1 /* Try to unroll loops, and split induction variables.
2 Copyright (C) 1992, 1993, 1994, 1995 Free Software Foundation, Inc.
3 Contributed by James E. Wilson, Cygnus Support/UC Berkeley.
5 This file is part of GNU CC.
7 GNU CC is free software; you can redistribute it and/or modify
8 it under the terms of the GNU General Public License as published by
9 the Free Software Foundation; either version 2, or (at your option)
12 GNU CC is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License for more details.
17 You should have received a copy of the GNU General Public License
18 along with GNU CC; see the file COPYING. If not, write to
19 the Free Software Foundation, 59 Temple Place - Suite 330,
20 Boston, MA 02111-1307, USA. */
22 /* Try to unroll a loop, and split induction variables.
24 Loops for which the number of iterations can be calculated exactly are
25 handled specially. If the number of iterations times the insn_count is
26 less than MAX_UNROLLED_INSNS, then the loop is unrolled completely.
27 Otherwise, we try to unroll the loop a number of times modulo the number
28 of iterations, so that only one exit test will be needed. It is unrolled
29 a number of times approximately equal to MAX_UNROLLED_INSNS divided by
32 Otherwise, if the number of iterations can be calculated exactly at
33 run time, and the loop is always entered at the top, then we try to
34 precondition the loop. That is, at run time, calculate how many times
35 the loop will execute, and then execute the loop body a few times so
36 that the remaining iterations will be some multiple of 4 (or 2 if the
37 loop is large). Then fall through to a loop unrolled 4 (or 2) times,
38 with only one exit test needed at the end of the loop.
40 Otherwise, if the number of iterations can not be calculated exactly,
41 not even at run time, then we still unroll the loop a number of times
42 approximately equal to MAX_UNROLLED_INSNS divided by the insn count,
43 but there must be an exit test after each copy of the loop body.
45 For each induction variable, which is dead outside the loop (replaceable)
46 or for which we can easily calculate the final value, if we can easily
47 calculate its value at each place where it is set as a function of the
48 current loop unroll count and the variable's value at loop entry, then
49 the induction variable is split into `N' different variables, one for
50 each copy of the loop body. One variable is live across the backward
51 branch, and the others are all calculated as a function of this variable.
52 This helps eliminate data dependencies, and leads to further opportunities
55 /* Possible improvements follow: */
57 /* ??? Add an extra pass somewhere to determine whether unrolling will
58 give any benefit. E.g. after generating all unrolled insns, compute the
59 cost of all insns and compare against cost of insns in rolled loop.
61 - On traditional architectures, unrolling a non-constant bound loop
62 is a win if there is a giv whose only use is in memory addresses, the
63 memory addresses can be split, and hence giv increments can be
65 - It is also a win if the loop is executed many times, and preconditioning
66 can be performed for the loop.
67 Add code to check for these and similar cases. */
69 /* ??? Improve control of which loops get unrolled. Could use profiling
70 info to only unroll the most commonly executed loops. Perhaps have
71 a user specifyable option to control the amount of code expansion,
72 or the percent of loops to consider for unrolling. Etc. */
74 /* ??? Look at the register copies inside the loop to see if they form a
75 simple permutation. If so, iterate the permutation until it gets back to
76 the start state. This is how many times we should unroll the loop, for
77 best results, because then all register copies can be eliminated.
78 For example, the lisp nreverse function should be unrolled 3 times
87 ??? The number of times to unroll the loop may also be based on data
88 references in the loop. For example, if we have a loop that references
89 x[i-1], x[i], and x[i+1], we should unroll it a multiple of 3 times. */
91 /* ??? Add some simple linear equation solving capability so that we can
92 determine the number of loop iterations for more complex loops.
93 For example, consider this loop from gdb
94 #define SWAP_TARGET_AND_HOST(buffer,len)
97 char *p = (char *) buffer;
98 char *q = ((char *) buffer) + len - 1;
99 int iterations = (len + 1) >> 1;
101 for (p; p < q; p++, q--;)
109 start value = p = &buffer + current_iteration
110 end value = q = &buffer + len - 1 - current_iteration
111 Given the loop exit test of "p < q", then there must be "q - p" iterations,
112 set equal to zero and solve for number of iterations:
113 q - p = len - 1 - 2*current_iteration = 0
114 current_iteration = (len - 1) / 2
115 Hence, there are (len - 1) / 2 (rounded up to the nearest integer)
116 iterations of this loop. */
118 /* ??? Currently, no labels are marked as loop invariant when doing loop
119 unrolling. This is because an insn inside the loop, that loads the address
120 of a label inside the loop into a register, could be moved outside the loop
121 by the invariant code motion pass if labels were invariant. If the loop
122 is subsequently unrolled, the code will be wrong because each unrolled
123 body of the loop will use the same address, whereas each actually needs a
124 different address. A case where this happens is when a loop containing
125 a switch statement is unrolled.
127 It would be better to let labels be considered invariant. When we
128 unroll loops here, check to see if any insns using a label local to the
129 loop were moved before the loop. If so, then correct the problem, by
130 moving the insn back into the loop, or perhaps replicate the insn before
131 the loop, one copy for each time the loop is unrolled. */
133 /* The prime factors looked for when trying to unroll a loop by some
134 number which is modulo the total number of iterations. Just checking
135 for these 4 prime factors will find at least one factor for 75% of
136 all numbers theoretically. Practically speaking, this will succeed
137 almost all of the time since loops are generally a multiple of 2
140 #define NUM_FACTORS 4
142 struct _factor
{ int factor
, count
; } factors
[NUM_FACTORS
]
143 = { {2, 0}, {3, 0}, {5, 0}, {7, 0}};
145 /* Describes the different types of loop unrolling performed. */
147 enum unroll_types
{ UNROLL_COMPLETELY
, UNROLL_MODULO
, UNROLL_NAIVE
};
151 #include "insn-config.h"
152 #include "integrate.h"
159 /* This controls which loops are unrolled, and by how much we unroll
162 #ifndef MAX_UNROLLED_INSNS
163 #define MAX_UNROLLED_INSNS 100
166 /* Indexed by register number, if non-zero, then it contains a pointer
167 to a struct induction for a DEST_REG giv which has been combined with
168 one of more address givs. This is needed because whenever such a DEST_REG
169 giv is modified, we must modify the value of all split address givs
170 that were combined with this DEST_REG giv. */
172 static struct induction
**addr_combined_regs
;
174 /* Indexed by register number, if this is a splittable induction variable,
175 then this will hold the current value of the register, which depends on the
178 static rtx
*splittable_regs
;
180 /* Indexed by register number, if this is a splittable induction variable,
181 then this will hold the number of instructions in the loop that modify
182 the induction variable. Used to ensure that only the last insn modifying
183 a split iv will update the original iv of the dest. */
185 static int *splittable_regs_updates
;
187 /* Values describing the current loop's iteration variable. These are set up
188 by loop_iterations, and used by precondition_loop_p. */
190 static rtx loop_iteration_var
;
191 static rtx loop_initial_value
;
192 static rtx loop_increment
;
193 static rtx loop_final_value
;
195 /* Forward declarations. */
197 static void init_reg_map
PROTO((struct inline_remap
*, int));
198 static int precondition_loop_p
PROTO((rtx
*, rtx
*, rtx
*, rtx
, rtx
));
199 static rtx calculate_giv_inc
PROTO((rtx
, rtx
, int));
200 static rtx initial_reg_note_copy
PROTO((rtx
, struct inline_remap
*));
201 static void final_reg_note_copy
PROTO((rtx
, struct inline_remap
*));
202 static void copy_loop_body
PROTO((rtx
, rtx
, struct inline_remap
*, rtx
, int,
203 enum unroll_types
, rtx
, rtx
, rtx
, rtx
));
204 static void iteration_info
PROTO((rtx
, rtx
*, rtx
*, rtx
, rtx
));
205 static rtx approx_final_value
PROTO((enum rtx_code
, rtx
, int *, int *));
206 static int find_splittable_regs
PROTO((enum unroll_types
, rtx
, rtx
, rtx
, int));
207 static int find_splittable_givs
PROTO((struct iv_class
*,enum unroll_types
,
208 rtx
, rtx
, rtx
, int));
209 static int reg_dead_after_loop
PROTO((rtx
, rtx
, rtx
));
210 static rtx fold_rtx_mult_add
PROTO((rtx
, rtx
, rtx
, enum machine_mode
));
211 static rtx remap_split_bivs
PROTO((rtx
));
213 /* Try to unroll one loop and split induction variables in the loop.
215 The loop is described by the arguments LOOP_END, INSN_COUNT, and
216 LOOP_START. END_INSERT_BEFORE indicates where insns should be added
217 which need to be executed when the loop falls through. STRENGTH_REDUCTION_P
218 indicates whether information generated in the strength reduction pass
221 This function is intended to be called from within `strength_reduce'
225 unroll_loop (loop_end
, insn_count
, loop_start
, end_insert_before
,
230 rtx end_insert_before
;
231 int strength_reduce_p
;
234 int unroll_number
= 1;
235 rtx copy_start
, copy_end
;
236 rtx insn
, copy
, sequence
, pattern
, tem
;
237 int max_labelno
, max_insnno
;
239 struct inline_remap
*map
;
247 int splitting_not_safe
= 0;
248 enum unroll_types unroll_type
;
249 int loop_preconditioned
= 0;
251 /* This points to the last real insn in the loop, which should be either
252 a JUMP_INSN (for conditional jumps) or a BARRIER (for unconditional
256 /* Don't bother unrolling huge loops. Since the minimum factor is
257 two, loops greater than one half of MAX_UNROLLED_INSNS will never
259 if (insn_count
> MAX_UNROLLED_INSNS
/ 2)
261 if (loop_dump_stream
)
262 fprintf (loop_dump_stream
, "Unrolling failure: Loop too big.\n");
266 /* When emitting debugger info, we can't unroll loops with unequal numbers
267 of block_beg and block_end notes, because that would unbalance the block
268 structure of the function. This can happen as a result of the
269 "if (foo) bar; else break;" optimization in jump.c. */
271 if (write_symbols
!= NO_DEBUG
)
273 int block_begins
= 0;
276 for (insn
= loop_start
; insn
!= loop_end
; insn
= NEXT_INSN (insn
))
278 if (GET_CODE (insn
) == NOTE
)
280 if (NOTE_LINE_NUMBER (insn
) == NOTE_INSN_BLOCK_BEG
)
282 else if (NOTE_LINE_NUMBER (insn
) == NOTE_INSN_BLOCK_END
)
287 if (block_begins
!= block_ends
)
289 if (loop_dump_stream
)
290 fprintf (loop_dump_stream
,
291 "Unrolling failure: Unbalanced block notes.\n");
296 /* Determine type of unroll to perform. Depends on the number of iterations
297 and the size of the loop. */
299 /* If there is no strength reduce info, then set loop_n_iterations to zero.
300 This can happen if strength_reduce can't find any bivs in the loop.
301 A value of zero indicates that the number of iterations could not be
304 if (! strength_reduce_p
)
305 loop_n_iterations
= 0;
307 if (loop_dump_stream
&& loop_n_iterations
> 0)
308 fprintf (loop_dump_stream
,
309 "Loop unrolling: %d iterations.\n", loop_n_iterations
);
311 /* Find and save a pointer to the last nonnote insn in the loop. */
313 last_loop_insn
= prev_nonnote_insn (loop_end
);
315 /* Calculate how many times to unroll the loop. Indicate whether or
316 not the loop is being completely unrolled. */
318 if (loop_n_iterations
== 1)
320 /* If number of iterations is exactly 1, then eliminate the compare and
321 branch at the end of the loop since they will never be taken.
322 Then return, since no other action is needed here. */
324 /* If the last instruction is not a BARRIER or a JUMP_INSN, then
325 don't do anything. */
327 if (GET_CODE (last_loop_insn
) == BARRIER
)
329 /* Delete the jump insn. This will delete the barrier also. */
330 delete_insn (PREV_INSN (last_loop_insn
));
332 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
335 /* The immediately preceding insn is a compare which must be
337 delete_insn (last_loop_insn
);
338 delete_insn (PREV_INSN (last_loop_insn
));
340 /* The immediately preceding insn may not be the compare, so don't
342 delete_insn (last_loop_insn
);
347 else if (loop_n_iterations
> 0
348 && loop_n_iterations
* insn_count
< MAX_UNROLLED_INSNS
)
350 unroll_number
= loop_n_iterations
;
351 unroll_type
= UNROLL_COMPLETELY
;
353 else if (loop_n_iterations
> 0)
355 /* Try to factor the number of iterations. Don't bother with the
356 general case, only using 2, 3, 5, and 7 will get 75% of all
357 numbers theoretically, and almost all in practice. */
359 for (i
= 0; i
< NUM_FACTORS
; i
++)
360 factors
[i
].count
= 0;
362 temp
= loop_n_iterations
;
363 for (i
= NUM_FACTORS
- 1; i
>= 0; i
--)
364 while (temp
% factors
[i
].factor
== 0)
367 temp
= temp
/ factors
[i
].factor
;
370 /* Start with the larger factors first so that we generally
371 get lots of unrolling. */
375 for (i
= 3; i
>= 0; i
--)
376 while (factors
[i
].count
--)
378 if (temp
* factors
[i
].factor
< MAX_UNROLLED_INSNS
)
380 unroll_number
*= factors
[i
].factor
;
381 temp
*= factors
[i
].factor
;
387 /* If we couldn't find any factors, then unroll as in the normal
389 if (unroll_number
== 1)
391 if (loop_dump_stream
)
392 fprintf (loop_dump_stream
,
393 "Loop unrolling: No factors found.\n");
396 unroll_type
= UNROLL_MODULO
;
400 /* Default case, calculate number of times to unroll loop based on its
402 if (unroll_number
== 1)
404 if (8 * insn_count
< MAX_UNROLLED_INSNS
)
406 else if (4 * insn_count
< MAX_UNROLLED_INSNS
)
411 unroll_type
= UNROLL_NAIVE
;
414 /* Now we know how many times to unroll the loop. */
416 if (loop_dump_stream
)
417 fprintf (loop_dump_stream
,
418 "Unrolling loop %d times.\n", unroll_number
);
421 if (unroll_type
== UNROLL_COMPLETELY
|| unroll_type
== UNROLL_MODULO
)
423 /* Loops of these types should never start with a jump down to
424 the exit condition test. For now, check for this case just to
425 be sure. UNROLL_NAIVE loops can be of this form, this case is
428 while (GET_CODE (insn
) != CODE_LABEL
&& GET_CODE (insn
) != JUMP_INSN
)
429 insn
= NEXT_INSN (insn
);
430 if (GET_CODE (insn
) == JUMP_INSN
)
434 if (unroll_type
== UNROLL_COMPLETELY
)
436 /* Completely unrolling the loop: Delete the compare and branch at
437 the end (the last two instructions). This delete must done at the
438 very end of loop unrolling, to avoid problems with calls to
439 back_branch_in_range_p, which is called by find_splittable_regs.
440 All increments of splittable bivs/givs are changed to load constant
443 copy_start
= loop_start
;
445 /* Set insert_before to the instruction immediately after the JUMP_INSN
446 (or BARRIER), so that any NOTEs between the JUMP_INSN and the end of
447 the loop will be correctly handled by copy_loop_body. */
448 insert_before
= NEXT_INSN (last_loop_insn
);
450 /* Set copy_end to the insn before the jump at the end of the loop. */
451 if (GET_CODE (last_loop_insn
) == BARRIER
)
452 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
453 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
456 /* The instruction immediately before the JUMP_INSN is a compare
457 instruction which we do not want to copy. */
458 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
460 /* The instruction immediately before the JUMP_INSN may not be the
461 compare, so we must copy it. */
462 copy_end
= PREV_INSN (last_loop_insn
);
467 /* We currently can't unroll a loop if it doesn't end with a
468 JUMP_INSN. There would need to be a mechanism that recognizes
469 this case, and then inserts a jump after each loop body, which
470 jumps to after the last loop body. */
471 if (loop_dump_stream
)
472 fprintf (loop_dump_stream
,
473 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
477 else if (unroll_type
== UNROLL_MODULO
)
479 /* Partially unrolling the loop: The compare and branch at the end
480 (the last two instructions) must remain. Don't copy the compare
481 and branch instructions at the end of the loop. Insert the unrolled
482 code immediately before the compare/branch at the end so that the
483 code will fall through to them as before. */
485 copy_start
= loop_start
;
487 /* Set insert_before to the jump insn at the end of the loop.
488 Set copy_end to before the jump insn at the end of the loop. */
489 if (GET_CODE (last_loop_insn
) == BARRIER
)
491 insert_before
= PREV_INSN (last_loop_insn
);
492 copy_end
= PREV_INSN (insert_before
);
494 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
497 /* The instruction immediately before the JUMP_INSN is a compare
498 instruction which we do not want to copy or delete. */
499 insert_before
= PREV_INSN (last_loop_insn
);
500 copy_end
= PREV_INSN (insert_before
);
502 /* The instruction immediately before the JUMP_INSN may not be the
503 compare, so we must copy it. */
504 insert_before
= last_loop_insn
;
505 copy_end
= PREV_INSN (last_loop_insn
);
510 /* We currently can't unroll a loop if it doesn't end with a
511 JUMP_INSN. There would need to be a mechanism that recognizes
512 this case, and then inserts a jump after each loop body, which
513 jumps to after the last loop body. */
514 if (loop_dump_stream
)
515 fprintf (loop_dump_stream
,
516 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
522 /* Normal case: Must copy the compare and branch instructions at the
525 if (GET_CODE (last_loop_insn
) == BARRIER
)
527 /* Loop ends with an unconditional jump and a barrier.
528 Handle this like above, don't copy jump and barrier.
529 This is not strictly necessary, but doing so prevents generating
530 unconditional jumps to an immediately following label.
532 This will be corrected below if the target of this jump is
533 not the start_label. */
535 insert_before
= PREV_INSN (last_loop_insn
);
536 copy_end
= PREV_INSN (insert_before
);
538 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
540 /* Set insert_before to immediately after the JUMP_INSN, so that
541 NOTEs at the end of the loop will be correctly handled by
543 insert_before
= NEXT_INSN (last_loop_insn
);
544 copy_end
= last_loop_insn
;
548 /* We currently can't unroll a loop if it doesn't end with a
549 JUMP_INSN. There would need to be a mechanism that recognizes
550 this case, and then inserts a jump after each loop body, which
551 jumps to after the last loop body. */
552 if (loop_dump_stream
)
553 fprintf (loop_dump_stream
,
554 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
558 /* If copying exit test branches because they can not be eliminated,
559 then must convert the fall through case of the branch to a jump past
560 the end of the loop. Create a label to emit after the loop and save
561 it for later use. Do not use the label after the loop, if any, since
562 it might be used by insns outside the loop, or there might be insns
563 added before it later by final_[bg]iv_value which must be after
564 the real exit label. */
565 exit_label
= gen_label_rtx ();
568 while (GET_CODE (insn
) != CODE_LABEL
&& GET_CODE (insn
) != JUMP_INSN
)
569 insn
= NEXT_INSN (insn
);
571 if (GET_CODE (insn
) == JUMP_INSN
)
573 /* The loop starts with a jump down to the exit condition test.
574 Start copying the loop after the barrier following this
576 copy_start
= NEXT_INSN (insn
);
578 /* Splitting induction variables doesn't work when the loop is
579 entered via a jump to the bottom, because then we end up doing
580 a comparison against a new register for a split variable, but
581 we did not execute the set insn for the new register because
582 it was skipped over. */
583 splitting_not_safe
= 1;
584 if (loop_dump_stream
)
585 fprintf (loop_dump_stream
,
586 "Splitting not safe, because loop not entered at top.\n");
589 copy_start
= loop_start
;
592 /* This should always be the first label in the loop. */
593 start_label
= NEXT_INSN (copy_start
);
594 /* There may be a line number note and/or a loop continue note here. */
595 while (GET_CODE (start_label
) == NOTE
)
596 start_label
= NEXT_INSN (start_label
);
597 if (GET_CODE (start_label
) != CODE_LABEL
)
599 /* This can happen as a result of jump threading. If the first insns in
600 the loop test the same condition as the loop's backward jump, or the
601 opposite condition, then the backward jump will be modified to point
602 to elsewhere, and the loop's start label is deleted.
604 This case currently can not be handled by the loop unrolling code. */
606 if (loop_dump_stream
)
607 fprintf (loop_dump_stream
,
608 "Unrolling failure: unknown insns between BEG note and loop label.\n");
611 if (LABEL_NAME (start_label
))
613 /* The jump optimization pass must have combined the original start label
614 with a named label for a goto. We can't unroll this case because
615 jumps which go to the named label must be handled differently than
616 jumps to the loop start, and it is impossible to differentiate them
618 if (loop_dump_stream
)
619 fprintf (loop_dump_stream
,
620 "Unrolling failure: loop start label is gone\n");
624 if (unroll_type
== UNROLL_NAIVE
625 && GET_CODE (last_loop_insn
) == BARRIER
626 && start_label
!= JUMP_LABEL (PREV_INSN (last_loop_insn
)))
628 /* In this case, we must copy the jump and barrier, because they will
629 not be converted to jumps to an immediately following label. */
631 insert_before
= NEXT_INSN (last_loop_insn
);
632 copy_end
= last_loop_insn
;
635 /* Allocate a translation table for the labels and insn numbers.
636 They will be filled in as we copy the insns in the loop. */
638 max_labelno
= max_label_num ();
639 max_insnno
= get_max_uid ();
641 map
= (struct inline_remap
*) alloca (sizeof (struct inline_remap
));
643 map
->integrating
= 0;
645 /* Allocate the label map. */
649 map
->label_map
= (rtx
*) alloca (max_labelno
* sizeof (rtx
));
651 local_label
= (char *) alloca (max_labelno
);
652 bzero (local_label
, max_labelno
);
657 /* Search the loop and mark all local labels, i.e. the ones which have to
658 be distinct labels when copied. For all labels which might be
659 non-local, set their label_map entries to point to themselves.
660 If they happen to be local their label_map entries will be overwritten
661 before the loop body is copied. The label_map entries for local labels
662 will be set to a different value each time the loop body is copied. */
664 for (insn
= copy_start
; insn
!= loop_end
; insn
= NEXT_INSN (insn
))
666 if (GET_CODE (insn
) == CODE_LABEL
)
667 local_label
[CODE_LABEL_NUMBER (insn
)] = 1;
668 else if (GET_CODE (insn
) == JUMP_INSN
)
670 if (JUMP_LABEL (insn
))
671 map
->label_map
[CODE_LABEL_NUMBER (JUMP_LABEL (insn
))]
673 else if (GET_CODE (PATTERN (insn
)) == ADDR_VEC
674 || GET_CODE (PATTERN (insn
)) == ADDR_DIFF_VEC
)
676 rtx pat
= PATTERN (insn
);
677 int diff_vec_p
= GET_CODE (PATTERN (insn
)) == ADDR_DIFF_VEC
;
678 int len
= XVECLEN (pat
, diff_vec_p
);
681 for (i
= 0; i
< len
; i
++)
683 label
= XEXP (XVECEXP (pat
, diff_vec_p
, i
), 0);
684 map
->label_map
[CODE_LABEL_NUMBER (label
)] = label
;
690 /* Allocate space for the insn map. */
692 map
->insn_map
= (rtx
*) alloca (max_insnno
* sizeof (rtx
));
694 /* Set this to zero, to indicate that we are doing loop unrolling,
695 not function inlining. */
696 map
->inline_target
= 0;
698 /* The register and constant maps depend on the number of registers
699 present, so the final maps can't be created until after
700 find_splittable_regs is called. However, they are needed for
701 preconditioning, so we create temporary maps when preconditioning
704 /* The preconditioning code may allocate two new pseudo registers. */
705 maxregnum
= max_reg_num ();
707 /* Allocate and zero out the splittable_regs and addr_combined_regs
708 arrays. These must be zeroed here because they will be used if
709 loop preconditioning is performed, and must be zero for that case.
711 It is safe to do this here, since the extra registers created by the
712 preconditioning code and find_splittable_regs will never be used
713 to access the splittable_regs[] and addr_combined_regs[] arrays. */
715 splittable_regs
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
716 bzero ((char *) splittable_regs
, maxregnum
* sizeof (rtx
));
717 splittable_regs_updates
= (int *) alloca (maxregnum
* sizeof (int));
718 bzero ((char *) splittable_regs_updates
, maxregnum
* sizeof (int));
720 = (struct induction
**) alloca (maxregnum
* sizeof (struct induction
*));
721 bzero ((char *) addr_combined_regs
, maxregnum
* sizeof (struct induction
*));
722 /* We must limit it to max_reg_before_loop, because only these pseudo
723 registers have valid regno_first_uid info. Any register created after
724 that is unlikely to be local to the loop anyways. */
725 local_regno
= (char *) alloca (max_reg_before_loop
);
726 bzero (local_regno
, max_reg_before_loop
);
728 /* Mark all local registers, i.e. the ones which are referenced only
730 if (INSN_UID (copy_end
) < max_uid_for_loop
)
732 int copy_start_luid
= INSN_LUID (copy_start
);
733 int copy_end_luid
= INSN_LUID (copy_end
);
735 /* If a register is used in the jump insn, we must not duplicate it
736 since it will also be used outside the loop. */
737 if (GET_CODE (copy_end
) == JUMP_INSN
)
739 /* If copy_start points to the NOTE that starts the loop, then we must
740 use the next luid, because invariant pseudo-regs moved out of the loop
741 have their lifetimes modified to start here, but they are not safe
743 if (copy_start
== loop_start
)
746 for (j
= FIRST_PSEUDO_REGISTER
; j
< max_reg_before_loop
; ++j
)
747 if (regno_first_uid
[j
] > 0 && regno_first_uid
[j
] <= max_uid_for_loop
748 && uid_luid
[regno_first_uid
[j
]] >= copy_start_luid
749 && regno_last_uid
[j
] > 0 && regno_last_uid
[j
] <= max_uid_for_loop
750 && uid_luid
[regno_last_uid
[j
]] <= copy_end_luid
)
754 /* If this loop requires exit tests when unrolled, check to see if we
755 can precondition the loop so as to make the exit tests unnecessary.
756 Just like variable splitting, this is not safe if the loop is entered
757 via a jump to the bottom. Also, can not do this if no strength
758 reduce info, because precondition_loop_p uses this info. */
760 /* Must copy the loop body for preconditioning before the following
761 find_splittable_regs call since that will emit insns which need to
762 be after the preconditioned loop copies, but immediately before the
763 unrolled loop copies. */
765 /* Also, it is not safe to split induction variables for the preconditioned
766 copies of the loop body. If we split induction variables, then the code
767 assumes that each induction variable can be represented as a function
768 of its initial value and the loop iteration number. This is not true
769 in this case, because the last preconditioned copy of the loop body
770 could be any iteration from the first up to the `unroll_number-1'th,
771 depending on the initial value of the iteration variable. Therefore
772 we can not split induction variables here, because we can not calculate
773 their value. Hence, this code must occur before find_splittable_regs
776 if (unroll_type
== UNROLL_NAIVE
&& ! splitting_not_safe
&& strength_reduce_p
)
778 rtx initial_value
, final_value
, increment
;
780 if (precondition_loop_p (&initial_value
, &final_value
, &increment
,
781 loop_start
, loop_end
))
783 register rtx diff
, temp
;
784 enum machine_mode mode
;
786 int abs_inc
, neg_inc
;
788 map
->reg_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
790 map
->const_equiv_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
791 map
->const_age_map
= (unsigned *) alloca (maxregnum
792 * sizeof (unsigned));
793 map
->const_equiv_map_size
= maxregnum
;
794 global_const_equiv_map
= map
->const_equiv_map
;
795 global_const_equiv_map_size
= maxregnum
;
797 init_reg_map (map
, maxregnum
);
799 /* Limit loop unrolling to 4, since this will make 7 copies of
801 if (unroll_number
> 4)
804 /* Save the absolute value of the increment, and also whether or
805 not it is negative. */
807 abs_inc
= INTVAL (increment
);
816 /* Decide what mode to do these calculations in. Choose the larger
817 of final_value's mode and initial_value's mode, or a full-word if
818 both are constants. */
819 mode
= GET_MODE (final_value
);
820 if (mode
== VOIDmode
)
822 mode
= GET_MODE (initial_value
);
823 if (mode
== VOIDmode
)
826 else if (mode
!= GET_MODE (initial_value
)
827 && (GET_MODE_SIZE (mode
)
828 < GET_MODE_SIZE (GET_MODE (initial_value
))))
829 mode
= GET_MODE (initial_value
);
831 /* Calculate the difference between the final and initial values.
832 Final value may be a (plus (reg x) (const_int 1)) rtx.
833 Let the following cse pass simplify this if initial value is
836 We must copy the final and initial values here to avoid
837 improperly shared rtl. */
839 diff
= expand_binop (mode
, sub_optab
, copy_rtx (final_value
),
840 copy_rtx (initial_value
), NULL_RTX
, 0,
843 /* Now calculate (diff % (unroll * abs (increment))) by using an
845 diff
= expand_binop (GET_MODE (diff
), and_optab
, diff
,
846 GEN_INT (unroll_number
* abs_inc
- 1),
847 NULL_RTX
, 0, OPTAB_LIB_WIDEN
);
849 /* Now emit a sequence of branches to jump to the proper precond
852 labels
= (rtx
*) alloca (sizeof (rtx
) * unroll_number
);
853 for (i
= 0; i
< unroll_number
; i
++)
854 labels
[i
] = gen_label_rtx ();
856 /* Assuming the unroll_number is 4, and the increment is 2, then
857 for a negative increment: for a positive increment:
858 diff = 0,1 precond 0 diff = 0,7 precond 0
859 diff = 2,3 precond 3 diff = 1,2 precond 1
860 diff = 4,5 precond 2 diff = 3,4 precond 2
861 diff = 6,7 precond 1 diff = 5,6 precond 3 */
863 /* We only need to emit (unroll_number - 1) branches here, the
864 last case just falls through to the following code. */
866 /* ??? This would give better code if we emitted a tree of branches
867 instead of the current linear list of branches. */
869 for (i
= 0; i
< unroll_number
- 1; i
++)
873 /* For negative increments, must invert the constant compared
874 against, except when comparing against zero. */
878 cmp_const
= unroll_number
- i
;
882 emit_cmp_insn (diff
, GEN_INT (abs_inc
* cmp_const
),
883 EQ
, NULL_RTX
, mode
, 0, 0);
886 emit_jump_insn (gen_beq (labels
[i
]));
888 emit_jump_insn (gen_bge (labels
[i
]));
890 emit_jump_insn (gen_ble (labels
[i
]));
891 JUMP_LABEL (get_last_insn ()) = labels
[i
];
892 LABEL_NUSES (labels
[i
])++;
895 /* If the increment is greater than one, then we need another branch,
896 to handle other cases equivalent to 0. */
898 /* ??? This should be merged into the code above somehow to help
899 simplify the code here, and reduce the number of branches emitted.
900 For the negative increment case, the branch here could easily
901 be merged with the `0' case branch above. For the positive
902 increment case, it is not clear how this can be simplified. */
909 cmp_const
= abs_inc
- 1;
911 cmp_const
= abs_inc
* (unroll_number
- 1) + 1;
913 emit_cmp_insn (diff
, GEN_INT (cmp_const
), EQ
, NULL_RTX
,
917 emit_jump_insn (gen_ble (labels
[0]));
919 emit_jump_insn (gen_bge (labels
[0]));
920 JUMP_LABEL (get_last_insn ()) = labels
[0];
921 LABEL_NUSES (labels
[0])++;
924 sequence
= gen_sequence ();
926 emit_insn_before (sequence
, loop_start
);
928 /* Only the last copy of the loop body here needs the exit
929 test, so set copy_end to exclude the compare/branch here,
930 and then reset it inside the loop when get to the last
933 if (GET_CODE (last_loop_insn
) == BARRIER
)
934 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
935 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
938 /* The immediately preceding insn is a compare which we do not
940 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
942 /* The immediately preceding insn may not be a compare, so we
944 copy_end
= PREV_INSN (last_loop_insn
);
950 for (i
= 1; i
< unroll_number
; i
++)
952 emit_label_after (labels
[unroll_number
- i
],
953 PREV_INSN (loop_start
));
955 bzero ((char *) map
->insn_map
, max_insnno
* sizeof (rtx
));
956 bzero ((char *) map
->const_equiv_map
, maxregnum
* sizeof (rtx
));
957 bzero ((char *) map
->const_age_map
,
958 maxregnum
* sizeof (unsigned));
961 for (j
= 0; j
< max_labelno
; j
++)
963 map
->label_map
[j
] = gen_label_rtx ();
965 for (j
= FIRST_PSEUDO_REGISTER
; j
< max_reg_before_loop
; j
++)
967 map
->reg_map
[j
] = gen_reg_rtx (GET_MODE (regno_reg_rtx
[j
]));
969 /* The last copy needs the compare/branch insns at the end,
970 so reset copy_end here if the loop ends with a conditional
973 if (i
== unroll_number
- 1)
975 if (GET_CODE (last_loop_insn
) == BARRIER
)
976 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
978 copy_end
= last_loop_insn
;
981 /* None of the copies are the `last_iteration', so just
982 pass zero for that parameter. */
983 copy_loop_body (copy_start
, copy_end
, map
, exit_label
, 0,
984 unroll_type
, start_label
, loop_end
,
985 loop_start
, copy_end
);
987 emit_label_after (labels
[0], PREV_INSN (loop_start
));
989 if (GET_CODE (last_loop_insn
) == BARRIER
)
991 insert_before
= PREV_INSN (last_loop_insn
);
992 copy_end
= PREV_INSN (insert_before
);
997 /* The immediately preceding insn is a compare which we do not
999 insert_before
= PREV_INSN (last_loop_insn
);
1000 copy_end
= PREV_INSN (insert_before
);
1002 /* The immediately preceding insn may not be a compare, so we
1004 insert_before
= last_loop_insn
;
1005 copy_end
= PREV_INSN (last_loop_insn
);
1009 /* Set unroll type to MODULO now. */
1010 unroll_type
= UNROLL_MODULO
;
1011 loop_preconditioned
= 1;
1015 /* If reach here, and the loop type is UNROLL_NAIVE, then don't unroll
1016 the loop unless all loops are being unrolled. */
1017 if (unroll_type
== UNROLL_NAIVE
&& ! flag_unroll_all_loops
)
1019 if (loop_dump_stream
)
1020 fprintf (loop_dump_stream
, "Unrolling failure: Naive unrolling not being done.\n");
1024 /* At this point, we are guaranteed to unroll the loop. */
1026 /* For each biv and giv, determine whether it can be safely split into
1027 a different variable for each unrolled copy of the loop body.
1028 We precalculate and save this info here, since computing it is
1031 Do this before deleting any instructions from the loop, so that
1032 back_branch_in_range_p will work correctly. */
1034 if (splitting_not_safe
)
1037 temp
= find_splittable_regs (unroll_type
, loop_start
, loop_end
,
1038 end_insert_before
, unroll_number
);
1040 /* find_splittable_regs may have created some new registers, so must
1041 reallocate the reg_map with the new larger size, and must realloc
1042 the constant maps also. */
1044 maxregnum
= max_reg_num ();
1045 map
->reg_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
1047 init_reg_map (map
, maxregnum
);
1049 /* Space is needed in some of the map for new registers, so new_maxregnum
1050 is an (over)estimate of how many registers will exist at the end. */
1051 new_maxregnum
= maxregnum
+ (temp
* unroll_number
* 2);
1053 /* Must realloc space for the constant maps, because the number of registers
1054 may have changed. */
1056 map
->const_equiv_map
= (rtx
*) alloca (new_maxregnum
* sizeof (rtx
));
1057 map
->const_age_map
= (unsigned *) alloca (new_maxregnum
* sizeof (unsigned));
1059 map
->const_equiv_map_size
= new_maxregnum
;
1060 global_const_equiv_map
= map
->const_equiv_map
;
1061 global_const_equiv_map_size
= new_maxregnum
;
1063 /* Search the list of bivs and givs to find ones which need to be remapped
1064 when split, and set their reg_map entry appropriately. */
1066 for (bl
= loop_iv_list
; bl
; bl
= bl
->next
)
1068 if (REGNO (bl
->biv
->src_reg
) != bl
->regno
)
1069 map
->reg_map
[bl
->regno
] = bl
->biv
->src_reg
;
1071 /* Currently, non-reduced/final-value givs are never split. */
1072 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
1073 if (REGNO (v
->src_reg
) != bl
->regno
)
1074 map
->reg_map
[REGNO (v
->dest_reg
)] = v
->src_reg
;
1078 /* If the loop is being partially unrolled, and the iteration variables
1079 are being split, and are being renamed for the split, then must fix up
1080 the compare/jump instruction at the end of the loop to refer to the new
1081 registers. This compare isn't copied, so the registers used in it
1082 will never be replaced if it isn't done here. */
1084 if (unroll_type
== UNROLL_MODULO
)
1086 insn
= NEXT_INSN (copy_end
);
1087 if (GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == JUMP_INSN
)
1088 PATTERN (insn
) = remap_split_bivs (PATTERN (insn
));
1091 /* For unroll_number - 1 times, make a copy of each instruction
1092 between copy_start and copy_end, and insert these new instructions
1093 before the end of the loop. */
1095 for (i
= 0; i
< unroll_number
; i
++)
1097 bzero ((char *) map
->insn_map
, max_insnno
* sizeof (rtx
));
1098 bzero ((char *) map
->const_equiv_map
, new_maxregnum
* sizeof (rtx
));
1099 bzero ((char *) map
->const_age_map
, new_maxregnum
* sizeof (unsigned));
1102 for (j
= 0; j
< max_labelno
; j
++)
1104 map
->label_map
[j
] = gen_label_rtx ();
1106 for (j
= FIRST_PSEUDO_REGISTER
; j
< max_reg_before_loop
; j
++)
1108 map
->reg_map
[j
] = gen_reg_rtx (GET_MODE (regno_reg_rtx
[j
]));
1110 /* If loop starts with a branch to the test, then fix it so that
1111 it points to the test of the first unrolled copy of the loop. */
1112 if (i
== 0 && loop_start
!= copy_start
)
1114 insn
= PREV_INSN (copy_start
);
1115 pattern
= PATTERN (insn
);
1117 tem
= map
->label_map
[CODE_LABEL_NUMBER
1118 (XEXP (SET_SRC (pattern
), 0))];
1119 SET_SRC (pattern
) = gen_rtx (LABEL_REF
, VOIDmode
, tem
);
1121 /* Set the jump label so that it can be used by later loop unrolling
1123 JUMP_LABEL (insn
) = tem
;
1124 LABEL_NUSES (tem
)++;
1127 copy_loop_body (copy_start
, copy_end
, map
, exit_label
,
1128 i
== unroll_number
- 1, unroll_type
, start_label
,
1129 loop_end
, insert_before
, insert_before
);
1132 /* Before deleting any insns, emit a CODE_LABEL immediately after the last
1133 insn to be deleted. This prevents any runaway delete_insn call from
1134 more insns that it should, as it always stops at a CODE_LABEL. */
1136 /* Delete the compare and branch at the end of the loop if completely
1137 unrolling the loop. Deleting the backward branch at the end also
1138 deletes the code label at the start of the loop. This is done at
1139 the very end to avoid problems with back_branch_in_range_p. */
1141 if (unroll_type
== UNROLL_COMPLETELY
)
1142 safety_label
= emit_label_after (gen_label_rtx (), last_loop_insn
);
1144 safety_label
= emit_label_after (gen_label_rtx (), copy_end
);
1146 /* Delete all of the original loop instructions. Don't delete the
1147 LOOP_BEG note, or the first code label in the loop. */
1149 insn
= NEXT_INSN (copy_start
);
1150 while (insn
!= safety_label
)
1152 if (insn
!= start_label
)
1153 insn
= delete_insn (insn
);
1155 insn
= NEXT_INSN (insn
);
1158 /* Can now delete the 'safety' label emitted to protect us from runaway
1159 delete_insn calls. */
1160 if (INSN_DELETED_P (safety_label
))
1162 delete_insn (safety_label
);
1164 /* If exit_label exists, emit it after the loop. Doing the emit here
1165 forces it to have a higher INSN_UID than any insn in the unrolled loop.
1166 This is needed so that mostly_true_jump in reorg.c will treat jumps
1167 to this loop end label correctly, i.e. predict that they are usually
1170 emit_label_after (exit_label
, loop_end
);
1173 /* Return true if the loop can be safely, and profitably, preconditioned
1174 so that the unrolled copies of the loop body don't need exit tests.
1176 This only works if final_value, initial_value and increment can be
1177 determined, and if increment is a constant power of 2.
1178 If increment is not a power of 2, then the preconditioning modulo
1179 operation would require a real modulo instead of a boolean AND, and this
1180 is not considered `profitable'. */
1182 /* ??? If the loop is known to be executed very many times, or the machine
1183 has a very cheap divide instruction, then preconditioning is a win even
1184 when the increment is not a power of 2. Use RTX_COST to compute
1185 whether divide is cheap. */
1188 precondition_loop_p (initial_value
, final_value
, increment
, loop_start
,
1190 rtx
*initial_value
, *final_value
, *increment
;
1191 rtx loop_start
, loop_end
;
1194 if (loop_n_iterations
> 0)
1196 *initial_value
= const0_rtx
;
1197 *increment
= const1_rtx
;
1198 *final_value
= GEN_INT (loop_n_iterations
);
1200 if (loop_dump_stream
)
1201 fprintf (loop_dump_stream
,
1202 "Preconditioning: Success, number of iterations known, %d.\n",
1207 if (loop_initial_value
== 0)
1209 if (loop_dump_stream
)
1210 fprintf (loop_dump_stream
,
1211 "Preconditioning: Could not find initial value.\n");
1214 else if (loop_increment
== 0)
1216 if (loop_dump_stream
)
1217 fprintf (loop_dump_stream
,
1218 "Preconditioning: Could not find increment value.\n");
1221 else if (GET_CODE (loop_increment
) != CONST_INT
)
1223 if (loop_dump_stream
)
1224 fprintf (loop_dump_stream
,
1225 "Preconditioning: Increment not a constant.\n");
1228 else if ((exact_log2 (INTVAL (loop_increment
)) < 0)
1229 && (exact_log2 (- INTVAL (loop_increment
)) < 0))
1231 if (loop_dump_stream
)
1232 fprintf (loop_dump_stream
,
1233 "Preconditioning: Increment not a constant power of 2.\n");
1237 /* Unsigned_compare and compare_dir can be ignored here, since they do
1238 not matter for preconditioning. */
1240 if (loop_final_value
== 0)
1242 if (loop_dump_stream
)
1243 fprintf (loop_dump_stream
,
1244 "Preconditioning: EQ comparison loop.\n");
1248 /* Must ensure that final_value is invariant, so call invariant_p to
1249 check. Before doing so, must check regno against max_reg_before_loop
1250 to make sure that the register is in the range covered by invariant_p.
1251 If it isn't, then it is most likely a biv/giv which by definition are
1253 if ((GET_CODE (loop_final_value
) == REG
1254 && REGNO (loop_final_value
) >= max_reg_before_loop
)
1255 || (GET_CODE (loop_final_value
) == PLUS
1256 && REGNO (XEXP (loop_final_value
, 0)) >= max_reg_before_loop
)
1257 || ! invariant_p (loop_final_value
))
1259 if (loop_dump_stream
)
1260 fprintf (loop_dump_stream
,
1261 "Preconditioning: Final value not invariant.\n");
1265 /* Fail for floating point values, since the caller of this function
1266 does not have code to deal with them. */
1267 if (GET_MODE_CLASS (GET_MODE (loop_final_value
)) == MODE_FLOAT
1268 || GET_MODE_CLASS (GET_MODE (loop_initial_value
)) == MODE_FLOAT
)
1270 if (loop_dump_stream
)
1271 fprintf (loop_dump_stream
,
1272 "Preconditioning: Floating point final or initial value.\n");
1276 /* Now set initial_value to be the iteration_var, since that may be a
1277 simpler expression, and is guaranteed to be correct if all of the
1278 above tests succeed.
1280 We can not use the initial_value as calculated, because it will be
1281 one too small for loops of the form "while (i-- > 0)". We can not
1282 emit code before the loop_skip_over insns to fix this problem as this
1283 will then give a number one too large for loops of the form
1286 Note that all loops that reach here are entered at the top, because
1287 this function is not called if the loop starts with a jump. */
1289 /* Fail if loop_iteration_var is not live before loop_start, since we need
1290 to test its value in the preconditioning code. */
1292 if (uid_luid
[regno_first_uid
[REGNO (loop_iteration_var
)]]
1293 > INSN_LUID (loop_start
))
1295 if (loop_dump_stream
)
1296 fprintf (loop_dump_stream
,
1297 "Preconditioning: Iteration var not live before loop start.\n");
1301 *initial_value
= loop_iteration_var
;
1302 *increment
= loop_increment
;
1303 *final_value
= loop_final_value
;
1306 if (loop_dump_stream
)
1307 fprintf (loop_dump_stream
, "Preconditioning: Successful.\n");
1312 /* All pseudo-registers must be mapped to themselves. Two hard registers
1313 must be mapped, VIRTUAL_STACK_VARS_REGNUM and VIRTUAL_INCOMING_ARGS_
1314 REGNUM, to avoid function-inlining specific conversions of these
1315 registers. All other hard regs can not be mapped because they may be
1320 init_reg_map (map
, maxregnum
)
1321 struct inline_remap
*map
;
1326 for (i
= maxregnum
- 1; i
> LAST_VIRTUAL_REGISTER
; i
--)
1327 map
->reg_map
[i
] = regno_reg_rtx
[i
];
1328 /* Just clear the rest of the entries. */
1329 for (i
= LAST_VIRTUAL_REGISTER
; i
>= 0; i
--)
1330 map
->reg_map
[i
] = 0;
1332 map
->reg_map
[VIRTUAL_STACK_VARS_REGNUM
]
1333 = regno_reg_rtx
[VIRTUAL_STACK_VARS_REGNUM
];
1334 map
->reg_map
[VIRTUAL_INCOMING_ARGS_REGNUM
]
1335 = regno_reg_rtx
[VIRTUAL_INCOMING_ARGS_REGNUM
];
1338 /* Strength-reduction will often emit code for optimized biv/givs which
1339 calculates their value in a temporary register, and then copies the result
1340 to the iv. This procedure reconstructs the pattern computing the iv;
1341 verifying that all operands are of the proper form.
1343 The return value is the amount that the giv is incremented by. */
1346 calculate_giv_inc (pattern
, src_insn
, regno
)
1347 rtx pattern
, src_insn
;
1351 rtx increment_total
= 0;
1355 /* Verify that we have an increment insn here. First check for a plus
1356 as the set source. */
1357 if (GET_CODE (SET_SRC (pattern
)) != PLUS
)
1359 /* SR sometimes computes the new giv value in a temp, then copies it
1361 src_insn
= PREV_INSN (src_insn
);
1362 pattern
= PATTERN (src_insn
);
1363 if (GET_CODE (SET_SRC (pattern
)) != PLUS
)
1366 /* The last insn emitted is not needed, so delete it to avoid confusing
1367 the second cse pass. This insn sets the giv unnecessarily. */
1368 delete_insn (get_last_insn ());
1371 /* Verify that we have a constant as the second operand of the plus. */
1372 increment
= XEXP (SET_SRC (pattern
), 1);
1373 if (GET_CODE (increment
) != CONST_INT
)
1375 /* SR sometimes puts the constant in a register, especially if it is
1376 too big to be an add immed operand. */
1377 src_insn
= PREV_INSN (src_insn
);
1378 increment
= SET_SRC (PATTERN (src_insn
));
1380 /* SR may have used LO_SUM to compute the constant if it is too large
1381 for a load immed operand. In this case, the constant is in operand
1382 one of the LO_SUM rtx. */
1383 if (GET_CODE (increment
) == LO_SUM
)
1384 increment
= XEXP (increment
, 1);
1385 else if (GET_CODE (increment
) == IOR
)
1387 /* The rs6000 port loads some constants with IOR. */
1388 rtx second_part
= XEXP (increment
, 1);
1390 src_insn
= PREV_INSN (src_insn
);
1391 increment
= SET_SRC (PATTERN (src_insn
));
1392 /* Don't need the last insn anymore. */
1393 delete_insn (get_last_insn ());
1395 if (GET_CODE (second_part
) != CONST_INT
1396 || GET_CODE (increment
) != CONST_INT
)
1399 increment
= GEN_INT (INTVAL (increment
) | INTVAL (second_part
));
1402 if (GET_CODE (increment
) != CONST_INT
)
1405 /* The insn loading the constant into a register is no longer needed,
1407 delete_insn (get_last_insn ());
1410 if (increment_total
)
1411 increment_total
= GEN_INT (INTVAL (increment_total
) + INTVAL (increment
));
1413 increment_total
= increment
;
1415 /* Check that the source register is the same as the register we expected
1416 to see as the source. If not, something is seriously wrong. */
1417 if (GET_CODE (XEXP (SET_SRC (pattern
), 0)) != REG
1418 || REGNO (XEXP (SET_SRC (pattern
), 0)) != regno
)
1420 /* Some machines (e.g. the romp), may emit two add instructions for
1421 certain constants, so lets try looking for another add immediately
1422 before this one if we have only seen one add insn so far. */
1428 src_insn
= PREV_INSN (src_insn
);
1429 pattern
= PATTERN (src_insn
);
1431 delete_insn (get_last_insn ());
1439 return increment_total
;
1442 /* Copy REG_NOTES, except for insn references, because not all insn_map
1443 entries are valid yet. We do need to copy registers now though, because
1444 the reg_map entries can change during copying. */
1447 initial_reg_note_copy (notes
, map
)
1449 struct inline_remap
*map
;
1456 copy
= rtx_alloc (GET_CODE (notes
));
1457 PUT_MODE (copy
, GET_MODE (notes
));
1459 if (GET_CODE (notes
) == EXPR_LIST
)
1460 XEXP (copy
, 0) = copy_rtx_and_substitute (XEXP (notes
, 0), map
);
1461 else if (GET_CODE (notes
) == INSN_LIST
)
1462 /* Don't substitute for these yet. */
1463 XEXP (copy
, 0) = XEXP (notes
, 0);
1467 XEXP (copy
, 1) = initial_reg_note_copy (XEXP (notes
, 1), map
);
1472 /* Fixup insn references in copied REG_NOTES. */
1475 final_reg_note_copy (notes
, map
)
1477 struct inline_remap
*map
;
1481 for (note
= notes
; note
; note
= XEXP (note
, 1))
1482 if (GET_CODE (note
) == INSN_LIST
)
1483 XEXP (note
, 0) = map
->insn_map
[INSN_UID (XEXP (note
, 0))];
1486 /* Copy each instruction in the loop, substituting from map as appropriate.
1487 This is very similar to a loop in expand_inline_function. */
1490 copy_loop_body (copy_start
, copy_end
, map
, exit_label
, last_iteration
,
1491 unroll_type
, start_label
, loop_end
, insert_before
,
1493 rtx copy_start
, copy_end
;
1494 struct inline_remap
*map
;
1497 enum unroll_types unroll_type
;
1498 rtx start_label
, loop_end
, insert_before
, copy_notes_from
;
1502 int dest_reg_was_split
, i
;
1504 rtx final_label
= 0;
1505 rtx giv_inc
, giv_dest_reg
, giv_src_reg
;
1507 /* If this isn't the last iteration, then map any references to the
1508 start_label to final_label. Final label will then be emitted immediately
1509 after the end of this loop body if it was ever used.
1511 If this is the last iteration, then map references to the start_label
1513 if (! last_iteration
)
1515 final_label
= gen_label_rtx ();
1516 map
->label_map
[CODE_LABEL_NUMBER (start_label
)] = final_label
;
1519 map
->label_map
[CODE_LABEL_NUMBER (start_label
)] = start_label
;
1526 insn
= NEXT_INSN (insn
);
1528 map
->orig_asm_operands_vector
= 0;
1530 switch (GET_CODE (insn
))
1533 pattern
= PATTERN (insn
);
1537 /* Check to see if this is a giv that has been combined with
1538 some split address givs. (Combined in the sense that
1539 `combine_givs' in loop.c has put two givs in the same register.)
1540 In this case, we must search all givs based on the same biv to
1541 find the address givs. Then split the address givs.
1542 Do this before splitting the giv, since that may map the
1543 SET_DEST to a new register. */
1545 if (GET_CODE (pattern
) == SET
1546 && GET_CODE (SET_DEST (pattern
)) == REG
1547 && addr_combined_regs
[REGNO (SET_DEST (pattern
))])
1549 struct iv_class
*bl
;
1550 struct induction
*v
, *tv
;
1551 int regno
= REGNO (SET_DEST (pattern
));
1553 v
= addr_combined_regs
[REGNO (SET_DEST (pattern
))];
1554 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
1556 /* Although the giv_inc amount is not needed here, we must call
1557 calculate_giv_inc here since it might try to delete the
1558 last insn emitted. If we wait until later to call it,
1559 we might accidentally delete insns generated immediately
1560 below by emit_unrolled_add. */
1562 giv_inc
= calculate_giv_inc (pattern
, insn
, regno
);
1564 /* Now find all address giv's that were combined with this
1566 for (tv
= bl
->giv
; tv
; tv
= tv
->next_iv
)
1567 if (tv
->giv_type
== DEST_ADDR
&& tv
->same
== v
)
1569 int this_giv_inc
= INTVAL (giv_inc
);
1571 /* Scale this_giv_inc if the multiplicative factors of
1572 the two givs are different. */
1573 if (tv
->mult_val
!= v
->mult_val
)
1574 this_giv_inc
= (this_giv_inc
/ INTVAL (v
->mult_val
)
1575 * INTVAL (tv
->mult_val
));
1577 tv
->dest_reg
= plus_constant (tv
->dest_reg
, this_giv_inc
);
1578 *tv
->location
= tv
->dest_reg
;
1580 if (last_iteration
&& unroll_type
!= UNROLL_COMPLETELY
)
1582 /* Must emit an insn to increment the split address
1583 giv. Add in the const_adjust field in case there
1584 was a constant eliminated from the address. */
1585 rtx value
, dest_reg
;
1587 /* tv->dest_reg will be either a bare register,
1588 or else a register plus a constant. */
1589 if (GET_CODE (tv
->dest_reg
) == REG
)
1590 dest_reg
= tv
->dest_reg
;
1592 dest_reg
= XEXP (tv
->dest_reg
, 0);
1594 /* Check for shared address givs, and avoid
1595 incrementing the shared pseudo reg more than
1597 if (! tv
->same_insn
)
1599 /* tv->dest_reg may actually be a (PLUS (REG)
1600 (CONST)) here, so we must call plus_constant
1601 to add the const_adjust amount before calling
1602 emit_unrolled_add below. */
1603 value
= plus_constant (tv
->dest_reg
,
1606 /* The constant could be too large for an add
1607 immediate, so can't directly emit an insn
1609 emit_unrolled_add (dest_reg
, XEXP (value
, 0),
1613 /* Reset the giv to be just the register again, in case
1614 it is used after the set we have just emitted.
1615 We must subtract the const_adjust factor added in
1617 tv
->dest_reg
= plus_constant (dest_reg
,
1618 - tv
->const_adjust
);
1619 *tv
->location
= tv
->dest_reg
;
1624 /* If this is a setting of a splittable variable, then determine
1625 how to split the variable, create a new set based on this split,
1626 and set up the reg_map so that later uses of the variable will
1627 use the new split variable. */
1629 dest_reg_was_split
= 0;
1631 if (GET_CODE (pattern
) == SET
1632 && GET_CODE (SET_DEST (pattern
)) == REG
1633 && splittable_regs
[REGNO (SET_DEST (pattern
))])
1635 int regno
= REGNO (SET_DEST (pattern
));
1637 dest_reg_was_split
= 1;
1639 /* Compute the increment value for the giv, if it wasn't
1640 already computed above. */
1643 giv_inc
= calculate_giv_inc (pattern
, insn
, regno
);
1644 giv_dest_reg
= SET_DEST (pattern
);
1645 giv_src_reg
= SET_DEST (pattern
);
1647 if (unroll_type
== UNROLL_COMPLETELY
)
1649 /* Completely unrolling the loop. Set the induction
1650 variable to a known constant value. */
1652 /* The value in splittable_regs may be an invariant
1653 value, so we must use plus_constant here. */
1654 splittable_regs
[regno
]
1655 = plus_constant (splittable_regs
[regno
], INTVAL (giv_inc
));
1657 if (GET_CODE (splittable_regs
[regno
]) == PLUS
)
1659 giv_src_reg
= XEXP (splittable_regs
[regno
], 0);
1660 giv_inc
= XEXP (splittable_regs
[regno
], 1);
1664 /* The splittable_regs value must be a REG or a
1665 CONST_INT, so put the entire value in the giv_src_reg
1667 giv_src_reg
= splittable_regs
[regno
];
1668 giv_inc
= const0_rtx
;
1673 /* Partially unrolling loop. Create a new pseudo
1674 register for the iteration variable, and set it to
1675 be a constant plus the original register. Except
1676 on the last iteration, when the result has to
1677 go back into the original iteration var register. */
1679 /* Handle bivs which must be mapped to a new register
1680 when split. This happens for bivs which need their
1681 final value set before loop entry. The new register
1682 for the biv was stored in the biv's first struct
1683 induction entry by find_splittable_regs. */
1685 if (regno
< max_reg_before_loop
1686 && reg_iv_type
[regno
] == BASIC_INDUCT
)
1688 giv_src_reg
= reg_biv_class
[regno
]->biv
->src_reg
;
1689 giv_dest_reg
= giv_src_reg
;
1693 /* If non-reduced/final-value givs were split, then
1694 this would have to remap those givs also. See
1695 find_splittable_regs. */
1698 splittable_regs
[regno
]
1699 = GEN_INT (INTVAL (giv_inc
)
1700 + INTVAL (splittable_regs
[regno
]));
1701 giv_inc
= splittable_regs
[regno
];
1703 /* Now split the induction variable by changing the dest
1704 of this insn to a new register, and setting its
1705 reg_map entry to point to this new register.
1707 If this is the last iteration, and this is the last insn
1708 that will update the iv, then reuse the original dest,
1709 to ensure that the iv will have the proper value when
1710 the loop exits or repeats.
1712 Using splittable_regs_updates here like this is safe,
1713 because it can only be greater than one if all
1714 instructions modifying the iv are always executed in
1717 if (! last_iteration
1718 || (splittable_regs_updates
[regno
]-- != 1))
1720 tem
= gen_reg_rtx (GET_MODE (giv_src_reg
));
1722 map
->reg_map
[regno
] = tem
;
1725 map
->reg_map
[regno
] = giv_src_reg
;
1728 /* The constant being added could be too large for an add
1729 immediate, so can't directly emit an insn here. */
1730 emit_unrolled_add (giv_dest_reg
, giv_src_reg
, giv_inc
);
1731 copy
= get_last_insn ();
1732 pattern
= PATTERN (copy
);
1736 pattern
= copy_rtx_and_substitute (pattern
, map
);
1737 copy
= emit_insn (pattern
);
1739 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
1742 /* If this insn is setting CC0, it may need to look at
1743 the insn that uses CC0 to see what type of insn it is.
1744 In that case, the call to recog via validate_change will
1745 fail. So don't substitute constants here. Instead,
1746 do it when we emit the following insn.
1748 For example, see the pyr.md file. That machine has signed and
1749 unsigned compares. The compare patterns must check the
1750 following branch insn to see which what kind of compare to
1753 If the previous insn set CC0, substitute constants on it as
1755 if (sets_cc0_p (PATTERN (copy
)) != 0)
1760 try_constants (cc0_insn
, map
);
1762 try_constants (copy
, map
);
1765 try_constants (copy
, map
);
1768 /* Make split induction variable constants `permanent' since we
1769 know there are no backward branches across iteration variable
1770 settings which would invalidate this. */
1771 if (dest_reg_was_split
)
1773 int regno
= REGNO (SET_DEST (pattern
));
1775 if (regno
< map
->const_equiv_map_size
1776 && map
->const_age_map
[regno
] == map
->const_age
)
1777 map
->const_age_map
[regno
] = -1;
1782 pattern
= copy_rtx_and_substitute (PATTERN (insn
), map
);
1783 copy
= emit_jump_insn (pattern
);
1784 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
1786 if (JUMP_LABEL (insn
) == start_label
&& insn
== copy_end
1787 && ! last_iteration
)
1789 /* This is a branch to the beginning of the loop; this is the
1790 last insn being copied; and this is not the last iteration.
1791 In this case, we want to change the original fall through
1792 case to be a branch past the end of the loop, and the
1793 original jump label case to fall_through. */
1795 if (invert_exp (pattern
, copy
))
1797 if (! redirect_exp (&pattern
,
1798 map
->label_map
[CODE_LABEL_NUMBER
1799 (JUMP_LABEL (insn
))],
1806 rtx lab
= gen_label_rtx ();
1807 /* Can't do it by reversing the jump (probably because we
1808 couldn't reverse the conditions), so emit a new
1809 jump_insn after COPY, and redirect the jump around
1811 jmp
= emit_jump_insn_after (gen_jump (exit_label
), copy
);
1812 jmp
= emit_barrier_after (jmp
);
1813 emit_label_after (lab
, jmp
);
1814 LABEL_NUSES (lab
) = 0;
1815 if (! redirect_exp (&pattern
,
1816 map
->label_map
[CODE_LABEL_NUMBER
1817 (JUMP_LABEL (insn
))],
1825 try_constants (cc0_insn
, map
);
1828 try_constants (copy
, map
);
1830 /* Set the jump label of COPY correctly to avoid problems with
1831 later passes of unroll_loop, if INSN had jump label set. */
1832 if (JUMP_LABEL (insn
))
1836 /* Can't use the label_map for every insn, since this may be
1837 the backward branch, and hence the label was not mapped. */
1838 if (GET_CODE (pattern
) == SET
)
1840 tem
= SET_SRC (pattern
);
1841 if (GET_CODE (tem
) == LABEL_REF
)
1842 label
= XEXP (tem
, 0);
1843 else if (GET_CODE (tem
) == IF_THEN_ELSE
)
1845 if (XEXP (tem
, 1) != pc_rtx
)
1846 label
= XEXP (XEXP (tem
, 1), 0);
1848 label
= XEXP (XEXP (tem
, 2), 0);
1852 if (label
&& GET_CODE (label
) == CODE_LABEL
)
1853 JUMP_LABEL (copy
) = label
;
1856 /* An unrecognizable jump insn, probably the entry jump
1857 for a switch statement. This label must have been mapped,
1858 so just use the label_map to get the new jump label. */
1860 = map
->label_map
[CODE_LABEL_NUMBER (JUMP_LABEL (insn
))];
1863 /* If this is a non-local jump, then must increase the label
1864 use count so that the label will not be deleted when the
1865 original jump is deleted. */
1866 LABEL_NUSES (JUMP_LABEL (copy
))++;
1868 else if (GET_CODE (PATTERN (copy
)) == ADDR_VEC
1869 || GET_CODE (PATTERN (copy
)) == ADDR_DIFF_VEC
)
1871 rtx pat
= PATTERN (copy
);
1872 int diff_vec_p
= GET_CODE (pat
) == ADDR_DIFF_VEC
;
1873 int len
= XVECLEN (pat
, diff_vec_p
);
1876 for (i
= 0; i
< len
; i
++)
1877 LABEL_NUSES (XEXP (XVECEXP (pat
, diff_vec_p
, i
), 0))++;
1880 /* If this used to be a conditional jump insn but whose branch
1881 direction is now known, we must do something special. */
1882 if (condjump_p (insn
) && !simplejump_p (insn
) && map
->last_pc_value
)
1885 /* The previous insn set cc0 for us. So delete it. */
1886 delete_insn (PREV_INSN (copy
));
1889 /* If this is now a no-op, delete it. */
1890 if (map
->last_pc_value
== pc_rtx
)
1892 /* Don't let delete_insn delete the label referenced here,
1893 because we might possibly need it later for some other
1894 instruction in the loop. */
1895 if (JUMP_LABEL (copy
))
1896 LABEL_NUSES (JUMP_LABEL (copy
))++;
1898 if (JUMP_LABEL (copy
))
1899 LABEL_NUSES (JUMP_LABEL (copy
))--;
1903 /* Otherwise, this is unconditional jump so we must put a
1904 BARRIER after it. We could do some dead code elimination
1905 here, but jump.c will do it just as well. */
1911 pattern
= copy_rtx_and_substitute (PATTERN (insn
), map
);
1912 copy
= emit_call_insn (pattern
);
1913 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
1915 /* Because the USAGE information potentially contains objects other
1916 than hard registers, we need to copy it. */
1917 CALL_INSN_FUNCTION_USAGE (copy
) =
1918 copy_rtx_and_substitute (CALL_INSN_FUNCTION_USAGE (insn
), map
);
1922 try_constants (cc0_insn
, map
);
1925 try_constants (copy
, map
);
1927 /* Be lazy and assume CALL_INSNs clobber all hard registers. */
1928 for (i
= 0; i
< FIRST_PSEUDO_REGISTER
; i
++)
1929 map
->const_equiv_map
[i
] = 0;
1933 /* If this is the loop start label, then we don't need to emit a
1934 copy of this label since no one will use it. */
1936 if (insn
!= start_label
)
1938 copy
= emit_label (map
->label_map
[CODE_LABEL_NUMBER (insn
)]);
1944 copy
= emit_barrier ();
1948 /* VTOP notes are valid only before the loop exit test. If placed
1949 anywhere else, loop may generate bad code. */
1951 if (NOTE_LINE_NUMBER (insn
) != NOTE_INSN_DELETED
1952 && (NOTE_LINE_NUMBER (insn
) != NOTE_INSN_LOOP_VTOP
1953 || (last_iteration
&& unroll_type
!= UNROLL_COMPLETELY
)))
1954 copy
= emit_note (NOTE_SOURCE_FILE (insn
),
1955 NOTE_LINE_NUMBER (insn
));
1965 map
->insn_map
[INSN_UID (insn
)] = copy
;
1967 while (insn
!= copy_end
);
1969 /* Now finish coping the REG_NOTES. */
1973 insn
= NEXT_INSN (insn
);
1974 if ((GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == JUMP_INSN
1975 || GET_CODE (insn
) == CALL_INSN
)
1976 && map
->insn_map
[INSN_UID (insn
)])
1977 final_reg_note_copy (REG_NOTES (map
->insn_map
[INSN_UID (insn
)]), map
);
1979 while (insn
!= copy_end
);
1981 /* There may be notes between copy_notes_from and loop_end. Emit a copy of
1982 each of these notes here, since there may be some important ones, such as
1983 NOTE_INSN_BLOCK_END notes, in this group. We don't do this on the last
1984 iteration, because the original notes won't be deleted.
1986 We can't use insert_before here, because when from preconditioning,
1987 insert_before points before the loop. We can't use copy_end, because
1988 there may be insns already inserted after it (which we don't want to
1989 copy) when not from preconditioning code. */
1991 if (! last_iteration
)
1993 for (insn
= copy_notes_from
; insn
!= loop_end
; insn
= NEXT_INSN (insn
))
1995 if (GET_CODE (insn
) == NOTE
1996 && NOTE_LINE_NUMBER (insn
) != NOTE_INSN_DELETED
)
1997 emit_note (NOTE_SOURCE_FILE (insn
), NOTE_LINE_NUMBER (insn
));
2001 if (final_label
&& LABEL_NUSES (final_label
) > 0)
2002 emit_label (final_label
);
2004 tem
= gen_sequence ();
2006 emit_insn_before (tem
, insert_before
);
2009 /* Emit an insn, using the expand_binop to ensure that a valid insn is
2010 emitted. This will correctly handle the case where the increment value
2011 won't fit in the immediate field of a PLUS insns. */
2014 emit_unrolled_add (dest_reg
, src_reg
, increment
)
2015 rtx dest_reg
, src_reg
, increment
;
2019 result
= expand_binop (GET_MODE (dest_reg
), add_optab
, src_reg
, increment
,
2020 dest_reg
, 0, OPTAB_LIB_WIDEN
);
2022 if (dest_reg
!= result
)
2023 emit_move_insn (dest_reg
, result
);
2026 /* Searches the insns between INSN and LOOP_END. Returns 1 if there
2027 is a backward branch in that range that branches to somewhere between
2028 LOOP_START and INSN. Returns 0 otherwise. */
2030 /* ??? This is quadratic algorithm. Could be rewritten to be linear.
2031 In practice, this is not a problem, because this function is seldom called,
2032 and uses a negligible amount of CPU time on average. */
2035 back_branch_in_range_p (insn
, loop_start
, loop_end
)
2037 rtx loop_start
, loop_end
;
2039 rtx p
, q
, target_insn
;
2041 /* Stop before we get to the backward branch at the end of the loop. */
2042 loop_end
= prev_nonnote_insn (loop_end
);
2043 if (GET_CODE (loop_end
) == BARRIER
)
2044 loop_end
= PREV_INSN (loop_end
);
2046 /* Check in case insn has been deleted, search forward for first non
2047 deleted insn following it. */
2048 while (INSN_DELETED_P (insn
))
2049 insn
= NEXT_INSN (insn
);
2051 /* Check for the case where insn is the last insn in the loop. */
2052 if (insn
== loop_end
)
2055 for (p
= NEXT_INSN (insn
); p
!= loop_end
; p
= NEXT_INSN (p
))
2057 if (GET_CODE (p
) == JUMP_INSN
)
2059 target_insn
= JUMP_LABEL (p
);
2061 /* Search from loop_start to insn, to see if one of them is
2062 the target_insn. We can't use INSN_LUID comparisons here,
2063 since insn may not have an LUID entry. */
2064 for (q
= loop_start
; q
!= insn
; q
= NEXT_INSN (q
))
2065 if (q
== target_insn
)
2073 /* Try to generate the simplest rtx for the expression
2074 (PLUS (MULT mult1 mult2) add1). This is used to calculate the initial
2078 fold_rtx_mult_add (mult1
, mult2
, add1
, mode
)
2079 rtx mult1
, mult2
, add1
;
2080 enum machine_mode mode
;
2085 /* The modes must all be the same. This should always be true. For now,
2086 check to make sure. */
2087 if ((GET_MODE (mult1
) != mode
&& GET_MODE (mult1
) != VOIDmode
)
2088 || (GET_MODE (mult2
) != mode
&& GET_MODE (mult2
) != VOIDmode
)
2089 || (GET_MODE (add1
) != mode
&& GET_MODE (add1
) != VOIDmode
))
2092 /* Ensure that if at least one of mult1/mult2 are constant, then mult2
2093 will be a constant. */
2094 if (GET_CODE (mult1
) == CONST_INT
)
2101 mult_res
= simplify_binary_operation (MULT
, mode
, mult1
, mult2
);
2103 mult_res
= gen_rtx (MULT
, mode
, mult1
, mult2
);
2105 /* Again, put the constant second. */
2106 if (GET_CODE (add1
) == CONST_INT
)
2113 result
= simplify_binary_operation (PLUS
, mode
, add1
, mult_res
);
2115 result
= gen_rtx (PLUS
, mode
, add1
, mult_res
);
2120 /* Searches the list of induction struct's for the biv BL, to try to calculate
2121 the total increment value for one iteration of the loop as a constant.
2123 Returns the increment value as an rtx, simplified as much as possible,
2124 if it can be calculated. Otherwise, returns 0. */
2127 biv_total_increment (bl
, loop_start
, loop_end
)
2128 struct iv_class
*bl
;
2129 rtx loop_start
, loop_end
;
2131 struct induction
*v
;
2134 /* For increment, must check every instruction that sets it. Each
2135 instruction must be executed only once each time through the loop.
2136 To verify this, we check that the the insn is always executed, and that
2137 there are no backward branches after the insn that branch to before it.
2138 Also, the insn must have a mult_val of one (to make sure it really is
2141 result
= const0_rtx
;
2142 for (v
= bl
->biv
; v
; v
= v
->next_iv
)
2144 if (v
->always_computable
&& v
->mult_val
== const1_rtx
2145 && ! back_branch_in_range_p (v
->insn
, loop_start
, loop_end
))
2146 result
= fold_rtx_mult_add (result
, const1_rtx
, v
->add_val
, v
->mode
);
2154 /* Determine the initial value of the iteration variable, and the amount
2155 that it is incremented each loop. Use the tables constructed by
2156 the strength reduction pass to calculate these values.
2158 Initial_value and/or increment are set to zero if their values could not
2162 iteration_info (iteration_var
, initial_value
, increment
, loop_start
, loop_end
)
2163 rtx iteration_var
, *initial_value
, *increment
;
2164 rtx loop_start
, loop_end
;
2166 struct iv_class
*bl
;
2167 struct induction
*v
, *b
;
2169 /* Clear the result values, in case no answer can be found. */
2173 /* The iteration variable can be either a giv or a biv. Check to see
2174 which it is, and compute the variable's initial value, and increment
2175 value if possible. */
2177 /* If this is a new register, can't handle it since we don't have any
2178 reg_iv_type entry for it. */
2179 if (REGNO (iteration_var
) >= max_reg_before_loop
)
2181 if (loop_dump_stream
)
2182 fprintf (loop_dump_stream
,
2183 "Loop unrolling: No reg_iv_type entry for iteration var.\n");
2186 /* Reject iteration variables larger than the host long size, since they
2187 could result in a number of iterations greater than the range of our
2188 `unsigned long' variable loop_n_iterations. */
2189 else if (GET_MODE_BITSIZE (GET_MODE (iteration_var
)) > HOST_BITS_PER_LONG
)
2191 if (loop_dump_stream
)
2192 fprintf (loop_dump_stream
,
2193 "Loop unrolling: Iteration var rejected because mode larger than host long.\n");
2196 else if (GET_MODE_CLASS (GET_MODE (iteration_var
)) != MODE_INT
)
2198 if (loop_dump_stream
)
2199 fprintf (loop_dump_stream
,
2200 "Loop unrolling: Iteration var not an integer.\n");
2203 else if (reg_iv_type
[REGNO (iteration_var
)] == BASIC_INDUCT
)
2205 /* Grab initial value, only useful if it is a constant. */
2206 bl
= reg_biv_class
[REGNO (iteration_var
)];
2207 *initial_value
= bl
->initial_value
;
2209 *increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2211 else if (reg_iv_type
[REGNO (iteration_var
)] == GENERAL_INDUCT
)
2214 /* ??? The code below does not work because the incorrect number of
2215 iterations is calculated when the biv is incremented after the giv
2216 is set (which is the usual case). This can probably be accounted
2217 for by biasing the initial_value by subtracting the amount of the
2218 increment that occurs between the giv set and the giv test. However,
2219 a giv as an iterator is very rare, so it does not seem worthwhile
2221 /* ??? An example failure is: i = 6; do {;} while (i++ < 9). */
2222 if (loop_dump_stream
)
2223 fprintf (loop_dump_stream
,
2224 "Loop unrolling: Giv iterators are not handled.\n");
2227 /* Initial value is mult_val times the biv's initial value plus
2228 add_val. Only useful if it is a constant. */
2229 v
= reg_iv_info
[REGNO (iteration_var
)];
2230 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
2231 *initial_value
= fold_rtx_mult_add (v
->mult_val
, bl
->initial_value
,
2232 v
->add_val
, v
->mode
);
2234 /* Increment value is mult_val times the increment value of the biv. */
2236 *increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2238 *increment
= fold_rtx_mult_add (v
->mult_val
, *increment
, const0_rtx
,
2244 if (loop_dump_stream
)
2245 fprintf (loop_dump_stream
,
2246 "Loop unrolling: Not basic or general induction var.\n");
2251 /* Calculate the approximate final value of the iteration variable
2252 which has an loop exit test with code COMPARISON_CODE and comparison value
2253 of COMPARISON_VALUE. Also returns an indication of whether the comparison
2254 was signed or unsigned, and the direction of the comparison. This info is
2255 needed to calculate the number of loop iterations. */
2258 approx_final_value (comparison_code
, comparison_value
, unsigned_p
, compare_dir
)
2259 enum rtx_code comparison_code
;
2260 rtx comparison_value
;
2264 /* Calculate the final value of the induction variable.
2265 The exact final value depends on the branch operator, and increment sign.
2266 This is only an approximate value. It will be wrong if the iteration
2267 variable is not incremented by one each time through the loop, and
2268 approx final value - start value % increment != 0. */
2271 switch (comparison_code
)
2277 return plus_constant (comparison_value
, 1);
2282 return plus_constant (comparison_value
, -1);
2284 /* Can not calculate a final value for this case. */
2291 return comparison_value
;
2297 return comparison_value
;
2300 return comparison_value
;
2306 /* For each biv and giv, determine whether it can be safely split into
2307 a different variable for each unrolled copy of the loop body. If it
2308 is safe to split, then indicate that by saving some useful info
2309 in the splittable_regs array.
2311 If the loop is being completely unrolled, then splittable_regs will hold
2312 the current value of the induction variable while the loop is unrolled.
2313 It must be set to the initial value of the induction variable here.
2314 Otherwise, splittable_regs will hold the difference between the current
2315 value of the induction variable and the value the induction variable had
2316 at the top of the loop. It must be set to the value 0 here.
2318 Returns the total number of instructions that set registers that are
2321 /* ?? If the loop is only unrolled twice, then most of the restrictions to
2322 constant values are unnecessary, since we can easily calculate increment
2323 values in this case even if nothing is constant. The increment value
2324 should not involve a multiply however. */
2326 /* ?? Even if the biv/giv increment values aren't constant, it may still
2327 be beneficial to split the variable if the loop is only unrolled a few
2328 times, since multiplies by small integers (1,2,3,4) are very cheap. */
2331 find_splittable_regs (unroll_type
, loop_start
, loop_end
, end_insert_before
,
2333 enum unroll_types unroll_type
;
2334 rtx loop_start
, loop_end
;
2335 rtx end_insert_before
;
2338 struct iv_class
*bl
;
2339 struct induction
*v
;
2341 rtx biv_final_value
;
2345 for (bl
= loop_iv_list
; bl
; bl
= bl
->next
)
2347 /* Biv_total_increment must return a constant value,
2348 otherwise we can not calculate the split values. */
2350 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2351 if (! increment
|| GET_CODE (increment
) != CONST_INT
)
2354 /* The loop must be unrolled completely, or else have a known number
2355 of iterations and only one exit, or else the biv must be dead
2356 outside the loop, or else the final value must be known. Otherwise,
2357 it is unsafe to split the biv since it may not have the proper
2358 value on loop exit. */
2360 /* loop_number_exit_count is non-zero if the loop has an exit other than
2361 a fall through at the end. */
2364 biv_final_value
= 0;
2365 if (unroll_type
!= UNROLL_COMPLETELY
2366 && (loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]]
2367 || unroll_type
== UNROLL_NAIVE
)
2368 && (uid_luid
[regno_last_uid
[bl
->regno
]] >= INSN_LUID (loop_end
)
2370 || INSN_UID (bl
->init_insn
) >= max_uid_for_loop
2371 || (uid_luid
[regno_first_uid
[bl
->regno
]]
2372 < INSN_LUID (bl
->init_insn
))
2373 || reg_mentioned_p (bl
->biv
->dest_reg
, SET_SRC (bl
->init_set
)))
2374 && ! (biv_final_value
= final_biv_value (bl
, loop_start
, loop_end
)))
2377 /* If any of the insns setting the BIV don't do so with a simple
2378 PLUS, we don't know how to split it. */
2379 for (v
= bl
->biv
; biv_splittable
&& v
; v
= v
->next_iv
)
2380 if ((tem
= single_set (v
->insn
)) == 0
2381 || GET_CODE (SET_DEST (tem
)) != REG
2382 || REGNO (SET_DEST (tem
)) != bl
->regno
2383 || GET_CODE (SET_SRC (tem
)) != PLUS
)
2386 /* If final value is non-zero, then must emit an instruction which sets
2387 the value of the biv to the proper value. This is done after
2388 handling all of the givs, since some of them may need to use the
2389 biv's value in their initialization code. */
2391 /* This biv is splittable. If completely unrolling the loop, save
2392 the biv's initial value. Otherwise, save the constant zero. */
2394 if (biv_splittable
== 1)
2396 if (unroll_type
== UNROLL_COMPLETELY
)
2398 /* If the initial value of the biv is itself (i.e. it is too
2399 complicated for strength_reduce to compute), or is a hard
2400 register, or it isn't invariant, then we must create a new
2401 pseudo reg to hold the initial value of the biv. */
2403 if (GET_CODE (bl
->initial_value
) == REG
2404 && (REGNO (bl
->initial_value
) == bl
->regno
2405 || REGNO (bl
->initial_value
) < FIRST_PSEUDO_REGISTER
2406 || ! invariant_p (bl
->initial_value
)))
2408 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2410 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2413 if (loop_dump_stream
)
2414 fprintf (loop_dump_stream
, "Biv %d initial value remapped to %d.\n",
2415 bl
->regno
, REGNO (tem
));
2417 splittable_regs
[bl
->regno
] = tem
;
2420 splittable_regs
[bl
->regno
] = bl
->initial_value
;
2423 splittable_regs
[bl
->regno
] = const0_rtx
;
2425 /* Save the number of instructions that modify the biv, so that
2426 we can treat the last one specially. */
2428 splittable_regs_updates
[bl
->regno
] = bl
->biv_count
;
2429 result
+= bl
->biv_count
;
2431 if (loop_dump_stream
)
2432 fprintf (loop_dump_stream
,
2433 "Biv %d safe to split.\n", bl
->regno
);
2436 /* Check every giv that depends on this biv to see whether it is
2437 splittable also. Even if the biv isn't splittable, givs which
2438 depend on it may be splittable if the biv is live outside the
2439 loop, and the givs aren't. */
2441 result
+= find_splittable_givs (bl
, unroll_type
, loop_start
, loop_end
,
2442 increment
, unroll_number
);
2444 /* If final value is non-zero, then must emit an instruction which sets
2445 the value of the biv to the proper value. This is done after
2446 handling all of the givs, since some of them may need to use the
2447 biv's value in their initialization code. */
2448 if (biv_final_value
)
2450 /* If the loop has multiple exits, emit the insns before the
2451 loop to ensure that it will always be executed no matter
2452 how the loop exits. Otherwise emit the insn after the loop,
2453 since this is slightly more efficient. */
2454 if (! loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]])
2455 emit_insn_before (gen_move_insn (bl
->biv
->src_reg
,
2460 /* Create a new register to hold the value of the biv, and then
2461 set the biv to its final value before the loop start. The biv
2462 is set to its final value before loop start to ensure that
2463 this insn will always be executed, no matter how the loop
2465 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2466 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2468 emit_insn_before (gen_move_insn (bl
->biv
->src_reg
,
2472 if (loop_dump_stream
)
2473 fprintf (loop_dump_stream
, "Biv %d mapped to %d for split.\n",
2474 REGNO (bl
->biv
->src_reg
), REGNO (tem
));
2476 /* Set up the mapping from the original biv register to the new
2478 bl
->biv
->src_reg
= tem
;
2485 /* Return 1 if the first and last unrolled copy of the address giv V is valid
2486 for the instruction that is using it. Do not make any changes to that
2490 verify_addresses (v
, giv_inc
, unroll_number
)
2491 struct induction
*v
;
2496 rtx orig_addr
= *v
->location
;
2497 rtx last_addr
= plus_constant (v
->dest_reg
,
2498 INTVAL (giv_inc
) * (unroll_number
- 1));
2500 /* First check to see if either address would fail. */
2501 if (! validate_change (v
->insn
, v
->location
, v
->dest_reg
, 0)
2502 || ! validate_change (v
->insn
, v
->location
, last_addr
, 0))
2505 /* Now put things back the way they were before. This will always
2507 validate_change (v
->insn
, v
->location
, orig_addr
, 0);
2512 /* For every giv based on the biv BL, check to determine whether it is
2513 splittable. This is a subroutine to find_splittable_regs ().
2515 Return the number of instructions that set splittable registers. */
2518 find_splittable_givs (bl
, unroll_type
, loop_start
, loop_end
, increment
,
2520 struct iv_class
*bl
;
2521 enum unroll_types unroll_type
;
2522 rtx loop_start
, loop_end
;
2526 struct induction
*v
, *v2
;
2531 /* Scan the list of givs, and set the same_insn field when there are
2532 multiple identical givs in the same insn. */
2533 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
2534 for (v2
= v
->next_iv
; v2
; v2
= v2
->next_iv
)
2535 if (v
->insn
== v2
->insn
&& rtx_equal_p (v
->new_reg
, v2
->new_reg
)
2539 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
2543 /* Only split the giv if it has already been reduced, or if the loop is
2544 being completely unrolled. */
2545 if (unroll_type
!= UNROLL_COMPLETELY
&& v
->ignore
)
2548 /* The giv can be split if the insn that sets the giv is executed once
2549 and only once on every iteration of the loop. */
2550 /* An address giv can always be split. v->insn is just a use not a set,
2551 and hence it does not matter whether it is always executed. All that
2552 matters is that all the biv increments are always executed, and we
2553 won't reach here if they aren't. */
2554 if (v
->giv_type
!= DEST_ADDR
2555 && (! v
->always_computable
2556 || back_branch_in_range_p (v
->insn
, loop_start
, loop_end
)))
2559 /* The giv increment value must be a constant. */
2560 giv_inc
= fold_rtx_mult_add (v
->mult_val
, increment
, const0_rtx
,
2562 if (! giv_inc
|| GET_CODE (giv_inc
) != CONST_INT
)
2565 /* The loop must be unrolled completely, or else have a known number of
2566 iterations and only one exit, or else the giv must be dead outside
2567 the loop, or else the final value of the giv must be known.
2568 Otherwise, it is not safe to split the giv since it may not have the
2569 proper value on loop exit. */
2571 /* The used outside loop test will fail for DEST_ADDR givs. They are
2572 never used outside the loop anyways, so it is always safe to split a
2576 if (unroll_type
!= UNROLL_COMPLETELY
2577 && (loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]]
2578 || unroll_type
== UNROLL_NAIVE
)
2579 && v
->giv_type
!= DEST_ADDR
2580 && ((regno_first_uid
[REGNO (v
->dest_reg
)] != INSN_UID (v
->insn
)
2581 /* Check for the case where the pseudo is set by a shift/add
2582 sequence, in which case the first insn setting the pseudo
2583 is the first insn of the shift/add sequence. */
2584 && (! (tem
= find_reg_note (v
->insn
, REG_RETVAL
, NULL_RTX
))
2585 || (regno_first_uid
[REGNO (v
->dest_reg
)]
2586 != INSN_UID (XEXP (tem
, 0)))))
2587 /* Line above always fails if INSN was moved by loop opt. */
2588 || (uid_luid
[regno_last_uid
[REGNO (v
->dest_reg
)]]
2589 >= INSN_LUID (loop_end
)))
2590 && ! (final_value
= v
->final_value
))
2594 /* Currently, non-reduced/final-value givs are never split. */
2595 /* Should emit insns after the loop if possible, as the biv final value
2598 /* If the final value is non-zero, and the giv has not been reduced,
2599 then must emit an instruction to set the final value. */
2600 if (final_value
&& !v
->new_reg
)
2602 /* Create a new register to hold the value of the giv, and then set
2603 the giv to its final value before the loop start. The giv is set
2604 to its final value before loop start to ensure that this insn
2605 will always be executed, no matter how we exit. */
2606 tem
= gen_reg_rtx (v
->mode
);
2607 emit_insn_before (gen_move_insn (tem
, v
->dest_reg
), loop_start
);
2608 emit_insn_before (gen_move_insn (v
->dest_reg
, final_value
),
2611 if (loop_dump_stream
)
2612 fprintf (loop_dump_stream
, "Giv %d mapped to %d for split.\n",
2613 REGNO (v
->dest_reg
), REGNO (tem
));
2619 /* This giv is splittable. If completely unrolling the loop, save the
2620 giv's initial value. Otherwise, save the constant zero for it. */
2622 if (unroll_type
== UNROLL_COMPLETELY
)
2624 /* It is not safe to use bl->initial_value here, because it may not
2625 be invariant. It is safe to use the initial value stored in
2626 the splittable_regs array if it is set. In rare cases, it won't
2627 be set, so then we do exactly the same thing as
2628 find_splittable_regs does to get a safe value. */
2629 rtx biv_initial_value
;
2631 if (splittable_regs
[bl
->regno
])
2632 biv_initial_value
= splittable_regs
[bl
->regno
];
2633 else if (GET_CODE (bl
->initial_value
) != REG
2634 || (REGNO (bl
->initial_value
) != bl
->regno
2635 && REGNO (bl
->initial_value
) >= FIRST_PSEUDO_REGISTER
))
2636 biv_initial_value
= bl
->initial_value
;
2639 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2641 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2643 biv_initial_value
= tem
;
2645 value
= fold_rtx_mult_add (v
->mult_val
, biv_initial_value
,
2646 v
->add_val
, v
->mode
);
2653 /* If a giv was combined with another giv, then we can only split
2654 this giv if the giv it was combined with was reduced. This
2655 is because the value of v->new_reg is meaningless in this
2657 if (v
->same
&& ! v
->same
->new_reg
)
2659 if (loop_dump_stream
)
2660 fprintf (loop_dump_stream
,
2661 "giv combined with unreduced giv not split.\n");
2664 /* If the giv is an address destination, it could be something other
2665 than a simple register, these have to be treated differently. */
2666 else if (v
->giv_type
== DEST_REG
)
2668 /* If value is not a constant, register, or register plus
2669 constant, then compute its value into a register before
2670 loop start. This prevents invalid rtx sharing, and should
2671 generate better code. We can use bl->initial_value here
2672 instead of splittable_regs[bl->regno] because this code
2673 is going before the loop start. */
2674 if (unroll_type
== UNROLL_COMPLETELY
2675 && GET_CODE (value
) != CONST_INT
2676 && GET_CODE (value
) != REG
2677 && (GET_CODE (value
) != PLUS
2678 || GET_CODE (XEXP (value
, 0)) != REG
2679 || GET_CODE (XEXP (value
, 1)) != CONST_INT
))
2681 rtx tem
= gen_reg_rtx (v
->mode
);
2682 emit_iv_add_mult (bl
->initial_value
, v
->mult_val
,
2683 v
->add_val
, tem
, loop_start
);
2687 splittable_regs
[REGNO (v
->new_reg
)] = value
;
2691 /* Splitting address givs is useful since it will often allow us
2692 to eliminate some increment insns for the base giv as
2695 /* If the addr giv is combined with a dest_reg giv, then all
2696 references to that dest reg will be remapped, which is NOT
2697 what we want for split addr regs. We always create a new
2698 register for the split addr giv, just to be safe. */
2700 /* ??? If there are multiple address givs which have been
2701 combined with the same dest_reg giv, then we may only need
2702 one new register for them. Pulling out constants below will
2703 catch some of the common cases of this. Currently, I leave
2704 the work of simplifying multiple address givs to the
2705 following cse pass. */
2707 /* As a special case, if we have multiple identical address givs
2708 within a single instruction, then we do use a single pseudo
2709 reg for both. This is necessary in case one is a match_dup
2712 v
->const_adjust
= 0;
2716 v
->dest_reg
= v
->same_insn
->dest_reg
;
2717 if (loop_dump_stream
)
2718 fprintf (loop_dump_stream
,
2719 "Sharing address givs in insn %d\n",
2720 INSN_UID (v
->insn
));
2722 else if (unroll_type
!= UNROLL_COMPLETELY
)
2724 /* If not completely unrolling the loop, then create a new
2725 register to hold the split value of the DEST_ADDR giv.
2726 Emit insn to initialize its value before loop start. */
2727 tem
= gen_reg_rtx (v
->mode
);
2729 /* If the address giv has a constant in its new_reg value,
2730 then this constant can be pulled out and put in value,
2731 instead of being part of the initialization code. */
2733 if (GET_CODE (v
->new_reg
) == PLUS
2734 && GET_CODE (XEXP (v
->new_reg
, 1)) == CONST_INT
)
2737 = plus_constant (tem
, INTVAL (XEXP (v
->new_reg
,1)));
2739 /* Only succeed if this will give valid addresses.
2740 Try to validate both the first and the last
2741 address resulting from loop unrolling, if
2742 one fails, then can't do const elim here. */
2743 if (! verify_addresses (v
, giv_inc
, unroll_number
))
2745 /* Save the negative of the eliminated const, so
2746 that we can calculate the dest_reg's increment
2748 v
->const_adjust
= - INTVAL (XEXP (v
->new_reg
, 1));
2750 v
->new_reg
= XEXP (v
->new_reg
, 0);
2751 if (loop_dump_stream
)
2752 fprintf (loop_dump_stream
,
2753 "Eliminating constant from giv %d\n",
2762 /* If the address hasn't been checked for validity yet, do so
2763 now, and fail completely if either the first or the last
2764 unrolled copy of the address is not a valid address
2765 for the instruction that uses it. */
2766 if (v
->dest_reg
== tem
2767 && ! verify_addresses (v
, giv_inc
, unroll_number
))
2769 if (loop_dump_stream
)
2770 fprintf (loop_dump_stream
,
2771 "Invalid address for giv at insn %d\n",
2772 INSN_UID (v
->insn
));
2776 /* To initialize the new register, just move the value of
2777 new_reg into it. This is not guaranteed to give a valid
2778 instruction on machines with complex addressing modes.
2779 If we can't recognize it, then delete it and emit insns
2780 to calculate the value from scratch. */
2781 emit_insn_before (gen_rtx (SET
, VOIDmode
, tem
,
2782 copy_rtx (v
->new_reg
)),
2784 if (recog_memoized (PREV_INSN (loop_start
)) < 0)
2788 /* We can't use bl->initial_value to compute the initial
2789 value, because the loop may have been preconditioned.
2790 We must calculate it from NEW_REG. Try using
2791 force_operand instead of emit_iv_add_mult. */
2792 delete_insn (PREV_INSN (loop_start
));
2795 ret
= force_operand (v
->new_reg
, tem
);
2797 emit_move_insn (tem
, ret
);
2798 sequence
= gen_sequence ();
2800 emit_insn_before (sequence
, loop_start
);
2802 if (loop_dump_stream
)
2803 fprintf (loop_dump_stream
,
2804 "Invalid init insn, rewritten.\n");
2809 v
->dest_reg
= value
;
2811 /* Check the resulting address for validity, and fail
2812 if the resulting address would be invalid. */
2813 if (! verify_addresses (v
, giv_inc
, unroll_number
))
2815 if (loop_dump_stream
)
2816 fprintf (loop_dump_stream
,
2817 "Invalid address for giv at insn %d\n",
2818 INSN_UID (v
->insn
));
2823 /* Store the value of dest_reg into the insn. This sharing
2824 will not be a problem as this insn will always be copied
2827 *v
->location
= v
->dest_reg
;
2829 /* If this address giv is combined with a dest reg giv, then
2830 save the base giv's induction pointer so that we will be
2831 able to handle this address giv properly. The base giv
2832 itself does not have to be splittable. */
2834 if (v
->same
&& v
->same
->giv_type
== DEST_REG
)
2835 addr_combined_regs
[REGNO (v
->same
->new_reg
)] = v
->same
;
2837 if (GET_CODE (v
->new_reg
) == REG
)
2839 /* This giv maybe hasn't been combined with any others.
2840 Make sure that it's giv is marked as splittable here. */
2842 splittable_regs
[REGNO (v
->new_reg
)] = value
;
2844 /* Make it appear to depend upon itself, so that the
2845 giv will be properly split in the main loop above. */
2849 addr_combined_regs
[REGNO (v
->new_reg
)] = v
;
2853 if (loop_dump_stream
)
2854 fprintf (loop_dump_stream
, "DEST_ADDR giv being split.\n");
2860 /* Currently, unreduced giv's can't be split. This is not too much
2861 of a problem since unreduced giv's are not live across loop
2862 iterations anyways. When unrolling a loop completely though,
2863 it makes sense to reduce&split givs when possible, as this will
2864 result in simpler instructions, and will not require that a reg
2865 be live across loop iterations. */
2867 splittable_regs
[REGNO (v
->dest_reg
)] = value
;
2868 fprintf (stderr
, "Giv %d at insn %d not reduced\n",
2869 REGNO (v
->dest_reg
), INSN_UID (v
->insn
));
2875 /* Givs are only updated once by definition. Mark it so if this is
2876 a splittable register. Don't need to do anything for address givs
2877 where this may not be a register. */
2879 if (GET_CODE (v
->new_reg
) == REG
)
2880 splittable_regs_updates
[REGNO (v
->new_reg
)] = 1;
2884 if (loop_dump_stream
)
2888 if (GET_CODE (v
->dest_reg
) == CONST_INT
)
2890 else if (GET_CODE (v
->dest_reg
) != REG
)
2891 regnum
= REGNO (XEXP (v
->dest_reg
, 0));
2893 regnum
= REGNO (v
->dest_reg
);
2894 fprintf (loop_dump_stream
, "Giv %d at insn %d safe to split.\n",
2895 regnum
, INSN_UID (v
->insn
));
2902 /* Try to prove that the register is dead after the loop exits. Trace every
2903 loop exit looking for an insn that will always be executed, which sets
2904 the register to some value, and appears before the first use of the register
2905 is found. If successful, then return 1, otherwise return 0. */
2907 /* ?? Could be made more intelligent in the handling of jumps, so that
2908 it can search past if statements and other similar structures. */
2911 reg_dead_after_loop (reg
, loop_start
, loop_end
)
2912 rtx reg
, loop_start
, loop_end
;
2917 int label_count
= 0;
2918 int this_loop_num
= uid_loop_num
[INSN_UID (loop_start
)];
2920 /* In addition to checking all exits of this loop, we must also check
2921 all exits of inner nested loops that would exit this loop. We don't
2922 have any way to identify those, so we just give up if there are any
2923 such inner loop exits. */
2925 for (label
= loop_number_exit_labels
[this_loop_num
]; label
;
2926 label
= LABEL_NEXTREF (label
))
2929 if (label_count
!= loop_number_exit_count
[this_loop_num
])
2932 /* HACK: Must also search the loop fall through exit, create a label_ref
2933 here which points to the loop_end, and append the loop_number_exit_labels
2935 label
= gen_rtx (LABEL_REF
, VOIDmode
, loop_end
);
2936 LABEL_NEXTREF (label
) = loop_number_exit_labels
[this_loop_num
];
2938 for ( ; label
; label
= LABEL_NEXTREF (label
))
2940 /* Succeed if find an insn which sets the biv or if reach end of
2941 function. Fail if find an insn that uses the biv, or if come to
2942 a conditional jump. */
2944 insn
= NEXT_INSN (XEXP (label
, 0));
2947 code
= GET_CODE (insn
);
2948 if (GET_RTX_CLASS (code
) == 'i')
2952 if (reg_referenced_p (reg
, PATTERN (insn
)))
2955 set
= single_set (insn
);
2956 if (set
&& rtx_equal_p (SET_DEST (set
), reg
))
2960 if (code
== JUMP_INSN
)
2962 if (GET_CODE (PATTERN (insn
)) == RETURN
)
2964 else if (! simplejump_p (insn
)
2965 /* Prevent infinite loop following infinite loops. */
2966 || jump_count
++ > 20)
2969 insn
= JUMP_LABEL (insn
);
2972 insn
= NEXT_INSN (insn
);
2976 /* Success, the register is dead on all loop exits. */
2980 /* Try to calculate the final value of the biv, the value it will have at
2981 the end of the loop. If we can do it, return that value. */
2984 final_biv_value (bl
, loop_start
, loop_end
)
2985 struct iv_class
*bl
;
2986 rtx loop_start
, loop_end
;
2990 /* ??? This only works for MODE_INT biv's. Reject all others for now. */
2992 if (GET_MODE_CLASS (bl
->biv
->mode
) != MODE_INT
)
2995 /* The final value for reversed bivs must be calculated differently than
2996 for ordinary bivs. In this case, there is already an insn after the
2997 loop which sets this biv's final value (if necessary), and there are
2998 no other loop exits, so we can return any value. */
3001 if (loop_dump_stream
)
3002 fprintf (loop_dump_stream
,
3003 "Final biv value for %d, reversed biv.\n", bl
->regno
);
3008 /* Try to calculate the final value as initial value + (number of iterations
3009 * increment). For this to work, increment must be invariant, the only
3010 exit from the loop must be the fall through at the bottom (otherwise
3011 it may not have its final value when the loop exits), and the initial
3012 value of the biv must be invariant. */
3014 if (loop_n_iterations
!= 0
3015 && ! loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]]
3016 && invariant_p (bl
->initial_value
))
3018 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
3020 if (increment
&& invariant_p (increment
))
3022 /* Can calculate the loop exit value, emit insns after loop
3023 end to calculate this value into a temporary register in
3024 case it is needed later. */
3026 tem
= gen_reg_rtx (bl
->biv
->mode
);
3027 /* Make sure loop_end is not the last insn. */
3028 if (NEXT_INSN (loop_end
) == 0)
3029 emit_note_after (NOTE_INSN_DELETED
, loop_end
);
3030 emit_iv_add_mult (increment
, GEN_INT (loop_n_iterations
),
3031 bl
->initial_value
, tem
, NEXT_INSN (loop_end
));
3033 if (loop_dump_stream
)
3034 fprintf (loop_dump_stream
,
3035 "Final biv value for %d, calculated.\n", bl
->regno
);
3041 /* Check to see if the biv is dead at all loop exits. */
3042 if (reg_dead_after_loop (bl
->biv
->src_reg
, loop_start
, loop_end
))
3044 if (loop_dump_stream
)
3045 fprintf (loop_dump_stream
,
3046 "Final biv value for %d, biv dead after loop exit.\n",
3055 /* Try to calculate the final value of the giv, the value it will have at
3056 the end of the loop. If we can do it, return that value. */
3059 final_giv_value (v
, loop_start
, loop_end
)
3060 struct induction
*v
;
3061 rtx loop_start
, loop_end
;
3063 struct iv_class
*bl
;
3066 rtx insert_before
, seq
;
3068 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
3070 /* The final value for givs which depend on reversed bivs must be calculated
3071 differently than for ordinary givs. In this case, there is already an
3072 insn after the loop which sets this giv's final value (if necessary),
3073 and there are no other loop exits, so we can return any value. */
3076 if (loop_dump_stream
)
3077 fprintf (loop_dump_stream
,
3078 "Final giv value for %d, depends on reversed biv\n",
3079 REGNO (v
->dest_reg
));
3083 /* Try to calculate the final value as a function of the biv it depends
3084 upon. The only exit from the loop must be the fall through at the bottom
3085 (otherwise it may not have its final value when the loop exits). */
3087 /* ??? Can calculate the final giv value by subtracting off the
3088 extra biv increments times the giv's mult_val. The loop must have
3089 only one exit for this to work, but the loop iterations does not need
3092 if (loop_n_iterations
!= 0
3093 && ! loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]])
3095 /* ?? It is tempting to use the biv's value here since these insns will
3096 be put after the loop, and hence the biv will have its final value
3097 then. However, this fails if the biv is subsequently eliminated.
3098 Perhaps determine whether biv's are eliminable before trying to
3099 determine whether giv's are replaceable so that we can use the
3100 biv value here if it is not eliminable. */
3102 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
3104 if (increment
&& invariant_p (increment
))
3106 /* Can calculate the loop exit value of its biv as
3107 (loop_n_iterations * increment) + initial_value */
3109 /* The loop exit value of the giv is then
3110 (final_biv_value - extra increments) * mult_val + add_val.
3111 The extra increments are any increments to the biv which
3112 occur in the loop after the giv's value is calculated.
3113 We must search from the insn that sets the giv to the end
3114 of the loop to calculate this value. */
3116 insert_before
= NEXT_INSN (loop_end
);
3118 /* Put the final biv value in tem. */
3119 tem
= gen_reg_rtx (bl
->biv
->mode
);
3120 emit_iv_add_mult (increment
, GEN_INT (loop_n_iterations
),
3121 bl
->initial_value
, tem
, insert_before
);
3123 /* Subtract off extra increments as we find them. */
3124 for (insn
= NEXT_INSN (v
->insn
); insn
!= loop_end
;
3125 insn
= NEXT_INSN (insn
))
3127 struct induction
*biv
;
3129 for (biv
= bl
->biv
; biv
; biv
= biv
->next_iv
)
3130 if (biv
->insn
== insn
)
3133 tem
= expand_binop (GET_MODE (tem
), sub_optab
, tem
,
3134 biv
->add_val
, NULL_RTX
, 0,
3136 seq
= gen_sequence ();
3138 emit_insn_before (seq
, insert_before
);
3142 /* Now calculate the giv's final value. */
3143 emit_iv_add_mult (tem
, v
->mult_val
, v
->add_val
, tem
,
3146 if (loop_dump_stream
)
3147 fprintf (loop_dump_stream
,
3148 "Final giv value for %d, calc from biv's value.\n",
3149 REGNO (v
->dest_reg
));
3155 /* Replaceable giv's should never reach here. */
3159 /* Check to see if the biv is dead at all loop exits. */
3160 if (reg_dead_after_loop (v
->dest_reg
, loop_start
, loop_end
))
3162 if (loop_dump_stream
)
3163 fprintf (loop_dump_stream
,
3164 "Final giv value for %d, giv dead after loop exit.\n",
3165 REGNO (v
->dest_reg
));
3174 /* Calculate the number of loop iterations. Returns the exact number of loop
3175 iterations if it can be calculated, otherwise returns zero. */
3177 unsigned HOST_WIDE_INT
3178 loop_iterations (loop_start
, loop_end
)
3179 rtx loop_start
, loop_end
;
3181 rtx comparison
, comparison_value
;
3182 rtx iteration_var
, initial_value
, increment
, final_value
;
3183 enum rtx_code comparison_code
;
3186 int unsigned_compare
, compare_dir
, final_larger
;
3187 unsigned long tempu
;
3190 /* First find the iteration variable. If the last insn is a conditional
3191 branch, and the insn before tests a register value, make that the
3192 iteration variable. */
3194 loop_initial_value
= 0;
3196 loop_final_value
= 0;
3197 loop_iteration_var
= 0;
3199 /* We used to use pren_nonnote_insn here, but that fails because it might
3200 accidentally get the branch for a contained loop if the branch for this
3201 loop was deleted. We can only trust branches immediately before the
3203 last_loop_insn
= PREV_INSN (loop_end
);
3205 comparison
= get_condition_for_loop (last_loop_insn
);
3206 if (comparison
== 0)
3208 if (loop_dump_stream
)
3209 fprintf (loop_dump_stream
,
3210 "Loop unrolling: No final conditional branch found.\n");
3214 /* ??? Get_condition may switch position of induction variable and
3215 invariant register when it canonicalizes the comparison. */
3217 comparison_code
= GET_CODE (comparison
);
3218 iteration_var
= XEXP (comparison
, 0);
3219 comparison_value
= XEXP (comparison
, 1);
3221 if (GET_CODE (iteration_var
) != REG
)
3223 if (loop_dump_stream
)
3224 fprintf (loop_dump_stream
,
3225 "Loop unrolling: Comparison not against register.\n");
3229 /* Loop iterations is always called before any new registers are created
3230 now, so this should never occur. */
3232 if (REGNO (iteration_var
) >= max_reg_before_loop
)
3235 iteration_info (iteration_var
, &initial_value
, &increment
,
3236 loop_start
, loop_end
);
3237 if (initial_value
== 0)
3238 /* iteration_info already printed a message. */
3241 /* If the comparison value is an invariant register, then try to find
3242 its value from the insns before the start of the loop. */
3244 if (GET_CODE (comparison_value
) == REG
&& invariant_p (comparison_value
))
3248 for (insn
= PREV_INSN (loop_start
); insn
; insn
= PREV_INSN (insn
))
3250 if (GET_CODE (insn
) == CODE_LABEL
)
3253 else if (GET_RTX_CLASS (GET_CODE (insn
)) == 'i'
3254 && reg_set_p (comparison_value
, insn
))
3256 /* We found the last insn before the loop that sets the register.
3257 If it sets the entire register, and has a REG_EQUAL note,
3258 then use the value of the REG_EQUAL note. */
3259 if ((set
= single_set (insn
))
3260 && (SET_DEST (set
) == comparison_value
))
3262 rtx note
= find_reg_note (insn
, REG_EQUAL
, NULL_RTX
);
3264 /* Only use the REG_EQUAL note if it is a constant.
3265 Other things, divide in particular, will cause
3266 problems later if we use them. */
3267 if (note
&& GET_CODE (XEXP (note
, 0)) != EXPR_LIST
3268 && CONSTANT_P (XEXP (note
, 0)))
3269 comparison_value
= XEXP (note
, 0);
3276 final_value
= approx_final_value (comparison_code
, comparison_value
,
3277 &unsigned_compare
, &compare_dir
);
3279 /* Save the calculated values describing this loop's bounds, in case
3280 precondition_loop_p will need them later. These values can not be
3281 recalculated inside precondition_loop_p because strength reduction
3282 optimizations may obscure the loop's structure. */
3284 loop_iteration_var
= iteration_var
;
3285 loop_initial_value
= initial_value
;
3286 loop_increment
= increment
;
3287 loop_final_value
= final_value
;
3291 if (loop_dump_stream
)
3292 fprintf (loop_dump_stream
,
3293 "Loop unrolling: Increment value can't be calculated.\n");
3296 else if (GET_CODE (increment
) != CONST_INT
)
3298 if (loop_dump_stream
)
3299 fprintf (loop_dump_stream
,
3300 "Loop unrolling: Increment value not constant.\n");
3303 else if (GET_CODE (initial_value
) != CONST_INT
)
3305 if (loop_dump_stream
)
3306 fprintf (loop_dump_stream
,
3307 "Loop unrolling: Initial value not constant.\n");
3310 else if (final_value
== 0)
3312 if (loop_dump_stream
)
3313 fprintf (loop_dump_stream
,
3314 "Loop unrolling: EQ comparison loop.\n");
3317 else if (GET_CODE (final_value
) != CONST_INT
)
3319 if (loop_dump_stream
)
3320 fprintf (loop_dump_stream
,
3321 "Loop unrolling: Final value not constant.\n");
3325 /* ?? Final value and initial value do not have to be constants.
3326 Only their difference has to be constant. When the iteration variable
3327 is an array address, the final value and initial value might both
3328 be addresses with the same base but different constant offsets.
3329 Final value must be invariant for this to work.
3331 To do this, need some way to find the values of registers which are
3334 /* Final_larger is 1 if final larger, 0 if they are equal, otherwise -1. */
3335 if (unsigned_compare
)
3337 = ((unsigned HOST_WIDE_INT
) INTVAL (final_value
)
3338 > (unsigned HOST_WIDE_INT
) INTVAL (initial_value
))
3339 - ((unsigned HOST_WIDE_INT
) INTVAL (final_value
)
3340 < (unsigned HOST_WIDE_INT
) INTVAL (initial_value
));
3342 final_larger
= (INTVAL (final_value
) > INTVAL (initial_value
))
3343 - (INTVAL (final_value
) < INTVAL (initial_value
));
3345 if (INTVAL (increment
) > 0)
3347 else if (INTVAL (increment
) == 0)
3352 /* There are 27 different cases: compare_dir = -1, 0, 1;
3353 final_larger = -1, 0, 1; increment_dir = -1, 0, 1.
3354 There are 4 normal cases, 4 reverse cases (where the iteration variable
3355 will overflow before the loop exits), 4 infinite loop cases, and 15
3356 immediate exit (0 or 1 iteration depending on loop type) cases.
3357 Only try to optimize the normal cases. */
3359 /* (compare_dir/final_larger/increment_dir)
3360 Normal cases: (0/-1/-1), (0/1/1), (-1/-1/-1), (1/1/1)
3361 Reverse cases: (0/-1/1), (0/1/-1), (-1/-1/1), (1/1/-1)
3362 Infinite loops: (0/-1/0), (0/1/0), (-1/-1/0), (1/1/0)
3363 Immediate exit: (0/0/X), (-1/0/X), (-1/1/X), (1/0/X), (1/-1/X) */
3365 /* ?? If the meaning of reverse loops (where the iteration variable
3366 will overflow before the loop exits) is undefined, then could
3367 eliminate all of these special checks, and just always assume
3368 the loops are normal/immediate/infinite. Note that this means
3369 the sign of increment_dir does not have to be known. Also,
3370 since it does not really hurt if immediate exit loops or infinite loops
3371 are optimized, then that case could be ignored also, and hence all
3372 loops can be optimized.
3374 According to ANSI Spec, the reverse loop case result is undefined,
3375 because the action on overflow is undefined.
3377 See also the special test for NE loops below. */
3379 if (final_larger
== increment_dir
&& final_larger
!= 0
3380 && (final_larger
== compare_dir
|| compare_dir
== 0))
3385 if (loop_dump_stream
)
3386 fprintf (loop_dump_stream
,
3387 "Loop unrolling: Not normal loop.\n");
3391 /* Calculate the number of iterations, final_value is only an approximation,
3392 so correct for that. Note that tempu and loop_n_iterations are
3393 unsigned, because they can be as large as 2^n - 1. */
3395 i
= INTVAL (increment
);
3397 tempu
= INTVAL (final_value
) - INTVAL (initial_value
);
3400 tempu
= INTVAL (initial_value
) - INTVAL (final_value
);
3406 /* For NE tests, make sure that the iteration variable won't miss the
3407 final value. If tempu mod i is not zero, then the iteration variable
3408 will overflow before the loop exits, and we can not calculate the
3409 number of iterations. */
3410 if (compare_dir
== 0 && (tempu
% i
) != 0)
3413 return tempu
/ i
+ ((tempu
% i
) != 0);
3416 /* Replace uses of split bivs with their split pseudo register. This is
3417 for original instructions which remain after loop unrolling without
3421 remap_split_bivs (x
)
3424 register enum rtx_code code
;
3431 code
= GET_CODE (x
);
3446 /* If non-reduced/final-value givs were split, then this would also
3447 have to remap those givs also. */
3449 if (REGNO (x
) < max_reg_before_loop
3450 && reg_iv_type
[REGNO (x
)] == BASIC_INDUCT
)
3451 return reg_biv_class
[REGNO (x
)]->biv
->src_reg
;
3454 fmt
= GET_RTX_FORMAT (code
);
3455 for (i
= GET_RTX_LENGTH (code
) - 1; i
>= 0; i
--)
3458 XEXP (x
, i
) = remap_split_bivs (XEXP (x
, i
));
3462 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
3463 XVECEXP (x
, i
, j
) = remap_split_bivs (XVECEXP (x
, i
, j
));