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 /* Forward declarations. */
191 static void init_reg_map
PROTO((struct inline_remap
*, int));
192 static rtx calculate_giv_inc
PROTO((rtx
, rtx
, int));
193 static rtx initial_reg_note_copy
PROTO((rtx
, struct inline_remap
*));
194 static void final_reg_note_copy
PROTO((rtx
, struct inline_remap
*));
195 static void copy_loop_body
PROTO((rtx
, rtx
, struct inline_remap
*, rtx
, int,
196 enum unroll_types
, rtx
, rtx
, rtx
, rtx
));
197 static void iteration_info
PROTO((rtx
, rtx
*, rtx
*, rtx
, rtx
));
198 static int find_splittable_regs
PROTO((enum unroll_types
, rtx
, rtx
, rtx
, int,
199 unsigned HOST_WIDE_INT
));
200 static int find_splittable_givs
PROTO((struct iv_class
*, enum unroll_types
,
201 rtx
, rtx
, rtx
, int));
202 static int reg_dead_after_loop
PROTO((rtx
, rtx
, rtx
));
203 static rtx fold_rtx_mult_add
PROTO((rtx
, rtx
, rtx
, enum machine_mode
));
204 static int verify_addresses
PROTO((struct induction
*, rtx
, int));
205 static rtx remap_split_bivs
PROTO((rtx
));
207 /* Try to unroll one loop and split induction variables in the loop.
209 The loop is described by the arguments LOOP_END, INSN_COUNT, and
210 LOOP_START. END_INSERT_BEFORE indicates where insns should be added
211 which need to be executed when the loop falls through. STRENGTH_REDUCTION_P
212 indicates whether information generated in the strength reduction pass
215 This function is intended to be called from within `strength_reduce'
219 unroll_loop (loop_end
, insn_count
, loop_start
, end_insert_before
,
220 loop_info
, strength_reduce_p
)
224 rtx end_insert_before
;
225 struct loop_info
*loop_info
;
226 int strength_reduce_p
;
229 int unroll_number
= 1;
230 rtx copy_start
, copy_end
;
231 rtx insn
, sequence
, pattern
, tem
;
232 int max_labelno
, max_insnno
;
234 struct inline_remap
*map
;
242 int splitting_not_safe
= 0;
243 enum unroll_types unroll_type
;
244 int loop_preconditioned
= 0;
246 /* This points to the last real insn in the loop, which should be either
247 a JUMP_INSN (for conditional jumps) or a BARRIER (for unconditional
251 /* Don't bother unrolling huge loops. Since the minimum factor is
252 two, loops greater than one half of MAX_UNROLLED_INSNS will never
254 if (insn_count
> MAX_UNROLLED_INSNS
/ 2)
256 if (loop_dump_stream
)
257 fprintf (loop_dump_stream
, "Unrolling failure: Loop too big.\n");
261 /* When emitting debugger info, we can't unroll loops with unequal numbers
262 of block_beg and block_end notes, because that would unbalance the block
263 structure of the function. This can happen as a result of the
264 "if (foo) bar; else break;" optimization in jump.c. */
265 /* ??? Gcc has a general policy that -g is never supposed to change the code
266 that the compiler emits, so we must disable this optimization always,
267 even if debug info is not being output. This is rare, so this should
268 not be a significant performance problem. */
270 if (1 /* write_symbols != NO_DEBUG */)
272 int block_begins
= 0;
275 for (insn
= loop_start
; insn
!= loop_end
; insn
= NEXT_INSN (insn
))
277 if (GET_CODE (insn
) == NOTE
)
279 if (NOTE_LINE_NUMBER (insn
) == NOTE_INSN_BLOCK_BEG
)
281 else if (NOTE_LINE_NUMBER (insn
) == NOTE_INSN_BLOCK_END
)
286 if (block_begins
!= block_ends
)
288 if (loop_dump_stream
)
289 fprintf (loop_dump_stream
,
290 "Unrolling failure: Unbalanced block notes.\n");
295 /* Determine type of unroll to perform. Depends on the number of iterations
296 and the size of the loop. */
298 /* If there is no strength reduce info, then set
299 loop_info->n_iterations to zero. This can happen if
300 strength_reduce can't find any bivs in the loop. A value of zero
301 indicates that the number of iterations could not be calculated. */
303 if (! strength_reduce_p
)
304 loop_info
->n_iterations
= 0;
306 if (loop_dump_stream
&& loop_info
->n_iterations
> 0)
308 fputs ("Loop unrolling: ", loop_dump_stream
);
309 fprintf (loop_dump_stream
, HOST_WIDE_INT_PRINT_DEC
,
310 loop_info
->n_iterations
);
311 fputs (" iterations.\n", loop_dump_stream
);
314 /* Find and save a pointer to the last nonnote insn in the loop. */
316 last_loop_insn
= prev_nonnote_insn (loop_end
);
318 /* Calculate how many times to unroll the loop. Indicate whether or
319 not the loop is being completely unrolled. */
321 if (loop_info
->n_iterations
== 1)
323 /* If number of iterations is exactly 1, then eliminate the compare and
324 branch at the end of the loop since they will never be taken.
325 Then return, since no other action is needed here. */
327 /* If the last instruction is not a BARRIER or a JUMP_INSN, then
328 don't do anything. */
330 if (GET_CODE (last_loop_insn
) == BARRIER
)
332 /* Delete the jump insn. This will delete the barrier also. */
333 delete_insn (PREV_INSN (last_loop_insn
));
335 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
338 /* The immediately preceding insn is a compare which must be
340 delete_insn (last_loop_insn
);
341 delete_insn (PREV_INSN (last_loop_insn
));
343 /* The immediately preceding insn may not be the compare, so don't
345 delete_insn (last_loop_insn
);
350 else if (loop_info
->n_iterations
> 0
351 && loop_info
->n_iterations
* insn_count
< MAX_UNROLLED_INSNS
)
353 unroll_number
= loop_info
->n_iterations
;
354 unroll_type
= UNROLL_COMPLETELY
;
356 else if (loop_info
->n_iterations
> 0)
358 /* Try to factor the number of iterations. Don't bother with the
359 general case, only using 2, 3, 5, and 7 will get 75% of all
360 numbers theoretically, and almost all in practice. */
362 for (i
= 0; i
< NUM_FACTORS
; i
++)
363 factors
[i
].count
= 0;
365 temp
= loop_info
->n_iterations
;
366 for (i
= NUM_FACTORS
- 1; i
>= 0; i
--)
367 while (temp
% factors
[i
].factor
== 0)
370 temp
= temp
/ factors
[i
].factor
;
373 /* Start with the larger factors first so that we generally
374 get lots of unrolling. */
378 for (i
= 3; i
>= 0; i
--)
379 while (factors
[i
].count
--)
381 if (temp
* factors
[i
].factor
< MAX_UNROLLED_INSNS
)
383 unroll_number
*= factors
[i
].factor
;
384 temp
*= factors
[i
].factor
;
390 /* If we couldn't find any factors, then unroll as in the normal
392 if (unroll_number
== 1)
394 if (loop_dump_stream
)
395 fprintf (loop_dump_stream
,
396 "Loop unrolling: No factors found.\n");
399 unroll_type
= UNROLL_MODULO
;
403 /* Default case, calculate number of times to unroll loop based on its
405 if (unroll_number
== 1)
407 if (8 * insn_count
< MAX_UNROLLED_INSNS
)
409 else if (4 * insn_count
< MAX_UNROLLED_INSNS
)
414 unroll_type
= UNROLL_NAIVE
;
417 /* Now we know how many times to unroll the loop. */
419 if (loop_dump_stream
)
420 fprintf (loop_dump_stream
,
421 "Unrolling loop %d times.\n", unroll_number
);
424 if (unroll_type
== UNROLL_COMPLETELY
|| unroll_type
== UNROLL_MODULO
)
426 /* Loops of these types can start with jump down to the exit condition
427 in rare circumstances.
429 Consider a pair of nested loops where the inner loop is part
430 of the exit code for the outer loop.
432 In this case jump.c will not duplicate the exit test for the outer
433 loop, so it will start with a jump to the exit code.
435 Then consider if the inner loop turns out to iterate once and
436 only once. We will end up deleting the jumps associated with
437 the inner loop. However, the loop notes are not removed from
438 the instruction stream.
440 And finally assume that we can compute the number of iterations
443 In this case unroll may want to unroll the outer loop even though
444 it starts with a jump to the outer loop's exit code.
446 We could try to optimize this case, but it hardly seems worth it.
447 Just return without unrolling the loop in such cases. */
450 while (GET_CODE (insn
) != CODE_LABEL
&& GET_CODE (insn
) != JUMP_INSN
)
451 insn
= NEXT_INSN (insn
);
452 if (GET_CODE (insn
) == JUMP_INSN
)
456 if (unroll_type
== UNROLL_COMPLETELY
)
458 /* Completely unrolling the loop: Delete the compare and branch at
459 the end (the last two instructions). This delete must done at the
460 very end of loop unrolling, to avoid problems with calls to
461 back_branch_in_range_p, which is called by find_splittable_regs.
462 All increments of splittable bivs/givs are changed to load constant
465 copy_start
= loop_start
;
467 /* Set insert_before to the instruction immediately after the JUMP_INSN
468 (or BARRIER), so that any NOTEs between the JUMP_INSN and the end of
469 the loop will be correctly handled by copy_loop_body. */
470 insert_before
= NEXT_INSN (last_loop_insn
);
472 /* Set copy_end to the insn before the jump at the end of the loop. */
473 if (GET_CODE (last_loop_insn
) == BARRIER
)
474 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
475 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
478 /* The instruction immediately before the JUMP_INSN is a compare
479 instruction which we do not want to copy. */
480 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
482 /* The instruction immediately before the JUMP_INSN may not be the
483 compare, so we must copy it. */
484 copy_end
= PREV_INSN (last_loop_insn
);
489 /* We currently can't unroll a loop if it doesn't end with a
490 JUMP_INSN. There would need to be a mechanism that recognizes
491 this case, and then inserts a jump after each loop body, which
492 jumps to after the last loop body. */
493 if (loop_dump_stream
)
494 fprintf (loop_dump_stream
,
495 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
499 else if (unroll_type
== UNROLL_MODULO
)
501 /* Partially unrolling the loop: The compare and branch at the end
502 (the last two instructions) must remain. Don't copy the compare
503 and branch instructions at the end of the loop. Insert the unrolled
504 code immediately before the compare/branch at the end so that the
505 code will fall through to them as before. */
507 copy_start
= loop_start
;
509 /* Set insert_before to the jump insn at the end of the loop.
510 Set copy_end to before the jump insn at the end of the loop. */
511 if (GET_CODE (last_loop_insn
) == BARRIER
)
513 insert_before
= PREV_INSN (last_loop_insn
);
514 copy_end
= PREV_INSN (insert_before
);
516 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
519 /* The instruction immediately before the JUMP_INSN is a compare
520 instruction which we do not want to copy or delete. */
521 insert_before
= PREV_INSN (last_loop_insn
);
522 copy_end
= PREV_INSN (insert_before
);
524 /* The instruction immediately before the JUMP_INSN may not be the
525 compare, so we must copy it. */
526 insert_before
= last_loop_insn
;
527 copy_end
= PREV_INSN (last_loop_insn
);
532 /* We currently can't unroll a loop if it doesn't end with a
533 JUMP_INSN. There would need to be a mechanism that recognizes
534 this case, and then inserts a jump after each loop body, which
535 jumps to after the last loop body. */
536 if (loop_dump_stream
)
537 fprintf (loop_dump_stream
,
538 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
544 /* Normal case: Must copy the compare and branch instructions at the
547 if (GET_CODE (last_loop_insn
) == BARRIER
)
549 /* Loop ends with an unconditional jump and a barrier.
550 Handle this like above, don't copy jump and barrier.
551 This is not strictly necessary, but doing so prevents generating
552 unconditional jumps to an immediately following label.
554 This will be corrected below if the target of this jump is
555 not the start_label. */
557 insert_before
= PREV_INSN (last_loop_insn
);
558 copy_end
= PREV_INSN (insert_before
);
560 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
562 /* Set insert_before to immediately after the JUMP_INSN, so that
563 NOTEs at the end of the loop will be correctly handled by
565 insert_before
= NEXT_INSN (last_loop_insn
);
566 copy_end
= last_loop_insn
;
570 /* We currently can't unroll a loop if it doesn't end with a
571 JUMP_INSN. There would need to be a mechanism that recognizes
572 this case, and then inserts a jump after each loop body, which
573 jumps to after the last loop body. */
574 if (loop_dump_stream
)
575 fprintf (loop_dump_stream
,
576 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
580 /* If copying exit test branches because they can not be eliminated,
581 then must convert the fall through case of the branch to a jump past
582 the end of the loop. Create a label to emit after the loop and save
583 it for later use. Do not use the label after the loop, if any, since
584 it might be used by insns outside the loop, or there might be insns
585 added before it later by final_[bg]iv_value which must be after
586 the real exit label. */
587 exit_label
= gen_label_rtx ();
590 while (GET_CODE (insn
) != CODE_LABEL
&& GET_CODE (insn
) != JUMP_INSN
)
591 insn
= NEXT_INSN (insn
);
593 if (GET_CODE (insn
) == JUMP_INSN
)
595 /* The loop starts with a jump down to the exit condition test.
596 Start copying the loop after the barrier following this
598 copy_start
= NEXT_INSN (insn
);
600 /* Splitting induction variables doesn't work when the loop is
601 entered via a jump to the bottom, because then we end up doing
602 a comparison against a new register for a split variable, but
603 we did not execute the set insn for the new register because
604 it was skipped over. */
605 splitting_not_safe
= 1;
606 if (loop_dump_stream
)
607 fprintf (loop_dump_stream
,
608 "Splitting not safe, because loop not entered at top.\n");
611 copy_start
= loop_start
;
614 /* This should always be the first label in the loop. */
615 start_label
= NEXT_INSN (copy_start
);
616 /* There may be a line number note and/or a loop continue note here. */
617 while (GET_CODE (start_label
) == NOTE
)
618 start_label
= NEXT_INSN (start_label
);
619 if (GET_CODE (start_label
) != CODE_LABEL
)
621 /* This can happen as a result of jump threading. If the first insns in
622 the loop test the same condition as the loop's backward jump, or the
623 opposite condition, then the backward jump will be modified to point
624 to elsewhere, and the loop's start label is deleted.
626 This case currently can not be handled by the loop unrolling code. */
628 if (loop_dump_stream
)
629 fprintf (loop_dump_stream
,
630 "Unrolling failure: unknown insns between BEG note and loop label.\n");
633 if (LABEL_NAME (start_label
))
635 /* The jump optimization pass must have combined the original start label
636 with a named label for a goto. We can't unroll this case because
637 jumps which go to the named label must be handled differently than
638 jumps to the loop start, and it is impossible to differentiate them
640 if (loop_dump_stream
)
641 fprintf (loop_dump_stream
,
642 "Unrolling failure: loop start label is gone\n");
646 if (unroll_type
== UNROLL_NAIVE
647 && GET_CODE (last_loop_insn
) == BARRIER
648 && start_label
!= JUMP_LABEL (PREV_INSN (last_loop_insn
)))
650 /* In this case, we must copy the jump and barrier, because they will
651 not be converted to jumps to an immediately following label. */
653 insert_before
= NEXT_INSN (last_loop_insn
);
654 copy_end
= last_loop_insn
;
657 if (unroll_type
== UNROLL_NAIVE
658 && GET_CODE (last_loop_insn
) == JUMP_INSN
659 && start_label
!= JUMP_LABEL (last_loop_insn
))
661 /* ??? The loop ends with a conditional branch that does not branch back
662 to the loop start label. In this case, we must emit an unconditional
663 branch to the loop exit after emitting the final branch.
664 copy_loop_body does not have support for this currently, so we
665 give up. It doesn't seem worthwhile to unroll anyways since
666 unrolling would increase the number of branch instructions
668 if (loop_dump_stream
)
669 fprintf (loop_dump_stream
,
670 "Unrolling failure: final conditional branch not to loop start\n");
674 /* Allocate a translation table for the labels and insn numbers.
675 They will be filled in as we copy the insns in the loop. */
677 max_labelno
= max_label_num ();
678 max_insnno
= get_max_uid ();
680 map
= (struct inline_remap
*) alloca (sizeof (struct inline_remap
));
682 map
->integrating
= 0;
684 /* Allocate the label map. */
688 map
->label_map
= (rtx
*) alloca (max_labelno
* sizeof (rtx
));
690 local_label
= (char *) alloca (max_labelno
);
691 bzero (local_label
, max_labelno
);
696 /* Search the loop and mark all local labels, i.e. the ones which have to
697 be distinct labels when copied. For all labels which might be
698 non-local, set their label_map entries to point to themselves.
699 If they happen to be local their label_map entries will be overwritten
700 before the loop body is copied. The label_map entries for local labels
701 will be set to a different value each time the loop body is copied. */
703 for (insn
= copy_start
; insn
!= loop_end
; insn
= NEXT_INSN (insn
))
707 if (GET_CODE (insn
) == CODE_LABEL
)
708 local_label
[CODE_LABEL_NUMBER (insn
)] = 1;
709 else if (GET_CODE (insn
) == JUMP_INSN
)
711 if (JUMP_LABEL (insn
))
712 set_label_in_map (map
,
713 CODE_LABEL_NUMBER (JUMP_LABEL (insn
)),
715 else if (GET_CODE (PATTERN (insn
)) == ADDR_VEC
716 || GET_CODE (PATTERN (insn
)) == ADDR_DIFF_VEC
)
718 rtx pat
= PATTERN (insn
);
719 int diff_vec_p
= GET_CODE (PATTERN (insn
)) == ADDR_DIFF_VEC
;
720 int len
= XVECLEN (pat
, diff_vec_p
);
723 for (i
= 0; i
< len
; i
++)
725 label
= XEXP (XVECEXP (pat
, diff_vec_p
, i
), 0);
726 set_label_in_map (map
,
727 CODE_LABEL_NUMBER (label
),
732 else if ((note
= find_reg_note (insn
, REG_LABEL
, NULL_RTX
)))
733 set_label_in_map (map
, CODE_LABEL_NUMBER (XEXP (note
, 0)),
737 /* Allocate space for the insn map. */
739 map
->insn_map
= (rtx
*) alloca (max_insnno
* sizeof (rtx
));
741 /* Set this to zero, to indicate that we are doing loop unrolling,
742 not function inlining. */
743 map
->inline_target
= 0;
745 /* The register and constant maps depend on the number of registers
746 present, so the final maps can't be created until after
747 find_splittable_regs is called. However, they are needed for
748 preconditioning, so we create temporary maps when preconditioning
751 /* The preconditioning code may allocate two new pseudo registers. */
752 maxregnum
= max_reg_num ();
754 /* Allocate and zero out the splittable_regs and addr_combined_regs
755 arrays. These must be zeroed here because they will be used if
756 loop preconditioning is performed, and must be zero for that case.
758 It is safe to do this here, since the extra registers created by the
759 preconditioning code and find_splittable_regs will never be used
760 to access the splittable_regs[] and addr_combined_regs[] arrays. */
762 splittable_regs
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
763 bzero ((char *) splittable_regs
, maxregnum
* sizeof (rtx
));
764 splittable_regs_updates
= (int *) alloca (maxregnum
* sizeof (int));
765 bzero ((char *) splittable_regs_updates
, maxregnum
* sizeof (int));
767 = (struct induction
**) alloca (maxregnum
* sizeof (struct induction
*));
768 bzero ((char *) addr_combined_regs
, maxregnum
* sizeof (struct induction
*));
769 /* We must limit it to max_reg_before_loop, because only these pseudo
770 registers have valid regno_first_uid info. Any register created after
771 that is unlikely to be local to the loop anyways. */
772 local_regno
= (char *) alloca (max_reg_before_loop
);
773 bzero (local_regno
, max_reg_before_loop
);
775 /* Mark all local registers, i.e. the ones which are referenced only
777 if (INSN_UID (copy_end
) < max_uid_for_loop
)
779 int copy_start_luid
= INSN_LUID (copy_start
);
780 int copy_end_luid
= INSN_LUID (copy_end
);
782 /* If a register is used in the jump insn, we must not duplicate it
783 since it will also be used outside the loop. */
784 if (GET_CODE (copy_end
) == JUMP_INSN
)
786 /* If copy_start points to the NOTE that starts the loop, then we must
787 use the next luid, because invariant pseudo-regs moved out of the loop
788 have their lifetimes modified to start here, but they are not safe
790 if (copy_start
== loop_start
)
793 /* If a pseudo's lifetime is entirely contained within this loop, then we
794 can use a different pseudo in each unrolled copy of the loop. This
795 results in better code. */
796 for (j
= FIRST_PSEUDO_REGISTER
; j
< max_reg_before_loop
; ++j
)
797 if (REGNO_FIRST_UID (j
) > 0 && REGNO_FIRST_UID (j
) <= max_uid_for_loop
798 && uid_luid
[REGNO_FIRST_UID (j
)] >= copy_start_luid
799 && REGNO_LAST_UID (j
) > 0 && REGNO_LAST_UID (j
) <= max_uid_for_loop
800 && uid_luid
[REGNO_LAST_UID (j
)] <= copy_end_luid
)
802 /* However, we must also check for loop-carried dependencies.
803 If the value the pseudo has at the end of iteration X is
804 used by iteration X+1, then we can not use a different pseudo
805 for each unrolled copy of the loop. */
806 /* A pseudo is safe if regno_first_uid is a set, and this
807 set dominates all instructions from regno_first_uid to
809 /* ??? This check is simplistic. We would get better code if
810 this check was more sophisticated. */
811 if (set_dominates_use (j
, REGNO_FIRST_UID (j
), REGNO_LAST_UID (j
),
812 copy_start
, copy_end
))
815 if (loop_dump_stream
)
818 fprintf (loop_dump_stream
, "Marked reg %d as local\n", j
);
820 fprintf (loop_dump_stream
, "Did not mark reg %d as local\n",
826 /* If this loop requires exit tests when unrolled, check to see if we
827 can precondition the loop so as to make the exit tests unnecessary.
828 Just like variable splitting, this is not safe if the loop is entered
829 via a jump to the bottom. Also, can not do this if no strength
830 reduce info, because precondition_loop_p uses this info. */
832 /* Must copy the loop body for preconditioning before the following
833 find_splittable_regs call since that will emit insns which need to
834 be after the preconditioned loop copies, but immediately before the
835 unrolled loop copies. */
837 /* Also, it is not safe to split induction variables for the preconditioned
838 copies of the loop body. If we split induction variables, then the code
839 assumes that each induction variable can be represented as a function
840 of its initial value and the loop iteration number. This is not true
841 in this case, because the last preconditioned copy of the loop body
842 could be any iteration from the first up to the `unroll_number-1'th,
843 depending on the initial value of the iteration variable. Therefore
844 we can not split induction variables here, because we can not calculate
845 their value. Hence, this code must occur before find_splittable_regs
848 if (unroll_type
== UNROLL_NAIVE
&& ! splitting_not_safe
&& strength_reduce_p
)
850 rtx initial_value
, final_value
, increment
;
851 enum machine_mode mode
;
853 if (precondition_loop_p (loop_start
, loop_info
,
854 &initial_value
, &final_value
, &increment
,
859 int abs_inc
, neg_inc
;
861 map
->reg_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
863 map
->const_equiv_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
864 map
->const_age_map
= (unsigned *) alloca (maxregnum
865 * sizeof (unsigned));
866 map
->const_equiv_map_size
= maxregnum
;
867 global_const_equiv_map
= map
->const_equiv_map
;
868 global_const_equiv_map_size
= maxregnum
;
870 init_reg_map (map
, maxregnum
);
872 /* Limit loop unrolling to 4, since this will make 7 copies of
874 if (unroll_number
> 4)
877 /* Save the absolute value of the increment, and also whether or
878 not it is negative. */
880 abs_inc
= INTVAL (increment
);
889 /* Calculate the difference between the final and initial values.
890 Final value may be a (plus (reg x) (const_int 1)) rtx.
891 Let the following cse pass simplify this if initial value is
894 We must copy the final and initial values here to avoid
895 improperly shared rtl. */
897 diff
= expand_binop (mode
, sub_optab
, copy_rtx (final_value
),
898 copy_rtx (initial_value
), NULL_RTX
, 0,
901 /* Now calculate (diff % (unroll * abs (increment))) by using an
903 diff
= expand_binop (GET_MODE (diff
), and_optab
, diff
,
904 GEN_INT (unroll_number
* abs_inc
- 1),
905 NULL_RTX
, 0, OPTAB_LIB_WIDEN
);
907 /* Now emit a sequence of branches to jump to the proper precond
910 labels
= (rtx
*) alloca (sizeof (rtx
) * unroll_number
);
911 for (i
= 0; i
< unroll_number
; i
++)
912 labels
[i
] = gen_label_rtx ();
914 /* Check for the case where the initial value is greater than or
915 equal to the final value. In that case, we want to execute
916 exactly one loop iteration. The code below will fail for this
917 case. This check does not apply if the loop has a NE
918 comparison at the end. */
920 if (loop_info
->comparison_code
!= NE
)
922 emit_cmp_insn (initial_value
, final_value
, neg_inc
? LE
: GE
,
923 NULL_RTX
, mode
, 0, 0);
925 emit_jump_insn (gen_ble (labels
[1]));
927 emit_jump_insn (gen_bge (labels
[1]));
928 JUMP_LABEL (get_last_insn ()) = labels
[1];
929 LABEL_NUSES (labels
[1])++;
932 /* Assuming the unroll_number is 4, and the increment is 2, then
933 for a negative increment: for a positive increment:
934 diff = 0,1 precond 0 diff = 0,7 precond 0
935 diff = 2,3 precond 3 diff = 1,2 precond 1
936 diff = 4,5 precond 2 diff = 3,4 precond 2
937 diff = 6,7 precond 1 diff = 5,6 precond 3 */
939 /* We only need to emit (unroll_number - 1) branches here, the
940 last case just falls through to the following code. */
942 /* ??? This would give better code if we emitted a tree of branches
943 instead of the current linear list of branches. */
945 for (i
= 0; i
< unroll_number
- 1; i
++)
948 enum rtx_code cmp_code
;
950 /* For negative increments, must invert the constant compared
951 against, except when comparing against zero. */
959 cmp_const
= unroll_number
- i
;
968 emit_cmp_insn (diff
, GEN_INT (abs_inc
* cmp_const
),
969 cmp_code
, NULL_RTX
, mode
, 0, 0);
972 emit_jump_insn (gen_beq (labels
[i
]));
974 emit_jump_insn (gen_bge (labels
[i
]));
976 emit_jump_insn (gen_ble (labels
[i
]));
977 JUMP_LABEL (get_last_insn ()) = labels
[i
];
978 LABEL_NUSES (labels
[i
])++;
981 /* If the increment is greater than one, then we need another branch,
982 to handle other cases equivalent to 0. */
984 /* ??? This should be merged into the code above somehow to help
985 simplify the code here, and reduce the number of branches emitted.
986 For the negative increment case, the branch here could easily
987 be merged with the `0' case branch above. For the positive
988 increment case, it is not clear how this can be simplified. */
993 enum rtx_code cmp_code
;
997 cmp_const
= abs_inc
- 1;
1002 cmp_const
= abs_inc
* (unroll_number
- 1) + 1;
1006 emit_cmp_insn (diff
, GEN_INT (cmp_const
), cmp_code
, NULL_RTX
,
1010 emit_jump_insn (gen_ble (labels
[0]));
1012 emit_jump_insn (gen_bge (labels
[0]));
1013 JUMP_LABEL (get_last_insn ()) = labels
[0];
1014 LABEL_NUSES (labels
[0])++;
1017 sequence
= gen_sequence ();
1019 emit_insn_before (sequence
, loop_start
);
1021 /* Only the last copy of the loop body here needs the exit
1022 test, so set copy_end to exclude the compare/branch here,
1023 and then reset it inside the loop when get to the last
1026 if (GET_CODE (last_loop_insn
) == BARRIER
)
1027 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
1028 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
1031 /* The immediately preceding insn is a compare which we do not
1033 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
1035 /* The immediately preceding insn may not be a compare, so we
1037 copy_end
= PREV_INSN (last_loop_insn
);
1043 for (i
= 1; i
< unroll_number
; i
++)
1045 emit_label_after (labels
[unroll_number
- i
],
1046 PREV_INSN (loop_start
));
1048 bzero ((char *) map
->insn_map
, max_insnno
* sizeof (rtx
));
1049 bzero ((char *) map
->const_equiv_map
, maxregnum
* sizeof (rtx
));
1050 bzero ((char *) map
->const_age_map
,
1051 maxregnum
* sizeof (unsigned));
1054 for (j
= 0; j
< max_labelno
; j
++)
1056 set_label_in_map (map
, j
, gen_label_rtx ());
1058 for (j
= FIRST_PSEUDO_REGISTER
; j
< max_reg_before_loop
; j
++)
1061 map
->reg_map
[j
] = gen_reg_rtx (GET_MODE (regno_reg_rtx
[j
]));
1062 record_base_value (REGNO (map
->reg_map
[j
]),
1063 regno_reg_rtx
[j
], 0);
1065 /* The last copy needs the compare/branch insns at the end,
1066 so reset copy_end here if the loop ends with a conditional
1069 if (i
== unroll_number
- 1)
1071 if (GET_CODE (last_loop_insn
) == BARRIER
)
1072 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
1074 copy_end
= last_loop_insn
;
1077 /* None of the copies are the `last_iteration', so just
1078 pass zero for that parameter. */
1079 copy_loop_body (copy_start
, copy_end
, map
, exit_label
, 0,
1080 unroll_type
, start_label
, loop_end
,
1081 loop_start
, copy_end
);
1083 emit_label_after (labels
[0], PREV_INSN (loop_start
));
1085 if (GET_CODE (last_loop_insn
) == BARRIER
)
1087 insert_before
= PREV_INSN (last_loop_insn
);
1088 copy_end
= PREV_INSN (insert_before
);
1093 /* The immediately preceding insn is a compare which we do not
1095 insert_before
= PREV_INSN (last_loop_insn
);
1096 copy_end
= PREV_INSN (insert_before
);
1098 /* The immediately preceding insn may not be a compare, so we
1100 insert_before
= last_loop_insn
;
1101 copy_end
= PREV_INSN (last_loop_insn
);
1105 /* Set unroll type to MODULO now. */
1106 unroll_type
= UNROLL_MODULO
;
1107 loop_preconditioned
= 1;
1111 /* If reach here, and the loop type is UNROLL_NAIVE, then don't unroll
1112 the loop unless all loops are being unrolled. */
1113 if (unroll_type
== UNROLL_NAIVE
&& ! flag_unroll_all_loops
)
1115 if (loop_dump_stream
)
1116 fprintf (loop_dump_stream
, "Unrolling failure: Naive unrolling not being done.\n");
1120 /* At this point, we are guaranteed to unroll the loop. */
1122 /* Keep track of the unroll factor for the loop. */
1123 if (unroll_type
== UNROLL_COMPLETELY
)
1124 loop_info
->unroll_number
= -1;
1126 loop_info
->unroll_number
= unroll_number
;
1129 /* For each biv and giv, determine whether it can be safely split into
1130 a different variable for each unrolled copy of the loop body.
1131 We precalculate and save this info here, since computing it is
1134 Do this before deleting any instructions from the loop, so that
1135 back_branch_in_range_p will work correctly. */
1137 if (splitting_not_safe
)
1140 temp
= find_splittable_regs (unroll_type
, loop_start
, loop_end
,
1141 end_insert_before
, unroll_number
,
1142 loop_info
->n_iterations
);
1144 /* find_splittable_regs may have created some new registers, so must
1145 reallocate the reg_map with the new larger size, and must realloc
1146 the constant maps also. */
1148 maxregnum
= max_reg_num ();
1149 map
->reg_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
1151 init_reg_map (map
, maxregnum
);
1153 /* Space is needed in some of the map for new registers, so new_maxregnum
1154 is an (over)estimate of how many registers will exist at the end. */
1155 new_maxregnum
= maxregnum
+ (temp
* unroll_number
* 2);
1157 /* Must realloc space for the constant maps, because the number of registers
1158 may have changed. */
1160 map
->const_equiv_map
= (rtx
*) alloca (new_maxregnum
* sizeof (rtx
));
1161 map
->const_age_map
= (unsigned *) alloca (new_maxregnum
* sizeof (unsigned));
1163 map
->const_equiv_map_size
= new_maxregnum
;
1164 global_const_equiv_map
= map
->const_equiv_map
;
1165 global_const_equiv_map_size
= new_maxregnum
;
1167 /* Search the list of bivs and givs to find ones which need to be remapped
1168 when split, and set their reg_map entry appropriately. */
1170 for (bl
= loop_iv_list
; bl
; bl
= bl
->next
)
1172 if (REGNO (bl
->biv
->src_reg
) != bl
->regno
)
1173 map
->reg_map
[bl
->regno
] = bl
->biv
->src_reg
;
1175 /* Currently, non-reduced/final-value givs are never split. */
1176 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
1177 if (REGNO (v
->src_reg
) != bl
->regno
)
1178 map
->reg_map
[REGNO (v
->dest_reg
)] = v
->src_reg
;
1182 /* Use our current register alignment and pointer flags. */
1183 map
->regno_pointer_flag
= regno_pointer_flag
;
1184 map
->regno_pointer_align
= regno_pointer_align
;
1186 /* If the loop is being partially unrolled, and the iteration variables
1187 are being split, and are being renamed for the split, then must fix up
1188 the compare/jump instruction at the end of the loop to refer to the new
1189 registers. This compare isn't copied, so the registers used in it
1190 will never be replaced if it isn't done here. */
1192 if (unroll_type
== UNROLL_MODULO
)
1194 insn
= NEXT_INSN (copy_end
);
1195 if (GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == JUMP_INSN
)
1196 PATTERN (insn
) = remap_split_bivs (PATTERN (insn
));
1199 /* For unroll_number times, make a copy of each instruction
1200 between copy_start and copy_end, and insert these new instructions
1201 before the end of the loop. */
1203 for (i
= 0; i
< unroll_number
; i
++)
1205 bzero ((char *) map
->insn_map
, max_insnno
* sizeof (rtx
));
1206 bzero ((char *) map
->const_equiv_map
, new_maxregnum
* sizeof (rtx
));
1207 bzero ((char *) map
->const_age_map
, new_maxregnum
* sizeof (unsigned));
1210 for (j
= 0; j
< max_labelno
; j
++)
1212 set_label_in_map (map
, j
, gen_label_rtx ());
1214 for (j
= FIRST_PSEUDO_REGISTER
; j
< max_reg_before_loop
; j
++)
1217 map
->reg_map
[j
] = gen_reg_rtx (GET_MODE (regno_reg_rtx
[j
]));
1218 record_base_value (REGNO (map
->reg_map
[j
]),
1219 regno_reg_rtx
[j
], 0);
1222 /* If loop starts with a branch to the test, then fix it so that
1223 it points to the test of the first unrolled copy of the loop. */
1224 if (i
== 0 && loop_start
!= copy_start
)
1226 insn
= PREV_INSN (copy_start
);
1227 pattern
= PATTERN (insn
);
1229 tem
= get_label_from_map (map
,
1231 (XEXP (SET_SRC (pattern
), 0)));
1232 SET_SRC (pattern
) = gen_rtx_LABEL_REF (VOIDmode
, tem
);
1234 /* Set the jump label so that it can be used by later loop unrolling
1236 JUMP_LABEL (insn
) = tem
;
1237 LABEL_NUSES (tem
)++;
1240 copy_loop_body (copy_start
, copy_end
, map
, exit_label
,
1241 i
== unroll_number
- 1, unroll_type
, start_label
,
1242 loop_end
, insert_before
, insert_before
);
1245 /* Before deleting any insns, emit a CODE_LABEL immediately after the last
1246 insn to be deleted. This prevents any runaway delete_insn call from
1247 more insns that it should, as it always stops at a CODE_LABEL. */
1249 /* Delete the compare and branch at the end of the loop if completely
1250 unrolling the loop. Deleting the backward branch at the end also
1251 deletes the code label at the start of the loop. This is done at
1252 the very end to avoid problems with back_branch_in_range_p. */
1254 if (unroll_type
== UNROLL_COMPLETELY
)
1255 safety_label
= emit_label_after (gen_label_rtx (), last_loop_insn
);
1257 safety_label
= emit_label_after (gen_label_rtx (), copy_end
);
1259 /* Delete all of the original loop instructions. Don't delete the
1260 LOOP_BEG note, or the first code label in the loop. */
1262 insn
= NEXT_INSN (copy_start
);
1263 while (insn
!= safety_label
)
1265 if (insn
!= start_label
)
1266 insn
= delete_insn (insn
);
1268 insn
= NEXT_INSN (insn
);
1271 /* Can now delete the 'safety' label emitted to protect us from runaway
1272 delete_insn calls. */
1273 if (INSN_DELETED_P (safety_label
))
1275 delete_insn (safety_label
);
1277 /* If exit_label exists, emit it after the loop. Doing the emit here
1278 forces it to have a higher INSN_UID than any insn in the unrolled loop.
1279 This is needed so that mostly_true_jump in reorg.c will treat jumps
1280 to this loop end label correctly, i.e. predict that they are usually
1283 emit_label_after (exit_label
, loop_end
);
1286 /* Return true if the loop can be safely, and profitably, preconditioned
1287 so that the unrolled copies of the loop body don't need exit tests.
1289 This only works if final_value, initial_value and increment can be
1290 determined, and if increment is a constant power of 2.
1291 If increment is not a power of 2, then the preconditioning modulo
1292 operation would require a real modulo instead of a boolean AND, and this
1293 is not considered `profitable'. */
1295 /* ??? If the loop is known to be executed very many times, or the machine
1296 has a very cheap divide instruction, then preconditioning is a win even
1297 when the increment is not a power of 2. Use RTX_COST to compute
1298 whether divide is cheap.
1299 ??? A divide by constant doesn't actually need a divide, look at
1300 expand_divmod. The reduced cost of this optimized modulo is not
1301 reflected in RTX_COST. */
1304 precondition_loop_p (loop_start
, loop_info
,
1305 initial_value
, final_value
, increment
, mode
)
1307 struct loop_info
*loop_info
;
1308 rtx
*initial_value
, *final_value
, *increment
;
1309 enum machine_mode
*mode
;
1312 if (loop_info
->n_iterations
> 0)
1314 *initial_value
= const0_rtx
;
1315 *increment
= const1_rtx
;
1316 *final_value
= GEN_INT (loop_info
->n_iterations
);
1319 if (loop_dump_stream
)
1321 fputs ("Preconditioning: Success, number of iterations known, ",
1323 fprintf (loop_dump_stream
, HOST_WIDE_INT_PRINT_DEC
,
1324 loop_info
->n_iterations
);
1325 fputs (".\n", loop_dump_stream
);
1330 if (loop_info
->initial_value
== 0)
1332 if (loop_dump_stream
)
1333 fprintf (loop_dump_stream
,
1334 "Preconditioning: Could not find initial value.\n");
1337 else if (loop_info
->increment
== 0)
1339 if (loop_dump_stream
)
1340 fprintf (loop_dump_stream
,
1341 "Preconditioning: Could not find increment value.\n");
1344 else if (GET_CODE (loop_info
->increment
) != CONST_INT
)
1346 if (loop_dump_stream
)
1347 fprintf (loop_dump_stream
,
1348 "Preconditioning: Increment not a constant.\n");
1351 else if ((exact_log2 (INTVAL (loop_info
->increment
)) < 0)
1352 && (exact_log2 (- INTVAL (loop_info
->increment
)) < 0))
1354 if (loop_dump_stream
)
1355 fprintf (loop_dump_stream
,
1356 "Preconditioning: Increment not a constant power of 2.\n");
1360 /* Unsigned_compare and compare_dir can be ignored here, since they do
1361 not matter for preconditioning. */
1363 if (loop_info
->final_value
== 0)
1365 if (loop_dump_stream
)
1366 fprintf (loop_dump_stream
,
1367 "Preconditioning: EQ comparison loop.\n");
1371 /* Must ensure that final_value is invariant, so call invariant_p to
1372 check. Before doing so, must check regno against max_reg_before_loop
1373 to make sure that the register is in the range covered by invariant_p.
1374 If it isn't, then it is most likely a biv/giv which by definition are
1376 if ((GET_CODE (loop_info
->final_value
) == REG
1377 && REGNO (loop_info
->final_value
) >= max_reg_before_loop
)
1378 || (GET_CODE (loop_info
->final_value
) == PLUS
1379 && REGNO (XEXP (loop_info
->final_value
, 0)) >= max_reg_before_loop
)
1380 || ! invariant_p (loop_info
->final_value
))
1382 if (loop_dump_stream
)
1383 fprintf (loop_dump_stream
,
1384 "Preconditioning: Final value not invariant.\n");
1388 /* Fail for floating point values, since the caller of this function
1389 does not have code to deal with them. */
1390 if (GET_MODE_CLASS (GET_MODE (loop_info
->final_value
)) == MODE_FLOAT
1391 || GET_MODE_CLASS (GET_MODE (loop_info
->initial_value
)) == MODE_FLOAT
)
1393 if (loop_dump_stream
)
1394 fprintf (loop_dump_stream
,
1395 "Preconditioning: Floating point final or initial value.\n");
1399 /* Fail if loop_info->iteration_var is not live before loop_start,
1400 since we need to test its value in the preconditioning code. */
1402 if (uid_luid
[REGNO_FIRST_UID (REGNO (loop_info
->iteration_var
))]
1403 > INSN_LUID (loop_start
))
1405 if (loop_dump_stream
)
1406 fprintf (loop_dump_stream
,
1407 "Preconditioning: Iteration var not live before loop start.\n");
1411 /* ??? Note that if iteration_info is modifed to allow GIV iterators
1412 such as "while (i-- > 0)", the initial value will be one too small.
1413 In this case, loop_iteration_var could be used to determine
1414 the correct initial value, provided the loop has not been reversed.
1416 Also note that the absolute values of initial_value and
1417 final_value are unimportant as only their difference is used for
1418 calculating the number of loop iterations. */
1419 *initial_value
= loop_info
->initial_value
;
1420 *increment
= loop_info
->increment
;
1421 *final_value
= loop_info
->final_value
;
1423 /* Decide what mode to do these calculations in. Choose the larger
1424 of final_value's mode and initial_value's mode, or a full-word if
1425 both are constants. */
1426 *mode
= GET_MODE (*final_value
);
1427 if (*mode
== VOIDmode
)
1429 *mode
= GET_MODE (*initial_value
);
1430 if (*mode
== VOIDmode
)
1433 else if (*mode
!= GET_MODE (*initial_value
)
1434 && (GET_MODE_SIZE (*mode
)
1435 < GET_MODE_SIZE (GET_MODE (*initial_value
))))
1436 *mode
= GET_MODE (*initial_value
);
1439 if (loop_dump_stream
)
1440 fprintf (loop_dump_stream
, "Preconditioning: Successful.\n");
1445 /* All pseudo-registers must be mapped to themselves. Two hard registers
1446 must be mapped, VIRTUAL_STACK_VARS_REGNUM and VIRTUAL_INCOMING_ARGS_
1447 REGNUM, to avoid function-inlining specific conversions of these
1448 registers. All other hard regs can not be mapped because they may be
1453 init_reg_map (map
, maxregnum
)
1454 struct inline_remap
*map
;
1459 for (i
= maxregnum
- 1; i
> LAST_VIRTUAL_REGISTER
; i
--)
1460 map
->reg_map
[i
] = regno_reg_rtx
[i
];
1461 /* Just clear the rest of the entries. */
1462 for (i
= LAST_VIRTUAL_REGISTER
; i
>= 0; i
--)
1463 map
->reg_map
[i
] = 0;
1465 map
->reg_map
[VIRTUAL_STACK_VARS_REGNUM
]
1466 = regno_reg_rtx
[VIRTUAL_STACK_VARS_REGNUM
];
1467 map
->reg_map
[VIRTUAL_INCOMING_ARGS_REGNUM
]
1468 = regno_reg_rtx
[VIRTUAL_INCOMING_ARGS_REGNUM
];
1471 /* Strength-reduction will often emit code for optimized biv/givs which
1472 calculates their value in a temporary register, and then copies the result
1473 to the iv. This procedure reconstructs the pattern computing the iv;
1474 verifying that all operands are of the proper form.
1476 PATTERN must be the result of single_set.
1477 The return value is the amount that the giv is incremented by. */
1480 calculate_giv_inc (pattern
, src_insn
, regno
)
1481 rtx pattern
, src_insn
;
1485 rtx increment_total
= 0;
1489 /* Verify that we have an increment insn here. First check for a plus
1490 as the set source. */
1491 if (GET_CODE (SET_SRC (pattern
)) != PLUS
)
1493 /* SR sometimes computes the new giv value in a temp, then copies it
1495 src_insn
= PREV_INSN (src_insn
);
1496 pattern
= PATTERN (src_insn
);
1497 if (GET_CODE (SET_SRC (pattern
)) != PLUS
)
1500 /* The last insn emitted is not needed, so delete it to avoid confusing
1501 the second cse pass. This insn sets the giv unnecessarily. */
1502 delete_insn (get_last_insn ());
1505 /* Verify that we have a constant as the second operand of the plus. */
1506 increment
= XEXP (SET_SRC (pattern
), 1);
1507 if (GET_CODE (increment
) != CONST_INT
)
1509 /* SR sometimes puts the constant in a register, especially if it is
1510 too big to be an add immed operand. */
1511 src_insn
= PREV_INSN (src_insn
);
1512 increment
= SET_SRC (PATTERN (src_insn
));
1514 /* SR may have used LO_SUM to compute the constant if it is too large
1515 for a load immed operand. In this case, the constant is in operand
1516 one of the LO_SUM rtx. */
1517 if (GET_CODE (increment
) == LO_SUM
)
1518 increment
= XEXP (increment
, 1);
1520 /* Some ports store large constants in memory and add a REG_EQUAL
1521 note to the store insn. */
1522 else if (GET_CODE (increment
) == MEM
)
1524 rtx note
= find_reg_note (src_insn
, REG_EQUAL
, 0);
1526 increment
= XEXP (note
, 0);
1529 else if (GET_CODE (increment
) == IOR
1530 || GET_CODE (increment
) == ASHIFT
1531 || GET_CODE (increment
) == PLUS
)
1533 /* The rs6000 port loads some constants with IOR.
1534 The alpha port loads some constants with ASHIFT and PLUS. */
1535 rtx second_part
= XEXP (increment
, 1);
1536 enum rtx_code code
= GET_CODE (increment
);
1538 src_insn
= PREV_INSN (src_insn
);
1539 increment
= SET_SRC (PATTERN (src_insn
));
1540 /* Don't need the last insn anymore. */
1541 delete_insn (get_last_insn ());
1543 if (GET_CODE (second_part
) != CONST_INT
1544 || GET_CODE (increment
) != CONST_INT
)
1548 increment
= GEN_INT (INTVAL (increment
) | INTVAL (second_part
));
1549 else if (code
== PLUS
)
1550 increment
= GEN_INT (INTVAL (increment
) + INTVAL (second_part
));
1552 increment
= GEN_INT (INTVAL (increment
) << INTVAL (second_part
));
1555 if (GET_CODE (increment
) != CONST_INT
)
1558 /* The insn loading the constant into a register is no longer needed,
1560 delete_insn (get_last_insn ());
1563 if (increment_total
)
1564 increment_total
= GEN_INT (INTVAL (increment_total
) + INTVAL (increment
));
1566 increment_total
= increment
;
1568 /* Check that the source register is the same as the register we expected
1569 to see as the source. If not, something is seriously wrong. */
1570 if (GET_CODE (XEXP (SET_SRC (pattern
), 0)) != REG
1571 || REGNO (XEXP (SET_SRC (pattern
), 0)) != regno
)
1573 /* Some machines (e.g. the romp), may emit two add instructions for
1574 certain constants, so lets try looking for another add immediately
1575 before this one if we have only seen one add insn so far. */
1581 src_insn
= PREV_INSN (src_insn
);
1582 pattern
= PATTERN (src_insn
);
1584 delete_insn (get_last_insn ());
1592 return increment_total
;
1595 /* Copy REG_NOTES, except for insn references, because not all insn_map
1596 entries are valid yet. We do need to copy registers now though, because
1597 the reg_map entries can change during copying. */
1600 initial_reg_note_copy (notes
, map
)
1602 struct inline_remap
*map
;
1609 copy
= rtx_alloc (GET_CODE (notes
));
1610 PUT_MODE (copy
, GET_MODE (notes
));
1612 if (GET_CODE (notes
) == EXPR_LIST
)
1613 XEXP (copy
, 0) = copy_rtx_and_substitute (XEXP (notes
, 0), map
);
1614 else if (GET_CODE (notes
) == INSN_LIST
)
1615 /* Don't substitute for these yet. */
1616 XEXP (copy
, 0) = XEXP (notes
, 0);
1620 XEXP (copy
, 1) = initial_reg_note_copy (XEXP (notes
, 1), map
);
1625 /* Fixup insn references in copied REG_NOTES. */
1628 final_reg_note_copy (notes
, map
)
1630 struct inline_remap
*map
;
1634 for (note
= notes
; note
; note
= XEXP (note
, 1))
1635 if (GET_CODE (note
) == INSN_LIST
)
1636 XEXP (note
, 0) = map
->insn_map
[INSN_UID (XEXP (note
, 0))];
1639 /* Copy each instruction in the loop, substituting from map as appropriate.
1640 This is very similar to a loop in expand_inline_function. */
1643 copy_loop_body (copy_start
, copy_end
, map
, exit_label
, last_iteration
,
1644 unroll_type
, start_label
, loop_end
, insert_before
,
1646 rtx copy_start
, copy_end
;
1647 struct inline_remap
*map
;
1650 enum unroll_types unroll_type
;
1651 rtx start_label
, loop_end
, insert_before
, copy_notes_from
;
1655 int dest_reg_was_split
, i
;
1659 rtx final_label
= 0;
1660 rtx giv_inc
, giv_dest_reg
, giv_src_reg
;
1662 /* If this isn't the last iteration, then map any references to the
1663 start_label to final_label. Final label will then be emitted immediately
1664 after the end of this loop body if it was ever used.
1666 If this is the last iteration, then map references to the start_label
1668 if (! last_iteration
)
1670 final_label
= gen_label_rtx ();
1671 set_label_in_map (map
, CODE_LABEL_NUMBER (start_label
),
1675 set_label_in_map (map
, CODE_LABEL_NUMBER (start_label
), start_label
);
1682 insn
= NEXT_INSN (insn
);
1684 map
->orig_asm_operands_vector
= 0;
1686 switch (GET_CODE (insn
))
1689 pattern
= PATTERN (insn
);
1693 /* Check to see if this is a giv that has been combined with
1694 some split address givs. (Combined in the sense that
1695 `combine_givs' in loop.c has put two givs in the same register.)
1696 In this case, we must search all givs based on the same biv to
1697 find the address givs. Then split the address givs.
1698 Do this before splitting the giv, since that may map the
1699 SET_DEST to a new register. */
1701 if ((set
= single_set (insn
))
1702 && GET_CODE (SET_DEST (set
)) == REG
1703 && addr_combined_regs
[REGNO (SET_DEST (set
))])
1705 struct iv_class
*bl
;
1706 struct induction
*v
, *tv
;
1707 int regno
= REGNO (SET_DEST (set
));
1709 v
= addr_combined_regs
[REGNO (SET_DEST (set
))];
1710 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
1712 /* Although the giv_inc amount is not needed here, we must call
1713 calculate_giv_inc here since it might try to delete the
1714 last insn emitted. If we wait until later to call it,
1715 we might accidentally delete insns generated immediately
1716 below by emit_unrolled_add. */
1718 giv_inc
= calculate_giv_inc (set
, insn
, regno
);
1720 /* Now find all address giv's that were combined with this
1722 for (tv
= bl
->giv
; tv
; tv
= tv
->next_iv
)
1723 if (tv
->giv_type
== DEST_ADDR
&& tv
->same
== v
)
1727 /* If this DEST_ADDR giv was not split, then ignore it. */
1728 if (*tv
->location
!= tv
->dest_reg
)
1731 /* Scale this_giv_inc if the multiplicative factors of
1732 the two givs are different. */
1733 this_giv_inc
= INTVAL (giv_inc
);
1734 if (tv
->mult_val
!= v
->mult_val
)
1735 this_giv_inc
= (this_giv_inc
/ INTVAL (v
->mult_val
)
1736 * INTVAL (tv
->mult_val
));
1738 tv
->dest_reg
= plus_constant (tv
->dest_reg
, this_giv_inc
);
1739 *tv
->location
= tv
->dest_reg
;
1741 if (last_iteration
&& unroll_type
!= UNROLL_COMPLETELY
)
1743 /* Must emit an insn to increment the split address
1744 giv. Add in the const_adjust field in case there
1745 was a constant eliminated from the address. */
1746 rtx value
, dest_reg
;
1748 /* tv->dest_reg will be either a bare register,
1749 or else a register plus a constant. */
1750 if (GET_CODE (tv
->dest_reg
) == REG
)
1751 dest_reg
= tv
->dest_reg
;
1753 dest_reg
= XEXP (tv
->dest_reg
, 0);
1755 /* Check for shared address givs, and avoid
1756 incrementing the shared pseudo reg more than
1758 if (! tv
->same_insn
&& ! tv
->shared
)
1760 /* tv->dest_reg may actually be a (PLUS (REG)
1761 (CONST)) here, so we must call plus_constant
1762 to add the const_adjust amount before calling
1763 emit_unrolled_add below. */
1764 value
= plus_constant (tv
->dest_reg
,
1767 /* The constant could be too large for an add
1768 immediate, so can't directly emit an insn
1770 emit_unrolled_add (dest_reg
, XEXP (value
, 0),
1774 /* Reset the giv to be just the register again, in case
1775 it is used after the set we have just emitted.
1776 We must subtract the const_adjust factor added in
1778 tv
->dest_reg
= plus_constant (dest_reg
,
1779 - tv
->const_adjust
);
1780 *tv
->location
= tv
->dest_reg
;
1785 /* If this is a setting of a splittable variable, then determine
1786 how to split the variable, create a new set based on this split,
1787 and set up the reg_map so that later uses of the variable will
1788 use the new split variable. */
1790 dest_reg_was_split
= 0;
1792 if ((set
= single_set (insn
))
1793 && GET_CODE (SET_DEST (set
)) == REG
1794 && splittable_regs
[REGNO (SET_DEST (set
))])
1796 int regno
= REGNO (SET_DEST (set
));
1798 dest_reg_was_split
= 1;
1800 /* Compute the increment value for the giv, if it wasn't
1801 already computed above. */
1804 giv_inc
= calculate_giv_inc (set
, insn
, regno
);
1805 giv_dest_reg
= SET_DEST (set
);
1806 giv_src_reg
= SET_DEST (set
);
1808 if (unroll_type
== UNROLL_COMPLETELY
)
1810 /* Completely unrolling the loop. Set the induction
1811 variable to a known constant value. */
1813 /* The value in splittable_regs may be an invariant
1814 value, so we must use plus_constant here. */
1815 splittable_regs
[regno
]
1816 = plus_constant (splittable_regs
[regno
], INTVAL (giv_inc
));
1818 if (GET_CODE (splittable_regs
[regno
]) == PLUS
)
1820 giv_src_reg
= XEXP (splittable_regs
[regno
], 0);
1821 giv_inc
= XEXP (splittable_regs
[regno
], 1);
1825 /* The splittable_regs value must be a REG or a
1826 CONST_INT, so put the entire value in the giv_src_reg
1828 giv_src_reg
= splittable_regs
[regno
];
1829 giv_inc
= const0_rtx
;
1834 /* Partially unrolling loop. Create a new pseudo
1835 register for the iteration variable, and set it to
1836 be a constant plus the original register. Except
1837 on the last iteration, when the result has to
1838 go back into the original iteration var register. */
1840 /* Handle bivs which must be mapped to a new register
1841 when split. This happens for bivs which need their
1842 final value set before loop entry. The new register
1843 for the biv was stored in the biv's first struct
1844 induction entry by find_splittable_regs. */
1846 if (regno
< max_reg_before_loop
1847 && reg_iv_type
[regno
] == BASIC_INDUCT
)
1849 giv_src_reg
= reg_biv_class
[regno
]->biv
->src_reg
;
1850 giv_dest_reg
= giv_src_reg
;
1854 /* If non-reduced/final-value givs were split, then
1855 this would have to remap those givs also. See
1856 find_splittable_regs. */
1859 splittable_regs
[regno
]
1860 = GEN_INT (INTVAL (giv_inc
)
1861 + INTVAL (splittable_regs
[regno
]));
1862 giv_inc
= splittable_regs
[regno
];
1864 /* Now split the induction variable by changing the dest
1865 of this insn to a new register, and setting its
1866 reg_map entry to point to this new register.
1868 If this is the last iteration, and this is the last insn
1869 that will update the iv, then reuse the original dest,
1870 to ensure that the iv will have the proper value when
1871 the loop exits or repeats.
1873 Using splittable_regs_updates here like this is safe,
1874 because it can only be greater than one if all
1875 instructions modifying the iv are always executed in
1878 if (! last_iteration
1879 || (splittable_regs_updates
[regno
]-- != 1))
1881 tem
= gen_reg_rtx (GET_MODE (giv_src_reg
));
1883 map
->reg_map
[regno
] = tem
;
1884 record_base_value (REGNO (tem
),
1885 giv_inc
== const0_rtx
1887 : gen_rtx_PLUS (GET_MODE (giv_src_reg
),
1888 giv_src_reg
, giv_inc
),
1892 map
->reg_map
[regno
] = giv_src_reg
;
1895 /* The constant being added could be too large for an add
1896 immediate, so can't directly emit an insn here. */
1897 emit_unrolled_add (giv_dest_reg
, giv_src_reg
, giv_inc
);
1898 copy
= get_last_insn ();
1899 pattern
= PATTERN (copy
);
1903 pattern
= copy_rtx_and_substitute (pattern
, map
);
1904 copy
= emit_insn (pattern
);
1906 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
1909 /* If this insn is setting CC0, it may need to look at
1910 the insn that uses CC0 to see what type of insn it is.
1911 In that case, the call to recog via validate_change will
1912 fail. So don't substitute constants here. Instead,
1913 do it when we emit the following insn.
1915 For example, see the pyr.md file. That machine has signed and
1916 unsigned compares. The compare patterns must check the
1917 following branch insn to see which what kind of compare to
1920 If the previous insn set CC0, substitute constants on it as
1922 if (sets_cc0_p (PATTERN (copy
)) != 0)
1927 try_constants (cc0_insn
, map
);
1929 try_constants (copy
, map
);
1932 try_constants (copy
, map
);
1935 /* Make split induction variable constants `permanent' since we
1936 know there are no backward branches across iteration variable
1937 settings which would invalidate this. */
1938 if (dest_reg_was_split
)
1940 int regno
= REGNO (SET_DEST (pattern
));
1942 if (regno
< map
->const_equiv_map_size
1943 && map
->const_age_map
[regno
] == map
->const_age
)
1944 map
->const_age_map
[regno
] = -1;
1949 pattern
= copy_rtx_and_substitute (PATTERN (insn
), map
);
1950 copy
= emit_jump_insn (pattern
);
1951 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
1953 if (JUMP_LABEL (insn
) == start_label
&& insn
== copy_end
1954 && ! last_iteration
)
1956 /* This is a branch to the beginning of the loop; this is the
1957 last insn being copied; and this is not the last iteration.
1958 In this case, we want to change the original fall through
1959 case to be a branch past the end of the loop, and the
1960 original jump label case to fall_through. */
1962 if (invert_exp (pattern
, copy
))
1964 if (! redirect_exp (&pattern
,
1965 get_label_from_map (map
,
1967 (JUMP_LABEL (insn
))),
1974 rtx lab
= gen_label_rtx ();
1975 /* Can't do it by reversing the jump (probably because we
1976 couldn't reverse the conditions), so emit a new
1977 jump_insn after COPY, and redirect the jump around
1979 jmp
= emit_jump_insn_after (gen_jump (exit_label
), copy
);
1980 jmp
= emit_barrier_after (jmp
);
1981 emit_label_after (lab
, jmp
);
1982 LABEL_NUSES (lab
) = 0;
1983 if (! redirect_exp (&pattern
,
1984 get_label_from_map (map
,
1986 (JUMP_LABEL (insn
))),
1994 try_constants (cc0_insn
, map
);
1997 try_constants (copy
, map
);
1999 /* Set the jump label of COPY correctly to avoid problems with
2000 later passes of unroll_loop, if INSN had jump label set. */
2001 if (JUMP_LABEL (insn
))
2005 /* Can't use the label_map for every insn, since this may be
2006 the backward branch, and hence the label was not mapped. */
2007 if ((set
= single_set (copy
)))
2009 tem
= SET_SRC (set
);
2010 if (GET_CODE (tem
) == LABEL_REF
)
2011 label
= XEXP (tem
, 0);
2012 else if (GET_CODE (tem
) == IF_THEN_ELSE
)
2014 if (XEXP (tem
, 1) != pc_rtx
)
2015 label
= XEXP (XEXP (tem
, 1), 0);
2017 label
= XEXP (XEXP (tem
, 2), 0);
2021 if (label
&& GET_CODE (label
) == CODE_LABEL
)
2022 JUMP_LABEL (copy
) = label
;
2025 /* An unrecognizable jump insn, probably the entry jump
2026 for a switch statement. This label must have been mapped,
2027 so just use the label_map to get the new jump label. */
2029 = get_label_from_map (map
,
2030 CODE_LABEL_NUMBER (JUMP_LABEL (insn
)));
2033 /* If this is a non-local jump, then must increase the label
2034 use count so that the label will not be deleted when the
2035 original jump is deleted. */
2036 LABEL_NUSES (JUMP_LABEL (copy
))++;
2038 else if (GET_CODE (PATTERN (copy
)) == ADDR_VEC
2039 || GET_CODE (PATTERN (copy
)) == ADDR_DIFF_VEC
)
2041 rtx pat
= PATTERN (copy
);
2042 int diff_vec_p
= GET_CODE (pat
) == ADDR_DIFF_VEC
;
2043 int len
= XVECLEN (pat
, diff_vec_p
);
2046 for (i
= 0; i
< len
; i
++)
2047 LABEL_NUSES (XEXP (XVECEXP (pat
, diff_vec_p
, i
), 0))++;
2050 /* If this used to be a conditional jump insn but whose branch
2051 direction is now known, we must do something special. */
2052 if (condjump_p (insn
) && !simplejump_p (insn
) && map
->last_pc_value
)
2055 /* The previous insn set cc0 for us. So delete it. */
2056 delete_insn (PREV_INSN (copy
));
2059 /* If this is now a no-op, delete it. */
2060 if (map
->last_pc_value
== pc_rtx
)
2062 /* Don't let delete_insn delete the label referenced here,
2063 because we might possibly need it later for some other
2064 instruction in the loop. */
2065 if (JUMP_LABEL (copy
))
2066 LABEL_NUSES (JUMP_LABEL (copy
))++;
2068 if (JUMP_LABEL (copy
))
2069 LABEL_NUSES (JUMP_LABEL (copy
))--;
2073 /* Otherwise, this is unconditional jump so we must put a
2074 BARRIER after it. We could do some dead code elimination
2075 here, but jump.c will do it just as well. */
2081 pattern
= copy_rtx_and_substitute (PATTERN (insn
), map
);
2082 copy
= emit_call_insn (pattern
);
2083 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
2085 /* Because the USAGE information potentially contains objects other
2086 than hard registers, we need to copy it. */
2087 CALL_INSN_FUNCTION_USAGE (copy
)
2088 = copy_rtx_and_substitute (CALL_INSN_FUNCTION_USAGE (insn
), map
);
2092 try_constants (cc0_insn
, map
);
2095 try_constants (copy
, map
);
2097 /* Be lazy and assume CALL_INSNs clobber all hard registers. */
2098 for (i
= 0; i
< FIRST_PSEUDO_REGISTER
; i
++)
2099 map
->const_equiv_map
[i
] = 0;
2103 /* If this is the loop start label, then we don't need to emit a
2104 copy of this label since no one will use it. */
2106 if (insn
!= start_label
)
2108 copy
= emit_label (get_label_from_map (map
,
2109 CODE_LABEL_NUMBER (insn
)));
2115 copy
= emit_barrier ();
2119 /* VTOP notes are valid only before the loop exit test. If placed
2120 anywhere else, loop may generate bad code. */
2122 if (NOTE_LINE_NUMBER (insn
) != NOTE_INSN_DELETED
2123 && (NOTE_LINE_NUMBER (insn
) != NOTE_INSN_LOOP_VTOP
2124 || (last_iteration
&& unroll_type
!= UNROLL_COMPLETELY
)))
2125 copy
= emit_note (NOTE_SOURCE_FILE (insn
),
2126 NOTE_LINE_NUMBER (insn
));
2136 map
->insn_map
[INSN_UID (insn
)] = copy
;
2138 while (insn
!= copy_end
);
2140 /* Now finish coping the REG_NOTES. */
2144 insn
= NEXT_INSN (insn
);
2145 if ((GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == JUMP_INSN
2146 || GET_CODE (insn
) == CALL_INSN
)
2147 && map
->insn_map
[INSN_UID (insn
)])
2148 final_reg_note_copy (REG_NOTES (map
->insn_map
[INSN_UID (insn
)]), map
);
2150 while (insn
!= copy_end
);
2152 /* There may be notes between copy_notes_from and loop_end. Emit a copy of
2153 each of these notes here, since there may be some important ones, such as
2154 NOTE_INSN_BLOCK_END notes, in this group. We don't do this on the last
2155 iteration, because the original notes won't be deleted.
2157 We can't use insert_before here, because when from preconditioning,
2158 insert_before points before the loop. We can't use copy_end, because
2159 there may be insns already inserted after it (which we don't want to
2160 copy) when not from preconditioning code. */
2162 if (! last_iteration
)
2164 for (insn
= copy_notes_from
; insn
!= loop_end
; insn
= NEXT_INSN (insn
))
2166 if (GET_CODE (insn
) == NOTE
2167 && NOTE_LINE_NUMBER (insn
) != NOTE_INSN_DELETED
)
2168 emit_note (NOTE_SOURCE_FILE (insn
), NOTE_LINE_NUMBER (insn
));
2172 if (final_label
&& LABEL_NUSES (final_label
) > 0)
2173 emit_label (final_label
);
2175 tem
= gen_sequence ();
2177 emit_insn_before (tem
, insert_before
);
2180 /* Emit an insn, using the expand_binop to ensure that a valid insn is
2181 emitted. This will correctly handle the case where the increment value
2182 won't fit in the immediate field of a PLUS insns. */
2185 emit_unrolled_add (dest_reg
, src_reg
, increment
)
2186 rtx dest_reg
, src_reg
, increment
;
2190 result
= expand_binop (GET_MODE (dest_reg
), add_optab
, src_reg
, increment
,
2191 dest_reg
, 0, OPTAB_LIB_WIDEN
);
2193 if (dest_reg
!= result
)
2194 emit_move_insn (dest_reg
, result
);
2197 /* Searches the insns between INSN and LOOP_END. Returns 1 if there
2198 is a backward branch in that range that branches to somewhere between
2199 LOOP_START and INSN. Returns 0 otherwise. */
2201 /* ??? This is quadratic algorithm. Could be rewritten to be linear.
2202 In practice, this is not a problem, because this function is seldom called,
2203 and uses a negligible amount of CPU time on average. */
2206 back_branch_in_range_p (insn
, loop_start
, loop_end
)
2208 rtx loop_start
, loop_end
;
2210 rtx p
, q
, target_insn
;
2211 rtx orig_loop_end
= loop_end
;
2213 /* Stop before we get to the backward branch at the end of the loop. */
2214 loop_end
= prev_nonnote_insn (loop_end
);
2215 if (GET_CODE (loop_end
) == BARRIER
)
2216 loop_end
= PREV_INSN (loop_end
);
2218 /* Check in case insn has been deleted, search forward for first non
2219 deleted insn following it. */
2220 while (INSN_DELETED_P (insn
))
2221 insn
= NEXT_INSN (insn
);
2223 /* Check for the case where insn is the last insn in the loop. Deal
2224 with the case where INSN was a deleted loop test insn, in which case
2225 it will now be the NOTE_LOOP_END. */
2226 if (insn
== loop_end
|| insn
== orig_loop_end
)
2229 for (p
= NEXT_INSN (insn
); p
!= loop_end
; p
= NEXT_INSN (p
))
2231 if (GET_CODE (p
) == JUMP_INSN
)
2233 target_insn
= JUMP_LABEL (p
);
2235 /* Search from loop_start to insn, to see if one of them is
2236 the target_insn. We can't use INSN_LUID comparisons here,
2237 since insn may not have an LUID entry. */
2238 for (q
= loop_start
; q
!= insn
; q
= NEXT_INSN (q
))
2239 if (q
== target_insn
)
2247 /* Try to generate the simplest rtx for the expression
2248 (PLUS (MULT mult1 mult2) add1). This is used to calculate the initial
2252 fold_rtx_mult_add (mult1
, mult2
, add1
, mode
)
2253 rtx mult1
, mult2
, add1
;
2254 enum machine_mode mode
;
2259 /* The modes must all be the same. This should always be true. For now,
2260 check to make sure. */
2261 if ((GET_MODE (mult1
) != mode
&& GET_MODE (mult1
) != VOIDmode
)
2262 || (GET_MODE (mult2
) != mode
&& GET_MODE (mult2
) != VOIDmode
)
2263 || (GET_MODE (add1
) != mode
&& GET_MODE (add1
) != VOIDmode
))
2266 /* Ensure that if at least one of mult1/mult2 are constant, then mult2
2267 will be a constant. */
2268 if (GET_CODE (mult1
) == CONST_INT
)
2275 mult_res
= simplify_binary_operation (MULT
, mode
, mult1
, mult2
);
2277 mult_res
= gen_rtx_MULT (mode
, mult1
, mult2
);
2279 /* Again, put the constant second. */
2280 if (GET_CODE (add1
) == CONST_INT
)
2287 result
= simplify_binary_operation (PLUS
, mode
, add1
, mult_res
);
2289 result
= gen_rtx_PLUS (mode
, add1
, mult_res
);
2294 /* Searches the list of induction struct's for the biv BL, to try to calculate
2295 the total increment value for one iteration of the loop as a constant.
2297 Returns the increment value as an rtx, simplified as much as possible,
2298 if it can be calculated. Otherwise, returns 0. */
2301 biv_total_increment (bl
, loop_start
, loop_end
)
2302 struct iv_class
*bl
;
2303 rtx loop_start
, loop_end
;
2305 struct induction
*v
;
2308 /* For increment, must check every instruction that sets it. Each
2309 instruction must be executed only once each time through the loop.
2310 To verify this, we check that the insn is always executed, and that
2311 there are no backward branches after the insn that branch to before it.
2312 Also, the insn must have a mult_val of one (to make sure it really is
2315 result
= const0_rtx
;
2316 for (v
= bl
->biv
; v
; v
= v
->next_iv
)
2318 if (v
->always_computable
&& v
->mult_val
== const1_rtx
2319 && ! v
->maybe_multiple
)
2320 result
= fold_rtx_mult_add (result
, const1_rtx
, v
->add_val
, v
->mode
);
2328 /* Determine the initial value of the iteration variable, and the amount
2329 that it is incremented each loop. Use the tables constructed by
2330 the strength reduction pass to calculate these values.
2332 Initial_value and/or increment are set to zero if their values could not
2336 iteration_info (iteration_var
, initial_value
, increment
, loop_start
, loop_end
)
2337 rtx iteration_var
, *initial_value
, *increment
;
2338 rtx loop_start
, loop_end
;
2340 struct iv_class
*bl
;
2342 struct induction
*v
;
2345 /* Clear the result values, in case no answer can be found. */
2349 /* The iteration variable can be either a giv or a biv. Check to see
2350 which it is, and compute the variable's initial value, and increment
2351 value if possible. */
2353 /* If this is a new register, can't handle it since we don't have any
2354 reg_iv_type entry for it. */
2355 if (REGNO (iteration_var
) >= max_reg_before_loop
)
2357 if (loop_dump_stream
)
2358 fprintf (loop_dump_stream
,
2359 "Loop unrolling: No reg_iv_type entry for iteration var.\n");
2363 /* Reject iteration variables larger than the host wide int size, since they
2364 could result in a number of iterations greater than the range of our
2365 `unsigned HOST_WIDE_INT' variable loop_info->n_iterations. */
2366 else if ((GET_MODE_BITSIZE (GET_MODE (iteration_var
))
2367 > HOST_BITS_PER_WIDE_INT
))
2369 if (loop_dump_stream
)
2370 fprintf (loop_dump_stream
,
2371 "Loop unrolling: Iteration var rejected because mode too large.\n");
2374 else if (GET_MODE_CLASS (GET_MODE (iteration_var
)) != MODE_INT
)
2376 if (loop_dump_stream
)
2377 fprintf (loop_dump_stream
,
2378 "Loop unrolling: Iteration var not an integer.\n");
2381 else if (reg_iv_type
[REGNO (iteration_var
)] == BASIC_INDUCT
)
2383 /* Grab initial value, only useful if it is a constant. */
2384 bl
= reg_biv_class
[REGNO (iteration_var
)];
2385 *initial_value
= bl
->initial_value
;
2387 *increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2389 else if (reg_iv_type
[REGNO (iteration_var
)] == GENERAL_INDUCT
)
2392 /* ??? The code below does not work because the incorrect number of
2393 iterations is calculated when the biv is incremented after the giv
2394 is set (which is the usual case). This can probably be accounted
2395 for by biasing the initial_value by subtracting the amount of the
2396 increment that occurs between the giv set and the giv test. However,
2397 a giv as an iterator is very rare, so it does not seem worthwhile
2399 /* ??? An example failure is: i = 6; do {;} while (i++ < 9). */
2400 if (loop_dump_stream
)
2401 fprintf (loop_dump_stream
,
2402 "Loop unrolling: Giv iterators are not handled.\n");
2405 /* Initial value is mult_val times the biv's initial value plus
2406 add_val. Only useful if it is a constant. */
2407 v
= reg_iv_info
[REGNO (iteration_var
)];
2408 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
2409 *initial_value
= fold_rtx_mult_add (v
->mult_val
, bl
->initial_value
,
2410 v
->add_val
, v
->mode
);
2412 /* Increment value is mult_val times the increment value of the biv. */
2414 *increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2416 *increment
= fold_rtx_mult_add (v
->mult_val
, *increment
, const0_rtx
,
2422 if (loop_dump_stream
)
2423 fprintf (loop_dump_stream
,
2424 "Loop unrolling: Not basic or general induction var.\n");
2430 /* For each biv and giv, determine whether it can be safely split into
2431 a different variable for each unrolled copy of the loop body. If it
2432 is safe to split, then indicate that by saving some useful info
2433 in the splittable_regs array.
2435 If the loop is being completely unrolled, then splittable_regs will hold
2436 the current value of the induction variable while the loop is unrolled.
2437 It must be set to the initial value of the induction variable here.
2438 Otherwise, splittable_regs will hold the difference between the current
2439 value of the induction variable and the value the induction variable had
2440 at the top of the loop. It must be set to the value 0 here.
2442 Returns the total number of instructions that set registers that are
2445 /* ?? If the loop is only unrolled twice, then most of the restrictions to
2446 constant values are unnecessary, since we can easily calculate increment
2447 values in this case even if nothing is constant. The increment value
2448 should not involve a multiply however. */
2450 /* ?? Even if the biv/giv increment values aren't constant, it may still
2451 be beneficial to split the variable if the loop is only unrolled a few
2452 times, since multiplies by small integers (1,2,3,4) are very cheap. */
2455 find_splittable_regs (unroll_type
, loop_start
, loop_end
, end_insert_before
,
2456 unroll_number
, n_iterations
)
2457 enum unroll_types unroll_type
;
2458 rtx loop_start
, loop_end
;
2459 rtx end_insert_before
;
2461 unsigned HOST_WIDE_INT n_iterations
;
2463 struct iv_class
*bl
;
2464 struct induction
*v
;
2466 rtx biv_final_value
;
2470 for (bl
= loop_iv_list
; bl
; bl
= bl
->next
)
2472 /* Biv_total_increment must return a constant value,
2473 otherwise we can not calculate the split values. */
2475 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2476 if (! increment
|| GET_CODE (increment
) != CONST_INT
)
2479 /* The loop must be unrolled completely, or else have a known number
2480 of iterations and only one exit, or else the biv must be dead
2481 outside the loop, or else the final value must be known. Otherwise,
2482 it is unsafe to split the biv since it may not have the proper
2483 value on loop exit. */
2485 /* loop_number_exit_count is non-zero if the loop has an exit other than
2486 a fall through at the end. */
2489 biv_final_value
= 0;
2490 if (unroll_type
!= UNROLL_COMPLETELY
2491 && (loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]]
2492 || unroll_type
== UNROLL_NAIVE
)
2493 && (uid_luid
[REGNO_LAST_UID (bl
->regno
)] >= INSN_LUID (loop_end
)
2495 || INSN_UID (bl
->init_insn
) >= max_uid_for_loop
2496 || (uid_luid
[REGNO_FIRST_UID (bl
->regno
)]
2497 < INSN_LUID (bl
->init_insn
))
2498 || reg_mentioned_p (bl
->biv
->dest_reg
, SET_SRC (bl
->init_set
)))
2499 && ! (biv_final_value
= final_biv_value (bl
, loop_start
, loop_end
,
2503 /* If any of the insns setting the BIV don't do so with a simple
2504 PLUS, we don't know how to split it. */
2505 for (v
= bl
->biv
; biv_splittable
&& v
; v
= v
->next_iv
)
2506 if ((tem
= single_set (v
->insn
)) == 0
2507 || GET_CODE (SET_DEST (tem
)) != REG
2508 || REGNO (SET_DEST (tem
)) != bl
->regno
2509 || GET_CODE (SET_SRC (tem
)) != PLUS
)
2512 /* If final value is non-zero, then must emit an instruction which sets
2513 the value of the biv to the proper value. This is done after
2514 handling all of the givs, since some of them may need to use the
2515 biv's value in their initialization code. */
2517 /* This biv is splittable. If completely unrolling the loop, save
2518 the biv's initial value. Otherwise, save the constant zero. */
2520 if (biv_splittable
== 1)
2522 if (unroll_type
== UNROLL_COMPLETELY
)
2524 /* If the initial value of the biv is itself (i.e. it is too
2525 complicated for strength_reduce to compute), or is a hard
2526 register, or it isn't invariant, then we must create a new
2527 pseudo reg to hold the initial value of the biv. */
2529 if (GET_CODE (bl
->initial_value
) == REG
2530 && (REGNO (bl
->initial_value
) == bl
->regno
2531 || REGNO (bl
->initial_value
) < FIRST_PSEUDO_REGISTER
2532 || ! invariant_p (bl
->initial_value
)))
2534 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2536 record_base_value (REGNO (tem
), bl
->biv
->add_val
, 0);
2537 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2540 if (loop_dump_stream
)
2541 fprintf (loop_dump_stream
, "Biv %d initial value remapped to %d.\n",
2542 bl
->regno
, REGNO (tem
));
2544 splittable_regs
[bl
->regno
] = tem
;
2547 splittable_regs
[bl
->regno
] = bl
->initial_value
;
2550 splittable_regs
[bl
->regno
] = const0_rtx
;
2552 /* Save the number of instructions that modify the biv, so that
2553 we can treat the last one specially. */
2555 splittable_regs_updates
[bl
->regno
] = bl
->biv_count
;
2556 result
+= bl
->biv_count
;
2558 if (loop_dump_stream
)
2559 fprintf (loop_dump_stream
,
2560 "Biv %d safe to split.\n", bl
->regno
);
2563 /* Check every giv that depends on this biv to see whether it is
2564 splittable also. Even if the biv isn't splittable, givs which
2565 depend on it may be splittable if the biv is live outside the
2566 loop, and the givs aren't. */
2568 result
+= find_splittable_givs (bl
, unroll_type
, loop_start
, loop_end
,
2569 increment
, unroll_number
);
2571 /* If final value is non-zero, then must emit an instruction which sets
2572 the value of the biv to the proper value. This is done after
2573 handling all of the givs, since some of them may need to use the
2574 biv's value in their initialization code. */
2575 if (biv_final_value
)
2577 /* If the loop has multiple exits, emit the insns before the
2578 loop to ensure that it will always be executed no matter
2579 how the loop exits. Otherwise emit the insn after the loop,
2580 since this is slightly more efficient. */
2581 if (! loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]])
2582 emit_insn_before (gen_move_insn (bl
->biv
->src_reg
,
2587 /* Create a new register to hold the value of the biv, and then
2588 set the biv to its final value before the loop start. The biv
2589 is set to its final value before loop start to ensure that
2590 this insn will always be executed, no matter how the loop
2592 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2593 record_base_value (REGNO (tem
), bl
->biv
->add_val
, 0);
2595 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2597 emit_insn_before (gen_move_insn (bl
->biv
->src_reg
,
2601 if (loop_dump_stream
)
2602 fprintf (loop_dump_stream
, "Biv %d mapped to %d for split.\n",
2603 REGNO (bl
->biv
->src_reg
), REGNO (tem
));
2605 /* Set up the mapping from the original biv register to the new
2607 bl
->biv
->src_reg
= tem
;
2614 /* Return 1 if the first and last unrolled copy of the address giv V is valid
2615 for the instruction that is using it. Do not make any changes to that
2619 verify_addresses (v
, giv_inc
, unroll_number
)
2620 struct induction
*v
;
2625 rtx orig_addr
= *v
->location
;
2626 rtx last_addr
= plus_constant (v
->dest_reg
,
2627 INTVAL (giv_inc
) * (unroll_number
- 1));
2629 /* First check to see if either address would fail. Handle the fact
2630 that we have may have a match_dup. */
2631 if (! validate_replace_rtx (*v
->location
, v
->dest_reg
, v
->insn
)
2632 || ! validate_replace_rtx (*v
->location
, last_addr
, v
->insn
))
2635 /* Now put things back the way they were before. This should always
2637 if (! validate_replace_rtx (*v
->location
, orig_addr
, v
->insn
))
2643 /* For every giv based on the biv BL, check to determine whether it is
2644 splittable. This is a subroutine to find_splittable_regs ().
2646 Return the number of instructions that set splittable registers. */
2649 find_splittable_givs (bl
, unroll_type
, loop_start
, loop_end
, increment
,
2651 struct iv_class
*bl
;
2652 enum unroll_types unroll_type
;
2653 rtx loop_start
, loop_end
;
2657 struct induction
*v
, *v2
;
2662 /* Scan the list of givs, and set the same_insn field when there are
2663 multiple identical givs in the same insn. */
2664 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
2665 for (v2
= v
->next_iv
; v2
; v2
= v2
->next_iv
)
2666 if (v
->insn
== v2
->insn
&& rtx_equal_p (v
->new_reg
, v2
->new_reg
)
2670 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
2674 /* Only split the giv if it has already been reduced, or if the loop is
2675 being completely unrolled. */
2676 if (unroll_type
!= UNROLL_COMPLETELY
&& v
->ignore
)
2679 /* The giv can be split if the insn that sets the giv is executed once
2680 and only once on every iteration of the loop. */
2681 /* An address giv can always be split. v->insn is just a use not a set,
2682 and hence it does not matter whether it is always executed. All that
2683 matters is that all the biv increments are always executed, and we
2684 won't reach here if they aren't. */
2685 if (v
->giv_type
!= DEST_ADDR
2686 && (! v
->always_computable
2687 || back_branch_in_range_p (v
->insn
, loop_start
, loop_end
)))
2690 /* The giv increment value must be a constant. */
2691 giv_inc
= fold_rtx_mult_add (v
->mult_val
, increment
, const0_rtx
,
2693 if (! giv_inc
|| GET_CODE (giv_inc
) != CONST_INT
)
2696 /* The loop must be unrolled completely, or else have a known number of
2697 iterations and only one exit, or else the giv must be dead outside
2698 the loop, or else the final value of the giv must be known.
2699 Otherwise, it is not safe to split the giv since it may not have the
2700 proper value on loop exit. */
2702 /* The used outside loop test will fail for DEST_ADDR givs. They are
2703 never used outside the loop anyways, so it is always safe to split a
2707 if (unroll_type
!= UNROLL_COMPLETELY
2708 && (loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]]
2709 || unroll_type
== UNROLL_NAIVE
)
2710 && v
->giv_type
!= DEST_ADDR
2711 /* The next part is true if the pseudo is used outside the loop.
2712 We assume that this is true for any pseudo created after loop
2713 starts, because we don't have a reg_n_info entry for them. */
2714 && (REGNO (v
->dest_reg
) >= max_reg_before_loop
2715 || (REGNO_FIRST_UID (REGNO (v
->dest_reg
)) != INSN_UID (v
->insn
)
2716 /* Check for the case where the pseudo is set by a shift/add
2717 sequence, in which case the first insn setting the pseudo
2718 is the first insn of the shift/add sequence. */
2719 && (! (tem
= find_reg_note (v
->insn
, REG_RETVAL
, NULL_RTX
))
2720 || (REGNO_FIRST_UID (REGNO (v
->dest_reg
))
2721 != INSN_UID (XEXP (tem
, 0)))))
2722 /* Line above always fails if INSN was moved by loop opt. */
2723 || (uid_luid
[REGNO_LAST_UID (REGNO (v
->dest_reg
))]
2724 >= INSN_LUID (loop_end
)))
2725 && ! (final_value
= v
->final_value
))
2729 /* Currently, non-reduced/final-value givs are never split. */
2730 /* Should emit insns after the loop if possible, as the biv final value
2733 /* If the final value is non-zero, and the giv has not been reduced,
2734 then must emit an instruction to set the final value. */
2735 if (final_value
&& !v
->new_reg
)
2737 /* Create a new register to hold the value of the giv, and then set
2738 the giv to its final value before the loop start. The giv is set
2739 to its final value before loop start to ensure that this insn
2740 will always be executed, no matter how we exit. */
2741 tem
= gen_reg_rtx (v
->mode
);
2742 emit_insn_before (gen_move_insn (tem
, v
->dest_reg
), loop_start
);
2743 emit_insn_before (gen_move_insn (v
->dest_reg
, final_value
),
2746 if (loop_dump_stream
)
2747 fprintf (loop_dump_stream
, "Giv %d mapped to %d for split.\n",
2748 REGNO (v
->dest_reg
), REGNO (tem
));
2754 /* This giv is splittable. If completely unrolling the loop, save the
2755 giv's initial value. Otherwise, save the constant zero for it. */
2757 if (unroll_type
== UNROLL_COMPLETELY
)
2759 /* It is not safe to use bl->initial_value here, because it may not
2760 be invariant. It is safe to use the initial value stored in
2761 the splittable_regs array if it is set. In rare cases, it won't
2762 be set, so then we do exactly the same thing as
2763 find_splittable_regs does to get a safe value. */
2764 rtx biv_initial_value
;
2766 if (splittable_regs
[bl
->regno
])
2767 biv_initial_value
= splittable_regs
[bl
->regno
];
2768 else if (GET_CODE (bl
->initial_value
) != REG
2769 || (REGNO (bl
->initial_value
) != bl
->regno
2770 && REGNO (bl
->initial_value
) >= FIRST_PSEUDO_REGISTER
))
2771 biv_initial_value
= bl
->initial_value
;
2774 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2776 record_base_value (REGNO (tem
), bl
->biv
->add_val
, 0);
2777 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2779 biv_initial_value
= tem
;
2781 value
= fold_rtx_mult_add (v
->mult_val
, biv_initial_value
,
2782 v
->add_val
, v
->mode
);
2789 /* If a giv was combined with another giv, then we can only split
2790 this giv if the giv it was combined with was reduced. This
2791 is because the value of v->new_reg is meaningless in this
2793 if (v
->same
&& ! v
->same
->new_reg
)
2795 if (loop_dump_stream
)
2796 fprintf (loop_dump_stream
,
2797 "giv combined with unreduced giv not split.\n");
2800 /* If the giv is an address destination, it could be something other
2801 than a simple register, these have to be treated differently. */
2802 else if (v
->giv_type
== DEST_REG
)
2804 /* If value is not a constant, register, or register plus
2805 constant, then compute its value into a register before
2806 loop start. This prevents invalid rtx sharing, and should
2807 generate better code. We can use bl->initial_value here
2808 instead of splittable_regs[bl->regno] because this code
2809 is going before the loop start. */
2810 if (unroll_type
== UNROLL_COMPLETELY
2811 && GET_CODE (value
) != CONST_INT
2812 && GET_CODE (value
) != REG
2813 && (GET_CODE (value
) != PLUS
2814 || GET_CODE (XEXP (value
, 0)) != REG
2815 || GET_CODE (XEXP (value
, 1)) != CONST_INT
))
2817 rtx tem
= gen_reg_rtx (v
->mode
);
2818 record_base_value (REGNO (tem
), v
->add_val
, 0);
2819 emit_iv_add_mult (bl
->initial_value
, v
->mult_val
,
2820 v
->add_val
, tem
, loop_start
);
2824 splittable_regs
[REGNO (v
->new_reg
)] = value
;
2828 /* Splitting address givs is useful since it will often allow us
2829 to eliminate some increment insns for the base giv as
2832 /* If the addr giv is combined with a dest_reg giv, then all
2833 references to that dest reg will be remapped, which is NOT
2834 what we want for split addr regs. We always create a new
2835 register for the split addr giv, just to be safe. */
2837 /* If we have multiple identical address givs within a
2838 single instruction, then use a single pseudo reg for
2839 both. This is necessary in case one is a match_dup
2842 v
->const_adjust
= 0;
2846 v
->dest_reg
= v
->same_insn
->dest_reg
;
2847 if (loop_dump_stream
)
2848 fprintf (loop_dump_stream
,
2849 "Sharing address givs in insn %d\n",
2850 INSN_UID (v
->insn
));
2852 /* If multiple address GIVs have been combined with the
2853 same dest_reg GIV, do not create a new register for
2855 else if (unroll_type
!= UNROLL_COMPLETELY
2856 && v
->giv_type
== DEST_ADDR
2857 && v
->same
&& v
->same
->giv_type
== DEST_ADDR
2858 && v
->same
->unrolled
2859 /* combine_givs_p may return true for some cases
2860 where the add and mult values are not equal.
2861 To share a register here, the values must be
2863 && rtx_equal_p (v
->same
->mult_val
, v
->mult_val
)
2864 && rtx_equal_p (v
->same
->add_val
, v
->add_val
))
2867 v
->dest_reg
= v
->same
->dest_reg
;
2870 else if (unroll_type
!= UNROLL_COMPLETELY
)
2872 /* If not completely unrolling the loop, then create a new
2873 register to hold the split value of the DEST_ADDR giv.
2874 Emit insn to initialize its value before loop start. */
2876 rtx tem
= gen_reg_rtx (v
->mode
);
2877 record_base_value (REGNO (tem
), v
->add_val
, 0);
2879 /* If the address giv has a constant in its new_reg value,
2880 then this constant can be pulled out and put in value,
2881 instead of being part of the initialization code. */
2883 if (GET_CODE (v
->new_reg
) == PLUS
2884 && GET_CODE (XEXP (v
->new_reg
, 1)) == CONST_INT
)
2887 = plus_constant (tem
, INTVAL (XEXP (v
->new_reg
,1)));
2889 /* Only succeed if this will give valid addresses.
2890 Try to validate both the first and the last
2891 address resulting from loop unrolling, if
2892 one fails, then can't do const elim here. */
2893 if (verify_addresses (v
, giv_inc
, unroll_number
))
2895 /* Save the negative of the eliminated const, so
2896 that we can calculate the dest_reg's increment
2898 v
->const_adjust
= - INTVAL (XEXP (v
->new_reg
, 1));
2900 v
->new_reg
= XEXP (v
->new_reg
, 0);
2901 if (loop_dump_stream
)
2902 fprintf (loop_dump_stream
,
2903 "Eliminating constant from giv %d\n",
2912 /* If the address hasn't been checked for validity yet, do so
2913 now, and fail completely if either the first or the last
2914 unrolled copy of the address is not a valid address
2915 for the instruction that uses it. */
2916 if (v
->dest_reg
== tem
2917 && ! verify_addresses (v
, giv_inc
, unroll_number
))
2919 for (v2
= v
->next_iv
; v2
; v2
= v2
->next_iv
)
2920 if (v2
->same_insn
== v
)
2923 if (loop_dump_stream
)
2924 fprintf (loop_dump_stream
,
2925 "Invalid address for giv at insn %d\n",
2926 INSN_UID (v
->insn
));
2930 /* We set this after the address check, to guarantee that
2931 the register will be initialized. */
2934 /* To initialize the new register, just move the value of
2935 new_reg into it. This is not guaranteed to give a valid
2936 instruction on machines with complex addressing modes.
2937 If we can't recognize it, then delete it and emit insns
2938 to calculate the value from scratch. */
2939 emit_insn_before (gen_rtx_SET (VOIDmode
, tem
,
2940 copy_rtx (v
->new_reg
)),
2942 if (recog_memoized (PREV_INSN (loop_start
)) < 0)
2946 /* We can't use bl->initial_value to compute the initial
2947 value, because the loop may have been preconditioned.
2948 We must calculate it from NEW_REG. Try using
2949 force_operand instead of emit_iv_add_mult. */
2950 delete_insn (PREV_INSN (loop_start
));
2953 ret
= force_operand (v
->new_reg
, tem
);
2955 emit_move_insn (tem
, ret
);
2956 sequence
= gen_sequence ();
2958 emit_insn_before (sequence
, loop_start
);
2960 if (loop_dump_stream
)
2961 fprintf (loop_dump_stream
,
2962 "Invalid init insn, rewritten.\n");
2967 v
->dest_reg
= value
;
2969 /* Check the resulting address for validity, and fail
2970 if the resulting address would be invalid. */
2971 if (! verify_addresses (v
, giv_inc
, unroll_number
))
2973 for (v2
= v
->next_iv
; v2
; v2
= v2
->next_iv
)
2974 if (v2
->same_insn
== v
)
2977 if (loop_dump_stream
)
2978 fprintf (loop_dump_stream
,
2979 "Invalid address for giv at insn %d\n",
2980 INSN_UID (v
->insn
));
2985 /* Store the value of dest_reg into the insn. This sharing
2986 will not be a problem as this insn will always be copied
2989 *v
->location
= v
->dest_reg
;
2991 /* If this address giv is combined with a dest reg giv, then
2992 save the base giv's induction pointer so that we will be
2993 able to handle this address giv properly. The base giv
2994 itself does not have to be splittable. */
2996 if (v
->same
&& v
->same
->giv_type
== DEST_REG
)
2997 addr_combined_regs
[REGNO (v
->same
->new_reg
)] = v
->same
;
2999 if (GET_CODE (v
->new_reg
) == REG
)
3001 /* This giv maybe hasn't been combined with any others.
3002 Make sure that it's giv is marked as splittable here. */
3004 splittable_regs
[REGNO (v
->new_reg
)] = value
;
3006 /* Make it appear to depend upon itself, so that the
3007 giv will be properly split in the main loop above. */
3011 addr_combined_regs
[REGNO (v
->new_reg
)] = v
;
3015 if (loop_dump_stream
)
3016 fprintf (loop_dump_stream
, "DEST_ADDR giv being split.\n");
3022 /* Currently, unreduced giv's can't be split. This is not too much
3023 of a problem since unreduced giv's are not live across loop
3024 iterations anyways. When unrolling a loop completely though,
3025 it makes sense to reduce&split givs when possible, as this will
3026 result in simpler instructions, and will not require that a reg
3027 be live across loop iterations. */
3029 splittable_regs
[REGNO (v
->dest_reg
)] = value
;
3030 fprintf (stderr
, "Giv %d at insn %d not reduced\n",
3031 REGNO (v
->dest_reg
), INSN_UID (v
->insn
));
3037 /* Unreduced givs are only updated once by definition. Reduced givs
3038 are updated as many times as their biv is. Mark it so if this is
3039 a splittable register. Don't need to do anything for address givs
3040 where this may not be a register. */
3042 if (GET_CODE (v
->new_reg
) == REG
)
3046 count
= reg_biv_class
[REGNO (v
->src_reg
)]->biv_count
;
3048 splittable_regs_updates
[REGNO (v
->new_reg
)] = count
;
3053 if (loop_dump_stream
)
3057 if (GET_CODE (v
->dest_reg
) == CONST_INT
)
3059 else if (GET_CODE (v
->dest_reg
) != REG
)
3060 regnum
= REGNO (XEXP (v
->dest_reg
, 0));
3062 regnum
= REGNO (v
->dest_reg
);
3063 fprintf (loop_dump_stream
, "Giv %d at insn %d safe to split.\n",
3064 regnum
, INSN_UID (v
->insn
));
3071 /* Try to prove that the register is dead after the loop exits. Trace every
3072 loop exit looking for an insn that will always be executed, which sets
3073 the register to some value, and appears before the first use of the register
3074 is found. If successful, then return 1, otherwise return 0. */
3076 /* ?? Could be made more intelligent in the handling of jumps, so that
3077 it can search past if statements and other similar structures. */
3080 reg_dead_after_loop (reg
, loop_start
, loop_end
)
3081 rtx reg
, loop_start
, loop_end
;
3086 int label_count
= 0;
3087 int this_loop_num
= uid_loop_num
[INSN_UID (loop_start
)];
3089 /* In addition to checking all exits of this loop, we must also check
3090 all exits of inner nested loops that would exit this loop. We don't
3091 have any way to identify those, so we just give up if there are any
3092 such inner loop exits. */
3094 for (label
= loop_number_exit_labels
[this_loop_num
]; label
;
3095 label
= LABEL_NEXTREF (label
))
3098 if (label_count
!= loop_number_exit_count
[this_loop_num
])
3101 /* HACK: Must also search the loop fall through exit, create a label_ref
3102 here which points to the loop_end, and append the loop_number_exit_labels
3104 label
= gen_rtx_LABEL_REF (VOIDmode
, loop_end
);
3105 LABEL_NEXTREF (label
) = loop_number_exit_labels
[this_loop_num
];
3107 for ( ; label
; label
= LABEL_NEXTREF (label
))
3109 /* Succeed if find an insn which sets the biv or if reach end of
3110 function. Fail if find an insn that uses the biv, or if come to
3111 a conditional jump. */
3113 insn
= NEXT_INSN (XEXP (label
, 0));
3116 code
= GET_CODE (insn
);
3117 if (GET_RTX_CLASS (code
) == 'i')
3121 if (reg_referenced_p (reg
, PATTERN (insn
)))
3124 set
= single_set (insn
);
3125 if (set
&& rtx_equal_p (SET_DEST (set
), reg
))
3129 if (code
== JUMP_INSN
)
3131 if (GET_CODE (PATTERN (insn
)) == RETURN
)
3133 else if (! simplejump_p (insn
)
3134 /* Prevent infinite loop following infinite loops. */
3135 || jump_count
++ > 20)
3138 insn
= JUMP_LABEL (insn
);
3141 insn
= NEXT_INSN (insn
);
3145 /* Success, the register is dead on all loop exits. */
3149 /* Try to calculate the final value of the biv, the value it will have at
3150 the end of the loop. If we can do it, return that value. */
3153 final_biv_value (bl
, loop_start
, loop_end
, n_iterations
)
3154 struct iv_class
*bl
;
3155 rtx loop_start
, loop_end
;
3156 unsigned HOST_WIDE_INT n_iterations
;
3160 /* ??? This only works for MODE_INT biv's. Reject all others for now. */
3162 if (GET_MODE_CLASS (bl
->biv
->mode
) != MODE_INT
)
3165 /* The final value for reversed bivs must be calculated differently than
3166 for ordinary bivs. In this case, there is already an insn after the
3167 loop which sets this biv's final value (if necessary), and there are
3168 no other loop exits, so we can return any value. */
3171 if (loop_dump_stream
)
3172 fprintf (loop_dump_stream
,
3173 "Final biv value for %d, reversed biv.\n", bl
->regno
);
3178 /* Try to calculate the final value as initial value + (number of iterations
3179 * increment). For this to work, increment must be invariant, the only
3180 exit from the loop must be the fall through at the bottom (otherwise
3181 it may not have its final value when the loop exits), and the initial
3182 value of the biv must be invariant. */
3184 if (n_iterations
!= 0
3185 && ! loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]]
3186 && invariant_p (bl
->initial_value
))
3188 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
3190 if (increment
&& invariant_p (increment
))
3192 /* Can calculate the loop exit value, emit insns after loop
3193 end to calculate this value into a temporary register in
3194 case it is needed later. */
3196 tem
= gen_reg_rtx (bl
->biv
->mode
);
3197 record_base_value (REGNO (tem
), bl
->biv
->add_val
, 0);
3198 /* Make sure loop_end is not the last insn. */
3199 if (NEXT_INSN (loop_end
) == 0)
3200 emit_note_after (NOTE_INSN_DELETED
, loop_end
);
3201 emit_iv_add_mult (increment
, GEN_INT (n_iterations
),
3202 bl
->initial_value
, tem
, NEXT_INSN (loop_end
));
3204 if (loop_dump_stream
)
3205 fprintf (loop_dump_stream
,
3206 "Final biv value for %d, calculated.\n", bl
->regno
);
3212 /* Check to see if the biv is dead at all loop exits. */
3213 if (reg_dead_after_loop (bl
->biv
->src_reg
, loop_start
, loop_end
))
3215 if (loop_dump_stream
)
3216 fprintf (loop_dump_stream
,
3217 "Final biv value for %d, biv dead after loop exit.\n",
3226 /* Try to calculate the final value of the giv, the value it will have at
3227 the end of the loop. If we can do it, return that value. */
3230 final_giv_value (v
, loop_start
, loop_end
, n_iterations
)
3231 struct induction
*v
;
3232 rtx loop_start
, loop_end
;
3233 unsigned HOST_WIDE_INT n_iterations
;
3235 struct iv_class
*bl
;
3238 rtx insert_before
, seq
;
3240 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
3242 /* The final value for givs which depend on reversed bivs must be calculated
3243 differently than for ordinary givs. In this case, there is already an
3244 insn after the loop which sets this giv's final value (if necessary),
3245 and there are no other loop exits, so we can return any value. */
3248 if (loop_dump_stream
)
3249 fprintf (loop_dump_stream
,
3250 "Final giv value for %d, depends on reversed biv\n",
3251 REGNO (v
->dest_reg
));
3255 /* Try to calculate the final value as a function of the biv it depends
3256 upon. The only exit from the loop must be the fall through at the bottom
3257 (otherwise it may not have its final value when the loop exits). */
3259 /* ??? Can calculate the final giv value by subtracting off the
3260 extra biv increments times the giv's mult_val. The loop must have
3261 only one exit for this to work, but the loop iterations does not need
3264 if (n_iterations
!= 0
3265 && ! loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]])
3267 /* ?? It is tempting to use the biv's value here since these insns will
3268 be put after the loop, and hence the biv will have its final value
3269 then. However, this fails if the biv is subsequently eliminated.
3270 Perhaps determine whether biv's are eliminable before trying to
3271 determine whether giv's are replaceable so that we can use the
3272 biv value here if it is not eliminable. */
3274 /* We are emitting code after the end of the loop, so we must make
3275 sure that bl->initial_value is still valid then. It will still
3276 be valid if it is invariant. */
3278 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
3280 if (increment
&& invariant_p (increment
)
3281 && invariant_p (bl
->initial_value
))
3283 /* Can calculate the loop exit value of its biv as
3284 (n_iterations * increment) + initial_value */
3286 /* The loop exit value of the giv is then
3287 (final_biv_value - extra increments) * mult_val + add_val.
3288 The extra increments are any increments to the biv which
3289 occur in the loop after the giv's value is calculated.
3290 We must search from the insn that sets the giv to the end
3291 of the loop to calculate this value. */
3293 insert_before
= NEXT_INSN (loop_end
);
3295 /* Put the final biv value in tem. */
3296 tem
= gen_reg_rtx (bl
->biv
->mode
);
3297 record_base_value (REGNO (tem
), bl
->biv
->add_val
, 0);
3298 emit_iv_add_mult (increment
, GEN_INT (n_iterations
),
3299 bl
->initial_value
, tem
, insert_before
);
3301 /* Subtract off extra increments as we find them. */
3302 for (insn
= NEXT_INSN (v
->insn
); insn
!= loop_end
;
3303 insn
= NEXT_INSN (insn
))
3305 struct induction
*biv
;
3307 for (biv
= bl
->biv
; biv
; biv
= biv
->next_iv
)
3308 if (biv
->insn
== insn
)
3311 tem
= expand_binop (GET_MODE (tem
), sub_optab
, tem
,
3312 biv
->add_val
, NULL_RTX
, 0,
3314 seq
= gen_sequence ();
3316 emit_insn_before (seq
, insert_before
);
3320 /* Now calculate the giv's final value. */
3321 emit_iv_add_mult (tem
, v
->mult_val
, v
->add_val
, tem
,
3324 if (loop_dump_stream
)
3325 fprintf (loop_dump_stream
,
3326 "Final giv value for %d, calc from biv's value.\n",
3327 REGNO (v
->dest_reg
));
3333 /* Replaceable giv's should never reach here. */
3337 /* Check to see if the biv is dead at all loop exits. */
3338 if (reg_dead_after_loop (v
->dest_reg
, loop_start
, loop_end
))
3340 if (loop_dump_stream
)
3341 fprintf (loop_dump_stream
,
3342 "Final giv value for %d, giv dead after loop exit.\n",
3343 REGNO (v
->dest_reg
));
3352 /* Look back before LOOP_START for then insn that sets REG and return
3353 the equivalent constant if there is a REG_EQUAL note otherwise just
3354 the SET_SRC of REG. */
3357 loop_find_equiv_value (loop_start
, reg
)
3365 for (insn
= PREV_INSN (loop_start
); insn
; insn
= PREV_INSN (insn
))
3367 if (GET_CODE (insn
) == CODE_LABEL
)
3370 else if (GET_RTX_CLASS (GET_CODE (insn
)) == 'i'
3371 && reg_set_p (reg
, insn
))
3373 /* We found the last insn before the loop that sets the register.
3374 If it sets the entire register, and has a REG_EQUAL note,
3375 then use the value of the REG_EQUAL note. */
3376 if ((set
= single_set (insn
))
3377 && (SET_DEST (set
) == reg
))
3379 rtx note
= find_reg_note (insn
, REG_EQUAL
, NULL_RTX
);
3381 /* Only use the REG_EQUAL note if it is a constant.
3382 Other things, divide in particular, will cause
3383 problems later if we use them. */
3384 if (note
&& GET_CODE (XEXP (note
, 0)) != EXPR_LIST
3385 && CONSTANT_P (XEXP (note
, 0)))
3386 ret
= XEXP (note
, 0);
3388 ret
= SET_SRC (set
);
3397 /* Return a simplified rtx for the expression OP - REG.
3399 REG must appear in OP, and OP must be a register or the sum of a register
3402 Thus, the return value must be const0_rtx or the second term.
3404 The caller is responsible for verifying that REG appears in OP and OP has
3408 subtract_reg_term (op
, reg
)
3413 if (GET_CODE (op
) == PLUS
)
3415 if (XEXP (op
, 0) == reg
)
3416 return XEXP (op
, 1);
3417 else if (XEXP (op
, 1) == reg
)
3418 return XEXP (op
, 0);
3420 /* OP does not contain REG as a term. */
3425 /* Find and return register term common to both expressions OP0 and
3426 OP1 or NULL_RTX if no such term exists. Each expression must be a
3427 REG or a PLUS of a REG. */
3430 find_common_reg_term (op0
, op1
)
3433 if ((GET_CODE (op0
) == REG
|| GET_CODE (op0
) == PLUS
)
3434 && (GET_CODE (op1
) == REG
|| GET_CODE (op1
) == PLUS
))
3441 if (GET_CODE (op0
) == PLUS
)
3442 op01
= XEXP (op0
, 1), op00
= XEXP (op0
, 0);
3444 op01
= const0_rtx
, op00
= op0
;
3446 if (GET_CODE (op1
) == PLUS
)
3447 op11
= XEXP (op1
, 1), op10
= XEXP (op1
, 0);
3449 op11
= const0_rtx
, op10
= op1
;
3451 /* Find and return common register term if present. */
3452 if (REG_P (op00
) && (op00
== op10
|| op00
== op11
))
3454 else if (REG_P (op01
) && (op01
== op10
|| op01
== op11
))
3458 /* No common register term found. */
3463 /* Calculate the number of loop iterations. Returns the exact number of loop
3464 iterations if it can be calculated, otherwise returns zero. */
3466 unsigned HOST_WIDE_INT
3467 loop_iterations (loop_start
, loop_end
, loop_info
)
3468 rtx loop_start
, loop_end
;
3469 struct loop_info
*loop_info
;
3471 rtx comparison
, comparison_value
;
3472 rtx iteration_var
, initial_value
, increment
, final_value
;
3473 enum rtx_code comparison_code
;
3474 HOST_WIDE_INT abs_inc
;
3475 unsigned HOST_WIDE_INT abs_diff
;
3478 int unsigned_p
, compare_dir
, final_larger
;
3483 loop_info
->n_iterations
= 0;
3484 loop_info
->initial_value
= 0;
3485 loop_info
->initial_equiv_value
= 0;
3486 loop_info
->comparison_value
= 0;
3487 loop_info
->final_value
= 0;
3488 loop_info
->final_equiv_value
= 0;
3489 loop_info
->increment
= 0;
3490 loop_info
->iteration_var
= 0;
3491 loop_info
->unroll_number
= 1;
3492 loop_info
->vtop
= 0;
3494 /* First find the iteration variable. If the last insn is a conditional
3495 branch, and the insn before tests a register value, make that the
3496 iteration variable. */
3498 /* We used to use prev_nonnote_insn here, but that fails because it might
3499 accidentally get the branch for a contained loop if the branch for this
3500 loop was deleted. We can only trust branches immediately before the
3502 last_loop_insn
= PREV_INSN (loop_end
);
3504 comparison
= get_condition_for_loop (last_loop_insn
);
3505 if (comparison
== 0)
3507 if (loop_dump_stream
)
3508 fprintf (loop_dump_stream
,
3509 "Loop iterations: No final conditional branch found.\n");
3513 /* ??? Get_condition may switch position of induction variable and
3514 invariant register when it canonicalizes the comparison. */
3516 comparison_code
= GET_CODE (comparison
);
3517 iteration_var
= XEXP (comparison
, 0);
3518 comparison_value
= XEXP (comparison
, 1);
3520 /* Check if there is a NOTE_INSN_LOOP_VTOP note. If there is,
3521 that means that this is a for or while style loop, with
3522 a loop exit test at the start. Thus, we can assume that
3523 the loop condition was true when the loop was entered.
3525 We start at the end and search backwards for the previous
3526 NOTE. If there is no NOTE_INSN_LOOP_VTOP for this loop,
3527 the search will stop at the NOTE_INSN_LOOP_CONT. */
3530 vtop
= PREV_INSN (vtop
);
3531 while (GET_CODE (vtop
) != NOTE
3532 || NOTE_LINE_NUMBER (vtop
) > 0
3533 || NOTE_LINE_NUMBER (vtop
) == NOTE_REPEATED_LINE_NUMBER
3534 || NOTE_LINE_NUMBER (vtop
) == NOTE_INSN_DELETED
);
3535 if (NOTE_LINE_NUMBER (vtop
) != NOTE_INSN_LOOP_VTOP
)
3537 loop_info
->vtop
= vtop
;
3539 if (GET_CODE (iteration_var
) != REG
)
3541 if (loop_dump_stream
)
3542 fprintf (loop_dump_stream
,
3543 "Loop iterations: Comparison not against register.\n");
3547 /* Loop iterations is always called before any new registers are created
3548 now, so this should never occur. */
3550 if (REGNO (iteration_var
) >= max_reg_before_loop
)
3553 iteration_info (iteration_var
, &initial_value
, &increment
,
3554 loop_start
, loop_end
);
3555 if (initial_value
== 0)
3556 /* iteration_info already printed a message. */
3561 switch (comparison_code
)
3576 /* Cannot determine loop iterations with this case. */
3595 /* If the comparison value is an invariant register, then try to find
3596 its value from the insns before the start of the loop. */
3598 final_value
= comparison_value
;
3599 if (GET_CODE (comparison_value
) == REG
&& invariant_p (comparison_value
))
3601 final_value
= loop_find_equiv_value (loop_start
, comparison_value
);
3602 /* If we don't get an invariant final value, we are better
3603 off with the original register. */
3604 if (!invariant_p (final_value
))
3605 final_value
= comparison_value
;
3608 /* Calculate the approximate final value of the induction variable
3609 (on the last successful iteration). The exact final value
3610 depends on the branch operator, and increment sign. It will be
3611 wrong if the iteration variable is not incremented by one each
3612 time through the loop and (comparison_value + off_by_one -
3613 initial_value) % increment != 0.
3614 ??? Note that the final_value may overflow and thus final_larger
3615 will be bogus. A potentially infinite loop will be classified
3616 as immediate, e.g. for (i = 0x7ffffff0; i <= 0x7fffffff; i++) */
3618 final_value
= plus_constant (final_value
, off_by_one
);
3620 /* Save the calculated values describing this loop's bounds, in case
3621 precondition_loop_p will need them later. These values can not be
3622 recalculated inside precondition_loop_p because strength reduction
3623 optimizations may obscure the loop's structure.
3625 These values are only required by precondition_loop_p and insert_bct
3626 whenever the number of iterations cannot be computed at compile time.
3627 Only the difference between final_value and initial_value is
3628 important. Note that final_value is only approximate. */
3629 loop_info
->initial_value
= initial_value
;
3630 loop_info
->comparison_value
= comparison_value
;
3631 loop_info
->final_value
= plus_constant (comparison_value
, off_by_one
);
3632 loop_info
->increment
= increment
;
3633 loop_info
->iteration_var
= iteration_var
;
3634 loop_info
->comparison_code
= comparison_code
;
3636 /* Try to determine the iteration count for loops such
3637 as (for i = init; i < init + const; i++). When running the
3638 loop optimization twice, the first pass often converts simple
3639 loops into this form. */
3641 if (REG_P (initial_value
))
3647 reg1
= initial_value
;
3648 if (GET_CODE (final_value
) == PLUS
)
3649 reg2
= XEXP (final_value
, 0), const2
= XEXP (final_value
, 1);
3651 reg2
= final_value
, const2
= const0_rtx
;
3653 /* Check for initial_value = reg1, final_value = reg2 + const2,
3654 where reg1 != reg2. */
3655 if (REG_P (reg2
) && reg2
!= reg1
)
3659 /* Find what reg1 is equivalent to. Hopefully it will
3660 either be reg2 or reg2 plus a constant. */
3661 temp
= loop_find_equiv_value (loop_start
, reg1
);
3662 if (find_common_reg_term (temp
, reg2
))
3663 initial_value
= temp
;
3666 /* Find what reg2 is equivalent to. Hopefully it will
3667 either be reg1 or reg1 plus a constant. Let's ignore
3668 the latter case for now since it is not so common. */
3669 temp
= loop_find_equiv_value (loop_start
, reg2
);
3670 if (temp
== loop_info
->iteration_var
)
3671 temp
= initial_value
;
3673 final_value
= (const2
== const0_rtx
)
3674 ? reg1
: gen_rtx_PLUS (GET_MODE (reg1
), reg1
, const2
);
3677 else if (loop_info
->vtop
&& GET_CODE (reg2
) == CONST_INT
)
3681 /* When running the loop optimizer twice, check_dbra_loop
3682 further obfuscates reversible loops of the form:
3683 for (i = init; i < init + const; i++). We often end up with
3684 final_value = 0, initial_value = temp, temp = temp2 - init,
3685 where temp2 = init + const. If the loop has a vtop we
3686 can replace initial_value with const. */
3688 temp
= loop_find_equiv_value (loop_start
, reg1
);
3689 if (GET_CODE (temp
) == MINUS
&& REG_P (XEXP (temp
, 0)))
3691 rtx temp2
= loop_find_equiv_value (loop_start
, XEXP (temp
, 0));
3692 if (GET_CODE (temp2
) == PLUS
3693 && XEXP (temp2
, 0) == XEXP (temp
, 1))
3694 initial_value
= XEXP (temp2
, 1);
3699 /* If have initial_value = reg + const1 and final_value = reg +
3700 const2, then replace initial_value with const1 and final_value
3701 with const2. This should be safe since we are protected by the
3702 initial comparison before entering the loop if we have a vtop.
3703 For example, a + b < a + c is not equivalent to b < c for all a
3704 when using modulo arithmetic.
3706 ??? Without a vtop we could still perform the optimization if we check
3707 the initial and final values carefully. */
3709 && (reg_term
= find_common_reg_term (initial_value
, final_value
)))
3711 initial_value
= subtract_reg_term (initial_value
, reg_term
);
3712 final_value
= subtract_reg_term (final_value
, reg_term
);
3715 loop_info
->initial_equiv_value
= initial_value
;
3716 loop_info
->final_equiv_value
= final_value
;
3720 if (loop_dump_stream
)
3721 fprintf (loop_dump_stream
,
3722 "Loop iterations: Increment value can't be calculated.\n");
3726 if (GET_CODE (increment
) != CONST_INT
)
3728 increment
= loop_find_equiv_value (loop_start
, increment
);
3730 if (GET_CODE (increment
) != CONST_INT
)
3732 if (loop_dump_stream
)
3734 fprintf (loop_dump_stream
,
3735 "Loop iterations: Increment value not constant ");
3736 print_rtl (loop_dump_stream
, increment
);
3737 fprintf (loop_dump_stream
, ".\n");
3741 loop_info
->increment
= increment
;
3744 if (GET_CODE (initial_value
) != CONST_INT
)
3746 if (loop_dump_stream
)
3748 fprintf (loop_dump_stream
,
3749 "Loop iterations: Initial value not constant ");
3750 print_rtl (loop_dump_stream
, initial_value
);
3751 fprintf (loop_dump_stream
, ".\n");
3755 else if (comparison_code
== EQ
)
3757 if (loop_dump_stream
)
3758 fprintf (loop_dump_stream
,
3759 "Loop iterations: EQ comparison loop.\n");
3762 else if (GET_CODE (final_value
) != CONST_INT
)
3764 if (loop_dump_stream
)
3766 fprintf (loop_dump_stream
,
3767 "Loop iterations: Final value not constant ");
3768 print_rtl (loop_dump_stream
, final_value
);
3769 fprintf (loop_dump_stream
, ".\n");
3774 /* Final_larger is 1 if final larger, 0 if they are equal, otherwise -1. */
3777 = ((unsigned HOST_WIDE_INT
) INTVAL (final_value
)
3778 > (unsigned HOST_WIDE_INT
) INTVAL (initial_value
))
3779 - ((unsigned HOST_WIDE_INT
) INTVAL (final_value
)
3780 < (unsigned HOST_WIDE_INT
) INTVAL (initial_value
));
3782 final_larger
= (INTVAL (final_value
) > INTVAL (initial_value
))
3783 - (INTVAL (final_value
) < INTVAL (initial_value
));
3785 if (INTVAL (increment
) > 0)
3787 else if (INTVAL (increment
) == 0)
3792 /* There are 27 different cases: compare_dir = -1, 0, 1;
3793 final_larger = -1, 0, 1; increment_dir = -1, 0, 1.
3794 There are 4 normal cases, 4 reverse cases (where the iteration variable
3795 will overflow before the loop exits), 4 infinite loop cases, and 15
3796 immediate exit (0 or 1 iteration depending on loop type) cases.
3797 Only try to optimize the normal cases. */
3799 /* (compare_dir/final_larger/increment_dir)
3800 Normal cases: (0/-1/-1), (0/1/1), (-1/-1/-1), (1/1/1)
3801 Reverse cases: (0/-1/1), (0/1/-1), (-1/-1/1), (1/1/-1)
3802 Infinite loops: (0/-1/0), (0/1/0), (-1/-1/0), (1/1/0)
3803 Immediate exit: (0/0/X), (-1/0/X), (-1/1/X), (1/0/X), (1/-1/X) */
3805 /* ?? If the meaning of reverse loops (where the iteration variable
3806 will overflow before the loop exits) is undefined, then could
3807 eliminate all of these special checks, and just always assume
3808 the loops are normal/immediate/infinite. Note that this means
3809 the sign of increment_dir does not have to be known. Also,
3810 since it does not really hurt if immediate exit loops or infinite loops
3811 are optimized, then that case could be ignored also, and hence all
3812 loops can be optimized.
3814 According to ANSI Spec, the reverse loop case result is undefined,
3815 because the action on overflow is undefined.
3817 See also the special test for NE loops below. */
3819 if (final_larger
== increment_dir
&& final_larger
!= 0
3820 && (final_larger
== compare_dir
|| compare_dir
== 0))
3825 if (loop_dump_stream
)
3826 fprintf (loop_dump_stream
,
3827 "Loop iterations: Not normal loop.\n");
3831 /* Calculate the number of iterations, final_value is only an approximation,
3832 so correct for that. Note that abs_diff and n_iterations are
3833 unsigned, because they can be as large as 2^n - 1. */
3835 abs_inc
= INTVAL (increment
);
3837 abs_diff
= INTVAL (final_value
) - INTVAL (initial_value
);
3838 else if (abs_inc
< 0)
3840 abs_diff
= INTVAL (initial_value
) - INTVAL (final_value
);
3846 /* For NE tests, make sure that the iteration variable won't miss
3847 the final value. If abs_diff mod abs_incr is not zero, then the
3848 iteration variable will overflow before the loop exits, and we
3849 can not calculate the number of iterations. */
3850 if (compare_dir
== 0 && (abs_diff
% abs_inc
) != 0)
3853 /* Note that the number of iterations could be calculated using
3854 (abs_diff + abs_inc - 1) / abs_inc, provided care was taken to
3855 handle potential overflow of the summation. */
3856 loop_info
->n_iterations
= abs_diff
/ abs_inc
+ ((abs_diff
% abs_inc
) != 0);
3857 return loop_info
->n_iterations
;
3861 /* Replace uses of split bivs with their split pseudo register. This is
3862 for original instructions which remain after loop unrolling without
3866 remap_split_bivs (x
)
3869 register enum rtx_code code
;
3876 code
= GET_CODE (x
);
3891 /* If non-reduced/final-value givs were split, then this would also
3892 have to remap those givs also. */
3894 if (REGNO (x
) < max_reg_before_loop
3895 && reg_iv_type
[REGNO (x
)] == BASIC_INDUCT
)
3896 return reg_biv_class
[REGNO (x
)]->biv
->src_reg
;
3903 fmt
= GET_RTX_FORMAT (code
);
3904 for (i
= GET_RTX_LENGTH (code
) - 1; i
>= 0; i
--)
3907 XEXP (x
, i
) = remap_split_bivs (XEXP (x
, i
));
3911 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
3912 XVECEXP (x
, i
, j
) = remap_split_bivs (XVECEXP (x
, i
, j
));
3918 /* If FIRST_UID is a set of REGNO, and FIRST_UID dominates LAST_UID (e.g.
3919 FIST_UID is always executed if LAST_UID is), then return 1. Otherwise
3920 return 0. COPY_START is where we can start looking for the insns
3921 FIRST_UID and LAST_UID. COPY_END is where we stop looking for these
3924 If there is no JUMP_INSN between LOOP_START and FIRST_UID, then FIRST_UID
3925 must dominate LAST_UID.
3927 If there is a CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID
3928 may not dominate LAST_UID.
3930 If there is no CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID
3931 must dominate LAST_UID. */
3934 set_dominates_use (regno
, first_uid
, last_uid
, copy_start
, copy_end
)
3941 int passed_jump
= 0;
3942 rtx p
= NEXT_INSN (copy_start
);
3944 while (INSN_UID (p
) != first_uid
)
3946 if (GET_CODE (p
) == JUMP_INSN
)
3948 /* Could not find FIRST_UID. */
3954 /* Verify that FIRST_UID is an insn that entirely sets REGNO. */
3955 if (GET_RTX_CLASS (GET_CODE (p
)) != 'i'
3956 || ! dead_or_set_regno_p (p
, regno
))
3959 /* FIRST_UID is always executed. */
3960 if (passed_jump
== 0)
3963 while (INSN_UID (p
) != last_uid
)
3965 /* If we see a CODE_LABEL between FIRST_UID and LAST_UID, then we
3966 can not be sure that FIRST_UID dominates LAST_UID. */
3967 if (GET_CODE (p
) == CODE_LABEL
)
3969 /* Could not find LAST_UID, but we reached the end of the loop, so
3971 else if (p
== copy_end
)
3976 /* FIRST_UID is always executed if LAST_UID is executed. */