1 /* Try to unroll loops, and split induction variables.
2 Copyright (C) 1992, 93, 94, 95, 97, 98, 1999 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 this indicates if it was made from a derived giv. */
184 static char *derived_regs
;
186 /* Indexed by register number, if this is a splittable induction variable,
187 then this will hold the number of instructions in the loop that modify
188 the induction variable. Used to ensure that only the last insn modifying
189 a split iv will update the original iv of the dest. */
191 static int *splittable_regs_updates
;
193 /* Forward declarations. */
195 static void init_reg_map
PROTO((struct inline_remap
*, int));
196 static rtx calculate_giv_inc
PROTO((rtx
, rtx
, int));
197 static rtx initial_reg_note_copy
PROTO((rtx
, struct inline_remap
*));
198 static void final_reg_note_copy
PROTO((rtx
, struct inline_remap
*));
199 static void copy_loop_body
PROTO((rtx
, rtx
, struct inline_remap
*, rtx
, int,
200 enum unroll_types
, rtx
, rtx
, rtx
, rtx
));
201 static void iteration_info
PROTO((rtx
, rtx
*, rtx
*, rtx
, rtx
));
202 static int find_splittable_regs
PROTO((enum unroll_types
, rtx
, rtx
, rtx
, int,
203 unsigned HOST_WIDE_INT
));
204 static int find_splittable_givs
PROTO((struct iv_class
*, enum unroll_types
,
205 rtx
, rtx
, rtx
, int));
206 static int reg_dead_after_loop
PROTO((rtx
, rtx
, rtx
));
207 static rtx fold_rtx_mult_add
PROTO((rtx
, rtx
, rtx
, enum machine_mode
));
208 static int verify_addresses
PROTO((struct induction
*, rtx
, int));
209 static rtx remap_split_bivs
PROTO((rtx
));
211 /* Try to unroll one loop and split induction variables in the loop.
213 The loop is described by the arguments LOOP_END, INSN_COUNT, and
214 LOOP_START. END_INSERT_BEFORE indicates where insns should be added
215 which need to be executed when the loop falls through. STRENGTH_REDUCTION_P
216 indicates whether information generated in the strength reduction pass
219 This function is intended to be called from within `strength_reduce'
223 unroll_loop (loop_end
, insn_count
, loop_start
, end_insert_before
,
224 loop_info
, strength_reduce_p
)
228 rtx end_insert_before
;
229 struct loop_info
*loop_info
;
230 int strength_reduce_p
;
233 int unroll_number
= 1;
234 rtx copy_start
, copy_end
;
235 rtx insn
, sequence
, pattern
, tem
;
236 int max_labelno
, max_insnno
;
238 struct inline_remap
*map
;
241 int max_local_regnum
;
246 int splitting_not_safe
= 0;
247 enum unroll_types unroll_type
;
248 int loop_preconditioned
= 0;
250 /* This points to the last real insn in the loop, which should be either
251 a JUMP_INSN (for conditional jumps) or a BARRIER (for unconditional
255 /* Don't bother unrolling huge loops. Since the minimum factor is
256 two, loops greater than one half of MAX_UNROLLED_INSNS will never
258 if (insn_count
> MAX_UNROLLED_INSNS
/ 2)
260 if (loop_dump_stream
)
261 fprintf (loop_dump_stream
, "Unrolling failure: Loop too big.\n");
265 /* When emitting debugger info, we can't unroll loops with unequal numbers
266 of block_beg and block_end notes, because that would unbalance the block
267 structure of the function. This can happen as a result of the
268 "if (foo) bar; else break;" optimization in jump.c. */
269 /* ??? Gcc has a general policy that -g is never supposed to change the code
270 that the compiler emits, so we must disable this optimization always,
271 even if debug info is not being output. This is rare, so this should
272 not be a significant performance problem. */
274 if (1 /* write_symbols != NO_DEBUG */)
276 int block_begins
= 0;
279 for (insn
= loop_start
; insn
!= loop_end
; insn
= NEXT_INSN (insn
))
281 if (GET_CODE (insn
) == NOTE
)
283 if (NOTE_LINE_NUMBER (insn
) == NOTE_INSN_BLOCK_BEG
)
285 else if (NOTE_LINE_NUMBER (insn
) == NOTE_INSN_BLOCK_END
)
290 if (block_begins
!= block_ends
)
292 if (loop_dump_stream
)
293 fprintf (loop_dump_stream
,
294 "Unrolling failure: Unbalanced block notes.\n");
299 /* Determine type of unroll to perform. Depends on the number of iterations
300 and the size of the loop. */
302 /* If there is no strength reduce info, then set
303 loop_info->n_iterations to zero. This can happen if
304 strength_reduce can't find any bivs in the loop. A value of zero
305 indicates that the number of iterations could not be calculated. */
307 if (! strength_reduce_p
)
308 loop_info
->n_iterations
= 0;
310 if (loop_dump_stream
&& loop_info
->n_iterations
> 0)
312 fputs ("Loop unrolling: ", loop_dump_stream
);
313 fprintf (loop_dump_stream
, HOST_WIDE_INT_PRINT_DEC
,
314 loop_info
->n_iterations
);
315 fputs (" iterations.\n", loop_dump_stream
);
318 /* Find and save a pointer to the last nonnote insn in the loop. */
320 last_loop_insn
= prev_nonnote_insn (loop_end
);
322 /* Calculate how many times to unroll the loop. Indicate whether or
323 not the loop is being completely unrolled. */
325 if (loop_info
->n_iterations
== 1)
327 /* If number of iterations is exactly 1, then eliminate the compare and
328 branch at the end of the loop since they will never be taken.
329 Then return, since no other action is needed here. */
331 /* If the last instruction is not a BARRIER or a JUMP_INSN, then
332 don't do anything. */
334 if (GET_CODE (last_loop_insn
) == BARRIER
)
336 /* Delete the jump insn. This will delete the barrier also. */
337 delete_insn (PREV_INSN (last_loop_insn
));
339 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
342 /* The immediately preceding insn is a compare which must be
344 delete_insn (last_loop_insn
);
345 delete_insn (PREV_INSN (last_loop_insn
));
347 /* The immediately preceding insn may not be the compare, so don't
349 delete_insn (last_loop_insn
);
354 else if (loop_info
->n_iterations
> 0
355 && loop_info
->n_iterations
* insn_count
< MAX_UNROLLED_INSNS
)
357 unroll_number
= loop_info
->n_iterations
;
358 unroll_type
= UNROLL_COMPLETELY
;
360 else if (loop_info
->n_iterations
> 0)
362 /* Try to factor the number of iterations. Don't bother with the
363 general case, only using 2, 3, 5, and 7 will get 75% of all
364 numbers theoretically, and almost all in practice. */
366 for (i
= 0; i
< NUM_FACTORS
; i
++)
367 factors
[i
].count
= 0;
369 temp
= loop_info
->n_iterations
;
370 for (i
= NUM_FACTORS
- 1; i
>= 0; i
--)
371 while (temp
% factors
[i
].factor
== 0)
374 temp
= temp
/ factors
[i
].factor
;
377 /* Start with the larger factors first so that we generally
378 get lots of unrolling. */
382 for (i
= 3; i
>= 0; i
--)
383 while (factors
[i
].count
--)
385 if (temp
* factors
[i
].factor
< MAX_UNROLLED_INSNS
)
387 unroll_number
*= factors
[i
].factor
;
388 temp
*= factors
[i
].factor
;
394 /* If we couldn't find any factors, then unroll as in the normal
396 if (unroll_number
== 1)
398 if (loop_dump_stream
)
399 fprintf (loop_dump_stream
,
400 "Loop unrolling: No factors found.\n");
403 unroll_type
= UNROLL_MODULO
;
407 /* Default case, calculate number of times to unroll loop based on its
409 if (unroll_number
== 1)
411 if (8 * insn_count
< MAX_UNROLLED_INSNS
)
413 else if (4 * insn_count
< MAX_UNROLLED_INSNS
)
418 unroll_type
= UNROLL_NAIVE
;
421 /* Now we know how many times to unroll the loop. */
423 if (loop_dump_stream
)
424 fprintf (loop_dump_stream
,
425 "Unrolling loop %d times.\n", unroll_number
);
428 if (unroll_type
== UNROLL_COMPLETELY
|| unroll_type
== UNROLL_MODULO
)
430 /* Loops of these types can start with jump down to the exit condition
431 in rare circumstances.
433 Consider a pair of nested loops where the inner loop is part
434 of the exit code for the outer loop.
436 In this case jump.c will not duplicate the exit test for the outer
437 loop, so it will start with a jump to the exit code.
439 Then consider if the inner loop turns out to iterate once and
440 only once. We will end up deleting the jumps associated with
441 the inner loop. However, the loop notes are not removed from
442 the instruction stream.
444 And finally assume that we can compute the number of iterations
447 In this case unroll may want to unroll the outer loop even though
448 it starts with a jump to the outer loop's exit code.
450 We could try to optimize this case, but it hardly seems worth it.
451 Just return without unrolling the loop in such cases. */
454 while (GET_CODE (insn
) != CODE_LABEL
&& GET_CODE (insn
) != JUMP_INSN
)
455 insn
= NEXT_INSN (insn
);
456 if (GET_CODE (insn
) == JUMP_INSN
)
460 if (unroll_type
== UNROLL_COMPLETELY
)
462 /* Completely unrolling the loop: Delete the compare and branch at
463 the end (the last two instructions). This delete must done at the
464 very end of loop unrolling, to avoid problems with calls to
465 back_branch_in_range_p, which is called by find_splittable_regs.
466 All increments of splittable bivs/givs are changed to load constant
469 copy_start
= loop_start
;
471 /* Set insert_before to the instruction immediately after the JUMP_INSN
472 (or BARRIER), so that any NOTEs between the JUMP_INSN and the end of
473 the loop will be correctly handled by copy_loop_body. */
474 insert_before
= NEXT_INSN (last_loop_insn
);
476 /* Set copy_end to the insn before the jump at the end of the loop. */
477 if (GET_CODE (last_loop_insn
) == BARRIER
)
478 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
479 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
482 /* The instruction immediately before the JUMP_INSN is a compare
483 instruction which we do not want to copy. */
484 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
486 /* The instruction immediately before the JUMP_INSN may not be the
487 compare, so we must copy it. */
488 copy_end
= PREV_INSN (last_loop_insn
);
493 /* We currently can't unroll a loop if it doesn't end with a
494 JUMP_INSN. There would need to be a mechanism that recognizes
495 this case, and then inserts a jump after each loop body, which
496 jumps to after the last loop body. */
497 if (loop_dump_stream
)
498 fprintf (loop_dump_stream
,
499 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
503 else if (unroll_type
== UNROLL_MODULO
)
505 /* Partially unrolling the loop: The compare and branch at the end
506 (the last two instructions) must remain. Don't copy the compare
507 and branch instructions at the end of the loop. Insert the unrolled
508 code immediately before the compare/branch at the end so that the
509 code will fall through to them as before. */
511 copy_start
= loop_start
;
513 /* Set insert_before to the jump insn at the end of the loop.
514 Set copy_end to before the jump insn at the end of the loop. */
515 if (GET_CODE (last_loop_insn
) == BARRIER
)
517 insert_before
= PREV_INSN (last_loop_insn
);
518 copy_end
= PREV_INSN (insert_before
);
520 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
523 /* The instruction immediately before the JUMP_INSN is a compare
524 instruction which we do not want to copy or delete. */
525 insert_before
= PREV_INSN (last_loop_insn
);
526 copy_end
= PREV_INSN (insert_before
);
528 /* The instruction immediately before the JUMP_INSN may not be the
529 compare, so we must copy it. */
530 insert_before
= last_loop_insn
;
531 copy_end
= PREV_INSN (last_loop_insn
);
536 /* We currently can't unroll a loop if it doesn't end with a
537 JUMP_INSN. There would need to be a mechanism that recognizes
538 this case, and then inserts a jump after each loop body, which
539 jumps to after the last loop body. */
540 if (loop_dump_stream
)
541 fprintf (loop_dump_stream
,
542 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
548 /* Normal case: Must copy the compare and branch instructions at the
551 if (GET_CODE (last_loop_insn
) == BARRIER
)
553 /* Loop ends with an unconditional jump and a barrier.
554 Handle this like above, don't copy jump and barrier.
555 This is not strictly necessary, but doing so prevents generating
556 unconditional jumps to an immediately following label.
558 This will be corrected below if the target of this jump is
559 not the start_label. */
561 insert_before
= PREV_INSN (last_loop_insn
);
562 copy_end
= PREV_INSN (insert_before
);
564 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
566 /* Set insert_before to immediately after the JUMP_INSN, so that
567 NOTEs at the end of the loop will be correctly handled by
569 insert_before
= NEXT_INSN (last_loop_insn
);
570 copy_end
= last_loop_insn
;
574 /* We currently can't unroll a loop if it doesn't end with a
575 JUMP_INSN. There would need to be a mechanism that recognizes
576 this case, and then inserts a jump after each loop body, which
577 jumps to after the last loop body. */
578 if (loop_dump_stream
)
579 fprintf (loop_dump_stream
,
580 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
584 /* If copying exit test branches because they can not be eliminated,
585 then must convert the fall through case of the branch to a jump past
586 the end of the loop. Create a label to emit after the loop and save
587 it for later use. Do not use the label after the loop, if any, since
588 it might be used by insns outside the loop, or there might be insns
589 added before it later by final_[bg]iv_value which must be after
590 the real exit label. */
591 exit_label
= gen_label_rtx ();
594 while (GET_CODE (insn
) != CODE_LABEL
&& GET_CODE (insn
) != JUMP_INSN
)
595 insn
= NEXT_INSN (insn
);
597 if (GET_CODE (insn
) == JUMP_INSN
)
599 /* The loop starts with a jump down to the exit condition test.
600 Start copying the loop after the barrier following this
602 copy_start
= NEXT_INSN (insn
);
604 /* Splitting induction variables doesn't work when the loop is
605 entered via a jump to the bottom, because then we end up doing
606 a comparison against a new register for a split variable, but
607 we did not execute the set insn for the new register because
608 it was skipped over. */
609 splitting_not_safe
= 1;
610 if (loop_dump_stream
)
611 fprintf (loop_dump_stream
,
612 "Splitting not safe, because loop not entered at top.\n");
615 copy_start
= loop_start
;
618 /* This should always be the first label in the loop. */
619 start_label
= NEXT_INSN (copy_start
);
620 /* There may be a line number note and/or a loop continue note here. */
621 while (GET_CODE (start_label
) == NOTE
)
622 start_label
= NEXT_INSN (start_label
);
623 if (GET_CODE (start_label
) != CODE_LABEL
)
625 /* This can happen as a result of jump threading. If the first insns in
626 the loop test the same condition as the loop's backward jump, or the
627 opposite condition, then the backward jump will be modified to point
628 to elsewhere, and the loop's start label is deleted.
630 This case currently can not be handled by the loop unrolling code. */
632 if (loop_dump_stream
)
633 fprintf (loop_dump_stream
,
634 "Unrolling failure: unknown insns between BEG note and loop label.\n");
637 if (LABEL_NAME (start_label
))
639 /* The jump optimization pass must have combined the original start label
640 with a named label for a goto. We can't unroll this case because
641 jumps which go to the named label must be handled differently than
642 jumps to the loop start, and it is impossible to differentiate them
644 if (loop_dump_stream
)
645 fprintf (loop_dump_stream
,
646 "Unrolling failure: loop start label is gone\n");
650 if (unroll_type
== UNROLL_NAIVE
651 && GET_CODE (last_loop_insn
) == BARRIER
652 && start_label
!= JUMP_LABEL (PREV_INSN (last_loop_insn
)))
654 /* In this case, we must copy the jump and barrier, because they will
655 not be converted to jumps to an immediately following label. */
657 insert_before
= NEXT_INSN (last_loop_insn
);
658 copy_end
= last_loop_insn
;
661 if (unroll_type
== UNROLL_NAIVE
662 && GET_CODE (last_loop_insn
) == JUMP_INSN
663 && start_label
!= JUMP_LABEL (last_loop_insn
))
665 /* ??? The loop ends with a conditional branch that does not branch back
666 to the loop start label. In this case, we must emit an unconditional
667 branch to the loop exit after emitting the final branch.
668 copy_loop_body does not have support for this currently, so we
669 give up. It doesn't seem worthwhile to unroll anyways since
670 unrolling would increase the number of branch instructions
672 if (loop_dump_stream
)
673 fprintf (loop_dump_stream
,
674 "Unrolling failure: final conditional branch not to loop start\n");
678 /* Allocate a translation table for the labels and insn numbers.
679 They will be filled in as we copy the insns in the loop. */
681 max_labelno
= max_label_num ();
682 max_insnno
= get_max_uid ();
684 map
= (struct inline_remap
*) alloca (sizeof (struct inline_remap
));
686 map
->integrating
= 0;
687 map
->const_equiv_varray
= 0;
689 /* Allocate the label map. */
693 map
->label_map
= (rtx
*) alloca (max_labelno
* sizeof (rtx
));
695 local_label
= (char *) alloca (max_labelno
);
696 bzero (local_label
, max_labelno
);
701 /* Search the loop and mark all local labels, i.e. the ones which have to
702 be distinct labels when copied. For all labels which might be
703 non-local, set their label_map entries to point to themselves.
704 If they happen to be local their label_map entries will be overwritten
705 before the loop body is copied. The label_map entries for local labels
706 will be set to a different value each time the loop body is copied. */
708 for (insn
= copy_start
; insn
!= loop_end
; insn
= NEXT_INSN (insn
))
712 if (GET_CODE (insn
) == CODE_LABEL
)
713 local_label
[CODE_LABEL_NUMBER (insn
)] = 1;
714 else if (GET_CODE (insn
) == JUMP_INSN
)
716 if (JUMP_LABEL (insn
))
717 set_label_in_map (map
,
718 CODE_LABEL_NUMBER (JUMP_LABEL (insn
)),
720 else if (GET_CODE (PATTERN (insn
)) == ADDR_VEC
721 || GET_CODE (PATTERN (insn
)) == ADDR_DIFF_VEC
)
723 rtx pat
= PATTERN (insn
);
724 int diff_vec_p
= GET_CODE (PATTERN (insn
)) == ADDR_DIFF_VEC
;
725 int len
= XVECLEN (pat
, diff_vec_p
);
728 for (i
= 0; i
< len
; i
++)
730 label
= XEXP (XVECEXP (pat
, diff_vec_p
, i
), 0);
731 set_label_in_map (map
,
732 CODE_LABEL_NUMBER (label
),
737 else if ((note
= find_reg_note (insn
, REG_LABEL
, NULL_RTX
)))
738 set_label_in_map (map
, CODE_LABEL_NUMBER (XEXP (note
, 0)),
742 /* Allocate space for the insn map. */
744 map
->insn_map
= (rtx
*) alloca (max_insnno
* sizeof (rtx
));
746 /* Set this to zero, to indicate that we are doing loop unrolling,
747 not function inlining. */
748 map
->inline_target
= 0;
750 /* The register and constant maps depend on the number of registers
751 present, so the final maps can't be created until after
752 find_splittable_regs is called. However, they are needed for
753 preconditioning, so we create temporary maps when preconditioning
756 /* The preconditioning code may allocate two new pseudo registers. */
757 maxregnum
= max_reg_num ();
759 /* local_regno is only valid for regnos < max_local_regnum. */
760 max_local_regnum
= maxregnum
;
762 /* Allocate and zero out the splittable_regs and addr_combined_regs
763 arrays. These must be zeroed here because they will be used if
764 loop preconditioning is performed, and must be zero for that case.
766 It is safe to do this here, since the extra registers created by the
767 preconditioning code and find_splittable_regs will never be used
768 to access the splittable_regs[] and addr_combined_regs[] arrays. */
770 splittable_regs
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
771 bzero ((char *) splittable_regs
, maxregnum
* sizeof (rtx
));
772 derived_regs
= alloca (maxregnum
);
773 bzero (derived_regs
, maxregnum
);
774 splittable_regs_updates
= (int *) alloca (maxregnum
* sizeof (int));
775 bzero ((char *) splittable_regs_updates
, maxregnum
* sizeof (int));
777 = (struct induction
**) alloca (maxregnum
* sizeof (struct induction
*));
778 bzero ((char *) addr_combined_regs
, maxregnum
* sizeof (struct induction
*));
779 local_regno
= (char *) alloca (maxregnum
);
780 bzero (local_regno
, maxregnum
);
782 /* Mark all local registers, i.e. the ones which are referenced only
784 if (INSN_UID (copy_end
) < max_uid_for_loop
)
786 int copy_start_luid
= INSN_LUID (copy_start
);
787 int copy_end_luid
= INSN_LUID (copy_end
);
789 /* If a register is used in the jump insn, we must not duplicate it
790 since it will also be used outside the loop. */
791 if (GET_CODE (copy_end
) == JUMP_INSN
)
794 /* If we have a target that uses cc0, then we also must not duplicate
795 the insn that sets cc0 before the jump insn. */
797 if (GET_CODE (copy_end
) == JUMP_INSN
)
801 /* If copy_start points to the NOTE that starts the loop, then we must
802 use the next luid, because invariant pseudo-regs moved out of the loop
803 have their lifetimes modified to start here, but they are not safe
805 if (copy_start
== loop_start
)
808 /* If a pseudo's lifetime is entirely contained within this loop, then we
809 can use a different pseudo in each unrolled copy of the loop. This
810 results in better code. */
811 /* We must limit the generic test to max_reg_before_loop, because only
812 these pseudo registers have valid regno_first_uid info. */
813 for (j
= FIRST_PSEUDO_REGISTER
; j
< max_reg_before_loop
; ++j
)
814 if (REGNO_FIRST_UID (j
) > 0 && REGNO_FIRST_UID (j
) <= max_uid_for_loop
815 && uid_luid
[REGNO_FIRST_UID (j
)] >= copy_start_luid
816 && REGNO_LAST_UID (j
) > 0 && REGNO_LAST_UID (j
) <= max_uid_for_loop
817 && uid_luid
[REGNO_LAST_UID (j
)] <= copy_end_luid
)
819 /* However, we must also check for loop-carried dependencies.
820 If the value the pseudo has at the end of iteration X is
821 used by iteration X+1, then we can not use a different pseudo
822 for each unrolled copy of the loop. */
823 /* A pseudo is safe if regno_first_uid is a set, and this
824 set dominates all instructions from regno_first_uid to
826 /* ??? This check is simplistic. We would get better code if
827 this check was more sophisticated. */
828 if (set_dominates_use (j
, REGNO_FIRST_UID (j
), REGNO_LAST_UID (j
),
829 copy_start
, copy_end
))
832 if (loop_dump_stream
)
835 fprintf (loop_dump_stream
, "Marked reg %d as local\n", j
);
837 fprintf (loop_dump_stream
, "Did not mark reg %d as local\n",
841 /* Givs that have been created from multiple biv increments always have
843 for (j
= first_increment_giv
; j
<= last_increment_giv
; j
++)
846 if (loop_dump_stream
)
847 fprintf (loop_dump_stream
, "Marked reg %d as local\n", j
);
851 /* If this loop requires exit tests when unrolled, check to see if we
852 can precondition the loop so as to make the exit tests unnecessary.
853 Just like variable splitting, this is not safe if the loop is entered
854 via a jump to the bottom. Also, can not do this if no strength
855 reduce info, because precondition_loop_p uses this info. */
857 /* Must copy the loop body for preconditioning before the following
858 find_splittable_regs call since that will emit insns which need to
859 be after the preconditioned loop copies, but immediately before the
860 unrolled loop copies. */
862 /* Also, it is not safe to split induction variables for the preconditioned
863 copies of the loop body. If we split induction variables, then the code
864 assumes that each induction variable can be represented as a function
865 of its initial value and the loop iteration number. This is not true
866 in this case, because the last preconditioned copy of the loop body
867 could be any iteration from the first up to the `unroll_number-1'th,
868 depending on the initial value of the iteration variable. Therefore
869 we can not split induction variables here, because we can not calculate
870 their value. Hence, this code must occur before find_splittable_regs
873 if (unroll_type
== UNROLL_NAIVE
&& ! splitting_not_safe
&& strength_reduce_p
)
875 rtx initial_value
, final_value
, increment
;
876 enum machine_mode mode
;
878 if (precondition_loop_p (loop_start
, loop_info
,
879 &initial_value
, &final_value
, &increment
,
884 int abs_inc
, neg_inc
;
886 map
->reg_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
888 VARRAY_CONST_EQUIV_INIT (map
->const_equiv_varray
, maxregnum
,
890 global_const_equiv_varray
= map
->const_equiv_varray
;
892 init_reg_map (map
, maxregnum
);
894 /* Limit loop unrolling to 4, since this will make 7 copies of
896 if (unroll_number
> 4)
899 /* Save the absolute value of the increment, and also whether or
900 not it is negative. */
902 abs_inc
= INTVAL (increment
);
911 /* Calculate the difference between the final and initial values.
912 Final value may be a (plus (reg x) (const_int 1)) rtx.
913 Let the following cse pass simplify this if initial value is
916 We must copy the final and initial values here to avoid
917 improperly shared rtl. */
919 diff
= expand_binop (mode
, sub_optab
, copy_rtx (final_value
),
920 copy_rtx (initial_value
), NULL_RTX
, 0,
923 /* Now calculate (diff % (unroll * abs (increment))) by using an
925 diff
= expand_binop (GET_MODE (diff
), and_optab
, diff
,
926 GEN_INT (unroll_number
* abs_inc
- 1),
927 NULL_RTX
, 0, OPTAB_LIB_WIDEN
);
929 /* Now emit a sequence of branches to jump to the proper precond
932 labels
= (rtx
*) alloca (sizeof (rtx
) * unroll_number
);
933 for (i
= 0; i
< unroll_number
; i
++)
934 labels
[i
] = gen_label_rtx ();
936 /* Check for the case where the initial value is greater than or
937 equal to the final value. In that case, we want to execute
938 exactly one loop iteration. The code below will fail for this
939 case. This check does not apply if the loop has a NE
940 comparison at the end. */
942 if (loop_info
->comparison_code
!= NE
)
944 emit_cmp_and_jump_insns (initial_value
, final_value
,
946 NULL_RTX
, mode
, 0, 0, labels
[1]);
947 JUMP_LABEL (get_last_insn ()) = labels
[1];
948 LABEL_NUSES (labels
[1])++;
951 /* Assuming the unroll_number is 4, and the increment is 2, then
952 for a negative increment: for a positive increment:
953 diff = 0,1 precond 0 diff = 0,7 precond 0
954 diff = 2,3 precond 3 diff = 1,2 precond 1
955 diff = 4,5 precond 2 diff = 3,4 precond 2
956 diff = 6,7 precond 1 diff = 5,6 precond 3 */
958 /* We only need to emit (unroll_number - 1) branches here, the
959 last case just falls through to the following code. */
961 /* ??? This would give better code if we emitted a tree of branches
962 instead of the current linear list of branches. */
964 for (i
= 0; i
< unroll_number
- 1; i
++)
967 enum rtx_code cmp_code
;
969 /* For negative increments, must invert the constant compared
970 against, except when comparing against zero. */
978 cmp_const
= unroll_number
- i
;
987 emit_cmp_and_jump_insns (diff
, GEN_INT (abs_inc
* cmp_const
),
988 cmp_code
, NULL_RTX
, mode
, 0, 0,
990 JUMP_LABEL (get_last_insn ()) = labels
[i
];
991 LABEL_NUSES (labels
[i
])++;
994 /* If the increment is greater than one, then we need another branch,
995 to handle other cases equivalent to 0. */
997 /* ??? This should be merged into the code above somehow to help
998 simplify the code here, and reduce the number of branches emitted.
999 For the negative increment case, the branch here could easily
1000 be merged with the `0' case branch above. For the positive
1001 increment case, it is not clear how this can be simplified. */
1006 enum rtx_code cmp_code
;
1010 cmp_const
= abs_inc
- 1;
1015 cmp_const
= abs_inc
* (unroll_number
- 1) + 1;
1019 emit_cmp_and_jump_insns (diff
, GEN_INT (cmp_const
), cmp_code
,
1020 NULL_RTX
, mode
, 0, 0, labels
[0]);
1021 JUMP_LABEL (get_last_insn ()) = labels
[0];
1022 LABEL_NUSES (labels
[0])++;
1025 sequence
= gen_sequence ();
1027 emit_insn_before (sequence
, loop_start
);
1029 /* Only the last copy of the loop body here needs the exit
1030 test, so set copy_end to exclude the compare/branch here,
1031 and then reset it inside the loop when get to the last
1034 if (GET_CODE (last_loop_insn
) == BARRIER
)
1035 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
1036 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
1039 /* The immediately preceding insn is a compare which we do not
1041 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
1043 /* The immediately preceding insn may not be a compare, so we
1045 copy_end
= PREV_INSN (last_loop_insn
);
1051 for (i
= 1; i
< unroll_number
; i
++)
1053 emit_label_after (labels
[unroll_number
- i
],
1054 PREV_INSN (loop_start
));
1056 bzero ((char *) map
->insn_map
, max_insnno
* sizeof (rtx
));
1057 bzero ((char *) &VARRAY_CONST_EQUIV (map
->const_equiv_varray
, 0),
1058 (VARRAY_SIZE (map
->const_equiv_varray
)
1059 * sizeof (struct const_equiv_data
)));
1062 for (j
= 0; j
< max_labelno
; j
++)
1064 set_label_in_map (map
, j
, gen_label_rtx ());
1066 for (j
= FIRST_PSEUDO_REGISTER
; j
< max_local_regnum
; j
++)
1069 map
->reg_map
[j
] = gen_reg_rtx (GET_MODE (regno_reg_rtx
[j
]));
1070 record_base_value (REGNO (map
->reg_map
[j
]),
1071 regno_reg_rtx
[j
], 0);
1073 /* The last copy needs the compare/branch insns at the end,
1074 so reset copy_end here if the loop ends with a conditional
1077 if (i
== unroll_number
- 1)
1079 if (GET_CODE (last_loop_insn
) == BARRIER
)
1080 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
1082 copy_end
= last_loop_insn
;
1085 /* None of the copies are the `last_iteration', so just
1086 pass zero for that parameter. */
1087 copy_loop_body (copy_start
, copy_end
, map
, exit_label
, 0,
1088 unroll_type
, start_label
, loop_end
,
1089 loop_start
, copy_end
);
1091 emit_label_after (labels
[0], PREV_INSN (loop_start
));
1093 if (GET_CODE (last_loop_insn
) == BARRIER
)
1095 insert_before
= PREV_INSN (last_loop_insn
);
1096 copy_end
= PREV_INSN (insert_before
);
1101 /* The immediately preceding insn is a compare which we do not
1103 insert_before
= PREV_INSN (last_loop_insn
);
1104 copy_end
= PREV_INSN (insert_before
);
1106 /* The immediately preceding insn may not be a compare, so we
1108 insert_before
= last_loop_insn
;
1109 copy_end
= PREV_INSN (last_loop_insn
);
1113 /* Set unroll type to MODULO now. */
1114 unroll_type
= UNROLL_MODULO
;
1115 loop_preconditioned
= 1;
1119 /* If reach here, and the loop type is UNROLL_NAIVE, then don't unroll
1120 the loop unless all loops are being unrolled. */
1121 if (unroll_type
== UNROLL_NAIVE
&& ! flag_unroll_all_loops
)
1123 if (loop_dump_stream
)
1124 fprintf (loop_dump_stream
, "Unrolling failure: Naive unrolling not being done.\n");
1128 /* At this point, we are guaranteed to unroll the loop. */
1130 /* Keep track of the unroll factor for the loop. */
1131 if (unroll_type
== UNROLL_COMPLETELY
)
1132 loop_info
->unroll_number
= -1;
1134 loop_info
->unroll_number
= unroll_number
;
1137 /* For each biv and giv, determine whether it can be safely split into
1138 a different variable for each unrolled copy of the loop body.
1139 We precalculate and save this info here, since computing it is
1142 Do this before deleting any instructions from the loop, so that
1143 back_branch_in_range_p will work correctly. */
1145 if (splitting_not_safe
)
1148 temp
= find_splittable_regs (unroll_type
, loop_start
, loop_end
,
1149 end_insert_before
, unroll_number
,
1150 loop_info
->n_iterations
);
1152 /* find_splittable_regs may have created some new registers, so must
1153 reallocate the reg_map with the new larger size, and must realloc
1154 the constant maps also. */
1156 maxregnum
= max_reg_num ();
1157 map
->reg_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
1159 init_reg_map (map
, maxregnum
);
1161 if (map
->const_equiv_varray
== 0)
1162 VARRAY_CONST_EQUIV_INIT (map
->const_equiv_varray
,
1163 maxregnum
+ temp
* unroll_number
* 2,
1165 global_const_equiv_varray
= map
->const_equiv_varray
;
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 *) &VARRAY_CONST_EQUIV (map
->const_equiv_varray
, 0),
1207 VARRAY_SIZE (map
->const_equiv_varray
) * sizeof (struct const_equiv_data
));
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_local_regnum
; 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 /* ??? We can't delete a NOTE_INSN_DELETED_LABEL unless we fix the
1266 DECL_RTL field of the associated LABEL_DECL to point to (one of)
1267 the new copies of the label. Otherwise, we hit an abort in
1268 dwarfout.c/dwarf2out.c. */
1269 if (insn
!= start_label
1270 && ! (GET_CODE (insn
) == NOTE
1271 && NOTE_LINE_NUMBER (insn
) == NOTE_INSN_DELETED_LABEL
))
1272 insn
= delete_insn (insn
);
1274 insn
= NEXT_INSN (insn
);
1277 /* Can now delete the 'safety' label emitted to protect us from runaway
1278 delete_insn calls. */
1279 if (INSN_DELETED_P (safety_label
))
1281 delete_insn (safety_label
);
1283 /* If exit_label exists, emit it after the loop. Doing the emit here
1284 forces it to have a higher INSN_UID than any insn in the unrolled loop.
1285 This is needed so that mostly_true_jump in reorg.c will treat jumps
1286 to this loop end label correctly, i.e. predict that they are usually
1289 emit_label_after (exit_label
, loop_end
);
1292 if (map
&& map
->const_equiv_varray
)
1293 VARRAY_FREE (map
->const_equiv_varray
);
1296 /* Return true if the loop can be safely, and profitably, preconditioned
1297 so that the unrolled copies of the loop body don't need exit tests.
1299 This only works if final_value, initial_value and increment can be
1300 determined, and if increment is a constant power of 2.
1301 If increment is not a power of 2, then the preconditioning modulo
1302 operation would require a real modulo instead of a boolean AND, and this
1303 is not considered `profitable'. */
1305 /* ??? If the loop is known to be executed very many times, or the machine
1306 has a very cheap divide instruction, then preconditioning is a win even
1307 when the increment is not a power of 2. Use RTX_COST to compute
1308 whether divide is cheap.
1309 ??? A divide by constant doesn't actually need a divide, look at
1310 expand_divmod. The reduced cost of this optimized modulo is not
1311 reflected in RTX_COST. */
1314 precondition_loop_p (loop_start
, loop_info
,
1315 initial_value
, final_value
, increment
, mode
)
1317 struct loop_info
*loop_info
;
1318 rtx
*initial_value
, *final_value
, *increment
;
1319 enum machine_mode
*mode
;
1322 if (loop_info
->n_iterations
> 0)
1324 *initial_value
= const0_rtx
;
1325 *increment
= const1_rtx
;
1326 *final_value
= GEN_INT (loop_info
->n_iterations
);
1329 if (loop_dump_stream
)
1331 fputs ("Preconditioning: Success, number of iterations known, ",
1333 fprintf (loop_dump_stream
, HOST_WIDE_INT_PRINT_DEC
,
1334 loop_info
->n_iterations
);
1335 fputs (".\n", loop_dump_stream
);
1340 if (loop_info
->initial_value
== 0)
1342 if (loop_dump_stream
)
1343 fprintf (loop_dump_stream
,
1344 "Preconditioning: Could not find initial value.\n");
1347 else if (loop_info
->increment
== 0)
1349 if (loop_dump_stream
)
1350 fprintf (loop_dump_stream
,
1351 "Preconditioning: Could not find increment value.\n");
1354 else if (GET_CODE (loop_info
->increment
) != CONST_INT
)
1356 if (loop_dump_stream
)
1357 fprintf (loop_dump_stream
,
1358 "Preconditioning: Increment not a constant.\n");
1361 else if ((exact_log2 (INTVAL (loop_info
->increment
)) < 0)
1362 && (exact_log2 (- INTVAL (loop_info
->increment
)) < 0))
1364 if (loop_dump_stream
)
1365 fprintf (loop_dump_stream
,
1366 "Preconditioning: Increment not a constant power of 2.\n");
1370 /* Unsigned_compare and compare_dir can be ignored here, since they do
1371 not matter for preconditioning. */
1373 if (loop_info
->final_value
== 0)
1375 if (loop_dump_stream
)
1376 fprintf (loop_dump_stream
,
1377 "Preconditioning: EQ comparison loop.\n");
1381 /* Must ensure that final_value is invariant, so call invariant_p to
1382 check. Before doing so, must check regno against max_reg_before_loop
1383 to make sure that the register is in the range covered by invariant_p.
1384 If it isn't, then it is most likely a biv/giv which by definition are
1386 if ((GET_CODE (loop_info
->final_value
) == REG
1387 && REGNO (loop_info
->final_value
) >= max_reg_before_loop
)
1388 || (GET_CODE (loop_info
->final_value
) == PLUS
1389 && REGNO (XEXP (loop_info
->final_value
, 0)) >= max_reg_before_loop
)
1390 || ! invariant_p (loop_info
->final_value
))
1392 if (loop_dump_stream
)
1393 fprintf (loop_dump_stream
,
1394 "Preconditioning: Final value not invariant.\n");
1398 /* Fail for floating point values, since the caller of this function
1399 does not have code to deal with them. */
1400 if (GET_MODE_CLASS (GET_MODE (loop_info
->final_value
)) == MODE_FLOAT
1401 || GET_MODE_CLASS (GET_MODE (loop_info
->initial_value
)) == MODE_FLOAT
)
1403 if (loop_dump_stream
)
1404 fprintf (loop_dump_stream
,
1405 "Preconditioning: Floating point final or initial value.\n");
1409 /* Fail if loop_info->iteration_var is not live before loop_start,
1410 since we need to test its value in the preconditioning code. */
1412 if (uid_luid
[REGNO_FIRST_UID (REGNO (loop_info
->iteration_var
))]
1413 > INSN_LUID (loop_start
))
1415 if (loop_dump_stream
)
1416 fprintf (loop_dump_stream
,
1417 "Preconditioning: Iteration var not live before loop start.\n");
1421 /* Note that iteration_info biases the initial value for GIV iterators
1422 such as "while (i-- > 0)" so that we can calculate the number of
1423 iterations just like for BIV iterators.
1425 Also note that the absolute values of initial_value and
1426 final_value are unimportant as only their difference is used for
1427 calculating the number of loop iterations. */
1428 *initial_value
= loop_info
->initial_value
;
1429 *increment
= loop_info
->increment
;
1430 *final_value
= loop_info
->final_value
;
1432 /* Decide what mode to do these calculations in. Choose the larger
1433 of final_value's mode and initial_value's mode, or a full-word if
1434 both are constants. */
1435 *mode
= GET_MODE (*final_value
);
1436 if (*mode
== VOIDmode
)
1438 *mode
= GET_MODE (*initial_value
);
1439 if (*mode
== VOIDmode
)
1442 else if (*mode
!= GET_MODE (*initial_value
)
1443 && (GET_MODE_SIZE (*mode
)
1444 < GET_MODE_SIZE (GET_MODE (*initial_value
))))
1445 *mode
= GET_MODE (*initial_value
);
1448 if (loop_dump_stream
)
1449 fprintf (loop_dump_stream
, "Preconditioning: Successful.\n");
1454 /* All pseudo-registers must be mapped to themselves. Two hard registers
1455 must be mapped, VIRTUAL_STACK_VARS_REGNUM and VIRTUAL_INCOMING_ARGS_
1456 REGNUM, to avoid function-inlining specific conversions of these
1457 registers. All other hard regs can not be mapped because they may be
1462 init_reg_map (map
, maxregnum
)
1463 struct inline_remap
*map
;
1468 for (i
= maxregnum
- 1; i
> LAST_VIRTUAL_REGISTER
; i
--)
1469 map
->reg_map
[i
] = regno_reg_rtx
[i
];
1470 /* Just clear the rest of the entries. */
1471 for (i
= LAST_VIRTUAL_REGISTER
; i
>= 0; i
--)
1472 map
->reg_map
[i
] = 0;
1474 map
->reg_map
[VIRTUAL_STACK_VARS_REGNUM
]
1475 = regno_reg_rtx
[VIRTUAL_STACK_VARS_REGNUM
];
1476 map
->reg_map
[VIRTUAL_INCOMING_ARGS_REGNUM
]
1477 = regno_reg_rtx
[VIRTUAL_INCOMING_ARGS_REGNUM
];
1480 /* Strength-reduction will often emit code for optimized biv/givs which
1481 calculates their value in a temporary register, and then copies the result
1482 to the iv. This procedure reconstructs the pattern computing the iv;
1483 verifying that all operands are of the proper form.
1485 PATTERN must be the result of single_set.
1486 The return value is the amount that the giv is incremented by. */
1489 calculate_giv_inc (pattern
, src_insn
, regno
)
1490 rtx pattern
, src_insn
;
1494 rtx increment_total
= 0;
1498 /* Verify that we have an increment insn here. First check for a plus
1499 as the set source. */
1500 if (GET_CODE (SET_SRC (pattern
)) != PLUS
)
1502 /* SR sometimes computes the new giv value in a temp, then copies it
1504 src_insn
= PREV_INSN (src_insn
);
1505 pattern
= PATTERN (src_insn
);
1506 if (GET_CODE (SET_SRC (pattern
)) != PLUS
)
1509 /* The last insn emitted is not needed, so delete it to avoid confusing
1510 the second cse pass. This insn sets the giv unnecessarily. */
1511 delete_insn (get_last_insn ());
1514 /* Verify that we have a constant as the second operand of the plus. */
1515 increment
= XEXP (SET_SRC (pattern
), 1);
1516 if (GET_CODE (increment
) != CONST_INT
)
1518 /* SR sometimes puts the constant in a register, especially if it is
1519 too big to be an add immed operand. */
1520 src_insn
= PREV_INSN (src_insn
);
1521 increment
= SET_SRC (PATTERN (src_insn
));
1523 /* SR may have used LO_SUM to compute the constant if it is too large
1524 for a load immed operand. In this case, the constant is in operand
1525 one of the LO_SUM rtx. */
1526 if (GET_CODE (increment
) == LO_SUM
)
1527 increment
= XEXP (increment
, 1);
1529 /* Some ports store large constants in memory and add a REG_EQUAL
1530 note to the store insn. */
1531 else if (GET_CODE (increment
) == MEM
)
1533 rtx note
= find_reg_note (src_insn
, REG_EQUAL
, 0);
1535 increment
= XEXP (note
, 0);
1538 else if (GET_CODE (increment
) == IOR
1539 || GET_CODE (increment
) == ASHIFT
1540 || GET_CODE (increment
) == PLUS
)
1542 /* The rs6000 port loads some constants with IOR.
1543 The alpha port loads some constants with ASHIFT and PLUS. */
1544 rtx second_part
= XEXP (increment
, 1);
1545 enum rtx_code code
= GET_CODE (increment
);
1547 src_insn
= PREV_INSN (src_insn
);
1548 increment
= SET_SRC (PATTERN (src_insn
));
1549 /* Don't need the last insn anymore. */
1550 delete_insn (get_last_insn ());
1552 if (GET_CODE (second_part
) != CONST_INT
1553 || GET_CODE (increment
) != CONST_INT
)
1557 increment
= GEN_INT (INTVAL (increment
) | INTVAL (second_part
));
1558 else if (code
== PLUS
)
1559 increment
= GEN_INT (INTVAL (increment
) + INTVAL (second_part
));
1561 increment
= GEN_INT (INTVAL (increment
) << INTVAL (second_part
));
1564 if (GET_CODE (increment
) != CONST_INT
)
1567 /* The insn loading the constant into a register is no longer needed,
1569 delete_insn (get_last_insn ());
1572 if (increment_total
)
1573 increment_total
= GEN_INT (INTVAL (increment_total
) + INTVAL (increment
));
1575 increment_total
= increment
;
1577 /* Check that the source register is the same as the register we expected
1578 to see as the source. If not, something is seriously wrong. */
1579 if (GET_CODE (XEXP (SET_SRC (pattern
), 0)) != REG
1580 || REGNO (XEXP (SET_SRC (pattern
), 0)) != regno
)
1582 /* Some machines (e.g. the romp), may emit two add instructions for
1583 certain constants, so lets try looking for another add immediately
1584 before this one if we have only seen one add insn so far. */
1590 src_insn
= PREV_INSN (src_insn
);
1591 pattern
= PATTERN (src_insn
);
1593 delete_insn (get_last_insn ());
1601 return increment_total
;
1604 /* Copy REG_NOTES, except for insn references, because not all insn_map
1605 entries are valid yet. We do need to copy registers now though, because
1606 the reg_map entries can change during copying. */
1609 initial_reg_note_copy (notes
, map
)
1611 struct inline_remap
*map
;
1618 copy
= rtx_alloc (GET_CODE (notes
));
1619 PUT_MODE (copy
, GET_MODE (notes
));
1621 if (GET_CODE (notes
) == EXPR_LIST
)
1622 XEXP (copy
, 0) = copy_rtx_and_substitute (XEXP (notes
, 0), map
);
1623 else if (GET_CODE (notes
) == INSN_LIST
)
1624 /* Don't substitute for these yet. */
1625 XEXP (copy
, 0) = XEXP (notes
, 0);
1629 XEXP (copy
, 1) = initial_reg_note_copy (XEXP (notes
, 1), map
);
1634 /* Fixup insn references in copied REG_NOTES. */
1637 final_reg_note_copy (notes
, map
)
1639 struct inline_remap
*map
;
1643 for (note
= notes
; note
; note
= XEXP (note
, 1))
1644 if (GET_CODE (note
) == INSN_LIST
)
1645 XEXP (note
, 0) = map
->insn_map
[INSN_UID (XEXP (note
, 0))];
1648 /* Copy each instruction in the loop, substituting from map as appropriate.
1649 This is very similar to a loop in expand_inline_function. */
1652 copy_loop_body (copy_start
, copy_end
, map
, exit_label
, last_iteration
,
1653 unroll_type
, start_label
, loop_end
, insert_before
,
1655 rtx copy_start
, copy_end
;
1656 struct inline_remap
*map
;
1659 enum unroll_types unroll_type
;
1660 rtx start_label
, loop_end
, insert_before
, copy_notes_from
;
1664 int dest_reg_was_split
, i
;
1668 rtx final_label
= 0;
1669 rtx giv_inc
, giv_dest_reg
, giv_src_reg
;
1671 /* If this isn't the last iteration, then map any references to the
1672 start_label to final_label. Final label will then be emitted immediately
1673 after the end of this loop body if it was ever used.
1675 If this is the last iteration, then map references to the start_label
1677 if (! last_iteration
)
1679 final_label
= gen_label_rtx ();
1680 set_label_in_map (map
, CODE_LABEL_NUMBER (start_label
),
1684 set_label_in_map (map
, CODE_LABEL_NUMBER (start_label
), start_label
);
1688 /* Emit a NOTE_INSN_DELETED to force at least two insns onto the sequence.
1689 Else gen_sequence could return a raw pattern for a jump which we pass
1690 off to emit_insn_before (instead of emit_jump_insn_before) which causes
1691 a variety of losing behaviors later. */
1692 emit_note (0, NOTE_INSN_DELETED
);
1697 insn
= NEXT_INSN (insn
);
1699 map
->orig_asm_operands_vector
= 0;
1701 switch (GET_CODE (insn
))
1704 pattern
= PATTERN (insn
);
1708 /* Check to see if this is a giv that has been combined with
1709 some split address givs. (Combined in the sense that
1710 `combine_givs' in loop.c has put two givs in the same register.)
1711 In this case, we must search all givs based on the same biv to
1712 find the address givs. Then split the address givs.
1713 Do this before splitting the giv, since that may map the
1714 SET_DEST to a new register. */
1716 if ((set
= single_set (insn
))
1717 && GET_CODE (SET_DEST (set
)) == REG
1718 && addr_combined_regs
[REGNO (SET_DEST (set
))])
1720 struct iv_class
*bl
;
1721 struct induction
*v
, *tv
;
1722 int regno
= REGNO (SET_DEST (set
));
1724 v
= addr_combined_regs
[REGNO (SET_DEST (set
))];
1725 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
1727 /* Although the giv_inc amount is not needed here, we must call
1728 calculate_giv_inc here since it might try to delete the
1729 last insn emitted. If we wait until later to call it,
1730 we might accidentally delete insns generated immediately
1731 below by emit_unrolled_add. */
1733 if (! derived_regs
[regno
])
1734 giv_inc
= calculate_giv_inc (set
, insn
, regno
);
1736 /* Now find all address giv's that were combined with this
1738 for (tv
= bl
->giv
; tv
; tv
= tv
->next_iv
)
1739 if (tv
->giv_type
== DEST_ADDR
&& tv
->same
== v
)
1743 /* If this DEST_ADDR giv was not split, then ignore it. */
1744 if (*tv
->location
!= tv
->dest_reg
)
1747 /* Scale this_giv_inc if the multiplicative factors of
1748 the two givs are different. */
1749 this_giv_inc
= INTVAL (giv_inc
);
1750 if (tv
->mult_val
!= v
->mult_val
)
1751 this_giv_inc
= (this_giv_inc
/ INTVAL (v
->mult_val
)
1752 * INTVAL (tv
->mult_val
));
1754 tv
->dest_reg
= plus_constant (tv
->dest_reg
, this_giv_inc
);
1755 *tv
->location
= tv
->dest_reg
;
1757 if (last_iteration
&& unroll_type
!= UNROLL_COMPLETELY
)
1759 /* Must emit an insn to increment the split address
1760 giv. Add in the const_adjust field in case there
1761 was a constant eliminated from the address. */
1762 rtx value
, dest_reg
;
1764 /* tv->dest_reg will be either a bare register,
1765 or else a register plus a constant. */
1766 if (GET_CODE (tv
->dest_reg
) == REG
)
1767 dest_reg
= tv
->dest_reg
;
1769 dest_reg
= XEXP (tv
->dest_reg
, 0);
1771 /* Check for shared address givs, and avoid
1772 incrementing the shared pseudo reg more than
1774 if (! tv
->same_insn
&& ! tv
->shared
)
1776 /* tv->dest_reg may actually be a (PLUS (REG)
1777 (CONST)) here, so we must call plus_constant
1778 to add the const_adjust amount before calling
1779 emit_unrolled_add below. */
1780 value
= plus_constant (tv
->dest_reg
,
1783 /* The constant could be too large for an add
1784 immediate, so can't directly emit an insn
1786 emit_unrolled_add (dest_reg
, XEXP (value
, 0),
1790 /* Reset the giv to be just the register again, in case
1791 it is used after the set we have just emitted.
1792 We must subtract the const_adjust factor added in
1794 tv
->dest_reg
= plus_constant (dest_reg
,
1795 - tv
->const_adjust
);
1796 *tv
->location
= tv
->dest_reg
;
1801 /* If this is a setting of a splittable variable, then determine
1802 how to split the variable, create a new set based on this split,
1803 and set up the reg_map so that later uses of the variable will
1804 use the new split variable. */
1806 dest_reg_was_split
= 0;
1808 if ((set
= single_set (insn
))
1809 && GET_CODE (SET_DEST (set
)) == REG
1810 && splittable_regs
[REGNO (SET_DEST (set
))])
1812 int regno
= REGNO (SET_DEST (set
));
1815 dest_reg_was_split
= 1;
1817 giv_dest_reg
= SET_DEST (set
);
1818 if (derived_regs
[regno
])
1820 /* ??? This relies on SET_SRC (SET) to be of
1821 the form (plus (reg) (const_int)), and thus
1822 forces recombine_givs to restrict the kind
1823 of giv derivations it does before unrolling. */
1824 giv_src_reg
= XEXP (SET_SRC (set
), 0);
1825 giv_inc
= XEXP (SET_SRC (set
), 1);
1829 giv_src_reg
= giv_dest_reg
;
1830 /* Compute the increment value for the giv, if it wasn't
1831 already computed above. */
1833 giv_inc
= calculate_giv_inc (set
, insn
, regno
);
1835 src_regno
= REGNO (giv_src_reg
);
1837 if (unroll_type
== UNROLL_COMPLETELY
)
1839 /* Completely unrolling the loop. Set the induction
1840 variable to a known constant value. */
1842 /* The value in splittable_regs may be an invariant
1843 value, so we must use plus_constant here. */
1844 splittable_regs
[regno
]
1845 = plus_constant (splittable_regs
[src_regno
],
1848 if (GET_CODE (splittable_regs
[regno
]) == PLUS
)
1850 giv_src_reg
= XEXP (splittable_regs
[regno
], 0);
1851 giv_inc
= XEXP (splittable_regs
[regno
], 1);
1855 /* The splittable_regs value must be a REG or a
1856 CONST_INT, so put the entire value in the giv_src_reg
1858 giv_src_reg
= splittable_regs
[regno
];
1859 giv_inc
= const0_rtx
;
1864 /* Partially unrolling loop. Create a new pseudo
1865 register for the iteration variable, and set it to
1866 be a constant plus the original register. Except
1867 on the last iteration, when the result has to
1868 go back into the original iteration var register. */
1870 /* Handle bivs which must be mapped to a new register
1871 when split. This happens for bivs which need their
1872 final value set before loop entry. The new register
1873 for the biv was stored in the biv's first struct
1874 induction entry by find_splittable_regs. */
1876 if (regno
< max_reg_before_loop
1877 && REG_IV_TYPE (regno
) == BASIC_INDUCT
)
1879 giv_src_reg
= reg_biv_class
[regno
]->biv
->src_reg
;
1880 giv_dest_reg
= giv_src_reg
;
1884 /* If non-reduced/final-value givs were split, then
1885 this would have to remap those givs also. See
1886 find_splittable_regs. */
1889 splittable_regs
[regno
]
1890 = GEN_INT (INTVAL (giv_inc
)
1891 + INTVAL (splittable_regs
[src_regno
]));
1892 giv_inc
= splittable_regs
[regno
];
1894 /* Now split the induction variable by changing the dest
1895 of this insn to a new register, and setting its
1896 reg_map entry to point to this new register.
1898 If this is the last iteration, and this is the last insn
1899 that will update the iv, then reuse the original dest,
1900 to ensure that the iv will have the proper value when
1901 the loop exits or repeats.
1903 Using splittable_regs_updates here like this is safe,
1904 because it can only be greater than one if all
1905 instructions modifying the iv are always executed in
1908 if (! last_iteration
1909 || (splittable_regs_updates
[regno
]-- != 1))
1911 tem
= gen_reg_rtx (GET_MODE (giv_src_reg
));
1913 map
->reg_map
[regno
] = tem
;
1914 record_base_value (REGNO (tem
),
1915 giv_inc
== const0_rtx
1917 : gen_rtx_PLUS (GET_MODE (giv_src_reg
),
1918 giv_src_reg
, giv_inc
),
1922 map
->reg_map
[regno
] = giv_src_reg
;
1925 /* The constant being added could be too large for an add
1926 immediate, so can't directly emit an insn here. */
1927 emit_unrolled_add (giv_dest_reg
, giv_src_reg
, giv_inc
);
1928 copy
= get_last_insn ();
1929 pattern
= PATTERN (copy
);
1933 pattern
= copy_rtx_and_substitute (pattern
, map
);
1934 copy
= emit_insn (pattern
);
1936 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
1939 /* If this insn is setting CC0, it may need to look at
1940 the insn that uses CC0 to see what type of insn it is.
1941 In that case, the call to recog via validate_change will
1942 fail. So don't substitute constants here. Instead,
1943 do it when we emit the following insn.
1945 For example, see the pyr.md file. That machine has signed and
1946 unsigned compares. The compare patterns must check the
1947 following branch insn to see which what kind of compare to
1950 If the previous insn set CC0, substitute constants on it as
1952 if (sets_cc0_p (PATTERN (copy
)) != 0)
1957 try_constants (cc0_insn
, map
);
1959 try_constants (copy
, map
);
1962 try_constants (copy
, map
);
1965 /* Make split induction variable constants `permanent' since we
1966 know there are no backward branches across iteration variable
1967 settings which would invalidate this. */
1968 if (dest_reg_was_split
)
1970 int regno
= REGNO (SET_DEST (pattern
));
1972 if (regno
< VARRAY_SIZE (map
->const_equiv_varray
)
1973 && (VARRAY_CONST_EQUIV (map
->const_equiv_varray
, regno
).age
1975 VARRAY_CONST_EQUIV (map
->const_equiv_varray
, regno
).age
= -1;
1980 pattern
= copy_rtx_and_substitute (PATTERN (insn
), map
);
1981 copy
= emit_jump_insn (pattern
);
1982 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
1984 if (JUMP_LABEL (insn
) == start_label
&& insn
== copy_end
1985 && ! last_iteration
)
1987 /* This is a branch to the beginning of the loop; this is the
1988 last insn being copied; and this is not the last iteration.
1989 In this case, we want to change the original fall through
1990 case to be a branch past the end of the loop, and the
1991 original jump label case to fall_through. */
1993 if (invert_exp (pattern
, copy
))
1995 if (! redirect_exp (&pattern
,
1996 get_label_from_map (map
,
1998 (JUMP_LABEL (insn
))),
2005 rtx lab
= gen_label_rtx ();
2006 /* Can't do it by reversing the jump (probably because we
2007 couldn't reverse the conditions), so emit a new
2008 jump_insn after COPY, and redirect the jump around
2010 jmp
= emit_jump_insn_after (gen_jump (exit_label
), copy
);
2011 jmp
= emit_barrier_after (jmp
);
2012 emit_label_after (lab
, jmp
);
2013 LABEL_NUSES (lab
) = 0;
2014 if (! redirect_exp (&pattern
,
2015 get_label_from_map (map
,
2017 (JUMP_LABEL (insn
))),
2025 try_constants (cc0_insn
, map
);
2028 try_constants (copy
, map
);
2030 /* Set the jump label of COPY correctly to avoid problems with
2031 later passes of unroll_loop, if INSN had jump label set. */
2032 if (JUMP_LABEL (insn
))
2036 /* Can't use the label_map for every insn, since this may be
2037 the backward branch, and hence the label was not mapped. */
2038 if ((set
= single_set (copy
)))
2040 tem
= SET_SRC (set
);
2041 if (GET_CODE (tem
) == LABEL_REF
)
2042 label
= XEXP (tem
, 0);
2043 else if (GET_CODE (tem
) == IF_THEN_ELSE
)
2045 if (XEXP (tem
, 1) != pc_rtx
)
2046 label
= XEXP (XEXP (tem
, 1), 0);
2048 label
= XEXP (XEXP (tem
, 2), 0);
2052 if (label
&& GET_CODE (label
) == CODE_LABEL
)
2053 JUMP_LABEL (copy
) = label
;
2056 /* An unrecognizable jump insn, probably the entry jump
2057 for a switch statement. This label must have been mapped,
2058 so just use the label_map to get the new jump label. */
2060 = get_label_from_map (map
,
2061 CODE_LABEL_NUMBER (JUMP_LABEL (insn
)));
2064 /* If this is a non-local jump, then must increase the label
2065 use count so that the label will not be deleted when the
2066 original jump is deleted. */
2067 LABEL_NUSES (JUMP_LABEL (copy
))++;
2069 else if (GET_CODE (PATTERN (copy
)) == ADDR_VEC
2070 || GET_CODE (PATTERN (copy
)) == ADDR_DIFF_VEC
)
2072 rtx pat
= PATTERN (copy
);
2073 int diff_vec_p
= GET_CODE (pat
) == ADDR_DIFF_VEC
;
2074 int len
= XVECLEN (pat
, diff_vec_p
);
2077 for (i
= 0; i
< len
; i
++)
2078 LABEL_NUSES (XEXP (XVECEXP (pat
, diff_vec_p
, i
), 0))++;
2081 /* If this used to be a conditional jump insn but whose branch
2082 direction is now known, we must do something special. */
2083 if (condjump_p (insn
) && !simplejump_p (insn
) && map
->last_pc_value
)
2086 /* The previous insn set cc0 for us. So delete it. */
2087 delete_insn (PREV_INSN (copy
));
2090 /* If this is now a no-op, delete it. */
2091 if (map
->last_pc_value
== pc_rtx
)
2093 /* Don't let delete_insn delete the label referenced here,
2094 because we might possibly need it later for some other
2095 instruction in the loop. */
2096 if (JUMP_LABEL (copy
))
2097 LABEL_NUSES (JUMP_LABEL (copy
))++;
2099 if (JUMP_LABEL (copy
))
2100 LABEL_NUSES (JUMP_LABEL (copy
))--;
2104 /* Otherwise, this is unconditional jump so we must put a
2105 BARRIER after it. We could do some dead code elimination
2106 here, but jump.c will do it just as well. */
2112 pattern
= copy_rtx_and_substitute (PATTERN (insn
), map
);
2113 copy
= emit_call_insn (pattern
);
2114 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
2116 /* Because the USAGE information potentially contains objects other
2117 than hard registers, we need to copy it. */
2118 CALL_INSN_FUNCTION_USAGE (copy
)
2119 = copy_rtx_and_substitute (CALL_INSN_FUNCTION_USAGE (insn
), map
);
2123 try_constants (cc0_insn
, map
);
2126 try_constants (copy
, map
);
2128 /* Be lazy and assume CALL_INSNs clobber all hard registers. */
2129 for (i
= 0; i
< FIRST_PSEUDO_REGISTER
; i
++)
2130 VARRAY_CONST_EQUIV (map
->const_equiv_varray
, i
).rtx
= 0;
2134 /* If this is the loop start label, then we don't need to emit a
2135 copy of this label since no one will use it. */
2137 if (insn
!= start_label
)
2139 copy
= emit_label (get_label_from_map (map
,
2140 CODE_LABEL_NUMBER (insn
)));
2146 copy
= emit_barrier ();
2150 /* VTOP and CONT notes are valid only before the loop exit test.
2151 If placed anywhere else, loop may generate bad code. */
2152 /* BASIC_BLOCK notes exist to stabilize basic block structures with
2153 the associated rtl. We do not want to share the structure in
2156 if (NOTE_LINE_NUMBER (insn
) != NOTE_INSN_DELETED
2157 && NOTE_LINE_NUMBER (insn
) != NOTE_INSN_BASIC_BLOCK
2158 && ((NOTE_LINE_NUMBER (insn
) != NOTE_INSN_LOOP_VTOP
2159 && NOTE_LINE_NUMBER (insn
) != NOTE_INSN_LOOP_CONT
)
2160 || (last_iteration
&& unroll_type
!= UNROLL_COMPLETELY
)))
2161 copy
= emit_note (NOTE_SOURCE_FILE (insn
),
2162 NOTE_LINE_NUMBER (insn
));
2172 map
->insn_map
[INSN_UID (insn
)] = copy
;
2174 while (insn
!= copy_end
);
2176 /* Now finish coping the REG_NOTES. */
2180 insn
= NEXT_INSN (insn
);
2181 if ((GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == JUMP_INSN
2182 || GET_CODE (insn
) == CALL_INSN
)
2183 && map
->insn_map
[INSN_UID (insn
)])
2184 final_reg_note_copy (REG_NOTES (map
->insn_map
[INSN_UID (insn
)]), map
);
2186 while (insn
!= copy_end
);
2188 /* There may be notes between copy_notes_from and loop_end. Emit a copy of
2189 each of these notes here, since there may be some important ones, such as
2190 NOTE_INSN_BLOCK_END notes, in this group. We don't do this on the last
2191 iteration, because the original notes won't be deleted.
2193 We can't use insert_before here, because when from preconditioning,
2194 insert_before points before the loop. We can't use copy_end, because
2195 there may be insns already inserted after it (which we don't want to
2196 copy) when not from preconditioning code. */
2198 if (! last_iteration
)
2200 for (insn
= copy_notes_from
; insn
!= loop_end
; insn
= NEXT_INSN (insn
))
2202 /* VTOP notes are valid only before the loop exit test.
2203 If placed anywhere else, loop may generate bad code.
2204 There is no need to test for NOTE_INSN_LOOP_CONT notes
2205 here, since COPY_NOTES_FROM will be at most one or two (for cc0)
2206 instructions before the last insn in the loop, and if the
2207 end test is that short, there will be a VTOP note between
2208 the CONT note and the test. */
2209 if (GET_CODE (insn
) == NOTE
2210 && NOTE_LINE_NUMBER (insn
) != NOTE_INSN_DELETED
2211 && NOTE_LINE_NUMBER (insn
) != NOTE_INSN_BASIC_BLOCK
2212 && NOTE_LINE_NUMBER (insn
) != NOTE_INSN_LOOP_VTOP
)
2213 emit_note (NOTE_SOURCE_FILE (insn
), NOTE_LINE_NUMBER (insn
));
2217 if (final_label
&& LABEL_NUSES (final_label
) > 0)
2218 emit_label (final_label
);
2220 tem
= gen_sequence ();
2222 emit_insn_before (tem
, insert_before
);
2225 /* Emit an insn, using the expand_binop to ensure that a valid insn is
2226 emitted. This will correctly handle the case where the increment value
2227 won't fit in the immediate field of a PLUS insns. */
2230 emit_unrolled_add (dest_reg
, src_reg
, increment
)
2231 rtx dest_reg
, src_reg
, increment
;
2235 result
= expand_binop (GET_MODE (dest_reg
), add_optab
, src_reg
, increment
,
2236 dest_reg
, 0, OPTAB_LIB_WIDEN
);
2238 if (dest_reg
!= result
)
2239 emit_move_insn (dest_reg
, result
);
2242 /* Searches the insns between INSN and LOOP_END. Returns 1 if there
2243 is a backward branch in that range that branches to somewhere between
2244 LOOP_START and INSN. Returns 0 otherwise. */
2246 /* ??? This is quadratic algorithm. Could be rewritten to be linear.
2247 In practice, this is not a problem, because this function is seldom called,
2248 and uses a negligible amount of CPU time on average. */
2251 back_branch_in_range_p (insn
, loop_start
, loop_end
)
2253 rtx loop_start
, loop_end
;
2255 rtx p
, q
, target_insn
;
2256 rtx orig_loop_end
= loop_end
;
2258 /* Stop before we get to the backward branch at the end of the loop. */
2259 loop_end
= prev_nonnote_insn (loop_end
);
2260 if (GET_CODE (loop_end
) == BARRIER
)
2261 loop_end
= PREV_INSN (loop_end
);
2263 /* Check in case insn has been deleted, search forward for first non
2264 deleted insn following it. */
2265 while (INSN_DELETED_P (insn
))
2266 insn
= NEXT_INSN (insn
);
2268 /* Check for the case where insn is the last insn in the loop. Deal
2269 with the case where INSN was a deleted loop test insn, in which case
2270 it will now be the NOTE_LOOP_END. */
2271 if (insn
== loop_end
|| insn
== orig_loop_end
)
2274 for (p
= NEXT_INSN (insn
); p
!= loop_end
; p
= NEXT_INSN (p
))
2276 if (GET_CODE (p
) == JUMP_INSN
)
2278 target_insn
= JUMP_LABEL (p
);
2280 /* Search from loop_start to insn, to see if one of them is
2281 the target_insn. We can't use INSN_LUID comparisons here,
2282 since insn may not have an LUID entry. */
2283 for (q
= loop_start
; q
!= insn
; q
= NEXT_INSN (q
))
2284 if (q
== target_insn
)
2292 /* Try to generate the simplest rtx for the expression
2293 (PLUS (MULT mult1 mult2) add1). This is used to calculate the initial
2297 fold_rtx_mult_add (mult1
, mult2
, add1
, mode
)
2298 rtx mult1
, mult2
, add1
;
2299 enum machine_mode mode
;
2304 /* The modes must all be the same. This should always be true. For now,
2305 check to make sure. */
2306 if ((GET_MODE (mult1
) != mode
&& GET_MODE (mult1
) != VOIDmode
)
2307 || (GET_MODE (mult2
) != mode
&& GET_MODE (mult2
) != VOIDmode
)
2308 || (GET_MODE (add1
) != mode
&& GET_MODE (add1
) != VOIDmode
))
2311 /* Ensure that if at least one of mult1/mult2 are constant, then mult2
2312 will be a constant. */
2313 if (GET_CODE (mult1
) == CONST_INT
)
2320 mult_res
= simplify_binary_operation (MULT
, mode
, mult1
, mult2
);
2322 mult_res
= gen_rtx_MULT (mode
, mult1
, mult2
);
2324 /* Again, put the constant second. */
2325 if (GET_CODE (add1
) == CONST_INT
)
2332 result
= simplify_binary_operation (PLUS
, mode
, add1
, mult_res
);
2334 result
= gen_rtx_PLUS (mode
, add1
, mult_res
);
2339 /* Searches the list of induction struct's for the biv BL, to try to calculate
2340 the total increment value for one iteration of the loop as a constant.
2342 Returns the increment value as an rtx, simplified as much as possible,
2343 if it can be calculated. Otherwise, returns 0. */
2346 biv_total_increment (bl
, loop_start
, loop_end
)
2347 struct iv_class
*bl
;
2348 rtx loop_start
, loop_end
;
2350 struct induction
*v
;
2353 /* For increment, must check every instruction that sets it. Each
2354 instruction must be executed only once each time through the loop.
2355 To verify this, we check that the insn is always executed, and that
2356 there are no backward branches after the insn that branch to before it.
2357 Also, the insn must have a mult_val of one (to make sure it really is
2360 result
= const0_rtx
;
2361 for (v
= bl
->biv
; v
; v
= v
->next_iv
)
2363 if (v
->always_computable
&& v
->mult_val
== const1_rtx
2364 && ! v
->maybe_multiple
)
2365 result
= fold_rtx_mult_add (result
, const1_rtx
, v
->add_val
, v
->mode
);
2373 /* Determine the initial value of the iteration variable, and the amount
2374 that it is incremented each loop. Use the tables constructed by
2375 the strength reduction pass to calculate these values.
2377 Initial_value and/or increment are set to zero if their values could not
2381 iteration_info (iteration_var
, initial_value
, increment
, loop_start
, loop_end
)
2382 rtx iteration_var
, *initial_value
, *increment
;
2383 rtx loop_start
, loop_end
;
2385 struct iv_class
*bl
;
2387 struct induction
*v
;
2390 /* Clear the result values, in case no answer can be found. */
2394 /* The iteration variable can be either a giv or a biv. Check to see
2395 which it is, and compute the variable's initial value, and increment
2396 value if possible. */
2398 /* If this is a new register, can't handle it since we don't have any
2399 reg_iv_type entry for it. */
2400 if ((unsigned) REGNO (iteration_var
) >= reg_iv_type
->num_elements
)
2402 if (loop_dump_stream
)
2403 fprintf (loop_dump_stream
,
2404 "Loop unrolling: No reg_iv_type entry for iteration var.\n");
2408 /* Reject iteration variables larger than the host wide int size, since they
2409 could result in a number of iterations greater than the range of our
2410 `unsigned HOST_WIDE_INT' variable loop_info->n_iterations. */
2411 else if ((GET_MODE_BITSIZE (GET_MODE (iteration_var
))
2412 > HOST_BITS_PER_WIDE_INT
))
2414 if (loop_dump_stream
)
2415 fprintf (loop_dump_stream
,
2416 "Loop unrolling: Iteration var rejected because mode too large.\n");
2419 else if (GET_MODE_CLASS (GET_MODE (iteration_var
)) != MODE_INT
)
2421 if (loop_dump_stream
)
2422 fprintf (loop_dump_stream
,
2423 "Loop unrolling: Iteration var not an integer.\n");
2426 else if (REG_IV_TYPE (REGNO (iteration_var
)) == BASIC_INDUCT
)
2428 /* When reg_iv_type / reg_iv_info is resized for biv increments
2429 that are turned into givs, reg_biv_class is not resized.
2430 So check here that we don't make an out-of-bounds access. */
2431 if (REGNO (iteration_var
) >= max_reg_before_loop
)
2434 /* Grab initial value, only useful if it is a constant. */
2435 bl
= reg_biv_class
[REGNO (iteration_var
)];
2436 *initial_value
= bl
->initial_value
;
2438 *increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2440 else if (REG_IV_TYPE (REGNO (iteration_var
)) == GENERAL_INDUCT
)
2442 HOST_WIDE_INT offset
= 0;
2443 struct induction
*v
= REG_IV_INFO (REGNO (iteration_var
));
2445 if (REGNO (v
->src_reg
) >= max_reg_before_loop
)
2448 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
2450 /* Increment value is mult_val times the increment value of the biv. */
2452 *increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2455 struct induction
*biv_inc
;
2458 = fold_rtx_mult_add (v
->mult_val
, *increment
, const0_rtx
, v
->mode
);
2459 /* The caller assumes that one full increment has occured at the
2460 first loop test. But that's not true when the biv is incremented
2461 after the giv is set (which is the usual case), e.g.:
2462 i = 6; do {;} while (i++ < 9) .
2463 Therefore, we bias the initial value by subtracting the amount of
2464 the increment that occurs between the giv set and the giv test. */
2465 for (biv_inc
= bl
->biv
; biv_inc
; biv_inc
= biv_inc
->next_iv
)
2467 if (loop_insn_first_p (v
->insn
, biv_inc
->insn
))
2468 offset
-= INTVAL (biv_inc
->add_val
);
2470 offset
*= INTVAL (v
->mult_val
);
2472 if (loop_dump_stream
)
2473 fprintf (loop_dump_stream
,
2474 "Loop unrolling: Giv iterator, initial value bias %ld.\n",
2476 /* Initial value is mult_val times the biv's initial value plus
2477 add_val. Only useful if it is a constant. */
2479 = fold_rtx_mult_add (v
->mult_val
,
2480 plus_constant (bl
->initial_value
, offset
),
2481 v
->add_val
, v
->mode
);
2485 if (loop_dump_stream
)
2486 fprintf (loop_dump_stream
,
2487 "Loop unrolling: Not basic or general induction var.\n");
2493 /* For each biv and giv, determine whether it can be safely split into
2494 a different variable for each unrolled copy of the loop body. If it
2495 is safe to split, then indicate that by saving some useful info
2496 in the splittable_regs array.
2498 If the loop is being completely unrolled, then splittable_regs will hold
2499 the current value of the induction variable while the loop is unrolled.
2500 It must be set to the initial value of the induction variable here.
2501 Otherwise, splittable_regs will hold the difference between the current
2502 value of the induction variable and the value the induction variable had
2503 at the top of the loop. It must be set to the value 0 here.
2505 Returns the total number of instructions that set registers that are
2508 /* ?? If the loop is only unrolled twice, then most of the restrictions to
2509 constant values are unnecessary, since we can easily calculate increment
2510 values in this case even if nothing is constant. The increment value
2511 should not involve a multiply however. */
2513 /* ?? Even if the biv/giv increment values aren't constant, it may still
2514 be beneficial to split the variable if the loop is only unrolled a few
2515 times, since multiplies by small integers (1,2,3,4) are very cheap. */
2518 find_splittable_regs (unroll_type
, loop_start
, loop_end
, end_insert_before
,
2519 unroll_number
, n_iterations
)
2520 enum unroll_types unroll_type
;
2521 rtx loop_start
, loop_end
;
2522 rtx end_insert_before
;
2524 unsigned HOST_WIDE_INT n_iterations
;
2526 struct iv_class
*bl
;
2527 struct induction
*v
;
2529 rtx biv_final_value
;
2533 for (bl
= loop_iv_list
; bl
; bl
= bl
->next
)
2535 /* Biv_total_increment must return a constant value,
2536 otherwise we can not calculate the split values. */
2538 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2539 if (! increment
|| GET_CODE (increment
) != CONST_INT
)
2542 /* The loop must be unrolled completely, or else have a known number
2543 of iterations and only one exit, or else the biv must be dead
2544 outside the loop, or else the final value must be known. Otherwise,
2545 it is unsafe to split the biv since it may not have the proper
2546 value on loop exit. */
2548 /* loop_number_exit_count is non-zero if the loop has an exit other than
2549 a fall through at the end. */
2552 biv_final_value
= 0;
2553 if (unroll_type
!= UNROLL_COMPLETELY
2554 && (loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]]
2555 || unroll_type
== UNROLL_NAIVE
)
2556 && (uid_luid
[REGNO_LAST_UID (bl
->regno
)] >= INSN_LUID (loop_end
)
2558 || INSN_UID (bl
->init_insn
) >= max_uid_for_loop
2559 || (uid_luid
[REGNO_FIRST_UID (bl
->regno
)]
2560 < INSN_LUID (bl
->init_insn
))
2561 || reg_mentioned_p (bl
->biv
->dest_reg
, SET_SRC (bl
->init_set
)))
2562 && ! (biv_final_value
= final_biv_value (bl
, loop_start
, loop_end
,
2566 /* If any of the insns setting the BIV don't do so with a simple
2567 PLUS, we don't know how to split it. */
2568 for (v
= bl
->biv
; biv_splittable
&& v
; v
= v
->next_iv
)
2569 if ((tem
= single_set (v
->insn
)) == 0
2570 || GET_CODE (SET_DEST (tem
)) != REG
2571 || REGNO (SET_DEST (tem
)) != bl
->regno
2572 || GET_CODE (SET_SRC (tem
)) != PLUS
)
2575 /* If final value is non-zero, then must emit an instruction which sets
2576 the value of the biv to the proper value. This is done after
2577 handling all of the givs, since some of them may need to use the
2578 biv's value in their initialization code. */
2580 /* This biv is splittable. If completely unrolling the loop, save
2581 the biv's initial value. Otherwise, save the constant zero. */
2583 if (biv_splittable
== 1)
2585 if (unroll_type
== UNROLL_COMPLETELY
)
2587 /* If the initial value of the biv is itself (i.e. it is too
2588 complicated for strength_reduce to compute), or is a hard
2589 register, or it isn't invariant, then we must create a new
2590 pseudo reg to hold the initial value of the biv. */
2592 if (GET_CODE (bl
->initial_value
) == REG
2593 && (REGNO (bl
->initial_value
) == bl
->regno
2594 || REGNO (bl
->initial_value
) < FIRST_PSEUDO_REGISTER
2595 || ! invariant_p (bl
->initial_value
)))
2597 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2599 record_base_value (REGNO (tem
), bl
->biv
->add_val
, 0);
2600 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2603 if (loop_dump_stream
)
2604 fprintf (loop_dump_stream
, "Biv %d initial value remapped to %d.\n",
2605 bl
->regno
, REGNO (tem
));
2607 splittable_regs
[bl
->regno
] = tem
;
2610 splittable_regs
[bl
->regno
] = bl
->initial_value
;
2613 splittable_regs
[bl
->regno
] = const0_rtx
;
2615 /* Save the number of instructions that modify the biv, so that
2616 we can treat the last one specially. */
2618 splittable_regs_updates
[bl
->regno
] = bl
->biv_count
;
2619 result
+= bl
->biv_count
;
2621 if (loop_dump_stream
)
2622 fprintf (loop_dump_stream
,
2623 "Biv %d safe to split.\n", bl
->regno
);
2626 /* Check every giv that depends on this biv to see whether it is
2627 splittable also. Even if the biv isn't splittable, givs which
2628 depend on it may be splittable if the biv is live outside the
2629 loop, and the givs aren't. */
2631 result
+= find_splittable_givs (bl
, unroll_type
, loop_start
, loop_end
,
2632 increment
, unroll_number
);
2634 /* If final value is non-zero, then must emit an instruction which sets
2635 the value of the biv to the proper value. This is done after
2636 handling all of the givs, since some of them may need to use the
2637 biv's value in their initialization code. */
2638 if (biv_final_value
)
2640 /* If the loop has multiple exits, emit the insns before the
2641 loop to ensure that it will always be executed no matter
2642 how the loop exits. Otherwise emit the insn after the loop,
2643 since this is slightly more efficient. */
2644 if (! loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]])
2645 emit_insn_before (gen_move_insn (bl
->biv
->src_reg
,
2650 /* Create a new register to hold the value of the biv, and then
2651 set the biv to its final value before the loop start. The biv
2652 is set to its final value before loop start to ensure that
2653 this insn will always be executed, no matter how the loop
2655 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2656 record_base_value (REGNO (tem
), bl
->biv
->add_val
, 0);
2658 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2660 emit_insn_before (gen_move_insn (bl
->biv
->src_reg
,
2664 if (loop_dump_stream
)
2665 fprintf (loop_dump_stream
, "Biv %d mapped to %d for split.\n",
2666 REGNO (bl
->biv
->src_reg
), REGNO (tem
));
2668 /* Set up the mapping from the original biv register to the new
2670 bl
->biv
->src_reg
= tem
;
2677 /* Return 1 if the first and last unrolled copy of the address giv V is valid
2678 for the instruction that is using it. Do not make any changes to that
2682 verify_addresses (v
, giv_inc
, unroll_number
)
2683 struct induction
*v
;
2688 rtx orig_addr
= *v
->location
;
2689 rtx last_addr
= plus_constant (v
->dest_reg
,
2690 INTVAL (giv_inc
) * (unroll_number
- 1));
2692 /* First check to see if either address would fail. Handle the fact
2693 that we have may have a match_dup. */
2694 if (! validate_replace_rtx (*v
->location
, v
->dest_reg
, v
->insn
)
2695 || ! validate_replace_rtx (*v
->location
, last_addr
, v
->insn
))
2698 /* Now put things back the way they were before. This should always
2700 if (! validate_replace_rtx (*v
->location
, orig_addr
, v
->insn
))
2706 /* For every giv based on the biv BL, check to determine whether it is
2707 splittable. This is a subroutine to find_splittable_regs ().
2709 Return the number of instructions that set splittable registers. */
2712 find_splittable_givs (bl
, unroll_type
, loop_start
, loop_end
, increment
,
2714 struct iv_class
*bl
;
2715 enum unroll_types unroll_type
;
2716 rtx loop_start
, loop_end
;
2720 struct induction
*v
, *v2
;
2725 /* Scan the list of givs, and set the same_insn field when there are
2726 multiple identical givs in the same insn. */
2727 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
2728 for (v2
= v
->next_iv
; v2
; v2
= v2
->next_iv
)
2729 if (v
->insn
== v2
->insn
&& rtx_equal_p (v
->new_reg
, v2
->new_reg
)
2733 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
2737 /* Only split the giv if it has already been reduced, or if the loop is
2738 being completely unrolled. */
2739 if (unroll_type
!= UNROLL_COMPLETELY
&& v
->ignore
)
2742 /* The giv can be split if the insn that sets the giv is executed once
2743 and only once on every iteration of the loop. */
2744 /* An address giv can always be split. v->insn is just a use not a set,
2745 and hence it does not matter whether it is always executed. All that
2746 matters is that all the biv increments are always executed, and we
2747 won't reach here if they aren't. */
2748 if (v
->giv_type
!= DEST_ADDR
2749 && (! v
->always_computable
2750 || back_branch_in_range_p (v
->insn
, loop_start
, loop_end
)))
2753 /* The giv increment value must be a constant. */
2754 giv_inc
= fold_rtx_mult_add (v
->mult_val
, increment
, const0_rtx
,
2756 if (! giv_inc
|| GET_CODE (giv_inc
) != CONST_INT
)
2759 /* The loop must be unrolled completely, or else have a known number of
2760 iterations and only one exit, or else the giv must be dead outside
2761 the loop, or else the final value of the giv must be known.
2762 Otherwise, it is not safe to split the giv since it may not have the
2763 proper value on loop exit. */
2765 /* The used outside loop test will fail for DEST_ADDR givs. They are
2766 never used outside the loop anyways, so it is always safe to split a
2770 if (unroll_type
!= UNROLL_COMPLETELY
2771 && (loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]]
2772 || unroll_type
== UNROLL_NAIVE
)
2773 && v
->giv_type
!= DEST_ADDR
2774 /* The next part is true if the pseudo is used outside the loop.
2775 We assume that this is true for any pseudo created after loop
2776 starts, because we don't have a reg_n_info entry for them. */
2777 && (REGNO (v
->dest_reg
) >= max_reg_before_loop
2778 || (REGNO_FIRST_UID (REGNO (v
->dest_reg
)) != INSN_UID (v
->insn
)
2779 /* Check for the case where the pseudo is set by a shift/add
2780 sequence, in which case the first insn setting the pseudo
2781 is the first insn of the shift/add sequence. */
2782 && (! (tem
= find_reg_note (v
->insn
, REG_RETVAL
, NULL_RTX
))
2783 || (REGNO_FIRST_UID (REGNO (v
->dest_reg
))
2784 != INSN_UID (XEXP (tem
, 0)))))
2785 /* Line above always fails if INSN was moved by loop opt. */
2786 || (uid_luid
[REGNO_LAST_UID (REGNO (v
->dest_reg
))]
2787 >= INSN_LUID (loop_end
)))
2788 /* Givs made from biv increments are missed by the above test, so
2789 test explicitly for them. */
2790 && (REGNO (v
->dest_reg
) < first_increment_giv
2791 || REGNO (v
->dest_reg
) > last_increment_giv
)
2792 && ! (final_value
= v
->final_value
))
2796 /* Currently, non-reduced/final-value givs are never split. */
2797 /* Should emit insns after the loop if possible, as the biv final value
2800 /* If the final value is non-zero, and the giv has not been reduced,
2801 then must emit an instruction to set the final value. */
2802 if (final_value
&& !v
->new_reg
)
2804 /* Create a new register to hold the value of the giv, and then set
2805 the giv to its final value before the loop start. The giv is set
2806 to its final value before loop start to ensure that this insn
2807 will always be executed, no matter how we exit. */
2808 tem
= gen_reg_rtx (v
->mode
);
2809 emit_insn_before (gen_move_insn (tem
, v
->dest_reg
), loop_start
);
2810 emit_insn_before (gen_move_insn (v
->dest_reg
, final_value
),
2813 if (loop_dump_stream
)
2814 fprintf (loop_dump_stream
, "Giv %d mapped to %d for split.\n",
2815 REGNO (v
->dest_reg
), REGNO (tem
));
2821 /* This giv is splittable. If completely unrolling the loop, save the
2822 giv's initial value. Otherwise, save the constant zero for it. */
2824 if (unroll_type
== UNROLL_COMPLETELY
)
2826 /* It is not safe to use bl->initial_value here, because it may not
2827 be invariant. It is safe to use the initial value stored in
2828 the splittable_regs array if it is set. In rare cases, it won't
2829 be set, so then we do exactly the same thing as
2830 find_splittable_regs does to get a safe value. */
2831 rtx biv_initial_value
;
2833 if (splittable_regs
[bl
->regno
])
2834 biv_initial_value
= splittable_regs
[bl
->regno
];
2835 else if (GET_CODE (bl
->initial_value
) != REG
2836 || (REGNO (bl
->initial_value
) != bl
->regno
2837 && REGNO (bl
->initial_value
) >= FIRST_PSEUDO_REGISTER
))
2838 biv_initial_value
= bl
->initial_value
;
2841 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2843 record_base_value (REGNO (tem
), bl
->biv
->add_val
, 0);
2844 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2846 biv_initial_value
= tem
;
2848 value
= fold_rtx_mult_add (v
->mult_val
, biv_initial_value
,
2849 v
->add_val
, v
->mode
);
2856 /* If a giv was combined with another giv, then we can only split
2857 this giv if the giv it was combined with was reduced. This
2858 is because the value of v->new_reg is meaningless in this
2860 if (v
->same
&& ! v
->same
->new_reg
)
2862 if (loop_dump_stream
)
2863 fprintf (loop_dump_stream
,
2864 "giv combined with unreduced giv not split.\n");
2867 /* If the giv is an address destination, it could be something other
2868 than a simple register, these have to be treated differently. */
2869 else if (v
->giv_type
== DEST_REG
)
2871 /* If value is not a constant, register, or register plus
2872 constant, then compute its value into a register before
2873 loop start. This prevents invalid rtx sharing, and should
2874 generate better code. We can use bl->initial_value here
2875 instead of splittable_regs[bl->regno] because this code
2876 is going before the loop start. */
2877 if (unroll_type
== UNROLL_COMPLETELY
2878 && GET_CODE (value
) != CONST_INT
2879 && GET_CODE (value
) != REG
2880 && (GET_CODE (value
) != PLUS
2881 || GET_CODE (XEXP (value
, 0)) != REG
2882 || GET_CODE (XEXP (value
, 1)) != CONST_INT
))
2884 rtx tem
= gen_reg_rtx (v
->mode
);
2885 record_base_value (REGNO (tem
), v
->add_val
, 0);
2886 emit_iv_add_mult (bl
->initial_value
, v
->mult_val
,
2887 v
->add_val
, tem
, loop_start
);
2891 splittable_regs
[REGNO (v
->new_reg
)] = value
;
2892 derived_regs
[REGNO (v
->new_reg
)] = v
->derived_from
!= 0;
2896 /* Splitting address givs is useful since it will often allow us
2897 to eliminate some increment insns for the base giv as
2900 /* If the addr giv is combined with a dest_reg giv, then all
2901 references to that dest reg will be remapped, which is NOT
2902 what we want for split addr regs. We always create a new
2903 register for the split addr giv, just to be safe. */
2905 /* If we have multiple identical address givs within a
2906 single instruction, then use a single pseudo reg for
2907 both. This is necessary in case one is a match_dup
2910 v
->const_adjust
= 0;
2914 v
->dest_reg
= v
->same_insn
->dest_reg
;
2915 if (loop_dump_stream
)
2916 fprintf (loop_dump_stream
,
2917 "Sharing address givs in insn %d\n",
2918 INSN_UID (v
->insn
));
2920 /* If multiple address GIVs have been combined with the
2921 same dest_reg GIV, do not create a new register for
2923 else if (unroll_type
!= UNROLL_COMPLETELY
2924 && v
->giv_type
== DEST_ADDR
2925 && v
->same
&& v
->same
->giv_type
== DEST_ADDR
2926 && v
->same
->unrolled
2927 /* combine_givs_p may return true for some cases
2928 where the add and mult values are not equal.
2929 To share a register here, the values must be
2931 && rtx_equal_p (v
->same
->mult_val
, v
->mult_val
)
2932 && rtx_equal_p (v
->same
->add_val
, v
->add_val
)
2933 /* If the memory references have different modes,
2934 then the address may not be valid and we must
2935 not share registers. */
2936 && verify_addresses (v
, giv_inc
, unroll_number
))
2938 v
->dest_reg
= v
->same
->dest_reg
;
2941 else if (unroll_type
!= UNROLL_COMPLETELY
)
2943 /* If not completely unrolling the loop, then create a new
2944 register to hold the split value of the DEST_ADDR giv.
2945 Emit insn to initialize its value before loop start. */
2947 rtx tem
= gen_reg_rtx (v
->mode
);
2948 struct induction
*same
= v
->same
;
2949 rtx new_reg
= v
->new_reg
;
2950 record_base_value (REGNO (tem
), v
->add_val
, 0);
2952 if (same
&& same
->derived_from
)
2954 /* calculate_giv_inc doesn't work for derived givs.
2955 copy_loop_body works around the problem for the
2956 DEST_REG givs themselves, but it can't handle
2957 DEST_ADDR givs that have been combined with
2958 a derived DEST_REG giv.
2959 So Handle V as if the giv from which V->SAME has
2960 been derived has been combined with V.
2961 recombine_givs only derives givs from givs that
2962 are reduced the ordinary, so we need not worry
2963 about same->derived_from being in turn derived. */
2965 same
= same
->derived_from
;
2966 new_reg
= express_from (same
, v
);
2967 new_reg
= replace_rtx (new_reg
, same
->dest_reg
,
2971 /* If the address giv has a constant in its new_reg value,
2972 then this constant can be pulled out and put in value,
2973 instead of being part of the initialization code. */
2975 if (GET_CODE (new_reg
) == PLUS
2976 && GET_CODE (XEXP (new_reg
, 1)) == CONST_INT
)
2979 = plus_constant (tem
, INTVAL (XEXP (new_reg
, 1)));
2981 /* Only succeed if this will give valid addresses.
2982 Try to validate both the first and the last
2983 address resulting from loop unrolling, if
2984 one fails, then can't do const elim here. */
2985 if (verify_addresses (v
, giv_inc
, unroll_number
))
2987 /* Save the negative of the eliminated const, so
2988 that we can calculate the dest_reg's increment
2990 v
->const_adjust
= - INTVAL (XEXP (new_reg
, 1));
2992 new_reg
= XEXP (new_reg
, 0);
2993 if (loop_dump_stream
)
2994 fprintf (loop_dump_stream
,
2995 "Eliminating constant from giv %d\n",
3004 /* If the address hasn't been checked for validity yet, do so
3005 now, and fail completely if either the first or the last
3006 unrolled copy of the address is not a valid address
3007 for the instruction that uses it. */
3008 if (v
->dest_reg
== tem
3009 && ! verify_addresses (v
, giv_inc
, unroll_number
))
3011 for (v2
= v
->next_iv
; v2
; v2
= v2
->next_iv
)
3012 if (v2
->same_insn
== v
)
3015 if (loop_dump_stream
)
3016 fprintf (loop_dump_stream
,
3017 "Invalid address for giv at insn %d\n",
3018 INSN_UID (v
->insn
));
3022 v
->new_reg
= new_reg
;
3025 /* We set this after the address check, to guarantee that
3026 the register will be initialized. */
3029 /* To initialize the new register, just move the value of
3030 new_reg into it. This is not guaranteed to give a valid
3031 instruction on machines with complex addressing modes.
3032 If we can't recognize it, then delete it and emit insns
3033 to calculate the value from scratch. */
3034 emit_insn_before (gen_rtx_SET (VOIDmode
, tem
,
3035 copy_rtx (v
->new_reg
)),
3037 if (recog_memoized (PREV_INSN (loop_start
)) < 0)
3041 /* We can't use bl->initial_value to compute the initial
3042 value, because the loop may have been preconditioned.
3043 We must calculate it from NEW_REG. Try using
3044 force_operand instead of emit_iv_add_mult. */
3045 delete_insn (PREV_INSN (loop_start
));
3048 ret
= force_operand (v
->new_reg
, tem
);
3050 emit_move_insn (tem
, ret
);
3051 sequence
= gen_sequence ();
3053 emit_insn_before (sequence
, loop_start
);
3055 if (loop_dump_stream
)
3056 fprintf (loop_dump_stream
,
3057 "Invalid init insn, rewritten.\n");
3062 v
->dest_reg
= value
;
3064 /* Check the resulting address for validity, and fail
3065 if the resulting address would be invalid. */
3066 if (! verify_addresses (v
, giv_inc
, unroll_number
))
3068 for (v2
= v
->next_iv
; v2
; v2
= v2
->next_iv
)
3069 if (v2
->same_insn
== v
)
3072 if (loop_dump_stream
)
3073 fprintf (loop_dump_stream
,
3074 "Invalid address for giv at insn %d\n",
3075 INSN_UID (v
->insn
));
3078 if (v
->same
&& v
->same
->derived_from
)
3080 /* Handle V as if the giv from which V->SAME has
3081 been derived has been combined with V. */
3083 v
->same
= v
->same
->derived_from
;
3084 v
->new_reg
= express_from (v
->same
, v
);
3085 v
->new_reg
= replace_rtx (v
->new_reg
, v
->same
->dest_reg
,
3091 /* Store the value of dest_reg into the insn. This sharing
3092 will not be a problem as this insn will always be copied
3095 *v
->location
= v
->dest_reg
;
3097 /* If this address giv is combined with a dest reg giv, then
3098 save the base giv's induction pointer so that we will be
3099 able to handle this address giv properly. The base giv
3100 itself does not have to be splittable. */
3102 if (v
->same
&& v
->same
->giv_type
== DEST_REG
)
3103 addr_combined_regs
[REGNO (v
->same
->new_reg
)] = v
->same
;
3105 if (GET_CODE (v
->new_reg
) == REG
)
3107 /* This giv maybe hasn't been combined with any others.
3108 Make sure that it's giv is marked as splittable here. */
3110 splittable_regs
[REGNO (v
->new_reg
)] = value
;
3111 derived_regs
[REGNO (v
->new_reg
)] = v
->derived_from
!= 0;
3113 /* Make it appear to depend upon itself, so that the
3114 giv will be properly split in the main loop above. */
3118 addr_combined_regs
[REGNO (v
->new_reg
)] = v
;
3122 if (loop_dump_stream
)
3123 fprintf (loop_dump_stream
, "DEST_ADDR giv being split.\n");
3129 /* Currently, unreduced giv's can't be split. This is not too much
3130 of a problem since unreduced giv's are not live across loop
3131 iterations anyways. When unrolling a loop completely though,
3132 it makes sense to reduce&split givs when possible, as this will
3133 result in simpler instructions, and will not require that a reg
3134 be live across loop iterations. */
3136 splittable_regs
[REGNO (v
->dest_reg
)] = value
;
3137 fprintf (stderr
, "Giv %d at insn %d not reduced\n",
3138 REGNO (v
->dest_reg
), INSN_UID (v
->insn
));
3144 /* Unreduced givs are only updated once by definition. Reduced givs
3145 are updated as many times as their biv is. Mark it so if this is
3146 a splittable register. Don't need to do anything for address givs
3147 where this may not be a register. */
3149 if (GET_CODE (v
->new_reg
) == REG
)
3153 count
= reg_biv_class
[REGNO (v
->src_reg
)]->biv_count
;
3155 if (count
> 1 && v
->derived_from
)
3156 /* In this case, there is one set where the giv insn was and one
3157 set each after each biv increment. (Most are likely dead.) */
3160 splittable_regs_updates
[REGNO (v
->new_reg
)] = count
;
3165 if (loop_dump_stream
)
3169 if (GET_CODE (v
->dest_reg
) == CONST_INT
)
3171 else if (GET_CODE (v
->dest_reg
) != REG
)
3172 regnum
= REGNO (XEXP (v
->dest_reg
, 0));
3174 regnum
= REGNO (v
->dest_reg
);
3175 fprintf (loop_dump_stream
, "Giv %d at insn %d safe to split.\n",
3176 regnum
, INSN_UID (v
->insn
));
3183 /* Try to prove that the register is dead after the loop exits. Trace every
3184 loop exit looking for an insn that will always be executed, which sets
3185 the register to some value, and appears before the first use of the register
3186 is found. If successful, then return 1, otherwise return 0. */
3188 /* ?? Could be made more intelligent in the handling of jumps, so that
3189 it can search past if statements and other similar structures. */
3192 reg_dead_after_loop (reg
, loop_start
, loop_end
)
3193 rtx reg
, loop_start
, loop_end
;
3198 int label_count
= 0;
3199 int this_loop_num
= uid_loop_num
[INSN_UID (loop_start
)];
3201 /* In addition to checking all exits of this loop, we must also check
3202 all exits of inner nested loops that would exit this loop. We don't
3203 have any way to identify those, so we just give up if there are any
3204 such inner loop exits. */
3206 for (label
= loop_number_exit_labels
[this_loop_num
]; label
;
3207 label
= LABEL_NEXTREF (label
))
3210 if (label_count
!= loop_number_exit_count
[this_loop_num
])
3213 /* HACK: Must also search the loop fall through exit, create a label_ref
3214 here which points to the loop_end, and append the loop_number_exit_labels
3216 label
= gen_rtx_LABEL_REF (VOIDmode
, loop_end
);
3217 LABEL_NEXTREF (label
) = loop_number_exit_labels
[this_loop_num
];
3219 for ( ; label
; label
= LABEL_NEXTREF (label
))
3221 /* Succeed if find an insn which sets the biv or if reach end of
3222 function. Fail if find an insn that uses the biv, or if come to
3223 a conditional jump. */
3225 insn
= NEXT_INSN (XEXP (label
, 0));
3228 code
= GET_CODE (insn
);
3229 if (GET_RTX_CLASS (code
) == 'i')
3233 if (reg_referenced_p (reg
, PATTERN (insn
)))
3236 set
= single_set (insn
);
3237 if (set
&& rtx_equal_p (SET_DEST (set
), reg
))
3241 if (code
== JUMP_INSN
)
3243 if (GET_CODE (PATTERN (insn
)) == RETURN
)
3245 else if (! simplejump_p (insn
)
3246 /* Prevent infinite loop following infinite loops. */
3247 || jump_count
++ > 20)
3250 insn
= JUMP_LABEL (insn
);
3253 insn
= NEXT_INSN (insn
);
3257 /* Success, the register is dead on all loop exits. */
3261 /* Try to calculate the final value of the biv, the value it will have at
3262 the end of the loop. If we can do it, return that value. */
3265 final_biv_value (bl
, loop_start
, loop_end
, n_iterations
)
3266 struct iv_class
*bl
;
3267 rtx loop_start
, loop_end
;
3268 unsigned HOST_WIDE_INT n_iterations
;
3272 /* ??? This only works for MODE_INT biv's. Reject all others for now. */
3274 if (GET_MODE_CLASS (bl
->biv
->mode
) != MODE_INT
)
3277 /* The final value for reversed bivs must be calculated differently than
3278 for ordinary bivs. In this case, there is already an insn after the
3279 loop which sets this biv's final value (if necessary), and there are
3280 no other loop exits, so we can return any value. */
3283 if (loop_dump_stream
)
3284 fprintf (loop_dump_stream
,
3285 "Final biv value for %d, reversed biv.\n", bl
->regno
);
3290 /* Try to calculate the final value as initial value + (number of iterations
3291 * increment). For this to work, increment must be invariant, the only
3292 exit from the loop must be the fall through at the bottom (otherwise
3293 it may not have its final value when the loop exits), and the initial
3294 value of the biv must be invariant. */
3296 if (n_iterations
!= 0
3297 && ! loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]]
3298 && invariant_p (bl
->initial_value
))
3300 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
3302 if (increment
&& invariant_p (increment
))
3304 /* Can calculate the loop exit value, emit insns after loop
3305 end to calculate this value into a temporary register in
3306 case it is needed later. */
3308 tem
= gen_reg_rtx (bl
->biv
->mode
);
3309 record_base_value (REGNO (tem
), bl
->biv
->add_val
, 0);
3310 /* Make sure loop_end is not the last insn. */
3311 if (NEXT_INSN (loop_end
) == 0)
3312 emit_note_after (NOTE_INSN_DELETED
, loop_end
);
3313 emit_iv_add_mult (increment
, GEN_INT (n_iterations
),
3314 bl
->initial_value
, tem
, NEXT_INSN (loop_end
));
3316 if (loop_dump_stream
)
3317 fprintf (loop_dump_stream
,
3318 "Final biv value for %d, calculated.\n", bl
->regno
);
3324 /* Check to see if the biv is dead at all loop exits. */
3325 if (reg_dead_after_loop (bl
->biv
->src_reg
, loop_start
, loop_end
))
3327 if (loop_dump_stream
)
3328 fprintf (loop_dump_stream
,
3329 "Final biv value for %d, biv dead after loop exit.\n",
3338 /* Try to calculate the final value of the giv, the value it will have at
3339 the end of the loop. If we can do it, return that value. */
3342 final_giv_value (v
, loop_start
, loop_end
, n_iterations
)
3343 struct induction
*v
;
3344 rtx loop_start
, loop_end
;
3345 unsigned HOST_WIDE_INT n_iterations
;
3347 struct iv_class
*bl
;
3350 rtx insert_before
, seq
;
3352 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
3354 /* The final value for givs which depend on reversed bivs must be calculated
3355 differently than for ordinary givs. In this case, there is already an
3356 insn after the loop which sets this giv's final value (if necessary),
3357 and there are no other loop exits, so we can return any value. */
3360 if (loop_dump_stream
)
3361 fprintf (loop_dump_stream
,
3362 "Final giv value for %d, depends on reversed biv\n",
3363 REGNO (v
->dest_reg
));
3367 /* Try to calculate the final value as a function of the biv it depends
3368 upon. The only exit from the loop must be the fall through at the bottom
3369 (otherwise it may not have its final value when the loop exits). */
3371 /* ??? Can calculate the final giv value by subtracting off the
3372 extra biv increments times the giv's mult_val. The loop must have
3373 only one exit for this to work, but the loop iterations does not need
3376 if (n_iterations
!= 0
3377 && ! loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]])
3379 /* ?? It is tempting to use the biv's value here since these insns will
3380 be put after the loop, and hence the biv will have its final value
3381 then. However, this fails if the biv is subsequently eliminated.
3382 Perhaps determine whether biv's are eliminable before trying to
3383 determine whether giv's are replaceable so that we can use the
3384 biv value here if it is not eliminable. */
3386 /* We are emitting code after the end of the loop, so we must make
3387 sure that bl->initial_value is still valid then. It will still
3388 be valid if it is invariant. */
3390 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
3392 if (increment
&& invariant_p (increment
)
3393 && invariant_p (bl
->initial_value
))
3395 /* Can calculate the loop exit value of its biv as
3396 (n_iterations * increment) + initial_value */
3398 /* The loop exit value of the giv is then
3399 (final_biv_value - extra increments) * mult_val + add_val.
3400 The extra increments are any increments to the biv which
3401 occur in the loop after the giv's value is calculated.
3402 We must search from the insn that sets the giv to the end
3403 of the loop to calculate this value. */
3405 insert_before
= NEXT_INSN (loop_end
);
3407 /* Put the final biv value in tem. */
3408 tem
= gen_reg_rtx (bl
->biv
->mode
);
3409 record_base_value (REGNO (tem
), bl
->biv
->add_val
, 0);
3410 emit_iv_add_mult (increment
, GEN_INT (n_iterations
),
3411 bl
->initial_value
, tem
, insert_before
);
3413 /* Subtract off extra increments as we find them. */
3414 for (insn
= NEXT_INSN (v
->insn
); insn
!= loop_end
;
3415 insn
= NEXT_INSN (insn
))
3417 struct induction
*biv
;
3419 for (biv
= bl
->biv
; biv
; biv
= biv
->next_iv
)
3420 if (biv
->insn
== insn
)
3423 tem
= expand_binop (GET_MODE (tem
), sub_optab
, tem
,
3424 biv
->add_val
, NULL_RTX
, 0,
3426 seq
= gen_sequence ();
3428 emit_insn_before (seq
, insert_before
);
3432 /* Now calculate the giv's final value. */
3433 emit_iv_add_mult (tem
, v
->mult_val
, v
->add_val
, tem
,
3436 if (loop_dump_stream
)
3437 fprintf (loop_dump_stream
,
3438 "Final giv value for %d, calc from biv's value.\n",
3439 REGNO (v
->dest_reg
));
3445 /* Replaceable giv's should never reach here. */
3449 /* Check to see if the biv is dead at all loop exits. */
3450 if (reg_dead_after_loop (v
->dest_reg
, loop_start
, loop_end
))
3452 if (loop_dump_stream
)
3453 fprintf (loop_dump_stream
,
3454 "Final giv value for %d, giv dead after loop exit.\n",
3455 REGNO (v
->dest_reg
));
3464 /* Look back before LOOP_START for then insn that sets REG and return
3465 the equivalent constant if there is a REG_EQUAL note otherwise just
3466 the SET_SRC of REG. */
3469 loop_find_equiv_value (loop_start
, reg
)
3477 for (insn
= PREV_INSN (loop_start
); insn
; insn
= PREV_INSN (insn
))
3479 if (GET_CODE (insn
) == CODE_LABEL
)
3482 else if (GET_RTX_CLASS (GET_CODE (insn
)) == 'i'
3483 && reg_set_p (reg
, insn
))
3485 /* We found the last insn before the loop that sets the register.
3486 If it sets the entire register, and has a REG_EQUAL note,
3487 then use the value of the REG_EQUAL note. */
3488 if ((set
= single_set (insn
))
3489 && (SET_DEST (set
) == reg
))
3491 rtx note
= find_reg_note (insn
, REG_EQUAL
, NULL_RTX
);
3493 /* Only use the REG_EQUAL note if it is a constant.
3494 Other things, divide in particular, will cause
3495 problems later if we use them. */
3496 if (note
&& GET_CODE (XEXP (note
, 0)) != EXPR_LIST
3497 && CONSTANT_P (XEXP (note
, 0)))
3498 ret
= XEXP (note
, 0);
3500 ret
= SET_SRC (set
);
3509 /* Return a simplified rtx for the expression OP - REG.
3511 REG must appear in OP, and OP must be a register or the sum of a register
3514 Thus, the return value must be const0_rtx or the second term.
3516 The caller is responsible for verifying that REG appears in OP and OP has
3520 subtract_reg_term (op
, reg
)
3525 if (GET_CODE (op
) == PLUS
)
3527 if (XEXP (op
, 0) == reg
)
3528 return XEXP (op
, 1);
3529 else if (XEXP (op
, 1) == reg
)
3530 return XEXP (op
, 0);
3532 /* OP does not contain REG as a term. */
3537 /* Find and return register term common to both expressions OP0 and
3538 OP1 or NULL_RTX if no such term exists. Each expression must be a
3539 REG or a PLUS of a REG. */
3542 find_common_reg_term (op0
, op1
)
3545 if ((GET_CODE (op0
) == REG
|| GET_CODE (op0
) == PLUS
)
3546 && (GET_CODE (op1
) == REG
|| GET_CODE (op1
) == PLUS
))
3553 if (GET_CODE (op0
) == PLUS
)
3554 op01
= XEXP (op0
, 1), op00
= XEXP (op0
, 0);
3556 op01
= const0_rtx
, op00
= op0
;
3558 if (GET_CODE (op1
) == PLUS
)
3559 op11
= XEXP (op1
, 1), op10
= XEXP (op1
, 0);
3561 op11
= const0_rtx
, op10
= op1
;
3563 /* Find and return common register term if present. */
3564 if (REG_P (op00
) && (op00
== op10
|| op00
== op11
))
3566 else if (REG_P (op01
) && (op01
== op10
|| op01
== op11
))
3570 /* No common register term found. */
3575 /* Calculate the number of loop iterations. Returns the exact number of loop
3576 iterations if it can be calculated, otherwise returns zero. */
3578 unsigned HOST_WIDE_INT
3579 loop_iterations (loop_start
, loop_end
, loop_info
)
3580 rtx loop_start
, loop_end
;
3581 struct loop_info
*loop_info
;
3583 rtx comparison
, comparison_value
;
3584 rtx iteration_var
, initial_value
, increment
, final_value
;
3585 enum rtx_code comparison_code
;
3586 HOST_WIDE_INT abs_inc
;
3587 unsigned HOST_WIDE_INT abs_diff
;
3590 int unsigned_p
, compare_dir
, final_larger
;
3595 loop_info
->n_iterations
= 0;
3596 loop_info
->initial_value
= 0;
3597 loop_info
->initial_equiv_value
= 0;
3598 loop_info
->comparison_value
= 0;
3599 loop_info
->final_value
= 0;
3600 loop_info
->final_equiv_value
= 0;
3601 loop_info
->increment
= 0;
3602 loop_info
->iteration_var
= 0;
3603 loop_info
->unroll_number
= 1;
3604 loop_info
->vtop
= 0;
3606 /* We used to use prev_nonnote_insn here, but that fails because it might
3607 accidentally get the branch for a contained loop if the branch for this
3608 loop was deleted. We can only trust branches immediately before the
3610 last_loop_insn
= PREV_INSN (loop_end
);
3612 /* ??? We should probably try harder to find the jump insn
3613 at the end of the loop. The following code assumes that
3614 the last loop insn is a jump to the top of the loop. */
3615 if (GET_CODE (last_loop_insn
) != JUMP_INSN
)
3617 if (loop_dump_stream
)
3618 fprintf (loop_dump_stream
,
3619 "Loop iterations: No final conditional branch found.\n");
3623 /* If there is a more than a single jump to the top of the loop
3624 we cannot (easily) determine the iteration count. */
3625 if (LABEL_NUSES (JUMP_LABEL (last_loop_insn
)) > 1)
3627 if (loop_dump_stream
)
3628 fprintf (loop_dump_stream
,
3629 "Loop iterations: Loop has multiple back edges.\n");
3633 /* Find the iteration variable. If the last insn is a conditional
3634 branch, and the insn before tests a register value, make that the
3635 iteration variable. */
3637 comparison
= get_condition_for_loop (last_loop_insn
);
3638 if (comparison
== 0)
3640 if (loop_dump_stream
)
3641 fprintf (loop_dump_stream
,
3642 "Loop iterations: No final comparison found.\n");
3646 /* ??? Get_condition may switch position of induction variable and
3647 invariant register when it canonicalizes the comparison. */
3649 comparison_code
= GET_CODE (comparison
);
3650 iteration_var
= XEXP (comparison
, 0);
3651 comparison_value
= XEXP (comparison
, 1);
3653 /* Check if there is a NOTE_INSN_LOOP_VTOP note. If there is,
3654 that means that this is a for or while style loop, with
3655 a loop exit test at the start. Thus, we can assume that
3656 the loop condition was true when the loop was entered.
3658 We start at the end and search backwards for the previous
3659 NOTE. If there is no NOTE_INSN_LOOP_VTOP for this loop,
3660 the search will stop at the NOTE_INSN_LOOP_CONT. */
3663 vtop
= PREV_INSN (vtop
);
3664 while (GET_CODE (vtop
) != NOTE
3665 || NOTE_LINE_NUMBER (vtop
) > 0
3666 || NOTE_LINE_NUMBER (vtop
) == NOTE_REPEATED_LINE_NUMBER
3667 || NOTE_LINE_NUMBER (vtop
) == NOTE_INSN_DELETED
);
3668 if (NOTE_LINE_NUMBER (vtop
) != NOTE_INSN_LOOP_VTOP
)
3670 loop_info
->vtop
= vtop
;
3672 if (GET_CODE (iteration_var
) != REG
)
3674 if (loop_dump_stream
)
3675 fprintf (loop_dump_stream
,
3676 "Loop iterations: Comparison not against register.\n");
3680 /* The only new registers that care created before loop iterations are
3681 givs made from biv increments, so this should never occur. */
3683 if ((unsigned) REGNO (iteration_var
) >= reg_iv_type
->num_elements
)
3686 iteration_info (iteration_var
, &initial_value
, &increment
,
3687 loop_start
, loop_end
);
3688 if (initial_value
== 0)
3689 /* iteration_info already printed a message. */
3694 switch (comparison_code
)
3709 /* Cannot determine loop iterations with this case. */
3728 /* If the comparison value is an invariant register, then try to find
3729 its value from the insns before the start of the loop. */
3731 final_value
= comparison_value
;
3732 if (GET_CODE (comparison_value
) == REG
&& invariant_p (comparison_value
))
3734 final_value
= loop_find_equiv_value (loop_start
, comparison_value
);
3735 /* If we don't get an invariant final value, we are better
3736 off with the original register. */
3737 if (!invariant_p (final_value
))
3738 final_value
= comparison_value
;
3741 /* Calculate the approximate final value of the induction variable
3742 (on the last successful iteration). The exact final value
3743 depends on the branch operator, and increment sign. It will be
3744 wrong if the iteration variable is not incremented by one each
3745 time through the loop and (comparison_value + off_by_one -
3746 initial_value) % increment != 0.
3747 ??? Note that the final_value may overflow and thus final_larger
3748 will be bogus. A potentially infinite loop will be classified
3749 as immediate, e.g. for (i = 0x7ffffff0; i <= 0x7fffffff; i++) */
3751 final_value
= plus_constant (final_value
, off_by_one
);
3753 /* Save the calculated values describing this loop's bounds, in case
3754 precondition_loop_p will need them later. These values can not be
3755 recalculated inside precondition_loop_p because strength reduction
3756 optimizations may obscure the loop's structure.
3758 These values are only required by precondition_loop_p and insert_bct
3759 whenever the number of iterations cannot be computed at compile time.
3760 Only the difference between final_value and initial_value is
3761 important. Note that final_value is only approximate. */
3762 loop_info
->initial_value
= initial_value
;
3763 loop_info
->comparison_value
= comparison_value
;
3764 loop_info
->final_value
= plus_constant (comparison_value
, off_by_one
);
3765 loop_info
->increment
= increment
;
3766 loop_info
->iteration_var
= iteration_var
;
3767 loop_info
->comparison_code
= comparison_code
;
3769 /* Try to determine the iteration count for loops such
3770 as (for i = init; i < init + const; i++). When running the
3771 loop optimization twice, the first pass often converts simple
3772 loops into this form. */
3774 if (REG_P (initial_value
))
3780 reg1
= initial_value
;
3781 if (GET_CODE (final_value
) == PLUS
)
3782 reg2
= XEXP (final_value
, 0), const2
= XEXP (final_value
, 1);
3784 reg2
= final_value
, const2
= const0_rtx
;
3786 /* Check for initial_value = reg1, final_value = reg2 + const2,
3787 where reg1 != reg2. */
3788 if (REG_P (reg2
) && reg2
!= reg1
)
3792 /* Find what reg1 is equivalent to. Hopefully it will
3793 either be reg2 or reg2 plus a constant. */
3794 temp
= loop_find_equiv_value (loop_start
, reg1
);
3795 if (find_common_reg_term (temp
, reg2
))
3796 initial_value
= temp
;
3799 /* Find what reg2 is equivalent to. Hopefully it will
3800 either be reg1 or reg1 plus a constant. Let's ignore
3801 the latter case for now since it is not so common. */
3802 temp
= loop_find_equiv_value (loop_start
, reg2
);
3803 if (temp
== loop_info
->iteration_var
)
3804 temp
= initial_value
;
3806 final_value
= (const2
== const0_rtx
)
3807 ? reg1
: gen_rtx_PLUS (GET_MODE (reg1
), reg1
, const2
);
3810 else if (loop_info
->vtop
&& GET_CODE (reg2
) == CONST_INT
)
3814 /* When running the loop optimizer twice, check_dbra_loop
3815 further obfuscates reversible loops of the form:
3816 for (i = init; i < init + const; i++). We often end up with
3817 final_value = 0, initial_value = temp, temp = temp2 - init,
3818 where temp2 = init + const. If the loop has a vtop we
3819 can replace initial_value with const. */
3821 temp
= loop_find_equiv_value (loop_start
, reg1
);
3822 if (GET_CODE (temp
) == MINUS
&& REG_P (XEXP (temp
, 0)))
3824 rtx temp2
= loop_find_equiv_value (loop_start
, XEXP (temp
, 0));
3825 if (GET_CODE (temp2
) == PLUS
3826 && XEXP (temp2
, 0) == XEXP (temp
, 1))
3827 initial_value
= XEXP (temp2
, 1);
3832 /* If have initial_value = reg + const1 and final_value = reg +
3833 const2, then replace initial_value with const1 and final_value
3834 with const2. This should be safe since we are protected by the
3835 initial comparison before entering the loop if we have a vtop.
3836 For example, a + b < a + c is not equivalent to b < c for all a
3837 when using modulo arithmetic.
3839 ??? Without a vtop we could still perform the optimization if we check
3840 the initial and final values carefully. */
3842 && (reg_term
= find_common_reg_term (initial_value
, final_value
)))
3844 initial_value
= subtract_reg_term (initial_value
, reg_term
);
3845 final_value
= subtract_reg_term (final_value
, reg_term
);
3848 loop_info
->initial_equiv_value
= initial_value
;
3849 loop_info
->final_equiv_value
= final_value
;
3851 /* For EQ comparison loops, we don't have a valid final value.
3852 Check this now so that we won't leave an invalid value if we
3853 return early for any other reason. */
3854 if (comparison_code
== EQ
)
3855 loop_info
->final_equiv_value
= loop_info
->final_value
= 0;
3859 if (loop_dump_stream
)
3860 fprintf (loop_dump_stream
,
3861 "Loop iterations: Increment value can't be calculated.\n");
3865 if (GET_CODE (increment
) != CONST_INT
)
3867 /* If we have a REG, check to see if REG holds a constant value. */
3868 /* ??? Other RTL, such as (neg (reg)) is possible here, but it isn't
3869 clear if it is worthwhile to try to handle such RTL. */
3870 if (GET_CODE (increment
) == REG
|| GET_CODE (increment
) == SUBREG
)
3871 increment
= loop_find_equiv_value (loop_start
, increment
);
3873 if (GET_CODE (increment
) != CONST_INT
)
3875 if (loop_dump_stream
)
3877 fprintf (loop_dump_stream
,
3878 "Loop iterations: Increment value not constant ");
3879 print_rtl (loop_dump_stream
, increment
);
3880 fprintf (loop_dump_stream
, ".\n");
3884 loop_info
->increment
= increment
;
3887 if (GET_CODE (initial_value
) != CONST_INT
)
3889 if (loop_dump_stream
)
3891 fprintf (loop_dump_stream
,
3892 "Loop iterations: Initial value not constant ");
3893 print_rtl (loop_dump_stream
, initial_value
);
3894 fprintf (loop_dump_stream
, ".\n");
3898 else if (comparison_code
== EQ
)
3900 if (loop_dump_stream
)
3901 fprintf (loop_dump_stream
,
3902 "Loop iterations: EQ comparison loop.\n");
3905 else if (GET_CODE (final_value
) != CONST_INT
)
3907 if (loop_dump_stream
)
3909 fprintf (loop_dump_stream
,
3910 "Loop iterations: Final value not constant ");
3911 print_rtl (loop_dump_stream
, final_value
);
3912 fprintf (loop_dump_stream
, ".\n");
3917 /* Final_larger is 1 if final larger, 0 if they are equal, otherwise -1. */
3920 = ((unsigned HOST_WIDE_INT
) INTVAL (final_value
)
3921 > (unsigned HOST_WIDE_INT
) INTVAL (initial_value
))
3922 - ((unsigned HOST_WIDE_INT
) INTVAL (final_value
)
3923 < (unsigned HOST_WIDE_INT
) INTVAL (initial_value
));
3925 final_larger
= (INTVAL (final_value
) > INTVAL (initial_value
))
3926 - (INTVAL (final_value
) < INTVAL (initial_value
));
3928 if (INTVAL (increment
) > 0)
3930 else if (INTVAL (increment
) == 0)
3935 /* There are 27 different cases: compare_dir = -1, 0, 1;
3936 final_larger = -1, 0, 1; increment_dir = -1, 0, 1.
3937 There are 4 normal cases, 4 reverse cases (where the iteration variable
3938 will overflow before the loop exits), 4 infinite loop cases, and 15
3939 immediate exit (0 or 1 iteration depending on loop type) cases.
3940 Only try to optimize the normal cases. */
3942 /* (compare_dir/final_larger/increment_dir)
3943 Normal cases: (0/-1/-1), (0/1/1), (-1/-1/-1), (1/1/1)
3944 Reverse cases: (0/-1/1), (0/1/-1), (-1/-1/1), (1/1/-1)
3945 Infinite loops: (0/-1/0), (0/1/0), (-1/-1/0), (1/1/0)
3946 Immediate exit: (0/0/X), (-1/0/X), (-1/1/X), (1/0/X), (1/-1/X) */
3948 /* ?? If the meaning of reverse loops (where the iteration variable
3949 will overflow before the loop exits) is undefined, then could
3950 eliminate all of these special checks, and just always assume
3951 the loops are normal/immediate/infinite. Note that this means
3952 the sign of increment_dir does not have to be known. Also,
3953 since it does not really hurt if immediate exit loops or infinite loops
3954 are optimized, then that case could be ignored also, and hence all
3955 loops can be optimized.
3957 According to ANSI Spec, the reverse loop case result is undefined,
3958 because the action on overflow is undefined.
3960 See also the special test for NE loops below. */
3962 if (final_larger
== increment_dir
&& final_larger
!= 0
3963 && (final_larger
== compare_dir
|| compare_dir
== 0))
3968 if (loop_dump_stream
)
3969 fprintf (loop_dump_stream
,
3970 "Loop iterations: Not normal loop.\n");
3974 /* Calculate the number of iterations, final_value is only an approximation,
3975 so correct for that. Note that abs_diff and n_iterations are
3976 unsigned, because they can be as large as 2^n - 1. */
3978 abs_inc
= INTVAL (increment
);
3980 abs_diff
= INTVAL (final_value
) - INTVAL (initial_value
);
3981 else if (abs_inc
< 0)
3983 abs_diff
= INTVAL (initial_value
) - INTVAL (final_value
);
3989 /* For NE tests, make sure that the iteration variable won't miss
3990 the final value. If abs_diff mod abs_incr is not zero, then the
3991 iteration variable will overflow before the loop exits, and we
3992 can not calculate the number of iterations. */
3993 if (compare_dir
== 0 && (abs_diff
% abs_inc
) != 0)
3996 /* Note that the number of iterations could be calculated using
3997 (abs_diff + abs_inc - 1) / abs_inc, provided care was taken to
3998 handle potential overflow of the summation. */
3999 loop_info
->n_iterations
= abs_diff
/ abs_inc
+ ((abs_diff
% abs_inc
) != 0);
4000 return loop_info
->n_iterations
;
4004 /* Replace uses of split bivs with their split pseudo register. This is
4005 for original instructions which remain after loop unrolling without
4009 remap_split_bivs (x
)
4012 register enum rtx_code code
;
4019 code
= GET_CODE (x
);
4034 /* If non-reduced/final-value givs were split, then this would also
4035 have to remap those givs also. */
4037 if (REGNO (x
) < max_reg_before_loop
4038 && REG_IV_TYPE (REGNO (x
)) == BASIC_INDUCT
)
4039 return reg_biv_class
[REGNO (x
)]->biv
->src_reg
;
4046 fmt
= GET_RTX_FORMAT (code
);
4047 for (i
= GET_RTX_LENGTH (code
) - 1; i
>= 0; i
--)
4050 XEXP (x
, i
) = remap_split_bivs (XEXP (x
, i
));
4054 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
4055 XVECEXP (x
, i
, j
) = remap_split_bivs (XVECEXP (x
, i
, j
));
4061 /* If FIRST_UID is a set of REGNO, and FIRST_UID dominates LAST_UID (e.g.
4062 FIST_UID is always executed if LAST_UID is), then return 1. Otherwise
4063 return 0. COPY_START is where we can start looking for the insns
4064 FIRST_UID and LAST_UID. COPY_END is where we stop looking for these
4067 If there is no JUMP_INSN between LOOP_START and FIRST_UID, then FIRST_UID
4068 must dominate LAST_UID.
4070 If there is a CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID
4071 may not dominate LAST_UID.
4073 If there is no CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID
4074 must dominate LAST_UID. */
4077 set_dominates_use (regno
, first_uid
, last_uid
, copy_start
, copy_end
)
4084 int passed_jump
= 0;
4085 rtx p
= NEXT_INSN (copy_start
);
4087 while (INSN_UID (p
) != first_uid
)
4089 if (GET_CODE (p
) == JUMP_INSN
)
4091 /* Could not find FIRST_UID. */
4097 /* Verify that FIRST_UID is an insn that entirely sets REGNO. */
4098 if (GET_RTX_CLASS (GET_CODE (p
)) != 'i'
4099 || ! dead_or_set_regno_p (p
, regno
))
4102 /* FIRST_UID is always executed. */
4103 if (passed_jump
== 0)
4106 while (INSN_UID (p
) != last_uid
)
4108 /* If we see a CODE_LABEL between FIRST_UID and LAST_UID, then we
4109 can not be sure that FIRST_UID dominates LAST_UID. */
4110 if (GET_CODE (p
) == CODE_LABEL
)
4112 /* Could not find LAST_UID, but we reached the end of the loop, so
4114 else if (p
== copy_end
)
4119 /* FIRST_UID is always executed if LAST_UID is executed. */