1 /* Try to unroll loops, and split induction variables.
2 Copyright (C) 1992, 1993, 1994, 1995, 1997, 1998, 1999, 2000, 2001, 2002
3 Free Software Foundation, Inc.
4 Contributed by James E. Wilson, Cygnus Support/UC Berkeley.
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify it under
9 the terms of the GNU General Public License as published by the Free
10 Software Foundation; either version 2, or (at your option) any later
13 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
14 WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING. If not, write to the Free
20 Software Foundation, 59 Temple Place - Suite 330, Boston, MA
23 /* Try to unroll a loop, and split induction variables.
25 Loops for which the number of iterations can be calculated exactly are
26 handled specially. If the number of iterations times the insn_count is
27 less than MAX_UNROLLED_INSNS, then the loop is unrolled completely.
28 Otherwise, we try to unroll the loop a number of times modulo the number
29 of iterations, so that only one exit test will be needed. It is unrolled
30 a number of times approximately equal to MAX_UNROLLED_INSNS divided by
33 Otherwise, if the number of iterations can be calculated exactly at
34 run time, and the loop is always entered at the top, then we try to
35 precondition the loop. That is, at run time, calculate how many times
36 the loop will execute, and then execute the loop body a few times so
37 that the remaining iterations will be some multiple of 4 (or 2 if the
38 loop is large). Then fall through to a loop unrolled 4 (or 2) times,
39 with only one exit test needed at the end of the loop.
41 Otherwise, if the number of iterations can not be calculated exactly,
42 not even at run time, then we still unroll the loop a number of times
43 approximately equal to MAX_UNROLLED_INSNS divided by the insn count,
44 but there must be an exit test after each copy of the loop body.
46 For each induction variable, which is dead outside the loop (replaceable)
47 or for which we can easily calculate the final value, if we can easily
48 calculate its value at each place where it is set as a function of the
49 current loop unroll count and the variable's value at loop entry, then
50 the induction variable is split into `N' different variables, one for
51 each copy of the loop body. One variable is live across the backward
52 branch, and the others are all calculated as a function of this variable.
53 This helps eliminate data dependencies, and leads to further opportunities
56 /* Possible improvements follow: */
58 /* ??? Add an extra pass somewhere to determine whether unrolling will
59 give any benefit. E.g. after generating all unrolled insns, compute the
60 cost of all insns and compare against cost of insns in rolled loop.
62 - On traditional architectures, unrolling a non-constant bound loop
63 is a win if there is a giv whose only use is in memory addresses, the
64 memory addresses can be split, and hence giv increments can be
66 - It is also a win if the loop is executed many times, and preconditioning
67 can be performed for the loop.
68 Add code to check for these and similar cases. */
70 /* ??? Improve control of which loops get unrolled. Could use profiling
71 info to only unroll the most commonly executed loops. Perhaps have
72 a user specifyable option to control the amount of code expansion,
73 or the percent of loops to consider for unrolling. Etc. */
75 /* ??? Look at the register copies inside the loop to see if they form a
76 simple permutation. If so, iterate the permutation until it gets back to
77 the start state. This is how many times we should unroll the loop, for
78 best results, because then all register copies can be eliminated.
79 For example, the lisp nreverse function should be unrolled 3 times
88 ??? The number of times to unroll the loop may also be based on data
89 references in the loop. For example, if we have a loop that references
90 x[i-1], x[i], and x[i+1], we should unroll it a multiple of 3 times. */
92 /* ??? Add some simple linear equation solving capability so that we can
93 determine the number of loop iterations for more complex loops.
94 For example, consider this loop from gdb
95 #define SWAP_TARGET_AND_HOST(buffer,len)
98 char *p = (char *) buffer;
99 char *q = ((char *) buffer) + len - 1;
100 int iterations = (len + 1) >> 1;
102 for (p; p < q; p++, q--;)
110 start value = p = &buffer + current_iteration
111 end value = q = &buffer + len - 1 - current_iteration
112 Given the loop exit test of "p < q", then there must be "q - p" iterations,
113 set equal to zero and solve for number of iterations:
114 q - p = len - 1 - 2*current_iteration = 0
115 current_iteration = (len - 1) / 2
116 Hence, there are (len - 1) / 2 (rounded up to the nearest integer)
117 iterations of this loop. */
119 /* ??? Currently, no labels are marked as loop invariant when doing loop
120 unrolling. This is because an insn inside the loop, that loads the address
121 of a label inside the loop into a register, could be moved outside the loop
122 by the invariant code motion pass if labels were invariant. If the loop
123 is subsequently unrolled, the code will be wrong because each unrolled
124 body of the loop will use the same address, whereas each actually needs a
125 different address. A case where this happens is when a loop containing
126 a switch statement is unrolled.
128 It would be better to let labels be considered invariant. When we
129 unroll loops here, check to see if any insns using a label local to the
130 loop were moved before the loop. If so, then correct the problem, by
131 moving the insn back into the loop, or perhaps replicate the insn before
132 the loop, one copy for each time the loop is unrolled. */
138 #include "insn-config.h"
139 #include "integrate.h"
143 #include "function.h"
147 #include "hard-reg-set.h"
148 #include "basic-block.h"
152 /* The prime factors looked for when trying to unroll a loop by some
153 number which is modulo the total number of iterations. Just checking
154 for these 4 prime factors will find at least one factor for 75% of
155 all numbers theoretically. Practically speaking, this will succeed
156 almost all of the time since loops are generally a multiple of 2
159 #define NUM_FACTORS 4
161 static struct _factor { const int factor; int count; }
162 factors[NUM_FACTORS] = { {2, 0}, {3, 0}, {5, 0}, {7, 0}};
164 /* Describes the different types of loop unrolling performed. */
173 /* Indexed by register number, if non-zero, then it contains a pointer
174 to a struct induction for a DEST_REG giv which has been combined with
175 one of more address givs. This is needed because whenever such a DEST_REG
176 giv is modified, we must modify the value of all split address givs
177 that were combined with this DEST_REG giv. */
179 static struct induction **addr_combined_regs;
181 /* Indexed by register number, if this is a splittable induction variable,
182 then this will hold the current value of the register, which depends on the
185 static rtx *splittable_regs;
187 /* Indexed by register number, if this is a splittable induction variable,
188 then this will hold the number of instructions in the loop that modify
189 the induction variable. Used to ensure that only the last insn modifying
190 a split iv will update the original iv of the dest. */
192 static int *splittable_regs_updates;
194 /* Forward declarations. */
196 static void init_reg_map PARAMS ((struct inline_remap *, int));
197 static rtx calculate_giv_inc PARAMS ((rtx, rtx, unsigned int));
198 static rtx initial_reg_note_copy PARAMS ((rtx, struct inline_remap *));
199 static void final_reg_note_copy PARAMS ((rtx *, struct inline_remap *));
200 static void copy_loop_body PARAMS ((struct loop *, rtx, rtx,
201 struct inline_remap *, rtx, int,
202 enum unroll_types, rtx, rtx, rtx, rtx));
203 static int find_splittable_regs PARAMS ((const struct loop *,
204 enum unroll_types, int));
205 static int find_splittable_givs PARAMS ((const struct loop *,
206 struct iv_class *, enum unroll_types,
208 static int reg_dead_after_loop PARAMS ((const struct loop *, rtx));
209 static rtx fold_rtx_mult_add PARAMS ((rtx, rtx, rtx, enum machine_mode));
210 static int verify_addresses PARAMS ((struct induction *, rtx, int));
211 static rtx remap_split_bivs PARAMS ((struct loop *, rtx));
212 static rtx find_common_reg_term PARAMS ((rtx, rtx));
213 static rtx subtract_reg_term PARAMS ((rtx, rtx));
214 static rtx loop_find_equiv_value PARAMS ((const struct loop *, rtx));
215 static rtx ujump_to_loop_cont PARAMS ((rtx, rtx));
217 /* Try to unroll one loop and split induction variables in the loop.
219 The loop is described by the arguments LOOP and INSN_COUNT.
220 STRENGTH_REDUCTION_P indicates whether information generated in the
221 strength reduction pass is available.
223 This function is intended to be called from within `strength_reduce'
227 unroll_loop (loop, insn_count, strength_reduce_p)
230 int strength_reduce_p;
232 struct loop_info *loop_info = LOOP_INFO (loop);
233 struct loop_ivs *ivs = LOOP_IVS (loop);
236 unsigned HOST_WIDE_INT temp;
237 int unroll_number = 1;
238 rtx copy_start, copy_end;
239 rtx insn, sequence, pattern, tem;
240 int max_labelno, max_insnno;
242 struct inline_remap *map;
243 char *local_label = NULL;
245 unsigned int max_local_regnum;
246 unsigned int maxregnum;
250 int splitting_not_safe = 0;
251 enum unroll_types unroll_type = UNROLL_NAIVE;
252 int loop_preconditioned = 0;
254 /* This points to the last real insn in the loop, which should be either
255 a JUMP_INSN (for conditional jumps) or a BARRIER (for unconditional
258 rtx loop_start = loop->start;
259 rtx loop_end = loop->end;
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 /* Determine type of unroll to perform. Depends on the number of iterations
272 and the size of the loop. */
274 /* If there is no strength reduce info, then set
275 loop_info->n_iterations to zero. This can happen if
276 strength_reduce can't find any bivs in the loop. A value of zero
277 indicates that the number of iterations could not be calculated. */
279 if (! strength_reduce_p)
280 loop_info->n_iterations = 0;
282 if (loop_dump_stream && loop_info->n_iterations > 0)
284 fputs ("Loop unrolling: ", loop_dump_stream);
285 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC,
286 loop_info->n_iterations);
287 fputs (" iterations.\n", loop_dump_stream);
290 /* Find and save a pointer to the last nonnote insn in the loop. */
292 last_loop_insn = prev_nonnote_insn (loop_end);
294 /* Calculate how many times to unroll the loop. Indicate whether or
295 not the loop is being completely unrolled. */
297 if (loop_info->n_iterations == 1)
299 /* Handle the case where the loop begins with an unconditional
300 jump to the loop condition. Make sure to delete the jump
301 insn, otherwise the loop body will never execute. */
303 rtx ujump = ujump_to_loop_cont (loop->start, loop->cont);
305 delete_related_insns (ujump);
307 /* If number of iterations is exactly 1, then eliminate the compare and
308 branch at the end of the loop since they will never be taken.
309 Then return, since no other action is needed here. */
311 /* If the last instruction is not a BARRIER or a JUMP_INSN, then
312 don't do anything. */
314 if (GET_CODE (last_loop_insn) == BARRIER)
316 /* Delete the jump insn. This will delete the barrier also. */
317 delete_related_insns (PREV_INSN (last_loop_insn));
319 else if (GET_CODE (last_loop_insn) == JUMP_INSN)
322 rtx prev = PREV_INSN (last_loop_insn);
324 delete_related_insns (last_loop_insn);
326 /* The immediately preceding insn may be a compare which must be
328 if (only_sets_cc0_p (prev))
329 delete_related_insns (prev);
333 /* Remove the loop notes since this is no longer a loop. */
335 delete_related_insns (loop->vtop);
337 delete_related_insns (loop->cont);
339 delete_related_insns (loop_start);
341 delete_related_insns (loop_end);
345 else if (loop_info->n_iterations > 0
346 /* Avoid overflow in the next expression. */
347 && loop_info->n_iterations < MAX_UNROLLED_INSNS
348 && loop_info->n_iterations * insn_count < MAX_UNROLLED_INSNS)
350 unroll_number = loop_info->n_iterations;
351 unroll_type = UNROLL_COMPLETELY;
353 else if (loop_info->n_iterations > 0)
355 /* Try to factor the number of iterations. Don't bother with the
356 general case, only using 2, 3, 5, and 7 will get 75% of all
357 numbers theoretically, and almost all in practice. */
359 for (i = 0; i < NUM_FACTORS; i++)
360 factors[i].count = 0;
362 temp = loop_info->n_iterations;
363 for (i = NUM_FACTORS - 1; i >= 0; i--)
364 while (temp % factors[i].factor == 0)
367 temp = temp / factors[i].factor;
370 /* Start with the larger factors first so that we generally
371 get lots of unrolling. */
375 for (i = 3; i >= 0; i--)
376 while (factors[i].count--)
378 if (temp * factors[i].factor < MAX_UNROLLED_INSNS)
380 unroll_number *= factors[i].factor;
381 temp *= factors[i].factor;
387 /* If we couldn't find any factors, then unroll as in the normal
389 if (unroll_number == 1)
391 if (loop_dump_stream)
392 fprintf (loop_dump_stream, "Loop unrolling: No factors found.\n");
395 unroll_type = UNROLL_MODULO;
398 /* Default case, calculate number of times to unroll loop based on its
400 if (unroll_type == UNROLL_NAIVE)
402 if (8 * insn_count < MAX_UNROLLED_INSNS)
404 else if (4 * insn_count < MAX_UNROLLED_INSNS)
410 /* Now we know how many times to unroll the loop. */
412 if (loop_dump_stream)
413 fprintf (loop_dump_stream, "Unrolling loop %d times.\n", unroll_number);
415 if (unroll_type == UNROLL_COMPLETELY || unroll_type == UNROLL_MODULO)
417 /* Loops of these types can start with jump down to the exit condition
418 in rare circumstances.
420 Consider a pair of nested loops where the inner loop is part
421 of the exit code for the outer loop.
423 In this case jump.c will not duplicate the exit test for the outer
424 loop, so it will start with a jump to the exit code.
426 Then consider if the inner loop turns out to iterate once and
427 only once. We will end up deleting the jumps associated with
428 the inner loop. However, the loop notes are not removed from
429 the instruction stream.
431 And finally assume that we can compute the number of iterations
434 In this case unroll may want to unroll the outer loop even though
435 it starts with a jump to the outer loop's exit code.
437 We could try to optimize this case, but it hardly seems worth it.
438 Just return without unrolling the loop in such cases. */
441 while (GET_CODE (insn) != CODE_LABEL && GET_CODE (insn) != JUMP_INSN)
442 insn = NEXT_INSN (insn);
443 if (GET_CODE (insn) == JUMP_INSN)
447 if (unroll_type == UNROLL_COMPLETELY)
449 /* Completely unrolling the loop: Delete the compare and branch at
450 the end (the last two instructions). This delete must done at the
451 very end of loop unrolling, to avoid problems with calls to
452 back_branch_in_range_p, which is called by find_splittable_regs.
453 All increments of splittable bivs/givs are changed to load constant
456 copy_start = loop_start;
458 /* Set insert_before to the instruction immediately after the JUMP_INSN
459 (or BARRIER), so that any NOTEs between the JUMP_INSN and the end of
460 the loop will be correctly handled by copy_loop_body. */
461 insert_before = NEXT_INSN (last_loop_insn);
463 /* Set copy_end to the insn before the jump at the end of the loop. */
464 if (GET_CODE (last_loop_insn) == BARRIER)
465 copy_end = PREV_INSN (PREV_INSN (last_loop_insn));
466 else if (GET_CODE (last_loop_insn) == JUMP_INSN)
468 copy_end = PREV_INSN (last_loop_insn);
470 /* The instruction immediately before the JUMP_INSN may be a compare
471 instruction which we do not want to copy. */
472 if (sets_cc0_p (PREV_INSN (copy_end)))
473 copy_end = PREV_INSN (copy_end);
478 /* We currently can't unroll a loop if it doesn't end with a
479 JUMP_INSN. There would need to be a mechanism that recognizes
480 this case, and then inserts a jump after each loop body, which
481 jumps to after the last loop body. */
482 if (loop_dump_stream)
483 fprintf (loop_dump_stream,
484 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
488 else if (unroll_type == UNROLL_MODULO)
490 /* Partially unrolling the loop: The compare and branch at the end
491 (the last two instructions) must remain. Don't copy the compare
492 and branch instructions at the end of the loop. Insert the unrolled
493 code immediately before the compare/branch at the end so that the
494 code will fall through to them as before. */
496 copy_start = loop_start;
498 /* Set insert_before to the jump insn at the end of the loop.
499 Set copy_end to before the jump insn at the end of the loop. */
500 if (GET_CODE (last_loop_insn) == BARRIER)
502 insert_before = PREV_INSN (last_loop_insn);
503 copy_end = PREV_INSN (insert_before);
505 else if (GET_CODE (last_loop_insn) == JUMP_INSN)
507 insert_before = last_loop_insn;
509 /* The instruction immediately before the JUMP_INSN may be a compare
510 instruction which we do not want to copy or delete. */
511 if (sets_cc0_p (PREV_INSN (insert_before)))
512 insert_before = PREV_INSN (insert_before);
514 copy_end = PREV_INSN (insert_before);
518 /* We currently can't unroll a loop if it doesn't end with a
519 JUMP_INSN. There would need to be a mechanism that recognizes
520 this case, and then inserts a jump after each loop body, which
521 jumps to after the last loop body. */
522 if (loop_dump_stream)
523 fprintf (loop_dump_stream,
524 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
530 /* Normal case: Must copy the compare and branch instructions at the
533 if (GET_CODE (last_loop_insn) == BARRIER)
535 /* Loop ends with an unconditional jump and a barrier.
536 Handle this like above, don't copy jump and barrier.
537 This is not strictly necessary, but doing so prevents generating
538 unconditional jumps to an immediately following label.
540 This will be corrected below if the target of this jump is
541 not the start_label. */
543 insert_before = PREV_INSN (last_loop_insn);
544 copy_end = PREV_INSN (insert_before);
546 else if (GET_CODE (last_loop_insn) == JUMP_INSN)
548 /* Set insert_before to immediately after the JUMP_INSN, so that
549 NOTEs at the end of the loop will be correctly handled by
551 insert_before = NEXT_INSN (last_loop_insn);
552 copy_end = last_loop_insn;
556 /* We currently can't unroll a loop if it doesn't end with a
557 JUMP_INSN. There would need to be a mechanism that recognizes
558 this case, and then inserts a jump after each loop body, which
559 jumps to after the last loop body. */
560 if (loop_dump_stream)
561 fprintf (loop_dump_stream,
562 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
566 /* If copying exit test branches because they can not be eliminated,
567 then must convert the fall through case of the branch to a jump past
568 the end of the loop. Create a label to emit after the loop and save
569 it for later use. Do not use the label after the loop, if any, since
570 it might be used by insns outside the loop, or there might be insns
571 added before it later by final_[bg]iv_value which must be after
572 the real exit label. */
573 exit_label = gen_label_rtx ();
576 while (GET_CODE (insn) != CODE_LABEL && GET_CODE (insn) != JUMP_INSN)
577 insn = NEXT_INSN (insn);
579 if (GET_CODE (insn) == JUMP_INSN)
581 /* The loop starts with a jump down to the exit condition test.
582 Start copying the loop after the barrier following this
584 copy_start = NEXT_INSN (insn);
586 /* Splitting induction variables doesn't work when the loop is
587 entered via a jump to the bottom, because then we end up doing
588 a comparison against a new register for a split variable, but
589 we did not execute the set insn for the new register because
590 it was skipped over. */
591 splitting_not_safe = 1;
592 if (loop_dump_stream)
593 fprintf (loop_dump_stream,
594 "Splitting not safe, because loop not entered at top.\n");
597 copy_start = loop_start;
600 /* This should always be the first label in the loop. */
601 start_label = NEXT_INSN (copy_start);
602 /* There may be a line number note and/or a loop continue note here. */
603 while (GET_CODE (start_label) == NOTE)
604 start_label = NEXT_INSN (start_label);
605 if (GET_CODE (start_label) != CODE_LABEL)
607 /* This can happen as a result of jump threading. If the first insns in
608 the loop test the same condition as the loop's backward jump, or the
609 opposite condition, then the backward jump will be modified to point
610 to elsewhere, and the loop's start label is deleted.
612 This case currently can not be handled by the loop unrolling code. */
614 if (loop_dump_stream)
615 fprintf (loop_dump_stream,
616 "Unrolling failure: unknown insns between BEG note and loop label.\n");
619 if (LABEL_NAME (start_label))
621 /* The jump optimization pass must have combined the original start label
622 with a named label for a goto. We can't unroll this case because
623 jumps which go to the named label must be handled differently than
624 jumps to the loop start, and it is impossible to differentiate them
626 if (loop_dump_stream)
627 fprintf (loop_dump_stream,
628 "Unrolling failure: loop start label is gone\n");
632 if (unroll_type == UNROLL_NAIVE
633 && GET_CODE (last_loop_insn) == BARRIER
634 && GET_CODE (PREV_INSN (last_loop_insn)) == JUMP_INSN
635 && start_label != JUMP_LABEL (PREV_INSN (last_loop_insn)))
637 /* In this case, we must copy the jump and barrier, because they will
638 not be converted to jumps to an immediately following label. */
640 insert_before = NEXT_INSN (last_loop_insn);
641 copy_end = last_loop_insn;
644 if (unroll_type == UNROLL_NAIVE
645 && GET_CODE (last_loop_insn) == JUMP_INSN
646 && start_label != JUMP_LABEL (last_loop_insn))
648 /* ??? The loop ends with a conditional branch that does not branch back
649 to the loop start label. In this case, we must emit an unconditional
650 branch to the loop exit after emitting the final branch.
651 copy_loop_body does not have support for this currently, so we
652 give up. It doesn't seem worthwhile to unroll anyways since
653 unrolling would increase the number of branch instructions
655 if (loop_dump_stream)
656 fprintf (loop_dump_stream,
657 "Unrolling failure: final conditional branch not to loop start\n");
661 /* Allocate a translation table for the labels and insn numbers.
662 They will be filled in as we copy the insns in the loop. */
664 max_labelno = max_label_num ();
665 max_insnno = get_max_uid ();
667 /* Various paths through the unroll code may reach the "egress" label
668 without initializing fields within the map structure.
670 To be safe, we use xcalloc to zero the memory. */
671 map = (struct inline_remap *) xcalloc (1, sizeof (struct inline_remap));
673 /* Allocate the label map. */
677 map->label_map = (rtx *) xcalloc (max_labelno, sizeof (rtx));
678 local_label = (char *) xcalloc (max_labelno, sizeof (char));
681 /* Search the loop and mark all local labels, i.e. the ones which have to
682 be distinct labels when copied. For all labels which might be
683 non-local, set their label_map entries to point to themselves.
684 If they happen to be local their label_map entries will be overwritten
685 before the loop body is copied. The label_map entries for local labels
686 will be set to a different value each time the loop body is copied. */
688 for (insn = copy_start; insn != loop_end; insn = NEXT_INSN (insn))
692 if (GET_CODE (insn) == CODE_LABEL)
693 local_label[CODE_LABEL_NUMBER (insn)] = 1;
694 else if (GET_CODE (insn) == JUMP_INSN)
696 if (JUMP_LABEL (insn))
697 set_label_in_map (map,
698 CODE_LABEL_NUMBER (JUMP_LABEL (insn)),
700 else if (GET_CODE (PATTERN (insn)) == ADDR_VEC
701 || GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC)
703 rtx pat = PATTERN (insn);
704 int diff_vec_p = GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC;
705 int len = XVECLEN (pat, diff_vec_p);
708 for (i = 0; i < len; i++)
710 label = XEXP (XVECEXP (pat, diff_vec_p, i), 0);
711 set_label_in_map (map, CODE_LABEL_NUMBER (label), label);
715 if ((note = find_reg_note (insn, REG_LABEL, NULL_RTX)))
716 set_label_in_map (map, CODE_LABEL_NUMBER (XEXP (note, 0)),
720 /* Allocate space for the insn map. */
722 map->insn_map = (rtx *) xmalloc (max_insnno * sizeof (rtx));
724 /* Set this to zero, to indicate that we are doing loop unrolling,
725 not function inlining. */
726 map->inline_target = 0;
728 /* The register and constant maps depend on the number of registers
729 present, so the final maps can't be created until after
730 find_splittable_regs is called. However, they are needed for
731 preconditioning, so we create temporary maps when preconditioning
734 /* The preconditioning code may allocate two new pseudo registers. */
735 maxregnum = max_reg_num ();
737 /* local_regno is only valid for regnos < max_local_regnum. */
738 max_local_regnum = maxregnum;
740 /* Allocate and zero out the splittable_regs and addr_combined_regs
741 arrays. These must be zeroed here because they will be used if
742 loop preconditioning is performed, and must be zero for that case.
744 It is safe to do this here, since the extra registers created by the
745 preconditioning code and find_splittable_regs will never be used
746 to access the splittable_regs[] and addr_combined_regs[] arrays. */
748 splittable_regs = (rtx *) xcalloc (maxregnum, sizeof (rtx));
749 splittable_regs_updates = (int *) xcalloc (maxregnum, sizeof (int));
751 = (struct induction **) xcalloc (maxregnum, sizeof (struct induction *));
752 local_regno = (char *) xcalloc (maxregnum, sizeof (char));
754 /* Mark all local registers, i.e. the ones which are referenced only
756 if (INSN_UID (copy_end) < max_uid_for_loop)
758 int copy_start_luid = INSN_LUID (copy_start);
759 int copy_end_luid = INSN_LUID (copy_end);
761 /* If a register is used in the jump insn, we must not duplicate it
762 since it will also be used outside the loop. */
763 if (GET_CODE (copy_end) == JUMP_INSN)
766 /* If we have a target that uses cc0, then we also must not duplicate
767 the insn that sets cc0 before the jump insn, if one is present. */
769 if (GET_CODE (copy_end) == JUMP_INSN
770 && sets_cc0_p (PREV_INSN (copy_end)))
774 /* If copy_start points to the NOTE that starts the loop, then we must
775 use the next luid, because invariant pseudo-regs moved out of the loop
776 have their lifetimes modified to start here, but they are not safe
778 if (copy_start == loop_start)
781 /* If a pseudo's lifetime is entirely contained within this loop, then we
782 can use a different pseudo in each unrolled copy of the loop. This
783 results in better code. */
784 /* We must limit the generic test to max_reg_before_loop, because only
785 these pseudo registers have valid regno_first_uid info. */
786 for (r = FIRST_PSEUDO_REGISTER; r < max_reg_before_loop; ++r)
787 if (REGNO_FIRST_UID (r) > 0 && REGNO_FIRST_UID (r) <= max_uid_for_loop
788 && REGNO_FIRST_LUID (r) >= copy_start_luid
789 && REGNO_LAST_UID (r) > 0 && REGNO_LAST_UID (r) <= max_uid_for_loop
790 && REGNO_LAST_LUID (r) <= copy_end_luid)
792 /* However, we must also check for loop-carried dependencies.
793 If the value the pseudo has at the end of iteration X is
794 used by iteration X+1, then we can not use a different pseudo
795 for each unrolled copy of the loop. */
796 /* A pseudo is safe if regno_first_uid is a set, and this
797 set dominates all instructions from regno_first_uid to
799 /* ??? This check is simplistic. We would get better code if
800 this check was more sophisticated. */
801 if (set_dominates_use (r, REGNO_FIRST_UID (r), REGNO_LAST_UID (r),
802 copy_start, copy_end))
805 if (loop_dump_stream)
808 fprintf (loop_dump_stream, "Marked reg %d as local\n", r);
810 fprintf (loop_dump_stream, "Did not mark reg %d as local\n",
816 /* If this loop requires exit tests when unrolled, check to see if we
817 can precondition the loop so as to make the exit tests unnecessary.
818 Just like variable splitting, this is not safe if the loop is entered
819 via a jump to the bottom. Also, can not do this if no strength
820 reduce info, because precondition_loop_p uses this info. */
822 /* Must copy the loop body for preconditioning before the following
823 find_splittable_regs call since that will emit insns which need to
824 be after the preconditioned loop copies, but immediately before the
825 unrolled loop copies. */
827 /* Also, it is not safe to split induction variables for the preconditioned
828 copies of the loop body. If we split induction variables, then the code
829 assumes that each induction variable can be represented as a function
830 of its initial value and the loop iteration number. This is not true
831 in this case, because the last preconditioned copy of the loop body
832 could be any iteration from the first up to the `unroll_number-1'th,
833 depending on the initial value of the iteration variable. Therefore
834 we can not split induction variables here, because we can not calculate
835 their value. Hence, this code must occur before find_splittable_regs
838 if (unroll_type == UNROLL_NAIVE && ! splitting_not_safe && strength_reduce_p)
840 rtx initial_value, final_value, increment;
841 enum machine_mode mode;
843 if (precondition_loop_p (loop,
844 &initial_value, &final_value, &increment,
849 int abs_inc, neg_inc;
850 enum rtx_code cc = loop_info->comparison_code;
851 int less_p = (cc == LE || cc == LEU || cc == LT || cc == LTU);
852 int unsigned_p = (cc == LEU || cc == GEU || cc == LTU || cc == GTU);
854 map->reg_map = (rtx *) xmalloc (maxregnum * sizeof (rtx));
856 VARRAY_CONST_EQUIV_INIT (map->const_equiv_varray, maxregnum,
857 "unroll_loop_precondition");
858 global_const_equiv_varray = map->const_equiv_varray;
860 init_reg_map (map, maxregnum);
862 /* Limit loop unrolling to 4, since this will make 7 copies of
864 if (unroll_number > 4)
867 /* Save the absolute value of the increment, and also whether or
868 not it is negative. */
870 abs_inc = INTVAL (increment);
879 /* Final value may have form of (PLUS val1 const1_rtx). We need
880 to convert it into general operand, so compute the real value. */
882 if (GET_CODE (final_value) == PLUS)
884 final_value = expand_simple_binop (mode, PLUS,
885 copy_rtx (XEXP (final_value, 0)),
886 copy_rtx (XEXP (final_value, 1)),
887 NULL_RTX, 0, OPTAB_LIB_WIDEN);
889 if (!nonmemory_operand (final_value, VOIDmode))
890 final_value = force_reg (mode, copy_rtx (final_value));
892 /* Calculate the difference between the final and initial values.
893 Final value may be a (plus (reg x) (const_int 1)) rtx.
894 Let the following cse pass simplify this if initial value is
897 We must copy the final and initial values here to avoid
898 improperly shared rtl.
900 We have to deal with for (i = 0; --i < 6;) type loops.
901 For such loops the real final value is the first time the
902 loop variable overflows, so the diff we calculate is the
903 distance from the overflow value. This is 0 or ~0 for
904 unsigned loops depending on the direction, or INT_MAX,
905 INT_MAX+1 for signed loops. We really do not need the
906 exact value, since we are only interested in the diff
907 modulo the increment, and the increment is a power of 2,
908 so we can pretend that the overflow value is 0/~0. */
910 if (cc == NE || less_p != neg_inc)
911 diff = expand_simple_binop (mode, MINUS, final_value,
912 copy_rtx (initial_value), NULL_RTX, 0,
915 diff = expand_simple_unop (mode, neg_inc ? NOT : NEG,
916 copy_rtx (initial_value), NULL_RTX, 0);
918 /* Now calculate (diff % (unroll * abs (increment))) by using an
920 diff = expand_simple_binop (GET_MODE (diff), AND, diff,
921 GEN_INT (unroll_number * abs_inc - 1),
922 NULL_RTX, 0, OPTAB_LIB_WIDEN);
924 /* Now emit a sequence of branches to jump to the proper precond
927 labels = (rtx *) xmalloc (sizeof (rtx) * unroll_number);
928 for (i = 0; i < unroll_number; i++)
929 labels[i] = gen_label_rtx ();
931 /* Check for the case where the initial value is greater than or
932 equal to the final value. In that case, we want to execute
933 exactly one loop iteration. The code below will fail for this
934 case. This check does not apply if the loop has a NE
935 comparison at the end. */
939 rtx incremented_initval;
940 incremented_initval = expand_simple_binop (mode, PLUS,
945 emit_cmp_and_jump_insns (incremented_initval, final_value,
946 less_p ? GE : LE, NULL_RTX,
947 mode, unsigned_p, labels[1]);
948 predict_insn_def (get_last_insn (), PRED_LOOP_CONDITION,
950 JUMP_LABEL (get_last_insn ()) = labels[1];
951 LABEL_NUSES (labels[1])++;
954 /* Assuming the unroll_number is 4, and the increment is 2, then
955 for a negative increment: for a positive increment:
956 diff = 0,1 precond 0 diff = 0,7 precond 0
957 diff = 2,3 precond 3 diff = 1,2 precond 1
958 diff = 4,5 precond 2 diff = 3,4 precond 2
959 diff = 6,7 precond 1 diff = 5,6 precond 3 */
961 /* We only need to emit (unroll_number - 1) branches here, the
962 last case just falls through to the following code. */
964 /* ??? This would give better code if we emitted a tree of branches
965 instead of the current linear list of branches. */
967 for (i = 0; i < unroll_number - 1; i++)
970 enum rtx_code cmp_code;
972 /* For negative increments, must invert the constant compared
973 against, except when comparing against zero. */
981 cmp_const = unroll_number - i;
990 emit_cmp_and_jump_insns (diff, GEN_INT (abs_inc * cmp_const),
991 cmp_code, NULL_RTX, mode, 0, labels[i]);
992 JUMP_LABEL (get_last_insn ()) = labels[i];
993 LABEL_NUSES (labels[i])++;
994 predict_insn (get_last_insn (), PRED_LOOP_PRECONDITIONING,
995 REG_BR_PROB_BASE / (unroll_number - i));
998 /* If the increment is greater than one, then we need another branch,
999 to handle other cases equivalent to 0. */
1001 /* ??? This should be merged into the code above somehow to help
1002 simplify the code here, and reduce the number of branches emitted.
1003 For the negative increment case, the branch here could easily
1004 be merged with the `0' case branch above. For the positive
1005 increment case, it is not clear how this can be simplified. */
1010 enum rtx_code cmp_code;
1014 cmp_const = abs_inc - 1;
1019 cmp_const = abs_inc * (unroll_number - 1) + 1;
1023 emit_cmp_and_jump_insns (diff, GEN_INT (cmp_const), cmp_code,
1024 NULL_RTX, mode, 0, labels[0]);
1025 JUMP_LABEL (get_last_insn ()) = labels[0];
1026 LABEL_NUSES (labels[0])++;
1029 sequence = gen_sequence ();
1031 loop_insn_hoist (loop, sequence);
1033 /* Only the last copy of the loop body here needs the exit
1034 test, so set copy_end to exclude the compare/branch here,
1035 and then reset it inside the loop when get to the last
1038 if (GET_CODE (last_loop_insn) == BARRIER)
1039 copy_end = PREV_INSN (PREV_INSN (last_loop_insn));
1040 else if (GET_CODE (last_loop_insn) == JUMP_INSN)
1042 copy_end = PREV_INSN (last_loop_insn);
1044 /* The immediately preceding insn may be a compare which
1045 we do not want to copy. */
1046 if (sets_cc0_p (PREV_INSN (copy_end)))
1047 copy_end = PREV_INSN (copy_end);
1053 for (i = 1; i < unroll_number; i++)
1055 emit_label_after (labels[unroll_number - i],
1056 PREV_INSN (loop_start));
1058 memset ((char *) map->insn_map, 0, max_insnno * sizeof (rtx));
1059 memset ((char *) &VARRAY_CONST_EQUIV (map->const_equiv_varray, 0),
1060 0, (VARRAY_SIZE (map->const_equiv_varray)
1061 * sizeof (struct const_equiv_data)));
1064 for (j = 0; j < max_labelno; j++)
1066 set_label_in_map (map, j, gen_label_rtx ());
1068 for (r = FIRST_PSEUDO_REGISTER; r < max_local_regnum; r++)
1072 = gen_reg_rtx (GET_MODE (regno_reg_rtx[r]));
1073 record_base_value (REGNO (map->reg_map[r]),
1074 regno_reg_rtx[r], 0);
1076 /* The last copy needs the compare/branch insns at the end,
1077 so reset copy_end here if the loop ends with a conditional
1080 if (i == unroll_number - 1)
1082 if (GET_CODE (last_loop_insn) == BARRIER)
1083 copy_end = PREV_INSN (PREV_INSN (last_loop_insn));
1085 copy_end = last_loop_insn;
1088 /* None of the copies are the `last_iteration', so just
1089 pass zero for that parameter. */
1090 copy_loop_body (loop, copy_start, copy_end, map, exit_label, 0,
1091 unroll_type, start_label, loop_end,
1092 loop_start, copy_end);
1094 emit_label_after (labels[0], PREV_INSN (loop_start));
1096 if (GET_CODE (last_loop_insn) == BARRIER)
1098 insert_before = PREV_INSN (last_loop_insn);
1099 copy_end = PREV_INSN (insert_before);
1103 insert_before = last_loop_insn;
1105 /* The instruction immediately before the JUMP_INSN may
1106 be a compare instruction which we do not want to copy
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,
1129 "Unrolling failure: Naive unrolling not being done.\n");
1133 /* At this point, we are guaranteed to unroll the loop. */
1135 /* Keep track of the unroll factor for the loop. */
1136 loop_info->unroll_number = unroll_number;
1138 /* For each biv and giv, determine whether it can be safely split into
1139 a different variable for each unrolled copy of the loop body.
1140 We precalculate and save this info here, since computing it is
1143 Do this before deleting any instructions from the loop, so that
1144 back_branch_in_range_p will work correctly. */
1146 if (splitting_not_safe)
1149 temp = find_splittable_regs (loop, unroll_type, unroll_number);
1151 /* find_splittable_regs may have created some new registers, so must
1152 reallocate the reg_map with the new larger size, and must realloc
1153 the constant maps also. */
1155 maxregnum = max_reg_num ();
1156 map->reg_map = (rtx *) xmalloc (maxregnum * sizeof (rtx));
1158 init_reg_map (map, maxregnum);
1160 if (map->const_equiv_varray == 0)
1161 VARRAY_CONST_EQUIV_INIT (map->const_equiv_varray,
1162 maxregnum + temp * unroll_number * 2,
1164 global_const_equiv_varray = map->const_equiv_varray;
1166 /* Search the list of bivs and givs to find ones which need to be remapped
1167 when split, and set their reg_map entry appropriately. */
1169 for (bl = ivs->list; bl; bl = bl->next)
1171 if (REGNO (bl->biv->src_reg) != bl->regno)
1172 map->reg_map[bl->regno] = bl->biv->src_reg;
1174 /* Currently, non-reduced/final-value givs are never split. */
1175 for (v = bl->giv; v; v = v->next_iv)
1176 if (REGNO (v->src_reg) != bl->regno)
1177 map->reg_map[REGNO (v->dest_reg)] = v->src_reg;
1181 /* Use our current register alignment and pointer flags. */
1182 map->regno_pointer_align = cfun->emit->regno_pointer_align;
1183 map->x_regno_reg_rtx = cfun->emit->x_regno_reg_rtx;
1185 /* If the loop is being partially unrolled, and the iteration variables
1186 are being split, and are being renamed for the split, then must fix up
1187 the compare/jump instruction at the end of the loop to refer to the new
1188 registers. This compare isn't copied, so the registers used in it
1189 will never be replaced if it isn't done here. */
1191 if (unroll_type == UNROLL_MODULO)
1193 insn = NEXT_INSN (copy_end);
1194 if (GET_CODE (insn) == INSN || GET_CODE (insn) == JUMP_INSN)
1195 PATTERN (insn) = remap_split_bivs (loop, PATTERN (insn));
1198 /* For unroll_number times, make a copy of each instruction
1199 between copy_start and copy_end, and insert these new instructions
1200 before the end of the loop. */
1202 for (i = 0; i < unroll_number; i++)
1204 memset ((char *) map->insn_map, 0, max_insnno * sizeof (rtx));
1205 memset ((char *) &VARRAY_CONST_EQUIV (map->const_equiv_varray, 0), 0,
1206 VARRAY_SIZE (map->const_equiv_varray) * sizeof (struct const_equiv_data));
1209 for (j = 0; j < max_labelno; j++)
1211 set_label_in_map (map, j, gen_label_rtx ());
1213 for (r = FIRST_PSEUDO_REGISTER; r < max_local_regnum; r++)
1216 map->reg_map[r] = gen_reg_rtx (GET_MODE (regno_reg_rtx[r]));
1217 record_base_value (REGNO (map->reg_map[r]),
1218 regno_reg_rtx[r], 0);
1221 /* If loop starts with a branch to the test, then fix it so that
1222 it points to the test of the first unrolled copy of the loop. */
1223 if (i == 0 && loop_start != copy_start)
1225 insn = PREV_INSN (copy_start);
1226 pattern = PATTERN (insn);
1228 tem = get_label_from_map (map,
1230 (XEXP (SET_SRC (pattern), 0)));
1231 SET_SRC (pattern) = gen_rtx_LABEL_REF (VOIDmode, tem);
1233 /* Set the jump label so that it can be used by later loop unrolling
1235 JUMP_LABEL (insn) = tem;
1236 LABEL_NUSES (tem)++;
1239 copy_loop_body (loop, copy_start, copy_end, map, exit_label,
1240 i == unroll_number - 1, unroll_type, start_label,
1241 loop_end, insert_before, insert_before);
1244 /* Before deleting any insns, emit a CODE_LABEL immediately after the last
1245 insn to be deleted. This prevents any runaway delete_insn call from
1246 more insns that it should, as it always stops at a CODE_LABEL. */
1248 /* Delete the compare and branch at the end of the loop if completely
1249 unrolling the loop. Deleting the backward branch at the end also
1250 deletes the code label at the start of the loop. This is done at
1251 the very end to avoid problems with back_branch_in_range_p. */
1253 if (unroll_type == UNROLL_COMPLETELY)
1254 safety_label = emit_label_after (gen_label_rtx (), last_loop_insn);
1256 safety_label = emit_label_after (gen_label_rtx (), copy_end);
1258 /* Delete all of the original loop instructions. Don't delete the
1259 LOOP_BEG note, or the first code label in the loop. */
1261 insn = NEXT_INSN (copy_start);
1262 while (insn != safety_label)
1264 /* ??? Don't delete named code labels. They will be deleted when the
1265 jump that references them is deleted. Otherwise, we end up deleting
1266 them twice, which causes them to completely disappear instead of turn
1267 into NOTE_INSN_DELETED_LABEL notes. This in turn causes aborts in
1268 dwarfout.c/dwarf2out.c. We could perhaps fix the dwarf*out.c files
1269 to handle deleted labels instead. Or perhaps fix DECL_RTL of the
1270 associated LABEL_DECL to point to one of the new label instances. */
1271 /* ??? Likewise, we can't delete a NOTE_INSN_DELETED_LABEL note. */
1272 if (insn != start_label
1273 && ! (GET_CODE (insn) == CODE_LABEL && LABEL_NAME (insn))
1274 && ! (GET_CODE (insn) == NOTE
1275 && NOTE_LINE_NUMBER (insn) == NOTE_INSN_DELETED_LABEL))
1276 insn = delete_related_insns (insn);
1278 insn = NEXT_INSN (insn);
1281 /* Can now delete the 'safety' label emitted to protect us from runaway
1282 delete_related_insns calls. */
1283 if (INSN_DELETED_P (safety_label))
1285 delete_related_insns (safety_label);
1287 /* If exit_label exists, emit it after the loop. Doing the emit here
1288 forces it to have a higher INSN_UID than any insn in the unrolled loop.
1289 This is needed so that mostly_true_jump in reorg.c will treat jumps
1290 to this loop end label correctly, i.e. predict that they are usually
1293 emit_label_after (exit_label, loop_end);
1296 if (unroll_type == UNROLL_COMPLETELY)
1298 /* Remove the loop notes since this is no longer a loop. */
1300 delete_related_insns (loop->vtop);
1302 delete_related_insns (loop->cont);
1304 delete_related_insns (loop_start);
1306 delete_related_insns (loop_end);
1309 if (map->const_equiv_varray)
1310 VARRAY_FREE (map->const_equiv_varray);
1313 free (map->label_map);
1316 free (map->insn_map);
1317 free (splittable_regs);
1318 free (splittable_regs_updates);
1319 free (addr_combined_regs);
1322 free (map->reg_map);
1326 /* Return true if the loop can be safely, and profitably, preconditioned
1327 so that the unrolled copies of the loop body don't need exit tests.
1329 This only works if final_value, initial_value and increment can be
1330 determined, and if increment is a constant power of 2.
1331 If increment is not a power of 2, then the preconditioning modulo
1332 operation would require a real modulo instead of a boolean AND, and this
1333 is not considered `profitable'. */
1335 /* ??? If the loop is known to be executed very many times, or the machine
1336 has a very cheap divide instruction, then preconditioning is a win even
1337 when the increment is not a power of 2. Use RTX_COST to compute
1338 whether divide is cheap.
1339 ??? A divide by constant doesn't actually need a divide, look at
1340 expand_divmod. The reduced cost of this optimized modulo is not
1341 reflected in RTX_COST. */
1344 precondition_loop_p (loop, initial_value, final_value, increment, mode)
1345 const struct loop *loop;
1346 rtx *initial_value, *final_value, *increment;
1347 enum machine_mode *mode;
1349 rtx loop_start = loop->start;
1350 struct loop_info *loop_info = LOOP_INFO (loop);
1352 if (loop_info->n_iterations > 0)
1354 if (INTVAL (loop_info->increment) > 0)
1356 *initial_value = const0_rtx;
1357 *increment = const1_rtx;
1358 *final_value = GEN_INT (loop_info->n_iterations);
1362 *initial_value = GEN_INT (loop_info->n_iterations);
1363 *increment = constm1_rtx;
1364 *final_value = const0_rtx;
1368 if (loop_dump_stream)
1370 fputs ("Preconditioning: Success, number of iterations known, ",
1372 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC,
1373 loop_info->n_iterations);
1374 fputs (".\n", loop_dump_stream);
1379 if (loop_info->iteration_var == 0)
1381 if (loop_dump_stream)
1382 fprintf (loop_dump_stream,
1383 "Preconditioning: Could not find iteration variable.\n");
1386 else if (loop_info->initial_value == 0)
1388 if (loop_dump_stream)
1389 fprintf (loop_dump_stream,
1390 "Preconditioning: Could not find initial value.\n");
1393 else if (loop_info->increment == 0)
1395 if (loop_dump_stream)
1396 fprintf (loop_dump_stream,
1397 "Preconditioning: Could not find increment value.\n");
1400 else if (GET_CODE (loop_info->increment) != CONST_INT)
1402 if (loop_dump_stream)
1403 fprintf (loop_dump_stream,
1404 "Preconditioning: Increment not a constant.\n");
1407 else if ((exact_log2 (INTVAL (loop_info->increment)) < 0)
1408 && (exact_log2 (-INTVAL (loop_info->increment)) < 0))
1410 if (loop_dump_stream)
1411 fprintf (loop_dump_stream,
1412 "Preconditioning: Increment not a constant power of 2.\n");
1416 /* Unsigned_compare and compare_dir can be ignored here, since they do
1417 not matter for preconditioning. */
1419 if (loop_info->final_value == 0)
1421 if (loop_dump_stream)
1422 fprintf (loop_dump_stream,
1423 "Preconditioning: EQ comparison loop.\n");
1427 /* Must ensure that final_value is invariant, so call
1428 loop_invariant_p to check. Before doing so, must check regno
1429 against max_reg_before_loop to make sure that the register is in
1430 the range covered by loop_invariant_p. If it isn't, then it is
1431 most likely a biv/giv which by definition are not invariant. */
1432 if ((GET_CODE (loop_info->final_value) == REG
1433 && REGNO (loop_info->final_value) >= max_reg_before_loop)
1434 || (GET_CODE (loop_info->final_value) == PLUS
1435 && REGNO (XEXP (loop_info->final_value, 0)) >= max_reg_before_loop)
1436 || ! loop_invariant_p (loop, loop_info->final_value))
1438 if (loop_dump_stream)
1439 fprintf (loop_dump_stream,
1440 "Preconditioning: Final value not invariant.\n");
1444 /* Fail for floating point values, since the caller of this function
1445 does not have code to deal with them. */
1446 if (GET_MODE_CLASS (GET_MODE (loop_info->final_value)) == MODE_FLOAT
1447 || GET_MODE_CLASS (GET_MODE (loop_info->initial_value)) == MODE_FLOAT)
1449 if (loop_dump_stream)
1450 fprintf (loop_dump_stream,
1451 "Preconditioning: Floating point final or initial value.\n");
1455 /* Fail if loop_info->iteration_var is not live before loop_start,
1456 since we need to test its value in the preconditioning code. */
1458 if (REGNO_FIRST_LUID (REGNO (loop_info->iteration_var))
1459 > INSN_LUID (loop_start))
1461 if (loop_dump_stream)
1462 fprintf (loop_dump_stream,
1463 "Preconditioning: Iteration var not live before loop start.\n");
1467 /* Note that loop_iterations biases the initial value for GIV iterators
1468 such as "while (i-- > 0)" so that we can calculate the number of
1469 iterations just like for BIV iterators.
1471 Also note that the absolute values of initial_value and
1472 final_value are unimportant as only their difference is used for
1473 calculating the number of loop iterations. */
1474 *initial_value = loop_info->initial_value;
1475 *increment = loop_info->increment;
1476 *final_value = loop_info->final_value;
1478 /* Decide what mode to do these calculations in. Choose the larger
1479 of final_value's mode and initial_value's mode, or a full-word if
1480 both are constants. */
1481 *mode = GET_MODE (*final_value);
1482 if (*mode == VOIDmode)
1484 *mode = GET_MODE (*initial_value);
1485 if (*mode == VOIDmode)
1488 else if (*mode != GET_MODE (*initial_value)
1489 && (GET_MODE_SIZE (*mode)
1490 < GET_MODE_SIZE (GET_MODE (*initial_value))))
1491 *mode = GET_MODE (*initial_value);
1494 if (loop_dump_stream)
1495 fprintf (loop_dump_stream, "Preconditioning: Successful.\n");
1499 /* All pseudo-registers must be mapped to themselves. Two hard registers
1500 must be mapped, VIRTUAL_STACK_VARS_REGNUM and VIRTUAL_INCOMING_ARGS_
1501 REGNUM, to avoid function-inlining specific conversions of these
1502 registers. All other hard regs can not be mapped because they may be
1507 init_reg_map (map, maxregnum)
1508 struct inline_remap *map;
1513 for (i = maxregnum - 1; i > LAST_VIRTUAL_REGISTER; i--)
1514 map->reg_map[i] = regno_reg_rtx[i];
1515 /* Just clear the rest of the entries. */
1516 for (i = LAST_VIRTUAL_REGISTER; i >= 0; i--)
1517 map->reg_map[i] = 0;
1519 map->reg_map[VIRTUAL_STACK_VARS_REGNUM]
1520 = regno_reg_rtx[VIRTUAL_STACK_VARS_REGNUM];
1521 map->reg_map[VIRTUAL_INCOMING_ARGS_REGNUM]
1522 = regno_reg_rtx[VIRTUAL_INCOMING_ARGS_REGNUM];
1525 /* Strength-reduction will often emit code for optimized biv/givs which
1526 calculates their value in a temporary register, and then copies the result
1527 to the iv. This procedure reconstructs the pattern computing the iv;
1528 verifying that all operands are of the proper form.
1530 PATTERN must be the result of single_set.
1531 The return value is the amount that the giv is incremented by. */
1534 calculate_giv_inc (pattern, src_insn, regno)
1535 rtx pattern, src_insn;
1539 rtx increment_total = 0;
1543 /* Verify that we have an increment insn here. First check for a plus
1544 as the set source. */
1545 if (GET_CODE (SET_SRC (pattern)) != PLUS)
1547 /* SR sometimes computes the new giv value in a temp, then copies it
1549 src_insn = PREV_INSN (src_insn);
1550 pattern = single_set (src_insn);
1551 if (GET_CODE (SET_SRC (pattern)) != PLUS)
1554 /* The last insn emitted is not needed, so delete it to avoid confusing
1555 the second cse pass. This insn sets the giv unnecessarily. */
1556 delete_related_insns (get_last_insn ());
1559 /* Verify that we have a constant as the second operand of the plus. */
1560 increment = XEXP (SET_SRC (pattern), 1);
1561 if (GET_CODE (increment) != CONST_INT)
1563 /* SR sometimes puts the constant in a register, especially if it is
1564 too big to be an add immed operand. */
1565 increment = find_last_value (increment, &src_insn, NULL_RTX, 0);
1567 /* SR may have used LO_SUM to compute the constant if it is too large
1568 for a load immed operand. In this case, the constant is in operand
1569 one of the LO_SUM rtx. */
1570 if (GET_CODE (increment) == LO_SUM)
1571 increment = XEXP (increment, 1);
1573 /* Some ports store large constants in memory and add a REG_EQUAL
1574 note to the store insn. */
1575 else if (GET_CODE (increment) == MEM)
1577 rtx note = find_reg_note (src_insn, REG_EQUAL, 0);
1579 increment = XEXP (note, 0);
1582 else if (GET_CODE (increment) == IOR
1583 || GET_CODE (increment) == ASHIFT
1584 || GET_CODE (increment) == PLUS)
1586 /* The rs6000 port loads some constants with IOR.
1587 The alpha port loads some constants with ASHIFT and PLUS. */
1588 rtx second_part = XEXP (increment, 1);
1589 enum rtx_code code = GET_CODE (increment);
1591 increment = find_last_value (XEXP (increment, 0),
1592 &src_insn, NULL_RTX, 0);
1593 /* Don't need the last insn anymore. */
1594 delete_related_insns (get_last_insn ());
1596 if (GET_CODE (second_part) != CONST_INT
1597 || GET_CODE (increment) != CONST_INT)
1601 increment = GEN_INT (INTVAL (increment) | INTVAL (second_part));
1602 else if (code == PLUS)
1603 increment = GEN_INT (INTVAL (increment) + INTVAL (second_part));
1605 increment = GEN_INT (INTVAL (increment) << INTVAL (second_part));
1608 if (GET_CODE (increment) != CONST_INT)
1611 /* The insn loading the constant into a register is no longer needed,
1613 delete_related_insns (get_last_insn ());
1616 if (increment_total)
1617 increment_total = GEN_INT (INTVAL (increment_total) + INTVAL (increment));
1619 increment_total = increment;
1621 /* Check that the source register is the same as the register we expected
1622 to see as the source. If not, something is seriously wrong. */
1623 if (GET_CODE (XEXP (SET_SRC (pattern), 0)) != REG
1624 || REGNO (XEXP (SET_SRC (pattern), 0)) != regno)
1626 /* Some machines (e.g. the romp), may emit two add instructions for
1627 certain constants, so lets try looking for another add immediately
1628 before this one if we have only seen one add insn so far. */
1634 src_insn = PREV_INSN (src_insn);
1635 pattern = single_set (src_insn);
1637 delete_related_insns (get_last_insn ());
1645 return increment_total;
1648 /* Copy REG_NOTES, except for insn references, because not all insn_map
1649 entries are valid yet. We do need to copy registers now though, because
1650 the reg_map entries can change during copying. */
1653 initial_reg_note_copy (notes, map)
1655 struct inline_remap *map;
1662 copy = rtx_alloc (GET_CODE (notes));
1663 PUT_REG_NOTE_KIND (copy, REG_NOTE_KIND (notes));
1665 if (GET_CODE (notes) == EXPR_LIST)
1666 XEXP (copy, 0) = copy_rtx_and_substitute (XEXP (notes, 0), map, 0);
1667 else if (GET_CODE (notes) == INSN_LIST)
1668 /* Don't substitute for these yet. */
1669 XEXP (copy, 0) = copy_rtx (XEXP (notes, 0));
1673 XEXP (copy, 1) = initial_reg_note_copy (XEXP (notes, 1), map);
1678 /* Fixup insn references in copied REG_NOTES. */
1681 final_reg_note_copy (notesp, map)
1683 struct inline_remap *map;
1689 if (GET_CODE (note) == INSN_LIST)
1691 /* Sometimes, we have a REG_WAS_0 note that points to a
1692 deleted instruction. In that case, we can just delete the
1694 if (REG_NOTE_KIND (note) == REG_WAS_0)
1696 *notesp = XEXP (note, 1);
1701 rtx insn = map->insn_map[INSN_UID (XEXP (note, 0))];
1703 /* If we failed to remap the note, something is awry.
1704 Allow REG_LABEL as it may reference label outside
1705 the unrolled loop. */
1708 if (REG_NOTE_KIND (note) != REG_LABEL)
1712 XEXP (note, 0) = insn;
1716 notesp = &XEXP (note, 1);
1720 /* Copy each instruction in the loop, substituting from map as appropriate.
1721 This is very similar to a loop in expand_inline_function. */
1724 copy_loop_body (loop, copy_start, copy_end, map, exit_label, last_iteration,
1725 unroll_type, start_label, loop_end, insert_before,
1728 rtx copy_start, copy_end;
1729 struct inline_remap *map;
1732 enum unroll_types unroll_type;
1733 rtx start_label, loop_end, insert_before, copy_notes_from;
1735 struct loop_ivs *ivs = LOOP_IVS (loop);
1737 rtx set, tem, copy = NULL_RTX;
1738 int dest_reg_was_split, i;
1742 rtx final_label = 0;
1743 rtx giv_inc, giv_dest_reg, giv_src_reg;
1745 /* If this isn't the last iteration, then map any references to the
1746 start_label to final_label. Final label will then be emitted immediately
1747 after the end of this loop body if it was ever used.
1749 If this is the last iteration, then map references to the start_label
1751 if (! last_iteration)
1753 final_label = gen_label_rtx ();
1754 set_label_in_map (map, CODE_LABEL_NUMBER (start_label), final_label);
1757 set_label_in_map (map, CODE_LABEL_NUMBER (start_label), start_label);
1761 /* Emit a NOTE_INSN_DELETED to force at least two insns onto the sequence.
1762 Else gen_sequence could return a raw pattern for a jump which we pass
1763 off to emit_insn_before (instead of emit_jump_insn_before) which causes
1764 a variety of losing behaviors later. */
1765 emit_note (0, NOTE_INSN_DELETED);
1770 insn = NEXT_INSN (insn);
1772 map->orig_asm_operands_vector = 0;
1774 switch (GET_CODE (insn))
1777 pattern = PATTERN (insn);
1781 /* Check to see if this is a giv that has been combined with
1782 some split address givs. (Combined in the sense that
1783 `combine_givs' in loop.c has put two givs in the same register.)
1784 In this case, we must search all givs based on the same biv to
1785 find the address givs. Then split the address givs.
1786 Do this before splitting the giv, since that may map the
1787 SET_DEST to a new register. */
1789 if ((set = single_set (insn))
1790 && GET_CODE (SET_DEST (set)) == REG
1791 && addr_combined_regs[REGNO (SET_DEST (set))])
1793 struct iv_class *bl;
1794 struct induction *v, *tv;
1795 unsigned int regno = REGNO (SET_DEST (set));
1797 v = addr_combined_regs[REGNO (SET_DEST (set))];
1798 bl = REG_IV_CLASS (ivs, REGNO (v->src_reg));
1800 /* Although the giv_inc amount is not needed here, we must call
1801 calculate_giv_inc here since it might try to delete the
1802 last insn emitted. If we wait until later to call it,
1803 we might accidentally delete insns generated immediately
1804 below by emit_unrolled_add. */
1806 giv_inc = calculate_giv_inc (set, insn, regno);
1808 /* Now find all address giv's that were combined with this
1810 for (tv = bl->giv; tv; tv = tv->next_iv)
1811 if (tv->giv_type == DEST_ADDR && tv->same == v)
1815 /* If this DEST_ADDR giv was not split, then ignore it. */
1816 if (*tv->location != tv->dest_reg)
1819 /* Scale this_giv_inc if the multiplicative factors of
1820 the two givs are different. */
1821 this_giv_inc = INTVAL (giv_inc);
1822 if (tv->mult_val != v->mult_val)
1823 this_giv_inc = (this_giv_inc / INTVAL (v->mult_val)
1824 * INTVAL (tv->mult_val));
1826 tv->dest_reg = plus_constant (tv->dest_reg, this_giv_inc);
1827 *tv->location = tv->dest_reg;
1829 if (last_iteration && unroll_type != UNROLL_COMPLETELY)
1831 /* Must emit an insn to increment the split address
1832 giv. Add in the const_adjust field in case there
1833 was a constant eliminated from the address. */
1834 rtx value, dest_reg;
1836 /* tv->dest_reg will be either a bare register,
1837 or else a register plus a constant. */
1838 if (GET_CODE (tv->dest_reg) == REG)
1839 dest_reg = tv->dest_reg;
1841 dest_reg = XEXP (tv->dest_reg, 0);
1843 /* Check for shared address givs, and avoid
1844 incrementing the shared pseudo reg more than
1846 if (! tv->same_insn && ! tv->shared)
1848 /* tv->dest_reg may actually be a (PLUS (REG)
1849 (CONST)) here, so we must call plus_constant
1850 to add the const_adjust amount before calling
1851 emit_unrolled_add below. */
1852 value = plus_constant (tv->dest_reg,
1855 if (GET_CODE (value) == PLUS)
1857 /* The constant could be too large for an add
1858 immediate, so can't directly emit an insn
1860 emit_unrolled_add (dest_reg, XEXP (value, 0),
1865 /* Reset the giv to be just the register again, in case
1866 it is used after the set we have just emitted.
1867 We must subtract the const_adjust factor added in
1869 tv->dest_reg = plus_constant (dest_reg,
1871 *tv->location = tv->dest_reg;
1876 /* If this is a setting of a splittable variable, then determine
1877 how to split the variable, create a new set based on this split,
1878 and set up the reg_map so that later uses of the variable will
1879 use the new split variable. */
1881 dest_reg_was_split = 0;
1883 if ((set = single_set (insn))
1884 && GET_CODE (SET_DEST (set)) == REG
1885 && splittable_regs[REGNO (SET_DEST (set))])
1887 unsigned int regno = REGNO (SET_DEST (set));
1888 unsigned int src_regno;
1890 dest_reg_was_split = 1;
1892 giv_dest_reg = SET_DEST (set);
1893 giv_src_reg = giv_dest_reg;
1894 /* Compute the increment value for the giv, if it wasn't
1895 already computed above. */
1897 giv_inc = calculate_giv_inc (set, insn, regno);
1899 src_regno = REGNO (giv_src_reg);
1901 if (unroll_type == UNROLL_COMPLETELY)
1903 /* Completely unrolling the loop. Set the induction
1904 variable to a known constant value. */
1906 /* The value in splittable_regs may be an invariant
1907 value, so we must use plus_constant here. */
1908 splittable_regs[regno]
1909 = plus_constant (splittable_regs[src_regno],
1912 if (GET_CODE (splittable_regs[regno]) == PLUS)
1914 giv_src_reg = XEXP (splittable_regs[regno], 0);
1915 giv_inc = XEXP (splittable_regs[regno], 1);
1919 /* The splittable_regs value must be a REG or a
1920 CONST_INT, so put the entire value in the giv_src_reg
1922 giv_src_reg = splittable_regs[regno];
1923 giv_inc = const0_rtx;
1928 /* Partially unrolling loop. Create a new pseudo
1929 register for the iteration variable, and set it to
1930 be a constant plus the original register. Except
1931 on the last iteration, when the result has to
1932 go back into the original iteration var register. */
1934 /* Handle bivs which must be mapped to a new register
1935 when split. This happens for bivs which need their
1936 final value set before loop entry. The new register
1937 for the biv was stored in the biv's first struct
1938 induction entry by find_splittable_regs. */
1940 if (regno < ivs->n_regs
1941 && REG_IV_TYPE (ivs, regno) == BASIC_INDUCT)
1943 giv_src_reg = REG_IV_CLASS (ivs, regno)->biv->src_reg;
1944 giv_dest_reg = giv_src_reg;
1948 /* If non-reduced/final-value givs were split, then
1949 this would have to remap those givs also. See
1950 find_splittable_regs. */
1953 splittable_regs[regno]
1954 = simplify_gen_binary (PLUS, GET_MODE (giv_src_reg),
1956 splittable_regs[src_regno]);
1957 giv_inc = splittable_regs[regno];
1959 /* Now split the induction variable by changing the dest
1960 of this insn to a new register, and setting its
1961 reg_map entry to point to this new register.
1963 If this is the last iteration, and this is the last insn
1964 that will update the iv, then reuse the original dest,
1965 to ensure that the iv will have the proper value when
1966 the loop exits or repeats.
1968 Using splittable_regs_updates here like this is safe,
1969 because it can only be greater than one if all
1970 instructions modifying the iv are always executed in
1973 if (! last_iteration
1974 || (splittable_regs_updates[regno]-- != 1))
1976 tem = gen_reg_rtx (GET_MODE (giv_src_reg));
1978 map->reg_map[regno] = tem;
1979 record_base_value (REGNO (tem),
1980 giv_inc == const0_rtx
1982 : gen_rtx_PLUS (GET_MODE (giv_src_reg),
1983 giv_src_reg, giv_inc),
1987 map->reg_map[regno] = giv_src_reg;
1990 /* The constant being added could be too large for an add
1991 immediate, so can't directly emit an insn here. */
1992 emit_unrolled_add (giv_dest_reg, giv_src_reg, giv_inc);
1993 copy = get_last_insn ();
1994 pattern = PATTERN (copy);
1998 pattern = copy_rtx_and_substitute (pattern, map, 0);
1999 copy = emit_insn (pattern);
2001 REG_NOTES (copy) = initial_reg_note_copy (REG_NOTES (insn), map);
2002 INSN_SCOPE (copy) = INSN_SCOPE (insn);
2005 /* If this insn is setting CC0, it may need to look at
2006 the insn that uses CC0 to see what type of insn it is.
2007 In that case, the call to recog via validate_change will
2008 fail. So don't substitute constants here. Instead,
2009 do it when we emit the following insn.
2011 For example, see the pyr.md file. That machine has signed and
2012 unsigned compares. The compare patterns must check the
2013 following branch insn to see which what kind of compare to
2016 If the previous insn set CC0, substitute constants on it as
2018 if (sets_cc0_p (PATTERN (copy)) != 0)
2023 try_constants (cc0_insn, map);
2025 try_constants (copy, map);
2028 try_constants (copy, map);
2031 /* Make split induction variable constants `permanent' since we
2032 know there are no backward branches across iteration variable
2033 settings which would invalidate this. */
2034 if (dest_reg_was_split)
2036 int regno = REGNO (SET_DEST (set));
2038 if ((size_t) regno < VARRAY_SIZE (map->const_equiv_varray)
2039 && (VARRAY_CONST_EQUIV (map->const_equiv_varray, regno).age
2041 VARRAY_CONST_EQUIV (map->const_equiv_varray, regno).age = -1;
2046 pattern = copy_rtx_and_substitute (PATTERN (insn), map, 0);
2047 copy = emit_jump_insn (pattern);
2048 REG_NOTES (copy) = initial_reg_note_copy (REG_NOTES (insn), map);
2049 INSN_SCOPE (copy) = INSN_SCOPE (insn);
2051 if (JUMP_LABEL (insn))
2053 JUMP_LABEL (copy) = get_label_from_map (map,
2055 (JUMP_LABEL (insn)));
2056 LABEL_NUSES (JUMP_LABEL (copy))++;
2058 if (JUMP_LABEL (insn) == start_label && insn == copy_end
2059 && ! last_iteration)
2062 /* This is a branch to the beginning of the loop; this is the
2063 last insn being copied; and this is not the last iteration.
2064 In this case, we want to change the original fall through
2065 case to be a branch past the end of the loop, and the
2066 original jump label case to fall_through. */
2068 if (!invert_jump (copy, exit_label, 0))
2071 rtx lab = gen_label_rtx ();
2072 /* Can't do it by reversing the jump (probably because we
2073 couldn't reverse the conditions), so emit a new
2074 jump_insn after COPY, and redirect the jump around
2076 jmp = emit_jump_insn_after (gen_jump (exit_label), copy);
2077 JUMP_LABEL (jmp) = exit_label;
2078 LABEL_NUSES (exit_label)++;
2079 jmp = emit_barrier_after (jmp);
2080 emit_label_after (lab, jmp);
2081 LABEL_NUSES (lab) = 0;
2082 if (!redirect_jump (copy, lab, 0))
2089 try_constants (cc0_insn, map);
2092 try_constants (copy, map);
2094 /* Set the jump label of COPY correctly to avoid problems with
2095 later passes of unroll_loop, if INSN had jump label set. */
2096 if (JUMP_LABEL (insn))
2100 /* Can't use the label_map for every insn, since this may be
2101 the backward branch, and hence the label was not mapped. */
2102 if ((set = single_set (copy)))
2104 tem = SET_SRC (set);
2105 if (GET_CODE (tem) == LABEL_REF)
2106 label = XEXP (tem, 0);
2107 else if (GET_CODE (tem) == IF_THEN_ELSE)
2109 if (XEXP (tem, 1) != pc_rtx)
2110 label = XEXP (XEXP (tem, 1), 0);
2112 label = XEXP (XEXP (tem, 2), 0);
2116 if (label && GET_CODE (label) == CODE_LABEL)
2117 JUMP_LABEL (copy) = label;
2120 /* An unrecognizable jump insn, probably the entry jump
2121 for a switch statement. This label must have been mapped,
2122 so just use the label_map to get the new jump label. */
2124 = get_label_from_map (map,
2125 CODE_LABEL_NUMBER (JUMP_LABEL (insn)));
2128 /* If this is a non-local jump, then must increase the label
2129 use count so that the label will not be deleted when the
2130 original jump is deleted. */
2131 LABEL_NUSES (JUMP_LABEL (copy))++;
2133 else if (GET_CODE (PATTERN (copy)) == ADDR_VEC
2134 || GET_CODE (PATTERN (copy)) == ADDR_DIFF_VEC)
2136 rtx pat = PATTERN (copy);
2137 int diff_vec_p = GET_CODE (pat) == ADDR_DIFF_VEC;
2138 int len = XVECLEN (pat, diff_vec_p);
2141 for (i = 0; i < len; i++)
2142 LABEL_NUSES (XEXP (XVECEXP (pat, diff_vec_p, i), 0))++;
2145 /* If this used to be a conditional jump insn but whose branch
2146 direction is now known, we must do something special. */
2147 if (any_condjump_p (insn) && onlyjump_p (insn) && map->last_pc_value)
2150 /* If the previous insn set cc0 for us, delete it. */
2151 if (only_sets_cc0_p (PREV_INSN (copy)))
2152 delete_related_insns (PREV_INSN (copy));
2155 /* If this is now a no-op, delete it. */
2156 if (map->last_pc_value == pc_rtx)
2162 /* Otherwise, this is unconditional jump so we must put a
2163 BARRIER after it. We could do some dead code elimination
2164 here, but jump.c will do it just as well. */
2170 pattern = copy_rtx_and_substitute (PATTERN (insn), map, 0);
2171 copy = emit_call_insn (pattern);
2172 REG_NOTES (copy) = initial_reg_note_copy (REG_NOTES (insn), map);
2173 INSN_SCOPE (copy) = INSN_SCOPE (insn);
2174 SIBLING_CALL_P (copy) = SIBLING_CALL_P (insn);
2176 /* Because the USAGE information potentially contains objects other
2177 than hard registers, we need to copy it. */
2178 CALL_INSN_FUNCTION_USAGE (copy)
2179 = copy_rtx_and_substitute (CALL_INSN_FUNCTION_USAGE (insn),
2184 try_constants (cc0_insn, map);
2187 try_constants (copy, map);
2189 /* Be lazy and assume CALL_INSNs clobber all hard registers. */
2190 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
2191 VARRAY_CONST_EQUIV (map->const_equiv_varray, i).rtx = 0;
2195 /* If this is the loop start label, then we don't need to emit a
2196 copy of this label since no one will use it. */
2198 if (insn != start_label)
2200 copy = emit_label (get_label_from_map (map,
2201 CODE_LABEL_NUMBER (insn)));
2207 copy = emit_barrier ();
2211 /* VTOP and CONT notes are valid only before the loop exit test.
2212 If placed anywhere else, loop may generate bad code. */
2213 /* BASIC_BLOCK notes exist to stabilize basic block structures with
2214 the associated rtl. We do not want to share the structure in
2217 if (NOTE_LINE_NUMBER (insn) != NOTE_INSN_DELETED
2218 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_DELETED_LABEL
2219 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_BASIC_BLOCK
2220 && ((NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_VTOP
2221 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_CONT)
2222 || (last_iteration && unroll_type != UNROLL_COMPLETELY)))
2223 copy = emit_note (NOTE_SOURCE_FILE (insn),
2224 NOTE_LINE_NUMBER (insn));
2233 map->insn_map[INSN_UID (insn)] = copy;
2235 while (insn != copy_end);
2237 /* Now finish coping the REG_NOTES. */
2241 insn = NEXT_INSN (insn);
2242 if ((GET_CODE (insn) == INSN || GET_CODE (insn) == JUMP_INSN
2243 || GET_CODE (insn) == CALL_INSN)
2244 && map->insn_map[INSN_UID (insn)])
2245 final_reg_note_copy (®_NOTES (map->insn_map[INSN_UID (insn)]), map);
2247 while (insn != copy_end);
2249 /* There may be notes between copy_notes_from and loop_end. Emit a copy of
2250 each of these notes here, since there may be some important ones, such as
2251 NOTE_INSN_BLOCK_END notes, in this group. We don't do this on the last
2252 iteration, because the original notes won't be deleted.
2254 We can't use insert_before here, because when from preconditioning,
2255 insert_before points before the loop. We can't use copy_end, because
2256 there may be insns already inserted after it (which we don't want to
2257 copy) when not from preconditioning code. */
2259 if (! last_iteration)
2261 for (insn = copy_notes_from; insn != loop_end; insn = NEXT_INSN (insn))
2263 /* VTOP notes are valid only before the loop exit test.
2264 If placed anywhere else, loop may generate bad code.
2265 There is no need to test for NOTE_INSN_LOOP_CONT notes
2266 here, since COPY_NOTES_FROM will be at most one or two (for cc0)
2267 instructions before the last insn in the loop, and if the
2268 end test is that short, there will be a VTOP note between
2269 the CONT note and the test. */
2270 if (GET_CODE (insn) == NOTE
2271 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_DELETED
2272 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_BASIC_BLOCK
2273 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_VTOP)
2274 emit_note (NOTE_SOURCE_FILE (insn), NOTE_LINE_NUMBER (insn));
2278 if (final_label && LABEL_NUSES (final_label) > 0)
2279 emit_label (final_label);
2281 tem = gen_sequence ();
2283 loop_insn_emit_before (loop, 0, insert_before, tem);
2286 /* Emit an insn, using the expand_binop to ensure that a valid insn is
2287 emitted. This will correctly handle the case where the increment value
2288 won't fit in the immediate field of a PLUS insns. */
2291 emit_unrolled_add (dest_reg, src_reg, increment)
2292 rtx dest_reg, src_reg, increment;
2296 result = expand_simple_binop (GET_MODE (dest_reg), PLUS, src_reg, increment,
2297 dest_reg, 0, OPTAB_LIB_WIDEN);
2299 if (dest_reg != result)
2300 emit_move_insn (dest_reg, result);
2303 /* Searches the insns between INSN and LOOP->END. Returns 1 if there
2304 is a backward branch in that range that branches to somewhere between
2305 LOOP->START and INSN. Returns 0 otherwise. */
2307 /* ??? This is quadratic algorithm. Could be rewritten to be linear.
2308 In practice, this is not a problem, because this function is seldom called,
2309 and uses a negligible amount of CPU time on average. */
2312 back_branch_in_range_p (loop, insn)
2313 const struct loop *loop;
2316 rtx p, q, target_insn;
2317 rtx loop_start = loop->start;
2318 rtx loop_end = loop->end;
2319 rtx orig_loop_end = loop->end;
2321 /* Stop before we get to the backward branch at the end of the loop. */
2322 loop_end = prev_nonnote_insn (loop_end);
2323 if (GET_CODE (loop_end) == BARRIER)
2324 loop_end = PREV_INSN (loop_end);
2326 /* Check in case insn has been deleted, search forward for first non
2327 deleted insn following it. */
2328 while (INSN_DELETED_P (insn))
2329 insn = NEXT_INSN (insn);
2331 /* Check for the case where insn is the last insn in the loop. Deal
2332 with the case where INSN was a deleted loop test insn, in which case
2333 it will now be the NOTE_LOOP_END. */
2334 if (insn == loop_end || insn == orig_loop_end)
2337 for (p = NEXT_INSN (insn); p != loop_end; p = NEXT_INSN (p))
2339 if (GET_CODE (p) == JUMP_INSN)
2341 target_insn = JUMP_LABEL (p);
2343 /* Search from loop_start to insn, to see if one of them is
2344 the target_insn. We can't use INSN_LUID comparisons here,
2345 since insn may not have an LUID entry. */
2346 for (q = loop_start; q != insn; q = NEXT_INSN (q))
2347 if (q == target_insn)
2355 /* Try to generate the simplest rtx for the expression
2356 (PLUS (MULT mult1 mult2) add1). This is used to calculate the initial
2360 fold_rtx_mult_add (mult1, mult2, add1, mode)
2361 rtx mult1, mult2, add1;
2362 enum machine_mode mode;
2367 /* The modes must all be the same. This should always be true. For now,
2368 check to make sure. */
2369 if ((GET_MODE (mult1) != mode && GET_MODE (mult1) != VOIDmode)
2370 || (GET_MODE (mult2) != mode && GET_MODE (mult2) != VOIDmode)
2371 || (GET_MODE (add1) != mode && GET_MODE (add1) != VOIDmode))
2374 /* Ensure that if at least one of mult1/mult2 are constant, then mult2
2375 will be a constant. */
2376 if (GET_CODE (mult1) == CONST_INT)
2383 mult_res = simplify_binary_operation (MULT, mode, mult1, mult2);
2385 mult_res = gen_rtx_MULT (mode, mult1, mult2);
2387 /* Again, put the constant second. */
2388 if (GET_CODE (add1) == CONST_INT)
2395 result = simplify_binary_operation (PLUS, mode, add1, mult_res);
2397 result = gen_rtx_PLUS (mode, add1, mult_res);
2402 /* Searches the list of induction struct's for the biv BL, to try to calculate
2403 the total increment value for one iteration of the loop as a constant.
2405 Returns the increment value as an rtx, simplified as much as possible,
2406 if it can be calculated. Otherwise, returns 0. */
2409 biv_total_increment (bl)
2410 const struct iv_class *bl;
2412 struct induction *v;
2415 /* For increment, must check every instruction that sets it. Each
2416 instruction must be executed only once each time through the loop.
2417 To verify this, we check that the insn is always executed, and that
2418 there are no backward branches after the insn that branch to before it.
2419 Also, the insn must have a mult_val of one (to make sure it really is
2422 result = const0_rtx;
2423 for (v = bl->biv; v; v = v->next_iv)
2425 if (v->always_computable && v->mult_val == const1_rtx
2426 && ! v->maybe_multiple)
2427 result = fold_rtx_mult_add (result, const1_rtx, v->add_val, v->mode);
2435 /* For each biv and giv, determine whether it can be safely split into
2436 a different variable for each unrolled copy of the loop body. If it
2437 is safe to split, then indicate that by saving some useful info
2438 in the splittable_regs array.
2440 If the loop is being completely unrolled, then splittable_regs will hold
2441 the current value of the induction variable while the loop is unrolled.
2442 It must be set to the initial value of the induction variable here.
2443 Otherwise, splittable_regs will hold the difference between the current
2444 value of the induction variable and the value the induction variable had
2445 at the top of the loop. It must be set to the value 0 here.
2447 Returns the total number of instructions that set registers that are
2450 /* ?? If the loop is only unrolled twice, then most of the restrictions to
2451 constant values are unnecessary, since we can easily calculate increment
2452 values in this case even if nothing is constant. The increment value
2453 should not involve a multiply however. */
2455 /* ?? Even if the biv/giv increment values aren't constant, it may still
2456 be beneficial to split the variable if the loop is only unrolled a few
2457 times, since multiplies by small integers (1,2,3,4) are very cheap. */
2460 find_splittable_regs (loop, unroll_type, unroll_number)
2461 const struct loop *loop;
2462 enum unroll_types unroll_type;
2465 struct loop_ivs *ivs = LOOP_IVS (loop);
2466 struct iv_class *bl;
2467 struct induction *v;
2469 rtx biv_final_value;
2473 for (bl = ivs->list; bl; bl = bl->next)
2475 /* Biv_total_increment must return a constant value,
2476 otherwise we can not calculate the split values. */
2478 increment = biv_total_increment (bl);
2479 if (! increment || GET_CODE (increment) != CONST_INT)
2482 /* The loop must be unrolled completely, or else have a known number
2483 of iterations and only one exit, or else the biv must be dead
2484 outside the loop, or else the final value must be known. Otherwise,
2485 it is unsafe to split the biv since it may not have the proper
2486 value on loop exit. */
2488 /* loop_number_exit_count is non-zero if the loop has an exit other than
2489 a fall through at the end. */
2492 biv_final_value = 0;
2493 if (unroll_type != UNROLL_COMPLETELY
2494 && (loop->exit_count || unroll_type == UNROLL_NAIVE)
2495 && (REGNO_LAST_LUID (bl->regno) >= INSN_LUID (loop->end)
2497 || INSN_UID (bl->init_insn) >= max_uid_for_loop
2498 || (REGNO_FIRST_LUID (bl->regno)
2499 < INSN_LUID (bl->init_insn))
2500 || reg_mentioned_p (bl->biv->dest_reg, SET_SRC (bl->init_set)))
2501 && ! (biv_final_value = final_biv_value (loop, bl)))
2504 /* If any of the insns setting the BIV don't do so with a simple
2505 PLUS, we don't know how to split it. */
2506 for (v = bl->biv; biv_splittable && v; v = v->next_iv)
2507 if ((tem = single_set (v->insn)) == 0
2508 || GET_CODE (SET_DEST (tem)) != REG
2509 || REGNO (SET_DEST (tem)) != bl->regno
2510 || GET_CODE (SET_SRC (tem)) != PLUS)
2513 /* If final value is non-zero, then must emit an instruction which sets
2514 the value of the biv to the proper value. This is done after
2515 handling all of the givs, since some of them may need to use the
2516 biv's value in their initialization code. */
2518 /* This biv is splittable. If completely unrolling the loop, save
2519 the biv's initial value. Otherwise, save the constant zero. */
2521 if (biv_splittable == 1)
2523 if (unroll_type == UNROLL_COMPLETELY)
2525 /* If the initial value of the biv is itself (i.e. it is too
2526 complicated for strength_reduce to compute), or is a hard
2527 register, or it isn't invariant, then we must create a new
2528 pseudo reg to hold the initial value of the biv. */
2530 if (GET_CODE (bl->initial_value) == REG
2531 && (REGNO (bl->initial_value) == bl->regno
2532 || REGNO (bl->initial_value) < FIRST_PSEUDO_REGISTER
2533 || ! loop_invariant_p (loop, bl->initial_value)))
2535 rtx tem = gen_reg_rtx (bl->biv->mode);
2537 record_base_value (REGNO (tem), bl->biv->add_val, 0);
2538 loop_insn_hoist (loop,
2539 gen_move_insn (tem, bl->biv->src_reg));
2541 if (loop_dump_stream)
2542 fprintf (loop_dump_stream,
2543 "Biv %d initial value remapped to %d.\n",
2544 bl->regno, REGNO (tem));
2546 splittable_regs[bl->regno] = tem;
2549 splittable_regs[bl->regno] = bl->initial_value;
2552 splittable_regs[bl->regno] = const0_rtx;
2554 /* Save the number of instructions that modify the biv, so that
2555 we can treat the last one specially. */
2557 splittable_regs_updates[bl->regno] = bl->biv_count;
2558 result += bl->biv_count;
2560 if (loop_dump_stream)
2561 fprintf (loop_dump_stream,
2562 "Biv %d safe to split.\n", bl->regno);
2565 /* Check every giv that depends on this biv to see whether it is
2566 splittable also. Even if the biv isn't splittable, givs which
2567 depend on it may be splittable if the biv is live outside the
2568 loop, and the givs aren't. */
2570 result += find_splittable_givs (loop, bl, unroll_type, increment,
2573 /* If final value is non-zero, then must emit an instruction which sets
2574 the value of the biv to the proper value. This is done after
2575 handling all of the givs, since some of them may need to use the
2576 biv's value in their initialization code. */
2577 if (biv_final_value)
2579 /* If the loop has multiple exits, emit the insns before the
2580 loop to ensure that it will always be executed no matter
2581 how the loop exits. Otherwise emit the insn after the loop,
2582 since this is slightly more efficient. */
2583 if (! loop->exit_count)
2584 loop_insn_sink (loop, gen_move_insn (bl->biv->src_reg,
2588 /* Create a new register to hold the value of the biv, and then
2589 set the biv to its final value before the loop start. The biv
2590 is set to its final value before loop start to ensure that
2591 this insn will always be executed, no matter how the loop
2593 rtx tem = gen_reg_rtx (bl->biv->mode);
2594 record_base_value (REGNO (tem), bl->biv->add_val, 0);
2596 loop_insn_hoist (loop, gen_move_insn (tem, bl->biv->src_reg));
2597 loop_insn_hoist (loop, gen_move_insn (bl->biv->src_reg,
2600 if (loop_dump_stream)
2601 fprintf (loop_dump_stream, "Biv %d mapped to %d for split.\n",
2602 REGNO (bl->biv->src_reg), REGNO (tem));
2604 /* Set up the mapping from the original biv register to the new
2606 bl->biv->src_reg = tem;
2613 /* Return 1 if the first and last unrolled copy of the address giv V is valid
2614 for the instruction that is using it. Do not make any changes to that
2618 verify_addresses (v, giv_inc, unroll_number)
2619 struct induction *v;
2624 rtx orig_addr = *v->location;
2625 rtx last_addr = plus_constant (v->dest_reg,
2626 INTVAL (giv_inc) * (unroll_number - 1));
2628 /* First check to see if either address would fail. Handle the fact
2629 that we have may have a match_dup. */
2630 if (! validate_replace_rtx (*v->location, v->dest_reg, v->insn)
2631 || ! validate_replace_rtx (*v->location, last_addr, v->insn))
2634 /* Now put things back the way they were before. This should always
2636 if (! validate_replace_rtx (*v->location, orig_addr, v->insn))
2642 /* For every giv based on the biv BL, check to determine whether it is
2643 splittable. This is a subroutine to find_splittable_regs ().
2645 Return the number of instructions that set splittable registers. */
2648 find_splittable_givs (loop, bl, unroll_type, increment, unroll_number)
2649 const struct loop *loop;
2650 struct iv_class *bl;
2651 enum unroll_types unroll_type;
2655 struct loop_ivs *ivs = LOOP_IVS (loop);
2656 struct induction *v, *v2;
2661 /* Scan the list of givs, and set the same_insn field when there are
2662 multiple identical givs in the same insn. */
2663 for (v = bl->giv; v; v = v->next_iv)
2664 for (v2 = v->next_iv; v2; v2 = v2->next_iv)
2665 if (v->insn == v2->insn && rtx_equal_p (v->new_reg, v2->new_reg)
2669 for (v = bl->giv; v; v = v->next_iv)
2673 /* Only split the giv if it has already been reduced, or if the loop is
2674 being completely unrolled. */
2675 if (unroll_type != UNROLL_COMPLETELY && v->ignore)
2678 /* The giv can be split if the insn that sets the giv is executed once
2679 and only once on every iteration of the loop. */
2680 /* An address giv can always be split. v->insn is just a use not a set,
2681 and hence it does not matter whether it is always executed. All that
2682 matters is that all the biv increments are always executed, and we
2683 won't reach here if they aren't. */
2684 if (v->giv_type != DEST_ADDR
2685 && (! v->always_computable
2686 || back_branch_in_range_p (loop, v->insn)))
2689 /* The giv increment value must be a constant. */
2690 giv_inc = fold_rtx_mult_add (v->mult_val, increment, const0_rtx,
2692 if (! giv_inc || GET_CODE (giv_inc) != CONST_INT)
2695 /* The loop must be unrolled completely, or else have a known number of
2696 iterations and only one exit, or else the giv must be dead outside
2697 the loop, or else the final value of the giv must be known.
2698 Otherwise, it is not safe to split the giv since it may not have the
2699 proper value on loop exit. */
2701 /* The used outside loop test will fail for DEST_ADDR givs. They are
2702 never used outside the loop anyways, so it is always safe to split a
2706 if (unroll_type != UNROLL_COMPLETELY
2707 && (loop->exit_count || unroll_type == UNROLL_NAIVE)
2708 && v->giv_type != DEST_ADDR
2709 /* The next part is true if the pseudo is used outside the loop.
2710 We assume that this is true for any pseudo created after loop
2711 starts, because we don't have a reg_n_info entry for them. */
2712 && (REGNO (v->dest_reg) >= max_reg_before_loop
2713 || (REGNO_FIRST_UID (REGNO (v->dest_reg)) != INSN_UID (v->insn)
2714 /* Check for the case where the pseudo is set by a shift/add
2715 sequence, in which case the first insn setting the pseudo
2716 is the first insn of the shift/add sequence. */
2717 && (! (tem = find_reg_note (v->insn, REG_RETVAL, NULL_RTX))
2718 || (REGNO_FIRST_UID (REGNO (v->dest_reg))
2719 != INSN_UID (XEXP (tem, 0)))))
2720 /* Line above always fails if INSN was moved by loop opt. */
2721 || (REGNO_LAST_LUID (REGNO (v->dest_reg))
2722 >= INSN_LUID (loop->end)))
2723 && ! (final_value = v->final_value))
2727 /* Currently, non-reduced/final-value givs are never split. */
2728 /* Should emit insns after the loop if possible, as the biv final value
2731 /* If the final value is non-zero, and the giv has not been reduced,
2732 then must emit an instruction to set the final value. */
2733 if (final_value && !v->new_reg)
2735 /* Create a new register to hold the value of the giv, and then set
2736 the giv to its final value before the loop start. The giv is set
2737 to its final value before loop start to ensure that this insn
2738 will always be executed, no matter how we exit. */
2739 tem = gen_reg_rtx (v->mode);
2740 loop_insn_hoist (loop, gen_move_insn (tem, v->dest_reg));
2741 loop_insn_hoist (loop, gen_move_insn (v->dest_reg, final_value));
2743 if (loop_dump_stream)
2744 fprintf (loop_dump_stream, "Giv %d mapped to %d for split.\n",
2745 REGNO (v->dest_reg), REGNO (tem));
2751 /* This giv is splittable. If completely unrolling the loop, save the
2752 giv's initial value. Otherwise, save the constant zero for it. */
2754 if (unroll_type == UNROLL_COMPLETELY)
2756 /* It is not safe to use bl->initial_value here, because it may not
2757 be invariant. It is safe to use the initial value stored in
2758 the splittable_regs array if it is set. In rare cases, it won't
2759 be set, so then we do exactly the same thing as
2760 find_splittable_regs does to get a safe value. */
2761 rtx biv_initial_value;
2763 if (splittable_regs[bl->regno])
2764 biv_initial_value = splittable_regs[bl->regno];
2765 else if (GET_CODE (bl->initial_value) != REG
2766 || (REGNO (bl->initial_value) != bl->regno
2767 && REGNO (bl->initial_value) >= FIRST_PSEUDO_REGISTER))
2768 biv_initial_value = bl->initial_value;
2771 rtx tem = gen_reg_rtx (bl->biv->mode);
2773 record_base_value (REGNO (tem), bl->biv->add_val, 0);
2774 loop_insn_hoist (loop, gen_move_insn (tem, bl->biv->src_reg));
2775 biv_initial_value = tem;
2777 biv_initial_value = extend_value_for_giv (v, biv_initial_value);
2778 value = fold_rtx_mult_add (v->mult_val, biv_initial_value,
2779 v->add_val, v->mode);
2786 /* If a giv was combined with another giv, then we can only split
2787 this giv if the giv it was combined with was reduced. This
2788 is because the value of v->new_reg is meaningless in this
2790 if (v->same && ! v->same->new_reg)
2792 if (loop_dump_stream)
2793 fprintf (loop_dump_stream,
2794 "giv combined with unreduced giv not split.\n");
2797 /* If the giv is an address destination, it could be something other
2798 than a simple register, these have to be treated differently. */
2799 else if (v->giv_type == DEST_REG)
2801 /* If value is not a constant, register, or register plus
2802 constant, then compute its value into a register before
2803 loop start. This prevents invalid rtx sharing, and should
2804 generate better code. We can use bl->initial_value here
2805 instead of splittable_regs[bl->regno] because this code
2806 is going before the loop start. */
2807 if (unroll_type == UNROLL_COMPLETELY
2808 && GET_CODE (value) != CONST_INT
2809 && GET_CODE (value) != REG
2810 && (GET_CODE (value) != PLUS
2811 || GET_CODE (XEXP (value, 0)) != REG
2812 || GET_CODE (XEXP (value, 1)) != CONST_INT))
2814 rtx tem = gen_reg_rtx (v->mode);
2815 record_base_value (REGNO (tem), v->add_val, 0);
2816 loop_iv_add_mult_hoist (loop, bl->initial_value, v->mult_val,
2821 splittable_regs[REGNO (v->new_reg)] = value;
2825 /* Splitting address givs is useful since it will often allow us
2826 to eliminate some increment insns for the base giv as
2829 /* If the addr giv is combined with a dest_reg giv, then all
2830 references to that dest reg will be remapped, which is NOT
2831 what we want for split addr regs. We always create a new
2832 register for the split addr giv, just to be safe. */
2834 /* If we have multiple identical address givs within a
2835 single instruction, then use a single pseudo reg for
2836 both. This is necessary in case one is a match_dup
2839 v->const_adjust = 0;
2843 v->dest_reg = v->same_insn->dest_reg;
2844 if (loop_dump_stream)
2845 fprintf (loop_dump_stream,
2846 "Sharing address givs in insn %d\n",
2847 INSN_UID (v->insn));
2849 /* If multiple address GIVs have been combined with the
2850 same dest_reg GIV, do not create a new register for
2852 else if (unroll_type != UNROLL_COMPLETELY
2853 && v->giv_type == DEST_ADDR
2854 && v->same && v->same->giv_type == DEST_ADDR
2855 && v->same->unrolled
2856 /* combine_givs_p may return true for some cases
2857 where the add and mult values are not equal.
2858 To share a register here, the values must be
2860 && rtx_equal_p (v->same->mult_val, v->mult_val)
2861 && rtx_equal_p (v->same->add_val, v->add_val)
2862 /* If the memory references have different modes,
2863 then the address may not be valid and we must
2864 not share registers. */
2865 && verify_addresses (v, giv_inc, unroll_number))
2867 v->dest_reg = v->same->dest_reg;
2870 else if (unroll_type != UNROLL_COMPLETELY)
2872 /* If not completely unrolling the loop, then create a new
2873 register to hold the split value of the DEST_ADDR giv.
2874 Emit insn to initialize its value before loop start. */
2876 rtx tem = gen_reg_rtx (v->mode);
2877 struct induction *same = v->same;
2878 rtx new_reg = v->new_reg;
2879 record_base_value (REGNO (tem), v->add_val, 0);
2881 /* If the address giv has a constant in its new_reg value,
2882 then this constant can be pulled out and put in value,
2883 instead of being part of the initialization code. */
2885 if (GET_CODE (new_reg) == PLUS
2886 && GET_CODE (XEXP (new_reg, 1)) == CONST_INT)
2889 = plus_constant (tem, INTVAL (XEXP (new_reg, 1)));
2891 /* Only succeed if this will give valid addresses.
2892 Try to validate both the first and the last
2893 address resulting from loop unrolling, if
2894 one fails, then can't do const elim here. */
2895 if (verify_addresses (v, giv_inc, unroll_number))
2897 /* Save the negative of the eliminated const, so
2898 that we can calculate the dest_reg's increment
2900 v->const_adjust = -INTVAL (XEXP (new_reg, 1));
2902 new_reg = XEXP (new_reg, 0);
2903 if (loop_dump_stream)
2904 fprintf (loop_dump_stream,
2905 "Eliminating constant from giv %d\n",
2914 /* If the address hasn't been checked for validity yet, do so
2915 now, and fail completely if either the first or the last
2916 unrolled copy of the address is not a valid address
2917 for the instruction that uses it. */
2918 if (v->dest_reg == tem
2919 && ! verify_addresses (v, giv_inc, unroll_number))
2921 for (v2 = v->next_iv; v2; v2 = v2->next_iv)
2922 if (v2->same_insn == v)
2925 if (loop_dump_stream)
2926 fprintf (loop_dump_stream,
2927 "Invalid address for giv at insn %d\n",
2928 INSN_UID (v->insn));
2932 v->new_reg = new_reg;
2935 /* We set this after the address check, to guarantee that
2936 the register will be initialized. */
2939 /* To initialize the new register, just move the value of
2940 new_reg into it. This is not guaranteed to give a valid
2941 instruction on machines with complex addressing modes.
2942 If we can't recognize it, then delete it and emit insns
2943 to calculate the value from scratch. */
2944 loop_insn_hoist (loop, gen_rtx_SET (VOIDmode, tem,
2945 copy_rtx (v->new_reg)));
2946 if (recog_memoized (PREV_INSN (loop->start)) < 0)
2950 /* We can't use bl->initial_value to compute the initial
2951 value, because the loop may have been preconditioned.
2952 We must calculate it from NEW_REG. */
2953 delete_related_insns (PREV_INSN (loop->start));
2956 ret = force_operand (v->new_reg, tem);
2958 emit_move_insn (tem, ret);
2959 sequence = gen_sequence ();
2961 loop_insn_hoist (loop, sequence);
2963 if (loop_dump_stream)
2964 fprintf (loop_dump_stream,
2965 "Invalid init insn, rewritten.\n");
2970 v->dest_reg = value;
2972 /* Check the resulting address for validity, and fail
2973 if the resulting address would be invalid. */
2974 if (! verify_addresses (v, giv_inc, unroll_number))
2976 for (v2 = v->next_iv; v2; v2 = v2->next_iv)
2977 if (v2->same_insn == v)
2980 if (loop_dump_stream)
2981 fprintf (loop_dump_stream,
2982 "Invalid address for giv at insn %d\n",
2983 INSN_UID (v->insn));
2988 /* Store the value of dest_reg into the insn. This sharing
2989 will not be a problem as this insn will always be copied
2992 *v->location = v->dest_reg;
2994 /* If this address giv is combined with a dest reg giv, then
2995 save the base giv's induction pointer so that we will be
2996 able to handle this address giv properly. The base giv
2997 itself does not have to be splittable. */
2999 if (v->same && v->same->giv_type == DEST_REG)
3000 addr_combined_regs[REGNO (v->same->new_reg)] = v->same;
3002 if (GET_CODE (v->new_reg) == REG)
3004 /* This giv maybe hasn't been combined with any others.
3005 Make sure that it's giv is marked as splittable here. */
3007 splittable_regs[REGNO (v->new_reg)] = value;
3009 /* Make it appear to depend upon itself, so that the
3010 giv will be properly split in the main loop above. */
3014 addr_combined_regs[REGNO (v->new_reg)] = v;
3018 if (loop_dump_stream)
3019 fprintf (loop_dump_stream, "DEST_ADDR giv being split.\n");
3025 /* Currently, unreduced giv's can't be split. This is not too much
3026 of a problem since unreduced giv's are not live across loop
3027 iterations anyways. When unrolling a loop completely though,
3028 it makes sense to reduce&split givs when possible, as this will
3029 result in simpler instructions, and will not require that a reg
3030 be live across loop iterations. */
3032 splittable_regs[REGNO (v->dest_reg)] = value;
3033 fprintf (stderr, "Giv %d at insn %d not reduced\n",
3034 REGNO (v->dest_reg), INSN_UID (v->insn));
3040 /* Unreduced givs are only updated once by definition. Reduced givs
3041 are updated as many times as their biv is. Mark it so if this is
3042 a splittable register. Don't need to do anything for address givs
3043 where this may not be a register. */
3045 if (GET_CODE (v->new_reg) == REG)
3049 count = REG_IV_CLASS (ivs, REGNO (v->src_reg))->biv_count;
3051 splittable_regs_updates[REGNO (v->new_reg)] = count;
3056 if (loop_dump_stream)
3060 if (GET_CODE (v->dest_reg) == CONST_INT)
3062 else if (GET_CODE (v->dest_reg) != REG)
3063 regnum = REGNO (XEXP (v->dest_reg, 0));
3065 regnum = REGNO (v->dest_reg);
3066 fprintf (loop_dump_stream, "Giv %d at insn %d safe to split.\n",
3067 regnum, INSN_UID (v->insn));
3074 /* Try to prove that the register is dead after the loop exits. Trace every
3075 loop exit looking for an insn that will always be executed, which sets
3076 the register to some value, and appears before the first use of the register
3077 is found. If successful, then return 1, otherwise return 0. */
3079 /* ?? Could be made more intelligent in the handling of jumps, so that
3080 it can search past if statements and other similar structures. */
3083 reg_dead_after_loop (loop, reg)
3084 const struct loop *loop;
3090 int label_count = 0;
3092 /* In addition to checking all exits of this loop, we must also check
3093 all exits of inner nested loops that would exit this loop. We don't
3094 have any way to identify those, so we just give up if there are any
3095 such inner loop exits. */
3097 for (label = loop->exit_labels; label; label = LABEL_NEXTREF (label))
3100 if (label_count != loop->exit_count)
3103 /* HACK: Must also search the loop fall through exit, create a label_ref
3104 here which points to the loop->end, and append the loop_number_exit_labels
3106 label = gen_rtx_LABEL_REF (VOIDmode, loop->end);
3107 LABEL_NEXTREF (label) = loop->exit_labels;
3109 for (; label; label = LABEL_NEXTREF (label))
3111 /* Succeed if find an insn which sets the biv or if reach end of
3112 function. Fail if find an insn that uses the biv, or if come to
3113 a conditional jump. */
3115 insn = NEXT_INSN (XEXP (label, 0));
3118 code = GET_CODE (insn);
3119 if (GET_RTX_CLASS (code) == 'i')
3123 if (reg_referenced_p (reg, PATTERN (insn)))
3126 set = single_set (insn);
3127 if (set && rtx_equal_p (SET_DEST (set), reg))
3131 if (code == JUMP_INSN)
3133 if (GET_CODE (PATTERN (insn)) == RETURN)
3135 else if (!any_uncondjump_p (insn)
3136 /* Prevent infinite loop following infinite loops. */
3137 || jump_count++ > 20)
3140 insn = JUMP_LABEL (insn);
3143 insn = NEXT_INSN (insn);
3147 /* Success, the register is dead on all loop exits. */
3151 /* Try to calculate the final value of the biv, the value it will have at
3152 the end of the loop. If we can do it, return that value. */
3155 final_biv_value (loop, bl)
3156 const struct loop *loop;
3157 struct iv_class *bl;
3159 unsigned HOST_WIDE_INT n_iterations = LOOP_INFO (loop)->n_iterations;
3162 /* ??? This only works for MODE_INT biv's. Reject all others for now. */
3164 if (GET_MODE_CLASS (bl->biv->mode) != MODE_INT)
3167 /* The final value for reversed bivs must be calculated differently than
3168 for ordinary bivs. In this case, there is already an insn after the
3169 loop which sets this biv's final value (if necessary), and there are
3170 no other loop exits, so we can return any value. */
3173 if (loop_dump_stream)
3174 fprintf (loop_dump_stream,
3175 "Final biv value for %d, reversed biv.\n", bl->regno);
3180 /* Try to calculate the final value as initial value + (number of iterations
3181 * increment). For this to work, increment must be invariant, the only
3182 exit from the loop must be the fall through at the bottom (otherwise
3183 it may not have its final value when the loop exits), and the initial
3184 value of the biv must be invariant. */
3186 if (n_iterations != 0
3187 && ! loop->exit_count
3188 && loop_invariant_p (loop, bl->initial_value))
3190 increment = biv_total_increment (bl);
3192 if (increment && loop_invariant_p (loop, increment))
3194 /* Can calculate the loop exit value, emit insns after loop
3195 end to calculate this value into a temporary register in
3196 case it is needed later. */
3198 tem = gen_reg_rtx (bl->biv->mode);
3199 record_base_value (REGNO (tem), bl->biv->add_val, 0);
3200 loop_iv_add_mult_sink (loop, increment, GEN_INT (n_iterations),
3201 bl->initial_value, tem);
3203 if (loop_dump_stream)
3204 fprintf (loop_dump_stream,
3205 "Final biv value for %d, calculated.\n", bl->regno);
3211 /* Check to see if the biv is dead at all loop exits. */
3212 if (reg_dead_after_loop (loop, bl->biv->src_reg))
3214 if (loop_dump_stream)
3215 fprintf (loop_dump_stream,
3216 "Final biv value for %d, biv dead after loop exit.\n",
3225 /* Try to calculate the final value of the giv, the value it will have at
3226 the end of the loop. If we can do it, return that value. */
3229 final_giv_value (loop, v)
3230 const struct loop *loop;
3231 struct induction *v;
3233 struct loop_ivs *ivs = LOOP_IVS (loop);
3234 struct iv_class *bl;
3238 rtx loop_end = loop->end;
3239 unsigned HOST_WIDE_INT n_iterations = LOOP_INFO (loop)->n_iterations;
3241 bl = REG_IV_CLASS (ivs, REGNO (v->src_reg));
3243 /* The final value for givs which depend on reversed bivs must be calculated
3244 differently than for ordinary givs. In this case, there is already an
3245 insn after the loop which sets this giv's final value (if necessary),
3246 and there are no other loop exits, so we can return any value. */
3249 if (loop_dump_stream)
3250 fprintf (loop_dump_stream,
3251 "Final giv value for %d, depends on reversed biv\n",
3252 REGNO (v->dest_reg));
3256 /* Try to calculate the final value as a function of the biv it depends
3257 upon. The only exit from the loop must be the fall through at the bottom
3258 (otherwise it may not have its final value when the loop exits). */
3260 /* ??? Can calculate the final giv value by subtracting off the
3261 extra biv increments times the giv's mult_val. The loop must have
3262 only one exit for this to work, but the loop iterations does not need
3265 if (n_iterations != 0
3266 && ! loop->exit_count)
3268 /* ?? It is tempting to use the biv's value here since these insns will
3269 be put after the loop, and hence the biv will have its final value
3270 then. However, this fails if the biv is subsequently eliminated.
3271 Perhaps determine whether biv's are eliminable before trying to
3272 determine whether giv's are replaceable so that we can use the
3273 biv value here if it is not eliminable. */
3275 /* We are emitting code after the end of the loop, so we must make
3276 sure that bl->initial_value is still valid then. It will still
3277 be valid if it is invariant. */
3279 increment = biv_total_increment (bl);
3281 if (increment && loop_invariant_p (loop, increment)
3282 && loop_invariant_p (loop, bl->initial_value))
3284 /* Can calculate the loop exit value of its biv as
3285 (n_iterations * increment) + initial_value */
3287 /* The loop exit value of the giv is then
3288 (final_biv_value - extra increments) * mult_val + add_val.
3289 The extra increments are any increments to the biv which
3290 occur in the loop after the giv's value is calculated.
3291 We must search from the insn that sets the giv to the end
3292 of the loop to calculate this value. */
3294 /* Put the final biv value in tem. */
3295 tem = gen_reg_rtx (v->mode);
3296 record_base_value (REGNO (tem), bl->biv->add_val, 0);
3297 loop_iv_add_mult_sink (loop, extend_value_for_giv (v, increment),
3298 GEN_INT (n_iterations),
3299 extend_value_for_giv (v, bl->initial_value),
3302 /* Subtract off extra increments as we find them. */
3303 for (insn = NEXT_INSN (v->insn); insn != loop_end;
3304 insn = NEXT_INSN (insn))
3306 struct induction *biv;
3308 for (biv = bl->biv; biv; biv = biv->next_iv)
3309 if (biv->insn == insn)
3312 tem = expand_simple_binop (GET_MODE (tem), MINUS, tem,
3313 biv->add_val, NULL_RTX, 0,
3315 seq = gen_sequence ();
3317 loop_insn_sink (loop, seq);
3321 /* Now calculate the giv's final value. */
3322 loop_iv_add_mult_sink (loop, tem, v->mult_val, v->add_val, tem);
3324 if (loop_dump_stream)
3325 fprintf (loop_dump_stream,
3326 "Final giv value for %d, calc from biv's value.\n",
3327 REGNO (v->dest_reg));
3333 /* Replaceable giv's should never reach here. */
3337 /* Check to see if the biv is dead at all loop exits. */
3338 if (reg_dead_after_loop (loop, v->dest_reg))
3340 if (loop_dump_stream)
3341 fprintf (loop_dump_stream,
3342 "Final giv value for %d, giv dead after loop exit.\n",
3343 REGNO (v->dest_reg));
3351 /* Look back before LOOP->START for the insn that sets REG and return
3352 the equivalent constant if there is a REG_EQUAL note otherwise just
3353 the SET_SRC of REG. */
3356 loop_find_equiv_value (loop, reg)
3357 const struct loop *loop;
3360 rtx loop_start = loop->start;
3365 for (insn = PREV_INSN (loop_start); insn; insn = PREV_INSN (insn))
3367 if (GET_CODE (insn) == CODE_LABEL)
3370 else if (INSN_P (insn) && reg_set_p (reg, insn))
3372 /* We found the last insn before the loop that sets the register.
3373 If it sets the entire register, and has a REG_EQUAL note,
3374 then use the value of the REG_EQUAL note. */
3375 if ((set = single_set (insn))
3376 && (SET_DEST (set) == reg))
3378 rtx note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
3380 /* Only use the REG_EQUAL note if it is a constant.
3381 Other things, divide in particular, will cause
3382 problems later if we use them. */
3383 if (note && GET_CODE (XEXP (note, 0)) != EXPR_LIST
3384 && CONSTANT_P (XEXP (note, 0)))
3385 ret = XEXP (note, 0);
3387 ret = SET_SRC (set);
3389 /* We cannot do this if it changes between the
3390 assignment and loop start though. */
3391 if (modified_between_p (ret, insn, loop_start))
3400 /* Return a simplified rtx for the expression OP - REG.
3402 REG must appear in OP, and OP must be a register or the sum of a register
3405 Thus, the return value must be const0_rtx or the second term.
3407 The caller is responsible for verifying that REG appears in OP and OP has
3411 subtract_reg_term (op, reg)
3416 if (GET_CODE (op) == PLUS)
3418 if (XEXP (op, 0) == reg)
3419 return XEXP (op, 1);
3420 else if (XEXP (op, 1) == reg)
3421 return XEXP (op, 0);
3423 /* OP does not contain REG as a term. */
3427 /* Find and return register term common to both expressions OP0 and
3428 OP1 or NULL_RTX if no such term exists. Each expression must be a
3429 REG or a PLUS of a REG. */
3432 find_common_reg_term (op0, op1)
3435 if ((GET_CODE (op0) == REG || GET_CODE (op0) == PLUS)
3436 && (GET_CODE (op1) == REG || GET_CODE (op1) == PLUS))
3443 if (GET_CODE (op0) == PLUS)
3444 op01 = XEXP (op0, 1), op00 = XEXP (op0, 0);
3446 op01 = const0_rtx, op00 = op0;
3448 if (GET_CODE (op1) == PLUS)
3449 op11 = XEXP (op1, 1), op10 = XEXP (op1, 0);
3451 op11 = const0_rtx, op10 = op1;
3453 /* Find and return common register term if present. */
3454 if (REG_P (op00) && (op00 == op10 || op00 == op11))
3456 else if (REG_P (op01) && (op01 == op10 || op01 == op11))
3460 /* No common register term found. */
3464 /* Determine the loop iterator and calculate the number of loop
3465 iterations. Returns the exact number of loop iterations if it can
3466 be calculated, otherwise returns zero. */
3468 unsigned HOST_WIDE_INT
3469 loop_iterations (loop)
3472 struct loop_info *loop_info = LOOP_INFO (loop);
3473 struct loop_ivs *ivs = LOOP_IVS (loop);
3474 rtx comparison, comparison_value;
3475 rtx iteration_var, initial_value, increment, final_value;
3476 enum rtx_code comparison_code;
3478 unsigned HOST_WIDE_INT abs_inc;
3479 unsigned HOST_WIDE_INT abs_diff;
3482 int unsigned_p, compare_dir, final_larger;
3485 struct iv_class *bl;
3487 loop_info->n_iterations = 0;
3488 loop_info->initial_value = 0;
3489 loop_info->initial_equiv_value = 0;
3490 loop_info->comparison_value = 0;
3491 loop_info->final_value = 0;
3492 loop_info->final_equiv_value = 0;
3493 loop_info->increment = 0;
3494 loop_info->iteration_var = 0;
3495 loop_info->unroll_number = 1;
3498 /* We used to use prev_nonnote_insn here, but that fails because it might
3499 accidentally get the branch for a contained loop if the branch for this
3500 loop was deleted. We can only trust branches immediately before the
3502 last_loop_insn = PREV_INSN (loop->end);
3504 /* ??? We should probably try harder to find the jump insn
3505 at the end of the loop. The following code assumes that
3506 the last loop insn is a jump to the top of the loop. */
3507 if (GET_CODE (last_loop_insn) != JUMP_INSN)
3509 if (loop_dump_stream)
3510 fprintf (loop_dump_stream,
3511 "Loop iterations: No final conditional branch found.\n");
3515 /* If there is a more than a single jump to the top of the loop
3516 we cannot (easily) determine the iteration count. */
3517 if (LABEL_NUSES (JUMP_LABEL (last_loop_insn)) > 1)
3519 if (loop_dump_stream)
3520 fprintf (loop_dump_stream,
3521 "Loop iterations: Loop has multiple back edges.\n");
3525 /* If there are multiple conditionalized loop exit tests, they may jump
3526 back to differing CODE_LABELs. */
3527 if (loop->top && loop->cont)
3529 rtx temp = PREV_INSN (last_loop_insn);
3533 if (GET_CODE (temp) == JUMP_INSN)
3535 /* There are some kinds of jumps we can't deal with easily. */
3536 if (JUMP_LABEL (temp) == 0)
3538 if (loop_dump_stream)
3541 "Loop iterations: Jump insn has null JUMP_LABEL.\n");
3545 if (/* Previous unrolling may have generated new insns not
3546 covered by the uid_luid array. */
3547 INSN_UID (JUMP_LABEL (temp)) < max_uid_for_loop
3548 /* Check if we jump back into the loop body. */
3549 && INSN_LUID (JUMP_LABEL (temp)) > INSN_LUID (loop->top)
3550 && INSN_LUID (JUMP_LABEL (temp)) < INSN_LUID (loop->cont))
3552 if (loop_dump_stream)
3555 "Loop iterations: Loop has multiple back edges.\n");
3560 while ((temp = PREV_INSN (temp)) != loop->cont);
3563 /* Find the iteration variable. If the last insn is a conditional
3564 branch, and the insn before tests a register value, make that the
3565 iteration variable. */
3567 comparison = get_condition_for_loop (loop, last_loop_insn);
3568 if (comparison == 0)
3570 if (loop_dump_stream)
3571 fprintf (loop_dump_stream,
3572 "Loop iterations: No final comparison found.\n");
3576 /* ??? Get_condition may switch position of induction variable and
3577 invariant register when it canonicalizes the comparison. */
3579 comparison_code = GET_CODE (comparison);
3580 iteration_var = XEXP (comparison, 0);
3581 comparison_value = XEXP (comparison, 1);
3583 if (GET_CODE (iteration_var) != REG)
3585 if (loop_dump_stream)
3586 fprintf (loop_dump_stream,
3587 "Loop iterations: Comparison not against register.\n");
3591 /* The only new registers that are created before loop iterations
3592 are givs made from biv increments or registers created by
3593 load_mems. In the latter case, it is possible that try_copy_prop
3594 will propagate a new pseudo into the old iteration register but
3595 this will be marked by having the REG_USERVAR_P bit set. */
3597 if ((unsigned) REGNO (iteration_var) >= ivs->n_regs
3598 && ! REG_USERVAR_P (iteration_var))
3601 /* Determine the initial value of the iteration variable, and the amount
3602 that it is incremented each loop. Use the tables constructed by
3603 the strength reduction pass to calculate these values. */
3605 /* Clear the result values, in case no answer can be found. */
3609 /* The iteration variable can be either a giv or a biv. Check to see
3610 which it is, and compute the variable's initial value, and increment
3611 value if possible. */
3613 /* If this is a new register, can't handle it since we don't have any
3614 reg_iv_type entry for it. */
3615 if ((unsigned) REGNO (iteration_var) >= ivs->n_regs)
3617 if (loop_dump_stream)
3618 fprintf (loop_dump_stream,
3619 "Loop iterations: No reg_iv_type entry for iteration var.\n");
3623 /* Reject iteration variables larger than the host wide int size, since they
3624 could result in a number of iterations greater than the range of our
3625 `unsigned HOST_WIDE_INT' variable loop_info->n_iterations. */
3626 else if ((GET_MODE_BITSIZE (GET_MODE (iteration_var))
3627 > HOST_BITS_PER_WIDE_INT))
3629 if (loop_dump_stream)
3630 fprintf (loop_dump_stream,
3631 "Loop iterations: Iteration var rejected because mode too large.\n");
3634 else if (GET_MODE_CLASS (GET_MODE (iteration_var)) != MODE_INT)
3636 if (loop_dump_stream)
3637 fprintf (loop_dump_stream,
3638 "Loop iterations: Iteration var not an integer.\n");
3641 else if (REG_IV_TYPE (ivs, REGNO (iteration_var)) == BASIC_INDUCT)
3643 if (REGNO (iteration_var) >= ivs->n_regs)
3646 /* Grab initial value, only useful if it is a constant. */
3647 bl = REG_IV_CLASS (ivs, REGNO (iteration_var));
3648 initial_value = bl->initial_value;
3649 if (!bl->biv->always_executed || bl->biv->maybe_multiple)
3651 if (loop_dump_stream)
3652 fprintf (loop_dump_stream,
3653 "Loop iterations: Basic induction var not set once in each iteration.\n");
3657 increment = biv_total_increment (bl);
3659 else if (REG_IV_TYPE (ivs, REGNO (iteration_var)) == GENERAL_INDUCT)
3661 HOST_WIDE_INT offset = 0;
3662 struct induction *v = REG_IV_INFO (ivs, REGNO (iteration_var));
3663 rtx biv_initial_value;
3665 if (REGNO (v->src_reg) >= ivs->n_regs)
3668 if (!v->always_executed || v->maybe_multiple)
3670 if (loop_dump_stream)
3671 fprintf (loop_dump_stream,
3672 "Loop iterations: General induction var not set once in each iteration.\n");
3676 bl = REG_IV_CLASS (ivs, REGNO (v->src_reg));
3678 /* Increment value is mult_val times the increment value of the biv. */
3680 increment = biv_total_increment (bl);
3683 struct induction *biv_inc;
3685 increment = fold_rtx_mult_add (v->mult_val,
3686 extend_value_for_giv (v, increment),
3687 const0_rtx, v->mode);
3688 /* The caller assumes that one full increment has occurred at the
3689 first loop test. But that's not true when the biv is incremented
3690 after the giv is set (which is the usual case), e.g.:
3691 i = 6; do {;} while (i++ < 9) .
3692 Therefore, we bias the initial value by subtracting the amount of
3693 the increment that occurs between the giv set and the giv test. */
3694 for (biv_inc = bl->biv; biv_inc; biv_inc = biv_inc->next_iv)
3696 if (loop_insn_first_p (v->insn, biv_inc->insn))
3698 if (REG_P (biv_inc->add_val))
3700 if (loop_dump_stream)
3701 fprintf (loop_dump_stream,
3702 "Loop iterations: Basic induction var add_val is REG %d.\n",
3703 REGNO (biv_inc->add_val));
3707 offset -= INTVAL (biv_inc->add_val);
3711 if (loop_dump_stream)
3712 fprintf (loop_dump_stream,
3713 "Loop iterations: Giv iterator, initial value bias %ld.\n",
3716 /* Initial value is mult_val times the biv's initial value plus
3717 add_val. Only useful if it is a constant. */
3718 biv_initial_value = extend_value_for_giv (v, bl->initial_value);
3720 = fold_rtx_mult_add (v->mult_val,
3721 plus_constant (biv_initial_value, offset),
3722 v->add_val, v->mode);
3726 if (loop_dump_stream)
3727 fprintf (loop_dump_stream,
3728 "Loop iterations: Not basic or general induction var.\n");
3732 if (initial_value == 0)
3737 switch (comparison_code)
3752 /* Cannot determine loop iterations with this case. */
3771 /* If the comparison value is an invariant register, then try to find
3772 its value from the insns before the start of the loop. */
3774 final_value = comparison_value;
3775 if (GET_CODE (comparison_value) == REG
3776 && loop_invariant_p (loop, comparison_value))
3778 final_value = loop_find_equiv_value (loop, comparison_value);
3780 /* If we don't get an invariant final value, we are better
3781 off with the original register. */
3782 if (! loop_invariant_p (loop, final_value))
3783 final_value = comparison_value;
3786 /* Calculate the approximate final value of the induction variable
3787 (on the last successful iteration). The exact final value
3788 depends on the branch operator, and increment sign. It will be
3789 wrong if the iteration variable is not incremented by one each
3790 time through the loop and (comparison_value + off_by_one -
3791 initial_value) % increment != 0.
3792 ??? Note that the final_value may overflow and thus final_larger
3793 will be bogus. A potentially infinite loop will be classified
3794 as immediate, e.g. for (i = 0x7ffffff0; i <= 0x7fffffff; i++) */
3796 final_value = plus_constant (final_value, off_by_one);
3798 /* Save the calculated values describing this loop's bounds, in case
3799 precondition_loop_p will need them later. These values can not be
3800 recalculated inside precondition_loop_p because strength reduction
3801 optimizations may obscure the loop's structure.
3803 These values are only required by precondition_loop_p and insert_bct
3804 whenever the number of iterations cannot be computed at compile time.
3805 Only the difference between final_value and initial_value is
3806 important. Note that final_value is only approximate. */
3807 loop_info->initial_value = initial_value;
3808 loop_info->comparison_value = comparison_value;
3809 loop_info->final_value = plus_constant (comparison_value, off_by_one);
3810 loop_info->increment = increment;
3811 loop_info->iteration_var = iteration_var;
3812 loop_info->comparison_code = comparison_code;
3815 /* Try to determine the iteration count for loops such
3816 as (for i = init; i < init + const; i++). When running the
3817 loop optimization twice, the first pass often converts simple
3818 loops into this form. */
3820 if (REG_P (initial_value))
3826 reg1 = initial_value;
3827 if (GET_CODE (final_value) == PLUS)
3828 reg2 = XEXP (final_value, 0), const2 = XEXP (final_value, 1);
3830 reg2 = final_value, const2 = const0_rtx;
3832 /* Check for initial_value = reg1, final_value = reg2 + const2,
3833 where reg1 != reg2. */
3834 if (REG_P (reg2) && reg2 != reg1)
3838 /* Find what reg1 is equivalent to. Hopefully it will
3839 either be reg2 or reg2 plus a constant. */
3840 temp = loop_find_equiv_value (loop, reg1);
3842 if (find_common_reg_term (temp, reg2))
3843 initial_value = temp;
3846 /* Find what reg2 is equivalent to. Hopefully it will
3847 either be reg1 or reg1 plus a constant. Let's ignore
3848 the latter case for now since it is not so common. */
3849 temp = loop_find_equiv_value (loop, reg2);
3851 if (temp == loop_info->iteration_var)
3852 temp = initial_value;
3854 final_value = (const2 == const0_rtx)
3855 ? reg1 : gen_rtx_PLUS (GET_MODE (reg1), reg1, const2);
3858 else if (loop->vtop && GET_CODE (reg2) == CONST_INT)
3862 /* When running the loop optimizer twice, check_dbra_loop
3863 further obfuscates reversible loops of the form:
3864 for (i = init; i < init + const; i++). We often end up with
3865 final_value = 0, initial_value = temp, temp = temp2 - init,
3866 where temp2 = init + const. If the loop has a vtop we
3867 can replace initial_value with const. */
3869 temp = loop_find_equiv_value (loop, reg1);
3871 if (GET_CODE (temp) == MINUS && REG_P (XEXP (temp, 0)))
3873 rtx temp2 = loop_find_equiv_value (loop, XEXP (temp, 0));
3875 if (GET_CODE (temp2) == PLUS
3876 && XEXP (temp2, 0) == XEXP (temp, 1))
3877 initial_value = XEXP (temp2, 1);
3882 /* If have initial_value = reg + const1 and final_value = reg +
3883 const2, then replace initial_value with const1 and final_value
3884 with const2. This should be safe since we are protected by the
3885 initial comparison before entering the loop if we have a vtop.
3886 For example, a + b < a + c is not equivalent to b < c for all a
3887 when using modulo arithmetic.
3889 ??? Without a vtop we could still perform the optimization if we check
3890 the initial and final values carefully. */
3892 && (reg_term = find_common_reg_term (initial_value, final_value)))
3894 initial_value = subtract_reg_term (initial_value, reg_term);
3895 final_value = subtract_reg_term (final_value, reg_term);
3898 loop_info->initial_equiv_value = initial_value;
3899 loop_info->final_equiv_value = final_value;
3901 /* For EQ comparison loops, we don't have a valid final value.
3902 Check this now so that we won't leave an invalid value if we
3903 return early for any other reason. */
3904 if (comparison_code == EQ)
3905 loop_info->final_equiv_value = loop_info->final_value = 0;
3909 if (loop_dump_stream)
3910 fprintf (loop_dump_stream,
3911 "Loop iterations: Increment value can't be calculated.\n");
3915 if (GET_CODE (increment) != CONST_INT)
3917 /* If we have a REG, check to see if REG holds a constant value. */
3918 /* ??? Other RTL, such as (neg (reg)) is possible here, but it isn't
3919 clear if it is worthwhile to try to handle such RTL. */
3920 if (GET_CODE (increment) == REG || GET_CODE (increment) == SUBREG)
3921 increment = loop_find_equiv_value (loop, increment);
3923 if (GET_CODE (increment) != CONST_INT)
3925 if (loop_dump_stream)
3927 fprintf (loop_dump_stream,
3928 "Loop iterations: Increment value not constant ");
3929 print_simple_rtl (loop_dump_stream, increment);
3930 fprintf (loop_dump_stream, ".\n");
3934 loop_info->increment = increment;
3937 if (GET_CODE (initial_value) != CONST_INT)
3939 if (loop_dump_stream)
3941 fprintf (loop_dump_stream,
3942 "Loop iterations: Initial value not constant ");
3943 print_simple_rtl (loop_dump_stream, initial_value);
3944 fprintf (loop_dump_stream, ".\n");
3948 else if (comparison_code == EQ)
3950 if (loop_dump_stream)
3951 fprintf (loop_dump_stream, "Loop iterations: EQ comparison loop.\n");
3954 else if (GET_CODE (final_value) != CONST_INT)
3956 if (loop_dump_stream)
3958 fprintf (loop_dump_stream,
3959 "Loop iterations: Final value not constant ");
3960 print_simple_rtl (loop_dump_stream, final_value);
3961 fprintf (loop_dump_stream, ".\n");
3966 /* Final_larger is 1 if final larger, 0 if they are equal, otherwise -1. */
3969 = ((unsigned HOST_WIDE_INT) INTVAL (final_value)
3970 > (unsigned HOST_WIDE_INT) INTVAL (initial_value))
3971 - ((unsigned HOST_WIDE_INT) INTVAL (final_value)
3972 < (unsigned HOST_WIDE_INT) INTVAL (initial_value));
3974 final_larger = (INTVAL (final_value) > INTVAL (initial_value))
3975 - (INTVAL (final_value) < INTVAL (initial_value));
3977 if (INTVAL (increment) > 0)
3979 else if (INTVAL (increment) == 0)
3984 /* There are 27 different cases: compare_dir = -1, 0, 1;
3985 final_larger = -1, 0, 1; increment_dir = -1, 0, 1.
3986 There are 4 normal cases, 4 reverse cases (where the iteration variable
3987 will overflow before the loop exits), 4 infinite loop cases, and 15
3988 immediate exit (0 or 1 iteration depending on loop type) cases.
3989 Only try to optimize the normal cases. */
3991 /* (compare_dir/final_larger/increment_dir)
3992 Normal cases: (0/-1/-1), (0/1/1), (-1/-1/-1), (1/1/1)
3993 Reverse cases: (0/-1/1), (0/1/-1), (-1/-1/1), (1/1/-1)
3994 Infinite loops: (0/-1/0), (0/1/0), (-1/-1/0), (1/1/0)
3995 Immediate exit: (0/0/X), (-1/0/X), (-1/1/X), (1/0/X), (1/-1/X) */
3997 /* ?? If the meaning of reverse loops (where the iteration variable
3998 will overflow before the loop exits) is undefined, then could
3999 eliminate all of these special checks, and just always assume
4000 the loops are normal/immediate/infinite. Note that this means
4001 the sign of increment_dir does not have to be known. Also,
4002 since it does not really hurt if immediate exit loops or infinite loops
4003 are optimized, then that case could be ignored also, and hence all
4004 loops can be optimized.
4006 According to ANSI Spec, the reverse loop case result is undefined,
4007 because the action on overflow is undefined.
4009 See also the special test for NE loops below. */
4011 if (final_larger == increment_dir && final_larger != 0
4012 && (final_larger == compare_dir || compare_dir == 0))
4017 if (loop_dump_stream)
4018 fprintf (loop_dump_stream, "Loop iterations: Not normal loop.\n");
4022 /* Calculate the number of iterations, final_value is only an approximation,
4023 so correct for that. Note that abs_diff and n_iterations are
4024 unsigned, because they can be as large as 2^n - 1. */
4026 inc = INTVAL (increment);
4029 abs_diff = INTVAL (final_value) - INTVAL (initial_value);
4034 abs_diff = INTVAL (initial_value) - INTVAL (final_value);
4040 /* Given that iteration_var is going to iterate over its own mode,
4041 not HOST_WIDE_INT, disregard higher bits that might have come
4042 into the picture due to sign extension of initial and final
4044 abs_diff &= ((unsigned HOST_WIDE_INT) 1
4045 << (GET_MODE_BITSIZE (GET_MODE (iteration_var)) - 1)
4048 /* For NE tests, make sure that the iteration variable won't miss
4049 the final value. If abs_diff mod abs_incr is not zero, then the
4050 iteration variable will overflow before the loop exits, and we
4051 can not calculate the number of iterations. */
4052 if (compare_dir == 0 && (abs_diff % abs_inc) != 0)
4055 /* Note that the number of iterations could be calculated using
4056 (abs_diff + abs_inc - 1) / abs_inc, provided care was taken to
4057 handle potential overflow of the summation. */
4058 loop_info->n_iterations = abs_diff / abs_inc + ((abs_diff % abs_inc) != 0);
4059 return loop_info->n_iterations;
4062 /* Replace uses of split bivs with their split pseudo register. This is
4063 for original instructions which remain after loop unrolling without
4067 remap_split_bivs (loop, x)
4071 struct loop_ivs *ivs = LOOP_IVS (loop);
4079 code = GET_CODE (x);
4094 /* If non-reduced/final-value givs were split, then this would also
4095 have to remap those givs also. */
4097 if (REGNO (x) < ivs->n_regs
4098 && REG_IV_TYPE (ivs, REGNO (x)) == BASIC_INDUCT)
4099 return REG_IV_CLASS (ivs, REGNO (x))->biv->src_reg;
4106 fmt = GET_RTX_FORMAT (code);
4107 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4110 XEXP (x, i) = remap_split_bivs (loop, XEXP (x, i));
4111 else if (fmt[i] == 'E')
4114 for (j = 0; j < XVECLEN (x, i); j++)
4115 XVECEXP (x, i, j) = remap_split_bivs (loop, XVECEXP (x, i, j));
4121 /* If FIRST_UID is a set of REGNO, and FIRST_UID dominates LAST_UID (e.g.
4122 FIST_UID is always executed if LAST_UID is), then return 1. Otherwise
4123 return 0. COPY_START is where we can start looking for the insns
4124 FIRST_UID and LAST_UID. COPY_END is where we stop looking for these
4127 If there is no JUMP_INSN between LOOP_START and FIRST_UID, then FIRST_UID
4128 must dominate LAST_UID.
4130 If there is a CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID
4131 may not dominate LAST_UID.
4133 If there is no CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID
4134 must dominate LAST_UID. */
4137 set_dominates_use (regno, first_uid, last_uid, copy_start, copy_end)
4144 int passed_jump = 0;
4145 rtx p = NEXT_INSN (copy_start);
4147 while (INSN_UID (p) != first_uid)
4149 if (GET_CODE (p) == JUMP_INSN)
4151 /* Could not find FIRST_UID. */
4157 /* Verify that FIRST_UID is an insn that entirely sets REGNO. */
4158 if (! INSN_P (p) || ! dead_or_set_regno_p (p, regno))
4161 /* FIRST_UID is always executed. */
4162 if (passed_jump == 0)
4165 while (INSN_UID (p) != last_uid)
4167 /* If we see a CODE_LABEL between FIRST_UID and LAST_UID, then we
4168 can not be sure that FIRST_UID dominates LAST_UID. */
4169 if (GET_CODE (p) == CODE_LABEL)
4171 /* Could not find LAST_UID, but we reached the end of the loop, so
4173 else if (p == copy_end)
4178 /* FIRST_UID is always executed if LAST_UID is executed. */
4182 /* This routine is called when the number of iterations for the unrolled
4183 loop is one. The goal is to identify a loop that begins with an
4184 unconditional branch to the loop continuation note (or a label just after).
4185 In this case, the unconditional branch that starts the loop needs to be
4186 deleted so that we execute the single iteration. */
4189 ujump_to_loop_cont (loop_start, loop_cont)
4193 rtx x, label, label_ref;
4195 /* See if loop start, or the next insn is an unconditional jump. */
4196 loop_start = next_nonnote_insn (loop_start);
4198 x = pc_set (loop_start);
4202 label_ref = SET_SRC (x);
4206 /* Examine insn after loop continuation note. Return if not a label. */
4207 label = next_nonnote_insn (loop_cont);
4208 if (label == 0 || GET_CODE (label) != CODE_LABEL)
4211 /* Return the loop start if the branch label matches the code label. */
4212 if (CODE_LABEL_NUMBER (label) == CODE_LABEL_NUMBER (XEXP (label_ref, 0)))