Use accessor macros to access arrays based on regno; move many of the arrays into...
[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
2765 if ((x == frame_pointer_rtx || x == hard_frame_pointer_rtx
2766 || x == arg_pointer_rtx)
2767 && ! current_function_has_nonlocal_goto)
2768 return 1;
2769
2770 if (loop_has_call
2771 && REGNO (x) < FIRST_PSEUDO_REGISTER && call_used_regs[REGNO (x)])
2772 return 0;
2773
2774 if (n_times_set[REGNO (x)] < 0)
2775 return 2;
2776
2777 return n_times_set[REGNO (x)] == 0;
2778
2779 case MEM:
2780 /* Volatile memory references must be rejected. Do this before
2781 checking for read-only items, so that volatile read-only items
2782 will be rejected also. */
2783 if (MEM_VOLATILE_P (x))
2784 return 0;
2785
2786 /* Read-only items (such as constants in a constant pool) are
2787 invariant if their address is. */
2788 if (RTX_UNCHANGING_P (x))
2789 break;
2790
2791 /* If we filled the table (or had a subroutine call), any location
2792 in memory could have been clobbered. */
2793 if (unknown_address_altered)
2794 return 0;
2795
2796 /* See if there is any dependence between a store and this load. */
2797 for (i = loop_store_mems_idx - 1; i >= 0; i--)
2798 if (true_dependence (loop_store_mems[i], x))
2799 return 0;
2800
2801 /* It's not invalidated by a store in memory
2802 but we must still verify the address is invariant. */
2803 break;
2804
2805 case ASM_OPERANDS:
2806 /* Don't mess with insns declared volatile. */
2807 if (MEM_VOLATILE_P (x))
2808 return 0;
2809 }
2810
2811 fmt = GET_RTX_FORMAT (code);
2812 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2813 {
2814 if (fmt[i] == 'e')
2815 {
2816 int tem = invariant_p (XEXP (x, i));
2817 if (tem == 0)
2818 return 0;
2819 if (tem == 2)
2820 conditional = 1;
2821 }
2822 else if (fmt[i] == 'E')
2823 {
2824 register int j;
2825 for (j = 0; j < XVECLEN (x, i); j++)
2826 {
2827 int tem = invariant_p (XVECEXP (x, i, j));
2828 if (tem == 0)
2829 return 0;
2830 if (tem == 2)
2831 conditional = 1;
2832 }
2833
2834 }
2835 }
2836
2837 return 1 + conditional;
2838 }
2839
2840 \f
2841 /* Return nonzero if all the insns in the loop that set REG
2842 are INSN and the immediately following insns,
2843 and if each of those insns sets REG in an invariant way
2844 (not counting uses of REG in them).
2845
2846 The value is 2 if some of these insns are only conditionally invariant.
2847
2848 We assume that INSN itself is the first set of REG
2849 and that its source is invariant. */
2850
2851 static int
2852 consec_sets_invariant_p (reg, n_sets, insn)
2853 int n_sets;
2854 rtx reg, insn;
2855 {
2856 register rtx p = insn;
2857 register int regno = REGNO (reg);
2858 rtx temp;
2859 /* Number of sets we have to insist on finding after INSN. */
2860 int count = n_sets - 1;
2861 int old = n_times_set[regno];
2862 int value = 0;
2863 int this;
2864
2865 /* If N_SETS hit the limit, we can't rely on its value. */
2866 if (n_sets == 127)
2867 return 0;
2868
2869 n_times_set[regno] = 0;
2870
2871 while (count > 0)
2872 {
2873 register enum rtx_code code;
2874 rtx set;
2875
2876 p = NEXT_INSN (p);
2877 code = GET_CODE (p);
2878
2879 /* If library call, skip to end of of it. */
2880 if (code == INSN && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
2881 p = XEXP (temp, 0);
2882
2883 this = 0;
2884 if (code == INSN
2885 && (set = single_set (p))
2886 && GET_CODE (SET_DEST (set)) == REG
2887 && REGNO (SET_DEST (set)) == regno)
2888 {
2889 this = invariant_p (SET_SRC (set));
2890 if (this != 0)
2891 value |= this;
2892 else if (temp = find_reg_note (p, REG_EQUAL, NULL_RTX))
2893 {
2894 /* If this is a libcall, then any invariant REG_EQUAL note is OK.
2895 If this is an ordinary insn, then only CONSTANT_P REG_EQUAL
2896 notes are OK. */
2897 this = (CONSTANT_P (XEXP (temp, 0))
2898 || (find_reg_note (p, REG_RETVAL, NULL_RTX)
2899 && invariant_p (XEXP (temp, 0))));
2900 if (this != 0)
2901 value |= this;
2902 }
2903 }
2904 if (this != 0)
2905 count--;
2906 else if (code != NOTE)
2907 {
2908 n_times_set[regno] = old;
2909 return 0;
2910 }
2911 }
2912
2913 n_times_set[regno] = old;
2914 /* If invariant_p ever returned 2, we return 2. */
2915 return 1 + (value & 2);
2916 }
2917
2918 #if 0
2919 /* I don't think this condition is sufficient to allow INSN
2920 to be moved, so we no longer test it. */
2921
2922 /* Return 1 if all insns in the basic block of INSN and following INSN
2923 that set REG are invariant according to TABLE. */
2924
2925 static int
2926 all_sets_invariant_p (reg, insn, table)
2927 rtx reg, insn;
2928 short *table;
2929 {
2930 register rtx p = insn;
2931 register int regno = REGNO (reg);
2932
2933 while (1)
2934 {
2935 register enum rtx_code code;
2936 p = NEXT_INSN (p);
2937 code = GET_CODE (p);
2938 if (code == CODE_LABEL || code == JUMP_INSN)
2939 return 1;
2940 if (code == INSN && GET_CODE (PATTERN (p)) == SET
2941 && GET_CODE (SET_DEST (PATTERN (p))) == REG
2942 && REGNO (SET_DEST (PATTERN (p))) == regno)
2943 {
2944 if (!invariant_p (SET_SRC (PATTERN (p)), table))
2945 return 0;
2946 }
2947 }
2948 }
2949 #endif /* 0 */
2950 \f
2951 /* Look at all uses (not sets) of registers in X. For each, if it is
2952 the single use, set USAGE[REGNO] to INSN; if there was a previous use in
2953 a different insn, set USAGE[REGNO] to const0_rtx. */
2954
2955 static void
2956 find_single_use_in_loop (insn, x, usage)
2957 rtx insn;
2958 rtx x;
2959 rtx *usage;
2960 {
2961 enum rtx_code code = GET_CODE (x);
2962 char *fmt = GET_RTX_FORMAT (code);
2963 int i, j;
2964
2965 if (code == REG)
2966 usage[REGNO (x)]
2967 = (usage[REGNO (x)] != 0 && usage[REGNO (x)] != insn)
2968 ? const0_rtx : insn;
2969
2970 else if (code == SET)
2971 {
2972 /* Don't count SET_DEST if it is a REG; otherwise count things
2973 in SET_DEST because if a register is partially modified, it won't
2974 show up as a potential movable so we don't care how USAGE is set
2975 for it. */
2976 if (GET_CODE (SET_DEST (x)) != REG)
2977 find_single_use_in_loop (insn, SET_DEST (x), usage);
2978 find_single_use_in_loop (insn, SET_SRC (x), usage);
2979 }
2980 else
2981 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2982 {
2983 if (fmt[i] == 'e' && XEXP (x, i) != 0)
2984 find_single_use_in_loop (insn, XEXP (x, i), usage);
2985 else if (fmt[i] == 'E')
2986 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
2987 find_single_use_in_loop (insn, XVECEXP (x, i, j), usage);
2988 }
2989 }
2990 \f
2991 /* Increment N_TIMES_SET at the index of each register
2992 that is modified by an insn between FROM and TO.
2993 If the value of an element of N_TIMES_SET becomes 127 or more,
2994 stop incrementing it, to avoid overflow.
2995
2996 Store in SINGLE_USAGE[I] the single insn in which register I is
2997 used, if it is only used once. Otherwise, it is set to 0 (for no
2998 uses) or const0_rtx for more than one use. This parameter may be zero,
2999 in which case this processing is not done.
3000
3001 Store in *COUNT_PTR the number of actual instruction
3002 in the loop. We use this to decide what is worth moving out. */
3003
3004 /* last_set[n] is nonzero iff reg n has been set in the current basic block.
3005 In that case, it is the insn that last set reg n. */
3006
3007 static void
3008 count_loop_regs_set (from, to, may_not_move, single_usage, count_ptr, nregs)
3009 register rtx from, to;
3010 char *may_not_move;
3011 rtx *single_usage;
3012 int *count_ptr;
3013 int nregs;
3014 {
3015 register rtx *last_set = (rtx *) alloca (nregs * sizeof (rtx));
3016 register rtx insn;
3017 register int count = 0;
3018 register rtx dest;
3019
3020 bzero ((char *) last_set, nregs * sizeof (rtx));
3021 for (insn = from; insn != to; insn = NEXT_INSN (insn))
3022 {
3023 if (GET_RTX_CLASS (GET_CODE (insn)) == 'i')
3024 {
3025 ++count;
3026
3027 /* If requested, record registers that have exactly one use. */
3028 if (single_usage)
3029 {
3030 find_single_use_in_loop (insn, PATTERN (insn), single_usage);
3031
3032 /* Include uses in REG_EQUAL notes. */
3033 if (REG_NOTES (insn))
3034 find_single_use_in_loop (insn, REG_NOTES (insn), single_usage);
3035 }
3036
3037 if (GET_CODE (PATTERN (insn)) == CLOBBER
3038 && GET_CODE (XEXP (PATTERN (insn), 0)) == REG)
3039 /* Don't move a reg that has an explicit clobber.
3040 We might do so sometimes, but it's not worth the pain. */
3041 may_not_move[REGNO (XEXP (PATTERN (insn), 0))] = 1;
3042
3043 if (GET_CODE (PATTERN (insn)) == SET
3044 || GET_CODE (PATTERN (insn)) == CLOBBER)
3045 {
3046 dest = SET_DEST (PATTERN (insn));
3047 while (GET_CODE (dest) == SUBREG
3048 || GET_CODE (dest) == ZERO_EXTRACT
3049 || GET_CODE (dest) == SIGN_EXTRACT
3050 || GET_CODE (dest) == STRICT_LOW_PART)
3051 dest = XEXP (dest, 0);
3052 if (GET_CODE (dest) == REG)
3053 {
3054 register int regno = REGNO (dest);
3055 /* If this is the first setting of this reg
3056 in current basic block, and it was set before,
3057 it must be set in two basic blocks, so it cannot
3058 be moved out of the loop. */
3059 if (n_times_set[regno] > 0 && last_set[regno] == 0)
3060 may_not_move[regno] = 1;
3061 /* If this is not first setting in current basic block,
3062 see if reg was used in between previous one and this.
3063 If so, neither one can be moved. */
3064 if (last_set[regno] != 0
3065 && reg_used_between_p (dest, last_set[regno], insn))
3066 may_not_move[regno] = 1;
3067 if (n_times_set[regno] < 127)
3068 ++n_times_set[regno];
3069 last_set[regno] = insn;
3070 }
3071 }
3072 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
3073 {
3074 register int i;
3075 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
3076 {
3077 register rtx x = XVECEXP (PATTERN (insn), 0, i);
3078 if (GET_CODE (x) == CLOBBER && GET_CODE (XEXP (x, 0)) == REG)
3079 /* Don't move a reg that has an explicit clobber.
3080 It's not worth the pain to try to do it correctly. */
3081 may_not_move[REGNO (XEXP (x, 0))] = 1;
3082
3083 if (GET_CODE (x) == SET || GET_CODE (x) == CLOBBER)
3084 {
3085 dest = SET_DEST (x);
3086 while (GET_CODE (dest) == SUBREG
3087 || GET_CODE (dest) == ZERO_EXTRACT
3088 || GET_CODE (dest) == SIGN_EXTRACT
3089 || GET_CODE (dest) == STRICT_LOW_PART)
3090 dest = XEXP (dest, 0);
3091 if (GET_CODE (dest) == REG)
3092 {
3093 register int regno = REGNO (dest);
3094 if (n_times_set[regno] > 0 && last_set[regno] == 0)
3095 may_not_move[regno] = 1;
3096 if (last_set[regno] != 0
3097 && reg_used_between_p (dest, last_set[regno], insn))
3098 may_not_move[regno] = 1;
3099 if (n_times_set[regno] < 127)
3100 ++n_times_set[regno];
3101 last_set[regno] = insn;
3102 }
3103 }
3104 }
3105 }
3106 }
3107
3108 if (GET_CODE (insn) == CODE_LABEL || GET_CODE (insn) == JUMP_INSN)
3109 bzero ((char *) last_set, nregs * sizeof (rtx));
3110 }
3111 *count_ptr = count;
3112 }
3113 \f
3114 /* Given a loop that is bounded by LOOP_START and LOOP_END
3115 and that is entered at SCAN_START,
3116 return 1 if the register set in SET contained in insn INSN is used by
3117 any insn that precedes INSN in cyclic order starting
3118 from the loop entry point.
3119
3120 We don't want to use INSN_LUID here because if we restrict INSN to those
3121 that have a valid INSN_LUID, it means we cannot move an invariant out
3122 from an inner loop past two loops. */
3123
3124 static int
3125 loop_reg_used_before_p (set, insn, loop_start, scan_start, loop_end)
3126 rtx set, insn, loop_start, scan_start, loop_end;
3127 {
3128 rtx reg = SET_DEST (set);
3129 rtx p;
3130
3131 /* Scan forward checking for register usage. If we hit INSN, we
3132 are done. Otherwise, if we hit LOOP_END, wrap around to LOOP_START. */
3133 for (p = scan_start; p != insn; p = NEXT_INSN (p))
3134 {
3135 if (GET_RTX_CLASS (GET_CODE (p)) == 'i'
3136 && reg_overlap_mentioned_p (reg, PATTERN (p)))
3137 return 1;
3138
3139 if (p == loop_end)
3140 p = loop_start;
3141 }
3142
3143 return 0;
3144 }
3145 \f
3146 /* A "basic induction variable" or biv is a pseudo reg that is set
3147 (within this loop) only by incrementing or decrementing it. */
3148 /* A "general induction variable" or giv is a pseudo reg whose
3149 value is a linear function of a biv. */
3150
3151 /* Bivs are recognized by `basic_induction_var';
3152 Givs by `general_induct_var'. */
3153
3154 /* Indexed by register number, indicates whether or not register is an
3155 induction variable, and if so what type. */
3156
3157 enum iv_mode *reg_iv_type;
3158
3159 /* Indexed by register number, contains pointer to `struct induction'
3160 if register is an induction variable. This holds general info for
3161 all induction variables. */
3162
3163 struct induction **reg_iv_info;
3164
3165 /* Indexed by register number, contains pointer to `struct iv_class'
3166 if register is a basic induction variable. This holds info describing
3167 the class (a related group) of induction variables that the biv belongs
3168 to. */
3169
3170 struct iv_class **reg_biv_class;
3171
3172 /* The head of a list which links together (via the next field)
3173 every iv class for the current loop. */
3174
3175 struct iv_class *loop_iv_list;
3176
3177 /* Communication with routines called via `note_stores'. */
3178
3179 static rtx note_insn;
3180
3181 /* Dummy register to have non-zero DEST_REG for DEST_ADDR type givs. */
3182
3183 static rtx addr_placeholder;
3184
3185 /* ??? Unfinished optimizations, and possible future optimizations,
3186 for the strength reduction code. */
3187
3188 /* ??? There is one more optimization you might be interested in doing: to
3189 allocate pseudo registers for frequently-accessed memory locations.
3190 If the same memory location is referenced each time around, it might
3191 be possible to copy it into a register before and out after.
3192 This is especially useful when the memory location is a variable which
3193 is in a stack slot because somewhere its address is taken. If the
3194 loop doesn't contain a function call and the variable isn't volatile,
3195 it is safe to keep the value in a register for the duration of the
3196 loop. One tricky thing is that the copying of the value back from the
3197 register has to be done on all exits from the loop. You need to check that
3198 all the exits from the loop go to the same place. */
3199
3200 /* ??? The interaction of biv elimination, and recognition of 'constant'
3201 bivs, may cause problems. */
3202
3203 /* ??? Add heuristics so that DEST_ADDR strength reduction does not cause
3204 performance problems.
3205
3206 Perhaps don't eliminate things that can be combined with an addressing
3207 mode. Find all givs that have the same biv, mult_val, and add_val;
3208 then for each giv, check to see if its only use dies in a following
3209 memory address. If so, generate a new memory address and check to see
3210 if it is valid. If it is valid, then store the modified memory address,
3211 otherwise, mark the giv as not done so that it will get its own iv. */
3212
3213 /* ??? Could try to optimize branches when it is known that a biv is always
3214 positive. */
3215
3216 /* ??? When replace a biv in a compare insn, we should replace with closest
3217 giv so that an optimized branch can still be recognized by the combiner,
3218 e.g. the VAX acb insn. */
3219
3220 /* ??? Many of the checks involving uid_luid could be simplified if regscan
3221 was rerun in loop_optimize whenever a register was added or moved.
3222 Also, some of the optimizations could be a little less conservative. */
3223 \f
3224 /* Perform strength reduction and induction variable elimination. */
3225
3226 /* Pseudo registers created during this function will be beyond the last
3227 valid index in several tables including n_times_set and regno_last_uid.
3228 This does not cause a problem here, because the added registers cannot be
3229 givs outside of their loop, and hence will never be reconsidered.
3230 But scan_loop must check regnos to make sure they are in bounds. */
3231
3232 static void
3233 strength_reduce (scan_start, end, loop_top, insn_count,
3234 loop_start, loop_end)
3235 rtx scan_start;
3236 rtx end;
3237 rtx loop_top;
3238 int insn_count;
3239 rtx loop_start;
3240 rtx loop_end;
3241 {
3242 rtx p;
3243 rtx set;
3244 rtx inc_val;
3245 rtx mult_val;
3246 rtx dest_reg;
3247 /* This is 1 if current insn is not executed at least once for every loop
3248 iteration. */
3249 int not_every_iteration = 0;
3250 /* This is 1 if current insn may be executed more than once for every
3251 loop iteration. */
3252 int maybe_multiple = 0;
3253 /* Temporary list pointers for traversing loop_iv_list. */
3254 struct iv_class *bl, **backbl;
3255 /* Ratio of extra register life span we can justify
3256 for saving an instruction. More if loop doesn't call subroutines
3257 since in that case saving an insn makes more difference
3258 and more registers are available. */
3259 /* ??? could set this to last value of threshold in move_movables */
3260 int threshold = (loop_has_call ? 1 : 2) * (3 + n_non_fixed_regs);
3261 /* Map of pseudo-register replacements. */
3262 rtx *reg_map;
3263 int call_seen;
3264 rtx test;
3265 rtx end_insert_before;
3266 int loop_depth = 0;
3267
3268 reg_iv_type = (enum iv_mode *) alloca (max_reg_before_loop
3269 * sizeof (enum iv_mode *));
3270 bzero ((char *) reg_iv_type, max_reg_before_loop * sizeof (enum iv_mode *));
3271 reg_iv_info = (struct induction **)
3272 alloca (max_reg_before_loop * sizeof (struct induction *));
3273 bzero ((char *) reg_iv_info, (max_reg_before_loop
3274 * sizeof (struct induction *)));
3275 reg_biv_class = (struct iv_class **)
3276 alloca (max_reg_before_loop * sizeof (struct iv_class *));
3277 bzero ((char *) reg_biv_class, (max_reg_before_loop
3278 * sizeof (struct iv_class *)));
3279
3280 loop_iv_list = 0;
3281 addr_placeholder = gen_reg_rtx (Pmode);
3282
3283 /* Save insn immediately after the loop_end. Insns inserted after loop_end
3284 must be put before this insn, so that they will appear in the right
3285 order (i.e. loop order).
3286
3287 If loop_end is the end of the current function, then emit a
3288 NOTE_INSN_DELETED after loop_end and set end_insert_before to the
3289 dummy note insn. */
3290 if (NEXT_INSN (loop_end) != 0)
3291 end_insert_before = NEXT_INSN (loop_end);
3292 else
3293 end_insert_before = emit_note_after (NOTE_INSN_DELETED, loop_end);
3294
3295 /* Scan through loop to find all possible bivs. */
3296
3297 p = scan_start;
3298 while (1)
3299 {
3300 p = NEXT_INSN (p);
3301 /* At end of a straight-in loop, we are done.
3302 At end of a loop entered at the bottom, scan the top. */
3303 if (p == scan_start)
3304 break;
3305 if (p == end)
3306 {
3307 if (loop_top != 0)
3308 p = loop_top;
3309 else
3310 break;
3311 if (p == scan_start)
3312 break;
3313 }
3314
3315 if (GET_CODE (p) == INSN
3316 && (set = single_set (p))
3317 && GET_CODE (SET_DEST (set)) == REG)
3318 {
3319 dest_reg = SET_DEST (set);
3320 if (REGNO (dest_reg) < max_reg_before_loop
3321 && REGNO (dest_reg) >= FIRST_PSEUDO_REGISTER
3322 && reg_iv_type[REGNO (dest_reg)] != NOT_BASIC_INDUCT)
3323 {
3324 if (basic_induction_var (SET_SRC (set), GET_MODE (SET_SRC (set)),
3325 dest_reg, p, &inc_val, &mult_val))
3326 {
3327 /* It is a possible basic induction variable.
3328 Create and initialize an induction structure for it. */
3329
3330 struct induction *v
3331 = (struct induction *) alloca (sizeof (struct induction));
3332
3333 record_biv (v, p, dest_reg, inc_val, mult_val,
3334 not_every_iteration, maybe_multiple);
3335 reg_iv_type[REGNO (dest_reg)] = BASIC_INDUCT;
3336 }
3337 else if (REGNO (dest_reg) < max_reg_before_loop)
3338 reg_iv_type[REGNO (dest_reg)] = NOT_BASIC_INDUCT;
3339 }
3340 }
3341
3342 /* Past CODE_LABEL, we get to insns that may be executed multiple
3343 times. The only way we can be sure that they can't is if every
3344 every jump insn between here and the end of the loop either
3345 returns, exits the loop, is a forward jump, or is a jump
3346 to the loop start. */
3347
3348 if (GET_CODE (p) == CODE_LABEL)
3349 {
3350 rtx insn = p;
3351
3352 maybe_multiple = 0;
3353
3354 while (1)
3355 {
3356 insn = NEXT_INSN (insn);
3357 if (insn == scan_start)
3358 break;
3359 if (insn == end)
3360 {
3361 if (loop_top != 0)
3362 insn = loop_top;
3363 else
3364 break;
3365 if (insn == scan_start)
3366 break;
3367 }
3368
3369 if (GET_CODE (insn) == JUMP_INSN
3370 && GET_CODE (PATTERN (insn)) != RETURN
3371 && (! condjump_p (insn)
3372 || (JUMP_LABEL (insn) != 0
3373 && JUMP_LABEL (insn) != scan_start
3374 && (INSN_UID (JUMP_LABEL (insn)) >= max_uid_for_loop
3375 || INSN_UID (insn) >= max_uid_for_loop
3376 || (INSN_LUID (JUMP_LABEL (insn))
3377 < INSN_LUID (insn))))))
3378 {
3379 maybe_multiple = 1;
3380 break;
3381 }
3382 }
3383 }
3384
3385 /* Past a jump, we get to insns for which we can't count
3386 on whether they will be executed during each iteration. */
3387 /* This code appears twice in strength_reduce. There is also similar
3388 code in scan_loop. */
3389 if (GET_CODE (p) == JUMP_INSN
3390 /* If we enter the loop in the middle, and scan around to the
3391 beginning, don't set not_every_iteration for that.
3392 This can be any kind of jump, since we want to know if insns
3393 will be executed if the loop is executed. */
3394 && ! (JUMP_LABEL (p) == loop_top
3395 && ((NEXT_INSN (NEXT_INSN (p)) == loop_end && simplejump_p (p))
3396 || (NEXT_INSN (p) == loop_end && condjump_p (p)))))
3397 {
3398 rtx label = 0;
3399
3400 /* If this is a jump outside the loop, then it also doesn't
3401 matter. Check to see if the target of this branch is on the
3402 loop_number_exits_labels list. */
3403
3404 for (label = loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]];
3405 label;
3406 label = LABEL_NEXTREF (label))
3407 if (XEXP (label, 0) == JUMP_LABEL (p))
3408 break;
3409
3410 if (! label)
3411 not_every_iteration = 1;
3412 }
3413
3414 else if (GET_CODE (p) == NOTE)
3415 {
3416 /* At the virtual top of a converted loop, insns are again known to
3417 be executed each iteration: logically, the loop begins here
3418 even though the exit code has been duplicated. */
3419 if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP && loop_depth == 0)
3420 not_every_iteration = 0;
3421 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
3422 loop_depth++;
3423 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END)
3424 loop_depth--;
3425 }
3426
3427 /* Unlike in the code motion pass where MAYBE_NEVER indicates that
3428 an insn may never be executed, NOT_EVERY_ITERATION indicates whether
3429 or not an insn is known to be executed each iteration of the
3430 loop, whether or not any iterations are known to occur.
3431
3432 Therefore, if we have just passed a label and have no more labels
3433 between here and the test insn of the loop, we know these insns
3434 will be executed each iteration. */
3435
3436 if (not_every_iteration && GET_CODE (p) == CODE_LABEL
3437 && no_labels_between_p (p, loop_end))
3438 not_every_iteration = 0;
3439 }
3440
3441 /* Scan loop_iv_list to remove all regs that proved not to be bivs.
3442 Make a sanity check against n_times_set. */
3443 for (backbl = &loop_iv_list, bl = *backbl; bl; bl = bl->next)
3444 {
3445 if (reg_iv_type[bl->regno] != BASIC_INDUCT
3446 /* Above happens if register modified by subreg, etc. */
3447 /* Make sure it is not recognized as a basic induction var: */
3448 || n_times_set[bl->regno] != bl->biv_count
3449 /* If never incremented, it is invariant that we decided not to
3450 move. So leave it alone. */
3451 || ! bl->incremented)
3452 {
3453 if (loop_dump_stream)
3454 fprintf (loop_dump_stream, "Reg %d: biv discarded, %s\n",
3455 bl->regno,
3456 (reg_iv_type[bl->regno] != BASIC_INDUCT
3457 ? "not induction variable"
3458 : (! bl->incremented ? "never incremented"
3459 : "count error")));
3460
3461 reg_iv_type[bl->regno] = NOT_BASIC_INDUCT;
3462 *backbl = bl->next;
3463 }
3464 else
3465 {
3466 backbl = &bl->next;
3467
3468 if (loop_dump_stream)
3469 fprintf (loop_dump_stream, "Reg %d: biv verified\n", bl->regno);
3470 }
3471 }
3472
3473 /* Exit if there are no bivs. */
3474 if (! loop_iv_list)
3475 {
3476 /* Can still unroll the loop anyways, but indicate that there is no
3477 strength reduction info available. */
3478 if (flag_unroll_loops)
3479 unroll_loop (loop_end, insn_count, loop_start, end_insert_before, 0);
3480
3481 return;
3482 }
3483
3484 /* Find initial value for each biv by searching backwards from loop_start,
3485 halting at first label. Also record any test condition. */
3486
3487 call_seen = 0;
3488 for (p = loop_start; p && GET_CODE (p) != CODE_LABEL; p = PREV_INSN (p))
3489 {
3490 note_insn = p;
3491
3492 if (GET_CODE (p) == CALL_INSN)
3493 call_seen = 1;
3494
3495 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
3496 || GET_CODE (p) == CALL_INSN)
3497 note_stores (PATTERN (p), record_initial);
3498
3499 /* Record any test of a biv that branches around the loop if no store
3500 between it and the start of loop. We only care about tests with
3501 constants and registers and only certain of those. */
3502 if (GET_CODE (p) == JUMP_INSN
3503 && JUMP_LABEL (p) != 0
3504 && next_real_insn (JUMP_LABEL (p)) == next_real_insn (loop_end)
3505 && (test = get_condition_for_loop (p)) != 0
3506 && GET_CODE (XEXP (test, 0)) == REG
3507 && REGNO (XEXP (test, 0)) < max_reg_before_loop
3508 && (bl = reg_biv_class[REGNO (XEXP (test, 0))]) != 0
3509 && valid_initial_value_p (XEXP (test, 1), p, call_seen, loop_start)
3510 && bl->init_insn == 0)
3511 {
3512 /* If an NE test, we have an initial value! */
3513 if (GET_CODE (test) == NE)
3514 {
3515 bl->init_insn = p;
3516 bl->init_set = gen_rtx (SET, VOIDmode,
3517 XEXP (test, 0), XEXP (test, 1));
3518 }
3519 else
3520 bl->initial_test = test;
3521 }
3522 }
3523
3524 /* Look at the each biv and see if we can say anything better about its
3525 initial value from any initializing insns set up above. (This is done
3526 in two passes to avoid missing SETs in a PARALLEL.) */
3527 for (bl = loop_iv_list; bl; bl = bl->next)
3528 {
3529 rtx src;
3530
3531 if (! bl->init_insn)
3532 continue;
3533
3534 src = SET_SRC (bl->init_set);
3535
3536 if (loop_dump_stream)
3537 fprintf (loop_dump_stream,
3538 "Biv %d initialized at insn %d: initial value ",
3539 bl->regno, INSN_UID (bl->init_insn));
3540
3541 if ((GET_MODE (src) == GET_MODE (regno_reg_rtx[bl->regno])
3542 || GET_MODE (src) == VOIDmode)
3543 && valid_initial_value_p (src, bl->init_insn, call_seen, loop_start))
3544 {
3545 bl->initial_value = src;
3546
3547 if (loop_dump_stream)
3548 {
3549 if (GET_CODE (src) == CONST_INT)
3550 fprintf (loop_dump_stream, "%d\n", INTVAL (src));
3551 else
3552 {
3553 print_rtl (loop_dump_stream, src);
3554 fprintf (loop_dump_stream, "\n");
3555 }
3556 }
3557 }
3558 else
3559 {
3560 /* Biv initial value is not simple move,
3561 so let it keep initial value of "itself". */
3562
3563 if (loop_dump_stream)
3564 fprintf (loop_dump_stream, "is complex\n");
3565 }
3566 }
3567
3568 /* Search the loop for general induction variables. */
3569
3570 /* A register is a giv if: it is only set once, it is a function of a
3571 biv and a constant (or invariant), and it is not a biv. */
3572
3573 not_every_iteration = 0;
3574 loop_depth = 0;
3575 p = scan_start;
3576 while (1)
3577 {
3578 p = NEXT_INSN (p);
3579 /* At end of a straight-in loop, we are done.
3580 At end of a loop entered at the bottom, scan the top. */
3581 if (p == scan_start)
3582 break;
3583 if (p == end)
3584 {
3585 if (loop_top != 0)
3586 p = loop_top;
3587 else
3588 break;
3589 if (p == scan_start)
3590 break;
3591 }
3592
3593 /* Look for a general induction variable in a register. */
3594 if (GET_CODE (p) == INSN
3595 && (set = single_set (p))
3596 && GET_CODE (SET_DEST (set)) == REG
3597 && ! may_not_optimize[REGNO (SET_DEST (set))])
3598 {
3599 rtx src_reg;
3600 rtx add_val;
3601 rtx mult_val;
3602 int benefit;
3603 rtx regnote = 0;
3604
3605 dest_reg = SET_DEST (set);
3606 if (REGNO (dest_reg) < FIRST_PSEUDO_REGISTER)
3607 continue;
3608
3609 if (/* SET_SRC is a giv. */
3610 ((benefit = general_induction_var (SET_SRC (set),
3611 &src_reg, &add_val,
3612 &mult_val))
3613 /* Equivalent expression is a giv. */
3614 || ((regnote = find_reg_note (p, REG_EQUAL, NULL_RTX))
3615 && (benefit = general_induction_var (XEXP (regnote, 0),
3616 &src_reg,
3617 &add_val, &mult_val))))
3618 /* Don't try to handle any regs made by loop optimization.
3619 We have nothing on them in regno_first_uid, etc. */
3620 && REGNO (dest_reg) < max_reg_before_loop
3621 /* Don't recognize a BASIC_INDUCT_VAR here. */
3622 && dest_reg != src_reg
3623 /* This must be the only place where the register is set. */
3624 && (n_times_set[REGNO (dest_reg)] == 1
3625 /* or all sets must be consecutive and make a giv. */
3626 || (benefit = consec_sets_giv (benefit, p,
3627 src_reg, dest_reg,
3628 &add_val, &mult_val))))
3629 {
3630 int count;
3631 struct induction *v
3632 = (struct induction *) alloca (sizeof (struct induction));
3633 rtx temp;
3634
3635 /* If this is a library call, increase benefit. */
3636 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
3637 benefit += libcall_benefit (p);
3638
3639 /* Skip the consecutive insns, if there are any. */
3640 for (count = n_times_set[REGNO (dest_reg)] - 1;
3641 count > 0; count--)
3642 {
3643 /* If first insn of libcall sequence, skip to end.
3644 Do this at start of loop, since INSN is guaranteed to
3645 be an insn here. */
3646 if (GET_CODE (p) != NOTE
3647 && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
3648 p = XEXP (temp, 0);
3649
3650 do p = NEXT_INSN (p);
3651 while (GET_CODE (p) == NOTE);
3652 }
3653
3654 record_giv (v, p, src_reg, dest_reg, mult_val, add_val, benefit,
3655 DEST_REG, not_every_iteration, NULL_PTR, loop_start,
3656 loop_end);
3657
3658 }
3659 }
3660
3661 #ifndef DONT_REDUCE_ADDR
3662 /* Look for givs which are memory addresses. */
3663 /* This resulted in worse code on a VAX 8600. I wonder if it
3664 still does. */
3665 if (GET_CODE (p) == INSN)
3666 find_mem_givs (PATTERN (p), p, not_every_iteration, loop_start,
3667 loop_end);
3668 #endif
3669
3670 /* Update the status of whether giv can derive other givs. This can
3671 change when we pass a label or an insn that updates a biv. */
3672 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
3673 || GET_CODE (p) == CODE_LABEL)
3674 update_giv_derive (p);
3675
3676 /* Past a jump, we get to insns for which we can't count
3677 on whether they will be executed during each iteration. */
3678 /* This code appears twice in strength_reduce. There is also similar
3679 code in scan_loop. */
3680 if (GET_CODE (p) == JUMP_INSN
3681 /* If we enter the loop in the middle, and scan around to the
3682 beginning, don't set not_every_iteration for that.
3683 This can be any kind of jump, since we want to know if insns
3684 will be executed if the loop is executed. */
3685 && ! (JUMP_LABEL (p) == loop_top
3686 && ((NEXT_INSN (NEXT_INSN (p)) == loop_end && simplejump_p (p))
3687 || (NEXT_INSN (p) == loop_end && condjump_p (p)))))
3688 {
3689 rtx label = 0;
3690
3691 /* If this is a jump outside the loop, then it also doesn't
3692 matter. Check to see if the target of this branch is on the
3693 loop_number_exits_labels list. */
3694
3695 for (label = loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]];
3696 label;
3697 label = LABEL_NEXTREF (label))
3698 if (XEXP (label, 0) == JUMP_LABEL (p))
3699 break;
3700
3701 if (! label)
3702 not_every_iteration = 1;
3703 }
3704
3705 else if (GET_CODE (p) == NOTE)
3706 {
3707 /* At the virtual top of a converted loop, insns are again known to
3708 be executed each iteration: logically, the loop begins here
3709 even though the exit code has been duplicated. */
3710 if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP && loop_depth == 0)
3711 not_every_iteration = 0;
3712 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
3713 loop_depth++;
3714 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END)
3715 loop_depth--;
3716 }
3717
3718 /* Unlike in the code motion pass where MAYBE_NEVER indicates that
3719 an insn may never be executed, NOT_EVERY_ITERATION indicates whether
3720 or not an insn is known to be executed each iteration of the
3721 loop, whether or not any iterations are known to occur.
3722
3723 Therefore, if we have just passed a label and have no more labels
3724 between here and the test insn of the loop, we know these insns
3725 will be executed each iteration. */
3726
3727 if (not_every_iteration && GET_CODE (p) == CODE_LABEL
3728 && no_labels_between_p (p, loop_end))
3729 not_every_iteration = 0;
3730 }
3731
3732 /* Try to calculate and save the number of loop iterations. This is
3733 set to zero if the actual number can not be calculated. This must
3734 be called after all giv's have been identified, since otherwise it may
3735 fail if the iteration variable is a giv. */
3736
3737 loop_n_iterations = loop_iterations (loop_start, loop_end);
3738
3739 /* Now for each giv for which we still don't know whether or not it is
3740 replaceable, check to see if it is replaceable because its final value
3741 can be calculated. This must be done after loop_iterations is called,
3742 so that final_giv_value will work correctly. */
3743
3744 for (bl = loop_iv_list; bl; bl = bl->next)
3745 {
3746 struct induction *v;
3747
3748 for (v = bl->giv; v; v = v->next_iv)
3749 if (! v->replaceable && ! v->not_replaceable)
3750 check_final_value (v, loop_start, loop_end);
3751 }
3752
3753 /* Try to prove that the loop counter variable (if any) is always
3754 nonnegative; if so, record that fact with a REG_NONNEG note
3755 so that "decrement and branch until zero" insn can be used. */
3756 check_dbra_loop (loop_end, insn_count, loop_start);
3757
3758 /* Create reg_map to hold substitutions for replaceable giv regs. */
3759 reg_map = (rtx *) alloca (max_reg_before_loop * sizeof (rtx));
3760 bzero ((char *) reg_map, max_reg_before_loop * sizeof (rtx));
3761
3762 /* Examine each iv class for feasibility of strength reduction/induction
3763 variable elimination. */
3764
3765 for (bl = loop_iv_list; bl; bl = bl->next)
3766 {
3767 struct induction *v;
3768 int benefit;
3769 int all_reduced;
3770 rtx final_value = 0;
3771
3772 /* Test whether it will be possible to eliminate this biv
3773 provided all givs are reduced. This is possible if either
3774 the reg is not used outside the loop, or we can compute
3775 what its final value will be.
3776
3777 For architectures with a decrement_and_branch_until_zero insn,
3778 don't do this if we put a REG_NONNEG note on the endtest for
3779 this biv. */
3780
3781 /* Compare against bl->init_insn rather than loop_start.
3782 We aren't concerned with any uses of the biv between
3783 init_insn and loop_start since these won't be affected
3784 by the value of the biv elsewhere in the function, so
3785 long as init_insn doesn't use the biv itself.
3786 March 14, 1989 -- self@bayes.arc.nasa.gov */
3787
3788 if ((uid_luid[REGNO_LAST_UID (bl->regno)] < INSN_LUID (loop_end)
3789 && bl->init_insn
3790 && INSN_UID (bl->init_insn) < max_uid_for_loop
3791 && uid_luid[REGNO_FIRST_UID (bl->regno)] >= INSN_LUID (bl->init_insn)
3792 #ifdef HAVE_decrement_and_branch_until_zero
3793 && ! bl->nonneg
3794 #endif
3795 && ! reg_mentioned_p (bl->biv->dest_reg, SET_SRC (bl->init_set)))
3796 || ((final_value = final_biv_value (bl, loop_start, loop_end))
3797 #ifdef HAVE_decrement_and_branch_until_zero
3798 && ! bl->nonneg
3799 #endif
3800 ))
3801 bl->eliminable = maybe_eliminate_biv (bl, loop_start, end, 0,
3802 threshold, insn_count);
3803 else
3804 {
3805 if (loop_dump_stream)
3806 {
3807 fprintf (loop_dump_stream,
3808 "Cannot eliminate biv %d.\n",
3809 bl->regno);
3810 fprintf (loop_dump_stream,
3811 "First use: insn %d, last use: insn %d.\n",
3812 REGNO_FIRST_UID (bl->regno),
3813 REGNO_LAST_UID (bl->regno));
3814 }
3815 }
3816
3817 /* Combine all giv's for this iv_class. */
3818 combine_givs (bl);
3819
3820 /* This will be true at the end, if all givs which depend on this
3821 biv have been strength reduced.
3822 We can't (currently) eliminate the biv unless this is so. */
3823 all_reduced = 1;
3824
3825 /* Check each giv in this class to see if we will benefit by reducing
3826 it. Skip giv's combined with others. */
3827 for (v = bl->giv; v; v = v->next_iv)
3828 {
3829 struct induction *tv;
3830
3831 if (v->ignore || v->same)
3832 continue;
3833
3834 benefit = v->benefit;
3835
3836 /* Reduce benefit if not replaceable, since we will insert
3837 a move-insn to replace the insn that calculates this giv.
3838 Don't do this unless the giv is a user variable, since it
3839 will often be marked non-replaceable because of the duplication
3840 of the exit code outside the loop. In such a case, the copies
3841 we insert are dead and will be deleted. So they don't have
3842 a cost. Similar situations exist. */
3843 /* ??? The new final_[bg]iv_value code does a much better job
3844 of finding replaceable giv's, and hence this code may no longer
3845 be necessary. */
3846 if (! v->replaceable && ! bl->eliminable
3847 && REG_USERVAR_P (v->dest_reg))
3848 benefit -= copy_cost;
3849
3850 /* Decrease the benefit to count the add-insns that we will
3851 insert to increment the reduced reg for the giv. */
3852 benefit -= add_cost * bl->biv_count;
3853
3854 /* Decide whether to strength-reduce this giv or to leave the code
3855 unchanged (recompute it from the biv each time it is used).
3856 This decision can be made independently for each giv. */
3857
3858 #ifdef AUTO_INC_DEC
3859 /* Attempt to guess whether autoincrement will handle some of the
3860 new add insns; if so, increase BENEFIT (undo the subtraction of
3861 add_cost that was done above). */
3862 if (v->giv_type == DEST_ADDR
3863 && GET_CODE (v->mult_val) == CONST_INT)
3864 {
3865 #if defined (HAVE_POST_INCREMENT) || defined (HAVE_PRE_INCREMENT)
3866 if (INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
3867 benefit += add_cost * bl->biv_count;
3868 #endif
3869 #if defined (HAVE_POST_DECREMENT) || defined (HAVE_PRE_DECREMENT)
3870 if (-INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
3871 benefit += add_cost * bl->biv_count;
3872 #endif
3873 }
3874 #endif
3875
3876 /* If an insn is not to be strength reduced, then set its ignore
3877 flag, and clear all_reduced. */
3878
3879 /* A giv that depends on a reversed biv must be reduced if it is
3880 used after the loop exit, otherwise, it would have the wrong
3881 value after the loop exit. To make it simple, just reduce all
3882 of such giv's whether or not we know they are used after the loop
3883 exit. */
3884
3885 if (v->lifetime * threshold * benefit < insn_count
3886 && ! bl->reversed)
3887 {
3888 if (loop_dump_stream)
3889 fprintf (loop_dump_stream,
3890 "giv of insn %d not worth while, %d vs %d.\n",
3891 INSN_UID (v->insn),
3892 v->lifetime * threshold * benefit, insn_count);
3893 v->ignore = 1;
3894 all_reduced = 0;
3895 }
3896 else
3897 {
3898 /* Check that we can increment the reduced giv without a
3899 multiply insn. If not, reject it. */
3900
3901 for (tv = bl->biv; tv; tv = tv->next_iv)
3902 if (tv->mult_val == const1_rtx
3903 && ! product_cheap_p (tv->add_val, v->mult_val))
3904 {
3905 if (loop_dump_stream)
3906 fprintf (loop_dump_stream,
3907 "giv of insn %d: would need a multiply.\n",
3908 INSN_UID (v->insn));
3909 v->ignore = 1;
3910 all_reduced = 0;
3911 break;
3912 }
3913 }
3914 }
3915
3916 /* Reduce each giv that we decided to reduce. */
3917
3918 for (v = bl->giv; v; v = v->next_iv)
3919 {
3920 struct induction *tv;
3921 if (! v->ignore && v->same == 0)
3922 {
3923 int auto_inc_opt = 0;
3924
3925 v->new_reg = gen_reg_rtx (v->mode);
3926
3927 #ifdef AUTO_INC_DEC
3928 /* If the target has auto-increment addressing modes, and
3929 this is an address giv, then try to put the increment
3930 immediately after its use, so that flow can create an
3931 auto-increment addressing mode. */
3932 if (v->giv_type == DEST_ADDR && bl->biv_count == 1
3933 && bl->biv->always_executed && ! bl->biv->maybe_multiple
3934 /* We don't handle reversed biv's because bl->biv->insn
3935 does not have a valid INSN_LUID. */
3936 && ! bl->reversed
3937 && v->always_executed && ! v->maybe_multiple)
3938 {
3939 /* If other giv's have been combined with this one, then
3940 this will work only if all uses of the other giv's occur
3941 before this giv's insn. This is difficult to check.
3942
3943 We simplify this by looking for the common case where
3944 there is one DEST_REG giv, and this giv's insn is the
3945 last use of the dest_reg of that DEST_REG giv. If the
3946 the increment occurs after the address giv, then we can
3947 perform the optimization. (Otherwise, the increment
3948 would have to go before other_giv, and we would not be
3949 able to combine it with the address giv to get an
3950 auto-inc address.) */
3951 if (v->combined_with)
3952 {
3953 struct induction *other_giv = 0;
3954
3955 for (tv = bl->giv; tv; tv = tv->next_iv)
3956 if (tv->same == v)
3957 {
3958 if (other_giv)
3959 break;
3960 else
3961 other_giv = tv;
3962 }
3963 if (! tv && other_giv
3964 && (REGNO_LAST_UID (REGNO (other_giv->dest_reg))
3965 == INSN_UID (v->insn))
3966 && INSN_LUID (v->insn) < INSN_LUID (bl->biv->insn))
3967 auto_inc_opt = 1;
3968 }
3969 /* Check for case where increment is before the the address
3970 giv. */
3971 else if (INSN_LUID (v->insn) > INSN_LUID (bl->biv->insn))
3972 auto_inc_opt = -1;
3973 else
3974 auto_inc_opt = 1;
3975
3976 #ifdef HAVE_cc0
3977 {
3978 rtx prev;
3979
3980 /* We can't put an insn immediately after one setting
3981 cc0, or immediately before one using cc0. */
3982 if ((auto_inc_opt == 1 && sets_cc0_p (PATTERN (v->insn)))
3983 || (auto_inc_opt == -1
3984 && (prev = prev_nonnote_insn (v->insn)) != 0
3985 && GET_RTX_CLASS (GET_CODE (prev)) == 'i'
3986 && sets_cc0_p (PATTERN (prev))))
3987 auto_inc_opt = 0;
3988 }
3989 #endif
3990
3991 if (auto_inc_opt)
3992 v->auto_inc_opt = 1;
3993 }
3994 #endif
3995
3996 /* For each place where the biv is incremented, add an insn
3997 to increment the new, reduced reg for the giv. */
3998 for (tv = bl->biv; tv; tv = tv->next_iv)
3999 {
4000 rtx insert_before;
4001
4002 if (! auto_inc_opt)
4003 insert_before = tv->insn;
4004 else if (auto_inc_opt == 1)
4005 insert_before = NEXT_INSN (v->insn);
4006 else
4007 insert_before = v->insn;
4008
4009 if (tv->mult_val == const1_rtx)
4010 emit_iv_add_mult (tv->add_val, v->mult_val,
4011 v->new_reg, v->new_reg, insert_before);
4012 else /* tv->mult_val == const0_rtx */
4013 /* A multiply is acceptable here
4014 since this is presumed to be seldom executed. */
4015 emit_iv_add_mult (tv->add_val, v->mult_val,
4016 v->add_val, v->new_reg, insert_before);
4017 }
4018
4019 /* Add code at loop start to initialize giv's reduced reg. */
4020
4021 emit_iv_add_mult (bl->initial_value, v->mult_val,
4022 v->add_val, v->new_reg, loop_start);
4023 }
4024 }
4025
4026 /* Rescan all givs. If a giv is the same as a giv not reduced, mark it
4027 as not reduced.
4028
4029 For each giv register that can be reduced now: if replaceable,
4030 substitute reduced reg wherever the old giv occurs;
4031 else add new move insn "giv_reg = reduced_reg".
4032
4033 Also check for givs whose first use is their definition and whose
4034 last use is the definition of another giv. If so, it is likely
4035 dead and should not be used to eliminate a biv. */
4036 for (v = bl->giv; v; v = v->next_iv)
4037 {
4038 if (v->same && v->same->ignore)
4039 v->ignore = 1;
4040
4041 if (v->ignore)
4042 continue;
4043
4044 if (v->giv_type == DEST_REG
4045 && REGNO_FIRST_UID (REGNO (v->dest_reg)) == INSN_UID (v->insn))
4046 {
4047 struct induction *v1;
4048
4049 for (v1 = bl->giv; v1; v1 = v1->next_iv)
4050 if (REGNO_LAST_UID (REGNO (v->dest_reg)) == INSN_UID (v1->insn))
4051 v->maybe_dead = 1;
4052 }
4053
4054 /* Update expression if this was combined, in case other giv was
4055 replaced. */
4056 if (v->same)
4057 v->new_reg = replace_rtx (v->new_reg,
4058 v->same->dest_reg, v->same->new_reg);
4059
4060 if (v->giv_type == DEST_ADDR)
4061 /* Store reduced reg as the address in the memref where we found
4062 this giv. */
4063 validate_change (v->insn, v->location, v->new_reg, 0);
4064 else if (v->replaceable)
4065 {
4066 reg_map[REGNO (v->dest_reg)] = v->new_reg;
4067
4068 #if 0
4069 /* I can no longer duplicate the original problem. Perhaps
4070 this is unnecessary now? */
4071
4072 /* Replaceable; it isn't strictly necessary to delete the old
4073 insn and emit a new one, because v->dest_reg is now dead.
4074
4075 However, especially when unrolling loops, the special
4076 handling for (set REG0 REG1) in the second cse pass may
4077 make v->dest_reg live again. To avoid this problem, emit
4078 an insn to set the original giv reg from the reduced giv.
4079 We can not delete the original insn, since it may be part
4080 of a LIBCALL, and the code in flow that eliminates dead
4081 libcalls will fail if it is deleted. */
4082 emit_insn_after (gen_move_insn (v->dest_reg, v->new_reg),
4083 v->insn);
4084 #endif
4085 }
4086 else
4087 {
4088 /* Not replaceable; emit an insn to set the original giv reg from
4089 the reduced giv, same as above. */
4090 emit_insn_after (gen_move_insn (v->dest_reg, v->new_reg),
4091 v->insn);
4092 }
4093
4094 /* When a loop is reversed, givs which depend on the reversed
4095 biv, and which are live outside the loop, must be set to their
4096 correct final value. This insn is only needed if the giv is
4097 not replaceable. The correct final value is the same as the
4098 value that the giv starts the reversed loop with. */
4099 if (bl->reversed && ! v->replaceable)
4100 emit_iv_add_mult (bl->initial_value, v->mult_val,
4101 v->add_val, v->dest_reg, end_insert_before);
4102 else if (v->final_value)
4103 {
4104 rtx insert_before;
4105
4106 /* If the loop has multiple exits, emit the insn before the
4107 loop to ensure that it will always be executed no matter
4108 how the loop exits. Otherwise, emit the insn after the loop,
4109 since this is slightly more efficient. */
4110 if (loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]])
4111 insert_before = loop_start;
4112 else
4113 insert_before = end_insert_before;
4114 emit_insn_before (gen_move_insn (v->dest_reg, v->final_value),
4115 insert_before);
4116
4117 #if 0
4118 /* If the insn to set the final value of the giv was emitted
4119 before the loop, then we must delete the insn inside the loop
4120 that sets it. If this is a LIBCALL, then we must delete
4121 every insn in the libcall. Note, however, that
4122 final_giv_value will only succeed when there are multiple
4123 exits if the giv is dead at each exit, hence it does not
4124 matter that the original insn remains because it is dead
4125 anyways. */
4126 /* Delete the insn inside the loop that sets the giv since
4127 the giv is now set before (or after) the loop. */
4128 delete_insn (v->insn);
4129 #endif
4130 }
4131
4132 if (loop_dump_stream)
4133 {
4134 fprintf (loop_dump_stream, "giv at %d reduced to ",
4135 INSN_UID (v->insn));
4136 print_rtl (loop_dump_stream, v->new_reg);
4137 fprintf (loop_dump_stream, "\n");
4138 }
4139 }
4140
4141 /* All the givs based on the biv bl have been reduced if they
4142 merit it. */
4143
4144 /* For each giv not marked as maybe dead that has been combined with a
4145 second giv, clear any "maybe dead" mark on that second giv.
4146 v->new_reg will either be or refer to the register of the giv it
4147 combined with.
4148
4149 Doing this clearing avoids problems in biv elimination where a
4150 giv's new_reg is a complex value that can't be put in the insn but
4151 the giv combined with (with a reg as new_reg) is marked maybe_dead.
4152 Since the register will be used in either case, we'd prefer it be
4153 used from the simpler giv. */
4154
4155 for (v = bl->giv; v; v = v->next_iv)
4156 if (! v->maybe_dead && v->same)
4157 v->same->maybe_dead = 0;
4158
4159 /* Try to eliminate the biv, if it is a candidate.
4160 This won't work if ! all_reduced,
4161 since the givs we planned to use might not have been reduced.
4162
4163 We have to be careful that we didn't initially think we could eliminate
4164 this biv because of a giv that we now think may be dead and shouldn't
4165 be used as a biv replacement.
4166
4167 Also, there is the possibility that we may have a giv that looks
4168 like it can be used to eliminate a biv, but the resulting insn
4169 isn't valid. This can happen, for example, on the 88k, where a
4170 JUMP_INSN can compare a register only with zero. Attempts to
4171 replace it with a compare with a constant will fail.
4172
4173 Note that in cases where this call fails, we may have replaced some
4174 of the occurrences of the biv with a giv, but no harm was done in
4175 doing so in the rare cases where it can occur. */
4176
4177 if (all_reduced == 1 && bl->eliminable
4178 && maybe_eliminate_biv (bl, loop_start, end, 1,
4179 threshold, insn_count))
4180
4181 {
4182 /* ?? If we created a new test to bypass the loop entirely,
4183 or otherwise drop straight in, based on this test, then
4184 we might want to rewrite it also. This way some later
4185 pass has more hope of removing the initialization of this
4186 biv entirely. */
4187
4188 /* If final_value != 0, then the biv may be used after loop end
4189 and we must emit an insn to set it just in case.
4190
4191 Reversed bivs already have an insn after the loop setting their
4192 value, so we don't need another one. We can't calculate the
4193 proper final value for such a biv here anyways. */
4194 if (final_value != 0 && ! bl->reversed)
4195 {
4196 rtx insert_before;
4197
4198 /* If the loop has multiple exits, emit the insn before the
4199 loop to ensure that it will always be executed no matter
4200 how the loop exits. Otherwise, emit the insn after the
4201 loop, since this is slightly more efficient. */
4202 if (loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]])
4203 insert_before = loop_start;
4204 else
4205 insert_before = end_insert_before;
4206
4207 emit_insn_before (gen_move_insn (bl->biv->dest_reg, final_value),
4208 end_insert_before);
4209 }
4210
4211 #if 0
4212 /* Delete all of the instructions inside the loop which set
4213 the biv, as they are all dead. If is safe to delete them,
4214 because an insn setting a biv will never be part of a libcall. */
4215 /* However, deleting them will invalidate the regno_last_uid info,
4216 so keeping them around is more convenient. Final_biv_value
4217 will only succeed when there are multiple exits if the biv
4218 is dead at each exit, hence it does not matter that the original
4219 insn remains, because it is dead anyways. */
4220 for (v = bl->biv; v; v = v->next_iv)
4221 delete_insn (v->insn);
4222 #endif
4223
4224 if (loop_dump_stream)
4225 fprintf (loop_dump_stream, "Reg %d: biv eliminated\n",
4226 bl->regno);
4227 }
4228 }
4229
4230 /* Go through all the instructions in the loop, making all the
4231 register substitutions scheduled in REG_MAP. */
4232
4233 for (p = loop_start; p != end; p = NEXT_INSN (p))
4234 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
4235 || GET_CODE (p) == CALL_INSN)
4236 {
4237 replace_regs (PATTERN (p), reg_map, max_reg_before_loop, 0);
4238 replace_regs (REG_NOTES (p), reg_map, max_reg_before_loop, 0);
4239 INSN_CODE (p) = -1;
4240 }
4241
4242 /* Unroll loops from within strength reduction so that we can use the
4243 induction variable information that strength_reduce has already
4244 collected. */
4245
4246 if (flag_unroll_loops)
4247 unroll_loop (loop_end, insn_count, loop_start, end_insert_before, 1);
4248
4249 if (loop_dump_stream)
4250 fprintf (loop_dump_stream, "\n");
4251 }
4252 \f
4253 /* Return 1 if X is a valid source for an initial value (or as value being
4254 compared against in an initial test).
4255
4256 X must be either a register or constant and must not be clobbered between
4257 the current insn and the start of the loop.
4258
4259 INSN is the insn containing X. */
4260
4261 static int
4262 valid_initial_value_p (x, insn, call_seen, loop_start)
4263 rtx x;
4264 rtx insn;
4265 int call_seen;
4266 rtx loop_start;
4267 {
4268 if (CONSTANT_P (x))
4269 return 1;
4270
4271 /* Only consider pseudos we know about initialized in insns whose luids
4272 we know. */
4273 if (GET_CODE (x) != REG
4274 || REGNO (x) >= max_reg_before_loop)
4275 return 0;
4276
4277 /* Don't use call-clobbered registers across a call which clobbers it. On
4278 some machines, don't use any hard registers at all. */
4279 if (REGNO (x) < FIRST_PSEUDO_REGISTER
4280 && (
4281 #ifdef SMALL_REGISTER_CLASSES
4282 SMALL_REGISTER_CLASSES
4283 #else
4284 0
4285 #endif
4286 || (call_used_regs[REGNO (x)] && call_seen))
4287 )
4288 return 0;
4289
4290 /* Don't use registers that have been clobbered before the start of the
4291 loop. */
4292 if (reg_set_between_p (x, insn, loop_start))
4293 return 0;
4294
4295 return 1;
4296 }
4297 \f
4298 /* Scan X for memory refs and check each memory address
4299 as a possible giv. INSN is the insn whose pattern X comes from.
4300 NOT_EVERY_ITERATION is 1 if the insn might not be executed during
4301 every loop iteration. */
4302
4303 static void
4304 find_mem_givs (x, insn, not_every_iteration, loop_start, loop_end)
4305 rtx x;
4306 rtx insn;
4307 int not_every_iteration;
4308 rtx loop_start, loop_end;
4309 {
4310 register int i, j;
4311 register enum rtx_code code;
4312 register char *fmt;
4313
4314 if (x == 0)
4315 return;
4316
4317 code = GET_CODE (x);
4318 switch (code)
4319 {
4320 case REG:
4321 case CONST_INT:
4322 case CONST:
4323 case CONST_DOUBLE:
4324 case SYMBOL_REF:
4325 case LABEL_REF:
4326 case PC:
4327 case CC0:
4328 case ADDR_VEC:
4329 case ADDR_DIFF_VEC:
4330 case USE:
4331 case CLOBBER:
4332 return;
4333
4334 case MEM:
4335 {
4336 rtx src_reg;
4337 rtx add_val;
4338 rtx mult_val;
4339 int benefit;
4340
4341 benefit = general_induction_var (XEXP (x, 0),
4342 &src_reg, &add_val, &mult_val);
4343
4344 /* Don't make a DEST_ADDR giv with mult_val == 1 && add_val == 0.
4345 Such a giv isn't useful. */
4346 if (benefit > 0 && (mult_val != const1_rtx || add_val != const0_rtx))
4347 {
4348 /* Found one; record it. */
4349 struct induction *v
4350 = (struct induction *) oballoc (sizeof (struct induction));
4351
4352 record_giv (v, insn, src_reg, addr_placeholder, mult_val,
4353 add_val, benefit, DEST_ADDR, not_every_iteration,
4354 &XEXP (x, 0), loop_start, loop_end);
4355
4356 v->mem_mode = GET_MODE (x);
4357 }
4358 return;
4359 }
4360 }
4361
4362 /* Recursively scan the subexpressions for other mem refs. */
4363
4364 fmt = GET_RTX_FORMAT (code);
4365 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4366 if (fmt[i] == 'e')
4367 find_mem_givs (XEXP (x, i), insn, not_every_iteration, loop_start,
4368 loop_end);
4369 else if (fmt[i] == 'E')
4370 for (j = 0; j < XVECLEN (x, i); j++)
4371 find_mem_givs (XVECEXP (x, i, j), insn, not_every_iteration,
4372 loop_start, loop_end);
4373 }
4374 \f
4375 /* Fill in the data about one biv update.
4376 V is the `struct induction' in which we record the biv. (It is
4377 allocated by the caller, with alloca.)
4378 INSN is the insn that sets it.
4379 DEST_REG is the biv's reg.
4380
4381 MULT_VAL is const1_rtx if the biv is being incremented here, in which case
4382 INC_VAL is the increment. Otherwise, MULT_VAL is const0_rtx and the biv is
4383 being set to INC_VAL.
4384
4385 NOT_EVERY_ITERATION is nonzero if this biv update is not know to be
4386 executed every iteration; MAYBE_MULTIPLE is nonzero if this biv update
4387 can be executed more than once per iteration. If MAYBE_MULTIPLE
4388 and NOT_EVERY_ITERATION are both zero, we know that the biv update is
4389 executed exactly once per iteration. */
4390
4391 static void
4392 record_biv (v, insn, dest_reg, inc_val, mult_val,
4393 not_every_iteration, maybe_multiple)
4394 struct induction *v;
4395 rtx insn;
4396 rtx dest_reg;
4397 rtx inc_val;
4398 rtx mult_val;
4399 int not_every_iteration;
4400 int maybe_multiple;
4401 {
4402 struct iv_class *bl;
4403
4404 v->insn = insn;
4405 v->src_reg = dest_reg;
4406 v->dest_reg = dest_reg;
4407 v->mult_val = mult_val;
4408 v->add_val = inc_val;
4409 v->mode = GET_MODE (dest_reg);
4410 v->always_computable = ! not_every_iteration;
4411 v->always_executed = ! not_every_iteration;
4412 v->maybe_multiple = maybe_multiple;
4413
4414 /* Add this to the reg's iv_class, creating a class
4415 if this is the first incrementation of the reg. */
4416
4417 bl = reg_biv_class[REGNO (dest_reg)];
4418 if (bl == 0)
4419 {
4420 /* Create and initialize new iv_class. */
4421
4422 bl = (struct iv_class *) oballoc (sizeof (struct iv_class));
4423
4424 bl->regno = REGNO (dest_reg);
4425 bl->biv = 0;
4426 bl->giv = 0;
4427 bl->biv_count = 0;
4428 bl->giv_count = 0;
4429
4430 /* Set initial value to the reg itself. */
4431 bl->initial_value = dest_reg;
4432 /* We haven't seen the initializing insn yet */
4433 bl->init_insn = 0;
4434 bl->init_set = 0;
4435 bl->initial_test = 0;
4436 bl->incremented = 0;
4437 bl->eliminable = 0;
4438 bl->nonneg = 0;
4439 bl->reversed = 0;
4440 bl->total_benefit = 0;
4441
4442 /* Add this class to loop_iv_list. */
4443 bl->next = loop_iv_list;
4444 loop_iv_list = bl;
4445
4446 /* Put it in the array of biv register classes. */
4447 reg_biv_class[REGNO (dest_reg)] = bl;
4448 }
4449
4450 /* Update IV_CLASS entry for this biv. */
4451 v->next_iv = bl->biv;
4452 bl->biv = v;
4453 bl->biv_count++;
4454 if (mult_val == const1_rtx)
4455 bl->incremented = 1;
4456
4457 if (loop_dump_stream)
4458 {
4459 fprintf (loop_dump_stream,
4460 "Insn %d: possible biv, reg %d,",
4461 INSN_UID (insn), REGNO (dest_reg));
4462 if (GET_CODE (inc_val) == CONST_INT)
4463 fprintf (loop_dump_stream, " const = %d\n",
4464 INTVAL (inc_val));
4465 else
4466 {
4467 fprintf (loop_dump_stream, " const = ");
4468 print_rtl (loop_dump_stream, inc_val);
4469 fprintf (loop_dump_stream, "\n");
4470 }
4471 }
4472 }
4473 \f
4474 /* Fill in the data about one giv.
4475 V is the `struct induction' in which we record the giv. (It is
4476 allocated by the caller, with alloca.)
4477 INSN is the insn that sets it.
4478 BENEFIT estimates the savings from deleting this insn.
4479 TYPE is DEST_REG or DEST_ADDR; it says whether the giv is computed
4480 into a register or is used as a memory address.
4481
4482 SRC_REG is the biv reg which the giv is computed from.
4483 DEST_REG is the giv's reg (if the giv is stored in a reg).
4484 MULT_VAL and ADD_VAL are the coefficients used to compute the giv.
4485 LOCATION points to the place where this giv's value appears in INSN. */
4486
4487 static void
4488 record_giv (v, insn, src_reg, dest_reg, mult_val, add_val, benefit,
4489 type, not_every_iteration, location, loop_start, loop_end)
4490 struct induction *v;
4491 rtx insn;
4492 rtx src_reg;
4493 rtx dest_reg;
4494 rtx mult_val, add_val;
4495 int benefit;
4496 enum g_types type;
4497 int not_every_iteration;
4498 rtx *location;
4499 rtx loop_start, loop_end;
4500 {
4501 struct induction *b;
4502 struct iv_class *bl;
4503 rtx set = single_set (insn);
4504 rtx p;
4505
4506 v->insn = insn;
4507 v->src_reg = src_reg;
4508 v->giv_type = type;
4509 v->dest_reg = dest_reg;
4510 v->mult_val = mult_val;
4511 v->add_val = add_val;
4512 v->benefit = benefit;
4513 v->location = location;
4514 v->cant_derive = 0;
4515 v->combined_with = 0;
4516 v->maybe_multiple = 0;
4517 v->maybe_dead = 0;
4518 v->derive_adjustment = 0;
4519 v->same = 0;
4520 v->ignore = 0;
4521 v->new_reg = 0;
4522 v->final_value = 0;
4523 v->same_insn = 0;
4524 v->auto_inc_opt = 0;
4525
4526 /* The v->always_computable field is used in update_giv_derive, to
4527 determine whether a giv can be used to derive another giv. For a
4528 DEST_REG giv, INSN computes a new value for the giv, so its value
4529 isn't computable if INSN insn't executed every iteration.
4530 However, for a DEST_ADDR giv, INSN merely uses the value of the giv;
4531 it does not compute a new value. Hence the value is always computable
4532 regardless of whether INSN is executed each iteration. */
4533
4534 if (type == DEST_ADDR)
4535 v->always_computable = 1;
4536 else
4537 v->always_computable = ! not_every_iteration;
4538
4539 v->always_executed = ! not_every_iteration;
4540
4541 if (type == DEST_ADDR)
4542 {
4543 v->mode = GET_MODE (*location);
4544 v->lifetime = 1;
4545 v->times_used = 1;
4546 }
4547 else /* type == DEST_REG */
4548 {
4549 v->mode = GET_MODE (SET_DEST (set));
4550
4551 v->lifetime = (uid_luid[REGNO_LAST_UID (REGNO (dest_reg))]
4552 - uid_luid[REGNO_FIRST_UID (REGNO (dest_reg))]);
4553
4554 v->times_used = n_times_used[REGNO (dest_reg)];
4555
4556 /* If the lifetime is zero, it means that this register is
4557 really a dead store. So mark this as a giv that can be
4558 ignored. This will not prevent the biv from being eliminated. */
4559 if (v->lifetime == 0)
4560 v->ignore = 1;
4561
4562 reg_iv_type[REGNO (dest_reg)] = GENERAL_INDUCT;
4563 reg_iv_info[REGNO (dest_reg)] = v;
4564 }
4565
4566 /* Add the giv to the class of givs computed from one biv. */
4567
4568 bl = reg_biv_class[REGNO (src_reg)];
4569 if (bl)
4570 {
4571 v->next_iv = bl->giv;
4572 bl->giv = v;
4573 /* Don't count DEST_ADDR. This is supposed to count the number of
4574 insns that calculate givs. */
4575 if (type == DEST_REG)
4576 bl->giv_count++;
4577 bl->total_benefit += benefit;
4578 }
4579 else
4580 /* Fatal error, biv missing for this giv? */
4581 abort ();
4582
4583 if (type == DEST_ADDR)
4584 v->replaceable = 1;
4585 else
4586 {
4587 /* The giv can be replaced outright by the reduced register only if all
4588 of the following conditions are true:
4589 - the insn that sets the giv is always executed on any iteration
4590 on which the giv is used at all
4591 (there are two ways to deduce this:
4592 either the insn is executed on every iteration,
4593 or all uses follow that insn in the same basic block),
4594 - the giv is not used outside the loop
4595 - no assignments to the biv occur during the giv's lifetime. */
4596
4597 if (REGNO_FIRST_UID (REGNO (dest_reg)) == INSN_UID (insn)
4598 /* Previous line always fails if INSN was moved by loop opt. */
4599 && uid_luid[REGNO_LAST_UID (REGNO (dest_reg))] < INSN_LUID (loop_end)
4600 && (! not_every_iteration
4601 || last_use_this_basic_block (dest_reg, insn)))
4602 {
4603 /* Now check that there are no assignments to the biv within the
4604 giv's lifetime. This requires two separate checks. */
4605
4606 /* Check each biv update, and fail if any are between the first
4607 and last use of the giv.
4608
4609 If this loop contains an inner loop that was unrolled, then
4610 the insn modifying the biv may have been emitted by the loop
4611 unrolling code, and hence does not have a valid luid. Just
4612 mark the biv as not replaceable in this case. It is not very
4613 useful as a biv, because it is used in two different loops.
4614 It is very unlikely that we would be able to optimize the giv
4615 using this biv anyways. */
4616
4617 v->replaceable = 1;
4618 for (b = bl->biv; b; b = b->next_iv)
4619 {
4620 if (INSN_UID (b->insn) >= max_uid_for_loop
4621 || ((uid_luid[INSN_UID (b->insn)]
4622 >= uid_luid[REGNO_FIRST_UID (REGNO (dest_reg))])
4623 && (uid_luid[INSN_UID (b->insn)]
4624 <= uid_luid[REGNO_LAST_UID (REGNO (dest_reg))])))
4625 {
4626 v->replaceable = 0;
4627 v->not_replaceable = 1;
4628 break;
4629 }
4630 }
4631
4632 /* If there are any backwards branches that go from after the
4633 biv update to before it, then this giv is not replaceable. */
4634 if (v->replaceable)
4635 for (b = bl->biv; b; b = b->next_iv)
4636 if (back_branch_in_range_p (b->insn, loop_start, loop_end))
4637 {
4638 v->replaceable = 0;
4639 v->not_replaceable = 1;
4640 break;
4641 }
4642 }
4643 else
4644 {
4645 /* May still be replaceable, we don't have enough info here to
4646 decide. */
4647 v->replaceable = 0;
4648 v->not_replaceable = 0;
4649 }
4650 }
4651
4652 if (loop_dump_stream)
4653 {
4654 if (type == DEST_REG)
4655 fprintf (loop_dump_stream, "Insn %d: giv reg %d",
4656 INSN_UID (insn), REGNO (dest_reg));
4657 else
4658 fprintf (loop_dump_stream, "Insn %d: dest address",
4659 INSN_UID (insn));
4660
4661 fprintf (loop_dump_stream, " src reg %d benefit %d",
4662 REGNO (src_reg), v->benefit);
4663 fprintf (loop_dump_stream, " used %d lifetime %d",
4664 v->times_used, v->lifetime);
4665
4666 if (v->replaceable)
4667 fprintf (loop_dump_stream, " replaceable");
4668
4669 if (GET_CODE (mult_val) == CONST_INT)
4670 fprintf (loop_dump_stream, " mult %d",
4671 INTVAL (mult_val));
4672 else
4673 {
4674 fprintf (loop_dump_stream, " mult ");
4675 print_rtl (loop_dump_stream, mult_val);
4676 }
4677
4678 if (GET_CODE (add_val) == CONST_INT)
4679 fprintf (loop_dump_stream, " add %d",
4680 INTVAL (add_val));
4681 else
4682 {
4683 fprintf (loop_dump_stream, " add ");
4684 print_rtl (loop_dump_stream, add_val);
4685 }
4686 }
4687
4688 if (loop_dump_stream)
4689 fprintf (loop_dump_stream, "\n");
4690
4691 }
4692
4693
4694 /* All this does is determine whether a giv can be made replaceable because
4695 its final value can be calculated. This code can not be part of record_giv
4696 above, because final_giv_value requires that the number of loop iterations
4697 be known, and that can not be accurately calculated until after all givs
4698 have been identified. */
4699
4700 static void
4701 check_final_value (v, loop_start, loop_end)
4702 struct induction *v;
4703 rtx loop_start, loop_end;
4704 {
4705 struct iv_class *bl;
4706 rtx final_value = 0;
4707
4708 bl = reg_biv_class[REGNO (v->src_reg)];
4709
4710 /* DEST_ADDR givs will never reach here, because they are always marked
4711 replaceable above in record_giv. */
4712
4713 /* The giv can be replaced outright by the reduced register only if all
4714 of the following conditions are true:
4715 - the insn that sets the giv is always executed on any iteration
4716 on which the giv is used at all
4717 (there are two ways to deduce this:
4718 either the insn is executed on every iteration,
4719 or all uses follow that insn in the same basic block),
4720 - its final value can be calculated (this condition is different
4721 than the one above in record_giv)
4722 - no assignments to the biv occur during the giv's lifetime. */
4723
4724 #if 0
4725 /* This is only called now when replaceable is known to be false. */
4726 /* Clear replaceable, so that it won't confuse final_giv_value. */
4727 v->replaceable = 0;
4728 #endif
4729
4730 if ((final_value = final_giv_value (v, loop_start, loop_end))
4731 && (v->always_computable || last_use_this_basic_block (v->dest_reg, v->insn)))
4732 {
4733 int biv_increment_seen = 0;
4734 rtx p = v->insn;
4735 rtx last_giv_use;
4736
4737 v->replaceable = 1;
4738
4739 /* When trying to determine whether or not a biv increment occurs
4740 during the lifetime of the giv, we can ignore uses of the variable
4741 outside the loop because final_value is true. Hence we can not
4742 use regno_last_uid and regno_first_uid as above in record_giv. */
4743
4744 /* Search the loop to determine whether any assignments to the
4745 biv occur during the giv's lifetime. Start with the insn
4746 that sets the giv, and search around the loop until we come
4747 back to that insn again.
4748
4749 Also fail if there is a jump within the giv's lifetime that jumps
4750 to somewhere outside the lifetime but still within the loop. This
4751 catches spaghetti code where the execution order is not linear, and
4752 hence the above test fails. Here we assume that the giv lifetime
4753 does not extend from one iteration of the loop to the next, so as
4754 to make the test easier. Since the lifetime isn't known yet,
4755 this requires two loops. See also record_giv above. */
4756
4757 last_giv_use = v->insn;
4758
4759 while (1)
4760 {
4761 p = NEXT_INSN (p);
4762 if (p == loop_end)
4763 p = NEXT_INSN (loop_start);
4764 if (p == v->insn)
4765 break;
4766
4767 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
4768 || GET_CODE (p) == CALL_INSN)
4769 {
4770 if (biv_increment_seen)
4771 {
4772 if (reg_mentioned_p (v->dest_reg, PATTERN (p)))
4773 {
4774 v->replaceable = 0;
4775 v->not_replaceable = 1;
4776 break;
4777 }
4778 }
4779 else if (GET_CODE (PATTERN (p)) == SET
4780 && SET_DEST (PATTERN (p)) == v->src_reg)
4781 biv_increment_seen = 1;
4782 else if (reg_mentioned_p (v->dest_reg, PATTERN (p)))
4783 last_giv_use = p;
4784 }
4785 }
4786
4787 /* Now that the lifetime of the giv is known, check for branches
4788 from within the lifetime to outside the lifetime if it is still
4789 replaceable. */
4790
4791 if (v->replaceable)
4792 {
4793 p = v->insn;
4794 while (1)
4795 {
4796 p = NEXT_INSN (p);
4797 if (p == loop_end)
4798 p = NEXT_INSN (loop_start);
4799 if (p == last_giv_use)
4800 break;
4801
4802 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p)
4803 && LABEL_NAME (JUMP_LABEL (p))
4804 && ((INSN_UID (JUMP_LABEL (p)) >= max_uid_for_loop)
4805 || (INSN_UID (v->insn) >= max_uid_for_loop)
4806 || (INSN_UID (last_giv_use) >= max_uid_for_loop)
4807 || (INSN_LUID (JUMP_LABEL (p)) < INSN_LUID (v->insn)
4808 && INSN_LUID (JUMP_LABEL (p)) > INSN_LUID (loop_start))
4809 || (INSN_LUID (JUMP_LABEL (p)) > INSN_LUID (last_giv_use)
4810 && INSN_LUID (JUMP_LABEL (p)) < INSN_LUID (loop_end))))
4811 {
4812 v->replaceable = 0;
4813 v->not_replaceable = 1;
4814
4815 if (loop_dump_stream)
4816 fprintf (loop_dump_stream,
4817 "Found branch outside giv lifetime.\n");
4818
4819 break;
4820 }
4821 }
4822 }
4823
4824 /* If it is replaceable, then save the final value. */
4825 if (v->replaceable)
4826 v->final_value = final_value;
4827 }
4828
4829 if (loop_dump_stream && v->replaceable)
4830 fprintf (loop_dump_stream, "Insn %d: giv reg %d final_value replaceable\n",
4831 INSN_UID (v->insn), REGNO (v->dest_reg));
4832 }
4833 \f
4834 /* Update the status of whether a giv can derive other givs.
4835
4836 We need to do something special if there is or may be an update to the biv
4837 between the time the giv is defined and the time it is used to derive
4838 another giv.
4839
4840 In addition, a giv that is only conditionally set is not allowed to
4841 derive another giv once a label has been passed.
4842
4843 The cases we look at are when a label or an update to a biv is passed. */
4844
4845 static void
4846 update_giv_derive (p)
4847 rtx p;
4848 {
4849 struct iv_class *bl;
4850 struct induction *biv, *giv;
4851 rtx tem;
4852 int dummy;
4853
4854 /* Search all IV classes, then all bivs, and finally all givs.
4855
4856 There are three cases we are concerned with. First we have the situation
4857 of a giv that is only updated conditionally. In that case, it may not
4858 derive any givs after a label is passed.
4859
4860 The second case is when a biv update occurs, or may occur, after the
4861 definition of a giv. For certain biv updates (see below) that are
4862 known to occur between the giv definition and use, we can adjust the
4863 giv definition. For others, or when the biv update is conditional,
4864 we must prevent the giv from deriving any other givs. There are two
4865 sub-cases within this case.
4866
4867 If this is a label, we are concerned with any biv update that is done
4868 conditionally, since it may be done after the giv is defined followed by
4869 a branch here (actually, we need to pass both a jump and a label, but
4870 this extra tracking doesn't seem worth it).
4871
4872 If this is a jump, we are concerned about any biv update that may be
4873 executed multiple times. We are actually only concerned about
4874 backward jumps, but it is probably not worth performing the test
4875 on the jump again here.
4876
4877 If this is a biv update, we must adjust the giv status to show that a
4878 subsequent biv update was performed. If this adjustment cannot be done,
4879 the giv cannot derive further givs. */
4880
4881 for (bl = loop_iv_list; bl; bl = bl->next)
4882 for (biv = bl->biv; biv; biv = biv->next_iv)
4883 if (GET_CODE (p) == CODE_LABEL || GET_CODE (p) == JUMP_INSN
4884 || biv->insn == p)
4885 {
4886 for (giv = bl->giv; giv; giv = giv->next_iv)
4887 {
4888 /* If cant_derive is already true, there is no point in
4889 checking all of these conditions again. */
4890 if (giv->cant_derive)
4891 continue;
4892
4893 /* If this giv is conditionally set and we have passed a label,
4894 it cannot derive anything. */
4895 if (GET_CODE (p) == CODE_LABEL && ! giv->always_computable)
4896 giv->cant_derive = 1;
4897
4898 /* Skip givs that have mult_val == 0, since
4899 they are really invariants. Also skip those that are
4900 replaceable, since we know their lifetime doesn't contain
4901 any biv update. */
4902 else if (giv->mult_val == const0_rtx || giv->replaceable)
4903 continue;
4904
4905 /* The only way we can allow this giv to derive another
4906 is if this is a biv increment and we can form the product
4907 of biv->add_val and giv->mult_val. In this case, we will
4908 be able to compute a compensation. */
4909 else if (biv->insn == p)
4910 {
4911 tem = 0;
4912
4913 if (biv->mult_val == const1_rtx)
4914 tem = simplify_giv_expr (gen_rtx (MULT, giv->mode,
4915 biv->add_val,
4916 giv->mult_val),
4917 &dummy);
4918
4919 if (tem && giv->derive_adjustment)
4920 tem = simplify_giv_expr (gen_rtx (PLUS, giv->mode, tem,
4921 giv->derive_adjustment),
4922 &dummy);
4923 if (tem)
4924 giv->derive_adjustment = tem;
4925 else
4926 giv->cant_derive = 1;
4927 }
4928 else if ((GET_CODE (p) == CODE_LABEL && ! biv->always_computable)
4929 || (GET_CODE (p) == JUMP_INSN && biv->maybe_multiple))
4930 giv->cant_derive = 1;
4931 }
4932 }
4933 }
4934 \f
4935 /* Check whether an insn is an increment legitimate for a basic induction var.
4936 X is the source of insn P, or a part of it.
4937 MODE is the mode in which X should be interpreted.
4938
4939 DEST_REG is the putative biv, also the destination of the insn.
4940 We accept patterns of these forms:
4941 REG = REG + INVARIANT (includes REG = REG - CONSTANT)
4942 REG = INVARIANT + REG
4943
4944 If X is suitable, we return 1, set *MULT_VAL to CONST1_RTX,
4945 and store the additive term into *INC_VAL.
4946
4947 If X is an assignment of an invariant into DEST_REG, we set
4948 *MULT_VAL to CONST0_RTX, and store the invariant into *INC_VAL.
4949
4950 We also want to detect a BIV when it corresponds to a variable
4951 whose mode was promoted via PROMOTED_MODE. In that case, an increment
4952 of the variable may be a PLUS that adds a SUBREG of that variable to
4953 an invariant and then sign- or zero-extends the result of the PLUS
4954 into the variable.
4955
4956 Most GIVs in such cases will be in the promoted mode, since that is the
4957 probably the natural computation mode (and almost certainly the mode
4958 used for addresses) on the machine. So we view the pseudo-reg containing
4959 the variable as the BIV, as if it were simply incremented.
4960
4961 Note that treating the entire pseudo as a BIV will result in making
4962 simple increments to any GIVs based on it. However, if the variable
4963 overflows in its declared mode but not its promoted mode, the result will
4964 be incorrect. This is acceptable if the variable is signed, since
4965 overflows in such cases are undefined, but not if it is unsigned, since
4966 those overflows are defined. So we only check for SIGN_EXTEND and
4967 not ZERO_EXTEND.
4968
4969 If we cannot find a biv, we return 0. */
4970
4971 static int
4972 basic_induction_var (x, mode, dest_reg, p, inc_val, mult_val)
4973 register rtx x;
4974 enum machine_mode mode;
4975 rtx p;
4976 rtx dest_reg;
4977 rtx *inc_val;
4978 rtx *mult_val;
4979 {
4980 register enum rtx_code code;
4981 rtx arg;
4982 rtx insn, set = 0;
4983
4984 code = GET_CODE (x);
4985 switch (code)
4986 {
4987 case PLUS:
4988 if (XEXP (x, 0) == dest_reg
4989 || (GET_CODE (XEXP (x, 0)) == SUBREG
4990 && SUBREG_PROMOTED_VAR_P (XEXP (x, 0))
4991 && SUBREG_REG (XEXP (x, 0)) == dest_reg))
4992 arg = XEXP (x, 1);
4993 else if (XEXP (x, 1) == dest_reg
4994 || (GET_CODE (XEXP (x, 1)) == SUBREG
4995 && SUBREG_PROMOTED_VAR_P (XEXP (x, 1))
4996 && SUBREG_REG (XEXP (x, 1)) == dest_reg))
4997 arg = XEXP (x, 0);
4998 else
4999 return 0;
5000
5001 if (invariant_p (arg) != 1)
5002 return 0;
5003
5004 *inc_val = convert_modes (GET_MODE (dest_reg), GET_MODE (x), arg, 0);
5005 *mult_val = const1_rtx;
5006 return 1;
5007
5008 case SUBREG:
5009 /* If this is a SUBREG for a promoted variable, check the inner
5010 value. */
5011 if (SUBREG_PROMOTED_VAR_P (x))
5012 return basic_induction_var (SUBREG_REG (x), GET_MODE (SUBREG_REG (x)),
5013 dest_reg, p, inc_val, mult_val);
5014 return 0;
5015
5016 case REG:
5017 /* If this register is assigned in the previous insn, look at its
5018 source, but don't go outside the loop or past a label. */
5019
5020 for (insn = PREV_INSN (p);
5021 (insn && GET_CODE (insn) == NOTE
5022 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_BEG);
5023 insn = PREV_INSN (insn))
5024 ;
5025
5026 if (insn)
5027 set = single_set (insn);
5028
5029 if (set != 0
5030 && (SET_DEST (set) == x
5031 || (GET_CODE (SET_DEST (set)) == SUBREG
5032 && (GET_MODE_SIZE (GET_MODE (SET_DEST (set)))
5033 <= UNITS_PER_WORD)
5034 && SUBREG_REG (SET_DEST (set)) == x)))
5035 return basic_induction_var (SET_SRC (set),
5036 (GET_MODE (SET_SRC (set)) == VOIDmode
5037 ? GET_MODE (x)
5038 : GET_MODE (SET_SRC (set))),
5039 dest_reg, insn,
5040 inc_val, mult_val);
5041 /* ... fall through ... */
5042
5043 /* Can accept constant setting of biv only when inside inner most loop.
5044 Otherwise, a biv of an inner loop may be incorrectly recognized
5045 as a biv of the outer loop,
5046 causing code to be moved INTO the inner loop. */
5047 case MEM:
5048 if (invariant_p (x) != 1)
5049 return 0;
5050 case CONST_INT:
5051 case SYMBOL_REF:
5052 case CONST:
5053 if (loops_enclosed == 1)
5054 {
5055 /* Possible bug here? Perhaps we don't know the mode of X. */
5056 *inc_val = convert_modes (GET_MODE (dest_reg), mode, x, 0);
5057 *mult_val = const0_rtx;
5058 return 1;
5059 }
5060 else
5061 return 0;
5062
5063 case SIGN_EXTEND:
5064 return basic_induction_var (XEXP (x, 0), GET_MODE (XEXP (x, 0)),
5065 dest_reg, p, inc_val, mult_val);
5066 case ASHIFTRT:
5067 /* Similar, since this can be a sign extension. */
5068 for (insn = PREV_INSN (p);
5069 (insn && GET_CODE (insn) == NOTE
5070 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_BEG);
5071 insn = PREV_INSN (insn))
5072 ;
5073
5074 if (insn)
5075 set = single_set (insn);
5076
5077 if (set && SET_DEST (set) == XEXP (x, 0)
5078 && GET_CODE (XEXP (x, 1)) == CONST_INT
5079 && INTVAL (XEXP (x, 1)) >= 0
5080 && GET_CODE (SET_SRC (set)) == ASHIFT
5081 && XEXP (x, 1) == XEXP (SET_SRC (set), 1))
5082 return basic_induction_var (XEXP (SET_SRC (set), 0),
5083 GET_MODE (XEXP (x, 0)),
5084 dest_reg, insn, inc_val, mult_val);
5085 return 0;
5086
5087 default:
5088 return 0;
5089 }
5090 }
5091 \f
5092 /* A general induction variable (giv) is any quantity that is a linear
5093 function of a basic induction variable,
5094 i.e. giv = biv * mult_val + add_val.
5095 The coefficients can be any loop invariant quantity.
5096 A giv need not be computed directly from the biv;
5097 it can be computed by way of other givs. */
5098
5099 /* Determine whether X computes a giv.
5100 If it does, return a nonzero value
5101 which is the benefit from eliminating the computation of X;
5102 set *SRC_REG to the register of the biv that it is computed from;
5103 set *ADD_VAL and *MULT_VAL to the coefficients,
5104 such that the value of X is biv * mult + add; */
5105
5106 static int
5107 general_induction_var (x, src_reg, add_val, mult_val)
5108 rtx x;
5109 rtx *src_reg;
5110 rtx *add_val;
5111 rtx *mult_val;
5112 {
5113 rtx orig_x = x;
5114 int benefit = 0;
5115 char *storage;
5116
5117 /* If this is an invariant, forget it, it isn't a giv. */
5118 if (invariant_p (x) == 1)
5119 return 0;
5120
5121 /* See if the expression could be a giv and get its form.
5122 Mark our place on the obstack in case we don't find a giv. */
5123 storage = (char *) oballoc (0);
5124 x = simplify_giv_expr (x, &benefit);
5125 if (x == 0)
5126 {
5127 obfree (storage);
5128 return 0;
5129 }
5130
5131 switch (GET_CODE (x))
5132 {
5133 case USE:
5134 case CONST_INT:
5135 /* Since this is now an invariant and wasn't before, it must be a giv
5136 with MULT_VAL == 0. It doesn't matter which BIV we associate this
5137 with. */
5138 *src_reg = loop_iv_list->biv->dest_reg;
5139 *mult_val = const0_rtx;
5140 *add_val = x;
5141 break;
5142
5143 case REG:
5144 /* This is equivalent to a BIV. */
5145 *src_reg = x;
5146 *mult_val = const1_rtx;
5147 *add_val = const0_rtx;
5148 break;
5149
5150 case PLUS:
5151 /* Either (plus (biv) (invar)) or
5152 (plus (mult (biv) (invar_1)) (invar_2)). */
5153 if (GET_CODE (XEXP (x, 0)) == MULT)
5154 {
5155 *src_reg = XEXP (XEXP (x, 0), 0);
5156 *mult_val = XEXP (XEXP (x, 0), 1);
5157 }
5158 else
5159 {
5160 *src_reg = XEXP (x, 0);
5161 *mult_val = const1_rtx;
5162 }
5163 *add_val = XEXP (x, 1);
5164 break;
5165
5166 case MULT:
5167 /* ADD_VAL is zero. */
5168 *src_reg = XEXP (x, 0);
5169 *mult_val = XEXP (x, 1);
5170 *add_val = const0_rtx;
5171 break;
5172
5173 default:
5174 abort ();
5175 }
5176
5177 /* Remove any enclosing USE from ADD_VAL and MULT_VAL (there will be
5178 unless they are CONST_INT). */
5179 if (GET_CODE (*add_val) == USE)
5180 *add_val = XEXP (*add_val, 0);
5181 if (GET_CODE (*mult_val) == USE)
5182 *mult_val = XEXP (*mult_val, 0);
5183
5184 benefit += rtx_cost (orig_x, SET);
5185
5186 /* Always return some benefit if this is a giv so it will be detected
5187 as such. This allows elimination of bivs that might otherwise
5188 not be eliminated. */
5189 return benefit == 0 ? 1 : benefit;
5190 }
5191 \f
5192 /* Given an expression, X, try to form it as a linear function of a biv.
5193 We will canonicalize it to be of the form
5194 (plus (mult (BIV) (invar_1))
5195 (invar_2))
5196 with possible degeneracies.
5197
5198 The invariant expressions must each be of a form that can be used as a
5199 machine operand. We surround then with a USE rtx (a hack, but localized
5200 and certainly unambiguous!) if not a CONST_INT for simplicity in this
5201 routine; it is the caller's responsibility to strip them.
5202
5203 If no such canonicalization is possible (i.e., two biv's are used or an
5204 expression that is neither invariant nor a biv or giv), this routine
5205 returns 0.
5206
5207 For a non-zero return, the result will have a code of CONST_INT, USE,
5208 REG (for a BIV), PLUS, or MULT. No other codes will occur.
5209
5210 *BENEFIT will be incremented by the benefit of any sub-giv encountered. */
5211
5212 static rtx
5213 simplify_giv_expr (x, benefit)
5214 rtx x;
5215 int *benefit;
5216 {
5217 enum machine_mode mode = GET_MODE (x);
5218 rtx arg0, arg1;
5219 rtx tem;
5220
5221 /* If this is not an integer mode, or if we cannot do arithmetic in this
5222 mode, this can't be a giv. */
5223 if (mode != VOIDmode
5224 && (GET_MODE_CLASS (mode) != MODE_INT
5225 || GET_MODE_BITSIZE (mode) > HOST_BITS_PER_WIDE_INT))
5226 return 0;
5227
5228 switch (GET_CODE (x))
5229 {
5230 case PLUS:
5231 arg0 = simplify_giv_expr (XEXP (x, 0), benefit);
5232 arg1 = simplify_giv_expr (XEXP (x, 1), benefit);
5233 if (arg0 == 0 || arg1 == 0)
5234 return 0;
5235
5236 /* Put constant last, CONST_INT last if both constant. */
5237 if ((GET_CODE (arg0) == USE
5238 || GET_CODE (arg0) == CONST_INT)
5239 && GET_CODE (arg1) != CONST_INT)
5240 tem = arg0, arg0 = arg1, arg1 = tem;
5241
5242 /* Handle addition of zero, then addition of an invariant. */
5243 if (arg1 == const0_rtx)
5244 return arg0;
5245 else if (GET_CODE (arg1) == CONST_INT || GET_CODE (arg1) == USE)
5246 switch (GET_CODE (arg0))
5247 {
5248 case CONST_INT:
5249 case USE:
5250 /* Both invariant. Only valid if sum is machine operand.
5251 First strip off possible USE on first operand. */
5252 if (GET_CODE (arg0) == USE)
5253 arg0 = XEXP (arg0, 0);
5254
5255 tem = 0;
5256 if (CONSTANT_P (arg0) && GET_CODE (arg1) == CONST_INT)
5257 {
5258 tem = plus_constant (arg0, INTVAL (arg1));
5259 if (GET_CODE (tem) != CONST_INT)
5260 tem = gen_rtx (USE, mode, tem);
5261 }
5262
5263 return tem;
5264
5265 case REG:
5266 case MULT:
5267 /* biv + invar or mult + invar. Return sum. */
5268 return gen_rtx (PLUS, mode, arg0, arg1);
5269
5270 case PLUS:
5271 /* (a + invar_1) + invar_2. Associate. */
5272 return simplify_giv_expr (gen_rtx (PLUS, mode,
5273 XEXP (arg0, 0),
5274 gen_rtx (PLUS, mode,
5275 XEXP (arg0, 1), arg1)),
5276 benefit);
5277
5278 default:
5279 abort ();
5280 }
5281
5282 /* Each argument must be either REG, PLUS, or MULT. Convert REG to
5283 MULT to reduce cases. */
5284 if (GET_CODE (arg0) == REG)
5285 arg0 = gen_rtx (MULT, mode, arg0, const1_rtx);
5286 if (GET_CODE (arg1) == REG)
5287 arg1 = gen_rtx (MULT, mode, arg1, const1_rtx);
5288
5289 /* Now have PLUS + PLUS, PLUS + MULT, MULT + PLUS, or MULT + MULT.
5290 Put a MULT first, leaving PLUS + PLUS, MULT + PLUS, or MULT + MULT.
5291 Recurse to associate the second PLUS. */
5292 if (GET_CODE (arg1) == MULT)
5293 tem = arg0, arg0 = arg1, arg1 = tem;
5294
5295 if (GET_CODE (arg1) == PLUS)
5296 return simplify_giv_expr (gen_rtx (PLUS, mode,
5297 gen_rtx (PLUS, mode,
5298 arg0, XEXP (arg1, 0)),
5299 XEXP (arg1, 1)),
5300 benefit);
5301
5302 /* Now must have MULT + MULT. Distribute if same biv, else not giv. */
5303 if (GET_CODE (arg0) != MULT || GET_CODE (arg1) != MULT)
5304 abort ();
5305
5306 if (XEXP (arg0, 0) != XEXP (arg1, 0))
5307 return 0;
5308
5309 return simplify_giv_expr (gen_rtx (MULT, mode,
5310 XEXP (arg0, 0),
5311 gen_rtx (PLUS, mode,
5312 XEXP (arg0, 1),
5313 XEXP (arg1, 1))),
5314 benefit);
5315
5316 case MINUS:
5317 /* Handle "a - b" as "a + b * (-1)". */
5318 return simplify_giv_expr (gen_rtx (PLUS, mode,
5319 XEXP (x, 0),
5320 gen_rtx (MULT, mode,
5321 XEXP (x, 1), constm1_rtx)),
5322 benefit);
5323
5324 case MULT:
5325 arg0 = simplify_giv_expr (XEXP (x, 0), benefit);
5326 arg1 = simplify_giv_expr (XEXP (x, 1), benefit);
5327 if (arg0 == 0 || arg1 == 0)
5328 return 0;
5329
5330 /* Put constant last, CONST_INT last if both constant. */
5331 if ((GET_CODE (arg0) == USE || GET_CODE (arg0) == CONST_INT)
5332 && GET_CODE (arg1) != CONST_INT)
5333 tem = arg0, arg0 = arg1, arg1 = tem;
5334
5335 /* If second argument is not now constant, not giv. */
5336 if (GET_CODE (arg1) != USE && GET_CODE (arg1) != CONST_INT)
5337 return 0;
5338
5339 /* Handle multiply by 0 or 1. */
5340 if (arg1 == const0_rtx)
5341 return const0_rtx;
5342
5343 else if (arg1 == const1_rtx)
5344 return arg0;
5345
5346 switch (GET_CODE (arg0))
5347 {
5348 case REG:
5349 /* biv * invar. Done. */
5350 return gen_rtx (MULT, mode, arg0, arg1);
5351
5352 case CONST_INT:
5353 /* Product of two constants. */
5354 return GEN_INT (INTVAL (arg0) * INTVAL (arg1));
5355
5356 case USE:
5357 /* invar * invar. Not giv. */
5358 return 0;
5359
5360 case MULT:
5361 /* (a * invar_1) * invar_2. Associate. */
5362 return simplify_giv_expr (gen_rtx (MULT, mode,
5363 XEXP (arg0, 0),
5364 gen_rtx (MULT, mode,
5365 XEXP (arg0, 1), arg1)),
5366 benefit);
5367
5368 case PLUS:
5369 /* (a + invar_1) * invar_2. Distribute. */
5370 return simplify_giv_expr (gen_rtx (PLUS, mode,
5371 gen_rtx (MULT, mode,
5372 XEXP (arg0, 0), arg1),
5373 gen_rtx (MULT, mode,
5374 XEXP (arg0, 1), arg1)),
5375 benefit);
5376
5377 default:
5378 abort ();
5379 }
5380
5381 case ASHIFT:
5382 /* Shift by constant is multiply by power of two. */
5383 if (GET_CODE (XEXP (x, 1)) != CONST_INT)
5384 return 0;
5385
5386 return simplify_giv_expr (gen_rtx (MULT, mode,
5387 XEXP (x, 0),
5388 GEN_INT ((HOST_WIDE_INT) 1
5389 << INTVAL (XEXP (x, 1)))),
5390 benefit);
5391
5392 case NEG:
5393 /* "-a" is "a * (-1)" */
5394 return simplify_giv_expr (gen_rtx (MULT, mode, XEXP (x, 0), constm1_rtx),
5395 benefit);
5396
5397 case NOT:
5398 /* "~a" is "-a - 1". Silly, but easy. */
5399 return simplify_giv_expr (gen_rtx (MINUS, mode,
5400 gen_rtx (NEG, mode, XEXP (x, 0)),
5401 const1_rtx),
5402 benefit);
5403
5404 case USE:
5405 /* Already in proper form for invariant. */
5406 return x;
5407
5408 case REG:
5409 /* If this is a new register, we can't deal with it. */
5410 if (REGNO (x) >= max_reg_before_loop)
5411 return 0;
5412
5413 /* Check for biv or giv. */
5414 switch (reg_iv_type[REGNO (x)])
5415 {
5416 case BASIC_INDUCT:
5417 return x;
5418 case GENERAL_INDUCT:
5419 {
5420 struct induction *v = reg_iv_info[REGNO (x)];
5421
5422 /* Form expression from giv and add benefit. Ensure this giv
5423 can derive another and subtract any needed adjustment if so. */
5424 *benefit += v->benefit;
5425 if (v->cant_derive)
5426 return 0;
5427
5428 tem = gen_rtx (PLUS, mode, gen_rtx (MULT, mode,
5429 v->src_reg, v->mult_val),
5430 v->add_val);
5431 if (v->derive_adjustment)
5432 tem = gen_rtx (MINUS, mode, tem, v->derive_adjustment);
5433 return simplify_giv_expr (tem, benefit);
5434 }
5435 }
5436
5437 /* Fall through to general case. */
5438 default:
5439 /* If invariant, return as USE (unless CONST_INT).
5440 Otherwise, not giv. */
5441 if (GET_CODE (x) == USE)
5442 x = XEXP (x, 0);
5443
5444 if (invariant_p (x) == 1)
5445 {
5446 if (GET_CODE (x) == CONST_INT)
5447 return x;
5448 else
5449 return gen_rtx (USE, mode, x);
5450 }
5451 else
5452 return 0;
5453 }
5454 }
5455 \f
5456 /* Help detect a giv that is calculated by several consecutive insns;
5457 for example,
5458 giv = biv * M
5459 giv = giv + A
5460 The caller has already identified the first insn P as having a giv as dest;
5461 we check that all other insns that set the same register follow
5462 immediately after P, that they alter nothing else,
5463 and that the result of the last is still a giv.
5464
5465 The value is 0 if the reg set in P is not really a giv.
5466 Otherwise, the value is the amount gained by eliminating
5467 all the consecutive insns that compute the value.
5468
5469 FIRST_BENEFIT is the amount gained by eliminating the first insn, P.
5470 SRC_REG is the reg of the biv; DEST_REG is the reg of the giv.
5471
5472 The coefficients of the ultimate giv value are stored in
5473 *MULT_VAL and *ADD_VAL. */
5474
5475 static int
5476 consec_sets_giv (first_benefit, p, src_reg, dest_reg,
5477 add_val, mult_val)
5478 int first_benefit;
5479 rtx p;
5480 rtx src_reg;
5481 rtx dest_reg;
5482 rtx *add_val;
5483 rtx *mult_val;
5484 {
5485 int count;
5486 enum rtx_code code;
5487 int benefit;
5488 rtx temp;
5489 rtx set;
5490
5491 /* Indicate that this is a giv so that we can update the value produced in
5492 each insn of the multi-insn sequence.
5493
5494 This induction structure will be used only by the call to
5495 general_induction_var below, so we can allocate it on our stack.
5496 If this is a giv, our caller will replace the induct var entry with
5497 a new induction structure. */
5498 struct induction *v
5499 = (struct induction *) alloca (sizeof (struct induction));
5500 v->src_reg = src_reg;
5501 v->mult_val = *mult_val;
5502 v->add_val = *add_val;
5503 v->benefit = first_benefit;
5504 v->cant_derive = 0;
5505 v->derive_adjustment = 0;
5506
5507 reg_iv_type[REGNO (dest_reg)] = GENERAL_INDUCT;
5508 reg_iv_info[REGNO (dest_reg)] = v;
5509
5510 count = n_times_set[REGNO (dest_reg)] - 1;
5511
5512 while (count > 0)
5513 {
5514 p = NEXT_INSN (p);
5515 code = GET_CODE (p);
5516
5517 /* If libcall, skip to end of call sequence. */
5518 if (code == INSN && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
5519 p = XEXP (temp, 0);
5520
5521 if (code == INSN
5522 && (set = single_set (p))
5523 && GET_CODE (SET_DEST (set)) == REG
5524 && SET_DEST (set) == dest_reg
5525 && ((benefit = general_induction_var (SET_SRC (set), &src_reg,
5526 add_val, mult_val))
5527 /* Giv created by equivalent expression. */
5528 || ((temp = find_reg_note (p, REG_EQUAL, NULL_RTX))
5529 && (benefit = general_induction_var (XEXP (temp, 0), &src_reg,
5530 add_val, mult_val))))
5531 && src_reg == v->src_reg)
5532 {
5533 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
5534 benefit += libcall_benefit (p);
5535
5536 count--;
5537 v->mult_val = *mult_val;
5538 v->add_val = *add_val;
5539 v->benefit = benefit;
5540 }
5541 else if (code != NOTE)
5542 {
5543 /* Allow insns that set something other than this giv to a
5544 constant. Such insns are needed on machines which cannot
5545 include long constants and should not disqualify a giv. */
5546 if (code == INSN
5547 && (set = single_set (p))
5548 && SET_DEST (set) != dest_reg
5549 && CONSTANT_P (SET_SRC (set)))
5550 continue;
5551
5552 reg_iv_type[REGNO (dest_reg)] = UNKNOWN_INDUCT;
5553 return 0;
5554 }
5555 }
5556
5557 return v->benefit;
5558 }
5559 \f
5560 /* Return an rtx, if any, that expresses giv G2 as a function of the register
5561 represented by G1. If no such expression can be found, or it is clear that
5562 it cannot possibly be a valid address, 0 is returned.
5563
5564 To perform the computation, we note that
5565 G1 = a * v + b and
5566 G2 = c * v + d
5567 where `v' is the biv.
5568
5569 So G2 = (c/a) * G1 + (d - b*c/a) */
5570
5571 #ifdef ADDRESS_COST
5572 static rtx
5573 express_from (g1, g2)
5574 struct induction *g1, *g2;
5575 {
5576 rtx mult, add;
5577
5578 /* The value that G1 will be multiplied by must be a constant integer. Also,
5579 the only chance we have of getting a valid address is if b*c/a (see above
5580 for notation) is also an integer. */
5581 if (GET_CODE (g1->mult_val) != CONST_INT
5582 || GET_CODE (g2->mult_val) != CONST_INT
5583 || GET_CODE (g1->add_val) != CONST_INT
5584 || g1->mult_val == const0_rtx
5585 || INTVAL (g2->mult_val) % INTVAL (g1->mult_val) != 0)
5586 return 0;
5587
5588 mult = GEN_INT (INTVAL (g2->mult_val) / INTVAL (g1->mult_val));
5589 add = plus_constant (g2->add_val, - INTVAL (g1->add_val) * INTVAL (mult));
5590
5591 /* Form simplified final result. */
5592 if (mult == const0_rtx)
5593 return add;
5594 else if (mult == const1_rtx)
5595 mult = g1->dest_reg;
5596 else
5597 mult = gen_rtx (MULT, g2->mode, g1->dest_reg, mult);
5598
5599 if (add == const0_rtx)
5600 return mult;
5601 else
5602 return gen_rtx (PLUS, g2->mode, mult, add);
5603 }
5604 #endif
5605 \f
5606 /* Return 1 if giv G2 can be combined with G1. This means that G2 can use
5607 (either directly or via an address expression) a register used to represent
5608 G1. Set g2->new_reg to a represtation of G1 (normally just
5609 g1->dest_reg). */
5610
5611 static int
5612 combine_givs_p (g1, g2)
5613 struct induction *g1, *g2;
5614 {
5615 rtx tem;
5616
5617 /* If these givs are identical, they can be combined. */
5618 if (rtx_equal_p (g1->mult_val, g2->mult_val)
5619 && rtx_equal_p (g1->add_val, g2->add_val))
5620 {
5621 g2->new_reg = g1->dest_reg;
5622 return 1;
5623 }
5624
5625 #ifdef ADDRESS_COST
5626 /* If G2 can be expressed as a function of G1 and that function is valid
5627 as an address and no more expensive than using a register for G2,
5628 the expression of G2 in terms of G1 can be used. */
5629 if (g2->giv_type == DEST_ADDR
5630 && (tem = express_from (g1, g2)) != 0
5631 && memory_address_p (g2->mem_mode, tem)
5632 && ADDRESS_COST (tem) <= ADDRESS_COST (*g2->location))
5633 {
5634 g2->new_reg = tem;
5635 return 1;
5636 }
5637 #endif
5638
5639 return 0;
5640 }
5641 \f
5642 #ifdef GIV_SORT_CRITERION
5643 /* Compare two givs and sort the most desirable one for combinations first.
5644 This is used only in one qsort call below. */
5645
5646 static int
5647 giv_sort (x, y)
5648 struct induction **x, **y;
5649 {
5650 GIV_SORT_CRITERION (*x, *y);
5651
5652 return 0;
5653 }
5654 #endif
5655
5656 /* Check all pairs of givs for iv_class BL and see if any can be combined with
5657 any other. If so, point SAME to the giv combined with and set NEW_REG to
5658 be an expression (in terms of the other giv's DEST_REG) equivalent to the
5659 giv. Also, update BENEFIT and related fields for cost/benefit analysis. */
5660
5661 static void
5662 combine_givs (bl)
5663 struct iv_class *bl;
5664 {
5665 struct induction *g1, *g2, **giv_array, *temp_iv;
5666 int i, j, giv_count, pass;
5667
5668 /* Count givs, because bl->giv_count is incorrect here. */
5669 giv_count = 0;
5670 for (g1 = bl->giv; g1; g1 = g1->next_iv)
5671 giv_count++;
5672
5673 giv_array
5674 = (struct induction **) alloca (giv_count * sizeof (struct induction *));
5675 i = 0;
5676 for (g1 = bl->giv; g1; g1 = g1->next_iv)
5677 giv_array[i++] = g1;
5678
5679 #ifdef GIV_SORT_CRITERION
5680 /* Sort the givs if GIV_SORT_CRITERION is defined.
5681 This is usually defined for processors which lack
5682 negative register offsets so more givs may be combined. */
5683
5684 if (loop_dump_stream)
5685 fprintf (loop_dump_stream, "%d givs counted, sorting...\n", giv_count);
5686
5687 qsort (giv_array, giv_count, sizeof (struct induction *), giv_sort);
5688 #endif
5689
5690 for (i = 0; i < giv_count; i++)
5691 {
5692 g1 = giv_array[i];
5693 for (pass = 0; pass <= 1; pass++)
5694 for (j = 0; j < giv_count; j++)
5695 {
5696 g2 = giv_array[j];
5697 if (g1 != g2
5698 /* First try to combine with replaceable givs, then all givs. */
5699 && (g1->replaceable || pass == 1)
5700 /* If either has already been combined or is to be ignored, can't
5701 combine. */
5702 && ! g1->ignore && ! g2->ignore && ! g1->same && ! g2->same
5703 /* If something has been based on G2, G2 cannot itself be based
5704 on something else. */
5705 && ! g2->combined_with
5706 && combine_givs_p (g1, g2))
5707 {
5708 /* g2->new_reg set by `combine_givs_p' */
5709 g2->same = g1;
5710 g1->combined_with = 1;
5711
5712 /* If one of these givs is a DEST_REG that was only used
5713 once, by the other giv, this is actually a single use.
5714 The DEST_REG has the correct cost, while the other giv
5715 counts the REG use too often. */
5716 if (g2->giv_type == DEST_REG
5717 && n_times_used[REGNO (g2->dest_reg)] == 1
5718 && reg_mentioned_p (g2->dest_reg, PATTERN (g1->insn)))
5719 g1->benefit = g2->benefit;
5720 else if (g1->giv_type != DEST_REG
5721 || n_times_used[REGNO (g1->dest_reg)] != 1
5722 || ! reg_mentioned_p (g1->dest_reg,
5723 PATTERN (g2->insn)))
5724 {
5725 g1->benefit += g2->benefit;
5726 g1->times_used += g2->times_used;
5727 }
5728 /* ??? The new final_[bg]iv_value code does a much better job
5729 of finding replaceable giv's, and hence this code may no
5730 longer be necessary. */
5731 if (! g2->replaceable && REG_USERVAR_P (g2->dest_reg))
5732 g1->benefit -= copy_cost;
5733 g1->lifetime += g2->lifetime;
5734
5735 if (loop_dump_stream)
5736 fprintf (loop_dump_stream, "giv at %d combined with giv at %d\n",
5737 INSN_UID (g2->insn), INSN_UID (g1->insn));
5738 }
5739 }
5740 }
5741 }
5742 \f
5743 /* EMIT code before INSERT_BEFORE to set REG = B * M + A. */
5744
5745 void
5746 emit_iv_add_mult (b, m, a, reg, insert_before)
5747 rtx b; /* initial value of basic induction variable */
5748 rtx m; /* multiplicative constant */
5749 rtx a; /* additive constant */
5750 rtx reg; /* destination register */
5751 rtx insert_before;
5752 {
5753 rtx seq;
5754 rtx result;
5755
5756 /* Prevent unexpected sharing of these rtx. */
5757 a = copy_rtx (a);
5758 b = copy_rtx (b);
5759
5760 /* Increase the lifetime of any invariants moved further in code. */
5761 update_reg_last_use (a, insert_before);
5762 update_reg_last_use (b, insert_before);
5763 update_reg_last_use (m, insert_before);
5764
5765 start_sequence ();
5766 result = expand_mult_add (b, reg, m, a, GET_MODE (reg), 0);
5767 if (reg != result)
5768 emit_move_insn (reg, result);
5769 seq = gen_sequence ();
5770 end_sequence ();
5771
5772 emit_insn_before (seq, insert_before);
5773 }
5774 \f
5775 /* Test whether A * B can be computed without
5776 an actual multiply insn. Value is 1 if so. */
5777
5778 static int
5779 product_cheap_p (a, b)
5780 rtx a;
5781 rtx b;
5782 {
5783 int i;
5784 rtx tmp;
5785 struct obstack *old_rtl_obstack = rtl_obstack;
5786 char *storage = (char *) obstack_alloc (&temp_obstack, 0);
5787 int win = 1;
5788
5789 /* If only one is constant, make it B. */
5790 if (GET_CODE (a) == CONST_INT)
5791 tmp = a, a = b, b = tmp;
5792
5793 /* If first constant, both constant, so don't need multiply. */
5794 if (GET_CODE (a) == CONST_INT)
5795 return 1;
5796
5797 /* If second not constant, neither is constant, so would need multiply. */
5798 if (GET_CODE (b) != CONST_INT)
5799 return 0;
5800
5801 /* One operand is constant, so might not need multiply insn. Generate the
5802 code for the multiply and see if a call or multiply, or long sequence
5803 of insns is generated. */
5804
5805 rtl_obstack = &temp_obstack;
5806 start_sequence ();
5807 expand_mult (GET_MODE (a), a, b, NULL_RTX, 0);
5808 tmp = gen_sequence ();
5809 end_sequence ();
5810
5811 if (GET_CODE (tmp) == SEQUENCE)
5812 {
5813 if (XVEC (tmp, 0) == 0)
5814 win = 1;
5815 else if (XVECLEN (tmp, 0) > 3)
5816 win = 0;
5817 else
5818 for (i = 0; i < XVECLEN (tmp, 0); i++)
5819 {
5820 rtx insn = XVECEXP (tmp, 0, i);
5821
5822 if (GET_CODE (insn) != INSN
5823 || (GET_CODE (PATTERN (insn)) == SET
5824 && GET_CODE (SET_SRC (PATTERN (insn))) == MULT)
5825 || (GET_CODE (PATTERN (insn)) == PARALLEL
5826 && GET_CODE (XVECEXP (PATTERN (insn), 0, 0)) == SET
5827 && GET_CODE (SET_SRC (XVECEXP (PATTERN (insn), 0, 0))) == MULT))
5828 {
5829 win = 0;
5830 break;
5831 }
5832 }
5833 }
5834 else if (GET_CODE (tmp) == SET
5835 && GET_CODE (SET_SRC (tmp)) == MULT)
5836 win = 0;
5837 else if (GET_CODE (tmp) == PARALLEL
5838 && GET_CODE (XVECEXP (tmp, 0, 0)) == SET
5839 && GET_CODE (SET_SRC (XVECEXP (tmp, 0, 0))) == MULT)
5840 win = 0;
5841
5842 /* Free any storage we obtained in generating this multiply and restore rtl
5843 allocation to its normal obstack. */
5844 obstack_free (&temp_obstack, storage);
5845 rtl_obstack = old_rtl_obstack;
5846
5847 return win;
5848 }
5849 \f
5850 /* Check to see if loop can be terminated by a "decrement and branch until
5851 zero" instruction. If so, add a REG_NONNEG note to the branch insn if so.
5852 Also try reversing an increment loop to a decrement loop
5853 to see if the optimization can be performed.
5854 Value is nonzero if optimization was performed. */
5855
5856 /* This is useful even if the architecture doesn't have such an insn,
5857 because it might change a loops which increments from 0 to n to a loop
5858 which decrements from n to 0. A loop that decrements to zero is usually
5859 faster than one that increments from zero. */
5860
5861 /* ??? This could be rewritten to use some of the loop unrolling procedures,
5862 such as approx_final_value, biv_total_increment, loop_iterations, and
5863 final_[bg]iv_value. */
5864
5865 static int
5866 check_dbra_loop (loop_end, insn_count, loop_start)
5867 rtx loop_end;
5868 int insn_count;
5869 rtx loop_start;
5870 {
5871 struct iv_class *bl;
5872 rtx reg;
5873 rtx jump_label;
5874 rtx final_value;
5875 rtx start_value;
5876 rtx new_add_val;
5877 rtx comparison;
5878 rtx before_comparison;
5879 rtx p;
5880
5881 /* If last insn is a conditional branch, and the insn before tests a
5882 register value, try to optimize it. Otherwise, we can't do anything. */
5883
5884 comparison = get_condition_for_loop (PREV_INSN (loop_end));
5885 if (comparison == 0)
5886 return 0;
5887
5888 /* Check all of the bivs to see if the compare uses one of them.
5889 Skip biv's set more than once because we can't guarantee that
5890 it will be zero on the last iteration. Also skip if the biv is
5891 used between its update and the test insn. */
5892
5893 for (bl = loop_iv_list; bl; bl = bl->next)
5894 {
5895 if (bl->biv_count == 1
5896 && bl->biv->dest_reg == XEXP (comparison, 0)
5897 && ! reg_used_between_p (regno_reg_rtx[bl->regno], bl->biv->insn,
5898 PREV_INSN (PREV_INSN (loop_end))))
5899 break;
5900 }
5901
5902 if (! bl)
5903 return 0;
5904
5905 /* Look for the case where the basic induction variable is always
5906 nonnegative, and equals zero on the last iteration.
5907 In this case, add a reg_note REG_NONNEG, which allows the
5908 m68k DBRA instruction to be used. */
5909
5910 if (((GET_CODE (comparison) == GT
5911 && GET_CODE (XEXP (comparison, 1)) == CONST_INT
5912 && INTVAL (XEXP (comparison, 1)) == -1)
5913 || (GET_CODE (comparison) == NE && XEXP (comparison, 1) == const0_rtx))
5914 && GET_CODE (bl->biv->add_val) == CONST_INT
5915 && INTVAL (bl->biv->add_val) < 0)
5916 {
5917 /* Initial value must be greater than 0,
5918 init_val % -dec_value == 0 to ensure that it equals zero on
5919 the last iteration */
5920
5921 if (GET_CODE (bl->initial_value) == CONST_INT
5922 && INTVAL (bl->initial_value) > 0
5923 && (INTVAL (bl->initial_value)
5924 % (-INTVAL (bl->biv->add_val))) == 0)
5925 {
5926 /* register always nonnegative, add REG_NOTE to branch */
5927 REG_NOTES (PREV_INSN (loop_end))
5928 = gen_rtx (EXPR_LIST, REG_NONNEG, NULL_RTX,
5929 REG_NOTES (PREV_INSN (loop_end)));
5930 bl->nonneg = 1;
5931
5932 return 1;
5933 }
5934
5935 /* If the decrement is 1 and the value was tested as >= 0 before
5936 the loop, then we can safely optimize. */
5937 for (p = loop_start; p; p = PREV_INSN (p))
5938 {
5939 if (GET_CODE (p) == CODE_LABEL)
5940 break;
5941 if (GET_CODE (p) != JUMP_INSN)
5942 continue;
5943
5944 before_comparison = get_condition_for_loop (p);
5945 if (before_comparison
5946 && XEXP (before_comparison, 0) == bl->biv->dest_reg
5947 && GET_CODE (before_comparison) == LT
5948 && XEXP (before_comparison, 1) == const0_rtx
5949 && ! reg_set_between_p (bl->biv->dest_reg, p, loop_start)
5950 && INTVAL (bl->biv->add_val) == -1)
5951 {
5952 REG_NOTES (PREV_INSN (loop_end))
5953 = gen_rtx (EXPR_LIST, REG_NONNEG, NULL_RTX,
5954 REG_NOTES (PREV_INSN (loop_end)));
5955 bl->nonneg = 1;
5956
5957 return 1;
5958 }
5959 }
5960 }
5961 else if (num_mem_sets <= 1)
5962 {
5963 /* Try to change inc to dec, so can apply above optimization. */
5964 /* Can do this if:
5965 all registers modified are induction variables or invariant,
5966 all memory references have non-overlapping addresses
5967 (obviously true if only one write)
5968 allow 2 insns for the compare/jump at the end of the loop. */
5969 /* Also, we must avoid any instructions which use both the reversed
5970 biv and another biv. Such instructions will fail if the loop is
5971 reversed. We meet this condition by requiring that either
5972 no_use_except_counting is true, or else that there is only
5973 one biv. */
5974 int num_nonfixed_reads = 0;
5975 /* 1 if the iteration var is used only to count iterations. */
5976 int no_use_except_counting = 0;
5977 /* 1 if the loop has no memory store, or it has a single memory store
5978 which is reversible. */
5979 int reversible_mem_store = 1;
5980
5981 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
5982 if (GET_RTX_CLASS (GET_CODE (p)) == 'i')
5983 num_nonfixed_reads += count_nonfixed_reads (PATTERN (p));
5984
5985 if (bl->giv_count == 0
5986 && ! loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]])
5987 {
5988 rtx bivreg = regno_reg_rtx[bl->regno];
5989
5990 /* If there are no givs for this biv, and the only exit is the
5991 fall through at the end of the the loop, then
5992 see if perhaps there are no uses except to count. */
5993 no_use_except_counting = 1;
5994 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
5995 if (GET_RTX_CLASS (GET_CODE (p)) == 'i')
5996 {
5997 rtx set = single_set (p);
5998
5999 if (set && GET_CODE (SET_DEST (set)) == REG
6000 && REGNO (SET_DEST (set)) == bl->regno)
6001 /* An insn that sets the biv is okay. */
6002 ;
6003 else if (p == prev_nonnote_insn (prev_nonnote_insn (loop_end))
6004 || p == prev_nonnote_insn (loop_end))
6005 /* Don't bother about the end test. */
6006 ;
6007 else if (reg_mentioned_p (bivreg, PATTERN (p)))
6008 /* Any other use of the biv is no good. */
6009 {
6010 no_use_except_counting = 0;
6011 break;
6012 }
6013 }
6014 }
6015
6016 /* If the loop has a single store, and the destination address is
6017 invariant, then we can't reverse the loop, because this address
6018 might then have the wrong value at loop exit.
6019 This would work if the source was invariant also, however, in that
6020 case, the insn should have been moved out of the loop. */
6021
6022 if (num_mem_sets == 1)
6023 reversible_mem_store
6024 = (! unknown_address_altered
6025 && ! invariant_p (XEXP (loop_store_mems[0], 0)));
6026
6027 /* This code only acts for innermost loops. Also it simplifies
6028 the memory address check by only reversing loops with
6029 zero or one memory access.
6030 Two memory accesses could involve parts of the same array,
6031 and that can't be reversed. */
6032
6033 if (num_nonfixed_reads <= 1
6034 && !loop_has_call
6035 && !loop_has_volatile
6036 && reversible_mem_store
6037 && (no_use_except_counting
6038 || ((bl->giv_count + bl->biv_count + num_mem_sets
6039 + num_movables + 2 == insn_count)
6040 && (bl == loop_iv_list && bl->next == 0))))
6041 {
6042 rtx tem;
6043
6044 /* Loop can be reversed. */
6045 if (loop_dump_stream)
6046 fprintf (loop_dump_stream, "Can reverse loop\n");
6047
6048 /* Now check other conditions:
6049 initial_value must be zero,
6050 final_value % add_val == 0, so that when reversed, the
6051 biv will be zero on the last iteration.
6052
6053 This test can probably be improved since +/- 1 in the constant
6054 can be obtained by changing LT to LE and vice versa; this is
6055 confusing. */
6056
6057 if (comparison && bl->initial_value == const0_rtx
6058 && GET_CODE (XEXP (comparison, 1)) == CONST_INT
6059 /* LE gets turned into LT */
6060 && GET_CODE (comparison) == LT
6061 && (INTVAL (XEXP (comparison, 1))
6062 % INTVAL (bl->biv->add_val)) == 0)
6063 {
6064 /* Register will always be nonnegative, with value
6065 0 on last iteration if loop reversed */
6066
6067 /* Save some info needed to produce the new insns. */
6068 reg = bl->biv->dest_reg;
6069 jump_label = XEXP (SET_SRC (PATTERN (PREV_INSN (loop_end))), 1);
6070 if (jump_label == pc_rtx)
6071 jump_label = XEXP (SET_SRC (PATTERN (PREV_INSN (loop_end))), 2);
6072 new_add_val = GEN_INT (- INTVAL (bl->biv->add_val));
6073
6074 final_value = XEXP (comparison, 1);
6075 start_value = GEN_INT (INTVAL (XEXP (comparison, 1))
6076 - INTVAL (bl->biv->add_val));
6077
6078 /* Initialize biv to start_value before loop start.
6079 The old initializing insn will be deleted as a
6080 dead store by flow.c. */
6081 emit_insn_before (gen_move_insn (reg, start_value), loop_start);
6082
6083 /* Add insn to decrement register, and delete insn
6084 that incremented the register. */
6085 p = emit_insn_before (gen_add2_insn (reg, new_add_val),
6086 bl->biv->insn);
6087 delete_insn (bl->biv->insn);
6088
6089 /* Update biv info to reflect its new status. */
6090 bl->biv->insn = p;
6091 bl->initial_value = start_value;
6092 bl->biv->add_val = new_add_val;
6093
6094 /* Inc LABEL_NUSES so that delete_insn will
6095 not delete the label. */
6096 LABEL_NUSES (XEXP (jump_label, 0)) ++;
6097
6098 /* Emit an insn after the end of the loop to set the biv's
6099 proper exit value if it is used anywhere outside the loop. */
6100 if ((REGNO_LAST_UID (bl->regno)
6101 != INSN_UID (PREV_INSN (PREV_INSN (loop_end))))
6102 || ! bl->init_insn
6103 || REGNO_FIRST_UID (bl->regno) != INSN_UID (bl->init_insn))
6104 emit_insn_after (gen_move_insn (reg, final_value),
6105 loop_end);
6106
6107 /* Delete compare/branch at end of loop. */
6108 delete_insn (PREV_INSN (loop_end));
6109 delete_insn (PREV_INSN (loop_end));
6110
6111 /* Add new compare/branch insn at end of loop. */
6112 start_sequence ();
6113 emit_cmp_insn (reg, const0_rtx, GE, NULL_RTX,
6114 GET_MODE (reg), 0, 0);
6115 emit_jump_insn (gen_bge (XEXP (jump_label, 0)));
6116 tem = gen_sequence ();
6117 end_sequence ();
6118 emit_jump_insn_before (tem, loop_end);
6119
6120 for (tem = PREV_INSN (loop_end);
6121 tem && GET_CODE (tem) != JUMP_INSN; tem = PREV_INSN (tem))
6122 ;
6123 if (tem)
6124 {
6125 JUMP_LABEL (tem) = XEXP (jump_label, 0);
6126
6127 /* Increment of LABEL_NUSES done above. */
6128 /* Register is now always nonnegative,
6129 so add REG_NONNEG note to the branch. */
6130 REG_NOTES (tem) = gen_rtx (EXPR_LIST, REG_NONNEG, NULL_RTX,
6131 REG_NOTES (tem));
6132 }
6133
6134 bl->nonneg = 1;
6135
6136 /* Mark that this biv has been reversed. Each giv which depends
6137 on this biv, and which is also live past the end of the loop
6138 will have to be fixed up. */
6139
6140 bl->reversed = 1;
6141
6142 if (loop_dump_stream)
6143 fprintf (loop_dump_stream,
6144 "Reversed loop and added reg_nonneg\n");
6145
6146 return 1;
6147 }
6148 }
6149 }
6150
6151 return 0;
6152 }
6153 \f
6154 /* Verify whether the biv BL appears to be eliminable,
6155 based on the insns in the loop that refer to it.
6156 LOOP_START is the first insn of the loop, and END is the end insn.
6157
6158 If ELIMINATE_P is non-zero, actually do the elimination.
6159
6160 THRESHOLD and INSN_COUNT are from loop_optimize and are used to
6161 determine whether invariant insns should be placed inside or at the
6162 start of the loop. */
6163
6164 static int
6165 maybe_eliminate_biv (bl, loop_start, end, eliminate_p, threshold, insn_count)
6166 struct iv_class *bl;
6167 rtx loop_start;
6168 rtx end;
6169 int eliminate_p;
6170 int threshold, insn_count;
6171 {
6172 rtx reg = bl->biv->dest_reg;
6173 rtx p;
6174
6175 /* Scan all insns in the loop, stopping if we find one that uses the
6176 biv in a way that we cannot eliminate. */
6177
6178 for (p = loop_start; p != end; p = NEXT_INSN (p))
6179 {
6180 enum rtx_code code = GET_CODE (p);
6181 rtx where = threshold >= insn_count ? loop_start : p;
6182
6183 if ((code == INSN || code == JUMP_INSN || code == CALL_INSN)
6184 && reg_mentioned_p (reg, PATTERN (p))
6185 && ! maybe_eliminate_biv_1 (PATTERN (p), p, bl, eliminate_p, where))
6186 {
6187 if (loop_dump_stream)
6188 fprintf (loop_dump_stream,
6189 "Cannot eliminate biv %d: biv used in insn %d.\n",
6190 bl->regno, INSN_UID (p));
6191 break;
6192 }
6193 }
6194
6195 if (p == end)
6196 {
6197 if (loop_dump_stream)
6198 fprintf (loop_dump_stream, "biv %d %s eliminated.\n",
6199 bl->regno, eliminate_p ? "was" : "can be");
6200 return 1;
6201 }
6202
6203 return 0;
6204 }
6205 \f
6206 /* If BL appears in X (part of the pattern of INSN), see if we can
6207 eliminate its use. If so, return 1. If not, return 0.
6208
6209 If BIV does not appear in X, return 1.
6210
6211 If ELIMINATE_P is non-zero, actually do the elimination. WHERE indicates
6212 where extra insns should be added. Depending on how many items have been
6213 moved out of the loop, it will either be before INSN or at the start of
6214 the loop. */
6215
6216 static int
6217 maybe_eliminate_biv_1 (x, insn, bl, eliminate_p, where)
6218 rtx x, insn;
6219 struct iv_class *bl;
6220 int eliminate_p;
6221 rtx where;
6222 {
6223 enum rtx_code code = GET_CODE (x);
6224 rtx reg = bl->biv->dest_reg;
6225 enum machine_mode mode = GET_MODE (reg);
6226 struct induction *v;
6227 rtx arg, new, tem;
6228 int arg_operand;
6229 char *fmt;
6230 int i, j;
6231
6232 switch (code)
6233 {
6234 case REG:
6235 /* If we haven't already been able to do something with this BIV,
6236 we can't eliminate it. */
6237 if (x == reg)
6238 return 0;
6239 return 1;
6240
6241 case SET:
6242 /* If this sets the BIV, it is not a problem. */
6243 if (SET_DEST (x) == reg)
6244 return 1;
6245
6246 /* If this is an insn that defines a giv, it is also ok because
6247 it will go away when the giv is reduced. */
6248 for (v = bl->giv; v; v = v->next_iv)
6249 if (v->giv_type == DEST_REG && SET_DEST (x) == v->dest_reg)
6250 return 1;
6251
6252 #ifdef HAVE_cc0
6253 if (SET_DEST (x) == cc0_rtx && SET_SRC (x) == reg)
6254 {
6255 /* Can replace with any giv that was reduced and
6256 that has (MULT_VAL != 0) and (ADD_VAL == 0).
6257 Require a constant for MULT_VAL, so we know it's nonzero.
6258 ??? We disable this optimization to avoid potential
6259 overflows. */
6260
6261 for (v = bl->giv; v; v = v->next_iv)
6262 if (CONSTANT_P (v->mult_val) && v->mult_val != const0_rtx
6263 && v->add_val == const0_rtx
6264 && ! v->ignore && ! v->maybe_dead && v->always_computable
6265 && v->mode == mode
6266 && 0)
6267 {
6268 /* If the giv V had the auto-inc address optimization applied
6269 to it, and INSN occurs between the giv insn and the biv
6270 insn, then we must adjust the value used here.
6271 This is rare, so we don't bother to do so. */
6272 if (v->auto_inc_opt
6273 && ((INSN_LUID (v->insn) < INSN_LUID (insn)
6274 && INSN_LUID (insn) < INSN_LUID (bl->biv->insn))
6275 || (INSN_LUID (v->insn) > INSN_LUID (insn)
6276 && INSN_LUID (insn) > INSN_LUID (bl->biv->insn))))
6277 continue;
6278
6279 if (! eliminate_p)
6280 return 1;
6281
6282 /* If the giv has the opposite direction of change,
6283 then reverse the comparison. */
6284 if (INTVAL (v->mult_val) < 0)
6285 new = gen_rtx (COMPARE, GET_MODE (v->new_reg),
6286 const0_rtx, v->new_reg);
6287 else
6288 new = v->new_reg;
6289
6290 /* We can probably test that giv's reduced reg. */
6291 if (validate_change (insn, &SET_SRC (x), new, 0))
6292 return 1;
6293 }
6294
6295 /* Look for a giv with (MULT_VAL != 0) and (ADD_VAL != 0);
6296 replace test insn with a compare insn (cmp REDUCED_GIV ADD_VAL).
6297 Require a constant for MULT_VAL, so we know it's nonzero.
6298 ??? Do this only if ADD_VAL is a pointer to avoid a potential
6299 overflow problem. */
6300
6301 for (v = bl->giv; v; v = v->next_iv)
6302 if (CONSTANT_P (v->mult_val) && v->mult_val != const0_rtx
6303 && ! v->ignore && ! v->maybe_dead && v->always_computable
6304 && v->mode == mode
6305 && (GET_CODE (v->add_val) == SYMBOL_REF
6306 || GET_CODE (v->add_val) == LABEL_REF
6307 || GET_CODE (v->add_val) == CONST
6308 || (GET_CODE (v->add_val) == REG
6309 && REGNO_POINTER_FLAG (REGNO (v->add_val)))))
6310 {
6311 /* If the giv V had the auto-inc address optimization applied
6312 to it, and INSN occurs between the giv insn and the biv
6313 insn, then we must adjust the value used here.
6314 This is rare, so we don't bother to do so. */
6315 if (v->auto_inc_opt
6316 && ((INSN_LUID (v->insn) < INSN_LUID (insn)
6317 && INSN_LUID (insn) < INSN_LUID (bl->biv->insn))
6318 || (INSN_LUID (v->insn) > INSN_LUID (insn)
6319 && INSN_LUID (insn) > INSN_LUID (bl->biv->insn))))
6320 continue;
6321
6322 if (! eliminate_p)
6323 return 1;
6324
6325 /* If the giv has the opposite direction of change,
6326 then reverse the comparison. */
6327 if (INTVAL (v->mult_val) < 0)
6328 new = gen_rtx (COMPARE, VOIDmode, copy_rtx (v->add_val),
6329 v->new_reg);
6330 else
6331 new = gen_rtx (COMPARE, VOIDmode, v->new_reg,
6332 copy_rtx (v->add_val));
6333
6334 /* Replace biv with the giv's reduced register. */
6335 update_reg_last_use (v->add_val, insn);
6336 if (validate_change (insn, &SET_SRC (PATTERN (insn)), new, 0))
6337 return 1;
6338
6339 /* Insn doesn't support that constant or invariant. Copy it
6340 into a register (it will be a loop invariant.) */
6341 tem = gen_reg_rtx (GET_MODE (v->new_reg));
6342
6343 emit_insn_before (gen_move_insn (tem, copy_rtx (v->add_val)),
6344 where);
6345
6346 /* Substitute the new register for its invariant value in
6347 the compare expression. */
6348 XEXP (new, (INTVAL (v->mult_val) < 0) ? 0 : 1) = tem;
6349 if (validate_change (insn, &SET_SRC (PATTERN (insn)), new, 0))
6350 return 1;
6351 }
6352 }
6353 #endif
6354 break;
6355
6356 case COMPARE:
6357 case EQ: case NE:
6358 case GT: case GE: case GTU: case GEU:
6359 case LT: case LE: case LTU: case LEU:
6360 /* See if either argument is the biv. */
6361 if (XEXP (x, 0) == reg)
6362 arg = XEXP (x, 1), arg_operand = 1;
6363 else if (XEXP (x, 1) == reg)
6364 arg = XEXP (x, 0), arg_operand = 0;
6365 else
6366 break;
6367
6368 if (CONSTANT_P (arg))
6369 {
6370 /* First try to replace with any giv that has constant positive
6371 mult_val and constant add_val. We might be able to support
6372 negative mult_val, but it seems complex to do it in general. */
6373
6374 for (v = bl->giv; v; v = v->next_iv)
6375 if (CONSTANT_P (v->mult_val) && INTVAL (v->mult_val) > 0
6376 && (GET_CODE (v->add_val) == SYMBOL_REF
6377 || GET_CODE (v->add_val) == LABEL_REF
6378 || GET_CODE (v->add_val) == CONST
6379 || (GET_CODE (v->add_val) == REG
6380 && REGNO_POINTER_FLAG (REGNO (v->add_val))))
6381 && ! v->ignore && ! v->maybe_dead && v->always_computable
6382 && v->mode == mode)
6383 {
6384 /* If the giv V had the auto-inc address optimization applied
6385 to it, and INSN occurs between the giv insn and the biv
6386 insn, then we must adjust the value used here.
6387 This is rare, so we don't bother to do so. */
6388 if (v->auto_inc_opt
6389 && ((INSN_LUID (v->insn) < INSN_LUID (insn)
6390 && INSN_LUID (insn) < INSN_LUID (bl->biv->insn))
6391 || (INSN_LUID (v->insn) > INSN_LUID (insn)
6392 && INSN_LUID (insn) > INSN_LUID (bl->biv->insn))))
6393 continue;
6394
6395 if (! eliminate_p)
6396 return 1;
6397
6398 /* Replace biv with the giv's reduced reg. */
6399 XEXP (x, 1-arg_operand) = v->new_reg;
6400
6401 /* If all constants are actually constant integers and
6402 the derived constant can be directly placed in the COMPARE,
6403 do so. */
6404 if (GET_CODE (arg) == CONST_INT
6405 && GET_CODE (v->mult_val) == CONST_INT
6406 && GET_CODE (v->add_val) == CONST_INT
6407 && validate_change (insn, &XEXP (x, arg_operand),
6408 GEN_INT (INTVAL (arg)
6409 * INTVAL (v->mult_val)
6410 + INTVAL (v->add_val)), 0))
6411 return 1;
6412
6413 /* Otherwise, load it into a register. */
6414 tem = gen_reg_rtx (mode);
6415 emit_iv_add_mult (arg, v->mult_val, v->add_val, tem, where);
6416 if (validate_change (insn, &XEXP (x, arg_operand), tem, 0))
6417 return 1;
6418
6419 /* If that failed, put back the change we made above. */
6420 XEXP (x, 1-arg_operand) = reg;
6421 }
6422
6423 /* Look for giv with positive constant mult_val and nonconst add_val.
6424 Insert insns to calculate new compare value.
6425 ??? Turn this off due to possible overflow. */
6426
6427 for (v = bl->giv; v; v = v->next_iv)
6428 if (CONSTANT_P (v->mult_val) && INTVAL (v->mult_val) > 0
6429 && ! v->ignore && ! v->maybe_dead && v->always_computable
6430 && v->mode == mode
6431 && 0)
6432 {
6433 rtx tem;
6434
6435 /* If the giv V had the auto-inc address optimization applied
6436 to it, and INSN occurs between the giv insn and the biv
6437 insn, then we must adjust the value used here.
6438 This is rare, so we don't bother to do so. */
6439 if (v->auto_inc_opt
6440 && ((INSN_LUID (v->insn) < INSN_LUID (insn)
6441 && INSN_LUID (insn) < INSN_LUID (bl->biv->insn))
6442 || (INSN_LUID (v->insn) > INSN_LUID (insn)
6443 && INSN_LUID (insn) > INSN_LUID (bl->biv->insn))))
6444 continue;
6445
6446 if (! eliminate_p)
6447 return 1;
6448
6449 tem = gen_reg_rtx (mode);
6450
6451 /* Replace biv with giv's reduced register. */
6452 validate_change (insn, &XEXP (x, 1 - arg_operand),
6453 v->new_reg, 1);
6454
6455 /* Compute value to compare against. */
6456 emit_iv_add_mult (arg, v->mult_val, v->add_val, tem, where);
6457 /* Use it in this insn. */
6458 validate_change (insn, &XEXP (x, arg_operand), tem, 1);
6459 if (apply_change_group ())
6460 return 1;
6461 }
6462 }
6463 else if (GET_CODE (arg) == REG || GET_CODE (arg) == MEM)
6464 {
6465 if (invariant_p (arg) == 1)
6466 {
6467 /* Look for giv with constant positive mult_val and nonconst
6468 add_val. Insert insns to compute new compare value.
6469 ??? Turn this off due to possible overflow. */
6470
6471 for (v = bl->giv; v; v = v->next_iv)
6472 if (CONSTANT_P (v->mult_val) && INTVAL (v->mult_val) > 0
6473 && ! v->ignore && ! v->maybe_dead && v->always_computable
6474 && v->mode == mode
6475 && 0)
6476 {
6477 rtx tem;
6478
6479 /* If the giv V had the auto-inc address optimization applied
6480 to it, and INSN occurs between the giv insn and the biv
6481 insn, then we must adjust the value used here.
6482 This is rare, so we don't bother to do so. */
6483 if (v->auto_inc_opt
6484 && ((INSN_LUID (v->insn) < INSN_LUID (insn)
6485 && INSN_LUID (insn) < INSN_LUID (bl->biv->insn))
6486 || (INSN_LUID (v->insn) > INSN_LUID (insn)
6487 && INSN_LUID (insn) > INSN_LUID (bl->biv->insn))))
6488 continue;
6489
6490 if (! eliminate_p)
6491 return 1;
6492
6493 tem = gen_reg_rtx (mode);
6494
6495 /* Replace biv with giv's reduced register. */
6496 validate_change (insn, &XEXP (x, 1 - arg_operand),
6497 v->new_reg, 1);
6498
6499 /* Compute value to compare against. */
6500 emit_iv_add_mult (arg, v->mult_val, v->add_val,
6501 tem, where);
6502 validate_change (insn, &XEXP (x, arg_operand), tem, 1);
6503 if (apply_change_group ())
6504 return 1;
6505 }
6506 }
6507
6508 /* This code has problems. Basically, you can't know when
6509 seeing if we will eliminate BL, whether a particular giv
6510 of ARG will be reduced. If it isn't going to be reduced,
6511 we can't eliminate BL. We can try forcing it to be reduced,
6512 but that can generate poor code.
6513
6514 The problem is that the benefit of reducing TV, below should
6515 be increased if BL can actually be eliminated, but this means
6516 we might have to do a topological sort of the order in which
6517 we try to process biv. It doesn't seem worthwhile to do
6518 this sort of thing now. */
6519
6520 #if 0
6521 /* Otherwise the reg compared with had better be a biv. */
6522 if (GET_CODE (arg) != REG
6523 || reg_iv_type[REGNO (arg)] != BASIC_INDUCT)
6524 return 0;
6525
6526 /* Look for a pair of givs, one for each biv,
6527 with identical coefficients. */
6528 for (v = bl->giv; v; v = v->next_iv)
6529 {
6530 struct induction *tv;
6531
6532 if (v->ignore || v->maybe_dead || v->mode != mode)
6533 continue;
6534
6535 for (tv = reg_biv_class[REGNO (arg)]->giv; tv; tv = tv->next_iv)
6536 if (! tv->ignore && ! tv->maybe_dead
6537 && rtx_equal_p (tv->mult_val, v->mult_val)
6538 && rtx_equal_p (tv->add_val, v->add_val)
6539 && tv->mode == mode)
6540 {
6541 /* If the giv V had the auto-inc address optimization applied
6542 to it, and INSN occurs between the giv insn and the biv
6543 insn, then we must adjust the value used here.
6544 This is rare, so we don't bother to do so. */
6545 if (v->auto_inc_opt
6546 && ((INSN_LUID (v->insn) < INSN_LUID (insn)
6547 && INSN_LUID (insn) < INSN_LUID (bl->biv->insn))
6548 || (INSN_LUID (v->insn) > INSN_LUID (insn)
6549 && INSN_LUID (insn) > INSN_LUID (bl->biv->insn))))
6550 continue;
6551
6552 if (! eliminate_p)
6553 return 1;
6554
6555 /* Replace biv with its giv's reduced reg. */
6556 XEXP (x, 1-arg_operand) = v->new_reg;
6557 /* Replace other operand with the other giv's
6558 reduced reg. */
6559 XEXP (x, arg_operand) = tv->new_reg;
6560 return 1;
6561 }
6562 }
6563 #endif
6564 }
6565
6566 /* If we get here, the biv can't be eliminated. */
6567 return 0;
6568
6569 case MEM:
6570 /* If this address is a DEST_ADDR giv, it doesn't matter if the
6571 biv is used in it, since it will be replaced. */
6572 for (v = bl->giv; v; v = v->next_iv)
6573 if (v->giv_type == DEST_ADDR && v->location == &XEXP (x, 0))
6574 return 1;
6575 break;
6576 }
6577
6578 /* See if any subexpression fails elimination. */
6579 fmt = GET_RTX_FORMAT (code);
6580 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
6581 {
6582 switch (fmt[i])
6583 {
6584 case 'e':
6585 if (! maybe_eliminate_biv_1 (XEXP (x, i), insn, bl,
6586 eliminate_p, where))
6587 return 0;
6588 break;
6589
6590 case 'E':
6591 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
6592 if (! maybe_eliminate_biv_1 (XVECEXP (x, i, j), insn, bl,
6593 eliminate_p, where))
6594 return 0;
6595 break;
6596 }
6597 }
6598
6599 return 1;
6600 }
6601 \f
6602 /* Return nonzero if the last use of REG
6603 is in an insn following INSN in the same basic block. */
6604
6605 static int
6606 last_use_this_basic_block (reg, insn)
6607 rtx reg;
6608 rtx insn;
6609 {
6610 rtx n;
6611 for (n = insn;
6612 n && GET_CODE (n) != CODE_LABEL && GET_CODE (n) != JUMP_INSN;
6613 n = NEXT_INSN (n))
6614 {
6615 if (REGNO_LAST_UID (REGNO (reg)) == INSN_UID (n))
6616 return 1;
6617 }
6618 return 0;
6619 }
6620 \f
6621 /* Called via `note_stores' to record the initial value of a biv. Here we
6622 just record the location of the set and process it later. */
6623
6624 static void
6625 record_initial (dest, set)
6626 rtx dest;
6627 rtx set;
6628 {
6629 struct iv_class *bl;
6630
6631 if (GET_CODE (dest) != REG
6632 || REGNO (dest) >= max_reg_before_loop
6633 || reg_iv_type[REGNO (dest)] != BASIC_INDUCT)
6634 return;
6635
6636 bl = reg_biv_class[REGNO (dest)];
6637
6638 /* If this is the first set found, record it. */
6639 if (bl->init_insn == 0)
6640 {
6641 bl->init_insn = note_insn;
6642 bl->init_set = set;
6643 }
6644 }
6645 \f
6646 /* If any of the registers in X are "old" and currently have a last use earlier
6647 than INSN, update them to have a last use of INSN. Their actual last use
6648 will be the previous insn but it will not have a valid uid_luid so we can't
6649 use it. */
6650
6651 static void
6652 update_reg_last_use (x, insn)
6653 rtx x;
6654 rtx insn;
6655 {
6656 /* Check for the case where INSN does not have a valid luid. In this case,
6657 there is no need to modify the regno_last_uid, as this can only happen
6658 when code is inserted after the loop_end to set a pseudo's final value,
6659 and hence this insn will never be the last use of x. */
6660 if (GET_CODE (x) == REG && REGNO (x) < max_reg_before_loop
6661 && INSN_UID (insn) < max_uid_for_loop
6662 && uid_luid[REGNO_LAST_UID (REGNO (x))] < uid_luid[INSN_UID (insn)])
6663 REGNO_LAST_UID (REGNO (x)) = INSN_UID (insn);
6664 else
6665 {
6666 register int i, j;
6667 register char *fmt = GET_RTX_FORMAT (GET_CODE (x));
6668 for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
6669 {
6670 if (fmt[i] == 'e')
6671 update_reg_last_use (XEXP (x, i), insn);
6672 else if (fmt[i] == 'E')
6673 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
6674 update_reg_last_use (XVECEXP (x, i, j), insn);
6675 }
6676 }
6677 }
6678 \f
6679 /* Given a jump insn JUMP, return the condition that will cause it to branch
6680 to its JUMP_LABEL. If the condition cannot be understood, or is an
6681 inequality floating-point comparison which needs to be reversed, 0 will
6682 be returned.
6683
6684 If EARLIEST is non-zero, it is a pointer to a place where the earliest
6685 insn used in locating the condition was found. If a replacement test
6686 of the condition is desired, it should be placed in front of that
6687 insn and we will be sure that the inputs are still valid.
6688
6689 The condition will be returned in a canonical form to simplify testing by
6690 callers. Specifically:
6691
6692 (1) The code will always be a comparison operation (EQ, NE, GT, etc.).
6693 (2) Both operands will be machine operands; (cc0) will have been replaced.
6694 (3) If an operand is a constant, it will be the second operand.
6695 (4) (LE x const) will be replaced with (LT x <const+1>) and similarly
6696 for GE, GEU, and LEU. */
6697
6698 rtx
6699 get_condition (jump, earliest)
6700 rtx jump;
6701 rtx *earliest;
6702 {
6703 enum rtx_code code;
6704 rtx prev = jump;
6705 rtx set;
6706 rtx tem;
6707 rtx op0, op1;
6708 int reverse_code = 0;
6709 int did_reverse_condition = 0;
6710
6711 /* If this is not a standard conditional jump, we can't parse it. */
6712 if (GET_CODE (jump) != JUMP_INSN
6713 || ! condjump_p (jump) || simplejump_p (jump))
6714 return 0;
6715
6716 code = GET_CODE (XEXP (SET_SRC (PATTERN (jump)), 0));
6717 op0 = XEXP (XEXP (SET_SRC (PATTERN (jump)), 0), 0);
6718 op1 = XEXP (XEXP (SET_SRC (PATTERN (jump)), 0), 1);
6719
6720 if (earliest)
6721 *earliest = jump;
6722
6723 /* If this branches to JUMP_LABEL when the condition is false, reverse
6724 the condition. */
6725 if (GET_CODE (XEXP (SET_SRC (PATTERN (jump)), 2)) == LABEL_REF
6726 && XEXP (XEXP (SET_SRC (PATTERN (jump)), 2), 0) == JUMP_LABEL (jump))
6727 code = reverse_condition (code), did_reverse_condition ^= 1;
6728
6729 /* If we are comparing a register with zero, see if the register is set
6730 in the previous insn to a COMPARE or a comparison operation. Perform
6731 the same tests as a function of STORE_FLAG_VALUE as find_comparison_args
6732 in cse.c */
6733
6734 while (GET_RTX_CLASS (code) == '<' && op1 == CONST0_RTX (GET_MODE (op0)))
6735 {
6736 /* Set non-zero when we find something of interest. */
6737 rtx x = 0;
6738
6739 #ifdef HAVE_cc0
6740 /* If comparison with cc0, import actual comparison from compare
6741 insn. */
6742 if (op0 == cc0_rtx)
6743 {
6744 if ((prev = prev_nonnote_insn (prev)) == 0
6745 || GET_CODE (prev) != INSN
6746 || (set = single_set (prev)) == 0
6747 || SET_DEST (set) != cc0_rtx)
6748 return 0;
6749
6750 op0 = SET_SRC (set);
6751 op1 = CONST0_RTX (GET_MODE (op0));
6752 if (earliest)
6753 *earliest = prev;
6754 }
6755 #endif
6756
6757 /* If this is a COMPARE, pick up the two things being compared. */
6758 if (GET_CODE (op0) == COMPARE)
6759 {
6760 op1 = XEXP (op0, 1);
6761 op0 = XEXP (op0, 0);
6762 continue;
6763 }
6764 else if (GET_CODE (op0) != REG)
6765 break;
6766
6767 /* Go back to the previous insn. Stop if it is not an INSN. We also
6768 stop if it isn't a single set or if it has a REG_INC note because
6769 we don't want to bother dealing with it. */
6770
6771 if ((prev = prev_nonnote_insn (prev)) == 0
6772 || GET_CODE (prev) != INSN
6773 || FIND_REG_INC_NOTE (prev, 0)
6774 || (set = single_set (prev)) == 0)
6775 break;
6776
6777 /* If this is setting OP0, get what it sets it to if it looks
6778 relevant. */
6779 if (rtx_equal_p (SET_DEST (set), op0))
6780 {
6781 enum machine_mode inner_mode = GET_MODE (SET_SRC (set));
6782
6783 if ((GET_CODE (SET_SRC (set)) == COMPARE
6784 || (((code == NE
6785 || (code == LT
6786 && GET_MODE_CLASS (inner_mode) == MODE_INT
6787 && (GET_MODE_BITSIZE (inner_mode)
6788 <= HOST_BITS_PER_WIDE_INT)
6789 && (STORE_FLAG_VALUE
6790 & ((HOST_WIDE_INT) 1
6791 << (GET_MODE_BITSIZE (inner_mode) - 1))))
6792 #ifdef FLOAT_STORE_FLAG_VALUE
6793 || (code == LT
6794 && GET_MODE_CLASS (inner_mode) == MODE_FLOAT
6795 && FLOAT_STORE_FLAG_VALUE < 0)
6796 #endif
6797 ))
6798 && GET_RTX_CLASS (GET_CODE (SET_SRC (set))) == '<')))
6799 x = SET_SRC (set);
6800 else if (((code == EQ
6801 || (code == GE
6802 && (GET_MODE_BITSIZE (inner_mode)
6803 <= HOST_BITS_PER_WIDE_INT)
6804 && GET_MODE_CLASS (inner_mode) == MODE_INT
6805 && (STORE_FLAG_VALUE
6806 & ((HOST_WIDE_INT) 1
6807 << (GET_MODE_BITSIZE (inner_mode) - 1))))
6808 #ifdef FLOAT_STORE_FLAG_VALUE
6809 || (code == GE
6810 && GET_MODE_CLASS (inner_mode) == MODE_FLOAT
6811 && FLOAT_STORE_FLAG_VALUE < 0)
6812 #endif
6813 ))
6814 && GET_RTX_CLASS (GET_CODE (SET_SRC (set))) == '<')
6815 {
6816 /* We might have reversed a LT to get a GE here. But this wasn't
6817 actually the comparison of data, so we don't flag that we
6818 have had to reverse the condition. */
6819 did_reverse_condition ^= 1;
6820 reverse_code = 1;
6821 x = SET_SRC (set);
6822 }
6823 else
6824 break;
6825 }
6826
6827 else if (reg_set_p (op0, prev))
6828 /* If this sets OP0, but not directly, we have to give up. */
6829 break;
6830
6831 if (x)
6832 {
6833 if (GET_RTX_CLASS (GET_CODE (x)) == '<')
6834 code = GET_CODE (x);
6835 if (reverse_code)
6836 {
6837 code = reverse_condition (code);
6838 did_reverse_condition ^= 1;
6839 reverse_code = 0;
6840 }
6841
6842 op0 = XEXP (x, 0), op1 = XEXP (x, 1);
6843 if (earliest)
6844 *earliest = prev;
6845 }
6846 }
6847
6848 /* If constant is first, put it last. */
6849 if (CONSTANT_P (op0))
6850 code = swap_condition (code), tem = op0, op0 = op1, op1 = tem;
6851
6852 /* If OP0 is the result of a comparison, we weren't able to find what
6853 was really being compared, so fail. */
6854 if (GET_MODE_CLASS (GET_MODE (op0)) == MODE_CC)
6855 return 0;
6856
6857 /* Canonicalize any ordered comparison with integers involving equality
6858 if we can do computations in the relevant mode and we do not
6859 overflow. */
6860
6861 if (GET_CODE (op1) == CONST_INT
6862 && GET_MODE (op0) != VOIDmode
6863 && GET_MODE_BITSIZE (GET_MODE (op0)) <= HOST_BITS_PER_WIDE_INT)
6864 {
6865 HOST_WIDE_INT const_val = INTVAL (op1);
6866 unsigned HOST_WIDE_INT uconst_val = const_val;
6867 unsigned HOST_WIDE_INT max_val
6868 = (unsigned HOST_WIDE_INT) GET_MODE_MASK (GET_MODE (op0));
6869
6870 switch (code)
6871 {
6872 case LE:
6873 if (const_val != max_val >> 1)
6874 code = LT, op1 = GEN_INT (const_val + 1);
6875 break;
6876
6877 case GE:
6878 if (const_val
6879 != (((HOST_WIDE_INT) 1
6880 << (GET_MODE_BITSIZE (GET_MODE (op0)) - 1))))
6881 code = GT, op1 = GEN_INT (const_val - 1);
6882 break;
6883
6884 case LEU:
6885 if (uconst_val != max_val)
6886 code = LTU, op1 = GEN_INT (uconst_val + 1);
6887 break;
6888
6889 case GEU:
6890 if (uconst_val != 0)
6891 code = GTU, op1 = GEN_INT (uconst_val - 1);
6892 break;
6893 }
6894 }
6895
6896 /* If this was floating-point and we reversed anything other than an
6897 EQ or NE, return zero. */
6898 if (TARGET_FLOAT_FORMAT == IEEE_FLOAT_FORMAT
6899 && did_reverse_condition && code != NE && code != EQ
6900 && ! flag_fast_math
6901 && GET_MODE_CLASS (GET_MODE (op0)) == MODE_FLOAT)
6902 return 0;
6903
6904 #ifdef HAVE_cc0
6905 /* Never return CC0; return zero instead. */
6906 if (op0 == cc0_rtx)
6907 return 0;
6908 #endif
6909
6910 return gen_rtx (code, VOIDmode, op0, op1);
6911 }
6912
6913 /* Similar to above routine, except that we also put an invariant last
6914 unless both operands are invariants. */
6915
6916 rtx
6917 get_condition_for_loop (x)
6918 rtx x;
6919 {
6920 rtx comparison = get_condition (x, NULL_PTR);
6921
6922 if (comparison == 0
6923 || ! invariant_p (XEXP (comparison, 0))
6924 || invariant_p (XEXP (comparison, 1)))
6925 return comparison;
6926
6927 return gen_rtx (swap_condition (GET_CODE (comparison)), VOIDmode,
6928 XEXP (comparison, 1), XEXP (comparison, 0));
6929 }