loop.c (strength_reduce): Adjust BENEFIT appropriately if an autoincrement memory...
[gcc.git] / gcc / loop.c
1 /* Perform various loop optimizations, including strength reduction.
2 Copyright (C) 1987, 88, 89, 91-6, 1997 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, 59 Temple Place - Suite 330,
19 Boston, MA 02111-1307, USA. */
20
21
22 /* This is the loop optimization pass of the compiler.
23 It finds invariant computations within loops and moves them
24 to the beginning of the loop. Then it identifies basic and
25 general induction variables. Strength reduction is applied to the general
26 induction variables, and induction variable elimination is applied to
27 the basic induction variables.
28
29 It also finds cases where
30 a register is set within the loop by zero-extending a narrower value
31 and changes these to zero the entire register once before the loop
32 and merely copy the low part within the loop.
33
34 Most of the complexity is in heuristics to decide when it is worth
35 while to do these things. */
36
37 #include <stdio.h>
38 #include "config.h"
39 #include "rtl.h"
40 #include "obstack.h"
41 #include "expr.h"
42 #include "insn-config.h"
43 #include "insn-flags.h"
44 #include "regs.h"
45 #include "hard-reg-set.h"
46 #include "recog.h"
47 #include "flags.h"
48 #include "real.h"
49 #include "loop.h"
50 #include "except.h"
51
52 /* Vector mapping INSN_UIDs to luids.
53 The luids are like uids but increase monotonically always.
54 We use them to see whether a jump comes from outside a given loop. */
55
56 int *uid_luid;
57
58 /* Indexed by INSN_UID, contains the ordinal giving the (innermost) loop
59 number the insn is contained in. */
60
61 int *uid_loop_num;
62
63 /* 1 + largest uid of any insn. */
64
65 int max_uid_for_loop;
66
67 /* 1 + luid of last insn. */
68
69 static int max_luid;
70
71 /* Number of loops detected in current function. Used as index to the
72 next few tables. */
73
74 static int max_loop_num;
75
76 /* Indexed by loop number, contains the first and last insn of each loop. */
77
78 static rtx *loop_number_loop_starts, *loop_number_loop_ends;
79
80 /* For each loop, gives the containing loop number, -1 if none. */
81
82 int *loop_outer_loop;
83
84 /* Indexed by loop number, contains a nonzero value if the "loop" isn't
85 really a loop (an insn outside the loop branches into it). */
86
87 static char *loop_invalid;
88
89 /* Indexed by loop number, links together all LABEL_REFs which refer to
90 code labels outside the loop. Used by routines that need to know all
91 loop exits, such as final_biv_value and final_giv_value.
92
93 This does not include loop exits due to return instructions. This is
94 because all bivs and givs are pseudos, and hence must be dead after a
95 return, so the presense of a return does not affect any of the
96 optimizations that use this info. It is simpler to just not include return
97 instructions on this list. */
98
99 rtx *loop_number_exit_labels;
100
101 /* Indexed by loop number, counts the number of LABEL_REFs on
102 loop_number_exit_labels for this loop and all loops nested inside it. */
103
104 int *loop_number_exit_count;
105
106 /* Holds the number of loop iterations. It is zero if the number could not be
107 calculated. Must be unsigned since the number of iterations can
108 be as high as 2^wordsize-1. For loops with a wider iterator, this number
109 will will be zero if the number of loop iterations is too large for an
110 unsigned integer to hold. */
111
112 unsigned HOST_WIDE_INT loop_n_iterations;
113
114 /* Nonzero if there is a subroutine call in the current loop.
115 (unknown_address_altered is also nonzero in this case.) */
116
117 static int loop_has_call;
118
119 /* Nonzero if there is a volatile memory reference in the current
120 loop. */
121
122 static int loop_has_volatile;
123
124 /* Added loop_continue which is the NOTE_INSN_LOOP_CONT of the
125 current loop. A continue statement will generate a branch to
126 NEXT_INSN (loop_continue). */
127
128 static rtx loop_continue;
129
130 /* Indexed by register number, contains the number of times the reg
131 is set during the loop being scanned.
132 During code motion, a negative value indicates a reg that has been
133 made a candidate; in particular -2 means that it is an candidate that
134 we know is equal to a constant and -1 means that it is an candidate
135 not known equal to a constant.
136 After code motion, regs moved have 0 (which is accurate now)
137 while the failed candidates have the original number of times set.
138
139 Therefore, at all times, == 0 indicates an invariant register;
140 < 0 a conditionally invariant one. */
141
142 static short *n_times_set;
143
144 /* Original value of n_times_set; same except that this value
145 is not set negative for a reg whose sets have been made candidates
146 and not set to 0 for a reg that is moved. */
147
148 static short *n_times_used;
149
150 /* Index by register number, 1 indicates that the register
151 cannot be moved or strength reduced. */
152
153 static char *may_not_optimize;
154
155 /* Nonzero means reg N has already been moved out of one loop.
156 This reduces the desire to move it out of another. */
157
158 static char *moved_once;
159
160 /* Array of MEMs that are stored in this loop. If there are too many to fit
161 here, we just turn on unknown_address_altered. */
162
163 #define NUM_STORES 20
164 static rtx loop_store_mems[NUM_STORES];
165
166 /* Index of first available slot in above array. */
167 static int loop_store_mems_idx;
168
169 /* Nonzero if we don't know what MEMs were changed in the current loop.
170 This happens if the loop contains a call (in which case `loop_has_call'
171 will also be set) or if we store into more than NUM_STORES MEMs. */
172
173 static int unknown_address_altered;
174
175 /* Count of movable (i.e. invariant) instructions discovered in the loop. */
176 static int num_movables;
177
178 /* Count of memory write instructions discovered in the loop. */
179 static int num_mem_sets;
180
181 /* Number of loops contained within the current one, including itself. */
182 static int loops_enclosed;
183
184 /* Bound on pseudo register number before loop optimization.
185 A pseudo has valid regscan info if its number is < max_reg_before_loop. */
186 int max_reg_before_loop;
187
188 /* This obstack is used in product_cheap_p to allocate its rtl. It
189 may call gen_reg_rtx which, in turn, may reallocate regno_reg_rtx.
190 If we used the same obstack that it did, we would be deallocating
191 that array. */
192
193 static struct obstack temp_obstack;
194
195 /* This is where the pointer to the obstack being used for RTL is stored. */
196
197 extern struct obstack *rtl_obstack;
198
199 #define obstack_chunk_alloc xmalloc
200 #define obstack_chunk_free free
201
202 extern char *oballoc ();
203 \f
204 /* During the analysis of a loop, a chain of `struct movable's
205 is made to record all the movable insns found.
206 Then the entire chain can be scanned to decide which to move. */
207
208 struct movable
209 {
210 rtx insn; /* A movable insn */
211 rtx set_src; /* The expression this reg is set from. */
212 rtx set_dest; /* The destination of this SET. */
213 rtx dependencies; /* When INSN is libcall, this is an EXPR_LIST
214 of any registers used within the LIBCALL. */
215 int consec; /* Number of consecutive following insns
216 that must be moved with this one. */
217 int regno; /* The register it sets */
218 short lifetime; /* lifetime of that register;
219 may be adjusted when matching movables
220 that load the same value are found. */
221 short savings; /* Number of insns we can move for this reg,
222 including other movables that force this
223 or match this one. */
224 unsigned int cond : 1; /* 1 if only conditionally movable */
225 unsigned int force : 1; /* 1 means MUST move this insn */
226 unsigned int global : 1; /* 1 means reg is live outside this loop */
227 /* If PARTIAL is 1, GLOBAL means something different:
228 that the reg is live outside the range from where it is set
229 to the following label. */
230 unsigned int done : 1; /* 1 inhibits further processing of this */
231
232 unsigned int partial : 1; /* 1 means this reg is used for zero-extending.
233 In particular, moving it does not make it
234 invariant. */
235 unsigned int move_insn : 1; /* 1 means that we call emit_move_insn to
236 load SRC, rather than copying INSN. */
237 unsigned int is_equiv : 1; /* 1 means a REG_EQUIV is present on INSN. */
238 enum machine_mode savemode; /* Nonzero means it is a mode for a low part
239 that we should avoid changing when clearing
240 the rest of the reg. */
241 struct movable *match; /* First entry for same value */
242 struct movable *forces; /* An insn that must be moved if this is */
243 struct movable *next;
244 };
245
246 FILE *loop_dump_stream;
247
248 /* Forward declarations. */
249
250 static void find_and_verify_loops ();
251 static void mark_loop_jump ();
252 static void prescan_loop ();
253 static int reg_in_basic_block_p ();
254 static int consec_sets_invariant_p ();
255 static rtx libcall_other_reg ();
256 static int labels_in_range_p ();
257 static void count_loop_regs_set ();
258 static void note_addr_stored ();
259 static int loop_reg_used_before_p ();
260 static void scan_loop ();
261 static void replace_call_address ();
262 static rtx skip_consec_insns ();
263 static int libcall_benefit ();
264 static void ignore_some_movables ();
265 static void force_movables ();
266 static void combine_movables ();
267 static int rtx_equal_for_loop_p ();
268 static void move_movables ();
269 static void strength_reduce ();
270 static int valid_initial_value_p ();
271 static void find_mem_givs ();
272 static void record_biv ();
273 static void check_final_value ();
274 static void record_giv ();
275 static void update_giv_derive ();
276 static int basic_induction_var ();
277 static rtx simplify_giv_expr ();
278 static int general_induction_var ();
279 static int consec_sets_giv ();
280 static int check_dbra_loop ();
281 static rtx express_from ();
282 static int combine_givs_p ();
283 static void combine_givs ();
284 static int product_cheap_p ();
285 static int maybe_eliminate_biv ();
286 static int maybe_eliminate_biv_1 ();
287 static int last_use_this_basic_block ();
288 static void record_initial ();
289 static void update_reg_last_use ();
290 \f
291 /* Relative gain of eliminating various kinds of operations. */
292 int add_cost;
293 #if 0
294 int shift_cost;
295 int mult_cost;
296 #endif
297
298 /* Benefit penalty, if a giv is not replaceable, i.e. must emit an insn to
299 copy the value of the strength reduced giv to its original register. */
300 int copy_cost;
301
302 void
303 init_loop ()
304 {
305 char *free_point = (char *) oballoc (1);
306 rtx reg = gen_rtx (REG, word_mode, LAST_VIRTUAL_REGISTER + 1);
307
308 add_cost = rtx_cost (gen_rtx (PLUS, word_mode, reg, reg), SET);
309
310 /* We multiply by 2 to reconcile the difference in scale between
311 these two ways of computing costs. Otherwise the cost of a copy
312 will be far less than the cost of an add. */
313
314 copy_cost = 2 * 2;
315
316 /* Free the objects we just allocated. */
317 obfree (free_point);
318
319 /* Initialize the obstack used for rtl in product_cheap_p. */
320 gcc_obstack_init (&temp_obstack);
321 }
322 \f
323 /* Entry point of this file. Perform loop optimization
324 on the current function. F is the first insn of the function
325 and DUMPFILE is a stream for output of a trace of actions taken
326 (or 0 if none should be output). */
327
328 void
329 loop_optimize (f, dumpfile)
330 /* f is the first instruction of a chain of insns for one function */
331 rtx f;
332 FILE *dumpfile;
333 {
334 register rtx insn;
335 register int i;
336 rtx last_insn;
337
338 loop_dump_stream = dumpfile;
339
340 init_recog_no_volatile ();
341 init_alias_analysis ();
342
343 max_reg_before_loop = max_reg_num ();
344
345 moved_once = (char *) alloca (max_reg_before_loop);
346 bzero (moved_once, max_reg_before_loop);
347
348 regs_may_share = 0;
349
350 /* Count the number of loops. */
351
352 max_loop_num = 0;
353 for (insn = f; insn; insn = NEXT_INSN (insn))
354 {
355 if (GET_CODE (insn) == NOTE
356 && NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG)
357 max_loop_num++;
358 }
359
360 /* Don't waste time if no loops. */
361 if (max_loop_num == 0)
362 return;
363
364 /* Get size to use for tables indexed by uids.
365 Leave some space for labels allocated by find_and_verify_loops. */
366 max_uid_for_loop = get_max_uid () + 1 + max_loop_num * 32;
367
368 uid_luid = (int *) alloca (max_uid_for_loop * sizeof (int));
369 uid_loop_num = (int *) alloca (max_uid_for_loop * sizeof (int));
370
371 bzero ((char *) uid_luid, max_uid_for_loop * sizeof (int));
372 bzero ((char *) uid_loop_num, max_uid_for_loop * sizeof (int));
373
374 /* Allocate tables for recording each loop. We set each entry, so they need
375 not be zeroed. */
376 loop_number_loop_starts = (rtx *) alloca (max_loop_num * sizeof (rtx));
377 loop_number_loop_ends = (rtx *) alloca (max_loop_num * sizeof (rtx));
378 loop_outer_loop = (int *) alloca (max_loop_num * sizeof (int));
379 loop_invalid = (char *) alloca (max_loop_num * sizeof (char));
380 loop_number_exit_labels = (rtx *) alloca (max_loop_num * sizeof (rtx));
381 loop_number_exit_count = (int *) alloca (max_loop_num * sizeof (int));
382
383 /* Find and process each loop.
384 First, find them, and record them in order of their beginnings. */
385 find_and_verify_loops (f);
386
387 /* Now find all register lifetimes. This must be done after
388 find_and_verify_loops, because it might reorder the insns in the
389 function. */
390 reg_scan (f, max_reg_num (), 1);
391
392 /* See if we went too far. */
393 if (get_max_uid () > max_uid_for_loop)
394 abort ();
395
396 /* Compute the mapping from uids to luids.
397 LUIDs are numbers assigned to insns, like uids,
398 except that luids increase monotonically through the code.
399 Don't assign luids to line-number NOTEs, so that the distance in luids
400 between two insns is not affected by -g. */
401
402 for (insn = f, i = 0; insn; insn = NEXT_INSN (insn))
403 {
404 last_insn = insn;
405 if (GET_CODE (insn) != NOTE
406 || NOTE_LINE_NUMBER (insn) <= 0)
407 uid_luid[INSN_UID (insn)] = ++i;
408 else
409 /* Give a line number note the same luid as preceding insn. */
410 uid_luid[INSN_UID (insn)] = i;
411 }
412
413 max_luid = i + 1;
414
415 /* Don't leave gaps in uid_luid for insns that have been
416 deleted. It is possible that the first or last insn
417 using some register has been deleted by cross-jumping.
418 Make sure that uid_luid for that former insn's uid
419 points to the general area where that insn used to be. */
420 for (i = 0; i < max_uid_for_loop; i++)
421 {
422 uid_luid[0] = uid_luid[i];
423 if (uid_luid[0] != 0)
424 break;
425 }
426 for (i = 0; i < max_uid_for_loop; i++)
427 if (uid_luid[i] == 0)
428 uid_luid[i] = uid_luid[i - 1];
429
430 /* Create a mapping from loops to BLOCK tree nodes. */
431 if (flag_unroll_loops && write_symbols != NO_DEBUG)
432 find_loop_tree_blocks ();
433
434 /* Now scan the loops, last ones first, since this means inner ones are done
435 before outer ones. */
436 for (i = max_loop_num-1; i >= 0; i--)
437 if (! loop_invalid[i] && loop_number_loop_ends[i])
438 scan_loop (loop_number_loop_starts[i], loop_number_loop_ends[i],
439 max_reg_num ());
440
441 /* If debugging and unrolling loops, we must replicate the tree nodes
442 corresponding to the blocks inside the loop, so that the original one
443 to one mapping will remain. */
444 if (flag_unroll_loops && write_symbols != NO_DEBUG)
445 unroll_block_trees ();
446 }
447 \f
448 /* Optimize one loop whose start is LOOP_START and end is END.
449 LOOP_START is the NOTE_INSN_LOOP_BEG and END is the matching
450 NOTE_INSN_LOOP_END. */
451
452 /* ??? Could also move memory writes out of loops if the destination address
453 is invariant, the source is invariant, the memory write is not volatile,
454 and if we can prove that no read inside the loop can read this address
455 before the write occurs. If there is a read of this address after the
456 write, then we can also mark the memory read as invariant. */
457
458 static void
459 scan_loop (loop_start, end, nregs)
460 rtx loop_start, end;
461 int nregs;
462 {
463 register int i;
464 register rtx p;
465 /* 1 if we are scanning insns that could be executed zero times. */
466 int maybe_never = 0;
467 /* 1 if we are scanning insns that might never be executed
468 due to a subroutine call which might exit before they are reached. */
469 int call_passed = 0;
470 /* For a rotated loop that is entered near the bottom,
471 this is the label at the top. Otherwise it is zero. */
472 rtx loop_top = 0;
473 /* Jump insn that enters the loop, or 0 if control drops in. */
474 rtx loop_entry_jump = 0;
475 /* Place in the loop where control enters. */
476 rtx scan_start;
477 /* Number of insns in the loop. */
478 int insn_count;
479 int in_libcall = 0;
480 int tem;
481 rtx temp;
482 /* The SET from an insn, if it is the only SET in the insn. */
483 rtx set, set1;
484 /* Chain describing insns movable in current loop. */
485 struct movable *movables = 0;
486 /* Last element in `movables' -- so we can add elements at the end. */
487 struct movable *last_movable = 0;
488 /* Ratio of extra register life span we can justify
489 for saving an instruction. More if loop doesn't call subroutines
490 since in that case saving an insn makes more difference
491 and more registers are available. */
492 int threshold;
493 /* If we have calls, contains the insn in which a register was used
494 if it was used exactly once; contains const0_rtx if it was used more
495 than once. */
496 rtx *reg_single_usage = 0;
497 /* Nonzero if we are scanning instructions in a sub-loop. */
498 int loop_depth = 0;
499
500 n_times_set = (short *) alloca (nregs * sizeof (short));
501 n_times_used = (short *) alloca (nregs * sizeof (short));
502 may_not_optimize = (char *) alloca (nregs);
503
504 /* Determine whether this loop starts with a jump down to a test at
505 the end. This will occur for a small number of loops with a test
506 that is too complex to duplicate in front of the loop.
507
508 We search for the first insn or label in the loop, skipping NOTEs.
509 However, we must be careful not to skip past a NOTE_INSN_LOOP_BEG
510 (because we might have a loop executed only once that contains a
511 loop which starts with a jump to its exit test) or a NOTE_INSN_LOOP_END
512 (in case we have a degenerate loop).
513
514 Note that if we mistakenly think that a loop is entered at the top
515 when, in fact, it is entered at the exit test, the only effect will be
516 slightly poorer optimization. Making the opposite error can generate
517 incorrect code. Since very few loops now start with a jump to the
518 exit test, the code here to detect that case is very conservative. */
519
520 for (p = NEXT_INSN (loop_start);
521 p != end
522 && GET_CODE (p) != CODE_LABEL && GET_RTX_CLASS (GET_CODE (p)) != 'i'
523 && (GET_CODE (p) != NOTE
524 || (NOTE_LINE_NUMBER (p) != NOTE_INSN_LOOP_BEG
525 && NOTE_LINE_NUMBER (p) != NOTE_INSN_LOOP_END));
526 p = NEXT_INSN (p))
527 ;
528
529 scan_start = p;
530
531 /* Set up variables describing this loop. */
532 prescan_loop (loop_start, end);
533 threshold = (loop_has_call ? 1 : 2) * (1 + n_non_fixed_regs);
534
535 /* If loop has a jump before the first label,
536 the true entry is the target of that jump.
537 Start scan from there.
538 But record in LOOP_TOP the place where the end-test jumps
539 back to so we can scan that after the end of the loop. */
540 if (GET_CODE (p) == JUMP_INSN)
541 {
542 loop_entry_jump = p;
543
544 /* Loop entry must be unconditional jump (and not a RETURN) */
545 if (simplejump_p (p)
546 && JUMP_LABEL (p) != 0
547 /* Check to see whether the jump actually
548 jumps out of the loop (meaning it's no loop).
549 This case can happen for things like
550 do {..} while (0). If this label was generated previously
551 by loop, we can't tell anything about it and have to reject
552 the loop. */
553 && INSN_UID (JUMP_LABEL (p)) < max_uid_for_loop
554 && INSN_LUID (JUMP_LABEL (p)) >= INSN_LUID (loop_start)
555 && INSN_LUID (JUMP_LABEL (p)) < INSN_LUID (end))
556 {
557 loop_top = next_label (scan_start);
558 scan_start = JUMP_LABEL (p);
559 }
560 }
561
562 /* If SCAN_START was an insn created by loop, we don't know its luid
563 as required by loop_reg_used_before_p. So skip such loops. (This
564 test may never be true, but it's best to play it safe.)
565
566 Also, skip loops where we do not start scanning at a label. This
567 test also rejects loops starting with a JUMP_INSN that failed the
568 test above. */
569
570 if (INSN_UID (scan_start) >= max_uid_for_loop
571 || GET_CODE (scan_start) != CODE_LABEL)
572 {
573 if (loop_dump_stream)
574 fprintf (loop_dump_stream, "\nLoop from %d to %d is phony.\n\n",
575 INSN_UID (loop_start), INSN_UID (end));
576 return;
577 }
578
579 /* Count number of times each reg is set during this loop.
580 Set may_not_optimize[I] if it is not safe to move out
581 the setting of register I. If this loop has calls, set
582 reg_single_usage[I]. */
583
584 bzero ((char *) n_times_set, nregs * sizeof (short));
585 bzero (may_not_optimize, nregs);
586
587 if (loop_has_call)
588 {
589 reg_single_usage = (rtx *) alloca (nregs * sizeof (rtx));
590 bzero ((char *) reg_single_usage, nregs * sizeof (rtx));
591 }
592
593 count_loop_regs_set (loop_top ? loop_top : loop_start, end,
594 may_not_optimize, reg_single_usage, &insn_count, nregs);
595
596 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
597 may_not_optimize[i] = 1, n_times_set[i] = 1;
598 bcopy ((char *) n_times_set, (char *) n_times_used, nregs * sizeof (short));
599
600 if (loop_dump_stream)
601 {
602 fprintf (loop_dump_stream, "\nLoop from %d to %d: %d real insns.\n",
603 INSN_UID (loop_start), INSN_UID (end), insn_count);
604 if (loop_continue)
605 fprintf (loop_dump_stream, "Continue at insn %d.\n",
606 INSN_UID (loop_continue));
607 }
608
609 /* Scan through the loop finding insns that are safe to move.
610 Set n_times_set negative for the reg being set, so that
611 this reg will be considered invariant for subsequent insns.
612 We consider whether subsequent insns use the reg
613 in deciding whether it is worth actually moving.
614
615 MAYBE_NEVER is nonzero if we have passed a conditional jump insn
616 and therefore it is possible that the insns we are scanning
617 would never be executed. At such times, we must make sure
618 that it is safe to execute the insn once instead of zero times.
619 When MAYBE_NEVER is 0, all insns will be executed at least once
620 so that is not a problem. */
621
622 p = scan_start;
623 while (1)
624 {
625 p = NEXT_INSN (p);
626 /* At end of a straight-in loop, we are done.
627 At end of a loop entered at the bottom, scan the top. */
628 if (p == scan_start)
629 break;
630 if (p == end)
631 {
632 if (loop_top != 0)
633 p = loop_top;
634 else
635 break;
636 if (p == scan_start)
637 break;
638 }
639
640 if (GET_RTX_CLASS (GET_CODE (p)) == 'i'
641 && find_reg_note (p, REG_LIBCALL, NULL_RTX))
642 in_libcall = 1;
643 else if (GET_RTX_CLASS (GET_CODE (p)) == 'i'
644 && find_reg_note (p, REG_RETVAL, NULL_RTX))
645 in_libcall = 0;
646
647 if (GET_CODE (p) == INSN
648 && (set = single_set (p))
649 && GET_CODE (SET_DEST (set)) == REG
650 && ! may_not_optimize[REGNO (SET_DEST (set))])
651 {
652 int tem1 = 0;
653 int tem2 = 0;
654 int move_insn = 0;
655 rtx src = SET_SRC (set);
656 rtx dependencies = 0;
657
658 /* Figure out what to use as a source of this insn. If a REG_EQUIV
659 note is given or if a REG_EQUAL note with a constant operand is
660 specified, use it as the source and mark that we should move
661 this insn by calling emit_move_insn rather that duplicating the
662 insn.
663
664 Otherwise, only use the REG_EQUAL contents if a REG_RETVAL note
665 is present. */
666 temp = find_reg_note (p, REG_EQUIV, NULL_RTX);
667 if (temp)
668 src = XEXP (temp, 0), move_insn = 1;
669 else
670 {
671 temp = find_reg_note (p, REG_EQUAL, NULL_RTX);
672 if (temp && CONSTANT_P (XEXP (temp, 0)))
673 src = XEXP (temp, 0), move_insn = 1;
674 if (temp && find_reg_note (p, REG_RETVAL, NULL_RTX))
675 {
676 src = XEXP (temp, 0);
677 /* A libcall block can use regs that don't appear in
678 the equivalent expression. To move the libcall,
679 we must move those regs too. */
680 dependencies = libcall_other_reg (p, src);
681 }
682 }
683
684 /* Don't try to optimize a register that was made
685 by loop-optimization for an inner loop.
686 We don't know its life-span, so we can't compute the benefit. */
687 if (REGNO (SET_DEST (set)) >= max_reg_before_loop)
688 ;
689 /* In order to move a register, we need to have one of three cases:
690 (1) it is used only in the same basic block as the set
691 (2) it is not a user variable and it is not used in the
692 exit test (this can cause the variable to be used
693 before it is set just like a user-variable).
694 (3) the set is guaranteed to be executed once the loop starts,
695 and the reg is not used until after that. */
696 else if (! ((! maybe_never
697 && ! loop_reg_used_before_p (set, p, loop_start,
698 scan_start, end))
699 || (! REG_USERVAR_P (SET_DEST (set))
700 && ! REG_LOOP_TEST_P (SET_DEST (set)))
701 || reg_in_basic_block_p (p, SET_DEST (set))))
702 ;
703 else if ((tem = invariant_p (src))
704 && (dependencies == 0
705 || (tem2 = invariant_p (dependencies)) != 0)
706 && (n_times_set[REGNO (SET_DEST (set))] == 1
707 || (tem1
708 = consec_sets_invariant_p (SET_DEST (set),
709 n_times_set[REGNO (SET_DEST (set))],
710 p)))
711 /* If the insn can cause a trap (such as divide by zero),
712 can't move it unless it's guaranteed to be executed
713 once loop is entered. Even a function call might
714 prevent the trap insn from being reached
715 (since it might exit!) */
716 && ! ((maybe_never || call_passed)
717 && may_trap_p (src)))
718 {
719 register struct movable *m;
720 register int regno = REGNO (SET_DEST (set));
721
722 /* A potential lossage is where we have a case where two insns
723 can be combined as long as they are both in the loop, but
724 we move one of them outside the loop. For large loops,
725 this can lose. The most common case of this is the address
726 of a function being called.
727
728 Therefore, if this register is marked as being used exactly
729 once if we are in a loop with calls (a "large loop"), see if
730 we can replace the usage of this register with the source
731 of this SET. If we can, delete this insn.
732
733 Don't do this if P has a REG_RETVAL note or if we have
734 SMALL_REGISTER_CLASSES and SET_SRC is a hard register. */
735
736 if (reg_single_usage && reg_single_usage[regno] != 0
737 && reg_single_usage[regno] != const0_rtx
738 && regno_first_uid[regno] == INSN_UID (p)
739 && (regno_last_uid[regno]
740 == INSN_UID (reg_single_usage[regno]))
741 && n_times_set[REGNO (SET_DEST (set))] == 1
742 && ! side_effects_p (SET_SRC (set))
743 && ! find_reg_note (p, REG_RETVAL, NULL_RTX)
744 #ifdef SMALL_REGISTER_CLASSES
745 && ! (SMALL_REGISTER_CLASSES
746 && GET_CODE (SET_SRC (set)) == REG
747 && REGNO (SET_SRC (set)) < FIRST_PSEUDO_REGISTER)
748 #endif
749 /* This test is not redundant; SET_SRC (set) might be
750 a call-clobbered register and the life of REGNO
751 might span a call. */
752 && ! modified_between_p (SET_SRC (set), p,
753 reg_single_usage[regno])
754 && no_labels_between_p (p, reg_single_usage[regno])
755 && validate_replace_rtx (SET_DEST (set), SET_SRC (set),
756 reg_single_usage[regno]))
757 {
758 /* Replace any usage in a REG_EQUAL note. Must copy the
759 new source, so that we don't get rtx sharing between the
760 SET_SOURCE and REG_NOTES of insn p. */
761 REG_NOTES (reg_single_usage[regno])
762 = replace_rtx (REG_NOTES (reg_single_usage[regno]),
763 SET_DEST (set), copy_rtx (SET_SRC (set)));
764
765 PUT_CODE (p, NOTE);
766 NOTE_LINE_NUMBER (p) = NOTE_INSN_DELETED;
767 NOTE_SOURCE_FILE (p) = 0;
768 n_times_set[regno] = 0;
769 continue;
770 }
771
772 m = (struct movable *) alloca (sizeof (struct movable));
773 m->next = 0;
774 m->insn = p;
775 m->set_src = src;
776 m->dependencies = dependencies;
777 m->set_dest = SET_DEST (set);
778 m->force = 0;
779 m->consec = n_times_set[REGNO (SET_DEST (set))] - 1;
780 m->done = 0;
781 m->forces = 0;
782 m->partial = 0;
783 m->move_insn = move_insn;
784 m->is_equiv = (find_reg_note (p, REG_EQUIV, NULL_RTX) != 0);
785 m->savemode = VOIDmode;
786 m->regno = regno;
787 /* Set M->cond if either invariant_p or consec_sets_invariant_p
788 returned 2 (only conditionally invariant). */
789 m->cond = ((tem | tem1 | tem2) > 1);
790 m->global = (uid_luid[regno_last_uid[regno]] > INSN_LUID (end)
791 || uid_luid[regno_first_uid[regno]] < INSN_LUID (loop_start));
792 m->match = 0;
793 m->lifetime = (uid_luid[regno_last_uid[regno]]
794 - uid_luid[regno_first_uid[regno]]);
795 m->savings = n_times_used[regno];
796 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
797 m->savings += libcall_benefit (p);
798 n_times_set[regno] = move_insn ? -2 : -1;
799 /* Add M to the end of the chain MOVABLES. */
800 if (movables == 0)
801 movables = m;
802 else
803 last_movable->next = m;
804 last_movable = m;
805
806 if (m->consec > 0)
807 {
808 /* Skip this insn, not checking REG_LIBCALL notes. */
809 p = next_nonnote_insn (p);
810 /* Skip the consecutive insns, if there are any. */
811 p = skip_consec_insns (p, m->consec);
812 /* Back up to the last insn of the consecutive group. */
813 p = prev_nonnote_insn (p);
814
815 /* We must now reset m->move_insn, m->is_equiv, and possibly
816 m->set_src to correspond to the effects of all the
817 insns. */
818 temp = find_reg_note (p, REG_EQUIV, NULL_RTX);
819 if (temp)
820 m->set_src = XEXP (temp, 0), m->move_insn = 1;
821 else
822 {
823 temp = find_reg_note (p, REG_EQUAL, NULL_RTX);
824 if (temp && CONSTANT_P (XEXP (temp, 0)))
825 m->set_src = XEXP (temp, 0), m->move_insn = 1;
826 else
827 m->move_insn = 0;
828
829 }
830 m->is_equiv = (find_reg_note (p, REG_EQUIV, NULL_RTX) != 0);
831 }
832 }
833 /* If this register is always set within a STRICT_LOW_PART
834 or set to zero, then its high bytes are constant.
835 So clear them outside the loop and within the loop
836 just load the low bytes.
837 We must check that the machine has an instruction to do so.
838 Also, if the value loaded into the register
839 depends on the same register, this cannot be done. */
840 else if (SET_SRC (set) == const0_rtx
841 && GET_CODE (NEXT_INSN (p)) == INSN
842 && (set1 = single_set (NEXT_INSN (p)))
843 && GET_CODE (set1) == SET
844 && (GET_CODE (SET_DEST (set1)) == STRICT_LOW_PART)
845 && (GET_CODE (XEXP (SET_DEST (set1), 0)) == SUBREG)
846 && (SUBREG_REG (XEXP (SET_DEST (set1), 0))
847 == SET_DEST (set))
848 && !reg_mentioned_p (SET_DEST (set), SET_SRC (set1)))
849 {
850 register int regno = REGNO (SET_DEST (set));
851 if (n_times_set[regno] == 2)
852 {
853 register struct movable *m;
854 m = (struct movable *) alloca (sizeof (struct movable));
855 m->next = 0;
856 m->insn = p;
857 m->set_dest = SET_DEST (set);
858 m->dependencies = 0;
859 m->force = 0;
860 m->consec = 0;
861 m->done = 0;
862 m->forces = 0;
863 m->move_insn = 0;
864 m->partial = 1;
865 /* If the insn may not be executed on some cycles,
866 we can't clear the whole reg; clear just high part.
867 Not even if the reg is used only within this loop.
868 Consider this:
869 while (1)
870 while (s != t) {
871 if (foo ()) x = *s;
872 use (x);
873 }
874 Clearing x before the inner loop could clobber a value
875 being saved from the last time around the outer loop.
876 However, if the reg is not used outside this loop
877 and all uses of the register are in the same
878 basic block as the store, there is no problem.
879
880 If this insn was made by loop, we don't know its
881 INSN_LUID and hence must make a conservative
882 assumption. */
883 m->global = (INSN_UID (p) >= max_uid_for_loop
884 || (uid_luid[regno_last_uid[regno]]
885 > INSN_LUID (end))
886 || (uid_luid[regno_first_uid[regno]]
887 < INSN_LUID (p))
888 || (labels_in_range_p
889 (p, uid_luid[regno_first_uid[regno]])));
890 if (maybe_never && m->global)
891 m->savemode = GET_MODE (SET_SRC (set1));
892 else
893 m->savemode = VOIDmode;
894 m->regno = regno;
895 m->cond = 0;
896 m->match = 0;
897 m->lifetime = (uid_luid[regno_last_uid[regno]]
898 - uid_luid[regno_first_uid[regno]]);
899 m->savings = 1;
900 n_times_set[regno] = -1;
901 /* Add M to the end of the chain MOVABLES. */
902 if (movables == 0)
903 movables = m;
904 else
905 last_movable->next = m;
906 last_movable = m;
907 }
908 }
909 }
910 /* Past a call insn, we get to insns which might not be executed
911 because the call might exit. This matters for insns that trap.
912 Call insns inside a REG_LIBCALL/REG_RETVAL block always return,
913 so they don't count. */
914 else if (GET_CODE (p) == CALL_INSN && ! in_libcall)
915 call_passed = 1;
916 /* Past a label or a jump, we get to insns for which we
917 can't count on whether or how many times they will be
918 executed during each iteration. Therefore, we can
919 only move out sets of trivial variables
920 (those not used after the loop). */
921 /* Similar code appears twice in strength_reduce. */
922 else if ((GET_CODE (p) == CODE_LABEL || GET_CODE (p) == JUMP_INSN)
923 /* If we enter the loop in the middle, and scan around to the
924 beginning, don't set maybe_never for that. This must be an
925 unconditional jump, otherwise the code at the top of the
926 loop might never be executed. Unconditional jumps are
927 followed a by barrier then loop end. */
928 && ! (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) == loop_top
929 && NEXT_INSN (NEXT_INSN (p)) == end
930 && simplejump_p (p)))
931 maybe_never = 1;
932 else if (GET_CODE (p) == NOTE)
933 {
934 /* At the virtual top of a converted loop, insns are again known to
935 be executed: logically, the loop begins here even though the exit
936 code has been duplicated. */
937 if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP && loop_depth == 0)
938 maybe_never = call_passed = 0;
939 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
940 loop_depth++;
941 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END)
942 loop_depth--;
943 }
944 }
945
946 /* If one movable subsumes another, ignore that other. */
947
948 ignore_some_movables (movables);
949
950 /* For each movable insn, see if the reg that it loads
951 leads when it dies right into another conditionally movable insn.
952 If so, record that the second insn "forces" the first one,
953 since the second can be moved only if the first is. */
954
955 force_movables (movables);
956
957 /* See if there are multiple movable insns that load the same value.
958 If there are, make all but the first point at the first one
959 through the `match' field, and add the priorities of them
960 all together as the priority of the first. */
961
962 combine_movables (movables, nregs);
963
964 /* Now consider each movable insn to decide whether it is worth moving.
965 Store 0 in n_times_set for each reg that is moved. */
966
967 move_movables (movables, threshold,
968 insn_count, loop_start, end, nregs);
969
970 /* Now candidates that still are negative are those not moved.
971 Change n_times_set to indicate that those are not actually invariant. */
972 for (i = 0; i < nregs; i++)
973 if (n_times_set[i] < 0)
974 n_times_set[i] = n_times_used[i];
975
976 if (flag_strength_reduce)
977 strength_reduce (scan_start, end, loop_top,
978 insn_count, loop_start, end);
979 }
980 \f
981 /* Add elements to *OUTPUT to record all the pseudo-regs
982 mentioned in IN_THIS but not mentioned in NOT_IN_THIS. */
983
984 void
985 record_excess_regs (in_this, not_in_this, output)
986 rtx in_this, not_in_this;
987 rtx *output;
988 {
989 enum rtx_code code;
990 char *fmt;
991 int i;
992
993 code = GET_CODE (in_this);
994
995 switch (code)
996 {
997 case PC:
998 case CC0:
999 case CONST_INT:
1000 case CONST_DOUBLE:
1001 case CONST:
1002 case SYMBOL_REF:
1003 case LABEL_REF:
1004 return;
1005
1006 case REG:
1007 if (REGNO (in_this) >= FIRST_PSEUDO_REGISTER
1008 && ! reg_mentioned_p (in_this, not_in_this))
1009 *output = gen_rtx (EXPR_LIST, VOIDmode, in_this, *output);
1010 return;
1011 }
1012
1013 fmt = GET_RTX_FORMAT (code);
1014 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1015 {
1016 int j;
1017
1018 switch (fmt[i])
1019 {
1020 case 'E':
1021 for (j = 0; j < XVECLEN (in_this, i); j++)
1022 record_excess_regs (XVECEXP (in_this, i, j), not_in_this, output);
1023 break;
1024
1025 case 'e':
1026 record_excess_regs (XEXP (in_this, i), not_in_this, output);
1027 break;
1028 }
1029 }
1030 }
1031 \f
1032 /* Check what regs are referred to in the libcall block ending with INSN,
1033 aside from those mentioned in the equivalent value.
1034 If there are none, return 0.
1035 If there are one or more, return an EXPR_LIST containing all of them. */
1036
1037 static rtx
1038 libcall_other_reg (insn, equiv)
1039 rtx insn, equiv;
1040 {
1041 rtx note = find_reg_note (insn, REG_RETVAL, NULL_RTX);
1042 rtx p = XEXP (note, 0);
1043 rtx output = 0;
1044
1045 /* First, find all the regs used in the libcall block
1046 that are not mentioned as inputs to the result. */
1047
1048 while (p != insn)
1049 {
1050 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
1051 || GET_CODE (p) == CALL_INSN)
1052 record_excess_regs (PATTERN (p), equiv, &output);
1053 p = NEXT_INSN (p);
1054 }
1055
1056 return output;
1057 }
1058 \f
1059 /* Return 1 if all uses of REG
1060 are between INSN and the end of the basic block. */
1061
1062 static int
1063 reg_in_basic_block_p (insn, reg)
1064 rtx insn, reg;
1065 {
1066 int regno = REGNO (reg);
1067 rtx p;
1068
1069 if (regno_first_uid[regno] != INSN_UID (insn))
1070 return 0;
1071
1072 /* Search this basic block for the already recorded last use of the reg. */
1073 for (p = insn; p; p = NEXT_INSN (p))
1074 {
1075 switch (GET_CODE (p))
1076 {
1077 case NOTE:
1078 break;
1079
1080 case INSN:
1081 case CALL_INSN:
1082 /* Ordinary insn: if this is the last use, we win. */
1083 if (regno_last_uid[regno] == INSN_UID (p))
1084 return 1;
1085 break;
1086
1087 case JUMP_INSN:
1088 /* Jump insn: if this is the last use, we win. */
1089 if (regno_last_uid[regno] == INSN_UID (p))
1090 return 1;
1091 /* Otherwise, it's the end of the basic block, so we lose. */
1092 return 0;
1093
1094 case CODE_LABEL:
1095 case BARRIER:
1096 /* It's the end of the basic block, so we lose. */
1097 return 0;
1098 }
1099 }
1100
1101 /* The "last use" doesn't follow the "first use"?? */
1102 abort ();
1103 }
1104 \f
1105 /* Compute the benefit of eliminating the insns in the block whose
1106 last insn is LAST. This may be a group of insns used to compute a
1107 value directly or can contain a library call. */
1108
1109 static int
1110 libcall_benefit (last)
1111 rtx last;
1112 {
1113 rtx insn;
1114 int benefit = 0;
1115
1116 for (insn = XEXP (find_reg_note (last, REG_RETVAL, NULL_RTX), 0);
1117 insn != last; insn = NEXT_INSN (insn))
1118 {
1119 if (GET_CODE (insn) == CALL_INSN)
1120 benefit += 10; /* Assume at least this many insns in a library
1121 routine. */
1122 else if (GET_CODE (insn) == INSN
1123 && GET_CODE (PATTERN (insn)) != USE
1124 && GET_CODE (PATTERN (insn)) != CLOBBER)
1125 benefit++;
1126 }
1127
1128 return benefit;
1129 }
1130 \f
1131 /* Skip COUNT insns from INSN, counting library calls as 1 insn. */
1132
1133 static rtx
1134 skip_consec_insns (insn, count)
1135 rtx insn;
1136 int count;
1137 {
1138 for (; count > 0; count--)
1139 {
1140 rtx temp;
1141
1142 /* If first insn of libcall sequence, skip to end. */
1143 /* Do this at start of loop, since INSN is guaranteed to
1144 be an insn here. */
1145 if (GET_CODE (insn) != NOTE
1146 && (temp = find_reg_note (insn, REG_LIBCALL, NULL_RTX)))
1147 insn = XEXP (temp, 0);
1148
1149 do insn = NEXT_INSN (insn);
1150 while (GET_CODE (insn) == NOTE);
1151 }
1152
1153 return insn;
1154 }
1155
1156 /* Ignore any movable whose insn falls within a libcall
1157 which is part of another movable.
1158 We make use of the fact that the movable for the libcall value
1159 was made later and so appears later on the chain. */
1160
1161 static void
1162 ignore_some_movables (movables)
1163 struct movable *movables;
1164 {
1165 register struct movable *m, *m1;
1166
1167 for (m = movables; m; m = m->next)
1168 {
1169 /* Is this a movable for the value of a libcall? */
1170 rtx note = find_reg_note (m->insn, REG_RETVAL, NULL_RTX);
1171 if (note)
1172 {
1173 rtx insn;
1174 /* Check for earlier movables inside that range,
1175 and mark them invalid. We cannot use LUIDs here because
1176 insns created by loop.c for prior loops don't have LUIDs.
1177 Rather than reject all such insns from movables, we just
1178 explicitly check each insn in the libcall (since invariant
1179 libcalls aren't that common). */
1180 for (insn = XEXP (note, 0); insn != m->insn; insn = NEXT_INSN (insn))
1181 for (m1 = movables; m1 != m; m1 = m1->next)
1182 if (m1->insn == insn)
1183 m1->done = 1;
1184 }
1185 }
1186 }
1187
1188 /* For each movable insn, see if the reg that it loads
1189 leads when it dies right into another conditionally movable insn.
1190 If so, record that the second insn "forces" the first one,
1191 since the second can be moved only if the first is. */
1192
1193 static void
1194 force_movables (movables)
1195 struct movable *movables;
1196 {
1197 register struct movable *m, *m1;
1198 for (m1 = movables; m1; m1 = m1->next)
1199 /* Omit this if moving just the (SET (REG) 0) of a zero-extend. */
1200 if (!m1->partial && !m1->done)
1201 {
1202 int regno = m1->regno;
1203 for (m = m1->next; m; m = m->next)
1204 /* ??? Could this be a bug? What if CSE caused the
1205 register of M1 to be used after this insn?
1206 Since CSE does not update regno_last_uid,
1207 this insn M->insn might not be where it dies.
1208 But very likely this doesn't matter; what matters is
1209 that M's reg is computed from M1's reg. */
1210 if (INSN_UID (m->insn) == regno_last_uid[regno]
1211 && !m->done)
1212 break;
1213 if (m != 0 && m->set_src == m1->set_dest
1214 /* If m->consec, m->set_src isn't valid. */
1215 && m->consec == 0)
1216 m = 0;
1217
1218 /* Increase the priority of the moving the first insn
1219 since it permits the second to be moved as well. */
1220 if (m != 0)
1221 {
1222 m->forces = m1;
1223 m1->lifetime += m->lifetime;
1224 m1->savings += m1->savings;
1225 }
1226 }
1227 }
1228 \f
1229 /* Find invariant expressions that are equal and can be combined into
1230 one register. */
1231
1232 static void
1233 combine_movables (movables, nregs)
1234 struct movable *movables;
1235 int nregs;
1236 {
1237 register struct movable *m;
1238 char *matched_regs = (char *) alloca (nregs);
1239 enum machine_mode mode;
1240
1241 /* Regs that are set more than once are not allowed to match
1242 or be matched. I'm no longer sure why not. */
1243 /* Perhaps testing m->consec_sets would be more appropriate here? */
1244
1245 for (m = movables; m; m = m->next)
1246 if (m->match == 0 && n_times_used[m->regno] == 1 && !m->partial)
1247 {
1248 register struct movable *m1;
1249 int regno = m->regno;
1250
1251 bzero (matched_regs, nregs);
1252 matched_regs[regno] = 1;
1253
1254 for (m1 = movables; m1; m1 = m1->next)
1255 if (m != m1 && m1->match == 0 && n_times_used[m1->regno] == 1
1256 /* A reg used outside the loop mustn't be eliminated. */
1257 && !m1->global
1258 /* A reg used for zero-extending mustn't be eliminated. */
1259 && !m1->partial
1260 && (matched_regs[m1->regno]
1261 ||
1262 (
1263 /* Can combine regs with different modes loaded from the
1264 same constant only if the modes are the same or
1265 if both are integer modes with M wider or the same
1266 width as M1. The check for integer is redundant, but
1267 safe, since the only case of differing destination
1268 modes with equal sources is when both sources are
1269 VOIDmode, i.e., CONST_INT. */
1270 (GET_MODE (m->set_dest) == GET_MODE (m1->set_dest)
1271 || (GET_MODE_CLASS (GET_MODE (m->set_dest)) == MODE_INT
1272 && GET_MODE_CLASS (GET_MODE (m1->set_dest)) == MODE_INT
1273 && (GET_MODE_BITSIZE (GET_MODE (m->set_dest))
1274 >= GET_MODE_BITSIZE (GET_MODE (m1->set_dest)))))
1275 /* See if the source of M1 says it matches M. */
1276 && ((GET_CODE (m1->set_src) == REG
1277 && matched_regs[REGNO (m1->set_src)])
1278 || rtx_equal_for_loop_p (m->set_src, m1->set_src,
1279 movables))))
1280 && ((m->dependencies == m1->dependencies)
1281 || rtx_equal_p (m->dependencies, m1->dependencies)))
1282 {
1283 m->lifetime += m1->lifetime;
1284 m->savings += m1->savings;
1285 m1->done = 1;
1286 m1->match = m;
1287 matched_regs[m1->regno] = 1;
1288 }
1289 }
1290
1291 /* Now combine the regs used for zero-extension.
1292 This can be done for those not marked `global'
1293 provided their lives don't overlap. */
1294
1295 for (mode = GET_CLASS_NARROWEST_MODE (MODE_INT); mode != VOIDmode;
1296 mode = GET_MODE_WIDER_MODE (mode))
1297 {
1298 register struct movable *m0 = 0;
1299
1300 /* Combine all the registers for extension from mode MODE.
1301 Don't combine any that are used outside this loop. */
1302 for (m = movables; m; m = m->next)
1303 if (m->partial && ! m->global
1304 && mode == GET_MODE (SET_SRC (PATTERN (NEXT_INSN (m->insn)))))
1305 {
1306 register struct movable *m1;
1307 int first = uid_luid[regno_first_uid[m->regno]];
1308 int last = uid_luid[regno_last_uid[m->regno]];
1309
1310 if (m0 == 0)
1311 {
1312 /* First one: don't check for overlap, just record it. */
1313 m0 = m;
1314 continue;
1315 }
1316
1317 /* Make sure they extend to the same mode.
1318 (Almost always true.) */
1319 if (GET_MODE (m->set_dest) != GET_MODE (m0->set_dest))
1320 continue;
1321
1322 /* We already have one: check for overlap with those
1323 already combined together. */
1324 for (m1 = movables; m1 != m; m1 = m1->next)
1325 if (m1 == m0 || (m1->partial && m1->match == m0))
1326 if (! (uid_luid[regno_first_uid[m1->regno]] > last
1327 || uid_luid[regno_last_uid[m1->regno]] < first))
1328 goto overlap;
1329
1330 /* No overlap: we can combine this with the others. */
1331 m0->lifetime += m->lifetime;
1332 m0->savings += m->savings;
1333 m->done = 1;
1334 m->match = m0;
1335
1336 overlap: ;
1337 }
1338 }
1339 }
1340 \f
1341 /* Return 1 if regs X and Y will become the same if moved. */
1342
1343 static int
1344 regs_match_p (x, y, movables)
1345 rtx x, y;
1346 struct movable *movables;
1347 {
1348 int xn = REGNO (x);
1349 int yn = REGNO (y);
1350 struct movable *mx, *my;
1351
1352 for (mx = movables; mx; mx = mx->next)
1353 if (mx->regno == xn)
1354 break;
1355
1356 for (my = movables; my; my = my->next)
1357 if (my->regno == yn)
1358 break;
1359
1360 return (mx && my
1361 && ((mx->match == my->match && mx->match != 0)
1362 || mx->match == my
1363 || mx == my->match));
1364 }
1365
1366 /* Return 1 if X and Y are identical-looking rtx's.
1367 This is the Lisp function EQUAL for rtx arguments.
1368
1369 If two registers are matching movables or a movable register and an
1370 equivalent constant, consider them equal. */
1371
1372 static int
1373 rtx_equal_for_loop_p (x, y, movables)
1374 rtx x, y;
1375 struct movable *movables;
1376 {
1377 register int i;
1378 register int j;
1379 register struct movable *m;
1380 register enum rtx_code code;
1381 register char *fmt;
1382
1383 if (x == y)
1384 return 1;
1385 if (x == 0 || y == 0)
1386 return 0;
1387
1388 code = GET_CODE (x);
1389
1390 /* If we have a register and a constant, they may sometimes be
1391 equal. */
1392 if (GET_CODE (x) == REG && n_times_set[REGNO (x)] == -2
1393 && CONSTANT_P (y))
1394 for (m = movables; m; m = m->next)
1395 if (m->move_insn && m->regno == REGNO (x)
1396 && rtx_equal_p (m->set_src, y))
1397 return 1;
1398
1399 else if (GET_CODE (y) == REG && n_times_set[REGNO (y)] == -2
1400 && CONSTANT_P (x))
1401 for (m = movables; m; m = m->next)
1402 if (m->move_insn && m->regno == REGNO (y)
1403 && rtx_equal_p (m->set_src, x))
1404 return 1;
1405
1406 /* Otherwise, rtx's of different codes cannot be equal. */
1407 if (code != GET_CODE (y))
1408 return 0;
1409
1410 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent.
1411 (REG:SI x) and (REG:HI x) are NOT equivalent. */
1412
1413 if (GET_MODE (x) != GET_MODE (y))
1414 return 0;
1415
1416 /* These three types of rtx's can be compared nonrecursively. */
1417 if (code == REG)
1418 return (REGNO (x) == REGNO (y) || regs_match_p (x, y, movables));
1419
1420 if (code == LABEL_REF)
1421 return XEXP (x, 0) == XEXP (y, 0);
1422 if (code == SYMBOL_REF)
1423 return XSTR (x, 0) == XSTR (y, 0);
1424
1425 /* Compare the elements. If any pair of corresponding elements
1426 fail to match, return 0 for the whole things. */
1427
1428 fmt = GET_RTX_FORMAT (code);
1429 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1430 {
1431 switch (fmt[i])
1432 {
1433 case 'w':
1434 if (XWINT (x, i) != XWINT (y, i))
1435 return 0;
1436 break;
1437
1438 case 'i':
1439 if (XINT (x, i) != XINT (y, i))
1440 return 0;
1441 break;
1442
1443 case 'E':
1444 /* Two vectors must have the same length. */
1445 if (XVECLEN (x, i) != XVECLEN (y, i))
1446 return 0;
1447
1448 /* And the corresponding elements must match. */
1449 for (j = 0; j < XVECLEN (x, i); j++)
1450 if (rtx_equal_for_loop_p (XVECEXP (x, i, j), XVECEXP (y, i, j), movables) == 0)
1451 return 0;
1452 break;
1453
1454 case 'e':
1455 if (rtx_equal_for_loop_p (XEXP (x, i), XEXP (y, i), movables) == 0)
1456 return 0;
1457 break;
1458
1459 case 's':
1460 if (strcmp (XSTR (x, i), XSTR (y, i)))
1461 return 0;
1462 break;
1463
1464 case 'u':
1465 /* These are just backpointers, so they don't matter. */
1466 break;
1467
1468 case '0':
1469 break;
1470
1471 /* It is believed that rtx's at this level will never
1472 contain anything but integers and other rtx's,
1473 except for within LABEL_REFs and SYMBOL_REFs. */
1474 default:
1475 abort ();
1476 }
1477 }
1478 return 1;
1479 }
1480 \f
1481 /* If X contains any LABEL_REF's, add REG_LABEL notes for them to all
1482 insns in INSNS which use thet reference. */
1483
1484 static void
1485 add_label_notes (x, insns)
1486 rtx x;
1487 rtx insns;
1488 {
1489 enum rtx_code code = GET_CODE (x);
1490 int i, j;
1491 char *fmt;
1492 rtx insn;
1493
1494 if (code == LABEL_REF && !LABEL_REF_NONLOCAL_P (x))
1495 {
1496 rtx next = next_real_insn (XEXP (x, 0));
1497
1498 /* Don't record labels that refer to dispatch tables.
1499 This is not necessary, since the tablejump references the same label.
1500 And if we did record them, flow.c would make worse code. */
1501 if (next == 0
1502 || ! (GET_CODE (next) == JUMP_INSN
1503 && (GET_CODE (PATTERN (next)) == ADDR_VEC
1504 || GET_CODE (PATTERN (next)) == ADDR_DIFF_VEC)))
1505 {
1506 for (insn = insns; insn; insn = NEXT_INSN (insn))
1507 if (reg_mentioned_p (XEXP (x, 0), insn))
1508 REG_NOTES (insn) = gen_rtx (EXPR_LIST, REG_LABEL, XEXP (x, 0),
1509 REG_NOTES (insn));
1510 }
1511 return;
1512 }
1513
1514 fmt = GET_RTX_FORMAT (code);
1515 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1516 {
1517 if (fmt[i] == 'e')
1518 add_label_notes (XEXP (x, i), insns);
1519 else if (fmt[i] == 'E')
1520 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
1521 add_label_notes (XVECEXP (x, i, j), insns);
1522 }
1523 }
1524 \f
1525 /* Scan MOVABLES, and move the insns that deserve to be moved.
1526 If two matching movables are combined, replace one reg with the
1527 other throughout. */
1528
1529 static void
1530 move_movables (movables, threshold, insn_count, loop_start, end, nregs)
1531 struct movable *movables;
1532 int threshold;
1533 int insn_count;
1534 rtx loop_start;
1535 rtx end;
1536 int nregs;
1537 {
1538 rtx new_start = 0;
1539 register struct movable *m;
1540 register rtx p;
1541 /* Map of pseudo-register replacements to handle combining
1542 when we move several insns that load the same value
1543 into different pseudo-registers. */
1544 rtx *reg_map = (rtx *) alloca (nregs * sizeof (rtx));
1545 char *already_moved = (char *) alloca (nregs);
1546
1547 bzero (already_moved, nregs);
1548 bzero ((char *) reg_map, nregs * sizeof (rtx));
1549
1550 num_movables = 0;
1551
1552 for (m = movables; m; m = m->next)
1553 {
1554 /* Describe this movable insn. */
1555
1556 if (loop_dump_stream)
1557 {
1558 fprintf (loop_dump_stream, "Insn %d: regno %d (life %d), ",
1559 INSN_UID (m->insn), m->regno, m->lifetime);
1560 if (m->consec > 0)
1561 fprintf (loop_dump_stream, "consec %d, ", m->consec);
1562 if (m->cond)
1563 fprintf (loop_dump_stream, "cond ");
1564 if (m->force)
1565 fprintf (loop_dump_stream, "force ");
1566 if (m->global)
1567 fprintf (loop_dump_stream, "global ");
1568 if (m->done)
1569 fprintf (loop_dump_stream, "done ");
1570 if (m->move_insn)
1571 fprintf (loop_dump_stream, "move-insn ");
1572 if (m->match)
1573 fprintf (loop_dump_stream, "matches %d ",
1574 INSN_UID (m->match->insn));
1575 if (m->forces)
1576 fprintf (loop_dump_stream, "forces %d ",
1577 INSN_UID (m->forces->insn));
1578 }
1579
1580 /* Count movables. Value used in heuristics in strength_reduce. */
1581 num_movables++;
1582
1583 /* Ignore the insn if it's already done (it matched something else).
1584 Otherwise, see if it is now safe to move. */
1585
1586 if (!m->done
1587 && (! m->cond
1588 || (1 == invariant_p (m->set_src)
1589 && (m->dependencies == 0
1590 || 1 == invariant_p (m->dependencies))
1591 && (m->consec == 0
1592 || 1 == consec_sets_invariant_p (m->set_dest,
1593 m->consec + 1,
1594 m->insn))))
1595 && (! m->forces || m->forces->done))
1596 {
1597 register int regno;
1598 register rtx p;
1599 int savings = m->savings;
1600
1601 /* We have an insn that is safe to move.
1602 Compute its desirability. */
1603
1604 p = m->insn;
1605 regno = m->regno;
1606
1607 if (loop_dump_stream)
1608 fprintf (loop_dump_stream, "savings %d ", savings);
1609
1610 if (moved_once[regno])
1611 {
1612 insn_count *= 2;
1613
1614 if (loop_dump_stream)
1615 fprintf (loop_dump_stream, "halved since already moved ");
1616 }
1617
1618 /* An insn MUST be moved if we already moved something else
1619 which is safe only if this one is moved too: that is,
1620 if already_moved[REGNO] is nonzero. */
1621
1622 /* An insn is desirable to move if the new lifetime of the
1623 register is no more than THRESHOLD times the old lifetime.
1624 If it's not desirable, it means the loop is so big
1625 that moving won't speed things up much,
1626 and it is liable to make register usage worse. */
1627
1628 /* It is also desirable to move if it can be moved at no
1629 extra cost because something else was already moved. */
1630
1631 if (already_moved[regno]
1632 || (threshold * savings * m->lifetime) >= insn_count
1633 || (m->forces && m->forces->done
1634 && n_times_used[m->forces->regno] == 1))
1635 {
1636 int count;
1637 register struct movable *m1;
1638 rtx first;
1639
1640 /* Now move the insns that set the reg. */
1641
1642 if (m->partial && m->match)
1643 {
1644 rtx newpat, i1;
1645 rtx r1, r2;
1646 /* Find the end of this chain of matching regs.
1647 Thus, we load each reg in the chain from that one reg.
1648 And that reg is loaded with 0 directly,
1649 since it has ->match == 0. */
1650 for (m1 = m; m1->match; m1 = m1->match);
1651 newpat = gen_move_insn (SET_DEST (PATTERN (m->insn)),
1652 SET_DEST (PATTERN (m1->insn)));
1653 i1 = emit_insn_before (newpat, loop_start);
1654
1655 /* Mark the moved, invariant reg as being allowed to
1656 share a hard reg with the other matching invariant. */
1657 REG_NOTES (i1) = REG_NOTES (m->insn);
1658 r1 = SET_DEST (PATTERN (m->insn));
1659 r2 = SET_DEST (PATTERN (m1->insn));
1660 regs_may_share = gen_rtx (EXPR_LIST, VOIDmode, r1,
1661 gen_rtx (EXPR_LIST, VOIDmode, r2,
1662 regs_may_share));
1663 delete_insn (m->insn);
1664
1665 if (new_start == 0)
1666 new_start = i1;
1667
1668 if (loop_dump_stream)
1669 fprintf (loop_dump_stream, " moved to %d", INSN_UID (i1));
1670 }
1671 /* If we are to re-generate the item being moved with a
1672 new move insn, first delete what we have and then emit
1673 the move insn before the loop. */
1674 else if (m->move_insn)
1675 {
1676 rtx i1, temp;
1677
1678 for (count = m->consec; count >= 0; count--)
1679 {
1680 /* If this is the first insn of a library call sequence,
1681 skip to the end. */
1682 if (GET_CODE (p) != NOTE
1683 && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
1684 p = XEXP (temp, 0);
1685
1686 /* If this is the last insn of a libcall sequence, then
1687 delete every insn in the sequence except the last.
1688 The last insn is handled in the normal manner. */
1689 if (GET_CODE (p) != NOTE
1690 && (temp = find_reg_note (p, REG_RETVAL, NULL_RTX)))
1691 {
1692 temp = XEXP (temp, 0);
1693 while (temp != p)
1694 temp = delete_insn (temp);
1695 }
1696
1697 p = delete_insn (p);
1698 while (p && GET_CODE (p) == NOTE)
1699 p = NEXT_INSN (p);
1700 }
1701
1702 start_sequence ();
1703 emit_move_insn (m->set_dest, m->set_src);
1704 temp = get_insns ();
1705 end_sequence ();
1706
1707 add_label_notes (m->set_src, temp);
1708
1709 i1 = emit_insns_before (temp, loop_start);
1710 if (! find_reg_note (i1, REG_EQUAL, NULL_RTX))
1711 REG_NOTES (i1)
1712 = gen_rtx (EXPR_LIST,
1713 m->is_equiv ? REG_EQUIV : REG_EQUAL,
1714 m->set_src, REG_NOTES (i1));
1715
1716 if (loop_dump_stream)
1717 fprintf (loop_dump_stream, " moved to %d", INSN_UID (i1));
1718
1719 /* The more regs we move, the less we like moving them. */
1720 threshold -= 3;
1721 }
1722 else
1723 {
1724 for (count = m->consec; count >= 0; count--)
1725 {
1726 rtx i1, temp;
1727
1728 /* If first insn of libcall sequence, skip to end. */
1729 /* Do this at start of loop, since p is guaranteed to
1730 be an insn here. */
1731 if (GET_CODE (p) != NOTE
1732 && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
1733 p = XEXP (temp, 0);
1734
1735 /* If last insn of libcall sequence, move all
1736 insns except the last before the loop. The last
1737 insn is handled in the normal manner. */
1738 if (GET_CODE (p) != NOTE
1739 && (temp = find_reg_note (p, REG_RETVAL, NULL_RTX)))
1740 {
1741 rtx fn_address = 0;
1742 rtx fn_reg = 0;
1743 rtx fn_address_insn = 0;
1744
1745 first = 0;
1746 for (temp = XEXP (temp, 0); temp != p;
1747 temp = NEXT_INSN (temp))
1748 {
1749 rtx body;
1750 rtx n;
1751 rtx next;
1752
1753 if (GET_CODE (temp) == NOTE)
1754 continue;
1755
1756 body = PATTERN (temp);
1757
1758 /* Find the next insn after TEMP,
1759 not counting USE or NOTE insns. */
1760 for (next = NEXT_INSN (temp); next != p;
1761 next = NEXT_INSN (next))
1762 if (! (GET_CODE (next) == INSN
1763 && GET_CODE (PATTERN (next)) == USE)
1764 && GET_CODE (next) != NOTE)
1765 break;
1766
1767 /* If that is the call, this may be the insn
1768 that loads the function address.
1769
1770 Extract the function address from the insn
1771 that loads it into a register.
1772 If this insn was cse'd, we get incorrect code.
1773
1774 So emit a new move insn that copies the
1775 function address into the register that the
1776 call insn will use. flow.c will delete any
1777 redundant stores that we have created. */
1778 if (GET_CODE (next) == CALL_INSN
1779 && GET_CODE (body) == SET
1780 && GET_CODE (SET_DEST (body)) == REG
1781 && (n = find_reg_note (temp, REG_EQUAL,
1782 NULL_RTX)))
1783 {
1784 fn_reg = SET_SRC (body);
1785 if (GET_CODE (fn_reg) != REG)
1786 fn_reg = SET_DEST (body);
1787 fn_address = XEXP (n, 0);
1788 fn_address_insn = temp;
1789 }
1790 /* We have the call insn.
1791 If it uses the register we suspect it might,
1792 load it with the correct address directly. */
1793 if (GET_CODE (temp) == CALL_INSN
1794 && fn_address != 0
1795 && reg_referenced_p (fn_reg, body))
1796 emit_insn_after (gen_move_insn (fn_reg,
1797 fn_address),
1798 fn_address_insn);
1799
1800 if (GET_CODE (temp) == CALL_INSN)
1801 {
1802 i1 = emit_call_insn_before (body, loop_start);
1803 /* Because the USAGE information potentially
1804 contains objects other than hard registers
1805 we need to copy it. */
1806 if (CALL_INSN_FUNCTION_USAGE (temp))
1807 CALL_INSN_FUNCTION_USAGE (i1) =
1808 copy_rtx (CALL_INSN_FUNCTION_USAGE (temp));
1809 }
1810 else
1811 i1 = emit_insn_before (body, loop_start);
1812 if (first == 0)
1813 first = i1;
1814 if (temp == fn_address_insn)
1815 fn_address_insn = i1;
1816 REG_NOTES (i1) = REG_NOTES (temp);
1817 delete_insn (temp);
1818 }
1819 }
1820 if (m->savemode != VOIDmode)
1821 {
1822 /* P sets REG to zero; but we should clear only
1823 the bits that are not covered by the mode
1824 m->savemode. */
1825 rtx reg = m->set_dest;
1826 rtx sequence;
1827 rtx tem;
1828
1829 start_sequence ();
1830 tem = expand_binop
1831 (GET_MODE (reg), and_optab, reg,
1832 GEN_INT ((((HOST_WIDE_INT) 1
1833 << GET_MODE_BITSIZE (m->savemode)))
1834 - 1),
1835 reg, 1, OPTAB_LIB_WIDEN);
1836 if (tem == 0)
1837 abort ();
1838 if (tem != reg)
1839 emit_move_insn (reg, tem);
1840 sequence = gen_sequence ();
1841 end_sequence ();
1842 i1 = emit_insn_before (sequence, loop_start);
1843 }
1844 else if (GET_CODE (p) == CALL_INSN)
1845 {
1846 i1 = emit_call_insn_before (PATTERN (p), loop_start);
1847 /* Because the USAGE information potentially
1848 contains objects other than hard registers
1849 we need to copy it. */
1850 if (CALL_INSN_FUNCTION_USAGE (p))
1851 CALL_INSN_FUNCTION_USAGE (i1) =
1852 copy_rtx (CALL_INSN_FUNCTION_USAGE (p));
1853 }
1854 else
1855 i1 = emit_insn_before (PATTERN (p), loop_start);
1856
1857 REG_NOTES (i1) = REG_NOTES (p);
1858
1859 /* If there is a REG_EQUAL note present whose value is
1860 not loop invariant, then delete it, since it may
1861 cause problems with later optimization passes.
1862 It is possible for cse to create such notes
1863 like this as a result of record_jump_cond. */
1864
1865 if ((temp = find_reg_note (i1, REG_EQUAL, NULL_RTX))
1866 && ! invariant_p (XEXP (temp, 0)))
1867 remove_note (i1, temp);
1868
1869 if (new_start == 0)
1870 new_start = i1;
1871
1872 if (loop_dump_stream)
1873 fprintf (loop_dump_stream, " moved to %d",
1874 INSN_UID (i1));
1875
1876 #if 0
1877 /* This isn't needed because REG_NOTES is copied
1878 below and is wrong since P might be a PARALLEL. */
1879 if (REG_NOTES (i1) == 0
1880 && ! m->partial /* But not if it's a zero-extend clr. */
1881 && ! m->global /* and not if used outside the loop
1882 (since it might get set outside). */
1883 && CONSTANT_P (SET_SRC (PATTERN (p))))
1884 REG_NOTES (i1)
1885 = gen_rtx (EXPR_LIST, REG_EQUAL,
1886 SET_SRC (PATTERN (p)), REG_NOTES (i1));
1887 #endif
1888
1889 /* If library call, now fix the REG_NOTES that contain
1890 insn pointers, namely REG_LIBCALL on FIRST
1891 and REG_RETVAL on I1. */
1892 if (temp = find_reg_note (i1, REG_RETVAL, NULL_RTX))
1893 {
1894 XEXP (temp, 0) = first;
1895 temp = find_reg_note (first, REG_LIBCALL, NULL_RTX);
1896 XEXP (temp, 0) = i1;
1897 }
1898
1899 delete_insn (p);
1900 do p = NEXT_INSN (p);
1901 while (p && GET_CODE (p) == NOTE);
1902 }
1903
1904 /* The more regs we move, the less we like moving them. */
1905 threshold -= 3;
1906 }
1907
1908 /* Any other movable that loads the same register
1909 MUST be moved. */
1910 already_moved[regno] = 1;
1911
1912 /* This reg has been moved out of one loop. */
1913 moved_once[regno] = 1;
1914
1915 /* The reg set here is now invariant. */
1916 if (! m->partial)
1917 n_times_set[regno] = 0;
1918
1919 m->done = 1;
1920
1921 /* Change the length-of-life info for the register
1922 to say it lives at least the full length of this loop.
1923 This will help guide optimizations in outer loops. */
1924
1925 if (uid_luid[regno_first_uid[regno]] > INSN_LUID (loop_start))
1926 /* This is the old insn before all the moved insns.
1927 We can't use the moved insn because it is out of range
1928 in uid_luid. Only the old insns have luids. */
1929 regno_first_uid[regno] = INSN_UID (loop_start);
1930 if (uid_luid[regno_last_uid[regno]] < INSN_LUID (end))
1931 regno_last_uid[regno] = INSN_UID (end);
1932
1933 /* Combine with this moved insn any other matching movables. */
1934
1935 if (! m->partial)
1936 for (m1 = movables; m1; m1 = m1->next)
1937 if (m1->match == m)
1938 {
1939 rtx temp;
1940
1941 /* Schedule the reg loaded by M1
1942 for replacement so that shares the reg of M.
1943 If the modes differ (only possible in restricted
1944 circumstances, make a SUBREG. */
1945 if (GET_MODE (m->set_dest) == GET_MODE (m1->set_dest))
1946 reg_map[m1->regno] = m->set_dest;
1947 else
1948 reg_map[m1->regno]
1949 = gen_lowpart_common (GET_MODE (m1->set_dest),
1950 m->set_dest);
1951
1952 /* Get rid of the matching insn
1953 and prevent further processing of it. */
1954 m1->done = 1;
1955
1956 /* if library call, delete all insn except last, which
1957 is deleted below */
1958 if (temp = find_reg_note (m1->insn, REG_RETVAL,
1959 NULL_RTX))
1960 {
1961 for (temp = XEXP (temp, 0); temp != m1->insn;
1962 temp = NEXT_INSN (temp))
1963 delete_insn (temp);
1964 }
1965 delete_insn (m1->insn);
1966
1967 /* Any other movable that loads the same register
1968 MUST be moved. */
1969 already_moved[m1->regno] = 1;
1970
1971 /* The reg merged here is now invariant,
1972 if the reg it matches is invariant. */
1973 if (! m->partial)
1974 n_times_set[m1->regno] = 0;
1975 }
1976 }
1977 else if (loop_dump_stream)
1978 fprintf (loop_dump_stream, "not desirable");
1979 }
1980 else if (loop_dump_stream && !m->match)
1981 fprintf (loop_dump_stream, "not safe");
1982
1983 if (loop_dump_stream)
1984 fprintf (loop_dump_stream, "\n");
1985 }
1986
1987 if (new_start == 0)
1988 new_start = loop_start;
1989
1990 /* Go through all the instructions in the loop, making
1991 all the register substitutions scheduled in REG_MAP. */
1992 for (p = new_start; p != end; p = NEXT_INSN (p))
1993 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
1994 || GET_CODE (p) == CALL_INSN)
1995 {
1996 replace_regs (PATTERN (p), reg_map, nregs, 0);
1997 replace_regs (REG_NOTES (p), reg_map, nregs, 0);
1998 INSN_CODE (p) = -1;
1999 }
2000 }
2001 \f
2002 #if 0
2003 /* Scan X and replace the address of any MEM in it with ADDR.
2004 REG is the address that MEM should have before the replacement. */
2005
2006 static void
2007 replace_call_address (x, reg, addr)
2008 rtx x, reg, addr;
2009 {
2010 register enum rtx_code code;
2011 register int i;
2012 register char *fmt;
2013
2014 if (x == 0)
2015 return;
2016 code = GET_CODE (x);
2017 switch (code)
2018 {
2019 case PC:
2020 case CC0:
2021 case CONST_INT:
2022 case CONST_DOUBLE:
2023 case CONST:
2024 case SYMBOL_REF:
2025 case LABEL_REF:
2026 case REG:
2027 return;
2028
2029 case SET:
2030 /* Short cut for very common case. */
2031 replace_call_address (XEXP (x, 1), reg, addr);
2032 return;
2033
2034 case CALL:
2035 /* Short cut for very common case. */
2036 replace_call_address (XEXP (x, 0), reg, addr);
2037 return;
2038
2039 case MEM:
2040 /* If this MEM uses a reg other than the one we expected,
2041 something is wrong. */
2042 if (XEXP (x, 0) != reg)
2043 abort ();
2044 XEXP (x, 0) = addr;
2045 return;
2046 }
2047
2048 fmt = GET_RTX_FORMAT (code);
2049 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2050 {
2051 if (fmt[i] == 'e')
2052 replace_call_address (XEXP (x, i), reg, addr);
2053 if (fmt[i] == 'E')
2054 {
2055 register int j;
2056 for (j = 0; j < XVECLEN (x, i); j++)
2057 replace_call_address (XVECEXP (x, i, j), reg, addr);
2058 }
2059 }
2060 }
2061 #endif
2062 \f
2063 /* Return the number of memory refs to addresses that vary
2064 in the rtx X. */
2065
2066 static int
2067 count_nonfixed_reads (x)
2068 rtx x;
2069 {
2070 register enum rtx_code code;
2071 register int i;
2072 register char *fmt;
2073 int value;
2074
2075 if (x == 0)
2076 return 0;
2077
2078 code = GET_CODE (x);
2079 switch (code)
2080 {
2081 case PC:
2082 case CC0:
2083 case CONST_INT:
2084 case CONST_DOUBLE:
2085 case CONST:
2086 case SYMBOL_REF:
2087 case LABEL_REF:
2088 case REG:
2089 return 0;
2090
2091 case MEM:
2092 return ((invariant_p (XEXP (x, 0)) != 1)
2093 + count_nonfixed_reads (XEXP (x, 0)));
2094 }
2095
2096 value = 0;
2097 fmt = GET_RTX_FORMAT (code);
2098 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2099 {
2100 if (fmt[i] == 'e')
2101 value += count_nonfixed_reads (XEXP (x, i));
2102 if (fmt[i] == 'E')
2103 {
2104 register int j;
2105 for (j = 0; j < XVECLEN (x, i); j++)
2106 value += count_nonfixed_reads (XVECEXP (x, i, j));
2107 }
2108 }
2109 return value;
2110 }
2111
2112 \f
2113 #if 0
2114 /* P is an instruction that sets a register to the result of a ZERO_EXTEND.
2115 Replace it with an instruction to load just the low bytes
2116 if the machine supports such an instruction,
2117 and insert above LOOP_START an instruction to clear the register. */
2118
2119 static void
2120 constant_high_bytes (p, loop_start)
2121 rtx p, loop_start;
2122 {
2123 register rtx new;
2124 register int insn_code_number;
2125
2126 /* Try to change (SET (REG ...) (ZERO_EXTEND (..:B ...)))
2127 to (SET (STRICT_LOW_PART (SUBREG:B (REG...))) ...). */
2128
2129 new = gen_rtx (SET, VOIDmode,
2130 gen_rtx (STRICT_LOW_PART, VOIDmode,
2131 gen_rtx (SUBREG, GET_MODE (XEXP (SET_SRC (PATTERN (p)), 0)),
2132 SET_DEST (PATTERN (p)),
2133 0)),
2134 XEXP (SET_SRC (PATTERN (p)), 0));
2135 insn_code_number = recog (new, p);
2136
2137 if (insn_code_number)
2138 {
2139 register int i;
2140
2141 /* Clear destination register before the loop. */
2142 emit_insn_before (gen_rtx (SET, VOIDmode,
2143 SET_DEST (PATTERN (p)),
2144 const0_rtx),
2145 loop_start);
2146
2147 /* Inside the loop, just load the low part. */
2148 PATTERN (p) = new;
2149 }
2150 }
2151 #endif
2152 \f
2153 /* Scan a loop setting the variables `unknown_address_altered',
2154 `num_mem_sets', `loop_continue', loops_enclosed', `loop_has_call',
2155 and `loop_has_volatile'.
2156 Also, fill in the array `loop_store_mems'. */
2157
2158 static void
2159 prescan_loop (start, end)
2160 rtx start, end;
2161 {
2162 register int level = 1;
2163 register rtx insn;
2164
2165 unknown_address_altered = 0;
2166 loop_has_call = 0;
2167 loop_has_volatile = 0;
2168 loop_store_mems_idx = 0;
2169
2170 num_mem_sets = 0;
2171 loops_enclosed = 1;
2172 loop_continue = 0;
2173
2174 for (insn = NEXT_INSN (start); insn != NEXT_INSN (end);
2175 insn = NEXT_INSN (insn))
2176 {
2177 if (GET_CODE (insn) == NOTE)
2178 {
2179 if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG)
2180 {
2181 ++level;
2182 /* Count number of loops contained in this one. */
2183 loops_enclosed++;
2184 }
2185 else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_END)
2186 {
2187 --level;
2188 if (level == 0)
2189 {
2190 end = insn;
2191 break;
2192 }
2193 }
2194 else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_CONT)
2195 {
2196 if (level == 1)
2197 loop_continue = insn;
2198 }
2199 }
2200 else if (GET_CODE (insn) == CALL_INSN)
2201 {
2202 unknown_address_altered = 1;
2203 loop_has_call = 1;
2204 }
2205 else
2206 {
2207 if (GET_CODE (insn) == INSN || GET_CODE (insn) == JUMP_INSN)
2208 {
2209 if (volatile_refs_p (PATTERN (insn)))
2210 loop_has_volatile = 1;
2211
2212 note_stores (PATTERN (insn), note_addr_stored);
2213 }
2214 }
2215 }
2216 }
2217 \f
2218 /* Scan the function looking for loops. Record the start and end of each loop.
2219 Also mark as invalid loops any loops that contain a setjmp or are branched
2220 to from outside the loop. */
2221
2222 static void
2223 find_and_verify_loops (f)
2224 rtx f;
2225 {
2226 rtx insn, label;
2227 int current_loop = -1;
2228 int next_loop = -1;
2229 int loop;
2230
2231 /* If there are jumps to undefined labels,
2232 treat them as jumps out of any/all loops.
2233 This also avoids writing past end of tables when there are no loops. */
2234 uid_loop_num[0] = -1;
2235
2236 /* Find boundaries of loops, mark which loops are contained within
2237 loops, and invalidate loops that have setjmp. */
2238
2239 for (insn = f; insn; insn = NEXT_INSN (insn))
2240 {
2241 if (GET_CODE (insn) == NOTE)
2242 switch (NOTE_LINE_NUMBER (insn))
2243 {
2244 case NOTE_INSN_LOOP_BEG:
2245 loop_number_loop_starts[++next_loop] = insn;
2246 loop_number_loop_ends[next_loop] = 0;
2247 loop_outer_loop[next_loop] = current_loop;
2248 loop_invalid[next_loop] = 0;
2249 loop_number_exit_labels[next_loop] = 0;
2250 loop_number_exit_count[next_loop] = 0;
2251 current_loop = next_loop;
2252 break;
2253
2254 case NOTE_INSN_SETJMP:
2255 /* In this case, we must invalidate our current loop and any
2256 enclosing loop. */
2257 for (loop = current_loop; loop != -1; loop = loop_outer_loop[loop])
2258 {
2259 loop_invalid[loop] = 1;
2260 if (loop_dump_stream)
2261 fprintf (loop_dump_stream,
2262 "\nLoop at %d ignored due to setjmp.\n",
2263 INSN_UID (loop_number_loop_starts[loop]));
2264 }
2265 break;
2266
2267 case NOTE_INSN_LOOP_END:
2268 if (current_loop == -1)
2269 abort ();
2270
2271 loop_number_loop_ends[current_loop] = insn;
2272 current_loop = loop_outer_loop[current_loop];
2273 break;
2274
2275 }
2276
2277 /* Note that this will mark the NOTE_INSN_LOOP_END note as being in the
2278 enclosing loop, but this doesn't matter. */
2279 uid_loop_num[INSN_UID (insn)] = current_loop;
2280 }
2281
2282 /* Any loop containing a label used in an initializer must be invalidated,
2283 because it can be jumped into from anywhere. */
2284
2285 for (label = forced_labels; label; label = XEXP (label, 1))
2286 {
2287 int loop_num;
2288
2289 for (loop_num = uid_loop_num[INSN_UID (XEXP (label, 0))];
2290 loop_num != -1;
2291 loop_num = loop_outer_loop[loop_num])
2292 loop_invalid[loop_num] = 1;
2293 }
2294
2295 /* Any loop containing a label used for an exception handler must be
2296 invalidated, because it can be jumped into from anywhere. */
2297
2298 for (label = exception_handler_labels; label; label = XEXP (label, 1))
2299 {
2300 int loop_num;
2301
2302 for (loop_num = uid_loop_num[INSN_UID (XEXP (label, 0))];
2303 loop_num != -1;
2304 loop_num = loop_outer_loop[loop_num])
2305 loop_invalid[loop_num] = 1;
2306 }
2307
2308 /* Now scan all insn's in the function. If any JUMP_INSN branches into a
2309 loop that it is not contained within, that loop is marked invalid.
2310 If any INSN or CALL_INSN uses a label's address, then the loop containing
2311 that label is marked invalid, because it could be jumped into from
2312 anywhere.
2313
2314 Also look for blocks of code ending in an unconditional branch that
2315 exits the loop. If such a block is surrounded by a conditional
2316 branch around the block, move the block elsewhere (see below) and
2317 invert the jump to point to the code block. This may eliminate a
2318 label in our loop and will simplify processing by both us and a
2319 possible second cse pass. */
2320
2321 for (insn = f; insn; insn = NEXT_INSN (insn))
2322 if (GET_RTX_CLASS (GET_CODE (insn)) == 'i')
2323 {
2324 int this_loop_num = uid_loop_num[INSN_UID (insn)];
2325
2326 if (GET_CODE (insn) == INSN || GET_CODE (insn) == CALL_INSN)
2327 {
2328 rtx note = find_reg_note (insn, REG_LABEL, NULL_RTX);
2329 if (note)
2330 {
2331 int loop_num;
2332
2333 for (loop_num = uid_loop_num[INSN_UID (XEXP (note, 0))];
2334 loop_num != -1;
2335 loop_num = loop_outer_loop[loop_num])
2336 loop_invalid[loop_num] = 1;
2337 }
2338 }
2339
2340 if (GET_CODE (insn) != JUMP_INSN)
2341 continue;
2342
2343 mark_loop_jump (PATTERN (insn), this_loop_num);
2344
2345 /* See if this is an unconditional branch outside the loop. */
2346 if (this_loop_num != -1
2347 && (GET_CODE (PATTERN (insn)) == RETURN
2348 || (simplejump_p (insn)
2349 && (uid_loop_num[INSN_UID (JUMP_LABEL (insn))]
2350 != this_loop_num)))
2351 && get_max_uid () < max_uid_for_loop)
2352 {
2353 rtx p;
2354 rtx our_next = next_real_insn (insn);
2355 int dest_loop;
2356 int outer_loop = -1;
2357
2358 /* Go backwards until we reach the start of the loop, a label,
2359 or a JUMP_INSN. */
2360 for (p = PREV_INSN (insn);
2361 GET_CODE (p) != CODE_LABEL
2362 && ! (GET_CODE (p) == NOTE
2363 && NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
2364 && GET_CODE (p) != JUMP_INSN;
2365 p = PREV_INSN (p))
2366 ;
2367
2368 /* Check for the case where we have a jump to an inner nested
2369 loop, and do not perform the optimization in that case. */
2370
2371 if (JUMP_LABEL (insn))
2372 {
2373 dest_loop = uid_loop_num[INSN_UID (JUMP_LABEL (insn))];
2374 if (dest_loop != -1)
2375 {
2376 for (outer_loop = dest_loop; outer_loop != -1;
2377 outer_loop = loop_outer_loop[outer_loop])
2378 if (outer_loop == this_loop_num)
2379 break;
2380 }
2381 }
2382
2383 /* Make sure that the target of P is within the current loop. */
2384
2385 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p)
2386 && uid_loop_num[INSN_UID (JUMP_LABEL (p))] != this_loop_num)
2387 outer_loop = this_loop_num;
2388
2389 /* If we stopped on a JUMP_INSN to the next insn after INSN,
2390 we have a block of code to try to move.
2391
2392 We look backward and then forward from the target of INSN
2393 to find a BARRIER at the same loop depth as the target.
2394 If we find such a BARRIER, we make a new label for the start
2395 of the block, invert the jump in P and point it to that label,
2396 and move the block of code to the spot we found. */
2397
2398 if (outer_loop == -1
2399 && GET_CODE (p) == JUMP_INSN
2400 && JUMP_LABEL (p) != 0
2401 /* Just ignore jumps to labels that were never emitted.
2402 These always indicate compilation errors. */
2403 && INSN_UID (JUMP_LABEL (p)) != 0
2404 && condjump_p (p)
2405 && ! simplejump_p (p)
2406 && next_real_insn (JUMP_LABEL (p)) == our_next)
2407 {
2408 rtx target
2409 = JUMP_LABEL (insn) ? JUMP_LABEL (insn) : get_last_insn ();
2410 int target_loop_num = uid_loop_num[INSN_UID (target)];
2411 rtx loc;
2412
2413 for (loc = target; loc; loc = PREV_INSN (loc))
2414 if (GET_CODE (loc) == BARRIER
2415 && uid_loop_num[INSN_UID (loc)] == target_loop_num)
2416 break;
2417
2418 if (loc == 0)
2419 for (loc = target; loc; loc = NEXT_INSN (loc))
2420 if (GET_CODE (loc) == BARRIER
2421 && uid_loop_num[INSN_UID (loc)] == target_loop_num)
2422 break;
2423
2424 if (loc)
2425 {
2426 rtx cond_label = JUMP_LABEL (p);
2427 rtx new_label = get_label_after (p);
2428
2429 /* Ensure our label doesn't go away. */
2430 LABEL_NUSES (cond_label)++;
2431
2432 /* Verify that uid_loop_num is large enough and that
2433 we can invert P. */
2434 if (invert_jump (p, new_label))
2435 {
2436 rtx q, r;
2437
2438 /* Include the BARRIER after INSN and copy the
2439 block after LOC. */
2440 new_label = squeeze_notes (new_label, NEXT_INSN (insn));
2441 reorder_insns (new_label, NEXT_INSN (insn), loc);
2442
2443 /* All those insns are now in TARGET_LOOP_NUM. */
2444 for (q = new_label; q != NEXT_INSN (NEXT_INSN (insn));
2445 q = NEXT_INSN (q))
2446 uid_loop_num[INSN_UID (q)] = target_loop_num;
2447
2448 /* The label jumped to by INSN is no longer a loop exit.
2449 Unless INSN does not have a label (e.g., it is a
2450 RETURN insn), search loop_number_exit_labels to find
2451 its label_ref, and remove it. Also turn off
2452 LABEL_OUTSIDE_LOOP_P bit. */
2453 if (JUMP_LABEL (insn))
2454 {
2455 int loop_num;
2456
2457 for (q = 0,
2458 r = loop_number_exit_labels[this_loop_num];
2459 r; q = r, r = LABEL_NEXTREF (r))
2460 if (XEXP (r, 0) == JUMP_LABEL (insn))
2461 {
2462 LABEL_OUTSIDE_LOOP_P (r) = 0;
2463 if (q)
2464 LABEL_NEXTREF (q) = LABEL_NEXTREF (r);
2465 else
2466 loop_number_exit_labels[this_loop_num]
2467 = LABEL_NEXTREF (r);
2468 break;
2469 }
2470
2471 for (loop_num = this_loop_num;
2472 loop_num != -1 && loop_num != target_loop_num;
2473 loop_num = loop_outer_loop[loop_num])
2474 loop_number_exit_count[loop_num]--;
2475
2476 /* If we didn't find it, then something is wrong. */
2477 if (! r)
2478 abort ();
2479 }
2480
2481 /* P is now a jump outside the loop, so it must be put
2482 in loop_number_exit_labels, and marked as such.
2483 The easiest way to do this is to just call
2484 mark_loop_jump again for P. */
2485 mark_loop_jump (PATTERN (p), this_loop_num);
2486
2487 /* If INSN now jumps to the insn after it,
2488 delete INSN. */
2489 if (JUMP_LABEL (insn) != 0
2490 && (next_real_insn (JUMP_LABEL (insn))
2491 == next_real_insn (insn)))
2492 delete_insn (insn);
2493 }
2494
2495 /* Continue the loop after where the conditional
2496 branch used to jump, since the only branch insn
2497 in the block (if it still remains) is an inter-loop
2498 branch and hence needs no processing. */
2499 insn = NEXT_INSN (cond_label);
2500
2501 if (--LABEL_NUSES (cond_label) == 0)
2502 delete_insn (cond_label);
2503
2504 /* This loop will be continued with NEXT_INSN (insn). */
2505 insn = PREV_INSN (insn);
2506 }
2507 }
2508 }
2509 }
2510 }
2511
2512 /* If any label in X jumps to a loop different from LOOP_NUM and any of the
2513 loops it is contained in, mark the target loop invalid.
2514
2515 For speed, we assume that X is part of a pattern of a JUMP_INSN. */
2516
2517 static void
2518 mark_loop_jump (x, loop_num)
2519 rtx x;
2520 int loop_num;
2521 {
2522 int dest_loop;
2523 int outer_loop;
2524 int i;
2525
2526 switch (GET_CODE (x))
2527 {
2528 case PC:
2529 case USE:
2530 case CLOBBER:
2531 case REG:
2532 case MEM:
2533 case CONST_INT:
2534 case CONST_DOUBLE:
2535 case RETURN:
2536 return;
2537
2538 case CONST:
2539 /* There could be a label reference in here. */
2540 mark_loop_jump (XEXP (x, 0), loop_num);
2541 return;
2542
2543 case PLUS:
2544 case MINUS:
2545 case MULT:
2546 mark_loop_jump (XEXP (x, 0), loop_num);
2547 mark_loop_jump (XEXP (x, 1), loop_num);
2548 return;
2549
2550 case SIGN_EXTEND:
2551 case ZERO_EXTEND:
2552 mark_loop_jump (XEXP (x, 0), loop_num);
2553 return;
2554
2555 case LABEL_REF:
2556 dest_loop = uid_loop_num[INSN_UID (XEXP (x, 0))];
2557
2558 /* Link together all labels that branch outside the loop. This
2559 is used by final_[bg]iv_value and the loop unrolling code. Also
2560 mark this LABEL_REF so we know that this branch should predict
2561 false. */
2562
2563 /* A check to make sure the label is not in an inner nested loop,
2564 since this does not count as a loop exit. */
2565 if (dest_loop != -1)
2566 {
2567 for (outer_loop = dest_loop; outer_loop != -1;
2568 outer_loop = loop_outer_loop[outer_loop])
2569 if (outer_loop == loop_num)
2570 break;
2571 }
2572 else
2573 outer_loop = -1;
2574
2575 if (loop_num != -1 && outer_loop == -1)
2576 {
2577 LABEL_OUTSIDE_LOOP_P (x) = 1;
2578 LABEL_NEXTREF (x) = loop_number_exit_labels[loop_num];
2579 loop_number_exit_labels[loop_num] = x;
2580
2581 for (outer_loop = loop_num;
2582 outer_loop != -1 && outer_loop != dest_loop;
2583 outer_loop = loop_outer_loop[outer_loop])
2584 loop_number_exit_count[outer_loop]++;
2585 }
2586
2587 /* If this is inside a loop, but not in the current loop or one enclosed
2588 by it, it invalidates at least one loop. */
2589
2590 if (dest_loop == -1)
2591 return;
2592
2593 /* We must invalidate every nested loop containing the target of this
2594 label, except those that also contain the jump insn. */
2595
2596 for (; dest_loop != -1; dest_loop = loop_outer_loop[dest_loop])
2597 {
2598 /* Stop when we reach a loop that also contains the jump insn. */
2599 for (outer_loop = loop_num; outer_loop != -1;
2600 outer_loop = loop_outer_loop[outer_loop])
2601 if (dest_loop == outer_loop)
2602 return;
2603
2604 /* If we get here, we know we need to invalidate a loop. */
2605 if (loop_dump_stream && ! loop_invalid[dest_loop])
2606 fprintf (loop_dump_stream,
2607 "\nLoop at %d ignored due to multiple entry points.\n",
2608 INSN_UID (loop_number_loop_starts[dest_loop]));
2609
2610 loop_invalid[dest_loop] = 1;
2611 }
2612 return;
2613
2614 case SET:
2615 /* If this is not setting pc, ignore. */
2616 if (SET_DEST (x) == pc_rtx)
2617 mark_loop_jump (SET_SRC (x), loop_num);
2618 return;
2619
2620 case IF_THEN_ELSE:
2621 mark_loop_jump (XEXP (x, 1), loop_num);
2622 mark_loop_jump (XEXP (x, 2), loop_num);
2623 return;
2624
2625 case PARALLEL:
2626 case ADDR_VEC:
2627 for (i = 0; i < XVECLEN (x, 0); i++)
2628 mark_loop_jump (XVECEXP (x, 0, i), loop_num);
2629 return;
2630
2631 case ADDR_DIFF_VEC:
2632 for (i = 0; i < XVECLEN (x, 1); i++)
2633 mark_loop_jump (XVECEXP (x, 1, i), loop_num);
2634 return;
2635
2636 default:
2637 /* Treat anything else (such as a symbol_ref)
2638 as a branch out of this loop, but not into any loop. */
2639
2640 if (loop_num != -1)
2641 {
2642 loop_number_exit_labels[loop_num] = x;
2643
2644 for (outer_loop = loop_num; outer_loop != -1;
2645 outer_loop = loop_outer_loop[outer_loop])
2646 loop_number_exit_count[outer_loop]++;
2647 }
2648 return;
2649 }
2650 }
2651 \f
2652 /* Return nonzero if there is a label in the range from
2653 insn INSN to and including the insn whose luid is END
2654 INSN must have an assigned luid (i.e., it must not have
2655 been previously created by loop.c). */
2656
2657 static int
2658 labels_in_range_p (insn, end)
2659 rtx insn;
2660 int end;
2661 {
2662 while (insn && INSN_LUID (insn) <= end)
2663 {
2664 if (GET_CODE (insn) == CODE_LABEL)
2665 return 1;
2666 insn = NEXT_INSN (insn);
2667 }
2668
2669 return 0;
2670 }
2671
2672 /* Record that a memory reference X is being set. */
2673
2674 static void
2675 note_addr_stored (x)
2676 rtx x;
2677 {
2678 register int i;
2679
2680 if (x == 0 || GET_CODE (x) != MEM)
2681 return;
2682
2683 /* Count number of memory writes.
2684 This affects heuristics in strength_reduce. */
2685 num_mem_sets++;
2686
2687 /* BLKmode MEM means all memory is clobbered. */
2688 if (GET_MODE (x) == BLKmode)
2689 unknown_address_altered = 1;
2690
2691 if (unknown_address_altered)
2692 return;
2693
2694 for (i = 0; i < loop_store_mems_idx; i++)
2695 if (rtx_equal_p (XEXP (loop_store_mems[i], 0), XEXP (x, 0))
2696 && MEM_IN_STRUCT_P (x) == MEM_IN_STRUCT_P (loop_store_mems[i]))
2697 {
2698 /* We are storing at the same address as previously noted. Save the
2699 wider reference. */
2700 if (GET_MODE_SIZE (GET_MODE (x))
2701 > GET_MODE_SIZE (GET_MODE (loop_store_mems[i])))
2702 loop_store_mems[i] = x;
2703 break;
2704 }
2705
2706 if (i == NUM_STORES)
2707 unknown_address_altered = 1;
2708
2709 else if (i == loop_store_mems_idx)
2710 loop_store_mems[loop_store_mems_idx++] = x;
2711 }
2712 \f
2713 /* Return nonzero if the rtx X is invariant over the current loop.
2714
2715 The value is 2 if we refer to something only conditionally invariant.
2716
2717 If `unknown_address_altered' is nonzero, no memory ref is invariant.
2718 Otherwise, a memory ref is invariant if it does not conflict with
2719 anything stored in `loop_store_mems'. */
2720
2721 int
2722 invariant_p (x)
2723 register rtx x;
2724 {
2725 register int i;
2726 register enum rtx_code code;
2727 register char *fmt;
2728 int conditional = 0;
2729
2730 if (x == 0)
2731 return 1;
2732 code = GET_CODE (x);
2733 switch (code)
2734 {
2735 case CONST_INT:
2736 case CONST_DOUBLE:
2737 case SYMBOL_REF:
2738 case CONST:
2739 return 1;
2740
2741 case LABEL_REF:
2742 /* A LABEL_REF is normally invariant, however, if we are unrolling
2743 loops, and this label is inside the loop, then it isn't invariant.
2744 This is because each unrolled copy of the loop body will have
2745 a copy of this label. If this was invariant, then an insn loading
2746 the address of this label into a register might get moved outside
2747 the loop, and then each loop body would end up using the same label.
2748
2749 We don't know the loop bounds here though, so just fail for all
2750 labels. */
2751 if (flag_unroll_loops)
2752 return 0;
2753 else
2754 return 1;
2755
2756 case PC:
2757 case CC0:
2758 case UNSPEC_VOLATILE:
2759 return 0;
2760
2761 case REG:
2762 /* We used to check RTX_UNCHANGING_P (x) here, but that is invalid
2763 since the reg might be set by initialization within the loop. */
2764 if (x == frame_pointer_rtx || x == hard_frame_pointer_rtx
2765 || x == arg_pointer_rtx)
2766 return 1;
2767 if (loop_has_call
2768 && REGNO (x) < FIRST_PSEUDO_REGISTER && call_used_regs[REGNO (x)])
2769 return 0;
2770 if (n_times_set[REGNO (x)] < 0)
2771 return 2;
2772 return n_times_set[REGNO (x)] == 0;
2773
2774 case MEM:
2775 /* Volatile memory references must be rejected. Do this before
2776 checking for read-only items, so that volatile read-only items
2777 will be rejected also. */
2778 if (MEM_VOLATILE_P (x))
2779 return 0;
2780
2781 /* Read-only items (such as constants in a constant pool) are
2782 invariant if their address is. */
2783 if (RTX_UNCHANGING_P (x))
2784 break;
2785
2786 /* If we filled the table (or had a subroutine call), any location
2787 in memory could have been clobbered. */
2788 if (unknown_address_altered)
2789 return 0;
2790
2791 /* See if there is any dependence between a store and this load. */
2792 for (i = loop_store_mems_idx - 1; i >= 0; i--)
2793 if (true_dependence (loop_store_mems[i], x))
2794 return 0;
2795
2796 /* It's not invalidated by a store in memory
2797 but we must still verify the address is invariant. */
2798 break;
2799
2800 case ASM_OPERANDS:
2801 /* Don't mess with insns declared volatile. */
2802 if (MEM_VOLATILE_P (x))
2803 return 0;
2804 }
2805
2806 fmt = GET_RTX_FORMAT (code);
2807 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2808 {
2809 if (fmt[i] == 'e')
2810 {
2811 int tem = invariant_p (XEXP (x, i));
2812 if (tem == 0)
2813 return 0;
2814 if (tem == 2)
2815 conditional = 1;
2816 }
2817 else if (fmt[i] == 'E')
2818 {
2819 register int j;
2820 for (j = 0; j < XVECLEN (x, i); j++)
2821 {
2822 int tem = invariant_p (XVECEXP (x, i, j));
2823 if (tem == 0)
2824 return 0;
2825 if (tem == 2)
2826 conditional = 1;
2827 }
2828
2829 }
2830 }
2831
2832 return 1 + conditional;
2833 }
2834
2835 \f
2836 /* Return nonzero if all the insns in the loop that set REG
2837 are INSN and the immediately following insns,
2838 and if each of those insns sets REG in an invariant way
2839 (not counting uses of REG in them).
2840
2841 The value is 2 if some of these insns are only conditionally invariant.
2842
2843 We assume that INSN itself is the first set of REG
2844 and that its source is invariant. */
2845
2846 static int
2847 consec_sets_invariant_p (reg, n_sets, insn)
2848 int n_sets;
2849 rtx reg, insn;
2850 {
2851 register rtx p = insn;
2852 register int regno = REGNO (reg);
2853 rtx temp;
2854 /* Number of sets we have to insist on finding after INSN. */
2855 int count = n_sets - 1;
2856 int old = n_times_set[regno];
2857 int value = 0;
2858 int this;
2859
2860 /* If N_SETS hit the limit, we can't rely on its value. */
2861 if (n_sets == 127)
2862 return 0;
2863
2864 n_times_set[regno] = 0;
2865
2866 while (count > 0)
2867 {
2868 register enum rtx_code code;
2869 rtx set;
2870
2871 p = NEXT_INSN (p);
2872 code = GET_CODE (p);
2873
2874 /* If library call, skip to end of of it. */
2875 if (code == INSN && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
2876 p = XEXP (temp, 0);
2877
2878 this = 0;
2879 if (code == INSN
2880 && (set = single_set (p))
2881 && GET_CODE (SET_DEST (set)) == REG
2882 && REGNO (SET_DEST (set)) == regno)
2883 {
2884 this = invariant_p (SET_SRC (set));
2885 if (this != 0)
2886 value |= this;
2887 else if (temp = find_reg_note (p, REG_EQUAL, NULL_RTX))
2888 {
2889 /* If this is a libcall, then any invariant REG_EQUAL note is OK.
2890 If this is an ordinary insn, then only CONSTANT_P REG_EQUAL
2891 notes are OK. */
2892 this = (CONSTANT_P (XEXP (temp, 0))
2893 || (find_reg_note (p, REG_RETVAL, NULL_RTX)
2894 && invariant_p (XEXP (temp, 0))));
2895 if (this != 0)
2896 value |= this;
2897 }
2898 }
2899 if (this != 0)
2900 count--;
2901 else if (code != NOTE)
2902 {
2903 n_times_set[regno] = old;
2904 return 0;
2905 }
2906 }
2907
2908 n_times_set[regno] = old;
2909 /* If invariant_p ever returned 2, we return 2. */
2910 return 1 + (value & 2);
2911 }
2912
2913 #if 0
2914 /* I don't think this condition is sufficient to allow INSN
2915 to be moved, so we no longer test it. */
2916
2917 /* Return 1 if all insns in the basic block of INSN and following INSN
2918 that set REG are invariant according to TABLE. */
2919
2920 static int
2921 all_sets_invariant_p (reg, insn, table)
2922 rtx reg, insn;
2923 short *table;
2924 {
2925 register rtx p = insn;
2926 register int regno = REGNO (reg);
2927
2928 while (1)
2929 {
2930 register enum rtx_code code;
2931 p = NEXT_INSN (p);
2932 code = GET_CODE (p);
2933 if (code == CODE_LABEL || code == JUMP_INSN)
2934 return 1;
2935 if (code == INSN && GET_CODE (PATTERN (p)) == SET
2936 && GET_CODE (SET_DEST (PATTERN (p))) == REG
2937 && REGNO (SET_DEST (PATTERN (p))) == regno)
2938 {
2939 if (!invariant_p (SET_SRC (PATTERN (p)), table))
2940 return 0;
2941 }
2942 }
2943 }
2944 #endif /* 0 */
2945 \f
2946 /* Look at all uses (not sets) of registers in X. For each, if it is
2947 the single use, set USAGE[REGNO] to INSN; if there was a previous use in
2948 a different insn, set USAGE[REGNO] to const0_rtx. */
2949
2950 static void
2951 find_single_use_in_loop (insn, x, usage)
2952 rtx insn;
2953 rtx x;
2954 rtx *usage;
2955 {
2956 enum rtx_code code = GET_CODE (x);
2957 char *fmt = GET_RTX_FORMAT (code);
2958 int i, j;
2959
2960 if (code == REG)
2961 usage[REGNO (x)]
2962 = (usage[REGNO (x)] != 0 && usage[REGNO (x)] != insn)
2963 ? const0_rtx : insn;
2964
2965 else if (code == SET)
2966 {
2967 /* Don't count SET_DEST if it is a REG; otherwise count things
2968 in SET_DEST because if a register is partially modified, it won't
2969 show up as a potential movable so we don't care how USAGE is set
2970 for it. */
2971 if (GET_CODE (SET_DEST (x)) != REG)
2972 find_single_use_in_loop (insn, SET_DEST (x), usage);
2973 find_single_use_in_loop (insn, SET_SRC (x), usage);
2974 }
2975 else
2976 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2977 {
2978 if (fmt[i] == 'e' && XEXP (x, i) != 0)
2979 find_single_use_in_loop (insn, XEXP (x, i), usage);
2980 else if (fmt[i] == 'E')
2981 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
2982 find_single_use_in_loop (insn, XVECEXP (x, i, j), usage);
2983 }
2984 }
2985 \f
2986 /* Increment N_TIMES_SET at the index of each register
2987 that is modified by an insn between FROM and TO.
2988 If the value of an element of N_TIMES_SET becomes 127 or more,
2989 stop incrementing it, to avoid overflow.
2990
2991 Store in SINGLE_USAGE[I] the single insn in which register I is
2992 used, if it is only used once. Otherwise, it is set to 0 (for no
2993 uses) or const0_rtx for more than one use. This parameter may be zero,
2994 in which case this processing is not done.
2995
2996 Store in *COUNT_PTR the number of actual instruction
2997 in the loop. We use this to decide what is worth moving out. */
2998
2999 /* last_set[n] is nonzero iff reg n has been set in the current basic block.
3000 In that case, it is the insn that last set reg n. */
3001
3002 static void
3003 count_loop_regs_set (from, to, may_not_move, single_usage, count_ptr, nregs)
3004 register rtx from, to;
3005 char *may_not_move;
3006 rtx *single_usage;
3007 int *count_ptr;
3008 int nregs;
3009 {
3010 register rtx *last_set = (rtx *) alloca (nregs * sizeof (rtx));
3011 register rtx insn;
3012 register int count = 0;
3013 register rtx dest;
3014
3015 bzero ((char *) last_set, nregs * sizeof (rtx));
3016 for (insn = from; insn != to; insn = NEXT_INSN (insn))
3017 {
3018 if (GET_RTX_CLASS (GET_CODE (insn)) == 'i')
3019 {
3020 ++count;
3021
3022 /* If requested, record registers that have exactly one use. */
3023 if (single_usage)
3024 {
3025 find_single_use_in_loop (insn, PATTERN (insn), single_usage);
3026
3027 /* Include uses in REG_EQUAL notes. */
3028 if (REG_NOTES (insn))
3029 find_single_use_in_loop (insn, REG_NOTES (insn), single_usage);
3030 }
3031
3032 if (GET_CODE (PATTERN (insn)) == CLOBBER
3033 && GET_CODE (XEXP (PATTERN (insn), 0)) == REG)
3034 /* Don't move a reg that has an explicit clobber.
3035 We might do so sometimes, but it's not worth the pain. */
3036 may_not_move[REGNO (XEXP (PATTERN (insn), 0))] = 1;
3037
3038 if (GET_CODE (PATTERN (insn)) == SET
3039 || GET_CODE (PATTERN (insn)) == CLOBBER)
3040 {
3041 dest = SET_DEST (PATTERN (insn));
3042 while (GET_CODE (dest) == SUBREG
3043 || GET_CODE (dest) == ZERO_EXTRACT
3044 || GET_CODE (dest) == SIGN_EXTRACT
3045 || GET_CODE (dest) == STRICT_LOW_PART)
3046 dest = XEXP (dest, 0);
3047 if (GET_CODE (dest) == REG)
3048 {
3049 register int regno = REGNO (dest);
3050 /* If this is the first setting of this reg
3051 in current basic block, and it was set before,
3052 it must be set in two basic blocks, so it cannot
3053 be moved out of the loop. */
3054 if (n_times_set[regno] > 0 && last_set[regno] == 0)
3055 may_not_move[regno] = 1;
3056 /* If this is not first setting in current basic block,
3057 see if reg was used in between previous one and this.
3058 If so, neither one can be moved. */
3059 if (last_set[regno] != 0
3060 && reg_used_between_p (dest, last_set[regno], insn))
3061 may_not_move[regno] = 1;
3062 if (n_times_set[regno] < 127)
3063 ++n_times_set[regno];
3064 last_set[regno] = insn;
3065 }
3066 }
3067 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
3068 {
3069 register int i;
3070 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
3071 {
3072 register rtx x = XVECEXP (PATTERN (insn), 0, i);
3073 if (GET_CODE (x) == CLOBBER && GET_CODE (XEXP (x, 0)) == REG)
3074 /* Don't move a reg that has an explicit clobber.
3075 It's not worth the pain to try to do it correctly. */
3076 may_not_move[REGNO (XEXP (x, 0))] = 1;
3077
3078 if (GET_CODE (x) == SET || GET_CODE (x) == CLOBBER)
3079 {
3080 dest = SET_DEST (x);
3081 while (GET_CODE (dest) == SUBREG
3082 || GET_CODE (dest) == ZERO_EXTRACT
3083 || GET_CODE (dest) == SIGN_EXTRACT
3084 || GET_CODE (dest) == STRICT_LOW_PART)
3085 dest = XEXP (dest, 0);
3086 if (GET_CODE (dest) == REG)
3087 {
3088 register int regno = REGNO (dest);
3089 if (n_times_set[regno] > 0 && last_set[regno] == 0)
3090 may_not_move[regno] = 1;
3091 if (last_set[regno] != 0
3092 && reg_used_between_p (dest, last_set[regno], insn))
3093 may_not_move[regno] = 1;
3094 if (n_times_set[regno] < 127)
3095 ++n_times_set[regno];
3096 last_set[regno] = insn;
3097 }
3098 }
3099 }
3100 }
3101 }
3102
3103 if (GET_CODE (insn) == CODE_LABEL || GET_CODE (insn) == JUMP_INSN)
3104 bzero ((char *) last_set, nregs * sizeof (rtx));
3105 }
3106 *count_ptr = count;
3107 }
3108 \f
3109 /* Given a loop that is bounded by LOOP_START and LOOP_END
3110 and that is entered at SCAN_START,
3111 return 1 if the register set in SET contained in insn INSN is used by
3112 any insn that precedes INSN in cyclic order starting
3113 from the loop entry point.
3114
3115 We don't want to use INSN_LUID here because if we restrict INSN to those
3116 that have a valid INSN_LUID, it means we cannot move an invariant out
3117 from an inner loop past two loops. */
3118
3119 static int
3120 loop_reg_used_before_p (set, insn, loop_start, scan_start, loop_end)
3121 rtx set, insn, loop_start, scan_start, loop_end;
3122 {
3123 rtx reg = SET_DEST (set);
3124 rtx p;
3125
3126 /* Scan forward checking for register usage. If we hit INSN, we
3127 are done. Otherwise, if we hit LOOP_END, wrap around to LOOP_START. */
3128 for (p = scan_start; p != insn; p = NEXT_INSN (p))
3129 {
3130 if (GET_RTX_CLASS (GET_CODE (p)) == 'i'
3131 && reg_overlap_mentioned_p (reg, PATTERN (p)))
3132 return 1;
3133
3134 if (p == loop_end)
3135 p = loop_start;
3136 }
3137
3138 return 0;
3139 }
3140 \f
3141 /* A "basic induction variable" or biv is a pseudo reg that is set
3142 (within this loop) only by incrementing or decrementing it. */
3143 /* A "general induction variable" or giv is a pseudo reg whose
3144 value is a linear function of a biv. */
3145
3146 /* Bivs are recognized by `basic_induction_var';
3147 Givs by `general_induct_var'. */
3148
3149 /* Indexed by register number, indicates whether or not register is an
3150 induction variable, and if so what type. */
3151
3152 enum iv_mode *reg_iv_type;
3153
3154 /* Indexed by register number, contains pointer to `struct induction'
3155 if register is an induction variable. This holds general info for
3156 all induction variables. */
3157
3158 struct induction **reg_iv_info;
3159
3160 /* Indexed by register number, contains pointer to `struct iv_class'
3161 if register is a basic induction variable. This holds info describing
3162 the class (a related group) of induction variables that the biv belongs
3163 to. */
3164
3165 struct iv_class **reg_biv_class;
3166
3167 /* The head of a list which links together (via the next field)
3168 every iv class for the current loop. */
3169
3170 struct iv_class *loop_iv_list;
3171
3172 /* Communication with routines called via `note_stores'. */
3173
3174 static rtx note_insn;
3175
3176 /* Dummy register to have non-zero DEST_REG for DEST_ADDR type givs. */
3177
3178 static rtx addr_placeholder;
3179
3180 /* ??? Unfinished optimizations, and possible future optimizations,
3181 for the strength reduction code. */
3182
3183 /* ??? There is one more optimization you might be interested in doing: to
3184 allocate pseudo registers for frequently-accessed memory locations.
3185 If the same memory location is referenced each time around, it might
3186 be possible to copy it into a register before and out after.
3187 This is especially useful when the memory location is a variable which
3188 is in a stack slot because somewhere its address is taken. If the
3189 loop doesn't contain a function call and the variable isn't volatile,
3190 it is safe to keep the value in a register for the duration of the
3191 loop. One tricky thing is that the copying of the value back from the
3192 register has to be done on all exits from the loop. You need to check that
3193 all the exits from the loop go to the same place. */
3194
3195 /* ??? The interaction of biv elimination, and recognition of 'constant'
3196 bivs, may cause problems. */
3197
3198 /* ??? Add heuristics so that DEST_ADDR strength reduction does not cause
3199 performance problems.
3200
3201 Perhaps don't eliminate things that can be combined with an addressing
3202 mode. Find all givs that have the same biv, mult_val, and add_val;
3203 then for each giv, check to see if its only use dies in a following
3204 memory address. If so, generate a new memory address and check to see
3205 if it is valid. If it is valid, then store the modified memory address,
3206 otherwise, mark the giv as not done so that it will get its own iv. */
3207
3208 /* ??? Could try to optimize branches when it is known that a biv is always
3209 positive. */
3210
3211 /* ??? When replace a biv in a compare insn, we should replace with closest
3212 giv so that an optimized branch can still be recognized by the combiner,
3213 e.g. the VAX acb insn. */
3214
3215 /* ??? Many of the checks involving uid_luid could be simplified if regscan
3216 was rerun in loop_optimize whenever a register was added or moved.
3217 Also, some of the optimizations could be a little less conservative. */
3218 \f
3219 /* Perform strength reduction and induction variable elimination. */
3220
3221 /* Pseudo registers created during this function will be beyond the last
3222 valid index in several tables including n_times_set and regno_last_uid.
3223 This does not cause a problem here, because the added registers cannot be
3224 givs outside of their loop, and hence will never be reconsidered.
3225 But scan_loop must check regnos to make sure they are in bounds. */
3226
3227 static void
3228 strength_reduce (scan_start, end, loop_top, insn_count,
3229 loop_start, loop_end)
3230 rtx scan_start;
3231 rtx end;
3232 rtx loop_top;
3233 int insn_count;
3234 rtx loop_start;
3235 rtx loop_end;
3236 {
3237 rtx p;
3238 rtx set;
3239 rtx inc_val;
3240 rtx mult_val;
3241 rtx dest_reg;
3242 /* This is 1 if current insn is not executed at least once for every loop
3243 iteration. */
3244 int not_every_iteration = 0;
3245 /* This is 1 if current insn may be executed more than once for every
3246 loop iteration. */
3247 int maybe_multiple = 0;
3248 /* Temporary list pointers for traversing loop_iv_list. */
3249 struct iv_class *bl, **backbl;
3250 /* Ratio of extra register life span we can justify
3251 for saving an instruction. More if loop doesn't call subroutines
3252 since in that case saving an insn makes more difference
3253 and more registers are available. */
3254 /* ??? could set this to last value of threshold in move_movables */
3255 int threshold = (loop_has_call ? 1 : 2) * (3 + n_non_fixed_regs);
3256 /* Map of pseudo-register replacements. */
3257 rtx *reg_map;
3258 int call_seen;
3259 rtx test;
3260 rtx end_insert_before;
3261 int loop_depth = 0;
3262
3263 reg_iv_type = (enum iv_mode *) alloca (max_reg_before_loop
3264 * sizeof (enum iv_mode *));
3265 bzero ((char *) reg_iv_type, max_reg_before_loop * sizeof (enum iv_mode *));
3266 reg_iv_info = (struct induction **)
3267 alloca (max_reg_before_loop * sizeof (struct induction *));
3268 bzero ((char *) reg_iv_info, (max_reg_before_loop
3269 * sizeof (struct induction *)));
3270 reg_biv_class = (struct iv_class **)
3271 alloca (max_reg_before_loop * sizeof (struct iv_class *));
3272 bzero ((char *) reg_biv_class, (max_reg_before_loop
3273 * sizeof (struct iv_class *)));
3274
3275 loop_iv_list = 0;
3276 addr_placeholder = gen_reg_rtx (Pmode);
3277
3278 /* Save insn immediately after the loop_end. Insns inserted after loop_end
3279 must be put before this insn, so that they will appear in the right
3280 order (i.e. loop order).
3281
3282 If loop_end is the end of the current function, then emit a
3283 NOTE_INSN_DELETED after loop_end and set end_insert_before to the
3284 dummy note insn. */
3285 if (NEXT_INSN (loop_end) != 0)
3286 end_insert_before = NEXT_INSN (loop_end);
3287 else
3288 end_insert_before = emit_note_after (NOTE_INSN_DELETED, loop_end);
3289
3290 /* Scan through loop to find all possible bivs. */
3291
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 = loop_top;
3304 else
3305 break;
3306 if (p == scan_start)
3307 break;
3308 }
3309
3310 if (GET_CODE (p) == INSN
3311 && (set = single_set (p))
3312 && GET_CODE (SET_DEST (set)) == REG)
3313 {
3314 dest_reg = SET_DEST (set);
3315 if (REGNO (dest_reg) < max_reg_before_loop
3316 && REGNO (dest_reg) >= FIRST_PSEUDO_REGISTER
3317 && reg_iv_type[REGNO (dest_reg)] != NOT_BASIC_INDUCT)
3318 {
3319 if (basic_induction_var (SET_SRC (set), GET_MODE (SET_SRC (set)),
3320 dest_reg, p, &inc_val, &mult_val))
3321 {
3322 /* It is a possible basic induction variable.
3323 Create and initialize an induction structure for it. */
3324
3325 struct induction *v
3326 = (struct induction *) alloca (sizeof (struct induction));
3327
3328 record_biv (v, p, dest_reg, inc_val, mult_val,
3329 not_every_iteration, maybe_multiple);
3330 reg_iv_type[REGNO (dest_reg)] = BASIC_INDUCT;
3331 }
3332 else if (REGNO (dest_reg) < max_reg_before_loop)
3333 reg_iv_type[REGNO (dest_reg)] = NOT_BASIC_INDUCT;
3334 }
3335 }
3336
3337 /* Past CODE_LABEL, we get to insns that may be executed multiple
3338 times. The only way we can be sure that they can't is if every
3339 every jump insn between here and the end of the loop either
3340 returns, exits the loop, is a forward jump, or is a jump
3341 to the loop start. */
3342
3343 if (GET_CODE (p) == CODE_LABEL)
3344 {
3345 rtx insn = p;
3346
3347 maybe_multiple = 0;
3348
3349 while (1)
3350 {
3351 insn = NEXT_INSN (insn);
3352 if (insn == scan_start)
3353 break;
3354 if (insn == end)
3355 {
3356 if (loop_top != 0)
3357 insn = loop_top;
3358 else
3359 break;
3360 if (insn == scan_start)
3361 break;
3362 }
3363
3364 if (GET_CODE (insn) == JUMP_INSN
3365 && GET_CODE (PATTERN (insn)) != RETURN
3366 && (! condjump_p (insn)
3367 || (JUMP_LABEL (insn) != 0
3368 && JUMP_LABEL (insn) != scan_start
3369 && (INSN_UID (JUMP_LABEL (insn)) >= max_uid_for_loop
3370 || INSN_UID (insn) >= max_uid_for_loop
3371 || (INSN_LUID (JUMP_LABEL (insn))
3372 < INSN_LUID (insn))))))
3373 {
3374 maybe_multiple = 1;
3375 break;
3376 }
3377 }
3378 }
3379
3380 /* Past a jump, we get to insns for which we can't count
3381 on whether they will be executed during each iteration. */
3382 /* This code appears twice in strength_reduce. There is also similar
3383 code in scan_loop. */
3384 if (GET_CODE (p) == JUMP_INSN
3385 /* If we enter the loop in the middle, and scan around to the
3386 beginning, don't set not_every_iteration for that.
3387 This can be any kind of jump, since we want to know if insns
3388 will be executed if the loop is executed. */
3389 && ! (JUMP_LABEL (p) == loop_top
3390 && ((NEXT_INSN (NEXT_INSN (p)) == loop_end && simplejump_p (p))
3391 || (NEXT_INSN (p) == loop_end && condjump_p (p)))))
3392 {
3393 rtx label = 0;
3394
3395 /* If this is a jump outside the loop, then it also doesn't
3396 matter. Check to see if the target of this branch is on the
3397 loop_number_exits_labels list. */
3398
3399 for (label = loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]];
3400 label;
3401 label = LABEL_NEXTREF (label))
3402 if (XEXP (label, 0) == JUMP_LABEL (p))
3403 break;
3404
3405 if (! label)
3406 not_every_iteration = 1;
3407 }
3408
3409 else if (GET_CODE (p) == NOTE)
3410 {
3411 /* At the virtual top of a converted loop, insns are again known to
3412 be executed each iteration: logically, the loop begins here
3413 even though the exit code has been duplicated. */
3414 if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP && loop_depth == 0)
3415 not_every_iteration = 0;
3416 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
3417 loop_depth++;
3418 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END)
3419 loop_depth--;
3420 }
3421
3422 /* Unlike in the code motion pass where MAYBE_NEVER indicates that
3423 an insn may never be executed, NOT_EVERY_ITERATION indicates whether
3424 or not an insn is known to be executed each iteration of the
3425 loop, whether or not any iterations are known to occur.
3426
3427 Therefore, if we have just passed a label and have no more labels
3428 between here and the test insn of the loop, we know these insns
3429 will be executed each iteration. */
3430
3431 if (not_every_iteration && GET_CODE (p) == CODE_LABEL
3432 && no_labels_between_p (p, loop_end))
3433 not_every_iteration = 0;
3434 }
3435
3436 /* Scan loop_iv_list to remove all regs that proved not to be bivs.
3437 Make a sanity check against n_times_set. */
3438 for (backbl = &loop_iv_list, bl = *backbl; bl; bl = bl->next)
3439 {
3440 if (reg_iv_type[bl->regno] != BASIC_INDUCT
3441 /* Above happens if register modified by subreg, etc. */
3442 /* Make sure it is not recognized as a basic induction var: */
3443 || n_times_set[bl->regno] != bl->biv_count
3444 /* If never incremented, it is invariant that we decided not to
3445 move. So leave it alone. */
3446 || ! bl->incremented)
3447 {
3448 if (loop_dump_stream)
3449 fprintf (loop_dump_stream, "Reg %d: biv discarded, %s\n",
3450 bl->regno,
3451 (reg_iv_type[bl->regno] != BASIC_INDUCT
3452 ? "not induction variable"
3453 : (! bl->incremented ? "never incremented"
3454 : "count error")));
3455
3456 reg_iv_type[bl->regno] = NOT_BASIC_INDUCT;
3457 *backbl = bl->next;
3458 }
3459 else
3460 {
3461 backbl = &bl->next;
3462
3463 if (loop_dump_stream)
3464 fprintf (loop_dump_stream, "Reg %d: biv verified\n", bl->regno);
3465 }
3466 }
3467
3468 /* Exit if there are no bivs. */
3469 if (! loop_iv_list)
3470 {
3471 /* Can still unroll the loop anyways, but indicate that there is no
3472 strength reduction info available. */
3473 if (flag_unroll_loops)
3474 unroll_loop (loop_end, insn_count, loop_start, end_insert_before, 0);
3475
3476 return;
3477 }
3478
3479 /* Find initial value for each biv by searching backwards from loop_start,
3480 halting at first label. Also record any test condition. */
3481
3482 call_seen = 0;
3483 for (p = loop_start; p && GET_CODE (p) != CODE_LABEL; p = PREV_INSN (p))
3484 {
3485 note_insn = p;
3486
3487 if (GET_CODE (p) == CALL_INSN)
3488 call_seen = 1;
3489
3490 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
3491 || GET_CODE (p) == CALL_INSN)
3492 note_stores (PATTERN (p), record_initial);
3493
3494 /* Record any test of a biv that branches around the loop if no store
3495 between it and the start of loop. We only care about tests with
3496 constants and registers and only certain of those. */
3497 if (GET_CODE (p) == JUMP_INSN
3498 && JUMP_LABEL (p) != 0
3499 && next_real_insn (JUMP_LABEL (p)) == next_real_insn (loop_end)
3500 && (test = get_condition_for_loop (p)) != 0
3501 && GET_CODE (XEXP (test, 0)) == REG
3502 && REGNO (XEXP (test, 0)) < max_reg_before_loop
3503 && (bl = reg_biv_class[REGNO (XEXP (test, 0))]) != 0
3504 && valid_initial_value_p (XEXP (test, 1), p, call_seen, loop_start)
3505 && bl->init_insn == 0)
3506 {
3507 /* If an NE test, we have an initial value! */
3508 if (GET_CODE (test) == NE)
3509 {
3510 bl->init_insn = p;
3511 bl->init_set = gen_rtx (SET, VOIDmode,
3512 XEXP (test, 0), XEXP (test, 1));
3513 }
3514 else
3515 bl->initial_test = test;
3516 }
3517 }
3518
3519 /* Look at the each biv and see if we can say anything better about its
3520 initial value from any initializing insns set up above. (This is done
3521 in two passes to avoid missing SETs in a PARALLEL.) */
3522 for (bl = loop_iv_list; bl; bl = bl->next)
3523 {
3524 rtx src;
3525
3526 if (! bl->init_insn)
3527 continue;
3528
3529 src = SET_SRC (bl->init_set);
3530
3531 if (loop_dump_stream)
3532 fprintf (loop_dump_stream,
3533 "Biv %d initialized at insn %d: initial value ",
3534 bl->regno, INSN_UID (bl->init_insn));
3535
3536 if ((GET_MODE (src) == GET_MODE (regno_reg_rtx[bl->regno])
3537 || GET_MODE (src) == VOIDmode)
3538 && valid_initial_value_p (src, bl->init_insn, call_seen, loop_start))
3539 {
3540 bl->initial_value = src;
3541
3542 if (loop_dump_stream)
3543 {
3544 if (GET_CODE (src) == CONST_INT)
3545 fprintf (loop_dump_stream, "%d\n", INTVAL (src));
3546 else
3547 {
3548 print_rtl (loop_dump_stream, src);
3549 fprintf (loop_dump_stream, "\n");
3550 }
3551 }
3552 }
3553 else
3554 {
3555 /* Biv initial value is not simple move,
3556 so let it keep initial value of "itself". */
3557
3558 if (loop_dump_stream)
3559 fprintf (loop_dump_stream, "is complex\n");
3560 }
3561 }
3562
3563 /* Search the loop for general induction variables. */
3564
3565 /* A register is a giv if: it is only set once, it is a function of a
3566 biv and a constant (or invariant), and it is not a biv. */
3567
3568 not_every_iteration = 0;
3569 loop_depth = 0;
3570 p = scan_start;
3571 while (1)
3572 {
3573 p = NEXT_INSN (p);
3574 /* At end of a straight-in loop, we are done.
3575 At end of a loop entered at the bottom, scan the top. */
3576 if (p == scan_start)
3577 break;
3578 if (p == end)
3579 {
3580 if (loop_top != 0)
3581 p = loop_top;
3582 else
3583 break;
3584 if (p == scan_start)
3585 break;
3586 }
3587
3588 /* Look for a general induction variable in a register. */
3589 if (GET_CODE (p) == INSN
3590 && (set = single_set (p))
3591 && GET_CODE (SET_DEST (set)) == REG
3592 && ! may_not_optimize[REGNO (SET_DEST (set))])
3593 {
3594 rtx src_reg;
3595 rtx add_val;
3596 rtx mult_val;
3597 int benefit;
3598 rtx regnote = 0;
3599
3600 dest_reg = SET_DEST (set);
3601 if (REGNO (dest_reg) < FIRST_PSEUDO_REGISTER)
3602 continue;
3603
3604 if (/* SET_SRC is a giv. */
3605 ((benefit = general_induction_var (SET_SRC (set),
3606 &src_reg, &add_val,
3607 &mult_val))
3608 /* Equivalent expression is a giv. */
3609 || ((regnote = find_reg_note (p, REG_EQUAL, NULL_RTX))
3610 && (benefit = general_induction_var (XEXP (regnote, 0),
3611 &src_reg,
3612 &add_val, &mult_val))))
3613 /* Don't try to handle any regs made by loop optimization.
3614 We have nothing on them in regno_first_uid, etc. */
3615 && REGNO (dest_reg) < max_reg_before_loop
3616 /* Don't recognize a BASIC_INDUCT_VAR here. */
3617 && dest_reg != src_reg
3618 /* This must be the only place where the register is set. */
3619 && (n_times_set[REGNO (dest_reg)] == 1
3620 /* or all sets must be consecutive and make a giv. */
3621 || (benefit = consec_sets_giv (benefit, p,
3622 src_reg, dest_reg,
3623 &add_val, &mult_val))))
3624 {
3625 int count;
3626 struct induction *v
3627 = (struct induction *) alloca (sizeof (struct induction));
3628 rtx temp;
3629
3630 /* If this is a library call, increase benefit. */
3631 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
3632 benefit += libcall_benefit (p);
3633
3634 /* Skip the consecutive insns, if there are any. */
3635 for (count = n_times_set[REGNO (dest_reg)] - 1;
3636 count > 0; count--)
3637 {
3638 /* If first insn of libcall sequence, skip to end.
3639 Do this at start of loop, since INSN is guaranteed to
3640 be an insn here. */
3641 if (GET_CODE (p) != NOTE
3642 && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
3643 p = XEXP (temp, 0);
3644
3645 do p = NEXT_INSN (p);
3646 while (GET_CODE (p) == NOTE);
3647 }
3648
3649 record_giv (v, p, src_reg, dest_reg, mult_val, add_val, benefit,
3650 DEST_REG, not_every_iteration, NULL_PTR, loop_start,
3651 loop_end);
3652
3653 }
3654 }
3655
3656 #ifndef DONT_REDUCE_ADDR
3657 /* Look for givs which are memory addresses. */
3658 /* This resulted in worse code on a VAX 8600. I wonder if it
3659 still does. */
3660 if (GET_CODE (p) == INSN)
3661 find_mem_givs (PATTERN (p), p, not_every_iteration, loop_start,
3662 loop_end);
3663 #endif
3664
3665 /* Update the status of whether giv can derive other givs. This can
3666 change when we pass a label or an insn that updates a biv. */
3667 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
3668 || GET_CODE (p) == CODE_LABEL)
3669 update_giv_derive (p);
3670
3671 /* Past a jump, we get to insns for which we can't count
3672 on whether they will be executed during each iteration. */
3673 /* This code appears twice in strength_reduce. There is also similar
3674 code in scan_loop. */
3675 if (GET_CODE (p) == JUMP_INSN
3676 /* If we enter the loop in the middle, and scan around to the
3677 beginning, don't set not_every_iteration for that.
3678 This can be any kind of jump, since we want to know if insns
3679 will be executed if the loop is executed. */
3680 && ! (JUMP_LABEL (p) == loop_top
3681 && ((NEXT_INSN (NEXT_INSN (p)) == loop_end && simplejump_p (p))
3682 || (NEXT_INSN (p) == loop_end && condjump_p (p)))))
3683 {
3684 rtx label = 0;
3685
3686 /* If this is a jump outside the loop, then it also doesn't
3687 matter. Check to see if the target of this branch is on the
3688 loop_number_exits_labels list. */
3689
3690 for (label = loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]];
3691 label;
3692 label = LABEL_NEXTREF (label))
3693 if (XEXP (label, 0) == JUMP_LABEL (p))
3694 break;
3695
3696 if (! label)
3697 not_every_iteration = 1;
3698 }
3699
3700 else if (GET_CODE (p) == NOTE)
3701 {
3702 /* At the virtual top of a converted loop, insns are again known to
3703 be executed each iteration: logically, the loop begins here
3704 even though the exit code has been duplicated. */
3705 if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP && loop_depth == 0)
3706 not_every_iteration = 0;
3707 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
3708 loop_depth++;
3709 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END)
3710 loop_depth--;
3711 }
3712
3713 /* Unlike in the code motion pass where MAYBE_NEVER indicates that
3714 an insn may never be executed, NOT_EVERY_ITERATION indicates whether
3715 or not an insn is known to be executed each iteration of the
3716 loop, whether or not any iterations are known to occur.
3717
3718 Therefore, if we have just passed a label and have no more labels
3719 between here and the test insn of the loop, we know these insns
3720 will be executed each iteration. */
3721
3722 if (not_every_iteration && GET_CODE (p) == CODE_LABEL
3723 && no_labels_between_p (p, loop_end))
3724 not_every_iteration = 0;
3725 }
3726
3727 /* Try to calculate and save the number of loop iterations. This is
3728 set to zero if the actual number can not be calculated. This must
3729 be called after all giv's have been identified, since otherwise it may
3730 fail if the iteration variable is a giv. */
3731
3732 loop_n_iterations = loop_iterations (loop_start, loop_end);
3733
3734 /* Now for each giv for which we still don't know whether or not it is
3735 replaceable, check to see if it is replaceable because its final value
3736 can be calculated. This must be done after loop_iterations is called,
3737 so that final_giv_value will work correctly. */
3738
3739 for (bl = loop_iv_list; bl; bl = bl->next)
3740 {
3741 struct induction *v;
3742
3743 for (v = bl->giv; v; v = v->next_iv)
3744 if (! v->replaceable && ! v->not_replaceable)
3745 check_final_value (v, loop_start, loop_end);
3746 }
3747
3748 /* Try to prove that the loop counter variable (if any) is always
3749 nonnegative; if so, record that fact with a REG_NONNEG note
3750 so that "decrement and branch until zero" insn can be used. */
3751 check_dbra_loop (loop_end, insn_count, loop_start);
3752
3753 /* Create reg_map to hold substitutions for replaceable giv regs. */
3754 reg_map = (rtx *) alloca (max_reg_before_loop * sizeof (rtx));
3755 bzero ((char *) reg_map, max_reg_before_loop * sizeof (rtx));
3756
3757 /* Examine each iv class for feasibility of strength reduction/induction
3758 variable elimination. */
3759
3760 for (bl = loop_iv_list; bl; bl = bl->next)
3761 {
3762 struct induction *v;
3763 int benefit;
3764 int all_reduced;
3765 rtx final_value = 0;
3766
3767 /* Test whether it will be possible to eliminate this biv
3768 provided all givs are reduced. This is possible if either
3769 the reg is not used outside the loop, or we can compute
3770 what its final value will be.
3771
3772 For architectures with a decrement_and_branch_until_zero insn,
3773 don't do this if we put a REG_NONNEG note on the endtest for
3774 this biv. */
3775
3776 /* Compare against bl->init_insn rather than loop_start.
3777 We aren't concerned with any uses of the biv between
3778 init_insn and loop_start since these won't be affected
3779 by the value of the biv elsewhere in the function, so
3780 long as init_insn doesn't use the biv itself.
3781 March 14, 1989 -- self@bayes.arc.nasa.gov */
3782
3783 if ((uid_luid[regno_last_uid[bl->regno]] < INSN_LUID (loop_end)
3784 && bl->init_insn
3785 && INSN_UID (bl->init_insn) < max_uid_for_loop
3786 && uid_luid[regno_first_uid[bl->regno]] >= INSN_LUID (bl->init_insn)
3787 #ifdef HAVE_decrement_and_branch_until_zero
3788 && ! bl->nonneg
3789 #endif
3790 && ! reg_mentioned_p (bl->biv->dest_reg, SET_SRC (bl->init_set)))
3791 || ((final_value = final_biv_value (bl, loop_start, loop_end))
3792 #ifdef HAVE_decrement_and_branch_until_zero
3793 && ! bl->nonneg
3794 #endif
3795 ))
3796 bl->eliminable = maybe_eliminate_biv (bl, loop_start, end, 0,
3797 threshold, insn_count);
3798 else
3799 {
3800 if (loop_dump_stream)
3801 {
3802 fprintf (loop_dump_stream,
3803 "Cannot eliminate biv %d.\n",
3804 bl->regno);
3805 fprintf (loop_dump_stream,
3806 "First use: insn %d, last use: insn %d.\n",
3807 regno_first_uid[bl->regno],
3808 regno_last_uid[bl->regno]);
3809 }
3810 }
3811
3812 /* Combine all giv's for this iv_class. */
3813 combine_givs (bl);
3814
3815 /* This will be true at the end, if all givs which depend on this
3816 biv have been strength reduced.
3817 We can't (currently) eliminate the biv unless this is so. */
3818 all_reduced = 1;
3819
3820 /* Check each giv in this class to see if we will benefit by reducing
3821 it. Skip giv's combined with others. */
3822 for (v = bl->giv; v; v = v->next_iv)
3823 {
3824 struct induction *tv;
3825
3826 if (v->ignore || v->same)
3827 continue;
3828
3829 benefit = v->benefit;
3830
3831 /* Reduce benefit if not replaceable, since we will insert
3832 a move-insn to replace the insn that calculates this giv.
3833 Don't do this unless the giv is a user variable, since it
3834 will often be marked non-replaceable because of the duplication
3835 of the exit code outside the loop. In such a case, the copies
3836 we insert are dead and will be deleted. So they don't have
3837 a cost. Similar situations exist. */
3838 /* ??? The new final_[bg]iv_value code does a much better job
3839 of finding replaceable giv's, and hence this code may no longer
3840 be necessary. */
3841 if (! v->replaceable && ! bl->eliminable
3842 && REG_USERVAR_P (v->dest_reg))
3843 benefit -= copy_cost;
3844
3845 /* Decrease the benefit to count the add-insns that we will
3846 insert to increment the reduced reg for the giv. */
3847 benefit -= add_cost * bl->biv_count;
3848
3849 /* Decide whether to strength-reduce this giv or to leave the code
3850 unchanged (recompute it from the biv each time it is used).
3851 This decision can be made independently for each giv. */
3852
3853 #ifdef AUTO_INC_DEC
3854 /* Attempt to guess whether autoincrement will handle some of the
3855 new add insns; if so, increase BENEFIT (undo the subtraction of
3856 add_cost that was done above). */
3857 if (v->giv_type == DEST_ADDR
3858 && GET_CODE (v->mult_val) == CONST_INT)
3859 {
3860 #if defined (HAVE_POST_INCREMENT) || defined (HAVE_PRE_INCREMENT)
3861 if (INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
3862 benefit += add_cost * bl->biv_count;
3863 #endif
3864 #if defined (HAVE_POST_DECREMENT) || defined (HAVE_PRE_DECREMENT)
3865 if (-INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
3866 benefit += add_cost * bl->biv_count;
3867 #endif
3868 }
3869 #endif
3870
3871 /* If an insn is not to be strength reduced, then set its ignore
3872 flag, and clear all_reduced. */
3873
3874 /* A giv that depends on a reversed biv must be reduced if it is
3875 used after the loop exit, otherwise, it would have the wrong
3876 value after the loop exit. To make it simple, just reduce all
3877 of such giv's whether or not we know they are used after the loop
3878 exit. */
3879
3880 if (v->lifetime * threshold * benefit < insn_count
3881 && ! bl->reversed)
3882 {
3883 if (loop_dump_stream)
3884 fprintf (loop_dump_stream,
3885 "giv of insn %d not worth while, %d vs %d.\n",
3886 INSN_UID (v->insn),
3887 v->lifetime * threshold * benefit, insn_count);
3888 v->ignore = 1;
3889 all_reduced = 0;
3890 }
3891 else
3892 {
3893 /* Check that we can increment the reduced giv without a
3894 multiply insn. If not, reject it. */
3895
3896 for (tv = bl->biv; tv; tv = tv->next_iv)
3897 if (tv->mult_val == const1_rtx
3898 && ! product_cheap_p (tv->add_val, v->mult_val))
3899 {
3900 if (loop_dump_stream)
3901 fprintf (loop_dump_stream,
3902 "giv of insn %d: would need a multiply.\n",
3903 INSN_UID (v->insn));
3904 v->ignore = 1;
3905 all_reduced = 0;
3906 break;
3907 }
3908 }
3909 }
3910
3911 /* Reduce each giv that we decided to reduce. */
3912
3913 for (v = bl->giv; v; v = v->next_iv)
3914 {
3915 struct induction *tv;
3916 if (! v->ignore && v->same == 0)
3917 {
3918 int auto_inc_opt = 0;
3919
3920 v->new_reg = gen_reg_rtx (v->mode);
3921
3922 #ifdef AUTO_INC_DEC
3923 /* If the target has auto-increment addressing modes, and
3924 this is an address giv, then try to put the increment
3925 immediately after its use, so that flow can create an
3926 auto-increment addressing mode. */
3927 if (v->giv_type == DEST_ADDR && bl->biv_count == 1
3928 && bl->biv->always_executed && ! bl->biv->maybe_multiple
3929 /* We don't handle reversed biv's because bl->biv->insn
3930 does not have a valid INSN_LUID. */
3931 && ! bl->reversed
3932 && v->always_executed && ! v->maybe_multiple)
3933 {
3934 /* If other giv's have been combined with this one, then
3935 this will work only if all uses of the other giv's occur
3936 before this giv's insn. This is difficult to check.
3937
3938 We simplify this by looking for the common case where
3939 there is one DEST_REG giv, and this giv's insn is the
3940 last use of the dest_reg of that DEST_REG giv. If the
3941 the increment occurs after the address giv, then we can
3942 perform the optimization. (Otherwise, the increment
3943 would have to go before other_giv, and we would not be
3944 able to combine it with the address giv to get an
3945 auto-inc address.) */
3946 if (v->combined_with)
3947 {
3948 struct induction *other_giv = 0;
3949
3950 for (tv = bl->giv; tv; tv = tv->next_iv)
3951 if (tv->same == v)
3952 {
3953 if (other_giv)
3954 break;
3955 else
3956 other_giv = tv;
3957 }
3958 if (! tv && other_giv
3959 && (regno_last_uid[REGNO (other_giv->dest_reg)]
3960 == INSN_UID (v->insn))
3961 && INSN_LUID (v->insn) < INSN_LUID (bl->biv->insn))
3962 auto_inc_opt = 1;
3963 }
3964 /* Check for case where increment is before the the address
3965 giv. */
3966 else if (INSN_LUID (v->insn) > INSN_LUID (bl->biv->insn))
3967 auto_inc_opt = -1;
3968 else
3969 auto_inc_opt = 1;
3970
3971 #ifdef HAVE_cc0
3972 {
3973 rtx prev;
3974
3975 /* We can't put an insn immediately after one setting
3976 cc0, or immediately before one using cc0. */
3977 if ((auto_inc_opt == 1 && sets_cc0_p (PATTERN (v->insn)))
3978 || (auto_inc_opt == -1
3979 && (prev = prev_nonnote_insn (v->insn)) != 0
3980 && GET_RTX_CLASS (GET_CODE (prev)) == 'i'
3981 && sets_cc0_p (PATTERN (prev))))
3982 auto_inc_opt = 0;
3983 }
3984 #endif
3985
3986 if (auto_inc_opt)
3987 v->auto_inc_opt = 1;
3988 }
3989 #endif
3990
3991 /* For each place where the biv is incremented, add an insn
3992 to increment the new, reduced reg for the giv. */
3993 for (tv = bl->biv; tv; tv = tv->next_iv)
3994 {
3995 rtx insert_before;
3996
3997 if (! auto_inc_opt)
3998 insert_before = tv->insn;
3999 else if (auto_inc_opt == 1)
4000 insert_before = NEXT_INSN (v->insn);
4001 else
4002 insert_before = v->insn;
4003
4004 if (tv->mult_val == const1_rtx)
4005 emit_iv_add_mult (tv->add_val, v->mult_val,
4006 v->new_reg, v->new_reg, insert_before);
4007 else /* tv->mult_val == const0_rtx */
4008 /* A multiply is acceptable here
4009 since this is presumed to be seldom executed. */
4010 emit_iv_add_mult (tv->add_val, v->mult_val,
4011 v->add_val, v->new_reg, insert_before);
4012 }
4013
4014 /* Add code at loop start to initialize giv's reduced reg. */
4015
4016 emit_iv_add_mult (bl->initial_value, v->mult_val,
4017 v->add_val, v->new_reg, loop_start);
4018 }
4019 }
4020
4021 /* Rescan all givs. If a giv is the same as a giv not reduced, mark it
4022 as not reduced.
4023
4024 For each giv register that can be reduced now: if replaceable,
4025 substitute reduced reg wherever the old giv occurs;
4026 else add new move insn "giv_reg = reduced_reg".
4027
4028 Also check for givs whose first use is their definition and whose
4029 last use is the definition of another giv. If so, it is likely
4030 dead and should not be used to eliminate a biv. */
4031 for (v = bl->giv; v; v = v->next_iv)
4032 {
4033 if (v->same && v->same->ignore)
4034 v->ignore = 1;
4035
4036 if (v->ignore)
4037 continue;
4038
4039 if (v->giv_type == DEST_REG
4040 && regno_first_uid[REGNO (v->dest_reg)] == INSN_UID (v->insn))
4041 {
4042 struct induction *v1;
4043
4044 for (v1 = bl->giv; v1; v1 = v1->next_iv)
4045 if (regno_last_uid[REGNO (v->dest_reg)] == INSN_UID (v1->insn))
4046 v->maybe_dead = 1;
4047 }
4048
4049 /* Update expression if this was combined, in case other giv was
4050 replaced. */
4051 if (v->same)
4052 v->new_reg = replace_rtx (v->new_reg,
4053 v->same->dest_reg, v->same->new_reg);
4054
4055 if (v->giv_type == DEST_ADDR)
4056 /* Store reduced reg as the address in the memref where we found
4057 this giv. */
4058 validate_change (v->insn, v->location, v->new_reg, 0);
4059 else if (v->replaceable)
4060 {
4061 reg_map[REGNO (v->dest_reg)] = v->new_reg;
4062
4063 #if 0
4064 /* I can no longer duplicate the original problem. Perhaps
4065 this is unnecessary now? */
4066
4067 /* Replaceable; it isn't strictly necessary to delete the old
4068 insn and emit a new one, because v->dest_reg is now dead.
4069
4070 However, especially when unrolling loops, the special
4071 handling for (set REG0 REG1) in the second cse pass may
4072 make v->dest_reg live again. To avoid this problem, emit
4073 an insn to set the original giv reg from the reduced giv.
4074 We can not delete the original insn, since it may be part
4075 of a LIBCALL, and the code in flow that eliminates dead
4076 libcalls will fail if it is deleted. */
4077 emit_insn_after (gen_move_insn (v->dest_reg, v->new_reg),
4078 v->insn);
4079 #endif
4080 }
4081 else
4082 {
4083 /* Not replaceable; emit an insn to set the original giv reg from
4084 the reduced giv, same as above. */
4085 emit_insn_after (gen_move_insn (v->dest_reg, v->new_reg),
4086 v->insn);
4087 }
4088
4089 /* When a loop is reversed, givs which depend on the reversed
4090 biv, and which are live outside the loop, must be set to their
4091 correct final value. This insn is only needed if the giv is
4092 not replaceable. The correct final value is the same as the
4093 value that the giv starts the reversed loop with. */
4094 if (bl->reversed && ! v->replaceable)
4095 emit_iv_add_mult (bl->initial_value, v->mult_val,
4096 v->add_val, v->dest_reg, end_insert_before);
4097 else if (v->final_value)
4098 {
4099 rtx insert_before;
4100
4101 /* If the loop has multiple exits, emit the insn before the
4102 loop to ensure that it will always be executed no matter
4103 how the loop exits. Otherwise, emit the insn after the loop,
4104 since this is slightly more efficient. */
4105 if (loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]])
4106 insert_before = loop_start;
4107 else
4108 insert_before = end_insert_before;
4109 emit_insn_before (gen_move_insn (v->dest_reg, v->final_value),
4110 insert_before);
4111
4112 #if 0
4113 /* If the insn to set the final value of the giv was emitted
4114 before the loop, then we must delete the insn inside the loop
4115 that sets it. If this is a LIBCALL, then we must delete
4116 every insn in the libcall. Note, however, that
4117 final_giv_value will only succeed when there are multiple
4118 exits if the giv is dead at each exit, hence it does not
4119 matter that the original insn remains because it is dead
4120 anyways. */
4121 /* Delete the insn inside the loop that sets the giv since
4122 the giv is now set before (or after) the loop. */
4123 delete_insn (v->insn);
4124 #endif
4125 }
4126
4127 if (loop_dump_stream)
4128 {
4129 fprintf (loop_dump_stream, "giv at %d reduced to ",
4130 INSN_UID (v->insn));
4131 print_rtl (loop_dump_stream, v->new_reg);
4132 fprintf (loop_dump_stream, "\n");
4133 }
4134 }
4135
4136 /* All the givs based on the biv bl have been reduced if they
4137 merit it. */
4138
4139 /* For each giv not marked as maybe dead that has been combined with a
4140 second giv, clear any "maybe dead" mark on that second giv.
4141 v->new_reg will either be or refer to the register of the giv it
4142 combined with.
4143
4144 Doing this clearing avoids problems in biv elimination where a
4145 giv's new_reg is a complex value that can't be put in the insn but
4146 the giv combined with (with a reg as new_reg) is marked maybe_dead.
4147 Since the register will be used in either case, we'd prefer it be
4148 used from the simpler giv. */
4149
4150 for (v = bl->giv; v; v = v->next_iv)
4151 if (! v->maybe_dead && v->same)
4152 v->same->maybe_dead = 0;
4153
4154 /* Try to eliminate the biv, if it is a candidate.
4155 This won't work if ! all_reduced,
4156 since the givs we planned to use might not have been reduced.
4157
4158 We have to be careful that we didn't initially think we could eliminate
4159 this biv because of a giv that we now think may be dead and shouldn't
4160 be used as a biv replacement.
4161
4162 Also, there is the possibility that we may have a giv that looks
4163 like it can be used to eliminate a biv, but the resulting insn
4164 isn't valid. This can happen, for example, on the 88k, where a
4165 JUMP_INSN can compare a register only with zero. Attempts to
4166 replace it with a compare with a constant will fail.
4167
4168 Note that in cases where this call fails, we may have replaced some
4169 of the occurrences of the biv with a giv, but no harm was done in
4170 doing so in the rare cases where it can occur. */
4171
4172 if (all_reduced == 1 && bl->eliminable
4173 && maybe_eliminate_biv (bl, loop_start, end, 1,
4174 threshold, insn_count))
4175
4176 {
4177 /* ?? If we created a new test to bypass the loop entirely,
4178 or otherwise drop straight in, based on this test, then
4179 we might want to rewrite it also. This way some later
4180 pass has more hope of removing the initialization of this
4181 biv entirely. */
4182
4183 /* If final_value != 0, then the biv may be used after loop end
4184 and we must emit an insn to set it just in case.
4185
4186 Reversed bivs already have an insn after the loop setting their
4187 value, so we don't need another one. We can't calculate the
4188 proper final value for such a biv here anyways. */
4189 if (final_value != 0 && ! bl->reversed)
4190 {
4191 rtx insert_before;
4192
4193 /* If the loop has multiple exits, emit the insn before the
4194 loop to ensure that it will always be executed no matter
4195 how the loop exits. Otherwise, emit the insn after the
4196 loop, since this is slightly more efficient. */
4197 if (loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]])
4198 insert_before = loop_start;
4199 else
4200 insert_before = end_insert_before;
4201
4202 emit_insn_before (gen_move_insn (bl->biv->dest_reg, final_value),
4203 end_insert_before);
4204 }
4205
4206 #if 0
4207 /* Delete all of the instructions inside the loop which set
4208 the biv, as they are all dead. If is safe to delete them,
4209 because an insn setting a biv will never be part of a libcall. */
4210 /* However, deleting them will invalidate the regno_last_uid info,
4211 so keeping them around is more convenient. Final_biv_value
4212 will only succeed when there are multiple exits if the biv
4213 is dead at each exit, hence it does not matter that the original
4214 insn remains, because it is dead anyways. */
4215 for (v = bl->biv; v; v = v->next_iv)
4216 delete_insn (v->insn);
4217 #endif
4218
4219 if (loop_dump_stream)
4220 fprintf (loop_dump_stream, "Reg %d: biv eliminated\n",
4221 bl->regno);
4222 }
4223 }
4224
4225 /* Go through all the instructions in the loop, making all the
4226 register substitutions scheduled in REG_MAP. */
4227
4228 for (p = loop_start; p != end; p = NEXT_INSN (p))
4229 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
4230 || GET_CODE (p) == CALL_INSN)
4231 {
4232 replace_regs (PATTERN (p), reg_map, max_reg_before_loop, 0);
4233 replace_regs (REG_NOTES (p), reg_map, max_reg_before_loop, 0);
4234 INSN_CODE (p) = -1;
4235 }
4236
4237 /* Unroll loops from within strength reduction so that we can use the
4238 induction variable information that strength_reduce has already
4239 collected. */
4240
4241 if (flag_unroll_loops)
4242 unroll_loop (loop_end, insn_count, loop_start, end_insert_before, 1);
4243
4244 if (loop_dump_stream)
4245 fprintf (loop_dump_stream, "\n");
4246 }
4247 \f
4248 /* Return 1 if X is a valid source for an initial value (or as value being
4249 compared against in an initial test).
4250
4251 X must be either a register or constant and must not be clobbered between
4252 the current insn and the start of the loop.
4253
4254 INSN is the insn containing X. */
4255
4256 static int
4257 valid_initial_value_p (x, insn, call_seen, loop_start)
4258 rtx x;
4259 rtx insn;
4260 int call_seen;
4261 rtx loop_start;
4262 {
4263 if (CONSTANT_P (x))
4264 return 1;
4265
4266 /* Only consider pseudos we know about initialized in insns whose luids
4267 we know. */
4268 if (GET_CODE (x) != REG
4269 || REGNO (x) >= max_reg_before_loop)
4270 return 0;
4271
4272 /* Don't use call-clobbered registers across a call which clobbers it. On
4273 some machines, don't use any hard registers at all. */
4274 if (REGNO (x) < FIRST_PSEUDO_REGISTER
4275 && (
4276 #ifdef SMALL_REGISTER_CLASSES
4277 SMALL_REGISTER_CLASSES
4278 #else
4279 0
4280 #endif
4281 || (call_used_regs[REGNO (x)] && call_seen))
4282 )
4283 return 0;
4284
4285 /* Don't use registers that have been clobbered before the start of the
4286 loop. */
4287 if (reg_set_between_p (x, insn, loop_start))
4288 return 0;
4289
4290 return 1;
4291 }
4292 \f
4293 /* Scan X for memory refs and check each memory address
4294 as a possible giv. INSN is the insn whose pattern X comes from.
4295 NOT_EVERY_ITERATION is 1 if the insn might not be executed during
4296 every loop iteration. */
4297
4298 static void
4299 find_mem_givs (x, insn, not_every_iteration, loop_start, loop_end)
4300 rtx x;
4301 rtx insn;
4302 int not_every_iteration;
4303 rtx loop_start, loop_end;
4304 {
4305 register int i, j;
4306 register enum rtx_code code;
4307 register char *fmt;
4308
4309 if (x == 0)
4310 return;
4311
4312 code = GET_CODE (x);
4313 switch (code)
4314 {
4315 case REG:
4316 case CONST_INT:
4317 case CONST:
4318 case CONST_DOUBLE:
4319 case SYMBOL_REF:
4320 case LABEL_REF:
4321 case PC:
4322 case CC0:
4323 case ADDR_VEC:
4324 case ADDR_DIFF_VEC:
4325 case USE:
4326 case CLOBBER:
4327 return;
4328
4329 case MEM:
4330 {
4331 rtx src_reg;
4332 rtx add_val;
4333 rtx mult_val;
4334 int benefit;
4335
4336 benefit = general_induction_var (XEXP (x, 0),
4337 &src_reg, &add_val, &mult_val);
4338
4339 /* Don't make a DEST_ADDR giv with mult_val == 1 && add_val == 0.
4340 Such a giv isn't useful. */
4341 if (benefit > 0 && (mult_val != const1_rtx || add_val != const0_rtx))
4342 {
4343 /* Found one; record it. */
4344 struct induction *v
4345 = (struct induction *) oballoc (sizeof (struct induction));
4346
4347 record_giv (v, insn, src_reg, addr_placeholder, mult_val,
4348 add_val, benefit, DEST_ADDR, not_every_iteration,
4349 &XEXP (x, 0), loop_start, loop_end);
4350
4351 v->mem_mode = GET_MODE (x);
4352 }
4353 return;
4354 }
4355 }
4356
4357 /* Recursively scan the subexpressions for other mem refs. */
4358
4359 fmt = GET_RTX_FORMAT (code);
4360 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4361 if (fmt[i] == 'e')
4362 find_mem_givs (XEXP (x, i), insn, not_every_iteration, loop_start,
4363 loop_end);
4364 else if (fmt[i] == 'E')
4365 for (j = 0; j < XVECLEN (x, i); j++)
4366 find_mem_givs (XVECEXP (x, i, j), insn, not_every_iteration,
4367 loop_start, loop_end);
4368 }
4369 \f
4370 /* Fill in the data about one biv update.
4371 V is the `struct induction' in which we record the biv. (It is
4372 allocated by the caller, with alloca.)
4373 INSN is the insn that sets it.
4374 DEST_REG is the biv's reg.
4375
4376 MULT_VAL is const1_rtx if the biv is being incremented here, in which case
4377 INC_VAL is the increment. Otherwise, MULT_VAL is const0_rtx and the biv is
4378 being set to INC_VAL.
4379
4380 NOT_EVERY_ITERATION is nonzero if this biv update is not know to be
4381 executed every iteration; MAYBE_MULTIPLE is nonzero if this biv update
4382 can be executed more than once per iteration. If MAYBE_MULTIPLE
4383 and NOT_EVERY_ITERATION are both zero, we know that the biv update is
4384 executed exactly once per iteration. */
4385
4386 static void
4387 record_biv (v, insn, dest_reg, inc_val, mult_val,
4388 not_every_iteration, maybe_multiple)
4389 struct induction *v;
4390 rtx insn;
4391 rtx dest_reg;
4392 rtx inc_val;
4393 rtx mult_val;
4394 int not_every_iteration;
4395 int maybe_multiple;
4396 {
4397 struct iv_class *bl;
4398
4399 v->insn = insn;
4400 v->src_reg = dest_reg;
4401 v->dest_reg = dest_reg;
4402 v->mult_val = mult_val;
4403 v->add_val = inc_val;
4404 v->mode = GET_MODE (dest_reg);
4405 v->always_computable = ! not_every_iteration;
4406 v->always_executed = ! not_every_iteration;
4407 v->maybe_multiple = maybe_multiple;
4408
4409 /* Add this to the reg's iv_class, creating a class
4410 if this is the first incrementation of the reg. */
4411
4412 bl = reg_biv_class[REGNO (dest_reg)];
4413 if (bl == 0)
4414 {
4415 /* Create and initialize new iv_class. */
4416
4417 bl = (struct iv_class *) oballoc (sizeof (struct iv_class));
4418
4419 bl->regno = REGNO (dest_reg);
4420 bl->biv = 0;
4421 bl->giv = 0;
4422 bl->biv_count = 0;
4423 bl->giv_count = 0;
4424
4425 /* Set initial value to the reg itself. */
4426 bl->initial_value = dest_reg;
4427 /* We haven't seen the initializing insn yet */
4428 bl->init_insn = 0;
4429 bl->init_set = 0;
4430 bl->initial_test = 0;
4431 bl->incremented = 0;
4432 bl->eliminable = 0;
4433 bl->nonneg = 0;
4434 bl->reversed = 0;
4435 bl->total_benefit = 0;
4436
4437 /* Add this class to loop_iv_list. */
4438 bl->next = loop_iv_list;
4439 loop_iv_list = bl;
4440
4441 /* Put it in the array of biv register classes. */
4442 reg_biv_class[REGNO (dest_reg)] = bl;
4443 }
4444
4445 /* Update IV_CLASS entry for this biv. */
4446 v->next_iv = bl->biv;
4447 bl->biv = v;
4448 bl->biv_count++;
4449 if (mult_val == const1_rtx)
4450 bl->incremented = 1;
4451
4452 if (loop_dump_stream)
4453 {
4454 fprintf (loop_dump_stream,
4455 "Insn %d: possible biv, reg %d,",
4456 INSN_UID (insn), REGNO (dest_reg));
4457 if (GET_CODE (inc_val) == CONST_INT)
4458 fprintf (loop_dump_stream, " const = %d\n",
4459 INTVAL (inc_val));
4460 else
4461 {
4462 fprintf (loop_dump_stream, " const = ");
4463 print_rtl (loop_dump_stream, inc_val);
4464 fprintf (loop_dump_stream, "\n");
4465 }
4466 }
4467 }
4468 \f
4469 /* Fill in the data about one giv.
4470 V is the `struct induction' in which we record the giv. (It is
4471 allocated by the caller, with alloca.)
4472 INSN is the insn that sets it.
4473 BENEFIT estimates the savings from deleting this insn.
4474 TYPE is DEST_REG or DEST_ADDR; it says whether the giv is computed
4475 into a register or is used as a memory address.
4476
4477 SRC_REG is the biv reg which the giv is computed from.
4478 DEST_REG is the giv's reg (if the giv is stored in a reg).
4479 MULT_VAL and ADD_VAL are the coefficients used to compute the giv.
4480 LOCATION points to the place where this giv's value appears in INSN. */
4481
4482 static void
4483 record_giv (v, insn, src_reg, dest_reg, mult_val, add_val, benefit,
4484 type, not_every_iteration, location, loop_start, loop_end)
4485 struct induction *v;
4486 rtx insn;
4487 rtx src_reg;
4488 rtx dest_reg;
4489 rtx mult_val, add_val;
4490 int benefit;
4491 enum g_types type;
4492 int not_every_iteration;
4493 rtx *location;
4494 rtx loop_start, loop_end;
4495 {
4496 struct induction *b;
4497 struct iv_class *bl;
4498 rtx set = single_set (insn);
4499 rtx p;
4500
4501 v->insn = insn;
4502 v->src_reg = src_reg;
4503 v->giv_type = type;
4504 v->dest_reg = dest_reg;
4505 v->mult_val = mult_val;
4506 v->add_val = add_val;
4507 v->benefit = benefit;
4508 v->location = location;
4509 v->cant_derive = 0;
4510 v->combined_with = 0;
4511 v->maybe_multiple = 0;
4512 v->maybe_dead = 0;
4513 v->derive_adjustment = 0;
4514 v->same = 0;
4515 v->ignore = 0;
4516 v->new_reg = 0;
4517 v->final_value = 0;
4518 v->same_insn = 0;
4519 v->auto_inc_opt = 0;
4520
4521 /* The v->always_computable field is used in update_giv_derive, to
4522 determine whether a giv can be used to derive another giv. For a
4523 DEST_REG giv, INSN computes a new value for the giv, so its value
4524 isn't computable if INSN insn't executed every iteration.
4525 However, for a DEST_ADDR giv, INSN merely uses the value of the giv;
4526 it does not compute a new value. Hence the value is always computable
4527 regardless of whether INSN is executed each iteration. */
4528
4529 if (type == DEST_ADDR)
4530 v->always_computable = 1;
4531 else
4532 v->always_computable = ! not_every_iteration;
4533
4534 v->always_executed = ! not_every_iteration;
4535
4536 if (type == DEST_ADDR)
4537 {
4538 v->mode = GET_MODE (*location);
4539 v->lifetime = 1;
4540 v->times_used = 1;
4541 }
4542 else /* type == DEST_REG */
4543 {
4544 v->mode = GET_MODE (SET_DEST (set));
4545
4546 v->lifetime = (uid_luid[regno_last_uid[REGNO (dest_reg)]]
4547 - uid_luid[regno_first_uid[REGNO (dest_reg)]]);
4548
4549 v->times_used = n_times_used[REGNO (dest_reg)];
4550
4551 /* If the lifetime is zero, it means that this register is
4552 really a dead store. So mark this as a giv that can be
4553 ignored. This will not prevent the biv from being eliminated. */
4554 if (v->lifetime == 0)
4555 v->ignore = 1;
4556
4557 reg_iv_type[REGNO (dest_reg)] = GENERAL_INDUCT;
4558 reg_iv_info[REGNO (dest_reg)] = v;
4559 }
4560
4561 /* Add the giv to the class of givs computed from one biv. */
4562
4563 bl = reg_biv_class[REGNO (src_reg)];
4564 if (bl)
4565 {
4566 v->next_iv = bl->giv;
4567 bl->giv = v;
4568 /* Don't count DEST_ADDR. This is supposed to count the number of
4569 insns that calculate givs. */
4570 if (type == DEST_REG)
4571 bl->giv_count++;
4572 bl->total_benefit += benefit;
4573 }
4574 else
4575 /* Fatal error, biv missing for this giv? */
4576 abort ();
4577
4578 if (type == DEST_ADDR)
4579 v->replaceable = 1;
4580 else
4581 {
4582 /* The giv can be replaced outright by the reduced register only if all
4583 of the following conditions are true:
4584 - the insn that sets the giv is always executed on any iteration
4585 on which the giv is used at all
4586 (there are two ways to deduce this:
4587 either the insn is executed on every iteration,
4588 or all uses follow that insn in the same basic block),
4589 - the giv is not used outside the loop
4590 - no assignments to the biv occur during the giv's lifetime. */
4591
4592 if (regno_first_uid[REGNO (dest_reg)] == INSN_UID (insn)
4593 /* Previous line always fails if INSN was moved by loop opt. */
4594 && uid_luid[regno_last_uid[REGNO (dest_reg)]] < INSN_LUID (loop_end)
4595 && (! not_every_iteration
4596 || last_use_this_basic_block (dest_reg, insn)))
4597 {
4598 /* Now check that there are no assignments to the biv within the
4599 giv's lifetime. This requires two separate checks. */
4600
4601 /* Check each biv update, and fail if any are between the first
4602 and last use of the giv.
4603
4604 If this loop contains an inner loop that was unrolled, then
4605 the insn modifying the biv may have been emitted by the loop
4606 unrolling code, and hence does not have a valid luid. Just
4607 mark the biv as not replaceable in this case. It is not very
4608 useful as a biv, because it is used in two different loops.
4609 It is very unlikely that we would be able to optimize the giv
4610 using this biv anyways. */
4611
4612 v->replaceable = 1;
4613 for (b = bl->biv; b; b = b->next_iv)
4614 {
4615 if (INSN_UID (b->insn) >= max_uid_for_loop
4616 || ((uid_luid[INSN_UID (b->insn)]
4617 >= uid_luid[regno_first_uid[REGNO (dest_reg)]])
4618 && (uid_luid[INSN_UID (b->insn)]
4619 <= uid_luid[regno_last_uid[REGNO (dest_reg)]])))
4620 {
4621 v->replaceable = 0;
4622 v->not_replaceable = 1;
4623 break;
4624 }
4625 }
4626
4627 /* If there are any backwards branches that go from after the
4628 biv update to before it, then this giv is not replaceable. */
4629 if (v->replaceable)
4630 for (b = bl->biv; b; b = b->next_iv)
4631 if (back_branch_in_range_p (b->insn, loop_start, loop_end))
4632 {
4633 v->replaceable = 0;
4634 v->not_replaceable = 1;
4635 break;
4636 }
4637 }
4638 else
4639 {
4640 /* May still be replaceable, we don't have enough info here to
4641 decide. */
4642 v->replaceable = 0;
4643 v->not_replaceable = 0;
4644 }
4645 }
4646
4647 if (loop_dump_stream)
4648 {
4649 if (type == DEST_REG)
4650 fprintf (loop_dump_stream, "Insn %d: giv reg %d",
4651 INSN_UID (insn), REGNO (dest_reg));
4652 else
4653 fprintf (loop_dump_stream, "Insn %d: dest address",
4654 INSN_UID (insn));
4655
4656 fprintf (loop_dump_stream, " src reg %d benefit %d",
4657 REGNO (src_reg), v->benefit);
4658 fprintf (loop_dump_stream, " used %d lifetime %d",
4659 v->times_used, v->lifetime);
4660
4661 if (v->replaceable)
4662 fprintf (loop_dump_stream, " replaceable");
4663
4664 if (GET_CODE (mult_val) == CONST_INT)
4665 fprintf (loop_dump_stream, " mult %d",
4666 INTVAL (mult_val));
4667 else
4668 {
4669 fprintf (loop_dump_stream, " mult ");
4670 print_rtl (loop_dump_stream, mult_val);
4671 }
4672
4673 if (GET_CODE (add_val) == CONST_INT)
4674 fprintf (loop_dump_stream, " add %d",
4675 INTVAL (add_val));
4676 else
4677 {
4678 fprintf (loop_dump_stream, " add ");
4679 print_rtl (loop_dump_stream, add_val);
4680 }
4681 }
4682
4683 if (loop_dump_stream)
4684 fprintf (loop_dump_stream, "\n");
4685
4686 }
4687
4688
4689 /* All this does is determine whether a giv can be made replaceable because
4690 its final value can be calculated. This code can not be part of record_giv
4691 above, because final_giv_value requires that the number of loop iterations
4692 be known, and that can not be accurately calculated until after all givs
4693 have been identified. */
4694
4695 static void
4696 check_final_value (v, loop_start, loop_end)
4697 struct induction *v;
4698 rtx loop_start, loop_end;
4699 {
4700 struct iv_class *bl;
4701 rtx final_value = 0;
4702
4703 bl = reg_biv_class[REGNO (v->src_reg)];
4704
4705 /* DEST_ADDR givs will never reach here, because they are always marked
4706 replaceable above in record_giv. */
4707
4708 /* The giv can be replaced outright by the reduced register only if all
4709 of the following conditions are true:
4710 - the insn that sets the giv is always executed on any iteration
4711 on which the giv is used at all
4712 (there are two ways to deduce this:
4713 either the insn is executed on every iteration,
4714 or all uses follow that insn in the same basic block),
4715 - its final value can be calculated (this condition is different
4716 than the one above in record_giv)
4717 - no assignments to the biv occur during the giv's lifetime. */
4718
4719 #if 0
4720 /* This is only called now when replaceable is known to be false. */
4721 /* Clear replaceable, so that it won't confuse final_giv_value. */
4722 v->replaceable = 0;
4723 #endif
4724
4725 if ((final_value = final_giv_value (v, loop_start, loop_end))
4726 && (v->always_computable || last_use_this_basic_block (v->dest_reg, v->insn)))
4727 {
4728 int biv_increment_seen = 0;
4729 rtx p = v->insn;
4730 rtx last_giv_use;
4731
4732 v->replaceable = 1;
4733
4734 /* When trying to determine whether or not a biv increment occurs
4735 during the lifetime of the giv, we can ignore uses of the variable
4736 outside the loop because final_value is true. Hence we can not
4737 use regno_last_uid and regno_first_uid as above in record_giv. */
4738
4739 /* Search the loop to determine whether any assignments to the
4740 biv occur during the giv's lifetime. Start with the insn
4741 that sets the giv, and search around the loop until we come
4742 back to that insn again.
4743
4744 Also fail if there is a jump within the giv's lifetime that jumps
4745 to somewhere outside the lifetime but still within the loop. This
4746 catches spaghetti code where the execution order is not linear, and
4747 hence the above test fails. Here we assume that the giv lifetime
4748 does not extend from one iteration of the loop to the next, so as
4749 to make the test easier. Since the lifetime isn't known yet,
4750 this requires two loops. See also record_giv above. */
4751
4752 last_giv_use = v->insn;
4753
4754 while (1)
4755 {
4756 p = NEXT_INSN (p);
4757 if (p == loop_end)
4758 p = NEXT_INSN (loop_start);
4759 if (p == v->insn)
4760 break;
4761
4762 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
4763 || GET_CODE (p) == CALL_INSN)
4764 {
4765 if (biv_increment_seen)
4766 {
4767 if (reg_mentioned_p (v->dest_reg, PATTERN (p)))
4768 {
4769 v->replaceable = 0;
4770 v->not_replaceable = 1;
4771 break;
4772 }
4773 }
4774 else if (GET_CODE (PATTERN (p)) == SET
4775 && SET_DEST (PATTERN (p)) == v->src_reg)
4776 biv_increment_seen = 1;
4777 else if (reg_mentioned_p (v->dest_reg, PATTERN (p)))
4778 last_giv_use = p;
4779 }
4780 }
4781
4782 /* Now that the lifetime of the giv is known, check for branches
4783 from within the lifetime to outside the lifetime if it is still
4784 replaceable. */
4785
4786 if (v->replaceable)
4787 {
4788 p = v->insn;
4789 while (1)
4790 {
4791 p = NEXT_INSN (p);
4792 if (p == loop_end)
4793 p = NEXT_INSN (loop_start);
4794 if (p == last_giv_use)
4795 break;
4796
4797 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p)
4798 && LABEL_NAME (JUMP_LABEL (p))
4799 && ((INSN_UID (JUMP_LABEL (p)) >= max_uid_for_loop)
4800 || (INSN_UID (v->insn) >= max_uid_for_loop)
4801 || (INSN_UID (last_giv_use) >= max_uid_for_loop)
4802 || (INSN_LUID (JUMP_LABEL (p)) < INSN_LUID (v->insn)
4803 && INSN_LUID (JUMP_LABEL (p)) > INSN_LUID (loop_start))
4804 || (INSN_LUID (JUMP_LABEL (p)) > INSN_LUID (last_giv_use)
4805 && INSN_LUID (JUMP_LABEL (p)) < INSN_LUID (loop_end))))
4806 {
4807 v->replaceable = 0;
4808 v->not_replaceable = 1;
4809
4810 if (loop_dump_stream)
4811 fprintf (loop_dump_stream,
4812 "Found branch outside giv lifetime.\n");
4813
4814 break;
4815 }
4816 }
4817 }
4818
4819 /* If it is replaceable, then save the final value. */
4820 if (v->replaceable)
4821 v->final_value = final_value;
4822 }
4823
4824 if (loop_dump_stream && v->replaceable)
4825 fprintf (loop_dump_stream, "Insn %d: giv reg %d final_value replaceable\n",
4826 INSN_UID (v->insn), REGNO (v->dest_reg));
4827 }
4828 \f
4829 /* Update the status of whether a giv can derive other givs.
4830
4831 We need to do something special if there is or may be an update to the biv
4832 between the time the giv is defined and the time it is used to derive
4833 another giv.
4834
4835 In addition, a giv that is only conditionally set is not allowed to
4836 derive another giv once a label has been passed.
4837
4838 The cases we look at are when a label or an update to a biv is passed. */
4839
4840 static void
4841 update_giv_derive (p)
4842 rtx p;
4843 {
4844 struct iv_class *bl;
4845 struct induction *biv, *giv;
4846 rtx tem;
4847 int dummy;
4848
4849 /* Search all IV classes, then all bivs, and finally all givs.
4850
4851 There are three cases we are concerned with. First we have the situation
4852 of a giv that is only updated conditionally. In that case, it may not
4853 derive any givs after a label is passed.
4854
4855 The second case is when a biv update occurs, or may occur, after the
4856 definition of a giv. For certain biv updates (see below) that are
4857 known to occur between the giv definition and use, we can adjust the
4858 giv definition. For others, or when the biv update is conditional,
4859 we must prevent the giv from deriving any other givs. There are two
4860 sub-cases within this case.
4861
4862 If this is a label, we are concerned with any biv update that is done
4863 conditionally, since it may be done after the giv is defined followed by
4864 a branch here (actually, we need to pass both a jump and a label, but
4865 this extra tracking doesn't seem worth it).
4866
4867 If this is a jump, we are concerned about any biv update that may be
4868 executed multiple times. We are actually only concerned about
4869 backward jumps, but it is probably not worth performing the test
4870 on the jump again here.
4871
4872 If this is a biv update, we must adjust the giv status to show that a
4873 subsequent biv update was performed. If this adjustment cannot be done,
4874 the giv cannot derive further givs. */
4875
4876 for (bl = loop_iv_list; bl; bl = bl->next)
4877 for (biv = bl->biv; biv; biv = biv->next_iv)
4878 if (GET_CODE (p) == CODE_LABEL || GET_CODE (p) == JUMP_INSN
4879 || biv->insn == p)
4880 {
4881 for (giv = bl->giv; giv; giv = giv->next_iv)
4882 {
4883 /* If cant_derive is already true, there is no point in
4884 checking all of these conditions again. */
4885 if (giv->cant_derive)
4886 continue;
4887
4888 /* If this giv is conditionally set and we have passed a label,
4889 it cannot derive anything. */
4890 if (GET_CODE (p) == CODE_LABEL && ! giv->always_computable)
4891 giv->cant_derive = 1;
4892
4893 /* Skip givs that have mult_val == 0, since
4894 they are really invariants. Also skip those that are
4895 replaceable, since we know their lifetime doesn't contain
4896 any biv update. */
4897 else if (giv->mult_val == const0_rtx || giv->replaceable)
4898 continue;
4899
4900 /* The only way we can allow this giv to derive another
4901 is if this is a biv increment and we can form the product
4902 of biv->add_val and giv->mult_val. In this case, we will
4903 be able to compute a compensation. */
4904 else if (biv->insn == p)
4905 {
4906 tem = 0;
4907
4908 if (biv->mult_val == const1_rtx)
4909 tem = simplify_giv_expr (gen_rtx (MULT, giv->mode,
4910 biv->add_val,
4911 giv->mult_val),
4912 &dummy);
4913
4914 if (tem && giv->derive_adjustment)
4915 tem = simplify_giv_expr (gen_rtx (PLUS, giv->mode, tem,
4916 giv->derive_adjustment),
4917 &dummy);
4918 if (tem)
4919 giv->derive_adjustment = tem;
4920 else
4921 giv->cant_derive = 1;
4922 }
4923 else if ((GET_CODE (p) == CODE_LABEL && ! biv->always_computable)
4924 || (GET_CODE (p) == JUMP_INSN && biv->maybe_multiple))
4925 giv->cant_derive = 1;
4926 }
4927 }
4928 }
4929 \f
4930 /* Check whether an insn is an increment legitimate for a basic induction var.
4931 X is the source of insn P, or a part of it.
4932 MODE is the mode in which X should be interpreted.
4933
4934 DEST_REG is the putative biv, also the destination of the insn.
4935 We accept patterns of these forms:
4936 REG = REG + INVARIANT (includes REG = REG - CONSTANT)
4937 REG = INVARIANT + REG
4938
4939 If X is suitable, we return 1, set *MULT_VAL to CONST1_RTX,
4940 and store the additive term into *INC_VAL.
4941
4942 If X is an assignment of an invariant into DEST_REG, we set
4943 *MULT_VAL to CONST0_RTX, and store the invariant into *INC_VAL.
4944
4945 We also want to detect a BIV when it corresponds to a variable
4946 whose mode was promoted via PROMOTED_MODE. In that case, an increment
4947 of the variable may be a PLUS that adds a SUBREG of that variable to
4948 an invariant and then sign- or zero-extends the result of the PLUS
4949 into the variable.
4950
4951 Most GIVs in such cases will be in the promoted mode, since that is the
4952 probably the natural computation mode (and almost certainly the mode
4953 used for addresses) on the machine. So we view the pseudo-reg containing
4954 the variable as the BIV, as if it were simply incremented.
4955
4956 Note that treating the entire pseudo as a BIV will result in making
4957 simple increments to any GIVs based on it. However, if the variable
4958 overflows in its declared mode but not its promoted mode, the result will
4959 be incorrect. This is acceptable if the variable is signed, since
4960 overflows in such cases are undefined, but not if it is unsigned, since
4961 those overflows are defined. So we only check for SIGN_EXTEND and
4962 not ZERO_EXTEND.
4963
4964 If we cannot find a biv, we return 0. */
4965
4966 static int
4967 basic_induction_var (x, mode, dest_reg, p, inc_val, mult_val)
4968 register rtx x;
4969 enum machine_mode mode;
4970 rtx p;
4971 rtx dest_reg;
4972 rtx *inc_val;
4973 rtx *mult_val;
4974 {
4975 register enum rtx_code code;
4976 rtx arg;
4977 rtx insn, set = 0;
4978
4979 code = GET_CODE (x);
4980 switch (code)
4981 {
4982 case PLUS:
4983 if (XEXP (x, 0) == dest_reg
4984 || (GET_CODE (XEXP (x, 0)) == SUBREG
4985 && SUBREG_PROMOTED_VAR_P (XEXP (x, 0))
4986 && SUBREG_REG (XEXP (x, 0)) == dest_reg))
4987 arg = XEXP (x, 1);
4988 else if (XEXP (x, 1) == dest_reg
4989 || (GET_CODE (XEXP (x, 1)) == SUBREG
4990 && SUBREG_PROMOTED_VAR_P (XEXP (x, 1))
4991 && SUBREG_REG (XEXP (x, 1)) == dest_reg))
4992 arg = XEXP (x, 0);
4993 else
4994 return 0;
4995
4996 if (invariant_p (arg) != 1)
4997 return 0;
4998
4999 *inc_val = convert_modes (GET_MODE (dest_reg), GET_MODE (x), arg, 0);
5000 *mult_val = const1_rtx;
5001 return 1;
5002
5003 case SUBREG:
5004 /* If this is a SUBREG for a promoted variable, check the inner
5005 value. */
5006 if (SUBREG_PROMOTED_VAR_P (x))
5007 return basic_induction_var (SUBREG_REG (x), GET_MODE (SUBREG_REG (x)),
5008 dest_reg, p, inc_val, mult_val);
5009 return 0;
5010
5011 case REG:
5012 /* If this register is assigned in the previous insn, look at its
5013 source, but don't go outside the loop or past a label. */
5014
5015 for (insn = PREV_INSN (p);
5016 (insn && GET_CODE (insn) == NOTE
5017 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_BEG);
5018 insn = PREV_INSN (insn))
5019 ;
5020
5021 if (insn)
5022 set = single_set (insn);
5023
5024 if (set != 0
5025 && (SET_DEST (set) == x
5026 || (GET_CODE (SET_DEST (set)) == SUBREG
5027 && (GET_MODE_SIZE (GET_MODE (SET_DEST (set)))
5028 <= UNITS_PER_WORD)
5029 && SUBREG_REG (SET_DEST (set)) == x)))
5030 return basic_induction_var (SET_SRC (set),
5031 (GET_MODE (SET_SRC (set)) == VOIDmode
5032 ? GET_MODE (x)
5033 : GET_MODE (SET_SRC (set))),
5034 dest_reg, insn,
5035 inc_val, mult_val);
5036 /* ... fall through ... */
5037
5038 /* Can accept constant setting of biv only when inside inner most loop.
5039 Otherwise, a biv of an inner loop may be incorrectly recognized
5040 as a biv of the outer loop,
5041 causing code to be moved INTO the inner loop. */
5042 case MEM:
5043 if (invariant_p (x) != 1)
5044 return 0;
5045 case CONST_INT:
5046 case SYMBOL_REF:
5047 case CONST:
5048 if (loops_enclosed == 1)
5049 {
5050 /* Possible bug here? Perhaps we don't know the mode of X. */
5051 *inc_val = convert_modes (GET_MODE (dest_reg), mode, x, 0);
5052 *mult_val = const0_rtx;
5053 return 1;
5054 }
5055 else
5056 return 0;
5057
5058 case SIGN_EXTEND:
5059 return basic_induction_var (XEXP (x, 0), GET_MODE (XEXP (x, 0)),
5060 dest_reg, p, inc_val, mult_val);
5061 case ASHIFTRT:
5062 /* Similar, since this can be a sign extension. */
5063 for (insn = PREV_INSN (p);
5064 (insn && GET_CODE (insn) == NOTE
5065 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_BEG);
5066 insn = PREV_INSN (insn))
5067 ;
5068
5069 if (insn)
5070 set = single_set (insn);
5071
5072 if (set && SET_DEST (set) == XEXP (x, 0)
5073 && GET_CODE (XEXP (x, 1)) == CONST_INT
5074 && INTVAL (XEXP (x, 1)) >= 0
5075 && GET_CODE (SET_SRC (set)) == ASHIFT
5076 && XEXP (x, 1) == XEXP (SET_SRC (set), 1))
5077 return basic_induction_var (XEXP (SET_SRC (set), 0),
5078 GET_MODE (XEXP (x, 0)),
5079 dest_reg, insn, inc_val, mult_val);
5080 return 0;
5081
5082 default:
5083 return 0;
5084 }
5085 }
5086 \f
5087 /* A general induction variable (giv) is any quantity that is a linear
5088 function of a basic induction variable,
5089 i.e. giv = biv * mult_val + add_val.
5090 The coefficients can be any loop invariant quantity.
5091 A giv need not be computed directly from the biv;
5092 it can be computed by way of other givs. */
5093
5094 /* Determine whether X computes a giv.
5095 If it does, return a nonzero value
5096 which is the benefit from eliminating the computation of X;
5097 set *SRC_REG to the register of the biv that it is computed from;
5098 set *ADD_VAL and *MULT_VAL to the coefficients,
5099 such that the value of X is biv * mult + add; */
5100
5101 static int
5102 general_induction_var (x, src_reg, add_val, mult_val)
5103 rtx x;
5104 rtx *src_reg;
5105 rtx *add_val;
5106 rtx *mult_val;
5107 {
5108 rtx orig_x = x;
5109 int benefit = 0;
5110 char *storage;
5111
5112 /* If this is an invariant, forget it, it isn't a giv. */
5113 if (invariant_p (x) == 1)
5114 return 0;
5115
5116 /* See if the expression could be a giv and get its form.
5117 Mark our place on the obstack in case we don't find a giv. */
5118 storage = (char *) oballoc (0);
5119 x = simplify_giv_expr (x, &benefit);
5120 if (x == 0)
5121 {
5122 obfree (storage);
5123 return 0;
5124 }
5125
5126 switch (GET_CODE (x))
5127 {
5128 case USE:
5129 case CONST_INT:
5130 /* Since this is now an invariant and wasn't before, it must be a giv
5131 with MULT_VAL == 0. It doesn't matter which BIV we associate this
5132 with. */
5133 *src_reg = loop_iv_list->biv->dest_reg;
5134 *mult_val = const0_rtx;
5135 *add_val = x;
5136 break;
5137
5138 case REG:
5139 /* This is equivalent to a BIV. */
5140 *src_reg = x;
5141 *mult_val = const1_rtx;
5142 *add_val = const0_rtx;
5143 break;
5144
5145 case PLUS:
5146 /* Either (plus (biv) (invar)) or
5147 (plus (mult (biv) (invar_1)) (invar_2)). */
5148 if (GET_CODE (XEXP (x, 0)) == MULT)
5149 {
5150 *src_reg = XEXP (XEXP (x, 0), 0);
5151 *mult_val = XEXP (XEXP (x, 0), 1);
5152 }
5153 else
5154 {
5155 *src_reg = XEXP (x, 0);
5156 *mult_val = const1_rtx;
5157 }
5158 *add_val = XEXP (x, 1);
5159 break;
5160
5161 case MULT:
5162 /* ADD_VAL is zero. */
5163 *src_reg = XEXP (x, 0);
5164 *mult_val = XEXP (x, 1);
5165 *add_val = const0_rtx;
5166 break;
5167
5168 default:
5169 abort ();
5170 }
5171
5172 /* Remove any enclosing USE from ADD_VAL and MULT_VAL (there will be
5173 unless they are CONST_INT). */
5174 if (GET_CODE (*add_val) == USE)
5175 *add_val = XEXP (*add_val, 0);
5176 if (GET_CODE (*mult_val) == USE)
5177 *mult_val = XEXP (*mult_val, 0);
5178
5179 benefit += rtx_cost (orig_x, SET);
5180
5181 /* Always return some benefit if this is a giv so it will be detected
5182 as such. This allows elimination of bivs that might otherwise
5183 not be eliminated. */
5184 return benefit == 0 ? 1 : benefit;
5185 }
5186 \f
5187 /* Given an expression, X, try to form it as a linear function of a biv.
5188 We will canonicalize it to be of the form
5189 (plus (mult (BIV) (invar_1))
5190 (invar_2))
5191 with possible degeneracies.
5192
5193 The invariant expressions must each be of a form that can be used as a
5194 machine operand. We surround then with a USE rtx (a hack, but localized
5195 and certainly unambiguous!) if not a CONST_INT for simplicity in this
5196 routine; it is the caller's responsibility to strip them.
5197
5198 If no such canonicalization is possible (i.e., two biv's are used or an
5199 expression that is neither invariant nor a biv or giv), this routine
5200 returns 0.
5201
5202 For a non-zero return, the result will have a code of CONST_INT, USE,
5203 REG (for a BIV), PLUS, or MULT. No other codes will occur.
5204
5205 *BENEFIT will be incremented by the benefit of any sub-giv encountered. */
5206
5207 static rtx
5208 simplify_giv_expr (x, benefit)
5209 rtx x;
5210 int *benefit;
5211 {
5212 enum machine_mode mode = GET_MODE (x);
5213 rtx arg0, arg1;
5214 rtx tem;
5215
5216 /* If this is not an integer mode, or if we cannot do arithmetic in this
5217 mode, this can't be a giv. */
5218 if (mode != VOIDmode
5219 && (GET_MODE_CLASS (mode) != MODE_INT
5220 || GET_MODE_BITSIZE (mode) > HOST_BITS_PER_WIDE_INT))
5221 return 0;
5222
5223 switch (GET_CODE (x))
5224 {
5225 case PLUS:
5226 arg0 = simplify_giv_expr (XEXP (x, 0), benefit);
5227 arg1 = simplify_giv_expr (XEXP (x, 1), benefit);
5228 if (arg0 == 0 || arg1 == 0)
5229 return 0;
5230
5231 /* Put constant last, CONST_INT last if both constant. */
5232 if ((GET_CODE (arg0) == USE
5233 || GET_CODE (arg0) == CONST_INT)
5234 && GET_CODE (arg1) != CONST_INT)
5235 tem = arg0, arg0 = arg1, arg1 = tem;
5236
5237 /* Handle addition of zero, then addition of an invariant. */
5238 if (arg1 == const0_rtx)
5239 return arg0;
5240 else if (GET_CODE (arg1) == CONST_INT || GET_CODE (arg1) == USE)
5241 switch (GET_CODE (arg0))
5242 {
5243 case CONST_INT:
5244 case USE:
5245 /* Both invariant. Only valid if sum is machine operand.
5246 First strip off possible USE on first operand. */
5247 if (GET_CODE (arg0) == USE)
5248 arg0 = XEXP (arg0, 0);
5249
5250 tem = 0;
5251 if (CONSTANT_P (arg0) && GET_CODE (arg1) == CONST_INT)
5252 {
5253 tem = plus_constant (arg0, INTVAL (arg1));
5254 if (GET_CODE (tem) != CONST_INT)
5255 tem = gen_rtx (USE, mode, tem);
5256 }
5257
5258 return tem;
5259
5260 case REG:
5261 case MULT:
5262 /* biv + invar or mult + invar. Return sum. */
5263 return gen_rtx (PLUS, mode, arg0, arg1);
5264
5265 case PLUS:
5266 /* (a + invar_1) + invar_2. Associate. */
5267 return simplify_giv_expr (gen_rtx (PLUS, mode,
5268 XEXP (arg0, 0),
5269 gen_rtx (PLUS, mode,
5270 XEXP (arg0, 1), arg1)),
5271 benefit);
5272
5273 default:
5274 abort ();
5275 }
5276
5277 /* Each argument must be either REG, PLUS, or MULT. Convert REG to
5278 MULT to reduce cases. */
5279 if (GET_CODE (arg0) == REG)
5280 arg0 = gen_rtx (MULT, mode, arg0, const1_rtx);
5281 if (GET_CODE (arg1) == REG)
5282 arg1 = gen_rtx (MULT, mode, arg1, const1_rtx);
5283
5284 /* Now have PLUS + PLUS, PLUS + MULT, MULT + PLUS, or MULT + MULT.
5285 Put a MULT first, leaving PLUS + PLUS, MULT + PLUS, or MULT + MULT.
5286 Recurse to associate the second PLUS. */
5287 if (GET_CODE (arg1) == MULT)
5288 tem = arg0, arg0 = arg1, arg1 = tem;
5289
5290 if (GET_CODE (arg1) == PLUS)
5291 return simplify_giv_expr (gen_rtx (PLUS, mode,
5292 gen_rtx (PLUS, mode,
5293 arg0, XEXP (arg1, 0)),
5294 XEXP (arg1, 1)),
5295 benefit);
5296
5297 /* Now must have MULT + MULT. Distribute if same biv, else not giv. */
5298 if (GET_CODE (arg0) != MULT || GET_CODE (arg1) != MULT)
5299 abort ();
5300
5301 if (XEXP (arg0, 0) != XEXP (arg1, 0))
5302 return 0;
5303
5304 return simplify_giv_expr (gen_rtx (MULT, mode,
5305 XEXP (arg0, 0),
5306 gen_rtx (PLUS, mode,
5307 XEXP (arg0, 1),
5308 XEXP (arg1, 1))),
5309 benefit);
5310
5311 case MINUS:
5312 /* Handle "a - b" as "a + b * (-1)". */
5313 return simplify_giv_expr (gen_rtx (PLUS, mode,
5314 XEXP (x, 0),
5315 gen_rtx (MULT, mode,
5316 XEXP (x, 1), constm1_rtx)),
5317 benefit);
5318
5319 case MULT:
5320 arg0 = simplify_giv_expr (XEXP (x, 0), benefit);
5321 arg1 = simplify_giv_expr (XEXP (x, 1), benefit);
5322 if (arg0 == 0 || arg1 == 0)
5323 return 0;
5324
5325 /* Put constant last, CONST_INT last if both constant. */
5326 if ((GET_CODE (arg0) == USE || GET_CODE (arg0) == CONST_INT)
5327 && GET_CODE (arg1) != CONST_INT)
5328 tem = arg0, arg0 = arg1, arg1 = tem;
5329
5330 /* If second argument is not now constant, not giv. */
5331 if (GET_CODE (arg1) != USE && GET_CODE (arg1) != CONST_INT)
5332 return 0;
5333
5334 /* Handle multiply by 0 or 1. */
5335 if (arg1 == const0_rtx)
5336 return const0_rtx;
5337
5338 else if (arg1 == const1_rtx)
5339 return arg0;
5340
5341 switch (GET_CODE (arg0))
5342 {
5343 case REG:
5344 /* biv * invar. Done. */
5345 return gen_rtx (MULT, mode, arg0, arg1);
5346
5347 case CONST_INT:
5348 /* Product of two constants. */
5349 return GEN_INT (INTVAL (arg0) * INTVAL (arg1));
5350
5351 case USE:
5352 /* invar * invar. Not giv. */
5353 return 0;
5354
5355 case MULT:
5356 /* (a * invar_1) * invar_2. Associate. */
5357 return simplify_giv_expr (gen_rtx (MULT, mode,
5358 XEXP (arg0, 0),
5359 gen_rtx (MULT, mode,
5360 XEXP (arg0, 1), arg1)),
5361 benefit);
5362
5363 case PLUS:
5364 /* (a + invar_1) * invar_2. Distribute. */
5365 return simplify_giv_expr (gen_rtx (PLUS, mode,
5366 gen_rtx (MULT, mode,
5367 XEXP (arg0, 0), arg1),
5368 gen_rtx (MULT, mode,
5369 XEXP (arg0, 1), arg1)),
5370 benefit);
5371
5372 default:
5373 abort ();
5374 }
5375
5376 case ASHIFT:
5377 /* Shift by constant is multiply by power of two. */
5378 if (GET_CODE (XEXP (x, 1)) != CONST_INT)
5379 return 0;
5380
5381 return simplify_giv_expr (gen_rtx (MULT, mode,
5382 XEXP (x, 0),
5383 GEN_INT ((HOST_WIDE_INT) 1
5384 << INTVAL (XEXP (x, 1)))),
5385 benefit);
5386
5387 case NEG:
5388 /* "-a" is "a * (-1)" */
5389 return simplify_giv_expr (gen_rtx (MULT, mode, XEXP (x, 0), constm1_rtx),
5390 benefit);
5391
5392 case NOT:
5393 /* "~a" is "-a - 1". Silly, but easy. */
5394 return simplify_giv_expr (gen_rtx (MINUS, mode,
5395 gen_rtx (NEG, mode, XEXP (x, 0)),
5396 const1_rtx),
5397 benefit);
5398
5399 case USE:
5400 /* Already in proper form for invariant. */
5401 return x;
5402
5403 case REG:
5404 /* If this is a new register, we can't deal with it. */
5405 if (REGNO (x) >= max_reg_before_loop)
5406 return 0;
5407
5408 /* Check for biv or giv. */
5409 switch (reg_iv_type[REGNO (x)])
5410 {
5411 case BASIC_INDUCT:
5412 return x;
5413 case GENERAL_INDUCT:
5414 {
5415 struct induction *v = reg_iv_info[REGNO (x)];
5416
5417 /* Form expression from giv and add benefit. Ensure this giv
5418 can derive another and subtract any needed adjustment if so. */
5419 *benefit += v->benefit;
5420 if (v->cant_derive)
5421 return 0;
5422
5423 tem = gen_rtx (PLUS, mode, gen_rtx (MULT, mode,
5424 v->src_reg, v->mult_val),
5425 v->add_val);
5426 if (v->derive_adjustment)
5427 tem = gen_rtx (MINUS, mode, tem, v->derive_adjustment);
5428 return simplify_giv_expr (tem, benefit);
5429 }
5430 }
5431
5432 /* Fall through to general case. */
5433 default:
5434 /* If invariant, return as USE (unless CONST_INT).
5435 Otherwise, not giv. */
5436 if (GET_CODE (x) == USE)
5437 x = XEXP (x, 0);
5438
5439 if (invariant_p (x) == 1)
5440 {
5441 if (GET_CODE (x) == CONST_INT)
5442 return x;
5443 else
5444 return gen_rtx (USE, mode, x);
5445 }
5446 else
5447 return 0;
5448 }
5449 }
5450 \f
5451 /* Help detect a giv that is calculated by several consecutive insns;
5452 for example,
5453 giv = biv * M
5454 giv = giv + A
5455 The caller has already identified the first insn P as having a giv as dest;
5456 we check that all other insns that set the same register follow
5457 immediately after P, that they alter nothing else,
5458 and that the result of the last is still a giv.
5459
5460 The value is 0 if the reg set in P is not really a giv.
5461 Otherwise, the value is the amount gained by eliminating
5462 all the consecutive insns that compute the value.
5463
5464 FIRST_BENEFIT is the amount gained by eliminating the first insn, P.
5465 SRC_REG is the reg of the biv; DEST_REG is the reg of the giv.
5466
5467 The coefficients of the ultimate giv value are stored in
5468 *MULT_VAL and *ADD_VAL. */
5469
5470 static int
5471 consec_sets_giv (first_benefit, p, src_reg, dest_reg,
5472 add_val, mult_val)
5473 int first_benefit;
5474 rtx p;
5475 rtx src_reg;
5476 rtx dest_reg;
5477 rtx *add_val;
5478 rtx *mult_val;
5479 {
5480 int count;
5481 enum rtx_code code;
5482 int benefit;
5483 rtx temp;
5484 rtx set;
5485
5486 /* Indicate that this is a giv so that we can update the value produced in
5487 each insn of the multi-insn sequence.
5488
5489 This induction structure will be used only by the call to
5490 general_induction_var below, so we can allocate it on our stack.
5491 If this is a giv, our caller will replace the induct var entry with
5492 a new induction structure. */
5493 struct induction *v
5494 = (struct induction *) alloca (sizeof (struct induction));
5495 v->src_reg = src_reg;
5496 v->mult_val = *mult_val;
5497 v->add_val = *add_val;
5498 v->benefit = first_benefit;
5499 v->cant_derive = 0;
5500 v->derive_adjustment = 0;
5501
5502 reg_iv_type[REGNO (dest_reg)] = GENERAL_INDUCT;
5503 reg_iv_info[REGNO (dest_reg)] = v;
5504
5505 count = n_times_set[REGNO (dest_reg)] - 1;
5506
5507 while (count > 0)
5508 {
5509 p = NEXT_INSN (p);
5510 code = GET_CODE (p);
5511
5512 /* If libcall, skip to end of call sequence. */
5513 if (code == INSN && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
5514 p = XEXP (temp, 0);
5515
5516 if (code == INSN
5517 && (set = single_set (p))
5518 && GET_CODE (SET_DEST (set)) == REG
5519 && SET_DEST (set) == dest_reg
5520 && ((benefit = general_induction_var (SET_SRC (set), &src_reg,
5521 add_val, mult_val))
5522 /* Giv created by equivalent expression. */
5523 || ((temp = find_reg_note (p, REG_EQUAL, NULL_RTX))
5524 && (benefit = general_induction_var (XEXP (temp, 0), &src_reg,
5525 add_val, mult_val))))
5526 && src_reg == v->src_reg)
5527 {
5528 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
5529 benefit += libcall_benefit (p);
5530
5531 count--;
5532 v->mult_val = *mult_val;
5533 v->add_val = *add_val;
5534 v->benefit = benefit;
5535 }
5536 else if (code != NOTE)
5537 {
5538 /* Allow insns that set something other than this giv to a
5539 constant. Such insns are needed on machines which cannot
5540 include long constants and should not disqualify a giv. */
5541 if (code == INSN
5542 && (set = single_set (p))
5543 && SET_DEST (set) != dest_reg
5544 && CONSTANT_P (SET_SRC (set)))
5545 continue;
5546
5547 reg_iv_type[REGNO (dest_reg)] = UNKNOWN_INDUCT;
5548 return 0;
5549 }
5550 }
5551
5552 return v->benefit;
5553 }
5554 \f
5555 /* Return an rtx, if any, that expresses giv G2 as a function of the register
5556 represented by G1. If no such expression can be found, or it is clear that
5557 it cannot possibly be a valid address, 0 is returned.
5558
5559 To perform the computation, we note that
5560 G1 = a * v + b and
5561 G2 = c * v + d
5562 where `v' is the biv.
5563
5564 So G2 = (c/a) * G1 + (d - b*c/a) */
5565
5566 #ifdef ADDRESS_COST
5567 static rtx
5568 express_from (g1, g2)
5569 struct induction *g1, *g2;
5570 {
5571 rtx mult, add;
5572
5573 /* The value that G1 will be multiplied by must be a constant integer. Also,
5574 the only chance we have of getting a valid address is if b*c/a (see above
5575 for notation) is also an integer. */
5576 if (GET_CODE (g1->mult_val) != CONST_INT
5577 || GET_CODE (g2->mult_val) != CONST_INT
5578 || GET_CODE (g1->add_val) != CONST_INT
5579 || g1->mult_val == const0_rtx
5580 || INTVAL (g2->mult_val) % INTVAL (g1->mult_val) != 0)
5581 return 0;
5582
5583 mult = GEN_INT (INTVAL (g2->mult_val) / INTVAL (g1->mult_val));
5584 add = plus_constant (g2->add_val, - INTVAL (g1->add_val) * INTVAL (mult));
5585
5586 /* Form simplified final result. */
5587 if (mult == const0_rtx)
5588 return add;
5589 else if (mult == const1_rtx)
5590 mult = g1->dest_reg;
5591 else
5592 mult = gen_rtx (MULT, g2->mode, g1->dest_reg, mult);
5593
5594 if (add == const0_rtx)
5595 return mult;
5596 else
5597 return gen_rtx (PLUS, g2->mode, mult, add);
5598 }
5599 #endif
5600 \f
5601 /* Return 1 if giv G2 can be combined with G1. This means that G2 can use
5602 (either directly or via an address expression) a register used to represent
5603 G1. Set g2->new_reg to a represtation of G1 (normally just
5604 g1->dest_reg). */
5605
5606 static int
5607 combine_givs_p (g1, g2)
5608 struct induction *g1, *g2;
5609 {
5610 rtx tem;
5611
5612 /* If these givs are identical, they can be combined. */
5613 if (rtx_equal_p (g1->mult_val, g2->mult_val)
5614 && rtx_equal_p (g1->add_val, g2->add_val))
5615 {
5616 g2->new_reg = g1->dest_reg;
5617 return 1;
5618 }
5619
5620 #ifdef ADDRESS_COST
5621 /* If G2 can be expressed as a function of G1 and that function is valid
5622 as an address and no more expensive than using a register for G2,
5623 the expression of G2 in terms of G1 can be used. */
5624 if (g2->giv_type == DEST_ADDR
5625 && (tem = express_from (g1, g2)) != 0
5626 && memory_address_p (g2->mem_mode, tem)
5627 && ADDRESS_COST (tem) <= ADDRESS_COST (*g2->location))
5628 {
5629 g2->new_reg = tem;
5630 return 1;
5631 }
5632 #endif
5633
5634 return 0;
5635 }
5636 \f
5637 #ifdef GIV_SORT_CRITERION
5638 /* Compare two givs and sort the most desirable one for combinations first.
5639 This is used only in one qsort call below. */
5640
5641 static int
5642 giv_sort (x, y)
5643 struct induction **x, **y;
5644 {
5645 GIV_SORT_CRITERION (*x, *y);
5646
5647 return 0;
5648 }
5649 #endif
5650
5651 /* Check all pairs of givs for iv_class BL and see if any can be combined with
5652 any other. If so, point SAME to the giv combined with and set NEW_REG to
5653 be an expression (in terms of the other giv's DEST_REG) equivalent to the
5654 giv. Also, update BENEFIT and related fields for cost/benefit analysis. */
5655
5656 static void
5657 combine_givs (bl)
5658 struct iv_class *bl;
5659 {
5660 struct induction *g1, *g2, **giv_array, *temp_iv;
5661 int i, j, giv_count, pass;
5662
5663 /* Count givs, because bl->giv_count is incorrect here. */
5664 giv_count = 0;
5665 for (g1 = bl->giv; g1; g1 = g1->next_iv)
5666 giv_count++;
5667
5668 giv_array
5669 = (struct induction **) alloca (giv_count * sizeof (struct induction *));
5670 i = 0;
5671 for (g1 = bl->giv; g1; g1 = g1->next_iv)
5672 giv_array[i++] = g1;
5673
5674 #ifdef GIV_SORT_CRITERION
5675 /* Sort the givs if GIV_SORT_CRITERION is defined.
5676 This is usually defined for processors which lack
5677 negative register offsets so more givs may be combined. */
5678
5679 if (loop_dump_stream)
5680 fprintf (loop_dump_stream, "%d givs counted, sorting...\n", giv_count);
5681
5682 qsort (giv_array, giv_count, sizeof (struct induction *), giv_sort);
5683 #endif
5684
5685 for (i = 0; i < giv_count; i++)
5686 {
5687 g1 = giv_array[i];
5688 for (pass = 0; pass <= 1; pass++)
5689 for (j = 0; j < giv_count; j++)
5690 {
5691 g2 = giv_array[j];
5692 if (g1 != g2
5693 /* First try to combine with replaceable givs, then all givs. */
5694 && (g1->replaceable || pass == 1)
5695 /* If either has already been combined or is to be ignored, can't
5696 combine. */
5697 && ! g1->ignore && ! g2->ignore && ! g1->same && ! g2->same
5698 /* If something has been based on G2, G2 cannot itself be based
5699 on something else. */
5700 && ! g2->combined_with
5701 && combine_givs_p (g1, g2))
5702 {
5703 /* g2->new_reg set by `combine_givs_p' */
5704 g2->same = g1;
5705 g1->combined_with = 1;
5706
5707 /* If one of these givs is a DEST_REG that was only used
5708 once, by the other giv, this is actually a single use.
5709 The DEST_REG has the correct cost, while the other giv
5710 counts the REG use too often. */
5711 if (g2->giv_type == DEST_REG
5712 && n_times_used[REGNO (g2->dest_reg)] == 1
5713 && reg_mentioned_p (g2->dest_reg, PATTERN (g1->insn)))
5714 g1->benefit = g2->benefit;
5715 else if (g1->giv_type != DEST_REG
5716 || n_times_used[REGNO (g1->dest_reg)] != 1
5717 || ! reg_mentioned_p (g1->dest_reg,
5718 PATTERN (g2->insn)))
5719 {
5720 g1->benefit += g2->benefit;
5721 g1->times_used += g2->times_used;
5722 }
5723 /* ??? The new final_[bg]iv_value code does a much better job
5724 of finding replaceable giv's, and hence this code may no
5725 longer be necessary. */
5726 if (! g2->replaceable && REG_USERVAR_P (g2->dest_reg))
5727 g1->benefit -= copy_cost;
5728 g1->lifetime += g2->lifetime;
5729
5730 if (loop_dump_stream)
5731 fprintf (loop_dump_stream, "giv at %d combined with giv at %d\n",
5732 INSN_UID (g2->insn), INSN_UID (g1->insn));
5733 }
5734 }
5735 }
5736 }
5737 \f
5738 /* EMIT code before INSERT_BEFORE to set REG = B * M + A. */
5739
5740 void
5741 emit_iv_add_mult (b, m, a, reg, insert_before)
5742 rtx b; /* initial value of basic induction variable */
5743 rtx m; /* multiplicative constant */
5744 rtx a; /* additive constant */
5745 rtx reg; /* destination register */
5746 rtx insert_before;
5747 {
5748 rtx seq;
5749 rtx result;
5750
5751 /* Prevent unexpected sharing of these rtx. */
5752 a = copy_rtx (a);
5753 b = copy_rtx (b);
5754
5755 /* Increase the lifetime of any invariants moved further in code. */
5756 update_reg_last_use (a, insert_before);
5757 update_reg_last_use (b, insert_before);
5758 update_reg_last_use (m, insert_before);
5759
5760 start_sequence ();
5761 result = expand_mult_add (b, reg, m, a, GET_MODE (reg), 0);
5762 if (reg != result)
5763 emit_move_insn (reg, result);
5764 seq = gen_sequence ();
5765 end_sequence ();
5766
5767 emit_insn_before (seq, insert_before);
5768 }
5769 \f
5770 /* Test whether A * B can be computed without
5771 an actual multiply insn. Value is 1 if so. */
5772
5773 static int
5774 product_cheap_p (a, b)
5775 rtx a;
5776 rtx b;
5777 {
5778 int i;
5779 rtx tmp;
5780 struct obstack *old_rtl_obstack = rtl_obstack;
5781 char *storage = (char *) obstack_alloc (&temp_obstack, 0);
5782 int win = 1;
5783
5784 /* If only one is constant, make it B. */
5785 if (GET_CODE (a) == CONST_INT)
5786 tmp = a, a = b, b = tmp;
5787
5788 /* If first constant, both constant, so don't need multiply. */
5789 if (GET_CODE (a) == CONST_INT)
5790 return 1;
5791
5792 /* If second not constant, neither is constant, so would need multiply. */
5793 if (GET_CODE (b) != CONST_INT)
5794 return 0;
5795
5796 /* One operand is constant, so might not need multiply insn. Generate the
5797 code for the multiply and see if a call or multiply, or long sequence
5798 of insns is generated. */
5799
5800 rtl_obstack = &temp_obstack;
5801 start_sequence ();
5802 expand_mult (GET_MODE (a), a, b, NULL_RTX, 0);
5803 tmp = gen_sequence ();
5804 end_sequence ();
5805
5806 if (GET_CODE (tmp) == SEQUENCE)
5807 {
5808 if (XVEC (tmp, 0) == 0)
5809 win = 1;
5810 else if (XVECLEN (tmp, 0) > 3)
5811 win = 0;
5812 else
5813 for (i = 0; i < XVECLEN (tmp, 0); i++)
5814 {
5815 rtx insn = XVECEXP (tmp, 0, i);
5816
5817 if (GET_CODE (insn) != INSN
5818 || (GET_CODE (PATTERN (insn)) == SET
5819 && GET_CODE (SET_SRC (PATTERN (insn))) == MULT)
5820 || (GET_CODE (PATTERN (insn)) == PARALLEL
5821 && GET_CODE (XVECEXP (PATTERN (insn), 0, 0)) == SET
5822 && GET_CODE (SET_SRC (XVECEXP (PATTERN (insn), 0, 0))) == MULT))
5823 {
5824 win = 0;
5825 break;
5826 }
5827 }
5828 }
5829 else if (GET_CODE (tmp) == SET
5830 && GET_CODE (SET_SRC (tmp)) == MULT)
5831 win = 0;
5832 else if (GET_CODE (tmp) == PARALLEL
5833 && GET_CODE (XVECEXP (tmp, 0, 0)) == SET
5834 && GET_CODE (SET_SRC (XVECEXP (tmp, 0, 0))) == MULT)
5835 win = 0;
5836
5837 /* Free any storage we obtained in generating this multiply and restore rtl
5838 allocation to its normal obstack. */
5839 obstack_free (&temp_obstack, storage);
5840 rtl_obstack = old_rtl_obstack;
5841
5842 return win;
5843 }
5844 \f
5845 /* Check to see if loop can be terminated by a "decrement and branch until
5846 zero" instruction. If so, add a REG_NONNEG note to the branch insn if so.
5847 Also try reversing an increment loop to a decrement loop
5848 to see if the optimization can be performed.
5849 Value is nonzero if optimization was performed. */
5850
5851 /* This is useful even if the architecture doesn't have such an insn,
5852 because it might change a loops which increments from 0 to n to a loop
5853 which decrements from n to 0. A loop that decrements to zero is usually
5854 faster than one that increments from zero. */
5855
5856 /* ??? This could be rewritten to use some of the loop unrolling procedures,
5857 such as approx_final_value, biv_total_increment, loop_iterations, and
5858 final_[bg]iv_value. */
5859
5860 static int
5861 check_dbra_loop (loop_end, insn_count, loop_start)
5862 rtx loop_end;
5863 int insn_count;
5864 rtx loop_start;
5865 {
5866 struct iv_class *bl;
5867 rtx reg;
5868 rtx jump_label;
5869 rtx final_value;
5870 rtx start_value;
5871 rtx new_add_val;
5872 rtx comparison;
5873 rtx before_comparison;
5874 rtx p;
5875
5876 /* If last insn is a conditional branch, and the insn before tests a
5877 register value, try to optimize it. Otherwise, we can't do anything. */
5878
5879 comparison = get_condition_for_loop (PREV_INSN (loop_end));
5880 if (comparison == 0)
5881 return 0;
5882
5883 /* Check all of the bivs to see if the compare uses one of them.
5884 Skip biv's set more than once because we can't guarantee that
5885 it will be zero on the last iteration. Also skip if the biv is
5886 used between its update and the test insn. */
5887
5888 for (bl = loop_iv_list; bl; bl = bl->next)
5889 {
5890 if (bl->biv_count == 1
5891 && bl->biv->dest_reg == XEXP (comparison, 0)
5892 && ! reg_used_between_p (regno_reg_rtx[bl->regno], bl->biv->insn,
5893 PREV_INSN (PREV_INSN (loop_end))))
5894 break;
5895 }
5896
5897 if (! bl)
5898 return 0;
5899
5900 /* Look for the case where the basic induction variable is always
5901 nonnegative, and equals zero on the last iteration.
5902 In this case, add a reg_note REG_NONNEG, which allows the
5903 m68k DBRA instruction to be used. */
5904
5905 if (((GET_CODE (comparison) == GT
5906 && GET_CODE (XEXP (comparison, 1)) == CONST_INT
5907 && INTVAL (XEXP (comparison, 1)) == -1)
5908 || (GET_CODE (comparison) == NE && XEXP (comparison, 1) == const0_rtx))
5909 && GET_CODE (bl->biv->add_val) == CONST_INT
5910 && INTVAL (bl->biv->add_val) < 0)
5911 {
5912 /* Initial value must be greater than 0,
5913 init_val % -dec_value == 0 to ensure that it equals zero on
5914 the last iteration */
5915
5916 if (GET_CODE (bl->initial_value) == CONST_INT
5917 && INTVAL (bl->initial_value) > 0
5918 && (INTVAL (bl->initial_value) %
5919 (-INTVAL (bl->biv->add_val))) == 0)
5920 {
5921 /* register always nonnegative, add REG_NOTE to branch */
5922 REG_NOTES (PREV_INSN (loop_end))
5923 = gen_rtx (EXPR_LIST, REG_NONNEG, NULL_RTX,
5924 REG_NOTES (PREV_INSN (loop_end)));
5925 bl->nonneg = 1;
5926
5927 return 1;
5928 }
5929
5930 /* If the decrement is 1 and the value was tested as >= 0 before
5931 the loop, then we can safely optimize. */
5932 for (p = loop_start; p; p = PREV_INSN (p))
5933 {
5934 if (GET_CODE (p) == CODE_LABEL)
5935 break;
5936 if (GET_CODE (p) != JUMP_INSN)
5937 continue;
5938
5939 before_comparison = get_condition_for_loop (p);
5940 if (before_comparison
5941 && XEXP (before_comparison, 0) == bl->biv->dest_reg
5942 && GET_CODE (before_comparison) == LT
5943 && XEXP (before_comparison, 1) == const0_rtx
5944 && ! reg_set_between_p (bl->biv->dest_reg, p, loop_start)
5945 && INTVAL (bl->biv->add_val) == -1)
5946 {
5947 REG_NOTES (PREV_INSN (loop_end))
5948 = gen_rtx (EXPR_LIST, REG_NONNEG, NULL_RTX,
5949 REG_NOTES (PREV_INSN (loop_end)));
5950 bl->nonneg = 1;
5951
5952 return 1;
5953 }
5954 }
5955 }
5956 else if (num_mem_sets <= 1)
5957 {
5958 /* Try to change inc to dec, so can apply above optimization. */
5959 /* Can do this if:
5960 all registers modified are induction variables or invariant,
5961 all memory references have non-overlapping addresses
5962 (obviously true if only one write)
5963 allow 2 insns for the compare/jump at the end of the loop. */
5964 /* Also, we must avoid any instructions which use both the reversed
5965 biv and another biv. Such instructions will fail if the loop is
5966 reversed. We meet this condition by requiring that either
5967 no_use_except_counting is true, or else that there is only
5968 one biv. */
5969 int num_nonfixed_reads = 0;
5970 /* 1 if the iteration var is used only to count iterations. */
5971 int no_use_except_counting = 0;
5972 /* 1 if the loop has no memory store, or it has a single memory store
5973 which is reversible. */
5974 int reversible_mem_store = 1;
5975
5976 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
5977 if (GET_RTX_CLASS (GET_CODE (p)) == 'i')
5978 num_nonfixed_reads += count_nonfixed_reads (PATTERN (p));
5979
5980 if (bl->giv_count == 0
5981 && ! loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]])
5982 {
5983 rtx bivreg = regno_reg_rtx[bl->regno];
5984
5985 /* If there are no givs for this biv, and the only exit is the
5986 fall through at the end of the the loop, then
5987 see if perhaps there are no uses except to count. */
5988 no_use_except_counting = 1;
5989 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
5990 if (GET_RTX_CLASS (GET_CODE (p)) == 'i')
5991 {
5992 rtx set = single_set (p);
5993
5994 if (set && GET_CODE (SET_DEST (set)) == REG
5995 && REGNO (SET_DEST (set)) == bl->regno)
5996 /* An insn that sets the biv is okay. */
5997 ;
5998 else if (p == prev_nonnote_insn (prev_nonnote_insn (loop_end))
5999 || p == prev_nonnote_insn (loop_end))
6000 /* Don't bother about the end test. */
6001 ;
6002 else if (reg_mentioned_p (bivreg, PATTERN (p)))
6003 /* Any other use of the biv is no good. */
6004 {
6005 no_use_except_counting = 0;
6006 break;
6007 }
6008 }
6009 }
6010
6011 /* If the loop has a single store, and the destination address is
6012 invariant, then we can't reverse the loop, because this address
6013 might then have the wrong value at loop exit.
6014 This would work if the source was invariant also, however, in that
6015 case, the insn should have been moved out of the loop. */
6016
6017 if (num_mem_sets == 1)
6018 reversible_mem_store
6019 = (! unknown_address_altered
6020 && ! invariant_p (XEXP (loop_store_mems[0], 0)));
6021
6022 /* This code only acts for innermost loops. Also it simplifies
6023 the memory address check by only reversing loops with
6024 zero or one memory access.
6025 Two memory accesses could involve parts of the same array,
6026 and that can't be reversed. */
6027
6028 if (num_nonfixed_reads <= 1
6029 && !loop_has_call
6030 && !loop_has_volatile
6031 && reversible_mem_store
6032 && (no_use_except_counting
6033 || ((bl->giv_count + bl->biv_count + num_mem_sets
6034 + num_movables + 2 == insn_count)
6035 && (bl == loop_iv_list && bl->next == 0))))
6036 {
6037 rtx tem;
6038
6039 /* Loop can be reversed. */
6040 if (loop_dump_stream)
6041 fprintf (loop_dump_stream, "Can reverse loop\n");
6042
6043 /* Now check other conditions:
6044 initial_value must be zero,
6045 final_value % add_val == 0, so that when reversed, the
6046 biv will be zero on the last iteration.
6047
6048 This test can probably be improved since +/- 1 in the constant
6049 can be obtained by changing LT to LE and vice versa; this is
6050 confusing. */
6051
6052 if (comparison && bl->initial_value == const0_rtx
6053 && GET_CODE (XEXP (comparison, 1)) == CONST_INT
6054 /* LE gets turned into LT */
6055 && GET_CODE (comparison) == LT
6056 && (INTVAL (XEXP (comparison, 1))
6057 % INTVAL (bl->biv->add_val)) == 0)
6058 {
6059 /* Register will always be nonnegative, with value
6060 0 on last iteration if loop reversed */
6061
6062 /* Save some info needed to produce the new insns. */
6063 reg = bl->biv->dest_reg;
6064 jump_label = XEXP (SET_SRC (PATTERN (PREV_INSN (loop_end))), 1);
6065 if (jump_label == pc_rtx)
6066 jump_label = XEXP (SET_SRC (PATTERN (PREV_INSN (loop_end))), 2);
6067 new_add_val = GEN_INT (- INTVAL (bl->biv->add_val));
6068
6069 final_value = XEXP (comparison, 1);
6070 start_value = GEN_INT (INTVAL (XEXP (comparison, 1))
6071 - INTVAL (bl->biv->add_val));
6072
6073 /* Initialize biv to start_value before loop start.
6074 The old initializing insn will be deleted as a
6075 dead store by flow.c. */
6076 emit_insn_before (gen_move_insn (reg, start_value), loop_start);
6077
6078 /* Add insn to decrement register, and delete insn
6079 that incremented the register. */
6080 p = emit_insn_before (gen_add2_insn (reg, new_add_val),
6081 bl->biv->insn);
6082 delete_insn (bl->biv->insn);
6083
6084 /* Update biv info to reflect its new status. */
6085 bl->biv->insn = p;
6086 bl->initial_value = start_value;
6087 bl->biv->add_val = new_add_val;
6088
6089 /* Inc LABEL_NUSES so that delete_insn will
6090 not delete the label. */
6091 LABEL_NUSES (XEXP (jump_label, 0)) ++;
6092
6093 /* Emit an insn after the end of the loop to set the biv's
6094 proper exit value if it is used anywhere outside the loop. */
6095 if ((regno_last_uid[bl->regno]
6096 != INSN_UID (PREV_INSN (PREV_INSN (loop_end))))
6097 || ! bl->init_insn
6098 || regno_first_uid[bl->regno] != INSN_UID (bl->init_insn))
6099 emit_insn_after (gen_move_insn (reg, final_value),
6100 loop_end);
6101
6102 /* Delete compare/branch at end of loop. */
6103 delete_insn (PREV_INSN (loop_end));
6104 delete_insn (PREV_INSN (loop_end));
6105
6106 /* Add new compare/branch insn at end of loop. */
6107 start_sequence ();
6108 emit_cmp_insn (reg, const0_rtx, GE, NULL_RTX,
6109 GET_MODE (reg), 0, 0);
6110 emit_jump_insn (gen_bge (XEXP (jump_label, 0)));
6111 tem = gen_sequence ();
6112 end_sequence ();
6113 emit_jump_insn_before (tem, loop_end);
6114
6115 for (tem = PREV_INSN (loop_end);
6116 tem && GET_CODE (tem) != JUMP_INSN; tem = PREV_INSN (tem))
6117 ;
6118 if (tem)
6119 {
6120 JUMP_LABEL (tem) = XEXP (jump_label, 0);
6121
6122 /* Increment of LABEL_NUSES done above. */
6123 /* Register is now always nonnegative,
6124 so add REG_NONNEG note to the branch. */
6125 REG_NOTES (tem) = gen_rtx (EXPR_LIST, REG_NONNEG, NULL_RTX,
6126 REG_NOTES (tem));
6127 }
6128
6129 bl->nonneg = 1;
6130
6131 /* Mark that this biv has been reversed. Each giv which depends
6132 on this biv, and which is also live past the end of the loop
6133 will have to be fixed up. */
6134
6135 bl->reversed = 1;
6136
6137 if (loop_dump_stream)
6138 fprintf (loop_dump_stream,
6139 "Reversed loop and added reg_nonneg\n");
6140
6141 return 1;
6142 }
6143 }
6144 }
6145
6146 return 0;
6147 }
6148 \f
6149 /* Verify whether the biv BL appears to be eliminable,
6150 based on the insns in the loop that refer to it.
6151 LOOP_START is the first insn of the loop, and END is the end insn.
6152
6153 If ELIMINATE_P is non-zero, actually do the elimination.
6154
6155 THRESHOLD and INSN_COUNT are from loop_optimize and are used to
6156 determine whether invariant insns should be placed inside or at the
6157 start of the loop. */
6158
6159 static int
6160 maybe_eliminate_biv (bl, loop_start, end, eliminate_p, threshold, insn_count)
6161 struct iv_class *bl;
6162 rtx loop_start;
6163 rtx end;
6164 int eliminate_p;
6165 int threshold, insn_count;
6166 {
6167 rtx reg = bl->biv->dest_reg;
6168 rtx p;
6169
6170 /* Scan all insns in the loop, stopping if we find one that uses the
6171 biv in a way that we cannot eliminate. */
6172
6173 for (p = loop_start; p != end; p = NEXT_INSN (p))
6174 {
6175 enum rtx_code code = GET_CODE (p);
6176 rtx where = threshold >= insn_count ? loop_start : p;
6177
6178 if ((code == INSN || code == JUMP_INSN || code == CALL_INSN)
6179 && reg_mentioned_p (reg, PATTERN (p))
6180 && ! maybe_eliminate_biv_1 (PATTERN (p), p, bl, eliminate_p, where))
6181 {
6182 if (loop_dump_stream)
6183 fprintf (loop_dump_stream,
6184 "Cannot eliminate biv %d: biv used in insn %d.\n",
6185 bl->regno, INSN_UID (p));
6186 break;
6187 }
6188 }
6189
6190 if (p == end)
6191 {
6192 if (loop_dump_stream)
6193 fprintf (loop_dump_stream, "biv %d %s eliminated.\n",
6194 bl->regno, eliminate_p ? "was" : "can be");
6195 return 1;
6196 }
6197
6198 return 0;
6199 }
6200 \f
6201 /* If BL appears in X (part of the pattern of INSN), see if we can
6202 eliminate its use. If so, return 1. If not, return 0.
6203
6204 If BIV does not appear in X, return 1.
6205
6206 If ELIMINATE_P is non-zero, actually do the elimination. WHERE indicates
6207 where extra insns should be added. Depending on how many items have been
6208 moved out of the loop, it will either be before INSN or at the start of
6209 the loop. */
6210
6211 static int
6212 maybe_eliminate_biv_1 (x, insn, bl, eliminate_p, where)
6213 rtx x, insn;
6214 struct iv_class *bl;
6215 int eliminate_p;
6216 rtx where;
6217 {
6218 enum rtx_code code = GET_CODE (x);
6219 rtx reg = bl->biv->dest_reg;
6220 enum machine_mode mode = GET_MODE (reg);
6221 struct induction *v;
6222 rtx arg, new, tem;
6223 int arg_operand;
6224 char *fmt;
6225 int i, j;
6226
6227 switch (code)
6228 {
6229 case REG:
6230 /* If we haven't already been able to do something with this BIV,
6231 we can't eliminate it. */
6232 if (x == reg)
6233 return 0;
6234 return 1;
6235
6236 case SET:
6237 /* If this sets the BIV, it is not a problem. */
6238 if (SET_DEST (x) == reg)
6239 return 1;
6240
6241 /* If this is an insn that defines a giv, it is also ok because
6242 it will go away when the giv is reduced. */
6243 for (v = bl->giv; v; v = v->next_iv)
6244 if (v->giv_type == DEST_REG && SET_DEST (x) == v->dest_reg)
6245 return 1;
6246
6247 #ifdef HAVE_cc0
6248 if (SET_DEST (x) == cc0_rtx && SET_SRC (x) == reg)
6249 {
6250 /* Can replace with any giv that was reduced and
6251 that has (MULT_VAL != 0) and (ADD_VAL == 0).
6252 Require a constant for MULT_VAL, so we know it's nonzero.
6253 ??? We disable this optimization to avoid potential
6254 overflows. */
6255
6256 for (v = bl->giv; v; v = v->next_iv)
6257 if (CONSTANT_P (v->mult_val) && v->mult_val != const0_rtx
6258 && v->add_val == const0_rtx
6259 && ! v->ignore && ! v->maybe_dead && v->always_computable
6260 && v->mode == mode
6261 && 0)
6262 {
6263 /* If the giv V had the auto-inc address optimization applied
6264 to it, and INSN occurs between the giv insn and the biv
6265 insn, then we must adjust the value used here.
6266 This is rare, so we don't bother to do so. */
6267 if (v->auto_inc_opt
6268 && ((INSN_LUID (v->insn) < INSN_LUID (insn)
6269 && INSN_LUID (insn) < INSN_LUID (bl->biv->insn))
6270 || (INSN_LUID (v->insn) > INSN_LUID (insn)
6271 && INSN_LUID (insn) > INSN_LUID (bl->biv->insn))))
6272 continue;
6273
6274 if (! eliminate_p)
6275 return 1;
6276
6277 /* If the giv has the opposite direction of change,
6278 then reverse the comparison. */
6279 if (INTVAL (v->mult_val) < 0)
6280 new = gen_rtx (COMPARE, GET_MODE (v->new_reg),
6281 const0_rtx, v->new_reg);
6282 else
6283 new = v->new_reg;
6284
6285 /* We can probably test that giv's reduced reg. */
6286 if (validate_change (insn, &SET_SRC (x), new, 0))
6287 return 1;
6288 }
6289
6290 /* Look for a giv with (MULT_VAL != 0) and (ADD_VAL != 0);
6291 replace test insn with a compare insn (cmp REDUCED_GIV ADD_VAL).
6292 Require a constant for MULT_VAL, so we know it's nonzero.
6293 ??? Do this only if ADD_VAL is a pointer to avoid a potential
6294 overflow problem. */
6295
6296 for (v = bl->giv; v; v = v->next_iv)
6297 if (CONSTANT_P (v->mult_val) && v->mult_val != const0_rtx
6298 && ! v->ignore && ! v->maybe_dead && v->always_computable
6299 && v->mode == mode
6300 && (GET_CODE (v->add_val) == SYMBOL_REF
6301 || GET_CODE (v->add_val) == LABEL_REF
6302 || GET_CODE (v->add_val) == CONST
6303 || (GET_CODE (v->add_val) == REG
6304 && REGNO_POINTER_FLAG (REGNO (v->add_val)))))
6305 {
6306 /* If the giv V had the auto-inc address optimization applied
6307 to it, and INSN occurs between the giv insn and the biv
6308 insn, then we must adjust the value used here.
6309 This is rare, so we don't bother to do so. */
6310 if (v->auto_inc_opt
6311 && ((INSN_LUID (v->insn) < INSN_LUID (insn)
6312 && INSN_LUID (insn) < INSN_LUID (bl->biv->insn))
6313 || (INSN_LUID (v->insn) > INSN_LUID (insn)
6314 && INSN_LUID (insn) > INSN_LUID (bl->biv->insn))))
6315 continue;
6316
6317 if (! eliminate_p)
6318 return 1;
6319
6320 /* If the giv has the opposite direction of change,
6321 then reverse the comparison. */
6322 if (INTVAL (v->mult_val) < 0)
6323 new = gen_rtx (COMPARE, VOIDmode, copy_rtx (v->add_val),
6324 v->new_reg);
6325 else
6326 new = gen_rtx (COMPARE, VOIDmode, v->new_reg,
6327 copy_rtx (v->add_val));
6328
6329 /* Replace biv with the giv's reduced register. */
6330 update_reg_last_use (v->add_val, insn);
6331 if (validate_change (insn, &SET_SRC (PATTERN (insn)), new, 0))
6332 return 1;
6333
6334 /* Insn doesn't support that constant or invariant. Copy it
6335 into a register (it will be a loop invariant.) */
6336 tem = gen_reg_rtx (GET_MODE (v->new_reg));
6337
6338 emit_insn_before (gen_move_insn (tem, copy_rtx (v->add_val)),
6339 where);
6340
6341 if (validate_change (insn, &SET_SRC (PATTERN (insn)),
6342 gen_rtx (COMPARE, VOIDmode,
6343 v->new_reg, tem), 0))
6344 return 1;
6345 }
6346 }
6347 #endif
6348 break;
6349
6350 case COMPARE:
6351 case EQ: case NE:
6352 case GT: case GE: case GTU: case GEU:
6353 case LT: case LE: case LTU: case LEU:
6354 /* See if either argument is the biv. */
6355 if (XEXP (x, 0) == reg)
6356 arg = XEXP (x, 1), arg_operand = 1;
6357 else if (XEXP (x, 1) == reg)
6358 arg = XEXP (x, 0), arg_operand = 0;
6359 else
6360 break;
6361
6362 if (CONSTANT_P (arg))
6363 {
6364 /* First try to replace with any giv that has constant positive
6365 mult_val and constant add_val. We might be able to support
6366 negative mult_val, but it seems complex to do it in general. */
6367
6368 for (v = bl->giv; v; v = v->next_iv)
6369 if (CONSTANT_P (v->mult_val) && INTVAL (v->mult_val) > 0
6370 && (GET_CODE (v->add_val) == SYMBOL_REF
6371 || GET_CODE (v->add_val) == LABEL_REF
6372 || GET_CODE (v->add_val) == CONST
6373 || (GET_CODE (v->add_val) == REG
6374 && REGNO_POINTER_FLAG (REGNO (v->add_val))))
6375 && ! v->ignore && ! v->maybe_dead && v->always_computable
6376 && v->mode == mode)
6377 {
6378 /* If the giv V had the auto-inc address optimization applied
6379 to it, and INSN occurs between the giv insn and the biv
6380 insn, then we must adjust the value used here.
6381 This is rare, so we don't bother to do so. */
6382 if (v->auto_inc_opt
6383 && ((INSN_LUID (v->insn) < INSN_LUID (insn)
6384 && INSN_LUID (insn) < INSN_LUID (bl->biv->insn))
6385 || (INSN_LUID (v->insn) > INSN_LUID (insn)
6386 && INSN_LUID (insn) > INSN_LUID (bl->biv->insn))))
6387 continue;
6388
6389 if (! eliminate_p)
6390 return 1;
6391
6392 /* Replace biv with the giv's reduced reg. */
6393 XEXP (x, 1-arg_operand) = v->new_reg;
6394
6395 /* If all constants are actually constant integers and
6396 the derived constant can be directly placed in the COMPARE,
6397 do so. */
6398 if (GET_CODE (arg) == CONST_INT
6399 && GET_CODE (v->mult_val) == CONST_INT
6400 && GET_CODE (v->add_val) == CONST_INT
6401 && validate_change (insn, &XEXP (x, arg_operand),
6402 GEN_INT (INTVAL (arg)
6403 * INTVAL (v->mult_val)
6404 + INTVAL (v->add_val)), 0))
6405 return 1;
6406
6407 /* Otherwise, load it into a register. */
6408 tem = gen_reg_rtx (mode);
6409 emit_iv_add_mult (arg, v->mult_val, v->add_val, tem, where);
6410 if (validate_change (insn, &XEXP (x, arg_operand), tem, 0))
6411 return 1;
6412
6413 /* If that failed, put back the change we made above. */
6414 XEXP (x, 1-arg_operand) = reg;
6415 }
6416
6417 /* Look for giv with positive constant mult_val and nonconst add_val.
6418 Insert insns to calculate new compare value.
6419 ??? Turn this off due to possible overflow. */
6420
6421 for (v = bl->giv; v; v = v->next_iv)
6422 if (CONSTANT_P (v->mult_val) && INTVAL (v->mult_val) > 0
6423 && ! v->ignore && ! v->maybe_dead && v->always_computable
6424 && v->mode == mode
6425 && 0)
6426 {
6427 rtx tem;
6428
6429 /* If the giv V had the auto-inc address optimization applied
6430 to it, and INSN occurs between the giv insn and the biv
6431 insn, then we must adjust the value used here.
6432 This is rare, so we don't bother to do so. */
6433 if (v->auto_inc_opt
6434 && ((INSN_LUID (v->insn) < INSN_LUID (insn)
6435 && INSN_LUID (insn) < INSN_LUID (bl->biv->insn))
6436 || (INSN_LUID (v->insn) > INSN_LUID (insn)
6437 && INSN_LUID (insn) > INSN_LUID (bl->biv->insn))))
6438 continue;
6439
6440 if (! eliminate_p)
6441 return 1;
6442
6443 tem = gen_reg_rtx (mode);
6444
6445 /* Replace biv with giv's reduced register. */
6446 validate_change (insn, &XEXP (x, 1 - arg_operand),
6447 v->new_reg, 1);
6448
6449 /* Compute value to compare against. */
6450 emit_iv_add_mult (arg, v->mult_val, v->add_val, tem, where);
6451 /* Use it in this insn. */
6452 validate_change (insn, &XEXP (x, arg_operand), tem, 1);
6453 if (apply_change_group ())
6454 return 1;
6455 }
6456 }
6457 else if (GET_CODE (arg) == REG || GET_CODE (arg) == MEM)
6458 {
6459 if (invariant_p (arg) == 1)
6460 {
6461 /* Look for giv with constant positive mult_val and nonconst
6462 add_val. Insert insns to compute new compare value.
6463 ??? Turn this off due to possible overflow. */
6464
6465 for (v = bl->giv; v; v = v->next_iv)
6466 if (CONSTANT_P (v->mult_val) && INTVAL (v->mult_val) > 0
6467 && ! v->ignore && ! v->maybe_dead && v->always_computable
6468 && v->mode == mode
6469 && 0)
6470 {
6471 rtx tem;
6472
6473 /* If the giv V had the auto-inc address optimization applied
6474 to it, and INSN occurs between the giv insn and the biv
6475 insn, then we must adjust the value used here.
6476 This is rare, so we don't bother to do so. */
6477 if (v->auto_inc_opt
6478 && ((INSN_LUID (v->insn) < INSN_LUID (insn)
6479 && INSN_LUID (insn) < INSN_LUID (bl->biv->insn))
6480 || (INSN_LUID (v->insn) > INSN_LUID (insn)
6481 && INSN_LUID (insn) > INSN_LUID (bl->biv->insn))))
6482 continue;
6483
6484 if (! eliminate_p)
6485 return 1;
6486
6487 tem = gen_reg_rtx (mode);
6488
6489 /* Replace biv with giv's reduced register. */
6490 validate_change (insn, &XEXP (x, 1 - arg_operand),
6491 v->new_reg, 1);
6492
6493 /* Compute value to compare against. */
6494 emit_iv_add_mult (arg, v->mult_val, v->add_val,
6495 tem, where);
6496 validate_change (insn, &XEXP (x, arg_operand), tem, 1);
6497 if (apply_change_group ())
6498 return 1;
6499 }
6500 }
6501
6502 /* This code has problems. Basically, you can't know when
6503 seeing if we will eliminate BL, whether a particular giv
6504 of ARG will be reduced. If it isn't going to be reduced,
6505 we can't eliminate BL. We can try forcing it to be reduced,
6506 but that can generate poor code.
6507
6508 The problem is that the benefit of reducing TV, below should
6509 be increased if BL can actually be eliminated, but this means
6510 we might have to do a topological sort of the order in which
6511 we try to process biv. It doesn't seem worthwhile to do
6512 this sort of thing now. */
6513
6514 #if 0
6515 /* Otherwise the reg compared with had better be a biv. */
6516 if (GET_CODE (arg) != REG
6517 || reg_iv_type[REGNO (arg)] != BASIC_INDUCT)
6518 return 0;
6519
6520 /* Look for a pair of givs, one for each biv,
6521 with identical coefficients. */
6522 for (v = bl->giv; v; v = v->next_iv)
6523 {
6524 struct induction *tv;
6525
6526 if (v->ignore || v->maybe_dead || v->mode != mode)
6527 continue;
6528
6529 for (tv = reg_biv_class[REGNO (arg)]->giv; tv; tv = tv->next_iv)
6530 if (! tv->ignore && ! tv->maybe_dead
6531 && rtx_equal_p (tv->mult_val, v->mult_val)
6532 && rtx_equal_p (tv->add_val, v->add_val)
6533 && tv->mode == mode)
6534 {
6535 /* If the giv V had the auto-inc address optimization applied
6536 to it, and INSN occurs between the giv insn and the biv
6537 insn, then we must adjust the value used here.
6538 This is rare, so we don't bother to do so. */
6539 if (v->auto_inc_opt
6540 && ((INSN_LUID (v->insn) < INSN_LUID (insn)
6541 && INSN_LUID (insn) < INSN_LUID (bl->biv->insn))
6542 || (INSN_LUID (v->insn) > INSN_LUID (insn)
6543 && INSN_LUID (insn) > INSN_LUID (bl->biv->insn))))
6544 continue;
6545
6546 if (! eliminate_p)
6547 return 1;
6548
6549 /* Replace biv with its giv's reduced reg. */
6550 XEXP (x, 1-arg_operand) = v->new_reg;
6551 /* Replace other operand with the other giv's
6552 reduced reg. */
6553 XEXP (x, arg_operand) = tv->new_reg;
6554 return 1;
6555 }
6556 }
6557 #endif
6558 }
6559
6560 /* If we get here, the biv can't be eliminated. */
6561 return 0;
6562
6563 case MEM:
6564 /* If this address is a DEST_ADDR giv, it doesn't matter if the
6565 biv is used in it, since it will be replaced. */
6566 for (v = bl->giv; v; v = v->next_iv)
6567 if (v->giv_type == DEST_ADDR && v->location == &XEXP (x, 0))
6568 return 1;
6569 break;
6570 }
6571
6572 /* See if any subexpression fails elimination. */
6573 fmt = GET_RTX_FORMAT (code);
6574 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
6575 {
6576 switch (fmt[i])
6577 {
6578 case 'e':
6579 if (! maybe_eliminate_biv_1 (XEXP (x, i), insn, bl,
6580 eliminate_p, where))
6581 return 0;
6582 break;
6583
6584 case 'E':
6585 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
6586 if (! maybe_eliminate_biv_1 (XVECEXP (x, i, j), insn, bl,
6587 eliminate_p, where))
6588 return 0;
6589 break;
6590 }
6591 }
6592
6593 return 1;
6594 }
6595 \f
6596 /* Return nonzero if the last use of REG
6597 is in an insn following INSN in the same basic block. */
6598
6599 static int
6600 last_use_this_basic_block (reg, insn)
6601 rtx reg;
6602 rtx insn;
6603 {
6604 rtx n;
6605 for (n = insn;
6606 n && GET_CODE (n) != CODE_LABEL && GET_CODE (n) != JUMP_INSN;
6607 n = NEXT_INSN (n))
6608 {
6609 if (regno_last_uid[REGNO (reg)] == INSN_UID (n))
6610 return 1;
6611 }
6612 return 0;
6613 }
6614 \f
6615 /* Called via `note_stores' to record the initial value of a biv. Here we
6616 just record the location of the set and process it later. */
6617
6618 static void
6619 record_initial (dest, set)
6620 rtx dest;
6621 rtx set;
6622 {
6623 struct iv_class *bl;
6624
6625 if (GET_CODE (dest) != REG
6626 || REGNO (dest) >= max_reg_before_loop
6627 || reg_iv_type[REGNO (dest)] != BASIC_INDUCT)
6628 return;
6629
6630 bl = reg_biv_class[REGNO (dest)];
6631
6632 /* If this is the first set found, record it. */
6633 if (bl->init_insn == 0)
6634 {
6635 bl->init_insn = note_insn;
6636 bl->init_set = set;
6637 }
6638 }
6639 \f
6640 /* If any of the registers in X are "old" and currently have a last use earlier
6641 than INSN, update them to have a last use of INSN. Their actual last use
6642 will be the previous insn but it will not have a valid uid_luid so we can't
6643 use it. */
6644
6645 static void
6646 update_reg_last_use (x, insn)
6647 rtx x;
6648 rtx insn;
6649 {
6650 /* Check for the case where INSN does not have a valid luid. In this case,
6651 there is no need to modify the regno_last_uid, as this can only happen
6652 when code is inserted after the loop_end to set a pseudo's final value,
6653 and hence this insn will never be the last use of x. */
6654 if (GET_CODE (x) == REG && REGNO (x) < max_reg_before_loop
6655 && INSN_UID (insn) < max_uid_for_loop
6656 && uid_luid[regno_last_uid[REGNO (x)]] < uid_luid[INSN_UID (insn)])
6657 regno_last_uid[REGNO (x)] = INSN_UID (insn);
6658 else
6659 {
6660 register int i, j;
6661 register char *fmt = GET_RTX_FORMAT (GET_CODE (x));
6662 for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
6663 {
6664 if (fmt[i] == 'e')
6665 update_reg_last_use (XEXP (x, i), insn);
6666 else if (fmt[i] == 'E')
6667 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
6668 update_reg_last_use (XVECEXP (x, i, j), insn);
6669 }
6670 }
6671 }
6672 \f
6673 /* Given a jump insn JUMP, return the condition that will cause it to branch
6674 to its JUMP_LABEL. If the condition cannot be understood, or is an
6675 inequality floating-point comparison which needs to be reversed, 0 will
6676 be returned.
6677
6678 If EARLIEST is non-zero, it is a pointer to a place where the earliest
6679 insn used in locating the condition was found. If a replacement test
6680 of the condition is desired, it should be placed in front of that
6681 insn and we will be sure that the inputs are still valid.
6682
6683 The condition will be returned in a canonical form to simplify testing by
6684 callers. Specifically:
6685
6686 (1) The code will always be a comparison operation (EQ, NE, GT, etc.).
6687 (2) Both operands will be machine operands; (cc0) will have been replaced.
6688 (3) If an operand is a constant, it will be the second operand.
6689 (4) (LE x const) will be replaced with (LT x <const+1>) and similarly
6690 for GE, GEU, and LEU. */
6691
6692 rtx
6693 get_condition (jump, earliest)
6694 rtx jump;
6695 rtx *earliest;
6696 {
6697 enum rtx_code code;
6698 rtx prev = jump;
6699 rtx set;
6700 rtx tem;
6701 rtx op0, op1;
6702 int reverse_code = 0;
6703 int did_reverse_condition = 0;
6704
6705 /* If this is not a standard conditional jump, we can't parse it. */
6706 if (GET_CODE (jump) != JUMP_INSN
6707 || ! condjump_p (jump) || simplejump_p (jump))
6708 return 0;
6709
6710 code = GET_CODE (XEXP (SET_SRC (PATTERN (jump)), 0));
6711 op0 = XEXP (XEXP (SET_SRC (PATTERN (jump)), 0), 0);
6712 op1 = XEXP (XEXP (SET_SRC (PATTERN (jump)), 0), 1);
6713
6714 if (earliest)
6715 *earliest = jump;
6716
6717 /* If this branches to JUMP_LABEL when the condition is false, reverse
6718 the condition. */
6719 if (GET_CODE (XEXP (SET_SRC (PATTERN (jump)), 2)) == LABEL_REF
6720 && XEXP (XEXP (SET_SRC (PATTERN (jump)), 2), 0) == JUMP_LABEL (jump))
6721 code = reverse_condition (code), did_reverse_condition ^= 1;
6722
6723 /* If we are comparing a register with zero, see if the register is set
6724 in the previous insn to a COMPARE or a comparison operation. Perform
6725 the same tests as a function of STORE_FLAG_VALUE as find_comparison_args
6726 in cse.c */
6727
6728 while (GET_RTX_CLASS (code) == '<' && op1 == CONST0_RTX (GET_MODE (op0)))
6729 {
6730 /* Set non-zero when we find something of interest. */
6731 rtx x = 0;
6732
6733 #ifdef HAVE_cc0
6734 /* If comparison with cc0, import actual comparison from compare
6735 insn. */
6736 if (op0 == cc0_rtx)
6737 {
6738 if ((prev = prev_nonnote_insn (prev)) == 0
6739 || GET_CODE (prev) != INSN
6740 || (set = single_set (prev)) == 0
6741 || SET_DEST (set) != cc0_rtx)
6742 return 0;
6743
6744 op0 = SET_SRC (set);
6745 op1 = CONST0_RTX (GET_MODE (op0));
6746 if (earliest)
6747 *earliest = prev;
6748 }
6749 #endif
6750
6751 /* If this is a COMPARE, pick up the two things being compared. */
6752 if (GET_CODE (op0) == COMPARE)
6753 {
6754 op1 = XEXP (op0, 1);
6755 op0 = XEXP (op0, 0);
6756 continue;
6757 }
6758 else if (GET_CODE (op0) != REG)
6759 break;
6760
6761 /* Go back to the previous insn. Stop if it is not an INSN. We also
6762 stop if it isn't a single set or if it has a REG_INC note because
6763 we don't want to bother dealing with it. */
6764
6765 if ((prev = prev_nonnote_insn (prev)) == 0
6766 || GET_CODE (prev) != INSN
6767 || FIND_REG_INC_NOTE (prev, 0)
6768 || (set = single_set (prev)) == 0)
6769 break;
6770
6771 /* If this is setting OP0, get what it sets it to if it looks
6772 relevant. */
6773 if (rtx_equal_p (SET_DEST (set), op0))
6774 {
6775 enum machine_mode inner_mode = GET_MODE (SET_SRC (set));
6776
6777 if ((GET_CODE (SET_SRC (set)) == COMPARE
6778 || (((code == NE
6779 || (code == LT
6780 && GET_MODE_CLASS (inner_mode) == MODE_INT
6781 && (GET_MODE_BITSIZE (inner_mode)
6782 <= HOST_BITS_PER_WIDE_INT)
6783 && (STORE_FLAG_VALUE
6784 & ((HOST_WIDE_INT) 1
6785 << (GET_MODE_BITSIZE (inner_mode) - 1))))
6786 #ifdef FLOAT_STORE_FLAG_VALUE
6787 || (code == LT
6788 && GET_MODE_CLASS (inner_mode) == MODE_FLOAT
6789 && FLOAT_STORE_FLAG_VALUE < 0)
6790 #endif
6791 ))
6792 && GET_RTX_CLASS (GET_CODE (SET_SRC (set))) == '<')))
6793 x = SET_SRC (set);
6794 else if (((code == EQ
6795 || (code == GE
6796 && (GET_MODE_BITSIZE (inner_mode)
6797 <= HOST_BITS_PER_WIDE_INT)
6798 && GET_MODE_CLASS (inner_mode) == MODE_INT
6799 && (STORE_FLAG_VALUE
6800 & ((HOST_WIDE_INT) 1
6801 << (GET_MODE_BITSIZE (inner_mode) - 1))))
6802 #ifdef FLOAT_STORE_FLAG_VALUE
6803 || (code == GE
6804 && GET_MODE_CLASS (inner_mode) == MODE_FLOAT
6805 && FLOAT_STORE_FLAG_VALUE < 0)
6806 #endif
6807 ))
6808 && GET_RTX_CLASS (GET_CODE (SET_SRC (set))) == '<')
6809 {
6810 /* We might have reversed a LT to get a GE here. But this wasn't
6811 actually the comparison of data, so we don't flag that we
6812 have had to reverse the condition. */
6813 did_reverse_condition ^= 1;
6814 reverse_code = 1;
6815 x = SET_SRC (set);
6816 }
6817 else
6818 break;
6819 }
6820
6821 else if (reg_set_p (op0, prev))
6822 /* If this sets OP0, but not directly, we have to give up. */
6823 break;
6824
6825 if (x)
6826 {
6827 if (GET_RTX_CLASS (GET_CODE (x)) == '<')
6828 code = GET_CODE (x);
6829 if (reverse_code)
6830 {
6831 code = reverse_condition (code);
6832 did_reverse_condition ^= 1;
6833 reverse_code = 0;
6834 }
6835
6836 op0 = XEXP (x, 0), op1 = XEXP (x, 1);
6837 if (earliest)
6838 *earliest = prev;
6839 }
6840 }
6841
6842 /* If constant is first, put it last. */
6843 if (CONSTANT_P (op0))
6844 code = swap_condition (code), tem = op0, op0 = op1, op1 = tem;
6845
6846 /* If OP0 is the result of a comparison, we weren't able to find what
6847 was really being compared, so fail. */
6848 if (GET_MODE_CLASS (GET_MODE (op0)) == MODE_CC)
6849 return 0;
6850
6851 /* Canonicalize any ordered comparison with integers involving equality
6852 if we can do computations in the relevant mode and we do not
6853 overflow. */
6854
6855 if (GET_CODE (op1) == CONST_INT
6856 && GET_MODE (op0) != VOIDmode
6857 && GET_MODE_BITSIZE (GET_MODE (op0)) <= HOST_BITS_PER_WIDE_INT)
6858 {
6859 HOST_WIDE_INT const_val = INTVAL (op1);
6860 unsigned HOST_WIDE_INT uconst_val = const_val;
6861 unsigned HOST_WIDE_INT max_val
6862 = (unsigned HOST_WIDE_INT) GET_MODE_MASK (GET_MODE (op0));
6863
6864 switch (code)
6865 {
6866 case LE:
6867 if (const_val != max_val >> 1)
6868 code = LT, op1 = GEN_INT (const_val + 1);
6869 break;
6870
6871 case GE:
6872 if (const_val
6873 != (((HOST_WIDE_INT) 1
6874 << (GET_MODE_BITSIZE (GET_MODE (op0)) - 1))))
6875 code = GT, op1 = GEN_INT (const_val - 1);
6876 break;
6877
6878 case LEU:
6879 if (uconst_val != max_val)
6880 code = LTU, op1 = GEN_INT (uconst_val + 1);
6881 break;
6882
6883 case GEU:
6884 if (uconst_val != 0)
6885 code = GTU, op1 = GEN_INT (uconst_val - 1);
6886 break;
6887 }
6888 }
6889
6890 /* If this was floating-point and we reversed anything other than an
6891 EQ or NE, return zero. */
6892 if (TARGET_FLOAT_FORMAT == IEEE_FLOAT_FORMAT
6893 && did_reverse_condition && code != NE && code != EQ
6894 && ! flag_fast_math
6895 && GET_MODE_CLASS (GET_MODE (op0)) == MODE_FLOAT)
6896 return 0;
6897
6898 #ifdef HAVE_cc0
6899 /* Never return CC0; return zero instead. */
6900 if (op0 == cc0_rtx)
6901 return 0;
6902 #endif
6903
6904 return gen_rtx (code, VOIDmode, op0, op1);
6905 }
6906
6907 /* Similar to above routine, except that we also put an invariant last
6908 unless both operands are invariants. */
6909
6910 rtx
6911 get_condition_for_loop (x)
6912 rtx x;
6913 {
6914 rtx comparison = get_condition (x, NULL_PTR);
6915
6916 if (comparison == 0
6917 || ! invariant_p (XEXP (comparison, 0))
6918 || invariant_p (XEXP (comparison, 1)))
6919 return comparison;
6920
6921 return gen_rtx (swap_condition (GET_CODE (comparison)), VOIDmode,
6922 XEXP (comparison, 1), XEXP (comparison, 0));
6923 }