]> gcc.gnu.org Git - gcc.git/blob - gcc/loop.c
*** empty log message ***
[gcc.git] / gcc / loop.c
1 /* Move constant computations out of loops.
2 Copyright (C) 1987, 1988, 1989, 1991 Free Software Foundation, Inc.
3
4 This file is part of GNU CC.
5
6 GNU CC is free software; you can redistribute it and/or modify
7 it under the terms of the GNU General Public License as published by
8 the Free Software Foundation; either version 2, or (at your option)
9 any later version.
10
11 GNU CC is distributed in the hope that it will be useful,
12 but WITHOUT ANY WARRANTY; without even the implied warranty of
13 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 GNU General Public License for more details.
15
16 You should have received a copy of the GNU General Public License
17 along with GNU CC; see the file COPYING. If not, write to
18 the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA. */
19
20
21 /* This is the loop optimization pass of the compiler.
22 It finds invariant computations within loops and moves them
23 to the beginning of the loop. Then it identifies basic and
24 general induction variables. Strength reduction is applied to the general
25 induction variables, and induction variable elimination is applied to
26 the basic induction variables.
27
28 It also finds cases where
29 a register is set within the loop by zero-extending a narrower value
30 and changes these to zero the entire register once before the loop
31 and merely copy the low part within the loop.
32
33 Most of the complexity is in heuristics to decide when it is worth
34 while to do these things. */
35
36 #include "config.h"
37 #include "rtl.h"
38 #include "obstack.h"
39 #include "expr.h"
40 #include "insn-config.h"
41 #include "insn-flags.h"
42 #include "regs.h"
43 #include "hard-reg-set.h"
44 #include "recog.h"
45 #include "flags.h"
46 #include "real.h"
47 #include <stdio.h>
48 #include "loop.h"
49
50 /* Vector mapping INSN_UIDs to luids.
51 The luids are like uids but increase monotonically always.
52 We use them to see whether a jump comes from outside a given loop. */
53
54 int *uid_luid;
55
56 /* Indexed by INSN_UID, contains the ordinal giving the (innermost) loop
57 number the insn is contained in. */
58
59 int *uid_loop_num;
60
61 /* 1 + largest uid of any insn. */
62
63 int max_uid_for_loop;
64
65 /* 1 + luid of last insn. */
66
67 static int max_luid;
68
69 /* Number of loops detected in current function. Used as index to the
70 next few tables. */
71
72 static int max_loop_num;
73
74 /* Indexed by loop number, contains the first and last insn of each loop. */
75
76 static rtx *loop_number_loop_starts, *loop_number_loop_ends;
77
78 /* For each loop, gives the containing loop number, -1 if none. */
79
80 int *loop_outer_loop;
81
82 /* Indexed by loop number, contains a nonzero value if the "loop" isn't
83 really a loop (an insn outside the loop branches into it). */
84
85 static char *loop_invalid;
86
87 /* Indexed by loop number, links together all LABEL_REFs which refer to
88 code labels outside the loop. Used by routines that need to know all
89 loop exits, such as final_biv_value and final_giv_value.
90
91 This does not include loop exits due to return instructions. This is
92 because all bivs and givs are pseudos, and hence must be dead after a
93 return, so the presense of a return does not affect any of the
94 optimizations that use this info. It is simpler to just not include return
95 instructions on this list. */
96
97 rtx *loop_number_exit_labels;
98
99 /* Holds the number of loop iterations. It is zero if the number could not be
100 calculated. Must be unsigned long since the number of iterations can
101 be as high as 2^31-1. For loops with a DImode iterator, this number will
102 will be zero if the number of loop iterations is too large for an
103 unsigned long to hold. */
104
105 unsigned long loop_n_iterations;
106
107 /* Nonzero if there is a subroutine call in the current loop.
108 (unknown_address_altered is also nonzero in this case.) */
109
110 static int loop_has_call;
111
112 /* Added loop_continue which is the NOTE_INSN_LOOP_CONT of the
113 current loop. A continue statement will generate a branch to
114 NEXT_INSN (loop_continue). */
115
116 static rtx loop_continue;
117
118 /* Indexed by register number, contains the number of times the reg
119 is set during the loop being scanned.
120 During code motion, a negative value indicates a reg that has been
121 made a candidate; in particular -2 means that it is an candidate that
122 we know is equal to a constant and -1 means that it is an candidate
123 not known equal to a constant.
124 After code motion, regs moved have 0 (which is accurate now)
125 while the failed candidates have the original number of times set.
126
127 Therefore, at all times, == 0 indicates an invariant register;
128 < 0 a conditionally invariant one. */
129
130 static short *n_times_set;
131
132 /* Original value of n_times_set; same except that this value
133 is not set negative for a reg whose sets have been made candidates
134 and not set to 0 for a reg that is moved. */
135
136 static short *n_times_used;
137
138 /* Index by register number, 1 indicates that the register
139 cannot be moved or strength reduced. */
140
141 static char *may_not_optimize;
142
143 /* Nonzero means reg N has already been moved out of one loop.
144 This reduces the desire to move it out of another. */
145
146 static char *moved_once;
147
148 /* Array of MEMs that are stored in this loop. If there are too many to fit
149 here, we just turn on unknown_address_altered. */
150
151 #define NUM_STORES 20
152 static rtx loop_store_mems[NUM_STORES];
153
154 /* Index of first available slot in above array. */
155 static int loop_store_mems_idx;
156
157 /* Nonzero if we don't know what MEMs were changed in the current loop.
158 This happens if the loop contains a call (in which call `loop_has_call'
159 will also be set) or if we store into more than NUM_STORES MEMs. */
160
161 static int unknown_address_altered;
162
163 /* Count of movable (i.e. invariant) instructions discovered in the loop. */
164 static int num_movables;
165
166 /* Count of memory write instructions discovered in the loop. */
167 static int num_mem_sets;
168
169 /* Number of loops contained within the current one, including itself. */
170 static int loops_enclosed;
171
172 /* Bound on pseudo register number before loop optimization.
173 A pseudo has valid regscan info if its number is < max_reg_before_loop. */
174 int max_reg_before_loop;
175
176 /* This obstack is used in product_cheap_p to allocate its rtl. It
177 may call gen_reg_rtx which, in turn, may reallocate regno_reg_rtx.
178 If we used the same obstack that it did, we would be deallocating
179 that array. */
180
181 static struct obstack temp_obstack;
182
183 /* This is where the pointer to the obstack being used for RTL is stored. */
184
185 extern struct obstack *rtl_obstack;
186
187 #define obstack_chunk_alloc xmalloc
188 #define obstack_chunk_free free
189
190 extern char *oballoc ();
191 extern int xmalloc ();
192 extern void free ();
193 \f
194 /* During the analysis of a loop, a chain of `struct movable's
195 is made to record all the movable insns found.
196 Then the entire chain can be scanned to decide which to move. */
197
198 struct movable
199 {
200 rtx insn; /* A movable insn */
201 rtx set_src; /* The expression this reg is set from. */
202 rtx set_dest; /* The destination of this SET. */
203 rtx dependencies; /* When INSN is libcall, this is an EXPR_LIST
204 of any registers used within the LIBCALL. */
205 int consec; /* Number of consecutive following insns
206 that must be moved with this one. */
207 int regno; /* The register it sets */
208 short lifetime; /* lifetime of that register;
209 may be adjusted when matching movables
210 that load the same value are found. */
211 short savings; /* Number of insns we can move for this reg,
212 including other movables that force this
213 or match this one. */
214 unsigned int cond : 1; /* 1 if only conditionally movable */
215 unsigned int force : 1; /* 1 means MUST move this insn */
216 unsigned int global : 1; /* 1 means reg is live outside this loop */
217 /* If PARTIAL is 1, GLOBAL means something different:
218 that the reg is live outside the range from where it is set
219 to the following label. */
220 unsigned int done : 1; /* 1 inhibits further processing of this */
221
222 unsigned int partial : 1; /* 1 means this reg is used for zero-extending.
223 In particular, moving it does not make it
224 invariant. */
225 unsigned int move_insn : 1; /* 1 means that we call emit_move_insn to
226 load SRC, rather than copying INSN. */
227 unsigned int is_equiv : 1; /* 1 means a REG_EQUIV is present on INSN. */
228 enum machine_mode savemode; /* Nonzero means it is a mode for a low part
229 that we should avoid changing when clearing
230 the rest of the reg. */
231 struct movable *match; /* First entry for same value */
232 struct movable *forces; /* An insn that must be moved if this is */
233 struct movable *next;
234 };
235
236 FILE *loop_dump_stream;
237
238 /* Forward declarations. */
239
240 static void find_and_verify_loops ();
241 static void mark_loop_jump ();
242 static void prescan_loop ();
243 static int reg_in_basic_block_p ();
244 static int consec_sets_invariant_p ();
245 static rtx libcall_other_reg ();
246 static int labels_in_range_p ();
247 static void count_loop_regs_set ();
248 static void note_addr_stored ();
249 static int loop_reg_used_before_p ();
250 static void scan_loop ();
251 static void replace_call_address ();
252 static rtx skip_consec_insns ();
253 static int libcall_benefit ();
254 static void ignore_some_movables ();
255 static void force_movables ();
256 static void combine_movables ();
257 static int rtx_equal_for_loop_p ();
258 static void move_movables ();
259 static void strength_reduce ();
260 static int valid_initial_value_p ();
261 static void find_mem_givs ();
262 static void record_biv ();
263 static void check_final_value ();
264 static void record_giv ();
265 static void update_giv_derive ();
266 static void delete_insn_forces ();
267 static int basic_induction_var ();
268 static rtx simplify_giv_expr ();
269 static int general_induction_var ();
270 static int consec_sets_giv ();
271 static int check_dbra_loop ();
272 static rtx express_from ();
273 static int combine_givs_p ();
274 static void combine_givs ();
275 static int product_cheap_p ();
276 static int maybe_eliminate_biv ();
277 static int maybe_eliminate_biv_1 ();
278 static int last_use_this_basic_block ();
279 static void record_initial ();
280 static void update_reg_last_use ();
281 \f
282 /* Relative gain of eliminating various kinds of operations. */
283 int add_cost;
284 #if 0
285 int shift_cost;
286 int mult_cost;
287 #endif
288
289 /* Benefit penalty, if a giv is not replaceable, i.e. must emit an insn to
290 copy the value of the strength reduced giv to its original register. */
291 int copy_cost;
292
293 void
294 init_loop ()
295 {
296 char *free_point = (char *) oballoc (1);
297 rtx reg = gen_rtx (REG, SImode, 0);
298 rtx pow2 = gen_rtx (CONST_INT, VOIDmode, 32);
299 rtx lea;
300 int i;
301
302 add_cost = rtx_cost (gen_rtx (PLUS, SImode, reg, reg), SET);
303
304 /* We multiply by 2 to reconcile the difference in scale between
305 these two ways of computing costs. Otherwise the cost of a copy
306 will be far less than the cost of an add. */
307 #ifdef REGISTER_MOVE_COST
308 copy_cost = REGISTER_MOVE_COST (GENERAL_REGS, GENERAL_REGS) * 2;
309 #else
310 copy_cost = 2 * 2;
311 #endif
312
313 /* Free the objects we just allocated. */
314 obfree (free_point);
315
316 /* Initialize the obstack used for rtl in product_cheap_p. */
317 gcc_obstack_init (&temp_obstack);
318 }
319 \f
320 /* Entry point of this file. Perform loop optimization
321 on the current function. F is the first insn of the function
322 and DUMPFILE is a stream for output of a trace of actions taken
323 (or 0 if none should be output). */
324
325 void
326 loop_optimize (f, dumpfile)
327 /* f is the first instruction of a chain of insns for one function */
328 rtx f;
329 FILE *dumpfile;
330 {
331 register rtx insn;
332 register int i;
333 rtx end;
334 rtx last_insn;
335
336 loop_dump_stream = dumpfile;
337
338 init_recog_no_volatile ();
339 init_alias_analysis ();
340
341 max_reg_before_loop = max_reg_num ();
342
343 moved_once = (char *) alloca (max_reg_before_loop);
344 bzero (moved_once, max_reg_before_loop);
345
346 regs_may_share = 0;
347
348 /* Count the number of loops. */
349
350 max_loop_num = 0;
351 for (insn = f; insn; insn = NEXT_INSN (insn))
352 {
353 if (GET_CODE (insn) == NOTE
354 && NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG)
355 max_loop_num++;
356 }
357
358 /* Don't waste time if no loops. */
359 if (max_loop_num == 0)
360 return;
361
362 /* Get size to use for tables indexed by uids.
363 Leave some space for labels allocated by find_and_verify_loops. */
364 max_uid_for_loop = get_max_uid () + 1 + max_loop_num * 4;
365
366 uid_luid = (int *) alloca (max_uid_for_loop * sizeof (int));
367 uid_loop_num = (int *) alloca (max_uid_for_loop * sizeof (int));
368
369 bzero (uid_luid, max_uid_for_loop * sizeof (int));
370 bzero (uid_loop_num, max_uid_for_loop * sizeof (int));
371
372 /* Allocate tables for recording each loop. We set each entry, so they need
373 not be zeroed. */
374 loop_number_loop_starts = (rtx *) alloca (max_loop_num * sizeof (rtx));
375 loop_number_loop_ends = (rtx *) alloca (max_loop_num * sizeof (rtx));
376 loop_outer_loop = (int *) alloca (max_loop_num * sizeof (int));
377 loop_invalid = (char *) alloca (max_loop_num * sizeof (char));
378 loop_number_exit_labels = (rtx *) alloca (max_loop_num * sizeof (rtx));
379
380 if (flag_unroll_loops && write_symbols != NO_DEBUG)
381 {
382 loop_number_first_block
383 = (union tree_node **) alloca (max_loop_num
384 * sizeof (union tree_node *));
385 loop_number_last_block
386 = (union tree_node **) alloca (max_loop_num
387 * sizeof (union tree_node *));
388 loop_number_block_level = (int *) alloca (max_loop_num * sizeof (int));
389 }
390
391 /* Find and process each loop.
392 First, find them, and record them in order of their beginnings. */
393 find_and_verify_loops (f);
394
395 /* Now find all register lifetimes. This must be done after
396 find_and_verify_loops, because it might reorder the insns in the
397 function. */
398 reg_scan (f, max_reg_num (), 1);
399
400 /* Compute the mapping from uids to luids.
401 LUIDs are numbers assigned to insns, like uids,
402 except that luids increase monotonically through the code.
403 Don't assign luids to line-number NOTEs, so that the distance in luids
404 between two insns is not affected by -g. */
405
406 for (insn = f, i = 0; insn; insn = NEXT_INSN (insn))
407 {
408 last_insn = insn;
409 if (GET_CODE (insn) != NOTE
410 || NOTE_LINE_NUMBER (insn) <= 0)
411 uid_luid[INSN_UID (insn)] = ++i;
412 else
413 /* Give a line number note the same luid as preceding insn. */
414 uid_luid[INSN_UID (insn)] = i;
415 }
416
417 max_luid = i + 1;
418
419 /* Don't leave gaps in uid_luid for insns that have been
420 deleted. It is possible that the first or last insn
421 using some register has been deleted by cross-jumping.
422 Make sure that uid_luid for that former insn's uid
423 points to the general area where that insn used to be. */
424 for (i = 0; i < max_uid_for_loop; i++)
425 {
426 uid_luid[0] = uid_luid[i];
427 if (uid_luid[0] != 0)
428 break;
429 }
430 for (i = 0; i < max_uid_for_loop; i++)
431 if (uid_luid[i] == 0)
432 uid_luid[i] = uid_luid[i - 1];
433
434 /* Create a mapping from loops to BLOCK tree nodes. */
435 if (flag_unroll_loops && write_symbols != NO_DEBUG)
436 find_loop_tree_blocks (f);
437
438 /* Now scan the loops, last ones first, since this means inner ones are done
439 before outer ones. */
440 for (i = max_loop_num-1; i >= 0; i--)
441 if (! loop_invalid[i] && loop_number_loop_ends[i])
442 scan_loop (loop_number_loop_starts[i], loop_number_loop_ends[i],
443 max_reg_num ());
444 }
445 \f
446 /* Optimize one loop whose start is LOOP_START and end is END.
447 LOOP_START is the NOTE_INSN_LOOP_BEG and END is the matching
448 NOTE_INSN_LOOP_END. */
449
450 /* ??? Could also move memory writes out of loops if the destination address
451 is invariant, the source is invariant, the memory write is not volatile,
452 and if we can prove that no read inside the loop can read this address
453 before the write occurs. If there is a read of this address after the
454 write, then we can also mark the memory read as invariant. */
455
456 static void
457 scan_loop (loop_start, end, nregs)
458 rtx loop_start, end;
459 int nregs;
460 {
461 register int i;
462 register rtx p;
463 /* 1 if we are scanning insns that could be executed zero times. */
464 int maybe_never = 0;
465 /* 1 if we are scanning insns that might never be executed
466 due to a subroutine call which might exit before they are reached. */
467 int call_passed = 0;
468 /* For a rotated loop that is entered near the bottom,
469 this is the label at the top. Otherwise it is zero. */
470 rtx loop_top = 0;
471 /* Jump insn that enters the loop, or 0 if control drops in. */
472 rtx loop_entry_jump = 0;
473 /* Place in the loop where control enters. */
474 rtx scan_start;
475 /* Number of insns in the loop. */
476 int insn_count;
477 int in_libcall = 0;
478 int tem;
479 rtx temp;
480 /* The SET from an insn, if it is the only SET in the insn. */
481 rtx set, set1;
482 /* Chain describing insns movable in current loop. */
483 struct movable *movables = 0;
484 /* Last element in `movables' -- so we can add elements at the end. */
485 struct movable *last_movable = 0;
486 /* Ratio of extra register life span we can justify
487 for saving an instruction. More if loop doesn't call subroutines
488 since in that case saving an insn makes more difference
489 and more registers are available. */
490 int threshold;
491 /* If we have calls, contains the insn in which a register was used
492 if it was used exactly once; contains const0_rtx if it was used more
493 than once. */
494 rtx *reg_single_usage = 0;
495
496 n_times_set = (short *) alloca (nregs * sizeof (short));
497 n_times_used = (short *) alloca (nregs * sizeof (short));
498 may_not_optimize = (char *) alloca (nregs);
499
500 /* Determine whether this loop starts with a jump down to a test at
501 the end. This will occur for a small number of loops with a test
502 that is too complex to duplicate in front of the loop.
503
504 We search for the first insn or label in the loop, skipping NOTEs.
505 However, we must be careful not to skip past a NOTE_INSN_LOOP_BEG
506 (because we might have a loop executed only once that contains a
507 loop which starts with a jump to its exit test) or a NOTE_INSN_LOOP_END
508 (in case we have a degenerate loop).
509
510 Note that if we mistakenly think that a loop is entered at the top
511 when, in fact, it is entered at the exit test, the only effect will be
512 slightly poorer optimization. Making the opposite error can generate
513 incorrect code. Since very few loops now start with a jump to the
514 exit test, the code here to detect that case is very conservative. */
515
516 for (p = NEXT_INSN (loop_start);
517 p != end
518 && GET_CODE (p) != CODE_LABEL && GET_RTX_CLASS (GET_CODE (p)) != 'i'
519 && (GET_CODE (p) != NOTE
520 || (NOTE_LINE_NUMBER (p) != NOTE_INSN_LOOP_BEG
521 && NOTE_LINE_NUMBER (p) != NOTE_INSN_LOOP_END));
522 p = NEXT_INSN (p))
523 ;
524
525 scan_start = p;
526
527 /* Set up variables describing this loop. */
528 prescan_loop (loop_start, end);
529 threshold = (loop_has_call ? 1 : 2) * (1 + n_non_fixed_regs);
530
531 /* If loop has a jump before the first label,
532 the true entry is the target of that jump.
533 Start scan from there.
534 But record in LOOP_TOP the place where the end-test jumps
535 back to so we can scan that after the end of the loop. */
536 if (GET_CODE (p) == JUMP_INSN)
537 {
538 loop_entry_jump = p;
539
540 /* Loop entry must be unconditional jump (and not a RETURN) */
541 if (simplejump_p (p)
542 && JUMP_LABEL (p) != 0
543 /* Check to see whether the jump actually
544 jumps out of the loop (meaning it's no loop).
545 This case can happen for things like
546 do {..} while (0). If this label was generated previously
547 by loop, we can't tell anything about it and have to reject
548 the loop. */
549 && INSN_UID (JUMP_LABEL (p)) < max_uid_for_loop
550 && INSN_LUID (JUMP_LABEL (p)) >= INSN_LUID (loop_start)
551 && INSN_LUID (JUMP_LABEL (p)) < INSN_LUID (end))
552 {
553 loop_top = next_label (scan_start);
554 scan_start = JUMP_LABEL (p);
555 }
556 }
557
558 /* If SCAN_START was an insn created by loop, we don't know its luid
559 as required by loop_reg_used_before_p. So skip such loops. (This
560 test may never be true, but it's best to play it safe.)
561
562 Also, skip loops where we do not start scanning at a label. This
563 test also rejects loops starting with a JUMP_INSN that failed the
564 test above. */
565
566 if (INSN_UID (scan_start) >= max_uid_for_loop
567 || GET_CODE (scan_start) != CODE_LABEL)
568 {
569 if (loop_dump_stream)
570 fprintf (loop_dump_stream, "\nLoop from %d to %d is phony.\n\n",
571 INSN_UID (loop_start), INSN_UID (end));
572 return;
573 }
574
575 /* Count number of times each reg is set during this loop.
576 Set may_not_optimize[I] if it is not safe to move out
577 the setting of register I. If this loop has calls, set
578 reg_single_usage[I]. */
579
580 bzero (n_times_set, nregs * sizeof (short));
581 bzero (may_not_optimize, nregs);
582
583 if (loop_has_call)
584 {
585 reg_single_usage = (rtx *) alloca (nregs * sizeof (rtx));
586 bzero (reg_single_usage, nregs * sizeof (rtx));
587 }
588
589 count_loop_regs_set (loop_top ? loop_top : loop_start, end,
590 may_not_optimize, reg_single_usage, &insn_count, nregs);
591
592 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
593 may_not_optimize[i] = 1, n_times_set[i] = 1;
594 bcopy (n_times_set, n_times_used, nregs * sizeof (short));
595
596 if (loop_dump_stream)
597 {
598 fprintf (loop_dump_stream, "\nLoop from %d to %d: %d real insns.\n",
599 INSN_UID (loop_start), INSN_UID (end), insn_count);
600 if (loop_continue)
601 fprintf (loop_dump_stream, "Continue at insn %d.\n",
602 INSN_UID (loop_continue));
603 }
604
605 /* Scan through the loop finding insns that are safe to move.
606 Set n_times_set negative for the reg being set, so that
607 this reg will be considered invariant for subsequent insns.
608 We consider whether subsequent insns use the reg
609 in deciding whether it is worth actually moving.
610
611 MAYBE_NEVER is nonzero if we have passed a conditional jump insn
612 and therefore it is possible that the insns we are scanning
613 would never be executed. At such times, we must make sure
614 that it is safe to execute the insn once instead of zero times.
615 When MAYBE_NEVER is 0, all insns will be executed at least once
616 so that is not a problem. */
617
618 p = scan_start;
619 while (1)
620 {
621 p = NEXT_INSN (p);
622 /* At end of a straight-in loop, we are done.
623 At end of a loop entered at the bottom, scan the top. */
624 if (p == scan_start)
625 break;
626 if (p == end)
627 {
628 if (loop_top != 0)
629 p = NEXT_INSN (loop_top);
630 else
631 break;
632 if (p == scan_start)
633 break;
634 }
635
636 if (GET_RTX_CLASS (GET_CODE (p)) == 'i'
637 && find_reg_note (p, REG_LIBCALL, 0))
638 in_libcall = 1;
639 else if (GET_RTX_CLASS (GET_CODE (p)) == 'i'
640 && find_reg_note (p, REG_RETVAL, 0))
641 in_libcall = 0;
642
643 if (GET_CODE (p) == INSN
644 && (set = single_set (p))
645 && GET_CODE (SET_DEST (set)) == REG
646 && ! may_not_optimize[REGNO (SET_DEST (set))])
647 {
648 int tem1 = 0;
649 int tem2 = 0;
650 int move_insn = 0;
651 rtx src = SET_SRC (set);
652 rtx dependencies = 0;
653
654 /* Figure out what to use as a source of this insn. If a REG_EQUIV
655 note is given or if a REG_EQUAL note with a constant operand is
656 specified, use it as the source and mark that we should move
657 this insn by calling emit_move_insn rather that duplicating the
658 insn.
659
660 Otherwise, only use the REG_EQUAL contents if a REG_RETVAL note
661 is present. */
662 temp = find_reg_note (p, REG_EQUIV, 0);
663 if (temp)
664 src = XEXP (temp, 0), move_insn = 1;
665 else
666 {
667 temp = find_reg_note (p, REG_EQUAL, 0);
668 if (temp && CONSTANT_P (XEXP (temp, 0)))
669 src = XEXP (temp, 0), move_insn = 1;
670 if (temp && find_reg_note (p, REG_RETVAL, 0))
671 {
672 src = XEXP (temp, 0);
673 /* A libcall block can use regs that don't appear in
674 the equivalent expression. To move the libcall,
675 we must move those regs too. */
676 dependencies = libcall_other_reg (p, src);
677 }
678 }
679
680 /* Don't try to optimize a register that was made
681 by loop-optimization for an inner loop.
682 We don't know its life-span, so we can't compute the benefit. */
683 if (REGNO (SET_DEST (set)) >= max_reg_before_loop)
684 ;
685 /* In order to move a register, we need to have one of three cases:
686 (1) it is used only in the same basic block as the set
687 (2) it is not a user variable.
688 (3) the set is guaranteed to be executed once the loop starts,
689 and the reg is not used until after that. */
690 else if (! ((! maybe_never
691 && ! loop_reg_used_before_p (set, p, loop_start,
692 scan_start, end))
693 || ! REG_USERVAR_P (SET_DEST (PATTERN (p)))
694 || reg_in_basic_block_p (p, SET_DEST (PATTERN (p)))))
695 ;
696 else if ((tem = invariant_p (src))
697 && (dependencies == 0
698 || (tem2 = invariant_p (dependencies)) != 0)
699 && (n_times_set[REGNO (SET_DEST (set))] == 1
700 || (tem1
701 = consec_sets_invariant_p (SET_DEST (set),
702 n_times_set[REGNO (SET_DEST (set))],
703 p)))
704 /* If the insn can cause a trap (such as divide by zero),
705 can't move it unless it's guaranteed to be executed
706 once loop is entered. Even a function call might
707 prevent the trap insn from being reached
708 (since it might exit!) */
709 && ! ((maybe_never || call_passed)
710 && may_trap_p (src)))
711 {
712 register struct movable *m;
713 register int regno = REGNO (SET_DEST (set));
714
715 /* A potential lossage is where we have a case where two insns
716 can be combined as long as they are both in the loop, but
717 we move one of them outside the loop. For large loops,
718 this can lose. The most common case of this is the address
719 of a function being called.
720
721 Therefore, if this register is marked as being used exactly
722 once if we are in a loop with calls (a "large loop"), see if
723 we can replace the usage of this register with the source
724 of this SET. If we can, delete this insn.
725
726 Don't do this if P has a REG_RETVAL note or if we have
727 SMALL_REGISTER_CLASSES and SET_SRC is a hard register. */
728
729 if (reg_single_usage && reg_single_usage[regno] != 0
730 && reg_single_usage[regno] != const0_rtx
731 && regno_first_uid[regno] == INSN_UID (p)
732 && (regno_last_uid[regno]
733 == INSN_UID (reg_single_usage[regno]))
734 && n_times_set[REGNO (SET_DEST (set))] == 1
735 && ! side_effects_p (SET_SRC (set))
736 && ! find_reg_note (p, REG_RETVAL, 0)
737 #ifdef SMALL_REGISTER_CLASSES
738 && ! (GET_CODE (SET_SRC (set)) == REG
739 && REGNO (SET_SRC (set)) < FIRST_PSEUDO_REGISTER)
740 #endif
741 /* This test is not redundant; SET_SRC (set) might be
742 a call-clobbered register and the life of REGNO
743 might span a call. */
744 && ! modified_between_p (SET_SRC (set), p,
745 reg_single_usage[regno])
746 && validate_replace_rtx (SET_DEST (set), SET_SRC (set),
747 reg_single_usage[regno]))
748 {
749 /* Replace any usage in a REG_EQUAL note. */
750 REG_NOTES (reg_single_usage[regno])
751 = replace_rtx (REG_NOTES (reg_single_usage[regno]),
752 SET_DEST (set), SET_SRC (set));
753
754 PUT_CODE (p, NOTE);
755 NOTE_LINE_NUMBER (p) = NOTE_INSN_DELETED;
756 NOTE_SOURCE_FILE (p) = 0;
757 n_times_set[regno] = 0;
758 continue;
759 }
760
761 m = (struct movable *) alloca (sizeof (struct movable));
762 m->next = 0;
763 m->insn = p;
764 m->set_src = src;
765 m->dependencies = dependencies;
766 m->set_dest = SET_DEST (set);
767 m->force = 0;
768 m->consec = n_times_set[REGNO (SET_DEST (set))] - 1;
769 m->done = 0;
770 m->forces = 0;
771 m->partial = 0;
772 m->move_insn = move_insn;
773 m->is_equiv = (find_reg_note (p, REG_EQUIV, 0) != 0);
774 m->savemode = VOIDmode;
775 m->regno = regno;
776 /* Set M->cond if either invariant_p or consec_sets_invariant_p
777 returned 2 (only conditionally invariant). */
778 m->cond = ((tem | tem1 | tem2) > 1);
779 m->global = (uid_luid[regno_last_uid[regno]] > INSN_LUID (end)
780 || uid_luid[regno_first_uid[regno]] < INSN_LUID (loop_start));
781 m->match = 0;
782 m->lifetime = (uid_luid[regno_last_uid[regno]]
783 - uid_luid[regno_first_uid[regno]]);
784 m->savings = n_times_used[regno];
785 if (find_reg_note (p, REG_RETVAL, 0))
786 m->savings += libcall_benefit (p);
787 n_times_set[regno] = move_insn ? -2 : -1;
788 /* Add M to the end of the chain MOVABLES. */
789 if (movables == 0)
790 movables = m;
791 else
792 last_movable->next = m;
793 last_movable = m;
794
795 if (m->consec > 0)
796 {
797 /* Skip this insn, not checking REG_LIBCALL notes. */
798 p = NEXT_INSN (p);
799 /* Skip the consecutive insns, if there are any. */
800 p = skip_consec_insns (p, m->consec);
801 /* Back up to the last insn of the consecutive group. */
802 p = prev_nonnote_insn (p);
803
804 /* We must now reset m->move_insn, m->is_equiv, and possibly
805 m->set_src to correspond to the effects of all the
806 insns. */
807 temp = find_reg_note (p, REG_EQUIV, 0);
808 if (temp)
809 m->set_src = XEXP (temp, 0), m->move_insn = 1;
810 else
811 {
812 temp = find_reg_note (p, REG_EQUAL, 0);
813 if (temp && CONSTANT_P (XEXP (temp, 0)))
814 m->set_src = XEXP (temp, 0), m->move_insn = 1;
815 else
816 m->move_insn = 0;
817
818 }
819 m->is_equiv = (find_reg_note (p, REG_EQUIV, 0) != 0);
820 }
821 }
822 /* If this register is always set within a STRICT_LOW_PART
823 or set to zero, then its high bytes are constant.
824 So clear them outside the loop and within the loop
825 just load the low bytes.
826 We must check that the machine has an instruction to do so.
827 Also, if the value loaded into the register
828 depends on the same register, this cannot be done. */
829 else if (SET_SRC (set) == const0_rtx
830 && GET_CODE (NEXT_INSN (p)) == INSN
831 && (set1 = single_set (NEXT_INSN (p)))
832 && GET_CODE (set1) == SET
833 && (GET_CODE (SET_DEST (set1)) == STRICT_LOW_PART)
834 && (GET_CODE (XEXP (SET_DEST (set1), 0)) == SUBREG)
835 && (SUBREG_REG (XEXP (SET_DEST (set1), 0))
836 == SET_DEST (set))
837 && !reg_mentioned_p (SET_DEST (set), SET_SRC (set1)))
838 {
839 register int regno = REGNO (SET_DEST (set));
840 if (n_times_set[regno] == 2)
841 {
842 register struct movable *m;
843 m = (struct movable *) alloca (sizeof (struct movable));
844 m->next = 0;
845 m->insn = p;
846 m->set_dest = SET_DEST (set);
847 m->dependencies = 0;
848 m->force = 0;
849 m->consec = 0;
850 m->done = 0;
851 m->forces = 0;
852 m->move_insn = 0;
853 m->partial = 1;
854 /* If the insn may not be executed on some cycles,
855 we can't clear the whole reg; clear just high part.
856 Not even if the reg is used only within this loop.
857 Consider this:
858 while (1)
859 while (s != t) {
860 if (foo ()) x = *s;
861 use (x);
862 }
863 Clearing x before the inner loop could clobber a value
864 being saved from the last time around the outer loop.
865 However, if the reg is not used outside this loop
866 and all uses of the register are in the same
867 basic block as the store, there is no problem.
868
869 If this insn was made by loop, we don't know its
870 INSN_LUID and hence must make a conservative
871 assumption. */
872 m->global = (INSN_UID (p) >= max_uid_for_loop
873 || (uid_luid[regno_last_uid[regno]]
874 > INSN_LUID (end))
875 || (uid_luid[regno_first_uid[regno]]
876 < INSN_LUID (p))
877 || (labels_in_range_p
878 (p, uid_luid[regno_first_uid[regno]])));
879 if (maybe_never && m->global)
880 m->savemode = GET_MODE (SET_SRC (set1));
881 else
882 m->savemode = VOIDmode;
883 m->regno = regno;
884 m->cond = 0;
885 m->match = 0;
886 m->lifetime = (uid_luid[regno_last_uid[regno]]
887 - uid_luid[regno_first_uid[regno]]);
888 m->savings = 1;
889 n_times_set[regno] = -1;
890 /* Add M to the end of the chain MOVABLES. */
891 if (movables == 0)
892 movables = m;
893 else
894 last_movable->next = m;
895 last_movable = m;
896 }
897 }
898 }
899 /* Past a call insn, we get to insns which might not be executed
900 because the call might exit. This matters for insns that trap.
901 Call insns inside a REG_LIBCALL/REG_RETVAL block always return,
902 so they don't count. */
903 else if (GET_CODE (p) == CALL_INSN && ! in_libcall)
904 call_passed = 1;
905 /* Past a label or a jump, we get to insns for which we
906 can't count on whether or how many times they will be
907 executed during each iteration. Therefore, we can
908 only move out sets of trivial variables
909 (those not used after the loop). */
910 /* This code appears in three places, once in scan_loop, and twice
911 in strength_reduce. */
912 else if ((GET_CODE (p) == CODE_LABEL || GET_CODE (p) == JUMP_INSN)
913 /* If we enter the loop in the middle, and scan around to the
914 beginning, don't set maybe_never for that. This must be an
915 unconditional jump, otherwise the code at the top of the
916 loop might never be executed. Unconditional jumps are
917 followed a by barrier then loop end. */
918 && ! (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) == loop_top
919 && NEXT_INSN (NEXT_INSN (p)) == end
920 && simplejump_p (p)))
921 maybe_never = 1;
922 /* At the virtual top of a converted loop, insns are again known to
923 be executed: logically, the loop begins here even though the exit
924 code has been duplicated. */
925 else if (GET_CODE (p) == NOTE
926 && NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP)
927 maybe_never = call_passed = 0;
928 }
929
930 /* If one movable subsumes another, ignore that other. */
931
932 ignore_some_movables (movables);
933
934 /* For each movable insn, see if the reg that it loads
935 leads when it dies right into another conditionally movable insn.
936 If so, record that the second insn "forces" the first one,
937 since the second can be moved only if the first is. */
938
939 force_movables (movables);
940
941 /* See if there are multiple movable insns that load the same value.
942 If there are, make all but the first point at the first one
943 through the `match' field, and add the priorities of them
944 all together as the priority of the first. */
945
946 combine_movables (movables, nregs);
947
948 /* Now consider each movable insn to decide whether it is worth moving.
949 Store 0 in n_times_set for each reg that is moved. */
950
951 move_movables (movables, threshold,
952 insn_count, loop_start, end, nregs);
953
954 /* Now candidates that still are negative are those not moved.
955 Change n_times_set to indicate that those are not actually invariant. */
956 for (i = 0; i < nregs; i++)
957 if (n_times_set[i] < 0)
958 n_times_set[i] = n_times_used[i];
959
960 if (flag_strength_reduce)
961 strength_reduce (scan_start, end, loop_top,
962 insn_count, loop_start, end);
963 }
964 \f
965 /* Add elements to *OUTPUT to record all the pseudo-regs
966 mentioned in IN_THIS but not mentioned in NOT_IN_THIS. */
967
968 void
969 record_excess_regs (in_this, not_in_this, output)
970 rtx in_this, not_in_this;
971 rtx *output;
972 {
973 enum rtx_code code;
974 char *fmt;
975 int i;
976
977 code = GET_CODE (in_this);
978
979 switch (code)
980 {
981 case PC:
982 case CC0:
983 case CONST_INT:
984 case CONST_DOUBLE:
985 case CONST:
986 case SYMBOL_REF:
987 case LABEL_REF:
988 return;
989
990 case REG:
991 if (REGNO (in_this) >= FIRST_PSEUDO_REGISTER
992 && ! reg_mentioned_p (in_this, not_in_this))
993 *output = gen_rtx (EXPR_LIST, VOIDmode, in_this, *output);
994 return;
995 }
996
997 fmt = GET_RTX_FORMAT (code);
998 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
999 {
1000 int j;
1001
1002 switch (fmt[i])
1003 {
1004 case 'E':
1005 for (j = 0; j < XVECLEN (in_this, i); j++)
1006 record_excess_regs (XVECEXP (in_this, i, j), not_in_this, output);
1007 break;
1008
1009 case 'e':
1010 record_excess_regs (XEXP (in_this, i), not_in_this, output);
1011 break;
1012 }
1013 }
1014 }
1015 \f
1016 /* Check what regs are referred to in the libcall block ending with INSN,
1017 aside from those mentioned in the equivalent value.
1018 If there are none, return 0.
1019 If there are one or more, return an EXPR_LIST containing all of them. */
1020
1021 static rtx
1022 libcall_other_reg (insn, equiv)
1023 rtx insn, equiv;
1024 {
1025 rtx note = find_reg_note (insn, REG_RETVAL, 0);
1026 rtx p = XEXP (note, 0);
1027 rtx output = 0;
1028
1029 /* First, find all the regs used in the libcall block
1030 that are not mentioned as inputs to the result. */
1031
1032 while (p != insn)
1033 {
1034 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
1035 || GET_CODE (p) == CALL_INSN)
1036 record_excess_regs (PATTERN (p), equiv, &output);
1037 p = NEXT_INSN (p);
1038 }
1039
1040 return output;
1041 }
1042 \f
1043 /* Return 1 if all uses of REG
1044 are between INSN and the end of the basic block. */
1045
1046 static int
1047 reg_in_basic_block_p (insn, reg)
1048 rtx insn, reg;
1049 {
1050 int regno = REGNO (reg);
1051 rtx p;
1052
1053 if (regno_first_uid[regno] != INSN_UID (insn))
1054 return 0;
1055
1056 /* Search this basic block for the already recorded last use of the reg. */
1057 for (p = insn; p; p = NEXT_INSN (p))
1058 {
1059 switch (GET_CODE (p))
1060 {
1061 case NOTE:
1062 break;
1063
1064 case INSN:
1065 case CALL_INSN:
1066 /* Ordinary insn: if this is the last use, we win. */
1067 if (regno_last_uid[regno] == INSN_UID (p))
1068 return 1;
1069 break;
1070
1071 case JUMP_INSN:
1072 /* Jump insn: if this is the last use, we win. */
1073 if (regno_last_uid[regno] == INSN_UID (p))
1074 return 1;
1075 /* Otherwise, it's the end of the basic block, so we lose. */
1076 return 0;
1077
1078 case CODE_LABEL:
1079 case BARRIER:
1080 /* It's the end of the basic block, so we lose. */
1081 return 0;
1082 }
1083 }
1084
1085 /* The "last use" doesn't follow the "first use"?? */
1086 abort ();
1087 }
1088 \f
1089 /* Compute the benefit of eliminating the insns in the block whose
1090 last insn is LAST. This may be a group of insns used to compute a
1091 value directly or can contain a library call. */
1092
1093 static int
1094 libcall_benefit (last)
1095 rtx last;
1096 {
1097 rtx insn;
1098 int benefit = 0;
1099
1100 for (insn = XEXP (find_reg_note (last, REG_RETVAL, 0), 0);
1101 insn != last; insn = NEXT_INSN (insn))
1102 {
1103 if (GET_CODE (insn) == CALL_INSN)
1104 benefit += 10; /* Assume at least this many insns in a library
1105 routine. */
1106 else if (GET_CODE (insn) == INSN
1107 && GET_CODE (PATTERN (insn)) != USE
1108 && GET_CODE (PATTERN (insn)) != CLOBBER)
1109 benefit++;
1110 }
1111
1112 return benefit;
1113 }
1114 \f
1115 /* Skip COUNT insns from INSN, counting library calls as 1 insn. */
1116
1117 static rtx
1118 skip_consec_insns (insn, count)
1119 rtx insn;
1120 int count;
1121 {
1122 for (; count > 0; count--)
1123 {
1124 rtx temp;
1125
1126 /* If first insn of libcall sequence, skip to end. */
1127 /* Do this at start of loop, since INSN is guaranteed to
1128 be an insn here. */
1129 if (GET_CODE (insn) != NOTE
1130 && (temp = find_reg_note (insn, REG_LIBCALL, 0)))
1131 insn = XEXP (temp, 0);
1132
1133 do insn = NEXT_INSN (insn);
1134 while (GET_CODE (insn) == NOTE);
1135 }
1136
1137 return insn;
1138 }
1139
1140 /* Ignore any movable whose insn falls within a libcall
1141 which is part of another movable.
1142 We make use of the fact that the movable for the libcall value
1143 was made later and so appears later on the chain. */
1144
1145 static void
1146 ignore_some_movables (movables)
1147 struct movable *movables;
1148 {
1149 register struct movable *m, *m1;
1150
1151 for (m = movables; m; m = m->next)
1152 {
1153 /* Is this a movable for the value of a libcall? */
1154 rtx note = find_reg_note (m->insn, REG_RETVAL, 0);
1155 if (note)
1156 {
1157 rtx insn;
1158 /* Check for earlier movables inside that range,
1159 and mark them invalid. We cannot use LUIDs here because
1160 insns created by loop.c for prior loops don't have LUIDs.
1161 Rather than reject all such insns from movables, we just
1162 explicitly check each insn in the libcall (since invariant
1163 libcalls aren't that common). */
1164 for (insn = XEXP (note, 0); insn != m->insn; insn = NEXT_INSN (insn))
1165 for (m1 = movables; m1 != m; m1 = m1->next)
1166 if (m1->insn == insn)
1167 m1->done = 1;
1168 }
1169 }
1170 }
1171
1172 /* For each movable insn, see if the reg that it loads
1173 leads when it dies right into another conditionally movable insn.
1174 If so, record that the second insn "forces" the first one,
1175 since the second can be moved only if the first is. */
1176
1177 static void
1178 force_movables (movables)
1179 struct movable *movables;
1180 {
1181 register struct movable *m, *m1;
1182 for (m1 = movables; m1; m1 = m1->next)
1183 /* Omit this if moving just the (SET (REG) 0) of a zero-extend. */
1184 if (!m1->partial && !m1->done)
1185 {
1186 int regno = m1->regno;
1187 for (m = m1->next; m; m = m->next)
1188 /* ??? Could this be a bug? What if CSE caused the
1189 register of M1 to be used after this insn?
1190 Since CSE does not update regno_last_uid,
1191 this insn M->insn might not be where it dies.
1192 But very likely this doesn't matter; what matters is
1193 that M's reg is computed from M1's reg. */
1194 if (INSN_UID (m->insn) == regno_last_uid[regno]
1195 && !m->done)
1196 break;
1197 if (m != 0 && m->set_src == m1->set_dest
1198 /* If m->consec, m->set_src isn't valid. */
1199 && m->consec == 0)
1200 m = 0;
1201
1202 /* Increase the priority of the moving the first insn
1203 since it permits the second to be moved as well. */
1204 if (m != 0)
1205 {
1206 m->forces = m1;
1207 m1->lifetime += m->lifetime;
1208 m1->savings += m1->savings;
1209 }
1210 }
1211 }
1212 \f
1213 /* Find invariant expressions that are equal and can be combined into
1214 one register. */
1215
1216 static void
1217 combine_movables (movables, nregs)
1218 struct movable *movables;
1219 int nregs;
1220 {
1221 register struct movable *m;
1222 char *matched_regs = (char *) alloca (nregs);
1223 enum machine_mode mode;
1224
1225 /* Regs that are set more than once are not allowed to match
1226 or be matched. I'm no longer sure why not. */
1227 /* Perhaps testing m->consec_sets would be more appropriate here? */
1228
1229 for (m = movables; m; m = m->next)
1230 if (m->match == 0 && n_times_used[m->regno] == 1 && !m->partial)
1231 {
1232 register struct movable *m1;
1233 int regno = m->regno;
1234 rtx reg_note, reg_note1;
1235
1236 bzero (matched_regs, nregs);
1237 matched_regs[regno] = 1;
1238
1239 for (m1 = movables; m1; m1 = m1->next)
1240 if (m != m1 && m1->match == 0 && n_times_used[m1->regno] == 1
1241 /* A reg used outside the loop mustn't be eliminated. */
1242 && !m1->global
1243 /* A reg used for zero-extending mustn't be eliminated. */
1244 && !m1->partial
1245 && (matched_regs[m1->regno]
1246 ||
1247 (
1248 /* Can combine regs with different modes loaded from the
1249 same constant only if the modes are the same or
1250 if both are integer modes with M wider or the same
1251 width as M1. The check for integer is redundant, but
1252 safe, since the only case of differing destination
1253 modes with equal sources is when both sources are
1254 VOIDmode, i.e., CONST_INT. */
1255 (GET_MODE (m->set_dest) == GET_MODE (m1->set_dest)
1256 || (GET_MODE_CLASS (GET_MODE (m->set_dest)) == MODE_INT
1257 && GET_MODE_CLASS (GET_MODE (m1->set_dest)) == MODE_INT
1258 && (GET_MODE_BITSIZE (GET_MODE (m->set_dest))
1259 >= GET_MODE_BITSIZE (GET_MODE (m1->set_dest)))))
1260 /* See if the source of M1 says it matches M. */
1261 && ((GET_CODE (m1->set_src) == REG
1262 && matched_regs[REGNO (m1->set_src)])
1263 || rtx_equal_for_loop_p (m->set_src, m1->set_src,
1264 movables))))
1265 && ((m->dependencies == m1->dependencies)
1266 || rtx_equal_p (m->dependencies, m1->dependencies)))
1267 {
1268 m->lifetime += m1->lifetime;
1269 m->savings += m1->savings;
1270 m1->done = 1;
1271 m1->match = m;
1272 matched_regs[m1->regno] = 1;
1273 }
1274 }
1275
1276 /* Now combine the regs used for zero-extension.
1277 This can be done for those not marked `global'
1278 provided their lives don't overlap. */
1279
1280 for (mode = GET_CLASS_NARROWEST_MODE (MODE_INT); mode != VOIDmode;
1281 mode = GET_MODE_WIDER_MODE (mode))
1282 {
1283 register struct movable *m0 = 0;
1284
1285 /* Combine all the registers for extension from mode MODE.
1286 Don't combine any that are used outside this loop. */
1287 for (m = movables; m; m = m->next)
1288 if (m->partial && ! m->global
1289 && mode == GET_MODE (SET_SRC (PATTERN (NEXT_INSN (m->insn)))))
1290 {
1291 register struct movable *m1;
1292 int first = uid_luid[regno_first_uid[m->regno]];
1293 int last = uid_luid[regno_last_uid[m->regno]];
1294
1295 if (m0 == 0)
1296 {
1297 /* First one: don't check for overlap, just record it. */
1298 m0 = m;
1299 continue;
1300 }
1301
1302 /* Make sure they extend to the same mode.
1303 (Almost always true.) */
1304 if (GET_MODE (m->set_dest) != GET_MODE (m0->set_dest))
1305 continue;
1306
1307 /* We already have one: check for overlap with those
1308 already combined together. */
1309 for (m1 = movables; m1 != m; m1 = m1->next)
1310 if (m1 == m0 || (m1->partial && m1->match == m0))
1311 if (! (uid_luid[regno_first_uid[m1->regno]] > last
1312 || uid_luid[regno_last_uid[m1->regno]] < first))
1313 goto overlap;
1314
1315 /* No overlap: we can combine this with the others. */
1316 m0->lifetime += m->lifetime;
1317 m0->savings += m->savings;
1318 m->done = 1;
1319 m->match = m0;
1320
1321 overlap: ;
1322 }
1323 }
1324 }
1325 \f
1326 /* Return 1 if regs X and Y will become the same if moved. */
1327
1328 static int
1329 regs_match_p (x, y, movables)
1330 rtx x, y;
1331 struct movable *movables;
1332 {
1333 int xn = REGNO (x);
1334 int yn = REGNO (y);
1335 struct movable *mx, *my;
1336
1337 for (mx = movables; mx; mx = mx->next)
1338 if (mx->regno == xn)
1339 break;
1340
1341 for (my = movables; my; my = my->next)
1342 if (my->regno == yn)
1343 break;
1344
1345 return (mx && my
1346 && ((mx->match == my->match && mx->match != 0)
1347 || mx->match == my
1348 || mx == my->match));
1349 }
1350
1351 /* Return 1 if X and Y are identical-looking rtx's.
1352 This is the Lisp function EQUAL for rtx arguments.
1353
1354 If two registers are matching movables or a movable register and an
1355 equivalent constant, consider them equal. */
1356
1357 static int
1358 rtx_equal_for_loop_p (x, y, movables)
1359 rtx x, y;
1360 struct movable *movables;
1361 {
1362 register int i;
1363 register int j;
1364 register struct movable *m;
1365 register enum rtx_code code;
1366 register char *fmt;
1367
1368 if (x == y)
1369 return 1;
1370 if (x == 0 || y == 0)
1371 return 0;
1372
1373 code = GET_CODE (x);
1374
1375 /* If we have a register and a constant, they may sometimes be
1376 equal. */
1377 if (GET_CODE (x) == REG && n_times_set[REGNO (x)] == -2
1378 && CONSTANT_P (y))
1379 for (m = movables; m; m = m->next)
1380 if (m->move_insn && m->regno == REGNO (x)
1381 && rtx_equal_p (m->set_src, y))
1382 return 1;
1383
1384 else if (GET_CODE (y) == REG && n_times_set[REGNO (y)] == -2
1385 && CONSTANT_P (x))
1386 for (m = movables; m; m = m->next)
1387 if (m->move_insn && m->regno == REGNO (y)
1388 && rtx_equal_p (m->set_src, x))
1389 return 1;
1390
1391 /* Otherwise, rtx's of different codes cannot be equal. */
1392 if (code != GET_CODE (y))
1393 return 0;
1394
1395 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent.
1396 (REG:SI x) and (REG:HI x) are NOT equivalent. */
1397
1398 if (GET_MODE (x) != GET_MODE (y))
1399 return 0;
1400
1401 /* These three types of rtx's can be compared nonrecursively. */
1402 if (code == REG)
1403 return (REGNO (x) == REGNO (y) || regs_match_p (x, y, movables));
1404
1405 if (code == LABEL_REF)
1406 return XEXP (x, 0) == XEXP (y, 0);
1407 if (code == SYMBOL_REF)
1408 return XSTR (x, 0) == XSTR (y, 0);
1409
1410 /* Compare the elements. If any pair of corresponding elements
1411 fail to match, return 0 for the whole things. */
1412
1413 fmt = GET_RTX_FORMAT (code);
1414 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1415 {
1416 switch (fmt[i])
1417 {
1418 case 'i':
1419 if (XINT (x, i) != XINT (y, i))
1420 return 0;
1421 break;
1422
1423 case 'E':
1424 /* Two vectors must have the same length. */
1425 if (XVECLEN (x, i) != XVECLEN (y, i))
1426 return 0;
1427
1428 /* And the corresponding elements must match. */
1429 for (j = 0; j < XVECLEN (x, i); j++)
1430 if (rtx_equal_for_loop_p (XVECEXP (x, i, j), XVECEXP (y, i, j), movables) == 0)
1431 return 0;
1432 break;
1433
1434 case 'e':
1435 if (rtx_equal_for_loop_p (XEXP (x, i), XEXP (y, i), movables) == 0)
1436 return 0;
1437 break;
1438
1439 case 's':
1440 if (strcmp (XSTR (x, i), XSTR (y, i)))
1441 return 0;
1442 break;
1443
1444 case 'u':
1445 /* These are just backpointers, so they don't matter. */
1446 break;
1447
1448 case '0':
1449 break;
1450
1451 /* It is believed that rtx's at this level will never
1452 contain anything but integers and other rtx's,
1453 except for within LABEL_REFs and SYMBOL_REFs. */
1454 default:
1455 abort ();
1456 }
1457 }
1458 return 1;
1459 }
1460 \f
1461 /* If X contains any LABEL_REF's, add REG_LABEL notes for them to all
1462 insns in INSNS which use thet reference. */
1463
1464 static void
1465 add_label_notes (x, insns)
1466 rtx x;
1467 rtx insns;
1468 {
1469 enum rtx_code code = GET_CODE (x);
1470 int i, j;
1471 char *fmt;
1472 rtx insn;
1473
1474 if (code == LABEL_REF)
1475 {
1476 for (insn = insns; insn; insn = NEXT_INSN (insn))
1477 if (reg_mentioned_p (XEXP (x, 0), insn))
1478 REG_NOTES (insn) = gen_rtx (EXPR_LIST, REG_LABEL, XEXP (x, 0),
1479 REG_NOTES (insn));
1480 return;
1481 }
1482
1483 fmt = GET_RTX_FORMAT (code);
1484 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1485 {
1486 if (fmt[i] == 'e')
1487 add_label_notes (XEXP (x, i), insns);
1488 else if (fmt[i] == 'E')
1489 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
1490 add_label_notes (XVECEXP (x, i, j), insns);
1491 }
1492 }
1493 \f
1494 /* Scan MOVABLES, and move the insns that deserve to be moved.
1495 If two matching movables are combined, replace one reg with the
1496 other throughout. */
1497
1498 static void
1499 move_movables (movables, threshold, insn_count, loop_start, end, nregs)
1500 struct movable *movables;
1501 int threshold;
1502 int insn_count;
1503 rtx loop_start;
1504 rtx end;
1505 int nregs;
1506 {
1507 rtx new_start = 0;
1508 register struct movable *m;
1509 register rtx p;
1510 /* Map of pseudo-register replacements to handle combining
1511 when we move several insns that load the same value
1512 into different pseudo-registers. */
1513 rtx *reg_map = (rtx *) alloca (nregs * sizeof (rtx));
1514 char *already_moved = (char *) alloca (nregs);
1515
1516 bzero (already_moved, nregs);
1517 bzero (reg_map, nregs * sizeof (rtx));
1518
1519 num_movables = 0;
1520
1521 for (m = movables; m; m = m->next)
1522 {
1523 /* Describe this movable insn. */
1524
1525 if (loop_dump_stream)
1526 {
1527 fprintf (loop_dump_stream, "Insn %d: regno %d (life %d), ",
1528 INSN_UID (m->insn), m->regno, m->lifetime);
1529 if (m->consec > 0)
1530 fprintf (loop_dump_stream, "consec %d, ", m->consec);
1531 if (m->cond)
1532 fprintf (loop_dump_stream, "cond ");
1533 if (m->force)
1534 fprintf (loop_dump_stream, "force ");
1535 if (m->global)
1536 fprintf (loop_dump_stream, "global ");
1537 if (m->done)
1538 fprintf (loop_dump_stream, "done ");
1539 if (m->move_insn)
1540 fprintf (loop_dump_stream, "move-insn ");
1541 if (m->match)
1542 fprintf (loop_dump_stream, "matches %d ",
1543 INSN_UID (m->match->insn));
1544 if (m->forces)
1545 fprintf (loop_dump_stream, "forces %d ",
1546 INSN_UID (m->forces->insn));
1547 }
1548
1549 /* Count movables. Value used in heuristics in strength_reduce. */
1550 num_movables++;
1551
1552 /* Ignore the insn if it's already done (it matched something else).
1553 Otherwise, see if it is now safe to move. */
1554
1555 if (!m->done
1556 && (! m->cond
1557 || (1 == invariant_p (m->set_src)
1558 && (m->dependencies == 0
1559 || 1 == invariant_p (m->dependencies))
1560 && (m->consec == 0
1561 || 1 == consec_sets_invariant_p (m->set_dest,
1562 m->consec + 1,
1563 m->insn))))
1564 && (! m->forces || m->forces->done))
1565 {
1566 register int regno;
1567 register rtx p;
1568 int savings = m->savings;
1569
1570 /* We have an insn that is safe to move.
1571 Compute its desirability. */
1572
1573 p = m->insn;
1574 regno = m->regno;
1575
1576 if (loop_dump_stream)
1577 fprintf (loop_dump_stream, "savings %d ", savings);
1578
1579 if (moved_once[regno])
1580 {
1581 insn_count *= 2;
1582
1583 if (loop_dump_stream)
1584 fprintf (loop_dump_stream, "halved since already moved ");
1585 }
1586
1587 /* An insn MUST be moved if we already moved something else
1588 which is safe only if this one is moved too: that is,
1589 if already_moved[REGNO] is nonzero. */
1590
1591 /* An insn is desirable to move if the new lifetime of the
1592 register is no more than THRESHOLD times the old lifetime.
1593 If it's not desirable, it means the loop is so big
1594 that moving won't speed things up much,
1595 and it is liable to make register usage worse. */
1596
1597 /* It is also desirable to move if it can be moved at no
1598 extra cost because something else was already moved. */
1599
1600 if (already_moved[regno]
1601 || (threshold * savings * m->lifetime) >= insn_count
1602 || (m->forces && m->forces->done
1603 && n_times_used[m->forces->regno] == 1))
1604 {
1605 int count;
1606 register struct movable *m1;
1607 rtx first;
1608
1609 /* Now move the insns that set the reg. */
1610
1611 if (m->partial && m->match)
1612 {
1613 rtx newpat, i1;
1614 rtx r1, r2;
1615 /* Find the end of this chain of matching regs.
1616 Thus, we load each reg in the chain from that one reg.
1617 And that reg is loaded with 0 directly,
1618 since it has ->match == 0. */
1619 for (m1 = m; m1->match; m1 = m1->match);
1620 newpat = gen_move_insn (SET_DEST (PATTERN (m->insn)),
1621 SET_DEST (PATTERN (m1->insn)));
1622 i1 = emit_insn_before (newpat, loop_start);
1623
1624 /* Mark the moved, invariant reg as being allowed to
1625 share a hard reg with the other matching invariant. */
1626 REG_NOTES (i1) = REG_NOTES (m->insn);
1627 r1 = SET_DEST (PATTERN (m->insn));
1628 r2 = SET_DEST (PATTERN (m1->insn));
1629 regs_may_share = gen_rtx (EXPR_LIST, VOIDmode, r1,
1630 gen_rtx (EXPR_LIST, VOIDmode, r2,
1631 regs_may_share));
1632 delete_insn (m->insn);
1633
1634 if (new_start == 0)
1635 new_start = i1;
1636
1637 if (loop_dump_stream)
1638 fprintf (loop_dump_stream, " moved to %d", INSN_UID (i1));
1639 }
1640 /* If we are to re-generate the item being moved with a
1641 new move insn, first delete what we have and then emit
1642 the move insn before the loop. */
1643 else if (m->move_insn)
1644 {
1645 rtx i1, temp;
1646
1647 for (count = m->consec; count >= 0; count--)
1648 {
1649 /* If this is the first insn of a library call sequence,
1650 skip to the end. */
1651 if (GET_CODE (p) != NOTE
1652 && (temp = find_reg_note (p, REG_LIBCALL, 0)))
1653 p = XEXP (temp, 0);
1654
1655 /* If this is the last insn of a libcall sequence, then
1656 delete every insn in the sequence except the last.
1657 The last insn is handled in the normal manner. */
1658 if (GET_CODE (p) != NOTE
1659 && (temp = find_reg_note (p, REG_RETVAL, 0)))
1660 {
1661 temp = XEXP (temp, 0);
1662 while (temp != p)
1663 temp = delete_insn (temp);
1664 }
1665
1666 p = delete_insn (p);
1667 }
1668
1669 start_sequence ();
1670 emit_move_insn (m->set_dest, m->set_src);
1671 temp = get_insns ();
1672 end_sequence ();
1673
1674 add_label_notes (m->set_src, temp);
1675
1676 i1 = emit_insns_before (temp, loop_start);
1677 if (! find_reg_note (i1, REG_EQUAL, 0))
1678 REG_NOTES (i1)
1679 = gen_rtx (EXPR_LIST,
1680 m->is_equiv ? REG_EQUIV : REG_EQUAL,
1681 m->set_src, REG_NOTES (i1));
1682
1683 if (loop_dump_stream)
1684 fprintf (loop_dump_stream, " moved to %d", INSN_UID (i1));
1685
1686 /* The more regs we move, the less we like moving them. */
1687 threshold -= 3;
1688 }
1689 else
1690 {
1691 for (count = m->consec; count >= 0; count--)
1692 {
1693 rtx i1, temp;
1694
1695 /* If first insn of libcall sequence, skip to end. */
1696 /* Do this at start of loop, since p is guaranteed to
1697 be an insn here. */
1698 if (GET_CODE (p) != NOTE
1699 && (temp = find_reg_note (p, REG_LIBCALL, 0)))
1700 p = XEXP (temp, 0);
1701
1702 /* If last insn of libcall sequence, move all
1703 insns except the last before the loop. The last
1704 insn is handled in the normal manner. */
1705 if (GET_CODE (p) != NOTE
1706 && (temp = find_reg_note (p, REG_RETVAL, 0)))
1707 {
1708 rtx fn_address = 0;
1709 rtx fn_reg = 0;
1710 rtx fn_address_insn = 0;
1711
1712 first = 0;
1713 for (temp = XEXP (temp, 0); temp != p;
1714 temp = NEXT_INSN (temp))
1715 {
1716 rtx body;
1717 rtx n;
1718 rtx next;
1719
1720 if (GET_CODE (temp) == NOTE)
1721 continue;
1722
1723 body = PATTERN (temp);
1724
1725 /* Find the next insn after TEMP,
1726 not counting USE or NOTE insns. */
1727 for (next = NEXT_INSN (temp); next != p;
1728 next = NEXT_INSN (next))
1729 if (! (GET_CODE (next) == INSN
1730 && GET_CODE (PATTERN (next)) == USE)
1731 && GET_CODE (next) != NOTE)
1732 break;
1733
1734 /* If that is the call, this may be the insn
1735 that loads the function address.
1736
1737 Extract the function address from the insn
1738 that loads it into a register.
1739 If this insn was cse'd, we get incorrect code.
1740
1741 So emit a new move insn that copies the
1742 function address into the register that the
1743 call insn will use. flow.c will delete any
1744 redundant stores that we have created. */
1745 if (GET_CODE (next) == CALL_INSN
1746 && GET_CODE (body) == SET
1747 && GET_CODE (SET_DEST (body)) == REG
1748 && (n = find_reg_note (temp, REG_EQUAL, 0)))
1749 {
1750 fn_reg = SET_SRC (body);
1751 if (GET_CODE (fn_reg) != REG)
1752 fn_reg = SET_DEST (body);
1753 fn_address = XEXP (n, 0);
1754 fn_address_insn = temp;
1755 }
1756 /* We have the call insn.
1757 If it uses the register we suspect it might,
1758 load it with the correct address directly. */
1759 if (GET_CODE (temp) == CALL_INSN
1760 && fn_address != 0
1761 && reg_mentioned_p (fn_reg, body))
1762 emit_insn_after (gen_move_insn (fn_reg,
1763 fn_address),
1764 fn_address_insn);
1765
1766 if (GET_CODE (temp) == CALL_INSN)
1767 i1 = emit_call_insn_before (body, loop_start);
1768 else
1769 i1 = emit_insn_before (body, loop_start);
1770 if (first == 0)
1771 first = i1;
1772 if (temp == fn_address_insn)
1773 fn_address_insn = i1;
1774 REG_NOTES (i1) = REG_NOTES (temp);
1775 delete_insn (temp);
1776 }
1777 }
1778 if (m->savemode != VOIDmode)
1779 {
1780 /* P sets REG to zero; but we should clear only
1781 the bits that are not covered by the mode
1782 m->savemode. */
1783 rtx reg = m->set_dest;
1784 rtx sequence;
1785 rtx tem;
1786
1787 start_sequence ();
1788 tem = expand_binop
1789 (GET_MODE (reg), and_optab, reg,
1790 gen_rtx (CONST_INT, VOIDmode,
1791 ((1 << GET_MODE_BITSIZE (m->savemode)))
1792 - 1),
1793 reg, 1, OPTAB_LIB_WIDEN);
1794 if (tem == 0)
1795 abort ();
1796 if (tem != reg)
1797 emit_move_insn (reg, tem);
1798 sequence = gen_sequence ();
1799 end_sequence ();
1800 i1 = emit_insn_before (sequence, loop_start);
1801 }
1802 else if (GET_CODE (p) == CALL_INSN)
1803 i1 = emit_call_insn_before (PATTERN (p), loop_start);
1804 else
1805 i1 = emit_insn_before (PATTERN (p), loop_start);
1806
1807 REG_NOTES (i1) = REG_NOTES (p);
1808
1809 if (new_start == 0)
1810 new_start = i1;
1811
1812 if (loop_dump_stream)
1813 fprintf (loop_dump_stream, " moved to %d",
1814 INSN_UID (i1));
1815
1816 #if 0
1817 /* This isn't needed because REG_NOTES is copied
1818 below and is wrong since P might be a PARALLEL. */
1819 if (REG_NOTES (i1) == 0
1820 && ! m->partial /* But not if it's a zero-extend clr. */
1821 && ! m->global /* and not if used outside the loop
1822 (since it might get set outside). */
1823 && CONSTANT_P (SET_SRC (PATTERN (p))))
1824 REG_NOTES (i1)
1825 = gen_rtx (EXPR_LIST, REG_EQUAL,
1826 SET_SRC (PATTERN (p)), REG_NOTES (i1));
1827 #endif
1828
1829 /* If library call, now fix the REG_NOTES that contain
1830 insn pointers, namely REG_LIBCALL on FIRST
1831 and REG_RETVAL on I1. */
1832 if (temp = find_reg_note (i1, REG_RETVAL, 0))
1833 {
1834 XEXP (temp, 0) = first;
1835 temp = find_reg_note (first, REG_LIBCALL, 0);
1836 XEXP (temp, 0) = i1;
1837 }
1838
1839 delete_insn (p);
1840 do p = NEXT_INSN (p);
1841 while (p && GET_CODE (p) == NOTE);
1842 }
1843
1844 /* The more regs we move, the less we like moving them. */
1845 threshold -= 3;
1846 }
1847
1848 /* Any other movable that loads the same register
1849 MUST be moved. */
1850 already_moved[regno] = 1;
1851
1852 /* This reg has been moved out of one loop. */
1853 moved_once[regno] = 1;
1854
1855 /* The reg set here is now invariant. */
1856 if (! m->partial)
1857 n_times_set[regno] = 0;
1858
1859 m->done = 1;
1860
1861 /* Change the length-of-life info for the register
1862 to say it lives at least the full length of this loop.
1863 This will help guide optimizations in outer loops. */
1864
1865 if (uid_luid[regno_first_uid[regno]] > INSN_LUID (loop_start))
1866 /* This is the old insn before all the moved insns.
1867 We can't use the moved insn because it is out of range
1868 in uid_luid. Only the old insns have luids. */
1869 regno_first_uid[regno] = INSN_UID (loop_start);
1870 if (uid_luid[regno_last_uid[regno]] < INSN_LUID (end))
1871 regno_last_uid[regno] = INSN_UID (end);
1872
1873 /* Combine with this moved insn any other matching movables. */
1874
1875 if (! m->partial)
1876 for (m1 = movables; m1; m1 = m1->next)
1877 if (m1->match == m)
1878 {
1879 rtx temp;
1880
1881 /* Schedule the reg loaded by M1
1882 for replacement so that shares the reg of M.
1883 If the modes differ (only possible in restricted
1884 circumstances, make a SUBREG. */
1885 if (GET_MODE (m->set_dest) == GET_MODE (m1->set_dest))
1886 reg_map[m1->regno] = m->set_dest;
1887 else
1888 reg_map[m1->regno]
1889 = gen_lowpart_common (GET_MODE (m1->set_dest),
1890 m->set_dest);
1891
1892 /* Get rid of the matching insn
1893 and prevent further processing of it. */
1894 m1->done = 1;
1895
1896 /* if library call, delete all insn except last, which
1897 is deleted below */
1898 if (temp = find_reg_note (m1->insn, REG_RETVAL, 0))
1899 {
1900 for (temp = XEXP (temp, 0); temp != m1->insn;
1901 temp = NEXT_INSN (temp))
1902 delete_insn (temp);
1903 }
1904 delete_insn (m1->insn);
1905
1906 /* Any other movable that loads the same register
1907 MUST be moved. */
1908 already_moved[m1->regno] = 1;
1909
1910 /* The reg merged here is now invariant,
1911 if the reg it matches is invariant. */
1912 if (! m->partial)
1913 n_times_set[m1->regno] = 0;
1914 }
1915 }
1916 else if (loop_dump_stream)
1917 fprintf (loop_dump_stream, "not desirable");
1918 }
1919 else if (loop_dump_stream && !m->match)
1920 fprintf (loop_dump_stream, "not safe");
1921
1922 if (loop_dump_stream)
1923 fprintf (loop_dump_stream, "\n");
1924 }
1925
1926 if (new_start == 0)
1927 new_start = loop_start;
1928
1929 /* Go through all the instructions in the loop, making
1930 all the register substitutions scheduled in REG_MAP. */
1931 for (p = new_start; p != end; p = NEXT_INSN (p))
1932 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
1933 || GET_CODE (p) == CALL_INSN)
1934 {
1935 replace_regs (PATTERN (p), reg_map, nregs, 0);
1936 replace_regs (REG_NOTES (p), reg_map, nregs, 0);
1937 }
1938 }
1939 \f
1940 #if 0
1941 /* Scan X and replace the address of any MEM in it with ADDR.
1942 REG is the address that MEM should have before the replacement. */
1943
1944 static void
1945 replace_call_address (x, reg, addr)
1946 rtx x, reg, addr;
1947 {
1948 register enum rtx_code code;
1949 register int i;
1950 register char *fmt;
1951
1952 if (x == 0)
1953 return;
1954 code = GET_CODE (x);
1955 switch (code)
1956 {
1957 case PC:
1958 case CC0:
1959 case CONST_INT:
1960 case CONST_DOUBLE:
1961 case CONST:
1962 case SYMBOL_REF:
1963 case LABEL_REF:
1964 case REG:
1965 return;
1966
1967 case SET:
1968 /* Short cut for very common case. */
1969 replace_call_address (XEXP (x, 1), reg, addr);
1970 return;
1971
1972 case CALL:
1973 /* Short cut for very common case. */
1974 replace_call_address (XEXP (x, 0), reg, addr);
1975 return;
1976
1977 case MEM:
1978 /* If this MEM uses a reg other than the one we expected,
1979 something is wrong. */
1980 if (XEXP (x, 0) != reg)
1981 abort ();
1982 XEXP (x, 0) = addr;
1983 return;
1984 }
1985
1986 fmt = GET_RTX_FORMAT (code);
1987 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1988 {
1989 if (fmt[i] == 'e')
1990 replace_call_address (XEXP (x, i), reg, addr);
1991 if (fmt[i] == 'E')
1992 {
1993 register int j;
1994 for (j = 0; j < XVECLEN (x, i); j++)
1995 replace_call_address (XVECEXP (x, i, j), reg, addr);
1996 }
1997 }
1998 }
1999 #endif
2000 \f
2001 /* Return the number of memory refs to addresses that vary
2002 in the rtx X. */
2003
2004 static int
2005 count_nonfixed_reads (x)
2006 rtx x;
2007 {
2008 register enum rtx_code code;
2009 register int i;
2010 register char *fmt;
2011 int value;
2012
2013 if (x == 0)
2014 return 0;
2015
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 0;
2028
2029 case MEM:
2030 return ((invariant_p (XEXP (x, 0)) != 1)
2031 + count_nonfixed_reads (XEXP (x, 0)));
2032 }
2033
2034 value = 0;
2035 fmt = GET_RTX_FORMAT (code);
2036 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2037 {
2038 if (fmt[i] == 'e')
2039 value += count_nonfixed_reads (XEXP (x, i));
2040 if (fmt[i] == 'E')
2041 {
2042 register int j;
2043 for (j = 0; j < XVECLEN (x, i); j++)
2044 value += count_nonfixed_reads (XVECEXP (x, i, j));
2045 }
2046 }
2047 return value;
2048 }
2049
2050 \f
2051 #if 0
2052 /* P is an instruction that sets a register to the result of a ZERO_EXTEND.
2053 Replace it with an instruction to load just the low bytes
2054 if the machine supports such an instruction,
2055 and insert above LOOP_START an instruction to clear the register. */
2056
2057 static void
2058 constant_high_bytes (p, loop_start)
2059 rtx p, loop_start;
2060 {
2061 register rtx new;
2062 register int insn_code_number;
2063
2064 /* Try to change (SET (REG ...) (ZERO_EXTEND (..:B ...)))
2065 to (SET (STRICT_LOW_PART (SUBREG:B (REG...))) ...). */
2066
2067 new = gen_rtx (SET, VOIDmode,
2068 gen_rtx (STRICT_LOW_PART, VOIDmode,
2069 gen_rtx (SUBREG, GET_MODE (XEXP (SET_SRC (PATTERN (p)), 0)),
2070 SET_DEST (PATTERN (p)),
2071 0)),
2072 XEXP (SET_SRC (PATTERN (p)), 0));
2073 insn_code_number = recog (new, p);
2074
2075 if (insn_code_number)
2076 {
2077 register int i;
2078
2079 /* Clear destination register before the loop. */
2080 emit_insn_before (gen_rtx (SET, VOIDmode,
2081 SET_DEST (PATTERN (p)),
2082 const0_rtx),
2083 loop_start);
2084
2085 /* Inside the loop, just load the low part. */
2086 PATTERN (p) = new;
2087 }
2088 }
2089 #endif
2090 \f
2091 /* Scan a loop setting the variables `unknown_address_altered',
2092 `num_mem_sets', `loop_continue', loops_enclosed' and `loop_has_call'.
2093 Also, fill in the array `loop_store_mems'. */
2094
2095 static void
2096 prescan_loop (start, end)
2097 rtx start, end;
2098 {
2099 register int level = 1;
2100 register rtx insn;
2101
2102 unknown_address_altered = 0;
2103 loop_has_call = 0;
2104 loop_store_mems_idx = 0;
2105
2106 num_mem_sets = 0;
2107 loops_enclosed = 1;
2108 loop_continue = 0;
2109
2110 for (insn = NEXT_INSN (start); insn != NEXT_INSN (end);
2111 insn = NEXT_INSN (insn))
2112 {
2113 if (GET_CODE (insn) == NOTE)
2114 {
2115 if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG)
2116 {
2117 ++level;
2118 /* Count number of loops contained in this one. */
2119 loops_enclosed++;
2120 }
2121 else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_END)
2122 {
2123 --level;
2124 if (level == 0)
2125 {
2126 end = insn;
2127 break;
2128 }
2129 }
2130 else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_CONT)
2131 {
2132 if (level == 1)
2133 loop_continue = insn;
2134 }
2135 }
2136 else if (GET_CODE (insn) == CALL_INSN)
2137 {
2138 unknown_address_altered = 1;
2139 loop_has_call = 1;
2140 }
2141 else
2142 {
2143 if (GET_CODE (insn) == INSN || GET_CODE (insn) == JUMP_INSN)
2144 note_stores (PATTERN (insn), note_addr_stored);
2145 }
2146 }
2147 }
2148 \f
2149 /* Scan the function looking for loops. Record the start and end of each loop.
2150 Also mark as invalid loops any loops that contain a setjmp or are branched
2151 to from outside the loop. */
2152
2153 static void
2154 find_and_verify_loops (f)
2155 rtx f;
2156 {
2157 rtx insn;
2158 int current_loop = -1;
2159 int next_loop = -1;
2160 int loop;
2161
2162 /* If there are jumps to undefined labels,
2163 treat them as jumps out of any/all loops.
2164 This also avoids writing past end of tables when there are no loops. */
2165 uid_loop_num[0] = -1;
2166
2167 /* Find boundaries of loops, mark which loops are contained within
2168 loops, and invalidate loops that have setjmp. */
2169
2170 for (insn = f; insn; insn = NEXT_INSN (insn))
2171 {
2172 if (GET_CODE (insn) == NOTE)
2173 switch (NOTE_LINE_NUMBER (insn))
2174 {
2175 case NOTE_INSN_LOOP_BEG:
2176 loop_number_loop_starts[++next_loop] = insn;
2177 loop_number_loop_ends[next_loop] = 0;
2178 loop_outer_loop[next_loop] = current_loop;
2179 loop_invalid[next_loop] = 0;
2180 loop_number_exit_labels[next_loop] = 0;
2181 current_loop = next_loop;
2182 break;
2183
2184 case NOTE_INSN_SETJMP:
2185 /* In this case, we must invalidate our current loop and any
2186 enclosing loop. */
2187 for (loop = current_loop; loop != -1; loop = loop_outer_loop[loop])
2188 {
2189 loop_invalid[loop] = 1;
2190 if (loop_dump_stream)
2191 fprintf (loop_dump_stream,
2192 "\nLoop at %d ignored due to setjmp.\n",
2193 INSN_UID (loop_number_loop_starts[loop]));
2194 }
2195 break;
2196
2197 case NOTE_INSN_LOOP_END:
2198 if (current_loop == -1)
2199 abort ();
2200
2201 loop_number_loop_ends[current_loop] = insn;
2202 current_loop = loop_outer_loop[current_loop];
2203 break;
2204
2205 }
2206
2207 /* Note that this will mark the NOTE_INSN_LOOP_END note as being in the
2208 enclosing loop, but this doesn't matter. */
2209 uid_loop_num[INSN_UID (insn)] = current_loop;
2210 }
2211
2212 /* Now scan all JUMP_INSN's in the function. If any branches into a loop
2213 that it is not contained within, that loop is marked invalid.
2214
2215 Also look for blocks of code ending in an unconditional branch that
2216 exits the loop. If such a block is surrounded by a conditional
2217 branch around the block, move the block elsewhere (see below) and
2218 invert the jump to point to the code block. This may eliminate a
2219 label in our loop and will simplify processing by both us and a
2220 possible second cse pass. */
2221
2222 for (insn = f; insn; insn = NEXT_INSN (insn))
2223 if (GET_CODE (insn) == JUMP_INSN)
2224 {
2225 int this_loop_num = uid_loop_num[INSN_UID (insn)];
2226
2227 mark_loop_jump (PATTERN (insn), this_loop_num);
2228
2229 /* See if this is an unconditional branch outside the loop. */
2230 if (this_loop_num != -1
2231 && (GET_CODE (PATTERN (insn)) == RETURN
2232 || (simplejump_p (insn)
2233 && (uid_loop_num[INSN_UID (JUMP_LABEL (insn))]
2234 != this_loop_num))))
2235 {
2236 rtx p;
2237 rtx our_next = next_real_insn (insn);
2238
2239 /* Go backwards until we reach the start of the loop, a label,
2240 or a JUMP_INSN. */
2241 for (p = PREV_INSN (insn);
2242 GET_CODE (p) != CODE_LABEL
2243 && ! (GET_CODE (p) == NOTE
2244 && NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
2245 && GET_CODE (p) != JUMP_INSN;
2246 p = PREV_INSN (p))
2247 ;
2248
2249 /* If we stopped on a JUMP_INSN to the next insn after INSN,
2250 we have a block of code to try to move.
2251
2252 We look backward and then forward from the target of INSN
2253 to find a BARRIER at the same loop depth as the target.
2254 If we find such a BARRIER, we make a new label for the start
2255 of the block, invert the jump in P and point it to that label,
2256 and move the block of code to the spot we found. */
2257
2258 if (GET_CODE (p) == JUMP_INSN
2259 && JUMP_LABEL (p) != 0
2260 && condjump_p (p)
2261 && ! simplejump_p (p)
2262 && next_real_insn (JUMP_LABEL (p)) == our_next)
2263 {
2264 rtx target
2265 = JUMP_LABEL (insn) ? JUMP_LABEL (insn) : get_last_insn ();
2266 int target_loop_num = uid_loop_num[INSN_UID (target)];
2267 rtx loc;
2268
2269 for (loc = target; loc; loc = PREV_INSN (loc))
2270 if (GET_CODE (loc) == BARRIER
2271 && uid_loop_num[INSN_UID (loc)] == target_loop_num)
2272 break;
2273
2274 if (loc == 0)
2275 for (loc = target; loc; loc = NEXT_INSN (loc))
2276 if (GET_CODE (loc) == BARRIER
2277 && uid_loop_num[INSN_UID (loc)] == target_loop_num)
2278 break;
2279
2280 if (loc)
2281 {
2282 rtx cond_label = JUMP_LABEL (p);
2283 rtx new_label = get_label_after (p);
2284
2285 /* Ensure our label doesn't go away. */
2286 LABEL_NUSES (cond_label)++;
2287
2288 /* Verify that uid_loop_num is large enough and that
2289 we can invert P. */
2290 if (INSN_UID (new_label) < max_uid_for_loop
2291 && invert_jump (p, new_label))
2292 {
2293 rtx q, r;
2294
2295 /* Include the BARRIER after INSN and copy the
2296 block after LOC. */
2297 new_label = squeeze_notes (new_label, NEXT_INSN (insn));
2298 reorder_insns (new_label, NEXT_INSN (insn), loc);
2299
2300 /* All those insns are now in TARGET_LOOP_NUM. */
2301 for (q = new_label; q != NEXT_INSN (NEXT_INSN (insn));
2302 q = NEXT_INSN (q))
2303 uid_loop_num[INSN_UID (q)] = target_loop_num;
2304
2305 /* The label jumped to by INSN is no longer a loop exit.
2306 Unless INSN does not have a label (e.g., it is a
2307 RETURN insn), search loop_number_exit_labels to find
2308 its label_ref, and remove it. Also turn off
2309 LABEL_OUTSIDE_LOOP_P bit. */
2310 if (JUMP_LABEL (insn))
2311 {
2312 for (q = 0,
2313 r = loop_number_exit_labels[this_loop_num];
2314 r; q = r, r = LABEL_NEXTREF (r))
2315 if (XEXP (r, 0) == JUMP_LABEL (insn))
2316 {
2317 LABEL_OUTSIDE_LOOP_P (r) = 0;
2318 if (q)
2319 LABEL_NEXTREF (q) = LABEL_NEXTREF (r);
2320 else
2321 loop_number_exit_labels[this_loop_num]
2322 = LABEL_NEXTREF (r);
2323 break;
2324 }
2325
2326 /* If we didn't find it, then something is wrong. */
2327 if (! r)
2328 abort ();
2329 }
2330
2331 /* P is now a jump outside the loop, so it must be put
2332 in loop_number_exit_labels, and marked as such.
2333 The easiest way to do this is to just call
2334 mark_loop_jump again for P. */
2335 mark_loop_jump (PATTERN (p), this_loop_num);
2336
2337 /* If INSN now jumps to the insn after it,
2338 delete INSN. */
2339 if (JUMP_LABEL (insn) != 0
2340 && (next_real_insn (JUMP_LABEL (insn))
2341 == next_real_insn (insn)))
2342 delete_insn (insn);
2343 }
2344
2345 /* Continue the loop after where the conditional
2346 branch used to jump, since the only branch insn
2347 in the block (if it still remains) is an inter-loop
2348 branch and hence needs no processing. */
2349 insn = NEXT_INSN (cond_label);
2350
2351 if (--LABEL_NUSES (cond_label) == 0)
2352 delete_insn (cond_label);
2353 }
2354 }
2355 }
2356 }
2357 }
2358
2359 /* If any label in X jumps to a loop different from LOOP_NUM and any of the
2360 loops it is contained in, mark the target loop invalid.
2361
2362 For speed, we assume that X is part of a pattern of a JUMP_INSN. */
2363
2364 static void
2365 mark_loop_jump (x, loop_num)
2366 rtx x;
2367 int loop_num;
2368 {
2369 int dest_loop;
2370 int outer_loop;
2371 int i;
2372
2373 switch (GET_CODE (x))
2374 {
2375 case PC:
2376 case USE:
2377 case CLOBBER:
2378 case REG:
2379 case MEM:
2380 case CONST_INT:
2381 case CONST_DOUBLE:
2382 case RETURN:
2383 return;
2384
2385 case CONST:
2386 /* There could be a label reference in here. */
2387 mark_loop_jump (XEXP (x, 0), loop_num);
2388 return;
2389
2390 case PLUS:
2391 case MINUS:
2392 case MULT:
2393 case LSHIFT:
2394 mark_loop_jump (XEXP (x, 0), loop_num);
2395 mark_loop_jump (XEXP (x, 1), loop_num);
2396 return;
2397
2398 case SIGN_EXTEND:
2399 case ZERO_EXTEND:
2400 mark_loop_jump (XEXP (x, 0), loop_num);
2401 return;
2402
2403 case LABEL_REF:
2404 dest_loop = uid_loop_num[INSN_UID (XEXP (x, 0))];
2405
2406 /* Link together all labels that branch outside the loop. This
2407 is used by final_[bg]iv_value and the loop unrolling code. Also
2408 mark this LABEL_REF so we know that this branch should predict
2409 false. */
2410
2411 if (dest_loop != loop_num && loop_num != -1)
2412 {
2413 LABEL_OUTSIDE_LOOP_P (x) = 1;
2414 LABEL_NEXTREF (x) = loop_number_exit_labels[loop_num];
2415 loop_number_exit_labels[loop_num] = x;
2416 }
2417
2418 /* If this is inside a loop, but not in the current loop or one enclosed
2419 by it, it invalidates at least one loop. */
2420
2421 if (dest_loop == -1)
2422 return;
2423
2424 /* We must invalidate every nested loop containing the target of this
2425 label, except those that also contain the jump insn. */
2426
2427 for (; dest_loop != -1; dest_loop = loop_outer_loop[dest_loop])
2428 {
2429 /* Stop when we reach a loop that also contains the jump insn. */
2430 for (outer_loop = loop_num; outer_loop != -1;
2431 outer_loop = loop_outer_loop[outer_loop])
2432 if (dest_loop == outer_loop)
2433 return;
2434
2435 /* If we get here, we know we need to invalidate a loop. */
2436 if (loop_dump_stream && ! loop_invalid[dest_loop])
2437 fprintf (loop_dump_stream,
2438 "\nLoop at %d ignored due to multiple entry points.\n",
2439 INSN_UID (loop_number_loop_starts[dest_loop]));
2440
2441 loop_invalid[dest_loop] = 1;
2442 }
2443 return;
2444
2445 case SET:
2446 /* If this is not setting pc, ignore. */
2447 if (SET_DEST (x) == pc_rtx)
2448 mark_loop_jump (SET_SRC (x), loop_num);
2449 return;
2450
2451 case IF_THEN_ELSE:
2452 mark_loop_jump (XEXP (x, 1), loop_num);
2453 mark_loop_jump (XEXP (x, 2), loop_num);
2454 return;
2455
2456 case PARALLEL:
2457 case ADDR_VEC:
2458 for (i = 0; i < XVECLEN (x, 0); i++)
2459 mark_loop_jump (XVECEXP (x, 0, i), loop_num);
2460 return;
2461
2462 case ADDR_DIFF_VEC:
2463 for (i = 0; i < XVECLEN (x, 1); i++)
2464 mark_loop_jump (XVECEXP (x, 1, i), loop_num);
2465 return;
2466
2467 default:
2468 /* Nothing else should occur in a JUMP_INSN. */
2469 abort ();
2470 }
2471 }
2472 \f
2473 /* Return nonzero if there is a label in the range from
2474 insn INSN to and including the insn whose luid is END
2475 INSN must have an assigned luid (i.e., it must not have
2476 been previously created by loop.c). */
2477
2478 static int
2479 labels_in_range_p (insn, end)
2480 rtx insn;
2481 int end;
2482 {
2483 while (insn && INSN_LUID (insn) <= end)
2484 {
2485 if (GET_CODE (insn) == CODE_LABEL)
2486 return 1;
2487 insn = NEXT_INSN (insn);
2488 }
2489
2490 return 0;
2491 }
2492
2493 /* Record that a memory reference X is being set. */
2494
2495 static void
2496 note_addr_stored (x)
2497 rtx x;
2498 {
2499 register int i;
2500
2501 if (x == 0 || GET_CODE (x) != MEM)
2502 return;
2503
2504 /* Count number of memory writes.
2505 This affects heuristics in strength_reduce. */
2506 num_mem_sets++;
2507
2508 if (unknown_address_altered)
2509 return;
2510
2511 for (i = 0; i < loop_store_mems_idx; i++)
2512 if (rtx_equal_p (XEXP (loop_store_mems[i], 0), XEXP (x, 0))
2513 && MEM_IN_STRUCT_P (x) == MEM_IN_STRUCT_P (loop_store_mems[i]))
2514 {
2515 /* We are storing at the same address as previously noted. Save the
2516 wider reference, treating BLKmode as wider. */
2517 if (GET_MODE (x) == BLKmode
2518 || (GET_MODE_SIZE (GET_MODE (x))
2519 > GET_MODE_SIZE (GET_MODE (loop_store_mems[i]))))
2520 loop_store_mems[i] = x;
2521 break;
2522 }
2523
2524 if (i == NUM_STORES)
2525 unknown_address_altered = 1;
2526
2527 else if (i == loop_store_mems_idx)
2528 loop_store_mems[loop_store_mems_idx++] = x;
2529 }
2530 \f
2531 /* Return nonzero if the rtx X is invariant over the current loop.
2532
2533 The value is 2 if we refer to something only conditionally invariant.
2534
2535 If `unknown_address_altered' is nonzero, no memory ref is invariant.
2536 Otherwise, a memory ref is invariant if it does not conflict with
2537 anything stored in `loop_store_mems'. */
2538
2539 int
2540 invariant_p (x)
2541 register rtx x;
2542 {
2543 register int i;
2544 register enum rtx_code code;
2545 register char *fmt;
2546 int conditional = 0;
2547
2548 if (x == 0)
2549 return 1;
2550 code = GET_CODE (x);
2551 switch (code)
2552 {
2553 case CONST_INT:
2554 case CONST_DOUBLE:
2555 case SYMBOL_REF:
2556 case CONST:
2557 return 1;
2558
2559 case LABEL_REF:
2560 /* A LABEL_REF is normally invariant, however, if we are unrolling
2561 loops, and this label is inside the loop, then it isn't invariant.
2562 This is because each unrolled copy of the loop body will have
2563 a copy of this label. If this was invariant, then an insn loading
2564 the address of this label into a register might get moved outside
2565 the loop, and then each loop body would end up using the same label.
2566
2567 We don't know the loop bounds here though, so just fail for all
2568 labels. */
2569 if (flag_unroll_loops)
2570 return 0;
2571 else
2572 return 1;
2573
2574 case PC:
2575 case CC0:
2576 case UNSPEC_VOLATILE:
2577 return 0;
2578
2579 case REG:
2580 /* We used to check RTX_UNCHANGING_P (x) here, but that is invalid
2581 since the reg might be set by initialization within the loop. */
2582 if (x == frame_pointer_rtx || x == arg_pointer_rtx)
2583 return 1;
2584 if (loop_has_call
2585 && REGNO (x) < FIRST_PSEUDO_REGISTER && call_used_regs[REGNO (x)])
2586 return 0;
2587 if (n_times_set[REGNO (x)] < 0)
2588 return 2;
2589 return n_times_set[REGNO (x)] == 0;
2590
2591 case MEM:
2592 /* Read-only items (such as constants in a constant pool) are
2593 invariant if their address is. */
2594 if (RTX_UNCHANGING_P (x))
2595 break;
2596
2597 /* If we filled the table (or had a subroutine call), any location
2598 in memory could have been clobbered. */
2599 if (unknown_address_altered
2600 /* Don't mess with volatile memory references. */
2601 || MEM_VOLATILE_P (x))
2602 return 0;
2603
2604 /* See if there is any dependence between a store and this load. */
2605 for (i = loop_store_mems_idx - 1; i >= 0; i--)
2606 if (true_dependence (loop_store_mems[i], x))
2607 return 0;
2608
2609 /* It's not invalidated by a store in memory
2610 but we must still verify the address is invariant. */
2611 break;
2612
2613 case ASM_OPERANDS:
2614 /* Don't mess with insns declared volatile. */
2615 if (MEM_VOLATILE_P (x))
2616 return 0;
2617 }
2618
2619 fmt = GET_RTX_FORMAT (code);
2620 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2621 {
2622 if (fmt[i] == 'e')
2623 {
2624 int tem = invariant_p (XEXP (x, i));
2625 if (tem == 0)
2626 return 0;
2627 if (tem == 2)
2628 conditional = 1;
2629 }
2630 else if (fmt[i] == 'E')
2631 {
2632 register int j;
2633 for (j = 0; j < XVECLEN (x, i); j++)
2634 {
2635 int tem = invariant_p (XVECEXP (x, i, j));
2636 if (tem == 0)
2637 return 0;
2638 if (tem == 2)
2639 conditional = 1;
2640 }
2641
2642 }
2643 }
2644
2645 return 1 + conditional;
2646 }
2647
2648 /* Return 1 if OTHER (a mem ref) overlaps the area of memory
2649 which is SIZE bytes starting at BASE. */
2650
2651 int
2652 addr_overlap_p (other, base, size)
2653 rtx other;
2654 rtx base;
2655 int size;
2656 {
2657 int start = 0, end;
2658
2659 if (GET_CODE (base) == CONST)
2660 base = XEXP (base, 0);
2661 if (GET_CODE (base) == PLUS
2662 && GET_CODE (XEXP (base, 1)) == CONST_INT)
2663 {
2664 start = INTVAL (XEXP (base, 1));
2665 base = XEXP (base, 0);
2666 }
2667
2668 end = start + size;
2669 return refers_to_mem_p (other, base, start, end);
2670 }
2671 \f
2672 /* Return nonzero if all the insns in the loop that set REG
2673 are INSN and the immediately following insns,
2674 and if each of those insns sets REG in an invariant way
2675 (not counting uses of REG in them).
2676
2677 The value is 2 if some of these insns are only conditionally invariant.
2678
2679 We assume that INSN itself is the first set of REG
2680 and that its source is invariant. */
2681
2682 static int
2683 consec_sets_invariant_p (reg, n_sets, insn)
2684 int n_sets;
2685 rtx reg, insn;
2686 {
2687 register rtx p = insn;
2688 register int regno = REGNO (reg);
2689 rtx temp;
2690 /* Number of sets we have to insist on finding after INSN. */
2691 int count = n_sets - 1;
2692 int old = n_times_set[regno];
2693 int value = 0;
2694 int this;
2695
2696 /* If N_SETS hit the limit, we can't rely on its value. */
2697 if (n_sets == 127)
2698 return 0;
2699
2700 n_times_set[regno] = 0;
2701
2702 while (count > 0)
2703 {
2704 register enum rtx_code code;
2705 rtx set;
2706
2707 p = NEXT_INSN (p);
2708 code = GET_CODE (p);
2709
2710 /* If library call, skip to end of of it. */
2711 if (code == INSN && (temp = find_reg_note (p, REG_LIBCALL, 0)))
2712 p = XEXP (temp, 0);
2713
2714 this = 0;
2715 if (code == INSN
2716 && (set = single_set (p))
2717 && GET_CODE (SET_DEST (set)) == REG
2718 && REGNO (SET_DEST (set)) == regno)
2719 {
2720 this = invariant_p (SET_SRC (set));
2721 if (this != 0)
2722 value |= this;
2723 else if (temp = find_reg_note (p, REG_EQUAL, 0))
2724 {
2725 this = invariant_p (XEXP (temp, 0));
2726 if (this != 0)
2727 value |= this;
2728 }
2729 }
2730 if (this != 0)
2731 count--;
2732 else if (code != NOTE)
2733 {
2734 n_times_set[regno] = old;
2735 return 0;
2736 }
2737 }
2738
2739 n_times_set[regno] = old;
2740 /* If invariant_p ever returned 2, we return 2. */
2741 return 1 + (value & 2);
2742 }
2743
2744 #if 0
2745 /* I don't think this condition is sufficient to allow INSN
2746 to be moved, so we no longer test it. */
2747
2748 /* Return 1 if all insns in the basic block of INSN and following INSN
2749 that set REG are invariant according to TABLE. */
2750
2751 static int
2752 all_sets_invariant_p (reg, insn, table)
2753 rtx reg, insn;
2754 short *table;
2755 {
2756 register rtx p = insn;
2757 register int regno = REGNO (reg);
2758
2759 while (1)
2760 {
2761 register enum rtx_code code;
2762 p = NEXT_INSN (p);
2763 code = GET_CODE (p);
2764 if (code == CODE_LABEL || code == JUMP_INSN)
2765 return 1;
2766 if (code == INSN && GET_CODE (PATTERN (p)) == SET
2767 && GET_CODE (SET_DEST (PATTERN (p))) == REG
2768 && REGNO (SET_DEST (PATTERN (p))) == regno)
2769 {
2770 if (!invariant_p (SET_SRC (PATTERN (p)), table))
2771 return 0;
2772 }
2773 }
2774 }
2775 #endif /* 0 */
2776 \f
2777 /* Look at all uses (not sets) of registers in X. For each, if it is
2778 the single use, set USAGE[REGNO] to INSN; if there was a previous use in
2779 a different insn, set USAGE[REGNO] to const0_rtx. */
2780
2781 static void
2782 find_single_use_in_loop (insn, x, usage)
2783 rtx insn;
2784 rtx x;
2785 rtx *usage;
2786 {
2787 enum rtx_code code = GET_CODE (x);
2788 char *fmt = GET_RTX_FORMAT (code);
2789 int i, j;
2790
2791 if (code == REG)
2792 usage[REGNO (x)]
2793 = (usage[REGNO (x)] != 0 && usage[REGNO (x)] != insn)
2794 ? const0_rtx : insn;
2795
2796 else if (code == SET)
2797 {
2798 /* Don't count SET_DEST if it is a REG; otherwise count things
2799 in SET_DEST because if a register is partially modified, it won't
2800 show up as a potential movable so we don't care how USAGE is set
2801 for it. */
2802 if (GET_CODE (SET_DEST (x)) != REG)
2803 find_single_use_in_loop (insn, SET_DEST (x), usage);
2804 find_single_use_in_loop (insn, SET_SRC (x), usage);
2805 }
2806 else
2807 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2808 {
2809 if (fmt[i] == 'e' && XEXP (x, i) != 0)
2810 find_single_use_in_loop (insn, XEXP (x, i), usage);
2811 else if (fmt[i] == 'E')
2812 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
2813 find_single_use_in_loop (insn, XVECEXP (x, i, j), usage);
2814 }
2815 }
2816 \f
2817 /* Increment N_TIMES_SET at the index of each register
2818 that is modified by an insn between FROM and TO.
2819 If the value of an element of N_TIMES_SET becomes 127 or more,
2820 stop incrementing it, to avoid overflow.
2821
2822 Store in SINGLE_USAGE[I] the single insn in which register I is
2823 used, if it is only used once. Otherwise, it is set to 0 (for no
2824 uses) or const0_rtx for more than one use. This parameter may be zero,
2825 in which case this processing is not done.
2826
2827 Store in *COUNT_PTR the number of actual instruction
2828 in the loop. We use this to decide what is worth moving out. */
2829
2830 /* last_set[n] is nonzero iff reg n has been set in the current basic block.
2831 In that case, it is the insn that last set reg n. */
2832
2833 static void
2834 count_loop_regs_set (from, to, may_not_move, single_usage, count_ptr, nregs)
2835 register rtx from, to;
2836 char *may_not_move;
2837 rtx *single_usage;
2838 int *count_ptr;
2839 int nregs;
2840 {
2841 register rtx *last_set = (rtx *) alloca (nregs * sizeof (rtx));
2842 register rtx insn;
2843 register int count = 0;
2844 register rtx dest;
2845
2846 bzero (last_set, nregs * sizeof (rtx));
2847 for (insn = from; insn != to; insn = NEXT_INSN (insn))
2848 {
2849 if (GET_RTX_CLASS (GET_CODE (insn)) == 'i')
2850 {
2851 ++count;
2852
2853 /* If requested, record registers that have exactly one use. */
2854 if (single_usage)
2855 {
2856 find_single_use_in_loop (insn, PATTERN (insn), single_usage);
2857
2858 /* Include uses in REG_EQUAL notes. */
2859 if (REG_NOTES (insn))
2860 find_single_use_in_loop (insn, REG_NOTES (insn), single_usage);
2861 }
2862
2863 if (GET_CODE (PATTERN (insn)) == CLOBBER
2864 && GET_CODE (XEXP (PATTERN (insn), 0)) == REG)
2865 /* Don't move a reg that has an explicit clobber.
2866 We might do so sometimes, but it's not worth the pain. */
2867 may_not_move[REGNO (XEXP (PATTERN (insn), 0))] = 1;
2868
2869 if (GET_CODE (PATTERN (insn)) == SET
2870 || GET_CODE (PATTERN (insn)) == CLOBBER)
2871 {
2872 dest = SET_DEST (PATTERN (insn));
2873 while (GET_CODE (dest) == SUBREG
2874 || GET_CODE (dest) == ZERO_EXTRACT
2875 || GET_CODE (dest) == SIGN_EXTRACT
2876 || GET_CODE (dest) == STRICT_LOW_PART)
2877 dest = XEXP (dest, 0);
2878 if (GET_CODE (dest) == REG)
2879 {
2880 register int regno = REGNO (dest);
2881 /* If this is the first setting of this reg
2882 in current basic block, and it was set before,
2883 it must be set in two basic blocks, so it cannot
2884 be moved out of the loop. */
2885 if (n_times_set[regno] > 0 && last_set[regno] == 0)
2886 may_not_move[regno] = 1;
2887 /* If this is not first setting in current basic block,
2888 see if reg was used in between previous one and this.
2889 If so, neither one can be moved. */
2890 if (last_set[regno] != 0
2891 && reg_used_between_p (dest, last_set[regno], insn))
2892 may_not_move[regno] = 1;
2893 if (n_times_set[regno] < 127)
2894 ++n_times_set[regno];
2895 last_set[regno] = insn;
2896 }
2897 }
2898 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
2899 {
2900 register int i;
2901 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
2902 {
2903 register rtx x = XVECEXP (PATTERN (insn), 0, i);
2904 if (GET_CODE (x) == CLOBBER && GET_CODE (XEXP (x, 0)) == REG)
2905 /* Don't move a reg that has an explicit clobber.
2906 It's not worth the pain to try to do it correctly. */
2907 may_not_move[REGNO (XEXP (x, 0))] = 1;
2908
2909 if (GET_CODE (x) == SET || GET_CODE (x) == CLOBBER)
2910 {
2911 dest = SET_DEST (x);
2912 while (GET_CODE (dest) == SUBREG
2913 || GET_CODE (dest) == ZERO_EXTRACT
2914 || GET_CODE (dest) == SIGN_EXTRACT
2915 || GET_CODE (dest) == STRICT_LOW_PART)
2916 dest = XEXP (dest, 0);
2917 if (GET_CODE (dest) == REG)
2918 {
2919 register int regno = REGNO (dest);
2920 if (n_times_set[regno] > 0 && last_set[regno] == 0)
2921 may_not_move[regno] = 1;
2922 if (last_set[regno] != 0
2923 && reg_used_between_p (dest, last_set[regno], insn))
2924 may_not_move[regno] = 1;
2925 if (n_times_set[regno] < 127)
2926 ++n_times_set[regno];
2927 last_set[regno] = insn;
2928 }
2929 }
2930 }
2931 }
2932 }
2933 if (GET_CODE (insn) == CODE_LABEL || GET_CODE (insn) == JUMP_INSN)
2934 bzero (last_set, nregs * sizeof (rtx));
2935 }
2936 *count_ptr = count;
2937 }
2938 \f
2939 /* Given a loop that is bounded by LOOP_START and LOOP_END
2940 and that is entered at SCAN_START,
2941 return 1 if the register set in SET contained in insn INSN is used by
2942 any insn that precedes INSN in cyclic order starting
2943 from the loop entry point.
2944
2945 We don't want to use INSN_LUID here because if we restrict INSN to those
2946 that have a valid INSN_LUID, it means we cannot move an invariant out
2947 from an inner loop past two loops. */
2948
2949 static int
2950 loop_reg_used_before_p (set, insn, loop_start, scan_start, loop_end)
2951 rtx set, insn, loop_start, scan_start, loop_end;
2952 {
2953 rtx reg = SET_DEST (set);
2954 rtx p;
2955
2956 /* Scan forward checking for register usage. If we hit INSN, we
2957 are done. Otherwise, if we hit LOOP_END, wrap around to LOOP_START. */
2958 for (p = scan_start; p != insn; p = NEXT_INSN (p))
2959 {
2960 if (GET_RTX_CLASS (GET_CODE (p)) == 'i'
2961 && reg_overlap_mentioned_p (reg, PATTERN (p)))
2962 return 1;
2963
2964 if (p == loop_end)
2965 p = loop_start;
2966 }
2967
2968 return 0;
2969 }
2970 \f
2971 /* A "basic induction variable" or biv is a pseudo reg that is set
2972 (within this loop) only by incrementing or decrementing it. */
2973 /* A "general induction variable" or giv is a pseudo reg whose
2974 value is a linear function of a biv. */
2975
2976 /* Bivs are recognized by `basic_induction_var';
2977 Givs by `general_induct_var'. */
2978
2979 /* Indexed by register number, indicates whether or not register is an
2980 induction variable, and if so what type. */
2981
2982 enum iv_mode *reg_iv_type;
2983
2984 /* Indexed by register number, contains pointer to `struct induction'
2985 if register is an induction variable. This holds general info for
2986 all induction variables. */
2987
2988 struct induction **reg_iv_info;
2989
2990 /* Indexed by register number, contains pointer to `struct iv_class'
2991 if register is a basic induction variable. This holds info describing
2992 the class (a related group) of induction variables that the biv belongs
2993 to. */
2994
2995 struct iv_class **reg_biv_class;
2996
2997 /* The head of a list which links together (via the next field)
2998 every iv class for the current loop. */
2999
3000 struct iv_class *loop_iv_list;
3001
3002 /* Communication with routines called via `note_stores'. */
3003
3004 static rtx note_insn;
3005
3006 /* Dummy register to have non-zero DEST_REG for DEST_ADDR type givs. */
3007
3008 static rtx addr_placeholder;
3009
3010 /* ??? Unfinished optimizations, and possible future optimizations,
3011 for the strength reduction code. */
3012
3013 /* ??? There is one more optimization you might be interested in doing: to
3014 allocate pseudo registers for frequently-accessed memory locations.
3015 If the same memory location is referenced each time around, it might
3016 be possible to copy it into a register before and out after.
3017 This is especially useful when the memory location is a variable which
3018 is in a stack slot because somewhere its address is taken. If the
3019 loop doesn't contain a function call and the variable isn't volatile,
3020 it is safe to keep the value in a register for the duration of the
3021 loop. One tricky thing is that the copying of the value back from the
3022 register has to be done on all exits from the loop. You need to check that
3023 all the exits from the loop go to the same place. */
3024
3025 /* ??? The interaction of biv elimination, and recognition of 'constant'
3026 bivs, may cause problems. */
3027
3028 /* ??? Add heuristics so that DEST_ADDR strength reduction does not cause
3029 performance problems.
3030
3031 Perhaps don't eliminate things that can be combined with an addressing
3032 mode. Find all givs that have the same biv, mult_val, and add_val;
3033 then for each giv, check to see if its only use dies in a following
3034 memory address. If so, generate a new memory address and check to see
3035 if it is valid. If it is valid, then store the modified memory address,
3036 otherwise, mark the giv as not done so that it will get its own iv. */
3037
3038 /* ??? Could try to optimize branches when it is known that a biv is always
3039 positive. */
3040
3041 /* ??? When replace a biv in a compare insn, we should replace with closest
3042 giv so that an optimized branch can still be recognized by the combiner,
3043 e.g. the VAX acb insn. */
3044
3045 /* ??? Many of the checks involving uid_luid could be simplified if regscan
3046 was rerun in loop_optimize whenever a register was added or moved.
3047 Also, some of the optimizations could be a little less conservative. */
3048 \f
3049 /* Perform strength reduction and induction variable elimination. */
3050
3051 /* Pseudo registers created during this function will be beyond the last
3052 valid index in several tables including n_times_set and regno_last_uid.
3053 This does not cause a problem here, because the added registers cannot be
3054 givs outside of their loop, and hence will never be reconsidered.
3055 But scan_loop must check regnos to make sure they are in bounds. */
3056
3057 static void
3058 strength_reduce (scan_start, end, loop_top, insn_count,
3059 loop_start, loop_end)
3060 rtx scan_start;
3061 rtx end;
3062 rtx loop_top;
3063 int insn_count;
3064 rtx loop_start;
3065 rtx loop_end;
3066 {
3067 rtx p;
3068 rtx set;
3069 rtx inc_val;
3070 rtx mult_val;
3071 rtx dest_reg;
3072 /* This is 1 if current insn is not executed at least once for every loop
3073 iteration. */
3074 int not_every_iteration = 0;
3075 /* This is 1 if current insn may be executed more than once for every
3076 loop iteration. */
3077 int maybe_multiple = 0;
3078 /* Temporary list pointers for traversing loop_iv_list. */
3079 struct iv_class *bl, **backbl;
3080 /* Ratio of extra register life span we can justify
3081 for saving an instruction. More if loop doesn't call subroutines
3082 since in that case saving an insn makes more difference
3083 and more registers are available. */
3084 /* ??? could set this to last value of threshold in move_movables */
3085 int threshold = (loop_has_call ? 1 : 2) * (3 + n_non_fixed_regs);
3086 /* Map of pseudo-register replacements. */
3087 rtx *reg_map;
3088 int call_seen;
3089 rtx test;
3090 rtx end_insert_before;
3091
3092 reg_iv_type = (enum iv_mode *) alloca (max_reg_before_loop
3093 * sizeof (enum iv_mode *));
3094 bzero ((char *) reg_iv_type, max_reg_before_loop * sizeof (enum iv_mode *));
3095 reg_iv_info = (struct induction **)
3096 alloca (max_reg_before_loop * sizeof (struct induction *));
3097 bzero ((char *) reg_iv_info, (max_reg_before_loop
3098 * sizeof (struct induction *)));
3099 reg_biv_class = (struct iv_class **)
3100 alloca (max_reg_before_loop * sizeof (struct iv_class *));
3101 bzero ((char *) reg_biv_class, (max_reg_before_loop
3102 * sizeof (struct iv_class *)));
3103
3104 loop_iv_list = 0;
3105 addr_placeholder = gen_reg_rtx (Pmode);
3106
3107 /* Save insn immediately after the loop_end. Insns inserted after loop_end
3108 must be put before this insn, so that they will appear in the right
3109 order (i.e. loop order). */
3110
3111 end_insert_before = NEXT_INSN (loop_end);
3112
3113 /* Scan through loop to find all possible bivs. */
3114
3115 p = scan_start;
3116 while (1)
3117 {
3118 p = NEXT_INSN (p);
3119 /* At end of a straight-in loop, we are done.
3120 At end of a loop entered at the bottom, scan the top. */
3121 if (p == scan_start)
3122 break;
3123 if (p == end)
3124 {
3125 if (loop_top != 0)
3126 p = NEXT_INSN (loop_top);
3127 else
3128 break;
3129 if (p == scan_start)
3130 break;
3131 }
3132
3133 if (GET_CODE (p) == INSN
3134 && (set = single_set (p))
3135 && GET_CODE (SET_DEST (set)) == REG)
3136 {
3137 dest_reg = SET_DEST (set);
3138 if (REGNO (dest_reg) < max_reg_before_loop
3139 && REGNO (dest_reg) >= FIRST_PSEUDO_REGISTER
3140 && reg_iv_type[REGNO (dest_reg)] != NOT_BASIC_INDUCT)
3141 {
3142 if (basic_induction_var (SET_SRC (set), dest_reg,
3143 &inc_val, &mult_val))
3144 {
3145 /* It is a possible basic induction variable.
3146 Create and initialize an induction structure for it. */
3147
3148 struct induction *v
3149 = (struct induction *) alloca (sizeof (struct induction));
3150
3151 record_biv (v, p, dest_reg, inc_val, mult_val,
3152 not_every_iteration, maybe_multiple);
3153 reg_iv_type[REGNO (dest_reg)] = BASIC_INDUCT;
3154 }
3155 else if (REGNO (dest_reg) < max_reg_before_loop)
3156 reg_iv_type[REGNO (dest_reg)] = NOT_BASIC_INDUCT;
3157 }
3158 }
3159
3160 /* Past CODE_LABEL, we get to insns that may be executed multiple
3161 times. The only way we can be sure that they can't is if every
3162 every jump insn between here and the end of the loop either
3163 returns, exits the loop, or is a forward jump. */
3164
3165 if (GET_CODE (p) == CODE_LABEL)
3166 {
3167 rtx insn = p;
3168
3169 maybe_multiple = 0;
3170
3171 while (1)
3172 {
3173 insn = NEXT_INSN (insn);
3174 if (insn == scan_start)
3175 break;
3176 if (insn == end)
3177 {
3178 if (loop_top != 0)
3179 insn = NEXT_INSN (loop_top);
3180 else
3181 break;
3182 if (insn == scan_start)
3183 break;
3184 }
3185
3186 if (GET_CODE (insn) == JUMP_INSN
3187 && GET_CODE (PATTERN (insn)) != RETURN
3188 && (! condjump_p (insn)
3189 || (JUMP_LABEL (insn) != 0
3190 && (INSN_UID (JUMP_LABEL (insn)) >= max_uid_for_loop
3191 || INSN_UID (insn) >= max_uid_for_loop
3192 || (INSN_LUID (JUMP_LABEL (insn))
3193 < INSN_LUID (insn))))))
3194 {
3195 maybe_multiple = 1;
3196 break;
3197 }
3198 }
3199 }
3200
3201 /* Past a label or a jump, we get to insns for which we can't count
3202 on whether or how many times they will be executed during each
3203 iteration. */
3204 /* This code appears in three places, once in scan_loop, and twice
3205 in strength_reduce. */
3206 if ((GET_CODE (p) == CODE_LABEL || GET_CODE (p) == JUMP_INSN)
3207 /* If we enter the loop in the middle, and scan around to the
3208 beginning, don't set not_every_iteration for that.
3209 This can be any kind of jump, since we want to know if insns
3210 will be executed if the loop is executed. */
3211 && ! (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) == loop_top
3212 && ((NEXT_INSN (NEXT_INSN (p)) == loop_end && simplejump_p (p))
3213 || (NEXT_INSN (p) == loop_end && condjump_p (p)))))
3214 not_every_iteration = 1;
3215
3216 /* At the virtual top of a converted loop, insns are again known to
3217 be executed each iteration: logically, the loop begins here
3218 even though the exit code has been duplicated. */
3219
3220 else if (GET_CODE (p) == NOTE
3221 && NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP)
3222 not_every_iteration = 0;
3223
3224 /* Unlike in the code motion pass where MAYBE_NEVER indicates that
3225 an insn may never be executed, NOT_EVERY_ITERATION indicates whether
3226 or not an insn is known to be executed each iteration of the
3227 loop, whether or not any iterations are known to occur.
3228
3229 Therefore, if we have just passed a label and have no more labels
3230 between here and the test insn of the loop, we know these insns
3231 will be executed each iteration. This can also happen if we
3232 have just passed a jump, for example, when there are nested loops. */
3233
3234 if (not_every_iteration && GET_CODE (p) == CODE_LABEL
3235 && no_labels_between_p (p, loop_end))
3236 not_every_iteration = 0;
3237 }
3238
3239 /* Scan loop_iv_list to remove all regs that proved not to be bivs.
3240 Make a sanity check against n_times_set. */
3241 for (backbl = &loop_iv_list, bl = *backbl; bl; bl = bl->next)
3242 {
3243 if (reg_iv_type[bl->regno] != BASIC_INDUCT
3244 /* Above happens if register modified by subreg, etc. */
3245 /* Make sure it is not recognized as a basic induction var: */
3246 || n_times_set[bl->regno] != bl->biv_count
3247 /* If never incremented, it is invariant that we decided not to
3248 move. So leave it alone. */
3249 || ! bl->incremented)
3250 {
3251 if (loop_dump_stream)
3252 fprintf (loop_dump_stream, "Reg %d: biv discarded, %s\n",
3253 bl->regno,
3254 (reg_iv_type[bl->regno] != BASIC_INDUCT
3255 ? "not induction variable"
3256 : (! bl->incremented ? "never incremented"
3257 : "count error")));
3258
3259 reg_iv_type[bl->regno] = NOT_BASIC_INDUCT;
3260 *backbl = bl->next;
3261 }
3262 else
3263 {
3264 backbl = &bl->next;
3265
3266 if (loop_dump_stream)
3267 fprintf (loop_dump_stream, "Reg %d: biv verified\n", bl->regno);
3268 }
3269 }
3270
3271 /* Exit if there are no bivs. */
3272 if (! loop_iv_list)
3273 {
3274 /* Can still unroll the loop anyways, but indicate that there is no
3275 strength reduction info available. */
3276 if (flag_unroll_loops)
3277 unroll_loop (loop_end, insn_count, loop_start, end_insert_before, 0);
3278
3279 return;
3280 }
3281
3282 /* Find initial value for each biv by searching backwards from loop_start,
3283 halting at first label. Also record any test condition. */
3284
3285 call_seen = 0;
3286 for (p = loop_start; p && GET_CODE (p) != CODE_LABEL; p = PREV_INSN (p))
3287 {
3288 note_insn = p;
3289
3290 if (GET_CODE (p) == CALL_INSN)
3291 call_seen = 1;
3292
3293 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
3294 || GET_CODE (p) == CALL_INSN)
3295 note_stores (PATTERN (p), record_initial);
3296
3297 /* Record any test of a biv that branches around the loop if no store
3298 between it and the start of loop. We only care about tests with
3299 constants and registers and only certain of those. */
3300 if (GET_CODE (p) == JUMP_INSN
3301 && JUMP_LABEL (p) != 0
3302 && next_real_insn (JUMP_LABEL (p)) == next_real_insn (loop_end)
3303 && (test = get_condition_for_loop (p)) != 0
3304 && GET_CODE (XEXP (test, 0)) == REG
3305 && REGNO (XEXP (test, 0)) < max_reg_before_loop
3306 && (bl = reg_biv_class[REGNO (XEXP (test, 0))]) != 0
3307 && valid_initial_value_p (XEXP (test, 1), p, call_seen, loop_start)
3308 && bl->init_insn == 0)
3309 {
3310 /* If an NE test, we have an initial value! */
3311 if (GET_CODE (test) == NE)
3312 {
3313 bl->init_insn = p;
3314 bl->init_set = gen_rtx (SET, VOIDmode,
3315 XEXP (test, 0), XEXP (test, 1));
3316 }
3317 else
3318 bl->initial_test = test;
3319 }
3320 }
3321
3322 /* Look at the each biv and see if we can say anything better about its
3323 initial value from any initializing insns set up above. (This is done
3324 in two passes to avoid missing SETs in a PARALLEL.) */
3325 for (bl = loop_iv_list; bl; bl = bl->next)
3326 {
3327 rtx src;
3328
3329 if (! bl->init_insn)
3330 continue;
3331
3332 src = SET_SRC (bl->init_set);
3333
3334 if (loop_dump_stream)
3335 fprintf (loop_dump_stream,
3336 "Biv %d initialized at insn %d: initial value ",
3337 bl->regno, INSN_UID (bl->init_insn));
3338
3339 if (valid_initial_value_p (src, bl->init_insn, call_seen, loop_start))
3340 {
3341 bl->initial_value = src;
3342
3343 if (loop_dump_stream)
3344 {
3345 if (GET_CODE (src) == CONST_INT)
3346 fprintf (loop_dump_stream, "%d\n", INTVAL (src));
3347 else
3348 {
3349 print_rtl (loop_dump_stream, src);
3350 fprintf (loop_dump_stream, "\n");
3351 }
3352 }
3353 }
3354 else
3355 {
3356 /* Biv initial value is not simple move,
3357 so let it keep initial value of "itself". */
3358
3359 if (loop_dump_stream)
3360 fprintf (loop_dump_stream, "is complex\n");
3361 }
3362 }
3363
3364 /* Search the loop for general induction variables. */
3365
3366 /* A register is a giv if: it is only set once, it is a function of a
3367 biv and a constant (or invariant), and it is not a biv. */
3368
3369 not_every_iteration = 0;
3370 p = scan_start;
3371 while (1)
3372 {
3373 p = NEXT_INSN (p);
3374 /* At end of a straight-in loop, we are done.
3375 At end of a loop entered at the bottom, scan the top. */
3376 if (p == scan_start)
3377 break;
3378 if (p == end)
3379 {
3380 if (loop_top != 0)
3381 p = NEXT_INSN (loop_top);
3382 else
3383 break;
3384 if (p == scan_start)
3385 break;
3386 }
3387
3388 /* Look for a general induction variable in a register. */
3389 if (GET_CODE (p) == INSN
3390 && (set = single_set (p))
3391 && GET_CODE (SET_DEST (set)) == REG
3392 && ! may_not_optimize[REGNO (SET_DEST (set))])
3393 {
3394 rtx src_reg;
3395 rtx add_val;
3396 rtx mult_val;
3397 int benefit;
3398 rtx regnote = 0;
3399
3400 dest_reg = SET_DEST (set);
3401 if (REGNO (dest_reg) < FIRST_PSEUDO_REGISTER)
3402 continue;
3403
3404 if (/* SET_SRC is a giv. */
3405 ((benefit = general_induction_var (SET_SRC (set),
3406 &src_reg, &add_val,
3407 &mult_val))
3408 /* Equivalent expression is a giv. */
3409 || ((regnote = find_reg_note (p, REG_EQUAL, 0))
3410 && (benefit = general_induction_var (XEXP (regnote, 0),
3411 &src_reg,
3412 &add_val, &mult_val))))
3413 /* Don't try to handle any regs made by loop optimization.
3414 We have nothing on them in regno_first_uid, etc. */
3415 && REGNO (dest_reg) < max_reg_before_loop
3416 /* Don't recognize a BASIC_INDUCT_VAR here. */
3417 && dest_reg != src_reg
3418 /* This must be the only place where the register is set. */
3419 && (n_times_set[REGNO (dest_reg)] == 1
3420 /* or all sets must be consecutive and make a giv. */
3421 || (benefit = consec_sets_giv (benefit, p,
3422 src_reg, dest_reg,
3423 &add_val, &mult_val))))
3424 {
3425 int count;
3426 struct induction *v
3427 = (struct induction *) alloca (sizeof (struct induction));
3428 rtx temp;
3429
3430 /* If this is a library call, increase benefit. */
3431 if (find_reg_note (p, REG_RETVAL, 0))
3432 benefit += libcall_benefit (p);
3433
3434 /* Skip the consecutive insns, if there are any. */
3435 for (count = n_times_set[REGNO (dest_reg)] - 1;
3436 count > 0; count--)
3437 {
3438 /* If first insn of libcall sequence, skip to end.
3439 Do this at start of loop, since INSN is guaranteed to
3440 be an insn here. */
3441 if (GET_CODE (p) != NOTE
3442 && (temp = find_reg_note (p, REG_LIBCALL, 0)))
3443 p = XEXP (temp, 0);
3444
3445 do p = NEXT_INSN (p);
3446 while (GET_CODE (p) == NOTE);
3447 }
3448
3449 record_giv (v, p, src_reg, dest_reg, mult_val, add_val, benefit,
3450 DEST_REG, not_every_iteration, 0, loop_start,
3451 loop_end);
3452
3453 }
3454 }
3455
3456 #ifndef DONT_REDUCE_ADDR
3457 /* Look for givs which are memory addresses. */
3458 /* This resulted in worse code on a VAX 8600. I wonder if it
3459 still does. */
3460 if (GET_CODE (p) == INSN)
3461 find_mem_givs (PATTERN (p), p, not_every_iteration, loop_start,
3462 loop_end);
3463 #endif
3464
3465 /* Update the status of whether giv can derive other givs. This can
3466 change when we pass a label or an insn that updates a biv. */
3467 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
3468 || GET_CODE (p) == CODE_LABEL)
3469 update_giv_derive (p);
3470
3471 /* Past a label or a jump, we get to insns for which we can't count
3472 on whether or how many times they will be executed during each
3473 iteration. */
3474 /* This code appears in three places, once in scan_loop, and twice
3475 in strength_reduce. */
3476 if ((GET_CODE (p) == CODE_LABEL || GET_CODE (p) == JUMP_INSN)
3477 /* If we enter the loop in the middle, and scan around
3478 to the beginning, don't set not_every_iteration for that.
3479 This can be any kind of jump, since we want to know if insns
3480 will be executed if the loop is executed. */
3481 && ! (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) == loop_top
3482 && ((NEXT_INSN (NEXT_INSN (p)) == loop_end && simplejump_p (p))
3483 || (NEXT_INSN (p) == loop_end && condjump_p (p)))))
3484 not_every_iteration = 1;
3485
3486 /* At the virtual top of a converted loop, insns are again known to
3487 be executed each iteration: logically, the loop begins here
3488 even though the exit code has been duplicated. */
3489
3490 else if (GET_CODE (p) == NOTE
3491 && NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP)
3492 not_every_iteration = 0;
3493
3494 /* Unlike in the code motion pass where MAYBE_NEVER indicates that
3495 an insn may never be executed, NOT_EVERY_ITERATION indicates whether
3496 or not an insn is known to be executed each iteration of the
3497 loop, whether or not any iterations are known to occur.
3498
3499 Therefore, if we have just passed a label and have no more labels
3500 between here and the test insn of the loop, we know these insns
3501 will be executed each iteration. */
3502
3503 if (not_every_iteration && GET_CODE (p) == CODE_LABEL
3504 && no_labels_between_p (p, loop_end))
3505 not_every_iteration = 0;
3506 }
3507
3508 /* Try to calculate and save the number of loop iterations. This is
3509 set to zero if the actual number can not be calculated. This must
3510 be called after all giv's have been identified, since otherwise it may
3511 fail if the iteration variable is a giv. */
3512
3513 loop_n_iterations = loop_iterations (loop_start, loop_end);
3514
3515 /* Now for each giv for which we still don't know whether or not it is
3516 replaceable, check to see if it is replaceable because its final value
3517 can be calculated. This must be done after loop_iterations is called,
3518 so that final_giv_value will work correctly. */
3519
3520 for (bl = loop_iv_list; bl; bl = bl->next)
3521 {
3522 struct induction *v;
3523
3524 for (v = bl->giv; v; v = v->next_iv)
3525 if (! v->replaceable && ! v->not_replaceable)
3526 check_final_value (v, loop_start, loop_end);
3527 }
3528
3529 /* Try to prove that the loop counter variable (if any) is always
3530 nonnegative; if so, record that fact with a REG_NONNEG note
3531 so that "decrement and branch until zero" insn can be used. */
3532 check_dbra_loop (loop_end, insn_count, loop_start);
3533
3534 /* Create reg_map to hold substitutions for replaceable giv regs. */
3535 reg_map = (rtx *) alloca (max_reg_before_loop * sizeof (rtx));
3536 bzero ((char *) reg_map, max_reg_before_loop * sizeof (rtx));
3537
3538 /* Examine each iv class for feasibility of strength reduction/induction
3539 variable elimination. */
3540
3541 for (bl = loop_iv_list; bl; bl = bl->next)
3542 {
3543 struct induction *v;
3544 int benefit;
3545 int all_reduced;
3546 rtx final_value = 0;
3547
3548 /* Test whether it will be possible to eliminate this biv
3549 provided all givs are reduced. This is possible if either
3550 the reg is not used outside the loop, or we can compute
3551 what its final value will be.
3552
3553 For architectures with a decrement_and_branch_until_zero insn,
3554 don't do this if we put a REG_NONNEG note on the endtest for
3555 this biv. */
3556
3557 /* Compare against bl->init_insn rather than loop_start.
3558 We aren't concerned with any uses of the biv between
3559 init_insn and loop_start since these won't be affected
3560 by the value of the biv elsewhere in the function, so
3561 long as init_insn doesn't use the biv itself.
3562 March 14, 1989 -- self@bayes.arc.nasa.gov */
3563
3564 if ((uid_luid[regno_last_uid[bl->regno]] < INSN_LUID (loop_end)
3565 && bl->init_insn
3566 && INSN_UID (bl->init_insn) < max_uid_for_loop
3567 && uid_luid[regno_first_uid[bl->regno]] >= INSN_LUID (bl->init_insn)
3568 #ifdef HAVE_decrement_and_branch_until_zero
3569 && ! bl->nonneg
3570 #endif
3571 && ! reg_mentioned_p (bl->biv->dest_reg, SET_SRC (bl->init_set)))
3572 || ((final_value = final_biv_value (bl, loop_start, loop_end))
3573 #ifdef HAVE_decrement_and_branch_until_zero
3574 && ! bl->nonneg
3575 #endif
3576 ))
3577 bl->eliminable = maybe_eliminate_biv (bl, loop_start, end, 0,
3578 threshold, insn_count);
3579 else
3580 {
3581 if (loop_dump_stream)
3582 {
3583 fprintf (loop_dump_stream,
3584 "Cannot eliminate biv %d.\n",
3585 bl->regno);
3586 fprintf (loop_dump_stream,
3587 "First use: insn %d, last use: insn %d.\n",
3588 regno_first_uid[bl->regno],
3589 regno_last_uid[bl->regno]);
3590 }
3591 }
3592
3593 /* Combine all giv's for this iv_class. */
3594 combine_givs (bl);
3595
3596 /* This will be true at the end, if all givs which depend on this
3597 biv have been strength reduced.
3598 We can't (currently) eliminate the biv unless this is so. */
3599 all_reduced = 1;
3600
3601 /* Check each giv in this class to see if we will benefit by reducing
3602 it. Skip giv's combined with others. */
3603 for (v = bl->giv; v; v = v->next_iv)
3604 {
3605 struct induction *tv;
3606
3607 if (v->ignore || v->same)
3608 continue;
3609
3610 benefit = v->benefit;
3611
3612 /* Reduce benefit if not replaceable, since we will insert
3613 a move-insn to replace the insn that calculates this giv.
3614 Don't do this unless the giv is a user variable, since it
3615 will often be marked non-replaceable because of the duplication
3616 of the exit code outside the loop. In such a case, the copies
3617 we insert are dead and will be deleted. So they don't have
3618 a cost. Similar situations exist. */
3619 /* ??? The new final_[bg]iv_value code does a much better job
3620 of finding replaceable giv's, and hence this code may no longer
3621 be necessary. */
3622 if (! v->replaceable && ! bl->eliminable
3623 && REG_USERVAR_P (v->dest_reg))
3624 benefit -= copy_cost;
3625
3626 /* Decrease the benefit to count the add-insns that we will
3627 insert to increment the reduced reg for the giv. */
3628 benefit -= add_cost * bl->biv_count;
3629
3630 /* Decide whether to strength-reduce this giv or to leave the code
3631 unchanged (recompute it from the biv each time it is used).
3632 This decision can be made independently for each giv. */
3633
3634 /* ??? Perhaps attempt to guess whether autoincrement will handle
3635 some of the new add insns; if so, can increase BENEFIT
3636 (undo the subtraction of add_cost that was done above). */
3637
3638 /* If an insn is not to be strength reduced, then set its ignore
3639 flag, and clear all_reduced. */
3640
3641 if (v->lifetime * threshold * benefit < insn_count)
3642 {
3643 if (loop_dump_stream)
3644 fprintf (loop_dump_stream,
3645 "giv of insn %d not worth while, %d vs %d.\n",
3646 INSN_UID (v->insn),
3647 v->lifetime * threshold * benefit, insn_count);
3648 v->ignore = 1;
3649 all_reduced = 0;
3650 }
3651 else
3652 {
3653 /* Check that we can increment the reduced giv without a
3654 multiply insn. If not, reject it. */
3655
3656 for (tv = bl->biv; tv; tv = tv->next_iv)
3657 if (tv->mult_val == const1_rtx
3658 && ! product_cheap_p (tv->add_val, v->mult_val))
3659 {
3660 if (loop_dump_stream)
3661 fprintf (loop_dump_stream,
3662 "giv of insn %d: would need a multiply.\n",
3663 INSN_UID (v->insn));
3664 v->ignore = 1;
3665 all_reduced = 0;
3666 break;
3667 }
3668 }
3669 }
3670
3671 /* Reduce each giv that we decided to reduce. */
3672
3673 for (v = bl->giv; v; v = v->next_iv)
3674 {
3675 struct induction *tv;
3676 if (! v->ignore && v->same == 0)
3677 {
3678 v->new_reg = gen_reg_rtx (v->mode);
3679
3680 /* For each place where the biv is incremented,
3681 add an insn to increment the new, reduced reg for the giv. */
3682 for (tv = bl->biv; tv; tv = tv->next_iv)
3683 {
3684 if (tv->mult_val == const1_rtx)
3685 emit_iv_add_mult (tv->add_val, v->mult_val,
3686 v->new_reg, v->new_reg, tv->insn);
3687 else /* tv->mult_val == const0_rtx */
3688 /* A multiply is acceptable here
3689 since this is presumed to be seldom executed. */
3690 emit_iv_add_mult (tv->add_val, v->mult_val,
3691 v->add_val, v->new_reg, tv->insn);
3692 }
3693
3694 /* Add code at loop start to initialize giv's reduced reg. */
3695
3696 emit_iv_add_mult (bl->initial_value, v->mult_val,
3697 v->add_val, v->new_reg, loop_start);
3698 }
3699 }
3700
3701 /* Rescan all givs. If a giv is the same as a giv not reduced, mark it
3702 as not reduced.
3703
3704 For each giv register that can be reduced now: if replaceable,
3705 substitute reduced reg wherever the old giv occurs;
3706 else add new move insn "giv_reg = reduced_reg".
3707
3708 Also check for givs whose first use is their definition and whose
3709 last use is the definition of another giv. If so, it is likely
3710 dead and should not be used to eliminate a biv. */
3711 for (v = bl->giv; v; v = v->next_iv)
3712 {
3713 if (v->same && v->same->ignore)
3714 v->ignore = 1;
3715
3716 if (v->ignore)
3717 continue;
3718
3719 if (v->giv_type == DEST_REG
3720 && regno_first_uid[REGNO (v->dest_reg)] == INSN_UID (v->insn))
3721 {
3722 struct induction *v1;
3723
3724 for (v1 = bl->giv; v1; v1 = v1->next_iv)
3725 if (regno_last_uid[REGNO (v->dest_reg)] == INSN_UID (v1->insn))
3726 v->maybe_dead = 1;
3727 }
3728
3729 /* Update expression if this was combined, in case other giv was
3730 replaced. */
3731 if (v->same)
3732 v->new_reg = replace_rtx (v->new_reg,
3733 v->same->dest_reg, v->same->new_reg);
3734
3735 if (v->giv_type == DEST_ADDR)
3736 /* Store reduced reg as the address in the memref where we found
3737 this giv. */
3738 *v->location = v->new_reg;
3739 else if (v->replaceable)
3740 {
3741 reg_map[REGNO (v->dest_reg)] = v->new_reg;
3742
3743 #if 0
3744 /* I can no longer duplicate the original problem. Perhaps
3745 this is unnecessary now? */
3746
3747 /* Replaceable; it isn't strictly necessary to delete the old
3748 insn and emit a new one, because v->dest_reg is now dead.
3749
3750 However, especially when unrolling loops, the special
3751 handling for (set REG0 REG1) in the second cse pass may
3752 make v->dest_reg live again. To avoid this problem, emit
3753 an insn to set the original giv reg from the reduced giv.
3754 We can not delete the original insn, since it may be part
3755 of a LIBCALL, and the code in flow that eliminates dead
3756 libcalls will fail if it is deleted. */
3757 emit_insn_after (gen_move_insn (v->dest_reg, v->new_reg),
3758 v->insn);
3759 #endif
3760 }
3761 else
3762 {
3763 /* Not replaceable; emit an insn to set the original giv reg from
3764 the reduced giv, same as above. */
3765 emit_insn_after (gen_move_insn (v->dest_reg, v->new_reg),
3766 v->insn);
3767 }
3768
3769 /* When a loop is reversed, givs which depend on the reversed
3770 biv, and which are live outside the loop, must be set to their
3771 correct final value. This insn is only needed if the giv is
3772 not replaceable. The correct final value is the same as the
3773 value that the giv starts the reversed loop with. */
3774 if (bl->reversed && ! v->replaceable)
3775 emit_iv_add_mult (bl->initial_value, v->mult_val,
3776 v->add_val, v->dest_reg, end_insert_before);
3777 else if (v->final_value)
3778 {
3779 rtx insert_before;
3780
3781 /* If the loop has multiple exits, emit the insn before the
3782 loop to ensure that it will always be executed no matter
3783 how the loop exits. Otherwise, emit the insn after the loop,
3784 since this is slightly more efficient. */
3785 if (loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]])
3786 insert_before = loop_start;
3787 else
3788 insert_before = end_insert_before;
3789 emit_insn_before (gen_move_insn (v->dest_reg, v->final_value),
3790 insert_before);
3791
3792 #if 0
3793 /* If the insn to set the final value of the giv was emitted
3794 before the loop, then we must delete the insn inside the loop
3795 that sets it. If this is a LIBCALL, then we must delete
3796 every insn in the libcall. Note, however, that
3797 final_giv_value will only succeed when there are multiple
3798 exits if the giv is dead at each exit, hence it does not
3799 matter that the original insn remains because it is dead
3800 anyways. */
3801 /* Delete the insn inside the loop that sets the giv since
3802 the giv is now set before (or after) the loop. */
3803 delete_insn (v->insn);
3804 #endif
3805 }
3806
3807 if (loop_dump_stream)
3808 {
3809 fprintf (loop_dump_stream, "giv at %d reduced to ",
3810 INSN_UID (v->insn));
3811 print_rtl (loop_dump_stream, v->new_reg);
3812 fprintf (loop_dump_stream, "\n");
3813 }
3814 }
3815
3816 /* All the givs based on the biv bl have been reduced if they
3817 merit it. */
3818
3819 /* For each giv not marked as maybe dead that has been combined with a
3820 second giv, clear any "maybe dead" mark on that second giv.
3821 v->new_reg will either be or refer to the register of the giv it
3822 combined with.
3823
3824 Doing this clearing avoids problems in biv elimination where a
3825 giv's new_reg is a complex value that can't be put in the insn but
3826 the giv combined with (with a reg as new_reg) is marked maybe_dead.
3827 Since the register will be used in either case, we'd prefer it be
3828 used from the simpler giv. */
3829
3830 for (v = bl->giv; v; v = v->next_iv)
3831 if (! v->maybe_dead && v->same)
3832 v->same->maybe_dead = 0;
3833
3834 /* Try to eliminate the biv, if it is a candidate.
3835 This won't work if ! all_reduced,
3836 since the givs we planned to use might not have been reduced.
3837
3838 We have to be careful that we didn't initially think we could eliminate
3839 this biv because of a giv that we now think may be dead and shouldn't
3840 be used as a biv replacement.
3841
3842 Also, there is the possibility that we may have a giv that looks
3843 like it can be used to eliminate a biv, but the resulting insn
3844 isn't valid. This can happen, for example, on the 88k, where a
3845 JUMP_INSN can compare a register only with zero. Attempts to
3846 replace it with a compare with a constant will fail.
3847
3848 Note that in cases where this call fails, we may have replaced some
3849 of the occurrences of the biv with a giv, but no harm was done in
3850 doing so in the rare cases where it can occur. */
3851
3852 if (all_reduced == 1 && bl->eliminable
3853 && maybe_eliminate_biv (bl, loop_start, end, 1,
3854 threshold, insn_count))
3855
3856 {
3857 /* ?? If we created a new test to bypass the loop entirely,
3858 or otherwise drop straight in, based on this test, then
3859 we might want to rewrite it also. This way some later
3860 pass has more hope of removing the initialization of this
3861 biv entirely. */
3862
3863 /* If final_value != 0, then the biv may be used after loop end
3864 and we must emit an insn to set it just in case.
3865
3866 Reversed bivs already have an insn after the loop setting their
3867 value, so we don't need another one. We can't calculate the
3868 proper final value for such a biv here anyways. */
3869 if (final_value != 0 && ! bl->reversed)
3870 {
3871 rtx insert_before;
3872
3873 /* If the loop has multiple exits, emit the insn before the
3874 loop to ensure that it will always be executed no matter
3875 how the loop exits. Otherwise, emit the insn after the
3876 loop, since this is slightly more efficient. */
3877 if (loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]])
3878 insert_before = loop_start;
3879 else
3880 insert_before = end_insert_before;
3881
3882 emit_insn_before (gen_move_insn (bl->biv->dest_reg, final_value),
3883 end_insert_before);
3884 }
3885
3886 #if 0
3887 /* Delete all of the instructions inside the loop which set
3888 the biv, as they are all dead. If is safe to delete them,
3889 because an insn setting a biv will never be part of a libcall. */
3890 /* However, deleting them will invalidate the regno_last_uid info,
3891 so keeping them around is more convenient. Final_biv_value
3892 will only succeed when there are multiple exits if the biv
3893 is dead at each exit, hence it does not matter that the original
3894 insn remains, because it is dead anyways. */
3895 for (v = bl->biv; v; v = v->next_iv)
3896 delete_insn (v->insn);
3897 #endif
3898
3899 if (loop_dump_stream)
3900 fprintf (loop_dump_stream, "Reg %d: biv eliminated\n",
3901 bl->regno);
3902 }
3903 }
3904
3905 /* Go through all the instructions in the loop, making all the
3906 register substitutions scheduled in REG_MAP. */
3907
3908 for (p = loop_start; p != end; p = NEXT_INSN (p))
3909 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
3910 || GET_CODE (p) == CALL_INSN)
3911 {
3912 replace_regs (PATTERN (p), reg_map, max_reg_before_loop, 0);
3913 replace_regs (REG_NOTES (p), reg_map, max_reg_before_loop, 0);
3914 }
3915
3916 /* Unroll loops from within strength reduction so that we can use the
3917 induction variable information that strength_reduce has already
3918 collected. */
3919
3920 if (flag_unroll_loops)
3921 unroll_loop (loop_end, insn_count, loop_start, end_insert_before, 1);
3922
3923 if (loop_dump_stream)
3924 fprintf (loop_dump_stream, "\n");
3925 }
3926 \f
3927 /* Return 1 if X is a valid source for an initial value (or as value being
3928 compared against in an initial test).
3929
3930 X must be either a register or constant and must not be clobbered between
3931 the current insn and the start of the loop.
3932
3933 INSN is the insn containing X. */
3934
3935 static int
3936 valid_initial_value_p (x, insn, call_seen, loop_start)
3937 rtx x;
3938 rtx insn;
3939 int call_seen;
3940 rtx loop_start;
3941 {
3942 if (CONSTANT_P (x))
3943 return 1;
3944
3945 /* Only consider pseudos we know about initialized in insns whose luids
3946 we know. */
3947 if (GET_CODE (x) != REG
3948 || REGNO (x) >= max_reg_before_loop)
3949 return 0;
3950
3951 /* Don't use call-clobbered registers across a call which clobbers it. On
3952 some machines, don't use any hard registers at all. */
3953 if (REGNO (x) < FIRST_PSEUDO_REGISTER
3954 #ifndef SMALL_REGISTER_CLASSES
3955 && call_used_regs[REGNO (x)] && call_seen
3956 #endif
3957 )
3958 return 0;
3959
3960 /* Don't use registers that have been clobbered before the start of the
3961 loop. */
3962 if (reg_set_between_p (x, insn, loop_start))
3963 return 0;
3964
3965 return 1;
3966 }
3967 \f
3968 /* Scan X for memory refs and check each memory address
3969 as a possible giv. INSN is the insn whose pattern X comes from.
3970 NOT_EVERY_ITERATION is 1 if the insn might not be executed during
3971 every loop iteration. */
3972
3973 static void
3974 find_mem_givs (x, insn, not_every_iteration, loop_start, loop_end)
3975 rtx x;
3976 rtx insn;
3977 int not_every_iteration;
3978 rtx loop_start, loop_end;
3979 {
3980 register int i, j;
3981 register enum rtx_code code;
3982 register char *fmt;
3983
3984 if (x == 0)
3985 return;
3986
3987 code = GET_CODE (x);
3988 switch (code)
3989 {
3990 case REG:
3991 case CONST_INT:
3992 case CONST:
3993 case CONST_DOUBLE:
3994 case SYMBOL_REF:
3995 case LABEL_REF:
3996 case PC:
3997 case CC0:
3998 case ADDR_VEC:
3999 case ADDR_DIFF_VEC:
4000 case USE:
4001 case CLOBBER:
4002 return;
4003
4004 case MEM:
4005 {
4006 rtx src_reg;
4007 rtx add_val;
4008 rtx mult_val;
4009 int benefit;
4010
4011 benefit = general_induction_var (XEXP (x, 0),
4012 &src_reg, &add_val, &mult_val);
4013
4014 /* Don't make a DEST_ADDR giv with mult_val == 1 && add_val == 0.
4015 Such a giv isn't useful. */
4016 if (benefit > 0 && (mult_val != const1_rtx || add_val != const0_rtx))
4017 {
4018 /* Found one; record it. */
4019 struct induction *v
4020 = (struct induction *) oballoc (sizeof (struct induction));
4021
4022 record_giv (v, insn, src_reg, addr_placeholder, mult_val,
4023 add_val, benefit, DEST_ADDR, not_every_iteration,
4024 &XEXP (x, 0), loop_start, loop_end);
4025
4026 v->mem_mode = GET_MODE (x);
4027 }
4028 return;
4029 }
4030 }
4031
4032 /* Recursively scan the subexpressions for other mem refs. */
4033
4034 fmt = GET_RTX_FORMAT (code);
4035 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4036 if (fmt[i] == 'e')
4037 find_mem_givs (XEXP (x, i), insn, not_every_iteration, loop_start,
4038 loop_end);
4039 else if (fmt[i] == 'E')
4040 for (j = 0; j < XVECLEN (x, i); j++)
4041 find_mem_givs (XVECEXP (x, i, j), insn, not_every_iteration,
4042 loop_start, loop_end);
4043 }
4044 \f
4045 /* Fill in the data about one biv update.
4046 V is the `struct induction' in which we record the biv. (It is
4047 allocated by the caller, with alloca.)
4048 INSN is the insn that sets it.
4049 DEST_REG is the biv's reg.
4050
4051 MULT_VAL is const1_rtx if the biv is being incremented here, in which case
4052 INC_VAL is the increment. Otherwise, MULT_VAL is const0_rtx and the biv is
4053 being set to INC_VAL.
4054
4055 NOT_EVERY_ITERATION is nonzero if this biv update is not know to be
4056 executed every iteration; MAYBE_MULTIPLE is nonzero if this biv update
4057 can be executed more than once per iteration. If MAYBE_MULTIPLE
4058 and NOT_EVERY_ITERATION are both zero, we know that the biv update is
4059 executed exactly once per iteration. */
4060
4061 static void
4062 record_biv (v, insn, dest_reg, inc_val, mult_val,
4063 not_every_iteration, maybe_multiple)
4064 struct induction *v;
4065 rtx insn;
4066 rtx dest_reg;
4067 rtx inc_val;
4068 rtx mult_val;
4069 int not_every_iteration;
4070 int maybe_multiple;
4071 {
4072 struct iv_class *bl;
4073
4074 v->insn = insn;
4075 v->src_reg = dest_reg;
4076 v->dest_reg = dest_reg;
4077 v->mult_val = mult_val;
4078 v->add_val = inc_val;
4079 v->mode = GET_MODE (dest_reg);
4080 v->always_computable = ! not_every_iteration;
4081 v->maybe_multiple = maybe_multiple;
4082
4083 /* Add this to the reg's iv_class, creating a class
4084 if this is the first incrementation of the reg. */
4085
4086 bl = reg_biv_class[REGNO (dest_reg)];
4087 if (bl == 0)
4088 {
4089 /* Create and initialize new iv_class. */
4090
4091 bl = (struct iv_class *) oballoc (sizeof (struct iv_class));
4092
4093 bl->regno = REGNO (dest_reg);
4094 bl->biv = 0;
4095 bl->giv = 0;
4096 bl->biv_count = 0;
4097 bl->giv_count = 0;
4098
4099 /* Set initial value to the reg itself. */
4100 bl->initial_value = dest_reg;
4101 /* We haven't seen the initializing insn yet */
4102 bl->init_insn = 0;
4103 bl->init_set = 0;
4104 bl->initial_test = 0;
4105 bl->incremented = 0;
4106 bl->eliminable = 0;
4107 bl->nonneg = 0;
4108 bl->reversed = 0;
4109
4110 /* Add this class to loop_iv_list. */
4111 bl->next = loop_iv_list;
4112 loop_iv_list = bl;
4113
4114 /* Put it in the array of biv register classes. */
4115 reg_biv_class[REGNO (dest_reg)] = bl;
4116 }
4117
4118 /* Update IV_CLASS entry for this biv. */
4119 v->next_iv = bl->biv;
4120 bl->biv = v;
4121 bl->biv_count++;
4122 if (mult_val == const1_rtx)
4123 bl->incremented = 1;
4124
4125 if (loop_dump_stream)
4126 {
4127 fprintf (loop_dump_stream,
4128 "Insn %d: possible biv, reg %d,",
4129 INSN_UID (insn), REGNO (dest_reg));
4130 if (GET_CODE (inc_val) == CONST_INT)
4131 fprintf (loop_dump_stream, " const = %d\n",
4132 INTVAL (inc_val));
4133 else
4134 {
4135 fprintf (loop_dump_stream, " const = ");
4136 print_rtl (loop_dump_stream, inc_val);
4137 fprintf (loop_dump_stream, "\n");
4138 }
4139 }
4140 }
4141 \f
4142 /* Fill in the data about one giv.
4143 V is the `struct induction' in which we record the giv. (It is
4144 allocated by the caller, with alloca.)
4145 INSN is the insn that sets it.
4146 BENEFIT estimates the savings from deleting this insn.
4147 TYPE is DEST_REG or DEST_ADDR; it says whether the giv is computed
4148 into a register or is used as a memory address.
4149
4150 SRC_REG is the biv reg which the giv is computed from.
4151 DEST_REG is the giv's reg (if the giv is stored in a reg).
4152 MULT_VAL and ADD_VAL are the coefficients used to compute the giv.
4153 LOCATION points to the place where this giv's value appears in INSN. */
4154
4155 static void
4156 record_giv (v, insn, src_reg, dest_reg, mult_val, add_val, benefit,
4157 type, not_every_iteration, location, loop_start, loop_end)
4158 struct induction *v;
4159 rtx insn;
4160 rtx src_reg;
4161 rtx dest_reg;
4162 rtx mult_val, add_val;
4163 int benefit;
4164 enum g_types type;
4165 int not_every_iteration;
4166 rtx *location;
4167 rtx loop_start, loop_end;
4168 {
4169 struct induction *b;
4170 struct iv_class *bl;
4171 rtx set = single_set (insn);
4172 rtx p;
4173
4174 v->insn = insn;
4175 v->src_reg = src_reg;
4176 v->giv_type = type;
4177 v->dest_reg = dest_reg;
4178 v->mult_val = mult_val;
4179 v->add_val = add_val;
4180 v->benefit = benefit;
4181 v->location = location;
4182 v->cant_derive = 0;
4183 v->combined_with = 0;
4184 v->maybe_multiple = 0;
4185 v->maybe_dead = 0;
4186 v->derive_adjustment = 0;
4187 v->same = 0;
4188 v->ignore = 0;
4189 v->new_reg = 0;
4190 v->final_value = 0;
4191
4192 /* The v->always_computable field is used in update_giv_derive, to
4193 determine whether a giv can be used to derive another giv. For a
4194 DEST_REG giv, INSN computes a new value for the giv, so its value
4195 isn't computable if INSN insn't executed every iteration.
4196 However, for a DEST_ADDR giv, INSN merely uses the value of the giv;
4197 it does not compute a new value. Hence the value is always computable
4198 regardless of whether INSN is executed each iteration. */
4199
4200 if (type == DEST_ADDR)
4201 v->always_computable = 1;
4202 else
4203 v->always_computable = ! not_every_iteration;
4204
4205 if (type == DEST_ADDR)
4206 {
4207 v->mode = GET_MODE (*location);
4208 v->lifetime = 1;
4209 v->times_used = 1;
4210 }
4211 else /* type == DEST_REG */
4212 {
4213 v->mode = GET_MODE (SET_DEST (set));
4214
4215 v->lifetime = (uid_luid[regno_last_uid[REGNO (dest_reg)]]
4216 - uid_luid[regno_first_uid[REGNO (dest_reg)]]);
4217
4218 v->times_used = n_times_used[REGNO (dest_reg)];
4219
4220 /* If the lifetime is zero, it means that this register is
4221 really a dead store. So mark this as a giv that can be
4222 ignored. This will not prevent the biv from being eliminated. */
4223 if (v->lifetime == 0)
4224 v->ignore = 1;
4225
4226 reg_iv_type[REGNO (dest_reg)] = GENERAL_INDUCT;
4227 reg_iv_info[REGNO (dest_reg)] = v;
4228 }
4229
4230 /* Add the giv to the class of givs computed from one biv. */
4231
4232 bl = reg_biv_class[REGNO (src_reg)];
4233 if (bl)
4234 {
4235 v->next_iv = bl->giv;
4236 bl->giv = v;
4237 /* Don't count DEST_ADDR. This is supposed to count the number of
4238 insns that calculate givs. */
4239 if (type == DEST_REG)
4240 bl->giv_count++;
4241 bl->total_benefit += benefit;
4242 }
4243 else
4244 /* Fatal error, biv missing for this giv? */
4245 abort ();
4246
4247 if (type == DEST_ADDR)
4248 v->replaceable = 1;
4249 else
4250 {
4251 /* The giv can be replaced outright by the reduced register only if all
4252 of the following conditions are true:
4253 - the insn that sets the giv is always executed on any iteration
4254 on which the giv is used at all
4255 (there are two ways to deduce this:
4256 either the insn is executed on every iteration,
4257 or all uses follow that insn in the same basic block),
4258 - the giv is not used outside the loop
4259 - no assignments to the biv occur during the giv's lifetime. */
4260
4261 if (regno_first_uid[REGNO (dest_reg)] == INSN_UID (insn)
4262 /* Previous line always fails if INSN was moved by loop opt. */
4263 && uid_luid[regno_last_uid[REGNO (dest_reg)]] < INSN_LUID (loop_end)
4264 && (! not_every_iteration
4265 || last_use_this_basic_block (dest_reg, insn)))
4266 {
4267 /* Now check that there are no assignments to the biv within the
4268 giv's lifetime. This requires two separate checks. */
4269
4270 /* Check each biv update, and fail if any are between the first
4271 and last use of the giv.
4272
4273 If this loop contains an inner loop that was unrolled, then
4274 the insn modifying the biv may have been emitted by the loop
4275 unrolling code, and hence does not have a valid luid. Just
4276 mark the biv as not replaceable in this case. It is not very
4277 useful as a biv, because it is used in two different loops.
4278 It is very unlikely that we would be able to optimize the giv
4279 using this biv anyways. */
4280
4281 v->replaceable = 1;
4282 for (b = bl->biv; b; b = b->next_iv)
4283 {
4284 if (INSN_UID (b->insn) >= max_uid_for_loop
4285 || ((uid_luid[INSN_UID (b->insn)]
4286 >= uid_luid[regno_first_uid[REGNO (dest_reg)]])
4287 && (uid_luid[INSN_UID (b->insn)]
4288 <= uid_luid[regno_last_uid[REGNO (dest_reg)]])))
4289 {
4290 v->replaceable = 0;
4291 v->not_replaceable = 1;
4292 break;
4293 }
4294 }
4295
4296 /* Check each insn between the first and last use of the giv,
4297 and fail if any of them are branches that jump to a named label
4298 outside this range, but still inside the loop. This catches
4299 cases of spaghetti code where the execution order of insns
4300 is not linear, and hence the above test fails. For example,
4301 in the following code, j is not replaceable:
4302 for (i = 0; i < 100; ) {
4303 L0: j = 4*i; goto L1;
4304 L2: k = j; goto L3;
4305 L1: i++; goto L2;
4306 L3: ; }
4307 printf ("k = %d\n", k); }
4308 This test is conservative, but this test succeeds rarely enough
4309 that it isn't a problem. See also check_final_value below. */
4310
4311 if (v->replaceable)
4312 for (p = insn;
4313 INSN_UID (p) >= max_uid_for_loop
4314 || INSN_LUID (p) < uid_luid[regno_last_uid[REGNO (dest_reg)]];
4315 p = NEXT_INSN (p))
4316 {
4317 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p)
4318 && LABEL_NAME (JUMP_LABEL (p))
4319 && ((INSN_LUID (JUMP_LABEL (p)) > INSN_LUID (loop_start)
4320 && (INSN_LUID (JUMP_LABEL (p))
4321 < uid_luid[regno_first_uid[REGNO (dest_reg)]]))
4322 || (INSN_LUID (JUMP_LABEL (p)) < INSN_LUID (loop_end)
4323 && (INSN_LUID (JUMP_LABEL (p))
4324 > uid_luid[regno_last_uid[REGNO (dest_reg)]]))))
4325 {
4326 v->replaceable = 0;
4327 v->not_replaceable = 1;
4328
4329 if (loop_dump_stream)
4330 fprintf (loop_dump_stream,
4331 "Found branch outside giv lifetime.\n");
4332
4333 break;
4334 }
4335 }
4336 }
4337 else
4338 {
4339 /* May still be replaceable, we don't have enough info here to
4340 decide. */
4341 v->replaceable = 0;
4342 v->not_replaceable = 0;
4343 }
4344 }
4345
4346 if (loop_dump_stream)
4347 {
4348 if (type == DEST_REG)
4349 fprintf (loop_dump_stream, "Insn %d: giv reg %d",
4350 INSN_UID (insn), REGNO (dest_reg));
4351 else
4352 fprintf (loop_dump_stream, "Insn %d: dest address",
4353 INSN_UID (insn));
4354
4355 fprintf (loop_dump_stream, " src reg %d benefit %d",
4356 REGNO (src_reg), v->benefit);
4357 fprintf (loop_dump_stream, " used %d lifetime %d",
4358 v->times_used, v->lifetime);
4359
4360 if (v->replaceable)
4361 fprintf (loop_dump_stream, " replaceable");
4362
4363 if (GET_CODE (mult_val) == CONST_INT)
4364 fprintf (loop_dump_stream, " mult %d",
4365 INTVAL (mult_val));
4366 else
4367 {
4368 fprintf (loop_dump_stream, " mult ");
4369 print_rtl (loop_dump_stream, mult_val);
4370 }
4371
4372 if (GET_CODE (add_val) == CONST_INT)
4373 fprintf (loop_dump_stream, " add %d",
4374 INTVAL (add_val));
4375 else
4376 {
4377 fprintf (loop_dump_stream, " add ");
4378 print_rtl (loop_dump_stream, add_val);
4379 }
4380 }
4381
4382 if (loop_dump_stream)
4383 fprintf (loop_dump_stream, "\n");
4384
4385 }
4386
4387
4388 /* All this does is determine whether a giv can be made replaceable because
4389 its final value can be calculated. This code can not be part of record_giv
4390 above, because final_giv_value requires that the number of loop iterations
4391 be known, and that can not be accurately calculated until after all givs
4392 have been identified. */
4393
4394 static void
4395 check_final_value (v, loop_start, loop_end)
4396 struct induction *v;
4397 rtx loop_start, loop_end;
4398 {
4399 struct iv_class *bl;
4400 rtx final_value = 0;
4401 rtx tem;
4402
4403 bl = reg_biv_class[REGNO (v->src_reg)];
4404
4405 /* DEST_ADDR givs will never reach here, because they are always marked
4406 replaceable above in record_giv. */
4407
4408 /* The giv can be replaced outright by the reduced register only if all
4409 of the following conditions are true:
4410 - the insn that sets the giv is always executed on any iteration
4411 on which the giv is used at all
4412 (there are two ways to deduce this:
4413 either the insn is executed on every iteration,
4414 or all uses follow that insn in the same basic block),
4415 - its final value can be calculated (this condition is different
4416 than the one above in record_giv)
4417 - no assignments to the biv occur during the giv's lifetime. */
4418
4419 #if 0
4420 /* This is only called now when replaceable is known to be false. */
4421 /* Clear replaceable, so that it won't confuse final_giv_value. */
4422 v->replaceable = 0;
4423 #endif
4424
4425 if ((final_value = final_giv_value (v, loop_start, loop_end))
4426 && (v->always_computable || last_use_this_basic_block (v->dest_reg, v->insn)))
4427 {
4428 int biv_increment_seen = 0;
4429 rtx p = v->insn;
4430 rtx last_giv_use;
4431
4432 v->replaceable = 1;
4433
4434 /* When trying to determine whether or not a biv increment occurs
4435 during the lifetime of the giv, we can ignore uses of the variable
4436 outside the loop because final_value is true. Hence we can not
4437 use regno_last_uid and regno_first_uid as above in record_giv. */
4438
4439 /* Search the loop to determine whether any assignments to the
4440 biv occur during the giv's lifetime. Start with the insn
4441 that sets the giv, and search around the loop until we come
4442 back to that insn again.
4443
4444 Also fail if there is a jump within the giv's lifetime that jumps
4445 to somewhere outside the lifetime but still within the loop. This
4446 catches spaghetti code where the execution order is not linear, and
4447 hence the above test fails. Here we assume that the giv lifetime
4448 does not extend from one iteration of the loop to the next, so as
4449 to make the test easier. Since the lifetime isn't known yet,
4450 this requires two loops. See also record_giv above. */
4451
4452 last_giv_use = v->insn;
4453
4454 while (1)
4455 {
4456 p = NEXT_INSN (p);
4457 if (p == loop_end)
4458 p = NEXT_INSN (loop_start);
4459 if (p == v->insn)
4460 break;
4461
4462 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
4463 || GET_CODE (p) == CALL_INSN)
4464 {
4465 if (biv_increment_seen)
4466 {
4467 if (reg_mentioned_p (v->dest_reg, PATTERN (p)))
4468 {
4469 v->replaceable = 0;
4470 v->not_replaceable = 1;
4471 break;
4472 }
4473 }
4474 else if (GET_CODE (PATTERN (p)) == SET
4475 && SET_DEST (PATTERN (p)) == v->src_reg)
4476 biv_increment_seen = 1;
4477 else if (reg_mentioned_p (v->dest_reg, PATTERN (p)))
4478 last_giv_use = p;
4479 }
4480 }
4481
4482 /* Now that the lifetime of the giv is known, check for branches
4483 from within the lifetime to outside the lifetime if it is still
4484 replaceable. */
4485
4486 if (v->replaceable)
4487 {
4488 p = v->insn;
4489 while (1)
4490 {
4491 p = NEXT_INSN (p);
4492 if (p == loop_end)
4493 p = NEXT_INSN (loop_start);
4494 if (p == last_giv_use)
4495 break;
4496
4497 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p)
4498 && LABEL_NAME (JUMP_LABEL (p))
4499 && ((INSN_LUID (JUMP_LABEL (p)) < INSN_LUID (v->insn)
4500 && INSN_LUID (JUMP_LABEL (p)) > INSN_LUID (loop_start))
4501 || (INSN_LUID (JUMP_LABEL (p)) > INSN_LUID (last_giv_use)
4502 && INSN_LUID (JUMP_LABEL (p)) < INSN_LUID (loop_end))))
4503 {
4504 v->replaceable = 0;
4505 v->not_replaceable = 1;
4506
4507 if (loop_dump_stream)
4508 fprintf (loop_dump_stream,
4509 "Found branch outside giv lifetime.\n");
4510
4511 break;
4512 }
4513 }
4514 }
4515
4516 /* If it is replaceable, then save the final value. */
4517 if (v->replaceable)
4518 v->final_value = final_value;
4519 }
4520
4521 if (loop_dump_stream && v->replaceable)
4522 fprintf (loop_dump_stream, "Insn %d: giv reg %d final_value replaceable\n",
4523 INSN_UID (v->insn), REGNO (v->dest_reg));
4524 }
4525 \f
4526 /* Update the status of whether a giv can derive other givs.
4527
4528 We need to do something special if there is or may be an update to the biv
4529 between the time the giv is defined and the time it is used to derive
4530 another giv.
4531
4532 In addition, a giv that is only conditionally set is not allowed to
4533 derive another giv once a label has been passed.
4534
4535 The cases we look at are when a label or an update to a biv is passed. */
4536
4537 static void
4538 update_giv_derive (p)
4539 rtx p;
4540 {
4541 struct iv_class *bl;
4542 struct induction *biv, *giv;
4543 rtx tem;
4544 int dummy;
4545
4546 /* Search all IV classes, then all bivs, and finally all givs.
4547
4548 There are three cases we are concerned with. First we have the situation
4549 of a giv that is only updated conditionally. In that case, it may not
4550 derive any givs after a label is passed.
4551
4552 The second case is when a biv update occurs, or may occur, after the
4553 definition of a giv. For certain biv updates (see below) that are
4554 known to occur between the giv definition and use, we can adjust the
4555 giv definition. For others, or when the biv update is conditional,
4556 we must prevent the giv from deriving any other givs. There are two
4557 sub-cases within this case.
4558
4559 If this is a label, we are concerned with any biv update that is done
4560 conditionally, since it may be done after the giv is defined followed by
4561 a branch here (actually, we need to pass both a jump and a label, but
4562 this extra tracking doesn't seem worth it).
4563
4564 If this is a jump, we are concerned about any biv update that may be
4565 executed multiple times. We are actually only concerned about
4566 backward jumps, but it is probably not worth performing the test
4567 on the jump again here.
4568
4569 If this is a biv update, we must adjust the giv status to show that a
4570 subsequent biv update was performed. If this adjustment cannot be done,
4571 the giv cannot derive further givs. */
4572
4573 for (bl = loop_iv_list; bl; bl = bl->next)
4574 for (biv = bl->biv; biv; biv = biv->next_iv)
4575 if (GET_CODE (p) == CODE_LABEL || GET_CODE (p) == JUMP_INSN
4576 || biv->insn == p)
4577 {
4578 for (giv = bl->giv; giv; giv = giv->next_iv)
4579 {
4580 /* If cant_derive is already true, there is no point in
4581 checking all of these conditions again. */
4582 if (giv->cant_derive)
4583 continue;
4584
4585 /* If this giv is conditionally set and we have passed a label,
4586 it cannot derive anything. */
4587 if (GET_CODE (p) == CODE_LABEL && ! giv->always_computable)
4588 giv->cant_derive = 1;
4589
4590 /* Skip givs that have mult_val == 0, since
4591 they are really invariants. Also skip those that are
4592 replaceable, since we know their lifetime doesn't contain
4593 any biv update. */
4594 else if (giv->mult_val == const0_rtx || giv->replaceable)
4595 continue;
4596
4597 /* The only way we can allow this giv to derive another
4598 is if this is a biv increment and we can form the product
4599 of biv->add_val and giv->mult_val. In this case, we will
4600 be able to compute a compensation. */
4601 else if (biv->insn == p)
4602 {
4603 tem = 0;
4604
4605 if (biv->mult_val == const1_rtx)
4606 tem = simplify_giv_expr (gen_rtx (MULT, giv->mode,
4607 biv->add_val,
4608 giv->mult_val),
4609 &dummy);
4610
4611 if (tem && giv->derive_adjustment)
4612 tem = simplify_giv_expr (gen_rtx (PLUS, giv->mode, tem,
4613 giv->derive_adjustment),
4614 &dummy);
4615 if (tem)
4616 giv->derive_adjustment = tem;
4617 else
4618 giv->cant_derive = 1;
4619 }
4620 else if ((GET_CODE (p) == CODE_LABEL && ! biv->always_computable)
4621 || (GET_CODE (p) == JUMP_INSN && biv->maybe_multiple))
4622 giv->cant_derive = 1;
4623 }
4624 }
4625 }
4626 \f
4627 /* Check whether an insn is an increment legitimate for a basic induction var.
4628 X is the source of the insn.
4629 DEST_REG is the putative biv, also the destination of the insn.
4630 We accept patterns of these forms:
4631 REG = REG + INVARIANT
4632 REG = INVARIANT + REG
4633 REG = REG - CONSTANT
4634
4635 If X is suitable, we return 1, set *MULT_VAL to CONST1_RTX,
4636 and store the additive term into *INC_VAL.
4637
4638 If X is an assignment of an invariant into DEST_REG, we set
4639 *MULT_VAL to CONST0_RTX, and store the invariant into *INC_VAL.
4640
4641 Otherwise we return 0. */
4642
4643 static int
4644 basic_induction_var (x, dest_reg, inc_val, mult_val)
4645 register rtx x;
4646 rtx dest_reg;
4647 rtx *inc_val;
4648 rtx *mult_val;
4649 {
4650 register enum rtx_code code;
4651 rtx arg;
4652
4653 code = GET_CODE (x);
4654 switch (code)
4655 {
4656 case PLUS:
4657 if (XEXP (x, 0) == dest_reg)
4658 arg = XEXP (x, 1);
4659 else if (XEXP (x, 1) == dest_reg)
4660 arg = XEXP (x, 0);
4661 else
4662 return 0;
4663
4664 if (invariant_p (arg) != 1)
4665 return 0;
4666
4667 *inc_val = arg;
4668 *mult_val = const1_rtx;
4669 return 1;
4670
4671 case MINUS:
4672 if (XEXP (x, 0) == dest_reg
4673 && GET_CODE (XEXP (x, 1)) == CONST_INT)
4674 *inc_val = gen_rtx (CONST_INT, VOIDmode,
4675 - INTVAL (XEXP (x, 1)));
4676 else
4677 return 0;
4678
4679 *mult_val = const1_rtx;
4680 return 1;
4681
4682 /* Can accept constant setting of biv only when inside inner most loop.
4683 Otherwise, a biv of an inner loop may be incorrectly recognized
4684 as a biv of the outer loop,
4685 causing code to be moved INTO the inner loop. */
4686 case MEM:
4687 case REG:
4688 if (invariant_p (x) != 1)
4689 return 0;
4690 case CONST_INT:
4691 case SYMBOL_REF:
4692 case CONST:
4693 if (loops_enclosed == 1)
4694 {
4695 *inc_val = x;
4696 *mult_val = const0_rtx;
4697 return 1;
4698 }
4699 else
4700 return 0;
4701
4702 default:
4703 return 0;
4704 }
4705 }
4706 \f
4707 /* A general induction variable (giv) is any quantity that is a linear
4708 function of a basic induction variable,
4709 i.e. giv = biv * mult_val + add_val.
4710 The coefficients can be any loop invariant quantity.
4711 A giv need not be computed directly from the biv;
4712 it can be computed by way of other givs. */
4713
4714 /* Determine whether X computes a giv.
4715 If it does, return a nonzero value
4716 which is the benefit from eliminating the computation of X;
4717 set *SRC_REG to the register of the biv that it is computed from;
4718 set *ADD_VAL and *MULT_VAL to the coefficients,
4719 such that the value of X is biv * mult + add; */
4720
4721 static int
4722 general_induction_var (x, src_reg, add_val, mult_val)
4723 rtx x;
4724 rtx *src_reg;
4725 rtx *add_val;
4726 rtx *mult_val;
4727 {
4728 rtx orig_x = x;
4729 int benefit = 0;
4730 char *storage;
4731
4732 /* If this is an invariant, forget it, it isn't a giv. */
4733 if (invariant_p (x) == 1)
4734 return 0;
4735
4736 /* See if the expression could be a giv and get its form.
4737 Mark our place on the obstack in case we don't find a giv. */
4738 storage = (char *) oballoc (0);
4739 x = simplify_giv_expr (x, &benefit);
4740 if (x == 0)
4741 {
4742 obfree (storage);
4743 return 0;
4744 }
4745
4746 switch (GET_CODE (x))
4747 {
4748 case USE:
4749 case CONST_INT:
4750 /* Since this is now an invariant and wasn't before, it must be a giv
4751 with MULT_VAL == 0. It doesn't matter which BIV we associate this
4752 with. */
4753 *src_reg = loop_iv_list->biv->dest_reg;
4754 *mult_val = const0_rtx;
4755 *add_val = x;
4756 break;
4757
4758 case REG:
4759 /* This is equivalent to a BIV. */
4760 *src_reg = x;
4761 *mult_val = const1_rtx;
4762 *add_val = const0_rtx;
4763 break;
4764
4765 case PLUS:
4766 /* Either (plus (biv) (invar)) or
4767 (plus (mult (biv) (invar_1)) (invar_2)). */
4768 if (GET_CODE (XEXP (x, 0)) == MULT)
4769 {
4770 *src_reg = XEXP (XEXP (x, 0), 0);
4771 *mult_val = XEXP (XEXP (x, 0), 1);
4772 }
4773 else
4774 {
4775 *src_reg = XEXP (x, 0);
4776 *mult_val = const1_rtx;
4777 }
4778 *add_val = XEXP (x, 1);
4779 break;
4780
4781 case MULT:
4782 /* ADD_VAL is zero. */
4783 *src_reg = XEXP (x, 0);
4784 *mult_val = XEXP (x, 1);
4785 *add_val = const0_rtx;
4786 break;
4787
4788 default:
4789 abort ();
4790 }
4791
4792 /* Remove any enclosing USE from ADD_VAL and MULT_VAL (there will be
4793 unless they are CONST_INT). */
4794 if (GET_CODE (*add_val) == USE)
4795 *add_val = XEXP (*add_val, 0);
4796 if (GET_CODE (*mult_val) == USE)
4797 *mult_val = XEXP (*mult_val, 0);
4798
4799 benefit += rtx_cost (orig_x, SET);
4800
4801 /* Always return some benefit if this is a giv so it will be detected
4802 as such. This allows elimination of bivs that might otherwise
4803 not be eliminated. */
4804 return benefit == 0 ? 1 : benefit;
4805 }
4806 \f
4807 /* Given an expression, X, try to form it as a linear function of a biv.
4808 We will canonicalize it to be of the form
4809 (plus (mult (BIV) (invar_1))
4810 (invar_2))
4811 with possible degeneracies.
4812
4813 The invariant expressions must each be of a form that can be used as a
4814 machine operand. We surround then with a USE rtx (a hack, but localized
4815 and certainly unambiguous!) if not a CONST_INT for simplicity in this
4816 routine; it is the caller's responsibility to strip them.
4817
4818 If no such canonicalization is possible (i.e., two biv's are used or an
4819 expression that is neither invariant nor a biv or giv), this routine
4820 returns 0.
4821
4822 For a non-zero return, the result will have a code of CONST_INT, USE,
4823 REG (for a BIV), PLUS, or MULT. No other codes will occur.
4824
4825 *BENEFIT will be incremented by the benefit of any sub-giv encountered. */
4826
4827 static rtx
4828 simplify_giv_expr (x, benefit)
4829 rtx x;
4830 int *benefit;
4831 {
4832 enum machine_mode mode = GET_MODE (x);
4833 rtx arg0, arg1;
4834 rtx tem;
4835
4836 /* If this is not an integer mode, or if we cannot do arithmetic in this
4837 mode, this can't be a giv. */
4838 if (mode != VOIDmode
4839 && (GET_MODE_CLASS (mode) != MODE_INT
4840 || GET_MODE_BITSIZE (mode) > HOST_BITS_PER_INT))
4841 return 0;
4842
4843 switch (GET_CODE (x))
4844 {
4845 case PLUS:
4846 arg0 = simplify_giv_expr (XEXP (x, 0), benefit);
4847 arg1 = simplify_giv_expr (XEXP (x, 1), benefit);
4848 if (arg0 == 0 || arg1 == 0)
4849 return 0;
4850
4851 /* Put constant last, CONST_INT last if both constant. */
4852 if ((GET_CODE (arg0) == USE
4853 || GET_CODE (arg0) == CONST_INT)
4854 && GET_CODE (arg1) != CONST_INT)
4855 tem = arg0, arg0 = arg1, arg1 = tem;
4856
4857 /* Handle addition of zero, then addition of an invariant. */
4858 if (arg1 == const0_rtx)
4859 return arg0;
4860 else if (GET_CODE (arg1) == CONST_INT || GET_CODE (arg1) == USE)
4861 switch (GET_CODE (arg0))
4862 {
4863 case CONST_INT:
4864 case USE:
4865 /* Both invariant. Only valid if sum is machine operand.
4866 First strip off possible USE on first operand. */
4867 if (GET_CODE (arg0) == USE)
4868 arg0 = XEXP (arg0, 0);
4869
4870 tem = 0;
4871 if (CONSTANT_P (arg0) && GET_CODE (arg1) == CONST_INT)
4872 {
4873 tem = plus_constant (arg0, INTVAL (arg1));
4874 if (GET_CODE (tem) != CONST_INT)
4875 tem = gen_rtx (USE, mode, tem);
4876 }
4877
4878 return tem;
4879
4880 case REG:
4881 case MULT:
4882 /* biv + invar or mult + invar. Return sum. */
4883 return gen_rtx (PLUS, mode, arg0, arg1);
4884
4885 case PLUS:
4886 /* (a + invar_1) + invar_2. Associate. */
4887 return simplify_giv_expr (gen_rtx (PLUS, mode,
4888 XEXP (arg0, 0),
4889 gen_rtx (PLUS, mode,
4890 XEXP (arg0, 1), arg1)),
4891 benefit);
4892
4893 default:
4894 abort ();
4895 }
4896
4897 /* Each argument must be either REG, PLUS, or MULT. Convert REG to
4898 MULT to reduce cases. */
4899 if (GET_CODE (arg0) == REG)
4900 arg0 = gen_rtx (MULT, mode, arg0, const1_rtx);
4901 if (GET_CODE (arg1) == REG)
4902 arg1 = gen_rtx (MULT, mode, arg1, const1_rtx);
4903
4904 /* Now have PLUS + PLUS, PLUS + MULT, MULT + PLUS, or MULT + MULT.
4905 Put a MULT first, leaving PLUS + PLUS, MULT + PLUS, or MULT + MULT.
4906 Recurse to associate the second PLUS. */
4907 if (GET_CODE (arg1) == MULT)
4908 tem = arg0, arg0 = arg1, arg1 = tem;
4909
4910 if (GET_CODE (arg1) == PLUS)
4911 return simplify_giv_expr (gen_rtx (PLUS, mode,
4912 gen_rtx (PLUS, mode,
4913 arg0, XEXP (arg1, 0)),
4914 XEXP (arg1, 1)),
4915 benefit);
4916
4917 /* Now must have MULT + MULT. Distribute if same biv, else not giv. */
4918 if (GET_CODE (arg0) != MULT || GET_CODE (arg1) != MULT)
4919 abort ();
4920
4921 if (XEXP (arg0, 0) != XEXP (arg1, 0))
4922 return 0;
4923
4924 return simplify_giv_expr (gen_rtx (MULT, mode,
4925 XEXP (arg0, 0),
4926 gen_rtx (PLUS, mode,
4927 XEXP (arg0, 1),
4928 XEXP (arg1, 1))),
4929 benefit);
4930
4931 case MINUS:
4932 /* Handle "a - b" as "a + b * (-1)". */
4933 return simplify_giv_expr (gen_rtx (PLUS, mode,
4934 XEXP (x, 0),
4935 gen_rtx (MULT, mode,
4936 XEXP (x, 1),
4937 gen_rtx (CONST_INT,
4938 VOIDmode, -1))),
4939 benefit);
4940
4941 case MULT:
4942 arg0 = simplify_giv_expr (XEXP (x, 0), benefit);
4943 arg1 = simplify_giv_expr (XEXP (x, 1), benefit);
4944 if (arg0 == 0 || arg1 == 0)
4945 return 0;
4946
4947 /* Put constant last, CONST_INT last if both constant. */
4948 if ((GET_CODE (arg0) == USE || GET_CODE (arg0) == CONST_INT)
4949 && GET_CODE (arg1) != CONST_INT)
4950 tem = arg0, arg0 = arg1, arg1 = tem;
4951
4952 /* If second argument is not now constant, not giv. */
4953 if (GET_CODE (arg1) != USE && GET_CODE (arg1) != CONST_INT)
4954 return 0;
4955
4956 /* Handle multiply by 0 or 1. */
4957 if (arg1 == const0_rtx)
4958 return const0_rtx;
4959
4960 else if (arg1 == const1_rtx)
4961 return arg0;
4962
4963 switch (GET_CODE (arg0))
4964 {
4965 case REG:
4966 /* biv * invar. Done. */
4967 return gen_rtx (MULT, mode, arg0, arg1);
4968
4969 case CONST_INT:
4970 /* Product of two constants. */
4971 return gen_rtx (CONST_INT, mode, INTVAL (arg0) * INTVAL (arg1));
4972
4973 case USE:
4974 /* invar * invar. Not giv. */
4975 return 0;
4976
4977 case MULT:
4978 /* (a * invar_1) * invar_2. Associate. */
4979 return simplify_giv_expr (gen_rtx (MULT, mode,
4980 XEXP (arg0, 0),
4981 gen_rtx (MULT, mode,
4982 XEXP (arg0, 1), arg1)),
4983 benefit);
4984
4985 case PLUS:
4986 /* (a + invar_1) * invar_2. Distribute. */
4987 return simplify_giv_expr (gen_rtx (PLUS, mode,
4988 gen_rtx (MULT, mode,
4989 XEXP (arg0, 0), arg1),
4990 gen_rtx (MULT, mode,
4991 XEXP (arg0, 1), arg1)),
4992 benefit);
4993
4994 default:
4995 abort ();
4996 }
4997
4998 case ASHIFT:
4999 case LSHIFT:
5000 /* Shift by constant is multiply by power of two. */
5001 if (GET_CODE (XEXP (x, 1)) != CONST_INT)
5002 return 0;
5003
5004 return simplify_giv_expr (gen_rtx (MULT, mode,
5005 XEXP (x, 0),
5006 gen_rtx (CONST_INT, VOIDmode,
5007 1 << INTVAL (XEXP (x, 1)))),
5008 benefit);
5009
5010 case NEG:
5011 /* "-a" is "a * (-1)" */
5012 return simplify_giv_expr (gen_rtx (MULT, mode,
5013 XEXP (x, 0),
5014 gen_rtx (CONST_INT, VOIDmode, -1)),
5015 benefit);
5016
5017 case NOT:
5018 /* "~a" is "-a - 1". Silly, but easy. */
5019 return simplify_giv_expr (gen_rtx (MINUS, mode,
5020 gen_rtx (NEG, mode, XEXP (x, 0)),
5021 const1_rtx),
5022 benefit);
5023
5024 case USE:
5025 /* Already in proper form for invariant. */
5026 return x;
5027
5028 case REG:
5029 /* If this is a new register, we can't deal with it. */
5030 if (REGNO (x) >= max_reg_before_loop)
5031 return 0;
5032
5033 /* Check for biv or giv. */
5034 switch (reg_iv_type[REGNO (x)])
5035 {
5036 case BASIC_INDUCT:
5037 return x;
5038 case GENERAL_INDUCT:
5039 {
5040 struct induction *v = reg_iv_info[REGNO (x)];
5041
5042 /* Form expression from giv and add benefit. Ensure this giv
5043 can derive another and subtract any needed adjustment if so. */
5044 *benefit += v->benefit;
5045 if (v->cant_derive)
5046 return 0;
5047
5048 tem = gen_rtx (PLUS, mode, gen_rtx (MULT, mode,
5049 v->src_reg, v->mult_val),
5050 v->add_val);
5051 if (v->derive_adjustment)
5052 tem = gen_rtx (MINUS, mode, tem, v->derive_adjustment);
5053 return simplify_giv_expr (tem, benefit);
5054 }
5055 }
5056
5057 /* Fall through to general case. */
5058 default:
5059 /* If invariant, return as USE (unless CONST_INT).
5060 Otherwise, not giv. */
5061 if (GET_CODE (x) == USE)
5062 x = XEXP (x, 0);
5063
5064 if (invariant_p (x) == 1)
5065 {
5066 if (GET_CODE (x) == CONST_INT)
5067 return x;
5068 else
5069 return gen_rtx (USE, mode, x);
5070 }
5071 else
5072 return 0;
5073 }
5074 }
5075 \f
5076 /* Help detect a giv that is calculated by several consecutive insns;
5077 for example,
5078 giv = biv * M
5079 giv = giv + A
5080 The caller has already identified the first insn P as having a giv as dest;
5081 we check that all other insns that set the same register follow
5082 immediately after P, that they alter nothing else,
5083 and that the result of the last is still a giv.
5084
5085 The value is 0 if the reg set in P is not really a giv.
5086 Otherwise, the value is the amount gained by eliminating
5087 all the consecutive insns that compute the value.
5088
5089 FIRST_BENEFIT is the amount gained by eliminating the first insn, P.
5090 SRC_REG is the reg of the biv; DEST_REG is the reg of the giv.
5091
5092 The coefficients of the ultimate giv value are stored in
5093 *MULT_VAL and *ADD_VAL. */
5094
5095 static int
5096 consec_sets_giv (first_benefit, p, src_reg, dest_reg,
5097 add_val, mult_val)
5098 int first_benefit;
5099 rtx p;
5100 rtx src_reg;
5101 rtx dest_reg;
5102 rtx *add_val;
5103 rtx *mult_val;
5104 {
5105 int count;
5106 enum rtx_code code;
5107 int benefit;
5108 rtx temp;
5109 rtx set;
5110
5111 /* Indicate that this is a giv so that we can update the value produced in
5112 each insn of the multi-insn sequence.
5113
5114 This induction structure will be used only by the call to
5115 general_induction_var below, so we can allocate it on our stack.
5116 If this is a giv, our caller will replace the induct var entry with
5117 a new induction structure. */
5118 struct induction *v
5119 = (struct induction *) alloca (sizeof (struct induction));
5120 v->src_reg = src_reg;
5121 v->mult_val = *mult_val;
5122 v->add_val = *add_val;
5123 v->benefit = first_benefit;
5124 v->cant_derive = 0;
5125 v->derive_adjustment = 0;
5126
5127 reg_iv_type[REGNO (dest_reg)] = GENERAL_INDUCT;
5128 reg_iv_info[REGNO (dest_reg)] = v;
5129
5130 count = n_times_set[REGNO (dest_reg)] - 1;
5131
5132 while (count > 0)
5133 {
5134 p = NEXT_INSN (p);
5135 code = GET_CODE (p);
5136
5137 /* If libcall, skip to end of call sequence. */
5138 if (code == INSN && (temp = find_reg_note (p, REG_LIBCALL, 0)))
5139 p = XEXP (temp, 0);
5140
5141 if (code == INSN
5142 && (set = single_set (p))
5143 && GET_CODE (SET_DEST (set)) == REG
5144 && SET_DEST (set) == dest_reg
5145 && ((benefit = general_induction_var (SET_SRC (set), &src_reg,
5146 add_val, mult_val))
5147 /* Giv created by equivalent expression. */
5148 || ((temp = find_reg_note (p, REG_EQUAL, 0))
5149 && (benefit = general_induction_var (XEXP (temp, 0), &src_reg,
5150 add_val, mult_val))))
5151 && src_reg == v->src_reg)
5152 {
5153 if (find_reg_note (p, REG_RETVAL, 0))
5154 benefit += libcall_benefit (p);
5155
5156 count--;
5157 v->mult_val = *mult_val;
5158 v->add_val = *add_val;
5159 v->benefit = benefit;
5160 }
5161 else if (code != NOTE)
5162 {
5163 /* Allow insns that set something other than this giv to a
5164 constant. Such insns are needed on machines which cannot
5165 include long constants and should not disqualify a giv. */
5166 if (code == INSN
5167 && (set = single_set (p))
5168 && SET_DEST (set) != dest_reg
5169 && CONSTANT_P (SET_SRC (set)))
5170 continue;
5171
5172 reg_iv_type[REGNO (dest_reg)] = UNKNOWN_INDUCT;
5173 return 0;
5174 }
5175 }
5176
5177 return v->benefit;
5178 }
5179 \f
5180 /* Return an rtx, if any, that expresses giv G2 as a function of the register
5181 represented by G1. If no such expression can be found, or it is clear that
5182 it cannot possibly be a valid address, 0 is returned.
5183
5184 To perform the computation, we note that
5185 G1 = a * v + b and
5186 G2 = c * v + d
5187 where `v' is the biv.
5188
5189 So G2 = (c/a) * G1 + (d - b*c/a) */
5190
5191 #ifdef ADDRESS_COST
5192 static rtx
5193 express_from (g1, g2)
5194 struct induction *g1, *g2;
5195 {
5196 rtx mult, add;
5197
5198 /* The value that G1 will be multiplied by must be a constant integer. Also,
5199 the only chance we have of getting a valid address is if b*c/a (see above
5200 for notation) is also an integer. */
5201 if (GET_CODE (g1->mult_val) != CONST_INT
5202 || GET_CODE (g2->mult_val) != CONST_INT
5203 || GET_CODE (g1->add_val) != CONST_INT
5204 || g1->mult_val == const0_rtx
5205 || INTVAL (g2->mult_val) % INTVAL (g1->mult_val) != 0)
5206 return 0;
5207
5208 mult = gen_rtx (CONST_INT, VOIDmode,
5209 INTVAL (g2->mult_val) / INTVAL (g1->mult_val));
5210 add = plus_constant (g2->add_val, - INTVAL (g1->add_val) * INTVAL (mult));
5211
5212 /* Form simplified final result. */
5213 if (mult == const0_rtx)
5214 return add;
5215 else if (mult == const1_rtx)
5216 mult = g1->dest_reg;
5217 else
5218 mult = gen_rtx (MULT, g2->mode, g1->dest_reg, mult);
5219
5220 if (add == const0_rtx)
5221 return mult;
5222 else
5223 return gen_rtx (PLUS, g2->mode, mult, add);
5224 }
5225 #endif
5226 \f
5227 /* Return 1 if giv G2 can be combined with G1. This means that G2 can use
5228 (either directly or via an address expression) a register used to represent
5229 G1. Set g2->new_reg to a represtation of G1 (normally just
5230 g1->dest_reg). */
5231
5232 static int
5233 combine_givs_p (g1, g2)
5234 struct induction *g1, *g2;
5235 {
5236 rtx tem;
5237
5238 /* If these givs are identical, they can be combined. */
5239 if (rtx_equal_p (g1->mult_val, g2->mult_val)
5240 && rtx_equal_p (g1->add_val, g2->add_val))
5241 {
5242 g2->new_reg = g1->dest_reg;
5243 return 1;
5244 }
5245
5246 #ifdef ADDRESS_COST
5247 /* If G2 can be expressed as a function of G1 and that function is valid
5248 as an address and no more expensive than using a register for G2,
5249 the expression of G2 in terms of G1 can be used. */
5250 if (g2->giv_type == DEST_ADDR
5251 && (tem = express_from (g1, g2)) != 0
5252 && memory_address_p (g2->mem_mode, tem)
5253 && ADDRESS_COST (tem) <= ADDRESS_COST (*g2->location))
5254 {
5255 g2->new_reg = tem;
5256 return 1;
5257 }
5258 #endif
5259
5260 return 0;
5261 }
5262 \f
5263 /* Check all pairs of givs for iv_class BL and see if any can be combined with
5264 any other. If so, point SAME to the giv combined with and set NEW_REG to
5265 be an expression (in terms of the other giv's DEST_REG) equivalent to the
5266 giv. Also, update BENEFIT and related fields for cost/benefit analysis. */
5267
5268 static void
5269 combine_givs (bl)
5270 struct iv_class *bl;
5271 {
5272 struct induction *g1, *g2;
5273 int pass;
5274
5275 for (g1 = bl->giv; g1; g1 = g1->next_iv)
5276 for (pass = 0; pass <= 1; pass++)
5277 for (g2 = bl->giv; g2; g2 = g2->next_iv)
5278 if (g1 != g2
5279 /* First try to combine with replaceable givs, then all givs. */
5280 && (g1->replaceable || pass == 1)
5281 /* If either has already been combined or is to be ignored, can't
5282 combine. */
5283 && ! g1->ignore && ! g2->ignore && ! g1->same && ! g2->same
5284 /* If something has been based on G2, G2 cannot itself be based
5285 on something else. */
5286 && ! g2->combined_with
5287 && combine_givs_p (g1, g2))
5288 {
5289 /* g2->new_reg set by `combine_givs_p' */
5290 g2->same = g1;
5291 g1->combined_with = 1;
5292 g1->benefit += g2->benefit;
5293 /* ??? The new final_[bg]iv_value code does a much better job
5294 of finding replaceable giv's, and hence this code may no
5295 longer be necessary. */
5296 if (! g2->replaceable && REG_USERVAR_P (g2->dest_reg))
5297 g1->benefit -= copy_cost;
5298 g1->lifetime += g2->lifetime;
5299 g1->times_used += g2->times_used;
5300
5301 if (loop_dump_stream)
5302 fprintf (loop_dump_stream, "giv at %d combined with giv at %d\n",
5303 INSN_UID (g2->insn), INSN_UID (g1->insn));
5304 }
5305 }
5306 \f
5307 /* EMIT code before INSERT_BEFORE to set REG = B * M + A. */
5308
5309 void
5310 emit_iv_add_mult (b, m, a, reg, insert_before)
5311 rtx b; /* initial value of basic induction variable */
5312 rtx m; /* multiplicative constant */
5313 rtx a; /* additive constant */
5314 rtx reg; /* destination register */
5315 rtx insert_before;
5316 {
5317 rtx seq;
5318 rtx result;
5319
5320 /* Prevent unexpected sharing of these rtx. */
5321 a = copy_rtx (a);
5322 b = copy_rtx (b);
5323
5324 /* Increase the lifetime of any invariants moved further in code. */
5325 update_reg_last_use (a, insert_before);
5326 update_reg_last_use (b, insert_before);
5327 update_reg_last_use (m, insert_before);
5328
5329 start_sequence ();
5330 result = expand_mult_add (b, reg, m, a, GET_MODE (reg), 0);
5331 if (reg != result)
5332 emit_move_insn (reg, result);
5333 seq = gen_sequence ();
5334 end_sequence ();
5335
5336 emit_insn_before (seq, insert_before);
5337 }
5338 \f
5339 /* Test whether A * B can be computed without
5340 an actual multiply insn. Value is 1 if so. */
5341
5342 static int
5343 product_cheap_p (a, b)
5344 rtx a;
5345 rtx b;
5346 {
5347 int i;
5348 rtx tmp;
5349 struct obstack *old_rtl_obstack = rtl_obstack;
5350 char *storage = (char *) obstack_alloc (&temp_obstack, 0);
5351 int win = 1;
5352
5353 /* If only one is constant, make it B. */
5354 if (GET_CODE (a) == CONST_INT)
5355 tmp = a, a = b, b = tmp;
5356
5357 /* If first constant, both constant, so don't need multiply. */
5358 if (GET_CODE (a) == CONST_INT)
5359 return 1;
5360
5361 /* If second not constant, neither is constant, so would need multiply. */
5362 if (GET_CODE (b) != CONST_INT)
5363 return 0;
5364
5365 /* One operand is constant, so might not need multiply insn. Generate the
5366 code for the multiply and see if a call or multiply, or long sequence
5367 of insns is generated. */
5368
5369 rtl_obstack = &temp_obstack;
5370 start_sequence ();
5371 expand_mult (GET_MODE (a), a, b, 0, 0);
5372 tmp = gen_sequence ();
5373 end_sequence ();
5374
5375 if (GET_CODE (tmp) == SEQUENCE)
5376 {
5377 if (XVEC (tmp, 0) == 0)
5378 win = 1;
5379 else if (XVECLEN (tmp, 0) > 3)
5380 win = 0;
5381 else
5382 for (i = 0; i < XVECLEN (tmp, 0); i++)
5383 {
5384 rtx insn = XVECEXP (tmp, 0, i);
5385
5386 if (GET_CODE (insn) != INSN
5387 || (GET_CODE (PATTERN (insn)) == SET
5388 && GET_CODE (SET_SRC (PATTERN (insn))) == MULT)
5389 || (GET_CODE (PATTERN (insn)) == PARALLEL
5390 && GET_CODE (XVECEXP (PATTERN (insn), 0, 0)) == SET
5391 && GET_CODE (SET_SRC (XVECEXP (PATTERN (insn), 0, 0))) == MULT))
5392 {
5393 win = 0;
5394 break;
5395 }
5396 }
5397 }
5398 else if (GET_CODE (tmp) == SET
5399 && GET_CODE (SET_SRC (tmp)) == MULT)
5400 win = 0;
5401 else if (GET_CODE (tmp) == PARALLEL
5402 && GET_CODE (XVECEXP (tmp, 0, 0)) == SET
5403 && GET_CODE (SET_SRC (XVECEXP (tmp, 0, 0))) == MULT)
5404 win = 0;
5405
5406 /* Free any storage we obtained in generating this multiply and restore rtl
5407 allocation to its normal obstack. */
5408 obstack_free (&temp_obstack, storage);
5409 rtl_obstack = old_rtl_obstack;
5410
5411 return win;
5412 }
5413 \f
5414 /* Check to see if loop can be terminated by a "decrement and branch until
5415 zero" instruction. If so, add a REG_NONNEG note to the branch insn if so.
5416 Also try reversing an increment loop to a decrement loop
5417 to see if the optimization can be performed.
5418 Value is nonzero if optimization was performed. */
5419
5420 /* This is useful even if the architecture doesn't have such an insn,
5421 because it might change a loops which increments from 0 to n to a loop
5422 which decrements from n to 0. A loop that decrements to zero is usually
5423 faster than one that increments from zero. */
5424
5425 /* ??? This could be rewritten to use some of the loop unrolling procedures,
5426 such as approx_final_value, biv_total_increment, loop_iterations, and
5427 final_[bg]iv_value. */
5428
5429 static int
5430 check_dbra_loop (loop_end, insn_count, loop_start)
5431 rtx loop_end;
5432 int insn_count;
5433 rtx loop_start;
5434 {
5435 struct iv_class *bl;
5436 rtx reg;
5437 rtx jump_label;
5438 rtx final_value;
5439 rtx start_value;
5440 enum rtx_code branch_code;
5441 rtx new_add_val;
5442 rtx comparison;
5443 rtx before_comparison;
5444 rtx p;
5445
5446 /* If last insn is a conditional branch, and the insn before tests a
5447 register value, try to optimize it. Otherwise, we can't do anything. */
5448
5449 comparison = get_condition_for_loop (PREV_INSN (loop_end));
5450 if (comparison == 0)
5451 return 0;
5452
5453 /* Check all of the bivs to see if the compare uses one of them.
5454 Skip biv's set more than once because we can't guarantee that
5455 it will be zero on the last iteration. Also skip if the biv is
5456 used between its update and the test insn. */
5457
5458 for (bl = loop_iv_list; bl; bl = bl->next)
5459 {
5460 if (bl->biv_count == 1
5461 && bl->biv->dest_reg == XEXP (comparison, 0)
5462 && ! reg_used_between_p (regno_reg_rtx[bl->regno], bl->biv->insn,
5463 PREV_INSN (PREV_INSN (loop_end))))
5464 break;
5465 }
5466
5467 if (! bl)
5468 return 0;
5469
5470 /* Look for the case where the basic induction variable is always
5471 nonnegative, and equals zero on the last iteration.
5472 In this case, add a reg_note REG_NONNEG, which allows the
5473 m68k DBRA instruction to be used. */
5474
5475 if (((GET_CODE (comparison) == GT
5476 && GET_CODE (XEXP (comparison, 1)) == CONST_INT
5477 && INTVAL (XEXP (comparison, 1)) == -1)
5478 || (GET_CODE (comparison) == NE && XEXP (comparison, 1) == const0_rtx))
5479 && GET_CODE (bl->biv->add_val) == CONST_INT
5480 && INTVAL (bl->biv->add_val) < 0)
5481 {
5482 /* Initial value must be greater than 0,
5483 init_val % -dec_value == 0 to ensure that it equals zero on
5484 the last iteration */
5485
5486 if (GET_CODE (bl->initial_value) == CONST_INT
5487 && INTVAL (bl->initial_value) > 0
5488 && (INTVAL (bl->initial_value) %
5489 (-INTVAL (bl->biv->add_val))) == 0)
5490 {
5491 /* register always nonnegative, add REG_NOTE to branch */
5492 REG_NOTES (PREV_INSN (loop_end))
5493 = gen_rtx (EXPR_LIST, REG_NONNEG, 0,
5494 REG_NOTES (PREV_INSN (loop_end)));
5495 bl->nonneg = 1;
5496
5497 return 1;
5498 }
5499
5500 /* If the decrement is 1 and the value was tested as >= 0 before
5501 the loop, then we can safely optimize. */
5502 for (p = loop_start; p; p = PREV_INSN (p))
5503 {
5504 if (GET_CODE (p) == CODE_LABEL)
5505 break;
5506 if (GET_CODE (p) != JUMP_INSN)
5507 continue;
5508
5509 before_comparison = get_condition_for_loop (p);
5510 if (before_comparison
5511 && XEXP (before_comparison, 0) == bl->biv->dest_reg
5512 && GET_CODE (before_comparison) == LT
5513 && XEXP (before_comparison, 1) == const0_rtx
5514 && ! reg_set_between_p (bl->biv->dest_reg, p, loop_start)
5515 && INTVAL (bl->biv->add_val) == -1)
5516 {
5517 REG_NOTES (PREV_INSN (loop_end))
5518 = gen_rtx (EXPR_LIST, REG_NONNEG, 0,
5519 REG_NOTES (PREV_INSN (loop_end)));
5520 bl->nonneg = 1;
5521
5522 return 1;
5523 }
5524 }
5525 }
5526 else if (num_mem_sets <= 1)
5527 {
5528 /* Try to change inc to dec, so can apply above optimization. */
5529 /* Can do this if:
5530 all registers modified are induction variables or invariant,
5531 all memory references have non-overlapping addresses
5532 (obviously true if only one write)
5533 allow 2 insns for the compare/jump at the end of the loop. */
5534 int num_nonfixed_reads = 0;
5535 /* 1 if the iteration var is used only to count iterations. */
5536 int no_use_except_counting = 0;
5537
5538 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
5539 if (GET_RTX_CLASS (GET_CODE (p)) == 'i')
5540 num_nonfixed_reads += count_nonfixed_reads (PATTERN (p));
5541
5542 if (bl->giv_count == 0
5543 && ! loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]])
5544 {
5545 rtx bivreg = regno_reg_rtx[bl->regno];
5546
5547 /* If there are no givs for this biv, and the only exit is the
5548 fall through at the end of the the loop, then
5549 see if perhaps there are no uses except to count. */
5550 no_use_except_counting = 1;
5551 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
5552 if (GET_RTX_CLASS (GET_CODE (p)) == 'i')
5553 {
5554 rtx set = single_set (p);
5555
5556 if (set && GET_CODE (SET_DEST (set)) == REG
5557 && REGNO (SET_DEST (set)) == bl->regno)
5558 /* An insn that sets the biv is okay. */
5559 ;
5560 else if (p == prev_nonnote_insn (prev_nonnote_insn (loop_end))
5561 || p == prev_nonnote_insn (loop_end))
5562 /* Don't bother about the end test. */
5563 ;
5564 else if (reg_mentioned_p (bivreg, PATTERN (p)))
5565 /* Any other use of the biv is no good. */
5566 {
5567 no_use_except_counting = 0;
5568 break;
5569 }
5570 }
5571 }
5572
5573 /* This code only acts for innermost loops. Also it simplifies
5574 the memory address check by only reversing loops with
5575 zero or one memory access.
5576 Two memory accesses could involve parts of the same array,
5577 and that can't be reversed. */
5578
5579 if (num_nonfixed_reads <= 1
5580 && !loop_has_call
5581 && (no_use_except_counting
5582 || (bl->giv_count + bl->biv_count + num_mem_sets
5583 + num_movables + 2 == insn_count)))
5584 {
5585 rtx condition = get_condition_for_loop (PREV_INSN (loop_end));
5586 int win;
5587 rtx tem;
5588
5589 /* Loop can be reversed. */
5590 if (loop_dump_stream)
5591 fprintf (loop_dump_stream, "Can reverse loop\n");
5592
5593 /* Now check other conditions:
5594 initial_value must be zero,
5595 final_value % add_val == 0, so that when reversed, the
5596 biv will be zero on the last iteration.
5597
5598 This test can probably be improved since +/- 1 in the constant
5599 can be obtained by changing LT to LE and vice versa; this is
5600 confusing. */
5601
5602 if (comparison && bl->initial_value == const0_rtx
5603 && GET_CODE (XEXP (comparison, 1)) == CONST_INT
5604 /* LE gets turned into LT */
5605 && GET_CODE (comparison) == LT
5606 && (INTVAL (XEXP (comparison, 1))
5607 % INTVAL (bl->biv->add_val)) == 0)
5608 {
5609 /* Register will always be nonnegative, with value
5610 0 on last iteration if loop reversed */
5611
5612 /* Save some info needed to produce the new insns. */
5613 reg = bl->biv->dest_reg;
5614 jump_label = XEXP (SET_SRC (PATTERN (PREV_INSN (loop_end))), 1);
5615 new_add_val = gen_rtx (CONST_INT, VOIDmode,
5616 - INTVAL (bl->biv->add_val));
5617
5618 final_value = XEXP (comparison, 1);
5619 start_value = gen_rtx (CONST_INT, VOIDmode,
5620 (INTVAL (XEXP (comparison, 1))
5621 - INTVAL (bl->biv->add_val)));
5622
5623 /* Initialize biv to start_value before loop start.
5624 The old initializing insn will be deleted as a
5625 dead store by flow.c. */
5626 emit_insn_before (gen_move_insn (reg, start_value), loop_start);
5627
5628 /* Add insn to decrement register, and delete insn
5629 that incremented the register. */
5630 p = emit_insn_before (gen_add2_insn (reg, new_add_val),
5631 bl->biv->insn);
5632 delete_insn (bl->biv->insn);
5633
5634 /* Update biv info to reflect its new status. */
5635 bl->biv->insn = p;
5636 bl->initial_value = start_value;
5637 bl->biv->add_val = new_add_val;
5638
5639 /* Inc LABEL_NUSES so that delete_insn will
5640 not delete the label. */
5641 LABEL_NUSES (XEXP (jump_label, 0)) ++;
5642
5643 /* Emit an insn after the end of the loop to set the biv's
5644 proper exit value if it is used anywhere outside the loop. */
5645 if ((regno_last_uid[bl->regno]
5646 != INSN_UID (PREV_INSN (PREV_INSN (loop_end))))
5647 || ! bl->init_insn
5648 || regno_first_uid[bl->regno] != INSN_UID (bl->init_insn))
5649 emit_insn_after (gen_move_insn (reg, final_value),
5650 loop_end);
5651
5652 /* Delete compare/branch at end of loop. */
5653 delete_insn (PREV_INSN (loop_end));
5654 delete_insn (PREV_INSN (loop_end));
5655
5656 /* Add new compare/branch insn at end of loop. */
5657 start_sequence ();
5658 emit_cmp_insn (reg, const0_rtx, GE, 0, GET_MODE (reg), 0, 0);
5659 emit_jump_insn (gen_bge (XEXP (jump_label, 0)));
5660 tem = gen_sequence ();
5661 end_sequence ();
5662 emit_jump_insn_before (tem, loop_end);
5663
5664 for (tem = PREV_INSN (loop_end);
5665 tem && GET_CODE (tem) != JUMP_INSN; tem = PREV_INSN (tem))
5666 ;
5667 if (tem)
5668 {
5669 JUMP_LABEL (tem) = XEXP (jump_label, 0);
5670
5671 /* Increment of LABEL_NUSES done above. */
5672 /* Register is now always nonnegative,
5673 so add REG_NONNEG note to the branch. */
5674 REG_NOTES (tem) = gen_rtx (EXPR_LIST, REG_NONNEG, 0,
5675 REG_NOTES (tem));
5676 }
5677
5678 bl->nonneg = 1;
5679
5680 /* Mark that this biv has been reversed. Each giv which depends
5681 on this biv, and which is also live past the end of the loop
5682 will have to be fixed up. */
5683
5684 bl->reversed = 1;
5685
5686 if (loop_dump_stream)
5687 fprintf (loop_dump_stream,
5688 "Reversed loop and added reg_nonneg\n");
5689
5690 return 1;
5691 }
5692 }
5693 }
5694
5695 return 0;
5696 }
5697 \f
5698 /* Verify whether the biv BL appears to be eliminable,
5699 based on the insns in the loop that refer to it.
5700 LOOP_START is the first insn of the loop, and END is the end insn.
5701
5702 If ELIMINATE_P is non-zero, actually do the elimination.
5703
5704 THRESHOLD and INSN_COUNT are from loop_optimize and are used to
5705 determine whether invariant insns should be placed inside or at the
5706 start of the loop. */
5707
5708 static int
5709 maybe_eliminate_biv (bl, loop_start, end, eliminate_p, threshold, insn_count)
5710 struct iv_class *bl;
5711 rtx loop_start;
5712 rtx end;
5713 int eliminate_p;
5714 int threshold, insn_count;
5715 {
5716 rtx reg = bl->biv->dest_reg;
5717 rtx p, set;
5718 struct induction *v;
5719
5720 /* Scan all insns in the loop, stopping if we find one that uses the
5721 biv in a way that we cannot eliminate. */
5722
5723 for (p = loop_start; p != end; p = NEXT_INSN (p))
5724 {
5725 enum rtx_code code = GET_CODE (p);
5726 rtx where = threshold >= insn_count ? loop_start : p;
5727
5728 if ((code == INSN || code == JUMP_INSN || code == CALL_INSN)
5729 && reg_mentioned_p (reg, PATTERN (p))
5730 && ! maybe_eliminate_biv_1 (PATTERN (p), p, bl, eliminate_p, where))
5731 {
5732 if (loop_dump_stream)
5733 fprintf (loop_dump_stream,
5734 "Cannot eliminate biv %d: biv used in insn %d.\n",
5735 bl->regno, INSN_UID (p));
5736 break;
5737 }
5738 }
5739
5740 if (p == end)
5741 {
5742 if (loop_dump_stream)
5743 fprintf (loop_dump_stream, "biv %d %s eliminated.\n",
5744 bl->regno, eliminate_p ? "was" : "can be");
5745 return 1;
5746 }
5747
5748 return 0;
5749 }
5750 \f
5751 /* If BL appears in X (part of the pattern of INSN), see if we can
5752 eliminate its use. If so, return 1. If not, return 0.
5753
5754 If BIV does not appear in X, return 1.
5755
5756 If ELIMINATE_P is non-zero, actually do the elimination. WHERE indicates
5757 where extra insns should be added. Depending on how many items have been
5758 moved out of the loop, it will either be before INSN or at the start of
5759 the loop. */
5760
5761 static int
5762 maybe_eliminate_biv_1 (x, insn, bl, eliminate_p, where)
5763 rtx x, insn;
5764 struct iv_class *bl;
5765 int eliminate_p;
5766 rtx where;
5767 {
5768 enum rtx_code code = GET_CODE (x);
5769 rtx reg = bl->biv->dest_reg;
5770 enum machine_mode mode = GET_MODE (reg);
5771 struct induction *v;
5772 rtx arg, new, tem;
5773 int arg_operand;
5774 char *fmt;
5775 int i, j;
5776
5777 switch (code)
5778 {
5779 case REG:
5780 /* If we haven't already been able to do something with this BIV,
5781 we can't eliminate it. */
5782 if (x == reg)
5783 return 0;
5784 return 1;
5785
5786 case SET:
5787 /* If this sets the BIV, it is not a problem. */
5788 if (SET_DEST (x) == reg)
5789 return 1;
5790
5791 /* If this is an insn that defines a giv, it is also ok because
5792 it will go away when the giv is reduced. */
5793 for (v = bl->giv; v; v = v->next_iv)
5794 if (v->giv_type == DEST_REG && SET_DEST (x) == v->dest_reg)
5795 return 1;
5796
5797 #ifdef HAVE_cc0
5798 if (SET_DEST (x) == cc0_rtx && SET_SRC (x) == reg)
5799 {
5800 /* Can replace with any giv that was reduced and
5801 that has (MULT_VAL != 0) and (ADD_VAL == 0).
5802 Require a constant for MULT_VAL, so we know it's nonzero. */
5803
5804 for (v = bl->giv; v; v = v->next_iv)
5805 if (CONSTANT_P (v->mult_val) && v->mult_val != const0_rtx
5806 && v->add_val == const0_rtx
5807 && ! v->ignore && ! v->maybe_dead
5808 && v->mode == mode)
5809 {
5810 if (! eliminate_p)
5811 return 1;
5812
5813 /* If the giv has the opposite direction of change,
5814 then reverse the comparison. */
5815 if (INTVAL (v->mult_val) < 0)
5816 new = gen_rtx (COMPARE, GET_MODE (v->new_reg),
5817 const0_rtx, v->new_reg);
5818 else
5819 new = v->new_reg;
5820
5821 /* We can probably test that giv's reduced reg. */
5822 if (validate_change (insn, &SET_SRC (x), new, 0))
5823 return 1;
5824 }
5825
5826 /* Look for a giv with (MULT_VAL != 0) and (ADD_VAL != 0);
5827 replace test insn with a compare insn (cmp REDUCED_GIV ADD_VAL).
5828 Require a constant for MULT_VAL, so we know it's nonzero. */
5829
5830 for (v = bl->giv; v; v = v->next_iv)
5831 if (CONSTANT_P (v->mult_val) && v->mult_val != const0_rtx
5832 && ! v->ignore && ! v->maybe_dead
5833 && v->mode == mode)
5834 {
5835 if (! eliminate_p)
5836 return 1;
5837
5838 /* If the giv has the opposite direction of change,
5839 then reverse the comparison. */
5840 if (INTVAL (v->mult_val) < 0)
5841 new = gen_rtx (COMPARE, VOIDmode, copy_rtx (v->add_val),
5842 v->new_reg);
5843 else
5844 new = gen_rtx (COMPARE, VOIDmode, v->new_reg,
5845 copy_rtx (v->add_val));
5846
5847 /* Replace biv with the giv's reduced register. */
5848 update_reg_last_use (v->add_val, insn);
5849 if (validate_change (insn, &SET_SRC (PATTERN (insn)), new, 0))
5850 return 1;
5851
5852 /* Insn doesn't support that constant or invariant. Copy it
5853 into a register (it will be a loop invariant.) */
5854 tem = gen_reg_rtx (GET_MODE (v->new_reg));
5855
5856 emit_insn_before (gen_move_insn (tem, copy_rtx (v->add_val)),
5857 where);
5858
5859 if (validate_change (insn, &SET_SRC (PATTERN (insn)),
5860 gen_rtx (COMPARE, VOIDmode,
5861 v->new_reg, tem), 0))
5862 return 1;
5863 }
5864 }
5865 #endif
5866 break;
5867
5868 case COMPARE:
5869 case EQ: case NE:
5870 case GT: case GE: case GTU: case GEU:
5871 case LT: case LE: case LTU: case LEU:
5872 /* See if either argument is the biv. */
5873 if (XEXP (x, 0) == reg)
5874 arg = XEXP (x, 1), arg_operand = 1;
5875 else if (XEXP (x, 1) == reg)
5876 arg = XEXP (x, 0), arg_operand = 0;
5877 else
5878 break;
5879
5880 if (CONSTANT_P (arg))
5881 {
5882 /* First try to replace with any giv that has constant positive
5883 mult_val and constant add_val. We might be able to support
5884 negative mult_val, but it seems complex to do it in general. */
5885
5886 for (v = bl->giv; v; v = v->next_iv)
5887 if (CONSTANT_P (v->mult_val) && INTVAL (v->mult_val) > 0
5888 && CONSTANT_P (v->add_val)
5889 && ! v->ignore && ! v->maybe_dead
5890 && v->mode == mode)
5891 {
5892 if (! eliminate_p)
5893 return 1;
5894
5895 /* Replace biv with the giv's reduced reg. */
5896 XEXP (x, 1-arg_operand) = v->new_reg;
5897
5898 /* If all constants are actually constant integers and
5899 the derived constant can be directly placed in the COMPARE,
5900 do so. */
5901 if (GET_CODE (arg) == CONST_INT
5902 && GET_CODE (v->mult_val) == CONST_INT
5903 && GET_CODE (v->add_val) == CONST_INT
5904 && validate_change (insn, &XEXP (x, arg_operand),
5905 gen_rtx (CONST_INT, VOIDmode,
5906 (INTVAL (arg)
5907 * INTVAL (v->mult_val)
5908 + INTVAL (v->add_val))), 0))
5909 return 1;
5910
5911 /* Otherwise, load it into a register. */
5912 tem = gen_reg_rtx (mode);
5913 emit_iv_add_mult (arg, v->mult_val, v->add_val, tem, where);
5914 if (validate_change (insn, &XEXP (x, arg_operand), tem, 0))
5915 return 1;
5916
5917 /* If that failed, put back the change we made above. */
5918 XEXP (x, 1-arg_operand) = reg;
5919 }
5920
5921 /* Look for giv with positive constant mult_val and nonconst add_val.
5922 Insert insns to calculate new compare value. */
5923
5924 for (v = bl->giv; v; v = v->next_iv)
5925 if (CONSTANT_P (v->mult_val) && INTVAL (v->mult_val) > 0
5926 && ! v->ignore && ! v->maybe_dead
5927 && v->mode == mode)
5928 {
5929 rtx tem;
5930
5931 if (! eliminate_p)
5932 return 1;
5933
5934 tem = gen_reg_rtx (mode);
5935
5936 /* Replace biv with giv's reduced register. */
5937 validate_change (insn, &XEXP (x, 1 - arg_operand),
5938 v->new_reg, 1);
5939
5940 /* Compute value to compare against. */
5941 emit_iv_add_mult (arg, v->mult_val, v->add_val, tem, where);
5942 /* Use it in this insn. */
5943 validate_change (insn, &XEXP (x, arg_operand), tem, 1);
5944 if (apply_change_group ())
5945 return 1;
5946 }
5947 }
5948 else if (GET_CODE (arg) == REG || GET_CODE (arg) == MEM)
5949 {
5950 if (invariant_p (arg) == 1)
5951 {
5952 /* Look for giv with constant positive mult_val and nonconst
5953 add_val. Insert insns to compute new compare value. */
5954
5955 for (v = bl->giv; v; v = v->next_iv)
5956 if (CONSTANT_P (v->mult_val) && INTVAL (v->mult_val) > 0
5957 && ! v->ignore && ! v->maybe_dead
5958 && v->mode == mode)
5959 {
5960 rtx tem;
5961
5962 if (! eliminate_p)
5963 return 1;
5964
5965 tem = gen_reg_rtx (mode);
5966
5967 /* Replace biv with giv's reduced register. */
5968 validate_change (insn, &XEXP (x, 1 - arg_operand),
5969 v->new_reg, 1);
5970
5971 /* Compute value to compare against. */
5972 emit_iv_add_mult (arg, v->mult_val, v->add_val,
5973 tem, where);
5974 validate_change (insn, &XEXP (x, arg_operand), tem, 1);
5975 if (apply_change_group ())
5976 return 1;
5977 }
5978 }
5979
5980 /* This code has problems. Basically, you can't know when
5981 seeing if we will eliminate BL, whether a particular giv
5982 of ARG will be reduced. If it isn't going to be reduced,
5983 we can't eliminate BL. We can try forcing it to be reduced,
5984 but that can generate poor code.
5985
5986 The problem is that the benefit of reducing TV, below should
5987 be increased if BL can actually be eliminated, but this means
5988 we might have to do a topological sort of the order in which
5989 we try to process biv. It doesn't seem worthwhile to do
5990 this sort of thing now. */
5991
5992 #if 0
5993 /* Otherwise the reg compared with had better be a biv. */
5994 if (GET_CODE (arg) != REG
5995 || reg_iv_type[REGNO (arg)] != BASIC_INDUCT)
5996 return 0;
5997
5998 /* Look for a pair of givs, one for each biv,
5999 with identical coefficients. */
6000 for (v = bl->giv; v; v = v->next_iv)
6001 {
6002 struct induction *tv;
6003
6004 if (v->ignore || v->maybe_dead || v->mode != mode)
6005 continue;
6006
6007 for (tv = reg_biv_class[REGNO (arg)]->giv; tv; tv = tv->next_iv)
6008 if (! tv->ignore && ! tv->maybe_dead
6009 && rtx_equal_p (tv->mult_val, v->mult_val)
6010 && rtx_equal_p (tv->add_val, v->add_val)
6011 && tv->mode == mode)
6012 {
6013 if (! eliminate_p)
6014 return 1;
6015
6016 /* Replace biv with its giv's reduced reg. */
6017 XEXP (x, 1-arg_operand) = v->new_reg;
6018 /* Replace other operand with the other giv's
6019 reduced reg. */
6020 XEXP (x, arg_operand) = tv->new_reg;
6021 return 1;
6022 }
6023 }
6024 #endif
6025 }
6026
6027 /* If we get here, the biv can't be eliminated. */
6028 return 0;
6029
6030 case MEM:
6031 /* If this address is a DEST_ADDR giv, it doesn't matter if the
6032 biv is used in it, since it will be replaced. */
6033 for (v = bl->giv; v; v = v->next_iv)
6034 if (v->giv_type == DEST_ADDR && v->location == &XEXP (x, 0))
6035 return 1;
6036 break;
6037 }
6038
6039 /* See if any subexpression fails elimination. */
6040 fmt = GET_RTX_FORMAT (code);
6041 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
6042 {
6043 switch (fmt[i])
6044 {
6045 case 'e':
6046 if (! maybe_eliminate_biv_1 (XEXP (x, i), insn, bl,
6047 eliminate_p, where))
6048 return 0;
6049 break;
6050
6051 case 'E':
6052 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
6053 if (! maybe_eliminate_biv_1 (XVECEXP (x, i, j), insn, bl,
6054 eliminate_p, where))
6055 return 0;
6056 break;
6057 }
6058 }
6059
6060 return 1;
6061 }
6062 \f
6063 /* Return nonzero if the last use of REG
6064 is in an insn following INSN in the same basic block. */
6065
6066 static int
6067 last_use_this_basic_block (reg, insn)
6068 rtx reg;
6069 rtx insn;
6070 {
6071 rtx n;
6072 for (n = insn;
6073 n && GET_CODE (n) != CODE_LABEL && GET_CODE (n) != JUMP_INSN;
6074 n = NEXT_INSN (n))
6075 {
6076 if (regno_last_uid[REGNO (reg)] == INSN_UID (n))
6077 return 1;
6078 }
6079 return 0;
6080 }
6081 \f
6082 /* Called via `note_stores' to record the initial value of a biv. Here we
6083 just record the location of the set and process it later. */
6084
6085 static void
6086 record_initial (dest, set)
6087 rtx dest;
6088 rtx set;
6089 {
6090 struct iv_class *bl;
6091
6092 if (GET_CODE (dest) != REG
6093 || REGNO (dest) >= max_reg_before_loop
6094 || reg_iv_type[REGNO (dest)] != BASIC_INDUCT)
6095 return;
6096
6097 bl = reg_biv_class[REGNO (dest)];
6098
6099 /* If this is the first set found, record it. */
6100 if (bl->init_insn == 0)
6101 {
6102 bl->init_insn = note_insn;
6103 bl->init_set = set;
6104 }
6105 }
6106 \f
6107 /* If any of the registers in X are "old" and currently have a last use earlier
6108 than INSN, update them to have a last use of INSN. Their actual last use
6109 will be the previous insn but it will not have a valid uid_luid so we can't
6110 use it. */
6111
6112 static void
6113 update_reg_last_use (x, insn)
6114 rtx x;
6115 rtx insn;
6116 {
6117 /* Check for the case where INSN does not have a valid luid. In this case,
6118 there is no need to modify the regno_last_uid, as this can only happen
6119 when code is inserted after the loop_end to set a pseudo's final value,
6120 and hence this insn will never be the last use of x. */
6121 if (GET_CODE (x) == REG && REGNO (x) < max_reg_before_loop
6122 && INSN_UID (insn) < max_uid_for_loop
6123 && uid_luid[regno_last_uid[REGNO (x)]] < uid_luid[INSN_UID (insn)])
6124 regno_last_uid[REGNO (x)] = INSN_UID (insn);
6125 else
6126 {
6127 register int i, j;
6128 register char *fmt = GET_RTX_FORMAT (GET_CODE (x));
6129 for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
6130 {
6131 if (fmt[i] == 'e')
6132 update_reg_last_use (XEXP (x, i), insn);
6133 else if (fmt[i] == 'E')
6134 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
6135 update_reg_last_use (XVECEXP (x, i, j), insn);
6136 }
6137 }
6138 }
6139 \f
6140 /* Given a jump insn JUMP, return the condition that will cause it to branch
6141 to its JUMP_LABEL. If the condition cannot be understood, or is an
6142 inequality floating-point comparison which needs to be reversed, 0 will
6143 be returned.
6144
6145 If EARLIEST is non-zero, it is a pointer to a place where the earliest
6146 insn used in locating the condition was found. If a replacement test
6147 of the condition is desired, it should be placed in front of that
6148 insn and we will be sure that the inputs are still valid.
6149
6150 The condition will be returned in a canonical form to simplify testing by
6151 callers. Specifically:
6152
6153 (1) The code will always be a comparison operation (EQ, NE, GT, etc.).
6154 (2) Both operands will be machine operands; (cc0) will have been replaced.
6155 (3) If an operand is a constant, it will be the second operand.
6156 (4) (LE x const) will be replaced with (LT x <const+1>) and similarly
6157 for GE, GEU, and LEU. */
6158
6159 rtx
6160 get_condition (jump, earliest)
6161 rtx jump;
6162 rtx *earliest;
6163 {
6164 enum rtx_code code;
6165 rtx prev = jump;
6166 rtx set;
6167 rtx tem;
6168 rtx op0, op1;
6169 int reverse_code = 0;
6170 int did_reverse_condition = 0;
6171
6172 /* If this is not a standard conditional jump, we can't parse it. */
6173 if (GET_CODE (jump) != JUMP_INSN
6174 || ! condjump_p (jump) || simplejump_p (jump))
6175 return 0;
6176
6177 code = GET_CODE (XEXP (SET_SRC (PATTERN (jump)), 0));
6178 op0 = XEXP (XEXP (SET_SRC (PATTERN (jump)), 0), 0);
6179 op1 = XEXP (XEXP (SET_SRC (PATTERN (jump)), 0), 1);
6180
6181 if (earliest)
6182 *earliest = jump;
6183
6184 /* If this branches to JUMP_LABEL when the condition is false, reverse
6185 the condition. */
6186 if (XEXP (XEXP (SET_SRC (PATTERN (jump)), 2), 0) == JUMP_LABEL (jump))
6187 code = reverse_condition (code), did_reverse_condition ^= 1;
6188
6189 /* If we are comparing a register with zero, see if the register is set
6190 in the previous insn to a COMPARE or a comparison operation. Perform
6191 the same tests as a function of STORE_FLAG_VALUE as find_comparison_args
6192 in cse.c */
6193
6194 while (GET_RTX_CLASS (code) == '<' && op1 == const0_rtx)
6195 {
6196 /* Set non-zero when we find something of interest. */
6197 rtx x = 0;
6198
6199 #ifdef HAVE_cc0
6200 /* If comparison with cc0, import actual comparison from compare
6201 insn. */
6202 if (op0 == cc0_rtx)
6203 {
6204 if ((prev = prev_nonnote_insn (prev)) == 0
6205 || GET_CODE (prev) != INSN
6206 || (set = single_set (prev)) == 0
6207 || SET_DEST (set) != cc0_rtx)
6208 return 0;
6209
6210 op0 = SET_SRC (set);
6211 op1 = CONST0_RTX (GET_MODE (op0));
6212 if (earliest)
6213 *earliest = prev;
6214 }
6215 #endif
6216
6217 /* If this is a COMPARE, pick up the two things being compared. */
6218 if (GET_CODE (op0) == COMPARE)
6219 {
6220 op1 = XEXP (op0, 1);
6221 op0 = XEXP (op0, 0);
6222 continue;
6223 }
6224 else if (GET_CODE (op0) != REG)
6225 break;
6226
6227 /* Go back to the previous insn. Stop if it is not an INSN. We also
6228 stop if it isn't a single set or if it has a REG_INC note because
6229 we don't want to bother dealing with it. */
6230
6231 if ((prev = prev_nonnote_insn (prev)) == 0
6232 || GET_CODE (prev) != INSN
6233 || FIND_REG_INC_NOTE (prev, 0)
6234 || (set = single_set (prev)) == 0)
6235 break;
6236
6237 /* If this is setting OP0, get what it sets it to if it looks
6238 relevant. */
6239 if (SET_DEST (set) == op0)
6240 {
6241 enum machine_mode inner_mode = GET_MODE (SET_SRC (set));
6242
6243 if ((GET_CODE (SET_SRC (set)) == COMPARE
6244 || ((code == NE
6245 || (code == LT
6246 && GET_MODE_BITSIZE (inner_mode) <= HOST_BITS_PER_INT
6247 && (STORE_FLAG_VALUE
6248 & (1 << (GET_MODE_BITSIZE (inner_mode) - 1)))))
6249 && GET_RTX_CLASS (GET_CODE (SET_SRC (set))) == '<')))
6250 x = SET_SRC (set);
6251 else if ((code == EQ
6252 || (code == GE
6253 && GET_MODE_BITSIZE (inner_mode) <= HOST_BITS_PER_INT
6254 && (STORE_FLAG_VALUE
6255 & (1 << (GET_MODE_BITSIZE (inner_mode) - 1)))))
6256 && GET_RTX_CLASS (GET_CODE (SET_SRC (set))) == '<')
6257 {
6258 /* We might have reversed a LT to get a GE here. But this wasn't
6259 actually the comparison of data, so we don't flag that we
6260 have had to reverse the condition. */
6261 did_reverse_condition ^= 1;
6262 reverse_code = 1;
6263 x = SET_SRC (set);
6264 }
6265 }
6266
6267 else if (reg_set_p (op0, prev))
6268 /* If this sets OP0, but not directly, we have to give up. */
6269 break;
6270
6271 if (x)
6272 {
6273 if (GET_RTX_CLASS (GET_CODE (x)) == '<')
6274 code = GET_CODE (x);
6275 if (reverse_code)
6276 {
6277 code = reverse_condition (code);
6278 did_reverse_condition ^= 1;
6279 reverse_code = 0;
6280 }
6281
6282 op0 = XEXP (x, 0), op1 = XEXP (x, 1);
6283 if (earliest)
6284 *earliest = prev;
6285 }
6286 }
6287
6288 /* If constant is first, put it last. */
6289 if (CONSTANT_P (op0))
6290 code = swap_condition (code), tem = op0, op0 = op1, op1 = tem;
6291
6292 /* If OP0 is the result of a comparison, we weren't able to find what
6293 was really being compared, so fail. */
6294 if (GET_MODE_CLASS (GET_MODE (op0)) == MODE_CC)
6295 return 0;
6296
6297 /* Canonicalize any ordered comparison with integers involving equality. */
6298 if (GET_CODE (op1) == CONST_INT)
6299 {
6300 int const_val = INTVAL (op1);
6301 unsigned uconst_val = (unsigned) const_val;
6302
6303 switch (code)
6304 {
6305 case LE:
6306 code = LT;
6307 op1 = gen_rtx (CONST_INT, VOIDmode, const_val + 1);
6308 break;
6309
6310 case GE:
6311 code = GT;
6312 op1 = gen_rtx (CONST_INT, VOIDmode, const_val - 1);
6313 break;
6314
6315 case LEU:
6316 code = LTU;
6317 op1 = gen_rtx (CONST_INT, VOIDmode, uconst_val + 1);
6318 break;
6319
6320 case GEU:
6321 code = GTU;
6322 op1 = gen_rtx (CONST_INT, VOIDmode, uconst_val - 1);
6323 break;
6324 }
6325 }
6326
6327 /* If this was floating-point and we reversed anything other than an
6328 EQ or NE, return zero. */
6329 if (TARGET_FLOAT_FORMAT == IEEE_FLOAT_FORMAT
6330 && did_reverse_condition && code != NE && code != EQ
6331 && GET_MODE_CLASS (GET_MODE (op0)) == MODE_FLOAT)
6332 return 0;
6333
6334 #ifdef HAVE_cc0
6335 /* Never return CC0; return zero instead. */
6336 if (op0 == cc0_rtx)
6337 return 0;
6338 #endif
6339
6340 return gen_rtx (code, VOIDmode, op0, op1);
6341 }
6342
6343 /* Similar to above routine, except that we also put an invariant last
6344 unless both operands are invariants. */
6345
6346 rtx
6347 get_condition_for_loop (x)
6348 rtx x;
6349 {
6350 rtx comparison = get_condition (x, 0);
6351
6352 if (comparison == 0
6353 || ! invariant_p (XEXP (comparison, 0))
6354 || invariant_p (XEXP (comparison, 1)))
6355 return comparison;
6356
6357 return gen_rtx (swap_condition (GET_CODE (comparison)), VOIDmode,
6358 XEXP (comparison, 1), XEXP (comparison, 0));
6359 }
This page took 0.337315 seconds and 5 git commands to generate.