1 /* Try to unroll loops, and split induction variables.
2 Copyright (C) 1992, 1993, 1994, 1995 Free Software Foundation, Inc.
3 Contributed by James E. Wilson, Cygnus Support/UC Berkeley.
5 This file is part of GNU CC.
7 GNU CC is free software; you can redistribute it and/or modify
8 it under the terms of the GNU General Public License as published by
9 the Free Software Foundation; either version 2, or (at your option)
12 GNU CC is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License for more details.
17 You should have received a copy of the GNU General Public License
18 along with GNU CC; see the file COPYING. If not, write to
19 the Free Software Foundation, 59 Temple Place - Suite 330,
20 Boston, MA 02111-1307, USA. */
22 /* Try to unroll a loop, and split induction variables.
24 Loops for which the number of iterations can be calculated exactly are
25 handled specially. If the number of iterations times the insn_count is
26 less than MAX_UNROLLED_INSNS, then the loop is unrolled completely.
27 Otherwise, we try to unroll the loop a number of times modulo the number
28 of iterations, so that only one exit test will be needed. It is unrolled
29 a number of times approximately equal to MAX_UNROLLED_INSNS divided by
32 Otherwise, if the number of iterations can be calculated exactly at
33 run time, and the loop is always entered at the top, then we try to
34 precondition the loop. That is, at run time, calculate how many times
35 the loop will execute, and then execute the loop body a few times so
36 that the remaining iterations will be some multiple of 4 (or 2 if the
37 loop is large). Then fall through to a loop unrolled 4 (or 2) times,
38 with only one exit test needed at the end of the loop.
40 Otherwise, if the number of iterations can not be calculated exactly,
41 not even at run time, then we still unroll the loop a number of times
42 approximately equal to MAX_UNROLLED_INSNS divided by the insn count,
43 but there must be an exit test after each copy of the loop body.
45 For each induction variable, which is dead outside the loop (replaceable)
46 or for which we can easily calculate the final value, if we can easily
47 calculate its value at each place where it is set as a function of the
48 current loop unroll count and the variable's value at loop entry, then
49 the induction variable is split into `N' different variables, one for
50 each copy of the loop body. One variable is live across the backward
51 branch, and the others are all calculated as a function of this variable.
52 This helps eliminate data dependencies, and leads to further opportunities
55 /* Possible improvements follow: */
57 /* ??? Add an extra pass somewhere to determine whether unrolling will
58 give any benefit. E.g. after generating all unrolled insns, compute the
59 cost of all insns and compare against cost of insns in rolled loop.
61 - On traditional architectures, unrolling a non-constant bound loop
62 is a win if there is a giv whose only use is in memory addresses, the
63 memory addresses can be split, and hence giv increments can be
65 - It is also a win if the loop is executed many times, and preconditioning
66 can be performed for the loop.
67 Add code to check for these and similar cases. */
69 /* ??? Improve control of which loops get unrolled. Could use profiling
70 info to only unroll the most commonly executed loops. Perhaps have
71 a user specifyable option to control the amount of code expansion,
72 or the percent of loops to consider for unrolling. Etc. */
74 /* ??? Look at the register copies inside the loop to see if they form a
75 simple permutation. If so, iterate the permutation until it gets back to
76 the start state. This is how many times we should unroll the loop, for
77 best results, because then all register copies can be eliminated.
78 For example, the lisp nreverse function should be unrolled 3 times
87 ??? The number of times to unroll the loop may also be based on data
88 references in the loop. For example, if we have a loop that references
89 x[i-1], x[i], and x[i+1], we should unroll it a multiple of 3 times. */
91 /* ??? Add some simple linear equation solving capability so that we can
92 determine the number of loop iterations for more complex loops.
93 For example, consider this loop from gdb
94 #define SWAP_TARGET_AND_HOST(buffer,len)
97 char *p = (char *) buffer;
98 char *q = ((char *) buffer) + len - 1;
99 int iterations = (len + 1) >> 1;
101 for (p; p < q; p++, q--;)
109 start value = p = &buffer + current_iteration
110 end value = q = &buffer + len - 1 - current_iteration
111 Given the loop exit test of "p < q", then there must be "q - p" iterations,
112 set equal to zero and solve for number of iterations:
113 q - p = len - 1 - 2*current_iteration = 0
114 current_iteration = (len - 1) / 2
115 Hence, there are (len - 1) / 2 (rounded up to the nearest integer)
116 iterations of this loop. */
118 /* ??? Currently, no labels are marked as loop invariant when doing loop
119 unrolling. This is because an insn inside the loop, that loads the address
120 of a label inside the loop into a register, could be moved outside the loop
121 by the invariant code motion pass if labels were invariant. If the loop
122 is subsequently unrolled, the code will be wrong because each unrolled
123 body of the loop will use the same address, whereas each actually needs a
124 different address. A case where this happens is when a loop containing
125 a switch statement is unrolled.
127 It would be better to let labels be considered invariant. When we
128 unroll loops here, check to see if any insns using a label local to the
129 loop were moved before the loop. If so, then correct the problem, by
130 moving the insn back into the loop, or perhaps replicate the insn before
131 the loop, one copy for each time the loop is unrolled. */
133 /* The prime factors looked for when trying to unroll a loop by some
134 number which is modulo the total number of iterations. Just checking
135 for these 4 prime factors will find at least one factor for 75% of
136 all numbers theoretically. Practically speaking, this will succeed
137 almost all of the time since loops are generally a multiple of 2
140 #define NUM_FACTORS 4
142 struct _factor
{ int factor
, count
; } factors
[NUM_FACTORS
]
143 = { {2, 0}, {3, 0}, {5, 0}, {7, 0}};
145 /* Describes the different types of loop unrolling performed. */
147 enum unroll_types
{ UNROLL_COMPLETELY
, UNROLL_MODULO
, UNROLL_NAIVE
};
151 #include "insn-config.h"
152 #include "integrate.h"
159 /* This controls which loops are unrolled, and by how much we unroll
162 #ifndef MAX_UNROLLED_INSNS
163 #define MAX_UNROLLED_INSNS 100
166 /* Indexed by register number, if non-zero, then it contains a pointer
167 to a struct induction for a DEST_REG giv which has been combined with
168 one of more address givs. This is needed because whenever such a DEST_REG
169 giv is modified, we must modify the value of all split address givs
170 that were combined with this DEST_REG giv. */
172 static struct induction
**addr_combined_regs
;
174 /* Indexed by register number, if this is a splittable induction variable,
175 then this will hold the current value of the register, which depends on the
178 static rtx
*splittable_regs
;
180 /* Indexed by register number, if this is a splittable induction variable,
181 then this will hold the number of instructions in the loop that modify
182 the induction variable. Used to ensure that only the last insn modifying
183 a split iv will update the original iv of the dest. */
185 static int *splittable_regs_updates
;
187 /* Values describing the current loop's iteration variable. These are set up
188 by loop_iterations, and used by precondition_loop_p. */
190 static rtx loop_iteration_var
;
191 static rtx loop_initial_value
;
192 static rtx loop_increment
;
193 static rtx loop_final_value
;
194 static enum rtx_code loop_comparison_code
;
196 /* Forward declarations. */
198 static void init_reg_map
PROTO((struct inline_remap
*, int));
199 static int precondition_loop_p
PROTO((rtx
*, rtx
*, rtx
*, rtx
, rtx
));
200 static rtx calculate_giv_inc
PROTO((rtx
, rtx
, int));
201 static rtx initial_reg_note_copy
PROTO((rtx
, struct inline_remap
*));
202 static void final_reg_note_copy
PROTO((rtx
, struct inline_remap
*));
203 static void copy_loop_body
PROTO((rtx
, rtx
, struct inline_remap
*, rtx
, int,
204 enum unroll_types
, rtx
, rtx
, rtx
, rtx
));
205 static void iteration_info
PROTO((rtx
, rtx
*, rtx
*, rtx
, rtx
));
206 static rtx approx_final_value
PROTO((enum rtx_code
, rtx
, int *, int *));
207 static int find_splittable_regs
PROTO((enum unroll_types
, rtx
, rtx
, rtx
, int));
208 static int find_splittable_givs
PROTO((struct iv_class
*,enum unroll_types
,
209 rtx
, rtx
, rtx
, int));
210 static int reg_dead_after_loop
PROTO((rtx
, rtx
, rtx
));
211 static rtx fold_rtx_mult_add
PROTO((rtx
, rtx
, rtx
, enum machine_mode
));
212 static rtx remap_split_bivs
PROTO((rtx
));
214 /* Try to unroll one loop and split induction variables in the loop.
216 The loop is described by the arguments LOOP_END, INSN_COUNT, and
217 LOOP_START. END_INSERT_BEFORE indicates where insns should be added
218 which need to be executed when the loop falls through. STRENGTH_REDUCTION_P
219 indicates whether information generated in the strength reduction pass
222 This function is intended to be called from within `strength_reduce'
226 unroll_loop (loop_end
, insn_count
, loop_start
, end_insert_before
,
231 rtx end_insert_before
;
232 int strength_reduce_p
;
235 int unroll_number
= 1;
236 rtx copy_start
, copy_end
;
237 rtx insn
, copy
, sequence
, pattern
, tem
;
238 int max_labelno
, max_insnno
;
240 struct inline_remap
*map
;
248 int splitting_not_safe
= 0;
249 enum unroll_types unroll_type
;
250 int loop_preconditioned
= 0;
252 /* This points to the last real insn in the loop, which should be either
253 a JUMP_INSN (for conditional jumps) or a BARRIER (for unconditional
257 /* Don't bother unrolling huge loops. Since the minimum factor is
258 two, loops greater than one half of MAX_UNROLLED_INSNS will never
260 if (insn_count
> MAX_UNROLLED_INSNS
/ 2)
262 if (loop_dump_stream
)
263 fprintf (loop_dump_stream
, "Unrolling failure: Loop too big.\n");
267 /* When emitting debugger info, we can't unroll loops with unequal numbers
268 of block_beg and block_end notes, because that would unbalance the block
269 structure of the function. This can happen as a result of the
270 "if (foo) bar; else break;" optimization in jump.c. */
271 /* ??? Gcc has a general policy that -g is never supposed to change the code
272 that the compiler emits, so we must disable this optimization always,
273 even if debug info is not being output. This is rare, so this should
274 not be a significant performance problem. */
276 if (1 /* write_symbols != NO_DEBUG */)
278 int block_begins
= 0;
281 for (insn
= loop_start
; insn
!= loop_end
; insn
= NEXT_INSN (insn
))
283 if (GET_CODE (insn
) == NOTE
)
285 if (NOTE_LINE_NUMBER (insn
) == NOTE_INSN_BLOCK_BEG
)
287 else if (NOTE_LINE_NUMBER (insn
) == NOTE_INSN_BLOCK_END
)
292 if (block_begins
!= block_ends
)
294 if (loop_dump_stream
)
295 fprintf (loop_dump_stream
,
296 "Unrolling failure: Unbalanced block notes.\n");
301 /* Determine type of unroll to perform. Depends on the number of iterations
302 and the size of the loop. */
304 /* If there is no strength reduce info, then set loop_n_iterations to zero.
305 This can happen if strength_reduce can't find any bivs in the loop.
306 A value of zero indicates that the number of iterations could not be
309 if (! strength_reduce_p
)
310 loop_n_iterations
= 0;
312 if (loop_dump_stream
&& loop_n_iterations
> 0)
313 fprintf (loop_dump_stream
,
314 "Loop unrolling: %d iterations.\n", loop_n_iterations
);
316 /* Find and save a pointer to the last nonnote insn in the loop. */
318 last_loop_insn
= prev_nonnote_insn (loop_end
);
320 /* Calculate how many times to unroll the loop. Indicate whether or
321 not the loop is being completely unrolled. */
323 if (loop_n_iterations
== 1)
325 /* If number of iterations is exactly 1, then eliminate the compare and
326 branch at the end of the loop since they will never be taken.
327 Then return, since no other action is needed here. */
329 /* If the last instruction is not a BARRIER or a JUMP_INSN, then
330 don't do anything. */
332 if (GET_CODE (last_loop_insn
) == BARRIER
)
334 /* Delete the jump insn. This will delete the barrier also. */
335 delete_insn (PREV_INSN (last_loop_insn
));
337 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
340 /* The immediately preceding insn is a compare which must be
342 delete_insn (last_loop_insn
);
343 delete_insn (PREV_INSN (last_loop_insn
));
345 /* The immediately preceding insn may not be the compare, so don't
347 delete_insn (last_loop_insn
);
352 else if (loop_n_iterations
> 0
353 && loop_n_iterations
* insn_count
< MAX_UNROLLED_INSNS
)
355 unroll_number
= loop_n_iterations
;
356 unroll_type
= UNROLL_COMPLETELY
;
358 else if (loop_n_iterations
> 0)
360 /* Try to factor the number of iterations. Don't bother with the
361 general case, only using 2, 3, 5, and 7 will get 75% of all
362 numbers theoretically, and almost all in practice. */
364 for (i
= 0; i
< NUM_FACTORS
; i
++)
365 factors
[i
].count
= 0;
367 temp
= loop_n_iterations
;
368 for (i
= NUM_FACTORS
- 1; i
>= 0; i
--)
369 while (temp
% factors
[i
].factor
== 0)
372 temp
= temp
/ factors
[i
].factor
;
375 /* Start with the larger factors first so that we generally
376 get lots of unrolling. */
380 for (i
= 3; i
>= 0; i
--)
381 while (factors
[i
].count
--)
383 if (temp
* factors
[i
].factor
< MAX_UNROLLED_INSNS
)
385 unroll_number
*= factors
[i
].factor
;
386 temp
*= factors
[i
].factor
;
392 /* If we couldn't find any factors, then unroll as in the normal
394 if (unroll_number
== 1)
396 if (loop_dump_stream
)
397 fprintf (loop_dump_stream
,
398 "Loop unrolling: No factors found.\n");
401 unroll_type
= UNROLL_MODULO
;
405 /* Default case, calculate number of times to unroll loop based on its
407 if (unroll_number
== 1)
409 if (8 * insn_count
< MAX_UNROLLED_INSNS
)
411 else if (4 * insn_count
< MAX_UNROLLED_INSNS
)
416 unroll_type
= UNROLL_NAIVE
;
419 /* Now we know how many times to unroll the loop. */
421 if (loop_dump_stream
)
422 fprintf (loop_dump_stream
,
423 "Unrolling loop %d times.\n", unroll_number
);
426 if (unroll_type
== UNROLL_COMPLETELY
|| unroll_type
== UNROLL_MODULO
)
428 /* Loops of these types should never start with a jump down to
429 the exit condition test. For now, check for this case just to
430 be sure. UNROLL_NAIVE loops can be of this form, this case is
433 while (GET_CODE (insn
) != CODE_LABEL
&& GET_CODE (insn
) != JUMP_INSN
)
434 insn
= NEXT_INSN (insn
);
435 if (GET_CODE (insn
) == JUMP_INSN
)
439 if (unroll_type
== UNROLL_COMPLETELY
)
441 /* Completely unrolling the loop: Delete the compare and branch at
442 the end (the last two instructions). This delete must done at the
443 very end of loop unrolling, to avoid problems with calls to
444 back_branch_in_range_p, which is called by find_splittable_regs.
445 All increments of splittable bivs/givs are changed to load constant
448 copy_start
= loop_start
;
450 /* Set insert_before to the instruction immediately after the JUMP_INSN
451 (or BARRIER), so that any NOTEs between the JUMP_INSN and the end of
452 the loop will be correctly handled by copy_loop_body. */
453 insert_before
= NEXT_INSN (last_loop_insn
);
455 /* Set copy_end to the insn before the jump at the end of the loop. */
456 if (GET_CODE (last_loop_insn
) == BARRIER
)
457 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
458 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
461 /* The instruction immediately before the JUMP_INSN is a compare
462 instruction which we do not want to copy. */
463 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
465 /* The instruction immediately before the JUMP_INSN may not be the
466 compare, so we must copy it. */
467 copy_end
= PREV_INSN (last_loop_insn
);
472 /* We currently can't unroll a loop if it doesn't end with a
473 JUMP_INSN. There would need to be a mechanism that recognizes
474 this case, and then inserts a jump after each loop body, which
475 jumps to after the last loop body. */
476 if (loop_dump_stream
)
477 fprintf (loop_dump_stream
,
478 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
482 else if (unroll_type
== UNROLL_MODULO
)
484 /* Partially unrolling the loop: The compare and branch at the end
485 (the last two instructions) must remain. Don't copy the compare
486 and branch instructions at the end of the loop. Insert the unrolled
487 code immediately before the compare/branch at the end so that the
488 code will fall through to them as before. */
490 copy_start
= loop_start
;
492 /* Set insert_before to the jump insn at the end of the loop.
493 Set copy_end to before the jump insn at the end of the loop. */
494 if (GET_CODE (last_loop_insn
) == BARRIER
)
496 insert_before
= PREV_INSN (last_loop_insn
);
497 copy_end
= PREV_INSN (insert_before
);
499 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
502 /* The instruction immediately before the JUMP_INSN is a compare
503 instruction which we do not want to copy or delete. */
504 insert_before
= PREV_INSN (last_loop_insn
);
505 copy_end
= PREV_INSN (insert_before
);
507 /* The instruction immediately before the JUMP_INSN may not be the
508 compare, so we must copy it. */
509 insert_before
= last_loop_insn
;
510 copy_end
= PREV_INSN (last_loop_insn
);
515 /* We currently can't unroll a loop if it doesn't end with a
516 JUMP_INSN. There would need to be a mechanism that recognizes
517 this case, and then inserts a jump after each loop body, which
518 jumps to after the last loop body. */
519 if (loop_dump_stream
)
520 fprintf (loop_dump_stream
,
521 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
527 /* Normal case: Must copy the compare and branch instructions at the
530 if (GET_CODE (last_loop_insn
) == BARRIER
)
532 /* Loop ends with an unconditional jump and a barrier.
533 Handle this like above, don't copy jump and barrier.
534 This is not strictly necessary, but doing so prevents generating
535 unconditional jumps to an immediately following label.
537 This will be corrected below if the target of this jump is
538 not the start_label. */
540 insert_before
= PREV_INSN (last_loop_insn
);
541 copy_end
= PREV_INSN (insert_before
);
543 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
545 /* Set insert_before to immediately after the JUMP_INSN, so that
546 NOTEs at the end of the loop will be correctly handled by
548 insert_before
= NEXT_INSN (last_loop_insn
);
549 copy_end
= last_loop_insn
;
553 /* We currently can't unroll a loop if it doesn't end with a
554 JUMP_INSN. There would need to be a mechanism that recognizes
555 this case, and then inserts a jump after each loop body, which
556 jumps to after the last loop body. */
557 if (loop_dump_stream
)
558 fprintf (loop_dump_stream
,
559 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
563 /* If copying exit test branches because they can not be eliminated,
564 then must convert the fall through case of the branch to a jump past
565 the end of the loop. Create a label to emit after the loop and save
566 it for later use. Do not use the label after the loop, if any, since
567 it might be used by insns outside the loop, or there might be insns
568 added before it later by final_[bg]iv_value which must be after
569 the real exit label. */
570 exit_label
= gen_label_rtx ();
573 while (GET_CODE (insn
) != CODE_LABEL
&& GET_CODE (insn
) != JUMP_INSN
)
574 insn
= NEXT_INSN (insn
);
576 if (GET_CODE (insn
) == JUMP_INSN
)
578 /* The loop starts with a jump down to the exit condition test.
579 Start copying the loop after the barrier following this
581 copy_start
= NEXT_INSN (insn
);
583 /* Splitting induction variables doesn't work when the loop is
584 entered via a jump to the bottom, because then we end up doing
585 a comparison against a new register for a split variable, but
586 we did not execute the set insn for the new register because
587 it was skipped over. */
588 splitting_not_safe
= 1;
589 if (loop_dump_stream
)
590 fprintf (loop_dump_stream
,
591 "Splitting not safe, because loop not entered at top.\n");
594 copy_start
= loop_start
;
597 /* This should always be the first label in the loop. */
598 start_label
= NEXT_INSN (copy_start
);
599 /* There may be a line number note and/or a loop continue note here. */
600 while (GET_CODE (start_label
) == NOTE
)
601 start_label
= NEXT_INSN (start_label
);
602 if (GET_CODE (start_label
) != CODE_LABEL
)
604 /* This can happen as a result of jump threading. If the first insns in
605 the loop test the same condition as the loop's backward jump, or the
606 opposite condition, then the backward jump will be modified to point
607 to elsewhere, and the loop's start label is deleted.
609 This case currently can not be handled by the loop unrolling code. */
611 if (loop_dump_stream
)
612 fprintf (loop_dump_stream
,
613 "Unrolling failure: unknown insns between BEG note and loop label.\n");
616 if (LABEL_NAME (start_label
))
618 /* The jump optimization pass must have combined the original start label
619 with a named label for a goto. We can't unroll this case because
620 jumps which go to the named label must be handled differently than
621 jumps to the loop start, and it is impossible to differentiate them
623 if (loop_dump_stream
)
624 fprintf (loop_dump_stream
,
625 "Unrolling failure: loop start label is gone\n");
629 if (unroll_type
== UNROLL_NAIVE
630 && GET_CODE (last_loop_insn
) == BARRIER
631 && start_label
!= JUMP_LABEL (PREV_INSN (last_loop_insn
)))
633 /* In this case, we must copy the jump and barrier, because they will
634 not be converted to jumps to an immediately following label. */
636 insert_before
= NEXT_INSN (last_loop_insn
);
637 copy_end
= last_loop_insn
;
640 /* Allocate a translation table for the labels and insn numbers.
641 They will be filled in as we copy the insns in the loop. */
643 max_labelno
= max_label_num ();
644 max_insnno
= get_max_uid ();
646 map
= (struct inline_remap
*) alloca (sizeof (struct inline_remap
));
648 map
->integrating
= 0;
650 /* Allocate the label map. */
654 map
->label_map
= (rtx
*) alloca (max_labelno
* sizeof (rtx
));
656 local_label
= (char *) alloca (max_labelno
);
657 bzero (local_label
, max_labelno
);
662 /* Search the loop and mark all local labels, i.e. the ones which have to
663 be distinct labels when copied. For all labels which might be
664 non-local, set their label_map entries to point to themselves.
665 If they happen to be local their label_map entries will be overwritten
666 before the loop body is copied. The label_map entries for local labels
667 will be set to a different value each time the loop body is copied. */
669 for (insn
= copy_start
; insn
!= loop_end
; insn
= NEXT_INSN (insn
))
671 if (GET_CODE (insn
) == CODE_LABEL
)
672 local_label
[CODE_LABEL_NUMBER (insn
)] = 1;
673 else if (GET_CODE (insn
) == JUMP_INSN
)
675 if (JUMP_LABEL (insn
))
676 map
->label_map
[CODE_LABEL_NUMBER (JUMP_LABEL (insn
))]
678 else if (GET_CODE (PATTERN (insn
)) == ADDR_VEC
679 || GET_CODE (PATTERN (insn
)) == ADDR_DIFF_VEC
)
681 rtx pat
= PATTERN (insn
);
682 int diff_vec_p
= GET_CODE (PATTERN (insn
)) == ADDR_DIFF_VEC
;
683 int len
= XVECLEN (pat
, diff_vec_p
);
686 for (i
= 0; i
< len
; i
++)
688 label
= XEXP (XVECEXP (pat
, diff_vec_p
, i
), 0);
689 map
->label_map
[CODE_LABEL_NUMBER (label
)] = label
;
695 /* Allocate space for the insn map. */
697 map
->insn_map
= (rtx
*) alloca (max_insnno
* sizeof (rtx
));
699 /* Set this to zero, to indicate that we are doing loop unrolling,
700 not function inlining. */
701 map
->inline_target
= 0;
703 /* The register and constant maps depend on the number of registers
704 present, so the final maps can't be created until after
705 find_splittable_regs is called. However, they are needed for
706 preconditioning, so we create temporary maps when preconditioning
709 /* The preconditioning code may allocate two new pseudo registers. */
710 maxregnum
= max_reg_num ();
712 /* Allocate and zero out the splittable_regs and addr_combined_regs
713 arrays. These must be zeroed here because they will be used if
714 loop preconditioning is performed, and must be zero for that case.
716 It is safe to do this here, since the extra registers created by the
717 preconditioning code and find_splittable_regs will never be used
718 to access the splittable_regs[] and addr_combined_regs[] arrays. */
720 splittable_regs
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
721 bzero ((char *) splittable_regs
, maxregnum
* sizeof (rtx
));
722 splittable_regs_updates
= (int *) alloca (maxregnum
* sizeof (int));
723 bzero ((char *) splittable_regs_updates
, maxregnum
* sizeof (int));
725 = (struct induction
**) alloca (maxregnum
* sizeof (struct induction
*));
726 bzero ((char *) addr_combined_regs
, maxregnum
* sizeof (struct induction
*));
727 /* We must limit it to max_reg_before_loop, because only these pseudo
728 registers have valid regno_first_uid info. Any register created after
729 that is unlikely to be local to the loop anyways. */
730 local_regno
= (char *) alloca (max_reg_before_loop
);
731 bzero (local_regno
, max_reg_before_loop
);
733 /* Mark all local registers, i.e. the ones which are referenced only
735 if (INSN_UID (copy_end
) < max_uid_for_loop
)
737 int copy_start_luid
= INSN_LUID (copy_start
);
738 int copy_end_luid
= INSN_LUID (copy_end
);
740 /* If a register is used in the jump insn, we must not duplicate it
741 since it will also be used outside the loop. */
742 if (GET_CODE (copy_end
) == JUMP_INSN
)
744 /* If copy_start points to the NOTE that starts the loop, then we must
745 use the next luid, because invariant pseudo-regs moved out of the loop
746 have their lifetimes modified to start here, but they are not safe
748 if (copy_start
== loop_start
)
751 /* If a pseudo's lifetime is entirely contained within this loop, then we
752 can use a different pseudo in each unrolled copy of the loop. This
753 results in better code. */
754 for (j
= FIRST_PSEUDO_REGISTER
; j
< max_reg_before_loop
; ++j
)
755 if (regno_first_uid
[j
] > 0 && regno_first_uid
[j
] <= max_uid_for_loop
756 && uid_luid
[regno_first_uid
[j
]] >= copy_start_luid
757 && regno_last_uid
[j
] > 0 && regno_last_uid
[j
] <= max_uid_for_loop
758 && uid_luid
[regno_last_uid
[j
]] <= copy_end_luid
)
760 /* However, we must also check for loop-carried dependencies.
761 If the value the pseudo has at the end of iteration X is
762 used by iteration X+1, then we can not use a different pseudo
763 for each unrolled copy of the loop. */
764 /* A pseudo is safe if regno_first_uid is a set, and this
765 set dominates all instructions from regno_first_uid to
767 /* ??? This check is simplistic. We would get better code if
768 this check was more sophisticated. */
769 if (set_dominates_use (j
, regno_first_uid
[j
], regno_last_uid
[j
],
770 copy_start
, copy_end
))
773 if (loop_dump_stream
)
776 fprintf (loop_dump_stream
, "Marked reg %d as local\n", j
);
778 fprintf (loop_dump_stream
, "Did not mark reg %d as local\n",
784 /* If this loop requires exit tests when unrolled, check to see if we
785 can precondition the loop so as to make the exit tests unnecessary.
786 Just like variable splitting, this is not safe if the loop is entered
787 via a jump to the bottom. Also, can not do this if no strength
788 reduce info, because precondition_loop_p uses this info. */
790 /* Must copy the loop body for preconditioning before the following
791 find_splittable_regs call since that will emit insns which need to
792 be after the preconditioned loop copies, but immediately before the
793 unrolled loop copies. */
795 /* Also, it is not safe to split induction variables for the preconditioned
796 copies of the loop body. If we split induction variables, then the code
797 assumes that each induction variable can be represented as a function
798 of its initial value and the loop iteration number. This is not true
799 in this case, because the last preconditioned copy of the loop body
800 could be any iteration from the first up to the `unroll_number-1'th,
801 depending on the initial value of the iteration variable. Therefore
802 we can not split induction variables here, because we can not calculate
803 their value. Hence, this code must occur before find_splittable_regs
806 if (unroll_type
== UNROLL_NAIVE
&& ! splitting_not_safe
&& strength_reduce_p
)
808 rtx initial_value
, final_value
, increment
;
810 if (precondition_loop_p (&initial_value
, &final_value
, &increment
,
811 loop_start
, loop_end
))
813 register rtx diff
, temp
;
814 enum machine_mode mode
;
816 int abs_inc
, neg_inc
;
818 map
->reg_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
820 map
->const_equiv_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
821 map
->const_age_map
= (unsigned *) alloca (maxregnum
822 * sizeof (unsigned));
823 map
->const_equiv_map_size
= maxregnum
;
824 global_const_equiv_map
= map
->const_equiv_map
;
825 global_const_equiv_map_size
= maxregnum
;
827 init_reg_map (map
, maxregnum
);
829 /* Limit loop unrolling to 4, since this will make 7 copies of
831 if (unroll_number
> 4)
834 /* Save the absolute value of the increment, and also whether or
835 not it is negative. */
837 abs_inc
= INTVAL (increment
);
846 /* Decide what mode to do these calculations in. Choose the larger
847 of final_value's mode and initial_value's mode, or a full-word if
848 both are constants. */
849 mode
= GET_MODE (final_value
);
850 if (mode
== VOIDmode
)
852 mode
= GET_MODE (initial_value
);
853 if (mode
== VOIDmode
)
856 else if (mode
!= GET_MODE (initial_value
)
857 && (GET_MODE_SIZE (mode
)
858 < GET_MODE_SIZE (GET_MODE (initial_value
))))
859 mode
= GET_MODE (initial_value
);
861 /* Calculate the difference between the final and initial values.
862 Final value may be a (plus (reg x) (const_int 1)) rtx.
863 Let the following cse pass simplify this if initial value is
866 We must copy the final and initial values here to avoid
867 improperly shared rtl. */
869 diff
= expand_binop (mode
, sub_optab
, copy_rtx (final_value
),
870 copy_rtx (initial_value
), NULL_RTX
, 0,
873 /* Now calculate (diff % (unroll * abs (increment))) by using an
875 diff
= expand_binop (GET_MODE (diff
), and_optab
, diff
,
876 GEN_INT (unroll_number
* abs_inc
- 1),
877 NULL_RTX
, 0, OPTAB_LIB_WIDEN
);
879 /* Now emit a sequence of branches to jump to the proper precond
882 labels
= (rtx
*) alloca (sizeof (rtx
) * unroll_number
);
883 for (i
= 0; i
< unroll_number
; i
++)
884 labels
[i
] = gen_label_rtx ();
886 /* Check for the case where the initial value is greater than or
887 equal to the final value. In that case, we want to execute
888 exactly one loop iteration. The code below will fail for this
889 case. This check does not apply if the loop has a NE
890 comparison at the end. */
892 if (loop_comparison_code
!= NE
)
894 emit_cmp_insn (initial_value
, final_value
, neg_inc
? LE
: GE
,
895 NULL_RTX
, mode
, 0, 0);
897 emit_jump_insn (gen_ble (labels
[1]));
899 emit_jump_insn (gen_bge (labels
[1]));
900 JUMP_LABEL (get_last_insn ()) = labels
[1];
901 LABEL_NUSES (labels
[1])++;
904 /* Assuming the unroll_number is 4, and the increment is 2, then
905 for a negative increment: for a positive increment:
906 diff = 0,1 precond 0 diff = 0,7 precond 0
907 diff = 2,3 precond 3 diff = 1,2 precond 1
908 diff = 4,5 precond 2 diff = 3,4 precond 2
909 diff = 6,7 precond 1 diff = 5,6 precond 3 */
911 /* We only need to emit (unroll_number - 1) branches here, the
912 last case just falls through to the following code. */
914 /* ??? This would give better code if we emitted a tree of branches
915 instead of the current linear list of branches. */
917 for (i
= 0; i
< unroll_number
- 1; i
++)
920 enum rtx_code cmp_code
;
922 /* For negative increments, must invert the constant compared
923 against, except when comparing against zero. */
931 cmp_const
= unroll_number
- i
;
940 emit_cmp_insn (diff
, GEN_INT (abs_inc
* cmp_const
),
941 cmp_code
, NULL_RTX
, mode
, 0, 0);
944 emit_jump_insn (gen_beq (labels
[i
]));
946 emit_jump_insn (gen_bge (labels
[i
]));
948 emit_jump_insn (gen_ble (labels
[i
]));
949 JUMP_LABEL (get_last_insn ()) = labels
[i
];
950 LABEL_NUSES (labels
[i
])++;
953 /* If the increment is greater than one, then we need another branch,
954 to handle other cases equivalent to 0. */
956 /* ??? This should be merged into the code above somehow to help
957 simplify the code here, and reduce the number of branches emitted.
958 For the negative increment case, the branch here could easily
959 be merged with the `0' case branch above. For the positive
960 increment case, it is not clear how this can be simplified. */
965 enum rtx_code cmp_code
;
969 cmp_const
= abs_inc
- 1;
974 cmp_const
= abs_inc
* (unroll_number
- 1) + 1;
978 emit_cmp_insn (diff
, GEN_INT (cmp_const
), cmp_code
, NULL_RTX
,
982 emit_jump_insn (gen_ble (labels
[0]));
984 emit_jump_insn (gen_bge (labels
[0]));
985 JUMP_LABEL (get_last_insn ()) = labels
[0];
986 LABEL_NUSES (labels
[0])++;
989 sequence
= gen_sequence ();
991 emit_insn_before (sequence
, loop_start
);
993 /* Only the last copy of the loop body here needs the exit
994 test, so set copy_end to exclude the compare/branch here,
995 and then reset it inside the loop when get to the last
998 if (GET_CODE (last_loop_insn
) == BARRIER
)
999 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
1000 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
1003 /* The immediately preceding insn is a compare which we do not
1005 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
1007 /* The immediately preceding insn may not be a compare, so we
1009 copy_end
= PREV_INSN (last_loop_insn
);
1015 for (i
= 1; i
< unroll_number
; i
++)
1017 emit_label_after (labels
[unroll_number
- i
],
1018 PREV_INSN (loop_start
));
1020 bzero ((char *) map
->insn_map
, max_insnno
* sizeof (rtx
));
1021 bzero ((char *) map
->const_equiv_map
, maxregnum
* sizeof (rtx
));
1022 bzero ((char *) map
->const_age_map
,
1023 maxregnum
* sizeof (unsigned));
1026 for (j
= 0; j
< max_labelno
; j
++)
1028 map
->label_map
[j
] = gen_label_rtx ();
1030 for (j
= FIRST_PSEUDO_REGISTER
; j
< max_reg_before_loop
; j
++)
1032 map
->reg_map
[j
] = gen_reg_rtx (GET_MODE (regno_reg_rtx
[j
]));
1034 /* The last copy needs the compare/branch insns at the end,
1035 so reset copy_end here if the loop ends with a conditional
1038 if (i
== unroll_number
- 1)
1040 if (GET_CODE (last_loop_insn
) == BARRIER
)
1041 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
1043 copy_end
= last_loop_insn
;
1046 /* None of the copies are the `last_iteration', so just
1047 pass zero for that parameter. */
1048 copy_loop_body (copy_start
, copy_end
, map
, exit_label
, 0,
1049 unroll_type
, start_label
, loop_end
,
1050 loop_start
, copy_end
);
1052 emit_label_after (labels
[0], PREV_INSN (loop_start
));
1054 if (GET_CODE (last_loop_insn
) == BARRIER
)
1056 insert_before
= PREV_INSN (last_loop_insn
);
1057 copy_end
= PREV_INSN (insert_before
);
1062 /* The immediately preceding insn is a compare which we do not
1064 insert_before
= PREV_INSN (last_loop_insn
);
1065 copy_end
= PREV_INSN (insert_before
);
1067 /* The immediately preceding insn may not be a compare, so we
1069 insert_before
= last_loop_insn
;
1070 copy_end
= PREV_INSN (last_loop_insn
);
1074 /* Set unroll type to MODULO now. */
1075 unroll_type
= UNROLL_MODULO
;
1076 loop_preconditioned
= 1;
1080 /* If reach here, and the loop type is UNROLL_NAIVE, then don't unroll
1081 the loop unless all loops are being unrolled. */
1082 if (unroll_type
== UNROLL_NAIVE
&& ! flag_unroll_all_loops
)
1084 if (loop_dump_stream
)
1085 fprintf (loop_dump_stream
, "Unrolling failure: Naive unrolling not being done.\n");
1089 /* At this point, we are guaranteed to unroll the loop. */
1091 /* For each biv and giv, determine whether it can be safely split into
1092 a different variable for each unrolled copy of the loop body.
1093 We precalculate and save this info here, since computing it is
1096 Do this before deleting any instructions from the loop, so that
1097 back_branch_in_range_p will work correctly. */
1099 if (splitting_not_safe
)
1102 temp
= find_splittable_regs (unroll_type
, loop_start
, loop_end
,
1103 end_insert_before
, unroll_number
);
1105 /* find_splittable_regs may have created some new registers, so must
1106 reallocate the reg_map with the new larger size, and must realloc
1107 the constant maps also. */
1109 maxregnum
= max_reg_num ();
1110 map
->reg_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
1112 init_reg_map (map
, maxregnum
);
1114 /* Space is needed in some of the map for new registers, so new_maxregnum
1115 is an (over)estimate of how many registers will exist at the end. */
1116 new_maxregnum
= maxregnum
+ (temp
* unroll_number
* 2);
1118 /* Must realloc space for the constant maps, because the number of registers
1119 may have changed. */
1121 map
->const_equiv_map
= (rtx
*) alloca (new_maxregnum
* sizeof (rtx
));
1122 map
->const_age_map
= (unsigned *) alloca (new_maxregnum
* sizeof (unsigned));
1124 map
->const_equiv_map_size
= new_maxregnum
;
1125 global_const_equiv_map
= map
->const_equiv_map
;
1126 global_const_equiv_map_size
= new_maxregnum
;
1128 /* Search the list of bivs and givs to find ones which need to be remapped
1129 when split, and set their reg_map entry appropriately. */
1131 for (bl
= loop_iv_list
; bl
; bl
= bl
->next
)
1133 if (REGNO (bl
->biv
->src_reg
) != bl
->regno
)
1134 map
->reg_map
[bl
->regno
] = bl
->biv
->src_reg
;
1136 /* Currently, non-reduced/final-value givs are never split. */
1137 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
1138 if (REGNO (v
->src_reg
) != bl
->regno
)
1139 map
->reg_map
[REGNO (v
->dest_reg
)] = v
->src_reg
;
1143 /* Use our current register alignment and pointer flags. */
1144 map
->regno_pointer_flag
= regno_pointer_flag
;
1145 map
->regno_pointer_align
= regno_pointer_align
;
1147 /* If the loop is being partially unrolled, and the iteration variables
1148 are being split, and are being renamed for the split, then must fix up
1149 the compare/jump instruction at the end of the loop to refer to the new
1150 registers. This compare isn't copied, so the registers used in it
1151 will never be replaced if it isn't done here. */
1153 if (unroll_type
== UNROLL_MODULO
)
1155 insn
= NEXT_INSN (copy_end
);
1156 if (GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == JUMP_INSN
)
1157 PATTERN (insn
) = remap_split_bivs (PATTERN (insn
));
1160 /* For unroll_number - 1 times, make a copy of each instruction
1161 between copy_start and copy_end, and insert these new instructions
1162 before the end of the loop. */
1164 for (i
= 0; i
< unroll_number
; i
++)
1166 bzero ((char *) map
->insn_map
, max_insnno
* sizeof (rtx
));
1167 bzero ((char *) map
->const_equiv_map
, new_maxregnum
* sizeof (rtx
));
1168 bzero ((char *) map
->const_age_map
, new_maxregnum
* sizeof (unsigned));
1171 for (j
= 0; j
< max_labelno
; j
++)
1173 map
->label_map
[j
] = gen_label_rtx ();
1175 for (j
= FIRST_PSEUDO_REGISTER
; j
< max_reg_before_loop
; j
++)
1177 map
->reg_map
[j
] = gen_reg_rtx (GET_MODE (regno_reg_rtx
[j
]));
1179 /* If loop starts with a branch to the test, then fix it so that
1180 it points to the test of the first unrolled copy of the loop. */
1181 if (i
== 0 && loop_start
!= copy_start
)
1183 insn
= PREV_INSN (copy_start
);
1184 pattern
= PATTERN (insn
);
1186 tem
= map
->label_map
[CODE_LABEL_NUMBER
1187 (XEXP (SET_SRC (pattern
), 0))];
1188 SET_SRC (pattern
) = gen_rtx (LABEL_REF
, VOIDmode
, tem
);
1190 /* Set the jump label so that it can be used by later loop unrolling
1192 JUMP_LABEL (insn
) = tem
;
1193 LABEL_NUSES (tem
)++;
1196 copy_loop_body (copy_start
, copy_end
, map
, exit_label
,
1197 i
== unroll_number
- 1, unroll_type
, start_label
,
1198 loop_end
, insert_before
, insert_before
);
1201 /* Before deleting any insns, emit a CODE_LABEL immediately after the last
1202 insn to be deleted. This prevents any runaway delete_insn call from
1203 more insns that it should, as it always stops at a CODE_LABEL. */
1205 /* Delete the compare and branch at the end of the loop if completely
1206 unrolling the loop. Deleting the backward branch at the end also
1207 deletes the code label at the start of the loop. This is done at
1208 the very end to avoid problems with back_branch_in_range_p. */
1210 if (unroll_type
== UNROLL_COMPLETELY
)
1211 safety_label
= emit_label_after (gen_label_rtx (), last_loop_insn
);
1213 safety_label
= emit_label_after (gen_label_rtx (), copy_end
);
1215 /* Delete all of the original loop instructions. Don't delete the
1216 LOOP_BEG note, or the first code label in the loop. */
1218 insn
= NEXT_INSN (copy_start
);
1219 while (insn
!= safety_label
)
1221 if (insn
!= start_label
)
1222 insn
= delete_insn (insn
);
1224 insn
= NEXT_INSN (insn
);
1227 /* Can now delete the 'safety' label emitted to protect us from runaway
1228 delete_insn calls. */
1229 if (INSN_DELETED_P (safety_label
))
1231 delete_insn (safety_label
);
1233 /* If exit_label exists, emit it after the loop. Doing the emit here
1234 forces it to have a higher INSN_UID than any insn in the unrolled loop.
1235 This is needed so that mostly_true_jump in reorg.c will treat jumps
1236 to this loop end label correctly, i.e. predict that they are usually
1239 emit_label_after (exit_label
, loop_end
);
1242 /* Return true if the loop can be safely, and profitably, preconditioned
1243 so that the unrolled copies of the loop body don't need exit tests.
1245 This only works if final_value, initial_value and increment can be
1246 determined, and if increment is a constant power of 2.
1247 If increment is not a power of 2, then the preconditioning modulo
1248 operation would require a real modulo instead of a boolean AND, and this
1249 is not considered `profitable'. */
1251 /* ??? If the loop is known to be executed very many times, or the machine
1252 has a very cheap divide instruction, then preconditioning is a win even
1253 when the increment is not a power of 2. Use RTX_COST to compute
1254 whether divide is cheap. */
1257 precondition_loop_p (initial_value
, final_value
, increment
, loop_start
,
1259 rtx
*initial_value
, *final_value
, *increment
;
1260 rtx loop_start
, loop_end
;
1263 if (loop_n_iterations
> 0)
1265 *initial_value
= const0_rtx
;
1266 *increment
= const1_rtx
;
1267 *final_value
= GEN_INT (loop_n_iterations
);
1269 if (loop_dump_stream
)
1270 fprintf (loop_dump_stream
,
1271 "Preconditioning: Success, number of iterations known, %d.\n",
1276 if (loop_initial_value
== 0)
1278 if (loop_dump_stream
)
1279 fprintf (loop_dump_stream
,
1280 "Preconditioning: Could not find initial value.\n");
1283 else if (loop_increment
== 0)
1285 if (loop_dump_stream
)
1286 fprintf (loop_dump_stream
,
1287 "Preconditioning: Could not find increment value.\n");
1290 else if (GET_CODE (loop_increment
) != CONST_INT
)
1292 if (loop_dump_stream
)
1293 fprintf (loop_dump_stream
,
1294 "Preconditioning: Increment not a constant.\n");
1297 else if ((exact_log2 (INTVAL (loop_increment
)) < 0)
1298 && (exact_log2 (- INTVAL (loop_increment
)) < 0))
1300 if (loop_dump_stream
)
1301 fprintf (loop_dump_stream
,
1302 "Preconditioning: Increment not a constant power of 2.\n");
1306 /* Unsigned_compare and compare_dir can be ignored here, since they do
1307 not matter for preconditioning. */
1309 if (loop_final_value
== 0)
1311 if (loop_dump_stream
)
1312 fprintf (loop_dump_stream
,
1313 "Preconditioning: EQ comparison loop.\n");
1317 /* Must ensure that final_value is invariant, so call invariant_p to
1318 check. Before doing so, must check regno against max_reg_before_loop
1319 to make sure that the register is in the range covered by invariant_p.
1320 If it isn't, then it is most likely a biv/giv which by definition are
1322 if ((GET_CODE (loop_final_value
) == REG
1323 && REGNO (loop_final_value
) >= max_reg_before_loop
)
1324 || (GET_CODE (loop_final_value
) == PLUS
1325 && REGNO (XEXP (loop_final_value
, 0)) >= max_reg_before_loop
)
1326 || ! invariant_p (loop_final_value
))
1328 if (loop_dump_stream
)
1329 fprintf (loop_dump_stream
,
1330 "Preconditioning: Final value not invariant.\n");
1334 /* Fail for floating point values, since the caller of this function
1335 does not have code to deal with them. */
1336 if (GET_MODE_CLASS (GET_MODE (loop_final_value
)) == MODE_FLOAT
1337 || GET_MODE_CLASS (GET_MODE (loop_initial_value
)) == MODE_FLOAT
)
1339 if (loop_dump_stream
)
1340 fprintf (loop_dump_stream
,
1341 "Preconditioning: Floating point final or initial value.\n");
1345 /* Now set initial_value to be the iteration_var, since that may be a
1346 simpler expression, and is guaranteed to be correct if all of the
1347 above tests succeed.
1349 We can not use the initial_value as calculated, because it will be
1350 one too small for loops of the form "while (i-- > 0)". We can not
1351 emit code before the loop_skip_over insns to fix this problem as this
1352 will then give a number one too large for loops of the form
1355 Note that all loops that reach here are entered at the top, because
1356 this function is not called if the loop starts with a jump. */
1358 /* Fail if loop_iteration_var is not live before loop_start, since we need
1359 to test its value in the preconditioning code. */
1361 if (uid_luid
[regno_first_uid
[REGNO (loop_iteration_var
)]]
1362 > INSN_LUID (loop_start
))
1364 if (loop_dump_stream
)
1365 fprintf (loop_dump_stream
,
1366 "Preconditioning: Iteration var not live before loop start.\n");
1370 *initial_value
= loop_iteration_var
;
1371 *increment
= loop_increment
;
1372 *final_value
= loop_final_value
;
1375 if (loop_dump_stream
)
1376 fprintf (loop_dump_stream
, "Preconditioning: Successful.\n");
1381 /* All pseudo-registers must be mapped to themselves. Two hard registers
1382 must be mapped, VIRTUAL_STACK_VARS_REGNUM and VIRTUAL_INCOMING_ARGS_
1383 REGNUM, to avoid function-inlining specific conversions of these
1384 registers. All other hard regs can not be mapped because they may be
1389 init_reg_map (map
, maxregnum
)
1390 struct inline_remap
*map
;
1395 for (i
= maxregnum
- 1; i
> LAST_VIRTUAL_REGISTER
; i
--)
1396 map
->reg_map
[i
] = regno_reg_rtx
[i
];
1397 /* Just clear the rest of the entries. */
1398 for (i
= LAST_VIRTUAL_REGISTER
; i
>= 0; i
--)
1399 map
->reg_map
[i
] = 0;
1401 map
->reg_map
[VIRTUAL_STACK_VARS_REGNUM
]
1402 = regno_reg_rtx
[VIRTUAL_STACK_VARS_REGNUM
];
1403 map
->reg_map
[VIRTUAL_INCOMING_ARGS_REGNUM
]
1404 = regno_reg_rtx
[VIRTUAL_INCOMING_ARGS_REGNUM
];
1407 /* Strength-reduction will often emit code for optimized biv/givs which
1408 calculates their value in a temporary register, and then copies the result
1409 to the iv. This procedure reconstructs the pattern computing the iv;
1410 verifying that all operands are of the proper form.
1412 The return value is the amount that the giv is incremented by. */
1415 calculate_giv_inc (pattern
, src_insn
, regno
)
1416 rtx pattern
, src_insn
;
1420 rtx increment_total
= 0;
1424 /* Verify that we have an increment insn here. First check for a plus
1425 as the set source. */
1426 if (GET_CODE (SET_SRC (pattern
)) != PLUS
)
1428 /* SR sometimes computes the new giv value in a temp, then copies it
1430 src_insn
= PREV_INSN (src_insn
);
1431 pattern
= PATTERN (src_insn
);
1432 if (GET_CODE (SET_SRC (pattern
)) != PLUS
)
1435 /* The last insn emitted is not needed, so delete it to avoid confusing
1436 the second cse pass. This insn sets the giv unnecessarily. */
1437 delete_insn (get_last_insn ());
1440 /* Verify that we have a constant as the second operand of the plus. */
1441 increment
= XEXP (SET_SRC (pattern
), 1);
1442 if (GET_CODE (increment
) != CONST_INT
)
1444 /* SR sometimes puts the constant in a register, especially if it is
1445 too big to be an add immed operand. */
1446 src_insn
= PREV_INSN (src_insn
);
1447 increment
= SET_SRC (PATTERN (src_insn
));
1449 /* SR may have used LO_SUM to compute the constant if it is too large
1450 for a load immed operand. In this case, the constant is in operand
1451 one of the LO_SUM rtx. */
1452 if (GET_CODE (increment
) == LO_SUM
)
1453 increment
= XEXP (increment
, 1);
1454 else if (GET_CODE (increment
) == IOR
1455 || GET_CODE (increment
) == ASHIFT
)
1457 /* The rs6000 port loads some constants with IOR.
1458 The alpha port loads some constants with ASHIFT. */
1459 rtx second_part
= XEXP (increment
, 1);
1460 enum rtx_code code
= GET_CODE (increment
);
1462 src_insn
= PREV_INSN (src_insn
);
1463 increment
= SET_SRC (PATTERN (src_insn
));
1464 /* Don't need the last insn anymore. */
1465 delete_insn (get_last_insn ());
1467 if (GET_CODE (second_part
) != CONST_INT
1468 || GET_CODE (increment
) != CONST_INT
)
1472 increment
= GEN_INT (INTVAL (increment
) | INTVAL (second_part
));
1474 increment
= GEN_INT (INTVAL (increment
) << INTVAL (second_part
));
1477 if (GET_CODE (increment
) != CONST_INT
)
1480 /* The insn loading the constant into a register is no longer needed,
1482 delete_insn (get_last_insn ());
1485 if (increment_total
)
1486 increment_total
= GEN_INT (INTVAL (increment_total
) + INTVAL (increment
));
1488 increment_total
= increment
;
1490 /* Check that the source register is the same as the register we expected
1491 to see as the source. If not, something is seriously wrong. */
1492 if (GET_CODE (XEXP (SET_SRC (pattern
), 0)) != REG
1493 || REGNO (XEXP (SET_SRC (pattern
), 0)) != regno
)
1495 /* Some machines (e.g. the romp), may emit two add instructions for
1496 certain constants, so lets try looking for another add immediately
1497 before this one if we have only seen one add insn so far. */
1503 src_insn
= PREV_INSN (src_insn
);
1504 pattern
= PATTERN (src_insn
);
1506 delete_insn (get_last_insn ());
1514 return increment_total
;
1517 /* Copy REG_NOTES, except for insn references, because not all insn_map
1518 entries are valid yet. We do need to copy registers now though, because
1519 the reg_map entries can change during copying. */
1522 initial_reg_note_copy (notes
, map
)
1524 struct inline_remap
*map
;
1531 copy
= rtx_alloc (GET_CODE (notes
));
1532 PUT_MODE (copy
, GET_MODE (notes
));
1534 if (GET_CODE (notes
) == EXPR_LIST
)
1535 XEXP (copy
, 0) = copy_rtx_and_substitute (XEXP (notes
, 0), map
);
1536 else if (GET_CODE (notes
) == INSN_LIST
)
1537 /* Don't substitute for these yet. */
1538 XEXP (copy
, 0) = XEXP (notes
, 0);
1542 XEXP (copy
, 1) = initial_reg_note_copy (XEXP (notes
, 1), map
);
1547 /* Fixup insn references in copied REG_NOTES. */
1550 final_reg_note_copy (notes
, map
)
1552 struct inline_remap
*map
;
1556 for (note
= notes
; note
; note
= XEXP (note
, 1))
1557 if (GET_CODE (note
) == INSN_LIST
)
1558 XEXP (note
, 0) = map
->insn_map
[INSN_UID (XEXP (note
, 0))];
1561 /* Copy each instruction in the loop, substituting from map as appropriate.
1562 This is very similar to a loop in expand_inline_function. */
1565 copy_loop_body (copy_start
, copy_end
, map
, exit_label
, last_iteration
,
1566 unroll_type
, start_label
, loop_end
, insert_before
,
1568 rtx copy_start
, copy_end
;
1569 struct inline_remap
*map
;
1572 enum unroll_types unroll_type
;
1573 rtx start_label
, loop_end
, insert_before
, copy_notes_from
;
1577 int dest_reg_was_split
, i
;
1579 rtx final_label
= 0;
1580 rtx giv_inc
, giv_dest_reg
, giv_src_reg
;
1582 /* If this isn't the last iteration, then map any references to the
1583 start_label to final_label. Final label will then be emitted immediately
1584 after the end of this loop body if it was ever used.
1586 If this is the last iteration, then map references to the start_label
1588 if (! last_iteration
)
1590 final_label
= gen_label_rtx ();
1591 map
->label_map
[CODE_LABEL_NUMBER (start_label
)] = final_label
;
1594 map
->label_map
[CODE_LABEL_NUMBER (start_label
)] = start_label
;
1601 insn
= NEXT_INSN (insn
);
1603 map
->orig_asm_operands_vector
= 0;
1605 switch (GET_CODE (insn
))
1608 pattern
= PATTERN (insn
);
1612 /* Check to see if this is a giv that has been combined with
1613 some split address givs. (Combined in the sense that
1614 `combine_givs' in loop.c has put two givs in the same register.)
1615 In this case, we must search all givs based on the same biv to
1616 find the address givs. Then split the address givs.
1617 Do this before splitting the giv, since that may map the
1618 SET_DEST to a new register. */
1620 if (GET_CODE (pattern
) == SET
1621 && GET_CODE (SET_DEST (pattern
)) == REG
1622 && addr_combined_regs
[REGNO (SET_DEST (pattern
))])
1624 struct iv_class
*bl
;
1625 struct induction
*v
, *tv
;
1626 int regno
= REGNO (SET_DEST (pattern
));
1628 v
= addr_combined_regs
[REGNO (SET_DEST (pattern
))];
1629 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
1631 /* Although the giv_inc amount is not needed here, we must call
1632 calculate_giv_inc here since it might try to delete the
1633 last insn emitted. If we wait until later to call it,
1634 we might accidentally delete insns generated immediately
1635 below by emit_unrolled_add. */
1637 giv_inc
= calculate_giv_inc (pattern
, insn
, regno
);
1639 /* Now find all address giv's that were combined with this
1641 for (tv
= bl
->giv
; tv
; tv
= tv
->next_iv
)
1642 if (tv
->giv_type
== DEST_ADDR
&& tv
->same
== v
)
1646 /* If this DEST_ADDR giv was not split, then ignore it. */
1647 if (*tv
->location
!= tv
->dest_reg
)
1650 /* Scale this_giv_inc if the multiplicative factors of
1651 the two givs are different. */
1652 this_giv_inc
= INTVAL (giv_inc
);
1653 if (tv
->mult_val
!= v
->mult_val
)
1654 this_giv_inc
= (this_giv_inc
/ INTVAL (v
->mult_val
)
1655 * INTVAL (tv
->mult_val
));
1657 tv
->dest_reg
= plus_constant (tv
->dest_reg
, this_giv_inc
);
1658 *tv
->location
= tv
->dest_reg
;
1660 if (last_iteration
&& unroll_type
!= UNROLL_COMPLETELY
)
1662 /* Must emit an insn to increment the split address
1663 giv. Add in the const_adjust field in case there
1664 was a constant eliminated from the address. */
1665 rtx value
, dest_reg
;
1667 /* tv->dest_reg will be either a bare register,
1668 or else a register plus a constant. */
1669 if (GET_CODE (tv
->dest_reg
) == REG
)
1670 dest_reg
= tv
->dest_reg
;
1672 dest_reg
= XEXP (tv
->dest_reg
, 0);
1674 /* Check for shared address givs, and avoid
1675 incrementing the shared pseudo reg more than
1677 if (! tv
->same_insn
)
1679 /* tv->dest_reg may actually be a (PLUS (REG)
1680 (CONST)) here, so we must call plus_constant
1681 to add the const_adjust amount before calling
1682 emit_unrolled_add below. */
1683 value
= plus_constant (tv
->dest_reg
,
1686 /* The constant could be too large for an add
1687 immediate, so can't directly emit an insn
1689 emit_unrolled_add (dest_reg
, XEXP (value
, 0),
1693 /* Reset the giv to be just the register again, in case
1694 it is used after the set we have just emitted.
1695 We must subtract the const_adjust factor added in
1697 tv
->dest_reg
= plus_constant (dest_reg
,
1698 - tv
->const_adjust
);
1699 *tv
->location
= tv
->dest_reg
;
1704 /* If this is a setting of a splittable variable, then determine
1705 how to split the variable, create a new set based on this split,
1706 and set up the reg_map so that later uses of the variable will
1707 use the new split variable. */
1709 dest_reg_was_split
= 0;
1711 if (GET_CODE (pattern
) == SET
1712 && GET_CODE (SET_DEST (pattern
)) == REG
1713 && splittable_regs
[REGNO (SET_DEST (pattern
))])
1715 int regno
= REGNO (SET_DEST (pattern
));
1717 dest_reg_was_split
= 1;
1719 /* Compute the increment value for the giv, if it wasn't
1720 already computed above. */
1723 giv_inc
= calculate_giv_inc (pattern
, insn
, regno
);
1724 giv_dest_reg
= SET_DEST (pattern
);
1725 giv_src_reg
= SET_DEST (pattern
);
1727 if (unroll_type
== UNROLL_COMPLETELY
)
1729 /* Completely unrolling the loop. Set the induction
1730 variable to a known constant value. */
1732 /* The value in splittable_regs may be an invariant
1733 value, so we must use plus_constant here. */
1734 splittable_regs
[regno
]
1735 = plus_constant (splittable_regs
[regno
], INTVAL (giv_inc
));
1737 if (GET_CODE (splittable_regs
[regno
]) == PLUS
)
1739 giv_src_reg
= XEXP (splittable_regs
[regno
], 0);
1740 giv_inc
= XEXP (splittable_regs
[regno
], 1);
1744 /* The splittable_regs value must be a REG or a
1745 CONST_INT, so put the entire value in the giv_src_reg
1747 giv_src_reg
= splittable_regs
[regno
];
1748 giv_inc
= const0_rtx
;
1753 /* Partially unrolling loop. Create a new pseudo
1754 register for the iteration variable, and set it to
1755 be a constant plus the original register. Except
1756 on the last iteration, when the result has to
1757 go back into the original iteration var register. */
1759 /* Handle bivs which must be mapped to a new register
1760 when split. This happens for bivs which need their
1761 final value set before loop entry. The new register
1762 for the biv was stored in the biv's first struct
1763 induction entry by find_splittable_regs. */
1765 if (regno
< max_reg_before_loop
1766 && reg_iv_type
[regno
] == BASIC_INDUCT
)
1768 giv_src_reg
= reg_biv_class
[regno
]->biv
->src_reg
;
1769 giv_dest_reg
= giv_src_reg
;
1773 /* If non-reduced/final-value givs were split, then
1774 this would have to remap those givs also. See
1775 find_splittable_regs. */
1778 splittable_regs
[regno
]
1779 = GEN_INT (INTVAL (giv_inc
)
1780 + INTVAL (splittable_regs
[regno
]));
1781 giv_inc
= splittable_regs
[regno
];
1783 /* Now split the induction variable by changing the dest
1784 of this insn to a new register, and setting its
1785 reg_map entry to point to this new register.
1787 If this is the last iteration, and this is the last insn
1788 that will update the iv, then reuse the original dest,
1789 to ensure that the iv will have the proper value when
1790 the loop exits or repeats.
1792 Using splittable_regs_updates here like this is safe,
1793 because it can only be greater than one if all
1794 instructions modifying the iv are always executed in
1797 if (! last_iteration
1798 || (splittable_regs_updates
[regno
]-- != 1))
1800 tem
= gen_reg_rtx (GET_MODE (giv_src_reg
));
1802 map
->reg_map
[regno
] = tem
;
1805 map
->reg_map
[regno
] = giv_src_reg
;
1808 /* The constant being added could be too large for an add
1809 immediate, so can't directly emit an insn here. */
1810 emit_unrolled_add (giv_dest_reg
, giv_src_reg
, giv_inc
);
1811 copy
= get_last_insn ();
1812 pattern
= PATTERN (copy
);
1816 pattern
= copy_rtx_and_substitute (pattern
, map
);
1817 copy
= emit_insn (pattern
);
1819 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
1822 /* If this insn is setting CC0, it may need to look at
1823 the insn that uses CC0 to see what type of insn it is.
1824 In that case, the call to recog via validate_change will
1825 fail. So don't substitute constants here. Instead,
1826 do it when we emit the following insn.
1828 For example, see the pyr.md file. That machine has signed and
1829 unsigned compares. The compare patterns must check the
1830 following branch insn to see which what kind of compare to
1833 If the previous insn set CC0, substitute constants on it as
1835 if (sets_cc0_p (PATTERN (copy
)) != 0)
1840 try_constants (cc0_insn
, map
);
1842 try_constants (copy
, map
);
1845 try_constants (copy
, map
);
1848 /* Make split induction variable constants `permanent' since we
1849 know there are no backward branches across iteration variable
1850 settings which would invalidate this. */
1851 if (dest_reg_was_split
)
1853 int regno
= REGNO (SET_DEST (pattern
));
1855 if (regno
< map
->const_equiv_map_size
1856 && map
->const_age_map
[regno
] == map
->const_age
)
1857 map
->const_age_map
[regno
] = -1;
1862 pattern
= copy_rtx_and_substitute (PATTERN (insn
), map
);
1863 copy
= emit_jump_insn (pattern
);
1864 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
1866 if (JUMP_LABEL (insn
) == start_label
&& insn
== copy_end
1867 && ! last_iteration
)
1869 /* This is a branch to the beginning of the loop; this is the
1870 last insn being copied; and this is not the last iteration.
1871 In this case, we want to change the original fall through
1872 case to be a branch past the end of the loop, and the
1873 original jump label case to fall_through. */
1875 if (invert_exp (pattern
, copy
))
1877 if (! redirect_exp (&pattern
,
1878 map
->label_map
[CODE_LABEL_NUMBER
1879 (JUMP_LABEL (insn
))],
1886 rtx lab
= gen_label_rtx ();
1887 /* Can't do it by reversing the jump (probably because we
1888 couldn't reverse the conditions), so emit a new
1889 jump_insn after COPY, and redirect the jump around
1891 jmp
= emit_jump_insn_after (gen_jump (exit_label
), copy
);
1892 jmp
= emit_barrier_after (jmp
);
1893 emit_label_after (lab
, jmp
);
1894 LABEL_NUSES (lab
) = 0;
1895 if (! redirect_exp (&pattern
,
1896 map
->label_map
[CODE_LABEL_NUMBER
1897 (JUMP_LABEL (insn
))],
1905 try_constants (cc0_insn
, map
);
1908 try_constants (copy
, map
);
1910 /* Set the jump label of COPY correctly to avoid problems with
1911 later passes of unroll_loop, if INSN had jump label set. */
1912 if (JUMP_LABEL (insn
))
1916 /* Can't use the label_map for every insn, since this may be
1917 the backward branch, and hence the label was not mapped. */
1918 if (GET_CODE (pattern
) == SET
)
1920 tem
= SET_SRC (pattern
);
1921 if (GET_CODE (tem
) == LABEL_REF
)
1922 label
= XEXP (tem
, 0);
1923 else if (GET_CODE (tem
) == IF_THEN_ELSE
)
1925 if (XEXP (tem
, 1) != pc_rtx
)
1926 label
= XEXP (XEXP (tem
, 1), 0);
1928 label
= XEXP (XEXP (tem
, 2), 0);
1932 if (label
&& GET_CODE (label
) == CODE_LABEL
)
1933 JUMP_LABEL (copy
) = label
;
1936 /* An unrecognizable jump insn, probably the entry jump
1937 for a switch statement. This label must have been mapped,
1938 so just use the label_map to get the new jump label. */
1940 = map
->label_map
[CODE_LABEL_NUMBER (JUMP_LABEL (insn
))];
1943 /* If this is a non-local jump, then must increase the label
1944 use count so that the label will not be deleted when the
1945 original jump is deleted. */
1946 LABEL_NUSES (JUMP_LABEL (copy
))++;
1948 else if (GET_CODE (PATTERN (copy
)) == ADDR_VEC
1949 || GET_CODE (PATTERN (copy
)) == ADDR_DIFF_VEC
)
1951 rtx pat
= PATTERN (copy
);
1952 int diff_vec_p
= GET_CODE (pat
) == ADDR_DIFF_VEC
;
1953 int len
= XVECLEN (pat
, diff_vec_p
);
1956 for (i
= 0; i
< len
; i
++)
1957 LABEL_NUSES (XEXP (XVECEXP (pat
, diff_vec_p
, i
), 0))++;
1960 /* If this used to be a conditional jump insn but whose branch
1961 direction is now known, we must do something special. */
1962 if (condjump_p (insn
) && !simplejump_p (insn
) && map
->last_pc_value
)
1965 /* The previous insn set cc0 for us. So delete it. */
1966 delete_insn (PREV_INSN (copy
));
1969 /* If this is now a no-op, delete it. */
1970 if (map
->last_pc_value
== pc_rtx
)
1972 /* Don't let delete_insn delete the label referenced here,
1973 because we might possibly need it later for some other
1974 instruction in the loop. */
1975 if (JUMP_LABEL (copy
))
1976 LABEL_NUSES (JUMP_LABEL (copy
))++;
1978 if (JUMP_LABEL (copy
))
1979 LABEL_NUSES (JUMP_LABEL (copy
))--;
1983 /* Otherwise, this is unconditional jump so we must put a
1984 BARRIER after it. We could do some dead code elimination
1985 here, but jump.c will do it just as well. */
1991 pattern
= copy_rtx_and_substitute (PATTERN (insn
), map
);
1992 copy
= emit_call_insn (pattern
);
1993 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
1995 /* Because the USAGE information potentially contains objects other
1996 than hard registers, we need to copy it. */
1997 CALL_INSN_FUNCTION_USAGE (copy
) =
1998 copy_rtx_and_substitute (CALL_INSN_FUNCTION_USAGE (insn
), map
);
2002 try_constants (cc0_insn
, map
);
2005 try_constants (copy
, map
);
2007 /* Be lazy and assume CALL_INSNs clobber all hard registers. */
2008 for (i
= 0; i
< FIRST_PSEUDO_REGISTER
; i
++)
2009 map
->const_equiv_map
[i
] = 0;
2013 /* If this is the loop start label, then we don't need to emit a
2014 copy of this label since no one will use it. */
2016 if (insn
!= start_label
)
2018 copy
= emit_label (map
->label_map
[CODE_LABEL_NUMBER (insn
)]);
2024 copy
= emit_barrier ();
2028 /* VTOP notes are valid only before the loop exit test. If placed
2029 anywhere else, loop may generate bad code. */
2031 if (NOTE_LINE_NUMBER (insn
) != NOTE_INSN_DELETED
2032 && (NOTE_LINE_NUMBER (insn
) != NOTE_INSN_LOOP_VTOP
2033 || (last_iteration
&& unroll_type
!= UNROLL_COMPLETELY
)))
2034 copy
= emit_note (NOTE_SOURCE_FILE (insn
),
2035 NOTE_LINE_NUMBER (insn
));
2045 map
->insn_map
[INSN_UID (insn
)] = copy
;
2047 while (insn
!= copy_end
);
2049 /* Now finish coping the REG_NOTES. */
2053 insn
= NEXT_INSN (insn
);
2054 if ((GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == JUMP_INSN
2055 || GET_CODE (insn
) == CALL_INSN
)
2056 && map
->insn_map
[INSN_UID (insn
)])
2057 final_reg_note_copy (REG_NOTES (map
->insn_map
[INSN_UID (insn
)]), map
);
2059 while (insn
!= copy_end
);
2061 /* There may be notes between copy_notes_from and loop_end. Emit a copy of
2062 each of these notes here, since there may be some important ones, such as
2063 NOTE_INSN_BLOCK_END notes, in this group. We don't do this on the last
2064 iteration, because the original notes won't be deleted.
2066 We can't use insert_before here, because when from preconditioning,
2067 insert_before points before the loop. We can't use copy_end, because
2068 there may be insns already inserted after it (which we don't want to
2069 copy) when not from preconditioning code. */
2071 if (! last_iteration
)
2073 for (insn
= copy_notes_from
; insn
!= loop_end
; insn
= NEXT_INSN (insn
))
2075 if (GET_CODE (insn
) == NOTE
2076 && NOTE_LINE_NUMBER (insn
) != NOTE_INSN_DELETED
)
2077 emit_note (NOTE_SOURCE_FILE (insn
), NOTE_LINE_NUMBER (insn
));
2081 if (final_label
&& LABEL_NUSES (final_label
) > 0)
2082 emit_label (final_label
);
2084 tem
= gen_sequence ();
2086 emit_insn_before (tem
, insert_before
);
2089 /* Emit an insn, using the expand_binop to ensure that a valid insn is
2090 emitted. This will correctly handle the case where the increment value
2091 won't fit in the immediate field of a PLUS insns. */
2094 emit_unrolled_add (dest_reg
, src_reg
, increment
)
2095 rtx dest_reg
, src_reg
, increment
;
2099 result
= expand_binop (GET_MODE (dest_reg
), add_optab
, src_reg
, increment
,
2100 dest_reg
, 0, OPTAB_LIB_WIDEN
);
2102 if (dest_reg
!= result
)
2103 emit_move_insn (dest_reg
, result
);
2106 /* Searches the insns between INSN and LOOP_END. Returns 1 if there
2107 is a backward branch in that range that branches to somewhere between
2108 LOOP_START and INSN. Returns 0 otherwise. */
2110 /* ??? This is quadratic algorithm. Could be rewritten to be linear.
2111 In practice, this is not a problem, because this function is seldom called,
2112 and uses a negligible amount of CPU time on average. */
2115 back_branch_in_range_p (insn
, loop_start
, loop_end
)
2117 rtx loop_start
, loop_end
;
2119 rtx p
, q
, target_insn
;
2121 /* Stop before we get to the backward branch at the end of the loop. */
2122 loop_end
= prev_nonnote_insn (loop_end
);
2123 if (GET_CODE (loop_end
) == BARRIER
)
2124 loop_end
= PREV_INSN (loop_end
);
2126 /* Check in case insn has been deleted, search forward for first non
2127 deleted insn following it. */
2128 while (INSN_DELETED_P (insn
))
2129 insn
= NEXT_INSN (insn
);
2131 /* Check for the case where insn is the last insn in the loop. */
2132 if (insn
== loop_end
)
2135 for (p
= NEXT_INSN (insn
); p
!= loop_end
; p
= NEXT_INSN (p
))
2137 if (GET_CODE (p
) == JUMP_INSN
)
2139 target_insn
= JUMP_LABEL (p
);
2141 /* Search from loop_start to insn, to see if one of them is
2142 the target_insn. We can't use INSN_LUID comparisons here,
2143 since insn may not have an LUID entry. */
2144 for (q
= loop_start
; q
!= insn
; q
= NEXT_INSN (q
))
2145 if (q
== target_insn
)
2153 /* Try to generate the simplest rtx for the expression
2154 (PLUS (MULT mult1 mult2) add1). This is used to calculate the initial
2158 fold_rtx_mult_add (mult1
, mult2
, add1
, mode
)
2159 rtx mult1
, mult2
, add1
;
2160 enum machine_mode mode
;
2165 /* The modes must all be the same. This should always be true. For now,
2166 check to make sure. */
2167 if ((GET_MODE (mult1
) != mode
&& GET_MODE (mult1
) != VOIDmode
)
2168 || (GET_MODE (mult2
) != mode
&& GET_MODE (mult2
) != VOIDmode
)
2169 || (GET_MODE (add1
) != mode
&& GET_MODE (add1
) != VOIDmode
))
2172 /* Ensure that if at least one of mult1/mult2 are constant, then mult2
2173 will be a constant. */
2174 if (GET_CODE (mult1
) == CONST_INT
)
2181 mult_res
= simplify_binary_operation (MULT
, mode
, mult1
, mult2
);
2183 mult_res
= gen_rtx (MULT
, mode
, mult1
, mult2
);
2185 /* Again, put the constant second. */
2186 if (GET_CODE (add1
) == CONST_INT
)
2193 result
= simplify_binary_operation (PLUS
, mode
, add1
, mult_res
);
2195 result
= gen_rtx (PLUS
, mode
, add1
, mult_res
);
2200 /* Searches the list of induction struct's for the biv BL, to try to calculate
2201 the total increment value for one iteration of the loop as a constant.
2203 Returns the increment value as an rtx, simplified as much as possible,
2204 if it can be calculated. Otherwise, returns 0. */
2207 biv_total_increment (bl
, loop_start
, loop_end
)
2208 struct iv_class
*bl
;
2209 rtx loop_start
, loop_end
;
2211 struct induction
*v
;
2214 /* For increment, must check every instruction that sets it. Each
2215 instruction must be executed only once each time through the loop.
2216 To verify this, we check that the the insn is always executed, and that
2217 there are no backward branches after the insn that branch to before it.
2218 Also, the insn must have a mult_val of one (to make sure it really is
2221 result
= const0_rtx
;
2222 for (v
= bl
->biv
; v
; v
= v
->next_iv
)
2224 if (v
->always_computable
&& v
->mult_val
== const1_rtx
2225 && ! back_branch_in_range_p (v
->insn
, loop_start
, loop_end
))
2226 result
= fold_rtx_mult_add (result
, const1_rtx
, v
->add_val
, v
->mode
);
2234 /* Determine the initial value of the iteration variable, and the amount
2235 that it is incremented each loop. Use the tables constructed by
2236 the strength reduction pass to calculate these values.
2238 Initial_value and/or increment are set to zero if their values could not
2242 iteration_info (iteration_var
, initial_value
, increment
, loop_start
, loop_end
)
2243 rtx iteration_var
, *initial_value
, *increment
;
2244 rtx loop_start
, loop_end
;
2246 struct iv_class
*bl
;
2247 struct induction
*v
, *b
;
2249 /* Clear the result values, in case no answer can be found. */
2253 /* The iteration variable can be either a giv or a biv. Check to see
2254 which it is, and compute the variable's initial value, and increment
2255 value if possible. */
2257 /* If this is a new register, can't handle it since we don't have any
2258 reg_iv_type entry for it. */
2259 if (REGNO (iteration_var
) >= max_reg_before_loop
)
2261 if (loop_dump_stream
)
2262 fprintf (loop_dump_stream
,
2263 "Loop unrolling: No reg_iv_type entry for iteration var.\n");
2266 /* Reject iteration variables larger than the host long size, since they
2267 could result in a number of iterations greater than the range of our
2268 `unsigned long' variable loop_n_iterations. */
2269 else if (GET_MODE_BITSIZE (GET_MODE (iteration_var
)) > HOST_BITS_PER_LONG
)
2271 if (loop_dump_stream
)
2272 fprintf (loop_dump_stream
,
2273 "Loop unrolling: Iteration var rejected because mode larger than host long.\n");
2276 else if (GET_MODE_CLASS (GET_MODE (iteration_var
)) != MODE_INT
)
2278 if (loop_dump_stream
)
2279 fprintf (loop_dump_stream
,
2280 "Loop unrolling: Iteration var not an integer.\n");
2283 else if (reg_iv_type
[REGNO (iteration_var
)] == BASIC_INDUCT
)
2285 /* Grab initial value, only useful if it is a constant. */
2286 bl
= reg_biv_class
[REGNO (iteration_var
)];
2287 *initial_value
= bl
->initial_value
;
2289 *increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2291 else if (reg_iv_type
[REGNO (iteration_var
)] == GENERAL_INDUCT
)
2294 /* ??? The code below does not work because the incorrect number of
2295 iterations is calculated when the biv is incremented after the giv
2296 is set (which is the usual case). This can probably be accounted
2297 for by biasing the initial_value by subtracting the amount of the
2298 increment that occurs between the giv set and the giv test. However,
2299 a giv as an iterator is very rare, so it does not seem worthwhile
2301 /* ??? An example failure is: i = 6; do {;} while (i++ < 9). */
2302 if (loop_dump_stream
)
2303 fprintf (loop_dump_stream
,
2304 "Loop unrolling: Giv iterators are not handled.\n");
2307 /* Initial value is mult_val times the biv's initial value plus
2308 add_val. Only useful if it is a constant. */
2309 v
= reg_iv_info
[REGNO (iteration_var
)];
2310 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
2311 *initial_value
= fold_rtx_mult_add (v
->mult_val
, bl
->initial_value
,
2312 v
->add_val
, v
->mode
);
2314 /* Increment value is mult_val times the increment value of the biv. */
2316 *increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2318 *increment
= fold_rtx_mult_add (v
->mult_val
, *increment
, const0_rtx
,
2324 if (loop_dump_stream
)
2325 fprintf (loop_dump_stream
,
2326 "Loop unrolling: Not basic or general induction var.\n");
2331 /* Calculate the approximate final value of the iteration variable
2332 which has an loop exit test with code COMPARISON_CODE and comparison value
2333 of COMPARISON_VALUE. Also returns an indication of whether the comparison
2334 was signed or unsigned, and the direction of the comparison. This info is
2335 needed to calculate the number of loop iterations. */
2338 approx_final_value (comparison_code
, comparison_value
, unsigned_p
, compare_dir
)
2339 enum rtx_code comparison_code
;
2340 rtx comparison_value
;
2344 /* Calculate the final value of the induction variable.
2345 The exact final value depends on the branch operator, and increment sign.
2346 This is only an approximate value. It will be wrong if the iteration
2347 variable is not incremented by one each time through the loop, and
2348 approx final value - start value % increment != 0. */
2351 switch (comparison_code
)
2357 return plus_constant (comparison_value
, 1);
2362 return plus_constant (comparison_value
, -1);
2364 /* Can not calculate a final value for this case. */
2371 return comparison_value
;
2377 return comparison_value
;
2380 return comparison_value
;
2386 /* For each biv and giv, determine whether it can be safely split into
2387 a different variable for each unrolled copy of the loop body. If it
2388 is safe to split, then indicate that by saving some useful info
2389 in the splittable_regs array.
2391 If the loop is being completely unrolled, then splittable_regs will hold
2392 the current value of the induction variable while the loop is unrolled.
2393 It must be set to the initial value of the induction variable here.
2394 Otherwise, splittable_regs will hold the difference between the current
2395 value of the induction variable and the value the induction variable had
2396 at the top of the loop. It must be set to the value 0 here.
2398 Returns the total number of instructions that set registers that are
2401 /* ?? If the loop is only unrolled twice, then most of the restrictions to
2402 constant values are unnecessary, since we can easily calculate increment
2403 values in this case even if nothing is constant. The increment value
2404 should not involve a multiply however. */
2406 /* ?? Even if the biv/giv increment values aren't constant, it may still
2407 be beneficial to split the variable if the loop is only unrolled a few
2408 times, since multiplies by small integers (1,2,3,4) are very cheap. */
2411 find_splittable_regs (unroll_type
, loop_start
, loop_end
, end_insert_before
,
2413 enum unroll_types unroll_type
;
2414 rtx loop_start
, loop_end
;
2415 rtx end_insert_before
;
2418 struct iv_class
*bl
;
2419 struct induction
*v
;
2421 rtx biv_final_value
;
2425 for (bl
= loop_iv_list
; bl
; bl
= bl
->next
)
2427 /* Biv_total_increment must return a constant value,
2428 otherwise we can not calculate the split values. */
2430 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2431 if (! increment
|| GET_CODE (increment
) != CONST_INT
)
2434 /* The loop must be unrolled completely, or else have a known number
2435 of iterations and only one exit, or else the biv must be dead
2436 outside the loop, or else the final value must be known. Otherwise,
2437 it is unsafe to split the biv since it may not have the proper
2438 value on loop exit. */
2440 /* loop_number_exit_count is non-zero if the loop has an exit other than
2441 a fall through at the end. */
2444 biv_final_value
= 0;
2445 if (unroll_type
!= UNROLL_COMPLETELY
2446 && (loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]]
2447 || unroll_type
== UNROLL_NAIVE
)
2448 && (uid_luid
[regno_last_uid
[bl
->regno
]] >= INSN_LUID (loop_end
)
2450 || INSN_UID (bl
->init_insn
) >= max_uid_for_loop
2451 || (uid_luid
[regno_first_uid
[bl
->regno
]]
2452 < INSN_LUID (bl
->init_insn
))
2453 || reg_mentioned_p (bl
->biv
->dest_reg
, SET_SRC (bl
->init_set
)))
2454 && ! (biv_final_value
= final_biv_value (bl
, loop_start
, loop_end
)))
2457 /* If any of the insns setting the BIV don't do so with a simple
2458 PLUS, we don't know how to split it. */
2459 for (v
= bl
->biv
; biv_splittable
&& v
; v
= v
->next_iv
)
2460 if ((tem
= single_set (v
->insn
)) == 0
2461 || GET_CODE (SET_DEST (tem
)) != REG
2462 || REGNO (SET_DEST (tem
)) != bl
->regno
2463 || GET_CODE (SET_SRC (tem
)) != PLUS
)
2466 /* If final value is non-zero, then must emit an instruction which sets
2467 the value of the biv to the proper value. This is done after
2468 handling all of the givs, since some of them may need to use the
2469 biv's value in their initialization code. */
2471 /* This biv is splittable. If completely unrolling the loop, save
2472 the biv's initial value. Otherwise, save the constant zero. */
2474 if (biv_splittable
== 1)
2476 if (unroll_type
== UNROLL_COMPLETELY
)
2478 /* If the initial value of the biv is itself (i.e. it is too
2479 complicated for strength_reduce to compute), or is a hard
2480 register, or it isn't invariant, then we must create a new
2481 pseudo reg to hold the initial value of the biv. */
2483 if (GET_CODE (bl
->initial_value
) == REG
2484 && (REGNO (bl
->initial_value
) == bl
->regno
2485 || REGNO (bl
->initial_value
) < FIRST_PSEUDO_REGISTER
2486 || ! invariant_p (bl
->initial_value
)))
2488 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2490 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2493 if (loop_dump_stream
)
2494 fprintf (loop_dump_stream
, "Biv %d initial value remapped to %d.\n",
2495 bl
->regno
, REGNO (tem
));
2497 splittable_regs
[bl
->regno
] = tem
;
2500 splittable_regs
[bl
->regno
] = bl
->initial_value
;
2503 splittable_regs
[bl
->regno
] = const0_rtx
;
2505 /* Save the number of instructions that modify the biv, so that
2506 we can treat the last one specially. */
2508 splittable_regs_updates
[bl
->regno
] = bl
->biv_count
;
2509 result
+= bl
->biv_count
;
2511 if (loop_dump_stream
)
2512 fprintf (loop_dump_stream
,
2513 "Biv %d safe to split.\n", bl
->regno
);
2516 /* Check every giv that depends on this biv to see whether it is
2517 splittable also. Even if the biv isn't splittable, givs which
2518 depend on it may be splittable if the biv is live outside the
2519 loop, and the givs aren't. */
2521 result
+= find_splittable_givs (bl
, unroll_type
, loop_start
, loop_end
,
2522 increment
, unroll_number
);
2524 /* If final value is non-zero, then must emit an instruction which sets
2525 the value of the biv to the proper value. This is done after
2526 handling all of the givs, since some of them may need to use the
2527 biv's value in their initialization code. */
2528 if (biv_final_value
)
2530 /* If the loop has multiple exits, emit the insns before the
2531 loop to ensure that it will always be executed no matter
2532 how the loop exits. Otherwise emit the insn after the loop,
2533 since this is slightly more efficient. */
2534 if (! loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]])
2535 emit_insn_before (gen_move_insn (bl
->biv
->src_reg
,
2540 /* Create a new register to hold the value of the biv, and then
2541 set the biv to its final value before the loop start. The biv
2542 is set to its final value before loop start to ensure that
2543 this insn will always be executed, no matter how the loop
2545 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2546 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2548 emit_insn_before (gen_move_insn (bl
->biv
->src_reg
,
2552 if (loop_dump_stream
)
2553 fprintf (loop_dump_stream
, "Biv %d mapped to %d for split.\n",
2554 REGNO (bl
->biv
->src_reg
), REGNO (tem
));
2556 /* Set up the mapping from the original biv register to the new
2558 bl
->biv
->src_reg
= tem
;
2565 /* Return 1 if the first and last unrolled copy of the address giv V is valid
2566 for the instruction that is using it. Do not make any changes to that
2570 verify_addresses (v
, giv_inc
, unroll_number
)
2571 struct induction
*v
;
2576 rtx orig_addr
= *v
->location
;
2577 rtx last_addr
= plus_constant (v
->dest_reg
,
2578 INTVAL (giv_inc
) * (unroll_number
- 1));
2580 /* First check to see if either address would fail. */
2581 if (! validate_change (v
->insn
, v
->location
, v
->dest_reg
, 0)
2582 || ! validate_change (v
->insn
, v
->location
, last_addr
, 0))
2585 /* Now put things back the way they were before. This will always
2587 validate_change (v
->insn
, v
->location
, orig_addr
, 0);
2592 /* For every giv based on the biv BL, check to determine whether it is
2593 splittable. This is a subroutine to find_splittable_regs ().
2595 Return the number of instructions that set splittable registers. */
2598 find_splittable_givs (bl
, unroll_type
, loop_start
, loop_end
, increment
,
2600 struct iv_class
*bl
;
2601 enum unroll_types unroll_type
;
2602 rtx loop_start
, loop_end
;
2606 struct induction
*v
, *v2
;
2611 /* Scan the list of givs, and set the same_insn field when there are
2612 multiple identical givs in the same insn. */
2613 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
2614 for (v2
= v
->next_iv
; v2
; v2
= v2
->next_iv
)
2615 if (v
->insn
== v2
->insn
&& rtx_equal_p (v
->new_reg
, v2
->new_reg
)
2619 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
2623 /* Only split the giv if it has already been reduced, or if the loop is
2624 being completely unrolled. */
2625 if (unroll_type
!= UNROLL_COMPLETELY
&& v
->ignore
)
2628 /* The giv can be split if the insn that sets the giv is executed once
2629 and only once on every iteration of the loop. */
2630 /* An address giv can always be split. v->insn is just a use not a set,
2631 and hence it does not matter whether it is always executed. All that
2632 matters is that all the biv increments are always executed, and we
2633 won't reach here if they aren't. */
2634 if (v
->giv_type
!= DEST_ADDR
2635 && (! v
->always_computable
2636 || back_branch_in_range_p (v
->insn
, loop_start
, loop_end
)))
2639 /* The giv increment value must be a constant. */
2640 giv_inc
= fold_rtx_mult_add (v
->mult_val
, increment
, const0_rtx
,
2642 if (! giv_inc
|| GET_CODE (giv_inc
) != CONST_INT
)
2645 /* The loop must be unrolled completely, or else have a known number of
2646 iterations and only one exit, or else the giv must be dead outside
2647 the loop, or else the final value of the giv must be known.
2648 Otherwise, it is not safe to split the giv since it may not have the
2649 proper value on loop exit. */
2651 /* The used outside loop test will fail for DEST_ADDR givs. They are
2652 never used outside the loop anyways, so it is always safe to split a
2656 if (unroll_type
!= UNROLL_COMPLETELY
2657 && (loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]]
2658 || unroll_type
== UNROLL_NAIVE
)
2659 && v
->giv_type
!= DEST_ADDR
2660 && ((regno_first_uid
[REGNO (v
->dest_reg
)] != INSN_UID (v
->insn
)
2661 /* Check for the case where the pseudo is set by a shift/add
2662 sequence, in which case the first insn setting the pseudo
2663 is the first insn of the shift/add sequence. */
2664 && (! (tem
= find_reg_note (v
->insn
, REG_RETVAL
, NULL_RTX
))
2665 || (regno_first_uid
[REGNO (v
->dest_reg
)]
2666 != INSN_UID (XEXP (tem
, 0)))))
2667 /* Line above always fails if INSN was moved by loop opt. */
2668 || (uid_luid
[regno_last_uid
[REGNO (v
->dest_reg
)]]
2669 >= INSN_LUID (loop_end
)))
2670 && ! (final_value
= v
->final_value
))
2674 /* Currently, non-reduced/final-value givs are never split. */
2675 /* Should emit insns after the loop if possible, as the biv final value
2678 /* If the final value is non-zero, and the giv has not been reduced,
2679 then must emit an instruction to set the final value. */
2680 if (final_value
&& !v
->new_reg
)
2682 /* Create a new register to hold the value of the giv, and then set
2683 the giv to its final value before the loop start. The giv is set
2684 to its final value before loop start to ensure that this insn
2685 will always be executed, no matter how we exit. */
2686 tem
= gen_reg_rtx (v
->mode
);
2687 emit_insn_before (gen_move_insn (tem
, v
->dest_reg
), loop_start
);
2688 emit_insn_before (gen_move_insn (v
->dest_reg
, final_value
),
2691 if (loop_dump_stream
)
2692 fprintf (loop_dump_stream
, "Giv %d mapped to %d for split.\n",
2693 REGNO (v
->dest_reg
), REGNO (tem
));
2699 /* This giv is splittable. If completely unrolling the loop, save the
2700 giv's initial value. Otherwise, save the constant zero for it. */
2702 if (unroll_type
== UNROLL_COMPLETELY
)
2704 /* It is not safe to use bl->initial_value here, because it may not
2705 be invariant. It is safe to use the initial value stored in
2706 the splittable_regs array if it is set. In rare cases, it won't
2707 be set, so then we do exactly the same thing as
2708 find_splittable_regs does to get a safe value. */
2709 rtx biv_initial_value
;
2711 if (splittable_regs
[bl
->regno
])
2712 biv_initial_value
= splittable_regs
[bl
->regno
];
2713 else if (GET_CODE (bl
->initial_value
) != REG
2714 || (REGNO (bl
->initial_value
) != bl
->regno
2715 && REGNO (bl
->initial_value
) >= FIRST_PSEUDO_REGISTER
))
2716 biv_initial_value
= bl
->initial_value
;
2719 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2721 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2723 biv_initial_value
= tem
;
2725 value
= fold_rtx_mult_add (v
->mult_val
, biv_initial_value
,
2726 v
->add_val
, v
->mode
);
2733 /* If a giv was combined with another giv, then we can only split
2734 this giv if the giv it was combined with was reduced. This
2735 is because the value of v->new_reg is meaningless in this
2737 if (v
->same
&& ! v
->same
->new_reg
)
2739 if (loop_dump_stream
)
2740 fprintf (loop_dump_stream
,
2741 "giv combined with unreduced giv not split.\n");
2744 /* If the giv is an address destination, it could be something other
2745 than a simple register, these have to be treated differently. */
2746 else if (v
->giv_type
== DEST_REG
)
2748 /* If value is not a constant, register, or register plus
2749 constant, then compute its value into a register before
2750 loop start. This prevents invalid rtx sharing, and should
2751 generate better code. We can use bl->initial_value here
2752 instead of splittable_regs[bl->regno] because this code
2753 is going before the loop start. */
2754 if (unroll_type
== UNROLL_COMPLETELY
2755 && GET_CODE (value
) != CONST_INT
2756 && GET_CODE (value
) != REG
2757 && (GET_CODE (value
) != PLUS
2758 || GET_CODE (XEXP (value
, 0)) != REG
2759 || GET_CODE (XEXP (value
, 1)) != CONST_INT
))
2761 rtx tem
= gen_reg_rtx (v
->mode
);
2762 emit_iv_add_mult (bl
->initial_value
, v
->mult_val
,
2763 v
->add_val
, tem
, loop_start
);
2767 splittable_regs
[REGNO (v
->new_reg
)] = value
;
2771 /* Splitting address givs is useful since it will often allow us
2772 to eliminate some increment insns for the base giv as
2775 /* If the addr giv is combined with a dest_reg giv, then all
2776 references to that dest reg will be remapped, which is NOT
2777 what we want for split addr regs. We always create a new
2778 register for the split addr giv, just to be safe. */
2780 /* ??? If there are multiple address givs which have been
2781 combined with the same dest_reg giv, then we may only need
2782 one new register for them. Pulling out constants below will
2783 catch some of the common cases of this. Currently, I leave
2784 the work of simplifying multiple address givs to the
2785 following cse pass. */
2787 /* As a special case, if we have multiple identical address givs
2788 within a single instruction, then we do use a single pseudo
2789 reg for both. This is necessary in case one is a match_dup
2792 v
->const_adjust
= 0;
2796 v
->dest_reg
= v
->same_insn
->dest_reg
;
2797 if (loop_dump_stream
)
2798 fprintf (loop_dump_stream
,
2799 "Sharing address givs in insn %d\n",
2800 INSN_UID (v
->insn
));
2802 else if (unroll_type
!= UNROLL_COMPLETELY
)
2804 /* If not completely unrolling the loop, then create a new
2805 register to hold the split value of the DEST_ADDR giv.
2806 Emit insn to initialize its value before loop start. */
2807 tem
= gen_reg_rtx (v
->mode
);
2809 /* If the address giv has a constant in its new_reg value,
2810 then this constant can be pulled out and put in value,
2811 instead of being part of the initialization code. */
2813 if (GET_CODE (v
->new_reg
) == PLUS
2814 && GET_CODE (XEXP (v
->new_reg
, 1)) == CONST_INT
)
2817 = plus_constant (tem
, INTVAL (XEXP (v
->new_reg
,1)));
2819 /* Only succeed if this will give valid addresses.
2820 Try to validate both the first and the last
2821 address resulting from loop unrolling, if
2822 one fails, then can't do const elim here. */
2823 if (verify_addresses (v
, giv_inc
, unroll_number
))
2825 /* Save the negative of the eliminated const, so
2826 that we can calculate the dest_reg's increment
2828 v
->const_adjust
= - INTVAL (XEXP (v
->new_reg
, 1));
2830 v
->new_reg
= XEXP (v
->new_reg
, 0);
2831 if (loop_dump_stream
)
2832 fprintf (loop_dump_stream
,
2833 "Eliminating constant from giv %d\n",
2842 /* If the address hasn't been checked for validity yet, do so
2843 now, and fail completely if either the first or the last
2844 unrolled copy of the address is not a valid address
2845 for the instruction that uses it. */
2846 if (v
->dest_reg
== tem
2847 && ! verify_addresses (v
, giv_inc
, unroll_number
))
2849 if (loop_dump_stream
)
2850 fprintf (loop_dump_stream
,
2851 "Invalid address for giv at insn %d\n",
2852 INSN_UID (v
->insn
));
2856 /* To initialize the new register, just move the value of
2857 new_reg into it. This is not guaranteed to give a valid
2858 instruction on machines with complex addressing modes.
2859 If we can't recognize it, then delete it and emit insns
2860 to calculate the value from scratch. */
2861 emit_insn_before (gen_rtx (SET
, VOIDmode
, tem
,
2862 copy_rtx (v
->new_reg
)),
2864 if (recog_memoized (PREV_INSN (loop_start
)) < 0)
2868 /* We can't use bl->initial_value to compute the initial
2869 value, because the loop may have been preconditioned.
2870 We must calculate it from NEW_REG. Try using
2871 force_operand instead of emit_iv_add_mult. */
2872 delete_insn (PREV_INSN (loop_start
));
2875 ret
= force_operand (v
->new_reg
, tem
);
2877 emit_move_insn (tem
, ret
);
2878 sequence
= gen_sequence ();
2880 emit_insn_before (sequence
, loop_start
);
2882 if (loop_dump_stream
)
2883 fprintf (loop_dump_stream
,
2884 "Invalid init insn, rewritten.\n");
2889 v
->dest_reg
= value
;
2891 /* Check the resulting address for validity, and fail
2892 if the resulting address would be invalid. */
2893 if (! verify_addresses (v
, giv_inc
, unroll_number
))
2895 if (loop_dump_stream
)
2896 fprintf (loop_dump_stream
,
2897 "Invalid address for giv at insn %d\n",
2898 INSN_UID (v
->insn
));
2903 /* Store the value of dest_reg into the insn. This sharing
2904 will not be a problem as this insn will always be copied
2907 *v
->location
= v
->dest_reg
;
2909 /* If this address giv is combined with a dest reg giv, then
2910 save the base giv's induction pointer so that we will be
2911 able to handle this address giv properly. The base giv
2912 itself does not have to be splittable. */
2914 if (v
->same
&& v
->same
->giv_type
== DEST_REG
)
2915 addr_combined_regs
[REGNO (v
->same
->new_reg
)] = v
->same
;
2917 if (GET_CODE (v
->new_reg
) == REG
)
2919 /* This giv maybe hasn't been combined with any others.
2920 Make sure that it's giv is marked as splittable here. */
2922 splittable_regs
[REGNO (v
->new_reg
)] = value
;
2924 /* Make it appear to depend upon itself, so that the
2925 giv will be properly split in the main loop above. */
2929 addr_combined_regs
[REGNO (v
->new_reg
)] = v
;
2933 if (loop_dump_stream
)
2934 fprintf (loop_dump_stream
, "DEST_ADDR giv being split.\n");
2940 /* Currently, unreduced giv's can't be split. This is not too much
2941 of a problem since unreduced giv's are not live across loop
2942 iterations anyways. When unrolling a loop completely though,
2943 it makes sense to reduce&split givs when possible, as this will
2944 result in simpler instructions, and will not require that a reg
2945 be live across loop iterations. */
2947 splittable_regs
[REGNO (v
->dest_reg
)] = value
;
2948 fprintf (stderr
, "Giv %d at insn %d not reduced\n",
2949 REGNO (v
->dest_reg
), INSN_UID (v
->insn
));
2955 /* Givs are only updated once by definition. Mark it so if this is
2956 a splittable register. Don't need to do anything for address givs
2957 where this may not be a register. */
2959 if (GET_CODE (v
->new_reg
) == REG
)
2960 splittable_regs_updates
[REGNO (v
->new_reg
)] = 1;
2964 if (loop_dump_stream
)
2968 if (GET_CODE (v
->dest_reg
) == CONST_INT
)
2970 else if (GET_CODE (v
->dest_reg
) != REG
)
2971 regnum
= REGNO (XEXP (v
->dest_reg
, 0));
2973 regnum
= REGNO (v
->dest_reg
);
2974 fprintf (loop_dump_stream
, "Giv %d at insn %d safe to split.\n",
2975 regnum
, INSN_UID (v
->insn
));
2982 /* Try to prove that the register is dead after the loop exits. Trace every
2983 loop exit looking for an insn that will always be executed, which sets
2984 the register to some value, and appears before the first use of the register
2985 is found. If successful, then return 1, otherwise return 0. */
2987 /* ?? Could be made more intelligent in the handling of jumps, so that
2988 it can search past if statements and other similar structures. */
2991 reg_dead_after_loop (reg
, loop_start
, loop_end
)
2992 rtx reg
, loop_start
, loop_end
;
2997 int label_count
= 0;
2998 int this_loop_num
= uid_loop_num
[INSN_UID (loop_start
)];
3000 /* In addition to checking all exits of this loop, we must also check
3001 all exits of inner nested loops that would exit this loop. We don't
3002 have any way to identify those, so we just give up if there are any
3003 such inner loop exits. */
3005 for (label
= loop_number_exit_labels
[this_loop_num
]; label
;
3006 label
= LABEL_NEXTREF (label
))
3009 if (label_count
!= loop_number_exit_count
[this_loop_num
])
3012 /* HACK: Must also search the loop fall through exit, create a label_ref
3013 here which points to the loop_end, and append the loop_number_exit_labels
3015 label
= gen_rtx (LABEL_REF
, VOIDmode
, loop_end
);
3016 LABEL_NEXTREF (label
) = loop_number_exit_labels
[this_loop_num
];
3018 for ( ; label
; label
= LABEL_NEXTREF (label
))
3020 /* Succeed if find an insn which sets the biv or if reach end of
3021 function. Fail if find an insn that uses the biv, or if come to
3022 a conditional jump. */
3024 insn
= NEXT_INSN (XEXP (label
, 0));
3027 code
= GET_CODE (insn
);
3028 if (GET_RTX_CLASS (code
) == 'i')
3032 if (reg_referenced_p (reg
, PATTERN (insn
)))
3035 set
= single_set (insn
);
3036 if (set
&& rtx_equal_p (SET_DEST (set
), reg
))
3040 if (code
== JUMP_INSN
)
3042 if (GET_CODE (PATTERN (insn
)) == RETURN
)
3044 else if (! simplejump_p (insn
)
3045 /* Prevent infinite loop following infinite loops. */
3046 || jump_count
++ > 20)
3049 insn
= JUMP_LABEL (insn
);
3052 insn
= NEXT_INSN (insn
);
3056 /* Success, the register is dead on all loop exits. */
3060 /* Try to calculate the final value of the biv, the value it will have at
3061 the end of the loop. If we can do it, return that value. */
3064 final_biv_value (bl
, loop_start
, loop_end
)
3065 struct iv_class
*bl
;
3066 rtx loop_start
, loop_end
;
3070 /* ??? This only works for MODE_INT biv's. Reject all others for now. */
3072 if (GET_MODE_CLASS (bl
->biv
->mode
) != MODE_INT
)
3075 /* The final value for reversed bivs must be calculated differently than
3076 for ordinary bivs. In this case, there is already an insn after the
3077 loop which sets this biv's final value (if necessary), and there are
3078 no other loop exits, so we can return any value. */
3081 if (loop_dump_stream
)
3082 fprintf (loop_dump_stream
,
3083 "Final biv value for %d, reversed biv.\n", bl
->regno
);
3088 /* Try to calculate the final value as initial value + (number of iterations
3089 * increment). For this to work, increment must be invariant, the only
3090 exit from the loop must be the fall through at the bottom (otherwise
3091 it may not have its final value when the loop exits), and the initial
3092 value of the biv must be invariant. */
3094 if (loop_n_iterations
!= 0
3095 && ! loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]]
3096 && invariant_p (bl
->initial_value
))
3098 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
3100 if (increment
&& invariant_p (increment
))
3102 /* Can calculate the loop exit value, emit insns after loop
3103 end to calculate this value into a temporary register in
3104 case it is needed later. */
3106 tem
= gen_reg_rtx (bl
->biv
->mode
);
3107 /* Make sure loop_end is not the last insn. */
3108 if (NEXT_INSN (loop_end
) == 0)
3109 emit_note_after (NOTE_INSN_DELETED
, loop_end
);
3110 emit_iv_add_mult (increment
, GEN_INT (loop_n_iterations
),
3111 bl
->initial_value
, tem
, NEXT_INSN (loop_end
));
3113 if (loop_dump_stream
)
3114 fprintf (loop_dump_stream
,
3115 "Final biv value for %d, calculated.\n", bl
->regno
);
3121 /* Check to see if the biv is dead at all loop exits. */
3122 if (reg_dead_after_loop (bl
->biv
->src_reg
, loop_start
, loop_end
))
3124 if (loop_dump_stream
)
3125 fprintf (loop_dump_stream
,
3126 "Final biv value for %d, biv dead after loop exit.\n",
3135 /* Try to calculate the final value of the giv, the value it will have at
3136 the end of the loop. If we can do it, return that value. */
3139 final_giv_value (v
, loop_start
, loop_end
)
3140 struct induction
*v
;
3141 rtx loop_start
, loop_end
;
3143 struct iv_class
*bl
;
3146 rtx insert_before
, seq
;
3148 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
3150 /* The final value for givs which depend on reversed bivs must be calculated
3151 differently than for ordinary givs. In this case, there is already an
3152 insn after the loop which sets this giv's final value (if necessary),
3153 and there are no other loop exits, so we can return any value. */
3156 if (loop_dump_stream
)
3157 fprintf (loop_dump_stream
,
3158 "Final giv value for %d, depends on reversed biv\n",
3159 REGNO (v
->dest_reg
));
3163 /* Try to calculate the final value as a function of the biv it depends
3164 upon. The only exit from the loop must be the fall through at the bottom
3165 (otherwise it may not have its final value when the loop exits). */
3167 /* ??? Can calculate the final giv value by subtracting off the
3168 extra biv increments times the giv's mult_val. The loop must have
3169 only one exit for this to work, but the loop iterations does not need
3172 if (loop_n_iterations
!= 0
3173 && ! loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]])
3175 /* ?? It is tempting to use the biv's value here since these insns will
3176 be put after the loop, and hence the biv will have its final value
3177 then. However, this fails if the biv is subsequently eliminated.
3178 Perhaps determine whether biv's are eliminable before trying to
3179 determine whether giv's are replaceable so that we can use the
3180 biv value here if it is not eliminable. */
3182 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
3184 if (increment
&& invariant_p (increment
))
3186 /* Can calculate the loop exit value of its biv as
3187 (loop_n_iterations * increment) + initial_value */
3189 /* The loop exit value of the giv is then
3190 (final_biv_value - extra increments) * mult_val + add_val.
3191 The extra increments are any increments to the biv which
3192 occur in the loop after the giv's value is calculated.
3193 We must search from the insn that sets the giv to the end
3194 of the loop to calculate this value. */
3196 insert_before
= NEXT_INSN (loop_end
);
3198 /* Put the final biv value in tem. */
3199 tem
= gen_reg_rtx (bl
->biv
->mode
);
3200 emit_iv_add_mult (increment
, GEN_INT (loop_n_iterations
),
3201 bl
->initial_value
, tem
, insert_before
);
3203 /* Subtract off extra increments as we find them. */
3204 for (insn
= NEXT_INSN (v
->insn
); insn
!= loop_end
;
3205 insn
= NEXT_INSN (insn
))
3207 struct induction
*biv
;
3209 for (biv
= bl
->biv
; biv
; biv
= biv
->next_iv
)
3210 if (biv
->insn
== insn
)
3213 tem
= expand_binop (GET_MODE (tem
), sub_optab
, tem
,
3214 biv
->add_val
, NULL_RTX
, 0,
3216 seq
= gen_sequence ();
3218 emit_insn_before (seq
, insert_before
);
3222 /* Now calculate the giv's final value. */
3223 emit_iv_add_mult (tem
, v
->mult_val
, v
->add_val
, tem
,
3226 if (loop_dump_stream
)
3227 fprintf (loop_dump_stream
,
3228 "Final giv value for %d, calc from biv's value.\n",
3229 REGNO (v
->dest_reg
));
3235 /* Replaceable giv's should never reach here. */
3239 /* Check to see if the biv is dead at all loop exits. */
3240 if (reg_dead_after_loop (v
->dest_reg
, loop_start
, loop_end
))
3242 if (loop_dump_stream
)
3243 fprintf (loop_dump_stream
,
3244 "Final giv value for %d, giv dead after loop exit.\n",
3245 REGNO (v
->dest_reg
));
3254 /* Calculate the number of loop iterations. Returns the exact number of loop
3255 iterations if it can be calculated, otherwise returns zero. */
3257 unsigned HOST_WIDE_INT
3258 loop_iterations (loop_start
, loop_end
)
3259 rtx loop_start
, loop_end
;
3261 rtx comparison
, comparison_value
;
3262 rtx iteration_var
, initial_value
, increment
, final_value
;
3263 enum rtx_code comparison_code
;
3266 int unsigned_compare
, compare_dir
, final_larger
;
3267 unsigned long tempu
;
3270 /* First find the iteration variable. If the last insn is a conditional
3271 branch, and the insn before tests a register value, make that the
3272 iteration variable. */
3274 loop_initial_value
= 0;
3276 loop_final_value
= 0;
3277 loop_iteration_var
= 0;
3279 /* We used to use pren_nonnote_insn here, but that fails because it might
3280 accidentally get the branch for a contained loop if the branch for this
3281 loop was deleted. We can only trust branches immediately before the
3283 last_loop_insn
= PREV_INSN (loop_end
);
3285 comparison
= get_condition_for_loop (last_loop_insn
);
3286 if (comparison
== 0)
3288 if (loop_dump_stream
)
3289 fprintf (loop_dump_stream
,
3290 "Loop unrolling: No final conditional branch found.\n");
3294 /* ??? Get_condition may switch position of induction variable and
3295 invariant register when it canonicalizes the comparison. */
3297 comparison_code
= GET_CODE (comparison
);
3298 iteration_var
= XEXP (comparison
, 0);
3299 comparison_value
= XEXP (comparison
, 1);
3301 if (GET_CODE (iteration_var
) != REG
)
3303 if (loop_dump_stream
)
3304 fprintf (loop_dump_stream
,
3305 "Loop unrolling: Comparison not against register.\n");
3309 /* Loop iterations is always called before any new registers are created
3310 now, so this should never occur. */
3312 if (REGNO (iteration_var
) >= max_reg_before_loop
)
3315 iteration_info (iteration_var
, &initial_value
, &increment
,
3316 loop_start
, loop_end
);
3317 if (initial_value
== 0)
3318 /* iteration_info already printed a message. */
3321 /* If the comparison value is an invariant register, then try to find
3322 its value from the insns before the start of the loop. */
3324 if (GET_CODE (comparison_value
) == REG
&& invariant_p (comparison_value
))
3328 for (insn
= PREV_INSN (loop_start
); insn
; insn
= PREV_INSN (insn
))
3330 if (GET_CODE (insn
) == CODE_LABEL
)
3333 else if (GET_RTX_CLASS (GET_CODE (insn
)) == 'i'
3334 && reg_set_p (comparison_value
, insn
))
3336 /* We found the last insn before the loop that sets the register.
3337 If it sets the entire register, and has a REG_EQUAL note,
3338 then use the value of the REG_EQUAL note. */
3339 if ((set
= single_set (insn
))
3340 && (SET_DEST (set
) == comparison_value
))
3342 rtx note
= find_reg_note (insn
, REG_EQUAL
, NULL_RTX
);
3344 /* Only use the REG_EQUAL note if it is a constant.
3345 Other things, divide in particular, will cause
3346 problems later if we use them. */
3347 if (note
&& GET_CODE (XEXP (note
, 0)) != EXPR_LIST
3348 && CONSTANT_P (XEXP (note
, 0)))
3349 comparison_value
= XEXP (note
, 0);
3356 final_value
= approx_final_value (comparison_code
, comparison_value
,
3357 &unsigned_compare
, &compare_dir
);
3359 /* Save the calculated values describing this loop's bounds, in case
3360 precondition_loop_p will need them later. These values can not be
3361 recalculated inside precondition_loop_p because strength reduction
3362 optimizations may obscure the loop's structure. */
3364 loop_iteration_var
= iteration_var
;
3365 loop_initial_value
= initial_value
;
3366 loop_increment
= increment
;
3367 loop_final_value
= final_value
;
3368 loop_comparison_code
= comparison_code
;
3372 if (loop_dump_stream
)
3373 fprintf (loop_dump_stream
,
3374 "Loop unrolling: Increment value can't be calculated.\n");
3377 else if (GET_CODE (increment
) != CONST_INT
)
3379 if (loop_dump_stream
)
3380 fprintf (loop_dump_stream
,
3381 "Loop unrolling: Increment value not constant.\n");
3384 else if (GET_CODE (initial_value
) != CONST_INT
)
3386 if (loop_dump_stream
)
3387 fprintf (loop_dump_stream
,
3388 "Loop unrolling: Initial value not constant.\n");
3391 else if (final_value
== 0)
3393 if (loop_dump_stream
)
3394 fprintf (loop_dump_stream
,
3395 "Loop unrolling: EQ comparison loop.\n");
3398 else if (GET_CODE (final_value
) != CONST_INT
)
3400 if (loop_dump_stream
)
3401 fprintf (loop_dump_stream
,
3402 "Loop unrolling: Final value not constant.\n");
3406 /* ?? Final value and initial value do not have to be constants.
3407 Only their difference has to be constant. When the iteration variable
3408 is an array address, the final value and initial value might both
3409 be addresses with the same base but different constant offsets.
3410 Final value must be invariant for this to work.
3412 To do this, need some way to find the values of registers which are
3415 /* Final_larger is 1 if final larger, 0 if they are equal, otherwise -1. */
3416 if (unsigned_compare
)
3418 = ((unsigned HOST_WIDE_INT
) INTVAL (final_value
)
3419 > (unsigned HOST_WIDE_INT
) INTVAL (initial_value
))
3420 - ((unsigned HOST_WIDE_INT
) INTVAL (final_value
)
3421 < (unsigned HOST_WIDE_INT
) INTVAL (initial_value
));
3423 final_larger
= (INTVAL (final_value
) > INTVAL (initial_value
))
3424 - (INTVAL (final_value
) < INTVAL (initial_value
));
3426 if (INTVAL (increment
) > 0)
3428 else if (INTVAL (increment
) == 0)
3433 /* There are 27 different cases: compare_dir = -1, 0, 1;
3434 final_larger = -1, 0, 1; increment_dir = -1, 0, 1.
3435 There are 4 normal cases, 4 reverse cases (where the iteration variable
3436 will overflow before the loop exits), 4 infinite loop cases, and 15
3437 immediate exit (0 or 1 iteration depending on loop type) cases.
3438 Only try to optimize the normal cases. */
3440 /* (compare_dir/final_larger/increment_dir)
3441 Normal cases: (0/-1/-1), (0/1/1), (-1/-1/-1), (1/1/1)
3442 Reverse cases: (0/-1/1), (0/1/-1), (-1/-1/1), (1/1/-1)
3443 Infinite loops: (0/-1/0), (0/1/0), (-1/-1/0), (1/1/0)
3444 Immediate exit: (0/0/X), (-1/0/X), (-1/1/X), (1/0/X), (1/-1/X) */
3446 /* ?? If the meaning of reverse loops (where the iteration variable
3447 will overflow before the loop exits) is undefined, then could
3448 eliminate all of these special checks, and just always assume
3449 the loops are normal/immediate/infinite. Note that this means
3450 the sign of increment_dir does not have to be known. Also,
3451 since it does not really hurt if immediate exit loops or infinite loops
3452 are optimized, then that case could be ignored also, and hence all
3453 loops can be optimized.
3455 According to ANSI Spec, the reverse loop case result is undefined,
3456 because the action on overflow is undefined.
3458 See also the special test for NE loops below. */
3460 if (final_larger
== increment_dir
&& final_larger
!= 0
3461 && (final_larger
== compare_dir
|| compare_dir
== 0))
3466 if (loop_dump_stream
)
3467 fprintf (loop_dump_stream
,
3468 "Loop unrolling: Not normal loop.\n");
3472 /* Calculate the number of iterations, final_value is only an approximation,
3473 so correct for that. Note that tempu and loop_n_iterations are
3474 unsigned, because they can be as large as 2^n - 1. */
3476 i
= INTVAL (increment
);
3478 tempu
= INTVAL (final_value
) - INTVAL (initial_value
);
3481 tempu
= INTVAL (initial_value
) - INTVAL (final_value
);
3487 /* For NE tests, make sure that the iteration variable won't miss the
3488 final value. If tempu mod i is not zero, then the iteration variable
3489 will overflow before the loop exits, and we can not calculate the
3490 number of iterations. */
3491 if (compare_dir
== 0 && (tempu
% i
) != 0)
3494 return tempu
/ i
+ ((tempu
% i
) != 0);
3497 /* Replace uses of split bivs with their split pseudo register. This is
3498 for original instructions which remain after loop unrolling without
3502 remap_split_bivs (x
)
3505 register enum rtx_code code
;
3512 code
= GET_CODE (x
);
3527 /* If non-reduced/final-value givs were split, then this would also
3528 have to remap those givs also. */
3530 if (REGNO (x
) < max_reg_before_loop
3531 && reg_iv_type
[REGNO (x
)] == BASIC_INDUCT
)
3532 return reg_biv_class
[REGNO (x
)]->biv
->src_reg
;
3535 fmt
= GET_RTX_FORMAT (code
);
3536 for (i
= GET_RTX_LENGTH (code
) - 1; i
>= 0; i
--)
3539 XEXP (x
, i
) = remap_split_bivs (XEXP (x
, i
));
3543 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
3544 XVECEXP (x
, i
, j
) = remap_split_bivs (XVECEXP (x
, i
, j
));
3550 /* If FIRST_UID is a set of REGNO, and FIRST_UID dominates LAST_UID (e.g.
3551 FIST_UID is always executed if LAST_UID is), then return 1. Otherwise
3552 return 0. COPY_START is where we can start looking for the insns
3553 FIRST_UID and LAST_UID. COPY_END is where we stop looking for these
3556 If there is no JUMP_INSN between LOOP_START and FIRST_UID, then FIRST_UID
3557 must dominate LAST_UID.
3559 If there is a CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID
3560 may not dominate LAST_UID.
3562 If there is no CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID
3563 must dominate LAST_UID. */
3566 set_dominates_use (regno
, first_uid
, last_uid
, copy_start
, copy_end
)
3573 int passed_jump
= 0;
3574 rtx p
= NEXT_INSN (copy_start
);
3576 while (INSN_UID (p
) != first_uid
)
3578 if (GET_CODE (p
) == JUMP_INSN
)
3580 /* Could not find FIRST_UID. */
3586 /* Verify that FIRST_UID is an insn that entirely sets REGNO. */
3587 if (GET_RTX_CLASS (GET_CODE (p
)) != 'i'
3588 || ! dead_or_set_regno_p (p
, regno
))
3591 /* FIRST_UID is always executed. */
3592 if (passed_jump
== 0)
3595 while (INSN_UID (p
) != last_uid
)
3597 /* If we see a CODE_LABEL between FIRST_UID and LAST_UID, then we
3598 can not be sure that FIRST_UID dominates LAST_UID. */
3599 if (GET_CODE (p
) == CODE_LABEL
)
3604 /* FIRST_UID is always executed if LAST_UID is executed. */