1 /* Try to unroll loops, and split induction variables.
2 Copyright (C) 1992, 93, 94, 95, 97, 1998 Free Software Foundation, Inc.
3 Contributed by James E. Wilson, Cygnus Support/UC Berkeley.
5 This file is part of GNU CC.
7 GNU CC is free software; you can redistribute it and/or modify
8 it under the terms of the GNU General Public License as published by
9 the Free Software Foundation; either version 2, or (at your option)
12 GNU CC is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License for more details.
17 You should have received a copy of the GNU General Public License
18 along with GNU CC; see the file COPYING. If not, write to
19 the Free Software Foundation, 59 Temple Place - Suite 330,
20 Boston, MA 02111-1307, USA. */
22 /* Try to unroll a loop, and split induction variables.
24 Loops for which the number of iterations can be calculated exactly are
25 handled specially. If the number of iterations times the insn_count is
26 less than MAX_UNROLLED_INSNS, then the loop is unrolled completely.
27 Otherwise, we try to unroll the loop a number of times modulo the number
28 of iterations, so that only one exit test will be needed. It is unrolled
29 a number of times approximately equal to MAX_UNROLLED_INSNS divided by
32 Otherwise, if the number of iterations can be calculated exactly at
33 run time, and the loop is always entered at the top, then we try to
34 precondition the loop. That is, at run time, calculate how many times
35 the loop will execute, and then execute the loop body a few times so
36 that the remaining iterations will be some multiple of 4 (or 2 if the
37 loop is large). Then fall through to a loop unrolled 4 (or 2) times,
38 with only one exit test needed at the end of the loop.
40 Otherwise, if the number of iterations can not be calculated exactly,
41 not even at run time, then we still unroll the loop a number of times
42 approximately equal to MAX_UNROLLED_INSNS divided by the insn count,
43 but there must be an exit test after each copy of the loop body.
45 For each induction variable, which is dead outside the loop (replaceable)
46 or for which we can easily calculate the final value, if we can easily
47 calculate its value at each place where it is set as a function of the
48 current loop unroll count and the variable's value at loop entry, then
49 the induction variable is split into `N' different variables, one for
50 each copy of the loop body. One variable is live across the backward
51 branch, and the others are all calculated as a function of this variable.
52 This helps eliminate data dependencies, and leads to further opportunities
55 /* Possible improvements follow: */
57 /* ??? Add an extra pass somewhere to determine whether unrolling will
58 give any benefit. E.g. after generating all unrolled insns, compute the
59 cost of all insns and compare against cost of insns in rolled loop.
61 - On traditional architectures, unrolling a non-constant bound loop
62 is a win if there is a giv whose only use is in memory addresses, the
63 memory addresses can be split, and hence giv increments can be
65 - It is also a win if the loop is executed many times, and preconditioning
66 can be performed for the loop.
67 Add code to check for these and similar cases. */
69 /* ??? Improve control of which loops get unrolled. Could use profiling
70 info to only unroll the most commonly executed loops. Perhaps have
71 a user specifyable option to control the amount of code expansion,
72 or the percent of loops to consider for unrolling. Etc. */
74 /* ??? Look at the register copies inside the loop to see if they form a
75 simple permutation. If so, iterate the permutation until it gets back to
76 the start state. This is how many times we should unroll the loop, for
77 best results, because then all register copies can be eliminated.
78 For example, the lisp nreverse function should be unrolled 3 times
87 ??? The number of times to unroll the loop may also be based on data
88 references in the loop. For example, if we have a loop that references
89 x[i-1], x[i], and x[i+1], we should unroll it a multiple of 3 times. */
91 /* ??? Add some simple linear equation solving capability so that we can
92 determine the number of loop iterations for more complex loops.
93 For example, consider this loop from gdb
94 #define SWAP_TARGET_AND_HOST(buffer,len)
97 char *p = (char *) buffer;
98 char *q = ((char *) buffer) + len - 1;
99 int iterations = (len + 1) >> 1;
101 for (p; p < q; p++, q--;)
109 start value = p = &buffer + current_iteration
110 end value = q = &buffer + len - 1 - current_iteration
111 Given the loop exit test of "p < q", then there must be "q - p" iterations,
112 set equal to zero and solve for number of iterations:
113 q - p = len - 1 - 2*current_iteration = 0
114 current_iteration = (len - 1) / 2
115 Hence, there are (len - 1) / 2 (rounded up to the nearest integer)
116 iterations of this loop. */
118 /* ??? Currently, no labels are marked as loop invariant when doing loop
119 unrolling. This is because an insn inside the loop, that loads the address
120 of a label inside the loop into a register, could be moved outside the loop
121 by the invariant code motion pass if labels were invariant. If the loop
122 is subsequently unrolled, the code will be wrong because each unrolled
123 body of the loop will use the same address, whereas each actually needs a
124 different address. A case where this happens is when a loop containing
125 a switch statement is unrolled.
127 It would be better to let labels be considered invariant. When we
128 unroll loops here, check to see if any insns using a label local to the
129 loop were moved before the loop. If so, then correct the problem, by
130 moving the insn back into the loop, or perhaps replicate the insn before
131 the loop, one copy for each time the loop is unrolled. */
133 /* The prime factors looked for when trying to unroll a loop by some
134 number which is modulo the total number of iterations. Just checking
135 for these 4 prime factors will find at least one factor for 75% of
136 all numbers theoretically. Practically speaking, this will succeed
137 almost all of the time since loops are generally a multiple of 2
140 #define NUM_FACTORS 4
142 struct _factor
{ int factor
, count
; } factors
[NUM_FACTORS
]
143 = { {2, 0}, {3, 0}, {5, 0}, {7, 0}};
145 /* Describes the different types of loop unrolling performed. */
147 enum unroll_types
{ UNROLL_COMPLETELY
, UNROLL_MODULO
, UNROLL_NAIVE
};
152 #include "insn-config.h"
153 #include "integrate.h"
161 /* This controls which loops are unrolled, and by how much we unroll
164 #ifndef MAX_UNROLLED_INSNS
165 #define MAX_UNROLLED_INSNS 100
168 /* Indexed by register number, if non-zero, then it contains a pointer
169 to a struct induction for a DEST_REG giv which has been combined with
170 one of more address givs. This is needed because whenever such a DEST_REG
171 giv is modified, we must modify the value of all split address givs
172 that were combined with this DEST_REG giv. */
174 static struct induction
**addr_combined_regs
;
176 /* Indexed by register number, if this is a splittable induction variable,
177 then this will hold the current value of the register, which depends on the
180 static rtx
*splittable_regs
;
182 /* Indexed by register number, if this is a splittable induction variable,
183 then this will hold the number of instructions in the loop that modify
184 the induction variable. Used to ensure that only the last insn modifying
185 a split iv will update the original iv of the dest. */
187 static int *splittable_regs_updates
;
189 /* Values describing the current loop's iteration variable. These are set up
190 by loop_iterations, and used by precondition_loop_p. */
192 static rtx loop_iteration_var
;
193 static rtx loop_initial_value
;
194 static rtx loop_increment
;
195 static rtx loop_final_value
;
196 static enum rtx_code loop_comparison_code
;
198 /* Forward declarations. */
200 static void init_reg_map
PROTO((struct inline_remap
*, int));
201 static int precondition_loop_p
PROTO((rtx
*, rtx
*, rtx
*, rtx
, rtx
));
202 static rtx calculate_giv_inc
PROTO((rtx
, rtx
, int));
203 static rtx initial_reg_note_copy
PROTO((rtx
, struct inline_remap
*));
204 static void final_reg_note_copy
PROTO((rtx
, struct inline_remap
*));
205 static void copy_loop_body
PROTO((rtx
, rtx
, struct inline_remap
*, rtx
, int,
206 enum unroll_types
, rtx
, rtx
, rtx
, rtx
));
207 void iteration_info
PROTO((rtx
, rtx
*, rtx
*, rtx
, rtx
));
208 static rtx approx_final_value
PROTO((enum rtx_code
, rtx
, int *, int *));
209 static int find_splittable_regs
PROTO((enum unroll_types
, rtx
, rtx
, rtx
, int));
210 static int find_splittable_givs
PROTO((struct iv_class
*,enum unroll_types
,
211 rtx
, rtx
, rtx
, int));
212 static int reg_dead_after_loop
PROTO((rtx
, rtx
, rtx
));
213 static rtx fold_rtx_mult_add
PROTO((rtx
, rtx
, rtx
, enum machine_mode
));
214 static int verify_addresses
PROTO((struct induction
*, rtx
, int));
215 static rtx remap_split_bivs
PROTO((rtx
));
217 /* Try to unroll one loop and split induction variables in the loop.
219 The loop is described by the arguments LOOP_END, INSN_COUNT, and
220 LOOP_START. END_INSERT_BEFORE indicates where insns should be added
221 which need to be executed when the loop falls through. STRENGTH_REDUCTION_P
222 indicates whether information generated in the strength reduction pass
225 This function is intended to be called from within `strength_reduce'
229 unroll_loop (loop_end
, insn_count
, loop_start
, end_insert_before
,
234 rtx end_insert_before
;
235 int strength_reduce_p
;
238 int unroll_number
= 1;
239 rtx copy_start
, copy_end
;
240 rtx insn
, sequence
, pattern
, tem
;
241 int max_labelno
, max_insnno
;
243 struct inline_remap
*map
;
251 int splitting_not_safe
= 0;
252 enum unroll_types unroll_type
;
253 int loop_preconditioned
= 0;
255 /* This points to the last real insn in the loop, which should be either
256 a JUMP_INSN (for conditional jumps) or a BARRIER (for unconditional
260 /* Don't bother unrolling huge loops. Since the minimum factor is
261 two, loops greater than one half of MAX_UNROLLED_INSNS will never
263 if (insn_count
> MAX_UNROLLED_INSNS
/ 2)
265 if (loop_dump_stream
)
266 fprintf (loop_dump_stream
, "Unrolling failure: Loop too big.\n");
270 /* When emitting debugger info, we can't unroll loops with unequal numbers
271 of block_beg and block_end notes, because that would unbalance the block
272 structure of the function. This can happen as a result of the
273 "if (foo) bar; else break;" optimization in jump.c. */
274 /* ??? Gcc has a general policy that -g is never supposed to change the code
275 that the compiler emits, so we must disable this optimization always,
276 even if debug info is not being output. This is rare, so this should
277 not be a significant performance problem. */
279 if (1 /* write_symbols != NO_DEBUG */)
281 int block_begins
= 0;
284 for (insn
= loop_start
; insn
!= loop_end
; insn
= NEXT_INSN (insn
))
286 if (GET_CODE (insn
) == NOTE
)
288 if (NOTE_LINE_NUMBER (insn
) == NOTE_INSN_BLOCK_BEG
)
290 else if (NOTE_LINE_NUMBER (insn
) == NOTE_INSN_BLOCK_END
)
295 if (block_begins
!= block_ends
)
297 if (loop_dump_stream
)
298 fprintf (loop_dump_stream
,
299 "Unrolling failure: Unbalanced block notes.\n");
304 /* Determine type of unroll to perform. Depends on the number of iterations
305 and the size of the loop. */
307 /* If there is no strength reduce info, then set loop_n_iterations to zero.
308 This can happen if strength_reduce can't find any bivs in the loop.
309 A value of zero indicates that the number of iterations could not be
312 if (! strength_reduce_p
)
313 loop_n_iterations
= 0;
315 if (loop_dump_stream
&& loop_n_iterations
> 0)
317 fputs ("Loop unrolling: ", loop_dump_stream
);
318 fprintf (loop_dump_stream
, HOST_WIDE_INT_PRINT_DEC
, loop_n_iterations
);
319 fputs (" iterations.\n", loop_dump_stream
);
322 /* Find and save a pointer to the last nonnote insn in the loop. */
324 last_loop_insn
= prev_nonnote_insn (loop_end
);
326 /* Calculate how many times to unroll the loop. Indicate whether or
327 not the loop is being completely unrolled. */
329 if (loop_n_iterations
== 1)
331 /* If number of iterations is exactly 1, then eliminate the compare and
332 branch at the end of the loop since they will never be taken.
333 Then return, since no other action is needed here. */
335 /* If the last instruction is not a BARRIER or a JUMP_INSN, then
336 don't do anything. */
338 if (GET_CODE (last_loop_insn
) == BARRIER
)
340 /* Delete the jump insn. This will delete the barrier also. */
341 delete_insn (PREV_INSN (last_loop_insn
));
343 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
346 /* The immediately preceding insn is a compare which must be
348 delete_insn (last_loop_insn
);
349 delete_insn (PREV_INSN (last_loop_insn
));
351 /* The immediately preceding insn may not be the compare, so don't
353 delete_insn (last_loop_insn
);
358 else if (loop_n_iterations
> 0
359 && loop_n_iterations
* insn_count
< MAX_UNROLLED_INSNS
)
361 unroll_number
= loop_n_iterations
;
362 unroll_type
= UNROLL_COMPLETELY
;
364 else if (loop_n_iterations
> 0)
366 /* Try to factor the number of iterations. Don't bother with the
367 general case, only using 2, 3, 5, and 7 will get 75% of all
368 numbers theoretically, and almost all in practice. */
370 for (i
= 0; i
< NUM_FACTORS
; i
++)
371 factors
[i
].count
= 0;
373 temp
= loop_n_iterations
;
374 for (i
= NUM_FACTORS
- 1; i
>= 0; i
--)
375 while (temp
% factors
[i
].factor
== 0)
378 temp
= temp
/ factors
[i
].factor
;
381 /* Start with the larger factors first so that we generally
382 get lots of unrolling. */
386 for (i
= 3; i
>= 0; i
--)
387 while (factors
[i
].count
--)
389 if (temp
* factors
[i
].factor
< MAX_UNROLLED_INSNS
)
391 unroll_number
*= factors
[i
].factor
;
392 temp
*= factors
[i
].factor
;
398 /* If we couldn't find any factors, then unroll as in the normal
400 if (unroll_number
== 1)
402 if (loop_dump_stream
)
403 fprintf (loop_dump_stream
,
404 "Loop unrolling: No factors found.\n");
407 unroll_type
= UNROLL_MODULO
;
411 /* Default case, calculate number of times to unroll loop based on its
413 if (unroll_number
== 1)
415 if (8 * insn_count
< MAX_UNROLLED_INSNS
)
417 else if (4 * insn_count
< MAX_UNROLLED_INSNS
)
422 unroll_type
= UNROLL_NAIVE
;
425 /* Now we know how many times to unroll the loop. */
427 if (loop_dump_stream
)
428 fprintf (loop_dump_stream
,
429 "Unrolling loop %d times.\n", unroll_number
);
432 if (unroll_type
== UNROLL_COMPLETELY
|| unroll_type
== UNROLL_MODULO
)
434 /* Loops of these types can start with jump down to the exit condition
435 in rare circumstances.
437 Consider a pair of nested loops where the inner loop is part
438 of the exit code for the outer loop.
440 In this case jump.c will not duplicate the exit test for the outer
441 loop, so it will start with a jump to the exit code.
443 Then consider if the inner loop turns out to iterate once and
444 only once. We will end up deleting the jumps associated with
445 the inner loop. However, the loop notes are not removed from
446 the instruction stream.
448 And finally assume that we can compute the number of iterations
451 In this case unroll may want to unroll the outer loop even though
452 it starts with a jump to the outer loop's exit code.
454 We could try to optimize this case, but it hardly seems worth it.
455 Just return without unrolling the loop in such cases. */
458 while (GET_CODE (insn
) != CODE_LABEL
&& GET_CODE (insn
) != JUMP_INSN
)
459 insn
= NEXT_INSN (insn
);
460 if (GET_CODE (insn
) == JUMP_INSN
)
464 if (unroll_type
== UNROLL_COMPLETELY
)
466 /* Completely unrolling the loop: Delete the compare and branch at
467 the end (the last two instructions). This delete must done at the
468 very end of loop unrolling, to avoid problems with calls to
469 back_branch_in_range_p, which is called by find_splittable_regs.
470 All increments of splittable bivs/givs are changed to load constant
473 copy_start
= loop_start
;
475 /* Set insert_before to the instruction immediately after the JUMP_INSN
476 (or BARRIER), so that any NOTEs between the JUMP_INSN and the end of
477 the loop will be correctly handled by copy_loop_body. */
478 insert_before
= NEXT_INSN (last_loop_insn
);
480 /* Set copy_end to the insn before the jump at the end of the loop. */
481 if (GET_CODE (last_loop_insn
) == BARRIER
)
482 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
483 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
486 /* The instruction immediately before the JUMP_INSN is a compare
487 instruction which we do not want to copy. */
488 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
490 /* The instruction immediately before the JUMP_INSN may not be the
491 compare, so we must copy it. */
492 copy_end
= PREV_INSN (last_loop_insn
);
497 /* We currently can't unroll a loop if it doesn't end with a
498 JUMP_INSN. There would need to be a mechanism that recognizes
499 this case, and then inserts a jump after each loop body, which
500 jumps to after the last loop body. */
501 if (loop_dump_stream
)
502 fprintf (loop_dump_stream
,
503 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
507 else if (unroll_type
== UNROLL_MODULO
)
509 /* Partially unrolling the loop: The compare and branch at the end
510 (the last two instructions) must remain. Don't copy the compare
511 and branch instructions at the end of the loop. Insert the unrolled
512 code immediately before the compare/branch at the end so that the
513 code will fall through to them as before. */
515 copy_start
= loop_start
;
517 /* Set insert_before to the jump insn at the end of the loop.
518 Set copy_end to before the jump insn at the end of the loop. */
519 if (GET_CODE (last_loop_insn
) == BARRIER
)
521 insert_before
= PREV_INSN (last_loop_insn
);
522 copy_end
= PREV_INSN (insert_before
);
524 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
527 /* The instruction immediately before the JUMP_INSN is a compare
528 instruction which we do not want to copy or delete. */
529 insert_before
= PREV_INSN (last_loop_insn
);
530 copy_end
= PREV_INSN (insert_before
);
532 /* The instruction immediately before the JUMP_INSN may not be the
533 compare, so we must copy it. */
534 insert_before
= last_loop_insn
;
535 copy_end
= PREV_INSN (last_loop_insn
);
540 /* We currently can't unroll a loop if it doesn't end with a
541 JUMP_INSN. There would need to be a mechanism that recognizes
542 this case, and then inserts a jump after each loop body, which
543 jumps to after the last loop body. */
544 if (loop_dump_stream
)
545 fprintf (loop_dump_stream
,
546 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
552 /* Normal case: Must copy the compare and branch instructions at the
555 if (GET_CODE (last_loop_insn
) == BARRIER
)
557 /* Loop ends with an unconditional jump and a barrier.
558 Handle this like above, don't copy jump and barrier.
559 This is not strictly necessary, but doing so prevents generating
560 unconditional jumps to an immediately following label.
562 This will be corrected below if the target of this jump is
563 not the start_label. */
565 insert_before
= PREV_INSN (last_loop_insn
);
566 copy_end
= PREV_INSN (insert_before
);
568 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
570 /* Set insert_before to immediately after the JUMP_INSN, so that
571 NOTEs at the end of the loop will be correctly handled by
573 insert_before
= NEXT_INSN (last_loop_insn
);
574 copy_end
= last_loop_insn
;
578 /* We currently can't unroll a loop if it doesn't end with a
579 JUMP_INSN. There would need to be a mechanism that recognizes
580 this case, and then inserts a jump after each loop body, which
581 jumps to after the last loop body. */
582 if (loop_dump_stream
)
583 fprintf (loop_dump_stream
,
584 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
588 /* If copying exit test branches because they can not be eliminated,
589 then must convert the fall through case of the branch to a jump past
590 the end of the loop. Create a label to emit after the loop and save
591 it for later use. Do not use the label after the loop, if any, since
592 it might be used by insns outside the loop, or there might be insns
593 added before it later by final_[bg]iv_value which must be after
594 the real exit label. */
595 exit_label
= gen_label_rtx ();
598 while (GET_CODE (insn
) != CODE_LABEL
&& GET_CODE (insn
) != JUMP_INSN
)
599 insn
= NEXT_INSN (insn
);
601 if (GET_CODE (insn
) == JUMP_INSN
)
603 /* The loop starts with a jump down to the exit condition test.
604 Start copying the loop after the barrier following this
606 copy_start
= NEXT_INSN (insn
);
608 /* Splitting induction variables doesn't work when the loop is
609 entered via a jump to the bottom, because then we end up doing
610 a comparison against a new register for a split variable, but
611 we did not execute the set insn for the new register because
612 it was skipped over. */
613 splitting_not_safe
= 1;
614 if (loop_dump_stream
)
615 fprintf (loop_dump_stream
,
616 "Splitting not safe, because loop not entered at top.\n");
619 copy_start
= loop_start
;
622 /* This should always be the first label in the loop. */
623 start_label
= NEXT_INSN (copy_start
);
624 /* There may be a line number note and/or a loop continue note here. */
625 while (GET_CODE (start_label
) == NOTE
)
626 start_label
= NEXT_INSN (start_label
);
627 if (GET_CODE (start_label
) != CODE_LABEL
)
629 /* This can happen as a result of jump threading. If the first insns in
630 the loop test the same condition as the loop's backward jump, or the
631 opposite condition, then the backward jump will be modified to point
632 to elsewhere, and the loop's start label is deleted.
634 This case currently can not be handled by the loop unrolling code. */
636 if (loop_dump_stream
)
637 fprintf (loop_dump_stream
,
638 "Unrolling failure: unknown insns between BEG note and loop label.\n");
641 if (LABEL_NAME (start_label
))
643 /* The jump optimization pass must have combined the original start label
644 with a named label for a goto. We can't unroll this case because
645 jumps which go to the named label must be handled differently than
646 jumps to the loop start, and it is impossible to differentiate them
648 if (loop_dump_stream
)
649 fprintf (loop_dump_stream
,
650 "Unrolling failure: loop start label is gone\n");
654 if (unroll_type
== UNROLL_NAIVE
655 && GET_CODE (last_loop_insn
) == BARRIER
656 && start_label
!= JUMP_LABEL (PREV_INSN (last_loop_insn
)))
658 /* In this case, we must copy the jump and barrier, because they will
659 not be converted to jumps to an immediately following label. */
661 insert_before
= NEXT_INSN (last_loop_insn
);
662 copy_end
= last_loop_insn
;
665 if (unroll_type
== UNROLL_NAIVE
666 && GET_CODE (last_loop_insn
) == JUMP_INSN
667 && start_label
!= JUMP_LABEL (last_loop_insn
))
669 /* ??? The loop ends with a conditional branch that does not branch back
670 to the loop start label. In this case, we must emit an unconditional
671 branch to the loop exit after emitting the final branch.
672 copy_loop_body does not have support for this currently, so we
673 give up. It doesn't seem worthwhile to unroll anyways since
674 unrolling would increase the number of branch instructions
676 if (loop_dump_stream
)
677 fprintf (loop_dump_stream
,
678 "Unrolling failure: final conditional branch not to loop start\n");
682 /* Allocate a translation table for the labels and insn numbers.
683 They will be filled in as we copy the insns in the loop. */
685 max_labelno
= max_label_num ();
686 max_insnno
= get_max_uid ();
688 map
= (struct inline_remap
*) alloca (sizeof (struct inline_remap
));
690 map
->integrating
= 0;
692 /* Allocate the label map. */
696 map
->label_map
= (rtx
*) alloca (max_labelno
* sizeof (rtx
));
698 local_label
= (char *) alloca (max_labelno
);
699 bzero (local_label
, max_labelno
);
704 /* Search the loop and mark all local labels, i.e. the ones which have to
705 be distinct labels when copied. For all labels which might be
706 non-local, set their label_map entries to point to themselves.
707 If they happen to be local their label_map entries will be overwritten
708 before the loop body is copied. The label_map entries for local labels
709 will be set to a different value each time the loop body is copied. */
711 for (insn
= copy_start
; insn
!= loop_end
; insn
= NEXT_INSN (insn
))
715 if (GET_CODE (insn
) == CODE_LABEL
)
716 local_label
[CODE_LABEL_NUMBER (insn
)] = 1;
717 else if (GET_CODE (insn
) == JUMP_INSN
)
719 if (JUMP_LABEL (insn
))
720 set_label_in_map (map
,
721 CODE_LABEL_NUMBER (JUMP_LABEL (insn
)),
723 else if (GET_CODE (PATTERN (insn
)) == ADDR_VEC
724 || GET_CODE (PATTERN (insn
)) == ADDR_DIFF_VEC
)
726 rtx pat
= PATTERN (insn
);
727 int diff_vec_p
= GET_CODE (PATTERN (insn
)) == ADDR_DIFF_VEC
;
728 int len
= XVECLEN (pat
, diff_vec_p
);
731 for (i
= 0; i
< len
; i
++)
733 label
= XEXP (XVECEXP (pat
, diff_vec_p
, i
), 0);
734 set_label_in_map (map
,
735 CODE_LABEL_NUMBER (label
),
740 else if ((note
= find_reg_note (insn
, REG_LABEL
, NULL_RTX
)))
741 set_label_in_map (map
, CODE_LABEL_NUMBER (XEXP (note
, 0)),
745 /* Allocate space for the insn map. */
747 map
->insn_map
= (rtx
*) alloca (max_insnno
* sizeof (rtx
));
749 /* Set this to zero, to indicate that we are doing loop unrolling,
750 not function inlining. */
751 map
->inline_target
= 0;
753 /* The register and constant maps depend on the number of registers
754 present, so the final maps can't be created until after
755 find_splittable_regs is called. However, they are needed for
756 preconditioning, so we create temporary maps when preconditioning
759 /* The preconditioning code may allocate two new pseudo registers. */
760 maxregnum
= max_reg_num ();
762 /* Allocate and zero out the splittable_regs and addr_combined_regs
763 arrays. These must be zeroed here because they will be used if
764 loop preconditioning is performed, and must be zero for that case.
766 It is safe to do this here, since the extra registers created by the
767 preconditioning code and find_splittable_regs will never be used
768 to access the splittable_regs[] and addr_combined_regs[] arrays. */
770 splittable_regs
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
771 bzero ((char *) splittable_regs
, maxregnum
* sizeof (rtx
));
772 splittable_regs_updates
= (int *) alloca (maxregnum
* sizeof (int));
773 bzero ((char *) splittable_regs_updates
, maxregnum
* sizeof (int));
775 = (struct induction
**) alloca (maxregnum
* sizeof (struct induction
*));
776 bzero ((char *) addr_combined_regs
, maxregnum
* sizeof (struct induction
*));
777 /* We must limit it to max_reg_before_loop, because only these pseudo
778 registers have valid regno_first_uid info. Any register created after
779 that is unlikely to be local to the loop anyways. */
780 local_regno
= (char *) alloca (max_reg_before_loop
);
781 bzero (local_regno
, max_reg_before_loop
);
783 /* Mark all local registers, i.e. the ones which are referenced only
785 if (INSN_UID (copy_end
) < max_uid_for_loop
)
787 int copy_start_luid
= INSN_LUID (copy_start
);
788 int copy_end_luid
= INSN_LUID (copy_end
);
790 /* If a register is used in the jump insn, we must not duplicate it
791 since it will also be used outside the loop. */
792 if (GET_CODE (copy_end
) == JUMP_INSN
)
794 /* If copy_start points to the NOTE that starts the loop, then we must
795 use the next luid, because invariant pseudo-regs moved out of the loop
796 have their lifetimes modified to start here, but they are not safe
798 if (copy_start
== loop_start
)
801 /* If a pseudo's lifetime is entirely contained within this loop, then we
802 can use a different pseudo in each unrolled copy of the loop. This
803 results in better code. */
804 for (j
= FIRST_PSEUDO_REGISTER
; j
< max_reg_before_loop
; ++j
)
805 if (REGNO_FIRST_UID (j
) > 0 && REGNO_FIRST_UID (j
) <= max_uid_for_loop
806 && uid_luid
[REGNO_FIRST_UID (j
)] >= copy_start_luid
807 && REGNO_LAST_UID (j
) > 0 && REGNO_LAST_UID (j
) <= max_uid_for_loop
808 && uid_luid
[REGNO_LAST_UID (j
)] <= copy_end_luid
)
810 /* However, we must also check for loop-carried dependencies.
811 If the value the pseudo has at the end of iteration X is
812 used by iteration X+1, then we can not use a different pseudo
813 for each unrolled copy of the loop. */
814 /* A pseudo is safe if regno_first_uid is a set, and this
815 set dominates all instructions from regno_first_uid to
817 /* ??? This check is simplistic. We would get better code if
818 this check was more sophisticated. */
819 if (set_dominates_use (j
, REGNO_FIRST_UID (j
), REGNO_LAST_UID (j
),
820 copy_start
, copy_end
))
823 if (loop_dump_stream
)
826 fprintf (loop_dump_stream
, "Marked reg %d as local\n", j
);
828 fprintf (loop_dump_stream
, "Did not mark reg %d as local\n",
834 /* If this loop requires exit tests when unrolled, check to see if we
835 can precondition the loop so as to make the exit tests unnecessary.
836 Just like variable splitting, this is not safe if the loop is entered
837 via a jump to the bottom. Also, can not do this if no strength
838 reduce info, because precondition_loop_p uses this info. */
840 /* Must copy the loop body for preconditioning before the following
841 find_splittable_regs call since that will emit insns which need to
842 be after the preconditioned loop copies, but immediately before the
843 unrolled loop copies. */
845 /* Also, it is not safe to split induction variables for the preconditioned
846 copies of the loop body. If we split induction variables, then the code
847 assumes that each induction variable can be represented as a function
848 of its initial value and the loop iteration number. This is not true
849 in this case, because the last preconditioned copy of the loop body
850 could be any iteration from the first up to the `unroll_number-1'th,
851 depending on the initial value of the iteration variable. Therefore
852 we can not split induction variables here, because we can not calculate
853 their value. Hence, this code must occur before find_splittable_regs
856 if (unroll_type
== UNROLL_NAIVE
&& ! splitting_not_safe
&& strength_reduce_p
)
858 rtx initial_value
, final_value
, increment
;
860 if (precondition_loop_p (&initial_value
, &final_value
, &increment
,
861 loop_start
, loop_end
))
864 enum machine_mode mode
;
866 int abs_inc
, neg_inc
;
868 map
->reg_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
870 map
->const_equiv_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
871 map
->const_age_map
= (unsigned *) alloca (maxregnum
872 * sizeof (unsigned));
873 map
->const_equiv_map_size
= maxregnum
;
874 global_const_equiv_map
= map
->const_equiv_map
;
875 global_const_equiv_map_size
= maxregnum
;
877 init_reg_map (map
, maxregnum
);
879 /* Limit loop unrolling to 4, since this will make 7 copies of
881 if (unroll_number
> 4)
884 /* Save the absolute value of the increment, and also whether or
885 not it is negative. */
887 abs_inc
= INTVAL (increment
);
896 /* Decide what mode to do these calculations in. Choose the larger
897 of final_value's mode and initial_value's mode, or a full-word if
898 both are constants. */
899 mode
= GET_MODE (final_value
);
900 if (mode
== VOIDmode
)
902 mode
= GET_MODE (initial_value
);
903 if (mode
== VOIDmode
)
906 else if (mode
!= GET_MODE (initial_value
)
907 && (GET_MODE_SIZE (mode
)
908 < GET_MODE_SIZE (GET_MODE (initial_value
))))
909 mode
= GET_MODE (initial_value
);
911 /* Calculate the difference between the final and initial values.
912 Final value may be a (plus (reg x) (const_int 1)) rtx.
913 Let the following cse pass simplify this if initial value is
916 We must copy the final and initial values here to avoid
917 improperly shared rtl. */
919 diff
= expand_binop (mode
, sub_optab
, copy_rtx (final_value
),
920 copy_rtx (initial_value
), NULL_RTX
, 0,
923 /* Now calculate (diff % (unroll * abs (increment))) by using an
925 diff
= expand_binop (GET_MODE (diff
), and_optab
, diff
,
926 GEN_INT (unroll_number
* abs_inc
- 1),
927 NULL_RTX
, 0, OPTAB_LIB_WIDEN
);
929 /* Now emit a sequence of branches to jump to the proper precond
932 labels
= (rtx
*) alloca (sizeof (rtx
) * unroll_number
);
933 for (i
= 0; i
< unroll_number
; i
++)
934 labels
[i
] = gen_label_rtx ();
936 /* Check for the case where the initial value is greater than or
937 equal to the final value. In that case, we want to execute
938 exactly one loop iteration. The code below will fail for this
939 case. This check does not apply if the loop has a NE
940 comparison at the end. */
942 if (loop_comparison_code
!= NE
)
944 emit_cmp_insn (initial_value
, final_value
, neg_inc
? LE
: GE
,
945 NULL_RTX
, mode
, 0, 0);
947 emit_jump_insn (gen_ble (labels
[1]));
949 emit_jump_insn (gen_bge (labels
[1]));
950 JUMP_LABEL (get_last_insn ()) = labels
[1];
951 LABEL_NUSES (labels
[1])++;
954 /* Assuming the unroll_number is 4, and the increment is 2, then
955 for a negative increment: for a positive increment:
956 diff = 0,1 precond 0 diff = 0,7 precond 0
957 diff = 2,3 precond 3 diff = 1,2 precond 1
958 diff = 4,5 precond 2 diff = 3,4 precond 2
959 diff = 6,7 precond 1 diff = 5,6 precond 3 */
961 /* We only need to emit (unroll_number - 1) branches here, the
962 last case just falls through to the following code. */
964 /* ??? This would give better code if we emitted a tree of branches
965 instead of the current linear list of branches. */
967 for (i
= 0; i
< unroll_number
- 1; i
++)
970 enum rtx_code cmp_code
;
972 /* For negative increments, must invert the constant compared
973 against, except when comparing against zero. */
981 cmp_const
= unroll_number
- i
;
990 emit_cmp_insn (diff
, GEN_INT (abs_inc
* cmp_const
),
991 cmp_code
, NULL_RTX
, mode
, 0, 0);
994 emit_jump_insn (gen_beq (labels
[i
]));
996 emit_jump_insn (gen_bge (labels
[i
]));
998 emit_jump_insn (gen_ble (labels
[i
]));
999 JUMP_LABEL (get_last_insn ()) = labels
[i
];
1000 LABEL_NUSES (labels
[i
])++;
1003 /* If the increment is greater than one, then we need another branch,
1004 to handle other cases equivalent to 0. */
1006 /* ??? This should be merged into the code above somehow to help
1007 simplify the code here, and reduce the number of branches emitted.
1008 For the negative increment case, the branch here could easily
1009 be merged with the `0' case branch above. For the positive
1010 increment case, it is not clear how this can be simplified. */
1015 enum rtx_code cmp_code
;
1019 cmp_const
= abs_inc
- 1;
1024 cmp_const
= abs_inc
* (unroll_number
- 1) + 1;
1028 emit_cmp_insn (diff
, GEN_INT (cmp_const
), cmp_code
, NULL_RTX
,
1032 emit_jump_insn (gen_ble (labels
[0]));
1034 emit_jump_insn (gen_bge (labels
[0]));
1035 JUMP_LABEL (get_last_insn ()) = labels
[0];
1036 LABEL_NUSES (labels
[0])++;
1039 sequence
= gen_sequence ();
1041 emit_insn_before (sequence
, loop_start
);
1043 /* Only the last copy of the loop body here needs the exit
1044 test, so set copy_end to exclude the compare/branch here,
1045 and then reset it inside the loop when get to the last
1048 if (GET_CODE (last_loop_insn
) == BARRIER
)
1049 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
1050 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
1053 /* The immediately preceding insn is a compare which we do not
1055 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
1057 /* The immediately preceding insn may not be a compare, so we
1059 copy_end
= PREV_INSN (last_loop_insn
);
1065 for (i
= 1; i
< unroll_number
; i
++)
1067 emit_label_after (labels
[unroll_number
- i
],
1068 PREV_INSN (loop_start
));
1070 bzero ((char *) map
->insn_map
, max_insnno
* sizeof (rtx
));
1071 bzero ((char *) map
->const_equiv_map
, maxregnum
* sizeof (rtx
));
1072 bzero ((char *) map
->const_age_map
,
1073 maxregnum
* sizeof (unsigned));
1076 for (j
= 0; j
< max_labelno
; j
++)
1078 set_label_in_map (map
, j
, gen_label_rtx ());
1080 for (j
= FIRST_PSEUDO_REGISTER
; j
< max_reg_before_loop
; j
++)
1083 map
->reg_map
[j
] = gen_reg_rtx (GET_MODE (regno_reg_rtx
[j
]));
1084 record_base_value (REGNO (map
->reg_map
[j
]),
1085 regno_reg_rtx
[j
], 0);
1087 /* The last copy needs the compare/branch insns at the end,
1088 so reset copy_end here if the loop ends with a conditional
1091 if (i
== unroll_number
- 1)
1093 if (GET_CODE (last_loop_insn
) == BARRIER
)
1094 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
1096 copy_end
= last_loop_insn
;
1099 /* None of the copies are the `last_iteration', so just
1100 pass zero for that parameter. */
1101 copy_loop_body (copy_start
, copy_end
, map
, exit_label
, 0,
1102 unroll_type
, start_label
, loop_end
,
1103 loop_start
, copy_end
);
1105 emit_label_after (labels
[0], PREV_INSN (loop_start
));
1107 if (GET_CODE (last_loop_insn
) == BARRIER
)
1109 insert_before
= PREV_INSN (last_loop_insn
);
1110 copy_end
= PREV_INSN (insert_before
);
1115 /* The immediately preceding insn is a compare which we do not
1117 insert_before
= PREV_INSN (last_loop_insn
);
1118 copy_end
= PREV_INSN (insert_before
);
1120 /* The immediately preceding insn may not be a compare, so we
1122 insert_before
= last_loop_insn
;
1123 copy_end
= PREV_INSN (last_loop_insn
);
1127 /* Set unroll type to MODULO now. */
1128 unroll_type
= UNROLL_MODULO
;
1129 loop_preconditioned
= 1;
1132 /* Fix the initial value for the loop as needed. */
1133 if (loop_n_iterations
<= 0)
1134 loop_start_value
[uid_loop_num
[INSN_UID (loop_start
)]]
1140 /* If reach here, and the loop type is UNROLL_NAIVE, then don't unroll
1141 the loop unless all loops are being unrolled. */
1142 if (unroll_type
== UNROLL_NAIVE
&& ! flag_unroll_all_loops
)
1144 if (loop_dump_stream
)
1145 fprintf (loop_dump_stream
, "Unrolling failure: Naive unrolling not being done.\n");
1149 /* At this point, we are guaranteed to unroll the loop. */
1151 /* Keep track of the unroll factor for each loop. */
1152 if (unroll_type
== UNROLL_COMPLETELY
)
1153 loop_unroll_factor
[uid_loop_num
[INSN_UID (loop_start
)]] = -1;
1155 loop_unroll_factor
[uid_loop_num
[INSN_UID (loop_start
)]] = unroll_number
;
1158 /* For each biv and giv, determine whether it can be safely split into
1159 a different variable for each unrolled copy of the loop body.
1160 We precalculate and save this info here, since computing it is
1163 Do this before deleting any instructions from the loop, so that
1164 back_branch_in_range_p will work correctly. */
1166 if (splitting_not_safe
)
1169 temp
= find_splittable_regs (unroll_type
, loop_start
, loop_end
,
1170 end_insert_before
, unroll_number
);
1172 /* find_splittable_regs may have created some new registers, so must
1173 reallocate the reg_map with the new larger size, and must realloc
1174 the constant maps also. */
1176 maxregnum
= max_reg_num ();
1177 map
->reg_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
1179 init_reg_map (map
, maxregnum
);
1181 /* Space is needed in some of the map for new registers, so new_maxregnum
1182 is an (over)estimate of how many registers will exist at the end. */
1183 new_maxregnum
= maxregnum
+ (temp
* unroll_number
* 2);
1185 /* Must realloc space for the constant maps, because the number of registers
1186 may have changed. */
1188 map
->const_equiv_map
= (rtx
*) alloca (new_maxregnum
* sizeof (rtx
));
1189 map
->const_age_map
= (unsigned *) alloca (new_maxregnum
* sizeof (unsigned));
1191 map
->const_equiv_map_size
= new_maxregnum
;
1192 global_const_equiv_map
= map
->const_equiv_map
;
1193 global_const_equiv_map_size
= new_maxregnum
;
1195 /* Search the list of bivs and givs to find ones which need to be remapped
1196 when split, and set their reg_map entry appropriately. */
1198 for (bl
= loop_iv_list
; bl
; bl
= bl
->next
)
1200 if (REGNO (bl
->biv
->src_reg
) != bl
->regno
)
1201 map
->reg_map
[bl
->regno
] = bl
->biv
->src_reg
;
1203 /* Currently, non-reduced/final-value givs are never split. */
1204 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
1205 if (REGNO (v
->src_reg
) != bl
->regno
)
1206 map
->reg_map
[REGNO (v
->dest_reg
)] = v
->src_reg
;
1210 /* Use our current register alignment and pointer flags. */
1211 map
->regno_pointer_flag
= regno_pointer_flag
;
1212 map
->regno_pointer_align
= regno_pointer_align
;
1214 /* If the loop is being partially unrolled, and the iteration variables
1215 are being split, and are being renamed for the split, then must fix up
1216 the compare/jump instruction at the end of the loop to refer to the new
1217 registers. This compare isn't copied, so the registers used in it
1218 will never be replaced if it isn't done here. */
1220 if (unroll_type
== UNROLL_MODULO
)
1222 insn
= NEXT_INSN (copy_end
);
1223 if (GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == JUMP_INSN
)
1224 PATTERN (insn
) = remap_split_bivs (PATTERN (insn
));
1227 /* For unroll_number - 1 times, make a copy of each instruction
1228 between copy_start and copy_end, and insert these new instructions
1229 before the end of the loop. */
1231 for (i
= 0; i
< unroll_number
; i
++)
1233 bzero ((char *) map
->insn_map
, max_insnno
* sizeof (rtx
));
1234 bzero ((char *) map
->const_equiv_map
, new_maxregnum
* sizeof (rtx
));
1235 bzero ((char *) map
->const_age_map
, new_maxregnum
* sizeof (unsigned));
1238 for (j
= 0; j
< max_labelno
; j
++)
1240 set_label_in_map (map
, j
, gen_label_rtx ());
1242 for (j
= FIRST_PSEUDO_REGISTER
; j
< max_reg_before_loop
; j
++)
1245 map
->reg_map
[j
] = gen_reg_rtx (GET_MODE (regno_reg_rtx
[j
]));
1246 record_base_value (REGNO (map
->reg_map
[j
]),
1247 regno_reg_rtx
[j
], 0);
1250 /* If loop starts with a branch to the test, then fix it so that
1251 it points to the test of the first unrolled copy of the loop. */
1252 if (i
== 0 && loop_start
!= copy_start
)
1254 insn
= PREV_INSN (copy_start
);
1255 pattern
= PATTERN (insn
);
1257 tem
= get_label_from_map (map
,
1259 (XEXP (SET_SRC (pattern
), 0)));
1260 SET_SRC (pattern
) = gen_rtx_LABEL_REF (VOIDmode
, tem
);
1262 /* Set the jump label so that it can be used by later loop unrolling
1264 JUMP_LABEL (insn
) = tem
;
1265 LABEL_NUSES (tem
)++;
1268 copy_loop_body (copy_start
, copy_end
, map
, exit_label
,
1269 i
== unroll_number
- 1, unroll_type
, start_label
,
1270 loop_end
, insert_before
, insert_before
);
1273 /* Before deleting any insns, emit a CODE_LABEL immediately after the last
1274 insn to be deleted. This prevents any runaway delete_insn call from
1275 more insns that it should, as it always stops at a CODE_LABEL. */
1277 /* Delete the compare and branch at the end of the loop if completely
1278 unrolling the loop. Deleting the backward branch at the end also
1279 deletes the code label at the start of the loop. This is done at
1280 the very end to avoid problems with back_branch_in_range_p. */
1282 if (unroll_type
== UNROLL_COMPLETELY
)
1283 safety_label
= emit_label_after (gen_label_rtx (), last_loop_insn
);
1285 safety_label
= emit_label_after (gen_label_rtx (), copy_end
);
1287 /* Delete all of the original loop instructions. Don't delete the
1288 LOOP_BEG note, or the first code label in the loop. */
1290 insn
= NEXT_INSN (copy_start
);
1291 while (insn
!= safety_label
)
1293 if (insn
!= start_label
)
1294 insn
= delete_insn (insn
);
1296 insn
= NEXT_INSN (insn
);
1299 /* Can now delete the 'safety' label emitted to protect us from runaway
1300 delete_insn calls. */
1301 if (INSN_DELETED_P (safety_label
))
1303 delete_insn (safety_label
);
1305 /* If exit_label exists, emit it after the loop. Doing the emit here
1306 forces it to have a higher INSN_UID than any insn in the unrolled loop.
1307 This is needed so that mostly_true_jump in reorg.c will treat jumps
1308 to this loop end label correctly, i.e. predict that they are usually
1311 emit_label_after (exit_label
, loop_end
);
1314 /* Return true if the loop can be safely, and profitably, preconditioned
1315 so that the unrolled copies of the loop body don't need exit tests.
1317 This only works if final_value, initial_value and increment can be
1318 determined, and if increment is a constant power of 2.
1319 If increment is not a power of 2, then the preconditioning modulo
1320 operation would require a real modulo instead of a boolean AND, and this
1321 is not considered `profitable'. */
1323 /* ??? If the loop is known to be executed very many times, or the machine
1324 has a very cheap divide instruction, then preconditioning is a win even
1325 when the increment is not a power of 2. Use RTX_COST to compute
1326 whether divide is cheap. */
1329 precondition_loop_p (initial_value
, final_value
, increment
, loop_start
,
1331 rtx
*initial_value
, *final_value
, *increment
;
1332 rtx loop_start
, loop_end
;
1335 if (loop_n_iterations
> 0)
1337 *initial_value
= const0_rtx
;
1338 *increment
= const1_rtx
;
1339 *final_value
= GEN_INT (loop_n_iterations
);
1341 if (loop_dump_stream
)
1343 fputs ("Preconditioning: Success, number of iterations known, ",
1345 fprintf (loop_dump_stream
, HOST_WIDE_INT_PRINT_DEC
,
1347 fputs (".\n", loop_dump_stream
);
1352 if (loop_initial_value
== 0)
1354 if (loop_dump_stream
)
1355 fprintf (loop_dump_stream
,
1356 "Preconditioning: Could not find initial value.\n");
1359 else if (loop_increment
== 0)
1361 if (loop_dump_stream
)
1362 fprintf (loop_dump_stream
,
1363 "Preconditioning: Could not find increment value.\n");
1366 else if (GET_CODE (loop_increment
) != CONST_INT
)
1368 if (loop_dump_stream
)
1369 fprintf (loop_dump_stream
,
1370 "Preconditioning: Increment not a constant.\n");
1373 else if ((exact_log2 (INTVAL (loop_increment
)) < 0)
1374 && (exact_log2 (- INTVAL (loop_increment
)) < 0))
1376 if (loop_dump_stream
)
1377 fprintf (loop_dump_stream
,
1378 "Preconditioning: Increment not a constant power of 2.\n");
1382 /* Unsigned_compare and compare_dir can be ignored here, since they do
1383 not matter for preconditioning. */
1385 if (loop_final_value
== 0)
1387 if (loop_dump_stream
)
1388 fprintf (loop_dump_stream
,
1389 "Preconditioning: EQ comparison loop.\n");
1393 /* Must ensure that final_value is invariant, so call invariant_p to
1394 check. Before doing so, must check regno against max_reg_before_loop
1395 to make sure that the register is in the range covered by invariant_p.
1396 If it isn't, then it is most likely a biv/giv which by definition are
1398 if ((GET_CODE (loop_final_value
) == REG
1399 && REGNO (loop_final_value
) >= max_reg_before_loop
)
1400 || (GET_CODE (loop_final_value
) == PLUS
1401 && REGNO (XEXP (loop_final_value
, 0)) >= max_reg_before_loop
)
1402 || ! invariant_p (loop_final_value
))
1404 if (loop_dump_stream
)
1405 fprintf (loop_dump_stream
,
1406 "Preconditioning: Final value not invariant.\n");
1410 /* Fail for floating point values, since the caller of this function
1411 does not have code to deal with them. */
1412 if (GET_MODE_CLASS (GET_MODE (loop_final_value
)) == MODE_FLOAT
1413 || GET_MODE_CLASS (GET_MODE (loop_initial_value
)) == MODE_FLOAT
)
1415 if (loop_dump_stream
)
1416 fprintf (loop_dump_stream
,
1417 "Preconditioning: Floating point final or initial value.\n");
1421 /* Now set initial_value to be the iteration_var, since that may be a
1422 simpler expression, and is guaranteed to be correct if all of the
1423 above tests succeed.
1425 We can not use the initial_value as calculated, because it will be
1426 one too small for loops of the form "while (i-- > 0)". We can not
1427 emit code before the loop_skip_over insns to fix this problem as this
1428 will then give a number one too large for loops of the form
1431 Note that all loops that reach here are entered at the top, because
1432 this function is not called if the loop starts with a jump. */
1434 /* Fail if loop_iteration_var is not live before loop_start, since we need
1435 to test its value in the preconditioning code. */
1437 if (uid_luid
[REGNO_FIRST_UID (REGNO (loop_iteration_var
))]
1438 > INSN_LUID (loop_start
))
1440 if (loop_dump_stream
)
1441 fprintf (loop_dump_stream
,
1442 "Preconditioning: Iteration var not live before loop start.\n");
1446 *initial_value
= loop_iteration_var
;
1447 *increment
= loop_increment
;
1448 *final_value
= loop_final_value
;
1451 if (loop_dump_stream
)
1452 fprintf (loop_dump_stream
, "Preconditioning: Successful.\n");
1457 /* All pseudo-registers must be mapped to themselves. Two hard registers
1458 must be mapped, VIRTUAL_STACK_VARS_REGNUM and VIRTUAL_INCOMING_ARGS_
1459 REGNUM, to avoid function-inlining specific conversions of these
1460 registers. All other hard regs can not be mapped because they may be
1465 init_reg_map (map
, maxregnum
)
1466 struct inline_remap
*map
;
1471 for (i
= maxregnum
- 1; i
> LAST_VIRTUAL_REGISTER
; i
--)
1472 map
->reg_map
[i
] = regno_reg_rtx
[i
];
1473 /* Just clear the rest of the entries. */
1474 for (i
= LAST_VIRTUAL_REGISTER
; i
>= 0; i
--)
1475 map
->reg_map
[i
] = 0;
1477 map
->reg_map
[VIRTUAL_STACK_VARS_REGNUM
]
1478 = regno_reg_rtx
[VIRTUAL_STACK_VARS_REGNUM
];
1479 map
->reg_map
[VIRTUAL_INCOMING_ARGS_REGNUM
]
1480 = regno_reg_rtx
[VIRTUAL_INCOMING_ARGS_REGNUM
];
1483 /* Strength-reduction will often emit code for optimized biv/givs which
1484 calculates their value in a temporary register, and then copies the result
1485 to the iv. This procedure reconstructs the pattern computing the iv;
1486 verifying that all operands are of the proper form.
1488 PATTERN must be the result of single_set.
1489 The return value is the amount that the giv is incremented by. */
1492 calculate_giv_inc (pattern
, src_insn
, regno
)
1493 rtx pattern
, src_insn
;
1497 rtx increment_total
= 0;
1501 /* Verify that we have an increment insn here. First check for a plus
1502 as the set source. */
1503 if (GET_CODE (SET_SRC (pattern
)) != PLUS
)
1505 /* SR sometimes computes the new giv value in a temp, then copies it
1507 src_insn
= PREV_INSN (src_insn
);
1508 pattern
= PATTERN (src_insn
);
1509 if (GET_CODE (SET_SRC (pattern
)) != PLUS
)
1512 /* The last insn emitted is not needed, so delete it to avoid confusing
1513 the second cse pass. This insn sets the giv unnecessarily. */
1514 delete_insn (get_last_insn ());
1517 /* Verify that we have a constant as the second operand of the plus. */
1518 increment
= XEXP (SET_SRC (pattern
), 1);
1519 if (GET_CODE (increment
) != CONST_INT
)
1521 /* SR sometimes puts the constant in a register, especially if it is
1522 too big to be an add immed operand. */
1523 src_insn
= PREV_INSN (src_insn
);
1524 increment
= SET_SRC (PATTERN (src_insn
));
1526 /* SR may have used LO_SUM to compute the constant if it is too large
1527 for a load immed operand. In this case, the constant is in operand
1528 one of the LO_SUM rtx. */
1529 if (GET_CODE (increment
) == LO_SUM
)
1530 increment
= XEXP (increment
, 1);
1532 /* Some ports store large constants in memory and add a REG_EQUAL
1533 note to the store insn. */
1534 else if (GET_CODE (increment
) == MEM
)
1536 rtx note
= find_reg_note (src_insn
, REG_EQUAL
, 0);
1538 increment
= XEXP (note
, 0);
1541 else if (GET_CODE (increment
) == IOR
1542 || GET_CODE (increment
) == ASHIFT
1543 || GET_CODE (increment
) == PLUS
)
1545 /* The rs6000 port loads some constants with IOR.
1546 The alpha port loads some constants with ASHIFT and PLUS. */
1547 rtx second_part
= XEXP (increment
, 1);
1548 enum rtx_code code
= GET_CODE (increment
);
1550 src_insn
= PREV_INSN (src_insn
);
1551 increment
= SET_SRC (PATTERN (src_insn
));
1552 /* Don't need the last insn anymore. */
1553 delete_insn (get_last_insn ());
1555 if (GET_CODE (second_part
) != CONST_INT
1556 || GET_CODE (increment
) != CONST_INT
)
1560 increment
= GEN_INT (INTVAL (increment
) | INTVAL (second_part
));
1561 else if (code
== PLUS
)
1562 increment
= GEN_INT (INTVAL (increment
) + INTVAL (second_part
));
1564 increment
= GEN_INT (INTVAL (increment
) << INTVAL (second_part
));
1567 if (GET_CODE (increment
) != CONST_INT
)
1570 /* The insn loading the constant into a register is no longer needed,
1572 delete_insn (get_last_insn ());
1575 if (increment_total
)
1576 increment_total
= GEN_INT (INTVAL (increment_total
) + INTVAL (increment
));
1578 increment_total
= increment
;
1580 /* Check that the source register is the same as the register we expected
1581 to see as the source. If not, something is seriously wrong. */
1582 if (GET_CODE (XEXP (SET_SRC (pattern
), 0)) != REG
1583 || REGNO (XEXP (SET_SRC (pattern
), 0)) != regno
)
1585 /* Some machines (e.g. the romp), may emit two add instructions for
1586 certain constants, so lets try looking for another add immediately
1587 before this one if we have only seen one add insn so far. */
1593 src_insn
= PREV_INSN (src_insn
);
1594 pattern
= PATTERN (src_insn
);
1596 delete_insn (get_last_insn ());
1604 return increment_total
;
1607 /* Copy REG_NOTES, except for insn references, because not all insn_map
1608 entries are valid yet. We do need to copy registers now though, because
1609 the reg_map entries can change during copying. */
1612 initial_reg_note_copy (notes
, map
)
1614 struct inline_remap
*map
;
1621 copy
= rtx_alloc (GET_CODE (notes
));
1622 PUT_MODE (copy
, GET_MODE (notes
));
1624 if (GET_CODE (notes
) == EXPR_LIST
)
1625 XEXP (copy
, 0) = copy_rtx_and_substitute (XEXP (notes
, 0), map
);
1626 else if (GET_CODE (notes
) == INSN_LIST
)
1627 /* Don't substitute for these yet. */
1628 XEXP (copy
, 0) = XEXP (notes
, 0);
1632 XEXP (copy
, 1) = initial_reg_note_copy (XEXP (notes
, 1), map
);
1637 /* Fixup insn references in copied REG_NOTES. */
1640 final_reg_note_copy (notes
, map
)
1642 struct inline_remap
*map
;
1646 for (note
= notes
; note
; note
= XEXP (note
, 1))
1647 if (GET_CODE (note
) == INSN_LIST
)
1648 XEXP (note
, 0) = map
->insn_map
[INSN_UID (XEXP (note
, 0))];
1651 /* Copy each instruction in the loop, substituting from map as appropriate.
1652 This is very similar to a loop in expand_inline_function. */
1655 copy_loop_body (copy_start
, copy_end
, map
, exit_label
, last_iteration
,
1656 unroll_type
, start_label
, loop_end
, insert_before
,
1658 rtx copy_start
, copy_end
;
1659 struct inline_remap
*map
;
1662 enum unroll_types unroll_type
;
1663 rtx start_label
, loop_end
, insert_before
, copy_notes_from
;
1667 int dest_reg_was_split
, i
;
1671 rtx final_label
= 0;
1672 rtx giv_inc
, giv_dest_reg
, giv_src_reg
;
1674 /* If this isn't the last iteration, then map any references to the
1675 start_label to final_label. Final label will then be emitted immediately
1676 after the end of this loop body if it was ever used.
1678 If this is the last iteration, then map references to the start_label
1680 if (! last_iteration
)
1682 final_label
= gen_label_rtx ();
1683 set_label_in_map (map
, CODE_LABEL_NUMBER (start_label
),
1687 set_label_in_map (map
, CODE_LABEL_NUMBER (start_label
), start_label
);
1694 insn
= NEXT_INSN (insn
);
1696 map
->orig_asm_operands_vector
= 0;
1698 switch (GET_CODE (insn
))
1701 pattern
= PATTERN (insn
);
1705 /* Check to see if this is a giv that has been combined with
1706 some split address givs. (Combined in the sense that
1707 `combine_givs' in loop.c has put two givs in the same register.)
1708 In this case, we must search all givs based on the same biv to
1709 find the address givs. Then split the address givs.
1710 Do this before splitting the giv, since that may map the
1711 SET_DEST to a new register. */
1713 if ((set
= single_set (insn
))
1714 && GET_CODE (SET_DEST (set
)) == REG
1715 && addr_combined_regs
[REGNO (SET_DEST (set
))])
1717 struct iv_class
*bl
;
1718 struct induction
*v
, *tv
;
1719 int regno
= REGNO (SET_DEST (set
));
1721 v
= addr_combined_regs
[REGNO (SET_DEST (set
))];
1722 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
1724 /* Although the giv_inc amount is not needed here, we must call
1725 calculate_giv_inc here since it might try to delete the
1726 last insn emitted. If we wait until later to call it,
1727 we might accidentally delete insns generated immediately
1728 below by emit_unrolled_add. */
1730 giv_inc
= calculate_giv_inc (set
, insn
, regno
);
1732 /* Now find all address giv's that were combined with this
1734 for (tv
= bl
->giv
; tv
; tv
= tv
->next_iv
)
1735 if (tv
->giv_type
== DEST_ADDR
&& tv
->same
== v
)
1739 /* If this DEST_ADDR giv was not split, then ignore it. */
1740 if (*tv
->location
!= tv
->dest_reg
)
1743 /* Scale this_giv_inc if the multiplicative factors of
1744 the two givs are different. */
1745 this_giv_inc
= INTVAL (giv_inc
);
1746 if (tv
->mult_val
!= v
->mult_val
)
1747 this_giv_inc
= (this_giv_inc
/ INTVAL (v
->mult_val
)
1748 * INTVAL (tv
->mult_val
));
1750 tv
->dest_reg
= plus_constant (tv
->dest_reg
, this_giv_inc
);
1751 *tv
->location
= tv
->dest_reg
;
1753 if (last_iteration
&& unroll_type
!= UNROLL_COMPLETELY
)
1755 /* Must emit an insn to increment the split address
1756 giv. Add in the const_adjust field in case there
1757 was a constant eliminated from the address. */
1758 rtx value
, dest_reg
;
1760 /* tv->dest_reg will be either a bare register,
1761 or else a register plus a constant. */
1762 if (GET_CODE (tv
->dest_reg
) == REG
)
1763 dest_reg
= tv
->dest_reg
;
1765 dest_reg
= XEXP (tv
->dest_reg
, 0);
1767 /* Check for shared address givs, and avoid
1768 incrementing the shared pseudo reg more than
1770 if (! tv
->same_insn
&& ! tv
->shared
)
1772 /* tv->dest_reg may actually be a (PLUS (REG)
1773 (CONST)) here, so we must call plus_constant
1774 to add the const_adjust amount before calling
1775 emit_unrolled_add below. */
1776 value
= plus_constant (tv
->dest_reg
,
1779 /* The constant could be too large for an add
1780 immediate, so can't directly emit an insn
1782 emit_unrolled_add (dest_reg
, XEXP (value
, 0),
1786 /* Reset the giv to be just the register again, in case
1787 it is used after the set we have just emitted.
1788 We must subtract the const_adjust factor added in
1790 tv
->dest_reg
= plus_constant (dest_reg
,
1791 - tv
->const_adjust
);
1792 *tv
->location
= tv
->dest_reg
;
1797 /* If this is a setting of a splittable variable, then determine
1798 how to split the variable, create a new set based on this split,
1799 and set up the reg_map so that later uses of the variable will
1800 use the new split variable. */
1802 dest_reg_was_split
= 0;
1804 if ((set
= single_set (insn
))
1805 && GET_CODE (SET_DEST (set
)) == REG
1806 && splittable_regs
[REGNO (SET_DEST (set
))])
1808 int regno
= REGNO (SET_DEST (set
));
1810 dest_reg_was_split
= 1;
1812 /* Compute the increment value for the giv, if it wasn't
1813 already computed above. */
1816 giv_inc
= calculate_giv_inc (set
, insn
, regno
);
1817 giv_dest_reg
= SET_DEST (set
);
1818 giv_src_reg
= SET_DEST (set
);
1820 if (unroll_type
== UNROLL_COMPLETELY
)
1822 /* Completely unrolling the loop. Set the induction
1823 variable to a known constant value. */
1825 /* The value in splittable_regs may be an invariant
1826 value, so we must use plus_constant here. */
1827 splittable_regs
[regno
]
1828 = plus_constant (splittable_regs
[regno
], INTVAL (giv_inc
));
1830 if (GET_CODE (splittable_regs
[regno
]) == PLUS
)
1832 giv_src_reg
= XEXP (splittable_regs
[regno
], 0);
1833 giv_inc
= XEXP (splittable_regs
[regno
], 1);
1837 /* The splittable_regs value must be a REG or a
1838 CONST_INT, so put the entire value in the giv_src_reg
1840 giv_src_reg
= splittable_regs
[regno
];
1841 giv_inc
= const0_rtx
;
1846 /* Partially unrolling loop. Create a new pseudo
1847 register for the iteration variable, and set it to
1848 be a constant plus the original register. Except
1849 on the last iteration, when the result has to
1850 go back into the original iteration var register. */
1852 /* Handle bivs which must be mapped to a new register
1853 when split. This happens for bivs which need their
1854 final value set before loop entry. The new register
1855 for the biv was stored in the biv's first struct
1856 induction entry by find_splittable_regs. */
1858 if (regno
< max_reg_before_loop
1859 && reg_iv_type
[regno
] == BASIC_INDUCT
)
1861 giv_src_reg
= reg_biv_class
[regno
]->biv
->src_reg
;
1862 giv_dest_reg
= giv_src_reg
;
1866 /* If non-reduced/final-value givs were split, then
1867 this would have to remap those givs also. See
1868 find_splittable_regs. */
1871 splittable_regs
[regno
]
1872 = GEN_INT (INTVAL (giv_inc
)
1873 + INTVAL (splittable_regs
[regno
]));
1874 giv_inc
= splittable_regs
[regno
];
1876 /* Now split the induction variable by changing the dest
1877 of this insn to a new register, and setting its
1878 reg_map entry to point to this new register.
1880 If this is the last iteration, and this is the last insn
1881 that will update the iv, then reuse the original dest,
1882 to ensure that the iv will have the proper value when
1883 the loop exits or repeats.
1885 Using splittable_regs_updates here like this is safe,
1886 because it can only be greater than one if all
1887 instructions modifying the iv are always executed in
1890 if (! last_iteration
1891 || (splittable_regs_updates
[regno
]-- != 1))
1893 tem
= gen_reg_rtx (GET_MODE (giv_src_reg
));
1895 map
->reg_map
[regno
] = tem
;
1896 record_base_value (REGNO (tem
),
1897 giv_inc
== const0_rtx
1899 : gen_rtx_PLUS (GET_MODE (giv_src_reg
),
1900 giv_src_reg
, giv_inc
),
1904 map
->reg_map
[regno
] = giv_src_reg
;
1907 /* The constant being added could be too large for an add
1908 immediate, so can't directly emit an insn here. */
1909 emit_unrolled_add (giv_dest_reg
, giv_src_reg
, giv_inc
);
1910 copy
= get_last_insn ();
1911 pattern
= PATTERN (copy
);
1915 pattern
= copy_rtx_and_substitute (pattern
, map
);
1916 copy
= emit_insn (pattern
);
1918 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
1921 /* If this insn is setting CC0, it may need to look at
1922 the insn that uses CC0 to see what type of insn it is.
1923 In that case, the call to recog via validate_change will
1924 fail. So don't substitute constants here. Instead,
1925 do it when we emit the following insn.
1927 For example, see the pyr.md file. That machine has signed and
1928 unsigned compares. The compare patterns must check the
1929 following branch insn to see which what kind of compare to
1932 If the previous insn set CC0, substitute constants on it as
1934 if (sets_cc0_p (PATTERN (copy
)) != 0)
1939 try_constants (cc0_insn
, map
);
1941 try_constants (copy
, map
);
1944 try_constants (copy
, map
);
1947 /* Make split induction variable constants `permanent' since we
1948 know there are no backward branches across iteration variable
1949 settings which would invalidate this. */
1950 if (dest_reg_was_split
)
1952 int regno
= REGNO (SET_DEST (pattern
));
1954 if (regno
< map
->const_equiv_map_size
1955 && map
->const_age_map
[regno
] == map
->const_age
)
1956 map
->const_age_map
[regno
] = -1;
1961 pattern
= copy_rtx_and_substitute (PATTERN (insn
), map
);
1962 copy
= emit_jump_insn (pattern
);
1963 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
1965 if (JUMP_LABEL (insn
) == start_label
&& insn
== copy_end
1966 && ! last_iteration
)
1968 /* This is a branch to the beginning of the loop; this is the
1969 last insn being copied; and this is not the last iteration.
1970 In this case, we want to change the original fall through
1971 case to be a branch past the end of the loop, and the
1972 original jump label case to fall_through. */
1974 if (invert_exp (pattern
, copy
))
1976 if (! redirect_exp (&pattern
,
1977 get_label_from_map (map
,
1979 (JUMP_LABEL (insn
))),
1986 rtx lab
= gen_label_rtx ();
1987 /* Can't do it by reversing the jump (probably because we
1988 couldn't reverse the conditions), so emit a new
1989 jump_insn after COPY, and redirect the jump around
1991 jmp
= emit_jump_insn_after (gen_jump (exit_label
), copy
);
1992 jmp
= emit_barrier_after (jmp
);
1993 emit_label_after (lab
, jmp
);
1994 LABEL_NUSES (lab
) = 0;
1995 if (! redirect_exp (&pattern
,
1996 get_label_from_map (map
,
1998 (JUMP_LABEL (insn
))),
2006 try_constants (cc0_insn
, map
);
2009 try_constants (copy
, map
);
2011 /* Set the jump label of COPY correctly to avoid problems with
2012 later passes of unroll_loop, if INSN had jump label set. */
2013 if (JUMP_LABEL (insn
))
2017 /* Can't use the label_map for every insn, since this may be
2018 the backward branch, and hence the label was not mapped. */
2019 if ((set
= single_set (copy
)))
2021 tem
= SET_SRC (set
);
2022 if (GET_CODE (tem
) == LABEL_REF
)
2023 label
= XEXP (tem
, 0);
2024 else if (GET_CODE (tem
) == IF_THEN_ELSE
)
2026 if (XEXP (tem
, 1) != pc_rtx
)
2027 label
= XEXP (XEXP (tem
, 1), 0);
2029 label
= XEXP (XEXP (tem
, 2), 0);
2033 if (label
&& GET_CODE (label
) == CODE_LABEL
)
2034 JUMP_LABEL (copy
) = label
;
2037 /* An unrecognizable jump insn, probably the entry jump
2038 for a switch statement. This label must have been mapped,
2039 so just use the label_map to get the new jump label. */
2041 = get_label_from_map (map
,
2042 CODE_LABEL_NUMBER (JUMP_LABEL (insn
)));
2045 /* If this is a non-local jump, then must increase the label
2046 use count so that the label will not be deleted when the
2047 original jump is deleted. */
2048 LABEL_NUSES (JUMP_LABEL (copy
))++;
2050 else if (GET_CODE (PATTERN (copy
)) == ADDR_VEC
2051 || GET_CODE (PATTERN (copy
)) == ADDR_DIFF_VEC
)
2053 rtx pat
= PATTERN (copy
);
2054 int diff_vec_p
= GET_CODE (pat
) == ADDR_DIFF_VEC
;
2055 int len
= XVECLEN (pat
, diff_vec_p
);
2058 for (i
= 0; i
< len
; i
++)
2059 LABEL_NUSES (XEXP (XVECEXP (pat
, diff_vec_p
, i
), 0))++;
2062 /* If this used to be a conditional jump insn but whose branch
2063 direction is now known, we must do something special. */
2064 if (condjump_p (insn
) && !simplejump_p (insn
) && map
->last_pc_value
)
2067 /* The previous insn set cc0 for us. So delete it. */
2068 delete_insn (PREV_INSN (copy
));
2071 /* If this is now a no-op, delete it. */
2072 if (map
->last_pc_value
== pc_rtx
)
2074 /* Don't let delete_insn delete the label referenced here,
2075 because we might possibly need it later for some other
2076 instruction in the loop. */
2077 if (JUMP_LABEL (copy
))
2078 LABEL_NUSES (JUMP_LABEL (copy
))++;
2080 if (JUMP_LABEL (copy
))
2081 LABEL_NUSES (JUMP_LABEL (copy
))--;
2085 /* Otherwise, this is unconditional jump so we must put a
2086 BARRIER after it. We could do some dead code elimination
2087 here, but jump.c will do it just as well. */
2093 pattern
= copy_rtx_and_substitute (PATTERN (insn
), map
);
2094 copy
= emit_call_insn (pattern
);
2095 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
2097 /* Because the USAGE information potentially contains objects other
2098 than hard registers, we need to copy it. */
2099 CALL_INSN_FUNCTION_USAGE (copy
)
2100 = copy_rtx_and_substitute (CALL_INSN_FUNCTION_USAGE (insn
), map
);
2104 try_constants (cc0_insn
, map
);
2107 try_constants (copy
, map
);
2109 /* Be lazy and assume CALL_INSNs clobber all hard registers. */
2110 for (i
= 0; i
< FIRST_PSEUDO_REGISTER
; i
++)
2111 map
->const_equiv_map
[i
] = 0;
2115 /* If this is the loop start label, then we don't need to emit a
2116 copy of this label since no one will use it. */
2118 if (insn
!= start_label
)
2120 copy
= emit_label (get_label_from_map (map
,
2121 CODE_LABEL_NUMBER (insn
)));
2127 copy
= emit_barrier ();
2131 /* VTOP notes are valid only before the loop exit test. If placed
2132 anywhere else, loop may generate bad code. */
2134 if (NOTE_LINE_NUMBER (insn
) != NOTE_INSN_DELETED
2135 && (NOTE_LINE_NUMBER (insn
) != NOTE_INSN_LOOP_VTOP
2136 || (last_iteration
&& unroll_type
!= UNROLL_COMPLETELY
)))
2137 copy
= emit_note (NOTE_SOURCE_FILE (insn
),
2138 NOTE_LINE_NUMBER (insn
));
2148 map
->insn_map
[INSN_UID (insn
)] = copy
;
2150 while (insn
!= copy_end
);
2152 /* Now finish coping the REG_NOTES. */
2156 insn
= NEXT_INSN (insn
);
2157 if ((GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == JUMP_INSN
2158 || GET_CODE (insn
) == CALL_INSN
)
2159 && map
->insn_map
[INSN_UID (insn
)])
2160 final_reg_note_copy (REG_NOTES (map
->insn_map
[INSN_UID (insn
)]), map
);
2162 while (insn
!= copy_end
);
2164 /* There may be notes between copy_notes_from and loop_end. Emit a copy of
2165 each of these notes here, since there may be some important ones, such as
2166 NOTE_INSN_BLOCK_END notes, in this group. We don't do this on the last
2167 iteration, because the original notes won't be deleted.
2169 We can't use insert_before here, because when from preconditioning,
2170 insert_before points before the loop. We can't use copy_end, because
2171 there may be insns already inserted after it (which we don't want to
2172 copy) when not from preconditioning code. */
2174 if (! last_iteration
)
2176 for (insn
= copy_notes_from
; insn
!= loop_end
; insn
= NEXT_INSN (insn
))
2178 if (GET_CODE (insn
) == NOTE
2179 && NOTE_LINE_NUMBER (insn
) != NOTE_INSN_DELETED
)
2180 emit_note (NOTE_SOURCE_FILE (insn
), NOTE_LINE_NUMBER (insn
));
2184 if (final_label
&& LABEL_NUSES (final_label
) > 0)
2185 emit_label (final_label
);
2187 tem
= gen_sequence ();
2189 emit_insn_before (tem
, insert_before
);
2192 /* Emit an insn, using the expand_binop to ensure that a valid insn is
2193 emitted. This will correctly handle the case where the increment value
2194 won't fit in the immediate field of a PLUS insns. */
2197 emit_unrolled_add (dest_reg
, src_reg
, increment
)
2198 rtx dest_reg
, src_reg
, increment
;
2202 result
= expand_binop (GET_MODE (dest_reg
), add_optab
, src_reg
, increment
,
2203 dest_reg
, 0, OPTAB_LIB_WIDEN
);
2205 if (dest_reg
!= result
)
2206 emit_move_insn (dest_reg
, result
);
2209 /* Searches the insns between INSN and LOOP_END. Returns 1 if there
2210 is a backward branch in that range that branches to somewhere between
2211 LOOP_START and INSN. Returns 0 otherwise. */
2213 /* ??? This is quadratic algorithm. Could be rewritten to be linear.
2214 In practice, this is not a problem, because this function is seldom called,
2215 and uses a negligible amount of CPU time on average. */
2218 back_branch_in_range_p (insn
, loop_start
, loop_end
)
2220 rtx loop_start
, loop_end
;
2222 rtx p
, q
, target_insn
;
2223 rtx orig_loop_end
= loop_end
;
2225 /* Stop before we get to the backward branch at the end of the loop. */
2226 loop_end
= prev_nonnote_insn (loop_end
);
2227 if (GET_CODE (loop_end
) == BARRIER
)
2228 loop_end
= PREV_INSN (loop_end
);
2230 /* Check in case insn has been deleted, search forward for first non
2231 deleted insn following it. */
2232 while (INSN_DELETED_P (insn
))
2233 insn
= NEXT_INSN (insn
);
2235 /* Check for the case where insn is the last insn in the loop. Deal
2236 with the case where INSN was a deleted loop test insn, in which case
2237 it will now be the NOTE_LOOP_END. */
2238 if (insn
== loop_end
|| insn
== orig_loop_end
)
2241 for (p
= NEXT_INSN (insn
); p
!= loop_end
; p
= NEXT_INSN (p
))
2243 if (GET_CODE (p
) == JUMP_INSN
)
2245 target_insn
= JUMP_LABEL (p
);
2247 /* Search from loop_start to insn, to see if one of them is
2248 the target_insn. We can't use INSN_LUID comparisons here,
2249 since insn may not have an LUID entry. */
2250 for (q
= loop_start
; q
!= insn
; q
= NEXT_INSN (q
))
2251 if (q
== target_insn
)
2259 /* Try to generate the simplest rtx for the expression
2260 (PLUS (MULT mult1 mult2) add1). This is used to calculate the initial
2264 fold_rtx_mult_add (mult1
, mult2
, add1
, mode
)
2265 rtx mult1
, mult2
, add1
;
2266 enum machine_mode mode
;
2271 /* The modes must all be the same. This should always be true. For now,
2272 check to make sure. */
2273 if ((GET_MODE (mult1
) != mode
&& GET_MODE (mult1
) != VOIDmode
)
2274 || (GET_MODE (mult2
) != mode
&& GET_MODE (mult2
) != VOIDmode
)
2275 || (GET_MODE (add1
) != mode
&& GET_MODE (add1
) != VOIDmode
))
2278 /* Ensure that if at least one of mult1/mult2 are constant, then mult2
2279 will be a constant. */
2280 if (GET_CODE (mult1
) == CONST_INT
)
2287 mult_res
= simplify_binary_operation (MULT
, mode
, mult1
, mult2
);
2289 mult_res
= gen_rtx_MULT (mode
, mult1
, mult2
);
2291 /* Again, put the constant second. */
2292 if (GET_CODE (add1
) == CONST_INT
)
2299 result
= simplify_binary_operation (PLUS
, mode
, add1
, mult_res
);
2301 result
= gen_rtx_PLUS (mode
, add1
, mult_res
);
2306 /* Searches the list of induction struct's for the biv BL, to try to calculate
2307 the total increment value for one iteration of the loop as a constant.
2309 Returns the increment value as an rtx, simplified as much as possible,
2310 if it can be calculated. Otherwise, returns 0. */
2313 biv_total_increment (bl
, loop_start
, loop_end
)
2314 struct iv_class
*bl
;
2315 rtx loop_start
, loop_end
;
2317 struct induction
*v
;
2320 /* For increment, must check every instruction that sets it. Each
2321 instruction must be executed only once each time through the loop.
2322 To verify this, we check that the insn is always executed, and that
2323 there are no backward branches after the insn that branch to before it.
2324 Also, the insn must have a mult_val of one (to make sure it really is
2327 result
= const0_rtx
;
2328 for (v
= bl
->biv
; v
; v
= v
->next_iv
)
2330 if (v
->always_computable
&& v
->mult_val
== const1_rtx
2331 && ! back_branch_in_range_p (v
->insn
, loop_start
, loop_end
))
2332 result
= fold_rtx_mult_add (result
, const1_rtx
, v
->add_val
, v
->mode
);
2340 /* Determine the initial value of the iteration variable, and the amount
2341 that it is incremented each loop. Use the tables constructed by
2342 the strength reduction pass to calculate these values.
2344 Initial_value and/or increment are set to zero if their values could not
2348 iteration_info (iteration_var
, initial_value
, increment
, loop_start
, loop_end
)
2349 rtx iteration_var
, *initial_value
, *increment
;
2350 rtx loop_start
, loop_end
;
2352 struct iv_class
*bl
;
2354 struct induction
*v
;
2357 /* Clear the result values, in case no answer can be found. */
2361 /* The iteration variable can be either a giv or a biv. Check to see
2362 which it is, and compute the variable's initial value, and increment
2363 value if possible. */
2365 /* If this is a new register, can't handle it since we don't have any
2366 reg_iv_type entry for it. */
2367 if (REGNO (iteration_var
) >= max_reg_before_loop
)
2369 if (loop_dump_stream
)
2370 fprintf (loop_dump_stream
,
2371 "Loop unrolling: No reg_iv_type entry for iteration var.\n");
2375 /* Reject iteration variables larger than the host wide int size, since they
2376 could result in a number of iterations greater than the range of our
2377 `unsigned HOST_WIDE_INT' variable loop_n_iterations. */
2378 else if ((GET_MODE_BITSIZE (GET_MODE (iteration_var
))
2379 > HOST_BITS_PER_WIDE_INT
))
2381 if (loop_dump_stream
)
2382 fprintf (loop_dump_stream
,
2383 "Loop unrolling: Iteration var rejected because mode too large.\n");
2386 else if (GET_MODE_CLASS (GET_MODE (iteration_var
)) != MODE_INT
)
2388 if (loop_dump_stream
)
2389 fprintf (loop_dump_stream
,
2390 "Loop unrolling: Iteration var not an integer.\n");
2393 else if (reg_iv_type
[REGNO (iteration_var
)] == BASIC_INDUCT
)
2395 /* Grab initial value, only useful if it is a constant. */
2396 bl
= reg_biv_class
[REGNO (iteration_var
)];
2397 *initial_value
= bl
->initial_value
;
2399 *increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2401 else if (reg_iv_type
[REGNO (iteration_var
)] == GENERAL_INDUCT
)
2404 /* ??? The code below does not work because the incorrect number of
2405 iterations is calculated when the biv is incremented after the giv
2406 is set (which is the usual case). This can probably be accounted
2407 for by biasing the initial_value by subtracting the amount of the
2408 increment that occurs between the giv set and the giv test. However,
2409 a giv as an iterator is very rare, so it does not seem worthwhile
2411 /* ??? An example failure is: i = 6; do {;} while (i++ < 9). */
2412 if (loop_dump_stream
)
2413 fprintf (loop_dump_stream
,
2414 "Loop unrolling: Giv iterators are not handled.\n");
2417 /* Initial value is mult_val times the biv's initial value plus
2418 add_val. Only useful if it is a constant. */
2419 v
= reg_iv_info
[REGNO (iteration_var
)];
2420 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
2421 *initial_value
= fold_rtx_mult_add (v
->mult_val
, bl
->initial_value
,
2422 v
->add_val
, v
->mode
);
2424 /* Increment value is mult_val times the increment value of the biv. */
2426 *increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2428 *increment
= fold_rtx_mult_add (v
->mult_val
, *increment
, const0_rtx
,
2434 if (loop_dump_stream
)
2435 fprintf (loop_dump_stream
,
2436 "Loop unrolling: Not basic or general induction var.\n");
2441 /* Calculate the approximate final value of the iteration variable
2442 which has an loop exit test with code COMPARISON_CODE and comparison value
2443 of COMPARISON_VALUE. Also returns an indication of whether the comparison
2444 was signed or unsigned, and the direction of the comparison. This info is
2445 needed to calculate the number of loop iterations. */
2448 approx_final_value (comparison_code
, comparison_value
, unsigned_p
, compare_dir
)
2449 enum rtx_code comparison_code
;
2450 rtx comparison_value
;
2454 /* Calculate the final value of the induction variable.
2455 The exact final value depends on the branch operator, and increment sign.
2456 This is only an approximate value. It will be wrong if the iteration
2457 variable is not incremented by one each time through the loop, and
2458 approx final value - start value % increment != 0. */
2461 switch (comparison_code
)
2467 return plus_constant (comparison_value
, 1);
2472 return plus_constant (comparison_value
, -1);
2474 /* Can not calculate a final value for this case. */
2481 return comparison_value
;
2487 return comparison_value
;
2490 return comparison_value
;
2496 /* For each biv and giv, determine whether it can be safely split into
2497 a different variable for each unrolled copy of the loop body. If it
2498 is safe to split, then indicate that by saving some useful info
2499 in the splittable_regs array.
2501 If the loop is being completely unrolled, then splittable_regs will hold
2502 the current value of the induction variable while the loop is unrolled.
2503 It must be set to the initial value of the induction variable here.
2504 Otherwise, splittable_regs will hold the difference between the current
2505 value of the induction variable and the value the induction variable had
2506 at the top of the loop. It must be set to the value 0 here.
2508 Returns the total number of instructions that set registers that are
2511 /* ?? If the loop is only unrolled twice, then most of the restrictions to
2512 constant values are unnecessary, since we can easily calculate increment
2513 values in this case even if nothing is constant. The increment value
2514 should not involve a multiply however. */
2516 /* ?? Even if the biv/giv increment values aren't constant, it may still
2517 be beneficial to split the variable if the loop is only unrolled a few
2518 times, since multiplies by small integers (1,2,3,4) are very cheap. */
2521 find_splittable_regs (unroll_type
, loop_start
, loop_end
, end_insert_before
,
2523 enum unroll_types unroll_type
;
2524 rtx loop_start
, loop_end
;
2525 rtx end_insert_before
;
2528 struct iv_class
*bl
;
2529 struct induction
*v
;
2531 rtx biv_final_value
;
2535 for (bl
= loop_iv_list
; bl
; bl
= bl
->next
)
2537 /* Biv_total_increment must return a constant value,
2538 otherwise we can not calculate the split values. */
2540 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2541 if (! increment
|| GET_CODE (increment
) != CONST_INT
)
2544 /* The loop must be unrolled completely, or else have a known number
2545 of iterations and only one exit, or else the biv must be dead
2546 outside the loop, or else the final value must be known. Otherwise,
2547 it is unsafe to split the biv since it may not have the proper
2548 value on loop exit. */
2550 /* loop_number_exit_count is non-zero if the loop has an exit other than
2551 a fall through at the end. */
2554 biv_final_value
= 0;
2555 if (unroll_type
!= UNROLL_COMPLETELY
2556 && (loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]]
2557 || unroll_type
== UNROLL_NAIVE
)
2558 && (uid_luid
[REGNO_LAST_UID (bl
->regno
)] >= INSN_LUID (loop_end
)
2560 || INSN_UID (bl
->init_insn
) >= max_uid_for_loop
2561 || (uid_luid
[REGNO_FIRST_UID (bl
->regno
)]
2562 < INSN_LUID (bl
->init_insn
))
2563 || reg_mentioned_p (bl
->biv
->dest_reg
, SET_SRC (bl
->init_set
)))
2564 && ! (biv_final_value
= final_biv_value (bl
, loop_start
, loop_end
)))
2567 /* If any of the insns setting the BIV don't do so with a simple
2568 PLUS, we don't know how to split it. */
2569 for (v
= bl
->biv
; biv_splittable
&& v
; v
= v
->next_iv
)
2570 if ((tem
= single_set (v
->insn
)) == 0
2571 || GET_CODE (SET_DEST (tem
)) != REG
2572 || REGNO (SET_DEST (tem
)) != bl
->regno
2573 || GET_CODE (SET_SRC (tem
)) != PLUS
)
2576 /* If final value is non-zero, then must emit an instruction which sets
2577 the value of the biv to the proper value. This is done after
2578 handling all of the givs, since some of them may need to use the
2579 biv's value in their initialization code. */
2581 /* This biv is splittable. If completely unrolling the loop, save
2582 the biv's initial value. Otherwise, save the constant zero. */
2584 if (biv_splittable
== 1)
2586 if (unroll_type
== UNROLL_COMPLETELY
)
2588 /* If the initial value of the biv is itself (i.e. it is too
2589 complicated for strength_reduce to compute), or is a hard
2590 register, or it isn't invariant, then we must create a new
2591 pseudo reg to hold the initial value of the biv. */
2593 if (GET_CODE (bl
->initial_value
) == REG
2594 && (REGNO (bl
->initial_value
) == bl
->regno
2595 || REGNO (bl
->initial_value
) < FIRST_PSEUDO_REGISTER
2596 || ! invariant_p (bl
->initial_value
)))
2598 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2600 record_base_value (REGNO (tem
), bl
->biv
->add_val
, 0);
2601 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2604 if (loop_dump_stream
)
2605 fprintf (loop_dump_stream
, "Biv %d initial value remapped to %d.\n",
2606 bl
->regno
, REGNO (tem
));
2608 splittable_regs
[bl
->regno
] = tem
;
2611 splittable_regs
[bl
->regno
] = bl
->initial_value
;
2614 splittable_regs
[bl
->regno
] = const0_rtx
;
2616 /* Save the number of instructions that modify the biv, so that
2617 we can treat the last one specially. */
2619 splittable_regs_updates
[bl
->regno
] = bl
->biv_count
;
2620 result
+= bl
->biv_count
;
2622 if (loop_dump_stream
)
2623 fprintf (loop_dump_stream
,
2624 "Biv %d safe to split.\n", bl
->regno
);
2627 /* Check every giv that depends on this biv to see whether it is
2628 splittable also. Even if the biv isn't splittable, givs which
2629 depend on it may be splittable if the biv is live outside the
2630 loop, and the givs aren't. */
2632 result
+= find_splittable_givs (bl
, unroll_type
, loop_start
, loop_end
,
2633 increment
, unroll_number
);
2635 /* If final value is non-zero, then must emit an instruction which sets
2636 the value of the biv to the proper value. This is done after
2637 handling all of the givs, since some of them may need to use the
2638 biv's value in their initialization code. */
2639 if (biv_final_value
)
2641 /* If the loop has multiple exits, emit the insns before the
2642 loop to ensure that it will always be executed no matter
2643 how the loop exits. Otherwise emit the insn after the loop,
2644 since this is slightly more efficient. */
2645 if (! loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]])
2646 emit_insn_before (gen_move_insn (bl
->biv
->src_reg
,
2651 /* Create a new register to hold the value of the biv, and then
2652 set the biv to its final value before the loop start. The biv
2653 is set to its final value before loop start to ensure that
2654 this insn will always be executed, no matter how the loop
2656 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2657 record_base_value (REGNO (tem
), bl
->biv
->add_val
, 0);
2659 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2661 emit_insn_before (gen_move_insn (bl
->biv
->src_reg
,
2665 if (loop_dump_stream
)
2666 fprintf (loop_dump_stream
, "Biv %d mapped to %d for split.\n",
2667 REGNO (bl
->biv
->src_reg
), REGNO (tem
));
2669 /* Set up the mapping from the original biv register to the new
2671 bl
->biv
->src_reg
= tem
;
2678 /* Return 1 if the first and last unrolled copy of the address giv V is valid
2679 for the instruction that is using it. Do not make any changes to that
2683 verify_addresses (v
, giv_inc
, unroll_number
)
2684 struct induction
*v
;
2689 rtx orig_addr
= *v
->location
;
2690 rtx last_addr
= plus_constant (v
->dest_reg
,
2691 INTVAL (giv_inc
) * (unroll_number
- 1));
2693 /* First check to see if either address would fail. Handle the fact
2694 that we have may have a match_dup. */
2695 if (! validate_replace_rtx (*v
->location
, v
->dest_reg
, v
->insn
)
2696 || ! validate_replace_rtx (*v
->location
, last_addr
, v
->insn
))
2699 /* Now put things back the way they were before. This should always
2701 if (! validate_replace_rtx (*v
->location
, orig_addr
, v
->insn
))
2707 /* For every giv based on the biv BL, check to determine whether it is
2708 splittable. This is a subroutine to find_splittable_regs ().
2710 Return the number of instructions that set splittable registers. */
2713 find_splittable_givs (bl
, unroll_type
, loop_start
, loop_end
, increment
,
2715 struct iv_class
*bl
;
2716 enum unroll_types unroll_type
;
2717 rtx loop_start
, loop_end
;
2721 struct induction
*v
, *v2
;
2726 /* Scan the list of givs, and set the same_insn field when there are
2727 multiple identical givs in the same insn. */
2728 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
2729 for (v2
= v
->next_iv
; v2
; v2
= v2
->next_iv
)
2730 if (v
->insn
== v2
->insn
&& rtx_equal_p (v
->new_reg
, v2
->new_reg
)
2734 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
2738 /* Only split the giv if it has already been reduced, or if the loop is
2739 being completely unrolled. */
2740 if (unroll_type
!= UNROLL_COMPLETELY
&& v
->ignore
)
2743 /* The giv can be split if the insn that sets the giv is executed once
2744 and only once on every iteration of the loop. */
2745 /* An address giv can always be split. v->insn is just a use not a set,
2746 and hence it does not matter whether it is always executed. All that
2747 matters is that all the biv increments are always executed, and we
2748 won't reach here if they aren't. */
2749 if (v
->giv_type
!= DEST_ADDR
2750 && (! v
->always_computable
2751 || back_branch_in_range_p (v
->insn
, loop_start
, loop_end
)))
2754 /* The giv increment value must be a constant. */
2755 giv_inc
= fold_rtx_mult_add (v
->mult_val
, increment
, const0_rtx
,
2757 if (! giv_inc
|| GET_CODE (giv_inc
) != CONST_INT
)
2760 /* The loop must be unrolled completely, or else have a known number of
2761 iterations and only one exit, or else the giv must be dead outside
2762 the loop, or else the final value of the giv must be known.
2763 Otherwise, it is not safe to split the giv since it may not have the
2764 proper value on loop exit. */
2766 /* The used outside loop test will fail for DEST_ADDR givs. They are
2767 never used outside the loop anyways, so it is always safe to split a
2771 if (unroll_type
!= UNROLL_COMPLETELY
2772 && (loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]]
2773 || unroll_type
== UNROLL_NAIVE
)
2774 && v
->giv_type
!= DEST_ADDR
2775 /* The next part is true if the pseudo is used outside the loop.
2776 We assume that this is true for any pseudo created after loop
2777 starts, because we don't have a reg_n_info entry for them. */
2778 && (REGNO (v
->dest_reg
) >= max_reg_before_loop
2779 || (REGNO_FIRST_UID (REGNO (v
->dest_reg
)) != INSN_UID (v
->insn
)
2780 /* Check for the case where the pseudo is set by a shift/add
2781 sequence, in which case the first insn setting the pseudo
2782 is the first insn of the shift/add sequence. */
2783 && (! (tem
= find_reg_note (v
->insn
, REG_RETVAL
, NULL_RTX
))
2784 || (REGNO_FIRST_UID (REGNO (v
->dest_reg
))
2785 != INSN_UID (XEXP (tem
, 0)))))
2786 /* Line above always fails if INSN was moved by loop opt. */
2787 || (uid_luid
[REGNO_LAST_UID (REGNO (v
->dest_reg
))]
2788 >= INSN_LUID (loop_end
)))
2789 && ! (final_value
= v
->final_value
))
2793 /* Currently, non-reduced/final-value givs are never split. */
2794 /* Should emit insns after the loop if possible, as the biv final value
2797 /* If the final value is non-zero, and the giv has not been reduced,
2798 then must emit an instruction to set the final value. */
2799 if (final_value
&& !v
->new_reg
)
2801 /* Create a new register to hold the value of the giv, and then set
2802 the giv to its final value before the loop start. The giv is set
2803 to its final value before loop start to ensure that this insn
2804 will always be executed, no matter how we exit. */
2805 tem
= gen_reg_rtx (v
->mode
);
2806 emit_insn_before (gen_move_insn (tem
, v
->dest_reg
), loop_start
);
2807 emit_insn_before (gen_move_insn (v
->dest_reg
, final_value
),
2810 if (loop_dump_stream
)
2811 fprintf (loop_dump_stream
, "Giv %d mapped to %d for split.\n",
2812 REGNO (v
->dest_reg
), REGNO (tem
));
2818 /* This giv is splittable. If completely unrolling the loop, save the
2819 giv's initial value. Otherwise, save the constant zero for it. */
2821 if (unroll_type
== UNROLL_COMPLETELY
)
2823 /* It is not safe to use bl->initial_value here, because it may not
2824 be invariant. It is safe to use the initial value stored in
2825 the splittable_regs array if it is set. In rare cases, it won't
2826 be set, so then we do exactly the same thing as
2827 find_splittable_regs does to get a safe value. */
2828 rtx biv_initial_value
;
2830 if (splittable_regs
[bl
->regno
])
2831 biv_initial_value
= splittable_regs
[bl
->regno
];
2832 else if (GET_CODE (bl
->initial_value
) != REG
2833 || (REGNO (bl
->initial_value
) != bl
->regno
2834 && REGNO (bl
->initial_value
) >= FIRST_PSEUDO_REGISTER
))
2835 biv_initial_value
= bl
->initial_value
;
2838 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2840 record_base_value (REGNO (tem
), bl
->biv
->add_val
, 0);
2841 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2843 biv_initial_value
= tem
;
2845 value
= fold_rtx_mult_add (v
->mult_val
, biv_initial_value
,
2846 v
->add_val
, v
->mode
);
2853 /* If a giv was combined with another giv, then we can only split
2854 this giv if the giv it was combined with was reduced. This
2855 is because the value of v->new_reg is meaningless in this
2857 if (v
->same
&& ! v
->same
->new_reg
)
2859 if (loop_dump_stream
)
2860 fprintf (loop_dump_stream
,
2861 "giv combined with unreduced giv not split.\n");
2864 /* If the giv is an address destination, it could be something other
2865 than a simple register, these have to be treated differently. */
2866 else if (v
->giv_type
== DEST_REG
)
2868 /* If value is not a constant, register, or register plus
2869 constant, then compute its value into a register before
2870 loop start. This prevents invalid rtx sharing, and should
2871 generate better code. We can use bl->initial_value here
2872 instead of splittable_regs[bl->regno] because this code
2873 is going before the loop start. */
2874 if (unroll_type
== UNROLL_COMPLETELY
2875 && GET_CODE (value
) != CONST_INT
2876 && GET_CODE (value
) != REG
2877 && (GET_CODE (value
) != PLUS
2878 || GET_CODE (XEXP (value
, 0)) != REG
2879 || GET_CODE (XEXP (value
, 1)) != CONST_INT
))
2881 rtx tem
= gen_reg_rtx (v
->mode
);
2882 record_base_value (REGNO (tem
), v
->add_val
, 0);
2883 emit_iv_add_mult (bl
->initial_value
, v
->mult_val
,
2884 v
->add_val
, tem
, loop_start
);
2888 splittable_regs
[REGNO (v
->new_reg
)] = value
;
2892 /* Splitting address givs is useful since it will often allow us
2893 to eliminate some increment insns for the base giv as
2896 /* If the addr giv is combined with a dest_reg giv, then all
2897 references to that dest reg will be remapped, which is NOT
2898 what we want for split addr regs. We always create a new
2899 register for the split addr giv, just to be safe. */
2901 /* If we have multiple identical address givs within a
2902 single instruction, then use a single pseudo reg for
2903 both. This is necessary in case one is a match_dup
2906 v
->const_adjust
= 0;
2910 v
->dest_reg
= v
->same_insn
->dest_reg
;
2911 if (loop_dump_stream
)
2912 fprintf (loop_dump_stream
,
2913 "Sharing address givs in insn %d\n",
2914 INSN_UID (v
->insn
));
2916 /* If multiple address GIVs have been combined with the
2917 same dest_reg GIV, do not create a new register for
2919 else if (unroll_type
!= UNROLL_COMPLETELY
2920 && v
->giv_type
== DEST_ADDR
2921 && v
->same
&& v
->same
->giv_type
== DEST_ADDR
2922 && v
->same
->unrolled
2923 /* combine_givs_p may return true for some cases
2924 where the add and mult values are not equal.
2925 To share a register here, the values must be
2927 && rtx_equal_p (v
->same
->mult_val
, v
->mult_val
)
2928 && rtx_equal_p (v
->same
->add_val
, v
->add_val
))
2931 v
->dest_reg
= v
->same
->dest_reg
;
2934 else if (unroll_type
!= UNROLL_COMPLETELY
)
2936 /* If not completely unrolling the loop, then create a new
2937 register to hold the split value of the DEST_ADDR giv.
2938 Emit insn to initialize its value before loop start. */
2940 rtx tem
= gen_reg_rtx (v
->mode
);
2941 record_base_value (REGNO (tem
), v
->add_val
, 0);
2943 /* If the address giv has a constant in its new_reg value,
2944 then this constant can be pulled out and put in value,
2945 instead of being part of the initialization code. */
2947 if (GET_CODE (v
->new_reg
) == PLUS
2948 && GET_CODE (XEXP (v
->new_reg
, 1)) == CONST_INT
)
2951 = plus_constant (tem
, INTVAL (XEXP (v
->new_reg
,1)));
2953 /* Only succeed if this will give valid addresses.
2954 Try to validate both the first and the last
2955 address resulting from loop unrolling, if
2956 one fails, then can't do const elim here. */
2957 if (verify_addresses (v
, giv_inc
, unroll_number
))
2959 /* Save the negative of the eliminated const, so
2960 that we can calculate the dest_reg's increment
2962 v
->const_adjust
= - INTVAL (XEXP (v
->new_reg
, 1));
2964 v
->new_reg
= XEXP (v
->new_reg
, 0);
2965 if (loop_dump_stream
)
2966 fprintf (loop_dump_stream
,
2967 "Eliminating constant from giv %d\n",
2976 /* If the address hasn't been checked for validity yet, do so
2977 now, and fail completely if either the first or the last
2978 unrolled copy of the address is not a valid address
2979 for the instruction that uses it. */
2980 if (v
->dest_reg
== tem
2981 && ! verify_addresses (v
, giv_inc
, unroll_number
))
2983 for (v2
= v
->next_iv
; v2
; v2
= v2
->next_iv
)
2984 if (v2
->same_insn
== v
)
2987 if (loop_dump_stream
)
2988 fprintf (loop_dump_stream
,
2989 "Invalid address for giv at insn %d\n",
2990 INSN_UID (v
->insn
));
2994 /* We set this after the address check, to guarantee that
2995 the register will be initialized. */
2998 /* To initialize the new register, just move the value of
2999 new_reg into it. This is not guaranteed to give a valid
3000 instruction on machines with complex addressing modes.
3001 If we can't recognize it, then delete it and emit insns
3002 to calculate the value from scratch. */
3003 emit_insn_before (gen_rtx_SET (VOIDmode
, tem
,
3004 copy_rtx (v
->new_reg
)),
3006 if (recog_memoized (PREV_INSN (loop_start
)) < 0)
3010 /* We can't use bl->initial_value to compute the initial
3011 value, because the loop may have been preconditioned.
3012 We must calculate it from NEW_REG. Try using
3013 force_operand instead of emit_iv_add_mult. */
3014 delete_insn (PREV_INSN (loop_start
));
3017 ret
= force_operand (v
->new_reg
, tem
);
3019 emit_move_insn (tem
, ret
);
3020 sequence
= gen_sequence ();
3022 emit_insn_before (sequence
, loop_start
);
3024 if (loop_dump_stream
)
3025 fprintf (loop_dump_stream
,
3026 "Invalid init insn, rewritten.\n");
3031 v
->dest_reg
= value
;
3033 /* Check the resulting address for validity, and fail
3034 if the resulting address would be invalid. */
3035 if (! verify_addresses (v
, giv_inc
, unroll_number
))
3037 for (v2
= v
->next_iv
; v2
; v2
= v2
->next_iv
)
3038 if (v2
->same_insn
== v
)
3041 if (loop_dump_stream
)
3042 fprintf (loop_dump_stream
,
3043 "Invalid address for giv at insn %d\n",
3044 INSN_UID (v
->insn
));
3049 /* Store the value of dest_reg into the insn. This sharing
3050 will not be a problem as this insn will always be copied
3053 *v
->location
= v
->dest_reg
;
3055 /* If this address giv is combined with a dest reg giv, then
3056 save the base giv's induction pointer so that we will be
3057 able to handle this address giv properly. The base giv
3058 itself does not have to be splittable. */
3060 if (v
->same
&& v
->same
->giv_type
== DEST_REG
)
3061 addr_combined_regs
[REGNO (v
->same
->new_reg
)] = v
->same
;
3063 if (GET_CODE (v
->new_reg
) == REG
)
3065 /* This giv maybe hasn't been combined with any others.
3066 Make sure that it's giv is marked as splittable here. */
3068 splittable_regs
[REGNO (v
->new_reg
)] = value
;
3070 /* Make it appear to depend upon itself, so that the
3071 giv will be properly split in the main loop above. */
3075 addr_combined_regs
[REGNO (v
->new_reg
)] = v
;
3079 if (loop_dump_stream
)
3080 fprintf (loop_dump_stream
, "DEST_ADDR giv being split.\n");
3086 /* Currently, unreduced giv's can't be split. This is not too much
3087 of a problem since unreduced giv's are not live across loop
3088 iterations anyways. When unrolling a loop completely though,
3089 it makes sense to reduce&split givs when possible, as this will
3090 result in simpler instructions, and will not require that a reg
3091 be live across loop iterations. */
3093 splittable_regs
[REGNO (v
->dest_reg
)] = value
;
3094 fprintf (stderr
, "Giv %d at insn %d not reduced\n",
3095 REGNO (v
->dest_reg
), INSN_UID (v
->insn
));
3101 /* Unreduced givs are only updated once by definition. Reduced givs
3102 are updated as many times as their biv is. Mark it so if this is
3103 a splittable register. Don't need to do anything for address givs
3104 where this may not be a register. */
3106 if (GET_CODE (v
->new_reg
) == REG
)
3110 count
= reg_biv_class
[REGNO (v
->src_reg
)]->biv_count
;
3112 splittable_regs_updates
[REGNO (v
->new_reg
)] = count
;
3117 if (loop_dump_stream
)
3121 if (GET_CODE (v
->dest_reg
) == CONST_INT
)
3123 else if (GET_CODE (v
->dest_reg
) != REG
)
3124 regnum
= REGNO (XEXP (v
->dest_reg
, 0));
3126 regnum
= REGNO (v
->dest_reg
);
3127 fprintf (loop_dump_stream
, "Giv %d at insn %d safe to split.\n",
3128 regnum
, INSN_UID (v
->insn
));
3135 /* Try to prove that the register is dead after the loop exits. Trace every
3136 loop exit looking for an insn that will always be executed, which sets
3137 the register to some value, and appears before the first use of the register
3138 is found. If successful, then return 1, otherwise return 0. */
3140 /* ?? Could be made more intelligent in the handling of jumps, so that
3141 it can search past if statements and other similar structures. */
3144 reg_dead_after_loop (reg
, loop_start
, loop_end
)
3145 rtx reg
, loop_start
, loop_end
;
3150 int label_count
= 0;
3151 int this_loop_num
= uid_loop_num
[INSN_UID (loop_start
)];
3153 /* In addition to checking all exits of this loop, we must also check
3154 all exits of inner nested loops that would exit this loop. We don't
3155 have any way to identify those, so we just give up if there are any
3156 such inner loop exits. */
3158 for (label
= loop_number_exit_labels
[this_loop_num
]; label
;
3159 label
= LABEL_NEXTREF (label
))
3162 if (label_count
!= loop_number_exit_count
[this_loop_num
])
3165 /* HACK: Must also search the loop fall through exit, create a label_ref
3166 here which points to the loop_end, and append the loop_number_exit_labels
3168 label
= gen_rtx_LABEL_REF (VOIDmode
, loop_end
);
3169 LABEL_NEXTREF (label
) = loop_number_exit_labels
[this_loop_num
];
3171 for ( ; label
; label
= LABEL_NEXTREF (label
))
3173 /* Succeed if find an insn which sets the biv or if reach end of
3174 function. Fail if find an insn that uses the biv, or if come to
3175 a conditional jump. */
3177 insn
= NEXT_INSN (XEXP (label
, 0));
3180 code
= GET_CODE (insn
);
3181 if (GET_RTX_CLASS (code
) == 'i')
3185 if (reg_referenced_p (reg
, PATTERN (insn
)))
3188 set
= single_set (insn
);
3189 if (set
&& rtx_equal_p (SET_DEST (set
), reg
))
3193 if (code
== JUMP_INSN
)
3195 if (GET_CODE (PATTERN (insn
)) == RETURN
)
3197 else if (! simplejump_p (insn
)
3198 /* Prevent infinite loop following infinite loops. */
3199 || jump_count
++ > 20)
3202 insn
= JUMP_LABEL (insn
);
3205 insn
= NEXT_INSN (insn
);
3209 /* Success, the register is dead on all loop exits. */
3213 /* Try to calculate the final value of the biv, the value it will have at
3214 the end of the loop. If we can do it, return that value. */
3217 final_biv_value (bl
, loop_start
, loop_end
)
3218 struct iv_class
*bl
;
3219 rtx loop_start
, loop_end
;
3223 /* ??? This only works for MODE_INT biv's. Reject all others for now. */
3225 if (GET_MODE_CLASS (bl
->biv
->mode
) != MODE_INT
)
3228 /* The final value for reversed bivs must be calculated differently than
3229 for ordinary bivs. In this case, there is already an insn after the
3230 loop which sets this biv's final value (if necessary), and there are
3231 no other loop exits, so we can return any value. */
3234 if (loop_dump_stream
)
3235 fprintf (loop_dump_stream
,
3236 "Final biv value for %d, reversed biv.\n", bl
->regno
);
3241 /* Try to calculate the final value as initial value + (number of iterations
3242 * increment). For this to work, increment must be invariant, the only
3243 exit from the loop must be the fall through at the bottom (otherwise
3244 it may not have its final value when the loop exits), and the initial
3245 value of the biv must be invariant. */
3247 if (loop_n_iterations
!= 0
3248 && ! loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]]
3249 && invariant_p (bl
->initial_value
))
3251 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
3253 if (increment
&& invariant_p (increment
))
3255 /* Can calculate the loop exit value, emit insns after loop
3256 end to calculate this value into a temporary register in
3257 case it is needed later. */
3259 tem
= gen_reg_rtx (bl
->biv
->mode
);
3260 record_base_value (REGNO (tem
), bl
->biv
->add_val
, 0);
3261 /* Make sure loop_end is not the last insn. */
3262 if (NEXT_INSN (loop_end
) == 0)
3263 emit_note_after (NOTE_INSN_DELETED
, loop_end
);
3264 emit_iv_add_mult (increment
, GEN_INT (loop_n_iterations
),
3265 bl
->initial_value
, tem
, NEXT_INSN (loop_end
));
3267 if (loop_dump_stream
)
3268 fprintf (loop_dump_stream
,
3269 "Final biv value for %d, calculated.\n", bl
->regno
);
3275 /* Check to see if the biv is dead at all loop exits. */
3276 if (reg_dead_after_loop (bl
->biv
->src_reg
, loop_start
, loop_end
))
3278 if (loop_dump_stream
)
3279 fprintf (loop_dump_stream
,
3280 "Final biv value for %d, biv dead after loop exit.\n",
3289 /* Try to calculate the final value of the giv, the value it will have at
3290 the end of the loop. If we can do it, return that value. */
3293 final_giv_value (v
, loop_start
, loop_end
)
3294 struct induction
*v
;
3295 rtx loop_start
, loop_end
;
3297 struct iv_class
*bl
;
3300 rtx insert_before
, seq
;
3302 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
3304 /* The final value for givs which depend on reversed bivs must be calculated
3305 differently than for ordinary givs. In this case, there is already an
3306 insn after the loop which sets this giv's final value (if necessary),
3307 and there are no other loop exits, so we can return any value. */
3310 if (loop_dump_stream
)
3311 fprintf (loop_dump_stream
,
3312 "Final giv value for %d, depends on reversed biv\n",
3313 REGNO (v
->dest_reg
));
3317 /* Try to calculate the final value as a function of the biv it depends
3318 upon. The only exit from the loop must be the fall through at the bottom
3319 (otherwise it may not have its final value when the loop exits). */
3321 /* ??? Can calculate the final giv value by subtracting off the
3322 extra biv increments times the giv's mult_val. The loop must have
3323 only one exit for this to work, but the loop iterations does not need
3326 if (loop_n_iterations
!= 0
3327 && ! loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]])
3329 /* ?? It is tempting to use the biv's value here since these insns will
3330 be put after the loop, and hence the biv will have its final value
3331 then. However, this fails if the biv is subsequently eliminated.
3332 Perhaps determine whether biv's are eliminable before trying to
3333 determine whether giv's are replaceable so that we can use the
3334 biv value here if it is not eliminable. */
3336 /* We are emitting code after the end of the loop, so we must make
3337 sure that bl->initial_value is still valid then. It will still
3338 be valid if it is invariant. */
3340 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
3342 if (increment
&& invariant_p (increment
)
3343 && invariant_p (bl
->initial_value
))
3345 /* Can calculate the loop exit value of its biv as
3346 (loop_n_iterations * increment) + initial_value */
3348 /* The loop exit value of the giv is then
3349 (final_biv_value - extra increments) * mult_val + add_val.
3350 The extra increments are any increments to the biv which
3351 occur in the loop after the giv's value is calculated.
3352 We must search from the insn that sets the giv to the end
3353 of the loop to calculate this value. */
3355 insert_before
= NEXT_INSN (loop_end
);
3357 /* Put the final biv value in tem. */
3358 tem
= gen_reg_rtx (bl
->biv
->mode
);
3359 record_base_value (REGNO (tem
), bl
->biv
->add_val
, 0);
3360 emit_iv_add_mult (increment
, GEN_INT (loop_n_iterations
),
3361 bl
->initial_value
, tem
, insert_before
);
3363 /* Subtract off extra increments as we find them. */
3364 for (insn
= NEXT_INSN (v
->insn
); insn
!= loop_end
;
3365 insn
= NEXT_INSN (insn
))
3367 struct induction
*biv
;
3369 for (biv
= bl
->biv
; biv
; biv
= biv
->next_iv
)
3370 if (biv
->insn
== insn
)
3373 tem
= expand_binop (GET_MODE (tem
), sub_optab
, tem
,
3374 biv
->add_val
, NULL_RTX
, 0,
3376 seq
= gen_sequence ();
3378 emit_insn_before (seq
, insert_before
);
3382 /* Now calculate the giv's final value. */
3383 emit_iv_add_mult (tem
, v
->mult_val
, v
->add_val
, tem
,
3386 if (loop_dump_stream
)
3387 fprintf (loop_dump_stream
,
3388 "Final giv value for %d, calc from biv's value.\n",
3389 REGNO (v
->dest_reg
));
3395 /* Replaceable giv's should never reach here. */
3399 /* Check to see if the biv is dead at all loop exits. */
3400 if (reg_dead_after_loop (v
->dest_reg
, loop_start
, loop_end
))
3402 if (loop_dump_stream
)
3403 fprintf (loop_dump_stream
,
3404 "Final giv value for %d, giv dead after loop exit.\n",
3405 REGNO (v
->dest_reg
));
3414 /* Calculate the number of loop iterations. Returns the exact number of loop
3415 iterations if it can be calculated, otherwise returns zero. */
3417 unsigned HOST_WIDE_INT
3418 loop_iterations (loop_start
, loop_end
)
3419 rtx loop_start
, loop_end
;
3421 rtx comparison
, comparison_value
;
3422 rtx iteration_var
, initial_value
, increment
, final_value
;
3423 enum rtx_code comparison_code
;
3426 int unsigned_compare
, compare_dir
, final_larger
;
3427 unsigned long tempu
;
3430 /* First find the iteration variable. If the last insn is a conditional
3431 branch, and the insn before tests a register value, make that the
3432 iteration variable. */
3434 loop_initial_value
= 0;
3436 loop_final_value
= 0;
3437 loop_iteration_var
= 0;
3439 /* We used to use pren_nonnote_insn here, but that fails because it might
3440 accidentally get the branch for a contained loop if the branch for this
3441 loop was deleted. We can only trust branches immediately before the
3443 last_loop_insn
= PREV_INSN (loop_end
);
3445 comparison
= get_condition_for_loop (last_loop_insn
);
3446 if (comparison
== 0)
3448 if (loop_dump_stream
)
3449 fprintf (loop_dump_stream
,
3450 "Loop unrolling: No final conditional branch found.\n");
3454 /* ??? Get_condition may switch position of induction variable and
3455 invariant register when it canonicalizes the comparison. */
3457 comparison_code
= GET_CODE (comparison
);
3458 iteration_var
= XEXP (comparison
, 0);
3459 comparison_value
= XEXP (comparison
, 1);
3461 if (GET_CODE (iteration_var
) != REG
)
3463 if (loop_dump_stream
)
3464 fprintf (loop_dump_stream
,
3465 "Loop unrolling: Comparison not against register.\n");
3469 /* Loop iterations is always called before any new registers are created
3470 now, so this should never occur. */
3472 if (REGNO (iteration_var
) >= max_reg_before_loop
)
3475 iteration_info (iteration_var
, &initial_value
, &increment
,
3476 loop_start
, loop_end
);
3477 if (initial_value
== 0)
3478 /* iteration_info already printed a message. */
3481 /* If the comparison value is an invariant register, then try to find
3482 its value from the insns before the start of the loop. */
3484 if (GET_CODE (comparison_value
) == REG
&& invariant_p (comparison_value
))
3488 for (insn
= PREV_INSN (loop_start
); insn
; insn
= PREV_INSN (insn
))
3490 if (GET_CODE (insn
) == CODE_LABEL
)
3493 else if (GET_RTX_CLASS (GET_CODE (insn
)) == 'i'
3494 && reg_set_p (comparison_value
, insn
))
3496 /* We found the last insn before the loop that sets the register.
3497 If it sets the entire register, and has a REG_EQUAL note,
3498 then use the value of the REG_EQUAL note. */
3499 if ((set
= single_set (insn
))
3500 && (SET_DEST (set
) == comparison_value
))
3502 rtx note
= find_reg_note (insn
, REG_EQUAL
, NULL_RTX
);
3504 /* Only use the REG_EQUAL note if it is a constant.
3505 Other things, divide in particular, will cause
3506 problems later if we use them. */
3507 if (note
&& GET_CODE (XEXP (note
, 0)) != EXPR_LIST
3508 && CONSTANT_P (XEXP (note
, 0)))
3509 comparison_value
= XEXP (note
, 0);
3516 final_value
= approx_final_value (comparison_code
, comparison_value
,
3517 &unsigned_compare
, &compare_dir
);
3519 /* Save the calculated values describing this loop's bounds, in case
3520 precondition_loop_p will need them later. These values can not be
3521 recalculated inside precondition_loop_p because strength reduction
3522 optimizations may obscure the loop's structure. */
3524 loop_iteration_var
= iteration_var
;
3525 loop_initial_value
= initial_value
;
3526 loop_increment
= increment
;
3527 loop_final_value
= final_value
;
3528 loop_comparison_code
= comparison_code
;
3532 if (loop_dump_stream
)
3533 fprintf (loop_dump_stream
,
3534 "Loop unrolling: Increment value can't be calculated.\n");
3537 else if (GET_CODE (increment
) != CONST_INT
)
3539 if (loop_dump_stream
)
3540 fprintf (loop_dump_stream
,
3541 "Loop unrolling: Increment value not constant.\n");
3544 else if (GET_CODE (initial_value
) != CONST_INT
)
3546 if (loop_dump_stream
)
3547 fprintf (loop_dump_stream
,
3548 "Loop unrolling: Initial value not constant.\n");
3551 else if (final_value
== 0)
3553 if (loop_dump_stream
)
3554 fprintf (loop_dump_stream
,
3555 "Loop unrolling: EQ comparison loop.\n");
3558 else if (GET_CODE (final_value
) != CONST_INT
)
3560 if (loop_dump_stream
)
3561 fprintf (loop_dump_stream
,
3562 "Loop unrolling: Final value not constant.\n");
3566 /* ?? Final value and initial value do not have to be constants.
3567 Only their difference has to be constant. When the iteration variable
3568 is an array address, the final value and initial value might both
3569 be addresses with the same base but different constant offsets.
3570 Final value must be invariant for this to work.
3572 To do this, need some way to find the values of registers which are
3575 /* Final_larger is 1 if final larger, 0 if they are equal, otherwise -1. */
3576 if (unsigned_compare
)
3578 = ((unsigned HOST_WIDE_INT
) INTVAL (final_value
)
3579 > (unsigned HOST_WIDE_INT
) INTVAL (initial_value
))
3580 - ((unsigned HOST_WIDE_INT
) INTVAL (final_value
)
3581 < (unsigned HOST_WIDE_INT
) INTVAL (initial_value
));
3583 final_larger
= (INTVAL (final_value
) > INTVAL (initial_value
))
3584 - (INTVAL (final_value
) < INTVAL (initial_value
));
3586 if (INTVAL (increment
) > 0)
3588 else if (INTVAL (increment
) == 0)
3593 /* There are 27 different cases: compare_dir = -1, 0, 1;
3594 final_larger = -1, 0, 1; increment_dir = -1, 0, 1.
3595 There are 4 normal cases, 4 reverse cases (where the iteration variable
3596 will overflow before the loop exits), 4 infinite loop cases, and 15
3597 immediate exit (0 or 1 iteration depending on loop type) cases.
3598 Only try to optimize the normal cases. */
3600 /* (compare_dir/final_larger/increment_dir)
3601 Normal cases: (0/-1/-1), (0/1/1), (-1/-1/-1), (1/1/1)
3602 Reverse cases: (0/-1/1), (0/1/-1), (-1/-1/1), (1/1/-1)
3603 Infinite loops: (0/-1/0), (0/1/0), (-1/-1/0), (1/1/0)
3604 Immediate exit: (0/0/X), (-1/0/X), (-1/1/X), (1/0/X), (1/-1/X) */
3606 /* ?? If the meaning of reverse loops (where the iteration variable
3607 will overflow before the loop exits) is undefined, then could
3608 eliminate all of these special checks, and just always assume
3609 the loops are normal/immediate/infinite. Note that this means
3610 the sign of increment_dir does not have to be known. Also,
3611 since it does not really hurt if immediate exit loops or infinite loops
3612 are optimized, then that case could be ignored also, and hence all
3613 loops can be optimized.
3615 According to ANSI Spec, the reverse loop case result is undefined,
3616 because the action on overflow is undefined.
3618 See also the special test for NE loops below. */
3620 if (final_larger
== increment_dir
&& final_larger
!= 0
3621 && (final_larger
== compare_dir
|| compare_dir
== 0))
3626 if (loop_dump_stream
)
3627 fprintf (loop_dump_stream
,
3628 "Loop unrolling: Not normal loop.\n");
3632 /* Calculate the number of iterations, final_value is only an approximation,
3633 so correct for that. Note that tempu and loop_n_iterations are
3634 unsigned, because they can be as large as 2^n - 1. */
3636 i
= INTVAL (increment
);
3638 tempu
= INTVAL (final_value
) - INTVAL (initial_value
);
3641 tempu
= INTVAL (initial_value
) - INTVAL (final_value
);
3647 /* For NE tests, make sure that the iteration variable won't miss the
3648 final value. If tempu mod i is not zero, then the iteration variable
3649 will overflow before the loop exits, and we can not calculate the
3650 number of iterations. */
3651 if (compare_dir
== 0 && (tempu
% i
) != 0)
3654 return tempu
/ i
+ ((tempu
% i
) != 0);
3657 /* Replace uses of split bivs with their split pseudo register. This is
3658 for original instructions which remain after loop unrolling without
3662 remap_split_bivs (x
)
3665 register enum rtx_code code
;
3672 code
= GET_CODE (x
);
3687 /* If non-reduced/final-value givs were split, then this would also
3688 have to remap those givs also. */
3690 if (REGNO (x
) < max_reg_before_loop
3691 && reg_iv_type
[REGNO (x
)] == BASIC_INDUCT
)
3692 return reg_biv_class
[REGNO (x
)]->biv
->src_reg
;
3699 fmt
= GET_RTX_FORMAT (code
);
3700 for (i
= GET_RTX_LENGTH (code
) - 1; i
>= 0; i
--)
3703 XEXP (x
, i
) = remap_split_bivs (XEXP (x
, i
));
3707 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
3708 XVECEXP (x
, i
, j
) = remap_split_bivs (XVECEXP (x
, i
, j
));
3714 /* If FIRST_UID is a set of REGNO, and FIRST_UID dominates LAST_UID (e.g.
3715 FIST_UID is always executed if LAST_UID is), then return 1. Otherwise
3716 return 0. COPY_START is where we can start looking for the insns
3717 FIRST_UID and LAST_UID. COPY_END is where we stop looking for these
3720 If there is no JUMP_INSN between LOOP_START and FIRST_UID, then FIRST_UID
3721 must dominate LAST_UID.
3723 If there is a CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID
3724 may not dominate LAST_UID.
3726 If there is no CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID
3727 must dominate LAST_UID. */
3730 set_dominates_use (regno
, first_uid
, last_uid
, copy_start
, copy_end
)
3737 int passed_jump
= 0;
3738 rtx p
= NEXT_INSN (copy_start
);
3740 while (INSN_UID (p
) != first_uid
)
3742 if (GET_CODE (p
) == JUMP_INSN
)
3744 /* Could not find FIRST_UID. */
3750 /* Verify that FIRST_UID is an insn that entirely sets REGNO. */
3751 if (GET_RTX_CLASS (GET_CODE (p
)) != 'i'
3752 || ! dead_or_set_regno_p (p
, regno
))
3755 /* FIRST_UID is always executed. */
3756 if (passed_jump
== 0)
3759 while (INSN_UID (p
) != last_uid
)
3761 /* If we see a CODE_LABEL between FIRST_UID and LAST_UID, then we
3762 can not be sure that FIRST_UID dominates LAST_UID. */
3763 if (GET_CODE (p
) == CODE_LABEL
)
3765 /* Could not find LAST_UID, but we reached the end of the loop, so
3767 else if (p
== copy_end
)
3772 /* FIRST_UID is always executed if LAST_UID is executed. */