1 /* Try to unroll loops, and split induction variables.
2 Copyright (C) 1992, 1993, 1994, 1995, 1997, 1998, 1999, 2000, 2001,
4 Free Software Foundation, Inc.
5 Contributed by James E. Wilson, Cygnus Support/UC Berkeley.
7 This file is part of GCC.
9 GCC is free software; you can redistribute it and/or modify it under
10 the terms of the GNU General Public License as published by the Free
11 Software Foundation; either version 2, or (at your option) any later
14 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
15 WARRANTY; without even the implied warranty of MERCHANTABILITY or
16 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
19 You should have received a copy of the GNU General Public License
20 along with GCC; see the file COPYING. If not, write to the Free
21 Software Foundation, 59 Temple Place - Suite 330, Boston, MA
24 /* Try to unroll a loop, and split induction variables.
26 Loops for which the number of iterations can be calculated exactly are
27 handled specially. If the number of iterations times the insn_count is
28 less than MAX_UNROLLED_INSNS, then the loop is unrolled completely.
29 Otherwise, we try to unroll the loop a number of times modulo the number
30 of iterations, so that only one exit test will be needed. It is unrolled
31 a number of times approximately equal to MAX_UNROLLED_INSNS divided by
34 Otherwise, if the number of iterations can be calculated exactly at
35 run time, and the loop is always entered at the top, then we try to
36 precondition the loop. That is, at run time, calculate how many times
37 the loop will execute, and then execute the loop body a few times so
38 that the remaining iterations will be some multiple of 4 (or 2 if the
39 loop is large). Then fall through to a loop unrolled 4 (or 2) times,
40 with only one exit test needed at the end of the loop.
42 Otherwise, if the number of iterations can not be calculated exactly,
43 not even at run time, then we still unroll the loop a number of times
44 approximately equal to MAX_UNROLLED_INSNS divided by the insn count,
45 but there must be an exit test after each copy of the loop body.
47 For each induction variable, which is dead outside the loop (replaceable)
48 or for which we can easily calculate the final value, if we can easily
49 calculate its value at each place where it is set as a function of the
50 current loop unroll count and the variable's value at loop entry, then
51 the induction variable is split into `N' different variables, one for
52 each copy of the loop body. One variable is live across the backward
53 branch, and the others are all calculated as a function of this variable.
54 This helps eliminate data dependencies, and leads to further opportunities
57 /* Possible improvements follow: */
59 /* ??? Add an extra pass somewhere to determine whether unrolling will
60 give any benefit. E.g. after generating all unrolled insns, compute the
61 cost of all insns and compare against cost of insns in rolled loop.
63 - On traditional architectures, unrolling a non-constant bound loop
64 is a win if there is a giv whose only use is in memory addresses, the
65 memory addresses can be split, and hence giv increments can be
67 - It is also a win if the loop is executed many times, and preconditioning
68 can be performed for the loop.
69 Add code to check for these and similar cases. */
71 /* ??? Improve control of which loops get unrolled. Could use profiling
72 info to only unroll the most commonly executed loops. Perhaps have
73 a user specifiable option to control the amount of code expansion,
74 or the percent of loops to consider for unrolling. Etc. */
76 /* ??? Look at the register copies inside the loop to see if they form a
77 simple permutation. If so, iterate the permutation until it gets back to
78 the start state. This is how many times we should unroll the loop, for
79 best results, because then all register copies can be eliminated.
80 For example, the lisp nreverse function should be unrolled 3 times
89 ??? The number of times to unroll the loop may also be based on data
90 references in the loop. For example, if we have a loop that references
91 x[i-1], x[i], and x[i+1], we should unroll it a multiple of 3 times. */
93 /* ??? Add some simple linear equation solving capability so that we can
94 determine the number of loop iterations for more complex loops.
95 For example, consider this loop from gdb
96 #define SWAP_TARGET_AND_HOST(buffer,len)
99 char *p = (char *) buffer;
100 char *q = ((char *) buffer) + len - 1;
101 int iterations = (len + 1) >> 1;
103 for (p; p < q; p++, q--;)
111 start value = p = &buffer + current_iteration
112 end value = q = &buffer + len - 1 - current_iteration
113 Given the loop exit test of "p < q", then there must be "q - p" iterations,
114 set equal to zero and solve for number of iterations:
115 q - p = len - 1 - 2*current_iteration = 0
116 current_iteration = (len - 1) / 2
117 Hence, there are (len - 1) / 2 (rounded up to the nearest integer)
118 iterations of this loop. */
120 /* ??? Currently, no labels are marked as loop invariant when doing loop
121 unrolling. This is because an insn inside the loop, that loads the address
122 of a label inside the loop into a register, could be moved outside the loop
123 by the invariant code motion pass if labels were invariant. If the loop
124 is subsequently unrolled, the code will be wrong because each unrolled
125 body of the loop will use the same address, whereas each actually needs a
126 different address. A case where this happens is when a loop containing
127 a switch statement is unrolled.
129 It would be better to let labels be considered invariant. When we
130 unroll loops here, check to see if any insns using a label local to the
131 loop were moved before the loop. If so, then correct the problem, by
132 moving the insn back into the loop, or perhaps replicate the insn before
133 the loop, one copy for each time the loop is unrolled. */
137 #include "coretypes.h"
141 #include "insn-config.h"
142 #include "integrate.h"
146 #include "function.h"
150 #include "hard-reg-set.h"
151 #include "basic-block.h"
156 /* The prime factors looked for when trying to unroll a loop by some
157 number which is modulo the total number of iterations. Just checking
158 for these 4 prime factors will find at least one factor for 75% of
159 all numbers theoretically. Practically speaking, this will succeed
160 almost all of the time since loops are generally a multiple of 2
163 #define NUM_FACTORS 4
165 static struct _factor { const int factor; int count; }
166 factors[NUM_FACTORS] = { {2, 0}, {3, 0}, {5, 0}, {7, 0}};
168 /* Describes the different types of loop unrolling performed. */
177 /* Indexed by register number, if nonzero, then it contains a pointer
178 to a struct induction for a DEST_REG giv which has been combined with
179 one of more address givs. This is needed because whenever such a DEST_REG
180 giv is modified, we must modify the value of all split address givs
181 that were combined with this DEST_REG giv. */
183 static struct induction **addr_combined_regs;
185 /* Indexed by register number, if this is a splittable induction variable,
186 then this will hold the current value of the register, which depends on the
189 static rtx *splittable_regs;
191 /* Indexed by register number, if this is a splittable induction variable,
192 then this will hold the number of instructions in the loop that modify
193 the induction variable. Used to ensure that only the last insn modifying
194 a split iv will update the original iv of the dest. */
196 static int *splittable_regs_updates;
198 /* Forward declarations. */
200 static rtx simplify_cmp_and_jump_insns (enum rtx_code, enum machine_mode,
202 static void init_reg_map (struct inline_remap *, int);
203 static rtx calculate_giv_inc (rtx, rtx, unsigned int);
204 static rtx initial_reg_note_copy (rtx, struct inline_remap *);
205 static void final_reg_note_copy (rtx *, struct inline_remap *);
206 static void copy_loop_body (struct loop *, rtx, rtx,
207 struct inline_remap *, rtx, int,
208 enum unroll_types, rtx, rtx, rtx, rtx);
209 static int find_splittable_regs (const struct loop *, enum unroll_types,
211 static int find_splittable_givs (const struct loop *, struct iv_class *,
212 enum unroll_types, rtx, int);
213 static int reg_dead_after_loop (const struct loop *, rtx);
214 static rtx fold_rtx_mult_add (rtx, rtx, rtx, enum machine_mode);
215 static rtx remap_split_bivs (struct loop *, rtx);
216 static rtx find_common_reg_term (rtx, rtx);
217 static rtx loop_find_equiv_value (const struct loop *, rtx);
219 /* Try to unroll one loop and split induction variables in the loop.
221 The loop is described by the arguments LOOP and INSN_COUNT.
222 STRENGTH_REDUCTION_P indicates whether information generated in the
223 strength reduction pass is available.
225 This function is intended to be called from within `strength_reduce'
229 unroll_loop (struct loop *loop, int insn_count, int strength_reduce_p)
231 struct loop_info *loop_info = LOOP_INFO (loop);
232 struct loop_ivs *ivs = LOOP_IVS (loop);
235 unsigned HOST_WIDE_INT temp;
236 int unroll_number = 1;
237 rtx copy_start, copy_end;
238 rtx insn, sequence, pattern, tem;
239 int max_labelno, max_insnno;
241 struct inline_remap *map;
242 char *local_label = NULL;
244 unsigned int max_local_regnum;
245 unsigned int maxregnum;
249 int splitting_not_safe = 0;
250 enum unroll_types unroll_type = UNROLL_NAIVE;
251 int loop_preconditioned = 0;
253 /* This points to the last real insn in the loop, which should be either
254 a JUMP_INSN (for conditional jumps) or a BARRIER (for unconditional
257 rtx loop_start = loop->start;
258 rtx loop_end = loop->end;
260 /* Don't bother unrolling huge loops. Since the minimum factor is
261 two, loops greater than one half of MAX_UNROLLED_INSNS will never
263 if (insn_count > MAX_UNROLLED_INSNS / 2)
265 if (loop_dump_stream)
266 fprintf (loop_dump_stream, "Unrolling failure: Loop too big.\n");
270 /* Determine type of unroll to perform. Depends on the number of iterations
271 and the size of the loop. */
273 /* If there is no strength reduce info, then set
274 loop_info->n_iterations to zero. This can happen if
275 strength_reduce can't find any bivs in the loop. A value of zero
276 indicates that the number of iterations could not be calculated. */
278 if (! strength_reduce_p)
279 loop_info->n_iterations = 0;
281 if (loop_dump_stream && loop_info->n_iterations > 0)
282 fprintf (loop_dump_stream, "Loop unrolling: " HOST_WIDE_INT_PRINT_DEC
283 " iterations.\n", loop_info->n_iterations);
285 /* Find and save a pointer to the last nonnote insn in the loop. */
287 last_loop_insn = prev_nonnote_insn (loop_end);
289 /* Calculate how many times to unroll the loop. Indicate whether or
290 not the loop is being completely unrolled. */
292 if (loop_info->n_iterations == 1)
294 /* Handle the case where the loop begins with an unconditional
295 jump to the loop condition. Make sure to delete the jump
296 insn, otherwise the loop body will never execute. */
298 /* If the last instruction is not a BARRIER or a JUMP_INSN, then
299 don't do anything. */
301 if (BARRIER_P (last_loop_insn))
303 /* Delete the jump insn. This will delete the barrier also. */
304 last_loop_insn = PREV_INSN (last_loop_insn);
308 if (loop_info->n_iterations > 0
309 /* Avoid overflow in the next expression. */
310 && loop_info->n_iterations < (unsigned) MAX_UNROLLED_INSNS
311 && loop_info->n_iterations * insn_count < (unsigned) MAX_UNROLLED_INSNS)
313 unroll_number = loop_info->n_iterations;
314 unroll_type = UNROLL_COMPLETELY;
316 else if (loop_info->n_iterations > 0)
318 /* Try to factor the number of iterations. Don't bother with the
319 general case, only using 2, 3, 5, and 7 will get 75% of all
320 numbers theoretically, and almost all in practice. */
322 for (i = 0; i < NUM_FACTORS; i++)
323 factors[i].count = 0;
325 temp = loop_info->n_iterations;
326 for (i = NUM_FACTORS - 1; i >= 0; i--)
327 while (temp % factors[i].factor == 0)
330 temp = temp / factors[i].factor;
333 /* Start with the larger factors first so that we generally
334 get lots of unrolling. */
338 for (i = 3; i >= 0; i--)
339 while (factors[i].count--)
341 if (temp * factors[i].factor < (unsigned) MAX_UNROLLED_INSNS)
343 unroll_number *= factors[i].factor;
344 temp *= factors[i].factor;
350 /* If we couldn't find any factors, then unroll as in the normal
352 if (unroll_number == 1)
354 if (loop_dump_stream)
355 fprintf (loop_dump_stream, "Loop unrolling: No factors found.\n");
358 unroll_type = UNROLL_MODULO;
361 /* Default case, calculate number of times to unroll loop based on its
363 if (unroll_type == UNROLL_NAIVE)
365 if (8 * insn_count < MAX_UNROLLED_INSNS)
367 else if (4 * insn_count < MAX_UNROLLED_INSNS)
373 /* Now we know how many times to unroll the loop. */
375 if (loop_dump_stream)
376 fprintf (loop_dump_stream, "Unrolling loop %d times.\n", unroll_number);
378 if (unroll_type == UNROLL_COMPLETELY || unroll_type == UNROLL_MODULO)
380 /* Loops of these types can start with jump down to the exit condition
381 in rare circumstances.
383 Consider a pair of nested loops where the inner loop is part
384 of the exit code for the outer loop.
386 In this case jump.c will not duplicate the exit test for the outer
387 loop, so it will start with a jump to the exit code.
389 Then consider if the inner loop turns out to iterate once and
390 only once. We will end up deleting the jumps associated with
391 the inner loop. However, the loop notes are not removed from
392 the instruction stream.
394 And finally assume that we can compute the number of iterations
397 In this case unroll may want to unroll the outer loop even though
398 it starts with a jump to the outer loop's exit code.
400 We could try to optimize this case, but it hardly seems worth it.
401 Just return without unrolling the loop in such cases. */
404 while (!LABEL_P (insn) && !JUMP_P (insn))
405 insn = NEXT_INSN (insn);
410 if (unroll_type == UNROLL_COMPLETELY)
412 /* Completely unrolling the loop: Delete the compare and branch at
413 the end (the last two instructions). This delete must done at the
414 very end of loop unrolling, to avoid problems with calls to
415 back_branch_in_range_p, which is called by find_splittable_regs.
416 All increments of splittable bivs/givs are changed to load constant
419 copy_start = loop_start;
421 /* Set insert_before to the instruction immediately after the JUMP_INSN
422 (or BARRIER), so that any NOTEs between the JUMP_INSN and the end of
423 the loop will be correctly handled by copy_loop_body. */
424 insert_before = NEXT_INSN (last_loop_insn);
426 /* Set copy_end to the insn before the jump at the end of the loop. */
427 if (BARRIER_P (last_loop_insn))
428 copy_end = PREV_INSN (PREV_INSN (last_loop_insn));
429 else if (JUMP_P (last_loop_insn))
431 copy_end = PREV_INSN (last_loop_insn);
433 /* The instruction immediately before the JUMP_INSN may be a compare
434 instruction which we do not want to copy. */
435 if (sets_cc0_p (PREV_INSN (copy_end)))
436 copy_end = PREV_INSN (copy_end);
441 /* We currently can't unroll a loop if it doesn't end with a
442 JUMP_INSN. There would need to be a mechanism that recognizes
443 this case, and then inserts a jump after each loop body, which
444 jumps to after the last loop body. */
445 if (loop_dump_stream)
446 fprintf (loop_dump_stream,
447 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
451 else if (unroll_type == UNROLL_MODULO)
453 /* Partially unrolling the loop: The compare and branch at the end
454 (the last two instructions) must remain. Don't copy the compare
455 and branch instructions at the end of the loop. Insert the unrolled
456 code immediately before the compare/branch at the end so that the
457 code will fall through to them as before. */
459 copy_start = loop_start;
461 /* Set insert_before to the jump insn at the end of the loop.
462 Set copy_end to before the jump insn at the end of the loop. */
463 if (BARRIER_P (last_loop_insn))
465 insert_before = PREV_INSN (last_loop_insn);
466 copy_end = PREV_INSN (insert_before);
468 else if (JUMP_P (last_loop_insn))
470 insert_before = last_loop_insn;
472 /* The instruction immediately before the JUMP_INSN may be a compare
473 instruction which we do not want to copy or delete. */
474 if (sets_cc0_p (PREV_INSN (insert_before)))
475 insert_before = PREV_INSN (insert_before);
477 copy_end = PREV_INSN (insert_before);
481 /* We currently can't unroll a loop if it doesn't end with a
482 JUMP_INSN. There would need to be a mechanism that recognizes
483 this case, and then inserts a jump after each loop body, which
484 jumps to after the last loop body. */
485 if (loop_dump_stream)
486 fprintf (loop_dump_stream,
487 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
493 /* Normal case: Must copy the compare and branch instructions at the
496 if (BARRIER_P (last_loop_insn))
498 /* Loop ends with an unconditional jump and a barrier.
499 Handle this like above, don't copy jump and barrier.
500 This is not strictly necessary, but doing so prevents generating
501 unconditional jumps to an immediately following label.
503 This will be corrected below if the target of this jump is
504 not the start_label. */
506 insert_before = PREV_INSN (last_loop_insn);
507 copy_end = PREV_INSN (insert_before);
509 else if (JUMP_P (last_loop_insn))
511 /* Set insert_before to immediately after the JUMP_INSN, so that
512 NOTEs at the end of the loop will be correctly handled by
514 insert_before = NEXT_INSN (last_loop_insn);
515 copy_end = last_loop_insn;
519 /* We currently can't unroll a loop if it doesn't end with a
520 JUMP_INSN. There would need to be a mechanism that recognizes
521 this case, and then inserts a jump after each loop body, which
522 jumps to after the last loop body. */
523 if (loop_dump_stream)
524 fprintf (loop_dump_stream,
525 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
529 /* If copying exit test branches because they can not be eliminated,
530 then must convert the fall through case of the branch to a jump past
531 the end of the loop. Create a label to emit after the loop and save
532 it for later use. Do not use the label after the loop, if any, since
533 it might be used by insns outside the loop, or there might be insns
534 added before it later by final_[bg]iv_value which must be after
535 the real exit label. */
536 exit_label = gen_label_rtx ();
539 while (!LABEL_P (insn) && !JUMP_P (insn))
540 insn = NEXT_INSN (insn);
544 /* The loop starts with a jump down to the exit condition test.
545 Start copying the loop after the barrier following this
547 copy_start = NEXT_INSN (insn);
549 /* Splitting induction variables doesn't work when the loop is
550 entered via a jump to the bottom, because then we end up doing
551 a comparison against a new register for a split variable, but
552 we did not execute the set insn for the new register because
553 it was skipped over. */
554 splitting_not_safe = 1;
555 if (loop_dump_stream)
556 fprintf (loop_dump_stream,
557 "Splitting not safe, because loop not entered at top.\n");
560 copy_start = loop_start;
563 /* This should always be the first label in the loop. */
564 start_label = NEXT_INSN (copy_start);
565 /* There may be a line number note and/or a loop continue note here. */
566 while (NOTE_P (start_label))
567 start_label = NEXT_INSN (start_label);
568 if (!LABEL_P (start_label))
570 /* This can happen as a result of jump threading. If the first insns in
571 the loop test the same condition as the loop's backward jump, or the
572 opposite condition, then the backward jump will be modified to point
573 to elsewhere, and the loop's start label is deleted.
575 This case currently can not be handled by the loop unrolling code. */
577 if (loop_dump_stream)
578 fprintf (loop_dump_stream,
579 "Unrolling failure: unknown insns between BEG note and loop label.\n");
582 if (LABEL_NAME (start_label))
584 /* The jump optimization pass must have combined the original start label
585 with a named label for a goto. We can't unroll this case because
586 jumps which go to the named label must be handled differently than
587 jumps to the loop start, and it is impossible to differentiate them
589 if (loop_dump_stream)
590 fprintf (loop_dump_stream,
591 "Unrolling failure: loop start label is gone\n");
595 if (unroll_type == UNROLL_NAIVE
596 && BARRIER_P (last_loop_insn)
597 && JUMP_P (PREV_INSN (last_loop_insn))
598 && start_label != JUMP_LABEL (PREV_INSN (last_loop_insn)))
600 /* In this case, we must copy the jump and barrier, because they will
601 not be converted to jumps to an immediately following label. */
603 insert_before = NEXT_INSN (last_loop_insn);
604 copy_end = last_loop_insn;
607 if (unroll_type == UNROLL_NAIVE
608 && JUMP_P (last_loop_insn)
609 && start_label != JUMP_LABEL (last_loop_insn))
611 /* ??? The loop ends with a conditional branch that does not branch back
612 to the loop start label. In this case, we must emit an unconditional
613 branch to the loop exit after emitting the final branch.
614 copy_loop_body does not have support for this currently, so we
615 give up. It doesn't seem worthwhile to unroll anyways since
616 unrolling would increase the number of branch instructions
618 if (loop_dump_stream)
619 fprintf (loop_dump_stream,
620 "Unrolling failure: final conditional branch not to loop start\n");
624 /* Allocate a translation table for the labels and insn numbers.
625 They will be filled in as we copy the insns in the loop. */
627 max_labelno = max_label_num ();
628 max_insnno = get_max_uid ();
630 /* Various paths through the unroll code may reach the "egress" label
631 without initializing fields within the map structure.
633 To be safe, we use xcalloc to zero the memory. */
634 map = xcalloc (1, sizeof (struct inline_remap));
636 /* Allocate the label map. */
640 map->label_map = xcalloc (max_labelno, sizeof (rtx));
641 local_label = xcalloc (max_labelno, sizeof (char));
644 /* Search the loop and mark all local labels, i.e. the ones which have to
645 be distinct labels when copied. For all labels which might be
646 non-local, set their label_map entries to point to themselves.
647 If they happen to be local their label_map entries will be overwritten
648 before the loop body is copied. The label_map entries for local labels
649 will be set to a different value each time the loop body is copied. */
651 for (insn = copy_start; insn != loop_end; insn = NEXT_INSN (insn))
656 local_label[CODE_LABEL_NUMBER (insn)] = 1;
657 else if (JUMP_P (insn))
659 if (JUMP_LABEL (insn))
660 set_label_in_map (map,
661 CODE_LABEL_NUMBER (JUMP_LABEL (insn)),
663 else if (GET_CODE (PATTERN (insn)) == ADDR_VEC
664 || GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC)
666 rtx pat = PATTERN (insn);
667 int diff_vec_p = GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC;
668 int len = XVECLEN (pat, diff_vec_p);
671 for (i = 0; i < len; i++)
673 label = XEXP (XVECEXP (pat, diff_vec_p, i), 0);
674 set_label_in_map (map, CODE_LABEL_NUMBER (label), label);
678 if ((note = find_reg_note (insn, REG_LABEL, NULL_RTX)))
679 set_label_in_map (map, CODE_LABEL_NUMBER (XEXP (note, 0)),
683 /* Allocate space for the insn map. */
685 map->insn_map = xmalloc (max_insnno * sizeof (rtx));
687 /* The register and constant maps depend on the number of registers
688 present, so the final maps can't be created until after
689 find_splittable_regs is called. However, they are needed for
690 preconditioning, so we create temporary maps when preconditioning
693 /* The preconditioning code may allocate two new pseudo registers. */
694 maxregnum = max_reg_num ();
696 /* local_regno is only valid for regnos < max_local_regnum. */
697 max_local_regnum = maxregnum;
699 /* Allocate and zero out the splittable_regs and addr_combined_regs
700 arrays. These must be zeroed here because they will be used if
701 loop preconditioning is performed, and must be zero for that case.
703 It is safe to do this here, since the extra registers created by the
704 preconditioning code and find_splittable_regs will never be used
705 to access the splittable_regs[] and addr_combined_regs[] arrays. */
707 splittable_regs = xcalloc (maxregnum, sizeof (rtx));
708 splittable_regs_updates = xcalloc (maxregnum, sizeof (int));
709 addr_combined_regs = xcalloc (maxregnum, sizeof (struct induction *));
710 local_regno = xcalloc (maxregnum, sizeof (char));
712 /* Mark all local registers, i.e. the ones which are referenced only
714 if (INSN_UID (copy_end) < max_uid_for_loop)
716 int copy_start_luid = INSN_LUID (copy_start);
717 int copy_end_luid = INSN_LUID (copy_end);
719 /* If a register is used in the jump insn, we must not duplicate it
720 since it will also be used outside the loop. */
721 if (JUMP_P (copy_end))
724 /* If we have a target that uses cc0, then we also must not duplicate
725 the insn that sets cc0 before the jump insn, if one is present. */
727 if (JUMP_P (copy_end)
728 && sets_cc0_p (PREV_INSN (copy_end)))
732 /* If copy_start points to the NOTE that starts the loop, then we must
733 use the next luid, because invariant pseudo-regs moved out of the loop
734 have their lifetimes modified to start here, but they are not safe
736 if (copy_start == loop_start)
739 /* If a pseudo's lifetime is entirely contained within this loop, then we
740 can use a different pseudo in each unrolled copy of the loop. This
741 results in better code. */
742 /* We must limit the generic test to max_reg_before_loop, because only
743 these pseudo registers have valid regno_first_uid info. */
744 for (r = FIRST_PSEUDO_REGISTER; r < max_reg_before_loop; ++r)
745 if (REGNO_FIRST_UID (r) > 0 && REGNO_FIRST_UID (r) < max_uid_for_loop
746 && REGNO_FIRST_LUID (r) >= copy_start_luid
747 && REGNO_LAST_UID (r) > 0 && REGNO_LAST_UID (r) < max_uid_for_loop
748 && REGNO_LAST_LUID (r) <= copy_end_luid)
750 /* However, we must also check for loop-carried dependencies.
751 If the value the pseudo has at the end of iteration X is
752 used by iteration X+1, then we can not use a different pseudo
753 for each unrolled copy of the loop. */
754 /* A pseudo is safe if regno_first_uid is a set, and this
755 set dominates all instructions from regno_first_uid to
757 /* ??? This check is simplistic. We would get better code if
758 this check was more sophisticated. */
759 if (set_dominates_use (r, REGNO_FIRST_UID (r), REGNO_LAST_UID (r),
760 copy_start, copy_end))
763 if (loop_dump_stream)
766 fprintf (loop_dump_stream, "Marked reg %d as local\n", r);
768 fprintf (loop_dump_stream, "Did not mark reg %d as local\n",
774 /* If this loop requires exit tests when unrolled, check to see if we
775 can precondition the loop so as to make the exit tests unnecessary.
776 Just like variable splitting, this is not safe if the loop is entered
777 via a jump to the bottom. Also, can not do this if no strength
778 reduce info, because precondition_loop_p uses this info. */
780 /* Must copy the loop body for preconditioning before the following
781 find_splittable_regs call since that will emit insns which need to
782 be after the preconditioned loop copies, but immediately before the
783 unrolled loop copies. */
785 /* Also, it is not safe to split induction variables for the preconditioned
786 copies of the loop body. If we split induction variables, then the code
787 assumes that each induction variable can be represented as a function
788 of its initial value and the loop iteration number. This is not true
789 in this case, because the last preconditioned copy of the loop body
790 could be any iteration from the first up to the `unroll_number-1'th,
791 depending on the initial value of the iteration variable. Therefore
792 we can not split induction variables here, because we can not calculate
793 their value. Hence, this code must occur before find_splittable_regs
796 if (unroll_type == UNROLL_NAIVE && ! splitting_not_safe && strength_reduce_p)
798 rtx initial_value, final_value, increment;
799 enum machine_mode mode;
801 if (precondition_loop_p (loop,
802 &initial_value, &final_value, &increment,
807 int abs_inc, neg_inc;
808 enum rtx_code cc = loop_info->comparison_code;
809 int less_p = (cc == LE || cc == LEU || cc == LT || cc == LTU);
810 int unsigned_p = (cc == LEU || cc == GEU || cc == LTU || cc == GTU);
812 map->reg_map = xmalloc (maxregnum * sizeof (rtx));
814 VARRAY_CONST_EQUIV_INIT (map->const_equiv_varray, maxregnum,
815 "unroll_loop_precondition");
816 global_const_equiv_varray = map->const_equiv_varray;
818 init_reg_map (map, maxregnum);
820 /* Limit loop unrolling to 4, since this will make 7 copies of
822 if (unroll_number > 4)
825 /* Save the absolute value of the increment, and also whether or
826 not it is negative. */
828 abs_inc = INTVAL (increment);
837 /* We must copy the final and initial values here to avoid
838 improperly shared rtl. */
839 final_value = copy_rtx (final_value);
840 initial_value = copy_rtx (initial_value);
842 /* Final value may have form of (PLUS val1 const1_rtx). We need
843 to convert it into general operand, so compute the real value. */
845 final_value = force_operand (final_value, NULL_RTX);
846 if (!nonmemory_operand (final_value, VOIDmode))
847 final_value = force_reg (mode, final_value);
849 /* Calculate the difference between the final and initial values.
850 Final value may be a (plus (reg x) (const_int 1)) rtx.
852 We have to deal with for (i = 0; --i < 6;) type loops.
853 For such loops the real final value is the first time the
854 loop variable overflows, so the diff we calculate is the
855 distance from the overflow value. This is 0 or ~0 for
856 unsigned loops depending on the direction, or INT_MAX,
857 INT_MAX+1 for signed loops. We really do not need the
858 exact value, since we are only interested in the diff
859 modulo the increment, and the increment is a power of 2,
860 so we can pretend that the overflow value is 0/~0. */
862 if (cc == NE || less_p != neg_inc)
863 diff = simplify_gen_binary (MINUS, mode, final_value,
866 diff = simplify_gen_unary (neg_inc ? NOT : NEG, mode,
867 initial_value, mode);
868 diff = force_operand (diff, NULL_RTX);
870 /* Now calculate (diff % (unroll * abs (increment))) by using an
872 diff = simplify_gen_binary (AND, mode, diff,
873 GEN_INT (unroll_number*abs_inc - 1));
874 diff = force_operand (diff, NULL_RTX);
876 /* Now emit a sequence of branches to jump to the proper precond
879 labels = xmalloc (sizeof (rtx) * unroll_number);
880 for (i = 0; i < unroll_number; i++)
881 labels[i] = gen_label_rtx ();
883 /* Check for the case where the initial value is greater than or
884 equal to the final value. In that case, we want to execute
885 exactly one loop iteration. The code below will fail for this
886 case. This check does not apply if the loop has a NE
887 comparison at the end. */
891 rtx incremented_initval;
892 enum rtx_code cmp_code;
895 = simplify_gen_binary (PLUS, mode, initial_value, increment);
897 = force_operand (incremented_initval, NULL_RTX);
900 ? (unsigned_p ? GEU : GE)
901 : (unsigned_p ? LEU : LE));
903 insn = simplify_cmp_and_jump_insns (cmp_code, mode,
905 final_value, labels[1]);
907 predict_insn_def (insn, PRED_LOOP_CONDITION, TAKEN);
910 /* Assuming the unroll_number is 4, and the increment is 2, then
911 for a negative increment: for a positive increment:
912 diff = 0,1 precond 0 diff = 0,7 precond 0
913 diff = 2,3 precond 3 diff = 1,2 precond 1
914 diff = 4,5 precond 2 diff = 3,4 precond 2
915 diff = 6,7 precond 1 diff = 5,6 precond 3 */
917 /* We only need to emit (unroll_number - 1) branches here, the
918 last case just falls through to the following code. */
920 /* ??? This would give better code if we emitted a tree of branches
921 instead of the current linear list of branches. */
923 for (i = 0; i < unroll_number - 1; i++)
926 enum rtx_code cmp_code;
928 /* For negative increments, must invert the constant compared
929 against, except when comparing against zero. */
937 cmp_const = unroll_number - i;
946 insn = simplify_cmp_and_jump_insns (cmp_code, mode, diff,
947 GEN_INT (abs_inc*cmp_const),
950 predict_insn (insn, PRED_LOOP_PRECONDITIONING,
951 REG_BR_PROB_BASE / (unroll_number - i));
954 /* If the increment is greater than one, then we need another branch,
955 to handle other cases equivalent to 0. */
957 /* ??? This should be merged into the code above somehow to help
958 simplify the code here, and reduce the number of branches emitted.
959 For the negative increment case, the branch here could easily
960 be merged with the `0' case branch above. For the positive
961 increment case, it is not clear how this can be simplified. */
966 enum rtx_code cmp_code;
970 cmp_const = abs_inc - 1;
975 cmp_const = abs_inc * (unroll_number - 1) + 1;
979 simplify_cmp_and_jump_insns (cmp_code, mode, diff,
980 GEN_INT (cmp_const), labels[0]);
983 sequence = get_insns ();
985 loop_insn_hoist (loop, sequence);
987 /* Only the last copy of the loop body here needs the exit
988 test, so set copy_end to exclude the compare/branch here,
989 and then reset it inside the loop when get to the last
992 if (BARRIER_P (last_loop_insn))
993 copy_end = PREV_INSN (PREV_INSN (last_loop_insn));
994 else if (JUMP_P (last_loop_insn))
996 copy_end = PREV_INSN (last_loop_insn);
998 /* The immediately preceding insn may be a compare which
999 we do not want to copy. */
1000 if (sets_cc0_p (PREV_INSN (copy_end)))
1001 copy_end = PREV_INSN (copy_end);
1007 for (i = 1; i < unroll_number; i++)
1009 emit_label_after (labels[unroll_number - i],
1010 PREV_INSN (loop_start));
1012 memset (map->insn_map, 0, max_insnno * sizeof (rtx));
1013 memset (&VARRAY_CONST_EQUIV (map->const_equiv_varray, 0),
1014 0, (VARRAY_SIZE (map->const_equiv_varray)
1015 * sizeof (struct const_equiv_data)));
1018 for (j = 0; j < max_labelno; j++)
1020 set_label_in_map (map, j, gen_label_rtx ());
1022 for (r = FIRST_PSEUDO_REGISTER; r < max_local_regnum; r++)
1026 = gen_reg_rtx (GET_MODE (regno_reg_rtx[r]));
1027 record_base_value (REGNO (map->reg_map[r]),
1028 regno_reg_rtx[r], 0);
1030 /* The last copy needs the compare/branch insns at the end,
1031 so reset copy_end here if the loop ends with a conditional
1034 if (i == unroll_number - 1)
1036 if (BARRIER_P (last_loop_insn))
1037 copy_end = PREV_INSN (PREV_INSN (last_loop_insn));
1039 copy_end = last_loop_insn;
1042 /* None of the copies are the `last_iteration', so just
1043 pass zero for that parameter. */
1044 copy_loop_body (loop, copy_start, copy_end, map, exit_label, 0,
1045 unroll_type, start_label, loop_end,
1046 loop_start, copy_end);
1048 emit_label_after (labels[0], PREV_INSN (loop_start));
1050 if (BARRIER_P (last_loop_insn))
1052 insert_before = PREV_INSN (last_loop_insn);
1053 copy_end = PREV_INSN (insert_before);
1057 insert_before = last_loop_insn;
1059 /* The instruction immediately before the JUMP_INSN may
1060 be a compare instruction which we do not want to copy
1062 if (sets_cc0_p (PREV_INSN (insert_before)))
1063 insert_before = PREV_INSN (insert_before);
1065 copy_end = PREV_INSN (insert_before);
1068 /* Set unroll type to MODULO now. */
1069 unroll_type = UNROLL_MODULO;
1070 loop_preconditioned = 1;
1077 /* If reach here, and the loop type is UNROLL_NAIVE, then don't unroll
1078 the loop unless all loops are being unrolled. */
1079 if (unroll_type == UNROLL_NAIVE && ! flag_old_unroll_all_loops)
1081 if (loop_dump_stream)
1082 fprintf (loop_dump_stream,
1083 "Unrolling failure: Naive unrolling not being done.\n");
1087 /* At this point, we are guaranteed to unroll the loop. */
1089 /* Keep track of the unroll factor for the loop. */
1090 loop_info->unroll_number = unroll_number;
1092 /* And whether the loop has been preconditioned. */
1093 loop_info->preconditioned = loop_preconditioned;
1095 /* Remember whether it was preconditioned for the second loop pass. */
1096 NOTE_PRECONDITIONED (loop->end) = loop_preconditioned;
1098 /* For each biv and giv, determine whether it can be safely split into
1099 a different variable for each unrolled copy of the loop body.
1100 We precalculate and save this info here, since computing it is
1103 Do this before deleting any instructions from the loop, so that
1104 back_branch_in_range_p will work correctly. */
1106 if (splitting_not_safe)
1109 temp = find_splittable_regs (loop, unroll_type, unroll_number);
1111 /* find_splittable_regs may have created some new registers, so must
1112 reallocate the reg_map with the new larger size, and must realloc
1113 the constant maps also. */
1115 maxregnum = max_reg_num ();
1116 map->reg_map = xmalloc (maxregnum * sizeof (rtx));
1118 init_reg_map (map, maxregnum);
1120 if (map->const_equiv_varray == 0)
1121 VARRAY_CONST_EQUIV_INIT (map->const_equiv_varray,
1122 maxregnum + temp * unroll_number * 2,
1124 global_const_equiv_varray = map->const_equiv_varray;
1126 /* Search the list of bivs and givs to find ones which need to be remapped
1127 when split, and set their reg_map entry appropriately. */
1129 for (bl = ivs->list; bl; bl = bl->next)
1131 if (REGNO (bl->biv->src_reg) != bl->regno)
1132 map->reg_map[bl->regno] = bl->biv->src_reg;
1134 /* Currently, non-reduced/final-value givs are never split. */
1135 for (v = bl->giv; v; v = v->next_iv)
1136 if (REGNO (v->src_reg) != bl->regno)
1137 map->reg_map[REGNO (v->dest_reg)] = v->src_reg;
1141 /* Use our current register alignment and pointer flags. */
1142 map->regno_pointer_align = cfun->emit->regno_pointer_align;
1143 map->x_regno_reg_rtx = cfun->emit->x_regno_reg_rtx;
1145 /* If the loop is being partially unrolled, and the iteration variables
1146 are being split, and are being renamed for the split, then must fix up
1147 the compare/jump instruction at the end of the loop to refer to the new
1148 registers. This compare isn't copied, so the registers used in it
1149 will never be replaced if it isn't done here. */
1151 if (unroll_type == UNROLL_MODULO)
1153 insn = NEXT_INSN (copy_end);
1154 if (NONJUMP_INSN_P (insn) || JUMP_P (insn))
1155 PATTERN (insn) = remap_split_bivs (loop, PATTERN (insn));
1158 /* For unroll_number times, make a copy of each instruction
1159 between copy_start and copy_end, and insert these new instructions
1160 before the end of the loop. */
1162 for (i = 0; i < unroll_number; i++)
1164 memset (map->insn_map, 0, max_insnno * sizeof (rtx));
1165 memset (&VARRAY_CONST_EQUIV (map->const_equiv_varray, 0), 0,
1166 VARRAY_SIZE (map->const_equiv_varray) * sizeof (struct const_equiv_data));
1169 for (j = 0; j < max_labelno; j++)
1171 set_label_in_map (map, j, gen_label_rtx ());
1173 for (r = FIRST_PSEUDO_REGISTER; r < max_local_regnum; r++)
1176 map->reg_map[r] = gen_reg_rtx (GET_MODE (regno_reg_rtx[r]));
1177 record_base_value (REGNO (map->reg_map[r]),
1178 regno_reg_rtx[r], 0);
1181 /* If loop starts with a branch to the test, then fix it so that
1182 it points to the test of the first unrolled copy of the loop. */
1183 if (i == 0 && loop_start != copy_start)
1185 insn = PREV_INSN (copy_start);
1186 pattern = PATTERN (insn);
1188 tem = get_label_from_map (map,
1190 (XEXP (SET_SRC (pattern), 0)));
1191 SET_SRC (pattern) = gen_rtx_LABEL_REF (VOIDmode, tem);
1193 /* Set the jump label so that it can be used by later loop unrolling
1195 JUMP_LABEL (insn) = tem;
1196 LABEL_NUSES (tem)++;
1199 copy_loop_body (loop, copy_start, copy_end, map, exit_label,
1200 i == unroll_number - 1, unroll_type, start_label,
1201 loop_end, insert_before, insert_before);
1204 /* Before deleting any insns, emit a CODE_LABEL immediately after the last
1205 insn to be deleted. This prevents any runaway delete_insn call from
1206 more insns that it should, as it always stops at a CODE_LABEL. */
1208 /* Delete the compare and branch at the end of the loop if completely
1209 unrolling the loop. Deleting the backward branch at the end also
1210 deletes the code label at the start of the loop. This is done at
1211 the very end to avoid problems with back_branch_in_range_p. */
1213 if (unroll_type == UNROLL_COMPLETELY)
1214 safety_label = emit_label_after (gen_label_rtx (), last_loop_insn);
1216 safety_label = emit_label_after (gen_label_rtx (), copy_end);
1218 /* Delete all of the original loop instructions. Don't delete the
1219 LOOP_BEG note, or the first code label in the loop. */
1221 insn = NEXT_INSN (copy_start);
1222 while (insn != safety_label)
1224 /* ??? Don't delete named code labels. They will be deleted when the
1225 jump that references them is deleted. Otherwise, we end up deleting
1226 them twice, which causes them to completely disappear instead of turn
1227 into NOTE_INSN_DELETED_LABEL notes. This in turn causes aborts in
1228 dwarfout.c/dwarf2out.c. We could perhaps fix the dwarf*out.c files
1229 to handle deleted labels instead. Or perhaps fix DECL_RTL of the
1230 associated LABEL_DECL to point to one of the new label instances. */
1231 /* ??? Likewise, we can't delete a NOTE_INSN_DELETED_LABEL note. */
1232 if (insn != start_label
1233 && ! (LABEL_P (insn) && LABEL_NAME (insn))
1235 && NOTE_LINE_NUMBER (insn) == NOTE_INSN_DELETED_LABEL))
1236 insn = delete_related_insns (insn);
1238 insn = NEXT_INSN (insn);
1241 /* Can now delete the 'safety' label emitted to protect us from runaway
1242 delete_related_insns calls. */
1243 if (INSN_DELETED_P (safety_label))
1245 delete_related_insns (safety_label);
1247 /* If exit_label exists, emit it after the loop. Doing the emit here
1248 forces it to have a higher INSN_UID than any insn in the unrolled loop.
1249 This is needed so that mostly_true_jump in reorg.c will treat jumps
1250 to this loop end label correctly, i.e. predict that they are usually
1253 emit_label_after (exit_label, loop_end);
1256 if (unroll_type == UNROLL_COMPLETELY)
1258 /* Remove the loop notes since this is no longer a loop. */
1260 delete_related_insns (loop_start);
1262 delete_related_insns (loop_end);
1265 if (map->const_equiv_varray)
1266 VARRAY_FREE (map->const_equiv_varray);
1269 free (map->label_map);
1272 free (map->insn_map);
1273 free (splittable_regs);
1274 free (splittable_regs_updates);
1275 free (addr_combined_regs);
1278 free (map->reg_map);
1282 /* A helper function for unroll_loop. Emit a compare and branch to
1283 satisfy (CMP OP1 OP2), but pass this through the simplifier first.
1284 If the branch turned out to be conditional, return it, otherwise
1288 simplify_cmp_and_jump_insns (enum rtx_code code, enum machine_mode mode,
1289 rtx op0, rtx op1, rtx label)
1293 t = simplify_const_relational_operation (code, mode, op0, op1);
1296 enum rtx_code scode = signed_condition (code);
1297 emit_cmp_and_jump_insns (op0, op1, scode, NULL_RTX, mode,
1298 code != scode, label);
1299 insn = get_last_insn ();
1301 JUMP_LABEL (insn) = label;
1302 LABEL_NUSES (label) += 1;
1306 else if (t == const_true_rtx)
1308 insn = emit_jump_insn (gen_jump (label));
1310 JUMP_LABEL (insn) = label;
1311 LABEL_NUSES (label) += 1;
1317 /* Return true if the loop can be safely, and profitably, preconditioned
1318 so that the unrolled copies of the loop body don't need exit tests.
1320 This only works if final_value, initial_value and increment can be
1321 determined, and if increment is a constant power of 2.
1322 If increment is not a power of 2, then the preconditioning modulo
1323 operation would require a real modulo instead of a boolean AND, and this
1324 is not considered `profitable'. */
1326 /* ??? If the loop is known to be executed very many times, or the machine
1327 has a very cheap divide instruction, then preconditioning is a win even
1328 when the increment is not a power of 2. Use RTX_COST to compute
1329 whether divide is cheap.
1330 ??? A divide by constant doesn't actually need a divide, look at
1331 expand_divmod. The reduced cost of this optimized modulo is not
1332 reflected in RTX_COST. */
1335 precondition_loop_p (const struct loop *loop, rtx *initial_value,
1336 rtx *final_value, rtx *increment,
1337 enum machine_mode *mode)
1339 rtx loop_start = loop->start;
1340 struct loop_info *loop_info = LOOP_INFO (loop);
1342 if (loop_info->n_iterations > 0)
1344 if (INTVAL (loop_info->increment) > 0)
1346 *initial_value = const0_rtx;
1347 *increment = const1_rtx;
1348 *final_value = GEN_INT (loop_info->n_iterations);
1352 *initial_value = GEN_INT (loop_info->n_iterations);
1353 *increment = constm1_rtx;
1354 *final_value = const0_rtx;
1358 if (loop_dump_stream)
1359 fprintf (loop_dump_stream,
1360 "Preconditioning: Success, number of iterations known, "
1361 HOST_WIDE_INT_PRINT_DEC ".\n",
1362 loop_info->n_iterations);
1366 if (loop_info->iteration_var == 0)
1368 if (loop_dump_stream)
1369 fprintf (loop_dump_stream,
1370 "Preconditioning: Could not find iteration variable.\n");
1373 else if (loop_info->initial_value == 0)
1375 if (loop_dump_stream)
1376 fprintf (loop_dump_stream,
1377 "Preconditioning: Could not find initial value.\n");
1380 else if (loop_info->increment == 0)
1382 if (loop_dump_stream)
1383 fprintf (loop_dump_stream,
1384 "Preconditioning: Could not find increment value.\n");
1387 else if (GET_CODE (loop_info->increment) != CONST_INT)
1389 if (loop_dump_stream)
1390 fprintf (loop_dump_stream,
1391 "Preconditioning: Increment not a constant.\n");
1394 else if ((exact_log2 (INTVAL (loop_info->increment)) < 0)
1395 && (exact_log2 (-INTVAL (loop_info->increment)) < 0))
1397 if (loop_dump_stream)
1398 fprintf (loop_dump_stream,
1399 "Preconditioning: Increment not a constant power of 2.\n");
1403 /* Unsigned_compare and compare_dir can be ignored here, since they do
1404 not matter for preconditioning. */
1406 if (loop_info->final_value == 0)
1408 if (loop_dump_stream)
1409 fprintf (loop_dump_stream,
1410 "Preconditioning: EQ comparison loop.\n");
1414 /* Must ensure that final_value is invariant, so call
1415 loop_invariant_p to check. Before doing so, must check regno
1416 against max_reg_before_loop to make sure that the register is in
1417 the range covered by loop_invariant_p. If it isn't, then it is
1418 most likely a biv/giv which by definition are not invariant. */
1419 if ((REG_P (loop_info->final_value)
1420 && REGNO (loop_info->final_value) >= max_reg_before_loop)
1421 || (GET_CODE (loop_info->final_value) == PLUS
1422 && REGNO (XEXP (loop_info->final_value, 0)) >= max_reg_before_loop)
1423 || ! loop_invariant_p (loop, loop_info->final_value))
1425 if (loop_dump_stream)
1426 fprintf (loop_dump_stream,
1427 "Preconditioning: Final value not invariant.\n");
1431 /* Fail for floating point values, since the caller of this function
1432 does not have code to deal with them. */
1433 if (GET_MODE_CLASS (GET_MODE (loop_info->final_value)) == MODE_FLOAT
1434 || GET_MODE_CLASS (GET_MODE (loop_info->initial_value)) == MODE_FLOAT)
1436 if (loop_dump_stream)
1437 fprintf (loop_dump_stream,
1438 "Preconditioning: Floating point final or initial value.\n");
1442 /* Fail if loop_info->iteration_var is not live before loop_start,
1443 since we need to test its value in the preconditioning code. */
1445 if (REGNO_FIRST_LUID (REGNO (loop_info->iteration_var))
1446 > INSN_LUID (loop_start))
1448 if (loop_dump_stream)
1449 fprintf (loop_dump_stream,
1450 "Preconditioning: Iteration var not live before loop start.\n");
1454 /* Note that loop_iterations biases the initial value for GIV iterators
1455 such as "while (i-- > 0)" so that we can calculate the number of
1456 iterations just like for BIV iterators.
1458 Also note that the absolute values of initial_value and
1459 final_value are unimportant as only their difference is used for
1460 calculating the number of loop iterations. */
1461 *initial_value = loop_info->initial_value;
1462 *increment = loop_info->increment;
1463 *final_value = loop_info->final_value;
1465 /* Decide what mode to do these calculations in. Choose the larger
1466 of final_value's mode and initial_value's mode, or a full-word if
1467 both are constants. */
1468 *mode = GET_MODE (*final_value);
1469 if (*mode == VOIDmode)
1471 *mode = GET_MODE (*initial_value);
1472 if (*mode == VOIDmode)
1475 else if (*mode != GET_MODE (*initial_value)
1476 && (GET_MODE_SIZE (*mode)
1477 < GET_MODE_SIZE (GET_MODE (*initial_value))))
1478 *mode = GET_MODE (*initial_value);
1481 if (loop_dump_stream)
1482 fprintf (loop_dump_stream, "Preconditioning: Successful.\n");
1486 /* All pseudo-registers must be mapped to themselves. Two hard registers
1487 must be mapped, VIRTUAL_STACK_VARS_REGNUM and VIRTUAL_INCOMING_ARGS_
1488 REGNUM, to avoid function-inlining specific conversions of these
1489 registers. All other hard regs can not be mapped because they may be
1494 init_reg_map (struct inline_remap *map, int maxregnum)
1498 for (i = maxregnum - 1; i > LAST_VIRTUAL_REGISTER; i--)
1499 map->reg_map[i] = regno_reg_rtx[i];
1500 /* Just clear the rest of the entries. */
1501 for (i = LAST_VIRTUAL_REGISTER; i >= 0; i--)
1502 map->reg_map[i] = 0;
1504 map->reg_map[VIRTUAL_STACK_VARS_REGNUM]
1505 = regno_reg_rtx[VIRTUAL_STACK_VARS_REGNUM];
1506 map->reg_map[VIRTUAL_INCOMING_ARGS_REGNUM]
1507 = regno_reg_rtx[VIRTUAL_INCOMING_ARGS_REGNUM];
1510 /* Strength-reduction will often emit code for optimized biv/givs which
1511 calculates their value in a temporary register, and then copies the result
1512 to the iv. This procedure reconstructs the pattern computing the iv;
1513 verifying that all operands are of the proper form.
1515 PATTERN must be the result of single_set.
1516 The return value is the amount that the giv is incremented by. */
1519 calculate_giv_inc (rtx pattern, rtx src_insn, unsigned int regno)
1522 rtx increment_total = 0;
1526 /* Verify that we have an increment insn here. First check for a plus
1527 as the set source. */
1528 if (GET_CODE (SET_SRC (pattern)) != PLUS)
1530 /* SR sometimes computes the new giv value in a temp, then copies it
1532 src_insn = PREV_INSN (src_insn);
1533 pattern = single_set (src_insn);
1534 if (GET_CODE (SET_SRC (pattern)) != PLUS)
1537 /* The last insn emitted is not needed, so delete it to avoid confusing
1538 the second cse pass. This insn sets the giv unnecessarily. */
1539 delete_related_insns (get_last_insn ());
1542 /* Verify that we have a constant as the second operand of the plus. */
1543 increment = XEXP (SET_SRC (pattern), 1);
1544 if (GET_CODE (increment) != CONST_INT)
1546 /* SR sometimes puts the constant in a register, especially if it is
1547 too big to be an add immed operand. */
1548 increment = find_last_value (increment, &src_insn, NULL_RTX, 0);
1550 /* SR may have used LO_SUM to compute the constant if it is too large
1551 for a load immed operand. In this case, the constant is in operand
1552 one of the LO_SUM rtx. */
1553 if (GET_CODE (increment) == LO_SUM)
1554 increment = XEXP (increment, 1);
1556 /* Some ports store large constants in memory and add a REG_EQUAL
1557 note to the store insn. */
1558 else if (MEM_P (increment))
1560 rtx note = find_reg_note (src_insn, REG_EQUAL, 0);
1562 increment = XEXP (note, 0);
1565 else if (GET_CODE (increment) == IOR
1566 || GET_CODE (increment) == PLUS
1567 || GET_CODE (increment) == ASHIFT
1568 || GET_CODE (increment) == LSHIFTRT)
1570 /* The rs6000 port loads some constants with IOR.
1571 The alpha port loads some constants with ASHIFT and PLUS.
1572 The sparc64 port loads some constants with LSHIFTRT. */
1573 rtx second_part = XEXP (increment, 1);
1574 enum rtx_code code = GET_CODE (increment);
1576 increment = find_last_value (XEXP (increment, 0),
1577 &src_insn, NULL_RTX, 0);
1578 /* Don't need the last insn anymore. */
1579 delete_related_insns (get_last_insn ());
1581 if (GET_CODE (second_part) != CONST_INT
1582 || GET_CODE (increment) != CONST_INT)
1586 increment = GEN_INT (INTVAL (increment) | INTVAL (second_part));
1587 else if (code == PLUS)
1588 increment = GEN_INT (INTVAL (increment) + INTVAL (second_part));
1589 else if (code == ASHIFT)
1590 increment = GEN_INT (INTVAL (increment) << INTVAL (second_part));
1592 increment = GEN_INT ((unsigned HOST_WIDE_INT) INTVAL (increment) >> INTVAL (second_part));
1595 if (GET_CODE (increment) != CONST_INT)
1598 /* The insn loading the constant into a register is no longer needed,
1600 delete_related_insns (get_last_insn ());
1603 if (increment_total)
1604 increment_total = GEN_INT (INTVAL (increment_total) + INTVAL (increment));
1606 increment_total = increment;
1608 /* Check that the source register is the same as the register we expected
1609 to see as the source. If not, something is seriously wrong. */
1610 if (!REG_P (XEXP (SET_SRC (pattern), 0))
1611 || REGNO (XEXP (SET_SRC (pattern), 0)) != regno)
1613 /* Some machines (e.g. the romp), may emit two add instructions for
1614 certain constants, so lets try looking for another add immediately
1615 before this one if we have only seen one add insn so far. */
1621 src_insn = PREV_INSN (src_insn);
1622 pattern = single_set (src_insn);
1624 delete_related_insns (get_last_insn ());
1632 return increment_total;
1635 /* Copy REG_NOTES, except for insn references, because not all insn_map
1636 entries are valid yet. We do need to copy registers now though, because
1637 the reg_map entries can change during copying. */
1640 initial_reg_note_copy (rtx notes, struct inline_remap *map)
1647 copy = rtx_alloc (GET_CODE (notes));
1648 PUT_REG_NOTE_KIND (copy, REG_NOTE_KIND (notes));
1650 if (GET_CODE (notes) == EXPR_LIST)
1651 XEXP (copy, 0) = copy_rtx_and_substitute (XEXP (notes, 0), map, 0);
1652 else if (GET_CODE (notes) == INSN_LIST)
1653 /* Don't substitute for these yet. */
1654 XEXP (copy, 0) = copy_rtx (XEXP (notes, 0));
1658 XEXP (copy, 1) = initial_reg_note_copy (XEXP (notes, 1), map);
1663 /* Fixup insn references in copied REG_NOTES. */
1666 final_reg_note_copy (rtx *notesp, struct inline_remap *map)
1672 if (GET_CODE (note) == INSN_LIST)
1674 rtx insn = map->insn_map[INSN_UID (XEXP (note, 0))];
1676 /* If we failed to remap the note, something is awry.
1677 Allow REG_LABEL as it may reference label outside
1678 the unrolled loop. */
1681 if (REG_NOTE_KIND (note) != REG_LABEL)
1685 XEXP (note, 0) = insn;
1688 notesp = &XEXP (note, 1);
1692 /* Copy each instruction in the loop, substituting from map as appropriate.
1693 This is very similar to a loop in expand_inline_function. */
1696 copy_loop_body (struct loop *loop, rtx copy_start, rtx copy_end,
1697 struct inline_remap *map, rtx exit_label,
1698 int last_iteration, enum unroll_types unroll_type,
1699 rtx start_label, rtx loop_end, rtx insert_before,
1700 rtx copy_notes_from)
1702 struct loop_ivs *ivs = LOOP_IVS (loop);
1704 rtx set, tem, copy = NULL_RTX;
1705 int dest_reg_was_split, i;
1709 rtx final_label = 0;
1710 rtx giv_inc, giv_dest_reg, giv_src_reg;
1712 /* If this isn't the last iteration, then map any references to the
1713 start_label to final_label. Final label will then be emitted immediately
1714 after the end of this loop body if it was ever used.
1716 If this is the last iteration, then map references to the start_label
1718 if (! last_iteration)
1720 final_label = gen_label_rtx ();
1721 set_label_in_map (map, CODE_LABEL_NUMBER (start_label), final_label);
1724 set_label_in_map (map, CODE_LABEL_NUMBER (start_label), start_label);
1731 insn = NEXT_INSN (insn);
1733 map->orig_asm_operands_vector = 0;
1735 switch (GET_CODE (insn))
1738 pattern = PATTERN (insn);
1742 /* Check to see if this is a giv that has been combined with
1743 some split address givs. (Combined in the sense that
1744 `combine_givs' in loop.c has put two givs in the same register.)
1745 In this case, we must search all givs based on the same biv to
1746 find the address givs. Then split the address givs.
1747 Do this before splitting the giv, since that may map the
1748 SET_DEST to a new register. */
1750 if ((set = single_set (insn))
1751 && REG_P (SET_DEST (set))
1752 && addr_combined_regs[REGNO (SET_DEST (set))])
1754 struct iv_class *bl;
1755 struct induction *v, *tv;
1756 unsigned int regno = REGNO (SET_DEST (set));
1758 v = addr_combined_regs[REGNO (SET_DEST (set))];
1759 bl = REG_IV_CLASS (ivs, REGNO (v->src_reg));
1761 /* Although the giv_inc amount is not needed here, we must call
1762 calculate_giv_inc here since it might try to delete the
1763 last insn emitted. If we wait until later to call it,
1764 we might accidentally delete insns generated immediately
1765 below by emit_unrolled_add. */
1767 giv_inc = calculate_giv_inc (set, insn, regno);
1769 /* Now find all address giv's that were combined with this
1771 for (tv = bl->giv; tv; tv = tv->next_iv)
1772 if (tv->giv_type == DEST_ADDR && tv->same == v)
1776 /* If this DEST_ADDR giv was not split, then ignore it. */
1777 if (*tv->location != tv->dest_reg)
1780 /* Scale this_giv_inc if the multiplicative factors of
1781 the two givs are different. */
1782 this_giv_inc = INTVAL (giv_inc);
1783 if (tv->mult_val != v->mult_val)
1784 this_giv_inc = (this_giv_inc / INTVAL (v->mult_val)
1785 * INTVAL (tv->mult_val));
1787 tv->dest_reg = plus_constant (tv->dest_reg, this_giv_inc);
1788 *tv->location = tv->dest_reg;
1790 if (last_iteration && unroll_type != UNROLL_COMPLETELY)
1792 /* Must emit an insn to increment the split address
1793 giv. Add in the const_adjust field in case there
1794 was a constant eliminated from the address. */
1795 rtx value, dest_reg;
1797 /* tv->dest_reg will be either a bare register,
1798 or else a register plus a constant. */
1799 if (REG_P (tv->dest_reg))
1800 dest_reg = tv->dest_reg;
1802 dest_reg = XEXP (tv->dest_reg, 0);
1804 /* Check for shared address givs, and avoid
1805 incrementing the shared pseudo reg more than
1807 if (! tv->same_insn && ! tv->shared)
1809 /* tv->dest_reg may actually be a (PLUS (REG)
1810 (CONST)) here, so we must call plus_constant
1811 to add the const_adjust amount before calling
1812 emit_unrolled_add below. */
1813 value = plus_constant (tv->dest_reg,
1816 if (GET_CODE (value) == PLUS)
1818 /* The constant could be too large for an add
1819 immediate, so can't directly emit an insn
1821 emit_unrolled_add (dest_reg, XEXP (value, 0),
1826 /* Reset the giv to be just the register again, in case
1827 it is used after the set we have just emitted.
1828 We must subtract the const_adjust factor added in
1830 tv->dest_reg = plus_constant (dest_reg,
1832 *tv->location = tv->dest_reg;
1837 /* If this is a setting of a splittable variable, then determine
1838 how to split the variable, create a new set based on this split,
1839 and set up the reg_map so that later uses of the variable will
1840 use the new split variable. */
1842 dest_reg_was_split = 0;
1844 if ((set = single_set (insn))
1845 && REG_P (SET_DEST (set))
1846 && splittable_regs[REGNO (SET_DEST (set))])
1848 unsigned int regno = REGNO (SET_DEST (set));
1849 unsigned int src_regno;
1851 dest_reg_was_split = 1;
1853 giv_dest_reg = SET_DEST (set);
1854 giv_src_reg = giv_dest_reg;
1855 /* Compute the increment value for the giv, if it wasn't
1856 already computed above. */
1858 giv_inc = calculate_giv_inc (set, insn, regno);
1860 src_regno = REGNO (giv_src_reg);
1862 if (unroll_type == UNROLL_COMPLETELY)
1864 /* Completely unrolling the loop. Set the induction
1865 variable to a known constant value. */
1867 /* The value in splittable_regs may be an invariant
1868 value, so we must use plus_constant here. */
1869 splittable_regs[regno]
1870 = plus_constant (splittable_regs[src_regno],
1873 if (GET_CODE (splittable_regs[regno]) == PLUS)
1875 giv_src_reg = XEXP (splittable_regs[regno], 0);
1876 giv_inc = XEXP (splittable_regs[regno], 1);
1880 /* The splittable_regs value must be a REG or a
1881 CONST_INT, so put the entire value in the giv_src_reg
1883 giv_src_reg = splittable_regs[regno];
1884 giv_inc = const0_rtx;
1889 /* Partially unrolling loop. Create a new pseudo
1890 register for the iteration variable, and set it to
1891 be a constant plus the original register. Except
1892 on the last iteration, when the result has to
1893 go back into the original iteration var register. */
1895 /* Handle bivs which must be mapped to a new register
1896 when split. This happens for bivs which need their
1897 final value set before loop entry. The new register
1898 for the biv was stored in the biv's first struct
1899 induction entry by find_splittable_regs. */
1901 if (regno < ivs->n_regs
1902 && REG_IV_TYPE (ivs, regno) == BASIC_INDUCT)
1904 giv_src_reg = REG_IV_CLASS (ivs, regno)->biv->src_reg;
1905 giv_dest_reg = giv_src_reg;
1909 /* If non-reduced/final-value givs were split, then
1910 this would have to remap those givs also. See
1911 find_splittable_regs. */
1914 splittable_regs[regno]
1915 = simplify_gen_binary (PLUS, GET_MODE (giv_src_reg),
1917 splittable_regs[src_regno]);
1918 giv_inc = splittable_regs[regno];
1920 /* Now split the induction variable by changing the dest
1921 of this insn to a new register, and setting its
1922 reg_map entry to point to this new register.
1924 If this is the last iteration, and this is the last insn
1925 that will update the iv, then reuse the original dest,
1926 to ensure that the iv will have the proper value when
1927 the loop exits or repeats.
1929 Using splittable_regs_updates here like this is safe,
1930 because it can only be greater than one if all
1931 instructions modifying the iv are always executed in
1934 if (! last_iteration
1935 || (splittable_regs_updates[regno]-- != 1))
1937 tem = gen_reg_rtx (GET_MODE (giv_src_reg));
1939 map->reg_map[regno] = tem;
1940 record_base_value (REGNO (tem),
1941 giv_inc == const0_rtx
1943 : gen_rtx_PLUS (GET_MODE (giv_src_reg),
1944 giv_src_reg, giv_inc),
1948 map->reg_map[regno] = giv_src_reg;
1951 /* The constant being added could be too large for an add
1952 immediate, so can't directly emit an insn here. */
1953 emit_unrolled_add (giv_dest_reg, giv_src_reg, giv_inc);
1954 copy = get_last_insn ();
1955 pattern = PATTERN (copy);
1959 pattern = copy_rtx_and_substitute (pattern, map, 0);
1960 copy = emit_insn (pattern);
1962 REG_NOTES (copy) = initial_reg_note_copy (REG_NOTES (insn), map);
1963 INSN_LOCATOR (copy) = INSN_LOCATOR (insn);
1965 /* If there is a REG_EQUAL note present whose value
1966 is not loop invariant, then delete it, since it
1967 may cause problems with later optimization passes. */
1968 if ((tem = find_reg_note (copy, REG_EQUAL, NULL_RTX))
1969 && !loop_invariant_p (loop, XEXP (tem, 0)))
1970 remove_note (copy, tem);
1973 /* If this insn is setting CC0, it may need to look at
1974 the insn that uses CC0 to see what type of insn it is.
1975 In that case, the call to recog via validate_change will
1976 fail. So don't substitute constants here. Instead,
1977 do it when we emit the following insn.
1979 For example, see the pyr.md file. That machine has signed and
1980 unsigned compares. The compare patterns must check the
1981 following branch insn to see which what kind of compare to
1984 If the previous insn set CC0, substitute constants on it as
1986 if (sets_cc0_p (PATTERN (copy)) != 0)
1991 try_constants (cc0_insn, map);
1993 try_constants (copy, map);
1996 try_constants (copy, map);
1999 /* Make split induction variable constants `permanent' since we
2000 know there are no backward branches across iteration variable
2001 settings which would invalidate this. */
2002 if (dest_reg_was_split)
2004 int regno = REGNO (SET_DEST (set));
2006 if ((size_t) regno < VARRAY_SIZE (map->const_equiv_varray)
2007 && (VARRAY_CONST_EQUIV (map->const_equiv_varray, regno).age
2009 VARRAY_CONST_EQUIV (map->const_equiv_varray, regno).age = -1;
2014 pattern = copy_rtx_and_substitute (PATTERN (insn), map, 0);
2015 copy = emit_jump_insn (pattern);
2016 REG_NOTES (copy) = initial_reg_note_copy (REG_NOTES (insn), map);
2017 INSN_LOCATOR (copy) = INSN_LOCATOR (insn);
2019 if (JUMP_LABEL (insn))
2021 JUMP_LABEL (copy) = get_label_from_map (map,
2023 (JUMP_LABEL (insn)));
2024 LABEL_NUSES (JUMP_LABEL (copy))++;
2026 if (JUMP_LABEL (insn) == start_label && insn == copy_end
2027 && ! last_iteration)
2030 /* This is a branch to the beginning of the loop; this is the
2031 last insn being copied; and this is not the last iteration.
2032 In this case, we want to change the original fall through
2033 case to be a branch past the end of the loop, and the
2034 original jump label case to fall_through. */
2036 if (!invert_jump (copy, exit_label, 0))
2039 rtx lab = gen_label_rtx ();
2040 /* Can't do it by reversing the jump (probably because we
2041 couldn't reverse the conditions), so emit a new
2042 jump_insn after COPY, and redirect the jump around
2044 jmp = emit_jump_insn_after (gen_jump (exit_label), copy);
2045 JUMP_LABEL (jmp) = exit_label;
2046 LABEL_NUSES (exit_label)++;
2047 jmp = emit_barrier_after (jmp);
2048 emit_label_after (lab, jmp);
2049 LABEL_NUSES (lab) = 0;
2050 if (!redirect_jump (copy, lab, 0))
2057 try_constants (cc0_insn, map);
2060 try_constants (copy, map);
2062 /* Set the jump label of COPY correctly to avoid problems with
2063 later passes of unroll_loop, if INSN had jump label set. */
2064 if (JUMP_LABEL (insn))
2068 /* Can't use the label_map for every insn, since this may be
2069 the backward branch, and hence the label was not mapped. */
2070 if ((set = single_set (copy)))
2072 tem = SET_SRC (set);
2073 if (GET_CODE (tem) == LABEL_REF)
2074 label = XEXP (tem, 0);
2075 else if (GET_CODE (tem) == IF_THEN_ELSE)
2077 if (XEXP (tem, 1) != pc_rtx)
2078 label = XEXP (XEXP (tem, 1), 0);
2080 label = XEXP (XEXP (tem, 2), 0);
2084 if (label && LABEL_P (label))
2085 JUMP_LABEL (copy) = label;
2088 /* An unrecognizable jump insn, probably the entry jump
2089 for a switch statement. This label must have been mapped,
2090 so just use the label_map to get the new jump label. */
2092 = get_label_from_map (map,
2093 CODE_LABEL_NUMBER (JUMP_LABEL (insn)));
2096 /* If this is a non-local jump, then must increase the label
2097 use count so that the label will not be deleted when the
2098 original jump is deleted. */
2099 LABEL_NUSES (JUMP_LABEL (copy))++;
2101 else if (GET_CODE (PATTERN (copy)) == ADDR_VEC
2102 || GET_CODE (PATTERN (copy)) == ADDR_DIFF_VEC)
2104 rtx pat = PATTERN (copy);
2105 int diff_vec_p = GET_CODE (pat) == ADDR_DIFF_VEC;
2106 int len = XVECLEN (pat, diff_vec_p);
2109 for (i = 0; i < len; i++)
2110 LABEL_NUSES (XEXP (XVECEXP (pat, diff_vec_p, i), 0))++;
2113 /* If this used to be a conditional jump insn but whose branch
2114 direction is now known, we must do something special. */
2115 if (any_condjump_p (insn) && onlyjump_p (insn) && map->last_pc_value)
2118 /* If the previous insn set cc0 for us, delete it. */
2119 if (only_sets_cc0_p (PREV_INSN (copy)))
2120 delete_related_insns (PREV_INSN (copy));
2123 /* If this is now a no-op, delete it. */
2124 if (map->last_pc_value == pc_rtx)
2130 /* Otherwise, this is unconditional jump so we must put a
2131 BARRIER after it. We could do some dead code elimination
2132 here, but jump.c will do it just as well. */
2138 pattern = copy_rtx_and_substitute (PATTERN (insn), map, 0);
2139 copy = emit_call_insn (pattern);
2140 REG_NOTES (copy) = initial_reg_note_copy (REG_NOTES (insn), map);
2141 INSN_LOCATOR (copy) = INSN_LOCATOR (insn);
2142 SIBLING_CALL_P (copy) = SIBLING_CALL_P (insn);
2143 CONST_OR_PURE_CALL_P (copy) = CONST_OR_PURE_CALL_P (insn);
2145 /* Because the USAGE information potentially contains objects other
2146 than hard registers, we need to copy it. */
2147 CALL_INSN_FUNCTION_USAGE (copy)
2148 = copy_rtx_and_substitute (CALL_INSN_FUNCTION_USAGE (insn),
2153 try_constants (cc0_insn, map);
2156 try_constants (copy, map);
2158 /* Be lazy and assume CALL_INSNs clobber all hard registers. */
2159 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
2160 VARRAY_CONST_EQUIV (map->const_equiv_varray, i).rtx = 0;
2164 /* If this is the loop start label, then we don't need to emit a
2165 copy of this label since no one will use it. */
2167 if (insn != start_label)
2169 copy = emit_label (get_label_from_map (map,
2170 CODE_LABEL_NUMBER (insn)));
2176 copy = emit_barrier ();
2180 /* BASIC_BLOCK notes exist to stabilize basic block structures with
2181 the associated rtl. We do not want to share the structure in
2184 if (NOTE_LINE_NUMBER (insn) != NOTE_INSN_DELETED
2185 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_DELETED_LABEL
2186 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_BASIC_BLOCK)
2187 copy = emit_note_copy (insn);
2196 map->insn_map[INSN_UID (insn)] = copy;
2198 while (insn != copy_end);
2200 /* Now finish coping the REG_NOTES. */
2204 insn = NEXT_INSN (insn);
2206 && map->insn_map[INSN_UID (insn)])
2207 final_reg_note_copy (®_NOTES (map->insn_map[INSN_UID (insn)]), map);
2209 while (insn != copy_end);
2211 /* There may be notes between copy_notes_from and loop_end. Emit a copy of
2212 each of these notes here, since there may be some important ones, such as
2213 NOTE_INSN_BLOCK_END notes, in this group. We don't do this on the last
2214 iteration, because the original notes won't be deleted.
2216 We can't use insert_before here, because when from preconditioning,
2217 insert_before points before the loop. We can't use copy_end, because
2218 there may be insns already inserted after it (which we don't want to
2219 copy) when not from preconditioning code. */
2221 if (! last_iteration)
2223 for (insn = copy_notes_from; insn != loop_end; insn = NEXT_INSN (insn))
2226 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_DELETED
2227 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_BASIC_BLOCK)
2228 emit_note_copy (insn);
2232 if (final_label && LABEL_NUSES (final_label) > 0)
2233 emit_label (final_label);
2237 loop_insn_emit_before (loop, 0, insert_before, tem);
2240 /* Emit an insn, using the expand_binop to ensure that a valid insn is
2241 emitted. This will correctly handle the case where the increment value
2242 won't fit in the immediate field of a PLUS insns. */
2245 emit_unrolled_add (rtx dest_reg, rtx src_reg, rtx increment)
2249 result = expand_simple_binop (GET_MODE (dest_reg), PLUS, src_reg, increment,
2250 dest_reg, 0, OPTAB_LIB_WIDEN);
2252 if (dest_reg != result)
2253 emit_move_insn (dest_reg, result);
2256 /* Searches the insns between INSN and LOOP->END. Returns 1 if there
2257 is a backward branch in that range that branches to somewhere between
2258 LOOP->START and INSN. Returns 0 otherwise. */
2260 /* ??? This is quadratic algorithm. Could be rewritten to be linear.
2261 In practice, this is not a problem, because this function is seldom called,
2262 and uses a negligible amount of CPU time on average. */
2265 back_branch_in_range_p (const struct loop *loop, rtx insn)
2267 rtx p, q, target_insn;
2268 rtx loop_start = loop->start;
2269 rtx loop_end = loop->end;
2270 rtx orig_loop_end = loop->end;
2272 /* Stop before we get to the backward branch at the end of the loop. */
2273 loop_end = prev_nonnote_insn (loop_end);
2274 if (BARRIER_P (loop_end))
2275 loop_end = PREV_INSN (loop_end);
2277 /* Check in case insn has been deleted, search forward for first non
2278 deleted insn following it. */
2279 while (INSN_DELETED_P (insn))
2280 insn = NEXT_INSN (insn);
2282 /* Check for the case where insn is the last insn in the loop. Deal
2283 with the case where INSN was a deleted loop test insn, in which case
2284 it will now be the NOTE_LOOP_END. */
2285 if (insn == loop_end || insn == orig_loop_end)
2288 for (p = NEXT_INSN (insn); p != loop_end; p = NEXT_INSN (p))
2292 target_insn = JUMP_LABEL (p);
2294 /* Search from loop_start to insn, to see if one of them is
2295 the target_insn. We can't use INSN_LUID comparisons here,
2296 since insn may not have an LUID entry. */
2297 for (q = loop_start; q != insn; q = NEXT_INSN (q))
2298 if (q == target_insn)
2306 /* Try to generate the simplest rtx for the expression
2307 (PLUS (MULT mult1 mult2) add1). This is used to calculate the initial
2311 fold_rtx_mult_add (rtx mult1, rtx mult2, rtx add1, enum machine_mode mode)
2316 /* The modes must all be the same. This should always be true. For now,
2317 check to make sure. */
2318 if ((GET_MODE (mult1) != mode && GET_MODE (mult1) != VOIDmode)
2319 || (GET_MODE (mult2) != mode && GET_MODE (mult2) != VOIDmode)
2320 || (GET_MODE (add1) != mode && GET_MODE (add1) != VOIDmode))
2323 /* Ensure that if at least one of mult1/mult2 are constant, then mult2
2324 will be a constant. */
2325 if (GET_CODE (mult1) == CONST_INT)
2332 mult_res = simplify_binary_operation (MULT, mode, mult1, mult2);
2334 mult_res = gen_rtx_MULT (mode, mult1, mult2);
2336 /* Again, put the constant second. */
2337 if (GET_CODE (add1) == CONST_INT)
2344 result = simplify_binary_operation (PLUS, mode, add1, mult_res);
2346 result = gen_rtx_PLUS (mode, add1, mult_res);
2351 /* Searches the list of induction struct's for the biv BL, to try to calculate
2352 the total increment value for one iteration of the loop as a constant.
2354 Returns the increment value as an rtx, simplified as much as possible,
2355 if it can be calculated. Otherwise, returns 0. */
2358 biv_total_increment (const struct iv_class *bl)
2360 struct induction *v;
2363 /* For increment, must check every instruction that sets it. Each
2364 instruction must be executed only once each time through the loop.
2365 To verify this, we check that the insn is always executed, and that
2366 there are no backward branches after the insn that branch to before it.
2367 Also, the insn must have a mult_val of one (to make sure it really is
2370 result = const0_rtx;
2371 for (v = bl->biv; v; v = v->next_iv)
2373 if (v->always_computable && v->mult_val == const1_rtx
2374 && ! v->maybe_multiple
2375 && SCALAR_INT_MODE_P (v->mode))
2377 /* If we have already counted it, skip it. */
2381 result = fold_rtx_mult_add (result, const1_rtx, v->add_val, v->mode);
2390 /* For each biv and giv, determine whether it can be safely split into
2391 a different variable for each unrolled copy of the loop body. If it
2392 is safe to split, then indicate that by saving some useful info
2393 in the splittable_regs array.
2395 If the loop is being completely unrolled, then splittable_regs will hold
2396 the current value of the induction variable while the loop is unrolled.
2397 It must be set to the initial value of the induction variable here.
2398 Otherwise, splittable_regs will hold the difference between the current
2399 value of the induction variable and the value the induction variable had
2400 at the top of the loop. It must be set to the value 0 here.
2402 Returns the total number of instructions that set registers that are
2405 /* ?? If the loop is only unrolled twice, then most of the restrictions to
2406 constant values are unnecessary, since we can easily calculate increment
2407 values in this case even if nothing is constant. The increment value
2408 should not involve a multiply however. */
2410 /* ?? Even if the biv/giv increment values aren't constant, it may still
2411 be beneficial to split the variable if the loop is only unrolled a few
2412 times, since multiplies by small integers (1,2,3,4) are very cheap. */
2415 find_splittable_regs (const struct loop *loop,
2416 enum unroll_types unroll_type, int unroll_number)
2418 struct loop_ivs *ivs = LOOP_IVS (loop);
2419 struct iv_class *bl;
2420 struct induction *v;
2422 rtx biv_final_value;
2426 for (bl = ivs->list; bl; bl = bl->next)
2428 /* Biv_total_increment must return a constant value,
2429 otherwise we can not calculate the split values. */
2431 increment = biv_total_increment (bl);
2432 if (! increment || GET_CODE (increment) != CONST_INT)
2435 /* The loop must be unrolled completely, or else have a known number
2436 of iterations and only one exit, or else the biv must be dead
2437 outside the loop, or else the final value must be known. Otherwise,
2438 it is unsafe to split the biv since it may not have the proper
2439 value on loop exit. */
2441 /* loop_number_exit_count is nonzero if the loop has an exit other than
2442 a fall through at the end. */
2445 biv_final_value = 0;
2446 if (unroll_type != UNROLL_COMPLETELY
2447 && (loop->exit_count || unroll_type == UNROLL_NAIVE)
2448 && (REGNO_LAST_LUID (bl->regno) >= INSN_LUID (loop->end)
2450 || INSN_UID (bl->init_insn) >= max_uid_for_loop
2451 || (REGNO_FIRST_LUID (bl->regno)
2452 < INSN_LUID (bl->init_insn))
2453 || reg_mentioned_p (bl->biv->dest_reg, SET_SRC (bl->init_set)))
2454 && ! (biv_final_value = final_biv_value (loop, bl)))
2457 /* If any of the insns setting the BIV don't do so with a simple
2458 PLUS, we don't know how to split it. */
2459 for (v = bl->biv; biv_splittable && v; v = v->next_iv)
2460 if ((tem = single_set (v->insn)) == 0
2461 || !REG_P (SET_DEST (tem))
2462 || REGNO (SET_DEST (tem)) != bl->regno
2463 || GET_CODE (SET_SRC (tem)) != PLUS)
2466 /* If final value is nonzero, then must emit an instruction which sets
2467 the value of the biv to the proper value. This is done after
2468 handling all of the givs, since some of them may need to use the
2469 biv's value in their initialization code. */
2471 /* This biv is splittable. If completely unrolling the loop, save
2472 the biv's initial value. Otherwise, save the constant zero. */
2474 if (biv_splittable == 1)
2476 if (unroll_type == UNROLL_COMPLETELY)
2478 /* If the initial value of the biv is itself (i.e. it is too
2479 complicated for strength_reduce to compute), or is a hard
2480 register, or it isn't invariant, then we must create a new
2481 pseudo reg to hold the initial value of the biv. */
2483 if (REG_P (bl->initial_value)
2484 && (REGNO (bl->initial_value) == bl->regno
2485 || REGNO (bl->initial_value) < FIRST_PSEUDO_REGISTER
2486 || ! loop_invariant_p (loop, bl->initial_value)))
2488 rtx tem = gen_reg_rtx (bl->biv->mode);
2490 record_base_value (REGNO (tem), bl->biv->add_val, 0);
2491 loop_insn_hoist (loop,
2492 gen_move_insn (tem, bl->biv->src_reg));
2494 if (loop_dump_stream)
2495 fprintf (loop_dump_stream,
2496 "Biv %d initial value remapped to %d.\n",
2497 bl->regno, REGNO (tem));
2499 splittable_regs[bl->regno] = tem;
2502 splittable_regs[bl->regno] = bl->initial_value;
2505 splittable_regs[bl->regno] = const0_rtx;
2507 /* Save the number of instructions that modify the biv, so that
2508 we can treat the last one specially. */
2510 splittable_regs_updates[bl->regno] = bl->biv_count;
2511 result += bl->biv_count;
2513 if (loop_dump_stream)
2514 fprintf (loop_dump_stream,
2515 "Biv %d safe to split.\n", bl->regno);
2518 /* Check every giv that depends on this biv to see whether it is
2519 splittable also. Even if the biv isn't splittable, givs which
2520 depend on it may be splittable if the biv is live outside the
2521 loop, and the givs aren't. */
2523 result += find_splittable_givs (loop, bl, unroll_type, increment,
2526 /* If final value is nonzero, then must emit an instruction which sets
2527 the value of the biv to the proper value. This is done after
2528 handling all of the givs, since some of them may need to use the
2529 biv's value in their initialization code. */
2530 if (biv_final_value)
2532 /* If the loop has multiple exits, emit the insns before the
2533 loop to ensure that it will always be executed no matter
2534 how the loop exits. Otherwise emit the insn after the loop,
2535 since this is slightly more efficient. */
2536 if (! loop->exit_count)
2537 loop_insn_sink (loop, gen_move_insn (bl->biv->src_reg,
2541 /* Create a new register to hold the value of the biv, and then
2542 set the biv to its final value before the loop start. The biv
2543 is set to its final value before loop start to ensure that
2544 this insn will always be executed, no matter how the loop
2546 rtx tem = gen_reg_rtx (bl->biv->mode);
2547 record_base_value (REGNO (tem), bl->biv->add_val, 0);
2549 loop_insn_hoist (loop, gen_move_insn (tem, bl->biv->src_reg));
2550 loop_insn_hoist (loop, gen_move_insn (bl->biv->src_reg,
2553 if (loop_dump_stream)
2554 fprintf (loop_dump_stream, "Biv %d mapped to %d for split.\n",
2555 REGNO (bl->biv->src_reg), REGNO (tem));
2557 /* Set up the mapping from the original biv register to the new
2559 bl->biv->src_reg = tem;
2566 /* For every giv based on the biv BL, check to determine whether it is
2567 splittable. This is a subroutine to find_splittable_regs ().
2569 Return the number of instructions that set splittable registers. */
2572 find_splittable_givs (const struct loop *loop, struct iv_class *bl,
2573 enum unroll_types unroll_type, rtx increment,
2574 int unroll_number ATTRIBUTE_UNUSED)
2576 struct loop_ivs *ivs = LOOP_IVS (loop);
2577 struct induction *v, *v2;
2582 /* Scan the list of givs, and set the same_insn field when there are
2583 multiple identical givs in the same insn. */
2584 for (v = bl->giv; v; v = v->next_iv)
2585 for (v2 = v->next_iv; v2; v2 = v2->next_iv)
2586 if (v->insn == v2->insn && rtx_equal_p (v->new_reg, v2->new_reg)
2590 for (v = bl->giv; v; v = v->next_iv)
2594 /* Only split the giv if it has already been reduced, or if the loop is
2595 being completely unrolled. */
2596 if (unroll_type != UNROLL_COMPLETELY && v->ignore)
2599 /* The giv can be split if the insn that sets the giv is executed once
2600 and only once on every iteration of the loop. */
2601 /* An address giv can always be split. v->insn is just a use not a set,
2602 and hence it does not matter whether it is always executed. All that
2603 matters is that all the biv increments are always executed, and we
2604 won't reach here if they aren't. */
2605 if (v->giv_type != DEST_ADDR
2606 && (! v->always_computable
2607 || back_branch_in_range_p (loop, v->insn)))
2610 /* The giv increment value must be a constant. */
2611 giv_inc = fold_rtx_mult_add (v->mult_val, increment, const0_rtx,
2613 if (! giv_inc || GET_CODE (giv_inc) != CONST_INT)
2616 /* The loop must be unrolled completely, or else have a known number of
2617 iterations and only one exit, or else the giv must be dead outside
2618 the loop, or else the final value of the giv must be known.
2619 Otherwise, it is not safe to split the giv since it may not have the
2620 proper value on loop exit. */
2622 /* The used outside loop test will fail for DEST_ADDR givs. They are
2623 never used outside the loop anyways, so it is always safe to split a
2627 if (unroll_type != UNROLL_COMPLETELY
2628 && (loop->exit_count || unroll_type == UNROLL_NAIVE)
2629 && v->giv_type != DEST_ADDR
2630 /* The next part is true if the pseudo is used outside the loop.
2631 We assume that this is true for any pseudo created after loop
2632 starts, because we don't have a reg_n_info entry for them. */
2633 && (REGNO (v->dest_reg) >= max_reg_before_loop
2634 || (REGNO_FIRST_UID (REGNO (v->dest_reg)) != INSN_UID (v->insn)
2635 /* Check for the case where the pseudo is set by a shift/add
2636 sequence, in which case the first insn setting the pseudo
2637 is the first insn of the shift/add sequence. */
2638 && (! (tem = find_reg_note (v->insn, REG_RETVAL, NULL_RTX))
2639 || (REGNO_FIRST_UID (REGNO (v->dest_reg))
2640 != INSN_UID (XEXP (tem, 0)))))
2641 /* Line above always fails if INSN was moved by loop opt. */
2642 || (REGNO_LAST_LUID (REGNO (v->dest_reg))
2643 >= INSN_LUID (loop->end)))
2644 && ! (final_value = v->final_value))
2648 /* Currently, non-reduced/final-value givs are never split. */
2649 /* Should emit insns after the loop if possible, as the biv final value
2652 /* If the final value is nonzero, and the giv has not been reduced,
2653 then must emit an instruction to set the final value. */
2654 if (final_value && !v->new_reg)
2656 /* Create a new register to hold the value of the giv, and then set
2657 the giv to its final value before the loop start. The giv is set
2658 to its final value before loop start to ensure that this insn
2659 will always be executed, no matter how we exit. */
2660 tem = gen_reg_rtx (v->mode);
2661 loop_insn_hoist (loop, gen_move_insn (tem, v->dest_reg));
2662 loop_insn_hoist (loop, gen_move_insn (v->dest_reg, final_value));
2664 if (loop_dump_stream)
2665 fprintf (loop_dump_stream, "Giv %d mapped to %d for split.\n",
2666 REGNO (v->dest_reg), REGNO (tem));
2672 /* This giv is splittable. If completely unrolling the loop, save the
2673 giv's initial value. Otherwise, save the constant zero for it. */
2675 if (unroll_type == UNROLL_COMPLETELY)
2677 /* It is not safe to use bl->initial_value here, because it may not
2678 be invariant. It is safe to use the initial value stored in
2679 the splittable_regs array if it is set. In rare cases, it won't
2680 be set, so then we do exactly the same thing as
2681 find_splittable_regs does to get a safe value. */
2682 rtx biv_initial_value;
2684 if (splittable_regs[bl->regno])
2685 biv_initial_value = splittable_regs[bl->regno];
2686 else if (!REG_P (bl->initial_value)
2687 || (REGNO (bl->initial_value) != bl->regno
2688 && REGNO (bl->initial_value) >= FIRST_PSEUDO_REGISTER))
2689 biv_initial_value = bl->initial_value;
2692 rtx tem = gen_reg_rtx (bl->biv->mode);
2694 record_base_value (REGNO (tem), bl->biv->add_val, 0);
2695 loop_insn_hoist (loop, gen_move_insn (tem, bl->biv->src_reg));
2696 biv_initial_value = tem;
2698 biv_initial_value = extend_value_for_giv (v, biv_initial_value);
2699 value = fold_rtx_mult_add (v->mult_val, biv_initial_value,
2700 v->add_val, v->mode);
2707 /* If a giv was combined with another giv, then we can only split
2708 this giv if the giv it was combined with was reduced. This
2709 is because the value of v->new_reg is meaningless in this
2711 if (v->same && ! v->same->new_reg)
2713 if (loop_dump_stream)
2714 fprintf (loop_dump_stream,
2715 "giv combined with unreduced giv not split.\n");
2718 /* If the giv is an address destination, it could be something other
2719 than a simple register, these have to be treated differently. */
2720 else if (v->giv_type == DEST_REG)
2722 /* If value is not a constant, register, or register plus
2723 constant, then compute its value into a register before
2724 loop start. This prevents invalid rtx sharing, and should
2725 generate better code. We can use bl->initial_value here
2726 instead of splittable_regs[bl->regno] because this code
2727 is going before the loop start. */
2728 if (unroll_type == UNROLL_COMPLETELY
2729 && GET_CODE (value) != CONST_INT
2731 && (GET_CODE (value) != PLUS
2732 || !REG_P (XEXP (value, 0))
2733 || GET_CODE (XEXP (value, 1)) != CONST_INT))
2735 rtx tem = gen_reg_rtx (v->mode);
2736 record_base_value (REGNO (tem), v->add_val, 0);
2737 loop_iv_add_mult_hoist (loop,
2738 extend_value_for_giv (v, bl->initial_value),
2739 v->mult_val, v->add_val, tem);
2743 splittable_regs[reg_or_subregno (v->new_reg)] = value;
2751 /* Currently, unreduced giv's can't be split. This is not too much
2752 of a problem since unreduced giv's are not live across loop
2753 iterations anyways. When unrolling a loop completely though,
2754 it makes sense to reduce&split givs when possible, as this will
2755 result in simpler instructions, and will not require that a reg
2756 be live across loop iterations. */
2758 splittable_regs[REGNO (v->dest_reg)] = value;
2759 fprintf (stderr, "Giv %d at insn %d not reduced\n",
2760 REGNO (v->dest_reg), INSN_UID (v->insn));
2766 /* Unreduced givs are only updated once by definition. Reduced givs
2767 are updated as many times as their biv is. Mark it so if this is
2768 a splittable register. Don't need to do anything for address givs
2769 where this may not be a register. */
2771 if (REG_P (v->new_reg))
2775 count = REG_IV_CLASS (ivs, REGNO (v->src_reg))->biv_count;
2777 splittable_regs_updates[reg_or_subregno (v->new_reg)] = count;
2782 if (loop_dump_stream)
2786 if (GET_CODE (v->dest_reg) == CONST_INT)
2788 else if (!REG_P (v->dest_reg))
2789 regnum = REGNO (XEXP (v->dest_reg, 0));
2791 regnum = REGNO (v->dest_reg);
2792 fprintf (loop_dump_stream, "Giv %d at insn %d safe to split.\n",
2793 regnum, INSN_UID (v->insn));
2800 /* Try to prove that the register is dead after the loop exits. Trace every
2801 loop exit looking for an insn that will always be executed, which sets
2802 the register to some value, and appears before the first use of the register
2803 is found. If successful, then return 1, otherwise return 0. */
2805 /* ?? Could be made more intelligent in the handling of jumps, so that
2806 it can search past if statements and other similar structures. */
2809 reg_dead_after_loop (const struct loop *loop, rtx reg)
2813 int label_count = 0;
2815 /* In addition to checking all exits of this loop, we must also check
2816 all exits of inner nested loops that would exit this loop. We don't
2817 have any way to identify those, so we just give up if there are any
2818 such inner loop exits. */
2820 for (label = loop->exit_labels; label; label = LABEL_NEXTREF (label))
2823 if (label_count != loop->exit_count)
2826 /* HACK: Must also search the loop fall through exit, create a label_ref
2827 here which points to the loop->end, and append the loop_number_exit_labels
2829 label = gen_rtx_LABEL_REF (VOIDmode, loop->end);
2830 LABEL_NEXTREF (label) = loop->exit_labels;
2832 for (; label; label = LABEL_NEXTREF (label))
2834 /* Succeed if find an insn which sets the biv or if reach end of
2835 function. Fail if find an insn that uses the biv, or if come to
2836 a conditional jump. */
2838 insn = NEXT_INSN (XEXP (label, 0));
2845 if (reg_referenced_p (reg, PATTERN (insn)))
2848 note = find_reg_equal_equiv_note (insn);
2849 if (note && reg_overlap_mentioned_p (reg, XEXP (note, 0)))
2852 set = single_set (insn);
2853 if (set && rtx_equal_p (SET_DEST (set), reg))
2858 if (GET_CODE (PATTERN (insn)) == RETURN)
2860 else if (!any_uncondjump_p (insn)
2861 /* Prevent infinite loop following infinite loops. */
2862 || jump_count++ > 20)
2865 insn = JUMP_LABEL (insn);
2869 insn = NEXT_INSN (insn);
2873 /* Success, the register is dead on all loop exits. */
2877 /* Try to calculate the final value of the biv, the value it will have at
2878 the end of the loop. If we can do it, return that value. */
2881 final_biv_value (const struct loop *loop, struct iv_class *bl)
2883 unsigned HOST_WIDE_INT n_iterations = LOOP_INFO (loop)->n_iterations;
2886 /* ??? This only works for MODE_INT biv's. Reject all others for now. */
2888 if (GET_MODE_CLASS (bl->biv->mode) != MODE_INT)
2891 /* The final value for reversed bivs must be calculated differently than
2892 for ordinary bivs. In this case, there is already an insn after the
2893 loop which sets this biv's final value (if necessary), and there are
2894 no other loop exits, so we can return any value. */
2897 if (loop_dump_stream)
2898 fprintf (loop_dump_stream,
2899 "Final biv value for %d, reversed biv.\n", bl->regno);
2904 /* Try to calculate the final value as initial value + (number of iterations
2905 * increment). For this to work, increment must be invariant, the only
2906 exit from the loop must be the fall through at the bottom (otherwise
2907 it may not have its final value when the loop exits), and the initial
2908 value of the biv must be invariant. */
2910 if (n_iterations != 0
2911 && ! loop->exit_count
2912 && loop_invariant_p (loop, bl->initial_value))
2914 increment = biv_total_increment (bl);
2916 if (increment && loop_invariant_p (loop, increment))
2918 /* Can calculate the loop exit value, emit insns after loop
2919 end to calculate this value into a temporary register in
2920 case it is needed later. */
2922 tem = gen_reg_rtx (bl->biv->mode);
2923 record_base_value (REGNO (tem), bl->biv->add_val, 0);
2924 loop_iv_add_mult_sink (loop, increment, GEN_INT (n_iterations),
2925 bl->initial_value, tem);
2927 if (loop_dump_stream)
2928 fprintf (loop_dump_stream,
2929 "Final biv value for %d, calculated.\n", bl->regno);
2935 /* Check to see if the biv is dead at all loop exits. */
2936 if (reg_dead_after_loop (loop, bl->biv->src_reg))
2938 if (loop_dump_stream)
2939 fprintf (loop_dump_stream,
2940 "Final biv value for %d, biv dead after loop exit.\n",
2949 /* Try to calculate the final value of the giv, the value it will have at
2950 the end of the loop. If we can do it, return that value. */
2953 final_giv_value (const struct loop *loop, struct induction *v)
2955 struct loop_ivs *ivs = LOOP_IVS (loop);
2956 struct iv_class *bl;
2960 rtx loop_end = loop->end;
2961 unsigned HOST_WIDE_INT n_iterations = LOOP_INFO (loop)->n_iterations;
2963 bl = REG_IV_CLASS (ivs, REGNO (v->src_reg));
2965 /* The final value for givs which depend on reversed bivs must be calculated
2966 differently than for ordinary givs. In this case, there is already an
2967 insn after the loop which sets this giv's final value (if necessary),
2968 and there are no other loop exits, so we can return any value. */
2971 if (loop_dump_stream)
2972 fprintf (loop_dump_stream,
2973 "Final giv value for %d, depends on reversed biv\n",
2974 REGNO (v->dest_reg));
2978 /* Try to calculate the final value as a function of the biv it depends
2979 upon. The only exit from the loop must be the fall through at the bottom
2980 and the insn that sets the giv must be executed on every iteration
2981 (otherwise the giv may not have its final value when the loop exits). */
2983 /* ??? Can calculate the final giv value by subtracting off the
2984 extra biv increments times the giv's mult_val. The loop must have
2985 only one exit for this to work, but the loop iterations does not need
2988 if (n_iterations != 0
2989 && ! loop->exit_count
2990 && v->always_executed)
2992 /* ?? It is tempting to use the biv's value here since these insns will
2993 be put after the loop, and hence the biv will have its final value
2994 then. However, this fails if the biv is subsequently eliminated.
2995 Perhaps determine whether biv's are eliminable before trying to
2996 determine whether giv's are replaceable so that we can use the
2997 biv value here if it is not eliminable. */
2999 /* We are emitting code after the end of the loop, so we must make
3000 sure that bl->initial_value is still valid then. It will still
3001 be valid if it is invariant. */
3003 increment = biv_total_increment (bl);
3005 if (increment && loop_invariant_p (loop, increment)
3006 && loop_invariant_p (loop, bl->initial_value))
3008 /* Can calculate the loop exit value of its biv as
3009 (n_iterations * increment) + initial_value */
3011 /* The loop exit value of the giv is then
3012 (final_biv_value - extra increments) * mult_val + add_val.
3013 The extra increments are any increments to the biv which
3014 occur in the loop after the giv's value is calculated.
3015 We must search from the insn that sets the giv to the end
3016 of the loop to calculate this value. */
3018 /* Put the final biv value in tem. */
3019 tem = gen_reg_rtx (v->mode);
3020 record_base_value (REGNO (tem), bl->biv->add_val, 0);
3021 loop_iv_add_mult_sink (loop, extend_value_for_giv (v, increment),
3022 GEN_INT (n_iterations),
3023 extend_value_for_giv (v, bl->initial_value),
3026 /* Subtract off extra increments as we find them. */
3027 for (insn = NEXT_INSN (v->insn); insn != loop_end;
3028 insn = NEXT_INSN (insn))
3030 struct induction *biv;
3032 for (biv = bl->biv; biv; biv = biv->next_iv)
3033 if (biv->insn == insn)
3036 tem = expand_simple_binop (GET_MODE (tem), MINUS, tem,
3037 biv->add_val, NULL_RTX, 0,
3041 loop_insn_sink (loop, seq);
3045 /* Now calculate the giv's final value. */
3046 loop_iv_add_mult_sink (loop, tem, v->mult_val, v->add_val, tem);
3048 if (loop_dump_stream)
3049 fprintf (loop_dump_stream,
3050 "Final giv value for %d, calc from biv's value.\n",
3051 REGNO (v->dest_reg));
3057 /* Replaceable giv's should never reach here. */
3061 /* Check to see if the biv is dead at all loop exits. */
3062 if (reg_dead_after_loop (loop, v->dest_reg))
3064 if (loop_dump_stream)
3065 fprintf (loop_dump_stream,
3066 "Final giv value for %d, giv dead after loop exit.\n",
3067 REGNO (v->dest_reg));
3075 /* Look back before LOOP->START for the insn that sets REG and return
3076 the equivalent constant if there is a REG_EQUAL note otherwise just
3077 the SET_SRC of REG. */
3080 loop_find_equiv_value (const struct loop *loop, rtx reg)
3082 rtx loop_start = loop->start;
3087 for (insn = PREV_INSN (loop_start); insn; insn = PREV_INSN (insn))
3092 else if (INSN_P (insn) && reg_set_p (reg, insn))
3094 /* We found the last insn before the loop that sets the register.
3095 If it sets the entire register, and has a REG_EQUAL note,
3096 then use the value of the REG_EQUAL note. */
3097 if ((set = single_set (insn))
3098 && (SET_DEST (set) == reg))
3100 rtx note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
3102 /* Only use the REG_EQUAL note if it is a constant.
3103 Other things, divide in particular, will cause
3104 problems later if we use them. */
3105 if (note && GET_CODE (XEXP (note, 0)) != EXPR_LIST
3106 && CONSTANT_P (XEXP (note, 0)))
3107 ret = XEXP (note, 0);
3109 ret = SET_SRC (set);
3111 /* We cannot do this if it changes between the
3112 assignment and loop start though. */
3113 if (modified_between_p (ret, insn, loop_start))
3122 /* Find and return register term common to both expressions OP0 and
3123 OP1 or NULL_RTX if no such term exists. Each expression must be a
3124 REG or a PLUS of a REG. */
3127 find_common_reg_term (rtx op0, rtx op1)
3129 if ((REG_P (op0) || GET_CODE (op0) == PLUS)
3130 && (REG_P (op1) || GET_CODE (op1) == PLUS))
3137 if (GET_CODE (op0) == PLUS)
3138 op01 = XEXP (op0, 1), op00 = XEXP (op0, 0);
3140 op01 = const0_rtx, op00 = op0;
3142 if (GET_CODE (op1) == PLUS)
3143 op11 = XEXP (op1, 1), op10 = XEXP (op1, 0);
3145 op11 = const0_rtx, op10 = op1;
3147 /* Find and return common register term if present. */
3148 if (REG_P (op00) && (op00 == op10 || op00 == op11))
3150 else if (REG_P (op01) && (op01 == op10 || op01 == op11))
3154 /* No common register term found. */
3158 /* Determine the loop iterator and calculate the number of loop
3159 iterations. Returns the exact number of loop iterations if it can
3160 be calculated, otherwise returns zero. */
3162 unsigned HOST_WIDE_INT
3163 loop_iterations (struct loop *loop)
3165 struct loop_info *loop_info = LOOP_INFO (loop);
3166 struct loop_ivs *ivs = LOOP_IVS (loop);
3167 rtx comparison, comparison_value;
3168 rtx iteration_var, initial_value, increment, final_value;
3169 enum rtx_code comparison_code;
3171 unsigned HOST_WIDE_INT abs_inc;
3172 unsigned HOST_WIDE_INT abs_diff;
3175 int unsigned_p, compare_dir, final_larger;
3177 struct iv_class *bl;
3179 loop_info->n_iterations = 0;
3180 loop_info->initial_value = 0;
3181 loop_info->initial_equiv_value = 0;
3182 loop_info->comparison_value = 0;
3183 loop_info->final_value = 0;
3184 loop_info->final_equiv_value = 0;
3185 loop_info->increment = 0;
3186 loop_info->iteration_var = 0;
3187 loop_info->unroll_number = 1;
3190 /* We used to use prev_nonnote_insn here, but that fails because it might
3191 accidentally get the branch for a contained loop if the branch for this
3192 loop was deleted. We can only trust branches immediately before the
3194 last_loop_insn = PREV_INSN (loop->end);
3196 /* ??? We should probably try harder to find the jump insn
3197 at the end of the loop. The following code assumes that
3198 the last loop insn is a jump to the top of the loop. */
3199 if (!JUMP_P (last_loop_insn))
3201 if (loop_dump_stream)
3202 fprintf (loop_dump_stream,
3203 "Loop iterations: No final conditional branch found.\n");
3207 /* If there is a more than a single jump to the top of the loop
3208 we cannot (easily) determine the iteration count. */
3209 if (LABEL_NUSES (JUMP_LABEL (last_loop_insn)) > 1)
3211 if (loop_dump_stream)
3212 fprintf (loop_dump_stream,
3213 "Loop iterations: Loop has multiple back edges.\n");
3217 /* Find the iteration variable. If the last insn is a conditional
3218 branch, and the insn before tests a register value, make that the
3219 iteration variable. */
3221 comparison = get_condition_for_loop (loop, last_loop_insn);
3222 if (comparison == 0)
3224 if (loop_dump_stream)
3225 fprintf (loop_dump_stream,
3226 "Loop iterations: No final comparison found.\n");
3230 /* ??? Get_condition may switch position of induction variable and
3231 invariant register when it canonicalizes the comparison. */
3233 comparison_code = GET_CODE (comparison);
3234 iteration_var = XEXP (comparison, 0);
3235 comparison_value = XEXP (comparison, 1);
3237 if (!REG_P (iteration_var))
3239 if (loop_dump_stream)
3240 fprintf (loop_dump_stream,
3241 "Loop iterations: Comparison not against register.\n");
3245 /* The only new registers that are created before loop iterations
3246 are givs made from biv increments or registers created by
3247 load_mems. In the latter case, it is possible that try_copy_prop
3248 will propagate a new pseudo into the old iteration register but
3249 this will be marked by having the REG_USERVAR_P bit set. */
3251 if ((unsigned) REGNO (iteration_var) >= ivs->n_regs
3252 && ! REG_USERVAR_P (iteration_var))
3255 /* Determine the initial value of the iteration variable, and the amount
3256 that it is incremented each loop. Use the tables constructed by
3257 the strength reduction pass to calculate these values. */
3259 /* Clear the result values, in case no answer can be found. */
3263 /* The iteration variable can be either a giv or a biv. Check to see
3264 which it is, and compute the variable's initial value, and increment
3265 value if possible. */
3267 /* If this is a new register, can't handle it since we don't have any
3268 reg_iv_type entry for it. */
3269 if ((unsigned) REGNO (iteration_var) >= ivs->n_regs)
3271 if (loop_dump_stream)
3272 fprintf (loop_dump_stream,
3273 "Loop iterations: No reg_iv_type entry for iteration var.\n");
3277 /* Reject iteration variables larger than the host wide int size, since they
3278 could result in a number of iterations greater than the range of our
3279 `unsigned HOST_WIDE_INT' variable loop_info->n_iterations. */
3280 else if ((GET_MODE_BITSIZE (GET_MODE (iteration_var))
3281 > HOST_BITS_PER_WIDE_INT))
3283 if (loop_dump_stream)
3284 fprintf (loop_dump_stream,
3285 "Loop iterations: Iteration var rejected because mode too large.\n");
3288 else if (GET_MODE_CLASS (GET_MODE (iteration_var)) != MODE_INT)
3290 if (loop_dump_stream)
3291 fprintf (loop_dump_stream,
3292 "Loop iterations: Iteration var not an integer.\n");
3296 /* Try swapping the comparison to identify a suitable iv. */
3297 if (REG_IV_TYPE (ivs, REGNO (iteration_var)) != BASIC_INDUCT
3298 && REG_IV_TYPE (ivs, REGNO (iteration_var)) != GENERAL_INDUCT
3299 && REG_P (comparison_value)
3300 && REGNO (comparison_value) < ivs->n_regs)
3302 rtx temp = comparison_value;
3303 comparison_code = swap_condition (comparison_code);
3304 comparison_value = iteration_var;
3305 iteration_var = temp;
3308 if (REG_IV_TYPE (ivs, REGNO (iteration_var)) == BASIC_INDUCT)
3310 if (REGNO (iteration_var) >= ivs->n_regs)
3313 /* Grab initial value, only useful if it is a constant. */
3314 bl = REG_IV_CLASS (ivs, REGNO (iteration_var));
3315 initial_value = bl->initial_value;
3316 if (!bl->biv->always_executed || bl->biv->maybe_multiple)
3318 if (loop_dump_stream)
3319 fprintf (loop_dump_stream,
3320 "Loop iterations: Basic induction var not set once in each iteration.\n");
3324 increment = biv_total_increment (bl);
3326 else if (REG_IV_TYPE (ivs, REGNO (iteration_var)) == GENERAL_INDUCT)
3328 HOST_WIDE_INT offset = 0;
3329 struct induction *v = REG_IV_INFO (ivs, REGNO (iteration_var));
3330 rtx biv_initial_value;
3332 if (REGNO (v->src_reg) >= ivs->n_regs)
3335 if (!v->always_executed || v->maybe_multiple)
3337 if (loop_dump_stream)
3338 fprintf (loop_dump_stream,
3339 "Loop iterations: General induction var not set once in each iteration.\n");
3343 bl = REG_IV_CLASS (ivs, REGNO (v->src_reg));
3345 /* Increment value is mult_val times the increment value of the biv. */
3347 increment = biv_total_increment (bl);
3350 struct induction *biv_inc;
3352 increment = fold_rtx_mult_add (v->mult_val,
3353 extend_value_for_giv (v, increment),
3354 const0_rtx, v->mode);
3355 /* The caller assumes that one full increment has occurred at the
3356 first loop test. But that's not true when the biv is incremented
3357 after the giv is set (which is the usual case), e.g.:
3358 i = 6; do {;} while (i++ < 9) .
3359 Therefore, we bias the initial value by subtracting the amount of
3360 the increment that occurs between the giv set and the giv test. */
3361 for (biv_inc = bl->biv; biv_inc; biv_inc = biv_inc->next_iv)
3363 if (loop_insn_first_p (v->insn, biv_inc->insn))
3365 if (REG_P (biv_inc->add_val))
3367 if (loop_dump_stream)
3368 fprintf (loop_dump_stream,
3369 "Loop iterations: Basic induction var add_val is REG %d.\n",
3370 REGNO (biv_inc->add_val));
3374 /* If we have already counted it, skip it. */
3378 offset -= INTVAL (biv_inc->add_val);
3382 if (loop_dump_stream)
3383 fprintf (loop_dump_stream,
3384 "Loop iterations: Giv iterator, initial value bias %ld.\n",
3387 /* Initial value is mult_val times the biv's initial value plus
3388 add_val. Only useful if it is a constant. */
3389 biv_initial_value = extend_value_for_giv (v, bl->initial_value);
3391 = fold_rtx_mult_add (v->mult_val,
3392 plus_constant (biv_initial_value, offset),
3393 v->add_val, v->mode);
3397 if (loop_dump_stream)
3398 fprintf (loop_dump_stream,
3399 "Loop iterations: Not basic or general induction var.\n");
3403 if (initial_value == 0)
3408 switch (comparison_code)
3423 /* Cannot determine loop iterations with this case. */
3443 /* If the comparison value is an invariant register, then try to find
3444 its value from the insns before the start of the loop. */
3446 final_value = comparison_value;
3447 if (REG_P (comparison_value)
3448 && loop_invariant_p (loop, comparison_value))
3450 final_value = loop_find_equiv_value (loop, comparison_value);
3452 /* If we don't get an invariant final value, we are better
3453 off with the original register. */
3454 if (! loop_invariant_p (loop, final_value))
3455 final_value = comparison_value;
3458 /* Calculate the approximate final value of the induction variable
3459 (on the last successful iteration). The exact final value
3460 depends on the branch operator, and increment sign. It will be
3461 wrong if the iteration variable is not incremented by one each
3462 time through the loop and (comparison_value + off_by_one -
3463 initial_value) % increment != 0.
3464 ??? Note that the final_value may overflow and thus final_larger
3465 will be bogus. A potentially infinite loop will be classified
3466 as immediate, e.g. for (i = 0x7ffffff0; i <= 0x7fffffff; i++) */
3468 final_value = plus_constant (final_value, off_by_one);
3470 /* Save the calculated values describing this loop's bounds, in case
3471 precondition_loop_p will need them later. These values can not be
3472 recalculated inside precondition_loop_p because strength reduction
3473 optimizations may obscure the loop's structure.
3475 These values are only required by precondition_loop_p and insert_bct
3476 whenever the number of iterations cannot be computed at compile time.
3477 Only the difference between final_value and initial_value is
3478 important. Note that final_value is only approximate. */
3479 loop_info->initial_value = initial_value;
3480 loop_info->comparison_value = comparison_value;
3481 loop_info->final_value = plus_constant (comparison_value, off_by_one);
3482 loop_info->increment = increment;
3483 loop_info->iteration_var = iteration_var;
3484 loop_info->comparison_code = comparison_code;
3487 /* Try to determine the iteration count for loops such
3488 as (for i = init; i < init + const; i++). When running the
3489 loop optimization twice, the first pass often converts simple
3490 loops into this form. */
3492 if (REG_P (initial_value))
3498 reg1 = initial_value;
3499 if (GET_CODE (final_value) == PLUS)
3500 reg2 = XEXP (final_value, 0), const2 = XEXP (final_value, 1);
3502 reg2 = final_value, const2 = const0_rtx;
3504 /* Check for initial_value = reg1, final_value = reg2 + const2,
3505 where reg1 != reg2. */
3506 if (REG_P (reg2) && reg2 != reg1)
3510 /* Find what reg1 is equivalent to. Hopefully it will
3511 either be reg2 or reg2 plus a constant. */
3512 temp = loop_find_equiv_value (loop, reg1);
3514 if (find_common_reg_term (temp, reg2))
3515 initial_value = temp;
3516 else if (loop_invariant_p (loop, reg2))
3518 /* Find what reg2 is equivalent to. Hopefully it will
3519 either be reg1 or reg1 plus a constant. Let's ignore
3520 the latter case for now since it is not so common. */
3521 temp = loop_find_equiv_value (loop, reg2);
3523 if (temp == loop_info->iteration_var)
3524 temp = initial_value;
3526 final_value = (const2 == const0_rtx)
3527 ? reg1 : gen_rtx_PLUS (GET_MODE (reg1), reg1, const2);
3532 loop_info->initial_equiv_value = initial_value;
3533 loop_info->final_equiv_value = final_value;
3535 /* For EQ comparison loops, we don't have a valid final value.
3536 Check this now so that we won't leave an invalid value if we
3537 return early for any other reason. */
3538 if (comparison_code == EQ)
3539 loop_info->final_equiv_value = loop_info->final_value = 0;
3543 if (loop_dump_stream)
3544 fprintf (loop_dump_stream,
3545 "Loop iterations: Increment value can't be calculated.\n");
3549 if (GET_CODE (increment) != CONST_INT)
3551 /* If we have a REG, check to see if REG holds a constant value. */
3552 /* ??? Other RTL, such as (neg (reg)) is possible here, but it isn't
3553 clear if it is worthwhile to try to handle such RTL. */
3554 if (REG_P (increment) || GET_CODE (increment) == SUBREG)
3555 increment = loop_find_equiv_value (loop, increment);
3557 if (GET_CODE (increment) != CONST_INT)
3559 if (loop_dump_stream)
3561 fprintf (loop_dump_stream,
3562 "Loop iterations: Increment value not constant ");
3563 print_simple_rtl (loop_dump_stream, increment);
3564 fprintf (loop_dump_stream, ".\n");
3568 loop_info->increment = increment;
3571 if (GET_CODE (initial_value) != CONST_INT)
3573 if (loop_dump_stream)
3575 fprintf (loop_dump_stream,
3576 "Loop iterations: Initial value not constant ");
3577 print_simple_rtl (loop_dump_stream, initial_value);
3578 fprintf (loop_dump_stream, ".\n");
3582 else if (GET_CODE (final_value) != CONST_INT)
3584 if (loop_dump_stream)
3586 fprintf (loop_dump_stream,
3587 "Loop iterations: Final value not constant ");
3588 print_simple_rtl (loop_dump_stream, final_value);
3589 fprintf (loop_dump_stream, ".\n");
3593 else if (comparison_code == EQ)
3597 if (loop_dump_stream)
3598 fprintf (loop_dump_stream, "Loop iterations: EQ comparison loop.\n");
3600 inc_once = gen_int_mode (INTVAL (initial_value) + INTVAL (increment),
3601 GET_MODE (iteration_var));
3603 if (inc_once == final_value)
3605 /* The iterator value once through the loop is equal to the
3606 comparison value. Either we have an infinite loop, or
3607 we'll loop twice. */
3608 if (increment == const0_rtx)
3610 loop_info->n_iterations = 2;
3613 loop_info->n_iterations = 1;
3615 if (GET_CODE (loop_info->initial_value) == CONST_INT)
3616 loop_info->final_value
3617 = gen_int_mode ((INTVAL (loop_info->initial_value)
3618 + loop_info->n_iterations * INTVAL (increment)),
3619 GET_MODE (iteration_var));
3621 loop_info->final_value
3622 = plus_constant (loop_info->initial_value,
3623 loop_info->n_iterations * INTVAL (increment));
3624 loop_info->final_equiv_value
3625 = gen_int_mode ((INTVAL (initial_value)
3626 + loop_info->n_iterations * INTVAL (increment)),
3627 GET_MODE (iteration_var));
3628 return loop_info->n_iterations;
3631 /* Final_larger is 1 if final larger, 0 if they are equal, otherwise -1. */
3634 = ((unsigned HOST_WIDE_INT) INTVAL (final_value)
3635 > (unsigned HOST_WIDE_INT) INTVAL (initial_value))
3636 - ((unsigned HOST_WIDE_INT) INTVAL (final_value)
3637 < (unsigned HOST_WIDE_INT) INTVAL (initial_value));
3639 final_larger = (INTVAL (final_value) > INTVAL (initial_value))
3640 - (INTVAL (final_value) < INTVAL (initial_value));
3642 if (INTVAL (increment) > 0)
3644 else if (INTVAL (increment) == 0)
3649 /* There are 27 different cases: compare_dir = -1, 0, 1;
3650 final_larger = -1, 0, 1; increment_dir = -1, 0, 1.
3651 There are 4 normal cases, 4 reverse cases (where the iteration variable
3652 will overflow before the loop exits), 4 infinite loop cases, and 15
3653 immediate exit (0 or 1 iteration depending on loop type) cases.
3654 Only try to optimize the normal cases. */
3656 /* (compare_dir/final_larger/increment_dir)
3657 Normal cases: (0/-1/-1), (0/1/1), (-1/-1/-1), (1/1/1)
3658 Reverse cases: (0/-1/1), (0/1/-1), (-1/-1/1), (1/1/-1)
3659 Infinite loops: (0/-1/0), (0/1/0), (-1/-1/0), (1/1/0)
3660 Immediate exit: (0/0/X), (-1/0/X), (-1/1/X), (1/0/X), (1/-1/X) */
3662 /* ?? If the meaning of reverse loops (where the iteration variable
3663 will overflow before the loop exits) is undefined, then could
3664 eliminate all of these special checks, and just always assume
3665 the loops are normal/immediate/infinite. Note that this means
3666 the sign of increment_dir does not have to be known. Also,
3667 since it does not really hurt if immediate exit loops or infinite loops
3668 are optimized, then that case could be ignored also, and hence all
3669 loops can be optimized.
3671 According to ANSI Spec, the reverse loop case result is undefined,
3672 because the action on overflow is undefined.
3674 See also the special test for NE loops below. */
3676 if (final_larger == increment_dir && final_larger != 0
3677 && (final_larger == compare_dir || compare_dir == 0))
3682 if (loop_dump_stream)
3683 fprintf (loop_dump_stream, "Loop iterations: Not normal loop.\n");
3687 /* Calculate the number of iterations, final_value is only an approximation,
3688 so correct for that. Note that abs_diff and n_iterations are
3689 unsigned, because they can be as large as 2^n - 1. */
3691 inc = INTVAL (increment);
3694 abs_diff = INTVAL (final_value) - INTVAL (initial_value);
3699 abs_diff = INTVAL (initial_value) - INTVAL (final_value);
3705 /* Given that iteration_var is going to iterate over its own mode,
3706 not HOST_WIDE_INT, disregard higher bits that might have come
3707 into the picture due to sign extension of initial and final
3709 abs_diff &= ((unsigned HOST_WIDE_INT) 1
3710 << (GET_MODE_BITSIZE (GET_MODE (iteration_var)) - 1)
3713 /* For NE tests, make sure that the iteration variable won't miss
3714 the final value. If abs_diff mod abs_incr is not zero, then the
3715 iteration variable will overflow before the loop exits, and we
3716 can not calculate the number of iterations. */
3717 if (compare_dir == 0 && (abs_diff % abs_inc) != 0)
3720 /* Note that the number of iterations could be calculated using
3721 (abs_diff + abs_inc - 1) / abs_inc, provided care was taken to
3722 handle potential overflow of the summation. */
3723 loop_info->n_iterations = abs_diff / abs_inc + ((abs_diff % abs_inc) != 0);
3724 return loop_info->n_iterations;
3727 /* Replace uses of split bivs with their split pseudo register. This is
3728 for original instructions which remain after loop unrolling without
3732 remap_split_bivs (struct loop *loop, rtx x)
3734 struct loop_ivs *ivs = LOOP_IVS (loop);
3742 code = GET_CODE (x);
3757 /* If non-reduced/final-value givs were split, then this would also
3758 have to remap those givs also. */
3760 if (REGNO (x) < ivs->n_regs
3761 && REG_IV_TYPE (ivs, REGNO (x)) == BASIC_INDUCT)
3762 return REG_IV_CLASS (ivs, REGNO (x))->biv->src_reg;
3769 fmt = GET_RTX_FORMAT (code);
3770 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3773 XEXP (x, i) = remap_split_bivs (loop, XEXP (x, i));
3774 else if (fmt[i] == 'E')
3777 for (j = 0; j < XVECLEN (x, i); j++)
3778 XVECEXP (x, i, j) = remap_split_bivs (loop, XVECEXP (x, i, j));
3784 /* If FIRST_UID is a set of REGNO, and FIRST_UID dominates LAST_UID (e.g.
3785 FIST_UID is always executed if LAST_UID is), then return 1. Otherwise
3786 return 0. COPY_START is where we can start looking for the insns
3787 FIRST_UID and LAST_UID. COPY_END is where we stop looking for these
3790 If there is no JUMP_INSN between LOOP_START and FIRST_UID, then FIRST_UID
3791 must dominate LAST_UID.
3793 If there is a CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID
3794 may not dominate LAST_UID.
3796 If there is no CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID
3797 must dominate LAST_UID. */
3800 set_dominates_use (int regno, int first_uid, int last_uid, rtx copy_start,
3803 int passed_jump = 0;
3804 rtx p = NEXT_INSN (copy_start);
3806 while (INSN_UID (p) != first_uid)
3810 /* Could not find FIRST_UID. */
3816 /* Verify that FIRST_UID is an insn that entirely sets REGNO. */
3817 if (! INSN_P (p) || ! dead_or_set_regno_p (p, regno))
3820 /* FIRST_UID is always executed. */
3821 if (passed_jump == 0)
3824 while (INSN_UID (p) != last_uid)
3826 /* If we see a CODE_LABEL between FIRST_UID and LAST_UID, then we
3827 can not be sure that FIRST_UID dominates LAST_UID. */
3830 /* Could not find LAST_UID, but we reached the end of the loop, so
3832 else if (p == copy_end)
3837 /* FIRST_UID is always executed if LAST_UID is executed. */