1 /* Try to unroll loops, and split induction variables.
2 Copyright (C) 1992, 1993, 1994, 1995 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
};
151 #include "insn-config.h"
152 #include "integrate.h"
159 /* This controls which loops are unrolled, and by how much we unroll
162 #ifndef MAX_UNROLLED_INSNS
163 #define MAX_UNROLLED_INSNS 100
166 /* Indexed by register number, if non-zero, then it contains a pointer
167 to a struct induction for a DEST_REG giv which has been combined with
168 one of more address givs. This is needed because whenever such a DEST_REG
169 giv is modified, we must modify the value of all split address givs
170 that were combined with this DEST_REG giv. */
172 static struct induction
**addr_combined_regs
;
174 /* Indexed by register number, if this is a splittable induction variable,
175 then this will hold the current value of the register, which depends on the
178 static rtx
*splittable_regs
;
180 /* Indexed by register number, if this is a splittable induction variable,
181 then this will hold the number of instructions in the loop that modify
182 the induction variable. Used to ensure that only the last insn modifying
183 a split iv will update the original iv of the dest. */
185 static int *splittable_regs_updates
;
187 /* Values describing the current loop's iteration variable. These are set up
188 by loop_iterations, and used by precondition_loop_p. */
190 static rtx loop_iteration_var
;
191 static rtx loop_initial_value
;
192 static rtx loop_increment
;
193 static rtx loop_final_value
;
194 static enum rtx_code loop_comparison_code
;
196 /* Forward declarations. */
198 static void init_reg_map
PROTO((struct inline_remap
*, int));
199 static int precondition_loop_p
PROTO((rtx
*, rtx
*, rtx
*, rtx
, rtx
));
200 static rtx calculate_giv_inc
PROTO((rtx
, rtx
, int));
201 static rtx initial_reg_note_copy
PROTO((rtx
, struct inline_remap
*));
202 static void final_reg_note_copy
PROTO((rtx
, struct inline_remap
*));
203 static void copy_loop_body
PROTO((rtx
, rtx
, struct inline_remap
*, rtx
, int,
204 enum unroll_types
, rtx
, rtx
, rtx
, rtx
));
205 void iteration_info
PROTO((rtx
, rtx
*, rtx
*, rtx
, rtx
));
206 static rtx approx_final_value
PROTO((enum rtx_code
, rtx
, int *, int *));
207 static int find_splittable_regs
PROTO((enum unroll_types
, rtx
, rtx
, rtx
, int));
208 static int find_splittable_givs
PROTO((struct iv_class
*,enum unroll_types
,
209 rtx
, rtx
, rtx
, int));
210 static int reg_dead_after_loop
PROTO((rtx
, rtx
, rtx
));
211 static rtx fold_rtx_mult_add
PROTO((rtx
, rtx
, rtx
, enum machine_mode
));
212 static rtx remap_split_bivs
PROTO((rtx
));
214 /* Try to unroll one loop and split induction variables in the loop.
216 The loop is described by the arguments LOOP_END, INSN_COUNT, and
217 LOOP_START. END_INSERT_BEFORE indicates where insns should be added
218 which need to be executed when the loop falls through. STRENGTH_REDUCTION_P
219 indicates whether information generated in the strength reduction pass
222 This function is intended to be called from within `strength_reduce'
226 unroll_loop (loop_end
, insn_count
, loop_start
, end_insert_before
,
231 rtx end_insert_before
;
232 int strength_reduce_p
;
235 int unroll_number
= 1;
236 rtx copy_start
, copy_end
;
237 rtx insn
, copy
, sequence
, pattern
, tem
;
238 int max_labelno
, max_insnno
;
240 struct inline_remap
*map
;
248 int splitting_not_safe
= 0;
249 enum unroll_types unroll_type
;
250 int loop_preconditioned
= 0;
252 /* This points to the last real insn in the loop, which should be either
253 a JUMP_INSN (for conditional jumps) or a BARRIER (for unconditional
257 /* Don't bother unrolling huge loops. Since the minimum factor is
258 two, loops greater than one half of MAX_UNROLLED_INSNS will never
260 if (insn_count
> MAX_UNROLLED_INSNS
/ 2)
262 if (loop_dump_stream
)
263 fprintf (loop_dump_stream
, "Unrolling failure: Loop too big.\n");
267 /* When emitting debugger info, we can't unroll loops with unequal numbers
268 of block_beg and block_end notes, because that would unbalance the block
269 structure of the function. This can happen as a result of the
270 "if (foo) bar; else break;" optimization in jump.c. */
271 /* ??? Gcc has a general policy that -g is never supposed to change the code
272 that the compiler emits, so we must disable this optimization always,
273 even if debug info is not being output. This is rare, so this should
274 not be a significant performance problem. */
276 if (1 /* write_symbols != NO_DEBUG */)
278 int block_begins
= 0;
281 for (insn
= loop_start
; insn
!= loop_end
; insn
= NEXT_INSN (insn
))
283 if (GET_CODE (insn
) == NOTE
)
285 if (NOTE_LINE_NUMBER (insn
) == NOTE_INSN_BLOCK_BEG
)
287 else if (NOTE_LINE_NUMBER (insn
) == NOTE_INSN_BLOCK_END
)
292 if (block_begins
!= block_ends
)
294 if (loop_dump_stream
)
295 fprintf (loop_dump_stream
,
296 "Unrolling failure: Unbalanced block notes.\n");
301 /* Determine type of unroll to perform. Depends on the number of iterations
302 and the size of the loop. */
304 /* If there is no strength reduce info, then set loop_n_iterations to zero.
305 This can happen if strength_reduce can't find any bivs in the loop.
306 A value of zero indicates that the number of iterations could not be
309 if (! strength_reduce_p
)
310 loop_n_iterations
= 0;
312 if (loop_dump_stream
&& loop_n_iterations
> 0)
313 fprintf (loop_dump_stream
,
314 "Loop unrolling: %d iterations.\n", loop_n_iterations
);
316 /* Find and save a pointer to the last nonnote insn in the loop. */
318 last_loop_insn
= prev_nonnote_insn (loop_end
);
320 /* Calculate how many times to unroll the loop. Indicate whether or
321 not the loop is being completely unrolled. */
323 if (loop_n_iterations
== 1)
325 /* If number of iterations is exactly 1, then eliminate the compare and
326 branch at the end of the loop since they will never be taken.
327 Then return, since no other action is needed here. */
329 /* If the last instruction is not a BARRIER or a JUMP_INSN, then
330 don't do anything. */
332 if (GET_CODE (last_loop_insn
) == BARRIER
)
334 /* Delete the jump insn. This will delete the barrier also. */
335 delete_insn (PREV_INSN (last_loop_insn
));
337 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
340 /* The immediately preceding insn is a compare which must be
342 delete_insn (last_loop_insn
);
343 delete_insn (PREV_INSN (last_loop_insn
));
345 /* The immediately preceding insn may not be the compare, so don't
347 delete_insn (last_loop_insn
);
352 else if (loop_n_iterations
> 0
353 && loop_n_iterations
* insn_count
< MAX_UNROLLED_INSNS
)
355 unroll_number
= loop_n_iterations
;
356 unroll_type
= UNROLL_COMPLETELY
;
358 else if (loop_n_iterations
> 0)
360 /* Try to factor the number of iterations. Don't bother with the
361 general case, only using 2, 3, 5, and 7 will get 75% of all
362 numbers theoretically, and almost all in practice. */
364 for (i
= 0; i
< NUM_FACTORS
; i
++)
365 factors
[i
].count
= 0;
367 temp
= loop_n_iterations
;
368 for (i
= NUM_FACTORS
- 1; i
>= 0; i
--)
369 while (temp
% factors
[i
].factor
== 0)
372 temp
= temp
/ factors
[i
].factor
;
375 /* Start with the larger factors first so that we generally
376 get lots of unrolling. */
380 for (i
= 3; i
>= 0; i
--)
381 while (factors
[i
].count
--)
383 if (temp
* factors
[i
].factor
< MAX_UNROLLED_INSNS
)
385 unroll_number
*= factors
[i
].factor
;
386 temp
*= factors
[i
].factor
;
392 /* If we couldn't find any factors, then unroll as in the normal
394 if (unroll_number
== 1)
396 if (loop_dump_stream
)
397 fprintf (loop_dump_stream
,
398 "Loop unrolling: No factors found.\n");
401 unroll_type
= UNROLL_MODULO
;
405 /* Default case, calculate number of times to unroll loop based on its
407 if (unroll_number
== 1)
409 if (8 * insn_count
< MAX_UNROLLED_INSNS
)
411 else if (4 * insn_count
< MAX_UNROLLED_INSNS
)
416 unroll_type
= UNROLL_NAIVE
;
419 /* Now we know how many times to unroll the loop. */
421 if (loop_dump_stream
)
422 fprintf (loop_dump_stream
,
423 "Unrolling loop %d times.\n", unroll_number
);
426 if (unroll_type
== UNROLL_COMPLETELY
|| unroll_type
== UNROLL_MODULO
)
428 /* Loops of these types should never start with a jump down to
429 the exit condition test. For now, check for this case just to
430 be sure. UNROLL_NAIVE loops can be of this form, this case is
433 while (GET_CODE (insn
) != CODE_LABEL
&& GET_CODE (insn
) != JUMP_INSN
)
434 insn
= NEXT_INSN (insn
);
435 if (GET_CODE (insn
) == JUMP_INSN
)
439 if (unroll_type
== UNROLL_COMPLETELY
)
441 /* Completely unrolling the loop: Delete the compare and branch at
442 the end (the last two instructions). This delete must done at the
443 very end of loop unrolling, to avoid problems with calls to
444 back_branch_in_range_p, which is called by find_splittable_regs.
445 All increments of splittable bivs/givs are changed to load constant
448 copy_start
= loop_start
;
450 /* Set insert_before to the instruction immediately after the JUMP_INSN
451 (or BARRIER), so that any NOTEs between the JUMP_INSN and the end of
452 the loop will be correctly handled by copy_loop_body. */
453 insert_before
= NEXT_INSN (last_loop_insn
);
455 /* Set copy_end to the insn before the jump at the end of the loop. */
456 if (GET_CODE (last_loop_insn
) == BARRIER
)
457 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
458 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
461 /* The instruction immediately before the JUMP_INSN is a compare
462 instruction which we do not want to copy. */
463 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
465 /* The instruction immediately before the JUMP_INSN may not be the
466 compare, so we must copy it. */
467 copy_end
= PREV_INSN (last_loop_insn
);
472 /* We currently can't unroll a loop if it doesn't end with a
473 JUMP_INSN. There would need to be a mechanism that recognizes
474 this case, and then inserts a jump after each loop body, which
475 jumps to after the last loop body. */
476 if (loop_dump_stream
)
477 fprintf (loop_dump_stream
,
478 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
482 else if (unroll_type
== UNROLL_MODULO
)
484 /* Partially unrolling the loop: The compare and branch at the end
485 (the last two instructions) must remain. Don't copy the compare
486 and branch instructions at the end of the loop. Insert the unrolled
487 code immediately before the compare/branch at the end so that the
488 code will fall through to them as before. */
490 copy_start
= loop_start
;
492 /* Set insert_before to the jump insn at the end of the loop.
493 Set copy_end to before the jump insn at the end of the loop. */
494 if (GET_CODE (last_loop_insn
) == BARRIER
)
496 insert_before
= PREV_INSN (last_loop_insn
);
497 copy_end
= PREV_INSN (insert_before
);
499 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
502 /* The instruction immediately before the JUMP_INSN is a compare
503 instruction which we do not want to copy or delete. */
504 insert_before
= PREV_INSN (last_loop_insn
);
505 copy_end
= PREV_INSN (insert_before
);
507 /* The instruction immediately before the JUMP_INSN may not be the
508 compare, so we must copy it. */
509 insert_before
= last_loop_insn
;
510 copy_end
= PREV_INSN (last_loop_insn
);
515 /* We currently can't unroll a loop if it doesn't end with a
516 JUMP_INSN. There would need to be a mechanism that recognizes
517 this case, and then inserts a jump after each loop body, which
518 jumps to after the last loop body. */
519 if (loop_dump_stream
)
520 fprintf (loop_dump_stream
,
521 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
527 /* Normal case: Must copy the compare and branch instructions at the
530 if (GET_CODE (last_loop_insn
) == BARRIER
)
532 /* Loop ends with an unconditional jump and a barrier.
533 Handle this like above, don't copy jump and barrier.
534 This is not strictly necessary, but doing so prevents generating
535 unconditional jumps to an immediately following label.
537 This will be corrected below if the target of this jump is
538 not the start_label. */
540 insert_before
= PREV_INSN (last_loop_insn
);
541 copy_end
= PREV_INSN (insert_before
);
543 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
545 /* Set insert_before to immediately after the JUMP_INSN, so that
546 NOTEs at the end of the loop will be correctly handled by
548 insert_before
= NEXT_INSN (last_loop_insn
);
549 copy_end
= last_loop_insn
;
553 /* We currently can't unroll a loop if it doesn't end with a
554 JUMP_INSN. There would need to be a mechanism that recognizes
555 this case, and then inserts a jump after each loop body, which
556 jumps to after the last loop body. */
557 if (loop_dump_stream
)
558 fprintf (loop_dump_stream
,
559 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
563 /* If copying exit test branches because they can not be eliminated,
564 then must convert the fall through case of the branch to a jump past
565 the end of the loop. Create a label to emit after the loop and save
566 it for later use. Do not use the label after the loop, if any, since
567 it might be used by insns outside the loop, or there might be insns
568 added before it later by final_[bg]iv_value which must be after
569 the real exit label. */
570 exit_label
= gen_label_rtx ();
573 while (GET_CODE (insn
) != CODE_LABEL
&& GET_CODE (insn
) != JUMP_INSN
)
574 insn
= NEXT_INSN (insn
);
576 if (GET_CODE (insn
) == JUMP_INSN
)
578 /* The loop starts with a jump down to the exit condition test.
579 Start copying the loop after the barrier following this
581 copy_start
= NEXT_INSN (insn
);
583 /* Splitting induction variables doesn't work when the loop is
584 entered via a jump to the bottom, because then we end up doing
585 a comparison against a new register for a split variable, but
586 we did not execute the set insn for the new register because
587 it was skipped over. */
588 splitting_not_safe
= 1;
589 if (loop_dump_stream
)
590 fprintf (loop_dump_stream
,
591 "Splitting not safe, because loop not entered at top.\n");
594 copy_start
= loop_start
;
597 /* This should always be the first label in the loop. */
598 start_label
= NEXT_INSN (copy_start
);
599 /* There may be a line number note and/or a loop continue note here. */
600 while (GET_CODE (start_label
) == NOTE
)
601 start_label
= NEXT_INSN (start_label
);
602 if (GET_CODE (start_label
) != CODE_LABEL
)
604 /* This can happen as a result of jump threading. If the first insns in
605 the loop test the same condition as the loop's backward jump, or the
606 opposite condition, then the backward jump will be modified to point
607 to elsewhere, and the loop's start label is deleted.
609 This case currently can not be handled by the loop unrolling code. */
611 if (loop_dump_stream
)
612 fprintf (loop_dump_stream
,
613 "Unrolling failure: unknown insns between BEG note and loop label.\n");
616 if (LABEL_NAME (start_label
))
618 /* The jump optimization pass must have combined the original start label
619 with a named label for a goto. We can't unroll this case because
620 jumps which go to the named label must be handled differently than
621 jumps to the loop start, and it is impossible to differentiate them
623 if (loop_dump_stream
)
624 fprintf (loop_dump_stream
,
625 "Unrolling failure: loop start label is gone\n");
629 if (unroll_type
== UNROLL_NAIVE
630 && GET_CODE (last_loop_insn
) == BARRIER
631 && start_label
!= JUMP_LABEL (PREV_INSN (last_loop_insn
)))
633 /* In this case, we must copy the jump and barrier, because they will
634 not be converted to jumps to an immediately following label. */
636 insert_before
= NEXT_INSN (last_loop_insn
);
637 copy_end
= last_loop_insn
;
640 if (unroll_type
== UNROLL_NAIVE
641 && GET_CODE (last_loop_insn
) == JUMP_INSN
642 && start_label
!= JUMP_LABEL (last_loop_insn
))
644 /* ??? The loop ends with a conditional branch that does not branch back
645 to the loop start label. In this case, we must emit an unconditional
646 branch to the loop exit after emitting the final branch.
647 copy_loop_body does not have support for this currently, so we
648 give up. It doesn't seem worthwhile to unroll anyways since
649 unrolling would increase the number of branch instructions
651 if (loop_dump_stream
)
652 fprintf (loop_dump_stream
,
653 "Unrolling failure: final conditional branch not to loop start\n");
657 /* Allocate a translation table for the labels and insn numbers.
658 They will be filled in as we copy the insns in the loop. */
660 max_labelno
= max_label_num ();
661 max_insnno
= get_max_uid ();
663 map
= (struct inline_remap
*) alloca (sizeof (struct inline_remap
));
665 map
->integrating
= 0;
667 /* Allocate the label map. */
671 map
->label_map
= (rtx
*) alloca (max_labelno
* sizeof (rtx
));
673 local_label
= (char *) alloca (max_labelno
);
674 bzero (local_label
, max_labelno
);
679 /* Search the loop and mark all local labels, i.e. the ones which have to
680 be distinct labels when copied. For all labels which might be
681 non-local, set their label_map entries to point to themselves.
682 If they happen to be local their label_map entries will be overwritten
683 before the loop body is copied. The label_map entries for local labels
684 will be set to a different value each time the loop body is copied. */
686 for (insn
= copy_start
; insn
!= loop_end
; insn
= NEXT_INSN (insn
))
688 if (GET_CODE (insn
) == CODE_LABEL
)
689 local_label
[CODE_LABEL_NUMBER (insn
)] = 1;
690 else if (GET_CODE (insn
) == JUMP_INSN
)
692 if (JUMP_LABEL (insn
))
693 map
->label_map
[CODE_LABEL_NUMBER (JUMP_LABEL (insn
))]
695 else if (GET_CODE (PATTERN (insn
)) == ADDR_VEC
696 || GET_CODE (PATTERN (insn
)) == ADDR_DIFF_VEC
)
698 rtx pat
= PATTERN (insn
);
699 int diff_vec_p
= GET_CODE (PATTERN (insn
)) == ADDR_DIFF_VEC
;
700 int len
= XVECLEN (pat
, diff_vec_p
);
703 for (i
= 0; i
< len
; i
++)
705 label
= XEXP (XVECEXP (pat
, diff_vec_p
, i
), 0);
706 map
->label_map
[CODE_LABEL_NUMBER (label
)] = label
;
712 /* Allocate space for the insn map. */
714 map
->insn_map
= (rtx
*) alloca (max_insnno
* sizeof (rtx
));
716 /* Set this to zero, to indicate that we are doing loop unrolling,
717 not function inlining. */
718 map
->inline_target
= 0;
720 /* The register and constant maps depend on the number of registers
721 present, so the final maps can't be created until after
722 find_splittable_regs is called. However, they are needed for
723 preconditioning, so we create temporary maps when preconditioning
726 /* The preconditioning code may allocate two new pseudo registers. */
727 maxregnum
= max_reg_num ();
729 /* Allocate and zero out the splittable_regs and addr_combined_regs
730 arrays. These must be zeroed here because they will be used if
731 loop preconditioning is performed, and must be zero for that case.
733 It is safe to do this here, since the extra registers created by the
734 preconditioning code and find_splittable_regs will never be used
735 to access the splittable_regs[] and addr_combined_regs[] arrays. */
737 splittable_regs
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
738 bzero ((char *) splittable_regs
, maxregnum
* sizeof (rtx
));
739 splittable_regs_updates
= (int *) alloca (maxregnum
* sizeof (int));
740 bzero ((char *) splittable_regs_updates
, maxregnum
* sizeof (int));
742 = (struct induction
**) alloca (maxregnum
* sizeof (struct induction
*));
743 bzero ((char *) addr_combined_regs
, maxregnum
* sizeof (struct induction
*));
744 /* We must limit it to max_reg_before_loop, because only these pseudo
745 registers have valid regno_first_uid info. Any register created after
746 that is unlikely to be local to the loop anyways. */
747 local_regno
= (char *) alloca (max_reg_before_loop
);
748 bzero (local_regno
, max_reg_before_loop
);
750 /* Mark all local registers, i.e. the ones which are referenced only
752 if (INSN_UID (copy_end
) < max_uid_for_loop
)
754 int copy_start_luid
= INSN_LUID (copy_start
);
755 int copy_end_luid
= INSN_LUID (copy_end
);
757 /* If a register is used in the jump insn, we must not duplicate it
758 since it will also be used outside the loop. */
759 if (GET_CODE (copy_end
) == JUMP_INSN
)
761 /* If copy_start points to the NOTE that starts the loop, then we must
762 use the next luid, because invariant pseudo-regs moved out of the loop
763 have their lifetimes modified to start here, but they are not safe
765 if (copy_start
== loop_start
)
768 /* If a pseudo's lifetime is entirely contained within this loop, then we
769 can use a different pseudo in each unrolled copy of the loop. This
770 results in better code. */
771 for (j
= FIRST_PSEUDO_REGISTER
; j
< max_reg_before_loop
; ++j
)
772 if (REGNO_FIRST_UID (j
) > 0 && REGNO_FIRST_UID (j
) <= max_uid_for_loop
773 && uid_luid
[REGNO_FIRST_UID (j
)] >= copy_start_luid
774 && REGNO_LAST_UID (j
) > 0 && REGNO_LAST_UID (j
) <= max_uid_for_loop
775 && uid_luid
[REGNO_LAST_UID (j
)] <= copy_end_luid
)
777 /* However, we must also check for loop-carried dependencies.
778 If the value the pseudo has at the end of iteration X is
779 used by iteration X+1, then we can not use a different pseudo
780 for each unrolled copy of the loop. */
781 /* A pseudo is safe if regno_first_uid is a set, and this
782 set dominates all instructions from regno_first_uid to
784 /* ??? This check is simplistic. We would get better code if
785 this check was more sophisticated. */
786 if (set_dominates_use (j
, REGNO_FIRST_UID (j
), REGNO_LAST_UID (j
),
787 copy_start
, copy_end
))
790 if (loop_dump_stream
)
793 fprintf (loop_dump_stream
, "Marked reg %d as local\n", j
);
795 fprintf (loop_dump_stream
, "Did not mark reg %d as local\n",
801 /* If this loop requires exit tests when unrolled, check to see if we
802 can precondition the loop so as to make the exit tests unnecessary.
803 Just like variable splitting, this is not safe if the loop is entered
804 via a jump to the bottom. Also, can not do this if no strength
805 reduce info, because precondition_loop_p uses this info. */
807 /* Must copy the loop body for preconditioning before the following
808 find_splittable_regs call since that will emit insns which need to
809 be after the preconditioned loop copies, but immediately before the
810 unrolled loop copies. */
812 /* Also, it is not safe to split induction variables for the preconditioned
813 copies of the loop body. If we split induction variables, then the code
814 assumes that each induction variable can be represented as a function
815 of its initial value and the loop iteration number. This is not true
816 in this case, because the last preconditioned copy of the loop body
817 could be any iteration from the first up to the `unroll_number-1'th,
818 depending on the initial value of the iteration variable. Therefore
819 we can not split induction variables here, because we can not calculate
820 their value. Hence, this code must occur before find_splittable_regs
823 if (unroll_type
== UNROLL_NAIVE
&& ! splitting_not_safe
&& strength_reduce_p
)
825 rtx initial_value
, final_value
, increment
;
827 if (precondition_loop_p (&initial_value
, &final_value
, &increment
,
828 loop_start
, loop_end
))
830 register rtx diff
, temp
;
831 enum machine_mode mode
;
833 int abs_inc
, neg_inc
;
835 map
->reg_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
837 map
->const_equiv_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
838 map
->const_age_map
= (unsigned *) alloca (maxregnum
839 * sizeof (unsigned));
840 map
->const_equiv_map_size
= maxregnum
;
841 global_const_equiv_map
= map
->const_equiv_map
;
842 global_const_equiv_map_size
= maxregnum
;
844 init_reg_map (map
, maxregnum
);
846 /* Limit loop unrolling to 4, since this will make 7 copies of
848 if (unroll_number
> 4)
851 /* Save the absolute value of the increment, and also whether or
852 not it is negative. */
854 abs_inc
= INTVAL (increment
);
863 /* Decide what mode to do these calculations in. Choose the larger
864 of final_value's mode and initial_value's mode, or a full-word if
865 both are constants. */
866 mode
= GET_MODE (final_value
);
867 if (mode
== VOIDmode
)
869 mode
= GET_MODE (initial_value
);
870 if (mode
== VOIDmode
)
873 else if (mode
!= GET_MODE (initial_value
)
874 && (GET_MODE_SIZE (mode
)
875 < GET_MODE_SIZE (GET_MODE (initial_value
))))
876 mode
= GET_MODE (initial_value
);
878 /* Calculate the difference between the final and initial values.
879 Final value may be a (plus (reg x) (const_int 1)) rtx.
880 Let the following cse pass simplify this if initial value is
883 We must copy the final and initial values here to avoid
884 improperly shared rtl. */
886 diff
= expand_binop (mode
, sub_optab
, copy_rtx (final_value
),
887 copy_rtx (initial_value
), NULL_RTX
, 0,
890 /* Now calculate (diff % (unroll * abs (increment))) by using an
892 diff
= expand_binop (GET_MODE (diff
), and_optab
, diff
,
893 GEN_INT (unroll_number
* abs_inc
- 1),
894 NULL_RTX
, 0, OPTAB_LIB_WIDEN
);
896 /* Now emit a sequence of branches to jump to the proper precond
899 labels
= (rtx
*) alloca (sizeof (rtx
) * unroll_number
);
900 for (i
= 0; i
< unroll_number
; i
++)
901 labels
[i
] = gen_label_rtx ();
903 /* Check for the case where the initial value is greater than or
904 equal to the final value. In that case, we want to execute
905 exactly one loop iteration. The code below will fail for this
906 case. This check does not apply if the loop has a NE
907 comparison at the end. */
909 if (loop_comparison_code
!= NE
)
911 emit_cmp_insn (initial_value
, final_value
, neg_inc
? LE
: GE
,
912 NULL_RTX
, mode
, 0, 0);
914 emit_jump_insn (gen_ble (labels
[1]));
916 emit_jump_insn (gen_bge (labels
[1]));
917 JUMP_LABEL (get_last_insn ()) = labels
[1];
918 LABEL_NUSES (labels
[1])++;
921 /* Assuming the unroll_number is 4, and the increment is 2, then
922 for a negative increment: for a positive increment:
923 diff = 0,1 precond 0 diff = 0,7 precond 0
924 diff = 2,3 precond 3 diff = 1,2 precond 1
925 diff = 4,5 precond 2 diff = 3,4 precond 2
926 diff = 6,7 precond 1 diff = 5,6 precond 3 */
928 /* We only need to emit (unroll_number - 1) branches here, the
929 last case just falls through to the following code. */
931 /* ??? This would give better code if we emitted a tree of branches
932 instead of the current linear list of branches. */
934 for (i
= 0; i
< unroll_number
- 1; i
++)
937 enum rtx_code cmp_code
;
939 /* For negative increments, must invert the constant compared
940 against, except when comparing against zero. */
948 cmp_const
= unroll_number
- i
;
957 emit_cmp_insn (diff
, GEN_INT (abs_inc
* cmp_const
),
958 cmp_code
, NULL_RTX
, mode
, 0, 0);
961 emit_jump_insn (gen_beq (labels
[i
]));
963 emit_jump_insn (gen_bge (labels
[i
]));
965 emit_jump_insn (gen_ble (labels
[i
]));
966 JUMP_LABEL (get_last_insn ()) = labels
[i
];
967 LABEL_NUSES (labels
[i
])++;
970 /* If the increment is greater than one, then we need another branch,
971 to handle other cases equivalent to 0. */
973 /* ??? This should be merged into the code above somehow to help
974 simplify the code here, and reduce the number of branches emitted.
975 For the negative increment case, the branch here could easily
976 be merged with the `0' case branch above. For the positive
977 increment case, it is not clear how this can be simplified. */
982 enum rtx_code cmp_code
;
986 cmp_const
= abs_inc
- 1;
991 cmp_const
= abs_inc
* (unroll_number
- 1) + 1;
995 emit_cmp_insn (diff
, GEN_INT (cmp_const
), cmp_code
, NULL_RTX
,
999 emit_jump_insn (gen_ble (labels
[0]));
1001 emit_jump_insn (gen_bge (labels
[0]));
1002 JUMP_LABEL (get_last_insn ()) = labels
[0];
1003 LABEL_NUSES (labels
[0])++;
1006 sequence
= gen_sequence ();
1008 emit_insn_before (sequence
, loop_start
);
1010 /* Only the last copy of the loop body here needs the exit
1011 test, so set copy_end to exclude the compare/branch here,
1012 and then reset it inside the loop when get to the last
1015 if (GET_CODE (last_loop_insn
) == BARRIER
)
1016 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
1017 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
1020 /* The immediately preceding insn is a compare which we do not
1022 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
1024 /* The immediately preceding insn may not be a compare, so we
1026 copy_end
= PREV_INSN (last_loop_insn
);
1032 for (i
= 1; i
< unroll_number
; i
++)
1034 emit_label_after (labels
[unroll_number
- i
],
1035 PREV_INSN (loop_start
));
1037 bzero ((char *) map
->insn_map
, max_insnno
* sizeof (rtx
));
1038 bzero ((char *) map
->const_equiv_map
, maxregnum
* sizeof (rtx
));
1039 bzero ((char *) map
->const_age_map
,
1040 maxregnum
* sizeof (unsigned));
1043 for (j
= 0; j
< max_labelno
; j
++)
1045 map
->label_map
[j
] = gen_label_rtx ();
1047 for (j
= FIRST_PSEUDO_REGISTER
; j
< max_reg_before_loop
; j
++)
1050 map
->reg_map
[j
] = gen_reg_rtx (GET_MODE (regno_reg_rtx
[j
]));
1051 record_base_value (REGNO (map
->reg_map
[j
]),
1054 /* The last copy needs the compare/branch insns at the end,
1055 so reset copy_end here if the loop ends with a conditional
1058 if (i
== unroll_number
- 1)
1060 if (GET_CODE (last_loop_insn
) == BARRIER
)
1061 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
1063 copy_end
= last_loop_insn
;
1066 /* None of the copies are the `last_iteration', so just
1067 pass zero for that parameter. */
1068 copy_loop_body (copy_start
, copy_end
, map
, exit_label
, 0,
1069 unroll_type
, start_label
, loop_end
,
1070 loop_start
, copy_end
);
1072 emit_label_after (labels
[0], PREV_INSN (loop_start
));
1074 if (GET_CODE (last_loop_insn
) == BARRIER
)
1076 insert_before
= PREV_INSN (last_loop_insn
);
1077 copy_end
= PREV_INSN (insert_before
);
1082 /* The immediately preceding insn is a compare which we do not
1084 insert_before
= PREV_INSN (last_loop_insn
);
1085 copy_end
= PREV_INSN (insert_before
);
1087 /* The immediately preceding insn may not be a compare, so we
1089 insert_before
= last_loop_insn
;
1090 copy_end
= PREV_INSN (last_loop_insn
);
1094 /* Set unroll type to MODULO now. */
1095 unroll_type
= UNROLL_MODULO
;
1096 loop_preconditioned
= 1;
1098 if (loop_n_iterations
> 0)
1099 loop_unroll_iter
[ loop_number(loop_start
, loop_end
) ]
1100 = (loop_n_iterations
1101 - loop_n_iterations
% (abs_inc
* unroll_number
));
1103 /* inform loop.c about the new initial value */
1104 loop_start_value
[loop_number(loop_start
, loop_end
)] = initial_value
;
1110 /* If reach here, and the loop type is UNROLL_NAIVE, then don't unroll
1111 the loop unless all loops are being unrolled. */
1112 if (unroll_type
== UNROLL_NAIVE
&& ! flag_unroll_all_loops
)
1114 if (loop_dump_stream
)
1115 fprintf (loop_dump_stream
, "Unrolling failure: Naive unrolling not being done.\n");
1119 /* At this point, we are guaranteed to unroll the loop. */
1122 /* inform loop.c about the factor of unrolling */
1123 if (unroll_type
== UNROLL_COMPLETELY
)
1124 loop_unroll_factor
[ loop_number(loop_start
, loop_end
) ] = -1;
1126 loop_unroll_factor
[ loop_number(loop_start
, loop_end
) ] = unroll_number
;
1130 /* For each biv and giv, determine whether it can be safely split into
1131 a different variable for each unrolled copy of the loop body.
1132 We precalculate and save this info here, since computing it is
1135 Do this before deleting any instructions from the loop, so that
1136 back_branch_in_range_p will work correctly. */
1138 if (splitting_not_safe
)
1141 temp
= find_splittable_regs (unroll_type
, loop_start
, loop_end
,
1142 end_insert_before
, unroll_number
);
1144 /* find_splittable_regs may have created some new registers, so must
1145 reallocate the reg_map with the new larger size, and must realloc
1146 the constant maps also. */
1148 maxregnum
= max_reg_num ();
1149 map
->reg_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
1151 init_reg_map (map
, maxregnum
);
1153 /* Space is needed in some of the map for new registers, so new_maxregnum
1154 is an (over)estimate of how many registers will exist at the end. */
1155 new_maxregnum
= maxregnum
+ (temp
* unroll_number
* 2);
1157 /* Must realloc space for the constant maps, because the number of registers
1158 may have changed. */
1160 map
->const_equiv_map
= (rtx
*) alloca (new_maxregnum
* sizeof (rtx
));
1161 map
->const_age_map
= (unsigned *) alloca (new_maxregnum
* sizeof (unsigned));
1163 map
->const_equiv_map_size
= new_maxregnum
;
1164 global_const_equiv_map
= map
->const_equiv_map
;
1165 global_const_equiv_map_size
= new_maxregnum
;
1167 /* Search the list of bivs and givs to find ones which need to be remapped
1168 when split, and set their reg_map entry appropriately. */
1170 for (bl
= loop_iv_list
; bl
; bl
= bl
->next
)
1172 if (REGNO (bl
->biv
->src_reg
) != bl
->regno
)
1173 map
->reg_map
[bl
->regno
] = bl
->biv
->src_reg
;
1175 /* Currently, non-reduced/final-value givs are never split. */
1176 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
1177 if (REGNO (v
->src_reg
) != bl
->regno
)
1178 map
->reg_map
[REGNO (v
->dest_reg
)] = v
->src_reg
;
1182 /* Use our current register alignment and pointer flags. */
1183 map
->regno_pointer_flag
= regno_pointer_flag
;
1184 map
->regno_pointer_align
= regno_pointer_align
;
1186 /* If the loop is being partially unrolled, and the iteration variables
1187 are being split, and are being renamed for the split, then must fix up
1188 the compare/jump instruction at the end of the loop to refer to the new
1189 registers. This compare isn't copied, so the registers used in it
1190 will never be replaced if it isn't done here. */
1192 if (unroll_type
== UNROLL_MODULO
)
1194 insn
= NEXT_INSN (copy_end
);
1195 if (GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == JUMP_INSN
)
1196 PATTERN (insn
) = remap_split_bivs (PATTERN (insn
));
1199 /* For unroll_number - 1 times, make a copy of each instruction
1200 between copy_start and copy_end, and insert these new instructions
1201 before the end of the loop. */
1203 for (i
= 0; i
< unroll_number
; i
++)
1205 bzero ((char *) map
->insn_map
, max_insnno
* sizeof (rtx
));
1206 bzero ((char *) map
->const_equiv_map
, new_maxregnum
* sizeof (rtx
));
1207 bzero ((char *) map
->const_age_map
, new_maxregnum
* sizeof (unsigned));
1210 for (j
= 0; j
< max_labelno
; j
++)
1212 map
->label_map
[j
] = gen_label_rtx ();
1214 for (j
= FIRST_PSEUDO_REGISTER
; j
< max_reg_before_loop
; j
++)
1217 map
->reg_map
[j
] = gen_reg_rtx (GET_MODE (regno_reg_rtx
[j
]));
1218 record_base_value (REGNO (map
->reg_map
[j
]),
1222 /* If loop starts with a branch to the test, then fix it so that
1223 it points to the test of the first unrolled copy of the loop. */
1224 if (i
== 0 && loop_start
!= copy_start
)
1226 insn
= PREV_INSN (copy_start
);
1227 pattern
= PATTERN (insn
);
1229 tem
= map
->label_map
[CODE_LABEL_NUMBER
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);
1497 else if (GET_CODE (increment
) == IOR
1498 || GET_CODE (increment
) == ASHIFT
)
1500 /* The rs6000 port loads some constants with IOR.
1501 The alpha port loads some constants with ASHIFT. */
1502 rtx second_part
= XEXP (increment
, 1);
1503 enum rtx_code code
= GET_CODE (increment
);
1505 src_insn
= PREV_INSN (src_insn
);
1506 increment
= SET_SRC (PATTERN (src_insn
));
1507 /* Don't need the last insn anymore. */
1508 delete_insn (get_last_insn ());
1510 if (GET_CODE (second_part
) != CONST_INT
1511 || GET_CODE (increment
) != CONST_INT
)
1515 increment
= GEN_INT (INTVAL (increment
) | INTVAL (second_part
));
1517 increment
= GEN_INT (INTVAL (increment
) << INTVAL (second_part
));
1520 if (GET_CODE (increment
) != CONST_INT
)
1523 /* The insn loading the constant into a register is no longer needed,
1525 delete_insn (get_last_insn ());
1528 if (increment_total
)
1529 increment_total
= GEN_INT (INTVAL (increment_total
) + INTVAL (increment
));
1531 increment_total
= increment
;
1533 /* Check that the source register is the same as the register we expected
1534 to see as the source. If not, something is seriously wrong. */
1535 if (GET_CODE (XEXP (SET_SRC (pattern
), 0)) != REG
1536 || REGNO (XEXP (SET_SRC (pattern
), 0)) != regno
)
1538 /* Some machines (e.g. the romp), may emit two add instructions for
1539 certain constants, so lets try looking for another add immediately
1540 before this one if we have only seen one add insn so far. */
1546 src_insn
= PREV_INSN (src_insn
);
1547 pattern
= PATTERN (src_insn
);
1549 delete_insn (get_last_insn ());
1557 return increment_total
;
1560 /* Copy REG_NOTES, except for insn references, because not all insn_map
1561 entries are valid yet. We do need to copy registers now though, because
1562 the reg_map entries can change during copying. */
1565 initial_reg_note_copy (notes
, map
)
1567 struct inline_remap
*map
;
1574 copy
= rtx_alloc (GET_CODE (notes
));
1575 PUT_MODE (copy
, GET_MODE (notes
));
1577 if (GET_CODE (notes
) == EXPR_LIST
)
1578 XEXP (copy
, 0) = copy_rtx_and_substitute (XEXP (notes
, 0), map
);
1579 else if (GET_CODE (notes
) == INSN_LIST
)
1580 /* Don't substitute for these yet. */
1581 XEXP (copy
, 0) = XEXP (notes
, 0);
1585 XEXP (copy
, 1) = initial_reg_note_copy (XEXP (notes
, 1), map
);
1590 /* Fixup insn references in copied REG_NOTES. */
1593 final_reg_note_copy (notes
, map
)
1595 struct inline_remap
*map
;
1599 for (note
= notes
; note
; note
= XEXP (note
, 1))
1600 if (GET_CODE (note
) == INSN_LIST
)
1601 XEXP (note
, 0) = map
->insn_map
[INSN_UID (XEXP (note
, 0))];
1604 /* Copy each instruction in the loop, substituting from map as appropriate.
1605 This is very similar to a loop in expand_inline_function. */
1608 copy_loop_body (copy_start
, copy_end
, map
, exit_label
, last_iteration
,
1609 unroll_type
, start_label
, loop_end
, insert_before
,
1611 rtx copy_start
, copy_end
;
1612 struct inline_remap
*map
;
1615 enum unroll_types unroll_type
;
1616 rtx start_label
, loop_end
, insert_before
, copy_notes_from
;
1620 int dest_reg_was_split
, i
;
1622 rtx final_label
= 0;
1623 rtx giv_inc
, giv_dest_reg
, giv_src_reg
;
1625 /* If this isn't the last iteration, then map any references to the
1626 start_label to final_label. Final label will then be emitted immediately
1627 after the end of this loop body if it was ever used.
1629 If this is the last iteration, then map references to the start_label
1631 if (! last_iteration
)
1633 final_label
= gen_label_rtx ();
1634 map
->label_map
[CODE_LABEL_NUMBER (start_label
)] = final_label
;
1637 map
->label_map
[CODE_LABEL_NUMBER (start_label
)] = start_label
;
1644 insn
= NEXT_INSN (insn
);
1646 map
->orig_asm_operands_vector
= 0;
1648 switch (GET_CODE (insn
))
1651 pattern
= PATTERN (insn
);
1655 /* Check to see if this is a giv that has been combined with
1656 some split address givs. (Combined in the sense that
1657 `combine_givs' in loop.c has put two givs in the same register.)
1658 In this case, we must search all givs based on the same biv to
1659 find the address givs. Then split the address givs.
1660 Do this before splitting the giv, since that may map the
1661 SET_DEST to a new register. */
1663 if (GET_CODE (pattern
) == SET
1664 && GET_CODE (SET_DEST (pattern
)) == REG
1665 && addr_combined_regs
[REGNO (SET_DEST (pattern
))])
1667 struct iv_class
*bl
;
1668 struct induction
*v
, *tv
;
1669 int regno
= REGNO (SET_DEST (pattern
));
1671 v
= addr_combined_regs
[REGNO (SET_DEST (pattern
))];
1672 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
1674 /* Although the giv_inc amount is not needed here, we must call
1675 calculate_giv_inc here since it might try to delete the
1676 last insn emitted. If we wait until later to call it,
1677 we might accidentally delete insns generated immediately
1678 below by emit_unrolled_add. */
1680 giv_inc
= calculate_giv_inc (pattern
, insn
, regno
);
1682 /* Now find all address giv's that were combined with this
1684 for (tv
= bl
->giv
; tv
; tv
= tv
->next_iv
)
1685 if (tv
->giv_type
== DEST_ADDR
&& tv
->same
== v
)
1689 /* If this DEST_ADDR giv was not split, then ignore it. */
1690 if (*tv
->location
!= tv
->dest_reg
)
1693 /* Scale this_giv_inc if the multiplicative factors of
1694 the two givs are different. */
1695 this_giv_inc
= INTVAL (giv_inc
);
1696 if (tv
->mult_val
!= v
->mult_val
)
1697 this_giv_inc
= (this_giv_inc
/ INTVAL (v
->mult_val
)
1698 * INTVAL (tv
->mult_val
));
1700 tv
->dest_reg
= plus_constant (tv
->dest_reg
, this_giv_inc
);
1701 *tv
->location
= tv
->dest_reg
;
1703 if (last_iteration
&& unroll_type
!= UNROLL_COMPLETELY
)
1705 /* Must emit an insn to increment the split address
1706 giv. Add in the const_adjust field in case there
1707 was a constant eliminated from the address. */
1708 rtx value
, dest_reg
;
1710 /* tv->dest_reg will be either a bare register,
1711 or else a register plus a constant. */
1712 if (GET_CODE (tv
->dest_reg
) == REG
)
1713 dest_reg
= tv
->dest_reg
;
1715 dest_reg
= XEXP (tv
->dest_reg
, 0);
1717 /* Check for shared address givs, and avoid
1718 incrementing the shared pseudo reg more than
1720 if (! tv
->same_insn
&& ! tv
->shared
)
1722 /* tv->dest_reg may actually be a (PLUS (REG)
1723 (CONST)) here, so we must call plus_constant
1724 to add the const_adjust amount before calling
1725 emit_unrolled_add below. */
1726 value
= plus_constant (tv
->dest_reg
,
1729 /* The constant could be too large for an add
1730 immediate, so can't directly emit an insn
1732 emit_unrolled_add (dest_reg
, XEXP (value
, 0),
1736 /* Reset the giv to be just the register again, in case
1737 it is used after the set we have just emitted.
1738 We must subtract the const_adjust factor added in
1740 tv
->dest_reg
= plus_constant (dest_reg
,
1741 - tv
->const_adjust
);
1742 *tv
->location
= tv
->dest_reg
;
1747 /* If this is a setting of a splittable variable, then determine
1748 how to split the variable, create a new set based on this split,
1749 and set up the reg_map so that later uses of the variable will
1750 use the new split variable. */
1752 dest_reg_was_split
= 0;
1754 if (GET_CODE (pattern
) == SET
1755 && GET_CODE (SET_DEST (pattern
)) == REG
1756 && splittable_regs
[REGNO (SET_DEST (pattern
))])
1758 int regno
= REGNO (SET_DEST (pattern
));
1760 dest_reg_was_split
= 1;
1762 /* Compute the increment value for the giv, if it wasn't
1763 already computed above. */
1766 giv_inc
= calculate_giv_inc (pattern
, insn
, regno
);
1767 giv_dest_reg
= SET_DEST (pattern
);
1768 giv_src_reg
= SET_DEST (pattern
);
1770 if (unroll_type
== UNROLL_COMPLETELY
)
1772 /* Completely unrolling the loop. Set the induction
1773 variable to a known constant value. */
1775 /* The value in splittable_regs may be an invariant
1776 value, so we must use plus_constant here. */
1777 splittable_regs
[regno
]
1778 = plus_constant (splittable_regs
[regno
], INTVAL (giv_inc
));
1780 if (GET_CODE (splittable_regs
[regno
]) == PLUS
)
1782 giv_src_reg
= XEXP (splittable_regs
[regno
], 0);
1783 giv_inc
= XEXP (splittable_regs
[regno
], 1);
1787 /* The splittable_regs value must be a REG or a
1788 CONST_INT, so put the entire value in the giv_src_reg
1790 giv_src_reg
= splittable_regs
[regno
];
1791 giv_inc
= const0_rtx
;
1796 /* Partially unrolling loop. Create a new pseudo
1797 register for the iteration variable, and set it to
1798 be a constant plus the original register. Except
1799 on the last iteration, when the result has to
1800 go back into the original iteration var register. */
1802 /* Handle bivs which must be mapped to a new register
1803 when split. This happens for bivs which need their
1804 final value set before loop entry. The new register
1805 for the biv was stored in the biv's first struct
1806 induction entry by find_splittable_regs. */
1808 if (regno
< max_reg_before_loop
1809 && reg_iv_type
[regno
] == BASIC_INDUCT
)
1811 giv_src_reg
= reg_biv_class
[regno
]->biv
->src_reg
;
1812 giv_dest_reg
= giv_src_reg
;
1816 /* If non-reduced/final-value givs were split, then
1817 this would have to remap those givs also. See
1818 find_splittable_regs. */
1821 splittable_regs
[regno
]
1822 = GEN_INT (INTVAL (giv_inc
)
1823 + INTVAL (splittable_regs
[regno
]));
1824 giv_inc
= splittable_regs
[regno
];
1826 /* Now split the induction variable by changing the dest
1827 of this insn to a new register, and setting its
1828 reg_map entry to point to this new register.
1830 If this is the last iteration, and this is the last insn
1831 that will update the iv, then reuse the original dest,
1832 to ensure that the iv will have the proper value when
1833 the loop exits or repeats.
1835 Using splittable_regs_updates here like this is safe,
1836 because it can only be greater than one if all
1837 instructions modifying the iv are always executed in
1840 if (! last_iteration
1841 || (splittable_regs_updates
[regno
]-- != 1))
1843 tem
= gen_reg_rtx (GET_MODE (giv_src_reg
));
1845 map
->reg_map
[regno
] = tem
;
1846 record_base_value (REGNO (tem
), giv_src_reg
);
1849 map
->reg_map
[regno
] = giv_src_reg
;
1852 /* The constant being added could be too large for an add
1853 immediate, so can't directly emit an insn here. */
1854 emit_unrolled_add (giv_dest_reg
, giv_src_reg
, giv_inc
);
1855 copy
= get_last_insn ();
1856 pattern
= PATTERN (copy
);
1860 pattern
= copy_rtx_and_substitute (pattern
, map
);
1861 copy
= emit_insn (pattern
);
1863 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
1866 /* If this insn is setting CC0, it may need to look at
1867 the insn that uses CC0 to see what type of insn it is.
1868 In that case, the call to recog via validate_change will
1869 fail. So don't substitute constants here. Instead,
1870 do it when we emit the following insn.
1872 For example, see the pyr.md file. That machine has signed and
1873 unsigned compares. The compare patterns must check the
1874 following branch insn to see which what kind of compare to
1877 If the previous insn set CC0, substitute constants on it as
1879 if (sets_cc0_p (PATTERN (copy
)) != 0)
1884 try_constants (cc0_insn
, map
);
1886 try_constants (copy
, map
);
1889 try_constants (copy
, map
);
1892 /* Make split induction variable constants `permanent' since we
1893 know there are no backward branches across iteration variable
1894 settings which would invalidate this. */
1895 if (dest_reg_was_split
)
1897 int regno
= REGNO (SET_DEST (pattern
));
1899 if (regno
< map
->const_equiv_map_size
1900 && map
->const_age_map
[regno
] == map
->const_age
)
1901 map
->const_age_map
[regno
] = -1;
1906 pattern
= copy_rtx_and_substitute (PATTERN (insn
), map
);
1907 copy
= emit_jump_insn (pattern
);
1908 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
1910 if (JUMP_LABEL (insn
) == start_label
&& insn
== copy_end
1911 && ! last_iteration
)
1913 /* This is a branch to the beginning of the loop; this is the
1914 last insn being copied; and this is not the last iteration.
1915 In this case, we want to change the original fall through
1916 case to be a branch past the end of the loop, and the
1917 original jump label case to fall_through. */
1919 if (invert_exp (pattern
, copy
))
1921 if (! redirect_exp (&pattern
,
1922 map
->label_map
[CODE_LABEL_NUMBER
1923 (JUMP_LABEL (insn
))],
1930 rtx lab
= gen_label_rtx ();
1931 /* Can't do it by reversing the jump (probably because we
1932 couldn't reverse the conditions), so emit a new
1933 jump_insn after COPY, and redirect the jump around
1935 jmp
= emit_jump_insn_after (gen_jump (exit_label
), copy
);
1936 jmp
= emit_barrier_after (jmp
);
1937 emit_label_after (lab
, jmp
);
1938 LABEL_NUSES (lab
) = 0;
1939 if (! redirect_exp (&pattern
,
1940 map
->label_map
[CODE_LABEL_NUMBER
1941 (JUMP_LABEL (insn
))],
1949 try_constants (cc0_insn
, map
);
1952 try_constants (copy
, map
);
1954 /* Set the jump label of COPY correctly to avoid problems with
1955 later passes of unroll_loop, if INSN had jump label set. */
1956 if (JUMP_LABEL (insn
))
1960 /* Can't use the label_map for every insn, since this may be
1961 the backward branch, and hence the label was not mapped. */
1962 if (GET_CODE (pattern
) == SET
)
1964 tem
= SET_SRC (pattern
);
1965 if (GET_CODE (tem
) == LABEL_REF
)
1966 label
= XEXP (tem
, 0);
1967 else if (GET_CODE (tem
) == IF_THEN_ELSE
)
1969 if (XEXP (tem
, 1) != pc_rtx
)
1970 label
= XEXP (XEXP (tem
, 1), 0);
1972 label
= XEXP (XEXP (tem
, 2), 0);
1976 if (label
&& GET_CODE (label
) == CODE_LABEL
)
1977 JUMP_LABEL (copy
) = label
;
1980 /* An unrecognizable jump insn, probably the entry jump
1981 for a switch statement. This label must have been mapped,
1982 so just use the label_map to get the new jump label. */
1984 = map
->label_map
[CODE_LABEL_NUMBER (JUMP_LABEL (insn
))];
1987 /* If this is a non-local jump, then must increase the label
1988 use count so that the label will not be deleted when the
1989 original jump is deleted. */
1990 LABEL_NUSES (JUMP_LABEL (copy
))++;
1992 else if (GET_CODE (PATTERN (copy
)) == ADDR_VEC
1993 || GET_CODE (PATTERN (copy
)) == ADDR_DIFF_VEC
)
1995 rtx pat
= PATTERN (copy
);
1996 int diff_vec_p
= GET_CODE (pat
) == ADDR_DIFF_VEC
;
1997 int len
= XVECLEN (pat
, diff_vec_p
);
2000 for (i
= 0; i
< len
; i
++)
2001 LABEL_NUSES (XEXP (XVECEXP (pat
, diff_vec_p
, i
), 0))++;
2004 /* If this used to be a conditional jump insn but whose branch
2005 direction is now known, we must do something special. */
2006 if (condjump_p (insn
) && !simplejump_p (insn
) && map
->last_pc_value
)
2009 /* The previous insn set cc0 for us. So delete it. */
2010 delete_insn (PREV_INSN (copy
));
2013 /* If this is now a no-op, delete it. */
2014 if (map
->last_pc_value
== pc_rtx
)
2016 /* Don't let delete_insn delete the label referenced here,
2017 because we might possibly need it later for some other
2018 instruction in the loop. */
2019 if (JUMP_LABEL (copy
))
2020 LABEL_NUSES (JUMP_LABEL (copy
))++;
2022 if (JUMP_LABEL (copy
))
2023 LABEL_NUSES (JUMP_LABEL (copy
))--;
2027 /* Otherwise, this is unconditional jump so we must put a
2028 BARRIER after it. We could do some dead code elimination
2029 here, but jump.c will do it just as well. */
2035 pattern
= copy_rtx_and_substitute (PATTERN (insn
), map
);
2036 copy
= emit_call_insn (pattern
);
2037 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
2039 /* Because the USAGE information potentially contains objects other
2040 than hard registers, we need to copy it. */
2041 CALL_INSN_FUNCTION_USAGE (copy
)
2042 = copy_rtx_and_substitute (CALL_INSN_FUNCTION_USAGE (insn
), map
);
2046 try_constants (cc0_insn
, map
);
2049 try_constants (copy
, map
);
2051 /* Be lazy and assume CALL_INSNs clobber all hard registers. */
2052 for (i
= 0; i
< FIRST_PSEUDO_REGISTER
; i
++)
2053 map
->const_equiv_map
[i
] = 0;
2057 /* If this is the loop start label, then we don't need to emit a
2058 copy of this label since no one will use it. */
2060 if (insn
!= start_label
)
2062 copy
= emit_label (map
->label_map
[CODE_LABEL_NUMBER (insn
)]);
2068 copy
= emit_barrier ();
2072 /* VTOP notes are valid only before the loop exit test. If placed
2073 anywhere else, loop may generate bad code. */
2075 if (NOTE_LINE_NUMBER (insn
) != NOTE_INSN_DELETED
2076 && (NOTE_LINE_NUMBER (insn
) != NOTE_INSN_LOOP_VTOP
2077 || (last_iteration
&& unroll_type
!= UNROLL_COMPLETELY
)))
2078 copy
= emit_note (NOTE_SOURCE_FILE (insn
),
2079 NOTE_LINE_NUMBER (insn
));
2089 map
->insn_map
[INSN_UID (insn
)] = copy
;
2091 while (insn
!= copy_end
);
2093 /* Now finish coping the REG_NOTES. */
2097 insn
= NEXT_INSN (insn
);
2098 if ((GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == JUMP_INSN
2099 || GET_CODE (insn
) == CALL_INSN
)
2100 && map
->insn_map
[INSN_UID (insn
)])
2101 final_reg_note_copy (REG_NOTES (map
->insn_map
[INSN_UID (insn
)]), map
);
2103 while (insn
!= copy_end
);
2105 /* There may be notes between copy_notes_from and loop_end. Emit a copy of
2106 each of these notes here, since there may be some important ones, such as
2107 NOTE_INSN_BLOCK_END notes, in this group. We don't do this on the last
2108 iteration, because the original notes won't be deleted.
2110 We can't use insert_before here, because when from preconditioning,
2111 insert_before points before the loop. We can't use copy_end, because
2112 there may be insns already inserted after it (which we don't want to
2113 copy) when not from preconditioning code. */
2115 if (! last_iteration
)
2117 for (insn
= copy_notes_from
; insn
!= loop_end
; insn
= NEXT_INSN (insn
))
2119 if (GET_CODE (insn
) == NOTE
2120 && NOTE_LINE_NUMBER (insn
) != NOTE_INSN_DELETED
)
2121 emit_note (NOTE_SOURCE_FILE (insn
), NOTE_LINE_NUMBER (insn
));
2125 if (final_label
&& LABEL_NUSES (final_label
) > 0)
2126 emit_label (final_label
);
2128 tem
= gen_sequence ();
2130 emit_insn_before (tem
, insert_before
);
2133 /* Emit an insn, using the expand_binop to ensure that a valid insn is
2134 emitted. This will correctly handle the case where the increment value
2135 won't fit in the immediate field of a PLUS insns. */
2138 emit_unrolled_add (dest_reg
, src_reg
, increment
)
2139 rtx dest_reg
, src_reg
, increment
;
2143 result
= expand_binop (GET_MODE (dest_reg
), add_optab
, src_reg
, increment
,
2144 dest_reg
, 0, OPTAB_LIB_WIDEN
);
2146 if (dest_reg
!= result
)
2147 emit_move_insn (dest_reg
, result
);
2150 /* Searches the insns between INSN and LOOP_END. Returns 1 if there
2151 is a backward branch in that range that branches to somewhere between
2152 LOOP_START and INSN. Returns 0 otherwise. */
2154 /* ??? This is quadratic algorithm. Could be rewritten to be linear.
2155 In practice, this is not a problem, because this function is seldom called,
2156 and uses a negligible amount of CPU time on average. */
2159 back_branch_in_range_p (insn
, loop_start
, loop_end
)
2161 rtx loop_start
, loop_end
;
2163 rtx p
, q
, target_insn
;
2165 /* Stop before we get to the backward branch at the end of the loop. */
2166 loop_end
= prev_nonnote_insn (loop_end
);
2167 if (GET_CODE (loop_end
) == BARRIER
)
2168 loop_end
= PREV_INSN (loop_end
);
2170 /* Check in case insn has been deleted, search forward for first non
2171 deleted insn following it. */
2172 while (INSN_DELETED_P (insn
))
2173 insn
= NEXT_INSN (insn
);
2175 /* Check for the case where insn is the last insn in the loop. */
2176 if (insn
== loop_end
)
2179 for (p
= NEXT_INSN (insn
); p
!= loop_end
; p
= NEXT_INSN (p
))
2181 if (GET_CODE (p
) == JUMP_INSN
)
2183 target_insn
= JUMP_LABEL (p
);
2185 /* Search from loop_start to insn, to see if one of them is
2186 the target_insn. We can't use INSN_LUID comparisons here,
2187 since insn may not have an LUID entry. */
2188 for (q
= loop_start
; q
!= insn
; q
= NEXT_INSN (q
))
2189 if (q
== target_insn
)
2197 /* Try to generate the simplest rtx for the expression
2198 (PLUS (MULT mult1 mult2) add1). This is used to calculate the initial
2202 fold_rtx_mult_add (mult1
, mult2
, add1
, mode
)
2203 rtx mult1
, mult2
, add1
;
2204 enum machine_mode mode
;
2209 /* The modes must all be the same. This should always be true. For now,
2210 check to make sure. */
2211 if ((GET_MODE (mult1
) != mode
&& GET_MODE (mult1
) != VOIDmode
)
2212 || (GET_MODE (mult2
) != mode
&& GET_MODE (mult2
) != VOIDmode
)
2213 || (GET_MODE (add1
) != mode
&& GET_MODE (add1
) != VOIDmode
))
2216 /* Ensure that if at least one of mult1/mult2 are constant, then mult2
2217 will be a constant. */
2218 if (GET_CODE (mult1
) == CONST_INT
)
2225 mult_res
= simplify_binary_operation (MULT
, mode
, mult1
, mult2
);
2227 mult_res
= gen_rtx (MULT
, mode
, mult1
, mult2
);
2229 /* Again, put the constant second. */
2230 if (GET_CODE (add1
) == CONST_INT
)
2237 result
= simplify_binary_operation (PLUS
, mode
, add1
, mult_res
);
2239 result
= gen_rtx (PLUS
, mode
, add1
, mult_res
);
2244 /* Searches the list of induction struct's for the biv BL, to try to calculate
2245 the total increment value for one iteration of the loop as a constant.
2247 Returns the increment value as an rtx, simplified as much as possible,
2248 if it can be calculated. Otherwise, returns 0. */
2251 biv_total_increment (bl
, loop_start
, loop_end
)
2252 struct iv_class
*bl
;
2253 rtx loop_start
, loop_end
;
2255 struct induction
*v
;
2258 /* For increment, must check every instruction that sets it. Each
2259 instruction must be executed only once each time through the loop.
2260 To verify this, we check that the the insn is always executed, and that
2261 there are no backward branches after the insn that branch to before it.
2262 Also, the insn must have a mult_val of one (to make sure it really is
2265 result
= const0_rtx
;
2266 for (v
= bl
->biv
; v
; v
= v
->next_iv
)
2268 if (v
->always_computable
&& v
->mult_val
== const1_rtx
2269 && ! back_branch_in_range_p (v
->insn
, loop_start
, loop_end
))
2270 result
= fold_rtx_mult_add (result
, const1_rtx
, v
->add_val
, v
->mode
);
2278 /* Determine the initial value of the iteration variable, and the amount
2279 that it is incremented each loop. Use the tables constructed by
2280 the strength reduction pass to calculate these values.
2282 Initial_value and/or increment are set to zero if their values could not
2286 iteration_info (iteration_var
, initial_value
, increment
, loop_start
, loop_end
)
2287 rtx iteration_var
, *initial_value
, *increment
;
2288 rtx loop_start
, loop_end
;
2290 struct iv_class
*bl
;
2291 struct induction
*v
, *b
;
2293 /* Clear the result values, in case no answer can be found. */
2297 /* The iteration variable can be either a giv or a biv. Check to see
2298 which it is, and compute the variable's initial value, and increment
2299 value if possible. */
2301 /* If this is a new register, can't handle it since we don't have any
2302 reg_iv_type entry for it. */
2303 if (REGNO (iteration_var
) >= max_reg_before_loop
)
2305 if (loop_dump_stream
)
2306 fprintf (loop_dump_stream
,
2307 "Loop unrolling: No reg_iv_type entry for iteration var.\n");
2311 /* Reject iteration variables larger than the host wide int size, since they
2312 could result in a number of iterations greater than the range of our
2313 `unsigned HOST_WIDE_INT' variable loop_n_iterations. */
2314 else if ((GET_MODE_BITSIZE (GET_MODE (iteration_var
))
2315 > HOST_BITS_PER_WIDE_INT
))
2317 if (loop_dump_stream
)
2318 fprintf (loop_dump_stream
,
2319 "Loop unrolling: Iteration var rejected because mode too large.\n");
2322 else if (GET_MODE_CLASS (GET_MODE (iteration_var
)) != MODE_INT
)
2324 if (loop_dump_stream
)
2325 fprintf (loop_dump_stream
,
2326 "Loop unrolling: Iteration var not an integer.\n");
2329 else if (reg_iv_type
[REGNO (iteration_var
)] == BASIC_INDUCT
)
2331 /* Grab initial value, only useful if it is a constant. */
2332 bl
= reg_biv_class
[REGNO (iteration_var
)];
2333 *initial_value
= bl
->initial_value
;
2335 *increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2337 else if (reg_iv_type
[REGNO (iteration_var
)] == GENERAL_INDUCT
)
2340 /* ??? The code below does not work because the incorrect number of
2341 iterations is calculated when the biv is incremented after the giv
2342 is set (which is the usual case). This can probably be accounted
2343 for by biasing the initial_value by subtracting the amount of the
2344 increment that occurs between the giv set and the giv test. However,
2345 a giv as an iterator is very rare, so it does not seem worthwhile
2347 /* ??? An example failure is: i = 6; do {;} while (i++ < 9). */
2348 if (loop_dump_stream
)
2349 fprintf (loop_dump_stream
,
2350 "Loop unrolling: Giv iterators are not handled.\n");
2353 /* Initial value is mult_val times the biv's initial value plus
2354 add_val. Only useful if it is a constant. */
2355 v
= reg_iv_info
[REGNO (iteration_var
)];
2356 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
2357 *initial_value
= fold_rtx_mult_add (v
->mult_val
, bl
->initial_value
,
2358 v
->add_val
, v
->mode
);
2360 /* Increment value is mult_val times the increment value of the biv. */
2362 *increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2364 *increment
= fold_rtx_mult_add (v
->mult_val
, *increment
, const0_rtx
,
2370 if (loop_dump_stream
)
2371 fprintf (loop_dump_stream
,
2372 "Loop unrolling: Not basic or general induction var.\n");
2377 /* Calculate the approximate final value of the iteration variable
2378 which has an loop exit test with code COMPARISON_CODE and comparison value
2379 of COMPARISON_VALUE. Also returns an indication of whether the comparison
2380 was signed or unsigned, and the direction of the comparison. This info is
2381 needed to calculate the number of loop iterations. */
2384 approx_final_value (comparison_code
, comparison_value
, unsigned_p
, compare_dir
)
2385 enum rtx_code comparison_code
;
2386 rtx comparison_value
;
2390 /* Calculate the final value of the induction variable.
2391 The exact final value depends on the branch operator, and increment sign.
2392 This is only an approximate value. It will be wrong if the iteration
2393 variable is not incremented by one each time through the loop, and
2394 approx final value - start value % increment != 0. */
2397 switch (comparison_code
)
2403 return plus_constant (comparison_value
, 1);
2408 return plus_constant (comparison_value
, -1);
2410 /* Can not calculate a final value for this case. */
2417 return comparison_value
;
2423 return comparison_value
;
2426 return comparison_value
;
2432 /* For each biv and giv, determine whether it can be safely split into
2433 a different variable for each unrolled copy of the loop body. If it
2434 is safe to split, then indicate that by saving some useful info
2435 in the splittable_regs array.
2437 If the loop is being completely unrolled, then splittable_regs will hold
2438 the current value of the induction variable while the loop is unrolled.
2439 It must be set to the initial value of the induction variable here.
2440 Otherwise, splittable_regs will hold the difference between the current
2441 value of the induction variable and the value the induction variable had
2442 at the top of the loop. It must be set to the value 0 here.
2444 Returns the total number of instructions that set registers that are
2447 /* ?? If the loop is only unrolled twice, then most of the restrictions to
2448 constant values are unnecessary, since we can easily calculate increment
2449 values in this case even if nothing is constant. The increment value
2450 should not involve a multiply however. */
2452 /* ?? Even if the biv/giv increment values aren't constant, it may still
2453 be beneficial to split the variable if the loop is only unrolled a few
2454 times, since multiplies by small integers (1,2,3,4) are very cheap. */
2457 find_splittable_regs (unroll_type
, loop_start
, loop_end
, end_insert_before
,
2459 enum unroll_types unroll_type
;
2460 rtx loop_start
, loop_end
;
2461 rtx end_insert_before
;
2464 struct iv_class
*bl
;
2465 struct induction
*v
;
2467 rtx biv_final_value
;
2471 for (bl
= loop_iv_list
; bl
; bl
= bl
->next
)
2473 /* Biv_total_increment must return a constant value,
2474 otherwise we can not calculate the split values. */
2476 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2477 if (! increment
|| GET_CODE (increment
) != CONST_INT
)
2480 /* The loop must be unrolled completely, or else have a known number
2481 of iterations and only one exit, or else the biv must be dead
2482 outside the loop, or else the final value must be known. Otherwise,
2483 it is unsafe to split the biv since it may not have the proper
2484 value on loop exit. */
2486 /* loop_number_exit_count is non-zero if the loop has an exit other than
2487 a fall through at the end. */
2490 biv_final_value
= 0;
2491 if (unroll_type
!= UNROLL_COMPLETELY
2492 && (loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]]
2493 || unroll_type
== UNROLL_NAIVE
)
2494 && (uid_luid
[REGNO_LAST_UID (bl
->regno
)] >= INSN_LUID (loop_end
)
2496 || INSN_UID (bl
->init_insn
) >= max_uid_for_loop
2497 || (uid_luid
[REGNO_FIRST_UID (bl
->regno
)]
2498 < INSN_LUID (bl
->init_insn
))
2499 || reg_mentioned_p (bl
->biv
->dest_reg
, SET_SRC (bl
->init_set
)))
2500 && ! (biv_final_value
= final_biv_value (bl
, loop_start
, loop_end
)))
2503 /* If any of the insns setting the BIV don't do so with a simple
2504 PLUS, we don't know how to split it. */
2505 for (v
= bl
->biv
; biv_splittable
&& v
; v
= v
->next_iv
)
2506 if ((tem
= single_set (v
->insn
)) == 0
2507 || GET_CODE (SET_DEST (tem
)) != REG
2508 || REGNO (SET_DEST (tem
)) != bl
->regno
2509 || GET_CODE (SET_SRC (tem
)) != PLUS
)
2512 /* If final value is non-zero, then must emit an instruction which sets
2513 the value of the biv to the proper value. This is done after
2514 handling all of the givs, since some of them may need to use the
2515 biv's value in their initialization code. */
2517 /* This biv is splittable. If completely unrolling the loop, save
2518 the biv's initial value. Otherwise, save the constant zero. */
2520 if (biv_splittable
== 1)
2522 if (unroll_type
== UNROLL_COMPLETELY
)
2524 /* If the initial value of the biv is itself (i.e. it is too
2525 complicated for strength_reduce to compute), or is a hard
2526 register, or it isn't invariant, then we must create a new
2527 pseudo reg to hold the initial value of the biv. */
2529 if (GET_CODE (bl
->initial_value
) == REG
2530 && (REGNO (bl
->initial_value
) == bl
->regno
2531 || REGNO (bl
->initial_value
) < FIRST_PSEUDO_REGISTER
2532 || ! invariant_p (bl
->initial_value
)))
2534 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2536 record_base_value (REGNO (tem
), bl
->biv
->add_val
);
2537 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2540 if (loop_dump_stream
)
2541 fprintf (loop_dump_stream
, "Biv %d initial value remapped to %d.\n",
2542 bl
->regno
, REGNO (tem
));
2544 splittable_regs
[bl
->regno
] = tem
;
2547 splittable_regs
[bl
->regno
] = bl
->initial_value
;
2550 splittable_regs
[bl
->regno
] = const0_rtx
;
2552 /* Save the number of instructions that modify the biv, so that
2553 we can treat the last one specially. */
2555 splittable_regs_updates
[bl
->regno
] = bl
->biv_count
;
2556 result
+= bl
->biv_count
;
2558 if (loop_dump_stream
)
2559 fprintf (loop_dump_stream
,
2560 "Biv %d safe to split.\n", bl
->regno
);
2563 /* Check every giv that depends on this biv to see whether it is
2564 splittable also. Even if the biv isn't splittable, givs which
2565 depend on it may be splittable if the biv is live outside the
2566 loop, and the givs aren't. */
2568 result
+= find_splittable_givs (bl
, unroll_type
, loop_start
, loop_end
,
2569 increment
, unroll_number
);
2571 /* If final value is non-zero, then must emit an instruction which sets
2572 the value of the biv to the proper value. This is done after
2573 handling all of the givs, since some of them may need to use the
2574 biv's value in their initialization code. */
2575 if (biv_final_value
)
2577 /* If the loop has multiple exits, emit the insns before the
2578 loop to ensure that it will always be executed no matter
2579 how the loop exits. Otherwise emit the insn after the loop,
2580 since this is slightly more efficient. */
2581 if (! loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]])
2582 emit_insn_before (gen_move_insn (bl
->biv
->src_reg
,
2587 /* Create a new register to hold the value of the biv, and then
2588 set the biv to its final value before the loop start. The biv
2589 is set to its final value before loop start to ensure that
2590 this insn will always be executed, no matter how the loop
2592 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2593 record_base_value (REGNO (tem
), bl
->biv
->add_val
);
2595 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2597 emit_insn_before (gen_move_insn (bl
->biv
->src_reg
,
2601 if (loop_dump_stream
)
2602 fprintf (loop_dump_stream
, "Biv %d mapped to %d for split.\n",
2603 REGNO (bl
->biv
->src_reg
), REGNO (tem
));
2605 /* Set up the mapping from the original biv register to the new
2607 bl
->biv
->src_reg
= tem
;
2614 /* Return 1 if the first and last unrolled copy of the address giv V is valid
2615 for the instruction that is using it. Do not make any changes to that
2619 verify_addresses (v
, giv_inc
, unroll_number
)
2620 struct induction
*v
;
2625 rtx orig_addr
= *v
->location
;
2626 rtx last_addr
= plus_constant (v
->dest_reg
,
2627 INTVAL (giv_inc
) * (unroll_number
- 1));
2629 /* First check to see if either address would fail. */
2630 if (! validate_change (v
->insn
, v
->location
, v
->dest_reg
, 0)
2631 || ! validate_change (v
->insn
, v
->location
, last_addr
, 0))
2634 /* Now put things back the way they were before. This will always
2636 validate_change (v
->insn
, v
->location
, orig_addr
, 0);
2641 /* For every giv based on the biv BL, check to determine whether it is
2642 splittable. This is a subroutine to find_splittable_regs ().
2644 Return the number of instructions that set splittable registers. */
2647 find_splittable_givs (bl
, unroll_type
, loop_start
, loop_end
, increment
,
2649 struct iv_class
*bl
;
2650 enum unroll_types unroll_type
;
2651 rtx loop_start
, loop_end
;
2655 struct induction
*v
, *v2
;
2660 /* Scan the list of givs, and set the same_insn field when there are
2661 multiple identical givs in the same insn. */
2662 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
2663 for (v2
= v
->next_iv
; v2
; v2
= v2
->next_iv
)
2664 if (v
->insn
== v2
->insn
&& rtx_equal_p (v
->new_reg
, v2
->new_reg
)
2668 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
2672 /* Only split the giv if it has already been reduced, or if the loop is
2673 being completely unrolled. */
2674 if (unroll_type
!= UNROLL_COMPLETELY
&& v
->ignore
)
2677 /* The giv can be split if the insn that sets the giv is executed once
2678 and only once on every iteration of the loop. */
2679 /* An address giv can always be split. v->insn is just a use not a set,
2680 and hence it does not matter whether it is always executed. All that
2681 matters is that all the biv increments are always executed, and we
2682 won't reach here if they aren't. */
2683 if (v
->giv_type
!= DEST_ADDR
2684 && (! v
->always_computable
2685 || back_branch_in_range_p (v
->insn
, loop_start
, loop_end
)))
2688 /* The giv increment value must be a constant. */
2689 giv_inc
= fold_rtx_mult_add (v
->mult_val
, increment
, const0_rtx
,
2691 if (! giv_inc
|| GET_CODE (giv_inc
) != CONST_INT
)
2694 /* The loop must be unrolled completely, or else have a known number of
2695 iterations and only one exit, or else the giv must be dead outside
2696 the loop, or else the final value of the giv must be known.
2697 Otherwise, it is not safe to split the giv since it may not have the
2698 proper value on loop exit. */
2700 /* The used outside loop test will fail for DEST_ADDR givs. They are
2701 never used outside the loop anyways, so it is always safe to split a
2705 if (unroll_type
!= UNROLL_COMPLETELY
2706 && (loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]]
2707 || unroll_type
== UNROLL_NAIVE
)
2708 && v
->giv_type
!= DEST_ADDR
2709 && ((REGNO_FIRST_UID (REGNO (v
->dest_reg
)) != INSN_UID (v
->insn
)
2710 /* Check for the case where the pseudo is set by a shift/add
2711 sequence, in which case the first insn setting the pseudo
2712 is the first insn of the shift/add sequence. */
2713 && (! (tem
= find_reg_note (v
->insn
, REG_RETVAL
, NULL_RTX
))
2714 || (REGNO_FIRST_UID (REGNO (v
->dest_reg
))
2715 != INSN_UID (XEXP (tem
, 0)))))
2716 /* Line above always fails if INSN was moved by loop opt. */
2717 || (uid_luid
[REGNO_LAST_UID (REGNO (v
->dest_reg
))]
2718 >= INSN_LUID (loop_end
)))
2719 && ! (final_value
= v
->final_value
))
2723 /* Currently, non-reduced/final-value givs are never split. */
2724 /* Should emit insns after the loop if possible, as the biv final value
2727 /* If the final value is non-zero, and the giv has not been reduced,
2728 then must emit an instruction to set the final value. */
2729 if (final_value
&& !v
->new_reg
)
2731 /* Create a new register to hold the value of the giv, and then set
2732 the giv to its final value before the loop start. The giv is set
2733 to its final value before loop start to ensure that this insn
2734 will always be executed, no matter how we exit. */
2735 tem
= gen_reg_rtx (v
->mode
);
2736 emit_insn_before (gen_move_insn (tem
, v
->dest_reg
), loop_start
);
2737 emit_insn_before (gen_move_insn (v
->dest_reg
, final_value
),
2740 if (loop_dump_stream
)
2741 fprintf (loop_dump_stream
, "Giv %d mapped to %d for split.\n",
2742 REGNO (v
->dest_reg
), REGNO (tem
));
2748 /* This giv is splittable. If completely unrolling the loop, save the
2749 giv's initial value. Otherwise, save the constant zero for it. */
2751 if (unroll_type
== UNROLL_COMPLETELY
)
2753 /* It is not safe to use bl->initial_value here, because it may not
2754 be invariant. It is safe to use the initial value stored in
2755 the splittable_regs array if it is set. In rare cases, it won't
2756 be set, so then we do exactly the same thing as
2757 find_splittable_regs does to get a safe value. */
2758 rtx biv_initial_value
;
2760 if (splittable_regs
[bl
->regno
])
2761 biv_initial_value
= splittable_regs
[bl
->regno
];
2762 else if (GET_CODE (bl
->initial_value
) != REG
2763 || (REGNO (bl
->initial_value
) != bl
->regno
2764 && REGNO (bl
->initial_value
) >= FIRST_PSEUDO_REGISTER
))
2765 biv_initial_value
= bl
->initial_value
;
2768 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2770 record_base_value (REGNO (tem
), bl
->biv
->add_val
);
2771 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2773 biv_initial_value
= tem
;
2775 value
= fold_rtx_mult_add (v
->mult_val
, biv_initial_value
,
2776 v
->add_val
, v
->mode
);
2783 /* If a giv was combined with another giv, then we can only split
2784 this giv if the giv it was combined with was reduced. This
2785 is because the value of v->new_reg is meaningless in this
2787 if (v
->same
&& ! v
->same
->new_reg
)
2789 if (loop_dump_stream
)
2790 fprintf (loop_dump_stream
,
2791 "giv combined with unreduced giv not split.\n");
2794 /* If the giv is an address destination, it could be something other
2795 than a simple register, these have to be treated differently. */
2796 else if (v
->giv_type
== DEST_REG
)
2798 /* If value is not a constant, register, or register plus
2799 constant, then compute its value into a register before
2800 loop start. This prevents invalid rtx sharing, and should
2801 generate better code. We can use bl->initial_value here
2802 instead of splittable_regs[bl->regno] because this code
2803 is going before the loop start. */
2804 if (unroll_type
== UNROLL_COMPLETELY
2805 && GET_CODE (value
) != CONST_INT
2806 && GET_CODE (value
) != REG
2807 && (GET_CODE (value
) != PLUS
2808 || GET_CODE (XEXP (value
, 0)) != REG
2809 || GET_CODE (XEXP (value
, 1)) != CONST_INT
))
2811 rtx tem
= gen_reg_rtx (v
->mode
);
2812 record_base_value (REGNO (tem
), v
->add_val
);
2813 emit_iv_add_mult (bl
->initial_value
, v
->mult_val
,
2814 v
->add_val
, tem
, loop_start
);
2818 splittable_regs
[REGNO (v
->new_reg
)] = value
;
2822 /* Splitting address givs is useful since it will often allow us
2823 to eliminate some increment insns for the base giv as
2826 /* If the addr giv is combined with a dest_reg giv, then all
2827 references to that dest reg will be remapped, which is NOT
2828 what we want for split addr regs. We always create a new
2829 register for the split addr giv, just to be safe. */
2831 /* If we have multiple identical address givs within a
2832 single instruction, then use a single pseudo reg for
2833 both. This is necessary in case one is a match_dup
2836 v
->const_adjust
= 0;
2840 v
->dest_reg
= v
->same_insn
->dest_reg
;
2841 if (loop_dump_stream
)
2842 fprintf (loop_dump_stream
,
2843 "Sharing address givs in insn %d\n",
2844 INSN_UID (v
->insn
));
2846 /* If multiple address GIVs have been combined with the
2847 same dest_reg GIV, do not create a new register for
2849 else if (unroll_type
!= UNROLL_COMPLETELY
2850 && v
->giv_type
== DEST_ADDR
2851 && v
->same
&& v
->same
->giv_type
== DEST_ADDR
2852 && v
->same
->unrolled
)
2854 v
->dest_reg
= v
->same
->dest_reg
;
2857 else if (unroll_type
!= UNROLL_COMPLETELY
)
2859 /* If not completely unrolling the loop, then create a new
2860 register to hold the split value of the DEST_ADDR giv.
2861 Emit insn to initialize its value before loop start. */
2863 rtx tem
= gen_reg_rtx (v
->mode
);
2864 record_base_value (REGNO (tem
), v
->add_val
);
2867 /* If the address giv has a constant in its new_reg value,
2868 then this constant can be pulled out and put in value,
2869 instead of being part of the initialization code. */
2871 if (GET_CODE (v
->new_reg
) == PLUS
2872 && GET_CODE (XEXP (v
->new_reg
, 1)) == CONST_INT
)
2875 = plus_constant (tem
, INTVAL (XEXP (v
->new_reg
,1)));
2877 /* Only succeed if this will give valid addresses.
2878 Try to validate both the first and the last
2879 address resulting from loop unrolling, if
2880 one fails, then can't do const elim here. */
2881 if (verify_addresses (v
, giv_inc
, unroll_number
))
2883 /* Save the negative of the eliminated const, so
2884 that we can calculate the dest_reg's increment
2886 v
->const_adjust
= - INTVAL (XEXP (v
->new_reg
, 1));
2888 v
->new_reg
= XEXP (v
->new_reg
, 0);
2889 if (loop_dump_stream
)
2890 fprintf (loop_dump_stream
,
2891 "Eliminating constant from giv %d\n",
2900 /* If the address hasn't been checked for validity yet, do so
2901 now, and fail completely if either the first or the last
2902 unrolled copy of the address is not a valid address
2903 for the instruction that uses it. */
2904 if (v
->dest_reg
== tem
2905 && ! verify_addresses (v
, giv_inc
, unroll_number
))
2907 if (loop_dump_stream
)
2908 fprintf (loop_dump_stream
,
2909 "Invalid address for giv at insn %d\n",
2910 INSN_UID (v
->insn
));
2914 /* To initialize the new register, just move the value of
2915 new_reg into it. This is not guaranteed to give a valid
2916 instruction on machines with complex addressing modes.
2917 If we can't recognize it, then delete it and emit insns
2918 to calculate the value from scratch. */
2919 emit_insn_before (gen_rtx (SET
, VOIDmode
, tem
,
2920 copy_rtx (v
->new_reg
)),
2922 if (recog_memoized (PREV_INSN (loop_start
)) < 0)
2926 /* We can't use bl->initial_value to compute the initial
2927 value, because the loop may have been preconditioned.
2928 We must calculate it from NEW_REG. Try using
2929 force_operand instead of emit_iv_add_mult. */
2930 delete_insn (PREV_INSN (loop_start
));
2933 ret
= force_operand (v
->new_reg
, tem
);
2935 emit_move_insn (tem
, ret
);
2936 sequence
= gen_sequence ();
2938 emit_insn_before (sequence
, loop_start
);
2940 if (loop_dump_stream
)
2941 fprintf (loop_dump_stream
,
2942 "Invalid init insn, rewritten.\n");
2947 v
->dest_reg
= value
;
2949 /* Check the resulting address for validity, and fail
2950 if the resulting address would be invalid. */
2951 if (! verify_addresses (v
, giv_inc
, unroll_number
))
2953 if (loop_dump_stream
)
2954 fprintf (loop_dump_stream
,
2955 "Invalid address for giv at insn %d\n",
2956 INSN_UID (v
->insn
));
2961 /* Store the value of dest_reg into the insn. This sharing
2962 will not be a problem as this insn will always be copied
2965 *v
->location
= v
->dest_reg
;
2967 /* If this address giv is combined with a dest reg giv, then
2968 save the base giv's induction pointer so that we will be
2969 able to handle this address giv properly. The base giv
2970 itself does not have to be splittable. */
2972 if (v
->same
&& v
->same
->giv_type
== DEST_REG
)
2973 addr_combined_regs
[REGNO (v
->same
->new_reg
)] = v
->same
;
2975 if (GET_CODE (v
->new_reg
) == REG
)
2977 /* This giv maybe hasn't been combined with any others.
2978 Make sure that it's giv is marked as splittable here. */
2980 splittable_regs
[REGNO (v
->new_reg
)] = value
;
2982 /* Make it appear to depend upon itself, so that the
2983 giv will be properly split in the main loop above. */
2987 addr_combined_regs
[REGNO (v
->new_reg
)] = v
;
2991 if (loop_dump_stream
)
2992 fprintf (loop_dump_stream
, "DEST_ADDR giv being split.\n");
2998 /* Currently, unreduced giv's can't be split. This is not too much
2999 of a problem since unreduced giv's are not live across loop
3000 iterations anyways. When unrolling a loop completely though,
3001 it makes sense to reduce&split givs when possible, as this will
3002 result in simpler instructions, and will not require that a reg
3003 be live across loop iterations. */
3005 splittable_regs
[REGNO (v
->dest_reg
)] = value
;
3006 fprintf (stderr
, "Giv %d at insn %d not reduced\n",
3007 REGNO (v
->dest_reg
), INSN_UID (v
->insn
));
3013 /* Unreduced givs are only updated once by definition. Reduced givs
3014 are updated as many times as their biv is. Mark it so if this is
3015 a splittable register. Don't need to do anything for address givs
3016 where this may not be a register. */
3018 if (GET_CODE (v
->new_reg
) == REG
)
3022 count
= reg_biv_class
[REGNO (v
->src_reg
)]->biv_count
;
3024 splittable_regs_updates
[REGNO (v
->new_reg
)] = count
;
3029 if (loop_dump_stream
)
3033 if (GET_CODE (v
->dest_reg
) == CONST_INT
)
3035 else if (GET_CODE (v
->dest_reg
) != REG
)
3036 regnum
= REGNO (XEXP (v
->dest_reg
, 0));
3038 regnum
= REGNO (v
->dest_reg
);
3039 fprintf (loop_dump_stream
, "Giv %d at insn %d safe to split.\n",
3040 regnum
, INSN_UID (v
->insn
));
3047 /* Try to prove that the register is dead after the loop exits. Trace every
3048 loop exit looking for an insn that will always be executed, which sets
3049 the register to some value, and appears before the first use of the register
3050 is found. If successful, then return 1, otherwise return 0. */
3052 /* ?? Could be made more intelligent in the handling of jumps, so that
3053 it can search past if statements and other similar structures. */
3056 reg_dead_after_loop (reg
, loop_start
, loop_end
)
3057 rtx reg
, loop_start
, loop_end
;
3062 int label_count
= 0;
3063 int this_loop_num
= uid_loop_num
[INSN_UID (loop_start
)];
3065 /* In addition to checking all exits of this loop, we must also check
3066 all exits of inner nested loops that would exit this loop. We don't
3067 have any way to identify those, so we just give up if there are any
3068 such inner loop exits. */
3070 for (label
= loop_number_exit_labels
[this_loop_num
]; label
;
3071 label
= LABEL_NEXTREF (label
))
3074 if (label_count
!= loop_number_exit_count
[this_loop_num
])
3077 /* HACK: Must also search the loop fall through exit, create a label_ref
3078 here which points to the loop_end, and append the loop_number_exit_labels
3080 label
= gen_rtx (LABEL_REF
, VOIDmode
, loop_end
);
3081 LABEL_NEXTREF (label
) = loop_number_exit_labels
[this_loop_num
];
3083 for ( ; label
; label
= LABEL_NEXTREF (label
))
3085 /* Succeed if find an insn which sets the biv or if reach end of
3086 function. Fail if find an insn that uses the biv, or if come to
3087 a conditional jump. */
3089 insn
= NEXT_INSN (XEXP (label
, 0));
3092 code
= GET_CODE (insn
);
3093 if (GET_RTX_CLASS (code
) == 'i')
3097 if (reg_referenced_p (reg
, PATTERN (insn
)))
3100 set
= single_set (insn
);
3101 if (set
&& rtx_equal_p (SET_DEST (set
), reg
))
3105 if (code
== JUMP_INSN
)
3107 if (GET_CODE (PATTERN (insn
)) == RETURN
)
3109 else if (! simplejump_p (insn
)
3110 /* Prevent infinite loop following infinite loops. */
3111 || jump_count
++ > 20)
3114 insn
= JUMP_LABEL (insn
);
3117 insn
= NEXT_INSN (insn
);
3121 /* Success, the register is dead on all loop exits. */
3125 /* Try to calculate the final value of the biv, the value it will have at
3126 the end of the loop. If we can do it, return that value. */
3129 final_biv_value (bl
, loop_start
, loop_end
)
3130 struct iv_class
*bl
;
3131 rtx loop_start
, loop_end
;
3135 /* ??? This only works for MODE_INT biv's. Reject all others for now. */
3137 if (GET_MODE_CLASS (bl
->biv
->mode
) != MODE_INT
)
3140 /* The final value for reversed bivs must be calculated differently than
3141 for ordinary bivs. In this case, there is already an insn after the
3142 loop which sets this biv's final value (if necessary), and there are
3143 no other loop exits, so we can return any value. */
3146 if (loop_dump_stream
)
3147 fprintf (loop_dump_stream
,
3148 "Final biv value for %d, reversed biv.\n", bl
->regno
);
3153 /* Try to calculate the final value as initial value + (number of iterations
3154 * increment). For this to work, increment must be invariant, the only
3155 exit from the loop must be the fall through at the bottom (otherwise
3156 it may not have its final value when the loop exits), and the initial
3157 value of the biv must be invariant. */
3159 if (loop_n_iterations
!= 0
3160 && ! loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]]
3161 && invariant_p (bl
->initial_value
))
3163 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
3165 if (increment
&& invariant_p (increment
))
3167 /* Can calculate the loop exit value, emit insns after loop
3168 end to calculate this value into a temporary register in
3169 case it is needed later. */
3171 tem
= gen_reg_rtx (bl
->biv
->mode
);
3172 record_base_value (REGNO (tem
), bl
->biv
->add_val
);
3173 /* Make sure loop_end is not the last insn. */
3174 if (NEXT_INSN (loop_end
) == 0)
3175 emit_note_after (NOTE_INSN_DELETED
, loop_end
);
3176 emit_iv_add_mult (increment
, GEN_INT (loop_n_iterations
),
3177 bl
->initial_value
, tem
, NEXT_INSN (loop_end
));
3179 if (loop_dump_stream
)
3180 fprintf (loop_dump_stream
,
3181 "Final biv value for %d, calculated.\n", bl
->regno
);
3187 /* Check to see if the biv is dead at all loop exits. */
3188 if (reg_dead_after_loop (bl
->biv
->src_reg
, loop_start
, loop_end
))
3190 if (loop_dump_stream
)
3191 fprintf (loop_dump_stream
,
3192 "Final biv value for %d, biv dead after loop exit.\n",
3201 /* Try to calculate the final value of the giv, the value it will have at
3202 the end of the loop. If we can do it, return that value. */
3205 final_giv_value (v
, loop_start
, loop_end
)
3206 struct induction
*v
;
3207 rtx loop_start
, loop_end
;
3209 struct iv_class
*bl
;
3212 rtx insert_before
, seq
;
3214 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
3216 /* The final value for givs which depend on reversed bivs must be calculated
3217 differently than for ordinary givs. In this case, there is already an
3218 insn after the loop which sets this giv's final value (if necessary),
3219 and there are no other loop exits, so we can return any value. */
3222 if (loop_dump_stream
)
3223 fprintf (loop_dump_stream
,
3224 "Final giv value for %d, depends on reversed biv\n",
3225 REGNO (v
->dest_reg
));
3229 /* Try to calculate the final value as a function of the biv it depends
3230 upon. The only exit from the loop must be the fall through at the bottom
3231 (otherwise it may not have its final value when the loop exits). */
3233 /* ??? Can calculate the final giv value by subtracting off the
3234 extra biv increments times the giv's mult_val. The loop must have
3235 only one exit for this to work, but the loop iterations does not need
3238 if (loop_n_iterations
!= 0
3239 && ! loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]])
3241 /* ?? It is tempting to use the biv's value here since these insns will
3242 be put after the loop, and hence the biv will have its final value
3243 then. However, this fails if the biv is subsequently eliminated.
3244 Perhaps determine whether biv's are eliminable before trying to
3245 determine whether giv's are replaceable so that we can use the
3246 biv value here if it is not eliminable. */
3248 /* We are emitting code after the end of the loop, so we must make
3249 sure that bl->initial_value is still valid then. It will still
3250 be valid if it is invariant. */
3252 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
3254 if (increment
&& invariant_p (increment
)
3255 && invariant_p (bl
->initial_value
))
3257 /* Can calculate the loop exit value of its biv as
3258 (loop_n_iterations * increment) + initial_value */
3260 /* The loop exit value of the giv is then
3261 (final_biv_value - extra increments) * mult_val + add_val.
3262 The extra increments are any increments to the biv which
3263 occur in the loop after the giv's value is calculated.
3264 We must search from the insn that sets the giv to the end
3265 of the loop to calculate this value. */
3267 insert_before
= NEXT_INSN (loop_end
);
3269 /* Put the final biv value in tem. */
3270 tem
= gen_reg_rtx (bl
->biv
->mode
);
3271 record_base_value (REGNO (tem
), bl
->biv
->add_val
);
3272 emit_iv_add_mult (increment
, GEN_INT (loop_n_iterations
),
3273 bl
->initial_value
, tem
, insert_before
);
3275 /* Subtract off extra increments as we find them. */
3276 for (insn
= NEXT_INSN (v
->insn
); insn
!= loop_end
;
3277 insn
= NEXT_INSN (insn
))
3279 struct induction
*biv
;
3281 for (biv
= bl
->biv
; biv
; biv
= biv
->next_iv
)
3282 if (biv
->insn
== insn
)
3285 tem
= expand_binop (GET_MODE (tem
), sub_optab
, tem
,
3286 biv
->add_val
, NULL_RTX
, 0,
3288 seq
= gen_sequence ();
3290 emit_insn_before (seq
, insert_before
);
3294 /* Now calculate the giv's final value. */
3295 emit_iv_add_mult (tem
, v
->mult_val
, v
->add_val
, tem
,
3298 if (loop_dump_stream
)
3299 fprintf (loop_dump_stream
,
3300 "Final giv value for %d, calc from biv's value.\n",
3301 REGNO (v
->dest_reg
));
3307 /* Replaceable giv's should never reach here. */
3311 /* Check to see if the biv is dead at all loop exits. */
3312 if (reg_dead_after_loop (v
->dest_reg
, loop_start
, loop_end
))
3314 if (loop_dump_stream
)
3315 fprintf (loop_dump_stream
,
3316 "Final giv value for %d, giv dead after loop exit.\n",
3317 REGNO (v
->dest_reg
));
3326 /* Calculate the number of loop iterations. Returns the exact number of loop
3327 iterations if it can be calculated, otherwise returns zero. */
3329 unsigned HOST_WIDE_INT
3330 loop_iterations (loop_start
, loop_end
)
3331 rtx loop_start
, loop_end
;
3333 rtx comparison
, comparison_value
;
3334 rtx iteration_var
, initial_value
, increment
, final_value
;
3335 enum rtx_code comparison_code
;
3338 int unsigned_compare
, compare_dir
, final_larger
;
3339 unsigned long tempu
;
3342 /* First find the iteration variable. If the last insn is a conditional
3343 branch, and the insn before tests a register value, make that the
3344 iteration variable. */
3346 loop_initial_value
= 0;
3348 loop_final_value
= 0;
3349 loop_iteration_var
= 0;
3351 /* We used to use pren_nonnote_insn here, but that fails because it might
3352 accidentally get the branch for a contained loop if the branch for this
3353 loop was deleted. We can only trust branches immediately before the
3355 last_loop_insn
= PREV_INSN (loop_end
);
3357 comparison
= get_condition_for_loop (last_loop_insn
);
3358 if (comparison
== 0)
3360 if (loop_dump_stream
)
3361 fprintf (loop_dump_stream
,
3362 "Loop unrolling: No final conditional branch found.\n");
3366 /* ??? Get_condition may switch position of induction variable and
3367 invariant register when it canonicalizes the comparison. */
3369 comparison_code
= GET_CODE (comparison
);
3370 iteration_var
= XEXP (comparison
, 0);
3371 comparison_value
= XEXP (comparison
, 1);
3373 if (GET_CODE (iteration_var
) != REG
)
3375 if (loop_dump_stream
)
3376 fprintf (loop_dump_stream
,
3377 "Loop unrolling: Comparison not against register.\n");
3381 /* Loop iterations is always called before any new registers are created
3382 now, so this should never occur. */
3384 if (REGNO (iteration_var
) >= max_reg_before_loop
)
3387 iteration_info (iteration_var
, &initial_value
, &increment
,
3388 loop_start
, loop_end
);
3389 if (initial_value
== 0)
3390 /* iteration_info already printed a message. */
3393 /* If the comparison value is an invariant register, then try to find
3394 its value from the insns before the start of the loop. */
3396 if (GET_CODE (comparison_value
) == REG
&& invariant_p (comparison_value
))
3400 for (insn
= PREV_INSN (loop_start
); insn
; insn
= PREV_INSN (insn
))
3402 if (GET_CODE (insn
) == CODE_LABEL
)
3405 else if (GET_RTX_CLASS (GET_CODE (insn
)) == 'i'
3406 && reg_set_p (comparison_value
, insn
))
3408 /* We found the last insn before the loop that sets the register.
3409 If it sets the entire register, and has a REG_EQUAL note,
3410 then use the value of the REG_EQUAL note. */
3411 if ((set
= single_set (insn
))
3412 && (SET_DEST (set
) == comparison_value
))
3414 rtx note
= find_reg_note (insn
, REG_EQUAL
, NULL_RTX
);
3416 /* Only use the REG_EQUAL note if it is a constant.
3417 Other things, divide in particular, will cause
3418 problems later if we use them. */
3419 if (note
&& GET_CODE (XEXP (note
, 0)) != EXPR_LIST
3420 && CONSTANT_P (XEXP (note
, 0)))
3421 comparison_value
= XEXP (note
, 0);
3428 final_value
= approx_final_value (comparison_code
, comparison_value
,
3429 &unsigned_compare
, &compare_dir
);
3431 /* Save the calculated values describing this loop's bounds, in case
3432 precondition_loop_p will need them later. These values can not be
3433 recalculated inside precondition_loop_p because strength reduction
3434 optimizations may obscure the loop's structure. */
3436 loop_iteration_var
= iteration_var
;
3437 loop_initial_value
= initial_value
;
3438 loop_increment
= increment
;
3439 loop_final_value
= final_value
;
3440 loop_comparison_code
= comparison_code
;
3444 if (loop_dump_stream
)
3445 fprintf (loop_dump_stream
,
3446 "Loop unrolling: Increment value can't be calculated.\n");
3449 else if (GET_CODE (increment
) != CONST_INT
)
3451 if (loop_dump_stream
)
3452 fprintf (loop_dump_stream
,
3453 "Loop unrolling: Increment value not constant.\n");
3456 else if (GET_CODE (initial_value
) != CONST_INT
)
3458 if (loop_dump_stream
)
3459 fprintf (loop_dump_stream
,
3460 "Loop unrolling: Initial value not constant.\n");
3463 else if (final_value
== 0)
3465 if (loop_dump_stream
)
3466 fprintf (loop_dump_stream
,
3467 "Loop unrolling: EQ comparison loop.\n");
3470 else if (GET_CODE (final_value
) != CONST_INT
)
3472 if (loop_dump_stream
)
3473 fprintf (loop_dump_stream
,
3474 "Loop unrolling: Final value not constant.\n");
3478 /* ?? Final value and initial value do not have to be constants.
3479 Only their difference has to be constant. When the iteration variable
3480 is an array address, the final value and initial value might both
3481 be addresses with the same base but different constant offsets.
3482 Final value must be invariant for this to work.
3484 To do this, need some way to find the values of registers which are
3487 /* Final_larger is 1 if final larger, 0 if they are equal, otherwise -1. */
3488 if (unsigned_compare
)
3490 = ((unsigned HOST_WIDE_INT
) INTVAL (final_value
)
3491 > (unsigned HOST_WIDE_INT
) INTVAL (initial_value
))
3492 - ((unsigned HOST_WIDE_INT
) INTVAL (final_value
)
3493 < (unsigned HOST_WIDE_INT
) INTVAL (initial_value
));
3495 final_larger
= (INTVAL (final_value
) > INTVAL (initial_value
))
3496 - (INTVAL (final_value
) < INTVAL (initial_value
));
3498 if (INTVAL (increment
) > 0)
3500 else if (INTVAL (increment
) == 0)
3505 /* There are 27 different cases: compare_dir = -1, 0, 1;
3506 final_larger = -1, 0, 1; increment_dir = -1, 0, 1.
3507 There are 4 normal cases, 4 reverse cases (where the iteration variable
3508 will overflow before the loop exits), 4 infinite loop cases, and 15
3509 immediate exit (0 or 1 iteration depending on loop type) cases.
3510 Only try to optimize the normal cases. */
3512 /* (compare_dir/final_larger/increment_dir)
3513 Normal cases: (0/-1/-1), (0/1/1), (-1/-1/-1), (1/1/1)
3514 Reverse cases: (0/-1/1), (0/1/-1), (-1/-1/1), (1/1/-1)
3515 Infinite loops: (0/-1/0), (0/1/0), (-1/-1/0), (1/1/0)
3516 Immediate exit: (0/0/X), (-1/0/X), (-1/1/X), (1/0/X), (1/-1/X) */
3518 /* ?? If the meaning of reverse loops (where the iteration variable
3519 will overflow before the loop exits) is undefined, then could
3520 eliminate all of these special checks, and just always assume
3521 the loops are normal/immediate/infinite. Note that this means
3522 the sign of increment_dir does not have to be known. Also,
3523 since it does not really hurt if immediate exit loops or infinite loops
3524 are optimized, then that case could be ignored also, and hence all
3525 loops can be optimized.
3527 According to ANSI Spec, the reverse loop case result is undefined,
3528 because the action on overflow is undefined.
3530 See also the special test for NE loops below. */
3532 if (final_larger
== increment_dir
&& final_larger
!= 0
3533 && (final_larger
== compare_dir
|| compare_dir
== 0))
3538 if (loop_dump_stream
)
3539 fprintf (loop_dump_stream
,
3540 "Loop unrolling: Not normal loop.\n");
3544 /* Calculate the number of iterations, final_value is only an approximation,
3545 so correct for that. Note that tempu and loop_n_iterations are
3546 unsigned, because they can be as large as 2^n - 1. */
3548 i
= INTVAL (increment
);
3550 tempu
= INTVAL (final_value
) - INTVAL (initial_value
);
3553 tempu
= INTVAL (initial_value
) - INTVAL (final_value
);
3559 /* For NE tests, make sure that the iteration variable won't miss the
3560 final value. If tempu mod i is not zero, then the iteration variable
3561 will overflow before the loop exits, and we can not calculate the
3562 number of iterations. */
3563 if (compare_dir
== 0 && (tempu
% i
) != 0)
3566 return tempu
/ i
+ ((tempu
% i
) != 0);
3569 /* Replace uses of split bivs with their split pseudo register. This is
3570 for original instructions which remain after loop unrolling without
3574 remap_split_bivs (x
)
3577 register enum rtx_code code
;
3584 code
= GET_CODE (x
);
3599 /* If non-reduced/final-value givs were split, then this would also
3600 have to remap those givs also. */
3602 if (REGNO (x
) < max_reg_before_loop
3603 && reg_iv_type
[REGNO (x
)] == BASIC_INDUCT
)
3604 return reg_biv_class
[REGNO (x
)]->biv
->src_reg
;
3607 fmt
= GET_RTX_FORMAT (code
);
3608 for (i
= GET_RTX_LENGTH (code
) - 1; i
>= 0; i
--)
3611 XEXP (x
, i
) = remap_split_bivs (XEXP (x
, i
));
3615 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
3616 XVECEXP (x
, i
, j
) = remap_split_bivs (XVECEXP (x
, i
, j
));
3622 /* If FIRST_UID is a set of REGNO, and FIRST_UID dominates LAST_UID (e.g.
3623 FIST_UID is always executed if LAST_UID is), then return 1. Otherwise
3624 return 0. COPY_START is where we can start looking for the insns
3625 FIRST_UID and LAST_UID. COPY_END is where we stop looking for these
3628 If there is no JUMP_INSN between LOOP_START and FIRST_UID, then FIRST_UID
3629 must dominate LAST_UID.
3631 If there is a CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID
3632 may not dominate LAST_UID.
3634 If there is no CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID
3635 must dominate LAST_UID. */
3638 set_dominates_use (regno
, first_uid
, last_uid
, copy_start
, copy_end
)
3645 int passed_jump
= 0;
3646 rtx p
= NEXT_INSN (copy_start
);
3648 while (INSN_UID (p
) != first_uid
)
3650 if (GET_CODE (p
) == JUMP_INSN
)
3652 /* Could not find FIRST_UID. */
3658 /* Verify that FIRST_UID is an insn that entirely sets REGNO. */
3659 if (GET_RTX_CLASS (GET_CODE (p
)) != 'i'
3660 || ! dead_or_set_regno_p (p
, regno
))
3663 /* FIRST_UID is always executed. */
3664 if (passed_jump
== 0)
3667 while (INSN_UID (p
) != last_uid
)
3669 /* If we see a CODE_LABEL between FIRST_UID and LAST_UID, then we
3670 can not be sure that FIRST_UID dominates LAST_UID. */
3671 if (GET_CODE (p
) == CODE_LABEL
)
3673 /* Could not find LAST_UID, but we reached the end of the loop, so
3675 else if (p
== copy_end
)
3680 /* FIRST_UID is always executed if LAST_UID is executed. */