1 /* Try to unroll loops, and split induction variables.
2 Copyright (C) 1992, 93, 94, 95, 97, 1998 Free Software Foundation, Inc.
3 Contributed by James E. Wilson, Cygnus Support/UC Berkeley.
5 This file is part of GNU CC.
7 GNU CC is free software; you can redistribute it and/or modify
8 it under the terms of the GNU General Public License as published by
9 the Free Software Foundation; either version 2, or (at your option)
12 GNU CC is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License for more details.
17 You should have received a copy of the GNU General Public License
18 along with GNU CC; see the file COPYING. If not, write to
19 the Free Software Foundation, 59 Temple Place - Suite 330,
20 Boston, MA 02111-1307, USA. */
22 /* Try to unroll a loop, and split induction variables.
24 Loops for which the number of iterations can be calculated exactly are
25 handled specially. If the number of iterations times the insn_count is
26 less than MAX_UNROLLED_INSNS, then the loop is unrolled completely.
27 Otherwise, we try to unroll the loop a number of times modulo the number
28 of iterations, so that only one exit test will be needed. It is unrolled
29 a number of times approximately equal to MAX_UNROLLED_INSNS divided by
32 Otherwise, if the number of iterations can be calculated exactly at
33 run time, and the loop is always entered at the top, then we try to
34 precondition the loop. That is, at run time, calculate how many times
35 the loop will execute, and then execute the loop body a few times so
36 that the remaining iterations will be some multiple of 4 (or 2 if the
37 loop is large). Then fall through to a loop unrolled 4 (or 2) times,
38 with only one exit test needed at the end of the loop.
40 Otherwise, if the number of iterations can not be calculated exactly,
41 not even at run time, then we still unroll the loop a number of times
42 approximately equal to MAX_UNROLLED_INSNS divided by the insn count,
43 but there must be an exit test after each copy of the loop body.
45 For each induction variable, which is dead outside the loop (replaceable)
46 or for which we can easily calculate the final value, if we can easily
47 calculate its value at each place where it is set as a function of the
48 current loop unroll count and the variable's value at loop entry, then
49 the induction variable is split into `N' different variables, one for
50 each copy of the loop body. One variable is live across the backward
51 branch, and the others are all calculated as a function of this variable.
52 This helps eliminate data dependencies, and leads to further opportunities
55 /* Possible improvements follow: */
57 /* ??? Add an extra pass somewhere to determine whether unrolling will
58 give any benefit. E.g. after generating all unrolled insns, compute the
59 cost of all insns and compare against cost of insns in rolled loop.
61 - On traditional architectures, unrolling a non-constant bound loop
62 is a win if there is a giv whose only use is in memory addresses, the
63 memory addresses can be split, and hence giv increments can be
65 - It is also a win if the loop is executed many times, and preconditioning
66 can be performed for the loop.
67 Add code to check for these and similar cases. */
69 /* ??? Improve control of which loops get unrolled. Could use profiling
70 info to only unroll the most commonly executed loops. Perhaps have
71 a user specifyable option to control the amount of code expansion,
72 or the percent of loops to consider for unrolling. Etc. */
74 /* ??? Look at the register copies inside the loop to see if they form a
75 simple permutation. If so, iterate the permutation until it gets back to
76 the start state. This is how many times we should unroll the loop, for
77 best results, because then all register copies can be eliminated.
78 For example, the lisp nreverse function should be unrolled 3 times
87 ??? The number of times to unroll the loop may also be based on data
88 references in the loop. For example, if we have a loop that references
89 x[i-1], x[i], and x[i+1], we should unroll it a multiple of 3 times. */
91 /* ??? Add some simple linear equation solving capability so that we can
92 determine the number of loop iterations for more complex loops.
93 For example, consider this loop from gdb
94 #define SWAP_TARGET_AND_HOST(buffer,len)
97 char *p = (char *) buffer;
98 char *q = ((char *) buffer) + len - 1;
99 int iterations = (len + 1) >> 1;
101 for (p; p < q; p++, q--;)
109 start value = p = &buffer + current_iteration
110 end value = q = &buffer + len - 1 - current_iteration
111 Given the loop exit test of "p < q", then there must be "q - p" iterations,
112 set equal to zero and solve for number of iterations:
113 q - p = len - 1 - 2*current_iteration = 0
114 current_iteration = (len - 1) / 2
115 Hence, there are (len - 1) / 2 (rounded up to the nearest integer)
116 iterations of this loop. */
118 /* ??? Currently, no labels are marked as loop invariant when doing loop
119 unrolling. This is because an insn inside the loop, that loads the address
120 of a label inside the loop into a register, could be moved outside the loop
121 by the invariant code motion pass if labels were invariant. If the loop
122 is subsequently unrolled, the code will be wrong because each unrolled
123 body of the loop will use the same address, whereas each actually needs a
124 different address. A case where this happens is when a loop containing
125 a switch statement is unrolled.
127 It would be better to let labels be considered invariant. When we
128 unroll loops here, check to see if any insns using a label local to the
129 loop were moved before the loop. If so, then correct the problem, by
130 moving the insn back into the loop, or perhaps replicate the insn before
131 the loop, one copy for each time the loop is unrolled. */
133 /* The prime factors looked for when trying to unroll a loop by some
134 number which is modulo the total number of iterations. Just checking
135 for these 4 prime factors will find at least one factor for 75% of
136 all numbers theoretically. Practically speaking, this will succeed
137 almost all of the time since loops are generally a multiple of 2
140 #define NUM_FACTORS 4
142 struct _factor
{ int factor
, count
; } factors
[NUM_FACTORS
]
143 = { {2, 0}, {3, 0}, {5, 0}, {7, 0}};
145 /* Describes the different types of loop unrolling performed. */
147 enum unroll_types
{ UNROLL_COMPLETELY
, UNROLL_MODULO
, UNROLL_NAIVE
};
152 #include "insn-config.h"
153 #include "integrate.h"
161 /* This controls which loops are unrolled, and by how much we unroll
164 #ifndef MAX_UNROLLED_INSNS
165 #define MAX_UNROLLED_INSNS 100
168 /* Indexed by register number, if non-zero, then it contains a pointer
169 to a struct induction for a DEST_REG giv which has been combined with
170 one of more address givs. This is needed because whenever such a DEST_REG
171 giv is modified, we must modify the value of all split address givs
172 that were combined with this DEST_REG giv. */
174 static struct induction
**addr_combined_regs
;
176 /* Indexed by register number, if this is a splittable induction variable,
177 then this will hold the current value of the register, which depends on the
180 static rtx
*splittable_regs
;
182 /* Indexed by register number, if this is a splittable induction variable,
183 then this will hold the number of instructions in the loop that modify
184 the induction variable. Used to ensure that only the last insn modifying
185 a split iv will update the original iv of the dest. */
187 static int *splittable_regs_updates
;
189 /* Values describing the current loop's iteration variable. These are set up
190 by loop_iterations, and used by precondition_loop_p. */
192 static rtx loop_iteration_var
;
193 static rtx loop_initial_value
;
194 static rtx loop_increment
;
195 static rtx loop_final_value
;
196 static enum rtx_code loop_comparison_code
;
198 /* Forward declarations. */
200 static void init_reg_map
PROTO((struct inline_remap
*, int));
201 static int precondition_loop_p
PROTO((rtx
*, rtx
*, rtx
*, rtx
, rtx
));
202 static rtx calculate_giv_inc
PROTO((rtx
, rtx
, int));
203 static rtx initial_reg_note_copy
PROTO((rtx
, struct inline_remap
*));
204 static void final_reg_note_copy
PROTO((rtx
, struct inline_remap
*));
205 static void copy_loop_body
PROTO((rtx
, rtx
, struct inline_remap
*, rtx
, int,
206 enum unroll_types
, rtx
, rtx
, rtx
, rtx
));
207 void iteration_info
PROTO((rtx
, rtx
*, rtx
*, rtx
, rtx
));
208 static rtx approx_final_value
PROTO((enum rtx_code
, rtx
, int *, int *));
209 static int find_splittable_regs
PROTO((enum unroll_types
, rtx
, rtx
, rtx
, int));
210 static int find_splittable_givs
PROTO((struct iv_class
*,enum unroll_types
,
211 rtx
, rtx
, rtx
, int));
212 static int reg_dead_after_loop
PROTO((rtx
, rtx
, rtx
));
213 static rtx fold_rtx_mult_add
PROTO((rtx
, rtx
, rtx
, enum machine_mode
));
214 static int verify_addresses
PROTO((struct induction
*, rtx
, int));
215 static rtx remap_split_bivs
PROTO((rtx
));
217 /* Try to unroll one loop and split induction variables in the loop.
219 The loop is described by the arguments LOOP_END, INSN_COUNT, and
220 LOOP_START. END_INSERT_BEFORE indicates where insns should be added
221 which need to be executed when the loop falls through. STRENGTH_REDUCTION_P
222 indicates whether information generated in the strength reduction pass
225 This function is intended to be called from within `strength_reduce'
229 unroll_loop (loop_end
, insn_count
, loop_start
, end_insert_before
,
234 rtx end_insert_before
;
235 int strength_reduce_p
;
238 int unroll_number
= 1;
239 rtx copy_start
, copy_end
;
240 rtx insn
, sequence
, pattern
, tem
;
241 int max_labelno
, max_insnno
;
243 struct inline_remap
*map
;
251 int splitting_not_safe
= 0;
252 enum unroll_types unroll_type
;
253 int loop_preconditioned
= 0;
255 /* This points to the last real insn in the loop, which should be either
256 a JUMP_INSN (for conditional jumps) or a BARRIER (for unconditional
260 /* Don't bother unrolling huge loops. Since the minimum factor is
261 two, loops greater than one half of MAX_UNROLLED_INSNS will never
263 if (insn_count
> MAX_UNROLLED_INSNS
/ 2)
265 if (loop_dump_stream
)
266 fprintf (loop_dump_stream
, "Unrolling failure: Loop too big.\n");
270 /* When emitting debugger info, we can't unroll loops with unequal numbers
271 of block_beg and block_end notes, because that would unbalance the block
272 structure of the function. This can happen as a result of the
273 "if (foo) bar; else break;" optimization in jump.c. */
274 /* ??? Gcc has a general policy that -g is never supposed to change the code
275 that the compiler emits, so we must disable this optimization always,
276 even if debug info is not being output. This is rare, so this should
277 not be a significant performance problem. */
279 if (1 /* write_symbols != NO_DEBUG */)
281 int block_begins
= 0;
284 for (insn
= loop_start
; insn
!= loop_end
; insn
= NEXT_INSN (insn
))
286 if (GET_CODE (insn
) == NOTE
)
288 if (NOTE_LINE_NUMBER (insn
) == NOTE_INSN_BLOCK_BEG
)
290 else if (NOTE_LINE_NUMBER (insn
) == NOTE_INSN_BLOCK_END
)
295 if (block_begins
!= block_ends
)
297 if (loop_dump_stream
)
298 fprintf (loop_dump_stream
,
299 "Unrolling failure: Unbalanced block notes.\n");
304 /* Determine type of unroll to perform. Depends on the number of iterations
305 and the size of the loop. */
307 /* If there is no strength reduce info, then set loop_n_iterations to zero.
308 This can happen if strength_reduce can't find any bivs in the loop.
309 A value of zero indicates that the number of iterations could not be
312 if (! strength_reduce_p
)
313 loop_n_iterations
= 0;
315 if (loop_dump_stream
&& loop_n_iterations
> 0)
317 fputs ("Loop unrolling: ", loop_dump_stream
);
318 fprintf (loop_dump_stream
, HOST_WIDE_INT_PRINT_DEC
, loop_n_iterations
);
319 fputs (" iterations.\n", loop_dump_stream
);
322 /* Find and save a pointer to the last nonnote insn in the loop. */
324 last_loop_insn
= prev_nonnote_insn (loop_end
);
326 /* Calculate how many times to unroll the loop. Indicate whether or
327 not the loop is being completely unrolled. */
329 if (loop_n_iterations
== 1)
331 /* If number of iterations is exactly 1, then eliminate the compare and
332 branch at the end of the loop since they will never be taken.
333 Then return, since no other action is needed here. */
335 /* If the last instruction is not a BARRIER or a JUMP_INSN, then
336 don't do anything. */
338 if (GET_CODE (last_loop_insn
) == BARRIER
)
340 /* Delete the jump insn. This will delete the barrier also. */
341 delete_insn (PREV_INSN (last_loop_insn
));
343 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
346 /* The immediately preceding insn is a compare which must be
348 delete_insn (last_loop_insn
);
349 delete_insn (PREV_INSN (last_loop_insn
));
351 /* The immediately preceding insn may not be the compare, so don't
353 delete_insn (last_loop_insn
);
358 else if (loop_n_iterations
> 0
359 && loop_n_iterations
* insn_count
< MAX_UNROLLED_INSNS
)
361 unroll_number
= loop_n_iterations
;
362 unroll_type
= UNROLL_COMPLETELY
;
364 else if (loop_n_iterations
> 0)
366 /* Try to factor the number of iterations. Don't bother with the
367 general case, only using 2, 3, 5, and 7 will get 75% of all
368 numbers theoretically, and almost all in practice. */
370 for (i
= 0; i
< NUM_FACTORS
; i
++)
371 factors
[i
].count
= 0;
373 temp
= loop_n_iterations
;
374 for (i
= NUM_FACTORS
- 1; i
>= 0; i
--)
375 while (temp
% factors
[i
].factor
== 0)
378 temp
= temp
/ factors
[i
].factor
;
381 /* Start with the larger factors first so that we generally
382 get lots of unrolling. */
386 for (i
= 3; i
>= 0; i
--)
387 while (factors
[i
].count
--)
389 if (temp
* factors
[i
].factor
< MAX_UNROLLED_INSNS
)
391 unroll_number
*= factors
[i
].factor
;
392 temp
*= factors
[i
].factor
;
398 /* If we couldn't find any factors, then unroll as in the normal
400 if (unroll_number
== 1)
402 if (loop_dump_stream
)
403 fprintf (loop_dump_stream
,
404 "Loop unrolling: No factors found.\n");
407 unroll_type
= UNROLL_MODULO
;
411 /* Default case, calculate number of times to unroll loop based on its
413 if (unroll_number
== 1)
415 if (8 * insn_count
< MAX_UNROLLED_INSNS
)
417 else if (4 * insn_count
< MAX_UNROLLED_INSNS
)
422 unroll_type
= UNROLL_NAIVE
;
425 /* Now we know how many times to unroll the loop. */
427 if (loop_dump_stream
)
428 fprintf (loop_dump_stream
,
429 "Unrolling loop %d times.\n", unroll_number
);
432 if (unroll_type
== UNROLL_COMPLETELY
|| unroll_type
== UNROLL_MODULO
)
434 /* Loops of these types should never start with a jump down to
435 the exit condition test. For now, check for this case just to
436 be sure. UNROLL_NAIVE loops can be of this form, this case is
439 while (GET_CODE (insn
) != CODE_LABEL
&& GET_CODE (insn
) != JUMP_INSN
)
440 insn
= NEXT_INSN (insn
);
441 if (GET_CODE (insn
) == JUMP_INSN
)
445 if (unroll_type
== UNROLL_COMPLETELY
)
447 /* Completely unrolling the loop: Delete the compare and branch at
448 the end (the last two instructions). This delete must done at the
449 very end of loop unrolling, to avoid problems with calls to
450 back_branch_in_range_p, which is called by find_splittable_regs.
451 All increments of splittable bivs/givs are changed to load constant
454 copy_start
= loop_start
;
456 /* Set insert_before to the instruction immediately after the JUMP_INSN
457 (or BARRIER), so that any NOTEs between the JUMP_INSN and the end of
458 the loop will be correctly handled by copy_loop_body. */
459 insert_before
= NEXT_INSN (last_loop_insn
);
461 /* Set copy_end to the insn before the jump at the end of the loop. */
462 if (GET_CODE (last_loop_insn
) == BARRIER
)
463 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
464 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
467 /* The instruction immediately before the JUMP_INSN is a compare
468 instruction which we do not want to copy. */
469 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
471 /* The instruction immediately before the JUMP_INSN may not be the
472 compare, so we must copy it. */
473 copy_end
= PREV_INSN (last_loop_insn
);
478 /* We currently can't unroll a loop if it doesn't end with a
479 JUMP_INSN. There would need to be a mechanism that recognizes
480 this case, and then inserts a jump after each loop body, which
481 jumps to after the last loop body. */
482 if (loop_dump_stream
)
483 fprintf (loop_dump_stream
,
484 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
488 else if (unroll_type
== UNROLL_MODULO
)
490 /* Partially unrolling the loop: The compare and branch at the end
491 (the last two instructions) must remain. Don't copy the compare
492 and branch instructions at the end of the loop. Insert the unrolled
493 code immediately before the compare/branch at the end so that the
494 code will fall through to them as before. */
496 copy_start
= loop_start
;
498 /* Set insert_before to the jump insn at the end of the loop.
499 Set copy_end to before the jump insn at the end of the loop. */
500 if (GET_CODE (last_loop_insn
) == BARRIER
)
502 insert_before
= PREV_INSN (last_loop_insn
);
503 copy_end
= PREV_INSN (insert_before
);
505 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
508 /* The instruction immediately before the JUMP_INSN is a compare
509 instruction which we do not want to copy or delete. */
510 insert_before
= PREV_INSN (last_loop_insn
);
511 copy_end
= PREV_INSN (insert_before
);
513 /* The instruction immediately before the JUMP_INSN may not be the
514 compare, so we must copy it. */
515 insert_before
= last_loop_insn
;
516 copy_end
= PREV_INSN (last_loop_insn
);
521 /* We currently can't unroll a loop if it doesn't end with a
522 JUMP_INSN. There would need to be a mechanism that recognizes
523 this case, and then inserts a jump after each loop body, which
524 jumps to after the last loop body. */
525 if (loop_dump_stream
)
526 fprintf (loop_dump_stream
,
527 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
533 /* Normal case: Must copy the compare and branch instructions at the
536 if (GET_CODE (last_loop_insn
) == BARRIER
)
538 /* Loop ends with an unconditional jump and a barrier.
539 Handle this like above, don't copy jump and barrier.
540 This is not strictly necessary, but doing so prevents generating
541 unconditional jumps to an immediately following label.
543 This will be corrected below if the target of this jump is
544 not the start_label. */
546 insert_before
= PREV_INSN (last_loop_insn
);
547 copy_end
= PREV_INSN (insert_before
);
549 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
551 /* Set insert_before to immediately after the JUMP_INSN, so that
552 NOTEs at the end of the loop will be correctly handled by
554 insert_before
= NEXT_INSN (last_loop_insn
);
555 copy_end
= last_loop_insn
;
559 /* We currently can't unroll a loop if it doesn't end with a
560 JUMP_INSN. There would need to be a mechanism that recognizes
561 this case, and then inserts a jump after each loop body, which
562 jumps to after the last loop body. */
563 if (loop_dump_stream
)
564 fprintf (loop_dump_stream
,
565 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
569 /* If copying exit test branches because they can not be eliminated,
570 then must convert the fall through case of the branch to a jump past
571 the end of the loop. Create a label to emit after the loop and save
572 it for later use. Do not use the label after the loop, if any, since
573 it might be used by insns outside the loop, or there might be insns
574 added before it later by final_[bg]iv_value which must be after
575 the real exit label. */
576 exit_label
= gen_label_rtx ();
579 while (GET_CODE (insn
) != CODE_LABEL
&& GET_CODE (insn
) != JUMP_INSN
)
580 insn
= NEXT_INSN (insn
);
582 if (GET_CODE (insn
) == JUMP_INSN
)
584 /* The loop starts with a jump down to the exit condition test.
585 Start copying the loop after the barrier following this
587 copy_start
= NEXT_INSN (insn
);
589 /* Splitting induction variables doesn't work when the loop is
590 entered via a jump to the bottom, because then we end up doing
591 a comparison against a new register for a split variable, but
592 we did not execute the set insn for the new register because
593 it was skipped over. */
594 splitting_not_safe
= 1;
595 if (loop_dump_stream
)
596 fprintf (loop_dump_stream
,
597 "Splitting not safe, because loop not entered at top.\n");
600 copy_start
= loop_start
;
603 /* This should always be the first label in the loop. */
604 start_label
= NEXT_INSN (copy_start
);
605 /* There may be a line number note and/or a loop continue note here. */
606 while (GET_CODE (start_label
) == NOTE
)
607 start_label
= NEXT_INSN (start_label
);
608 if (GET_CODE (start_label
) != CODE_LABEL
)
610 /* This can happen as a result of jump threading. If the first insns in
611 the loop test the same condition as the loop's backward jump, or the
612 opposite condition, then the backward jump will be modified to point
613 to elsewhere, and the loop's start label is deleted.
615 This case currently can not be handled by the loop unrolling code. */
617 if (loop_dump_stream
)
618 fprintf (loop_dump_stream
,
619 "Unrolling failure: unknown insns between BEG note and loop label.\n");
622 if (LABEL_NAME (start_label
))
624 /* The jump optimization pass must have combined the original start label
625 with a named label for a goto. We can't unroll this case because
626 jumps which go to the named label must be handled differently than
627 jumps to the loop start, and it is impossible to differentiate them
629 if (loop_dump_stream
)
630 fprintf (loop_dump_stream
,
631 "Unrolling failure: loop start label is gone\n");
635 if (unroll_type
== UNROLL_NAIVE
636 && GET_CODE (last_loop_insn
) == BARRIER
637 && start_label
!= JUMP_LABEL (PREV_INSN (last_loop_insn
)))
639 /* In this case, we must copy the jump and barrier, because they will
640 not be converted to jumps to an immediately following label. */
642 insert_before
= NEXT_INSN (last_loop_insn
);
643 copy_end
= last_loop_insn
;
646 if (unroll_type
== UNROLL_NAIVE
647 && GET_CODE (last_loop_insn
) == JUMP_INSN
648 && start_label
!= JUMP_LABEL (last_loop_insn
))
650 /* ??? The loop ends with a conditional branch that does not branch back
651 to the loop start label. In this case, we must emit an unconditional
652 branch to the loop exit after emitting the final branch.
653 copy_loop_body does not have support for this currently, so we
654 give up. It doesn't seem worthwhile to unroll anyways since
655 unrolling would increase the number of branch instructions
657 if (loop_dump_stream
)
658 fprintf (loop_dump_stream
,
659 "Unrolling failure: final conditional branch not to loop start\n");
663 /* Allocate a translation table for the labels and insn numbers.
664 They will be filled in as we copy the insns in the loop. */
666 max_labelno
= max_label_num ();
667 max_insnno
= get_max_uid ();
669 map
= (struct inline_remap
*) alloca (sizeof (struct inline_remap
));
671 map
->integrating
= 0;
673 /* Allocate the label map. */
677 map
->label_map
= (rtx
*) alloca (max_labelno
* sizeof (rtx
));
679 local_label
= (char *) alloca (max_labelno
);
680 bzero (local_label
, max_labelno
);
685 /* Search the loop and mark all local labels, i.e. the ones which have to
686 be distinct labels when copied. For all labels which might be
687 non-local, set their label_map entries to point to themselves.
688 If they happen to be local their label_map entries will be overwritten
689 before the loop body is copied. The label_map entries for local labels
690 will be set to a different value each time the loop body is copied. */
692 for (insn
= copy_start
; insn
!= loop_end
; insn
= NEXT_INSN (insn
))
696 if (GET_CODE (insn
) == CODE_LABEL
)
697 local_label
[CODE_LABEL_NUMBER (insn
)] = 1;
698 else if (GET_CODE (insn
) == JUMP_INSN
)
700 if (JUMP_LABEL (insn
))
701 set_label_in_map (map
,
702 CODE_LABEL_NUMBER (JUMP_LABEL (insn
)),
704 else if (GET_CODE (PATTERN (insn
)) == ADDR_VEC
705 || GET_CODE (PATTERN (insn
)) == ADDR_DIFF_VEC
)
707 rtx pat
= PATTERN (insn
);
708 int diff_vec_p
= GET_CODE (PATTERN (insn
)) == ADDR_DIFF_VEC
;
709 int len
= XVECLEN (pat
, diff_vec_p
);
712 for (i
= 0; i
< len
; i
++)
714 label
= XEXP (XVECEXP (pat
, diff_vec_p
, i
), 0);
715 set_label_in_map (map
,
716 CODE_LABEL_NUMBER (label
),
721 else if (note
= find_reg_note (insn
, REG_LABEL
, NULL_RTX
))
722 set_label_in_map (map
, CODE_LABEL_NUMBER (XEXP (note
, 0)),
726 /* Allocate space for the insn map. */
728 map
->insn_map
= (rtx
*) alloca (max_insnno
* sizeof (rtx
));
730 /* Set this to zero, to indicate that we are doing loop unrolling,
731 not function inlining. */
732 map
->inline_target
= 0;
734 /* The register and constant maps depend on the number of registers
735 present, so the final maps can't be created until after
736 find_splittable_regs is called. However, they are needed for
737 preconditioning, so we create temporary maps when preconditioning
740 /* The preconditioning code may allocate two new pseudo registers. */
741 maxregnum
= max_reg_num ();
743 /* Allocate and zero out the splittable_regs and addr_combined_regs
744 arrays. These must be zeroed here because they will be used if
745 loop preconditioning is performed, and must be zero for that case.
747 It is safe to do this here, since the extra registers created by the
748 preconditioning code and find_splittable_regs will never be used
749 to access the splittable_regs[] and addr_combined_regs[] arrays. */
751 splittable_regs
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
752 bzero ((char *) splittable_regs
, maxregnum
* sizeof (rtx
));
753 splittable_regs_updates
= (int *) alloca (maxregnum
* sizeof (int));
754 bzero ((char *) splittable_regs_updates
, maxregnum
* sizeof (int));
756 = (struct induction
**) alloca (maxregnum
* sizeof (struct induction
*));
757 bzero ((char *) addr_combined_regs
, maxregnum
* sizeof (struct induction
*));
758 /* We must limit it to max_reg_before_loop, because only these pseudo
759 registers have valid regno_first_uid info. Any register created after
760 that is unlikely to be local to the loop anyways. */
761 local_regno
= (char *) alloca (max_reg_before_loop
);
762 bzero (local_regno
, max_reg_before_loop
);
764 /* Mark all local registers, i.e. the ones which are referenced only
766 if (INSN_UID (copy_end
) < max_uid_for_loop
)
768 int copy_start_luid
= INSN_LUID (copy_start
);
769 int copy_end_luid
= INSN_LUID (copy_end
);
771 /* If a register is used in the jump insn, we must not duplicate it
772 since it will also be used outside the loop. */
773 if (GET_CODE (copy_end
) == JUMP_INSN
)
775 /* If copy_start points to the NOTE that starts the loop, then we must
776 use the next luid, because invariant pseudo-regs moved out of the loop
777 have their lifetimes modified to start here, but they are not safe
779 if (copy_start
== loop_start
)
782 /* If a pseudo's lifetime is entirely contained within this loop, then we
783 can use a different pseudo in each unrolled copy of the loop. This
784 results in better code. */
785 for (j
= FIRST_PSEUDO_REGISTER
; j
< max_reg_before_loop
; ++j
)
786 if (REGNO_FIRST_UID (j
) > 0 && REGNO_FIRST_UID (j
) <= max_uid_for_loop
787 && uid_luid
[REGNO_FIRST_UID (j
)] >= copy_start_luid
788 && REGNO_LAST_UID (j
) > 0 && REGNO_LAST_UID (j
) <= max_uid_for_loop
789 && uid_luid
[REGNO_LAST_UID (j
)] <= copy_end_luid
)
791 /* However, we must also check for loop-carried dependencies.
792 If the value the pseudo has at the end of iteration X is
793 used by iteration X+1, then we can not use a different pseudo
794 for each unrolled copy of the loop. */
795 /* A pseudo is safe if regno_first_uid is a set, and this
796 set dominates all instructions from regno_first_uid to
798 /* ??? This check is simplistic. We would get better code if
799 this check was more sophisticated. */
800 if (set_dominates_use (j
, REGNO_FIRST_UID (j
), REGNO_LAST_UID (j
),
801 copy_start
, copy_end
))
804 if (loop_dump_stream
)
807 fprintf (loop_dump_stream
, "Marked reg %d as local\n", j
);
809 fprintf (loop_dump_stream
, "Did not mark reg %d as local\n",
815 /* If this loop requires exit tests when unrolled, check to see if we
816 can precondition the loop so as to make the exit tests unnecessary.
817 Just like variable splitting, this is not safe if the loop is entered
818 via a jump to the bottom. Also, can not do this if no strength
819 reduce info, because precondition_loop_p uses this info. */
821 /* Must copy the loop body for preconditioning before the following
822 find_splittable_regs call since that will emit insns which need to
823 be after the preconditioned loop copies, but immediately before the
824 unrolled loop copies. */
826 /* Also, it is not safe to split induction variables for the preconditioned
827 copies of the loop body. If we split induction variables, then the code
828 assumes that each induction variable can be represented as a function
829 of its initial value and the loop iteration number. This is not true
830 in this case, because the last preconditioned copy of the loop body
831 could be any iteration from the first up to the `unroll_number-1'th,
832 depending on the initial value of the iteration variable. Therefore
833 we can not split induction variables here, because we can not calculate
834 their value. Hence, this code must occur before find_splittable_regs
837 if (unroll_type
== UNROLL_NAIVE
&& ! splitting_not_safe
&& strength_reduce_p
)
839 rtx initial_value
, final_value
, increment
;
841 if (precondition_loop_p (&initial_value
, &final_value
, &increment
,
842 loop_start
, loop_end
))
845 enum machine_mode mode
;
847 int abs_inc
, neg_inc
;
849 map
->reg_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
851 map
->const_equiv_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
852 map
->const_age_map
= (unsigned *) alloca (maxregnum
853 * sizeof (unsigned));
854 map
->const_equiv_map_size
= maxregnum
;
855 global_const_equiv_map
= map
->const_equiv_map
;
856 global_const_equiv_map_size
= maxregnum
;
858 init_reg_map (map
, maxregnum
);
860 /* Limit loop unrolling to 4, since this will make 7 copies of
862 if (unroll_number
> 4)
865 /* Save the absolute value of the increment, and also whether or
866 not it is negative. */
868 abs_inc
= INTVAL (increment
);
877 /* Decide what mode to do these calculations in. Choose the larger
878 of final_value's mode and initial_value's mode, or a full-word if
879 both are constants. */
880 mode
= GET_MODE (final_value
);
881 if (mode
== VOIDmode
)
883 mode
= GET_MODE (initial_value
);
884 if (mode
== VOIDmode
)
887 else if (mode
!= GET_MODE (initial_value
)
888 && (GET_MODE_SIZE (mode
)
889 < GET_MODE_SIZE (GET_MODE (initial_value
))))
890 mode
= GET_MODE (initial_value
);
892 /* Calculate the difference between the final and initial values.
893 Final value may be a (plus (reg x) (const_int 1)) rtx.
894 Let the following cse pass simplify this if initial value is
897 We must copy the final and initial values here to avoid
898 improperly shared rtl. */
900 diff
= expand_binop (mode
, sub_optab
, copy_rtx (final_value
),
901 copy_rtx (initial_value
), NULL_RTX
, 0,
904 /* Now calculate (diff % (unroll * abs (increment))) by using an
906 diff
= expand_binop (GET_MODE (diff
), and_optab
, diff
,
907 GEN_INT (unroll_number
* abs_inc
- 1),
908 NULL_RTX
, 0, OPTAB_LIB_WIDEN
);
910 /* Now emit a sequence of branches to jump to the proper precond
913 labels
= (rtx
*) alloca (sizeof (rtx
) * unroll_number
);
914 for (i
= 0; i
< unroll_number
; i
++)
915 labels
[i
] = gen_label_rtx ();
917 /* Check for the case where the initial value is greater than or
918 equal to the final value. In that case, we want to execute
919 exactly one loop iteration. The code below will fail for this
920 case. This check does not apply if the loop has a NE
921 comparison at the end. */
923 if (loop_comparison_code
!= NE
)
925 emit_cmp_insn (initial_value
, final_value
, neg_inc
? LE
: GE
,
926 NULL_RTX
, mode
, 0, 0);
928 emit_jump_insn (gen_ble (labels
[1]));
930 emit_jump_insn (gen_bge (labels
[1]));
931 JUMP_LABEL (get_last_insn ()) = labels
[1];
932 LABEL_NUSES (labels
[1])++;
935 /* Assuming the unroll_number is 4, and the increment is 2, then
936 for a negative increment: for a positive increment:
937 diff = 0,1 precond 0 diff = 0,7 precond 0
938 diff = 2,3 precond 3 diff = 1,2 precond 1
939 diff = 4,5 precond 2 diff = 3,4 precond 2
940 diff = 6,7 precond 1 diff = 5,6 precond 3 */
942 /* We only need to emit (unroll_number - 1) branches here, the
943 last case just falls through to the following code. */
945 /* ??? This would give better code if we emitted a tree of branches
946 instead of the current linear list of branches. */
948 for (i
= 0; i
< unroll_number
- 1; i
++)
951 enum rtx_code cmp_code
;
953 /* For negative increments, must invert the constant compared
954 against, except when comparing against zero. */
962 cmp_const
= unroll_number
- i
;
971 emit_cmp_insn (diff
, GEN_INT (abs_inc
* cmp_const
),
972 cmp_code
, NULL_RTX
, mode
, 0, 0);
975 emit_jump_insn (gen_beq (labels
[i
]));
977 emit_jump_insn (gen_bge (labels
[i
]));
979 emit_jump_insn (gen_ble (labels
[i
]));
980 JUMP_LABEL (get_last_insn ()) = labels
[i
];
981 LABEL_NUSES (labels
[i
])++;
984 /* If the increment is greater than one, then we need another branch,
985 to handle other cases equivalent to 0. */
987 /* ??? This should be merged into the code above somehow to help
988 simplify the code here, and reduce the number of branches emitted.
989 For the negative increment case, the branch here could easily
990 be merged with the `0' case branch above. For the positive
991 increment case, it is not clear how this can be simplified. */
996 enum rtx_code cmp_code
;
1000 cmp_const
= abs_inc
- 1;
1005 cmp_const
= abs_inc
* (unroll_number
- 1) + 1;
1009 emit_cmp_insn (diff
, GEN_INT (cmp_const
), cmp_code
, NULL_RTX
,
1013 emit_jump_insn (gen_ble (labels
[0]));
1015 emit_jump_insn (gen_bge (labels
[0]));
1016 JUMP_LABEL (get_last_insn ()) = labels
[0];
1017 LABEL_NUSES (labels
[0])++;
1020 sequence
= gen_sequence ();
1022 emit_insn_before (sequence
, loop_start
);
1024 /* Only the last copy of the loop body here needs the exit
1025 test, so set copy_end to exclude the compare/branch here,
1026 and then reset it inside the loop when get to the last
1029 if (GET_CODE (last_loop_insn
) == BARRIER
)
1030 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
1031 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
1034 /* The immediately preceding insn is a compare which we do not
1036 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
1038 /* The immediately preceding insn may not be a compare, so we
1040 copy_end
= PREV_INSN (last_loop_insn
);
1046 for (i
= 1; i
< unroll_number
; i
++)
1048 emit_label_after (labels
[unroll_number
- i
],
1049 PREV_INSN (loop_start
));
1051 bzero ((char *) map
->insn_map
, max_insnno
* sizeof (rtx
));
1052 bzero ((char *) map
->const_equiv_map
, maxregnum
* sizeof (rtx
));
1053 bzero ((char *) map
->const_age_map
,
1054 maxregnum
* sizeof (unsigned));
1057 for (j
= 0; j
< max_labelno
; j
++)
1059 set_label_in_map (map
, j
, gen_label_rtx ());
1061 for (j
= FIRST_PSEUDO_REGISTER
; j
< max_reg_before_loop
; j
++)
1064 map
->reg_map
[j
] = gen_reg_rtx (GET_MODE (regno_reg_rtx
[j
]));
1065 record_base_value (REGNO (map
->reg_map
[j
]),
1066 regno_reg_rtx
[j
], 0);
1068 /* The last copy needs the compare/branch insns at the end,
1069 so reset copy_end here if the loop ends with a conditional
1072 if (i
== unroll_number
- 1)
1074 if (GET_CODE (last_loop_insn
) == BARRIER
)
1075 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
1077 copy_end
= last_loop_insn
;
1080 /* None of the copies are the `last_iteration', so just
1081 pass zero for that parameter. */
1082 copy_loop_body (copy_start
, copy_end
, map
, exit_label
, 0,
1083 unroll_type
, start_label
, loop_end
,
1084 loop_start
, copy_end
);
1086 emit_label_after (labels
[0], PREV_INSN (loop_start
));
1088 if (GET_CODE (last_loop_insn
) == BARRIER
)
1090 insert_before
= PREV_INSN (last_loop_insn
);
1091 copy_end
= PREV_INSN (insert_before
);
1096 /* The immediately preceding insn is a compare which we do not
1098 insert_before
= PREV_INSN (last_loop_insn
);
1099 copy_end
= PREV_INSN (insert_before
);
1101 /* The immediately preceding insn may not be a compare, so we
1103 insert_before
= last_loop_insn
;
1104 copy_end
= PREV_INSN (last_loop_insn
);
1108 /* Set unroll type to MODULO now. */
1109 unroll_type
= UNROLL_MODULO
;
1110 loop_preconditioned
= 1;
1113 /* Fix the initial value for the loop as needed. */
1114 if (loop_n_iterations
<= 0)
1115 loop_start_value
[uid_loop_num
[INSN_UID (loop_start
)]]
1121 /* If reach here, and the loop type is UNROLL_NAIVE, then don't unroll
1122 the loop unless all loops are being unrolled. */
1123 if (unroll_type
== UNROLL_NAIVE
&& ! flag_unroll_all_loops
)
1125 if (loop_dump_stream
)
1126 fprintf (loop_dump_stream
, "Unrolling failure: Naive unrolling not being done.\n");
1130 /* At this point, we are guaranteed to unroll the loop. */
1132 /* Keep track of the unroll factor for each loop. */
1133 if (unroll_type
== UNROLL_COMPLETELY
)
1134 loop_unroll_factor
[uid_loop_num
[INSN_UID (loop_start
)]] = -1;
1136 loop_unroll_factor
[uid_loop_num
[INSN_UID (loop_start
)]] = unroll_number
;
1139 /* For each biv and giv, determine whether it can be safely split into
1140 a different variable for each unrolled copy of the loop body.
1141 We precalculate and save this info here, since computing it is
1144 Do this before deleting any instructions from the loop, so that
1145 back_branch_in_range_p will work correctly. */
1147 if (splitting_not_safe
)
1150 temp
= find_splittable_regs (unroll_type
, loop_start
, loop_end
,
1151 end_insert_before
, unroll_number
);
1153 /* find_splittable_regs may have created some new registers, so must
1154 reallocate the reg_map with the new larger size, and must realloc
1155 the constant maps also. */
1157 maxregnum
= max_reg_num ();
1158 map
->reg_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
1160 init_reg_map (map
, maxregnum
);
1162 /* Space is needed in some of the map for new registers, so new_maxregnum
1163 is an (over)estimate of how many registers will exist at the end. */
1164 new_maxregnum
= maxregnum
+ (temp
* unroll_number
* 2);
1166 /* Must realloc space for the constant maps, because the number of registers
1167 may have changed. */
1169 map
->const_equiv_map
= (rtx
*) alloca (new_maxregnum
* sizeof (rtx
));
1170 map
->const_age_map
= (unsigned *) alloca (new_maxregnum
* sizeof (unsigned));
1172 map
->const_equiv_map_size
= new_maxregnum
;
1173 global_const_equiv_map
= map
->const_equiv_map
;
1174 global_const_equiv_map_size
= new_maxregnum
;
1176 /* Search the list of bivs and givs to find ones which need to be remapped
1177 when split, and set their reg_map entry appropriately. */
1179 for (bl
= loop_iv_list
; bl
; bl
= bl
->next
)
1181 if (REGNO (bl
->biv
->src_reg
) != bl
->regno
)
1182 map
->reg_map
[bl
->regno
] = bl
->biv
->src_reg
;
1184 /* Currently, non-reduced/final-value givs are never split. */
1185 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
1186 if (REGNO (v
->src_reg
) != bl
->regno
)
1187 map
->reg_map
[REGNO (v
->dest_reg
)] = v
->src_reg
;
1191 /* Use our current register alignment and pointer flags. */
1192 map
->regno_pointer_flag
= regno_pointer_flag
;
1193 map
->regno_pointer_align
= regno_pointer_align
;
1195 /* If the loop is being partially unrolled, and the iteration variables
1196 are being split, and are being renamed for the split, then must fix up
1197 the compare/jump instruction at the end of the loop to refer to the new
1198 registers. This compare isn't copied, so the registers used in it
1199 will never be replaced if it isn't done here. */
1201 if (unroll_type
== UNROLL_MODULO
)
1203 insn
= NEXT_INSN (copy_end
);
1204 if (GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == JUMP_INSN
)
1205 PATTERN (insn
) = remap_split_bivs (PATTERN (insn
));
1208 /* For unroll_number - 1 times, make a copy of each instruction
1209 between copy_start and copy_end, and insert these new instructions
1210 before the end of the loop. */
1212 for (i
= 0; i
< unroll_number
; i
++)
1214 bzero ((char *) map
->insn_map
, max_insnno
* sizeof (rtx
));
1215 bzero ((char *) map
->const_equiv_map
, new_maxregnum
* sizeof (rtx
));
1216 bzero ((char *) map
->const_age_map
, new_maxregnum
* sizeof (unsigned));
1219 for (j
= 0; j
< max_labelno
; j
++)
1221 set_label_in_map (map
, j
, gen_label_rtx ());
1223 for (j
= FIRST_PSEUDO_REGISTER
; j
< max_reg_before_loop
; j
++)
1226 map
->reg_map
[j
] = gen_reg_rtx (GET_MODE (regno_reg_rtx
[j
]));
1227 record_base_value (REGNO (map
->reg_map
[j
]),
1228 regno_reg_rtx
[j
], 0);
1231 /* If loop starts with a branch to the test, then fix it so that
1232 it points to the test of the first unrolled copy of the loop. */
1233 if (i
== 0 && loop_start
!= copy_start
)
1235 insn
= PREV_INSN (copy_start
);
1236 pattern
= PATTERN (insn
);
1238 tem
= get_label_from_map (map
,
1240 (XEXP (SET_SRC (pattern
), 0)));
1241 SET_SRC (pattern
) = gen_rtx_LABEL_REF (VOIDmode
, tem
);
1243 /* Set the jump label so that it can be used by later loop unrolling
1245 JUMP_LABEL (insn
) = tem
;
1246 LABEL_NUSES (tem
)++;
1249 copy_loop_body (copy_start
, copy_end
, map
, exit_label
,
1250 i
== unroll_number
- 1, unroll_type
, start_label
,
1251 loop_end
, insert_before
, insert_before
);
1254 /* Before deleting any insns, emit a CODE_LABEL immediately after the last
1255 insn to be deleted. This prevents any runaway delete_insn call from
1256 more insns that it should, as it always stops at a CODE_LABEL. */
1258 /* Delete the compare and branch at the end of the loop if completely
1259 unrolling the loop. Deleting the backward branch at the end also
1260 deletes the code label at the start of the loop. This is done at
1261 the very end to avoid problems with back_branch_in_range_p. */
1263 if (unroll_type
== UNROLL_COMPLETELY
)
1264 safety_label
= emit_label_after (gen_label_rtx (), last_loop_insn
);
1266 safety_label
= emit_label_after (gen_label_rtx (), copy_end
);
1268 /* Delete all of the original loop instructions. Don't delete the
1269 LOOP_BEG note, or the first code label in the loop. */
1271 insn
= NEXT_INSN (copy_start
);
1272 while (insn
!= safety_label
)
1274 if (insn
!= start_label
)
1275 insn
= delete_insn (insn
);
1277 insn
= NEXT_INSN (insn
);
1280 /* Can now delete the 'safety' label emitted to protect us from runaway
1281 delete_insn calls. */
1282 if (INSN_DELETED_P (safety_label
))
1284 delete_insn (safety_label
);
1286 /* If exit_label exists, emit it after the loop. Doing the emit here
1287 forces it to have a higher INSN_UID than any insn in the unrolled loop.
1288 This is needed so that mostly_true_jump in reorg.c will treat jumps
1289 to this loop end label correctly, i.e. predict that they are usually
1292 emit_label_after (exit_label
, loop_end
);
1295 /* Return true if the loop can be safely, and profitably, preconditioned
1296 so that the unrolled copies of the loop body don't need exit tests.
1298 This only works if final_value, initial_value and increment can be
1299 determined, and if increment is a constant power of 2.
1300 If increment is not a power of 2, then the preconditioning modulo
1301 operation would require a real modulo instead of a boolean AND, and this
1302 is not considered `profitable'. */
1304 /* ??? If the loop is known to be executed very many times, or the machine
1305 has a very cheap divide instruction, then preconditioning is a win even
1306 when the increment is not a power of 2. Use RTX_COST to compute
1307 whether divide is cheap. */
1310 precondition_loop_p (initial_value
, final_value
, increment
, loop_start
,
1312 rtx
*initial_value
, *final_value
, *increment
;
1313 rtx loop_start
, loop_end
;
1316 if (loop_n_iterations
> 0)
1318 *initial_value
= const0_rtx
;
1319 *increment
= const1_rtx
;
1320 *final_value
= GEN_INT (loop_n_iterations
);
1322 if (loop_dump_stream
)
1324 fputs ("Preconditioning: Success, number of iterations known, ",
1326 fprintf (loop_dump_stream
, HOST_WIDE_INT_PRINT_DEC
,
1328 fputs (".\n", loop_dump_stream
);
1333 if (loop_initial_value
== 0)
1335 if (loop_dump_stream
)
1336 fprintf (loop_dump_stream
,
1337 "Preconditioning: Could not find initial value.\n");
1340 else if (loop_increment
== 0)
1342 if (loop_dump_stream
)
1343 fprintf (loop_dump_stream
,
1344 "Preconditioning: Could not find increment value.\n");
1347 else if (GET_CODE (loop_increment
) != CONST_INT
)
1349 if (loop_dump_stream
)
1350 fprintf (loop_dump_stream
,
1351 "Preconditioning: Increment not a constant.\n");
1354 else if ((exact_log2 (INTVAL (loop_increment
)) < 0)
1355 && (exact_log2 (- INTVAL (loop_increment
)) < 0))
1357 if (loop_dump_stream
)
1358 fprintf (loop_dump_stream
,
1359 "Preconditioning: Increment not a constant power of 2.\n");
1363 /* Unsigned_compare and compare_dir can be ignored here, since they do
1364 not matter for preconditioning. */
1366 if (loop_final_value
== 0)
1368 if (loop_dump_stream
)
1369 fprintf (loop_dump_stream
,
1370 "Preconditioning: EQ comparison loop.\n");
1374 /* Must ensure that final_value is invariant, so call invariant_p to
1375 check. Before doing so, must check regno against max_reg_before_loop
1376 to make sure that the register is in the range covered by invariant_p.
1377 If it isn't, then it is most likely a biv/giv which by definition are
1379 if ((GET_CODE (loop_final_value
) == REG
1380 && REGNO (loop_final_value
) >= max_reg_before_loop
)
1381 || (GET_CODE (loop_final_value
) == PLUS
1382 && REGNO (XEXP (loop_final_value
, 0)) >= max_reg_before_loop
)
1383 || ! invariant_p (loop_final_value
))
1385 if (loop_dump_stream
)
1386 fprintf (loop_dump_stream
,
1387 "Preconditioning: Final value not invariant.\n");
1391 /* Fail for floating point values, since the caller of this function
1392 does not have code to deal with them. */
1393 if (GET_MODE_CLASS (GET_MODE (loop_final_value
)) == MODE_FLOAT
1394 || GET_MODE_CLASS (GET_MODE (loop_initial_value
)) == MODE_FLOAT
)
1396 if (loop_dump_stream
)
1397 fprintf (loop_dump_stream
,
1398 "Preconditioning: Floating point final or initial value.\n");
1402 /* Now set initial_value to be the iteration_var, since that may be a
1403 simpler expression, and is guaranteed to be correct if all of the
1404 above tests succeed.
1406 We can not use the initial_value as calculated, because it will be
1407 one too small for loops of the form "while (i-- > 0)". We can not
1408 emit code before the loop_skip_over insns to fix this problem as this
1409 will then give a number one too large for loops of the form
1412 Note that all loops that reach here are entered at the top, because
1413 this function is not called if the loop starts with a jump. */
1415 /* Fail if loop_iteration_var is not live before loop_start, since we need
1416 to test its value in the preconditioning code. */
1418 if (uid_luid
[REGNO_FIRST_UID (REGNO (loop_iteration_var
))]
1419 > INSN_LUID (loop_start
))
1421 if (loop_dump_stream
)
1422 fprintf (loop_dump_stream
,
1423 "Preconditioning: Iteration var not live before loop start.\n");
1427 *initial_value
= loop_iteration_var
;
1428 *increment
= loop_increment
;
1429 *final_value
= loop_final_value
;
1432 if (loop_dump_stream
)
1433 fprintf (loop_dump_stream
, "Preconditioning: Successful.\n");
1438 /* All pseudo-registers must be mapped to themselves. Two hard registers
1439 must be mapped, VIRTUAL_STACK_VARS_REGNUM and VIRTUAL_INCOMING_ARGS_
1440 REGNUM, to avoid function-inlining specific conversions of these
1441 registers. All other hard regs can not be mapped because they may be
1446 init_reg_map (map
, maxregnum
)
1447 struct inline_remap
*map
;
1452 for (i
= maxregnum
- 1; i
> LAST_VIRTUAL_REGISTER
; i
--)
1453 map
->reg_map
[i
] = regno_reg_rtx
[i
];
1454 /* Just clear the rest of the entries. */
1455 for (i
= LAST_VIRTUAL_REGISTER
; i
>= 0; i
--)
1456 map
->reg_map
[i
] = 0;
1458 map
->reg_map
[VIRTUAL_STACK_VARS_REGNUM
]
1459 = regno_reg_rtx
[VIRTUAL_STACK_VARS_REGNUM
];
1460 map
->reg_map
[VIRTUAL_INCOMING_ARGS_REGNUM
]
1461 = regno_reg_rtx
[VIRTUAL_INCOMING_ARGS_REGNUM
];
1464 /* Strength-reduction will often emit code for optimized biv/givs which
1465 calculates their value in a temporary register, and then copies the result
1466 to the iv. This procedure reconstructs the pattern computing the iv;
1467 verifying that all operands are of the proper form.
1469 PATTERN must be the result of single_set.
1470 The return value is the amount that the giv is incremented by. */
1473 calculate_giv_inc (pattern
, src_insn
, regno
)
1474 rtx pattern
, src_insn
;
1478 rtx increment_total
= 0;
1482 /* Verify that we have an increment insn here. First check for a plus
1483 as the set source. */
1484 if (GET_CODE (SET_SRC (pattern
)) != PLUS
)
1486 /* SR sometimes computes the new giv value in a temp, then copies it
1488 src_insn
= PREV_INSN (src_insn
);
1489 pattern
= PATTERN (src_insn
);
1490 if (GET_CODE (SET_SRC (pattern
)) != PLUS
)
1493 /* The last insn emitted is not needed, so delete it to avoid confusing
1494 the second cse pass. This insn sets the giv unnecessarily. */
1495 delete_insn (get_last_insn ());
1498 /* Verify that we have a constant as the second operand of the plus. */
1499 increment
= XEXP (SET_SRC (pattern
), 1);
1500 if (GET_CODE (increment
) != CONST_INT
)
1502 /* SR sometimes puts the constant in a register, especially if it is
1503 too big to be an add immed operand. */
1504 src_insn
= PREV_INSN (src_insn
);
1505 increment
= SET_SRC (PATTERN (src_insn
));
1507 /* SR may have used LO_SUM to compute the constant if it is too large
1508 for a load immed operand. In this case, the constant is in operand
1509 one of the LO_SUM rtx. */
1510 if (GET_CODE (increment
) == LO_SUM
)
1511 increment
= XEXP (increment
, 1);
1513 /* Some ports store large constants in memory and add a REG_EQUAL
1514 note to the store insn. */
1515 else if (GET_CODE (increment
) == MEM
)
1517 rtx note
= find_reg_note (src_insn
, REG_EQUAL
, 0);
1519 increment
= XEXP (note
, 0);
1522 else if (GET_CODE (increment
) == IOR
1523 || GET_CODE (increment
) == ASHIFT
1524 || GET_CODE (increment
) == PLUS
)
1526 /* The rs6000 port loads some constants with IOR.
1527 The alpha port loads some constants with ASHIFT and PLUS. */
1528 rtx second_part
= XEXP (increment
, 1);
1529 enum rtx_code code
= GET_CODE (increment
);
1531 src_insn
= PREV_INSN (src_insn
);
1532 increment
= SET_SRC (PATTERN (src_insn
));
1533 /* Don't need the last insn anymore. */
1534 delete_insn (get_last_insn ());
1536 if (GET_CODE (second_part
) != CONST_INT
1537 || GET_CODE (increment
) != CONST_INT
)
1541 increment
= GEN_INT (INTVAL (increment
) | INTVAL (second_part
));
1542 else if (code
== PLUS
)
1543 increment
= GEN_INT (INTVAL (increment
) + INTVAL (second_part
));
1545 increment
= GEN_INT (INTVAL (increment
) << INTVAL (second_part
));
1548 if (GET_CODE (increment
) != CONST_INT
)
1551 /* The insn loading the constant into a register is no longer needed,
1553 delete_insn (get_last_insn ());
1556 if (increment_total
)
1557 increment_total
= GEN_INT (INTVAL (increment_total
) + INTVAL (increment
));
1559 increment_total
= increment
;
1561 /* Check that the source register is the same as the register we expected
1562 to see as the source. If not, something is seriously wrong. */
1563 if (GET_CODE (XEXP (SET_SRC (pattern
), 0)) != REG
1564 || REGNO (XEXP (SET_SRC (pattern
), 0)) != regno
)
1566 /* Some machines (e.g. the romp), may emit two add instructions for
1567 certain constants, so lets try looking for another add immediately
1568 before this one if we have only seen one add insn so far. */
1574 src_insn
= PREV_INSN (src_insn
);
1575 pattern
= PATTERN (src_insn
);
1577 delete_insn (get_last_insn ());
1585 return increment_total
;
1588 /* Copy REG_NOTES, except for insn references, because not all insn_map
1589 entries are valid yet. We do need to copy registers now though, because
1590 the reg_map entries can change during copying. */
1593 initial_reg_note_copy (notes
, map
)
1595 struct inline_remap
*map
;
1602 copy
= rtx_alloc (GET_CODE (notes
));
1603 PUT_MODE (copy
, GET_MODE (notes
));
1605 if (GET_CODE (notes
) == EXPR_LIST
)
1606 XEXP (copy
, 0) = copy_rtx_and_substitute (XEXP (notes
, 0), map
);
1607 else if (GET_CODE (notes
) == INSN_LIST
)
1608 /* Don't substitute for these yet. */
1609 XEXP (copy
, 0) = XEXP (notes
, 0);
1613 XEXP (copy
, 1) = initial_reg_note_copy (XEXP (notes
, 1), map
);
1618 /* Fixup insn references in copied REG_NOTES. */
1621 final_reg_note_copy (notes
, map
)
1623 struct inline_remap
*map
;
1627 for (note
= notes
; note
; note
= XEXP (note
, 1))
1628 if (GET_CODE (note
) == INSN_LIST
)
1629 XEXP (note
, 0) = map
->insn_map
[INSN_UID (XEXP (note
, 0))];
1632 /* Copy each instruction in the loop, substituting from map as appropriate.
1633 This is very similar to a loop in expand_inline_function. */
1636 copy_loop_body (copy_start
, copy_end
, map
, exit_label
, last_iteration
,
1637 unroll_type
, start_label
, loop_end
, insert_before
,
1639 rtx copy_start
, copy_end
;
1640 struct inline_remap
*map
;
1643 enum unroll_types unroll_type
;
1644 rtx start_label
, loop_end
, insert_before
, copy_notes_from
;
1648 int dest_reg_was_split
, i
;
1652 rtx final_label
= 0;
1653 rtx giv_inc
, giv_dest_reg
, giv_src_reg
;
1655 /* If this isn't the last iteration, then map any references to the
1656 start_label to final_label. Final label will then be emitted immediately
1657 after the end of this loop body if it was ever used.
1659 If this is the last iteration, then map references to the start_label
1661 if (! last_iteration
)
1663 final_label
= gen_label_rtx ();
1664 set_label_in_map (map
, CODE_LABEL_NUMBER (start_label
),
1668 set_label_in_map (map
, CODE_LABEL_NUMBER (start_label
), start_label
);
1675 insn
= NEXT_INSN (insn
);
1677 map
->orig_asm_operands_vector
= 0;
1679 switch (GET_CODE (insn
))
1682 pattern
= PATTERN (insn
);
1686 /* Check to see if this is a giv that has been combined with
1687 some split address givs. (Combined in the sense that
1688 `combine_givs' in loop.c has put two givs in the same register.)
1689 In this case, we must search all givs based on the same biv to
1690 find the address givs. Then split the address givs.
1691 Do this before splitting the giv, since that may map the
1692 SET_DEST to a new register. */
1694 if ((set
= single_set (insn
))
1695 && GET_CODE (SET_DEST (set
)) == REG
1696 && addr_combined_regs
[REGNO (SET_DEST (set
))])
1698 struct iv_class
*bl
;
1699 struct induction
*v
, *tv
;
1700 int regno
= REGNO (SET_DEST (set
));
1702 v
= addr_combined_regs
[REGNO (SET_DEST (set
))];
1703 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
1705 /* Although the giv_inc amount is not needed here, we must call
1706 calculate_giv_inc here since it might try to delete the
1707 last insn emitted. If we wait until later to call it,
1708 we might accidentally delete insns generated immediately
1709 below by emit_unrolled_add. */
1711 giv_inc
= calculate_giv_inc (set
, insn
, regno
);
1713 /* Now find all address giv's that were combined with this
1715 for (tv
= bl
->giv
; tv
; tv
= tv
->next_iv
)
1716 if (tv
->giv_type
== DEST_ADDR
&& tv
->same
== v
)
1720 /* If this DEST_ADDR giv was not split, then ignore it. */
1721 if (*tv
->location
!= tv
->dest_reg
)
1724 /* Scale this_giv_inc if the multiplicative factors of
1725 the two givs are different. */
1726 this_giv_inc
= INTVAL (giv_inc
);
1727 if (tv
->mult_val
!= v
->mult_val
)
1728 this_giv_inc
= (this_giv_inc
/ INTVAL (v
->mult_val
)
1729 * INTVAL (tv
->mult_val
));
1731 tv
->dest_reg
= plus_constant (tv
->dest_reg
, this_giv_inc
);
1732 *tv
->location
= tv
->dest_reg
;
1734 if (last_iteration
&& unroll_type
!= UNROLL_COMPLETELY
)
1736 /* Must emit an insn to increment the split address
1737 giv. Add in the const_adjust field in case there
1738 was a constant eliminated from the address. */
1739 rtx value
, dest_reg
;
1741 /* tv->dest_reg will be either a bare register,
1742 or else a register plus a constant. */
1743 if (GET_CODE (tv
->dest_reg
) == REG
)
1744 dest_reg
= tv
->dest_reg
;
1746 dest_reg
= XEXP (tv
->dest_reg
, 0);
1748 /* Check for shared address givs, and avoid
1749 incrementing the shared pseudo reg more than
1751 if (! tv
->same_insn
&& ! tv
->shared
)
1753 /* tv->dest_reg may actually be a (PLUS (REG)
1754 (CONST)) here, so we must call plus_constant
1755 to add the const_adjust amount before calling
1756 emit_unrolled_add below. */
1757 value
= plus_constant (tv
->dest_reg
,
1760 /* The constant could be too large for an add
1761 immediate, so can't directly emit an insn
1763 emit_unrolled_add (dest_reg
, XEXP (value
, 0),
1767 /* Reset the giv to be just the register again, in case
1768 it is used after the set we have just emitted.
1769 We must subtract the const_adjust factor added in
1771 tv
->dest_reg
= plus_constant (dest_reg
,
1772 - tv
->const_adjust
);
1773 *tv
->location
= tv
->dest_reg
;
1778 /* If this is a setting of a splittable variable, then determine
1779 how to split the variable, create a new set based on this split,
1780 and set up the reg_map so that later uses of the variable will
1781 use the new split variable. */
1783 dest_reg_was_split
= 0;
1785 if ((set
= single_set (insn
))
1786 && GET_CODE (SET_DEST (set
)) == REG
1787 && splittable_regs
[REGNO (SET_DEST (set
))])
1789 int regno
= REGNO (SET_DEST (set
));
1791 dest_reg_was_split
= 1;
1793 /* Compute the increment value for the giv, if it wasn't
1794 already computed above. */
1797 giv_inc
= calculate_giv_inc (set
, insn
, regno
);
1798 giv_dest_reg
= SET_DEST (set
);
1799 giv_src_reg
= SET_DEST (set
);
1801 if (unroll_type
== UNROLL_COMPLETELY
)
1803 /* Completely unrolling the loop. Set the induction
1804 variable to a known constant value. */
1806 /* The value in splittable_regs may be an invariant
1807 value, so we must use plus_constant here. */
1808 splittable_regs
[regno
]
1809 = plus_constant (splittable_regs
[regno
], INTVAL (giv_inc
));
1811 if (GET_CODE (splittable_regs
[regno
]) == PLUS
)
1813 giv_src_reg
= XEXP (splittable_regs
[regno
], 0);
1814 giv_inc
= XEXP (splittable_regs
[regno
], 1);
1818 /* The splittable_regs value must be a REG or a
1819 CONST_INT, so put the entire value in the giv_src_reg
1821 giv_src_reg
= splittable_regs
[regno
];
1822 giv_inc
= const0_rtx
;
1827 /* Partially unrolling loop. Create a new pseudo
1828 register for the iteration variable, and set it to
1829 be a constant plus the original register. Except
1830 on the last iteration, when the result has to
1831 go back into the original iteration var register. */
1833 /* Handle bivs which must be mapped to a new register
1834 when split. This happens for bivs which need their
1835 final value set before loop entry. The new register
1836 for the biv was stored in the biv's first struct
1837 induction entry by find_splittable_regs. */
1839 if (regno
< max_reg_before_loop
1840 && reg_iv_type
[regno
] == BASIC_INDUCT
)
1842 giv_src_reg
= reg_biv_class
[regno
]->biv
->src_reg
;
1843 giv_dest_reg
= giv_src_reg
;
1847 /* If non-reduced/final-value givs were split, then
1848 this would have to remap those givs also. See
1849 find_splittable_regs. */
1852 splittable_regs
[regno
]
1853 = GEN_INT (INTVAL (giv_inc
)
1854 + INTVAL (splittable_regs
[regno
]));
1855 giv_inc
= splittable_regs
[regno
];
1857 /* Now split the induction variable by changing the dest
1858 of this insn to a new register, and setting its
1859 reg_map entry to point to this new register.
1861 If this is the last iteration, and this is the last insn
1862 that will update the iv, then reuse the original dest,
1863 to ensure that the iv will have the proper value when
1864 the loop exits or repeats.
1866 Using splittable_regs_updates here like this is safe,
1867 because it can only be greater than one if all
1868 instructions modifying the iv are always executed in
1871 if (! last_iteration
1872 || (splittable_regs_updates
[regno
]-- != 1))
1874 tem
= gen_reg_rtx (GET_MODE (giv_src_reg
));
1876 map
->reg_map
[regno
] = tem
;
1877 record_base_value (REGNO (tem
),
1878 giv_inc
== const0_rtx
1880 : gen_rtx_PLUS (GET_MODE (giv_src_reg
),
1881 giv_src_reg
, giv_inc
),
1885 map
->reg_map
[regno
] = giv_src_reg
;
1888 /* The constant being added could be too large for an add
1889 immediate, so can't directly emit an insn here. */
1890 emit_unrolled_add (giv_dest_reg
, giv_src_reg
, giv_inc
);
1891 copy
= get_last_insn ();
1892 pattern
= PATTERN (copy
);
1896 pattern
= copy_rtx_and_substitute (pattern
, map
);
1897 copy
= emit_insn (pattern
);
1899 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
1902 /* If this insn is setting CC0, it may need to look at
1903 the insn that uses CC0 to see what type of insn it is.
1904 In that case, the call to recog via validate_change will
1905 fail. So don't substitute constants here. Instead,
1906 do it when we emit the following insn.
1908 For example, see the pyr.md file. That machine has signed and
1909 unsigned compares. The compare patterns must check the
1910 following branch insn to see which what kind of compare to
1913 If the previous insn set CC0, substitute constants on it as
1915 if (sets_cc0_p (PATTERN (copy
)) != 0)
1920 try_constants (cc0_insn
, map
);
1922 try_constants (copy
, map
);
1925 try_constants (copy
, map
);
1928 /* Make split induction variable constants `permanent' since we
1929 know there are no backward branches across iteration variable
1930 settings which would invalidate this. */
1931 if (dest_reg_was_split
)
1933 int regno
= REGNO (SET_DEST (pattern
));
1935 if (regno
< map
->const_equiv_map_size
1936 && map
->const_age_map
[regno
] == map
->const_age
)
1937 map
->const_age_map
[regno
] = -1;
1942 pattern
= copy_rtx_and_substitute (PATTERN (insn
), map
);
1943 copy
= emit_jump_insn (pattern
);
1944 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
1946 if (JUMP_LABEL (insn
) == start_label
&& insn
== copy_end
1947 && ! last_iteration
)
1949 /* This is a branch to the beginning of the loop; this is the
1950 last insn being copied; and this is not the last iteration.
1951 In this case, we want to change the original fall through
1952 case to be a branch past the end of the loop, and the
1953 original jump label case to fall_through. */
1955 if (invert_exp (pattern
, copy
))
1957 if (! redirect_exp (&pattern
,
1958 get_label_from_map (map
,
1960 (JUMP_LABEL (insn
))),
1967 rtx lab
= gen_label_rtx ();
1968 /* Can't do it by reversing the jump (probably because we
1969 couldn't reverse the conditions), so emit a new
1970 jump_insn after COPY, and redirect the jump around
1972 jmp
= emit_jump_insn_after (gen_jump (exit_label
), copy
);
1973 jmp
= emit_barrier_after (jmp
);
1974 emit_label_after (lab
, jmp
);
1975 LABEL_NUSES (lab
) = 0;
1976 if (! redirect_exp (&pattern
,
1977 get_label_from_map (map
,
1979 (JUMP_LABEL (insn
))),
1987 try_constants (cc0_insn
, map
);
1990 try_constants (copy
, map
);
1992 /* Set the jump label of COPY correctly to avoid problems with
1993 later passes of unroll_loop, if INSN had jump label set. */
1994 if (JUMP_LABEL (insn
))
1998 /* Can't use the label_map for every insn, since this may be
1999 the backward branch, and hence the label was not mapped. */
2000 if ((set
= single_set (copy
)))
2002 tem
= SET_SRC (set
);
2003 if (GET_CODE (tem
) == LABEL_REF
)
2004 label
= XEXP (tem
, 0);
2005 else if (GET_CODE (tem
) == IF_THEN_ELSE
)
2007 if (XEXP (tem
, 1) != pc_rtx
)
2008 label
= XEXP (XEXP (tem
, 1), 0);
2010 label
= XEXP (XEXP (tem
, 2), 0);
2014 if (label
&& GET_CODE (label
) == CODE_LABEL
)
2015 JUMP_LABEL (copy
) = label
;
2018 /* An unrecognizable jump insn, probably the entry jump
2019 for a switch statement. This label must have been mapped,
2020 so just use the label_map to get the new jump label. */
2022 = get_label_from_map (map
,
2023 CODE_LABEL_NUMBER (JUMP_LABEL (insn
)));
2026 /* If this is a non-local jump, then must increase the label
2027 use count so that the label will not be deleted when the
2028 original jump is deleted. */
2029 LABEL_NUSES (JUMP_LABEL (copy
))++;
2031 else if (GET_CODE (PATTERN (copy
)) == ADDR_VEC
2032 || GET_CODE (PATTERN (copy
)) == ADDR_DIFF_VEC
)
2034 rtx pat
= PATTERN (copy
);
2035 int diff_vec_p
= GET_CODE (pat
) == ADDR_DIFF_VEC
;
2036 int len
= XVECLEN (pat
, diff_vec_p
);
2039 for (i
= 0; i
< len
; i
++)
2040 LABEL_NUSES (XEXP (XVECEXP (pat
, diff_vec_p
, i
), 0))++;
2043 /* If this used to be a conditional jump insn but whose branch
2044 direction is now known, we must do something special. */
2045 if (condjump_p (insn
) && !simplejump_p (insn
) && map
->last_pc_value
)
2048 /* The previous insn set cc0 for us. So delete it. */
2049 delete_insn (PREV_INSN (copy
));
2052 /* If this is now a no-op, delete it. */
2053 if (map
->last_pc_value
== pc_rtx
)
2055 /* Don't let delete_insn delete the label referenced here,
2056 because we might possibly need it later for some other
2057 instruction in the loop. */
2058 if (JUMP_LABEL (copy
))
2059 LABEL_NUSES (JUMP_LABEL (copy
))++;
2061 if (JUMP_LABEL (copy
))
2062 LABEL_NUSES (JUMP_LABEL (copy
))--;
2066 /* Otherwise, this is unconditional jump so we must put a
2067 BARRIER after it. We could do some dead code elimination
2068 here, but jump.c will do it just as well. */
2074 pattern
= copy_rtx_and_substitute (PATTERN (insn
), map
);
2075 copy
= emit_call_insn (pattern
);
2076 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
2078 /* Because the USAGE information potentially contains objects other
2079 than hard registers, we need to copy it. */
2080 CALL_INSN_FUNCTION_USAGE (copy
)
2081 = copy_rtx_and_substitute (CALL_INSN_FUNCTION_USAGE (insn
), map
);
2085 try_constants (cc0_insn
, map
);
2088 try_constants (copy
, map
);
2090 /* Be lazy and assume CALL_INSNs clobber all hard registers. */
2091 for (i
= 0; i
< FIRST_PSEUDO_REGISTER
; i
++)
2092 map
->const_equiv_map
[i
] = 0;
2096 /* If this is the loop start label, then we don't need to emit a
2097 copy of this label since no one will use it. */
2099 if (insn
!= start_label
)
2101 copy
= emit_label (get_label_from_map (map
,
2102 CODE_LABEL_NUMBER (insn
)));
2108 copy
= emit_barrier ();
2112 /* VTOP notes are valid only before the loop exit test. If placed
2113 anywhere else, loop may generate bad code. */
2115 if (NOTE_LINE_NUMBER (insn
) != NOTE_INSN_DELETED
2116 && (NOTE_LINE_NUMBER (insn
) != NOTE_INSN_LOOP_VTOP
2117 || (last_iteration
&& unroll_type
!= UNROLL_COMPLETELY
)))
2118 copy
= emit_note (NOTE_SOURCE_FILE (insn
),
2119 NOTE_LINE_NUMBER (insn
));
2129 map
->insn_map
[INSN_UID (insn
)] = copy
;
2131 while (insn
!= copy_end
);
2133 /* Now finish coping the REG_NOTES. */
2137 insn
= NEXT_INSN (insn
);
2138 if ((GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == JUMP_INSN
2139 || GET_CODE (insn
) == CALL_INSN
)
2140 && map
->insn_map
[INSN_UID (insn
)])
2141 final_reg_note_copy (REG_NOTES (map
->insn_map
[INSN_UID (insn
)]), map
);
2143 while (insn
!= copy_end
);
2145 /* There may be notes between copy_notes_from and loop_end. Emit a copy of
2146 each of these notes here, since there may be some important ones, such as
2147 NOTE_INSN_BLOCK_END notes, in this group. We don't do this on the last
2148 iteration, because the original notes won't be deleted.
2150 We can't use insert_before here, because when from preconditioning,
2151 insert_before points before the loop. We can't use copy_end, because
2152 there may be insns already inserted after it (which we don't want to
2153 copy) when not from preconditioning code. */
2155 if (! last_iteration
)
2157 for (insn
= copy_notes_from
; insn
!= loop_end
; insn
= NEXT_INSN (insn
))
2159 if (GET_CODE (insn
) == NOTE
2160 && NOTE_LINE_NUMBER (insn
) != NOTE_INSN_DELETED
)
2161 emit_note (NOTE_SOURCE_FILE (insn
), NOTE_LINE_NUMBER (insn
));
2165 if (final_label
&& LABEL_NUSES (final_label
) > 0)
2166 emit_label (final_label
);
2168 tem
= gen_sequence ();
2170 emit_insn_before (tem
, insert_before
);
2173 /* Emit an insn, using the expand_binop to ensure that a valid insn is
2174 emitted. This will correctly handle the case where the increment value
2175 won't fit in the immediate field of a PLUS insns. */
2178 emit_unrolled_add (dest_reg
, src_reg
, increment
)
2179 rtx dest_reg
, src_reg
, increment
;
2183 result
= expand_binop (GET_MODE (dest_reg
), add_optab
, src_reg
, increment
,
2184 dest_reg
, 0, OPTAB_LIB_WIDEN
);
2186 if (dest_reg
!= result
)
2187 emit_move_insn (dest_reg
, result
);
2190 /* Searches the insns between INSN and LOOP_END. Returns 1 if there
2191 is a backward branch in that range that branches to somewhere between
2192 LOOP_START and INSN. Returns 0 otherwise. */
2194 /* ??? This is quadratic algorithm. Could be rewritten to be linear.
2195 In practice, this is not a problem, because this function is seldom called,
2196 and uses a negligible amount of CPU time on average. */
2199 back_branch_in_range_p (insn
, loop_start
, loop_end
)
2201 rtx loop_start
, loop_end
;
2203 rtx p
, q
, target_insn
;
2204 rtx orig_loop_end
= loop_end
;
2206 /* Stop before we get to the backward branch at the end of the loop. */
2207 loop_end
= prev_nonnote_insn (loop_end
);
2208 if (GET_CODE (loop_end
) == BARRIER
)
2209 loop_end
= PREV_INSN (loop_end
);
2211 /* Check in case insn has been deleted, search forward for first non
2212 deleted insn following it. */
2213 while (INSN_DELETED_P (insn
))
2214 insn
= NEXT_INSN (insn
);
2216 /* Check for the case where insn is the last insn in the loop. Deal
2217 with the case where INSN was a deleted loop test insn, in which case
2218 it will now be the NOTE_LOOP_END. */
2219 if (insn
== loop_end
|| insn
== orig_loop_end
)
2222 for (p
= NEXT_INSN (insn
); p
!= loop_end
; p
= NEXT_INSN (p
))
2224 if (GET_CODE (p
) == JUMP_INSN
)
2226 target_insn
= JUMP_LABEL (p
);
2228 /* Search from loop_start to insn, to see if one of them is
2229 the target_insn. We can't use INSN_LUID comparisons here,
2230 since insn may not have an LUID entry. */
2231 for (q
= loop_start
; q
!= insn
; q
= NEXT_INSN (q
))
2232 if (q
== target_insn
)
2240 /* Try to generate the simplest rtx for the expression
2241 (PLUS (MULT mult1 mult2) add1). This is used to calculate the initial
2245 fold_rtx_mult_add (mult1
, mult2
, add1
, mode
)
2246 rtx mult1
, mult2
, add1
;
2247 enum machine_mode mode
;
2252 /* The modes must all be the same. This should always be true. For now,
2253 check to make sure. */
2254 if ((GET_MODE (mult1
) != mode
&& GET_MODE (mult1
) != VOIDmode
)
2255 || (GET_MODE (mult2
) != mode
&& GET_MODE (mult2
) != VOIDmode
)
2256 || (GET_MODE (add1
) != mode
&& GET_MODE (add1
) != VOIDmode
))
2259 /* Ensure that if at least one of mult1/mult2 are constant, then mult2
2260 will be a constant. */
2261 if (GET_CODE (mult1
) == CONST_INT
)
2268 mult_res
= simplify_binary_operation (MULT
, mode
, mult1
, mult2
);
2270 mult_res
= gen_rtx_MULT (mode
, mult1
, mult2
);
2272 /* Again, put the constant second. */
2273 if (GET_CODE (add1
) == CONST_INT
)
2280 result
= simplify_binary_operation (PLUS
, mode
, add1
, mult_res
);
2282 result
= gen_rtx_PLUS (mode
, add1
, mult_res
);
2287 /* Searches the list of induction struct's for the biv BL, to try to calculate
2288 the total increment value for one iteration of the loop as a constant.
2290 Returns the increment value as an rtx, simplified as much as possible,
2291 if it can be calculated. Otherwise, returns 0. */
2294 biv_total_increment (bl
, loop_start
, loop_end
)
2295 struct iv_class
*bl
;
2296 rtx loop_start
, loop_end
;
2298 struct induction
*v
;
2301 /* For increment, must check every instruction that sets it. Each
2302 instruction must be executed only once each time through the loop.
2303 To verify this, we check that the insn is always executed, and that
2304 there are no backward branches after the insn that branch to before it.
2305 Also, the insn must have a mult_val of one (to make sure it really is
2308 result
= const0_rtx
;
2309 for (v
= bl
->biv
; v
; v
= v
->next_iv
)
2311 if (v
->always_computable
&& v
->mult_val
== const1_rtx
2312 && ! back_branch_in_range_p (v
->insn
, loop_start
, loop_end
))
2313 result
= fold_rtx_mult_add (result
, const1_rtx
, v
->add_val
, v
->mode
);
2321 /* Determine the initial value of the iteration variable, and the amount
2322 that it is incremented each loop. Use the tables constructed by
2323 the strength reduction pass to calculate these values.
2325 Initial_value and/or increment are set to zero if their values could not
2329 iteration_info (iteration_var
, initial_value
, increment
, loop_start
, loop_end
)
2330 rtx iteration_var
, *initial_value
, *increment
;
2331 rtx loop_start
, loop_end
;
2333 struct iv_class
*bl
;
2335 struct induction
*v
;
2338 /* Clear the result values, in case no answer can be found. */
2342 /* The iteration variable can be either a giv or a biv. Check to see
2343 which it is, and compute the variable's initial value, and increment
2344 value if possible. */
2346 /* If this is a new register, can't handle it since we don't have any
2347 reg_iv_type entry for it. */
2348 if (REGNO (iteration_var
) >= max_reg_before_loop
)
2350 if (loop_dump_stream
)
2351 fprintf (loop_dump_stream
,
2352 "Loop unrolling: No reg_iv_type entry for iteration var.\n");
2356 /* Reject iteration variables larger than the host wide int size, since they
2357 could result in a number of iterations greater than the range of our
2358 `unsigned HOST_WIDE_INT' variable loop_n_iterations. */
2359 else if ((GET_MODE_BITSIZE (GET_MODE (iteration_var
))
2360 > HOST_BITS_PER_WIDE_INT
))
2362 if (loop_dump_stream
)
2363 fprintf (loop_dump_stream
,
2364 "Loop unrolling: Iteration var rejected because mode too large.\n");
2367 else if (GET_MODE_CLASS (GET_MODE (iteration_var
)) != MODE_INT
)
2369 if (loop_dump_stream
)
2370 fprintf (loop_dump_stream
,
2371 "Loop unrolling: Iteration var not an integer.\n");
2374 else if (reg_iv_type
[REGNO (iteration_var
)] == BASIC_INDUCT
)
2376 /* Grab initial value, only useful if it is a constant. */
2377 bl
= reg_biv_class
[REGNO (iteration_var
)];
2378 *initial_value
= bl
->initial_value
;
2380 *increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2382 else if (reg_iv_type
[REGNO (iteration_var
)] == GENERAL_INDUCT
)
2385 /* ??? The code below does not work because the incorrect number of
2386 iterations is calculated when the biv is incremented after the giv
2387 is set (which is the usual case). This can probably be accounted
2388 for by biasing the initial_value by subtracting the amount of the
2389 increment that occurs between the giv set and the giv test. However,
2390 a giv as an iterator is very rare, so it does not seem worthwhile
2392 /* ??? An example failure is: i = 6; do {;} while (i++ < 9). */
2393 if (loop_dump_stream
)
2394 fprintf (loop_dump_stream
,
2395 "Loop unrolling: Giv iterators are not handled.\n");
2398 /* Initial value is mult_val times the biv's initial value plus
2399 add_val. Only useful if it is a constant. */
2400 v
= reg_iv_info
[REGNO (iteration_var
)];
2401 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
2402 *initial_value
= fold_rtx_mult_add (v
->mult_val
, bl
->initial_value
,
2403 v
->add_val
, v
->mode
);
2405 /* Increment value is mult_val times the increment value of the biv. */
2407 *increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2409 *increment
= fold_rtx_mult_add (v
->mult_val
, *increment
, const0_rtx
,
2415 if (loop_dump_stream
)
2416 fprintf (loop_dump_stream
,
2417 "Loop unrolling: Not basic or general induction var.\n");
2422 /* Calculate the approximate final value of the iteration variable
2423 which has an loop exit test with code COMPARISON_CODE and comparison value
2424 of COMPARISON_VALUE. Also returns an indication of whether the comparison
2425 was signed or unsigned, and the direction of the comparison. This info is
2426 needed to calculate the number of loop iterations. */
2429 approx_final_value (comparison_code
, comparison_value
, unsigned_p
, compare_dir
)
2430 enum rtx_code comparison_code
;
2431 rtx comparison_value
;
2435 /* Calculate the final value of the induction variable.
2436 The exact final value depends on the branch operator, and increment sign.
2437 This is only an approximate value. It will be wrong if the iteration
2438 variable is not incremented by one each time through the loop, and
2439 approx final value - start value % increment != 0. */
2442 switch (comparison_code
)
2448 return plus_constant (comparison_value
, 1);
2453 return plus_constant (comparison_value
, -1);
2455 /* Can not calculate a final value for this case. */
2462 return comparison_value
;
2468 return comparison_value
;
2471 return comparison_value
;
2477 /* For each biv and giv, determine whether it can be safely split into
2478 a different variable for each unrolled copy of the loop body. If it
2479 is safe to split, then indicate that by saving some useful info
2480 in the splittable_regs array.
2482 If the loop is being completely unrolled, then splittable_regs will hold
2483 the current value of the induction variable while the loop is unrolled.
2484 It must be set to the initial value of the induction variable here.
2485 Otherwise, splittable_regs will hold the difference between the current
2486 value of the induction variable and the value the induction variable had
2487 at the top of the loop. It must be set to the value 0 here.
2489 Returns the total number of instructions that set registers that are
2492 /* ?? If the loop is only unrolled twice, then most of the restrictions to
2493 constant values are unnecessary, since we can easily calculate increment
2494 values in this case even if nothing is constant. The increment value
2495 should not involve a multiply however. */
2497 /* ?? Even if the biv/giv increment values aren't constant, it may still
2498 be beneficial to split the variable if the loop is only unrolled a few
2499 times, since multiplies by small integers (1,2,3,4) are very cheap. */
2502 find_splittable_regs (unroll_type
, loop_start
, loop_end
, end_insert_before
,
2504 enum unroll_types unroll_type
;
2505 rtx loop_start
, loop_end
;
2506 rtx end_insert_before
;
2509 struct iv_class
*bl
;
2510 struct induction
*v
;
2512 rtx biv_final_value
;
2516 for (bl
= loop_iv_list
; bl
; bl
= bl
->next
)
2518 /* Biv_total_increment must return a constant value,
2519 otherwise we can not calculate the split values. */
2521 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2522 if (! increment
|| GET_CODE (increment
) != CONST_INT
)
2525 /* The loop must be unrolled completely, or else have a known number
2526 of iterations and only one exit, or else the biv must be dead
2527 outside the loop, or else the final value must be known. Otherwise,
2528 it is unsafe to split the biv since it may not have the proper
2529 value on loop exit. */
2531 /* loop_number_exit_count is non-zero if the loop has an exit other than
2532 a fall through at the end. */
2535 biv_final_value
= 0;
2536 if (unroll_type
!= UNROLL_COMPLETELY
2537 && (loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]]
2538 || unroll_type
== UNROLL_NAIVE
)
2539 && (uid_luid
[REGNO_LAST_UID (bl
->regno
)] >= INSN_LUID (loop_end
)
2541 || INSN_UID (bl
->init_insn
) >= max_uid_for_loop
2542 || (uid_luid
[REGNO_FIRST_UID (bl
->regno
)]
2543 < INSN_LUID (bl
->init_insn
))
2544 || reg_mentioned_p (bl
->biv
->dest_reg
, SET_SRC (bl
->init_set
)))
2545 && ! (biv_final_value
= final_biv_value (bl
, loop_start
, loop_end
)))
2548 /* If any of the insns setting the BIV don't do so with a simple
2549 PLUS, we don't know how to split it. */
2550 for (v
= bl
->biv
; biv_splittable
&& v
; v
= v
->next_iv
)
2551 if ((tem
= single_set (v
->insn
)) == 0
2552 || GET_CODE (SET_DEST (tem
)) != REG
2553 || REGNO (SET_DEST (tem
)) != bl
->regno
2554 || GET_CODE (SET_SRC (tem
)) != PLUS
)
2557 /* If final value is non-zero, then must emit an instruction which sets
2558 the value of the biv to the proper value. This is done after
2559 handling all of the givs, since some of them may need to use the
2560 biv's value in their initialization code. */
2562 /* This biv is splittable. If completely unrolling the loop, save
2563 the biv's initial value. Otherwise, save the constant zero. */
2565 if (biv_splittable
== 1)
2567 if (unroll_type
== UNROLL_COMPLETELY
)
2569 /* If the initial value of the biv is itself (i.e. it is too
2570 complicated for strength_reduce to compute), or is a hard
2571 register, or it isn't invariant, then we must create a new
2572 pseudo reg to hold the initial value of the biv. */
2574 if (GET_CODE (bl
->initial_value
) == REG
2575 && (REGNO (bl
->initial_value
) == bl
->regno
2576 || REGNO (bl
->initial_value
) < FIRST_PSEUDO_REGISTER
2577 || ! invariant_p (bl
->initial_value
)))
2579 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2581 record_base_value (REGNO (tem
), bl
->biv
->add_val
, 0);
2582 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2585 if (loop_dump_stream
)
2586 fprintf (loop_dump_stream
, "Biv %d initial value remapped to %d.\n",
2587 bl
->regno
, REGNO (tem
));
2589 splittable_regs
[bl
->regno
] = tem
;
2592 splittable_regs
[bl
->regno
] = bl
->initial_value
;
2595 splittable_regs
[bl
->regno
] = const0_rtx
;
2597 /* Save the number of instructions that modify the biv, so that
2598 we can treat the last one specially. */
2600 splittable_regs_updates
[bl
->regno
] = bl
->biv_count
;
2601 result
+= bl
->biv_count
;
2603 if (loop_dump_stream
)
2604 fprintf (loop_dump_stream
,
2605 "Biv %d safe to split.\n", bl
->regno
);
2608 /* Check every giv that depends on this biv to see whether it is
2609 splittable also. Even if the biv isn't splittable, givs which
2610 depend on it may be splittable if the biv is live outside the
2611 loop, and the givs aren't. */
2613 result
+= find_splittable_givs (bl
, unroll_type
, loop_start
, loop_end
,
2614 increment
, unroll_number
);
2616 /* If final value is non-zero, then must emit an instruction which sets
2617 the value of the biv to the proper value. This is done after
2618 handling all of the givs, since some of them may need to use the
2619 biv's value in their initialization code. */
2620 if (biv_final_value
)
2622 /* If the loop has multiple exits, emit the insns before the
2623 loop to ensure that it will always be executed no matter
2624 how the loop exits. Otherwise emit the insn after the loop,
2625 since this is slightly more efficient. */
2626 if (! loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]])
2627 emit_insn_before (gen_move_insn (bl
->biv
->src_reg
,
2632 /* Create a new register to hold the value of the biv, and then
2633 set the biv to its final value before the loop start. The biv
2634 is set to its final value before loop start to ensure that
2635 this insn will always be executed, no matter how the loop
2637 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2638 record_base_value (REGNO (tem
), bl
->biv
->add_val
, 0);
2640 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2642 emit_insn_before (gen_move_insn (bl
->biv
->src_reg
,
2646 if (loop_dump_stream
)
2647 fprintf (loop_dump_stream
, "Biv %d mapped to %d for split.\n",
2648 REGNO (bl
->biv
->src_reg
), REGNO (tem
));
2650 /* Set up the mapping from the original biv register to the new
2652 bl
->biv
->src_reg
= tem
;
2659 /* Return 1 if the first and last unrolled copy of the address giv V is valid
2660 for the instruction that is using it. Do not make any changes to that
2664 verify_addresses (v
, giv_inc
, unroll_number
)
2665 struct induction
*v
;
2670 rtx orig_addr
= *v
->location
;
2671 rtx last_addr
= plus_constant (v
->dest_reg
,
2672 INTVAL (giv_inc
) * (unroll_number
- 1));
2674 /* First check to see if either address would fail. Handle the fact
2675 that we have may have a match_dup. */
2676 if (! validate_replace_rtx (*v
->location
, v
->dest_reg
, v
->insn
)
2677 || ! validate_replace_rtx (*v
->location
, last_addr
, v
->insn
))
2680 /* Now put things back the way they were before. This should always
2682 if (! validate_replace_rtx (*v
->location
, orig_addr
, v
->insn
))
2688 /* For every giv based on the biv BL, check to determine whether it is
2689 splittable. This is a subroutine to find_splittable_regs ().
2691 Return the number of instructions that set splittable registers. */
2694 find_splittable_givs (bl
, unroll_type
, loop_start
, loop_end
, increment
,
2696 struct iv_class
*bl
;
2697 enum unroll_types unroll_type
;
2698 rtx loop_start
, loop_end
;
2702 struct induction
*v
, *v2
;
2707 /* Scan the list of givs, and set the same_insn field when there are
2708 multiple identical givs in the same insn. */
2709 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
2710 for (v2
= v
->next_iv
; v2
; v2
= v2
->next_iv
)
2711 if (v
->insn
== v2
->insn
&& rtx_equal_p (v
->new_reg
, v2
->new_reg
)
2715 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
2719 /* Only split the giv if it has already been reduced, or if the loop is
2720 being completely unrolled. */
2721 if (unroll_type
!= UNROLL_COMPLETELY
&& v
->ignore
)
2724 /* The giv can be split if the insn that sets the giv is executed once
2725 and only once on every iteration of the loop. */
2726 /* An address giv can always be split. v->insn is just a use not a set,
2727 and hence it does not matter whether it is always executed. All that
2728 matters is that all the biv increments are always executed, and we
2729 won't reach here if they aren't. */
2730 if (v
->giv_type
!= DEST_ADDR
2731 && (! v
->always_computable
2732 || back_branch_in_range_p (v
->insn
, loop_start
, loop_end
)))
2735 /* The giv increment value must be a constant. */
2736 giv_inc
= fold_rtx_mult_add (v
->mult_val
, increment
, const0_rtx
,
2738 if (! giv_inc
|| GET_CODE (giv_inc
) != CONST_INT
)
2741 /* The loop must be unrolled completely, or else have a known number of
2742 iterations and only one exit, or else the giv must be dead outside
2743 the loop, or else the final value of the giv must be known.
2744 Otherwise, it is not safe to split the giv since it may not have the
2745 proper value on loop exit. */
2747 /* The used outside loop test will fail for DEST_ADDR givs. They are
2748 never used outside the loop anyways, so it is always safe to split a
2752 if (unroll_type
!= UNROLL_COMPLETELY
2753 && (loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]]
2754 || unroll_type
== UNROLL_NAIVE
)
2755 && v
->giv_type
!= DEST_ADDR
2756 /* The next part is true if the pseudo is used outside the loop.
2757 We assume that this is true for any pseudo created after loop
2758 starts, because we don't have a reg_n_info entry for them. */
2759 && (REGNO (v
->dest_reg
) >= max_reg_before_loop
2760 || (REGNO_FIRST_UID (REGNO (v
->dest_reg
)) != INSN_UID (v
->insn
)
2761 /* Check for the case where the pseudo is set by a shift/add
2762 sequence, in which case the first insn setting the pseudo
2763 is the first insn of the shift/add sequence. */
2764 && (! (tem
= find_reg_note (v
->insn
, REG_RETVAL
, NULL_RTX
))
2765 || (REGNO_FIRST_UID (REGNO (v
->dest_reg
))
2766 != INSN_UID (XEXP (tem
, 0)))))
2767 /* Line above always fails if INSN was moved by loop opt. */
2768 || (uid_luid
[REGNO_LAST_UID (REGNO (v
->dest_reg
))]
2769 >= INSN_LUID (loop_end
)))
2770 && ! (final_value
= v
->final_value
))
2774 /* Currently, non-reduced/final-value givs are never split. */
2775 /* Should emit insns after the loop if possible, as the biv final value
2778 /* If the final value is non-zero, and the giv has not been reduced,
2779 then must emit an instruction to set the final value. */
2780 if (final_value
&& !v
->new_reg
)
2782 /* Create a new register to hold the value of the giv, and then set
2783 the giv to its final value before the loop start. The giv is set
2784 to its final value before loop start to ensure that this insn
2785 will always be executed, no matter how we exit. */
2786 tem
= gen_reg_rtx (v
->mode
);
2787 emit_insn_before (gen_move_insn (tem
, v
->dest_reg
), loop_start
);
2788 emit_insn_before (gen_move_insn (v
->dest_reg
, final_value
),
2791 if (loop_dump_stream
)
2792 fprintf (loop_dump_stream
, "Giv %d mapped to %d for split.\n",
2793 REGNO (v
->dest_reg
), REGNO (tem
));
2799 /* This giv is splittable. If completely unrolling the loop, save the
2800 giv's initial value. Otherwise, save the constant zero for it. */
2802 if (unroll_type
== UNROLL_COMPLETELY
)
2804 /* It is not safe to use bl->initial_value here, because it may not
2805 be invariant. It is safe to use the initial value stored in
2806 the splittable_regs array if it is set. In rare cases, it won't
2807 be set, so then we do exactly the same thing as
2808 find_splittable_regs does to get a safe value. */
2809 rtx biv_initial_value
;
2811 if (splittable_regs
[bl
->regno
])
2812 biv_initial_value
= splittable_regs
[bl
->regno
];
2813 else if (GET_CODE (bl
->initial_value
) != REG
2814 || (REGNO (bl
->initial_value
) != bl
->regno
2815 && REGNO (bl
->initial_value
) >= FIRST_PSEUDO_REGISTER
))
2816 biv_initial_value
= bl
->initial_value
;
2819 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2821 record_base_value (REGNO (tem
), bl
->biv
->add_val
, 0);
2822 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2824 biv_initial_value
= tem
;
2826 value
= fold_rtx_mult_add (v
->mult_val
, biv_initial_value
,
2827 v
->add_val
, v
->mode
);
2834 /* If a giv was combined with another giv, then we can only split
2835 this giv if the giv it was combined with was reduced. This
2836 is because the value of v->new_reg is meaningless in this
2838 if (v
->same
&& ! v
->same
->new_reg
)
2840 if (loop_dump_stream
)
2841 fprintf (loop_dump_stream
,
2842 "giv combined with unreduced giv not split.\n");
2845 /* If the giv is an address destination, it could be something other
2846 than a simple register, these have to be treated differently. */
2847 else if (v
->giv_type
== DEST_REG
)
2849 /* If value is not a constant, register, or register plus
2850 constant, then compute its value into a register before
2851 loop start. This prevents invalid rtx sharing, and should
2852 generate better code. We can use bl->initial_value here
2853 instead of splittable_regs[bl->regno] because this code
2854 is going before the loop start. */
2855 if (unroll_type
== UNROLL_COMPLETELY
2856 && GET_CODE (value
) != CONST_INT
2857 && GET_CODE (value
) != REG
2858 && (GET_CODE (value
) != PLUS
2859 || GET_CODE (XEXP (value
, 0)) != REG
2860 || GET_CODE (XEXP (value
, 1)) != CONST_INT
))
2862 rtx tem
= gen_reg_rtx (v
->mode
);
2863 record_base_value (REGNO (tem
), v
->add_val
, 0);
2864 emit_iv_add_mult (bl
->initial_value
, v
->mult_val
,
2865 v
->add_val
, tem
, loop_start
);
2869 splittable_regs
[REGNO (v
->new_reg
)] = value
;
2873 /* Splitting address givs is useful since it will often allow us
2874 to eliminate some increment insns for the base giv as
2877 /* If the addr giv is combined with a dest_reg giv, then all
2878 references to that dest reg will be remapped, which is NOT
2879 what we want for split addr regs. We always create a new
2880 register for the split addr giv, just to be safe. */
2882 /* If we have multiple identical address givs within a
2883 single instruction, then use a single pseudo reg for
2884 both. This is necessary in case one is a match_dup
2887 v
->const_adjust
= 0;
2891 v
->dest_reg
= v
->same_insn
->dest_reg
;
2892 if (loop_dump_stream
)
2893 fprintf (loop_dump_stream
,
2894 "Sharing address givs in insn %d\n",
2895 INSN_UID (v
->insn
));
2897 /* If multiple address GIVs have been combined with the
2898 same dest_reg GIV, do not create a new register for
2900 else if (unroll_type
!= UNROLL_COMPLETELY
2901 && v
->giv_type
== DEST_ADDR
2902 && v
->same
&& v
->same
->giv_type
== DEST_ADDR
2903 && v
->same
->unrolled
2904 /* combine_givs_p may return true for some cases
2905 where the add and mult values are not equal.
2906 To share a register here, the values must be
2908 && rtx_equal_p (v
->same
->mult_val
, v
->mult_val
)
2909 && rtx_equal_p (v
->same
->add_val
, v
->add_val
))
2912 v
->dest_reg
= v
->same
->dest_reg
;
2915 else if (unroll_type
!= UNROLL_COMPLETELY
)
2917 /* If not completely unrolling the loop, then create a new
2918 register to hold the split value of the DEST_ADDR giv.
2919 Emit insn to initialize its value before loop start. */
2921 rtx tem
= gen_reg_rtx (v
->mode
);
2922 record_base_value (REGNO (tem
), v
->add_val
, 0);
2924 /* If the address giv has a constant in its new_reg value,
2925 then this constant can be pulled out and put in value,
2926 instead of being part of the initialization code. */
2928 if (GET_CODE (v
->new_reg
) == PLUS
2929 && GET_CODE (XEXP (v
->new_reg
, 1)) == CONST_INT
)
2932 = plus_constant (tem
, INTVAL (XEXP (v
->new_reg
,1)));
2934 /* Only succeed if this will give valid addresses.
2935 Try to validate both the first and the last
2936 address resulting from loop unrolling, if
2937 one fails, then can't do const elim here. */
2938 if (verify_addresses (v
, giv_inc
, unroll_number
))
2940 /* Save the negative of the eliminated const, so
2941 that we can calculate the dest_reg's increment
2943 v
->const_adjust
= - INTVAL (XEXP (v
->new_reg
, 1));
2945 v
->new_reg
= XEXP (v
->new_reg
, 0);
2946 if (loop_dump_stream
)
2947 fprintf (loop_dump_stream
,
2948 "Eliminating constant from giv %d\n",
2957 /* If the address hasn't been checked for validity yet, do so
2958 now, and fail completely if either the first or the last
2959 unrolled copy of the address is not a valid address
2960 for the instruction that uses it. */
2961 if (v
->dest_reg
== tem
2962 && ! verify_addresses (v
, giv_inc
, unroll_number
))
2964 for (v2
= v
->next_iv
; v2
; v2
= v2
->next_iv
)
2965 if (v2
->same_insn
== v
)
2968 if (loop_dump_stream
)
2969 fprintf (loop_dump_stream
,
2970 "Invalid address for giv at insn %d\n",
2971 INSN_UID (v
->insn
));
2975 /* We set this after the address check, to guarantee that
2976 the register will be initialized. */
2979 /* To initialize the new register, just move the value of
2980 new_reg into it. This is not guaranteed to give a valid
2981 instruction on machines with complex addressing modes.
2982 If we can't recognize it, then delete it and emit insns
2983 to calculate the value from scratch. */
2984 emit_insn_before (gen_rtx_SET (VOIDmode
, tem
,
2985 copy_rtx (v
->new_reg
)),
2987 if (recog_memoized (PREV_INSN (loop_start
)) < 0)
2991 /* We can't use bl->initial_value to compute the initial
2992 value, because the loop may have been preconditioned.
2993 We must calculate it from NEW_REG. Try using
2994 force_operand instead of emit_iv_add_mult. */
2995 delete_insn (PREV_INSN (loop_start
));
2998 ret
= force_operand (v
->new_reg
, tem
);
3000 emit_move_insn (tem
, ret
);
3001 sequence
= gen_sequence ();
3003 emit_insn_before (sequence
, loop_start
);
3005 if (loop_dump_stream
)
3006 fprintf (loop_dump_stream
,
3007 "Invalid init insn, rewritten.\n");
3012 v
->dest_reg
= value
;
3014 /* Check the resulting address for validity, and fail
3015 if the resulting address would be invalid. */
3016 if (! verify_addresses (v
, giv_inc
, unroll_number
))
3018 for (v2
= v
->next_iv
; v2
; v2
= v2
->next_iv
)
3019 if (v2
->same_insn
== v
)
3022 if (loop_dump_stream
)
3023 fprintf (loop_dump_stream
,
3024 "Invalid address for giv at insn %d\n",
3025 INSN_UID (v
->insn
));
3030 /* Store the value of dest_reg into the insn. This sharing
3031 will not be a problem as this insn will always be copied
3034 *v
->location
= v
->dest_reg
;
3036 /* If this address giv is combined with a dest reg giv, then
3037 save the base giv's induction pointer so that we will be
3038 able to handle this address giv properly. The base giv
3039 itself does not have to be splittable. */
3041 if (v
->same
&& v
->same
->giv_type
== DEST_REG
)
3042 addr_combined_regs
[REGNO (v
->same
->new_reg
)] = v
->same
;
3044 if (GET_CODE (v
->new_reg
) == REG
)
3046 /* This giv maybe hasn't been combined with any others.
3047 Make sure that it's giv is marked as splittable here. */
3049 splittable_regs
[REGNO (v
->new_reg
)] = value
;
3051 /* Make it appear to depend upon itself, so that the
3052 giv will be properly split in the main loop above. */
3056 addr_combined_regs
[REGNO (v
->new_reg
)] = v
;
3060 if (loop_dump_stream
)
3061 fprintf (loop_dump_stream
, "DEST_ADDR giv being split.\n");
3067 /* Currently, unreduced giv's can't be split. This is not too much
3068 of a problem since unreduced giv's are not live across loop
3069 iterations anyways. When unrolling a loop completely though,
3070 it makes sense to reduce&split givs when possible, as this will
3071 result in simpler instructions, and will not require that a reg
3072 be live across loop iterations. */
3074 splittable_regs
[REGNO (v
->dest_reg
)] = value
;
3075 fprintf (stderr
, "Giv %d at insn %d not reduced\n",
3076 REGNO (v
->dest_reg
), INSN_UID (v
->insn
));
3082 /* Unreduced givs are only updated once by definition. Reduced givs
3083 are updated as many times as their biv is. Mark it so if this is
3084 a splittable register. Don't need to do anything for address givs
3085 where this may not be a register. */
3087 if (GET_CODE (v
->new_reg
) == REG
)
3091 count
= reg_biv_class
[REGNO (v
->src_reg
)]->biv_count
;
3093 splittable_regs_updates
[REGNO (v
->new_reg
)] = count
;
3098 if (loop_dump_stream
)
3102 if (GET_CODE (v
->dest_reg
) == CONST_INT
)
3104 else if (GET_CODE (v
->dest_reg
) != REG
)
3105 regnum
= REGNO (XEXP (v
->dest_reg
, 0));
3107 regnum
= REGNO (v
->dest_reg
);
3108 fprintf (loop_dump_stream
, "Giv %d at insn %d safe to split.\n",
3109 regnum
, INSN_UID (v
->insn
));
3116 /* Try to prove that the register is dead after the loop exits. Trace every
3117 loop exit looking for an insn that will always be executed, which sets
3118 the register to some value, and appears before the first use of the register
3119 is found. If successful, then return 1, otherwise return 0. */
3121 /* ?? Could be made more intelligent in the handling of jumps, so that
3122 it can search past if statements and other similar structures. */
3125 reg_dead_after_loop (reg
, loop_start
, loop_end
)
3126 rtx reg
, loop_start
, loop_end
;
3131 int label_count
= 0;
3132 int this_loop_num
= uid_loop_num
[INSN_UID (loop_start
)];
3134 /* In addition to checking all exits of this loop, we must also check
3135 all exits of inner nested loops that would exit this loop. We don't
3136 have any way to identify those, so we just give up if there are any
3137 such inner loop exits. */
3139 for (label
= loop_number_exit_labels
[this_loop_num
]; label
;
3140 label
= LABEL_NEXTREF (label
))
3143 if (label_count
!= loop_number_exit_count
[this_loop_num
])
3146 /* HACK: Must also search the loop fall through exit, create a label_ref
3147 here which points to the loop_end, and append the loop_number_exit_labels
3149 label
= gen_rtx_LABEL_REF (VOIDmode
, loop_end
);
3150 LABEL_NEXTREF (label
) = loop_number_exit_labels
[this_loop_num
];
3152 for ( ; label
; label
= LABEL_NEXTREF (label
))
3154 /* Succeed if find an insn which sets the biv or if reach end of
3155 function. Fail if find an insn that uses the biv, or if come to
3156 a conditional jump. */
3158 insn
= NEXT_INSN (XEXP (label
, 0));
3161 code
= GET_CODE (insn
);
3162 if (GET_RTX_CLASS (code
) == 'i')
3166 if (reg_referenced_p (reg
, PATTERN (insn
)))
3169 set
= single_set (insn
);
3170 if (set
&& rtx_equal_p (SET_DEST (set
), reg
))
3174 if (code
== JUMP_INSN
)
3176 if (GET_CODE (PATTERN (insn
)) == RETURN
)
3178 else if (! simplejump_p (insn
)
3179 /* Prevent infinite loop following infinite loops. */
3180 || jump_count
++ > 20)
3183 insn
= JUMP_LABEL (insn
);
3186 insn
= NEXT_INSN (insn
);
3190 /* Success, the register is dead on all loop exits. */
3194 /* Try to calculate the final value of the biv, the value it will have at
3195 the end of the loop. If we can do it, return that value. */
3198 final_biv_value (bl
, loop_start
, loop_end
)
3199 struct iv_class
*bl
;
3200 rtx loop_start
, loop_end
;
3204 /* ??? This only works for MODE_INT biv's. Reject all others for now. */
3206 if (GET_MODE_CLASS (bl
->biv
->mode
) != MODE_INT
)
3209 /* The final value for reversed bivs must be calculated differently than
3210 for ordinary bivs. In this case, there is already an insn after the
3211 loop which sets this biv's final value (if necessary), and there are
3212 no other loop exits, so we can return any value. */
3215 if (loop_dump_stream
)
3216 fprintf (loop_dump_stream
,
3217 "Final biv value for %d, reversed biv.\n", bl
->regno
);
3222 /* Try to calculate the final value as initial value + (number of iterations
3223 * increment). For this to work, increment must be invariant, the only
3224 exit from the loop must be the fall through at the bottom (otherwise
3225 it may not have its final value when the loop exits), and the initial
3226 value of the biv must be invariant. */
3228 if (loop_n_iterations
!= 0
3229 && ! loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]]
3230 && invariant_p (bl
->initial_value
))
3232 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
3234 if (increment
&& invariant_p (increment
))
3236 /* Can calculate the loop exit value, emit insns after loop
3237 end to calculate this value into a temporary register in
3238 case it is needed later. */
3240 tem
= gen_reg_rtx (bl
->biv
->mode
);
3241 record_base_value (REGNO (tem
), bl
->biv
->add_val
, 0);
3242 /* Make sure loop_end is not the last insn. */
3243 if (NEXT_INSN (loop_end
) == 0)
3244 emit_note_after (NOTE_INSN_DELETED
, loop_end
);
3245 emit_iv_add_mult (increment
, GEN_INT (loop_n_iterations
),
3246 bl
->initial_value
, tem
, NEXT_INSN (loop_end
));
3248 if (loop_dump_stream
)
3249 fprintf (loop_dump_stream
,
3250 "Final biv value for %d, calculated.\n", bl
->regno
);
3256 /* Check to see if the biv is dead at all loop exits. */
3257 if (reg_dead_after_loop (bl
->biv
->src_reg
, loop_start
, loop_end
))
3259 if (loop_dump_stream
)
3260 fprintf (loop_dump_stream
,
3261 "Final biv value for %d, biv dead after loop exit.\n",
3270 /* Try to calculate the final value of the giv, the value it will have at
3271 the end of the loop. If we can do it, return that value. */
3274 final_giv_value (v
, loop_start
, loop_end
)
3275 struct induction
*v
;
3276 rtx loop_start
, loop_end
;
3278 struct iv_class
*bl
;
3281 rtx insert_before
, seq
;
3283 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
3285 /* The final value for givs which depend on reversed bivs must be calculated
3286 differently than for ordinary givs. In this case, there is already an
3287 insn after the loop which sets this giv's final value (if necessary),
3288 and there are no other loop exits, so we can return any value. */
3291 if (loop_dump_stream
)
3292 fprintf (loop_dump_stream
,
3293 "Final giv value for %d, depends on reversed biv\n",
3294 REGNO (v
->dest_reg
));
3298 /* Try to calculate the final value as a function of the biv it depends
3299 upon. The only exit from the loop must be the fall through at the bottom
3300 (otherwise it may not have its final value when the loop exits). */
3302 /* ??? Can calculate the final giv value by subtracting off the
3303 extra biv increments times the giv's mult_val. The loop must have
3304 only one exit for this to work, but the loop iterations does not need
3307 if (loop_n_iterations
!= 0
3308 && ! loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]])
3310 /* ?? It is tempting to use the biv's value here since these insns will
3311 be put after the loop, and hence the biv will have its final value
3312 then. However, this fails if the biv is subsequently eliminated.
3313 Perhaps determine whether biv's are eliminable before trying to
3314 determine whether giv's are replaceable so that we can use the
3315 biv value here if it is not eliminable. */
3317 /* We are emitting code after the end of the loop, so we must make
3318 sure that bl->initial_value is still valid then. It will still
3319 be valid if it is invariant. */
3321 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
3323 if (increment
&& invariant_p (increment
)
3324 && invariant_p (bl
->initial_value
))
3326 /* Can calculate the loop exit value of its biv as
3327 (loop_n_iterations * increment) + initial_value */
3329 /* The loop exit value of the giv is then
3330 (final_biv_value - extra increments) * mult_val + add_val.
3331 The extra increments are any increments to the biv which
3332 occur in the loop after the giv's value is calculated.
3333 We must search from the insn that sets the giv to the end
3334 of the loop to calculate this value. */
3336 insert_before
= NEXT_INSN (loop_end
);
3338 /* Put the final biv value in tem. */
3339 tem
= gen_reg_rtx (bl
->biv
->mode
);
3340 record_base_value (REGNO (tem
), bl
->biv
->add_val
, 0);
3341 emit_iv_add_mult (increment
, GEN_INT (loop_n_iterations
),
3342 bl
->initial_value
, tem
, insert_before
);
3344 /* Subtract off extra increments as we find them. */
3345 for (insn
= NEXT_INSN (v
->insn
); insn
!= loop_end
;
3346 insn
= NEXT_INSN (insn
))
3348 struct induction
*biv
;
3350 for (biv
= bl
->biv
; biv
; biv
= biv
->next_iv
)
3351 if (biv
->insn
== insn
)
3354 tem
= expand_binop (GET_MODE (tem
), sub_optab
, tem
,
3355 biv
->add_val
, NULL_RTX
, 0,
3357 seq
= gen_sequence ();
3359 emit_insn_before (seq
, insert_before
);
3363 /* Now calculate the giv's final value. */
3364 emit_iv_add_mult (tem
, v
->mult_val
, v
->add_val
, tem
,
3367 if (loop_dump_stream
)
3368 fprintf (loop_dump_stream
,
3369 "Final giv value for %d, calc from biv's value.\n",
3370 REGNO (v
->dest_reg
));
3376 /* Replaceable giv's should never reach here. */
3380 /* Check to see if the biv is dead at all loop exits. */
3381 if (reg_dead_after_loop (v
->dest_reg
, loop_start
, loop_end
))
3383 if (loop_dump_stream
)
3384 fprintf (loop_dump_stream
,
3385 "Final giv value for %d, giv dead after loop exit.\n",
3386 REGNO (v
->dest_reg
));
3395 /* Calculate the number of loop iterations. Returns the exact number of loop
3396 iterations if it can be calculated, otherwise returns zero. */
3398 unsigned HOST_WIDE_INT
3399 loop_iterations (loop_start
, loop_end
)
3400 rtx loop_start
, loop_end
;
3402 rtx comparison
, comparison_value
;
3403 rtx iteration_var
, initial_value
, increment
, final_value
;
3404 enum rtx_code comparison_code
;
3407 int unsigned_compare
, compare_dir
, final_larger
;
3408 unsigned long tempu
;
3411 /* First find the iteration variable. If the last insn is a conditional
3412 branch, and the insn before tests a register value, make that the
3413 iteration variable. */
3415 loop_initial_value
= 0;
3417 loop_final_value
= 0;
3418 loop_iteration_var
= 0;
3420 /* We used to use pren_nonnote_insn here, but that fails because it might
3421 accidentally get the branch for a contained loop if the branch for this
3422 loop was deleted. We can only trust branches immediately before the
3424 last_loop_insn
= PREV_INSN (loop_end
);
3426 comparison
= get_condition_for_loop (last_loop_insn
);
3427 if (comparison
== 0)
3429 if (loop_dump_stream
)
3430 fprintf (loop_dump_stream
,
3431 "Loop unrolling: No final conditional branch found.\n");
3435 /* ??? Get_condition may switch position of induction variable and
3436 invariant register when it canonicalizes the comparison. */
3438 comparison_code
= GET_CODE (comparison
);
3439 iteration_var
= XEXP (comparison
, 0);
3440 comparison_value
= XEXP (comparison
, 1);
3442 if (GET_CODE (iteration_var
) != REG
)
3444 if (loop_dump_stream
)
3445 fprintf (loop_dump_stream
,
3446 "Loop unrolling: Comparison not against register.\n");
3450 /* Loop iterations is always called before any new registers are created
3451 now, so this should never occur. */
3453 if (REGNO (iteration_var
) >= max_reg_before_loop
)
3456 iteration_info (iteration_var
, &initial_value
, &increment
,
3457 loop_start
, loop_end
);
3458 if (initial_value
== 0)
3459 /* iteration_info already printed a message. */
3462 /* If the comparison value is an invariant register, then try to find
3463 its value from the insns before the start of the loop. */
3465 if (GET_CODE (comparison_value
) == REG
&& invariant_p (comparison_value
))
3469 for (insn
= PREV_INSN (loop_start
); insn
; insn
= PREV_INSN (insn
))
3471 if (GET_CODE (insn
) == CODE_LABEL
)
3474 else if (GET_RTX_CLASS (GET_CODE (insn
)) == 'i'
3475 && reg_set_p (comparison_value
, insn
))
3477 /* We found the last insn before the loop that sets the register.
3478 If it sets the entire register, and has a REG_EQUAL note,
3479 then use the value of the REG_EQUAL note. */
3480 if ((set
= single_set (insn
))
3481 && (SET_DEST (set
) == comparison_value
))
3483 rtx note
= find_reg_note (insn
, REG_EQUAL
, NULL_RTX
);
3485 /* Only use the REG_EQUAL note if it is a constant.
3486 Other things, divide in particular, will cause
3487 problems later if we use them. */
3488 if (note
&& GET_CODE (XEXP (note
, 0)) != EXPR_LIST
3489 && CONSTANT_P (XEXP (note
, 0)))
3490 comparison_value
= XEXP (note
, 0);
3497 final_value
= approx_final_value (comparison_code
, comparison_value
,
3498 &unsigned_compare
, &compare_dir
);
3500 /* Save the calculated values describing this loop's bounds, in case
3501 precondition_loop_p will need them later. These values can not be
3502 recalculated inside precondition_loop_p because strength reduction
3503 optimizations may obscure the loop's structure. */
3505 loop_iteration_var
= iteration_var
;
3506 loop_initial_value
= initial_value
;
3507 loop_increment
= increment
;
3508 loop_final_value
= final_value
;
3509 loop_comparison_code
= comparison_code
;
3513 if (loop_dump_stream
)
3514 fprintf (loop_dump_stream
,
3515 "Loop unrolling: Increment value can't be calculated.\n");
3518 else if (GET_CODE (increment
) != CONST_INT
)
3520 if (loop_dump_stream
)
3521 fprintf (loop_dump_stream
,
3522 "Loop unrolling: Increment value not constant.\n");
3525 else if (GET_CODE (initial_value
) != CONST_INT
)
3527 if (loop_dump_stream
)
3528 fprintf (loop_dump_stream
,
3529 "Loop unrolling: Initial value not constant.\n");
3532 else if (final_value
== 0)
3534 if (loop_dump_stream
)
3535 fprintf (loop_dump_stream
,
3536 "Loop unrolling: EQ comparison loop.\n");
3539 else if (GET_CODE (final_value
) != CONST_INT
)
3541 if (loop_dump_stream
)
3542 fprintf (loop_dump_stream
,
3543 "Loop unrolling: Final value not constant.\n");
3547 /* ?? Final value and initial value do not have to be constants.
3548 Only their difference has to be constant. When the iteration variable
3549 is an array address, the final value and initial value might both
3550 be addresses with the same base but different constant offsets.
3551 Final value must be invariant for this to work.
3553 To do this, need some way to find the values of registers which are
3556 /* Final_larger is 1 if final larger, 0 if they are equal, otherwise -1. */
3557 if (unsigned_compare
)
3559 = ((unsigned HOST_WIDE_INT
) INTVAL (final_value
)
3560 > (unsigned HOST_WIDE_INT
) INTVAL (initial_value
))
3561 - ((unsigned HOST_WIDE_INT
) INTVAL (final_value
)
3562 < (unsigned HOST_WIDE_INT
) INTVAL (initial_value
));
3564 final_larger
= (INTVAL (final_value
) > INTVAL (initial_value
))
3565 - (INTVAL (final_value
) < INTVAL (initial_value
));
3567 if (INTVAL (increment
) > 0)
3569 else if (INTVAL (increment
) == 0)
3574 /* There are 27 different cases: compare_dir = -1, 0, 1;
3575 final_larger = -1, 0, 1; increment_dir = -1, 0, 1.
3576 There are 4 normal cases, 4 reverse cases (where the iteration variable
3577 will overflow before the loop exits), 4 infinite loop cases, and 15
3578 immediate exit (0 or 1 iteration depending on loop type) cases.
3579 Only try to optimize the normal cases. */
3581 /* (compare_dir/final_larger/increment_dir)
3582 Normal cases: (0/-1/-1), (0/1/1), (-1/-1/-1), (1/1/1)
3583 Reverse cases: (0/-1/1), (0/1/-1), (-1/-1/1), (1/1/-1)
3584 Infinite loops: (0/-1/0), (0/1/0), (-1/-1/0), (1/1/0)
3585 Immediate exit: (0/0/X), (-1/0/X), (-1/1/X), (1/0/X), (1/-1/X) */
3587 /* ?? If the meaning of reverse loops (where the iteration variable
3588 will overflow before the loop exits) is undefined, then could
3589 eliminate all of these special checks, and just always assume
3590 the loops are normal/immediate/infinite. Note that this means
3591 the sign of increment_dir does not have to be known. Also,
3592 since it does not really hurt if immediate exit loops or infinite loops
3593 are optimized, then that case could be ignored also, and hence all
3594 loops can be optimized.
3596 According to ANSI Spec, the reverse loop case result is undefined,
3597 because the action on overflow is undefined.
3599 See also the special test for NE loops below. */
3601 if (final_larger
== increment_dir
&& final_larger
!= 0
3602 && (final_larger
== compare_dir
|| compare_dir
== 0))
3607 if (loop_dump_stream
)
3608 fprintf (loop_dump_stream
,
3609 "Loop unrolling: Not normal loop.\n");
3613 /* Calculate the number of iterations, final_value is only an approximation,
3614 so correct for that. Note that tempu and loop_n_iterations are
3615 unsigned, because they can be as large as 2^n - 1. */
3617 i
= INTVAL (increment
);
3619 tempu
= INTVAL (final_value
) - INTVAL (initial_value
);
3622 tempu
= INTVAL (initial_value
) - INTVAL (final_value
);
3628 /* For NE tests, make sure that the iteration variable won't miss the
3629 final value. If tempu mod i is not zero, then the iteration variable
3630 will overflow before the loop exits, and we can not calculate the
3631 number of iterations. */
3632 if (compare_dir
== 0 && (tempu
% i
) != 0)
3635 return tempu
/ i
+ ((tempu
% i
) != 0);
3638 /* Replace uses of split bivs with their split pseudo register. This is
3639 for original instructions which remain after loop unrolling without
3643 remap_split_bivs (x
)
3646 register enum rtx_code code
;
3653 code
= GET_CODE (x
);
3668 /* If non-reduced/final-value givs were split, then this would also
3669 have to remap those givs also. */
3671 if (REGNO (x
) < max_reg_before_loop
3672 && reg_iv_type
[REGNO (x
)] == BASIC_INDUCT
)
3673 return reg_biv_class
[REGNO (x
)]->biv
->src_reg
;
3680 fmt
= GET_RTX_FORMAT (code
);
3681 for (i
= GET_RTX_LENGTH (code
) - 1; i
>= 0; i
--)
3684 XEXP (x
, i
) = remap_split_bivs (XEXP (x
, i
));
3688 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
3689 XVECEXP (x
, i
, j
) = remap_split_bivs (XVECEXP (x
, i
, j
));
3695 /* If FIRST_UID is a set of REGNO, and FIRST_UID dominates LAST_UID (e.g.
3696 FIST_UID is always executed if LAST_UID is), then return 1. Otherwise
3697 return 0. COPY_START is where we can start looking for the insns
3698 FIRST_UID and LAST_UID. COPY_END is where we stop looking for these
3701 If there is no JUMP_INSN between LOOP_START and FIRST_UID, then FIRST_UID
3702 must dominate LAST_UID.
3704 If there is a CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID
3705 may not dominate LAST_UID.
3707 If there is no CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID
3708 must dominate LAST_UID. */
3711 set_dominates_use (regno
, first_uid
, last_uid
, copy_start
, copy_end
)
3718 int passed_jump
= 0;
3719 rtx p
= NEXT_INSN (copy_start
);
3721 while (INSN_UID (p
) != first_uid
)
3723 if (GET_CODE (p
) == JUMP_INSN
)
3725 /* Could not find FIRST_UID. */
3731 /* Verify that FIRST_UID is an insn that entirely sets REGNO. */
3732 if (GET_RTX_CLASS (GET_CODE (p
)) != 'i'
3733 || ! dead_or_set_regno_p (p
, regno
))
3736 /* FIRST_UID is always executed. */
3737 if (passed_jump
== 0)
3740 while (INSN_UID (p
) != last_uid
)
3742 /* If we see a CODE_LABEL between FIRST_UID and LAST_UID, then we
3743 can not be sure that FIRST_UID dominates LAST_UID. */
3744 if (GET_CODE (p
) == CODE_LABEL
)
3746 /* Could not find LAST_UID, but we reached the end of the loop, so
3748 else if (p
== copy_end
)
3753 /* FIRST_UID is always executed if LAST_UID is executed. */