1 /* Try to unroll loops, and split induction variables.
2 Copyright (C) 1992, 93, 94, 95, 97, 98, 1999 Free Software Foundation, Inc.
3 Contributed by James E. Wilson, Cygnus Support/UC Berkeley.
5 This file is part of GNU CC.
7 GNU CC is free software; you can redistribute it and/or modify
8 it under the terms of the GNU General Public License as published by
9 the Free Software Foundation; either version 2, or (at your option)
12 GNU CC is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License for more details.
17 You should have received a copy of the GNU General Public License
18 along with GNU CC; see the file COPYING. If not, write to
19 the Free Software Foundation, 59 Temple Place - Suite 330,
20 Boston, MA 02111-1307, USA. */
22 /* Try to unroll a loop, and split induction variables.
24 Loops for which the number of iterations can be calculated exactly are
25 handled specially. If the number of iterations times the insn_count is
26 less than MAX_UNROLLED_INSNS, then the loop is unrolled completely.
27 Otherwise, we try to unroll the loop a number of times modulo the number
28 of iterations, so that only one exit test will be needed. It is unrolled
29 a number of times approximately equal to MAX_UNROLLED_INSNS divided by
32 Otherwise, if the number of iterations can be calculated exactly at
33 run time, and the loop is always entered at the top, then we try to
34 precondition the loop. That is, at run time, calculate how many times
35 the loop will execute, and then execute the loop body a few times so
36 that the remaining iterations will be some multiple of 4 (or 2 if the
37 loop is large). Then fall through to a loop unrolled 4 (or 2) times,
38 with only one exit test needed at the end of the loop.
40 Otherwise, if the number of iterations can not be calculated exactly,
41 not even at run time, then we still unroll the loop a number of times
42 approximately equal to MAX_UNROLLED_INSNS divided by the insn count,
43 but there must be an exit test after each copy of the loop body.
45 For each induction variable, which is dead outside the loop (replaceable)
46 or for which we can easily calculate the final value, if we can easily
47 calculate its value at each place where it is set as a function of the
48 current loop unroll count and the variable's value at loop entry, then
49 the induction variable is split into `N' different variables, one for
50 each copy of the loop body. One variable is live across the backward
51 branch, and the others are all calculated as a function of this variable.
52 This helps eliminate data dependencies, and leads to further opportunities
55 /* Possible improvements follow: */
57 /* ??? Add an extra pass somewhere to determine whether unrolling will
58 give any benefit. E.g. after generating all unrolled insns, compute the
59 cost of all insns and compare against cost of insns in rolled loop.
61 - On traditional architectures, unrolling a non-constant bound loop
62 is a win if there is a giv whose only use is in memory addresses, the
63 memory addresses can be split, and hence giv increments can be
65 - It is also a win if the loop is executed many times, and preconditioning
66 can be performed for the loop.
67 Add code to check for these and similar cases. */
69 /* ??? Improve control of which loops get unrolled. Could use profiling
70 info to only unroll the most commonly executed loops. Perhaps have
71 a user specifyable option to control the amount of code expansion,
72 or the percent of loops to consider for unrolling. Etc. */
74 /* ??? Look at the register copies inside the loop to see if they form a
75 simple permutation. If so, iterate the permutation until it gets back to
76 the start state. This is how many times we should unroll the loop, for
77 best results, because then all register copies can be eliminated.
78 For example, the lisp nreverse function should be unrolled 3 times
87 ??? The number of times to unroll the loop may also be based on data
88 references in the loop. For example, if we have a loop that references
89 x[i-1], x[i], and x[i+1], we should unroll it a multiple of 3 times. */
91 /* ??? Add some simple linear equation solving capability so that we can
92 determine the number of loop iterations for more complex loops.
93 For example, consider this loop from gdb
94 #define SWAP_TARGET_AND_HOST(buffer,len)
97 char *p = (char *) buffer;
98 char *q = ((char *) buffer) + len - 1;
99 int iterations = (len + 1) >> 1;
101 for (p; p < q; p++, q--;)
109 start value = p = &buffer + current_iteration
110 end value = q = &buffer + len - 1 - current_iteration
111 Given the loop exit test of "p < q", then there must be "q - p" iterations,
112 set equal to zero and solve for number of iterations:
113 q - p = len - 1 - 2*current_iteration = 0
114 current_iteration = (len - 1) / 2
115 Hence, there are (len - 1) / 2 (rounded up to the nearest integer)
116 iterations of this loop. */
118 /* ??? Currently, no labels are marked as loop invariant when doing loop
119 unrolling. This is because an insn inside the loop, that loads the address
120 of a label inside the loop into a register, could be moved outside the loop
121 by the invariant code motion pass if labels were invariant. If the loop
122 is subsequently unrolled, the code will be wrong because each unrolled
123 body of the loop will use the same address, whereas each actually needs a
124 different address. A case where this happens is when a loop containing
125 a switch statement is unrolled.
127 It would be better to let labels be considered invariant. When we
128 unroll loops here, check to see if any insns using a label local to the
129 loop were moved before the loop. If so, then correct the problem, by
130 moving the insn back into the loop, or perhaps replicate the insn before
131 the loop, one copy for each time the loop is unrolled. */
133 /* The prime factors looked for when trying to unroll a loop by some
134 number which is modulo the total number of iterations. Just checking
135 for these 4 prime factors will find at least one factor for 75% of
136 all numbers theoretically. Practically speaking, this will succeed
137 almost all of the time since loops are generally a multiple of 2
140 #define NUM_FACTORS 4
142 struct _factor { int factor, count; } factors[NUM_FACTORS]
143 = { {2, 0}, {3, 0}, {5, 0}, {7, 0}};
145 /* Describes the different types of loop unrolling performed. */
147 enum unroll_types { UNROLL_COMPLETELY, UNROLL_MODULO, UNROLL_NAIVE };
153 #include "insn-config.h"
154 #include "integrate.h"
158 #include "function.h"
163 /* This controls which loops are unrolled, and by how much we unroll
166 #ifndef MAX_UNROLLED_INSNS
167 #define MAX_UNROLLED_INSNS 100
170 /* Indexed by register number, if non-zero, then it contains a pointer
171 to a struct induction for a DEST_REG giv which has been combined with
172 one of more address givs. This is needed because whenever such a DEST_REG
173 giv is modified, we must modify the value of all split address givs
174 that were combined with this DEST_REG giv. */
176 static struct induction **addr_combined_regs;
178 /* Indexed by register number, if this is a splittable induction variable,
179 then this will hold the current value of the register, which depends on the
182 static rtx *splittable_regs;
184 /* Indexed by register number, if this is a splittable induction variable,
185 this indicates if it was made from a derived giv. */
186 static char *derived_regs;
188 /* Indexed by register number, if this is a splittable induction variable,
189 then this will hold the number of instructions in the loop that modify
190 the induction variable. Used to ensure that only the last insn modifying
191 a split iv will update the original iv of the dest. */
193 static int *splittable_regs_updates;
195 /* Forward declarations. */
197 static void init_reg_map PROTO((struct inline_remap *, int));
198 static rtx calculate_giv_inc PROTO((rtx, rtx, int));
199 static rtx initial_reg_note_copy PROTO((rtx, struct inline_remap *));
200 static void final_reg_note_copy PROTO((rtx, struct inline_remap *));
201 static void copy_loop_body PROTO((rtx, rtx, struct inline_remap *, rtx, int,
202 enum unroll_types, rtx, rtx, rtx, rtx));
203 static void iteration_info PROTO((rtx, rtx *, rtx *, rtx, rtx));
204 static int find_splittable_regs PROTO((enum unroll_types, rtx, rtx, rtx, int,
205 unsigned HOST_WIDE_INT));
206 static int find_splittable_givs PROTO((struct iv_class *, enum unroll_types,
207 rtx, rtx, rtx, int));
208 static int reg_dead_after_loop PROTO((rtx, rtx, rtx));
209 static rtx fold_rtx_mult_add PROTO((rtx, rtx, rtx, enum machine_mode));
210 static int verify_addresses PROTO((struct induction *, rtx, int));
211 static rtx remap_split_bivs PROTO((rtx));
212 static rtx find_common_reg_term PROTO((rtx, rtx));
213 static rtx subtract_reg_term PROTO((rtx, rtx));
214 static rtx loop_find_equiv_value PROTO((rtx, rtx));
216 /* Try to unroll one loop and split induction variables in the loop.
218 The loop is described by the arguments LOOP_END, INSN_COUNT, and
219 LOOP_START. END_INSERT_BEFORE indicates where insns should be added
220 which need to be executed when the loop falls through. STRENGTH_REDUCTION_P
221 indicates whether information generated in the strength reduction pass
224 This function is intended to be called from within `strength_reduce'
228 unroll_loop (loop_end, insn_count, loop_start, end_insert_before,
229 loop_info, strength_reduce_p)
233 rtx end_insert_before;
234 struct loop_info *loop_info;
235 int strength_reduce_p;
238 unsigned HOST_WIDE_INT temp;
239 int unroll_number = 1;
240 rtx copy_start, copy_end;
241 rtx insn, sequence, pattern, tem;
242 int max_labelno, max_insnno;
244 struct inline_remap *map;
245 char *local_label = NULL;
247 int max_local_regnum;
252 int splitting_not_safe = 0;
253 enum unroll_types unroll_type;
254 int loop_preconditioned = 0;
256 /* This points to the last real insn in the loop, which should be either
257 a JUMP_INSN (for conditional jumps) or a BARRIER (for unconditional
261 /* Don't bother unrolling huge loops. Since the minimum factor is
262 two, loops greater than one half of MAX_UNROLLED_INSNS will never
264 if (insn_count > MAX_UNROLLED_INSNS / 2)
266 if (loop_dump_stream)
267 fprintf (loop_dump_stream, "Unrolling failure: Loop too big.\n");
271 /* When emitting debugger info, we can't unroll loops with unequal numbers
272 of block_beg and block_end notes, because that would unbalance the block
273 structure of the function. This can happen as a result of the
274 "if (foo) bar; else break;" optimization in jump.c. */
275 /* ??? Gcc has a general policy that -g is never supposed to change the code
276 that the compiler emits, so we must disable this optimization always,
277 even if debug info is not being output. This is rare, so this should
278 not be a significant performance problem. */
280 if (1 /* write_symbols != NO_DEBUG */)
282 int block_begins = 0;
285 for (insn = loop_start; insn != loop_end; insn = NEXT_INSN (insn))
287 if (GET_CODE (insn) == NOTE)
289 if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_BLOCK_BEG)
291 else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_BLOCK_END)
296 if (block_begins != block_ends)
298 if (loop_dump_stream)
299 fprintf (loop_dump_stream,
300 "Unrolling failure: Unbalanced block notes.\n");
305 /* Determine type of unroll to perform. Depends on the number of iterations
306 and the size of the loop. */
308 /* If there is no strength reduce info, then set
309 loop_info->n_iterations to zero. This can happen if
310 strength_reduce can't find any bivs in the loop. A value of zero
311 indicates that the number of iterations could not be calculated. */
313 if (! strength_reduce_p)
314 loop_info->n_iterations = 0;
316 if (loop_dump_stream && loop_info->n_iterations > 0)
318 fputs ("Loop unrolling: ", loop_dump_stream);
319 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC,
320 loop_info->n_iterations);
321 fputs (" iterations.\n", loop_dump_stream);
324 /* Find and save a pointer to the last nonnote insn in the loop. */
326 last_loop_insn = prev_nonnote_insn (loop_end);
328 /* Calculate how many times to unroll the loop. Indicate whether or
329 not the loop is being completely unrolled. */
331 if (loop_info->n_iterations == 1)
333 /* If number of iterations is exactly 1, then eliminate the compare and
334 branch at the end of the loop since they will never be taken.
335 Then return, since no other action is needed here. */
337 /* If the last instruction is not a BARRIER or a JUMP_INSN, then
338 don't do anything. */
340 if (GET_CODE (last_loop_insn) == BARRIER)
342 /* Delete the jump insn. This will delete the barrier also. */
343 delete_insn (PREV_INSN (last_loop_insn));
345 else if (GET_CODE (last_loop_insn) == JUMP_INSN)
348 rtx prev = PREV_INSN (last_loop_insn);
350 delete_insn (last_loop_insn);
352 /* The immediately preceding insn may be a compare which must be
354 if (sets_cc0_p (prev))
359 /* Remove the loop notes since this is no longer a loop. */
361 delete_insn (loop_info->vtop);
363 delete_insn (loop_info->cont);
365 delete_insn (loop_start);
367 delete_insn (loop_end);
371 else if (loop_info->n_iterations > 0
372 && loop_info->n_iterations * insn_count < MAX_UNROLLED_INSNS)
374 unroll_number = loop_info->n_iterations;
375 unroll_type = UNROLL_COMPLETELY;
377 else if (loop_info->n_iterations > 0)
379 /* Try to factor the number of iterations. Don't bother with the
380 general case, only using 2, 3, 5, and 7 will get 75% of all
381 numbers theoretically, and almost all in practice. */
383 for (i = 0; i < NUM_FACTORS; i++)
384 factors[i].count = 0;
386 temp = loop_info->n_iterations;
387 for (i = NUM_FACTORS - 1; i >= 0; i--)
388 while (temp % factors[i].factor == 0)
391 temp = temp / factors[i].factor;
394 /* Start with the larger factors first so that we generally
395 get lots of unrolling. */
399 for (i = 3; i >= 0; i--)
400 while (factors[i].count--)
402 if (temp * factors[i].factor < MAX_UNROLLED_INSNS)
404 unroll_number *= factors[i].factor;
405 temp *= factors[i].factor;
411 /* If we couldn't find any factors, then unroll as in the normal
413 if (unroll_number == 1)
415 if (loop_dump_stream)
416 fprintf (loop_dump_stream,
417 "Loop unrolling: No factors found.\n");
420 unroll_type = UNROLL_MODULO;
424 /* Default case, calculate number of times to unroll loop based on its
426 if (unroll_number == 1)
428 if (8 * insn_count < MAX_UNROLLED_INSNS)
430 else if (4 * insn_count < MAX_UNROLLED_INSNS)
435 unroll_type = UNROLL_NAIVE;
438 /* Now we know how many times to unroll the loop. */
440 if (loop_dump_stream)
441 fprintf (loop_dump_stream,
442 "Unrolling loop %d times.\n", unroll_number);
445 if (unroll_type == UNROLL_COMPLETELY || unroll_type == UNROLL_MODULO)
447 /* Loops of these types can start with jump down to the exit condition
448 in rare circumstances.
450 Consider a pair of nested loops where the inner loop is part
451 of the exit code for the outer loop.
453 In this case jump.c will not duplicate the exit test for the outer
454 loop, so it will start with a jump to the exit code.
456 Then consider if the inner loop turns out to iterate once and
457 only once. We will end up deleting the jumps associated with
458 the inner loop. However, the loop notes are not removed from
459 the instruction stream.
461 And finally assume that we can compute the number of iterations
464 In this case unroll may want to unroll the outer loop even though
465 it starts with a jump to the outer loop's exit code.
467 We could try to optimize this case, but it hardly seems worth it.
468 Just return without unrolling the loop in such cases. */
471 while (GET_CODE (insn) != CODE_LABEL && GET_CODE (insn) != JUMP_INSN)
472 insn = NEXT_INSN (insn);
473 if (GET_CODE (insn) == JUMP_INSN)
477 if (unroll_type == UNROLL_COMPLETELY)
479 /* Completely unrolling the loop: Delete the compare and branch at
480 the end (the last two instructions). This delete must done at the
481 very end of loop unrolling, to avoid problems with calls to
482 back_branch_in_range_p, which is called by find_splittable_regs.
483 All increments of splittable bivs/givs are changed to load constant
486 copy_start = loop_start;
488 /* Set insert_before to the instruction immediately after the JUMP_INSN
489 (or BARRIER), so that any NOTEs between the JUMP_INSN and the end of
490 the loop will be correctly handled by copy_loop_body. */
491 insert_before = NEXT_INSN (last_loop_insn);
493 /* Set copy_end to the insn before the jump at the end of the loop. */
494 if (GET_CODE (last_loop_insn) == BARRIER)
495 copy_end = PREV_INSN (PREV_INSN (last_loop_insn));
496 else if (GET_CODE (last_loop_insn) == JUMP_INSN)
498 copy_end = PREV_INSN (last_loop_insn);
500 /* The instruction immediately before the JUMP_INSN may be a compare
501 instruction which we do not want to copy. */
502 if (sets_cc0_p (PREV_INSN (copy_end)))
503 copy_end = PREV_INSN (copy_end);
508 /* We currently can't unroll a loop if it doesn't end with a
509 JUMP_INSN. There would need to be a mechanism that recognizes
510 this case, and then inserts a jump after each loop body, which
511 jumps to after the last loop body. */
512 if (loop_dump_stream)
513 fprintf (loop_dump_stream,
514 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
518 else if (unroll_type == UNROLL_MODULO)
520 /* Partially unrolling the loop: The compare and branch at the end
521 (the last two instructions) must remain. Don't copy the compare
522 and branch instructions at the end of the loop. Insert the unrolled
523 code immediately before the compare/branch at the end so that the
524 code will fall through to them as before. */
526 copy_start = loop_start;
528 /* Set insert_before to the jump insn at the end of the loop.
529 Set copy_end to before the jump insn at the end of the loop. */
530 if (GET_CODE (last_loop_insn) == BARRIER)
532 insert_before = PREV_INSN (last_loop_insn);
533 copy_end = PREV_INSN (insert_before);
535 else if (GET_CODE (last_loop_insn) == JUMP_INSN)
537 insert_before = last_loop_insn;
539 /* The instruction immediately before the JUMP_INSN may be a compare
540 instruction which we do not want to copy or delete. */
541 if (sets_cc0_p (PREV_INSN (insert_before)))
542 insert_before = PREV_INSN (insert_before);
544 copy_end = PREV_INSN (insert_before);
548 /* We currently can't unroll a loop if it doesn't end with a
549 JUMP_INSN. There would need to be a mechanism that recognizes
550 this case, and then inserts a jump after each loop body, which
551 jumps to after the last loop body. */
552 if (loop_dump_stream)
553 fprintf (loop_dump_stream,
554 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
560 /* Normal case: Must copy the compare and branch instructions at the
563 if (GET_CODE (last_loop_insn) == BARRIER)
565 /* Loop ends with an unconditional jump and a barrier.
566 Handle this like above, don't copy jump and barrier.
567 This is not strictly necessary, but doing so prevents generating
568 unconditional jumps to an immediately following label.
570 This will be corrected below if the target of this jump is
571 not the start_label. */
573 insert_before = PREV_INSN (last_loop_insn);
574 copy_end = PREV_INSN (insert_before);
576 else if (GET_CODE (last_loop_insn) == JUMP_INSN)
578 /* Set insert_before to immediately after the JUMP_INSN, so that
579 NOTEs at the end of the loop will be correctly handled by
581 insert_before = NEXT_INSN (last_loop_insn);
582 copy_end = last_loop_insn;
586 /* We currently can't unroll a loop if it doesn't end with a
587 JUMP_INSN. There would need to be a mechanism that recognizes
588 this case, and then inserts a jump after each loop body, which
589 jumps to after the last loop body. */
590 if (loop_dump_stream)
591 fprintf (loop_dump_stream,
592 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
596 /* If copying exit test branches because they can not be eliminated,
597 then must convert the fall through case of the branch to a jump past
598 the end of the loop. Create a label to emit after the loop and save
599 it for later use. Do not use the label after the loop, if any, since
600 it might be used by insns outside the loop, or there might be insns
601 added before it later by final_[bg]iv_value which must be after
602 the real exit label. */
603 exit_label = gen_label_rtx ();
606 while (GET_CODE (insn) != CODE_LABEL && GET_CODE (insn) != JUMP_INSN)
607 insn = NEXT_INSN (insn);
609 if (GET_CODE (insn) == JUMP_INSN)
611 /* The loop starts with a jump down to the exit condition test.
612 Start copying the loop after the barrier following this
614 copy_start = NEXT_INSN (insn);
616 /* Splitting induction variables doesn't work when the loop is
617 entered via a jump to the bottom, because then we end up doing
618 a comparison against a new register for a split variable, but
619 we did not execute the set insn for the new register because
620 it was skipped over. */
621 splitting_not_safe = 1;
622 if (loop_dump_stream)
623 fprintf (loop_dump_stream,
624 "Splitting not safe, because loop not entered at top.\n");
627 copy_start = loop_start;
630 /* This should always be the first label in the loop. */
631 start_label = NEXT_INSN (copy_start);
632 /* There may be a line number note and/or a loop continue note here. */
633 while (GET_CODE (start_label) == NOTE)
634 start_label = NEXT_INSN (start_label);
635 if (GET_CODE (start_label) != CODE_LABEL)
637 /* This can happen as a result of jump threading. If the first insns in
638 the loop test the same condition as the loop's backward jump, or the
639 opposite condition, then the backward jump will be modified to point
640 to elsewhere, and the loop's start label is deleted.
642 This case currently can not be handled by the loop unrolling code. */
644 if (loop_dump_stream)
645 fprintf (loop_dump_stream,
646 "Unrolling failure: unknown insns between BEG note and loop label.\n");
649 if (LABEL_NAME (start_label))
651 /* The jump optimization pass must have combined the original start label
652 with a named label for a goto. We can't unroll this case because
653 jumps which go to the named label must be handled differently than
654 jumps to the loop start, and it is impossible to differentiate them
656 if (loop_dump_stream)
657 fprintf (loop_dump_stream,
658 "Unrolling failure: loop start label is gone\n");
662 if (unroll_type == UNROLL_NAIVE
663 && GET_CODE (last_loop_insn) == BARRIER
664 && GET_CODE (PREV_INSN (last_loop_insn)) == JUMP_INSN
665 && start_label != JUMP_LABEL (PREV_INSN (last_loop_insn)))
667 /* In this case, we must copy the jump and barrier, because they will
668 not be converted to jumps to an immediately following label. */
670 insert_before = NEXT_INSN (last_loop_insn);
671 copy_end = last_loop_insn;
674 if (unroll_type == UNROLL_NAIVE
675 && GET_CODE (last_loop_insn) == JUMP_INSN
676 && start_label != JUMP_LABEL (last_loop_insn))
678 /* ??? The loop ends with a conditional branch that does not branch back
679 to the loop start label. In this case, we must emit an unconditional
680 branch to the loop exit after emitting the final branch.
681 copy_loop_body does not have support for this currently, so we
682 give up. It doesn't seem worthwhile to unroll anyways since
683 unrolling would increase the number of branch instructions
685 if (loop_dump_stream)
686 fprintf (loop_dump_stream,
687 "Unrolling failure: final conditional branch not to loop start\n");
691 /* Allocate a translation table for the labels and insn numbers.
692 They will be filled in as we copy the insns in the loop. */
694 max_labelno = max_label_num ();
695 max_insnno = get_max_uid ();
697 /* Various paths through the unroll code may reach the "egress" label
698 without initializing fields within the map structure.
700 To be safe, we use xcalloc to zero the memory. */
701 map = (struct inline_remap *) xcalloc (1, sizeof (struct inline_remap));
703 /* Allocate the label map. */
707 map->label_map = (rtx *) xmalloc (max_labelno * sizeof (rtx));
709 local_label = (char *) xcalloc (max_labelno, sizeof (char));
712 /* Search the loop and mark all local labels, i.e. the ones which have to
713 be distinct labels when copied. For all labels which might be
714 non-local, set their label_map entries to point to themselves.
715 If they happen to be local their label_map entries will be overwritten
716 before the loop body is copied. The label_map entries for local labels
717 will be set to a different value each time the loop body is copied. */
719 for (insn = copy_start; insn != loop_end; insn = NEXT_INSN (insn))
723 if (GET_CODE (insn) == CODE_LABEL)
724 local_label[CODE_LABEL_NUMBER (insn)] = 1;
725 else if (GET_CODE (insn) == JUMP_INSN)
727 if (JUMP_LABEL (insn))
728 set_label_in_map (map,
729 CODE_LABEL_NUMBER (JUMP_LABEL (insn)),
731 else if (GET_CODE (PATTERN (insn)) == ADDR_VEC
732 || GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC)
734 rtx pat = PATTERN (insn);
735 int diff_vec_p = GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC;
736 int len = XVECLEN (pat, diff_vec_p);
739 for (i = 0; i < len; i++)
741 label = XEXP (XVECEXP (pat, diff_vec_p, i), 0);
742 set_label_in_map (map,
743 CODE_LABEL_NUMBER (label),
748 else if ((note = find_reg_note (insn, REG_LABEL, NULL_RTX)))
749 set_label_in_map (map, CODE_LABEL_NUMBER (XEXP (note, 0)),
753 /* Allocate space for the insn map. */
755 map->insn_map = (rtx *) xmalloc (max_insnno * sizeof (rtx));
757 /* Set this to zero, to indicate that we are doing loop unrolling,
758 not function inlining. */
759 map->inline_target = 0;
761 /* The register and constant maps depend on the number of registers
762 present, so the final maps can't be created until after
763 find_splittable_regs is called. However, they are needed for
764 preconditioning, so we create temporary maps when preconditioning
767 /* The preconditioning code may allocate two new pseudo registers. */
768 maxregnum = max_reg_num ();
770 /* local_regno is only valid for regnos < max_local_regnum. */
771 max_local_regnum = maxregnum;
773 /* Allocate and zero out the splittable_regs and addr_combined_regs
774 arrays. These must be zeroed here because they will be used if
775 loop preconditioning is performed, and must be zero for that case.
777 It is safe to do this here, since the extra registers created by the
778 preconditioning code and find_splittable_regs will never be used
779 to access the splittable_regs[] and addr_combined_regs[] arrays. */
781 splittable_regs = (rtx *) xcalloc (maxregnum, sizeof (rtx));
782 derived_regs = (char *) xcalloc (maxregnum, sizeof (char));
783 splittable_regs_updates = (int *) xcalloc (maxregnum, sizeof (int));
785 = (struct induction **) xcalloc (maxregnum, sizeof (struct induction *));
786 local_regno = (char *) xcalloc (maxregnum, sizeof (char));
788 /* Mark all local registers, i.e. the ones which are referenced only
790 if (INSN_UID (copy_end) < max_uid_for_loop)
792 int copy_start_luid = INSN_LUID (copy_start);
793 int copy_end_luid = INSN_LUID (copy_end);
795 /* If a register is used in the jump insn, we must not duplicate it
796 since it will also be used outside the loop. */
797 if (GET_CODE (copy_end) == JUMP_INSN)
800 /* If we have a target that uses cc0, then we also must not duplicate
801 the insn that sets cc0 before the jump insn, if one is present. */
803 if (GET_CODE (copy_end) == JUMP_INSN && sets_cc0_p (PREV_INSN (copy_end)))
807 /* If copy_start points to the NOTE that starts the loop, then we must
808 use the next luid, because invariant pseudo-regs moved out of the loop
809 have their lifetimes modified to start here, but they are not safe
811 if (copy_start == loop_start)
814 /* If a pseudo's lifetime is entirely contained within this loop, then we
815 can use a different pseudo in each unrolled copy of the loop. This
816 results in better code. */
817 /* We must limit the generic test to max_reg_before_loop, because only
818 these pseudo registers have valid regno_first_uid info. */
819 for (j = FIRST_PSEUDO_REGISTER; j < max_reg_before_loop; ++j)
820 if (REGNO_FIRST_UID (j) > 0 && REGNO_FIRST_UID (j) <= max_uid_for_loop
821 && uid_luid[REGNO_FIRST_UID (j)] >= copy_start_luid
822 && REGNO_LAST_UID (j) > 0 && REGNO_LAST_UID (j) <= max_uid_for_loop
823 && uid_luid[REGNO_LAST_UID (j)] <= copy_end_luid)
825 /* However, we must also check for loop-carried dependencies.
826 If the value the pseudo has at the end of iteration X is
827 used by iteration X+1, then we can not use a different pseudo
828 for each unrolled copy of the loop. */
829 /* A pseudo is safe if regno_first_uid is a set, and this
830 set dominates all instructions from regno_first_uid to
832 /* ??? This check is simplistic. We would get better code if
833 this check was more sophisticated. */
834 if (set_dominates_use (j, REGNO_FIRST_UID (j), REGNO_LAST_UID (j),
835 copy_start, copy_end))
838 if (loop_dump_stream)
841 fprintf (loop_dump_stream, "Marked reg %d as local\n", j);
843 fprintf (loop_dump_stream, "Did not mark reg %d as local\n",
847 /* Givs that have been created from multiple biv increments always have
849 for (j = first_increment_giv; j <= last_increment_giv; j++)
852 if (loop_dump_stream)
853 fprintf (loop_dump_stream, "Marked reg %d as local\n", j);
857 /* If this loop requires exit tests when unrolled, check to see if we
858 can precondition the loop so as to make the exit tests unnecessary.
859 Just like variable splitting, this is not safe if the loop is entered
860 via a jump to the bottom. Also, can not do this if no strength
861 reduce info, because precondition_loop_p uses this info. */
863 /* Must copy the loop body for preconditioning before the following
864 find_splittable_regs call since that will emit insns which need to
865 be after the preconditioned loop copies, but immediately before the
866 unrolled loop copies. */
868 /* Also, it is not safe to split induction variables for the preconditioned
869 copies of the loop body. If we split induction variables, then the code
870 assumes that each induction variable can be represented as a function
871 of its initial value and the loop iteration number. This is not true
872 in this case, because the last preconditioned copy of the loop body
873 could be any iteration from the first up to the `unroll_number-1'th,
874 depending on the initial value of the iteration variable. Therefore
875 we can not split induction variables here, because we can not calculate
876 their value. Hence, this code must occur before find_splittable_regs
879 if (unroll_type == UNROLL_NAIVE && ! splitting_not_safe && strength_reduce_p)
881 rtx initial_value, final_value, increment;
882 enum machine_mode mode;
884 if (precondition_loop_p (loop_start, loop_info,
885 &initial_value, &final_value, &increment,
890 int abs_inc, neg_inc;
892 map->reg_map = (rtx *) xmalloc (maxregnum * sizeof (rtx));
894 VARRAY_CONST_EQUIV_INIT (map->const_equiv_varray, maxregnum,
896 global_const_equiv_varray = map->const_equiv_varray;
898 init_reg_map (map, maxregnum);
900 /* Limit loop unrolling to 4, since this will make 7 copies of
902 if (unroll_number > 4)
905 /* Save the absolute value of the increment, and also whether or
906 not it is negative. */
908 abs_inc = INTVAL (increment);
917 /* Calculate the difference between the final and initial values.
918 Final value may be a (plus (reg x) (const_int 1)) rtx.
919 Let the following cse pass simplify this if initial value is
922 We must copy the final and initial values here to avoid
923 improperly shared rtl. */
925 diff = expand_binop (mode, sub_optab, copy_rtx (final_value),
926 copy_rtx (initial_value), NULL_RTX, 0,
929 /* Now calculate (diff % (unroll * abs (increment))) by using an
931 diff = expand_binop (GET_MODE (diff), and_optab, diff,
932 GEN_INT (unroll_number * abs_inc - 1),
933 NULL_RTX, 0, OPTAB_LIB_WIDEN);
935 /* Now emit a sequence of branches to jump to the proper precond
938 labels = (rtx *) xmalloc (sizeof (rtx) * unroll_number);
939 for (i = 0; i < unroll_number; i++)
940 labels[i] = gen_label_rtx ();
942 /* Check for the case where the initial value is greater than or
943 equal to the final value. In that case, we want to execute
944 exactly one loop iteration. The code below will fail for this
945 case. This check does not apply if the loop has a NE
946 comparison at the end. */
948 if (loop_info->comparison_code != NE)
950 emit_cmp_and_jump_insns (initial_value, final_value,
952 NULL_RTX, mode, 0, 0, labels[1]);
953 JUMP_LABEL (get_last_insn ()) = labels[1];
954 LABEL_NUSES (labels[1])++;
957 /* Assuming the unroll_number is 4, and the increment is 2, then
958 for a negative increment: for a positive increment:
959 diff = 0,1 precond 0 diff = 0,7 precond 0
960 diff = 2,3 precond 3 diff = 1,2 precond 1
961 diff = 4,5 precond 2 diff = 3,4 precond 2
962 diff = 6,7 precond 1 diff = 5,6 precond 3 */
964 /* We only need to emit (unroll_number - 1) branches here, the
965 last case just falls through to the following code. */
967 /* ??? This would give better code if we emitted a tree of branches
968 instead of the current linear list of branches. */
970 for (i = 0; i < unroll_number - 1; i++)
973 enum rtx_code cmp_code;
975 /* For negative increments, must invert the constant compared
976 against, except when comparing against zero. */
984 cmp_const = unroll_number - i;
993 emit_cmp_and_jump_insns (diff, GEN_INT (abs_inc * cmp_const),
994 cmp_code, NULL_RTX, mode, 0, 0,
996 JUMP_LABEL (get_last_insn ()) = labels[i];
997 LABEL_NUSES (labels[i])++;
1000 /* If the increment is greater than one, then we need another branch,
1001 to handle other cases equivalent to 0. */
1003 /* ??? This should be merged into the code above somehow to help
1004 simplify the code here, and reduce the number of branches emitted.
1005 For the negative increment case, the branch here could easily
1006 be merged with the `0' case branch above. For the positive
1007 increment case, it is not clear how this can be simplified. */
1012 enum rtx_code cmp_code;
1016 cmp_const = abs_inc - 1;
1021 cmp_const = abs_inc * (unroll_number - 1) + 1;
1025 emit_cmp_and_jump_insns (diff, GEN_INT (cmp_const), cmp_code,
1026 NULL_RTX, mode, 0, 0, labels[0]);
1027 JUMP_LABEL (get_last_insn ()) = labels[0];
1028 LABEL_NUSES (labels[0])++;
1031 sequence = gen_sequence ();
1033 emit_insn_before (sequence, loop_start);
1035 /* Only the last copy of the loop body here needs the exit
1036 test, so set copy_end to exclude the compare/branch here,
1037 and then reset it inside the loop when get to the last
1040 if (GET_CODE (last_loop_insn) == BARRIER)
1041 copy_end = PREV_INSN (PREV_INSN (last_loop_insn));
1042 else if (GET_CODE (last_loop_insn) == JUMP_INSN)
1044 copy_end = PREV_INSN (last_loop_insn);
1046 /* The immediately preceding insn may be a compare which we do not
1048 if (sets_cc0_p (PREV_INSN (copy_end)))
1049 copy_end = PREV_INSN (copy_end);
1055 for (i = 1; i < unroll_number; i++)
1057 emit_label_after (labels[unroll_number - i],
1058 PREV_INSN (loop_start));
1060 bzero ((char *) map->insn_map, max_insnno * sizeof (rtx));
1061 bzero ((char *) &VARRAY_CONST_EQUIV (map->const_equiv_varray, 0),
1062 (VARRAY_SIZE (map->const_equiv_varray)
1063 * sizeof (struct const_equiv_data)));
1066 for (j = 0; j < max_labelno; j++)
1068 set_label_in_map (map, j, gen_label_rtx ());
1070 for (j = FIRST_PSEUDO_REGISTER; j < max_local_regnum; j++)
1073 map->reg_map[j] = gen_reg_rtx (GET_MODE (regno_reg_rtx[j]));
1074 record_base_value (REGNO (map->reg_map[j]),
1075 regno_reg_rtx[j], 0);
1077 /* The last copy needs the compare/branch insns at the end,
1078 so reset copy_end here if the loop ends with a conditional
1081 if (i == unroll_number - 1)
1083 if (GET_CODE (last_loop_insn) == BARRIER)
1084 copy_end = PREV_INSN (PREV_INSN (last_loop_insn));
1086 copy_end = last_loop_insn;
1089 /* None of the copies are the `last_iteration', so just
1090 pass zero for that parameter. */
1091 copy_loop_body (copy_start, copy_end, map, exit_label, 0,
1092 unroll_type, start_label, loop_end,
1093 loop_start, copy_end);
1095 emit_label_after (labels[0], PREV_INSN (loop_start));
1097 if (GET_CODE (last_loop_insn) == BARRIER)
1099 insert_before = PREV_INSN (last_loop_insn);
1100 copy_end = PREV_INSN (insert_before);
1104 insert_before = last_loop_insn;
1106 /* The instruction immediately before the JUMP_INSN may be a compare
1107 instruction which we do not want to copy or delete. */
1108 if (sets_cc0_p (PREV_INSN (insert_before)))
1109 insert_before = PREV_INSN (insert_before);
1111 copy_end = PREV_INSN (insert_before);
1114 /* Set unroll type to MODULO now. */
1115 unroll_type = UNROLL_MODULO;
1116 loop_preconditioned = 1;
1123 /* If reach here, and the loop type is UNROLL_NAIVE, then don't unroll
1124 the loop unless all loops are being unrolled. */
1125 if (unroll_type == UNROLL_NAIVE && ! flag_unroll_all_loops)
1127 if (loop_dump_stream)
1128 fprintf (loop_dump_stream, "Unrolling failure: Naive unrolling not being done.\n");
1132 /* At this point, we are guaranteed to unroll the loop. */
1134 /* Keep track of the unroll factor for the loop. */
1135 loop_info->unroll_number = unroll_number;
1137 /* For each biv and giv, determine whether it can be safely split into
1138 a different variable for each unrolled copy of the loop body.
1139 We precalculate and save this info here, since computing it is
1142 Do this before deleting any instructions from the loop, so that
1143 back_branch_in_range_p will work correctly. */
1145 if (splitting_not_safe)
1148 temp = find_splittable_regs (unroll_type, loop_start, loop_end,
1149 end_insert_before, unroll_number,
1150 loop_info->n_iterations);
1152 /* find_splittable_regs may have created some new registers, so must
1153 reallocate the reg_map with the new larger size, and must realloc
1154 the constant maps also. */
1156 maxregnum = max_reg_num ();
1157 map->reg_map = (rtx *) xmalloc (maxregnum * sizeof (rtx));
1159 init_reg_map (map, maxregnum);
1161 if (map->const_equiv_varray == 0)
1162 VARRAY_CONST_EQUIV_INIT (map->const_equiv_varray,
1163 maxregnum + temp * unroll_number * 2,
1165 global_const_equiv_varray = map->const_equiv_varray;
1167 /* Search the list of bivs and givs to find ones which need to be remapped
1168 when split, and set their reg_map entry appropriately. */
1170 for (bl = loop_iv_list; bl; bl = bl->next)
1172 if (REGNO (bl->biv->src_reg) != bl->regno)
1173 map->reg_map[bl->regno] = bl->biv->src_reg;
1175 /* Currently, non-reduced/final-value givs are never split. */
1176 for (v = bl->giv; v; v = v->next_iv)
1177 if (REGNO (v->src_reg) != bl->regno)
1178 map->reg_map[REGNO (v->dest_reg)] = v->src_reg;
1182 /* Use our current register alignment and pointer flags. */
1183 map->regno_pointer_flag = cfun->emit->regno_pointer_flag;
1184 map->regno_pointer_align = cfun->emit->regno_pointer_align;
1186 /* If the loop is being partially unrolled, and the iteration variables
1187 are being split, and are being renamed for the split, then must fix up
1188 the compare/jump instruction at the end of the loop to refer to the new
1189 registers. This compare isn't copied, so the registers used in it
1190 will never be replaced if it isn't done here. */
1192 if (unroll_type == UNROLL_MODULO)
1194 insn = NEXT_INSN (copy_end);
1195 if (GET_CODE (insn) == INSN || GET_CODE (insn) == JUMP_INSN)
1196 PATTERN (insn) = remap_split_bivs (PATTERN (insn));
1199 /* For unroll_number times, make a copy of each instruction
1200 between copy_start and copy_end, and insert these new instructions
1201 before the end of the loop. */
1203 for (i = 0; i < unroll_number; i++)
1205 bzero ((char *) map->insn_map, max_insnno * sizeof (rtx));
1206 bzero ((char *) &VARRAY_CONST_EQUIV (map->const_equiv_varray, 0),
1207 VARRAY_SIZE (map->const_equiv_varray) * sizeof (struct const_equiv_data));
1210 for (j = 0; j < max_labelno; j++)
1212 set_label_in_map (map, j, gen_label_rtx ());
1214 for (j = FIRST_PSEUDO_REGISTER; j < max_local_regnum; j++)
1217 map->reg_map[j] = gen_reg_rtx (GET_MODE (regno_reg_rtx[j]));
1218 record_base_value (REGNO (map->reg_map[j]),
1219 regno_reg_rtx[j], 0);
1222 /* If loop starts with a branch to the test, then fix it so that
1223 it points to the test of the first unrolled copy of the loop. */
1224 if (i == 0 && loop_start != copy_start)
1226 insn = PREV_INSN (copy_start);
1227 pattern = PATTERN (insn);
1229 tem = get_label_from_map (map,
1231 (XEXP (SET_SRC (pattern), 0)));
1232 SET_SRC (pattern) = gen_rtx_LABEL_REF (VOIDmode, tem);
1234 /* Set the jump label so that it can be used by later loop unrolling
1236 JUMP_LABEL (insn) = tem;
1237 LABEL_NUSES (tem)++;
1240 copy_loop_body (copy_start, copy_end, map, exit_label,
1241 i == unroll_number - 1, unroll_type, start_label,
1242 loop_end, insert_before, insert_before);
1245 /* Before deleting any insns, emit a CODE_LABEL immediately after the last
1246 insn to be deleted. This prevents any runaway delete_insn call from
1247 more insns that it should, as it always stops at a CODE_LABEL. */
1249 /* Delete the compare and branch at the end of the loop if completely
1250 unrolling the loop. Deleting the backward branch at the end also
1251 deletes the code label at the start of the loop. This is done at
1252 the very end to avoid problems with back_branch_in_range_p. */
1254 if (unroll_type == UNROLL_COMPLETELY)
1255 safety_label = emit_label_after (gen_label_rtx (), last_loop_insn);
1257 safety_label = emit_label_after (gen_label_rtx (), copy_end);
1259 /* Delete all of the original loop instructions. Don't delete the
1260 LOOP_BEG note, or the first code label in the loop. */
1262 insn = NEXT_INSN (copy_start);
1263 while (insn != safety_label)
1265 /* ??? Don't delete named code labels. They will be deleted when the
1266 jump that references them is deleted. Otherwise, we end up deleting
1267 them twice, which causes them to completely disappear instead of turn
1268 into NOTE_INSN_DELETED_LABEL notes. This in turn causes aborts in
1269 dwarfout.c/dwarf2out.c. We could perhaps fix the dwarf*out.c files
1270 to handle deleted labels instead. Or perhaps fix DECL_RTL of the
1271 associated LABEL_DECL to point to one of the new label instances. */
1272 /* ??? Likewise, we can't delete a NOTE_INSN_DELETED_LABEL note. */
1273 if (insn != start_label
1274 && ! (GET_CODE (insn) == CODE_LABEL && LABEL_NAME (insn))
1275 && ! (GET_CODE (insn) == NOTE
1276 && NOTE_LINE_NUMBER (insn) == NOTE_INSN_DELETED_LABEL))
1277 insn = delete_insn (insn);
1279 insn = NEXT_INSN (insn);
1282 /* Can now delete the 'safety' label emitted to protect us from runaway
1283 delete_insn calls. */
1284 if (INSN_DELETED_P (safety_label))
1286 delete_insn (safety_label);
1288 /* If exit_label exists, emit it after the loop. Doing the emit here
1289 forces it to have a higher INSN_UID than any insn in the unrolled loop.
1290 This is needed so that mostly_true_jump in reorg.c will treat jumps
1291 to this loop end label correctly, i.e. predict that they are usually
1294 emit_label_after (exit_label, loop_end);
1297 if (unroll_type == UNROLL_COMPLETELY)
1299 /* Remove the loop notes since this is no longer a loop. */
1300 if (loop_info->vtop)
1301 delete_insn (loop_info->vtop);
1302 if (loop_info->cont)
1303 delete_insn (loop_info->cont);
1305 delete_insn (loop_start);
1307 delete_insn (loop_end);
1310 if (map->const_equiv_varray)
1311 VARRAY_FREE (map->const_equiv_varray);
1314 free (map->label_map);
1317 free (map->insn_map);
1318 free (splittable_regs);
1319 free (derived_regs);
1320 free (splittable_regs_updates);
1321 free (addr_combined_regs);
1324 free (map->reg_map);
1328 /* Return true if the loop can be safely, and profitably, preconditioned
1329 so that the unrolled copies of the loop body don't need exit tests.
1331 This only works if final_value, initial_value and increment can be
1332 determined, and if increment is a constant power of 2.
1333 If increment is not a power of 2, then the preconditioning modulo
1334 operation would require a real modulo instead of a boolean AND, and this
1335 is not considered `profitable'. */
1337 /* ??? If the loop is known to be executed very many times, or the machine
1338 has a very cheap divide instruction, then preconditioning is a win even
1339 when the increment is not a power of 2. Use RTX_COST to compute
1340 whether divide is cheap.
1341 ??? A divide by constant doesn't actually need a divide, look at
1342 expand_divmod. The reduced cost of this optimized modulo is not
1343 reflected in RTX_COST. */
1346 precondition_loop_p (loop_start, loop_info,
1347 initial_value, final_value, increment, mode)
1349 struct loop_info *loop_info;
1350 rtx *initial_value, *final_value, *increment;
1351 enum machine_mode *mode;
1354 if (loop_info->n_iterations > 0)
1356 *initial_value = const0_rtx;
1357 *increment = const1_rtx;
1358 *final_value = GEN_INT (loop_info->n_iterations);
1361 if (loop_dump_stream)
1363 fputs ("Preconditioning: Success, number of iterations known, ",
1365 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC,
1366 loop_info->n_iterations);
1367 fputs (".\n", loop_dump_stream);
1372 if (loop_info->initial_value == 0)
1374 if (loop_dump_stream)
1375 fprintf (loop_dump_stream,
1376 "Preconditioning: Could not find initial value.\n");
1379 else if (loop_info->increment == 0)
1381 if (loop_dump_stream)
1382 fprintf (loop_dump_stream,
1383 "Preconditioning: Could not find increment value.\n");
1386 else if (GET_CODE (loop_info->increment) != CONST_INT)
1388 if (loop_dump_stream)
1389 fprintf (loop_dump_stream,
1390 "Preconditioning: Increment not a constant.\n");
1393 else if ((exact_log2 (INTVAL (loop_info->increment)) < 0)
1394 && (exact_log2 (- INTVAL (loop_info->increment)) < 0))
1396 if (loop_dump_stream)
1397 fprintf (loop_dump_stream,
1398 "Preconditioning: Increment not a constant power of 2.\n");
1402 /* Unsigned_compare and compare_dir can be ignored here, since they do
1403 not matter for preconditioning. */
1405 if (loop_info->final_value == 0)
1407 if (loop_dump_stream)
1408 fprintf (loop_dump_stream,
1409 "Preconditioning: EQ comparison loop.\n");
1413 /* Must ensure that final_value is invariant, so call invariant_p to
1414 check. Before doing so, must check regno against max_reg_before_loop
1415 to make sure that the register is in the range covered by invariant_p.
1416 If it isn't, then it is most likely a biv/giv which by definition are
1418 if ((GET_CODE (loop_info->final_value) == REG
1419 && REGNO (loop_info->final_value) >= max_reg_before_loop)
1420 || (GET_CODE (loop_info->final_value) == PLUS
1421 && REGNO (XEXP (loop_info->final_value, 0)) >= max_reg_before_loop)
1422 || ! invariant_p (loop_info->final_value))
1424 if (loop_dump_stream)
1425 fprintf (loop_dump_stream,
1426 "Preconditioning: Final value not invariant.\n");
1430 /* Fail for floating point values, since the caller of this function
1431 does not have code to deal with them. */
1432 if (GET_MODE_CLASS (GET_MODE (loop_info->final_value)) == MODE_FLOAT
1433 || GET_MODE_CLASS (GET_MODE (loop_info->initial_value)) == MODE_FLOAT)
1435 if (loop_dump_stream)
1436 fprintf (loop_dump_stream,
1437 "Preconditioning: Floating point final or initial value.\n");
1441 /* Fail if loop_info->iteration_var is not live before loop_start,
1442 since we need to test its value in the preconditioning code. */
1444 if (uid_luid[REGNO_FIRST_UID (REGNO (loop_info->iteration_var))]
1445 > INSN_LUID (loop_start))
1447 if (loop_dump_stream)
1448 fprintf (loop_dump_stream,
1449 "Preconditioning: Iteration var not live before loop start.\n");
1453 /* Note that iteration_info biases the initial value for GIV iterators
1454 such as "while (i-- > 0)" so that we can calculate the number of
1455 iterations just like for BIV iterators.
1457 Also note that the absolute values of initial_value and
1458 final_value are unimportant as only their difference is used for
1459 calculating the number of loop iterations. */
1460 *initial_value = loop_info->initial_value;
1461 *increment = loop_info->increment;
1462 *final_value = loop_info->final_value;
1464 /* Decide what mode to do these calculations in. Choose the larger
1465 of final_value's mode and initial_value's mode, or a full-word if
1466 both are constants. */
1467 *mode = GET_MODE (*final_value);
1468 if (*mode == VOIDmode)
1470 *mode = GET_MODE (*initial_value);
1471 if (*mode == VOIDmode)
1474 else if (*mode != GET_MODE (*initial_value)
1475 && (GET_MODE_SIZE (*mode)
1476 < GET_MODE_SIZE (GET_MODE (*initial_value))))
1477 *mode = GET_MODE (*initial_value);
1480 if (loop_dump_stream)
1481 fprintf (loop_dump_stream, "Preconditioning: Successful.\n");
1486 /* All pseudo-registers must be mapped to themselves. Two hard registers
1487 must be mapped, VIRTUAL_STACK_VARS_REGNUM and VIRTUAL_INCOMING_ARGS_
1488 REGNUM, to avoid function-inlining specific conversions of these
1489 registers. All other hard regs can not be mapped because they may be
1494 init_reg_map (map, maxregnum)
1495 struct inline_remap *map;
1500 for (i = maxregnum - 1; i > LAST_VIRTUAL_REGISTER; i--)
1501 map->reg_map[i] = regno_reg_rtx[i];
1502 /* Just clear the rest of the entries. */
1503 for (i = LAST_VIRTUAL_REGISTER; i >= 0; i--)
1504 map->reg_map[i] = 0;
1506 map->reg_map[VIRTUAL_STACK_VARS_REGNUM]
1507 = regno_reg_rtx[VIRTUAL_STACK_VARS_REGNUM];
1508 map->reg_map[VIRTUAL_INCOMING_ARGS_REGNUM]
1509 = regno_reg_rtx[VIRTUAL_INCOMING_ARGS_REGNUM];
1512 /* Strength-reduction will often emit code for optimized biv/givs which
1513 calculates their value in a temporary register, and then copies the result
1514 to the iv. This procedure reconstructs the pattern computing the iv;
1515 verifying that all operands are of the proper form.
1517 PATTERN must be the result of single_set.
1518 The return value is the amount that the giv is incremented by. */
1521 calculate_giv_inc (pattern, src_insn, regno)
1522 rtx pattern, src_insn;
1526 rtx increment_total = 0;
1530 /* Verify that we have an increment insn here. First check for a plus
1531 as the set source. */
1532 if (GET_CODE (SET_SRC (pattern)) != PLUS)
1534 /* SR sometimes computes the new giv value in a temp, then copies it
1536 src_insn = PREV_INSN (src_insn);
1537 pattern = PATTERN (src_insn);
1538 if (GET_CODE (SET_SRC (pattern)) != PLUS)
1541 /* The last insn emitted is not needed, so delete it to avoid confusing
1542 the second cse pass. This insn sets the giv unnecessarily. */
1543 delete_insn (get_last_insn ());
1546 /* Verify that we have a constant as the second operand of the plus. */
1547 increment = XEXP (SET_SRC (pattern), 1);
1548 if (GET_CODE (increment) != CONST_INT)
1550 /* SR sometimes puts the constant in a register, especially if it is
1551 too big to be an add immed operand. */
1552 src_insn = PREV_INSN (src_insn);
1553 increment = SET_SRC (PATTERN (src_insn));
1555 /* SR may have used LO_SUM to compute the constant if it is too large
1556 for a load immed operand. In this case, the constant is in operand
1557 one of the LO_SUM rtx. */
1558 if (GET_CODE (increment) == LO_SUM)
1559 increment = XEXP (increment, 1);
1561 /* Some ports store large constants in memory and add a REG_EQUAL
1562 note to the store insn. */
1563 else if (GET_CODE (increment) == MEM)
1565 rtx note = find_reg_note (src_insn, REG_EQUAL, 0);
1567 increment = XEXP (note, 0);
1570 else if (GET_CODE (increment) == IOR
1571 || GET_CODE (increment) == ASHIFT
1572 || GET_CODE (increment) == PLUS)
1574 /* The rs6000 port loads some constants with IOR.
1575 The alpha port loads some constants with ASHIFT and PLUS. */
1576 rtx second_part = XEXP (increment, 1);
1577 enum rtx_code code = GET_CODE (increment);
1579 src_insn = PREV_INSN (src_insn);
1580 increment = SET_SRC (PATTERN (src_insn));
1581 /* Don't need the last insn anymore. */
1582 delete_insn (get_last_insn ());
1584 if (GET_CODE (second_part) != CONST_INT
1585 || GET_CODE (increment) != CONST_INT)
1589 increment = GEN_INT (INTVAL (increment) | INTVAL (second_part));
1590 else if (code == PLUS)
1591 increment = GEN_INT (INTVAL (increment) + INTVAL (second_part));
1593 increment = GEN_INT (INTVAL (increment) << INTVAL (second_part));
1596 if (GET_CODE (increment) != CONST_INT)
1599 /* The insn loading the constant into a register is no longer needed,
1601 delete_insn (get_last_insn ());
1604 if (increment_total)
1605 increment_total = GEN_INT (INTVAL (increment_total) + INTVAL (increment));
1607 increment_total = increment;
1609 /* Check that the source register is the same as the register we expected
1610 to see as the source. If not, something is seriously wrong. */
1611 if (GET_CODE (XEXP (SET_SRC (pattern), 0)) != REG
1612 || REGNO (XEXP (SET_SRC (pattern), 0)) != regno)
1614 /* Some machines (e.g. the romp), may emit two add instructions for
1615 certain constants, so lets try looking for another add immediately
1616 before this one if we have only seen one add insn so far. */
1622 src_insn = PREV_INSN (src_insn);
1623 pattern = PATTERN (src_insn);
1625 delete_insn (get_last_insn ());
1633 return increment_total;
1636 /* Copy REG_NOTES, except for insn references, because not all insn_map
1637 entries are valid yet. We do need to copy registers now though, because
1638 the reg_map entries can change during copying. */
1641 initial_reg_note_copy (notes, map)
1643 struct inline_remap *map;
1650 copy = rtx_alloc (GET_CODE (notes));
1651 PUT_MODE (copy, GET_MODE (notes));
1653 if (GET_CODE (notes) == EXPR_LIST)
1654 XEXP (copy, 0) = copy_rtx_and_substitute (XEXP (notes, 0), map, 0);
1655 else if (GET_CODE (notes) == INSN_LIST)
1656 /* Don't substitute for these yet. */
1657 XEXP (copy, 0) = XEXP (notes, 0);
1661 XEXP (copy, 1) = initial_reg_note_copy (XEXP (notes, 1), map);
1666 /* Fixup insn references in copied REG_NOTES. */
1669 final_reg_note_copy (notes, map)
1671 struct inline_remap *map;
1675 for (note = notes; note; note = XEXP (note, 1))
1676 if (GET_CODE (note) == INSN_LIST)
1677 XEXP (note, 0) = map->insn_map[INSN_UID (XEXP (note, 0))];
1680 /* Copy each instruction in the loop, substituting from map as appropriate.
1681 This is very similar to a loop in expand_inline_function. */
1684 copy_loop_body (copy_start, copy_end, map, exit_label, last_iteration,
1685 unroll_type, start_label, loop_end, insert_before,
1687 rtx copy_start, copy_end;
1688 struct inline_remap *map;
1691 enum unroll_types unroll_type;
1692 rtx start_label, loop_end, insert_before, copy_notes_from;
1696 int dest_reg_was_split, i;
1700 rtx final_label = 0;
1701 rtx giv_inc, giv_dest_reg, giv_src_reg;
1703 /* If this isn't the last iteration, then map any references to the
1704 start_label to final_label. Final label will then be emitted immediately
1705 after the end of this loop body if it was ever used.
1707 If this is the last iteration, then map references to the start_label
1709 if (! last_iteration)
1711 final_label = gen_label_rtx ();
1712 set_label_in_map (map, CODE_LABEL_NUMBER (start_label),
1716 set_label_in_map (map, CODE_LABEL_NUMBER (start_label), start_label);
1720 /* Emit a NOTE_INSN_DELETED to force at least two insns onto the sequence.
1721 Else gen_sequence could return a raw pattern for a jump which we pass
1722 off to emit_insn_before (instead of emit_jump_insn_before) which causes
1723 a variety of losing behaviors later. */
1724 emit_note (0, NOTE_INSN_DELETED);
1729 insn = NEXT_INSN (insn);
1731 map->orig_asm_operands_vector = 0;
1733 switch (GET_CODE (insn))
1736 pattern = PATTERN (insn);
1740 /* Check to see if this is a giv that has been combined with
1741 some split address givs. (Combined in the sense that
1742 `combine_givs' in loop.c has put two givs in the same register.)
1743 In this case, we must search all givs based on the same biv to
1744 find the address givs. Then split the address givs.
1745 Do this before splitting the giv, since that may map the
1746 SET_DEST to a new register. */
1748 if ((set = single_set (insn))
1749 && GET_CODE (SET_DEST (set)) == REG
1750 && addr_combined_regs[REGNO (SET_DEST (set))])
1752 struct iv_class *bl;
1753 struct induction *v, *tv;
1754 int regno = REGNO (SET_DEST (set));
1756 v = addr_combined_regs[REGNO (SET_DEST (set))];
1757 bl = reg_biv_class[REGNO (v->src_reg)];
1759 /* Although the giv_inc amount is not needed here, we must call
1760 calculate_giv_inc here since it might try to delete the
1761 last insn emitted. If we wait until later to call it,
1762 we might accidentally delete insns generated immediately
1763 below by emit_unrolled_add. */
1765 if (! derived_regs[regno])
1766 giv_inc = calculate_giv_inc (set, insn, regno);
1768 /* Now find all address giv's that were combined with this
1770 for (tv = bl->giv; tv; tv = tv->next_iv)
1771 if (tv->giv_type == DEST_ADDR && tv->same == v)
1775 /* If this DEST_ADDR giv was not split, then ignore it. */
1776 if (*tv->location != tv->dest_reg)
1779 /* Scale this_giv_inc if the multiplicative factors of
1780 the two givs are different. */
1781 this_giv_inc = INTVAL (giv_inc);
1782 if (tv->mult_val != v->mult_val)
1783 this_giv_inc = (this_giv_inc / INTVAL (v->mult_val)
1784 * INTVAL (tv->mult_val));
1786 tv->dest_reg = plus_constant (tv->dest_reg, this_giv_inc);
1787 *tv->location = tv->dest_reg;
1789 if (last_iteration && unroll_type != UNROLL_COMPLETELY)
1791 /* Must emit an insn to increment the split address
1792 giv. Add in the const_adjust field in case there
1793 was a constant eliminated from the address. */
1794 rtx value, dest_reg;
1796 /* tv->dest_reg will be either a bare register,
1797 or else a register plus a constant. */
1798 if (GET_CODE (tv->dest_reg) == REG)
1799 dest_reg = tv->dest_reg;
1801 dest_reg = XEXP (tv->dest_reg, 0);
1803 /* Check for shared address givs, and avoid
1804 incrementing the shared pseudo reg more than
1806 if (! tv->same_insn && ! tv->shared)
1808 /* tv->dest_reg may actually be a (PLUS (REG)
1809 (CONST)) here, so we must call plus_constant
1810 to add the const_adjust amount before calling
1811 emit_unrolled_add below. */
1812 value = plus_constant (tv->dest_reg,
1815 if (GET_CODE (value) == PLUS)
1817 /* The constant could be too large for an add
1818 immediate, so can't directly emit an insn
1820 emit_unrolled_add (dest_reg, XEXP (value, 0),
1825 /* Reset the giv to be just the register again, in case
1826 it is used after the set we have just emitted.
1827 We must subtract the const_adjust factor added in
1829 tv->dest_reg = plus_constant (dest_reg,
1830 - tv->const_adjust);
1831 *tv->location = tv->dest_reg;
1836 /* If this is a setting of a splittable variable, then determine
1837 how to split the variable, create a new set based on this split,
1838 and set up the reg_map so that later uses of the variable will
1839 use the new split variable. */
1841 dest_reg_was_split = 0;
1843 if ((set = single_set (insn))
1844 && GET_CODE (SET_DEST (set)) == REG
1845 && splittable_regs[REGNO (SET_DEST (set))])
1847 int regno = REGNO (SET_DEST (set));
1850 dest_reg_was_split = 1;
1852 giv_dest_reg = SET_DEST (set);
1853 if (derived_regs[regno])
1855 /* ??? This relies on SET_SRC (SET) to be of
1856 the form (plus (reg) (const_int)), and thus
1857 forces recombine_givs to restrict the kind
1858 of giv derivations it does before unrolling. */
1859 giv_src_reg = XEXP (SET_SRC (set), 0);
1860 giv_inc = XEXP (SET_SRC (set), 1);
1864 giv_src_reg = giv_dest_reg;
1865 /* Compute the increment value for the giv, if it wasn't
1866 already computed above. */
1868 giv_inc = calculate_giv_inc (set, insn, regno);
1870 src_regno = REGNO (giv_src_reg);
1872 if (unroll_type == UNROLL_COMPLETELY)
1874 /* Completely unrolling the loop. Set the induction
1875 variable to a known constant value. */
1877 /* The value in splittable_regs may be an invariant
1878 value, so we must use plus_constant here. */
1879 splittable_regs[regno]
1880 = plus_constant (splittable_regs[src_regno],
1883 if (GET_CODE (splittable_regs[regno]) == PLUS)
1885 giv_src_reg = XEXP (splittable_regs[regno], 0);
1886 giv_inc = XEXP (splittable_regs[regno], 1);
1890 /* The splittable_regs value must be a REG or a
1891 CONST_INT, so put the entire value in the giv_src_reg
1893 giv_src_reg = splittable_regs[regno];
1894 giv_inc = const0_rtx;
1899 /* Partially unrolling loop. Create a new pseudo
1900 register for the iteration variable, and set it to
1901 be a constant plus the original register. Except
1902 on the last iteration, when the result has to
1903 go back into the original iteration var register. */
1905 /* Handle bivs which must be mapped to a new register
1906 when split. This happens for bivs which need their
1907 final value set before loop entry. The new register
1908 for the biv was stored in the biv's first struct
1909 induction entry by find_splittable_regs. */
1911 if (regno < max_reg_before_loop
1912 && REG_IV_TYPE (regno) == BASIC_INDUCT)
1914 giv_src_reg = reg_biv_class[regno]->biv->src_reg;
1915 giv_dest_reg = giv_src_reg;
1919 /* If non-reduced/final-value givs were split, then
1920 this would have to remap those givs also. See
1921 find_splittable_regs. */
1924 splittable_regs[regno]
1925 = GEN_INT (INTVAL (giv_inc)
1926 + INTVAL (splittable_regs[src_regno]));
1927 giv_inc = splittable_regs[regno];
1929 /* Now split the induction variable by changing the dest
1930 of this insn to a new register, and setting its
1931 reg_map entry to point to this new register.
1933 If this is the last iteration, and this is the last insn
1934 that will update the iv, then reuse the original dest,
1935 to ensure that the iv will have the proper value when
1936 the loop exits or repeats.
1938 Using splittable_regs_updates here like this is safe,
1939 because it can only be greater than one if all
1940 instructions modifying the iv are always executed in
1943 if (! last_iteration
1944 || (splittable_regs_updates[regno]-- != 1))
1946 tem = gen_reg_rtx (GET_MODE (giv_src_reg));
1948 map->reg_map[regno] = tem;
1949 record_base_value (REGNO (tem),
1950 giv_inc == const0_rtx
1952 : gen_rtx_PLUS (GET_MODE (giv_src_reg),
1953 giv_src_reg, giv_inc),
1957 map->reg_map[regno] = giv_src_reg;
1960 /* The constant being added could be too large for an add
1961 immediate, so can't directly emit an insn here. */
1962 emit_unrolled_add (giv_dest_reg, giv_src_reg, giv_inc);
1963 copy = get_last_insn ();
1964 pattern = PATTERN (copy);
1968 pattern = copy_rtx_and_substitute (pattern, map, 0);
1969 copy = emit_insn (pattern);
1971 REG_NOTES (copy) = initial_reg_note_copy (REG_NOTES (insn), map);
1974 /* If this insn is setting CC0, it may need to look at
1975 the insn that uses CC0 to see what type of insn it is.
1976 In that case, the call to recog via validate_change will
1977 fail. So don't substitute constants here. Instead,
1978 do it when we emit the following insn.
1980 For example, see the pyr.md file. That machine has signed and
1981 unsigned compares. The compare patterns must check the
1982 following branch insn to see which what kind of compare to
1985 If the previous insn set CC0, substitute constants on it as
1987 if (sets_cc0_p (PATTERN (copy)) != 0)
1992 try_constants (cc0_insn, map);
1994 try_constants (copy, map);
1997 try_constants (copy, map);
2000 /* Make split induction variable constants `permanent' since we
2001 know there are no backward branches across iteration variable
2002 settings which would invalidate this. */
2003 if (dest_reg_was_split)
2005 int regno = REGNO (SET_DEST (set));
2007 if ((size_t) regno < VARRAY_SIZE (map->const_equiv_varray)
2008 && (VARRAY_CONST_EQUIV (map->const_equiv_varray, regno).age
2010 VARRAY_CONST_EQUIV (map->const_equiv_varray, regno).age = -1;
2015 pattern = copy_rtx_and_substitute (PATTERN (insn), map, 0);
2016 copy = emit_jump_insn (pattern);
2017 REG_NOTES (copy) = initial_reg_note_copy (REG_NOTES (insn), map);
2019 if (JUMP_LABEL (insn) == start_label && insn == copy_end
2020 && ! last_iteration)
2022 /* This is a branch to the beginning of the loop; this is the
2023 last insn being copied; and this is not the last iteration.
2024 In this case, we want to change the original fall through
2025 case to be a branch past the end of the loop, and the
2026 original jump label case to fall_through. */
2028 if (invert_exp (pattern, copy))
2030 if (! redirect_exp (&pattern,
2031 get_label_from_map (map,
2033 (JUMP_LABEL (insn))),
2040 rtx lab = gen_label_rtx ();
2041 /* Can't do it by reversing the jump (probably because we
2042 couldn't reverse the conditions), so emit a new
2043 jump_insn after COPY, and redirect the jump around
2045 jmp = emit_jump_insn_after (gen_jump (exit_label), copy);
2046 jmp = emit_barrier_after (jmp);
2047 emit_label_after (lab, jmp);
2048 LABEL_NUSES (lab) = 0;
2049 if (! redirect_exp (&pattern,
2050 get_label_from_map (map,
2052 (JUMP_LABEL (insn))),
2060 try_constants (cc0_insn, map);
2063 try_constants (copy, map);
2065 /* Set the jump label of COPY correctly to avoid problems with
2066 later passes of unroll_loop, if INSN had jump label set. */
2067 if (JUMP_LABEL (insn))
2071 /* Can't use the label_map for every insn, since this may be
2072 the backward branch, and hence the label was not mapped. */
2073 if ((set = single_set (copy)))
2075 tem = SET_SRC (set);
2076 if (GET_CODE (tem) == LABEL_REF)
2077 label = XEXP (tem, 0);
2078 else if (GET_CODE (tem) == IF_THEN_ELSE)
2080 if (XEXP (tem, 1) != pc_rtx)
2081 label = XEXP (XEXP (tem, 1), 0);
2083 label = XEXP (XEXP (tem, 2), 0);
2087 if (label && GET_CODE (label) == CODE_LABEL)
2088 JUMP_LABEL (copy) = label;
2091 /* An unrecognizable jump insn, probably the entry jump
2092 for a switch statement. This label must have been mapped,
2093 so just use the label_map to get the new jump label. */
2095 = get_label_from_map (map,
2096 CODE_LABEL_NUMBER (JUMP_LABEL (insn)));
2099 /* If this is a non-local jump, then must increase the label
2100 use count so that the label will not be deleted when the
2101 original jump is deleted. */
2102 LABEL_NUSES (JUMP_LABEL (copy))++;
2104 else if (GET_CODE (PATTERN (copy)) == ADDR_VEC
2105 || GET_CODE (PATTERN (copy)) == ADDR_DIFF_VEC)
2107 rtx pat = PATTERN (copy);
2108 int diff_vec_p = GET_CODE (pat) == ADDR_DIFF_VEC;
2109 int len = XVECLEN (pat, diff_vec_p);
2112 for (i = 0; i < len; i++)
2113 LABEL_NUSES (XEXP (XVECEXP (pat, diff_vec_p, i), 0))++;
2116 /* If this used to be a conditional jump insn but whose branch
2117 direction is now known, we must do something special. */
2118 if (condjump_p (insn) && !simplejump_p (insn) && map->last_pc_value)
2121 /* If the previous insn set cc0 for us, delete it. */
2122 if (sets_cc0_p (PREV_INSN (copy)))
2123 delete_insn (PREV_INSN (copy));
2126 /* If this is now a no-op, delete it. */
2127 if (map->last_pc_value == pc_rtx)
2129 /* Don't let delete_insn delete the label referenced here,
2130 because we might possibly need it later for some other
2131 instruction in the loop. */
2132 if (JUMP_LABEL (copy))
2133 LABEL_NUSES (JUMP_LABEL (copy))++;
2135 if (JUMP_LABEL (copy))
2136 LABEL_NUSES (JUMP_LABEL (copy))--;
2140 /* Otherwise, this is unconditional jump so we must put a
2141 BARRIER after it. We could do some dead code elimination
2142 here, but jump.c will do it just as well. */
2148 pattern = copy_rtx_and_substitute (PATTERN (insn), map, 0);
2149 copy = emit_call_insn (pattern);
2150 REG_NOTES (copy) = initial_reg_note_copy (REG_NOTES (insn), map);
2152 /* Because the USAGE information potentially contains objects other
2153 than hard registers, we need to copy it. */
2154 CALL_INSN_FUNCTION_USAGE (copy)
2155 = copy_rtx_and_substitute (CALL_INSN_FUNCTION_USAGE (insn),
2160 try_constants (cc0_insn, map);
2163 try_constants (copy, map);
2165 /* Be lazy and assume CALL_INSNs clobber all hard registers. */
2166 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
2167 VARRAY_CONST_EQUIV (map->const_equiv_varray, i).rtx = 0;
2171 /* If this is the loop start label, then we don't need to emit a
2172 copy of this label since no one will use it. */
2174 if (insn != start_label)
2176 copy = emit_label (get_label_from_map (map,
2177 CODE_LABEL_NUMBER (insn)));
2183 copy = emit_barrier ();
2187 /* VTOP and CONT notes are valid only before the loop exit test.
2188 If placed anywhere else, loop may generate bad code. */
2189 /* BASIC_BLOCK notes exist to stabilize basic block structures with
2190 the associated rtl. We do not want to share the structure in
2193 if (NOTE_LINE_NUMBER (insn) != NOTE_INSN_DELETED
2194 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_BASIC_BLOCK
2195 && ((NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_VTOP
2196 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_CONT)
2197 || (last_iteration && unroll_type != UNROLL_COMPLETELY)))
2198 copy = emit_note (NOTE_SOURCE_FILE (insn),
2199 NOTE_LINE_NUMBER (insn));
2208 map->insn_map[INSN_UID (insn)] = copy;
2210 while (insn != copy_end);
2212 /* Now finish coping the REG_NOTES. */
2216 insn = NEXT_INSN (insn);
2217 if ((GET_CODE (insn) == INSN || GET_CODE (insn) == JUMP_INSN
2218 || GET_CODE (insn) == CALL_INSN)
2219 && map->insn_map[INSN_UID (insn)])
2220 final_reg_note_copy (REG_NOTES (map->insn_map[INSN_UID (insn)]), map);
2222 while (insn != copy_end);
2224 /* There may be notes between copy_notes_from and loop_end. Emit a copy of
2225 each of these notes here, since there may be some important ones, such as
2226 NOTE_INSN_BLOCK_END notes, in this group. We don't do this on the last
2227 iteration, because the original notes won't be deleted.
2229 We can't use insert_before here, because when from preconditioning,
2230 insert_before points before the loop. We can't use copy_end, because
2231 there may be insns already inserted after it (which we don't want to
2232 copy) when not from preconditioning code. */
2234 if (! last_iteration)
2236 for (insn = copy_notes_from; insn != loop_end; insn = NEXT_INSN (insn))
2238 /* VTOP notes are valid only before the loop exit test.
2239 If placed anywhere else, loop may generate bad code.
2240 There is no need to test for NOTE_INSN_LOOP_CONT notes
2241 here, since COPY_NOTES_FROM will be at most one or two (for cc0)
2242 instructions before the last insn in the loop, and if the
2243 end test is that short, there will be a VTOP note between
2244 the CONT note and the test. */
2245 if (GET_CODE (insn) == NOTE
2246 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_DELETED
2247 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_BASIC_BLOCK
2248 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_VTOP)
2249 emit_note (NOTE_SOURCE_FILE (insn), NOTE_LINE_NUMBER (insn));
2253 if (final_label && LABEL_NUSES (final_label) > 0)
2254 emit_label (final_label);
2256 tem = gen_sequence ();
2258 emit_insn_before (tem, insert_before);
2261 /* Emit an insn, using the expand_binop to ensure that a valid insn is
2262 emitted. This will correctly handle the case where the increment value
2263 won't fit in the immediate field of a PLUS insns. */
2266 emit_unrolled_add (dest_reg, src_reg, increment)
2267 rtx dest_reg, src_reg, increment;
2271 result = expand_binop (GET_MODE (dest_reg), add_optab, src_reg, increment,
2272 dest_reg, 0, OPTAB_LIB_WIDEN);
2274 if (dest_reg != result)
2275 emit_move_insn (dest_reg, result);
2278 /* Searches the insns between INSN and LOOP_END. Returns 1 if there
2279 is a backward branch in that range that branches to somewhere between
2280 LOOP_START and INSN. Returns 0 otherwise. */
2282 /* ??? This is quadratic algorithm. Could be rewritten to be linear.
2283 In practice, this is not a problem, because this function is seldom called,
2284 and uses a negligible amount of CPU time on average. */
2287 back_branch_in_range_p (insn, loop_start, loop_end)
2289 rtx loop_start, loop_end;
2291 rtx p, q, target_insn;
2292 rtx orig_loop_end = loop_end;
2294 /* Stop before we get to the backward branch at the end of the loop. */
2295 loop_end = prev_nonnote_insn (loop_end);
2296 if (GET_CODE (loop_end) == BARRIER)
2297 loop_end = PREV_INSN (loop_end);
2299 /* Check in case insn has been deleted, search forward for first non
2300 deleted insn following it. */
2301 while (INSN_DELETED_P (insn))
2302 insn = NEXT_INSN (insn);
2304 /* Check for the case where insn is the last insn in the loop. Deal
2305 with the case where INSN was a deleted loop test insn, in which case
2306 it will now be the NOTE_LOOP_END. */
2307 if (insn == loop_end || insn == orig_loop_end)
2310 for (p = NEXT_INSN (insn); p != loop_end; p = NEXT_INSN (p))
2312 if (GET_CODE (p) == JUMP_INSN)
2314 target_insn = JUMP_LABEL (p);
2316 /* Search from loop_start to insn, to see if one of them is
2317 the target_insn. We can't use INSN_LUID comparisons here,
2318 since insn may not have an LUID entry. */
2319 for (q = loop_start; q != insn; q = NEXT_INSN (q))
2320 if (q == target_insn)
2328 /* Try to generate the simplest rtx for the expression
2329 (PLUS (MULT mult1 mult2) add1). This is used to calculate the initial
2333 fold_rtx_mult_add (mult1, mult2, add1, mode)
2334 rtx mult1, mult2, add1;
2335 enum machine_mode mode;
2340 /* The modes must all be the same. This should always be true. For now,
2341 check to make sure. */
2342 if ((GET_MODE (mult1) != mode && GET_MODE (mult1) != VOIDmode)
2343 || (GET_MODE (mult2) != mode && GET_MODE (mult2) != VOIDmode)
2344 || (GET_MODE (add1) != mode && GET_MODE (add1) != VOIDmode))
2347 /* Ensure that if at least one of mult1/mult2 are constant, then mult2
2348 will be a constant. */
2349 if (GET_CODE (mult1) == CONST_INT)
2356 mult_res = simplify_binary_operation (MULT, mode, mult1, mult2);
2358 mult_res = gen_rtx_MULT (mode, mult1, mult2);
2360 /* Again, put the constant second. */
2361 if (GET_CODE (add1) == CONST_INT)
2368 result = simplify_binary_operation (PLUS, mode, add1, mult_res);
2370 result = gen_rtx_PLUS (mode, add1, mult_res);
2375 /* Searches the list of induction struct's for the biv BL, to try to calculate
2376 the total increment value for one iteration of the loop as a constant.
2378 Returns the increment value as an rtx, simplified as much as possible,
2379 if it can be calculated. Otherwise, returns 0. */
2382 biv_total_increment (bl, loop_start, loop_end)
2383 struct iv_class *bl;
2384 rtx loop_start, loop_end;
2386 struct induction *v;
2389 /* For increment, must check every instruction that sets it. Each
2390 instruction must be executed only once each time through the loop.
2391 To verify this, we check that the insn is always executed, and that
2392 there are no backward branches after the insn that branch to before it.
2393 Also, the insn must have a mult_val of one (to make sure it really is
2396 result = const0_rtx;
2397 for (v = bl->biv; v; v = v->next_iv)
2399 if (v->always_computable && v->mult_val == const1_rtx
2400 && ! v->maybe_multiple)
2401 result = fold_rtx_mult_add (result, const1_rtx, v->add_val, v->mode);
2409 /* Determine the initial value of the iteration variable, and the amount
2410 that it is incremented each loop. Use the tables constructed by
2411 the strength reduction pass to calculate these values.
2413 Initial_value and/or increment are set to zero if their values could not
2417 iteration_info (iteration_var, initial_value, increment, loop_start, loop_end)
2418 rtx iteration_var, *initial_value, *increment;
2419 rtx loop_start, loop_end;
2421 struct iv_class *bl;
2423 struct induction *v;
2426 /* Clear the result values, in case no answer can be found. */
2430 /* The iteration variable can be either a giv or a biv. Check to see
2431 which it is, and compute the variable's initial value, and increment
2432 value if possible. */
2434 /* If this is a new register, can't handle it since we don't have any
2435 reg_iv_type entry for it. */
2436 if ((unsigned) REGNO (iteration_var) >= reg_iv_type->num_elements)
2438 if (loop_dump_stream)
2439 fprintf (loop_dump_stream,
2440 "Loop unrolling: No reg_iv_type entry for iteration var.\n");
2444 /* Reject iteration variables larger than the host wide int size, since they
2445 could result in a number of iterations greater than the range of our
2446 `unsigned HOST_WIDE_INT' variable loop_info->n_iterations. */
2447 else if ((GET_MODE_BITSIZE (GET_MODE (iteration_var))
2448 > HOST_BITS_PER_WIDE_INT))
2450 if (loop_dump_stream)
2451 fprintf (loop_dump_stream,
2452 "Loop unrolling: Iteration var rejected because mode too large.\n");
2455 else if (GET_MODE_CLASS (GET_MODE (iteration_var)) != MODE_INT)
2457 if (loop_dump_stream)
2458 fprintf (loop_dump_stream,
2459 "Loop unrolling: Iteration var not an integer.\n");
2462 else if (REG_IV_TYPE (REGNO (iteration_var)) == BASIC_INDUCT)
2464 /* When reg_iv_type / reg_iv_info is resized for biv increments
2465 that are turned into givs, reg_biv_class is not resized.
2466 So check here that we don't make an out-of-bounds access. */
2467 if (REGNO (iteration_var) >= max_reg_before_loop)
2470 /* Grab initial value, only useful if it is a constant. */
2471 bl = reg_biv_class[REGNO (iteration_var)];
2472 *initial_value = bl->initial_value;
2474 *increment = biv_total_increment (bl, loop_start, loop_end);
2476 else if (REG_IV_TYPE (REGNO (iteration_var)) == GENERAL_INDUCT)
2478 HOST_WIDE_INT offset = 0;
2479 struct induction *v = REG_IV_INFO (REGNO (iteration_var));
2481 if (REGNO (v->src_reg) >= max_reg_before_loop)
2484 bl = reg_biv_class[REGNO (v->src_reg)];
2486 /* Increment value is mult_val times the increment value of the biv. */
2488 *increment = biv_total_increment (bl, loop_start, loop_end);
2491 struct induction *biv_inc;
2494 = fold_rtx_mult_add (v->mult_val, *increment, const0_rtx, v->mode);
2495 /* The caller assumes that one full increment has occured at the
2496 first loop test. But that's not true when the biv is incremented
2497 after the giv is set (which is the usual case), e.g.:
2498 i = 6; do {;} while (i++ < 9) .
2499 Therefore, we bias the initial value by subtracting the amount of
2500 the increment that occurs between the giv set and the giv test. */
2501 for (biv_inc = bl->biv; biv_inc; biv_inc = biv_inc->next_iv)
2503 if (loop_insn_first_p (v->insn, biv_inc->insn))
2504 offset -= INTVAL (biv_inc->add_val);
2506 offset *= INTVAL (v->mult_val);
2508 if (loop_dump_stream)
2509 fprintf (loop_dump_stream,
2510 "Loop unrolling: Giv iterator, initial value bias %ld.\n",
2512 /* Initial value is mult_val times the biv's initial value plus
2513 add_val. Only useful if it is a constant. */
2515 = fold_rtx_mult_add (v->mult_val,
2516 plus_constant (bl->initial_value, offset),
2517 v->add_val, v->mode);
2521 if (loop_dump_stream)
2522 fprintf (loop_dump_stream,
2523 "Loop unrolling: Not basic or general induction var.\n");
2529 /* For each biv and giv, determine whether it can be safely split into
2530 a different variable for each unrolled copy of the loop body. If it
2531 is safe to split, then indicate that by saving some useful info
2532 in the splittable_regs array.
2534 If the loop is being completely unrolled, then splittable_regs will hold
2535 the current value of the induction variable while the loop is unrolled.
2536 It must be set to the initial value of the induction variable here.
2537 Otherwise, splittable_regs will hold the difference between the current
2538 value of the induction variable and the value the induction variable had
2539 at the top of the loop. It must be set to the value 0 here.
2541 Returns the total number of instructions that set registers that are
2544 /* ?? If the loop is only unrolled twice, then most of the restrictions to
2545 constant values are unnecessary, since we can easily calculate increment
2546 values in this case even if nothing is constant. The increment value
2547 should not involve a multiply however. */
2549 /* ?? Even if the biv/giv increment values aren't constant, it may still
2550 be beneficial to split the variable if the loop is only unrolled a few
2551 times, since multiplies by small integers (1,2,3,4) are very cheap. */
2554 find_splittable_regs (unroll_type, loop_start, loop_end, end_insert_before,
2555 unroll_number, n_iterations)
2556 enum unroll_types unroll_type;
2557 rtx loop_start, loop_end;
2558 rtx end_insert_before;
2560 unsigned HOST_WIDE_INT n_iterations;
2562 struct iv_class *bl;
2563 struct induction *v;
2565 rtx biv_final_value;
2569 for (bl = loop_iv_list; bl; bl = bl->next)
2571 /* Biv_total_increment must return a constant value,
2572 otherwise we can not calculate the split values. */
2574 increment = biv_total_increment (bl, loop_start, loop_end);
2575 if (! increment || GET_CODE (increment) != CONST_INT)
2578 /* The loop must be unrolled completely, or else have a known number
2579 of iterations and only one exit, or else the biv must be dead
2580 outside the loop, or else the final value must be known. Otherwise,
2581 it is unsafe to split the biv since it may not have the proper
2582 value on loop exit. */
2584 /* loop_number_exit_count is non-zero if the loop has an exit other than
2585 a fall through at the end. */
2588 biv_final_value = 0;
2589 if (unroll_type != UNROLL_COMPLETELY
2590 && (loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]]
2591 || unroll_type == UNROLL_NAIVE)
2592 && (uid_luid[REGNO_LAST_UID (bl->regno)] >= INSN_LUID (loop_end)
2594 || INSN_UID (bl->init_insn) >= max_uid_for_loop
2595 || (uid_luid[REGNO_FIRST_UID (bl->regno)]
2596 < INSN_LUID (bl->init_insn))
2597 || reg_mentioned_p (bl->biv->dest_reg, SET_SRC (bl->init_set)))
2598 && ! (biv_final_value = final_biv_value (bl, loop_start, loop_end,
2602 /* If any of the insns setting the BIV don't do so with a simple
2603 PLUS, we don't know how to split it. */
2604 for (v = bl->biv; biv_splittable && v; v = v->next_iv)
2605 if ((tem = single_set (v->insn)) == 0
2606 || GET_CODE (SET_DEST (tem)) != REG
2607 || REGNO (SET_DEST (tem)) != bl->regno
2608 || GET_CODE (SET_SRC (tem)) != PLUS)
2611 /* If final value is non-zero, then must emit an instruction which sets
2612 the value of the biv to the proper value. This is done after
2613 handling all of the givs, since some of them may need to use the
2614 biv's value in their initialization code. */
2616 /* This biv is splittable. If completely unrolling the loop, save
2617 the biv's initial value. Otherwise, save the constant zero. */
2619 if (biv_splittable == 1)
2621 if (unroll_type == UNROLL_COMPLETELY)
2623 /* If the initial value of the biv is itself (i.e. it is too
2624 complicated for strength_reduce to compute), or is a hard
2625 register, or it isn't invariant, then we must create a new
2626 pseudo reg to hold the initial value of the biv. */
2628 if (GET_CODE (bl->initial_value) == REG
2629 && (REGNO (bl->initial_value) == bl->regno
2630 || REGNO (bl->initial_value) < FIRST_PSEUDO_REGISTER
2631 || ! invariant_p (bl->initial_value)))
2633 rtx tem = gen_reg_rtx (bl->biv->mode);
2635 record_base_value (REGNO (tem), bl->biv->add_val, 0);
2636 emit_insn_before (gen_move_insn (tem, bl->biv->src_reg),
2639 if (loop_dump_stream)
2640 fprintf (loop_dump_stream, "Biv %d initial value remapped to %d.\n",
2641 bl->regno, REGNO (tem));
2643 splittable_regs[bl->regno] = tem;
2646 splittable_regs[bl->regno] = bl->initial_value;
2649 splittable_regs[bl->regno] = const0_rtx;
2651 /* Save the number of instructions that modify the biv, so that
2652 we can treat the last one specially. */
2654 splittable_regs_updates[bl->regno] = bl->biv_count;
2655 result += bl->biv_count;
2657 if (loop_dump_stream)
2658 fprintf (loop_dump_stream,
2659 "Biv %d safe to split.\n", bl->regno);
2662 /* Check every giv that depends on this biv to see whether it is
2663 splittable also. Even if the biv isn't splittable, givs which
2664 depend on it may be splittable if the biv is live outside the
2665 loop, and the givs aren't. */
2667 result += find_splittable_givs (bl, unroll_type, loop_start, loop_end,
2668 increment, unroll_number);
2670 /* If final value is non-zero, then must emit an instruction which sets
2671 the value of the biv to the proper value. This is done after
2672 handling all of the givs, since some of them may need to use the
2673 biv's value in their initialization code. */
2674 if (biv_final_value)
2676 /* If the loop has multiple exits, emit the insns before the
2677 loop to ensure that it will always be executed no matter
2678 how the loop exits. Otherwise emit the insn after the loop,
2679 since this is slightly more efficient. */
2680 if (! loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]])
2681 emit_insn_before (gen_move_insn (bl->biv->src_reg,
2686 /* Create a new register to hold the value of the biv, and then
2687 set the biv to its final value before the loop start. The biv
2688 is set to its final value before loop start to ensure that
2689 this insn will always be executed, no matter how the loop
2691 rtx tem = gen_reg_rtx (bl->biv->mode);
2692 record_base_value (REGNO (tem), bl->biv->add_val, 0);
2694 emit_insn_before (gen_move_insn (tem, bl->biv->src_reg),
2696 emit_insn_before (gen_move_insn (bl->biv->src_reg,
2700 if (loop_dump_stream)
2701 fprintf (loop_dump_stream, "Biv %d mapped to %d for split.\n",
2702 REGNO (bl->biv->src_reg), REGNO (tem));
2704 /* Set up the mapping from the original biv register to the new
2706 bl->biv->src_reg = tem;
2713 /* Return 1 if the first and last unrolled copy of the address giv V is valid
2714 for the instruction that is using it. Do not make any changes to that
2718 verify_addresses (v, giv_inc, unroll_number)
2719 struct induction *v;
2724 rtx orig_addr = *v->location;
2725 rtx last_addr = plus_constant (v->dest_reg,
2726 INTVAL (giv_inc) * (unroll_number - 1));
2728 /* First check to see if either address would fail. Handle the fact
2729 that we have may have a match_dup. */
2730 if (! validate_replace_rtx (*v->location, v->dest_reg, v->insn)
2731 || ! validate_replace_rtx (*v->location, last_addr, v->insn))
2734 /* Now put things back the way they were before. This should always
2736 if (! validate_replace_rtx (*v->location, orig_addr, v->insn))
2742 /* For every giv based on the biv BL, check to determine whether it is
2743 splittable. This is a subroutine to find_splittable_regs ().
2745 Return the number of instructions that set splittable registers. */
2748 find_splittable_givs (bl, unroll_type, loop_start, loop_end, increment,
2750 struct iv_class *bl;
2751 enum unroll_types unroll_type;
2752 rtx loop_start, loop_end;
2756 struct induction *v, *v2;
2761 /* Scan the list of givs, and set the same_insn field when there are
2762 multiple identical givs in the same insn. */
2763 for (v = bl->giv; v; v = v->next_iv)
2764 for (v2 = v->next_iv; v2; v2 = v2->next_iv)
2765 if (v->insn == v2->insn && rtx_equal_p (v->new_reg, v2->new_reg)
2769 for (v = bl->giv; v; v = v->next_iv)
2773 /* Only split the giv if it has already been reduced, or if the loop is
2774 being completely unrolled. */
2775 if (unroll_type != UNROLL_COMPLETELY && v->ignore)
2778 /* The giv can be split if the insn that sets the giv is executed once
2779 and only once on every iteration of the loop. */
2780 /* An address giv can always be split. v->insn is just a use not a set,
2781 and hence it does not matter whether it is always executed. All that
2782 matters is that all the biv increments are always executed, and we
2783 won't reach here if they aren't. */
2784 if (v->giv_type != DEST_ADDR
2785 && (! v->always_computable
2786 || back_branch_in_range_p (v->insn, loop_start, loop_end)))
2789 /* The giv increment value must be a constant. */
2790 giv_inc = fold_rtx_mult_add (v->mult_val, increment, const0_rtx,
2792 if (! giv_inc || GET_CODE (giv_inc) != CONST_INT)
2795 /* The loop must be unrolled completely, or else have a known number of
2796 iterations and only one exit, or else the giv must be dead outside
2797 the loop, or else the final value of the giv must be known.
2798 Otherwise, it is not safe to split the giv since it may not have the
2799 proper value on loop exit. */
2801 /* The used outside loop test will fail for DEST_ADDR givs. They are
2802 never used outside the loop anyways, so it is always safe to split a
2806 if (unroll_type != UNROLL_COMPLETELY
2807 && (loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]]
2808 || unroll_type == UNROLL_NAIVE)
2809 && v->giv_type != DEST_ADDR
2810 /* The next part is true if the pseudo is used outside the loop.
2811 We assume that this is true for any pseudo created after loop
2812 starts, because we don't have a reg_n_info entry for them. */
2813 && (REGNO (v->dest_reg) >= max_reg_before_loop
2814 || (REGNO_FIRST_UID (REGNO (v->dest_reg)) != INSN_UID (v->insn)
2815 /* Check for the case where the pseudo is set by a shift/add
2816 sequence, in which case the first insn setting the pseudo
2817 is the first insn of the shift/add sequence. */
2818 && (! (tem = find_reg_note (v->insn, REG_RETVAL, NULL_RTX))
2819 || (REGNO_FIRST_UID (REGNO (v->dest_reg))
2820 != INSN_UID (XEXP (tem, 0)))))
2821 /* Line above always fails if INSN was moved by loop opt. */
2822 || (uid_luid[REGNO_LAST_UID (REGNO (v->dest_reg))]
2823 >= INSN_LUID (loop_end)))
2824 /* Givs made from biv increments are missed by the above test, so
2825 test explicitly for them. */
2826 && (REGNO (v->dest_reg) < first_increment_giv
2827 || REGNO (v->dest_reg) > last_increment_giv)
2828 && ! (final_value = v->final_value))
2832 /* Currently, non-reduced/final-value givs are never split. */
2833 /* Should emit insns after the loop if possible, as the biv final value
2836 /* If the final value is non-zero, and the giv has not been reduced,
2837 then must emit an instruction to set the final value. */
2838 if (final_value && !v->new_reg)
2840 /* Create a new register to hold the value of the giv, and then set
2841 the giv to its final value before the loop start. The giv is set
2842 to its final value before loop start to ensure that this insn
2843 will always be executed, no matter how we exit. */
2844 tem = gen_reg_rtx (v->mode);
2845 emit_insn_before (gen_move_insn (tem, v->dest_reg), loop_start);
2846 emit_insn_before (gen_move_insn (v->dest_reg, final_value),
2849 if (loop_dump_stream)
2850 fprintf (loop_dump_stream, "Giv %d mapped to %d for split.\n",
2851 REGNO (v->dest_reg), REGNO (tem));
2857 /* This giv is splittable. If completely unrolling the loop, save the
2858 giv's initial value. Otherwise, save the constant zero for it. */
2860 if (unroll_type == UNROLL_COMPLETELY)
2862 /* It is not safe to use bl->initial_value here, because it may not
2863 be invariant. It is safe to use the initial value stored in
2864 the splittable_regs array if it is set. In rare cases, it won't
2865 be set, so then we do exactly the same thing as
2866 find_splittable_regs does to get a safe value. */
2867 rtx biv_initial_value;
2869 if (splittable_regs[bl->regno])
2870 biv_initial_value = splittable_regs[bl->regno];
2871 else if (GET_CODE (bl->initial_value) != REG
2872 || (REGNO (bl->initial_value) != bl->regno
2873 && REGNO (bl->initial_value) >= FIRST_PSEUDO_REGISTER))
2874 biv_initial_value = bl->initial_value;
2877 rtx tem = gen_reg_rtx (bl->biv->mode);
2879 record_base_value (REGNO (tem), bl->biv->add_val, 0);
2880 emit_insn_before (gen_move_insn (tem, bl->biv->src_reg),
2882 biv_initial_value = tem;
2884 value = fold_rtx_mult_add (v->mult_val, biv_initial_value,
2885 v->add_val, v->mode);
2892 /* If a giv was combined with another giv, then we can only split
2893 this giv if the giv it was combined with was reduced. This
2894 is because the value of v->new_reg is meaningless in this
2896 if (v->same && ! v->same->new_reg)
2898 if (loop_dump_stream)
2899 fprintf (loop_dump_stream,
2900 "giv combined with unreduced giv not split.\n");
2903 /* If the giv is an address destination, it could be something other
2904 than a simple register, these have to be treated differently. */
2905 else if (v->giv_type == DEST_REG)
2907 /* If value is not a constant, register, or register plus
2908 constant, then compute its value into a register before
2909 loop start. This prevents invalid rtx sharing, and should
2910 generate better code. We can use bl->initial_value here
2911 instead of splittable_regs[bl->regno] because this code
2912 is going before the loop start. */
2913 if (unroll_type == UNROLL_COMPLETELY
2914 && GET_CODE (value) != CONST_INT
2915 && GET_CODE (value) != REG
2916 && (GET_CODE (value) != PLUS
2917 || GET_CODE (XEXP (value, 0)) != REG
2918 || GET_CODE (XEXP (value, 1)) != CONST_INT))
2920 rtx tem = gen_reg_rtx (v->mode);
2921 record_base_value (REGNO (tem), v->add_val, 0);
2922 emit_iv_add_mult (bl->initial_value, v->mult_val,
2923 v->add_val, tem, loop_start);
2927 splittable_regs[REGNO (v->new_reg)] = value;
2928 derived_regs[REGNO (v->new_reg)] = v->derived_from != 0;
2932 /* Splitting address givs is useful since it will often allow us
2933 to eliminate some increment insns for the base giv as
2936 /* If the addr giv is combined with a dest_reg giv, then all
2937 references to that dest reg will be remapped, which is NOT
2938 what we want for split addr regs. We always create a new
2939 register for the split addr giv, just to be safe. */
2941 /* If we have multiple identical address givs within a
2942 single instruction, then use a single pseudo reg for
2943 both. This is necessary in case one is a match_dup
2946 v->const_adjust = 0;
2950 v->dest_reg = v->same_insn->dest_reg;
2951 if (loop_dump_stream)
2952 fprintf (loop_dump_stream,
2953 "Sharing address givs in insn %d\n",
2954 INSN_UID (v->insn));
2956 /* If multiple address GIVs have been combined with the
2957 same dest_reg GIV, do not create a new register for
2959 else if (unroll_type != UNROLL_COMPLETELY
2960 && v->giv_type == DEST_ADDR
2961 && v->same && v->same->giv_type == DEST_ADDR
2962 && v->same->unrolled
2963 /* combine_givs_p may return true for some cases
2964 where the add and mult values are not equal.
2965 To share a register here, the values must be
2967 && rtx_equal_p (v->same->mult_val, v->mult_val)
2968 && rtx_equal_p (v->same->add_val, v->add_val)
2969 /* If the memory references have different modes,
2970 then the address may not be valid and we must
2971 not share registers. */
2972 && verify_addresses (v, giv_inc, unroll_number))
2974 v->dest_reg = v->same->dest_reg;
2977 else if (unroll_type != UNROLL_COMPLETELY)
2979 /* If not completely unrolling the loop, then create a new
2980 register to hold the split value of the DEST_ADDR giv.
2981 Emit insn to initialize its value before loop start. */
2983 rtx tem = gen_reg_rtx (v->mode);
2984 struct induction *same = v->same;
2985 rtx new_reg = v->new_reg;
2986 record_base_value (REGNO (tem), v->add_val, 0);
2988 if (same && same->derived_from)
2990 /* calculate_giv_inc doesn't work for derived givs.
2991 copy_loop_body works around the problem for the
2992 DEST_REG givs themselves, but it can't handle
2993 DEST_ADDR givs that have been combined with
2994 a derived DEST_REG giv.
2995 So Handle V as if the giv from which V->SAME has
2996 been derived has been combined with V.
2997 recombine_givs only derives givs from givs that
2998 are reduced the ordinary, so we need not worry
2999 about same->derived_from being in turn derived. */
3001 same = same->derived_from;
3002 new_reg = express_from (same, v);
3003 new_reg = replace_rtx (new_reg, same->dest_reg,
3007 /* If the address giv has a constant in its new_reg value,
3008 then this constant can be pulled out and put in value,
3009 instead of being part of the initialization code. */
3011 if (GET_CODE (new_reg) == PLUS
3012 && GET_CODE (XEXP (new_reg, 1)) == CONST_INT)
3015 = plus_constant (tem, INTVAL (XEXP (new_reg, 1)));
3017 /* Only succeed if this will give valid addresses.
3018 Try to validate both the first and the last
3019 address resulting from loop unrolling, if
3020 one fails, then can't do const elim here. */
3021 if (verify_addresses (v, giv_inc, unroll_number))
3023 /* Save the negative of the eliminated const, so
3024 that we can calculate the dest_reg's increment
3026 v->const_adjust = - INTVAL (XEXP (new_reg, 1));
3028 new_reg = XEXP (new_reg, 0);
3029 if (loop_dump_stream)
3030 fprintf (loop_dump_stream,
3031 "Eliminating constant from giv %d\n",
3040 /* If the address hasn't been checked for validity yet, do so
3041 now, and fail completely if either the first or the last
3042 unrolled copy of the address is not a valid address
3043 for the instruction that uses it. */
3044 if (v->dest_reg == tem
3045 && ! verify_addresses (v, giv_inc, unroll_number))
3047 for (v2 = v->next_iv; v2; v2 = v2->next_iv)
3048 if (v2->same_insn == v)
3051 if (loop_dump_stream)
3052 fprintf (loop_dump_stream,
3053 "Invalid address for giv at insn %d\n",
3054 INSN_UID (v->insn));
3058 v->new_reg = new_reg;
3061 /* We set this after the address check, to guarantee that
3062 the register will be initialized. */
3065 /* To initialize the new register, just move the value of
3066 new_reg into it. This is not guaranteed to give a valid
3067 instruction on machines with complex addressing modes.
3068 If we can't recognize it, then delete it and emit insns
3069 to calculate the value from scratch. */
3070 emit_insn_before (gen_rtx_SET (VOIDmode, tem,
3071 copy_rtx (v->new_reg)),
3073 if (recog_memoized (PREV_INSN (loop_start)) < 0)
3077 /* We can't use bl->initial_value to compute the initial
3078 value, because the loop may have been preconditioned.
3079 We must calculate it from NEW_REG. Try using
3080 force_operand instead of emit_iv_add_mult. */
3081 delete_insn (PREV_INSN (loop_start));
3084 ret = force_operand (v->new_reg, tem);
3086 emit_move_insn (tem, ret);
3087 sequence = gen_sequence ();
3089 emit_insn_before (sequence, loop_start);
3091 if (loop_dump_stream)
3092 fprintf (loop_dump_stream,
3093 "Invalid init insn, rewritten.\n");
3098 v->dest_reg = value;
3100 /* Check the resulting address for validity, and fail
3101 if the resulting address would be invalid. */
3102 if (! verify_addresses (v, giv_inc, unroll_number))
3104 for (v2 = v->next_iv; v2; v2 = v2->next_iv)
3105 if (v2->same_insn == v)
3108 if (loop_dump_stream)
3109 fprintf (loop_dump_stream,
3110 "Invalid address for giv at insn %d\n",
3111 INSN_UID (v->insn));
3114 if (v->same && v->same->derived_from)
3116 /* Handle V as if the giv from which V->SAME has
3117 been derived has been combined with V. */
3119 v->same = v->same->derived_from;
3120 v->new_reg = express_from (v->same, v);
3121 v->new_reg = replace_rtx (v->new_reg, v->same->dest_reg,
3127 /* Store the value of dest_reg into the insn. This sharing
3128 will not be a problem as this insn will always be copied
3131 *v->location = v->dest_reg;
3133 /* If this address giv is combined with a dest reg giv, then
3134 save the base giv's induction pointer so that we will be
3135 able to handle this address giv properly. The base giv
3136 itself does not have to be splittable. */
3138 if (v->same && v->same->giv_type == DEST_REG)
3139 addr_combined_regs[REGNO (v->same->new_reg)] = v->same;
3141 if (GET_CODE (v->new_reg) == REG)
3143 /* This giv maybe hasn't been combined with any others.
3144 Make sure that it's giv is marked as splittable here. */
3146 splittable_regs[REGNO (v->new_reg)] = value;
3147 derived_regs[REGNO (v->new_reg)] = v->derived_from != 0;
3149 /* Make it appear to depend upon itself, so that the
3150 giv will be properly split in the main loop above. */
3154 addr_combined_regs[REGNO (v->new_reg)] = v;
3158 if (loop_dump_stream)
3159 fprintf (loop_dump_stream, "DEST_ADDR giv being split.\n");
3165 /* Currently, unreduced giv's can't be split. This is not too much
3166 of a problem since unreduced giv's are not live across loop
3167 iterations anyways. When unrolling a loop completely though,
3168 it makes sense to reduce&split givs when possible, as this will
3169 result in simpler instructions, and will not require that a reg
3170 be live across loop iterations. */
3172 splittable_regs[REGNO (v->dest_reg)] = value;
3173 fprintf (stderr, "Giv %d at insn %d not reduced\n",
3174 REGNO (v->dest_reg), INSN_UID (v->insn));
3180 /* Unreduced givs are only updated once by definition. Reduced givs
3181 are updated as many times as their biv is. Mark it so if this is
3182 a splittable register. Don't need to do anything for address givs
3183 where this may not be a register. */
3185 if (GET_CODE (v->new_reg) == REG)
3189 count = reg_biv_class[REGNO (v->src_reg)]->biv_count;
3191 if (count > 1 && v->derived_from)
3192 /* In this case, there is one set where the giv insn was and one
3193 set each after each biv increment. (Most are likely dead.) */
3196 splittable_regs_updates[REGNO (v->new_reg)] = count;
3201 if (loop_dump_stream)
3205 if (GET_CODE (v->dest_reg) == CONST_INT)
3207 else if (GET_CODE (v->dest_reg) != REG)
3208 regnum = REGNO (XEXP (v->dest_reg, 0));
3210 regnum = REGNO (v->dest_reg);
3211 fprintf (loop_dump_stream, "Giv %d at insn %d safe to split.\n",
3212 regnum, INSN_UID (v->insn));
3219 /* Try to prove that the register is dead after the loop exits. Trace every
3220 loop exit looking for an insn that will always be executed, which sets
3221 the register to some value, and appears before the first use of the register
3222 is found. If successful, then return 1, otherwise return 0. */
3224 /* ?? Could be made more intelligent in the handling of jumps, so that
3225 it can search past if statements and other similar structures. */
3228 reg_dead_after_loop (reg, loop_start, loop_end)
3229 rtx reg, loop_start, loop_end;
3234 int label_count = 0;
3235 int this_loop_num = uid_loop_num[INSN_UID (loop_start)];
3237 /* In addition to checking all exits of this loop, we must also check
3238 all exits of inner nested loops that would exit this loop. We don't
3239 have any way to identify those, so we just give up if there are any
3240 such inner loop exits. */
3242 for (label = loop_number_exit_labels[this_loop_num]; label;
3243 label = LABEL_NEXTREF (label))
3246 if (label_count != loop_number_exit_count[this_loop_num])
3249 /* HACK: Must also search the loop fall through exit, create a label_ref
3250 here which points to the loop_end, and append the loop_number_exit_labels
3252 label = gen_rtx_LABEL_REF (VOIDmode, loop_end);
3253 LABEL_NEXTREF (label) = loop_number_exit_labels[this_loop_num];
3255 for ( ; label; label = LABEL_NEXTREF (label))
3257 /* Succeed if find an insn which sets the biv or if reach end of
3258 function. Fail if find an insn that uses the biv, or if come to
3259 a conditional jump. */
3261 insn = NEXT_INSN (XEXP (label, 0));
3264 code = GET_CODE (insn);
3265 if (GET_RTX_CLASS (code) == 'i')
3269 if (reg_referenced_p (reg, PATTERN (insn)))
3272 set = single_set (insn);
3273 if (set && rtx_equal_p (SET_DEST (set), reg))
3277 if (code == JUMP_INSN)
3279 if (GET_CODE (PATTERN (insn)) == RETURN)
3281 else if (! simplejump_p (insn)
3282 /* Prevent infinite loop following infinite loops. */
3283 || jump_count++ > 20)
3286 insn = JUMP_LABEL (insn);
3289 insn = NEXT_INSN (insn);
3293 /* Success, the register is dead on all loop exits. */
3297 /* Try to calculate the final value of the biv, the value it will have at
3298 the end of the loop. If we can do it, return that value. */
3301 final_biv_value (bl, loop_start, loop_end, n_iterations)
3302 struct iv_class *bl;
3303 rtx loop_start, loop_end;
3304 unsigned HOST_WIDE_INT n_iterations;
3308 /* ??? This only works for MODE_INT biv's. Reject all others for now. */
3310 if (GET_MODE_CLASS (bl->biv->mode) != MODE_INT)
3313 /* The final value for reversed bivs must be calculated differently than
3314 for ordinary bivs. In this case, there is already an insn after the
3315 loop which sets this biv's final value (if necessary), and there are
3316 no other loop exits, so we can return any value. */
3319 if (loop_dump_stream)
3320 fprintf (loop_dump_stream,
3321 "Final biv value for %d, reversed biv.\n", bl->regno);
3326 /* Try to calculate the final value as initial value + (number of iterations
3327 * increment). For this to work, increment must be invariant, the only
3328 exit from the loop must be the fall through at the bottom (otherwise
3329 it may not have its final value when the loop exits), and the initial
3330 value of the biv must be invariant. */
3332 if (n_iterations != 0
3333 && ! loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]]
3334 && invariant_p (bl->initial_value))
3336 increment = biv_total_increment (bl, loop_start, loop_end);
3338 if (increment && invariant_p (increment))
3340 /* Can calculate the loop exit value, emit insns after loop
3341 end to calculate this value into a temporary register in
3342 case it is needed later. */
3344 tem = gen_reg_rtx (bl->biv->mode);
3345 record_base_value (REGNO (tem), bl->biv->add_val, 0);
3346 /* Make sure loop_end is not the last insn. */
3347 if (NEXT_INSN (loop_end) == 0)
3348 emit_note_after (NOTE_INSN_DELETED, loop_end);
3349 emit_iv_add_mult (increment, GEN_INT (n_iterations),
3350 bl->initial_value, tem, NEXT_INSN (loop_end));
3352 if (loop_dump_stream)
3353 fprintf (loop_dump_stream,
3354 "Final biv value for %d, calculated.\n", bl->regno);
3360 /* Check to see if the biv is dead at all loop exits. */
3361 if (reg_dead_after_loop (bl->biv->src_reg, loop_start, loop_end))
3363 if (loop_dump_stream)
3364 fprintf (loop_dump_stream,
3365 "Final biv value for %d, biv dead after loop exit.\n",
3374 /* Try to calculate the final value of the giv, the value it will have at
3375 the end of the loop. If we can do it, return that value. */
3378 final_giv_value (v, loop_start, loop_end, n_iterations)
3379 struct induction *v;
3380 rtx loop_start, loop_end;
3381 unsigned HOST_WIDE_INT n_iterations;
3383 struct iv_class *bl;
3386 rtx insert_before, seq;
3388 bl = reg_biv_class[REGNO (v->src_reg)];
3390 /* The final value for givs which depend on reversed bivs must be calculated
3391 differently than for ordinary givs. In this case, there is already an
3392 insn after the loop which sets this giv's final value (if necessary),
3393 and there are no other loop exits, so we can return any value. */
3396 if (loop_dump_stream)
3397 fprintf (loop_dump_stream,
3398 "Final giv value for %d, depends on reversed biv\n",
3399 REGNO (v->dest_reg));
3403 /* Try to calculate the final value as a function of the biv it depends
3404 upon. The only exit from the loop must be the fall through at the bottom
3405 (otherwise it may not have its final value when the loop exits). */
3407 /* ??? Can calculate the final giv value by subtracting off the
3408 extra biv increments times the giv's mult_val. The loop must have
3409 only one exit for this to work, but the loop iterations does not need
3412 if (n_iterations != 0
3413 && ! loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]])
3415 /* ?? It is tempting to use the biv's value here since these insns will
3416 be put after the loop, and hence the biv will have its final value
3417 then. However, this fails if the biv is subsequently eliminated.
3418 Perhaps determine whether biv's are eliminable before trying to
3419 determine whether giv's are replaceable so that we can use the
3420 biv value here if it is not eliminable. */
3422 /* We are emitting code after the end of the loop, so we must make
3423 sure that bl->initial_value is still valid then. It will still
3424 be valid if it is invariant. */
3426 increment = biv_total_increment (bl, loop_start, loop_end);
3428 if (increment && invariant_p (increment)
3429 && invariant_p (bl->initial_value))
3431 /* Can calculate the loop exit value of its biv as
3432 (n_iterations * increment) + initial_value */
3434 /* The loop exit value of the giv is then
3435 (final_biv_value - extra increments) * mult_val + add_val.
3436 The extra increments are any increments to the biv which
3437 occur in the loop after the giv's value is calculated.
3438 We must search from the insn that sets the giv to the end
3439 of the loop to calculate this value. */
3441 insert_before = NEXT_INSN (loop_end);
3443 /* Put the final biv value in tem. */
3444 tem = gen_reg_rtx (bl->biv->mode);
3445 record_base_value (REGNO (tem), bl->biv->add_val, 0);
3446 emit_iv_add_mult (increment, GEN_INT (n_iterations),
3447 bl->initial_value, tem, insert_before);
3449 /* Subtract off extra increments as we find them. */
3450 for (insn = NEXT_INSN (v->insn); insn != loop_end;
3451 insn = NEXT_INSN (insn))
3453 struct induction *biv;
3455 for (biv = bl->biv; biv; biv = biv->next_iv)
3456 if (biv->insn == insn)
3459 tem = expand_binop (GET_MODE (tem), sub_optab, tem,
3460 biv->add_val, NULL_RTX, 0,
3462 seq = gen_sequence ();
3464 emit_insn_before (seq, insert_before);
3468 /* Now calculate the giv's final value. */
3469 emit_iv_add_mult (tem, v->mult_val, v->add_val, tem,
3472 if (loop_dump_stream)
3473 fprintf (loop_dump_stream,
3474 "Final giv value for %d, calc from biv's value.\n",
3475 REGNO (v->dest_reg));
3481 /* Replaceable giv's should never reach here. */
3485 /* Check to see if the biv is dead at all loop exits. */
3486 if (reg_dead_after_loop (v->dest_reg, loop_start, loop_end))
3488 if (loop_dump_stream)
3489 fprintf (loop_dump_stream,
3490 "Final giv value for %d, giv dead after loop exit.\n",
3491 REGNO (v->dest_reg));
3500 /* Look back before LOOP_START for then insn that sets REG and return
3501 the equivalent constant if there is a REG_EQUAL note otherwise just
3502 the SET_SRC of REG. */
3505 loop_find_equiv_value (loop_start, reg)
3513 for (insn = PREV_INSN (loop_start); insn ; insn = PREV_INSN (insn))
3515 if (GET_CODE (insn) == CODE_LABEL)
3518 else if (GET_RTX_CLASS (GET_CODE (insn)) == 'i'
3519 && reg_set_p (reg, insn))
3521 /* We found the last insn before the loop that sets the register.
3522 If it sets the entire register, and has a REG_EQUAL note,
3523 then use the value of the REG_EQUAL note. */
3524 if ((set = single_set (insn))
3525 && (SET_DEST (set) == reg))
3527 rtx note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
3529 /* Only use the REG_EQUAL note if it is a constant.
3530 Other things, divide in particular, will cause
3531 problems later if we use them. */
3532 if (note && GET_CODE (XEXP (note, 0)) != EXPR_LIST
3533 && CONSTANT_P (XEXP (note, 0)))
3534 ret = XEXP (note, 0);
3536 ret = SET_SRC (set);
3544 /* Return a simplified rtx for the expression OP - REG.
3546 REG must appear in OP, and OP must be a register or the sum of a register
3549 Thus, the return value must be const0_rtx or the second term.
3551 The caller is responsible for verifying that REG appears in OP and OP has
3555 subtract_reg_term (op, reg)
3560 if (GET_CODE (op) == PLUS)
3562 if (XEXP (op, 0) == reg)
3563 return XEXP (op, 1);
3564 else if (XEXP (op, 1) == reg)
3565 return XEXP (op, 0);
3567 /* OP does not contain REG as a term. */
3572 /* Find and return register term common to both expressions OP0 and
3573 OP1 or NULL_RTX if no such term exists. Each expression must be a
3574 REG or a PLUS of a REG. */
3577 find_common_reg_term (op0, op1)
3580 if ((GET_CODE (op0) == REG || GET_CODE (op0) == PLUS)
3581 && (GET_CODE (op1) == REG || GET_CODE (op1) == PLUS))
3588 if (GET_CODE (op0) == PLUS)
3589 op01 = XEXP (op0, 1), op00 = XEXP (op0, 0);
3591 op01 = const0_rtx, op00 = op0;
3593 if (GET_CODE (op1) == PLUS)
3594 op11 = XEXP (op1, 1), op10 = XEXP (op1, 0);
3596 op11 = const0_rtx, op10 = op1;
3598 /* Find and return common register term if present. */
3599 if (REG_P (op00) && (op00 == op10 || op00 == op11))
3601 else if (REG_P (op01) && (op01 == op10 || op01 == op11))
3605 /* No common register term found. */
3609 /* Calculate the number of loop iterations. Returns the exact number of loop
3610 iterations if it can be calculated, otherwise returns zero. */
3612 unsigned HOST_WIDE_INT
3613 loop_iterations (loop_start, loop_end, loop_info)
3614 rtx loop_start, loop_end;
3615 struct loop_info *loop_info;
3617 rtx comparison, comparison_value;
3618 rtx iteration_var, initial_value, increment, final_value;
3619 enum rtx_code comparison_code;
3620 HOST_WIDE_INT abs_inc;
3621 unsigned HOST_WIDE_INT abs_diff;
3624 int unsigned_p, compare_dir, final_larger;
3628 loop_info->n_iterations = 0;
3629 loop_info->initial_value = 0;
3630 loop_info->initial_equiv_value = 0;
3631 loop_info->comparison_value = 0;
3632 loop_info->final_value = 0;
3633 loop_info->final_equiv_value = 0;
3634 loop_info->increment = 0;
3635 loop_info->iteration_var = 0;
3636 loop_info->unroll_number = 1;
3638 /* We used to use prev_nonnote_insn here, but that fails because it might
3639 accidentally get the branch for a contained loop if the branch for this
3640 loop was deleted. We can only trust branches immediately before the
3642 last_loop_insn = PREV_INSN (loop_end);
3644 /* ??? We should probably try harder to find the jump insn
3645 at the end of the loop. The following code assumes that
3646 the last loop insn is a jump to the top of the loop. */
3647 if (GET_CODE (last_loop_insn) != JUMP_INSN)
3649 if (loop_dump_stream)
3650 fprintf (loop_dump_stream,
3651 "Loop iterations: No final conditional branch found.\n");
3655 /* If there is a more than a single jump to the top of the loop
3656 we cannot (easily) determine the iteration count. */
3657 if (LABEL_NUSES (JUMP_LABEL (last_loop_insn)) > 1)
3659 if (loop_dump_stream)
3660 fprintf (loop_dump_stream,
3661 "Loop iterations: Loop has multiple back edges.\n");
3665 /* Find the iteration variable. If the last insn is a conditional
3666 branch, and the insn before tests a register value, make that the
3667 iteration variable. */
3669 comparison = get_condition_for_loop (last_loop_insn);
3670 if (comparison == 0)
3672 if (loop_dump_stream)
3673 fprintf (loop_dump_stream,
3674 "Loop iterations: No final comparison found.\n");
3678 /* ??? Get_condition may switch position of induction variable and
3679 invariant register when it canonicalizes the comparison. */
3681 comparison_code = GET_CODE (comparison);
3682 iteration_var = XEXP (comparison, 0);
3683 comparison_value = XEXP (comparison, 1);
3685 if (GET_CODE (iteration_var) != REG)
3687 if (loop_dump_stream)
3688 fprintf (loop_dump_stream,
3689 "Loop iterations: Comparison not against register.\n");
3693 /* This can happen due to optimization in load_mems. */
3694 if ((unsigned) REGNO (iteration_var) >= reg_iv_type->num_elements)
3697 iteration_info (iteration_var, &initial_value, &increment,
3698 loop_start, loop_end);
3699 if (initial_value == 0)
3700 /* iteration_info already printed a message. */
3705 switch (comparison_code)
3720 /* Cannot determine loop iterations with this case. */
3739 /* If the comparison value is an invariant register, then try to find
3740 its value from the insns before the start of the loop. */
3742 final_value = comparison_value;
3743 if (GET_CODE (comparison_value) == REG && invariant_p (comparison_value))
3745 final_value = loop_find_equiv_value (loop_start, comparison_value);
3746 /* If we don't get an invariant final value, we are better
3747 off with the original register. */
3748 if (!invariant_p (final_value))
3749 final_value = comparison_value;
3752 /* Calculate the approximate final value of the induction variable
3753 (on the last successful iteration). The exact final value
3754 depends on the branch operator, and increment sign. It will be
3755 wrong if the iteration variable is not incremented by one each
3756 time through the loop and (comparison_value + off_by_one -
3757 initial_value) % increment != 0.
3758 ??? Note that the final_value may overflow and thus final_larger
3759 will be bogus. A potentially infinite loop will be classified
3760 as immediate, e.g. for (i = 0x7ffffff0; i <= 0x7fffffff; i++) */
3762 final_value = plus_constant (final_value, off_by_one);
3764 /* Save the calculated values describing this loop's bounds, in case
3765 precondition_loop_p will need them later. These values can not be
3766 recalculated inside precondition_loop_p because strength reduction
3767 optimizations may obscure the loop's structure.
3769 These values are only required by precondition_loop_p and insert_bct
3770 whenever the number of iterations cannot be computed at compile time.
3771 Only the difference between final_value and initial_value is
3772 important. Note that final_value is only approximate. */
3773 loop_info->initial_value = initial_value;
3774 loop_info->comparison_value = comparison_value;
3775 loop_info->final_value = plus_constant (comparison_value, off_by_one);
3776 loop_info->increment = increment;
3777 loop_info->iteration_var = iteration_var;
3778 loop_info->comparison_code = comparison_code;
3780 /* Try to determine the iteration count for loops such
3781 as (for i = init; i < init + const; i++). When running the
3782 loop optimization twice, the first pass often converts simple
3783 loops into this form. */
3785 if (REG_P (initial_value))
3791 reg1 = initial_value;
3792 if (GET_CODE (final_value) == PLUS)
3793 reg2 = XEXP (final_value, 0), const2 = XEXP (final_value, 1);
3795 reg2 = final_value, const2 = const0_rtx;
3797 /* Check for initial_value = reg1, final_value = reg2 + const2,
3798 where reg1 != reg2. */
3799 if (REG_P (reg2) && reg2 != reg1)
3803 /* Find what reg1 is equivalent to. Hopefully it will
3804 either be reg2 or reg2 plus a constant. */
3805 temp = loop_find_equiv_value (loop_start, reg1);
3806 if (find_common_reg_term (temp, reg2))
3807 initial_value = temp;
3810 /* Find what reg2 is equivalent to. Hopefully it will
3811 either be reg1 or reg1 plus a constant. Let's ignore
3812 the latter case for now since it is not so common. */
3813 temp = loop_find_equiv_value (loop_start, reg2);
3814 if (temp == loop_info->iteration_var)
3815 temp = initial_value;
3817 final_value = (const2 == const0_rtx)
3818 ? reg1 : gen_rtx_PLUS (GET_MODE (reg1), reg1, const2);
3821 else if (loop_info->vtop && GET_CODE (reg2) == CONST_INT)
3825 /* When running the loop optimizer twice, check_dbra_loop
3826 further obfuscates reversible loops of the form:
3827 for (i = init; i < init + const; i++). We often end up with
3828 final_value = 0, initial_value = temp, temp = temp2 - init,
3829 where temp2 = init + const. If the loop has a vtop we
3830 can replace initial_value with const. */
3832 temp = loop_find_equiv_value (loop_start, reg1);
3833 if (GET_CODE (temp) == MINUS && REG_P (XEXP (temp, 0)))
3835 rtx temp2 = loop_find_equiv_value (loop_start, XEXP (temp, 0));
3836 if (GET_CODE (temp2) == PLUS
3837 && XEXP (temp2, 0) == XEXP (temp, 1))
3838 initial_value = XEXP (temp2, 1);
3843 /* If have initial_value = reg + const1 and final_value = reg +
3844 const2, then replace initial_value with const1 and final_value
3845 with const2. This should be safe since we are protected by the
3846 initial comparison before entering the loop if we have a vtop.
3847 For example, a + b < a + c is not equivalent to b < c for all a
3848 when using modulo arithmetic.
3850 ??? Without a vtop we could still perform the optimization if we check
3851 the initial and final values carefully. */
3853 && (reg_term = find_common_reg_term (initial_value, final_value)))
3855 initial_value = subtract_reg_term (initial_value, reg_term);
3856 final_value = subtract_reg_term (final_value, reg_term);
3859 loop_info->initial_equiv_value = initial_value;
3860 loop_info->final_equiv_value = final_value;
3862 /* For EQ comparison loops, we don't have a valid final value.
3863 Check this now so that we won't leave an invalid value if we
3864 return early for any other reason. */
3865 if (comparison_code == EQ)
3866 loop_info->final_equiv_value = loop_info->final_value = 0;
3870 if (loop_dump_stream)
3871 fprintf (loop_dump_stream,
3872 "Loop iterations: Increment value can't be calculated.\n");
3876 if (GET_CODE (increment) != CONST_INT)
3878 /* If we have a REG, check to see if REG holds a constant value. */
3879 /* ??? Other RTL, such as (neg (reg)) is possible here, but it isn't
3880 clear if it is worthwhile to try to handle such RTL. */
3881 if (GET_CODE (increment) == REG || GET_CODE (increment) == SUBREG)
3882 increment = loop_find_equiv_value (loop_start, increment);
3884 if (GET_CODE (increment) != CONST_INT)
3886 if (loop_dump_stream)
3888 fprintf (loop_dump_stream,
3889 "Loop iterations: Increment value not constant ");
3890 print_rtl (loop_dump_stream, increment);
3891 fprintf (loop_dump_stream, ".\n");
3895 loop_info->increment = increment;
3898 if (GET_CODE (initial_value) != CONST_INT)
3900 if (loop_dump_stream)
3902 fprintf (loop_dump_stream,
3903 "Loop iterations: Initial value not constant ");
3904 print_rtl (loop_dump_stream, initial_value);
3905 fprintf (loop_dump_stream, ".\n");
3909 else if (comparison_code == EQ)
3911 if (loop_dump_stream)
3912 fprintf (loop_dump_stream,
3913 "Loop iterations: EQ comparison loop.\n");
3916 else if (GET_CODE (final_value) != CONST_INT)
3918 if (loop_dump_stream)
3920 fprintf (loop_dump_stream,
3921 "Loop iterations: Final value not constant ");
3922 print_rtl (loop_dump_stream, final_value);
3923 fprintf (loop_dump_stream, ".\n");
3928 /* Final_larger is 1 if final larger, 0 if they are equal, otherwise -1. */
3931 = ((unsigned HOST_WIDE_INT) INTVAL (final_value)
3932 > (unsigned HOST_WIDE_INT) INTVAL (initial_value))
3933 - ((unsigned HOST_WIDE_INT) INTVAL (final_value)
3934 < (unsigned HOST_WIDE_INT) INTVAL (initial_value));
3936 final_larger = (INTVAL (final_value) > INTVAL (initial_value))
3937 - (INTVAL (final_value) < INTVAL (initial_value));
3939 if (INTVAL (increment) > 0)
3941 else if (INTVAL (increment) == 0)
3946 /* There are 27 different cases: compare_dir = -1, 0, 1;
3947 final_larger = -1, 0, 1; increment_dir = -1, 0, 1.
3948 There are 4 normal cases, 4 reverse cases (where the iteration variable
3949 will overflow before the loop exits), 4 infinite loop cases, and 15
3950 immediate exit (0 or 1 iteration depending on loop type) cases.
3951 Only try to optimize the normal cases. */
3953 /* (compare_dir/final_larger/increment_dir)
3954 Normal cases: (0/-1/-1), (0/1/1), (-1/-1/-1), (1/1/1)
3955 Reverse cases: (0/-1/1), (0/1/-1), (-1/-1/1), (1/1/-1)
3956 Infinite loops: (0/-1/0), (0/1/0), (-1/-1/0), (1/1/0)
3957 Immediate exit: (0/0/X), (-1/0/X), (-1/1/X), (1/0/X), (1/-1/X) */
3959 /* ?? If the meaning of reverse loops (where the iteration variable
3960 will overflow before the loop exits) is undefined, then could
3961 eliminate all of these special checks, and just always assume
3962 the loops are normal/immediate/infinite. Note that this means
3963 the sign of increment_dir does not have to be known. Also,
3964 since it does not really hurt if immediate exit loops or infinite loops
3965 are optimized, then that case could be ignored also, and hence all
3966 loops can be optimized.
3968 According to ANSI Spec, the reverse loop case result is undefined,
3969 because the action on overflow is undefined.
3971 See also the special test for NE loops below. */
3973 if (final_larger == increment_dir && final_larger != 0
3974 && (final_larger == compare_dir || compare_dir == 0))
3979 if (loop_dump_stream)
3980 fprintf (loop_dump_stream,
3981 "Loop iterations: Not normal loop.\n");
3985 /* Calculate the number of iterations, final_value is only an approximation,
3986 so correct for that. Note that abs_diff and n_iterations are
3987 unsigned, because they can be as large as 2^n - 1. */
3989 abs_inc = INTVAL (increment);
3991 abs_diff = INTVAL (final_value) - INTVAL (initial_value);
3992 else if (abs_inc < 0)
3994 abs_diff = INTVAL (initial_value) - INTVAL (final_value);
4000 /* For NE tests, make sure that the iteration variable won't miss
4001 the final value. If abs_diff mod abs_incr is not zero, then the
4002 iteration variable will overflow before the loop exits, and we
4003 can not calculate the number of iterations. */
4004 if (compare_dir == 0 && (abs_diff % abs_inc) != 0)
4007 /* Note that the number of iterations could be calculated using
4008 (abs_diff + abs_inc - 1) / abs_inc, provided care was taken to
4009 handle potential overflow of the summation. */
4010 loop_info->n_iterations = abs_diff / abs_inc + ((abs_diff % abs_inc) != 0);
4011 return loop_info->n_iterations;
4015 /* Replace uses of split bivs with their split pseudo register. This is
4016 for original instructions which remain after loop unrolling without
4020 remap_split_bivs (x)
4023 register enum rtx_code code;
4025 register const char *fmt;
4030 code = GET_CODE (x);
4045 /* If non-reduced/final-value givs were split, then this would also
4046 have to remap those givs also. */
4048 if (REGNO (x) < max_reg_before_loop
4049 && REG_IV_TYPE (REGNO (x)) == BASIC_INDUCT)
4050 return reg_biv_class[REGNO (x)]->biv->src_reg;
4057 fmt = GET_RTX_FORMAT (code);
4058 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4061 XEXP (x, i) = remap_split_bivs (XEXP (x, i));
4065 for (j = 0; j < XVECLEN (x, i); j++)
4066 XVECEXP (x, i, j) = remap_split_bivs (XVECEXP (x, i, j));
4072 /* If FIRST_UID is a set of REGNO, and FIRST_UID dominates LAST_UID (e.g.
4073 FIST_UID is always executed if LAST_UID is), then return 1. Otherwise
4074 return 0. COPY_START is where we can start looking for the insns
4075 FIRST_UID and LAST_UID. COPY_END is where we stop looking for these
4078 If there is no JUMP_INSN between LOOP_START and FIRST_UID, then FIRST_UID
4079 must dominate LAST_UID.
4081 If there is a CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID
4082 may not dominate LAST_UID.
4084 If there is no CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID
4085 must dominate LAST_UID. */
4088 set_dominates_use (regno, first_uid, last_uid, copy_start, copy_end)
4095 int passed_jump = 0;
4096 rtx p = NEXT_INSN (copy_start);
4098 while (INSN_UID (p) != first_uid)
4100 if (GET_CODE (p) == JUMP_INSN)
4102 /* Could not find FIRST_UID. */
4108 /* Verify that FIRST_UID is an insn that entirely sets REGNO. */
4109 if (GET_RTX_CLASS (GET_CODE (p)) != 'i'
4110 || ! dead_or_set_regno_p (p, regno))
4113 /* FIRST_UID is always executed. */
4114 if (passed_jump == 0)
4117 while (INSN_UID (p) != last_uid)
4119 /* If we see a CODE_LABEL between FIRST_UID and LAST_UID, then we
4120 can not be sure that FIRST_UID dominates LAST_UID. */
4121 if (GET_CODE (p) == CODE_LABEL)
4123 /* Could not find LAST_UID, but we reached the end of the loop, so
4125 else if (p == copy_end)
4130 /* FIRST_UID is always executed if LAST_UID is executed. */