1 /* Try to unroll loops, and split induction variables.
2 Copyright (C) 1992 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, 675 Mass Ave, Cambridge, MA 02139, USA. */
21 /* Try to unroll a loop, and split induction variables.
23 Loops for which the number of iterations can be calculated exactly are
24 handled specially. If the number of iterations times the insn_count is
25 less than MAX_UNROLLED_INSNS, then the loop is unrolled completely.
26 Otherwise, we try to unroll the loop a number of times modulo the number
27 of iterations, so that only one exit test will be needed. It is unrolled
28 a number of times approximately equal to MAX_UNROLLED_INSNS divided by
31 Otherwise, if the number of iterations can be calculated exactly at
32 run time, and the loop is always entered at the top, then we try to
33 precondition the loop. That is, at run time, calculate how many times
34 the loop will execute, and then execute the loop body a few times so
35 that the remaining iterations will be some multiple of 4 (or 2 if the
36 loop is large). Then fall through to a loop unrolled 4 (or 2) times,
37 with only one exit test needed at the end of the loop.
39 Otherwise, if the number of iterations can not be calculated exactly,
40 not even at run time, then we still unroll the loop a number of times
41 approximately equal to MAX_UNROLLED_INSNS divided by the insn count,
42 but there must be an exit test after each copy of the loop body.
44 For each induction variable, which is dead outside the loop (replaceable)
45 or for which we can easily calculate the final value, if we can easily
46 calculate its value at each place where it is set as a function of the
47 current loop unroll count and the variable's value at loop entry, then
48 the induction variable is split into `N' different variables, one for
49 each copy of the loop body. One variable is live across the backward
50 branch, and the others are all calculated as a function of this variable.
51 This helps eliminate data dependencies, and leads to further opportunities
54 /* Possible improvements follow: */
56 /* ??? Add an extra pass somewhere to determine whether unrolling will
57 give any benefit. E.g. after generating all unrolled insns, compute the
58 cost of all insns and compare against cost of insns in rolled loop.
60 - On traditional architectures, unrolling a non-constant bound loop
61 is a win if there is a giv whose only use is in memory addresses, the
62 memory addresses can be split, and hence giv increments can be
64 - It is also a win if the loop is executed many times, and preconditioning
65 can be performed for the loop.
66 Add code to check for these and similar cases. */
68 /* ??? Improve control of which loops get unrolled. Could use profiling
69 info to only unroll the most commonly executed loops. Perhaps have
70 a user specifyable option to control the amount of code expansion,
71 or the percent of loops to consider for unrolling. Etc. */
73 /* ??? Look at the register copies inside the loop to see if they form a
74 simple permutation. If so, iterate the permutation until it gets back to
75 the start state. This is how many times we should unroll the loop, for
76 best results, because then all register copies can be eliminated.
77 For example, the lisp nreverse function should be unrolled 3 times
86 ??? The number of times to unroll the loop may also be based on data
87 references in the loop. For example, if we have a loop that references
88 x[i-1], x[i], and x[i+1], we should unroll it a multiple of 3 times. */
90 /* ??? Add some simple linear equation solving capability so that we can
91 determine the number of loop iterations for more complex loops.
92 For example, consider this loop from gdb
93 #define SWAP_TARGET_AND_HOST(buffer,len)
96 char *p = (char *) buffer;
97 char *q = ((char *) buffer) + len - 1;
98 int iterations = (len + 1) >> 1;
100 for (p; p < q; p++, q--;)
108 start value = p = &buffer + current_iteration
109 end value = q = &buffer + len - 1 - current_iteration
110 Given the loop exit test of "p < q", then there must be "q - p" iterations,
111 set equal to zero and solve for number of iterations:
112 q - p = len - 1 - 2*current_iteration = 0
113 current_iteration = (len - 1) / 2
114 Hence, there are (len - 1) / 2 (rounded up to the nearest integer)
115 iterations of this loop. */
117 /* ??? Currently, no labels are marked as loop invariant when doing loop
118 unrolling. This is because an insn inside the loop, that loads the address
119 of a label inside the loop into a register, could be moved outside the loop
120 by the invariant code motion pass if labels were invariant. If the loop
121 is subsequently unrolled, the code will be wrong because each unrolled
122 body of the loop will use the same address, whereas each actually needs a
123 different address. A case where this happens is when a loop containing
124 a switch statement is unrolled.
126 It would be better to let labels be considered invariant. When we
127 unroll loops here, check to see if any insns using a label local to the
128 loop were moved before the loop. If so, then correct the problem, by
129 moving the insn back into the loop, or perhaps replicate the insn before
130 the loop, one copy for each time the loop is unrolled. */
132 /* The prime factors looked for when trying to unroll a loop by some
133 number which is modulo the total number of iterations. Just checking
134 for these 4 prime factors will find at least one factor for 75% of
135 all numbers theoretically. Practically speaking, this will succeed
136 almost all of the time since loops are generally a multiple of 2
139 #define NUM_FACTORS 4
141 struct _factor
{ int factor
, count
; } factors
[NUM_FACTORS
]
142 = { {2, 0}, {3, 0}, {5, 0}, {7, 0}};
144 /* Describes the different types of loop unrolling performed. */
146 enum unroll_types
{ UNROLL_COMPLETELY
, UNROLL_MODULO
, UNROLL_NAIVE
};
150 #include "insn-config.h"
151 #include "integrate.h"
158 /* This controls which loops are unrolled, and by how much we unroll
161 #ifndef MAX_UNROLLED_INSNS
162 #define MAX_UNROLLED_INSNS 100
165 /* Indexed by register number, if non-zero, then it contains a pointer
166 to a struct induction for a DEST_REG giv which has been combined with
167 one of more address givs. This is needed because whenever such a DEST_REG
168 giv is modified, we must modify the value of all split address givs
169 that were combined with this DEST_REG giv. */
171 static struct induction
**addr_combined_regs
;
173 /* Indexed by register number, if this is a splittable induction variable,
174 then this will hold the current value of the register, which depends on the
177 static rtx
*splittable_regs
;
179 /* Indexed by register number, if this is a splittable induction variable,
180 then this will hold the number of instructions in the loop that modify
181 the induction variable. Used to ensure that only the last insn modifying
182 a split iv will update the original iv of the dest. */
184 static int *splittable_regs_updates
;
186 /* Values describing the current loop's iteration variable. These are set up
187 by loop_iterations, and used by precondition_loop_p. */
189 static rtx loop_iteration_var
;
190 static rtx loop_initial_value
;
191 static rtx loop_increment
;
192 static rtx loop_final_value
;
194 /* Forward declarations. */
196 static void init_reg_map ();
197 static int precondition_loop_p ();
198 static void copy_loop_body ();
199 static void iteration_info ();
200 static rtx
approx_final_value ();
201 static int find_splittable_regs ();
202 static int find_splittable_givs ();
203 static rtx
fold_rtx_mult_add ();
205 /* Try to unroll one loop and split induction variables in the loop.
207 The loop is described by the arguments LOOP_END, INSN_COUNT, and
208 LOOP_START. END_INSERT_BEFORE indicates where insns should be added
209 which need to be executed when the loop falls through. STRENGTH_REDUCTION_P
210 indicates whether information generated in the strength reduction pass
213 This function is intended to be called from within `strength_reduce'
217 unroll_loop (loop_end
, insn_count
, loop_start
, end_insert_before
,
222 rtx end_insert_before
;
223 int strength_reduce_p
;
226 int unroll_number
= 1;
227 rtx copy_start
, copy_end
;
228 rtx insn
, copy
, sequence
, pattern
, tem
;
229 int max_labelno
, max_insnno
;
231 struct inline_remap
*map
;
239 int splitting_not_safe
= 0;
240 enum unroll_types unroll_type
;
241 int loop_preconditioned
= 0;
243 /* This points to the last real insn in the loop, which should be either
244 a JUMP_INSN (for conditional jumps) or a BARRIER (for unconditional
248 /* Don't bother unrolling huge loops. Since the minimum factor is
249 two, loops greater than one half of MAX_UNROLLED_INSNS will never
251 if (insn_count
> MAX_UNROLLED_INSNS
/ 2)
253 if (loop_dump_stream
)
254 fprintf (loop_dump_stream
, "Unrolling failure: Loop too big.\n");
258 /* When emitting debugger info, we can't unroll loops with unequal numbers
259 of block_beg and block_end notes, because that would unbalance the block
260 structure of the function. This can happen as a result of the
261 "if (foo) bar; else break;" optimization in jump.c. */
263 if (write_symbols
!= NO_DEBUG
)
265 int block_begins
= 0;
268 for (insn
= loop_start
; insn
!= loop_end
; insn
= NEXT_INSN (insn
))
270 if (GET_CODE (insn
) == NOTE
)
272 if (NOTE_LINE_NUMBER (insn
) == NOTE_INSN_BLOCK_BEG
)
274 else if (NOTE_LINE_NUMBER (insn
) == NOTE_INSN_BLOCK_END
)
279 if (block_begins
!= block_ends
)
281 if (loop_dump_stream
)
282 fprintf (loop_dump_stream
,
283 "Unrolling failure: Unbalanced block notes.\n");
288 /* Determine type of unroll to perform. Depends on the number of iterations
289 and the size of the loop. */
291 /* If there is no strength reduce info, then set loop_n_iterations to zero.
292 This can happen if strength_reduce can't find any bivs in the loop.
293 A value of zero indicates that the number of iterations could not be
296 if (! strength_reduce_p
)
297 loop_n_iterations
= 0;
299 if (loop_dump_stream
&& loop_n_iterations
> 0)
300 fprintf (loop_dump_stream
,
301 "Loop unrolling: %d iterations.\n", loop_n_iterations
);
303 /* Find and save a pointer to the last nonnote insn in the loop. */
305 last_loop_insn
= prev_nonnote_insn (loop_end
);
307 /* Calculate how many times to unroll the loop. Indicate whether or
308 not the loop is being completely unrolled. */
310 if (loop_n_iterations
== 1)
312 /* If number of iterations is exactly 1, then eliminate the compare and
313 branch at the end of the loop since they will never be taken.
314 Then return, since no other action is needed here. */
316 /* If the last instruction is not a BARRIER or a JUMP_INSN, then
317 don't do anything. */
319 if (GET_CODE (last_loop_insn
) == BARRIER
)
321 /* Delete the jump insn. This will delete the barrier also. */
322 delete_insn (PREV_INSN (last_loop_insn
));
324 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
327 /* The immediately preceding insn is a compare which must be
329 delete_insn (last_loop_insn
);
330 delete_insn (PREV_INSN (last_loop_insn
));
332 /* The immediately preceding insn may not be the compare, so don't
334 delete_insn (last_loop_insn
);
339 else if (loop_n_iterations
> 0
340 && loop_n_iterations
* insn_count
< MAX_UNROLLED_INSNS
)
342 unroll_number
= loop_n_iterations
;
343 unroll_type
= UNROLL_COMPLETELY
;
345 else if (loop_n_iterations
> 0)
347 /* Try to factor the number of iterations. Don't bother with the
348 general case, only using 2, 3, 5, and 7 will get 75% of all
349 numbers theoretically, and almost all in practice. */
351 for (i
= 0; i
< NUM_FACTORS
; i
++)
352 factors
[i
].count
= 0;
354 temp
= loop_n_iterations
;
355 for (i
= NUM_FACTORS
- 1; i
>= 0; i
--)
356 while (temp
% factors
[i
].factor
== 0)
359 temp
= temp
/ factors
[i
].factor
;
362 /* Start with the larger factors first so that we generally
363 get lots of unrolling. */
367 for (i
= 3; i
>= 0; i
--)
368 while (factors
[i
].count
--)
370 if (temp
* factors
[i
].factor
< MAX_UNROLLED_INSNS
)
372 unroll_number
*= factors
[i
].factor
;
373 temp
*= factors
[i
].factor
;
379 /* If we couldn't find any factors, then unroll as in the normal
381 if (unroll_number
== 1)
383 if (loop_dump_stream
)
384 fprintf (loop_dump_stream
,
385 "Loop unrolling: No factors found.\n");
388 unroll_type
= UNROLL_MODULO
;
392 /* Default case, calculate number of times to unroll loop based on its
394 if (unroll_number
== 1)
396 if (8 * insn_count
< MAX_UNROLLED_INSNS
)
398 else if (4 * insn_count
< MAX_UNROLLED_INSNS
)
403 unroll_type
= UNROLL_NAIVE
;
406 /* Now we know how many times to unroll the loop. */
408 if (loop_dump_stream
)
409 fprintf (loop_dump_stream
,
410 "Unrolling loop %d times.\n", unroll_number
);
413 if (unroll_type
== UNROLL_COMPLETELY
|| unroll_type
== UNROLL_MODULO
)
415 /* Loops of these types should never start with a jump down to
416 the exit condition test. For now, check for this case just to
417 be sure. UNROLL_NAIVE loops can be of this form, this case is
420 while (GET_CODE (insn
) != CODE_LABEL
&& GET_CODE (insn
) != JUMP_INSN
)
421 insn
= NEXT_INSN (insn
);
422 if (GET_CODE (insn
) == JUMP_INSN
)
426 if (unroll_type
== UNROLL_COMPLETELY
)
428 /* Completely unrolling the loop: Delete the compare and branch at
429 the end (the last two instructions). This delete must done at the
430 very end of loop unrolling, to avoid problems with calls to
431 back_branch_in_range_p, which is called by find_splittable_regs.
432 All increments of splittable bivs/givs are changed to load constant
435 copy_start
= loop_start
;
437 /* Set insert_before to the instruction immediately after the JUMP_INSN
438 (or BARRIER), so that any NOTEs between the JUMP_INSN and the end of
439 the loop will be correctly handled by copy_loop_body. */
440 insert_before
= NEXT_INSN (last_loop_insn
);
442 /* Set copy_end to the insn before the jump at the end of the loop. */
443 if (GET_CODE (last_loop_insn
) == BARRIER
)
444 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
445 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
448 /* The instruction immediately before the JUMP_INSN is a compare
449 instruction which we do not want to copy. */
450 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
452 /* The instruction immediately before the JUMP_INSN may not be the
453 compare, so we must copy it. */
454 copy_end
= PREV_INSN (last_loop_insn
);
459 /* We currently can't unroll a loop if it doesn't end with a
460 JUMP_INSN. There would need to be a mechanism that recognizes
461 this case, and then inserts a jump after each loop body, which
462 jumps to after the last loop body. */
463 if (loop_dump_stream
)
464 fprintf (loop_dump_stream
,
465 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
469 else if (unroll_type
== UNROLL_MODULO
)
471 /* Partially unrolling the loop: The compare and branch at the end
472 (the last two instructions) must remain. Don't copy the compare
473 and branch instructions at the end of the loop. Insert the unrolled
474 code immediately before the compare/branch at the end so that the
475 code will fall through to them as before. */
477 copy_start
= loop_start
;
479 /* Set insert_before to the jump insn at the end of the loop.
480 Set copy_end to before the jump insn at the end of the loop. */
481 if (GET_CODE (last_loop_insn
) == BARRIER
)
483 insert_before
= PREV_INSN (last_loop_insn
);
484 copy_end
= PREV_INSN (insert_before
);
486 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
489 /* The instruction immediately before the JUMP_INSN is a compare
490 instruction which we do not want to copy or delete. */
491 insert_before
= PREV_INSN (last_loop_insn
);
492 copy_end
= PREV_INSN (insert_before
);
494 /* The instruction immediately before the JUMP_INSN may not be the
495 compare, so we must copy it. */
496 insert_before
= last_loop_insn
;
497 copy_end
= PREV_INSN (last_loop_insn
);
502 /* We currently can't unroll a loop if it doesn't end with a
503 JUMP_INSN. There would need to be a mechanism that recognizes
504 this case, and then inserts a jump after each loop body, which
505 jumps to after the last loop body. */
506 if (loop_dump_stream
)
507 fprintf (loop_dump_stream
,
508 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
514 /* Normal case: Must copy the compare and branch instructions at the
517 if (GET_CODE (last_loop_insn
) == BARRIER
)
519 /* Loop ends with an unconditional jump and a barrier.
520 Handle this like above, don't copy jump and barrier.
521 This is not strictly necessary, but doing so prevents generating
522 unconditional jumps to an immediately following label.
524 This will be corrected below if the target of this jump is
525 not the start_label. */
527 insert_before
= PREV_INSN (last_loop_insn
);
528 copy_end
= PREV_INSN (insert_before
);
530 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
532 /* Set insert_before to immediately after the JUMP_INSN, so that
533 NOTEs at the end of the loop will be correctly handled by
535 insert_before
= NEXT_INSN (last_loop_insn
);
536 copy_end
= last_loop_insn
;
540 /* We currently can't unroll a loop if it doesn't end with a
541 JUMP_INSN. There would need to be a mechanism that recognizes
542 this case, and then inserts a jump after each loop body, which
543 jumps to after the last loop body. */
544 if (loop_dump_stream
)
545 fprintf (loop_dump_stream
,
546 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
550 /* If copying exit test branches because they can not be eliminated,
551 then must convert the fall through case of the branch to a jump past
552 the end of the loop. Create a label to emit after the loop and save
553 it for later use. Do not use the label after the loop, if any, since
554 it might be used by insns outside the loop, or there might be insns
555 added before it later by final_[bg]iv_value which must be after
556 the real exit label. */
557 exit_label
= gen_label_rtx ();
560 while (GET_CODE (insn
) != CODE_LABEL
&& GET_CODE (insn
) != JUMP_INSN
)
561 insn
= NEXT_INSN (insn
);
563 if (GET_CODE (insn
) == JUMP_INSN
)
565 /* The loop starts with a jump down to the exit condition test.
566 Start copying the loop after the barrier following this
568 copy_start
= NEXT_INSN (insn
);
570 /* Splitting induction variables doesn't work when the loop is
571 entered via a jump to the bottom, because then we end up doing
572 a comparison against a new register for a split variable, but
573 we did not execute the set insn for the new register because
574 it was skipped over. */
575 splitting_not_safe
= 1;
576 if (loop_dump_stream
)
577 fprintf (loop_dump_stream
,
578 "Splitting not safe, because loop not entered at top.\n");
581 copy_start
= loop_start
;
584 /* This should always be the first label in the loop. */
585 start_label
= NEXT_INSN (copy_start
);
586 /* There may be a line number note and/or a loop continue note here. */
587 while (GET_CODE (start_label
) == NOTE
)
588 start_label
= NEXT_INSN (start_label
);
589 if (GET_CODE (start_label
) != CODE_LABEL
)
591 /* This can happen as a result of jump threading. If the first insns in
592 the loop test the same condition as the loop's backward jump, or the
593 opposite condition, then the backward jump will be modified to point
594 to elsewhere, and the loop's start label is deleted.
596 This case currently can not be handled by the loop unrolling code. */
598 if (loop_dump_stream
)
599 fprintf (loop_dump_stream
,
600 "Unrolling failure: unknown insns between BEG note and loop label.\n");
604 if (unroll_type
== UNROLL_NAIVE
605 && GET_CODE (last_loop_insn
) == BARRIER
606 && start_label
!= JUMP_LABEL (PREV_INSN (last_loop_insn
)))
608 /* In this case, we must copy the jump and barrier, because they will
609 not be converted to jumps to an immediately following label. */
611 insert_before
= NEXT_INSN (last_loop_insn
);
612 copy_end
= last_loop_insn
;
615 /* Allocate a translation table for the labels and insn numbers.
616 They will be filled in as we copy the insns in the loop. */
618 max_labelno
= max_label_num ();
619 max_insnno
= get_max_uid ();
621 map
= (struct inline_remap
*) alloca (sizeof (struct inline_remap
));
623 /* Allocate the label map. */
627 map
->label_map
= (rtx
*) alloca (max_labelno
* sizeof (rtx
));
629 local_label
= (char *) alloca (max_labelno
);
630 bzero (local_label
, max_labelno
);
635 /* Search the loop and mark all local labels, i.e. the ones which have to
636 be distinct labels when copied. For all labels which might be
637 non-local, set their label_map entries to point to themselves.
638 If they happen to be local their label_map entries will be overwritten
639 before the loop body is copied. The label_map entries for local labels
640 will be set to a different value each time the loop body is copied. */
642 for (insn
= copy_start
; insn
!= loop_end
; insn
= NEXT_INSN (insn
))
644 if (GET_CODE (insn
) == CODE_LABEL
)
645 local_label
[CODE_LABEL_NUMBER (insn
)] = 1;
646 else if (GET_CODE (insn
) == JUMP_INSN
)
648 if (JUMP_LABEL (insn
))
649 map
->label_map
[CODE_LABEL_NUMBER (JUMP_LABEL (insn
))]
651 else if (GET_CODE (PATTERN (insn
)) == ADDR_VEC
652 || GET_CODE (PATTERN (insn
)) == ADDR_DIFF_VEC
)
654 rtx pat
= PATTERN (insn
);
655 int diff_vec_p
= GET_CODE (PATTERN (insn
)) == ADDR_DIFF_VEC
;
656 int len
= XVECLEN (pat
, diff_vec_p
);
659 for (i
= 0; i
< len
; i
++)
661 label
= XEXP (XVECEXP (pat
, diff_vec_p
, i
), 0);
662 map
->label_map
[CODE_LABEL_NUMBER (label
)] = label
;
668 /* Allocate space for the insn map. */
670 map
->insn_map
= (rtx
*) alloca (max_insnno
* sizeof (rtx
));
672 /* Set this to zero, to indicate that we are doing loop unrolling,
673 not function inlining. */
674 map
->inline_target
= 0;
676 /* The register and constant maps depend on the number of registers
677 present, so the final maps can't be created until after
678 find_splittable_regs is called. However, they are needed for
679 preconditioning, so we create temporary maps when preconditioning
682 /* The preconditioning code may allocate two new pseudo registers. */
683 maxregnum
= max_reg_num ();
685 /* Allocate and zero out the splittable_regs and addr_combined_regs
686 arrays. These must be zeroed here because they will be used if
687 loop preconditioning is performed, and must be zero for that case.
689 It is safe to do this here, since the extra registers created by the
690 preconditioning code and find_splittable_regs will never be used
691 to access the splittable_regs[] and addr_combined_regs[] arrays. */
693 splittable_regs
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
694 bzero (splittable_regs
, maxregnum
* sizeof (rtx
));
695 splittable_regs_updates
= (int *) alloca (maxregnum
* sizeof (int));
696 bzero (splittable_regs_updates
, maxregnum
* sizeof (int));
698 = (struct induction
**) alloca (maxregnum
* sizeof (struct induction
*));
699 bzero (addr_combined_regs
, maxregnum
* sizeof (struct induction
*));
701 /* If this loop requires exit tests when unrolled, check to see if we
702 can precondition the loop so as to make the exit tests unnecessary.
703 Just like variable splitting, this is not safe if the loop is entered
704 via a jump to the bottom. Also, can not do this if no strength
705 reduce info, because precondition_loop_p uses this info. */
707 /* Must copy the loop body for preconditioning before the following
708 find_splittable_regs call since that will emit insns which need to
709 be after the preconditioned loop copies, but immediately before the
710 unrolled loop copies. */
712 /* Also, it is not safe to split induction variables for the preconditioned
713 copies of the loop body. If we split induction variables, then the code
714 assumes that each induction variable can be represented as a function
715 of its initial value and the loop iteration number. This is not true
716 in this case, because the last preconditioned copy of the loop body
717 could be any iteration from the first up to the `unroll_number-1'th,
718 depending on the initial value of the iteration variable. Therefore
719 we can not split induction variables here, because we can not calculate
720 their value. Hence, this code must occur before find_splittable_regs
723 if (unroll_type
== UNROLL_NAIVE
&& ! splitting_not_safe
&& strength_reduce_p
)
725 rtx initial_value
, final_value
, increment
;
727 if (precondition_loop_p (&initial_value
, &final_value
, &increment
,
728 loop_start
, loop_end
))
730 register rtx diff
, temp
;
731 enum machine_mode mode
;
733 int abs_inc
, neg_inc
;
735 map
->reg_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
737 map
->const_equiv_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
738 map
->const_age_map
= (unsigned *) alloca (maxregnum
739 * sizeof (unsigned));
740 map
->const_equiv_map_size
= maxregnum
;
741 global_const_equiv_map
= map
->const_equiv_map
;
743 init_reg_map (map
, maxregnum
);
745 /* Limit loop unrolling to 4, since this will make 7 copies of
747 if (unroll_number
> 4)
750 /* Save the absolute value of the increment, and also whether or
751 not it is negative. */
753 abs_inc
= INTVAL (increment
);
762 /* Decide what mode to do these calculations in. Choose the larger
763 of final_value's mode and initial_value's mode, or a full-word if
764 both are constants. */
765 mode
= GET_MODE (final_value
);
766 if (mode
== VOIDmode
)
768 mode
= GET_MODE (initial_value
);
769 if (mode
== VOIDmode
)
772 else if (mode
!= GET_MODE (initial_value
)
773 && (GET_MODE_SIZE (mode
)
774 < GET_MODE_SIZE (GET_MODE (initial_value
))))
775 mode
= GET_MODE (initial_value
);
777 /* Calculate the difference between the final and initial values.
778 Final value may be a (plus (reg x) (const_int 1)) rtx.
779 Let the following cse pass simplify this if initial value is
782 We must copy the final and initial values here to avoid
783 improperly shared rtl. */
785 diff
= expand_binop (mode
, sub_optab
, copy_rtx (final_value
),
786 copy_rtx (initial_value
), NULL_RTX
, 0,
789 /* Now calculate (diff % (unroll * abs (increment))) by using an
791 diff
= expand_binop (GET_MODE (diff
), and_optab
, diff
,
792 GEN_INT (unroll_number
* abs_inc
- 1),
793 NULL_RTX
, 0, OPTAB_LIB_WIDEN
);
795 /* Now emit a sequence of branches to jump to the proper precond
798 labels
= (rtx
*) alloca (sizeof (rtx
) * unroll_number
);
799 for (i
= 0; i
< unroll_number
; i
++)
800 labels
[i
] = gen_label_rtx ();
802 /* Assuming the unroll_number is 4, and the increment is 2, then
803 for a negative increment: for a positive increment:
804 diff = 0,1 precond 0 diff = 0,7 precond 0
805 diff = 2,3 precond 3 diff = 1,2 precond 1
806 diff = 4,5 precond 2 diff = 3,4 precond 2
807 diff = 6,7 precond 1 diff = 5,6 precond 3 */
809 /* We only need to emit (unroll_number - 1) branches here, the
810 last case just falls through to the following code. */
812 /* ??? This would give better code if we emitted a tree of branches
813 instead of the current linear list of branches. */
815 for (i
= 0; i
< unroll_number
- 1; i
++)
819 /* For negative increments, must invert the constant compared
820 against, except when comparing against zero. */
824 cmp_const
= unroll_number
- i
;
828 emit_cmp_insn (diff
, GEN_INT (abs_inc
* cmp_const
),
829 EQ
, NULL_RTX
, mode
, 0, 0);
832 emit_jump_insn (gen_beq (labels
[i
]));
834 emit_jump_insn (gen_bge (labels
[i
]));
836 emit_jump_insn (gen_ble (labels
[i
]));
837 JUMP_LABEL (get_last_insn ()) = labels
[i
];
838 LABEL_NUSES (labels
[i
])++;
841 /* If the increment is greater than one, then we need another branch,
842 to handle other cases equivalent to 0. */
844 /* ??? This should be merged into the code above somehow to help
845 simplify the code here, and reduce the number of branches emitted.
846 For the negative increment case, the branch here could easily
847 be merged with the `0' case branch above. For the positive
848 increment case, it is not clear how this can be simplified. */
855 cmp_const
= abs_inc
- 1;
857 cmp_const
= abs_inc
* (unroll_number
- 1) + 1;
859 emit_cmp_insn (diff
, GEN_INT (cmp_const
), EQ
, NULL_RTX
,
863 emit_jump_insn (gen_ble (labels
[0]));
865 emit_jump_insn (gen_bge (labels
[0]));
866 JUMP_LABEL (get_last_insn ()) = labels
[0];
867 LABEL_NUSES (labels
[0])++;
870 sequence
= gen_sequence ();
872 emit_insn_before (sequence
, loop_start
);
874 /* Only the last copy of the loop body here needs the exit
875 test, so set copy_end to exclude the compare/branch here,
876 and then reset it inside the loop when get to the last
879 if (GET_CODE (last_loop_insn
) == BARRIER
)
880 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
881 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
884 /* The immediately preceding insn is a compare which we do not
886 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
888 /* The immediately preceding insn may not be a compare, so we
890 copy_end
= PREV_INSN (last_loop_insn
);
896 for (i
= 1; i
< unroll_number
; i
++)
898 emit_label_after (labels
[unroll_number
- i
],
899 PREV_INSN (loop_start
));
901 bzero (map
->insn_map
, max_insnno
* sizeof (rtx
));
902 bzero (map
->const_equiv_map
, maxregnum
* sizeof (rtx
));
903 bzero (map
->const_age_map
, maxregnum
* sizeof (unsigned));
906 for (j
= 0; j
< max_labelno
; j
++)
908 map
->label_map
[j
] = gen_label_rtx ();
910 /* The last copy needs the compare/branch insns at the end,
911 so reset copy_end here if the loop ends with a conditional
914 if (i
== unroll_number
- 1)
916 if (GET_CODE (last_loop_insn
) == BARRIER
)
917 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
919 copy_end
= last_loop_insn
;
922 /* None of the copies are the `last_iteration', so just
923 pass zero for that parameter. */
924 copy_loop_body (copy_start
, copy_end
, map
, exit_label
, 0,
925 unroll_type
, start_label
, loop_end
,
926 loop_start
, copy_end
);
928 emit_label_after (labels
[0], PREV_INSN (loop_start
));
930 if (GET_CODE (last_loop_insn
) == BARRIER
)
932 insert_before
= PREV_INSN (last_loop_insn
);
933 copy_end
= PREV_INSN (insert_before
);
938 /* The immediately preceding insn is a compare which we do not
940 insert_before
= PREV_INSN (last_loop_insn
);
941 copy_end
= PREV_INSN (insert_before
);
943 /* The immediately preceding insn may not be a compare, so we
945 insert_before
= last_loop_insn
;
946 copy_end
= PREV_INSN (last_loop_insn
);
950 /* Set unroll type to MODULO now. */
951 unroll_type
= UNROLL_MODULO
;
952 loop_preconditioned
= 1;
956 /* If reach here, and the loop type is UNROLL_NAIVE, then don't unroll
957 the loop unless all loops are being unrolled. */
958 if (unroll_type
== UNROLL_NAIVE
&& ! flag_unroll_all_loops
)
960 if (loop_dump_stream
)
961 fprintf (loop_dump_stream
, "Unrolling failure: Naive unrolling not being done.\n");
965 /* At this point, we are guaranteed to unroll the loop. */
967 /* For each biv and giv, determine whether it can be safely split into
968 a different variable for each unrolled copy of the loop body.
969 We precalculate and save this info here, since computing it is
972 Do this before deleting any instructions from the loop, so that
973 back_branch_in_range_p will work correctly. */
975 if (splitting_not_safe
)
978 temp
= find_splittable_regs (unroll_type
, loop_start
, loop_end
,
979 end_insert_before
, unroll_number
);
981 /* find_splittable_regs may have created some new registers, so must
982 reallocate the reg_map with the new larger size, and must realloc
983 the constant maps also. */
985 maxregnum
= max_reg_num ();
986 map
->reg_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
988 init_reg_map (map
, maxregnum
);
990 /* Space is needed in some of the map for new registers, so new_maxregnum
991 is an (over)estimate of how many registers will exist at the end. */
992 new_maxregnum
= maxregnum
+ (temp
* unroll_number
* 2);
994 /* Must realloc space for the constant maps, because the number of registers
997 map
->const_equiv_map
= (rtx
*) alloca (new_maxregnum
* sizeof (rtx
));
998 map
->const_age_map
= (unsigned *) alloca (new_maxregnum
* sizeof (unsigned));
1000 global_const_equiv_map
= map
->const_equiv_map
;
1002 /* Search the list of bivs and givs to find ones which need to be remapped
1003 when split, and set their reg_map entry appropriately. */
1005 for (bl
= loop_iv_list
; bl
; bl
= bl
->next
)
1007 if (REGNO (bl
->biv
->src_reg
) != bl
->regno
)
1008 map
->reg_map
[bl
->regno
] = bl
->biv
->src_reg
;
1010 /* Currently, non-reduced/final-value givs are never split. */
1011 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
1012 if (REGNO (v
->src_reg
) != bl
->regno
)
1013 map
->reg_map
[REGNO (v
->dest_reg
)] = v
->src_reg
;
1017 /* If the loop is being partially unrolled, and the iteration variables
1018 are being split, and are being renamed for the split, then must fix up
1019 the compare instruction at the end of the loop to refer to the new
1020 registers. This compare isn't copied, so the registers used in it
1021 will never be replaced if it isn't done here. */
1023 if (unroll_type
== UNROLL_MODULO
)
1025 insn
= NEXT_INSN (copy_end
);
1026 if (GET_CODE (insn
) == INSN
&& GET_CODE (PATTERN (insn
)) == SET
)
1029 /* If non-reduced/final-value givs were split, then this would also
1030 have to remap those givs. */
1033 tem
= SET_SRC (PATTERN (insn
));
1034 /* The set source is a register. */
1035 if (GET_CODE (tem
) == REG
)
1037 if (REGNO (tem
) < max_reg_before_loop
1038 && reg_iv_type
[REGNO (tem
)] == BASIC_INDUCT
)
1039 SET_SRC (PATTERN (insn
))
1040 = reg_biv_class
[REGNO (tem
)]->biv
->src_reg
;
1044 /* The set source is a compare of some sort. */
1045 tem
= XEXP (SET_SRC (PATTERN (insn
)), 0);
1046 if (GET_CODE (tem
) == REG
1047 && REGNO (tem
) < max_reg_before_loop
1048 && reg_iv_type
[REGNO (tem
)] == BASIC_INDUCT
)
1049 XEXP (SET_SRC (PATTERN (insn
)), 0)
1050 = reg_biv_class
[REGNO (tem
)]->biv
->src_reg
;
1052 tem
= XEXP (SET_SRC (PATTERN (insn
)), 1);
1053 if (GET_CODE (tem
) == REG
1054 && REGNO (tem
) < max_reg_before_loop
1055 && reg_iv_type
[REGNO (tem
)] == BASIC_INDUCT
)
1056 XEXP (SET_SRC (PATTERN (insn
)), 1)
1057 = reg_biv_class
[REGNO (tem
)]->biv
->src_reg
;
1062 /* For unroll_number - 1 times, make a copy of each instruction
1063 between copy_start and copy_end, and insert these new instructions
1064 before the end of the loop. */
1066 for (i
= 0; i
< unroll_number
; i
++)
1068 bzero (map
->insn_map
, max_insnno
* sizeof (rtx
));
1069 bzero (map
->const_equiv_map
, new_maxregnum
* sizeof (rtx
));
1070 bzero (map
->const_age_map
, new_maxregnum
* sizeof (unsigned));
1073 for (j
= 0; j
< max_labelno
; j
++)
1075 map
->label_map
[j
] = gen_label_rtx ();
1077 /* If loop starts with a branch to the test, then fix it so that
1078 it points to the test of the first unrolled copy of the loop. */
1079 if (i
== 0 && loop_start
!= copy_start
)
1081 insn
= PREV_INSN (copy_start
);
1082 pattern
= PATTERN (insn
);
1084 tem
= map
->label_map
[CODE_LABEL_NUMBER
1085 (XEXP (SET_SRC (pattern
), 0))];
1086 SET_SRC (pattern
) = gen_rtx (LABEL_REF
, VOIDmode
, tem
);
1088 /* Set the jump label so that it can be used by later loop unrolling
1090 JUMP_LABEL (insn
) = tem
;
1091 LABEL_NUSES (tem
)++;
1094 copy_loop_body (copy_start
, copy_end
, map
, exit_label
,
1095 i
== unroll_number
- 1, unroll_type
, start_label
,
1096 loop_end
, insert_before
, insert_before
);
1099 /* Before deleting any insns, emit a CODE_LABEL immediately after the last
1100 insn to be deleted. This prevents any runaway delete_insn call from
1101 more insns that it should, as it always stops at a CODE_LABEL. */
1103 /* Delete the compare and branch at the end of the loop if completely
1104 unrolling the loop. Deleting the backward branch at the end also
1105 deletes the code label at the start of the loop. This is done at
1106 the very end to avoid problems with back_branch_in_range_p. */
1108 if (unroll_type
== UNROLL_COMPLETELY
)
1109 safety_label
= emit_label_after (gen_label_rtx (), last_loop_insn
);
1111 safety_label
= emit_label_after (gen_label_rtx (), copy_end
);
1113 /* Delete all of the original loop instructions. Don't delete the
1114 LOOP_BEG note, or the first code label in the loop. */
1116 insn
= NEXT_INSN (copy_start
);
1117 while (insn
!= safety_label
)
1119 if (insn
!= start_label
)
1120 insn
= delete_insn (insn
);
1122 insn
= NEXT_INSN (insn
);
1125 /* Can now delete the 'safety' label emitted to protect us from runaway
1126 delete_insn calls. */
1127 if (INSN_DELETED_P (safety_label
))
1129 delete_insn (safety_label
);
1131 /* If exit_label exists, emit it after the loop. Doing the emit here
1132 forces it to have a higher INSN_UID than any insn in the unrolled loop.
1133 This is needed so that mostly_true_jump in reorg.c will treat jumps
1134 to this loop end label correctly, i.e. predict that they are usually
1137 emit_label_after (exit_label
, loop_end
);
1140 /* Return true if the loop can be safely, and profitably, preconditioned
1141 so that the unrolled copies of the loop body don't need exit tests.
1143 This only works if final_value, initial_value and increment can be
1144 determined, and if increment is a constant power of 2.
1145 If increment is not a power of 2, then the preconditioning modulo
1146 operation would require a real modulo instead of a boolean AND, and this
1147 is not considered `profitable'. */
1149 /* ??? If the loop is known to be executed very many times, or the machine
1150 has a very cheap divide instruction, then preconditioning is a win even
1151 when the increment is not a power of 2. Use RTX_COST to compute
1152 whether divide is cheap. */
1155 precondition_loop_p (initial_value
, final_value
, increment
, loop_start
,
1157 rtx
*initial_value
, *final_value
, *increment
;
1158 rtx loop_start
, loop_end
;
1160 int unsigned_compare
, compare_dir
;
1162 if (loop_n_iterations
> 0)
1164 *initial_value
= const0_rtx
;
1165 *increment
= const1_rtx
;
1166 *final_value
= GEN_INT (loop_n_iterations
);
1168 if (loop_dump_stream
)
1169 fprintf (loop_dump_stream
,
1170 "Preconditioning: Success, number of iterations known, %d.\n",
1175 if (loop_initial_value
== 0)
1177 if (loop_dump_stream
)
1178 fprintf (loop_dump_stream
,
1179 "Preconditioning: Could not find initial value.\n");
1182 else if (loop_increment
== 0)
1184 if (loop_dump_stream
)
1185 fprintf (loop_dump_stream
,
1186 "Preconditioning: Could not find increment value.\n");
1189 else if (GET_CODE (loop_increment
) != CONST_INT
)
1191 if (loop_dump_stream
)
1192 fprintf (loop_dump_stream
,
1193 "Preconditioning: Increment not a constant.\n");
1196 else if ((exact_log2 (INTVAL (loop_increment
)) < 0)
1197 && (exact_log2 (- INTVAL (loop_increment
)) < 0))
1199 if (loop_dump_stream
)
1200 fprintf (loop_dump_stream
,
1201 "Preconditioning: Increment not a constant power of 2.\n");
1205 /* Unsigned_compare and compare_dir can be ignored here, since they do
1206 not matter for preconditioning. */
1208 if (loop_final_value
== 0)
1210 if (loop_dump_stream
)
1211 fprintf (loop_dump_stream
,
1212 "Preconditioning: EQ comparison loop.\n");
1216 /* Must ensure that final_value is invariant, so call invariant_p to
1217 check. Before doing so, must check regno against max_reg_before_loop
1218 to make sure that the register is in the range covered by invariant_p.
1219 If it isn't, then it is most likely a biv/giv which by definition are
1221 if ((GET_CODE (loop_final_value
) == REG
1222 && REGNO (loop_final_value
) >= max_reg_before_loop
)
1223 || (GET_CODE (loop_final_value
) == PLUS
1224 && REGNO (XEXP (loop_final_value
, 0)) >= max_reg_before_loop
)
1225 || ! invariant_p (loop_final_value
))
1227 if (loop_dump_stream
)
1228 fprintf (loop_dump_stream
,
1229 "Preconditioning: Final value not invariant.\n");
1233 /* Fail for floating point values, since the caller of this function
1234 does not have code to deal with them. */
1235 if (GET_MODE_CLASS (GET_MODE (loop_final_value
)) == MODE_FLOAT
1236 || GET_MODE_CLASS (GET_MODE (loop_initial_value
)) == MODE_FLOAT
)
1238 if (loop_dump_stream
)
1239 fprintf (loop_dump_stream
,
1240 "Preconditioning: Floating point final or initial value.\n");
1244 /* Now set initial_value to be the iteration_var, since that may be a
1245 simpler expression, and is guaranteed to be correct if all of the
1246 above tests succeed.
1248 We can not use the initial_value as calculated, because it will be
1249 one too small for loops of the form "while (i-- > 0)". We can not
1250 emit code before the loop_skip_over insns to fix this problem as this
1251 will then give a number one too large for loops of the form
1254 Note that all loops that reach here are entered at the top, because
1255 this function is not called if the loop starts with a jump. */
1257 /* Fail if loop_iteration_var is not live before loop_start, since we need
1258 to test its value in the preconditioning code. */
1260 if (uid_luid
[regno_first_uid
[REGNO (loop_iteration_var
)]]
1261 > INSN_LUID (loop_start
))
1263 if (loop_dump_stream
)
1264 fprintf (loop_dump_stream
,
1265 "Preconditioning: Iteration var not live before loop start.\n");
1269 *initial_value
= loop_iteration_var
;
1270 *increment
= loop_increment
;
1271 *final_value
= loop_final_value
;
1274 if (loop_dump_stream
)
1275 fprintf (loop_dump_stream
, "Preconditioning: Successful.\n");
1280 /* All pseudo-registers must be mapped to themselves. Two hard registers
1281 must be mapped, VIRTUAL_STACK_VARS_REGNUM and VIRTUAL_INCOMING_ARGS_
1282 REGNUM, to avoid function-inlining specific conversions of these
1283 registers. All other hard regs can not be mapped because they may be
1288 init_reg_map (map
, maxregnum
)
1289 struct inline_remap
*map
;
1294 for (i
= maxregnum
- 1; i
> LAST_VIRTUAL_REGISTER
; i
--)
1295 map
->reg_map
[i
] = regno_reg_rtx
[i
];
1296 /* Just clear the rest of the entries. */
1297 for (i
= LAST_VIRTUAL_REGISTER
; i
>= 0; i
--)
1298 map
->reg_map
[i
] = 0;
1300 map
->reg_map
[VIRTUAL_STACK_VARS_REGNUM
]
1301 = regno_reg_rtx
[VIRTUAL_STACK_VARS_REGNUM
];
1302 map
->reg_map
[VIRTUAL_INCOMING_ARGS_REGNUM
]
1303 = regno_reg_rtx
[VIRTUAL_INCOMING_ARGS_REGNUM
];
1306 /* Strength-reduction will often emit code for optimized biv/givs which
1307 calculates their value in a temporary register, and then copies the result
1308 to the iv. This procedure reconstructs the pattern computing the iv;
1309 verifying that all operands are of the proper form.
1311 The return value is the amount that the giv is incremented by. */
1314 calculate_giv_inc (pattern
, src_insn
, regno
)
1315 rtx pattern
, src_insn
;
1320 /* Verify that we have an increment insn here. First check for a plus
1321 as the set source. */
1322 if (GET_CODE (SET_SRC (pattern
)) != PLUS
)
1324 /* SR sometimes computes the new giv value in a temp, then copies it
1326 src_insn
= PREV_INSN (src_insn
);
1327 pattern
= PATTERN (src_insn
);
1328 if (GET_CODE (SET_SRC (pattern
)) != PLUS
)
1331 /* The last insn emitted is not needed, so delete it to avoid confusing
1332 the second cse pass. This insn sets the giv unnecessarily. */
1333 delete_insn (get_last_insn ());
1336 /* Verify that we have a constant as the second operand of the plus. */
1337 increment
= XEXP (SET_SRC (pattern
), 1);
1338 if (GET_CODE (increment
) != CONST_INT
)
1340 /* SR sometimes puts the constant in a register, especially if it is
1341 too big to be an add immed operand. */
1342 increment
= SET_SRC (PATTERN (PREV_INSN (src_insn
)));
1344 /* SR may have used LO_SUM to compute the constant if it is too large
1345 for a load immed operand. In this case, the constant is in operand
1346 one of the LO_SUM rtx. */
1347 if (GET_CODE (increment
) == LO_SUM
)
1348 increment
= XEXP (increment
, 1);
1350 if (GET_CODE (increment
) != CONST_INT
)
1353 /* The insn loading the constant into a register is not longer needed,
1355 delete_insn (get_last_insn ());
1358 /* Check that the source register is the same as the dest register. */
1359 if (GET_CODE (XEXP (SET_SRC (pattern
), 0)) != REG
1360 || REGNO (XEXP (SET_SRC (pattern
), 0)) != regno
)
1367 /* Copy each instruction in the loop, substituting from map as appropriate.
1368 This is very similar to a loop in expand_inline_function. */
1371 copy_loop_body (copy_start
, copy_end
, map
, exit_label
, last_iteration
,
1372 unroll_type
, start_label
, loop_end
, insert_before
,
1374 rtx copy_start
, copy_end
;
1375 struct inline_remap
*map
;
1378 enum unroll_types unroll_type
;
1379 rtx start_label
, loop_end
, insert_before
, copy_notes_from
;
1383 int dest_reg_was_split
, i
;
1385 rtx final_label
= 0;
1386 rtx giv_inc
, giv_dest_reg
, giv_src_reg
;
1388 /* If this isn't the last iteration, then map any references to the
1389 start_label to final_label. Final label will then be emitted immediately
1390 after the end of this loop body if it was ever used.
1392 If this is the last iteration, then map references to the start_label
1394 if (! last_iteration
)
1396 final_label
= gen_label_rtx ();
1397 map
->label_map
[CODE_LABEL_NUMBER (start_label
)] = final_label
;
1400 map
->label_map
[CODE_LABEL_NUMBER (start_label
)] = start_label
;
1407 insn
= NEXT_INSN (insn
);
1409 map
->orig_asm_operands_vector
= 0;
1411 switch (GET_CODE (insn
))
1414 pattern
= PATTERN (insn
);
1418 /* Check to see if this is a giv that has been combined with
1419 some split address givs. (Combined in the sense that
1420 `combine_givs' in loop.c has put two givs in the same register.)
1421 In this case, we must search all givs based on the same biv to
1422 find the address givs. Then split the address givs.
1423 Do this before splitting the giv, since that may map the
1424 SET_DEST to a new register. */
1426 if (GET_CODE (pattern
) == SET
1427 && GET_CODE (SET_DEST (pattern
)) == REG
1428 && addr_combined_regs
[REGNO (SET_DEST (pattern
))])
1430 struct iv_class
*bl
;
1431 struct induction
*v
, *tv
;
1432 int regno
= REGNO (SET_DEST (pattern
));
1434 v
= addr_combined_regs
[REGNO (SET_DEST (pattern
))];
1435 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
1437 /* Although the giv_inc amount is not needed here, we must call
1438 calculate_giv_inc here since it might try to delete the
1439 last insn emitted. If we wait until later to call it,
1440 we might accidentally delete insns generated immediately
1441 below by emit_unrolled_add. */
1443 giv_inc
= calculate_giv_inc (pattern
, insn
, regno
);
1445 /* Now find all address giv's that were combined with this
1447 for (tv
= bl
->giv
; tv
; tv
= tv
->next_iv
)
1448 if (tv
->giv_type
== DEST_ADDR
&& tv
->same
== v
)
1450 int this_giv_inc
= INTVAL (giv_inc
);
1452 /* Scale this_giv_inc if the multiplicative factors of
1453 the two givs are different. */
1454 if (tv
->mult_val
!= v
->mult_val
)
1455 this_giv_inc
= (this_giv_inc
/ INTVAL (v
->mult_val
)
1456 * INTVAL (tv
->mult_val
));
1458 tv
->dest_reg
= plus_constant (tv
->dest_reg
, this_giv_inc
);
1459 *tv
->location
= tv
->dest_reg
;
1461 if (last_iteration
&& unroll_type
!= UNROLL_COMPLETELY
)
1463 /* Must emit an insn to increment the split address
1464 giv. Add in the const_adjust field in case there
1465 was a constant eliminated from the address. */
1466 rtx value
, dest_reg
;
1468 /* tv->dest_reg will be either a bare register,
1469 or else a register plus a constant. */
1470 if (GET_CODE (tv
->dest_reg
) == REG
)
1471 dest_reg
= tv
->dest_reg
;
1473 dest_reg
= XEXP (tv
->dest_reg
, 0);
1475 /* tv->dest_reg may actually be a (PLUS (REG) (CONST))
1476 here, so we must call plus_constant to add
1477 the const_adjust amount before calling
1478 emit_unrolled_add below. */
1479 value
= plus_constant (tv
->dest_reg
, tv
->const_adjust
);
1481 /* The constant could be too large for an add
1482 immediate, so can't directly emit an insn here. */
1483 emit_unrolled_add (dest_reg
, XEXP (value
, 0),
1486 /* Reset the giv to be just the register again, in case
1487 it is used after the set we have just emitted.
1488 We must subtract the const_adjust factor added in
1490 tv
->dest_reg
= plus_constant (dest_reg
,
1491 - tv
->const_adjust
);
1492 *tv
->location
= tv
->dest_reg
;
1497 /* If this is a setting of a splittable variable, then determine
1498 how to split the variable, create a new set based on this split,
1499 and set up the reg_map so that later uses of the variable will
1500 use the new split variable. */
1502 dest_reg_was_split
= 0;
1504 if (GET_CODE (pattern
) == SET
1505 && GET_CODE (SET_DEST (pattern
)) == REG
1506 && splittable_regs
[REGNO (SET_DEST (pattern
))])
1508 int regno
= REGNO (SET_DEST (pattern
));
1510 dest_reg_was_split
= 1;
1512 /* Compute the increment value for the giv, if it wasn't
1513 already computed above. */
1516 giv_inc
= calculate_giv_inc (pattern
, insn
, regno
);
1517 giv_dest_reg
= SET_DEST (pattern
);
1518 giv_src_reg
= SET_DEST (pattern
);
1520 if (unroll_type
== UNROLL_COMPLETELY
)
1522 /* Completely unrolling the loop. Set the induction
1523 variable to a known constant value. */
1525 /* The value in splittable_regs may be an invariant
1526 value, so we must use plus_constant here. */
1527 splittable_regs
[regno
]
1528 = plus_constant (splittable_regs
[regno
], INTVAL (giv_inc
));
1530 if (GET_CODE (splittable_regs
[regno
]) == PLUS
)
1532 giv_src_reg
= XEXP (splittable_regs
[regno
], 0);
1533 giv_inc
= XEXP (splittable_regs
[regno
], 1);
1537 /* The splittable_regs value must be a REG or a
1538 CONST_INT, so put the entire value in the giv_src_reg
1540 giv_src_reg
= splittable_regs
[regno
];
1541 giv_inc
= const0_rtx
;
1546 /* Partially unrolling loop. Create a new pseudo
1547 register for the iteration variable, and set it to
1548 be a constant plus the original register. Except
1549 on the last iteration, when the result has to
1550 go back into the original iteration var register. */
1552 /* Handle bivs which must be mapped to a new register
1553 when split. This happens for bivs which need their
1554 final value set before loop entry. The new register
1555 for the biv was stored in the biv's first struct
1556 induction entry by find_splittable_regs. */
1558 if (regno
< max_reg_before_loop
1559 && reg_iv_type
[regno
] == BASIC_INDUCT
)
1561 giv_src_reg
= reg_biv_class
[regno
]->biv
->src_reg
;
1562 giv_dest_reg
= giv_src_reg
;
1566 /* If non-reduced/final-value givs were split, then
1567 this would have to remap those givs also. See
1568 find_splittable_regs. */
1571 splittable_regs
[regno
]
1572 = GEN_INT (INTVAL (giv_inc
)
1573 + INTVAL (splittable_regs
[regno
]));
1574 giv_inc
= splittable_regs
[regno
];
1576 /* Now split the induction variable by changing the dest
1577 of this insn to a new register, and setting its
1578 reg_map entry to point to this new register.
1580 If this is the last iteration, and this is the last insn
1581 that will update the iv, then reuse the original dest,
1582 to ensure that the iv will have the proper value when
1583 the loop exits or repeats.
1585 Using splittable_regs_updates here like this is safe,
1586 because it can only be greater than one if all
1587 instructions modifying the iv are always executed in
1590 if (! last_iteration
1591 || (splittable_regs_updates
[regno
]-- != 1))
1593 tem
= gen_reg_rtx (GET_MODE (giv_src_reg
));
1595 map
->reg_map
[regno
] = tem
;
1598 map
->reg_map
[regno
] = giv_src_reg
;
1601 /* The constant being added could be too large for an add
1602 immediate, so can't directly emit an insn here. */
1603 emit_unrolled_add (giv_dest_reg
, giv_src_reg
, giv_inc
);
1604 copy
= get_last_insn ();
1605 pattern
= PATTERN (copy
);
1609 pattern
= copy_rtx_and_substitute (pattern
, map
);
1610 copy
= emit_insn (pattern
);
1612 /* REG_NOTES will be copied later. */
1615 /* If this insn is setting CC0, it may need to look at
1616 the insn that uses CC0 to see what type of insn it is.
1617 In that case, the call to recog via validate_change will
1618 fail. So don't substitute constants here. Instead,
1619 do it when we emit the following insn.
1621 For example, see the pyr.md file. That machine has signed and
1622 unsigned compares. The compare patterns must check the
1623 following branch insn to see which what kind of compare to
1626 If the previous insn set CC0, substitute constants on it as
1628 if (sets_cc0_p (copy
) != 0)
1633 try_constants (cc0_insn
, map
);
1635 try_constants (copy
, map
);
1638 try_constants (copy
, map
);
1641 /* Make split induction variable constants `permanent' since we
1642 know there are no backward branches across iteration variable
1643 settings which would invalidate this. */
1644 if (dest_reg_was_split
)
1646 int regno
= REGNO (SET_DEST (pattern
));
1648 if (map
->const_age_map
[regno
] == map
->const_age
)
1649 map
->const_age_map
[regno
] = -1;
1654 pattern
= copy_rtx_and_substitute (PATTERN (insn
), map
);
1655 copy
= emit_jump_insn (pattern
);
1657 if (JUMP_LABEL (insn
) == start_label
&& insn
== copy_end
1658 && ! last_iteration
)
1660 /* This is a branch to the beginning of the loop; this is the
1661 last insn being copied; and this is not the last iteration.
1662 In this case, we want to change the original fall through
1663 case to be a branch past the end of the loop, and the
1664 original jump label case to fall_through. */
1666 if (! invert_exp (pattern
, copy
)
1667 || ! redirect_exp (&pattern
,
1668 map
->label_map
[CODE_LABEL_NUMBER
1669 (JUMP_LABEL (insn
))],
1676 try_constants (cc0_insn
, map
);
1679 try_constants (copy
, map
);
1681 /* Set the jump label of COPY correctly to avoid problems with
1682 later passes of unroll_loop, if INSN had jump label set. */
1683 if (JUMP_LABEL (insn
))
1687 /* Can't use the label_map for every insn, since this may be
1688 the backward branch, and hence the label was not mapped. */
1689 if (GET_CODE (pattern
) == SET
)
1691 tem
= SET_SRC (pattern
);
1692 if (GET_CODE (tem
) == LABEL_REF
)
1693 label
= XEXP (tem
, 0);
1694 else if (GET_CODE (tem
) == IF_THEN_ELSE
)
1696 if (XEXP (tem
, 1) != pc_rtx
)
1697 label
= XEXP (XEXP (tem
, 1), 0);
1699 label
= XEXP (XEXP (tem
, 2), 0);
1703 if (label
&& GET_CODE (label
) == CODE_LABEL
)
1704 JUMP_LABEL (copy
) = label
;
1707 /* An unrecognizable jump insn, probably the entry jump
1708 for a switch statement. This label must have been mapped,
1709 so just use the label_map to get the new jump label. */
1710 JUMP_LABEL (copy
) = map
->label_map
[CODE_LABEL_NUMBER
1711 (JUMP_LABEL (insn
))];
1714 /* If this is a non-local jump, then must increase the label
1715 use count so that the label will not be deleted when the
1716 original jump is deleted. */
1717 LABEL_NUSES (JUMP_LABEL (copy
))++;
1719 else if (GET_CODE (PATTERN (copy
)) == ADDR_VEC
1720 || GET_CODE (PATTERN (copy
)) == ADDR_DIFF_VEC
)
1722 rtx pat
= PATTERN (copy
);
1723 int diff_vec_p
= GET_CODE (pat
) == ADDR_DIFF_VEC
;
1724 int len
= XVECLEN (pat
, diff_vec_p
);
1727 for (i
= 0; i
< len
; i
++)
1728 LABEL_NUSES (XEXP (XVECEXP (pat
, diff_vec_p
, i
), 0))++;
1731 /* If this used to be a conditional jump insn but whose branch
1732 direction is now known, we must do something special. */
1733 if (condjump_p (insn
) && !simplejump_p (insn
) && map
->last_pc_value
)
1736 /* The previous insn set cc0 for us. So delete it. */
1737 delete_insn (PREV_INSN (copy
));
1740 /* If this is now a no-op, delete it. */
1741 if (map
->last_pc_value
== pc_rtx
)
1747 /* Otherwise, this is unconditional jump so we must put a
1748 BARRIER after it. We could do some dead code elimination
1749 here, but jump.c will do it just as well. */
1755 pattern
= copy_rtx_and_substitute (PATTERN (insn
), map
);
1756 copy
= emit_call_insn (pattern
);
1760 try_constants (cc0_insn
, map
);
1763 try_constants (copy
, map
);
1765 /* Be lazy and assume CALL_INSNs clobber all hard registers. */
1766 for (i
= 0; i
< FIRST_PSEUDO_REGISTER
; i
++)
1767 map
->const_equiv_map
[i
] = 0;
1771 /* If this is the loop start label, then we don't need to emit a
1772 copy of this label since no one will use it. */
1774 if (insn
!= start_label
)
1776 copy
= emit_label (map
->label_map
[CODE_LABEL_NUMBER (insn
)]);
1782 copy
= emit_barrier ();
1786 /* VTOP notes are valid only before the loop exit test. If placed
1787 anywhere else, loop may generate bad code. */
1789 if (NOTE_LINE_NUMBER (insn
) != NOTE_INSN_DELETED
1790 && (NOTE_LINE_NUMBER (insn
) != NOTE_INSN_LOOP_VTOP
1791 || (last_iteration
&& unroll_type
!= UNROLL_COMPLETELY
)))
1792 copy
= emit_note (NOTE_SOURCE_FILE (insn
),
1793 NOTE_LINE_NUMBER (insn
));
1803 map
->insn_map
[INSN_UID (insn
)] = copy
;
1805 while (insn
!= copy_end
);
1807 /* Now copy the REG_NOTES. */
1811 insn
= NEXT_INSN (insn
);
1812 if ((GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == JUMP_INSN
1813 || GET_CODE (insn
) == CALL_INSN
)
1814 && map
->insn_map
[INSN_UID (insn
)])
1815 REG_NOTES (map
->insn_map
[INSN_UID (insn
)])
1816 = copy_rtx_and_substitute (REG_NOTES (insn
), map
);
1818 while (insn
!= copy_end
);
1820 /* There may be notes between copy_notes_from and loop_end. Emit a copy of
1821 each of these notes here, since there may be some important ones, such as
1822 NOTE_INSN_BLOCK_END notes, in this group. We don't do this on the last
1823 iteration, because the original notes won't be deleted.
1825 We can't use insert_before here, because when from preconditioning,
1826 insert_before points before the loop. We can't use copy_end, because
1827 there may be insns already inserted after it (which we don't want to
1828 copy) when not from preconditioning code. */
1830 if (! last_iteration
)
1832 for (insn
= copy_notes_from
; insn
!= loop_end
; insn
= NEXT_INSN (insn
))
1834 if (GET_CODE (insn
) == NOTE
1835 && NOTE_LINE_NUMBER (insn
) != NOTE_INSN_DELETED
)
1836 emit_note (NOTE_SOURCE_FILE (insn
), NOTE_LINE_NUMBER (insn
));
1840 if (final_label
&& LABEL_NUSES (final_label
) > 0)
1841 emit_label (final_label
);
1843 tem
= gen_sequence ();
1845 emit_insn_before (tem
, insert_before
);
1848 /* Emit an insn, using the expand_binop to ensure that a valid insn is
1849 emitted. This will correctly handle the case where the increment value
1850 won't fit in the immediate field of a PLUS insns. */
1853 emit_unrolled_add (dest_reg
, src_reg
, increment
)
1854 rtx dest_reg
, src_reg
, increment
;
1858 result
= expand_binop (GET_MODE (dest_reg
), add_optab
, src_reg
, increment
,
1859 dest_reg
, 0, OPTAB_LIB_WIDEN
);
1861 if (dest_reg
!= result
)
1862 emit_move_insn (dest_reg
, result
);
1865 /* Searches the insns between INSN and LOOP_END. Returns 1 if there
1866 is a backward branch in that range that branches to somewhere between
1867 LOOP_START and INSN. Returns 0 otherwise. */
1869 /* ??? This is quadratic algorithm. Could be rewritten to be linear.
1870 In practice, this is not a problem, because this function is seldom called,
1871 and uses a negligible amount of CPU time on average. */
1874 back_branch_in_range_p (insn
, loop_start
, loop_end
)
1876 rtx loop_start
, loop_end
;
1878 rtx p
, q
, target_insn
;
1880 /* Stop before we get to the backward branch at the end of the loop. */
1881 loop_end
= prev_nonnote_insn (loop_end
);
1882 if (GET_CODE (loop_end
) == BARRIER
)
1883 loop_end
= PREV_INSN (loop_end
);
1885 /* Check in case insn has been deleted, search forward for first non
1886 deleted insn following it. */
1887 while (INSN_DELETED_P (insn
))
1888 insn
= NEXT_INSN (insn
);
1890 /* Check for the case where insn is the last insn in the loop. */
1891 if (insn
== loop_end
)
1894 for (p
= NEXT_INSN (insn
); p
!= loop_end
; p
= NEXT_INSN (p
))
1896 if (GET_CODE (p
) == JUMP_INSN
)
1898 target_insn
= JUMP_LABEL (p
);
1900 /* Search from loop_start to insn, to see if one of them is
1901 the target_insn. We can't use INSN_LUID comparisons here,
1902 since insn may not have an LUID entry. */
1903 for (q
= loop_start
; q
!= insn
; q
= NEXT_INSN (q
))
1904 if (q
== target_insn
)
1912 /* Try to generate the simplest rtx for the expression
1913 (PLUS (MULT mult1 mult2) add1). This is used to calculate the initial
1917 fold_rtx_mult_add (mult1
, mult2
, add1
, mode
)
1918 rtx mult1
, mult2
, add1
;
1919 enum machine_mode mode
;
1924 /* The modes must all be the same. This should always be true. For now,
1925 check to make sure. */
1926 if ((GET_MODE (mult1
) != mode
&& GET_MODE (mult1
) != VOIDmode
)
1927 || (GET_MODE (mult2
) != mode
&& GET_MODE (mult2
) != VOIDmode
)
1928 || (GET_MODE (add1
) != mode
&& GET_MODE (add1
) != VOIDmode
))
1931 /* Ensure that if at least one of mult1/mult2 are constant, then mult2
1932 will be a constant. */
1933 if (GET_CODE (mult1
) == CONST_INT
)
1940 mult_res
= simplify_binary_operation (MULT
, mode
, mult1
, mult2
);
1942 mult_res
= gen_rtx (MULT
, mode
, mult1
, mult2
);
1944 /* Again, put the constant second. */
1945 if (GET_CODE (add1
) == CONST_INT
)
1952 result
= simplify_binary_operation (PLUS
, mode
, add1
, mult_res
);
1954 result
= gen_rtx (PLUS
, mode
, add1
, mult_res
);
1959 /* Searches the list of induction struct's for the biv BL, to try to calculate
1960 the total increment value for one iteration of the loop as a constant.
1962 Returns the increment value as an rtx, simplified as much as possible,
1963 if it can be calculated. Otherwise, returns 0. */
1966 biv_total_increment (bl
, loop_start
, loop_end
)
1967 struct iv_class
*bl
;
1968 rtx loop_start
, loop_end
;
1970 struct induction
*v
;
1973 /* For increment, must check every instruction that sets it. Each
1974 instruction must be executed only once each time through the loop.
1975 To verify this, we check that the the insn is always executed, and that
1976 there are no backward branches after the insn that branch to before it.
1977 Also, the insn must have a mult_val of one (to make sure it really is
1980 result
= const0_rtx
;
1981 for (v
= bl
->biv
; v
; v
= v
->next_iv
)
1983 if (v
->always_computable
&& v
->mult_val
== const1_rtx
1984 && ! back_branch_in_range_p (v
->insn
, loop_start
, loop_end
))
1985 result
= fold_rtx_mult_add (result
, const1_rtx
, v
->add_val
, v
->mode
);
1993 /* Determine the initial value of the iteration variable, and the amount
1994 that it is incremented each loop. Use the tables constructed by
1995 the strength reduction pass to calculate these values.
1997 Initial_value and/or increment are set to zero if their values could not
2001 iteration_info (iteration_var
, initial_value
, increment
, loop_start
, loop_end
)
2002 rtx iteration_var
, *initial_value
, *increment
;
2003 rtx loop_start
, loop_end
;
2005 struct iv_class
*bl
;
2006 struct induction
*v
, *b
;
2008 /* Clear the result values, in case no answer can be found. */
2012 /* The iteration variable can be either a giv or a biv. Check to see
2013 which it is, and compute the variable's initial value, and increment
2014 value if possible. */
2016 /* If this is a new register, can't handle it since we don't have any
2017 reg_iv_type entry for it. */
2018 if (REGNO (iteration_var
) >= max_reg_before_loop
)
2020 if (loop_dump_stream
)
2021 fprintf (loop_dump_stream
,
2022 "Loop unrolling: No reg_iv_type entry for iteration var.\n");
2025 /* Reject iteration variables larger than the host long size, since they
2026 could result in a number of iterations greater than the range of our
2027 `unsigned long' variable loop_n_iterations. */
2028 else if (GET_MODE_BITSIZE (GET_MODE (iteration_var
)) > HOST_BITS_PER_LONG
)
2030 if (loop_dump_stream
)
2031 fprintf (loop_dump_stream
,
2032 "Loop unrolling: Iteration var rejected because mode larger than host long.\n");
2035 else if (GET_MODE_CLASS (GET_MODE (iteration_var
)) != MODE_INT
)
2037 if (loop_dump_stream
)
2038 fprintf (loop_dump_stream
,
2039 "Loop unrolling: Iteration var not an integer.\n");
2042 else if (reg_iv_type
[REGNO (iteration_var
)] == BASIC_INDUCT
)
2044 /* Grab initial value, only useful if it is a constant. */
2045 bl
= reg_biv_class
[REGNO (iteration_var
)];
2046 *initial_value
= bl
->initial_value
;
2048 *increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2050 else if (reg_iv_type
[REGNO (iteration_var
)] == GENERAL_INDUCT
)
2053 /* ??? The code below does not work because the incorrect number of
2054 iterations is calculated when the biv is incremented after the giv
2055 is set (which is the usual case). This can probably be accounted
2056 for by biasing the initial_value by subtracting the amount of the
2057 increment that occurs between the giv set and the giv test. However,
2058 a giv as an iterator is very rare, so it does not seem worthwhile
2060 /* ??? An example failure is: i = 6; do {;} while (i++ < 9). */
2061 if (loop_dump_stream
)
2062 fprintf (loop_dump_stream
,
2063 "Loop unrolling: Giv iterators are not handled.\n");
2066 /* Initial value is mult_val times the biv's initial value plus
2067 add_val. Only useful if it is a constant. */
2068 v
= reg_iv_info
[REGNO (iteration_var
)];
2069 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
2070 *initial_value
= fold_rtx_mult_add (v
->mult_val
, bl
->initial_value
,
2071 v
->add_val
, v
->mode
);
2073 /* Increment value is mult_val times the increment value of the biv. */
2075 *increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2077 *increment
= fold_rtx_mult_add (v
->mult_val
, *increment
, const0_rtx
,
2083 if (loop_dump_stream
)
2084 fprintf (loop_dump_stream
,
2085 "Loop unrolling: Not basic or general induction var.\n");
2090 /* Calculate the approximate final value of the iteration variable
2091 which has an loop exit test with code COMPARISON_CODE and comparison value
2092 of COMPARISON_VALUE. Also returns an indication of whether the comparison
2093 was signed or unsigned, and the direction of the comparison. This info is
2094 needed to calculate the number of loop iterations. */
2097 approx_final_value (comparison_code
, comparison_value
, unsigned_p
, compare_dir
)
2098 enum rtx_code comparison_code
;
2099 rtx comparison_value
;
2103 /* Calculate the final value of the induction variable.
2104 The exact final value depends on the branch operator, and increment sign.
2105 This is only an approximate value. It will be wrong if the iteration
2106 variable is not incremented by one each time through the loop, and
2107 approx final value - start value % increment != 0. */
2110 switch (comparison_code
)
2116 return plus_constant (comparison_value
, 1);
2121 return plus_constant (comparison_value
, -1);
2123 /* Can not calculate a final value for this case. */
2130 return comparison_value
;
2136 return comparison_value
;
2139 return comparison_value
;
2145 /* For each biv and giv, determine whether it can be safely split into
2146 a different variable for each unrolled copy of the loop body. If it
2147 is safe to split, then indicate that by saving some useful info
2148 in the splittable_regs array.
2150 If the loop is being completely unrolled, then splittable_regs will hold
2151 the current value of the induction variable while the loop is unrolled.
2152 It must be set to the initial value of the induction variable here.
2153 Otherwise, splittable_regs will hold the difference between the current
2154 value of the induction variable and the value the induction variable had
2155 at the top of the loop. It must be set to the value 0 here. */
2157 /* ?? If the loop is only unrolled twice, then most of the restrictions to
2158 constant values are unnecessary, since we can easily calculate increment
2159 values in this case even if nothing is constant. The increment value
2160 should not involve a multiply however. */
2162 /* ?? Even if the biv/giv increment values aren't constant, it may still
2163 be beneficial to split the variable if the loop is only unrolled a few
2164 times, since multiplies by small integers (1,2,3,4) are very cheap. */
2167 find_splittable_regs (unroll_type
, loop_start
, loop_end
, end_insert_before
,
2169 enum unroll_types unroll_type
;
2170 rtx loop_start
, loop_end
;
2171 rtx end_insert_before
;
2174 struct iv_class
*bl
;
2175 struct induction
*v
;
2177 rtx biv_final_value
;
2181 for (bl
= loop_iv_list
; bl
; bl
= bl
->next
)
2183 /* Biv_total_increment must return a constant value,
2184 otherwise we can not calculate the split values. */
2186 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2187 if (! increment
|| GET_CODE (increment
) != CONST_INT
)
2190 /* The loop must be unrolled completely, or else have a known number
2191 of iterations and only one exit, or else the biv must be dead
2192 outside the loop, or else the final value must be known. Otherwise,
2193 it is unsafe to split the biv since it may not have the proper
2194 value on loop exit. */
2196 /* loop_number_exit_labels is non-zero if the loop has an exit other than
2197 a fall through at the end. */
2200 biv_final_value
= 0;
2201 if (unroll_type
!= UNROLL_COMPLETELY
2202 && (loop_number_exit_labels
[uid_loop_num
[INSN_UID (loop_start
)]]
2203 || unroll_type
== UNROLL_NAIVE
)
2204 && (uid_luid
[regno_last_uid
[bl
->regno
]] >= INSN_LUID (loop_end
)
2206 || INSN_UID (bl
->init_insn
) >= max_uid_for_loop
2207 || (uid_luid
[regno_first_uid
[bl
->regno
]]
2208 < INSN_LUID (bl
->init_insn
))
2209 || reg_mentioned_p (bl
->biv
->dest_reg
, SET_SRC (bl
->init_set
)))
2210 && ! (biv_final_value
= final_biv_value (bl
, loop_start
, loop_end
)))
2213 /* If any of the insns setting the BIV don't do so with a simple
2214 PLUS, we don't know how to split it. */
2215 for (v
= bl
->biv
; biv_splittable
&& v
; v
= v
->next_iv
)
2216 if ((tem
= single_set (v
->insn
)) == 0
2217 || GET_CODE (SET_DEST (tem
)) != REG
2218 || REGNO (SET_DEST (tem
)) != bl
->regno
2219 || GET_CODE (SET_SRC (tem
)) != PLUS
)
2222 /* If final value is non-zero, then must emit an instruction which sets
2223 the value of the biv to the proper value. This is done after
2224 handling all of the givs, since some of them may need to use the
2225 biv's value in their initialization code. */
2227 /* This biv is splittable. If completely unrolling the loop, save
2228 the biv's initial value. Otherwise, save the constant zero. */
2230 if (biv_splittable
== 1)
2232 if (unroll_type
== UNROLL_COMPLETELY
)
2234 /* If the initial value of the biv is itself (i.e. it is too
2235 complicated for strength_reduce to compute), or is a hard
2236 register, then we must create a new pseudo reg to hold the
2237 initial value of the biv. */
2239 if (GET_CODE (bl
->initial_value
) == REG
2240 && (REGNO (bl
->initial_value
) == bl
->regno
2241 || REGNO (bl
->initial_value
) < FIRST_PSEUDO_REGISTER
))
2243 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2245 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2248 if (loop_dump_stream
)
2249 fprintf (loop_dump_stream
, "Biv %d initial value remapped to %d.\n",
2250 bl
->regno
, REGNO (tem
));
2252 splittable_regs
[bl
->regno
] = tem
;
2255 splittable_regs
[bl
->regno
] = bl
->initial_value
;
2258 splittable_regs
[bl
->regno
] = const0_rtx
;
2260 /* Save the number of instructions that modify the biv, so that
2261 we can treat the last one specially. */
2263 splittable_regs_updates
[bl
->regno
] = bl
->biv_count
;
2267 if (loop_dump_stream
)
2268 fprintf (loop_dump_stream
,
2269 "Biv %d safe to split.\n", bl
->regno
);
2272 /* Check every giv that depends on this biv to see whether it is
2273 splittable also. Even if the biv isn't splittable, givs which
2274 depend on it may be splittable if the biv is live outside the
2275 loop, and the givs aren't. */
2277 result
= find_splittable_givs (bl
, unroll_type
, loop_start
, loop_end
,
2278 increment
, unroll_number
, result
);
2280 /* If final value is non-zero, then must emit an instruction which sets
2281 the value of the biv to the proper value. This is done after
2282 handling all of the givs, since some of them may need to use the
2283 biv's value in their initialization code. */
2284 if (biv_final_value
)
2286 /* If the loop has multiple exits, emit the insns before the
2287 loop to ensure that it will always be executed no matter
2288 how the loop exits. Otherwise emit the insn after the loop,
2289 since this is slightly more efficient. */
2290 if (! loop_number_exit_labels
[uid_loop_num
[INSN_UID (loop_start
)]])
2291 emit_insn_before (gen_move_insn (bl
->biv
->src_reg
,
2296 /* Create a new register to hold the value of the biv, and then
2297 set the biv to its final value before the loop start. The biv
2298 is set to its final value before loop start to ensure that
2299 this insn will always be executed, no matter how the loop
2301 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2302 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2304 emit_insn_before (gen_move_insn (bl
->biv
->src_reg
,
2308 if (loop_dump_stream
)
2309 fprintf (loop_dump_stream
, "Biv %d mapped to %d for split.\n",
2310 REGNO (bl
->biv
->src_reg
), REGNO (tem
));
2312 /* Set up the mapping from the original biv register to the new
2314 bl
->biv
->src_reg
= tem
;
2321 /* For every giv based on the biv BL, check to determine whether it is
2322 splittable. This is a subroutine to find_splittable_regs (). */
2325 find_splittable_givs (bl
, unroll_type
, loop_start
, loop_end
, increment
,
2326 unroll_number
, result
)
2327 struct iv_class
*bl
;
2328 enum unroll_types unroll_type
;
2329 rtx loop_start
, loop_end
;
2331 int unroll_number
, result
;
2333 struct induction
*v
;
2337 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
2341 /* Only split the giv if it has already been reduced, or if the loop is
2342 being completely unrolled. */
2343 if (unroll_type
!= UNROLL_COMPLETELY
&& v
->ignore
)
2346 /* The giv can be split if the insn that sets the giv is executed once
2347 and only once on every iteration of the loop. */
2348 /* An address giv can always be split. v->insn is just a use not a set,
2349 and hence it does not matter whether it is always executed. All that
2350 matters is that all the biv increments are always executed, and we
2351 won't reach here if they aren't. */
2352 if (v
->giv_type
!= DEST_ADDR
2353 && (! v
->always_computable
2354 || back_branch_in_range_p (v
->insn
, loop_start
, loop_end
)))
2357 /* The giv increment value must be a constant. */
2358 giv_inc
= fold_rtx_mult_add (v
->mult_val
, increment
, const0_rtx
,
2360 if (! giv_inc
|| GET_CODE (giv_inc
) != CONST_INT
)
2363 /* The loop must be unrolled completely, or else have a known number of
2364 iterations and only one exit, or else the giv must be dead outside
2365 the loop, or else the final value of the giv must be known.
2366 Otherwise, it is not safe to split the giv since it may not have the
2367 proper value on loop exit. */
2369 /* The used outside loop test will fail for DEST_ADDR givs. They are
2370 never used outside the loop anyways, so it is always safe to split a
2374 if (unroll_type
!= UNROLL_COMPLETELY
2375 && (loop_number_exit_labels
[uid_loop_num
[INSN_UID (loop_start
)]]
2376 || unroll_type
== UNROLL_NAIVE
)
2377 && v
->giv_type
!= DEST_ADDR
2378 && ((regno_first_uid
[REGNO (v
->dest_reg
)] != INSN_UID (v
->insn
)
2379 /* Check for the case where the pseudo is set by a shift/add
2380 sequence, in which case the first insn setting the pseudo
2381 is the first insn of the shift/add sequence. */
2382 && (! (tem
= find_reg_note (v
->insn
, REG_RETVAL
, NULL_RTX
))
2383 || (regno_first_uid
[REGNO (v
->dest_reg
)]
2384 != INSN_UID (XEXP (tem
, 0)))))
2385 /* Line above always fails if INSN was moved by loop opt. */
2386 || (uid_luid
[regno_last_uid
[REGNO (v
->dest_reg
)]]
2387 >= INSN_LUID (loop_end
)))
2388 && ! (final_value
= v
->final_value
))
2392 /* Currently, non-reduced/final-value givs are never split. */
2393 /* Should emit insns after the loop if possible, as the biv final value
2396 /* If the final value is non-zero, and the giv has not been reduced,
2397 then must emit an instruction to set the final value. */
2398 if (final_value
&& !v
->new_reg
)
2400 /* Create a new register to hold the value of the giv, and then set
2401 the giv to its final value before the loop start. The giv is set
2402 to its final value before loop start to ensure that this insn
2403 will always be executed, no matter how we exit. */
2404 tem
= gen_reg_rtx (v
->mode
);
2405 emit_insn_before (gen_move_insn (tem
, v
->dest_reg
), loop_start
);
2406 emit_insn_before (gen_move_insn (v
->dest_reg
, final_value
),
2409 if (loop_dump_stream
)
2410 fprintf (loop_dump_stream
, "Giv %d mapped to %d for split.\n",
2411 REGNO (v
->dest_reg
), REGNO (tem
));
2417 /* This giv is splittable. If completely unrolling the loop, save the
2418 giv's initial value. Otherwise, save the constant zero for it. */
2420 if (unroll_type
== UNROLL_COMPLETELY
)
2422 /* It is not safe to use bl->initial_value here, because it may not
2423 be invariant. It is safe to use the initial value stored in
2424 the splittable_regs array if it is set. In rare cases, it won't
2425 be set, so then we do exactly the same thing as
2426 find_splittable_regs does to get a safe value. */
2427 rtx biv_initial_value
;
2429 if (splittable_regs
[bl
->regno
])
2430 biv_initial_value
= splittable_regs
[bl
->regno
];
2431 else if (GET_CODE (bl
->initial_value
) != REG
2432 || (REGNO (bl
->initial_value
) != bl
->regno
2433 && REGNO (bl
->initial_value
) >= FIRST_PSEUDO_REGISTER
))
2434 biv_initial_value
= bl
->initial_value
;
2437 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2439 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2441 biv_initial_value
= tem
;
2443 value
= fold_rtx_mult_add (v
->mult_val
, biv_initial_value
,
2444 v
->add_val
, v
->mode
);
2451 /* If a giv was combined with another giv, then we can only split
2452 this giv if the giv it was combined with was reduced. This
2453 is because the value of v->new_reg is meaningless in this
2455 if (v
->same
&& ! v
->same
->new_reg
)
2457 if (loop_dump_stream
)
2458 fprintf (loop_dump_stream
,
2459 "giv combined with unreduced giv not split.\n");
2462 /* If the giv is an address destination, it could be something other
2463 than a simple register, these have to be treated differently. */
2464 else if (v
->giv_type
== DEST_REG
)
2466 /* If value is not a constant, register, or register plus
2467 constant, then compute its value into a register before
2468 loop start. This prevents illegal rtx sharing, and should
2469 generate better code. We can use bl->initial_value here
2470 instead of splittable_regs[bl->regno] because this code
2471 is going before the loop start. */
2472 if (unroll_type
== UNROLL_COMPLETELY
2473 && GET_CODE (value
) != CONST_INT
2474 && GET_CODE (value
) != REG
2475 && (GET_CODE (value
) != PLUS
2476 || GET_CODE (XEXP (value
, 0)) != REG
2477 || GET_CODE (XEXP (value
, 1)) != CONST_INT
))
2479 rtx tem
= gen_reg_rtx (v
->mode
);
2480 emit_iv_add_mult (bl
->initial_value
, v
->mult_val
,
2481 v
->add_val
, tem
, loop_start
);
2485 splittable_regs
[REGNO (v
->new_reg
)] = value
;
2489 /* Splitting address givs is useful since it will often allow us
2490 to eliminate some increment insns for the base giv as
2493 /* If the addr giv is combined with a dest_reg giv, then all
2494 references to that dest reg will be remapped, which is NOT
2495 what we want for split addr regs. We always create a new
2496 register for the split addr giv, just to be safe. */
2498 /* ??? If there are multiple address givs which have been
2499 combined with the same dest_reg giv, then we may only need
2500 one new register for them. Pulling out constants below will
2501 catch some of the common cases of this. Currently, I leave
2502 the work of simplifying multiple address givs to the
2503 following cse pass. */
2505 v
->const_adjust
= 0;
2506 if (unroll_type
!= UNROLL_COMPLETELY
)
2508 /* If not completely unrolling the loop, then create a new
2509 register to hold the split value of the DEST_ADDR giv.
2510 Emit insn to initialize its value before loop start. */
2511 tem
= gen_reg_rtx (v
->mode
);
2513 /* If the address giv has a constant in its new_reg value,
2514 then this constant can be pulled out and put in value,
2515 instead of being part of the initialization code. */
2517 if (GET_CODE (v
->new_reg
) == PLUS
2518 && GET_CODE (XEXP (v
->new_reg
, 1)) == CONST_INT
)
2521 = plus_constant (tem
, INTVAL (XEXP (v
->new_reg
,1)));
2523 /* Only succeed if this will give valid addresses.
2524 Try to validate both the first and the last
2525 address resulting from loop unrolling, if
2526 one fails, then can't do const elim here. */
2527 if (memory_address_p (v
->mem_mode
, v
->dest_reg
)
2528 && memory_address_p (v
->mem_mode
,
2529 plus_constant (v
->dest_reg
,
2531 * (unroll_number
- 1))))
2533 /* Save the negative of the eliminated const, so
2534 that we can calculate the dest_reg's increment
2536 v
->const_adjust
= - INTVAL (XEXP (v
->new_reg
, 1));
2538 v
->new_reg
= XEXP (v
->new_reg
, 0);
2539 if (loop_dump_stream
)
2540 fprintf (loop_dump_stream
,
2541 "Eliminating constant from giv %d\n",
2550 /* If the address hasn't been checked for validity yet, do so
2551 now, and fail completely if either the first or the last
2552 unrolled copy of the address is not a valid address. */
2553 if (v
->dest_reg
== tem
2554 && (! memory_address_p (v
->mem_mode
, v
->dest_reg
)
2555 || ! memory_address_p (v
->mem_mode
,
2556 plus_constant (v
->dest_reg
,
2558 * (unroll_number
-1)))))
2560 if (loop_dump_stream
)
2561 fprintf (loop_dump_stream
,
2562 "Illegal address for giv at insn %d\n",
2563 INSN_UID (v
->insn
));
2567 /* To initialize the new register, just move the value of
2568 new_reg into it. This is not guaranteed to give a valid
2569 instruction on machines with complex addressing modes.
2570 If we can't recognize it, then delete it and emit insns
2571 to calculate the value from scratch. */
2572 emit_insn_before (gen_rtx (SET
, VOIDmode
, tem
,
2573 copy_rtx (v
->new_reg
)),
2575 if (recog_memoized (PREV_INSN (loop_start
)) < 0)
2577 delete_insn (PREV_INSN (loop_start
));
2578 emit_iv_add_mult (bl
->initial_value
, v
->mult_val
,
2579 v
->add_val
, tem
, loop_start
);
2580 if (loop_dump_stream
)
2581 fprintf (loop_dump_stream
,
2582 "Illegal init insn, rewritten.\n");
2587 v
->dest_reg
= value
;
2589 /* Check the resulting address for validity, and fail
2590 if the resulting address would be illegal. */
2591 if (! memory_address_p (v
->mem_mode
, v
->dest_reg
)
2592 || ! memory_address_p (v
->mem_mode
,
2593 plus_constant (v
->dest_reg
,
2595 (unroll_number
-1))))
2597 if (loop_dump_stream
)
2598 fprintf (loop_dump_stream
,
2599 "Illegal address for giv at insn %d\n",
2600 INSN_UID (v
->insn
));
2605 /* Store the value of dest_reg into the insn. This sharing
2606 will not be a problem as this insn will always be copied
2609 *v
->location
= v
->dest_reg
;
2611 /* If this address giv is combined with a dest reg giv, then
2612 save the base giv's induction pointer so that we will be
2613 able to handle this address giv properly. The base giv
2614 itself does not have to be splittable. */
2616 if (v
->same
&& v
->same
->giv_type
== DEST_REG
)
2617 addr_combined_regs
[REGNO (v
->same
->new_reg
)] = v
->same
;
2619 if (GET_CODE (v
->new_reg
) == REG
)
2621 /* This giv maybe hasn't been combined with any others.
2622 Make sure that it's giv is marked as splittable here. */
2624 splittable_regs
[REGNO (v
->new_reg
)] = value
;
2626 /* Make it appear to depend upon itself, so that the
2627 giv will be properly split in the main loop above. */
2631 addr_combined_regs
[REGNO (v
->new_reg
)] = v
;
2635 if (loop_dump_stream
)
2636 fprintf (loop_dump_stream
, "DEST_ADDR giv being split.\n");
2642 /* Currently, unreduced giv's can't be split. This is not too much
2643 of a problem since unreduced giv's are not live across loop
2644 iterations anyways. When unrolling a loop completely though,
2645 it makes sense to reduce&split givs when possible, as this will
2646 result in simpler instructions, and will not require that a reg
2647 be live across loop iterations. */
2649 splittable_regs
[REGNO (v
->dest_reg
)] = value
;
2650 fprintf (stderr
, "Giv %d at insn %d not reduced\n",
2651 REGNO (v
->dest_reg
), INSN_UID (v
->insn
));
2657 /* Givs are only updated once by definition. Mark it so if this is
2658 a splittable register. Don't need to do anything for address givs
2659 where this may not be a register. */
2661 if (GET_CODE (v
->new_reg
) == REG
)
2662 splittable_regs_updates
[REGNO (v
->new_reg
)] = 1;
2666 if (loop_dump_stream
)
2670 if (GET_CODE (v
->dest_reg
) == CONST_INT
)
2672 else if (GET_CODE (v
->dest_reg
) != REG
)
2673 regnum
= REGNO (XEXP (v
->dest_reg
, 0));
2675 regnum
= REGNO (v
->dest_reg
);
2676 fprintf (loop_dump_stream
, "Giv %d at insn %d safe to split.\n",
2677 regnum
, INSN_UID (v
->insn
));
2684 /* Try to prove that the register is dead after the loop exits. Trace every
2685 loop exit looking for an insn that will always be executed, which sets
2686 the register to some value, and appears before the first use of the register
2687 is found. If successful, then return 1, otherwise return 0. */
2689 /* ?? Could be made more intelligent in the handling of jumps, so that
2690 it can search past if statements and other similar structures. */
2693 reg_dead_after_loop (reg
, loop_start
, loop_end
)
2694 rtx reg
, loop_start
, loop_end
;
2700 /* HACK: Must also search the loop fall through exit, create a label_ref
2701 here which points to the loop_end, and append the loop_number_exit_labels
2703 label
= gen_rtx (LABEL_REF
, VOIDmode
, loop_end
);
2704 LABEL_NEXTREF (label
)
2705 = loop_number_exit_labels
[uid_loop_num
[INSN_UID (loop_start
)]];
2707 for ( ; label
; label
= LABEL_NEXTREF (label
))
2709 /* Succeed if find an insn which sets the biv or if reach end of
2710 function. Fail if find an insn that uses the biv, or if come to
2711 a conditional jump. */
2713 insn
= NEXT_INSN (XEXP (label
, 0));
2716 code
= GET_CODE (insn
);
2717 if (GET_RTX_CLASS (code
) == 'i')
2721 if (reg_referenced_p (reg
, PATTERN (insn
)))
2724 set
= single_set (insn
);
2725 if (set
&& rtx_equal_p (SET_DEST (set
), reg
))
2729 if (code
== JUMP_INSN
)
2731 if (GET_CODE (PATTERN (insn
)) == RETURN
)
2733 else if (! simplejump_p (insn
)
2734 /* Prevent infinite loop following infinite loops. */
2735 || jump_count
++ > 20)
2738 insn
= JUMP_LABEL (insn
);
2741 insn
= NEXT_INSN (insn
);
2745 /* Success, the register is dead on all loop exits. */
2749 /* Try to calculate the final value of the biv, the value it will have at
2750 the end of the loop. If we can do it, return that value. */
2753 final_biv_value (bl
, loop_start
, loop_end
)
2754 struct iv_class
*bl
;
2755 rtx loop_start
, loop_end
;
2759 /* ??? This only works for MODE_INT biv's. Reject all others for now. */
2761 if (GET_MODE_CLASS (bl
->biv
->mode
) != MODE_INT
)
2764 /* The final value for reversed bivs must be calculated differently than
2765 for ordinary bivs. In this case, there is already an insn after the
2766 loop which sets this biv's final value (if necessary), and there are
2767 no other loop exits, so we can return any value. */
2770 if (loop_dump_stream
)
2771 fprintf (loop_dump_stream
,
2772 "Final biv value for %d, reversed biv.\n", bl
->regno
);
2777 /* Try to calculate the final value as initial value + (number of iterations
2778 * increment). For this to work, increment must be invariant, the only
2779 exit from the loop must be the fall through at the bottom (otherwise
2780 it may not have its final value when the loop exits), and the initial
2781 value of the biv must be invariant. */
2783 if (loop_n_iterations
!= 0
2784 && ! loop_number_exit_labels
[uid_loop_num
[INSN_UID (loop_start
)]]
2785 && invariant_p (bl
->initial_value
))
2787 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2789 if (increment
&& invariant_p (increment
))
2791 /* Can calculate the loop exit value, emit insns after loop
2792 end to calculate this value into a temporary register in
2793 case it is needed later. */
2795 tem
= gen_reg_rtx (bl
->biv
->mode
);
2796 /* Make sure loop_end is not the last insn. */
2797 if (NEXT_INSN (loop_end
) == 0)
2798 emit_note_after (NOTE_INSN_DELETED
, loop_end
);
2799 emit_iv_add_mult (increment
, GEN_INT (loop_n_iterations
),
2800 bl
->initial_value
, tem
, NEXT_INSN (loop_end
));
2802 if (loop_dump_stream
)
2803 fprintf (loop_dump_stream
,
2804 "Final biv value for %d, calculated.\n", bl
->regno
);
2810 /* Check to see if the biv is dead at all loop exits. */
2811 if (reg_dead_after_loop (bl
->biv
->src_reg
, loop_start
, loop_end
))
2813 if (loop_dump_stream
)
2814 fprintf (loop_dump_stream
,
2815 "Final biv value for %d, biv dead after loop exit.\n",
2824 /* Try to calculate the final value of the giv, the value it will have at
2825 the end of the loop. If we can do it, return that value. */
2828 final_giv_value (v
, loop_start
, loop_end
)
2829 struct induction
*v
;
2830 rtx loop_start
, loop_end
;
2832 struct iv_class
*bl
;
2836 rtx insert_before
, seq
;
2838 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
2840 /* The final value for givs which depend on reversed bivs must be calculated
2841 differently than for ordinary givs. In this case, there is already an
2842 insn after the loop which sets this giv's final value (if necessary),
2843 and there are no other loop exits, so we can return any value. */
2846 if (loop_dump_stream
)
2847 fprintf (loop_dump_stream
,
2848 "Final giv value for %d, depends on reversed biv\n",
2849 REGNO (v
->dest_reg
));
2853 /* Try to calculate the final value as a function of the biv it depends
2854 upon. The only exit from the loop must be the fall through at the bottom
2855 (otherwise it may not have its final value when the loop exits). */
2857 /* ??? Can calculate the final giv value by subtracting off the
2858 extra biv increments times the giv's mult_val. The loop must have
2859 only one exit for this to work, but the loop iterations does not need
2862 if (loop_n_iterations
!= 0
2863 && ! loop_number_exit_labels
[uid_loop_num
[INSN_UID (loop_start
)]])
2865 /* ?? It is tempting to use the biv's value here since these insns will
2866 be put after the loop, and hence the biv will have its final value
2867 then. However, this fails if the biv is subsequently eliminated.
2868 Perhaps determine whether biv's are eliminable before trying to
2869 determine whether giv's are replaceable so that we can use the
2870 biv value here if it is not eliminable. */
2872 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2874 if (increment
&& invariant_p (increment
))
2876 /* Can calculate the loop exit value of its biv as
2877 (loop_n_iterations * increment) + initial_value */
2879 /* The loop exit value of the giv is then
2880 (final_biv_value - extra increments) * mult_val + add_val.
2881 The extra increments are any increments to the biv which
2882 occur in the loop after the giv's value is calculated.
2883 We must search from the insn that sets the giv to the end
2884 of the loop to calculate this value. */
2886 insert_before
= NEXT_INSN (loop_end
);
2888 /* Put the final biv value in tem. */
2889 tem
= gen_reg_rtx (bl
->biv
->mode
);
2890 emit_iv_add_mult (increment
, GEN_INT (loop_n_iterations
),
2891 bl
->initial_value
, tem
, insert_before
);
2893 /* Subtract off extra increments as we find them. */
2894 for (insn
= NEXT_INSN (v
->insn
); insn
!= loop_end
;
2895 insn
= NEXT_INSN (insn
))
2897 struct induction
*biv
;
2899 for (biv
= bl
->biv
; biv
; biv
= biv
->next_iv
)
2900 if (biv
->insn
== insn
)
2903 tem
= expand_binop (GET_MODE (tem
), sub_optab
, tem
,
2904 biv
->add_val
, NULL_RTX
, 0,
2906 seq
= gen_sequence ();
2908 emit_insn_before (seq
, insert_before
);
2912 /* Now calculate the giv's final value. */
2913 emit_iv_add_mult (tem
, v
->mult_val
, v
->add_val
, tem
,
2916 if (loop_dump_stream
)
2917 fprintf (loop_dump_stream
,
2918 "Final giv value for %d, calc from biv's value.\n",
2919 REGNO (v
->dest_reg
));
2925 /* Replaceable giv's should never reach here. */
2929 /* Check to see if the biv is dead at all loop exits. */
2930 if (reg_dead_after_loop (v
->dest_reg
, loop_start
, loop_end
))
2932 if (loop_dump_stream
)
2933 fprintf (loop_dump_stream
,
2934 "Final giv value for %d, giv dead after loop exit.\n",
2935 REGNO (v
->dest_reg
));
2944 /* Calculate the number of loop iterations. Returns the exact number of loop
2945 iterations if it can be calculated, otherwise returns zero. */
2947 unsigned HOST_WIDE_INT
2948 loop_iterations (loop_start
, loop_end
)
2949 rtx loop_start
, loop_end
;
2951 rtx comparison
, comparison_value
;
2952 rtx iteration_var
, initial_value
, increment
, final_value
;
2953 enum rtx_code comparison_code
;
2956 int unsigned_compare
, compare_dir
, final_larger
;
2957 unsigned long tempu
;
2960 /* First find the iteration variable. If the last insn is a conditional
2961 branch, and the insn before tests a register value, make that the
2962 iteration variable. */
2964 loop_initial_value
= 0;
2966 loop_final_value
= 0;
2967 loop_iteration_var
= 0;
2969 last_loop_insn
= prev_nonnote_insn (loop_end
);
2971 comparison
= get_condition_for_loop (last_loop_insn
);
2972 if (comparison
== 0)
2974 if (loop_dump_stream
)
2975 fprintf (loop_dump_stream
,
2976 "Loop unrolling: No final conditional branch found.\n");
2980 /* ??? Get_condition may switch position of induction variable and
2981 invariant register when it canonicalizes the comparison. */
2983 comparison_code
= GET_CODE (comparison
);
2984 iteration_var
= XEXP (comparison
, 0);
2985 comparison_value
= XEXP (comparison
, 1);
2987 if (GET_CODE (iteration_var
) != REG
)
2989 if (loop_dump_stream
)
2990 fprintf (loop_dump_stream
,
2991 "Loop unrolling: Comparison not against register.\n");
2995 /* Loop iterations is always called before any new registers are created
2996 now, so this should never occur. */
2998 if (REGNO (iteration_var
) >= max_reg_before_loop
)
3001 iteration_info (iteration_var
, &initial_value
, &increment
,
3002 loop_start
, loop_end
);
3003 if (initial_value
== 0)
3004 /* iteration_info already printed a message. */
3009 if (loop_dump_stream
)
3010 fprintf (loop_dump_stream
,
3011 "Loop unrolling: Increment value can't be calculated.\n");
3014 if (GET_CODE (increment
) != CONST_INT
)
3016 if (loop_dump_stream
)
3017 fprintf (loop_dump_stream
,
3018 "Loop unrolling: Increment value not constant.\n");
3021 if (GET_CODE (initial_value
) != CONST_INT
)
3023 if (loop_dump_stream
)
3024 fprintf (loop_dump_stream
,
3025 "Loop unrolling: Initial value not constant.\n");
3029 /* If the comparison value is an invariant register, then try to find
3030 its value from the insns before the start of the loop. */
3032 if (GET_CODE (comparison_value
) == REG
&& invariant_p (comparison_value
))
3036 for (insn
= PREV_INSN (loop_start
); insn
; insn
= PREV_INSN (insn
))
3038 if (GET_CODE (insn
) == CODE_LABEL
)
3041 else if (GET_RTX_CLASS (GET_CODE (insn
)) == 'i'
3042 && reg_set_p (comparison_value
, insn
))
3044 /* We found the last insn before the loop that sets the register.
3045 If it sets the entire register, and has a REG_EQUAL note,
3046 then use the value of the REG_EQUAL note. */
3047 if ((set
= single_set (insn
))
3048 && (SET_DEST (set
) == comparison_value
))
3050 rtx note
= find_reg_note (insn
, REG_EQUAL
, NULL_RTX
);
3052 if (note
&& GET_CODE (XEXP (note
, 0)) != EXPR_LIST
)
3053 comparison_value
= XEXP (note
, 0);
3060 final_value
= approx_final_value (comparison_code
, comparison_value
,
3061 &unsigned_compare
, &compare_dir
);
3063 /* Save the calculated values describing this loop's bounds, in case
3064 precondition_loop_p will need them later. These values can not be
3065 recalculated inside precondition_loop_p because strength reduction
3066 optimizations may obscure the loop's structure. */
3068 loop_iteration_var
= iteration_var
;
3069 loop_initial_value
= initial_value
;
3070 loop_increment
= increment
;
3071 loop_final_value
= final_value
;
3073 if (final_value
== 0)
3075 if (loop_dump_stream
)
3076 fprintf (loop_dump_stream
,
3077 "Loop unrolling: EQ comparison loop.\n");
3080 else if (GET_CODE (final_value
) != CONST_INT
)
3082 if (loop_dump_stream
)
3083 fprintf (loop_dump_stream
,
3084 "Loop unrolling: Final value not constant.\n");
3088 /* ?? Final value and initial value do not have to be constants.
3089 Only their difference has to be constant. When the iteration variable
3090 is an array address, the final value and initial value might both
3091 be addresses with the same base but different constant offsets.
3092 Final value must be invariant for this to work.
3094 To do this, need some way to find the values of registers which are
3097 /* Final_larger is 1 if final larger, 0 if they are equal, otherwise -1. */
3098 if (unsigned_compare
)
3100 = ((unsigned HOST_WIDE_INT
) INTVAL (final_value
)
3101 > (unsigned HOST_WIDE_INT
) INTVAL (initial_value
))
3102 - ((unsigned HOST_WIDE_INT
) INTVAL (final_value
)
3103 < (unsigned HOST_WIDE_INT
) INTVAL (initial_value
));
3105 final_larger
= (INTVAL (final_value
) > INTVAL (initial_value
))
3106 - (INTVAL (final_value
) < INTVAL (initial_value
));
3108 if (INTVAL (increment
) > 0)
3110 else if (INTVAL (increment
) == 0)
3115 /* There are 27 different cases: compare_dir = -1, 0, 1;
3116 final_larger = -1, 0, 1; increment_dir = -1, 0, 1.
3117 There are 4 normal cases, 4 reverse cases (where the iteration variable
3118 will overflow before the loop exits), 4 infinite loop cases, and 15
3119 immediate exit (0 or 1 iteration depending on loop type) cases.
3120 Only try to optimize the normal cases. */
3122 /* (compare_dir/final_larger/increment_dir)
3123 Normal cases: (0/-1/-1), (0/1/1), (-1/-1/-1), (1/1/1)
3124 Reverse cases: (0/-1/1), (0/1/-1), (-1/-1/1), (1/1/-1)
3125 Infinite loops: (0/-1/0), (0/1/0), (-1/-1/0), (1/1/0)
3126 Immediate exit: (0/0/X), (-1/0/X), (-1/1/X), (1/0/X), (1/-1/X) */
3128 /* ?? If the meaning of reverse loops (where the iteration variable
3129 will overflow before the loop exits) is undefined, then could
3130 eliminate all of these special checks, and just always assume
3131 the loops are normal/immediate/infinite. Note that this means
3132 the sign of increment_dir does not have to be known. Also,
3133 since it does not really hurt if immediate exit loops or infinite loops
3134 are optimized, then that case could be ignored also, and hence all
3135 loops can be optimized.
3137 According to ANSI Spec, the reverse loop case result is undefined,
3138 because the action on overflow is undefined.
3140 See also the special test for NE loops below. */
3142 if (final_larger
== increment_dir
&& final_larger
!= 0
3143 && (final_larger
== compare_dir
|| compare_dir
== 0))
3148 if (loop_dump_stream
)
3149 fprintf (loop_dump_stream
,
3150 "Loop unrolling: Not normal loop.\n");
3154 /* Calculate the number of iterations, final_value is only an approximation,
3155 so correct for that. Note that tempu and loop_n_iterations are
3156 unsigned, because they can be as large as 2^n - 1. */
3158 i
= INTVAL (increment
);
3160 tempu
= INTVAL (final_value
) - INTVAL (initial_value
);
3163 tempu
= INTVAL (initial_value
) - INTVAL (final_value
);
3169 /* For NE tests, make sure that the iteration variable won't miss the
3170 final value. If tempu mod i is not zero, then the iteration variable
3171 will overflow before the loop exits, and we can not calculate the
3172 number of iterations. */
3173 if (compare_dir
== 0 && (tempu
% i
) != 0)
3176 return tempu
/ i
+ ((tempu
% i
) != 0);