1 /* Try to unroll loops, and split induction variables.
2 Copyright (C) 1992, 93-95, 1997, 1998 Free Software Foundation, Inc.
3 Contributed by James E. Wilson, Cygnus Support/UC Berkeley.
5 This file is part of GNU CC.
7 GNU CC is free software; you can redistribute it and/or modify
8 it under the terms of the GNU General Public License as published by
9 the Free Software Foundation; either version 2, or (at your option)
12 GNU CC is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License for more details.
17 You should have received a copy of the GNU General Public License
18 along with GNU CC; see the file COPYING. If not, write to
19 the Free Software Foundation, 59 Temple Place - Suite 330,
20 Boston, MA 02111-1307, USA. */
22 /* Try to unroll a loop, and split induction variables.
24 Loops for which the number of iterations can be calculated exactly are
25 handled specially. If the number of iterations times the insn_count is
26 less than MAX_UNROLLED_INSNS, then the loop is unrolled completely.
27 Otherwise, we try to unroll the loop a number of times modulo the number
28 of iterations, so that only one exit test will be needed. It is unrolled
29 a number of times approximately equal to MAX_UNROLLED_INSNS divided by
32 Otherwise, if the number of iterations can be calculated exactly at
33 run time, and the loop is always entered at the top, then we try to
34 precondition the loop. That is, at run time, calculate how many times
35 the loop will execute, and then execute the loop body a few times so
36 that the remaining iterations will be some multiple of 4 (or 2 if the
37 loop is large). Then fall through to a loop unrolled 4 (or 2) times,
38 with only one exit test needed at the end of the loop.
40 Otherwise, if the number of iterations can not be calculated exactly,
41 not even at run time, then we still unroll the loop a number of times
42 approximately equal to MAX_UNROLLED_INSNS divided by the insn count,
43 but there must be an exit test after each copy of the loop body.
45 For each induction variable, which is dead outside the loop (replaceable)
46 or for which we can easily calculate the final value, if we can easily
47 calculate its value at each place where it is set as a function of the
48 current loop unroll count and the variable's value at loop entry, then
49 the induction variable is split into `N' different variables, one for
50 each copy of the loop body. One variable is live across the backward
51 branch, and the others are all calculated as a function of this variable.
52 This helps eliminate data dependencies, and leads to further opportunities
55 /* Possible improvements follow: */
57 /* ??? Add an extra pass somewhere to determine whether unrolling will
58 give any benefit. E.g. after generating all unrolled insns, compute the
59 cost of all insns and compare against cost of insns in rolled loop.
61 - On traditional architectures, unrolling a non-constant bound loop
62 is a win if there is a giv whose only use is in memory addresses, the
63 memory addresses can be split, and hence giv increments can be
65 - It is also a win if the loop is executed many times, and preconditioning
66 can be performed for the loop.
67 Add code to check for these and similar cases. */
69 /* ??? Improve control of which loops get unrolled. Could use profiling
70 info to only unroll the most commonly executed loops. Perhaps have
71 a user specifyable option to control the amount of code expansion,
72 or the percent of loops to consider for unrolling. Etc. */
74 /* ??? Look at the register copies inside the loop to see if they form a
75 simple permutation. If so, iterate the permutation until it gets back to
76 the start state. This is how many times we should unroll the loop, for
77 best results, because then all register copies can be eliminated.
78 For example, the lisp nreverse function should be unrolled 3 times
87 ??? The number of times to unroll the loop may also be based on data
88 references in the loop. For example, if we have a loop that references
89 x[i-1], x[i], and x[i+1], we should unroll it a multiple of 3 times. */
91 /* ??? Add some simple linear equation solving capability so that we can
92 determine the number of loop iterations for more complex loops.
93 For example, consider this loop from gdb
94 #define SWAP_TARGET_AND_HOST(buffer,len)
97 char *p = (char *) buffer;
98 char *q = ((char *) buffer) + len - 1;
99 int iterations = (len + 1) >> 1;
101 for (p; p < q; p++, q--;)
109 start value = p = &buffer + current_iteration
110 end value = q = &buffer + len - 1 - current_iteration
111 Given the loop exit test of "p < q", then there must be "q - p" iterations,
112 set equal to zero and solve for number of iterations:
113 q - p = len - 1 - 2*current_iteration = 0
114 current_iteration = (len - 1) / 2
115 Hence, there are (len - 1) / 2 (rounded up to the nearest integer)
116 iterations of this loop. */
118 /* ??? Currently, no labels are marked as loop invariant when doing loop
119 unrolling. This is because an insn inside the loop, that loads the address
120 of a label inside the loop into a register, could be moved outside the loop
121 by the invariant code motion pass if labels were invariant. If the loop
122 is subsequently unrolled, the code will be wrong because each unrolled
123 body of the loop will use the same address, whereas each actually needs a
124 different address. A case where this happens is when a loop containing
125 a switch statement is unrolled.
127 It would be better to let labels be considered invariant. When we
128 unroll loops here, check to see if any insns using a label local to the
129 loop were moved before the loop. If so, then correct the problem, by
130 moving the insn back into the loop, or perhaps replicate the insn before
131 the loop, one copy for each time the loop is unrolled. */
133 /* The prime factors looked for when trying to unroll a loop by some
134 number which is modulo the total number of iterations. Just checking
135 for these 4 prime factors will find at least one factor for 75% of
136 all numbers theoretically. Practically speaking, this will succeed
137 almost all of the time since loops are generally a multiple of 2
140 #define NUM_FACTORS 4
142 struct _factor
{ int factor
, count
; } factors
[NUM_FACTORS
]
143 = { {2, 0}, {3, 0}, {5, 0}, {7, 0}};
145 /* Describes the different types of loop unrolling performed. */
147 enum unroll_types
{ UNROLL_COMPLETELY
, UNROLL_MODULO
, UNROLL_NAIVE
};
152 #include "insn-config.h"
153 #include "integrate.h"
160 /* This controls which loops are unrolled, and by how much we unroll
163 #ifndef MAX_UNROLLED_INSNS
164 #define MAX_UNROLLED_INSNS 100
167 /* Indexed by register number, if non-zero, then it contains a pointer
168 to a struct induction for a DEST_REG giv which has been combined with
169 one of more address givs. This is needed because whenever such a DEST_REG
170 giv is modified, we must modify the value of all split address givs
171 that were combined with this DEST_REG giv. */
173 static struct induction
**addr_combined_regs
;
175 /* Indexed by register number, if this is a splittable induction variable,
176 then this will hold the current value of the register, which depends on the
179 static rtx
*splittable_regs
;
181 /* Indexed by register number, if this is a splittable induction variable,
182 then this will hold the number of instructions in the loop that modify
183 the induction variable. Used to ensure that only the last insn modifying
184 a split iv will update the original iv of the dest. */
186 static int *splittable_regs_updates
;
188 /* Values describing the current loop's iteration variable. These are set up
189 by loop_iterations, and used by precondition_loop_p. */
191 static rtx loop_iteration_var
;
192 static rtx loop_initial_value
;
193 static rtx loop_increment
;
194 static rtx loop_final_value
;
195 static enum rtx_code loop_comparison_code
;
197 /* Forward declarations. */
199 static void init_reg_map
PROTO((struct inline_remap
*, int));
200 static int precondition_loop_p
PROTO((rtx
*, rtx
*, rtx
*, rtx
, rtx
));
201 static rtx calculate_giv_inc
PROTO((rtx
, rtx
, int));
202 static rtx initial_reg_note_copy
PROTO((rtx
, struct inline_remap
*));
203 static void final_reg_note_copy
PROTO((rtx
, struct inline_remap
*));
204 static void copy_loop_body
PROTO((rtx
, rtx
, struct inline_remap
*, rtx
, int,
205 enum unroll_types
, rtx
, rtx
, rtx
, rtx
));
206 void iteration_info
PROTO((rtx
, rtx
*, rtx
*, rtx
, rtx
));
207 static rtx approx_final_value
PROTO((enum rtx_code
, rtx
, int *, int *));
208 static int find_splittable_regs
PROTO((enum unroll_types
, rtx
, rtx
, rtx
, int));
209 static int find_splittable_givs
PROTO((struct iv_class
*,enum unroll_types
,
210 rtx
, rtx
, rtx
, int));
211 static int reg_dead_after_loop
PROTO((rtx
, rtx
, rtx
));
212 static rtx fold_rtx_mult_add
PROTO((rtx
, rtx
, rtx
, enum machine_mode
));
213 static rtx remap_split_bivs
PROTO((rtx
));
215 /* Try to unroll one loop and split induction variables in the loop.
217 The loop is described by the arguments LOOP_END, INSN_COUNT, and
218 LOOP_START. END_INSERT_BEFORE indicates where insns should be added
219 which need to be executed when the loop falls through. STRENGTH_REDUCTION_P
220 indicates whether information generated in the strength reduction pass
223 This function is intended to be called from within `strength_reduce'
227 unroll_loop (loop_end
, insn_count
, loop_start
, end_insert_before
,
232 rtx end_insert_before
;
233 int strength_reduce_p
;
236 int unroll_number
= 1;
237 rtx copy_start
, copy_end
;
238 rtx insn
, sequence
, pattern
, tem
;
239 int max_labelno
, max_insnno
;
241 struct inline_remap
*map
;
249 int splitting_not_safe
= 0;
250 enum unroll_types unroll_type
;
251 int loop_preconditioned
= 0;
253 /* This points to the last real insn in the loop, which should be either
254 a JUMP_INSN (for conditional jumps) or a BARRIER (for unconditional
258 /* Don't bother unrolling huge loops. Since the minimum factor is
259 two, loops greater than one half of MAX_UNROLLED_INSNS will never
261 if (insn_count
> MAX_UNROLLED_INSNS
/ 2)
263 if (loop_dump_stream
)
264 fprintf (loop_dump_stream
, "Unrolling failure: Loop too big.\n");
268 /* When emitting debugger info, we can't unroll loops with unequal numbers
269 of block_beg and block_end notes, because that would unbalance the block
270 structure of the function. This can happen as a result of the
271 "if (foo) bar; else break;" optimization in jump.c. */
272 /* ??? Gcc has a general policy that -g is never supposed to change the code
273 that the compiler emits, so we must disable this optimization always,
274 even if debug info is not being output. This is rare, so this should
275 not be a significant performance problem. */
277 if (1 /* write_symbols != NO_DEBUG */)
279 int block_begins
= 0;
282 for (insn
= loop_start
; insn
!= loop_end
; insn
= NEXT_INSN (insn
))
284 if (GET_CODE (insn
) == NOTE
)
286 if (NOTE_LINE_NUMBER (insn
) == NOTE_INSN_BLOCK_BEG
)
288 else if (NOTE_LINE_NUMBER (insn
) == NOTE_INSN_BLOCK_END
)
293 if (block_begins
!= block_ends
)
295 if (loop_dump_stream
)
296 fprintf (loop_dump_stream
,
297 "Unrolling failure: Unbalanced block notes.\n");
302 /* Determine type of unroll to perform. Depends on the number of iterations
303 and the size of the loop. */
305 /* If there is no strength reduce info, then set loop_n_iterations to zero.
306 This can happen if strength_reduce can't find any bivs in the loop.
307 A value of zero indicates that the number of iterations could not be
310 if (! strength_reduce_p
)
311 loop_n_iterations
= 0;
313 if (loop_dump_stream
&& loop_n_iterations
> 0)
314 fprintf (loop_dump_stream
,
315 "Loop unrolling: %d iterations.\n", loop_n_iterations
);
317 /* Find and save a pointer to the last nonnote insn in the loop. */
319 last_loop_insn
= prev_nonnote_insn (loop_end
);
321 /* Calculate how many times to unroll the loop. Indicate whether or
322 not the loop is being completely unrolled. */
324 if (loop_n_iterations
== 1)
326 /* If number of iterations is exactly 1, then eliminate the compare and
327 branch at the end of the loop since they will never be taken.
328 Then return, since no other action is needed here. */
330 /* If the last instruction is not a BARRIER or a JUMP_INSN, then
331 don't do anything. */
333 if (GET_CODE (last_loop_insn
) == BARRIER
)
335 /* Delete the jump insn. This will delete the barrier also. */
336 delete_insn (PREV_INSN (last_loop_insn
));
338 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
341 /* The immediately preceding insn is a compare which must be
343 delete_insn (last_loop_insn
);
344 delete_insn (PREV_INSN (last_loop_insn
));
346 /* The immediately preceding insn may not be the compare, so don't
348 delete_insn (last_loop_insn
);
353 else if (loop_n_iterations
> 0
354 && loop_n_iterations
* insn_count
< MAX_UNROLLED_INSNS
)
356 unroll_number
= loop_n_iterations
;
357 unroll_type
= UNROLL_COMPLETELY
;
359 else if (loop_n_iterations
> 0)
361 /* Try to factor the number of iterations. Don't bother with the
362 general case, only using 2, 3, 5, and 7 will get 75% of all
363 numbers theoretically, and almost all in practice. */
365 for (i
= 0; i
< NUM_FACTORS
; i
++)
366 factors
[i
].count
= 0;
368 temp
= loop_n_iterations
;
369 for (i
= NUM_FACTORS
- 1; i
>= 0; i
--)
370 while (temp
% factors
[i
].factor
== 0)
373 temp
= temp
/ factors
[i
].factor
;
376 /* Start with the larger factors first so that we generally
377 get lots of unrolling. */
381 for (i
= 3; i
>= 0; i
--)
382 while (factors
[i
].count
--)
384 if (temp
* factors
[i
].factor
< MAX_UNROLLED_INSNS
)
386 unroll_number
*= factors
[i
].factor
;
387 temp
*= factors
[i
].factor
;
393 /* If we couldn't find any factors, then unroll as in the normal
395 if (unroll_number
== 1)
397 if (loop_dump_stream
)
398 fprintf (loop_dump_stream
,
399 "Loop unrolling: No factors found.\n");
402 unroll_type
= UNROLL_MODULO
;
406 /* Default case, calculate number of times to unroll loop based on its
408 if (unroll_number
== 1)
410 if (8 * insn_count
< MAX_UNROLLED_INSNS
)
412 else if (4 * insn_count
< MAX_UNROLLED_INSNS
)
417 unroll_type
= UNROLL_NAIVE
;
420 /* Now we know how many times to unroll the loop. */
422 if (loop_dump_stream
)
423 fprintf (loop_dump_stream
,
424 "Unrolling loop %d times.\n", unroll_number
);
427 if (unroll_type
== UNROLL_COMPLETELY
|| unroll_type
== UNROLL_MODULO
)
429 /* Loops of these types should never start with a jump down to
430 the exit condition test. For now, check for this case just to
431 be sure. UNROLL_NAIVE loops can be of this form, this case is
434 while (GET_CODE (insn
) != CODE_LABEL
&& GET_CODE (insn
) != JUMP_INSN
)
435 insn
= NEXT_INSN (insn
);
436 if (GET_CODE (insn
) == JUMP_INSN
)
440 if (unroll_type
== UNROLL_COMPLETELY
)
442 /* Completely unrolling the loop: Delete the compare and branch at
443 the end (the last two instructions). This delete must done at the
444 very end of loop unrolling, to avoid problems with calls to
445 back_branch_in_range_p, which is called by find_splittable_regs.
446 All increments of splittable bivs/givs are changed to load constant
449 copy_start
= loop_start
;
451 /* Set insert_before to the instruction immediately after the JUMP_INSN
452 (or BARRIER), so that any NOTEs between the JUMP_INSN and the end of
453 the loop will be correctly handled by copy_loop_body. */
454 insert_before
= NEXT_INSN (last_loop_insn
);
456 /* Set copy_end to the insn before the jump at the end of the loop. */
457 if (GET_CODE (last_loop_insn
) == BARRIER
)
458 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
459 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
462 /* The instruction immediately before the JUMP_INSN is a compare
463 instruction which we do not want to copy. */
464 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
466 /* The instruction immediately before the JUMP_INSN may not be the
467 compare, so we must copy it. */
468 copy_end
= PREV_INSN (last_loop_insn
);
473 /* We currently can't unroll a loop if it doesn't end with a
474 JUMP_INSN. There would need to be a mechanism that recognizes
475 this case, and then inserts a jump after each loop body, which
476 jumps to after the last loop body. */
477 if (loop_dump_stream
)
478 fprintf (loop_dump_stream
,
479 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
483 else if (unroll_type
== UNROLL_MODULO
)
485 /* Partially unrolling the loop: The compare and branch at the end
486 (the last two instructions) must remain. Don't copy the compare
487 and branch instructions at the end of the loop. Insert the unrolled
488 code immediately before the compare/branch at the end so that the
489 code will fall through to them as before. */
491 copy_start
= loop_start
;
493 /* Set insert_before to the jump insn at the end of the loop.
494 Set copy_end to before the jump insn at the end of the loop. */
495 if (GET_CODE (last_loop_insn
) == BARRIER
)
497 insert_before
= PREV_INSN (last_loop_insn
);
498 copy_end
= PREV_INSN (insert_before
);
500 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
503 /* The instruction immediately before the JUMP_INSN is a compare
504 instruction which we do not want to copy or delete. */
505 insert_before
= PREV_INSN (last_loop_insn
);
506 copy_end
= PREV_INSN (insert_before
);
508 /* The instruction immediately before the JUMP_INSN may not be the
509 compare, so we must copy it. */
510 insert_before
= last_loop_insn
;
511 copy_end
= PREV_INSN (last_loop_insn
);
516 /* We currently can't unroll a loop if it doesn't end with a
517 JUMP_INSN. There would need to be a mechanism that recognizes
518 this case, and then inserts a jump after each loop body, which
519 jumps to after the last loop body. */
520 if (loop_dump_stream
)
521 fprintf (loop_dump_stream
,
522 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
528 /* Normal case: Must copy the compare and branch instructions at the
531 if (GET_CODE (last_loop_insn
) == BARRIER
)
533 /* Loop ends with an unconditional jump and a barrier.
534 Handle this like above, don't copy jump and barrier.
535 This is not strictly necessary, but doing so prevents generating
536 unconditional jumps to an immediately following label.
538 This will be corrected below if the target of this jump is
539 not the start_label. */
541 insert_before
= PREV_INSN (last_loop_insn
);
542 copy_end
= PREV_INSN (insert_before
);
544 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
546 /* Set insert_before to immediately after the JUMP_INSN, so that
547 NOTEs at the end of the loop will be correctly handled by
549 insert_before
= NEXT_INSN (last_loop_insn
);
550 copy_end
= last_loop_insn
;
554 /* We currently can't unroll a loop if it doesn't end with a
555 JUMP_INSN. There would need to be a mechanism that recognizes
556 this case, and then inserts a jump after each loop body, which
557 jumps to after the last loop body. */
558 if (loop_dump_stream
)
559 fprintf (loop_dump_stream
,
560 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
564 /* If copying exit test branches because they can not be eliminated,
565 then must convert the fall through case of the branch to a jump past
566 the end of the loop. Create a label to emit after the loop and save
567 it for later use. Do not use the label after the loop, if any, since
568 it might be used by insns outside the loop, or there might be insns
569 added before it later by final_[bg]iv_value which must be after
570 the real exit label. */
571 exit_label
= gen_label_rtx ();
574 while (GET_CODE (insn
) != CODE_LABEL
&& GET_CODE (insn
) != JUMP_INSN
)
575 insn
= NEXT_INSN (insn
);
577 if (GET_CODE (insn
) == JUMP_INSN
)
579 /* The loop starts with a jump down to the exit condition test.
580 Start copying the loop after the barrier following this
582 copy_start
= NEXT_INSN (insn
);
584 /* Splitting induction variables doesn't work when the loop is
585 entered via a jump to the bottom, because then we end up doing
586 a comparison against a new register for a split variable, but
587 we did not execute the set insn for the new register because
588 it was skipped over. */
589 splitting_not_safe
= 1;
590 if (loop_dump_stream
)
591 fprintf (loop_dump_stream
,
592 "Splitting not safe, because loop not entered at top.\n");
595 copy_start
= loop_start
;
598 /* This should always be the first label in the loop. */
599 start_label
= NEXT_INSN (copy_start
);
600 /* There may be a line number note and/or a loop continue note here. */
601 while (GET_CODE (start_label
) == NOTE
)
602 start_label
= NEXT_INSN (start_label
);
603 if (GET_CODE (start_label
) != CODE_LABEL
)
605 /* This can happen as a result of jump threading. If the first insns in
606 the loop test the same condition as the loop's backward jump, or the
607 opposite condition, then the backward jump will be modified to point
608 to elsewhere, and the loop's start label is deleted.
610 This case currently can not be handled by the loop unrolling code. */
612 if (loop_dump_stream
)
613 fprintf (loop_dump_stream
,
614 "Unrolling failure: unknown insns between BEG note and loop label.\n");
617 if (LABEL_NAME (start_label
))
619 /* The jump optimization pass must have combined the original start label
620 with a named label for a goto. We can't unroll this case because
621 jumps which go to the named label must be handled differently than
622 jumps to the loop start, and it is impossible to differentiate them
624 if (loop_dump_stream
)
625 fprintf (loop_dump_stream
,
626 "Unrolling failure: loop start label is gone\n");
630 if (unroll_type
== UNROLL_NAIVE
631 && GET_CODE (last_loop_insn
) == BARRIER
632 && start_label
!= JUMP_LABEL (PREV_INSN (last_loop_insn
)))
634 /* In this case, we must copy the jump and barrier, because they will
635 not be converted to jumps to an immediately following label. */
637 insert_before
= NEXT_INSN (last_loop_insn
);
638 copy_end
= last_loop_insn
;
641 if (unroll_type
== UNROLL_NAIVE
642 && GET_CODE (last_loop_insn
) == JUMP_INSN
643 && start_label
!= JUMP_LABEL (last_loop_insn
))
645 /* ??? The loop ends with a conditional branch that does not branch back
646 to the loop start label. In this case, we must emit an unconditional
647 branch to the loop exit after emitting the final branch.
648 copy_loop_body does not have support for this currently, so we
649 give up. It doesn't seem worthwhile to unroll anyways since
650 unrolling would increase the number of branch instructions
652 if (loop_dump_stream
)
653 fprintf (loop_dump_stream
,
654 "Unrolling failure: final conditional branch not to loop start\n");
658 /* Allocate a translation table for the labels and insn numbers.
659 They will be filled in as we copy the insns in the loop. */
661 max_labelno
= max_label_num ();
662 max_insnno
= get_max_uid ();
664 map
= (struct inline_remap
*) alloca (sizeof (struct inline_remap
));
666 map
->integrating
= 0;
668 /* Allocate the label map. */
672 map
->label_map
= (rtx
*) alloca (max_labelno
* sizeof (rtx
));
674 local_label
= (char *) alloca (max_labelno
);
675 bzero (local_label
, max_labelno
);
680 /* Search the loop and mark all local labels, i.e. the ones which have to
681 be distinct labels when copied. For all labels which might be
682 non-local, set their label_map entries to point to themselves.
683 If they happen to be local their label_map entries will be overwritten
684 before the loop body is copied. The label_map entries for local labels
685 will be set to a different value each time the loop body is copied. */
687 for (insn
= copy_start
; insn
!= loop_end
; insn
= NEXT_INSN (insn
))
689 if (GET_CODE (insn
) == CODE_LABEL
)
690 local_label
[CODE_LABEL_NUMBER (insn
)] = 1;
691 else if (GET_CODE (insn
) == JUMP_INSN
)
693 if (JUMP_LABEL (insn
))
694 set_label_in_map (map
,
695 CODE_LABEL_NUMBER (JUMP_LABEL (insn
)),
697 else if (GET_CODE (PATTERN (insn
)) == ADDR_VEC
698 || GET_CODE (PATTERN (insn
)) == ADDR_DIFF_VEC
)
700 rtx pat
= PATTERN (insn
);
701 int diff_vec_p
= GET_CODE (PATTERN (insn
)) == ADDR_DIFF_VEC
;
702 int len
= XVECLEN (pat
, diff_vec_p
);
705 for (i
= 0; i
< len
; i
++)
707 label
= XEXP (XVECEXP (pat
, diff_vec_p
, i
), 0);
708 set_label_in_map (map
,
709 CODE_LABEL_NUMBER (label
),
716 /* Allocate space for the insn map. */
718 map
->insn_map
= (rtx
*) alloca (max_insnno
* sizeof (rtx
));
720 /* Set this to zero, to indicate that we are doing loop unrolling,
721 not function inlining. */
722 map
->inline_target
= 0;
724 /* The register and constant maps depend on the number of registers
725 present, so the final maps can't be created until after
726 find_splittable_regs is called. However, they are needed for
727 preconditioning, so we create temporary maps when preconditioning
730 /* The preconditioning code may allocate two new pseudo registers. */
731 maxregnum
= max_reg_num ();
733 /* Allocate and zero out the splittable_regs and addr_combined_regs
734 arrays. These must be zeroed here because they will be used if
735 loop preconditioning is performed, and must be zero for that case.
737 It is safe to do this here, since the extra registers created by the
738 preconditioning code and find_splittable_regs will never be used
739 to access the splittable_regs[] and addr_combined_regs[] arrays. */
741 splittable_regs
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
742 bzero ((char *) splittable_regs
, maxregnum
* sizeof (rtx
));
743 splittable_regs_updates
= (int *) alloca (maxregnum
* sizeof (int));
744 bzero ((char *) splittable_regs_updates
, maxregnum
* sizeof (int));
746 = (struct induction
**) alloca (maxregnum
* sizeof (struct induction
*));
747 bzero ((char *) addr_combined_regs
, maxregnum
* sizeof (struct induction
*));
748 /* We must limit it to max_reg_before_loop, because only these pseudo
749 registers have valid regno_first_uid info. Any register created after
750 that is unlikely to be local to the loop anyways. */
751 local_regno
= (char *) alloca (max_reg_before_loop
);
752 bzero (local_regno
, max_reg_before_loop
);
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
)
765 /* If copy_start points to the NOTE that starts the loop, then we must
766 use the next luid, because invariant pseudo-regs moved out of the loop
767 have their lifetimes modified to start here, but they are not safe
769 if (copy_start
== loop_start
)
772 /* If a pseudo's lifetime is entirely contained within this loop, then we
773 can use a different pseudo in each unrolled copy of the loop. This
774 results in better code. */
775 for (j
= FIRST_PSEUDO_REGISTER
; j
< max_reg_before_loop
; ++j
)
776 if (REGNO_FIRST_UID (j
) > 0 && REGNO_FIRST_UID (j
) <= max_uid_for_loop
777 && uid_luid
[REGNO_FIRST_UID (j
)] >= copy_start_luid
778 && REGNO_LAST_UID (j
) > 0 && REGNO_LAST_UID (j
) <= max_uid_for_loop
779 && uid_luid
[REGNO_LAST_UID (j
)] <= copy_end_luid
)
781 /* However, we must also check for loop-carried dependencies.
782 If the value the pseudo has at the end of iteration X is
783 used by iteration X+1, then we can not use a different pseudo
784 for each unrolled copy of the loop. */
785 /* A pseudo is safe if regno_first_uid is a set, and this
786 set dominates all instructions from regno_first_uid to
788 /* ??? This check is simplistic. We would get better code if
789 this check was more sophisticated. */
790 if (set_dominates_use (j
, REGNO_FIRST_UID (j
), REGNO_LAST_UID (j
),
791 copy_start
, copy_end
))
794 if (loop_dump_stream
)
797 fprintf (loop_dump_stream
, "Marked reg %d as local\n", j
);
799 fprintf (loop_dump_stream
, "Did not mark reg %d as local\n",
805 /* If this loop requires exit tests when unrolled, check to see if we
806 can precondition the loop so as to make the exit tests unnecessary.
807 Just like variable splitting, this is not safe if the loop is entered
808 via a jump to the bottom. Also, can not do this if no strength
809 reduce info, because precondition_loop_p uses this info. */
811 /* Must copy the loop body for preconditioning before the following
812 find_splittable_regs call since that will emit insns which need to
813 be after the preconditioned loop copies, but immediately before the
814 unrolled loop copies. */
816 /* Also, it is not safe to split induction variables for the preconditioned
817 copies of the loop body. If we split induction variables, then the code
818 assumes that each induction variable can be represented as a function
819 of its initial value and the loop iteration number. This is not true
820 in this case, because the last preconditioned copy of the loop body
821 could be any iteration from the first up to the `unroll_number-1'th,
822 depending on the initial value of the iteration variable. Therefore
823 we can not split induction variables here, because we can not calculate
824 their value. Hence, this code must occur before find_splittable_regs
827 if (unroll_type
== UNROLL_NAIVE
&& ! splitting_not_safe
&& strength_reduce_p
)
829 rtx initial_value
, final_value
, increment
;
831 if (precondition_loop_p (&initial_value
, &final_value
, &increment
,
832 loop_start
, loop_end
))
835 enum machine_mode mode
;
837 int abs_inc
, neg_inc
;
839 map
->reg_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
841 map
->const_equiv_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
842 map
->const_age_map
= (unsigned *) alloca (maxregnum
843 * sizeof (unsigned));
844 map
->const_equiv_map_size
= maxregnum
;
845 global_const_equiv_map
= map
->const_equiv_map
;
846 global_const_equiv_map_size
= maxregnum
;
848 init_reg_map (map
, maxregnum
);
850 /* Limit loop unrolling to 4, since this will make 7 copies of
852 if (unroll_number
> 4)
855 /* Save the absolute value of the increment, and also whether or
856 not it is negative. */
858 abs_inc
= INTVAL (increment
);
867 /* Decide what mode to do these calculations in. Choose the larger
868 of final_value's mode and initial_value's mode, or a full-word if
869 both are constants. */
870 mode
= GET_MODE (final_value
);
871 if (mode
== VOIDmode
)
873 mode
= GET_MODE (initial_value
);
874 if (mode
== VOIDmode
)
877 else if (mode
!= GET_MODE (initial_value
)
878 && (GET_MODE_SIZE (mode
)
879 < GET_MODE_SIZE (GET_MODE (initial_value
))))
880 mode
= GET_MODE (initial_value
);
882 /* Calculate the difference between the final and initial values.
883 Final value may be a (plus (reg x) (const_int 1)) rtx.
884 Let the following cse pass simplify this if initial value is
887 We must copy the final and initial values here to avoid
888 improperly shared rtl. */
890 diff
= expand_binop (mode
, sub_optab
, copy_rtx (final_value
),
891 copy_rtx (initial_value
), NULL_RTX
, 0,
894 /* Now calculate (diff % (unroll * abs (increment))) by using an
896 diff
= expand_binop (GET_MODE (diff
), and_optab
, diff
,
897 GEN_INT (unroll_number
* abs_inc
- 1),
898 NULL_RTX
, 0, OPTAB_LIB_WIDEN
);
900 /* Now emit a sequence of branches to jump to the proper precond
903 labels
= (rtx
*) alloca (sizeof (rtx
) * unroll_number
);
904 for (i
= 0; i
< unroll_number
; i
++)
905 labels
[i
] = gen_label_rtx ();
907 /* Check for the case where the initial value is greater than or
908 equal to the final value. In that case, we want to execute
909 exactly one loop iteration. The code below will fail for this
910 case. This check does not apply if the loop has a NE
911 comparison at the end. */
913 if (loop_comparison_code
!= NE
)
915 emit_cmp_insn (initial_value
, final_value
, neg_inc
? LE
: GE
,
916 NULL_RTX
, mode
, 0, 0);
918 emit_jump_insn (gen_ble (labels
[1]));
920 emit_jump_insn (gen_bge (labels
[1]));
921 JUMP_LABEL (get_last_insn ()) = labels
[1];
922 LABEL_NUSES (labels
[1])++;
925 /* Assuming the unroll_number is 4, and the increment is 2, then
926 for a negative increment: for a positive increment:
927 diff = 0,1 precond 0 diff = 0,7 precond 0
928 diff = 2,3 precond 3 diff = 1,2 precond 1
929 diff = 4,5 precond 2 diff = 3,4 precond 2
930 diff = 6,7 precond 1 diff = 5,6 precond 3 */
932 /* We only need to emit (unroll_number - 1) branches here, the
933 last case just falls through to the following code. */
935 /* ??? This would give better code if we emitted a tree of branches
936 instead of the current linear list of branches. */
938 for (i
= 0; i
< unroll_number
- 1; i
++)
941 enum rtx_code cmp_code
;
943 /* For negative increments, must invert the constant compared
944 against, except when comparing against zero. */
952 cmp_const
= unroll_number
- i
;
961 emit_cmp_insn (diff
, GEN_INT (abs_inc
* cmp_const
),
962 cmp_code
, NULL_RTX
, mode
, 0, 0);
965 emit_jump_insn (gen_beq (labels
[i
]));
967 emit_jump_insn (gen_bge (labels
[i
]));
969 emit_jump_insn (gen_ble (labels
[i
]));
970 JUMP_LABEL (get_last_insn ()) = labels
[i
];
971 LABEL_NUSES (labels
[i
])++;
974 /* If the increment is greater than one, then we need another branch,
975 to handle other cases equivalent to 0. */
977 /* ??? This should be merged into the code above somehow to help
978 simplify the code here, and reduce the number of branches emitted.
979 For the negative increment case, the branch here could easily
980 be merged with the `0' case branch above. For the positive
981 increment case, it is not clear how this can be simplified. */
986 enum rtx_code cmp_code
;
990 cmp_const
= abs_inc
- 1;
995 cmp_const
= abs_inc
* (unroll_number
- 1) + 1;
999 emit_cmp_insn (diff
, GEN_INT (cmp_const
), cmp_code
, NULL_RTX
,
1003 emit_jump_insn (gen_ble (labels
[0]));
1005 emit_jump_insn (gen_bge (labels
[0]));
1006 JUMP_LABEL (get_last_insn ()) = labels
[0];
1007 LABEL_NUSES (labels
[0])++;
1010 sequence
= gen_sequence ();
1012 emit_insn_before (sequence
, loop_start
);
1014 /* Only the last copy of the loop body here needs the exit
1015 test, so set copy_end to exclude the compare/branch here,
1016 and then reset it inside the loop when get to the last
1019 if (GET_CODE (last_loop_insn
) == BARRIER
)
1020 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
1021 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
1024 /* The immediately preceding insn is a compare which we do not
1026 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
1028 /* The immediately preceding insn may not be a compare, so we
1030 copy_end
= PREV_INSN (last_loop_insn
);
1036 for (i
= 1; i
< unroll_number
; i
++)
1038 emit_label_after (labels
[unroll_number
- i
],
1039 PREV_INSN (loop_start
));
1041 bzero ((char *) map
->insn_map
, max_insnno
* sizeof (rtx
));
1042 bzero ((char *) map
->const_equiv_map
, maxregnum
* sizeof (rtx
));
1043 bzero ((char *) map
->const_age_map
,
1044 maxregnum
* sizeof (unsigned));
1047 for (j
= 0; j
< max_labelno
; j
++)
1049 set_label_in_map (map
, j
, gen_label_rtx ());
1051 for (j
= FIRST_PSEUDO_REGISTER
; j
< max_reg_before_loop
; j
++)
1054 map
->reg_map
[j
] = gen_reg_rtx (GET_MODE (regno_reg_rtx
[j
]));
1055 record_base_value (REGNO (map
->reg_map
[j
]),
1058 /* The last copy needs the compare/branch insns at the end,
1059 so reset copy_end here if the loop ends with a conditional
1062 if (i
== unroll_number
- 1)
1064 if (GET_CODE (last_loop_insn
) == BARRIER
)
1065 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
1067 copy_end
= last_loop_insn
;
1070 /* None of the copies are the `last_iteration', so just
1071 pass zero for that parameter. */
1072 copy_loop_body (copy_start
, copy_end
, map
, exit_label
, 0,
1073 unroll_type
, start_label
, loop_end
,
1074 loop_start
, copy_end
);
1076 emit_label_after (labels
[0], PREV_INSN (loop_start
));
1078 if (GET_CODE (last_loop_insn
) == BARRIER
)
1080 insert_before
= PREV_INSN (last_loop_insn
);
1081 copy_end
= PREV_INSN (insert_before
);
1086 /* The immediately preceding insn is a compare which we do not
1088 insert_before
= PREV_INSN (last_loop_insn
);
1089 copy_end
= PREV_INSN (insert_before
);
1091 /* The immediately preceding insn may not be a compare, so we
1093 insert_before
= last_loop_insn
;
1094 copy_end
= PREV_INSN (last_loop_insn
);
1098 /* Set unroll type to MODULO now. */
1099 unroll_type
= UNROLL_MODULO
;
1100 loop_preconditioned
= 1;
1103 /* Fix the initial value for the loop as needed. */
1104 if (loop_n_iterations
<= 0)
1105 loop_start_value
[uid_loop_num
[INSN_UID (loop_start
)]]
1111 /* If reach here, and the loop type is UNROLL_NAIVE, then don't unroll
1112 the loop unless all loops are being unrolled. */
1113 if (unroll_type
== UNROLL_NAIVE
&& ! flag_unroll_all_loops
)
1115 if (loop_dump_stream
)
1116 fprintf (loop_dump_stream
, "Unrolling failure: Naive unrolling not being done.\n");
1120 /* At this point, we are guaranteed to unroll the loop. */
1122 /* Keep track of the unroll factor for each loop. */
1123 if (unroll_type
== UNROLL_COMPLETELY
)
1124 loop_unroll_factor
[uid_loop_num
[INSN_UID (loop_start
)]] = -1;
1126 loop_unroll_factor
[uid_loop_num
[INSN_UID (loop_start
)]] = unroll_number
;
1129 /* For each biv and giv, determine whether it can be safely split into
1130 a different variable for each unrolled copy of the loop body.
1131 We precalculate and save this info here, since computing it is
1134 Do this before deleting any instructions from the loop, so that
1135 back_branch_in_range_p will work correctly. */
1137 if (splitting_not_safe
)
1140 temp
= find_splittable_regs (unroll_type
, loop_start
, loop_end
,
1141 end_insert_before
, unroll_number
);
1143 /* find_splittable_regs may have created some new registers, so must
1144 reallocate the reg_map with the new larger size, and must realloc
1145 the constant maps also. */
1147 maxregnum
= max_reg_num ();
1148 map
->reg_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
1150 init_reg_map (map
, maxregnum
);
1152 /* Space is needed in some of the map for new registers, so new_maxregnum
1153 is an (over)estimate of how many registers will exist at the end. */
1154 new_maxregnum
= maxregnum
+ (temp
* unroll_number
* 2);
1156 /* Must realloc space for the constant maps, because the number of registers
1157 may have changed. */
1159 map
->const_equiv_map
= (rtx
*) alloca (new_maxregnum
* sizeof (rtx
));
1160 map
->const_age_map
= (unsigned *) alloca (new_maxregnum
* sizeof (unsigned));
1162 map
->const_equiv_map_size
= new_maxregnum
;
1163 global_const_equiv_map
= map
->const_equiv_map
;
1164 global_const_equiv_map_size
= new_maxregnum
;
1166 /* Search the list of bivs and givs to find ones which need to be remapped
1167 when split, and set their reg_map entry appropriately. */
1169 for (bl
= loop_iv_list
; bl
; bl
= bl
->next
)
1171 if (REGNO (bl
->biv
->src_reg
) != bl
->regno
)
1172 map
->reg_map
[bl
->regno
] = bl
->biv
->src_reg
;
1174 /* Currently, non-reduced/final-value givs are never split. */
1175 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
1176 if (REGNO (v
->src_reg
) != bl
->regno
)
1177 map
->reg_map
[REGNO (v
->dest_reg
)] = v
->src_reg
;
1181 /* Use our current register alignment and pointer flags. */
1182 map
->regno_pointer_flag
= regno_pointer_flag
;
1183 map
->regno_pointer_align
= regno_pointer_align
;
1185 /* If the loop is being partially unrolled, and the iteration variables
1186 are being split, and are being renamed for the split, then must fix up
1187 the compare/jump instruction at the end of the loop to refer to the new
1188 registers. This compare isn't copied, so the registers used in it
1189 will never be replaced if it isn't done here. */
1191 if (unroll_type
== UNROLL_MODULO
)
1193 insn
= NEXT_INSN (copy_end
);
1194 if (GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == JUMP_INSN
)
1195 PATTERN (insn
) = remap_split_bivs (PATTERN (insn
));
1198 /* For unroll_number - 1 times, make a copy of each instruction
1199 between copy_start and copy_end, and insert these new instructions
1200 before the end of the loop. */
1202 for (i
= 0; i
< unroll_number
; i
++)
1204 bzero ((char *) map
->insn_map
, max_insnno
* sizeof (rtx
));
1205 bzero ((char *) map
->const_equiv_map
, new_maxregnum
* sizeof (rtx
));
1206 bzero ((char *) map
->const_age_map
, new_maxregnum
* sizeof (unsigned));
1209 for (j
= 0; j
< max_labelno
; j
++)
1211 set_label_in_map (map
, j
, gen_label_rtx ());
1213 for (j
= FIRST_PSEUDO_REGISTER
; j
< max_reg_before_loop
; j
++)
1216 map
->reg_map
[j
] = gen_reg_rtx (GET_MODE (regno_reg_rtx
[j
]));
1217 record_base_value (REGNO (map
->reg_map
[j
]),
1221 /* If loop starts with a branch to the test, then fix it so that
1222 it points to the test of the first unrolled copy of the loop. */
1223 if (i
== 0 && loop_start
!= copy_start
)
1225 insn
= PREV_INSN (copy_start
);
1226 pattern
= PATTERN (insn
);
1228 tem
= get_label_from_map (map
,
1230 (XEXP (SET_SRC (pattern
), 0)));
1231 SET_SRC (pattern
) = gen_rtx_LABEL_REF (VOIDmode
, tem
);
1233 /* Set the jump label so that it can be used by later loop unrolling
1235 JUMP_LABEL (insn
) = tem
;
1236 LABEL_NUSES (tem
)++;
1239 copy_loop_body (copy_start
, copy_end
, map
, exit_label
,
1240 i
== unroll_number
- 1, unroll_type
, start_label
,
1241 loop_end
, insert_before
, insert_before
);
1244 /* Before deleting any insns, emit a CODE_LABEL immediately after the last
1245 insn to be deleted. This prevents any runaway delete_insn call from
1246 more insns that it should, as it always stops at a CODE_LABEL. */
1248 /* Delete the compare and branch at the end of the loop if completely
1249 unrolling the loop. Deleting the backward branch at the end also
1250 deletes the code label at the start of the loop. This is done at
1251 the very end to avoid problems with back_branch_in_range_p. */
1253 if (unroll_type
== UNROLL_COMPLETELY
)
1254 safety_label
= emit_label_after (gen_label_rtx (), last_loop_insn
);
1256 safety_label
= emit_label_after (gen_label_rtx (), copy_end
);
1258 /* Delete all of the original loop instructions. Don't delete the
1259 LOOP_BEG note, or the first code label in the loop. */
1261 insn
= NEXT_INSN (copy_start
);
1262 while (insn
!= safety_label
)
1264 if (insn
!= start_label
)
1265 insn
= delete_insn (insn
);
1267 insn
= NEXT_INSN (insn
);
1270 /* Can now delete the 'safety' label emitted to protect us from runaway
1271 delete_insn calls. */
1272 if (INSN_DELETED_P (safety_label
))
1274 delete_insn (safety_label
);
1276 /* If exit_label exists, emit it after the loop. Doing the emit here
1277 forces it to have a higher INSN_UID than any insn in the unrolled loop.
1278 This is needed so that mostly_true_jump in reorg.c will treat jumps
1279 to this loop end label correctly, i.e. predict that they are usually
1282 emit_label_after (exit_label
, loop_end
);
1285 /* Return true if the loop can be safely, and profitably, preconditioned
1286 so that the unrolled copies of the loop body don't need exit tests.
1288 This only works if final_value, initial_value and increment can be
1289 determined, and if increment is a constant power of 2.
1290 If increment is not a power of 2, then the preconditioning modulo
1291 operation would require a real modulo instead of a boolean AND, and this
1292 is not considered `profitable'. */
1294 /* ??? If the loop is known to be executed very many times, or the machine
1295 has a very cheap divide instruction, then preconditioning is a win even
1296 when the increment is not a power of 2. Use RTX_COST to compute
1297 whether divide is cheap. */
1300 precondition_loop_p (initial_value
, final_value
, increment
, loop_start
,
1302 rtx
*initial_value
, *final_value
, *increment
;
1303 rtx loop_start
, loop_end
;
1306 if (loop_n_iterations
> 0)
1308 *initial_value
= const0_rtx
;
1309 *increment
= const1_rtx
;
1310 *final_value
= GEN_INT (loop_n_iterations
);
1312 if (loop_dump_stream
)
1313 fprintf (loop_dump_stream
,
1314 "Preconditioning: Success, number of iterations known, %d.\n",
1319 if (loop_initial_value
== 0)
1321 if (loop_dump_stream
)
1322 fprintf (loop_dump_stream
,
1323 "Preconditioning: Could not find initial value.\n");
1326 else if (loop_increment
== 0)
1328 if (loop_dump_stream
)
1329 fprintf (loop_dump_stream
,
1330 "Preconditioning: Could not find increment value.\n");
1333 else if (GET_CODE (loop_increment
) != CONST_INT
)
1335 if (loop_dump_stream
)
1336 fprintf (loop_dump_stream
,
1337 "Preconditioning: Increment not a constant.\n");
1340 else if ((exact_log2 (INTVAL (loop_increment
)) < 0)
1341 && (exact_log2 (- INTVAL (loop_increment
)) < 0))
1343 if (loop_dump_stream
)
1344 fprintf (loop_dump_stream
,
1345 "Preconditioning: Increment not a constant power of 2.\n");
1349 /* Unsigned_compare and compare_dir can be ignored here, since they do
1350 not matter for preconditioning. */
1352 if (loop_final_value
== 0)
1354 if (loop_dump_stream
)
1355 fprintf (loop_dump_stream
,
1356 "Preconditioning: EQ comparison loop.\n");
1360 /* Must ensure that final_value is invariant, so call invariant_p to
1361 check. Before doing so, must check regno against max_reg_before_loop
1362 to make sure that the register is in the range covered by invariant_p.
1363 If it isn't, then it is most likely a biv/giv which by definition are
1365 if ((GET_CODE (loop_final_value
) == REG
1366 && REGNO (loop_final_value
) >= max_reg_before_loop
)
1367 || (GET_CODE (loop_final_value
) == PLUS
1368 && REGNO (XEXP (loop_final_value
, 0)) >= max_reg_before_loop
)
1369 || ! invariant_p (loop_final_value
))
1371 if (loop_dump_stream
)
1372 fprintf (loop_dump_stream
,
1373 "Preconditioning: Final value not invariant.\n");
1377 /* Fail for floating point values, since the caller of this function
1378 does not have code to deal with them. */
1379 if (GET_MODE_CLASS (GET_MODE (loop_final_value
)) == MODE_FLOAT
1380 || GET_MODE_CLASS (GET_MODE (loop_initial_value
)) == MODE_FLOAT
)
1382 if (loop_dump_stream
)
1383 fprintf (loop_dump_stream
,
1384 "Preconditioning: Floating point final or initial value.\n");
1388 /* Now set initial_value to be the iteration_var, since that may be a
1389 simpler expression, and is guaranteed to be correct if all of the
1390 above tests succeed.
1392 We can not use the initial_value as calculated, because it will be
1393 one too small for loops of the form "while (i-- > 0)". We can not
1394 emit code before the loop_skip_over insns to fix this problem as this
1395 will then give a number one too large for loops of the form
1398 Note that all loops that reach here are entered at the top, because
1399 this function is not called if the loop starts with a jump. */
1401 /* Fail if loop_iteration_var is not live before loop_start, since we need
1402 to test its value in the preconditioning code. */
1404 if (uid_luid
[REGNO_FIRST_UID (REGNO (loop_iteration_var
))]
1405 > INSN_LUID (loop_start
))
1407 if (loop_dump_stream
)
1408 fprintf (loop_dump_stream
,
1409 "Preconditioning: Iteration var not live before loop start.\n");
1413 *initial_value
= loop_iteration_var
;
1414 *increment
= loop_increment
;
1415 *final_value
= loop_final_value
;
1418 if (loop_dump_stream
)
1419 fprintf (loop_dump_stream
, "Preconditioning: Successful.\n");
1424 /* All pseudo-registers must be mapped to themselves. Two hard registers
1425 must be mapped, VIRTUAL_STACK_VARS_REGNUM and VIRTUAL_INCOMING_ARGS_
1426 REGNUM, to avoid function-inlining specific conversions of these
1427 registers. All other hard regs can not be mapped because they may be
1432 init_reg_map (map
, maxregnum
)
1433 struct inline_remap
*map
;
1438 for (i
= maxregnum
- 1; i
> LAST_VIRTUAL_REGISTER
; i
--)
1439 map
->reg_map
[i
] = regno_reg_rtx
[i
];
1440 /* Just clear the rest of the entries. */
1441 for (i
= LAST_VIRTUAL_REGISTER
; i
>= 0; i
--)
1442 map
->reg_map
[i
] = 0;
1444 map
->reg_map
[VIRTUAL_STACK_VARS_REGNUM
]
1445 = regno_reg_rtx
[VIRTUAL_STACK_VARS_REGNUM
];
1446 map
->reg_map
[VIRTUAL_INCOMING_ARGS_REGNUM
]
1447 = regno_reg_rtx
[VIRTUAL_INCOMING_ARGS_REGNUM
];
1450 /* Strength-reduction will often emit code for optimized biv/givs which
1451 calculates their value in a temporary register, and then copies the result
1452 to the iv. This procedure reconstructs the pattern computing the iv;
1453 verifying that all operands are of the proper form.
1455 The return value is the amount that the giv is incremented by. */
1458 calculate_giv_inc (pattern
, src_insn
, regno
)
1459 rtx pattern
, src_insn
;
1463 rtx increment_total
= 0;
1467 /* Verify that we have an increment insn here. First check for a plus
1468 as the set source. */
1469 if (GET_CODE (SET_SRC (pattern
)) != PLUS
)
1471 /* SR sometimes computes the new giv value in a temp, then copies it
1473 src_insn
= PREV_INSN (src_insn
);
1474 pattern
= PATTERN (src_insn
);
1475 if (GET_CODE (SET_SRC (pattern
)) != PLUS
)
1478 /* The last insn emitted is not needed, so delete it to avoid confusing
1479 the second cse pass. This insn sets the giv unnecessarily. */
1480 delete_insn (get_last_insn ());
1483 /* Verify that we have a constant as the second operand of the plus. */
1484 increment
= XEXP (SET_SRC (pattern
), 1);
1485 if (GET_CODE (increment
) != CONST_INT
)
1487 /* SR sometimes puts the constant in a register, especially if it is
1488 too big to be an add immed operand. */
1489 src_insn
= PREV_INSN (src_insn
);
1490 increment
= SET_SRC (PATTERN (src_insn
));
1492 /* SR may have used LO_SUM to compute the constant if it is too large
1493 for a load immed operand. In this case, the constant is in operand
1494 one of the LO_SUM rtx. */
1495 if (GET_CODE (increment
) == LO_SUM
)
1496 increment
= XEXP (increment
, 1);
1498 /* Some ports store large constants in memory and add a REG_EQUAL
1499 note to the store insn. */
1500 else if (GET_CODE (increment
) == MEM
)
1502 rtx note
= find_reg_note (src_insn
, REG_EQUAL
, 0);
1504 increment
= XEXP (note
, 0);
1507 else if (GET_CODE (increment
) == IOR
1508 || GET_CODE (increment
) == ASHIFT
1509 || GET_CODE (increment
) == PLUS
)
1511 /* The rs6000 port loads some constants with IOR.
1512 The alpha port loads some constants with ASHIFT and PLUS. */
1513 rtx second_part
= XEXP (increment
, 1);
1514 enum rtx_code code
= GET_CODE (increment
);
1516 src_insn
= PREV_INSN (src_insn
);
1517 increment
= SET_SRC (PATTERN (src_insn
));
1518 /* Don't need the last insn anymore. */
1519 delete_insn (get_last_insn ());
1521 if (GET_CODE (second_part
) != CONST_INT
1522 || GET_CODE (increment
) != CONST_INT
)
1526 increment
= GEN_INT (INTVAL (increment
) | INTVAL (second_part
));
1527 else if (code
== PLUS
)
1528 increment
= GEN_INT (INTVAL (increment
) + INTVAL (second_part
));
1530 increment
= GEN_INT (INTVAL (increment
) << INTVAL (second_part
));
1533 if (GET_CODE (increment
) != CONST_INT
)
1536 /* The insn loading the constant into a register is no longer needed,
1538 delete_insn (get_last_insn ());
1541 if (increment_total
)
1542 increment_total
= GEN_INT (INTVAL (increment_total
) + INTVAL (increment
));
1544 increment_total
= increment
;
1546 /* Check that the source register is the same as the register we expected
1547 to see as the source. If not, something is seriously wrong. */
1548 if (GET_CODE (XEXP (SET_SRC (pattern
), 0)) != REG
1549 || REGNO (XEXP (SET_SRC (pattern
), 0)) != regno
)
1551 /* Some machines (e.g. the romp), may emit two add instructions for
1552 certain constants, so lets try looking for another add immediately
1553 before this one if we have only seen one add insn so far. */
1559 src_insn
= PREV_INSN (src_insn
);
1560 pattern
= PATTERN (src_insn
);
1562 delete_insn (get_last_insn ());
1570 return increment_total
;
1573 /* Copy REG_NOTES, except for insn references, because not all insn_map
1574 entries are valid yet. We do need to copy registers now though, because
1575 the reg_map entries can change during copying. */
1578 initial_reg_note_copy (notes
, map
)
1580 struct inline_remap
*map
;
1587 copy
= rtx_alloc (GET_CODE (notes
));
1588 PUT_MODE (copy
, GET_MODE (notes
));
1590 if (GET_CODE (notes
) == EXPR_LIST
)
1591 XEXP (copy
, 0) = copy_rtx_and_substitute (XEXP (notes
, 0), map
);
1592 else if (GET_CODE (notes
) == INSN_LIST
)
1593 /* Don't substitute for these yet. */
1594 XEXP (copy
, 0) = XEXP (notes
, 0);
1598 XEXP (copy
, 1) = initial_reg_note_copy (XEXP (notes
, 1), map
);
1603 /* Fixup insn references in copied REG_NOTES. */
1606 final_reg_note_copy (notes
, map
)
1608 struct inline_remap
*map
;
1612 for (note
= notes
; note
; note
= XEXP (note
, 1))
1613 if (GET_CODE (note
) == INSN_LIST
)
1614 XEXP (note
, 0) = map
->insn_map
[INSN_UID (XEXP (note
, 0))];
1617 /* Copy each instruction in the loop, substituting from map as appropriate.
1618 This is very similar to a loop in expand_inline_function. */
1621 copy_loop_body (copy_start
, copy_end
, map
, exit_label
, last_iteration
,
1622 unroll_type
, start_label
, loop_end
, insert_before
,
1624 rtx copy_start
, copy_end
;
1625 struct inline_remap
*map
;
1628 enum unroll_types unroll_type
;
1629 rtx start_label
, loop_end
, insert_before
, copy_notes_from
;
1633 int dest_reg_was_split
, i
;
1637 rtx final_label
= 0;
1638 rtx giv_inc
, giv_dest_reg
, giv_src_reg
;
1640 /* If this isn't the last iteration, then map any references to the
1641 start_label to final_label. Final label will then be emitted immediately
1642 after the end of this loop body if it was ever used.
1644 If this is the last iteration, then map references to the start_label
1646 if (! last_iteration
)
1648 final_label
= gen_label_rtx ();
1649 set_label_in_map (map
, CODE_LABEL_NUMBER (start_label
),
1653 set_label_in_map (map
, CODE_LABEL_NUMBER (start_label
), start_label
);
1660 insn
= NEXT_INSN (insn
);
1662 map
->orig_asm_operands_vector
= 0;
1664 switch (GET_CODE (insn
))
1667 pattern
= PATTERN (insn
);
1671 /* Check to see if this is a giv that has been combined with
1672 some split address givs. (Combined in the sense that
1673 `combine_givs' in loop.c has put two givs in the same register.)
1674 In this case, we must search all givs based on the same biv to
1675 find the address givs. Then split the address givs.
1676 Do this before splitting the giv, since that may map the
1677 SET_DEST to a new register. */
1679 if (GET_CODE (pattern
) == SET
1680 && GET_CODE (SET_DEST (pattern
)) == REG
1681 && addr_combined_regs
[REGNO (SET_DEST (pattern
))])
1683 struct iv_class
*bl
;
1684 struct induction
*v
, *tv
;
1685 int regno
= REGNO (SET_DEST (pattern
));
1687 v
= addr_combined_regs
[REGNO (SET_DEST (pattern
))];
1688 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
1690 /* Although the giv_inc amount is not needed here, we must call
1691 calculate_giv_inc here since it might try to delete the
1692 last insn emitted. If we wait until later to call it,
1693 we might accidentally delete insns generated immediately
1694 below by emit_unrolled_add. */
1696 giv_inc
= calculate_giv_inc (pattern
, insn
, regno
);
1698 /* Now find all address giv's that were combined with this
1700 for (tv
= bl
->giv
; tv
; tv
= tv
->next_iv
)
1701 if (tv
->giv_type
== DEST_ADDR
&& tv
->same
== v
)
1705 /* If this DEST_ADDR giv was not split, then ignore it. */
1706 if (*tv
->location
!= tv
->dest_reg
)
1709 /* Scale this_giv_inc if the multiplicative factors of
1710 the two givs are different. */
1711 this_giv_inc
= INTVAL (giv_inc
);
1712 if (tv
->mult_val
!= v
->mult_val
)
1713 this_giv_inc
= (this_giv_inc
/ INTVAL (v
->mult_val
)
1714 * INTVAL (tv
->mult_val
));
1716 tv
->dest_reg
= plus_constant (tv
->dest_reg
, this_giv_inc
);
1717 *tv
->location
= tv
->dest_reg
;
1719 if (last_iteration
&& unroll_type
!= UNROLL_COMPLETELY
)
1721 /* Must emit an insn to increment the split address
1722 giv. Add in the const_adjust field in case there
1723 was a constant eliminated from the address. */
1724 rtx value
, dest_reg
;
1726 /* tv->dest_reg will be either a bare register,
1727 or else a register plus a constant. */
1728 if (GET_CODE (tv
->dest_reg
) == REG
)
1729 dest_reg
= tv
->dest_reg
;
1731 dest_reg
= XEXP (tv
->dest_reg
, 0);
1733 /* Check for shared address givs, and avoid
1734 incrementing the shared pseudo reg more than
1736 if (! tv
->same_insn
&& ! tv
->shared
)
1738 /* tv->dest_reg may actually be a (PLUS (REG)
1739 (CONST)) here, so we must call plus_constant
1740 to add the const_adjust amount before calling
1741 emit_unrolled_add below. */
1742 value
= plus_constant (tv
->dest_reg
,
1745 /* The constant could be too large for an add
1746 immediate, so can't directly emit an insn
1748 emit_unrolled_add (dest_reg
, XEXP (value
, 0),
1752 /* Reset the giv to be just the register again, in case
1753 it is used after the set we have just emitted.
1754 We must subtract the const_adjust factor added in
1756 tv
->dest_reg
= plus_constant (dest_reg
,
1757 - tv
->const_adjust
);
1758 *tv
->location
= tv
->dest_reg
;
1763 /* If this is a setting of a splittable variable, then determine
1764 how to split the variable, create a new set based on this split,
1765 and set up the reg_map so that later uses of the variable will
1766 use the new split variable. */
1768 dest_reg_was_split
= 0;
1770 if (GET_CODE (pattern
) == SET
1771 && GET_CODE (SET_DEST (pattern
)) == REG
1772 && splittable_regs
[REGNO (SET_DEST (pattern
))])
1774 int regno
= REGNO (SET_DEST (pattern
));
1776 dest_reg_was_split
= 1;
1778 /* Compute the increment value for the giv, if it wasn't
1779 already computed above. */
1782 giv_inc
= calculate_giv_inc (pattern
, insn
, regno
);
1783 giv_dest_reg
= SET_DEST (pattern
);
1784 giv_src_reg
= SET_DEST (pattern
);
1786 if (unroll_type
== UNROLL_COMPLETELY
)
1788 /* Completely unrolling the loop. Set the induction
1789 variable to a known constant value. */
1791 /* The value in splittable_regs may be an invariant
1792 value, so we must use plus_constant here. */
1793 splittable_regs
[regno
]
1794 = plus_constant (splittable_regs
[regno
], INTVAL (giv_inc
));
1796 if (GET_CODE (splittable_regs
[regno
]) == PLUS
)
1798 giv_src_reg
= XEXP (splittable_regs
[regno
], 0);
1799 giv_inc
= XEXP (splittable_regs
[regno
], 1);
1803 /* The splittable_regs value must be a REG or a
1804 CONST_INT, so put the entire value in the giv_src_reg
1806 giv_src_reg
= splittable_regs
[regno
];
1807 giv_inc
= const0_rtx
;
1812 /* Partially unrolling loop. Create a new pseudo
1813 register for the iteration variable, and set it to
1814 be a constant plus the original register. Except
1815 on the last iteration, when the result has to
1816 go back into the original iteration var register. */
1818 /* Handle bivs which must be mapped to a new register
1819 when split. This happens for bivs which need their
1820 final value set before loop entry. The new register
1821 for the biv was stored in the biv's first struct
1822 induction entry by find_splittable_regs. */
1824 if (regno
< max_reg_before_loop
1825 && reg_iv_type
[regno
] == BASIC_INDUCT
)
1827 giv_src_reg
= reg_biv_class
[regno
]->biv
->src_reg
;
1828 giv_dest_reg
= giv_src_reg
;
1832 /* If non-reduced/final-value givs were split, then
1833 this would have to remap those givs also. See
1834 find_splittable_regs. */
1837 splittable_regs
[regno
]
1838 = GEN_INT (INTVAL (giv_inc
)
1839 + INTVAL (splittable_regs
[regno
]));
1840 giv_inc
= splittable_regs
[regno
];
1842 /* Now split the induction variable by changing the dest
1843 of this insn to a new register, and setting its
1844 reg_map entry to point to this new register.
1846 If this is the last iteration, and this is the last insn
1847 that will update the iv, then reuse the original dest,
1848 to ensure that the iv will have the proper value when
1849 the loop exits or repeats.
1851 Using splittable_regs_updates here like this is safe,
1852 because it can only be greater than one if all
1853 instructions modifying the iv are always executed in
1856 if (! last_iteration
1857 || (splittable_regs_updates
[regno
]-- != 1))
1859 tem
= gen_reg_rtx (GET_MODE (giv_src_reg
));
1861 map
->reg_map
[regno
] = tem
;
1862 record_base_value (REGNO (tem
), giv_src_reg
);
1865 map
->reg_map
[regno
] = giv_src_reg
;
1868 /* The constant being added could be too large for an add
1869 immediate, so can't directly emit an insn here. */
1870 emit_unrolled_add (giv_dest_reg
, giv_src_reg
, giv_inc
);
1871 copy
= get_last_insn ();
1872 pattern
= PATTERN (copy
);
1876 pattern
= copy_rtx_and_substitute (pattern
, map
);
1877 copy
= emit_insn (pattern
);
1879 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
1882 /* If this insn is setting CC0, it may need to look at
1883 the insn that uses CC0 to see what type of insn it is.
1884 In that case, the call to recog via validate_change will
1885 fail. So don't substitute constants here. Instead,
1886 do it when we emit the following insn.
1888 For example, see the pyr.md file. That machine has signed and
1889 unsigned compares. The compare patterns must check the
1890 following branch insn to see which what kind of compare to
1893 If the previous insn set CC0, substitute constants on it as
1895 if (sets_cc0_p (PATTERN (copy
)) != 0)
1900 try_constants (cc0_insn
, map
);
1902 try_constants (copy
, map
);
1905 try_constants (copy
, map
);
1908 /* Make split induction variable constants `permanent' since we
1909 know there are no backward branches across iteration variable
1910 settings which would invalidate this. */
1911 if (dest_reg_was_split
)
1913 int regno
= REGNO (SET_DEST (pattern
));
1915 if (regno
< map
->const_equiv_map_size
1916 && map
->const_age_map
[regno
] == map
->const_age
)
1917 map
->const_age_map
[regno
] = -1;
1922 pattern
= copy_rtx_and_substitute (PATTERN (insn
), map
);
1923 copy
= emit_jump_insn (pattern
);
1924 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
1926 if (JUMP_LABEL (insn
) == start_label
&& insn
== copy_end
1927 && ! last_iteration
)
1929 /* This is a branch to the beginning of the loop; this is the
1930 last insn being copied; and this is not the last iteration.
1931 In this case, we want to change the original fall through
1932 case to be a branch past the end of the loop, and the
1933 original jump label case to fall_through. */
1935 if (invert_exp (pattern
, copy
))
1937 if (! redirect_exp (&pattern
,
1938 get_label_from_map (map
,
1940 (JUMP_LABEL (insn
))),
1947 rtx lab
= gen_label_rtx ();
1948 /* Can't do it by reversing the jump (probably because we
1949 couldn't reverse the conditions), so emit a new
1950 jump_insn after COPY, and redirect the jump around
1952 jmp
= emit_jump_insn_after (gen_jump (exit_label
), copy
);
1953 jmp
= emit_barrier_after (jmp
);
1954 emit_label_after (lab
, jmp
);
1955 LABEL_NUSES (lab
) = 0;
1956 if (! redirect_exp (&pattern
,
1957 get_label_from_map (map
,
1959 (JUMP_LABEL (insn
))),
1967 try_constants (cc0_insn
, map
);
1970 try_constants (copy
, map
);
1972 /* Set the jump label of COPY correctly to avoid problems with
1973 later passes of unroll_loop, if INSN had jump label set. */
1974 if (JUMP_LABEL (insn
))
1978 /* Can't use the label_map for every insn, since this may be
1979 the backward branch, and hence the label was not mapped. */
1980 if (GET_CODE (pattern
) == SET
)
1982 tem
= SET_SRC (pattern
);
1983 if (GET_CODE (tem
) == LABEL_REF
)
1984 label
= XEXP (tem
, 0);
1985 else if (GET_CODE (tem
) == IF_THEN_ELSE
)
1987 if (XEXP (tem
, 1) != pc_rtx
)
1988 label
= XEXP (XEXP (tem
, 1), 0);
1990 label
= XEXP (XEXP (tem
, 2), 0);
1994 if (label
&& GET_CODE (label
) == CODE_LABEL
)
1995 JUMP_LABEL (copy
) = label
;
1998 /* An unrecognizable jump insn, probably the entry jump
1999 for a switch statement. This label must have been mapped,
2000 so just use the label_map to get the new jump label. */
2002 = get_label_from_map (map
,
2003 CODE_LABEL_NUMBER (JUMP_LABEL (insn
)));
2006 /* If this is a non-local jump, then must increase the label
2007 use count so that the label will not be deleted when the
2008 original jump is deleted. */
2009 LABEL_NUSES (JUMP_LABEL (copy
))++;
2011 else if (GET_CODE (PATTERN (copy
)) == ADDR_VEC
2012 || GET_CODE (PATTERN (copy
)) == ADDR_DIFF_VEC
)
2014 rtx pat
= PATTERN (copy
);
2015 int diff_vec_p
= GET_CODE (pat
) == ADDR_DIFF_VEC
;
2016 int len
= XVECLEN (pat
, diff_vec_p
);
2019 for (i
= 0; i
< len
; i
++)
2020 LABEL_NUSES (XEXP (XVECEXP (pat
, diff_vec_p
, i
), 0))++;
2023 /* If this used to be a conditional jump insn but whose branch
2024 direction is now known, we must do something special. */
2025 if (condjump_p (insn
) && !simplejump_p (insn
) && map
->last_pc_value
)
2028 /* The previous insn set cc0 for us. So delete it. */
2029 delete_insn (PREV_INSN (copy
));
2032 /* If this is now a no-op, delete it. */
2033 if (map
->last_pc_value
== pc_rtx
)
2035 /* Don't let delete_insn delete the label referenced here,
2036 because we might possibly need it later for some other
2037 instruction in the loop. */
2038 if (JUMP_LABEL (copy
))
2039 LABEL_NUSES (JUMP_LABEL (copy
))++;
2041 if (JUMP_LABEL (copy
))
2042 LABEL_NUSES (JUMP_LABEL (copy
))--;
2046 /* Otherwise, this is unconditional jump so we must put a
2047 BARRIER after it. We could do some dead code elimination
2048 here, but jump.c will do it just as well. */
2054 pattern
= copy_rtx_and_substitute (PATTERN (insn
), map
);
2055 copy
= emit_call_insn (pattern
);
2056 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
2058 /* Because the USAGE information potentially contains objects other
2059 than hard registers, we need to copy it. */
2060 CALL_INSN_FUNCTION_USAGE (copy
)
2061 = copy_rtx_and_substitute (CALL_INSN_FUNCTION_USAGE (insn
), map
);
2065 try_constants (cc0_insn
, map
);
2068 try_constants (copy
, map
);
2070 /* Be lazy and assume CALL_INSNs clobber all hard registers. */
2071 for (i
= 0; i
< FIRST_PSEUDO_REGISTER
; i
++)
2072 map
->const_equiv_map
[i
] = 0;
2076 /* If this is the loop start label, then we don't need to emit a
2077 copy of this label since no one will use it. */
2079 if (insn
!= start_label
)
2081 copy
= emit_label (get_label_from_map (map
,
2082 CODE_LABEL_NUMBER (insn
)));
2088 copy
= emit_barrier ();
2092 /* VTOP notes are valid only before the loop exit test. If placed
2093 anywhere else, loop may generate bad code. */
2095 if (NOTE_LINE_NUMBER (insn
) != NOTE_INSN_DELETED
2096 && (NOTE_LINE_NUMBER (insn
) != NOTE_INSN_LOOP_VTOP
2097 || (last_iteration
&& unroll_type
!= UNROLL_COMPLETELY
)))
2098 copy
= emit_note (NOTE_SOURCE_FILE (insn
),
2099 NOTE_LINE_NUMBER (insn
));
2109 map
->insn_map
[INSN_UID (insn
)] = copy
;
2111 while (insn
!= copy_end
);
2113 /* Now finish coping the REG_NOTES. */
2117 insn
= NEXT_INSN (insn
);
2118 if ((GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == JUMP_INSN
2119 || GET_CODE (insn
) == CALL_INSN
)
2120 && map
->insn_map
[INSN_UID (insn
)])
2121 final_reg_note_copy (REG_NOTES (map
->insn_map
[INSN_UID (insn
)]), map
);
2123 while (insn
!= copy_end
);
2125 /* There may be notes between copy_notes_from and loop_end. Emit a copy of
2126 each of these notes here, since there may be some important ones, such as
2127 NOTE_INSN_BLOCK_END notes, in this group. We don't do this on the last
2128 iteration, because the original notes won't be deleted.
2130 We can't use insert_before here, because when from preconditioning,
2131 insert_before points before the loop. We can't use copy_end, because
2132 there may be insns already inserted after it (which we don't want to
2133 copy) when not from preconditioning code. */
2135 if (! last_iteration
)
2137 for (insn
= copy_notes_from
; insn
!= loop_end
; insn
= NEXT_INSN (insn
))
2139 if (GET_CODE (insn
) == NOTE
2140 && NOTE_LINE_NUMBER (insn
) != NOTE_INSN_DELETED
)
2141 emit_note (NOTE_SOURCE_FILE (insn
), NOTE_LINE_NUMBER (insn
));
2145 if (final_label
&& LABEL_NUSES (final_label
) > 0)
2146 emit_label (final_label
);
2148 tem
= gen_sequence ();
2150 emit_insn_before (tem
, insert_before
);
2153 /* Emit an insn, using the expand_binop to ensure that a valid insn is
2154 emitted. This will correctly handle the case where the increment value
2155 won't fit in the immediate field of a PLUS insns. */
2158 emit_unrolled_add (dest_reg
, src_reg
, increment
)
2159 rtx dest_reg
, src_reg
, increment
;
2163 result
= expand_binop (GET_MODE (dest_reg
), add_optab
, src_reg
, increment
,
2164 dest_reg
, 0, OPTAB_LIB_WIDEN
);
2166 if (dest_reg
!= result
)
2167 emit_move_insn (dest_reg
, result
);
2170 /* Searches the insns between INSN and LOOP_END. Returns 1 if there
2171 is a backward branch in that range that branches to somewhere between
2172 LOOP_START and INSN. Returns 0 otherwise. */
2174 /* ??? This is quadratic algorithm. Could be rewritten to be linear.
2175 In practice, this is not a problem, because this function is seldom called,
2176 and uses a negligible amount of CPU time on average. */
2179 back_branch_in_range_p (insn
, loop_start
, loop_end
)
2181 rtx loop_start
, loop_end
;
2183 rtx p
, q
, target_insn
;
2184 rtx orig_loop_end
= loop_end
;
2186 /* Stop before we get to the backward branch at the end of the loop. */
2187 loop_end
= prev_nonnote_insn (loop_end
);
2188 if (GET_CODE (loop_end
) == BARRIER
)
2189 loop_end
= PREV_INSN (loop_end
);
2191 /* Check in case insn has been deleted, search forward for first non
2192 deleted insn following it. */
2193 while (INSN_DELETED_P (insn
))
2194 insn
= NEXT_INSN (insn
);
2196 /* Check for the case where insn is the last insn in the loop. Deal
2197 with the case where INSN was a deleted loop test insn, in which case
2198 it will now be the NOTE_LOOP_END. */
2199 if (insn
== loop_end
|| insn
== orig_loop_end
)
2202 for (p
= NEXT_INSN (insn
); p
!= loop_end
; p
= NEXT_INSN (p
))
2204 if (GET_CODE (p
) == JUMP_INSN
)
2206 target_insn
= JUMP_LABEL (p
);
2208 /* Search from loop_start to insn, to see if one of them is
2209 the target_insn. We can't use INSN_LUID comparisons here,
2210 since insn may not have an LUID entry. */
2211 for (q
= loop_start
; q
!= insn
; q
= NEXT_INSN (q
))
2212 if (q
== target_insn
)
2220 /* Try to generate the simplest rtx for the expression
2221 (PLUS (MULT mult1 mult2) add1). This is used to calculate the initial
2225 fold_rtx_mult_add (mult1
, mult2
, add1
, mode
)
2226 rtx mult1
, mult2
, add1
;
2227 enum machine_mode mode
;
2232 /* The modes must all be the same. This should always be true. For now,
2233 check to make sure. */
2234 if ((GET_MODE (mult1
) != mode
&& GET_MODE (mult1
) != VOIDmode
)
2235 || (GET_MODE (mult2
) != mode
&& GET_MODE (mult2
) != VOIDmode
)
2236 || (GET_MODE (add1
) != mode
&& GET_MODE (add1
) != VOIDmode
))
2239 /* Ensure that if at least one of mult1/mult2 are constant, then mult2
2240 will be a constant. */
2241 if (GET_CODE (mult1
) == CONST_INT
)
2248 mult_res
= simplify_binary_operation (MULT
, mode
, mult1
, mult2
);
2250 mult_res
= gen_rtx_MULT (mode
, mult1
, mult2
);
2252 /* Again, put the constant second. */
2253 if (GET_CODE (add1
) == CONST_INT
)
2260 result
= simplify_binary_operation (PLUS
, mode
, add1
, mult_res
);
2262 result
= gen_rtx_PLUS (mode
, add1
, mult_res
);
2267 /* Searches the list of induction struct's for the biv BL, to try to calculate
2268 the total increment value for one iteration of the loop as a constant.
2270 Returns the increment value as an rtx, simplified as much as possible,
2271 if it can be calculated. Otherwise, returns 0. */
2274 biv_total_increment (bl
, loop_start
, loop_end
)
2275 struct iv_class
*bl
;
2276 rtx loop_start
, loop_end
;
2278 struct induction
*v
;
2281 /* For increment, must check every instruction that sets it. Each
2282 instruction must be executed only once each time through the loop.
2283 To verify this, we check that the the insn is always executed, and that
2284 there are no backward branches after the insn that branch to before it.
2285 Also, the insn must have a mult_val of one (to make sure it really is
2288 result
= const0_rtx
;
2289 for (v
= bl
->biv
; v
; v
= v
->next_iv
)
2291 if (v
->always_computable
&& v
->mult_val
== const1_rtx
2292 && ! back_branch_in_range_p (v
->insn
, loop_start
, loop_end
))
2293 result
= fold_rtx_mult_add (result
, const1_rtx
, v
->add_val
, v
->mode
);
2301 /* Determine the initial value of the iteration variable, and the amount
2302 that it is incremented each loop. Use the tables constructed by
2303 the strength reduction pass to calculate these values.
2305 Initial_value and/or increment are set to zero if their values could not
2309 iteration_info (iteration_var
, initial_value
, increment
, loop_start
, loop_end
)
2310 rtx iteration_var
, *initial_value
, *increment
;
2311 rtx loop_start
, loop_end
;
2313 struct iv_class
*bl
;
2315 struct induction
*v
;
2318 /* Clear the result values, in case no answer can be found. */
2322 /* The iteration variable can be either a giv or a biv. Check to see
2323 which it is, and compute the variable's initial value, and increment
2324 value if possible. */
2326 /* If this is a new register, can't handle it since we don't have any
2327 reg_iv_type entry for it. */
2328 if (REGNO (iteration_var
) >= max_reg_before_loop
)
2330 if (loop_dump_stream
)
2331 fprintf (loop_dump_stream
,
2332 "Loop unrolling: No reg_iv_type entry for iteration var.\n");
2336 /* Reject iteration variables larger than the host wide int size, since they
2337 could result in a number of iterations greater than the range of our
2338 `unsigned HOST_WIDE_INT' variable loop_n_iterations. */
2339 else if ((GET_MODE_BITSIZE (GET_MODE (iteration_var
))
2340 > HOST_BITS_PER_WIDE_INT
))
2342 if (loop_dump_stream
)
2343 fprintf (loop_dump_stream
,
2344 "Loop unrolling: Iteration var rejected because mode too large.\n");
2347 else if (GET_MODE_CLASS (GET_MODE (iteration_var
)) != MODE_INT
)
2349 if (loop_dump_stream
)
2350 fprintf (loop_dump_stream
,
2351 "Loop unrolling: Iteration var not an integer.\n");
2354 else if (reg_iv_type
[REGNO (iteration_var
)] == BASIC_INDUCT
)
2356 /* Grab initial value, only useful if it is a constant. */
2357 bl
= reg_biv_class
[REGNO (iteration_var
)];
2358 *initial_value
= bl
->initial_value
;
2360 *increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2362 else if (reg_iv_type
[REGNO (iteration_var
)] == GENERAL_INDUCT
)
2365 /* ??? The code below does not work because the incorrect number of
2366 iterations is calculated when the biv is incremented after the giv
2367 is set (which is the usual case). This can probably be accounted
2368 for by biasing the initial_value by subtracting the amount of the
2369 increment that occurs between the giv set and the giv test. However,
2370 a giv as an iterator is very rare, so it does not seem worthwhile
2372 /* ??? An example failure is: i = 6; do {;} while (i++ < 9). */
2373 if (loop_dump_stream
)
2374 fprintf (loop_dump_stream
,
2375 "Loop unrolling: Giv iterators are not handled.\n");
2378 /* Initial value is mult_val times the biv's initial value plus
2379 add_val. Only useful if it is a constant. */
2380 v
= reg_iv_info
[REGNO (iteration_var
)];
2381 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
2382 *initial_value
= fold_rtx_mult_add (v
->mult_val
, bl
->initial_value
,
2383 v
->add_val
, v
->mode
);
2385 /* Increment value is mult_val times the increment value of the biv. */
2387 *increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2389 *increment
= fold_rtx_mult_add (v
->mult_val
, *increment
, const0_rtx
,
2395 if (loop_dump_stream
)
2396 fprintf (loop_dump_stream
,
2397 "Loop unrolling: Not basic or general induction var.\n");
2402 /* Calculate the approximate final value of the iteration variable
2403 which has an loop exit test with code COMPARISON_CODE and comparison value
2404 of COMPARISON_VALUE. Also returns an indication of whether the comparison
2405 was signed or unsigned, and the direction of the comparison. This info is
2406 needed to calculate the number of loop iterations. */
2409 approx_final_value (comparison_code
, comparison_value
, unsigned_p
, compare_dir
)
2410 enum rtx_code comparison_code
;
2411 rtx comparison_value
;
2415 /* Calculate the final value of the induction variable.
2416 The exact final value depends on the branch operator, and increment sign.
2417 This is only an approximate value. It will be wrong if the iteration
2418 variable is not incremented by one each time through the loop, and
2419 approx final value - start value % increment != 0. */
2422 switch (comparison_code
)
2428 return plus_constant (comparison_value
, 1);
2433 return plus_constant (comparison_value
, -1);
2435 /* Can not calculate a final value for this case. */
2442 return comparison_value
;
2448 return comparison_value
;
2451 return comparison_value
;
2457 /* For each biv and giv, determine whether it can be safely split into
2458 a different variable for each unrolled copy of the loop body. If it
2459 is safe to split, then indicate that by saving some useful info
2460 in the splittable_regs array.
2462 If the loop is being completely unrolled, then splittable_regs will hold
2463 the current value of the induction variable while the loop is unrolled.
2464 It must be set to the initial value of the induction variable here.
2465 Otherwise, splittable_regs will hold the difference between the current
2466 value of the induction variable and the value the induction variable had
2467 at the top of the loop. It must be set to the value 0 here.
2469 Returns the total number of instructions that set registers that are
2472 /* ?? If the loop is only unrolled twice, then most of the restrictions to
2473 constant values are unnecessary, since we can easily calculate increment
2474 values in this case even if nothing is constant. The increment value
2475 should not involve a multiply however. */
2477 /* ?? Even if the biv/giv increment values aren't constant, it may still
2478 be beneficial to split the variable if the loop is only unrolled a few
2479 times, since multiplies by small integers (1,2,3,4) are very cheap. */
2482 find_splittable_regs (unroll_type
, loop_start
, loop_end
, end_insert_before
,
2484 enum unroll_types unroll_type
;
2485 rtx loop_start
, loop_end
;
2486 rtx end_insert_before
;
2489 struct iv_class
*bl
;
2490 struct induction
*v
;
2492 rtx biv_final_value
;
2496 for (bl
= loop_iv_list
; bl
; bl
= bl
->next
)
2498 /* Biv_total_increment must return a constant value,
2499 otherwise we can not calculate the split values. */
2501 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2502 if (! increment
|| GET_CODE (increment
) != CONST_INT
)
2505 /* The loop must be unrolled completely, or else have a known number
2506 of iterations and only one exit, or else the biv must be dead
2507 outside the loop, or else the final value must be known. Otherwise,
2508 it is unsafe to split the biv since it may not have the proper
2509 value on loop exit. */
2511 /* loop_number_exit_count is non-zero if the loop has an exit other than
2512 a fall through at the end. */
2515 biv_final_value
= 0;
2516 if (unroll_type
!= UNROLL_COMPLETELY
2517 && (loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]]
2518 || unroll_type
== UNROLL_NAIVE
)
2519 && (uid_luid
[REGNO_LAST_UID (bl
->regno
)] >= INSN_LUID (loop_end
)
2521 || INSN_UID (bl
->init_insn
) >= max_uid_for_loop
2522 || (uid_luid
[REGNO_FIRST_UID (bl
->regno
)]
2523 < INSN_LUID (bl
->init_insn
))
2524 || reg_mentioned_p (bl
->biv
->dest_reg
, SET_SRC (bl
->init_set
)))
2525 && ! (biv_final_value
= final_biv_value (bl
, loop_start
, loop_end
)))
2528 /* If any of the insns setting the BIV don't do so with a simple
2529 PLUS, we don't know how to split it. */
2530 for (v
= bl
->biv
; biv_splittable
&& v
; v
= v
->next_iv
)
2531 if ((tem
= single_set (v
->insn
)) == 0
2532 || GET_CODE (SET_DEST (tem
)) != REG
2533 || REGNO (SET_DEST (tem
)) != bl
->regno
2534 || GET_CODE (SET_SRC (tem
)) != PLUS
)
2537 /* If final value is non-zero, then must emit an instruction which sets
2538 the value of the biv to the proper value. This is done after
2539 handling all of the givs, since some of them may need to use the
2540 biv's value in their initialization code. */
2542 /* This biv is splittable. If completely unrolling the loop, save
2543 the biv's initial value. Otherwise, save the constant zero. */
2545 if (biv_splittable
== 1)
2547 if (unroll_type
== UNROLL_COMPLETELY
)
2549 /* If the initial value of the biv is itself (i.e. it is too
2550 complicated for strength_reduce to compute), or is a hard
2551 register, or it isn't invariant, then we must create a new
2552 pseudo reg to hold the initial value of the biv. */
2554 if (GET_CODE (bl
->initial_value
) == REG
2555 && (REGNO (bl
->initial_value
) == bl
->regno
2556 || REGNO (bl
->initial_value
) < FIRST_PSEUDO_REGISTER
2557 || ! invariant_p (bl
->initial_value
)))
2559 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2561 record_base_value (REGNO (tem
), bl
->biv
->add_val
);
2562 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2565 if (loop_dump_stream
)
2566 fprintf (loop_dump_stream
, "Biv %d initial value remapped to %d.\n",
2567 bl
->regno
, REGNO (tem
));
2569 splittable_regs
[bl
->regno
] = tem
;
2572 splittable_regs
[bl
->regno
] = bl
->initial_value
;
2575 splittable_regs
[bl
->regno
] = const0_rtx
;
2577 /* Save the number of instructions that modify the biv, so that
2578 we can treat the last one specially. */
2580 splittable_regs_updates
[bl
->regno
] = bl
->biv_count
;
2581 result
+= bl
->biv_count
;
2583 if (loop_dump_stream
)
2584 fprintf (loop_dump_stream
,
2585 "Biv %d safe to split.\n", bl
->regno
);
2588 /* Check every giv that depends on this biv to see whether it is
2589 splittable also. Even if the biv isn't splittable, givs which
2590 depend on it may be splittable if the biv is live outside the
2591 loop, and the givs aren't. */
2593 result
+= find_splittable_givs (bl
, unroll_type
, loop_start
, loop_end
,
2594 increment
, unroll_number
);
2596 /* If final value is non-zero, then must emit an instruction which sets
2597 the value of the biv to the proper value. This is done after
2598 handling all of the givs, since some of them may need to use the
2599 biv's value in their initialization code. */
2600 if (biv_final_value
)
2602 /* If the loop has multiple exits, emit the insns before the
2603 loop to ensure that it will always be executed no matter
2604 how the loop exits. Otherwise emit the insn after the loop,
2605 since this is slightly more efficient. */
2606 if (! loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]])
2607 emit_insn_before (gen_move_insn (bl
->biv
->src_reg
,
2612 /* Create a new register to hold the value of the biv, and then
2613 set the biv to its final value before the loop start. The biv
2614 is set to its final value before loop start to ensure that
2615 this insn will always be executed, no matter how the loop
2617 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2618 record_base_value (REGNO (tem
), bl
->biv
->add_val
);
2620 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2622 emit_insn_before (gen_move_insn (bl
->biv
->src_reg
,
2626 if (loop_dump_stream
)
2627 fprintf (loop_dump_stream
, "Biv %d mapped to %d for split.\n",
2628 REGNO (bl
->biv
->src_reg
), REGNO (tem
));
2630 /* Set up the mapping from the original biv register to the new
2632 bl
->biv
->src_reg
= tem
;
2639 /* Return 1 if the first and last unrolled copy of the address giv V is valid
2640 for the instruction that is using it. Do not make any changes to that
2644 verify_addresses (v
, giv_inc
, unroll_number
)
2645 struct induction
*v
;
2650 rtx orig_addr
= *v
->location
;
2651 rtx last_addr
= plus_constant (v
->dest_reg
,
2652 INTVAL (giv_inc
) * (unroll_number
- 1));
2654 /* First check to see if either address would fail. */
2655 if (! validate_change (v
->insn
, v
->location
, v
->dest_reg
, 0)
2656 || ! validate_change (v
->insn
, v
->location
, last_addr
, 0))
2659 /* Now put things back the way they were before. This will always
2661 validate_change (v
->insn
, v
->location
, orig_addr
, 0);
2666 /* For every giv based on the biv BL, check to determine whether it is
2667 splittable. This is a subroutine to find_splittable_regs ().
2669 Return the number of instructions that set splittable registers. */
2672 find_splittable_givs (bl
, unroll_type
, loop_start
, loop_end
, increment
,
2674 struct iv_class
*bl
;
2675 enum unroll_types unroll_type
;
2676 rtx loop_start
, loop_end
;
2680 struct induction
*v
, *v2
;
2685 /* Scan the list of givs, and set the same_insn field when there are
2686 multiple identical givs in the same insn. */
2687 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
2688 for (v2
= v
->next_iv
; v2
; v2
= v2
->next_iv
)
2689 if (v
->insn
== v2
->insn
&& rtx_equal_p (v
->new_reg
, v2
->new_reg
)
2693 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
2697 /* Only split the giv if it has already been reduced, or if the loop is
2698 being completely unrolled. */
2699 if (unroll_type
!= UNROLL_COMPLETELY
&& v
->ignore
)
2702 /* The giv can be split if the insn that sets the giv is executed once
2703 and only once on every iteration of the loop. */
2704 /* An address giv can always be split. v->insn is just a use not a set,
2705 and hence it does not matter whether it is always executed. All that
2706 matters is that all the biv increments are always executed, and we
2707 won't reach here if they aren't. */
2708 if (v
->giv_type
!= DEST_ADDR
2709 && (! v
->always_computable
2710 || back_branch_in_range_p (v
->insn
, loop_start
, loop_end
)))
2713 /* The giv increment value must be a constant. */
2714 giv_inc
= fold_rtx_mult_add (v
->mult_val
, increment
, const0_rtx
,
2716 if (! giv_inc
|| GET_CODE (giv_inc
) != CONST_INT
)
2719 /* The loop must be unrolled completely, or else have a known number of
2720 iterations and only one exit, or else the giv must be dead outside
2721 the loop, or else the final value of the giv must be known.
2722 Otherwise, it is not safe to split the giv since it may not have the
2723 proper value on loop exit. */
2725 /* The used outside loop test will fail for DEST_ADDR givs. They are
2726 never used outside the loop anyways, so it is always safe to split a
2730 if (unroll_type
!= UNROLL_COMPLETELY
2731 && (loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]]
2732 || unroll_type
== UNROLL_NAIVE
)
2733 && v
->giv_type
!= DEST_ADDR
2734 /* The next part is true if the pseudo is used outside the loop.
2735 We assume that this is true for any pseudo created after loop
2736 starts, because we don't have a reg_n_info entry for them. */
2737 && (REGNO (v
->dest_reg
) >= max_reg_before_loop
2738 || (REGNO_FIRST_UID (REGNO (v
->dest_reg
)) != INSN_UID (v
->insn
)
2739 /* Check for the case where the pseudo is set by a shift/add
2740 sequence, in which case the first insn setting the pseudo
2741 is the first insn of the shift/add sequence. */
2742 && (! (tem
= find_reg_note (v
->insn
, REG_RETVAL
, NULL_RTX
))
2743 || (REGNO_FIRST_UID (REGNO (v
->dest_reg
))
2744 != INSN_UID (XEXP (tem
, 0)))))
2745 /* Line above always fails if INSN was moved by loop opt. */
2746 || (uid_luid
[REGNO_LAST_UID (REGNO (v
->dest_reg
))]
2747 >= INSN_LUID (loop_end
)))
2748 && ! (final_value
= v
->final_value
))
2752 /* Currently, non-reduced/final-value givs are never split. */
2753 /* Should emit insns after the loop if possible, as the biv final value
2756 /* If the final value is non-zero, and the giv has not been reduced,
2757 then must emit an instruction to set the final value. */
2758 if (final_value
&& !v
->new_reg
)
2760 /* Create a new register to hold the value of the giv, and then set
2761 the giv to its final value before the loop start. The giv is set
2762 to its final value before loop start to ensure that this insn
2763 will always be executed, no matter how we exit. */
2764 tem
= gen_reg_rtx (v
->mode
);
2765 emit_insn_before (gen_move_insn (tem
, v
->dest_reg
), loop_start
);
2766 emit_insn_before (gen_move_insn (v
->dest_reg
, final_value
),
2769 if (loop_dump_stream
)
2770 fprintf (loop_dump_stream
, "Giv %d mapped to %d for split.\n",
2771 REGNO (v
->dest_reg
), REGNO (tem
));
2777 /* This giv is splittable. If completely unrolling the loop, save the
2778 giv's initial value. Otherwise, save the constant zero for it. */
2780 if (unroll_type
== UNROLL_COMPLETELY
)
2782 /* It is not safe to use bl->initial_value here, because it may not
2783 be invariant. It is safe to use the initial value stored in
2784 the splittable_regs array if it is set. In rare cases, it won't
2785 be set, so then we do exactly the same thing as
2786 find_splittable_regs does to get a safe value. */
2787 rtx biv_initial_value
;
2789 if (splittable_regs
[bl
->regno
])
2790 biv_initial_value
= splittable_regs
[bl
->regno
];
2791 else if (GET_CODE (bl
->initial_value
) != REG
2792 || (REGNO (bl
->initial_value
) != bl
->regno
2793 && REGNO (bl
->initial_value
) >= FIRST_PSEUDO_REGISTER
))
2794 biv_initial_value
= bl
->initial_value
;
2797 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2799 record_base_value (REGNO (tem
), bl
->biv
->add_val
);
2800 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2802 biv_initial_value
= tem
;
2804 value
= fold_rtx_mult_add (v
->mult_val
, biv_initial_value
,
2805 v
->add_val
, v
->mode
);
2812 /* If a giv was combined with another giv, then we can only split
2813 this giv if the giv it was combined with was reduced. This
2814 is because the value of v->new_reg is meaningless in this
2816 if (v
->same
&& ! v
->same
->new_reg
)
2818 if (loop_dump_stream
)
2819 fprintf (loop_dump_stream
,
2820 "giv combined with unreduced giv not split.\n");
2823 /* If the giv is an address destination, it could be something other
2824 than a simple register, these have to be treated differently. */
2825 else if (v
->giv_type
== DEST_REG
)
2827 /* If value is not a constant, register, or register plus
2828 constant, then compute its value into a register before
2829 loop start. This prevents invalid rtx sharing, and should
2830 generate better code. We can use bl->initial_value here
2831 instead of splittable_regs[bl->regno] because this code
2832 is going before the loop start. */
2833 if (unroll_type
== UNROLL_COMPLETELY
2834 && GET_CODE (value
) != CONST_INT
2835 && GET_CODE (value
) != REG
2836 && (GET_CODE (value
) != PLUS
2837 || GET_CODE (XEXP (value
, 0)) != REG
2838 || GET_CODE (XEXP (value
, 1)) != CONST_INT
))
2840 rtx tem
= gen_reg_rtx (v
->mode
);
2841 record_base_value (REGNO (tem
), v
->add_val
);
2842 emit_iv_add_mult (bl
->initial_value
, v
->mult_val
,
2843 v
->add_val
, tem
, loop_start
);
2847 splittable_regs
[REGNO (v
->new_reg
)] = value
;
2851 /* Splitting address givs is useful since it will often allow us
2852 to eliminate some increment insns for the base giv as
2855 /* If the addr giv is combined with a dest_reg giv, then all
2856 references to that dest reg will be remapped, which is NOT
2857 what we want for split addr regs. We always create a new
2858 register for the split addr giv, just to be safe. */
2860 /* If we have multiple identical address givs within a
2861 single instruction, then use a single pseudo reg for
2862 both. This is necessary in case one is a match_dup
2865 v
->const_adjust
= 0;
2869 v
->dest_reg
= v
->same_insn
->dest_reg
;
2870 if (loop_dump_stream
)
2871 fprintf (loop_dump_stream
,
2872 "Sharing address givs in insn %d\n",
2873 INSN_UID (v
->insn
));
2875 /* If multiple address GIVs have been combined with the
2876 same dest_reg GIV, do not create a new register for
2878 else if (unroll_type
!= UNROLL_COMPLETELY
2879 && v
->giv_type
== DEST_ADDR
2880 && v
->same
&& v
->same
->giv_type
== DEST_ADDR
2881 && v
->same
->unrolled
2882 /* combine_givs_p may return true for some cases
2883 where the add and mult values are not equal.
2884 To share a register here, the values must be
2886 && rtx_equal_p (v
->same
->mult_val
, v
->mult_val
)
2887 && rtx_equal_p (v
->same
->add_val
, v
->add_val
))
2890 v
->dest_reg
= v
->same
->dest_reg
;
2893 else if (unroll_type
!= UNROLL_COMPLETELY
)
2895 /* If not completely unrolling the loop, then create a new
2896 register to hold the split value of the DEST_ADDR giv.
2897 Emit insn to initialize its value before loop start. */
2899 rtx tem
= gen_reg_rtx (v
->mode
);
2900 record_base_value (REGNO (tem
), v
->add_val
);
2903 /* If the address giv has a constant in its new_reg value,
2904 then this constant can be pulled out and put in value,
2905 instead of being part of the initialization code. */
2907 if (GET_CODE (v
->new_reg
) == PLUS
2908 && GET_CODE (XEXP (v
->new_reg
, 1)) == CONST_INT
)
2911 = plus_constant (tem
, INTVAL (XEXP (v
->new_reg
,1)));
2913 /* Only succeed if this will give valid addresses.
2914 Try to validate both the first and the last
2915 address resulting from loop unrolling, if
2916 one fails, then can't do const elim here. */
2917 if (verify_addresses (v
, giv_inc
, unroll_number
))
2919 /* Save the negative of the eliminated const, so
2920 that we can calculate the dest_reg's increment
2922 v
->const_adjust
= - INTVAL (XEXP (v
->new_reg
, 1));
2924 v
->new_reg
= XEXP (v
->new_reg
, 0);
2925 if (loop_dump_stream
)
2926 fprintf (loop_dump_stream
,
2927 "Eliminating constant from giv %d\n",
2936 /* If the address hasn't been checked for validity yet, do so
2937 now, and fail completely if either the first or the last
2938 unrolled copy of the address is not a valid address
2939 for the instruction that uses it. */
2940 if (v
->dest_reg
== tem
2941 && ! verify_addresses (v
, giv_inc
, unroll_number
))
2943 if (loop_dump_stream
)
2944 fprintf (loop_dump_stream
,
2945 "Invalid address for giv at insn %d\n",
2946 INSN_UID (v
->insn
));
2950 /* To initialize the new register, just move the value of
2951 new_reg into it. This is not guaranteed to give a valid
2952 instruction on machines with complex addressing modes.
2953 If we can't recognize it, then delete it and emit insns
2954 to calculate the value from scratch. */
2955 emit_insn_before (gen_rtx_SET (VOIDmode
, tem
,
2956 copy_rtx (v
->new_reg
)),
2958 if (recog_memoized (PREV_INSN (loop_start
)) < 0)
2962 /* We can't use bl->initial_value to compute the initial
2963 value, because the loop may have been preconditioned.
2964 We must calculate it from NEW_REG. Try using
2965 force_operand instead of emit_iv_add_mult. */
2966 delete_insn (PREV_INSN (loop_start
));
2969 ret
= force_operand (v
->new_reg
, tem
);
2971 emit_move_insn (tem
, ret
);
2972 sequence
= gen_sequence ();
2974 emit_insn_before (sequence
, loop_start
);
2976 if (loop_dump_stream
)
2977 fprintf (loop_dump_stream
,
2978 "Invalid init insn, rewritten.\n");
2983 v
->dest_reg
= value
;
2985 /* Check the resulting address for validity, and fail
2986 if the resulting address would be invalid. */
2987 if (! verify_addresses (v
, giv_inc
, unroll_number
))
2989 if (loop_dump_stream
)
2990 fprintf (loop_dump_stream
,
2991 "Invalid address for giv at insn %d\n",
2992 INSN_UID (v
->insn
));
2997 /* Store the value of dest_reg into the insn. This sharing
2998 will not be a problem as this insn will always be copied
3001 *v
->location
= v
->dest_reg
;
3003 /* If this address giv is combined with a dest reg giv, then
3004 save the base giv's induction pointer so that we will be
3005 able to handle this address giv properly. The base giv
3006 itself does not have to be splittable. */
3008 if (v
->same
&& v
->same
->giv_type
== DEST_REG
)
3009 addr_combined_regs
[REGNO (v
->same
->new_reg
)] = v
->same
;
3011 if (GET_CODE (v
->new_reg
) == REG
)
3013 /* This giv maybe hasn't been combined with any others.
3014 Make sure that it's giv is marked as splittable here. */
3016 splittable_regs
[REGNO (v
->new_reg
)] = value
;
3018 /* Make it appear to depend upon itself, so that the
3019 giv will be properly split in the main loop above. */
3023 addr_combined_regs
[REGNO (v
->new_reg
)] = v
;
3027 if (loop_dump_stream
)
3028 fprintf (loop_dump_stream
, "DEST_ADDR giv being split.\n");
3034 /* Currently, unreduced giv's can't be split. This is not too much
3035 of a problem since unreduced giv's are not live across loop
3036 iterations anyways. When unrolling a loop completely though,
3037 it makes sense to reduce&split givs when possible, as this will
3038 result in simpler instructions, and will not require that a reg
3039 be live across loop iterations. */
3041 splittable_regs
[REGNO (v
->dest_reg
)] = value
;
3042 fprintf (stderr
, "Giv %d at insn %d not reduced\n",
3043 REGNO (v
->dest_reg
), INSN_UID (v
->insn
));
3049 /* Unreduced givs are only updated once by definition. Reduced givs
3050 are updated as many times as their biv is. Mark it so if this is
3051 a splittable register. Don't need to do anything for address givs
3052 where this may not be a register. */
3054 if (GET_CODE (v
->new_reg
) == REG
)
3058 count
= reg_biv_class
[REGNO (v
->src_reg
)]->biv_count
;
3060 splittable_regs_updates
[REGNO (v
->new_reg
)] = count
;
3065 if (loop_dump_stream
)
3069 if (GET_CODE (v
->dest_reg
) == CONST_INT
)
3071 else if (GET_CODE (v
->dest_reg
) != REG
)
3072 regnum
= REGNO (XEXP (v
->dest_reg
, 0));
3074 regnum
= REGNO (v
->dest_reg
);
3075 fprintf (loop_dump_stream
, "Giv %d at insn %d safe to split.\n",
3076 regnum
, INSN_UID (v
->insn
));
3083 /* Try to prove that the register is dead after the loop exits. Trace every
3084 loop exit looking for an insn that will always be executed, which sets
3085 the register to some value, and appears before the first use of the register
3086 is found. If successful, then return 1, otherwise return 0. */
3088 /* ?? Could be made more intelligent in the handling of jumps, so that
3089 it can search past if statements and other similar structures. */
3092 reg_dead_after_loop (reg
, loop_start
, loop_end
)
3093 rtx reg
, loop_start
, loop_end
;
3098 int label_count
= 0;
3099 int this_loop_num
= uid_loop_num
[INSN_UID (loop_start
)];
3101 /* In addition to checking all exits of this loop, we must also check
3102 all exits of inner nested loops that would exit this loop. We don't
3103 have any way to identify those, so we just give up if there are any
3104 such inner loop exits. */
3106 for (label
= loop_number_exit_labels
[this_loop_num
]; label
;
3107 label
= LABEL_NEXTREF (label
))
3110 if (label_count
!= loop_number_exit_count
[this_loop_num
])
3113 /* HACK: Must also search the loop fall through exit, create a label_ref
3114 here which points to the loop_end, and append the loop_number_exit_labels
3116 label
= gen_rtx_LABEL_REF (VOIDmode
, loop_end
);
3117 LABEL_NEXTREF (label
) = loop_number_exit_labels
[this_loop_num
];
3119 for ( ; label
; label
= LABEL_NEXTREF (label
))
3121 /* Succeed if find an insn which sets the biv or if reach end of
3122 function. Fail if find an insn that uses the biv, or if come to
3123 a conditional jump. */
3125 insn
= NEXT_INSN (XEXP (label
, 0));
3128 code
= GET_CODE (insn
);
3129 if (GET_RTX_CLASS (code
) == 'i')
3133 if (reg_referenced_p (reg
, PATTERN (insn
)))
3136 set
= single_set (insn
);
3137 if (set
&& rtx_equal_p (SET_DEST (set
), reg
))
3141 if (code
== JUMP_INSN
)
3143 if (GET_CODE (PATTERN (insn
)) == RETURN
)
3145 else if (! simplejump_p (insn
)
3146 /* Prevent infinite loop following infinite loops. */
3147 || jump_count
++ > 20)
3150 insn
= JUMP_LABEL (insn
);
3153 insn
= NEXT_INSN (insn
);
3157 /* Success, the register is dead on all loop exits. */
3161 /* Try to calculate the final value of the biv, the value it will have at
3162 the end of the loop. If we can do it, return that value. */
3165 final_biv_value (bl
, loop_start
, loop_end
)
3166 struct iv_class
*bl
;
3167 rtx loop_start
, loop_end
;
3171 /* ??? This only works for MODE_INT biv's. Reject all others for now. */
3173 if (GET_MODE_CLASS (bl
->biv
->mode
) != MODE_INT
)
3176 /* The final value for reversed bivs must be calculated differently than
3177 for ordinary bivs. In this case, there is already an insn after the
3178 loop which sets this biv's final value (if necessary), and there are
3179 no other loop exits, so we can return any value. */
3182 if (loop_dump_stream
)
3183 fprintf (loop_dump_stream
,
3184 "Final biv value for %d, reversed biv.\n", bl
->regno
);
3189 /* Try to calculate the final value as initial value + (number of iterations
3190 * increment). For this to work, increment must be invariant, the only
3191 exit from the loop must be the fall through at the bottom (otherwise
3192 it may not have its final value when the loop exits), and the initial
3193 value of the biv must be invariant. */
3195 if (loop_n_iterations
!= 0
3196 && ! loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]]
3197 && invariant_p (bl
->initial_value
))
3199 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
3201 if (increment
&& invariant_p (increment
))
3203 /* Can calculate the loop exit value, emit insns after loop
3204 end to calculate this value into a temporary register in
3205 case it is needed later. */
3207 tem
= gen_reg_rtx (bl
->biv
->mode
);
3208 record_base_value (REGNO (tem
), bl
->biv
->add_val
);
3209 /* Make sure loop_end is not the last insn. */
3210 if (NEXT_INSN (loop_end
) == 0)
3211 emit_note_after (NOTE_INSN_DELETED
, loop_end
);
3212 emit_iv_add_mult (increment
, GEN_INT (loop_n_iterations
),
3213 bl
->initial_value
, tem
, NEXT_INSN (loop_end
));
3215 if (loop_dump_stream
)
3216 fprintf (loop_dump_stream
,
3217 "Final biv value for %d, calculated.\n", bl
->regno
);
3223 /* Check to see if the biv is dead at all loop exits. */
3224 if (reg_dead_after_loop (bl
->biv
->src_reg
, loop_start
, loop_end
))
3226 if (loop_dump_stream
)
3227 fprintf (loop_dump_stream
,
3228 "Final biv value for %d, biv dead after loop exit.\n",
3237 /* Try to calculate the final value of the giv, the value it will have at
3238 the end of the loop. If we can do it, return that value. */
3241 final_giv_value (v
, loop_start
, loop_end
)
3242 struct induction
*v
;
3243 rtx loop_start
, loop_end
;
3245 struct iv_class
*bl
;
3248 rtx insert_before
, seq
;
3250 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
3252 /* The final value for givs which depend on reversed bivs must be calculated
3253 differently than for ordinary givs. In this case, there is already an
3254 insn after the loop which sets this giv's final value (if necessary),
3255 and there are no other loop exits, so we can return any value. */
3258 if (loop_dump_stream
)
3259 fprintf (loop_dump_stream
,
3260 "Final giv value for %d, depends on reversed biv\n",
3261 REGNO (v
->dest_reg
));
3265 /* Try to calculate the final value as a function of the biv it depends
3266 upon. The only exit from the loop must be the fall through at the bottom
3267 (otherwise it may not have its final value when the loop exits). */
3269 /* ??? Can calculate the final giv value by subtracting off the
3270 extra biv increments times the giv's mult_val. The loop must have
3271 only one exit for this to work, but the loop iterations does not need
3274 if (loop_n_iterations
!= 0
3275 && ! loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]])
3277 /* ?? It is tempting to use the biv's value here since these insns will
3278 be put after the loop, and hence the biv will have its final value
3279 then. However, this fails if the biv is subsequently eliminated.
3280 Perhaps determine whether biv's are eliminable before trying to
3281 determine whether giv's are replaceable so that we can use the
3282 biv value here if it is not eliminable. */
3284 /* We are emitting code after the end of the loop, so we must make
3285 sure that bl->initial_value is still valid then. It will still
3286 be valid if it is invariant. */
3288 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
3290 if (increment
&& invariant_p (increment
)
3291 && invariant_p (bl
->initial_value
))
3293 /* Can calculate the loop exit value of its biv as
3294 (loop_n_iterations * increment) + initial_value */
3296 /* The loop exit value of the giv is then
3297 (final_biv_value - extra increments) * mult_val + add_val.
3298 The extra increments are any increments to the biv which
3299 occur in the loop after the giv's value is calculated.
3300 We must search from the insn that sets the giv to the end
3301 of the loop to calculate this value. */
3303 insert_before
= NEXT_INSN (loop_end
);
3305 /* Put the final biv value in tem. */
3306 tem
= gen_reg_rtx (bl
->biv
->mode
);
3307 record_base_value (REGNO (tem
), bl
->biv
->add_val
);
3308 emit_iv_add_mult (increment
, GEN_INT (loop_n_iterations
),
3309 bl
->initial_value
, tem
, insert_before
);
3311 /* Subtract off extra increments as we find them. */
3312 for (insn
= NEXT_INSN (v
->insn
); insn
!= loop_end
;
3313 insn
= NEXT_INSN (insn
))
3315 struct induction
*biv
;
3317 for (biv
= bl
->biv
; biv
; biv
= biv
->next_iv
)
3318 if (biv
->insn
== insn
)
3321 tem
= expand_binop (GET_MODE (tem
), sub_optab
, tem
,
3322 biv
->add_val
, NULL_RTX
, 0,
3324 seq
= gen_sequence ();
3326 emit_insn_before (seq
, insert_before
);
3330 /* Now calculate the giv's final value. */
3331 emit_iv_add_mult (tem
, v
->mult_val
, v
->add_val
, tem
,
3334 if (loop_dump_stream
)
3335 fprintf (loop_dump_stream
,
3336 "Final giv value for %d, calc from biv's value.\n",
3337 REGNO (v
->dest_reg
));
3343 /* Replaceable giv's should never reach here. */
3347 /* Check to see if the biv is dead at all loop exits. */
3348 if (reg_dead_after_loop (v
->dest_reg
, loop_start
, loop_end
))
3350 if (loop_dump_stream
)
3351 fprintf (loop_dump_stream
,
3352 "Final giv value for %d, giv dead after loop exit.\n",
3353 REGNO (v
->dest_reg
));
3362 /* Calculate the number of loop iterations. Returns the exact number of loop
3363 iterations if it can be calculated, otherwise returns zero. */
3365 unsigned HOST_WIDE_INT
3366 loop_iterations (loop_start
, loop_end
)
3367 rtx loop_start
, loop_end
;
3369 rtx comparison
, comparison_value
;
3370 rtx iteration_var
, initial_value
, increment
, final_value
;
3371 enum rtx_code comparison_code
;
3374 int unsigned_compare
, compare_dir
, final_larger
;
3375 unsigned long tempu
;
3378 /* First find the iteration variable. If the last insn is a conditional
3379 branch, and the insn before tests a register value, make that the
3380 iteration variable. */
3382 loop_initial_value
= 0;
3384 loop_final_value
= 0;
3385 loop_iteration_var
= 0;
3387 /* We used to use pren_nonnote_insn here, but that fails because it might
3388 accidentally get the branch for a contained loop if the branch for this
3389 loop was deleted. We can only trust branches immediately before the
3391 last_loop_insn
= PREV_INSN (loop_end
);
3393 comparison
= get_condition_for_loop (last_loop_insn
);
3394 if (comparison
== 0)
3396 if (loop_dump_stream
)
3397 fprintf (loop_dump_stream
,
3398 "Loop unrolling: No final conditional branch found.\n");
3402 /* ??? Get_condition may switch position of induction variable and
3403 invariant register when it canonicalizes the comparison. */
3405 comparison_code
= GET_CODE (comparison
);
3406 iteration_var
= XEXP (comparison
, 0);
3407 comparison_value
= XEXP (comparison
, 1);
3409 if (GET_CODE (iteration_var
) != REG
)
3411 if (loop_dump_stream
)
3412 fprintf (loop_dump_stream
,
3413 "Loop unrolling: Comparison not against register.\n");
3417 /* Loop iterations is always called before any new registers are created
3418 now, so this should never occur. */
3420 if (REGNO (iteration_var
) >= max_reg_before_loop
)
3423 iteration_info (iteration_var
, &initial_value
, &increment
,
3424 loop_start
, loop_end
);
3425 if (initial_value
== 0)
3426 /* iteration_info already printed a message. */
3429 /* If the comparison value is an invariant register, then try to find
3430 its value from the insns before the start of the loop. */
3432 if (GET_CODE (comparison_value
) == REG
&& invariant_p (comparison_value
))
3436 for (insn
= PREV_INSN (loop_start
); insn
; insn
= PREV_INSN (insn
))
3438 if (GET_CODE (insn
) == CODE_LABEL
)
3441 else if (GET_RTX_CLASS (GET_CODE (insn
)) == 'i'
3442 && reg_set_p (comparison_value
, insn
))
3444 /* We found the last insn before the loop that sets the register.
3445 If it sets the entire register, and has a REG_EQUAL note,
3446 then use the value of the REG_EQUAL note. */
3447 if ((set
= single_set (insn
))
3448 && (SET_DEST (set
) == comparison_value
))
3450 rtx note
= find_reg_note (insn
, REG_EQUAL
, NULL_RTX
);
3452 /* Only use the REG_EQUAL note if it is a constant.
3453 Other things, divide in particular, will cause
3454 problems later if we use them. */
3455 if (note
&& GET_CODE (XEXP (note
, 0)) != EXPR_LIST
3456 && CONSTANT_P (XEXP (note
, 0)))
3457 comparison_value
= XEXP (note
, 0);
3464 final_value
= approx_final_value (comparison_code
, comparison_value
,
3465 &unsigned_compare
, &compare_dir
);
3467 /* Save the calculated values describing this loop's bounds, in case
3468 precondition_loop_p will need them later. These values can not be
3469 recalculated inside precondition_loop_p because strength reduction
3470 optimizations may obscure the loop's structure. */
3472 loop_iteration_var
= iteration_var
;
3473 loop_initial_value
= initial_value
;
3474 loop_increment
= increment
;
3475 loop_final_value
= final_value
;
3476 loop_comparison_code
= comparison_code
;
3480 if (loop_dump_stream
)
3481 fprintf (loop_dump_stream
,
3482 "Loop unrolling: Increment value can't be calculated.\n");
3485 else if (GET_CODE (increment
) != CONST_INT
)
3487 if (loop_dump_stream
)
3488 fprintf (loop_dump_stream
,
3489 "Loop unrolling: Increment value not constant.\n");
3492 else if (GET_CODE (initial_value
) != CONST_INT
)
3494 if (loop_dump_stream
)
3495 fprintf (loop_dump_stream
,
3496 "Loop unrolling: Initial value not constant.\n");
3499 else if (final_value
== 0)
3501 if (loop_dump_stream
)
3502 fprintf (loop_dump_stream
,
3503 "Loop unrolling: EQ comparison loop.\n");
3506 else if (GET_CODE (final_value
) != CONST_INT
)
3508 if (loop_dump_stream
)
3509 fprintf (loop_dump_stream
,
3510 "Loop unrolling: Final value not constant.\n");
3514 /* ?? Final value and initial value do not have to be constants.
3515 Only their difference has to be constant. When the iteration variable
3516 is an array address, the final value and initial value might both
3517 be addresses with the same base but different constant offsets.
3518 Final value must be invariant for this to work.
3520 To do this, need some way to find the values of registers which are
3523 /* Final_larger is 1 if final larger, 0 if they are equal, otherwise -1. */
3524 if (unsigned_compare
)
3526 = ((unsigned HOST_WIDE_INT
) INTVAL (final_value
)
3527 > (unsigned HOST_WIDE_INT
) INTVAL (initial_value
))
3528 - ((unsigned HOST_WIDE_INT
) INTVAL (final_value
)
3529 < (unsigned HOST_WIDE_INT
) INTVAL (initial_value
));
3531 final_larger
= (INTVAL (final_value
) > INTVAL (initial_value
))
3532 - (INTVAL (final_value
) < INTVAL (initial_value
));
3534 if (INTVAL (increment
) > 0)
3536 else if (INTVAL (increment
) == 0)
3541 /* There are 27 different cases: compare_dir = -1, 0, 1;
3542 final_larger = -1, 0, 1; increment_dir = -1, 0, 1.
3543 There are 4 normal cases, 4 reverse cases (where the iteration variable
3544 will overflow before the loop exits), 4 infinite loop cases, and 15
3545 immediate exit (0 or 1 iteration depending on loop type) cases.
3546 Only try to optimize the normal cases. */
3548 /* (compare_dir/final_larger/increment_dir)
3549 Normal cases: (0/-1/-1), (0/1/1), (-1/-1/-1), (1/1/1)
3550 Reverse cases: (0/-1/1), (0/1/-1), (-1/-1/1), (1/1/-1)
3551 Infinite loops: (0/-1/0), (0/1/0), (-1/-1/0), (1/1/0)
3552 Immediate exit: (0/0/X), (-1/0/X), (-1/1/X), (1/0/X), (1/-1/X) */
3554 /* ?? If the meaning of reverse loops (where the iteration variable
3555 will overflow before the loop exits) is undefined, then could
3556 eliminate all of these special checks, and just always assume
3557 the loops are normal/immediate/infinite. Note that this means
3558 the sign of increment_dir does not have to be known. Also,
3559 since it does not really hurt if immediate exit loops or infinite loops
3560 are optimized, then that case could be ignored also, and hence all
3561 loops can be optimized.
3563 According to ANSI Spec, the reverse loop case result is undefined,
3564 because the action on overflow is undefined.
3566 See also the special test for NE loops below. */
3568 if (final_larger
== increment_dir
&& final_larger
!= 0
3569 && (final_larger
== compare_dir
|| compare_dir
== 0))
3574 if (loop_dump_stream
)
3575 fprintf (loop_dump_stream
,
3576 "Loop unrolling: Not normal loop.\n");
3580 /* Calculate the number of iterations, final_value is only an approximation,
3581 so correct for that. Note that tempu and loop_n_iterations are
3582 unsigned, because they can be as large as 2^n - 1. */
3584 i
= INTVAL (increment
);
3586 tempu
= INTVAL (final_value
) - INTVAL (initial_value
);
3589 tempu
= INTVAL (initial_value
) - INTVAL (final_value
);
3595 /* For NE tests, make sure that the iteration variable won't miss the
3596 final value. If tempu mod i is not zero, then the iteration variable
3597 will overflow before the loop exits, and we can not calculate the
3598 number of iterations. */
3599 if (compare_dir
== 0 && (tempu
% i
) != 0)
3602 return tempu
/ i
+ ((tempu
% i
) != 0);
3605 /* Replace uses of split bivs with their split pseudo register. This is
3606 for original instructions which remain after loop unrolling without
3610 remap_split_bivs (x
)
3613 register enum rtx_code code
;
3620 code
= GET_CODE (x
);
3635 /* If non-reduced/final-value givs were split, then this would also
3636 have to remap those givs also. */
3638 if (REGNO (x
) < max_reg_before_loop
3639 && reg_iv_type
[REGNO (x
)] == BASIC_INDUCT
)
3640 return reg_biv_class
[REGNO (x
)]->biv
->src_reg
;
3647 fmt
= GET_RTX_FORMAT (code
);
3648 for (i
= GET_RTX_LENGTH (code
) - 1; i
>= 0; i
--)
3651 XEXP (x
, i
) = remap_split_bivs (XEXP (x
, i
));
3655 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
3656 XVECEXP (x
, i
, j
) = remap_split_bivs (XVECEXP (x
, i
, j
));
3662 /* If FIRST_UID is a set of REGNO, and FIRST_UID dominates LAST_UID (e.g.
3663 FIST_UID is always executed if LAST_UID is), then return 1. Otherwise
3664 return 0. COPY_START is where we can start looking for the insns
3665 FIRST_UID and LAST_UID. COPY_END is where we stop looking for these
3668 If there is no JUMP_INSN between LOOP_START and FIRST_UID, then FIRST_UID
3669 must dominate LAST_UID.
3671 If there is a CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID
3672 may not dominate LAST_UID.
3674 If there is no CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID
3675 must dominate LAST_UID. */
3678 set_dominates_use (regno
, first_uid
, last_uid
, copy_start
, copy_end
)
3685 int passed_jump
= 0;
3686 rtx p
= NEXT_INSN (copy_start
);
3688 while (INSN_UID (p
) != first_uid
)
3690 if (GET_CODE (p
) == JUMP_INSN
)
3692 /* Could not find FIRST_UID. */
3698 /* Verify that FIRST_UID is an insn that entirely sets REGNO. */
3699 if (GET_RTX_CLASS (GET_CODE (p
)) != 'i'
3700 || ! dead_or_set_regno_p (p
, regno
))
3703 /* FIRST_UID is always executed. */
3704 if (passed_jump
== 0)
3707 while (INSN_UID (p
) != last_uid
)
3709 /* If we see a CODE_LABEL between FIRST_UID and LAST_UID, then we
3710 can not be sure that FIRST_UID dominates LAST_UID. */
3711 if (GET_CODE (p
) == CODE_LABEL
)
3713 /* Could not find LAST_UID, but we reached the end of the loop, so
3715 else if (p
== copy_end
)
3720 /* FIRST_UID is always executed if LAST_UID is executed. */