]> gcc.gnu.org Git - gcc.git/blob - gcc/loop.c
formatting tweaks
[gcc.git] / gcc / loop.c
1 /* Perform various loop optimizations, including strength reduction.
2 Copyright (C) 1987, 88, 89, 91-4, 1995 Free Software Foundation, Inc.
3
4 This file is part of GNU CC.
5
6 GNU CC is free software; you can redistribute it and/or modify
7 it under the terms of the GNU General Public License as published by
8 the Free Software Foundation; either version 2, or (at your option)
9 any later version.
10
11 GNU CC is distributed in the hope that it will be useful,
12 but WITHOUT ANY WARRANTY; without even the implied warranty of
13 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 GNU General Public License for more details.
15
16 You should have received a copy of the GNU General Public License
17 along with GNU CC; see the file COPYING. If not, write to
18 the Free Software Foundation, 59 Temple Place - Suite 330,
19 Boston, MA 02111-1307, USA. */
20
21
22 /* This is the loop optimization pass of the compiler.
23 It finds invariant computations within loops and moves them
24 to the beginning of the loop. Then it identifies basic and
25 general induction variables. Strength reduction is applied to the general
26 induction variables, and induction variable elimination is applied to
27 the basic induction variables.
28
29 It also finds cases where
30 a register is set within the loop by zero-extending a narrower value
31 and changes these to zero the entire register once before the loop
32 and merely copy the low part within the loop.
33
34 Most of the complexity is in heuristics to decide when it is worth
35 while to do these things. */
36
37 #include <stdio.h>
38 #include "config.h"
39 #include "rtl.h"
40 #include "obstack.h"
41 #include "expr.h"
42 #include "insn-config.h"
43 #include "insn-flags.h"
44 #include "regs.h"
45 #include "hard-reg-set.h"
46 #include "recog.h"
47 #include "flags.h"
48 #include "real.h"
49 #include "loop.h"
50
51 /* Vector mapping INSN_UIDs to luids.
52 The luids are like uids but increase monotonically always.
53 We use them to see whether a jump comes from outside a given loop. */
54
55 int *uid_luid;
56
57 /* Indexed by INSN_UID, contains the ordinal giving the (innermost) loop
58 number the insn is contained in. */
59
60 int *uid_loop_num;
61
62 /* 1 + largest uid of any insn. */
63
64 int max_uid_for_loop;
65
66 /* 1 + luid of last insn. */
67
68 static int max_luid;
69
70 /* Number of loops detected in current function. Used as index to the
71 next few tables. */
72
73 static int max_loop_num;
74
75 /* Indexed by loop number, contains the first and last insn of each loop. */
76
77 static rtx *loop_number_loop_starts, *loop_number_loop_ends;
78
79 /* For each loop, gives the containing loop number, -1 if none. */
80
81 int *loop_outer_loop;
82
83 /* Indexed by loop number, contains a nonzero value if the "loop" isn't
84 really a loop (an insn outside the loop branches into it). */
85
86 static char *loop_invalid;
87
88 /* Indexed by loop number, links together all LABEL_REFs which refer to
89 code labels outside the loop. Used by routines that need to know all
90 loop exits, such as final_biv_value and final_giv_value.
91
92 This does not include loop exits due to return instructions. This is
93 because all bivs and givs are pseudos, and hence must be dead after a
94 return, so the presense of a return does not affect any of the
95 optimizations that use this info. It is simpler to just not include return
96 instructions on this list. */
97
98 rtx *loop_number_exit_labels;
99
100 /* Indexed by loop number, counts the number of LABEL_REFs on
101 loop_number_exit_labels for this loop and all loops nested inside it. */
102
103 int *loop_number_exit_count;
104
105 /* Holds the number of loop iterations. It is zero if the number could not be
106 calculated. Must be unsigned since the number of iterations can
107 be as high as 2^wordsize-1. For loops with a wider iterator, this number
108 will will be zero if the number of loop iterations is too large for an
109 unsigned integer to hold. */
110
111 unsigned HOST_WIDE_INT loop_n_iterations;
112
113 /* Nonzero if there is a subroutine call in the current loop.
114 (unknown_address_altered is also nonzero in this case.) */
115
116 static int loop_has_call;
117
118 /* Nonzero if there is a volatile memory reference in the current
119 loop. */
120
121 static int loop_has_volatile;
122
123 /* Added loop_continue which is the NOTE_INSN_LOOP_CONT of the
124 current loop. A continue statement will generate a branch to
125 NEXT_INSN (loop_continue). */
126
127 static rtx loop_continue;
128
129 /* Indexed by register number, contains the number of times the reg
130 is set during the loop being scanned.
131 During code motion, a negative value indicates a reg that has been
132 made a candidate; in particular -2 means that it is an candidate that
133 we know is equal to a constant and -1 means that it is an candidate
134 not known equal to a constant.
135 After code motion, regs moved have 0 (which is accurate now)
136 while the failed candidates have the original number of times set.
137
138 Therefore, at all times, == 0 indicates an invariant register;
139 < 0 a conditionally invariant one. */
140
141 static short *n_times_set;
142
143 /* Original value of n_times_set; same except that this value
144 is not set negative for a reg whose sets have been made candidates
145 and not set to 0 for a reg that is moved. */
146
147 static short *n_times_used;
148
149 /* Index by register number, 1 indicates that the register
150 cannot be moved or strength reduced. */
151
152 static char *may_not_optimize;
153
154 /* Nonzero means reg N has already been moved out of one loop.
155 This reduces the desire to move it out of another. */
156
157 static char *moved_once;
158
159 /* Array of MEMs that are stored in this loop. If there are too many to fit
160 here, we just turn on unknown_address_altered. */
161
162 #define NUM_STORES 20
163 static rtx loop_store_mems[NUM_STORES];
164
165 /* Index of first available slot in above array. */
166 static int loop_store_mems_idx;
167
168 /* Nonzero if we don't know what MEMs were changed in the current loop.
169 This happens if the loop contains a call (in which case `loop_has_call'
170 will also be set) or if we store into more than NUM_STORES MEMs. */
171
172 static int unknown_address_altered;
173
174 /* Count of movable (i.e. invariant) instructions discovered in the loop. */
175 static int num_movables;
176
177 /* Count of memory write instructions discovered in the loop. */
178 static int num_mem_sets;
179
180 /* Number of loops contained within the current one, including itself. */
181 static int loops_enclosed;
182
183 /* Bound on pseudo register number before loop optimization.
184 A pseudo has valid regscan info if its number is < max_reg_before_loop. */
185 int max_reg_before_loop;
186
187 /* This obstack is used in product_cheap_p to allocate its rtl. It
188 may call gen_reg_rtx which, in turn, may reallocate regno_reg_rtx.
189 If we used the same obstack that it did, we would be deallocating
190 that array. */
191
192 static struct obstack temp_obstack;
193
194 /* This is where the pointer to the obstack being used for RTL is stored. */
195
196 extern struct obstack *rtl_obstack;
197
198 #define obstack_chunk_alloc xmalloc
199 #define obstack_chunk_free free
200
201 extern char *oballoc ();
202 \f
203 /* During the analysis of a loop, a chain of `struct movable's
204 is made to record all the movable insns found.
205 Then the entire chain can be scanned to decide which to move. */
206
207 struct movable
208 {
209 rtx insn; /* A movable insn */
210 rtx set_src; /* The expression this reg is set from. */
211 rtx set_dest; /* The destination of this SET. */
212 rtx dependencies; /* When INSN is libcall, this is an EXPR_LIST
213 of any registers used within the LIBCALL. */
214 int consec; /* Number of consecutive following insns
215 that must be moved with this one. */
216 int regno; /* The register it sets */
217 short lifetime; /* lifetime of that register;
218 may be adjusted when matching movables
219 that load the same value are found. */
220 short savings; /* Number of insns we can move for this reg,
221 including other movables that force this
222 or match this one. */
223 unsigned int cond : 1; /* 1 if only conditionally movable */
224 unsigned int force : 1; /* 1 means MUST move this insn */
225 unsigned int global : 1; /* 1 means reg is live outside this loop */
226 /* If PARTIAL is 1, GLOBAL means something different:
227 that the reg is live outside the range from where it is set
228 to the following label. */
229 unsigned int done : 1; /* 1 inhibits further processing of this */
230
231 unsigned int partial : 1; /* 1 means this reg is used for zero-extending.
232 In particular, moving it does not make it
233 invariant. */
234 unsigned int move_insn : 1; /* 1 means that we call emit_move_insn to
235 load SRC, rather than copying INSN. */
236 unsigned int is_equiv : 1; /* 1 means a REG_EQUIV is present on INSN. */
237 enum machine_mode savemode; /* Nonzero means it is a mode for a low part
238 that we should avoid changing when clearing
239 the rest of the reg. */
240 struct movable *match; /* First entry for same value */
241 struct movable *forces; /* An insn that must be moved if this is */
242 struct movable *next;
243 };
244
245 FILE *loop_dump_stream;
246
247 /* Forward declarations. */
248
249 static void find_and_verify_loops ();
250 static void mark_loop_jump ();
251 static void prescan_loop ();
252 static int reg_in_basic_block_p ();
253 static int consec_sets_invariant_p ();
254 static rtx libcall_other_reg ();
255 static int labels_in_range_p ();
256 static void count_loop_regs_set ();
257 static void note_addr_stored ();
258 static int loop_reg_used_before_p ();
259 static void scan_loop ();
260 static void replace_call_address ();
261 static rtx skip_consec_insns ();
262 static int libcall_benefit ();
263 static void ignore_some_movables ();
264 static void force_movables ();
265 static void combine_movables ();
266 static int rtx_equal_for_loop_p ();
267 static void move_movables ();
268 static void strength_reduce ();
269 static int valid_initial_value_p ();
270 static void find_mem_givs ();
271 static void record_biv ();
272 static void check_final_value ();
273 static void record_giv ();
274 static void update_giv_derive ();
275 static int basic_induction_var ();
276 static rtx simplify_giv_expr ();
277 static int general_induction_var ();
278 static int consec_sets_giv ();
279 static int check_dbra_loop ();
280 static rtx express_from ();
281 static int combine_givs_p ();
282 static void combine_givs ();
283 static int product_cheap_p ();
284 static int maybe_eliminate_biv ();
285 static int maybe_eliminate_biv_1 ();
286 static int last_use_this_basic_block ();
287 static void record_initial ();
288 static void update_reg_last_use ();
289 \f
290 /* Relative gain of eliminating various kinds of operations. */
291 int add_cost;
292 #if 0
293 int shift_cost;
294 int mult_cost;
295 #endif
296
297 /* Benefit penalty, if a giv is not replaceable, i.e. must emit an insn to
298 copy the value of the strength reduced giv to its original register. */
299 int copy_cost;
300
301 void
302 init_loop ()
303 {
304 char *free_point = (char *) oballoc (1);
305 rtx reg = gen_rtx (REG, word_mode, LAST_VIRTUAL_REGISTER + 1);
306
307 add_cost = rtx_cost (gen_rtx (PLUS, word_mode, reg, reg), SET);
308
309 /* We multiply by 2 to reconcile the difference in scale between
310 these two ways of computing costs. Otherwise the cost of a copy
311 will be far less than the cost of an add. */
312
313 copy_cost = 2 * 2;
314
315 /* Free the objects we just allocated. */
316 obfree (free_point);
317
318 /* Initialize the obstack used for rtl in product_cheap_p. */
319 gcc_obstack_init (&temp_obstack);
320 }
321 \f
322 /* Entry point of this file. Perform loop optimization
323 on the current function. F is the first insn of the function
324 and DUMPFILE is a stream for output of a trace of actions taken
325 (or 0 if none should be output). */
326
327 void
328 loop_optimize (f, dumpfile)
329 /* f is the first instruction of a chain of insns for one function */
330 rtx f;
331 FILE *dumpfile;
332 {
333 register rtx insn;
334 register int i;
335 rtx last_insn;
336
337 loop_dump_stream = dumpfile;
338
339 init_recog_no_volatile ();
340 init_alias_analysis ();
341
342 max_reg_before_loop = max_reg_num ();
343
344 moved_once = (char *) alloca (max_reg_before_loop);
345 bzero (moved_once, max_reg_before_loop);
346
347 regs_may_share = 0;
348
349 /* Count the number of loops. */
350
351 max_loop_num = 0;
352 for (insn = f; insn; insn = NEXT_INSN (insn))
353 {
354 if (GET_CODE (insn) == NOTE
355 && NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG)
356 max_loop_num++;
357 }
358
359 /* Don't waste time if no loops. */
360 if (max_loop_num == 0)
361 return;
362
363 /* Get size to use for tables indexed by uids.
364 Leave some space for labels allocated by find_and_verify_loops. */
365 max_uid_for_loop = get_max_uid () + 1 + max_loop_num * 32;
366
367 uid_luid = (int *) alloca (max_uid_for_loop * sizeof (int));
368 uid_loop_num = (int *) alloca (max_uid_for_loop * sizeof (int));
369
370 bzero ((char *) uid_luid, max_uid_for_loop * sizeof (int));
371 bzero ((char *) uid_loop_num, max_uid_for_loop * sizeof (int));
372
373 /* Allocate tables for recording each loop. We set each entry, so they need
374 not be zeroed. */
375 loop_number_loop_starts = (rtx *) alloca (max_loop_num * sizeof (rtx));
376 loop_number_loop_ends = (rtx *) alloca (max_loop_num * sizeof (rtx));
377 loop_outer_loop = (int *) alloca (max_loop_num * sizeof (int));
378 loop_invalid = (char *) alloca (max_loop_num * sizeof (char));
379 loop_number_exit_labels = (rtx *) alloca (max_loop_num * sizeof (rtx));
380 loop_number_exit_count = (int *) alloca (max_loop_num * sizeof (int));
381
382 /* Find and process each loop.
383 First, find them, and record them in order of their beginnings. */
384 find_and_verify_loops (f);
385
386 /* Now find all register lifetimes. This must be done after
387 find_and_verify_loops, because it might reorder the insns in the
388 function. */
389 reg_scan (f, max_reg_num (), 1);
390
391 /* See if we went too far. */
392 if (get_max_uid () > max_uid_for_loop)
393 abort ();
394
395 /* Compute the mapping from uids to luids.
396 LUIDs are numbers assigned to insns, like uids,
397 except that luids increase monotonically through the code.
398 Don't assign luids to line-number NOTEs, so that the distance in luids
399 between two insns is not affected by -g. */
400
401 for (insn = f, i = 0; insn; insn = NEXT_INSN (insn))
402 {
403 last_insn = insn;
404 if (GET_CODE (insn) != NOTE
405 || NOTE_LINE_NUMBER (insn) <= 0)
406 uid_luid[INSN_UID (insn)] = ++i;
407 else
408 /* Give a line number note the same luid as preceding insn. */
409 uid_luid[INSN_UID (insn)] = i;
410 }
411
412 max_luid = i + 1;
413
414 /* Don't leave gaps in uid_luid for insns that have been
415 deleted. It is possible that the first or last insn
416 using some register has been deleted by cross-jumping.
417 Make sure that uid_luid for that former insn's uid
418 points to the general area where that insn used to be. */
419 for (i = 0; i < max_uid_for_loop; i++)
420 {
421 uid_luid[0] = uid_luid[i];
422 if (uid_luid[0] != 0)
423 break;
424 }
425 for (i = 0; i < max_uid_for_loop; i++)
426 if (uid_luid[i] == 0)
427 uid_luid[i] = uid_luid[i - 1];
428
429 /* Create a mapping from loops to BLOCK tree nodes. */
430 if (flag_unroll_loops && write_symbols != NO_DEBUG)
431 find_loop_tree_blocks ();
432
433 /* Now scan the loops, last ones first, since this means inner ones are done
434 before outer ones. */
435 for (i = max_loop_num-1; i >= 0; i--)
436 if (! loop_invalid[i] && loop_number_loop_ends[i])
437 scan_loop (loop_number_loop_starts[i], loop_number_loop_ends[i],
438 max_reg_num ());
439
440 /* If debugging and unrolling loops, we must replicate the tree nodes
441 corresponding to the blocks inside the loop, so that the original one
442 to one mapping will remain. */
443 if (flag_unroll_loops && write_symbols != NO_DEBUG)
444 unroll_block_trees ();
445 }
446 \f
447 /* Optimize one loop whose start is LOOP_START and end is END.
448 LOOP_START is the NOTE_INSN_LOOP_BEG and END is the matching
449 NOTE_INSN_LOOP_END. */
450
451 /* ??? Could also move memory writes out of loops if the destination address
452 is invariant, the source is invariant, the memory write is not volatile,
453 and if we can prove that no read inside the loop can read this address
454 before the write occurs. If there is a read of this address after the
455 write, then we can also mark the memory read as invariant. */
456
457 static void
458 scan_loop (loop_start, end, nregs)
459 rtx loop_start, end;
460 int nregs;
461 {
462 register int i;
463 register rtx p;
464 /* 1 if we are scanning insns that could be executed zero times. */
465 int maybe_never = 0;
466 /* 1 if we are scanning insns that might never be executed
467 due to a subroutine call which might exit before they are reached. */
468 int call_passed = 0;
469 /* For a rotated loop that is entered near the bottom,
470 this is the label at the top. Otherwise it is zero. */
471 rtx loop_top = 0;
472 /* Jump insn that enters the loop, or 0 if control drops in. */
473 rtx loop_entry_jump = 0;
474 /* Place in the loop where control enters. */
475 rtx scan_start;
476 /* Number of insns in the loop. */
477 int insn_count;
478 int in_libcall = 0;
479 int tem;
480 rtx temp;
481 /* The SET from an insn, if it is the only SET in the insn. */
482 rtx set, set1;
483 /* Chain describing insns movable in current loop. */
484 struct movable *movables = 0;
485 /* Last element in `movables' -- so we can add elements at the end. */
486 struct movable *last_movable = 0;
487 /* Ratio of extra register life span we can justify
488 for saving an instruction. More if loop doesn't call subroutines
489 since in that case saving an insn makes more difference
490 and more registers are available. */
491 int threshold;
492 /* If we have calls, contains the insn in which a register was used
493 if it was used exactly once; contains const0_rtx if it was used more
494 than once. */
495 rtx *reg_single_usage = 0;
496 /* Nonzero if we are scanning instructions in a sub-loop. */
497 int loop_depth = 0;
498
499 n_times_set = (short *) alloca (nregs * sizeof (short));
500 n_times_used = (short *) alloca (nregs * sizeof (short));
501 may_not_optimize = (char *) alloca (nregs);
502
503 /* Determine whether this loop starts with a jump down to a test at
504 the end. This will occur for a small number of loops with a test
505 that is too complex to duplicate in front of the loop.
506
507 We search for the first insn or label in the loop, skipping NOTEs.
508 However, we must be careful not to skip past a NOTE_INSN_LOOP_BEG
509 (because we might have a loop executed only once that contains a
510 loop which starts with a jump to its exit test) or a NOTE_INSN_LOOP_END
511 (in case we have a degenerate loop).
512
513 Note that if we mistakenly think that a loop is entered at the top
514 when, in fact, it is entered at the exit test, the only effect will be
515 slightly poorer optimization. Making the opposite error can generate
516 incorrect code. Since very few loops now start with a jump to the
517 exit test, the code here to detect that case is very conservative. */
518
519 for (p = NEXT_INSN (loop_start);
520 p != end
521 && GET_CODE (p) != CODE_LABEL && GET_RTX_CLASS (GET_CODE (p)) != 'i'
522 && (GET_CODE (p) != NOTE
523 || (NOTE_LINE_NUMBER (p) != NOTE_INSN_LOOP_BEG
524 && NOTE_LINE_NUMBER (p) != NOTE_INSN_LOOP_END));
525 p = NEXT_INSN (p))
526 ;
527
528 scan_start = p;
529
530 /* Set up variables describing this loop. */
531 prescan_loop (loop_start, end);
532 threshold = (loop_has_call ? 1 : 2) * (1 + n_non_fixed_regs);
533
534 /* If loop has a jump before the first label,
535 the true entry is the target of that jump.
536 Start scan from there.
537 But record in LOOP_TOP the place where the end-test jumps
538 back to so we can scan that after the end of the loop. */
539 if (GET_CODE (p) == JUMP_INSN)
540 {
541 loop_entry_jump = p;
542
543 /* Loop entry must be unconditional jump (and not a RETURN) */
544 if (simplejump_p (p)
545 && JUMP_LABEL (p) != 0
546 /* Check to see whether the jump actually
547 jumps out of the loop (meaning it's no loop).
548 This case can happen for things like
549 do {..} while (0). If this label was generated previously
550 by loop, we can't tell anything about it and have to reject
551 the loop. */
552 && INSN_UID (JUMP_LABEL (p)) < max_uid_for_loop
553 && INSN_LUID (JUMP_LABEL (p)) >= INSN_LUID (loop_start)
554 && INSN_LUID (JUMP_LABEL (p)) < INSN_LUID (end))
555 {
556 loop_top = next_label (scan_start);
557 scan_start = JUMP_LABEL (p);
558 }
559 }
560
561 /* If SCAN_START was an insn created by loop, we don't know its luid
562 as required by loop_reg_used_before_p. So skip such loops. (This
563 test may never be true, but it's best to play it safe.)
564
565 Also, skip loops where we do not start scanning at a label. This
566 test also rejects loops starting with a JUMP_INSN that failed the
567 test above. */
568
569 if (INSN_UID (scan_start) >= max_uid_for_loop
570 || GET_CODE (scan_start) != CODE_LABEL)
571 {
572 if (loop_dump_stream)
573 fprintf (loop_dump_stream, "\nLoop from %d to %d is phony.\n\n",
574 INSN_UID (loop_start), INSN_UID (end));
575 return;
576 }
577
578 /* Count number of times each reg is set during this loop.
579 Set may_not_optimize[I] if it is not safe to move out
580 the setting of register I. If this loop has calls, set
581 reg_single_usage[I]. */
582
583 bzero ((char *) n_times_set, nregs * sizeof (short));
584 bzero (may_not_optimize, nregs);
585
586 if (loop_has_call)
587 {
588 reg_single_usage = (rtx *) alloca (nregs * sizeof (rtx));
589 bzero ((char *) reg_single_usage, nregs * sizeof (rtx));
590 }
591
592 count_loop_regs_set (loop_top ? loop_top : loop_start, end,
593 may_not_optimize, reg_single_usage, &insn_count, nregs);
594
595 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
596 may_not_optimize[i] = 1, n_times_set[i] = 1;
597 bcopy ((char *) n_times_set, (char *) n_times_used, nregs * sizeof (short));
598
599 if (loop_dump_stream)
600 {
601 fprintf (loop_dump_stream, "\nLoop from %d to %d: %d real insns.\n",
602 INSN_UID (loop_start), INSN_UID (end), insn_count);
603 if (loop_continue)
604 fprintf (loop_dump_stream, "Continue at insn %d.\n",
605 INSN_UID (loop_continue));
606 }
607
608 /* Scan through the loop finding insns that are safe to move.
609 Set n_times_set negative for the reg being set, so that
610 this reg will be considered invariant for subsequent insns.
611 We consider whether subsequent insns use the reg
612 in deciding whether it is worth actually moving.
613
614 MAYBE_NEVER is nonzero if we have passed a conditional jump insn
615 and therefore it is possible that the insns we are scanning
616 would never be executed. At such times, we must make sure
617 that it is safe to execute the insn once instead of zero times.
618 When MAYBE_NEVER is 0, all insns will be executed at least once
619 so that is not a problem. */
620
621 p = scan_start;
622 while (1)
623 {
624 p = NEXT_INSN (p);
625 /* At end of a straight-in loop, we are done.
626 At end of a loop entered at the bottom, scan the top. */
627 if (p == scan_start)
628 break;
629 if (p == end)
630 {
631 if (loop_top != 0)
632 p = loop_top;
633 else
634 break;
635 if (p == scan_start)
636 break;
637 }
638
639 if (GET_RTX_CLASS (GET_CODE (p)) == 'i'
640 && find_reg_note (p, REG_LIBCALL, NULL_RTX))
641 in_libcall = 1;
642 else if (GET_RTX_CLASS (GET_CODE (p)) == 'i'
643 && find_reg_note (p, REG_RETVAL, NULL_RTX))
644 in_libcall = 0;
645
646 if (GET_CODE (p) == INSN
647 && (set = single_set (p))
648 && GET_CODE (SET_DEST (set)) == REG
649 && ! may_not_optimize[REGNO (SET_DEST (set))])
650 {
651 int tem1 = 0;
652 int tem2 = 0;
653 int move_insn = 0;
654 rtx src = SET_SRC (set);
655 rtx dependencies = 0;
656
657 /* Figure out what to use as a source of this insn. If a REG_EQUIV
658 note is given or if a REG_EQUAL note with a constant operand is
659 specified, use it as the source and mark that we should move
660 this insn by calling emit_move_insn rather that duplicating the
661 insn.
662
663 Otherwise, only use the REG_EQUAL contents if a REG_RETVAL note
664 is present. */
665 temp = find_reg_note (p, REG_EQUIV, NULL_RTX);
666 if (temp)
667 src = XEXP (temp, 0), move_insn = 1;
668 else
669 {
670 temp = find_reg_note (p, REG_EQUAL, NULL_RTX);
671 if (temp && CONSTANT_P (XEXP (temp, 0)))
672 src = XEXP (temp, 0), move_insn = 1;
673 if (temp && find_reg_note (p, REG_RETVAL, NULL_RTX))
674 {
675 src = XEXP (temp, 0);
676 /* A libcall block can use regs that don't appear in
677 the equivalent expression. To move the libcall,
678 we must move those regs too. */
679 dependencies = libcall_other_reg (p, src);
680 }
681 }
682
683 /* Don't try to optimize a register that was made
684 by loop-optimization for an inner loop.
685 We don't know its life-span, so we can't compute the benefit. */
686 if (REGNO (SET_DEST (set)) >= max_reg_before_loop)
687 ;
688 /* In order to move a register, we need to have one of three cases:
689 (1) it is used only in the same basic block as the set
690 (2) it is not a user variable and it is not used in the
691 exit test (this can cause the variable to be used
692 before it is set just like a user-variable).
693 (3) the set is guaranteed to be executed once the loop starts,
694 and the reg is not used until after that. */
695 else if (! ((! maybe_never
696 && ! loop_reg_used_before_p (set, p, loop_start,
697 scan_start, end))
698 || (! REG_USERVAR_P (SET_DEST (set))
699 && ! REG_LOOP_TEST_P (SET_DEST (set)))
700 || reg_in_basic_block_p (p, SET_DEST (set))))
701 ;
702 else if ((tem = invariant_p (src))
703 && (dependencies == 0
704 || (tem2 = invariant_p (dependencies)) != 0)
705 && (n_times_set[REGNO (SET_DEST (set))] == 1
706 || (tem1
707 = consec_sets_invariant_p (SET_DEST (set),
708 n_times_set[REGNO (SET_DEST (set))],
709 p)))
710 /* If the insn can cause a trap (such as divide by zero),
711 can't move it unless it's guaranteed to be executed
712 once loop is entered. Even a function call might
713 prevent the trap insn from being reached
714 (since it might exit!) */
715 && ! ((maybe_never || call_passed)
716 && may_trap_p (src)))
717 {
718 register struct movable *m;
719 register int regno = REGNO (SET_DEST (set));
720
721 /* A potential lossage is where we have a case where two insns
722 can be combined as long as they are both in the loop, but
723 we move one of them outside the loop. For large loops,
724 this can lose. The most common case of this is the address
725 of a function being called.
726
727 Therefore, if this register is marked as being used exactly
728 once if we are in a loop with calls (a "large loop"), see if
729 we can replace the usage of this register with the source
730 of this SET. If we can, delete this insn.
731
732 Don't do this if P has a REG_RETVAL note or if we have
733 SMALL_REGISTER_CLASSES and SET_SRC is a hard register. */
734
735 if (reg_single_usage && reg_single_usage[regno] != 0
736 && reg_single_usage[regno] != const0_rtx
737 && regno_first_uid[regno] == INSN_UID (p)
738 && (regno_last_uid[regno]
739 == INSN_UID (reg_single_usage[regno]))
740 && n_times_set[REGNO (SET_DEST (set))] == 1
741 && ! side_effects_p (SET_SRC (set))
742 && ! find_reg_note (p, REG_RETVAL, NULL_RTX)
743 #ifdef SMALL_REGISTER_CLASSES
744 && ! (GET_CODE (SET_SRC (set)) == REG
745 && REGNO (SET_SRC (set)) < FIRST_PSEUDO_REGISTER)
746 #endif
747 /* This test is not redundant; SET_SRC (set) might be
748 a call-clobbered register and the life of REGNO
749 might span a call. */
750 && ! modified_between_p (SET_SRC (set), p,
751 reg_single_usage[regno])
752 && no_labels_between_p (p, reg_single_usage[regno])
753 && validate_replace_rtx (SET_DEST (set), SET_SRC (set),
754 reg_single_usage[regno]))
755 {
756 /* Replace any usage in a REG_EQUAL note. Must copy the
757 new source, so that we don't get rtx sharing between the
758 SET_SOURCE and REG_NOTES of insn p. */
759 REG_NOTES (reg_single_usage[regno])
760 = replace_rtx (REG_NOTES (reg_single_usage[regno]),
761 SET_DEST (set), copy_rtx (SET_SRC (set)));
762
763 PUT_CODE (p, NOTE);
764 NOTE_LINE_NUMBER (p) = NOTE_INSN_DELETED;
765 NOTE_SOURCE_FILE (p) = 0;
766 n_times_set[regno] = 0;
767 continue;
768 }
769
770 m = (struct movable *) alloca (sizeof (struct movable));
771 m->next = 0;
772 m->insn = p;
773 m->set_src = src;
774 m->dependencies = dependencies;
775 m->set_dest = SET_DEST (set);
776 m->force = 0;
777 m->consec = n_times_set[REGNO (SET_DEST (set))] - 1;
778 m->done = 0;
779 m->forces = 0;
780 m->partial = 0;
781 m->move_insn = move_insn;
782 m->is_equiv = (find_reg_note (p, REG_EQUIV, NULL_RTX) != 0);
783 m->savemode = VOIDmode;
784 m->regno = regno;
785 /* Set M->cond if either invariant_p or consec_sets_invariant_p
786 returned 2 (only conditionally invariant). */
787 m->cond = ((tem | tem1 | tem2) > 1);
788 m->global = (uid_luid[regno_last_uid[regno]] > INSN_LUID (end)
789 || uid_luid[regno_first_uid[regno]] < INSN_LUID (loop_start));
790 m->match = 0;
791 m->lifetime = (uid_luid[regno_last_uid[regno]]
792 - uid_luid[regno_first_uid[regno]]);
793 m->savings = n_times_used[regno];
794 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
795 m->savings += libcall_benefit (p);
796 n_times_set[regno] = move_insn ? -2 : -1;
797 /* Add M to the end of the chain MOVABLES. */
798 if (movables == 0)
799 movables = m;
800 else
801 last_movable->next = m;
802 last_movable = m;
803
804 if (m->consec > 0)
805 {
806 /* Skip this insn, not checking REG_LIBCALL notes. */
807 p = next_nonnote_insn (p);
808 /* Skip the consecutive insns, if there are any. */
809 p = skip_consec_insns (p, m->consec);
810 /* Back up to the last insn of the consecutive group. */
811 p = prev_nonnote_insn (p);
812
813 /* We must now reset m->move_insn, m->is_equiv, and possibly
814 m->set_src to correspond to the effects of all the
815 insns. */
816 temp = find_reg_note (p, REG_EQUIV, NULL_RTX);
817 if (temp)
818 m->set_src = XEXP (temp, 0), m->move_insn = 1;
819 else
820 {
821 temp = find_reg_note (p, REG_EQUAL, NULL_RTX);
822 if (temp && CONSTANT_P (XEXP (temp, 0)))
823 m->set_src = XEXP (temp, 0), m->move_insn = 1;
824 else
825 m->move_insn = 0;
826
827 }
828 m->is_equiv = (find_reg_note (p, REG_EQUIV, NULL_RTX) != 0);
829 }
830 }
831 /* If this register is always set within a STRICT_LOW_PART
832 or set to zero, then its high bytes are constant.
833 So clear them outside the loop and within the loop
834 just load the low bytes.
835 We must check that the machine has an instruction to do so.
836 Also, if the value loaded into the register
837 depends on the same register, this cannot be done. */
838 else if (SET_SRC (set) == const0_rtx
839 && GET_CODE (NEXT_INSN (p)) == INSN
840 && (set1 = single_set (NEXT_INSN (p)))
841 && GET_CODE (set1) == SET
842 && (GET_CODE (SET_DEST (set1)) == STRICT_LOW_PART)
843 && (GET_CODE (XEXP (SET_DEST (set1), 0)) == SUBREG)
844 && (SUBREG_REG (XEXP (SET_DEST (set1), 0))
845 == SET_DEST (set))
846 && !reg_mentioned_p (SET_DEST (set), SET_SRC (set1)))
847 {
848 register int regno = REGNO (SET_DEST (set));
849 if (n_times_set[regno] == 2)
850 {
851 register struct movable *m;
852 m = (struct movable *) alloca (sizeof (struct movable));
853 m->next = 0;
854 m->insn = p;
855 m->set_dest = SET_DEST (set);
856 m->dependencies = 0;
857 m->force = 0;
858 m->consec = 0;
859 m->done = 0;
860 m->forces = 0;
861 m->move_insn = 0;
862 m->partial = 1;
863 /* If the insn may not be executed on some cycles,
864 we can't clear the whole reg; clear just high part.
865 Not even if the reg is used only within this loop.
866 Consider this:
867 while (1)
868 while (s != t) {
869 if (foo ()) x = *s;
870 use (x);
871 }
872 Clearing x before the inner loop could clobber a value
873 being saved from the last time around the outer loop.
874 However, if the reg is not used outside this loop
875 and all uses of the register are in the same
876 basic block as the store, there is no problem.
877
878 If this insn was made by loop, we don't know its
879 INSN_LUID and hence must make a conservative
880 assumption. */
881 m->global = (INSN_UID (p) >= max_uid_for_loop
882 || (uid_luid[regno_last_uid[regno]]
883 > INSN_LUID (end))
884 || (uid_luid[regno_first_uid[regno]]
885 < INSN_LUID (p))
886 || (labels_in_range_p
887 (p, uid_luid[regno_first_uid[regno]])));
888 if (maybe_never && m->global)
889 m->savemode = GET_MODE (SET_SRC (set1));
890 else
891 m->savemode = VOIDmode;
892 m->regno = regno;
893 m->cond = 0;
894 m->match = 0;
895 m->lifetime = (uid_luid[regno_last_uid[regno]]
896 - uid_luid[regno_first_uid[regno]]);
897 m->savings = 1;
898 n_times_set[regno] = -1;
899 /* Add M to the end of the chain MOVABLES. */
900 if (movables == 0)
901 movables = m;
902 else
903 last_movable->next = m;
904 last_movable = m;
905 }
906 }
907 }
908 /* Past a call insn, we get to insns which might not be executed
909 because the call might exit. This matters for insns that trap.
910 Call insns inside a REG_LIBCALL/REG_RETVAL block always return,
911 so they don't count. */
912 else if (GET_CODE (p) == CALL_INSN && ! in_libcall)
913 call_passed = 1;
914 /* Past a label or a jump, we get to insns for which we
915 can't count on whether or how many times they will be
916 executed during each iteration. Therefore, we can
917 only move out sets of trivial variables
918 (those not used after the loop). */
919 /* Similar code appears twice in strength_reduce. */
920 else if ((GET_CODE (p) == CODE_LABEL || GET_CODE (p) == JUMP_INSN)
921 /* If we enter the loop in the middle, and scan around to the
922 beginning, don't set maybe_never for that. This must be an
923 unconditional jump, otherwise the code at the top of the
924 loop might never be executed. Unconditional jumps are
925 followed a by barrier then loop end. */
926 && ! (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) == loop_top
927 && NEXT_INSN (NEXT_INSN (p)) == end
928 && simplejump_p (p)))
929 maybe_never = 1;
930 else if (GET_CODE (p) == NOTE)
931 {
932 /* At the virtual top of a converted loop, insns are again known to
933 be executed: logically, the loop begins here even though the exit
934 code has been duplicated. */
935 if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP && loop_depth == 0)
936 maybe_never = call_passed = 0;
937 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
938 loop_depth++;
939 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END)
940 loop_depth--;
941 }
942 }
943
944 /* If one movable subsumes another, ignore that other. */
945
946 ignore_some_movables (movables);
947
948 /* For each movable insn, see if the reg that it loads
949 leads when it dies right into another conditionally movable insn.
950 If so, record that the second insn "forces" the first one,
951 since the second can be moved only if the first is. */
952
953 force_movables (movables);
954
955 /* See if there are multiple movable insns that load the same value.
956 If there are, make all but the first point at the first one
957 through the `match' field, and add the priorities of them
958 all together as the priority of the first. */
959
960 combine_movables (movables, nregs);
961
962 /* Now consider each movable insn to decide whether it is worth moving.
963 Store 0 in n_times_set for each reg that is moved. */
964
965 move_movables (movables, threshold,
966 insn_count, loop_start, end, nregs);
967
968 /* Now candidates that still are negative are those not moved.
969 Change n_times_set to indicate that those are not actually invariant. */
970 for (i = 0; i < nregs; i++)
971 if (n_times_set[i] < 0)
972 n_times_set[i] = n_times_used[i];
973
974 if (flag_strength_reduce)
975 strength_reduce (scan_start, end, loop_top,
976 insn_count, loop_start, end);
977 }
978 \f
979 /* Add elements to *OUTPUT to record all the pseudo-regs
980 mentioned in IN_THIS but not mentioned in NOT_IN_THIS. */
981
982 void
983 record_excess_regs (in_this, not_in_this, output)
984 rtx in_this, not_in_this;
985 rtx *output;
986 {
987 enum rtx_code code;
988 char *fmt;
989 int i;
990
991 code = GET_CODE (in_this);
992
993 switch (code)
994 {
995 case PC:
996 case CC0:
997 case CONST_INT:
998 case CONST_DOUBLE:
999 case CONST:
1000 case SYMBOL_REF:
1001 case LABEL_REF:
1002 return;
1003
1004 case REG:
1005 if (REGNO (in_this) >= FIRST_PSEUDO_REGISTER
1006 && ! reg_mentioned_p (in_this, not_in_this))
1007 *output = gen_rtx (EXPR_LIST, VOIDmode, in_this, *output);
1008 return;
1009 }
1010
1011 fmt = GET_RTX_FORMAT (code);
1012 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1013 {
1014 int j;
1015
1016 switch (fmt[i])
1017 {
1018 case 'E':
1019 for (j = 0; j < XVECLEN (in_this, i); j++)
1020 record_excess_regs (XVECEXP (in_this, i, j), not_in_this, output);
1021 break;
1022
1023 case 'e':
1024 record_excess_regs (XEXP (in_this, i), not_in_this, output);
1025 break;
1026 }
1027 }
1028 }
1029 \f
1030 /* Check what regs are referred to in the libcall block ending with INSN,
1031 aside from those mentioned in the equivalent value.
1032 If there are none, return 0.
1033 If there are one or more, return an EXPR_LIST containing all of them. */
1034
1035 static rtx
1036 libcall_other_reg (insn, equiv)
1037 rtx insn, equiv;
1038 {
1039 rtx note = find_reg_note (insn, REG_RETVAL, NULL_RTX);
1040 rtx p = XEXP (note, 0);
1041 rtx output = 0;
1042
1043 /* First, find all the regs used in the libcall block
1044 that are not mentioned as inputs to the result. */
1045
1046 while (p != insn)
1047 {
1048 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
1049 || GET_CODE (p) == CALL_INSN)
1050 record_excess_regs (PATTERN (p), equiv, &output);
1051 p = NEXT_INSN (p);
1052 }
1053
1054 return output;
1055 }
1056 \f
1057 /* Return 1 if all uses of REG
1058 are between INSN and the end of the basic block. */
1059
1060 static int
1061 reg_in_basic_block_p (insn, reg)
1062 rtx insn, reg;
1063 {
1064 int regno = REGNO (reg);
1065 rtx p;
1066
1067 if (regno_first_uid[regno] != INSN_UID (insn))
1068 return 0;
1069
1070 /* Search this basic block for the already recorded last use of the reg. */
1071 for (p = insn; p; p = NEXT_INSN (p))
1072 {
1073 switch (GET_CODE (p))
1074 {
1075 case NOTE:
1076 break;
1077
1078 case INSN:
1079 case CALL_INSN:
1080 /* Ordinary insn: if this is the last use, we win. */
1081 if (regno_last_uid[regno] == INSN_UID (p))
1082 return 1;
1083 break;
1084
1085 case JUMP_INSN:
1086 /* Jump insn: if this is the last use, we win. */
1087 if (regno_last_uid[regno] == INSN_UID (p))
1088 return 1;
1089 /* Otherwise, it's the end of the basic block, so we lose. */
1090 return 0;
1091
1092 case CODE_LABEL:
1093 case BARRIER:
1094 /* It's the end of the basic block, so we lose. */
1095 return 0;
1096 }
1097 }
1098
1099 /* The "last use" doesn't follow the "first use"?? */
1100 abort ();
1101 }
1102 \f
1103 /* Compute the benefit of eliminating the insns in the block whose
1104 last insn is LAST. This may be a group of insns used to compute a
1105 value directly or can contain a library call. */
1106
1107 static int
1108 libcall_benefit (last)
1109 rtx last;
1110 {
1111 rtx insn;
1112 int benefit = 0;
1113
1114 for (insn = XEXP (find_reg_note (last, REG_RETVAL, NULL_RTX), 0);
1115 insn != last; insn = NEXT_INSN (insn))
1116 {
1117 if (GET_CODE (insn) == CALL_INSN)
1118 benefit += 10; /* Assume at least this many insns in a library
1119 routine. */
1120 else if (GET_CODE (insn) == INSN
1121 && GET_CODE (PATTERN (insn)) != USE
1122 && GET_CODE (PATTERN (insn)) != CLOBBER)
1123 benefit++;
1124 }
1125
1126 return benefit;
1127 }
1128 \f
1129 /* Skip COUNT insns from INSN, counting library calls as 1 insn. */
1130
1131 static rtx
1132 skip_consec_insns (insn, count)
1133 rtx insn;
1134 int count;
1135 {
1136 for (; count > 0; count--)
1137 {
1138 rtx temp;
1139
1140 /* If first insn of libcall sequence, skip to end. */
1141 /* Do this at start of loop, since INSN is guaranteed to
1142 be an insn here. */
1143 if (GET_CODE (insn) != NOTE
1144 && (temp = find_reg_note (insn, REG_LIBCALL, NULL_RTX)))
1145 insn = XEXP (temp, 0);
1146
1147 do insn = NEXT_INSN (insn);
1148 while (GET_CODE (insn) == NOTE);
1149 }
1150
1151 return insn;
1152 }
1153
1154 /* Ignore any movable whose insn falls within a libcall
1155 which is part of another movable.
1156 We make use of the fact that the movable for the libcall value
1157 was made later and so appears later on the chain. */
1158
1159 static void
1160 ignore_some_movables (movables)
1161 struct movable *movables;
1162 {
1163 register struct movable *m, *m1;
1164
1165 for (m = movables; m; m = m->next)
1166 {
1167 /* Is this a movable for the value of a libcall? */
1168 rtx note = find_reg_note (m->insn, REG_RETVAL, NULL_RTX);
1169 if (note)
1170 {
1171 rtx insn;
1172 /* Check for earlier movables inside that range,
1173 and mark them invalid. We cannot use LUIDs here because
1174 insns created by loop.c for prior loops don't have LUIDs.
1175 Rather than reject all such insns from movables, we just
1176 explicitly check each insn in the libcall (since invariant
1177 libcalls aren't that common). */
1178 for (insn = XEXP (note, 0); insn != m->insn; insn = NEXT_INSN (insn))
1179 for (m1 = movables; m1 != m; m1 = m1->next)
1180 if (m1->insn == insn)
1181 m1->done = 1;
1182 }
1183 }
1184 }
1185
1186 /* For each movable insn, see if the reg that it loads
1187 leads when it dies right into another conditionally movable insn.
1188 If so, record that the second insn "forces" the first one,
1189 since the second can be moved only if the first is. */
1190
1191 static void
1192 force_movables (movables)
1193 struct movable *movables;
1194 {
1195 register struct movable *m, *m1;
1196 for (m1 = movables; m1; m1 = m1->next)
1197 /* Omit this if moving just the (SET (REG) 0) of a zero-extend. */
1198 if (!m1->partial && !m1->done)
1199 {
1200 int regno = m1->regno;
1201 for (m = m1->next; m; m = m->next)
1202 /* ??? Could this be a bug? What if CSE caused the
1203 register of M1 to be used after this insn?
1204 Since CSE does not update regno_last_uid,
1205 this insn M->insn might not be where it dies.
1206 But very likely this doesn't matter; what matters is
1207 that M's reg is computed from M1's reg. */
1208 if (INSN_UID (m->insn) == regno_last_uid[regno]
1209 && !m->done)
1210 break;
1211 if (m != 0 && m->set_src == m1->set_dest
1212 /* If m->consec, m->set_src isn't valid. */
1213 && m->consec == 0)
1214 m = 0;
1215
1216 /* Increase the priority of the moving the first insn
1217 since it permits the second to be moved as well. */
1218 if (m != 0)
1219 {
1220 m->forces = m1;
1221 m1->lifetime += m->lifetime;
1222 m1->savings += m1->savings;
1223 }
1224 }
1225 }
1226 \f
1227 /* Find invariant expressions that are equal and can be combined into
1228 one register. */
1229
1230 static void
1231 combine_movables (movables, nregs)
1232 struct movable *movables;
1233 int nregs;
1234 {
1235 register struct movable *m;
1236 char *matched_regs = (char *) alloca (nregs);
1237 enum machine_mode mode;
1238
1239 /* Regs that are set more than once are not allowed to match
1240 or be matched. I'm no longer sure why not. */
1241 /* Perhaps testing m->consec_sets would be more appropriate here? */
1242
1243 for (m = movables; m; m = m->next)
1244 if (m->match == 0 && n_times_used[m->regno] == 1 && !m->partial)
1245 {
1246 register struct movable *m1;
1247 int regno = m->regno;
1248
1249 bzero (matched_regs, nregs);
1250 matched_regs[regno] = 1;
1251
1252 for (m1 = movables; m1; m1 = m1->next)
1253 if (m != m1 && m1->match == 0 && n_times_used[m1->regno] == 1
1254 /* A reg used outside the loop mustn't be eliminated. */
1255 && !m1->global
1256 /* A reg used for zero-extending mustn't be eliminated. */
1257 && !m1->partial
1258 && (matched_regs[m1->regno]
1259 ||
1260 (
1261 /* Can combine regs with different modes loaded from the
1262 same constant only if the modes are the same or
1263 if both are integer modes with M wider or the same
1264 width as M1. The check for integer is redundant, but
1265 safe, since the only case of differing destination
1266 modes with equal sources is when both sources are
1267 VOIDmode, i.e., CONST_INT. */
1268 (GET_MODE (m->set_dest) == GET_MODE (m1->set_dest)
1269 || (GET_MODE_CLASS (GET_MODE (m->set_dest)) == MODE_INT
1270 && GET_MODE_CLASS (GET_MODE (m1->set_dest)) == MODE_INT
1271 && (GET_MODE_BITSIZE (GET_MODE (m->set_dest))
1272 >= GET_MODE_BITSIZE (GET_MODE (m1->set_dest)))))
1273 /* See if the source of M1 says it matches M. */
1274 && ((GET_CODE (m1->set_src) == REG
1275 && matched_regs[REGNO (m1->set_src)])
1276 || rtx_equal_for_loop_p (m->set_src, m1->set_src,
1277 movables))))
1278 && ((m->dependencies == m1->dependencies)
1279 || rtx_equal_p (m->dependencies, m1->dependencies)))
1280 {
1281 m->lifetime += m1->lifetime;
1282 m->savings += m1->savings;
1283 m1->done = 1;
1284 m1->match = m;
1285 matched_regs[m1->regno] = 1;
1286 }
1287 }
1288
1289 /* Now combine the regs used for zero-extension.
1290 This can be done for those not marked `global'
1291 provided their lives don't overlap. */
1292
1293 for (mode = GET_CLASS_NARROWEST_MODE (MODE_INT); mode != VOIDmode;
1294 mode = GET_MODE_WIDER_MODE (mode))
1295 {
1296 register struct movable *m0 = 0;
1297
1298 /* Combine all the registers for extension from mode MODE.
1299 Don't combine any that are used outside this loop. */
1300 for (m = movables; m; m = m->next)
1301 if (m->partial && ! m->global
1302 && mode == GET_MODE (SET_SRC (PATTERN (NEXT_INSN (m->insn)))))
1303 {
1304 register struct movable *m1;
1305 int first = uid_luid[regno_first_uid[m->regno]];
1306 int last = uid_luid[regno_last_uid[m->regno]];
1307
1308 if (m0 == 0)
1309 {
1310 /* First one: don't check for overlap, just record it. */
1311 m0 = m;
1312 continue;
1313 }
1314
1315 /* Make sure they extend to the same mode.
1316 (Almost always true.) */
1317 if (GET_MODE (m->set_dest) != GET_MODE (m0->set_dest))
1318 continue;
1319
1320 /* We already have one: check for overlap with those
1321 already combined together. */
1322 for (m1 = movables; m1 != m; m1 = m1->next)
1323 if (m1 == m0 || (m1->partial && m1->match == m0))
1324 if (! (uid_luid[regno_first_uid[m1->regno]] > last
1325 || uid_luid[regno_last_uid[m1->regno]] < first))
1326 goto overlap;
1327
1328 /* No overlap: we can combine this with the others. */
1329 m0->lifetime += m->lifetime;
1330 m0->savings += m->savings;
1331 m->done = 1;
1332 m->match = m0;
1333
1334 overlap: ;
1335 }
1336 }
1337 }
1338 \f
1339 /* Return 1 if regs X and Y will become the same if moved. */
1340
1341 static int
1342 regs_match_p (x, y, movables)
1343 rtx x, y;
1344 struct movable *movables;
1345 {
1346 int xn = REGNO (x);
1347 int yn = REGNO (y);
1348 struct movable *mx, *my;
1349
1350 for (mx = movables; mx; mx = mx->next)
1351 if (mx->regno == xn)
1352 break;
1353
1354 for (my = movables; my; my = my->next)
1355 if (my->regno == yn)
1356 break;
1357
1358 return (mx && my
1359 && ((mx->match == my->match && mx->match != 0)
1360 || mx->match == my
1361 || mx == my->match));
1362 }
1363
1364 /* Return 1 if X and Y are identical-looking rtx's.
1365 This is the Lisp function EQUAL for rtx arguments.
1366
1367 If two registers are matching movables or a movable register and an
1368 equivalent constant, consider them equal. */
1369
1370 static int
1371 rtx_equal_for_loop_p (x, y, movables)
1372 rtx x, y;
1373 struct movable *movables;
1374 {
1375 register int i;
1376 register int j;
1377 register struct movable *m;
1378 register enum rtx_code code;
1379 register char *fmt;
1380
1381 if (x == y)
1382 return 1;
1383 if (x == 0 || y == 0)
1384 return 0;
1385
1386 code = GET_CODE (x);
1387
1388 /* If we have a register and a constant, they may sometimes be
1389 equal. */
1390 if (GET_CODE (x) == REG && n_times_set[REGNO (x)] == -2
1391 && CONSTANT_P (y))
1392 for (m = movables; m; m = m->next)
1393 if (m->move_insn && m->regno == REGNO (x)
1394 && rtx_equal_p (m->set_src, y))
1395 return 1;
1396
1397 else if (GET_CODE (y) == REG && n_times_set[REGNO (y)] == -2
1398 && CONSTANT_P (x))
1399 for (m = movables; m; m = m->next)
1400 if (m->move_insn && m->regno == REGNO (y)
1401 && rtx_equal_p (m->set_src, x))
1402 return 1;
1403
1404 /* Otherwise, rtx's of different codes cannot be equal. */
1405 if (code != GET_CODE (y))
1406 return 0;
1407
1408 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent.
1409 (REG:SI x) and (REG:HI x) are NOT equivalent. */
1410
1411 if (GET_MODE (x) != GET_MODE (y))
1412 return 0;
1413
1414 /* These three types of rtx's can be compared nonrecursively. */
1415 if (code == REG)
1416 return (REGNO (x) == REGNO (y) || regs_match_p (x, y, movables));
1417
1418 if (code == LABEL_REF)
1419 return XEXP (x, 0) == XEXP (y, 0);
1420 if (code == SYMBOL_REF)
1421 return XSTR (x, 0) == XSTR (y, 0);
1422
1423 /* Compare the elements. If any pair of corresponding elements
1424 fail to match, return 0 for the whole things. */
1425
1426 fmt = GET_RTX_FORMAT (code);
1427 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1428 {
1429 switch (fmt[i])
1430 {
1431 case 'w':
1432 if (XWINT (x, i) != XWINT (y, i))
1433 return 0;
1434 break;
1435
1436 case 'i':
1437 if (XINT (x, i) != XINT (y, i))
1438 return 0;
1439 break;
1440
1441 case 'E':
1442 /* Two vectors must have the same length. */
1443 if (XVECLEN (x, i) != XVECLEN (y, i))
1444 return 0;
1445
1446 /* And the corresponding elements must match. */
1447 for (j = 0; j < XVECLEN (x, i); j++)
1448 if (rtx_equal_for_loop_p (XVECEXP (x, i, j), XVECEXP (y, i, j), movables) == 0)
1449 return 0;
1450 break;
1451
1452 case 'e':
1453 if (rtx_equal_for_loop_p (XEXP (x, i), XEXP (y, i), movables) == 0)
1454 return 0;
1455 break;
1456
1457 case 's':
1458 if (strcmp (XSTR (x, i), XSTR (y, i)))
1459 return 0;
1460 break;
1461
1462 case 'u':
1463 /* These are just backpointers, so they don't matter. */
1464 break;
1465
1466 case '0':
1467 break;
1468
1469 /* It is believed that rtx's at this level will never
1470 contain anything but integers and other rtx's,
1471 except for within LABEL_REFs and SYMBOL_REFs. */
1472 default:
1473 abort ();
1474 }
1475 }
1476 return 1;
1477 }
1478 \f
1479 /* If X contains any LABEL_REF's, add REG_LABEL notes for them to all
1480 insns in INSNS which use thet reference. */
1481
1482 static void
1483 add_label_notes (x, insns)
1484 rtx x;
1485 rtx insns;
1486 {
1487 enum rtx_code code = GET_CODE (x);
1488 int i, j;
1489 char *fmt;
1490 rtx insn;
1491
1492 if (code == LABEL_REF && !LABEL_REF_NONLOCAL_P (x))
1493 {
1494 rtx next = next_real_insn (XEXP (x, 0));
1495
1496 /* Don't record labels that refer to dispatch tables.
1497 This is not necessary, since the tablejump references the same label.
1498 And if we did record them, flow.c would make worse code. */
1499 if (next == 0
1500 || ! (GET_CODE (next) == JUMP_INSN
1501 && (GET_CODE (PATTERN (next)) == ADDR_VEC
1502 || GET_CODE (PATTERN (next)) == ADDR_DIFF_VEC)))
1503 {
1504 for (insn = insns; insn; insn = NEXT_INSN (insn))
1505 if (reg_mentioned_p (XEXP (x, 0), insn))
1506 REG_NOTES (insn) = gen_rtx (EXPR_LIST, REG_LABEL, XEXP (x, 0),
1507 REG_NOTES (insn));
1508 }
1509 return;
1510 }
1511
1512 fmt = GET_RTX_FORMAT (code);
1513 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1514 {
1515 if (fmt[i] == 'e')
1516 add_label_notes (XEXP (x, i), insns);
1517 else if (fmt[i] == 'E')
1518 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
1519 add_label_notes (XVECEXP (x, i, j), insns);
1520 }
1521 }
1522 \f
1523 /* Scan MOVABLES, and move the insns that deserve to be moved.
1524 If two matching movables are combined, replace one reg with the
1525 other throughout. */
1526
1527 static void
1528 move_movables (movables, threshold, insn_count, loop_start, end, nregs)
1529 struct movable *movables;
1530 int threshold;
1531 int insn_count;
1532 rtx loop_start;
1533 rtx end;
1534 int nregs;
1535 {
1536 rtx new_start = 0;
1537 register struct movable *m;
1538 register rtx p;
1539 /* Map of pseudo-register replacements to handle combining
1540 when we move several insns that load the same value
1541 into different pseudo-registers. */
1542 rtx *reg_map = (rtx *) alloca (nregs * sizeof (rtx));
1543 char *already_moved = (char *) alloca (nregs);
1544
1545 bzero (already_moved, nregs);
1546 bzero ((char *) reg_map, nregs * sizeof (rtx));
1547
1548 num_movables = 0;
1549
1550 for (m = movables; m; m = m->next)
1551 {
1552 /* Describe this movable insn. */
1553
1554 if (loop_dump_stream)
1555 {
1556 fprintf (loop_dump_stream, "Insn %d: regno %d (life %d), ",
1557 INSN_UID (m->insn), m->regno, m->lifetime);
1558 if (m->consec > 0)
1559 fprintf (loop_dump_stream, "consec %d, ", m->consec);
1560 if (m->cond)
1561 fprintf (loop_dump_stream, "cond ");
1562 if (m->force)
1563 fprintf (loop_dump_stream, "force ");
1564 if (m->global)
1565 fprintf (loop_dump_stream, "global ");
1566 if (m->done)
1567 fprintf (loop_dump_stream, "done ");
1568 if (m->move_insn)
1569 fprintf (loop_dump_stream, "move-insn ");
1570 if (m->match)
1571 fprintf (loop_dump_stream, "matches %d ",
1572 INSN_UID (m->match->insn));
1573 if (m->forces)
1574 fprintf (loop_dump_stream, "forces %d ",
1575 INSN_UID (m->forces->insn));
1576 }
1577
1578 /* Count movables. Value used in heuristics in strength_reduce. */
1579 num_movables++;
1580
1581 /* Ignore the insn if it's already done (it matched something else).
1582 Otherwise, see if it is now safe to move. */
1583
1584 if (!m->done
1585 && (! m->cond
1586 || (1 == invariant_p (m->set_src)
1587 && (m->dependencies == 0
1588 || 1 == invariant_p (m->dependencies))
1589 && (m->consec == 0
1590 || 1 == consec_sets_invariant_p (m->set_dest,
1591 m->consec + 1,
1592 m->insn))))
1593 && (! m->forces || m->forces->done))
1594 {
1595 register int regno;
1596 register rtx p;
1597 int savings = m->savings;
1598
1599 /* We have an insn that is safe to move.
1600 Compute its desirability. */
1601
1602 p = m->insn;
1603 regno = m->regno;
1604
1605 if (loop_dump_stream)
1606 fprintf (loop_dump_stream, "savings %d ", savings);
1607
1608 if (moved_once[regno])
1609 {
1610 insn_count *= 2;
1611
1612 if (loop_dump_stream)
1613 fprintf (loop_dump_stream, "halved since already moved ");
1614 }
1615
1616 /* An insn MUST be moved if we already moved something else
1617 which is safe only if this one is moved too: that is,
1618 if already_moved[REGNO] is nonzero. */
1619
1620 /* An insn is desirable to move if the new lifetime of the
1621 register is no more than THRESHOLD times the old lifetime.
1622 If it's not desirable, it means the loop is so big
1623 that moving won't speed things up much,
1624 and it is liable to make register usage worse. */
1625
1626 /* It is also desirable to move if it can be moved at no
1627 extra cost because something else was already moved. */
1628
1629 if (already_moved[regno]
1630 || (threshold * savings * m->lifetime) >= insn_count
1631 || (m->forces && m->forces->done
1632 && n_times_used[m->forces->regno] == 1))
1633 {
1634 int count;
1635 register struct movable *m1;
1636 rtx first;
1637
1638 /* Now move the insns that set the reg. */
1639
1640 if (m->partial && m->match)
1641 {
1642 rtx newpat, i1;
1643 rtx r1, r2;
1644 /* Find the end of this chain of matching regs.
1645 Thus, we load each reg in the chain from that one reg.
1646 And that reg is loaded with 0 directly,
1647 since it has ->match == 0. */
1648 for (m1 = m; m1->match; m1 = m1->match);
1649 newpat = gen_move_insn (SET_DEST (PATTERN (m->insn)),
1650 SET_DEST (PATTERN (m1->insn)));
1651 i1 = emit_insn_before (newpat, loop_start);
1652
1653 /* Mark the moved, invariant reg as being allowed to
1654 share a hard reg with the other matching invariant. */
1655 REG_NOTES (i1) = REG_NOTES (m->insn);
1656 r1 = SET_DEST (PATTERN (m->insn));
1657 r2 = SET_DEST (PATTERN (m1->insn));
1658 regs_may_share = gen_rtx (EXPR_LIST, VOIDmode, r1,
1659 gen_rtx (EXPR_LIST, VOIDmode, r2,
1660 regs_may_share));
1661 delete_insn (m->insn);
1662
1663 if (new_start == 0)
1664 new_start = i1;
1665
1666 if (loop_dump_stream)
1667 fprintf (loop_dump_stream, " moved to %d", INSN_UID (i1));
1668 }
1669 /* If we are to re-generate the item being moved with a
1670 new move insn, first delete what we have and then emit
1671 the move insn before the loop. */
1672 else if (m->move_insn)
1673 {
1674 rtx i1, temp;
1675
1676 for (count = m->consec; count >= 0; count--)
1677 {
1678 /* If this is the first insn of a library call sequence,
1679 skip to the end. */
1680 if (GET_CODE (p) != NOTE
1681 && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
1682 p = XEXP (temp, 0);
1683
1684 /* If this is the last insn of a libcall sequence, then
1685 delete every insn in the sequence except the last.
1686 The last insn is handled in the normal manner. */
1687 if (GET_CODE (p) != NOTE
1688 && (temp = find_reg_note (p, REG_RETVAL, NULL_RTX)))
1689 {
1690 temp = XEXP (temp, 0);
1691 while (temp != p)
1692 temp = delete_insn (temp);
1693 }
1694
1695 p = delete_insn (p);
1696 while (p && GET_CODE (p) == NOTE)
1697 p = NEXT_INSN (p);
1698 }
1699
1700 start_sequence ();
1701 emit_move_insn (m->set_dest, m->set_src);
1702 temp = get_insns ();
1703 end_sequence ();
1704
1705 add_label_notes (m->set_src, temp);
1706
1707 i1 = emit_insns_before (temp, loop_start);
1708 if (! find_reg_note (i1, REG_EQUAL, NULL_RTX))
1709 REG_NOTES (i1)
1710 = gen_rtx (EXPR_LIST,
1711 m->is_equiv ? REG_EQUIV : REG_EQUAL,
1712 m->set_src, REG_NOTES (i1));
1713
1714 if (loop_dump_stream)
1715 fprintf (loop_dump_stream, " moved to %d", INSN_UID (i1));
1716
1717 /* The more regs we move, the less we like moving them. */
1718 threshold -= 3;
1719 }
1720 else
1721 {
1722 for (count = m->consec; count >= 0; count--)
1723 {
1724 rtx i1, temp;
1725
1726 /* If first insn of libcall sequence, skip to end. */
1727 /* Do this at start of loop, since p is guaranteed to
1728 be an insn here. */
1729 if (GET_CODE (p) != NOTE
1730 && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
1731 p = XEXP (temp, 0);
1732
1733 /* If last insn of libcall sequence, move all
1734 insns except the last before the loop. The last
1735 insn is handled in the normal manner. */
1736 if (GET_CODE (p) != NOTE
1737 && (temp = find_reg_note (p, REG_RETVAL, NULL_RTX)))
1738 {
1739 rtx fn_address = 0;
1740 rtx fn_reg = 0;
1741 rtx fn_address_insn = 0;
1742
1743 first = 0;
1744 for (temp = XEXP (temp, 0); temp != p;
1745 temp = NEXT_INSN (temp))
1746 {
1747 rtx body;
1748 rtx n;
1749 rtx next;
1750
1751 if (GET_CODE (temp) == NOTE)
1752 continue;
1753
1754 body = PATTERN (temp);
1755
1756 /* Find the next insn after TEMP,
1757 not counting USE or NOTE insns. */
1758 for (next = NEXT_INSN (temp); next != p;
1759 next = NEXT_INSN (next))
1760 if (! (GET_CODE (next) == INSN
1761 && GET_CODE (PATTERN (next)) == USE)
1762 && GET_CODE (next) != NOTE)
1763 break;
1764
1765 /* If that is the call, this may be the insn
1766 that loads the function address.
1767
1768 Extract the function address from the insn
1769 that loads it into a register.
1770 If this insn was cse'd, we get incorrect code.
1771
1772 So emit a new move insn that copies the
1773 function address into the register that the
1774 call insn will use. flow.c will delete any
1775 redundant stores that we have created. */
1776 if (GET_CODE (next) == CALL_INSN
1777 && GET_CODE (body) == SET
1778 && GET_CODE (SET_DEST (body)) == REG
1779 && (n = find_reg_note (temp, REG_EQUAL,
1780 NULL_RTX)))
1781 {
1782 fn_reg = SET_SRC (body);
1783 if (GET_CODE (fn_reg) != REG)
1784 fn_reg = SET_DEST (body);
1785 fn_address = XEXP (n, 0);
1786 fn_address_insn = temp;
1787 }
1788 /* We have the call insn.
1789 If it uses the register we suspect it might,
1790 load it with the correct address directly. */
1791 if (GET_CODE (temp) == CALL_INSN
1792 && fn_address != 0
1793 && reg_referenced_p (fn_reg, body))
1794 emit_insn_after (gen_move_insn (fn_reg,
1795 fn_address),
1796 fn_address_insn);
1797
1798 if (GET_CODE (temp) == CALL_INSN)
1799 {
1800 i1 = emit_call_insn_before (body, loop_start);
1801 /* Because the USAGE information potentially
1802 contains objects other than hard registers
1803 we need to copy it. */
1804 if (CALL_INSN_FUNCTION_USAGE (temp))
1805 CALL_INSN_FUNCTION_USAGE (i1) =
1806 copy_rtx (CALL_INSN_FUNCTION_USAGE (temp));
1807 }
1808 else
1809 i1 = emit_insn_before (body, loop_start);
1810 if (first == 0)
1811 first = i1;
1812 if (temp == fn_address_insn)
1813 fn_address_insn = i1;
1814 REG_NOTES (i1) = REG_NOTES (temp);
1815 delete_insn (temp);
1816 }
1817 }
1818 if (m->savemode != VOIDmode)
1819 {
1820 /* P sets REG to zero; but we should clear only
1821 the bits that are not covered by the mode
1822 m->savemode. */
1823 rtx reg = m->set_dest;
1824 rtx sequence;
1825 rtx tem;
1826
1827 start_sequence ();
1828 tem = expand_binop
1829 (GET_MODE (reg), and_optab, reg,
1830 GEN_INT ((((HOST_WIDE_INT) 1
1831 << GET_MODE_BITSIZE (m->savemode)))
1832 - 1),
1833 reg, 1, OPTAB_LIB_WIDEN);
1834 if (tem == 0)
1835 abort ();
1836 if (tem != reg)
1837 emit_move_insn (reg, tem);
1838 sequence = gen_sequence ();
1839 end_sequence ();
1840 i1 = emit_insn_before (sequence, loop_start);
1841 }
1842 else if (GET_CODE (p) == CALL_INSN)
1843 {
1844 i1 = emit_call_insn_before (PATTERN (p), loop_start);
1845 /* Because the USAGE information potentially
1846 contains objects other than hard registers
1847 we need to copy it. */
1848 if (CALL_INSN_FUNCTION_USAGE (p))
1849 CALL_INSN_FUNCTION_USAGE (i1) =
1850 copy_rtx (CALL_INSN_FUNCTION_USAGE (p));
1851 }
1852 else
1853 i1 = emit_insn_before (PATTERN (p), loop_start);
1854
1855 REG_NOTES (i1) = REG_NOTES (p);
1856
1857 /* If there is a REG_EQUAL note present whose value is
1858 not loop invariant, then delete it, since it may
1859 cause problems with later optimization passes.
1860 It is possible for cse to create such notes
1861 like this as a result of record_jump_cond. */
1862
1863 if ((temp = find_reg_note (i1, REG_EQUAL, NULL_RTX))
1864 && ! invariant_p (XEXP (temp, 0)))
1865 remove_note (i1, temp);
1866
1867 if (new_start == 0)
1868 new_start = i1;
1869
1870 if (loop_dump_stream)
1871 fprintf (loop_dump_stream, " moved to %d",
1872 INSN_UID (i1));
1873
1874 #if 0
1875 /* This isn't needed because REG_NOTES is copied
1876 below and is wrong since P might be a PARALLEL. */
1877 if (REG_NOTES (i1) == 0
1878 && ! m->partial /* But not if it's a zero-extend clr. */
1879 && ! m->global /* and not if used outside the loop
1880 (since it might get set outside). */
1881 && CONSTANT_P (SET_SRC (PATTERN (p))))
1882 REG_NOTES (i1)
1883 = gen_rtx (EXPR_LIST, REG_EQUAL,
1884 SET_SRC (PATTERN (p)), REG_NOTES (i1));
1885 #endif
1886
1887 /* If library call, now fix the REG_NOTES that contain
1888 insn pointers, namely REG_LIBCALL on FIRST
1889 and REG_RETVAL on I1. */
1890 if (temp = find_reg_note (i1, REG_RETVAL, NULL_RTX))
1891 {
1892 XEXP (temp, 0) = first;
1893 temp = find_reg_note (first, REG_LIBCALL, NULL_RTX);
1894 XEXP (temp, 0) = i1;
1895 }
1896
1897 delete_insn (p);
1898 do p = NEXT_INSN (p);
1899 while (p && GET_CODE (p) == NOTE);
1900 }
1901
1902 /* The more regs we move, the less we like moving them. */
1903 threshold -= 3;
1904 }
1905
1906 /* Any other movable that loads the same register
1907 MUST be moved. */
1908 already_moved[regno] = 1;
1909
1910 /* This reg has been moved out of one loop. */
1911 moved_once[regno] = 1;
1912
1913 /* The reg set here is now invariant. */
1914 if (! m->partial)
1915 n_times_set[regno] = 0;
1916
1917 m->done = 1;
1918
1919 /* Change the length-of-life info for the register
1920 to say it lives at least the full length of this loop.
1921 This will help guide optimizations in outer loops. */
1922
1923 if (uid_luid[regno_first_uid[regno]] > INSN_LUID (loop_start))
1924 /* This is the old insn before all the moved insns.
1925 We can't use the moved insn because it is out of range
1926 in uid_luid. Only the old insns have luids. */
1927 regno_first_uid[regno] = INSN_UID (loop_start);
1928 if (uid_luid[regno_last_uid[regno]] < INSN_LUID (end))
1929 regno_last_uid[regno] = INSN_UID (end);
1930
1931 /* Combine with this moved insn any other matching movables. */
1932
1933 if (! m->partial)
1934 for (m1 = movables; m1; m1 = m1->next)
1935 if (m1->match == m)
1936 {
1937 rtx temp;
1938
1939 /* Schedule the reg loaded by M1
1940 for replacement so that shares the reg of M.
1941 If the modes differ (only possible in restricted
1942 circumstances, make a SUBREG. */
1943 if (GET_MODE (m->set_dest) == GET_MODE (m1->set_dest))
1944 reg_map[m1->regno] = m->set_dest;
1945 else
1946 reg_map[m1->regno]
1947 = gen_lowpart_common (GET_MODE (m1->set_dest),
1948 m->set_dest);
1949
1950 /* Get rid of the matching insn
1951 and prevent further processing of it. */
1952 m1->done = 1;
1953
1954 /* if library call, delete all insn except last, which
1955 is deleted below */
1956 if (temp = find_reg_note (m1->insn, REG_RETVAL,
1957 NULL_RTX))
1958 {
1959 for (temp = XEXP (temp, 0); temp != m1->insn;
1960 temp = NEXT_INSN (temp))
1961 delete_insn (temp);
1962 }
1963 delete_insn (m1->insn);
1964
1965 /* Any other movable that loads the same register
1966 MUST be moved. */
1967 already_moved[m1->regno] = 1;
1968
1969 /* The reg merged here is now invariant,
1970 if the reg it matches is invariant. */
1971 if (! m->partial)
1972 n_times_set[m1->regno] = 0;
1973 }
1974 }
1975 else if (loop_dump_stream)
1976 fprintf (loop_dump_stream, "not desirable");
1977 }
1978 else if (loop_dump_stream && !m->match)
1979 fprintf (loop_dump_stream, "not safe");
1980
1981 if (loop_dump_stream)
1982 fprintf (loop_dump_stream, "\n");
1983 }
1984
1985 if (new_start == 0)
1986 new_start = loop_start;
1987
1988 /* Go through all the instructions in the loop, making
1989 all the register substitutions scheduled in REG_MAP. */
1990 for (p = new_start; p != end; p = NEXT_INSN (p))
1991 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
1992 || GET_CODE (p) == CALL_INSN)
1993 {
1994 replace_regs (PATTERN (p), reg_map, nregs, 0);
1995 replace_regs (REG_NOTES (p), reg_map, nregs, 0);
1996 INSN_CODE (p) = -1;
1997 }
1998 }
1999 \f
2000 #if 0
2001 /* Scan X and replace the address of any MEM in it with ADDR.
2002 REG is the address that MEM should have before the replacement. */
2003
2004 static void
2005 replace_call_address (x, reg, addr)
2006 rtx x, reg, addr;
2007 {
2008 register enum rtx_code code;
2009 register int i;
2010 register char *fmt;
2011
2012 if (x == 0)
2013 return;
2014 code = GET_CODE (x);
2015 switch (code)
2016 {
2017 case PC:
2018 case CC0:
2019 case CONST_INT:
2020 case CONST_DOUBLE:
2021 case CONST:
2022 case SYMBOL_REF:
2023 case LABEL_REF:
2024 case REG:
2025 return;
2026
2027 case SET:
2028 /* Short cut for very common case. */
2029 replace_call_address (XEXP (x, 1), reg, addr);
2030 return;
2031
2032 case CALL:
2033 /* Short cut for very common case. */
2034 replace_call_address (XEXP (x, 0), reg, addr);
2035 return;
2036
2037 case MEM:
2038 /* If this MEM uses a reg other than the one we expected,
2039 something is wrong. */
2040 if (XEXP (x, 0) != reg)
2041 abort ();
2042 XEXP (x, 0) = addr;
2043 return;
2044 }
2045
2046 fmt = GET_RTX_FORMAT (code);
2047 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2048 {
2049 if (fmt[i] == 'e')
2050 replace_call_address (XEXP (x, i), reg, addr);
2051 if (fmt[i] == 'E')
2052 {
2053 register int j;
2054 for (j = 0; j < XVECLEN (x, i); j++)
2055 replace_call_address (XVECEXP (x, i, j), reg, addr);
2056 }
2057 }
2058 }
2059 #endif
2060 \f
2061 /* Return the number of memory refs to addresses that vary
2062 in the rtx X. */
2063
2064 static int
2065 count_nonfixed_reads (x)
2066 rtx x;
2067 {
2068 register enum rtx_code code;
2069 register int i;
2070 register char *fmt;
2071 int value;
2072
2073 if (x == 0)
2074 return 0;
2075
2076 code = GET_CODE (x);
2077 switch (code)
2078 {
2079 case PC:
2080 case CC0:
2081 case CONST_INT:
2082 case CONST_DOUBLE:
2083 case CONST:
2084 case SYMBOL_REF:
2085 case LABEL_REF:
2086 case REG:
2087 return 0;
2088
2089 case MEM:
2090 return ((invariant_p (XEXP (x, 0)) != 1)
2091 + count_nonfixed_reads (XEXP (x, 0)));
2092 }
2093
2094 value = 0;
2095 fmt = GET_RTX_FORMAT (code);
2096 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2097 {
2098 if (fmt[i] == 'e')
2099 value += count_nonfixed_reads (XEXP (x, i));
2100 if (fmt[i] == 'E')
2101 {
2102 register int j;
2103 for (j = 0; j < XVECLEN (x, i); j++)
2104 value += count_nonfixed_reads (XVECEXP (x, i, j));
2105 }
2106 }
2107 return value;
2108 }
2109
2110 \f
2111 #if 0
2112 /* P is an instruction that sets a register to the result of a ZERO_EXTEND.
2113 Replace it with an instruction to load just the low bytes
2114 if the machine supports such an instruction,
2115 and insert above LOOP_START an instruction to clear the register. */
2116
2117 static void
2118 constant_high_bytes (p, loop_start)
2119 rtx p, loop_start;
2120 {
2121 register rtx new;
2122 register int insn_code_number;
2123
2124 /* Try to change (SET (REG ...) (ZERO_EXTEND (..:B ...)))
2125 to (SET (STRICT_LOW_PART (SUBREG:B (REG...))) ...). */
2126
2127 new = gen_rtx (SET, VOIDmode,
2128 gen_rtx (STRICT_LOW_PART, VOIDmode,
2129 gen_rtx (SUBREG, GET_MODE (XEXP (SET_SRC (PATTERN (p)), 0)),
2130 SET_DEST (PATTERN (p)),
2131 0)),
2132 XEXP (SET_SRC (PATTERN (p)), 0));
2133 insn_code_number = recog (new, p);
2134
2135 if (insn_code_number)
2136 {
2137 register int i;
2138
2139 /* Clear destination register before the loop. */
2140 emit_insn_before (gen_rtx (SET, VOIDmode,
2141 SET_DEST (PATTERN (p)),
2142 const0_rtx),
2143 loop_start);
2144
2145 /* Inside the loop, just load the low part. */
2146 PATTERN (p) = new;
2147 }
2148 }
2149 #endif
2150 \f
2151 /* Scan a loop setting the variables `unknown_address_altered',
2152 `num_mem_sets', `loop_continue', loops_enclosed', `loop_has_call',
2153 and `loop_has_volatile'.
2154 Also, fill in the array `loop_store_mems'. */
2155
2156 static void
2157 prescan_loop (start, end)
2158 rtx start, end;
2159 {
2160 register int level = 1;
2161 register rtx insn;
2162
2163 unknown_address_altered = 0;
2164 loop_has_call = 0;
2165 loop_has_volatile = 0;
2166 loop_store_mems_idx = 0;
2167
2168 num_mem_sets = 0;
2169 loops_enclosed = 1;
2170 loop_continue = 0;
2171
2172 for (insn = NEXT_INSN (start); insn != NEXT_INSN (end);
2173 insn = NEXT_INSN (insn))
2174 {
2175 if (GET_CODE (insn) == NOTE)
2176 {
2177 if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG)
2178 {
2179 ++level;
2180 /* Count number of loops contained in this one. */
2181 loops_enclosed++;
2182 }
2183 else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_END)
2184 {
2185 --level;
2186 if (level == 0)
2187 {
2188 end = insn;
2189 break;
2190 }
2191 }
2192 else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_CONT)
2193 {
2194 if (level == 1)
2195 loop_continue = insn;
2196 }
2197 }
2198 else if (GET_CODE (insn) == CALL_INSN)
2199 {
2200 unknown_address_altered = 1;
2201 loop_has_call = 1;
2202 }
2203 else
2204 {
2205 if (GET_CODE (insn) == INSN || GET_CODE (insn) == JUMP_INSN)
2206 {
2207 if (volatile_refs_p (PATTERN (insn)))
2208 loop_has_volatile = 1;
2209
2210 note_stores (PATTERN (insn), note_addr_stored);
2211 }
2212 }
2213 }
2214 }
2215 \f
2216 /* Scan the function looking for loops. Record the start and end of each loop.
2217 Also mark as invalid loops any loops that contain a setjmp or are branched
2218 to from outside the loop. */
2219
2220 static void
2221 find_and_verify_loops (f)
2222 rtx f;
2223 {
2224 rtx insn, label;
2225 int current_loop = -1;
2226 int next_loop = -1;
2227 int loop;
2228
2229 /* If there are jumps to undefined labels,
2230 treat them as jumps out of any/all loops.
2231 This also avoids writing past end of tables when there are no loops. */
2232 uid_loop_num[0] = -1;
2233
2234 /* Find boundaries of loops, mark which loops are contained within
2235 loops, and invalidate loops that have setjmp. */
2236
2237 for (insn = f; insn; insn = NEXT_INSN (insn))
2238 {
2239 if (GET_CODE (insn) == NOTE)
2240 switch (NOTE_LINE_NUMBER (insn))
2241 {
2242 case NOTE_INSN_LOOP_BEG:
2243 loop_number_loop_starts[++next_loop] = insn;
2244 loop_number_loop_ends[next_loop] = 0;
2245 loop_outer_loop[next_loop] = current_loop;
2246 loop_invalid[next_loop] = 0;
2247 loop_number_exit_labels[next_loop] = 0;
2248 loop_number_exit_count[next_loop] = 0;
2249 current_loop = next_loop;
2250 break;
2251
2252 case NOTE_INSN_SETJMP:
2253 /* In this case, we must invalidate our current loop and any
2254 enclosing loop. */
2255 for (loop = current_loop; loop != -1; loop = loop_outer_loop[loop])
2256 {
2257 loop_invalid[loop] = 1;
2258 if (loop_dump_stream)
2259 fprintf (loop_dump_stream,
2260 "\nLoop at %d ignored due to setjmp.\n",
2261 INSN_UID (loop_number_loop_starts[loop]));
2262 }
2263 break;
2264
2265 case NOTE_INSN_LOOP_END:
2266 if (current_loop == -1)
2267 abort ();
2268
2269 loop_number_loop_ends[current_loop] = insn;
2270 current_loop = loop_outer_loop[current_loop];
2271 break;
2272
2273 }
2274
2275 /* Note that this will mark the NOTE_INSN_LOOP_END note as being in the
2276 enclosing loop, but this doesn't matter. */
2277 uid_loop_num[INSN_UID (insn)] = current_loop;
2278 }
2279
2280 /* Any loop containing a label used in an initializer must be invalidated,
2281 because it can be jumped into from anywhere. */
2282
2283 for (label = forced_labels; label; label = XEXP (label, 1))
2284 {
2285 int loop_num;
2286
2287 for (loop_num = uid_loop_num[INSN_UID (XEXP (label, 0))];
2288 loop_num != -1;
2289 loop_num = loop_outer_loop[loop_num])
2290 loop_invalid[loop_num] = 1;
2291 }
2292
2293 /* Now scan all insn's in the function. If any JUMP_INSN branches into a
2294 loop that it is not contained within, that loop is marked invalid.
2295 If any INSN or CALL_INSN uses a label's address, then the loop containing
2296 that label is marked invalid, because it could be jumped into from
2297 anywhere.
2298
2299 Also look for blocks of code ending in an unconditional branch that
2300 exits the loop. If such a block is surrounded by a conditional
2301 branch around the block, move the block elsewhere (see below) and
2302 invert the jump to point to the code block. This may eliminate a
2303 label in our loop and will simplify processing by both us and a
2304 possible second cse pass. */
2305
2306 for (insn = f; insn; insn = NEXT_INSN (insn))
2307 if (GET_RTX_CLASS (GET_CODE (insn)) == 'i')
2308 {
2309 int this_loop_num = uid_loop_num[INSN_UID (insn)];
2310
2311 if (GET_CODE (insn) == INSN || GET_CODE (insn) == CALL_INSN)
2312 {
2313 rtx note = find_reg_note (insn, REG_LABEL, NULL_RTX);
2314 if (note)
2315 {
2316 int loop_num;
2317
2318 for (loop_num = uid_loop_num[INSN_UID (XEXP (note, 0))];
2319 loop_num != -1;
2320 loop_num = loop_outer_loop[loop_num])
2321 loop_invalid[loop_num] = 1;
2322 }
2323 }
2324
2325 if (GET_CODE (insn) != JUMP_INSN)
2326 continue;
2327
2328 mark_loop_jump (PATTERN (insn), this_loop_num);
2329
2330 /* See if this is an unconditional branch outside the loop. */
2331 if (this_loop_num != -1
2332 && (GET_CODE (PATTERN (insn)) == RETURN
2333 || (simplejump_p (insn)
2334 && (uid_loop_num[INSN_UID (JUMP_LABEL (insn))]
2335 != this_loop_num)))
2336 && get_max_uid () < max_uid_for_loop)
2337 {
2338 rtx p;
2339 rtx our_next = next_real_insn (insn);
2340 int dest_loop;
2341 int outer_loop = -1;
2342
2343 /* Go backwards until we reach the start of the loop, a label,
2344 or a JUMP_INSN. */
2345 for (p = PREV_INSN (insn);
2346 GET_CODE (p) != CODE_LABEL
2347 && ! (GET_CODE (p) == NOTE
2348 && NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
2349 && GET_CODE (p) != JUMP_INSN;
2350 p = PREV_INSN (p))
2351 ;
2352
2353 /* Check for the case where we have a jump to an inner nested
2354 loop, and do not perform the optimization in that case. */
2355
2356 if (JUMP_LABEL (insn))
2357 {
2358 dest_loop = uid_loop_num[INSN_UID (JUMP_LABEL (insn))];
2359 if (dest_loop != -1)
2360 {
2361 for (outer_loop = dest_loop; outer_loop != -1;
2362 outer_loop = loop_outer_loop[outer_loop])
2363 if (outer_loop == this_loop_num)
2364 break;
2365 }
2366 }
2367
2368 /* Make sure that the target of P is within the current loop. */
2369
2370 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p)
2371 && uid_loop_num[INSN_UID (JUMP_LABEL (p))] != this_loop_num)
2372 outer_loop = this_loop_num;
2373
2374 /* If we stopped on a JUMP_INSN to the next insn after INSN,
2375 we have a block of code to try to move.
2376
2377 We look backward and then forward from the target of INSN
2378 to find a BARRIER at the same loop depth as the target.
2379 If we find such a BARRIER, we make a new label for the start
2380 of the block, invert the jump in P and point it to that label,
2381 and move the block of code to the spot we found. */
2382
2383 if (outer_loop == -1
2384 && GET_CODE (p) == JUMP_INSN
2385 && JUMP_LABEL (p) != 0
2386 /* Just ignore jumps to labels that were never emitted.
2387 These always indicate compilation errors. */
2388 && INSN_UID (JUMP_LABEL (p)) != 0
2389 && condjump_p (p)
2390 && ! simplejump_p (p)
2391 && next_real_insn (JUMP_LABEL (p)) == our_next)
2392 {
2393 rtx target
2394 = JUMP_LABEL (insn) ? JUMP_LABEL (insn) : get_last_insn ();
2395 int target_loop_num = uid_loop_num[INSN_UID (target)];
2396 rtx loc;
2397
2398 for (loc = target; loc; loc = PREV_INSN (loc))
2399 if (GET_CODE (loc) == BARRIER
2400 && uid_loop_num[INSN_UID (loc)] == target_loop_num)
2401 break;
2402
2403 if (loc == 0)
2404 for (loc = target; loc; loc = NEXT_INSN (loc))
2405 if (GET_CODE (loc) == BARRIER
2406 && uid_loop_num[INSN_UID (loc)] == target_loop_num)
2407 break;
2408
2409 if (loc)
2410 {
2411 rtx cond_label = JUMP_LABEL (p);
2412 rtx new_label = get_label_after (p);
2413
2414 /* Ensure our label doesn't go away. */
2415 LABEL_NUSES (cond_label)++;
2416
2417 /* Verify that uid_loop_num is large enough and that
2418 we can invert P. */
2419 if (invert_jump (p, new_label))
2420 {
2421 rtx q, r;
2422
2423 /* Include the BARRIER after INSN and copy the
2424 block after LOC. */
2425 new_label = squeeze_notes (new_label, NEXT_INSN (insn));
2426 reorder_insns (new_label, NEXT_INSN (insn), loc);
2427
2428 /* All those insns are now in TARGET_LOOP_NUM. */
2429 for (q = new_label; q != NEXT_INSN (NEXT_INSN (insn));
2430 q = NEXT_INSN (q))
2431 uid_loop_num[INSN_UID (q)] = target_loop_num;
2432
2433 /* The label jumped to by INSN is no longer a loop exit.
2434 Unless INSN does not have a label (e.g., it is a
2435 RETURN insn), search loop_number_exit_labels to find
2436 its label_ref, and remove it. Also turn off
2437 LABEL_OUTSIDE_LOOP_P bit. */
2438 if (JUMP_LABEL (insn))
2439 {
2440 int loop_num;
2441
2442 for (q = 0,
2443 r = loop_number_exit_labels[this_loop_num];
2444 r; q = r, r = LABEL_NEXTREF (r))
2445 if (XEXP (r, 0) == JUMP_LABEL (insn))
2446 {
2447 LABEL_OUTSIDE_LOOP_P (r) = 0;
2448 if (q)
2449 LABEL_NEXTREF (q) = LABEL_NEXTREF (r);
2450 else
2451 loop_number_exit_labels[this_loop_num]
2452 = LABEL_NEXTREF (r);
2453 break;
2454 }
2455
2456 for (loop_num = this_loop_num;
2457 loop_num != -1 && loop_num != target_loop_num;
2458 loop_num = loop_outer_loop[loop_num])
2459 loop_number_exit_count[loop_num]--;
2460
2461 /* If we didn't find it, then something is wrong. */
2462 if (! r)
2463 abort ();
2464 }
2465
2466 /* P is now a jump outside the loop, so it must be put
2467 in loop_number_exit_labels, and marked as such.
2468 The easiest way to do this is to just call
2469 mark_loop_jump again for P. */
2470 mark_loop_jump (PATTERN (p), this_loop_num);
2471
2472 /* If INSN now jumps to the insn after it,
2473 delete INSN. */
2474 if (JUMP_LABEL (insn) != 0
2475 && (next_real_insn (JUMP_LABEL (insn))
2476 == next_real_insn (insn)))
2477 delete_insn (insn);
2478 }
2479
2480 /* Continue the loop after where the conditional
2481 branch used to jump, since the only branch insn
2482 in the block (if it still remains) is an inter-loop
2483 branch and hence needs no processing. */
2484 insn = NEXT_INSN (cond_label);
2485
2486 if (--LABEL_NUSES (cond_label) == 0)
2487 delete_insn (cond_label);
2488
2489 /* This loop will be continued with NEXT_INSN (insn). */
2490 insn = PREV_INSN (insn);
2491 }
2492 }
2493 }
2494 }
2495 }
2496
2497 /* If any label in X jumps to a loop different from LOOP_NUM and any of the
2498 loops it is contained in, mark the target loop invalid.
2499
2500 For speed, we assume that X is part of a pattern of a JUMP_INSN. */
2501
2502 static void
2503 mark_loop_jump (x, loop_num)
2504 rtx x;
2505 int loop_num;
2506 {
2507 int dest_loop;
2508 int outer_loop;
2509 int i;
2510
2511 switch (GET_CODE (x))
2512 {
2513 case PC:
2514 case USE:
2515 case CLOBBER:
2516 case REG:
2517 case MEM:
2518 case CONST_INT:
2519 case CONST_DOUBLE:
2520 case RETURN:
2521 return;
2522
2523 case CONST:
2524 /* There could be a label reference in here. */
2525 mark_loop_jump (XEXP (x, 0), loop_num);
2526 return;
2527
2528 case PLUS:
2529 case MINUS:
2530 case MULT:
2531 mark_loop_jump (XEXP (x, 0), loop_num);
2532 mark_loop_jump (XEXP (x, 1), loop_num);
2533 return;
2534
2535 case SIGN_EXTEND:
2536 case ZERO_EXTEND:
2537 mark_loop_jump (XEXP (x, 0), loop_num);
2538 return;
2539
2540 case LABEL_REF:
2541 dest_loop = uid_loop_num[INSN_UID (XEXP (x, 0))];
2542
2543 /* Link together all labels that branch outside the loop. This
2544 is used by final_[bg]iv_value and the loop unrolling code. Also
2545 mark this LABEL_REF so we know that this branch should predict
2546 false. */
2547
2548 /* A check to make sure the label is not in an inner nested loop,
2549 since this does not count as a loop exit. */
2550 if (dest_loop != -1)
2551 {
2552 for (outer_loop = dest_loop; outer_loop != -1;
2553 outer_loop = loop_outer_loop[outer_loop])
2554 if (outer_loop == loop_num)
2555 break;
2556 }
2557 else
2558 outer_loop = -1;
2559
2560 if (loop_num != -1 && outer_loop == -1)
2561 {
2562 LABEL_OUTSIDE_LOOP_P (x) = 1;
2563 LABEL_NEXTREF (x) = loop_number_exit_labels[loop_num];
2564 loop_number_exit_labels[loop_num] = x;
2565
2566 for (outer_loop = loop_num;
2567 outer_loop != -1 && outer_loop != dest_loop;
2568 outer_loop = loop_outer_loop[outer_loop])
2569 loop_number_exit_count[outer_loop]++;
2570 }
2571
2572 /* If this is inside a loop, but not in the current loop or one enclosed
2573 by it, it invalidates at least one loop. */
2574
2575 if (dest_loop == -1)
2576 return;
2577
2578 /* We must invalidate every nested loop containing the target of this
2579 label, except those that also contain the jump insn. */
2580
2581 for (; dest_loop != -1; dest_loop = loop_outer_loop[dest_loop])
2582 {
2583 /* Stop when we reach a loop that also contains the jump insn. */
2584 for (outer_loop = loop_num; outer_loop != -1;
2585 outer_loop = loop_outer_loop[outer_loop])
2586 if (dest_loop == outer_loop)
2587 return;
2588
2589 /* If we get here, we know we need to invalidate a loop. */
2590 if (loop_dump_stream && ! loop_invalid[dest_loop])
2591 fprintf (loop_dump_stream,
2592 "\nLoop at %d ignored due to multiple entry points.\n",
2593 INSN_UID (loop_number_loop_starts[dest_loop]));
2594
2595 loop_invalid[dest_loop] = 1;
2596 }
2597 return;
2598
2599 case SET:
2600 /* If this is not setting pc, ignore. */
2601 if (SET_DEST (x) == pc_rtx)
2602 mark_loop_jump (SET_SRC (x), loop_num);
2603 return;
2604
2605 case IF_THEN_ELSE:
2606 mark_loop_jump (XEXP (x, 1), loop_num);
2607 mark_loop_jump (XEXP (x, 2), loop_num);
2608 return;
2609
2610 case PARALLEL:
2611 case ADDR_VEC:
2612 for (i = 0; i < XVECLEN (x, 0); i++)
2613 mark_loop_jump (XVECEXP (x, 0, i), loop_num);
2614 return;
2615
2616 case ADDR_DIFF_VEC:
2617 for (i = 0; i < XVECLEN (x, 1); i++)
2618 mark_loop_jump (XVECEXP (x, 1, i), loop_num);
2619 return;
2620
2621 default:
2622 /* Treat anything else (such as a symbol_ref)
2623 as a branch out of this loop, but not into any loop. */
2624
2625 if (loop_num != -1)
2626 {
2627 loop_number_exit_labels[loop_num] = x;
2628
2629 for (outer_loop = loop_num; outer_loop != -1;
2630 outer_loop = loop_outer_loop[outer_loop])
2631 loop_number_exit_count[outer_loop]++;
2632 }
2633 return;
2634 }
2635 }
2636 \f
2637 /* Return nonzero if there is a label in the range from
2638 insn INSN to and including the insn whose luid is END
2639 INSN must have an assigned luid (i.e., it must not have
2640 been previously created by loop.c). */
2641
2642 static int
2643 labels_in_range_p (insn, end)
2644 rtx insn;
2645 int end;
2646 {
2647 while (insn && INSN_LUID (insn) <= end)
2648 {
2649 if (GET_CODE (insn) == CODE_LABEL)
2650 return 1;
2651 insn = NEXT_INSN (insn);
2652 }
2653
2654 return 0;
2655 }
2656
2657 /* Record that a memory reference X is being set. */
2658
2659 static void
2660 note_addr_stored (x)
2661 rtx x;
2662 {
2663 register int i;
2664
2665 if (x == 0 || GET_CODE (x) != MEM)
2666 return;
2667
2668 /* Count number of memory writes.
2669 This affects heuristics in strength_reduce. */
2670 num_mem_sets++;
2671
2672 /* BLKmode MEM means all memory is clobbered. */
2673 if (GET_MODE (x) == BLKmode)
2674 unknown_address_altered = 1;
2675
2676 if (unknown_address_altered)
2677 return;
2678
2679 for (i = 0; i < loop_store_mems_idx; i++)
2680 if (rtx_equal_p (XEXP (loop_store_mems[i], 0), XEXP (x, 0))
2681 && MEM_IN_STRUCT_P (x) == MEM_IN_STRUCT_P (loop_store_mems[i]))
2682 {
2683 /* We are storing at the same address as previously noted. Save the
2684 wider reference. */
2685 if (GET_MODE_SIZE (GET_MODE (x))
2686 > GET_MODE_SIZE (GET_MODE (loop_store_mems[i])))
2687 loop_store_mems[i] = x;
2688 break;
2689 }
2690
2691 if (i == NUM_STORES)
2692 unknown_address_altered = 1;
2693
2694 else if (i == loop_store_mems_idx)
2695 loop_store_mems[loop_store_mems_idx++] = x;
2696 }
2697 \f
2698 /* Return nonzero if the rtx X is invariant over the current loop.
2699
2700 The value is 2 if we refer to something only conditionally invariant.
2701
2702 If `unknown_address_altered' is nonzero, no memory ref is invariant.
2703 Otherwise, a memory ref is invariant if it does not conflict with
2704 anything stored in `loop_store_mems'. */
2705
2706 int
2707 invariant_p (x)
2708 register rtx x;
2709 {
2710 register int i;
2711 register enum rtx_code code;
2712 register char *fmt;
2713 int conditional = 0;
2714
2715 if (x == 0)
2716 return 1;
2717 code = GET_CODE (x);
2718 switch (code)
2719 {
2720 case CONST_INT:
2721 case CONST_DOUBLE:
2722 case SYMBOL_REF:
2723 case CONST:
2724 return 1;
2725
2726 case LABEL_REF:
2727 /* A LABEL_REF is normally invariant, however, if we are unrolling
2728 loops, and this label is inside the loop, then it isn't invariant.
2729 This is because each unrolled copy of the loop body will have
2730 a copy of this label. If this was invariant, then an insn loading
2731 the address of this label into a register might get moved outside
2732 the loop, and then each loop body would end up using the same label.
2733
2734 We don't know the loop bounds here though, so just fail for all
2735 labels. */
2736 if (flag_unroll_loops)
2737 return 0;
2738 else
2739 return 1;
2740
2741 case PC:
2742 case CC0:
2743 case UNSPEC_VOLATILE:
2744 return 0;
2745
2746 case REG:
2747 /* We used to check RTX_UNCHANGING_P (x) here, but that is invalid
2748 since the reg might be set by initialization within the loop. */
2749 if (x == frame_pointer_rtx || x == hard_frame_pointer_rtx
2750 || x == arg_pointer_rtx)
2751 return 1;
2752 if (loop_has_call
2753 && REGNO (x) < FIRST_PSEUDO_REGISTER && call_used_regs[REGNO (x)])
2754 return 0;
2755 if (n_times_set[REGNO (x)] < 0)
2756 return 2;
2757 return n_times_set[REGNO (x)] == 0;
2758
2759 case MEM:
2760 /* Volatile memory references must be rejected. Do this before
2761 checking for read-only items, so that volatile read-only items
2762 will be rejected also. */
2763 if (MEM_VOLATILE_P (x))
2764 return 0;
2765
2766 /* Read-only items (such as constants in a constant pool) are
2767 invariant if their address is. */
2768 if (RTX_UNCHANGING_P (x))
2769 break;
2770
2771 /* If we filled the table (or had a subroutine call), any location
2772 in memory could have been clobbered. */
2773 if (unknown_address_altered)
2774 return 0;
2775
2776 /* See if there is any dependence between a store and this load. */
2777 for (i = loop_store_mems_idx - 1; i >= 0; i--)
2778 if (true_dependence (loop_store_mems[i], x))
2779 return 0;
2780
2781 /* It's not invalidated by a store in memory
2782 but we must still verify the address is invariant. */
2783 break;
2784
2785 case ASM_OPERANDS:
2786 /* Don't mess with insns declared volatile. */
2787 if (MEM_VOLATILE_P (x))
2788 return 0;
2789 }
2790
2791 fmt = GET_RTX_FORMAT (code);
2792 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2793 {
2794 if (fmt[i] == 'e')
2795 {
2796 int tem = invariant_p (XEXP (x, i));
2797 if (tem == 0)
2798 return 0;
2799 if (tem == 2)
2800 conditional = 1;
2801 }
2802 else if (fmt[i] == 'E')
2803 {
2804 register int j;
2805 for (j = 0; j < XVECLEN (x, i); j++)
2806 {
2807 int tem = invariant_p (XVECEXP (x, i, j));
2808 if (tem == 0)
2809 return 0;
2810 if (tem == 2)
2811 conditional = 1;
2812 }
2813
2814 }
2815 }
2816
2817 return 1 + conditional;
2818 }
2819
2820 \f
2821 /* Return nonzero if all the insns in the loop that set REG
2822 are INSN and the immediately following insns,
2823 and if each of those insns sets REG in an invariant way
2824 (not counting uses of REG in them).
2825
2826 The value is 2 if some of these insns are only conditionally invariant.
2827
2828 We assume that INSN itself is the first set of REG
2829 and that its source is invariant. */
2830
2831 static int
2832 consec_sets_invariant_p (reg, n_sets, insn)
2833 int n_sets;
2834 rtx reg, insn;
2835 {
2836 register rtx p = insn;
2837 register int regno = REGNO (reg);
2838 rtx temp;
2839 /* Number of sets we have to insist on finding after INSN. */
2840 int count = n_sets - 1;
2841 int old = n_times_set[regno];
2842 int value = 0;
2843 int this;
2844
2845 /* If N_SETS hit the limit, we can't rely on its value. */
2846 if (n_sets == 127)
2847 return 0;
2848
2849 n_times_set[regno] = 0;
2850
2851 while (count > 0)
2852 {
2853 register enum rtx_code code;
2854 rtx set;
2855
2856 p = NEXT_INSN (p);
2857 code = GET_CODE (p);
2858
2859 /* If library call, skip to end of of it. */
2860 if (code == INSN && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
2861 p = XEXP (temp, 0);
2862
2863 this = 0;
2864 if (code == INSN
2865 && (set = single_set (p))
2866 && GET_CODE (SET_DEST (set)) == REG
2867 && REGNO (SET_DEST (set)) == regno)
2868 {
2869 this = invariant_p (SET_SRC (set));
2870 if (this != 0)
2871 value |= this;
2872 else if (temp = find_reg_note (p, REG_EQUAL, NULL_RTX))
2873 {
2874 /* If this is a libcall, then any invariant REG_EQUAL note is OK.
2875 If this is an ordinary insn, then only CONSTANT_P REG_EQUAL
2876 notes are OK. */
2877 this = (CONSTANT_P (XEXP (temp, 0))
2878 || (find_reg_note (p, REG_RETVAL, NULL_RTX)
2879 && invariant_p (XEXP (temp, 0))));
2880 if (this != 0)
2881 value |= this;
2882 }
2883 }
2884 if (this != 0)
2885 count--;
2886 else if (code != NOTE)
2887 {
2888 n_times_set[regno] = old;
2889 return 0;
2890 }
2891 }
2892
2893 n_times_set[regno] = old;
2894 /* If invariant_p ever returned 2, we return 2. */
2895 return 1 + (value & 2);
2896 }
2897
2898 #if 0
2899 /* I don't think this condition is sufficient to allow INSN
2900 to be moved, so we no longer test it. */
2901
2902 /* Return 1 if all insns in the basic block of INSN and following INSN
2903 that set REG are invariant according to TABLE. */
2904
2905 static int
2906 all_sets_invariant_p (reg, insn, table)
2907 rtx reg, insn;
2908 short *table;
2909 {
2910 register rtx p = insn;
2911 register int regno = REGNO (reg);
2912
2913 while (1)
2914 {
2915 register enum rtx_code code;
2916 p = NEXT_INSN (p);
2917 code = GET_CODE (p);
2918 if (code == CODE_LABEL || code == JUMP_INSN)
2919 return 1;
2920 if (code == INSN && GET_CODE (PATTERN (p)) == SET
2921 && GET_CODE (SET_DEST (PATTERN (p))) == REG
2922 && REGNO (SET_DEST (PATTERN (p))) == regno)
2923 {
2924 if (!invariant_p (SET_SRC (PATTERN (p)), table))
2925 return 0;
2926 }
2927 }
2928 }
2929 #endif /* 0 */
2930 \f
2931 /* Look at all uses (not sets) of registers in X. For each, if it is
2932 the single use, set USAGE[REGNO] to INSN; if there was a previous use in
2933 a different insn, set USAGE[REGNO] to const0_rtx. */
2934
2935 static void
2936 find_single_use_in_loop (insn, x, usage)
2937 rtx insn;
2938 rtx x;
2939 rtx *usage;
2940 {
2941 enum rtx_code code = GET_CODE (x);
2942 char *fmt = GET_RTX_FORMAT (code);
2943 int i, j;
2944
2945 if (code == REG)
2946 usage[REGNO (x)]
2947 = (usage[REGNO (x)] != 0 && usage[REGNO (x)] != insn)
2948 ? const0_rtx : insn;
2949
2950 else if (code == SET)
2951 {
2952 /* Don't count SET_DEST if it is a REG; otherwise count things
2953 in SET_DEST because if a register is partially modified, it won't
2954 show up as a potential movable so we don't care how USAGE is set
2955 for it. */
2956 if (GET_CODE (SET_DEST (x)) != REG)
2957 find_single_use_in_loop (insn, SET_DEST (x), usage);
2958 find_single_use_in_loop (insn, SET_SRC (x), usage);
2959 }
2960 else
2961 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2962 {
2963 if (fmt[i] == 'e' && XEXP (x, i) != 0)
2964 find_single_use_in_loop (insn, XEXP (x, i), usage);
2965 else if (fmt[i] == 'E')
2966 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
2967 find_single_use_in_loop (insn, XVECEXP (x, i, j), usage);
2968 }
2969 }
2970 \f
2971 /* Increment N_TIMES_SET at the index of each register
2972 that is modified by an insn between FROM and TO.
2973 If the value of an element of N_TIMES_SET becomes 127 or more,
2974 stop incrementing it, to avoid overflow.
2975
2976 Store in SINGLE_USAGE[I] the single insn in which register I is
2977 used, if it is only used once. Otherwise, it is set to 0 (for no
2978 uses) or const0_rtx for more than one use. This parameter may be zero,
2979 in which case this processing is not done.
2980
2981 Store in *COUNT_PTR the number of actual instruction
2982 in the loop. We use this to decide what is worth moving out. */
2983
2984 /* last_set[n] is nonzero iff reg n has been set in the current basic block.
2985 In that case, it is the insn that last set reg n. */
2986
2987 static void
2988 count_loop_regs_set (from, to, may_not_move, single_usage, count_ptr, nregs)
2989 register rtx from, to;
2990 char *may_not_move;
2991 rtx *single_usage;
2992 int *count_ptr;
2993 int nregs;
2994 {
2995 register rtx *last_set = (rtx *) alloca (nregs * sizeof (rtx));
2996 register rtx insn;
2997 register int count = 0;
2998 register rtx dest;
2999
3000 bzero ((char *) last_set, nregs * sizeof (rtx));
3001 for (insn = from; insn != to; insn = NEXT_INSN (insn))
3002 {
3003 if (GET_RTX_CLASS (GET_CODE (insn)) == 'i')
3004 {
3005 ++count;
3006
3007 /* If requested, record registers that have exactly one use. */
3008 if (single_usage)
3009 {
3010 find_single_use_in_loop (insn, PATTERN (insn), single_usage);
3011
3012 /* Include uses in REG_EQUAL notes. */
3013 if (REG_NOTES (insn))
3014 find_single_use_in_loop (insn, REG_NOTES (insn), single_usage);
3015 }
3016
3017 if (GET_CODE (PATTERN (insn)) == CLOBBER
3018 && GET_CODE (XEXP (PATTERN (insn), 0)) == REG)
3019 /* Don't move a reg that has an explicit clobber.
3020 We might do so sometimes, but it's not worth the pain. */
3021 may_not_move[REGNO (XEXP (PATTERN (insn), 0))] = 1;
3022
3023 if (GET_CODE (PATTERN (insn)) == SET
3024 || GET_CODE (PATTERN (insn)) == CLOBBER)
3025 {
3026 dest = SET_DEST (PATTERN (insn));
3027 while (GET_CODE (dest) == SUBREG
3028 || GET_CODE (dest) == ZERO_EXTRACT
3029 || GET_CODE (dest) == SIGN_EXTRACT
3030 || GET_CODE (dest) == STRICT_LOW_PART)
3031 dest = XEXP (dest, 0);
3032 if (GET_CODE (dest) == REG)
3033 {
3034 register int regno = REGNO (dest);
3035 /* If this is the first setting of this reg
3036 in current basic block, and it was set before,
3037 it must be set in two basic blocks, so it cannot
3038 be moved out of the loop. */
3039 if (n_times_set[regno] > 0 && last_set[regno] == 0)
3040 may_not_move[regno] = 1;
3041 /* If this is not first setting in current basic block,
3042 see if reg was used in between previous one and this.
3043 If so, neither one can be moved. */
3044 if (last_set[regno] != 0
3045 && reg_used_between_p (dest, last_set[regno], insn))
3046 may_not_move[regno] = 1;
3047 if (n_times_set[regno] < 127)
3048 ++n_times_set[regno];
3049 last_set[regno] = insn;
3050 }
3051 }
3052 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
3053 {
3054 register int i;
3055 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
3056 {
3057 register rtx x = XVECEXP (PATTERN (insn), 0, i);
3058 if (GET_CODE (x) == CLOBBER && GET_CODE (XEXP (x, 0)) == REG)
3059 /* Don't move a reg that has an explicit clobber.
3060 It's not worth the pain to try to do it correctly. */
3061 may_not_move[REGNO (XEXP (x, 0))] = 1;
3062
3063 if (GET_CODE (x) == SET || GET_CODE (x) == CLOBBER)
3064 {
3065 dest = SET_DEST (x);
3066 while (GET_CODE (dest) == SUBREG
3067 || GET_CODE (dest) == ZERO_EXTRACT
3068 || GET_CODE (dest) == SIGN_EXTRACT
3069 || GET_CODE (dest) == STRICT_LOW_PART)
3070 dest = XEXP (dest, 0);
3071 if (GET_CODE (dest) == REG)
3072 {
3073 register int regno = REGNO (dest);
3074 if (n_times_set[regno] > 0 && last_set[regno] == 0)
3075 may_not_move[regno] = 1;
3076 if (last_set[regno] != 0
3077 && reg_used_between_p (dest, last_set[regno], insn))
3078 may_not_move[regno] = 1;
3079 if (n_times_set[regno] < 127)
3080 ++n_times_set[regno];
3081 last_set[regno] = insn;
3082 }
3083 }
3084 }
3085 }
3086 }
3087
3088 if (GET_CODE (insn) == CODE_LABEL || GET_CODE (insn) == JUMP_INSN)
3089 bzero ((char *) last_set, nregs * sizeof (rtx));
3090 }
3091 *count_ptr = count;
3092 }
3093 \f
3094 /* Given a loop that is bounded by LOOP_START and LOOP_END
3095 and that is entered at SCAN_START,
3096 return 1 if the register set in SET contained in insn INSN is used by
3097 any insn that precedes INSN in cyclic order starting
3098 from the loop entry point.
3099
3100 We don't want to use INSN_LUID here because if we restrict INSN to those
3101 that have a valid INSN_LUID, it means we cannot move an invariant out
3102 from an inner loop past two loops. */
3103
3104 static int
3105 loop_reg_used_before_p (set, insn, loop_start, scan_start, loop_end)
3106 rtx set, insn, loop_start, scan_start, loop_end;
3107 {
3108 rtx reg = SET_DEST (set);
3109 rtx p;
3110
3111 /* Scan forward checking for register usage. If we hit INSN, we
3112 are done. Otherwise, if we hit LOOP_END, wrap around to LOOP_START. */
3113 for (p = scan_start; p != insn; p = NEXT_INSN (p))
3114 {
3115 if (GET_RTX_CLASS (GET_CODE (p)) == 'i'
3116 && reg_overlap_mentioned_p (reg, PATTERN (p)))
3117 return 1;
3118
3119 if (p == loop_end)
3120 p = loop_start;
3121 }
3122
3123 return 0;
3124 }
3125 \f
3126 /* A "basic induction variable" or biv is a pseudo reg that is set
3127 (within this loop) only by incrementing or decrementing it. */
3128 /* A "general induction variable" or giv is a pseudo reg whose
3129 value is a linear function of a biv. */
3130
3131 /* Bivs are recognized by `basic_induction_var';
3132 Givs by `general_induct_var'. */
3133
3134 /* Indexed by register number, indicates whether or not register is an
3135 induction variable, and if so what type. */
3136
3137 enum iv_mode *reg_iv_type;
3138
3139 /* Indexed by register number, contains pointer to `struct induction'
3140 if register is an induction variable. This holds general info for
3141 all induction variables. */
3142
3143 struct induction **reg_iv_info;
3144
3145 /* Indexed by register number, contains pointer to `struct iv_class'
3146 if register is a basic induction variable. This holds info describing
3147 the class (a related group) of induction variables that the biv belongs
3148 to. */
3149
3150 struct iv_class **reg_biv_class;
3151
3152 /* The head of a list which links together (via the next field)
3153 every iv class for the current loop. */
3154
3155 struct iv_class *loop_iv_list;
3156
3157 /* Communication with routines called via `note_stores'. */
3158
3159 static rtx note_insn;
3160
3161 /* Dummy register to have non-zero DEST_REG for DEST_ADDR type givs. */
3162
3163 static rtx addr_placeholder;
3164
3165 /* ??? Unfinished optimizations, and possible future optimizations,
3166 for the strength reduction code. */
3167
3168 /* ??? There is one more optimization you might be interested in doing: to
3169 allocate pseudo registers for frequently-accessed memory locations.
3170 If the same memory location is referenced each time around, it might
3171 be possible to copy it into a register before and out after.
3172 This is especially useful when the memory location is a variable which
3173 is in a stack slot because somewhere its address is taken. If the
3174 loop doesn't contain a function call and the variable isn't volatile,
3175 it is safe to keep the value in a register for the duration of the
3176 loop. One tricky thing is that the copying of the value back from the
3177 register has to be done on all exits from the loop. You need to check that
3178 all the exits from the loop go to the same place. */
3179
3180 /* ??? The interaction of biv elimination, and recognition of 'constant'
3181 bivs, may cause problems. */
3182
3183 /* ??? Add heuristics so that DEST_ADDR strength reduction does not cause
3184 performance problems.
3185
3186 Perhaps don't eliminate things that can be combined with an addressing
3187 mode. Find all givs that have the same biv, mult_val, and add_val;
3188 then for each giv, check to see if its only use dies in a following
3189 memory address. If so, generate a new memory address and check to see
3190 if it is valid. If it is valid, then store the modified memory address,
3191 otherwise, mark the giv as not done so that it will get its own iv. */
3192
3193 /* ??? Could try to optimize branches when it is known that a biv is always
3194 positive. */
3195
3196 /* ??? When replace a biv in a compare insn, we should replace with closest
3197 giv so that an optimized branch can still be recognized by the combiner,
3198 e.g. the VAX acb insn. */
3199
3200 /* ??? Many of the checks involving uid_luid could be simplified if regscan
3201 was rerun in loop_optimize whenever a register was added or moved.
3202 Also, some of the optimizations could be a little less conservative. */
3203 \f
3204 /* Perform strength reduction and induction variable elimination. */
3205
3206 /* Pseudo registers created during this function will be beyond the last
3207 valid index in several tables including n_times_set and regno_last_uid.
3208 This does not cause a problem here, because the added registers cannot be
3209 givs outside of their loop, and hence will never be reconsidered.
3210 But scan_loop must check regnos to make sure they are in bounds. */
3211
3212 static void
3213 strength_reduce (scan_start, end, loop_top, insn_count,
3214 loop_start, loop_end)
3215 rtx scan_start;
3216 rtx end;
3217 rtx loop_top;
3218 int insn_count;
3219 rtx loop_start;
3220 rtx loop_end;
3221 {
3222 rtx p;
3223 rtx set;
3224 rtx inc_val;
3225 rtx mult_val;
3226 rtx dest_reg;
3227 /* This is 1 if current insn is not executed at least once for every loop
3228 iteration. */
3229 int not_every_iteration = 0;
3230 /* This is 1 if current insn may be executed more than once for every
3231 loop iteration. */
3232 int maybe_multiple = 0;
3233 /* Temporary list pointers for traversing loop_iv_list. */
3234 struct iv_class *bl, **backbl;
3235 /* Ratio of extra register life span we can justify
3236 for saving an instruction. More if loop doesn't call subroutines
3237 since in that case saving an insn makes more difference
3238 and more registers are available. */
3239 /* ??? could set this to last value of threshold in move_movables */
3240 int threshold = (loop_has_call ? 1 : 2) * (3 + n_non_fixed_regs);
3241 /* Map of pseudo-register replacements. */
3242 rtx *reg_map;
3243 int call_seen;
3244 rtx test;
3245 rtx end_insert_before;
3246 int loop_depth = 0;
3247
3248 reg_iv_type = (enum iv_mode *) alloca (max_reg_before_loop
3249 * sizeof (enum iv_mode *));
3250 bzero ((char *) reg_iv_type, max_reg_before_loop * sizeof (enum iv_mode *));
3251 reg_iv_info = (struct induction **)
3252 alloca (max_reg_before_loop * sizeof (struct induction *));
3253 bzero ((char *) reg_iv_info, (max_reg_before_loop
3254 * sizeof (struct induction *)));
3255 reg_biv_class = (struct iv_class **)
3256 alloca (max_reg_before_loop * sizeof (struct iv_class *));
3257 bzero ((char *) reg_biv_class, (max_reg_before_loop
3258 * sizeof (struct iv_class *)));
3259
3260 loop_iv_list = 0;
3261 addr_placeholder = gen_reg_rtx (Pmode);
3262
3263 /* Save insn immediately after the loop_end. Insns inserted after loop_end
3264 must be put before this insn, so that they will appear in the right
3265 order (i.e. loop order).
3266
3267 If loop_end is the end of the current function, then emit a
3268 NOTE_INSN_DELETED after loop_end and set end_insert_before to the
3269 dummy note insn. */
3270 if (NEXT_INSN (loop_end) != 0)
3271 end_insert_before = NEXT_INSN (loop_end);
3272 else
3273 end_insert_before = emit_note_after (NOTE_INSN_DELETED, loop_end);
3274
3275 /* Scan through loop to find all possible bivs. */
3276
3277 p = scan_start;
3278 while (1)
3279 {
3280 p = NEXT_INSN (p);
3281 /* At end of a straight-in loop, we are done.
3282 At end of a loop entered at the bottom, scan the top. */
3283 if (p == scan_start)
3284 break;
3285 if (p == end)
3286 {
3287 if (loop_top != 0)
3288 p = loop_top;
3289 else
3290 break;
3291 if (p == scan_start)
3292 break;
3293 }
3294
3295 if (GET_CODE (p) == INSN
3296 && (set = single_set (p))
3297 && GET_CODE (SET_DEST (set)) == REG)
3298 {
3299 dest_reg = SET_DEST (set);
3300 if (REGNO (dest_reg) < max_reg_before_loop
3301 && REGNO (dest_reg) >= FIRST_PSEUDO_REGISTER
3302 && reg_iv_type[REGNO (dest_reg)] != NOT_BASIC_INDUCT)
3303 {
3304 if (basic_induction_var (SET_SRC (set), GET_MODE (SET_SRC (set)),
3305 dest_reg, p, &inc_val, &mult_val))
3306 {
3307 /* It is a possible basic induction variable.
3308 Create and initialize an induction structure for it. */
3309
3310 struct induction *v
3311 = (struct induction *) alloca (sizeof (struct induction));
3312
3313 record_biv (v, p, dest_reg, inc_val, mult_val,
3314 not_every_iteration, maybe_multiple);
3315 reg_iv_type[REGNO (dest_reg)] = BASIC_INDUCT;
3316 }
3317 else if (REGNO (dest_reg) < max_reg_before_loop)
3318 reg_iv_type[REGNO (dest_reg)] = NOT_BASIC_INDUCT;
3319 }
3320 }
3321
3322 /* Past CODE_LABEL, we get to insns that may be executed multiple
3323 times. The only way we can be sure that they can't is if every
3324 every jump insn between here and the end of the loop either
3325 returns, exits the loop, is a forward jump, or is a jump
3326 to the loop start. */
3327
3328 if (GET_CODE (p) == CODE_LABEL)
3329 {
3330 rtx insn = p;
3331
3332 maybe_multiple = 0;
3333
3334 while (1)
3335 {
3336 insn = NEXT_INSN (insn);
3337 if (insn == scan_start)
3338 break;
3339 if (insn == end)
3340 {
3341 if (loop_top != 0)
3342 insn = loop_top;
3343 else
3344 break;
3345 if (insn == scan_start)
3346 break;
3347 }
3348
3349 if (GET_CODE (insn) == JUMP_INSN
3350 && GET_CODE (PATTERN (insn)) != RETURN
3351 && (! condjump_p (insn)
3352 || (JUMP_LABEL (insn) != 0
3353 && JUMP_LABEL (insn) != scan_start
3354 && (INSN_UID (JUMP_LABEL (insn)) >= max_uid_for_loop
3355 || INSN_UID (insn) >= max_uid_for_loop
3356 || (INSN_LUID (JUMP_LABEL (insn))
3357 < INSN_LUID (insn))))))
3358 {
3359 maybe_multiple = 1;
3360 break;
3361 }
3362 }
3363 }
3364
3365 /* Past a jump, we get to insns for which we can't count
3366 on whether they will be executed during each iteration. */
3367 /* This code appears twice in strength_reduce. There is also similar
3368 code in scan_loop. */
3369 if (GET_CODE (p) == JUMP_INSN
3370 /* If we enter the loop in the middle, and scan around to the
3371 beginning, don't set not_every_iteration for that.
3372 This can be any kind of jump, since we want to know if insns
3373 will be executed if the loop is executed. */
3374 && ! (JUMP_LABEL (p) == loop_top
3375 && ((NEXT_INSN (NEXT_INSN (p)) == loop_end && simplejump_p (p))
3376 || (NEXT_INSN (p) == loop_end && condjump_p (p)))))
3377 {
3378 rtx label = 0;
3379
3380 /* If this is a jump outside the loop, then it also doesn't
3381 matter. Check to see if the target of this branch is on the
3382 loop_number_exits_labels list. */
3383
3384 for (label = loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]];
3385 label;
3386 label = LABEL_NEXTREF (label))
3387 if (XEXP (label, 0) == JUMP_LABEL (p))
3388 break;
3389
3390 if (! label)
3391 not_every_iteration = 1;
3392 }
3393
3394 else if (GET_CODE (p) == NOTE)
3395 {
3396 /* At the virtual top of a converted loop, insns are again known to
3397 be executed each iteration: logically, the loop begins here
3398 even though the exit code has been duplicated. */
3399 if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP && loop_depth == 0)
3400 not_every_iteration = 0;
3401 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
3402 loop_depth++;
3403 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END)
3404 loop_depth--;
3405 }
3406
3407 /* Unlike in the code motion pass where MAYBE_NEVER indicates that
3408 an insn may never be executed, NOT_EVERY_ITERATION indicates whether
3409 or not an insn is known to be executed each iteration of the
3410 loop, whether or not any iterations are known to occur.
3411
3412 Therefore, if we have just passed a label and have no more labels
3413 between here and the test insn of the loop, we know these insns
3414 will be executed each iteration. */
3415
3416 if (not_every_iteration && GET_CODE (p) == CODE_LABEL
3417 && no_labels_between_p (p, loop_end))
3418 not_every_iteration = 0;
3419 }
3420
3421 /* Scan loop_iv_list to remove all regs that proved not to be bivs.
3422 Make a sanity check against n_times_set. */
3423 for (backbl = &loop_iv_list, bl = *backbl; bl; bl = bl->next)
3424 {
3425 if (reg_iv_type[bl->regno] != BASIC_INDUCT
3426 /* Above happens if register modified by subreg, etc. */
3427 /* Make sure it is not recognized as a basic induction var: */
3428 || n_times_set[bl->regno] != bl->biv_count
3429 /* If never incremented, it is invariant that we decided not to
3430 move. So leave it alone. */
3431 || ! bl->incremented)
3432 {
3433 if (loop_dump_stream)
3434 fprintf (loop_dump_stream, "Reg %d: biv discarded, %s\n",
3435 bl->regno,
3436 (reg_iv_type[bl->regno] != BASIC_INDUCT
3437 ? "not induction variable"
3438 : (! bl->incremented ? "never incremented"
3439 : "count error")));
3440
3441 reg_iv_type[bl->regno] = NOT_BASIC_INDUCT;
3442 *backbl = bl->next;
3443 }
3444 else
3445 {
3446 backbl = &bl->next;
3447
3448 if (loop_dump_stream)
3449 fprintf (loop_dump_stream, "Reg %d: biv verified\n", bl->regno);
3450 }
3451 }
3452
3453 /* Exit if there are no bivs. */
3454 if (! loop_iv_list)
3455 {
3456 /* Can still unroll the loop anyways, but indicate that there is no
3457 strength reduction info available. */
3458 if (flag_unroll_loops)
3459 unroll_loop (loop_end, insn_count, loop_start, end_insert_before, 0);
3460
3461 return;
3462 }
3463
3464 /* Find initial value for each biv by searching backwards from loop_start,
3465 halting at first label. Also record any test condition. */
3466
3467 call_seen = 0;
3468 for (p = loop_start; p && GET_CODE (p) != CODE_LABEL; p = PREV_INSN (p))
3469 {
3470 note_insn = p;
3471
3472 if (GET_CODE (p) == CALL_INSN)
3473 call_seen = 1;
3474
3475 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
3476 || GET_CODE (p) == CALL_INSN)
3477 note_stores (PATTERN (p), record_initial);
3478
3479 /* Record any test of a biv that branches around the loop if no store
3480 between it and the start of loop. We only care about tests with
3481 constants and registers and only certain of those. */
3482 if (GET_CODE (p) == JUMP_INSN
3483 && JUMP_LABEL (p) != 0
3484 && next_real_insn (JUMP_LABEL (p)) == next_real_insn (loop_end)
3485 && (test = get_condition_for_loop (p)) != 0
3486 && GET_CODE (XEXP (test, 0)) == REG
3487 && REGNO (XEXP (test, 0)) < max_reg_before_loop
3488 && (bl = reg_biv_class[REGNO (XEXP (test, 0))]) != 0
3489 && valid_initial_value_p (XEXP (test, 1), p, call_seen, loop_start)
3490 && bl->init_insn == 0)
3491 {
3492 /* If an NE test, we have an initial value! */
3493 if (GET_CODE (test) == NE)
3494 {
3495 bl->init_insn = p;
3496 bl->init_set = gen_rtx (SET, VOIDmode,
3497 XEXP (test, 0), XEXP (test, 1));
3498 }
3499 else
3500 bl->initial_test = test;
3501 }
3502 }
3503
3504 /* Look at the each biv and see if we can say anything better about its
3505 initial value from any initializing insns set up above. (This is done
3506 in two passes to avoid missing SETs in a PARALLEL.) */
3507 for (bl = loop_iv_list; bl; bl = bl->next)
3508 {
3509 rtx src;
3510
3511 if (! bl->init_insn)
3512 continue;
3513
3514 src = SET_SRC (bl->init_set);
3515
3516 if (loop_dump_stream)
3517 fprintf (loop_dump_stream,
3518 "Biv %d initialized at insn %d: initial value ",
3519 bl->regno, INSN_UID (bl->init_insn));
3520
3521 if ((GET_MODE (src) == GET_MODE (regno_reg_rtx[bl->regno])
3522 || GET_MODE (src) == VOIDmode)
3523 && valid_initial_value_p (src, bl->init_insn, call_seen, loop_start))
3524 {
3525 bl->initial_value = src;
3526
3527 if (loop_dump_stream)
3528 {
3529 if (GET_CODE (src) == CONST_INT)
3530 fprintf (loop_dump_stream, "%d\n", INTVAL (src));
3531 else
3532 {
3533 print_rtl (loop_dump_stream, src);
3534 fprintf (loop_dump_stream, "\n");
3535 }
3536 }
3537 }
3538 else
3539 {
3540 /* Biv initial value is not simple move,
3541 so let it keep initial value of "itself". */
3542
3543 if (loop_dump_stream)
3544 fprintf (loop_dump_stream, "is complex\n");
3545 }
3546 }
3547
3548 /* Search the loop for general induction variables. */
3549
3550 /* A register is a giv if: it is only set once, it is a function of a
3551 biv and a constant (or invariant), and it is not a biv. */
3552
3553 not_every_iteration = 0;
3554 loop_depth = 0;
3555 p = scan_start;
3556 while (1)
3557 {
3558 p = NEXT_INSN (p);
3559 /* At end of a straight-in loop, we are done.
3560 At end of a loop entered at the bottom, scan the top. */
3561 if (p == scan_start)
3562 break;
3563 if (p == end)
3564 {
3565 if (loop_top != 0)
3566 p = loop_top;
3567 else
3568 break;
3569 if (p == scan_start)
3570 break;
3571 }
3572
3573 /* Look for a general induction variable in a register. */
3574 if (GET_CODE (p) == INSN
3575 && (set = single_set (p))
3576 && GET_CODE (SET_DEST (set)) == REG
3577 && ! may_not_optimize[REGNO (SET_DEST (set))])
3578 {
3579 rtx src_reg;
3580 rtx add_val;
3581 rtx mult_val;
3582 int benefit;
3583 rtx regnote = 0;
3584
3585 dest_reg = SET_DEST (set);
3586 if (REGNO (dest_reg) < FIRST_PSEUDO_REGISTER)
3587 continue;
3588
3589 if (/* SET_SRC is a giv. */
3590 ((benefit = general_induction_var (SET_SRC (set),
3591 &src_reg, &add_val,
3592 &mult_val))
3593 /* Equivalent expression is a giv. */
3594 || ((regnote = find_reg_note (p, REG_EQUAL, NULL_RTX))
3595 && (benefit = general_induction_var (XEXP (regnote, 0),
3596 &src_reg,
3597 &add_val, &mult_val))))
3598 /* Don't try to handle any regs made by loop optimization.
3599 We have nothing on them in regno_first_uid, etc. */
3600 && REGNO (dest_reg) < max_reg_before_loop
3601 /* Don't recognize a BASIC_INDUCT_VAR here. */
3602 && dest_reg != src_reg
3603 /* This must be the only place where the register is set. */
3604 && (n_times_set[REGNO (dest_reg)] == 1
3605 /* or all sets must be consecutive and make a giv. */
3606 || (benefit = consec_sets_giv (benefit, p,
3607 src_reg, dest_reg,
3608 &add_val, &mult_val))))
3609 {
3610 int count;
3611 struct induction *v
3612 = (struct induction *) alloca (sizeof (struct induction));
3613 rtx temp;
3614
3615 /* If this is a library call, increase benefit. */
3616 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
3617 benefit += libcall_benefit (p);
3618
3619 /* Skip the consecutive insns, if there are any. */
3620 for (count = n_times_set[REGNO (dest_reg)] - 1;
3621 count > 0; count--)
3622 {
3623 /* If first insn of libcall sequence, skip to end.
3624 Do this at start of loop, since INSN is guaranteed to
3625 be an insn here. */
3626 if (GET_CODE (p) != NOTE
3627 && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
3628 p = XEXP (temp, 0);
3629
3630 do p = NEXT_INSN (p);
3631 while (GET_CODE (p) == NOTE);
3632 }
3633
3634 record_giv (v, p, src_reg, dest_reg, mult_val, add_val, benefit,
3635 DEST_REG, not_every_iteration, NULL_PTR, loop_start,
3636 loop_end);
3637
3638 }
3639 }
3640
3641 #ifndef DONT_REDUCE_ADDR
3642 /* Look for givs which are memory addresses. */
3643 /* This resulted in worse code on a VAX 8600. I wonder if it
3644 still does. */
3645 if (GET_CODE (p) == INSN)
3646 find_mem_givs (PATTERN (p), p, not_every_iteration, loop_start,
3647 loop_end);
3648 #endif
3649
3650 /* Update the status of whether giv can derive other givs. This can
3651 change when we pass a label or an insn that updates a biv. */
3652 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
3653 || GET_CODE (p) == CODE_LABEL)
3654 update_giv_derive (p);
3655
3656 /* Past a jump, we get to insns for which we can't count
3657 on whether they will be executed during each iteration. */
3658 /* This code appears twice in strength_reduce. There is also similar
3659 code in scan_loop. */
3660 if (GET_CODE (p) == JUMP_INSN
3661 /* If we enter the loop in the middle, and scan around to the
3662 beginning, don't set not_every_iteration for that.
3663 This can be any kind of jump, since we want to know if insns
3664 will be executed if the loop is executed. */
3665 && ! (JUMP_LABEL (p) == loop_top
3666 && ((NEXT_INSN (NEXT_INSN (p)) == loop_end && simplejump_p (p))
3667 || (NEXT_INSN (p) == loop_end && condjump_p (p)))))
3668 {
3669 rtx label = 0;
3670
3671 /* If this is a jump outside the loop, then it also doesn't
3672 matter. Check to see if the target of this branch is on the
3673 loop_number_exits_labels list. */
3674
3675 for (label = loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]];
3676 label;
3677 label = LABEL_NEXTREF (label))
3678 if (XEXP (label, 0) == JUMP_LABEL (p))
3679 break;
3680
3681 if (! label)
3682 not_every_iteration = 1;
3683 }
3684
3685 else if (GET_CODE (p) == NOTE)
3686 {
3687 /* At the virtual top of a converted loop, insns are again known to
3688 be executed each iteration: logically, the loop begins here
3689 even though the exit code has been duplicated. */
3690 if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP && loop_depth == 0)
3691 not_every_iteration = 0;
3692 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
3693 loop_depth++;
3694 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END)
3695 loop_depth--;
3696 }
3697
3698 /* Unlike in the code motion pass where MAYBE_NEVER indicates that
3699 an insn may never be executed, NOT_EVERY_ITERATION indicates whether
3700 or not an insn is known to be executed each iteration of the
3701 loop, whether or not any iterations are known to occur.
3702
3703 Therefore, if we have just passed a label and have no more labels
3704 between here and the test insn of the loop, we know these insns
3705 will be executed each iteration. */
3706
3707 if (not_every_iteration && GET_CODE (p) == CODE_LABEL
3708 && no_labels_between_p (p, loop_end))
3709 not_every_iteration = 0;
3710 }
3711
3712 /* Try to calculate and save the number of loop iterations. This is
3713 set to zero if the actual number can not be calculated. This must
3714 be called after all giv's have been identified, since otherwise it may
3715 fail if the iteration variable is a giv. */
3716
3717 loop_n_iterations = loop_iterations (loop_start, loop_end);
3718
3719 /* Now for each giv for which we still don't know whether or not it is
3720 replaceable, check to see if it is replaceable because its final value
3721 can be calculated. This must be done after loop_iterations is called,
3722 so that final_giv_value will work correctly. */
3723
3724 for (bl = loop_iv_list; bl; bl = bl->next)
3725 {
3726 struct induction *v;
3727
3728 for (v = bl->giv; v; v = v->next_iv)
3729 if (! v->replaceable && ! v->not_replaceable)
3730 check_final_value (v, loop_start, loop_end);
3731 }
3732
3733 /* Try to prove that the loop counter variable (if any) is always
3734 nonnegative; if so, record that fact with a REG_NONNEG note
3735 so that "decrement and branch until zero" insn can be used. */
3736 check_dbra_loop (loop_end, insn_count, loop_start);
3737
3738 /* Create reg_map to hold substitutions for replaceable giv regs. */
3739 reg_map = (rtx *) alloca (max_reg_before_loop * sizeof (rtx));
3740 bzero ((char *) reg_map, max_reg_before_loop * sizeof (rtx));
3741
3742 /* Examine each iv class for feasibility of strength reduction/induction
3743 variable elimination. */
3744
3745 for (bl = loop_iv_list; bl; bl = bl->next)
3746 {
3747 struct induction *v;
3748 int benefit;
3749 int all_reduced;
3750 rtx final_value = 0;
3751
3752 /* Test whether it will be possible to eliminate this biv
3753 provided all givs are reduced. This is possible if either
3754 the reg is not used outside the loop, or we can compute
3755 what its final value will be.
3756
3757 For architectures with a decrement_and_branch_until_zero insn,
3758 don't do this if we put a REG_NONNEG note on the endtest for
3759 this biv. */
3760
3761 /* Compare against bl->init_insn rather than loop_start.
3762 We aren't concerned with any uses of the biv between
3763 init_insn and loop_start since these won't be affected
3764 by the value of the biv elsewhere in the function, so
3765 long as init_insn doesn't use the biv itself.
3766 March 14, 1989 -- self@bayes.arc.nasa.gov */
3767
3768 if ((uid_luid[regno_last_uid[bl->regno]] < INSN_LUID (loop_end)
3769 && bl->init_insn
3770 && INSN_UID (bl->init_insn) < max_uid_for_loop
3771 && uid_luid[regno_first_uid[bl->regno]] >= INSN_LUID (bl->init_insn)
3772 #ifdef HAVE_decrement_and_branch_until_zero
3773 && ! bl->nonneg
3774 #endif
3775 && ! reg_mentioned_p (bl->biv->dest_reg, SET_SRC (bl->init_set)))
3776 || ((final_value = final_biv_value (bl, loop_start, loop_end))
3777 #ifdef HAVE_decrement_and_branch_until_zero
3778 && ! bl->nonneg
3779 #endif
3780 ))
3781 bl->eliminable = maybe_eliminate_biv (bl, loop_start, end, 0,
3782 threshold, insn_count);
3783 else
3784 {
3785 if (loop_dump_stream)
3786 {
3787 fprintf (loop_dump_stream,
3788 "Cannot eliminate biv %d.\n",
3789 bl->regno);
3790 fprintf (loop_dump_stream,
3791 "First use: insn %d, last use: insn %d.\n",
3792 regno_first_uid[bl->regno],
3793 regno_last_uid[bl->regno]);
3794 }
3795 }
3796
3797 /* Combine all giv's for this iv_class. */
3798 combine_givs (bl);
3799
3800 /* This will be true at the end, if all givs which depend on this
3801 biv have been strength reduced.
3802 We can't (currently) eliminate the biv unless this is so. */
3803 all_reduced = 1;
3804
3805 /* Check each giv in this class to see if we will benefit by reducing
3806 it. Skip giv's combined with others. */
3807 for (v = bl->giv; v; v = v->next_iv)
3808 {
3809 struct induction *tv;
3810
3811 if (v->ignore || v->same)
3812 continue;
3813
3814 benefit = v->benefit;
3815
3816 /* Reduce benefit if not replaceable, since we will insert
3817 a move-insn to replace the insn that calculates this giv.
3818 Don't do this unless the giv is a user variable, since it
3819 will often be marked non-replaceable because of the duplication
3820 of the exit code outside the loop. In such a case, the copies
3821 we insert are dead and will be deleted. So they don't have
3822 a cost. Similar situations exist. */
3823 /* ??? The new final_[bg]iv_value code does a much better job
3824 of finding replaceable giv's, and hence this code may no longer
3825 be necessary. */
3826 if (! v->replaceable && ! bl->eliminable
3827 && REG_USERVAR_P (v->dest_reg))
3828 benefit -= copy_cost;
3829
3830 /* Decrease the benefit to count the add-insns that we will
3831 insert to increment the reduced reg for the giv. */
3832 benefit -= add_cost * bl->biv_count;
3833
3834 /* Decide whether to strength-reduce this giv or to leave the code
3835 unchanged (recompute it from the biv each time it is used).
3836 This decision can be made independently for each giv. */
3837
3838 /* ??? Perhaps attempt to guess whether autoincrement will handle
3839 some of the new add insns; if so, can increase BENEFIT
3840 (undo the subtraction of add_cost that was done above). */
3841
3842 /* If an insn is not to be strength reduced, then set its ignore
3843 flag, and clear all_reduced. */
3844
3845 /* A giv that depends on a reversed biv must be reduced if it is
3846 used after the loop exit, otherwise, it would have the wrong
3847 value after the loop exit. To make it simple, just reduce all
3848 of such giv's whether or not we know they are used after the loop
3849 exit. */
3850
3851 if (v->lifetime * threshold * benefit < insn_count
3852 && ! bl->reversed)
3853 {
3854 if (loop_dump_stream)
3855 fprintf (loop_dump_stream,
3856 "giv of insn %d not worth while, %d vs %d.\n",
3857 INSN_UID (v->insn),
3858 v->lifetime * threshold * benefit, insn_count);
3859 v->ignore = 1;
3860 all_reduced = 0;
3861 }
3862 else
3863 {
3864 /* Check that we can increment the reduced giv without a
3865 multiply insn. If not, reject it. */
3866
3867 for (tv = bl->biv; tv; tv = tv->next_iv)
3868 if (tv->mult_val == const1_rtx
3869 && ! product_cheap_p (tv->add_val, v->mult_val))
3870 {
3871 if (loop_dump_stream)
3872 fprintf (loop_dump_stream,
3873 "giv of insn %d: would need a multiply.\n",
3874 INSN_UID (v->insn));
3875 v->ignore = 1;
3876 all_reduced = 0;
3877 break;
3878 }
3879 }
3880 }
3881
3882 /* Reduce each giv that we decided to reduce. */
3883
3884 for (v = bl->giv; v; v = v->next_iv)
3885 {
3886 struct induction *tv;
3887 if (! v->ignore && v->same == 0)
3888 {
3889 int auto_inc_opt = 0;
3890
3891 v->new_reg = gen_reg_rtx (v->mode);
3892
3893 #ifdef AUTO_INC_DEC
3894 /* If the target has auto-increment addressing modes, and
3895 this is an address giv, then try to put the increment
3896 immediately after its use, so that flow can create an
3897 auto-increment addressing mode. */
3898 if (v->giv_type == DEST_ADDR && bl->biv_count == 1
3899 && bl->biv->always_executed
3900 && ! bl->biv->maybe_multiple
3901 && v->always_executed && ! v->maybe_multiple)
3902 {
3903 /* If other giv's have been combined with this one, then
3904 this will work only if all uses of the other giv's occur
3905 before this giv's insn. This is difficult to check.
3906
3907 We simplify this by looking for the common case where
3908 there is one DEST_REG giv, and this giv's insn is the
3909 last use of the dest_reg of that DEST_REG giv. If the
3910 the increment occurs after the address giv, then we can
3911 perform the optimization. (Otherwise, the increment
3912 would have to go before other_giv, and we would not be
3913 able to combine it with the address giv to get an
3914 auto-inc address.) */
3915 if (v->combined_with)
3916 {
3917 struct induction *other_giv = 0;
3918
3919 for (tv = bl->giv; tv; tv = tv->next_iv)
3920 if (tv->same == v)
3921 {
3922 if (other_giv)
3923 break;
3924 else
3925 other_giv = tv;
3926 }
3927 if (! tv && other_giv
3928 && (regno_last_uid[REGNO (other_giv->dest_reg)]
3929 == INSN_UID (v->insn))
3930 && INSN_LUID (v->insn) < INSN_LUID (bl->biv->insn))
3931 auto_inc_opt = 1;
3932 }
3933 /* Check for case where increment is before the the address
3934 giv. */
3935 else if (INSN_LUID (v->insn) > INSN_LUID (bl->biv->insn))
3936 auto_inc_opt = -1;
3937 else
3938 auto_inc_opt = 1;
3939
3940 #ifdef HAVE_cc0
3941 /* We can't put an insn immediately after one setting
3942 cc0, or immediately before one using cc0. */
3943 if ((auto_inc_opt == 1 && sets_cc0_p (PATTERN (v->insn)))
3944 || (auto_inc_opt == -1
3945 && sets_cc0_p (PATTERN (prev_nonnote_insn (v->insn)))))
3946 auto_inc_opt = 0;
3947 #endif
3948
3949 if (auto_inc_opt)
3950 v->auto_inc_opt = 1;
3951 }
3952 #endif
3953
3954 /* For each place where the biv is incremented, add an insn
3955 to increment the new, reduced reg for the giv. */
3956 for (tv = bl->biv; tv; tv = tv->next_iv)
3957 {
3958 rtx insert_before;
3959
3960 if (! auto_inc_opt)
3961 insert_before = tv->insn;
3962 else if (auto_inc_opt == 1)
3963 insert_before = NEXT_INSN (v->insn);
3964 else
3965 insert_before = v->insn;
3966
3967 if (tv->mult_val == const1_rtx)
3968 emit_iv_add_mult (tv->add_val, v->mult_val,
3969 v->new_reg, v->new_reg, insert_before);
3970 else /* tv->mult_val == const0_rtx */
3971 /* A multiply is acceptable here
3972 since this is presumed to be seldom executed. */
3973 emit_iv_add_mult (tv->add_val, v->mult_val,
3974 v->add_val, v->new_reg, insert_before);
3975 }
3976
3977 /* Add code at loop start to initialize giv's reduced reg. */
3978
3979 emit_iv_add_mult (bl->initial_value, v->mult_val,
3980 v->add_val, v->new_reg, loop_start);
3981 }
3982 }
3983
3984 /* Rescan all givs. If a giv is the same as a giv not reduced, mark it
3985 as not reduced.
3986
3987 For each giv register that can be reduced now: if replaceable,
3988 substitute reduced reg wherever the old giv occurs;
3989 else add new move insn "giv_reg = reduced_reg".
3990
3991 Also check for givs whose first use is their definition and whose
3992 last use is the definition of another giv. If so, it is likely
3993 dead and should not be used to eliminate a biv. */
3994 for (v = bl->giv; v; v = v->next_iv)
3995 {
3996 if (v->same && v->same->ignore)
3997 v->ignore = 1;
3998
3999 if (v->ignore)
4000 continue;
4001
4002 if (v->giv_type == DEST_REG
4003 && regno_first_uid[REGNO (v->dest_reg)] == INSN_UID (v->insn))
4004 {
4005 struct induction *v1;
4006
4007 for (v1 = bl->giv; v1; v1 = v1->next_iv)
4008 if (regno_last_uid[REGNO (v->dest_reg)] == INSN_UID (v1->insn))
4009 v->maybe_dead = 1;
4010 }
4011
4012 /* Update expression if this was combined, in case other giv was
4013 replaced. */
4014 if (v->same)
4015 v->new_reg = replace_rtx (v->new_reg,
4016 v->same->dest_reg, v->same->new_reg);
4017
4018 if (v->giv_type == DEST_ADDR)
4019 /* Store reduced reg as the address in the memref where we found
4020 this giv. */
4021 validate_change (v->insn, v->location, v->new_reg, 0);
4022 else if (v->replaceable)
4023 {
4024 reg_map[REGNO (v->dest_reg)] = v->new_reg;
4025
4026 #if 0
4027 /* I can no longer duplicate the original problem. Perhaps
4028 this is unnecessary now? */
4029
4030 /* Replaceable; it isn't strictly necessary to delete the old
4031 insn and emit a new one, because v->dest_reg is now dead.
4032
4033 However, especially when unrolling loops, the special
4034 handling for (set REG0 REG1) in the second cse pass may
4035 make v->dest_reg live again. To avoid this problem, emit
4036 an insn to set the original giv reg from the reduced giv.
4037 We can not delete the original insn, since it may be part
4038 of a LIBCALL, and the code in flow that eliminates dead
4039 libcalls will fail if it is deleted. */
4040 emit_insn_after (gen_move_insn (v->dest_reg, v->new_reg),
4041 v->insn);
4042 #endif
4043 }
4044 else
4045 {
4046 /* Not replaceable; emit an insn to set the original giv reg from
4047 the reduced giv, same as above. */
4048 emit_insn_after (gen_move_insn (v->dest_reg, v->new_reg),
4049 v->insn);
4050 }
4051
4052 /* When a loop is reversed, givs which depend on the reversed
4053 biv, and which are live outside the loop, must be set to their
4054 correct final value. This insn is only needed if the giv is
4055 not replaceable. The correct final value is the same as the
4056 value that the giv starts the reversed loop with. */
4057 if (bl->reversed && ! v->replaceable)
4058 emit_iv_add_mult (bl->initial_value, v->mult_val,
4059 v->add_val, v->dest_reg, end_insert_before);
4060 else if (v->final_value)
4061 {
4062 rtx insert_before;
4063
4064 /* If the loop has multiple exits, emit the insn before the
4065 loop to ensure that it will always be executed no matter
4066 how the loop exits. Otherwise, emit the insn after the loop,
4067 since this is slightly more efficient. */
4068 if (loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]])
4069 insert_before = loop_start;
4070 else
4071 insert_before = end_insert_before;
4072 emit_insn_before (gen_move_insn (v->dest_reg, v->final_value),
4073 insert_before);
4074
4075 #if 0
4076 /* If the insn to set the final value of the giv was emitted
4077 before the loop, then we must delete the insn inside the loop
4078 that sets it. If this is a LIBCALL, then we must delete
4079 every insn in the libcall. Note, however, that
4080 final_giv_value will only succeed when there are multiple
4081 exits if the giv is dead at each exit, hence it does not
4082 matter that the original insn remains because it is dead
4083 anyways. */
4084 /* Delete the insn inside the loop that sets the giv since
4085 the giv is now set before (or after) the loop. */
4086 delete_insn (v->insn);
4087 #endif
4088 }
4089
4090 if (loop_dump_stream)
4091 {
4092 fprintf (loop_dump_stream, "giv at %d reduced to ",
4093 INSN_UID (v->insn));
4094 print_rtl (loop_dump_stream, v->new_reg);
4095 fprintf (loop_dump_stream, "\n");
4096 }
4097 }
4098
4099 /* All the givs based on the biv bl have been reduced if they
4100 merit it. */
4101
4102 /* For each giv not marked as maybe dead that has been combined with a
4103 second giv, clear any "maybe dead" mark on that second giv.
4104 v->new_reg will either be or refer to the register of the giv it
4105 combined with.
4106
4107 Doing this clearing avoids problems in biv elimination where a
4108 giv's new_reg is a complex value that can't be put in the insn but
4109 the giv combined with (with a reg as new_reg) is marked maybe_dead.
4110 Since the register will be used in either case, we'd prefer it be
4111 used from the simpler giv. */
4112
4113 for (v = bl->giv; v; v = v->next_iv)
4114 if (! v->maybe_dead && v->same)
4115 v->same->maybe_dead = 0;
4116
4117 /* Try to eliminate the biv, if it is a candidate.
4118 This won't work if ! all_reduced,
4119 since the givs we planned to use might not have been reduced.
4120
4121 We have to be careful that we didn't initially think we could eliminate
4122 this biv because of a giv that we now think may be dead and shouldn't
4123 be used as a biv replacement.
4124
4125 Also, there is the possibility that we may have a giv that looks
4126 like it can be used to eliminate a biv, but the resulting insn
4127 isn't valid. This can happen, for example, on the 88k, where a
4128 JUMP_INSN can compare a register only with zero. Attempts to
4129 replace it with a compare with a constant will fail.
4130
4131 Note that in cases where this call fails, we may have replaced some
4132 of the occurrences of the biv with a giv, but no harm was done in
4133 doing so in the rare cases where it can occur. */
4134
4135 if (all_reduced == 1 && bl->eliminable
4136 && maybe_eliminate_biv (bl, loop_start, end, 1,
4137 threshold, insn_count))
4138
4139 {
4140 /* ?? If we created a new test to bypass the loop entirely,
4141 or otherwise drop straight in, based on this test, then
4142 we might want to rewrite it also. This way some later
4143 pass has more hope of removing the initialization of this
4144 biv entirely. */
4145
4146 /* If final_value != 0, then the biv may be used after loop end
4147 and we must emit an insn to set it just in case.
4148
4149 Reversed bivs already have an insn after the loop setting their
4150 value, so we don't need another one. We can't calculate the
4151 proper final value for such a biv here anyways. */
4152 if (final_value != 0 && ! bl->reversed)
4153 {
4154 rtx insert_before;
4155
4156 /* If the loop has multiple exits, emit the insn before the
4157 loop to ensure that it will always be executed no matter
4158 how the loop exits. Otherwise, emit the insn after the
4159 loop, since this is slightly more efficient. */
4160 if (loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]])
4161 insert_before = loop_start;
4162 else
4163 insert_before = end_insert_before;
4164
4165 emit_insn_before (gen_move_insn (bl->biv->dest_reg, final_value),
4166 end_insert_before);
4167 }
4168
4169 #if 0
4170 /* Delete all of the instructions inside the loop which set
4171 the biv, as they are all dead. If is safe to delete them,
4172 because an insn setting a biv will never be part of a libcall. */
4173 /* However, deleting them will invalidate the regno_last_uid info,
4174 so keeping them around is more convenient. Final_biv_value
4175 will only succeed when there are multiple exits if the biv
4176 is dead at each exit, hence it does not matter that the original
4177 insn remains, because it is dead anyways. */
4178 for (v = bl->biv; v; v = v->next_iv)
4179 delete_insn (v->insn);
4180 #endif
4181
4182 if (loop_dump_stream)
4183 fprintf (loop_dump_stream, "Reg %d: biv eliminated\n",
4184 bl->regno);
4185 }
4186 }
4187
4188 /* Go through all the instructions in the loop, making all the
4189 register substitutions scheduled in REG_MAP. */
4190
4191 for (p = loop_start; p != end; p = NEXT_INSN (p))
4192 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
4193 || GET_CODE (p) == CALL_INSN)
4194 {
4195 replace_regs (PATTERN (p), reg_map, max_reg_before_loop, 0);
4196 replace_regs (REG_NOTES (p), reg_map, max_reg_before_loop, 0);
4197 INSN_CODE (p) = -1;
4198 }
4199
4200 /* Unroll loops from within strength reduction so that we can use the
4201 induction variable information that strength_reduce has already
4202 collected. */
4203
4204 if (flag_unroll_loops)
4205 unroll_loop (loop_end, insn_count, loop_start, end_insert_before, 1);
4206
4207 if (loop_dump_stream)
4208 fprintf (loop_dump_stream, "\n");
4209 }
4210 \f
4211 /* Return 1 if X is a valid source for an initial value (or as value being
4212 compared against in an initial test).
4213
4214 X must be either a register or constant and must not be clobbered between
4215 the current insn and the start of the loop.
4216
4217 INSN is the insn containing X. */
4218
4219 static int
4220 valid_initial_value_p (x, insn, call_seen, loop_start)
4221 rtx x;
4222 rtx insn;
4223 int call_seen;
4224 rtx loop_start;
4225 {
4226 if (CONSTANT_P (x))
4227 return 1;
4228
4229 /* Only consider pseudos we know about initialized in insns whose luids
4230 we know. */
4231 if (GET_CODE (x) != REG
4232 || REGNO (x) >= max_reg_before_loop)
4233 return 0;
4234
4235 /* Don't use call-clobbered registers across a call which clobbers it. On
4236 some machines, don't use any hard registers at all. */
4237 if (REGNO (x) < FIRST_PSEUDO_REGISTER
4238 #ifndef SMALL_REGISTER_CLASSES
4239 && call_used_regs[REGNO (x)] && call_seen
4240 #endif
4241 )
4242 return 0;
4243
4244 /* Don't use registers that have been clobbered before the start of the
4245 loop. */
4246 if (reg_set_between_p (x, insn, loop_start))
4247 return 0;
4248
4249 return 1;
4250 }
4251 \f
4252 /* Scan X for memory refs and check each memory address
4253 as a possible giv. INSN is the insn whose pattern X comes from.
4254 NOT_EVERY_ITERATION is 1 if the insn might not be executed during
4255 every loop iteration. */
4256
4257 static void
4258 find_mem_givs (x, insn, not_every_iteration, loop_start, loop_end)
4259 rtx x;
4260 rtx insn;
4261 int not_every_iteration;
4262 rtx loop_start, loop_end;
4263 {
4264 register int i, j;
4265 register enum rtx_code code;
4266 register char *fmt;
4267
4268 if (x == 0)
4269 return;
4270
4271 code = GET_CODE (x);
4272 switch (code)
4273 {
4274 case REG:
4275 case CONST_INT:
4276 case CONST:
4277 case CONST_DOUBLE:
4278 case SYMBOL_REF:
4279 case LABEL_REF:
4280 case PC:
4281 case CC0:
4282 case ADDR_VEC:
4283 case ADDR_DIFF_VEC:
4284 case USE:
4285 case CLOBBER:
4286 return;
4287
4288 case MEM:
4289 {
4290 rtx src_reg;
4291 rtx add_val;
4292 rtx mult_val;
4293 int benefit;
4294
4295 benefit = general_induction_var (XEXP (x, 0),
4296 &src_reg, &add_val, &mult_val);
4297
4298 /* Don't make a DEST_ADDR giv with mult_val == 1 && add_val == 0.
4299 Such a giv isn't useful. */
4300 if (benefit > 0 && (mult_val != const1_rtx || add_val != const0_rtx))
4301 {
4302 /* Found one; record it. */
4303 struct induction *v
4304 = (struct induction *) oballoc (sizeof (struct induction));
4305
4306 record_giv (v, insn, src_reg, addr_placeholder, mult_val,
4307 add_val, benefit, DEST_ADDR, not_every_iteration,
4308 &XEXP (x, 0), loop_start, loop_end);
4309
4310 v->mem_mode = GET_MODE (x);
4311 }
4312 return;
4313 }
4314 }
4315
4316 /* Recursively scan the subexpressions for other mem refs. */
4317
4318 fmt = GET_RTX_FORMAT (code);
4319 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4320 if (fmt[i] == 'e')
4321 find_mem_givs (XEXP (x, i), insn, not_every_iteration, loop_start,
4322 loop_end);
4323 else if (fmt[i] == 'E')
4324 for (j = 0; j < XVECLEN (x, i); j++)
4325 find_mem_givs (XVECEXP (x, i, j), insn, not_every_iteration,
4326 loop_start, loop_end);
4327 }
4328 \f
4329 /* Fill in the data about one biv update.
4330 V is the `struct induction' in which we record the biv. (It is
4331 allocated by the caller, with alloca.)
4332 INSN is the insn that sets it.
4333 DEST_REG is the biv's reg.
4334
4335 MULT_VAL is const1_rtx if the biv is being incremented here, in which case
4336 INC_VAL is the increment. Otherwise, MULT_VAL is const0_rtx and the biv is
4337 being set to INC_VAL.
4338
4339 NOT_EVERY_ITERATION is nonzero if this biv update is not know to be
4340 executed every iteration; MAYBE_MULTIPLE is nonzero if this biv update
4341 can be executed more than once per iteration. If MAYBE_MULTIPLE
4342 and NOT_EVERY_ITERATION are both zero, we know that the biv update is
4343 executed exactly once per iteration. */
4344
4345 static void
4346 record_biv (v, insn, dest_reg, inc_val, mult_val,
4347 not_every_iteration, maybe_multiple)
4348 struct induction *v;
4349 rtx insn;
4350 rtx dest_reg;
4351 rtx inc_val;
4352 rtx mult_val;
4353 int not_every_iteration;
4354 int maybe_multiple;
4355 {
4356 struct iv_class *bl;
4357
4358 v->insn = insn;
4359 v->src_reg = dest_reg;
4360 v->dest_reg = dest_reg;
4361 v->mult_val = mult_val;
4362 v->add_val = inc_val;
4363 v->mode = GET_MODE (dest_reg);
4364 v->always_computable = ! not_every_iteration;
4365 v->always_executed = ! not_every_iteration;
4366 v->maybe_multiple = maybe_multiple;
4367
4368 /* Add this to the reg's iv_class, creating a class
4369 if this is the first incrementation of the reg. */
4370
4371 bl = reg_biv_class[REGNO (dest_reg)];
4372 if (bl == 0)
4373 {
4374 /* Create and initialize new iv_class. */
4375
4376 bl = (struct iv_class *) oballoc (sizeof (struct iv_class));
4377
4378 bl->regno = REGNO (dest_reg);
4379 bl->biv = 0;
4380 bl->giv = 0;
4381 bl->biv_count = 0;
4382 bl->giv_count = 0;
4383
4384 /* Set initial value to the reg itself. */
4385 bl->initial_value = dest_reg;
4386 /* We haven't seen the initializing insn yet */
4387 bl->init_insn = 0;
4388 bl->init_set = 0;
4389 bl->initial_test = 0;
4390 bl->incremented = 0;
4391 bl->eliminable = 0;
4392 bl->nonneg = 0;
4393 bl->reversed = 0;
4394 bl->total_benefit = 0;
4395
4396 /* Add this class to loop_iv_list. */
4397 bl->next = loop_iv_list;
4398 loop_iv_list = bl;
4399
4400 /* Put it in the array of biv register classes. */
4401 reg_biv_class[REGNO (dest_reg)] = bl;
4402 }
4403
4404 /* Update IV_CLASS entry for this biv. */
4405 v->next_iv = bl->biv;
4406 bl->biv = v;
4407 bl->biv_count++;
4408 if (mult_val == const1_rtx)
4409 bl->incremented = 1;
4410
4411 if (loop_dump_stream)
4412 {
4413 fprintf (loop_dump_stream,
4414 "Insn %d: possible biv, reg %d,",
4415 INSN_UID (insn), REGNO (dest_reg));
4416 if (GET_CODE (inc_val) == CONST_INT)
4417 fprintf (loop_dump_stream, " const = %d\n",
4418 INTVAL (inc_val));
4419 else
4420 {
4421 fprintf (loop_dump_stream, " const = ");
4422 print_rtl (loop_dump_stream, inc_val);
4423 fprintf (loop_dump_stream, "\n");
4424 }
4425 }
4426 }
4427 \f
4428 /* Fill in the data about one giv.
4429 V is the `struct induction' in which we record the giv. (It is
4430 allocated by the caller, with alloca.)
4431 INSN is the insn that sets it.
4432 BENEFIT estimates the savings from deleting this insn.
4433 TYPE is DEST_REG or DEST_ADDR; it says whether the giv is computed
4434 into a register or is used as a memory address.
4435
4436 SRC_REG is the biv reg which the giv is computed from.
4437 DEST_REG is the giv's reg (if the giv is stored in a reg).
4438 MULT_VAL and ADD_VAL are the coefficients used to compute the giv.
4439 LOCATION points to the place where this giv's value appears in INSN. */
4440
4441 static void
4442 record_giv (v, insn, src_reg, dest_reg, mult_val, add_val, benefit,
4443 type, not_every_iteration, location, loop_start, loop_end)
4444 struct induction *v;
4445 rtx insn;
4446 rtx src_reg;
4447 rtx dest_reg;
4448 rtx mult_val, add_val;
4449 int benefit;
4450 enum g_types type;
4451 int not_every_iteration;
4452 rtx *location;
4453 rtx loop_start, loop_end;
4454 {
4455 struct induction *b;
4456 struct iv_class *bl;
4457 rtx set = single_set (insn);
4458 rtx p;
4459
4460 v->insn = insn;
4461 v->src_reg = src_reg;
4462 v->giv_type = type;
4463 v->dest_reg = dest_reg;
4464 v->mult_val = mult_val;
4465 v->add_val = add_val;
4466 v->benefit = benefit;
4467 v->location = location;
4468 v->cant_derive = 0;
4469 v->combined_with = 0;
4470 v->maybe_multiple = 0;
4471 v->maybe_dead = 0;
4472 v->derive_adjustment = 0;
4473 v->same = 0;
4474 v->ignore = 0;
4475 v->new_reg = 0;
4476 v->final_value = 0;
4477 v->same_insn = 0;
4478 v->auto_inc_opt = 0;
4479
4480 /* The v->always_computable field is used in update_giv_derive, to
4481 determine whether a giv can be used to derive another giv. For a
4482 DEST_REG giv, INSN computes a new value for the giv, so its value
4483 isn't computable if INSN insn't executed every iteration.
4484 However, for a DEST_ADDR giv, INSN merely uses the value of the giv;
4485 it does not compute a new value. Hence the value is always computable
4486 regardless of whether INSN is executed each iteration. */
4487
4488 if (type == DEST_ADDR)
4489 v->always_computable = 1;
4490 else
4491 v->always_computable = ! not_every_iteration;
4492
4493 v->always_executed = ! not_every_iteration;
4494
4495 if (type == DEST_ADDR)
4496 {
4497 v->mode = GET_MODE (*location);
4498 v->lifetime = 1;
4499 v->times_used = 1;
4500 }
4501 else /* type == DEST_REG */
4502 {
4503 v->mode = GET_MODE (SET_DEST (set));
4504
4505 v->lifetime = (uid_luid[regno_last_uid[REGNO (dest_reg)]]
4506 - uid_luid[regno_first_uid[REGNO (dest_reg)]]);
4507
4508 v->times_used = n_times_used[REGNO (dest_reg)];
4509
4510 /* If the lifetime is zero, it means that this register is
4511 really a dead store. So mark this as a giv that can be
4512 ignored. This will not prevent the biv from being eliminated. */
4513 if (v->lifetime == 0)
4514 v->ignore = 1;
4515
4516 reg_iv_type[REGNO (dest_reg)] = GENERAL_INDUCT;
4517 reg_iv_info[REGNO (dest_reg)] = v;
4518 }
4519
4520 /* Add the giv to the class of givs computed from one biv. */
4521
4522 bl = reg_biv_class[REGNO (src_reg)];
4523 if (bl)
4524 {
4525 v->next_iv = bl->giv;
4526 bl->giv = v;
4527 /* Don't count DEST_ADDR. This is supposed to count the number of
4528 insns that calculate givs. */
4529 if (type == DEST_REG)
4530 bl->giv_count++;
4531 bl->total_benefit += benefit;
4532 }
4533 else
4534 /* Fatal error, biv missing for this giv? */
4535 abort ();
4536
4537 if (type == DEST_ADDR)
4538 v->replaceable = 1;
4539 else
4540 {
4541 /* The giv can be replaced outright by the reduced register only if all
4542 of the following conditions are true:
4543 - the insn that sets the giv is always executed on any iteration
4544 on which the giv is used at all
4545 (there are two ways to deduce this:
4546 either the insn is executed on every iteration,
4547 or all uses follow that insn in the same basic block),
4548 - the giv is not used outside the loop
4549 - no assignments to the biv occur during the giv's lifetime. */
4550
4551 if (regno_first_uid[REGNO (dest_reg)] == INSN_UID (insn)
4552 /* Previous line always fails if INSN was moved by loop opt. */
4553 && uid_luid[regno_last_uid[REGNO (dest_reg)]] < INSN_LUID (loop_end)
4554 && (! not_every_iteration
4555 || last_use_this_basic_block (dest_reg, insn)))
4556 {
4557 /* Now check that there are no assignments to the biv within the
4558 giv's lifetime. This requires two separate checks. */
4559
4560 /* Check each biv update, and fail if any are between the first
4561 and last use of the giv.
4562
4563 If this loop contains an inner loop that was unrolled, then
4564 the insn modifying the biv may have been emitted by the loop
4565 unrolling code, and hence does not have a valid luid. Just
4566 mark the biv as not replaceable in this case. It is not very
4567 useful as a biv, because it is used in two different loops.
4568 It is very unlikely that we would be able to optimize the giv
4569 using this biv anyways. */
4570
4571 v->replaceable = 1;
4572 for (b = bl->biv; b; b = b->next_iv)
4573 {
4574 if (INSN_UID (b->insn) >= max_uid_for_loop
4575 || ((uid_luid[INSN_UID (b->insn)]
4576 >= uid_luid[regno_first_uid[REGNO (dest_reg)]])
4577 && (uid_luid[INSN_UID (b->insn)]
4578 <= uid_luid[regno_last_uid[REGNO (dest_reg)]])))
4579 {
4580 v->replaceable = 0;
4581 v->not_replaceable = 1;
4582 break;
4583 }
4584 }
4585
4586 /* If there are any backwards branches that go from after the
4587 biv update to before it, then this giv is not replaceable. */
4588 if (v->replaceable)
4589 for (b = bl->biv; b; b = b->next_iv)
4590 if (back_branch_in_range_p (b->insn, loop_start, loop_end))
4591 {
4592 v->replaceable = 0;
4593 v->not_replaceable = 1;
4594 break;
4595 }
4596 }
4597 else
4598 {
4599 /* May still be replaceable, we don't have enough info here to
4600 decide. */
4601 v->replaceable = 0;
4602 v->not_replaceable = 0;
4603 }
4604 }
4605
4606 if (loop_dump_stream)
4607 {
4608 if (type == DEST_REG)
4609 fprintf (loop_dump_stream, "Insn %d: giv reg %d",
4610 INSN_UID (insn), REGNO (dest_reg));
4611 else
4612 fprintf (loop_dump_stream, "Insn %d: dest address",
4613 INSN_UID (insn));
4614
4615 fprintf (loop_dump_stream, " src reg %d benefit %d",
4616 REGNO (src_reg), v->benefit);
4617 fprintf (loop_dump_stream, " used %d lifetime %d",
4618 v->times_used, v->lifetime);
4619
4620 if (v->replaceable)
4621 fprintf (loop_dump_stream, " replaceable");
4622
4623 if (GET_CODE (mult_val) == CONST_INT)
4624 fprintf (loop_dump_stream, " mult %d",
4625 INTVAL (mult_val));
4626 else
4627 {
4628 fprintf (loop_dump_stream, " mult ");
4629 print_rtl (loop_dump_stream, mult_val);
4630 }
4631
4632 if (GET_CODE (add_val) == CONST_INT)
4633 fprintf (loop_dump_stream, " add %d",
4634 INTVAL (add_val));
4635 else
4636 {
4637 fprintf (loop_dump_stream, " add ");
4638 print_rtl (loop_dump_stream, add_val);
4639 }
4640 }
4641
4642 if (loop_dump_stream)
4643 fprintf (loop_dump_stream, "\n");
4644
4645 }
4646
4647
4648 /* All this does is determine whether a giv can be made replaceable because
4649 its final value can be calculated. This code can not be part of record_giv
4650 above, because final_giv_value requires that the number of loop iterations
4651 be known, and that can not be accurately calculated until after all givs
4652 have been identified. */
4653
4654 static void
4655 check_final_value (v, loop_start, loop_end)
4656 struct induction *v;
4657 rtx loop_start, loop_end;
4658 {
4659 struct iv_class *bl;
4660 rtx final_value = 0;
4661
4662 bl = reg_biv_class[REGNO (v->src_reg)];
4663
4664 /* DEST_ADDR givs will never reach here, because they are always marked
4665 replaceable above in record_giv. */
4666
4667 /* The giv can be replaced outright by the reduced register only if all
4668 of the following conditions are true:
4669 - the insn that sets the giv is always executed on any iteration
4670 on which the giv is used at all
4671 (there are two ways to deduce this:
4672 either the insn is executed on every iteration,
4673 or all uses follow that insn in the same basic block),
4674 - its final value can be calculated (this condition is different
4675 than the one above in record_giv)
4676 - no assignments to the biv occur during the giv's lifetime. */
4677
4678 #if 0
4679 /* This is only called now when replaceable is known to be false. */
4680 /* Clear replaceable, so that it won't confuse final_giv_value. */
4681 v->replaceable = 0;
4682 #endif
4683
4684 if ((final_value = final_giv_value (v, loop_start, loop_end))
4685 && (v->always_computable || last_use_this_basic_block (v->dest_reg, v->insn)))
4686 {
4687 int biv_increment_seen = 0;
4688 rtx p = v->insn;
4689 rtx last_giv_use;
4690
4691 v->replaceable = 1;
4692
4693 /* When trying to determine whether or not a biv increment occurs
4694 during the lifetime of the giv, we can ignore uses of the variable
4695 outside the loop because final_value is true. Hence we can not
4696 use regno_last_uid and regno_first_uid as above in record_giv. */
4697
4698 /* Search the loop to determine whether any assignments to the
4699 biv occur during the giv's lifetime. Start with the insn
4700 that sets the giv, and search around the loop until we come
4701 back to that insn again.
4702
4703 Also fail if there is a jump within the giv's lifetime that jumps
4704 to somewhere outside the lifetime but still within the loop. This
4705 catches spaghetti code where the execution order is not linear, and
4706 hence the above test fails. Here we assume that the giv lifetime
4707 does not extend from one iteration of the loop to the next, so as
4708 to make the test easier. Since the lifetime isn't known yet,
4709 this requires two loops. See also record_giv above. */
4710
4711 last_giv_use = v->insn;
4712
4713 while (1)
4714 {
4715 p = NEXT_INSN (p);
4716 if (p == loop_end)
4717 p = NEXT_INSN (loop_start);
4718 if (p == v->insn)
4719 break;
4720
4721 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
4722 || GET_CODE (p) == CALL_INSN)
4723 {
4724 if (biv_increment_seen)
4725 {
4726 if (reg_mentioned_p (v->dest_reg, PATTERN (p)))
4727 {
4728 v->replaceable = 0;
4729 v->not_replaceable = 1;
4730 break;
4731 }
4732 }
4733 else if (GET_CODE (PATTERN (p)) == SET
4734 && SET_DEST (PATTERN (p)) == v->src_reg)
4735 biv_increment_seen = 1;
4736 else if (reg_mentioned_p (v->dest_reg, PATTERN (p)))
4737 last_giv_use = p;
4738 }
4739 }
4740
4741 /* Now that the lifetime of the giv is known, check for branches
4742 from within the lifetime to outside the lifetime if it is still
4743 replaceable. */
4744
4745 if (v->replaceable)
4746 {
4747 p = v->insn;
4748 while (1)
4749 {
4750 p = NEXT_INSN (p);
4751 if (p == loop_end)
4752 p = NEXT_INSN (loop_start);
4753 if (p == last_giv_use)
4754 break;
4755
4756 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p)
4757 && LABEL_NAME (JUMP_LABEL (p))
4758 && ((INSN_LUID (JUMP_LABEL (p)) < INSN_LUID (v->insn)
4759 && INSN_LUID (JUMP_LABEL (p)) > INSN_LUID (loop_start))
4760 || (INSN_LUID (JUMP_LABEL (p)) > INSN_LUID (last_giv_use)
4761 && INSN_LUID (JUMP_LABEL (p)) < INSN_LUID (loop_end))))
4762 {
4763 v->replaceable = 0;
4764 v->not_replaceable = 1;
4765
4766 if (loop_dump_stream)
4767 fprintf (loop_dump_stream,
4768 "Found branch outside giv lifetime.\n");
4769
4770 break;
4771 }
4772 }
4773 }
4774
4775 /* If it is replaceable, then save the final value. */
4776 if (v->replaceable)
4777 v->final_value = final_value;
4778 }
4779
4780 if (loop_dump_stream && v->replaceable)
4781 fprintf (loop_dump_stream, "Insn %d: giv reg %d final_value replaceable\n",
4782 INSN_UID (v->insn), REGNO (v->dest_reg));
4783 }
4784 \f
4785 /* Update the status of whether a giv can derive other givs.
4786
4787 We need to do something special if there is or may be an update to the biv
4788 between the time the giv is defined and the time it is used to derive
4789 another giv.
4790
4791 In addition, a giv that is only conditionally set is not allowed to
4792 derive another giv once a label has been passed.
4793
4794 The cases we look at are when a label or an update to a biv is passed. */
4795
4796 static void
4797 update_giv_derive (p)
4798 rtx p;
4799 {
4800 struct iv_class *bl;
4801 struct induction *biv, *giv;
4802 rtx tem;
4803 int dummy;
4804
4805 /* Search all IV classes, then all bivs, and finally all givs.
4806
4807 There are three cases we are concerned with. First we have the situation
4808 of a giv that is only updated conditionally. In that case, it may not
4809 derive any givs after a label is passed.
4810
4811 The second case is when a biv update occurs, or may occur, after the
4812 definition of a giv. For certain biv updates (see below) that are
4813 known to occur between the giv definition and use, we can adjust the
4814 giv definition. For others, or when the biv update is conditional,
4815 we must prevent the giv from deriving any other givs. There are two
4816 sub-cases within this case.
4817
4818 If this is a label, we are concerned with any biv update that is done
4819 conditionally, since it may be done after the giv is defined followed by
4820 a branch here (actually, we need to pass both a jump and a label, but
4821 this extra tracking doesn't seem worth it).
4822
4823 If this is a jump, we are concerned about any biv update that may be
4824 executed multiple times. We are actually only concerned about
4825 backward jumps, but it is probably not worth performing the test
4826 on the jump again here.
4827
4828 If this is a biv update, we must adjust the giv status to show that a
4829 subsequent biv update was performed. If this adjustment cannot be done,
4830 the giv cannot derive further givs. */
4831
4832 for (bl = loop_iv_list; bl; bl = bl->next)
4833 for (biv = bl->biv; biv; biv = biv->next_iv)
4834 if (GET_CODE (p) == CODE_LABEL || GET_CODE (p) == JUMP_INSN
4835 || biv->insn == p)
4836 {
4837 for (giv = bl->giv; giv; giv = giv->next_iv)
4838 {
4839 /* If cant_derive is already true, there is no point in
4840 checking all of these conditions again. */
4841 if (giv->cant_derive)
4842 continue;
4843
4844 /* If this giv is conditionally set and we have passed a label,
4845 it cannot derive anything. */
4846 if (GET_CODE (p) == CODE_LABEL && ! giv->always_computable)
4847 giv->cant_derive = 1;
4848
4849 /* Skip givs that have mult_val == 0, since
4850 they are really invariants. Also skip those that are
4851 replaceable, since we know their lifetime doesn't contain
4852 any biv update. */
4853 else if (giv->mult_val == const0_rtx || giv->replaceable)
4854 continue;
4855
4856 /* The only way we can allow this giv to derive another
4857 is if this is a biv increment and we can form the product
4858 of biv->add_val and giv->mult_val. In this case, we will
4859 be able to compute a compensation. */
4860 else if (biv->insn == p)
4861 {
4862 tem = 0;
4863
4864 if (biv->mult_val == const1_rtx)
4865 tem = simplify_giv_expr (gen_rtx (MULT, giv->mode,
4866 biv->add_val,
4867 giv->mult_val),
4868 &dummy);
4869
4870 if (tem && giv->derive_adjustment)
4871 tem = simplify_giv_expr (gen_rtx (PLUS, giv->mode, tem,
4872 giv->derive_adjustment),
4873 &dummy);
4874 if (tem)
4875 giv->derive_adjustment = tem;
4876 else
4877 giv->cant_derive = 1;
4878 }
4879 else if ((GET_CODE (p) == CODE_LABEL && ! biv->always_computable)
4880 || (GET_CODE (p) == JUMP_INSN && biv->maybe_multiple))
4881 giv->cant_derive = 1;
4882 }
4883 }
4884 }
4885 \f
4886 /* Check whether an insn is an increment legitimate for a basic induction var.
4887 X is the source of insn P, or a part of it.
4888 MODE is the mode in which X should be interpreted.
4889
4890 DEST_REG is the putative biv, also the destination of the insn.
4891 We accept patterns of these forms:
4892 REG = REG + INVARIANT (includes REG = REG - CONSTANT)
4893 REG = INVARIANT + REG
4894
4895 If X is suitable, we return 1, set *MULT_VAL to CONST1_RTX,
4896 and store the additive term into *INC_VAL.
4897
4898 If X is an assignment of an invariant into DEST_REG, we set
4899 *MULT_VAL to CONST0_RTX, and store the invariant into *INC_VAL.
4900
4901 We also want to detect a BIV when it corresponds to a variable
4902 whose mode was promoted via PROMOTED_MODE. In that case, an increment
4903 of the variable may be a PLUS that adds a SUBREG of that variable to
4904 an invariant and then sign- or zero-extends the result of the PLUS
4905 into the variable.
4906
4907 Most GIVs in such cases will be in the promoted mode, since that is the
4908 probably the natural computation mode (and almost certainly the mode
4909 used for addresses) on the machine. So we view the pseudo-reg containing
4910 the variable as the BIV, as if it were simply incremented.
4911
4912 Note that treating the entire pseudo as a BIV will result in making
4913 simple increments to any GIVs based on it. However, if the variable
4914 overflows in its declared mode but not its promoted mode, the result will
4915 be incorrect. This is acceptable if the variable is signed, since
4916 overflows in such cases are undefined, but not if it is unsigned, since
4917 those overflows are defined. So we only check for SIGN_EXTEND and
4918 not ZERO_EXTEND.
4919
4920 If we cannot find a biv, we return 0. */
4921
4922 static int
4923 basic_induction_var (x, mode, dest_reg, p, inc_val, mult_val)
4924 register rtx x;
4925 enum machine_mode mode;
4926 rtx p;
4927 rtx dest_reg;
4928 rtx *inc_val;
4929 rtx *mult_val;
4930 {
4931 register enum rtx_code code;
4932 rtx arg;
4933 rtx insn, set = 0;
4934
4935 code = GET_CODE (x);
4936 switch (code)
4937 {
4938 case PLUS:
4939 if (XEXP (x, 0) == dest_reg
4940 || (GET_CODE (XEXP (x, 0)) == SUBREG
4941 && SUBREG_PROMOTED_VAR_P (XEXP (x, 0))
4942 && SUBREG_REG (XEXP (x, 0)) == dest_reg))
4943 arg = XEXP (x, 1);
4944 else if (XEXP (x, 1) == dest_reg
4945 || (GET_CODE (XEXP (x, 1)) == SUBREG
4946 && SUBREG_PROMOTED_VAR_P (XEXP (x, 1))
4947 && SUBREG_REG (XEXP (x, 1)) == dest_reg))
4948 arg = XEXP (x, 0);
4949 else
4950 return 0;
4951
4952 if (invariant_p (arg) != 1)
4953 return 0;
4954
4955 *inc_val = convert_modes (GET_MODE (dest_reg), GET_MODE (x), arg, 0);
4956 *mult_val = const1_rtx;
4957 return 1;
4958
4959 case SUBREG:
4960 /* If this is a SUBREG for a promoted variable, check the inner
4961 value. */
4962 if (SUBREG_PROMOTED_VAR_P (x))
4963 return basic_induction_var (SUBREG_REG (x), GET_MODE (SUBREG_REG (x)),
4964 dest_reg, p, inc_val, mult_val);
4965
4966 case REG:
4967 /* If this register is assigned in the previous insn, look at its
4968 source, but don't go outside the loop or past a label. */
4969
4970 for (insn = PREV_INSN (p);
4971 (insn && GET_CODE (insn) == NOTE
4972 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_BEG);
4973 insn = PREV_INSN (insn))
4974 ;
4975
4976 if (insn)
4977 set = single_set (insn);
4978
4979 if (set != 0
4980 && (SET_DEST (set) == x
4981 || (GET_CODE (SET_DEST (set)) == SUBREG
4982 && (GET_MODE_SIZE (GET_MODE (SET_DEST (set)))
4983 <= UNITS_PER_WORD)
4984 && SUBREG_REG (SET_DEST (set)) == x)))
4985 return basic_induction_var (SET_SRC (set),
4986 (GET_MODE (SET_SRC (set)) == VOIDmode
4987 ? GET_MODE (x)
4988 : GET_MODE (SET_SRC (set))),
4989 dest_reg, insn,
4990 inc_val, mult_val);
4991 /* ... fall through ... */
4992
4993 /* Can accept constant setting of biv only when inside inner most loop.
4994 Otherwise, a biv of an inner loop may be incorrectly recognized
4995 as a biv of the outer loop,
4996 causing code to be moved INTO the inner loop. */
4997 case MEM:
4998 if (invariant_p (x) != 1)
4999 return 0;
5000 case CONST_INT:
5001 case SYMBOL_REF:
5002 case CONST:
5003 if (loops_enclosed == 1)
5004 {
5005 /* Possible bug here? Perhaps we don't know the mode of X. */
5006 *inc_val = convert_modes (GET_MODE (dest_reg), mode, x, 0);
5007 *mult_val = const0_rtx;
5008 return 1;
5009 }
5010 else
5011 return 0;
5012
5013 case SIGN_EXTEND:
5014 return basic_induction_var (XEXP (x, 0), GET_MODE (XEXP (x, 0)),
5015 dest_reg, p, inc_val, mult_val);
5016 case ASHIFTRT:
5017 /* Similar, since this can be a sign extension. */
5018 for (insn = PREV_INSN (p);
5019 (insn && GET_CODE (insn) == NOTE
5020 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_BEG);
5021 insn = PREV_INSN (insn))
5022 ;
5023
5024 if (insn)
5025 set = single_set (insn);
5026
5027 if (set && SET_DEST (set) == XEXP (x, 0)
5028 && GET_CODE (XEXP (x, 1)) == CONST_INT
5029 && INTVAL (XEXP (x, 1)) >= 0
5030 && GET_CODE (SET_SRC (set)) == ASHIFT
5031 && XEXP (x, 1) == XEXP (SET_SRC (set), 1))
5032 return basic_induction_var (XEXP (SET_SRC (set), 0),
5033 GET_MODE (XEXP (x, 0)),
5034 dest_reg, insn, inc_val, mult_val);
5035 return 0;
5036
5037 default:
5038 return 0;
5039 }
5040 }
5041 \f
5042 /* A general induction variable (giv) is any quantity that is a linear
5043 function of a basic induction variable,
5044 i.e. giv = biv * mult_val + add_val.
5045 The coefficients can be any loop invariant quantity.
5046 A giv need not be computed directly from the biv;
5047 it can be computed by way of other givs. */
5048
5049 /* Determine whether X computes a giv.
5050 If it does, return a nonzero value
5051 which is the benefit from eliminating the computation of X;
5052 set *SRC_REG to the register of the biv that it is computed from;
5053 set *ADD_VAL and *MULT_VAL to the coefficients,
5054 such that the value of X is biv * mult + add; */
5055
5056 static int
5057 general_induction_var (x, src_reg, add_val, mult_val)
5058 rtx x;
5059 rtx *src_reg;
5060 rtx *add_val;
5061 rtx *mult_val;
5062 {
5063 rtx orig_x = x;
5064 int benefit = 0;
5065 char *storage;
5066
5067 /* If this is an invariant, forget it, it isn't a giv. */
5068 if (invariant_p (x) == 1)
5069 return 0;
5070
5071 /* See if the expression could be a giv and get its form.
5072 Mark our place on the obstack in case we don't find a giv. */
5073 storage = (char *) oballoc (0);
5074 x = simplify_giv_expr (x, &benefit);
5075 if (x == 0)
5076 {
5077 obfree (storage);
5078 return 0;
5079 }
5080
5081 switch (GET_CODE (x))
5082 {
5083 case USE:
5084 case CONST_INT:
5085 /* Since this is now an invariant and wasn't before, it must be a giv
5086 with MULT_VAL == 0. It doesn't matter which BIV we associate this
5087 with. */
5088 *src_reg = loop_iv_list->biv->dest_reg;
5089 *mult_val = const0_rtx;
5090 *add_val = x;
5091 break;
5092
5093 case REG:
5094 /* This is equivalent to a BIV. */
5095 *src_reg = x;
5096 *mult_val = const1_rtx;
5097 *add_val = const0_rtx;
5098 break;
5099
5100 case PLUS:
5101 /* Either (plus (biv) (invar)) or
5102 (plus (mult (biv) (invar_1)) (invar_2)). */
5103 if (GET_CODE (XEXP (x, 0)) == MULT)
5104 {
5105 *src_reg = XEXP (XEXP (x, 0), 0);
5106 *mult_val = XEXP (XEXP (x, 0), 1);
5107 }
5108 else
5109 {
5110 *src_reg = XEXP (x, 0);
5111 *mult_val = const1_rtx;
5112 }
5113 *add_val = XEXP (x, 1);
5114 break;
5115
5116 case MULT:
5117 /* ADD_VAL is zero. */
5118 *src_reg = XEXP (x, 0);
5119 *mult_val = XEXP (x, 1);
5120 *add_val = const0_rtx;
5121 break;
5122
5123 default:
5124 abort ();
5125 }
5126
5127 /* Remove any enclosing USE from ADD_VAL and MULT_VAL (there will be
5128 unless they are CONST_INT). */
5129 if (GET_CODE (*add_val) == USE)
5130 *add_val = XEXP (*add_val, 0);
5131 if (GET_CODE (*mult_val) == USE)
5132 *mult_val = XEXP (*mult_val, 0);
5133
5134 benefit += rtx_cost (orig_x, SET);
5135
5136 /* Always return some benefit if this is a giv so it will be detected
5137 as such. This allows elimination of bivs that might otherwise
5138 not be eliminated. */
5139 return benefit == 0 ? 1 : benefit;
5140 }
5141 \f
5142 /* Given an expression, X, try to form it as a linear function of a biv.
5143 We will canonicalize it to be of the form
5144 (plus (mult (BIV) (invar_1))
5145 (invar_2))
5146 with possible degeneracies.
5147
5148 The invariant expressions must each be of a form that can be used as a
5149 machine operand. We surround then with a USE rtx (a hack, but localized
5150 and certainly unambiguous!) if not a CONST_INT for simplicity in this
5151 routine; it is the caller's responsibility to strip them.
5152
5153 If no such canonicalization is possible (i.e., two biv's are used or an
5154 expression that is neither invariant nor a biv or giv), this routine
5155 returns 0.
5156
5157 For a non-zero return, the result will have a code of CONST_INT, USE,
5158 REG (for a BIV), PLUS, or MULT. No other codes will occur.
5159
5160 *BENEFIT will be incremented by the benefit of any sub-giv encountered. */
5161
5162 static rtx
5163 simplify_giv_expr (x, benefit)
5164 rtx x;
5165 int *benefit;
5166 {
5167 enum machine_mode mode = GET_MODE (x);
5168 rtx arg0, arg1;
5169 rtx tem;
5170
5171 /* If this is not an integer mode, or if we cannot do arithmetic in this
5172 mode, this can't be a giv. */
5173 if (mode != VOIDmode
5174 && (GET_MODE_CLASS (mode) != MODE_INT
5175 || GET_MODE_BITSIZE (mode) > HOST_BITS_PER_WIDE_INT))
5176 return 0;
5177
5178 switch (GET_CODE (x))
5179 {
5180 case PLUS:
5181 arg0 = simplify_giv_expr (XEXP (x, 0), benefit);
5182 arg1 = simplify_giv_expr (XEXP (x, 1), benefit);
5183 if (arg0 == 0 || arg1 == 0)
5184 return 0;
5185
5186 /* Put constant last, CONST_INT last if both constant. */
5187 if ((GET_CODE (arg0) == USE
5188 || GET_CODE (arg0) == CONST_INT)
5189 && GET_CODE (arg1) != CONST_INT)
5190 tem = arg0, arg0 = arg1, arg1 = tem;
5191
5192 /* Handle addition of zero, then addition of an invariant. */
5193 if (arg1 == const0_rtx)
5194 return arg0;
5195 else if (GET_CODE (arg1) == CONST_INT || GET_CODE (arg1) == USE)
5196 switch (GET_CODE (arg0))
5197 {
5198 case CONST_INT:
5199 case USE:
5200 /* Both invariant. Only valid if sum is machine operand.
5201 First strip off possible USE on first operand. */
5202 if (GET_CODE (arg0) == USE)
5203 arg0 = XEXP (arg0, 0);
5204
5205 tem = 0;
5206 if (CONSTANT_P (arg0) && GET_CODE (arg1) == CONST_INT)
5207 {
5208 tem = plus_constant (arg0, INTVAL (arg1));
5209 if (GET_CODE (tem) != CONST_INT)
5210 tem = gen_rtx (USE, mode, tem);
5211 }
5212
5213 return tem;
5214
5215 case REG:
5216 case MULT:
5217 /* biv + invar or mult + invar. Return sum. */
5218 return gen_rtx (PLUS, mode, arg0, arg1);
5219
5220 case PLUS:
5221 /* (a + invar_1) + invar_2. Associate. */
5222 return simplify_giv_expr (gen_rtx (PLUS, mode,
5223 XEXP (arg0, 0),
5224 gen_rtx (PLUS, mode,
5225 XEXP (arg0, 1), arg1)),
5226 benefit);
5227
5228 default:
5229 abort ();
5230 }
5231
5232 /* Each argument must be either REG, PLUS, or MULT. Convert REG to
5233 MULT to reduce cases. */
5234 if (GET_CODE (arg0) == REG)
5235 arg0 = gen_rtx (MULT, mode, arg0, const1_rtx);
5236 if (GET_CODE (arg1) == REG)
5237 arg1 = gen_rtx (MULT, mode, arg1, const1_rtx);
5238
5239 /* Now have PLUS + PLUS, PLUS + MULT, MULT + PLUS, or MULT + MULT.
5240 Put a MULT first, leaving PLUS + PLUS, MULT + PLUS, or MULT + MULT.
5241 Recurse to associate the second PLUS. */
5242 if (GET_CODE (arg1) == MULT)
5243 tem = arg0, arg0 = arg1, arg1 = tem;
5244
5245 if (GET_CODE (arg1) == PLUS)
5246 return simplify_giv_expr (gen_rtx (PLUS, mode,
5247 gen_rtx (PLUS, mode,
5248 arg0, XEXP (arg1, 0)),
5249 XEXP (arg1, 1)),
5250 benefit);
5251
5252 /* Now must have MULT + MULT. Distribute if same biv, else not giv. */
5253 if (GET_CODE (arg0) != MULT || GET_CODE (arg1) != MULT)
5254 abort ();
5255
5256 if (XEXP (arg0, 0) != XEXP (arg1, 0))
5257 return 0;
5258
5259 return simplify_giv_expr (gen_rtx (MULT, mode,
5260 XEXP (arg0, 0),
5261 gen_rtx (PLUS, mode,
5262 XEXP (arg0, 1),
5263 XEXP (arg1, 1))),
5264 benefit);
5265
5266 case MINUS:
5267 /* Handle "a - b" as "a + b * (-1)". */
5268 return simplify_giv_expr (gen_rtx (PLUS, mode,
5269 XEXP (x, 0),
5270 gen_rtx (MULT, mode,
5271 XEXP (x, 1), constm1_rtx)),
5272 benefit);
5273
5274 case MULT:
5275 arg0 = simplify_giv_expr (XEXP (x, 0), benefit);
5276 arg1 = simplify_giv_expr (XEXP (x, 1), benefit);
5277 if (arg0 == 0 || arg1 == 0)
5278 return 0;
5279
5280 /* Put constant last, CONST_INT last if both constant. */
5281 if ((GET_CODE (arg0) == USE || GET_CODE (arg0) == CONST_INT)
5282 && GET_CODE (arg1) != CONST_INT)
5283 tem = arg0, arg0 = arg1, arg1 = tem;
5284
5285 /* If second argument is not now constant, not giv. */
5286 if (GET_CODE (arg1) != USE && GET_CODE (arg1) != CONST_INT)
5287 return 0;
5288
5289 /* Handle multiply by 0 or 1. */
5290 if (arg1 == const0_rtx)
5291 return const0_rtx;
5292
5293 else if (arg1 == const1_rtx)
5294 return arg0;
5295
5296 switch (GET_CODE (arg0))
5297 {
5298 case REG:
5299 /* biv * invar. Done. */
5300 return gen_rtx (MULT, mode, arg0, arg1);
5301
5302 case CONST_INT:
5303 /* Product of two constants. */
5304 return GEN_INT (INTVAL (arg0) * INTVAL (arg1));
5305
5306 case USE:
5307 /* invar * invar. Not giv. */
5308 return 0;
5309
5310 case MULT:
5311 /* (a * invar_1) * invar_2. Associate. */
5312 return simplify_giv_expr (gen_rtx (MULT, mode,
5313 XEXP (arg0, 0),
5314 gen_rtx (MULT, mode,
5315 XEXP (arg0, 1), arg1)),
5316 benefit);
5317
5318 case PLUS:
5319 /* (a + invar_1) * invar_2. Distribute. */
5320 return simplify_giv_expr (gen_rtx (PLUS, mode,
5321 gen_rtx (MULT, mode,
5322 XEXP (arg0, 0), arg1),
5323 gen_rtx (MULT, mode,
5324 XEXP (arg0, 1), arg1)),
5325 benefit);
5326
5327 default:
5328 abort ();
5329 }
5330
5331 case ASHIFT:
5332 /* Shift by constant is multiply by power of two. */
5333 if (GET_CODE (XEXP (x, 1)) != CONST_INT)
5334 return 0;
5335
5336 return simplify_giv_expr (gen_rtx (MULT, mode,
5337 XEXP (x, 0),
5338 GEN_INT ((HOST_WIDE_INT) 1
5339 << INTVAL (XEXP (x, 1)))),
5340 benefit);
5341
5342 case NEG:
5343 /* "-a" is "a * (-1)" */
5344 return simplify_giv_expr (gen_rtx (MULT, mode, XEXP (x, 0), constm1_rtx),
5345 benefit);
5346
5347 case NOT:
5348 /* "~a" is "-a - 1". Silly, but easy. */
5349 return simplify_giv_expr (gen_rtx (MINUS, mode,
5350 gen_rtx (NEG, mode, XEXP (x, 0)),
5351 const1_rtx),
5352 benefit);
5353
5354 case USE:
5355 /* Already in proper form for invariant. */
5356 return x;
5357
5358 case REG:
5359 /* If this is a new register, we can't deal with it. */
5360 if (REGNO (x) >= max_reg_before_loop)
5361 return 0;
5362
5363 /* Check for biv or giv. */
5364 switch (reg_iv_type[REGNO (x)])
5365 {
5366 case BASIC_INDUCT:
5367 return x;
5368 case GENERAL_INDUCT:
5369 {
5370 struct induction *v = reg_iv_info[REGNO (x)];
5371
5372 /* Form expression from giv and add benefit. Ensure this giv
5373 can derive another and subtract any needed adjustment if so. */
5374 *benefit += v->benefit;
5375 if (v->cant_derive)
5376 return 0;
5377
5378 tem = gen_rtx (PLUS, mode, gen_rtx (MULT, mode,
5379 v->src_reg, v->mult_val),
5380 v->add_val);
5381 if (v->derive_adjustment)
5382 tem = gen_rtx (MINUS, mode, tem, v->derive_adjustment);
5383 return simplify_giv_expr (tem, benefit);
5384 }
5385 }
5386
5387 /* Fall through to general case. */
5388 default:
5389 /* If invariant, return as USE (unless CONST_INT).
5390 Otherwise, not giv. */
5391 if (GET_CODE (x) == USE)
5392 x = XEXP (x, 0);
5393
5394 if (invariant_p (x) == 1)
5395 {
5396 if (GET_CODE (x) == CONST_INT)
5397 return x;
5398 else
5399 return gen_rtx (USE, mode, x);
5400 }
5401 else
5402 return 0;
5403 }
5404 }
5405 \f
5406 /* Help detect a giv that is calculated by several consecutive insns;
5407 for example,
5408 giv = biv * M
5409 giv = giv + A
5410 The caller has already identified the first insn P as having a giv as dest;
5411 we check that all other insns that set the same register follow
5412 immediately after P, that they alter nothing else,
5413 and that the result of the last is still a giv.
5414
5415 The value is 0 if the reg set in P is not really a giv.
5416 Otherwise, the value is the amount gained by eliminating
5417 all the consecutive insns that compute the value.
5418
5419 FIRST_BENEFIT is the amount gained by eliminating the first insn, P.
5420 SRC_REG is the reg of the biv; DEST_REG is the reg of the giv.
5421
5422 The coefficients of the ultimate giv value are stored in
5423 *MULT_VAL and *ADD_VAL. */
5424
5425 static int
5426 consec_sets_giv (first_benefit, p, src_reg, dest_reg,
5427 add_val, mult_val)
5428 int first_benefit;
5429 rtx p;
5430 rtx src_reg;
5431 rtx dest_reg;
5432 rtx *add_val;
5433 rtx *mult_val;
5434 {
5435 int count;
5436 enum rtx_code code;
5437 int benefit;
5438 rtx temp;
5439 rtx set;
5440
5441 /* Indicate that this is a giv so that we can update the value produced in
5442 each insn of the multi-insn sequence.
5443
5444 This induction structure will be used only by the call to
5445 general_induction_var below, so we can allocate it on our stack.
5446 If this is a giv, our caller will replace the induct var entry with
5447 a new induction structure. */
5448 struct induction *v
5449 = (struct induction *) alloca (sizeof (struct induction));
5450 v->src_reg = src_reg;
5451 v->mult_val = *mult_val;
5452 v->add_val = *add_val;
5453 v->benefit = first_benefit;
5454 v->cant_derive = 0;
5455 v->derive_adjustment = 0;
5456
5457 reg_iv_type[REGNO (dest_reg)] = GENERAL_INDUCT;
5458 reg_iv_info[REGNO (dest_reg)] = v;
5459
5460 count = n_times_set[REGNO (dest_reg)] - 1;
5461
5462 while (count > 0)
5463 {
5464 p = NEXT_INSN (p);
5465 code = GET_CODE (p);
5466
5467 /* If libcall, skip to end of call sequence. */
5468 if (code == INSN && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
5469 p = XEXP (temp, 0);
5470
5471 if (code == INSN
5472 && (set = single_set (p))
5473 && GET_CODE (SET_DEST (set)) == REG
5474 && SET_DEST (set) == dest_reg
5475 && ((benefit = general_induction_var (SET_SRC (set), &src_reg,
5476 add_val, mult_val))
5477 /* Giv created by equivalent expression. */
5478 || ((temp = find_reg_note (p, REG_EQUAL, NULL_RTX))
5479 && (benefit = general_induction_var (XEXP (temp, 0), &src_reg,
5480 add_val, mult_val))))
5481 && src_reg == v->src_reg)
5482 {
5483 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
5484 benefit += libcall_benefit (p);
5485
5486 count--;
5487 v->mult_val = *mult_val;
5488 v->add_val = *add_val;
5489 v->benefit = benefit;
5490 }
5491 else if (code != NOTE)
5492 {
5493 /* Allow insns that set something other than this giv to a
5494 constant. Such insns are needed on machines which cannot
5495 include long constants and should not disqualify a giv. */
5496 if (code == INSN
5497 && (set = single_set (p))
5498 && SET_DEST (set) != dest_reg
5499 && CONSTANT_P (SET_SRC (set)))
5500 continue;
5501
5502 reg_iv_type[REGNO (dest_reg)] = UNKNOWN_INDUCT;
5503 return 0;
5504 }
5505 }
5506
5507 return v->benefit;
5508 }
5509 \f
5510 /* Return an rtx, if any, that expresses giv G2 as a function of the register
5511 represented by G1. If no such expression can be found, or it is clear that
5512 it cannot possibly be a valid address, 0 is returned.
5513
5514 To perform the computation, we note that
5515 G1 = a * v + b and
5516 G2 = c * v + d
5517 where `v' is the biv.
5518
5519 So G2 = (c/a) * G1 + (d - b*c/a) */
5520
5521 #ifdef ADDRESS_COST
5522 static rtx
5523 express_from (g1, g2)
5524 struct induction *g1, *g2;
5525 {
5526 rtx mult, add;
5527
5528 /* The value that G1 will be multiplied by must be a constant integer. Also,
5529 the only chance we have of getting a valid address is if b*c/a (see above
5530 for notation) is also an integer. */
5531 if (GET_CODE (g1->mult_val) != CONST_INT
5532 || GET_CODE (g2->mult_val) != CONST_INT
5533 || GET_CODE (g1->add_val) != CONST_INT
5534 || g1->mult_val == const0_rtx
5535 || INTVAL (g2->mult_val) % INTVAL (g1->mult_val) != 0)
5536 return 0;
5537
5538 mult = GEN_INT (INTVAL (g2->mult_val) / INTVAL (g1->mult_val));
5539 add = plus_constant (g2->add_val, - INTVAL (g1->add_val) * INTVAL (mult));
5540
5541 /* Form simplified final result. */
5542 if (mult == const0_rtx)
5543 return add;
5544 else if (mult == const1_rtx)
5545 mult = g1->dest_reg;
5546 else
5547 mult = gen_rtx (MULT, g2->mode, g1->dest_reg, mult);
5548
5549 if (add == const0_rtx)
5550 return mult;
5551 else
5552 return gen_rtx (PLUS, g2->mode, mult, add);
5553 }
5554 #endif
5555 \f
5556 /* Return 1 if giv G2 can be combined with G1. This means that G2 can use
5557 (either directly or via an address expression) a register used to represent
5558 G1. Set g2->new_reg to a represtation of G1 (normally just
5559 g1->dest_reg). */
5560
5561 static int
5562 combine_givs_p (g1, g2)
5563 struct induction *g1, *g2;
5564 {
5565 rtx tem;
5566
5567 /* If these givs are identical, they can be combined. */
5568 if (rtx_equal_p (g1->mult_val, g2->mult_val)
5569 && rtx_equal_p (g1->add_val, g2->add_val))
5570 {
5571 g2->new_reg = g1->dest_reg;
5572 return 1;
5573 }
5574
5575 #ifdef ADDRESS_COST
5576 /* If G2 can be expressed as a function of G1 and that function is valid
5577 as an address and no more expensive than using a register for G2,
5578 the expression of G2 in terms of G1 can be used. */
5579 if (g2->giv_type == DEST_ADDR
5580 && (tem = express_from (g1, g2)) != 0
5581 && memory_address_p (g2->mem_mode, tem)
5582 && ADDRESS_COST (tem) <= ADDRESS_COST (*g2->location))
5583 {
5584 g2->new_reg = tem;
5585 return 1;
5586 }
5587 #endif
5588
5589 return 0;
5590 }
5591 \f
5592 #ifdef GIV_SORT_CRITERION
5593 /* Compare two givs and sort the most desirable one for combinations first.
5594 This is used only in one qsort call below. */
5595
5596 static int
5597 giv_sort (x, y)
5598 struct induction **x, **y;
5599 {
5600 GIV_SORT_CRITERION (*x, *y);
5601
5602 return 0;
5603 }
5604 #endif
5605
5606 /* Check all pairs of givs for iv_class BL and see if any can be combined with
5607 any other. If so, point SAME to the giv combined with and set NEW_REG to
5608 be an expression (in terms of the other giv's DEST_REG) equivalent to the
5609 giv. Also, update BENEFIT and related fields for cost/benefit analysis. */
5610
5611 static void
5612 combine_givs (bl)
5613 struct iv_class *bl;
5614 {
5615 struct induction *g1, *g2, **giv_array, *temp_iv;
5616 int i, j, giv_count, pass;
5617
5618 /* Count givs, because bl->giv_count is incorrect here. */
5619 giv_count = 0;
5620 for (g1 = bl->giv; g1; g1 = g1->next_iv)
5621 giv_count++;
5622
5623 giv_array
5624 = (struct induction **) alloca (giv_count * sizeof (struct induction *));
5625 i = 0;
5626 for (g1 = bl->giv; g1; g1 = g1->next_iv)
5627 giv_array[i++] = g1;
5628
5629 #ifdef GIV_SORT_CRITERION
5630 /* Sort the givs if GIV_SORT_CRITERION is defined.
5631 This is usually defined for processors which lack
5632 negative register offsets so more givs may be combined. */
5633
5634 if (loop_dump_stream)
5635 fprintf (loop_dump_stream, "%d givs counted, sorting...\n", giv_count);
5636
5637 qsort (giv_array, giv_count, sizeof (struct induction *), giv_sort);
5638 #endif
5639
5640 for (i = 0; i < giv_count; i++)
5641 {
5642 g1 = giv_array[i];
5643 for (pass = 0; pass <= 1; pass++)
5644 for (j = 0; j < giv_count; j++)
5645 {
5646 g2 = giv_array[j];
5647 if (g1 != g2
5648 /* First try to combine with replaceable givs, then all givs. */
5649 && (g1->replaceable || pass == 1)
5650 /* If either has already been combined or is to be ignored, can't
5651 combine. */
5652 && ! g1->ignore && ! g2->ignore && ! g1->same && ! g2->same
5653 /* If something has been based on G2, G2 cannot itself be based
5654 on something else. */
5655 && ! g2->combined_with
5656 && combine_givs_p (g1, g2))
5657 {
5658 /* g2->new_reg set by `combine_givs_p' */
5659 g2->same = g1;
5660 g1->combined_with = 1;
5661 g1->benefit += g2->benefit;
5662 /* ??? The new final_[bg]iv_value code does a much better job
5663 of finding replaceable giv's, and hence this code may no
5664 longer be necessary. */
5665 if (! g2->replaceable && REG_USERVAR_P (g2->dest_reg))
5666 g1->benefit -= copy_cost;
5667 g1->lifetime += g2->lifetime;
5668 g1->times_used += g2->times_used;
5669
5670 if (loop_dump_stream)
5671 fprintf (loop_dump_stream, "giv at %d combined with giv at %d\n",
5672 INSN_UID (g2->insn), INSN_UID (g1->insn));
5673 }
5674 }
5675 }
5676 }
5677 \f
5678 /* EMIT code before INSERT_BEFORE to set REG = B * M + A. */
5679
5680 void
5681 emit_iv_add_mult (b, m, a, reg, insert_before)
5682 rtx b; /* initial value of basic induction variable */
5683 rtx m; /* multiplicative constant */
5684 rtx a; /* additive constant */
5685 rtx reg; /* destination register */
5686 rtx insert_before;
5687 {
5688 rtx seq;
5689 rtx result;
5690
5691 /* Prevent unexpected sharing of these rtx. */
5692 a = copy_rtx (a);
5693 b = copy_rtx (b);
5694
5695 /* Increase the lifetime of any invariants moved further in code. */
5696 update_reg_last_use (a, insert_before);
5697 update_reg_last_use (b, insert_before);
5698 update_reg_last_use (m, insert_before);
5699
5700 start_sequence ();
5701 result = expand_mult_add (b, reg, m, a, GET_MODE (reg), 0);
5702 if (reg != result)
5703 emit_move_insn (reg, result);
5704 seq = gen_sequence ();
5705 end_sequence ();
5706
5707 emit_insn_before (seq, insert_before);
5708 }
5709 \f
5710 /* Test whether A * B can be computed without
5711 an actual multiply insn. Value is 1 if so. */
5712
5713 static int
5714 product_cheap_p (a, b)
5715 rtx a;
5716 rtx b;
5717 {
5718 int i;
5719 rtx tmp;
5720 struct obstack *old_rtl_obstack = rtl_obstack;
5721 char *storage = (char *) obstack_alloc (&temp_obstack, 0);
5722 int win = 1;
5723
5724 /* If only one is constant, make it B. */
5725 if (GET_CODE (a) == CONST_INT)
5726 tmp = a, a = b, b = tmp;
5727
5728 /* If first constant, both constant, so don't need multiply. */
5729 if (GET_CODE (a) == CONST_INT)
5730 return 1;
5731
5732 /* If second not constant, neither is constant, so would need multiply. */
5733 if (GET_CODE (b) != CONST_INT)
5734 return 0;
5735
5736 /* One operand is constant, so might not need multiply insn. Generate the
5737 code for the multiply and see if a call or multiply, or long sequence
5738 of insns is generated. */
5739
5740 rtl_obstack = &temp_obstack;
5741 start_sequence ();
5742 expand_mult (GET_MODE (a), a, b, NULL_RTX, 0);
5743 tmp = gen_sequence ();
5744 end_sequence ();
5745
5746 if (GET_CODE (tmp) == SEQUENCE)
5747 {
5748 if (XVEC (tmp, 0) == 0)
5749 win = 1;
5750 else if (XVECLEN (tmp, 0) > 3)
5751 win = 0;
5752 else
5753 for (i = 0; i < XVECLEN (tmp, 0); i++)
5754 {
5755 rtx insn = XVECEXP (tmp, 0, i);
5756
5757 if (GET_CODE (insn) != INSN
5758 || (GET_CODE (PATTERN (insn)) == SET
5759 && GET_CODE (SET_SRC (PATTERN (insn))) == MULT)
5760 || (GET_CODE (PATTERN (insn)) == PARALLEL
5761 && GET_CODE (XVECEXP (PATTERN (insn), 0, 0)) == SET
5762 && GET_CODE (SET_SRC (XVECEXP (PATTERN (insn), 0, 0))) == MULT))
5763 {
5764 win = 0;
5765 break;
5766 }
5767 }
5768 }
5769 else if (GET_CODE (tmp) == SET
5770 && GET_CODE (SET_SRC (tmp)) == MULT)
5771 win = 0;
5772 else if (GET_CODE (tmp) == PARALLEL
5773 && GET_CODE (XVECEXP (tmp, 0, 0)) == SET
5774 && GET_CODE (SET_SRC (XVECEXP (tmp, 0, 0))) == MULT)
5775 win = 0;
5776
5777 /* Free any storage we obtained in generating this multiply and restore rtl
5778 allocation to its normal obstack. */
5779 obstack_free (&temp_obstack, storage);
5780 rtl_obstack = old_rtl_obstack;
5781
5782 return win;
5783 }
5784 \f
5785 /* Check to see if loop can be terminated by a "decrement and branch until
5786 zero" instruction. If so, add a REG_NONNEG note to the branch insn if so.
5787 Also try reversing an increment loop to a decrement loop
5788 to see if the optimization can be performed.
5789 Value is nonzero if optimization was performed. */
5790
5791 /* This is useful even if the architecture doesn't have such an insn,
5792 because it might change a loops which increments from 0 to n to a loop
5793 which decrements from n to 0. A loop that decrements to zero is usually
5794 faster than one that increments from zero. */
5795
5796 /* ??? This could be rewritten to use some of the loop unrolling procedures,
5797 such as approx_final_value, biv_total_increment, loop_iterations, and
5798 final_[bg]iv_value. */
5799
5800 static int
5801 check_dbra_loop (loop_end, insn_count, loop_start)
5802 rtx loop_end;
5803 int insn_count;
5804 rtx loop_start;
5805 {
5806 struct iv_class *bl;
5807 rtx reg;
5808 rtx jump_label;
5809 rtx final_value;
5810 rtx start_value;
5811 rtx new_add_val;
5812 rtx comparison;
5813 rtx before_comparison;
5814 rtx p;
5815
5816 /* If last insn is a conditional branch, and the insn before tests a
5817 register value, try to optimize it. Otherwise, we can't do anything. */
5818
5819 comparison = get_condition_for_loop (PREV_INSN (loop_end));
5820 if (comparison == 0)
5821 return 0;
5822
5823 /* Check all of the bivs to see if the compare uses one of them.
5824 Skip biv's set more than once because we can't guarantee that
5825 it will be zero on the last iteration. Also skip if the biv is
5826 used between its update and the test insn. */
5827
5828 for (bl = loop_iv_list; bl; bl = bl->next)
5829 {
5830 if (bl->biv_count == 1
5831 && bl->biv->dest_reg == XEXP (comparison, 0)
5832 && ! reg_used_between_p (regno_reg_rtx[bl->regno], bl->biv->insn,
5833 PREV_INSN (PREV_INSN (loop_end))))
5834 break;
5835 }
5836
5837 if (! bl)
5838 return 0;
5839
5840 /* Look for the case where the basic induction variable is always
5841 nonnegative, and equals zero on the last iteration.
5842 In this case, add a reg_note REG_NONNEG, which allows the
5843 m68k DBRA instruction to be used. */
5844
5845 if (((GET_CODE (comparison) == GT
5846 && GET_CODE (XEXP (comparison, 1)) == CONST_INT
5847 && INTVAL (XEXP (comparison, 1)) == -1)
5848 || (GET_CODE (comparison) == NE && XEXP (comparison, 1) == const0_rtx))
5849 && GET_CODE (bl->biv->add_val) == CONST_INT
5850 && INTVAL (bl->biv->add_val) < 0)
5851 {
5852 /* Initial value must be greater than 0,
5853 init_val % -dec_value == 0 to ensure that it equals zero on
5854 the last iteration */
5855
5856 if (GET_CODE (bl->initial_value) == CONST_INT
5857 && INTVAL (bl->initial_value) > 0
5858 && (INTVAL (bl->initial_value) %
5859 (-INTVAL (bl->biv->add_val))) == 0)
5860 {
5861 /* register always nonnegative, add REG_NOTE to branch */
5862 REG_NOTES (PREV_INSN (loop_end))
5863 = gen_rtx (EXPR_LIST, REG_NONNEG, NULL_RTX,
5864 REG_NOTES (PREV_INSN (loop_end)));
5865 bl->nonneg = 1;
5866
5867 return 1;
5868 }
5869
5870 /* If the decrement is 1 and the value was tested as >= 0 before
5871 the loop, then we can safely optimize. */
5872 for (p = loop_start; p; p = PREV_INSN (p))
5873 {
5874 if (GET_CODE (p) == CODE_LABEL)
5875 break;
5876 if (GET_CODE (p) != JUMP_INSN)
5877 continue;
5878
5879 before_comparison = get_condition_for_loop (p);
5880 if (before_comparison
5881 && XEXP (before_comparison, 0) == bl->biv->dest_reg
5882 && GET_CODE (before_comparison) == LT
5883 && XEXP (before_comparison, 1) == const0_rtx
5884 && ! reg_set_between_p (bl->biv->dest_reg, p, loop_start)
5885 && INTVAL (bl->biv->add_val) == -1)
5886 {
5887 REG_NOTES (PREV_INSN (loop_end))
5888 = gen_rtx (EXPR_LIST, REG_NONNEG, NULL_RTX,
5889 REG_NOTES (PREV_INSN (loop_end)));
5890 bl->nonneg = 1;
5891
5892 return 1;
5893 }
5894 }
5895 }
5896 else if (num_mem_sets <= 1)
5897 {
5898 /* Try to change inc to dec, so can apply above optimization. */
5899 /* Can do this if:
5900 all registers modified are induction variables or invariant,
5901 all memory references have non-overlapping addresses
5902 (obviously true if only one write)
5903 allow 2 insns for the compare/jump at the end of the loop. */
5904 /* Also, we must avoid any instructions which use both the reversed
5905 biv and another biv. Such instructions will fail if the loop is
5906 reversed. We meet this condition by requiring that either
5907 no_use_except_counting is true, or else that there is only
5908 one biv. */
5909 int num_nonfixed_reads = 0;
5910 /* 1 if the iteration var is used only to count iterations. */
5911 int no_use_except_counting = 0;
5912 /* 1 if the loop has no memory store, or it has a single memory store
5913 which is reversible. */
5914 int reversible_mem_store = 1;
5915
5916 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
5917 if (GET_RTX_CLASS (GET_CODE (p)) == 'i')
5918 num_nonfixed_reads += count_nonfixed_reads (PATTERN (p));
5919
5920 if (bl->giv_count == 0
5921 && ! loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]])
5922 {
5923 rtx bivreg = regno_reg_rtx[bl->regno];
5924
5925 /* If there are no givs for this biv, and the only exit is the
5926 fall through at the end of the the loop, then
5927 see if perhaps there are no uses except to count. */
5928 no_use_except_counting = 1;
5929 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
5930 if (GET_RTX_CLASS (GET_CODE (p)) == 'i')
5931 {
5932 rtx set = single_set (p);
5933
5934 if (set && GET_CODE (SET_DEST (set)) == REG
5935 && REGNO (SET_DEST (set)) == bl->regno)
5936 /* An insn that sets the biv is okay. */
5937 ;
5938 else if (p == prev_nonnote_insn (prev_nonnote_insn (loop_end))
5939 || p == prev_nonnote_insn (loop_end))
5940 /* Don't bother about the end test. */
5941 ;
5942 else if (reg_mentioned_p (bivreg, PATTERN (p)))
5943 /* Any other use of the biv is no good. */
5944 {
5945 no_use_except_counting = 0;
5946 break;
5947 }
5948 }
5949 }
5950
5951 /* If the loop has a single store, and the destination address is
5952 invariant, then we can't reverse the loop, because this address
5953 might then have the wrong value at loop exit.
5954 This would work if the source was invariant also, however, in that
5955 case, the insn should have been moved out of the loop. */
5956
5957 if (num_mem_sets == 1)
5958 reversible_mem_store
5959 = (! unknown_address_altered
5960 && ! invariant_p (XEXP (loop_store_mems[0], 0)));
5961
5962 /* This code only acts for innermost loops. Also it simplifies
5963 the memory address check by only reversing loops with
5964 zero or one memory access.
5965 Two memory accesses could involve parts of the same array,
5966 and that can't be reversed. */
5967
5968 if (num_nonfixed_reads <= 1
5969 && !loop_has_call
5970 && !loop_has_volatile
5971 && reversible_mem_store
5972 && (no_use_except_counting
5973 || ((bl->giv_count + bl->biv_count + num_mem_sets
5974 + num_movables + 2 == insn_count)
5975 && (bl == loop_iv_list && bl->next == 0))))
5976 {
5977 rtx tem;
5978
5979 /* Loop can be reversed. */
5980 if (loop_dump_stream)
5981 fprintf (loop_dump_stream, "Can reverse loop\n");
5982
5983 /* Now check other conditions:
5984 initial_value must be zero,
5985 final_value % add_val == 0, so that when reversed, the
5986 biv will be zero on the last iteration.
5987
5988 This test can probably be improved since +/- 1 in the constant
5989 can be obtained by changing LT to LE and vice versa; this is
5990 confusing. */
5991
5992 if (comparison && bl->initial_value == const0_rtx
5993 && GET_CODE (XEXP (comparison, 1)) == CONST_INT
5994 /* LE gets turned into LT */
5995 && GET_CODE (comparison) == LT
5996 && (INTVAL (XEXP (comparison, 1))
5997 % INTVAL (bl->biv->add_val)) == 0)
5998 {
5999 /* Register will always be nonnegative, with value
6000 0 on last iteration if loop reversed */
6001
6002 /* Save some info needed to produce the new insns. */
6003 reg = bl->biv->dest_reg;
6004 jump_label = XEXP (SET_SRC (PATTERN (PREV_INSN (loop_end))), 1);
6005 new_add_val = GEN_INT (- INTVAL (bl->biv->add_val));
6006
6007 final_value = XEXP (comparison, 1);
6008 start_value = GEN_INT (INTVAL (XEXP (comparison, 1))
6009 - INTVAL (bl->biv->add_val));
6010
6011 /* Initialize biv to start_value before loop start.
6012 The old initializing insn will be deleted as a
6013 dead store by flow.c. */
6014 emit_insn_before (gen_move_insn (reg, start_value), loop_start);
6015
6016 /* Add insn to decrement register, and delete insn
6017 that incremented the register. */
6018 p = emit_insn_before (gen_add2_insn (reg, new_add_val),
6019 bl->biv->insn);
6020 delete_insn (bl->biv->insn);
6021
6022 /* Update biv info to reflect its new status. */
6023 bl->biv->insn = p;
6024 bl->initial_value = start_value;
6025 bl->biv->add_val = new_add_val;
6026
6027 /* Inc LABEL_NUSES so that delete_insn will
6028 not delete the label. */
6029 LABEL_NUSES (XEXP (jump_label, 0)) ++;
6030
6031 /* Emit an insn after the end of the loop to set the biv's
6032 proper exit value if it is used anywhere outside the loop. */
6033 if ((regno_last_uid[bl->regno]
6034 != INSN_UID (PREV_INSN (PREV_INSN (loop_end))))
6035 || ! bl->init_insn
6036 || regno_first_uid[bl->regno] != INSN_UID (bl->init_insn))
6037 emit_insn_after (gen_move_insn (reg, final_value),
6038 loop_end);
6039
6040 /* Delete compare/branch at end of loop. */
6041 delete_insn (PREV_INSN (loop_end));
6042 delete_insn (PREV_INSN (loop_end));
6043
6044 /* Add new compare/branch insn at end of loop. */
6045 start_sequence ();
6046 emit_cmp_insn (reg, const0_rtx, GE, NULL_RTX,
6047 GET_MODE (reg), 0, 0);
6048 emit_jump_insn (gen_bge (XEXP (jump_label, 0)));
6049 tem = gen_sequence ();
6050 end_sequence ();
6051 emit_jump_insn_before (tem, loop_end);
6052
6053 for (tem = PREV_INSN (loop_end);
6054 tem && GET_CODE (tem) != JUMP_INSN; tem = PREV_INSN (tem))
6055 ;
6056 if (tem)
6057 {
6058 JUMP_LABEL (tem) = XEXP (jump_label, 0);
6059
6060 /* Increment of LABEL_NUSES done above. */
6061 /* Register is now always nonnegative,
6062 so add REG_NONNEG note to the branch. */
6063 REG_NOTES (tem) = gen_rtx (EXPR_LIST, REG_NONNEG, NULL_RTX,
6064 REG_NOTES (tem));
6065 }
6066
6067 bl->nonneg = 1;
6068
6069 /* Mark that this biv has been reversed. Each giv which depends
6070 on this biv, and which is also live past the end of the loop
6071 will have to be fixed up. */
6072
6073 bl->reversed = 1;
6074
6075 if (loop_dump_stream)
6076 fprintf (loop_dump_stream,
6077 "Reversed loop and added reg_nonneg\n");
6078
6079 return 1;
6080 }
6081 }
6082 }
6083
6084 return 0;
6085 }
6086 \f
6087 /* Verify whether the biv BL appears to be eliminable,
6088 based on the insns in the loop that refer to it.
6089 LOOP_START is the first insn of the loop, and END is the end insn.
6090
6091 If ELIMINATE_P is non-zero, actually do the elimination.
6092
6093 THRESHOLD and INSN_COUNT are from loop_optimize and are used to
6094 determine whether invariant insns should be placed inside or at the
6095 start of the loop. */
6096
6097 static int
6098 maybe_eliminate_biv (bl, loop_start, end, eliminate_p, threshold, insn_count)
6099 struct iv_class *bl;
6100 rtx loop_start;
6101 rtx end;
6102 int eliminate_p;
6103 int threshold, insn_count;
6104 {
6105 rtx reg = bl->biv->dest_reg;
6106 rtx p;
6107
6108 /* Scan all insns in the loop, stopping if we find one that uses the
6109 biv in a way that we cannot eliminate. */
6110
6111 for (p = loop_start; p != end; p = NEXT_INSN (p))
6112 {
6113 enum rtx_code code = GET_CODE (p);
6114 rtx where = threshold >= insn_count ? loop_start : p;
6115
6116 if ((code == INSN || code == JUMP_INSN || code == CALL_INSN)
6117 && reg_mentioned_p (reg, PATTERN (p))
6118 && ! maybe_eliminate_biv_1 (PATTERN (p), p, bl, eliminate_p, where))
6119 {
6120 if (loop_dump_stream)
6121 fprintf (loop_dump_stream,
6122 "Cannot eliminate biv %d: biv used in insn %d.\n",
6123 bl->regno, INSN_UID (p));
6124 break;
6125 }
6126 }
6127
6128 if (p == end)
6129 {
6130 if (loop_dump_stream)
6131 fprintf (loop_dump_stream, "biv %d %s eliminated.\n",
6132 bl->regno, eliminate_p ? "was" : "can be");
6133 return 1;
6134 }
6135
6136 return 0;
6137 }
6138 \f
6139 /* If BL appears in X (part of the pattern of INSN), see if we can
6140 eliminate its use. If so, return 1. If not, return 0.
6141
6142 If BIV does not appear in X, return 1.
6143
6144 If ELIMINATE_P is non-zero, actually do the elimination. WHERE indicates
6145 where extra insns should be added. Depending on how many items have been
6146 moved out of the loop, it will either be before INSN or at the start of
6147 the loop. */
6148
6149 static int
6150 maybe_eliminate_biv_1 (x, insn, bl, eliminate_p, where)
6151 rtx x, insn;
6152 struct iv_class *bl;
6153 int eliminate_p;
6154 rtx where;
6155 {
6156 enum rtx_code code = GET_CODE (x);
6157 rtx reg = bl->biv->dest_reg;
6158 enum machine_mode mode = GET_MODE (reg);
6159 struct induction *v;
6160 rtx arg, new, tem;
6161 int arg_operand;
6162 char *fmt;
6163 int i, j;
6164
6165 switch (code)
6166 {
6167 case REG:
6168 /* If we haven't already been able to do something with this BIV,
6169 we can't eliminate it. */
6170 if (x == reg)
6171 return 0;
6172 return 1;
6173
6174 case SET:
6175 /* If this sets the BIV, it is not a problem. */
6176 if (SET_DEST (x) == reg)
6177 return 1;
6178
6179 /* If this is an insn that defines a giv, it is also ok because
6180 it will go away when the giv is reduced. */
6181 for (v = bl->giv; v; v = v->next_iv)
6182 if (v->giv_type == DEST_REG && SET_DEST (x) == v->dest_reg)
6183 return 1;
6184
6185 #ifdef HAVE_cc0
6186 if (SET_DEST (x) == cc0_rtx && SET_SRC (x) == reg)
6187 {
6188 /* Can replace with any giv that was reduced and
6189 that has (MULT_VAL != 0) and (ADD_VAL == 0).
6190 Require a constant for MULT_VAL, so we know it's nonzero.
6191 ??? We disable this optimization to avoid potential
6192 overflows. */
6193
6194 for (v = bl->giv; v; v = v->next_iv)
6195 if (CONSTANT_P (v->mult_val) && v->mult_val != const0_rtx
6196 && v->add_val == const0_rtx
6197 && ! v->ignore && ! v->maybe_dead && v->always_computable
6198 && v->mode == mode
6199 && 0)
6200 {
6201 /* If the giv V had the auto-inc address optimization applied
6202 to it, and INSN occurs between the giv insn and the biv
6203 insn, then we must adjust the value used here.
6204 This is rare, so we don't bother to do so. */
6205 if (v->auto_inc_opt
6206 && ((INSN_LUID (v->insn) < INSN_LUID (insn)
6207 && INSN_LUID (insn) < INSN_LUID (bl->biv->insn))
6208 || (INSN_LUID (v->insn) > INSN_LUID (insn)
6209 && INSN_LUID (insn) > INSN_LUID (bl->biv->insn))))
6210 continue;
6211
6212 if (! eliminate_p)
6213 return 1;
6214
6215 /* If the giv has the opposite direction of change,
6216 then reverse the comparison. */
6217 if (INTVAL (v->mult_val) < 0)
6218 new = gen_rtx (COMPARE, GET_MODE (v->new_reg),
6219 const0_rtx, v->new_reg);
6220 else
6221 new = v->new_reg;
6222
6223 /* We can probably test that giv's reduced reg. */
6224 if (validate_change (insn, &SET_SRC (x), new, 0))
6225 return 1;
6226 }
6227
6228 /* Look for a giv with (MULT_VAL != 0) and (ADD_VAL != 0);
6229 replace test insn with a compare insn (cmp REDUCED_GIV ADD_VAL).
6230 Require a constant for MULT_VAL, so we know it's nonzero.
6231 ??? Do this only if ADD_VAL is a pointer to avoid a potential
6232 overflow problem. */
6233
6234 for (v = bl->giv; v; v = v->next_iv)
6235 if (CONSTANT_P (v->mult_val) && v->mult_val != const0_rtx
6236 && ! v->ignore && ! v->maybe_dead && v->always_computable
6237 && v->mode == mode
6238 && (GET_CODE (v->add_val) == SYMBOL_REF
6239 || GET_CODE (v->add_val) == LABEL_REF
6240 || GET_CODE (v->add_val) == CONST
6241 || (GET_CODE (v->add_val) == REG
6242 && REGNO_POINTER_FLAG (REGNO (v->add_val)))))
6243 {
6244 /* If the giv V had the auto-inc address optimization applied
6245 to it, and INSN occurs between the giv insn and the biv
6246 insn, then we must adjust the value used here.
6247 This is rare, so we don't bother to do so. */
6248 if (v->auto_inc_opt
6249 && ((INSN_LUID (v->insn) < INSN_LUID (insn)
6250 && INSN_LUID (insn) < INSN_LUID (bl->biv->insn))
6251 || (INSN_LUID (v->insn) > INSN_LUID (insn)
6252 && INSN_LUID (insn) > INSN_LUID (bl->biv->insn))))
6253 continue;
6254
6255 if (! eliminate_p)
6256 return 1;
6257
6258 /* If the giv has the opposite direction of change,
6259 then reverse the comparison. */
6260 if (INTVAL (v->mult_val) < 0)
6261 new = gen_rtx (COMPARE, VOIDmode, copy_rtx (v->add_val),
6262 v->new_reg);
6263 else
6264 new = gen_rtx (COMPARE, VOIDmode, v->new_reg,
6265 copy_rtx (v->add_val));
6266
6267 /* Replace biv with the giv's reduced register. */
6268 update_reg_last_use (v->add_val, insn);
6269 if (validate_change (insn, &SET_SRC (PATTERN (insn)), new, 0))
6270 return 1;
6271
6272 /* Insn doesn't support that constant or invariant. Copy it
6273 into a register (it will be a loop invariant.) */
6274 tem = gen_reg_rtx (GET_MODE (v->new_reg));
6275
6276 emit_insn_before (gen_move_insn (tem, copy_rtx (v->add_val)),
6277 where);
6278
6279 if (validate_change (insn, &SET_SRC (PATTERN (insn)),
6280 gen_rtx (COMPARE, VOIDmode,
6281 v->new_reg, tem), 0))
6282 return 1;
6283 }
6284 }
6285 #endif
6286 break;
6287
6288 case COMPARE:
6289 case EQ: case NE:
6290 case GT: case GE: case GTU: case GEU:
6291 case LT: case LE: case LTU: case LEU:
6292 /* See if either argument is the biv. */
6293 if (XEXP (x, 0) == reg)
6294 arg = XEXP (x, 1), arg_operand = 1;
6295 else if (XEXP (x, 1) == reg)
6296 arg = XEXP (x, 0), arg_operand = 0;
6297 else
6298 break;
6299
6300 if (CONSTANT_P (arg))
6301 {
6302 /* First try to replace with any giv that has constant positive
6303 mult_val and constant add_val. We might be able to support
6304 negative mult_val, but it seems complex to do it in general. */
6305
6306 for (v = bl->giv; v; v = v->next_iv)
6307 if (CONSTANT_P (v->mult_val) && INTVAL (v->mult_val) > 0
6308 && (GET_CODE (v->add_val) == SYMBOL_REF
6309 || GET_CODE (v->add_val) == LABEL_REF
6310 || GET_CODE (v->add_val) == CONST
6311 || (GET_CODE (v->add_val) == REG
6312 && REGNO_POINTER_FLAG (REGNO (v->add_val))))
6313 && ! v->ignore && ! v->maybe_dead && v->always_computable
6314 && v->mode == mode)
6315 {
6316 /* If the giv V had the auto-inc address optimization applied
6317 to it, and INSN occurs between the giv insn and the biv
6318 insn, then we must adjust the value used here.
6319 This is rare, so we don't bother to do so. */
6320 if (v->auto_inc_opt
6321 && ((INSN_LUID (v->insn) < INSN_LUID (insn)
6322 && INSN_LUID (insn) < INSN_LUID (bl->biv->insn))
6323 || (INSN_LUID (v->insn) > INSN_LUID (insn)
6324 && INSN_LUID (insn) > INSN_LUID (bl->biv->insn))))
6325 continue;
6326
6327 if (! eliminate_p)
6328 return 1;
6329
6330 /* Replace biv with the giv's reduced reg. */
6331 XEXP (x, 1-arg_operand) = v->new_reg;
6332
6333 /* If all constants are actually constant integers and
6334 the derived constant can be directly placed in the COMPARE,
6335 do so. */
6336 if (GET_CODE (arg) == CONST_INT
6337 && GET_CODE (v->mult_val) == CONST_INT
6338 && GET_CODE (v->add_val) == CONST_INT
6339 && validate_change (insn, &XEXP (x, arg_operand),
6340 GEN_INT (INTVAL (arg)
6341 * INTVAL (v->mult_val)
6342 + INTVAL (v->add_val)), 0))
6343 return 1;
6344
6345 /* Otherwise, load it into a register. */
6346 tem = gen_reg_rtx (mode);
6347 emit_iv_add_mult (arg, v->mult_val, v->add_val, tem, where);
6348 if (validate_change (insn, &XEXP (x, arg_operand), tem, 0))
6349 return 1;
6350
6351 /* If that failed, put back the change we made above. */
6352 XEXP (x, 1-arg_operand) = reg;
6353 }
6354
6355 /* Look for giv with positive constant mult_val and nonconst add_val.
6356 Insert insns to calculate new compare value.
6357 ??? Turn this off due to possible overflow. */
6358
6359 for (v = bl->giv; v; v = v->next_iv)
6360 if (CONSTANT_P (v->mult_val) && INTVAL (v->mult_val) > 0
6361 && ! v->ignore && ! v->maybe_dead && v->always_computable
6362 && v->mode == mode
6363 && 0)
6364 {
6365 rtx tem;
6366
6367 /* If the giv V had the auto-inc address optimization applied
6368 to it, and INSN occurs between the giv insn and the biv
6369 insn, then we must adjust the value used here.
6370 This is rare, so we don't bother to do so. */
6371 if (v->auto_inc_opt
6372 && ((INSN_LUID (v->insn) < INSN_LUID (insn)
6373 && INSN_LUID (insn) < INSN_LUID (bl->biv->insn))
6374 || (INSN_LUID (v->insn) > INSN_LUID (insn)
6375 && INSN_LUID (insn) > INSN_LUID (bl->biv->insn))))
6376 continue;
6377
6378 if (! eliminate_p)
6379 return 1;
6380
6381 tem = gen_reg_rtx (mode);
6382
6383 /* Replace biv with giv's reduced register. */
6384 validate_change (insn, &XEXP (x, 1 - arg_operand),
6385 v->new_reg, 1);
6386
6387 /* Compute value to compare against. */
6388 emit_iv_add_mult (arg, v->mult_val, v->add_val, tem, where);
6389 /* Use it in this insn. */
6390 validate_change (insn, &XEXP (x, arg_operand), tem, 1);
6391 if (apply_change_group ())
6392 return 1;
6393 }
6394 }
6395 else if (GET_CODE (arg) == REG || GET_CODE (arg) == MEM)
6396 {
6397 if (invariant_p (arg) == 1)
6398 {
6399 /* Look for giv with constant positive mult_val and nonconst
6400 add_val. Insert insns to compute new compare value.
6401 ??? Turn this off due to possible overflow. */
6402
6403 for (v = bl->giv; v; v = v->next_iv)
6404 if (CONSTANT_P (v->mult_val) && INTVAL (v->mult_val) > 0
6405 && ! v->ignore && ! v->maybe_dead && v->always_computable
6406 && v->mode == mode
6407 && 0)
6408 {
6409 rtx tem;
6410
6411 /* If the giv V had the auto-inc address optimization applied
6412 to it, and INSN occurs between the giv insn and the biv
6413 insn, then we must adjust the value used here.
6414 This is rare, so we don't bother to do so. */
6415 if (v->auto_inc_opt
6416 && ((INSN_LUID (v->insn) < INSN_LUID (insn)
6417 && INSN_LUID (insn) < INSN_LUID (bl->biv->insn))
6418 || (INSN_LUID (v->insn) > INSN_LUID (insn)
6419 && INSN_LUID (insn) > INSN_LUID (bl->biv->insn))))
6420 continue;
6421
6422 if (! eliminate_p)
6423 return 1;
6424
6425 tem = gen_reg_rtx (mode);
6426
6427 /* Replace biv with giv's reduced register. */
6428 validate_change (insn, &XEXP (x, 1 - arg_operand),
6429 v->new_reg, 1);
6430
6431 /* Compute value to compare against. */
6432 emit_iv_add_mult (arg, v->mult_val, v->add_val,
6433 tem, where);
6434 validate_change (insn, &XEXP (x, arg_operand), tem, 1);
6435 if (apply_change_group ())
6436 return 1;
6437 }
6438 }
6439
6440 /* This code has problems. Basically, you can't know when
6441 seeing if we will eliminate BL, whether a particular giv
6442 of ARG will be reduced. If it isn't going to be reduced,
6443 we can't eliminate BL. We can try forcing it to be reduced,
6444 but that can generate poor code.
6445
6446 The problem is that the benefit of reducing TV, below should
6447 be increased if BL can actually be eliminated, but this means
6448 we might have to do a topological sort of the order in which
6449 we try to process biv. It doesn't seem worthwhile to do
6450 this sort of thing now. */
6451
6452 #if 0
6453 /* Otherwise the reg compared with had better be a biv. */
6454 if (GET_CODE (arg) != REG
6455 || reg_iv_type[REGNO (arg)] != BASIC_INDUCT)
6456 return 0;
6457
6458 /* Look for a pair of givs, one for each biv,
6459 with identical coefficients. */
6460 for (v = bl->giv; v; v = v->next_iv)
6461 {
6462 struct induction *tv;
6463
6464 if (v->ignore || v->maybe_dead || v->mode != mode)
6465 continue;
6466
6467 for (tv = reg_biv_class[REGNO (arg)]->giv; tv; tv = tv->next_iv)
6468 if (! tv->ignore && ! tv->maybe_dead
6469 && rtx_equal_p (tv->mult_val, v->mult_val)
6470 && rtx_equal_p (tv->add_val, v->add_val)
6471 && tv->mode == mode)
6472 {
6473 /* If the giv V had the auto-inc address optimization applied
6474 to it, and INSN occurs between the giv insn and the biv
6475 insn, then we must adjust the value used here.
6476 This is rare, so we don't bother to do so. */
6477 if (v->auto_inc_opt
6478 && ((INSN_LUID (v->insn) < INSN_LUID (insn)
6479 && INSN_LUID (insn) < INSN_LUID (bl->biv->insn))
6480 || (INSN_LUID (v->insn) > INSN_LUID (insn)
6481 && INSN_LUID (insn) > INSN_LUID (bl->biv->insn))))
6482 continue;
6483
6484 if (! eliminate_p)
6485 return 1;
6486
6487 /* Replace biv with its giv's reduced reg. */
6488 XEXP (x, 1-arg_operand) = v->new_reg;
6489 /* Replace other operand with the other giv's
6490 reduced reg. */
6491 XEXP (x, arg_operand) = tv->new_reg;
6492 return 1;
6493 }
6494 }
6495 #endif
6496 }
6497
6498 /* If we get here, the biv can't be eliminated. */
6499 return 0;
6500
6501 case MEM:
6502 /* If this address is a DEST_ADDR giv, it doesn't matter if the
6503 biv is used in it, since it will be replaced. */
6504 for (v = bl->giv; v; v = v->next_iv)
6505 if (v->giv_type == DEST_ADDR && v->location == &XEXP (x, 0))
6506 return 1;
6507 break;
6508 }
6509
6510 /* See if any subexpression fails elimination. */
6511 fmt = GET_RTX_FORMAT (code);
6512 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
6513 {
6514 switch (fmt[i])
6515 {
6516 case 'e':
6517 if (! maybe_eliminate_biv_1 (XEXP (x, i), insn, bl,
6518 eliminate_p, where))
6519 return 0;
6520 break;
6521
6522 case 'E':
6523 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
6524 if (! maybe_eliminate_biv_1 (XVECEXP (x, i, j), insn, bl,
6525 eliminate_p, where))
6526 return 0;
6527 break;
6528 }
6529 }
6530
6531 return 1;
6532 }
6533 \f
6534 /* Return nonzero if the last use of REG
6535 is in an insn following INSN in the same basic block. */
6536
6537 static int
6538 last_use_this_basic_block (reg, insn)
6539 rtx reg;
6540 rtx insn;
6541 {
6542 rtx n;
6543 for (n = insn;
6544 n && GET_CODE (n) != CODE_LABEL && GET_CODE (n) != JUMP_INSN;
6545 n = NEXT_INSN (n))
6546 {
6547 if (regno_last_uid[REGNO (reg)] == INSN_UID (n))
6548 return 1;
6549 }
6550 return 0;
6551 }
6552 \f
6553 /* Called via `note_stores' to record the initial value of a biv. Here we
6554 just record the location of the set and process it later. */
6555
6556 static void
6557 record_initial (dest, set)
6558 rtx dest;
6559 rtx set;
6560 {
6561 struct iv_class *bl;
6562
6563 if (GET_CODE (dest) != REG
6564 || REGNO (dest) >= max_reg_before_loop
6565 || reg_iv_type[REGNO (dest)] != BASIC_INDUCT)
6566 return;
6567
6568 bl = reg_biv_class[REGNO (dest)];
6569
6570 /* If this is the first set found, record it. */
6571 if (bl->init_insn == 0)
6572 {
6573 bl->init_insn = note_insn;
6574 bl->init_set = set;
6575 }
6576 }
6577 \f
6578 /* If any of the registers in X are "old" and currently have a last use earlier
6579 than INSN, update them to have a last use of INSN. Their actual last use
6580 will be the previous insn but it will not have a valid uid_luid so we can't
6581 use it. */
6582
6583 static void
6584 update_reg_last_use (x, insn)
6585 rtx x;
6586 rtx insn;
6587 {
6588 /* Check for the case where INSN does not have a valid luid. In this case,
6589 there is no need to modify the regno_last_uid, as this can only happen
6590 when code is inserted after the loop_end to set a pseudo's final value,
6591 and hence this insn will never be the last use of x. */
6592 if (GET_CODE (x) == REG && REGNO (x) < max_reg_before_loop
6593 && INSN_UID (insn) < max_uid_for_loop
6594 && uid_luid[regno_last_uid[REGNO (x)]] < uid_luid[INSN_UID (insn)])
6595 regno_last_uid[REGNO (x)] = INSN_UID (insn);
6596 else
6597 {
6598 register int i, j;
6599 register char *fmt = GET_RTX_FORMAT (GET_CODE (x));
6600 for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
6601 {
6602 if (fmt[i] == 'e')
6603 update_reg_last_use (XEXP (x, i), insn);
6604 else if (fmt[i] == 'E')
6605 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
6606 update_reg_last_use (XVECEXP (x, i, j), insn);
6607 }
6608 }
6609 }
6610 \f
6611 /* Given a jump insn JUMP, return the condition that will cause it to branch
6612 to its JUMP_LABEL. If the condition cannot be understood, or is an
6613 inequality floating-point comparison which needs to be reversed, 0 will
6614 be returned.
6615
6616 If EARLIEST is non-zero, it is a pointer to a place where the earliest
6617 insn used in locating the condition was found. If a replacement test
6618 of the condition is desired, it should be placed in front of that
6619 insn and we will be sure that the inputs are still valid.
6620
6621 The condition will be returned in a canonical form to simplify testing by
6622 callers. Specifically:
6623
6624 (1) The code will always be a comparison operation (EQ, NE, GT, etc.).
6625 (2) Both operands will be machine operands; (cc0) will have been replaced.
6626 (3) If an operand is a constant, it will be the second operand.
6627 (4) (LE x const) will be replaced with (LT x <const+1>) and similarly
6628 for GE, GEU, and LEU. */
6629
6630 rtx
6631 get_condition (jump, earliest)
6632 rtx jump;
6633 rtx *earliest;
6634 {
6635 enum rtx_code code;
6636 rtx prev = jump;
6637 rtx set;
6638 rtx tem;
6639 rtx op0, op1;
6640 int reverse_code = 0;
6641 int did_reverse_condition = 0;
6642
6643 /* If this is not a standard conditional jump, we can't parse it. */
6644 if (GET_CODE (jump) != JUMP_INSN
6645 || ! condjump_p (jump) || simplejump_p (jump))
6646 return 0;
6647
6648 code = GET_CODE (XEXP (SET_SRC (PATTERN (jump)), 0));
6649 op0 = XEXP (XEXP (SET_SRC (PATTERN (jump)), 0), 0);
6650 op1 = XEXP (XEXP (SET_SRC (PATTERN (jump)), 0), 1);
6651
6652 if (earliest)
6653 *earliest = jump;
6654
6655 /* If this branches to JUMP_LABEL when the condition is false, reverse
6656 the condition. */
6657 if (GET_CODE (XEXP (SET_SRC (PATTERN (jump)), 2)) == LABEL_REF
6658 && XEXP (XEXP (SET_SRC (PATTERN (jump)), 2), 0) == JUMP_LABEL (jump))
6659 code = reverse_condition (code), did_reverse_condition ^= 1;
6660
6661 /* If we are comparing a register with zero, see if the register is set
6662 in the previous insn to a COMPARE or a comparison operation. Perform
6663 the same tests as a function of STORE_FLAG_VALUE as find_comparison_args
6664 in cse.c */
6665
6666 while (GET_RTX_CLASS (code) == '<' && op1 == CONST0_RTX (GET_MODE (op0)))
6667 {
6668 /* Set non-zero when we find something of interest. */
6669 rtx x = 0;
6670
6671 #ifdef HAVE_cc0
6672 /* If comparison with cc0, import actual comparison from compare
6673 insn. */
6674 if (op0 == cc0_rtx)
6675 {
6676 if ((prev = prev_nonnote_insn (prev)) == 0
6677 || GET_CODE (prev) != INSN
6678 || (set = single_set (prev)) == 0
6679 || SET_DEST (set) != cc0_rtx)
6680 return 0;
6681
6682 op0 = SET_SRC (set);
6683 op1 = CONST0_RTX (GET_MODE (op0));
6684 if (earliest)
6685 *earliest = prev;
6686 }
6687 #endif
6688
6689 /* If this is a COMPARE, pick up the two things being compared. */
6690 if (GET_CODE (op0) == COMPARE)
6691 {
6692 op1 = XEXP (op0, 1);
6693 op0 = XEXP (op0, 0);
6694 continue;
6695 }
6696 else if (GET_CODE (op0) != REG)
6697 break;
6698
6699 /* Go back to the previous insn. Stop if it is not an INSN. We also
6700 stop if it isn't a single set or if it has a REG_INC note because
6701 we don't want to bother dealing with it. */
6702
6703 if ((prev = prev_nonnote_insn (prev)) == 0
6704 || GET_CODE (prev) != INSN
6705 || FIND_REG_INC_NOTE (prev, 0)
6706 || (set = single_set (prev)) == 0)
6707 break;
6708
6709 /* If this is setting OP0, get what it sets it to if it looks
6710 relevant. */
6711 if (SET_DEST (set) == op0)
6712 {
6713 enum machine_mode inner_mode = GET_MODE (SET_SRC (set));
6714
6715 if ((GET_CODE (SET_SRC (set)) == COMPARE
6716 || (((code == NE
6717 || (code == LT
6718 && GET_MODE_CLASS (inner_mode) == MODE_INT
6719 && (GET_MODE_BITSIZE (inner_mode)
6720 <= HOST_BITS_PER_WIDE_INT)
6721 && (STORE_FLAG_VALUE
6722 & ((HOST_WIDE_INT) 1
6723 << (GET_MODE_BITSIZE (inner_mode) - 1))))
6724 #ifdef FLOAT_STORE_FLAG_VALUE
6725 || (code == LT
6726 && GET_MODE_CLASS (inner_mode) == MODE_FLOAT
6727 && FLOAT_STORE_FLAG_VALUE < 0)
6728 #endif
6729 ))
6730 && GET_RTX_CLASS (GET_CODE (SET_SRC (set))) == '<')))
6731 x = SET_SRC (set);
6732 else if (((code == EQ
6733 || (code == GE
6734 && (GET_MODE_BITSIZE (inner_mode)
6735 <= HOST_BITS_PER_WIDE_INT)
6736 && GET_MODE_CLASS (inner_mode) == MODE_INT
6737 && (STORE_FLAG_VALUE
6738 & ((HOST_WIDE_INT) 1
6739 << (GET_MODE_BITSIZE (inner_mode) - 1))))
6740 #ifdef FLOAT_STORE_FLAG_VALUE
6741 || (code == GE
6742 && GET_MODE_CLASS (inner_mode) == MODE_FLOAT
6743 && FLOAT_STORE_FLAG_VALUE < 0)
6744 #endif
6745 ))
6746 && GET_RTX_CLASS (GET_CODE (SET_SRC (set))) == '<')
6747 {
6748 /* We might have reversed a LT to get a GE here. But this wasn't
6749 actually the comparison of data, so we don't flag that we
6750 have had to reverse the condition. */
6751 did_reverse_condition ^= 1;
6752 reverse_code = 1;
6753 x = SET_SRC (set);
6754 }
6755 else
6756 break;
6757 }
6758
6759 else if (reg_set_p (op0, prev))
6760 /* If this sets OP0, but not directly, we have to give up. */
6761 break;
6762
6763 if (x)
6764 {
6765 if (GET_RTX_CLASS (GET_CODE (x)) == '<')
6766 code = GET_CODE (x);
6767 if (reverse_code)
6768 {
6769 code = reverse_condition (code);
6770 did_reverse_condition ^= 1;
6771 reverse_code = 0;
6772 }
6773
6774 op0 = XEXP (x, 0), op1 = XEXP (x, 1);
6775 if (earliest)
6776 *earliest = prev;
6777 }
6778 }
6779
6780 /* If constant is first, put it last. */
6781 if (CONSTANT_P (op0))
6782 code = swap_condition (code), tem = op0, op0 = op1, op1 = tem;
6783
6784 /* If OP0 is the result of a comparison, we weren't able to find what
6785 was really being compared, so fail. */
6786 if (GET_MODE_CLASS (GET_MODE (op0)) == MODE_CC)
6787 return 0;
6788
6789 /* Canonicalize any ordered comparison with integers involving equality
6790 if we can do computations in the relevant mode and we do not
6791 overflow. */
6792
6793 if (GET_CODE (op1) == CONST_INT
6794 && GET_MODE (op0) != VOIDmode
6795 && GET_MODE_BITSIZE (GET_MODE (op0)) <= HOST_BITS_PER_WIDE_INT)
6796 {
6797 HOST_WIDE_INT const_val = INTVAL (op1);
6798 unsigned HOST_WIDE_INT uconst_val = const_val;
6799 unsigned HOST_WIDE_INT max_val
6800 = (unsigned HOST_WIDE_INT) GET_MODE_MASK (GET_MODE (op0));
6801
6802 switch (code)
6803 {
6804 case LE:
6805 if (const_val != max_val >> 1)
6806 code = LT, op1 = GEN_INT (const_val + 1);
6807 break;
6808
6809 case GE:
6810 if (const_val
6811 != (((HOST_WIDE_INT) 1
6812 << (GET_MODE_BITSIZE (GET_MODE (op0)) - 1))))
6813 code = GT, op1 = GEN_INT (const_val - 1);
6814 break;
6815
6816 case LEU:
6817 if (uconst_val != max_val)
6818 code = LTU, op1 = GEN_INT (uconst_val + 1);
6819 break;
6820
6821 case GEU:
6822 if (uconst_val != 0)
6823 code = GTU, op1 = GEN_INT (uconst_val - 1);
6824 break;
6825 }
6826 }
6827
6828 /* If this was floating-point and we reversed anything other than an
6829 EQ or NE, return zero. */
6830 if (TARGET_FLOAT_FORMAT == IEEE_FLOAT_FORMAT
6831 && did_reverse_condition && code != NE && code != EQ
6832 && ! flag_fast_math
6833 && GET_MODE_CLASS (GET_MODE (op0)) == MODE_FLOAT)
6834 return 0;
6835
6836 #ifdef HAVE_cc0
6837 /* Never return CC0; return zero instead. */
6838 if (op0 == cc0_rtx)
6839 return 0;
6840 #endif
6841
6842 return gen_rtx (code, VOIDmode, op0, op1);
6843 }
6844
6845 /* Similar to above routine, except that we also put an invariant last
6846 unless both operands are invariants. */
6847
6848 rtx
6849 get_condition_for_loop (x)
6850 rtx x;
6851 {
6852 rtx comparison = get_condition (x, NULL_PTR);
6853
6854 if (comparison == 0
6855 || ! invariant_p (XEXP (comparison, 0))
6856 || invariant_p (XEXP (comparison, 1)))
6857 return comparison;
6858
6859 return gen_rtx (swap_condition (GET_CODE (comparison)), VOIDmode,
6860 XEXP (comparison, 1), XEXP (comparison, 0));
6861 }
This page took 0.340713 seconds and 5 git commands to generate.