1 /* Try to unroll loops, and split induction variables.
2 Copyright (C) 1992, 1993, 1994, 1995, 1997, 1998, 1999, 2000, 2001
3 Free Software Foundation, Inc.
4 Contributed by James E. Wilson, Cygnus Support/UC Berkeley.
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify it under
9 the terms of the GNU General Public License as published by the Free
10 Software Foundation; either version 2, or (at your option) any later
13 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
14 WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING. If not, write to the Free
20 Software Foundation, 59 Temple Place - Suite 330, Boston, MA
23 /* Try to unroll a loop, and split induction variables.
25 Loops for which the number of iterations can be calculated exactly are
26 handled specially. If the number of iterations times the insn_count is
27 less than MAX_UNROLLED_INSNS, then the loop is unrolled completely.
28 Otherwise, we try to unroll the loop a number of times modulo the number
29 of iterations, so that only one exit test will be needed. It is unrolled
30 a number of times approximately equal to MAX_UNROLLED_INSNS divided by
33 Otherwise, if the number of iterations can be calculated exactly at
34 run time, and the loop is always entered at the top, then we try to
35 precondition the loop. That is, at run time, calculate how many times
36 the loop will execute, and then execute the loop body a few times so
37 that the remaining iterations will be some multiple of 4 (or 2 if the
38 loop is large). Then fall through to a loop unrolled 4 (or 2) times,
39 with only one exit test needed at the end of the loop.
41 Otherwise, if the number of iterations can not be calculated exactly,
42 not even at run time, then we still unroll the loop a number of times
43 approximately equal to MAX_UNROLLED_INSNS divided by the insn count,
44 but there must be an exit test after each copy of the loop body.
46 For each induction variable, which is dead outside the loop (replaceable)
47 or for which we can easily calculate the final value, if we can easily
48 calculate its value at each place where it is set as a function of the
49 current loop unroll count and the variable's value at loop entry, then
50 the induction variable is split into `N' different variables, one for
51 each copy of the loop body. One variable is live across the backward
52 branch, and the others are all calculated as a function of this variable.
53 This helps eliminate data dependencies, and leads to further opportunities
56 /* Possible improvements follow: */
58 /* ??? Add an extra pass somewhere to determine whether unrolling will
59 give any benefit. E.g. after generating all unrolled insns, compute the
60 cost of all insns and compare against cost of insns in rolled loop.
62 - On traditional architectures, unrolling a non-constant bound loop
63 is a win if there is a giv whose only use is in memory addresses, the
64 memory addresses can be split, and hence giv increments can be
66 - It is also a win if the loop is executed many times, and preconditioning
67 can be performed for the loop.
68 Add code to check for these and similar cases. */
70 /* ??? Improve control of which loops get unrolled. Could use profiling
71 info to only unroll the most commonly executed loops. Perhaps have
72 a user specifyable option to control the amount of code expansion,
73 or the percent of loops to consider for unrolling. Etc. */
75 /* ??? Look at the register copies inside the loop to see if they form a
76 simple permutation. If so, iterate the permutation until it gets back to
77 the start state. This is how many times we should unroll the loop, for
78 best results, because then all register copies can be eliminated.
79 For example, the lisp nreverse function should be unrolled 3 times
88 ??? The number of times to unroll the loop may also be based on data
89 references in the loop. For example, if we have a loop that references
90 x[i-1], x[i], and x[i+1], we should unroll it a multiple of 3 times. */
92 /* ??? Add some simple linear equation solving capability so that we can
93 determine the number of loop iterations for more complex loops.
94 For example, consider this loop from gdb
95 #define SWAP_TARGET_AND_HOST(buffer,len)
98 char *p = (char *) buffer;
99 char *q = ((char *) buffer) + len - 1;
100 int iterations = (len + 1) >> 1;
102 for (p; p < q; p++, q--;)
110 start value = p = &buffer + current_iteration
111 end value = q = &buffer + len - 1 - current_iteration
112 Given the loop exit test of "p < q", then there must be "q - p" iterations,
113 set equal to zero and solve for number of iterations:
114 q - p = len - 1 - 2*current_iteration = 0
115 current_iteration = (len - 1) / 2
116 Hence, there are (len - 1) / 2 (rounded up to the nearest integer)
117 iterations of this loop. */
119 /* ??? Currently, no labels are marked as loop invariant when doing loop
120 unrolling. This is because an insn inside the loop, that loads the address
121 of a label inside the loop into a register, could be moved outside the loop
122 by the invariant code motion pass if labels were invariant. If the loop
123 is subsequently unrolled, the code will be wrong because each unrolled
124 body of the loop will use the same address, whereas each actually needs a
125 different address. A case where this happens is when a loop containing
126 a switch statement is unrolled.
128 It would be better to let labels be considered invariant. When we
129 unroll loops here, check to see if any insns using a label local to the
130 loop were moved before the loop. If so, then correct the problem, by
131 moving the insn back into the loop, or perhaps replicate the insn before
132 the loop, one copy for each time the loop is unrolled. */
134 /* The prime factors looked for when trying to unroll a loop by some
135 number which is modulo the total number of iterations. Just checking
136 for these 4 prime factors will find at least one factor for 75% of
137 all numbers theoretically. Practically speaking, this will succeed
138 almost all of the time since loops are generally a multiple of 2
141 #define NUM_FACTORS 4
143 struct _factor { int factor, count; }
144 factors[NUM_FACTORS] = { {2, 0}, {3, 0}, {5, 0}, {7, 0}};
146 /* Describes the different types of loop unrolling performed. */
159 #include "insn-config.h"
160 #include "integrate.h"
164 #include "function.h"
168 #include "hard-reg-set.h"
169 #include "basic-block.h"
172 /* This controls which loops are unrolled, and by how much we unroll
175 #ifndef MAX_UNROLLED_INSNS
176 #define MAX_UNROLLED_INSNS 100
179 /* Indexed by register number, if non-zero, then it contains a pointer
180 to a struct induction for a DEST_REG giv which has been combined with
181 one of more address givs. This is needed because whenever such a DEST_REG
182 giv is modified, we must modify the value of all split address givs
183 that were combined with this DEST_REG giv. */
185 static struct induction **addr_combined_regs;
187 /* Indexed by register number, if this is a splittable induction variable,
188 then this will hold the current value of the register, which depends on the
191 static rtx *splittable_regs;
193 /* Indexed by register number, if this is a splittable induction variable,
194 then this will hold the number of instructions in the loop that modify
195 the induction variable. Used to ensure that only the last insn modifying
196 a split iv will update the original iv of the dest. */
198 static int *splittable_regs_updates;
200 /* Forward declarations. */
202 static void init_reg_map PARAMS ((struct inline_remap *, int));
203 static rtx calculate_giv_inc PARAMS ((rtx, rtx, unsigned int));
204 static rtx initial_reg_note_copy PARAMS ((rtx, struct inline_remap *));
205 static void final_reg_note_copy PARAMS ((rtx *, struct inline_remap *));
206 static void copy_loop_body PARAMS ((struct loop *, rtx, rtx,
207 struct inline_remap *, rtx, int,
208 enum unroll_types, rtx, rtx, rtx, rtx));
209 static int find_splittable_regs PARAMS ((const struct loop *,
210 enum unroll_types, int));
211 static int find_splittable_givs PARAMS ((const struct loop *,
212 struct iv_class *, enum unroll_types,
214 static int reg_dead_after_loop PARAMS ((const struct loop *, rtx));
215 static rtx fold_rtx_mult_add PARAMS ((rtx, rtx, rtx, enum machine_mode));
216 static int verify_addresses PARAMS ((struct induction *, rtx, int));
217 static rtx remap_split_bivs PARAMS ((struct loop *, rtx));
218 static rtx find_common_reg_term PARAMS ((rtx, rtx));
219 static rtx subtract_reg_term PARAMS ((rtx, rtx));
220 static rtx loop_find_equiv_value PARAMS ((const struct loop *, rtx));
221 static rtx ujump_to_loop_cont PARAMS ((rtx, rtx));
223 /* Try to unroll one loop and split induction variables in the loop.
225 The loop is described by the arguments LOOP and INSN_COUNT.
226 STRENGTH_REDUCTION_P indicates whether information generated in the
227 strength reduction pass is available.
229 This function is intended to be called from within `strength_reduce'
233 unroll_loop (loop, insn_count, strength_reduce_p)
236 int strength_reduce_p;
238 struct loop_info *loop_info = LOOP_INFO (loop);
239 struct loop_ivs *ivs = LOOP_IVS (loop);
242 unsigned HOST_WIDE_INT temp;
243 int unroll_number = 1;
244 rtx copy_start, copy_end;
245 rtx insn, sequence, pattern, tem;
246 int max_labelno, max_insnno;
248 struct inline_remap *map;
249 char *local_label = NULL;
251 unsigned int max_local_regnum;
252 unsigned int maxregnum;
256 int splitting_not_safe = 0;
257 enum unroll_types unroll_type = UNROLL_NAIVE;
258 int loop_preconditioned = 0;
260 /* This points to the last real insn in the loop, which should be either
261 a JUMP_INSN (for conditional jumps) or a BARRIER (for unconditional
264 rtx loop_start = loop->start;
265 rtx loop_end = loop->end;
267 /* Don't bother unrolling huge loops. Since the minimum factor is
268 two, loops greater than one half of MAX_UNROLLED_INSNS will never
270 if (insn_count > MAX_UNROLLED_INSNS / 2)
272 if (loop_dump_stream)
273 fprintf (loop_dump_stream, "Unrolling failure: Loop too big.\n");
277 /* When emitting debugger info, we can't unroll loops with unequal numbers
278 of block_beg and block_end notes, because that would unbalance the block
279 structure of the function. This can happen as a result of the
280 "if (foo) bar; else break;" optimization in jump.c. */
281 /* ??? Gcc has a general policy that -g is never supposed to change the code
282 that the compiler emits, so we must disable this optimization always,
283 even if debug info is not being output. This is rare, so this should
284 not be a significant performance problem. */
286 if (1 /* write_symbols != NO_DEBUG */)
288 int block_begins = 0;
291 for (insn = loop_start; insn != loop_end; insn = NEXT_INSN (insn))
293 if (GET_CODE (insn) == NOTE)
295 if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_BLOCK_BEG)
297 else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_BLOCK_END)
299 if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_EH_REGION_BEG
300 || NOTE_LINE_NUMBER (insn) == NOTE_INSN_EH_REGION_END)
302 /* Note, would be nice to add code to unroll EH
303 regions, but until that time, we punt (don't
304 unroll). For the proper way of doing it, see
305 expand_inline_function. */
307 if (loop_dump_stream)
308 fprintf (loop_dump_stream,
309 "Unrolling failure: cannot unroll EH regions.\n");
315 if (block_begins != block_ends)
317 if (loop_dump_stream)
318 fprintf (loop_dump_stream,
319 "Unrolling failure: Unbalanced block notes.\n");
324 /* Determine type of unroll to perform. Depends on the number of iterations
325 and the size of the loop. */
327 /* If there is no strength reduce info, then set
328 loop_info->n_iterations to zero. This can happen if
329 strength_reduce can't find any bivs in the loop. A value of zero
330 indicates that the number of iterations could not be calculated. */
332 if (! strength_reduce_p)
333 loop_info->n_iterations = 0;
335 if (loop_dump_stream && loop_info->n_iterations > 0)
337 fputs ("Loop unrolling: ", loop_dump_stream);
338 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC,
339 loop_info->n_iterations);
340 fputs (" iterations.\n", loop_dump_stream);
343 /* Find and save a pointer to the last nonnote insn in the loop. */
345 last_loop_insn = prev_nonnote_insn (loop_end);
347 /* Calculate how many times to unroll the loop. Indicate whether or
348 not the loop is being completely unrolled. */
350 if (loop_info->n_iterations == 1)
352 /* Handle the case where the loop begins with an unconditional
353 jump to the loop condition. Make sure to delete the jump
354 insn, otherwise the loop body will never execute. */
356 rtx ujump = ujump_to_loop_cont (loop->start, loop->cont);
358 delete_related_insns (ujump);
360 /* If number of iterations is exactly 1, then eliminate the compare and
361 branch at the end of the loop since they will never be taken.
362 Then return, since no other action is needed here. */
364 /* If the last instruction is not a BARRIER or a JUMP_INSN, then
365 don't do anything. */
367 if (GET_CODE (last_loop_insn) == BARRIER)
369 /* Delete the jump insn. This will delete the barrier also. */
370 delete_related_insns (PREV_INSN (last_loop_insn));
372 else if (GET_CODE (last_loop_insn) == JUMP_INSN)
375 rtx prev = PREV_INSN (last_loop_insn);
377 delete_related_insns (last_loop_insn);
379 /* The immediately preceding insn may be a compare which must be
381 if (only_sets_cc0_p (prev))
382 delete_related_insns (prev);
386 /* Remove the loop notes since this is no longer a loop. */
388 delete_related_insns (loop->vtop);
390 delete_related_insns (loop->cont);
392 delete_related_insns (loop_start);
394 delete_related_insns (loop_end);
398 else if (loop_info->n_iterations > 0
399 /* Avoid overflow in the next expression. */
400 && loop_info->n_iterations < MAX_UNROLLED_INSNS
401 && loop_info->n_iterations * insn_count < MAX_UNROLLED_INSNS)
403 unroll_number = loop_info->n_iterations;
404 unroll_type = UNROLL_COMPLETELY;
406 else if (loop_info->n_iterations > 0)
408 /* Try to factor the number of iterations. Don't bother with the
409 general case, only using 2, 3, 5, and 7 will get 75% of all
410 numbers theoretically, and almost all in practice. */
412 for (i = 0; i < NUM_FACTORS; i++)
413 factors[i].count = 0;
415 temp = loop_info->n_iterations;
416 for (i = NUM_FACTORS - 1; i >= 0; i--)
417 while (temp % factors[i].factor == 0)
420 temp = temp / factors[i].factor;
423 /* Start with the larger factors first so that we generally
424 get lots of unrolling. */
428 for (i = 3; i >= 0; i--)
429 while (factors[i].count--)
431 if (temp * factors[i].factor < MAX_UNROLLED_INSNS)
433 unroll_number *= factors[i].factor;
434 temp *= factors[i].factor;
440 /* If we couldn't find any factors, then unroll as in the normal
442 if (unroll_number == 1)
444 if (loop_dump_stream)
445 fprintf (loop_dump_stream, "Loop unrolling: No factors found.\n");
448 unroll_type = UNROLL_MODULO;
451 /* Default case, calculate number of times to unroll loop based on its
453 if (unroll_type == UNROLL_NAIVE)
455 if (8 * insn_count < MAX_UNROLLED_INSNS)
457 else if (4 * insn_count < MAX_UNROLLED_INSNS)
463 /* Now we know how many times to unroll the loop. */
465 if (loop_dump_stream)
466 fprintf (loop_dump_stream, "Unrolling loop %d times.\n", unroll_number);
468 if (unroll_type == UNROLL_COMPLETELY || unroll_type == UNROLL_MODULO)
470 /* Loops of these types can start with jump down to the exit condition
471 in rare circumstances.
473 Consider a pair of nested loops where the inner loop is part
474 of the exit code for the outer loop.
476 In this case jump.c will not duplicate the exit test for the outer
477 loop, so it will start with a jump to the exit code.
479 Then consider if the inner loop turns out to iterate once and
480 only once. We will end up deleting the jumps associated with
481 the inner loop. However, the loop notes are not removed from
482 the instruction stream.
484 And finally assume that we can compute the number of iterations
487 In this case unroll may want to unroll the outer loop even though
488 it starts with a jump to the outer loop's exit code.
490 We could try to optimize this case, but it hardly seems worth it.
491 Just return without unrolling the loop in such cases. */
494 while (GET_CODE (insn) != CODE_LABEL && GET_CODE (insn) != JUMP_INSN)
495 insn = NEXT_INSN (insn);
496 if (GET_CODE (insn) == JUMP_INSN)
500 if (unroll_type == UNROLL_COMPLETELY)
502 /* Completely unrolling the loop: Delete the compare and branch at
503 the end (the last two instructions). This delete must done at the
504 very end of loop unrolling, to avoid problems with calls to
505 back_branch_in_range_p, which is called by find_splittable_regs.
506 All increments of splittable bivs/givs are changed to load constant
509 copy_start = loop_start;
511 /* Set insert_before to the instruction immediately after the JUMP_INSN
512 (or BARRIER), so that any NOTEs between the JUMP_INSN and the end of
513 the loop will be correctly handled by copy_loop_body. */
514 insert_before = NEXT_INSN (last_loop_insn);
516 /* Set copy_end to the insn before the jump at the end of the loop. */
517 if (GET_CODE (last_loop_insn) == BARRIER)
518 copy_end = PREV_INSN (PREV_INSN (last_loop_insn));
519 else if (GET_CODE (last_loop_insn) == JUMP_INSN)
521 copy_end = PREV_INSN (last_loop_insn);
523 /* The instruction immediately before the JUMP_INSN may be a compare
524 instruction which we do not want to copy. */
525 if (sets_cc0_p (PREV_INSN (copy_end)))
526 copy_end = PREV_INSN (copy_end);
531 /* We currently can't unroll a loop if it doesn't end with a
532 JUMP_INSN. There would need to be a mechanism that recognizes
533 this case, and then inserts a jump after each loop body, which
534 jumps to after the last loop body. */
535 if (loop_dump_stream)
536 fprintf (loop_dump_stream,
537 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
541 else if (unroll_type == UNROLL_MODULO)
543 /* Partially unrolling the loop: The compare and branch at the end
544 (the last two instructions) must remain. Don't copy the compare
545 and branch instructions at the end of the loop. Insert the unrolled
546 code immediately before the compare/branch at the end so that the
547 code will fall through to them as before. */
549 copy_start = loop_start;
551 /* Set insert_before to the jump insn at the end of the loop.
552 Set copy_end to before the jump insn at the end of the loop. */
553 if (GET_CODE (last_loop_insn) == BARRIER)
555 insert_before = PREV_INSN (last_loop_insn);
556 copy_end = PREV_INSN (insert_before);
558 else if (GET_CODE (last_loop_insn) == JUMP_INSN)
560 insert_before = last_loop_insn;
562 /* The instruction immediately before the JUMP_INSN may be a compare
563 instruction which we do not want to copy or delete. */
564 if (sets_cc0_p (PREV_INSN (insert_before)))
565 insert_before = PREV_INSN (insert_before);
567 copy_end = PREV_INSN (insert_before);
571 /* We currently can't unroll a loop if it doesn't end with a
572 JUMP_INSN. There would need to be a mechanism that recognizes
573 this case, and then inserts a jump after each loop body, which
574 jumps to after the last loop body. */
575 if (loop_dump_stream)
576 fprintf (loop_dump_stream,
577 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
583 /* Normal case: Must copy the compare and branch instructions at the
586 if (GET_CODE (last_loop_insn) == BARRIER)
588 /* Loop ends with an unconditional jump and a barrier.
589 Handle this like above, don't copy jump and barrier.
590 This is not strictly necessary, but doing so prevents generating
591 unconditional jumps to an immediately following label.
593 This will be corrected below if the target of this jump is
594 not the start_label. */
596 insert_before = PREV_INSN (last_loop_insn);
597 copy_end = PREV_INSN (insert_before);
599 else if (GET_CODE (last_loop_insn) == JUMP_INSN)
601 /* Set insert_before to immediately after the JUMP_INSN, so that
602 NOTEs at the end of the loop will be correctly handled by
604 insert_before = NEXT_INSN (last_loop_insn);
605 copy_end = last_loop_insn;
609 /* We currently can't unroll a loop if it doesn't end with a
610 JUMP_INSN. There would need to be a mechanism that recognizes
611 this case, and then inserts a jump after each loop body, which
612 jumps to after the last loop body. */
613 if (loop_dump_stream)
614 fprintf (loop_dump_stream,
615 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
619 /* If copying exit test branches because they can not be eliminated,
620 then must convert the fall through case of the branch to a jump past
621 the end of the loop. Create a label to emit after the loop and save
622 it for later use. Do not use the label after the loop, if any, since
623 it might be used by insns outside the loop, or there might be insns
624 added before it later by final_[bg]iv_value which must be after
625 the real exit label. */
626 exit_label = gen_label_rtx ();
629 while (GET_CODE (insn) != CODE_LABEL && GET_CODE (insn) != JUMP_INSN)
630 insn = NEXT_INSN (insn);
632 if (GET_CODE (insn) == JUMP_INSN)
634 /* The loop starts with a jump down to the exit condition test.
635 Start copying the loop after the barrier following this
637 copy_start = NEXT_INSN (insn);
639 /* Splitting induction variables doesn't work when the loop is
640 entered via a jump to the bottom, because then we end up doing
641 a comparison against a new register for a split variable, but
642 we did not execute the set insn for the new register because
643 it was skipped over. */
644 splitting_not_safe = 1;
645 if (loop_dump_stream)
646 fprintf (loop_dump_stream,
647 "Splitting not safe, because loop not entered at top.\n");
650 copy_start = loop_start;
653 /* This should always be the first label in the loop. */
654 start_label = NEXT_INSN (copy_start);
655 /* There may be a line number note and/or a loop continue note here. */
656 while (GET_CODE (start_label) == NOTE)
657 start_label = NEXT_INSN (start_label);
658 if (GET_CODE (start_label) != CODE_LABEL)
660 /* This can happen as a result of jump threading. If the first insns in
661 the loop test the same condition as the loop's backward jump, or the
662 opposite condition, then the backward jump will be modified to point
663 to elsewhere, and the loop's start label is deleted.
665 This case currently can not be handled by the loop unrolling code. */
667 if (loop_dump_stream)
668 fprintf (loop_dump_stream,
669 "Unrolling failure: unknown insns between BEG note and loop label.\n");
672 if (LABEL_NAME (start_label))
674 /* The jump optimization pass must have combined the original start label
675 with a named label for a goto. We can't unroll this case because
676 jumps which go to the named label must be handled differently than
677 jumps to the loop start, and it is impossible to differentiate them
679 if (loop_dump_stream)
680 fprintf (loop_dump_stream,
681 "Unrolling failure: loop start label is gone\n");
685 if (unroll_type == UNROLL_NAIVE
686 && GET_CODE (last_loop_insn) == BARRIER
687 && GET_CODE (PREV_INSN (last_loop_insn)) == JUMP_INSN
688 && start_label != JUMP_LABEL (PREV_INSN (last_loop_insn)))
690 /* In this case, we must copy the jump and barrier, because they will
691 not be converted to jumps to an immediately following label. */
693 insert_before = NEXT_INSN (last_loop_insn);
694 copy_end = last_loop_insn;
697 if (unroll_type == UNROLL_NAIVE
698 && GET_CODE (last_loop_insn) == JUMP_INSN
699 && start_label != JUMP_LABEL (last_loop_insn))
701 /* ??? The loop ends with a conditional branch that does not branch back
702 to the loop start label. In this case, we must emit an unconditional
703 branch to the loop exit after emitting the final branch.
704 copy_loop_body does not have support for this currently, so we
705 give up. It doesn't seem worthwhile to unroll anyways since
706 unrolling would increase the number of branch instructions
708 if (loop_dump_stream)
709 fprintf (loop_dump_stream,
710 "Unrolling failure: final conditional branch not to loop start\n");
714 /* Allocate a translation table for the labels and insn numbers.
715 They will be filled in as we copy the insns in the loop. */
717 max_labelno = max_label_num ();
718 max_insnno = get_max_uid ();
720 /* Various paths through the unroll code may reach the "egress" label
721 without initializing fields within the map structure.
723 To be safe, we use xcalloc to zero the memory. */
724 map = (struct inline_remap *) xcalloc (1, sizeof (struct inline_remap));
726 /* Allocate the label map. */
730 map->label_map = (rtx *) xmalloc (max_labelno * sizeof (rtx));
732 local_label = (char *) xcalloc (max_labelno, sizeof (char));
735 /* Search the loop and mark all local labels, i.e. the ones which have to
736 be distinct labels when copied. For all labels which might be
737 non-local, set their label_map entries to point to themselves.
738 If they happen to be local their label_map entries will be overwritten
739 before the loop body is copied. The label_map entries for local labels
740 will be set to a different value each time the loop body is copied. */
742 for (insn = copy_start; insn != loop_end; insn = NEXT_INSN (insn))
746 if (GET_CODE (insn) == CODE_LABEL)
747 local_label[CODE_LABEL_NUMBER (insn)] = 1;
748 else if (GET_CODE (insn) == JUMP_INSN)
750 if (JUMP_LABEL (insn))
751 set_label_in_map (map,
752 CODE_LABEL_NUMBER (JUMP_LABEL (insn)),
754 else if (GET_CODE (PATTERN (insn)) == ADDR_VEC
755 || GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC)
757 rtx pat = PATTERN (insn);
758 int diff_vec_p = GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC;
759 int len = XVECLEN (pat, diff_vec_p);
762 for (i = 0; i < len; i++)
764 label = XEXP (XVECEXP (pat, diff_vec_p, i), 0);
765 set_label_in_map (map, CODE_LABEL_NUMBER (label), label);
769 if ((note = find_reg_note (insn, REG_LABEL, NULL_RTX)))
770 set_label_in_map (map, CODE_LABEL_NUMBER (XEXP (note, 0)),
774 /* Allocate space for the insn map. */
776 map->insn_map = (rtx *) xmalloc (max_insnno * sizeof (rtx));
778 /* Set this to zero, to indicate that we are doing loop unrolling,
779 not function inlining. */
780 map->inline_target = 0;
782 /* The register and constant maps depend on the number of registers
783 present, so the final maps can't be created until after
784 find_splittable_regs is called. However, they are needed for
785 preconditioning, so we create temporary maps when preconditioning
788 /* The preconditioning code may allocate two new pseudo registers. */
789 maxregnum = max_reg_num ();
791 /* local_regno is only valid for regnos < max_local_regnum. */
792 max_local_regnum = maxregnum;
794 /* Allocate and zero out the splittable_regs and addr_combined_regs
795 arrays. These must be zeroed here because they will be used if
796 loop preconditioning is performed, and must be zero for that case.
798 It is safe to do this here, since the extra registers created by the
799 preconditioning code and find_splittable_regs will never be used
800 to access the splittable_regs[] and addr_combined_regs[] arrays. */
802 splittable_regs = (rtx *) xcalloc (maxregnum, sizeof (rtx));
803 splittable_regs_updates = (int *) xcalloc (maxregnum, sizeof (int));
805 = (struct induction **) xcalloc (maxregnum, sizeof (struct induction *));
806 local_regno = (char *) xcalloc (maxregnum, sizeof (char));
808 /* Mark all local registers, i.e. the ones which are referenced only
810 if (INSN_UID (copy_end) < max_uid_for_loop)
812 int copy_start_luid = INSN_LUID (copy_start);
813 int copy_end_luid = INSN_LUID (copy_end);
815 /* If a register is used in the jump insn, we must not duplicate it
816 since it will also be used outside the loop. */
817 if (GET_CODE (copy_end) == JUMP_INSN)
820 /* If we have a target that uses cc0, then we also must not duplicate
821 the insn that sets cc0 before the jump insn, if one is present. */
823 if (GET_CODE (copy_end) == JUMP_INSN
824 && sets_cc0_p (PREV_INSN (copy_end)))
828 /* If copy_start points to the NOTE that starts the loop, then we must
829 use the next luid, because invariant pseudo-regs moved out of the loop
830 have their lifetimes modified to start here, but they are not safe
832 if (copy_start == loop_start)
835 /* If a pseudo's lifetime is entirely contained within this loop, then we
836 can use a different pseudo in each unrolled copy of the loop. This
837 results in better code. */
838 /* We must limit the generic test to max_reg_before_loop, because only
839 these pseudo registers have valid regno_first_uid info. */
840 for (r = FIRST_PSEUDO_REGISTER; r < max_reg_before_loop; ++r)
841 if (REGNO_FIRST_UID (r) > 0 && REGNO_FIRST_UID (r) <= max_uid_for_loop
842 && REGNO_FIRST_LUID (r) >= copy_start_luid
843 && REGNO_LAST_UID (r) > 0 && REGNO_LAST_UID (r) <= max_uid_for_loop
844 && REGNO_LAST_LUID (r) <= copy_end_luid)
846 /* However, we must also check for loop-carried dependencies.
847 If the value the pseudo has at the end of iteration X is
848 used by iteration X+1, then we can not use a different pseudo
849 for each unrolled copy of the loop. */
850 /* A pseudo is safe if regno_first_uid is a set, and this
851 set dominates all instructions from regno_first_uid to
853 /* ??? This check is simplistic. We would get better code if
854 this check was more sophisticated. */
855 if (set_dominates_use (r, REGNO_FIRST_UID (r), REGNO_LAST_UID (r),
856 copy_start, copy_end))
859 if (loop_dump_stream)
862 fprintf (loop_dump_stream, "Marked reg %d as local\n", r);
864 fprintf (loop_dump_stream, "Did not mark reg %d as local\n",
870 /* If this loop requires exit tests when unrolled, check to see if we
871 can precondition the loop so as to make the exit tests unnecessary.
872 Just like variable splitting, this is not safe if the loop is entered
873 via a jump to the bottom. Also, can not do this if no strength
874 reduce info, because precondition_loop_p uses this info. */
876 /* Must copy the loop body for preconditioning before the following
877 find_splittable_regs call since that will emit insns which need to
878 be after the preconditioned loop copies, but immediately before the
879 unrolled loop copies. */
881 /* Also, it is not safe to split induction variables for the preconditioned
882 copies of the loop body. If we split induction variables, then the code
883 assumes that each induction variable can be represented as a function
884 of its initial value and the loop iteration number. This is not true
885 in this case, because the last preconditioned copy of the loop body
886 could be any iteration from the first up to the `unroll_number-1'th,
887 depending on the initial value of the iteration variable. Therefore
888 we can not split induction variables here, because we can not calculate
889 their value. Hence, this code must occur before find_splittable_regs
892 if (unroll_type == UNROLL_NAIVE && ! splitting_not_safe && strength_reduce_p)
894 rtx initial_value, final_value, increment;
895 enum machine_mode mode;
897 if (precondition_loop_p (loop,
898 &initial_value, &final_value, &increment,
903 int abs_inc, neg_inc;
904 enum rtx_code cc = loop_info->comparison_code;
905 int less_p = (cc == LE || cc == LEU || cc == LT || cc == LTU);
906 int unsigned_p = (cc == LEU || cc == GEU || cc == LTU || cc == GTU);
908 map->reg_map = (rtx *) xmalloc (maxregnum * sizeof (rtx));
910 VARRAY_CONST_EQUIV_INIT (map->const_equiv_varray, maxregnum,
911 "unroll_loop_precondition");
912 global_const_equiv_varray = map->const_equiv_varray;
914 init_reg_map (map, maxregnum);
916 /* Limit loop unrolling to 4, since this will make 7 copies of
918 if (unroll_number > 4)
921 /* Save the absolute value of the increment, and also whether or
922 not it is negative. */
924 abs_inc = INTVAL (increment);
933 /* Calculate the difference between the final and initial values.
934 Final value may be a (plus (reg x) (const_int 1)) rtx.
935 Let the following cse pass simplify this if initial value is
938 We must copy the final and initial values here to avoid
939 improperly shared rtl.
941 We have to deal with for (i = 0; --i < 6;) type loops.
942 For such loops the real final value is the first time the
943 loop variable overflows, so the diff we calculate is the
944 distance from the overflow value. This is 0 or ~0 for
945 unsigned loops depending on the direction, or INT_MAX,
946 INT_MAX+1 for signed loops. We really do not need the
947 exact value, since we are only interested in the diff
948 modulo the increment, and the increment is a power of 2,
949 so we can pretend that the overflow value is 0/~0. */
951 if (cc == NE || less_p != neg_inc)
952 diff = expand_simple_binop (mode, MINUS, copy_rtx (final_value),
953 copy_rtx (initial_value), NULL_RTX, 0,
956 diff = expand_simple_unop (mode, neg_inc ? NOT : NEG,
957 copy_rtx (initial_value), NULL_RTX, 0);
959 /* Now calculate (diff % (unroll * abs (increment))) by using an
961 diff = expand_simple_binop (GET_MODE (diff), AND, diff,
962 GEN_INT (unroll_number * abs_inc - 1),
963 NULL_RTX, 0, OPTAB_LIB_WIDEN);
965 /* Now emit a sequence of branches to jump to the proper precond
968 labels = (rtx *) xmalloc (sizeof (rtx) * unroll_number);
969 for (i = 0; i < unroll_number; i++)
970 labels[i] = gen_label_rtx ();
972 /* Check for the case where the initial value is greater than or
973 equal to the final value. In that case, we want to execute
974 exactly one loop iteration. The code below will fail for this
975 case. This check does not apply if the loop has a NE
976 comparison at the end. */
980 rtx incremented_initval;
981 incremented_initval = expand_simple_binop (mode, PLUS,
986 emit_cmp_and_jump_insns (incremented_initval, final_value,
987 less_p ? GE : LE, NULL_RTX,
988 mode, unsigned_p, labels[1]);
989 predict_insn_def (get_last_insn (), PRED_LOOP_CONDITION,
991 JUMP_LABEL (get_last_insn ()) = labels[1];
992 LABEL_NUSES (labels[1])++;
995 /* Assuming the unroll_number is 4, and the increment is 2, then
996 for a negative increment: for a positive increment:
997 diff = 0,1 precond 0 diff = 0,7 precond 0
998 diff = 2,3 precond 3 diff = 1,2 precond 1
999 diff = 4,5 precond 2 diff = 3,4 precond 2
1000 diff = 6,7 precond 1 diff = 5,6 precond 3 */
1002 /* We only need to emit (unroll_number - 1) branches here, the
1003 last case just falls through to the following code. */
1005 /* ??? This would give better code if we emitted a tree of branches
1006 instead of the current linear list of branches. */
1008 for (i = 0; i < unroll_number - 1; i++)
1011 enum rtx_code cmp_code;
1013 /* For negative increments, must invert the constant compared
1014 against, except when comparing against zero. */
1022 cmp_const = unroll_number - i;
1031 emit_cmp_and_jump_insns (diff, GEN_INT (abs_inc * cmp_const),
1032 cmp_code, NULL_RTX, mode, 0, labels[i]);
1033 JUMP_LABEL (get_last_insn ()) = labels[i];
1034 LABEL_NUSES (labels[i])++;
1035 predict_insn (get_last_insn (), PRED_LOOP_PRECONDITIONING,
1036 REG_BR_PROB_BASE / (unroll_number - i));
1039 /* If the increment is greater than one, then we need another branch,
1040 to handle other cases equivalent to 0. */
1042 /* ??? This should be merged into the code above somehow to help
1043 simplify the code here, and reduce the number of branches emitted.
1044 For the negative increment case, the branch here could easily
1045 be merged with the `0' case branch above. For the positive
1046 increment case, it is not clear how this can be simplified. */
1051 enum rtx_code cmp_code;
1055 cmp_const = abs_inc - 1;
1060 cmp_const = abs_inc * (unroll_number - 1) + 1;
1064 emit_cmp_and_jump_insns (diff, GEN_INT (cmp_const), cmp_code,
1065 NULL_RTX, mode, 0, labels[0]);
1066 JUMP_LABEL (get_last_insn ()) = labels[0];
1067 LABEL_NUSES (labels[0])++;
1070 sequence = gen_sequence ();
1072 loop_insn_hoist (loop, sequence);
1074 /* Only the last copy of the loop body here needs the exit
1075 test, so set copy_end to exclude the compare/branch here,
1076 and then reset it inside the loop when get to the last
1079 if (GET_CODE (last_loop_insn) == BARRIER)
1080 copy_end = PREV_INSN (PREV_INSN (last_loop_insn));
1081 else if (GET_CODE (last_loop_insn) == JUMP_INSN)
1083 copy_end = PREV_INSN (last_loop_insn);
1085 /* The immediately preceding insn may be a compare which
1086 we do not want to copy. */
1087 if (sets_cc0_p (PREV_INSN (copy_end)))
1088 copy_end = PREV_INSN (copy_end);
1094 for (i = 1; i < unroll_number; i++)
1096 emit_label_after (labels[unroll_number - i],
1097 PREV_INSN (loop_start));
1099 memset ((char *) map->insn_map, 0, max_insnno * sizeof (rtx));
1100 memset ((char *) &VARRAY_CONST_EQUIV (map->const_equiv_varray, 0),
1101 0, (VARRAY_SIZE (map->const_equiv_varray)
1102 * sizeof (struct const_equiv_data)));
1105 for (j = 0; j < max_labelno; j++)
1107 set_label_in_map (map, j, gen_label_rtx ());
1109 for (r = FIRST_PSEUDO_REGISTER; r < max_local_regnum; r++)
1113 = gen_reg_rtx (GET_MODE (regno_reg_rtx[r]));
1114 record_base_value (REGNO (map->reg_map[r]),
1115 regno_reg_rtx[r], 0);
1117 /* The last copy needs the compare/branch insns at the end,
1118 so reset copy_end here if the loop ends with a conditional
1121 if (i == unroll_number - 1)
1123 if (GET_CODE (last_loop_insn) == BARRIER)
1124 copy_end = PREV_INSN (PREV_INSN (last_loop_insn));
1126 copy_end = last_loop_insn;
1129 /* None of the copies are the `last_iteration', so just
1130 pass zero for that parameter. */
1131 copy_loop_body (loop, copy_start, copy_end, map, exit_label, 0,
1132 unroll_type, start_label, loop_end,
1133 loop_start, copy_end);
1135 emit_label_after (labels[0], PREV_INSN (loop_start));
1137 if (GET_CODE (last_loop_insn) == BARRIER)
1139 insert_before = PREV_INSN (last_loop_insn);
1140 copy_end = PREV_INSN (insert_before);
1144 insert_before = last_loop_insn;
1146 /* The instruction immediately before the JUMP_INSN may
1147 be a compare instruction which we do not want to copy
1149 if (sets_cc0_p (PREV_INSN (insert_before)))
1150 insert_before = PREV_INSN (insert_before);
1152 copy_end = PREV_INSN (insert_before);
1155 /* Set unroll type to MODULO now. */
1156 unroll_type = UNROLL_MODULO;
1157 loop_preconditioned = 1;
1164 /* If reach here, and the loop type is UNROLL_NAIVE, then don't unroll
1165 the loop unless all loops are being unrolled. */
1166 if (unroll_type == UNROLL_NAIVE && ! flag_unroll_all_loops)
1168 if (loop_dump_stream)
1169 fprintf (loop_dump_stream,
1170 "Unrolling failure: Naive unrolling not being done.\n");
1174 /* At this point, we are guaranteed to unroll the loop. */
1176 /* Keep track of the unroll factor for the loop. */
1177 loop_info->unroll_number = unroll_number;
1179 /* For each biv and giv, determine whether it can be safely split into
1180 a different variable for each unrolled copy of the loop body.
1181 We precalculate and save this info here, since computing it is
1184 Do this before deleting any instructions from the loop, so that
1185 back_branch_in_range_p will work correctly. */
1187 if (splitting_not_safe)
1190 temp = find_splittable_regs (loop, unroll_type, unroll_number);
1192 /* find_splittable_regs may have created some new registers, so must
1193 reallocate the reg_map with the new larger size, and must realloc
1194 the constant maps also. */
1196 maxregnum = max_reg_num ();
1197 map->reg_map = (rtx *) xmalloc (maxregnum * sizeof (rtx));
1199 init_reg_map (map, maxregnum);
1201 if (map->const_equiv_varray == 0)
1202 VARRAY_CONST_EQUIV_INIT (map->const_equiv_varray,
1203 maxregnum + temp * unroll_number * 2,
1205 global_const_equiv_varray = map->const_equiv_varray;
1207 /* Search the list of bivs and givs to find ones which need to be remapped
1208 when split, and set their reg_map entry appropriately. */
1210 for (bl = ivs->list; bl; bl = bl->next)
1212 if (REGNO (bl->biv->src_reg) != bl->regno)
1213 map->reg_map[bl->regno] = bl->biv->src_reg;
1215 /* Currently, non-reduced/final-value givs are never split. */
1216 for (v = bl->giv; v; v = v->next_iv)
1217 if (REGNO (v->src_reg) != bl->regno)
1218 map->reg_map[REGNO (v->dest_reg)] = v->src_reg;
1222 /* Use our current register alignment and pointer flags. */
1223 map->regno_pointer_align = cfun->emit->regno_pointer_align;
1224 map->x_regno_reg_rtx = cfun->emit->x_regno_reg_rtx;
1226 /* If the loop is being partially unrolled, and the iteration variables
1227 are being split, and are being renamed for the split, then must fix up
1228 the compare/jump instruction at the end of the loop to refer to the new
1229 registers. This compare isn't copied, so the registers used in it
1230 will never be replaced if it isn't done here. */
1232 if (unroll_type == UNROLL_MODULO)
1234 insn = NEXT_INSN (copy_end);
1235 if (GET_CODE (insn) == INSN || GET_CODE (insn) == JUMP_INSN)
1236 PATTERN (insn) = remap_split_bivs (loop, PATTERN (insn));
1239 /* For unroll_number times, make a copy of each instruction
1240 between copy_start and copy_end, and insert these new instructions
1241 before the end of the loop. */
1243 for (i = 0; i < unroll_number; i++)
1245 memset ((char *) map->insn_map, 0, max_insnno * sizeof (rtx));
1246 memset ((char *) &VARRAY_CONST_EQUIV (map->const_equiv_varray, 0), 0,
1247 VARRAY_SIZE (map->const_equiv_varray) * sizeof (struct const_equiv_data));
1250 for (j = 0; j < max_labelno; j++)
1252 set_label_in_map (map, j, gen_label_rtx ());
1254 for (r = FIRST_PSEUDO_REGISTER; r < max_local_regnum; r++)
1257 map->reg_map[r] = gen_reg_rtx (GET_MODE (regno_reg_rtx[r]));
1258 record_base_value (REGNO (map->reg_map[r]),
1259 regno_reg_rtx[r], 0);
1262 /* If loop starts with a branch to the test, then fix it so that
1263 it points to the test of the first unrolled copy of the loop. */
1264 if (i == 0 && loop_start != copy_start)
1266 insn = PREV_INSN (copy_start);
1267 pattern = PATTERN (insn);
1269 tem = get_label_from_map (map,
1271 (XEXP (SET_SRC (pattern), 0)));
1272 SET_SRC (pattern) = gen_rtx_LABEL_REF (VOIDmode, tem);
1274 /* Set the jump label so that it can be used by later loop unrolling
1276 JUMP_LABEL (insn) = tem;
1277 LABEL_NUSES (tem)++;
1280 copy_loop_body (loop, copy_start, copy_end, map, exit_label,
1281 i == unroll_number - 1, unroll_type, start_label,
1282 loop_end, insert_before, insert_before);
1285 /* Before deleting any insns, emit a CODE_LABEL immediately after the last
1286 insn to be deleted. This prevents any runaway delete_insn call from
1287 more insns that it should, as it always stops at a CODE_LABEL. */
1289 /* Delete the compare and branch at the end of the loop if completely
1290 unrolling the loop. Deleting the backward branch at the end also
1291 deletes the code label at the start of the loop. This is done at
1292 the very end to avoid problems with back_branch_in_range_p. */
1294 if (unroll_type == UNROLL_COMPLETELY)
1295 safety_label = emit_label_after (gen_label_rtx (), last_loop_insn);
1297 safety_label = emit_label_after (gen_label_rtx (), copy_end);
1299 /* Delete all of the original loop instructions. Don't delete the
1300 LOOP_BEG note, or the first code label in the loop. */
1302 insn = NEXT_INSN (copy_start);
1303 while (insn != safety_label)
1305 /* ??? Don't delete named code labels. They will be deleted when the
1306 jump that references them is deleted. Otherwise, we end up deleting
1307 them twice, which causes them to completely disappear instead of turn
1308 into NOTE_INSN_DELETED_LABEL notes. This in turn causes aborts in
1309 dwarfout.c/dwarf2out.c. We could perhaps fix the dwarf*out.c files
1310 to handle deleted labels instead. Or perhaps fix DECL_RTL of the
1311 associated LABEL_DECL to point to one of the new label instances. */
1312 /* ??? Likewise, we can't delete a NOTE_INSN_DELETED_LABEL note. */
1313 if (insn != start_label
1314 && ! (GET_CODE (insn) == CODE_LABEL && LABEL_NAME (insn))
1315 && ! (GET_CODE (insn) == NOTE
1316 && NOTE_LINE_NUMBER (insn) == NOTE_INSN_DELETED_LABEL))
1317 insn = delete_related_insns (insn);
1319 insn = NEXT_INSN (insn);
1322 /* Can now delete the 'safety' label emitted to protect us from runaway
1323 delete_related_insns calls. */
1324 if (INSN_DELETED_P (safety_label))
1326 delete_related_insns (safety_label);
1328 /* If exit_label exists, emit it after the loop. Doing the emit here
1329 forces it to have a higher INSN_UID than any insn in the unrolled loop.
1330 This is needed so that mostly_true_jump in reorg.c will treat jumps
1331 to this loop end label correctly, i.e. predict that they are usually
1334 emit_label_after (exit_label, loop_end);
1337 if (unroll_type == UNROLL_COMPLETELY)
1339 /* Remove the loop notes since this is no longer a loop. */
1341 delete_related_insns (loop->vtop);
1343 delete_related_insns (loop->cont);
1345 delete_related_insns (loop_start);
1347 delete_related_insns (loop_end);
1350 if (map->const_equiv_varray)
1351 VARRAY_FREE (map->const_equiv_varray);
1354 free (map->label_map);
1357 free (map->insn_map);
1358 free (splittable_regs);
1359 free (splittable_regs_updates);
1360 free (addr_combined_regs);
1363 free (map->reg_map);
1367 /* Return true if the loop can be safely, and profitably, preconditioned
1368 so that the unrolled copies of the loop body don't need exit tests.
1370 This only works if final_value, initial_value and increment can be
1371 determined, and if increment is a constant power of 2.
1372 If increment is not a power of 2, then the preconditioning modulo
1373 operation would require a real modulo instead of a boolean AND, and this
1374 is not considered `profitable'. */
1376 /* ??? If the loop is known to be executed very many times, or the machine
1377 has a very cheap divide instruction, then preconditioning is a win even
1378 when the increment is not a power of 2. Use RTX_COST to compute
1379 whether divide is cheap.
1380 ??? A divide by constant doesn't actually need a divide, look at
1381 expand_divmod. The reduced cost of this optimized modulo is not
1382 reflected in RTX_COST. */
1385 precondition_loop_p (loop, initial_value, final_value, increment, mode)
1386 const struct loop *loop;
1387 rtx *initial_value, *final_value, *increment;
1388 enum machine_mode *mode;
1390 rtx loop_start = loop->start;
1391 struct loop_info *loop_info = LOOP_INFO (loop);
1393 if (loop_info->n_iterations > 0)
1395 *initial_value = const0_rtx;
1396 *increment = const1_rtx;
1397 *final_value = GEN_INT (loop_info->n_iterations);
1400 if (loop_dump_stream)
1402 fputs ("Preconditioning: Success, number of iterations known, ",
1404 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC,
1405 loop_info->n_iterations);
1406 fputs (".\n", loop_dump_stream);
1411 if (loop_info->iteration_var == 0)
1413 if (loop_dump_stream)
1414 fprintf (loop_dump_stream,
1415 "Preconditioning: Could not find iteration variable.\n");
1418 else if (loop_info->initial_value == 0)
1420 if (loop_dump_stream)
1421 fprintf (loop_dump_stream,
1422 "Preconditioning: Could not find initial value.\n");
1425 else if (loop_info->increment == 0)
1427 if (loop_dump_stream)
1428 fprintf (loop_dump_stream,
1429 "Preconditioning: Could not find increment value.\n");
1432 else if (GET_CODE (loop_info->increment) != CONST_INT)
1434 if (loop_dump_stream)
1435 fprintf (loop_dump_stream,
1436 "Preconditioning: Increment not a constant.\n");
1439 else if ((exact_log2 (INTVAL (loop_info->increment)) < 0)
1440 && (exact_log2 (-INTVAL (loop_info->increment)) < 0))
1442 if (loop_dump_stream)
1443 fprintf (loop_dump_stream,
1444 "Preconditioning: Increment not a constant power of 2.\n");
1448 /* Unsigned_compare and compare_dir can be ignored here, since they do
1449 not matter for preconditioning. */
1451 if (loop_info->final_value == 0)
1453 if (loop_dump_stream)
1454 fprintf (loop_dump_stream,
1455 "Preconditioning: EQ comparison loop.\n");
1459 /* Must ensure that final_value is invariant, so call
1460 loop_invariant_p to check. Before doing so, must check regno
1461 against max_reg_before_loop to make sure that the register is in
1462 the range covered by loop_invariant_p. If it isn't, then it is
1463 most likely a biv/giv which by definition are not invariant. */
1464 if ((GET_CODE (loop_info->final_value) == REG
1465 && REGNO (loop_info->final_value) >= max_reg_before_loop)
1466 || (GET_CODE (loop_info->final_value) == PLUS
1467 && REGNO (XEXP (loop_info->final_value, 0)) >= max_reg_before_loop)
1468 || ! loop_invariant_p (loop, loop_info->final_value))
1470 if (loop_dump_stream)
1471 fprintf (loop_dump_stream,
1472 "Preconditioning: Final value not invariant.\n");
1476 /* Fail for floating point values, since the caller of this function
1477 does not have code to deal with them. */
1478 if (GET_MODE_CLASS (GET_MODE (loop_info->final_value)) == MODE_FLOAT
1479 || GET_MODE_CLASS (GET_MODE (loop_info->initial_value)) == MODE_FLOAT)
1481 if (loop_dump_stream)
1482 fprintf (loop_dump_stream,
1483 "Preconditioning: Floating point final or initial value.\n");
1487 /* Fail if loop_info->iteration_var is not live before loop_start,
1488 since we need to test its value in the preconditioning code. */
1490 if (REGNO_FIRST_LUID (REGNO (loop_info->iteration_var))
1491 > INSN_LUID (loop_start))
1493 if (loop_dump_stream)
1494 fprintf (loop_dump_stream,
1495 "Preconditioning: Iteration var not live before loop start.\n");
1499 /* Note that loop_iterations biases the initial value for GIV iterators
1500 such as "while (i-- > 0)" so that we can calculate the number of
1501 iterations just like for BIV iterators.
1503 Also note that the absolute values of initial_value and
1504 final_value are unimportant as only their difference is used for
1505 calculating the number of loop iterations. */
1506 *initial_value = loop_info->initial_value;
1507 *increment = loop_info->increment;
1508 *final_value = loop_info->final_value;
1510 /* Decide what mode to do these calculations in. Choose the larger
1511 of final_value's mode and initial_value's mode, or a full-word if
1512 both are constants. */
1513 *mode = GET_MODE (*final_value);
1514 if (*mode == VOIDmode)
1516 *mode = GET_MODE (*initial_value);
1517 if (*mode == VOIDmode)
1520 else if (*mode != GET_MODE (*initial_value)
1521 && (GET_MODE_SIZE (*mode)
1522 < GET_MODE_SIZE (GET_MODE (*initial_value))))
1523 *mode = GET_MODE (*initial_value);
1526 if (loop_dump_stream)
1527 fprintf (loop_dump_stream, "Preconditioning: Successful.\n");
1531 /* All pseudo-registers must be mapped to themselves. Two hard registers
1532 must be mapped, VIRTUAL_STACK_VARS_REGNUM and VIRTUAL_INCOMING_ARGS_
1533 REGNUM, to avoid function-inlining specific conversions of these
1534 registers. All other hard regs can not be mapped because they may be
1539 init_reg_map (map, maxregnum)
1540 struct inline_remap *map;
1545 for (i = maxregnum - 1; i > LAST_VIRTUAL_REGISTER; i--)
1546 map->reg_map[i] = regno_reg_rtx[i];
1547 /* Just clear the rest of the entries. */
1548 for (i = LAST_VIRTUAL_REGISTER; i >= 0; i--)
1549 map->reg_map[i] = 0;
1551 map->reg_map[VIRTUAL_STACK_VARS_REGNUM]
1552 = regno_reg_rtx[VIRTUAL_STACK_VARS_REGNUM];
1553 map->reg_map[VIRTUAL_INCOMING_ARGS_REGNUM]
1554 = regno_reg_rtx[VIRTUAL_INCOMING_ARGS_REGNUM];
1557 /* Strength-reduction will often emit code for optimized biv/givs which
1558 calculates their value in a temporary register, and then copies the result
1559 to the iv. This procedure reconstructs the pattern computing the iv;
1560 verifying that all operands are of the proper form.
1562 PATTERN must be the result of single_set.
1563 The return value is the amount that the giv is incremented by. */
1566 calculate_giv_inc (pattern, src_insn, regno)
1567 rtx pattern, src_insn;
1571 rtx increment_total = 0;
1575 /* Verify that we have an increment insn here. First check for a plus
1576 as the set source. */
1577 if (GET_CODE (SET_SRC (pattern)) != PLUS)
1579 /* SR sometimes computes the new giv value in a temp, then copies it
1581 src_insn = PREV_INSN (src_insn);
1582 pattern = single_set (src_insn);
1583 if (GET_CODE (SET_SRC (pattern)) != PLUS)
1586 /* The last insn emitted is not needed, so delete it to avoid confusing
1587 the second cse pass. This insn sets the giv unnecessarily. */
1588 delete_related_insns (get_last_insn ());
1591 /* Verify that we have a constant as the second operand of the plus. */
1592 increment = XEXP (SET_SRC (pattern), 1);
1593 if (GET_CODE (increment) != CONST_INT)
1595 /* SR sometimes puts the constant in a register, especially if it is
1596 too big to be an add immed operand. */
1597 increment = find_last_value (increment, &src_insn, NULL_RTX, 0);
1599 /* SR may have used LO_SUM to compute the constant if it is too large
1600 for a load immed operand. In this case, the constant is in operand
1601 one of the LO_SUM rtx. */
1602 if (GET_CODE (increment) == LO_SUM)
1603 increment = XEXP (increment, 1);
1605 /* Some ports store large constants in memory and add a REG_EQUAL
1606 note to the store insn. */
1607 else if (GET_CODE (increment) == MEM)
1609 rtx note = find_reg_note (src_insn, REG_EQUAL, 0);
1611 increment = XEXP (note, 0);
1614 else if (GET_CODE (increment) == IOR
1615 || GET_CODE (increment) == ASHIFT
1616 || GET_CODE (increment) == PLUS)
1618 /* The rs6000 port loads some constants with IOR.
1619 The alpha port loads some constants with ASHIFT and PLUS. */
1620 rtx second_part = XEXP (increment, 1);
1621 enum rtx_code code = GET_CODE (increment);
1623 increment = find_last_value (XEXP (increment, 0),
1624 &src_insn, NULL_RTX, 0);
1625 /* Don't need the last insn anymore. */
1626 delete_related_insns (get_last_insn ());
1628 if (GET_CODE (second_part) != CONST_INT
1629 || GET_CODE (increment) != CONST_INT)
1633 increment = GEN_INT (INTVAL (increment) | INTVAL (second_part));
1634 else if (code == PLUS)
1635 increment = GEN_INT (INTVAL (increment) + INTVAL (second_part));
1637 increment = GEN_INT (INTVAL (increment) << INTVAL (second_part));
1640 if (GET_CODE (increment) != CONST_INT)
1643 /* The insn loading the constant into a register is no longer needed,
1645 delete_related_insns (get_last_insn ());
1648 if (increment_total)
1649 increment_total = GEN_INT (INTVAL (increment_total) + INTVAL (increment));
1651 increment_total = increment;
1653 /* Check that the source register is the same as the register we expected
1654 to see as the source. If not, something is seriously wrong. */
1655 if (GET_CODE (XEXP (SET_SRC (pattern), 0)) != REG
1656 || REGNO (XEXP (SET_SRC (pattern), 0)) != regno)
1658 /* Some machines (e.g. the romp), may emit two add instructions for
1659 certain constants, so lets try looking for another add immediately
1660 before this one if we have only seen one add insn so far. */
1666 src_insn = PREV_INSN (src_insn);
1667 pattern = single_set (src_insn);
1669 delete_related_insns (get_last_insn ());
1677 return increment_total;
1680 /* Copy REG_NOTES, except for insn references, because not all insn_map
1681 entries are valid yet. We do need to copy registers now though, because
1682 the reg_map entries can change during copying. */
1685 initial_reg_note_copy (notes, map)
1687 struct inline_remap *map;
1694 copy = rtx_alloc (GET_CODE (notes));
1695 PUT_REG_NOTE_KIND (copy, REG_NOTE_KIND (notes));
1697 if (GET_CODE (notes) == EXPR_LIST)
1698 XEXP (copy, 0) = copy_rtx_and_substitute (XEXP (notes, 0), map, 0);
1699 else if (GET_CODE (notes) == INSN_LIST)
1700 /* Don't substitute for these yet. */
1701 XEXP (copy, 0) = copy_rtx (XEXP (notes, 0));
1705 XEXP (copy, 1) = initial_reg_note_copy (XEXP (notes, 1), map);
1710 /* Fixup insn references in copied REG_NOTES. */
1713 final_reg_note_copy (notesp, map)
1715 struct inline_remap *map;
1721 if (GET_CODE (note) == INSN_LIST)
1723 /* Sometimes, we have a REG_WAS_0 note that points to a
1724 deleted instruction. In that case, we can just delete the
1726 if (REG_NOTE_KIND (note) == REG_WAS_0)
1728 *notesp = XEXP (note, 1);
1733 rtx insn = map->insn_map[INSN_UID (XEXP (note, 0))];
1735 /* If we failed to remap the note, something is awry. */
1739 XEXP (note, 0) = insn;
1743 notesp = &XEXP (note, 1);
1747 /* Copy each instruction in the loop, substituting from map as appropriate.
1748 This is very similar to a loop in expand_inline_function. */
1751 copy_loop_body (loop, copy_start, copy_end, map, exit_label, last_iteration,
1752 unroll_type, start_label, loop_end, insert_before,
1755 rtx copy_start, copy_end;
1756 struct inline_remap *map;
1759 enum unroll_types unroll_type;
1760 rtx start_label, loop_end, insert_before, copy_notes_from;
1762 struct loop_ivs *ivs = LOOP_IVS (loop);
1764 rtx set, tem, copy = NULL_RTX;
1765 int dest_reg_was_split, i;
1769 rtx final_label = 0;
1770 rtx giv_inc, giv_dest_reg, giv_src_reg;
1772 /* If this isn't the last iteration, then map any references to the
1773 start_label to final_label. Final label will then be emitted immediately
1774 after the end of this loop body if it was ever used.
1776 If this is the last iteration, then map references to the start_label
1778 if (! last_iteration)
1780 final_label = gen_label_rtx ();
1781 set_label_in_map (map, CODE_LABEL_NUMBER (start_label), final_label);
1784 set_label_in_map (map, CODE_LABEL_NUMBER (start_label), start_label);
1788 /* Emit a NOTE_INSN_DELETED to force at least two insns onto the sequence.
1789 Else gen_sequence could return a raw pattern for a jump which we pass
1790 off to emit_insn_before (instead of emit_jump_insn_before) which causes
1791 a variety of losing behaviors later. */
1792 emit_note (0, NOTE_INSN_DELETED);
1797 insn = NEXT_INSN (insn);
1799 map->orig_asm_operands_vector = 0;
1801 switch (GET_CODE (insn))
1804 pattern = PATTERN (insn);
1808 /* Check to see if this is a giv that has been combined with
1809 some split address givs. (Combined in the sense that
1810 `combine_givs' in loop.c has put two givs in the same register.)
1811 In this case, we must search all givs based on the same biv to
1812 find the address givs. Then split the address givs.
1813 Do this before splitting the giv, since that may map the
1814 SET_DEST to a new register. */
1816 if ((set = single_set (insn))
1817 && GET_CODE (SET_DEST (set)) == REG
1818 && addr_combined_regs[REGNO (SET_DEST (set))])
1820 struct iv_class *bl;
1821 struct induction *v, *tv;
1822 unsigned int regno = REGNO (SET_DEST (set));
1824 v = addr_combined_regs[REGNO (SET_DEST (set))];
1825 bl = REG_IV_CLASS (ivs, REGNO (v->src_reg));
1827 /* Although the giv_inc amount is not needed here, we must call
1828 calculate_giv_inc here since it might try to delete the
1829 last insn emitted. If we wait until later to call it,
1830 we might accidentally delete insns generated immediately
1831 below by emit_unrolled_add. */
1833 giv_inc = calculate_giv_inc (set, insn, regno);
1835 /* Now find all address giv's that were combined with this
1837 for (tv = bl->giv; tv; tv = tv->next_iv)
1838 if (tv->giv_type == DEST_ADDR && tv->same == v)
1842 /* If this DEST_ADDR giv was not split, then ignore it. */
1843 if (*tv->location != tv->dest_reg)
1846 /* Scale this_giv_inc if the multiplicative factors of
1847 the two givs are different. */
1848 this_giv_inc = INTVAL (giv_inc);
1849 if (tv->mult_val != v->mult_val)
1850 this_giv_inc = (this_giv_inc / INTVAL (v->mult_val)
1851 * INTVAL (tv->mult_val));
1853 tv->dest_reg = plus_constant (tv->dest_reg, this_giv_inc);
1854 *tv->location = tv->dest_reg;
1856 if (last_iteration && unroll_type != UNROLL_COMPLETELY)
1858 /* Must emit an insn to increment the split address
1859 giv. Add in the const_adjust field in case there
1860 was a constant eliminated from the address. */
1861 rtx value, dest_reg;
1863 /* tv->dest_reg will be either a bare register,
1864 or else a register plus a constant. */
1865 if (GET_CODE (tv->dest_reg) == REG)
1866 dest_reg = tv->dest_reg;
1868 dest_reg = XEXP (tv->dest_reg, 0);
1870 /* Check for shared address givs, and avoid
1871 incrementing the shared pseudo reg more than
1873 if (! tv->same_insn && ! tv->shared)
1875 /* tv->dest_reg may actually be a (PLUS (REG)
1876 (CONST)) here, so we must call plus_constant
1877 to add the const_adjust amount before calling
1878 emit_unrolled_add below. */
1879 value = plus_constant (tv->dest_reg,
1882 if (GET_CODE (value) == PLUS)
1884 /* The constant could be too large for an add
1885 immediate, so can't directly emit an insn
1887 emit_unrolled_add (dest_reg, XEXP (value, 0),
1892 /* Reset the giv to be just the register again, in case
1893 it is used after the set we have just emitted.
1894 We must subtract the const_adjust factor added in
1896 tv->dest_reg = plus_constant (dest_reg,
1898 *tv->location = tv->dest_reg;
1903 /* If this is a setting of a splittable variable, then determine
1904 how to split the variable, create a new set based on this split,
1905 and set up the reg_map so that later uses of the variable will
1906 use the new split variable. */
1908 dest_reg_was_split = 0;
1910 if ((set = single_set (insn))
1911 && GET_CODE (SET_DEST (set)) == REG
1912 && splittable_regs[REGNO (SET_DEST (set))])
1914 unsigned int regno = REGNO (SET_DEST (set));
1915 unsigned int src_regno;
1917 dest_reg_was_split = 1;
1919 giv_dest_reg = SET_DEST (set);
1920 giv_src_reg = giv_dest_reg;
1921 /* Compute the increment value for the giv, if it wasn't
1922 already computed above. */
1924 giv_inc = calculate_giv_inc (set, insn, regno);
1926 src_regno = REGNO (giv_src_reg);
1928 if (unroll_type == UNROLL_COMPLETELY)
1930 /* Completely unrolling the loop. Set the induction
1931 variable to a known constant value. */
1933 /* The value in splittable_regs may be an invariant
1934 value, so we must use plus_constant here. */
1935 splittable_regs[regno]
1936 = plus_constant (splittable_regs[src_regno],
1939 if (GET_CODE (splittable_regs[regno]) == PLUS)
1941 giv_src_reg = XEXP (splittable_regs[regno], 0);
1942 giv_inc = XEXP (splittable_regs[regno], 1);
1946 /* The splittable_regs value must be a REG or a
1947 CONST_INT, so put the entire value in the giv_src_reg
1949 giv_src_reg = splittable_regs[regno];
1950 giv_inc = const0_rtx;
1955 /* Partially unrolling loop. Create a new pseudo
1956 register for the iteration variable, and set it to
1957 be a constant plus the original register. Except
1958 on the last iteration, when the result has to
1959 go back into the original iteration var register. */
1961 /* Handle bivs which must be mapped to a new register
1962 when split. This happens for bivs which need their
1963 final value set before loop entry. The new register
1964 for the biv was stored in the biv's first struct
1965 induction entry by find_splittable_regs. */
1967 if (regno < ivs->n_regs
1968 && REG_IV_TYPE (ivs, regno) == BASIC_INDUCT)
1970 giv_src_reg = REG_IV_CLASS (ivs, regno)->biv->src_reg;
1971 giv_dest_reg = giv_src_reg;
1975 /* If non-reduced/final-value givs were split, then
1976 this would have to remap those givs also. See
1977 find_splittable_regs. */
1980 splittable_regs[regno]
1981 = simplify_gen_binary (PLUS, GET_MODE (giv_src_reg),
1983 splittable_regs[src_regno]);
1984 giv_inc = splittable_regs[regno];
1986 /* Now split the induction variable by changing the dest
1987 of this insn to a new register, and setting its
1988 reg_map entry to point to this new register.
1990 If this is the last iteration, and this is the last insn
1991 that will update the iv, then reuse the original dest,
1992 to ensure that the iv will have the proper value when
1993 the loop exits or repeats.
1995 Using splittable_regs_updates here like this is safe,
1996 because it can only be greater than one if all
1997 instructions modifying the iv are always executed in
2000 if (! last_iteration
2001 || (splittable_regs_updates[regno]-- != 1))
2003 tem = gen_reg_rtx (GET_MODE (giv_src_reg));
2005 map->reg_map[regno] = tem;
2006 record_base_value (REGNO (tem),
2007 giv_inc == const0_rtx
2009 : gen_rtx_PLUS (GET_MODE (giv_src_reg),
2010 giv_src_reg, giv_inc),
2014 map->reg_map[regno] = giv_src_reg;
2017 /* The constant being added could be too large for an add
2018 immediate, so can't directly emit an insn here. */
2019 emit_unrolled_add (giv_dest_reg, giv_src_reg, giv_inc);
2020 copy = get_last_insn ();
2021 pattern = PATTERN (copy);
2025 pattern = copy_rtx_and_substitute (pattern, map, 0);
2026 copy = emit_insn (pattern);
2028 REG_NOTES (copy) = initial_reg_note_copy (REG_NOTES (insn), map);
2031 /* If this insn is setting CC0, it may need to look at
2032 the insn that uses CC0 to see what type of insn it is.
2033 In that case, the call to recog via validate_change will
2034 fail. So don't substitute constants here. Instead,
2035 do it when we emit the following insn.
2037 For example, see the pyr.md file. That machine has signed and
2038 unsigned compares. The compare patterns must check the
2039 following branch insn to see which what kind of compare to
2042 If the previous insn set CC0, substitute constants on it as
2044 if (sets_cc0_p (PATTERN (copy)) != 0)
2049 try_constants (cc0_insn, map);
2051 try_constants (copy, map);
2054 try_constants (copy, map);
2057 /* Make split induction variable constants `permanent' since we
2058 know there are no backward branches across iteration variable
2059 settings which would invalidate this. */
2060 if (dest_reg_was_split)
2062 int regno = REGNO (SET_DEST (set));
2064 if ((size_t) regno < VARRAY_SIZE (map->const_equiv_varray)
2065 && (VARRAY_CONST_EQUIV (map->const_equiv_varray, regno).age
2067 VARRAY_CONST_EQUIV (map->const_equiv_varray, regno).age = -1;
2072 pattern = copy_rtx_and_substitute (PATTERN (insn), map, 0);
2073 copy = emit_jump_insn (pattern);
2074 REG_NOTES (copy) = initial_reg_note_copy (REG_NOTES (insn), map);
2076 if (JUMP_LABEL (insn) == start_label && insn == copy_end
2077 && ! last_iteration)
2079 /* Update JUMP_LABEL make invert_jump work correctly. */
2080 JUMP_LABEL (copy) = get_label_from_map (map,
2082 (JUMP_LABEL (insn)));
2083 LABEL_NUSES (JUMP_LABEL (copy))++;
2085 /* This is a branch to the beginning of the loop; this is the
2086 last insn being copied; and this is not the last iteration.
2087 In this case, we want to change the original fall through
2088 case to be a branch past the end of the loop, and the
2089 original jump label case to fall_through. */
2091 if (!invert_jump (copy, exit_label, 0))
2094 rtx lab = gen_label_rtx ();
2095 /* Can't do it by reversing the jump (probably because we
2096 couldn't reverse the conditions), so emit a new
2097 jump_insn after COPY, and redirect the jump around
2099 jmp = emit_jump_insn_after (gen_jump (exit_label), copy);
2100 JUMP_LABEL (jmp) = exit_label;
2101 LABEL_NUSES (exit_label)++;
2102 jmp = emit_barrier_after (jmp);
2103 emit_label_after (lab, jmp);
2104 LABEL_NUSES (lab) = 0;
2105 if (!redirect_jump (copy, lab, 0))
2112 try_constants (cc0_insn, map);
2115 try_constants (copy, map);
2117 /* Set the jump label of COPY correctly to avoid problems with
2118 later passes of unroll_loop, if INSN had jump label set. */
2119 if (JUMP_LABEL (insn))
2123 /* Can't use the label_map for every insn, since this may be
2124 the backward branch, and hence the label was not mapped. */
2125 if ((set = single_set (copy)))
2127 tem = SET_SRC (set);
2128 if (GET_CODE (tem) == LABEL_REF)
2129 label = XEXP (tem, 0);
2130 else if (GET_CODE (tem) == IF_THEN_ELSE)
2132 if (XEXP (tem, 1) != pc_rtx)
2133 label = XEXP (XEXP (tem, 1), 0);
2135 label = XEXP (XEXP (tem, 2), 0);
2139 if (label && GET_CODE (label) == CODE_LABEL)
2140 JUMP_LABEL (copy) = label;
2143 /* An unrecognizable jump insn, probably the entry jump
2144 for a switch statement. This label must have been mapped,
2145 so just use the label_map to get the new jump label. */
2147 = get_label_from_map (map,
2148 CODE_LABEL_NUMBER (JUMP_LABEL (insn)));
2151 /* If this is a non-local jump, then must increase the label
2152 use count so that the label will not be deleted when the
2153 original jump is deleted. */
2154 LABEL_NUSES (JUMP_LABEL (copy))++;
2156 else if (GET_CODE (PATTERN (copy)) == ADDR_VEC
2157 || GET_CODE (PATTERN (copy)) == ADDR_DIFF_VEC)
2159 rtx pat = PATTERN (copy);
2160 int diff_vec_p = GET_CODE (pat) == ADDR_DIFF_VEC;
2161 int len = XVECLEN (pat, diff_vec_p);
2164 for (i = 0; i < len; i++)
2165 LABEL_NUSES (XEXP (XVECEXP (pat, diff_vec_p, i), 0))++;
2168 /* If this used to be a conditional jump insn but whose branch
2169 direction is now known, we must do something special. */
2170 if (any_condjump_p (insn) && onlyjump_p (insn) && map->last_pc_value)
2173 /* If the previous insn set cc0 for us, delete it. */
2174 if (only_sets_cc0_p (PREV_INSN (copy)))
2175 delete_related_insns (PREV_INSN (copy));
2178 /* If this is now a no-op, delete it. */
2179 if (map->last_pc_value == pc_rtx)
2185 /* Otherwise, this is unconditional jump so we must put a
2186 BARRIER after it. We could do some dead code elimination
2187 here, but jump.c will do it just as well. */
2193 pattern = copy_rtx_and_substitute (PATTERN (insn), map, 0);
2194 copy = emit_call_insn (pattern);
2195 REG_NOTES (copy) = initial_reg_note_copy (REG_NOTES (insn), map);
2197 /* Because the USAGE information potentially contains objects other
2198 than hard registers, we need to copy it. */
2199 CALL_INSN_FUNCTION_USAGE (copy)
2200 = copy_rtx_and_substitute (CALL_INSN_FUNCTION_USAGE (insn),
2205 try_constants (cc0_insn, map);
2208 try_constants (copy, map);
2210 /* Be lazy and assume CALL_INSNs clobber all hard registers. */
2211 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
2212 VARRAY_CONST_EQUIV (map->const_equiv_varray, i).rtx = 0;
2216 /* If this is the loop start label, then we don't need to emit a
2217 copy of this label since no one will use it. */
2219 if (insn != start_label)
2221 copy = emit_label (get_label_from_map (map,
2222 CODE_LABEL_NUMBER (insn)));
2228 copy = emit_barrier ();
2232 /* VTOP and CONT notes are valid only before the loop exit test.
2233 If placed anywhere else, loop may generate bad code. */
2234 /* BASIC_BLOCK notes exist to stabilize basic block structures with
2235 the associated rtl. We do not want to share the structure in
2238 if (NOTE_LINE_NUMBER (insn) != NOTE_INSN_DELETED
2239 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_DELETED_LABEL
2240 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_BASIC_BLOCK
2241 && ((NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_VTOP
2242 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_CONT)
2243 || (last_iteration && unroll_type != UNROLL_COMPLETELY)))
2244 copy = emit_note (NOTE_SOURCE_FILE (insn),
2245 NOTE_LINE_NUMBER (insn));
2254 map->insn_map[INSN_UID (insn)] = copy;
2256 while (insn != copy_end);
2258 /* Now finish coping the REG_NOTES. */
2262 insn = NEXT_INSN (insn);
2263 if ((GET_CODE (insn) == INSN || GET_CODE (insn) == JUMP_INSN
2264 || GET_CODE (insn) == CALL_INSN)
2265 && map->insn_map[INSN_UID (insn)])
2266 final_reg_note_copy (®_NOTES (map->insn_map[INSN_UID (insn)]), map);
2268 while (insn != copy_end);
2270 /* There may be notes between copy_notes_from and loop_end. Emit a copy of
2271 each of these notes here, since there may be some important ones, such as
2272 NOTE_INSN_BLOCK_END notes, in this group. We don't do this on the last
2273 iteration, because the original notes won't be deleted.
2275 We can't use insert_before here, because when from preconditioning,
2276 insert_before points before the loop. We can't use copy_end, because
2277 there may be insns already inserted after it (which we don't want to
2278 copy) when not from preconditioning code. */
2280 if (! last_iteration)
2282 for (insn = copy_notes_from; insn != loop_end; insn = NEXT_INSN (insn))
2284 /* VTOP notes are valid only before the loop exit test.
2285 If placed anywhere else, loop may generate bad code.
2286 There is no need to test for NOTE_INSN_LOOP_CONT notes
2287 here, since COPY_NOTES_FROM will be at most one or two (for cc0)
2288 instructions before the last insn in the loop, and if the
2289 end test is that short, there will be a VTOP note between
2290 the CONT note and the test. */
2291 if (GET_CODE (insn) == NOTE
2292 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_DELETED
2293 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_BASIC_BLOCK
2294 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_VTOP)
2295 emit_note (NOTE_SOURCE_FILE (insn), NOTE_LINE_NUMBER (insn));
2299 if (final_label && LABEL_NUSES (final_label) > 0)
2300 emit_label (final_label);
2302 tem = gen_sequence ();
2304 loop_insn_emit_before (loop, 0, insert_before, tem);
2307 /* Emit an insn, using the expand_binop to ensure that a valid insn is
2308 emitted. This will correctly handle the case where the increment value
2309 won't fit in the immediate field of a PLUS insns. */
2312 emit_unrolled_add (dest_reg, src_reg, increment)
2313 rtx dest_reg, src_reg, increment;
2317 result = expand_simple_binop (GET_MODE (dest_reg), PLUS, src_reg, increment,
2318 dest_reg, 0, OPTAB_LIB_WIDEN);
2320 if (dest_reg != result)
2321 emit_move_insn (dest_reg, result);
2324 /* Searches the insns between INSN and LOOP->END. Returns 1 if there
2325 is a backward branch in that range that branches to somewhere between
2326 LOOP->START and INSN. Returns 0 otherwise. */
2328 /* ??? This is quadratic algorithm. Could be rewritten to be linear.
2329 In practice, this is not a problem, because this function is seldom called,
2330 and uses a negligible amount of CPU time on average. */
2333 back_branch_in_range_p (loop, insn)
2334 const struct loop *loop;
2337 rtx p, q, target_insn;
2338 rtx loop_start = loop->start;
2339 rtx loop_end = loop->end;
2340 rtx orig_loop_end = loop->end;
2342 /* Stop before we get to the backward branch at the end of the loop. */
2343 loop_end = prev_nonnote_insn (loop_end);
2344 if (GET_CODE (loop_end) == BARRIER)
2345 loop_end = PREV_INSN (loop_end);
2347 /* Check in case insn has been deleted, search forward for first non
2348 deleted insn following it. */
2349 while (INSN_DELETED_P (insn))
2350 insn = NEXT_INSN (insn);
2352 /* Check for the case where insn is the last insn in the loop. Deal
2353 with the case where INSN was a deleted loop test insn, in which case
2354 it will now be the NOTE_LOOP_END. */
2355 if (insn == loop_end || insn == orig_loop_end)
2358 for (p = NEXT_INSN (insn); p != loop_end; p = NEXT_INSN (p))
2360 if (GET_CODE (p) == JUMP_INSN)
2362 target_insn = JUMP_LABEL (p);
2364 /* Search from loop_start to insn, to see if one of them is
2365 the target_insn. We can't use INSN_LUID comparisons here,
2366 since insn may not have an LUID entry. */
2367 for (q = loop_start; q != insn; q = NEXT_INSN (q))
2368 if (q == target_insn)
2376 /* Try to generate the simplest rtx for the expression
2377 (PLUS (MULT mult1 mult2) add1). This is used to calculate the initial
2381 fold_rtx_mult_add (mult1, mult2, add1, mode)
2382 rtx mult1, mult2, add1;
2383 enum machine_mode mode;
2388 /* The modes must all be the same. This should always be true. For now,
2389 check to make sure. */
2390 if ((GET_MODE (mult1) != mode && GET_MODE (mult1) != VOIDmode)
2391 || (GET_MODE (mult2) != mode && GET_MODE (mult2) != VOIDmode)
2392 || (GET_MODE (add1) != mode && GET_MODE (add1) != VOIDmode))
2395 /* Ensure that if at least one of mult1/mult2 are constant, then mult2
2396 will be a constant. */
2397 if (GET_CODE (mult1) == CONST_INT)
2404 mult_res = simplify_binary_operation (MULT, mode, mult1, mult2);
2406 mult_res = gen_rtx_MULT (mode, mult1, mult2);
2408 /* Again, put the constant second. */
2409 if (GET_CODE (add1) == CONST_INT)
2416 result = simplify_binary_operation (PLUS, mode, add1, mult_res);
2418 result = gen_rtx_PLUS (mode, add1, mult_res);
2423 /* Searches the list of induction struct's for the biv BL, to try to calculate
2424 the total increment value for one iteration of the loop as a constant.
2426 Returns the increment value as an rtx, simplified as much as possible,
2427 if it can be calculated. Otherwise, returns 0. */
2430 biv_total_increment (bl)
2431 const struct iv_class *bl;
2433 struct induction *v;
2436 /* For increment, must check every instruction that sets it. Each
2437 instruction must be executed only once each time through the loop.
2438 To verify this, we check that the insn is always executed, and that
2439 there are no backward branches after the insn that branch to before it.
2440 Also, the insn must have a mult_val of one (to make sure it really is
2443 result = const0_rtx;
2444 for (v = bl->biv; v; v = v->next_iv)
2446 if (v->always_computable && v->mult_val == const1_rtx
2447 && ! v->maybe_multiple)
2448 result = fold_rtx_mult_add (result, const1_rtx, v->add_val, v->mode);
2456 /* For each biv and giv, determine whether it can be safely split into
2457 a different variable for each unrolled copy of the loop body. If it
2458 is safe to split, then indicate that by saving some useful info
2459 in the splittable_regs array.
2461 If the loop is being completely unrolled, then splittable_regs will hold
2462 the current value of the induction variable while the loop is unrolled.
2463 It must be set to the initial value of the induction variable here.
2464 Otherwise, splittable_regs will hold the difference between the current
2465 value of the induction variable and the value the induction variable had
2466 at the top of the loop. It must be set to the value 0 here.
2468 Returns the total number of instructions that set registers that are
2471 /* ?? If the loop is only unrolled twice, then most of the restrictions to
2472 constant values are unnecessary, since we can easily calculate increment
2473 values in this case even if nothing is constant. The increment value
2474 should not involve a multiply however. */
2476 /* ?? Even if the biv/giv increment values aren't constant, it may still
2477 be beneficial to split the variable if the loop is only unrolled a few
2478 times, since multiplies by small integers (1,2,3,4) are very cheap. */
2481 find_splittable_regs (loop, unroll_type, unroll_number)
2482 const struct loop *loop;
2483 enum unroll_types unroll_type;
2486 struct loop_ivs *ivs = LOOP_IVS (loop);
2487 struct iv_class *bl;
2488 struct induction *v;
2490 rtx biv_final_value;
2494 for (bl = ivs->list; bl; bl = bl->next)
2496 /* Biv_total_increment must return a constant value,
2497 otherwise we can not calculate the split values. */
2499 increment = biv_total_increment (bl);
2500 if (! increment || GET_CODE (increment) != CONST_INT)
2503 /* The loop must be unrolled completely, or else have a known number
2504 of iterations and only one exit, or else the biv must be dead
2505 outside the loop, or else the final value must be known. Otherwise,
2506 it is unsafe to split the biv since it may not have the proper
2507 value on loop exit. */
2509 /* loop_number_exit_count is non-zero if the loop has an exit other than
2510 a fall through at the end. */
2513 biv_final_value = 0;
2514 if (unroll_type != UNROLL_COMPLETELY
2515 && (loop->exit_count || unroll_type == UNROLL_NAIVE)
2516 && (REGNO_LAST_LUID (bl->regno) >= INSN_LUID (loop->end)
2518 || INSN_UID (bl->init_insn) >= max_uid_for_loop
2519 || (REGNO_FIRST_LUID (bl->regno)
2520 < INSN_LUID (bl->init_insn))
2521 || reg_mentioned_p (bl->biv->dest_reg, SET_SRC (bl->init_set)))
2522 && ! (biv_final_value = final_biv_value (loop, bl)))
2525 /* If any of the insns setting the BIV don't do so with a simple
2526 PLUS, we don't know how to split it. */
2527 for (v = bl->biv; biv_splittable && v; v = v->next_iv)
2528 if ((tem = single_set (v->insn)) == 0
2529 || GET_CODE (SET_DEST (tem)) != REG
2530 || REGNO (SET_DEST (tem)) != bl->regno
2531 || GET_CODE (SET_SRC (tem)) != PLUS)
2534 /* If final value is non-zero, then must emit an instruction which sets
2535 the value of the biv to the proper value. This is done after
2536 handling all of the givs, since some of them may need to use the
2537 biv's value in their initialization code. */
2539 /* This biv is splittable. If completely unrolling the loop, save
2540 the biv's initial value. Otherwise, save the constant zero. */
2542 if (biv_splittable == 1)
2544 if (unroll_type == UNROLL_COMPLETELY)
2546 /* If the initial value of the biv is itself (i.e. it is too
2547 complicated for strength_reduce to compute), or is a hard
2548 register, or it isn't invariant, then we must create a new
2549 pseudo reg to hold the initial value of the biv. */
2551 if (GET_CODE (bl->initial_value) == REG
2552 && (REGNO (bl->initial_value) == bl->regno
2553 || REGNO (bl->initial_value) < FIRST_PSEUDO_REGISTER
2554 || ! loop_invariant_p (loop, bl->initial_value)))
2556 rtx tem = gen_reg_rtx (bl->biv->mode);
2558 record_base_value (REGNO (tem), bl->biv->add_val, 0);
2559 loop_insn_hoist (loop,
2560 gen_move_insn (tem, bl->biv->src_reg));
2562 if (loop_dump_stream)
2563 fprintf (loop_dump_stream,
2564 "Biv %d initial value remapped to %d.\n",
2565 bl->regno, REGNO (tem));
2567 splittable_regs[bl->regno] = tem;
2570 splittable_regs[bl->regno] = bl->initial_value;
2573 splittable_regs[bl->regno] = const0_rtx;
2575 /* Save the number of instructions that modify the biv, so that
2576 we can treat the last one specially. */
2578 splittable_regs_updates[bl->regno] = bl->biv_count;
2579 result += bl->biv_count;
2581 if (loop_dump_stream)
2582 fprintf (loop_dump_stream,
2583 "Biv %d safe to split.\n", bl->regno);
2586 /* Check every giv that depends on this biv to see whether it is
2587 splittable also. Even if the biv isn't splittable, givs which
2588 depend on it may be splittable if the biv is live outside the
2589 loop, and the givs aren't. */
2591 result += find_splittable_givs (loop, bl, unroll_type, increment,
2594 /* If final value is non-zero, then must emit an instruction which sets
2595 the value of the biv to the proper value. This is done after
2596 handling all of the givs, since some of them may need to use the
2597 biv's value in their initialization code. */
2598 if (biv_final_value)
2600 /* If the loop has multiple exits, emit the insns before the
2601 loop to ensure that it will always be executed no matter
2602 how the loop exits. Otherwise emit the insn after the loop,
2603 since this is slightly more efficient. */
2604 if (! loop->exit_count)
2605 loop_insn_sink (loop, gen_move_insn (bl->biv->src_reg,
2609 /* Create a new register to hold the value of the biv, and then
2610 set the biv to its final value before the loop start. The biv
2611 is set to its final value before loop start to ensure that
2612 this insn will always be executed, no matter how the loop
2614 rtx tem = gen_reg_rtx (bl->biv->mode);
2615 record_base_value (REGNO (tem), bl->biv->add_val, 0);
2617 loop_insn_hoist (loop, gen_move_insn (tem, bl->biv->src_reg));
2618 loop_insn_hoist (loop, gen_move_insn (bl->biv->src_reg,
2621 if (loop_dump_stream)
2622 fprintf (loop_dump_stream, "Biv %d mapped to %d for split.\n",
2623 REGNO (bl->biv->src_reg), REGNO (tem));
2625 /* Set up the mapping from the original biv register to the new
2627 bl->biv->src_reg = tem;
2634 /* Return 1 if the first and last unrolled copy of the address giv V is valid
2635 for the instruction that is using it. Do not make any changes to that
2639 verify_addresses (v, giv_inc, unroll_number)
2640 struct induction *v;
2645 rtx orig_addr = *v->location;
2646 rtx last_addr = plus_constant (v->dest_reg,
2647 INTVAL (giv_inc) * (unroll_number - 1));
2649 /* First check to see if either address would fail. Handle the fact
2650 that we have may have a match_dup. */
2651 if (! validate_replace_rtx (*v->location, v->dest_reg, v->insn)
2652 || ! validate_replace_rtx (*v->location, last_addr, v->insn))
2655 /* Now put things back the way they were before. This should always
2657 if (! validate_replace_rtx (*v->location, orig_addr, v->insn))
2663 /* For every giv based on the biv BL, check to determine whether it is
2664 splittable. This is a subroutine to find_splittable_regs ().
2666 Return the number of instructions that set splittable registers. */
2669 find_splittable_givs (loop, bl, unroll_type, increment, unroll_number)
2670 const struct loop *loop;
2671 struct iv_class *bl;
2672 enum unroll_types unroll_type;
2676 struct loop_ivs *ivs = LOOP_IVS (loop);
2677 struct induction *v, *v2;
2682 /* Scan the list of givs, and set the same_insn field when there are
2683 multiple identical givs in the same insn. */
2684 for (v = bl->giv; v; v = v->next_iv)
2685 for (v2 = v->next_iv; v2; v2 = v2->next_iv)
2686 if (v->insn == v2->insn && rtx_equal_p (v->new_reg, v2->new_reg)
2690 for (v = bl->giv; v; v = v->next_iv)
2694 /* Only split the giv if it has already been reduced, or if the loop is
2695 being completely unrolled. */
2696 if (unroll_type != UNROLL_COMPLETELY && v->ignore)
2699 /* The giv can be split if the insn that sets the giv is executed once
2700 and only once on every iteration of the loop. */
2701 /* An address giv can always be split. v->insn is just a use not a set,
2702 and hence it does not matter whether it is always executed. All that
2703 matters is that all the biv increments are always executed, and we
2704 won't reach here if they aren't. */
2705 if (v->giv_type != DEST_ADDR
2706 && (! v->always_computable
2707 || back_branch_in_range_p (loop, v->insn)))
2710 /* The giv increment value must be a constant. */
2711 giv_inc = fold_rtx_mult_add (v->mult_val, increment, const0_rtx,
2713 if (! giv_inc || GET_CODE (giv_inc) != CONST_INT)
2716 /* The loop must be unrolled completely, or else have a known number of
2717 iterations and only one exit, or else the giv must be dead outside
2718 the loop, or else the final value of the giv must be known.
2719 Otherwise, it is not safe to split the giv since it may not have the
2720 proper value on loop exit. */
2722 /* The used outside loop test will fail for DEST_ADDR givs. They are
2723 never used outside the loop anyways, so it is always safe to split a
2727 if (unroll_type != UNROLL_COMPLETELY
2728 && (loop->exit_count || unroll_type == UNROLL_NAIVE)
2729 && v->giv_type != DEST_ADDR
2730 /* The next part is true if the pseudo is used outside the loop.
2731 We assume that this is true for any pseudo created after loop
2732 starts, because we don't have a reg_n_info entry for them. */
2733 && (REGNO (v->dest_reg) >= max_reg_before_loop
2734 || (REGNO_FIRST_UID (REGNO (v->dest_reg)) != INSN_UID (v->insn)
2735 /* Check for the case where the pseudo is set by a shift/add
2736 sequence, in which case the first insn setting the pseudo
2737 is the first insn of the shift/add sequence. */
2738 && (! (tem = find_reg_note (v->insn, REG_RETVAL, NULL_RTX))
2739 || (REGNO_FIRST_UID (REGNO (v->dest_reg))
2740 != INSN_UID (XEXP (tem, 0)))))
2741 /* Line above always fails if INSN was moved by loop opt. */
2742 || (REGNO_LAST_LUID (REGNO (v->dest_reg))
2743 >= INSN_LUID (loop->end)))
2744 && ! (final_value = v->final_value))
2748 /* Currently, non-reduced/final-value givs are never split. */
2749 /* Should emit insns after the loop if possible, as the biv final value
2752 /* If the final value is non-zero, and the giv has not been reduced,
2753 then must emit an instruction to set the final value. */
2754 if (final_value && !v->new_reg)
2756 /* Create a new register to hold the value of the giv, and then set
2757 the giv to its final value before the loop start. The giv is set
2758 to its final value before loop start to ensure that this insn
2759 will always be executed, no matter how we exit. */
2760 tem = gen_reg_rtx (v->mode);
2761 loop_insn_hoist (loop, gen_move_insn (tem, v->dest_reg));
2762 loop_insn_hoist (loop, gen_move_insn (v->dest_reg, final_value));
2764 if (loop_dump_stream)
2765 fprintf (loop_dump_stream, "Giv %d mapped to %d for split.\n",
2766 REGNO (v->dest_reg), REGNO (tem));
2772 /* This giv is splittable. If completely unrolling the loop, save the
2773 giv's initial value. Otherwise, save the constant zero for it. */
2775 if (unroll_type == UNROLL_COMPLETELY)
2777 /* It is not safe to use bl->initial_value here, because it may not
2778 be invariant. It is safe to use the initial value stored in
2779 the splittable_regs array if it is set. In rare cases, it won't
2780 be set, so then we do exactly the same thing as
2781 find_splittable_regs does to get a safe value. */
2782 rtx biv_initial_value;
2784 if (splittable_regs[bl->regno])
2785 biv_initial_value = splittable_regs[bl->regno];
2786 else if (GET_CODE (bl->initial_value) != REG
2787 || (REGNO (bl->initial_value) != bl->regno
2788 && REGNO (bl->initial_value) >= FIRST_PSEUDO_REGISTER))
2789 biv_initial_value = bl->initial_value;
2792 rtx tem = gen_reg_rtx (bl->biv->mode);
2794 record_base_value (REGNO (tem), bl->biv->add_val, 0);
2795 loop_insn_hoist (loop, gen_move_insn (tem, bl->biv->src_reg));
2796 biv_initial_value = tem;
2798 biv_initial_value = extend_value_for_giv (v, biv_initial_value);
2799 value = fold_rtx_mult_add (v->mult_val, biv_initial_value,
2800 v->add_val, v->mode);
2807 /* If a giv was combined with another giv, then we can only split
2808 this giv if the giv it was combined with was reduced. This
2809 is because the value of v->new_reg is meaningless in this
2811 if (v->same && ! v->same->new_reg)
2813 if (loop_dump_stream)
2814 fprintf (loop_dump_stream,
2815 "giv combined with unreduced giv not split.\n");
2818 /* If the giv is an address destination, it could be something other
2819 than a simple register, these have to be treated differently. */
2820 else if (v->giv_type == DEST_REG)
2822 /* If value is not a constant, register, or register plus
2823 constant, then compute its value into a register before
2824 loop start. This prevents invalid rtx sharing, and should
2825 generate better code. We can use bl->initial_value here
2826 instead of splittable_regs[bl->regno] because this code
2827 is going before the loop start. */
2828 if (unroll_type == UNROLL_COMPLETELY
2829 && GET_CODE (value) != CONST_INT
2830 && GET_CODE (value) != REG
2831 && (GET_CODE (value) != PLUS
2832 || GET_CODE (XEXP (value, 0)) != REG
2833 || GET_CODE (XEXP (value, 1)) != CONST_INT))
2835 rtx tem = gen_reg_rtx (v->mode);
2836 record_base_value (REGNO (tem), v->add_val, 0);
2837 loop_iv_add_mult_hoist (loop, bl->initial_value, v->mult_val,
2842 splittable_regs[REGNO (v->new_reg)] = value;
2846 /* Splitting address givs is useful since it will often allow us
2847 to eliminate some increment insns for the base giv as
2850 /* If the addr giv is combined with a dest_reg giv, then all
2851 references to that dest reg will be remapped, which is NOT
2852 what we want for split addr regs. We always create a new
2853 register for the split addr giv, just to be safe. */
2855 /* If we have multiple identical address givs within a
2856 single instruction, then use a single pseudo reg for
2857 both. This is necessary in case one is a match_dup
2860 v->const_adjust = 0;
2864 v->dest_reg = v->same_insn->dest_reg;
2865 if (loop_dump_stream)
2866 fprintf (loop_dump_stream,
2867 "Sharing address givs in insn %d\n",
2868 INSN_UID (v->insn));
2870 /* If multiple address GIVs have been combined with the
2871 same dest_reg GIV, do not create a new register for
2873 else if (unroll_type != UNROLL_COMPLETELY
2874 && v->giv_type == DEST_ADDR
2875 && v->same && v->same->giv_type == DEST_ADDR
2876 && v->same->unrolled
2877 /* combine_givs_p may return true for some cases
2878 where the add and mult values are not equal.
2879 To share a register here, the values must be
2881 && rtx_equal_p (v->same->mult_val, v->mult_val)
2882 && rtx_equal_p (v->same->add_val, v->add_val)
2883 /* If the memory references have different modes,
2884 then the address may not be valid and we must
2885 not share registers. */
2886 && verify_addresses (v, giv_inc, unroll_number))
2888 v->dest_reg = v->same->dest_reg;
2891 else if (unroll_type != UNROLL_COMPLETELY)
2893 /* If not completely unrolling the loop, then create a new
2894 register to hold the split value of the DEST_ADDR giv.
2895 Emit insn to initialize its value before loop start. */
2897 rtx tem = gen_reg_rtx (v->mode);
2898 struct induction *same = v->same;
2899 rtx new_reg = v->new_reg;
2900 record_base_value (REGNO (tem), v->add_val, 0);
2902 /* If the address giv has a constant in its new_reg value,
2903 then this constant can be pulled out and put in value,
2904 instead of being part of the initialization code. */
2906 if (GET_CODE (new_reg) == PLUS
2907 && GET_CODE (XEXP (new_reg, 1)) == CONST_INT)
2910 = plus_constant (tem, INTVAL (XEXP (new_reg, 1)));
2912 /* Only succeed if this will give valid addresses.
2913 Try to validate both the first and the last
2914 address resulting from loop unrolling, if
2915 one fails, then can't do const elim here. */
2916 if (verify_addresses (v, giv_inc, unroll_number))
2918 /* Save the negative of the eliminated const, so
2919 that we can calculate the dest_reg's increment
2921 v->const_adjust = -INTVAL (XEXP (new_reg, 1));
2923 new_reg = XEXP (new_reg, 0);
2924 if (loop_dump_stream)
2925 fprintf (loop_dump_stream,
2926 "Eliminating constant from giv %d\n",
2935 /* If the address hasn't been checked for validity yet, do so
2936 now, and fail completely if either the first or the last
2937 unrolled copy of the address is not a valid address
2938 for the instruction that uses it. */
2939 if (v->dest_reg == tem
2940 && ! verify_addresses (v, giv_inc, unroll_number))
2942 for (v2 = v->next_iv; v2; v2 = v2->next_iv)
2943 if (v2->same_insn == v)
2946 if (loop_dump_stream)
2947 fprintf (loop_dump_stream,
2948 "Invalid address for giv at insn %d\n",
2949 INSN_UID (v->insn));
2953 v->new_reg = new_reg;
2956 /* We set this after the address check, to guarantee that
2957 the register will be initialized. */
2960 /* To initialize the new register, just move the value of
2961 new_reg into it. This is not guaranteed to give a valid
2962 instruction on machines with complex addressing modes.
2963 If we can't recognize it, then delete it and emit insns
2964 to calculate the value from scratch. */
2965 loop_insn_hoist (loop, gen_rtx_SET (VOIDmode, tem,
2966 copy_rtx (v->new_reg)));
2967 if (recog_memoized (PREV_INSN (loop->start)) < 0)
2971 /* We can't use bl->initial_value to compute the initial
2972 value, because the loop may have been preconditioned.
2973 We must calculate it from NEW_REG. */
2974 delete_related_insns (PREV_INSN (loop->start));
2977 ret = force_operand (v->new_reg, tem);
2979 emit_move_insn (tem, ret);
2980 sequence = gen_sequence ();
2982 loop_insn_hoist (loop, sequence);
2984 if (loop_dump_stream)
2985 fprintf (loop_dump_stream,
2986 "Invalid init insn, rewritten.\n");
2991 v->dest_reg = value;
2993 /* Check the resulting address for validity, and fail
2994 if the resulting address would be invalid. */
2995 if (! verify_addresses (v, giv_inc, unroll_number))
2997 for (v2 = v->next_iv; v2; v2 = v2->next_iv)
2998 if (v2->same_insn == v)
3001 if (loop_dump_stream)
3002 fprintf (loop_dump_stream,
3003 "Invalid address for giv at insn %d\n",
3004 INSN_UID (v->insn));
3009 /* Store the value of dest_reg into the insn. This sharing
3010 will not be a problem as this insn will always be copied
3013 *v->location = v->dest_reg;
3015 /* If this address giv is combined with a dest reg giv, then
3016 save the base giv's induction pointer so that we will be
3017 able to handle this address giv properly. The base giv
3018 itself does not have to be splittable. */
3020 if (v->same && v->same->giv_type == DEST_REG)
3021 addr_combined_regs[REGNO (v->same->new_reg)] = v->same;
3023 if (GET_CODE (v->new_reg) == REG)
3025 /* This giv maybe hasn't been combined with any others.
3026 Make sure that it's giv is marked as splittable here. */
3028 splittable_regs[REGNO (v->new_reg)] = value;
3030 /* Make it appear to depend upon itself, so that the
3031 giv will be properly split in the main loop above. */
3035 addr_combined_regs[REGNO (v->new_reg)] = v;
3039 if (loop_dump_stream)
3040 fprintf (loop_dump_stream, "DEST_ADDR giv being split.\n");
3046 /* Currently, unreduced giv's can't be split. This is not too much
3047 of a problem since unreduced giv's are not live across loop
3048 iterations anyways. When unrolling a loop completely though,
3049 it makes sense to reduce&split givs when possible, as this will
3050 result in simpler instructions, and will not require that a reg
3051 be live across loop iterations. */
3053 splittable_regs[REGNO (v->dest_reg)] = value;
3054 fprintf (stderr, "Giv %d at insn %d not reduced\n",
3055 REGNO (v->dest_reg), INSN_UID (v->insn));
3061 /* Unreduced givs are only updated once by definition. Reduced givs
3062 are updated as many times as their biv is. Mark it so if this is
3063 a splittable register. Don't need to do anything for address givs
3064 where this may not be a register. */
3066 if (GET_CODE (v->new_reg) == REG)
3070 count = REG_IV_CLASS (ivs, REGNO (v->src_reg))->biv_count;
3072 splittable_regs_updates[REGNO (v->new_reg)] = count;
3077 if (loop_dump_stream)
3081 if (GET_CODE (v->dest_reg) == CONST_INT)
3083 else if (GET_CODE (v->dest_reg) != REG)
3084 regnum = REGNO (XEXP (v->dest_reg, 0));
3086 regnum = REGNO (v->dest_reg);
3087 fprintf (loop_dump_stream, "Giv %d at insn %d safe to split.\n",
3088 regnum, INSN_UID (v->insn));
3095 /* Try to prove that the register is dead after the loop exits. Trace every
3096 loop exit looking for an insn that will always be executed, which sets
3097 the register to some value, and appears before the first use of the register
3098 is found. If successful, then return 1, otherwise return 0. */
3100 /* ?? Could be made more intelligent in the handling of jumps, so that
3101 it can search past if statements and other similar structures. */
3104 reg_dead_after_loop (loop, reg)
3105 const struct loop *loop;
3111 int label_count = 0;
3113 /* In addition to checking all exits of this loop, we must also check
3114 all exits of inner nested loops that would exit this loop. We don't
3115 have any way to identify those, so we just give up if there are any
3116 such inner loop exits. */
3118 for (label = loop->exit_labels; label; label = LABEL_NEXTREF (label))
3121 if (label_count != loop->exit_count)
3124 /* HACK: Must also search the loop fall through exit, create a label_ref
3125 here which points to the loop->end, and append the loop_number_exit_labels
3127 label = gen_rtx_LABEL_REF (VOIDmode, loop->end);
3128 LABEL_NEXTREF (label) = loop->exit_labels;
3130 for (; label; label = LABEL_NEXTREF (label))
3132 /* Succeed if find an insn which sets the biv or if reach end of
3133 function. Fail if find an insn that uses the biv, or if come to
3134 a conditional jump. */
3136 insn = NEXT_INSN (XEXP (label, 0));
3139 code = GET_CODE (insn);
3140 if (GET_RTX_CLASS (code) == 'i')
3144 if (reg_referenced_p (reg, PATTERN (insn)))
3147 set = single_set (insn);
3148 if (set && rtx_equal_p (SET_DEST (set), reg))
3152 if (code == JUMP_INSN)
3154 if (GET_CODE (PATTERN (insn)) == RETURN)
3156 else if (!any_uncondjump_p (insn)
3157 /* Prevent infinite loop following infinite loops. */
3158 || jump_count++ > 20)
3161 insn = JUMP_LABEL (insn);
3164 insn = NEXT_INSN (insn);
3168 /* Success, the register is dead on all loop exits. */
3172 /* Try to calculate the final value of the biv, the value it will have at
3173 the end of the loop. If we can do it, return that value. */
3176 final_biv_value (loop, bl)
3177 const struct loop *loop;
3178 struct iv_class *bl;
3180 unsigned HOST_WIDE_INT n_iterations = LOOP_INFO (loop)->n_iterations;
3183 /* ??? This only works for MODE_INT biv's. Reject all others for now. */
3185 if (GET_MODE_CLASS (bl->biv->mode) != MODE_INT)
3188 /* The final value for reversed bivs must be calculated differently than
3189 for ordinary bivs. In this case, there is already an insn after the
3190 loop which sets this biv's final value (if necessary), and there are
3191 no other loop exits, so we can return any value. */
3194 if (loop_dump_stream)
3195 fprintf (loop_dump_stream,
3196 "Final biv value for %d, reversed biv.\n", bl->regno);
3201 /* Try to calculate the final value as initial value + (number of iterations
3202 * increment). For this to work, increment must be invariant, the only
3203 exit from the loop must be the fall through at the bottom (otherwise
3204 it may not have its final value when the loop exits), and the initial
3205 value of the biv must be invariant. */
3207 if (n_iterations != 0
3208 && ! loop->exit_count
3209 && loop_invariant_p (loop, bl->initial_value))
3211 increment = biv_total_increment (bl);
3213 if (increment && loop_invariant_p (loop, increment))
3215 /* Can calculate the loop exit value, emit insns after loop
3216 end to calculate this value into a temporary register in
3217 case it is needed later. */
3219 tem = gen_reg_rtx (bl->biv->mode);
3220 record_base_value (REGNO (tem), bl->biv->add_val, 0);
3221 loop_iv_add_mult_sink (loop, increment, GEN_INT (n_iterations),
3222 bl->initial_value, tem);
3224 if (loop_dump_stream)
3225 fprintf (loop_dump_stream,
3226 "Final biv value for %d, calculated.\n", bl->regno);
3232 /* Check to see if the biv is dead at all loop exits. */
3233 if (reg_dead_after_loop (loop, bl->biv->src_reg))
3235 if (loop_dump_stream)
3236 fprintf (loop_dump_stream,
3237 "Final biv value for %d, biv dead after loop exit.\n",
3246 /* Try to calculate the final value of the giv, the value it will have at
3247 the end of the loop. If we can do it, return that value. */
3250 final_giv_value (loop, v)
3251 const struct loop *loop;
3252 struct induction *v;
3254 struct loop_ivs *ivs = LOOP_IVS (loop);
3255 struct iv_class *bl;
3259 rtx loop_end = loop->end;
3260 unsigned HOST_WIDE_INT n_iterations = LOOP_INFO (loop)->n_iterations;
3262 bl = REG_IV_CLASS (ivs, REGNO (v->src_reg));
3264 /* The final value for givs which depend on reversed bivs must be calculated
3265 differently than for ordinary givs. In this case, there is already an
3266 insn after the loop which sets this giv's final value (if necessary),
3267 and there are no other loop exits, so we can return any value. */
3270 if (loop_dump_stream)
3271 fprintf (loop_dump_stream,
3272 "Final giv value for %d, depends on reversed biv\n",
3273 REGNO (v->dest_reg));
3277 /* Try to calculate the final value as a function of the biv it depends
3278 upon. The only exit from the loop must be the fall through at the bottom
3279 (otherwise it may not have its final value when the loop exits). */
3281 /* ??? Can calculate the final giv value by subtracting off the
3282 extra biv increments times the giv's mult_val. The loop must have
3283 only one exit for this to work, but the loop iterations does not need
3286 if (n_iterations != 0
3287 && ! loop->exit_count)
3289 /* ?? It is tempting to use the biv's value here since these insns will
3290 be put after the loop, and hence the biv will have its final value
3291 then. However, this fails if the biv is subsequently eliminated.
3292 Perhaps determine whether biv's are eliminable before trying to
3293 determine whether giv's are replaceable so that we can use the
3294 biv value here if it is not eliminable. */
3296 /* We are emitting code after the end of the loop, so we must make
3297 sure that bl->initial_value is still valid then. It will still
3298 be valid if it is invariant. */
3300 increment = biv_total_increment (bl);
3302 if (increment && loop_invariant_p (loop, increment)
3303 && loop_invariant_p (loop, bl->initial_value))
3305 /* Can calculate the loop exit value of its biv as
3306 (n_iterations * increment) + initial_value */
3308 /* The loop exit value of the giv is then
3309 (final_biv_value - extra increments) * mult_val + add_val.
3310 The extra increments are any increments to the biv which
3311 occur in the loop after the giv's value is calculated.
3312 We must search from the insn that sets the giv to the end
3313 of the loop to calculate this value. */
3315 /* Put the final biv value in tem. */
3316 tem = gen_reg_rtx (v->mode);
3317 record_base_value (REGNO (tem), bl->biv->add_val, 0);
3318 loop_iv_add_mult_sink (loop, extend_value_for_giv (v, increment),
3319 GEN_INT (n_iterations),
3320 extend_value_for_giv (v, bl->initial_value),
3323 /* Subtract off extra increments as we find them. */
3324 for (insn = NEXT_INSN (v->insn); insn != loop_end;
3325 insn = NEXT_INSN (insn))
3327 struct induction *biv;
3329 for (biv = bl->biv; biv; biv = biv->next_iv)
3330 if (biv->insn == insn)
3333 tem = expand_simple_binop (GET_MODE (tem), MINUS, tem,
3334 biv->add_val, NULL_RTX, 0,
3336 seq = gen_sequence ();
3338 loop_insn_sink (loop, seq);
3342 /* Now calculate the giv's final value. */
3343 loop_iv_add_mult_sink (loop, tem, v->mult_val, v->add_val, tem);
3345 if (loop_dump_stream)
3346 fprintf (loop_dump_stream,
3347 "Final giv value for %d, calc from biv's value.\n",
3348 REGNO (v->dest_reg));
3354 /* Replaceable giv's should never reach here. */
3358 /* Check to see if the biv is dead at all loop exits. */
3359 if (reg_dead_after_loop (loop, v->dest_reg))
3361 if (loop_dump_stream)
3362 fprintf (loop_dump_stream,
3363 "Final giv value for %d, giv dead after loop exit.\n",
3364 REGNO (v->dest_reg));
3372 /* Look back before LOOP->START for the insn that sets REG and return
3373 the equivalent constant if there is a REG_EQUAL note otherwise just
3374 the SET_SRC of REG. */
3377 loop_find_equiv_value (loop, reg)
3378 const struct loop *loop;
3381 rtx loop_start = loop->start;
3386 for (insn = PREV_INSN (loop_start); insn; insn = PREV_INSN (insn))
3388 if (GET_CODE (insn) == CODE_LABEL)
3391 else if (INSN_P (insn) && reg_set_p (reg, insn))
3393 /* We found the last insn before the loop that sets the register.
3394 If it sets the entire register, and has a REG_EQUAL note,
3395 then use the value of the REG_EQUAL note. */
3396 if ((set = single_set (insn))
3397 && (SET_DEST (set) == reg))
3399 rtx note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
3401 /* Only use the REG_EQUAL note if it is a constant.
3402 Other things, divide in particular, will cause
3403 problems later if we use them. */
3404 if (note && GET_CODE (XEXP (note, 0)) != EXPR_LIST
3405 && CONSTANT_P (XEXP (note, 0)))
3406 ret = XEXP (note, 0);
3408 ret = SET_SRC (set);
3410 /* We cannot do this if it changes between the
3411 assignment and loop start though. */
3412 if (modified_between_p (ret, insn, loop_start))
3421 /* Return a simplified rtx for the expression OP - REG.
3423 REG must appear in OP, and OP must be a register or the sum of a register
3426 Thus, the return value must be const0_rtx or the second term.
3428 The caller is responsible for verifying that REG appears in OP and OP has
3432 subtract_reg_term (op, reg)
3437 if (GET_CODE (op) == PLUS)
3439 if (XEXP (op, 0) == reg)
3440 return XEXP (op, 1);
3441 else if (XEXP (op, 1) == reg)
3442 return XEXP (op, 0);
3444 /* OP does not contain REG as a term. */
3448 /* Find and return register term common to both expressions OP0 and
3449 OP1 or NULL_RTX if no such term exists. Each expression must be a
3450 REG or a PLUS of a REG. */
3453 find_common_reg_term (op0, op1)
3456 if ((GET_CODE (op0) == REG || GET_CODE (op0) == PLUS)
3457 && (GET_CODE (op1) == REG || GET_CODE (op1) == PLUS))
3464 if (GET_CODE (op0) == PLUS)
3465 op01 = XEXP (op0, 1), op00 = XEXP (op0, 0);
3467 op01 = const0_rtx, op00 = op0;
3469 if (GET_CODE (op1) == PLUS)
3470 op11 = XEXP (op1, 1), op10 = XEXP (op1, 0);
3472 op11 = const0_rtx, op10 = op1;
3474 /* Find and return common register term if present. */
3475 if (REG_P (op00) && (op00 == op10 || op00 == op11))
3477 else if (REG_P (op01) && (op01 == op10 || op01 == op11))
3481 /* No common register term found. */
3485 /* Determine the loop iterator and calculate the number of loop
3486 iterations. Returns the exact number of loop iterations if it can
3487 be calculated, otherwise returns zero. */
3489 unsigned HOST_WIDE_INT
3490 loop_iterations (loop)
3493 struct loop_info *loop_info = LOOP_INFO (loop);
3494 struct loop_ivs *ivs = LOOP_IVS (loop);
3495 rtx comparison, comparison_value;
3496 rtx iteration_var, initial_value, increment, final_value;
3497 enum rtx_code comparison_code;
3499 unsigned HOST_WIDE_INT abs_inc;
3500 unsigned HOST_WIDE_INT abs_diff;
3503 int unsigned_p, compare_dir, final_larger;
3506 struct iv_class *bl;
3508 loop_info->n_iterations = 0;
3509 loop_info->initial_value = 0;
3510 loop_info->initial_equiv_value = 0;
3511 loop_info->comparison_value = 0;
3512 loop_info->final_value = 0;
3513 loop_info->final_equiv_value = 0;
3514 loop_info->increment = 0;
3515 loop_info->iteration_var = 0;
3516 loop_info->unroll_number = 1;
3519 /* We used to use prev_nonnote_insn here, but that fails because it might
3520 accidentally get the branch for a contained loop if the branch for this
3521 loop was deleted. We can only trust branches immediately before the
3523 last_loop_insn = PREV_INSN (loop->end);
3525 /* ??? We should probably try harder to find the jump insn
3526 at the end of the loop. The following code assumes that
3527 the last loop insn is a jump to the top of the loop. */
3528 if (GET_CODE (last_loop_insn) != JUMP_INSN)
3530 if (loop_dump_stream)
3531 fprintf (loop_dump_stream,
3532 "Loop iterations: No final conditional branch found.\n");
3536 /* If there is a more than a single jump to the top of the loop
3537 we cannot (easily) determine the iteration count. */
3538 if (LABEL_NUSES (JUMP_LABEL (last_loop_insn)) > 1)
3540 if (loop_dump_stream)
3541 fprintf (loop_dump_stream,
3542 "Loop iterations: Loop has multiple back edges.\n");
3546 /* If there are multiple conditionalized loop exit tests, they may jump
3547 back to differing CODE_LABELs. */
3548 if (loop->top && loop->cont)
3550 rtx temp = PREV_INSN (last_loop_insn);
3554 if (GET_CODE (temp) == JUMP_INSN
3555 /* Previous unrolling may have generated new insns not covered
3556 by the uid_luid array. */
3557 && INSN_UID (JUMP_LABEL (temp)) < max_uid_for_loop
3558 /* Check if we jump back into the loop body. */
3559 && INSN_LUID (JUMP_LABEL (temp)) > INSN_LUID (loop->top)
3560 && INSN_LUID (JUMP_LABEL (temp)) < INSN_LUID (loop->cont))
3562 if (loop_dump_stream)
3563 fprintf (loop_dump_stream,
3564 "Loop iterations: Loop has multiple back edges.\n");
3568 while ((temp = PREV_INSN (temp)) != loop->cont);
3571 /* Find the iteration variable. If the last insn is a conditional
3572 branch, and the insn before tests a register value, make that the
3573 iteration variable. */
3575 comparison = get_condition_for_loop (loop, last_loop_insn);
3576 if (comparison == 0)
3578 if (loop_dump_stream)
3579 fprintf (loop_dump_stream,
3580 "Loop iterations: No final comparison found.\n");
3584 /* ??? Get_condition may switch position of induction variable and
3585 invariant register when it canonicalizes the comparison. */
3587 comparison_code = GET_CODE (comparison);
3588 iteration_var = XEXP (comparison, 0);
3589 comparison_value = XEXP (comparison, 1);
3591 if (GET_CODE (iteration_var) != REG)
3593 if (loop_dump_stream)
3594 fprintf (loop_dump_stream,
3595 "Loop iterations: Comparison not against register.\n");
3599 /* The only new registers that are created before loop iterations
3600 are givs made from biv increments or registers created by
3601 load_mems. In the latter case, it is possible that try_copy_prop
3602 will propagate a new pseudo into the old iteration register but
3603 this will be marked by having the REG_USERVAR_P bit set. */
3605 if ((unsigned) REGNO (iteration_var) >= ivs->n_regs
3606 && ! REG_USERVAR_P (iteration_var))
3609 /* Determine the initial value of the iteration variable, and the amount
3610 that it is incremented each loop. Use the tables constructed by
3611 the strength reduction pass to calculate these values. */
3613 /* Clear the result values, in case no answer can be found. */
3617 /* The iteration variable can be either a giv or a biv. Check to see
3618 which it is, and compute the variable's initial value, and increment
3619 value if possible. */
3621 /* If this is a new register, can't handle it since we don't have any
3622 reg_iv_type entry for it. */
3623 if ((unsigned) REGNO (iteration_var) >= ivs->n_regs)
3625 if (loop_dump_stream)
3626 fprintf (loop_dump_stream,
3627 "Loop iterations: No reg_iv_type entry for iteration var.\n");
3631 /* Reject iteration variables larger than the host wide int size, since they
3632 could result in a number of iterations greater than the range of our
3633 `unsigned HOST_WIDE_INT' variable loop_info->n_iterations. */
3634 else if ((GET_MODE_BITSIZE (GET_MODE (iteration_var))
3635 > HOST_BITS_PER_WIDE_INT))
3637 if (loop_dump_stream)
3638 fprintf (loop_dump_stream,
3639 "Loop iterations: Iteration var rejected because mode too large.\n");
3642 else if (GET_MODE_CLASS (GET_MODE (iteration_var)) != MODE_INT)
3644 if (loop_dump_stream)
3645 fprintf (loop_dump_stream,
3646 "Loop iterations: Iteration var not an integer.\n");
3649 else if (REG_IV_TYPE (ivs, REGNO (iteration_var)) == BASIC_INDUCT)
3651 if (REGNO (iteration_var) >= ivs->n_regs)
3654 /* Grab initial value, only useful if it is a constant. */
3655 bl = REG_IV_CLASS (ivs, REGNO (iteration_var));
3656 initial_value = bl->initial_value;
3657 if (!bl->biv->always_executed || bl->biv->maybe_multiple)
3659 if (loop_dump_stream)
3660 fprintf (loop_dump_stream,
3661 "Loop iterations: Basic induction var not set once in each iteration.\n");
3665 increment = biv_total_increment (bl);
3667 else if (REG_IV_TYPE (ivs, REGNO (iteration_var)) == GENERAL_INDUCT)
3669 HOST_WIDE_INT offset = 0;
3670 struct induction *v = REG_IV_INFO (ivs, REGNO (iteration_var));
3671 rtx biv_initial_value;
3673 if (REGNO (v->src_reg) >= ivs->n_regs)
3676 if (!v->always_executed || v->maybe_multiple)
3678 if (loop_dump_stream)
3679 fprintf (loop_dump_stream,
3680 "Loop iterations: General induction var not set once in each iteration.\n");
3684 bl = REG_IV_CLASS (ivs, REGNO (v->src_reg));
3686 /* Increment value is mult_val times the increment value of the biv. */
3688 increment = biv_total_increment (bl);
3691 struct induction *biv_inc;
3693 increment = fold_rtx_mult_add (v->mult_val,
3694 extend_value_for_giv (v, increment),
3695 const0_rtx, v->mode);
3696 /* The caller assumes that one full increment has occurred at the
3697 first loop test. But that's not true when the biv is incremented
3698 after the giv is set (which is the usual case), e.g.:
3699 i = 6; do {;} while (i++ < 9) .
3700 Therefore, we bias the initial value by subtracting the amount of
3701 the increment that occurs between the giv set and the giv test. */
3702 for (biv_inc = bl->biv; biv_inc; biv_inc = biv_inc->next_iv)
3704 if (loop_insn_first_p (v->insn, biv_inc->insn))
3705 offset -= INTVAL (biv_inc->add_val);
3708 if (loop_dump_stream)
3709 fprintf (loop_dump_stream,
3710 "Loop iterations: Giv iterator, initial value bias %ld.\n",
3713 /* Initial value is mult_val times the biv's initial value plus
3714 add_val. Only useful if it is a constant. */
3715 biv_initial_value = extend_value_for_giv (v, bl->initial_value);
3717 = fold_rtx_mult_add (v->mult_val,
3718 plus_constant (biv_initial_value, offset),
3719 v->add_val, v->mode);
3723 if (loop_dump_stream)
3724 fprintf (loop_dump_stream,
3725 "Loop iterations: Not basic or general induction var.\n");
3729 if (initial_value == 0)
3734 switch (comparison_code)
3749 /* Cannot determine loop iterations with this case. */
3768 /* If the comparison value is an invariant register, then try to find
3769 its value from the insns before the start of the loop. */
3771 final_value = comparison_value;
3772 if (GET_CODE (comparison_value) == REG
3773 && loop_invariant_p (loop, comparison_value))
3775 final_value = loop_find_equiv_value (loop, comparison_value);
3777 /* If we don't get an invariant final value, we are better
3778 off with the original register. */
3779 if (! loop_invariant_p (loop, final_value))
3780 final_value = comparison_value;
3783 /* Calculate the approximate final value of the induction variable
3784 (on the last successful iteration). The exact final value
3785 depends on the branch operator, and increment sign. It will be
3786 wrong if the iteration variable is not incremented by one each
3787 time through the loop and (comparison_value + off_by_one -
3788 initial_value) % increment != 0.
3789 ??? Note that the final_value may overflow and thus final_larger
3790 will be bogus. A potentially infinite loop will be classified
3791 as immediate, e.g. for (i = 0x7ffffff0; i <= 0x7fffffff; i++) */
3793 final_value = plus_constant (final_value, off_by_one);
3795 /* Save the calculated values describing this loop's bounds, in case
3796 precondition_loop_p will need them later. These values can not be
3797 recalculated inside precondition_loop_p because strength reduction
3798 optimizations may obscure the loop's structure.
3800 These values are only required by precondition_loop_p and insert_bct
3801 whenever the number of iterations cannot be computed at compile time.
3802 Only the difference between final_value and initial_value is
3803 important. Note that final_value is only approximate. */
3804 loop_info->initial_value = initial_value;
3805 loop_info->comparison_value = comparison_value;
3806 loop_info->final_value = plus_constant (comparison_value, off_by_one);
3807 loop_info->increment = increment;
3808 loop_info->iteration_var = iteration_var;
3809 loop_info->comparison_code = comparison_code;
3812 /* Try to determine the iteration count for loops such
3813 as (for i = init; i < init + const; i++). When running the
3814 loop optimization twice, the first pass often converts simple
3815 loops into this form. */
3817 if (REG_P (initial_value))
3823 reg1 = initial_value;
3824 if (GET_CODE (final_value) == PLUS)
3825 reg2 = XEXP (final_value, 0), const2 = XEXP (final_value, 1);
3827 reg2 = final_value, const2 = const0_rtx;
3829 /* Check for initial_value = reg1, final_value = reg2 + const2,
3830 where reg1 != reg2. */
3831 if (REG_P (reg2) && reg2 != reg1)
3835 /* Find what reg1 is equivalent to. Hopefully it will
3836 either be reg2 or reg2 plus a constant. */
3837 temp = loop_find_equiv_value (loop, reg1);
3839 if (find_common_reg_term (temp, reg2))
3840 initial_value = temp;
3843 /* Find what reg2 is equivalent to. Hopefully it will
3844 either be reg1 or reg1 plus a constant. Let's ignore
3845 the latter case for now since it is not so common. */
3846 temp = loop_find_equiv_value (loop, reg2);
3848 if (temp == loop_info->iteration_var)
3849 temp = initial_value;
3851 final_value = (const2 == const0_rtx)
3852 ? reg1 : gen_rtx_PLUS (GET_MODE (reg1), reg1, const2);
3855 else if (loop->vtop && GET_CODE (reg2) == CONST_INT)
3859 /* When running the loop optimizer twice, check_dbra_loop
3860 further obfuscates reversible loops of the form:
3861 for (i = init; i < init + const; i++). We often end up with
3862 final_value = 0, initial_value = temp, temp = temp2 - init,
3863 where temp2 = init + const. If the loop has a vtop we
3864 can replace initial_value with const. */
3866 temp = loop_find_equiv_value (loop, reg1);
3868 if (GET_CODE (temp) == MINUS && REG_P (XEXP (temp, 0)))
3870 rtx temp2 = loop_find_equiv_value (loop, XEXP (temp, 0));
3872 if (GET_CODE (temp2) == PLUS
3873 && XEXP (temp2, 0) == XEXP (temp, 1))
3874 initial_value = XEXP (temp2, 1);
3879 /* If have initial_value = reg + const1 and final_value = reg +
3880 const2, then replace initial_value with const1 and final_value
3881 with const2. This should be safe since we are protected by the
3882 initial comparison before entering the loop if we have a vtop.
3883 For example, a + b < a + c is not equivalent to b < c for all a
3884 when using modulo arithmetic.
3886 ??? Without a vtop we could still perform the optimization if we check
3887 the initial and final values carefully. */
3889 && (reg_term = find_common_reg_term (initial_value, final_value)))
3891 initial_value = subtract_reg_term (initial_value, reg_term);
3892 final_value = subtract_reg_term (final_value, reg_term);
3895 loop_info->initial_equiv_value = initial_value;
3896 loop_info->final_equiv_value = final_value;
3898 /* For EQ comparison loops, we don't have a valid final value.
3899 Check this now so that we won't leave an invalid value if we
3900 return early for any other reason. */
3901 if (comparison_code == EQ)
3902 loop_info->final_equiv_value = loop_info->final_value = 0;
3906 if (loop_dump_stream)
3907 fprintf (loop_dump_stream,
3908 "Loop iterations: Increment value can't be calculated.\n");
3912 if (GET_CODE (increment) != CONST_INT)
3914 /* If we have a REG, check to see if REG holds a constant value. */
3915 /* ??? Other RTL, such as (neg (reg)) is possible here, but it isn't
3916 clear if it is worthwhile to try to handle such RTL. */
3917 if (GET_CODE (increment) == REG || GET_CODE (increment) == SUBREG)
3918 increment = loop_find_equiv_value (loop, increment);
3920 if (GET_CODE (increment) != CONST_INT)
3922 if (loop_dump_stream)
3924 fprintf (loop_dump_stream,
3925 "Loop iterations: Increment value not constant ");
3926 print_simple_rtl (loop_dump_stream, increment);
3927 fprintf (loop_dump_stream, ".\n");
3931 loop_info->increment = increment;
3934 if (GET_CODE (initial_value) != CONST_INT)
3936 if (loop_dump_stream)
3938 fprintf (loop_dump_stream,
3939 "Loop iterations: Initial value not constant ");
3940 print_simple_rtl (loop_dump_stream, initial_value);
3941 fprintf (loop_dump_stream, ".\n");
3945 else if (comparison_code == EQ)
3947 if (loop_dump_stream)
3948 fprintf (loop_dump_stream, "Loop iterations: EQ comparison loop.\n");
3951 else if (GET_CODE (final_value) != CONST_INT)
3953 if (loop_dump_stream)
3955 fprintf (loop_dump_stream,
3956 "Loop iterations: Final value not constant ");
3957 print_simple_rtl (loop_dump_stream, final_value);
3958 fprintf (loop_dump_stream, ".\n");
3963 /* Final_larger is 1 if final larger, 0 if they are equal, otherwise -1. */
3966 = ((unsigned HOST_WIDE_INT) INTVAL (final_value)
3967 > (unsigned HOST_WIDE_INT) INTVAL (initial_value))
3968 - ((unsigned HOST_WIDE_INT) INTVAL (final_value)
3969 < (unsigned HOST_WIDE_INT) INTVAL (initial_value));
3971 final_larger = (INTVAL (final_value) > INTVAL (initial_value))
3972 - (INTVAL (final_value) < INTVAL (initial_value));
3974 if (INTVAL (increment) > 0)
3976 else if (INTVAL (increment) == 0)
3981 /* There are 27 different cases: compare_dir = -1, 0, 1;
3982 final_larger = -1, 0, 1; increment_dir = -1, 0, 1.
3983 There are 4 normal cases, 4 reverse cases (where the iteration variable
3984 will overflow before the loop exits), 4 infinite loop cases, and 15
3985 immediate exit (0 or 1 iteration depending on loop type) cases.
3986 Only try to optimize the normal cases. */
3988 /* (compare_dir/final_larger/increment_dir)
3989 Normal cases: (0/-1/-1), (0/1/1), (-1/-1/-1), (1/1/1)
3990 Reverse cases: (0/-1/1), (0/1/-1), (-1/-1/1), (1/1/-1)
3991 Infinite loops: (0/-1/0), (0/1/0), (-1/-1/0), (1/1/0)
3992 Immediate exit: (0/0/X), (-1/0/X), (-1/1/X), (1/0/X), (1/-1/X) */
3994 /* ?? If the meaning of reverse loops (where the iteration variable
3995 will overflow before the loop exits) is undefined, then could
3996 eliminate all of these special checks, and just always assume
3997 the loops are normal/immediate/infinite. Note that this means
3998 the sign of increment_dir does not have to be known. Also,
3999 since it does not really hurt if immediate exit loops or infinite loops
4000 are optimized, then that case could be ignored also, and hence all
4001 loops can be optimized.
4003 According to ANSI Spec, the reverse loop case result is undefined,
4004 because the action on overflow is undefined.
4006 See also the special test for NE loops below. */
4008 if (final_larger == increment_dir && final_larger != 0
4009 && (final_larger == compare_dir || compare_dir == 0))
4014 if (loop_dump_stream)
4015 fprintf (loop_dump_stream, "Loop iterations: Not normal loop.\n");
4019 /* Calculate the number of iterations, final_value is only an approximation,
4020 so correct for that. Note that abs_diff and n_iterations are
4021 unsigned, because they can be as large as 2^n - 1. */
4023 inc = INTVAL (increment);
4026 abs_diff = INTVAL (final_value) - INTVAL (initial_value);
4031 abs_diff = INTVAL (initial_value) - INTVAL (final_value);
4037 /* Given that iteration_var is going to iterate over its own mode,
4038 not HOST_WIDE_INT, disregard higher bits that might have come
4039 into the picture due to sign extension of initial and final
4041 abs_diff &= ((unsigned HOST_WIDE_INT)1
4042 << (GET_MODE_BITSIZE (GET_MODE (iteration_var)) - 1)
4045 /* For NE tests, make sure that the iteration variable won't miss
4046 the final value. If abs_diff mod abs_incr is not zero, then the
4047 iteration variable will overflow before the loop exits, and we
4048 can not calculate the number of iterations. */
4049 if (compare_dir == 0 && (abs_diff % abs_inc) != 0)
4052 /* Note that the number of iterations could be calculated using
4053 (abs_diff + abs_inc - 1) / abs_inc, provided care was taken to
4054 handle potential overflow of the summation. */
4055 loop_info->n_iterations = abs_diff / abs_inc + ((abs_diff % abs_inc) != 0);
4056 return loop_info->n_iterations;
4059 /* Replace uses of split bivs with their split pseudo register. This is
4060 for original instructions which remain after loop unrolling without
4064 remap_split_bivs (loop, x)
4068 struct loop_ivs *ivs = LOOP_IVS (loop);
4076 code = GET_CODE (x);
4091 /* If non-reduced/final-value givs were split, then this would also
4092 have to remap those givs also. */
4094 if (REGNO (x) < ivs->n_regs
4095 && REG_IV_TYPE (ivs, REGNO (x)) == BASIC_INDUCT)
4096 return REG_IV_CLASS (ivs, REGNO (x))->biv->src_reg;
4103 fmt = GET_RTX_FORMAT (code);
4104 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4107 XEXP (x, i) = remap_split_bivs (loop, XEXP (x, i));
4108 else if (fmt[i] == 'E')
4111 for (j = 0; j < XVECLEN (x, i); j++)
4112 XVECEXP (x, i, j) = remap_split_bivs (loop, XVECEXP (x, i, j));
4118 /* If FIRST_UID is a set of REGNO, and FIRST_UID dominates LAST_UID (e.g.
4119 FIST_UID is always executed if LAST_UID is), then return 1. Otherwise
4120 return 0. COPY_START is where we can start looking for the insns
4121 FIRST_UID and LAST_UID. COPY_END is where we stop looking for these
4124 If there is no JUMP_INSN between LOOP_START and FIRST_UID, then FIRST_UID
4125 must dominate LAST_UID.
4127 If there is a CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID
4128 may not dominate LAST_UID.
4130 If there is no CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID
4131 must dominate LAST_UID. */
4134 set_dominates_use (regno, first_uid, last_uid, copy_start, copy_end)
4141 int passed_jump = 0;
4142 rtx p = NEXT_INSN (copy_start);
4144 while (INSN_UID (p) != first_uid)
4146 if (GET_CODE (p) == JUMP_INSN)
4148 /* Could not find FIRST_UID. */
4154 /* Verify that FIRST_UID is an insn that entirely sets REGNO. */
4155 if (! INSN_P (p) || ! dead_or_set_regno_p (p, regno))
4158 /* FIRST_UID is always executed. */
4159 if (passed_jump == 0)
4162 while (INSN_UID (p) != last_uid)
4164 /* If we see a CODE_LABEL between FIRST_UID and LAST_UID, then we
4165 can not be sure that FIRST_UID dominates LAST_UID. */
4166 if (GET_CODE (p) == CODE_LABEL)
4168 /* Could not find LAST_UID, but we reached the end of the loop, so
4170 else if (p == copy_end)
4175 /* FIRST_UID is always executed if LAST_UID is executed. */
4179 /* This routine is called when the number of iterations for the unrolled
4180 loop is one. The goal is to identify a loop that begins with an
4181 unconditional branch to the loop continuation note (or a label just after).
4182 In this case, the unconditional branch that starts the loop needs to be
4183 deleted so that we execute the single iteration. */
4186 ujump_to_loop_cont (loop_start, loop_cont)
4190 rtx x, label, label_ref;
4192 /* See if loop start, or the next insn is an unconditional jump. */
4193 loop_start = next_nonnote_insn (loop_start);
4195 x = pc_set (loop_start);
4199 label_ref = SET_SRC (x);
4203 /* Examine insn after loop continuation note. Return if not a label. */
4204 label = next_nonnote_insn (loop_cont);
4205 if (label == 0 || GET_CODE (label) != CODE_LABEL)
4208 /* Return the loop start if the branch label matches the code label. */
4209 if (CODE_LABEL_NUMBER (label) == CODE_LABEL_NUMBER (XEXP (label_ref, 0)))