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
)
739 for (j
= FIRST_PSEUDO_REGISTER
; j
< max_reg_before_loop
; ++j
)
740 if (regno_first_uid
[j
] > 0 && regno_first_uid
[j
] <= max_uid_for_loop
741 && uid_luid
[regno_first_uid
[j
]] >= copy_start_luid
742 && regno_last_uid
[j
] > 0 && regno_last_uid
[j
] <= max_uid_for_loop
743 && uid_luid
[regno_last_uid
[j
]] <= copy_end_luid
)
747 /* If this loop requires exit tests when unrolled, check to see if we
748 can precondition the loop so as to make the exit tests unnecessary.
749 Just like variable splitting, this is not safe if the loop is entered
750 via a jump to the bottom. Also, can not do this if no strength
751 reduce info, because precondition_loop_p uses this info. */
753 /* Must copy the loop body for preconditioning before the following
754 find_splittable_regs call since that will emit insns which need to
755 be after the preconditioned loop copies, but immediately before the
756 unrolled loop copies. */
758 /* Also, it is not safe to split induction variables for the preconditioned
759 copies of the loop body. If we split induction variables, then the code
760 assumes that each induction variable can be represented as a function
761 of its initial value and the loop iteration number. This is not true
762 in this case, because the last preconditioned copy of the loop body
763 could be any iteration from the first up to the `unroll_number-1'th,
764 depending on the initial value of the iteration variable. Therefore
765 we can not split induction variables here, because we can not calculate
766 their value. Hence, this code must occur before find_splittable_regs
769 if (unroll_type
== UNROLL_NAIVE
&& ! splitting_not_safe
&& strength_reduce_p
)
771 rtx initial_value
, final_value
, increment
;
773 if (precondition_loop_p (&initial_value
, &final_value
, &increment
,
774 loop_start
, loop_end
))
776 register rtx diff
, temp
;
777 enum machine_mode mode
;
779 int abs_inc
, neg_inc
;
781 map
->reg_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
783 map
->const_equiv_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
784 map
->const_age_map
= (unsigned *) alloca (maxregnum
785 * sizeof (unsigned));
786 map
->const_equiv_map_size
= maxregnum
;
787 global_const_equiv_map
= map
->const_equiv_map
;
788 global_const_equiv_map_size
= maxregnum
;
790 init_reg_map (map
, maxregnum
);
792 /* Limit loop unrolling to 4, since this will make 7 copies of
794 if (unroll_number
> 4)
797 /* Save the absolute value of the increment, and also whether or
798 not it is negative. */
800 abs_inc
= INTVAL (increment
);
809 /* Decide what mode to do these calculations in. Choose the larger
810 of final_value's mode and initial_value's mode, or a full-word if
811 both are constants. */
812 mode
= GET_MODE (final_value
);
813 if (mode
== VOIDmode
)
815 mode
= GET_MODE (initial_value
);
816 if (mode
== VOIDmode
)
819 else if (mode
!= GET_MODE (initial_value
)
820 && (GET_MODE_SIZE (mode
)
821 < GET_MODE_SIZE (GET_MODE (initial_value
))))
822 mode
= GET_MODE (initial_value
);
824 /* Calculate the difference between the final and initial values.
825 Final value may be a (plus (reg x) (const_int 1)) rtx.
826 Let the following cse pass simplify this if initial value is
829 We must copy the final and initial values here to avoid
830 improperly shared rtl. */
832 diff
= expand_binop (mode
, sub_optab
, copy_rtx (final_value
),
833 copy_rtx (initial_value
), NULL_RTX
, 0,
836 /* Now calculate (diff % (unroll * abs (increment))) by using an
838 diff
= expand_binop (GET_MODE (diff
), and_optab
, diff
,
839 GEN_INT (unroll_number
* abs_inc
- 1),
840 NULL_RTX
, 0, OPTAB_LIB_WIDEN
);
842 /* Now emit a sequence of branches to jump to the proper precond
845 labels
= (rtx
*) alloca (sizeof (rtx
) * unroll_number
);
846 for (i
= 0; i
< unroll_number
; i
++)
847 labels
[i
] = gen_label_rtx ();
849 /* Assuming the unroll_number is 4, and the increment is 2, then
850 for a negative increment: for a positive increment:
851 diff = 0,1 precond 0 diff = 0,7 precond 0
852 diff = 2,3 precond 3 diff = 1,2 precond 1
853 diff = 4,5 precond 2 diff = 3,4 precond 2
854 diff = 6,7 precond 1 diff = 5,6 precond 3 */
856 /* We only need to emit (unroll_number - 1) branches here, the
857 last case just falls through to the following code. */
859 /* ??? This would give better code if we emitted a tree of branches
860 instead of the current linear list of branches. */
862 for (i
= 0; i
< unroll_number
- 1; i
++)
866 /* For negative increments, must invert the constant compared
867 against, except when comparing against zero. */
871 cmp_const
= unroll_number
- i
;
875 emit_cmp_insn (diff
, GEN_INT (abs_inc
* cmp_const
),
876 EQ
, NULL_RTX
, mode
, 0, 0);
879 emit_jump_insn (gen_beq (labels
[i
]));
881 emit_jump_insn (gen_bge (labels
[i
]));
883 emit_jump_insn (gen_ble (labels
[i
]));
884 JUMP_LABEL (get_last_insn ()) = labels
[i
];
885 LABEL_NUSES (labels
[i
])++;
888 /* If the increment is greater than one, then we need another branch,
889 to handle other cases equivalent to 0. */
891 /* ??? This should be merged into the code above somehow to help
892 simplify the code here, and reduce the number of branches emitted.
893 For the negative increment case, the branch here could easily
894 be merged with the `0' case branch above. For the positive
895 increment case, it is not clear how this can be simplified. */
902 cmp_const
= abs_inc
- 1;
904 cmp_const
= abs_inc
* (unroll_number
- 1) + 1;
906 emit_cmp_insn (diff
, GEN_INT (cmp_const
), EQ
, NULL_RTX
,
910 emit_jump_insn (gen_ble (labels
[0]));
912 emit_jump_insn (gen_bge (labels
[0]));
913 JUMP_LABEL (get_last_insn ()) = labels
[0];
914 LABEL_NUSES (labels
[0])++;
917 sequence
= gen_sequence ();
919 emit_insn_before (sequence
, loop_start
);
921 /* Only the last copy of the loop body here needs the exit
922 test, so set copy_end to exclude the compare/branch here,
923 and then reset it inside the loop when get to the last
926 if (GET_CODE (last_loop_insn
) == BARRIER
)
927 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
928 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
931 /* The immediately preceding insn is a compare which we do not
933 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
935 /* The immediately preceding insn may not be a compare, so we
937 copy_end
= PREV_INSN (last_loop_insn
);
943 for (i
= 1; i
< unroll_number
; i
++)
945 emit_label_after (labels
[unroll_number
- i
],
946 PREV_INSN (loop_start
));
948 bzero ((char *) map
->insn_map
, max_insnno
* sizeof (rtx
));
949 bzero ((char *) map
->const_equiv_map
, maxregnum
* sizeof (rtx
));
950 bzero ((char *) map
->const_age_map
,
951 maxregnum
* sizeof (unsigned));
954 for (j
= 0; j
< max_labelno
; j
++)
956 map
->label_map
[j
] = gen_label_rtx ();
958 for (j
= FIRST_PSEUDO_REGISTER
; j
< max_reg_before_loop
; j
++)
960 map
->reg_map
[j
] = gen_reg_rtx (GET_MODE (regno_reg_rtx
[j
]));
962 /* The last copy needs the compare/branch insns at the end,
963 so reset copy_end here if the loop ends with a conditional
966 if (i
== unroll_number
- 1)
968 if (GET_CODE (last_loop_insn
) == BARRIER
)
969 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
971 copy_end
= last_loop_insn
;
974 /* None of the copies are the `last_iteration', so just
975 pass zero for that parameter. */
976 copy_loop_body (copy_start
, copy_end
, map
, exit_label
, 0,
977 unroll_type
, start_label
, loop_end
,
978 loop_start
, copy_end
);
980 emit_label_after (labels
[0], PREV_INSN (loop_start
));
982 if (GET_CODE (last_loop_insn
) == BARRIER
)
984 insert_before
= PREV_INSN (last_loop_insn
);
985 copy_end
= PREV_INSN (insert_before
);
990 /* The immediately preceding insn is a compare which we do not
992 insert_before
= PREV_INSN (last_loop_insn
);
993 copy_end
= PREV_INSN (insert_before
);
995 /* The immediately preceding insn may not be a compare, so we
997 insert_before
= last_loop_insn
;
998 copy_end
= PREV_INSN (last_loop_insn
);
1002 /* Set unroll type to MODULO now. */
1003 unroll_type
= UNROLL_MODULO
;
1004 loop_preconditioned
= 1;
1008 /* If reach here, and the loop type is UNROLL_NAIVE, then don't unroll
1009 the loop unless all loops are being unrolled. */
1010 if (unroll_type
== UNROLL_NAIVE
&& ! flag_unroll_all_loops
)
1012 if (loop_dump_stream
)
1013 fprintf (loop_dump_stream
, "Unrolling failure: Naive unrolling not being done.\n");
1017 /* At this point, we are guaranteed to unroll the loop. */
1019 /* For each biv and giv, determine whether it can be safely split into
1020 a different variable for each unrolled copy of the loop body.
1021 We precalculate and save this info here, since computing it is
1024 Do this before deleting any instructions from the loop, so that
1025 back_branch_in_range_p will work correctly. */
1027 if (splitting_not_safe
)
1030 temp
= find_splittable_regs (unroll_type
, loop_start
, loop_end
,
1031 end_insert_before
, unroll_number
);
1033 /* find_splittable_regs may have created some new registers, so must
1034 reallocate the reg_map with the new larger size, and must realloc
1035 the constant maps also. */
1037 maxregnum
= max_reg_num ();
1038 map
->reg_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
1040 init_reg_map (map
, maxregnum
);
1042 /* Space is needed in some of the map for new registers, so new_maxregnum
1043 is an (over)estimate of how many registers will exist at the end. */
1044 new_maxregnum
= maxregnum
+ (temp
* unroll_number
* 2);
1046 /* Must realloc space for the constant maps, because the number of registers
1047 may have changed. */
1049 map
->const_equiv_map
= (rtx
*) alloca (new_maxregnum
* sizeof (rtx
));
1050 map
->const_age_map
= (unsigned *) alloca (new_maxregnum
* sizeof (unsigned));
1052 map
->const_equiv_map_size
= new_maxregnum
;
1053 global_const_equiv_map
= map
->const_equiv_map
;
1054 global_const_equiv_map_size
= new_maxregnum
;
1056 /* Search the list of bivs and givs to find ones which need to be remapped
1057 when split, and set their reg_map entry appropriately. */
1059 for (bl
= loop_iv_list
; bl
; bl
= bl
->next
)
1061 if (REGNO (bl
->biv
->src_reg
) != bl
->regno
)
1062 map
->reg_map
[bl
->regno
] = bl
->biv
->src_reg
;
1064 /* Currently, non-reduced/final-value givs are never split. */
1065 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
1066 if (REGNO (v
->src_reg
) != bl
->regno
)
1067 map
->reg_map
[REGNO (v
->dest_reg
)] = v
->src_reg
;
1071 /* If the loop is being partially unrolled, and the iteration variables
1072 are being split, and are being renamed for the split, then must fix up
1073 the compare/jump instruction at the end of the loop to refer to the new
1074 registers. This compare isn't copied, so the registers used in it
1075 will never be replaced if it isn't done here. */
1077 if (unroll_type
== UNROLL_MODULO
)
1079 insn
= NEXT_INSN (copy_end
);
1080 if (GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == JUMP_INSN
)
1081 PATTERN (insn
) = remap_split_bivs (PATTERN (insn
));
1084 /* For unroll_number - 1 times, make a copy of each instruction
1085 between copy_start and copy_end, and insert these new instructions
1086 before the end of the loop. */
1088 for (i
= 0; i
< unroll_number
; i
++)
1090 bzero ((char *) map
->insn_map
, max_insnno
* sizeof (rtx
));
1091 bzero ((char *) map
->const_equiv_map
, new_maxregnum
* sizeof (rtx
));
1092 bzero ((char *) map
->const_age_map
, new_maxregnum
* sizeof (unsigned));
1095 for (j
= 0; j
< max_labelno
; j
++)
1097 map
->label_map
[j
] = gen_label_rtx ();
1099 for (j
= FIRST_PSEUDO_REGISTER
; j
< max_reg_before_loop
; j
++)
1101 map
->reg_map
[j
] = gen_reg_rtx (GET_MODE (regno_reg_rtx
[j
]));
1103 /* If loop starts with a branch to the test, then fix it so that
1104 it points to the test of the first unrolled copy of the loop. */
1105 if (i
== 0 && loop_start
!= copy_start
)
1107 insn
= PREV_INSN (copy_start
);
1108 pattern
= PATTERN (insn
);
1110 tem
= map
->label_map
[CODE_LABEL_NUMBER
1111 (XEXP (SET_SRC (pattern
), 0))];
1112 SET_SRC (pattern
) = gen_rtx (LABEL_REF
, VOIDmode
, tem
);
1114 /* Set the jump label so that it can be used by later loop unrolling
1116 JUMP_LABEL (insn
) = tem
;
1117 LABEL_NUSES (tem
)++;
1120 copy_loop_body (copy_start
, copy_end
, map
, exit_label
,
1121 i
== unroll_number
- 1, unroll_type
, start_label
,
1122 loop_end
, insert_before
, insert_before
);
1125 /* Before deleting any insns, emit a CODE_LABEL immediately after the last
1126 insn to be deleted. This prevents any runaway delete_insn call from
1127 more insns that it should, as it always stops at a CODE_LABEL. */
1129 /* Delete the compare and branch at the end of the loop if completely
1130 unrolling the loop. Deleting the backward branch at the end also
1131 deletes the code label at the start of the loop. This is done at
1132 the very end to avoid problems with back_branch_in_range_p. */
1134 if (unroll_type
== UNROLL_COMPLETELY
)
1135 safety_label
= emit_label_after (gen_label_rtx (), last_loop_insn
);
1137 safety_label
= emit_label_after (gen_label_rtx (), copy_end
);
1139 /* Delete all of the original loop instructions. Don't delete the
1140 LOOP_BEG note, or the first code label in the loop. */
1142 insn
= NEXT_INSN (copy_start
);
1143 while (insn
!= safety_label
)
1145 if (insn
!= start_label
)
1146 insn
= delete_insn (insn
);
1148 insn
= NEXT_INSN (insn
);
1151 /* Can now delete the 'safety' label emitted to protect us from runaway
1152 delete_insn calls. */
1153 if (INSN_DELETED_P (safety_label
))
1155 delete_insn (safety_label
);
1157 /* If exit_label exists, emit it after the loop. Doing the emit here
1158 forces it to have a higher INSN_UID than any insn in the unrolled loop.
1159 This is needed so that mostly_true_jump in reorg.c will treat jumps
1160 to this loop end label correctly, i.e. predict that they are usually
1163 emit_label_after (exit_label
, loop_end
);
1166 /* Return true if the loop can be safely, and profitably, preconditioned
1167 so that the unrolled copies of the loop body don't need exit tests.
1169 This only works if final_value, initial_value and increment can be
1170 determined, and if increment is a constant power of 2.
1171 If increment is not a power of 2, then the preconditioning modulo
1172 operation would require a real modulo instead of a boolean AND, and this
1173 is not considered `profitable'. */
1175 /* ??? If the loop is known to be executed very many times, or the machine
1176 has a very cheap divide instruction, then preconditioning is a win even
1177 when the increment is not a power of 2. Use RTX_COST to compute
1178 whether divide is cheap. */
1181 precondition_loop_p (initial_value
, final_value
, increment
, loop_start
,
1183 rtx
*initial_value
, *final_value
, *increment
;
1184 rtx loop_start
, loop_end
;
1187 if (loop_n_iterations
> 0)
1189 *initial_value
= const0_rtx
;
1190 *increment
= const1_rtx
;
1191 *final_value
= GEN_INT (loop_n_iterations
);
1193 if (loop_dump_stream
)
1194 fprintf (loop_dump_stream
,
1195 "Preconditioning: Success, number of iterations known, %d.\n",
1200 if (loop_initial_value
== 0)
1202 if (loop_dump_stream
)
1203 fprintf (loop_dump_stream
,
1204 "Preconditioning: Could not find initial value.\n");
1207 else if (loop_increment
== 0)
1209 if (loop_dump_stream
)
1210 fprintf (loop_dump_stream
,
1211 "Preconditioning: Could not find increment value.\n");
1214 else if (GET_CODE (loop_increment
) != CONST_INT
)
1216 if (loop_dump_stream
)
1217 fprintf (loop_dump_stream
,
1218 "Preconditioning: Increment not a constant.\n");
1221 else if ((exact_log2 (INTVAL (loop_increment
)) < 0)
1222 && (exact_log2 (- INTVAL (loop_increment
)) < 0))
1224 if (loop_dump_stream
)
1225 fprintf (loop_dump_stream
,
1226 "Preconditioning: Increment not a constant power of 2.\n");
1230 /* Unsigned_compare and compare_dir can be ignored here, since they do
1231 not matter for preconditioning. */
1233 if (loop_final_value
== 0)
1235 if (loop_dump_stream
)
1236 fprintf (loop_dump_stream
,
1237 "Preconditioning: EQ comparison loop.\n");
1241 /* Must ensure that final_value is invariant, so call invariant_p to
1242 check. Before doing so, must check regno against max_reg_before_loop
1243 to make sure that the register is in the range covered by invariant_p.
1244 If it isn't, then it is most likely a biv/giv which by definition are
1246 if ((GET_CODE (loop_final_value
) == REG
1247 && REGNO (loop_final_value
) >= max_reg_before_loop
)
1248 || (GET_CODE (loop_final_value
) == PLUS
1249 && REGNO (XEXP (loop_final_value
, 0)) >= max_reg_before_loop
)
1250 || ! invariant_p (loop_final_value
))
1252 if (loop_dump_stream
)
1253 fprintf (loop_dump_stream
,
1254 "Preconditioning: Final value not invariant.\n");
1258 /* Fail for floating point values, since the caller of this function
1259 does not have code to deal with them. */
1260 if (GET_MODE_CLASS (GET_MODE (loop_final_value
)) == MODE_FLOAT
1261 || GET_MODE_CLASS (GET_MODE (loop_initial_value
)) == MODE_FLOAT
)
1263 if (loop_dump_stream
)
1264 fprintf (loop_dump_stream
,
1265 "Preconditioning: Floating point final or initial value.\n");
1269 /* Now set initial_value to be the iteration_var, since that may be a
1270 simpler expression, and is guaranteed to be correct if all of the
1271 above tests succeed.
1273 We can not use the initial_value as calculated, because it will be
1274 one too small for loops of the form "while (i-- > 0)". We can not
1275 emit code before the loop_skip_over insns to fix this problem as this
1276 will then give a number one too large for loops of the form
1279 Note that all loops that reach here are entered at the top, because
1280 this function is not called if the loop starts with a jump. */
1282 /* Fail if loop_iteration_var is not live before loop_start, since we need
1283 to test its value in the preconditioning code. */
1285 if (uid_luid
[regno_first_uid
[REGNO (loop_iteration_var
)]]
1286 > INSN_LUID (loop_start
))
1288 if (loop_dump_stream
)
1289 fprintf (loop_dump_stream
,
1290 "Preconditioning: Iteration var not live before loop start.\n");
1294 *initial_value
= loop_iteration_var
;
1295 *increment
= loop_increment
;
1296 *final_value
= loop_final_value
;
1299 if (loop_dump_stream
)
1300 fprintf (loop_dump_stream
, "Preconditioning: Successful.\n");
1305 /* All pseudo-registers must be mapped to themselves. Two hard registers
1306 must be mapped, VIRTUAL_STACK_VARS_REGNUM and VIRTUAL_INCOMING_ARGS_
1307 REGNUM, to avoid function-inlining specific conversions of these
1308 registers. All other hard regs can not be mapped because they may be
1313 init_reg_map (map
, maxregnum
)
1314 struct inline_remap
*map
;
1319 for (i
= maxregnum
- 1; i
> LAST_VIRTUAL_REGISTER
; i
--)
1320 map
->reg_map
[i
] = regno_reg_rtx
[i
];
1321 /* Just clear the rest of the entries. */
1322 for (i
= LAST_VIRTUAL_REGISTER
; i
>= 0; i
--)
1323 map
->reg_map
[i
] = 0;
1325 map
->reg_map
[VIRTUAL_STACK_VARS_REGNUM
]
1326 = regno_reg_rtx
[VIRTUAL_STACK_VARS_REGNUM
];
1327 map
->reg_map
[VIRTUAL_INCOMING_ARGS_REGNUM
]
1328 = regno_reg_rtx
[VIRTUAL_INCOMING_ARGS_REGNUM
];
1331 /* Strength-reduction will often emit code for optimized biv/givs which
1332 calculates their value in a temporary register, and then copies the result
1333 to the iv. This procedure reconstructs the pattern computing the iv;
1334 verifying that all operands are of the proper form.
1336 The return value is the amount that the giv is incremented by. */
1339 calculate_giv_inc (pattern
, src_insn
, regno
)
1340 rtx pattern
, src_insn
;
1344 rtx increment_total
= 0;
1348 /* Verify that we have an increment insn here. First check for a plus
1349 as the set source. */
1350 if (GET_CODE (SET_SRC (pattern
)) != PLUS
)
1352 /* SR sometimes computes the new giv value in a temp, then copies it
1354 src_insn
= PREV_INSN (src_insn
);
1355 pattern
= PATTERN (src_insn
);
1356 if (GET_CODE (SET_SRC (pattern
)) != PLUS
)
1359 /* The last insn emitted is not needed, so delete it to avoid confusing
1360 the second cse pass. This insn sets the giv unnecessarily. */
1361 delete_insn (get_last_insn ());
1364 /* Verify that we have a constant as the second operand of the plus. */
1365 increment
= XEXP (SET_SRC (pattern
), 1);
1366 if (GET_CODE (increment
) != CONST_INT
)
1368 /* SR sometimes puts the constant in a register, especially if it is
1369 too big to be an add immed operand. */
1370 src_insn
= PREV_INSN (src_insn
);
1371 increment
= SET_SRC (PATTERN (src_insn
));
1373 /* SR may have used LO_SUM to compute the constant if it is too large
1374 for a load immed operand. In this case, the constant is in operand
1375 one of the LO_SUM rtx. */
1376 if (GET_CODE (increment
) == LO_SUM
)
1377 increment
= XEXP (increment
, 1);
1378 else if (GET_CODE (increment
) == IOR
)
1380 /* The rs6000 port loads some constants with IOR. */
1381 rtx second_part
= XEXP (increment
, 1);
1383 src_insn
= PREV_INSN (src_insn
);
1384 increment
= SET_SRC (PATTERN (src_insn
));
1385 /* Don't need the last insn anymore. */
1386 delete_insn (get_last_insn ());
1388 if (GET_CODE (second_part
) != CONST_INT
1389 || GET_CODE (increment
) != CONST_INT
)
1392 increment
= GEN_INT (INTVAL (increment
) | INTVAL (second_part
));
1395 if (GET_CODE (increment
) != CONST_INT
)
1398 /* The insn loading the constant into a register is no longer needed,
1400 delete_insn (get_last_insn ());
1403 if (increment_total
)
1404 increment_total
= GEN_INT (INTVAL (increment_total
) + INTVAL (increment
));
1406 increment_total
= increment
;
1408 /* Check that the source register is the same as the register we expected
1409 to see as the source. If not, something is seriously wrong. */
1410 if (GET_CODE (XEXP (SET_SRC (pattern
), 0)) != REG
1411 || REGNO (XEXP (SET_SRC (pattern
), 0)) != regno
)
1413 /* Some machines (e.g. the romp), may emit two add instructions for
1414 certain constants, so lets try looking for another add immediately
1415 before this one if we have only seen one add insn so far. */
1421 src_insn
= PREV_INSN (src_insn
);
1422 pattern
= PATTERN (src_insn
);
1424 delete_insn (get_last_insn ());
1432 return increment_total
;
1435 /* Copy REG_NOTES, except for insn references, because not all insn_map
1436 entries are valid yet. We do need to copy registers now though, because
1437 the reg_map entries can change during copying. */
1440 initial_reg_note_copy (notes
, map
)
1442 struct inline_remap
*map
;
1449 copy
= rtx_alloc (GET_CODE (notes
));
1450 PUT_MODE (copy
, GET_MODE (notes
));
1452 if (GET_CODE (notes
) == EXPR_LIST
)
1453 XEXP (copy
, 0) = copy_rtx_and_substitute (XEXP (notes
, 0), map
);
1454 else if (GET_CODE (notes
) == INSN_LIST
)
1455 /* Don't substitute for these yet. */
1456 XEXP (copy
, 0) = XEXP (notes
, 0);
1460 XEXP (copy
, 1) = initial_reg_note_copy (XEXP (notes
, 1), map
);
1465 /* Fixup insn references in copied REG_NOTES. */
1468 final_reg_note_copy (notes
, map
)
1470 struct inline_remap
*map
;
1474 for (note
= notes
; note
; note
= XEXP (note
, 1))
1475 if (GET_CODE (note
) == INSN_LIST
)
1476 XEXP (note
, 0) = map
->insn_map
[INSN_UID (XEXP (note
, 0))];
1479 /* Copy each instruction in the loop, substituting from map as appropriate.
1480 This is very similar to a loop in expand_inline_function. */
1483 copy_loop_body (copy_start
, copy_end
, map
, exit_label
, last_iteration
,
1484 unroll_type
, start_label
, loop_end
, insert_before
,
1486 rtx copy_start
, copy_end
;
1487 struct inline_remap
*map
;
1490 enum unroll_types unroll_type
;
1491 rtx start_label
, loop_end
, insert_before
, copy_notes_from
;
1495 int dest_reg_was_split
, i
;
1497 rtx final_label
= 0;
1498 rtx giv_inc
, giv_dest_reg
, giv_src_reg
;
1500 /* If this isn't the last iteration, then map any references to the
1501 start_label to final_label. Final label will then be emitted immediately
1502 after the end of this loop body if it was ever used.
1504 If this is the last iteration, then map references to the start_label
1506 if (! last_iteration
)
1508 final_label
= gen_label_rtx ();
1509 map
->label_map
[CODE_LABEL_NUMBER (start_label
)] = final_label
;
1512 map
->label_map
[CODE_LABEL_NUMBER (start_label
)] = start_label
;
1519 insn
= NEXT_INSN (insn
);
1521 map
->orig_asm_operands_vector
= 0;
1523 switch (GET_CODE (insn
))
1526 pattern
= PATTERN (insn
);
1530 /* Check to see if this is a giv that has been combined with
1531 some split address givs. (Combined in the sense that
1532 `combine_givs' in loop.c has put two givs in the same register.)
1533 In this case, we must search all givs based on the same biv to
1534 find the address givs. Then split the address givs.
1535 Do this before splitting the giv, since that may map the
1536 SET_DEST to a new register. */
1538 if (GET_CODE (pattern
) == SET
1539 && GET_CODE (SET_DEST (pattern
)) == REG
1540 && addr_combined_regs
[REGNO (SET_DEST (pattern
))])
1542 struct iv_class
*bl
;
1543 struct induction
*v
, *tv
;
1544 int regno
= REGNO (SET_DEST (pattern
));
1546 v
= addr_combined_regs
[REGNO (SET_DEST (pattern
))];
1547 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
1549 /* Although the giv_inc amount is not needed here, we must call
1550 calculate_giv_inc here since it might try to delete the
1551 last insn emitted. If we wait until later to call it,
1552 we might accidentally delete insns generated immediately
1553 below by emit_unrolled_add. */
1555 giv_inc
= calculate_giv_inc (pattern
, insn
, regno
);
1557 /* Now find all address giv's that were combined with this
1559 for (tv
= bl
->giv
; tv
; tv
= tv
->next_iv
)
1560 if (tv
->giv_type
== DEST_ADDR
&& tv
->same
== v
)
1562 int this_giv_inc
= INTVAL (giv_inc
);
1564 /* Scale this_giv_inc if the multiplicative factors of
1565 the two givs are different. */
1566 if (tv
->mult_val
!= v
->mult_val
)
1567 this_giv_inc
= (this_giv_inc
/ INTVAL (v
->mult_val
)
1568 * INTVAL (tv
->mult_val
));
1570 tv
->dest_reg
= plus_constant (tv
->dest_reg
, this_giv_inc
);
1571 *tv
->location
= tv
->dest_reg
;
1573 if (last_iteration
&& unroll_type
!= UNROLL_COMPLETELY
)
1575 /* Must emit an insn to increment the split address
1576 giv. Add in the const_adjust field in case there
1577 was a constant eliminated from the address. */
1578 rtx value
, dest_reg
;
1580 /* tv->dest_reg will be either a bare register,
1581 or else a register plus a constant. */
1582 if (GET_CODE (tv
->dest_reg
) == REG
)
1583 dest_reg
= tv
->dest_reg
;
1585 dest_reg
= XEXP (tv
->dest_reg
, 0);
1587 /* Check for shared address givs, and avoid
1588 incrementing the shared psuedo reg more than
1590 if (! tv
->same_insn
)
1592 /* tv->dest_reg may actually be a (PLUS (REG)
1593 (CONST)) here, so we must call plus_constant
1594 to add the const_adjust amount before calling
1595 emit_unrolled_add below. */
1596 value
= plus_constant (tv
->dest_reg
,
1599 /* The constant could be too large for an add
1600 immediate, so can't directly emit an insn
1602 emit_unrolled_add (dest_reg
, XEXP (value
, 0),
1606 /* Reset the giv to be just the register again, in case
1607 it is used after the set we have just emitted.
1608 We must subtract the const_adjust factor added in
1610 tv
->dest_reg
= plus_constant (dest_reg
,
1611 - tv
->const_adjust
);
1612 *tv
->location
= tv
->dest_reg
;
1617 /* If this is a setting of a splittable variable, then determine
1618 how to split the variable, create a new set based on this split,
1619 and set up the reg_map so that later uses of the variable will
1620 use the new split variable. */
1622 dest_reg_was_split
= 0;
1624 if (GET_CODE (pattern
) == SET
1625 && GET_CODE (SET_DEST (pattern
)) == REG
1626 && splittable_regs
[REGNO (SET_DEST (pattern
))])
1628 int regno
= REGNO (SET_DEST (pattern
));
1630 dest_reg_was_split
= 1;
1632 /* Compute the increment value for the giv, if it wasn't
1633 already computed above. */
1636 giv_inc
= calculate_giv_inc (pattern
, insn
, regno
);
1637 giv_dest_reg
= SET_DEST (pattern
);
1638 giv_src_reg
= SET_DEST (pattern
);
1640 if (unroll_type
== UNROLL_COMPLETELY
)
1642 /* Completely unrolling the loop. Set the induction
1643 variable to a known constant value. */
1645 /* The value in splittable_regs may be an invariant
1646 value, so we must use plus_constant here. */
1647 splittable_regs
[regno
]
1648 = plus_constant (splittable_regs
[regno
], INTVAL (giv_inc
));
1650 if (GET_CODE (splittable_regs
[regno
]) == PLUS
)
1652 giv_src_reg
= XEXP (splittable_regs
[regno
], 0);
1653 giv_inc
= XEXP (splittable_regs
[regno
], 1);
1657 /* The splittable_regs value must be a REG or a
1658 CONST_INT, so put the entire value in the giv_src_reg
1660 giv_src_reg
= splittable_regs
[regno
];
1661 giv_inc
= const0_rtx
;
1666 /* Partially unrolling loop. Create a new pseudo
1667 register for the iteration variable, and set it to
1668 be a constant plus the original register. Except
1669 on the last iteration, when the result has to
1670 go back into the original iteration var register. */
1672 /* Handle bivs which must be mapped to a new register
1673 when split. This happens for bivs which need their
1674 final value set before loop entry. The new register
1675 for the biv was stored in the biv's first struct
1676 induction entry by find_splittable_regs. */
1678 if (regno
< max_reg_before_loop
1679 && reg_iv_type
[regno
] == BASIC_INDUCT
)
1681 giv_src_reg
= reg_biv_class
[regno
]->biv
->src_reg
;
1682 giv_dest_reg
= giv_src_reg
;
1686 /* If non-reduced/final-value givs were split, then
1687 this would have to remap those givs also. See
1688 find_splittable_regs. */
1691 splittable_regs
[regno
]
1692 = GEN_INT (INTVAL (giv_inc
)
1693 + INTVAL (splittable_regs
[regno
]));
1694 giv_inc
= splittable_regs
[regno
];
1696 /* Now split the induction variable by changing the dest
1697 of this insn to a new register, and setting its
1698 reg_map entry to point to this new register.
1700 If this is the last iteration, and this is the last insn
1701 that will update the iv, then reuse the original dest,
1702 to ensure that the iv will have the proper value when
1703 the loop exits or repeats.
1705 Using splittable_regs_updates here like this is safe,
1706 because it can only be greater than one if all
1707 instructions modifying the iv are always executed in
1710 if (! last_iteration
1711 || (splittable_regs_updates
[regno
]-- != 1))
1713 tem
= gen_reg_rtx (GET_MODE (giv_src_reg
));
1715 map
->reg_map
[regno
] = tem
;
1718 map
->reg_map
[regno
] = giv_src_reg
;
1721 /* The constant being added could be too large for an add
1722 immediate, so can't directly emit an insn here. */
1723 emit_unrolled_add (giv_dest_reg
, giv_src_reg
, giv_inc
);
1724 copy
= get_last_insn ();
1725 pattern
= PATTERN (copy
);
1729 pattern
= copy_rtx_and_substitute (pattern
, map
);
1730 copy
= emit_insn (pattern
);
1732 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
1735 /* If this insn is setting CC0, it may need to look at
1736 the insn that uses CC0 to see what type of insn it is.
1737 In that case, the call to recog via validate_change will
1738 fail. So don't substitute constants here. Instead,
1739 do it when we emit the following insn.
1741 For example, see the pyr.md file. That machine has signed and
1742 unsigned compares. The compare patterns must check the
1743 following branch insn to see which what kind of compare to
1746 If the previous insn set CC0, substitute constants on it as
1748 if (sets_cc0_p (copy
) != 0)
1753 try_constants (cc0_insn
, map
);
1755 try_constants (copy
, map
);
1758 try_constants (copy
, map
);
1761 /* Make split induction variable constants `permanent' since we
1762 know there are no backward branches across iteration variable
1763 settings which would invalidate this. */
1764 if (dest_reg_was_split
)
1766 int regno
= REGNO (SET_DEST (pattern
));
1768 if (regno
< map
->const_equiv_map_size
1769 && map
->const_age_map
[regno
] == map
->const_age
)
1770 map
->const_age_map
[regno
] = -1;
1775 pattern
= copy_rtx_and_substitute (PATTERN (insn
), map
);
1776 copy
= emit_jump_insn (pattern
);
1777 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
1779 if (JUMP_LABEL (insn
) == start_label
&& insn
== copy_end
1780 && ! last_iteration
)
1782 /* This is a branch to the beginning of the loop; this is the
1783 last insn being copied; and this is not the last iteration.
1784 In this case, we want to change the original fall through
1785 case to be a branch past the end of the loop, and the
1786 original jump label case to fall_through. */
1788 if (invert_exp (pattern
, copy
))
1790 if (! redirect_exp (&pattern
,
1791 map
->label_map
[CODE_LABEL_NUMBER
1792 (JUMP_LABEL (insn
))],
1799 rtx lab
= gen_label_rtx ();
1800 /* Can't do it by reversing the jump (probably becasue we
1801 couln't reverse the conditions), so emit a new
1802 jump_insn after COPY, and redirect the jump around
1804 jmp
= emit_jump_insn_after (gen_jump (exit_label
), copy
);
1805 jmp
= emit_barrier_after (jmp
);
1806 emit_label_after (lab
, jmp
);
1807 LABEL_NUSES (lab
) = 0;
1808 if (! redirect_exp (&pattern
,
1809 map
->label_map
[CODE_LABEL_NUMBER
1810 (JUMP_LABEL (insn
))],
1818 try_constants (cc0_insn
, map
);
1821 try_constants (copy
, map
);
1823 /* Set the jump label of COPY correctly to avoid problems with
1824 later passes of unroll_loop, if INSN had jump label set. */
1825 if (JUMP_LABEL (insn
))
1829 /* Can't use the label_map for every insn, since this may be
1830 the backward branch, and hence the label was not mapped. */
1831 if (GET_CODE (pattern
) == SET
)
1833 tem
= SET_SRC (pattern
);
1834 if (GET_CODE (tem
) == LABEL_REF
)
1835 label
= XEXP (tem
, 0);
1836 else if (GET_CODE (tem
) == IF_THEN_ELSE
)
1838 if (XEXP (tem
, 1) != pc_rtx
)
1839 label
= XEXP (XEXP (tem
, 1), 0);
1841 label
= XEXP (XEXP (tem
, 2), 0);
1845 if (label
&& GET_CODE (label
) == CODE_LABEL
)
1846 JUMP_LABEL (copy
) = label
;
1849 /* An unrecognizable jump insn, probably the entry jump
1850 for a switch statement. This label must have been mapped,
1851 so just use the label_map to get the new jump label. */
1853 = map
->label_map
[CODE_LABEL_NUMBER (JUMP_LABEL (insn
))];
1856 /* If this is a non-local jump, then must increase the label
1857 use count so that the label will not be deleted when the
1858 original jump is deleted. */
1859 LABEL_NUSES (JUMP_LABEL (copy
))++;
1861 else if (GET_CODE (PATTERN (copy
)) == ADDR_VEC
1862 || GET_CODE (PATTERN (copy
)) == ADDR_DIFF_VEC
)
1864 rtx pat
= PATTERN (copy
);
1865 int diff_vec_p
= GET_CODE (pat
) == ADDR_DIFF_VEC
;
1866 int len
= XVECLEN (pat
, diff_vec_p
);
1869 for (i
= 0; i
< len
; i
++)
1870 LABEL_NUSES (XEXP (XVECEXP (pat
, diff_vec_p
, i
), 0))++;
1873 /* If this used to be a conditional jump insn but whose branch
1874 direction is now known, we must do something special. */
1875 if (condjump_p (insn
) && !simplejump_p (insn
) && map
->last_pc_value
)
1878 /* The previous insn set cc0 for us. So delete it. */
1879 delete_insn (PREV_INSN (copy
));
1882 /* If this is now a no-op, delete it. */
1883 if (map
->last_pc_value
== pc_rtx
)
1885 /* Don't let delete_insn delete the label referenced here,
1886 because we might possibly need it later for some other
1887 instruction in the loop. */
1888 if (JUMP_LABEL (copy
))
1889 LABEL_NUSES (JUMP_LABEL (copy
))++;
1891 if (JUMP_LABEL (copy
))
1892 LABEL_NUSES (JUMP_LABEL (copy
))--;
1896 /* Otherwise, this is unconditional jump so we must put a
1897 BARRIER after it. We could do some dead code elimination
1898 here, but jump.c will do it just as well. */
1904 pattern
= copy_rtx_and_substitute (PATTERN (insn
), map
);
1905 copy
= emit_call_insn (pattern
);
1906 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
1908 /* Because the USAGE information potentially contains objects other
1909 than hard registers, we need to copy it. */
1910 CALL_INSN_FUNCTION_USAGE (copy
) =
1911 copy_rtx_and_substitute (CALL_INSN_FUNCTION_USAGE (insn
), map
);
1915 try_constants (cc0_insn
, map
);
1918 try_constants (copy
, map
);
1920 /* Be lazy and assume CALL_INSNs clobber all hard registers. */
1921 for (i
= 0; i
< FIRST_PSEUDO_REGISTER
; i
++)
1922 map
->const_equiv_map
[i
] = 0;
1926 /* If this is the loop start label, then we don't need to emit a
1927 copy of this label since no one will use it. */
1929 if (insn
!= start_label
)
1931 copy
= emit_label (map
->label_map
[CODE_LABEL_NUMBER (insn
)]);
1937 copy
= emit_barrier ();
1941 /* VTOP notes are valid only before the loop exit test. If placed
1942 anywhere else, loop may generate bad code. */
1944 if (NOTE_LINE_NUMBER (insn
) != NOTE_INSN_DELETED
1945 && (NOTE_LINE_NUMBER (insn
) != NOTE_INSN_LOOP_VTOP
1946 || (last_iteration
&& unroll_type
!= UNROLL_COMPLETELY
)))
1947 copy
= emit_note (NOTE_SOURCE_FILE (insn
),
1948 NOTE_LINE_NUMBER (insn
));
1958 map
->insn_map
[INSN_UID (insn
)] = copy
;
1960 while (insn
!= copy_end
);
1962 /* Now finish coping the REG_NOTES. */
1966 insn
= NEXT_INSN (insn
);
1967 if ((GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == JUMP_INSN
1968 || GET_CODE (insn
) == CALL_INSN
)
1969 && map
->insn_map
[INSN_UID (insn
)])
1970 final_reg_note_copy (REG_NOTES (map
->insn_map
[INSN_UID (insn
)]), map
);
1972 while (insn
!= copy_end
);
1974 /* There may be notes between copy_notes_from and loop_end. Emit a copy of
1975 each of these notes here, since there may be some important ones, such as
1976 NOTE_INSN_BLOCK_END notes, in this group. We don't do this on the last
1977 iteration, because the original notes won't be deleted.
1979 We can't use insert_before here, because when from preconditioning,
1980 insert_before points before the loop. We can't use copy_end, because
1981 there may be insns already inserted after it (which we don't want to
1982 copy) when not from preconditioning code. */
1984 if (! last_iteration
)
1986 for (insn
= copy_notes_from
; insn
!= loop_end
; insn
= NEXT_INSN (insn
))
1988 if (GET_CODE (insn
) == NOTE
1989 && NOTE_LINE_NUMBER (insn
) != NOTE_INSN_DELETED
)
1990 emit_note (NOTE_SOURCE_FILE (insn
), NOTE_LINE_NUMBER (insn
));
1994 if (final_label
&& LABEL_NUSES (final_label
) > 0)
1995 emit_label (final_label
);
1997 tem
= gen_sequence ();
1999 emit_insn_before (tem
, insert_before
);
2002 /* Emit an insn, using the expand_binop to ensure that a valid insn is
2003 emitted. This will correctly handle the case where the increment value
2004 won't fit in the immediate field of a PLUS insns. */
2007 emit_unrolled_add (dest_reg
, src_reg
, increment
)
2008 rtx dest_reg
, src_reg
, increment
;
2012 result
= expand_binop (GET_MODE (dest_reg
), add_optab
, src_reg
, increment
,
2013 dest_reg
, 0, OPTAB_LIB_WIDEN
);
2015 if (dest_reg
!= result
)
2016 emit_move_insn (dest_reg
, result
);
2019 /* Searches the insns between INSN and LOOP_END. Returns 1 if there
2020 is a backward branch in that range that branches to somewhere between
2021 LOOP_START and INSN. Returns 0 otherwise. */
2023 /* ??? This is quadratic algorithm. Could be rewritten to be linear.
2024 In practice, this is not a problem, because this function is seldom called,
2025 and uses a negligible amount of CPU time on average. */
2028 back_branch_in_range_p (insn
, loop_start
, loop_end
)
2030 rtx loop_start
, loop_end
;
2032 rtx p
, q
, target_insn
;
2034 /* Stop before we get to the backward branch at the end of the loop. */
2035 loop_end
= prev_nonnote_insn (loop_end
);
2036 if (GET_CODE (loop_end
) == BARRIER
)
2037 loop_end
= PREV_INSN (loop_end
);
2039 /* Check in case insn has been deleted, search forward for first non
2040 deleted insn following it. */
2041 while (INSN_DELETED_P (insn
))
2042 insn
= NEXT_INSN (insn
);
2044 /* Check for the case where insn is the last insn in the loop. */
2045 if (insn
== loop_end
)
2048 for (p
= NEXT_INSN (insn
); p
!= loop_end
; p
= NEXT_INSN (p
))
2050 if (GET_CODE (p
) == JUMP_INSN
)
2052 target_insn
= JUMP_LABEL (p
);
2054 /* Search from loop_start to insn, to see if one of them is
2055 the target_insn. We can't use INSN_LUID comparisons here,
2056 since insn may not have an LUID entry. */
2057 for (q
= loop_start
; q
!= insn
; q
= NEXT_INSN (q
))
2058 if (q
== target_insn
)
2066 /* Try to generate the simplest rtx for the expression
2067 (PLUS (MULT mult1 mult2) add1). This is used to calculate the initial
2071 fold_rtx_mult_add (mult1
, mult2
, add1
, mode
)
2072 rtx mult1
, mult2
, add1
;
2073 enum machine_mode mode
;
2078 /* The modes must all be the same. This should always be true. For now,
2079 check to make sure. */
2080 if ((GET_MODE (mult1
) != mode
&& GET_MODE (mult1
) != VOIDmode
)
2081 || (GET_MODE (mult2
) != mode
&& GET_MODE (mult2
) != VOIDmode
)
2082 || (GET_MODE (add1
) != mode
&& GET_MODE (add1
) != VOIDmode
))
2085 /* Ensure that if at least one of mult1/mult2 are constant, then mult2
2086 will be a constant. */
2087 if (GET_CODE (mult1
) == CONST_INT
)
2094 mult_res
= simplify_binary_operation (MULT
, mode
, mult1
, mult2
);
2096 mult_res
= gen_rtx (MULT
, mode
, mult1
, mult2
);
2098 /* Again, put the constant second. */
2099 if (GET_CODE (add1
) == CONST_INT
)
2106 result
= simplify_binary_operation (PLUS
, mode
, add1
, mult_res
);
2108 result
= gen_rtx (PLUS
, mode
, add1
, mult_res
);
2113 /* Searches the list of induction struct's for the biv BL, to try to calculate
2114 the total increment value for one iteration of the loop as a constant.
2116 Returns the increment value as an rtx, simplified as much as possible,
2117 if it can be calculated. Otherwise, returns 0. */
2120 biv_total_increment (bl
, loop_start
, loop_end
)
2121 struct iv_class
*bl
;
2122 rtx loop_start
, loop_end
;
2124 struct induction
*v
;
2127 /* For increment, must check every instruction that sets it. Each
2128 instruction must be executed only once each time through the loop.
2129 To verify this, we check that the the insn is always executed, and that
2130 there are no backward branches after the insn that branch to before it.
2131 Also, the insn must have a mult_val of one (to make sure it really is
2134 result
= const0_rtx
;
2135 for (v
= bl
->biv
; v
; v
= v
->next_iv
)
2137 if (v
->always_computable
&& v
->mult_val
== const1_rtx
2138 && ! back_branch_in_range_p (v
->insn
, loop_start
, loop_end
))
2139 result
= fold_rtx_mult_add (result
, const1_rtx
, v
->add_val
, v
->mode
);
2147 /* Determine the initial value of the iteration variable, and the amount
2148 that it is incremented each loop. Use the tables constructed by
2149 the strength reduction pass to calculate these values.
2151 Initial_value and/or increment are set to zero if their values could not
2155 iteration_info (iteration_var
, initial_value
, increment
, loop_start
, loop_end
)
2156 rtx iteration_var
, *initial_value
, *increment
;
2157 rtx loop_start
, loop_end
;
2159 struct iv_class
*bl
;
2160 struct induction
*v
, *b
;
2162 /* Clear the result values, in case no answer can be found. */
2166 /* The iteration variable can be either a giv or a biv. Check to see
2167 which it is, and compute the variable's initial value, and increment
2168 value if possible. */
2170 /* If this is a new register, can't handle it since we don't have any
2171 reg_iv_type entry for it. */
2172 if (REGNO (iteration_var
) >= max_reg_before_loop
)
2174 if (loop_dump_stream
)
2175 fprintf (loop_dump_stream
,
2176 "Loop unrolling: No reg_iv_type entry for iteration var.\n");
2179 /* Reject iteration variables larger than the host long size, since they
2180 could result in a number of iterations greater than the range of our
2181 `unsigned long' variable loop_n_iterations. */
2182 else if (GET_MODE_BITSIZE (GET_MODE (iteration_var
)) > HOST_BITS_PER_LONG
)
2184 if (loop_dump_stream
)
2185 fprintf (loop_dump_stream
,
2186 "Loop unrolling: Iteration var rejected because mode larger than host long.\n");
2189 else if (GET_MODE_CLASS (GET_MODE (iteration_var
)) != MODE_INT
)
2191 if (loop_dump_stream
)
2192 fprintf (loop_dump_stream
,
2193 "Loop unrolling: Iteration var not an integer.\n");
2196 else if (reg_iv_type
[REGNO (iteration_var
)] == BASIC_INDUCT
)
2198 /* Grab initial value, only useful if it is a constant. */
2199 bl
= reg_biv_class
[REGNO (iteration_var
)];
2200 *initial_value
= bl
->initial_value
;
2202 *increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2204 else if (reg_iv_type
[REGNO (iteration_var
)] == GENERAL_INDUCT
)
2207 /* ??? The code below does not work because the incorrect number of
2208 iterations is calculated when the biv is incremented after the giv
2209 is set (which is the usual case). This can probably be accounted
2210 for by biasing the initial_value by subtracting the amount of the
2211 increment that occurs between the giv set and the giv test. However,
2212 a giv as an iterator is very rare, so it does not seem worthwhile
2214 /* ??? An example failure is: i = 6; do {;} while (i++ < 9). */
2215 if (loop_dump_stream
)
2216 fprintf (loop_dump_stream
,
2217 "Loop unrolling: Giv iterators are not handled.\n");
2220 /* Initial value is mult_val times the biv's initial value plus
2221 add_val. Only useful if it is a constant. */
2222 v
= reg_iv_info
[REGNO (iteration_var
)];
2223 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
2224 *initial_value
= fold_rtx_mult_add (v
->mult_val
, bl
->initial_value
,
2225 v
->add_val
, v
->mode
);
2227 /* Increment value is mult_val times the increment value of the biv. */
2229 *increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2231 *increment
= fold_rtx_mult_add (v
->mult_val
, *increment
, const0_rtx
,
2237 if (loop_dump_stream
)
2238 fprintf (loop_dump_stream
,
2239 "Loop unrolling: Not basic or general induction var.\n");
2244 /* Calculate the approximate final value of the iteration variable
2245 which has an loop exit test with code COMPARISON_CODE and comparison value
2246 of COMPARISON_VALUE. Also returns an indication of whether the comparison
2247 was signed or unsigned, and the direction of the comparison. This info is
2248 needed to calculate the number of loop iterations. */
2251 approx_final_value (comparison_code
, comparison_value
, unsigned_p
, compare_dir
)
2252 enum rtx_code comparison_code
;
2253 rtx comparison_value
;
2257 /* Calculate the final value of the induction variable.
2258 The exact final value depends on the branch operator, and increment sign.
2259 This is only an approximate value. It will be wrong if the iteration
2260 variable is not incremented by one each time through the loop, and
2261 approx final value - start value % increment != 0. */
2264 switch (comparison_code
)
2270 return plus_constant (comparison_value
, 1);
2275 return plus_constant (comparison_value
, -1);
2277 /* Can not calculate a final value for this case. */
2284 return comparison_value
;
2290 return comparison_value
;
2293 return comparison_value
;
2299 /* For each biv and giv, determine whether it can be safely split into
2300 a different variable for each unrolled copy of the loop body. If it
2301 is safe to split, then indicate that by saving some useful info
2302 in the splittable_regs array.
2304 If the loop is being completely unrolled, then splittable_regs will hold
2305 the current value of the induction variable while the loop is unrolled.
2306 It must be set to the initial value of the induction variable here.
2307 Otherwise, splittable_regs will hold the difference between the current
2308 value of the induction variable and the value the induction variable had
2309 at the top of the loop. It must be set to the value 0 here.
2311 Returns the total number of instructions that set registers that are
2314 /* ?? If the loop is only unrolled twice, then most of the restrictions to
2315 constant values are unnecessary, since we can easily calculate increment
2316 values in this case even if nothing is constant. The increment value
2317 should not involve a multiply however. */
2319 /* ?? Even if the biv/giv increment values aren't constant, it may still
2320 be beneficial to split the variable if the loop is only unrolled a few
2321 times, since multiplies by small integers (1,2,3,4) are very cheap. */
2324 find_splittable_regs (unroll_type
, loop_start
, loop_end
, end_insert_before
,
2326 enum unroll_types unroll_type
;
2327 rtx loop_start
, loop_end
;
2328 rtx end_insert_before
;
2331 struct iv_class
*bl
;
2332 struct induction
*v
;
2334 rtx biv_final_value
;
2338 for (bl
= loop_iv_list
; bl
; bl
= bl
->next
)
2340 /* Biv_total_increment must return a constant value,
2341 otherwise we can not calculate the split values. */
2343 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2344 if (! increment
|| GET_CODE (increment
) != CONST_INT
)
2347 /* The loop must be unrolled completely, or else have a known number
2348 of iterations and only one exit, or else the biv must be dead
2349 outside the loop, or else the final value must be known. Otherwise,
2350 it is unsafe to split the biv since it may not have the proper
2351 value on loop exit. */
2353 /* loop_number_exit_labels is non-zero if the loop has an exit other than
2354 a fall through at the end. */
2357 biv_final_value
= 0;
2358 if (unroll_type
!= UNROLL_COMPLETELY
2359 && (loop_number_exit_labels
[uid_loop_num
[INSN_UID (loop_start
)]]
2360 || unroll_type
== UNROLL_NAIVE
)
2361 && (uid_luid
[regno_last_uid
[bl
->regno
]] >= INSN_LUID (loop_end
)
2363 || INSN_UID (bl
->init_insn
) >= max_uid_for_loop
2364 || (uid_luid
[regno_first_uid
[bl
->regno
]]
2365 < INSN_LUID (bl
->init_insn
))
2366 || reg_mentioned_p (bl
->biv
->dest_reg
, SET_SRC (bl
->init_set
)))
2367 && ! (biv_final_value
= final_biv_value (bl
, loop_start
, loop_end
)))
2370 /* If any of the insns setting the BIV don't do so with a simple
2371 PLUS, we don't know how to split it. */
2372 for (v
= bl
->biv
; biv_splittable
&& v
; v
= v
->next_iv
)
2373 if ((tem
= single_set (v
->insn
)) == 0
2374 || GET_CODE (SET_DEST (tem
)) != REG
2375 || REGNO (SET_DEST (tem
)) != bl
->regno
2376 || GET_CODE (SET_SRC (tem
)) != PLUS
)
2379 /* If final value is non-zero, then must emit an instruction which sets
2380 the value of the biv to the proper value. This is done after
2381 handling all of the givs, since some of them may need to use the
2382 biv's value in their initialization code. */
2384 /* This biv is splittable. If completely unrolling the loop, save
2385 the biv's initial value. Otherwise, save the constant zero. */
2387 if (biv_splittable
== 1)
2389 if (unroll_type
== UNROLL_COMPLETELY
)
2391 /* If the initial value of the biv is itself (i.e. it is too
2392 complicated for strength_reduce to compute), or is a hard
2393 register, then we must create a new pseudo reg to hold the
2394 initial value of the biv. */
2396 if (GET_CODE (bl
->initial_value
) == REG
2397 && (REGNO (bl
->initial_value
) == bl
->regno
2398 || REGNO (bl
->initial_value
) < FIRST_PSEUDO_REGISTER
))
2400 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2402 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2405 if (loop_dump_stream
)
2406 fprintf (loop_dump_stream
, "Biv %d initial value remapped to %d.\n",
2407 bl
->regno
, REGNO (tem
));
2409 splittable_regs
[bl
->regno
] = tem
;
2412 splittable_regs
[bl
->regno
] = bl
->initial_value
;
2415 splittable_regs
[bl
->regno
] = const0_rtx
;
2417 /* Save the number of instructions that modify the biv, so that
2418 we can treat the last one specially. */
2420 splittable_regs_updates
[bl
->regno
] = bl
->biv_count
;
2421 result
+= bl
->biv_count
;
2423 if (loop_dump_stream
)
2424 fprintf (loop_dump_stream
,
2425 "Biv %d safe to split.\n", bl
->regno
);
2428 /* Check every giv that depends on this biv to see whether it is
2429 splittable also. Even if the biv isn't splittable, givs which
2430 depend on it may be splittable if the biv is live outside the
2431 loop, and the givs aren't. */
2433 result
+= find_splittable_givs (bl
, unroll_type
, loop_start
, loop_end
,
2434 increment
, unroll_number
);
2436 /* If final value is non-zero, then must emit an instruction which sets
2437 the value of the biv to the proper value. This is done after
2438 handling all of the givs, since some of them may need to use the
2439 biv's value in their initialization code. */
2440 if (biv_final_value
)
2442 /* If the loop has multiple exits, emit the insns before the
2443 loop to ensure that it will always be executed no matter
2444 how the loop exits. Otherwise emit the insn after the loop,
2445 since this is slightly more efficient. */
2446 if (! loop_number_exit_labels
[uid_loop_num
[INSN_UID (loop_start
)]])
2447 emit_insn_before (gen_move_insn (bl
->biv
->src_reg
,
2452 /* Create a new register to hold the value of the biv, and then
2453 set the biv to its final value before the loop start. The biv
2454 is set to its final value before loop start to ensure that
2455 this insn will always be executed, no matter how the loop
2457 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2458 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2460 emit_insn_before (gen_move_insn (bl
->biv
->src_reg
,
2464 if (loop_dump_stream
)
2465 fprintf (loop_dump_stream
, "Biv %d mapped to %d for split.\n",
2466 REGNO (bl
->biv
->src_reg
), REGNO (tem
));
2468 /* Set up the mapping from the original biv register to the new
2470 bl
->biv
->src_reg
= tem
;
2477 /* For every giv based on the biv BL, check to determine whether it is
2478 splittable. This is a subroutine to find_splittable_regs ().
2480 Return the number of instructions that set splittable registers. */
2483 find_splittable_givs (bl
, unroll_type
, loop_start
, loop_end
, increment
,
2485 struct iv_class
*bl
;
2486 enum unroll_types unroll_type
;
2487 rtx loop_start
, loop_end
;
2491 struct induction
*v
, *v2
;
2496 /* Scan the list of givs, and set the same_insn field when there are
2497 multiple identical givs in the same insn. */
2498 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
2499 for (v2
= v
->next_iv
; v2
; v2
= v2
->next_iv
)
2500 if (v
->insn
== v2
->insn
&& rtx_equal_p (v
->new_reg
, v2
->new_reg
)
2504 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
2508 /* Only split the giv if it has already been reduced, or if the loop is
2509 being completely unrolled. */
2510 if (unroll_type
!= UNROLL_COMPLETELY
&& v
->ignore
)
2513 /* The giv can be split if the insn that sets the giv is executed once
2514 and only once on every iteration of the loop. */
2515 /* An address giv can always be split. v->insn is just a use not a set,
2516 and hence it does not matter whether it is always executed. All that
2517 matters is that all the biv increments are always executed, and we
2518 won't reach here if they aren't. */
2519 if (v
->giv_type
!= DEST_ADDR
2520 && (! v
->always_computable
2521 || back_branch_in_range_p (v
->insn
, loop_start
, loop_end
)))
2524 /* The giv increment value must be a constant. */
2525 giv_inc
= fold_rtx_mult_add (v
->mult_val
, increment
, const0_rtx
,
2527 if (! giv_inc
|| GET_CODE (giv_inc
) != CONST_INT
)
2530 /* The loop must be unrolled completely, or else have a known number of
2531 iterations and only one exit, or else the giv must be dead outside
2532 the loop, or else the final value of the giv must be known.
2533 Otherwise, it is not safe to split the giv since it may not have the
2534 proper value on loop exit. */
2536 /* The used outside loop test will fail for DEST_ADDR givs. They are
2537 never used outside the loop anyways, so it is always safe to split a
2541 if (unroll_type
!= UNROLL_COMPLETELY
2542 && (loop_number_exit_labels
[uid_loop_num
[INSN_UID (loop_start
)]]
2543 || unroll_type
== UNROLL_NAIVE
)
2544 && v
->giv_type
!= DEST_ADDR
2545 && ((regno_first_uid
[REGNO (v
->dest_reg
)] != INSN_UID (v
->insn
)
2546 /* Check for the case where the pseudo is set by a shift/add
2547 sequence, in which case the first insn setting the pseudo
2548 is the first insn of the shift/add sequence. */
2549 && (! (tem
= find_reg_note (v
->insn
, REG_RETVAL
, NULL_RTX
))
2550 || (regno_first_uid
[REGNO (v
->dest_reg
)]
2551 != INSN_UID (XEXP (tem
, 0)))))
2552 /* Line above always fails if INSN was moved by loop opt. */
2553 || (uid_luid
[regno_last_uid
[REGNO (v
->dest_reg
)]]
2554 >= INSN_LUID (loop_end
)))
2555 && ! (final_value
= v
->final_value
))
2559 /* Currently, non-reduced/final-value givs are never split. */
2560 /* Should emit insns after the loop if possible, as the biv final value
2563 /* If the final value is non-zero, and the giv has not been reduced,
2564 then must emit an instruction to set the final value. */
2565 if (final_value
&& !v
->new_reg
)
2567 /* Create a new register to hold the value of the giv, and then set
2568 the giv to its final value before the loop start. The giv is set
2569 to its final value before loop start to ensure that this insn
2570 will always be executed, no matter how we exit. */
2571 tem
= gen_reg_rtx (v
->mode
);
2572 emit_insn_before (gen_move_insn (tem
, v
->dest_reg
), loop_start
);
2573 emit_insn_before (gen_move_insn (v
->dest_reg
, final_value
),
2576 if (loop_dump_stream
)
2577 fprintf (loop_dump_stream
, "Giv %d mapped to %d for split.\n",
2578 REGNO (v
->dest_reg
), REGNO (tem
));
2584 /* This giv is splittable. If completely unrolling the loop, save the
2585 giv's initial value. Otherwise, save the constant zero for it. */
2587 if (unroll_type
== UNROLL_COMPLETELY
)
2589 /* It is not safe to use bl->initial_value here, because it may not
2590 be invariant. It is safe to use the initial value stored in
2591 the splittable_regs array if it is set. In rare cases, it won't
2592 be set, so then we do exactly the same thing as
2593 find_splittable_regs does to get a safe value. */
2594 rtx biv_initial_value
;
2596 if (splittable_regs
[bl
->regno
])
2597 biv_initial_value
= splittable_regs
[bl
->regno
];
2598 else if (GET_CODE (bl
->initial_value
) != REG
2599 || (REGNO (bl
->initial_value
) != bl
->regno
2600 && REGNO (bl
->initial_value
) >= FIRST_PSEUDO_REGISTER
))
2601 biv_initial_value
= bl
->initial_value
;
2604 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2606 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2608 biv_initial_value
= tem
;
2610 value
= fold_rtx_mult_add (v
->mult_val
, biv_initial_value
,
2611 v
->add_val
, v
->mode
);
2618 /* If a giv was combined with another giv, then we can only split
2619 this giv if the giv it was combined with was reduced. This
2620 is because the value of v->new_reg is meaningless in this
2622 if (v
->same
&& ! v
->same
->new_reg
)
2624 if (loop_dump_stream
)
2625 fprintf (loop_dump_stream
,
2626 "giv combined with unreduced giv not split.\n");
2629 /* If the giv is an address destination, it could be something other
2630 than a simple register, these have to be treated differently. */
2631 else if (v
->giv_type
== DEST_REG
)
2633 /* If value is not a constant, register, or register plus
2634 constant, then compute its value into a register before
2635 loop start. This prevents invalid rtx sharing, and should
2636 generate better code. We can use bl->initial_value here
2637 instead of splittable_regs[bl->regno] because this code
2638 is going before the loop start. */
2639 if (unroll_type
== UNROLL_COMPLETELY
2640 && GET_CODE (value
) != CONST_INT
2641 && GET_CODE (value
) != REG
2642 && (GET_CODE (value
) != PLUS
2643 || GET_CODE (XEXP (value
, 0)) != REG
2644 || GET_CODE (XEXP (value
, 1)) != CONST_INT
))
2646 rtx tem
= gen_reg_rtx (v
->mode
);
2647 emit_iv_add_mult (bl
->initial_value
, v
->mult_val
,
2648 v
->add_val
, tem
, loop_start
);
2652 splittable_regs
[REGNO (v
->new_reg
)] = value
;
2656 /* Splitting address givs is useful since it will often allow us
2657 to eliminate some increment insns for the base giv as
2660 /* If the addr giv is combined with a dest_reg giv, then all
2661 references to that dest reg will be remapped, which is NOT
2662 what we want for split addr regs. We always create a new
2663 register for the split addr giv, just to be safe. */
2665 /* ??? If there are multiple address givs which have been
2666 combined with the same dest_reg giv, then we may only need
2667 one new register for them. Pulling out constants below will
2668 catch some of the common cases of this. Currently, I leave
2669 the work of simplifying multiple address givs to the
2670 following cse pass. */
2672 /* As a special case, if we have multiple identical address givs
2673 within a single instruction, then we do use a single psuedo
2674 reg for both. This is necessary in case one is a match_dup
2677 v
->const_adjust
= 0;
2681 v
->dest_reg
= v
->same_insn
->dest_reg
;
2682 if (loop_dump_stream
)
2683 fprintf (loop_dump_stream
,
2684 "Sharing address givs in insn %d\n",
2685 INSN_UID (v
->insn
));
2687 else if (unroll_type
!= UNROLL_COMPLETELY
)
2689 /* If not completely unrolling the loop, then create a new
2690 register to hold the split value of the DEST_ADDR giv.
2691 Emit insn to initialize its value before loop start. */
2692 tem
= gen_reg_rtx (v
->mode
);
2694 /* If the address giv has a constant in its new_reg value,
2695 then this constant can be pulled out and put in value,
2696 instead of being part of the initialization code. */
2698 if (GET_CODE (v
->new_reg
) == PLUS
2699 && GET_CODE (XEXP (v
->new_reg
, 1)) == CONST_INT
)
2702 = plus_constant (tem
, INTVAL (XEXP (v
->new_reg
,1)));
2704 /* Only succeed if this will give valid addresses.
2705 Try to validate both the first and the last
2706 address resulting from loop unrolling, if
2707 one fails, then can't do const elim here. */
2708 if (memory_address_p (v
->mem_mode
, v
->dest_reg
)
2709 && memory_address_p (v
->mem_mode
,
2710 plus_constant (v
->dest_reg
,
2712 * (unroll_number
- 1))))
2714 /* Save the negative of the eliminated const, so
2715 that we can calculate the dest_reg's increment
2717 v
->const_adjust
= - INTVAL (XEXP (v
->new_reg
, 1));
2719 v
->new_reg
= XEXP (v
->new_reg
, 0);
2720 if (loop_dump_stream
)
2721 fprintf (loop_dump_stream
,
2722 "Eliminating constant from giv %d\n",
2731 /* If the address hasn't been checked for validity yet, do so
2732 now, and fail completely if either the first or the last
2733 unrolled copy of the address is not a valid address. */
2734 if (v
->dest_reg
== tem
2735 && (! memory_address_p (v
->mem_mode
, v
->dest_reg
)
2736 || ! memory_address_p (v
->mem_mode
,
2737 plus_constant (v
->dest_reg
,
2739 * (unroll_number
-1)))))
2741 if (loop_dump_stream
)
2742 fprintf (loop_dump_stream
,
2743 "Invalid address for giv at insn %d\n",
2744 INSN_UID (v
->insn
));
2748 /* To initialize the new register, just move the value of
2749 new_reg into it. This is not guaranteed to give a valid
2750 instruction on machines with complex addressing modes.
2751 If we can't recognize it, then delete it and emit insns
2752 to calculate the value from scratch. */
2753 emit_insn_before (gen_rtx (SET
, VOIDmode
, tem
,
2754 copy_rtx (v
->new_reg
)),
2756 if (recog_memoized (PREV_INSN (loop_start
)) < 0)
2760 /* We can't use bl->initial_value to compute the initial
2761 value, because the loop may have been preconditioned.
2762 We must calculate it from NEW_REG. Try using
2763 force_operand instead of emit_iv_add_mult. */
2764 delete_insn (PREV_INSN (loop_start
));
2767 ret
= force_operand (v
->new_reg
, tem
);
2769 emit_move_insn (tem
, ret
);
2770 sequence
= gen_sequence ();
2772 emit_insn_before (sequence
, loop_start
);
2774 if (loop_dump_stream
)
2775 fprintf (loop_dump_stream
,
2776 "Invalid init insn, rewritten.\n");
2781 v
->dest_reg
= value
;
2783 /* Check the resulting address for validity, and fail
2784 if the resulting address would be invalid. */
2785 if (! memory_address_p (v
->mem_mode
, v
->dest_reg
)
2786 || ! memory_address_p (v
->mem_mode
,
2787 plus_constant (v
->dest_reg
,
2789 (unroll_number
-1))))
2791 if (loop_dump_stream
)
2792 fprintf (loop_dump_stream
,
2793 "Invalid address for giv at insn %d\n",
2794 INSN_UID (v
->insn
));
2799 /* Store the value of dest_reg into the insn. This sharing
2800 will not be a problem as this insn will always be copied
2803 *v
->location
= v
->dest_reg
;
2805 /* If this address giv is combined with a dest reg giv, then
2806 save the base giv's induction pointer so that we will be
2807 able to handle this address giv properly. The base giv
2808 itself does not have to be splittable. */
2810 if (v
->same
&& v
->same
->giv_type
== DEST_REG
)
2811 addr_combined_regs
[REGNO (v
->same
->new_reg
)] = v
->same
;
2813 if (GET_CODE (v
->new_reg
) == REG
)
2815 /* This giv maybe hasn't been combined with any others.
2816 Make sure that it's giv is marked as splittable here. */
2818 splittable_regs
[REGNO (v
->new_reg
)] = value
;
2820 /* Make it appear to depend upon itself, so that the
2821 giv will be properly split in the main loop above. */
2825 addr_combined_regs
[REGNO (v
->new_reg
)] = v
;
2829 if (loop_dump_stream
)
2830 fprintf (loop_dump_stream
, "DEST_ADDR giv being split.\n");
2836 /* Currently, unreduced giv's can't be split. This is not too much
2837 of a problem since unreduced giv's are not live across loop
2838 iterations anyways. When unrolling a loop completely though,
2839 it makes sense to reduce&split givs when possible, as this will
2840 result in simpler instructions, and will not require that a reg
2841 be live across loop iterations. */
2843 splittable_regs
[REGNO (v
->dest_reg
)] = value
;
2844 fprintf (stderr
, "Giv %d at insn %d not reduced\n",
2845 REGNO (v
->dest_reg
), INSN_UID (v
->insn
));
2851 /* Givs are only updated once by definition. Mark it so if this is
2852 a splittable register. Don't need to do anything for address givs
2853 where this may not be a register. */
2855 if (GET_CODE (v
->new_reg
) == REG
)
2856 splittable_regs_updates
[REGNO (v
->new_reg
)] = 1;
2860 if (loop_dump_stream
)
2864 if (GET_CODE (v
->dest_reg
) == CONST_INT
)
2866 else if (GET_CODE (v
->dest_reg
) != REG
)
2867 regnum
= REGNO (XEXP (v
->dest_reg
, 0));
2869 regnum
= REGNO (v
->dest_reg
);
2870 fprintf (loop_dump_stream
, "Giv %d at insn %d safe to split.\n",
2871 regnum
, INSN_UID (v
->insn
));
2878 /* Try to prove that the register is dead after the loop exits. Trace every
2879 loop exit looking for an insn that will always be executed, which sets
2880 the register to some value, and appears before the first use of the register
2881 is found. If successful, then return 1, otherwise return 0. */
2883 /* ?? Could be made more intelligent in the handling of jumps, so that
2884 it can search past if statements and other similar structures. */
2887 reg_dead_after_loop (reg
, loop_start
, loop_end
)
2888 rtx reg
, loop_start
, loop_end
;
2894 /* HACK: Must also search the loop fall through exit, create a label_ref
2895 here which points to the loop_end, and append the loop_number_exit_labels
2897 label
= gen_rtx (LABEL_REF
, VOIDmode
, loop_end
);
2898 LABEL_NEXTREF (label
)
2899 = loop_number_exit_labels
[uid_loop_num
[INSN_UID (loop_start
)]];
2901 for ( ; label
; label
= LABEL_NEXTREF (label
))
2903 /* Succeed if find an insn which sets the biv or if reach end of
2904 function. Fail if find an insn that uses the biv, or if come to
2905 a conditional jump. */
2907 insn
= NEXT_INSN (XEXP (label
, 0));
2910 code
= GET_CODE (insn
);
2911 if (GET_RTX_CLASS (code
) == 'i')
2915 if (reg_referenced_p (reg
, PATTERN (insn
)))
2918 set
= single_set (insn
);
2919 if (set
&& rtx_equal_p (SET_DEST (set
), reg
))
2923 if (code
== JUMP_INSN
)
2925 if (GET_CODE (PATTERN (insn
)) == RETURN
)
2927 else if (! simplejump_p (insn
)
2928 /* Prevent infinite loop following infinite loops. */
2929 || jump_count
++ > 20)
2932 insn
= JUMP_LABEL (insn
);
2935 insn
= NEXT_INSN (insn
);
2939 /* Success, the register is dead on all loop exits. */
2943 /* Try to calculate the final value of the biv, the value it will have at
2944 the end of the loop. If we can do it, return that value. */
2947 final_biv_value (bl
, loop_start
, loop_end
)
2948 struct iv_class
*bl
;
2949 rtx loop_start
, loop_end
;
2953 /* ??? This only works for MODE_INT biv's. Reject all others for now. */
2955 if (GET_MODE_CLASS (bl
->biv
->mode
) != MODE_INT
)
2958 /* The final value for reversed bivs must be calculated differently than
2959 for ordinary bivs. In this case, there is already an insn after the
2960 loop which sets this biv's final value (if necessary), and there are
2961 no other loop exits, so we can return any value. */
2964 if (loop_dump_stream
)
2965 fprintf (loop_dump_stream
,
2966 "Final biv value for %d, reversed biv.\n", bl
->regno
);
2971 /* Try to calculate the final value as initial value + (number of iterations
2972 * increment). For this to work, increment must be invariant, the only
2973 exit from the loop must be the fall through at the bottom (otherwise
2974 it may not have its final value when the loop exits), and the initial
2975 value of the biv must be invariant. */
2977 if (loop_n_iterations
!= 0
2978 && ! loop_number_exit_labels
[uid_loop_num
[INSN_UID (loop_start
)]]
2979 && invariant_p (bl
->initial_value
))
2981 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2983 if (increment
&& invariant_p (increment
))
2985 /* Can calculate the loop exit value, emit insns after loop
2986 end to calculate this value into a temporary register in
2987 case it is needed later. */
2989 tem
= gen_reg_rtx (bl
->biv
->mode
);
2990 /* Make sure loop_end is not the last insn. */
2991 if (NEXT_INSN (loop_end
) == 0)
2992 emit_note_after (NOTE_INSN_DELETED
, loop_end
);
2993 emit_iv_add_mult (increment
, GEN_INT (loop_n_iterations
),
2994 bl
->initial_value
, tem
, NEXT_INSN (loop_end
));
2996 if (loop_dump_stream
)
2997 fprintf (loop_dump_stream
,
2998 "Final biv value for %d, calculated.\n", bl
->regno
);
3004 /* Check to see if the biv is dead at all loop exits. */
3005 if (reg_dead_after_loop (bl
->biv
->src_reg
, loop_start
, loop_end
))
3007 if (loop_dump_stream
)
3008 fprintf (loop_dump_stream
,
3009 "Final biv value for %d, biv dead after loop exit.\n",
3018 /* Try to calculate the final value of the giv, the value it will have at
3019 the end of the loop. If we can do it, return that value. */
3022 final_giv_value (v
, loop_start
, loop_end
)
3023 struct induction
*v
;
3024 rtx loop_start
, loop_end
;
3026 struct iv_class
*bl
;
3029 rtx insert_before
, seq
;
3031 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
3033 /* The final value for givs which depend on reversed bivs must be calculated
3034 differently than for ordinary givs. In this case, there is already an
3035 insn after the loop which sets this giv's final value (if necessary),
3036 and there are no other loop exits, so we can return any value. */
3039 if (loop_dump_stream
)
3040 fprintf (loop_dump_stream
,
3041 "Final giv value for %d, depends on reversed biv\n",
3042 REGNO (v
->dest_reg
));
3046 /* Try to calculate the final value as a function of the biv it depends
3047 upon. The only exit from the loop must be the fall through at the bottom
3048 (otherwise it may not have its final value when the loop exits). */
3050 /* ??? Can calculate the final giv value by subtracting off the
3051 extra biv increments times the giv's mult_val. The loop must have
3052 only one exit for this to work, but the loop iterations does not need
3055 if (loop_n_iterations
!= 0
3056 && ! loop_number_exit_labels
[uid_loop_num
[INSN_UID (loop_start
)]])
3058 /* ?? It is tempting to use the biv's value here since these insns will
3059 be put after the loop, and hence the biv will have its final value
3060 then. However, this fails if the biv is subsequently eliminated.
3061 Perhaps determine whether biv's are eliminable before trying to
3062 determine whether giv's are replaceable so that we can use the
3063 biv value here if it is not eliminable. */
3065 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
3067 if (increment
&& invariant_p (increment
))
3069 /* Can calculate the loop exit value of its biv as
3070 (loop_n_iterations * increment) + initial_value */
3072 /* The loop exit value of the giv is then
3073 (final_biv_value - extra increments) * mult_val + add_val.
3074 The extra increments are any increments to the biv which
3075 occur in the loop after the giv's value is calculated.
3076 We must search from the insn that sets the giv to the end
3077 of the loop to calculate this value. */
3079 insert_before
= NEXT_INSN (loop_end
);
3081 /* Put the final biv value in tem. */
3082 tem
= gen_reg_rtx (bl
->biv
->mode
);
3083 emit_iv_add_mult (increment
, GEN_INT (loop_n_iterations
),
3084 bl
->initial_value
, tem
, insert_before
);
3086 /* Subtract off extra increments as we find them. */
3087 for (insn
= NEXT_INSN (v
->insn
); insn
!= loop_end
;
3088 insn
= NEXT_INSN (insn
))
3090 struct induction
*biv
;
3092 for (biv
= bl
->biv
; biv
; biv
= biv
->next_iv
)
3093 if (biv
->insn
== insn
)
3096 tem
= expand_binop (GET_MODE (tem
), sub_optab
, tem
,
3097 biv
->add_val
, NULL_RTX
, 0,
3099 seq
= gen_sequence ();
3101 emit_insn_before (seq
, insert_before
);
3105 /* Now calculate the giv's final value. */
3106 emit_iv_add_mult (tem
, v
->mult_val
, v
->add_val
, tem
,
3109 if (loop_dump_stream
)
3110 fprintf (loop_dump_stream
,
3111 "Final giv value for %d, calc from biv's value.\n",
3112 REGNO (v
->dest_reg
));
3118 /* Replaceable giv's should never reach here. */
3122 /* Check to see if the biv is dead at all loop exits. */
3123 if (reg_dead_after_loop (v
->dest_reg
, loop_start
, loop_end
))
3125 if (loop_dump_stream
)
3126 fprintf (loop_dump_stream
,
3127 "Final giv value for %d, giv dead after loop exit.\n",
3128 REGNO (v
->dest_reg
));
3137 /* Calculate the number of loop iterations. Returns the exact number of loop
3138 iterations if it can be calculated, otherwise returns zero. */
3140 unsigned HOST_WIDE_INT
3141 loop_iterations (loop_start
, loop_end
)
3142 rtx loop_start
, loop_end
;
3144 rtx comparison
, comparison_value
;
3145 rtx iteration_var
, initial_value
, increment
, final_value
;
3146 enum rtx_code comparison_code
;
3149 int unsigned_compare
, compare_dir
, final_larger
;
3150 unsigned long tempu
;
3153 /* First find the iteration variable. If the last insn is a conditional
3154 branch, and the insn before tests a register value, make that the
3155 iteration variable. */
3157 loop_initial_value
= 0;
3159 loop_final_value
= 0;
3160 loop_iteration_var
= 0;
3162 /* We used to use pren_nonnote_insn here, but that fails because it might
3163 accidentally get the branch for a contained loop if the branch for this
3164 loop was deleted. We can only trust branches immediately before the
3166 last_loop_insn
= PREV_INSN (loop_end
);
3168 comparison
= get_condition_for_loop (last_loop_insn
);
3169 if (comparison
== 0)
3171 if (loop_dump_stream
)
3172 fprintf (loop_dump_stream
,
3173 "Loop unrolling: No final conditional branch found.\n");
3177 /* ??? Get_condition may switch position of induction variable and
3178 invariant register when it canonicalizes the comparison. */
3180 comparison_code
= GET_CODE (comparison
);
3181 iteration_var
= XEXP (comparison
, 0);
3182 comparison_value
= XEXP (comparison
, 1);
3184 if (GET_CODE (iteration_var
) != REG
)
3186 if (loop_dump_stream
)
3187 fprintf (loop_dump_stream
,
3188 "Loop unrolling: Comparison not against register.\n");
3192 /* Loop iterations is always called before any new registers are created
3193 now, so this should never occur. */
3195 if (REGNO (iteration_var
) >= max_reg_before_loop
)
3198 iteration_info (iteration_var
, &initial_value
, &increment
,
3199 loop_start
, loop_end
);
3200 if (initial_value
== 0)
3201 /* iteration_info already printed a message. */
3204 /* If the comparison value is an invariant register, then try to find
3205 its value from the insns before the start of the loop. */
3207 if (GET_CODE (comparison_value
) == REG
&& invariant_p (comparison_value
))
3211 for (insn
= PREV_INSN (loop_start
); insn
; insn
= PREV_INSN (insn
))
3213 if (GET_CODE (insn
) == CODE_LABEL
)
3216 else if (GET_RTX_CLASS (GET_CODE (insn
)) == 'i'
3217 && reg_set_p (comparison_value
, insn
))
3219 /* We found the last insn before the loop that sets the register.
3220 If it sets the entire register, and has a REG_EQUAL note,
3221 then use the value of the REG_EQUAL note. */
3222 if ((set
= single_set (insn
))
3223 && (SET_DEST (set
) == comparison_value
))
3225 rtx note
= find_reg_note (insn
, REG_EQUAL
, NULL_RTX
);
3227 /* Only use the REG_EQUAL note if it is a constant.
3228 Other things, divide in particular, will cause
3229 problems later if we use them. */
3230 if (note
&& GET_CODE (XEXP (note
, 0)) != EXPR_LIST
3231 && CONSTANT_P (XEXP (note
, 0)))
3232 comparison_value
= XEXP (note
, 0);
3239 final_value
= approx_final_value (comparison_code
, comparison_value
,
3240 &unsigned_compare
, &compare_dir
);
3242 /* Save the calculated values describing this loop's bounds, in case
3243 precondition_loop_p will need them later. These values can not be
3244 recalculated inside precondition_loop_p because strength reduction
3245 optimizations may obscure the loop's structure. */
3247 loop_iteration_var
= iteration_var
;
3248 loop_initial_value
= initial_value
;
3249 loop_increment
= increment
;
3250 loop_final_value
= final_value
;
3254 if (loop_dump_stream
)
3255 fprintf (loop_dump_stream
,
3256 "Loop unrolling: Increment value can't be calculated.\n");
3259 else if (GET_CODE (increment
) != CONST_INT
)
3261 if (loop_dump_stream
)
3262 fprintf (loop_dump_stream
,
3263 "Loop unrolling: Increment value not constant.\n");
3266 else if (GET_CODE (initial_value
) != CONST_INT
)
3268 if (loop_dump_stream
)
3269 fprintf (loop_dump_stream
,
3270 "Loop unrolling: Initial value not constant.\n");
3273 else if (final_value
== 0)
3275 if (loop_dump_stream
)
3276 fprintf (loop_dump_stream
,
3277 "Loop unrolling: EQ comparison loop.\n");
3280 else if (GET_CODE (final_value
) != CONST_INT
)
3282 if (loop_dump_stream
)
3283 fprintf (loop_dump_stream
,
3284 "Loop unrolling: Final value not constant.\n");
3288 /* ?? Final value and initial value do not have to be constants.
3289 Only their difference has to be constant. When the iteration variable
3290 is an array address, the final value and initial value might both
3291 be addresses with the same base but different constant offsets.
3292 Final value must be invariant for this to work.
3294 To do this, need some way to find the values of registers which are
3297 /* Final_larger is 1 if final larger, 0 if they are equal, otherwise -1. */
3298 if (unsigned_compare
)
3300 = ((unsigned HOST_WIDE_INT
) INTVAL (final_value
)
3301 > (unsigned HOST_WIDE_INT
) INTVAL (initial_value
))
3302 - ((unsigned HOST_WIDE_INT
) INTVAL (final_value
)
3303 < (unsigned HOST_WIDE_INT
) INTVAL (initial_value
));
3305 final_larger
= (INTVAL (final_value
) > INTVAL (initial_value
))
3306 - (INTVAL (final_value
) < INTVAL (initial_value
));
3308 if (INTVAL (increment
) > 0)
3310 else if (INTVAL (increment
) == 0)
3315 /* There are 27 different cases: compare_dir = -1, 0, 1;
3316 final_larger = -1, 0, 1; increment_dir = -1, 0, 1.
3317 There are 4 normal cases, 4 reverse cases (where the iteration variable
3318 will overflow before the loop exits), 4 infinite loop cases, and 15
3319 immediate exit (0 or 1 iteration depending on loop type) cases.
3320 Only try to optimize the normal cases. */
3322 /* (compare_dir/final_larger/increment_dir)
3323 Normal cases: (0/-1/-1), (0/1/1), (-1/-1/-1), (1/1/1)
3324 Reverse cases: (0/-1/1), (0/1/-1), (-1/-1/1), (1/1/-1)
3325 Infinite loops: (0/-1/0), (0/1/0), (-1/-1/0), (1/1/0)
3326 Immediate exit: (0/0/X), (-1/0/X), (-1/1/X), (1/0/X), (1/-1/X) */
3328 /* ?? If the meaning of reverse loops (where the iteration variable
3329 will overflow before the loop exits) is undefined, then could
3330 eliminate all of these special checks, and just always assume
3331 the loops are normal/immediate/infinite. Note that this means
3332 the sign of increment_dir does not have to be known. Also,
3333 since it does not really hurt if immediate exit loops or infinite loops
3334 are optimized, then that case could be ignored also, and hence all
3335 loops can be optimized.
3337 According to ANSI Spec, the reverse loop case result is undefined,
3338 because the action on overflow is undefined.
3340 See also the special test for NE loops below. */
3342 if (final_larger
== increment_dir
&& final_larger
!= 0
3343 && (final_larger
== compare_dir
|| compare_dir
== 0))
3348 if (loop_dump_stream
)
3349 fprintf (loop_dump_stream
,
3350 "Loop unrolling: Not normal loop.\n");
3354 /* Calculate the number of iterations, final_value is only an approximation,
3355 so correct for that. Note that tempu and loop_n_iterations are
3356 unsigned, because they can be as large as 2^n - 1. */
3358 i
= INTVAL (increment
);
3360 tempu
= INTVAL (final_value
) - INTVAL (initial_value
);
3363 tempu
= INTVAL (initial_value
) - INTVAL (final_value
);
3369 /* For NE tests, make sure that the iteration variable won't miss the
3370 final value. If tempu mod i is not zero, then the iteration variable
3371 will overflow before the loop exits, and we can not calculate the
3372 number of iterations. */
3373 if (compare_dir
== 0 && (tempu
% i
) != 0)
3376 return tempu
/ i
+ ((tempu
% i
) != 0);
3379 /* Replace uses of split bivs with their split psuedo register. This is
3380 for original instructions which remain after loop unrolling without
3384 remap_split_bivs (x
)
3387 register enum rtx_code code
;
3394 code
= GET_CODE (x
);
3409 /* If non-reduced/final-value givs were split, then this would also
3410 have to remap those givs also. */
3412 if (REGNO (x
) < max_reg_before_loop
3413 && reg_iv_type
[REGNO (x
)] == BASIC_INDUCT
)
3414 return reg_biv_class
[REGNO (x
)]->biv
->src_reg
;
3417 fmt
= GET_RTX_FORMAT (code
);
3418 for (i
= GET_RTX_LENGTH (code
) - 1; i
>= 0; i
--)
3421 XEXP (x
, i
) = remap_split_bivs (XEXP (x
, i
));
3425 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
3426 XVECEXP (x
, i
, j
) = remap_split_bivs (XVECEXP (x
, i
, j
));