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 /* We must limit it to max_reg_before_loop, because only these pseudo
722 registers have valid regno_first_uid info. Any register created after
723 that is unlikely to be local to the loop anyways. */
724 local_regno
= (char *) alloca (max_reg_before_loop
);
725 bzero (local_regno
, max_reg_before_loop
);
727 /* Mark all local registers, i.e. the ones which are referenced only
729 if (INSN_UID (copy_end
) < max_uid_for_loop
)
731 int copy_start_luid
= INSN_LUID (copy_start
);
732 int copy_end_luid
= INSN_LUID (copy_end
);
734 /* If a register is used in the jump insn, we must not duplicate it
735 since it will also be used outside the loop. */
736 if (GET_CODE (copy_end
) == JUMP_INSN
)
738 /* If copy_start points to the NOTE that starts the loop, then we must
739 use the next luid, because invariant pseudo-regs moved out of the loop
740 have their lifetimes modified to start here, but they are not safe
742 if (copy_start
== loop_start
)
745 for (j
= FIRST_PSEUDO_REGISTER
; j
< max_reg_before_loop
; ++j
)
746 if (regno_first_uid
[j
] > 0 && regno_first_uid
[j
] <= max_uid_for_loop
747 && uid_luid
[regno_first_uid
[j
]] >= copy_start_luid
748 && regno_last_uid
[j
] > 0 && regno_last_uid
[j
] <= max_uid_for_loop
749 && uid_luid
[regno_last_uid
[j
]] <= copy_end_luid
)
753 /* If this loop requires exit tests when unrolled, check to see if we
754 can precondition the loop so as to make the exit tests unnecessary.
755 Just like variable splitting, this is not safe if the loop is entered
756 via a jump to the bottom. Also, can not do this if no strength
757 reduce info, because precondition_loop_p uses this info. */
759 /* Must copy the loop body for preconditioning before the following
760 find_splittable_regs call since that will emit insns which need to
761 be after the preconditioned loop copies, but immediately before the
762 unrolled loop copies. */
764 /* Also, it is not safe to split induction variables for the preconditioned
765 copies of the loop body. If we split induction variables, then the code
766 assumes that each induction variable can be represented as a function
767 of its initial value and the loop iteration number. This is not true
768 in this case, because the last preconditioned copy of the loop body
769 could be any iteration from the first up to the `unroll_number-1'th,
770 depending on the initial value of the iteration variable. Therefore
771 we can not split induction variables here, because we can not calculate
772 their value. Hence, this code must occur before find_splittable_regs
775 if (unroll_type
== UNROLL_NAIVE
&& ! splitting_not_safe
&& strength_reduce_p
)
777 rtx initial_value
, final_value
, increment
;
779 if (precondition_loop_p (&initial_value
, &final_value
, &increment
,
780 loop_start
, loop_end
))
782 register rtx diff
, temp
;
783 enum machine_mode mode
;
785 int abs_inc
, neg_inc
;
787 map
->reg_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
789 map
->const_equiv_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
790 map
->const_age_map
= (unsigned *) alloca (maxregnum
791 * sizeof (unsigned));
792 map
->const_equiv_map_size
= maxregnum
;
793 global_const_equiv_map
= map
->const_equiv_map
;
794 global_const_equiv_map_size
= maxregnum
;
796 init_reg_map (map
, maxregnum
);
798 /* Limit loop unrolling to 4, since this will make 7 copies of
800 if (unroll_number
> 4)
803 /* Save the absolute value of the increment, and also whether or
804 not it is negative. */
806 abs_inc
= INTVAL (increment
);
815 /* Decide what mode to do these calculations in. Choose the larger
816 of final_value's mode and initial_value's mode, or a full-word if
817 both are constants. */
818 mode
= GET_MODE (final_value
);
819 if (mode
== VOIDmode
)
821 mode
= GET_MODE (initial_value
);
822 if (mode
== VOIDmode
)
825 else if (mode
!= GET_MODE (initial_value
)
826 && (GET_MODE_SIZE (mode
)
827 < GET_MODE_SIZE (GET_MODE (initial_value
))))
828 mode
= GET_MODE (initial_value
);
830 /* Calculate the difference between the final and initial values.
831 Final value may be a (plus (reg x) (const_int 1)) rtx.
832 Let the following cse pass simplify this if initial value is
835 We must copy the final and initial values here to avoid
836 improperly shared rtl. */
838 diff
= expand_binop (mode
, sub_optab
, copy_rtx (final_value
),
839 copy_rtx (initial_value
), NULL_RTX
, 0,
842 /* Now calculate (diff % (unroll * abs (increment))) by using an
844 diff
= expand_binop (GET_MODE (diff
), and_optab
, diff
,
845 GEN_INT (unroll_number
* abs_inc
- 1),
846 NULL_RTX
, 0, OPTAB_LIB_WIDEN
);
848 /* Now emit a sequence of branches to jump to the proper precond
851 labels
= (rtx
*) alloca (sizeof (rtx
) * unroll_number
);
852 for (i
= 0; i
< unroll_number
; i
++)
853 labels
[i
] = gen_label_rtx ();
855 /* Assuming the unroll_number is 4, and the increment is 2, then
856 for a negative increment: for a positive increment:
857 diff = 0,1 precond 0 diff = 0,7 precond 0
858 diff = 2,3 precond 3 diff = 1,2 precond 1
859 diff = 4,5 precond 2 diff = 3,4 precond 2
860 diff = 6,7 precond 1 diff = 5,6 precond 3 */
862 /* We only need to emit (unroll_number - 1) branches here, the
863 last case just falls through to the following code. */
865 /* ??? This would give better code if we emitted a tree of branches
866 instead of the current linear list of branches. */
868 for (i
= 0; i
< unroll_number
- 1; i
++)
872 /* For negative increments, must invert the constant compared
873 against, except when comparing against zero. */
877 cmp_const
= unroll_number
- i
;
881 emit_cmp_insn (diff
, GEN_INT (abs_inc
* cmp_const
),
882 EQ
, NULL_RTX
, mode
, 0, 0);
885 emit_jump_insn (gen_beq (labels
[i
]));
887 emit_jump_insn (gen_bge (labels
[i
]));
889 emit_jump_insn (gen_ble (labels
[i
]));
890 JUMP_LABEL (get_last_insn ()) = labels
[i
];
891 LABEL_NUSES (labels
[i
])++;
894 /* If the increment is greater than one, then we need another branch,
895 to handle other cases equivalent to 0. */
897 /* ??? This should be merged into the code above somehow to help
898 simplify the code here, and reduce the number of branches emitted.
899 For the negative increment case, the branch here could easily
900 be merged with the `0' case branch above. For the positive
901 increment case, it is not clear how this can be simplified. */
908 cmp_const
= abs_inc
- 1;
910 cmp_const
= abs_inc
* (unroll_number
- 1) + 1;
912 emit_cmp_insn (diff
, GEN_INT (cmp_const
), EQ
, NULL_RTX
,
916 emit_jump_insn (gen_ble (labels
[0]));
918 emit_jump_insn (gen_bge (labels
[0]));
919 JUMP_LABEL (get_last_insn ()) = labels
[0];
920 LABEL_NUSES (labels
[0])++;
923 sequence
= gen_sequence ();
925 emit_insn_before (sequence
, loop_start
);
927 /* Only the last copy of the loop body here needs the exit
928 test, so set copy_end to exclude the compare/branch here,
929 and then reset it inside the loop when get to the last
932 if (GET_CODE (last_loop_insn
) == BARRIER
)
933 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
934 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
937 /* The immediately preceding insn is a compare which we do not
939 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
941 /* The immediately preceding insn may not be a compare, so we
943 copy_end
= PREV_INSN (last_loop_insn
);
949 for (i
= 1; i
< unroll_number
; i
++)
951 emit_label_after (labels
[unroll_number
- i
],
952 PREV_INSN (loop_start
));
954 bzero ((char *) map
->insn_map
, max_insnno
* sizeof (rtx
));
955 bzero ((char *) map
->const_equiv_map
, maxregnum
* sizeof (rtx
));
956 bzero ((char *) map
->const_age_map
,
957 maxregnum
* sizeof (unsigned));
960 for (j
= 0; j
< max_labelno
; j
++)
962 map
->label_map
[j
] = gen_label_rtx ();
964 for (j
= FIRST_PSEUDO_REGISTER
; j
< max_reg_before_loop
; j
++)
966 map
->reg_map
[j
] = gen_reg_rtx (GET_MODE (regno_reg_rtx
[j
]));
968 /* The last copy needs the compare/branch insns at the end,
969 so reset copy_end here if the loop ends with a conditional
972 if (i
== unroll_number
- 1)
974 if (GET_CODE (last_loop_insn
) == BARRIER
)
975 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
977 copy_end
= last_loop_insn
;
980 /* None of the copies are the `last_iteration', so just
981 pass zero for that parameter. */
982 copy_loop_body (copy_start
, copy_end
, map
, exit_label
, 0,
983 unroll_type
, start_label
, loop_end
,
984 loop_start
, copy_end
);
986 emit_label_after (labels
[0], PREV_INSN (loop_start
));
988 if (GET_CODE (last_loop_insn
) == BARRIER
)
990 insert_before
= PREV_INSN (last_loop_insn
);
991 copy_end
= PREV_INSN (insert_before
);
996 /* The immediately preceding insn is a compare which we do not
998 insert_before
= PREV_INSN (last_loop_insn
);
999 copy_end
= PREV_INSN (insert_before
);
1001 /* The immediately preceding insn may not be a compare, so we
1003 insert_before
= last_loop_insn
;
1004 copy_end
= PREV_INSN (last_loop_insn
);
1008 /* Set unroll type to MODULO now. */
1009 unroll_type
= UNROLL_MODULO
;
1010 loop_preconditioned
= 1;
1014 /* If reach here, and the loop type is UNROLL_NAIVE, then don't unroll
1015 the loop unless all loops are being unrolled. */
1016 if (unroll_type
== UNROLL_NAIVE
&& ! flag_unroll_all_loops
)
1018 if (loop_dump_stream
)
1019 fprintf (loop_dump_stream
, "Unrolling failure: Naive unrolling not being done.\n");
1023 /* At this point, we are guaranteed to unroll the loop. */
1025 /* For each biv and giv, determine whether it can be safely split into
1026 a different variable for each unrolled copy of the loop body.
1027 We precalculate and save this info here, since computing it is
1030 Do this before deleting any instructions from the loop, so that
1031 back_branch_in_range_p will work correctly. */
1033 if (splitting_not_safe
)
1036 temp
= find_splittable_regs (unroll_type
, loop_start
, loop_end
,
1037 end_insert_before
, unroll_number
);
1039 /* find_splittable_regs may have created some new registers, so must
1040 reallocate the reg_map with the new larger size, and must realloc
1041 the constant maps also. */
1043 maxregnum
= max_reg_num ();
1044 map
->reg_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
1046 init_reg_map (map
, maxregnum
);
1048 /* Space is needed in some of the map for new registers, so new_maxregnum
1049 is an (over)estimate of how many registers will exist at the end. */
1050 new_maxregnum
= maxregnum
+ (temp
* unroll_number
* 2);
1052 /* Must realloc space for the constant maps, because the number of registers
1053 may have changed. */
1055 map
->const_equiv_map
= (rtx
*) alloca (new_maxregnum
* sizeof (rtx
));
1056 map
->const_age_map
= (unsigned *) alloca (new_maxregnum
* sizeof (unsigned));
1058 map
->const_equiv_map_size
= new_maxregnum
;
1059 global_const_equiv_map
= map
->const_equiv_map
;
1060 global_const_equiv_map_size
= new_maxregnum
;
1062 /* Search the list of bivs and givs to find ones which need to be remapped
1063 when split, and set their reg_map entry appropriately. */
1065 for (bl
= loop_iv_list
; bl
; bl
= bl
->next
)
1067 if (REGNO (bl
->biv
->src_reg
) != bl
->regno
)
1068 map
->reg_map
[bl
->regno
] = bl
->biv
->src_reg
;
1070 /* Currently, non-reduced/final-value givs are never split. */
1071 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
1072 if (REGNO (v
->src_reg
) != bl
->regno
)
1073 map
->reg_map
[REGNO (v
->dest_reg
)] = v
->src_reg
;
1077 /* If the loop is being partially unrolled, and the iteration variables
1078 are being split, and are being renamed for the split, then must fix up
1079 the compare/jump instruction at the end of the loop to refer to the new
1080 registers. This compare isn't copied, so the registers used in it
1081 will never be replaced if it isn't done here. */
1083 if (unroll_type
== UNROLL_MODULO
)
1085 insn
= NEXT_INSN (copy_end
);
1086 if (GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == JUMP_INSN
)
1087 PATTERN (insn
) = remap_split_bivs (PATTERN (insn
));
1090 /* For unroll_number - 1 times, make a copy of each instruction
1091 between copy_start and copy_end, and insert these new instructions
1092 before the end of the loop. */
1094 for (i
= 0; i
< unroll_number
; i
++)
1096 bzero ((char *) map
->insn_map
, max_insnno
* sizeof (rtx
));
1097 bzero ((char *) map
->const_equiv_map
, new_maxregnum
* sizeof (rtx
));
1098 bzero ((char *) map
->const_age_map
, new_maxregnum
* sizeof (unsigned));
1101 for (j
= 0; j
< max_labelno
; j
++)
1103 map
->label_map
[j
] = gen_label_rtx ();
1105 for (j
= FIRST_PSEUDO_REGISTER
; j
< max_reg_before_loop
; j
++)
1107 map
->reg_map
[j
] = gen_reg_rtx (GET_MODE (regno_reg_rtx
[j
]));
1109 /* If loop starts with a branch to the test, then fix it so that
1110 it points to the test of the first unrolled copy of the loop. */
1111 if (i
== 0 && loop_start
!= copy_start
)
1113 insn
= PREV_INSN (copy_start
);
1114 pattern
= PATTERN (insn
);
1116 tem
= map
->label_map
[CODE_LABEL_NUMBER
1117 (XEXP (SET_SRC (pattern
), 0))];
1118 SET_SRC (pattern
) = gen_rtx (LABEL_REF
, VOIDmode
, tem
);
1120 /* Set the jump label so that it can be used by later loop unrolling
1122 JUMP_LABEL (insn
) = tem
;
1123 LABEL_NUSES (tem
)++;
1126 copy_loop_body (copy_start
, copy_end
, map
, exit_label
,
1127 i
== unroll_number
- 1, unroll_type
, start_label
,
1128 loop_end
, insert_before
, insert_before
);
1131 /* Before deleting any insns, emit a CODE_LABEL immediately after the last
1132 insn to be deleted. This prevents any runaway delete_insn call from
1133 more insns that it should, as it always stops at a CODE_LABEL. */
1135 /* Delete the compare and branch at the end of the loop if completely
1136 unrolling the loop. Deleting the backward branch at the end also
1137 deletes the code label at the start of the loop. This is done at
1138 the very end to avoid problems with back_branch_in_range_p. */
1140 if (unroll_type
== UNROLL_COMPLETELY
)
1141 safety_label
= emit_label_after (gen_label_rtx (), last_loop_insn
);
1143 safety_label
= emit_label_after (gen_label_rtx (), copy_end
);
1145 /* Delete all of the original loop instructions. Don't delete the
1146 LOOP_BEG note, or the first code label in the loop. */
1148 insn
= NEXT_INSN (copy_start
);
1149 while (insn
!= safety_label
)
1151 if (insn
!= start_label
)
1152 insn
= delete_insn (insn
);
1154 insn
= NEXT_INSN (insn
);
1157 /* Can now delete the 'safety' label emitted to protect us from runaway
1158 delete_insn calls. */
1159 if (INSN_DELETED_P (safety_label
))
1161 delete_insn (safety_label
);
1163 /* If exit_label exists, emit it after the loop. Doing the emit here
1164 forces it to have a higher INSN_UID than any insn in the unrolled loop.
1165 This is needed so that mostly_true_jump in reorg.c will treat jumps
1166 to this loop end label correctly, i.e. predict that they are usually
1169 emit_label_after (exit_label
, loop_end
);
1172 /* Return true if the loop can be safely, and profitably, preconditioned
1173 so that the unrolled copies of the loop body don't need exit tests.
1175 This only works if final_value, initial_value and increment can be
1176 determined, and if increment is a constant power of 2.
1177 If increment is not a power of 2, then the preconditioning modulo
1178 operation would require a real modulo instead of a boolean AND, and this
1179 is not considered `profitable'. */
1181 /* ??? If the loop is known to be executed very many times, or the machine
1182 has a very cheap divide instruction, then preconditioning is a win even
1183 when the increment is not a power of 2. Use RTX_COST to compute
1184 whether divide is cheap. */
1187 precondition_loop_p (initial_value
, final_value
, increment
, loop_start
,
1189 rtx
*initial_value
, *final_value
, *increment
;
1190 rtx loop_start
, loop_end
;
1193 if (loop_n_iterations
> 0)
1195 *initial_value
= const0_rtx
;
1196 *increment
= const1_rtx
;
1197 *final_value
= GEN_INT (loop_n_iterations
);
1199 if (loop_dump_stream
)
1200 fprintf (loop_dump_stream
,
1201 "Preconditioning: Success, number of iterations known, %d.\n",
1206 if (loop_initial_value
== 0)
1208 if (loop_dump_stream
)
1209 fprintf (loop_dump_stream
,
1210 "Preconditioning: Could not find initial value.\n");
1213 else if (loop_increment
== 0)
1215 if (loop_dump_stream
)
1216 fprintf (loop_dump_stream
,
1217 "Preconditioning: Could not find increment value.\n");
1220 else if (GET_CODE (loop_increment
) != CONST_INT
)
1222 if (loop_dump_stream
)
1223 fprintf (loop_dump_stream
,
1224 "Preconditioning: Increment not a constant.\n");
1227 else if ((exact_log2 (INTVAL (loop_increment
)) < 0)
1228 && (exact_log2 (- INTVAL (loop_increment
)) < 0))
1230 if (loop_dump_stream
)
1231 fprintf (loop_dump_stream
,
1232 "Preconditioning: Increment not a constant power of 2.\n");
1236 /* Unsigned_compare and compare_dir can be ignored here, since they do
1237 not matter for preconditioning. */
1239 if (loop_final_value
== 0)
1241 if (loop_dump_stream
)
1242 fprintf (loop_dump_stream
,
1243 "Preconditioning: EQ comparison loop.\n");
1247 /* Must ensure that final_value is invariant, so call invariant_p to
1248 check. Before doing so, must check regno against max_reg_before_loop
1249 to make sure that the register is in the range covered by invariant_p.
1250 If it isn't, then it is most likely a biv/giv which by definition are
1252 if ((GET_CODE (loop_final_value
) == REG
1253 && REGNO (loop_final_value
) >= max_reg_before_loop
)
1254 || (GET_CODE (loop_final_value
) == PLUS
1255 && REGNO (XEXP (loop_final_value
, 0)) >= max_reg_before_loop
)
1256 || ! invariant_p (loop_final_value
))
1258 if (loop_dump_stream
)
1259 fprintf (loop_dump_stream
,
1260 "Preconditioning: Final value not invariant.\n");
1264 /* Fail for floating point values, since the caller of this function
1265 does not have code to deal with them. */
1266 if (GET_MODE_CLASS (GET_MODE (loop_final_value
)) == MODE_FLOAT
1267 || GET_MODE_CLASS (GET_MODE (loop_initial_value
)) == MODE_FLOAT
)
1269 if (loop_dump_stream
)
1270 fprintf (loop_dump_stream
,
1271 "Preconditioning: Floating point final or initial value.\n");
1275 /* Now set initial_value to be the iteration_var, since that may be a
1276 simpler expression, and is guaranteed to be correct if all of the
1277 above tests succeed.
1279 We can not use the initial_value as calculated, because it will be
1280 one too small for loops of the form "while (i-- > 0)". We can not
1281 emit code before the loop_skip_over insns to fix this problem as this
1282 will then give a number one too large for loops of the form
1285 Note that all loops that reach here are entered at the top, because
1286 this function is not called if the loop starts with a jump. */
1288 /* Fail if loop_iteration_var is not live before loop_start, since we need
1289 to test its value in the preconditioning code. */
1291 if (uid_luid
[regno_first_uid
[REGNO (loop_iteration_var
)]]
1292 > INSN_LUID (loop_start
))
1294 if (loop_dump_stream
)
1295 fprintf (loop_dump_stream
,
1296 "Preconditioning: Iteration var not live before loop start.\n");
1300 *initial_value
= loop_iteration_var
;
1301 *increment
= loop_increment
;
1302 *final_value
= loop_final_value
;
1305 if (loop_dump_stream
)
1306 fprintf (loop_dump_stream
, "Preconditioning: Successful.\n");
1311 /* All pseudo-registers must be mapped to themselves. Two hard registers
1312 must be mapped, VIRTUAL_STACK_VARS_REGNUM and VIRTUAL_INCOMING_ARGS_
1313 REGNUM, to avoid function-inlining specific conversions of these
1314 registers. All other hard regs can not be mapped because they may be
1319 init_reg_map (map
, maxregnum
)
1320 struct inline_remap
*map
;
1325 for (i
= maxregnum
- 1; i
> LAST_VIRTUAL_REGISTER
; i
--)
1326 map
->reg_map
[i
] = regno_reg_rtx
[i
];
1327 /* Just clear the rest of the entries. */
1328 for (i
= LAST_VIRTUAL_REGISTER
; i
>= 0; i
--)
1329 map
->reg_map
[i
] = 0;
1331 map
->reg_map
[VIRTUAL_STACK_VARS_REGNUM
]
1332 = regno_reg_rtx
[VIRTUAL_STACK_VARS_REGNUM
];
1333 map
->reg_map
[VIRTUAL_INCOMING_ARGS_REGNUM
]
1334 = regno_reg_rtx
[VIRTUAL_INCOMING_ARGS_REGNUM
];
1337 /* Strength-reduction will often emit code for optimized biv/givs which
1338 calculates their value in a temporary register, and then copies the result
1339 to the iv. This procedure reconstructs the pattern computing the iv;
1340 verifying that all operands are of the proper form.
1342 The return value is the amount that the giv is incremented by. */
1345 calculate_giv_inc (pattern
, src_insn
, regno
)
1346 rtx pattern
, src_insn
;
1350 rtx increment_total
= 0;
1354 /* Verify that we have an increment insn here. First check for a plus
1355 as the set source. */
1356 if (GET_CODE (SET_SRC (pattern
)) != PLUS
)
1358 /* SR sometimes computes the new giv value in a temp, then copies it
1360 src_insn
= PREV_INSN (src_insn
);
1361 pattern
= PATTERN (src_insn
);
1362 if (GET_CODE (SET_SRC (pattern
)) != PLUS
)
1365 /* The last insn emitted is not needed, so delete it to avoid confusing
1366 the second cse pass. This insn sets the giv unnecessarily. */
1367 delete_insn (get_last_insn ());
1370 /* Verify that we have a constant as the second operand of the plus. */
1371 increment
= XEXP (SET_SRC (pattern
), 1);
1372 if (GET_CODE (increment
) != CONST_INT
)
1374 /* SR sometimes puts the constant in a register, especially if it is
1375 too big to be an add immed operand. */
1376 src_insn
= PREV_INSN (src_insn
);
1377 increment
= SET_SRC (PATTERN (src_insn
));
1379 /* SR may have used LO_SUM to compute the constant if it is too large
1380 for a load immed operand. In this case, the constant is in operand
1381 one of the LO_SUM rtx. */
1382 if (GET_CODE (increment
) == LO_SUM
)
1383 increment
= XEXP (increment
, 1);
1384 else if (GET_CODE (increment
) == IOR
)
1386 /* The rs6000 port loads some constants with IOR. */
1387 rtx second_part
= XEXP (increment
, 1);
1389 src_insn
= PREV_INSN (src_insn
);
1390 increment
= SET_SRC (PATTERN (src_insn
));
1391 /* Don't need the last insn anymore. */
1392 delete_insn (get_last_insn ());
1394 if (GET_CODE (second_part
) != CONST_INT
1395 || GET_CODE (increment
) != CONST_INT
)
1398 increment
= GEN_INT (INTVAL (increment
) | INTVAL (second_part
));
1401 if (GET_CODE (increment
) != CONST_INT
)
1404 /* The insn loading the constant into a register is no longer needed,
1406 delete_insn (get_last_insn ());
1409 if (increment_total
)
1410 increment_total
= GEN_INT (INTVAL (increment_total
) + INTVAL (increment
));
1412 increment_total
= increment
;
1414 /* Check that the source register is the same as the register we expected
1415 to see as the source. If not, something is seriously wrong. */
1416 if (GET_CODE (XEXP (SET_SRC (pattern
), 0)) != REG
1417 || REGNO (XEXP (SET_SRC (pattern
), 0)) != regno
)
1419 /* Some machines (e.g. the romp), may emit two add instructions for
1420 certain constants, so lets try looking for another add immediately
1421 before this one if we have only seen one add insn so far. */
1427 src_insn
= PREV_INSN (src_insn
);
1428 pattern
= PATTERN (src_insn
);
1430 delete_insn (get_last_insn ());
1438 return increment_total
;
1441 /* Copy REG_NOTES, except for insn references, because not all insn_map
1442 entries are valid yet. We do need to copy registers now though, because
1443 the reg_map entries can change during copying. */
1446 initial_reg_note_copy (notes
, map
)
1448 struct inline_remap
*map
;
1455 copy
= rtx_alloc (GET_CODE (notes
));
1456 PUT_MODE (copy
, GET_MODE (notes
));
1458 if (GET_CODE (notes
) == EXPR_LIST
)
1459 XEXP (copy
, 0) = copy_rtx_and_substitute (XEXP (notes
, 0), map
);
1460 else if (GET_CODE (notes
) == INSN_LIST
)
1461 /* Don't substitute for these yet. */
1462 XEXP (copy
, 0) = XEXP (notes
, 0);
1466 XEXP (copy
, 1) = initial_reg_note_copy (XEXP (notes
, 1), map
);
1471 /* Fixup insn references in copied REG_NOTES. */
1474 final_reg_note_copy (notes
, map
)
1476 struct inline_remap
*map
;
1480 for (note
= notes
; note
; note
= XEXP (note
, 1))
1481 if (GET_CODE (note
) == INSN_LIST
)
1482 XEXP (note
, 0) = map
->insn_map
[INSN_UID (XEXP (note
, 0))];
1485 /* Copy each instruction in the loop, substituting from map as appropriate.
1486 This is very similar to a loop in expand_inline_function. */
1489 copy_loop_body (copy_start
, copy_end
, map
, exit_label
, last_iteration
,
1490 unroll_type
, start_label
, loop_end
, insert_before
,
1492 rtx copy_start
, copy_end
;
1493 struct inline_remap
*map
;
1496 enum unroll_types unroll_type
;
1497 rtx start_label
, loop_end
, insert_before
, copy_notes_from
;
1501 int dest_reg_was_split
, i
;
1503 rtx final_label
= 0;
1504 rtx giv_inc
, giv_dest_reg
, giv_src_reg
;
1506 /* If this isn't the last iteration, then map any references to the
1507 start_label to final_label. Final label will then be emitted immediately
1508 after the end of this loop body if it was ever used.
1510 If this is the last iteration, then map references to the start_label
1512 if (! last_iteration
)
1514 final_label
= gen_label_rtx ();
1515 map
->label_map
[CODE_LABEL_NUMBER (start_label
)] = final_label
;
1518 map
->label_map
[CODE_LABEL_NUMBER (start_label
)] = start_label
;
1525 insn
= NEXT_INSN (insn
);
1527 map
->orig_asm_operands_vector
= 0;
1529 switch (GET_CODE (insn
))
1532 pattern
= PATTERN (insn
);
1536 /* Check to see if this is a giv that has been combined with
1537 some split address givs. (Combined in the sense that
1538 `combine_givs' in loop.c has put two givs in the same register.)
1539 In this case, we must search all givs based on the same biv to
1540 find the address givs. Then split the address givs.
1541 Do this before splitting the giv, since that may map the
1542 SET_DEST to a new register. */
1544 if (GET_CODE (pattern
) == SET
1545 && GET_CODE (SET_DEST (pattern
)) == REG
1546 && addr_combined_regs
[REGNO (SET_DEST (pattern
))])
1548 struct iv_class
*bl
;
1549 struct induction
*v
, *tv
;
1550 int regno
= REGNO (SET_DEST (pattern
));
1552 v
= addr_combined_regs
[REGNO (SET_DEST (pattern
))];
1553 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
1555 /* Although the giv_inc amount is not needed here, we must call
1556 calculate_giv_inc here since it might try to delete the
1557 last insn emitted. If we wait until later to call it,
1558 we might accidentally delete insns generated immediately
1559 below by emit_unrolled_add. */
1561 giv_inc
= calculate_giv_inc (pattern
, insn
, regno
);
1563 /* Now find all address giv's that were combined with this
1565 for (tv
= bl
->giv
; tv
; tv
= tv
->next_iv
)
1566 if (tv
->giv_type
== DEST_ADDR
&& tv
->same
== v
)
1568 int this_giv_inc
= INTVAL (giv_inc
);
1570 /* Scale this_giv_inc if the multiplicative factors of
1571 the two givs are different. */
1572 if (tv
->mult_val
!= v
->mult_val
)
1573 this_giv_inc
= (this_giv_inc
/ INTVAL (v
->mult_val
)
1574 * INTVAL (tv
->mult_val
));
1576 tv
->dest_reg
= plus_constant (tv
->dest_reg
, this_giv_inc
);
1577 *tv
->location
= tv
->dest_reg
;
1579 if (last_iteration
&& unroll_type
!= UNROLL_COMPLETELY
)
1581 /* Must emit an insn to increment the split address
1582 giv. Add in the const_adjust field in case there
1583 was a constant eliminated from the address. */
1584 rtx value
, dest_reg
;
1586 /* tv->dest_reg will be either a bare register,
1587 or else a register plus a constant. */
1588 if (GET_CODE (tv
->dest_reg
) == REG
)
1589 dest_reg
= tv
->dest_reg
;
1591 dest_reg
= XEXP (tv
->dest_reg
, 0);
1593 /* Check for shared address givs, and avoid
1594 incrementing the shared pseudo reg more than
1596 if (! tv
->same_insn
)
1598 /* tv->dest_reg may actually be a (PLUS (REG)
1599 (CONST)) here, so we must call plus_constant
1600 to add the const_adjust amount before calling
1601 emit_unrolled_add below. */
1602 value
= plus_constant (tv
->dest_reg
,
1605 /* The constant could be too large for an add
1606 immediate, so can't directly emit an insn
1608 emit_unrolled_add (dest_reg
, XEXP (value
, 0),
1612 /* Reset the giv to be just the register again, in case
1613 it is used after the set we have just emitted.
1614 We must subtract the const_adjust factor added in
1616 tv
->dest_reg
= plus_constant (dest_reg
,
1617 - tv
->const_adjust
);
1618 *tv
->location
= tv
->dest_reg
;
1623 /* If this is a setting of a splittable variable, then determine
1624 how to split the variable, create a new set based on this split,
1625 and set up the reg_map so that later uses of the variable will
1626 use the new split variable. */
1628 dest_reg_was_split
= 0;
1630 if (GET_CODE (pattern
) == SET
1631 && GET_CODE (SET_DEST (pattern
)) == REG
1632 && splittable_regs
[REGNO (SET_DEST (pattern
))])
1634 int regno
= REGNO (SET_DEST (pattern
));
1636 dest_reg_was_split
= 1;
1638 /* Compute the increment value for the giv, if it wasn't
1639 already computed above. */
1642 giv_inc
= calculate_giv_inc (pattern
, insn
, regno
);
1643 giv_dest_reg
= SET_DEST (pattern
);
1644 giv_src_reg
= SET_DEST (pattern
);
1646 if (unroll_type
== UNROLL_COMPLETELY
)
1648 /* Completely unrolling the loop. Set the induction
1649 variable to a known constant value. */
1651 /* The value in splittable_regs may be an invariant
1652 value, so we must use plus_constant here. */
1653 splittable_regs
[regno
]
1654 = plus_constant (splittable_regs
[regno
], INTVAL (giv_inc
));
1656 if (GET_CODE (splittable_regs
[regno
]) == PLUS
)
1658 giv_src_reg
= XEXP (splittable_regs
[regno
], 0);
1659 giv_inc
= XEXP (splittable_regs
[regno
], 1);
1663 /* The splittable_regs value must be a REG or a
1664 CONST_INT, so put the entire value in the giv_src_reg
1666 giv_src_reg
= splittable_regs
[regno
];
1667 giv_inc
= const0_rtx
;
1672 /* Partially unrolling loop. Create a new pseudo
1673 register for the iteration variable, and set it to
1674 be a constant plus the original register. Except
1675 on the last iteration, when the result has to
1676 go back into the original iteration var register. */
1678 /* Handle bivs which must be mapped to a new register
1679 when split. This happens for bivs which need their
1680 final value set before loop entry. The new register
1681 for the biv was stored in the biv's first struct
1682 induction entry by find_splittable_regs. */
1684 if (regno
< max_reg_before_loop
1685 && reg_iv_type
[regno
] == BASIC_INDUCT
)
1687 giv_src_reg
= reg_biv_class
[regno
]->biv
->src_reg
;
1688 giv_dest_reg
= giv_src_reg
;
1692 /* If non-reduced/final-value givs were split, then
1693 this would have to remap those givs also. See
1694 find_splittable_regs. */
1697 splittable_regs
[regno
]
1698 = GEN_INT (INTVAL (giv_inc
)
1699 + INTVAL (splittable_regs
[regno
]));
1700 giv_inc
= splittable_regs
[regno
];
1702 /* Now split the induction variable by changing the dest
1703 of this insn to a new register, and setting its
1704 reg_map entry to point to this new register.
1706 If this is the last iteration, and this is the last insn
1707 that will update the iv, then reuse the original dest,
1708 to ensure that the iv will have the proper value when
1709 the loop exits or repeats.
1711 Using splittable_regs_updates here like this is safe,
1712 because it can only be greater than one if all
1713 instructions modifying the iv are always executed in
1716 if (! last_iteration
1717 || (splittable_regs_updates
[regno
]-- != 1))
1719 tem
= gen_reg_rtx (GET_MODE (giv_src_reg
));
1721 map
->reg_map
[regno
] = tem
;
1724 map
->reg_map
[regno
] = giv_src_reg
;
1727 /* The constant being added could be too large for an add
1728 immediate, so can't directly emit an insn here. */
1729 emit_unrolled_add (giv_dest_reg
, giv_src_reg
, giv_inc
);
1730 copy
= get_last_insn ();
1731 pattern
= PATTERN (copy
);
1735 pattern
= copy_rtx_and_substitute (pattern
, map
);
1736 copy
= emit_insn (pattern
);
1738 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
1741 /* If this insn is setting CC0, it may need to look at
1742 the insn that uses CC0 to see what type of insn it is.
1743 In that case, the call to recog via validate_change will
1744 fail. So don't substitute constants here. Instead,
1745 do it when we emit the following insn.
1747 For example, see the pyr.md file. That machine has signed and
1748 unsigned compares. The compare patterns must check the
1749 following branch insn to see which what kind of compare to
1752 If the previous insn set CC0, substitute constants on it as
1754 if (sets_cc0_p (PATTERN (copy
)) != 0)
1759 try_constants (cc0_insn
, map
);
1761 try_constants (copy
, map
);
1764 try_constants (copy
, map
);
1767 /* Make split induction variable constants `permanent' since we
1768 know there are no backward branches across iteration variable
1769 settings which would invalidate this. */
1770 if (dest_reg_was_split
)
1772 int regno
= REGNO (SET_DEST (pattern
));
1774 if (regno
< map
->const_equiv_map_size
1775 && map
->const_age_map
[regno
] == map
->const_age
)
1776 map
->const_age_map
[regno
] = -1;
1781 pattern
= copy_rtx_and_substitute (PATTERN (insn
), map
);
1782 copy
= emit_jump_insn (pattern
);
1783 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
1785 if (JUMP_LABEL (insn
) == start_label
&& insn
== copy_end
1786 && ! last_iteration
)
1788 /* This is a branch to the beginning of the loop; this is the
1789 last insn being copied; and this is not the last iteration.
1790 In this case, we want to change the original fall through
1791 case to be a branch past the end of the loop, and the
1792 original jump label case to fall_through. */
1794 if (invert_exp (pattern
, copy
))
1796 if (! redirect_exp (&pattern
,
1797 map
->label_map
[CODE_LABEL_NUMBER
1798 (JUMP_LABEL (insn
))],
1805 rtx lab
= gen_label_rtx ();
1806 /* Can't do it by reversing the jump (probably because we
1807 couldn't reverse the conditions), so emit a new
1808 jump_insn after COPY, and redirect the jump around
1810 jmp
= emit_jump_insn_after (gen_jump (exit_label
), copy
);
1811 jmp
= emit_barrier_after (jmp
);
1812 emit_label_after (lab
, jmp
);
1813 LABEL_NUSES (lab
) = 0;
1814 if (! redirect_exp (&pattern
,
1815 map
->label_map
[CODE_LABEL_NUMBER
1816 (JUMP_LABEL (insn
))],
1824 try_constants (cc0_insn
, map
);
1827 try_constants (copy
, map
);
1829 /* Set the jump label of COPY correctly to avoid problems with
1830 later passes of unroll_loop, if INSN had jump label set. */
1831 if (JUMP_LABEL (insn
))
1835 /* Can't use the label_map for every insn, since this may be
1836 the backward branch, and hence the label was not mapped. */
1837 if (GET_CODE (pattern
) == SET
)
1839 tem
= SET_SRC (pattern
);
1840 if (GET_CODE (tem
) == LABEL_REF
)
1841 label
= XEXP (tem
, 0);
1842 else if (GET_CODE (tem
) == IF_THEN_ELSE
)
1844 if (XEXP (tem
, 1) != pc_rtx
)
1845 label
= XEXP (XEXP (tem
, 1), 0);
1847 label
= XEXP (XEXP (tem
, 2), 0);
1851 if (label
&& GET_CODE (label
) == CODE_LABEL
)
1852 JUMP_LABEL (copy
) = label
;
1855 /* An unrecognizable jump insn, probably the entry jump
1856 for a switch statement. This label must have been mapped,
1857 so just use the label_map to get the new jump label. */
1859 = map
->label_map
[CODE_LABEL_NUMBER (JUMP_LABEL (insn
))];
1862 /* If this is a non-local jump, then must increase the label
1863 use count so that the label will not be deleted when the
1864 original jump is deleted. */
1865 LABEL_NUSES (JUMP_LABEL (copy
))++;
1867 else if (GET_CODE (PATTERN (copy
)) == ADDR_VEC
1868 || GET_CODE (PATTERN (copy
)) == ADDR_DIFF_VEC
)
1870 rtx pat
= PATTERN (copy
);
1871 int diff_vec_p
= GET_CODE (pat
) == ADDR_DIFF_VEC
;
1872 int len
= XVECLEN (pat
, diff_vec_p
);
1875 for (i
= 0; i
< len
; i
++)
1876 LABEL_NUSES (XEXP (XVECEXP (pat
, diff_vec_p
, i
), 0))++;
1879 /* If this used to be a conditional jump insn but whose branch
1880 direction is now known, we must do something special. */
1881 if (condjump_p (insn
) && !simplejump_p (insn
) && map
->last_pc_value
)
1884 /* The previous insn set cc0 for us. So delete it. */
1885 delete_insn (PREV_INSN (copy
));
1888 /* If this is now a no-op, delete it. */
1889 if (map
->last_pc_value
== pc_rtx
)
1891 /* Don't let delete_insn delete the label referenced here,
1892 because we might possibly need it later for some other
1893 instruction in the loop. */
1894 if (JUMP_LABEL (copy
))
1895 LABEL_NUSES (JUMP_LABEL (copy
))++;
1897 if (JUMP_LABEL (copy
))
1898 LABEL_NUSES (JUMP_LABEL (copy
))--;
1902 /* Otherwise, this is unconditional jump so we must put a
1903 BARRIER after it. We could do some dead code elimination
1904 here, but jump.c will do it just as well. */
1910 pattern
= copy_rtx_and_substitute (PATTERN (insn
), map
);
1911 copy
= emit_call_insn (pattern
);
1912 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
1914 /* Because the USAGE information potentially contains objects other
1915 than hard registers, we need to copy it. */
1916 CALL_INSN_FUNCTION_USAGE (copy
) =
1917 copy_rtx_and_substitute (CALL_INSN_FUNCTION_USAGE (insn
), map
);
1921 try_constants (cc0_insn
, map
);
1924 try_constants (copy
, map
);
1926 /* Be lazy and assume CALL_INSNs clobber all hard registers. */
1927 for (i
= 0; i
< FIRST_PSEUDO_REGISTER
; i
++)
1928 map
->const_equiv_map
[i
] = 0;
1932 /* If this is the loop start label, then we don't need to emit a
1933 copy of this label since no one will use it. */
1935 if (insn
!= start_label
)
1937 copy
= emit_label (map
->label_map
[CODE_LABEL_NUMBER (insn
)]);
1943 copy
= emit_barrier ();
1947 /* VTOP notes are valid only before the loop exit test. If placed
1948 anywhere else, loop may generate bad code. */
1950 if (NOTE_LINE_NUMBER (insn
) != NOTE_INSN_DELETED
1951 && (NOTE_LINE_NUMBER (insn
) != NOTE_INSN_LOOP_VTOP
1952 || (last_iteration
&& unroll_type
!= UNROLL_COMPLETELY
)))
1953 copy
= emit_note (NOTE_SOURCE_FILE (insn
),
1954 NOTE_LINE_NUMBER (insn
));
1964 map
->insn_map
[INSN_UID (insn
)] = copy
;
1966 while (insn
!= copy_end
);
1968 /* Now finish coping the REG_NOTES. */
1972 insn
= NEXT_INSN (insn
);
1973 if ((GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == JUMP_INSN
1974 || GET_CODE (insn
) == CALL_INSN
)
1975 && map
->insn_map
[INSN_UID (insn
)])
1976 final_reg_note_copy (REG_NOTES (map
->insn_map
[INSN_UID (insn
)]), map
);
1978 while (insn
!= copy_end
);
1980 /* There may be notes between copy_notes_from and loop_end. Emit a copy of
1981 each of these notes here, since there may be some important ones, such as
1982 NOTE_INSN_BLOCK_END notes, in this group. We don't do this on the last
1983 iteration, because the original notes won't be deleted.
1985 We can't use insert_before here, because when from preconditioning,
1986 insert_before points before the loop. We can't use copy_end, because
1987 there may be insns already inserted after it (which we don't want to
1988 copy) when not from preconditioning code. */
1990 if (! last_iteration
)
1992 for (insn
= copy_notes_from
; insn
!= loop_end
; insn
= NEXT_INSN (insn
))
1994 if (GET_CODE (insn
) == NOTE
1995 && NOTE_LINE_NUMBER (insn
) != NOTE_INSN_DELETED
)
1996 emit_note (NOTE_SOURCE_FILE (insn
), NOTE_LINE_NUMBER (insn
));
2000 if (final_label
&& LABEL_NUSES (final_label
) > 0)
2001 emit_label (final_label
);
2003 tem
= gen_sequence ();
2005 emit_insn_before (tem
, insert_before
);
2008 /* Emit an insn, using the expand_binop to ensure that a valid insn is
2009 emitted. This will correctly handle the case where the increment value
2010 won't fit in the immediate field of a PLUS insns. */
2013 emit_unrolled_add (dest_reg
, src_reg
, increment
)
2014 rtx dest_reg
, src_reg
, increment
;
2018 result
= expand_binop (GET_MODE (dest_reg
), add_optab
, src_reg
, increment
,
2019 dest_reg
, 0, OPTAB_LIB_WIDEN
);
2021 if (dest_reg
!= result
)
2022 emit_move_insn (dest_reg
, result
);
2025 /* Searches the insns between INSN and LOOP_END. Returns 1 if there
2026 is a backward branch in that range that branches to somewhere between
2027 LOOP_START and INSN. Returns 0 otherwise. */
2029 /* ??? This is quadratic algorithm. Could be rewritten to be linear.
2030 In practice, this is not a problem, because this function is seldom called,
2031 and uses a negligible amount of CPU time on average. */
2034 back_branch_in_range_p (insn
, loop_start
, loop_end
)
2036 rtx loop_start
, loop_end
;
2038 rtx p
, q
, target_insn
;
2040 /* Stop before we get to the backward branch at the end of the loop. */
2041 loop_end
= prev_nonnote_insn (loop_end
);
2042 if (GET_CODE (loop_end
) == BARRIER
)
2043 loop_end
= PREV_INSN (loop_end
);
2045 /* Check in case insn has been deleted, search forward for first non
2046 deleted insn following it. */
2047 while (INSN_DELETED_P (insn
))
2048 insn
= NEXT_INSN (insn
);
2050 /* Check for the case where insn is the last insn in the loop. */
2051 if (insn
== loop_end
)
2054 for (p
= NEXT_INSN (insn
); p
!= loop_end
; p
= NEXT_INSN (p
))
2056 if (GET_CODE (p
) == JUMP_INSN
)
2058 target_insn
= JUMP_LABEL (p
);
2060 /* Search from loop_start to insn, to see if one of them is
2061 the target_insn. We can't use INSN_LUID comparisons here,
2062 since insn may not have an LUID entry. */
2063 for (q
= loop_start
; q
!= insn
; q
= NEXT_INSN (q
))
2064 if (q
== target_insn
)
2072 /* Try to generate the simplest rtx for the expression
2073 (PLUS (MULT mult1 mult2) add1). This is used to calculate the initial
2077 fold_rtx_mult_add (mult1
, mult2
, add1
, mode
)
2078 rtx mult1
, mult2
, add1
;
2079 enum machine_mode mode
;
2084 /* The modes must all be the same. This should always be true. For now,
2085 check to make sure. */
2086 if ((GET_MODE (mult1
) != mode
&& GET_MODE (mult1
) != VOIDmode
)
2087 || (GET_MODE (mult2
) != mode
&& GET_MODE (mult2
) != VOIDmode
)
2088 || (GET_MODE (add1
) != mode
&& GET_MODE (add1
) != VOIDmode
))
2091 /* Ensure that if at least one of mult1/mult2 are constant, then mult2
2092 will be a constant. */
2093 if (GET_CODE (mult1
) == CONST_INT
)
2100 mult_res
= simplify_binary_operation (MULT
, mode
, mult1
, mult2
);
2102 mult_res
= gen_rtx (MULT
, mode
, mult1
, mult2
);
2104 /* Again, put the constant second. */
2105 if (GET_CODE (add1
) == CONST_INT
)
2112 result
= simplify_binary_operation (PLUS
, mode
, add1
, mult_res
);
2114 result
= gen_rtx (PLUS
, mode
, add1
, mult_res
);
2119 /* Searches the list of induction struct's for the biv BL, to try to calculate
2120 the total increment value for one iteration of the loop as a constant.
2122 Returns the increment value as an rtx, simplified as much as possible,
2123 if it can be calculated. Otherwise, returns 0. */
2126 biv_total_increment (bl
, loop_start
, loop_end
)
2127 struct iv_class
*bl
;
2128 rtx loop_start
, loop_end
;
2130 struct induction
*v
;
2133 /* For increment, must check every instruction that sets it. Each
2134 instruction must be executed only once each time through the loop.
2135 To verify this, we check that the the insn is always executed, and that
2136 there are no backward branches after the insn that branch to before it.
2137 Also, the insn must have a mult_val of one (to make sure it really is
2140 result
= const0_rtx
;
2141 for (v
= bl
->biv
; v
; v
= v
->next_iv
)
2143 if (v
->always_computable
&& v
->mult_val
== const1_rtx
2144 && ! back_branch_in_range_p (v
->insn
, loop_start
, loop_end
))
2145 result
= fold_rtx_mult_add (result
, const1_rtx
, v
->add_val
, v
->mode
);
2153 /* Determine the initial value of the iteration variable, and the amount
2154 that it is incremented each loop. Use the tables constructed by
2155 the strength reduction pass to calculate these values.
2157 Initial_value and/or increment are set to zero if their values could not
2161 iteration_info (iteration_var
, initial_value
, increment
, loop_start
, loop_end
)
2162 rtx iteration_var
, *initial_value
, *increment
;
2163 rtx loop_start
, loop_end
;
2165 struct iv_class
*bl
;
2166 struct induction
*v
, *b
;
2168 /* Clear the result values, in case no answer can be found. */
2172 /* The iteration variable can be either a giv or a biv. Check to see
2173 which it is, and compute the variable's initial value, and increment
2174 value if possible. */
2176 /* If this is a new register, can't handle it since we don't have any
2177 reg_iv_type entry for it. */
2178 if (REGNO (iteration_var
) >= max_reg_before_loop
)
2180 if (loop_dump_stream
)
2181 fprintf (loop_dump_stream
,
2182 "Loop unrolling: No reg_iv_type entry for iteration var.\n");
2185 /* Reject iteration variables larger than the host long size, since they
2186 could result in a number of iterations greater than the range of our
2187 `unsigned long' variable loop_n_iterations. */
2188 else if (GET_MODE_BITSIZE (GET_MODE (iteration_var
)) > HOST_BITS_PER_LONG
)
2190 if (loop_dump_stream
)
2191 fprintf (loop_dump_stream
,
2192 "Loop unrolling: Iteration var rejected because mode larger than host long.\n");
2195 else if (GET_MODE_CLASS (GET_MODE (iteration_var
)) != MODE_INT
)
2197 if (loop_dump_stream
)
2198 fprintf (loop_dump_stream
,
2199 "Loop unrolling: Iteration var not an integer.\n");
2202 else if (reg_iv_type
[REGNO (iteration_var
)] == BASIC_INDUCT
)
2204 /* Grab initial value, only useful if it is a constant. */
2205 bl
= reg_biv_class
[REGNO (iteration_var
)];
2206 *initial_value
= bl
->initial_value
;
2208 *increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2210 else if (reg_iv_type
[REGNO (iteration_var
)] == GENERAL_INDUCT
)
2213 /* ??? The code below does not work because the incorrect number of
2214 iterations is calculated when the biv is incremented after the giv
2215 is set (which is the usual case). This can probably be accounted
2216 for by biasing the initial_value by subtracting the amount of the
2217 increment that occurs between the giv set and the giv test. However,
2218 a giv as an iterator is very rare, so it does not seem worthwhile
2220 /* ??? An example failure is: i = 6; do {;} while (i++ < 9). */
2221 if (loop_dump_stream
)
2222 fprintf (loop_dump_stream
,
2223 "Loop unrolling: Giv iterators are not handled.\n");
2226 /* Initial value is mult_val times the biv's initial value plus
2227 add_val. Only useful if it is a constant. */
2228 v
= reg_iv_info
[REGNO (iteration_var
)];
2229 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
2230 *initial_value
= fold_rtx_mult_add (v
->mult_val
, bl
->initial_value
,
2231 v
->add_val
, v
->mode
);
2233 /* Increment value is mult_val times the increment value of the biv. */
2235 *increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2237 *increment
= fold_rtx_mult_add (v
->mult_val
, *increment
, const0_rtx
,
2243 if (loop_dump_stream
)
2244 fprintf (loop_dump_stream
,
2245 "Loop unrolling: Not basic or general induction var.\n");
2250 /* Calculate the approximate final value of the iteration variable
2251 which has an loop exit test with code COMPARISON_CODE and comparison value
2252 of COMPARISON_VALUE. Also returns an indication of whether the comparison
2253 was signed or unsigned, and the direction of the comparison. This info is
2254 needed to calculate the number of loop iterations. */
2257 approx_final_value (comparison_code
, comparison_value
, unsigned_p
, compare_dir
)
2258 enum rtx_code comparison_code
;
2259 rtx comparison_value
;
2263 /* Calculate the final value of the induction variable.
2264 The exact final value depends on the branch operator, and increment sign.
2265 This is only an approximate value. It will be wrong if the iteration
2266 variable is not incremented by one each time through the loop, and
2267 approx final value - start value % increment != 0. */
2270 switch (comparison_code
)
2276 return plus_constant (comparison_value
, 1);
2281 return plus_constant (comparison_value
, -1);
2283 /* Can not calculate a final value for this case. */
2290 return comparison_value
;
2296 return comparison_value
;
2299 return comparison_value
;
2305 /* For each biv and giv, determine whether it can be safely split into
2306 a different variable for each unrolled copy of the loop body. If it
2307 is safe to split, then indicate that by saving some useful info
2308 in the splittable_regs array.
2310 If the loop is being completely unrolled, then splittable_regs will hold
2311 the current value of the induction variable while the loop is unrolled.
2312 It must be set to the initial value of the induction variable here.
2313 Otherwise, splittable_regs will hold the difference between the current
2314 value of the induction variable and the value the induction variable had
2315 at the top of the loop. It must be set to the value 0 here.
2317 Returns the total number of instructions that set registers that are
2320 /* ?? If the loop is only unrolled twice, then most of the restrictions to
2321 constant values are unnecessary, since we can easily calculate increment
2322 values in this case even if nothing is constant. The increment value
2323 should not involve a multiply however. */
2325 /* ?? Even if the biv/giv increment values aren't constant, it may still
2326 be beneficial to split the variable if the loop is only unrolled a few
2327 times, since multiplies by small integers (1,2,3,4) are very cheap. */
2330 find_splittable_regs (unroll_type
, loop_start
, loop_end
, end_insert_before
,
2332 enum unroll_types unroll_type
;
2333 rtx loop_start
, loop_end
;
2334 rtx end_insert_before
;
2337 struct iv_class
*bl
;
2338 struct induction
*v
;
2340 rtx biv_final_value
;
2344 for (bl
= loop_iv_list
; bl
; bl
= bl
->next
)
2346 /* Biv_total_increment must return a constant value,
2347 otherwise we can not calculate the split values. */
2349 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2350 if (! increment
|| GET_CODE (increment
) != CONST_INT
)
2353 /* The loop must be unrolled completely, or else have a known number
2354 of iterations and only one exit, or else the biv must be dead
2355 outside the loop, or else the final value must be known. Otherwise,
2356 it is unsafe to split the biv since it may not have the proper
2357 value on loop exit. */
2359 /* loop_number_exit_labels is non-zero if the loop has an exit other than
2360 a fall through at the end. */
2363 biv_final_value
= 0;
2364 if (unroll_type
!= UNROLL_COMPLETELY
2365 && (loop_number_exit_labels
[uid_loop_num
[INSN_UID (loop_start
)]]
2366 || unroll_type
== UNROLL_NAIVE
)
2367 && (uid_luid
[regno_last_uid
[bl
->regno
]] >= INSN_LUID (loop_end
)
2369 || INSN_UID (bl
->init_insn
) >= max_uid_for_loop
2370 || (uid_luid
[regno_first_uid
[bl
->regno
]]
2371 < INSN_LUID (bl
->init_insn
))
2372 || reg_mentioned_p (bl
->biv
->dest_reg
, SET_SRC (bl
->init_set
)))
2373 && ! (biv_final_value
= final_biv_value (bl
, loop_start
, loop_end
)))
2376 /* If any of the insns setting the BIV don't do so with a simple
2377 PLUS, we don't know how to split it. */
2378 for (v
= bl
->biv
; biv_splittable
&& v
; v
= v
->next_iv
)
2379 if ((tem
= single_set (v
->insn
)) == 0
2380 || GET_CODE (SET_DEST (tem
)) != REG
2381 || REGNO (SET_DEST (tem
)) != bl
->regno
2382 || GET_CODE (SET_SRC (tem
)) != PLUS
)
2385 /* If final value is non-zero, then must emit an instruction which sets
2386 the value of the biv to the proper value. This is done after
2387 handling all of the givs, since some of them may need to use the
2388 biv's value in their initialization code. */
2390 /* This biv is splittable. If completely unrolling the loop, save
2391 the biv's initial value. Otherwise, save the constant zero. */
2393 if (biv_splittable
== 1)
2395 if (unroll_type
== UNROLL_COMPLETELY
)
2397 /* If the initial value of the biv is itself (i.e. it is too
2398 complicated for strength_reduce to compute), or is a hard
2399 register, then we must create a new pseudo reg to hold the
2400 initial value of the biv. */
2402 if (GET_CODE (bl
->initial_value
) == REG
2403 && (REGNO (bl
->initial_value
) == bl
->regno
2404 || REGNO (bl
->initial_value
) < FIRST_PSEUDO_REGISTER
))
2406 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2408 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2411 if (loop_dump_stream
)
2412 fprintf (loop_dump_stream
, "Biv %d initial value remapped to %d.\n",
2413 bl
->regno
, REGNO (tem
));
2415 splittable_regs
[bl
->regno
] = tem
;
2418 splittable_regs
[bl
->regno
] = bl
->initial_value
;
2421 splittable_regs
[bl
->regno
] = const0_rtx
;
2423 /* Save the number of instructions that modify the biv, so that
2424 we can treat the last one specially. */
2426 splittable_regs_updates
[bl
->regno
] = bl
->biv_count
;
2427 result
+= bl
->biv_count
;
2429 if (loop_dump_stream
)
2430 fprintf (loop_dump_stream
,
2431 "Biv %d safe to split.\n", bl
->regno
);
2434 /* Check every giv that depends on this biv to see whether it is
2435 splittable also. Even if the biv isn't splittable, givs which
2436 depend on it may be splittable if the biv is live outside the
2437 loop, and the givs aren't. */
2439 result
+= find_splittable_givs (bl
, unroll_type
, loop_start
, loop_end
,
2440 increment
, unroll_number
);
2442 /* If final value is non-zero, then must emit an instruction which sets
2443 the value of the biv to the proper value. This is done after
2444 handling all of the givs, since some of them may need to use the
2445 biv's value in their initialization code. */
2446 if (biv_final_value
)
2448 /* If the loop has multiple exits, emit the insns before the
2449 loop to ensure that it will always be executed no matter
2450 how the loop exits. Otherwise emit the insn after the loop,
2451 since this is slightly more efficient. */
2452 if (! loop_number_exit_labels
[uid_loop_num
[INSN_UID (loop_start
)]])
2453 emit_insn_before (gen_move_insn (bl
->biv
->src_reg
,
2458 /* Create a new register to hold the value of the biv, and then
2459 set the biv to its final value before the loop start. The biv
2460 is set to its final value before loop start to ensure that
2461 this insn will always be executed, no matter how the loop
2463 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2464 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2466 emit_insn_before (gen_move_insn (bl
->biv
->src_reg
,
2470 if (loop_dump_stream
)
2471 fprintf (loop_dump_stream
, "Biv %d mapped to %d for split.\n",
2472 REGNO (bl
->biv
->src_reg
), REGNO (tem
));
2474 /* Set up the mapping from the original biv register to the new
2476 bl
->biv
->src_reg
= tem
;
2483 /* Return 1 if the first and last unrolled copy of the address giv V is valid
2484 for the instruction that is using it. Do not make any changes to that
2488 verify_addresses (v
, giv_inc
, unroll_number
)
2489 struct induction
*v
;
2494 rtx orig_addr
= *v
->location
;
2495 rtx last_addr
= plus_constant (v
->dest_reg
,
2496 INTVAL (giv_inc
) * (unroll_number
- 1));
2498 /* First check to see if either address would fail. */
2499 if (! validate_change (v
->insn
, v
->location
, v
->dest_reg
, 0)
2500 || ! validate_change (v
->insn
, v
->location
, last_addr
, 0))
2503 /* Now put things back the way they were before. This will always
2505 validate_change (v
->insn
, v
->location
, orig_addr
, 0);
2510 /* For every giv based on the biv BL, check to determine whether it is
2511 splittable. This is a subroutine to find_splittable_regs ().
2513 Return the number of instructions that set splittable registers. */
2516 find_splittable_givs (bl
, unroll_type
, loop_start
, loop_end
, increment
,
2518 struct iv_class
*bl
;
2519 enum unroll_types unroll_type
;
2520 rtx loop_start
, loop_end
;
2524 struct induction
*v
, *v2
;
2529 /* Scan the list of givs, and set the same_insn field when there are
2530 multiple identical givs in the same insn. */
2531 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
2532 for (v2
= v
->next_iv
; v2
; v2
= v2
->next_iv
)
2533 if (v
->insn
== v2
->insn
&& rtx_equal_p (v
->new_reg
, v2
->new_reg
)
2537 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
2541 /* Only split the giv if it has already been reduced, or if the loop is
2542 being completely unrolled. */
2543 if (unroll_type
!= UNROLL_COMPLETELY
&& v
->ignore
)
2546 /* The giv can be split if the insn that sets the giv is executed once
2547 and only once on every iteration of the loop. */
2548 /* An address giv can always be split. v->insn is just a use not a set,
2549 and hence it does not matter whether it is always executed. All that
2550 matters is that all the biv increments are always executed, and we
2551 won't reach here if they aren't. */
2552 if (v
->giv_type
!= DEST_ADDR
2553 && (! v
->always_computable
2554 || back_branch_in_range_p (v
->insn
, loop_start
, loop_end
)))
2557 /* The giv increment value must be a constant. */
2558 giv_inc
= fold_rtx_mult_add (v
->mult_val
, increment
, const0_rtx
,
2560 if (! giv_inc
|| GET_CODE (giv_inc
) != CONST_INT
)
2563 /* The loop must be unrolled completely, or else have a known number of
2564 iterations and only one exit, or else the giv must be dead outside
2565 the loop, or else the final value of the giv must be known.
2566 Otherwise, it is not safe to split the giv since it may not have the
2567 proper value on loop exit. */
2569 /* The used outside loop test will fail for DEST_ADDR givs. They are
2570 never used outside the loop anyways, so it is always safe to split a
2574 if (unroll_type
!= UNROLL_COMPLETELY
2575 && (loop_number_exit_labels
[uid_loop_num
[INSN_UID (loop_start
)]]
2576 || unroll_type
== UNROLL_NAIVE
)
2577 && v
->giv_type
!= DEST_ADDR
2578 && ((regno_first_uid
[REGNO (v
->dest_reg
)] != INSN_UID (v
->insn
)
2579 /* Check for the case where the pseudo is set by a shift/add
2580 sequence, in which case the first insn setting the pseudo
2581 is the first insn of the shift/add sequence. */
2582 && (! (tem
= find_reg_note (v
->insn
, REG_RETVAL
, NULL_RTX
))
2583 || (regno_first_uid
[REGNO (v
->dest_reg
)]
2584 != INSN_UID (XEXP (tem
, 0)))))
2585 /* Line above always fails if INSN was moved by loop opt. */
2586 || (uid_luid
[regno_last_uid
[REGNO (v
->dest_reg
)]]
2587 >= INSN_LUID (loop_end
)))
2588 && ! (final_value
= v
->final_value
))
2592 /* Currently, non-reduced/final-value givs are never split. */
2593 /* Should emit insns after the loop if possible, as the biv final value
2596 /* If the final value is non-zero, and the giv has not been reduced,
2597 then must emit an instruction to set the final value. */
2598 if (final_value
&& !v
->new_reg
)
2600 /* Create a new register to hold the value of the giv, and then set
2601 the giv to its final value before the loop start. The giv is set
2602 to its final value before loop start to ensure that this insn
2603 will always be executed, no matter how we exit. */
2604 tem
= gen_reg_rtx (v
->mode
);
2605 emit_insn_before (gen_move_insn (tem
, v
->dest_reg
), loop_start
);
2606 emit_insn_before (gen_move_insn (v
->dest_reg
, final_value
),
2609 if (loop_dump_stream
)
2610 fprintf (loop_dump_stream
, "Giv %d mapped to %d for split.\n",
2611 REGNO (v
->dest_reg
), REGNO (tem
));
2617 /* This giv is splittable. If completely unrolling the loop, save the
2618 giv's initial value. Otherwise, save the constant zero for it. */
2620 if (unroll_type
== UNROLL_COMPLETELY
)
2622 /* It is not safe to use bl->initial_value here, because it may not
2623 be invariant. It is safe to use the initial value stored in
2624 the splittable_regs array if it is set. In rare cases, it won't
2625 be set, so then we do exactly the same thing as
2626 find_splittable_regs does to get a safe value. */
2627 rtx biv_initial_value
;
2629 if (splittable_regs
[bl
->regno
])
2630 biv_initial_value
= splittable_regs
[bl
->regno
];
2631 else if (GET_CODE (bl
->initial_value
) != REG
2632 || (REGNO (bl
->initial_value
) != bl
->regno
2633 && REGNO (bl
->initial_value
) >= FIRST_PSEUDO_REGISTER
))
2634 biv_initial_value
= bl
->initial_value
;
2637 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2639 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2641 biv_initial_value
= tem
;
2643 value
= fold_rtx_mult_add (v
->mult_val
, biv_initial_value
,
2644 v
->add_val
, v
->mode
);
2651 /* If a giv was combined with another giv, then we can only split
2652 this giv if the giv it was combined with was reduced. This
2653 is because the value of v->new_reg is meaningless in this
2655 if (v
->same
&& ! v
->same
->new_reg
)
2657 if (loop_dump_stream
)
2658 fprintf (loop_dump_stream
,
2659 "giv combined with unreduced giv not split.\n");
2662 /* If the giv is an address destination, it could be something other
2663 than a simple register, these have to be treated differently. */
2664 else if (v
->giv_type
== DEST_REG
)
2666 /* If value is not a constant, register, or register plus
2667 constant, then compute its value into a register before
2668 loop start. This prevents invalid rtx sharing, and should
2669 generate better code. We can use bl->initial_value here
2670 instead of splittable_regs[bl->regno] because this code
2671 is going before the loop start. */
2672 if (unroll_type
== UNROLL_COMPLETELY
2673 && GET_CODE (value
) != CONST_INT
2674 && GET_CODE (value
) != REG
2675 && (GET_CODE (value
) != PLUS
2676 || GET_CODE (XEXP (value
, 0)) != REG
2677 || GET_CODE (XEXP (value
, 1)) != CONST_INT
))
2679 rtx tem
= gen_reg_rtx (v
->mode
);
2680 emit_iv_add_mult (bl
->initial_value
, v
->mult_val
,
2681 v
->add_val
, tem
, loop_start
);
2685 splittable_regs
[REGNO (v
->new_reg
)] = value
;
2689 /* Splitting address givs is useful since it will often allow us
2690 to eliminate some increment insns for the base giv as
2693 /* If the addr giv is combined with a dest_reg giv, then all
2694 references to that dest reg will be remapped, which is NOT
2695 what we want for split addr regs. We always create a new
2696 register for the split addr giv, just to be safe. */
2698 /* ??? If there are multiple address givs which have been
2699 combined with the same dest_reg giv, then we may only need
2700 one new register for them. Pulling out constants below will
2701 catch some of the common cases of this. Currently, I leave
2702 the work of simplifying multiple address givs to the
2703 following cse pass. */
2705 /* As a special case, if we have multiple identical address givs
2706 within a single instruction, then we do use a single pseudo
2707 reg for both. This is necessary in case one is a match_dup
2710 v
->const_adjust
= 0;
2714 v
->dest_reg
= v
->same_insn
->dest_reg
;
2715 if (loop_dump_stream
)
2716 fprintf (loop_dump_stream
,
2717 "Sharing address givs in insn %d\n",
2718 INSN_UID (v
->insn
));
2720 else if (unroll_type
!= UNROLL_COMPLETELY
)
2722 /* If not completely unrolling the loop, then create a new
2723 register to hold the split value of the DEST_ADDR giv.
2724 Emit insn to initialize its value before loop start. */
2725 tem
= gen_reg_rtx (v
->mode
);
2727 /* If the address giv has a constant in its new_reg value,
2728 then this constant can be pulled out and put in value,
2729 instead of being part of the initialization code. */
2731 if (GET_CODE (v
->new_reg
) == PLUS
2732 && GET_CODE (XEXP (v
->new_reg
, 1)) == CONST_INT
)
2735 = plus_constant (tem
, INTVAL (XEXP (v
->new_reg
,1)));
2737 /* Only succeed if this will give valid addresses.
2738 Try to validate both the first and the last
2739 address resulting from loop unrolling, if
2740 one fails, then can't do const elim here. */
2741 if (! verify_addresses (v
, giv_inc
, unroll_number
))
2743 /* Save the negative of the eliminated const, so
2744 that we can calculate the dest_reg's increment
2746 v
->const_adjust
= - INTVAL (XEXP (v
->new_reg
, 1));
2748 v
->new_reg
= XEXP (v
->new_reg
, 0);
2749 if (loop_dump_stream
)
2750 fprintf (loop_dump_stream
,
2751 "Eliminating constant from giv %d\n",
2760 /* If the address hasn't been checked for validity yet, do so
2761 now, and fail completely if either the first or the last
2762 unrolled copy of the address is not a valid address
2763 for the instruction that uses it. */
2764 if (v
->dest_reg
== tem
2765 && ! verify_addresses (v
, giv_inc
, unroll_number
))
2767 if (loop_dump_stream
)
2768 fprintf (loop_dump_stream
,
2769 "Invalid address for giv at insn %d\n",
2770 INSN_UID (v
->insn
));
2774 /* To initialize the new register, just move the value of
2775 new_reg into it. This is not guaranteed to give a valid
2776 instruction on machines with complex addressing modes.
2777 If we can't recognize it, then delete it and emit insns
2778 to calculate the value from scratch. */
2779 emit_insn_before (gen_rtx (SET
, VOIDmode
, tem
,
2780 copy_rtx (v
->new_reg
)),
2782 if (recog_memoized (PREV_INSN (loop_start
)) < 0)
2786 /* We can't use bl->initial_value to compute the initial
2787 value, because the loop may have been preconditioned.
2788 We must calculate it from NEW_REG. Try using
2789 force_operand instead of emit_iv_add_mult. */
2790 delete_insn (PREV_INSN (loop_start
));
2793 ret
= force_operand (v
->new_reg
, tem
);
2795 emit_move_insn (tem
, ret
);
2796 sequence
= gen_sequence ();
2798 emit_insn_before (sequence
, loop_start
);
2800 if (loop_dump_stream
)
2801 fprintf (loop_dump_stream
,
2802 "Invalid init insn, rewritten.\n");
2807 v
->dest_reg
= value
;
2809 /* Check the resulting address for validity, and fail
2810 if the resulting address would be invalid. */
2811 if (! verify_addresses (v
, giv_inc
, unroll_number
))
2813 if (loop_dump_stream
)
2814 fprintf (loop_dump_stream
,
2815 "Invalid address for giv at insn %d\n",
2816 INSN_UID (v
->insn
));
2821 /* Store the value of dest_reg into the insn. This sharing
2822 will not be a problem as this insn will always be copied
2825 *v
->location
= v
->dest_reg
;
2827 /* If this address giv is combined with a dest reg giv, then
2828 save the base giv's induction pointer so that we will be
2829 able to handle this address giv properly. The base giv
2830 itself does not have to be splittable. */
2832 if (v
->same
&& v
->same
->giv_type
== DEST_REG
)
2833 addr_combined_regs
[REGNO (v
->same
->new_reg
)] = v
->same
;
2835 if (GET_CODE (v
->new_reg
) == REG
)
2837 /* This giv maybe hasn't been combined with any others.
2838 Make sure that it's giv is marked as splittable here. */
2840 splittable_regs
[REGNO (v
->new_reg
)] = value
;
2842 /* Make it appear to depend upon itself, so that the
2843 giv will be properly split in the main loop above. */
2847 addr_combined_regs
[REGNO (v
->new_reg
)] = v
;
2851 if (loop_dump_stream
)
2852 fprintf (loop_dump_stream
, "DEST_ADDR giv being split.\n");
2858 /* Currently, unreduced giv's can't be split. This is not too much
2859 of a problem since unreduced giv's are not live across loop
2860 iterations anyways. When unrolling a loop completely though,
2861 it makes sense to reduce&split givs when possible, as this will
2862 result in simpler instructions, and will not require that a reg
2863 be live across loop iterations. */
2865 splittable_regs
[REGNO (v
->dest_reg
)] = value
;
2866 fprintf (stderr
, "Giv %d at insn %d not reduced\n",
2867 REGNO (v
->dest_reg
), INSN_UID (v
->insn
));
2873 /* Givs are only updated once by definition. Mark it so if this is
2874 a splittable register. Don't need to do anything for address givs
2875 where this may not be a register. */
2877 if (GET_CODE (v
->new_reg
) == REG
)
2878 splittable_regs_updates
[REGNO (v
->new_reg
)] = 1;
2882 if (loop_dump_stream
)
2886 if (GET_CODE (v
->dest_reg
) == CONST_INT
)
2888 else if (GET_CODE (v
->dest_reg
) != REG
)
2889 regnum
= REGNO (XEXP (v
->dest_reg
, 0));
2891 regnum
= REGNO (v
->dest_reg
);
2892 fprintf (loop_dump_stream
, "Giv %d at insn %d safe to split.\n",
2893 regnum
, INSN_UID (v
->insn
));
2900 /* Try to prove that the register is dead after the loop exits. Trace every
2901 loop exit looking for an insn that will always be executed, which sets
2902 the register to some value, and appears before the first use of the register
2903 is found. If successful, then return 1, otherwise return 0. */
2905 /* ?? Could be made more intelligent in the handling of jumps, so that
2906 it can search past if statements and other similar structures. */
2909 reg_dead_after_loop (reg
, loop_start
, loop_end
)
2910 rtx reg
, loop_start
, loop_end
;
2916 /* HACK: Must also search the loop fall through exit, create a label_ref
2917 here which points to the loop_end, and append the loop_number_exit_labels
2919 label
= gen_rtx (LABEL_REF
, VOIDmode
, loop_end
);
2920 LABEL_NEXTREF (label
)
2921 = loop_number_exit_labels
[uid_loop_num
[INSN_UID (loop_start
)]];
2923 for ( ; label
; label
= LABEL_NEXTREF (label
))
2925 /* Succeed if find an insn which sets the biv or if reach end of
2926 function. Fail if find an insn that uses the biv, or if come to
2927 a conditional jump. */
2929 insn
= NEXT_INSN (XEXP (label
, 0));
2932 code
= GET_CODE (insn
);
2933 if (GET_RTX_CLASS (code
) == 'i')
2937 if (reg_referenced_p (reg
, PATTERN (insn
)))
2940 set
= single_set (insn
);
2941 if (set
&& rtx_equal_p (SET_DEST (set
), reg
))
2945 if (code
== JUMP_INSN
)
2947 if (GET_CODE (PATTERN (insn
)) == RETURN
)
2949 else if (! simplejump_p (insn
)
2950 /* Prevent infinite loop following infinite loops. */
2951 || jump_count
++ > 20)
2954 insn
= JUMP_LABEL (insn
);
2957 insn
= NEXT_INSN (insn
);
2961 /* Success, the register is dead on all loop exits. */
2965 /* Try to calculate the final value of the biv, the value it will have at
2966 the end of the loop. If we can do it, return that value. */
2969 final_biv_value (bl
, loop_start
, loop_end
)
2970 struct iv_class
*bl
;
2971 rtx loop_start
, loop_end
;
2975 /* ??? This only works for MODE_INT biv's. Reject all others for now. */
2977 if (GET_MODE_CLASS (bl
->biv
->mode
) != MODE_INT
)
2980 /* The final value for reversed bivs must be calculated differently than
2981 for ordinary bivs. In this case, there is already an insn after the
2982 loop which sets this biv's final value (if necessary), and there are
2983 no other loop exits, so we can return any value. */
2986 if (loop_dump_stream
)
2987 fprintf (loop_dump_stream
,
2988 "Final biv value for %d, reversed biv.\n", bl
->regno
);
2993 /* Try to calculate the final value as initial value + (number of iterations
2994 * increment). For this to work, increment must be invariant, the only
2995 exit from the loop must be the fall through at the bottom (otherwise
2996 it may not have its final value when the loop exits), and the initial
2997 value of the biv must be invariant. */
2999 if (loop_n_iterations
!= 0
3000 && ! loop_number_exit_labels
[uid_loop_num
[INSN_UID (loop_start
)]]
3001 && invariant_p (bl
->initial_value
))
3003 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
3005 if (increment
&& invariant_p (increment
))
3007 /* Can calculate the loop exit value, emit insns after loop
3008 end to calculate this value into a temporary register in
3009 case it is needed later. */
3011 tem
= gen_reg_rtx (bl
->biv
->mode
);
3012 /* Make sure loop_end is not the last insn. */
3013 if (NEXT_INSN (loop_end
) == 0)
3014 emit_note_after (NOTE_INSN_DELETED
, loop_end
);
3015 emit_iv_add_mult (increment
, GEN_INT (loop_n_iterations
),
3016 bl
->initial_value
, tem
, NEXT_INSN (loop_end
));
3018 if (loop_dump_stream
)
3019 fprintf (loop_dump_stream
,
3020 "Final biv value for %d, calculated.\n", bl
->regno
);
3026 /* Check to see if the biv is dead at all loop exits. */
3027 if (reg_dead_after_loop (bl
->biv
->src_reg
, loop_start
, loop_end
))
3029 if (loop_dump_stream
)
3030 fprintf (loop_dump_stream
,
3031 "Final biv value for %d, biv dead after loop exit.\n",
3040 /* Try to calculate the final value of the giv, the value it will have at
3041 the end of the loop. If we can do it, return that value. */
3044 final_giv_value (v
, loop_start
, loop_end
)
3045 struct induction
*v
;
3046 rtx loop_start
, loop_end
;
3048 struct iv_class
*bl
;
3051 rtx insert_before
, seq
;
3053 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
3055 /* The final value for givs which depend on reversed bivs must be calculated
3056 differently than for ordinary givs. In this case, there is already an
3057 insn after the loop which sets this giv's final value (if necessary),
3058 and there are no other loop exits, so we can return any value. */
3061 if (loop_dump_stream
)
3062 fprintf (loop_dump_stream
,
3063 "Final giv value for %d, depends on reversed biv\n",
3064 REGNO (v
->dest_reg
));
3068 /* Try to calculate the final value as a function of the biv it depends
3069 upon. The only exit from the loop must be the fall through at the bottom
3070 (otherwise it may not have its final value when the loop exits). */
3072 /* ??? Can calculate the final giv value by subtracting off the
3073 extra biv increments times the giv's mult_val. The loop must have
3074 only one exit for this to work, but the loop iterations does not need
3077 if (loop_n_iterations
!= 0
3078 && ! loop_number_exit_labels
[uid_loop_num
[INSN_UID (loop_start
)]])
3080 /* ?? It is tempting to use the biv's value here since these insns will
3081 be put after the loop, and hence the biv will have its final value
3082 then. However, this fails if the biv is subsequently eliminated.
3083 Perhaps determine whether biv's are eliminable before trying to
3084 determine whether giv's are replaceable so that we can use the
3085 biv value here if it is not eliminable. */
3087 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
3089 if (increment
&& invariant_p (increment
))
3091 /* Can calculate the loop exit value of its biv as
3092 (loop_n_iterations * increment) + initial_value */
3094 /* The loop exit value of the giv is then
3095 (final_biv_value - extra increments) * mult_val + add_val.
3096 The extra increments are any increments to the biv which
3097 occur in the loop after the giv's value is calculated.
3098 We must search from the insn that sets the giv to the end
3099 of the loop to calculate this value. */
3101 insert_before
= NEXT_INSN (loop_end
);
3103 /* Put the final biv value in tem. */
3104 tem
= gen_reg_rtx (bl
->biv
->mode
);
3105 emit_iv_add_mult (increment
, GEN_INT (loop_n_iterations
),
3106 bl
->initial_value
, tem
, insert_before
);
3108 /* Subtract off extra increments as we find them. */
3109 for (insn
= NEXT_INSN (v
->insn
); insn
!= loop_end
;
3110 insn
= NEXT_INSN (insn
))
3112 struct induction
*biv
;
3114 for (biv
= bl
->biv
; biv
; biv
= biv
->next_iv
)
3115 if (biv
->insn
== insn
)
3118 tem
= expand_binop (GET_MODE (tem
), sub_optab
, tem
,
3119 biv
->add_val
, NULL_RTX
, 0,
3121 seq
= gen_sequence ();
3123 emit_insn_before (seq
, insert_before
);
3127 /* Now calculate the giv's final value. */
3128 emit_iv_add_mult (tem
, v
->mult_val
, v
->add_val
, tem
,
3131 if (loop_dump_stream
)
3132 fprintf (loop_dump_stream
,
3133 "Final giv value for %d, calc from biv's value.\n",
3134 REGNO (v
->dest_reg
));
3140 /* Replaceable giv's should never reach here. */
3144 /* Check to see if the biv is dead at all loop exits. */
3145 if (reg_dead_after_loop (v
->dest_reg
, loop_start
, loop_end
))
3147 if (loop_dump_stream
)
3148 fprintf (loop_dump_stream
,
3149 "Final giv value for %d, giv dead after loop exit.\n",
3150 REGNO (v
->dest_reg
));
3159 /* Calculate the number of loop iterations. Returns the exact number of loop
3160 iterations if it can be calculated, otherwise returns zero. */
3162 unsigned HOST_WIDE_INT
3163 loop_iterations (loop_start
, loop_end
)
3164 rtx loop_start
, loop_end
;
3166 rtx comparison
, comparison_value
;
3167 rtx iteration_var
, initial_value
, increment
, final_value
;
3168 enum rtx_code comparison_code
;
3171 int unsigned_compare
, compare_dir
, final_larger
;
3172 unsigned long tempu
;
3175 /* First find the iteration variable. If the last insn is a conditional
3176 branch, and the insn before tests a register value, make that the
3177 iteration variable. */
3179 loop_initial_value
= 0;
3181 loop_final_value
= 0;
3182 loop_iteration_var
= 0;
3184 /* We used to use pren_nonnote_insn here, but that fails because it might
3185 accidentally get the branch for a contained loop if the branch for this
3186 loop was deleted. We can only trust branches immediately before the
3188 last_loop_insn
= PREV_INSN (loop_end
);
3190 comparison
= get_condition_for_loop (last_loop_insn
);
3191 if (comparison
== 0)
3193 if (loop_dump_stream
)
3194 fprintf (loop_dump_stream
,
3195 "Loop unrolling: No final conditional branch found.\n");
3199 /* ??? Get_condition may switch position of induction variable and
3200 invariant register when it canonicalizes the comparison. */
3202 comparison_code
= GET_CODE (comparison
);
3203 iteration_var
= XEXP (comparison
, 0);
3204 comparison_value
= XEXP (comparison
, 1);
3206 if (GET_CODE (iteration_var
) != REG
)
3208 if (loop_dump_stream
)
3209 fprintf (loop_dump_stream
,
3210 "Loop unrolling: Comparison not against register.\n");
3214 /* Loop iterations is always called before any new registers are created
3215 now, so this should never occur. */
3217 if (REGNO (iteration_var
) >= max_reg_before_loop
)
3220 iteration_info (iteration_var
, &initial_value
, &increment
,
3221 loop_start
, loop_end
);
3222 if (initial_value
== 0)
3223 /* iteration_info already printed a message. */
3226 /* If the comparison value is an invariant register, then try to find
3227 its value from the insns before the start of the loop. */
3229 if (GET_CODE (comparison_value
) == REG
&& invariant_p (comparison_value
))
3233 for (insn
= PREV_INSN (loop_start
); insn
; insn
= PREV_INSN (insn
))
3235 if (GET_CODE (insn
) == CODE_LABEL
)
3238 else if (GET_RTX_CLASS (GET_CODE (insn
)) == 'i'
3239 && reg_set_p (comparison_value
, insn
))
3241 /* We found the last insn before the loop that sets the register.
3242 If it sets the entire register, and has a REG_EQUAL note,
3243 then use the value of the REG_EQUAL note. */
3244 if ((set
= single_set (insn
))
3245 && (SET_DEST (set
) == comparison_value
))
3247 rtx note
= find_reg_note (insn
, REG_EQUAL
, NULL_RTX
);
3249 /* Only use the REG_EQUAL note if it is a constant.
3250 Other things, divide in particular, will cause
3251 problems later if we use them. */
3252 if (note
&& GET_CODE (XEXP (note
, 0)) != EXPR_LIST
3253 && CONSTANT_P (XEXP (note
, 0)))
3254 comparison_value
= XEXP (note
, 0);
3261 final_value
= approx_final_value (comparison_code
, comparison_value
,
3262 &unsigned_compare
, &compare_dir
);
3264 /* Save the calculated values describing this loop's bounds, in case
3265 precondition_loop_p will need them later. These values can not be
3266 recalculated inside precondition_loop_p because strength reduction
3267 optimizations may obscure the loop's structure. */
3269 loop_iteration_var
= iteration_var
;
3270 loop_initial_value
= initial_value
;
3271 loop_increment
= increment
;
3272 loop_final_value
= final_value
;
3276 if (loop_dump_stream
)
3277 fprintf (loop_dump_stream
,
3278 "Loop unrolling: Increment value can't be calculated.\n");
3281 else if (GET_CODE (increment
) != CONST_INT
)
3283 if (loop_dump_stream
)
3284 fprintf (loop_dump_stream
,
3285 "Loop unrolling: Increment value not constant.\n");
3288 else if (GET_CODE (initial_value
) != CONST_INT
)
3290 if (loop_dump_stream
)
3291 fprintf (loop_dump_stream
,
3292 "Loop unrolling: Initial value not constant.\n");
3295 else if (final_value
== 0)
3297 if (loop_dump_stream
)
3298 fprintf (loop_dump_stream
,
3299 "Loop unrolling: EQ comparison loop.\n");
3302 else if (GET_CODE (final_value
) != CONST_INT
)
3304 if (loop_dump_stream
)
3305 fprintf (loop_dump_stream
,
3306 "Loop unrolling: Final value not constant.\n");
3310 /* ?? Final value and initial value do not have to be constants.
3311 Only their difference has to be constant. When the iteration variable
3312 is an array address, the final value and initial value might both
3313 be addresses with the same base but different constant offsets.
3314 Final value must be invariant for this to work.
3316 To do this, need some way to find the values of registers which are
3319 /* Final_larger is 1 if final larger, 0 if they are equal, otherwise -1. */
3320 if (unsigned_compare
)
3322 = ((unsigned HOST_WIDE_INT
) INTVAL (final_value
)
3323 > (unsigned HOST_WIDE_INT
) INTVAL (initial_value
))
3324 - ((unsigned HOST_WIDE_INT
) INTVAL (final_value
)
3325 < (unsigned HOST_WIDE_INT
) INTVAL (initial_value
));
3327 final_larger
= (INTVAL (final_value
) > INTVAL (initial_value
))
3328 - (INTVAL (final_value
) < INTVAL (initial_value
));
3330 if (INTVAL (increment
) > 0)
3332 else if (INTVAL (increment
) == 0)
3337 /* There are 27 different cases: compare_dir = -1, 0, 1;
3338 final_larger = -1, 0, 1; increment_dir = -1, 0, 1.
3339 There are 4 normal cases, 4 reverse cases (where the iteration variable
3340 will overflow before the loop exits), 4 infinite loop cases, and 15
3341 immediate exit (0 or 1 iteration depending on loop type) cases.
3342 Only try to optimize the normal cases. */
3344 /* (compare_dir/final_larger/increment_dir)
3345 Normal cases: (0/-1/-1), (0/1/1), (-1/-1/-1), (1/1/1)
3346 Reverse cases: (0/-1/1), (0/1/-1), (-1/-1/1), (1/1/-1)
3347 Infinite loops: (0/-1/0), (0/1/0), (-1/-1/0), (1/1/0)
3348 Immediate exit: (0/0/X), (-1/0/X), (-1/1/X), (1/0/X), (1/-1/X) */
3350 /* ?? If the meaning of reverse loops (where the iteration variable
3351 will overflow before the loop exits) is undefined, then could
3352 eliminate all of these special checks, and just always assume
3353 the loops are normal/immediate/infinite. Note that this means
3354 the sign of increment_dir does not have to be known. Also,
3355 since it does not really hurt if immediate exit loops or infinite loops
3356 are optimized, then that case could be ignored also, and hence all
3357 loops can be optimized.
3359 According to ANSI Spec, the reverse loop case result is undefined,
3360 because the action on overflow is undefined.
3362 See also the special test for NE loops below. */
3364 if (final_larger
== increment_dir
&& final_larger
!= 0
3365 && (final_larger
== compare_dir
|| compare_dir
== 0))
3370 if (loop_dump_stream
)
3371 fprintf (loop_dump_stream
,
3372 "Loop unrolling: Not normal loop.\n");
3376 /* Calculate the number of iterations, final_value is only an approximation,
3377 so correct for that. Note that tempu and loop_n_iterations are
3378 unsigned, because they can be as large as 2^n - 1. */
3380 i
= INTVAL (increment
);
3382 tempu
= INTVAL (final_value
) - INTVAL (initial_value
);
3385 tempu
= INTVAL (initial_value
) - INTVAL (final_value
);
3391 /* For NE tests, make sure that the iteration variable won't miss the
3392 final value. If tempu mod i is not zero, then the iteration variable
3393 will overflow before the loop exits, and we can not calculate the
3394 number of iterations. */
3395 if (compare_dir
== 0 && (tempu
% i
) != 0)
3398 return tempu
/ i
+ ((tempu
% i
) != 0);
3401 /* Replace uses of split bivs with their split pseudo register. This is
3402 for original instructions which remain after loop unrolling without
3406 remap_split_bivs (x
)
3409 register enum rtx_code code
;
3416 code
= GET_CODE (x
);
3431 /* If non-reduced/final-value givs were split, then this would also
3432 have to remap those givs also. */
3434 if (REGNO (x
) < max_reg_before_loop
3435 && reg_iv_type
[REGNO (x
)] == BASIC_INDUCT
)
3436 return reg_biv_class
[REGNO (x
)]->biv
->src_reg
;
3439 fmt
= GET_RTX_FORMAT (code
);
3440 for (i
= GET_RTX_LENGTH (code
) - 1; i
>= 0; i
--)
3443 XEXP (x
, i
) = remap_split_bivs (XEXP (x
, i
));
3447 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
3448 XVECEXP (x
, i
, j
) = remap_split_bivs (XVECEXP (x
, i
, j
));