1 /* Try to unroll loops, and split induction variables.
2 Copyright (C) 1992, 1993, 1994 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 ();
204 static rtx
remap_split_bivs ();
206 /* Try to unroll one loop and split induction variables in the loop.
208 The loop is described by the arguments LOOP_END, INSN_COUNT, and
209 LOOP_START. END_INSERT_BEFORE indicates where insns should be added
210 which need to be executed when the loop falls through. STRENGTH_REDUCTION_P
211 indicates whether information generated in the strength reduction pass
214 This function is intended to be called from within `strength_reduce'
218 unroll_loop (loop_end
, insn_count
, loop_start
, end_insert_before
,
223 rtx end_insert_before
;
224 int strength_reduce_p
;
227 int unroll_number
= 1;
228 rtx copy_start
, copy_end
;
229 rtx insn
, copy
, sequence
, pattern
, tem
;
230 int max_labelno
, max_insnno
;
232 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");
603 if (LABEL_NAME (start_label
))
605 /* The jump optimization pass must have combined the original start label
606 with a named label for a goto. We can't unroll this case because
607 jumps which go to the named label must be handled differently than
608 jumps to the loop start, and it is impossible to differentiate them
610 if (loop_dump_stream
)
611 fprintf (loop_dump_stream
,
612 "Unrolling failure: loop start label is gone\n");
616 if (unroll_type
== UNROLL_NAIVE
617 && GET_CODE (last_loop_insn
) == BARRIER
618 && start_label
!= JUMP_LABEL (PREV_INSN (last_loop_insn
)))
620 /* In this case, we must copy the jump and barrier, because they will
621 not be converted to jumps to an immediately following label. */
623 insert_before
= NEXT_INSN (last_loop_insn
);
624 copy_end
= last_loop_insn
;
627 /* Allocate a translation table for the labels and insn numbers.
628 They will be filled in as we copy the insns in the loop. */
630 max_labelno
= max_label_num ();
631 max_insnno
= get_max_uid ();
633 map
= (struct inline_remap
*) alloca (sizeof (struct inline_remap
));
635 map
->integrating
= 0;
637 /* Allocate the label map. */
641 map
->label_map
= (rtx
*) alloca (max_labelno
* sizeof (rtx
));
643 local_label
= (char *) alloca (max_labelno
);
644 bzero (local_label
, max_labelno
);
649 /* Search the loop and mark all local labels, i.e. the ones which have to
650 be distinct labels when copied. For all labels which might be
651 non-local, set their label_map entries to point to themselves.
652 If they happen to be local their label_map entries will be overwritten
653 before the loop body is copied. The label_map entries for local labels
654 will be set to a different value each time the loop body is copied. */
656 for (insn
= copy_start
; insn
!= loop_end
; insn
= NEXT_INSN (insn
))
658 if (GET_CODE (insn
) == CODE_LABEL
)
659 local_label
[CODE_LABEL_NUMBER (insn
)] = 1;
660 else if (GET_CODE (insn
) == JUMP_INSN
)
662 if (JUMP_LABEL (insn
))
663 map
->label_map
[CODE_LABEL_NUMBER (JUMP_LABEL (insn
))]
665 else if (GET_CODE (PATTERN (insn
)) == ADDR_VEC
666 || GET_CODE (PATTERN (insn
)) == ADDR_DIFF_VEC
)
668 rtx pat
= PATTERN (insn
);
669 int diff_vec_p
= GET_CODE (PATTERN (insn
)) == ADDR_DIFF_VEC
;
670 int len
= XVECLEN (pat
, diff_vec_p
);
673 for (i
= 0; i
< len
; i
++)
675 label
= XEXP (XVECEXP (pat
, diff_vec_p
, i
), 0);
676 map
->label_map
[CODE_LABEL_NUMBER (label
)] = label
;
682 /* Allocate space for the insn map. */
684 map
->insn_map
= (rtx
*) alloca (max_insnno
* sizeof (rtx
));
686 /* Set this to zero, to indicate that we are doing loop unrolling,
687 not function inlining. */
688 map
->inline_target
= 0;
690 /* The register and constant maps depend on the number of registers
691 present, so the final maps can't be created until after
692 find_splittable_regs is called. However, they are needed for
693 preconditioning, so we create temporary maps when preconditioning
696 /* The preconditioning code may allocate two new pseudo registers. */
697 maxregnum
= max_reg_num ();
699 /* Allocate and zero out the splittable_regs and addr_combined_regs
700 arrays. These must be zeroed here because they will be used if
701 loop preconditioning is performed, and must be zero for that case.
703 It is safe to do this here, since the extra registers created by the
704 preconditioning code and find_splittable_regs will never be used
705 to access the splittable_regs[] and addr_combined_regs[] arrays. */
707 splittable_regs
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
708 bzero ((char *) splittable_regs
, maxregnum
* sizeof (rtx
));
709 splittable_regs_updates
= (int *) alloca (maxregnum
* sizeof (int));
710 bzero ((char *) splittable_regs_updates
, maxregnum
* sizeof (int));
712 = (struct induction
**) alloca (maxregnum
* sizeof (struct induction
*));
713 bzero ((char *) addr_combined_regs
, maxregnum
* sizeof (struct induction
*));
715 /* If this loop requires exit tests when unrolled, check to see if we
716 can precondition the loop so as to make the exit tests unnecessary.
717 Just like variable splitting, this is not safe if the loop is entered
718 via a jump to the bottom. Also, can not do this if no strength
719 reduce info, because precondition_loop_p uses this info. */
721 /* Must copy the loop body for preconditioning before the following
722 find_splittable_regs call since that will emit insns which need to
723 be after the preconditioned loop copies, but immediately before the
724 unrolled loop copies. */
726 /* Also, it is not safe to split induction variables for the preconditioned
727 copies of the loop body. If we split induction variables, then the code
728 assumes that each induction variable can be represented as a function
729 of its initial value and the loop iteration number. This is not true
730 in this case, because the last preconditioned copy of the loop body
731 could be any iteration from the first up to the `unroll_number-1'th,
732 depending on the initial value of the iteration variable. Therefore
733 we can not split induction variables here, because we can not calculate
734 their value. Hence, this code must occur before find_splittable_regs
737 if (unroll_type
== UNROLL_NAIVE
&& ! splitting_not_safe
&& strength_reduce_p
)
739 rtx initial_value
, final_value
, increment
;
741 if (precondition_loop_p (&initial_value
, &final_value
, &increment
,
742 loop_start
, loop_end
))
744 register rtx diff
, temp
;
745 enum machine_mode mode
;
747 int abs_inc
, neg_inc
;
749 map
->reg_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
751 map
->const_equiv_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
752 map
->const_age_map
= (unsigned *) alloca (maxregnum
753 * sizeof (unsigned));
754 map
->const_equiv_map_size
= maxregnum
;
755 global_const_equiv_map
= map
->const_equiv_map
;
756 global_const_equiv_map_size
= maxregnum
;
758 init_reg_map (map
, maxregnum
);
760 /* Limit loop unrolling to 4, since this will make 7 copies of
762 if (unroll_number
> 4)
765 /* Save the absolute value of the increment, and also whether or
766 not it is negative. */
768 abs_inc
= INTVAL (increment
);
777 /* Decide what mode to do these calculations in. Choose the larger
778 of final_value's mode and initial_value's mode, or a full-word if
779 both are constants. */
780 mode
= GET_MODE (final_value
);
781 if (mode
== VOIDmode
)
783 mode
= GET_MODE (initial_value
);
784 if (mode
== VOIDmode
)
787 else if (mode
!= GET_MODE (initial_value
)
788 && (GET_MODE_SIZE (mode
)
789 < GET_MODE_SIZE (GET_MODE (initial_value
))))
790 mode
= GET_MODE (initial_value
);
792 /* Calculate the difference between the final and initial values.
793 Final value may be a (plus (reg x) (const_int 1)) rtx.
794 Let the following cse pass simplify this if initial value is
797 We must copy the final and initial values here to avoid
798 improperly shared rtl. */
800 diff
= expand_binop (mode
, sub_optab
, copy_rtx (final_value
),
801 copy_rtx (initial_value
), NULL_RTX
, 0,
804 /* Now calculate (diff % (unroll * abs (increment))) by using an
806 diff
= expand_binop (GET_MODE (diff
), and_optab
, diff
,
807 GEN_INT (unroll_number
* abs_inc
- 1),
808 NULL_RTX
, 0, OPTAB_LIB_WIDEN
);
810 /* Now emit a sequence of branches to jump to the proper precond
813 labels
= (rtx
*) alloca (sizeof (rtx
) * unroll_number
);
814 for (i
= 0; i
< unroll_number
; i
++)
815 labels
[i
] = gen_label_rtx ();
817 /* Assuming the unroll_number is 4, and the increment is 2, then
818 for a negative increment: for a positive increment:
819 diff = 0,1 precond 0 diff = 0,7 precond 0
820 diff = 2,3 precond 3 diff = 1,2 precond 1
821 diff = 4,5 precond 2 diff = 3,4 precond 2
822 diff = 6,7 precond 1 diff = 5,6 precond 3 */
824 /* We only need to emit (unroll_number - 1) branches here, the
825 last case just falls through to the following code. */
827 /* ??? This would give better code if we emitted a tree of branches
828 instead of the current linear list of branches. */
830 for (i
= 0; i
< unroll_number
- 1; i
++)
834 /* For negative increments, must invert the constant compared
835 against, except when comparing against zero. */
839 cmp_const
= unroll_number
- i
;
843 emit_cmp_insn (diff
, GEN_INT (abs_inc
* cmp_const
),
844 EQ
, NULL_RTX
, mode
, 0, 0);
847 emit_jump_insn (gen_beq (labels
[i
]));
849 emit_jump_insn (gen_bge (labels
[i
]));
851 emit_jump_insn (gen_ble (labels
[i
]));
852 JUMP_LABEL (get_last_insn ()) = labels
[i
];
853 LABEL_NUSES (labels
[i
])++;
856 /* If the increment is greater than one, then we need another branch,
857 to handle other cases equivalent to 0. */
859 /* ??? This should be merged into the code above somehow to help
860 simplify the code here, and reduce the number of branches emitted.
861 For the negative increment case, the branch here could easily
862 be merged with the `0' case branch above. For the positive
863 increment case, it is not clear how this can be simplified. */
870 cmp_const
= abs_inc
- 1;
872 cmp_const
= abs_inc
* (unroll_number
- 1) + 1;
874 emit_cmp_insn (diff
, GEN_INT (cmp_const
), EQ
, NULL_RTX
,
878 emit_jump_insn (gen_ble (labels
[0]));
880 emit_jump_insn (gen_bge (labels
[0]));
881 JUMP_LABEL (get_last_insn ()) = labels
[0];
882 LABEL_NUSES (labels
[0])++;
885 sequence
= gen_sequence ();
887 emit_insn_before (sequence
, loop_start
);
889 /* Only the last copy of the loop body here needs the exit
890 test, so set copy_end to exclude the compare/branch here,
891 and then reset it inside the loop when get to the last
894 if (GET_CODE (last_loop_insn
) == BARRIER
)
895 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
896 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
899 /* The immediately preceding insn is a compare which we do not
901 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
903 /* The immediately preceding insn may not be a compare, so we
905 copy_end
= PREV_INSN (last_loop_insn
);
911 for (i
= 1; i
< unroll_number
; i
++)
913 emit_label_after (labels
[unroll_number
- i
],
914 PREV_INSN (loop_start
));
916 bzero ((char *) map
->insn_map
, max_insnno
* sizeof (rtx
));
917 bzero ((char *) map
->const_equiv_map
, maxregnum
* sizeof (rtx
));
918 bzero ((char *) map
->const_age_map
,
919 maxregnum
* sizeof (unsigned));
922 for (j
= 0; j
< max_labelno
; j
++)
924 map
->label_map
[j
] = gen_label_rtx ();
926 /* The last copy needs the compare/branch insns at the end,
927 so reset copy_end here if the loop ends with a conditional
930 if (i
== unroll_number
- 1)
932 if (GET_CODE (last_loop_insn
) == BARRIER
)
933 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
935 copy_end
= last_loop_insn
;
938 /* None of the copies are the `last_iteration', so just
939 pass zero for that parameter. */
940 copy_loop_body (copy_start
, copy_end
, map
, exit_label
, 0,
941 unroll_type
, start_label
, loop_end
,
942 loop_start
, copy_end
);
944 emit_label_after (labels
[0], PREV_INSN (loop_start
));
946 if (GET_CODE (last_loop_insn
) == BARRIER
)
948 insert_before
= PREV_INSN (last_loop_insn
);
949 copy_end
= PREV_INSN (insert_before
);
954 /* The immediately preceding insn is a compare which we do not
956 insert_before
= PREV_INSN (last_loop_insn
);
957 copy_end
= PREV_INSN (insert_before
);
959 /* The immediately preceding insn may not be a compare, so we
961 insert_before
= last_loop_insn
;
962 copy_end
= PREV_INSN (last_loop_insn
);
966 /* Set unroll type to MODULO now. */
967 unroll_type
= UNROLL_MODULO
;
968 loop_preconditioned
= 1;
972 /* If reach here, and the loop type is UNROLL_NAIVE, then don't unroll
973 the loop unless all loops are being unrolled. */
974 if (unroll_type
== UNROLL_NAIVE
&& ! flag_unroll_all_loops
)
976 if (loop_dump_stream
)
977 fprintf (loop_dump_stream
, "Unrolling failure: Naive unrolling not being done.\n");
981 /* At this point, we are guaranteed to unroll the loop. */
983 /* For each biv and giv, determine whether it can be safely split into
984 a different variable for each unrolled copy of the loop body.
985 We precalculate and save this info here, since computing it is
988 Do this before deleting any instructions from the loop, so that
989 back_branch_in_range_p will work correctly. */
991 if (splitting_not_safe
)
994 temp
= find_splittable_regs (unroll_type
, loop_start
, loop_end
,
995 end_insert_before
, unroll_number
);
997 /* find_splittable_regs may have created some new registers, so must
998 reallocate the reg_map with the new larger size, and must realloc
999 the constant maps also. */
1001 maxregnum
= max_reg_num ();
1002 map
->reg_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
1004 init_reg_map (map
, maxregnum
);
1006 /* Space is needed in some of the map for new registers, so new_maxregnum
1007 is an (over)estimate of how many registers will exist at the end. */
1008 new_maxregnum
= maxregnum
+ (temp
* unroll_number
* 2);
1010 /* Must realloc space for the constant maps, because the number of registers
1011 may have changed. */
1013 map
->const_equiv_map
= (rtx
*) alloca (new_maxregnum
* sizeof (rtx
));
1014 map
->const_age_map
= (unsigned *) alloca (new_maxregnum
* sizeof (unsigned));
1016 map
->const_equiv_map_size
= new_maxregnum
;
1017 global_const_equiv_map
= map
->const_equiv_map
;
1018 global_const_equiv_map_size
= new_maxregnum
;
1020 /* Search the list of bivs and givs to find ones which need to be remapped
1021 when split, and set their reg_map entry appropriately. */
1023 for (bl
= loop_iv_list
; bl
; bl
= bl
->next
)
1025 if (REGNO (bl
->biv
->src_reg
) != bl
->regno
)
1026 map
->reg_map
[bl
->regno
] = bl
->biv
->src_reg
;
1028 /* Currently, non-reduced/final-value givs are never split. */
1029 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
1030 if (REGNO (v
->src_reg
) != bl
->regno
)
1031 map
->reg_map
[REGNO (v
->dest_reg
)] = v
->src_reg
;
1035 /* If the loop is being partially unrolled, and the iteration variables
1036 are being split, and are being renamed for the split, then must fix up
1037 the compare/jump instruction at the end of the loop to refer to the new
1038 registers. This compare isn't copied, so the registers used in it
1039 will never be replaced if it isn't done here. */
1041 if (unroll_type
== UNROLL_MODULO
)
1043 insn
= NEXT_INSN (copy_end
);
1044 if (GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == JUMP_INSN
)
1045 PATTERN (insn
) = remap_split_bivs (PATTERN (insn
));
1048 /* For unroll_number - 1 times, make a copy of each instruction
1049 between copy_start and copy_end, and insert these new instructions
1050 before the end of the loop. */
1052 for (i
= 0; i
< unroll_number
; i
++)
1054 bzero ((char *) map
->insn_map
, max_insnno
* sizeof (rtx
));
1055 bzero ((char *) map
->const_equiv_map
, new_maxregnum
* sizeof (rtx
));
1056 bzero ((char *) map
->const_age_map
, new_maxregnum
* sizeof (unsigned));
1059 for (j
= 0; j
< max_labelno
; j
++)
1061 map
->label_map
[j
] = gen_label_rtx ();
1063 /* If loop starts with a branch to the test, then fix it so that
1064 it points to the test of the first unrolled copy of the loop. */
1065 if (i
== 0 && loop_start
!= copy_start
)
1067 insn
= PREV_INSN (copy_start
);
1068 pattern
= PATTERN (insn
);
1070 tem
= map
->label_map
[CODE_LABEL_NUMBER
1071 (XEXP (SET_SRC (pattern
), 0))];
1072 SET_SRC (pattern
) = gen_rtx (LABEL_REF
, VOIDmode
, tem
);
1074 /* Set the jump label so that it can be used by later loop unrolling
1076 JUMP_LABEL (insn
) = tem
;
1077 LABEL_NUSES (tem
)++;
1080 copy_loop_body (copy_start
, copy_end
, map
, exit_label
,
1081 i
== unroll_number
- 1, unroll_type
, start_label
,
1082 loop_end
, insert_before
, insert_before
);
1085 /* Before deleting any insns, emit a CODE_LABEL immediately after the last
1086 insn to be deleted. This prevents any runaway delete_insn call from
1087 more insns that it should, as it always stops at a CODE_LABEL. */
1089 /* Delete the compare and branch at the end of the loop if completely
1090 unrolling the loop. Deleting the backward branch at the end also
1091 deletes the code label at the start of the loop. This is done at
1092 the very end to avoid problems with back_branch_in_range_p. */
1094 if (unroll_type
== UNROLL_COMPLETELY
)
1095 safety_label
= emit_label_after (gen_label_rtx (), last_loop_insn
);
1097 safety_label
= emit_label_after (gen_label_rtx (), copy_end
);
1099 /* Delete all of the original loop instructions. Don't delete the
1100 LOOP_BEG note, or the first code label in the loop. */
1102 insn
= NEXT_INSN (copy_start
);
1103 while (insn
!= safety_label
)
1105 if (insn
!= start_label
)
1106 insn
= delete_insn (insn
);
1108 insn
= NEXT_INSN (insn
);
1111 /* Can now delete the 'safety' label emitted to protect us from runaway
1112 delete_insn calls. */
1113 if (INSN_DELETED_P (safety_label
))
1115 delete_insn (safety_label
);
1117 /* If exit_label exists, emit it after the loop. Doing the emit here
1118 forces it to have a higher INSN_UID than any insn in the unrolled loop.
1119 This is needed so that mostly_true_jump in reorg.c will treat jumps
1120 to this loop end label correctly, i.e. predict that they are usually
1123 emit_label_after (exit_label
, loop_end
);
1126 /* Return true if the loop can be safely, and profitably, preconditioned
1127 so that the unrolled copies of the loop body don't need exit tests.
1129 This only works if final_value, initial_value and increment can be
1130 determined, and if increment is a constant power of 2.
1131 If increment is not a power of 2, then the preconditioning modulo
1132 operation would require a real modulo instead of a boolean AND, and this
1133 is not considered `profitable'. */
1135 /* ??? If the loop is known to be executed very many times, or the machine
1136 has a very cheap divide instruction, then preconditioning is a win even
1137 when the increment is not a power of 2. Use RTX_COST to compute
1138 whether divide is cheap. */
1141 precondition_loop_p (initial_value
, final_value
, increment
, loop_start
,
1143 rtx
*initial_value
, *final_value
, *increment
;
1144 rtx loop_start
, loop_end
;
1147 if (loop_n_iterations
> 0)
1149 *initial_value
= const0_rtx
;
1150 *increment
= const1_rtx
;
1151 *final_value
= GEN_INT (loop_n_iterations
);
1153 if (loop_dump_stream
)
1154 fprintf (loop_dump_stream
,
1155 "Preconditioning: Success, number of iterations known, %d.\n",
1160 if (loop_initial_value
== 0)
1162 if (loop_dump_stream
)
1163 fprintf (loop_dump_stream
,
1164 "Preconditioning: Could not find initial value.\n");
1167 else if (loop_increment
== 0)
1169 if (loop_dump_stream
)
1170 fprintf (loop_dump_stream
,
1171 "Preconditioning: Could not find increment value.\n");
1174 else if (GET_CODE (loop_increment
) != CONST_INT
)
1176 if (loop_dump_stream
)
1177 fprintf (loop_dump_stream
,
1178 "Preconditioning: Increment not a constant.\n");
1181 else if ((exact_log2 (INTVAL (loop_increment
)) < 0)
1182 && (exact_log2 (- INTVAL (loop_increment
)) < 0))
1184 if (loop_dump_stream
)
1185 fprintf (loop_dump_stream
,
1186 "Preconditioning: Increment not a constant power of 2.\n");
1190 /* Unsigned_compare and compare_dir can be ignored here, since they do
1191 not matter for preconditioning. */
1193 if (loop_final_value
== 0)
1195 if (loop_dump_stream
)
1196 fprintf (loop_dump_stream
,
1197 "Preconditioning: EQ comparison loop.\n");
1201 /* Must ensure that final_value is invariant, so call invariant_p to
1202 check. Before doing so, must check regno against max_reg_before_loop
1203 to make sure that the register is in the range covered by invariant_p.
1204 If it isn't, then it is most likely a biv/giv which by definition are
1206 if ((GET_CODE (loop_final_value
) == REG
1207 && REGNO (loop_final_value
) >= max_reg_before_loop
)
1208 || (GET_CODE (loop_final_value
) == PLUS
1209 && REGNO (XEXP (loop_final_value
, 0)) >= max_reg_before_loop
)
1210 || ! invariant_p (loop_final_value
))
1212 if (loop_dump_stream
)
1213 fprintf (loop_dump_stream
,
1214 "Preconditioning: Final value not invariant.\n");
1218 /* Fail for floating point values, since the caller of this function
1219 does not have code to deal with them. */
1220 if (GET_MODE_CLASS (GET_MODE (loop_final_value
)) == MODE_FLOAT
1221 || GET_MODE_CLASS (GET_MODE (loop_initial_value
)) == MODE_FLOAT
)
1223 if (loop_dump_stream
)
1224 fprintf (loop_dump_stream
,
1225 "Preconditioning: Floating point final or initial value.\n");
1229 /* Now set initial_value to be the iteration_var, since that may be a
1230 simpler expression, and is guaranteed to be correct if all of the
1231 above tests succeed.
1233 We can not use the initial_value as calculated, because it will be
1234 one too small for loops of the form "while (i-- > 0)". We can not
1235 emit code before the loop_skip_over insns to fix this problem as this
1236 will then give a number one too large for loops of the form
1239 Note that all loops that reach here are entered at the top, because
1240 this function is not called if the loop starts with a jump. */
1242 /* Fail if loop_iteration_var is not live before loop_start, since we need
1243 to test its value in the preconditioning code. */
1245 if (uid_luid
[regno_first_uid
[REGNO (loop_iteration_var
)]]
1246 > INSN_LUID (loop_start
))
1248 if (loop_dump_stream
)
1249 fprintf (loop_dump_stream
,
1250 "Preconditioning: Iteration var not live before loop start.\n");
1254 *initial_value
= loop_iteration_var
;
1255 *increment
= loop_increment
;
1256 *final_value
= loop_final_value
;
1259 if (loop_dump_stream
)
1260 fprintf (loop_dump_stream
, "Preconditioning: Successful.\n");
1265 /* All pseudo-registers must be mapped to themselves. Two hard registers
1266 must be mapped, VIRTUAL_STACK_VARS_REGNUM and VIRTUAL_INCOMING_ARGS_
1267 REGNUM, to avoid function-inlining specific conversions of these
1268 registers. All other hard regs can not be mapped because they may be
1273 init_reg_map (map
, maxregnum
)
1274 struct inline_remap
*map
;
1279 for (i
= maxregnum
- 1; i
> LAST_VIRTUAL_REGISTER
; i
--)
1280 map
->reg_map
[i
] = regno_reg_rtx
[i
];
1281 /* Just clear the rest of the entries. */
1282 for (i
= LAST_VIRTUAL_REGISTER
; i
>= 0; i
--)
1283 map
->reg_map
[i
] = 0;
1285 map
->reg_map
[VIRTUAL_STACK_VARS_REGNUM
]
1286 = regno_reg_rtx
[VIRTUAL_STACK_VARS_REGNUM
];
1287 map
->reg_map
[VIRTUAL_INCOMING_ARGS_REGNUM
]
1288 = regno_reg_rtx
[VIRTUAL_INCOMING_ARGS_REGNUM
];
1291 /* Strength-reduction will often emit code for optimized biv/givs which
1292 calculates their value in a temporary register, and then copies the result
1293 to the iv. This procedure reconstructs the pattern computing the iv;
1294 verifying that all operands are of the proper form.
1296 The return value is the amount that the giv is incremented by. */
1299 calculate_giv_inc (pattern
, src_insn
, regno
)
1300 rtx pattern
, src_insn
;
1304 rtx increment_total
= 0;
1308 /* Verify that we have an increment insn here. First check for a plus
1309 as the set source. */
1310 if (GET_CODE (SET_SRC (pattern
)) != PLUS
)
1312 /* SR sometimes computes the new giv value in a temp, then copies it
1314 src_insn
= PREV_INSN (src_insn
);
1315 pattern
= PATTERN (src_insn
);
1316 if (GET_CODE (SET_SRC (pattern
)) != PLUS
)
1319 /* The last insn emitted is not needed, so delete it to avoid confusing
1320 the second cse pass. This insn sets the giv unnecessarily. */
1321 delete_insn (get_last_insn ());
1324 /* Verify that we have a constant as the second operand of the plus. */
1325 increment
= XEXP (SET_SRC (pattern
), 1);
1326 if (GET_CODE (increment
) != CONST_INT
)
1328 /* SR sometimes puts the constant in a register, especially if it is
1329 too big to be an add immed operand. */
1330 src_insn
= PREV_INSN (src_insn
);
1331 increment
= SET_SRC (PATTERN (src_insn
));
1333 /* SR may have used LO_SUM to compute the constant if it is too large
1334 for a load immed operand. In this case, the constant is in operand
1335 one of the LO_SUM rtx. */
1336 if (GET_CODE (increment
) == LO_SUM
)
1337 increment
= XEXP (increment
, 1);
1338 else if (GET_CODE (increment
) == IOR
)
1340 /* The rs6000 port loads some constants with IOR. */
1341 rtx second_part
= XEXP (increment
, 1);
1343 src_insn
= PREV_INSN (src_insn
);
1344 increment
= SET_SRC (PATTERN (src_insn
));
1345 /* Don't need the last insn anymore. */
1346 delete_insn (get_last_insn ());
1348 if (GET_CODE (second_part
) != CONST_INT
1349 || GET_CODE (increment
) != CONST_INT
)
1352 increment
= GEN_INT (INTVAL (increment
) | INTVAL (second_part
));
1355 if (GET_CODE (increment
) != CONST_INT
)
1358 /* The insn loading the constant into a register is no longer needed,
1360 delete_insn (get_last_insn ());
1363 if (increment_total
)
1364 increment_total
= GEN_INT (INTVAL (increment_total
) + INTVAL (increment
));
1366 increment_total
= increment
;
1368 /* Check that the source register is the same as the register we expected
1369 to see as the source. If not, something is seriously wrong. */
1370 if (GET_CODE (XEXP (SET_SRC (pattern
), 0)) != REG
1371 || REGNO (XEXP (SET_SRC (pattern
), 0)) != regno
)
1373 /* Some machines (e.g. the romp), may emit two add instructions for
1374 certain constants, so lets try looking for another add immediately
1375 before this one if we have only seen one add insn so far. */
1381 src_insn
= PREV_INSN (src_insn
);
1382 pattern
= PATTERN (src_insn
);
1384 delete_insn (get_last_insn ());
1392 return increment_total
;
1395 /* Copy REG_NOTES, except for insn references, because not all insn_map
1396 entries are valid yet. We do need to copy registers now though, because
1397 the reg_map entries can change during copying. */
1400 initial_reg_note_copy (notes
, map
)
1402 struct inline_remap
*map
;
1409 copy
= rtx_alloc (GET_CODE (notes
));
1410 PUT_MODE (copy
, GET_MODE (notes
));
1412 if (GET_CODE (notes
) == EXPR_LIST
)
1413 XEXP (copy
, 0) = copy_rtx_and_substitute (XEXP (notes
, 0), map
);
1414 else if (GET_CODE (notes
) == INSN_LIST
)
1415 /* Don't substitute for these yet. */
1416 XEXP (copy
, 0) = XEXP (notes
, 0);
1420 XEXP (copy
, 1) = initial_reg_note_copy (XEXP (notes
, 1), map
);
1425 /* Fixup insn references in copied REG_NOTES. */
1428 final_reg_note_copy (notes
, map
)
1430 struct inline_remap
*map
;
1434 for (note
= notes
; note
; note
= XEXP (note
, 1))
1435 if (GET_CODE (note
) == INSN_LIST
)
1436 XEXP (note
, 0) = map
->insn_map
[INSN_UID (XEXP (note
, 0))];
1439 /* Copy each instruction in the loop, substituting from map as appropriate.
1440 This is very similar to a loop in expand_inline_function. */
1443 copy_loop_body (copy_start
, copy_end
, map
, exit_label
, last_iteration
,
1444 unroll_type
, start_label
, loop_end
, insert_before
,
1446 rtx copy_start
, copy_end
;
1447 struct inline_remap
*map
;
1450 enum unroll_types unroll_type
;
1451 rtx start_label
, loop_end
, insert_before
, copy_notes_from
;
1455 int dest_reg_was_split
, i
;
1457 rtx final_label
= 0;
1458 rtx giv_inc
, giv_dest_reg
, giv_src_reg
;
1460 /* If this isn't the last iteration, then map any references to the
1461 start_label to final_label. Final label will then be emitted immediately
1462 after the end of this loop body if it was ever used.
1464 If this is the last iteration, then map references to the start_label
1466 if (! last_iteration
)
1468 final_label
= gen_label_rtx ();
1469 map
->label_map
[CODE_LABEL_NUMBER (start_label
)] = final_label
;
1472 map
->label_map
[CODE_LABEL_NUMBER (start_label
)] = start_label
;
1479 insn
= NEXT_INSN (insn
);
1481 map
->orig_asm_operands_vector
= 0;
1483 switch (GET_CODE (insn
))
1486 pattern
= PATTERN (insn
);
1490 /* Check to see if this is a giv that has been combined with
1491 some split address givs. (Combined in the sense that
1492 `combine_givs' in loop.c has put two givs in the same register.)
1493 In this case, we must search all givs based on the same biv to
1494 find the address givs. Then split the address givs.
1495 Do this before splitting the giv, since that may map the
1496 SET_DEST to a new register. */
1498 if (GET_CODE (pattern
) == SET
1499 && GET_CODE (SET_DEST (pattern
)) == REG
1500 && addr_combined_regs
[REGNO (SET_DEST (pattern
))])
1502 struct iv_class
*bl
;
1503 struct induction
*v
, *tv
;
1504 int regno
= REGNO (SET_DEST (pattern
));
1506 v
= addr_combined_regs
[REGNO (SET_DEST (pattern
))];
1507 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
1509 /* Although the giv_inc amount is not needed here, we must call
1510 calculate_giv_inc here since it might try to delete the
1511 last insn emitted. If we wait until later to call it,
1512 we might accidentally delete insns generated immediately
1513 below by emit_unrolled_add. */
1515 giv_inc
= calculate_giv_inc (pattern
, insn
, regno
);
1517 /* Now find all address giv's that were combined with this
1519 for (tv
= bl
->giv
; tv
; tv
= tv
->next_iv
)
1520 if (tv
->giv_type
== DEST_ADDR
&& tv
->same
== v
)
1522 int this_giv_inc
= INTVAL (giv_inc
);
1524 /* Scale this_giv_inc if the multiplicative factors of
1525 the two givs are different. */
1526 if (tv
->mult_val
!= v
->mult_val
)
1527 this_giv_inc
= (this_giv_inc
/ INTVAL (v
->mult_val
)
1528 * INTVAL (tv
->mult_val
));
1530 tv
->dest_reg
= plus_constant (tv
->dest_reg
, this_giv_inc
);
1531 *tv
->location
= tv
->dest_reg
;
1533 if (last_iteration
&& unroll_type
!= UNROLL_COMPLETELY
)
1535 /* Must emit an insn to increment the split address
1536 giv. Add in the const_adjust field in case there
1537 was a constant eliminated from the address. */
1538 rtx value
, dest_reg
;
1540 /* tv->dest_reg will be either a bare register,
1541 or else a register plus a constant. */
1542 if (GET_CODE (tv
->dest_reg
) == REG
)
1543 dest_reg
= tv
->dest_reg
;
1545 dest_reg
= XEXP (tv
->dest_reg
, 0);
1547 /* Check for shared address givs, and avoid
1548 incrementing the shared psuedo reg more than
1550 if (! (tv
!= v
&& tv
->insn
== v
->insn
1551 && tv
->new_reg
== v
->new_reg
))
1553 /* tv->dest_reg may actually be a (PLUS (REG)
1554 (CONST)) here, so we must call plus_constant
1555 to add the const_adjust amount before calling
1556 emit_unrolled_add below. */
1557 value
= plus_constant (tv
->dest_reg
,
1560 /* The constant could be too large for an add
1561 immediate, so can't directly emit an insn
1563 emit_unrolled_add (dest_reg
, XEXP (value
, 0),
1567 /* Reset the giv to be just the register again, in case
1568 it is used after the set we have just emitted.
1569 We must subtract the const_adjust factor added in
1571 tv
->dest_reg
= plus_constant (dest_reg
,
1572 - tv
->const_adjust
);
1573 *tv
->location
= tv
->dest_reg
;
1578 /* If this is a setting of a splittable variable, then determine
1579 how to split the variable, create a new set based on this split,
1580 and set up the reg_map so that later uses of the variable will
1581 use the new split variable. */
1583 dest_reg_was_split
= 0;
1585 if (GET_CODE (pattern
) == SET
1586 && GET_CODE (SET_DEST (pattern
)) == REG
1587 && splittable_regs
[REGNO (SET_DEST (pattern
))])
1589 int regno
= REGNO (SET_DEST (pattern
));
1591 dest_reg_was_split
= 1;
1593 /* Compute the increment value for the giv, if it wasn't
1594 already computed above. */
1597 giv_inc
= calculate_giv_inc (pattern
, insn
, regno
);
1598 giv_dest_reg
= SET_DEST (pattern
);
1599 giv_src_reg
= SET_DEST (pattern
);
1601 if (unroll_type
== UNROLL_COMPLETELY
)
1603 /* Completely unrolling the loop. Set the induction
1604 variable to a known constant value. */
1606 /* The value in splittable_regs may be an invariant
1607 value, so we must use plus_constant here. */
1608 splittable_regs
[regno
]
1609 = plus_constant (splittable_regs
[regno
], INTVAL (giv_inc
));
1611 if (GET_CODE (splittable_regs
[regno
]) == PLUS
)
1613 giv_src_reg
= XEXP (splittable_regs
[regno
], 0);
1614 giv_inc
= XEXP (splittable_regs
[regno
], 1);
1618 /* The splittable_regs value must be a REG or a
1619 CONST_INT, so put the entire value in the giv_src_reg
1621 giv_src_reg
= splittable_regs
[regno
];
1622 giv_inc
= const0_rtx
;
1627 /* Partially unrolling loop. Create a new pseudo
1628 register for the iteration variable, and set it to
1629 be a constant plus the original register. Except
1630 on the last iteration, when the result has to
1631 go back into the original iteration var register. */
1633 /* Handle bivs which must be mapped to a new register
1634 when split. This happens for bivs which need their
1635 final value set before loop entry. The new register
1636 for the biv was stored in the biv's first struct
1637 induction entry by find_splittable_regs. */
1639 if (regno
< max_reg_before_loop
1640 && reg_iv_type
[regno
] == BASIC_INDUCT
)
1642 giv_src_reg
= reg_biv_class
[regno
]->biv
->src_reg
;
1643 giv_dest_reg
= giv_src_reg
;
1647 /* If non-reduced/final-value givs were split, then
1648 this would have to remap those givs also. See
1649 find_splittable_regs. */
1652 splittable_regs
[regno
]
1653 = GEN_INT (INTVAL (giv_inc
)
1654 + INTVAL (splittable_regs
[regno
]));
1655 giv_inc
= splittable_regs
[regno
];
1657 /* Now split the induction variable by changing the dest
1658 of this insn to a new register, and setting its
1659 reg_map entry to point to this new register.
1661 If this is the last iteration, and this is the last insn
1662 that will update the iv, then reuse the original dest,
1663 to ensure that the iv will have the proper value when
1664 the loop exits or repeats.
1666 Using splittable_regs_updates here like this is safe,
1667 because it can only be greater than one if all
1668 instructions modifying the iv are always executed in
1671 if (! last_iteration
1672 || (splittable_regs_updates
[regno
]-- != 1))
1674 tem
= gen_reg_rtx (GET_MODE (giv_src_reg
));
1676 map
->reg_map
[regno
] = tem
;
1679 map
->reg_map
[regno
] = giv_src_reg
;
1682 /* The constant being added could be too large for an add
1683 immediate, so can't directly emit an insn here. */
1684 emit_unrolled_add (giv_dest_reg
, giv_src_reg
, giv_inc
);
1685 copy
= get_last_insn ();
1686 pattern
= PATTERN (copy
);
1690 pattern
= copy_rtx_and_substitute (pattern
, map
);
1691 copy
= emit_insn (pattern
);
1693 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
1696 /* If this insn is setting CC0, it may need to look at
1697 the insn that uses CC0 to see what type of insn it is.
1698 In that case, the call to recog via validate_change will
1699 fail. So don't substitute constants here. Instead,
1700 do it when we emit the following insn.
1702 For example, see the pyr.md file. That machine has signed and
1703 unsigned compares. The compare patterns must check the
1704 following branch insn to see which what kind of compare to
1707 If the previous insn set CC0, substitute constants on it as
1709 if (sets_cc0_p (copy
) != 0)
1714 try_constants (cc0_insn
, map
);
1716 try_constants (copy
, map
);
1719 try_constants (copy
, map
);
1722 /* Make split induction variable constants `permanent' since we
1723 know there are no backward branches across iteration variable
1724 settings which would invalidate this. */
1725 if (dest_reg_was_split
)
1727 int regno
= REGNO (SET_DEST (pattern
));
1729 if (regno
< map
->const_equiv_map_size
1730 && map
->const_age_map
[regno
] == map
->const_age
)
1731 map
->const_age_map
[regno
] = -1;
1736 pattern
= copy_rtx_and_substitute (PATTERN (insn
), map
);
1737 copy
= emit_jump_insn (pattern
);
1738 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
1740 if (JUMP_LABEL (insn
) == start_label
&& insn
== copy_end
1741 && ! last_iteration
)
1743 /* This is a branch to the beginning of the loop; this is the
1744 last insn being copied; and this is not the last iteration.
1745 In this case, we want to change the original fall through
1746 case to be a branch past the end of the loop, and the
1747 original jump label case to fall_through. */
1749 if (! invert_exp (pattern
, copy
)
1750 || ! redirect_exp (&pattern
,
1751 map
->label_map
[CODE_LABEL_NUMBER
1752 (JUMP_LABEL (insn
))],
1759 try_constants (cc0_insn
, map
);
1762 try_constants (copy
, map
);
1764 /* Set the jump label of COPY correctly to avoid problems with
1765 later passes of unroll_loop, if INSN had jump label set. */
1766 if (JUMP_LABEL (insn
))
1770 /* Can't use the label_map for every insn, since this may be
1771 the backward branch, and hence the label was not mapped. */
1772 if (GET_CODE (pattern
) == SET
)
1774 tem
= SET_SRC (pattern
);
1775 if (GET_CODE (tem
) == LABEL_REF
)
1776 label
= XEXP (tem
, 0);
1777 else if (GET_CODE (tem
) == IF_THEN_ELSE
)
1779 if (XEXP (tem
, 1) != pc_rtx
)
1780 label
= XEXP (XEXP (tem
, 1), 0);
1782 label
= XEXP (XEXP (tem
, 2), 0);
1786 if (label
&& GET_CODE (label
) == CODE_LABEL
)
1787 JUMP_LABEL (copy
) = label
;
1790 /* An unrecognizable jump insn, probably the entry jump
1791 for a switch statement. This label must have been mapped,
1792 so just use the label_map to get the new jump label. */
1793 JUMP_LABEL (copy
) = map
->label_map
[CODE_LABEL_NUMBER
1794 (JUMP_LABEL (insn
))];
1797 /* If this is a non-local jump, then must increase the label
1798 use count so that the label will not be deleted when the
1799 original jump is deleted. */
1800 LABEL_NUSES (JUMP_LABEL (copy
))++;
1802 else if (GET_CODE (PATTERN (copy
)) == ADDR_VEC
1803 || GET_CODE (PATTERN (copy
)) == ADDR_DIFF_VEC
)
1805 rtx pat
= PATTERN (copy
);
1806 int diff_vec_p
= GET_CODE (pat
) == ADDR_DIFF_VEC
;
1807 int len
= XVECLEN (pat
, diff_vec_p
);
1810 for (i
= 0; i
< len
; i
++)
1811 LABEL_NUSES (XEXP (XVECEXP (pat
, diff_vec_p
, i
), 0))++;
1814 /* If this used to be a conditional jump insn but whose branch
1815 direction is now known, we must do something special. */
1816 if (condjump_p (insn
) && !simplejump_p (insn
) && map
->last_pc_value
)
1819 /* The previous insn set cc0 for us. So delete it. */
1820 delete_insn (PREV_INSN (copy
));
1823 /* If this is now a no-op, delete it. */
1824 if (map
->last_pc_value
== pc_rtx
)
1826 /* Don't let delete_insn delete the label referenced here,
1827 because we might possibly need it later for some other
1828 instruction in the loop. */
1829 if (JUMP_LABEL (copy
))
1830 LABEL_NUSES (JUMP_LABEL (copy
))++;
1832 if (JUMP_LABEL (copy
))
1833 LABEL_NUSES (JUMP_LABEL (copy
))--;
1837 /* Otherwise, this is unconditional jump so we must put a
1838 BARRIER after it. We could do some dead code elimination
1839 here, but jump.c will do it just as well. */
1845 pattern
= copy_rtx_and_substitute (PATTERN (insn
), map
);
1846 copy
= emit_call_insn (pattern
);
1847 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
1849 /* Because the USAGE information potentially contains objects other
1850 than hard registers, we need to copy it. */
1851 CALL_INSN_FUNCTION_USAGE (copy
) =
1852 copy_rtx_and_substitute (CALL_INSN_FUNCTION_USAGE (insn
), map
);
1856 try_constants (cc0_insn
, map
);
1859 try_constants (copy
, map
);
1861 /* Be lazy and assume CALL_INSNs clobber all hard registers. */
1862 for (i
= 0; i
< FIRST_PSEUDO_REGISTER
; i
++)
1863 map
->const_equiv_map
[i
] = 0;
1867 /* If this is the loop start label, then we don't need to emit a
1868 copy of this label since no one will use it. */
1870 if (insn
!= start_label
)
1872 copy
= emit_label (map
->label_map
[CODE_LABEL_NUMBER (insn
)]);
1878 copy
= emit_barrier ();
1882 /* VTOP notes are valid only before the loop exit test. If placed
1883 anywhere else, loop may generate bad code. */
1885 if (NOTE_LINE_NUMBER (insn
) != NOTE_INSN_DELETED
1886 && (NOTE_LINE_NUMBER (insn
) != NOTE_INSN_LOOP_VTOP
1887 || (last_iteration
&& unroll_type
!= UNROLL_COMPLETELY
)))
1888 copy
= emit_note (NOTE_SOURCE_FILE (insn
),
1889 NOTE_LINE_NUMBER (insn
));
1899 map
->insn_map
[INSN_UID (insn
)] = copy
;
1901 while (insn
!= copy_end
);
1903 /* Now finish coping the REG_NOTES. */
1907 insn
= NEXT_INSN (insn
);
1908 if ((GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == JUMP_INSN
1909 || GET_CODE (insn
) == CALL_INSN
)
1910 && map
->insn_map
[INSN_UID (insn
)])
1911 final_reg_note_copy (REG_NOTES (map
->insn_map
[INSN_UID (insn
)]), map
);
1913 while (insn
!= copy_end
);
1915 /* There may be notes between copy_notes_from and loop_end. Emit a copy of
1916 each of these notes here, since there may be some important ones, such as
1917 NOTE_INSN_BLOCK_END notes, in this group. We don't do this on the last
1918 iteration, because the original notes won't be deleted.
1920 We can't use insert_before here, because when from preconditioning,
1921 insert_before points before the loop. We can't use copy_end, because
1922 there may be insns already inserted after it (which we don't want to
1923 copy) when not from preconditioning code. */
1925 if (! last_iteration
)
1927 for (insn
= copy_notes_from
; insn
!= loop_end
; insn
= NEXT_INSN (insn
))
1929 if (GET_CODE (insn
) == NOTE
1930 && NOTE_LINE_NUMBER (insn
) != NOTE_INSN_DELETED
)
1931 emit_note (NOTE_SOURCE_FILE (insn
), NOTE_LINE_NUMBER (insn
));
1935 if (final_label
&& LABEL_NUSES (final_label
) > 0)
1936 emit_label (final_label
);
1938 tem
= gen_sequence ();
1940 emit_insn_before (tem
, insert_before
);
1943 /* Emit an insn, using the expand_binop to ensure that a valid insn is
1944 emitted. This will correctly handle the case where the increment value
1945 won't fit in the immediate field of a PLUS insns. */
1948 emit_unrolled_add (dest_reg
, src_reg
, increment
)
1949 rtx dest_reg
, src_reg
, increment
;
1953 result
= expand_binop (GET_MODE (dest_reg
), add_optab
, src_reg
, increment
,
1954 dest_reg
, 0, OPTAB_LIB_WIDEN
);
1956 if (dest_reg
!= result
)
1957 emit_move_insn (dest_reg
, result
);
1960 /* Searches the insns between INSN and LOOP_END. Returns 1 if there
1961 is a backward branch in that range that branches to somewhere between
1962 LOOP_START and INSN. Returns 0 otherwise. */
1964 /* ??? This is quadratic algorithm. Could be rewritten to be linear.
1965 In practice, this is not a problem, because this function is seldom called,
1966 and uses a negligible amount of CPU time on average. */
1969 back_branch_in_range_p (insn
, loop_start
, loop_end
)
1971 rtx loop_start
, loop_end
;
1973 rtx p
, q
, target_insn
;
1975 /* Stop before we get to the backward branch at the end of the loop. */
1976 loop_end
= prev_nonnote_insn (loop_end
);
1977 if (GET_CODE (loop_end
) == BARRIER
)
1978 loop_end
= PREV_INSN (loop_end
);
1980 /* Check in case insn has been deleted, search forward for first non
1981 deleted insn following it. */
1982 while (INSN_DELETED_P (insn
))
1983 insn
= NEXT_INSN (insn
);
1985 /* Check for the case where insn is the last insn in the loop. */
1986 if (insn
== loop_end
)
1989 for (p
= NEXT_INSN (insn
); p
!= loop_end
; p
= NEXT_INSN (p
))
1991 if (GET_CODE (p
) == JUMP_INSN
)
1993 target_insn
= JUMP_LABEL (p
);
1995 /* Search from loop_start to insn, to see if one of them is
1996 the target_insn. We can't use INSN_LUID comparisons here,
1997 since insn may not have an LUID entry. */
1998 for (q
= loop_start
; q
!= insn
; q
= NEXT_INSN (q
))
1999 if (q
== target_insn
)
2007 /* Try to generate the simplest rtx for the expression
2008 (PLUS (MULT mult1 mult2) add1). This is used to calculate the initial
2012 fold_rtx_mult_add (mult1
, mult2
, add1
, mode
)
2013 rtx mult1
, mult2
, add1
;
2014 enum machine_mode mode
;
2019 /* The modes must all be the same. This should always be true. For now,
2020 check to make sure. */
2021 if ((GET_MODE (mult1
) != mode
&& GET_MODE (mult1
) != VOIDmode
)
2022 || (GET_MODE (mult2
) != mode
&& GET_MODE (mult2
) != VOIDmode
)
2023 || (GET_MODE (add1
) != mode
&& GET_MODE (add1
) != VOIDmode
))
2026 /* Ensure that if at least one of mult1/mult2 are constant, then mult2
2027 will be a constant. */
2028 if (GET_CODE (mult1
) == CONST_INT
)
2035 mult_res
= simplify_binary_operation (MULT
, mode
, mult1
, mult2
);
2037 mult_res
= gen_rtx (MULT
, mode
, mult1
, mult2
);
2039 /* Again, put the constant second. */
2040 if (GET_CODE (add1
) == CONST_INT
)
2047 result
= simplify_binary_operation (PLUS
, mode
, add1
, mult_res
);
2049 result
= gen_rtx (PLUS
, mode
, add1
, mult_res
);
2054 /* Searches the list of induction struct's for the biv BL, to try to calculate
2055 the total increment value for one iteration of the loop as a constant.
2057 Returns the increment value as an rtx, simplified as much as possible,
2058 if it can be calculated. Otherwise, returns 0. */
2061 biv_total_increment (bl
, loop_start
, loop_end
)
2062 struct iv_class
*bl
;
2063 rtx loop_start
, loop_end
;
2065 struct induction
*v
;
2068 /* For increment, must check every instruction that sets it. Each
2069 instruction must be executed only once each time through the loop.
2070 To verify this, we check that the the insn is always executed, and that
2071 there are no backward branches after the insn that branch to before it.
2072 Also, the insn must have a mult_val of one (to make sure it really is
2075 result
= const0_rtx
;
2076 for (v
= bl
->biv
; v
; v
= v
->next_iv
)
2078 if (v
->always_computable
&& v
->mult_val
== const1_rtx
2079 && ! back_branch_in_range_p (v
->insn
, loop_start
, loop_end
))
2080 result
= fold_rtx_mult_add (result
, const1_rtx
, v
->add_val
, v
->mode
);
2088 /* Determine the initial value of the iteration variable, and the amount
2089 that it is incremented each loop. Use the tables constructed by
2090 the strength reduction pass to calculate these values.
2092 Initial_value and/or increment are set to zero if their values could not
2096 iteration_info (iteration_var
, initial_value
, increment
, loop_start
, loop_end
)
2097 rtx iteration_var
, *initial_value
, *increment
;
2098 rtx loop_start
, loop_end
;
2100 struct iv_class
*bl
;
2101 struct induction
*v
, *b
;
2103 /* Clear the result values, in case no answer can be found. */
2107 /* The iteration variable can be either a giv or a biv. Check to see
2108 which it is, and compute the variable's initial value, and increment
2109 value if possible. */
2111 /* If this is a new register, can't handle it since we don't have any
2112 reg_iv_type entry for it. */
2113 if (REGNO (iteration_var
) >= max_reg_before_loop
)
2115 if (loop_dump_stream
)
2116 fprintf (loop_dump_stream
,
2117 "Loop unrolling: No reg_iv_type entry for iteration var.\n");
2120 /* Reject iteration variables larger than the host long size, since they
2121 could result in a number of iterations greater than the range of our
2122 `unsigned long' variable loop_n_iterations. */
2123 else if (GET_MODE_BITSIZE (GET_MODE (iteration_var
)) > HOST_BITS_PER_LONG
)
2125 if (loop_dump_stream
)
2126 fprintf (loop_dump_stream
,
2127 "Loop unrolling: Iteration var rejected because mode larger than host long.\n");
2130 else if (GET_MODE_CLASS (GET_MODE (iteration_var
)) != MODE_INT
)
2132 if (loop_dump_stream
)
2133 fprintf (loop_dump_stream
,
2134 "Loop unrolling: Iteration var not an integer.\n");
2137 else if (reg_iv_type
[REGNO (iteration_var
)] == BASIC_INDUCT
)
2139 /* Grab initial value, only useful if it is a constant. */
2140 bl
= reg_biv_class
[REGNO (iteration_var
)];
2141 *initial_value
= bl
->initial_value
;
2143 *increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2145 else if (reg_iv_type
[REGNO (iteration_var
)] == GENERAL_INDUCT
)
2148 /* ??? The code below does not work because the incorrect number of
2149 iterations is calculated when the biv is incremented after the giv
2150 is set (which is the usual case). This can probably be accounted
2151 for by biasing the initial_value by subtracting the amount of the
2152 increment that occurs between the giv set and the giv test. However,
2153 a giv as an iterator is very rare, so it does not seem worthwhile
2155 /* ??? An example failure is: i = 6; do {;} while (i++ < 9). */
2156 if (loop_dump_stream
)
2157 fprintf (loop_dump_stream
,
2158 "Loop unrolling: Giv iterators are not handled.\n");
2161 /* Initial value is mult_val times the biv's initial value plus
2162 add_val. Only useful if it is a constant. */
2163 v
= reg_iv_info
[REGNO (iteration_var
)];
2164 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
2165 *initial_value
= fold_rtx_mult_add (v
->mult_val
, bl
->initial_value
,
2166 v
->add_val
, v
->mode
);
2168 /* Increment value is mult_val times the increment value of the biv. */
2170 *increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2172 *increment
= fold_rtx_mult_add (v
->mult_val
, *increment
, const0_rtx
,
2178 if (loop_dump_stream
)
2179 fprintf (loop_dump_stream
,
2180 "Loop unrolling: Not basic or general induction var.\n");
2185 /* Calculate the approximate final value of the iteration variable
2186 which has an loop exit test with code COMPARISON_CODE and comparison value
2187 of COMPARISON_VALUE. Also returns an indication of whether the comparison
2188 was signed or unsigned, and the direction of the comparison. This info is
2189 needed to calculate the number of loop iterations. */
2192 approx_final_value (comparison_code
, comparison_value
, unsigned_p
, compare_dir
)
2193 enum rtx_code comparison_code
;
2194 rtx comparison_value
;
2198 /* Calculate the final value of the induction variable.
2199 The exact final value depends on the branch operator, and increment sign.
2200 This is only an approximate value. It will be wrong if the iteration
2201 variable is not incremented by one each time through the loop, and
2202 approx final value - start value % increment != 0. */
2205 switch (comparison_code
)
2211 return plus_constant (comparison_value
, 1);
2216 return plus_constant (comparison_value
, -1);
2218 /* Can not calculate a final value for this case. */
2225 return comparison_value
;
2231 return comparison_value
;
2234 return comparison_value
;
2240 /* For each biv and giv, determine whether it can be safely split into
2241 a different variable for each unrolled copy of the loop body. If it
2242 is safe to split, then indicate that by saving some useful info
2243 in the splittable_regs array.
2245 If the loop is being completely unrolled, then splittable_regs will hold
2246 the current value of the induction variable while the loop is unrolled.
2247 It must be set to the initial value of the induction variable here.
2248 Otherwise, splittable_regs will hold the difference between the current
2249 value of the induction variable and the value the induction variable had
2250 at the top of the loop. It must be set to the value 0 here.
2252 Returns the total number of instructions that set registers that are
2255 /* ?? If the loop is only unrolled twice, then most of the restrictions to
2256 constant values are unnecessary, since we can easily calculate increment
2257 values in this case even if nothing is constant. The increment value
2258 should not involve a multiply however. */
2260 /* ?? Even if the biv/giv increment values aren't constant, it may still
2261 be beneficial to split the variable if the loop is only unrolled a few
2262 times, since multiplies by small integers (1,2,3,4) are very cheap. */
2265 find_splittable_regs (unroll_type
, loop_start
, loop_end
, end_insert_before
,
2267 enum unroll_types unroll_type
;
2268 rtx loop_start
, loop_end
;
2269 rtx end_insert_before
;
2272 struct iv_class
*bl
;
2273 struct induction
*v
;
2275 rtx biv_final_value
;
2279 for (bl
= loop_iv_list
; bl
; bl
= bl
->next
)
2281 /* Biv_total_increment must return a constant value,
2282 otherwise we can not calculate the split values. */
2284 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2285 if (! increment
|| GET_CODE (increment
) != CONST_INT
)
2288 /* The loop must be unrolled completely, or else have a known number
2289 of iterations and only one exit, or else the biv must be dead
2290 outside the loop, or else the final value must be known. Otherwise,
2291 it is unsafe to split the biv since it may not have the proper
2292 value on loop exit. */
2294 /* loop_number_exit_labels is non-zero if the loop has an exit other than
2295 a fall through at the end. */
2298 biv_final_value
= 0;
2299 if (unroll_type
!= UNROLL_COMPLETELY
2300 && (loop_number_exit_labels
[uid_loop_num
[INSN_UID (loop_start
)]]
2301 || unroll_type
== UNROLL_NAIVE
)
2302 && (uid_luid
[regno_last_uid
[bl
->regno
]] >= INSN_LUID (loop_end
)
2304 || INSN_UID (bl
->init_insn
) >= max_uid_for_loop
2305 || (uid_luid
[regno_first_uid
[bl
->regno
]]
2306 < INSN_LUID (bl
->init_insn
))
2307 || reg_mentioned_p (bl
->biv
->dest_reg
, SET_SRC (bl
->init_set
)))
2308 && ! (biv_final_value
= final_biv_value (bl
, loop_start
, loop_end
)))
2311 /* If any of the insns setting the BIV don't do so with a simple
2312 PLUS, we don't know how to split it. */
2313 for (v
= bl
->biv
; biv_splittable
&& v
; v
= v
->next_iv
)
2314 if ((tem
= single_set (v
->insn
)) == 0
2315 || GET_CODE (SET_DEST (tem
)) != REG
2316 || REGNO (SET_DEST (tem
)) != bl
->regno
2317 || GET_CODE (SET_SRC (tem
)) != PLUS
)
2320 /* If final value is non-zero, then must emit an instruction which sets
2321 the value of the biv to the proper value. This is done after
2322 handling all of the givs, since some of them may need to use the
2323 biv's value in their initialization code. */
2325 /* This biv is splittable. If completely unrolling the loop, save
2326 the biv's initial value. Otherwise, save the constant zero. */
2328 if (biv_splittable
== 1)
2330 if (unroll_type
== UNROLL_COMPLETELY
)
2332 /* If the initial value of the biv is itself (i.e. it is too
2333 complicated for strength_reduce to compute), or is a hard
2334 register, then we must create a new pseudo reg to hold the
2335 initial value of the biv. */
2337 if (GET_CODE (bl
->initial_value
) == REG
2338 && (REGNO (bl
->initial_value
) == bl
->regno
2339 || REGNO (bl
->initial_value
) < FIRST_PSEUDO_REGISTER
))
2341 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2343 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2346 if (loop_dump_stream
)
2347 fprintf (loop_dump_stream
, "Biv %d initial value remapped to %d.\n",
2348 bl
->regno
, REGNO (tem
));
2350 splittable_regs
[bl
->regno
] = tem
;
2353 splittable_regs
[bl
->regno
] = bl
->initial_value
;
2356 splittable_regs
[bl
->regno
] = const0_rtx
;
2358 /* Save the number of instructions that modify the biv, so that
2359 we can treat the last one specially. */
2361 splittable_regs_updates
[bl
->regno
] = bl
->biv_count
;
2362 result
+= bl
->biv_count
;
2364 if (loop_dump_stream
)
2365 fprintf (loop_dump_stream
,
2366 "Biv %d safe to split.\n", bl
->regno
);
2369 /* Check every giv that depends on this biv to see whether it is
2370 splittable also. Even if the biv isn't splittable, givs which
2371 depend on it may be splittable if the biv is live outside the
2372 loop, and the givs aren't. */
2374 result
+= find_splittable_givs (bl
, unroll_type
, loop_start
, loop_end
,
2375 increment
, unroll_number
);
2377 /* If final value is non-zero, then must emit an instruction which sets
2378 the value of the biv to the proper value. This is done after
2379 handling all of the givs, since some of them may need to use the
2380 biv's value in their initialization code. */
2381 if (biv_final_value
)
2383 /* If the loop has multiple exits, emit the insns before the
2384 loop to ensure that it will always be executed no matter
2385 how the loop exits. Otherwise emit the insn after the loop,
2386 since this is slightly more efficient. */
2387 if (! loop_number_exit_labels
[uid_loop_num
[INSN_UID (loop_start
)]])
2388 emit_insn_before (gen_move_insn (bl
->biv
->src_reg
,
2393 /* Create a new register to hold the value of the biv, and then
2394 set the biv to its final value before the loop start. The biv
2395 is set to its final value before loop start to ensure that
2396 this insn will always be executed, no matter how the loop
2398 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2399 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2401 emit_insn_before (gen_move_insn (bl
->biv
->src_reg
,
2405 if (loop_dump_stream
)
2406 fprintf (loop_dump_stream
, "Biv %d mapped to %d for split.\n",
2407 REGNO (bl
->biv
->src_reg
), REGNO (tem
));
2409 /* Set up the mapping from the original biv register to the new
2411 bl
->biv
->src_reg
= tem
;
2418 /* For every giv based on the biv BL, check to determine whether it is
2419 splittable. This is a subroutine to find_splittable_regs ().
2421 Return the number of instructions that set splittable registers. */
2424 find_splittable_givs (bl
, unroll_type
, loop_start
, loop_end
, increment
,
2426 struct iv_class
*bl
;
2427 enum unroll_types unroll_type
;
2428 rtx loop_start
, loop_end
;
2432 struct induction
*v
;
2437 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
2441 /* Only split the giv if it has already been reduced, or if the loop is
2442 being completely unrolled. */
2443 if (unroll_type
!= UNROLL_COMPLETELY
&& v
->ignore
)
2446 /* The giv can be split if the insn that sets the giv is executed once
2447 and only once on every iteration of the loop. */
2448 /* An address giv can always be split. v->insn is just a use not a set,
2449 and hence it does not matter whether it is always executed. All that
2450 matters is that all the biv increments are always executed, and we
2451 won't reach here if they aren't. */
2452 if (v
->giv_type
!= DEST_ADDR
2453 && (! v
->always_computable
2454 || back_branch_in_range_p (v
->insn
, loop_start
, loop_end
)))
2457 /* The giv increment value must be a constant. */
2458 giv_inc
= fold_rtx_mult_add (v
->mult_val
, increment
, const0_rtx
,
2460 if (! giv_inc
|| GET_CODE (giv_inc
) != CONST_INT
)
2463 /* The loop must be unrolled completely, or else have a known number of
2464 iterations and only one exit, or else the giv must be dead outside
2465 the loop, or else the final value of the giv must be known.
2466 Otherwise, it is not safe to split the giv since it may not have the
2467 proper value on loop exit. */
2469 /* The used outside loop test will fail for DEST_ADDR givs. They are
2470 never used outside the loop anyways, so it is always safe to split a
2474 if (unroll_type
!= UNROLL_COMPLETELY
2475 && (loop_number_exit_labels
[uid_loop_num
[INSN_UID (loop_start
)]]
2476 || unroll_type
== UNROLL_NAIVE
)
2477 && v
->giv_type
!= DEST_ADDR
2478 && ((regno_first_uid
[REGNO (v
->dest_reg
)] != INSN_UID (v
->insn
)
2479 /* Check for the case where the pseudo is set by a shift/add
2480 sequence, in which case the first insn setting the pseudo
2481 is the first insn of the shift/add sequence. */
2482 && (! (tem
= find_reg_note (v
->insn
, REG_RETVAL
, NULL_RTX
))
2483 || (regno_first_uid
[REGNO (v
->dest_reg
)]
2484 != INSN_UID (XEXP (tem
, 0)))))
2485 /* Line above always fails if INSN was moved by loop opt. */
2486 || (uid_luid
[regno_last_uid
[REGNO (v
->dest_reg
)]]
2487 >= INSN_LUID (loop_end
)))
2488 && ! (final_value
= v
->final_value
))
2492 /* Currently, non-reduced/final-value givs are never split. */
2493 /* Should emit insns after the loop if possible, as the biv final value
2496 /* If the final value is non-zero, and the giv has not been reduced,
2497 then must emit an instruction to set the final value. */
2498 if (final_value
&& !v
->new_reg
)
2500 /* Create a new register to hold the value of the giv, and then set
2501 the giv to its final value before the loop start. The giv is set
2502 to its final value before loop start to ensure that this insn
2503 will always be executed, no matter how we exit. */
2504 tem
= gen_reg_rtx (v
->mode
);
2505 emit_insn_before (gen_move_insn (tem
, v
->dest_reg
), loop_start
);
2506 emit_insn_before (gen_move_insn (v
->dest_reg
, final_value
),
2509 if (loop_dump_stream
)
2510 fprintf (loop_dump_stream
, "Giv %d mapped to %d for split.\n",
2511 REGNO (v
->dest_reg
), REGNO (tem
));
2517 /* This giv is splittable. If completely unrolling the loop, save the
2518 giv's initial value. Otherwise, save the constant zero for it. */
2520 if (unroll_type
== UNROLL_COMPLETELY
)
2522 /* It is not safe to use bl->initial_value here, because it may not
2523 be invariant. It is safe to use the initial value stored in
2524 the splittable_regs array if it is set. In rare cases, it won't
2525 be set, so then we do exactly the same thing as
2526 find_splittable_regs does to get a safe value. */
2527 rtx biv_initial_value
;
2529 if (splittable_regs
[bl
->regno
])
2530 biv_initial_value
= splittable_regs
[bl
->regno
];
2531 else if (GET_CODE (bl
->initial_value
) != REG
2532 || (REGNO (bl
->initial_value
) != bl
->regno
2533 && REGNO (bl
->initial_value
) >= FIRST_PSEUDO_REGISTER
))
2534 biv_initial_value
= bl
->initial_value
;
2537 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2539 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2541 biv_initial_value
= tem
;
2543 value
= fold_rtx_mult_add (v
->mult_val
, biv_initial_value
,
2544 v
->add_val
, v
->mode
);
2551 /* If a giv was combined with another giv, then we can only split
2552 this giv if the giv it was combined with was reduced. This
2553 is because the value of v->new_reg is meaningless in this
2555 if (v
->same
&& ! v
->same
->new_reg
)
2557 if (loop_dump_stream
)
2558 fprintf (loop_dump_stream
,
2559 "giv combined with unreduced giv not split.\n");
2562 /* If the giv is an address destination, it could be something other
2563 than a simple register, these have to be treated differently. */
2564 else if (v
->giv_type
== DEST_REG
)
2566 /* If value is not a constant, register, or register plus
2567 constant, then compute its value into a register before
2568 loop start. This prevents illegal rtx sharing, and should
2569 generate better code. We can use bl->initial_value here
2570 instead of splittable_regs[bl->regno] because this code
2571 is going before the loop start. */
2572 if (unroll_type
== UNROLL_COMPLETELY
2573 && GET_CODE (value
) != CONST_INT
2574 && GET_CODE (value
) != REG
2575 && (GET_CODE (value
) != PLUS
2576 || GET_CODE (XEXP (value
, 0)) != REG
2577 || GET_CODE (XEXP (value
, 1)) != CONST_INT
))
2579 rtx tem
= gen_reg_rtx (v
->mode
);
2580 emit_iv_add_mult (bl
->initial_value
, v
->mult_val
,
2581 v
->add_val
, tem
, loop_start
);
2585 splittable_regs
[REGNO (v
->new_reg
)] = value
;
2589 /* Splitting address givs is useful since it will often allow us
2590 to eliminate some increment insns for the base giv as
2593 /* If the addr giv is combined with a dest_reg giv, then all
2594 references to that dest reg will be remapped, which is NOT
2595 what we want for split addr regs. We always create a new
2596 register for the split addr giv, just to be safe. */
2598 /* ??? If there are multiple address givs which have been
2599 combined with the same dest_reg giv, then we may only need
2600 one new register for them. Pulling out constants below will
2601 catch some of the common cases of this. Currently, I leave
2602 the work of simplifying multiple address givs to the
2603 following cse pass. */
2605 /* As a special case, if we have multiple identical address givs
2606 within a single instruction, then we do use a single psuedo
2607 reg for both. This is necessary in case one is a match_dup
2610 v
->const_adjust
= 0;
2612 if (v
->same
&& v
->same
->insn
== v
->insn
2613 && v
->new_reg
== v
->same
->new_reg
)
2615 v
->dest_reg
= v
->same
->dest_reg
;
2616 if (loop_dump_stream
)
2617 fprintf (loop_dump_stream
,
2618 "Sharing address givs with reg %d\n",
2619 REGNO (v
->dest_reg
));
2621 else if (unroll_type
!= UNROLL_COMPLETELY
)
2623 /* If not completely unrolling the loop, then create a new
2624 register to hold the split value of the DEST_ADDR giv.
2625 Emit insn to initialize its value before loop start. */
2626 tem
= gen_reg_rtx (v
->mode
);
2628 /* If the address giv has a constant in its new_reg value,
2629 then this constant can be pulled out and put in value,
2630 instead of being part of the initialization code. */
2632 if (GET_CODE (v
->new_reg
) == PLUS
2633 && GET_CODE (XEXP (v
->new_reg
, 1)) == CONST_INT
)
2636 = plus_constant (tem
, INTVAL (XEXP (v
->new_reg
,1)));
2638 /* Only succeed if this will give valid addresses.
2639 Try to validate both the first and the last
2640 address resulting from loop unrolling, if
2641 one fails, then can't do const elim here. */
2642 if (memory_address_p (v
->mem_mode
, v
->dest_reg
)
2643 && memory_address_p (v
->mem_mode
,
2644 plus_constant (v
->dest_reg
,
2646 * (unroll_number
- 1))))
2648 /* Save the negative of the eliminated const, so
2649 that we can calculate the dest_reg's increment
2651 v
->const_adjust
= - INTVAL (XEXP (v
->new_reg
, 1));
2653 v
->new_reg
= XEXP (v
->new_reg
, 0);
2654 if (loop_dump_stream
)
2655 fprintf (loop_dump_stream
,
2656 "Eliminating constant from giv %d\n",
2665 /* If the address hasn't been checked for validity yet, do so
2666 now, and fail completely if either the first or the last
2667 unrolled copy of the address is not a valid address. */
2668 if (v
->dest_reg
== tem
2669 && (! memory_address_p (v
->mem_mode
, v
->dest_reg
)
2670 || ! memory_address_p (v
->mem_mode
,
2671 plus_constant (v
->dest_reg
,
2673 * (unroll_number
-1)))))
2675 if (loop_dump_stream
)
2676 fprintf (loop_dump_stream
,
2677 "Illegal address for giv at insn %d\n",
2678 INSN_UID (v
->insn
));
2682 /* To initialize the new register, just move the value of
2683 new_reg into it. This is not guaranteed to give a valid
2684 instruction on machines with complex addressing modes.
2685 If we can't recognize it, then delete it and emit insns
2686 to calculate the value from scratch. */
2687 emit_insn_before (gen_rtx (SET
, VOIDmode
, tem
,
2688 copy_rtx (v
->new_reg
)),
2690 if (recog_memoized (PREV_INSN (loop_start
)) < 0)
2694 /* We can't use bl->initial_value to compute the initial
2695 value, because the loop may have been preconditioned.
2696 We must calculate it from NEW_REG. Try using
2697 force_operand instead of emit_iv_add_mult. */
2698 delete_insn (PREV_INSN (loop_start
));
2701 ret
= force_operand (v
->new_reg
, tem
);
2703 emit_move_insn (tem
, ret
);
2704 sequence
= gen_sequence ();
2706 emit_insn_before (sequence
, loop_start
);
2708 if (loop_dump_stream
)
2709 fprintf (loop_dump_stream
,
2710 "Illegal init insn, rewritten.\n");
2715 v
->dest_reg
= value
;
2717 /* Check the resulting address for validity, and fail
2718 if the resulting address would be illegal. */
2719 if (! memory_address_p (v
->mem_mode
, v
->dest_reg
)
2720 || ! memory_address_p (v
->mem_mode
,
2721 plus_constant (v
->dest_reg
,
2723 (unroll_number
-1))))
2725 if (loop_dump_stream
)
2726 fprintf (loop_dump_stream
,
2727 "Illegal address for giv at insn %d\n",
2728 INSN_UID (v
->insn
));
2733 /* Store the value of dest_reg into the insn. This sharing
2734 will not be a problem as this insn will always be copied
2737 *v
->location
= v
->dest_reg
;
2739 /* If this address giv is combined with a dest reg giv, then
2740 save the base giv's induction pointer so that we will be
2741 able to handle this address giv properly. The base giv
2742 itself does not have to be splittable. */
2744 if (v
->same
&& v
->same
->giv_type
== DEST_REG
)
2745 addr_combined_regs
[REGNO (v
->same
->new_reg
)] = v
->same
;
2747 if (GET_CODE (v
->new_reg
) == REG
)
2749 /* This giv maybe hasn't been combined with any others.
2750 Make sure that it's giv is marked as splittable here. */
2752 splittable_regs
[REGNO (v
->new_reg
)] = value
;
2754 /* Make it appear to depend upon itself, so that the
2755 giv will be properly split in the main loop above. */
2759 addr_combined_regs
[REGNO (v
->new_reg
)] = v
;
2763 if (loop_dump_stream
)
2764 fprintf (loop_dump_stream
, "DEST_ADDR giv being split.\n");
2770 /* Currently, unreduced giv's can't be split. This is not too much
2771 of a problem since unreduced giv's are not live across loop
2772 iterations anyways. When unrolling a loop completely though,
2773 it makes sense to reduce&split givs when possible, as this will
2774 result in simpler instructions, and will not require that a reg
2775 be live across loop iterations. */
2777 splittable_regs
[REGNO (v
->dest_reg
)] = value
;
2778 fprintf (stderr
, "Giv %d at insn %d not reduced\n",
2779 REGNO (v
->dest_reg
), INSN_UID (v
->insn
));
2785 /* Givs are only updated once by definition. Mark it so if this is
2786 a splittable register. Don't need to do anything for address givs
2787 where this may not be a register. */
2789 if (GET_CODE (v
->new_reg
) == REG
)
2790 splittable_regs_updates
[REGNO (v
->new_reg
)] = 1;
2794 if (loop_dump_stream
)
2798 if (GET_CODE (v
->dest_reg
) == CONST_INT
)
2800 else if (GET_CODE (v
->dest_reg
) != REG
)
2801 regnum
= REGNO (XEXP (v
->dest_reg
, 0));
2803 regnum
= REGNO (v
->dest_reg
);
2804 fprintf (loop_dump_stream
, "Giv %d at insn %d safe to split.\n",
2805 regnum
, INSN_UID (v
->insn
));
2812 /* Try to prove that the register is dead after the loop exits. Trace every
2813 loop exit looking for an insn that will always be executed, which sets
2814 the register to some value, and appears before the first use of the register
2815 is found. If successful, then return 1, otherwise return 0. */
2817 /* ?? Could be made more intelligent in the handling of jumps, so that
2818 it can search past if statements and other similar structures. */
2821 reg_dead_after_loop (reg
, loop_start
, loop_end
)
2822 rtx reg
, loop_start
, loop_end
;
2828 /* HACK: Must also search the loop fall through exit, create a label_ref
2829 here which points to the loop_end, and append the loop_number_exit_labels
2831 label
= gen_rtx (LABEL_REF
, VOIDmode
, loop_end
);
2832 LABEL_NEXTREF (label
)
2833 = loop_number_exit_labels
[uid_loop_num
[INSN_UID (loop_start
)]];
2835 for ( ; label
; label
= LABEL_NEXTREF (label
))
2837 /* Succeed if find an insn which sets the biv or if reach end of
2838 function. Fail if find an insn that uses the biv, or if come to
2839 a conditional jump. */
2841 insn
= NEXT_INSN (XEXP (label
, 0));
2844 code
= GET_CODE (insn
);
2845 if (GET_RTX_CLASS (code
) == 'i')
2849 if (reg_referenced_p (reg
, PATTERN (insn
)))
2852 set
= single_set (insn
);
2853 if (set
&& rtx_equal_p (SET_DEST (set
), reg
))
2857 if (code
== JUMP_INSN
)
2859 if (GET_CODE (PATTERN (insn
)) == RETURN
)
2861 else if (! simplejump_p (insn
)
2862 /* Prevent infinite loop following infinite loops. */
2863 || jump_count
++ > 20)
2866 insn
= JUMP_LABEL (insn
);
2869 insn
= NEXT_INSN (insn
);
2873 /* Success, the register is dead on all loop exits. */
2877 /* Try to calculate the final value of the biv, the value it will have at
2878 the end of the loop. If we can do it, return that value. */
2881 final_biv_value (bl
, loop_start
, loop_end
)
2882 struct iv_class
*bl
;
2883 rtx loop_start
, loop_end
;
2887 /* ??? This only works for MODE_INT biv's. Reject all others for now. */
2889 if (GET_MODE_CLASS (bl
->biv
->mode
) != MODE_INT
)
2892 /* The final value for reversed bivs must be calculated differently than
2893 for ordinary bivs. In this case, there is already an insn after the
2894 loop which sets this biv's final value (if necessary), and there are
2895 no other loop exits, so we can return any value. */
2898 if (loop_dump_stream
)
2899 fprintf (loop_dump_stream
,
2900 "Final biv value for %d, reversed biv.\n", bl
->regno
);
2905 /* Try to calculate the final value as initial value + (number of iterations
2906 * increment). For this to work, increment must be invariant, the only
2907 exit from the loop must be the fall through at the bottom (otherwise
2908 it may not have its final value when the loop exits), and the initial
2909 value of the biv must be invariant. */
2911 if (loop_n_iterations
!= 0
2912 && ! loop_number_exit_labels
[uid_loop_num
[INSN_UID (loop_start
)]]
2913 && invariant_p (bl
->initial_value
))
2915 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2917 if (increment
&& invariant_p (increment
))
2919 /* Can calculate the loop exit value, emit insns after loop
2920 end to calculate this value into a temporary register in
2921 case it is needed later. */
2923 tem
= gen_reg_rtx (bl
->biv
->mode
);
2924 /* Make sure loop_end is not the last insn. */
2925 if (NEXT_INSN (loop_end
) == 0)
2926 emit_note_after (NOTE_INSN_DELETED
, loop_end
);
2927 emit_iv_add_mult (increment
, GEN_INT (loop_n_iterations
),
2928 bl
->initial_value
, tem
, NEXT_INSN (loop_end
));
2930 if (loop_dump_stream
)
2931 fprintf (loop_dump_stream
,
2932 "Final biv value for %d, calculated.\n", bl
->regno
);
2938 /* Check to see if the biv is dead at all loop exits. */
2939 if (reg_dead_after_loop (bl
->biv
->src_reg
, loop_start
, loop_end
))
2941 if (loop_dump_stream
)
2942 fprintf (loop_dump_stream
,
2943 "Final biv value for %d, biv dead after loop exit.\n",
2952 /* Try to calculate the final value of the giv, the value it will have at
2953 the end of the loop. If we can do it, return that value. */
2956 final_giv_value (v
, loop_start
, loop_end
)
2957 struct induction
*v
;
2958 rtx loop_start
, loop_end
;
2960 struct iv_class
*bl
;
2963 rtx insert_before
, seq
;
2965 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
2967 /* The final value for givs which depend on reversed bivs must be calculated
2968 differently than for ordinary givs. In this case, there is already an
2969 insn after the loop which sets this giv's final value (if necessary),
2970 and there are no other loop exits, so we can return any value. */
2973 if (loop_dump_stream
)
2974 fprintf (loop_dump_stream
,
2975 "Final giv value for %d, depends on reversed biv\n",
2976 REGNO (v
->dest_reg
));
2980 /* Try to calculate the final value as a function of the biv it depends
2981 upon. The only exit from the loop must be the fall through at the bottom
2982 (otherwise it may not have its final value when the loop exits). */
2984 /* ??? Can calculate the final giv value by subtracting off the
2985 extra biv increments times the giv's mult_val. The loop must have
2986 only one exit for this to work, but the loop iterations does not need
2989 if (loop_n_iterations
!= 0
2990 && ! loop_number_exit_labels
[uid_loop_num
[INSN_UID (loop_start
)]])
2992 /* ?? It is tempting to use the biv's value here since these insns will
2993 be put after the loop, and hence the biv will have its final value
2994 then. However, this fails if the biv is subsequently eliminated.
2995 Perhaps determine whether biv's are eliminable before trying to
2996 determine whether giv's are replaceable so that we can use the
2997 biv value here if it is not eliminable. */
2999 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
3001 if (increment
&& invariant_p (increment
))
3003 /* Can calculate the loop exit value of its biv as
3004 (loop_n_iterations * increment) + initial_value */
3006 /* The loop exit value of the giv is then
3007 (final_biv_value - extra increments) * mult_val + add_val.
3008 The extra increments are any increments to the biv which
3009 occur in the loop after the giv's value is calculated.
3010 We must search from the insn that sets the giv to the end
3011 of the loop to calculate this value. */
3013 insert_before
= NEXT_INSN (loop_end
);
3015 /* Put the final biv value in tem. */
3016 tem
= gen_reg_rtx (bl
->biv
->mode
);
3017 emit_iv_add_mult (increment
, GEN_INT (loop_n_iterations
),
3018 bl
->initial_value
, tem
, insert_before
);
3020 /* Subtract off extra increments as we find them. */
3021 for (insn
= NEXT_INSN (v
->insn
); insn
!= loop_end
;
3022 insn
= NEXT_INSN (insn
))
3024 struct induction
*biv
;
3026 for (biv
= bl
->biv
; biv
; biv
= biv
->next_iv
)
3027 if (biv
->insn
== insn
)
3030 tem
= expand_binop (GET_MODE (tem
), sub_optab
, tem
,
3031 biv
->add_val
, NULL_RTX
, 0,
3033 seq
= gen_sequence ();
3035 emit_insn_before (seq
, insert_before
);
3039 /* Now calculate the giv's final value. */
3040 emit_iv_add_mult (tem
, v
->mult_val
, v
->add_val
, tem
,
3043 if (loop_dump_stream
)
3044 fprintf (loop_dump_stream
,
3045 "Final giv value for %d, calc from biv's value.\n",
3046 REGNO (v
->dest_reg
));
3052 /* Replaceable giv's should never reach here. */
3056 /* Check to see if the biv is dead at all loop exits. */
3057 if (reg_dead_after_loop (v
->dest_reg
, loop_start
, loop_end
))
3059 if (loop_dump_stream
)
3060 fprintf (loop_dump_stream
,
3061 "Final giv value for %d, giv dead after loop exit.\n",
3062 REGNO (v
->dest_reg
));
3071 /* Calculate the number of loop iterations. Returns the exact number of loop
3072 iterations if it can be calculated, otherwise returns zero. */
3074 unsigned HOST_WIDE_INT
3075 loop_iterations (loop_start
, loop_end
)
3076 rtx loop_start
, loop_end
;
3078 rtx comparison
, comparison_value
;
3079 rtx iteration_var
, initial_value
, increment
, final_value
;
3080 enum rtx_code comparison_code
;
3083 int unsigned_compare
, compare_dir
, final_larger
;
3084 unsigned long tempu
;
3087 /* First find the iteration variable. If the last insn is a conditional
3088 branch, and the insn before tests a register value, make that the
3089 iteration variable. */
3091 loop_initial_value
= 0;
3093 loop_final_value
= 0;
3094 loop_iteration_var
= 0;
3096 /* We used to use pren_nonnote_insn here, but that fails because it might
3097 accidentally get the branch for a contained loop if the branch for this
3098 loop was deleted. We can only trust branches immediately before the
3100 last_loop_insn
= PREV_INSN (loop_end
);
3102 comparison
= get_condition_for_loop (last_loop_insn
);
3103 if (comparison
== 0)
3105 if (loop_dump_stream
)
3106 fprintf (loop_dump_stream
,
3107 "Loop unrolling: No final conditional branch found.\n");
3111 /* ??? Get_condition may switch position of induction variable and
3112 invariant register when it canonicalizes the comparison. */
3114 comparison_code
= GET_CODE (comparison
);
3115 iteration_var
= XEXP (comparison
, 0);
3116 comparison_value
= XEXP (comparison
, 1);
3118 if (GET_CODE (iteration_var
) != REG
)
3120 if (loop_dump_stream
)
3121 fprintf (loop_dump_stream
,
3122 "Loop unrolling: Comparison not against register.\n");
3126 /* Loop iterations is always called before any new registers are created
3127 now, so this should never occur. */
3129 if (REGNO (iteration_var
) >= max_reg_before_loop
)
3132 iteration_info (iteration_var
, &initial_value
, &increment
,
3133 loop_start
, loop_end
);
3134 if (initial_value
== 0)
3135 /* iteration_info already printed a message. */
3140 if (loop_dump_stream
)
3141 fprintf (loop_dump_stream
,
3142 "Loop unrolling: Increment value can't be calculated.\n");
3145 if (GET_CODE (increment
) != CONST_INT
)
3147 if (loop_dump_stream
)
3148 fprintf (loop_dump_stream
,
3149 "Loop unrolling: Increment value not constant.\n");
3152 if (GET_CODE (initial_value
) != CONST_INT
)
3154 if (loop_dump_stream
)
3155 fprintf (loop_dump_stream
,
3156 "Loop unrolling: Initial value not constant.\n");
3160 /* If the comparison value is an invariant register, then try to find
3161 its value from the insns before the start of the loop. */
3163 if (GET_CODE (comparison_value
) == REG
&& invariant_p (comparison_value
))
3167 for (insn
= PREV_INSN (loop_start
); insn
; insn
= PREV_INSN (insn
))
3169 if (GET_CODE (insn
) == CODE_LABEL
)
3172 else if (GET_RTX_CLASS (GET_CODE (insn
)) == 'i'
3173 && reg_set_p (comparison_value
, insn
))
3175 /* We found the last insn before the loop that sets the register.
3176 If it sets the entire register, and has a REG_EQUAL note,
3177 then use the value of the REG_EQUAL note. */
3178 if ((set
= single_set (insn
))
3179 && (SET_DEST (set
) == comparison_value
))
3181 rtx note
= find_reg_note (insn
, REG_EQUAL
, NULL_RTX
);
3183 /* Only use the REG_EQUAL note if it is a constant.
3184 Other things, divide in particular, will cause
3185 problems later if we use them. */
3186 if (note
&& GET_CODE (XEXP (note
, 0)) != EXPR_LIST
3187 && CONSTANT_P (XEXP (note
, 0)))
3188 comparison_value
= XEXP (note
, 0);
3195 final_value
= approx_final_value (comparison_code
, comparison_value
,
3196 &unsigned_compare
, &compare_dir
);
3198 /* Save the calculated values describing this loop's bounds, in case
3199 precondition_loop_p will need them later. These values can not be
3200 recalculated inside precondition_loop_p because strength reduction
3201 optimizations may obscure the loop's structure. */
3203 loop_iteration_var
= iteration_var
;
3204 loop_initial_value
= initial_value
;
3205 loop_increment
= increment
;
3206 loop_final_value
= final_value
;
3208 if (final_value
== 0)
3210 if (loop_dump_stream
)
3211 fprintf (loop_dump_stream
,
3212 "Loop unrolling: EQ comparison loop.\n");
3215 else if (GET_CODE (final_value
) != CONST_INT
)
3217 if (loop_dump_stream
)
3218 fprintf (loop_dump_stream
,
3219 "Loop unrolling: Final value not constant.\n");
3223 /* ?? Final value and initial value do not have to be constants.
3224 Only their difference has to be constant. When the iteration variable
3225 is an array address, the final value and initial value might both
3226 be addresses with the same base but different constant offsets.
3227 Final value must be invariant for this to work.
3229 To do this, need some way to find the values of registers which are
3232 /* Final_larger is 1 if final larger, 0 if they are equal, otherwise -1. */
3233 if (unsigned_compare
)
3235 = ((unsigned HOST_WIDE_INT
) INTVAL (final_value
)
3236 > (unsigned HOST_WIDE_INT
) INTVAL (initial_value
))
3237 - ((unsigned HOST_WIDE_INT
) INTVAL (final_value
)
3238 < (unsigned HOST_WIDE_INT
) INTVAL (initial_value
));
3240 final_larger
= (INTVAL (final_value
) > INTVAL (initial_value
))
3241 - (INTVAL (final_value
) < INTVAL (initial_value
));
3243 if (INTVAL (increment
) > 0)
3245 else if (INTVAL (increment
) == 0)
3250 /* There are 27 different cases: compare_dir = -1, 0, 1;
3251 final_larger = -1, 0, 1; increment_dir = -1, 0, 1.
3252 There are 4 normal cases, 4 reverse cases (where the iteration variable
3253 will overflow before the loop exits), 4 infinite loop cases, and 15
3254 immediate exit (0 or 1 iteration depending on loop type) cases.
3255 Only try to optimize the normal cases. */
3257 /* (compare_dir/final_larger/increment_dir)
3258 Normal cases: (0/-1/-1), (0/1/1), (-1/-1/-1), (1/1/1)
3259 Reverse cases: (0/-1/1), (0/1/-1), (-1/-1/1), (1/1/-1)
3260 Infinite loops: (0/-1/0), (0/1/0), (-1/-1/0), (1/1/0)
3261 Immediate exit: (0/0/X), (-1/0/X), (-1/1/X), (1/0/X), (1/-1/X) */
3263 /* ?? If the meaning of reverse loops (where the iteration variable
3264 will overflow before the loop exits) is undefined, then could
3265 eliminate all of these special checks, and just always assume
3266 the loops are normal/immediate/infinite. Note that this means
3267 the sign of increment_dir does not have to be known. Also,
3268 since it does not really hurt if immediate exit loops or infinite loops
3269 are optimized, then that case could be ignored also, and hence all
3270 loops can be optimized.
3272 According to ANSI Spec, the reverse loop case result is undefined,
3273 because the action on overflow is undefined.
3275 See also the special test for NE loops below. */
3277 if (final_larger
== increment_dir
&& final_larger
!= 0
3278 && (final_larger
== compare_dir
|| compare_dir
== 0))
3283 if (loop_dump_stream
)
3284 fprintf (loop_dump_stream
,
3285 "Loop unrolling: Not normal loop.\n");
3289 /* Calculate the number of iterations, final_value is only an approximation,
3290 so correct for that. Note that tempu and loop_n_iterations are
3291 unsigned, because they can be as large as 2^n - 1. */
3293 i
= INTVAL (increment
);
3295 tempu
= INTVAL (final_value
) - INTVAL (initial_value
);
3298 tempu
= INTVAL (initial_value
) - INTVAL (final_value
);
3304 /* For NE tests, make sure that the iteration variable won't miss the
3305 final value. If tempu mod i is not zero, then the iteration variable
3306 will overflow before the loop exits, and we can not calculate the
3307 number of iterations. */
3308 if (compare_dir
== 0 && (tempu
% i
) != 0)
3311 return tempu
/ i
+ ((tempu
% i
) != 0);
3314 /* Replace uses of split bivs with their split psuedo register. This is
3315 for original instructions which remain after loop unrolling without
3319 remap_split_bivs (x
)
3322 register enum rtx_code code
;
3329 code
= GET_CODE (x
);
3344 /* If non-reduced/final-value givs were split, then this would also
3345 have to remap those givs also. */
3347 if (REGNO (x
) < max_reg_before_loop
3348 && reg_iv_type
[REGNO (x
)] == BASIC_INDUCT
)
3349 return reg_biv_class
[REGNO (x
)]->biv
->src_reg
;
3352 fmt
= GET_RTX_FORMAT (code
);
3353 for (i
= GET_RTX_LENGTH (code
) - 1; i
>= 0; i
--)
3356 XEXP (x
, i
) = remap_split_bivs (XEXP (x
, i
));
3360 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
3361 XVECEXP (x
, i
, j
) = remap_split_bivs (XVECEXP (x
, i
, j
));