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
= get_insns ();
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 /* And whether the loop has been preconditioned. */
1139 loop_info
->preconditioned
= loop_preconditioned
;
1141 /* For each biv and giv, determine whether it can be safely split into
1142 a different variable for each unrolled copy of the loop body.
1143 We precalculate and save this info here, since computing it is
1146 Do this before deleting any instructions from the loop, so that
1147 back_branch_in_range_p will work correctly. */
1149 if (splitting_not_safe
)
1152 temp
= find_splittable_regs (loop
, unroll_type
, unroll_number
);
1154 /* find_splittable_regs may have created some new registers, so must
1155 reallocate the reg_map with the new larger size, and must realloc
1156 the constant maps also. */
1158 maxregnum
= max_reg_num ();
1159 map
->reg_map
= (rtx
*) xmalloc (maxregnum
* sizeof (rtx
));
1161 init_reg_map (map
, maxregnum
);
1163 if (map
->const_equiv_varray
== 0)
1164 VARRAY_CONST_EQUIV_INIT (map
->const_equiv_varray
,
1165 maxregnum
+ temp
* unroll_number
* 2,
1167 global_const_equiv_varray
= map
->const_equiv_varray
;
1169 /* Search the list of bivs and givs to find ones which need to be remapped
1170 when split, and set their reg_map entry appropriately. */
1172 for (bl
= ivs
->list
; bl
; bl
= bl
->next
)
1174 if (REGNO (bl
->biv
->src_reg
) != bl
->regno
)
1175 map
->reg_map
[bl
->regno
] = bl
->biv
->src_reg
;
1177 /* Currently, non-reduced/final-value givs are never split. */
1178 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
1179 if (REGNO (v
->src_reg
) != bl
->regno
)
1180 map
->reg_map
[REGNO (v
->dest_reg
)] = v
->src_reg
;
1184 /* Use our current register alignment and pointer flags. */
1185 map
->regno_pointer_align
= cfun
->emit
->regno_pointer_align
;
1186 map
->x_regno_reg_rtx
= cfun
->emit
->x_regno_reg_rtx
;
1188 /* If the loop is being partially unrolled, and the iteration variables
1189 are being split, and are being renamed for the split, then must fix up
1190 the compare/jump instruction at the end of the loop to refer to the new
1191 registers. This compare isn't copied, so the registers used in it
1192 will never be replaced if it isn't done here. */
1194 if (unroll_type
== UNROLL_MODULO
)
1196 insn
= NEXT_INSN (copy_end
);
1197 if (GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == JUMP_INSN
)
1198 PATTERN (insn
) = remap_split_bivs (loop
, PATTERN (insn
));
1201 /* For unroll_number times, make a copy of each instruction
1202 between copy_start and copy_end, and insert these new instructions
1203 before the end of the loop. */
1205 for (i
= 0; i
< unroll_number
; i
++)
1207 memset ((char *) map
->insn_map
, 0, max_insnno
* sizeof (rtx
));
1208 memset ((char *) &VARRAY_CONST_EQUIV (map
->const_equiv_varray
, 0), 0,
1209 VARRAY_SIZE (map
->const_equiv_varray
) * sizeof (struct const_equiv_data
));
1212 for (j
= 0; j
< max_labelno
; j
++)
1214 set_label_in_map (map
, j
, gen_label_rtx ());
1216 for (r
= FIRST_PSEUDO_REGISTER
; r
< max_local_regnum
; r
++)
1219 map
->reg_map
[r
] = gen_reg_rtx (GET_MODE (regno_reg_rtx
[r
]));
1220 record_base_value (REGNO (map
->reg_map
[r
]),
1221 regno_reg_rtx
[r
], 0);
1224 /* If loop starts with a branch to the test, then fix it so that
1225 it points to the test of the first unrolled copy of the loop. */
1226 if (i
== 0 && loop_start
!= copy_start
)
1228 insn
= PREV_INSN (copy_start
);
1229 pattern
= PATTERN (insn
);
1231 tem
= get_label_from_map (map
,
1233 (XEXP (SET_SRC (pattern
), 0)));
1234 SET_SRC (pattern
) = gen_rtx_LABEL_REF (VOIDmode
, tem
);
1236 /* Set the jump label so that it can be used by later loop unrolling
1238 JUMP_LABEL (insn
) = tem
;
1239 LABEL_NUSES (tem
)++;
1242 copy_loop_body (loop
, copy_start
, copy_end
, map
, exit_label
,
1243 i
== unroll_number
- 1, unroll_type
, start_label
,
1244 loop_end
, insert_before
, insert_before
);
1247 /* Before deleting any insns, emit a CODE_LABEL immediately after the last
1248 insn to be deleted. This prevents any runaway delete_insn call from
1249 more insns that it should, as it always stops at a CODE_LABEL. */
1251 /* Delete the compare and branch at the end of the loop if completely
1252 unrolling the loop. Deleting the backward branch at the end also
1253 deletes the code label at the start of the loop. This is done at
1254 the very end to avoid problems with back_branch_in_range_p. */
1256 if (unroll_type
== UNROLL_COMPLETELY
)
1257 safety_label
= emit_label_after (gen_label_rtx (), last_loop_insn
);
1259 safety_label
= emit_label_after (gen_label_rtx (), copy_end
);
1261 /* Delete all of the original loop instructions. Don't delete the
1262 LOOP_BEG note, or the first code label in the loop. */
1264 insn
= NEXT_INSN (copy_start
);
1265 while (insn
!= safety_label
)
1267 /* ??? Don't delete named code labels. They will be deleted when the
1268 jump that references them is deleted. Otherwise, we end up deleting
1269 them twice, which causes them to completely disappear instead of turn
1270 into NOTE_INSN_DELETED_LABEL notes. This in turn causes aborts in
1271 dwarfout.c/dwarf2out.c. We could perhaps fix the dwarf*out.c files
1272 to handle deleted labels instead. Or perhaps fix DECL_RTL of the
1273 associated LABEL_DECL to point to one of the new label instances. */
1274 /* ??? Likewise, we can't delete a NOTE_INSN_DELETED_LABEL note. */
1275 if (insn
!= start_label
1276 && ! (GET_CODE (insn
) == CODE_LABEL
&& LABEL_NAME (insn
))
1277 && ! (GET_CODE (insn
) == NOTE
1278 && NOTE_LINE_NUMBER (insn
) == NOTE_INSN_DELETED_LABEL
))
1279 insn
= delete_related_insns (insn
);
1281 insn
= NEXT_INSN (insn
);
1284 /* Can now delete the 'safety' label emitted to protect us from runaway
1285 delete_related_insns calls. */
1286 if (INSN_DELETED_P (safety_label
))
1288 delete_related_insns (safety_label
);
1290 /* If exit_label exists, emit it after the loop. Doing the emit here
1291 forces it to have a higher INSN_UID than any insn in the unrolled loop.
1292 This is needed so that mostly_true_jump in reorg.c will treat jumps
1293 to this loop end label correctly, i.e. predict that they are usually
1296 emit_label_after (exit_label
, loop_end
);
1299 if (unroll_type
== UNROLL_COMPLETELY
)
1301 /* Remove the loop notes since this is no longer a loop. */
1303 delete_related_insns (loop
->vtop
);
1305 delete_related_insns (loop
->cont
);
1307 delete_related_insns (loop_start
);
1309 delete_related_insns (loop_end
);
1312 if (map
->const_equiv_varray
)
1313 VARRAY_FREE (map
->const_equiv_varray
);
1316 free (map
->label_map
);
1319 free (map
->insn_map
);
1320 free (splittable_regs
);
1321 free (splittable_regs_updates
);
1322 free (addr_combined_regs
);
1325 free (map
->reg_map
);
1329 /* Return true if the loop can be safely, and profitably, preconditioned
1330 so that the unrolled copies of the loop body don't need exit tests.
1332 This only works if final_value, initial_value and increment can be
1333 determined, and if increment is a constant power of 2.
1334 If increment is not a power of 2, then the preconditioning modulo
1335 operation would require a real modulo instead of a boolean AND, and this
1336 is not considered `profitable'. */
1338 /* ??? If the loop is known to be executed very many times, or the machine
1339 has a very cheap divide instruction, then preconditioning is a win even
1340 when the increment is not a power of 2. Use RTX_COST to compute
1341 whether divide is cheap.
1342 ??? A divide by constant doesn't actually need a divide, look at
1343 expand_divmod. The reduced cost of this optimized modulo is not
1344 reflected in RTX_COST. */
1347 precondition_loop_p (loop
, initial_value
, final_value
, increment
, mode
)
1348 const struct loop
*loop
;
1349 rtx
*initial_value
, *final_value
, *increment
;
1350 enum machine_mode
*mode
;
1352 rtx loop_start
= loop
->start
;
1353 struct loop_info
*loop_info
= LOOP_INFO (loop
);
1355 if (loop_info
->n_iterations
> 0)
1357 if (INTVAL (loop_info
->increment
) > 0)
1359 *initial_value
= const0_rtx
;
1360 *increment
= const1_rtx
;
1361 *final_value
= GEN_INT (loop_info
->n_iterations
);
1365 *initial_value
= GEN_INT (loop_info
->n_iterations
);
1366 *increment
= constm1_rtx
;
1367 *final_value
= const0_rtx
;
1371 if (loop_dump_stream
)
1373 fputs ("Preconditioning: Success, number of iterations known, ",
1375 fprintf (loop_dump_stream
, HOST_WIDE_INT_PRINT_DEC
,
1376 loop_info
->n_iterations
);
1377 fputs (".\n", loop_dump_stream
);
1382 if (loop_info
->iteration_var
== 0)
1384 if (loop_dump_stream
)
1385 fprintf (loop_dump_stream
,
1386 "Preconditioning: Could not find iteration variable.\n");
1389 else if (loop_info
->initial_value
== 0)
1391 if (loop_dump_stream
)
1392 fprintf (loop_dump_stream
,
1393 "Preconditioning: Could not find initial value.\n");
1396 else if (loop_info
->increment
== 0)
1398 if (loop_dump_stream
)
1399 fprintf (loop_dump_stream
,
1400 "Preconditioning: Could not find increment value.\n");
1403 else if (GET_CODE (loop_info
->increment
) != CONST_INT
)
1405 if (loop_dump_stream
)
1406 fprintf (loop_dump_stream
,
1407 "Preconditioning: Increment not a constant.\n");
1410 else if ((exact_log2 (INTVAL (loop_info
->increment
)) < 0)
1411 && (exact_log2 (-INTVAL (loop_info
->increment
)) < 0))
1413 if (loop_dump_stream
)
1414 fprintf (loop_dump_stream
,
1415 "Preconditioning: Increment not a constant power of 2.\n");
1419 /* Unsigned_compare and compare_dir can be ignored here, since they do
1420 not matter for preconditioning. */
1422 if (loop_info
->final_value
== 0)
1424 if (loop_dump_stream
)
1425 fprintf (loop_dump_stream
,
1426 "Preconditioning: EQ comparison loop.\n");
1430 /* Must ensure that final_value is invariant, so call
1431 loop_invariant_p to check. Before doing so, must check regno
1432 against max_reg_before_loop to make sure that the register is in
1433 the range covered by loop_invariant_p. If it isn't, then it is
1434 most likely a biv/giv which by definition are not invariant. */
1435 if ((GET_CODE (loop_info
->final_value
) == REG
1436 && REGNO (loop_info
->final_value
) >= max_reg_before_loop
)
1437 || (GET_CODE (loop_info
->final_value
) == PLUS
1438 && REGNO (XEXP (loop_info
->final_value
, 0)) >= max_reg_before_loop
)
1439 || ! loop_invariant_p (loop
, loop_info
->final_value
))
1441 if (loop_dump_stream
)
1442 fprintf (loop_dump_stream
,
1443 "Preconditioning: Final value not invariant.\n");
1447 /* Fail for floating point values, since the caller of this function
1448 does not have code to deal with them. */
1449 if (GET_MODE_CLASS (GET_MODE (loop_info
->final_value
)) == MODE_FLOAT
1450 || GET_MODE_CLASS (GET_MODE (loop_info
->initial_value
)) == MODE_FLOAT
)
1452 if (loop_dump_stream
)
1453 fprintf (loop_dump_stream
,
1454 "Preconditioning: Floating point final or initial value.\n");
1458 /* Fail if loop_info->iteration_var is not live before loop_start,
1459 since we need to test its value in the preconditioning code. */
1461 if (REGNO_FIRST_LUID (REGNO (loop_info
->iteration_var
))
1462 > INSN_LUID (loop_start
))
1464 if (loop_dump_stream
)
1465 fprintf (loop_dump_stream
,
1466 "Preconditioning: Iteration var not live before loop start.\n");
1470 /* Note that loop_iterations biases the initial value for GIV iterators
1471 such as "while (i-- > 0)" so that we can calculate the number of
1472 iterations just like for BIV iterators.
1474 Also note that the absolute values of initial_value and
1475 final_value are unimportant as only their difference is used for
1476 calculating the number of loop iterations. */
1477 *initial_value
= loop_info
->initial_value
;
1478 *increment
= loop_info
->increment
;
1479 *final_value
= loop_info
->final_value
;
1481 /* Decide what mode to do these calculations in. Choose the larger
1482 of final_value's mode and initial_value's mode, or a full-word if
1483 both are constants. */
1484 *mode
= GET_MODE (*final_value
);
1485 if (*mode
== VOIDmode
)
1487 *mode
= GET_MODE (*initial_value
);
1488 if (*mode
== VOIDmode
)
1491 else if (*mode
!= GET_MODE (*initial_value
)
1492 && (GET_MODE_SIZE (*mode
)
1493 < GET_MODE_SIZE (GET_MODE (*initial_value
))))
1494 *mode
= GET_MODE (*initial_value
);
1497 if (loop_dump_stream
)
1498 fprintf (loop_dump_stream
, "Preconditioning: Successful.\n");
1502 /* All pseudo-registers must be mapped to themselves. Two hard registers
1503 must be mapped, VIRTUAL_STACK_VARS_REGNUM and VIRTUAL_INCOMING_ARGS_
1504 REGNUM, to avoid function-inlining specific conversions of these
1505 registers. All other hard regs can not be mapped because they may be
1510 init_reg_map (map
, maxregnum
)
1511 struct inline_remap
*map
;
1516 for (i
= maxregnum
- 1; i
> LAST_VIRTUAL_REGISTER
; i
--)
1517 map
->reg_map
[i
] = regno_reg_rtx
[i
];
1518 /* Just clear the rest of the entries. */
1519 for (i
= LAST_VIRTUAL_REGISTER
; i
>= 0; i
--)
1520 map
->reg_map
[i
] = 0;
1522 map
->reg_map
[VIRTUAL_STACK_VARS_REGNUM
]
1523 = regno_reg_rtx
[VIRTUAL_STACK_VARS_REGNUM
];
1524 map
->reg_map
[VIRTUAL_INCOMING_ARGS_REGNUM
]
1525 = regno_reg_rtx
[VIRTUAL_INCOMING_ARGS_REGNUM
];
1528 /* Strength-reduction will often emit code for optimized biv/givs which
1529 calculates their value in a temporary register, and then copies the result
1530 to the iv. This procedure reconstructs the pattern computing the iv;
1531 verifying that all operands are of the proper form.
1533 PATTERN must be the result of single_set.
1534 The return value is the amount that the giv is incremented by. */
1537 calculate_giv_inc (pattern
, src_insn
, regno
)
1538 rtx pattern
, src_insn
;
1542 rtx increment_total
= 0;
1546 /* Verify that we have an increment insn here. First check for a plus
1547 as the set source. */
1548 if (GET_CODE (SET_SRC (pattern
)) != PLUS
)
1550 /* SR sometimes computes the new giv value in a temp, then copies it
1552 src_insn
= PREV_INSN (src_insn
);
1553 pattern
= single_set (src_insn
);
1554 if (GET_CODE (SET_SRC (pattern
)) != PLUS
)
1557 /* The last insn emitted is not needed, so delete it to avoid confusing
1558 the second cse pass. This insn sets the giv unnecessarily. */
1559 delete_related_insns (get_last_insn ());
1562 /* Verify that we have a constant as the second operand of the plus. */
1563 increment
= XEXP (SET_SRC (pattern
), 1);
1564 if (GET_CODE (increment
) != CONST_INT
)
1566 /* SR sometimes puts the constant in a register, especially if it is
1567 too big to be an add immed operand. */
1568 increment
= find_last_value (increment
, &src_insn
, NULL_RTX
, 0);
1570 /* SR may have used LO_SUM to compute the constant if it is too large
1571 for a load immed operand. In this case, the constant is in operand
1572 one of the LO_SUM rtx. */
1573 if (GET_CODE (increment
) == LO_SUM
)
1574 increment
= XEXP (increment
, 1);
1576 /* Some ports store large constants in memory and add a REG_EQUAL
1577 note to the store insn. */
1578 else if (GET_CODE (increment
) == MEM
)
1580 rtx note
= find_reg_note (src_insn
, REG_EQUAL
, 0);
1582 increment
= XEXP (note
, 0);
1585 else if (GET_CODE (increment
) == IOR
1586 || GET_CODE (increment
) == ASHIFT
1587 || GET_CODE (increment
) == PLUS
)
1589 /* The rs6000 port loads some constants with IOR.
1590 The alpha port loads some constants with ASHIFT and PLUS. */
1591 rtx second_part
= XEXP (increment
, 1);
1592 enum rtx_code code
= GET_CODE (increment
);
1594 increment
= find_last_value (XEXP (increment
, 0),
1595 &src_insn
, NULL_RTX
, 0);
1596 /* Don't need the last insn anymore. */
1597 delete_related_insns (get_last_insn ());
1599 if (GET_CODE (second_part
) != CONST_INT
1600 || GET_CODE (increment
) != CONST_INT
)
1604 increment
= GEN_INT (INTVAL (increment
) | INTVAL (second_part
));
1605 else if (code
== PLUS
)
1606 increment
= GEN_INT (INTVAL (increment
) + INTVAL (second_part
));
1608 increment
= GEN_INT (INTVAL (increment
) << INTVAL (second_part
));
1611 if (GET_CODE (increment
) != CONST_INT
)
1614 /* The insn loading the constant into a register is no longer needed,
1616 delete_related_insns (get_last_insn ());
1619 if (increment_total
)
1620 increment_total
= GEN_INT (INTVAL (increment_total
) + INTVAL (increment
));
1622 increment_total
= increment
;
1624 /* Check that the source register is the same as the register we expected
1625 to see as the source. If not, something is seriously wrong. */
1626 if (GET_CODE (XEXP (SET_SRC (pattern
), 0)) != REG
1627 || REGNO (XEXP (SET_SRC (pattern
), 0)) != regno
)
1629 /* Some machines (e.g. the romp), may emit two add instructions for
1630 certain constants, so lets try looking for another add immediately
1631 before this one if we have only seen one add insn so far. */
1637 src_insn
= PREV_INSN (src_insn
);
1638 pattern
= single_set (src_insn
);
1640 delete_related_insns (get_last_insn ());
1648 return increment_total
;
1651 /* Copy REG_NOTES, except for insn references, because not all insn_map
1652 entries are valid yet. We do need to copy registers now though, because
1653 the reg_map entries can change during copying. */
1656 initial_reg_note_copy (notes
, map
)
1658 struct inline_remap
*map
;
1665 copy
= rtx_alloc (GET_CODE (notes
));
1666 PUT_REG_NOTE_KIND (copy
, REG_NOTE_KIND (notes
));
1668 if (GET_CODE (notes
) == EXPR_LIST
)
1669 XEXP (copy
, 0) = copy_rtx_and_substitute (XEXP (notes
, 0), map
, 0);
1670 else if (GET_CODE (notes
) == INSN_LIST
)
1671 /* Don't substitute for these yet. */
1672 XEXP (copy
, 0) = copy_rtx (XEXP (notes
, 0));
1676 XEXP (copy
, 1) = initial_reg_note_copy (XEXP (notes
, 1), map
);
1681 /* Fixup insn references in copied REG_NOTES. */
1684 final_reg_note_copy (notesp
, map
)
1686 struct inline_remap
*map
;
1692 if (GET_CODE (note
) == INSN_LIST
)
1694 /* Sometimes, we have a REG_WAS_0 note that points to a
1695 deleted instruction. In that case, we can just delete the
1697 if (REG_NOTE_KIND (note
) == REG_WAS_0
)
1699 *notesp
= XEXP (note
, 1);
1704 rtx insn
= map
->insn_map
[INSN_UID (XEXP (note
, 0))];
1706 /* If we failed to remap the note, something is awry.
1707 Allow REG_LABEL as it may reference label outside
1708 the unrolled loop. */
1711 if (REG_NOTE_KIND (note
) != REG_LABEL
)
1715 XEXP (note
, 0) = insn
;
1719 notesp
= &XEXP (note
, 1);
1723 /* Copy each instruction in the loop, substituting from map as appropriate.
1724 This is very similar to a loop in expand_inline_function. */
1727 copy_loop_body (loop
, copy_start
, copy_end
, map
, exit_label
, last_iteration
,
1728 unroll_type
, start_label
, loop_end
, insert_before
,
1731 rtx copy_start
, copy_end
;
1732 struct inline_remap
*map
;
1735 enum unroll_types unroll_type
;
1736 rtx start_label
, loop_end
, insert_before
, copy_notes_from
;
1738 struct loop_ivs
*ivs
= LOOP_IVS (loop
);
1740 rtx set
, tem
, copy
= NULL_RTX
;
1741 int dest_reg_was_split
, i
;
1745 rtx final_label
= 0;
1746 rtx giv_inc
, giv_dest_reg
, giv_src_reg
;
1748 /* If this isn't the last iteration, then map any references to the
1749 start_label to final_label. Final label will then be emitted immediately
1750 after the end of this loop body if it was ever used.
1752 If this is the last iteration, then map references to the start_label
1754 if (! last_iteration
)
1756 final_label
= gen_label_rtx ();
1757 set_label_in_map (map
, CODE_LABEL_NUMBER (start_label
), final_label
);
1760 set_label_in_map (map
, CODE_LABEL_NUMBER (start_label
), start_label
);
1767 insn
= NEXT_INSN (insn
);
1769 map
->orig_asm_operands_vector
= 0;
1771 switch (GET_CODE (insn
))
1774 pattern
= PATTERN (insn
);
1778 /* Check to see if this is a giv that has been combined with
1779 some split address givs. (Combined in the sense that
1780 `combine_givs' in loop.c has put two givs in the same register.)
1781 In this case, we must search all givs based on the same biv to
1782 find the address givs. Then split the address givs.
1783 Do this before splitting the giv, since that may map the
1784 SET_DEST to a new register. */
1786 if ((set
= single_set (insn
))
1787 && GET_CODE (SET_DEST (set
)) == REG
1788 && addr_combined_regs
[REGNO (SET_DEST (set
))])
1790 struct iv_class
*bl
;
1791 struct induction
*v
, *tv
;
1792 unsigned int regno
= REGNO (SET_DEST (set
));
1794 v
= addr_combined_regs
[REGNO (SET_DEST (set
))];
1795 bl
= REG_IV_CLASS (ivs
, REGNO (v
->src_reg
));
1797 /* Although the giv_inc amount is not needed here, we must call
1798 calculate_giv_inc here since it might try to delete the
1799 last insn emitted. If we wait until later to call it,
1800 we might accidentally delete insns generated immediately
1801 below by emit_unrolled_add. */
1803 giv_inc
= calculate_giv_inc (set
, insn
, regno
);
1805 /* Now find all address giv's that were combined with this
1807 for (tv
= bl
->giv
; tv
; tv
= tv
->next_iv
)
1808 if (tv
->giv_type
== DEST_ADDR
&& tv
->same
== v
)
1812 /* If this DEST_ADDR giv was not split, then ignore it. */
1813 if (*tv
->location
!= tv
->dest_reg
)
1816 /* Scale this_giv_inc if the multiplicative factors of
1817 the two givs are different. */
1818 this_giv_inc
= INTVAL (giv_inc
);
1819 if (tv
->mult_val
!= v
->mult_val
)
1820 this_giv_inc
= (this_giv_inc
/ INTVAL (v
->mult_val
)
1821 * INTVAL (tv
->mult_val
));
1823 tv
->dest_reg
= plus_constant (tv
->dest_reg
, this_giv_inc
);
1824 *tv
->location
= tv
->dest_reg
;
1826 if (last_iteration
&& unroll_type
!= UNROLL_COMPLETELY
)
1828 /* Must emit an insn to increment the split address
1829 giv. Add in the const_adjust field in case there
1830 was a constant eliminated from the address. */
1831 rtx value
, dest_reg
;
1833 /* tv->dest_reg will be either a bare register,
1834 or else a register plus a constant. */
1835 if (GET_CODE (tv
->dest_reg
) == REG
)
1836 dest_reg
= tv
->dest_reg
;
1838 dest_reg
= XEXP (tv
->dest_reg
, 0);
1840 /* Check for shared address givs, and avoid
1841 incrementing the shared pseudo reg more than
1843 if (! tv
->same_insn
&& ! tv
->shared
)
1845 /* tv->dest_reg may actually be a (PLUS (REG)
1846 (CONST)) here, so we must call plus_constant
1847 to add the const_adjust amount before calling
1848 emit_unrolled_add below. */
1849 value
= plus_constant (tv
->dest_reg
,
1852 if (GET_CODE (value
) == PLUS
)
1854 /* The constant could be too large for an add
1855 immediate, so can't directly emit an insn
1857 emit_unrolled_add (dest_reg
, XEXP (value
, 0),
1862 /* Reset the giv to be just the register again, in case
1863 it is used after the set we have just emitted.
1864 We must subtract the const_adjust factor added in
1866 tv
->dest_reg
= plus_constant (dest_reg
,
1868 *tv
->location
= tv
->dest_reg
;
1873 /* If this is a setting of a splittable variable, then determine
1874 how to split the variable, create a new set based on this split,
1875 and set up the reg_map so that later uses of the variable will
1876 use the new split variable. */
1878 dest_reg_was_split
= 0;
1880 if ((set
= single_set (insn
))
1881 && GET_CODE (SET_DEST (set
)) == REG
1882 && splittable_regs
[REGNO (SET_DEST (set
))])
1884 unsigned int regno
= REGNO (SET_DEST (set
));
1885 unsigned int src_regno
;
1887 dest_reg_was_split
= 1;
1889 giv_dest_reg
= SET_DEST (set
);
1890 giv_src_reg
= giv_dest_reg
;
1891 /* Compute the increment value for the giv, if it wasn't
1892 already computed above. */
1894 giv_inc
= calculate_giv_inc (set
, insn
, regno
);
1896 src_regno
= REGNO (giv_src_reg
);
1898 if (unroll_type
== UNROLL_COMPLETELY
)
1900 /* Completely unrolling the loop. Set the induction
1901 variable to a known constant value. */
1903 /* The value in splittable_regs may be an invariant
1904 value, so we must use plus_constant here. */
1905 splittable_regs
[regno
]
1906 = plus_constant (splittable_regs
[src_regno
],
1909 if (GET_CODE (splittable_regs
[regno
]) == PLUS
)
1911 giv_src_reg
= XEXP (splittable_regs
[regno
], 0);
1912 giv_inc
= XEXP (splittable_regs
[regno
], 1);
1916 /* The splittable_regs value must be a REG or a
1917 CONST_INT, so put the entire value in the giv_src_reg
1919 giv_src_reg
= splittable_regs
[regno
];
1920 giv_inc
= const0_rtx
;
1925 /* Partially unrolling loop. Create a new pseudo
1926 register for the iteration variable, and set it to
1927 be a constant plus the original register. Except
1928 on the last iteration, when the result has to
1929 go back into the original iteration var register. */
1931 /* Handle bivs which must be mapped to a new register
1932 when split. This happens for bivs which need their
1933 final value set before loop entry. The new register
1934 for the biv was stored in the biv's first struct
1935 induction entry by find_splittable_regs. */
1937 if (regno
< ivs
->n_regs
1938 && REG_IV_TYPE (ivs
, regno
) == BASIC_INDUCT
)
1940 giv_src_reg
= REG_IV_CLASS (ivs
, regno
)->biv
->src_reg
;
1941 giv_dest_reg
= giv_src_reg
;
1945 /* If non-reduced/final-value givs were split, then
1946 this would have to remap those givs also. See
1947 find_splittable_regs. */
1950 splittable_regs
[regno
]
1951 = simplify_gen_binary (PLUS
, GET_MODE (giv_src_reg
),
1953 splittable_regs
[src_regno
]);
1954 giv_inc
= splittable_regs
[regno
];
1956 /* Now split the induction variable by changing the dest
1957 of this insn to a new register, and setting its
1958 reg_map entry to point to this new register.
1960 If this is the last iteration, and this is the last insn
1961 that will update the iv, then reuse the original dest,
1962 to ensure that the iv will have the proper value when
1963 the loop exits or repeats.
1965 Using splittable_regs_updates here like this is safe,
1966 because it can only be greater than one if all
1967 instructions modifying the iv are always executed in
1970 if (! last_iteration
1971 || (splittable_regs_updates
[regno
]-- != 1))
1973 tem
= gen_reg_rtx (GET_MODE (giv_src_reg
));
1975 map
->reg_map
[regno
] = tem
;
1976 record_base_value (REGNO (tem
),
1977 giv_inc
== const0_rtx
1979 : gen_rtx_PLUS (GET_MODE (giv_src_reg
),
1980 giv_src_reg
, giv_inc
),
1984 map
->reg_map
[regno
] = giv_src_reg
;
1987 /* The constant being added could be too large for an add
1988 immediate, so can't directly emit an insn here. */
1989 emit_unrolled_add (giv_dest_reg
, giv_src_reg
, giv_inc
);
1990 copy
= get_last_insn ();
1991 pattern
= PATTERN (copy
);
1995 pattern
= copy_rtx_and_substitute (pattern
, map
, 0);
1996 copy
= emit_insn (pattern
);
1998 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
1999 INSN_SCOPE (copy
) = INSN_SCOPE (insn
);
2002 /* If this insn is setting CC0, it may need to look at
2003 the insn that uses CC0 to see what type of insn it is.
2004 In that case, the call to recog via validate_change will
2005 fail. So don't substitute constants here. Instead,
2006 do it when we emit the following insn.
2008 For example, see the pyr.md file. That machine has signed and
2009 unsigned compares. The compare patterns must check the
2010 following branch insn to see which what kind of compare to
2013 If the previous insn set CC0, substitute constants on it as
2015 if (sets_cc0_p (PATTERN (copy
)) != 0)
2020 try_constants (cc0_insn
, map
);
2022 try_constants (copy
, map
);
2025 try_constants (copy
, map
);
2028 /* Make split induction variable constants `permanent' since we
2029 know there are no backward branches across iteration variable
2030 settings which would invalidate this. */
2031 if (dest_reg_was_split
)
2033 int regno
= REGNO (SET_DEST (set
));
2035 if ((size_t) regno
< VARRAY_SIZE (map
->const_equiv_varray
)
2036 && (VARRAY_CONST_EQUIV (map
->const_equiv_varray
, regno
).age
2038 VARRAY_CONST_EQUIV (map
->const_equiv_varray
, regno
).age
= -1;
2043 pattern
= copy_rtx_and_substitute (PATTERN (insn
), map
, 0);
2044 copy
= emit_jump_insn (pattern
);
2045 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
2046 INSN_SCOPE (copy
) = INSN_SCOPE (insn
);
2048 if (JUMP_LABEL (insn
))
2050 JUMP_LABEL (copy
) = get_label_from_map (map
,
2052 (JUMP_LABEL (insn
)));
2053 LABEL_NUSES (JUMP_LABEL (copy
))++;
2055 if (JUMP_LABEL (insn
) == start_label
&& insn
== copy_end
2056 && ! last_iteration
)
2059 /* This is a branch to the beginning of the loop; this is the
2060 last insn being copied; and this is not the last iteration.
2061 In this case, we want to change the original fall through
2062 case to be a branch past the end of the loop, and the
2063 original jump label case to fall_through. */
2065 if (!invert_jump (copy
, exit_label
, 0))
2068 rtx lab
= gen_label_rtx ();
2069 /* Can't do it by reversing the jump (probably because we
2070 couldn't reverse the conditions), so emit a new
2071 jump_insn after COPY, and redirect the jump around
2073 jmp
= emit_jump_insn_after (gen_jump (exit_label
), copy
);
2074 JUMP_LABEL (jmp
) = exit_label
;
2075 LABEL_NUSES (exit_label
)++;
2076 jmp
= emit_barrier_after (jmp
);
2077 emit_label_after (lab
, jmp
);
2078 LABEL_NUSES (lab
) = 0;
2079 if (!redirect_jump (copy
, lab
, 0))
2086 try_constants (cc0_insn
, map
);
2089 try_constants (copy
, map
);
2091 /* Set the jump label of COPY correctly to avoid problems with
2092 later passes of unroll_loop, if INSN had jump label set. */
2093 if (JUMP_LABEL (insn
))
2097 /* Can't use the label_map for every insn, since this may be
2098 the backward branch, and hence the label was not mapped. */
2099 if ((set
= single_set (copy
)))
2101 tem
= SET_SRC (set
);
2102 if (GET_CODE (tem
) == LABEL_REF
)
2103 label
= XEXP (tem
, 0);
2104 else if (GET_CODE (tem
) == IF_THEN_ELSE
)
2106 if (XEXP (tem
, 1) != pc_rtx
)
2107 label
= XEXP (XEXP (tem
, 1), 0);
2109 label
= XEXP (XEXP (tem
, 2), 0);
2113 if (label
&& GET_CODE (label
) == CODE_LABEL
)
2114 JUMP_LABEL (copy
) = label
;
2117 /* An unrecognizable jump insn, probably the entry jump
2118 for a switch statement. This label must have been mapped,
2119 so just use the label_map to get the new jump label. */
2121 = get_label_from_map (map
,
2122 CODE_LABEL_NUMBER (JUMP_LABEL (insn
)));
2125 /* If this is a non-local jump, then must increase the label
2126 use count so that the label will not be deleted when the
2127 original jump is deleted. */
2128 LABEL_NUSES (JUMP_LABEL (copy
))++;
2130 else if (GET_CODE (PATTERN (copy
)) == ADDR_VEC
2131 || GET_CODE (PATTERN (copy
)) == ADDR_DIFF_VEC
)
2133 rtx pat
= PATTERN (copy
);
2134 int diff_vec_p
= GET_CODE (pat
) == ADDR_DIFF_VEC
;
2135 int len
= XVECLEN (pat
, diff_vec_p
);
2138 for (i
= 0; i
< len
; i
++)
2139 LABEL_NUSES (XEXP (XVECEXP (pat
, diff_vec_p
, i
), 0))++;
2142 /* If this used to be a conditional jump insn but whose branch
2143 direction is now known, we must do something special. */
2144 if (any_condjump_p (insn
) && onlyjump_p (insn
) && map
->last_pc_value
)
2147 /* If the previous insn set cc0 for us, delete it. */
2148 if (only_sets_cc0_p (PREV_INSN (copy
)))
2149 delete_related_insns (PREV_INSN (copy
));
2152 /* If this is now a no-op, delete it. */
2153 if (map
->last_pc_value
== pc_rtx
)
2159 /* Otherwise, this is unconditional jump so we must put a
2160 BARRIER after it. We could do some dead code elimination
2161 here, but jump.c will do it just as well. */
2167 pattern
= copy_rtx_and_substitute (PATTERN (insn
), map
, 0);
2168 copy
= emit_call_insn (pattern
);
2169 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
2170 INSN_SCOPE (copy
) = INSN_SCOPE (insn
);
2171 SIBLING_CALL_P (copy
) = SIBLING_CALL_P (insn
);
2173 /* Because the USAGE information potentially contains objects other
2174 than hard registers, we need to copy it. */
2175 CALL_INSN_FUNCTION_USAGE (copy
)
2176 = copy_rtx_and_substitute (CALL_INSN_FUNCTION_USAGE (insn
),
2181 try_constants (cc0_insn
, map
);
2184 try_constants (copy
, map
);
2186 /* Be lazy and assume CALL_INSNs clobber all hard registers. */
2187 for (i
= 0; i
< FIRST_PSEUDO_REGISTER
; i
++)
2188 VARRAY_CONST_EQUIV (map
->const_equiv_varray
, i
).rtx
= 0;
2192 /* If this is the loop start label, then we don't need to emit a
2193 copy of this label since no one will use it. */
2195 if (insn
!= start_label
)
2197 copy
= emit_label (get_label_from_map (map
,
2198 CODE_LABEL_NUMBER (insn
)));
2204 copy
= emit_barrier ();
2208 /* VTOP and CONT notes are valid only before the loop exit test.
2209 If placed anywhere else, loop may generate bad code. */
2210 /* BASIC_BLOCK notes exist to stabilize basic block structures with
2211 the associated rtl. We do not want to share the structure in
2214 if (NOTE_LINE_NUMBER (insn
) != NOTE_INSN_DELETED
2215 && NOTE_LINE_NUMBER (insn
) != NOTE_INSN_DELETED_LABEL
2216 && NOTE_LINE_NUMBER (insn
) != NOTE_INSN_BASIC_BLOCK
2217 && ((NOTE_LINE_NUMBER (insn
) != NOTE_INSN_LOOP_VTOP
2218 && NOTE_LINE_NUMBER (insn
) != NOTE_INSN_LOOP_CONT
)
2219 || (last_iteration
&& unroll_type
!= UNROLL_COMPLETELY
)))
2220 copy
= emit_note (NOTE_SOURCE_FILE (insn
),
2221 NOTE_LINE_NUMBER (insn
));
2230 map
->insn_map
[INSN_UID (insn
)] = copy
;
2232 while (insn
!= copy_end
);
2234 /* Now finish coping the REG_NOTES. */
2238 insn
= NEXT_INSN (insn
);
2239 if ((GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == JUMP_INSN
2240 || GET_CODE (insn
) == CALL_INSN
)
2241 && map
->insn_map
[INSN_UID (insn
)])
2242 final_reg_note_copy (®_NOTES (map
->insn_map
[INSN_UID (insn
)]), map
);
2244 while (insn
!= copy_end
);
2246 /* There may be notes between copy_notes_from and loop_end. Emit a copy of
2247 each of these notes here, since there may be some important ones, such as
2248 NOTE_INSN_BLOCK_END notes, in this group. We don't do this on the last
2249 iteration, because the original notes won't be deleted.
2251 We can't use insert_before here, because when from preconditioning,
2252 insert_before points before the loop. We can't use copy_end, because
2253 there may be insns already inserted after it (which we don't want to
2254 copy) when not from preconditioning code. */
2256 if (! last_iteration
)
2258 for (insn
= copy_notes_from
; insn
!= loop_end
; insn
= NEXT_INSN (insn
))
2260 /* VTOP notes are valid only before the loop exit test.
2261 If placed anywhere else, loop may generate bad code.
2262 There is no need to test for NOTE_INSN_LOOP_CONT notes
2263 here, since COPY_NOTES_FROM will be at most one or two (for cc0)
2264 instructions before the last insn in the loop, and if the
2265 end test is that short, there will be a VTOP note between
2266 the CONT note and the test. */
2267 if (GET_CODE (insn
) == NOTE
2268 && NOTE_LINE_NUMBER (insn
) != NOTE_INSN_DELETED
2269 && NOTE_LINE_NUMBER (insn
) != NOTE_INSN_BASIC_BLOCK
2270 && NOTE_LINE_NUMBER (insn
) != NOTE_INSN_LOOP_VTOP
)
2271 emit_note (NOTE_SOURCE_FILE (insn
), NOTE_LINE_NUMBER (insn
));
2275 if (final_label
&& LABEL_NUSES (final_label
) > 0)
2276 emit_label (final_label
);
2280 loop_insn_emit_before (loop
, 0, insert_before
, tem
);
2283 /* Emit an insn, using the expand_binop to ensure that a valid insn is
2284 emitted. This will correctly handle the case where the increment value
2285 won't fit in the immediate field of a PLUS insns. */
2288 emit_unrolled_add (dest_reg
, src_reg
, increment
)
2289 rtx dest_reg
, src_reg
, increment
;
2293 result
= expand_simple_binop (GET_MODE (dest_reg
), PLUS
, src_reg
, increment
,
2294 dest_reg
, 0, OPTAB_LIB_WIDEN
);
2296 if (dest_reg
!= result
)
2297 emit_move_insn (dest_reg
, result
);
2300 /* Searches the insns between INSN and LOOP->END. Returns 1 if there
2301 is a backward branch in that range that branches to somewhere between
2302 LOOP->START and INSN. Returns 0 otherwise. */
2304 /* ??? This is quadratic algorithm. Could be rewritten to be linear.
2305 In practice, this is not a problem, because this function is seldom called,
2306 and uses a negligible amount of CPU time on average. */
2309 back_branch_in_range_p (loop
, insn
)
2310 const struct loop
*loop
;
2313 rtx p
, q
, target_insn
;
2314 rtx loop_start
= loop
->start
;
2315 rtx loop_end
= loop
->end
;
2316 rtx orig_loop_end
= loop
->end
;
2318 /* Stop before we get to the backward branch at the end of the loop. */
2319 loop_end
= prev_nonnote_insn (loop_end
);
2320 if (GET_CODE (loop_end
) == BARRIER
)
2321 loop_end
= PREV_INSN (loop_end
);
2323 /* Check in case insn has been deleted, search forward for first non
2324 deleted insn following it. */
2325 while (INSN_DELETED_P (insn
))
2326 insn
= NEXT_INSN (insn
);
2328 /* Check for the case where insn is the last insn in the loop. Deal
2329 with the case where INSN was a deleted loop test insn, in which case
2330 it will now be the NOTE_LOOP_END. */
2331 if (insn
== loop_end
|| insn
== orig_loop_end
)
2334 for (p
= NEXT_INSN (insn
); p
!= loop_end
; p
= NEXT_INSN (p
))
2336 if (GET_CODE (p
) == JUMP_INSN
)
2338 target_insn
= JUMP_LABEL (p
);
2340 /* Search from loop_start to insn, to see if one of them is
2341 the target_insn. We can't use INSN_LUID comparisons here,
2342 since insn may not have an LUID entry. */
2343 for (q
= loop_start
; q
!= insn
; q
= NEXT_INSN (q
))
2344 if (q
== target_insn
)
2352 /* Try to generate the simplest rtx for the expression
2353 (PLUS (MULT mult1 mult2) add1). This is used to calculate the initial
2357 fold_rtx_mult_add (mult1
, mult2
, add1
, mode
)
2358 rtx mult1
, mult2
, add1
;
2359 enum machine_mode mode
;
2364 /* The modes must all be the same. This should always be true. For now,
2365 check to make sure. */
2366 if ((GET_MODE (mult1
) != mode
&& GET_MODE (mult1
) != VOIDmode
)
2367 || (GET_MODE (mult2
) != mode
&& GET_MODE (mult2
) != VOIDmode
)
2368 || (GET_MODE (add1
) != mode
&& GET_MODE (add1
) != VOIDmode
))
2371 /* Ensure that if at least one of mult1/mult2 are constant, then mult2
2372 will be a constant. */
2373 if (GET_CODE (mult1
) == CONST_INT
)
2380 mult_res
= simplify_binary_operation (MULT
, mode
, mult1
, mult2
);
2382 mult_res
= gen_rtx_MULT (mode
, mult1
, mult2
);
2384 /* Again, put the constant second. */
2385 if (GET_CODE (add1
) == CONST_INT
)
2392 result
= simplify_binary_operation (PLUS
, mode
, add1
, mult_res
);
2394 result
= gen_rtx_PLUS (mode
, add1
, mult_res
);
2399 /* Searches the list of induction struct's for the biv BL, to try to calculate
2400 the total increment value for one iteration of the loop as a constant.
2402 Returns the increment value as an rtx, simplified as much as possible,
2403 if it can be calculated. Otherwise, returns 0. */
2406 biv_total_increment (bl
)
2407 const struct iv_class
*bl
;
2409 struct induction
*v
;
2412 /* For increment, must check every instruction that sets it. Each
2413 instruction must be executed only once each time through the loop.
2414 To verify this, we check that the insn is always executed, and that
2415 there are no backward branches after the insn that branch to before it.
2416 Also, the insn must have a mult_val of one (to make sure it really is
2419 result
= const0_rtx
;
2420 for (v
= bl
->biv
; v
; v
= v
->next_iv
)
2422 if (v
->always_computable
&& v
->mult_val
== const1_rtx
2423 && ! v
->maybe_multiple
)
2424 result
= fold_rtx_mult_add (result
, const1_rtx
, v
->add_val
, v
->mode
);
2432 /* For each biv and giv, determine whether it can be safely split into
2433 a different variable for each unrolled copy of the loop body. If it
2434 is safe to split, then indicate that by saving some useful info
2435 in the splittable_regs array.
2437 If the loop is being completely unrolled, then splittable_regs will hold
2438 the current value of the induction variable while the loop is unrolled.
2439 It must be set to the initial value of the induction variable here.
2440 Otherwise, splittable_regs will hold the difference between the current
2441 value of the induction variable and the value the induction variable had
2442 at the top of the loop. It must be set to the value 0 here.
2444 Returns the total number of instructions that set registers that are
2447 /* ?? If the loop is only unrolled twice, then most of the restrictions to
2448 constant values are unnecessary, since we can easily calculate increment
2449 values in this case even if nothing is constant. The increment value
2450 should not involve a multiply however. */
2452 /* ?? Even if the biv/giv increment values aren't constant, it may still
2453 be beneficial to split the variable if the loop is only unrolled a few
2454 times, since multiplies by small integers (1,2,3,4) are very cheap. */
2457 find_splittable_regs (loop
, unroll_type
, unroll_number
)
2458 const struct loop
*loop
;
2459 enum unroll_types unroll_type
;
2462 struct loop_ivs
*ivs
= LOOP_IVS (loop
);
2463 struct iv_class
*bl
;
2464 struct induction
*v
;
2466 rtx biv_final_value
;
2470 for (bl
= ivs
->list
; bl
; bl
= bl
->next
)
2472 /* Biv_total_increment must return a constant value,
2473 otherwise we can not calculate the split values. */
2475 increment
= biv_total_increment (bl
);
2476 if (! increment
|| GET_CODE (increment
) != CONST_INT
)
2479 /* The loop must be unrolled completely, or else have a known number
2480 of iterations and only one exit, or else the biv must be dead
2481 outside the loop, or else the final value must be known. Otherwise,
2482 it is unsafe to split the biv since it may not have the proper
2483 value on loop exit. */
2485 /* loop_number_exit_count is non-zero if the loop has an exit other than
2486 a fall through at the end. */
2489 biv_final_value
= 0;
2490 if (unroll_type
!= UNROLL_COMPLETELY
2491 && (loop
->exit_count
|| unroll_type
== UNROLL_NAIVE
)
2492 && (REGNO_LAST_LUID (bl
->regno
) >= INSN_LUID (loop
->end
)
2494 || INSN_UID (bl
->init_insn
) >= max_uid_for_loop
2495 || (REGNO_FIRST_LUID (bl
->regno
)
2496 < INSN_LUID (bl
->init_insn
))
2497 || reg_mentioned_p (bl
->biv
->dest_reg
, SET_SRC (bl
->init_set
)))
2498 && ! (biv_final_value
= final_biv_value (loop
, bl
)))
2501 /* If any of the insns setting the BIV don't do so with a simple
2502 PLUS, we don't know how to split it. */
2503 for (v
= bl
->biv
; biv_splittable
&& v
; v
= v
->next_iv
)
2504 if ((tem
= single_set (v
->insn
)) == 0
2505 || GET_CODE (SET_DEST (tem
)) != REG
2506 || REGNO (SET_DEST (tem
)) != bl
->regno
2507 || GET_CODE (SET_SRC (tem
)) != PLUS
)
2510 /* If final value is non-zero, then must emit an instruction which sets
2511 the value of the biv to the proper value. This is done after
2512 handling all of the givs, since some of them may need to use the
2513 biv's value in their initialization code. */
2515 /* This biv is splittable. If completely unrolling the loop, save
2516 the biv's initial value. Otherwise, save the constant zero. */
2518 if (biv_splittable
== 1)
2520 if (unroll_type
== UNROLL_COMPLETELY
)
2522 /* If the initial value of the biv is itself (i.e. it is too
2523 complicated for strength_reduce to compute), or is a hard
2524 register, or it isn't invariant, then we must create a new
2525 pseudo reg to hold the initial value of the biv. */
2527 if (GET_CODE (bl
->initial_value
) == REG
2528 && (REGNO (bl
->initial_value
) == bl
->regno
2529 || REGNO (bl
->initial_value
) < FIRST_PSEUDO_REGISTER
2530 || ! loop_invariant_p (loop
, bl
->initial_value
)))
2532 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2534 record_base_value (REGNO (tem
), bl
->biv
->add_val
, 0);
2535 loop_insn_hoist (loop
,
2536 gen_move_insn (tem
, bl
->biv
->src_reg
));
2538 if (loop_dump_stream
)
2539 fprintf (loop_dump_stream
,
2540 "Biv %d initial value remapped to %d.\n",
2541 bl
->regno
, REGNO (tem
));
2543 splittable_regs
[bl
->regno
] = tem
;
2546 splittable_regs
[bl
->regno
] = bl
->initial_value
;
2549 splittable_regs
[bl
->regno
] = const0_rtx
;
2551 /* Save the number of instructions that modify the biv, so that
2552 we can treat the last one specially. */
2554 splittable_regs_updates
[bl
->regno
] = bl
->biv_count
;
2555 result
+= bl
->biv_count
;
2557 if (loop_dump_stream
)
2558 fprintf (loop_dump_stream
,
2559 "Biv %d safe to split.\n", bl
->regno
);
2562 /* Check every giv that depends on this biv to see whether it is
2563 splittable also. Even if the biv isn't splittable, givs which
2564 depend on it may be splittable if the biv is live outside the
2565 loop, and the givs aren't. */
2567 result
+= find_splittable_givs (loop
, bl
, unroll_type
, increment
,
2570 /* If final value is non-zero, then must emit an instruction which sets
2571 the value of the biv to the proper value. This is done after
2572 handling all of the givs, since some of them may need to use the
2573 biv's value in their initialization code. */
2574 if (biv_final_value
)
2576 /* If the loop has multiple exits, emit the insns before the
2577 loop to ensure that it will always be executed no matter
2578 how the loop exits. Otherwise emit the insn after the loop,
2579 since this is slightly more efficient. */
2580 if (! loop
->exit_count
)
2581 loop_insn_sink (loop
, gen_move_insn (bl
->biv
->src_reg
,
2585 /* Create a new register to hold the value of the biv, and then
2586 set the biv to its final value before the loop start. The biv
2587 is set to its final value before loop start to ensure that
2588 this insn will always be executed, no matter how the loop
2590 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2591 record_base_value (REGNO (tem
), bl
->biv
->add_val
, 0);
2593 loop_insn_hoist (loop
, gen_move_insn (tem
, bl
->biv
->src_reg
));
2594 loop_insn_hoist (loop
, gen_move_insn (bl
->biv
->src_reg
,
2597 if (loop_dump_stream
)
2598 fprintf (loop_dump_stream
, "Biv %d mapped to %d for split.\n",
2599 REGNO (bl
->biv
->src_reg
), REGNO (tem
));
2601 /* Set up the mapping from the original biv register to the new
2603 bl
->biv
->src_reg
= tem
;
2610 /* Return 1 if the first and last unrolled copy of the address giv V is valid
2611 for the instruction that is using it. Do not make any changes to that
2615 verify_addresses (v
, giv_inc
, unroll_number
)
2616 struct induction
*v
;
2621 rtx orig_addr
= *v
->location
;
2622 rtx last_addr
= plus_constant (v
->dest_reg
,
2623 INTVAL (giv_inc
) * (unroll_number
- 1));
2625 /* First check to see if either address would fail. Handle the fact
2626 that we have may have a match_dup. */
2627 if (! validate_replace_rtx (*v
->location
, v
->dest_reg
, v
->insn
)
2628 || ! validate_replace_rtx (*v
->location
, last_addr
, v
->insn
))
2631 /* Now put things back the way they were before. This should always
2633 if (! validate_replace_rtx (*v
->location
, orig_addr
, v
->insn
))
2639 /* For every giv based on the biv BL, check to determine whether it is
2640 splittable. This is a subroutine to find_splittable_regs ().
2642 Return the number of instructions that set splittable registers. */
2645 find_splittable_givs (loop
, bl
, unroll_type
, increment
, unroll_number
)
2646 const struct loop
*loop
;
2647 struct iv_class
*bl
;
2648 enum unroll_types unroll_type
;
2652 struct loop_ivs
*ivs
= LOOP_IVS (loop
);
2653 struct induction
*v
, *v2
;
2658 /* Scan the list of givs, and set the same_insn field when there are
2659 multiple identical givs in the same insn. */
2660 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
2661 for (v2
= v
->next_iv
; v2
; v2
= v2
->next_iv
)
2662 if (v
->insn
== v2
->insn
&& rtx_equal_p (v
->new_reg
, v2
->new_reg
)
2666 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
2670 /* Only split the giv if it has already been reduced, or if the loop is
2671 being completely unrolled. */
2672 if (unroll_type
!= UNROLL_COMPLETELY
&& v
->ignore
)
2675 /* The giv can be split if the insn that sets the giv is executed once
2676 and only once on every iteration of the loop. */
2677 /* An address giv can always be split. v->insn is just a use not a set,
2678 and hence it does not matter whether it is always executed. All that
2679 matters is that all the biv increments are always executed, and we
2680 won't reach here if they aren't. */
2681 if (v
->giv_type
!= DEST_ADDR
2682 && (! v
->always_computable
2683 || back_branch_in_range_p (loop
, v
->insn
)))
2686 /* The giv increment value must be a constant. */
2687 giv_inc
= fold_rtx_mult_add (v
->mult_val
, increment
, const0_rtx
,
2689 if (! giv_inc
|| GET_CODE (giv_inc
) != CONST_INT
)
2692 /* The loop must be unrolled completely, or else have a known number of
2693 iterations and only one exit, or else the giv must be dead outside
2694 the loop, or else the final value of the giv must be known.
2695 Otherwise, it is not safe to split the giv since it may not have the
2696 proper value on loop exit. */
2698 /* The used outside loop test will fail for DEST_ADDR givs. They are
2699 never used outside the loop anyways, so it is always safe to split a
2703 if (unroll_type
!= UNROLL_COMPLETELY
2704 && (loop
->exit_count
|| unroll_type
== UNROLL_NAIVE
)
2705 && v
->giv_type
!= DEST_ADDR
2706 /* The next part is true if the pseudo is used outside the loop.
2707 We assume that this is true for any pseudo created after loop
2708 starts, because we don't have a reg_n_info entry for them. */
2709 && (REGNO (v
->dest_reg
) >= max_reg_before_loop
2710 || (REGNO_FIRST_UID (REGNO (v
->dest_reg
)) != INSN_UID (v
->insn
)
2711 /* Check for the case where the pseudo is set by a shift/add
2712 sequence, in which case the first insn setting the pseudo
2713 is the first insn of the shift/add sequence. */
2714 && (! (tem
= find_reg_note (v
->insn
, REG_RETVAL
, NULL_RTX
))
2715 || (REGNO_FIRST_UID (REGNO (v
->dest_reg
))
2716 != INSN_UID (XEXP (tem
, 0)))))
2717 /* Line above always fails if INSN was moved by loop opt. */
2718 || (REGNO_LAST_LUID (REGNO (v
->dest_reg
))
2719 >= INSN_LUID (loop
->end
)))
2720 && ! (final_value
= v
->final_value
))
2724 /* Currently, non-reduced/final-value givs are never split. */
2725 /* Should emit insns after the loop if possible, as the biv final value
2728 /* If the final value is non-zero, and the giv has not been reduced,
2729 then must emit an instruction to set the final value. */
2730 if (final_value
&& !v
->new_reg
)
2732 /* Create a new register to hold the value of the giv, and then set
2733 the giv to its final value before the loop start. The giv is set
2734 to its final value before loop start to ensure that this insn
2735 will always be executed, no matter how we exit. */
2736 tem
= gen_reg_rtx (v
->mode
);
2737 loop_insn_hoist (loop
, gen_move_insn (tem
, v
->dest_reg
));
2738 loop_insn_hoist (loop
, gen_move_insn (v
->dest_reg
, final_value
));
2740 if (loop_dump_stream
)
2741 fprintf (loop_dump_stream
, "Giv %d mapped to %d for split.\n",
2742 REGNO (v
->dest_reg
), REGNO (tem
));
2748 /* This giv is splittable. If completely unrolling the loop, save the
2749 giv's initial value. Otherwise, save the constant zero for it. */
2751 if (unroll_type
== UNROLL_COMPLETELY
)
2753 /* It is not safe to use bl->initial_value here, because it may not
2754 be invariant. It is safe to use the initial value stored in
2755 the splittable_regs array if it is set. In rare cases, it won't
2756 be set, so then we do exactly the same thing as
2757 find_splittable_regs does to get a safe value. */
2758 rtx biv_initial_value
;
2760 if (splittable_regs
[bl
->regno
])
2761 biv_initial_value
= splittable_regs
[bl
->regno
];
2762 else if (GET_CODE (bl
->initial_value
) != REG
2763 || (REGNO (bl
->initial_value
) != bl
->regno
2764 && REGNO (bl
->initial_value
) >= FIRST_PSEUDO_REGISTER
))
2765 biv_initial_value
= bl
->initial_value
;
2768 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2770 record_base_value (REGNO (tem
), bl
->biv
->add_val
, 0);
2771 loop_insn_hoist (loop
, gen_move_insn (tem
, bl
->biv
->src_reg
));
2772 biv_initial_value
= tem
;
2774 biv_initial_value
= extend_value_for_giv (v
, biv_initial_value
);
2775 value
= fold_rtx_mult_add (v
->mult_val
, biv_initial_value
,
2776 v
->add_val
, v
->mode
);
2783 /* If a giv was combined with another giv, then we can only split
2784 this giv if the giv it was combined with was reduced. This
2785 is because the value of v->new_reg is meaningless in this
2787 if (v
->same
&& ! v
->same
->new_reg
)
2789 if (loop_dump_stream
)
2790 fprintf (loop_dump_stream
,
2791 "giv combined with unreduced giv not split.\n");
2794 /* If the giv is an address destination, it could be something other
2795 than a simple register, these have to be treated differently. */
2796 else if (v
->giv_type
== DEST_REG
)
2798 /* If value is not a constant, register, or register plus
2799 constant, then compute its value into a register before
2800 loop start. This prevents invalid rtx sharing, and should
2801 generate better code. We can use bl->initial_value here
2802 instead of splittable_regs[bl->regno] because this code
2803 is going before the loop start. */
2804 if (unroll_type
== UNROLL_COMPLETELY
2805 && GET_CODE (value
) != CONST_INT
2806 && GET_CODE (value
) != REG
2807 && (GET_CODE (value
) != PLUS
2808 || GET_CODE (XEXP (value
, 0)) != REG
2809 || GET_CODE (XEXP (value
, 1)) != CONST_INT
))
2811 rtx tem
= gen_reg_rtx (v
->mode
);
2812 record_base_value (REGNO (tem
), v
->add_val
, 0);
2813 loop_iv_add_mult_hoist (loop
, bl
->initial_value
, v
->mult_val
,
2818 splittable_regs
[REGNO (v
->new_reg
)] = value
;
2822 /* Splitting address givs is useful since it will often allow us
2823 to eliminate some increment insns for the base giv as
2826 /* If the addr giv is combined with a dest_reg giv, then all
2827 references to that dest reg will be remapped, which is NOT
2828 what we want for split addr regs. We always create a new
2829 register for the split addr giv, just to be safe. */
2831 /* If we have multiple identical address givs within a
2832 single instruction, then use a single pseudo reg for
2833 both. This is necessary in case one is a match_dup
2836 v
->const_adjust
= 0;
2840 v
->dest_reg
= v
->same_insn
->dest_reg
;
2841 if (loop_dump_stream
)
2842 fprintf (loop_dump_stream
,
2843 "Sharing address givs in insn %d\n",
2844 INSN_UID (v
->insn
));
2846 /* If multiple address GIVs have been combined with the
2847 same dest_reg GIV, do not create a new register for
2849 else if (unroll_type
!= UNROLL_COMPLETELY
2850 && v
->giv_type
== DEST_ADDR
2851 && v
->same
&& v
->same
->giv_type
== DEST_ADDR
2852 && v
->same
->unrolled
2853 /* combine_givs_p may return true for some cases
2854 where the add and mult values are not equal.
2855 To share a register here, the values must be
2857 && rtx_equal_p (v
->same
->mult_val
, v
->mult_val
)
2858 && rtx_equal_p (v
->same
->add_val
, v
->add_val
)
2859 /* If the memory references have different modes,
2860 then the address may not be valid and we must
2861 not share registers. */
2862 && verify_addresses (v
, giv_inc
, unroll_number
))
2864 v
->dest_reg
= v
->same
->dest_reg
;
2867 else if (unroll_type
== UNROLL_COMPLETELY
)
2869 v
->dest_reg
= value
;
2871 /* Check the resulting address for validity, and fail
2872 if the resulting address would be invalid. */
2873 if (! verify_addresses (v
, giv_inc
, unroll_number
))
2875 for (v2
= v
->next_iv
; v2
; v2
= v2
->next_iv
)
2876 if (v2
->same_insn
== v
)
2879 if (loop_dump_stream
)
2880 fprintf (loop_dump_stream
,
2881 "Invalid address for giv at insn %d\n",
2882 INSN_UID (v
->insn
));
2889 /* Store the value of dest_reg into the insn. This sharing
2890 will not be a problem as this insn will always be copied
2893 *v
->location
= v
->dest_reg
;
2895 /* If this address giv is combined with a dest reg giv, then
2896 save the base giv's induction pointer so that we will be
2897 able to handle this address giv properly. The base giv
2898 itself does not have to be splittable. */
2900 if (v
->same
&& v
->same
->giv_type
== DEST_REG
)
2901 addr_combined_regs
[REGNO (v
->same
->new_reg
)] = v
->same
;
2903 if (GET_CODE (v
->new_reg
) == REG
)
2905 /* This giv maybe hasn't been combined with any others.
2906 Make sure that it's giv is marked as splittable here. */
2908 splittable_regs
[REGNO (v
->new_reg
)] = value
;
2910 /* Make it appear to depend upon itself, so that the
2911 giv will be properly split in the main loop above. */
2915 addr_combined_regs
[REGNO (v
->new_reg
)] = v
;
2919 if (loop_dump_stream
)
2920 fprintf (loop_dump_stream
, "DEST_ADDR giv being split.\n");
2926 /* Currently, unreduced giv's can't be split. This is not too much
2927 of a problem since unreduced giv's are not live across loop
2928 iterations anyways. When unrolling a loop completely though,
2929 it makes sense to reduce&split givs when possible, as this will
2930 result in simpler instructions, and will not require that a reg
2931 be live across loop iterations. */
2933 splittable_regs
[REGNO (v
->dest_reg
)] = value
;
2934 fprintf (stderr
, "Giv %d at insn %d not reduced\n",
2935 REGNO (v
->dest_reg
), INSN_UID (v
->insn
));
2941 /* Unreduced givs are only updated once by definition. Reduced givs
2942 are updated as many times as their biv is. Mark it so if this is
2943 a splittable register. Don't need to do anything for address givs
2944 where this may not be a register. */
2946 if (GET_CODE (v
->new_reg
) == REG
)
2950 count
= REG_IV_CLASS (ivs
, REGNO (v
->src_reg
))->biv_count
;
2952 splittable_regs_updates
[REGNO (v
->new_reg
)] = count
;
2957 if (loop_dump_stream
)
2961 if (GET_CODE (v
->dest_reg
) == CONST_INT
)
2963 else if (GET_CODE (v
->dest_reg
) != REG
)
2964 regnum
= REGNO (XEXP (v
->dest_reg
, 0));
2966 regnum
= REGNO (v
->dest_reg
);
2967 fprintf (loop_dump_stream
, "Giv %d at insn %d safe to split.\n",
2968 regnum
, INSN_UID (v
->insn
));
2975 /* Try to prove that the register is dead after the loop exits. Trace every
2976 loop exit looking for an insn that will always be executed, which sets
2977 the register to some value, and appears before the first use of the register
2978 is found. If successful, then return 1, otherwise return 0. */
2980 /* ?? Could be made more intelligent in the handling of jumps, so that
2981 it can search past if statements and other similar structures. */
2984 reg_dead_after_loop (loop
, reg
)
2985 const struct loop
*loop
;
2991 int label_count
= 0;
2993 /* In addition to checking all exits of this loop, we must also check
2994 all exits of inner nested loops that would exit this loop. We don't
2995 have any way to identify those, so we just give up if there are any
2996 such inner loop exits. */
2998 for (label
= loop
->exit_labels
; label
; label
= LABEL_NEXTREF (label
))
3001 if (label_count
!= loop
->exit_count
)
3004 /* HACK: Must also search the loop fall through exit, create a label_ref
3005 here which points to the loop->end, and append the loop_number_exit_labels
3007 label
= gen_rtx_LABEL_REF (VOIDmode
, loop
->end
);
3008 LABEL_NEXTREF (label
) = loop
->exit_labels
;
3010 for (; label
; label
= LABEL_NEXTREF (label
))
3012 /* Succeed if find an insn which sets the biv or if reach end of
3013 function. Fail if find an insn that uses the biv, or if come to
3014 a conditional jump. */
3016 insn
= NEXT_INSN (XEXP (label
, 0));
3019 code
= GET_CODE (insn
);
3020 if (GET_RTX_CLASS (code
) == 'i')
3024 if (reg_referenced_p (reg
, PATTERN (insn
)))
3027 set
= single_set (insn
);
3028 if (set
&& rtx_equal_p (SET_DEST (set
), reg
))
3032 if (code
== JUMP_INSN
)
3034 if (GET_CODE (PATTERN (insn
)) == RETURN
)
3036 else if (!any_uncondjump_p (insn
)
3037 /* Prevent infinite loop following infinite loops. */
3038 || jump_count
++ > 20)
3041 insn
= JUMP_LABEL (insn
);
3044 insn
= NEXT_INSN (insn
);
3048 /* Success, the register is dead on all loop exits. */
3052 /* Try to calculate the final value of the biv, the value it will have at
3053 the end of the loop. If we can do it, return that value. */
3056 final_biv_value (loop
, bl
)
3057 const struct loop
*loop
;
3058 struct iv_class
*bl
;
3060 unsigned HOST_WIDE_INT n_iterations
= LOOP_INFO (loop
)->n_iterations
;
3063 /* ??? This only works for MODE_INT biv's. Reject all others for now. */
3065 if (GET_MODE_CLASS (bl
->biv
->mode
) != MODE_INT
)
3068 /* The final value for reversed bivs must be calculated differently than
3069 for ordinary bivs. In this case, there is already an insn after the
3070 loop which sets this biv's final value (if necessary), and there are
3071 no other loop exits, so we can return any value. */
3074 if (loop_dump_stream
)
3075 fprintf (loop_dump_stream
,
3076 "Final biv value for %d, reversed biv.\n", bl
->regno
);
3081 /* Try to calculate the final value as initial value + (number of iterations
3082 * increment). For this to work, increment must be invariant, the only
3083 exit from the loop must be the fall through at the bottom (otherwise
3084 it may not have its final value when the loop exits), and the initial
3085 value of the biv must be invariant. */
3087 if (n_iterations
!= 0
3088 && ! loop
->exit_count
3089 && loop_invariant_p (loop
, bl
->initial_value
))
3091 increment
= biv_total_increment (bl
);
3093 if (increment
&& loop_invariant_p (loop
, increment
))
3095 /* Can calculate the loop exit value, emit insns after loop
3096 end to calculate this value into a temporary register in
3097 case it is needed later. */
3099 tem
= gen_reg_rtx (bl
->biv
->mode
);
3100 record_base_value (REGNO (tem
), bl
->biv
->add_val
, 0);
3101 loop_iv_add_mult_sink (loop
, increment
, GEN_INT (n_iterations
),
3102 bl
->initial_value
, tem
);
3104 if (loop_dump_stream
)
3105 fprintf (loop_dump_stream
,
3106 "Final biv value for %d, calculated.\n", bl
->regno
);
3112 /* Check to see if the biv is dead at all loop exits. */
3113 if (reg_dead_after_loop (loop
, bl
->biv
->src_reg
))
3115 if (loop_dump_stream
)
3116 fprintf (loop_dump_stream
,
3117 "Final biv value for %d, biv dead after loop exit.\n",
3126 /* Try to calculate the final value of the giv, the value it will have at
3127 the end of the loop. If we can do it, return that value. */
3130 final_giv_value (loop
, v
)
3131 const struct loop
*loop
;
3132 struct induction
*v
;
3134 struct loop_ivs
*ivs
= LOOP_IVS (loop
);
3135 struct iv_class
*bl
;
3139 rtx loop_end
= loop
->end
;
3140 unsigned HOST_WIDE_INT n_iterations
= LOOP_INFO (loop
)->n_iterations
;
3142 bl
= REG_IV_CLASS (ivs
, REGNO (v
->src_reg
));
3144 /* The final value for givs which depend on reversed bivs must be calculated
3145 differently than for ordinary givs. In this case, there is already an
3146 insn after the loop which sets this giv's final value (if necessary),
3147 and there are no other loop exits, so we can return any value. */
3150 if (loop_dump_stream
)
3151 fprintf (loop_dump_stream
,
3152 "Final giv value for %d, depends on reversed biv\n",
3153 REGNO (v
->dest_reg
));
3157 /* Try to calculate the final value as a function of the biv it depends
3158 upon. The only exit from the loop must be the fall through at the bottom
3159 and the insn that sets the giv must be executed on every iteration
3160 (otherwise the giv may not have its final value when the loop exits). */
3162 /* ??? Can calculate the final giv value by subtracting off the
3163 extra biv increments times the giv's mult_val. The loop must have
3164 only one exit for this to work, but the loop iterations does not need
3167 if (n_iterations
!= 0
3168 && ! loop
->exit_count
3169 && v
->always_executed
)
3171 /* ?? It is tempting to use the biv's value here since these insns will
3172 be put after the loop, and hence the biv will have its final value
3173 then. However, this fails if the biv is subsequently eliminated.
3174 Perhaps determine whether biv's are eliminable before trying to
3175 determine whether giv's are replaceable so that we can use the
3176 biv value here if it is not eliminable. */
3178 /* We are emitting code after the end of the loop, so we must make
3179 sure that bl->initial_value is still valid then. It will still
3180 be valid if it is invariant. */
3182 increment
= biv_total_increment (bl
);
3184 if (increment
&& loop_invariant_p (loop
, increment
)
3185 && loop_invariant_p (loop
, bl
->initial_value
))
3187 /* Can calculate the loop exit value of its biv as
3188 (n_iterations * increment) + initial_value */
3190 /* The loop exit value of the giv is then
3191 (final_biv_value - extra increments) * mult_val + add_val.
3192 The extra increments are any increments to the biv which
3193 occur in the loop after the giv's value is calculated.
3194 We must search from the insn that sets the giv to the end
3195 of the loop to calculate this value. */
3197 /* Put the final biv value in tem. */
3198 tem
= gen_reg_rtx (v
->mode
);
3199 record_base_value (REGNO (tem
), bl
->biv
->add_val
, 0);
3200 loop_iv_add_mult_sink (loop
, extend_value_for_giv (v
, increment
),
3201 GEN_INT (n_iterations
),
3202 extend_value_for_giv (v
, bl
->initial_value
),
3205 /* Subtract off extra increments as we find them. */
3206 for (insn
= NEXT_INSN (v
->insn
); insn
!= loop_end
;
3207 insn
= NEXT_INSN (insn
))
3209 struct induction
*biv
;
3211 for (biv
= bl
->biv
; biv
; biv
= biv
->next_iv
)
3212 if (biv
->insn
== insn
)
3215 tem
= expand_simple_binop (GET_MODE (tem
), MINUS
, tem
,
3216 biv
->add_val
, NULL_RTX
, 0,
3220 loop_insn_sink (loop
, seq
);
3224 /* Now calculate the giv's final value. */
3225 loop_iv_add_mult_sink (loop
, tem
, v
->mult_val
, v
->add_val
, tem
);
3227 if (loop_dump_stream
)
3228 fprintf (loop_dump_stream
,
3229 "Final giv value for %d, calc from biv's value.\n",
3230 REGNO (v
->dest_reg
));
3236 /* Replaceable giv's should never reach here. */
3240 /* Check to see if the biv is dead at all loop exits. */
3241 if (reg_dead_after_loop (loop
, v
->dest_reg
))
3243 if (loop_dump_stream
)
3244 fprintf (loop_dump_stream
,
3245 "Final giv value for %d, giv dead after loop exit.\n",
3246 REGNO (v
->dest_reg
));
3254 /* Look back before LOOP->START for the insn that sets REG and return
3255 the equivalent constant if there is a REG_EQUAL note otherwise just
3256 the SET_SRC of REG. */
3259 loop_find_equiv_value (loop
, reg
)
3260 const struct loop
*loop
;
3263 rtx loop_start
= loop
->start
;
3268 for (insn
= PREV_INSN (loop_start
); insn
; insn
= PREV_INSN (insn
))
3270 if (GET_CODE (insn
) == CODE_LABEL
)
3273 else if (INSN_P (insn
) && reg_set_p (reg
, insn
))
3275 /* We found the last insn before the loop that sets the register.
3276 If it sets the entire register, and has a REG_EQUAL note,
3277 then use the value of the REG_EQUAL note. */
3278 if ((set
= single_set (insn
))
3279 && (SET_DEST (set
) == reg
))
3281 rtx note
= find_reg_note (insn
, REG_EQUAL
, NULL_RTX
);
3283 /* Only use the REG_EQUAL note if it is a constant.
3284 Other things, divide in particular, will cause
3285 problems later if we use them. */
3286 if (note
&& GET_CODE (XEXP (note
, 0)) != EXPR_LIST
3287 && CONSTANT_P (XEXP (note
, 0)))
3288 ret
= XEXP (note
, 0);
3290 ret
= SET_SRC (set
);
3292 /* We cannot do this if it changes between the
3293 assignment and loop start though. */
3294 if (modified_between_p (ret
, insn
, loop_start
))
3303 /* Return a simplified rtx for the expression OP - REG.
3305 REG must appear in OP, and OP must be a register or the sum of a register
3308 Thus, the return value must be const0_rtx or the second term.
3310 The caller is responsible for verifying that REG appears in OP and OP has
3314 subtract_reg_term (op
, reg
)
3319 if (GET_CODE (op
) == PLUS
)
3321 if (XEXP (op
, 0) == reg
)
3322 return XEXP (op
, 1);
3323 else if (XEXP (op
, 1) == reg
)
3324 return XEXP (op
, 0);
3326 /* OP does not contain REG as a term. */
3330 /* Find and return register term common to both expressions OP0 and
3331 OP1 or NULL_RTX if no such term exists. Each expression must be a
3332 REG or a PLUS of a REG. */
3335 find_common_reg_term (op0
, op1
)
3338 if ((GET_CODE (op0
) == REG
|| GET_CODE (op0
) == PLUS
)
3339 && (GET_CODE (op1
) == REG
|| GET_CODE (op1
) == PLUS
))
3346 if (GET_CODE (op0
) == PLUS
)
3347 op01
= XEXP (op0
, 1), op00
= XEXP (op0
, 0);
3349 op01
= const0_rtx
, op00
= op0
;
3351 if (GET_CODE (op1
) == PLUS
)
3352 op11
= XEXP (op1
, 1), op10
= XEXP (op1
, 0);
3354 op11
= const0_rtx
, op10
= op1
;
3356 /* Find and return common register term if present. */
3357 if (REG_P (op00
) && (op00
== op10
|| op00
== op11
))
3359 else if (REG_P (op01
) && (op01
== op10
|| op01
== op11
))
3363 /* No common register term found. */
3367 /* Determine the loop iterator and calculate the number of loop
3368 iterations. Returns the exact number of loop iterations if it can
3369 be calculated, otherwise returns zero. */
3371 unsigned HOST_WIDE_INT
3372 loop_iterations (loop
)
3375 struct loop_info
*loop_info
= LOOP_INFO (loop
);
3376 struct loop_ivs
*ivs
= LOOP_IVS (loop
);
3377 rtx comparison
, comparison_value
;
3378 rtx iteration_var
, initial_value
, increment
, final_value
;
3379 enum rtx_code comparison_code
;
3381 unsigned HOST_WIDE_INT abs_inc
;
3382 unsigned HOST_WIDE_INT abs_diff
;
3385 int unsigned_p
, compare_dir
, final_larger
;
3388 struct iv_class
*bl
;
3390 loop_info
->n_iterations
= 0;
3391 loop_info
->initial_value
= 0;
3392 loop_info
->initial_equiv_value
= 0;
3393 loop_info
->comparison_value
= 0;
3394 loop_info
->final_value
= 0;
3395 loop_info
->final_equiv_value
= 0;
3396 loop_info
->increment
= 0;
3397 loop_info
->iteration_var
= 0;
3398 loop_info
->unroll_number
= 1;
3401 /* We used to use prev_nonnote_insn here, but that fails because it might
3402 accidentally get the branch for a contained loop if the branch for this
3403 loop was deleted. We can only trust branches immediately before the
3405 last_loop_insn
= PREV_INSN (loop
->end
);
3407 /* ??? We should probably try harder to find the jump insn
3408 at the end of the loop. The following code assumes that
3409 the last loop insn is a jump to the top of the loop. */
3410 if (GET_CODE (last_loop_insn
) != JUMP_INSN
)
3412 if (loop_dump_stream
)
3413 fprintf (loop_dump_stream
,
3414 "Loop iterations: No final conditional branch found.\n");
3418 /* If there is a more than a single jump to the top of the loop
3419 we cannot (easily) determine the iteration count. */
3420 if (LABEL_NUSES (JUMP_LABEL (last_loop_insn
)) > 1)
3422 if (loop_dump_stream
)
3423 fprintf (loop_dump_stream
,
3424 "Loop iterations: Loop has multiple back edges.\n");
3428 /* If there are multiple conditionalized loop exit tests, they may jump
3429 back to differing CODE_LABELs. */
3430 if (loop
->top
&& loop
->cont
)
3432 rtx temp
= PREV_INSN (last_loop_insn
);
3436 if (GET_CODE (temp
) == JUMP_INSN
)
3438 /* There are some kinds of jumps we can't deal with easily. */
3439 if (JUMP_LABEL (temp
) == 0)
3441 if (loop_dump_stream
)
3444 "Loop iterations: Jump insn has null JUMP_LABEL.\n");
3448 if (/* Previous unrolling may have generated new insns not
3449 covered by the uid_luid array. */
3450 INSN_UID (JUMP_LABEL (temp
)) < max_uid_for_loop
3451 /* Check if we jump back into the loop body. */
3452 && INSN_LUID (JUMP_LABEL (temp
)) > INSN_LUID (loop
->top
)
3453 && INSN_LUID (JUMP_LABEL (temp
)) < INSN_LUID (loop
->cont
))
3455 if (loop_dump_stream
)
3458 "Loop iterations: Loop has multiple back edges.\n");
3463 while ((temp
= PREV_INSN (temp
)) != loop
->cont
);
3466 /* Find the iteration variable. If the last insn is a conditional
3467 branch, and the insn before tests a register value, make that the
3468 iteration variable. */
3470 comparison
= get_condition_for_loop (loop
, last_loop_insn
);
3471 if (comparison
== 0)
3473 if (loop_dump_stream
)
3474 fprintf (loop_dump_stream
,
3475 "Loop iterations: No final comparison found.\n");
3479 /* ??? Get_condition may switch position of induction variable and
3480 invariant register when it canonicalizes the comparison. */
3482 comparison_code
= GET_CODE (comparison
);
3483 iteration_var
= XEXP (comparison
, 0);
3484 comparison_value
= XEXP (comparison
, 1);
3486 if (GET_CODE (iteration_var
) != REG
)
3488 if (loop_dump_stream
)
3489 fprintf (loop_dump_stream
,
3490 "Loop iterations: Comparison not against register.\n");
3494 /* The only new registers that are created before loop iterations
3495 are givs made from biv increments or registers created by
3496 load_mems. In the latter case, it is possible that try_copy_prop
3497 will propagate a new pseudo into the old iteration register but
3498 this will be marked by having the REG_USERVAR_P bit set. */
3500 if ((unsigned) REGNO (iteration_var
) >= ivs
->n_regs
3501 && ! REG_USERVAR_P (iteration_var
))
3504 /* Determine the initial value of the iteration variable, and the amount
3505 that it is incremented each loop. Use the tables constructed by
3506 the strength reduction pass to calculate these values. */
3508 /* Clear the result values, in case no answer can be found. */
3512 /* The iteration variable can be either a giv or a biv. Check to see
3513 which it is, and compute the variable's initial value, and increment
3514 value if possible. */
3516 /* If this is a new register, can't handle it since we don't have any
3517 reg_iv_type entry for it. */
3518 if ((unsigned) REGNO (iteration_var
) >= ivs
->n_regs
)
3520 if (loop_dump_stream
)
3521 fprintf (loop_dump_stream
,
3522 "Loop iterations: No reg_iv_type entry for iteration var.\n");
3526 /* Reject iteration variables larger than the host wide int size, since they
3527 could result in a number of iterations greater than the range of our
3528 `unsigned HOST_WIDE_INT' variable loop_info->n_iterations. */
3529 else if ((GET_MODE_BITSIZE (GET_MODE (iteration_var
))
3530 > HOST_BITS_PER_WIDE_INT
))
3532 if (loop_dump_stream
)
3533 fprintf (loop_dump_stream
,
3534 "Loop iterations: Iteration var rejected because mode too large.\n");
3537 else if (GET_MODE_CLASS (GET_MODE (iteration_var
)) != MODE_INT
)
3539 if (loop_dump_stream
)
3540 fprintf (loop_dump_stream
,
3541 "Loop iterations: Iteration var not an integer.\n");
3544 else if (REG_IV_TYPE (ivs
, REGNO (iteration_var
)) == BASIC_INDUCT
)
3546 if (REGNO (iteration_var
) >= ivs
->n_regs
)
3549 /* Grab initial value, only useful if it is a constant. */
3550 bl
= REG_IV_CLASS (ivs
, REGNO (iteration_var
));
3551 initial_value
= bl
->initial_value
;
3552 if (!bl
->biv
->always_executed
|| bl
->biv
->maybe_multiple
)
3554 if (loop_dump_stream
)
3555 fprintf (loop_dump_stream
,
3556 "Loop iterations: Basic induction var not set once in each iteration.\n");
3560 increment
= biv_total_increment (bl
);
3562 else if (REG_IV_TYPE (ivs
, REGNO (iteration_var
)) == GENERAL_INDUCT
)
3564 HOST_WIDE_INT offset
= 0;
3565 struct induction
*v
= REG_IV_INFO (ivs
, REGNO (iteration_var
));
3566 rtx biv_initial_value
;
3568 if (REGNO (v
->src_reg
) >= ivs
->n_regs
)
3571 if (!v
->always_executed
|| v
->maybe_multiple
)
3573 if (loop_dump_stream
)
3574 fprintf (loop_dump_stream
,
3575 "Loop iterations: General induction var not set once in each iteration.\n");
3579 bl
= REG_IV_CLASS (ivs
, REGNO (v
->src_reg
));
3581 /* Increment value is mult_val times the increment value of the biv. */
3583 increment
= biv_total_increment (bl
);
3586 struct induction
*biv_inc
;
3588 increment
= fold_rtx_mult_add (v
->mult_val
,
3589 extend_value_for_giv (v
, increment
),
3590 const0_rtx
, v
->mode
);
3591 /* The caller assumes that one full increment has occurred at the
3592 first loop test. But that's not true when the biv is incremented
3593 after the giv is set (which is the usual case), e.g.:
3594 i = 6; do {;} while (i++ < 9) .
3595 Therefore, we bias the initial value by subtracting the amount of
3596 the increment that occurs between the giv set and the giv test. */
3597 for (biv_inc
= bl
->biv
; biv_inc
; biv_inc
= biv_inc
->next_iv
)
3599 if (loop_insn_first_p (v
->insn
, biv_inc
->insn
))
3601 if (REG_P (biv_inc
->add_val
))
3603 if (loop_dump_stream
)
3604 fprintf (loop_dump_stream
,
3605 "Loop iterations: Basic induction var add_val is REG %d.\n",
3606 REGNO (biv_inc
->add_val
));
3610 offset
-= INTVAL (biv_inc
->add_val
);
3614 if (loop_dump_stream
)
3615 fprintf (loop_dump_stream
,
3616 "Loop iterations: Giv iterator, initial value bias %ld.\n",
3619 /* Initial value is mult_val times the biv's initial value plus
3620 add_val. Only useful if it is a constant. */
3621 biv_initial_value
= extend_value_for_giv (v
, bl
->initial_value
);
3623 = fold_rtx_mult_add (v
->mult_val
,
3624 plus_constant (biv_initial_value
, offset
),
3625 v
->add_val
, v
->mode
);
3629 if (loop_dump_stream
)
3630 fprintf (loop_dump_stream
,
3631 "Loop iterations: Not basic or general induction var.\n");
3635 if (initial_value
== 0)
3640 switch (comparison_code
)
3655 /* Cannot determine loop iterations with this case. */
3674 /* If the comparison value is an invariant register, then try to find
3675 its value from the insns before the start of the loop. */
3677 final_value
= comparison_value
;
3678 if (GET_CODE (comparison_value
) == REG
3679 && loop_invariant_p (loop
, comparison_value
))
3681 final_value
= loop_find_equiv_value (loop
, comparison_value
);
3683 /* If we don't get an invariant final value, we are better
3684 off with the original register. */
3685 if (! loop_invariant_p (loop
, final_value
))
3686 final_value
= comparison_value
;
3689 /* Calculate the approximate final value of the induction variable
3690 (on the last successful iteration). The exact final value
3691 depends on the branch operator, and increment sign. It will be
3692 wrong if the iteration variable is not incremented by one each
3693 time through the loop and (comparison_value + off_by_one -
3694 initial_value) % increment != 0.
3695 ??? Note that the final_value may overflow and thus final_larger
3696 will be bogus. A potentially infinite loop will be classified
3697 as immediate, e.g. for (i = 0x7ffffff0; i <= 0x7fffffff; i++) */
3699 final_value
= plus_constant (final_value
, off_by_one
);
3701 /* Save the calculated values describing this loop's bounds, in case
3702 precondition_loop_p will need them later. These values can not be
3703 recalculated inside precondition_loop_p because strength reduction
3704 optimizations may obscure the loop's structure.
3706 These values are only required by precondition_loop_p and insert_bct
3707 whenever the number of iterations cannot be computed at compile time.
3708 Only the difference between final_value and initial_value is
3709 important. Note that final_value is only approximate. */
3710 loop_info
->initial_value
= initial_value
;
3711 loop_info
->comparison_value
= comparison_value
;
3712 loop_info
->final_value
= plus_constant (comparison_value
, off_by_one
);
3713 loop_info
->increment
= increment
;
3714 loop_info
->iteration_var
= iteration_var
;
3715 loop_info
->comparison_code
= comparison_code
;
3718 /* Try to determine the iteration count for loops such
3719 as (for i = init; i < init + const; i++). When running the
3720 loop optimization twice, the first pass often converts simple
3721 loops into this form. */
3723 if (REG_P (initial_value
))
3729 reg1
= initial_value
;
3730 if (GET_CODE (final_value
) == PLUS
)
3731 reg2
= XEXP (final_value
, 0), const2
= XEXP (final_value
, 1);
3733 reg2
= final_value
, const2
= const0_rtx
;
3735 /* Check for initial_value = reg1, final_value = reg2 + const2,
3736 where reg1 != reg2. */
3737 if (REG_P (reg2
) && reg2
!= reg1
)
3741 /* Find what reg1 is equivalent to. Hopefully it will
3742 either be reg2 or reg2 plus a constant. */
3743 temp
= loop_find_equiv_value (loop
, reg1
);
3745 if (find_common_reg_term (temp
, reg2
))
3746 initial_value
= temp
;
3749 /* Find what reg2 is equivalent to. Hopefully it will
3750 either be reg1 or reg1 plus a constant. Let's ignore
3751 the latter case for now since it is not so common. */
3752 temp
= loop_find_equiv_value (loop
, reg2
);
3754 if (temp
== loop_info
->iteration_var
)
3755 temp
= initial_value
;
3757 final_value
= (const2
== const0_rtx
)
3758 ? reg1
: gen_rtx_PLUS (GET_MODE (reg1
), reg1
, const2
);
3761 else if (loop
->vtop
&& GET_CODE (reg2
) == CONST_INT
)
3765 /* When running the loop optimizer twice, check_dbra_loop
3766 further obfuscates reversible loops of the form:
3767 for (i = init; i < init + const; i++). We often end up with
3768 final_value = 0, initial_value = temp, temp = temp2 - init,
3769 where temp2 = init + const. If the loop has a vtop we
3770 can replace initial_value with const. */
3772 temp
= loop_find_equiv_value (loop
, reg1
);
3774 if (GET_CODE (temp
) == MINUS
&& REG_P (XEXP (temp
, 0)))
3776 rtx temp2
= loop_find_equiv_value (loop
, XEXP (temp
, 0));
3778 if (GET_CODE (temp2
) == PLUS
3779 && XEXP (temp2
, 0) == XEXP (temp
, 1))
3780 initial_value
= XEXP (temp2
, 1);
3785 /* If have initial_value = reg + const1 and final_value = reg +
3786 const2, then replace initial_value with const1 and final_value
3787 with const2. This should be safe since we are protected by the
3788 initial comparison before entering the loop if we have a vtop.
3789 For example, a + b < a + c is not equivalent to b < c for all a
3790 when using modulo arithmetic.
3792 ??? Without a vtop we could still perform the optimization if we check
3793 the initial and final values carefully. */
3795 && (reg_term
= find_common_reg_term (initial_value
, final_value
)))
3797 initial_value
= subtract_reg_term (initial_value
, reg_term
);
3798 final_value
= subtract_reg_term (final_value
, reg_term
);
3801 loop_info
->initial_equiv_value
= initial_value
;
3802 loop_info
->final_equiv_value
= final_value
;
3804 /* For EQ comparison loops, we don't have a valid final value.
3805 Check this now so that we won't leave an invalid value if we
3806 return early for any other reason. */
3807 if (comparison_code
== EQ
)
3808 loop_info
->final_equiv_value
= loop_info
->final_value
= 0;
3812 if (loop_dump_stream
)
3813 fprintf (loop_dump_stream
,
3814 "Loop iterations: Increment value can't be calculated.\n");
3818 if (GET_CODE (increment
) != CONST_INT
)
3820 /* If we have a REG, check to see if REG holds a constant value. */
3821 /* ??? Other RTL, such as (neg (reg)) is possible here, but it isn't
3822 clear if it is worthwhile to try to handle such RTL. */
3823 if (GET_CODE (increment
) == REG
|| GET_CODE (increment
) == SUBREG
)
3824 increment
= loop_find_equiv_value (loop
, increment
);
3826 if (GET_CODE (increment
) != CONST_INT
)
3828 if (loop_dump_stream
)
3830 fprintf (loop_dump_stream
,
3831 "Loop iterations: Increment value not constant ");
3832 print_simple_rtl (loop_dump_stream
, increment
);
3833 fprintf (loop_dump_stream
, ".\n");
3837 loop_info
->increment
= increment
;
3840 if (GET_CODE (initial_value
) != CONST_INT
)
3842 if (loop_dump_stream
)
3844 fprintf (loop_dump_stream
,
3845 "Loop iterations: Initial value not constant ");
3846 print_simple_rtl (loop_dump_stream
, initial_value
);
3847 fprintf (loop_dump_stream
, ".\n");
3851 else if (GET_CODE (final_value
) != CONST_INT
)
3853 if (loop_dump_stream
)
3855 fprintf (loop_dump_stream
,
3856 "Loop iterations: Final value not constant ");
3857 print_simple_rtl (loop_dump_stream
, final_value
);
3858 fprintf (loop_dump_stream
, ".\n");
3862 else if (comparison_code
== EQ
)
3866 if (loop_dump_stream
)
3867 fprintf (loop_dump_stream
, "Loop iterations: EQ comparison loop.\n");
3869 inc_once
= gen_int_mode (INTVAL (initial_value
) + INTVAL (increment
),
3870 GET_MODE (iteration_var
));
3872 if (inc_once
== final_value
)
3874 /* The iterator value once through the loop is equal to the
3875 comparision value. Either we have an infinite loop, or
3876 we'll loop twice. */
3877 if (increment
== const0_rtx
)
3879 loop_info
->n_iterations
= 2;
3882 loop_info
->n_iterations
= 1;
3884 if (GET_CODE (loop_info
->initial_value
) == CONST_INT
)
3885 loop_info
->final_value
3886 = gen_int_mode ((INTVAL (loop_info
->initial_value
)
3887 + loop_info
->n_iterations
* INTVAL (increment
)),
3888 GET_MODE (iteration_var
));
3890 loop_info
->final_value
3891 = plus_constant (loop_info
->initial_value
,
3892 loop_info
->n_iterations
* INTVAL (increment
));
3893 loop_info
->final_equiv_value
3894 = gen_int_mode ((INTVAL (initial_value
)
3895 + loop_info
->n_iterations
* INTVAL (increment
)),
3896 GET_MODE (iteration_var
));
3897 return loop_info
->n_iterations
;
3900 /* Final_larger is 1 if final larger, 0 if they are equal, otherwise -1. */
3903 = ((unsigned HOST_WIDE_INT
) INTVAL (final_value
)
3904 > (unsigned HOST_WIDE_INT
) INTVAL (initial_value
))
3905 - ((unsigned HOST_WIDE_INT
) INTVAL (final_value
)
3906 < (unsigned HOST_WIDE_INT
) INTVAL (initial_value
));
3908 final_larger
= (INTVAL (final_value
) > INTVAL (initial_value
))
3909 - (INTVAL (final_value
) < INTVAL (initial_value
));
3911 if (INTVAL (increment
) > 0)
3913 else if (INTVAL (increment
) == 0)
3918 /* There are 27 different cases: compare_dir = -1, 0, 1;
3919 final_larger = -1, 0, 1; increment_dir = -1, 0, 1.
3920 There are 4 normal cases, 4 reverse cases (where the iteration variable
3921 will overflow before the loop exits), 4 infinite loop cases, and 15
3922 immediate exit (0 or 1 iteration depending on loop type) cases.
3923 Only try to optimize the normal cases. */
3925 /* (compare_dir/final_larger/increment_dir)
3926 Normal cases: (0/-1/-1), (0/1/1), (-1/-1/-1), (1/1/1)
3927 Reverse cases: (0/-1/1), (0/1/-1), (-1/-1/1), (1/1/-1)
3928 Infinite loops: (0/-1/0), (0/1/0), (-1/-1/0), (1/1/0)
3929 Immediate exit: (0/0/X), (-1/0/X), (-1/1/X), (1/0/X), (1/-1/X) */
3931 /* ?? If the meaning of reverse loops (where the iteration variable
3932 will overflow before the loop exits) is undefined, then could
3933 eliminate all of these special checks, and just always assume
3934 the loops are normal/immediate/infinite. Note that this means
3935 the sign of increment_dir does not have to be known. Also,
3936 since it does not really hurt if immediate exit loops or infinite loops
3937 are optimized, then that case could be ignored also, and hence all
3938 loops can be optimized.
3940 According to ANSI Spec, the reverse loop case result is undefined,
3941 because the action on overflow is undefined.
3943 See also the special test for NE loops below. */
3945 if (final_larger
== increment_dir
&& final_larger
!= 0
3946 && (final_larger
== compare_dir
|| compare_dir
== 0))
3951 if (loop_dump_stream
)
3952 fprintf (loop_dump_stream
, "Loop iterations: Not normal loop.\n");
3956 /* Calculate the number of iterations, final_value is only an approximation,
3957 so correct for that. Note that abs_diff and n_iterations are
3958 unsigned, because they can be as large as 2^n - 1. */
3960 inc
= INTVAL (increment
);
3963 abs_diff
= INTVAL (final_value
) - INTVAL (initial_value
);
3968 abs_diff
= INTVAL (initial_value
) - INTVAL (final_value
);
3974 /* Given that iteration_var is going to iterate over its own mode,
3975 not HOST_WIDE_INT, disregard higher bits that might have come
3976 into the picture due to sign extension of initial and final
3978 abs_diff
&= ((unsigned HOST_WIDE_INT
) 1
3979 << (GET_MODE_BITSIZE (GET_MODE (iteration_var
)) - 1)
3982 /* For NE tests, make sure that the iteration variable won't miss
3983 the final value. If abs_diff mod abs_incr is not zero, then the
3984 iteration variable will overflow before the loop exits, and we
3985 can not calculate the number of iterations. */
3986 if (compare_dir
== 0 && (abs_diff
% abs_inc
) != 0)
3989 /* Note that the number of iterations could be calculated using
3990 (abs_diff + abs_inc - 1) / abs_inc, provided care was taken to
3991 handle potential overflow of the summation. */
3992 loop_info
->n_iterations
= abs_diff
/ abs_inc
+ ((abs_diff
% abs_inc
) != 0);
3993 return loop_info
->n_iterations
;
3996 /* Replace uses of split bivs with their split pseudo register. This is
3997 for original instructions which remain after loop unrolling without
4001 remap_split_bivs (loop
, x
)
4005 struct loop_ivs
*ivs
= LOOP_IVS (loop
);
4013 code
= GET_CODE (x
);
4028 /* If non-reduced/final-value givs were split, then this would also
4029 have to remap those givs also. */
4031 if (REGNO (x
) < ivs
->n_regs
4032 && REG_IV_TYPE (ivs
, REGNO (x
)) == BASIC_INDUCT
)
4033 return REG_IV_CLASS (ivs
, REGNO (x
))->biv
->src_reg
;
4040 fmt
= GET_RTX_FORMAT (code
);
4041 for (i
= GET_RTX_LENGTH (code
) - 1; i
>= 0; i
--)
4044 XEXP (x
, i
) = remap_split_bivs (loop
, XEXP (x
, i
));
4045 else if (fmt
[i
] == 'E')
4048 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
4049 XVECEXP (x
, i
, j
) = remap_split_bivs (loop
, XVECEXP (x
, i
, j
));
4055 /* If FIRST_UID is a set of REGNO, and FIRST_UID dominates LAST_UID (e.g.
4056 FIST_UID is always executed if LAST_UID is), then return 1. Otherwise
4057 return 0. COPY_START is where we can start looking for the insns
4058 FIRST_UID and LAST_UID. COPY_END is where we stop looking for these
4061 If there is no JUMP_INSN between LOOP_START and FIRST_UID, then FIRST_UID
4062 must dominate LAST_UID.
4064 If there is a CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID
4065 may not dominate LAST_UID.
4067 If there is no CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID
4068 must dominate LAST_UID. */
4071 set_dominates_use (regno
, first_uid
, last_uid
, copy_start
, copy_end
)
4078 int passed_jump
= 0;
4079 rtx p
= NEXT_INSN (copy_start
);
4081 while (INSN_UID (p
) != first_uid
)
4083 if (GET_CODE (p
) == JUMP_INSN
)
4085 /* Could not find FIRST_UID. */
4091 /* Verify that FIRST_UID is an insn that entirely sets REGNO. */
4092 if (! INSN_P (p
) || ! dead_or_set_regno_p (p
, regno
))
4095 /* FIRST_UID is always executed. */
4096 if (passed_jump
== 0)
4099 while (INSN_UID (p
) != last_uid
)
4101 /* If we see a CODE_LABEL between FIRST_UID and LAST_UID, then we
4102 can not be sure that FIRST_UID dominates LAST_UID. */
4103 if (GET_CODE (p
) == CODE_LABEL
)
4105 /* Could not find LAST_UID, but we reached the end of the loop, so
4107 else if (p
== copy_end
)
4112 /* FIRST_UID is always executed if LAST_UID is executed. */
4116 /* This routine is called when the number of iterations for the unrolled
4117 loop is one. The goal is to identify a loop that begins with an
4118 unconditional branch to the loop continuation note (or a label just after).
4119 In this case, the unconditional branch that starts the loop needs to be
4120 deleted so that we execute the single iteration. */
4123 ujump_to_loop_cont (loop_start
, loop_cont
)
4127 rtx x
, label
, label_ref
;
4129 /* See if loop start, or the next insn is an unconditional jump. */
4130 loop_start
= next_nonnote_insn (loop_start
);
4132 x
= pc_set (loop_start
);
4136 label_ref
= SET_SRC (x
);
4140 /* Examine insn after loop continuation note. Return if not a label. */
4141 label
= next_nonnote_insn (loop_cont
);
4142 if (label
== 0 || GET_CODE (label
) != CODE_LABEL
)
4145 /* Return the loop start if the branch label matches the code label. */
4146 if (CODE_LABEL_NUMBER (label
) == CODE_LABEL_NUMBER (XEXP (label_ref
, 0)))