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, 675 Mass Ave, Cambridge, MA 02139, USA. */
21 /* Try to unroll a loop, and split induction variables.
23 Loops for which the number of iterations can be calculated exactly are
24 handled specially. If the number of iterations times the insn_count is
25 less than MAX_UNROLLED_INSNS, then the loop is unrolled completely.
26 Otherwise, we try to unroll the loop a number of times modulo the number
27 of iterations, so that only one exit test will be needed. It is unrolled
28 a number of times approximately equal to MAX_UNROLLED_INSNS divided by
31 Otherwise, if the number of iterations can be calculated exactly at
32 run time, and the loop is always entered at the top, then we try to
33 precondition the loop. That is, at run time, calculate how many times
34 the loop will execute, and then execute the loop body a few times so
35 that the remaining iterations will be some multiple of 4 (or 2 if the
36 loop is large). Then fall through to a loop unrolled 4 (or 2) times,
37 with only one exit test needed at the end of the loop.
39 Otherwise, if the number of iterations can not be calculated exactly,
40 not even at run time, then we still unroll the loop a number of times
41 approximately equal to MAX_UNROLLED_INSNS divided by the insn count,
42 but there must be an exit test after each copy of the loop body.
44 For each induction variable, which is dead outside the loop (replaceable)
45 or for which we can easily calculate the final value, if we can easily
46 calculate its value at each place where it is set as a function of the
47 current loop unroll count and the variable's value at loop entry, then
48 the induction variable is split into `N' different variables, one for
49 each copy of the loop body. One variable is live across the backward
50 branch, and the others are all calculated as a function of this variable.
51 This helps eliminate data dependencies, and leads to further opportunities
54 /* Possible improvements follow: */
56 /* ??? Add an extra pass somewhere to determine whether unrolling will
57 give any benefit. E.g. after generating all unrolled insns, compute the
58 cost of all insns and compare against cost of insns in rolled loop.
60 - On traditional architectures, unrolling a non-constant bound loop
61 is a win if there is a giv whose only use is in memory addresses, the
62 memory addresses can be split, and hence giv increments can be
64 - It is also a win if the loop is executed many times, and preconditioning
65 can be performed for the loop.
66 Add code to check for these and similar cases. */
68 /* ??? Improve control of which loops get unrolled. Could use profiling
69 info to only unroll the most commonly executed loops. Perhaps have
70 a user specifyable option to control the amount of code expansion,
71 or the percent of loops to consider for unrolling. Etc. */
73 /* ??? Look at the register copies inside the loop to see if they form a
74 simple permutation. If so, iterate the permutation until it gets back to
75 the start state. This is how many times we should unroll the loop, for
76 best results, because then all register copies can be eliminated.
77 For example, the lisp nreverse function should be unrolled 3 times
86 ??? The number of times to unroll the loop may also be based on data
87 references in the loop. For example, if we have a loop that references
88 x[i-1], x[i], and x[i+1], we should unroll it a multiple of 3 times. */
90 /* ??? Add some simple linear equation solving capability so that we can
91 determine the number of loop iterations for more complex loops.
92 For example, consider this loop from gdb
93 #define SWAP_TARGET_AND_HOST(buffer,len)
96 char *p = (char *) buffer;
97 char *q = ((char *) buffer) + len - 1;
98 int iterations = (len + 1) >> 1;
100 for (p; p < q; p++, q--;)
108 start value = p = &buffer + current_iteration
109 end value = q = &buffer + len - 1 - current_iteration
110 Given the loop exit test of "p < q", then there must be "q - p" iterations,
111 set equal to zero and solve for number of iterations:
112 q - p = len - 1 - 2*current_iteration = 0
113 current_iteration = (len - 1) / 2
114 Hence, there are (len - 1) / 2 (rounded up to the nearest integer)
115 iterations of this loop. */
117 /* ??? Currently, no labels are marked as loop invariant when doing loop
118 unrolling. This is because an insn inside the loop, that loads the address
119 of a label inside the loop into a register, could be moved outside the loop
120 by the invariant code motion pass if labels were invariant. If the loop
121 is subsequently unrolled, the code will be wrong because each unrolled
122 body of the loop will use the same address, whereas each actually needs a
123 different address. A case where this happens is when a loop containing
124 a switch statement is unrolled.
126 It would be better to let labels be considered invariant. When we
127 unroll loops here, check to see if any insns using a label local to the
128 loop were moved before the loop. If so, then correct the problem, by
129 moving the insn back into the loop, or perhaps replicate the insn before
130 the loop, one copy for each time the loop is unrolled. */
132 /* The prime factors looked for when trying to unroll a loop by some
133 number which is modulo the total number of iterations. Just checking
134 for these 4 prime factors will find at least one factor for 75% of
135 all numbers theoretically. Practically speaking, this will succeed
136 almost all of the time since loops are generally a multiple of 2
139 #define NUM_FACTORS 4
141 struct _factor
{ int factor
, count
; } factors
[NUM_FACTORS
]
142 = { {2, 0}, {3, 0}, {5, 0}, {7, 0}};
144 /* Describes the different types of loop unrolling performed. */
146 enum unroll_types
{ UNROLL_COMPLETELY
, UNROLL_MODULO
, UNROLL_NAIVE
};
150 #include "insn-config.h"
151 #include "integrate.h"
158 /* This controls which loops are unrolled, and by how much we unroll
161 #ifndef MAX_UNROLLED_INSNS
162 #define MAX_UNROLLED_INSNS 100
165 /* Indexed by register number, if non-zero, then it contains a pointer
166 to a struct induction for a DEST_REG giv which has been combined with
167 one of more address givs. This is needed because whenever such a DEST_REG
168 giv is modified, we must modify the value of all split address givs
169 that were combined with this DEST_REG giv. */
171 static struct induction
**addr_combined_regs
;
173 /* Indexed by register number, if this is a splittable induction variable,
174 then this will hold the current value of the register, which depends on the
177 static rtx
*splittable_regs
;
179 /* Indexed by register number, if this is a splittable induction variable,
180 then this will hold the number of instructions in the loop that modify
181 the induction variable. Used to ensure that only the last insn modifying
182 a split iv will update the original iv of the dest. */
184 static int *splittable_regs_updates
;
186 /* Values describing the current loop's iteration variable. These are set up
187 by loop_iterations, and used by precondition_loop_p. */
189 static rtx loop_iteration_var
;
190 static rtx loop_initial_value
;
191 static rtx loop_increment
;
192 static rtx loop_final_value
;
194 /* Forward declarations. */
196 static void init_reg_map
PROTO((struct inline_remap
*, int));
197 static int precondition_loop_p
PROTO((rtx
*, rtx
*, rtx
*, rtx
, rtx
));
198 static rtx calculate_giv_inc
PROTO((rtx
, rtx
, int));
199 static rtx initial_reg_note_copy
PROTO((rtx
, struct inline_remap
*));
200 static void final_reg_note_copy
PROTO((rtx
, struct inline_remap
*));
201 static void copy_loop_body
PROTO((rtx
, rtx
, struct inline_remap
*, rtx
, int,
202 enum unroll_types
, rtx
, rtx
, rtx
, rtx
));
203 static void iteration_info
PROTO((rtx
, rtx
*, rtx
*, rtx
, rtx
));
204 static rtx approx_final_value
PROTO((enum rtx_code
, rtx
, int *, int *));
205 static int find_splittable_regs
PROTO((enum unroll_types
, rtx
, rtx
, rtx
, int));
206 static int find_splittable_givs
PROTO((struct iv_class
*,enum unroll_types
,
207 rtx
, rtx
, rtx
, int));
208 static int reg_dead_after_loop
PROTO((rtx
, rtx
, rtx
));
209 static rtx fold_rtx_mult_add
PROTO((rtx
, rtx
, rtx
, enum machine_mode
));
210 static rtx remap_split_bivs
PROTO((rtx
));
212 /* Try to unroll one loop and split induction variables in the loop.
214 The loop is described by the arguments LOOP_END, INSN_COUNT, and
215 LOOP_START. END_INSERT_BEFORE indicates where insns should be added
216 which need to be executed when the loop falls through. STRENGTH_REDUCTION_P
217 indicates whether information generated in the strength reduction pass
220 This function is intended to be called from within `strength_reduce'
224 unroll_loop (loop_end
, insn_count
, loop_start
, end_insert_before
,
229 rtx end_insert_before
;
230 int strength_reduce_p
;
233 int unroll_number
= 1;
234 rtx copy_start
, copy_end
;
235 rtx insn
, copy
, sequence
, pattern
, tem
;
236 int max_labelno
, max_insnno
;
238 struct inline_remap
*map
;
246 int splitting_not_safe
= 0;
247 enum unroll_types unroll_type
;
248 int loop_preconditioned
= 0;
250 /* This points to the last real insn in the loop, which should be either
251 a JUMP_INSN (for conditional jumps) or a BARRIER (for unconditional
255 /* Don't bother unrolling huge loops. Since the minimum factor is
256 two, loops greater than one half of MAX_UNROLLED_INSNS will never
258 if (insn_count
> MAX_UNROLLED_INSNS
/ 2)
260 if (loop_dump_stream
)
261 fprintf (loop_dump_stream
, "Unrolling failure: Loop too big.\n");
265 /* When emitting debugger info, we can't unroll loops with unequal numbers
266 of block_beg and block_end notes, because that would unbalance the block
267 structure of the function. This can happen as a result of the
268 "if (foo) bar; else break;" optimization in jump.c. */
270 if (write_symbols
!= NO_DEBUG
)
272 int block_begins
= 0;
275 for (insn
= loop_start
; insn
!= loop_end
; insn
= NEXT_INSN (insn
))
277 if (GET_CODE (insn
) == NOTE
)
279 if (NOTE_LINE_NUMBER (insn
) == NOTE_INSN_BLOCK_BEG
)
281 else if (NOTE_LINE_NUMBER (insn
) == NOTE_INSN_BLOCK_END
)
286 if (block_begins
!= block_ends
)
288 if (loop_dump_stream
)
289 fprintf (loop_dump_stream
,
290 "Unrolling failure: Unbalanced block notes.\n");
295 /* Determine type of unroll to perform. Depends on the number of iterations
296 and the size of the loop. */
298 /* If there is no strength reduce info, then set loop_n_iterations to zero.
299 This can happen if strength_reduce can't find any bivs in the loop.
300 A value of zero indicates that the number of iterations could not be
303 if (! strength_reduce_p
)
304 loop_n_iterations
= 0;
306 if (loop_dump_stream
&& loop_n_iterations
> 0)
307 fprintf (loop_dump_stream
,
308 "Loop unrolling: %d iterations.\n", loop_n_iterations
);
310 /* Find and save a pointer to the last nonnote insn in the loop. */
312 last_loop_insn
= prev_nonnote_insn (loop_end
);
314 /* Calculate how many times to unroll the loop. Indicate whether or
315 not the loop is being completely unrolled. */
317 if (loop_n_iterations
== 1)
319 /* If number of iterations is exactly 1, then eliminate the compare and
320 branch at the end of the loop since they will never be taken.
321 Then return, since no other action is needed here. */
323 /* If the last instruction is not a BARRIER or a JUMP_INSN, then
324 don't do anything. */
326 if (GET_CODE (last_loop_insn
) == BARRIER
)
328 /* Delete the jump insn. This will delete the barrier also. */
329 delete_insn (PREV_INSN (last_loop_insn
));
331 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
334 /* The immediately preceding insn is a compare which must be
336 delete_insn (last_loop_insn
);
337 delete_insn (PREV_INSN (last_loop_insn
));
339 /* The immediately preceding insn may not be the compare, so don't
341 delete_insn (last_loop_insn
);
346 else if (loop_n_iterations
> 0
347 && loop_n_iterations
* insn_count
< MAX_UNROLLED_INSNS
)
349 unroll_number
= loop_n_iterations
;
350 unroll_type
= UNROLL_COMPLETELY
;
352 else if (loop_n_iterations
> 0)
354 /* Try to factor the number of iterations. Don't bother with the
355 general case, only using 2, 3, 5, and 7 will get 75% of all
356 numbers theoretically, and almost all in practice. */
358 for (i
= 0; i
< NUM_FACTORS
; i
++)
359 factors
[i
].count
= 0;
361 temp
= loop_n_iterations
;
362 for (i
= NUM_FACTORS
- 1; i
>= 0; i
--)
363 while (temp
% factors
[i
].factor
== 0)
366 temp
= temp
/ factors
[i
].factor
;
369 /* Start with the larger factors first so that we generally
370 get lots of unrolling. */
374 for (i
= 3; i
>= 0; i
--)
375 while (factors
[i
].count
--)
377 if (temp
* factors
[i
].factor
< MAX_UNROLLED_INSNS
)
379 unroll_number
*= factors
[i
].factor
;
380 temp
*= factors
[i
].factor
;
386 /* If we couldn't find any factors, then unroll as in the normal
388 if (unroll_number
== 1)
390 if (loop_dump_stream
)
391 fprintf (loop_dump_stream
,
392 "Loop unrolling: No factors found.\n");
395 unroll_type
= UNROLL_MODULO
;
399 /* Default case, calculate number of times to unroll loop based on its
401 if (unroll_number
== 1)
403 if (8 * insn_count
< MAX_UNROLLED_INSNS
)
405 else if (4 * insn_count
< MAX_UNROLLED_INSNS
)
410 unroll_type
= UNROLL_NAIVE
;
413 /* Now we know how many times to unroll the loop. */
415 if (loop_dump_stream
)
416 fprintf (loop_dump_stream
,
417 "Unrolling loop %d times.\n", unroll_number
);
420 if (unroll_type
== UNROLL_COMPLETELY
|| unroll_type
== UNROLL_MODULO
)
422 /* Loops of these types should never start with a jump down to
423 the exit condition test. For now, check for this case just to
424 be sure. UNROLL_NAIVE loops can be of this form, this case is
427 while (GET_CODE (insn
) != CODE_LABEL
&& GET_CODE (insn
) != JUMP_INSN
)
428 insn
= NEXT_INSN (insn
);
429 if (GET_CODE (insn
) == JUMP_INSN
)
433 if (unroll_type
== UNROLL_COMPLETELY
)
435 /* Completely unrolling the loop: Delete the compare and branch at
436 the end (the last two instructions). This delete must done at the
437 very end of loop unrolling, to avoid problems with calls to
438 back_branch_in_range_p, which is called by find_splittable_regs.
439 All increments of splittable bivs/givs are changed to load constant
442 copy_start
= loop_start
;
444 /* Set insert_before to the instruction immediately after the JUMP_INSN
445 (or BARRIER), so that any NOTEs between the JUMP_INSN and the end of
446 the loop will be correctly handled by copy_loop_body. */
447 insert_before
= NEXT_INSN (last_loop_insn
);
449 /* Set copy_end to the insn before the jump at the end of the loop. */
450 if (GET_CODE (last_loop_insn
) == BARRIER
)
451 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
452 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
455 /* The instruction immediately before the JUMP_INSN is a compare
456 instruction which we do not want to copy. */
457 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
459 /* The instruction immediately before the JUMP_INSN may not be the
460 compare, so we must copy it. */
461 copy_end
= PREV_INSN (last_loop_insn
);
466 /* We currently can't unroll a loop if it doesn't end with a
467 JUMP_INSN. There would need to be a mechanism that recognizes
468 this case, and then inserts a jump after each loop body, which
469 jumps to after the last loop body. */
470 if (loop_dump_stream
)
471 fprintf (loop_dump_stream
,
472 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
476 else if (unroll_type
== UNROLL_MODULO
)
478 /* Partially unrolling the loop: The compare and branch at the end
479 (the last two instructions) must remain. Don't copy the compare
480 and branch instructions at the end of the loop. Insert the unrolled
481 code immediately before the compare/branch at the end so that the
482 code will fall through to them as before. */
484 copy_start
= loop_start
;
486 /* Set insert_before to the jump insn at the end of the loop.
487 Set copy_end to before the jump insn at the end of the loop. */
488 if (GET_CODE (last_loop_insn
) == BARRIER
)
490 insert_before
= PREV_INSN (last_loop_insn
);
491 copy_end
= PREV_INSN (insert_before
);
493 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
496 /* The instruction immediately before the JUMP_INSN is a compare
497 instruction which we do not want to copy or delete. */
498 insert_before
= PREV_INSN (last_loop_insn
);
499 copy_end
= PREV_INSN (insert_before
);
501 /* The instruction immediately before the JUMP_INSN may not be the
502 compare, so we must copy it. */
503 insert_before
= last_loop_insn
;
504 copy_end
= PREV_INSN (last_loop_insn
);
509 /* We currently can't unroll a loop if it doesn't end with a
510 JUMP_INSN. There would need to be a mechanism that recognizes
511 this case, and then inserts a jump after each loop body, which
512 jumps to after the last loop body. */
513 if (loop_dump_stream
)
514 fprintf (loop_dump_stream
,
515 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
521 /* Normal case: Must copy the compare and branch instructions at the
524 if (GET_CODE (last_loop_insn
) == BARRIER
)
526 /* Loop ends with an unconditional jump and a barrier.
527 Handle this like above, don't copy jump and barrier.
528 This is not strictly necessary, but doing so prevents generating
529 unconditional jumps to an immediately following label.
531 This will be corrected below if the target of this jump is
532 not the start_label. */
534 insert_before
= PREV_INSN (last_loop_insn
);
535 copy_end
= PREV_INSN (insert_before
);
537 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
539 /* Set insert_before to immediately after the JUMP_INSN, so that
540 NOTEs at the end of the loop will be correctly handled by
542 insert_before
= NEXT_INSN (last_loop_insn
);
543 copy_end
= last_loop_insn
;
547 /* We currently can't unroll a loop if it doesn't end with a
548 JUMP_INSN. There would need to be a mechanism that recognizes
549 this case, and then inserts a jump after each loop body, which
550 jumps to after the last loop body. */
551 if (loop_dump_stream
)
552 fprintf (loop_dump_stream
,
553 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
557 /* If copying exit test branches because they can not be eliminated,
558 then must convert the fall through case of the branch to a jump past
559 the end of the loop. Create a label to emit after the loop and save
560 it for later use. Do not use the label after the loop, if any, since
561 it might be used by insns outside the loop, or there might be insns
562 added before it later by final_[bg]iv_value which must be after
563 the real exit label. */
564 exit_label
= gen_label_rtx ();
567 while (GET_CODE (insn
) != CODE_LABEL
&& GET_CODE (insn
) != JUMP_INSN
)
568 insn
= NEXT_INSN (insn
);
570 if (GET_CODE (insn
) == JUMP_INSN
)
572 /* The loop starts with a jump down to the exit condition test.
573 Start copying the loop after the barrier following this
575 copy_start
= NEXT_INSN (insn
);
577 /* Splitting induction variables doesn't work when the loop is
578 entered via a jump to the bottom, because then we end up doing
579 a comparison against a new register for a split variable, but
580 we did not execute the set insn for the new register because
581 it was skipped over. */
582 splitting_not_safe
= 1;
583 if (loop_dump_stream
)
584 fprintf (loop_dump_stream
,
585 "Splitting not safe, because loop not entered at top.\n");
588 copy_start
= loop_start
;
591 /* This should always be the first label in the loop. */
592 start_label
= NEXT_INSN (copy_start
);
593 /* There may be a line number note and/or a loop continue note here. */
594 while (GET_CODE (start_label
) == NOTE
)
595 start_label
= NEXT_INSN (start_label
);
596 if (GET_CODE (start_label
) != CODE_LABEL
)
598 /* This can happen as a result of jump threading. If the first insns in
599 the loop test the same condition as the loop's backward jump, or the
600 opposite condition, then the backward jump will be modified to point
601 to elsewhere, and the loop's start label is deleted.
603 This case currently can not be handled by the loop unrolling code. */
605 if (loop_dump_stream
)
606 fprintf (loop_dump_stream
,
607 "Unrolling failure: unknown insns between BEG note and loop label.\n");
610 if (LABEL_NAME (start_label
))
612 /* The jump optimization pass must have combined the original start label
613 with a named label for a goto. We can't unroll this case because
614 jumps which go to the named label must be handled differently than
615 jumps to the loop start, and it is impossible to differentiate them
617 if (loop_dump_stream
)
618 fprintf (loop_dump_stream
,
619 "Unrolling failure: loop start label is gone\n");
623 if (unroll_type
== UNROLL_NAIVE
624 && GET_CODE (last_loop_insn
) == BARRIER
625 && start_label
!= JUMP_LABEL (PREV_INSN (last_loop_insn
)))
627 /* In this case, we must copy the jump and barrier, because they will
628 not be converted to jumps to an immediately following label. */
630 insert_before
= NEXT_INSN (last_loop_insn
);
631 copy_end
= last_loop_insn
;
634 /* Allocate a translation table for the labels and insn numbers.
635 They will be filled in as we copy the insns in the loop. */
637 max_labelno
= max_label_num ();
638 max_insnno
= get_max_uid ();
640 map
= (struct inline_remap
*) alloca (sizeof (struct inline_remap
));
642 map
->integrating
= 0;
644 /* Allocate the label map. */
648 map
->label_map
= (rtx
*) alloca (max_labelno
* sizeof (rtx
));
650 local_label
= (char *) alloca (max_labelno
);
651 bzero (local_label
, max_labelno
);
656 /* Search the loop and mark all local labels, i.e. the ones which have to
657 be distinct labels when copied. For all labels which might be
658 non-local, set their label_map entries to point to themselves.
659 If they happen to be local their label_map entries will be overwritten
660 before the loop body is copied. The label_map entries for local labels
661 will be set to a different value each time the loop body is copied. */
663 for (insn
= copy_start
; insn
!= loop_end
; insn
= NEXT_INSN (insn
))
665 if (GET_CODE (insn
) == CODE_LABEL
)
666 local_label
[CODE_LABEL_NUMBER (insn
)] = 1;
667 else if (GET_CODE (insn
) == JUMP_INSN
)
669 if (JUMP_LABEL (insn
))
670 map
->label_map
[CODE_LABEL_NUMBER (JUMP_LABEL (insn
))]
672 else if (GET_CODE (PATTERN (insn
)) == ADDR_VEC
673 || GET_CODE (PATTERN (insn
)) == ADDR_DIFF_VEC
)
675 rtx pat
= PATTERN (insn
);
676 int diff_vec_p
= GET_CODE (PATTERN (insn
)) == ADDR_DIFF_VEC
;
677 int len
= XVECLEN (pat
, diff_vec_p
);
680 for (i
= 0; i
< len
; i
++)
682 label
= XEXP (XVECEXP (pat
, diff_vec_p
, i
), 0);
683 map
->label_map
[CODE_LABEL_NUMBER (label
)] = label
;
689 /* Allocate space for the insn map. */
691 map
->insn_map
= (rtx
*) alloca (max_insnno
* sizeof (rtx
));
693 /* Set this to zero, to indicate that we are doing loop unrolling,
694 not function inlining. */
695 map
->inline_target
= 0;
697 /* The register and constant maps depend on the number of registers
698 present, so the final maps can't be created until after
699 find_splittable_regs is called. However, they are needed for
700 preconditioning, so we create temporary maps when preconditioning
703 /* The preconditioning code may allocate two new pseudo registers. */
704 maxregnum
= max_reg_num ();
706 /* Allocate and zero out the splittable_regs and addr_combined_regs
707 arrays. These must be zeroed here because they will be used if
708 loop preconditioning is performed, and must be zero for that case.
710 It is safe to do this here, since the extra registers created by the
711 preconditioning code and find_splittable_regs will never be used
712 to access the splittable_regs[] and addr_combined_regs[] arrays. */
714 splittable_regs
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
715 bzero ((char *) splittable_regs
, maxregnum
* sizeof (rtx
));
716 splittable_regs_updates
= (int *) alloca (maxregnum
* sizeof (int));
717 bzero ((char *) splittable_regs_updates
, maxregnum
* sizeof (int));
719 = (struct induction
**) alloca (maxregnum
* sizeof (struct induction
*));
720 bzero ((char *) addr_combined_regs
, maxregnum
* sizeof (struct induction
*));
721 local_regno
= (char *) alloca (maxregnum
);
722 bzero (local_regno
, maxregnum
);
724 /* Mark all local registers, i.e. the ones which are referenced only
727 int copy_start_luid
= INSN_LUID (copy_start
);
728 int copy_end_luid
= INSN_LUID (copy_end
);
730 /* If a register is used in the jump insn, we must not duplicate it
731 since it will also be used outside the loop. */
732 if (GET_CODE (copy_end
) == JUMP_INSN
)
735 for (j
= FIRST_PSEUDO_REGISTER
; j
< maxregnum
; ++j
)
736 if (regno_first_uid
[j
] > 0 && regno_first_uid
[j
] <= max_uid_for_loop
737 && uid_luid
[regno_first_uid
[j
]] >= copy_start_luid
738 && regno_last_uid
[j
] > 0 && regno_last_uid
[j
] <= max_uid_for_loop
739 && uid_luid
[regno_last_uid
[j
]] <= copy_end_luid
)
743 /* If this loop requires exit tests when unrolled, check to see if we
744 can precondition the loop so as to make the exit tests unnecessary.
745 Just like variable splitting, this is not safe if the loop is entered
746 via a jump to the bottom. Also, can not do this if no strength
747 reduce info, because precondition_loop_p uses this info. */
749 /* Must copy the loop body for preconditioning before the following
750 find_splittable_regs call since that will emit insns which need to
751 be after the preconditioned loop copies, but immediately before the
752 unrolled loop copies. */
754 /* Also, it is not safe to split induction variables for the preconditioned
755 copies of the loop body. If we split induction variables, then the code
756 assumes that each induction variable can be represented as a function
757 of its initial value and the loop iteration number. This is not true
758 in this case, because the last preconditioned copy of the loop body
759 could be any iteration from the first up to the `unroll_number-1'th,
760 depending on the initial value of the iteration variable. Therefore
761 we can not split induction variables here, because we can not calculate
762 their value. Hence, this code must occur before find_splittable_regs
765 if (unroll_type
== UNROLL_NAIVE
&& ! splitting_not_safe
&& strength_reduce_p
)
767 rtx initial_value
, final_value
, increment
;
769 if (precondition_loop_p (&initial_value
, &final_value
, &increment
,
770 loop_start
, loop_end
))
772 register rtx diff
, temp
;
773 enum machine_mode mode
;
775 int abs_inc
, neg_inc
;
777 map
->reg_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
779 map
->const_equiv_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
780 map
->const_age_map
= (unsigned *) alloca (maxregnum
781 * sizeof (unsigned));
782 map
->const_equiv_map_size
= maxregnum
;
783 global_const_equiv_map
= map
->const_equiv_map
;
784 global_const_equiv_map_size
= maxregnum
;
786 init_reg_map (map
, maxregnum
);
788 /* Limit loop unrolling to 4, since this will make 7 copies of
790 if (unroll_number
> 4)
793 /* Save the absolute value of the increment, and also whether or
794 not it is negative. */
796 abs_inc
= INTVAL (increment
);
805 /* Decide what mode to do these calculations in. Choose the larger
806 of final_value's mode and initial_value's mode, or a full-word if
807 both are constants. */
808 mode
= GET_MODE (final_value
);
809 if (mode
== VOIDmode
)
811 mode
= GET_MODE (initial_value
);
812 if (mode
== VOIDmode
)
815 else if (mode
!= GET_MODE (initial_value
)
816 && (GET_MODE_SIZE (mode
)
817 < GET_MODE_SIZE (GET_MODE (initial_value
))))
818 mode
= GET_MODE (initial_value
);
820 /* Calculate the difference between the final and initial values.
821 Final value may be a (plus (reg x) (const_int 1)) rtx.
822 Let the following cse pass simplify this if initial value is
825 We must copy the final and initial values here to avoid
826 improperly shared rtl. */
828 diff
= expand_binop (mode
, sub_optab
, copy_rtx (final_value
),
829 copy_rtx (initial_value
), NULL_RTX
, 0,
832 /* Now calculate (diff % (unroll * abs (increment))) by using an
834 diff
= expand_binop (GET_MODE (diff
), and_optab
, diff
,
835 GEN_INT (unroll_number
* abs_inc
- 1),
836 NULL_RTX
, 0, OPTAB_LIB_WIDEN
);
838 /* Now emit a sequence of branches to jump to the proper precond
841 labels
= (rtx
*) alloca (sizeof (rtx
) * unroll_number
);
842 for (i
= 0; i
< unroll_number
; i
++)
843 labels
[i
] = gen_label_rtx ();
845 /* Assuming the unroll_number is 4, and the increment is 2, then
846 for a negative increment: for a positive increment:
847 diff = 0,1 precond 0 diff = 0,7 precond 0
848 diff = 2,3 precond 3 diff = 1,2 precond 1
849 diff = 4,5 precond 2 diff = 3,4 precond 2
850 diff = 6,7 precond 1 diff = 5,6 precond 3 */
852 /* We only need to emit (unroll_number - 1) branches here, the
853 last case just falls through to the following code. */
855 /* ??? This would give better code if we emitted a tree of branches
856 instead of the current linear list of branches. */
858 for (i
= 0; i
< unroll_number
- 1; i
++)
862 /* For negative increments, must invert the constant compared
863 against, except when comparing against zero. */
867 cmp_const
= unroll_number
- i
;
871 emit_cmp_insn (diff
, GEN_INT (abs_inc
* cmp_const
),
872 EQ
, NULL_RTX
, mode
, 0, 0);
875 emit_jump_insn (gen_beq (labels
[i
]));
877 emit_jump_insn (gen_bge (labels
[i
]));
879 emit_jump_insn (gen_ble (labels
[i
]));
880 JUMP_LABEL (get_last_insn ()) = labels
[i
];
881 LABEL_NUSES (labels
[i
])++;
884 /* If the increment is greater than one, then we need another branch,
885 to handle other cases equivalent to 0. */
887 /* ??? This should be merged into the code above somehow to help
888 simplify the code here, and reduce the number of branches emitted.
889 For the negative increment case, the branch here could easily
890 be merged with the `0' case branch above. For the positive
891 increment case, it is not clear how this can be simplified. */
898 cmp_const
= abs_inc
- 1;
900 cmp_const
= abs_inc
* (unroll_number
- 1) + 1;
902 emit_cmp_insn (diff
, GEN_INT (cmp_const
), EQ
, NULL_RTX
,
906 emit_jump_insn (gen_ble (labels
[0]));
908 emit_jump_insn (gen_bge (labels
[0]));
909 JUMP_LABEL (get_last_insn ()) = labels
[0];
910 LABEL_NUSES (labels
[0])++;
913 sequence
= gen_sequence ();
915 emit_insn_before (sequence
, loop_start
);
917 /* Only the last copy of the loop body here needs the exit
918 test, so set copy_end to exclude the compare/branch here,
919 and then reset it inside the loop when get to the last
922 if (GET_CODE (last_loop_insn
) == BARRIER
)
923 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
924 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
927 /* The immediately preceding insn is a compare which we do not
929 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
931 /* The immediately preceding insn may not be a compare, so we
933 copy_end
= PREV_INSN (last_loop_insn
);
939 for (i
= 1; i
< unroll_number
; i
++)
941 emit_label_after (labels
[unroll_number
- i
],
942 PREV_INSN (loop_start
));
944 bzero ((char *) map
->insn_map
, max_insnno
* sizeof (rtx
));
945 bzero ((char *) map
->const_equiv_map
, maxregnum
* sizeof (rtx
));
946 bzero ((char *) map
->const_age_map
,
947 maxregnum
* sizeof (unsigned));
950 for (j
= 0; j
< max_labelno
; j
++)
952 map
->label_map
[j
] = gen_label_rtx ();
954 for (j
= FIRST_PSEUDO_REGISTER
; j
< maxregnum
; j
++)
956 map
->reg_map
[j
] = gen_reg_rtx (GET_MODE (regno_reg_rtx
[j
]));
958 /* The last copy needs the compare/branch insns at the end,
959 so reset copy_end here if the loop ends with a conditional
962 if (i
== unroll_number
- 1)
964 if (GET_CODE (last_loop_insn
) == BARRIER
)
965 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
967 copy_end
= last_loop_insn
;
970 /* None of the copies are the `last_iteration', so just
971 pass zero for that parameter. */
972 copy_loop_body (copy_start
, copy_end
, map
, exit_label
, 0,
973 unroll_type
, start_label
, loop_end
,
974 loop_start
, copy_end
);
976 emit_label_after (labels
[0], PREV_INSN (loop_start
));
978 if (GET_CODE (last_loop_insn
) == BARRIER
)
980 insert_before
= PREV_INSN (last_loop_insn
);
981 copy_end
= PREV_INSN (insert_before
);
986 /* The immediately preceding insn is a compare which we do not
988 insert_before
= PREV_INSN (last_loop_insn
);
989 copy_end
= PREV_INSN (insert_before
);
991 /* The immediately preceding insn may not be a compare, so we
993 insert_before
= last_loop_insn
;
994 copy_end
= PREV_INSN (last_loop_insn
);
998 /* Set unroll type to MODULO now. */
999 unroll_type
= UNROLL_MODULO
;
1000 loop_preconditioned
= 1;
1004 /* If reach here, and the loop type is UNROLL_NAIVE, then don't unroll
1005 the loop unless all loops are being unrolled. */
1006 if (unroll_type
== UNROLL_NAIVE
&& ! flag_unroll_all_loops
)
1008 if (loop_dump_stream
)
1009 fprintf (loop_dump_stream
, "Unrolling failure: Naive unrolling not being done.\n");
1013 /* At this point, we are guaranteed to unroll the loop. */
1015 /* For each biv and giv, determine whether it can be safely split into
1016 a different variable for each unrolled copy of the loop body.
1017 We precalculate and save this info here, since computing it is
1020 Do this before deleting any instructions from the loop, so that
1021 back_branch_in_range_p will work correctly. */
1023 if (splitting_not_safe
)
1026 temp
= find_splittable_regs (unroll_type
, loop_start
, loop_end
,
1027 end_insert_before
, unroll_number
);
1029 /* find_splittable_regs may have created some new registers, so must
1030 reallocate the reg_map with the new larger size, and must realloc
1031 the constant maps also. */
1033 maxregnum
= max_reg_num ();
1034 map
->reg_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
1036 init_reg_map (map
, maxregnum
);
1038 /* Space is needed in some of the map for new registers, so new_maxregnum
1039 is an (over)estimate of how many registers will exist at the end. */
1040 new_maxregnum
= maxregnum
+ (temp
* unroll_number
* 2);
1042 /* Must realloc space for the constant maps, because the number of registers
1043 may have changed. */
1045 map
->const_equiv_map
= (rtx
*) alloca (new_maxregnum
* sizeof (rtx
));
1046 map
->const_age_map
= (unsigned *) alloca (new_maxregnum
* sizeof (unsigned));
1048 map
->const_equiv_map_size
= new_maxregnum
;
1049 global_const_equiv_map
= map
->const_equiv_map
;
1050 global_const_equiv_map_size
= new_maxregnum
;
1052 /* Search the list of bivs and givs to find ones which need to be remapped
1053 when split, and set their reg_map entry appropriately. */
1055 for (bl
= loop_iv_list
; bl
; bl
= bl
->next
)
1057 if (REGNO (bl
->biv
->src_reg
) != bl
->regno
)
1058 map
->reg_map
[bl
->regno
] = bl
->biv
->src_reg
;
1060 /* Currently, non-reduced/final-value givs are never split. */
1061 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
1062 if (REGNO (v
->src_reg
) != bl
->regno
)
1063 map
->reg_map
[REGNO (v
->dest_reg
)] = v
->src_reg
;
1067 /* If the loop is being partially unrolled, and the iteration variables
1068 are being split, and are being renamed for the split, then must fix up
1069 the compare/jump instruction at the end of the loop to refer to the new
1070 registers. This compare isn't copied, so the registers used in it
1071 will never be replaced if it isn't done here. */
1073 if (unroll_type
== UNROLL_MODULO
)
1075 insn
= NEXT_INSN (copy_end
);
1076 if (GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == JUMP_INSN
)
1077 PATTERN (insn
) = remap_split_bivs (PATTERN (insn
));
1080 /* For unroll_number - 1 times, make a copy of each instruction
1081 between copy_start and copy_end, and insert these new instructions
1082 before the end of the loop. */
1084 for (i
= 0; i
< unroll_number
; i
++)
1086 bzero ((char *) map
->insn_map
, max_insnno
* sizeof (rtx
));
1087 bzero ((char *) map
->const_equiv_map
, new_maxregnum
* sizeof (rtx
));
1088 bzero ((char *) map
->const_age_map
, new_maxregnum
* sizeof (unsigned));
1091 for (j
= 0; j
< max_labelno
; j
++)
1093 map
->label_map
[j
] = gen_label_rtx ();
1095 for (j
= FIRST_PSEUDO_REGISTER
; j
< maxregnum
; j
++)
1097 map
->reg_map
[j
] = gen_reg_rtx (GET_MODE (regno_reg_rtx
[j
]));
1099 /* If loop starts with a branch to the test, then fix it so that
1100 it points to the test of the first unrolled copy of the loop. */
1101 if (i
== 0 && loop_start
!= copy_start
)
1103 insn
= PREV_INSN (copy_start
);
1104 pattern
= PATTERN (insn
);
1106 tem
= map
->label_map
[CODE_LABEL_NUMBER
1107 (XEXP (SET_SRC (pattern
), 0))];
1108 SET_SRC (pattern
) = gen_rtx (LABEL_REF
, VOIDmode
, tem
);
1110 /* Set the jump label so that it can be used by later loop unrolling
1112 JUMP_LABEL (insn
) = tem
;
1113 LABEL_NUSES (tem
)++;
1116 copy_loop_body (copy_start
, copy_end
, map
, exit_label
,
1117 i
== unroll_number
- 1, unroll_type
, start_label
,
1118 loop_end
, insert_before
, insert_before
);
1121 /* Before deleting any insns, emit a CODE_LABEL immediately after the last
1122 insn to be deleted. This prevents any runaway delete_insn call from
1123 more insns that it should, as it always stops at a CODE_LABEL. */
1125 /* Delete the compare and branch at the end of the loop if completely
1126 unrolling the loop. Deleting the backward branch at the end also
1127 deletes the code label at the start of the loop. This is done at
1128 the very end to avoid problems with back_branch_in_range_p. */
1130 if (unroll_type
== UNROLL_COMPLETELY
)
1131 safety_label
= emit_label_after (gen_label_rtx (), last_loop_insn
);
1133 safety_label
= emit_label_after (gen_label_rtx (), copy_end
);
1135 /* Delete all of the original loop instructions. Don't delete the
1136 LOOP_BEG note, or the first code label in the loop. */
1138 insn
= NEXT_INSN (copy_start
);
1139 while (insn
!= safety_label
)
1141 if (insn
!= start_label
)
1142 insn
= delete_insn (insn
);
1144 insn
= NEXT_INSN (insn
);
1147 /* Can now delete the 'safety' label emitted to protect us from runaway
1148 delete_insn calls. */
1149 if (INSN_DELETED_P (safety_label
))
1151 delete_insn (safety_label
);
1153 /* If exit_label exists, emit it after the loop. Doing the emit here
1154 forces it to have a higher INSN_UID than any insn in the unrolled loop.
1155 This is needed so that mostly_true_jump in reorg.c will treat jumps
1156 to this loop end label correctly, i.e. predict that they are usually
1159 emit_label_after (exit_label
, loop_end
);
1162 /* Return true if the loop can be safely, and profitably, preconditioned
1163 so that the unrolled copies of the loop body don't need exit tests.
1165 This only works if final_value, initial_value and increment can be
1166 determined, and if increment is a constant power of 2.
1167 If increment is not a power of 2, then the preconditioning modulo
1168 operation would require a real modulo instead of a boolean AND, and this
1169 is not considered `profitable'. */
1171 /* ??? If the loop is known to be executed very many times, or the machine
1172 has a very cheap divide instruction, then preconditioning is a win even
1173 when the increment is not a power of 2. Use RTX_COST to compute
1174 whether divide is cheap. */
1177 precondition_loop_p (initial_value
, final_value
, increment
, loop_start
,
1179 rtx
*initial_value
, *final_value
, *increment
;
1180 rtx loop_start
, loop_end
;
1183 if (loop_n_iterations
> 0)
1185 *initial_value
= const0_rtx
;
1186 *increment
= const1_rtx
;
1187 *final_value
= GEN_INT (loop_n_iterations
);
1189 if (loop_dump_stream
)
1190 fprintf (loop_dump_stream
,
1191 "Preconditioning: Success, number of iterations known, %d.\n",
1196 if (loop_initial_value
== 0)
1198 if (loop_dump_stream
)
1199 fprintf (loop_dump_stream
,
1200 "Preconditioning: Could not find initial value.\n");
1203 else if (loop_increment
== 0)
1205 if (loop_dump_stream
)
1206 fprintf (loop_dump_stream
,
1207 "Preconditioning: Could not find increment value.\n");
1210 else if (GET_CODE (loop_increment
) != CONST_INT
)
1212 if (loop_dump_stream
)
1213 fprintf (loop_dump_stream
,
1214 "Preconditioning: Increment not a constant.\n");
1217 else if ((exact_log2 (INTVAL (loop_increment
)) < 0)
1218 && (exact_log2 (- INTVAL (loop_increment
)) < 0))
1220 if (loop_dump_stream
)
1221 fprintf (loop_dump_stream
,
1222 "Preconditioning: Increment not a constant power of 2.\n");
1226 /* Unsigned_compare and compare_dir can be ignored here, since they do
1227 not matter for preconditioning. */
1229 if (loop_final_value
== 0)
1231 if (loop_dump_stream
)
1232 fprintf (loop_dump_stream
,
1233 "Preconditioning: EQ comparison loop.\n");
1237 /* Must ensure that final_value is invariant, so call invariant_p to
1238 check. Before doing so, must check regno against max_reg_before_loop
1239 to make sure that the register is in the range covered by invariant_p.
1240 If it isn't, then it is most likely a biv/giv which by definition are
1242 if ((GET_CODE (loop_final_value
) == REG
1243 && REGNO (loop_final_value
) >= max_reg_before_loop
)
1244 || (GET_CODE (loop_final_value
) == PLUS
1245 && REGNO (XEXP (loop_final_value
, 0)) >= max_reg_before_loop
)
1246 || ! invariant_p (loop_final_value
))
1248 if (loop_dump_stream
)
1249 fprintf (loop_dump_stream
,
1250 "Preconditioning: Final value not invariant.\n");
1254 /* Fail for floating point values, since the caller of this function
1255 does not have code to deal with them. */
1256 if (GET_MODE_CLASS (GET_MODE (loop_final_value
)) == MODE_FLOAT
1257 || GET_MODE_CLASS (GET_MODE (loop_initial_value
)) == MODE_FLOAT
)
1259 if (loop_dump_stream
)
1260 fprintf (loop_dump_stream
,
1261 "Preconditioning: Floating point final or initial value.\n");
1265 /* Now set initial_value to be the iteration_var, since that may be a
1266 simpler expression, and is guaranteed to be correct if all of the
1267 above tests succeed.
1269 We can not use the initial_value as calculated, because it will be
1270 one too small for loops of the form "while (i-- > 0)". We can not
1271 emit code before the loop_skip_over insns to fix this problem as this
1272 will then give a number one too large for loops of the form
1275 Note that all loops that reach here are entered at the top, because
1276 this function is not called if the loop starts with a jump. */
1278 /* Fail if loop_iteration_var is not live before loop_start, since we need
1279 to test its value in the preconditioning code. */
1281 if (uid_luid
[regno_first_uid
[REGNO (loop_iteration_var
)]]
1282 > INSN_LUID (loop_start
))
1284 if (loop_dump_stream
)
1285 fprintf (loop_dump_stream
,
1286 "Preconditioning: Iteration var not live before loop start.\n");
1290 *initial_value
= loop_iteration_var
;
1291 *increment
= loop_increment
;
1292 *final_value
= loop_final_value
;
1295 if (loop_dump_stream
)
1296 fprintf (loop_dump_stream
, "Preconditioning: Successful.\n");
1301 /* All pseudo-registers must be mapped to themselves. Two hard registers
1302 must be mapped, VIRTUAL_STACK_VARS_REGNUM and VIRTUAL_INCOMING_ARGS_
1303 REGNUM, to avoid function-inlining specific conversions of these
1304 registers. All other hard regs can not be mapped because they may be
1309 init_reg_map (map
, maxregnum
)
1310 struct inline_remap
*map
;
1315 for (i
= maxregnum
- 1; i
> LAST_VIRTUAL_REGISTER
; i
--)
1316 map
->reg_map
[i
] = regno_reg_rtx
[i
];
1317 /* Just clear the rest of the entries. */
1318 for (i
= LAST_VIRTUAL_REGISTER
; i
>= 0; i
--)
1319 map
->reg_map
[i
] = 0;
1321 map
->reg_map
[VIRTUAL_STACK_VARS_REGNUM
]
1322 = regno_reg_rtx
[VIRTUAL_STACK_VARS_REGNUM
];
1323 map
->reg_map
[VIRTUAL_INCOMING_ARGS_REGNUM
]
1324 = regno_reg_rtx
[VIRTUAL_INCOMING_ARGS_REGNUM
];
1327 /* Strength-reduction will often emit code for optimized biv/givs which
1328 calculates their value in a temporary register, and then copies the result
1329 to the iv. This procedure reconstructs the pattern computing the iv;
1330 verifying that all operands are of the proper form.
1332 The return value is the amount that the giv is incremented by. */
1335 calculate_giv_inc (pattern
, src_insn
, regno
)
1336 rtx pattern
, src_insn
;
1340 rtx increment_total
= 0;
1344 /* Verify that we have an increment insn here. First check for a plus
1345 as the set source. */
1346 if (GET_CODE (SET_SRC (pattern
)) != PLUS
)
1348 /* SR sometimes computes the new giv value in a temp, then copies it
1350 src_insn
= PREV_INSN (src_insn
);
1351 pattern
= PATTERN (src_insn
);
1352 if (GET_CODE (SET_SRC (pattern
)) != PLUS
)
1355 /* The last insn emitted is not needed, so delete it to avoid confusing
1356 the second cse pass. This insn sets the giv unnecessarily. */
1357 delete_insn (get_last_insn ());
1360 /* Verify that we have a constant as the second operand of the plus. */
1361 increment
= XEXP (SET_SRC (pattern
), 1);
1362 if (GET_CODE (increment
) != CONST_INT
)
1364 /* SR sometimes puts the constant in a register, especially if it is
1365 too big to be an add immed operand. */
1366 src_insn
= PREV_INSN (src_insn
);
1367 increment
= SET_SRC (PATTERN (src_insn
));
1369 /* SR may have used LO_SUM to compute the constant if it is too large
1370 for a load immed operand. In this case, the constant is in operand
1371 one of the LO_SUM rtx. */
1372 if (GET_CODE (increment
) == LO_SUM
)
1373 increment
= XEXP (increment
, 1);
1374 else if (GET_CODE (increment
) == IOR
)
1376 /* The rs6000 port loads some constants with IOR. */
1377 rtx second_part
= XEXP (increment
, 1);
1379 src_insn
= PREV_INSN (src_insn
);
1380 increment
= SET_SRC (PATTERN (src_insn
));
1381 /* Don't need the last insn anymore. */
1382 delete_insn (get_last_insn ());
1384 if (GET_CODE (second_part
) != CONST_INT
1385 || GET_CODE (increment
) != CONST_INT
)
1388 increment
= GEN_INT (INTVAL (increment
) | INTVAL (second_part
));
1391 if (GET_CODE (increment
) != CONST_INT
)
1394 /* The insn loading the constant into a register is no longer needed,
1396 delete_insn (get_last_insn ());
1399 if (increment_total
)
1400 increment_total
= GEN_INT (INTVAL (increment_total
) + INTVAL (increment
));
1402 increment_total
= increment
;
1404 /* Check that the source register is the same as the register we expected
1405 to see as the source. If not, something is seriously wrong. */
1406 if (GET_CODE (XEXP (SET_SRC (pattern
), 0)) != REG
1407 || REGNO (XEXP (SET_SRC (pattern
), 0)) != regno
)
1409 /* Some machines (e.g. the romp), may emit two add instructions for
1410 certain constants, so lets try looking for another add immediately
1411 before this one if we have only seen one add insn so far. */
1417 src_insn
= PREV_INSN (src_insn
);
1418 pattern
= PATTERN (src_insn
);
1420 delete_insn (get_last_insn ());
1428 return increment_total
;
1431 /* Copy REG_NOTES, except for insn references, because not all insn_map
1432 entries are valid yet. We do need to copy registers now though, because
1433 the reg_map entries can change during copying. */
1436 initial_reg_note_copy (notes
, map
)
1438 struct inline_remap
*map
;
1445 copy
= rtx_alloc (GET_CODE (notes
));
1446 PUT_MODE (copy
, GET_MODE (notes
));
1448 if (GET_CODE (notes
) == EXPR_LIST
)
1449 XEXP (copy
, 0) = copy_rtx_and_substitute (XEXP (notes
, 0), map
);
1450 else if (GET_CODE (notes
) == INSN_LIST
)
1451 /* Don't substitute for these yet. */
1452 XEXP (copy
, 0) = XEXP (notes
, 0);
1456 XEXP (copy
, 1) = initial_reg_note_copy (XEXP (notes
, 1), map
);
1461 /* Fixup insn references in copied REG_NOTES. */
1464 final_reg_note_copy (notes
, map
)
1466 struct inline_remap
*map
;
1470 for (note
= notes
; note
; note
= XEXP (note
, 1))
1471 if (GET_CODE (note
) == INSN_LIST
)
1472 XEXP (note
, 0) = map
->insn_map
[INSN_UID (XEXP (note
, 0))];
1475 /* Copy each instruction in the loop, substituting from map as appropriate.
1476 This is very similar to a loop in expand_inline_function. */
1479 copy_loop_body (copy_start
, copy_end
, map
, exit_label
, last_iteration
,
1480 unroll_type
, start_label
, loop_end
, insert_before
,
1482 rtx copy_start
, copy_end
;
1483 struct inline_remap
*map
;
1486 enum unroll_types unroll_type
;
1487 rtx start_label
, loop_end
, insert_before
, copy_notes_from
;
1491 int dest_reg_was_split
, i
;
1493 rtx final_label
= 0;
1494 rtx giv_inc
, giv_dest_reg
, giv_src_reg
;
1496 /* If this isn't the last iteration, then map any references to the
1497 start_label to final_label. Final label will then be emitted immediately
1498 after the end of this loop body if it was ever used.
1500 If this is the last iteration, then map references to the start_label
1502 if (! last_iteration
)
1504 final_label
= gen_label_rtx ();
1505 map
->label_map
[CODE_LABEL_NUMBER (start_label
)] = final_label
;
1508 map
->label_map
[CODE_LABEL_NUMBER (start_label
)] = start_label
;
1515 insn
= NEXT_INSN (insn
);
1517 map
->orig_asm_operands_vector
= 0;
1519 switch (GET_CODE (insn
))
1522 pattern
= PATTERN (insn
);
1526 /* Check to see if this is a giv that has been combined with
1527 some split address givs. (Combined in the sense that
1528 `combine_givs' in loop.c has put two givs in the same register.)
1529 In this case, we must search all givs based on the same biv to
1530 find the address givs. Then split the address givs.
1531 Do this before splitting the giv, since that may map the
1532 SET_DEST to a new register. */
1534 if (GET_CODE (pattern
) == SET
1535 && GET_CODE (SET_DEST (pattern
)) == REG
1536 && addr_combined_regs
[REGNO (SET_DEST (pattern
))])
1538 struct iv_class
*bl
;
1539 struct induction
*v
, *tv
;
1540 int regno
= REGNO (SET_DEST (pattern
));
1542 v
= addr_combined_regs
[REGNO (SET_DEST (pattern
))];
1543 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
1545 /* Although the giv_inc amount is not needed here, we must call
1546 calculate_giv_inc here since it might try to delete the
1547 last insn emitted. If we wait until later to call it,
1548 we might accidentally delete insns generated immediately
1549 below by emit_unrolled_add. */
1551 giv_inc
= calculate_giv_inc (pattern
, insn
, regno
);
1553 /* Now find all address giv's that were combined with this
1555 for (tv
= bl
->giv
; tv
; tv
= tv
->next_iv
)
1556 if (tv
->giv_type
== DEST_ADDR
&& tv
->same
== v
)
1558 int this_giv_inc
= INTVAL (giv_inc
);
1560 /* Scale this_giv_inc if the multiplicative factors of
1561 the two givs are different. */
1562 if (tv
->mult_val
!= v
->mult_val
)
1563 this_giv_inc
= (this_giv_inc
/ INTVAL (v
->mult_val
)
1564 * INTVAL (tv
->mult_val
));
1566 tv
->dest_reg
= plus_constant (tv
->dest_reg
, this_giv_inc
);
1567 *tv
->location
= tv
->dest_reg
;
1569 if (last_iteration
&& unroll_type
!= UNROLL_COMPLETELY
)
1571 /* Must emit an insn to increment the split address
1572 giv. Add in the const_adjust field in case there
1573 was a constant eliminated from the address. */
1574 rtx value
, dest_reg
;
1576 /* tv->dest_reg will be either a bare register,
1577 or else a register plus a constant. */
1578 if (GET_CODE (tv
->dest_reg
) == REG
)
1579 dest_reg
= tv
->dest_reg
;
1581 dest_reg
= XEXP (tv
->dest_reg
, 0);
1583 /* Check for shared address givs, and avoid
1584 incrementing the shared psuedo reg more than
1586 if (! tv
->same_insn
)
1588 /* tv->dest_reg may actually be a (PLUS (REG)
1589 (CONST)) here, so we must call plus_constant
1590 to add the const_adjust amount before calling
1591 emit_unrolled_add below. */
1592 value
= plus_constant (tv
->dest_reg
,
1595 /* The constant could be too large for an add
1596 immediate, so can't directly emit an insn
1598 emit_unrolled_add (dest_reg
, XEXP (value
, 0),
1602 /* Reset the giv to be just the register again, in case
1603 it is used after the set we have just emitted.
1604 We must subtract the const_adjust factor added in
1606 tv
->dest_reg
= plus_constant (dest_reg
,
1607 - tv
->const_adjust
);
1608 *tv
->location
= tv
->dest_reg
;
1613 /* If this is a setting of a splittable variable, then determine
1614 how to split the variable, create a new set based on this split,
1615 and set up the reg_map so that later uses of the variable will
1616 use the new split variable. */
1618 dest_reg_was_split
= 0;
1620 if (GET_CODE (pattern
) == SET
1621 && GET_CODE (SET_DEST (pattern
)) == REG
1622 && splittable_regs
[REGNO (SET_DEST (pattern
))])
1624 int regno
= REGNO (SET_DEST (pattern
));
1626 dest_reg_was_split
= 1;
1628 /* Compute the increment value for the giv, if it wasn't
1629 already computed above. */
1632 giv_inc
= calculate_giv_inc (pattern
, insn
, regno
);
1633 giv_dest_reg
= SET_DEST (pattern
);
1634 giv_src_reg
= SET_DEST (pattern
);
1636 if (unroll_type
== UNROLL_COMPLETELY
)
1638 /* Completely unrolling the loop. Set the induction
1639 variable to a known constant value. */
1641 /* The value in splittable_regs may be an invariant
1642 value, so we must use plus_constant here. */
1643 splittable_regs
[regno
]
1644 = plus_constant (splittable_regs
[regno
], INTVAL (giv_inc
));
1646 if (GET_CODE (splittable_regs
[regno
]) == PLUS
)
1648 giv_src_reg
= XEXP (splittable_regs
[regno
], 0);
1649 giv_inc
= XEXP (splittable_regs
[regno
], 1);
1653 /* The splittable_regs value must be a REG or a
1654 CONST_INT, so put the entire value in the giv_src_reg
1656 giv_src_reg
= splittable_regs
[regno
];
1657 giv_inc
= const0_rtx
;
1662 /* Partially unrolling loop. Create a new pseudo
1663 register for the iteration variable, and set it to
1664 be a constant plus the original register. Except
1665 on the last iteration, when the result has to
1666 go back into the original iteration var register. */
1668 /* Handle bivs which must be mapped to a new register
1669 when split. This happens for bivs which need their
1670 final value set before loop entry. The new register
1671 for the biv was stored in the biv's first struct
1672 induction entry by find_splittable_regs. */
1674 if (regno
< max_reg_before_loop
1675 && reg_iv_type
[regno
] == BASIC_INDUCT
)
1677 giv_src_reg
= reg_biv_class
[regno
]->biv
->src_reg
;
1678 giv_dest_reg
= giv_src_reg
;
1682 /* If non-reduced/final-value givs were split, then
1683 this would have to remap those givs also. See
1684 find_splittable_regs. */
1687 splittable_regs
[regno
]
1688 = GEN_INT (INTVAL (giv_inc
)
1689 + INTVAL (splittable_regs
[regno
]));
1690 giv_inc
= splittable_regs
[regno
];
1692 /* Now split the induction variable by changing the dest
1693 of this insn to a new register, and setting its
1694 reg_map entry to point to this new register.
1696 If this is the last iteration, and this is the last insn
1697 that will update the iv, then reuse the original dest,
1698 to ensure that the iv will have the proper value when
1699 the loop exits or repeats.
1701 Using splittable_regs_updates here like this is safe,
1702 because it can only be greater than one if all
1703 instructions modifying the iv are always executed in
1706 if (! last_iteration
1707 || (splittable_regs_updates
[regno
]-- != 1))
1709 tem
= gen_reg_rtx (GET_MODE (giv_src_reg
));
1711 map
->reg_map
[regno
] = tem
;
1714 map
->reg_map
[regno
] = giv_src_reg
;
1717 /* The constant being added could be too large for an add
1718 immediate, so can't directly emit an insn here. */
1719 emit_unrolled_add (giv_dest_reg
, giv_src_reg
, giv_inc
);
1720 copy
= get_last_insn ();
1721 pattern
= PATTERN (copy
);
1725 pattern
= copy_rtx_and_substitute (pattern
, map
);
1726 copy
= emit_insn (pattern
);
1728 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
1731 /* If this insn is setting CC0, it may need to look at
1732 the insn that uses CC0 to see what type of insn it is.
1733 In that case, the call to recog via validate_change will
1734 fail. So don't substitute constants here. Instead,
1735 do it when we emit the following insn.
1737 For example, see the pyr.md file. That machine has signed and
1738 unsigned compares. The compare patterns must check the
1739 following branch insn to see which what kind of compare to
1742 If the previous insn set CC0, substitute constants on it as
1744 if (sets_cc0_p (copy
) != 0)
1749 try_constants (cc0_insn
, map
);
1751 try_constants (copy
, map
);
1754 try_constants (copy
, map
);
1757 /* Make split induction variable constants `permanent' since we
1758 know there are no backward branches across iteration variable
1759 settings which would invalidate this. */
1760 if (dest_reg_was_split
)
1762 int regno
= REGNO (SET_DEST (pattern
));
1764 if (regno
< map
->const_equiv_map_size
1765 && map
->const_age_map
[regno
] == map
->const_age
)
1766 map
->const_age_map
[regno
] = -1;
1771 pattern
= copy_rtx_and_substitute (PATTERN (insn
), map
);
1772 copy
= emit_jump_insn (pattern
);
1773 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
1775 if (JUMP_LABEL (insn
) == start_label
&& insn
== copy_end
1776 && ! last_iteration
)
1778 /* This is a branch to the beginning of the loop; this is the
1779 last insn being copied; and this is not the last iteration.
1780 In this case, we want to change the original fall through
1781 case to be a branch past the end of the loop, and the
1782 original jump label case to fall_through. */
1784 if (invert_exp (pattern
, copy
))
1786 if (! redirect_exp (&pattern
,
1787 map
->label_map
[CODE_LABEL_NUMBER
1788 (JUMP_LABEL (insn
))],
1795 rtx lab
= gen_label_rtx ();
1796 /* Can't do it by reversing the jump (probably becasue we
1797 couln't reverse the conditions), so emit a new
1798 jump_insn after COPY, and redirect the jump around
1800 jmp
= emit_jump_insn_after (gen_jump (exit_label
), copy
);
1801 jmp
= emit_barrier_after (jmp
);
1802 emit_label_after (lab
, jmp
);
1803 LABEL_NUSES (lab
) = 0;
1804 if (! redirect_exp (&pattern
,
1805 map
->label_map
[CODE_LABEL_NUMBER
1806 (JUMP_LABEL (insn
))],
1814 try_constants (cc0_insn
, map
);
1817 try_constants (copy
, map
);
1819 /* Set the jump label of COPY correctly to avoid problems with
1820 later passes of unroll_loop, if INSN had jump label set. */
1821 if (JUMP_LABEL (insn
))
1825 /* Can't use the label_map for every insn, since this may be
1826 the backward branch, and hence the label was not mapped. */
1827 if (GET_CODE (pattern
) == SET
)
1829 tem
= SET_SRC (pattern
);
1830 if (GET_CODE (tem
) == LABEL_REF
)
1831 label
= XEXP (tem
, 0);
1832 else if (GET_CODE (tem
) == IF_THEN_ELSE
)
1834 if (XEXP (tem
, 1) != pc_rtx
)
1835 label
= XEXP (XEXP (tem
, 1), 0);
1837 label
= XEXP (XEXP (tem
, 2), 0);
1841 if (label
&& GET_CODE (label
) == CODE_LABEL
)
1842 JUMP_LABEL (copy
) = label
;
1845 /* An unrecognizable jump insn, probably the entry jump
1846 for a switch statement. This label must have been mapped,
1847 so just use the label_map to get the new jump label. */
1849 = map
->label_map
[CODE_LABEL_NUMBER (JUMP_LABEL (insn
))];
1852 /* If this is a non-local jump, then must increase the label
1853 use count so that the label will not be deleted when the
1854 original jump is deleted. */
1855 LABEL_NUSES (JUMP_LABEL (copy
))++;
1857 else if (GET_CODE (PATTERN (copy
)) == ADDR_VEC
1858 || GET_CODE (PATTERN (copy
)) == ADDR_DIFF_VEC
)
1860 rtx pat
= PATTERN (copy
);
1861 int diff_vec_p
= GET_CODE (pat
) == ADDR_DIFF_VEC
;
1862 int len
= XVECLEN (pat
, diff_vec_p
);
1865 for (i
= 0; i
< len
; i
++)
1866 LABEL_NUSES (XEXP (XVECEXP (pat
, diff_vec_p
, i
), 0))++;
1869 /* If this used to be a conditional jump insn but whose branch
1870 direction is now known, we must do something special. */
1871 if (condjump_p (insn
) && !simplejump_p (insn
) && map
->last_pc_value
)
1874 /* The previous insn set cc0 for us. So delete it. */
1875 delete_insn (PREV_INSN (copy
));
1878 /* If this is now a no-op, delete it. */
1879 if (map
->last_pc_value
== pc_rtx
)
1881 /* Don't let delete_insn delete the label referenced here,
1882 because we might possibly need it later for some other
1883 instruction in the loop. */
1884 if (JUMP_LABEL (copy
))
1885 LABEL_NUSES (JUMP_LABEL (copy
))++;
1887 if (JUMP_LABEL (copy
))
1888 LABEL_NUSES (JUMP_LABEL (copy
))--;
1892 /* Otherwise, this is unconditional jump so we must put a
1893 BARRIER after it. We could do some dead code elimination
1894 here, but jump.c will do it just as well. */
1900 pattern
= copy_rtx_and_substitute (PATTERN (insn
), map
);
1901 copy
= emit_call_insn (pattern
);
1902 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
1904 /* Because the USAGE information potentially contains objects other
1905 than hard registers, we need to copy it. */
1906 CALL_INSN_FUNCTION_USAGE (copy
) =
1907 copy_rtx_and_substitute (CALL_INSN_FUNCTION_USAGE (insn
), map
);
1911 try_constants (cc0_insn
, map
);
1914 try_constants (copy
, map
);
1916 /* Be lazy and assume CALL_INSNs clobber all hard registers. */
1917 for (i
= 0; i
< FIRST_PSEUDO_REGISTER
; i
++)
1918 map
->const_equiv_map
[i
] = 0;
1922 /* If this is the loop start label, then we don't need to emit a
1923 copy of this label since no one will use it. */
1925 if (insn
!= start_label
)
1927 copy
= emit_label (map
->label_map
[CODE_LABEL_NUMBER (insn
)]);
1933 copy
= emit_barrier ();
1937 /* VTOP notes are valid only before the loop exit test. If placed
1938 anywhere else, loop may generate bad code. */
1940 if (NOTE_LINE_NUMBER (insn
) != NOTE_INSN_DELETED
1941 && (NOTE_LINE_NUMBER (insn
) != NOTE_INSN_LOOP_VTOP
1942 || (last_iteration
&& unroll_type
!= UNROLL_COMPLETELY
)))
1943 copy
= emit_note (NOTE_SOURCE_FILE (insn
),
1944 NOTE_LINE_NUMBER (insn
));
1954 map
->insn_map
[INSN_UID (insn
)] = copy
;
1956 while (insn
!= copy_end
);
1958 /* Now finish coping the REG_NOTES. */
1962 insn
= NEXT_INSN (insn
);
1963 if ((GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == JUMP_INSN
1964 || GET_CODE (insn
) == CALL_INSN
)
1965 && map
->insn_map
[INSN_UID (insn
)])
1966 final_reg_note_copy (REG_NOTES (map
->insn_map
[INSN_UID (insn
)]), map
);
1968 while (insn
!= copy_end
);
1970 /* There may be notes between copy_notes_from and loop_end. Emit a copy of
1971 each of these notes here, since there may be some important ones, such as
1972 NOTE_INSN_BLOCK_END notes, in this group. We don't do this on the last
1973 iteration, because the original notes won't be deleted.
1975 We can't use insert_before here, because when from preconditioning,
1976 insert_before points before the loop. We can't use copy_end, because
1977 there may be insns already inserted after it (which we don't want to
1978 copy) when not from preconditioning code. */
1980 if (! last_iteration
)
1982 for (insn
= copy_notes_from
; insn
!= loop_end
; insn
= NEXT_INSN (insn
))
1984 if (GET_CODE (insn
) == NOTE
1985 && NOTE_LINE_NUMBER (insn
) != NOTE_INSN_DELETED
)
1986 emit_note (NOTE_SOURCE_FILE (insn
), NOTE_LINE_NUMBER (insn
));
1990 if (final_label
&& LABEL_NUSES (final_label
) > 0)
1991 emit_label (final_label
);
1993 tem
= gen_sequence ();
1995 emit_insn_before (tem
, insert_before
);
1998 /* Emit an insn, using the expand_binop to ensure that a valid insn is
1999 emitted. This will correctly handle the case where the increment value
2000 won't fit in the immediate field of a PLUS insns. */
2003 emit_unrolled_add (dest_reg
, src_reg
, increment
)
2004 rtx dest_reg
, src_reg
, increment
;
2008 result
= expand_binop (GET_MODE (dest_reg
), add_optab
, src_reg
, increment
,
2009 dest_reg
, 0, OPTAB_LIB_WIDEN
);
2011 if (dest_reg
!= result
)
2012 emit_move_insn (dest_reg
, result
);
2015 /* Searches the insns between INSN and LOOP_END. Returns 1 if there
2016 is a backward branch in that range that branches to somewhere between
2017 LOOP_START and INSN. Returns 0 otherwise. */
2019 /* ??? This is quadratic algorithm. Could be rewritten to be linear.
2020 In practice, this is not a problem, because this function is seldom called,
2021 and uses a negligible amount of CPU time on average. */
2024 back_branch_in_range_p (insn
, loop_start
, loop_end
)
2026 rtx loop_start
, loop_end
;
2028 rtx p
, q
, target_insn
;
2030 /* Stop before we get to the backward branch at the end of the loop. */
2031 loop_end
= prev_nonnote_insn (loop_end
);
2032 if (GET_CODE (loop_end
) == BARRIER
)
2033 loop_end
= PREV_INSN (loop_end
);
2035 /* Check in case insn has been deleted, search forward for first non
2036 deleted insn following it. */
2037 while (INSN_DELETED_P (insn
))
2038 insn
= NEXT_INSN (insn
);
2040 /* Check for the case where insn is the last insn in the loop. */
2041 if (insn
== loop_end
)
2044 for (p
= NEXT_INSN (insn
); p
!= loop_end
; p
= NEXT_INSN (p
))
2046 if (GET_CODE (p
) == JUMP_INSN
)
2048 target_insn
= JUMP_LABEL (p
);
2050 /* Search from loop_start to insn, to see if one of them is
2051 the target_insn. We can't use INSN_LUID comparisons here,
2052 since insn may not have an LUID entry. */
2053 for (q
= loop_start
; q
!= insn
; q
= NEXT_INSN (q
))
2054 if (q
== target_insn
)
2062 /* Try to generate the simplest rtx for the expression
2063 (PLUS (MULT mult1 mult2) add1). This is used to calculate the initial
2067 fold_rtx_mult_add (mult1
, mult2
, add1
, mode
)
2068 rtx mult1
, mult2
, add1
;
2069 enum machine_mode mode
;
2074 /* The modes must all be the same. This should always be true. For now,
2075 check to make sure. */
2076 if ((GET_MODE (mult1
) != mode
&& GET_MODE (mult1
) != VOIDmode
)
2077 || (GET_MODE (mult2
) != mode
&& GET_MODE (mult2
) != VOIDmode
)
2078 || (GET_MODE (add1
) != mode
&& GET_MODE (add1
) != VOIDmode
))
2081 /* Ensure that if at least one of mult1/mult2 are constant, then mult2
2082 will be a constant. */
2083 if (GET_CODE (mult1
) == CONST_INT
)
2090 mult_res
= simplify_binary_operation (MULT
, mode
, mult1
, mult2
);
2092 mult_res
= gen_rtx (MULT
, mode
, mult1
, mult2
);
2094 /* Again, put the constant second. */
2095 if (GET_CODE (add1
) == CONST_INT
)
2102 result
= simplify_binary_operation (PLUS
, mode
, add1
, mult_res
);
2104 result
= gen_rtx (PLUS
, mode
, add1
, mult_res
);
2109 /* Searches the list of induction struct's for the biv BL, to try to calculate
2110 the total increment value for one iteration of the loop as a constant.
2112 Returns the increment value as an rtx, simplified as much as possible,
2113 if it can be calculated. Otherwise, returns 0. */
2116 biv_total_increment (bl
, loop_start
, loop_end
)
2117 struct iv_class
*bl
;
2118 rtx loop_start
, loop_end
;
2120 struct induction
*v
;
2123 /* For increment, must check every instruction that sets it. Each
2124 instruction must be executed only once each time through the loop.
2125 To verify this, we check that the the insn is always executed, and that
2126 there are no backward branches after the insn that branch to before it.
2127 Also, the insn must have a mult_val of one (to make sure it really is
2130 result
= const0_rtx
;
2131 for (v
= bl
->biv
; v
; v
= v
->next_iv
)
2133 if (v
->always_computable
&& v
->mult_val
== const1_rtx
2134 && ! back_branch_in_range_p (v
->insn
, loop_start
, loop_end
))
2135 result
= fold_rtx_mult_add (result
, const1_rtx
, v
->add_val
, v
->mode
);
2143 /* Determine the initial value of the iteration variable, and the amount
2144 that it is incremented each loop. Use the tables constructed by
2145 the strength reduction pass to calculate these values.
2147 Initial_value and/or increment are set to zero if their values could not
2151 iteration_info (iteration_var
, initial_value
, increment
, loop_start
, loop_end
)
2152 rtx iteration_var
, *initial_value
, *increment
;
2153 rtx loop_start
, loop_end
;
2155 struct iv_class
*bl
;
2156 struct induction
*v
, *b
;
2158 /* Clear the result values, in case no answer can be found. */
2162 /* The iteration variable can be either a giv or a biv. Check to see
2163 which it is, and compute the variable's initial value, and increment
2164 value if possible. */
2166 /* If this is a new register, can't handle it since we don't have any
2167 reg_iv_type entry for it. */
2168 if (REGNO (iteration_var
) >= max_reg_before_loop
)
2170 if (loop_dump_stream
)
2171 fprintf (loop_dump_stream
,
2172 "Loop unrolling: No reg_iv_type entry for iteration var.\n");
2175 /* Reject iteration variables larger than the host long size, since they
2176 could result in a number of iterations greater than the range of our
2177 `unsigned long' variable loop_n_iterations. */
2178 else if (GET_MODE_BITSIZE (GET_MODE (iteration_var
)) > HOST_BITS_PER_LONG
)
2180 if (loop_dump_stream
)
2181 fprintf (loop_dump_stream
,
2182 "Loop unrolling: Iteration var rejected because mode larger than host long.\n");
2185 else if (GET_MODE_CLASS (GET_MODE (iteration_var
)) != MODE_INT
)
2187 if (loop_dump_stream
)
2188 fprintf (loop_dump_stream
,
2189 "Loop unrolling: Iteration var not an integer.\n");
2192 else if (reg_iv_type
[REGNO (iteration_var
)] == BASIC_INDUCT
)
2194 /* Grab initial value, only useful if it is a constant. */
2195 bl
= reg_biv_class
[REGNO (iteration_var
)];
2196 *initial_value
= bl
->initial_value
;
2198 *increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2200 else if (reg_iv_type
[REGNO (iteration_var
)] == GENERAL_INDUCT
)
2203 /* ??? The code below does not work because the incorrect number of
2204 iterations is calculated when the biv is incremented after the giv
2205 is set (which is the usual case). This can probably be accounted
2206 for by biasing the initial_value by subtracting the amount of the
2207 increment that occurs between the giv set and the giv test. However,
2208 a giv as an iterator is very rare, so it does not seem worthwhile
2210 /* ??? An example failure is: i = 6; do {;} while (i++ < 9). */
2211 if (loop_dump_stream
)
2212 fprintf (loop_dump_stream
,
2213 "Loop unrolling: Giv iterators are not handled.\n");
2216 /* Initial value is mult_val times the biv's initial value plus
2217 add_val. Only useful if it is a constant. */
2218 v
= reg_iv_info
[REGNO (iteration_var
)];
2219 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
2220 *initial_value
= fold_rtx_mult_add (v
->mult_val
, bl
->initial_value
,
2221 v
->add_val
, v
->mode
);
2223 /* Increment value is mult_val times the increment value of the biv. */
2225 *increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2227 *increment
= fold_rtx_mult_add (v
->mult_val
, *increment
, const0_rtx
,
2233 if (loop_dump_stream
)
2234 fprintf (loop_dump_stream
,
2235 "Loop unrolling: Not basic or general induction var.\n");
2240 /* Calculate the approximate final value of the iteration variable
2241 which has an loop exit test with code COMPARISON_CODE and comparison value
2242 of COMPARISON_VALUE. Also returns an indication of whether the comparison
2243 was signed or unsigned, and the direction of the comparison. This info is
2244 needed to calculate the number of loop iterations. */
2247 approx_final_value (comparison_code
, comparison_value
, unsigned_p
, compare_dir
)
2248 enum rtx_code comparison_code
;
2249 rtx comparison_value
;
2253 /* Calculate the final value of the induction variable.
2254 The exact final value depends on the branch operator, and increment sign.
2255 This is only an approximate value. It will be wrong if the iteration
2256 variable is not incremented by one each time through the loop, and
2257 approx final value - start value % increment != 0. */
2260 switch (comparison_code
)
2266 return plus_constant (comparison_value
, 1);
2271 return plus_constant (comparison_value
, -1);
2273 /* Can not calculate a final value for this case. */
2280 return comparison_value
;
2286 return comparison_value
;
2289 return comparison_value
;
2295 /* For each biv and giv, determine whether it can be safely split into
2296 a different variable for each unrolled copy of the loop body. If it
2297 is safe to split, then indicate that by saving some useful info
2298 in the splittable_regs array.
2300 If the loop is being completely unrolled, then splittable_regs will hold
2301 the current value of the induction variable while the loop is unrolled.
2302 It must be set to the initial value of the induction variable here.
2303 Otherwise, splittable_regs will hold the difference between the current
2304 value of the induction variable and the value the induction variable had
2305 at the top of the loop. It must be set to the value 0 here.
2307 Returns the total number of instructions that set registers that are
2310 /* ?? If the loop is only unrolled twice, then most of the restrictions to
2311 constant values are unnecessary, since we can easily calculate increment
2312 values in this case even if nothing is constant. The increment value
2313 should not involve a multiply however. */
2315 /* ?? Even if the biv/giv increment values aren't constant, it may still
2316 be beneficial to split the variable if the loop is only unrolled a few
2317 times, since multiplies by small integers (1,2,3,4) are very cheap. */
2320 find_splittable_regs (unroll_type
, loop_start
, loop_end
, end_insert_before
,
2322 enum unroll_types unroll_type
;
2323 rtx loop_start
, loop_end
;
2324 rtx end_insert_before
;
2327 struct iv_class
*bl
;
2328 struct induction
*v
;
2330 rtx biv_final_value
;
2334 for (bl
= loop_iv_list
; bl
; bl
= bl
->next
)
2336 /* Biv_total_increment must return a constant value,
2337 otherwise we can not calculate the split values. */
2339 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2340 if (! increment
|| GET_CODE (increment
) != CONST_INT
)
2343 /* The loop must be unrolled completely, or else have a known number
2344 of iterations and only one exit, or else the biv must be dead
2345 outside the loop, or else the final value must be known. Otherwise,
2346 it is unsafe to split the biv since it may not have the proper
2347 value on loop exit. */
2349 /* loop_number_exit_labels is non-zero if the loop has an exit other than
2350 a fall through at the end. */
2353 biv_final_value
= 0;
2354 if (unroll_type
!= UNROLL_COMPLETELY
2355 && (loop_number_exit_labels
[uid_loop_num
[INSN_UID (loop_start
)]]
2356 || unroll_type
== UNROLL_NAIVE
)
2357 && (uid_luid
[regno_last_uid
[bl
->regno
]] >= INSN_LUID (loop_end
)
2359 || INSN_UID (bl
->init_insn
) >= max_uid_for_loop
2360 || (uid_luid
[regno_first_uid
[bl
->regno
]]
2361 < INSN_LUID (bl
->init_insn
))
2362 || reg_mentioned_p (bl
->biv
->dest_reg
, SET_SRC (bl
->init_set
)))
2363 && ! (biv_final_value
= final_biv_value (bl
, loop_start
, loop_end
)))
2366 /* If any of the insns setting the BIV don't do so with a simple
2367 PLUS, we don't know how to split it. */
2368 for (v
= bl
->biv
; biv_splittable
&& v
; v
= v
->next_iv
)
2369 if ((tem
= single_set (v
->insn
)) == 0
2370 || GET_CODE (SET_DEST (tem
)) != REG
2371 || REGNO (SET_DEST (tem
)) != bl
->regno
2372 || GET_CODE (SET_SRC (tem
)) != PLUS
)
2375 /* If final value is non-zero, then must emit an instruction which sets
2376 the value of the biv to the proper value. This is done after
2377 handling all of the givs, since some of them may need to use the
2378 biv's value in their initialization code. */
2380 /* This biv is splittable. If completely unrolling the loop, save
2381 the biv's initial value. Otherwise, save the constant zero. */
2383 if (biv_splittable
== 1)
2385 if (unroll_type
== UNROLL_COMPLETELY
)
2387 /* If the initial value of the biv is itself (i.e. it is too
2388 complicated for strength_reduce to compute), or is a hard
2389 register, then we must create a new pseudo reg to hold the
2390 initial value of the biv. */
2392 if (GET_CODE (bl
->initial_value
) == REG
2393 && (REGNO (bl
->initial_value
) == bl
->regno
2394 || REGNO (bl
->initial_value
) < FIRST_PSEUDO_REGISTER
))
2396 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2398 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2401 if (loop_dump_stream
)
2402 fprintf (loop_dump_stream
, "Biv %d initial value remapped to %d.\n",
2403 bl
->regno
, REGNO (tem
));
2405 splittable_regs
[bl
->regno
] = tem
;
2408 splittable_regs
[bl
->regno
] = bl
->initial_value
;
2411 splittable_regs
[bl
->regno
] = const0_rtx
;
2413 /* Save the number of instructions that modify the biv, so that
2414 we can treat the last one specially. */
2416 splittable_regs_updates
[bl
->regno
] = bl
->biv_count
;
2417 result
+= bl
->biv_count
;
2419 if (loop_dump_stream
)
2420 fprintf (loop_dump_stream
,
2421 "Biv %d safe to split.\n", bl
->regno
);
2424 /* Check every giv that depends on this biv to see whether it is
2425 splittable also. Even if the biv isn't splittable, givs which
2426 depend on it may be splittable if the biv is live outside the
2427 loop, and the givs aren't. */
2429 result
+= find_splittable_givs (bl
, unroll_type
, loop_start
, loop_end
,
2430 increment
, unroll_number
);
2432 /* If final value is non-zero, then must emit an instruction which sets
2433 the value of the biv to the proper value. This is done after
2434 handling all of the givs, since some of them may need to use the
2435 biv's value in their initialization code. */
2436 if (biv_final_value
)
2438 /* If the loop has multiple exits, emit the insns before the
2439 loop to ensure that it will always be executed no matter
2440 how the loop exits. Otherwise emit the insn after the loop,
2441 since this is slightly more efficient. */
2442 if (! loop_number_exit_labels
[uid_loop_num
[INSN_UID (loop_start
)]])
2443 emit_insn_before (gen_move_insn (bl
->biv
->src_reg
,
2448 /* Create a new register to hold the value of the biv, and then
2449 set the biv to its final value before the loop start. The biv
2450 is set to its final value before loop start to ensure that
2451 this insn will always be executed, no matter how the loop
2453 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2454 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2456 emit_insn_before (gen_move_insn (bl
->biv
->src_reg
,
2460 if (loop_dump_stream
)
2461 fprintf (loop_dump_stream
, "Biv %d mapped to %d for split.\n",
2462 REGNO (bl
->biv
->src_reg
), REGNO (tem
));
2464 /* Set up the mapping from the original biv register to the new
2466 bl
->biv
->src_reg
= tem
;
2473 /* For every giv based on the biv BL, check to determine whether it is
2474 splittable. This is a subroutine to find_splittable_regs ().
2476 Return the number of instructions that set splittable registers. */
2479 find_splittable_givs (bl
, unroll_type
, loop_start
, loop_end
, increment
,
2481 struct iv_class
*bl
;
2482 enum unroll_types unroll_type
;
2483 rtx loop_start
, loop_end
;
2487 struct induction
*v
, *v2
;
2492 /* Scan the list of givs, and set the same_insn field when there are
2493 multiple identical givs in the same insn. */
2494 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
2495 for (v2
= v
->next_iv
; v2
; v2
= v2
->next_iv
)
2496 if (v
->insn
== v2
->insn
&& rtx_equal_p (v
->new_reg
, v2
->new_reg
)
2500 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
2504 /* Only split the giv if it has already been reduced, or if the loop is
2505 being completely unrolled. */
2506 if (unroll_type
!= UNROLL_COMPLETELY
&& v
->ignore
)
2509 /* The giv can be split if the insn that sets the giv is executed once
2510 and only once on every iteration of the loop. */
2511 /* An address giv can always be split. v->insn is just a use not a set,
2512 and hence it does not matter whether it is always executed. All that
2513 matters is that all the biv increments are always executed, and we
2514 won't reach here if they aren't. */
2515 if (v
->giv_type
!= DEST_ADDR
2516 && (! v
->always_computable
2517 || back_branch_in_range_p (v
->insn
, loop_start
, loop_end
)))
2520 /* The giv increment value must be a constant. */
2521 giv_inc
= fold_rtx_mult_add (v
->mult_val
, increment
, const0_rtx
,
2523 if (! giv_inc
|| GET_CODE (giv_inc
) != CONST_INT
)
2526 /* The loop must be unrolled completely, or else have a known number of
2527 iterations and only one exit, or else the giv must be dead outside
2528 the loop, or else the final value of the giv must be known.
2529 Otherwise, it is not safe to split the giv since it may not have the
2530 proper value on loop exit. */
2532 /* The used outside loop test will fail for DEST_ADDR givs. They are
2533 never used outside the loop anyways, so it is always safe to split a
2537 if (unroll_type
!= UNROLL_COMPLETELY
2538 && (loop_number_exit_labels
[uid_loop_num
[INSN_UID (loop_start
)]]
2539 || unroll_type
== UNROLL_NAIVE
)
2540 && v
->giv_type
!= DEST_ADDR
2541 && ((regno_first_uid
[REGNO (v
->dest_reg
)] != INSN_UID (v
->insn
)
2542 /* Check for the case where the pseudo is set by a shift/add
2543 sequence, in which case the first insn setting the pseudo
2544 is the first insn of the shift/add sequence. */
2545 && (! (tem
= find_reg_note (v
->insn
, REG_RETVAL
, NULL_RTX
))
2546 || (regno_first_uid
[REGNO (v
->dest_reg
)]
2547 != INSN_UID (XEXP (tem
, 0)))))
2548 /* Line above always fails if INSN was moved by loop opt. */
2549 || (uid_luid
[regno_last_uid
[REGNO (v
->dest_reg
)]]
2550 >= INSN_LUID (loop_end
)))
2551 && ! (final_value
= v
->final_value
))
2555 /* Currently, non-reduced/final-value givs are never split. */
2556 /* Should emit insns after the loop if possible, as the biv final value
2559 /* If the final value is non-zero, and the giv has not been reduced,
2560 then must emit an instruction to set the final value. */
2561 if (final_value
&& !v
->new_reg
)
2563 /* Create a new register to hold the value of the giv, and then set
2564 the giv to its final value before the loop start. The giv is set
2565 to its final value before loop start to ensure that this insn
2566 will always be executed, no matter how we exit. */
2567 tem
= gen_reg_rtx (v
->mode
);
2568 emit_insn_before (gen_move_insn (tem
, v
->dest_reg
), loop_start
);
2569 emit_insn_before (gen_move_insn (v
->dest_reg
, final_value
),
2572 if (loop_dump_stream
)
2573 fprintf (loop_dump_stream
, "Giv %d mapped to %d for split.\n",
2574 REGNO (v
->dest_reg
), REGNO (tem
));
2580 /* This giv is splittable. If completely unrolling the loop, save the
2581 giv's initial value. Otherwise, save the constant zero for it. */
2583 if (unroll_type
== UNROLL_COMPLETELY
)
2585 /* It is not safe to use bl->initial_value here, because it may not
2586 be invariant. It is safe to use the initial value stored in
2587 the splittable_regs array if it is set. In rare cases, it won't
2588 be set, so then we do exactly the same thing as
2589 find_splittable_regs does to get a safe value. */
2590 rtx biv_initial_value
;
2592 if (splittable_regs
[bl
->regno
])
2593 biv_initial_value
= splittable_regs
[bl
->regno
];
2594 else if (GET_CODE (bl
->initial_value
) != REG
2595 || (REGNO (bl
->initial_value
) != bl
->regno
2596 && REGNO (bl
->initial_value
) >= FIRST_PSEUDO_REGISTER
))
2597 biv_initial_value
= bl
->initial_value
;
2600 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2602 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2604 biv_initial_value
= tem
;
2606 value
= fold_rtx_mult_add (v
->mult_val
, biv_initial_value
,
2607 v
->add_val
, v
->mode
);
2614 /* If a giv was combined with another giv, then we can only split
2615 this giv if the giv it was combined with was reduced. This
2616 is because the value of v->new_reg is meaningless in this
2618 if (v
->same
&& ! v
->same
->new_reg
)
2620 if (loop_dump_stream
)
2621 fprintf (loop_dump_stream
,
2622 "giv combined with unreduced giv not split.\n");
2625 /* If the giv is an address destination, it could be something other
2626 than a simple register, these have to be treated differently. */
2627 else if (v
->giv_type
== DEST_REG
)
2629 /* If value is not a constant, register, or register plus
2630 constant, then compute its value into a register before
2631 loop start. This prevents invalid rtx sharing, and should
2632 generate better code. We can use bl->initial_value here
2633 instead of splittable_regs[bl->regno] because this code
2634 is going before the loop start. */
2635 if (unroll_type
== UNROLL_COMPLETELY
2636 && GET_CODE (value
) != CONST_INT
2637 && GET_CODE (value
) != REG
2638 && (GET_CODE (value
) != PLUS
2639 || GET_CODE (XEXP (value
, 0)) != REG
2640 || GET_CODE (XEXP (value
, 1)) != CONST_INT
))
2642 rtx tem
= gen_reg_rtx (v
->mode
);
2643 emit_iv_add_mult (bl
->initial_value
, v
->mult_val
,
2644 v
->add_val
, tem
, loop_start
);
2648 splittable_regs
[REGNO (v
->new_reg
)] = value
;
2652 /* Splitting address givs is useful since it will often allow us
2653 to eliminate some increment insns for the base giv as
2656 /* If the addr giv is combined with a dest_reg giv, then all
2657 references to that dest reg will be remapped, which is NOT
2658 what we want for split addr regs. We always create a new
2659 register for the split addr giv, just to be safe. */
2661 /* ??? If there are multiple address givs which have been
2662 combined with the same dest_reg giv, then we may only need
2663 one new register for them. Pulling out constants below will
2664 catch some of the common cases of this. Currently, I leave
2665 the work of simplifying multiple address givs to the
2666 following cse pass. */
2668 /* As a special case, if we have multiple identical address givs
2669 within a single instruction, then we do use a single psuedo
2670 reg for both. This is necessary in case one is a match_dup
2673 v
->const_adjust
= 0;
2677 v
->dest_reg
= v
->same_insn
->dest_reg
;
2678 if (loop_dump_stream
)
2679 fprintf (loop_dump_stream
,
2680 "Sharing address givs in insn %d\n",
2681 INSN_UID (v
->insn
));
2683 else if (unroll_type
!= UNROLL_COMPLETELY
)
2685 /* If not completely unrolling the loop, then create a new
2686 register to hold the split value of the DEST_ADDR giv.
2687 Emit insn to initialize its value before loop start. */
2688 tem
= gen_reg_rtx (v
->mode
);
2690 /* If the address giv has a constant in its new_reg value,
2691 then this constant can be pulled out and put in value,
2692 instead of being part of the initialization code. */
2694 if (GET_CODE (v
->new_reg
) == PLUS
2695 && GET_CODE (XEXP (v
->new_reg
, 1)) == CONST_INT
)
2698 = plus_constant (tem
, INTVAL (XEXP (v
->new_reg
,1)));
2700 /* Only succeed if this will give valid addresses.
2701 Try to validate both the first and the last
2702 address resulting from loop unrolling, if
2703 one fails, then can't do const elim here. */
2704 if (memory_address_p (v
->mem_mode
, v
->dest_reg
)
2705 && memory_address_p (v
->mem_mode
,
2706 plus_constant (v
->dest_reg
,
2708 * (unroll_number
- 1))))
2710 /* Save the negative of the eliminated const, so
2711 that we can calculate the dest_reg's increment
2713 v
->const_adjust
= - INTVAL (XEXP (v
->new_reg
, 1));
2715 v
->new_reg
= XEXP (v
->new_reg
, 0);
2716 if (loop_dump_stream
)
2717 fprintf (loop_dump_stream
,
2718 "Eliminating constant from giv %d\n",
2727 /* If the address hasn't been checked for validity yet, do so
2728 now, and fail completely if either the first or the last
2729 unrolled copy of the address is not a valid address. */
2730 if (v
->dest_reg
== tem
2731 && (! memory_address_p (v
->mem_mode
, v
->dest_reg
)
2732 || ! memory_address_p (v
->mem_mode
,
2733 plus_constant (v
->dest_reg
,
2735 * (unroll_number
-1)))))
2737 if (loop_dump_stream
)
2738 fprintf (loop_dump_stream
,
2739 "Invalid address for giv at insn %d\n",
2740 INSN_UID (v
->insn
));
2744 /* To initialize the new register, just move the value of
2745 new_reg into it. This is not guaranteed to give a valid
2746 instruction on machines with complex addressing modes.
2747 If we can't recognize it, then delete it and emit insns
2748 to calculate the value from scratch. */
2749 emit_insn_before (gen_rtx (SET
, VOIDmode
, tem
,
2750 copy_rtx (v
->new_reg
)),
2752 if (recog_memoized (PREV_INSN (loop_start
)) < 0)
2756 /* We can't use bl->initial_value to compute the initial
2757 value, because the loop may have been preconditioned.
2758 We must calculate it from NEW_REG. Try using
2759 force_operand instead of emit_iv_add_mult. */
2760 delete_insn (PREV_INSN (loop_start
));
2763 ret
= force_operand (v
->new_reg
, tem
);
2765 emit_move_insn (tem
, ret
);
2766 sequence
= gen_sequence ();
2768 emit_insn_before (sequence
, loop_start
);
2770 if (loop_dump_stream
)
2771 fprintf (loop_dump_stream
,
2772 "Invalid init insn, rewritten.\n");
2777 v
->dest_reg
= value
;
2779 /* Check the resulting address for validity, and fail
2780 if the resulting address would be invalid. */
2781 if (! memory_address_p (v
->mem_mode
, v
->dest_reg
)
2782 || ! memory_address_p (v
->mem_mode
,
2783 plus_constant (v
->dest_reg
,
2785 (unroll_number
-1))))
2787 if (loop_dump_stream
)
2788 fprintf (loop_dump_stream
,
2789 "Invalid address for giv at insn %d\n",
2790 INSN_UID (v
->insn
));
2795 /* Store the value of dest_reg into the insn. This sharing
2796 will not be a problem as this insn will always be copied
2799 *v
->location
= v
->dest_reg
;
2801 /* If this address giv is combined with a dest reg giv, then
2802 save the base giv's induction pointer so that we will be
2803 able to handle this address giv properly. The base giv
2804 itself does not have to be splittable. */
2806 if (v
->same
&& v
->same
->giv_type
== DEST_REG
)
2807 addr_combined_regs
[REGNO (v
->same
->new_reg
)] = v
->same
;
2809 if (GET_CODE (v
->new_reg
) == REG
)
2811 /* This giv maybe hasn't been combined with any others.
2812 Make sure that it's giv is marked as splittable here. */
2814 splittable_regs
[REGNO (v
->new_reg
)] = value
;
2816 /* Make it appear to depend upon itself, so that the
2817 giv will be properly split in the main loop above. */
2821 addr_combined_regs
[REGNO (v
->new_reg
)] = v
;
2825 if (loop_dump_stream
)
2826 fprintf (loop_dump_stream
, "DEST_ADDR giv being split.\n");
2832 /* Currently, unreduced giv's can't be split. This is not too much
2833 of a problem since unreduced giv's are not live across loop
2834 iterations anyways. When unrolling a loop completely though,
2835 it makes sense to reduce&split givs when possible, as this will
2836 result in simpler instructions, and will not require that a reg
2837 be live across loop iterations. */
2839 splittable_regs
[REGNO (v
->dest_reg
)] = value
;
2840 fprintf (stderr
, "Giv %d at insn %d not reduced\n",
2841 REGNO (v
->dest_reg
), INSN_UID (v
->insn
));
2847 /* Givs are only updated once by definition. Mark it so if this is
2848 a splittable register. Don't need to do anything for address givs
2849 where this may not be a register. */
2851 if (GET_CODE (v
->new_reg
) == REG
)
2852 splittable_regs_updates
[REGNO (v
->new_reg
)] = 1;
2856 if (loop_dump_stream
)
2860 if (GET_CODE (v
->dest_reg
) == CONST_INT
)
2862 else if (GET_CODE (v
->dest_reg
) != REG
)
2863 regnum
= REGNO (XEXP (v
->dest_reg
, 0));
2865 regnum
= REGNO (v
->dest_reg
);
2866 fprintf (loop_dump_stream
, "Giv %d at insn %d safe to split.\n",
2867 regnum
, INSN_UID (v
->insn
));
2874 /* Try to prove that the register is dead after the loop exits. Trace every
2875 loop exit looking for an insn that will always be executed, which sets
2876 the register to some value, and appears before the first use of the register
2877 is found. If successful, then return 1, otherwise return 0. */
2879 /* ?? Could be made more intelligent in the handling of jumps, so that
2880 it can search past if statements and other similar structures. */
2883 reg_dead_after_loop (reg
, loop_start
, loop_end
)
2884 rtx reg
, loop_start
, loop_end
;
2890 /* HACK: Must also search the loop fall through exit, create a label_ref
2891 here which points to the loop_end, and append the loop_number_exit_labels
2893 label
= gen_rtx (LABEL_REF
, VOIDmode
, loop_end
);
2894 LABEL_NEXTREF (label
)
2895 = loop_number_exit_labels
[uid_loop_num
[INSN_UID (loop_start
)]];
2897 for ( ; label
; label
= LABEL_NEXTREF (label
))
2899 /* Succeed if find an insn which sets the biv or if reach end of
2900 function. Fail if find an insn that uses the biv, or if come to
2901 a conditional jump. */
2903 insn
= NEXT_INSN (XEXP (label
, 0));
2906 code
= GET_CODE (insn
);
2907 if (GET_RTX_CLASS (code
) == 'i')
2911 if (reg_referenced_p (reg
, PATTERN (insn
)))
2914 set
= single_set (insn
);
2915 if (set
&& rtx_equal_p (SET_DEST (set
), reg
))
2919 if (code
== JUMP_INSN
)
2921 if (GET_CODE (PATTERN (insn
)) == RETURN
)
2923 else if (! simplejump_p (insn
)
2924 /* Prevent infinite loop following infinite loops. */
2925 || jump_count
++ > 20)
2928 insn
= JUMP_LABEL (insn
);
2931 insn
= NEXT_INSN (insn
);
2935 /* Success, the register is dead on all loop exits. */
2939 /* Try to calculate the final value of the biv, the value it will have at
2940 the end of the loop. If we can do it, return that value. */
2943 final_biv_value (bl
, loop_start
, loop_end
)
2944 struct iv_class
*bl
;
2945 rtx loop_start
, loop_end
;
2949 /* ??? This only works for MODE_INT biv's. Reject all others for now. */
2951 if (GET_MODE_CLASS (bl
->biv
->mode
) != MODE_INT
)
2954 /* The final value for reversed bivs must be calculated differently than
2955 for ordinary bivs. In this case, there is already an insn after the
2956 loop which sets this biv's final value (if necessary), and there are
2957 no other loop exits, so we can return any value. */
2960 if (loop_dump_stream
)
2961 fprintf (loop_dump_stream
,
2962 "Final biv value for %d, reversed biv.\n", bl
->regno
);
2967 /* Try to calculate the final value as initial value + (number of iterations
2968 * increment). For this to work, increment must be invariant, the only
2969 exit from the loop must be the fall through at the bottom (otherwise
2970 it may not have its final value when the loop exits), and the initial
2971 value of the biv must be invariant. */
2973 if (loop_n_iterations
!= 0
2974 && ! loop_number_exit_labels
[uid_loop_num
[INSN_UID (loop_start
)]]
2975 && invariant_p (bl
->initial_value
))
2977 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2979 if (increment
&& invariant_p (increment
))
2981 /* Can calculate the loop exit value, emit insns after loop
2982 end to calculate this value into a temporary register in
2983 case it is needed later. */
2985 tem
= gen_reg_rtx (bl
->biv
->mode
);
2986 /* Make sure loop_end is not the last insn. */
2987 if (NEXT_INSN (loop_end
) == 0)
2988 emit_note_after (NOTE_INSN_DELETED
, loop_end
);
2989 emit_iv_add_mult (increment
, GEN_INT (loop_n_iterations
),
2990 bl
->initial_value
, tem
, NEXT_INSN (loop_end
));
2992 if (loop_dump_stream
)
2993 fprintf (loop_dump_stream
,
2994 "Final biv value for %d, calculated.\n", bl
->regno
);
3000 /* Check to see if the biv is dead at all loop exits. */
3001 if (reg_dead_after_loop (bl
->biv
->src_reg
, loop_start
, loop_end
))
3003 if (loop_dump_stream
)
3004 fprintf (loop_dump_stream
,
3005 "Final biv value for %d, biv dead after loop exit.\n",
3014 /* Try to calculate the final value of the giv, the value it will have at
3015 the end of the loop. If we can do it, return that value. */
3018 final_giv_value (v
, loop_start
, loop_end
)
3019 struct induction
*v
;
3020 rtx loop_start
, loop_end
;
3022 struct iv_class
*bl
;
3025 rtx insert_before
, seq
;
3027 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
3029 /* The final value for givs which depend on reversed bivs must be calculated
3030 differently than for ordinary givs. In this case, there is already an
3031 insn after the loop which sets this giv's final value (if necessary),
3032 and there are no other loop exits, so we can return any value. */
3035 if (loop_dump_stream
)
3036 fprintf (loop_dump_stream
,
3037 "Final giv value for %d, depends on reversed biv\n",
3038 REGNO (v
->dest_reg
));
3042 /* Try to calculate the final value as a function of the biv it depends
3043 upon. The only exit from the loop must be the fall through at the bottom
3044 (otherwise it may not have its final value when the loop exits). */
3046 /* ??? Can calculate the final giv value by subtracting off the
3047 extra biv increments times the giv's mult_val. The loop must have
3048 only one exit for this to work, but the loop iterations does not need
3051 if (loop_n_iterations
!= 0
3052 && ! loop_number_exit_labels
[uid_loop_num
[INSN_UID (loop_start
)]])
3054 /* ?? It is tempting to use the biv's value here since these insns will
3055 be put after the loop, and hence the biv will have its final value
3056 then. However, this fails if the biv is subsequently eliminated.
3057 Perhaps determine whether biv's are eliminable before trying to
3058 determine whether giv's are replaceable so that we can use the
3059 biv value here if it is not eliminable. */
3061 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
3063 if (increment
&& invariant_p (increment
))
3065 /* Can calculate the loop exit value of its biv as
3066 (loop_n_iterations * increment) + initial_value */
3068 /* The loop exit value of the giv is then
3069 (final_biv_value - extra increments) * mult_val + add_val.
3070 The extra increments are any increments to the biv which
3071 occur in the loop after the giv's value is calculated.
3072 We must search from the insn that sets the giv to the end
3073 of the loop to calculate this value. */
3075 insert_before
= NEXT_INSN (loop_end
);
3077 /* Put the final biv value in tem. */
3078 tem
= gen_reg_rtx (bl
->biv
->mode
);
3079 emit_iv_add_mult (increment
, GEN_INT (loop_n_iterations
),
3080 bl
->initial_value
, tem
, insert_before
);
3082 /* Subtract off extra increments as we find them. */
3083 for (insn
= NEXT_INSN (v
->insn
); insn
!= loop_end
;
3084 insn
= NEXT_INSN (insn
))
3086 struct induction
*biv
;
3088 for (biv
= bl
->biv
; biv
; biv
= biv
->next_iv
)
3089 if (biv
->insn
== insn
)
3092 tem
= expand_binop (GET_MODE (tem
), sub_optab
, tem
,
3093 biv
->add_val
, NULL_RTX
, 0,
3095 seq
= gen_sequence ();
3097 emit_insn_before (seq
, insert_before
);
3101 /* Now calculate the giv's final value. */
3102 emit_iv_add_mult (tem
, v
->mult_val
, v
->add_val
, tem
,
3105 if (loop_dump_stream
)
3106 fprintf (loop_dump_stream
,
3107 "Final giv value for %d, calc from biv's value.\n",
3108 REGNO (v
->dest_reg
));
3114 /* Replaceable giv's should never reach here. */
3118 /* Check to see if the biv is dead at all loop exits. */
3119 if (reg_dead_after_loop (v
->dest_reg
, loop_start
, loop_end
))
3121 if (loop_dump_stream
)
3122 fprintf (loop_dump_stream
,
3123 "Final giv value for %d, giv dead after loop exit.\n",
3124 REGNO (v
->dest_reg
));
3133 /* Calculate the number of loop iterations. Returns the exact number of loop
3134 iterations if it can be calculated, otherwise returns zero. */
3136 unsigned HOST_WIDE_INT
3137 loop_iterations (loop_start
, loop_end
)
3138 rtx loop_start
, loop_end
;
3140 rtx comparison
, comparison_value
;
3141 rtx iteration_var
, initial_value
, increment
, final_value
;
3142 enum rtx_code comparison_code
;
3145 int unsigned_compare
, compare_dir
, final_larger
;
3146 unsigned long tempu
;
3149 /* First find the iteration variable. If the last insn is a conditional
3150 branch, and the insn before tests a register value, make that the
3151 iteration variable. */
3153 loop_initial_value
= 0;
3155 loop_final_value
= 0;
3156 loop_iteration_var
= 0;
3158 /* We used to use pren_nonnote_insn here, but that fails because it might
3159 accidentally get the branch for a contained loop if the branch for this
3160 loop was deleted. We can only trust branches immediately before the
3162 last_loop_insn
= PREV_INSN (loop_end
);
3164 comparison
= get_condition_for_loop (last_loop_insn
);
3165 if (comparison
== 0)
3167 if (loop_dump_stream
)
3168 fprintf (loop_dump_stream
,
3169 "Loop unrolling: No final conditional branch found.\n");
3173 /* ??? Get_condition may switch position of induction variable and
3174 invariant register when it canonicalizes the comparison. */
3176 comparison_code
= GET_CODE (comparison
);
3177 iteration_var
= XEXP (comparison
, 0);
3178 comparison_value
= XEXP (comparison
, 1);
3180 if (GET_CODE (iteration_var
) != REG
)
3182 if (loop_dump_stream
)
3183 fprintf (loop_dump_stream
,
3184 "Loop unrolling: Comparison not against register.\n");
3188 /* Loop iterations is always called before any new registers are created
3189 now, so this should never occur. */
3191 if (REGNO (iteration_var
) >= max_reg_before_loop
)
3194 iteration_info (iteration_var
, &initial_value
, &increment
,
3195 loop_start
, loop_end
);
3196 if (initial_value
== 0)
3197 /* iteration_info already printed a message. */
3200 /* If the comparison value is an invariant register, then try to find
3201 its value from the insns before the start of the loop. */
3203 if (GET_CODE (comparison_value
) == REG
&& invariant_p (comparison_value
))
3207 for (insn
= PREV_INSN (loop_start
); insn
; insn
= PREV_INSN (insn
))
3209 if (GET_CODE (insn
) == CODE_LABEL
)
3212 else if (GET_RTX_CLASS (GET_CODE (insn
)) == 'i'
3213 && reg_set_p (comparison_value
, insn
))
3215 /* We found the last insn before the loop that sets the register.
3216 If it sets the entire register, and has a REG_EQUAL note,
3217 then use the value of the REG_EQUAL note. */
3218 if ((set
= single_set (insn
))
3219 && (SET_DEST (set
) == comparison_value
))
3221 rtx note
= find_reg_note (insn
, REG_EQUAL
, NULL_RTX
);
3223 /* Only use the REG_EQUAL note if it is a constant.
3224 Other things, divide in particular, will cause
3225 problems later if we use them. */
3226 if (note
&& GET_CODE (XEXP (note
, 0)) != EXPR_LIST
3227 && CONSTANT_P (XEXP (note
, 0)))
3228 comparison_value
= XEXP (note
, 0);
3235 final_value
= approx_final_value (comparison_code
, comparison_value
,
3236 &unsigned_compare
, &compare_dir
);
3238 /* Save the calculated values describing this loop's bounds, in case
3239 precondition_loop_p will need them later. These values can not be
3240 recalculated inside precondition_loop_p because strength reduction
3241 optimizations may obscure the loop's structure. */
3243 loop_iteration_var
= iteration_var
;
3244 loop_initial_value
= initial_value
;
3245 loop_increment
= increment
;
3246 loop_final_value
= final_value
;
3250 if (loop_dump_stream
)
3251 fprintf (loop_dump_stream
,
3252 "Loop unrolling: Increment value can't be calculated.\n");
3255 else if (GET_CODE (increment
) != CONST_INT
)
3257 if (loop_dump_stream
)
3258 fprintf (loop_dump_stream
,
3259 "Loop unrolling: Increment value not constant.\n");
3262 else if (GET_CODE (initial_value
) != CONST_INT
)
3264 if (loop_dump_stream
)
3265 fprintf (loop_dump_stream
,
3266 "Loop unrolling: Initial value not constant.\n");
3269 else if (final_value
== 0)
3271 if (loop_dump_stream
)
3272 fprintf (loop_dump_stream
,
3273 "Loop unrolling: EQ comparison loop.\n");
3276 else if (GET_CODE (final_value
) != CONST_INT
)
3278 if (loop_dump_stream
)
3279 fprintf (loop_dump_stream
,
3280 "Loop unrolling: Final value not constant.\n");
3284 /* ?? Final value and initial value do not have to be constants.
3285 Only their difference has to be constant. When the iteration variable
3286 is an array address, the final value and initial value might both
3287 be addresses with the same base but different constant offsets.
3288 Final value must be invariant for this to work.
3290 To do this, need some way to find the values of registers which are
3293 /* Final_larger is 1 if final larger, 0 if they are equal, otherwise -1. */
3294 if (unsigned_compare
)
3296 = ((unsigned HOST_WIDE_INT
) INTVAL (final_value
)
3297 > (unsigned HOST_WIDE_INT
) INTVAL (initial_value
))
3298 - ((unsigned HOST_WIDE_INT
) INTVAL (final_value
)
3299 < (unsigned HOST_WIDE_INT
) INTVAL (initial_value
));
3301 final_larger
= (INTVAL (final_value
) > INTVAL (initial_value
))
3302 - (INTVAL (final_value
) < INTVAL (initial_value
));
3304 if (INTVAL (increment
) > 0)
3306 else if (INTVAL (increment
) == 0)
3311 /* There are 27 different cases: compare_dir = -1, 0, 1;
3312 final_larger = -1, 0, 1; increment_dir = -1, 0, 1.
3313 There are 4 normal cases, 4 reverse cases (where the iteration variable
3314 will overflow before the loop exits), 4 infinite loop cases, and 15
3315 immediate exit (0 or 1 iteration depending on loop type) cases.
3316 Only try to optimize the normal cases. */
3318 /* (compare_dir/final_larger/increment_dir)
3319 Normal cases: (0/-1/-1), (0/1/1), (-1/-1/-1), (1/1/1)
3320 Reverse cases: (0/-1/1), (0/1/-1), (-1/-1/1), (1/1/-1)
3321 Infinite loops: (0/-1/0), (0/1/0), (-1/-1/0), (1/1/0)
3322 Immediate exit: (0/0/X), (-1/0/X), (-1/1/X), (1/0/X), (1/-1/X) */
3324 /* ?? If the meaning of reverse loops (where the iteration variable
3325 will overflow before the loop exits) is undefined, then could
3326 eliminate all of these special checks, and just always assume
3327 the loops are normal/immediate/infinite. Note that this means
3328 the sign of increment_dir does not have to be known. Also,
3329 since it does not really hurt if immediate exit loops or infinite loops
3330 are optimized, then that case could be ignored also, and hence all
3331 loops can be optimized.
3333 According to ANSI Spec, the reverse loop case result is undefined,
3334 because the action on overflow is undefined.
3336 See also the special test for NE loops below. */
3338 if (final_larger
== increment_dir
&& final_larger
!= 0
3339 && (final_larger
== compare_dir
|| compare_dir
== 0))
3344 if (loop_dump_stream
)
3345 fprintf (loop_dump_stream
,
3346 "Loop unrolling: Not normal loop.\n");
3350 /* Calculate the number of iterations, final_value is only an approximation,
3351 so correct for that. Note that tempu and loop_n_iterations are
3352 unsigned, because they can be as large as 2^n - 1. */
3354 i
= INTVAL (increment
);
3356 tempu
= INTVAL (final_value
) - INTVAL (initial_value
);
3359 tempu
= INTVAL (initial_value
) - INTVAL (final_value
);
3365 /* For NE tests, make sure that the iteration variable won't miss the
3366 final value. If tempu mod i is not zero, then the iteration variable
3367 will overflow before the loop exits, and we can not calculate the
3368 number of iterations. */
3369 if (compare_dir
== 0 && (tempu
% i
) != 0)
3372 return tempu
/ i
+ ((tempu
% i
) != 0);
3375 /* Replace uses of split bivs with their split psuedo register. This is
3376 for original instructions which remain after loop unrolling without
3380 remap_split_bivs (x
)
3383 register enum rtx_code code
;
3390 code
= GET_CODE (x
);
3405 /* If non-reduced/final-value givs were split, then this would also
3406 have to remap those givs also. */
3408 if (REGNO (x
) < max_reg_before_loop
3409 && reg_iv_type
[REGNO (x
)] == BASIC_INDUCT
)
3410 return reg_biv_class
[REGNO (x
)]->biv
->src_reg
;
3413 fmt
= GET_RTX_FORMAT (code
);
3414 for (i
= GET_RTX_LENGTH (code
) - 1; i
>= 0; i
--)
3417 XEXP (x
, i
) = remap_split_bivs (XEXP (x
, i
));
3421 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
3422 XVECEXP (x
, i
, j
) = remap_split_bivs (XVECEXP (x
, i
, j
));