPARE, pick up the two things being compared. */
6114 if (GET_CODE (op0) == COMPARE)
6115 {
6116 op1 = XEXP (op0, 1);
6117 op0 = XEXP (op0, 0);
6118 continue;
6119 }
6120 else if (GET_CODE (op0) != REG)
6121 break;
6122
6123 /* Go back to the previous insn. Stop if it is not an INSN. We also
6124 stop if it isn't a single set or if it has a REG_INC note because
6125 we don't want to bother dealing with it. */
6126
6127 if ((prev = prev_nonnote_insn (prev)) == 0
6128 || GET_CODE (prev) != INSN
6129 || FIND_REG_INC_NOTE (prev, 0)
6130 || (set = single_set (prev)) == 0)
6131 break;
6132
6133 /* If this is setting OP0, get what it sets it to if it looks
6134 relevant. */
6135 if (SET_DEST (set) == op0)
6136 {
6137 enum machine_mode inner_mode = GET_MODE (SET_SRC (set));
6138
6139 if ((GET_CODE (SET_SRC (set)) == COMPARE
6140 || ((code == NE
6141 || (code == LT
6142 && GET_MODE_BITSIZE (inner_mode) <= HOST_BITS_PER_INT
6143 && (STORE_FLAG_VALUE
6144 & (1 << (GET_MODE_BITSIZE (inner_mode) - 1)))))
6145 && GET_RTX_CLASS (GET_CODE (SET_SRC (set))) == '<')))
6146 x = SET_SRC (set);
6147 else if ((code == EQ
6148 || (code == GE
6149 && GET_MODE_BITSIZE (inner_mode) <= HOST_BITS_PER_INT
6150 && (STORE_FLAG_VALUE
6151 & (1 << (GET_MODE_BITSIZE (inner_mode) - 1)))))
6152 && GET_RTX_CLASS (GET_CODE (SET_SRC (set))) == '<')
6153 {
6154 /* We might have reversed a LT to get a GE here. But this wasn't
6155 actually the comparison of data, so we don't flag that we
6156 have had to reverse the condition. */
6157 did_reverse_condition ^= 1;
6158 reverse_code = 1;
6159 x = SET_SRC (set);
6160 }
6161 }
6162
6163 else if (reg_set_p (op0, prev))
6164 /* If this sets OP0, but not directly, we have to give up. */
6165 break;
6166
6167 if (x)
6168 {
6169 if (GET_RTX_CLASS (GET_CODE (x)) == '<')
6170 code = GET_CODE (x);
6171 if (reverse_code)
6172 {
6173 code = reverse_condition (code);
6174 did_reverse_condition ^= 1;
6175 reverse_code = 0;
6176 }
6177
6178 op0 = XEXP (x, 0), op1 = XEXP (x, 1);
6179 if (earliest)
6180 *earliest = prev;
6181 }
6182 }
6183
6184 /* If constant is first, put it last. */
6185 if (CONSTANT_P (op0))
6186 code = swap_condition (code), tem = op0, op0 = op1, op1 = tem;
6187
6188 /* If OP0 is the result of a comparison, we weren't able to find what
6189 was really being compared, so fail. */
6190 if (GET_MODE_CLASS (GET_MODE (op0)) == MODE_CC)
6191 return 0;
6192
6193 /* Canonicalize any ordered comparison with integers involving equality. */
6194 if (GET_CODE (op1) == CONST_INT)
6195 {
6196 int const_val = INTVAL (op1);
6197 unsigned uconst_val = (unsigned) const_val;
6198
6199 switch (code)
6200 {
6201 case LE:
6202 code = LT;
6203 op1 = gen_rtx (CONST_INT, VOIDmode, const_val + 1);
6204 break;
6205
6206 case GE:
6207 code = GT;
6208 op1 = gen_rtx (CONST_INT, VOIDmode, const_val - 1);
6209 break;
6210
6211 case LEU:
6212 code = LTU;
6213 op1 = gen_rtx (CONST_INT, VOIDmode, uconst_val + 1);
6214 break;
6215
6216 case GEU:
6217 code = GTU;
6218 op1 = gen_rtx (CONST_INT, VOIDmode, uconst_val - 1);
6219 break;
6220 }
6221 }
6222
6223 /* If this was floating-point and we reversed anything other than an
6224 EQ or NE, return zero. */
6225 if (TARGET_FLOAT_FORMAT == IEEE_FLOAT_FORMAT
6226 && did_reverse_condition && code != NE && code != EQ
6227 && GET_MODE_CLASS (GET_MODE (op0)) == MODE_FLOAT)
6228 return 0;
6229
6230 #ifdef HAVE_cc0
6231 /* Never return CC0; return zero instead. */
6232 if (op0 == cc0_rtx)
6233 return 0;
6234 #endif
6235
6236 return gen_rtx (code, VOIDmode, op0, op1);
6237 }
6238
6239 /* Similar to above routine, except that we also put an invariant last
6240 unless both operands are invariants. */
6241
6242 rtx
6243 get_condition_for_loop (x)
6244 rtx x;
6245 {
6246 rtx comparison = get_condition (x, 0);
6247
6248 if (comparison == 0
6249 || ! invariant_p (XEXP (comparison, 0))
6250 || invariant_p (XEXP (comparison, 1)))
6251 return comparison;
6252
6253 return gen_rtx (swap_condition (GET_CODE (comparison)), VOIDmode,
6254 XEXP (comparison, 1), XEXP (comparison, 0));
6255 }
This page took 0.338426 seconds and 5 git commands to generate.